Datasets:

timaeus
/

pile-arxiv

Dataset card Data Studio Files Files and versions Community

Split (1)

train · 100k rows

text stringlengths 4 2.78M	meta dict
--- abstract: \| Die vorliegende Arbeit beschäftigt sich mit der BFV-Reduktion von Hamiltonschen Systemen mit erstklassigen Zwangsbedingungen im Rahmen der klassischen Hamiltonschen Mechanik und im Rahmen der Deformationsquantisierung. Besondere Aufmerksamkeit wird dabei Zwangsbedingungen zuteil, die als Nullfaser singulärer äquivarianter Impulsabbildungen entstehen. Es ist schon länger bekannt, da[ß]{} für Nullfasern regulärer äquivarianter Impulsabbildungen die in der theoretischen Physik gebräuchliche Methode der BFV-Reduktion zur Phasenraumreduktion nach Marsden/Weinstein äquivalent ist. In [@BHW] konnte gezeigt werden, da[ß]{} in dieser Situation die BFV-Reduktion sich auch im Rahmen der Deformationsquantisierung natürlich formulieren lä[ß]{}t und erfolgreich zur Konstruktion von Sternprodukten auf Marsden/Weinstein-Quotienten verwendet werden kann. Ein Hauptergebnis der vorliegenden Arbeit besteht in der Verallgemeinerung der Ergebnisse aus [@BHW] auf den Fall singulärer Impulsabbildungen, deren Komponenten 1.) das Verschwindungsideal der Zwangsfläche erzeugen und 2.) einen vollständigen Durchschnitt bilden. Die Argumentation von [@BHW] wird durch Gebrauch der Störungslemmata aus dem Anhang \[HPT\] systematisiert und vereinfacht. Zum Existenzbeweis von stetigen Homotopien und stetiger Fortsetzungsabbildung für die Koszulauflösung werden der Zerfällungssatz und der Fortsetzungssatz von Bierstone und Schwarz [@Schwarzbier] benutzt. Au[ß]{}erdem wird ein ’Jacobisches Kriterium’ für die Überprüfung von Bedingung 2.) angegeben. Basierend auf diesem Kriterium und Techniken aus [@AGJ] werden die Bedingungen 1.) und 2.) an einer Reihe von Beispielen getestet. Als Korollar erhält man den Beweis dafür, da[ß]{} es symplektisch stratifizierte Räume gibt, die keine Orbifaltigkeiten sind und dennoch eine stetige Deformationsquantisierung zulassen. Ferner wird (ähnlich zu [@Sevost]) eine konzeptionielle Erklärung dafür gegeben, warum im Fall vollständiger Durchschnitte das Problem der Quantisierung der BRST-Ladung eine so einfache Lösung hat. Bildet die Impulsabbildung eine erstklassige Zwangsbedingung, ist aber kein vollständiger Durchschnitt, dann ist es im allgemeinen nicht bekannt, wie entsprechende Quantenreduktionsresultate zu erzielen sind. Ein Hauptaugenmerk der Untersuchung wird es deshalb sein, in dieser Situation die klassische BFV-Reduktion besser zu verstehen – natürlich in der Hoffnung, Grundlagen für eine etwaige (Deformations-)Quantisierung zu liefern. Wir werden feststellen, da[ß]{} es zwei Gründe gibt, die Tate-Erzeuger (alias: Antigeister höheren Niveaus) notwendig machen: die Topologie der Zwangsfläche und die Singularitätentheorie der Impulsabbildung. Die Zahl der Tate-Erzeuger kann durch Übergang zu projektiven Tate-Erzeugern, also Vektorbündeln, verringert werden. Allerdings sorgt Halperins Starrheitssatz [@Halperin] dafür, da[ß]{} im wesentlichen alle Fälle, für die die Zwangsfläche kein lokal vollständiger Durchschnitt ist, zu unendlich vielen Tate-Erzeugern führen. Erzeugen die Komponenten einer Impulsabbildung einer linearen symplektischen Gruppenwirkung das Verschwindungsideal der Zwangsfläche, so kann man eine lokal endliche Tate-Auflösung finden. Diese besitzt nach dem Fortsetzungssatz und dem Zerfällungssatz von Bierstone und Schwarz stetige, kontrahierende Homotopien. Ausgehend von einer solchen Tate-Auflösung konstruieren wir, die klassische BFV-Konstruktion für vollständige Durchschnitte verallgemeinernd, eine graduierte superkommutative Algebra. Wir können zeigen, da[ß]{} diese graduierte Algebra auch im Vektorbündelfall eine graduierte Poissonklammer besitzt, die sogenannte Rothstein-Poissonklammer. Die Existenz einer solchen Poissonklammer war bereits von Rothstein [@Rothstein] für die einfachere Situation einer symplektischen Supermannigfaltigkeit bewiesen worden. Darüberhinaus werden wir sehen, da[ß]{} es auch im Vektorbündelfall eine BRST-Ladung gibt. Diese sieht im Fall von Impulsabbildungen etwas einfacher aus als für allgemeine erstklassige Zwangsbedingungen. Insgesamt wird also die klassische BFV-Konstruktion [@Stashbull] auf den Fall projektiver Tate-Erzeuger verallgemeinert, und als eine Homotopieäquivalenz in der additiven Kategorie der Fréchet-Räume interpretiert. bibliography: - 'mainpromo.bib' - 'stratbrst.bib' - 'neumaier.bib' --- \ \ \ \ \ \ \ \ \ (D 30) Vom Fachbereich Mathematik\ der Johann Wolfgang Goethe – Universität\ als Dissertation angenommen.\ Prodekan: Prof. Dr. Klaus Johannson\ Gutachter: Prof. Dr. Markus Pflaum (Frankfurt) und\ Prof. Dr. Martin Bordemann (Mulhouse, Frankreich)\ Datum der Disputation: Introduction ============ If we would adopt the terminology of the physicists the topic of this work would probably most succinctly be described as ‘BRST-deformation quantization’. Let us read this term backwards, and say a few words about ‘quantization’ first. In physics, by quantization one loosely means a rule how to assign to a classical mechanical system a quantum mechanical system. For example, the quantum mechanical analog of the two body gravitational system (the Kepler system) is the hydrogen atom, i.e., the system of a proton and an electron subjected to electromagnetic force. In classical mechanics the states of a system are the points of the phase space (or, more generally, probability distributions on the phase space) and the observable quantities (for short observables), are the (smooth) functions on the phase space. Every observable makes the phase space into a dynamical system, the dynamics is governed by the Hamiltonian equations of motion, which is a nonlinear first order ordinary differential equation. In quantum mechanics the states of a system are (ray equivalence classes of) vectors in a separable Hilbert space (or, more generally, density matrices). An observable is a self adjoint operator acting on this Hilbert space. The dynamics in quantum mechanics is determined by the Schrödinger equation, which is a linear partial differential equation. One wishes that the quantization procedure should be structurally clear, e.g., symmetry properties of the system should be preserved. This is already an interesting issue for the hydrogen atom (see [@Kepler]). At this point the consensus ends and the dissent begins. About thirty years ago somebody coined the sentence ‘First quantization is a mystery, but second quantization is a functor!’ [@ReedSimon]. It is the opinion of the author that the situation did not change too dramatically ever since. Up to now there is no completely satisfactory, mathematical well-defined and universally applicable theory of (‘first’) quantization. This difficulty is already notorious if one restricts to systems with finitely many degrees of freedom. The quantum mechanics textbook approach, which is usually called ‘canonical quantization’, is clearly satisfactory for the working physicist, but should be viewed merely as heuristics with a fairly limited domain of applicability rather than a true theory. More than fifty years after its invention by R. Feynman, the path integral approach, which is highly used in theoretical physics and which is certainly pretty universal, still deserves full mathematical justification. Due to the tremendous success of applying path intergral techniques to deep problems in pure mathematics in the last two decades the attitude of the mathematical world to the path integral changed from brusque rejection to some sort of (neurotic) admiration. An extensively studied, rigorous approach to the quantization problem is geometric quantization[^1], which goes back to works of B. Kostant and J.-M. Soriau (see for example the monograph [@Woodhouse]). A serious drawback of this method is that the set of observables which can be quantized is, in general, too small. One should also mention the operator algebraic approach of strict quantization which goes back to M. Rieffel (being advocated in the monograph [@Landsman]). The approach to quantization that will be pursued in this work is that of formal deformation quantization (the ‘formal’ will be dropped for convenience). In deformation quantization one deliberately neglects all functional-analytic and convergence questions, and uses the gained freedom to focus on the algebraic content of the quantization problem. In this way one obtains a (not completely satisfactory) mathematically well-defined, universally applicable theory of quantization of mechanical systems with finitely many degrees of freedom. Inspired by the symbol calculus of differential operators and the deformation theory of associative algebras [@Gerstenhaber] the founding fathers [@BFFLS77; @BFFLSI78; @BFFLSII78] of the theory proposed to view the quantization problem as a deformation problem for the algebra of smooth functions on the phase space seen as an associative algebra. Accordingly, the basic objects of study are so-called star products which are associative formal deformations of the algebra of smooth functions on a Poisson manifold given by formal series of bidifferential operators and which reproduce the original Poisson structure as a semiclassical limit. The classification of star products on symplectic manifolds has been achieved by De Wilde and Lecomte [@dWL] and Deligne [@Deligne] using sheaf theoretic methods and by Fedosov [@Fed] using global, geometric methods. In 1997 in a preprint (which is meanwhile published [@Kontsevich]) Kontsevich obtained a proof of his formality theorem, which says that the differential Hochschild cochain complex of the algebra of smooth functions on a smooth manifold is $L_\infty$-quasiisomorphic to its cohomology, i.e., the $\mathbbm Z$-graded Lie algebra of polyvector fields. For the flat space the $L_\infty$-quasiisomorphism is given by a remarkable, explicit formula. The formality theorem entails the classification of star products for Poisson manifolds. Ironically, the ‘magic’ of Kontsevich’s universal quantization formula appears to be natural in the light of topological quantum field theory. As it has been explained by Cattaneo and Felder [@CF1] the associativity of the Kontsevich star product can be understood as a Ward identity in the perturbative expansion of a certain topological sigma model. A more refined version of the formality theorem with a more conceptual method of proof has been found by Tamarkin [@Tamarkin] (see also the recent preprint [@DolgTamTsy]). Since the appearance of the Kontsevich formality theorem the theory underwent a noteworthy expansion/metamorphosis[^2], which makes it somehow difficult to give a clear picture of the current status. For a more detailed exposition and references we refer to [@Dito]. At this point we would like to emphasize that (even though some progress has been made in the case of orbifolds [@PflDQSO]), in general, the above mentioned methods do not work if the phase space has singularities. It is a delicate task to find out the precise number of publications[^3] which employ BRST-like methods, since they are part of the collective subconsciousness of particle physics. Roughly speaking, the idea attributed to Becchi, Rouet, Stora [@BRS1; @BRS2; @BRS3] and Tyutin [@Tyutin] was to tackle the problem of gauge invariance, which makes the scattering amplitudes in the perturbative expansion of a gauge theory formally infinite, by introducing new artifical fermionic field variables, the so-called ‘ghosts’, and to exploit a certain transformation (the BRST-symmetry) on the field variables, which leaves the action invariant and is of square zero. Still in a field theoretic spirit, in a series of papers Batalin, Fradkin and Vilkovisky [@BF; @BV1; @BV2; @BV3] formulated a quite general Hamiltonian version of the theory. In the second half of the eighties it has been realized by the workers in the field that for mechanical systems of finitely many degrees of freedom, this BFV-reduction method is an incarnation of phase space reduction. For example, in the seminal article of Kostant and Sternberg [@KostStern] it has been explained that the reduced algebra of a regular Marsden-Weinstein reduction at zero level of the moment map can be interpreted as the zeroth cohomology of an irreducible first class BRST-algebra. Moreover, it is shown in [@KostStern] that, in this case, the important problem of the quantization of the BRST-charge has a straight forward solution. As a consequence, in the regular case the techniques of BFV-reduction have been successfully applied to quantum phase space reduction in the context of geometric quantization program, see e.g., [@DET]. Inspired by these works in [@BHW] the authors have been able to show that the BFV-technique can be successfully employed to construct differential star products on phase spaces which are obtained by regular Marsden-Weinstein reduction with respect to a proper (locally) free Hamiltonian Lie group action (a similar result has been proved before by Fedosov [@Fed98] using other methods). Based on techniques akin to standard homological perturbation theory (for HPT see e.g., [@HuebKad] and the references therein) the reduced star product was given by a formula involving a series of differential operators, which are recursively determined. The key ingredients for the proof that this star product is in fact differential have been that 1.) in the regular case, there are explicit integral formulas for contracting homotopies of the Koszul-resolution, 2.) there is a multiplicative prolongation map. Moreover, in certain examples the above recursion can be solved, and the reduced star product can be given by an explicit formula. In this paper we will see that the techniques of [@BHW], suitably modified, apply also to certain cases of singular phase space reduction. More specifically, we will be concerned with a symplectic manifold $M$ with a Hamiltonian action of a compact, connected Lie group $G$ (with Lie algebra $\mathfrak g$), such that the moment map $J:M\to\mathfrak g^$ satisfies the following two conditions 1. \[gh1\]generating hypothesis:* the components of the moment map generate the ideal of the zero fibre $Z=J^{-1}(0)$, 2. \[ch\]complete intersection hypothesis: the Koszul complex on $J$ is acyclic. The generating hypothesis entails that the constraint set $Z$ is first class. In case the moment map satisfies condition \[ch\]., according to the physicist’s terminology, one also says that $J$ is an irreducible constraint. Hence, the above setup is a special case of what is called in the physics literature an irreducible first class constraint. It is well-known (and not difficult to see) that these conditions are fulfilled if $0\in\mathfrak g^$ is a regular value of $J$. However, if $0\in\mathfrak g^$ is a singular value of $J$ it is not at all a straight forward to check the above conditions (in particular condition 1.). In order to decide wether the generating hypothesis holds we will make use of the techniques developed in the seminal paper of Arms, Gotay and Jennings [@AGJ]. Accordingly, the generating hypothesis can be reduced to an elementary problem in algebraic geometry. In order to check the complete intersection hypothesis we have found a nice ‘Jacobian criterion’ (Theorem \[acyckrit\]), which applies if the generating hypothesis is true. In contrast to the regular case, in the singular case the continuous prolongation map and the continuous contracting homotopies of the Koszul resolution are not explicitly given, but appear as Deus ex machina as a consequence of the extension and the division theorem of [@Schwarzbier]. Moreover, in the singular case the prolongation map is not multiplicative. As a result, if condition \[gh1\].) and \[ch\].) are fulfilled, we observe that the Koszul resolution on the moment map is actually a contraction in the additive category of Fréchet spaces (cf. Appendix \[HPT\]). This contraction is the main ingredient for the construction of the classical BRST-algebra (cf. Section \[classred\]). This BRST-algebra is a differential graded Poisson algebra. At the same time it contracts to the Lie algebra cohomology of the $\mathfrak g$-module $\mathcal C^\infty(Z)$. Therefore, $\mathrm H^\bullet(\mathfrak g,\mathcal C^\infty(Z))$ aquires a $\mathbbm Z$-graded Poisson algebra structure. If the generating hypothesis is true it will be explained (cf. Proposition \[redcomparison\]) that the Poisson subalgebra $\mathrm H^0(\mathfrak g,\mathcal C^\infty(Z))=C^\infty(Z)^{\mathfrak g}$ is isomorphic to the Poisson algebra of functions of the stratified symplectic space $Z/G$ (cf. [@SjLerm]). Along the lines of Kostant/Sternberg [@KostStern] and [@BHW] we construct in Section \[brstalgconstr\] a deformation quantized version of the classical BRST-algebra, the main ingredients being a star product on $M$ and a quantum moment map. To this end we follow the observation made in [@BHW] that it is most comfortable to use the standard ordered (aka normal ordered) Clifford multiplication. Afterwards (cf. Section \[extalg\]) we give a conceptual explanation (similar to [@Sevost]) why this construction of a quantum BRST algebra works, i.e., why the ‘miracle’ of quantization of the BRST-charge happens. In the final Section \[brstcomp\] we show, using the homological perturbation theory techniques of appendix \[HPT\], that the quantum BRST-complex contracts to a Lie algebra cohomology complex of a certain deformed respresentation. Unfortunately, the space of invariants of this representation does, in general, not coincide with the topological free module generated by the space of invariants of the classical moment map. However, we are able to show that this problem can be circumvented if either we choose a strongly invariant star product, or the Lie group $G$ is a compact, connected and semisimple. In these cases we obtain continuous star products which deform the Poisson algebra of smooth functions on the singular reduced space $Z/G$. In order to illustrate our reduction methods we spend considerable effort in checking the generating and complete intersection hypothesis on a number of examples (all of them appeared in the literature before), see the list of examples at the end of Subsection \[hamgract\]. As a matter of principle, the generating hypothesis is much more easy to verify for torus actions than for nonabelian group actions (cf. Proposition \[nonpcond\] and Theorem \[torus\]). In particular, if a moment map of an $S^1$-action changes sign in a open neighborhood of $z$ in $M$ for every $z\in Z$, then the generating hypothesis and the complete intersection hypothesis are true. An example of such a moment map is provided by the so-called $(1,1,-1,-1)$-resonance of [@CushSjam], for which it is known that the reduced space is not an orbifold, but a genuine stratified symplectic space. Based on results of [@BrenPiVas] we have also found a nonabelian example, a so-called commuting variety, for which the generating and the complete intersection hypothesis hold. Note that the results of [@BrenPiVas] have been generalized to moment maps of the isotropy representations of symmetric spaces of maximal rank [@Pany]. This closes the first circle of ideas which will be discussed here. The second topic of this work addresses the question what happens if the complete intersection hypothesis is dropped, i.e., if, in the physicist’s terminology, $J$ is a first class reducible constraint. A prominent example, where this happens is given by the system of one particle of zero angular momentum in dimension $\ge 3$. The original suggestion of Batalin, Fradkin and Vilkovisky [@BF; @BV1; @BV2; @BV3] was to adjoin successively higher antighost variables, in order to kill the homology degree by degree. In this way one acquires a resolution of the $\mathcal C^\infty(M)$-module $\mathcal C^\infty(Z)$, which shares important structural properties with the Koszul complex. It was noted by J. Stasheff [@Stashbull] that this adjunction process already appeared in a work of J. Tate from 1957 [@Tate]. Following the customs of mathematical physics, we shall call these infinite resolutions Koszul-Tate resolutions. The essential structural properties of Koszul-Tate resolution seem to be that 1.) it is a (semifree) super-commutative differential graded algebra resolution of $\mathcal C^\infty(Z)$ and 2.) this algebra structure is part of a super-commutative, super-cocommutative $\mathcal C^\infty(M)$-bialgebra structure. The Koszul complex over a moment map is, due to the equivariance, naturally a $\mathfrak g$-module. It is not known to the author, whether a Koszul-Tate resolution may have an analogous feature. The next step in the BFV-construction is to adjoin ghost variables, which are dual to the antighosts, and to extend, by using dual pairing, the original Poisson bracket on the base manifold to acquire a $\mathbbm Z$-graded super Poisson bracket. The main theorem of classical BFV-reduction [@Stashbull] is that this Poisson algebra possesses a homological Hamiltonian vector field, the so-called BRST-differential, which is a perturbation of the Koszul-Tate differential. This differential depends on the Poisson structure of the base manifold and is not $\mathcal C^\infty(M)$-linear. The associated Hamilton function, is called the BRST-charge. As a consequence, one obtains a differential graded Poisson algebra, the so-called BRST-algebra, whose zeroth cohomology is isomorphic to the singular reduced Poisson algebra. The whole construction is done in a purely formal manner, i.e., it does not make use of any particular feature of the Koszul-Tate resolution. At this point we would like to mention that the Koszul-Tate-resolution is linked to the singularity theory of constraint surface (e.g. to the homotopy Lie algebra of the constraint). In particular, the number of antighosts (aka Tate generators) is bounded from below by certain homological invariants of the ring, the so-called deviations. There is an important theorem of Halperin [@Halperin] (which appears to be unmentioned in the physics literature), which essentially says that if any of the deviations vanish the variety in question has to be a complete intersection. This means that in essentially all cases which force us to introduce anitghosts of higher level the adjunction procedure will not terminate. In the following we will see that the BFV-construction can also be done in the vector bundle setting. Actually, this appears to be a new result already for the case of projective Koszul resolutions. In fact, the passage from free to projective Koszul-Tate resolutions (being defined in Section \[projkostate\]) can be used to reduce the number of generators. This is illustrated by the special case of closed submanifolds, see Section \[projkos\]. However, the lower bounds given by the deviations, being of local nature, govern the construction in the projective case as well. In order to construct the BRST-algebra the first nontrivial task is to find a meaningful way how to adjoin to the Tate generators (=antighosts) momenta (=ghosts). This will be explained in Section \[ggmf\]. Besides, even though the necessity of a ‘completion’ is noticed in [@Kimura] we have not been able to find in the literature a clear definition of the BRST-algebra as a space. The next step is to make the BRST-algebra into a $\mathbbm Z$-graded super-Poisson algebra. It has been shown by M.Rothstein [@Rothstein] that one can lift a symplectic Poisson structure from a manifold $M$ to a super-Poisson structure on a Gra[ß]{}mann-algebra bundle over $M$. From this explicit formula it is easy to guess a formula for the $\mathbbm Z$-graded super-Poisson bracket we are looking for. In order to prove that this Rothstein-bracket is actually a $\mathbbm Z$-graded super-Poisson bracket it is most comfortable to work with the Schouten-algebra over the BRST-algebra and to use the derived bracket construction of Koszul and Kosmann-Schwarzbach (this construction will be recalled in Section \[derbrsection\]). Actually, this computation, which is done in Section \[RothPBfin\], does also apply if the base manifold is a genuine Poisson manifold. Unfortunately, it is well-defined merely in the case of finitely many Tate generators. The point is that neither geometric nor the algebraic part of the bracket is a genuine bivector field (this problem is also present in the free version of the theory). The way out is to consider a slightly bigger Gerstenhaber algebra which we call the algebra of multiderivations, which contains the Schouten-algebra of polyvector fields as a subalgebra (see Section \[derbrsection\]) and to view the Rothstein-Poisson bracket as a derived bracket for this bigger algebra. In Section \[RothPBinfin\] we try to convince the reader that the Rothstein bracket is a super-Poisson bracket also in the infinitely generated case. Admittedly, the argument is still merely heuristics. Next, we show that the ’Existence of Charge’-theorem holds also in the projective setup (see Theorem \[chargeexists\]). This will be done by refining the argument of Kimura [@Kimura] (which has been a refinement of the argument of Stasheff [@Stashbull] by itself). In fact, we need not only approximations to the Tate-differential, but also approximations to the Rothstein bracket. Next we show that for linear Hamiltonian group actions we find a locally finite Koszul-Tate resolution together with continuous contracting homotopies (see Theorem \[existsfiniteTate\]). Using the homological perturbation theory techniques of appendix \[HPT\] we show that the BRST-algebra contracts to a complex which we call the vertical complex, which generalizes the Lie algebra cohomology complex of the $\mathfrak{g}$-module $\mathcal C^\infty(Z)$. In this way, the vertical cohomology aquires the structure of a $\mathbbm Z$-graded super-Poisson algebra. The question of uniqueness of the BRST-algebra (the vertical complex, respectively) has yet to be settled (we conjecture that it is unique up to $P_\infty$-quasiisomorphism in the sense of [@relform].) Moreover, in the special case of a projective Koszul resolution there are alternative constructions of a $\mathbbm Z$-graded Poisson structure on the vertical complex, one being the derived bracket of Theorem \[redderbr\] the other the $P_\infty$-structure of [@relform], the relations still have to be clarified. To the authors knowledge, up to now, there have been no notable attempts to find a quantization of the BRST-charge in the reducible (aka non-complete intersection) situation (a possible way could be to proceed along the lines of [@relform]). The main reason seems to be that the nature of the ghost and antighost variables is not well enough understood. In particular, in the reducible(=noncomplete intersection) case, the author does not know of any interesting example for which the vertical cohomology, has been computed. Another important open question is, whether the formality theorem for super-manifolds [@relform] generalizes to the above setup. ##### How to read this paper. We suggest that the reader takes first of all a look on the material of appendix \[HPT\] since the techniques explained there will be used at several places. If the reader is primarily interested in the complete intersection case, then he or she may skip all sections of chapter \[BFVchapter\], except the Sections \[kc\] and \[classred\]. This reading strategy might also be useful for the first time reading, if the reader is not familiar with the concepts discussed here. In Section \[classred\] and Chapter \[qbrst\] we use a slightly different super-Poisson structure than in the sections before. This causes the traditional (but somehow unaesthetic) factor of $2$ in the decomposition of the BRST differential ${\mbox{$\mathcal D$}}=2\partial+\delta$. For readers who dislike this factor of $2$ there is an easy way to get rid of it: halve the odd part of the BRST-Poisson structure (equation (\[oddpoiss\])) and replace $\theta=-\frac{1}{4}\sum_{abc} f_{ab}^c\xi^a\xi^b\xi_c+\sum_a J_a\xi^a$ with $-\frac{1}{2}\sum_{abc} f_{ab}^c\xi^a\xi^b\xi_c+\sum_a J_a\xi^a$. Analogously, cancel the $2$ in the exponent of formula (\[clmult\]) and replace ${\mbox{\boldmath $\theta$}}:= -\frac{1}{4}\sum_{a,b,c}\:f_{ab}^c\:\xi^a\xi^b\xi_c+ \sum_a{\mbox{\boldmath{$J$}}}_a\:\xi^a+\frac{\nu}{2}\sum_{a,b} f^b_{ab}\:\xi^a$ with $-\frac{1}{2}\sum_{a,b,c}\:f_{ab}^c\:\xi^a\xi^b\xi_c+ \sum_a{\mbox{\boldmath{$J$}}}_a\:\xi^a+\frac{\nu}{2}\sum_{a,b} f^b_{ab}\:\xi^a$. We assume that the reader is familiar with the basics of differential geometry and homological algebra. In particular, we tacitely assume some familiarity with the cohomology of Lie algebras and the Theorem of Serre and Swan. The letter $\mathbb K$ will stand for the field of real numbers $\mathbbm R$ or the field of complex numbers $\mathbbm C$. ##### Acknowledgments. First of all, I would like to express my gratitude to Markus Pflaum who gave me the opportunity (and freedom) to rethink these seemingly oldfashioned concepts from a different, and in my eyes very interesting perspective. Not least, he has been able to realize excellent working conditions and create a productive and open-minded athmosphere. Second, I would like to thank Martin Bordemann for maintaining the dialogue on mathematics (and somehow physics) also through times of stagnation and frustration. Many of the threads of this work originate from our ongoing discourse. I would like to thank Lucho Avramov and Srikanth Iyengar for encouraging my attempts to earth the BFV-construction to the solid grounds of commutative algebra. A good portion of the project has been financed by the Deutsche Forschungsgemeinschaft. I would also like to thank the Herrmann-Willkomm-Stiftung for financial support. It is taken for granted that the proof-reader is responsible for every error that has survived. Preparatory material ==================== Hamiltonian reduction --------------------- In this section we recall basic notions of classical Hamiltonian mechanics. We will exhibit a list of examples of (mostly singular) moment maps, which will serve as an illustration to the methods presented in the course this work. We will recall and compare two different notions of singular reduction: universal reduction and Dirac reduction. We will present the toolbox of Arms, Gotay and Jennings [@AGJ] in order to investigate the $\mathcal C^\infty$-algebraic geometry of the singular moment maps from our list. ### Hamiltonian group actions {#hamgract} In order to fix notation and sign conventions, let us recall the some basic notions from Hamiltonian mechanics. Even though a considerable part of this work applies only for symplectic manifolds, let us talk for the moment, more generally, about Poisson manifolds. A Poisson manifold is a manifold $M$, which carries a Poisson tensor $\Pi\in\Gamma^\infty(M,\wedge^2 TM)$. This Poisson tensor has to satisfy the following first order differential equation $$\begin{aligned} \label{master} [\Pi,\Pi]=0,\end{aligned}$$ where the bracket $[\:,\:]:\Gamma^\infty(M,\wedge^i TM)\times\Gamma^\infty(M,\wedge^j TM)\to \Gamma^\infty(M,\wedge^{i+j-1} TM)$ is the Schouten-Nijenhuis bracket. Using local coordinates $x^1,\dots, x^n$ we write $\Pi={\frac{1}{2}}\sum_{i,j}\Pi^{ij}\frac{\partial}{\partial x^i}\wedge\frac{\partial}{\partial x^j}$ and rewrite equation (\[master\]) as $$\begin{aligned} \sum_{m}\Big(\Pi^{im}\frac{\partial}{\partial x^m}\Pi^{jk}+\Pi^{jm}\frac{\partial}{\partial x^m}\Pi^{ki}+\Pi^{km}\frac{\partial}{\partial x^m}\Pi^{ij}\Big)=0.\end{aligned}$$ From the Poisson tensor we derive a Poisson bracket $\{\:,\:\}$. This is a bilinear antisymmetric map $\mathcal C^\infty(M)\times \mathcal C^\infty(M)\to \mathcal C^\infty(M)$. For the definition we again use the Schouten-Nijenhuis bracket $$\begin{aligned} \{f,g\}:=-[[\Pi,f],g],\end{aligned}$$ where $f,g \in \mathcal C^\infty(M)$. In local cordinates the Poisson bracket is given by $$\begin{aligned} \{f,g\}=\sum_{i<j}\Pi^{ij}\Big(\frac{\partial f}{\partial x^i}\frac{\partial g}{\partial x^j}-\frac{\partial g}{\partial x^i}\frac{\partial f}{\partial x^j}\Big).\end{aligned}$$ As an easy consequence of the Gerstenhaber algebra identities for $[\:,\:]$ the bracket $\{\:,\:\}$ satisfies Leibniz rule in every argument. Equation (\[master\]) implies that $\{\:,\:\}$ satisfies Jacobi identity and vice versa. Summerizing, we say that $(\mathcal C^\infty(M),\cdot,\{\:,\:\})$ satisfies the axioms of a Poisson algebra, i.e., it is a commutative algebra with a bilinear composition $\{\:,\:\}$ such that 1. $\{f,g\}=-\{g,f\}$, 2. $\{f,gh\}=\{f,g\}h+g\{f,h\}$, 3. $\{f,\{g,h\}\}+\{g,\{h,f\}\}+\{h,\{f,g\}\}=0$ for all $f,g,h\in \mathcal C^\infty(M)$. Given such a Poisson algebra the center of the Lie algebra $(\mathcal C^\infty(M),\{\:\})$ is called the space of Casimir functions, or, for short, of Casimirs. The space of Casimir functions will be denoted by $\mathrm H^0_\Pi(M)$. Using the Poisson structure one may associate to every function $f\in\mathcal C^\infty(M)$ a vector field $$\begin{aligned} X_f:=-[\Pi,f]=\{f,\:\},\end{aligned}$$ which is called the Hamiltonian vector field associated to $f$. If $X=X_f$ then $f$, which is unique up to a Casimir, is sometimes called a Hamiltonian function for $X$. An easy computation yields, that this assignment is actually a morphism of Lie algebras, i.e., we have $$\begin{aligned} [X_f,X_g]=X_{\{f,g\}}\end{aligned}$$ for all $f,g \in \mathcal C^\infty(M)$. Hamiltonian vector fields commute with the Poisson bivector field. Vector fields $X$ with this property, i.e., $[X,\Pi]=0$, are called Poisson vector fields. They are infinitisemal versions of Poisson diffeomorphism. A Poisson diffeomorphism $\Phi$ of a Poisson manifold is a diffeomorphism, whose pullback $\Phi^$ is an automorphism of the Lie algebra $(\mathcal C^\infty(M),\{\:,\:\})$. Symplectic manifolds* constitute an important subclass of the class of Poisson manifolds. They arise if the Poisson tensor is everywhere nondegenerate. The inverse of the Poisson tensor is the symplectic form $\omega\in\Omega^2(M)$, which is nondegenerate and is uniquely determined by the requirement $$\begin{aligned} i(X_f)\:\omega=df \quad\forall f\in\mathcal C^\infty(M).\end{aligned}$$ Equation (\[master\]) is equivalent to $\omega$ being closed. Since $\omega$ is nondegenerate the dimension of $M$ has to be even. Examples of noncompact symplectic manifolds are provided by the cotangent bundle $T^N$ of an arbitrary smooth manifold $N$. Using coordinates $q^1,\dots,q^n$ for $N$ there is a canonical frame $dq^1,\dots,dq^n$ for $T^N$. The bundle coordinates with respect to this frame provide the canonical coordinates $q^1,\dots,q^n,p_1,\dots,p_n$ for $T^N$. The symplectic form $\omega=-d\theta_0$ on $T^N$ is given (up to a sign) by the exterior differential of the canonical one form, which is given in canonical coordinates by $\theta_0=\sum_i p_i\:dq^i$. In many examples, the symplectic manifold $(M,\omega)$ carries in addition an almost complex structure $I\in\Gamma ^\infty(M,{\operatorname{End}}TM)$, $I^2=-{\operatorname{id}}$, which is compatible with $\omega$, i.e., $$\begin{aligned} g(X,Y):=\omega(X,IY)\end{aligned}$$ defines a Riemannian metric on $M$, where $X,Y\in\Gamma^\infty(M,TM)$. In this case $M$ is called an almost Kähler manifold. If the complex structure is integrable, i.e. $[I,I]_{RN}=0$, where $[\:,\:]_{RN}$ denotes the Richardson-Nijenhuis bracket, $M$ is called a Kähler manifold. In particular, if $M$ is Kähler, it follows from the Newlander-Nirenberg theorem, that $M$ is a complex manifold and $I$ coincides with the standard complex structure. Typical examples of Kähler manifolds are the affine space $(\mathbbm C^n,\omega)$ and the complex projective space $(\mathbbm C P^n,\omega_{FS})$. More precisely, the standard Kähler structure on $\mathbbm C^n$ is given by $\omega=\frac{{\sqrt{-1}}}{2}\sum_i dz_i\wedge d\bar{z}_i$. The Fubini-Study form $\omega_{FS}$ on $\mathbbm C P^n$ is given as follows. Let $\pi:\mathbbm C^{n+1}-\{0\}\to \mathbbm C P^n$ be the standard projection and let $Z=(Z_0,\dots,Z_n): \mathbbm C P^n\to\mathbbm C^{n+1}-\{0\}$ be a holomorphic lift for $\pi$, i.e. $\pi\:Z={\operatorname{id}}$. Then $$\begin{aligned} \omega_{FS}:=\frac{{\sqrt{-1}}}{2}\partial\bar{\partial}\ln\Big(\sum_{i=0}^n\ Z_i^2\Big)\end{aligned}$$ does not depend on the choice of $Z$ and defines an ($U(n+1)$-invariant) Kähler structure on $\mathbbm C P^n$. The most basic examples of Poisson manifolds which are not symplectic are the constant and the linear Poisson structures. Linear Poisson structures arise as follows. The phase space $M=\mathfrak h^$ is the linear dual space of a finite dimensional real Lie algebra $\mathfrak h$. The Lie bracket $[\:,\:]$ can be naturally interpreted as a linear bivector field $\Pi\in\Gamma^\infty(M,\wedge^2 TM)$. Equation (\[master\]) is equivalent to the Jacobi identity for $[\:,\:]$. More specifically, if we use a basis $e^1,\dots,e^n$ of $\mathfrak h^$ in order to write $x=\sum_{i=1}^n x_i e^i\in \mathfrak h^$, then $\Pi={\frac{1}{2}}\sum_{i,j,k=1}^n f_{ij}^k\:x_k\:\frac{\partial}{\partial x_i}\wedge\frac{\partial}{\partial x_j}$. Here, $f_{ij}^k$ denote the structure constants of the Lie algebra $\mathfrak h$. It is well known that the symplectic leaves of $\mathfrak h^$ are precisely the orbits under the coadjoint action of $H$ (a connected Lie goup with Lie algebra $\mathfrak h$) on $\mathfrak h^$ (see e.g. [@Vaisman section 3.1]). A (left) action* $\Phi$ of a Lie group $G$ on a manifold $M$ is a group homomorphism from $G$ to the diffeomorphism group of $M$. We will assume that this group action is effective, i.e. the aforementioned group homomorphism is injective. The action of an element $g$ of $G$ on a point $m\in M$ will be written $m\mapsto g.m:=\Phi_g(m)$. An action of the Lie group $G$ on $M$ induces a morphism of Lie algebras from the Lie algebra $\mathfrak g$ of $G$ into the Lie algebra of $\Gamma^\infty(M,TM)$ vector fields of $M$ $$\begin{aligned} \mathfrak g\to\Gamma^\infty(M,TM),\quad X\mapsto X_M(m):=\frac{d}{dt}_{\|t=0}\Phi_{\exp (-tX)}(m).\end{aligned}$$ The image $X_M$ of a vector $X\in\mathfrak g$ is called the fundamental vector field associated to X. If $M$ is a Poisson manifold, then we are interested in group actions preserving the Poisson structure, i.e. acting by Poisson diffeomorphisms. Here, most useful is the situation where the fundamental vector fields of the action are given by Hamiltonian vector fields, such that the Hamiltonian functions can be chosen in a coherent way. An action of a Lie group $G$ on a Poisson manifold $M$ is called a Hamiltonian action with an moment map $J: M\to \mathfrak g^$* if the following conditions are true: 1. $\xi_M=X_{J(\xi)}$ for all $\xi\in\mathfrak g$. 2. $J$ is equivariant (here $G$ acts on $\mathfrak g^$ via the coadjoint action). Here $J(\xi)$ denotes the function, which is obtained by evaluating $J$ on $\xi\in \mathfrak g$. In the same way, if we are given merely a $\mathfrak g$-action on $M$, i.e. a morphism of Lie algebras $\mathfrak g\to\Gamma^\infty(M,TM)$, we say that this action is Hamiltonian, if the two conditions above are true (clearly, in 2. we have to replace $G$-equivariance by $\mathfrak g$-equivariance). At this point, let us stipulate that, unless otherwise stated, for a Hamiltonian action as above the Lie group $G$ will be assumed to be connected* and, hence, there is no need to distinguish between $\mathfrak g$- and $G$-equivariance. For our purposes the most useful form of the equivariance property is $$\begin{aligned} \label{JLiemorph} \{J(\xi),J(\eta)\}=J([\xi,\eta])\quad\forall \xi,\eta\in \mathfrak g.\end{aligned}$$ This means, that $J$ may equally well be considered as a Lie algebra morphism from $\mathfrak g$ to the Lie algebra $(\mathcal C^\infty(M),\{\:,\:\})$. Note that a moment map is a Poisson morphism from $M$ to $\mathfrak g^$, where $\mathfrak g^$ is endowed with the linear Poisson structure. For a proof of this fact and of equation (\[JLiemorph\]) see e.g. [@Vaisman Proposition 7.30]. Obviously, we have the freedom to add to $J$ a Casimir function, which vanishes on $[\mathfrak g,\mathfrak g]$, or, in other words, a Lie algebra 1-cocyle from $Z^1\big(\mathfrak g,\mathrm H^0_{\Pi}(M)\big)$. An example of such a cocycle is given by the trace form $\chi(\xi):={\frac{1}{2}}\operatorname{trace}({\operatorname{ad}}(\xi))$. Given a Poisson action of a Lie group $G$ on a Poisson manifold $M$, there are some obstructions for the existence of a moment map (see e.g. [@CdSWein Part III, section 7]). For instance, if $M$ is symplectic it is sufficient that $\mathrm H^1(\mathfrak g,\mathbbm R)=0=\mathrm H^2(\mathfrak g,\mathbbm R)$, or that $M$ is compact and $\mathrm H^1_{dR}(M,\mathbbm R)=0$, for a moment map to exist. Another class of examples of Hamiltonian actions is provided by cotangent lifted actions, which arise as follows. Every diffeomorphism $\varphi:N\to N$ of the base manifold $N$ can be lifted to a diffeomorphism of the cotangent bundle $T^\varphi:T^N\to T^N$. In local coordinates coordinates this cotangent lift* of $\varphi$ is given by $T^\varphi:(q^1,\dots ,q^n,p_1,\dots p_n)\mapsto (Q^1,\dots ,Q^n,P_1,\dots ,P_n)$, where $Q^i$ is the $i$th component of $\varphi(q^1,\dots ,q^n,p_1,\dots p_n)$ and $p_i=\sum_{j=1}^n \frac{\partial Q^j}{\partial q^i} P_j$. Since such cotangent lifts (which are also known as point transformations) preserve the canonical one-form $(T^\varphi)^\theta_0= \theta_0$, they are in fact Poisson diffeomorphism. They obey the composition rule $T^(\varphi\circ\psi)=T^\psi\circ T^\varphi$. Accordingly, a left action $\Phi:G\times N\to N$ of the Lie group $G$ induces $g\mapsto T^\Phi_{g^{-1}}$ a left action of $G$ on $T^N$ by Poisson diffeomorphisms, the so-called cotangent lifted action. Note that according to our definition the Poisson bracket on the cotangent bundle writes $\{f,g\}=\sum_i \frac{\partial f}{\partial p_i} \frac{\partial g}{\partial q^i}- \frac{\partial f}{\partial q^i} \frac{\partial g}{\partial p_i}$, this is $-1$ times the physicist’s convention. Any cotangent lifted $G$-action is Hamiltonian. In the notation as above, a moment map $J:T^N\to\mathfrak g^$ is provided in canonical coordinates $({\mbox{\boldmath $q$}},{\mbox{\boldmath $p$}})=(q^1,\dots,q^n,p_1\dots,p_n)$ by the formula $$\begin{aligned} J(\xi)({\mbox{\boldmath $q$}},{\mbox{\boldmath $p$}})=-\sum_i p_i\:\xi^i_N({\mbox{\boldmath $q$}}),\end{aligned}$$ where $\xi_N=\sum_i \xi^i_N\frac{\partial}{\partial q^i}$ is the fundamental vector field for the action $\xi\in\mathfrak g$ on $N$. See [@AM78 p.282–283]. \ A symmetry can be used to reduce the number of degrees of freedom of the phase space by taking only “gauge invariant quantities” into account. In Hamiltonian mechanics this idea involves a two step procedure, which is illustrated by the following reduction theorem, attributed to Marsden and Weinstein [@MW] and Meyer [@Meyer]. Let $M$ be a symplectic manifold with a proper free Hamiltonian action of a Lie group $G$ with moment map $J:M\to \mathfrak g^$. Let $\mu\in \mathfrak g^$ be a regular value of $J$ and $G_\mu$ the isotropy group of $\mu$. Then the reduced space $M_\mu:=J^{-1}(\mu)/G_\mu$ is a symplectic manifold with symplectic form $\omega_\mu$, which is uniquely determined by the requirement $\pi_\mu^\omega_\mu=i_\mu^\omega$. Here $\pi_\mu$ and $i_\mu$ $$\begin{aligned} \xymatrix{ &\qquad J^{-1}(\mu)\ar[dl]_{\pi_\mu}\ar@{^{(}->}[dr]^{i_\mu}\qquad &\\ M_\mu & &M}\end{aligned}$$ are the obvious projection and injection, repsectively. See e.g. [@AM78 p.299–300]. \ There are numerous generalizations and versions of this theorem. For instance, there is a generalization to Poisson manifolds [@Vaisman Theorem 7.31]. Instead of taking $\mu$ one may take the preimage of a whole coadjoint orbit $J^{-1}(\mathcal O_\mu)$ and divide out by the action of $G$. This approach is called orbit reduction, which is a special case of coisotropic reduction. Regular orbit and point reduction essentially coincide [@symptech chapter II, section 26]. If one drops the freeness assumption, the reduced space will be a symplectic orbifold. For an intelligible treatment of this issue the reader may consult [@BatesCush]. In the following we shall address the case, when the regularity assumption is dropped. Even though the set of regular values is generic, the singular values $\mu \in \mathfrak g^$ are particularly interesting. One of the reasons is that the points in the $\mu$-fibre $J^{-1}(\mu)$ tend to have larger isotropy groups. Since we will use the BRST-method, we are forced to treat merely the case of reduction at $0\in\mathfrak g^$. If $G$ is abelian this does not cause a restriction at all, since we are free to add a constant $\mu\in\mathfrak g^$ to $J$. For nonabelian $G$ one has to use the shifting trick, i.e. adjoin a coadjoint orbit $\mathcal O_\mu$. For the shifting trick in singular reduction we refer to [@CushSjam]. We close this section by giving a list of examples of moment maps, which we will frequently refer to in the course of this work. Note that for all of these examples, except Examples \[irrflow\] and \[linPoiss\], zero is a singular value. Furthermore, in all examples, except Example \[standardex\], the group $G$ is connected. The elementary Examples \[harmosc\],\[allwrong\] and \[irrflow\] and the physically interesting Example \[drehimpuls\] will serve to illustrate the limitations of the methods presented in this article. Example \[standardex\] will be important in connection with the normal form theorem, Theorem \[normcoord\]. The remaining examples will turn out as instances, where our ultimate goal, i.e., a quantum phase space reduction will be achieved. \[harmosc\] Consider the $S^1=\mathbbm R/2\pi \mathbbm Z$-action on $\mathbbm C$ given by $(\vartheta,z)\mapsto {\operatorname{e}}^{{\sqrt{-1}}\vartheta} z$. The moment map for this action is $J(z)={\frac{1}{2}}\|z\|^2$, which has $0$ as the only critical value. \[allwrong\] Here the phase space is $T^\mathbbm R=\mathbbm R^2$. The $\mathbbm R$-action which is generated by the kinetic energy of the free particle $J(q,p)={\frac{1}{2}}p^2$ is given by $(q,p)\mapsto (q+tp,p)$. \[irrflow\] Let $M:=T^\mathbbm T^2$ be the cotangent bundle of the 2-torus $\mathbbm T^2=(\mathbbm R/{2\pi \mathbbm Z})^2$. An element $t\in\mathbbm R$ acts on $\mathbbm T^2$ by the formula $(\vartheta_1,\vartheta_2)\mapsto (\vartheta_1+2\pi t,\vartheta_1+\alpha 2\pi t)$, where the slope parameter $\alpha\in \mathbbm R\backslash\mathbbm Q$ is irrational. Every orbit of this $\mathbbm R$-action is dense. A moment map for the cotangent lifted $\mathbbm R$-action on $T^\mathbbm T^2$ is given by $J( \vartheta_1\vartheta_2,p_1,p_2)=2\pi(p_1+\alpha p_2)$. \[standardex\] Let $M:=\mathbbm C^n$ with its standard Kähler structure. Let $G$ be any subgroup of the unitary group $U(n)$. We identify the Lie algebra $\mathfrak u(n)$ with the space of skew hermitian matrices $\xi=(\xi_{ij})$, $\bar{\xi}_{ij}=-\xi_{ji}$. The moment map for the action is given by the formula $$\begin{aligned} J(\xi)=\frac{{\sqrt{-1}}}{2}\sum_{ij}\xi_{ij}\bar z_i z_j,\end{aligned}$$ for $\xi=(\xi_{ij})\in \mathfrak g\subset \mathfrak u(n)$. Using real coordinates $z_i=x_i+{\sqrt{-1}}y_i$ we obtain $$\begin{aligned} J(\xi)=-\frac{1}{2}\sum_{i,j} \big(A_{ij}(x_iy_j-x_jy_i)+ S_{ij}(x_iy_j+x_jy_i)\big),\end{aligned}$$ where $A_{ij}=-A_{ji}$ and $S_{ij}=S_{ji}$ are the real and imaginary part of $\xi_{ij}$, respectively. In all cases, except the trivial case when $G$ is discrete, zero is a singular value of $J$. \[drehimpuls\] Let $M:=T^\mathbbm R^n$ be the cotangent bundle of the euclidean space $\mathbbm R^n$ together with the cotangent lift of the obvious $SO(n,\mathbbm R)$-action on $\mathbbm R^n$. Let us write the canonical ccordinates ${\mbox{\boldmath $q$}}=(q^1,\dots,q^n)^t$ and ${\mbox{\boldmath $p$}}=(p_1,\dots,p_n)^t$ in column vector form. The cotangent lifted action is just the diagonal $SO(n)$-action on $T^\mathbbm R^n=\mathbbm R^n\times\mathbbm R^n$ ,i.e., an orthogonal matrix $O\in SO(n)$ acts according to $({\mbox{\boldmath $q$}},{\mbox{\boldmath $p$}})\mapsto (O{\mbox{\boldmath $q$}},O{\mbox{\boldmath $p$}})$. Using the euclidian scalar product $<,>$ on $\mathbbm R^n$ we identify $\wedge^2\mathbbm R^n$ and $\mathfrak{so}(n)$ by letting $u\wedge v\in\wedge^2\mathbbm R^n$ act on $w\in \mathbbm R^n$ according to $(u\wedge v)w:=<u,w>v-<v,w>u$. Using the invariant, positive definite scalar product $(\:,\:)$ on $\mathfrak{so}(n)$, $(A,B):=-{\frac{1}{2}}\operatorname{tr}(AB)$, we identify $\mathfrak{so}(n)$ and $\mathfrak{so}(n)^$. With these identifications understood the moment map, the so-called ‘angular momentum’, writes $$\begin{aligned} J:T^\mathbbm R^n\to \mathfrak{so}(n)^\cong \wedge^2\mathbbm R^n,\quad({\mbox{\boldmath $q$}},{\mbox{\boldmath $p$}})\mapsto{\mbox{\boldmath $q$}}\wedge{\mbox{\boldmath $p$}}.\end{aligned}$$ Of course, by identifying the cotangent lifted action on $T^\mathbbm R^n$ with the compexified $SO(n)$-action on $\mathbbm C^n$ this example can be viewed as a special case of the preceding Example \[standardex\]. For completeness, let us mention the physically important special case $n=3$. Here, accidentally $\mathbbm R^3$ and $\mathfrak{so}(3)$ are isomorphic vector spaces. More specifically, the linear isomorphism $$\begin{aligned} (v_1,v_2,v_3)\mapsto \left(\begin{array}{ccc} 0&-v_3&v_2\\ v_3&0&-v_1\\ -v_2&v_1&0 \end{array} \right)\end{aligned}$$ is in fact an isomorphism of metric Lie algebras $\left(\mathfrak{so}(3),[\:,\:],(\:,\:)\right)$ and $\left(\mathbbm R^3,\times,<,>\right)$, where $\times$ is the well known vector product. The angular momentum now is the $\mathbbm R^3$-valued function $J({\mbox{\boldmath $q$}},{\mbox{\boldmath $p$}})={\mbox{\boldmath $q$}}\times{\mbox{\boldmath $p$}}$. \[mpart\] We consider the system of $m$ particles in $\mathbbm R^2$ with zero total angular momentum. More precisely, the phase space is $M:=(T^\mathbbm R^2)^m$ and we let $SO(2,\mathbbm R)\cong S^1$ act on it by lifting the diagonal action, i.e., $$\begin{aligned} SO(2)\times M&\to& M\\ (g,({\mbox{\boldmath $q$}}_1,{\mbox{\boldmath $p$}}^1,\dots,{\mbox{\boldmath $q$}}_m,{\mbox{\boldmath $p$}}^m))&\mapsto& (g{\mbox{\boldmath $q$}}_1,g{\mbox{\boldmath $p$}}^1,\dots,g{\mbox{\boldmath $q$}}_m,g{\mbox{\boldmath $p$}}^m),\end{aligned}$$ where ${\mbox{\boldmath $q$}}_i=(q^1_i,q^2_i)^t$ and ${\mbox{\boldmath $p$}}^i=(p_1^i,p_2^i)^t$ for $i=1,\dots,m$. The moment map $J:M\to\mathfrak{so}(2)^=\mathbbm R$ is given by $J({\mbox{\boldmath $q$}},{\mbox{\boldmath $p$}})=\sum_{i=1}^m q^1_ip_2^i-q^2_ip_1^i$. \[lemon\] Let $S^1=\mathbbm R/2\pi\mathbbm Z$ act on the product $M:=\mathbbm C P^1\times\mathbbm C P^1$ according to the formula $$\begin{aligned} \left((z_1:z_2),(z_3:z_4)\right)\mapsto\left(({\operatorname{e}}^{{\sqrt{-1}}\vartheta}z_1:{\operatorname{e}}^{-{\sqrt{-1}}\vartheta}z_2),({\operatorname{e}}^{{\sqrt{-1}}\vartheta}z_3:{\operatorname{e}}^{-{\sqrt{-1}}\vartheta}z_4)\right).\end{aligned}$$ The fix points of this action are $((1:0),(1:0))$, $((0:1),(0:1))$, $((1:0),(0:1))$ and $((0:1),(1:0))$. A moment map for this action is $$\begin{aligned} J\big((z_1:z_2),(z_3:z_4)\big)={\frac{1}{2}}\left(\frac{\|z_1\|^2-\|z_2\|^2}{\|z_1\|^2+\|z_2\|^2}+\frac{\|z_3\|^2-\|z_4\|^2}{\|z_3\|^2+\|z_4\|^2}\right).\end{aligned}$$ The critical set of $J$ constists of the points $-1,0$ and $1$. \[stratres\]Consider the $S^1$-action on $\mathbbm C^4$, endowed with the standard Kähler structure, given by $(z_1,z_2,z_3,z_4)\mapsto ({\operatorname{e}}^{{\sqrt{-1}}\vartheta}z_1,{\operatorname{e}}^{{\sqrt{-1}}\vartheta}z_2, {\operatorname{e}}^{-{\sqrt{-1}}\vartheta}z_3,{\operatorname{e}}^{-{\sqrt{-1}}\vartheta}z_4)$. The moment map for the action is $$J(z_1,z_2,z_3,z_4)=\frac{1}{2}(\|z_1\|^2+\|z_2\|^2-\|z_3\|^2-\|z_4\|^2).$$ \[tzwo\] The action of $\mathbbm T^2=(\mathbbm R/2\pi \mathbbm Z)^2$ on $\mathbbm C^4$ is given by, $$\begin{aligned} ((\vartheta_1,\vartheta_2),(z_1,z_2,z_3,z_4)) \mapsto ({\operatorname{e}}^{{\sqrt{-1}}(\alpha \vartheta_1+\beta\vartheta_2)}z_1, {\operatorname{e}}^{-{\sqrt{-1}}\vartheta_2}z_2,{\operatorname{e}}^{{\sqrt{-1}}\vartheta_1}z_3, {\operatorname{e}}^{-{\sqrt{-1}}\vartheta_2}z_4)\end{aligned}$$ for $\alpha,\beta \in \mathbbm Z$. A moment map for the action is $$\begin{aligned} J:\mathbbm C^4\to \mathbbm R^2,\quad J(z_1,z_2,z_3,z_4):= -\frac {1}{2}(\alpha\|z_1\|^2+\|z_3\|^2,\beta\|z_1\|^2-\|z_2\|^2+\|z_4\|^2). \end{aligned}$$ \[commvar\] Let $S$ the space of symmetric $n\times n$-matrices with real entries. We let $SO(n)$ act on $S$ by conjugation and we lift this action to an action of $SO(n)$ on the cotangent bundle $T^S=S\times S$. This action is Hamiltonian with the moment map $$\begin{aligned} J:S\times S&\to&\wedge^2\mathbbm R^n=\mathfrak{so}(n)^\\ (Q,P)&\mapsto&[Q,P].\end{aligned}$$ \[Linear Poisson structures\]\[linPoiss\]Consider $M:=\mathfrak h^$ the dual space of an $n$-dimensional $\mathbbm R$-Lie algebra $\mathfrak h$ with Poisson linear structure, i.e., $$\begin{aligned} \Pi(x_1,\dots,x_n)={\frac{1}{2}}\sum_{i,j,k} f_{ij}^k\:x_k\:{\frac{\partial }{\partial x_{i}}}\wedge{\frac{\partial }{\partial x_{j}}},\end{aligned}$$ where $x_1,\dots,x_n$ are the linear coordinates with respect to a chosen basis $e^1,\dots,e^n$ of $\mathfrak h^$ and $f_{ij}^k=<e^k,[e_i,e_j]>$ are the structure constants with respect to the dual basis $e_1,\dots,e_n$ of $\mathfrak h$. Given a Lie subalgebra $\mathfrak g$ of $\mathfrak h$ we obtain an $\mathfrak h$-action on $M$ by restricting the coadjoint $\mathfrak h$-action on $M$ to $\mathfrak g$. For any connected Lie group $G$ with Lie algebra $\mathfrak g$ this $\mathfrak g$-action integrates to a Hamiltonian $G$-action on $M$. The moment map for this action is given by restriction $J:M=\mathfrak h^\to \mathfrak g^$, $\alpha\mapsto \alpha_{\|\mathfrak g}$. ### Universal reduction Contrary to the regular case, in the case when zero is a singular value of the moment map, there are several possible approaches to phase space reduction (see e.g. [@Dirac; @SnW; @AGJ; @Henneaux]). In the beginning of the nineties, the attempts to find the most ‘correct’ reduction procedure lead to the notion of universal reduction of Arms, Cushman and Gotay [@ACG], who noticed that $\mathcal C^\infty(M)^G/I_Z^G$ is a good candidate for the reduced Poisson algebra. Here we shall present a (slightly different) version $\mathcal C^\infty(M)^{\mathfrak g}/I_Z^{\mathfrak g}$ of this algebra, where we have replaced $G$-invariance by $\mathfrak g$-invariance. Due to our overall assumption that $G$ is connected, this will make no difference for us. Universal reduction does not make the other aproaches – in particular Dirac reduction – obsolete, but serves rather as a benchmark. \[geomRed\] The space of invariant functions $$\mathcal C^\infty(M)^{\mathfrak g}\subset\mathcal C^\infty(M)$$ is a Poisson subalgebra, which contains the ideal of invariant functions vanishing on $Z$ $$I^\mathfrak g_Z:=\{f\in\mathcal C^\infty(M)^{\mathfrak g}\:\|\: f(z)=0\:\forall z\in Z\}$$ as a Poisson ideal. Hence the quotient $\mathcal C^\infty(M)^{\mathfrak g}/I_Z^\mathfrak g$ is in a natural way a Poisson algebra. More precisely, if $f$ and $g\in \mathcal C^\infty(M)^{\mathfrak g}$ are representatives of $[f]$ and $[g] \in \mathcal C^\infty(M)^{\mathfrak g}/I_Z^\mathfrak g$, respectively, then the Poisson bracket is defined as $$\begin{aligned} \{[f],[g]\}:=\big[\{f,g\}\big].\end{aligned}$$ If $M$ is a symplectic manifold and the action of $G$ on M is proper, then this Poisson structure is nondegenerate, i.e., there are no nontrivial Casimirs. Since $\xi\in\mathfrak g$ acts on $\mathcal C^\infty(M)$ via $\{J(\xi),\:\}$, the claim that $\mathcal C^\infty(M)^\mathfrak g\subset \mathcal C^\infty(M)$ is a Poisson subalgebra follows immediately from the Jacobi identity and the Leibniz rule for $\{\:,\:\}$. Let $f$ be in $\mathcal C^\infty(M)^\mathfrak g$. This means, that $$\begin{aligned} \{J(X),f\}=0\quad \forall X\in\mathfrak g.\end{aligned}$$ Hence, for all $X\in\mathfrak g$ the function $J(X)$ is constant along the integral curve of $X_f$. In particular, we conclude that the set $Z=J^{-1}(0)$ is stable under the local flow of $X_f$. Now, let $z\in Z$ and $\gamma:]-\epsilon,\epsilon[\to Z\subset M$ be an intergral curve for $X_f$ such that $\gamma(0)=z$. It follows that for every $g\in I_Z$ we have $$\begin{aligned} 0=\frac{d}{dt}_{\|t=0}\:g(\gamma(t))=\{f,g\}(z).\end{aligned}$$ This implies that $I_Z^\mathfrak g\subset \mathcal C^\infty(M)^\mathfrak g$ is a Poisson ideal. A proof for the nondegeneracy of the reduced Poisson structure can be found in [@ACG section 3]. \ The universal reduction may be interpreted geometrically as follows. We have the commutative diagram $$\begin{aligned} \xymatrix{Z=J^{-1}(0)\:\ar@{^{(}->}[r]\ar[d]&M\ar[d]\\ M_{{\operatorname{red}}}=Z/G\:\ar@{^{(}->}[r]^{\quad i_0}&M/G.}\end{aligned}$$ Since $0\in \mathfrak g^$ may be a singular value of $J$, the zero fibre $Z$ is no longer a submanifold, but, let us say at least, a closed subset of $M$. The (not necessarily free) action of $G$ on $Z$ and on $M$ yield quotient spaces which are singular. The map $i_0$ associates to every orbit in $Z$ the corresponding orbit in $M$. It may be shown to be a homoemorphism onto its image. If the action of $G$ on $M$ is not too pathological (e.g. if it is proper), then the space of invariant functions $\mathcal C^\infty(M)^G=\mathcal C^\infty(M)^\mathfrak g$ is a good substitute for the space of functions on $M/G$. The ideal of invariant functions $I^\mathfrak g_Z$ vanishing on $Z$ then may be thought of as the defining ideal of the “subvariety” $M_{{\operatorname{red}}}$ of $M/G$. The above theorem says that $i_0$ is an embedding of the Poisson “variety” $M_{{\operatorname{red}}}$ into the Poisson “variety” $M/G$. Note, that there are some subtleties in case the image of $i_0$ is not closed. It is easy to give examples of nonproper Hamiltonian group actions, for which the space of invariant functions is too small to give a meaningful description of the quotient spaces $M/G$ and $Z/G$, respectively. For instance, one may consider the cotangent lift of an irrational torus, Example \[irrflow\]. Here for any value of the moment map the reduced space is $M_{{\operatorname{red}}}=(\mathbbm T^2/\mathbbm R)\times \mathbbm R$ and, since the orbit of the $\mathbbm R$-action is dense, the reduced algebra is identified with $\mathcal C^\infty(\mathbbm R)$ with the trivial Poisson structure. The great advantage of universal reduction is that it is always applicable and gives sensible results for proper Hamiltonian group actions. For symplectic manifolds with a Hamiltonian action of a compact Lie group the nature of the singular geometry of the reduced space $M_{{\operatorname{red}}}$ has been clarified by the important singular reduction theorem of Sjamaar and Lerman [@SjLerm]. \[SL\] Let $(M,\omega)$ be a symplectic manifold and $G$ be a compact Lie group acting on $M$ in a Hamiltonian fashion with moment map $J:M\to \mathfrak g^$ and let $Z=J^{-1}(0)$. Then for every subgroup $H\subset G$ the intersection $M_{(H)}\cap Z$ and the quotient space $$\begin{aligned} (M_{{\operatorname{red}}})_{(H)}:=(M_{(H)}\cap Z)/G,\end{aligned}$$ are manifolds. Here $M_{(H)}$ is the set of points in $M$ whose isotropy group is conjugate to $H$. There exists a unique symplectic form $\omega_{(H)}$ on $(M_{{\operatorname{red}}})_{(H)}$ such that the pullback of $\omega_{(H)}$ to $M_{(H)}\cap Z$ coincides with the restriction of $\omega$. The disjoint union over the conjugacy classes $(H)$ of subgroups $H\subset G$ $$\begin{aligned} M_{{\operatorname{red}}}=\coprod (M_{{\operatorname{red}}})_{(H)}\end{aligned}$$ is in fact a symplectic stratification. Here the set of conjugacy classes is understood to be ordered by reverse subconjugacy. The Poisson algebra of smooth functions on $M_{{\operatorname{red}}}$ is isomorphic to the Poisson algebra obtained by universal reduction $\mathcal C^\infty(M)^G/I_Z^G$. For a detailed discussion of symplectic stratifications we refer to [@SjLerm; @LerMontSj] and the monographs [@PflaumHab; @OrtRat]. Note that in the orbit type decompositions as above we allow the pieces to have components of different dimension. Furthermore, let us mention that the symplectic pieces $(M_{{\operatorname{red}}})_{(H)}$ can also be obtained by regular Marsden-Weinstein reduction with respect to the Hamiltonian action of a, in general smaller, Lie group. Theorem \[SL\] has been generalized to singular orbit reduction of proper Hamiltonian Lie group actions [@BatesLerman]. Let us look at our list of examples in the light of Thereom \[SL\]. In the case of one particle of angular momentum zero in dimension $n\ge 2$, Example \[drehimpuls\], it is well known that the reduced phase space is the symplectic orbifold $\mathbbm R^2/\mathbbm Z_2$, where $\mathbbm Z_2$ acts on $\mathbbm R^2$ by $(x_1,x_2)\mapsto (-x_1,-x_2)$. The first systematic treatment of this example, appears to be [@BosGotay]. For the commuting variety, Example \[commvar\], the reduced space has been identified in [@LerMontSj] as the symplectic orbifold $\mathbbm R^n\times \mathbbm R^n/S_n$, where the symmetric group $S_n$ is understood to act diagonally. Example \[lemon\] (the ‘lemon’) has been discussed at length in [@CushSjam]. Accordingly, the reduced spaces at the singular values $\pm 1$ are points. The reduced spaces for regular values of $J$ in the open intervals $]-1,0[$ and $]0,1[$ are $\mathbbm C P^1$. Finally, the reduced space at singular value $0$ (being homeomorphic to $\mathbbm C P^1$) is in fact a symplectic orbifold, which can be pictured as two copies of a quadratic half-cone in $\mathbbm R^3$ being glued together (whence the name). An example of a singular momentum map whose reduced space is not an orbifold is provided by the $(1,1,-1,-1)$-resonance, Example \[stratres\], which has been discussed in [@CushSjam Example 2.4]. As a result the reduced space is a real cone over the fibred product $S^3\times_{S^1}S^3$, where $S^1$ acts on the first half $S^3\subset \mathbbm C^2$ by $(z_1,z_2)\mapsto ({\operatorname{e}}^{{\sqrt{-1}}\vartheta}z_1, {\operatorname{e}}^{{\sqrt{-1}}\vartheta}z_2)$ and on the second half by $(z_3,z_4)\mapsto ({\operatorname{e}}^{-{\sqrt{-1}}\vartheta}z_3, {\operatorname{e}}^{-{\sqrt{-1}}\vartheta}z_4)$. In [@CushSjam] it has been argued that this reduced space is not a rational homology manifold, and hence no orbifold. Even though we will make no use of it, let us briefly explain how invariant theory can be utilized to describe the geometry of singular reduced spaces. Here, we have to restrict to the case of a linear Hamiltonian action of the compact (for the moment, not necessarily connected) Lie group $G$ on a symplectic vector space $M=\mathbbm R^{2n}$. This is not a severe restriction since, due to a theorem of Gotay and Tuynman [@GotayTuynman], every symplectic Hamiltonian $G$-manifold with finitely generated homology can be equivariantly obtained by phase space reduction from flat space. The algebra of $G$-invariant real polynomials on $M$ is finitely generated. A system $\sigma_1,\dots,\sigma_k$ generators for this algebra is called a real Hilbert basis. According to [@Schwarz] $\sigma_1,\dots,\sigma_k$ generate the algebra of smooth invariant functions $\mathcal C^\infty(M)^G$. The Hilbert map $$\begin{aligned} \sigma=(\sigma_1,\dots,\sigma_k):M=\mathbbm R^{2n}\to \mathbbm R^k,\quad m\mapsto (\sigma_1(m),\dots,\sigma_k(m))\end{aligned}$$ seperates $G$-orbits, and hence gives rise to an injective map $\bar{\sigma}:M/G\to \mathbbm R^k$. Being the image of a real polynomial map the image of $\sigma$ is a semi-algebraic set as a consequence of the Tarski-Seidenberg theorem. For a comparison of the associated Whitney stratification and the orbit type stratification see [@Bierstone75; @PflaumHab]. Furthermore, it can be shown that the map $\sigma$ is proper and the pullback $\sigma^:\mathcal C^\infty(\mathbbm R^k)\to \mathcal C^\infty(M)$ is a split surjective map of Fréchet spaces [@Mather]. In many cases (but not all, see [@Egilsson]) there is a Poisson structure on $\mathbbm R^k$, such that $\sigma^$ is a Poisson algebra morphism. Real Hilbert bases are known for essentially all linear examples from our list. In [@Egilsson section 5] we find a formula for the real Hilbert basis of a general linear Hamiltonian circle action. For a real Hilbert basis for Example \[tzwo\] see [@AGJ Example 7.7], and for a real Hilbert basis for Example \[commvar\] see [@LerMontSj p.145]. Most favorable is the situation, when the algebra of invariant polynomials is generated by quadratic polynomials. In this case the Hilbert map itself is a moment map of a Hamiltonian action of a certain Lie subgroup $H$ of $Sp(n,\mathbbm R)$. If, moreover, $G$ and $H$ form a reductive dual pair. i.e., $G$ and $H$ are reductive Lie subgroups of $Sp(n,\mathbbm R)$ which centralize each other in $Sp(n,\mathbbm R)$, then the orbit-reduced spaces for the $G$-action are bijectively mapped via $\bar{\sigma}$ onto closures of coadjoint orbits in $\mathfrak h^$. This map is compatible with the stratifications and Poisson structures in an appropriate sense (for details see [@LerMontSj Theorem 4.4]). In particular, if $H$ is semisimple, then the reduced space at zero $M_{{\operatorname{red}}}=J^{-1}(0)/G$ is bijectively mapped onto the closure of a nilpotent coadjoint orbit in $\mathfrak h^$. An important example of a reductive dual pair is $O(d),Sp(m,\mathbbm R)\subset Sp(md,\mathbbm R)$, which corresponds to total angular momentum of $m$ particles in dimension $d$, and which generalizes Examples \[drehimpuls\] and \[mpart\] (for more details see [@LerMontSj section 5]). It is indicated in [@Kepler], that the $S^1$-action of Example \[stratres\] is linked to the Kepler problem via the reductive dual pair $U(1),SU(2,2)\subset Sp(4,\mathbbm R)$. ### Dirac reduction {#Diracsubsection} There is a second notion of algebraic phase space reduction, which goes back to works of Dirac [@Dirac], and considerably predates the universal reduction. In the following we will exclusively be concerned with the Dirac reduction in its most simple form, i.e., the first class formalism. Let us recall the requisite terminology. Let $Z$ be a closed subset of $M$. The vanishing ideal of $Z$ is defined to be $$I_Z:=\{f\in\mathcal C^\infty(M)\:\|\:f(z)=0\quad\forall z\in Z\}.$$ A function on $Z$ is called smooth if it arises as a restriction of a smooth function on $M$. The space of smooth functions on $Z$ is denoted by $\mathcal C^\infty(Z)$. Sometimes $\mathcal C^\infty(Z)$ is also callled the space of Whitney smooth functions on $Z$ – not to be confused with the Whitney functions. Clearly, the restriction map ${\operatorname{res}}:\mathcal C^\infty(M)\to \mathcal C^\infty(Z)$ is onto, and a morphism of $\mathbbm K$-algebras. The kernel of ${\operatorname{res}}$ being $I_Z$, we obtain a short exact sequence $$0\to I_Z\to\mathcal C^\infty (M)\stackrel{{\operatorname{res}}}{\to}\mathcal C^\infty (Z)\to 0$$ of commutative $\mathbbm K$-algebras. Let $Z$ be a closed subset of a Poisson manifold $M$. A smooth function $f\in \mathcal C^\infty(M)$ is called a constraint if its restriction to $Z$ vanishes, i.e., $f\in I_Z$. It is called first class if $\{\:f,I_Z\:\}\subset I_Z$, otherwise it is called second class. If $I_Z$ consists of first class functions, i.e., if it is a Poisson subalgebra of $\mathcal C^\infty(M)$, then we call the ideal $I_Z$ coisotropic and $Z$ first class. If $Z$ is first class, then there is a canonical action of the Lie algebra $I_Z$ on $\mathcal C^\infty (Z)=\mathcal C^\infty(M)/I_Z$. If $F\in \mathcal C^\infty (M)$ is a representative of $f\in \mathcal C^\infty(Z)$, i.e., $f={\operatorname{res}}\:F$, then $h\in I_Z$ acts on $f$ via $h.f:={\operatorname{res}}\:(\{h,F\})$. For the space of invariants of this action we write $\mathcal C^\infty(Z)^{I_Z}$. According to Dirac it is a Poisson algebra, and we call it the Dirac reduced algebra. \[Dirac reduction\] Given a first class constraint set $Z\subset M$, then $\mathcal C^\infty(Z)^{I_Z}$ carries a canonical Poisson bracket, which is given as follows. Let $f,g\in \mathcal C^\infty(Z)^{I_Z}$ and let $F$ and $G$ be smooth functions on $M$ such that $f={\operatorname{res}}\:F$ and $g={\operatorname{res}}\:G$. Then the bracket is given by $\{f,g\}:={\operatorname{res}}\{F,G\}$. The only thing we have to check is, that the bracket is well defined. Note, that if $f={\operatorname{res}}F$ is invariant, then $F$ has to be in the Lie normalizer $N(I_Z):=\{\:h\in\mathcal C^\infty(M)\:\|\:\{h,I_Z,\}\subset I_Z\:\}$ of the ideal $I_Z$. By Jacobi’s identity $N(I_Z)$ is a Lie subalgebra of $\mathcal C^\infty(M)$. Since $Z$ is first class, it is in fact a Poisson subalgebra, which contains $I_Z$ as a Poisson ideal. In fact, if $f,h\in\mathcal C^\infty(Z)^{I_Z}$ are represented by ${\operatorname{res}}F=f={\operatorname{res}}F'$ and $h={\operatorname{res}}H$, then $F-F'\in I_Z$ and it follows that ${\operatorname{res}}(\{\:F-F',H\:\})=0$ since $H$ is in $N(I_Z)$. \ Of primary interest is, of course, the situation, when $Z=J^{-1}(0)$ is the zero fibre of an equivariant moment map $J$. If $0\in \mathfrak g^$ is a regular value of $J$, then $Z$ is a coisotropic submanifold of $M$, that is a first class constraint set. Unfortunately, for many singular moment maps of interest $Z$ is not* first class. An easy example of a moment map with a second class constraint set is the harmonic oscillator (example \[harmosc\]). For this reason Dirac reduction is not called universal. At least for compact group actions it is quite obvious, that if Dirac reduction works both procedures give the same result. \[redcomparison\]Let $M$ be a Poisson manifold with a Hamiltonian action of a compact, connected Lie group $G$ with equivariant moment map $J:M\to \mathfrak g^$, such that $Z:=J^{-1}(0)$ is first class. Then the Dirac reduced algebra is isomorphic to the Poisson algebra obtained by universal reduction. It suffices to show that every smooth invariant function $f$ on $Z$ has an invariant representative $F\in\mathcal C^\infty(M)$. But that can be easily obtained by averaging. ### Normal coordinates The basic requisite to make Dirac reduction a powerful tool is a good description of the vanishing ideal, e.g. by giving a set of generators. In [@AGJ] the authors developed techniques for deciding whether the components of a given moment map $J_1,\dots,J_\ell$ are a set of generators for the ideal of the zero fibre $Z=J^{-1}(0)$. They rely on the following notion of normal coordinates for the moment map. \[normcoord\] Let $M$ be an almost Kähler manifold of dimension $2n$ and let the Lie group $G$ act properly on $M$ by automorphisms. Moreover, let $J$ be an equivariant moment map for this action, and $z\in Z:=J^{-1}(0)$. Let us choose a complementary subspace $\mathfrak h\subset \mathfrak g$ to the isotropy subalgebra $\mathfrak g_z\subset \mathfrak g$. We fix a basis $v_1,\dots,v_d$ for $\mathfrak g_z$ and a basis $w_1,\dots,w_e$ for $\mathfrak h$ and denote by $v^1,\dots,v^d$ and $w^1,\dots,w^e$ the corresponding dual basis for $\mathfrak g_z^$ and $\mathfrak h^$, respectively. Then there is a neighborhood $U\subset M$ of $z$ and a local coordinate system around $z$ $$\begin{aligned} \phi:=(x,y,s,t)=(x_1,\dots,x_m,y_1,\dots,y_m,s_1,\dots,s_e,t_1,\dots,t_e):U\to \mathbbm R^{2n}, z\mapsto 0,\end{aligned}$$ such that the following conditions are true. 1. The subset $\mathcal M:=U\cap \phi^{-1}\big(\{(x,y,0,0)\in\mathbbm R^{2n}\}\big)$ is a $G_z$-invariant symplectic submanifold of $U$, which is called the linear reduced space. There is a representation $G_z\to U(m)$ such that the coordinate map $(x,y):\mathcal M\to \mathbbm R^{2m}$ intertwines the $G_z$ actions. Here we think of $\mathbbm R^{2m}$ provided with the standard Kähler structure with standard $U(m)$-action (see example \[standardex\]). Accordingly, the moment map $J_z$ for the $G_z$-action on $\mathcal M$ is given by the formula: $$\begin{aligned} \label{polpart} J_z(x,y)=-{\frac{1}{2}}\sum _{a=1}^d\sum_{i,j=1}^m \big(A^{ij}_a(x_iy_j-x_jy_i)+S^{ij}_a(x_iy_j+x_jy_i)\big)v^a,\end{aligned}$$ where $A_a=(A^{ij}_a)$ and $S_a=(S^{ij}_a)$ are real antisymmetric (resp. symmetric) $m\times m$-matrices for $a=1,\dots,d$. 2. The restriction of $J$ to $U$ is given by the formula: $$\begin{aligned} J(x,y,s,t)=\operatorname{Ad}^\natural_{\operatorname{exp}(-\sum_{a=1}^e t^a w_a)}\big(J_z(x,y)+\sum_{a=1}^e s^a w^a\big).\end{aligned}$$ Here $\operatorname{exp}$ stands for the exponential map from $\mathfrak g$ to $G$, and $\operatorname{Ad}^\natural$ denotes the coadjoint action of $G$ on $\mathfrak g^$. 3. Each of the sets $\{(x,y,s,t)\:\|\: x=x_0,y=y_0, s=s_0\}\subset U$ is contained in a single $G$-orbit. Such a coordinate system will be called normal coordinates centered at $z\in Z$. If $z'\in \mathcal M\cap Z$, then there exists an analytic change of normal coordinates centered at $z$ to coordinates centered at $z'$. See [@AGJ p.62–65]. \ It is well known that every symplectic manifold $M$ admits a compatible almost complex structure. For a proof of this fact see e.g. [@BatesWein]. If in addition a compact Lie group acts on $M$ then this almost complex structure can be made equivariant by averaging over the Haar measure. Note, that in the examples \[harmosc\],\[standardex\],\[drehimpuls\],\[mpart\],\[stratres\],\[tzwo\] and \[commvar\] the moment map is already given in (global) normal coordinates around the fix point zero. ### The generating hypothesis {#ghsubsection} In this subsection we shall explain the tools developed in [@AGJ] to decide, whether the components of a given moment map $J:M\to \mathfrak g^$ generate the vanishing ideal $I_Z\subset \mathcal C^\infty(M)$. In this case we will say for short that $J$ satisfies the generating hypothesis. For completeness and better intelligibility let us mention the following criteria for the constraint set of a moment map to be first class [@AGJ Proposition 5.2]. \[frstclasscrit\] Let $M$ be a symplectic manifold with a Hamiltonian action of a compact Lie group $G$ and let $Z=J^{-1}(0)$ be the zero fibre of the moment map $J$. Then the following statements are equivalent 1. The spanning condition* $T_z Z=\ker T_z J$ holds at every $z\in Z$. 2. $Z$ is first class. 3. For every $z\in Z$ all polynomial constraints on the linear reduced space at $z$ (see Theorem \[normcoord\]) are first class. Here the tangent space at $z\in Z$ is the linear span of tangent vectors , which are obtained by taking the derivatives of smooth curves $\gamma:[0,\epsilon[\to Z$, with $\gamma(0)=z$. Note that there are examples of noncompact group actions for which Theorem \[frstclasscrit\] is wrong. The spanning condition is quite a practical tool to sort out second class examples, like the harmonic oscillator. Moreover, the theorem says that the question of $Z$ being a first class constraint set is actually a question of the local real algebraic geometry of the moment map. It is important to note that there are examples of moment maps with first class constraint set, which do not satisfy the generating hypothesis (see [@AGJ Example 7.13]). Another necessary criterion, which is even more easy to check, is the following nonpositivity condition [@AGJ Proposition 6.7]. \[nonpcond\] If the spanning condition holds at $z\in Z$, then $J$ fulfills the following nonpositivity condition at $z\in Z$: for every $\xi\in\mathfrak g$ one has either 1. $J(\xi)=0$ in a neighborhood $U\subset M$ of $z$, or 2. in every neighborhood $U\subset M$ of $z$ the function $J(\xi)$ takes strictly positive as well as strictly negative values. If $G$ is a torus, then, due to the following theorem (see [@AGJ Theorem 6.8]), the nonpositivity condition is also sufficient for $J$ to satisfy the generating hypothesis. \[torus\]Let $M$ be a symplectic manifold with a Hamiltonian action of a torus $G$, with moment map $J:M\to\mathfrak g^$. Then $J$ satisfies the generating hypothesis $\Leftrightarrow$ the spanning condition holds for all $z\in Z=J^{-1}(0)$ $\Leftrightarrow$ the nonpositivity condition holds for all $z\in Z$. Now we are ready to address the torus actions from our list of examples. As we already indicated the harmonic oscillator, Example \[harmosc\], does not satisfy the nonpositivity condition. The same is true for Example \[lemon\] (the ‘lemon’) at the singular values $\pm 1$. It is easy to see, that at the singular value $0$ the nonpositivity condition is fulfilled. Furthermore, the nonpositivity condition is clearly true for zero angular momentum for $m$ particles in the plane, Example \[mpart\], and for the $(1,1,-1,-1)$-resonance, Example \[stratres\]. For the $\mathbbm T^2$-action of Example \[tzwo\] the nonpositivity condition holds iff $\alpha<0$. As the nonpositivity condition, in the case of nonabelian group actions, is only necessary* for the ideal $I_Z\subset \mathcal C^{\infty}(M)$ to be generated by $J_1,\dots ,J_\ell$, the reasoning here is usually more intricate. As a first step, one notices that it is enough to analyse the problem locally. In fact, with the assumptions of Theorem \[normcoord\] the following statements are equivalent [@AGJ Corollary 4.6] 1. the components of the moment map $J$ generate the vanishing ideal $I_Z\subset \mathcal C^\infty(M)$ of $Z$ in $M$, 2. \[locgh\]for every $z\in Z$, in a normal coordinate system around $z$ the components of the moment map $J_z$ (cf. equation (\[polpart\])) generate the ideal $I_{J_z^{-1}(0)}\subset \mathcal C^\infty(\mathcal M)$ of the zero level set $J_z^{-1}(0)$ in the linear reduced space $\mathcal M$. Of course, it is tempting to view the quadratic polynomial function $J_z$ as a polynomial in the polynomial ring $\mathbbm R[x_1,\dots,x_m,y_1,\dots,y_m]$, where $m={\frac{1}{2}}\dim M+\dim G_z-\dim G$. As a matter of fact, if $I_{pol}(J_z)\subset \mathbbm R[x_1,\dots,x_m,y_1,\dots,y_m]$ denotes the ideal generated polynomial $J_z$ then it follows from the proof of [@AGJ Theorem 6.3] that if 3. \[realcrit\] the ideal $I_{pol}(J_z)$ is a real ideal in the sense of real algebraic geometry, then the local generating hypothesis \[locgh\]. above is true in $z$ ,i.e., $J_z$ generates the ideal $I_{J_z^{-1}(0)}\subset\mathcal C^\infty(\mathcal M)$. Let us recall that an ideal $I$ in an $\mathbbm R$-algebra $R$ is defined to be a real ideal, if it coincides with its real radical $$\begin{split} \sqrt[\mathbbm R]{I} := \big\{f\in R \mid f^{2i}+\sum_{s=1}^j g_s^2 \in I\text{ for some $i,j$ and } g_1,\dots , g_j\in R \big\}. \end{split}$$ If $R=\mathbbm R[x_1,\dots,x_k]$ the real Nullstellensatz says that an ideal $I\subset R$ is real iff it coincides with the ideal $I_{V(I)}=\{f\in R \mid f_{\|V(I)}=0\}$ of its real locus $V(I)=\{(a_1,\dots,a_k)\in\mathbbm R^k\mid f(a_1,\dots, a_k)=0\quad \forall f\in I\}$. The real Nulstellensatz is not valid for ideals in the ring of smooth functions of some manifold. The full statement of [@AGJ Theorem 6.3] is that for the above argument also the converse is true: the components of the moment map $J$ generate the ideal $I_Z\in \mathcal C^\infty(M)$ $\Leftrightarrow$ they generate a real ideal in $\mathcal C^\infty(M)$ $\Leftrightarrow$ for all $z\in Z$ the ideal $I_{pol}(J_z)$ is a real ideal in $\mathbbm R[x_1,\dots,x_m,y_1,\dots,y_m]$. As a last step, we follow the advice of [@AGJ] and translate the reality criterion \[realcrit\]. above into the more amenable language of complex algebraic geometry. \[GotayThm\] Let $I$ be an ideal in $\mathbbm R[x_1,\dots, x_k]$. Then $I$ is real, if and only if the following two conditions hold: 1. $I_\mathbbm C:=I\otimes_\mathbbm R \mathbbm C$ is radical in $\mathbbm C[x_1,\dots, x_k]$, and 2. for every irreducible component $W\subset \mathbbm C^m$ of the (complex) locus of $I_\mathbbm C$ we have that $\dim_{\mathbbm R} (W\cap \mathbbm R^m)=\dim_{\mathbbm C}(W)$.[^4] In other words, in order to check the generating hypothesis it is necessary to gain detailed insight into the complex algebraic geometry behind the scene (e.g. knowing the primary decomposition of $I_\mathbbm C$). Regardless of the fact that the varieties in question are cones, there is no straightforward way to provide this information. A basic and physically interesting example, for which the algebraic geometry is well-studied is the system of one particle in dimension $n$ with zero angular momentum, Example \[drehimpuls\]. Since the components of the moment map can (up to a sign) be written as the $2\times 2$-minors of the $2\times n$-matrix $$\begin{aligned} \left( \begin{array}{cccc} q^1&q^2&\dots&q^n\\ p_1&p_2&\dots&p_n \end{array} \right)\end{aligned}$$ the ideal $I_{\mathbbm C}$ which is generated by the components of the moment map is an instance of a determinantal ideal. By a theorem of Hochster [@Hochster] the ideal $I_\mathbbm C$ is prime, and the complex locus is of dimension $n+1$. It is easy to see [@AGJ Example 7.10], that the dimension condition is true as well. In fact, the $2\times 2$-minors of the above matrix are zero if and only if the vectors $(q^1,\dots,q^n)$ and $(p_1,\dots,p_n)$ are proportional. In particular, for any $(q^1,\dots q^n)\in\mathbbm R^n$ and $\lambda\in \mathbbm R^\times$ the vector $(\lambda q^1,\lambda q^2,\dots,\lambda q^n, ,\lambda^{-1}q^1,\lambda^{-1}q^2,\dots,\lambda^{-1}q^n)\in T^\mathbbm R^n$ is in the real locus, whose dimension is thus $\ge n+1$. We conclude that the ideal $I$ is real. Unfortunately, this example is not a complete intersection for $n\ge3$ (more on this in section \[kc\]). The only class of nonabelian examples, which the author is aware of, where the generating and (as we will see in section \[kc\]) the complete intersection hypothesis are true at the same time, is Example \[commvar\]. The complex locus $Z_{\mathbbm C}$ defined by these $\frac{1}{2}n(n-1)$ quadratic equations is an instance of what is called a commuting variety. In [@BrenPiVas] it was shown that $Z_{\mathbbm C}$ is irreducible of codimension $\frac{1}{2}n(n-1)$, and the ideal generated by the coefficients of $J$ in the complex polynomial ring is prime. Let $S_{\mathrm{reg}}\subset S$ be the open subset of symmetric matrices with pairwise distinct eigenvalues. Since the action of $SO(n)$ on $T^S_{\mathrm{reg}}$ is locally free, it follows that $Z\cap T^S_{\mathrm{reg}}$ is of codimension $\frac{1}{2}n(n-1)$ likewise. As a consequence of Theorem \[GotayThm\], the components of $J$ generate the vanishing ideal $I_Z$ in $\mathcal C^{\infty}(T^S)$. Summarizing, we have seen that the generating hypothesis is true for the Examples \[drehimpuls\], \[mpart\], \[lemon\], \[stratres\], \[tzwo\] for $\alpha<0$ and \[commvar\]. We would like to close this subsection with a discussion of the elementary example of the free particle on the line, Example \[allwrong\], where essentially all what we have done so far goes wrong. Clearly, $0$ is a singular value of $J(q,p)={\frac{1}{2}}p^2$. The constraint surface $Z=J^{-1}(0)=\{(q,0)\in T^* \mathbbm R\}=\mathbbm R$ is in fact a Lagrangian submanifold, hence first class. The action of $\mathbbm R$ on $Z$ is trivial, therefore the action of $\mathbbm R$ on $T^\mathbbm R$ can not be proper (proper actions have compact isotropy groups). The orbits of the $\mathbbm R$-action on $T^\mathbbm R$ are closed, but the quotient space $T^\mathbbm R/\mathbbm R$ is not Hausdorff. Furthermore, the quotient space $Z/\mathbbm R=Z$ is for obvious reasons not a stratified symplectic space, i.e., the theorem of Sjamaar and Lerman does not apply. Any function on $Z$ is invariant, but the only smooth function which can be extended to a smooth invariant functions on $T^ \mathbbm R$ are the constants. Hence, the Poisson algebra obtained by universal reduction is just $\mathbbm K$. In contrast, the Dirac reduced algebra is $\mathcal C^\infty(\mathbbm R)$ with trivial Poisson structure. Finally, $p$ is a constraint which is not a multiple of $J$. According to Śniatycki and Weinstein [@SnW] $\mathcal C^\infty(T^\mathbbm R)/p^2\mathcal C^\infty(T^\mathbbm R)$ is naturally a Poisson algebra, which in this example differs from the universal and Dirac Poisson reduced structure. Deformation quantization ------------------------ In this section we recall some basic notions and examples from the theory of deformation quantization. We recall Fedosov’s construction of star products on symplectic manifolds. We discuss the various invariance properties of star products, which will be the basic requisites for quantum phase space reduction in chapter \[qbrst\]. ### Formal deformations of associative algebras Let $A$ be an associative $\mathbbm K$-algebra with unit $\mathbbm 1$ and let us denote the multiplication map by $\mu:A\otimes_\mathbbm K A\to A$,, $a\otimes b\mapsto ab$. A formal deformation of the algebra $A$ is a sequence of linear operations $\mu_i:A\otimes_\mathbbm K A\to A$ for $i\ge 1$, such that the deformed multiplication $$ $$\begin{aligned} \label{formaldef} ab:=ab+\sum_{i\ge 1}\nu^i\mu_i(a,b), \end{aligned}$$ where $a,b\in A$ defines (by $\nu$-linear extension) an associative algebra structure on $A{[[\nu]]}$ with unit $\mathbbm 1$. The variable $\nu$ is called the formal parameter. Two formal deformations $$ and $'$ are called equivalent if there is a sequence of linear maps $S_i:A\to A$, for $i\ge 1$, such that $S:={\operatorname{id}}+\sum_{i\ge 1}\nu^i S_i$ defines an isomorphism of unital $\mathbbm K{[[\nu]]}$-algebras from $(A{[[\nu]]},)$ to $(A{[[\nu]]},')$, i.e. $$\begin{aligned} \label{equiv} S(ab)=S(a)'S(b)\end{aligned}$$ for all $a,b\in A$. If the algebra $A$ is commutative, it is easy to show, that the semiclassical limit of a formal deformation of $A$ makes $A$ into a Poisson algebra, i.e. $$\begin{aligned} \{a,b\}:=\mu_1(a, b)-\mu_1(b, a),\quad \mbox{ for }a,b\in A\end{aligned}$$ is a Poisson bracket. Conversely, if we start with a Poisson algebra $(A,\cdot,\{\:\:\})$, then a formal deformation of $(A,\cdot)$ is said to be a deformation of this Poisson algebra, if the semiclassical limit reproduces the original Poisson structure. An easy example of such a formal deformation of a commutative algebra $(A,\mu)$ arises if there is a family of pairwise commuting derivations $D_i$ , $i=1,\dots,n$, of $A$. Then for any tensor $P^{ij}\in \mathbbm K$, $i,j=1\dots,n$, $$\begin{aligned} ab:=\mu\circ {\operatorname{e}}^{\nu\sum_{i,j=1}^nP^{ij}D_i\otimes D_j}(a\otimes b)\end{aligned}$$ defines an associative deformation of $(A,\mu)$. If $A$ is the real polynomial ring $\mathbbm R[x^1,\dots,x^n]$ or $\mathcal C^\infty(\mathbbm R^n)$ and the $D_i$ are the partial derivatives $\frac{\partial}{\partial x^i}$ then the above multiplication is called the Moyal-Weyl-multiplication. The semicassical limit of the Moyal-Weyl multiplication is the Poisson structure corresponding to the constant Poisson bivector field ${\frac{1}{2}}\sum_{ij}\Pi^{ij}\frac{\partial}{\partial x^i}\wedge\frac{\partial}{\partial x^i}$, where $\Pi^{ij}=P^{ij}-P^{ji}$. Another important example of an associative deformation arises if $A=S\mathfrak h$ is the symmetric algebra of an real Lie algebra $\mathfrak h$ from the PBW-symmetrization map (more on this in subsection \[qmom\]). The notion of a formal deformation of an associative algebra, straightforwardly generalizes to $\mathbbm Z_2$ or $\mathbbm Z$-graded algebras (with Koszul sign rule). The purely odd analogon of Moyal-Weyl product is Clifford multiplication (see also equation (\[clmult\])). We leave it to the reader to fill in the details. It should be mentioned that the stepwise obstructions to construct a formal deformation of the multiplication $\mu$ lie in the third Hochschild cohomology group ${HH}^3(A,A)$ of the algebra $A$. In fact, the space $C^n(A,A)$ of Hochschild $n$-cochains is just the space linear maps $\varphi: A^{\otimes n}=A\otimes \dots \otimes A\to A$ and the differential of Hochschild cohomology $\delta_H:C^n(A,A)\to C^{n+1}(A,A)$ is given by the formula $$\begin{aligned} (\delta_H\varphi)(a_1,\dots, a_{n+1})&=&a_1\varphi(a_2,\dots, a_n)+\sum_{i=1}^{n}(-1)^i\varphi(a_1,\dots, a_ia_{i+1},\dots,a_{n+1})\nonumber\\ &&\qquad\qquad\qquad\qquad+(-1)^{n+1}\varphi(a_1,\dots, a_n)\:a_{n+1}.\label{Hochdiff}\end{aligned}$$ for $a_1,\dots, a_{n+1}\in A$. The associativity of the product (\[formaldef\]) in degree $\nu^k$ can be written as $$\begin{aligned} (\delta_H \mu_k)(a_1,a_2,a_3)=\sum_{i=1}^{k-1} \mu_i(\mu_{k-i}(a_1,a_2),a_3)-\mu_i(a_1,\mu_{k-i}(a_2,a_3))\qquad\forall a_1,a_2,a_3\in A.\end{aligned}$$ Therefore, if we would construct such a product inductively, then we would have to assure that the right hand side of this equation is a Hochschild 3-coboundary at every step. Using the differential graded Lie algebra structure on the shifted Hochschild cochain complex $C^\bullet(A,A)[1]$ which is given by the so-called Gerstenhaber bracket, the set of equivalence classes of formal deformations of $\mu$ can be described in terms of deformation functors and Maurer-Cartan equations. As we will not use this slightly more sophisticated language we refer the interested reader to the exposition [@KellerMini]. ### Star products If we are looking for a formal deformation of the Poisson algebra $A:=\mathcal C^\infty(M)$ of smooth functions on a Poisson manifold $(M,\Pi)$, then the general opinion is that full Hochschild cochain complex $C^\bullet (A,A)$ is too big for the deformation problem to make sense. As a consequence one usually considers the subcomplex $C^\bullet_{{\operatorname{diff}}}(A,A)$ of differential* Hochschild cochains $$\begin{aligned} C^k_{{\operatorname{diff}}}(A, A)=\{D:\underbrace{A\otimes\dots\otimes A}_{\mbox{k times}},\to A\:\|\:D \mbox{ is a polydifferential operator}\}\end{aligned}$$ instead. Accordingly, a formal deformation $$\begin{aligned} fg:=fg+\sum_{i\ge 1}\nu^i\mu_i(f\otimes g),\qquad f,g\in \mathcal C^\infty(M)\end{aligned}$$ of the Poisson algebra $(\mathcal C^\infty(M),\{\:,\})$ is defined to be a star product, if the bilinear operations $\mu_i$, $i=1,2,\dots$, are in fact bidifferential operators. Moreover, two such star products $$ and $'$ are defined to be equivalent, if the equivalence transformation $S:={\operatorname{id}}+\sum_{i\ge 1}\nu^i S_i$ of equation (\[equiv\]) is in fact a series of differential operators. Essentially all known explicit constructions of star products yield bidifferential operators $\mu_i$, which are differential operators of order at most $i$ in each argument. These star products are also called natural. It can be shown [@GuttRawnsley03] that an equivalence transformation $S$ between two natural star products $$ and $'$ is of the form $S=\operatorname{exp}(\sum_{i=1}^\infty \nu^iD_i)$, where the $D_i$ are differential operators of order at most $i+1$. Moreover, in the symplectic case every star product is equivalent to a Fedosov star product (see the next subsection) and, hence, to a natural one. Instead of considering differential star products one can also exploit the fact that $\mathcal C^\infty(M)$ is a nuclear Fréchet algebra and look at continuous* star products (defined below). Even though the meaning of continuity is somehow obscure, this approach has the great advantage that it applies for other interesting nuclear Fréchet algebras such as the singular reduced algebra $\mathcal C^\infty(Z)^{\mathfrak g}$ of subsection \[Diracsubsection\]. In fact, $\mathcal C^\infty(Z)$, being a quotient of a nuclear Fréchet space modulo a closed subspace, is nuclear Fréchet (cf. [@Treves Proposition 50.1]), as well as $\mathcal C^\infty(Z)^{\mathfrak g}$, being a closed complemented subspace of a nuclear Fréchet space (remember that $G$ is assumed to be compact and connected). If $A$ is a nuclear Fréchet algebra with commutative multiplication $\mu$, then the space $C^\bullet_{{\operatorname{cont}}}(A, A)$ of continuous Hochschild cochains as follows $$\begin{aligned} C^k_{{\operatorname{cont}}}(A,A)=\{D:\underbrace{A\widehat{\otimes}\dots\widehat{\otimes} A}_{\mbox{k times}},\to A\:\|\:D \mbox{ is linear continuous}\}.\end{aligned}$$ Here, $\widehat{\otimes}$ denotes the topological tensor product (since $A$ is a nuclear space all topological tensor products coincide). Accordingly, by a continuous star product we mean a formal deformation as in formula (\[formaldef\]), such that the operations $\mu_i:A\widehat{\otimes}A\to A$ are continuous. An equivalence transformation between continuous star products such as in formula (\[equiv\]) is said to be continuous if the operations $S_i:A \to A$ are continuous. The notion of a continuous star product has been studied before in [@BFGP], together with the topological version of Gerstenhaber deformation theory. Actually, for the main application we have in mind, i.e., the singular reduced algebra $\mathcal C^\infty(Z)^\mathfrak g$, it is not at all obvious what the ‘correct’ definition of (multi-) differential operator should be (see [@PflaumHab]), in order to define a feasible notion of a differential star product. As an indication that the continuous setup is not too weak, let us mention the well known fact that the natural map $C_{{\operatorname{diff}}}(A,A)\to C_{{\operatorname{cont}}}(A,A)$ is a quasiisomorphism if $A=\mathcal C^\infty(M)$ is the algebra of smooth functions on a smooth manifold $M$. The statement that the cohomology of $C_{{\operatorname{diff}}}(A,A)$ (respectively $C_{{\operatorname{cont}}}(A,A)$) is isomorphic to the space of polyvector fields on $M$ is just the well known differential (resp. continuous) Hochschild-Kostant-Rosenberg theorem. ### Fedosov construction For the sake of completeness and because it is a nice application of the perturbation lemma \[BPL1\], let us recall Fedosov’s construction of star products on a symplectic manifold $(M,\omega)$. The point of departure of Fedosov’s construction is an algebraic, fibrewise version of the de Rham complex. This complex occurs at many places in mathematics, e.g. in [@Bourbdix §9 No.3 exemple 1], but viewed the opposite way as a Koszul complex of a ‘linear constraint’. More precisely, we consider $S\otimes \wedge \:\:T^M\to M$, the tensor product of the symmetric and Grassmann algebra bundle of $T^M$, the cotangent bundle of the manifold $M$. A choice of local coordinates $x^1,\dots,x^n$ for $M$ gives a local frame $dx^1,\dots,dx^n$ for $T^M=\wedge^1 T^M$. The corresponding frame for $T^M=S^1 T^M$ will be written as $y^1,\dots,y^n$. We will interprete a section of $S^k\otimes \wedge^l \:\:T^M\to M$ as a polynomial valued $l$-form, i.e. the local frames are written as $y^{i_1}\dots y^{i_k}dx^{j_1}\wedge\dots\wedge dx^{j_l}$ for $i_1\le i_2,\dots\le i_k$ and $j_1<j_2\dots<j_l$. We introduce the algebraic (fibrewise) de Rham differential $$\delta:=\sum_i dx^i\wedge \frac{\partial}{\partial y^i}:\Gamma^\infty(M,S^k\otimes \wedge^l\:T^M)\to\Gamma^\infty(M,S^{k-1}\otimes \wedge^{l+1}\:T^M).$$ Obviously, $\big(\Gamma^\infty(M,S\otimes \wedge^\bullet\:T^M),\wedge,\delta\big)$ is a $\mathbbm Z$-graded super-commutative differential graded algebra. Moreover, there is an algebraic Poincaré lemma, i.e., the differential $\delta$ is acyclic. A contracting homotopy for $\delta$ is given as follows. We introduce the Koszul differential $\delta^=\sum_i y^i i(\partial/\partial x^i):\Gamma^\infty(M,S^k\otimes \wedge^l\:T^M)\to\Gamma^\infty(M,S^{k+1}\otimes \wedge^{l-1}\:T^M)$. An easy calculation yields the commutation relation $\delta\delta^+\delta^\delta=(k+l){\operatorname{id}}:\Gamma^\infty(M,S^k\otimes \wedge^l\:T^M)\to\Gamma^\infty(M,S^k\otimes \wedge^l\:T^M)$. Renormalizing the Koszul differential $\delta ^$, we obtain a contracting homotopy: $\delta^{-1}:=(k+l)^{-1}\delta^:\Gamma^\infty(M,S^k\otimes \wedge^l\:T^M)\to\Gamma^\infty(M,S^{k+1}\otimes \wedge^{l-1}\:T^M)$ for $k+l>0$ and for $k=0=l$ we define $\delta^{-1}$ to zero. Let us introduce the canonical projection $\pi:\Gamma^\infty(M,S\otimes \wedge\:T^M)\to \mathcal C^\infty(M)$ and the canonical injection $\iota:\mathcal C^\infty(M)\to\Gamma^\infty(M,S\otimes \wedge\:T^M)$. then the commutation relation above can be rewritten as $\delta\delta^{-1}+\delta^{-1}\delta={\operatorname{id}}-\iota\pi$. Traditionally $\iota\pi$ is denoted by $\sigma$. What we have done so far may be neatly subsumed in the language of appendix \[HPT\]: $$\begin{aligned} \label{elemcontr} \xymatrix{(\mathcal C^\infty(M),0)\ar@<-0.6ex>[r]_{\iota\quad\qquad}^{}&(\Gamma^\infty(M,S\otimes \wedge T^M),\delta),\delta^{-1} \ar@<-0.6ex>[l]_{\pi\quad\qquad}^{}},\end{aligned}$$ is a contraction fulfilling all side conditions. Moreover, $\iota$ and $\pi$ are homomorphisms of super-commutative algebras and $\delta$ is a derivation. It is clear, that if we replace the symmetric algebra part by the algebra of formal power series (i.e., we take the completion with respect to the ideal generated by $\Gamma^\infty(M, S^1\otimes \wedge^0 T^M)$), we do not spoil the contraction (\[elemcontr\]). The same applies, if we adjoin a formal variable $\nu$. Therefore, we may replace the above de Rham complex by the so called Weylalgebra* $$\begin{aligned} \mathcal W\otimes \Omega^\bullet(M):=\prod_{k\ge 0}^\infty \Gamma^\infty(M,S^k\otimes \wedge^ \bullet T^M){[[\nu]]},\end{aligned}$$ which is in an obvious way a $\mathbbm Z$-graded super-commutative $\mathbbm K{[[\nu]]}$-algebra. Again, we have a contraction $$\begin{aligned} \label{formelemcontr} \xymatrix{(\mathcal C^\infty(M){[[\nu]]},0)\ar@<-0.6ex>[r]_{\iota\quad}^{}&\big(\mathcal W\otimes \Omega^\bullet(M),\delta\big),\delta^{-1} \ar@<-0.6ex>[l]_{\pi\quad}^{}},\end{aligned}$$ which fulfills all side conditions. The next ingredient is a formal deformation $\circ$ of the super-commutative product on $\mathcal W=\prod_{k\ge 0}^\infty \Gamma^\infty(M,S^k T^M){[[\nu]]}$ into an associative multiplication. It is given by a fibrewise Moyal-Weyl-multiplication $$\begin{aligned} a\circ b:= \mu \exp\Big(\frac{\nu}{2}\sum_{i,j}\Pi^{ij}\frac{\partial}{\partial y^i}\otimes\frac{\partial}{\partial y^j}\Big)(a\otimes b),\end{aligned}$$ where $a,b\in\mathcal W$ and $\mu$ denotes the commutative multiplication. Recall that according to our sign convention $\sum_k\Pi^{ik}\omega_{kj}=\delta_j^i$. It is clear, that the Moyal-Weyl-multiplication $\circ$ respects the form degree. But in fact, there is a second $\mathbbm Z$-grading for which $\circ$ is graded. This total degree is given by counting the symmetric degree and twice the $\nu$ degree simultaneously. The homogeneous components of this grading are given by the eigenspaces $\mathcal W^{(k)}=\{a\in \mathcal W\mid \operatorname{Deg}a=ka\}$ of the derivation $\operatorname{Deg}:=\sum_i y^i\frac{\partial}{\partial y^i}+ 2\nu\frac{\partial}{\partial \nu}$, which is a derivation of $\circ$. For the descending filtration induced by $\operatorname{Deg}$ we write $\mathcal W_i:=\prod_{k\ge i}^\infty \mathcal W^{(k)}$. The multiplication $\circ$ extends naturally to $\mathbbm Z$-graded multiplication $\circ$ for $\mathcal W\otimes\Omega^\bullet(M)$. The super-center of the algebra $(\mathcal W\otimes\Omega^\bullet(M);\circ)$ is just $\Omega(M)[[\nu]]$. The derivation $\delta$ can be written as an inner derivation of $\mathcal W\otimes\Omega^\bullet(M)$: $\delta=\nu^{-1}{\operatorname{ad}}(\widetilde{\omega})$, where is $\widetilde{\omega}=\sum_{i,j}\omega_{ij}y^i dx^j$ and ${\operatorname{ad}}(\widetilde \omega)$ means taking the super-commutator with $\widetilde \omega$. Let $\nabla$ be a torsion free symplectic connection on the tangent bundle of $M$, i.e.,$\nabla_X Y-\nabla_Y X=[X,Y]$ for all $X,Y\in \Gamma^\infty(M,TM)$ and $\nabla \omega =0$. Such a connection can always be found. In contrast to the Riemannian case a symplectic connection is not uniquely determined. Let $\hat R\in \Gamma^\infty(M, {\operatorname{End}}(TM)\otimes \wedge^2 T^M$ denote the curvature endomorphism of the connection: $\hat R(X,Y)Z=\nabla_X\nabla_Y Z-\nabla_Y\nabla_X Z-\nabla_{[X,Y]}Z$. Because $\nabla$ is symplectic one can show that $$R(X,Y,Z,W):=\omega(X,\hat R(Z,W),Y)$$ is in fact symmetric in $X$ and $Y$. This means that $R$ is in fact a section in $\Gamma^\infty(M,S^2T^M\otimes\wedge ^2 T^M)$, i.e., an element of $\mathcal W\otimes \Omega$. We extend $\nabla$ in a standard fashion to $$\begin{aligned} \nabla=\sum_i dx^i\wedge\nabla_{\frac{\partial}{\partial x^i}}:\mathcal W\otimes \Omega ^\bullet\to \mathcal W\otimes \Omega ^{\bullet+1}.\end{aligned}$$ A little calculation yields that $$\nabla^2=\nu^{-1}{\operatorname{ad}}_\circ (R).$$ Since the connection is symplectic, one can prove that $\nabla$ is an odd derivation of $\circ$ and that $\nabla\widetilde{\omega}=0$. The latter equation entails that $[\nabla,\delta]=\nabla\delta+\delta\nabla=0$. From $[\delta,[\nabla,\nabla]]=0$ one derives $\delta R=0$, which is also known as the first Bianchi identity. From $[\nabla,[\nabla,\nabla]]=0$ one infers that $\nabla R=0$, which is known as the second Bianchi identity. The beautiful insight of Fedosov [@Fed94] has been that even though the derivation $-\delta +\nabla$ is, in general, not of square zero, it can be made into a differential by adding an (almost) inner derivation $\frac{1}{\nu}{\operatorname{ad}}_\circ(r)$. This $r$ can be found recursively. \[Fedosov\] Let $\Omega=\sum_{i\ge1}^\infty\nu^i \Omega_i\in \nu Z_{dR}^2(M)[[\nu]]$ be a series of closed 2-forms on $M$ and $s\in \mathcal W_3\otimes \Omega^0(M)$ such that $\pi(s)=0$. Then there is a unique $r\in \mathcal W_2\otimes \Omega^1(M)$ such that $\delta r=R+\nabla r+\frac{1}{\nu}r\circ r+\Omega$ and $\delta^{-1}r=s$. It follows that the Fedosov derivation* $$\begin{aligned} D:=-\delta+\nabla+\frac{1}{\nu}{\operatorname{ad}}_\circ(r)\end{aligned}$$ is a differential $\mathcal W\otimes \Omega^\bullet(M)\to\mathcal W\otimes \Omega^{\bullet+1}(M)$, i.e., $D^2=0$. The original proof (for $s=0$) can be found in [@Fed]. For a more elaborate exposition the reader may, e.g., consult [@Waldmannsoevre subsection 6.4.2]. \ The Fedosov derivation can be seen as a perturbation of the differential $-\delta$ of the contraction (\[formelemcontr\]). The filtration in question is the aforementioned one, which is associated to the degree $\operatorname{Deg}$. Clearly, this perturbation complies with the premises of perturbation lemma \[contrEins\]. As a result we obtain a contraction $$\begin{aligned} \label{Fedcontr} \xymatrix{(\mathcal C^\infty(M)[[\nu]],0)\ar@<-0.6ex>[r]_{\tau\quad}^{}&(\mathcal W\otimes \Omega(M),D),D^{-1} \ar@<-0.6ex>[l]_{\pi\quad}^{}}.\end{aligned}$$ The contracting homotopy is given by $D^{-1}=-\delta^{-1}(-D\delta^{-1}-\delta^{-1}D)^{-1}$. A little calculation yields that the fomula for the map $I$ in Lemma \[contrEins\] reproduces the well-known fomula for the Fedosov Taylor series $\tau=(-D\delta^{-1}-\delta^{-1}D)^{-1}\iota$. The contraction (\[Fedcontr\]) is used to transfer the associative algebra structure from the differential graded associative algebra $(\mathcal W\otimes \Omega(M),\circ,D)$ to $\mathcal C^\infty(M)[[\nu]]$, i.e., we obtain an associative product $$\begin{aligned} f* g:=\pi (\tau(f)\circ\tau(g)),\end{aligned}$$ for $f,g \in \mathcal C^\infty(M)[[\nu]]$. The associativity of such a product follows from general considerations (cf. also the computation (\[assoccomp\]) in chapter \[qbrst\]). The product $$ will be called the Fedosov star product* obtained from the data $(\nabla,\Omega,s)$. The name is justified by the following well known theorem. If $$ is obtained from the data $(\nabla,\Omega,s)$ then it is natural star product and the equivalence class of $$ does only depend on the cohomology class $[\Omega]\in\mathrm \nu H_{dR}^2(M)[[\nu]]$. Moreover, every star product on $M$ is equivalent to some Fedosov star product. An elementary proof for the first statments can be found in [@Waldmannsoevre]. The last statement follows from classification results of the thesis of N. Neumaier [@Neudiss]. \ We would also like to mention the following nice result on derivations of Fedosov star products. \[nicethm\] Let $M$ be a symplectic manifold with a Fedosov star product $$, which is obtained from the data $(\nabla, \Omega,s)$. If $X$ is a symplectic vector field on $M$, then $X$ is a derivation of $$ if and only if $\nabla$ is affine with respect to $X$ and $\mathcal L_X\Omega=0=\mathcal L_X s$. Such a vector field can be written as an (almost) inner derivation: $$\begin{aligned} X(h)=\frac{1}{\nu}ad_(f) h \qquad\forall h\in\mathcal C^\infty(M)\end{aligned}$$ for some series $f=\sum_{i\ge 0}\nu^i\,f_i\in \mathcal C^\infty(M){[[\nu]]}$ if and only if $f$ is a solution of $$\begin{aligned} df=i_X(\omega+\Omega).\end{aligned}$$ In this case we have $X=X_{f_0}$. A proof can be found in [@Neubahns section 3]. ### Quantum moment maps and strong invariance {#qmom} There are several inequivalent notions of compatibility of a star product $$ on $M$ with action of some Lie group $G$ on $M$. For example, if the Lie group $G$ acts by automorphisms of the algebra $(\mathcal C^\infty(M){[[\nu]]},)$, i.e., $(\Phi_g^f)(\Phi_g^h)=\Phi_g^(fh)$ for all $f,h\in\mathcal C^\infty(M)$ and $g\in G$ one simply says that $$ is $G$-invariant. The infinitisemal version of this notion is that of $\mathfrak g$-invariance* of the star product $$. Here one requires that the fundamental vector fields act by derivations. If the group $G$ is connected both notions are clearly equivalent. A slightly stronger notion, which will become important for us (cf. chapter \[qbrst\]), is the notion of a strongly invariant star product: $$ is said to be strongly invariant with respect to a Hamiltonian action of a Lie group $G$ with moment map $J:M\to \mathfrak g^$ if $$\begin{aligned} J(X)f-fJ(X)=\nu \{J(X),f\}\qquad \forall X\in \mathfrak g,f\in\mathcal C^\infty(M). \end{aligned}$$ It follows easily from Theorem \[nicethm\] that for symplectic manifolds with a Hamiltonian action of a compact Lie group a strongly invariant star product can always be found. In fact, the symplectic connection can be made invariant by averaging, the same is true for $s$ (alternatively, one can assume $s=0$). Finally, one has to assure that $i_{X_{J(\xi)}}\Omega=0$ for all $\xi\in \mathfrak g$. Another example of a strongly invariant star product is provided by the BCH star product [@Gutt83] on the dual space $\mathfrak h^$ of a $\mathbbm R$-Lie algebra $\mathfrak h$. Let $\mathfrak h_\nu:=\mathfrak h{[[\nu]]}$ be the $\mathbbm R{[[\nu]]}$-Lie algebra with the modified bracket $[\:,\:]_\nu:=\nu[\:,\:]$ and let $U\mathfrak h_\nu$ be the universal enveloping algebra of this $\mathbbm R{[[\nu]]}$-Lie algebra. Let us denote the canonical multiplication on $U\mathfrak h_\nu$ by $\cdot_\nu$. The (rescaled) PBW symmetrization map $\sigma_\nu$ is defined on monomials $X_1X_2\dots X_k\in S^k \mathfrak h$ by the formula $$\begin{aligned} \label{PBWsymm} \sigma_\nu(X_1X_2\dots X_k)=\frac{\nu^k}{k!}\sum_{\tau\in S_k} X_{\tau(1)}\cdot_\nu X_{\tau(2)}\cdot_\nu\dots\cdot_\nu X_{\tau(n)}.\end{aligned}$$ It is well-known that $\sigma_\nu$ extends to an injective map of $\mathbbm R{[[\nu]]}$-modules $\sigma _\nu:S\mathfrak h{[[\nu]]}\to U\mathfrak h_\nu$, such that $\sigma_\nu\left(S\mathfrak h{[[\nu]]}\right)$ is a subalgebra. It induces an associative multiplication $_{BCH}$ on $ S\mathfrak h{[[\nu]]}$ which is uniquely defined by the formula $$\begin{aligned} \label{BCHstar} \sigma_\nu(f_{BCH}g)=\sigma_\nu(f)\cdot_\nu\sigma_\nu(g).\end{aligned}$$ In fact, the bilinear composition $_{BCH}$ is given by a series of bidifferential operators for the algebra $S\mathfrak h$. Viewing $S\mathfrak h$ as polynomial functions on $\mathfrak h^$, the BCH product $_{BCH}$ extends uniquely to a star product on $\mathcal C^\infty(\mathfrak h^){[[\nu]]}$, which deforms the linear Poisson structure. By an elementary calculation one can show that for any $X\in \mathfrak h$ and any monomial $X_1X_2\dots X_k\in S^k\mathfrak h$ $$\begin{aligned} X\cdot_\nu \sigma_\nu(X_1X_2\dots X_k)-\sigma_\nu(X_1X_2\dots X_k)\cdot_\nu X=\nu \sigma_\nu\left( \{X,X_1X_2\dots X_k\}\right),\end{aligned}$$ where $\{,\}$ is the Poisson bracket arising from the linear Poisson structure. It follows that $_{BCH}$ is strongly invariant for any moment map as in Example \[linPoiss\]. There is yet another notion of compatibility of the $G$-action with the star product $$, that is interesting for our purposes (cf. chapter \[qbrst\]). This is the notion of a quantum moment map introduced by Xu [@Xuqmom], which is an deformed analog of equation (\[JLiemorph\]). If $M$ is a symplectic manifold with a Hamiltonian action of a Lie group $G$ with moment map $J:M\to \mathfrak g^$ then a quantum moment map* is a linear map ${\mbox{\boldmath{$J$}}}:\mathfrak g\to \mathcal C^\infty(M){[[\nu]]}$, such that ${\mbox{\boldmath{$J$}}}(\xi)=J(\xi)+\sum_{i\ge 1}\nu^i\,J_i(\xi)$ and we have $$\begin{aligned} \label{qmomprop} {\mbox{\boldmath{$J$}}}(X){\mbox{\boldmath{$J$}}}(Y)-{\mbox{\boldmath{$J$}}}(Y) {\mbox{\boldmath{$J$}}}(X)=\nu{\mbox{\boldmath{$J$}}}([X,Y])\quad \forall X,Y\in \mathfrak g.\end{aligned}$$ If we have found such a ${\mbox{\boldmath{$J$}}}$, then we say that $$ is quantum covariant with moment map ${\mbox{\boldmath{$J$}}}$. It follows that the linear map ${\mbox{\boldmath{$\operatorname{L}$}}}$ which associates to $X\in \mathfrak g$ the operator ${\mbox{\boldmath{$\operatorname{L}$}}}_X:=\frac{1}{\nu}{\operatorname{ad}}_({\mbox{\boldmath{$J$}}}(X))$ makes $\mathcal C^\infty(M){[[\nu]]}$ into a $\mathfrak g$-module. Note that we do not assume that this representation has to coincide with the representation given by the classical moment map (the latter condition is defined to be part of the data, e.g., in [@Xuqmom; @Neubahns]). The existence and uniqueness question for quantum moment maps has been discussed in detail so far merely for the case when these representations coincide, see [@Neubahns] (presumably most of the statements generalize somhow to the above situation). It is clear that one has always the freedom to add to a quantum moment map a Lie algebra 1-cocycle with values in the center of the algebra $(\mathcal C^\infty(M){[[\nu]]},)$. Besides, a quantum moment map gives rise to a ring homomorphism $U\mathfrak g_\nu\to \mathcal C^\infty(M){[[\nu]]}$. A general formalism to treat quantum reduction for such change of ring maps has been proposed in [@Sevost]. The classical BFV-construction {#BFVchapter} ============================== In this section we collect some results and techniques, which are related to the classical BFV-construction. First we recall the basic rules of multilinear super-algebra. We recall the notion of graded Poisson/Gerstenhaber algebras and review the derived bracket construction introduced by Koszul and Kosmann-Schwarzbach [@KosSchw96]. We study (generalized) graded manifolds in the setting of graded Lie-Rinehart pairs. We propose a method how to adjoin momenta for the ‘antighost variables’. We generalize the construction of the Rothstein Poisson-bracket to generalized graded manifolds, and discuss thereby the finitely and infinitely generated case seperately. We give a criterion how to check the acyclicity of a Koszul complex of an analytic map over the ring of smooth functions and show that in this situation there are continuous contracting homotopies. We apply the criterion to our list of examples. We introduce the notion of a projective Koszul-Tate resolution. We generalize the ‘Existence of the BRST-charge’-theorem to the vector bundle setting. We discuss the special cases of coisotropic submanifolds and of moment maps which satisfy the generating and the complete intersection hypothesis. Multilinear super-algebra ------------------------- Let $A$ be a commutative $\mathbbm K$-algebra and $\Gamma$ be one of the abelian groups $\mathbbm Z_2=\mathbbm Z/2 \mathbbm Z$ or $\mathbbm Z$. Since $A$ is commutative there is no distinction between left and right modules. The tensor product and the space of morphisms of two modules are $A$-modules in a canoncical way. We would like to recall the abelian tensor category ${{A}\mathbf{-Mod}^{\Gamma}}$ of $\Gamma$-graded $A$-modules. Objects in ${{A}\mathbf{-Mod}^{\Gamma}}$ are $\Gamma$-graded $A$-modules $\mathcal V=\oplus_{n\in \Gamma} \mathcal V^k$. Elements of $v\in\mathcal V$ such that $v\in V^k$ for some $k\in \Gamma$ are said to be homogeneous of degree $\|v\|:=k$. The sign of a homogeneous element $v$ is defined to be $(-1)^{\|v\|}$. Homogeneous elements with sign 1 are said to be even* and homogeneous elements with sign -1 are said to be odd. Given two graded $A$-modules $\mathcal V$ and $\mathcal W$ the space of linear maps of degree $i\in \Gamma$ is defined to be $${\operatorname{Hom}}^i_{{{A}\mathbf{-Mod}^{\Gamma}}}(\mathcal V,\mathcal W)=\{\varphi\in{\operatorname{Hom}}_{A}(\mathcal V, \mathcal W)\:\|\: \varphi(\mathcal V^j)\subset \mathcal W^{j+i}\quad\forall j\in \Gamma\}.$$ The space of morphisms in ${{A}\mathbf{-Mod}^{\Gamma}}$ between two graded $A$-modules $\mathcal V$ and $\mathcal W$ is defined to be $${\operatorname{Hom}}_{{{A}\mathbf{-Mod}^{\Gamma}}}(\mathcal V,\mathcal W):=\bigoplus_{i\in \Gamma}{\operatorname{Hom}}^i_{{{A}\mathbf{-Mod}^{\Gamma}}}(\mathcal V,\mathcal W).$$ Clearly, the Hom-sets are $A$-modules and the composition of morphisms is $A$-linear. Moreover, every morphism has a kernel and a cokernel. The tensor product $\mathcal V\otimes \mathcal W:=\oplus _{n\in \Gamma}(\mathcal V\otimes \mathcal W)^n$ of two objects $\mathcal V$ and $\mathcal W$ in ${{A}\mathbf{-Mod}^{\Gamma}}$ is the direct sum of its homogenous components $$(\mathcal V\otimes \mathcal W)^n:=\bigoplus_{i,j\in \Gamma,\:i+j=n}\mathcal V^i\otimes _A \mathcal W^j.$$ The tensor product is a biadditive bifunctor in ${{A}\mathbf{-Mod}^{\Gamma}}$. The commutativity morphism is given by $$\begin{aligned} \tau_{\mathcal V,\mathcal W}:\mathcal V\otimes \mathcal W\to\mathcal W\otimes \mathcal V, \quad v\otimes w\mapsto (-1)^{\|v\|\|w\|}w\otimes v \end{aligned}$$ for homogeneous $v$ and $w$. Together with the obvious associativity morphism $\sigma_{\mathcal U,\mathcal V,\mathcal W}:(\mathcal U\otimes \mathcal V)\otimes \mathcal W\to \mathcal U\otimes(\mathcal V\otimes\mathcal W)$ it satisfies the triangle, pentagon and the hexagon axiom (for details see e.g. [@MacLaneWorking]). Note that there is a forgetful functor of tensor categories ${{A}\mathbf{-Mod}^{\mathbbm Z}}\to {{A}\mathbf{-Mod}^{\mathbbm Z_2}}$ which associates to a $\mathbbm Z$-graded $A$-module $\mathcal V=\oplus_{i\in \mathbbm Z} \mathcal V^i$ the $\mathbbm Z_2$-graded $A$-module $\mathcal V_{\underline 0}\oplus\mathcal V_{\underline 1}$, where $\mathcal V_{\underline 0}:=\oplus_{i\text{ even}} \mathcal V^i$ and $\mathcal V_{\underline 1}:=\oplus_{i\text{ odd}} \mathcal V^i$. For every $j\in \mathbbm Z$ the shift ${{A}\mathbf{-Mod}^{\mathbbm Z}}\to{{A}\mathbf{-Mod}^{\mathbbm Z}}$, $\mathcal V\mapsto \mathcal V[j]$ is given by $$\begin{aligned} \mathcal V[j]^i:=\mathcal V^{i+j} \end{aligned}$$ for all $i\in \mathbbm Z$. It is compatible with the tensor product in the following sense: $\mathcal V[i]\otimes \mathcal W[j]=\mathcal V\otimes \mathcal W[i+j]$. The canonical map $\mathcal V\to \mathcal V[j]$ has degree $-j$. The analog of the shift functor $[1]$ in ${{A}\mathbf{-Mod}^{\mathbbm Z_2}}$ is the parity change. By considering a graded $A$-module as a complex with zero differential the category ${{A}\mathbf{-Mod}^{\mathbbm Z}}$ is a full subcategory of the category of complexes ${{A}\mathbf{\mathbf{-Compl}}}$ over $A$. Here we adopt the convention that maps of chain complexes (which will be cochain complexes if not otherwise specified) may carry a nonzero degree. On the other hand, there is a forgetful functor ${{A}\mathbf{\mathbf{-Compl}}}\to {{A}\mathbf{-Mod}^{\mathbbm Z}}$. Note that ${{A}\mathbf{\mathbf{-Compl}}}$ has also the structure of a tensor category. The differential $d_{X\otimes Y}$ on the tensor product $X\otimes Y$ (understood in ${{A}\mathbf{-Mod}^{\mathbbm Z}}$) of the complexes $(Y,d_Y)$ and $(X,d_X)$ is given by the formula $$\begin{aligned} d_{X\otimes Y}(x\otimes y):=d_X x \otimes y+(-1)^{\|x\|}x\otimes d_Y y\end{aligned}$$ for homogeneous $x\in X$ and $y\in Y$. The forgetful functor is a functor of tensor categories. Furthermore, the space of graded $A$-linear morphisms ${\operatorname{Hom}}_{{{A}\mathbf{-Mod}^{\mathbbm Z}}}(X,Y)$ between two complexes $(X,d_X)$ and $(Y,d_Y)$ is in a natural way a complex with differential: $$\begin{aligned} D(\varphi):=d_Y \varphi-(-1)^{\|\varphi\|}\varphi\: d_X\end{aligned}$$ for homogeneous $\varphi\in{\operatorname{Hom}}_{{{A}\mathbf{-Mod}^{\mathbbm Z}}}(X,Y)$. The chain maps ${\operatorname{Hom}}^0_{{{A}\mathbf{\mathbf{-Compl}}}}(X,Y)$ are precisely the $0$-cycles of this complex. Two chain maps are homotopic iff they are homologous. The shift functor extends to ${{A}\mathbf{\mathbf{-Compl}}}$ by setting $d_{X[j]}:=(-1)^j d_X$. By reinterpreting the structural diagrams of the basic algebraic structures, such as that of a commutative algebra, Lie algebra, module etc., these notions straightforwardly translate into the language of tensor categories. For instance, super-commutative $\mathbbm K$-algebras and $\mathbbm K$-Lie algebras are nothing but commutative algebras and Lie algebras in the category ${{{\mathbbm K}\mathbf{-Mod}^{\mathbbm Z_2}}}$, respectively. A differential graded associative (Lie) algebra over $\mathbbm K$ is a associative (Lie) algebra in ${{\mathbbm K}\mathbf{\mathbf{-Compl}}}$. Given an object $\mathcal V$ in ${{A}\mathbf{-Mod}^{\Gamma}}$ we can define the tensor algebra $T_A\mathcal V:=\oplus_{k\ge 0} T_A^k \mathcal V:=\oplus_{k\ge 0} \mathcal V^{\otimes k}$ generated by $\mathcal V$. This is an associative algebra in ${{A}\mathbf{-Mod}^{\Gamma}}$, which carries an additional $\mathbbm Z$-grading given by tensor power. The symmetric algebra $S_A\mathcal V=T_A\mathcal V/<v\otimes w-(-1)^{\|v\|\|w\|}w\otimes v>$ generated by $\mathcal V$ is obtained by dividing out the two sided ideal generated by expressions of the form $v\otimes w-(-1)^{\|v\|\|w\|}w\otimes v$ for homogeneous $v,w \in \mathcal V$. Similarly, the Gra[ß]{}mann algebra generated by $\mathcal V$ is given by the quotient $\wedge_A\mathcal V=T_A\mathcal V/<v\otimes w+(-1)^{\|v\|\|w\|}w\otimes v>$. In addition to their natural $\Gamma$-grading, which is refered to as total degree, $S_A\mathcal V$ and $\wedge_A\mathcal V$ inherit from the tensor algebra $\mathbbm Z$-grading, which we refer to as tensor power. It is well known that $T_A\mathcal V$, $S_A\mathcal V$ and $\wedge_A\mathcal V$ are actually bialgebras in ${{A}\mathbf{-Mod}^{\Gamma}}$. In all three cases the comultiplication $\Delta$ is uniquely determined by the requirement that $\mathcal V$ is the space of primitives, i.e., $\Delta(v)=v\otimes 1+1\otimes v$ for all $v\in \mathcal V$. The symmetric group $S_n={\operatorname{Aut}}_{Set}(\{1,\dots,n\})$ acts on the set of multiindices $\mathbbm Z^n$ from the right by $\sigma(x_1,\dots,x_n):=(x_{\sigma(1)},\dots,x_{\sigma(n)})$. Given such a multiindex $x=(x_1,\dots,x_n)$ the Koszul sign ${\operatorname{sign}}(\tau_{i,i+1},x)$ of the transposition $\tau_{i,i+1}\in S_n$, which interchanges $i$ and $i+1$ is defined to be $(-1)^{x_ix_{i+1}}$. It is well known that according to the rule $$\begin{aligned} {\operatorname{sign}}(\sigma\circ\sigma',x)={\operatorname{sign}}(\sigma,\sigma'(x)){\operatorname{sign}}(\sigma',x)\end{aligned}$$ the Koszul sign unambiguously extends to a map ${\operatorname{sign}}(\sigma,x):S_n\to\{\pm1\}$. Instead of the Koszul sign of $x=(x_1,\dots,x_n)$ one may also consider the Koszul sign of the shifted multiindex $x[1]:=(x_1+1,\dots,x_n+1)$. Derived brackets, Gerstenhaber algebras, etc. {#derbrsection} --------------------------------------------- A left Leibniz (or Loday) bracket of degree $n$ on a $\mathbbm Z$-graded vector space $L=\oplus_i L^i$ is a graded linear map $[\:,\:]:L\otimes L\to L$ of degree $n$, such that the following Leibniz rule holds for all homogeneous $a,b,c \in L$: $$\begin{aligned} \label{leibniz} [a,[b,c]]=[[a,b],c]+(-1)^{(\|a\|+n)(\|b\|+n)}[b,[a,c]].\end{aligned}$$ If the bracket $[\:,\:]$ is graded antisymmetric, i.e., $[a,b]=-(-1)^{\|a\|+n)(\|b\|+n)}[b,a]$, then it is a Lie bracket of degree $n$; in other words:, $L[-n]$ is a graded Lie algebra. If the space $L$ is, in addition, a super-commutative $\mathbbm Z$-graded algebra with multiplication $\mu(a\otimes b)=ab$, such that the following Leibniz rule holds $$\begin{aligned} \label{leibgerst} [a,bc]=[a,b]c+(-1)^{\|b\|(\|a\|+n)}b[a,c],\end{aligned}$$ then we say that $[\:,\:]$ is a Poisson bracket of degree $n$. Under these circumstances we also say: $\left(L,\mu,[\:,\:]\right)$ is an $n$-Poisson algebra. Thus a $0$-Poisson algebra is just a $\mathbbm Z$-graded Poisson algebra and a $(-1)$-Poisson algebra is what is usually called a Gerstenhaber algebra. \[derKS\] Let $(L,[\:,\:])$ be a $\mathbbm Z$-graded left Leibniz algebra, where the bracket $[\:,\:]$ is of degree $n$, and let $d:L^\bullet\to L^{\bullet+m}$ a differential (i.e., $d^2=0$) which is a derivation of $[\:,\:]$, i.e., $$\begin{aligned} \label{mder} d[a,b]=[da,b]+(-1)^{m(\|a\|+n)}[a,db]\end{aligned}$$ for all homogeneous $a,b\in L$. 1. Then the derived bracket $$\begin{aligned} [a,b]_d:=(-1)^{m(n+\|a\|)+1}[da,b]\end{aligned}$$ is a left Leibniz bracket of degree $n+m$. Moreover, $d$ is a derivation of $[\:,\:]_d$. 2. \[derbr1\]If, more specifically, $[\:,\:]$ is a Lie bracket of degree $n$ and $L_0$ is an abelian subalgebra of $L$, such that $[L_0,L_0]_d\subset L_0$ and $m$ is odd, then $[\:,\:]_d$ is a Lie bracket on $L_0$ of degree $n+m$. 3. \[derbr2\] There is an natural left Leibniz bracket of degree $n+m$ on $L/dL$: $$\begin{aligned} \label{coex} [\bar{a},\bar{b}]_d:=(-1)^{m(n+\|a\|)+1}\overline{[d{a},{b}]},\end{aligned}$$ where the bar indicates taking classes in $L/dL$. The left Leibniz algebra $L/dL$ contains the homology space $\mathrm H=\mathrm Z/dL$, where $\mathrm Z=\ker(d)$, naturally as a graded subalgebra. If the original bracket is a graded Lie bracket of degree $n$ and $m$ is odd, then the derived bracket $[\:,\:]_d$ is a Lie bracket of degree $n+m$. In this case $\big(L/dL,[\:,\:]_d\big)$ is also called the Lie algebra of co-exact elements of $L$. If the original bracket is a Poisson bracket of degree $n$ and $d$ is an odd derivation of this Poisson structure, then the derived bracket defined in \[derbr1\]. is a Poisson bracket of degree $n+m$. Moreover, the derived bracket of equation (\[coex\]) restricted to the homology subalgebra $\mathrm H=\mathrm Z/dL$ a Poisson bracket of degree $n+m$. We proof these statements along the lines of [@KosSchw96]. First of all, let us write ${\operatorname{ad}}: L\to{\operatorname{End}}_{{{\mathbbm K}\mathbf{-Mod}^{\mathbbm Z}}}(L)$, $a\mapsto [a,\:]$ for the adjoint action of $L$ on itself. The degree of the endomorphism ${\operatorname{ad}}(a)$ is $\|a\|+n$. We rewrite equation (\[leibniz\]) as $[{\operatorname{ad}}(a),{\operatorname{ad}}(b)]={\operatorname{ad}}\left({\operatorname{ad}}(a)b\right)$, this time $[\:,\:]$ denotes the super-commutator of graded endomorphisms. Analogously, we rewrite equation (\[mder\]) as $[d,{\operatorname{ad}}(a)]={\operatorname{ad}}(da)$. For the adjoint action corresponding to the derived bracket $[\:,\:]_d$ we write ${\operatorname{ad}}_d(a):=[a,\:]_d$. The degree of the endomorphism ${\operatorname{ad}}_d(a)$ is $\|a\|+n+m$. In fact, using super-commutators we have ${\operatorname{ad}}_d(a)=[{\operatorname{ad}}(a),d]$. An easy calculation using the Jacobi identitiy for the super-commutator and $[d,d]=2d^2=0$ yields that $d$ is a derivation of the derived bracket: $$\begin{aligned} [d,{\operatorname{ad}}_d(a)]=\big[d,[{\operatorname{ad}}(a),d]\big]=\big[[d,{\operatorname{ad}}(a)],d\big]+0=[{\operatorname{ad}}(da),d]={\operatorname{ad}}_d(da).\end{aligned}$$ Let us proof that $[\:,\:]_d$ is a left Leibniz bracket: $$\begin{aligned} \lefteqn{\left[{\operatorname{ad}}_d(a),{\operatorname{ad}}_d(b)\right]-{\operatorname{ad}}_d\left({\operatorname{ad}}_d(a)b\right)}\\ &=&\big[[{\operatorname{ad}}(a),d],[{\operatorname{ad}}(b),d]\big]-\big[{\operatorname{ad}}\big([{\operatorname{ad}}(a),d]b\big),d\big]\\ &=& (-1)^{m(\|a\|+\|b\|)}\big[{\operatorname{ad}}(da),{\operatorname{ad}}(db)\big]+(-1)^{m(\|a\|+n)}\big[{\operatorname{ad}}\big({\operatorname{ad}}(da)b\big),d\big]\\ &=&(-1)^{m(\|a\|+\|b\|)}{\operatorname{ad}}\big({\operatorname{ad}}(da)db\big)-(-1)^{m(\|a\|+n)}(-1)^{m(m+2n+\|a\|+\|b\|)}{\operatorname{ad}}\big(d{\operatorname{ad}}(da)b\big)\\ &=&(-1)^{m(\|a\|+\|b\|)}{\operatorname{ad}}\big({\operatorname{ad}}(da)db\big)-(-1)^{m(\|b\|+n+1)}(-1)^{m(m+n+\|a\|)}{\operatorname{ad}}\big({\operatorname{ad}}(da)db\big)=0.\end{aligned}$$ In order to proof item \[derbr1\].) we only need to show that the restriction of the derived bracket to the subalgebra is graded antisymmetric. In fact, since $L_0$ is an abelian with respect to the graded Lie bracket $[\:,\:]$ we have: $$\begin{aligned} \label{triveq} 0=d[a,b]=[da,b]+(-1)^{m(\|a\|+n)}[a,db]=[da,b]-(-1)^{(\|a\|+n)(\|b\|+n)}[db,a].\end{aligned}$$ Thus we have to see that $(-1)^{m(\|a\|+n)+1}[da,b]=(-1)^{m(\|a\|+n)+1}(-1)^{(\|a\|+n)(\|b\|+n)}[db,a]$ coincides with $-(-1)^{m(\|b\|+n)+1}(-1)^{(\|a\|+n+m)(\|b\|+n+m)}[db,a]$. These signs match iff $m$ is odd. In order to proof item \[derbr2\].) we observe that $dL$ is a two-sided ideal in $L$: $[dL,L]_d\subset dL\supset [L,dL]_d$. We conclude that the derived bracket of equation (\[coex\]) is well-defined on $L/dL$. Moreover, since the space of cycles $Z=\ker(d)$ is obviously a subalgebra of $(L,[\:,\:]_d)$, and hence $H=Z/dL$ is a subalgebra of $(L/dL,[\:,\:]_d)$. If $[\:,\:]$ is in fact a Lie bracket of degree $n$ and $m$ is odd we again have, due to equation (\[triveq\]), that the derived bracket is Lie of degree $n+m$. Finally, let us address the question what happens in the Poisson case. Let $\mu: L\otimes L\to L, a\otimes b\mapsto ab$ be a super-commutative multiplication of degree zero. Accordingly, the operator of left multiplication $\lambda: L\to{\operatorname{End}}_{{{\mathbbm K}\mathbf{-Mod}^{\mathbbm Z}}}(L)$, $\lambda(a)b:=ab$ for $b\in L$, is of degree zero. We impose the Leibniz rules $[d,\lambda(a)]=\lambda(da)$ (i.e., $d$ is a derivation of $\mu$) and $[{\operatorname{ad}}(a),\lambda(b)]=\lambda({\operatorname{ad}}(a)b)$, which is equivalent to (\[leibgerst\]). The analogue of equation (\[leibgerst\]) for the derived bracket works out as follows $$\begin{aligned} \big[{\operatorname{ad}}_d(a),\lambda(b)\big]&=&\big[[{\operatorname{ad}}(a),d],\lambda(b)\big]=\big[{\operatorname{ad}}(a),[d,\lambda(b)]\big]+(-1)^{m\|b\|}\big[[{\operatorname{ad}}(a),\lambda(b)],d\big]\\ &=&\big[{\operatorname{ad}}(a),\lambda(db)\big]+(-1)^{m\|b\|}\big[\lambda({\operatorname{ad}}(a)b),d\big]\\ &=&\lambda({\operatorname{ad}}(a)db)-(-1)^{m\|b\|}(-1)^{m(\|a\|+\|b\|+n)}\lambda(d{\operatorname{ad}}(a)b)\\ &=&\lambda\big(({\operatorname{ad}}(a)d-(-1)^{m(\|a\|+n)}d{\operatorname{ad}}(a))b\big)=\lambda({\operatorname{ad}}_d(a)b).\end{aligned}$$ Consequently, if the abelian subalgebra $L_0$ of item \[derbr1\].) is also a subalgebra for the multiplication $\mu$, then the derived bracket is a Poisson bracket. In general, the subspace $dL\subset L$ is not an ideal for the multiplication $\mu$, but $dL\subset \mathrm Z$ actually is an ideal! For every $x\in dL\,\mathrm Z$ can be written as $x=\sum_{i=1}^k da_i\:b_i=d\big(\sum_i^k a_ib_i\big)$ for some $a_1,\dots a_k\in L$ and $b_1,\dots b_k\in Z$. \ There is yet another way to produce a derived bracket, which generalizes the construction of item \[derbr2\].) of the above theorem. Let $(L,\mu,[\:,\:],d)$ be a Poisson algebra with a bracket of degree $n$ and differential of degree $m$ (assumed to be $\pm 1$), such that $d$ is a derivation of the supercommutative multiplication $\mu(a\otimes b)=ab$ and of the bracket $[\:,\:]$. Under these circumstances we also we call $(L,\mu,[\:,\:],d)$ a differential graded n-Poisson algebra. We will call a graded subspace $K\subset L$ a coisotropic ideal if the following conditions are true: $$\begin{aligned} K\cdot L&\subset&K\\ \:[K, K]&\subset& K\\ dK&\subset&K.\end{aligned}$$ Given such a coisotropic ideal $K\subset L$ the quotient space $V:=L/K$ is a differential graded commutative algebra. In particular $V$ is a complex. For the differential we write just $d$ and for the graded spaces of cycles and boundaries of this complex we will write $\mathrm Z\,V$ and $dV$, the homology is $\mathrm H\,V:=\mathrm Z\,V/dV$. Let us write $\widetilde a$ for the image of $a\in L$ under the projection onto $V$. Note that a cycle in $V$ is a class $\widetilde a$ of an $a\in V$ such that $da\in K$. \[Reduced bracket\]\[redderbr\] Let $a,b\in L$ represent the cycles $\widetilde a,\widetilde b\in \mathrm Z\,V$ then $$\begin{aligned} [\widetilde a,\widetilde b]_{d,K}:=(-1)^{m(\|a\|+n)+1}\widetilde{[da,b]}+dV\end{aligned}$$ is a well defined class in $\mathrm H\,V$. $[\:,\:]_{d,K}$ is a Poisson bracket of degree $n+m$ on the super-commutative $\mathbbm Z$-graded algebra $\mathrm H\, V$. In order to see, that the bracket is in fact well defined, it is more comfortable to write the cycles as $\widetilde{a}=a+K$, $\widetilde{b}=b+K$. Now we get $$\begin{aligned} [d(a+K),b+K]=[da,b]+[dK,K]+[da,K]+[dK, b].\end{aligned}$$ The second term is obviously in $K$. The third term is in $K$ since $\widetilde{a}$ is a cycle. Finally, writing the last term as $\pm d[K,b]\pm[K,db]$, it is in $K$ up to a boundary since $\widetilde{b}$ is a cycle. Note that $[da,b]$ is a cycle since $d[da,b]=\pm[da,db]\subset[K,K]$. Again, $dV$ is an ideal in $\mathrm Z\,V$ with respect to the commutative multiplication. The remaining statements follow from the proof of Theorem \[derKS\]. \ In any of the cases when the derived bracket $[\:,\:]_d$ is a Poisson bracket of degree $n+m$ the opposite bracket $[a,b]^{{\operatorname{opp}}}_d:=(-1)^{(\|a\|+n+m)(\|b\|+n+m)}[b,a]_d$, for homogeneous $a,b \in L$, is a Poisson bracket of degree $n+m$ as well (see Appendix \[oppPoiss\]). Since we have $$\begin{aligned} [a,db]&=&-(-1)^{(\|b\|+m+n)(\|a\|+n)}[db,a]=(-1)^{m(\|b\|+n)}(-1)^{(\|b\|+m+n)(\|a\|+n)}[b,a]_d\\ &=&(-1)^{(\|a\|+n+m)(\|b\|+n+m)}(-1)^m [b,a]_d,\end{aligned}$$ we conclude that the opposite derived bracket is just $$\begin{aligned} [a,b]^{{\operatorname{opp}}}_d:=(-1)^m[a,db].\end{aligned}$$ An important example of Poisson bracket of degree $-1$ is given by the Schouten-Nijenhuis bracket on the symmetric algebra $S_A(L[-1])$ of a Lie-Rinehart pair $(A,L)$ in the category ${{\mathbbm K}\mathbf{-Mod}^{\mathbbm Z}}$. For convenience of the reader we have collected some basic material on Lie-Rinehart pairs in Appendix \[LieRinehart\]. We caution the reader that for certain applications that we have in mind, this notion of polyvector field is a little too restricitive. \[SchoutenNijenhuis\] If $(A,L)$ is a Lie-Rinehart pair in the tensor category ${{\mathbbm K}\mathbf{-Mod}^{\mathbbm Z}}$, then there is a unique Lie bracket $[\:,\:]$ on $S_A(L[-1])$ extending the structure maps on $(A,L)$ and making $S_A(L[-1])$ into a Gerstenhaber algebra. More precisely, the bracket is the unique Gerstenhaber bracket determined by the requirements 1. $A\subset S_A(L[-1])$ is an abelian subalgebra, i.e., $[A,A]=0$, 2. $[X,a]=-(-1)^{(\|X\|+1)\|a\|}[a,X]=X(a)$, 3. if $[\:,\:]'$ denotes, for the moment, the bracket in $L$ then we have $[X,Y]'=[X,Y]$ for all homogeneous $X,Y\in L$ and $a\in A$. Moreover, any morphism from $(A,L)$ to another Lie Rinehart pair $(A',L')$ in ${{\mathbbm K}\mathbf{-Mod}^{\mathbbm Z}}$ extends uniquely to a morphism of Gerstenhaber algebras $S_A(L[-1])\to S_{A'}(L'[-1])$. Presumably, there is some simple ‘operadic’ reason for the statement. Nonetheless, we have included an elementary proof in the appendix \[Schoutenproof\]. \ We will call this Gerstenhaber algebra the algebra of polyvector fields of the Lie-Rinehart pair $(A,L)$. We will use for it the notation $\mathfrak X(A,L):=S_A(L[-1])$. It carries two degrees: the total degree and the tensor power. For the tensor power we will also use the term arity. For the subspace of arity $k$ we will write $${\mathfrak X}^k(A,L)=S^k_A(L[-1]).$$ If $a\in A$ and $X_1,\dots,X_i\in L$ are homogeneous, then the total degree of a monomial $a X_1\dots X_i$ is $\|a\|+\|X_1\|+\dots+\|X_i\|+i$. The space of polyvector fields of total degree $k$ will be written as $${\mathfrak X}(A,L)^k.$$ Of course, the degree for which ${\mathfrak X}(A,L)$ is a Gerstenhaber algebra is the total degree. However, we also have an inclusion $[{\mathfrak X}^k(A,L),{\mathfrak X}^l(A,L)]\subset {\mathfrak X}^{k+l-1}(A,L)$ for all $k,l\ge 0$. For the important special case, where $A=\mathcal C^\infty(M)$ is the ring of smooth functions on some manifold $M$ and $L=\Gamma^\infty(M,TM)$ is the space of vector fields we will use the shorthand ${\mathfrak X}(M)$. If $\Pi\in{\mathfrak X}^2(M)$ is a Poisson tensor, i.e., $[\Pi,\Pi]=0$, then $$\begin{aligned} \delta_{\Pi}:=[\Pi,\:]:{\mathfrak X}^\bullet(M)\to {\mathfrak X}^{\bullet+1}(M)\end{aligned}$$ is a codifferential. The cohomology of this differential is known as Lichnerowicz-Poisson cohomology. The Poisson bracket corresponding to $\Pi$ is just the derived bracket of $-\delta_\Pi$, or, equivalently, the opposite derived bracket of $\delta_\Pi$. Here, in the ungraded case, we will use the standard notations and identify ${\mathfrak X}^\bullet(M)$ with the space of sections $\Gamma^\infty(M,\wedge^\bullet TM)$ of the Grassmann-algebra bundle of the tangent bundle. The Lichnerowicz-Poisson cohomological complex $({\mathfrak X}^\bullet(M),\delta_\Pi)$ can alternatively be interpreted as the space of cochains of Lie algebroid cohomology of the Lie algebroid $(T^M,\#_\Pi,\{\:,\:\}_{KB})$ associated to $\Pi$. The anchor $\#_\Pi:T^M\to TM$ of this Lie algebroid is given the ‘musical map’ $\#_\Pi(\alpha):=\alpha^\#:=i(\alpha)\Pi$, where $i$ denotes the insertation derivation. The bracket $\{\:,\:\}_{KB}$ is the Koszul-Brylinski bracket which is given by the formula $\{\alpha,\beta\}_{KB}=L_{\alpha^\#}\beta-L_{\beta^\#}\alpha-d\Pi(\alpha,\beta)$. Given a commutative algebra $A$ in ${{\mathbbm K}\mathbf{-Mod}^{\mathbbm Z}}$ there is still another Gerstenhaber algebra $\mathrm{Der}(A)=\oplus_{n\ge 0}\mathrm{Der}^n(A)$, the algebra of multiderivations, which is related to the algebra of polyvector fields. A multiderivation $D\in \mathrm{Der}^n(A)$ of arity $n$ is by definition a graded linear map $D:A^{\otimes n}\to A$ which is graded symmetric in the following sense $$\begin{aligned} \label{symmprop} D(a_1,\dots,a_i,a_{i+1},\dots,a_n)=(-1)^{(\|a_i\|+1)(\|a_{i+1}\|+1)}D(a_1,\dots,a_{i+1},a_{i},\dots,a_n)\end{aligned}$$ for $i=1,\dots,n-1$ and which is a derivation in every argument, i.e., $$\begin{aligned} \label{derprop} D(a_1,\dots,a_{n-1},ab)=D(a_1,\dots,a_{n-1},a)b+ (-1)^{\|a\|\|b\|}D(a_1,\dots,a_{n-1},b)a\end{aligned}$$ for all homogeneous $a,b,a_1,\dots,a_n\in A$. If $D$ is a multiderivation of arity $n$ and degree $k$, then we define the total degree of $D$ to be $\|D\|:=n+k$. If $(A,L)$ is a Lie-Rinehart pair in ${{A}\mathbf{-Mod}^{\mathbbm Z}}$, then a polyvector field $X\in {\mathfrak X}^n(A,L)$ can be interpreted as a multiderivation. More precisely, we let $X$ act on $(a_1,\dots,a_n)$ by taking iterated commutators $$\begin{aligned} \label{itcom} B_X(a_1,\dots,a_n):= X(a_1,\dots,a_n):=[\dots[[X,a_1],a_2,],\dots ,a_n].\end{aligned}$$ Let us now define an analogue $\cup$ of the super-commutative product of polyvector fields $$\begin{aligned} \label{cupform} D\cup E(a_1,\dots,a_{k+l})&:=&\sum_{\sigma \in S_{k,l}}{\operatorname{sign}}(\sigma,\|a\|[1])(-1)^{\|E\|(\|a_{\sigma(1)}\|+\|a_{\sigma(2)}\|+\dots +\|a_{\sigma(k)}\|+k)}\nonumber\\ &&\qquad\qquad D(a_{\sigma(1)},\dots,a_{\sigma(k)})E(a_{\sigma(k+1)},\dots, a_{\sigma(k+l)}),\end{aligned}$$ where $D$ and $E$ are homogeneous multiderivations of arity $k$ and $l$, respectively, and $S_{k,l}$ is the set of $(k,l)$-unshuffle permutations, i.e., $\sigma(1)<\sigma(2)<\dots<\sigma(k)$ and $\sigma(k+1)<\sigma(k+2)<\dots<\sigma(k+l)$. In general, by $\|a\|:=(\|a_1\|,\dots,\|a_n\|)\in \mathbbm Z^{n}$ we mean the multiindex of the $n$-tuple $(a_1,\dots,a_n)$ of homogeneous elements of the graded space $A$. Next we define a bilinear operation $\bullet:\mathrm{Der}^k(A)\times\mathrm{Der}^l(A)\to\mathrm{Der}^{k+l-1}(A)$ as follows $$\begin{aligned} \label{bullform} D\bullet E(a_1,\dots, a_{k+l-1}):=\sum_{\sigma\in S_{l,k-1}}{\operatorname{sign}}(\sigma,\|a\|[1])D(E(a_{\sigma(1)},\dots,a_{\sigma(l)}),a_{\sigma(l+1)},\dots, a_{\sigma(k+l-1)}).\end{aligned}$$ The Richardson-Nijenhuis bracket [@RichNij] of two multiderivations is defined by $$\begin{aligned} \label{RNformula} [D,E]_{RN}:=D\bullet E-(-1)^{(\|D\|-1)(\|E\|-1)}E\bullet D.\end{aligned}$$ \[multiderisgerst\]$\left({\operatorname{Der}}(A),\cup,[\:,\:]_{RN}\right)$ is a Gerstenhaber algebra. The iterated bracket map $B$ (cf. equation (\[itcom\])) is a morphism of Gerstenhaber algebras, i.e., for all $X,Y \in {\mathfrak X}(A,L)$ we have $B_{XY}=B_X\cup B_Y$ and $B_{[X,Y]}=[B_X,B_Y]_{RN}$. It is well known (see e.g. [@Manchon Théorème III.2.1]) that the space $$\mathscr L:=\prod_n{\operatorname{Sym}}_{\mathbbm K}^n(A[1],A[1])$$ of graded $\mathbbm K$-linear maps $A[1]^{\otimes n}\to A[1]$ which have the symmetry property (\[symmprop\]) above is isomorphic to the graded Lie algebra ${\operatorname{Coder}}(S_{\mathbbm K}(A[1]))$ of coderivations of the free symmetric $\mathbbm K$-coalgebra cogenerated by $A[1]$. More specifically, if $D\in{\operatorname{Sym}}_{\mathbbm K}^n(A[1],A[1])$ then the corresponding coderivation $D$ of $S_{\mathbbm K}(A[1])$ is given by the formula $$\begin{aligned} D(a_1\dots a_m):=\sum_{\sigma\in S_{n,m-n}}{\operatorname{sign}}(\sigma,\|a\|[1])D(a_{\sigma(1)},\dots,a_{\sigma(n)})\:a_{\sigma(n+1)}\dots a_{\sigma(n+m)}.\end{aligned}$$ The induced Lie bracket on $\mathscr L$ is just the bracket of equation (\[RNformula\]). It is not difficult to see that the space of multiderivations ${\operatorname{Der}}(A)$ is in fact a Lie subalgebra of $\mathscr L$. In this way a multiderivation $D$ of degree $\|D\|$ is a coderivation of degree $\|D\|-1$. In order to show the Leibniz rule for $[\:,\:]_{RN}$ it suffices to prove $$\begin{aligned} D\bullet(E\cup F)&=&(D\bullet E)\cup F+(-1)^{(\|D\|-1)\|E\|}E\cup (E\bullet F)\qquad \mbox{ and} \\ (E\cup F)\bullet D&=&E\cup (F\bullet D)+(-1)^{(\|D\|-1)\|F\|}(E\bullet D)\cup F.\end{aligned}$$ These relations can be shown by an unpleasant but straightforward computation. The compatibility of $B$ with the cup product $B_{XY}=B_X\cup B_Y$ is straight forward to check. In order to prove $B_{[X,Y]}=[B_X,B_Y]_{RN}$ one can now use induction over the arities of $X$ and $Y$. To start the induction we need only to check the formulas $B_{[X,Y]}=[B_X,B_Y]_{RN}$ and $[X,a]=[B_X,a]_{RN}$ for $X,Y\in{\mathfrak X}^1(A,L)$ and $a\in A$, which are obviously fulfilled. Generalized graded manifolds {#ggmf} ---------------------------- There are several possible ways to define the notion of a super-manifold (see e.g. [@superkostant; @tuynman; @Eckel]). The most basic version is that of a space of sections $\Gamma^\infty(M,\wedge E)$ of a Gra[ß]{}mann algebra bundle $\wedge E\to M$ of a finite dimensional vector bundle over $M$ seen as a $\mathbbm Z$-graded super-commutative $\mathcal C^\infty(M)$-algebra. It is known that many geometric constructions, such as tangent vectors, the cotangent bundle, the de Rham complex and vector bundles with connections generalize to super-manifolds. In view of the applications we have in mind (cf. Sections \[RothPB\], \[projkostate\] and \[BRSTchargesection\]) we are forced to consider more general $\mathbbm Z$-graded algebras which serve as algebras of super-functions. We will propose a way to define the notion of cotangent space for these algebras and introduce the notion of polyvector fields in the spirit of the preceding section. Besides, we do not think that we have found a general geometric theory for this more general graded algebras (and it may well be that there are important references related to these questions, which we are not aware of). Let $M$ be a smooth manifold and let $V=\oplus_{k \in \mathbbm Z} V^k$ be the direct sum of finite rank vector bundles $V^k$ over $M$. For simplicity, let us always assume that these vector bundles admit simultaneous bundle charts. Let us write $\mathcal V$ for the space $\Gamma^\infty(M,V)$ of smooth sections of $V$. For convenience we use the shorthand notation $A:=C^\infty(M)$. According to the theorem of Serre and Swan we have that $\mathcal V=\oplus_{k\in \mathbbm Z} \mathcal V^k$ is the direct sum of the finitely generated projective $A$-modules $\mathcal V^k=\Gamma^\infty(M,V^k)$. The support of $\mathcal V$ is defined to be $\operatorname{supp}(\mathcal V):=\{k\in\mathbbm Z\mid \mathcal V^k< 0\}$. Assigning to each $\mathcal V^k$ the degree $k$ we will consider $\mathcal V$ as an object in ${{A}\mathbf{-Mod}^{\mathbbm Z}}$. Let $S_A\mathcal V=\oplus_{i\ge 0} S^i_A \mathcal V$ be the symmetric algebra in ${{A}\mathbf{-Mod}^{\mathbbm Z}}$ generated by the $A$-module $\mathcal V$. In fact, $S_A\mathcal V$ is a free commutative algebra in the category ${{A}\mathbf{-Mod}^{\mathbbm Z}}$. It is easy to see, that the $A$-linear derivations ${\operatorname{Der}}_A(S_A\mathcal V)^k={\operatorname{Der}}^1_A(S_A\mathcal V)^k$ of degree $k$ of the $\mathbbm Z$-graded algebra $S_A\mathcal V$ may be identified with the $A$-module $\prod_i(S_A\mathcal V)^{k+i} \otimes_A {\mathcal V^i}^$. In every degree $k$ we have an exact sequence $$\begin{aligned} \label{derseq} 0\to{\operatorname{Der}}_A(S_A\mathcal V)^k\to{\operatorname{Der}}_{\mathbbm K}(S_A\mathcal V)^k\stackrel{D\mapsto D_{\|A}}{\to}{\operatorname{Der}}_{\mathbbm K}(A,S_A\mathcal V)^k\to 0.\end{aligned}$$ Let us now choose for every vector bundle $V^i$ a connection $\nabla^{V^i}$. For the derivation of $S_A\mathcal V$ which gives for every section $v_j\in\mathcal V^j$ and every vector field $X\in {\mathfrak X}(M)$ the value $\nabla^{V^j}_{X}v_j$ we will write $$\begin{aligned} \nabla_{X}=\sum_{j\in \mathbbm Z}\nabla^{V^j}_{X}.\end{aligned}$$ Using this family of connections we can identify ${\operatorname{Der}}_{\mathbbm K}(A,S_A\mathcal V)^k=(S_A\mathcal V)^k\otimes {\operatorname{Der}}_{\mathbbm K}A=(S_A\mathcal V)^k\otimes {\mathfrak X}^1(M)$ with a submodule of ${\operatorname{Der}}_{\mathbbm K}(S_A\mathcal V)^k$ which is complementary to ${\operatorname{Der}}_A(S_A\mathcal V)^k$, i.e., we use the connections to split the sequence (\[derseq\]). More precisely, this split is the $S_A\mathcal V$-linear extension of the map which associates to every vector field $X$ the derivation $\nabla_X$. This complementary subspace to the space of $A$-linear derivations will be refered to as the space of geometric derivations. Using the exact sequence (\[derseq\]) one can show that $\mathrm{Der}_\mathbbm K(S_A\mathcal V)$ is a projective $S_A\mathcal V$-module. Hence, one can use the machinery of Appendix \[LieRinehart\] to define the Lie-Rinehart cohomology of the graded Lie-Rinehart pair $(S_A\mathcal V,\mathrm{Der}_\mathbbm K(S_A\mathcal V))$. The reader may have noticed that, in general, the $A$-linear derivations as well as the geometric derviations appear to be infinite sums. We therefore introduce the space of finite type derivations, which are derivations as above, such that the sums are actually finite sums. In order to give the precise definition let us write the derivations in local coordinates. Let $(x_1,\dots,x_n)$ be local coordinates for $M$ and let ${\xi^{(i)}_{1}},\dots ,{\xi^{(i)}_{\ell_i}}$ be local frames for the bundles $V^i$ for all $i\in \mathbbm Z$. Then a finite type derivations* of degree $k$ is a derivation $X$ which writes locally as $$\begin{aligned} {\sum_{j\in \mathbbm Z}}^\prime \sum_{a=1}^{\ell_j} X_a^j\:{\frac{\partial}{\partial \xi^{(i)}_{a}}}+{\sum_{j\in \mathbbm Z}}^\prime\sum_{i=1}^n X_i\:\nabla^{V^j}_{\frac{\partial}{\partial x^i}},\end{aligned}$$ where the $^\prime$ indicates that all except finitely many summands vanish. In the above formula ${\frac{\partial}{\partial \xi^{(i)}_{a}}}$ is the unique $A$-linear derivation such that$$\begin{aligned} \label{algebraicder} \frac{\partial {\xi^{(j)}_{b}}}{\partial {\xi^{(i)}_{a}}}:={\frac{\partial}{\partial \xi^{(i)}_{a}}}{\xi^{(j)}_{b}}=\delta_a^b\:\delta_i^j\qquad \forall i,j\in\mathbbm Z,\:a=1,\dots \ell_i,\:b=1,\dots,\ell_j.\end{aligned}$$ Of course, the coefficients $X_a^j$ are in $(S_A\mathcal V)^{j+k}$ and the $X_i$ have to be in $(S_A\mathcal V)^k$. We write for the graded space of finite type derivations ${\mathfrak X}^1(S_A\mathcal V)$. It is clear that $(S_A\mathcal V,{\mathfrak X}^1(S_A\mathcal V))$ form a graded Lie-Rinehart pair. Of course, if $\|\operatorname{supp}(\mathcal V)\|<\infty$ then every graded derivation is of finite type. Note that if $\|\operatorname{supp}(\mathcal V)\|=\infty$ the Euler vector field $$\begin{aligned} \sum_{i\in \mathbbm Z} \sum_{a=1}^{\ell_i} {\xi^{(i)}_{a}}\:{\frac{\partial}{\partial \xi^{(i)}_{a}}}\end{aligned}$$ is not a finite type derivation. Analogously, $\nabla_X$ is a finite type derivation if and only if $\|\operatorname {supp}(\mathcal V)\|<\infty$. For the applications we have in mind (cf. Section \[projkostate\]) we have that the module $\mathcal V=\oplus_{k\ge 1}\mathcal V^k$ is positively graded. This entails that the graded components of $S_A\mathcal V$ are actually finitely generated. In fact, the dimensions can be determined recursively from the Poincaré series $$\begin{aligned} \sum_{i\ge 0}\operatorname{rank}((S_A\mathcal V )^i)\:t^i=\prod_{j\ge 1}\big(1-(-t)^j\big)^{(-1)^{j+1}\operatorname{rank}(\mathcal V^j)}.\end{aligned}$$ The product on the right hand side converges in the $t$-adic topology of $\mathbbm Z[[t]]$. By a generalized graded manifold with base manifold $M$, $A:=\mathcal C^\infty(M)$, we mean a Lie-Rinehart pair in the category ${{A}\mathbf{-Mod}^{\mathbbm Z}}$of the form $\big(S_A\mathcal V,{\mathfrak X}^1(S_A\mathcal V)\big)$, where $\mathcal V=\oplus_{k\ge 1}\mathcal V^k$ is positively graded. We will say for short: $S_A \mathcal V$ is a graded manifold. The generalized graded manifold is said to be finitely generated of level $L:=\inf\{j\:\|\: \mathcal V^i=0\:\forall i>j\}$ if $L<\infty$. Otherwise it is said to be infinitely generated. In the finitely generated of level $L$ case we will sometimes write $\mathcal V^{\le L}$ instead of $\mathcal V$. If all the vector bundles $V^k$, $k=1,2,\dots$ are trivial, we will say that the generalized graded manifold is freely generated. For the BFV-construction we also need to adjoin extra variables which are dual to the local frames ${\xi^{(i)}_{1}},\dots ,{\xi^{(i)}_{\ell_i}}$ for the vector bundles $ V^i$, $i=1,2,\dots$. This construction is a graded analog of taking the cotangent bundle of a manifold. Let ${\mathcal V^i}^=\Gamma^\infty(M,{V^i}^)$ be space of sections of the dual vector bundle of $V^i$ with dual local frames ${\xi_{(i)}^{1}},\dots ,{\xi_{(i)}^{\ell_i}}$ (these sections will be also called momenta). The naive guess for an analog of cotangent space for the graded manifold $S_A\mathcal V$ would probably be the $\mathbbm Z$-graded algebra $S_A(\mathcal V\oplus \mathcal V^)=S_A\mathcal V^\otimes S_A\mathcal V$, where $\mathcal V^=\oplus_{i\ge i}{\mathcal V^i}^$. Unfortunately, this proposal does not work for our purposes (i.e. the constructions of the BRST charge in Section \[BRSTchargesection\] and of the Rothstein bracket in Section \[RothPB\]). The ‘correct’ cotangent space ${\mathscr A_{\mathcal V}}=\oplus_{k\in\mathbbm Z} {\mathscr A_{\mathcal V}}^k$, which is considerably bigger, is given by $$\begin{aligned} {\mathscr A_{\mathcal V}}^k=\prod_{j\ge 0} (S_A\mathcal V^)^{k+j}\otimes (S_A\mathcal V)^j.\end{aligned}$$ It is important to note that in this definition we have reversed the natural grading: if the variables ${\xi^{(i)}_{1}},\dots ,{\xi^{(i)}_{\ell_i}}$ are considered as elements of ${\mathscr A_{\mathcal V}}$ they have degree $-i$. Analogously, if the momenta ${\xi_{(i)}^{1}},\dots ,{\xi_{(i)}^{\ell_i}}$ are considered as elements of ${\mathscr A_{\mathcal V}}$ they have dgeree $i$ (which is $-1$ times their natural degree). The multiplication for $S_A(\mathcal V\oplus \mathcal V^)$ extends naturally to a $\mathbbm Z$-graded super-commutative multiplication on ${\mathscr A_{\mathcal V}}$. More specifically, if $\sum_{i\ge 0}\alpha_{i+n}\otimes v_i\in {\mathscr A_{\mathcal V}}^n$ and $\sum_{j\ge 0}\beta_{j+m}\otimes w_j\in {\mathscr A_{\mathcal V}}^m$ then the product is defined to be $$\begin{aligned} \left(\sum_{i\ge 0}\alpha_{i+n}\otimes v_i\right)\left(\sum_{j\ge 0}\beta_{j+m}\otimes w_j\right):=\sum_{k\ge 0}\sum_{i+j=k} (-1)^{i(j+m)} \alpha_{i+n}\beta_{j+m}\otimes v_i w_j.\end{aligned}$$ In calculations we will usually drop the $\otimes$-sign for convenience. Even though we will not make use of it, let us mention the following interpretation of the algebra ${\mathscr A_{\mathcal V}}$. Recall that the symmetric algebra $S_A\mathcal V$ is actually super-commutative (let us write for the multiplication $\mu$), super-cocommutative bialgebra (for the comultiplication we write $\Delta$). Moreover, there is a canonical isomorphism between the $\mathbbm Z$-graded $A$-modules ${\mathscr A_{\mathcal V}}^k$ and ${\operatorname{End}}_{{{A}\mathbf{-Mod}^{\mathbbm Z}}}^{-k}(S_A\mathcal V)$. The bialgebra structure on $S_A\mathcal V$ induces on ${\operatorname{End}}_{{{A}\mathbf{-Mod}^{\mathbbm Z}}}^{-\bullet}(S_A\mathcal V)$ the structure of a super-commutative algebra in ${{A}\mathbf{-Mod}^{\mathbbm Z}}$ with respect to the convolution product $\varphi\diamond \psi:=\mu\circ(\varphi\otimes \psi)\circ \Delta$. We claim that the aforementioned canonical isomorphism is an isomorphism of graded algebras. Let us now address the more subtle issue of derivations of the algebra ${\mathscr A_{\mathcal V}}$. In analogy to the derivations ${\frac{\partial}{\partial \xi^{(i)}_{1}}},\dots,{\frac{\partial}{\partial \xi^{(i)}_{\ell_i}}}$ for $i=1,2,\dots$ (defined in equation (\[algebraicder\])) we now have in addition the $A$-linear derivations $$\begin{aligned} {\frac{\partial}{\partial \xi_{(i)}^{1}}},\dots,{\frac{\partial}{\partial \xi_{(i)}^{\ell_i}}}\qquad \mbox{for }i=1,2,\dots. \end{aligned}$$ The above $A$-linear derivations do pairwise super-commute. Let us stipulate that the connections $\nabla^{V^i}$ and $\nabla ^{{V^i}^}$ are chosen in such a way that $$\begin{aligned} X<\alpha,v>=<\alpha,\nabla^{V^i}_Xv>+<\nabla ^{{V^i}^}_X\alpha,v>\end{aligned}$$ for all $X\in{\mathfrak X}^1(M)$, $\alpha\in{\mathcal V^i}^$ and $v\in\nabla^{V^i}$. In other words, the connection $\nabla^i:=\nabla^{V^i}+\nabla ^{{V^i}^}$ on the super-Riemannian vector bundle $({V^i}^\oplus V^i,g:=<,>)$ is required to be metric. Here, super-Riemannian means that $g:=<,>$ is symmetric for all odd $i$ and antisymmetric for even $i$. The above derivations are derivations of $S_A(\mathcal V\oplus \mathcal V^)$ and ${\mathscr A_{\mathcal V}}$ as well. Unfortunately, not every graded (and infinite) linear combination over ${\mathscr A_{\mathcal V}}$ of the above derivations gives a well-defined operation. In order to define the sensible tangent space to ${\mathscr A_{\mathcal V}}$ we therefore consider again finite type derivations. A derivation $X$ of ${\mathscr A_{\mathcal V}}$ on degree $k$ is said to be of finite type if it writes in local coordinates as follows $$\begin{aligned} {\sum_{j\in \mathbbm Z}}^\prime \sum_{a=1}^{\ell_j} X_a^j\:{\frac{\partial}{\partial \xi^{(j)}_{a}}}+ {\sum_{j\in \mathbbm Z}}^\prime \sum_{a=1}^{\ell_j} X^a_j\:{\frac{\partial}{\partial \xi_{(j)}^{a}}}+ {\sum_{j\in \mathbbm Z}}^\prime\sum_{i=1}^n X_i\:\nabla^j_{\frac{\partial}{\partial x^i}}.\end{aligned}$$ Once again, the prime indicates that all except finitely many summands vanish. For the coefficients we have $X_a^j\in {\mathscr A_{\mathcal V}}^{k-j}$, $X^a_j\in {\mathscr A_{\mathcal V}}^{k+j}$ and $X_i\in {\mathscr A_{\mathcal V}}^{k}$. The graded $A$-module of finite type derivations of ${\mathscr A_{\mathcal V}}$ will be denoted by ${\mathfrak X}^1({\mathscr A_{\mathcal V}})$. Of course, $({\mathscr A_{\mathcal V}},{\mathfrak X}^1({\mathscr A_{\mathcal V}}))$ form a Lie-Rinehart pair in ${{A}\mathbf{-Mod}^{\mathbbm Z}}$. If $(S_A\mathcal V,{\mathfrak X}^1(S_A\mathcal V))$ is a generalized graded manifold then its ghost-cotangent space is defined to be the Lie-Rinehart pair $({\mathscr A_{\mathcal V}},{\mathfrak X}^1({\mathscr A_{\mathcal V}}))$ in the category ${{A}\mathbf{-Mod}^{\mathbbm Z}}$. In the infinitely generated case this definition has some drawbacks. In fact, the Euler vector field $$\begin{aligned} \sum_{i\in Z}\sum_{a=1}^{\ell_i} {\xi^{(i)}_{a}} {\frac{\partial}{\partial \xi^{(i)}_{a}}}+{\xi_{(i)}^{a}} {\frac{\partial}{\partial \xi_{(i)}^{a}}}\end{aligned}$$ is a finite type derivation if and only if $\|\operatorname{supp}(\mathcal V)\|<\infty$. The same is true for the (even) derivations which write locally as follows $$\begin{aligned} \frac{\partial}{\partial \xi^i}:=\sum_{j\in\mathbbm Z}\nabla^j_{\frac{\partial}{\partial x^i}}\qquad i=1,\dots,n=\dim M.\end{aligned}$$ Note that there are several canonical morphism of Lie-Rinehart pairs around. For example, there are the obvious inclusions $$\begin{aligned} (S_A\mathcal V, {\mathfrak X}^1 (S_A\mathcal V))\hookrightarrow(S_A(\mathcal V\oplus\mathcal V^), {\mathfrak X}^1 (S_A(\mathcal V\oplus\mathcal V^)))\hookrightarrow({\mathscr A_{\mathcal V}},{\mathfrak X}^1({\mathscr A_{\mathcal V}})).\end{aligned}$$ Moreover, we have an obvious surjection $$\begin{aligned} ({\mathscr A_{\mathcal V}},{\mathfrak X}^1({\mathscr A_{\mathcal V}}))\stackrel{\epsilon^\prime}{\to} (A=\mathcal C^\infty(M),{\mathfrak X}^1(M)).\end{aligned}$$ More specifically, $\epsilon^\prime_{\|{\mathscr A_{\mathcal V}}}$ is just the canonical augmentation map ${\mathscr A_{\mathcal V}}\to A$. The image under $\epsilon^\prime$ of an $A$-linear derivation is defined to be zero. Finally, if $\mathcal V^j\ne 0$ then we define $$\begin{aligned} \epsilon^\prime(\nabla^j_{\frac{\partial}{\partial x^i}}):=\frac{\partial}{\partial x^i}\end{aligned}$$ (otherwise the latter is defined to be zero). Note that this map is in fact compatible with the brackets, since the curvature terms are killed by the augmentation map. Following the ideas of Section \[derbrsection\], we define the space of polyvector fields ${\mathfrak X}^\bullet({\mathscr A_{\mathcal V}})$ to be that which associated to the Lie-Rinehart pair $({\mathscr A_{\mathcal V}},{\mathfrak X}^1({\mathscr A_{\mathcal V}}))$. It will become clear, that (at least for the infinitely generated case) the more correct definition would be the space of multiderivations ${\operatorname{Der}}^{\bullet}({\mathscr A_{\mathcal V}})$ of the $\mathbbm Z$-graded algebra ${\mathscr A_{\mathcal V}}$. Unfortunately, computations in this object tend to be clumsy. Note that, according to Theorem \[SchoutenNijenhuis\], the morphism $\epsilon'$ extends to a morphism of Gerstenhaber algebras $$\begin{aligned} \epsilon': {\mathfrak X}^\bullet({\mathscr A_{\mathcal V}})\to {\mathfrak X}^\bullet(M).\end{aligned}$$ In the finitely generated case we prefer to work with a morphism $\epsilon$ which is obtained from $\epsilon^\prime$ by a renormalization. This $\epsilon$ is defined in the same way as $\epsilon^\prime$ except that $$\begin{aligned} \epsilon(\nabla^j_{\frac{\partial}{\partial x^i}})=\|\operatorname{supp}(\mathcal V)\|^{-{\frac{1}{2}}}\frac{\partial}{\partial x^i}.\end{aligned}$$ The Rothstein-Poisson bracket {#RothPB} ------------------------------ In [@Rothstein] M. Rothstein has shown, that on every super-manifold with symplectic base there is a super-symplectic 2-form. This fact has been rediscovered by M. Bordemann [@SuperB] as a byproduct of the Fedosov construction for super-manifolds. Here we shall extend this result to ghost-cotangent spaces of generalized graded manifolds with a Poisson base manifold. We will treat first the case when the graded manifold is finitely generated. The reason for that is that in the finitely generated case the computations can be done in the Gerstenhaber algebra of polyvector fields defined in the preceding section. In contrast, in the infinitely generated case they make sense merely in the Gerstenhaber algebra of multiderivations. In the infinitely generated case our argument is still incomplete. ### The finitely generated case {#RothPBfin} \[Roth\] Let $M$ be a Poisson manifold with Poisson tensor $\Pi\in{\mathfrak X}^2(M)$ and let $V=V^{\le L}=\oplus_{k=1}^L V^k$ a positively graded finite dimensional vector bundle over $M$. Let us write $A:=\mathcal C^\infty(M)$ for the algebra of smooth functions on $M$ and $\mathcal V=\oplus_{i\ge 1}^L\mathcal V^i$ for the space of smooth sections of $V$. Then there exists a Poisson tensor $\Pi_R\in {\mathfrak X}^2({\mathscr A_{\mathcal V^{\le{L}}}})$ on the ghost cotangent space of the finitely generated graded manifold $S_A\mathcal V^{\le L}$, such that 1. \[Poiss\] $[\Pi_R,\Pi_R]=0$, 2. \[degzero\]$\Pi_R$ is of total degree $2$, 3. \[proj\]the image under $\epsilon$ of $\Pi_R$ in ${\mathfrak X}^2(M)$ is $\Pi$. The proof will be done by checking, that the explicit solution $\Pi_R$, called the Rothstein Poisson tensor, which can be read off from [@Rothstein] does in fact fulfill the requirements in this more general context. The first thing that we will do is to introduce a more condensed notation using the super-Riemannian metric $g$ which we already mentioned in the preceeding section. We denote by $\xi_1^{(j)},\xi_2^{(j)}\dots \xi_{\ell_j}^{(j)}$, $j=1,\dots ,L$, a local frame for $V^j$ and by $\xi^1_{(j)},\xi^2_{(j)}\dots \xi^{\ell_j}_{(j)}$ the corresponding dual frame for $V^{j}$. Recall that $\ell_j$ is the dimension of $V^j$ and $\ell:=\sum_j \ell_j$ is the dimension of $V=\oplus_j V^j$. On $V\oplus V^$ metric the $g$ is given by: $$\begin{aligned} g\big({\xi_{(i)}^{a}},{\xi^{(j)}_{b}}\big):=\delta_j^i\:\delta^a_b=:(-1)^{i+1}g\big({\xi^{(j)}_{b}},{\xi_{(i)}^{a}}\big)\\ g\big({\xi^{(i)}_{a}},{\xi^{(j)}_{b}}\big):=0=:g\big({\xi_{(i)}^{a}},{\xi_{(j)}^{b}}\big),\end{aligned}$$ for indices $a=1,\dots ,\ell_i$, $b=1,\dots,\ell_j$ and $i,j=1,\dots,r$. Recall that we have chosen a connection $\nabla=\sum_j \nabla^j$ on $V\oplus V^$ which is metric with respect to $g$. Imposing the graded Leibniz rule, we let $$\begin{aligned} {\frac{\partial }{\partial \xi^{i}}}:=\nabla_{\frac{\partial}{\partial x^i}}\end{aligned}$$ act on the super-commutative algebra ${\mathscr A_{\mathcal V^{\le{L}}}}$ as even derivations. In order to simplify the computations, we reindex the above frames into a frame $\xi^1,\xi^2,\dots,\xi^{2\ell}$ for the vector bundle $V\oplus V^$ as follows: $$\begin{aligned} \xi^A:=\xi_a^{(i)},\qquad\xi^{\ell+A}:=\xi^a_{(i)}.\end{aligned}$$ for $A=a+\sum_k^{i-1}\ell_k$ (here we set $\ell_0:=0$). The parity $p(A)$ of the index $A$ is defined to be the parity of $\xi^A$, which is $i(\operatorname{mod}2)$. These indices will run through the capital letters $A,B,C$, etc. Quite importantly, we have the following commutation relations $$\begin{aligned} \Big[\:{\frac{\partial }{\partial \xi^{i}}},{\frac{\partial }{\partial \xi^{j}}}\:\Big]=\sum_{A,B}R^B_{Aij}\,\xi^A\,{\frac{\partial }{\partial \xi^{B}}}\label{commrel1}\\ \:\Big[\:{\frac{\partial }{\partial \xi^{i}}},{\frac{\partial }{\partial \xi^{A}}}\:\Big]=\sum_{B}\Gamma_{iA}^B\,{\frac{\partial }{\partial \xi^{B}}}\label{commrel2}\end{aligned}$$ for the ‘super-coordinate’ vector fields, where $\Gamma_{iA}^B$ and $R^B_{Aij}$ are the Christoffel symbols and the components of the curvature tensor of the connection $\nabla$. Using the inverse of the metric tensor $\sum_A g_{AB}\:g^{AC}=\delta_B^C$ we define the following algebraic bivector field $$\begin{aligned} \Pi_0:={\frac{1}{2}}\sum_{A,B}g^{AB}{\frac{\partial }{\partial \xi^{A}}}{\frac{\partial }{\partial \xi^{B}}}\in {\mathfrak X}^2({\mathscr A_{\mathcal V^{\le{L}}}}).\end{aligned}$$ We need to define the insertation derivations $i({\xi_A})$ and $i(\frac{\partial}{\partial \xi_A})$, for $A=1,\dots,2\ell$, of the supercommutative algebra structure on ${\mathfrak X}({\mathscr A_{\mathcal V^{\le{L}}}})$, which extend the dual pairings. More precisely, we set $i({\xi_A})\xi^B:=\delta^B_A$ and $i(\frac{\partial}{\partial \xi_A}){\frac{\partial }{\partial \xi^{B}}}:=\delta_B^A$. Of course, the remaining generators will be killed by these derivations. Moreover, let us introduce the derivations $d:=[\Pi_0,\:]$ and $d^:=(-1)^{p(A)+1}\sum_{A,B}g_{AB}\:\xi^A\: i(\frac{\partial}{\partial \xi_B})$ and the degree derivations ${\operatorname{deg}}_\xi:=\sum_A \xi^A\:i({\xi_A})$ and ${\operatorname{deg}}_{\partial\xi}:=\sum_A {\frac{\partial }{\partial \xi^{A}}}\:i({\frac{\partial}{\partial \xi_A}})$. Finally, we introduce the operator $d^{-1}$ as follows. If $X\in {\mathfrak X}({\mathscr A_{\mathcal V^{\le{L}}}})$ is a polyvector field such that ${\operatorname{deg}}_\xi X=m X$ and ${\operatorname{deg}}_{\partial\xi}X=n X$, then we define $d^{-1}X:=(m+n)^{-1}d^X$. On $\operatorname{ker}({\operatorname{deg}}_\xi)\cap\operatorname{ker}({\operatorname{deg}}_{\partial\xi})$ we define $d^{-1}$ to be zero. \[pinull\] Since the connection $\nabla$ is metric, we have $$\begin{aligned} d=[\Pi_0,\:]=\sum_{A,B}g^{AB}{\frac{\partial }{\partial \xi^{A}}}\:i(\xi_B)\label{adpi}\\ d^2=0.\label{Komplex}\end{aligned}$$ Moreover, we obtain a “Hodge identity”: $dd^+d^ d={\operatorname{deg}}_\xi+{\operatorname{deg}}_{\partial\xi}$, which implies that $d^{-1}$ is a contracting homotopy for $d$. More specifically, we have that $$\begin{aligned} \label{dcontr} \xymatrix{({\mathfrak X}^\bullet(M),0)\ar@<-0.6ex>[r]_{\iota\quad}^{}&({\mathfrak X}^\bullet({\mathscr A_{\mathcal V^{\le{L}}}}),d),d^{-1} \ar@<-0.6ex>[l]_{\epsilon\quad}^{}}\end{aligned}$$ is a contraction fulfilling all side conditions. Since $d$ is a derivation of the super-commutative product, it is sufficient to evaluate $d$ on the generators. Since the connection is metric we have $[\Pi_0,{\frac{\partial }{\partial \xi^{i}}}]=0$ for all $i=1,\dots,n$. Moreover, $\Pi_0$ commutes with all ${\frac{\partial }{\partial \xi^{C}}}$ and functions $f\in\mathcal C^\infty(M)$. Hence the only nonvanishing contribution comes from: $$\begin{aligned} \Big[\Pi_0,\xi^{C}\Big]={\frac{1}{2}}\Big[\sum_{A,B}g^{AB}{\frac{\partial }{\partial \xi^{A}}}{\frac{\partial }{\partial \xi^{B}}},\xi^{C}\Big]=\sum_{AB}g^{AB}{\frac{\partial }{\partial \xi^{A}}}\:\delta^C_B.\end{aligned}$$ This proves equation (\[adpi\]). Equation (\[Komplex\]) follows immediately. The Hodge identity follows from the supercommutator: $[{\frac{\partial }{\partial \xi^{A}}}\:i_{\xi_B},\xi^C\: i_{\frac{\partial}{\partial \xi_D}}]=\delta_B^C\:{\frac{\partial }{\partial \xi^{A}}}i_{\frac{\partial}{\partial \xi_D}}+\delta_A^D\:\xi^C\:i_{\xi_B}$ for $p(A)=p(B)$ and $p(C)=p(D)$. This is straightforward to check (however, one has to be careful with the signs). All the remaining statements are obvious. \ Next, we introduce the following matrix of super-functions, which has incorporated the Poisson structure and curvature as well: $$\begin{aligned} \hat{R}_j^i:=-{\frac{1}{2}}\sum_k\sum_{A,B,C}\:\Pi^{ik}\:R^A_{Bkj}\:g_{AC}\:\xi^B\xi^C\quad\in {\mathscr A_{\mathcal V^{\le{L}}}}^0.\end{aligned}$$ One of the reasons to work with the completion, i.e. ${\mathscr A_{\mathcal V}}$ instead of $S_A(\mathcal V^\oplus{\mathcal V}^)$, is that otherwise the matrix $({\operatorname{id}}-\hat R)^{-1}$ is well defined only under special circumstances. More specifically, the matrices $$\begin{aligned} \mathcal A^i_j:=\big(({\operatorname{id}}-\hat{R})^{-{\frac{1}{2}}}\big)^i_j =\big({\operatorname{id}}+{\frac{1}{2}}\:\hat{R}+\frac{1\cdot 3}{2\cdot 4}\:\hat{R}^2+\frac{1\cdot 3\cdot 5}{2\cdot 4\cdot 8}\:\hat{R}^3+\dots\big)_j^i \quad\in {\mathscr A_{\mathcal V^{\le{L}}}}^0\\ \mathcal B^i_j:=\sum_k \mathcal A^i_k \mathcal A^k_j=\big(({\operatorname{id}}-\hat{R})^{-1}\big)^i_j =\big({\operatorname{id}}+\hat{R}+\hat{R}^2+\hat{R}^3+\dots\big)_j^i\quad\in {\mathscr A_{\mathcal V^{\le{L}}}}^0\end{aligned}$$ of superfunctions will play a vital role in what follows. Let us establish some calculation rules. We have the following local formulas: $$\begin{aligned} \sum _j\Pi^{ij}\,\hat{R}^k_j=-\sum_j\Pi^{kj}\,\hat{R}_k^i \label{sym1},\\ \sum_{i,j}\Pi^{ij}\,\mathcal A_i^k\mathcal A_j^l=\sum_j \Pi^{kj}\mathcal B_j^l=-\sum_j \Pi^{lj}\mathcal B_j^k \label{sym2},\\ {\frac{\partial }{\partial \xi^{i}}}\Big(\sum_{A,B,C} \:R^A_{Bkj}\:g_{AC}\:\xi^B\xi^C\Big)=0\label{Bianchi},\\ d\hat R_j^i= - \sum_{k,A,B}\Pi^{ik}\: R_{Bkj}^A\:\xi^B{\frac{\partial }{\partial \xi^{A}}}\label{ortho1},\\ d\mathcal B_j^i=- \sum_{k,m,n,A,B}\Pi^{mk}\: \mathcal B^i_m\:\mathcal B^n_j\:R_{Bkn}^A\:\xi^B{\frac{\partial }{\partial \xi^{A}}}\label{ortho2},\\ {\frac{\partial }{\partial \xi^{k}}}\Big(\sum_l \Pi^{il} \mathcal B_l^j\Big)=\sum_{m,n} (\partial_k\Pi ^{mn})\,\mathcal B^i_m\,\mathcal B^j_n.\label{fortune}\end{aligned}$$ Throughout the proof we use Einstein summation convention. Equation (\[sym1\]) is an immediate consequence of the symmetry properties of the curvature tensor and (\[sym2\]) follows easily. Equation (\[Bianchi\]) is equivalent to Bianchi’s identity. In order to check identity (\[ortho1\]) we note that since $\nabla$ is metric, the curvature is orthogonal with respect to $g$ (regardless of the symmetry properties of $g$): $$\begin{aligned} g^{AC}\:R^B_{Cij}=g(\xi^A,R(\partial_i,\partial_j)\xi^B)=-g(R(\partial_i,\partial_j)\xi^A,\xi^B)=-g^{CB} \:R^A_{Cij}.\end{aligned}$$ By inverting $g$ we obtain $R^C_{Dij}=-g^{AC}\:g_{BD} R_{Aij}^B$. After these preparations we get $$\begin{aligned} d\hat R_j^i&=& -{\frac{1}{2}}\:\Pi^{il}\:R_{Blj}^A\: g_{AC}d(\xi^B\xi^C)\\ &=&-{\frac{1}{2}}\:\Pi^{il}\:R_{Blj}^A\Big(g_{AC}\:g^{DE}\:{\frac{\partial }{\partial \xi^{D}}}\:\delta^B_E\:\xi^C+(-1)^{p(B)}g_{AC}\:g^{DE}\:{\frac{\partial }{\partial \xi^{D}}}\xi^B\:\delta^C_E\Big)\\ &=&-{\frac{1}{2}}\:\Pi^{il}\:R_{Blj}^A\Big((-1)^{p(B)}g_{AC}\:g^{DB}\:\xi^C+(-1)^{2p(B)}g_{AC}\:g^{DC}\xi^B \:\Big){\frac{\partial }{\partial \xi^{D}}}\\ &=& - \Pi^{ik}\: R_{Bkj}^A\:\xi^B{\frac{\partial }{\partial \xi^{A}}}.\end{aligned}$$ Note the following easy to prove matrix identity $$\begin{aligned} d\big(({\operatorname{id}}-\hat{R})^{-1}\big)=({\operatorname{id}}-\hat{R})^{-1}(d\hat{R})({\operatorname{id}}-\hat{R})^{-1}\label{geomid}.\end{aligned}$$ This and the analogous statement for ${\frac{\partial }{\partial \xi^{i}}}$ are the key to the remaining computations. Equation (\[ortho2\]) follows from the equations (\[ortho1\]) and (\[geomid\]). Finally we show equation (\[fortune\]): $$\begin{aligned} {\frac{\partial }{\partial \xi^{k}}}\Big(\Pi^{il}\mathcal B_l^j\Big)&\stackrel{(\ref{geomid})}{=}&\partial_k \Pi^{il}\:\mathcal B^j_l+\Pi^{il}\mathcal B^m_l\Big({\frac{\partial }{\partial \xi^{k}}}\hat R_m^n\Big)\mathcal B_n^j\\ &\stackrel{(\ref{Bianchi})}{=}&\partial_k \Pi^{il}\:\mathcal B^j_l-{\frac{1}{2}}\:\Pi^{il}\mathcal B^m_l\Big(\partial_k \Pi^{nr}\: R_{Brm}^A\:g_{AC}\:\xi^B\xi^C\Big)\mathcal B_n^j\\ &\stackrel{(\ref{sym2})}{=}&\partial_k \Pi^{il}\:\mathcal B^j_l-{\frac{1}{2}}\:\Pi^{lm}\mathcal B^i_l\Big(\partial_k \Pi^{nr}\: R_{Brm}^A\:g_{AC}\:\xi^B\xi^C\Big)\mathcal B_n^j\\ &=&\partial_k \Pi^{il}\:\mathcal B^j_l-\hat R ^l_r\,\mathcal B_l^i\,\mathcal B_n^j\:\partial_k \Pi^{nr}\\ &\stackrel{()}{=}&\partial_k \Pi^{il}\:\mathcal B^j_l-(\mathcal B^i_r-\delta_r^i)\mathcal B_n^j\:\partial_k \Pi^{nr}=-\mathcal B^i_r\,\mathcal B_n^j\:\partial_k \Pi^{nr}.\end{aligned}$$ At step (\) we again have used the geometric series $\hat R({\operatorname{id}}-\hat R)^{-1}=({\operatorname{id}}-\hat R)^{-1}-{\operatorname{id}}$. \ Now we are ready to prove that $$\begin{aligned} \widetilde{\Pi}:={\frac{1}{2}}\sum_{i,j,k,l}\Pi^{ij}\,\mathcal A^k_i\mathcal A^l_j\:{\frac{\partial }{\partial \xi^{k}}}{\frac{\partial }{\partial \xi^{l}}}={\frac{1}{2}}\sum_{j,k,l}\:\Pi^{kj}\: \mathcal B_j^l\:{\frac{\partial }{\partial \xi^{k}}}{\frac{\partial }{\partial \xi^{l}}}\end{aligned}$$ is a solution of the Maurer-Cartan equation \[MCeq\] $d\widetilde{\Pi}+{\frac{1}{2}}\,[\widetilde{\Pi},\widetilde{\Pi}]=0$. Again, Einstein summation convention is in force. Using equation (\[ortho2\]) we obtain $$\begin{aligned} d\widetilde{\Pi}&=&{\frac{1}{2}}\:\Pi^{kj}\:d\mathcal B_j^l\:{\frac{\partial }{\partial \xi^{k}}}{\frac{\partial }{\partial \xi^{l}}}=-{\frac{1}{2}}\:\Pi^{kj}\:\Pi^{mn}\:\mathcal B_m^l\:\mathcal B_j^i\:R^A_{Bni}\:\xi^B{\frac{\partial }{\partial \xi^{A}}}{\frac{\partial }{\partial \xi^{k}}}{\frac{\partial }{\partial \xi^{l}}}\\ &=&{\frac{1}{2}}\:\Pi^{kj}\:\Pi^{ln}\:\mathcal B_m^l\mathcal B_j^i\:R^A_{Bni}\:\xi^B{\frac{\partial }{\partial \xi^{A}}}{\frac{\partial }{\partial \xi^{k}}}{\frac{\partial }{\partial \xi^{l}}}.\label{sieheunten}\end{aligned}$$ It remains to compute $[\widetilde{\Pi},\widetilde{\Pi}]$. Note this is a linear combination of two terms. The first is proportional to $\Pi^{k_1i_1}\mathcal B_{i_1}^{l_1}\big({\frac{\partial }{\partial \xi^{k_1}}}(\Pi^{k_2i_2}\mathcal B_{i_2}^{l_2})\big){\frac{\partial }{\partial \xi^{l_1}}}{\frac{\partial }{\partial \xi^{k_2}}}{\frac{\partial }{\partial \xi^{l_2}}}$ and we show using equation (\[fortune\]) that this contribution vanishes: $$\begin{aligned} \Pi^{k_1i_1}\mathcal B_{i_1}^{l_1}\Big({\frac{\partial }{\partial \xi^{k_1}}}(\Pi^{k_2i_2}\mathcal B_{i_2}^{l_2})\Big){\frac{\partial }{\partial \xi^{l_1}}}{\frac{\partial }{\partial \xi^{k_2}}}{\frac{\partial }{\partial \xi^{l_2}}}=\Pi^{k_1i_1}\:\partial_{k_1}\Pi^{mn}\:\mathcal B_{i_1}^{l_1}\mathcal B_{m}^{k_2}\mathcal B_{n}^{l_2}{\frac{\partial }{\partial \xi^{l_1}}}{\frac{\partial }{\partial \xi^{k_2}}}{\frac{\partial }{\partial \xi^{l_2}}}=0,\end{aligned}$$ since $\Pi^{k_1i_1}\partial_{k_1}\Pi^{mn}+cycl(i_1,m,n)=0$. Therefore, using (\[commrel1\]) we conclude that $$\begin{aligned} [\widetilde{\Pi},\widetilde{\Pi}]&=&\frac{1}{4}\:\Pi^{k_1j_1}\mathcal B_{j_1}^{l_1}\:\Pi^{k_2j_2}\mathcal B_{j_2}^{l_2}\:\Big[{\frac{\partial }{\partial \xi^{k_1}}}{\frac{\partial }{\partial \xi^{l_1}}},{\frac{\partial }{\partial \xi^{k_2}}}{\frac{\partial }{\partial \xi^{l_2}}}\Big]\\ &=&-\Pi^{k_1j_1}\Pi^{k_2j_2}\:\mathcal B_{j_1}^{l_1}\mathcal B_{j_2}^{l_2}\:R_{Bl_1l_2}^A\:{\frac{\partial }{\partial \xi^{k_1}}}\:\xi^B{\frac{\partial }{\partial \xi^{A}}}\:{\frac{\partial }{\partial \xi^{k_2}}}\\ &=&\Pi^{k_1j_1}\Pi^{k_2j_2}\:\mathcal B_{j_1}^{l_1}\mathcal B_{j_2}^{l_2}\:R_{Bl_1l_2}^A\:\xi^B{\frac{\partial }{\partial \xi^{A}}}\:{\frac{\partial }{\partial \xi^{k_1}}}{\frac{\partial }{\partial \xi^{k_2}}},\end{aligned}$$ which after comparison with equation (\[sieheunten\]) yields the Maurer-Cartan equation. \ We conclude that the Rothstein Poisson tensor* $$\begin{aligned} \Pi_R:=\Pi_0+\widetilde{\Pi} \end{aligned}$$ has the desired properties \[Poiss\], \[degzero\] and \[proj\] of Theorem \[Roth\]. Therefore, ${\mathfrak X}^\bullet({\mathscr A_{\mathcal V^{\le{L}}}})$ is a cochain complex with differential $\delta_{\Pi_R}:=[\Pi_R,\:]$, which is an analog of Lichnerowicz-Poisson cohomology. Since ${\mathscr A_{\mathcal V^{\le{L}}}}$ is an abelian subalgebra of the Gerstenhaber algebra ${\mathfrak X}({\mathscr A_{\mathcal V^{\le{L}}}})$, by the derived bracket construction \[derbr1\].) of Theorem \[derKS\], ${\mathscr A_{\mathcal V^{\le{L}}}}$ acquires the structure of a $\mathbbm Z$-graded Poisson algebra. More precisely, we define the Rothstein bracket $\{\:,\:\}_R$ to be the opposite derived bracket of $\delta_{\Pi_R}$, i.e., for two homogeneous elements $\alpha,\beta\in{\mathscr A_{\mathcal V^{\le{L}}}}$ we have : $$\begin{aligned} \{\alpha,\beta\}_R:=[\alpha,\beta]^{{\operatorname{opp}}}_{\delta_{\Pi_R}}=-[\alpha,\delta_{\Pi_R}\beta]=-[\alpha,[\Pi_R,\beta]]\end{aligned}$$ In local coordinates the Rothstein bracket writes as follows $$\begin{aligned} \{\alpha,\beta\}_R&=&-[\alpha,[\widetilde{\Pi},\beta]]-[\alpha,[\Pi_0,\beta]] \label{rothbr}\\ &=&\sum_{j,k,l}\:\Pi^{kj}\: \mathcal B_j^l\:\frac{\partial\alpha}{\partial \xi^k} \frac{\partial\beta}{\partial \xi^l} +\sum_{i=1}^L(-1)^{\|\alpha\|i+1}\sum_{a=1}^{\ell_i}\Big(\frac{\partial\alpha}{\partial {\xi^{(i)}_{a}}}\frac{\partial\beta}{\partial {\xi_{(i)}^{a}}}+ (-1)^i\frac{\partial\alpha}{\partial {\xi_{(i)}^{a}}}\frac{\partial\beta}{\partial {\xi^{(i)}_{a}}}\Big).\nonumber\end{aligned}$$ We would like to emphasize that $\mathcal C^\infty(M)$ is a Poisson subalgebra of ${\mathscr A_{\mathcal V^{\le{L}}}}$ if and only if the curvature terms vanish, i.e., the bundle $V=\oplus_i V^i$ is flat. Nonetheless, the Lichnerowicz-Poisson cohomologies of the Poisson algebras $({\mathscr A_{\mathcal V^{\le{L}}}},\{\,,\,\}_R)$ and $(\mathcal C^\infty(M),\{\,,\,\})$ are always quasiisomorphic as Theorem \[Lichcontr\] below shows. If the bundle $V=\oplus_i V^i$ is the trivial bundle with the canonical flat connection, then the $\frac{\partial}{\partial \xi^k}$ are just the ordinary partial derivatives, the curvature term $\mathcal B_j^l$ disappears and the Rothstein bracket boils down to the well known $$\begin{aligned} \label{ordbr} \{\alpha,\beta\}_R=\sum_{k,l}\:\Pi^{kl}\:\frac{\partial\alpha}{\partial \xi^k} \frac{\partial\beta}{\partial \xi^l} +\sum_{i=1}^L\sum_{a=1}^{\ell_i} \alpha\frac{\overleftarrow{\partial}}{\partial {\xi_{(i)}^{a}}}\frac{\overrightarrow{\partial}}{\partial {\xi^{(i)}_{a}}}\beta-(-1)^{\|\alpha\|\|\beta\|}\beta\frac{\overleftarrow{\partial}}{\partial {\xi_{(i)}^{a}}}\frac{\overrightarrow{\partial}}{\partial {\xi^{(i)}_{a}}}\alpha.\end{aligned}$$ Here we have used the physicist’s denotation[^5]: $$\begin{aligned} \alpha\frac{\overleftarrow{\partial}}{\partial {\xi_{(i)}^{a}}}:=-\Big[\alpha,\frac{\partial}{\partial {\xi_{(i)}^{a}}}\Big]=(-1)^{(\|\alpha\|+1)i}\frac{\partial\alpha}{\partial {\xi_{(i)}^{a}}},\quad \frac{\overrightarrow{\partial}}{\partial {\xi^{(i)}_{a}}}\alpha=\frac{\partial\alpha}{\partial {\xi^{(i)}_{a}}}.\end{aligned}$$ In order to proof equation (\[rothbr\]) and (\[ordbr\]) we essentially have to compute $$\begin{aligned} [\alpha,[\Pi_0,\beta]]&=&\sum_{i=1}^L\sum_{a=1}^{\ell_i} \Big[\alpha,\Big[{\frac{\partial}{\partial \xi_{(i)}^{a}}}{\frac{\partial}{\partial \xi^{(i)}_{a}}},\beta\Big]\Big]\\ &=&\sum_{i=1}^L\sum_{a=1}^{\ell_i}\Big[\alpha,{\frac{\partial}{\partial \xi_{(i)}^{a}}}\Big[{\frac{\partial}{\partial \xi^{(i)}_{a}}},\beta\Big]\Big]+(-1)^{(\|\beta\|+1)(i+1)}\Big[\alpha,\Big[{\frac{\partial}{\partial \xi_{(i)}^{a}}},\beta\Big]{\frac{\partial}{\partial \xi^{(i)}_{a}}}\Big]\\ &=&\sum_{i=1}^L\sum_{a=1}^{\ell_i}\Big[\alpha,{\frac{\partial}{\partial \xi_{(i)}^{a}}}\Big]\Big[{\frac{\partial}{\partial \xi^{(i)}_{a}}},\beta\Big]-(-1)^{(\|\beta\|+1)(i+1)}(-1)^{(\|\alpha\|+1)\|\beta\|}\Big[{\frac{\partial}{\partial \xi_{(i)}^{a}}},\beta\Big]\Big[{\frac{\partial}{\partial \xi^{(i)}_{a}}},\alpha\Big]\\ &=&\sum_{i=1}^L\sum_{a=1}^{\ell_i}\Big[\alpha,{\frac{\partial}{\partial \xi_{(i)}^{a}}}\Big]\Big[{\frac{\partial}{\partial \xi^{(i)}_{a}}},\beta\Big]-(-1)^{\|\alpha\|\|\beta\|+i(\|\beta\|+1)}(-1)^{(\|\beta\|+1)i+1}\Big[\beta,{\frac{\partial}{\partial \xi_{(i)}^{a}}}\Big] \Big[{\frac{\partial}{\partial \xi^{(i)}_{a}}},\alpha\Big]\\ &=&\sum_{i=1}^L\sum_{a=1}^{\ell_i}\Big[\alpha,{\frac{\partial}{\partial \xi_{(i)}^{a}}}\Big]\Big[{\frac{\partial}{\partial \xi^{(i)}_{a}}},\beta\Big]-(-1)^{\|\alpha\|\|\beta\|}\Big[\beta,{\frac{\partial}{\partial \xi_{(i)}^{a}}}\Big] \Big[{\frac{\partial}{\partial \xi^{(i)}_{a}}},\alpha\Big].\end{aligned}$$ \[Lichcontr\] There are linear maps $\mathcal J: {\mathfrak X}^\bullet(M)\to {\mathfrak X}^\bullet({\mathscr A_{\mathcal V^{\le{L}}}})$ and $\Sigma: {\mathfrak X}^\bullet({\mathscr A_{\mathcal V^{\le{L}}}})\to {\mathfrak X}^{\bullet+1}({\mathscr A_{\mathcal V^{\le{L}}}})$ such that $$\begin{aligned} \xymatrix{({\mathfrak X}^\bullet(M),\delta_{\Pi})\ar@<-0.6ex>[r]_{\mathcal J\quad}^{}&({\mathfrak X}^\bullet({\mathscr A_{\mathcal V^{\le{L}}}}),\delta_{{\Pi}_R}),\Sigma \ar@<-0.6ex>[l]_{\epsilon\quad}^{}}\end{aligned}$$ is a contraction. In particular, the Poisson cohomologies of $\Pi$ and $\Pi_R$ are isomorphic. Since $\epsilon$ is a morphism of Gerstenhaber algebras and $\epsilon(\Pi_R)=\Pi$, we have $\epsilon\:\delta_{\Pi_R}=\delta_{\Pi}\:\epsilon$. Hence we may apply perturbation lemma \[BPL1\]. ### The infinitely generated case {#RothPBinfin} Let us now address the infinitely generated case $\mathcal V= \oplus_{i\ge 1} \mathcal V^i$, where $\|\operatorname{supp}(\mathcal V)\|=\infty$. In this case the argument of the previous section seem to break down, first of all, due to the fact that $\Pi_0$ is no longer a true bivector field $$\begin{aligned} \Pi_0=\sum_{i=1}^\infty\sum_{a=1}^{\ell_i} {\frac{\partial}{\partial \xi_{(i)}^{a}}}{\frac{\partial}{\partial \xi^{(i)}_{a}}}\quad \notin {\mathfrak X}^2({\mathscr A_{\mathcal V}}).\end{aligned}$$ In fact, there seems to be no way to write $\Pi_0$ as a second symmetric power. Secondly, $\widetilde{\Pi}$ is no bivector field in the sense of our definition $$\begin{aligned} \widetilde{\Pi}={\frac{1}{2}}\sum_{j,k,l}\:\Pi^{kj}\: \mathcal B_j^l\:{\frac{\partial }{\partial \xi^{k}}}{\frac{\partial }{\partial \xi^{l}}}\notin {\mathfrak X}^2({\mathscr A_{\mathcal V}}).\end{aligned}$$ This is because the innocent looking derivations $\partial/\partial \xi^i$, $i=1,\dots,n=\dim(M)$ are not derivations of finite type. We conclude that, in the infinitely generated case, the Gerstenhaber algebra ${\mathfrak X}({\mathscr A_{\mathcal V}})$ of polyvector fields is too small for our purposes and we need to look for an appropriate replacement. The only proposal that we are aware of and that seems to make sense is to take the Gerstenhaber algebra $\left({\operatorname{Der}}({\mathscr A_{\mathcal V}}),\cup,[\:,\:]_{RN}\right)$ of multiderivations of ${\mathscr A_{\mathcal V}}$ (cf. Theorem \[multiderisgerst\]). In fact, one can show that the formal analog of the two-fold bracket $$\begin{aligned} ``\:B_{\Pi_0}(\alpha,\beta)=[[\Pi_0,\alpha],\beta]\:``=\sum_{i=0}^\infty (-1)^{(\|\beta\|+1)i+\alpha}\sum_{a=1}^{\ell_i}\Big(\frac{\partial\alpha}{\partial {\xi_{(i)}^{a}}}\frac{\partial\beta}{\partial {\xi^{(i)}_{a}}}+ (-1)^i\frac{\partial\alpha}{\partial {\xi^{(i)}_{a}}}\frac{\partial\beta}{\partial {\xi_{(i)}^{a}}}\Big)\end{aligned}$$ is a well-defined operation (this is somewhat surprising) and we conclude that $\Pi_0\in {\operatorname{Der}}^2({\mathscr A_{\mathcal V}})^0$. It is also not difficult to prove that $\Pi_0\bullet \Pi_0={\frac{1}{2}}[\Pi_0,\Pi_0]_{RN}=0$. The term $\widetilde{\Pi}$ is evidently a member of ${\operatorname{Der}}^2({\mathscr A_{\mathcal V}})^0$. We strongly believe that all calculations which lead us to the Maurer-Cartan equation of Lemma \[MCeq\], and which now involve certain infinite summations, make sense in the algebra of multiderivations. (even though we do not know whether there is an analog of the contraction (\[dcontr\])). As a consequence, we conjecture that also in the infinitely generated case the Rothstein bracket $$\begin{aligned} \{\alpha,\beta\}_R=\sum_{j,k,l}\:\Pi^{kj}\: \mathcal B_j^l\:\frac{\partial\alpha}{\partial \xi^k} \frac{\partial\beta}{\partial \xi^l} +\sum_{i=1}^\infty\sum_{a=1}^{\ell_i}\alpha\frac{\overleftarrow{\partial}}{\partial {\xi_{(i)}^{a}}}\frac{\overrightarrow{\partial}}{\partial {\xi^{(i)}_{a}}}\beta-(-1)^{\|\alpha\|\|\beta\|}\beta\frac{\overleftarrow{\partial}}{\partial {\xi_{(i)}^{a}}}\frac{\overrightarrow{\partial}}{\partial {\xi^{(i)}_{a}}}\alpha.\end{aligned}$$ is a $\mathbbm Z$-graded super-Poisson structure, which can be seen as the opposite derived bracket of the abelian subalgebra ${\mathscr A_{\mathcal V}}\subset \oplus_{n\ge 0} {\operatorname{Der}}^n({\mathscr A_{\mathcal V}})$ with respect to the differential $\delta_{\Pi_R}:=[\Pi_0+\widetilde{\Pi},\:]_{RN}$. Clearly, the virtues of the derived bracket construction (i.e. keeping the signs simple) disappear if we do calculations in the algebra of multiderivation. A better strategy to prove the above statement seems to be to use the concept of approximation (cf. Section \[BRSTchargesection\]) and some ‘clever’ continuity argument. In the freely generated case with a flat connection the Jacobiidentity for the bracket $\{\:,\:\}_R$ seems to be folklore. The Koszul complex {#kc} ------------------ Given a smooth map $J:M\to \mathbbm R^\ell=:V^$ we consider the Koszul holomogical complex of the sequence of ring elements $J_1,\dots, J_\ell\in \mathcal C^{\infty}(M)$. In other words, we define the space of chains to be $$K_{i}:=K_{i}(\mathcal C^\infty(M),J):=S^i_{\mathcal C^{\infty}(M)}(V[-1]),$$ i.e., the free (super)symmetric $\mathcal C^{\infty}(M)$-algebra generated by the graded vector space $V[-1]$, where we consider $V$ to be concentrated in degree zero. $K_{\bullet}$ may also be viewed as the space of sections of the trivial vector bundle over $M$ with fibre $\wedge^\bullet V$. Denoting by $\xi^1,\dots, \xi^\ell$ the canonical bases of $V[-1]$ for the dual space $V$ of $V^=\mathbbm R^\ell$, we define the Koszul differential $$\begin{aligned} \partial:=\sum_a J_a {\frac{\partial }{\partial \xi^{a}}},\end{aligned}$$ where the ${\frac{\partial }{\partial \xi^{a}}}$ , $a=1,\dots,\ell$, are the derivations extending the dual pairing. We will say, in accordance with [@Bourbdix], that the sequence of ring elements $J_1,\dots, J_\ell\in \mathcal C^{\infty}(M)$ is a complete intersection, if the homology of the Koszul complex vanishes in degree $\ne 0$. If zero is a regular value of $J$ is well known (and follows from Theorem \[acyckrit\] below) that $J_1,\dots, J_\ell\in \mathcal C^{\infty}(M)$ is a complete intersection. An elementary example of a noncomplete intersection is provided by the moment map for one particle of zero angular momentum, Example \[drehimpuls\], in dimension $n=3$. In this case, using the physicist’s denotation, the Koszul complex can be rewritten as $$\begin{aligned} 0\leftarrow\mathcal C^\infty(T^\mathbbm R^3)\stackrel{<J,\:>}{\leftarrow}\mathbbm R^3\otimes C^\infty(T^\mathbbm R^3)\stackrel{J\times}{\leftarrow}\mathbbm R^3\otimes C^\infty(T^\mathbbm R^3)\stackrel{J\cdot}{\leftarrow}\mathcal C^\infty(T^\mathbbm R^3)\leftarrow 0.\end{aligned}$$ Here $q=(q^1,q^2,q^3)$ and $p=(p_1,p_2,p_3)$ are interpreted as vector valued functions on $T^\mathbbm R^3$ and the angular momentum is $J=q\times p$, where $\times$ denote the vector product in $\mathbbm R^3$. Since $q$ and $p$ are orthogonal to $J$, i.e., the euclidian scalar products $<q,J>=0=<p,J>$ vanish, $q$ and $p$ are one-cycles. In fact, they cannot be boundaries, for if $q=J\times\text{'something'}$ it had to vanish as a function at $p=0$ and vice versa. Since such findings are merely accidental, we would like to have a more systematic way to decide whether a moment map is a complete intersection. One way could be to make a detour and use methods from the theory of commutative Noetherian ring together with flatness arguments. Instead, if we already know that the generating hypothesis is true, then the following ‘Jacobian criterion’ yields a more convenient method. \[acyckrit\] Let M be an analytic manifold and $J:M\to \mathbbm R^\ell$ an analytic map, such that the following conditions are true 1. \[gh\]$(J_1,\dots,J_\ell)$ generate the vanishing ideal of $Z:=J^{-1}(0)$ in $\mathcal C^\infty(M)$, 2. \[jc\]the regular stratum $Z_r:=\{z\in Z\:\|\: T_zJ \mbox{ is surjective}\}$ is dense in $Z:=J^{-1}(0)$. Then the Koszul complex $K:=K(\mathcal C^\infty(M),J)$ is acyclic and $H_0=\mathcal C^\infty(Z)$. We will show that the Koszul complex $K(\mathcal C^\omega_x(M),J) $ is acyclic for the ring $\mathcal C^\omega_x(M)$ of germs in $x$ of real analytic functions. Then it will follow that the Koszul complex $K(\mathcal C^{\infty}(M),J)$ is acyclic, since the ring of germs of smooth functions $\mathcal C^\infty_x(M)$ is (faithfully) flat over $\mathcal C^\omega_x(M)$ (see [@Malgrange Corollary VI 1.12]), and the sheaf of smooth functions on $M$ is fine. Since $\mathcal C^\omega_x(M)$ is Noetherian, Krull’s intersection theorem says that $\cap_{r\ge0} I^r_x=0$, where $I_x$ is the ideal of germs of analytic functions vanishing on $Z$. According to [@Bourbdix A X.160], it is therefore sufficient to show that $H_1(\mathcal C^\omega_x(M),J)=0$. Note that since $J$ generates the vanishing ideal of $Z$ in $\mathcal C^\infty(M)$, it also generates the vanishing ideal of $Z$ in $\mathcal C^\omega_x(M)$. This can easily be seen using M. Artin’s approximation theorem (see e.g.[@Ruiz]).[^6] Suppose $f=\sum_a f^a\:e_a\in K_1$ is a cycle, i.e. $\partial f=\sum_a J_af^a=0$. Since the restriction to $Z$ of the Jacobi matrix $D(\sum_a J_af^a)$ vanishes, we conclude (using condition b)) that $f^a_{\|Z}=0$ for all $a=1,\dots,\ell$. Since $J$ generates the vanishing ideal, we find an $\ell\times\ell$-matrix $F= (F^{ab})$ with smooth (resp. analytic) entries such that $f^a=\sum_b F^{ab}J_b$.[^7] It remains to be shown, that this matrix can be choosen to be antisymmetric. We have to distinguish two cases. If $x\notin Z$, the claim is obvious, since then one can take for example $F^{ab}:=(\sum_a J_a^2)^{-1}(J_bf^a-J_a f^b)$. So let us consider the case $x \in Z$. We then introduce some formalism to avoid tedious symmetrization arguments. Let $E$ denote the free $k:=\mathcal C^\omega_x(M)$-module on $\ell$ generators, and consider the Koszul-type complex $SE\otimes \wedge E$. Generators of the symmetric part will be denoted by $\mu_1,\dots,\mu_\ell$, generators of the Grassmann part by $e_1,\dots,e_\ell$, respectively. We have two derivations $\delta:=\sum_a e_a\wedge\frac{\partial}{\partial \mu_a}:S^nE\otimes \wedge^m E\to S^{n-1}E\otimes \wedge^{m+1}E$, and $\delta^:=\sum_a \mu_a i(e^a):S^nE\otimes \wedge^m E\to S^{n+1}E\otimes \wedge^{m-1}E$. They satisfy the well known identities: $\delta^2=0$, $(\delta^)^2=0$ and $\delta\delta^+\delta^\delta=(m+n){\operatorname{id}}$. Furthermore, we introduce the two commuting derivations $i_J:=\sum_a J_a i(e^a)$ and $d_J=\sum_a J_a\frac{\partial}{\partial \mu_a}$. They obey the identities $i_J^2=0$, $[i_J,\delta]=d_J$, $[d_J,\delta^]=i_J$ and $[i_J,\delta^]=0=[d_J,\delta]$. We interprete the cycle $f$ above as being in $E\otimes k$ and the matrix $F$ as a member of $E\otimes E$. We already know that $d_Jf=0$ implies $f=i_JF$. This argument may be generalized as follows: if $a\in S^nE\otimes k$ obeys $d_J^na=0$, then there is an $A\in S^nE\otimes E$ such that $a=i_J A$. The proof is easily provided by taking all $n$-fold partial derivatives of $d_J^na=0$, evaluating the result on $Z$ and using conditions a) and b). We now claim that there is a sequence of $F_{(n)}\in S^{n+1}E\otimes E$, $n\ge 0$, such that $F=F_{(0)}$, $\delta^F_{(n)}=(n+2)i_JF_{(n+1)}$ and $$\begin{aligned} \label{nteGl} f=d_J^ni_JF_{(n)}+i_J\delta^\underbrace{\Big(\sum_{i=0}^{n-1}\frac{1}{i+2}\:d_J^i\delta F_{(i)}\Big)}_{=:B_{n-1}}\quad\mbox{ for all }n\ge1.\end{aligned}$$ We prove this by induction. Setting $B_{-1}:=0$, we may start the induction with $n=0$, where nothing has to be done. Suppose now, that the claim is true for $F_{(0)},\dots ,F_{(n)}$. We obtain $f=\frac{1}{n+2}d_J^ni_J\big(\delta \delta^F_{(n)}+\delta^\delta F_{(n)}\big)+ i_J\delta^B_{n-1}=\frac{1}{n+2}d_J^{n+1}\delta^F_{(n)}+i_J\delta^B_n$, where we made use of the relations $[d^n_Ji_J,\delta^]=0$ and $[d^n_Ji_J,\delta]=d^{n+1}_J$. Since $0=d_Jf=d_J^{n+2}\delta^F_{(n)}$, we find an $F_{(n+1)}$ such that $\frac{1}{n+2}\delta^F_{(n)}=i_JF_{(n+1)}$, and the claim is proven. Finally, we would like to take the limit of equation (\[nteGl\]) as $n$ goes to $\infty$. For this limit to make sense, we have to change the ring to the ring of formal power series. Let us denote this change of rings by $\hat{}:\mathcal C_x^\omega(M)\to\mathbbm K[[x^1,\dots,x^n]]$. Since by Krull’s intersection theorem $\cap_{r\ge0} \hat{I}^r=0$ ($\hat I$ the ideal generated by $\hat J_1,\dots,\hat J_{\ell}$), we obtain a formal solution of the problem: $\hat f=i_{\hat J}\delta^B_{\infty}$, where $B_{\infty}:=\sum_{i=0}^{\infty}\frac{1}{i+2}\: d_{\hat J}^i\delta \hat F_{(i)}$ is well defined since $\hat I$ contains the maximal ideal. Applying M. Artin’s approximation theorem yields an analytic solution, and we are done. \ The above reasoning can be considered to be folklore, as the subtlety of finding an antisymmetric source term is often swept under the rug in semirigorous arguments. We do not know, whether, if condition \[gh\].) in Theorem \[acyckrit\] holds, \[jc\].) is also sufficient for the acyclicity of the Koszul complex. Nonetheless, let us, as a plausibility check, reconsider the system of one particle of zero angular momentum in dimension $n\ge 3$, Example \[drehimpuls\]. Here, the Jacobi matrix $T_z J$ is a ${n\choose 2}\times 2n$-matrix. Since the zero fibre $Z$ is the set of points where $p$ and $q$ are proportional, it follows easily that $T_z J$ has the same rank as a certain ${n\choose 2}\times n$ submatrix, which in fact occurs twice in $T_z J$. With a little more effort, one may proof that the rank of this submatrix is in fact $\le n-1$. Since ${n\choose 2}>n-1$ for $n\ge 3$ the regular stratum is empty here. For the Examples \[mpart\], \[lemon\], \[stratres\], \[tzwo\] (for $\alpha<0$) and \[commvar\] we already know from subsection \[ghsubsection\] that condition \[gh\].) of Therorem \[acyckrit\], i.e., the generating hypothesis, is fulfilled. Let us now check that condition \[jc\].) of Theorem \[acyckrit\] holds for these examples as well. For Example \[mpart\], i.e., the system of $m\ge 1$ particles of zero angular momentum in the plane, and for the $(1,1,-1,-1)$-resonance, Example \[stratres\], the Jacobi matrix $T_x J$ is clearly not onto iff $x=0$, i.e., the singularity is isolated in $M$. Clearly, in both cases $Z\ne \{0\}$ and hence condition \[jc\].) is true. In the case of the ‘lemon’, Example \[lemon\], an easy calculation using affine coordinates yields that the singular points of the moment map are precisely the fix points of the $S^1$-action: $((1:0),(1:0))$, $((0:1),(0:1))$, $((1:0),(0:1))$ and $((0:1),(1:0))$. The latter two are isolated points in the zero level $Z=J^{-1}(0)$ (which is obviously nondiscrete), and hence the requisites of Theroem \[acyckrit\] are fulfilled here. For the $\mathbbm T^2$-action of Example \[tzwo\], $\alpha<0$, the Jacobi matrix of $J$ works out as follows $$\begin{aligned} \left( \begin{array}{cccccccc} \alpha \bar{z}_1&\alpha z_1 & 0 &0 &\bar{z}_3&z_3&0 &0\\ \beta \bar{z}_1&\beta z_1 &\bar{z}_2&z_2& 0 & 0 &\bar{z}_4&z_4\\ \end{array} \right).\end{aligned}$$ The set of points where this matrix has not the full rank is just the union $\cup _{i=1}^4 L_i$ of the coordinate lines $L_1:=\{(z,0,0,0)\:\|\:z\in \mathbbm C\}$, $L_2:=\{(0,z,0,0)\:\|\:z\in \mathbbm C\}$, $L_3:=\{(0,0,z,0)\:\|\:z\in \mathbbm C\}$ and $L_4:=\{(0,0,0,z)\:\|\:z\in \mathbbm C\}$. Clearly, $Z\cap\cup_i L_i=\{0\}\ne Z$ and hence condition \[jc\].) is true. Finally, let us address Example \[commvar\], i.e., the commuting variety. Let $(Q,P)\in S\times S=T^S$ be a pair of symmetric $n\times n$-matrices. Then the Jacobi-matrix of $J$ applied on the tangent vector $(V,W)\in S\oplus S$ yields $$\begin{aligned} TJ_{(Q,P)}(V,W)=\frac{d}{dt}[Q+tV,P+tW]_{\|t=0}=[V,P]+[Q,W].\end{aligned}$$ It follows easily that $TJ_{(Q,P)}$ is surjective for $(Q,P)$ from an open dense subset of $S\times S$. In fact, let $Q\in S_{\operatorname{reg}}$ have pairwise distinct Eigenvalues $q_i$, $q_i\neq q_j$ for $i\ne j$. After an orthogonal change of the basis we may assume that $Q$ is diagonal: $Q=\sum_i q_i E_{ii}$. Here, $\left(E_{ij}\right)_{kl}=\delta_{ik}\delta_{jl}$ is the matrix with the only nonzero entry $1$ at the $i$th row and $j$th column. For any $P\in S$ we have $TJ_{(Q,P)}(E_{ij}+E_{ji},0)=[Q,E_{ij}+E_{ji}]=(q_i-q_j)(E_{ij}+E_{ji})$. Since $(E_{ij}+E_{ji})_{ij}$ constitute is a basis for $S$, it follows that $TJ_{\|S_{\operatorname{reg}}\times S}$ is surjective. Since $(Q,P)\in Z\cap S_{\operatorname{reg}}\times S$ $\Leftrightarrow$ $Q\in S_{\operatorname{reg}}$ and $P\in S$ are simultaneously diagonalizable, the claim follows. The next theorem is a consequence of rather deep analytic results. The problem of splitting the Koszul resolution in the context of Fréchet spaces has been also addressed in [@DomJac] from a different perspective. Let $M$ be a smooth manifold, $J:M\to \mathbbm R^\ell$ be a smooth map such that around every $m\in M$ there is a local chart in which $J$ is real analytic. Moreover, assume that the Koszul complex $K=K(M,J)$ is a resolution of $\mathcal C^\infty(Z)$, $Z=J^{-1}(0)$. Then there are a prolongation map ${\operatorname{prol}}: \mathcal C^{\infty}(Z)\to \mathcal C^{\infty}(M)$ and contracting homotopies $h_i: K_i\to K_{i+1}$, $i\ge 0$, which are continuous in the respective Fréchet topologies, such that $$\begin{aligned} \label{KosContr} \big(\mathcal C^{\infty}(Z),0\big)\: \begin{array}{c}{{\operatorname{res}}}\\\leftrightarrows\\{{\operatorname{prol}}} \end{array}\:\big(K,\partial\big),h\end{aligned}$$ is a contraction, i.e., ${\operatorname{res}}$ and ${\operatorname{prol}}$ are chain maps and ${\operatorname{res}}\:{\operatorname{prol}}={\operatorname{id}}$ and ${\operatorname{id}}-{\operatorname{prol}}\:{\operatorname{res}}=\partial h+h\partial$. If necessary, these can be adjusted in such a way, that the side conditions (see Appendix \[HPT\]) $h_0\:{\operatorname{prol}}=0$ and $h_{i+1}\:h_i=0$ are fulfilled. If, moreover, a compact Lie group $G$ acts smoothly on $M$, $G$ is represented on $\mathbbm R^\ell$ and $J:M\to \mathbbm R^\ell$ is equivariant, then ${\operatorname{prol}}$ and $h$ can additionally be chosen to be equivariant. A closed subset $X\subset\mathbbm R^n$ is defined to have the extension property, if there is a continuous linear map $\lambda:\mathcal C^\infty(X)\to \mathcal C^\infty(\mathbbm R^n)$, such that ${\operatorname{res}}\: \lambda={\operatorname{id}}$. The extension theorem of E. Bierstone and G. W. Schwarz, [@Schwarzbier Theorem 0.2.1] says that Nash subanalytic sets (and hence closed analytic sets) have the extension property. Using a partition of unity, we get a continuous linear map $\lambda:\mathcal C^\infty(Z)\to \mathcal C^\infty(M)$, such that ${\operatorname{res}}\: \lambda={\operatorname{id}}$. In the same reference, one finds a “division theorem” (Theorem 0.1.3.), which says that for a matrix $\varphi\in \mathcal C^\omega(\mathbbm R^n)^{r,s}$ of analytic functions the image of $\varphi:\mathcal C^\infty(\mathbbm R^n)^s\to C^\infty(\mathbbm R^n)^r$ is closed, and there is a continuous split $\sigma:\mathrm{im}\:\varphi\to\mathcal C^\infty(\mathbbm R^n)^s$ such that $\varphi\:\sigma={\operatorname{id}}$. Using a partition of unity, we conclude that there are linear continuous splits $\sigma_i:\mathrm{im}\:\partial_{i+1}\to K_{i+1}$ for the Koszul differentials $\partial_{i+1}:K_{i+1}\to K_i$ for $i\ge 0$, i.e., $\partial_{i+1}\:\sigma_i={\operatorname{id}}$. We observe that $\mathrm{im}\:\lambda\oplus\mathrm{im}\:\partial_1=K_0$, since for every $x\in K_0$ the difference $x-\lambda\:{\operatorname{res}}x$ is a boundary due to exactness and the sum is apparantly direct. Similarly, we show that $\mathrm{im}\:\sigma_i\oplus\mathrm{im}\:\partial_{i+2}=K_{i+1}$ for $i\ge 0$. The next step is to show that $\mathrm{im}\:\sigma_i$ is a closed* subspace of $K_0$. Therefor we assume that $(x_n)_{n\in\mathbbm N}$ is a sequence in $\mathrm{im}\:\partial_{i+1}$ such that $\sigma_i(x_n)$ converges to $y\in K_{i+1}$. Then $x_n=\partial_{i+1}\sigma_i(x_n)$ converges to $\partial_{i+1} y$, since $\partial_{i+1}$ is continuous. Since $\partial_{i+1} y$ is in the domain of $\sigma_i$, we obtain that $\sigma_i(x_n)$ converges to $\sigma_i\partial_{i+1} y=y\in \mathrm{im}\:\sigma_i$. Similarly, we have that $\mathrm{im}\:\lambda$ is a closed subspace of $K_0$. Altogether, it is feasible to extend $\sigma_i$ to a linear continuous map $K_i\to K_{i+1}$ (cf. [@Rudin p.133]). If necessary, $\lambda$ and $\sigma_i$ can be made equivariant by averaging over $G$, since ${\operatorname{res}}$ and $\partial$ are equivariant. We observe that we have $\lambda {\operatorname{res}}_{\|\mathrm{im}\lambda}={\operatorname{id}}$ and $\lambda {\operatorname{res}}_{\|\mathrm{im}\partial_1}=0$ and analogous equations in higher degrees. We now replace $\lambda$ by ${\operatorname{prol}}:=\lambda-\partial_1\sigma_0 \lambda$ and $\sigma_i$ by $h_i:=\sigma_i-\partial_{i+2}\sigma_{i+1}\sigma_i$ for $i\ge0$. These maps share all of the above mentioned properties with $\lambda$ and $\sigma_i$. Additionally, we have ${\partial_1 h_0}_{\|\mathrm{im}({\operatorname{prol}})}=0$ and ${\partial_{i+2}h_{i+1}}_{\|\mathrm{im}(h_i)}=0$ for $i\ge0$. This concludes the construction of (\[KosContr\]). The side conditions can be achieved by algebraic manipulations (see Appendix \[HPT\]). Note that these modifications do not ruin the equivariance. \ If the constraint surface $Z$ is singular it seems to be hopeless to find explicit formulas for ${\operatorname{prol}}$ and $h$. However, if $Z$ is a closed submanifold there is a general recipe [@BHW] to produce such formulas. We sketch this procedure for the case of linear Poisson structure, Example \[linPoiss\], which essentially contains already the whole idea. Recall that, if $x_1,\dots,x_n$ are linear coordinates for $M=\mathfrak h^$, the moment map is given by the projection $J:M\to\mathfrak g^$, $J(x_1,\dots,x_n)=(x_1,\dots ,x_\ell)$. The zero fibre $Z=J^{-1}(0)$ is the linear subspace on which the first $\ell$ coordinates vanish. The prolongation map ${\operatorname{prol}}:\mathcal C^\infty(Z)\to\mathcal C^\infty(M)$ is just $$\begin{aligned} \label{prolformula} ({\operatorname{prol}}f)(x_1,\dots,x_n):=f(x_{\ell+1},\dots,x_n).\end{aligned}$$ If $f\in\mathcal C^\infty(M)$ and $v\in S^k_\mathbbm K(\mathfrak g[1])$ then the contracting homotopy is given by the formula $$\begin{aligned} \label{hformula} h(fv):=\sum_{a=1}^\ell\left(\int_0^1 dt\:t^k\frac{\partial f}{\partial x_a}(tx_1\dots,tx_\ell,x_{\ell+1},\dots,x_n)\right)\:\xi_av.\end{aligned}$$ By linear extension this defines the contracting homotopy $h:K_k\to K_{k+1}$. An easy calulation yields that in this case (\[KosContr\]) is a contraction fulfilling the side conditions $h_0{\operatorname{prol}}=0$ and $h^2=0$. For completeness, we conclude this section by mentioning two consequences of the Koszul resolution, which are of course well known to commutative algebraists. \[freebasis\]Let $J=(J_1,\dots,J_\ell): M\to \mathbbm R^\ell=:V^$ such that the Koszul complex $K(A,J)$ is a resolution of $\mathcal C^\infty(Z)=\mathcal C^\infty(M)/I$. Then the image of $J_1,\dots,J_\ell$ under the projection to $I/I^2$ is a free system of generators for the $A/I$-module $I/I^2$. Let $\sum_a f^a [J_a]=0$, where $f^1\dots, f^\ell\in C^\infty(Z)$ and $[J_a]$ are the representatives of the $J_a$’s in $I/I^2$. By choosing $F^1, \dots F^\ell\in C^\infty(M)$ such that $f_a={F_a}_{\|Z}$ for $a=1\dots\ell$, we may rewrite this as $\sum_a F_a J_a\in I^2$. Defining $F:=\sum_a F^a e_a \in K_1(A,J)$ we obtain $\partial F=\sum_{a,b}G^{ab}J_aJ_b\in I^2$, for some $G^{ab}\in \mathcal C^\infty(M)$. So $F-\sum_{ab}G^{ab}J_ae_b$ is a 1-cycle. Since the first homology of $K(A,J)$ vanishes, there is an $H:=\sum H^{ab} e_ae_b\in K_2(A,J)$ such that $F=\sum_{ab}G^{ab}J_ae_b+\partial H$, and thus all $F^1,\dots ,F^\ell$ are in $I$. \[exttor\] Let $J=(J_1,\dots,J_\ell)$ be as above. Then there are isomorphisms of $\mathbbm K$-vector spaces $$\begin{aligned} {\operatorname{Tor}}^A_i(A/I,A/I)&\cong&\wedge^i_{A/I}I/I^2\quad\mbox{and}\\ {\operatorname{Ext}}^i_A(A/I,A/I)&\cong& {\operatorname{Hom}}_{A/I}(\wedge^i_{A/I}I/I^2,A/I)\end{aligned}$$ for all $i\ge 0$. In order to compute ${\operatorname{Tor}}^A_i(A/I,A/I)$, we have to compute the homology of the complex $(K(A,J)\otimes _A A/I,\partial \otimes 1)$. But this is evidently the homology of the Koszul complex $K(A/I,(0,\dots,0))$. The claim follows from Theorem \[freebasis\]. It is easy to see, that ${\operatorname{Ext}}^i_A(A/I,A/I)$ is the cohomology of the complex ${\operatorname{End}}_A(K(A,J))$ with differential $D$ equal to the super-commutator with $\partial$. We may identify ${\operatorname{End}}_{{{A}\mathbf{-Mod}^{\mathbbm Z}}}(K(A,J))={\operatorname{End}}_{{{A}\mathbf{-Mod}^{\mathbbm Z}}}(\wedge v\otimes A)={\operatorname{End}}_{{{\mathbbm K}\mathbf{-Mod}^{\mathbbm Z}}}(\wedge V)\otimes A$ with the Clifford algebra bundle over $M$ with fibre $\wedge V^\otimes \wedge V$. The latter is identified with ${\operatorname{End}}_{\mathbbm K}(\wedge V)$ using the standard representation: $x\in V$ acts on $\wedge V$ by left multiplication, and $\alpha\in V^$ acts on $\wedge V$ by $-1$ times the insertation derivation $i(\alpha)$. The analog of composition in ${\operatorname{End}}_{\mathbbm K}(\wedge V)$ is the Clifford multiplication $\mu\circ(-\sum_a i(e^a)\otimes i(e_a))$, where $e_1,\dots,e_\ell$ is a basis for $V$ and $e^1,\dots,e^\ell$ is the corresponding dual basis. Now the analog of $D$ in the Clifford algebra is just the commutator with $\sum_a J_a e^a$. The latter actually is the total differential of a double complex with acyclic rows and trivial columns. \ Actually, one can make the above statement more precise. It is well known [@Bourbdix §7], that ${\operatorname{Ext}}^\bullet_A(A/I,A/I)$ is $\mathbbm Z$-graded $\mathbbm K$-algebra and ${\operatorname{Tor}}^A_\bullet(A/I,A/I)$ is a graded module for this algebra. The isomorphisms of corollary \[exttor\] reflect this structure. The “produit de composition” is given up to a sign by the wedge product and module structure by insertation, respectively. In particular, the above isomorphisms are isomorphisms of $A/I$-modules. The projective Koszul resolution of a closed submanifold {#projkos} -------------------------------------------------------- Even in the regular case not every constraint surface admits a Koszul resolution. For a closed codimension $\ell$ submanifold $C$ of the manifold $M$, we know from Theorem \[freebasis\], that if there is a Koszul complex, which is a resolution of $\mathcal C^\infty(C)$ then the conormal bundle of $C$ in $M$ is trivial. Since from the homological point of view projective modules are as good as free modules, one may ask whether the situation improves if one also accepts projective Koszul resolutions. We will see that after restricting to an appropriate open neighborhood $U$ of $C$ in $M$, there is a $\mathcal C^\infty(U)$-projective resolution for every closed submanifold $C$. Let $TC$ be the tangent bundle of $C$ and let $TM_{\|C}$ and $TM^_{\|C}$ be the restrictions of the tangent bundle and the cotangent bundle of $M$ to $C$, respectively. The annihilator bundle $TC^{{\operatorname{ann}}}$ of $C$ is the subbundle of $TM^_{\|C}$ consisting of all 1-forms vanishing on $TC$. The dual bundle ${TC^{{\operatorname{ann}}}}^$ to the annihilator bundle is canonically isomorphic to $TM_{\|C}/TC$. It is well known that the space of sections of the annihilator bundle has a nice algebraic description. \[conormallemma\] The the map $f\mapsto {\mathrm d}f_c$ which associates to a smooth function $f\in \mathcal C^\infty(M)$ its differential evaluated at the point $c\in C$ induces an isomorphism of $\mathcal C^\infty(C)$-modules of conormal module $I/I^2$ of the vanishing ideal $I$ of $C$ and the space of sections $\Gamma^\infty(C,TC^{{\operatorname{ann}}})$ of the annihilator bundle $TC^{{\operatorname{ann}}}$ of the submanifold $C$. For a detailed proof which uses the Koszul complex constructed below see [@opusmagnum]. \ Let us recall the tubular neighborhood theorem for the submanifold $C\subset M$. There exists a subfibre bundle $N_C$ of $TM_{\|C}$ which is complementary to $TC$ $$\begin{aligned} TM_{\|C}=TC\oplus N_C,\end{aligned}$$ an open neighborhood $U'$ of the zero section of $N_C$ which is a disk bundle over $C$ and a diffeomorphism $\varphi$ $$\begin{aligned} \xymatrix {N_C\supset U'\qquad\ar[rr]^{\operatorname{\varphi}}&&\qquad U\subset M\\ &\quad C\quad \ar@{^{(}->}[ul]\ar@{^{(}->}[ur]&}\end{aligned}$$ into an open neighborhood $U\subset M$ whose restriction to $C$ is the identity. In fact, $\varphi$ may be given as a restriction of the exponential map corresponding to an auxiliary Riemannian metric on $M$. In this way we obtain a (noncanonical) identification of $N_C$ and ${TC^{{\operatorname{ann}}}}^$. Moreover, we have a submersion $$\begin{aligned} \tau:U\to C.\end{aligned}$$ Since $U'$ is fibrewise convex it makes sense to take a convex linear combination of any set of points of $U$ which lie in the same fibre of $\tau$. In particular for any $t\in [0,1]$ we have a fibre preserving shrinking map $\Phi_t:U\to U$, which is obtained from the convex combination $tu+(1-t)u_0$, where $u_0\in C\subset U$ and $u$ are in the same fibre. Let $V^:= {\operatorname{ker}}T\tau$ be the vertical subbundle of $TU$ of the projection $\tau:U\to C$. There is a section $J\in\Gamma^\infty(U,V^)$, which we will call the tautological section, such that the submanifold $C$ is the zero locus of $J$. It is given by the image under $T\varphi$ of the Euler vector field $\in\Gamma^\infty(U',TN_C)$ of the vector bundle $N_C$. Note that the restriction $V^_{\|C}$ to $C$ of the bundle $V^$ is isomorphic to ${TC^{{{\operatorname{ann}}}}}^$. The projective (homological) Koszul complex $(K_{\bullet}(U,J),\partial)$ on the tautological section $J$ is defined as follows. The space of chains is $K_{\bullet}(U,J):=S_{\mathcal C^\infty(U)}(V[-1])$. In more conventional terminology, it also may be viewed as $\Gamma^\infty(U,\wedge^\bullet V)$. The differential is the unique $\mathcal C^\infty(U)$-linear superderivation, such that $\partial(f)=<J,f>$ for all $f\in K_1(U,J)$, where $<\:,\:>$ denotes the dual pairing. By augmenting this complex with the restriction map we obtain a sequence $$\begin{aligned} 0\longleftarrow\mathcal C^\infty(C)\stackrel{{\operatorname{res}}}{\longleftarrow}\mathcal C^\infty(U)=K_0(U,J)\stackrel{\partial}{\longleftarrow}K_1(U,J)\stackrel{\partial}{\longleftarrow}K_2(U,J)\longleftarrow\dots\end{aligned}$$ In fact, this sequence is exact, which can of course be proven locally without difficulty. We will exhibit explicit contracting homotopies similar to the equations (\[prolformula\]) and (\[hformula\]) for the linear case. For the prolongation map $${\operatorname{prol}}:=\tau^:\mathcal C^\infty(C)\to\mathcal C^\infty(U)$$ we simply take the pull back of the projection $\tau:U\to C$. The contracting homotopies $h_i:K_i(U,J)\to K_{i+1}(U,J)$ for $i=0,1,\dots,\ell$ are defined by the formula (which is due to M. Bordemann [@opusmagnum]) $$h_i(a)(u):=\int_0^1dt\: (\Phi_t^(da))(u)$$ for $a\in K_i(U,J)$ and $u\in U$. In order to understand the right hand side of this equation note that the space $K_i(U,J)$ can be identified with the space of cochains of Lie algebroid cohomology with coefficients in $\mathcal C^\infty(U)$ of the Lie algebroid $V^:= {\operatorname{ker}}T\pi\to U$, the differential is denoted by $d:K_i(U,J)\to K_{i+1}(U,J)$. Since $\Phi_t$ is fibre preserving, the pullback $\Phi_t^$ with respect to $\Phi_t$ for differential forms descends to a well defined map $\Phi_t^:K_i(U,J)\to K_i(U,J)$ for all $i$. The maps ${\operatorname{prol}}$ and $h$ defined as above are continuous $\mathbbm K$-linear maps, such that $$\begin{aligned} \label{prKosContr} \big(\mathcal C^{\infty}(C),0\big)\: \begin{array}{c}{{\operatorname{res}}}\\\leftrightarrows\\{{\operatorname{prol}}} \end{array}\:\big(K_\bullet(U,J),\partial\big),h\end{aligned}$$ is a contraction. The proof can be reduced to a local computation, which is well-known. \ Similar to Corollary \[exttor\] we can use the projective Koszul resolution to compute $$\begin{aligned} {\operatorname{Tor}}_\bullet^{\:\mathcal C^\infty(U)}\big(\mathcal C^\infty(C),\mathcal C^\infty(C)\big)&=&\Gamma^\infty(C,\wedge^\bullet\:TC^{{\operatorname{ann}}})\\ {\operatorname{Ext}}^\bullet_{\:\mathcal C^\infty(U)}\big(\mathcal C^\infty(C),\mathcal C^\infty(C)\big)&=&\Gamma^\infty(C,\wedge^\bullet\:{TC^{{\operatorname{ann}}}}^).\end{aligned}$$ Projective Koszul-Tate complexes {#projkostate} -------------------------------- For the moment, let $A$ be an arbitrary commutative $\mathbbm K$-algebra, think of it as the algebra of smooth functions on a manifold or the algebra of germs of real analytic functions etc., and $I$ be an ideal in $A$. In the spirit of the preceding subsection we will need the notion of a projective presentation of an $A$-module $\mathcal W$, that is, a short exact sequence of $A$-modules $$\mathcal V\stackrel{J}{\to}\mathcal W\to 0,$$ where $\mathcal V$ is a projective $A$-module. It is called finite if $\mathcal V$ is finitely generated. We will call $J$ a system of projective generators for $\mathcal W$. A typical example is the tautological section of the preceding subsection, which is a system of projective generators for the vanishing ideal of the submanifold $C$. In the following we will be exclusively interested in the situation, where either 1. $A$ is arbitrary, but $\mathcal V$ is a finitely generated free $A$-module, or 2. \[vectorbundle\] $A=\mathcal C^\infty(M)$ and $\mathcal V$ is a finitely generated projective $A$-module, i.e., the space of sections of a vector bundle $V$ over $M$. In any case, there is a good notion of a basis for $\mathcal V$. In case \[vectorbundle\]. this will be a local frame for $V$. Generalizing the notion of Koszul resolution for modules over a (commutative) Noetherian ring, Tate [@Tate] introduced what is nowadays called a Tate resolution. For an exository article on the theory of Tate-resolutions for Noetherian (local) rings we refer to [@Avramov section 6 and 7]. Since we are concerned with the case $\mathbbm Q\subset \mathbbm K$ we can use a slightly simplified version of this construction (we replace the algebra of divided powers by the symmetric algebra). On the other hand, the ring of primary interest for us, $A=\mathcal C^\infty(M)$, is not a Noetherian local ring. Therefore, it makes sense to slightly generalize the construction by considering projective Tate generators. The potential of Tate resolutions for being useful in phase space reduction has already been recognized by the theoretical physicists Batalin, Fradkin and Vilkovisky in the early 80s [@BF; @BV1; @BV2; @BV3]. In [@Stashbull] Stasheff recognized that their “ghost for ghost” procedure and that of Tate [@Tate] essentially coincide. Traditionally in the mathematical physics literature Tate resolutions are called Koszul-Tate resolutions. A projective Koszul-Tate resolution of the $A$-module $A/I$ is a direct limit of a directed system of projective Koszul-Tate complexes $(KT_\bullet^{\le i},\partial^{\le i})$ of level $i\ge 1$. These are defined inductively as follows. The (nonunique) projective Koszul-Tate complex of level $i\ge 1$ over the module $A/I$ for an ideal $I$ of the commutative ring $A$ is a complex $(KT_\bullet^{\le i},\partial^{\le i})$ $$\begin{aligned} 0 \leftarrow A=KT^{\le i}_0\stackrel{\partial^{\le i}}{\leftarrow}KT^{\le i}_1\stackrel{\partial^{\le i}}{\leftarrow}KT^{\le i}_2\leftarrow\dots \leftarrow KT^{\le i}_i\leftarrow KT^{\le i}_{i+1}\leftarrow\dots,\end{aligned}$$ such that 1. $KT^{\le i}_k$ is the degree $k$ part of a graded symmetric algebra $S_A(\mathcal V^{\le i})$ generated by an $\mathbbm N$-graded projective $A$-module $\mathcal V^{\le i}=\oplus_{j= 1}^i \mathcal V^j$. 2. $\partial^{\le i}$ is an $A$-linear derivation of $KT^{\le i}=S_A(\mathcal V^{\le i})$. 3. The restriction map ${\operatorname{res}}:A\to A/I$ gives an isomorphism $\operatorname H_0 KT^{\le i}\cong A/I$. 4. $KT^{\le i}_k$ is acyclic up to degree $i-1$, that is $\operatorname H_k KT^{\le i}=0$ for $k=1,\dots ,i-1$. 5. For all $i\le j$ the obvious injections $KT^{\le i}\to KT^{\le j}$ are maps of supercommutative differential graded $\mathbbm K$-algebras. 6. For $i=1$ the restriction $J_{(1)}$ of $\partial ^{\le 1}$ to $\mathcal V^1=KT^{\le 1}_1$ is a projective presentation of the $A$-module $I$. For $i\ge2$ the restriction $J_{(i)}$ of $\partial^{\le i}$ to $\mathcal V^i\subset KT^{\le i}_i$ is a system of projective generators for the $A$-module $\operatorname H_i KT^{\le i}$. More precisely, we lift a projective presentation $\mathcal V^i$ of the $A$-module $\operatorname H_{i-1} KT^{\le i-1}$ to the space of cycles $$\begin{aligned} \label{ghostforghost} \xymatrix{&&&\mathcal V^i\ar[dl]_{J_{(i)}}\ar[d]&\\ 0\ar[r]&\operatorname B_{i-1} KT^{\le i-1}\ar[r]&\operatorname Z_{i-1} KT^{\le i-1}\ar[r]&\operatorname H_{i-1} KT^{\le i-1}\ar[r]\ar[d]&0.\\ &&&0&}\end{aligned}$$ The elements of $\mathcal V^j$ are called Tate generators of level $j$. In physics they are called antighosts of level $j$. Note, that $(KT^{\le 1}_\bullet,\partial^{\le1})$ coincides with (a projective version of) the usual Koszul complex for $A/I$. The process described by diagram (\[ghostforghost\]), which is called by the physicists the “ghost for ghost”-procedure, is named by the commutative algebraists “killing cycles by adding variables”. Given the data $(KT^{\le i}_\bullet,\partial^{\le i})$ for $i=1,2,\dots$ there is a unique differential $\partial$ on $KT_\bullet:=S_A(\oplus_{i\ge 1}\mathcal V^i)$ such that the obvious injections $(KT^{\le i}_\bullet,\partial^{\le i})\to (KT_\bullet,\partial)$ are morphisms of supercommutative differential graded algebras. By construction it is a projective resolution of $A/I$ and will be called the Koszul-Tate resolution henceforth. Sometimes we would like to stress the dependence on the ring and the presentation $J=J_{(1)}$ of the ideal $I$ and will write in this case $KT_\bullet=KT_\bullet(A,J)$. Even though the above definition applies quite generally, there seems to be no a priori reason for the Koszul-Tate complexes to be complexes. We still have to convince ourselves that the operators $\partial^{\le i}$ and $\partial$ are of square zero. At this point we use a (local) basis $\xi_1^{(i)},\dots,\xi_{\ell_i}^{(i)}$ for the $A$-module[^8] $\mathcal V^i$ for $i=1,2,\dots$. The Koszul-Tate differentials $\partial^{\le i}$ and $\partial$ are given by the (local) formulas $$\begin{aligned} \partial^{\le i}=\sum_{j=1}^i\sum _{a=1}^{\ell_j} J^{(j)}_a{\frac{\partial}{\partial \xi^{(j)}_{a}}},\\ \partial=\sum_{j=1}^\infty\sum _{a=1}^{\ell_j} J^{(j)}_a{\frac{\partial}{\partial \xi^{(j)}_{a}}},\end{aligned}$$ where $J^{(j)}_1,\dots, J^{(j)}_{\ell_j}\in KT_{j-1}^{\le j-1}$ are the components of the map $J_{(j)}$ in diagram (\[ghostforghost\]). We proof by induction that $\partial^{\le i}$ is of square zero for $i=1,2,\dots$. For $i=1$ this is obvious. For $i\ge 2$ we obtain $$\begin{aligned} \label{dsquareiszero} (\partial^{\le i})^2=(\partial^{\le i-1})^2+\sum_{a=1}^{\ell_{i}}\big(\partial^{\le i}J^{(i)}_a\big) {\frac{\partial}{\partial \xi^{(i)}_{a}}}+\sum_{a,b=1}^{\ell_{i}}J^{(i)}_a\frac{\partial J_b^{(i)}}{\partial\xi_a^{(i)}}\:{\frac{\partial}{\partial \xi^{(i)}_{b}}}=0.\end{aligned}$$ In the above formula the first term on the right hand side vanishes by induction. The second term vanishes due to diagram (\[ghostforghost\]). Finally, the third term vanishes since $J_1^{(i)},\dots ,J_{\ell_i}^{(i)}\in KT_{i-1}^{\le i-1}$ clearly do not depend on $\xi_1^{(i)},\dots,\xi_{\ell_i}^{(i)}$. Since for every homogeneuos $x\in KT_k$ there is an $i\in\mathbbm N$ such that $x$ is in the subcomplex $KT_k^{\le i}$ we conclude that $\partial ^2=0$. It seems to be an interesting question, whether equation (\[dsquareiszero\]) holds in more complicated situations. Obviously, the Koszul-Tate resolution of $A/I$ is nonunique. However, if $(A,\mathfrak m)$ is a Noetherian local ring[^9](e.g. the ring of germs of real analytic functions), then there is a distinguished Tate resolution, the so-called minimal model of $A/I$, which is uniquely determined up to isomorphism of differential graded algebras [@Avramov Proposition 7.2.4]. Moreover, the number of Tate generators is bounded from below by invariants of the ring $(A,\mathfrak m)$, the so-called deviations. The lower bounds are realized by the minimal model [@Avramov Proposition 7.2.5]. The deviations can be read off from the Poincaré series of $A/I$. It is known [@AvramovPrivate] for the important example \[drehimpuls\] of one particle in dimension $n\ge 3$ with angular momentum zero that the number of Tate generators grows exponentially with the level (in the polynomial setup one can show that $A/I$ is a Golod ring). In fact, due to the rigidity theorem of S. Halperin [@Halperin] we know that if $A/I$ is not a complete intersection, then none of the deviations vanish. Before we show that in the case of moment maps there are reasonable Koszul-Tate resolutions let us introduce some terminology. If there is a smallest integer $L$ such that $\mathcal V^i=0$ for all $i>L$, then $KT_\bullet$ is said to be finitely generated of level $L$. Otherwise it is said to be infinitely generated. It is clear from what is said above that, in general, there is no reason to expect that the Koszul-Tate resolution of a singular moment map is finitely generated. More reasonable is the following property. The Koszul-Tate resolution is said to be locally finite if all the modules $\mathcal V_i$ are finitely generated projective $A$-modules. In this case the rank of the projective module $KT_i$ can be recursively determined from the ranks of the projective modules $\mathcal V^j$, $1\le j\le i$ according to the product formula $$\begin{aligned} \sum_{i\ge 0}\operatorname{rank}(KT_i)\:t^i=\prod_{j\ge 1}\big(1-(-t)^j\big)^{(-1)^{j+1}\operatorname{rank}(\mathcal V^j)}.\end{aligned}$$ The product on the right hand side converges in the $t$-adic topology of $\mathbbm Z[[t]]$. If all the modules $\mathcal V_i$ are free $A$-modules, we say that the Koszul-Tate resolution is free. \[existsfiniteTate\] Let $J:M\to \mathfrak g^$ be the moment map of a linear Hamiltonian action of a compact Lie group $G$ on the real symplectic vector space $M$. Moreover, suppose that $J$ generates the vanishing ideal $I_Z \subset \mathcal C^\infty(M)$ of the zero fibre $Z=J^{-1}(0)$. Then there is a locally finite free Koszul-Tate resolution $\left(KT_\bullet=KT_\bullet(C^\infty(M),J),\partial\right)$ of the ring of smooth functions $C^\infty(Z)$ in $Z$ and a continuous contracting homotopy $h:KT_\bullet\to KT_{\bullet+1}$, a continuous prolongation map ${\operatorname{prol}}:\mathcal C^\infty(Z)\to KT_0$, such that $$\begin{aligned} \label{KTContr} \big(\mathcal C^{\infty}(Z),0\big)\: \begin{array}{c}{{\operatorname{res}}}\\\leftrightarrows\\{{\operatorname{prol}}} \end{array}\:\big(KT,\partial\big),h\end{aligned}$$ is a contraction. First of all, let us identify $M$ with $\mathbbm R^{2n}$ and note that the components of the moment map are quadratic polynomial functions with respect to the canonical coordinates. Since $\mathbbm R[x_1,\dots,x_{2n}]$ is a Noetherian ring, there is a free locally finite Koszul-Tate resolution $KT_\bullet (\mathbbm R[x_1,\dots,x_{2n}]),J)$ of the $\mathbbm R[x_1,\dots,x_{2n}]$-algebra $\mathbbm R[x_1,\dots,x_{2n}]/<J_1,\dots,J_\ell>$ (see [@Avramov]). According to subsection \[ghsubsection\] the latter is nothing but the coordinate ring of the real-variety determined by $J$. The complex $KT_\bullet(C^\infty(M),J)$ we are looking for is obtained by tensoring the complex $KT_\bullet (\mathbbm R[x_1,\dots,x_{2n}]),J)$ with the $\mathbbm R[x_1,\dots,x_{2n}]$-module $\mathcal C^\infty(M)$, and we have to show that this complex is still a resolution of the $\mathcal C^\infty(M)$-module $\mathcal C^\infty(Z)$. By a standard result [@Eisenbud Theorem 7.2] for every $x\in M$ the ring of formal power series $\mathcal F_x$ around $x$, being the completion of the real polynomial ring with respect to the maximal ideal corresponding to $x$, is a flat $\mathbbm R[x_1,\dots,x_{2n}]$-module. Next, $\mathcal F_x$ is a faithfully flat $\mathcal C^\omega_x(M)$-module [@Malgrange Proposition III 4.10]. We conclude that $KT_i(C^\omega_x(M),J)$ is exact for $i\ge 1$. Using a partition of unity and the fact that $\mathcal C^\infty_x(M)$ is a (faithfully) flat $\mathcal C^\omega_x(M)$-module [@Malgrange Corollary VI 1.12], it follows that $KT_i(C^\infty(M),J)$ is exact for all $i\ge 1$. It remains to proof the existence of the continuous prolongation map and the contracting homotopies. But this follows from the results of Bierstone and Schwarz [@Schwarzbier] precisely along the lines of the proof of Theorem \[KosContr\]. \ We expect that the conclusion of the above proposition holds for essentially all moment maps of compact Hamiltonian group actions on symplectic manifolds. The BRST-charge {#BRSTchargesection} --------------- We are now ready to introduce the BRST-algebra, which will turn out to be a differential graded Poisson algebra $\big({\mathscr A_{\mathcal V}},\{\:,\:\}_R,{\mbox{$\mathcal D$}}\big)$. To this end, we view the space of chains $KT=S_A\mathcal V$ of Koszul-Tate resolution of the preceding section as a generalized graded manifold with base manifold $M$. Note that in the infinitely generated case the Koszul-Tate differential is clearly not a derivation of finite type. We will therefore not use the module of finite type derivations in this section. The $\mathbbm Z$-graded algebra ${\mathscr A_{\mathcal V}}=\oplus_{i\in\mathbbm Z}{\mathscr A_{\mathcal V}}^i$ underlying the ghost-cotangent space of $KT=S_A\mathcal V$ is given by $$\begin{aligned} {\mathscr A_{\mathcal V}}^i=\prod_{j\ge 0}KT^{i+j}\otimes KT_j.\end{aligned}$$ Here we have used the shorthand notation $$\begin{aligned} KT^j:=(S_A \mathcal V^)^j\end{aligned}$$ for the degree $j$ part of the symmetric algebra $S_A \mathcal V^$ over the (positively graded) module $\mathcal V^=\oplus_{i\ge 1} {\mathcal V^i}^$, which is dual to the module of Tate-generators. We will use local frames ${\xi_{(i)}^{1}},\dots,{\xi_{(i)}^{\ell_i}}$ for the modules ${\mathcal V^i}^$, the so-called momenta of level $i$ or ghosts of level i, which are dual to the Tate generators ${\xi^{(i)}_{1}},\dots,{\xi^{(i)}_{\ell_i}}$ of level $i$. The Koszul-Tate differential of the preceeding section naturally extends to a derivation $$\begin{aligned} \partial=\sum_{j=0}^\infty \sum_{a=1} ^{\ell_j}J^{(j)}_a{\frac{\partial}{\partial \xi^{(j)}_{a}}},\qquad J^{(j)}_a\in KT^{\le j-1}_{j-1} \text{ for } a=1,\dots \ell_j\end{aligned}$$ of the algebra ${\mathscr A_{\mathcal V}}$. Slightly abusing the language, we introduce the filtration by ghost degree $F^k{\mathscr A_{\mathcal V}}=\oplus_{i\in \mathbbm Z}F^k{\mathscr A_{\mathcal V}}^i$, where $$\begin{aligned} F^k{\mathscr A_{\mathcal V}}^i=\prod_{i+j\ge k}KT^{i+j}\otimes KT_j\end{aligned}$$ $F^k{\mathscr A_{\mathcal V}}$ can be identified with the space of endomorphism which annihilate $\oplus_{i=0}^{k-1}KT_{i}$. The spaces $F^k{\mathscr A_{\mathcal V}}$ form a descending Hausdorff filtration $$\begin{aligned} F^0{\mathscr A_{\mathcal V}}={\mathscr A_{\mathcal V}}\supset F^1{\mathscr A_{\mathcal V}}\supset\dots\supset F^k{\mathscr A_{\mathcal V}}\supset F^{k+1}{\mathscr A_{\mathcal V}}\supset\dots,\end{aligned}$$ which is preserved by the supercommutative multiplication: $F^k{\mathscr A_{\mathcal V}}\:F^l{\mathscr A_{\mathcal V}}\subset F^{k+l}{\mathscr A_{\mathcal V}}$. A $\mathbbm K$-linear map $\varphi:{\mathscr A_{\mathcal V}}\to{\mathscr A_{\mathcal V}}$ is said to be of filtration degree $l$ if $\varphi(F^k{\mathscr A_{\mathcal V}})\subset F^{k+l}{\mathscr A_{\mathcal V}}$. In Section \[RothPB\] we have introduced the Rothstein bracket $\{\:,\:\}_R$. In the finitely generated case we have seen that $\{\:,\:\}_R$ defines a $\mathbb Z$-graded super-Poisson bracket on ${\mathscr A_{\mathcal V}}$. In the infinitely generated case a full proof of the Jacobiidentity has yet to be given. In the following we postulate that the Jacobiidentity fulfilled. (Curiously, for the construction of the BRST-charge we will merely use the Jacobiidentity for the approximating brackets defined below). The filtration by ghost degree and the Rothstein Poisson bracket $\{\:,\:\}_R$ are, in general, not compatible in the following sense: there is no $r\in \mathbb N$ such that $\{F ^k{\mathscr A_{\mathcal V}},F ^l{\mathscr A_{\mathcal V}}\}_R\subset F^{k+l-r}{\mathscr A_{\mathcal V}}$ for all $k,l\in\mathbbm Z$. This phenomenon is due to the algebraic part of the bracket: a ghost of arbitrarily high level can be killed if it is paired with an antighost of the same level. The geometric part of the Rothstein Poisson bracket, however, is filtered: $[[\widetilde{\Pi},F ^k{\mathscr A_{\mathcal V}}],F^l{\mathscr A_{\mathcal V}}]\subset F^{k+l}{\mathscr A_{\mathcal V}}$ (here $[\:,\:]$ denotes the Schouten-Nijenhuis bracket). In order to prove the main result of this section, Theorem \[chargeexists\] below, we need a more refined analysis, which we will explain next. For free Tate resolutions Theorem \[chargeexists\] is a well known result, which goes back to [@BF; @BV1; @BV2; @BV3]. The first rigorous proof seems to be due to J. Stasheff and is sketched in [@Stashbull]. We have also benefited from the more elaborate exposition in [@Kimura]. The refinement consists in examining successively the level $i$ BRST algebras ${\mathscr A_{\mathcal V^{\le{i}}}}^\bullet$ together with the level $i$ Koszul-Tate differentials $$\begin{aligned} \partial ^{\le i}=\sum_{j=0}^i \sum_{a=1} ^{\ell_j}J^{(j)}_a{\frac{\partial}{\partial \xi^{(j)}_{a}}}\end{aligned}$$ and their level $i$ Rothstein Poisson brackets, which will be denoted by $\{\:,\:\}_{\le i}$ for $i=1,2,\dots$. In this manner we obtain a family of differential graded Poisson algebras $$\begin{aligned} \big({\mathscr A_{\mathcal V^{\le{i}}}}^\bullet,\partial^{\le i},\{\:,\:\}_{\le i}\big)_{i=1,2,\dots}. \end{aligned}$$ For $i<j$ the inclusion ${\mathscr A_{\mathcal V^{\le{i}}}}\hookrightarrow{\mathscr A_{\mathcal V^{\le{j}}}}$ is a map of differential graded commutative algebras, which is compatible with the filtration by ghost degree. We emphasize that, due to the presence of curvature, ${\mathscr A_{\mathcal V^{\le{i}}}}\hookrightarrow{\mathscr A_{\mathcal V^{\le{j}}}}$ is, in general, not a Poisson subalgebra! The same remarks apply to the inclusions ${\mathscr A_{\mathcal V^{\le{i}}}}\hookrightarrow{\mathscr A_{\mathcal V}}$ for $i=1,2,\dots$. \[chargeexists\] Let $Z\subset M$ be a first class constraint set, and let $I:=I_Z\subset \mathcal C^\infty (M)=:A$ be the vanishing ideal of $Z$. Furthermore, let $(KT_\bullet(A,J),\partial)$ be a locally finite projective Koszul-Tate resolution of the $A$-module $A/I$. Let ${\mathscr A_{\mathcal V}}$ be the corresponding BRST-algebra together with the Poisson bracket $\{\:,\:\}_R$. Then there exists an element $\theta\in{\mathscr A_{\mathcal V}}^1$ such that 1. $\{\theta,\:\}_R=\partial + \mbox{higher order terms}$, 2. $\{\theta,\theta\}_R=0$. Here, “higher order terms” stands for a $\mathbbm K$-linear derivation of filtration degree 1. We will construct $\theta$ by induction. To this end we make the following Ansatz $\theta_{\le i}:=\sum_{j=1}^i \theta_j$, where $$\begin{aligned} \label{Ansatz} \theta_j=\sum_{a=1}^{\ell_j} J^{(j)}_a\:\xi^a_{(j)}+Q_{(j)}\in {\mathscr A_{\mathcal V^{\le{j}}}}^1,\end{aligned}$$ such that $Q_{(j)}\in {KT_{\le j}^{j+1}}\otimes KT^{\le j}_j$. Note that $Q_{(j)}$ is at least quadratic in the momenta. We will show that the $Q_{(i)}$, $i=1,2,\dots$, may be successively chosen, such that $$\begin{aligned} \label{filtprop} \{\theta_{\le i},\theta_{\le i}\}_{\le i}\in F^{i+1}{\mathscr A_{\mathcal V^{\le{i}}}}^2.\end{aligned}$$ for all $i=1,2,\dots$. From the Ansatz (\[Ansatz\]) it is clear that the $\theta_i$ add up to a well defined $\theta=\sum_{j=1}^\infty \theta_j\in {\mathscr A_{\mathcal V}}^1$. By an argument similar to that in equation (\[rothcompare\]) below we conclude that $$\begin{aligned} \{\theta_{\le i},\theta_{\le i}\}_R-\{\theta_{\le i},\theta_{\le i}\}_{\le i}\in F^{i+1}{\mathscr A_{\mathcal V}}^2.\end{aligned}$$ Since the filtration by ghost degree is Hausdorff we obtain $\{\theta,\theta\}_R=0$. First of all, let us take a closer look at the derivations $${\mbox{$\mathcal D$}}^{\le i}:=\{\theta_{\le i},\:\}_{\le i}:{\mathscr A_{\mathcal V^{\le{i}}}}^\bullet\to{\mathscr A_{\mathcal V^{\le{i}}}}^{\bullet+1}$$ for $i=1,2,\dots$. In fact, the operators ${\mbox{$\mathcal D$}}^{\le i}$ are (as we will see in a moment) of filtration degree $0$. We decompose them into its homogeneous components ${\mbox{$\mathcal D$}}^{\le i}=\sum_{j=0}^\infty {\mbox{$\mathcal D$}}^{\le i}_j$ (see also diagram (\[brsdecomp\]) below), where ${{\mbox{$\mathcal D$}}_j^{\le i}}\big({KT^k_{\le i}}\otimes KT_l^{\le i}\big)\subset {KT^{k+j}_{\le i}}\otimes KT_{l+j-1}^{\le i}$ for all $k,l\ge 0$. A crucial observation is that the term of lowest order in this decomposition is exactly the level $i$ Koszul-Tate differential $$\begin{aligned} \label{filtrationisgood} \partial^{\le i}={\mbox{$\mathcal D$}}^{\le i}_0.\end{aligned}$$ This can be proven as follows. As we have already indicated, the part of ${\mbox{$\mathcal D$}}^{\le i}$ which originates from the geometric part of the Rothstein bracket, which equals $-[\theta_{\le i},[\widetilde{\Pi}_{\le i},\:]]$, is of filtration degree 1. Therefore, we have for every $\alpha \in F^k{\mathscr A_{\mathcal V^{\le{i}}}}^n$ $$\begin{aligned} {\mbox{$\mathcal D$}}^{\le i}(\alpha)=\{\theta_{\le i},\alpha\}_{\le i}\in&-&\sum_{j=1}^i\sum_{a=1}^{\ell_j}\Big[\theta_{\le i},\Big[{\frac{\partial}{\partial \xi_{(j)}^{a}}}{\frac{\partial}{\partial \xi^{(j)}_{a}}},\alpha\Big]\Big]+ F^{k+1}{\mathscr A_{\mathcal V^{\le{i}}}}^{n+1}\nonumber\\ &=&\sum_{j=1}^i\sum_{a=1}^{\ell_j} \Big(\Big(J^{(j)}_a+\sum_{l=1}^i\frac{\partial Q_{(l)}}{\partial{\xi_{(j)}^{a}}}\Big)\frac{\partial \alpha}{\partial {\xi^{(j)}_{a}}}\nonumber\\ &&\qquad\qquad+(-1)^{(n+1)(i+1)}\frac{\partial \alpha}{\partial {\xi_{(j)}^{a}}}\frac{\partial \theta_{\le i}}{\partial {\xi^{(j)}_{a}}}\Big) +F^{k+1}{\mathscr A_{\mathcal V^{\le{i}}}}^{n+1}\nonumber\\ &\stackrel{()}{=}&\partial^{\le i}(\alpha)+\sum_{l=1}^i\sum_{j=1}^{l-1}\sum_{a=1}^{\ell_j}(-1)^{(n+1)(i+1)}\frac{\partial \alpha}{\partial {\xi_{(j)}^{a}}}\frac{\partial J_a^{(l)}}{\partial {\xi^{(j)}_{a}}}{\xi_{(l)}^{a}}\label{remainingterms}\\ &&+\sum_{l=1}^i\sum_{j=1}^{l}\sum_{a=1}^{\ell_j}(-1)^{(n+1)(i+1)}\frac{\partial \alpha}{\partial {\xi_{(j)}^{a}}}\frac{\partial Q_{(l)}}{\partial {\xi^{(j)}_{a}}}+F^{k+1}{\mathscr A_{\mathcal V^{\le{i}}}}^{n+1}.\nonumber\end{aligned}$$ At step $()$ we have used the fact that since $Q_{(l)}$ is at least quadratic in the momenta $\partial Q_{(l)}/\partial {\xi_{(j)}^{a}}\in F^1{\mathscr A_{\mathcal V^{\le{l}}}}^{1-j}$. We claim that the remaining terms in equation (\[remainingterms\]) are also in $F^{k+1}{\mathscr A_{\mathcal V^{\le{i}}}}^{n+1}$. In fact for $j+1\le l\le i$ we have $$\begin{aligned} \frac{\partial \alpha}{\partial {\xi_{(j)}^{a}}}\frac{\partial J_a^{(l)}}{\partial {\xi^{(j)}_{a}}}{\xi_{(l)}^{a}}\in (F^{k-j}{\mathscr A_{\mathcal V^{\le{i}}}}^{n-j})({\mathscr A_{\mathcal V^{\le{l-1}}}}^{n+1-l+j})(F^{l}{\mathscr A_{\mathcal V^{\le{l}}}}^l)\subset F^{k+1}{\mathscr A_{\mathcal V^{\le{i}}}}^{n+1}.\end{aligned}$$ Finally, for $j\le l\le i$ we have $$\begin{aligned} \frac{\partial \alpha}{\partial {\xi_{(j)}^{a}}}\frac{\partial Q_{(l)}}{\partial {\xi^{(j)}_{a}}}\in (F^{k-j}{\mathscr A_{\mathcal V^{\le{i}}}}^{n-j})(F^{l+1}{\mathscr A_{\mathcal V^{\le{l}}}}^{1+j})\subset F^{k+1}{\mathscr A_{\mathcal V^{\le{i}}}}^{n+1},\end{aligned}$$ and the proof of equation (\[filtrationisgood\]) is finished. In order to start the induction we look at the projective presentation $$\begin{aligned} \mathcal V^1=KT_1\stackrel{J=J_{(1)}}{\longrightarrow}I\to 0.\end{aligned}$$ Since the ideal $I$ is first class the term $r_1$ of lowest order in ${\frac{1}{2}}\{J_{(1)},J_{(1)}\}_{\le 1}$ vanishes when restricted to $Z$. By definition of the Koszul-Tate resolution there is a $Q_{(1)}\in KT^2_{\le 1}\otimes KT_1^{\le 1}$ such that $r_1=-\partial^{\le 1}Q_{(1)}$. Setting $\theta_{\le 1}=\theta_1=J_{(1)}+Q_{(1)}$, we obtain $$\begin{aligned} \{\theta_1,\theta_1\}_{\le 1}&=&\{J_{(1)},J_{(1)}\}_{\le 1}+2\:\{J_{(1)},Q_{(1)}\}_{\le 1}+\{Q_{(1)},Q_{(1)}\}_{\le 1}\\ &\in&\{J_{(1)},J_{(1)}\}_{\le 1}+2\partial^{\le 1}q_1+F^2{\mathscr A_{\mathcal V^{\le{1}}}}\\ &\in&2r_1+2\partial^{\le 1}Q_{(1)}+F^2{\mathscr A_{\mathcal V^{\le{1}}}}\end{aligned}$$ and we are done. Note that we may write the system of projective generators in local coordinates $J_{(1)}=\sum_{a=1}^{\ell_1}J^{(1)}_a\xi^a_{(1)}$. Since the ideal is first class, we have $\{J^{(1)}_a,J^{(1)}_b\}=\sum_{c=1}^{\ell_1} f_{ab}^c\:J^{(1)}_c$ for some (nonunique) functions $f_{ab}^c$, the so-called structure functions. Locally, we can write $Q_{(1)}$ in terms of the structure functions: $Q_{(1)}=-{\frac{1}{2}}\sum_{a,b,c=1}^{\ell_1} f_{ab}^c\: \xi_{(1)}^a\xi_{(1)}^b\xi^{(1)}_c$. Let us now assume that the Ansatz (\[Ansatz\]) fulfills equation (\[filtprop\]) for $i$. We are looking for a $Q_{(i+1)}$ such that equation (\[filtprop\]) is true for $i\to i+1$. Taking advantage of the Jacobi identity for the bracket $\{\:,\:\}_{\le i}$ $$\begin{aligned} 0=\{\theta_{\le i},\{\theta_{\le i},\theta_{\le i}\}_{\le i}\}_{\le i}={\mbox{$\mathcal D$}}^{\le i}\{\theta_{\le i},\theta_{\le i}\}_{\le i},\end{aligned}$$ we deduce $$\begin{aligned} \label{icycle} \partial^{\le i}\{\theta_{\le i},\theta_{\le i}\}_{\le i}=-\sum_{j=1}^{\infty}{\mbox{$\mathcal D$}}^{\le i}_j\{\theta_{\le i},\theta_{\le i}\}_{\le i}\qquad\in F^{i+2}{\mathscr A_{\mathcal V^{\le{i}}}}^3.\end{aligned}$$ Let $r_i\in {KT_{\le i}^{i+2}}\otimes KT^{\le i}_i$ be the term of lowest degree in ${\frac{1}{2}}\{\theta_{\le i},\theta_{\le i}\}_{\le i}$. It follows from equation (\[icycle\]) that $$\partial^{\le i} r_i=0.$$ By construction of the Koszul-Tate resolution there is a $Q_{(i+1)}\in {KT_{\le i}^{i+2}}\otimes KT^{\le i+1}_{i+1}$, such that $$r_i=-\partial^{\le i+1} Q_{(i+1)}.$$ Setting $\theta_{i+1}:=\sum_{a=1}^{\ell_{i+1}}J_a^{(i+1)}\xi^a_{(i+1)}+Q_{(i+1)}$ we have to make sure that $\{\theta_{\le i+1},\theta_{\le i+1}\}_{\le i+1}\in F^{i+2}{\mathscr A_{\mathcal V^{\le{i+1}}}}^2$. First of all, note that $$\partial^{\le i+1} \theta_{i+1}=\sum_{a=1}^{\ell_{i+1}}\big(\partial^{\le i+1}J^{(i+1)}_a\big)\:\xi_{(i+1)}^a+\partial^{\le i+1}Q_{(i+1)}=-r_i.$$ We conclude that $\partial^{\le i+1}\theta_{i+1}+{\frac{1}{2}}\{\theta_{\le i},\theta_{\le i}\}_{\le i}\in F^{i+2}{\mathscr A_{\mathcal V^{\le{i}}}}^2\subset F^{i+2}{\mathscr A_{\mathcal V^{\le{i+1}}}}^2$. It is important to note that from the very definition of the Rothstein Poisson bracket we have $$\begin{aligned} \label{rothcompare}\{\theta_{\le i},\theta_{\le i}\}_{\le i+1}-\{\theta_{\le i},\theta_{\le i}\}_{\le i}\:\in F^{i+2}{\mathscr A_{\mathcal V^{\le{i+1}}}}^2.\end{aligned}$$ More precisely, this difference does merely originate from the geometric part of the Rothstein Poisson bracket. The lowest order contributions involve the curvature of the vector bundle $V^{i+1}$ of level $i+1$ Tate generators and are generated by $\xi^a_{(i+1)}\xi_b^{(i+1)}\xi^c_{(1)}$ for $a,b=1,\dots ,\ell_{i+1}$ and $c= 1,\dots, \ell_1$. As a result we obtain $$\begin{aligned} \{\theta_{\le i+1},\theta_{\le i+1}\}_{\le i+1}&=&\{\theta_{\le i},\theta_{\le i}\}_{\le i+1}+2\:\{\theta_{\le i},\theta_{i+1}\}_{\le i+1}+\{\theta_{i+1},\theta_{i+1}\}_{\le i+1}\\ &=&\{\theta_{\le i},\theta_{\le i}\}_{\le i+1}+2\:{\mbox{$\mathcal D$}}^{\le i+1}(\theta_{i+1})-\{\theta_{i+1},\theta_{i+1}\}_{\le i+1}\\ &\in&\{\theta_{\le i},\theta_{\le i}\}_{\le i}+2\:{\mbox{$\mathcal D$}}^{\le i+1}(\theta_{i+1})-\{\theta_{i+1},\theta_{i+1}\}_{\le i+1}+F^{i+2}{\mathscr A_{\mathcal V^{\le{i+1}}}}^2\\ &\in& 2\sum_{j=1}^\infty{\mbox{$\mathcal D$}}^{\le i+1}_j(\theta_{i+1})-\{\theta_{i+1},\theta_{i+1}\}_{\le i+1}+F^{i+2}{\mathscr A_{\mathcal V^{\le{i+1}}}}^2.\end{aligned}$$ From the very definition (\[Ansatz\]) we see that $\theta_{i+1}\in F^{i+1}{\mathscr A_{\mathcal V^{\le{i+1}}}}^1$. Since $\sum_{j=1}^\infty{\mbox{$\mathcal D$}}^{\le i+1}_j$ is of filtration degree $1$ we obtain $\sum_{j=1}^\infty{\mbox{$\mathcal D$}}^{\le i+1}_j(\theta_{i+1})\in F^{i+2}{\mathscr A_{\mathcal V^{\le{i+1}}}}^2$. \ The element $\theta$ is called the BRST charge. Note that in the case of a moment map $J=J_{(1)}$ the first two terms of the series for the charge read $$\begin{aligned} \theta_1&=&\sum_{a=1}^{\ell_1} J_a\:{\xi_{(1)}^{a}}-{\frac{1}{2}}\sum_{a,b,c=1}^{\ell_1}f_{ab}^c\: {\xi_{(1)}^{a}}{\xi_{(1)}^{b}}{\xi^{(1)}_{c}},\\ \theta_2&=&\sum_{a=1}^{\ell_2} J^{(2)}_a\:{\xi_{(2)}^{a}}\:,\end{aligned}$$ where $f_{ab}^c$ are the structure constants of the Lie algebra $\mathfrak g$. In other words, the quadratic term $Q_{(2)}$ vanishes in this case. On the other hand, in the case of a projective Koszul-resolution (cf. section \[projkos\]) the $\theta_i$ for $2\le i\le \ell$ consist merely of the quadratic term (here we have of course $\theta_i=0$ for $i>\ell$). Associated to the charge there is a homological Hamiltonian vector field $$\begin{aligned} {\mbox{$\mathcal D$}}:=\{\theta,\:\}_R\end{aligned}$$ which is called the BRST differential. Although the homogeneous components of ${\mathscr A_{\mathcal V}}^\bullet$ are not direct sums, there is a unique decomposition ${\mbox{$\mathcal D$}}=\sum_{i=0}^\infty {\mbox{$\mathcal D$}}_i=\partial+\sum_{i=1}^\infty {\mbox{$\mathcal D$}}_i$ of the BRST differential, such that ${{\mbox{$\mathcal D$}}_i}\big(KT^j\otimes KT_k\big)\subset KT^{j+i}\otimes KT_{k+i-1}$: $$\begin{aligned} \label{brsdecomp} \xymatrix{\cdot&\cdot&\cdot&\cdot&\\ \cdot&\cdot&\cdot&\cdot&\\ \cdot&\cdot&\cdot&\cdot&\\ \cdot&\cdot\ar[l]^{{\mbox{$\mathcal D$}}_0}\ar[u]^{{\mbox{$\mathcal D$}}_1}\ar[uur]^{{\mbox{$\mathcal D$}}_2}\ar[uuurr]_{\dots}&\cdot&\cdot&}\end{aligned}$$ The equation ${\mbox{$\mathcal D$}}^2=0$ translates into a sequence of relations, starting with $$\begin{aligned} \partial^2&=&0\\ \partial\: {\mbox{$\mathcal D$}}_1+ {\mbox{$\mathcal D$}}_1\:\partial&=&0\\ \partial\: {\mbox{$\mathcal D$}}_2+ {\mbox{$\mathcal D$}}_2\:\partial+{\mbox{$\mathcal D$}}_1^2&=&0.\label{Dsquared2}\end{aligned}$$ The restriction map ${\operatorname{res}}:\mathcal C^\infty(M)\to \mathcal C^\infty(Z)$, which is an augmentation for the Koszul-Tate resolution extends to a degree zero map $$\begin{aligned} {\operatorname{res}}: {\mathscr A_{\mathcal V}}^\bullet\to KT^\bullet_{\|Z}.\end{aligned}$$ Here, $KT^\bullet_{\|Z}$ is the space of smooth sections of the restriction of the bundle of ghosts to $Z$ and ${\operatorname{res}}$ is defined to vanish on antighosts and acts on ghost by restricting the coefficients. By definition the smooth sections of this bundle are those which arise by restriction. Hence the restriction map ${\operatorname{res}}$ is onto. Therefore, the formula $d{\operatorname{res}}={\operatorname{res}}{\mbox{$\mathcal D$}}_1$ defines uniquely a $\mathbbm K$-linear degree 1 map $KT^\bullet_{\|Z}\to KT^{\bullet+1}_{\|Z}$. If ${\mbox{$\mathcal D$}}_1^2$ is evaluated on an element containing no antighosts, the result will be, due to equation (\[Dsquared2\]), in the kernel of ${\operatorname{res}}$. It follows that $d^2=0$. We will call the cochain complex $$(KT^\bullet_{\|Z},d)$$ the vertical complex. A priori the vertical complex depends upon choices made for the space of generators $\mathcal V$, for the connection on $\mathcal V$, for the differential $\partial$, for $\theta$ and for ${\operatorname{prol}}$. It is conjectured that the homotopy class of the vertical complex – in a sense yet to be specified – does not depend on these choices. The cochain complex $(KT^\bullet_{\|Z},d)$ is well understood for certain special cases (see section \[cois\] and section \[classred\]). It may be shown by a spectral sequence argument, that ${\operatorname{res}}$ is in fact a quasiisomorphism of complexes (see also Figure 3.1). \[quasiisom\] ![The map ${\operatorname{res}}$ is a continuous quasiisomorphism from the BRST complex to the vertical complex, which has a continuous split. Hence the zeroth cohomology algebra of the BRST complex is toplogically isomorphic to the space of invariant functions $\mathcal C^\infty(Z)^I$, i.e., the Dirac reduced algebra of the first class constraint. We emphasize that the BRST complex is, in general, not a double complex.](BRSTcontr.eps) In fact, if we are given a continuous prolongation map ${\operatorname{prol}}:\mathcal C^\infty(Z)\to \mathcal C^\infty(M)$ and continuous contracting homotopies $h_i:KT_i\to KT_{i+1}$ for the Koszul-Tate resolution we can be more specific. Note that these maps naturally extend to maps ${\operatorname{prol}}:KT^\bullet_{\|Z}\to KT^\bullet$ and $h_i:KT^\bullet\otimes KT_i\to KT^\bullet\otimes KT_{i+1}$. Assume that the premises for Theorem \[chargeexists\] are true and, in addition, we have a contraction (\[KTContr\]) as in Proposition \[existsfiniteTate\]. Then there are continuous $\mathbbm K$-linear maps $\Phi:KT_{\|Z}^\bullet\to {\mathscr A_{\mathcal V}}^\bullet$ and $H:{\mathscr A_{\mathcal V}}^{\bullet}\to {\mathscr A_{\mathcal V}}^{\bullet +1}$ such that $$\begin{aligned} \label{KTbrstcontr} \big(KT^\bullet_{\|Z},d\big)\: \begin{array}{c} {{\operatorname{res}}}\\\leftrightarrows\\ {\Phi} \end{array} \:({\mathscr A_{\mathcal V}}^\bullet,{\mbox{$\mathcal D$}}),H\end{aligned}$$ is a contraction. If $[a],[b]\in \mathrm H\: KT_{\|Z}$ are the cohomology classes of the cocycles $a,b \in\mathrm Z\: KT_{\|Z}$, then $$\begin{aligned} \label{HPTredpoiss} \{[a],[b]\}:=[{\operatorname{res}}\{\Phi(a),\Phi(b)\}_R]\end{aligned}$$ defines a $\mathbbm Z$-graded Poisson algebra structure on $\mathrm H^\bullet KT_{\|Z}$. In degree zero $\mathrm H^0KT_{\|Z}=\mathcal C^\infty(M)^I$ this Poisson structure coincides with the Dirac reduced Poisson structure. We apply perturbation lemma \[BPL1\] to the contraction (\[KTContr\]) and obtain the contraction (\[KTbrstcontr\]). It is a straightforward matter to check that the bracket is well defined. The Jacobi identity follows immediatly from the computation $$\begin{aligned} \{[a],\{[b],[c]\}\}&=&\Big[{\operatorname{res}}\big\{\Phi(a),\Phi\big({\operatorname{res}}\{\Phi(b),\Phi(c)\}_R\big)\big\}_R\Big]\\ &=&\Big[{\operatorname{res}}\big\{\Phi(a),\{\Phi(b),\Phi(c)\}_R\big\}_R\Big]-\Big[{\operatorname{res}}\big\{\Phi(a),{\mbox{$\mathcal D$}}H\{\Phi(b),\Phi(c)\}_R\big\}_R\Big]\\ &&-\Big[{\operatorname{res}}\big\{\Phi(a),H\{\Phi(db),\Phi(c)\}_R+H\{\Phi(b),\Phi(dc)\}_R\big\}_R\Big]\\ &=&\Big[{\operatorname{res}}\big\{\Phi(a),\{\Phi(b),\Phi(c)\}_R\big\}_R\Big]-\Big[d {\operatorname{res}}\big\{\Phi(a), H\{\Phi(b),\Phi(c)\}_R\big\}_R\Big]\\ &=&\Big[{\operatorname{res}}\big\{\Phi(a),\{\Phi(b),\Phi(c)\}_R\big\}_R\Big]\\\end{aligned}$$ for cocycles $a,b,c\in \mathrm Z\:KT_{\|Z}$. The Leibniz rule is a consequence of the following consideration. Given two cocycles $b,c \in \mathrm Z\: KT_{\|Z}$ the difference $\Delta:=\Phi(bc)-\Phi(b)\Phi(c)$ is, in general, nonzero since $\Phi$ (as well as ${\operatorname{prol}}$) is not multiplicative. Nonetheless, because ${\operatorname{res}}$ is multiplicative and ${\operatorname{res}}\:\Phi={\operatorname{id}}$, we have $\Delta\in {\operatorname{ker}}({\operatorname{res}})$. On the other hand, it is closed, since $\Phi$ is a chain map. It follows that $\Delta= ({\mbox{$\mathcal D$}}H+ H {\mbox{$\mathcal D$}}+\Phi{\operatorname{res}})\Delta={\mbox{$\mathcal D$}}(H( \Delta))$ is a coboundary, and consequently $$\begin{aligned} \{[a],[b][c]\}=\big[{\operatorname{res}}\{\Phi(a),\Phi(bc)\}_R\big]=\big[{\operatorname{res}}\{\Phi(a),\Phi(b)\Phi(c)\}_R\big]+\underbrace{\big[d{\operatorname{res}}\{\Phi(a),H(\Delta\}_R\big]}_{=0}.\end{aligned}$$ Since in degree zero the differential $d:\mathcal C^\infty(Z)\to KT_{\|Z}^1$ is given by the formula $d(a)={\operatorname{res}}\{J_{(1)},{\operatorname{prol}}(a)\}_R$, it is clear that $\mathrm H^0 KT_{\|Z}$ is the space of $I$ invariant smooth functions on $Z$. In order to see that the induced bracket on $\mathrm H^0 KT_{\|Z}$ is the Dirac reduced bracket, note that the image $\Phi(a)\subset {\operatorname{prol}}(a)+F^1{\mathscr A_{\mathcal V}}^0$ of a function $a\in\mathcal C^\infty(Z)$ starts with ${\operatorname{prol}}(a)$. Since $\{F^1{\mathscr A_{\mathcal V}}^0,F^1{\mathscr A_{\mathcal V}}^0\}_R\subset F^1{\mathscr A_{\mathcal V}}^0\subset{\operatorname{ker}}({\operatorname{res}})$ and $\{\mathcal C^\infty(M),F^1{\mathscr A_{\mathcal V}}^0\}_R\subset F^1{\mathscr A_{\mathcal V}}^0$, the higher order terms do not contribute, and the reduced bracket of two invariant functions $a,b\in \mathcal C^\infty(Z)^I$ is given by the formula $\{a,b\}={\operatorname{res}}\{{\operatorname{prol}}a,{\operatorname{prol}}b\}$. \ We would like to stress that the algebraic properties of the contraction (\[KTbrstcontr\]) which enable us to transfer the Poisson structure are: 1.) the right hand side is a differential graded Poisson algebra and 2.) ${\operatorname{res}}$ is a map of graded commutative agebras. Coisotropic submanifolds {#cois} ------------------------ First class constraint sets, which are closed submanifolds are, per definition, coisotropic submanifolds. It can be easily seen that a closed submanifold $C$ of a Poisson manifold $(M,\Pi)$ is coisotropic iff $$\begin{aligned} \label{coiscond} \Pi(\alpha,\beta)(c)=0 \quad \forall c\in C, \quad\forall \alpha,\beta\in TC^{{\operatorname{ann}}}.\end{aligned}$$ Another equivalent characterization of coisotropy of the closed submanifold $C$ is that $$\begin{aligned} \label{anchor} \#_\Pi(TC^{{\operatorname{ann}}})\subset TC,\end{aligned}$$ where $\#_\Pi$ is the restriction to $C$ of the anchor map $\#_\Pi:T^M \to TM$, $\#_\Pi(\alpha)=i(\alpha)\Pi$ of the Lie algebroid associated to $\Pi$. It is well-known that $\#_\Pi$ in equation (\[anchor\]) is an anchor map of a Lie algebroid over $C$. The bracket on $TC^{{\operatorname{ann}}}$ (cf. Lemma \[conormallemma\]) is given by the formula $[\mathrm df,\mathrm dg]_{\|c}:=\mathrm d\{f,g\}_{\|c}$, where $c\in C$ and $f,g\in \mathcal C^\infty(M)$ are functions vanishing on $C$. Coisotropic submanifolds arise naturally in many geometric situations. For example, the graph of a Poisson map $\varphi:M\to N$ is a coisotropic submanifold in $M\times\bar{N}$, where $\bar{N}$ means $N$ understood with the opposite Poisson structure. Furthermore, if $J: M\to \mathfrak g^$ is the moment map of a Hamiltonian $G$-action and if $J$ intersects the coadjoint orbit $\mathcal O\subset \mathfrak g^$ cleanly, then $J^{-1}(\mathcal O)\subset M$ is a coisotropic submanifold. Given a coisotropic submanifold $C\subset M$ there is an important subspace ${\mathfrak X}(C,M)$ of the Gerstenhaber algebra of polyvector fields ${\mathfrak X}(M)$, which is defined as follows: $$\begin{aligned} {\mathfrak X}^0(C,M)&:=&I_C=\{f\in\mathcal C^\infty(M)\mid f_{\|C}=0\}, \mbox{ and}\\ {\mathfrak X}^k(C,M)&:=&\{\:X\in {\mathfrak X}^k(M)\mid X_c(\alpha_1,\dots,\alpha_k)=0 \:\:\forall c\in C\mbox{ and }\alpha_1,\dots,\alpha_k\in T_cC^{{{\operatorname{ann}}}}\:\}\end{aligned}$$ for $k\ge 1$. Moreover, we consider the canonical map $\Psi:\Gamma^\infty(M,TM)\to \Gamma^\infty(C,TM_{\|C}/TC)$, which restricts a vector field to $C$ and takes the residue class of the result modulo $TC$. After identifying $TM_{\|C}/TC$ with the dual of the annihilator bundle we obtain a map $$\wedge ^k\Psi:{\mathfrak X}^k(M)\to\Gamma^\infty(C,\wedge^k{TC^{{\operatorname{ann}}}}^).$$ The latter space is just the space of cochains of Lie algebroid cohomology of $TC^{{\operatorname{ann}}}$. In fact, it follows from the next proposition that ${\mathfrak X}(C,M)$ is a coisotropic ideal in the differential Gerstenhaber algebra $({\mathfrak X}^\bullet(M),\wedge,[\:,\:],\delta_\Pi)$, in the sense of the definition which was given in subsection \[derbrsection\]. It is a curious fact that in this picture the Poisson tensor is a ‘first class constraint’ by itself. ${\mathfrak X}(C,M)$ has the following properties: 1. \[wedgeideal\]${\mathfrak X}(C,M)^i\wedge {\mathfrak X}^j(M)\subset{\mathfrak X}^{i+j}(C,M)$, 2. \[firstclassprop\]$[{\mathfrak X}^i(C,M),{\mathfrak X}^j(C,M)]\subset{\mathfrak X}^{i+j-1}(C,M)$, 3. \[subcomplex\]$\delta_\Pi{\mathfrak X}^i(C,M)\subset{\mathfrak X}^{i+1}(C,M)$ for all $i,j\ge0$. Moreover, $\wedge ^k\Psi$ is onto and the kernel of $\wedge ^k \Psi$ is just ${\mathfrak X}^k(C,M)$ for all $k\ge 0$. Thus, we can identify the quotient space ${\mathfrak X}^k(M)/{\mathfrak X}^k(C,M)$ with $\Gamma^\infty(C,\wedge^k{TC^{{\operatorname{ann}}}}^)$ for all $k\ge 0$. The induced differential on $\Gamma^\infty(C,\wedge^k{TC^{{\operatorname{ann}}}}^)$ coincides with the differential of Lie algebroid cohomology. \[wedgeideal\].) and \[firstclassprop\].) follow from the formulas (\[cupform\]), (\[bullform\]) and Theorem \[multiderisgerst\] since for the Lie Rinehart pair $\left(\mathcal C^\infty(M),\Gamma^\infty(M,TM)\right)$ the Gerstenhaber algebras of polyvector fields and multiderivations coincide. Note that for \[firstclassprop\].) one also makes use of Lemma \[conormallemma\]. Statement \[subcomplex\].) is a consequence of \[firstclassprop\].). The claim that $\wedge^k\Psi$ is onto can easily be proven by a partition of unity argument. The last statement follows straightforwardly from the fact that the Koszul-Brylinsky bracket of two exact forms $\mathrm df$ and $\mathrm dg$ is given by $\mathrm d\{f,g\}$. \[derbrPoissonstructure\] There is a natural $\mathbbm Z$-graded super-Poisson structure on the cohomology of the Lie algebroid $TC^{{\operatorname{ann}}}$. The induced Poisson strucure in degree zero coincides with the Dirac reduced bracket. Use the derived bracket of Theorem \[redderbr\]. \ Alternatively, we could use a projective Koszul resolution (\[prKosContr\]) of $\mathcal C^\infty(C)$ for some tubular neighborhood $U$ of $C$ in $M$ and the construction of Section \[BRSTchargesection\] to aquire the cohomology of the vertical complex $KT_{\|C}$ with a $\mathbbm Z$-graded super-Poisson structure according to formula (\[HPTredpoiss\]). It is not difficult to prove that the vertical complex coincides with the cochain complex of Lie algebroid cohomology of the Lie algebroid $TC^{{{\operatorname{ann}}}}$ as above. However, it is not clear to the author whether the $\mathbbm Z$-graded Poisson structure of formula (\[HPTredpoiss\]) coincides with that of Corollary \[derbrPoissonstructure\]. Moreover, these Poisson structures still have to be compared with the $P_\infty$-algebra structure on the vertical complex which has been introduced in [@relform]. Classical BRST-algebra for Hamiltonian group actions {#classred} ---------------------------------------------------- In the remainder of this work we will exclusively be concerned with the following important, and most simple, special case of the BFV-construction. We will consider a Hamiltonian $G$-space $M$, $G$ compact and connected, with a moment map $J:M\to \mathfrak g^$ for which the Koszul complex $$\begin{aligned} K_{-\bullet}=K_{-\bullet}(C^\infty(M),J)=S^\bullet_{C^\infty(M)}\big(\mathfrak g[1]\big),\qquad\partial=\sum_{a=1}^\ell J_a{\frac{\partial }{\partial \xi_{a}}},\end{aligned}$$ is a free resolution of the $A:=\mathcal C^\infty(M)$-module $\mathcal C^\infty(Z)$ of smooth functions on the zero fibre $Z=J^{-1}(0)$. Here $\xi_1,\dots,\xi_\ell$ denote a basis for $\mathfrak g[1]$ and ${\frac{\partial }{\partial \xi_{1}}},\dots,{\frac{\partial }{\partial \xi_{\ell}}}$ the corresponding algebraic vector fields on the graded manifold $S^\bullet_{A}\big(\mathfrak g[1]\big)$. We have readily seen in Subsection \[ghsubsection\] and Section \[kc\] that the Examples \[mpart\], \[lemon\], \[stratres\], \[tzwo\] for $\alpha<0$ and \[commvar\] provide in fact examples of such moment maps. Since we do not need to mention the level here, we will simply call the elements of $\mathfrak g^[-1]$ and $\mathfrak g[1]$ ghosts and antighosts, respectively. Dually to the basis $\xi_1,\dots ,\xi_\ell$ for $\mathfrak g[1]$, we also will need a basis $\xi^1,\dots ,\xi^\ell$ for $\mathfrak g^[-1]$. The respective indices will run over latin letters: $a,b,\dots$. Since all ghost variables are purely odd, we will identify the BRST algebra ${\mathscr A }:={\mathscr A }_{\mathcal V}$, where $\mathcal V=\Gamma^\infty\left(M,\mathfrak g[-1]\times M\right)$ is the space of sections of the trivial vector bundle with fibre $\mathfrak g[-1]$, with the space of polynomials in the ghosts and antighosts: $$\begin{aligned} {\mathscr A }=S_A\left(\mathfrak g[1]\oplus \mathfrak g^[-1]\right).\end{aligned}$$ Alternatively, we could identify ${\mathscr A }$ with the space of sections of the trivial vector bundle over $M$ with fibre $\wedge(\mathfrak g\oplus \mathfrak g^)=\wedge\mathfrak g^\otimes \wedge\mathfrak g$ understood with the appropriate grading. There is an even graded Poisson bracket on $S_{\mathbbm K}\left(\mathfrak g[1]\oplus \mathfrak g^[-1]\right)$, which is defined by the formula $$\begin{aligned} \label{oddpoiss} \{v,w\}=-2\mu\circ \left(\sum_{a=1}^\ell{\frac{\partial }{\partial \xi_{a}}}\otimes {\frac{\partial }{\partial \xi^{a}}}+{\frac{\partial }{\partial \xi^{a}}}\otimes{\frac{\partial }{\partial \xi_{a}}}\right)(v\otimes w).\end{aligned}$$ Here, $\mu$ denotes the super-commutative multiplication in $S_{\mathbbm K}\left(\mathfrak g[1]\oplus \mathfrak g^[-1]\right)$ and $\otimes$ is the graded tensor product. This Poisson bracket is the unique even super-Poisson bracket such that $\{\xi_a,\xi^b\}=2\delta_a^b$ and $\{\xi_a,\xi_b\}=0=\{\xi^a,\xi^b\}$ for all $a,b=1,\dots,\ell$. Note the slight change of convention: the above bracket differs from Rothstein bracket by a factor of $2$! If $f,g\in\mathcal C^\infty(M)$ and $v,w\in S_{\mathbbm K}\left(\mathfrak g[1]\oplus \mathfrak g^[-1]\right)$, then the formula $$\begin{aligned} \{fv,gw\}:=\{f,g\}vw+fg\{v,w\}\end{aligned}$$ defines a super-Poisson bracket on ${\mathscr A }$. From the Lie bracket of $\mathfrak g$ and the moment map we build the so-called BRST-charge* $$\begin{aligned} \theta:= -\frac{1}{4}\sum_{a,b,c}\:f_{ab}^c\:\xi^a\xi^b\xi_c+\sum_a J_a \xi^a\in\mathscr A^1,\end{aligned}$$ where the $f_{ab}^c$ are the structure constants of $\mathfrak g$. An easy calculation yields $\{\theta,\theta\}=0$. In other words, the recursion of Theorem \[chargeexists\] breaks off after one step here. Hence, the classical BRST-differential $${\mbox{$\mathcal D$}}:=\{\theta,\:\}$$ is in fact of square zero. Summing up, we obtain a differential graded Poisson algebra $$({\mathscr A },\{,\},{\mbox{$\mathcal D$}}=\{\theta,\:\}),$$ which is called the classical BRST algebra henceforth. Closer examination shows that $({\mathscr A },{\mbox{$\mathcal D$}})$ is the total complex of a certain double complex. In fact, there is a canonical identification ${\mathscr A }=S_A(\mathfrak g^[-1])\otimes S_A(\mathfrak g[1])=C^\bullet\left(\mathfrak g,S_A(\mathfrak g[1])\right)$, where the latter denotes the space of Lie algebra cochains with values in the $\mathfrak g$-module $S_A\mathfrak g[1]$. This representation will be denoted henceforth by $\mathrm L$. In this way ${\mathscr A }=\oplus_{i,j=0}^\ell{\mathscr A }^{i,j}$ acquires a $\mathbbm Z\times\mathbbm Z$ -grading $${\mathscr A }^{i,j}:=S^i_A(\mathfrak g^[-1])\otimes S^j_A(\mathfrak g[1])$$ The differential of the Lie algebra cohomology $\delta:{\mathscr A }^{i,j}\to{\mathscr A }^{i+1,j} $ corresponding to the $\mathfrak g$-module $S_{A}(\mathfrak g[1])$ can be writen as a super-differential operator $$\begin{aligned} \delta=-\frac{1}{2}\sum_{a,b,c}f_{ab}^c\:\:\xi^a\xi^b\:{\frac{\partial }{\partial \xi_{c}}} +\sum_{a,b,c}f_{ab}^c\:\xi^a\xi_c\:{\frac{\partial }{\partial \xi^{b}}}+\sum_a\:\xi^a\{J_a,\:\}.\end{aligned}$$ The Koszul-differential $\partial=\sum_{a=1}^\ell J_a{\frac{\partial }{\partial \xi^{a}}}$ extends naturally to a differential $\partial:{\mathscr A }^{i,j}\to{\mathscr A }^{i,j-1}$. Because $J$ is equivariant, these two differentials super-commute: $\partial \delta+\delta\partial=0$. An easy calculation yields that $$\begin{aligned} {\mbox{$\mathcal D$}}=2\partial+\delta.\end{aligned}$$ We will view ${\mbox{$\mathcal D$}}$ as a perturbation (see Appendix \[HPT\]) of the acyclic differential $2\partial$. We extend the restriction map ${\operatorname{res}}$ to a map ${\operatorname{res}}:\mathscr A\to S_{\mathcal C^\infty(Z)}(\mathfrak g^[-1])$ by setting it zero for all terms containing antighosts and restricting the coefficients. In the same fashion, we extend ${\operatorname{prol}}$ to a map $S_{\mathcal C^\infty(Z)}(\mathfrak g^[-1])\to\mathscr A$ extending the coefficients. Since the moment map $J$ is $G$-equivariant, $G$ acts on $Z=J^{-1}(0)$. Hence $\mathcal C^\infty(Z)$ is a $\mathfrak g$-module, this representation will be denoted by $\mathrm{L}^z$. Note that $\mathrm{L}^z_X={\operatorname{res}}\;\mathrm L_X\;{\operatorname{prol}}$ for all $X\in\mathfrak g$. We identify $S_{\mathcal C^\infty(Z)}(\mathfrak g^[-1])$ with the space of cochains of Lie algebra cohomology $C^\bullet\big(\mathfrak g,\mathcal C^\infty(Z)\big)$. Let us denote $d:C^\bullet\big(\mathfrak g,\mathcal C^\infty(Z)\big)\to C^{\bullet+1} \big(\mathfrak g,\mathcal C^\infty(Z)\big)$ the codifferential of Lie algebra cohomology coresponding to $\mathrm{L}^z$. Since ${\operatorname{res}}$ is a morphism of $\mathfrak g$-modules we obtain $d\;{\operatorname{res}}={\operatorname{res}}\;\delta$. There are $\mathbb K$-linear maps $\Phi:C^\bullet\big(\mathfrak g, \mathcal C^{\infty}(Z)\big)\to \mathscr A^\bullet$ and $H: \mathscr{A}^\bullet\to \mathscr{A}^{\bullet-1}$ which are continuous in the respective Fréchet topologies such that $$\begin{aligned} \label{cbrstSDR} \Big(C^\bullet\big(\mathfrak g, \mathcal C^{\infty}(Z)\big),d\Big)\: \begin{array}{c} {{\operatorname{res}}}\\\leftrightarrows\\ {\Phi} \end{array} \:(\mathscr A^\bullet,{\mbox{$\mathcal D$}}),H\end{aligned}$$ is a contraction. Apply lemma \[BPL1\] to the perturbation ${\mbox{$\mathcal D$}}$ of $2\partial$. Explicitly, we get $$\begin{aligned} H=\frac{1} {2} h\sum_{j=0}^{\ell}\Big(-\frac{1}{2}\Big)^j(h\delta+\delta h)^j,\\ \Phi={\operatorname{prol}}-H(\delta\:{\operatorname{prol}}-{\operatorname{prol}}\:d),\end{aligned}$$ which are obviously Fréchet continuous. Note that from $h{\operatorname{prol}}=0$ and $h^2=0$ it follows that $H\Phi=0$ and $H^2=0$. If ${\operatorname{prol}}$ is chosen to be equivariant, then the expression for $\Phi$ simplifies to $\Phi={\operatorname{prol}}$. In the same way one gets $H=\frac{1}{2}h$, if $h$ is equivariant. There is a graded Poisson structure on $\mathrm H^\bullet\big(\mathfrak g,\mathcal C^\infty(Z)\big)$. If $[a],[b]$ are the cohomology classes of $a,b\in C^\bullet\big(\mathfrak g, \mathcal C^{\infty}(Z)\big)$, then the bracket is given by $$\begin{aligned} \{[a],[b]\}:=[{\operatorname{res}}\{\Phi(a),\Phi(b)\}].\end{aligned}$$ The restriction of this bracket to $\mathrm H^0\big(\mathfrak g,\mathcal C^\infty(Z)\big)=\mathcal C^\infty (Z)^{\mathfrak g}$ coincides with the Dirac reduced Poisson structure. Quantum BRST reduction {#qbrst} ====================== In this chapter we will construct a deformation quantization of the classical BRST algebra for the situation, when the moment map satisfies the generating and complete intersection hypothesis, cf. section \[classred\]. In order to define the quantum BRST algebra it is sufficient to assume that there is some quantum moment map, which deforms the original moment map. It has been observed in [@BHW], that it is most convenient to use for the ghost variables the so-called standard ordered Clifford multiplication instead of the Weyl-ordered multiplication, since this renders the quantum BRST-complex a double complex. We will see that the quantum Koszul differential can also be found using standard homological algebra. This entails that the quantum BRST-algebra can essentially be viewed as an Ext-algebra. We are able to compute the BRST-cohomology if 1.) the star product is assumed to be strongly invariant or 2.) the group is compact and semisimple. In both cases we find deformation quantizations for the classical reduced algebra. In these cases continuous star products, which deform the reduced Poisson algebra, can be found. The quantum BRST algebra {#brstalgconstr} ------------------------ In this section we will introduce the quantum BRST algebra, which is $\mathbbm K[[\nu]]$-differential graded associative algebra $(\mathscr{A}^\bullet[[\nu]],,{\mbox{\boldmath $\mathcal D$}})$ deforming the differential graded Poisson algebra $(\mathscr A^\bullet,\{,\},\mathscr D)$. In order to define a graded product $$ on $\mathscr A[[\nu]]$, we use on the one hand a formal Clifford multiplication $$\begin{aligned} v\cdot w&:=&\mu\big(\operatorname{e}^{-2\nu\: \sum_a{\frac{\partial }{\partial \xi^{a}}}\otimes {\frac{\partial }{\partial \xi_{a}}}}(v\otimes w)\big)\nonumber\\ &=&\sum_{i=0}^\infty \frac{(-2\nu)^i}{i!}\sum_{a_1,\dots,a_i}(-1)^{i\|x\|}\frac{\partial ^i v}{\partial \xi^{a_1}\dots\partial \xi^{a_i}}\frac{\partial ^i w}{\partial \xi_{a_1}\dots\partial \xi_{a_i}}\label{clmult}\end{aligned}$$ for homogeneous $v,w\in S_{\mathbbm K}(\mathfrak g[1]\oplus \mathfrak g^[-1])$. Here $\mu$ denotes the super-commutative multiplication and $\otimes$ is the graded tensor product. The product $\cdot$ satisfies the Clifford relation $\xi_a\cdot \xi^b+\xi^b\cdot\xi_a=2\nu \delta_a^b$ for all $a,b=1\dots \ell$. There are of course other (equivalent) ways to define the Clifford multiplication, e.g. by symmetrisation. The above product, which is sometimes called the antistandard ordered product, has the advantage to render the quantum BRST-complex a double complex (see Theorem \[doppelcompl\]). On the other hand, we will need a quantum covariant star product $\star$ on $M$ with quantum moment map ${\mbox{\boldmath{$J$}}}$ (cf. Subsection \[qmom\]). Here the quantum moment map will be viewed as an element of the BRST algebra: ${\mbox{\boldmath{$J$}}}=J+\sum_{i\ge1}\nu^i J_i \in \mathfrak g^[-1]\otimes \mathcal C^\infty(M){[[\nu]]}\subset \mathscr A^1{[[\nu]]}$. We will frequently refer to a basis $\xi_1,\dots,\xi_\ell$ of $\mathfrak g[1]$ and write for short $<{\mbox{\boldmath{$J$}}},\xi_a>={\mbox{\boldmath{$J$}}}_a$ for $a=1,\dots,\ell$. The representation property (\[qmomprop\]) can be rewritten as $$\begin{aligned} \label{qmomentmap} {\mbox{\boldmath{$J$}}}_a\star {\mbox{\boldmath{$J$}}}_b-{\mbox{\boldmath{$J$}}}_b\star {\mbox{\boldmath{$J$}}}_a=\nu\sum_c f_{ab}^c\:{\mbox{\boldmath{$J$}}}_c\qquad \mbox{ for }a,b=1,\dots,\ell,\end{aligned}$$ where $f_{ab}^c$ are the structure constants of the Lie algebra $\mathfrak g$. For $f,g\in \mathcal C^{\infty}(M)$ and $v,w \in S\big(\mathfrak g[1]\oplus \mathfrak g^[-1]\big)$ we define $$(fv)(gw):=(f\star g)\: (v\cdot w).$$ Note, that $$ defines a $\mathbbm Z$-graded associative product, which is a formal deformation of the super-Poisson structure of Section \[classred\]. The next step is to quantize the BRST-charge. It was observed by Kostant and Sternberg [@KostStern] that $${\mbox{\boldmath $\theta$}}:= -\frac{1}{4}\sum_{a,b,c}\:f_{ab}^c\:\xi^a\xi^b\xi_c+ \mathbbm \sum_a{\mbox{\boldmath{$J$}}}_a\:\xi^a+\frac{\nu}{2}\sum_{a,b} f^b_{ab}\:\xi^a \:\in\mathscr A^1[[\nu]],$$ is a good guess. Here $f_{ab}^c$ denote the structure constants of the Lie algebra. Note, that the trace term is a side effect of the operator ordering. It may be absorbed by redefining the quantum moment map ${\mbox{\boldmath{$J$}}}':={\mbox{\boldmath{$J$}}}+ \frac{\nu}{2}\sum_{a,b} f^b_{ab}\:\xi^a$. In fact, since the trace of the adjoint representation vanishes on commutators, this modification does not spoil the representation property (\[qmomentmap\]). To start with ${\mbox{\boldmath{$J$}}}'$ from scratch is considered as slightly incorrect from the point of view of representation theory of deformed algebras. Nevertheless, we will sometimes do it and will indicate that by the prime. \[qsqzero\] ${\mbox{\boldmath $\theta$}}{\mbox{\boldmath $\theta$}}=0$. Let us write for short ${\mbox{\boldmath $\theta$}}=Q+{\mbox{\boldmath{$J$}}}'$, where $Q =-\frac{1}{4}\sum_{a,b,c}f_{ab}^c\:\xi_c\xi^a\xi^b$ and ${\mbox{\boldmath{$J$}}}'=\sum_a{\mbox{\boldmath{$J$}}}_a\xi^a+{\frac{1}{2}}\sum_{a,b}f_{ab}^b\xi^a$. We have $$\begin{aligned} QQ=Q^2+2\nu\:\frac{1}{4^2}\:2 \sum_{a_1,b_1,c_1,a_2,b_2,c_2}f_{a_1b_1}^{c_1}\:f_{c_1b_2}^{c_2}\:\xi_{c_2}\xi^{a_1}\xi^{b_1}\xi^{b_2}=0,\end{aligned}$$ since $Q$ is odd and the bracket satisfies the Jacobi identity. The terms of higher order in $\nu$ vanish due to degree reasons. Furthermore, we have $$\begin{aligned} {\mbox{\boldmath{$J$}}}'{\mbox{\boldmath{$J$}}}'=\sum_{a,b}{\mbox{\boldmath{$J$}}}'_a\star{\mbox{\boldmath{$J$}}}'_b\:\xi^a\xi^b={\frac{1}{2}}\sum_{a,b}[{\mbox{\boldmath{$J$}}}'_a,{\mbox{\boldmath{$J$}}}'_b]_{\star}\:\xi^a\xi^b=\frac{\nu}{2}\sum_{a,b}f_{ab}^c\:{\mbox{\boldmath{$J$}}}'_c\:\xi^a\xi^b.\end{aligned}$$ Hence, it remains to compute $$\begin{aligned} Q{\mbox{\boldmath{$J$}}}'+{\mbox{\boldmath{$J$}}}'Q=Q{\mbox{\boldmath{$J$}}}'- 2\nu \frac{1}{4}\sum_{a,b,c} f_{ab}^c\:{\mbox{\boldmath{$J$}}}'_c \:\xi^a\xi^b+{\mbox{\boldmath{$J$}}}'Q=-\frac{\nu}{2}\sum_{a,b}f_{ab}^c\:{\mbox{\boldmath{$J$}}}'_c\:\xi^a\xi^b.\end{aligned}$$ It follows that ${\mbox{\boldmath $\theta$}}{\mbox{\boldmath $\theta$}}=QQ+Q{\mbox{\boldmath{$J$}}}'+{\mbox{\boldmath{$J$}}}'Q+{\mbox{\boldmath{$J$}}}'{\mbox{\boldmath{$J$}}}'=0$. \ Now we are ready to define the quantum BRST differential to be $${\mbox{\boldmath $\mathcal D$}}:=\frac{1}{\nu}{\operatorname{ad}}_({\mbox{\boldmath $\theta$}}).$$ Before we take a closer look at ${\mbox{\boldmath $\mathcal D$}}$, let us introduce some terminology. We define the super-differential operators ${\mbox{\boldmath$\delta$}},\mathscr R,q,u:\mathscr A^\bullet\to\mathscr A^{\bullet+1},$ $$\begin{aligned} {\mbox{\boldmath$\delta$}}(fv)&:=&-\frac{1}{2}\sum_{a,b,c}f_{ab}^c\:f\:\xi^a\xi^b\:{\frac{\partial }{\partial \xi_{c}}} v+\sum_{a,b,c}f_{ab}^c\:f\:\xi^a\xi_c\:{\frac{\partial }{\partial \xi_{b}}}v+\sum_a\:\frac{ 1}{\nu}[{\mbox{\boldmath{$J$}}}_a,f]_\:\xi^av, \\ \mathscr R(fv)&:=&\sum_a f * {\mbox{\boldmath{$J$}}}_a\:{\frac{\partial }{\partial \xi^{a}}}v,\quad\quad\quad\quad\quad\quad\quad\mbox{``right multiplication''}\\ q(fv)&:=&-\frac{1}{2}\sum_{a,b,c}f\:f_{ab}^c\:\xi_c\:\frac{\partial^2 v}{\partial\xi^a\partial\xi^b},\quad\quad\quad\quad\quad\quad\mbox {``quadratic ...''} \\ u(fv)&:=&\sum_{a,b}f\:f_{ab}^b\:{\frac{\partial }{\partial \xi^{a}}}v,\quad\quad\quad\quad\quad\quad \quad\quad\mbox{``unimodular term''}\end{aligned}$$ for $f\in \mathcal C^\infty(M)$ and $v\in S_{\mathbbm K}(\mathfrak g[1]\oplus\mathfrak g^[-1])$. Note that ${\mbox{\boldmath$\delta$}}$ is the coboundary operator of Lie algebra cohomology corresponding to the representation $$\begin{aligned} {\mbox{\boldmath{$\operatorname{L}$}}}_X:S_{\mathcal C^\infty(M)}(\mathfrak g[1])[[\nu]]&\to& S_{\mathcal C^\infty(M)}(\mathfrak g[1])[[\nu]],\nonumber \\ fv&\mapsto& f({\operatorname{ad}}_X(v)) +\nu^{-1}({\mbox{\boldmath{$J$}}}(X)\star f- f\star {\mbox{\boldmath{$J$}}}(X))v,\end{aligned}$$ where $X\in \mathfrak g$, $v\in S_{\mathbbm K}(\mathfrak g[1])$ and $f\in \mathcal C^\infty(M)[[\nu]]$. Finally, we set $$\begin{aligned} \label{defkos} {\mbox{\boldmath${\partial}$}}:=\mathscr R+\nu\Big(\frac{1}{2}u-q\Big).\end{aligned}$$ This operator will be called the deformed* or quantum Koszul differential. Clearly, adding to ${\mbox{\boldmath{$J$}}}$ a scalar multiple of the trace form does only have an effect on ${\mbox{\boldmath${\partial}$}}$, the Lie algebra differential ${\mbox{\boldmath$\delta$}}$ stays unchanged. If we would have started with ${\mbox{\boldmath{$J$}}}'$, then the unimodular term would not occur in formula (\[defkos\]). It will become clear later, that the unimodular term does not have an effect on the homology of ${\mbox{\boldmath${\partial}$}}$. \[doppelcompl\]The quantum BRST differential $$\begin{aligned} \label{total} {\mbox{\boldmath $\mathcal D$}}={\mbox{\boldmath$\delta$}}+2{\mbox{\boldmath${\partial}$}}\end{aligned}$$ is a linear combination of two super-commuting differentials ${\mbox{\boldmath$\delta$}}$ and ${\mbox{\boldmath${\partial}$}}$. With the shorthand notation as in the proof of Theorem \[qsqzero\] we compute $$\begin{aligned} [{\mbox{\boldmath{$J$}}}',fv]&=&{\mbox{\boldmath{$J$}}}'fv-(-1)^{\|v\|}fv{\mbox{\boldmath{$J$}}}'\\ &=&\sum_a ({\mbox{\boldmath{$J$}}}'_a\star f-f\star{\mbox{\boldmath{$J$}}}'_a)\xi^av+2\nu\sum_a(f\star {\mbox{\boldmath{$J$}}}'_a){\frac{\partial }{\partial \xi^{a}}}v,\end{aligned}$$ where $f\in \mathcal C^\infty(M)$ and $v\in S_{\mathbbm K}(\mathfrak g[1]\oplus\mathfrak g^[-1])$. One also has to compute the Clifford commutator with the element $Q=-\frac{1}{4}\sum_{a,b,c}f_{ab}^c\:\xi_c\xi^a\xi^b$ with a homogeneous element $v\in S_{\mathbbm K}(\mathfrak g[1]\oplus\mathfrak g^[-1])$: $$\begin{aligned} [Q,v]_\cdot&=&Q\cdot v-(-1)^{\|v\|}v\cdot Q\\ &=&Q v+2\nu\sum_{a,b,c}\frac{-1}{4} f_{ab}^c\:\xi^a\xi^b{\frac{\partial }{\partial \xi_{c}}}v\\ &&-(-1)^{\|v\|}\Big(vQ+2\nu(-1)^{\|v\|}\sum_{a,b,c}\frac{-1}{4}\:f_{ab}^c\:2\xi_c\xi^b {\frac{\partial }{\partial \xi^{a}}}v+\frac{4\nu^2}{2}\sum_{a,b,c}\frac{-1}{4}\:f_{ab}^c\:2\frac{\partial^2 v}{\partial \xi^b\partial \xi^a}\xi_c\Big)\\ &=&-\frac{\nu}{2}\sum_{a,b,c}f_{ab}^c\:\xi^a\xi^b{\frac{\partial }{\partial \xi_{c}}}v+\nu\sum_{a,b,c}f_{ab}^c\:f\:\xi^a\xi_c\:{\frac{\partial }{\partial \xi_{b}}}v+2\nu^2 q.\end{aligned}$$ Collecting the terms, equation (\[total\]) follows easily. Since each of the three terms in $0={\mbox{\boldmath $\mathcal D$}}^2=4{\mbox{\boldmath${\partial}$}}^2+2({\mbox{\boldmath${\partial}$}}{\mbox{\boldmath$\delta$}}+{\mbox{\boldmath$\delta$}}{\mbox{\boldmath${\partial}$}})+{\mbox{\boldmath$\delta$}}^2$ lives in different degrees (see below) the claim follows. For all $X\in\mathfrak g$ we have $$\begin{aligned} \label{qkosequiv} {\mbox{\boldmath{$\operatorname{L}$}}}_X\:{\mbox{\boldmath${\partial}$}}-{\mbox{\boldmath${\partial}$}}\:{\mbox{\boldmath{$\operatorname{L}$}}}_X=[{\mbox{\boldmath{$\operatorname{L}$}}}_X,{\mbox{\boldmath${\partial}$}}]=0.\end{aligned}$$ Let us write $X=\sum_a X^a\xi_a\in\mathfrak g$ with respect to a basis $\xi_1,\dots,\xi_\ell\in\mathfrak g$. For the insertation derivation $i_X:=\sum_a X^a{\frac{\partial }{\partial \xi_{a}}}$ we have the well known Cartan homotopy formula ${\mbox{\boldmath{$\operatorname{L}$}}}_X=i_X\:{\mbox{\boldmath$\delta$}}+{\mbox{\boldmath$\delta$}}\:i_X=[i_X,{\mbox{\boldmath$\delta$}}]$. Since ${\mbox{\boldmath${\partial}$}}$ obviously commutes with $i_X$ the claim follows. \ We conclude that the deformed Koszul complex $(K_\bullet[[\nu]],{\mbox{\boldmath${\partial}$}})$ is in fact a complex of $\mathfrak g$-modules. The BRST-cochains may be identified with the Lie algebra cochains of this representation. More precisely, we have $\mathscr A^n[[\nu]]=\oplus_{i,j,n=i-j}\mathscr A^{i,j}[[\nu]]$, where $\mathscr A^{i,j}[[\nu]]=C^i\big(\mathfrak g,K_j[[\nu]]\big)$. The quantum BRST-differential is (up to a trivial factor of 2) the total differential of the double complex formed be the deformed Lie algebra cohomology differential ${\mbox{\boldmath$\delta$}}:\mathscr A^{i,j}[[\nu]]\to \mathscr A^{i+1,j}[[\nu]]$ and the deformed Koszul differential ${\mbox{\boldmath${\partial}$}}: \mathscr A^{i,j}[[\nu]]\to \mathscr A^{i,j-1}[[\nu]]$. Quantum BRST as an Ext-algebra {#extalg} ------------------------------ In this section we will give a conceptual explanation for the quantum BRST algebra similar to [@Sevost]. As the material is not needed in the following, the reader may skip this section. We will work over the field $\mathbbm K{((\nu))}$ of formal Laurent series, mainly because the standard reprensentation $\rho$ of Lemma \[strep\] is not onto for formal power series in $\nu$. This has the drawback, that the classical limit makes no sense. Nonetheless, for the cases under consideration the classical limit is already at hand. From the quantum symmetry point of view it is more natural to replace the Lie bracket $[\,,\,]$ on $\mathfrak g$ by $\nu[\,,\,]$. More precisely, we consider the $\mathbbm K{[[\nu]]}$-Lie algebra $\mathfrak g{[[\nu]]}$ with bracket $[\,,\,]_\nu:=\nu[\,,\,]$ and, accordingly, the universal envelopping algebra $U\mathfrak g_\nu=T\mathfrak g{[[\nu]]}/<x\otimes y-y\otimes x-\nu [x,y]>$. This is an augmented $\mathbbm K{[[\nu]]}$-algebra, the augementation map $\epsilon:U\mathfrak g_\nu\to\mathbbm K {[[\nu]]}$ is induced by the obvious augmentation of $T\mathfrak g{[[\nu]]}$. We consider the complex $X^\nu_\bullet=U\mathfrak g_\nu\otimes_{\mathbbm K} \wedge^\bullet \mathfrak g[[\nu]]$ with differential $$\begin{aligned} d(u\otimes x_1\wedge\cdots \wedge x_n)&:=&\sum_{i=1}^n(-1)^{i+1}ux_i\otimes x_1\wedge\cdots\widehat{x_i}\cdots\wedge x_n\label{cliffmult}\\ &&+\sum_{1\le i<j\le n}(-1)^{i+j}u\otimes [x_i,x_j]_\nu\wedge x_1\wedge\cdots\widehat{x_i}\cdots\widehat{x_j}\cdots\wedge x_n.\end{aligned}$$ In the literature this complex is frequently called Koszul resolution. For obvious reasons we refrain from using this terminology. $(X^\nu_\bullet,d)$ is a free resolution of the $U\mathfrak g_\nu$-module $\mathbbm K[[\nu]]$. The proof (see e.g. [@CE p.279–282]) relies on the PBW-theorem, which applies since $\mathfrak g[[\nu]]$ is a free $\mathbbm K[[\nu]]$-module. \ It is tempting to interprete the Clifford algebra as an algebra of super-differential operators. However, in the formal situation not every super-differential operator arises in this way, since the partial derivative is decorated with the formal parameter $\nu$. We solve this problem by brute force by formally inverting $\nu$. Secondly, we have to take opposite ${\cdot_{\operatorname{opp}}}$ of the Clifford algebra multiplication (\[cliffmult\]). Here opposite is understood in the graded sense, i.e., $v{\cdot_{\operatorname{opp}}}w=(-1)^{\|v\|\|w\|}w\cdot v$. Let us identify $S_{\mathbbm K}(\mathfrak g[1]\oplus \mathfrak g^[-1])$ with $\wedge(\mathfrak g\oplus \mathfrak g^)$, which is understood with the induced grading. We will use the symbols $e_1,\dots,e_{\ell}$ to denote a basis of $\mathfrak g$ and $e^1,\dots,e^{\ell}$ will denote the corresponding dual basis. We will write $i(\alpha)$ for the super-derivation of $\wedge \mathfrak g$ which extends dual pairing with $\alpha\in\mathfrak g^$. The Clifford multiplication now writes $\mu \exp\big(-2\nu \sum_a i(e^a)\otimes i(e_a)\big)$. \[strep\]The so called standard representation $$\begin{aligned} \rho :\wedge(\mathfrak g\oplus \mathfrak g^){((\nu))}&\to& {\operatorname{End}}_\mathbbm K(\wedge \mathfrak g){((\nu))}\\ (x_1\wedge\dots\wedge x_n\otimes \alpha_1\wedge\dots\wedge\alpha_m&\mapsto& \big (v\mapsto(-2\nu)^m x_1\wedge\dots\wedge x_n\wedge i(\alpha_1)\circ\dots\circ i(\alpha_m)v\big),\end{aligned}$$ where $x_1,\dots,x_n\in\mathfrak g$ and $\alpha_1,\dots,\alpha_m\in\mathfrak g$, is an isomorphism of algebras for the reversed Clifford multiplication ${\cdot_{\operatorname{opp}}}$, i.e $\rho(a{\cdot_{\operatorname{opp}}}b)=\rho(a)\circ\rho(b)$ for all $a,b\in\wedge(\mathfrak g\oplus \mathfrak g^){((\nu))}$. In fact, if we reverse the canonical $\mathbbm Z$-grading on ${\operatorname{End}}_\mathbbm K(\wedge \mathfrak g){((\nu))}$ then $\rho$ is $\mathbbm Z$-graded. A quantum moment map ${\mbox{\boldmath{$J$}}}'$ gives rise to an algebra morphism $U\mathfrak g{[[\nu]]}\to\mathcal C^\infty(M){((\nu))}$. Applying the functor $\mathcal C^\infty(M){((\nu))}\otimes_{U\mathfrak g_\nu}-$ on the complex $X^\nu_\bullet$ yields the complex $$\mathcal C^\infty(M){((\nu))}\otimes_{U\mathfrak g_\nu}X^\nu_\bullet=\mathcal C^\infty(M)\otimes \wedge^\bullet \mathfrak g{((\nu))}=K_\bullet{((\nu))}$$ with differential $$\begin{aligned} \label{dtilde} \widetilde d (f\otimes x)=\sum_a f\star {\mbox{\boldmath{$J$}}}'(e_a)\otimes i(e^a)x-\frac{\nu}{2}\sum_{a,b,c}f_{ab}^c\:f\otimes e_c\wedge i(e^a)i(e^b)x,\end{aligned}$$ where $f\in\mathcal C^\infty(M)$ and $x\in\wedge\mathfrak g$. In fact, this differential essentially coincides with the quantum Koszul differential ${\mbox{\boldmath${\partial}$}}$. There is an isomorphism $\rho$ (extending the standard representation) of $\mathbbm Z$-graded algebras between the BRST algebra $(\mathscr A{((\nu))},{_{\operatorname{opp}}})$ with the reversed multiplication and the algebra of endomorphisms of the Koszul complex $\big({\operatorname{End}}_{\mathcal C^\infty(M){((\nu))}}(K^\bullet{((\nu))}),\circ\big)$, such that $\rho({\mbox{\boldmath $\theta$}})=-2\nu\:\widetilde{d}$. Let us write for short $A:=C^\infty(M)$. There is a canonical isomorphism of algebras $${\operatorname{End}}_{A{((\nu))}}(K^\bullet{((\nu))})={\operatorname{End}}_{A{((\nu))}}(A\otimes\wedge \mathfrak g{((\nu))})\cong A{((\nu))}^{{\operatorname{opp}}}\otimes{\operatorname{End}}_{\mathbbm K{((\nu))}}(\wedge \mathfrak g{((\nu))}).$$ The left hand side is an algebra with respect to composition and the right hand side is isomorphic to $({\mathscr A_{\mathcal V}}{[[\nu]]},^{{\operatorname{opp}}})$. If we view the algebra on the left hand side as a $\mathbbm Z$-graded algebra with the reverse of the canonical grading, then this isomorphism is actually an isomorphism of $\mathbbm Z$-graded algebras. The formula $\rho({\mbox{\boldmath $\theta$}})=-2\nu\:\widetilde{d}$ follows by inspection. If the zeroth order term $J$ of the quantum moment map ${\mbox{\boldmath{$J$}}}'$ is a moment map satisfying the generating and complete intersection hypothesis, then the opposite of the BRST-cohomology algebra $\mathrm H^\bullet\mathscr A{((\nu))}$ is isomorphic to ${\operatorname{Ext}}^\bullet_{\mathcal C^\infty(M){((\nu))}}(\mathcal B,\mathcal B)$ with composition product, where $\mathcal B$ is the left $\mathcal C^\infty(M){((\nu))}$-module which is obtained by dividing out the left ideal which is generated by ${\mbox{\boldmath{$J$}}}_1',\dots,{\mbox{\boldmath{$J$}}}_{\ell}'$. It is clear (see Proposition \[qkoscontr\] below) that the complex of equation (\[dtilde\]) is a resolution of $\mathcal B$. Therefore, the cohomology algebra of the differential graded algebra $$\left({\operatorname{End}}^{-\bullet}_{\mathcal C^\infty(M){((\nu))}}\big(K_\bullet{((\nu))},\big),\circ,[\widetilde{d},\:]\right)$$ is just the $\mathbbm Z$-graded algebra ${\operatorname{Ext}}^\bullet_{\mathcal C^\infty(M){((\nu))}}(\mathcal B,\mathcal B)$ (cf. [@Bourbdix §7]). Computation of the quantum BRST-Cohomology {#brstcomp} ------------------------------------------ The main idea which we follow in order to compute the quantum BRST cohomology (i.e., the cohomology of $(\mathscr A[[\nu]],{\mbox{\boldmath $\mathcal D$}})$), is to provide a deformed version of the contraction (\[cbrstSDR\]). This will be done by applying Lemma \[BPL2\] to the contraction (\[KosContr\]) for the perturbation ${\mbox{\boldmath${\partial}$}}$ of $\partial$ and then applying Lemma \[BPL1\] for the perturbation ${\mbox{\boldmath $\mathcal D$}}$ of $2{\mbox{\boldmath${\partial}$}}$. We will also need to examine a deformed version of the representation $\mathrm{L}^z$ of $\mathfrak g$ on $\mathcal C^\infty(Z)$. \[qkoscontr\] If we choose $h_0$ such that $h_0{\operatorname{prol}}=0$, then there are deformations of the restriction map ${\mbox{\boldmath{$\operatorname{res}$}}}={\operatorname{res}}+\sum_{i\ge 1}\nu^i\;{\operatorname{res}}_{i}:\mathcal C^{\infty}(M)\to \mathcal C^{\infty}(Z)[[\nu]]$ and of the contracting homotopies ${\mbox{\boldmath $h$}}=h+\sum_{j\ge 1}\nu^j\;h_{(j)}:K_\bullet[[\nu]]\to K_{\bullet+1}[[\nu]]$, which are a formal power series of Fréchet continuous maps and such that $$\begin{aligned} \label{qKosContr} \big(\mathcal C^{\infty}(Z)[[\nu]],0\big)\:\begin{array}{c} {{\mbox{\boldmath{$\operatorname{res}$}}}}\\\leftrightarrows\\ {{\operatorname{prol}}} \end{array}\:\big(K[[\nu]],{\mbox{\boldmath${\partial}$}}\big),{\mbox{\boldmath $h$}}\end{aligned}$$ is a contraction with ${\mbox{\boldmath $h$}}_0\;{\operatorname{prol}}=0$. Explicitly, we have $${\mbox{\boldmath{$\operatorname{res}$}}}:={\operatorname{res}}\;({\operatorname{id}}+({\mbox{\boldmath${\partial}$}}_1 -\partial_1)h_0)^{-1}.$$ If we choose $h$ to be $\mathfrak g$-equivariant, the same is true for ${\mbox{\boldmath $h$}}$. Apply lemma \[BPL2\] to the perturbation ${\mbox{\boldmath${\partial}$}}$ of $\partial$. \ We are now ready to define the quantized representation ${\mbox{\boldmath{$\operatorname{L}$}}}^z$ of $\mathfrak g$ on $\mathcal C^\infty (Z)[[\nu]]$ by setting $${\mbox{\boldmath{$\operatorname{L}$}}}^z_X:={\mbox{\boldmath{$\operatorname{res}$}}}\;{\mbox{\boldmath{$\operatorname{L}$}}}_X \;{\operatorname{prol}}\quad\mbox{ for $X\in \mathfrak g$.}$$ We have ${\mbox{\boldmath{$\operatorname{L}$}}}^z_X\:{\mbox{\boldmath{$\operatorname{L}$}}}^z_Y -{\mbox{\boldmath{$\operatorname{L}$}}}^z_X\:{\mbox{\boldmath{$\operatorname{L}$}}}^z_Y={\mbox{\boldmath{$\operatorname{L}$}}}^z_{[X,Y]}$ for all $X,Y\in\mathfrak g$. The claim follows from $$\begin{aligned} {\mbox{\boldmath{$\operatorname{L}$}}}^z_X\:{\mbox{\boldmath{$\operatorname{L}$}}}^z_Y&=&{\mbox{\boldmath{$\operatorname{res}$}}}\:{\mbox{\boldmath{$\operatorname{L}$}}}_X\:{\operatorname{prol}}\:{\mbox{\boldmath{$\operatorname{res}$}}}\:{\mbox{\boldmath{$\operatorname{L}$}}}_X\:{\operatorname{prol}}\\ &\stackrel{(\ref{qKosContr})}{=}&{\mbox{\boldmath{$\operatorname{res}$}}}\:{\mbox{\boldmath{$\operatorname{L}$}}}_X({\operatorname{id}}-{\mbox{\boldmath${\partial}$}}{\mbox{\boldmath $h$}}){\mbox{\boldmath{$\operatorname{L}$}}}_Y\:{\operatorname{prol}}\:\stackrel{(\ref{qkosequiv})}{=}{\mbox{\boldmath{$\operatorname{res}$}}}({\operatorname{id}}-{\mbox{\boldmath${\partial}$}}{\mbox{\boldmath $h$}}){\mbox{\boldmath{$\operatorname{L}$}}}_X\:{\mbox{\boldmath{$\operatorname{L}$}}}_Y\:{\operatorname{prol}}\stackrel{(\ref{qKosContr})}{=}{\mbox{\boldmath{$\operatorname{res}$}}}\:{\mbox{\boldmath{$\operatorname{L}$}}}_X\:{\mbox{\boldmath{$\operatorname{L}$}}}_Y\:{\operatorname{prol}}\end{aligned}$$ and the fact that ${\mbox{\boldmath{$\operatorname{L}$}}}$ is a representation. \ In the same fashion as in Section \[classred\], we define ${\mbox{\boldmath $d$}}:C^\bullet(\mathfrak g,\mathcal C^\infty (Z)[[\nu]])\to C^{\bullet+1} (\mathfrak g,\mathcal C^\infty (Z)[[\nu]])$ to be the differential of Lie algebra cohomology of the representation ${\mbox{\boldmath{$\operatorname{L}$}}}^z$, i.e., ${\mbox{\boldmath $d$}}\;{\mbox{\boldmath{$\operatorname{res}$}}}={\mbox{\boldmath{$\operatorname{res}$}}}\;{\mbox{\boldmath$\delta$}}$. In the same manner, we extend ${\mbox{\boldmath{$\operatorname{res}$}}}$ and ${\mbox{\boldmath $h$}}$ as in Section \[classred\] to maps ${\mbox{\boldmath{$\operatorname{res}$}}}:\mathscr A\to C(\mathfrak g,\mathcal C^{\infty}(Z)[[\nu]]\big)$ and ${\mbox{\boldmath $h$}}:\mathscr A^\bullet[[\nu]]\to\mathscr A^{\bullet-1}[[\nu]]$. \[Hauptsatz\]\[maincontr\] There are $\mathbbm K[[\nu]]$-linear maps ${\mbox{\boldmath{$\Phi$}}}:C^\bullet\big(\mathfrak g, \mathcal C^{\infty}(Z)\big)\to \mathscr A^\bullet[[\nu]]$ and ${\mbox{\boldmath $H$}}: \mathscr{A}^\bullet\to \mathscr{A}^{\bullet-1}[[\nu]]$, which are series of Fréchet continuous maps such that $$\begin{aligned} \label{maincontrformula} \Big(C^\bullet\big(\mathfrak g, \mathcal C^{\infty}(Z)[[\nu]]\big),{\mbox{\boldmath $d$}}\Big)\: \begin{array}{c} {{\mbox{\boldmath{$\operatorname{res}$}}}}\\\leftrightarrows\\ {{\mbox{\boldmath{$\Phi$}}}} \end{array} \:(\mathscr A^\bullet[[\nu]],{\mbox{\boldmath $\mathcal D$}}),{\mbox{\boldmath $H$}}\end{aligned}$$ is a contraction. Since the requisite condition ${\mbox{\boldmath{$\operatorname{res}$}}}\;{\mbox{\boldmath $h$}}=0$ is obviously fulfilled, we apply Lemma \[BPL1\] to the perturbation ${\mbox{\boldmath $\mathcal D$}}$ of $2{\mbox{\boldmath${\partial}$}}$. Explicitly, this means that ${\mbox{\boldmath $H$}}:=\frac{1} {2} {\mbox{\boldmath $h$}}\sum_{j=0}^\ell(-\frac{1}{2})^j({\mbox{\boldmath $h$}}{\mbox{\boldmath$\delta$}}+{\mbox{\boldmath$\delta$}}{\mbox{\boldmath $h$}})^j $ and ${\mbox{\boldmath{$\Phi$}}}={\operatorname{prol}}-{\mbox{\boldmath $H$}}({\mbox{\boldmath$\delta$}}\;{\operatorname{prol}}-{\operatorname{prol}}\:{\mbox{\boldmath $d$}})$, which are obviously series of Fréchet continuous maps. Note that from $h_0 {\operatorname{prol}}=0$ and $h^2=0$, we get ${\mbox{\boldmath $H$}}{\mbox{\boldmath{$\Phi$}}}=0$ and ${\mbox{\boldmath $H$}}^2=0$. If ${\operatorname{prol}}$ is chosen to be equivariant, then the expression for ${\mbox{\boldmath{$\Phi$}}}$ simplifies to ${\mbox{\boldmath{$\Phi$}}}={\operatorname{prol}}$. If $h$ and (hence ${\mbox{\boldmath $h$}}$) is equivariant, then it follows that ${\mbox{\boldmath $H$}}=\frac{1}{2}{\mbox{\boldmath $h$}}$. \ For better intelligibility, let us mention that the above argument is a strengthening of the well known tic-tac-toe lemma [@BottTu p.135]: The homology of the total differential of a double complex with acyclic rows is isomorphic to the homology of the differential, which is induced by the action of the vertical differential on the horizontal homology. The double complex in question is of course that of Theorem \[doppelcompl\]. It is depicted in the following diagram from the second column onwards. $$\begin{aligned} \xymatrix{ \cdots&\cdots&\cdots&\\ C^1\big(\mathfrak g, \mathcal C^{\infty}(Z)[[\nu]]\big)\ar[u]^{\mbox{\boldmath $d$}}&\mathscr A^{1,0}[[\nu]]\ar[l]^{\qquad{\mbox{\boldmath{$\operatorname{res}$}}}}\ar[u]^{\mbox{\boldmath$\delta$}}&\mathscr A^{1,1}[[\nu]]\ar[l]^{\:\:\:\:2{\mbox{\boldmath${\partial}$}}}\ar[u]&\qquad\cdots\qquad\ar[l]\\ C^0\big(\mathfrak g, \mathcal C^{\infty}(Z)[[\nu]]\big)\ar[u]^{\mbox{\boldmath $d$}}&K_0(J,M)[[\nu]]\ar[l]^{\quad{\mbox{\boldmath{$\operatorname{res}$}}}}\ar[u]^{\mbox{\boldmath$\delta$}}&K_1(J,M)[[\nu]]\ar[l]^{\:\:\:\:2{\mbox{\boldmath${\partial}$}}}\ar[u]^{\mbox{\boldmath$\delta$}}&K_2(J,M)[[\nu]]\cdots\ar[l]^{2 {\mbox{\boldmath${\partial}$}}}}\end{aligned}$$ The first column is the Lie algebra cochain complex of the $\mathfrak g$-module $\mathcal C^\infty(Z)[[\nu]]$, which is quasiisomorphic to the BRST complex via the quasiisomorphism ${\mbox{\boldmath{$\operatorname{res}$}}}$. We use the contraction (\[maincontrformula\]) to transfer the associative algebra structure from $\mathscr A[[\nu]]$ to the Lie algebra cohomology $\mathrm H^\bullet\big(\mathfrak g,\mathcal C^\infty(Z)[[\nu]]\big)$ of the representation ${\mbox{\boldmath{$\operatorname{L}$}}}^z$ by setting $$\begin{aligned} \label{Formel1} [a][b]:=[{\mbox{\boldmath{$\operatorname{res}$}}}\big({\mbox{\boldmath{$\Phi$}}}(a){\mbox{\boldmath{$\Phi$}}}(b)\big)],\end{aligned}$$ where $[a],[b]$ denote the cohomology classes of the cocyles $a,b\in Z^\bullet\big(\mathfrak g, \mathcal C^{\infty}(Z)[[\nu]]\big)$. The associativity of this operation follows from $$\begin{aligned} [a]([b][c])&=&\Big[{\mbox{\boldmath{$\operatorname{res}$}}}\Big({\mbox{\boldmath{$\Phi$}}}(a){\mbox{\boldmath{$\Phi$}}}\big({\mbox{\boldmath{$\operatorname{res}$}}}\big({\mbox{\boldmath{$\Phi$}}}(b){\mbox{\boldmath{$\Phi$}}}(c)\big)\big)\Big)\Big]\nonumber\\ &=&\Big[{\mbox{\boldmath{$\operatorname{res}$}}}\big({\mbox{\boldmath{$\Phi$}}}(a){\mbox{\boldmath{$\Phi$}}}(b){\mbox{\boldmath{$\Phi$}}}(c)\big)\Big]-\Big[{\mbox{\boldmath{$\operatorname{res}$}}}\Big({\mbox{\boldmath{$\Phi$}}}(a)\big({\mbox{\boldmath $\mathcal D$}}{\mbox{\boldmath $H$}}\big({\mbox{\boldmath{$\Phi$}}}(b){\mbox{\boldmath{$\Phi$}}}(c)\big)\big)\Big)\Big]\nonumber\\ &&-\Big[{\mbox{\boldmath{$\operatorname{res}$}}}\Big({\mbox{\boldmath{$\Phi$}}}(a)\big({\mbox{\boldmath $H$}}{\mbox{\boldmath $\mathcal D$}}\big({\mbox{\boldmath{$\Phi$}}}(b){\mbox{\boldmath{$\Phi$}}}(c)\big)\big)\Big)\Big]\nonumber\\ &=&\Big[{\mbox{\boldmath{$\operatorname{res}$}}}\big({\mbox{\boldmath{$\Phi$}}}(a){\mbox{\boldmath{$\Phi$}}}(b){\mbox{\boldmath{$\Phi$}}}(c)\big)\Big]-\Big[{\mbox{\boldmath $d$}}\:{\mbox{\boldmath{$\operatorname{res}$}}}\Big({\mbox{\boldmath{$\Phi$}}}(a)\big({\mbox{\boldmath $H$}}\big({\mbox{\boldmath{$\Phi$}}}(b){\mbox{\boldmath{$\Phi$}}}(c)\big)\big)\Big)\Big]\nonumber\\ &&+\Big[{\mbox{\boldmath{$\operatorname{res}$}}}\Big({\mbox{\boldmath{$\Phi$}}}({\mbox{\boldmath $d$}}a)\big({\mbox{\boldmath $H$}}\big({\mbox{\boldmath{$\Phi$}}}(b){\mbox{\boldmath{$\Phi$}}}(c)\big)\big)\Big)\Big]\nonumber\\ &&-\Big[{\mbox{\boldmath{$\operatorname{res}$}}}\Big({\mbox{\boldmath{$\Phi$}}}(a)\big({\mbox{\boldmath $H$}}\big({\mbox{\boldmath{$\Phi$}}}({\mbox{\boldmath $d$}}b){\mbox{\boldmath{$\Phi$}}}(c)+{\mbox{\boldmath{$\Phi$}}}(b){\mbox{\boldmath{$\Phi$}}}({\mbox{\boldmath $d$}}c)\big)\big)\Big)\Big]\nonumber\\ &=&\Big[{\mbox{\boldmath{$\operatorname{res}$}}}\big({\mbox{\boldmath{$\Phi$}}}(a){\mbox{\boldmath{$\Phi$}}}(b){\mbox{\boldmath{$\Phi$}}}(c)\big)\Big]\label{assoccomp},\end{aligned}$$ which coincides, as a result of a similar calculation, with $([a][b])[c]$. However, this is not exactly, what we want to accomplish. The primary obstacle on the way to the main result, Corollary \[redprod\], is that, in general, we have $$\begin{aligned} \label{anomaly}\mathrm H^0\big(\mathfrak g,\mathcal C^\infty(Z)[[\nu]]\big)\neq \mathrm H^0\big(\mathfrak g,\mathcal C^\infty(Z)\big)[[\nu]],\end{aligned}$$ since there is no a priori reason that the representations ${\mbox{\boldmath{$\operatorname{L}$}}}^z$ and $\mathrm L^z$ have the same space of invariants. An example where this phenomenon does in fact occur has been given in [@BHW section 7]. One way out is to sharpen the compatibility condition (\[qmomentmap\]). We require, that ${\mbox{\boldmath{$J$}}}=J$ and $$J(X)\star f- f \star J(X)=\nu\{J(X),f\}\quad\mbox{ for all }X\in \mathfrak g, f\in \mathcal C^{\infty}(M).$$ This property, which has been discussed in Subsection \[qmom\], is referred to as strong invariance of the star product $\star$ with respect to the Lie algebra action. It can always be achieved for the cases under consideration. Of course, now the representations ${\mbox{\boldmath{$\operatorname{L}$}}}$ and $\mathrm L$ coincide and we get $\delta={\mbox{\boldmath$\delta$}}$. But with some mild restrictions on the contracting homotopy $h$ of the Koszul resolution we also have the following. If $h_0$ is $\mathfrak g$-equivariant and $h_0{\operatorname{prol}}=0$, then ${\mbox{\boldmath{$\operatorname{L}$}}}^z=\mathrm L^z$. For $X\in\mathfrak g$ we have ${\mbox{\boldmath{$\operatorname{L}$}}}^z_X={\mbox{\boldmath{$\operatorname{res}$}}}\;\mathrm L_X \;{\operatorname{prol}}={\operatorname{res}}\;({\operatorname{id}}+({\mbox{\boldmath${\partial}$}}_1 -\partial_1)h_0)^{-1} \mathrm L_X\;{\operatorname{prol}}$. Since $\mathrm L_X$ commutes with ${\mbox{\boldmath${\partial}$}}_1$, $\partial_1$ and $h_0$, the last expression can be written as ${\operatorname{res}}\;\mathrm L_X({\operatorname{id}}+({\mbox{\boldmath${\partial}$}}_1 -\partial_1)h_0)^{-1}{\operatorname{prol}}={\operatorname{res}}\;\mathrm L_X \;{\operatorname{prol}}$. \[redprod\] With the assumptions made above, the product defined by equation (\[Formel1\]) makes $\mathrm H^\bullet\big(\mathfrak g,\mathcal C^\infty(Z)\big)[[\nu]]$ into a $\mathbbm Z$-graded associative algebra. For the subalgebra of invariants $\mathrm H^{0}\big(\mathfrak g, \mathcal C^{\infty}(Z)\big)[[\nu]]= \big(\mathcal C^{\infty}(Z)\big)^{\mathfrak g}[[\nu]]$ this formula simplifies to $$\begin{aligned} \label{Formel2} fg:={\mbox{\boldmath{$\operatorname{res}$}}}\big({\operatorname{prol}}(f){\operatorname{prol}}(g)\big)\quad\mbox{ for }f,g\in\big(\mathcal C^{\infty}(Z)\big)^{\mathfrak g}.\end{aligned}$$ Since $\big(\mathcal C^{\infty}(Z)\big)^{\mathfrak g}[[\nu]]$ is $\mathbbm K[[\nu]]$-linearly isomorphic to the algebra of smooth functions on the symplectic stratified space $M_{\mathsf{red}}$, we obtain an associative product on $\mathcal C^\infty(M_{\mathsf{red}})[[\nu]]$ which gives rise to a continuous Hochschild cochain. There is another strategy to attack problem (\[anomaly\]). If $\mathrm H^1(\mathfrak g, C^\infty(Z))$ vanishes, it is possible to find a topologically linear isomorphism between the spaces of invariants for the classical and the deformed representation. Let $G$ be a compact, connected semisimple Lie group acting on the Poisson manifold $M$ in a Hamiltonian fashion. Assume that the equivariant moment map $J$ satisfies the generating and complete intersection hypothesis. Then for any star product $$ on $M$ with quantum moment map ${\mbox{\boldmath{$J$}}}$ there is a invertible sequence of continuous maps $$S=\sum_{i\ge 0}\nu^i\:S_i :\mathrm H^0(\mathfrak g, \mathcal C^\infty(Z))[[\nu]]=\mathcal C^\infty(Z)^\mathfrak g[[\nu]]\to \mathrm H^0(\mathfrak g,\mathcal C^\infty(Z)[[\nu]])$$ such that the formula $$\begin{aligned} f g:= S^{-1}\big(S(f)S(g)\big)=S^{-1}\Big({\mbox{\boldmath{$\operatorname{res}$}}}\big({\mbox{\boldmath{$\Phi$}}}(S(f)){\mbox{\boldmath{$\Phi$}}}(S(g))\big)\Big) \end{aligned}$$ defines a continuous formal deformation of the Poisson algebra $\mathcal C^\infty(Z)^\mathfrak g$ into an associative algebra. According to Viktor L. Ginzburg (see [@Ginz98 Theorem 2.13]) we have for any compact connected Lie group $G$ with a smooth representation on a Fréchet space $W$ an isomorphism $$\begin{aligned} \mathrm H^\bullet(\mathfrak g,W)\cong \mathrm H^\bullet(\mathfrak g,\mathbbm K)\otimes W^\mathfrak g.\end{aligned}$$ In particular, this implies that if $\mathfrak g$ is semisimple the first and the second cohomology groups of the $\mathfrak g$-module $\mathcal C^\infty(Z)$ vanish (for the so called Whitehead lemmata see e.g. [@Hilton]). Note that the Fréchet subspace of invariant functions $\mathcal C^\infty(Z)^\mathfrak g\subset \mathcal C^\infty(Z)$ has a closed complementary subspace $V$ $$\mathcal C^\infty(Z)=\mathcal C^\infty(Z)^\mathfrak g\oplus V.$$ This can be achieved by taking $V$ to be the kernel of the averaging projection $\pi:\mathcal C^\infty(Z)\to \mathcal C^\infty(Z)^G$, $\pi(f)(x):=\operatorname{vol}(G)^{-1}\int_G f(gx)dg$. Hence, the restriction of the Lie algebra cohomology differential $d$ to the closed complementary subspace $V$ is a bijection onto the closed supspace $Z^1(\mathfrak g,\mathcal C^\infty(Z))\subset C^1(\mathfrak g,\mathcal C^\infty(Z))$. Since every continuous linear bijection of Fréchet spaces has a continuous inverse (see [@Rudin corollary 2.12]), we have a continuous inverse map, which we call $d^{-1}$. $$\begin{aligned} \xymatrix{\mathcal C^\infty(Z)\ar[r]^{d\quad}&\mathfrak g^\otimes \mathcal C^\infty(Z)\\ V\ar@{^{(}->}[u]\ar[r]^{\cong\qquad}&Z^1(\mathfrak g,\mathcal C^\infty(Z))\ar@{^{(}->}[u]\ar@/^2pc/[l]^{d^{-1}} }\end{aligned}$$ Let $\varphi_0\in\mathcal C^\infty(Z)^\mathfrak g$. We will inductively construct an element $\varphi=\sum_i\nu^i\varphi_i$ which is invariant for the deformed representation, i.e., ${\mbox{\boldmath $d$}}\varphi=0$. Let us assume that we have found $\varphi_0,\varphi_1,\dots,\varphi_n\in \mathcal C^\infty(Z)$, such that $$\begin{aligned} \label{IV} \sum_{j=0}^i d_j\varphi_{i-j}=0\quad\forall i=0,\dots,n.\end{aligned}$$ We are looking for an element $\varphi_{n+1}$, such that $\sum_{j=0}^{n+1}d_j \varphi_{n-j}=0$. Rewriting the equation ${\mbox{\boldmath $d$}}^2=(\sum_{i=0}^\infty \nu^i d_i)^2$ order by order in the powers of $\nu$ we obtain $d_0d_j=-\sum_{i=0}^{j-1}d_{j-i}d_i$. Now, an easy calculation yields that $\sum_{i=0}^nd_{i+1}\varphi_{n-i}\in Z^1(\mathfrak g,\mathcal C^\infty(Z))$: $$\begin{aligned} d_0\big(\sum_{i=0}^nd_{i+1}\varphi_{n-i}\big)=-\sum_{i=0}^n\sum_{k=0}^id_{i-k}d_k\varphi_{n-i}=-\sum_{r=0}^n d_{n-r}\sum_{s=0}^{n-r}d_s\varphi_{r-s}\stackrel{(\ref{IV})}{=}0.\end{aligned}$$ Setting $\varphi_{n+1}:=-d^{-1}(\sum_{i=0}^nd_{i+1}\varphi_{n-i})$, we are done. Obviously, $\varphi_1=-d^{-1}d_1\varphi_0=:S_1(\varphi_0)$, $\varphi_2=-d^{-1}(d_2\varphi_0-d^{-1}d_1\varphi_0)=:S_2(\varphi_0)$, etc. arise by successive continuous operations acting on $\varphi_0$ and we acquire the desired sequence $S=\sum_{i=0}^\infty\nu^i S_i$. Conversely, let $\varphi=\sum_i\nu^i \varphi_i\in \mathrm H^0(\mathfrak g,\mathcal C^\infty(Z)[[\nu]])$. Rewriting ${\mbox{\boldmath $d$}}\varphi=0$ order by order in powers of $\nu$ we get $\sum_{j=0}^n d_j\varphi_{n-j}=0$ for all $n\ge0$. In particular we have $d_0(\sum_{j=1}^nd_j\varphi_{n-j})=0$. Setting $\psi_n:=\varphi+d^{-1}(\sum_{j=1}^nd_j\varphi_{n-j})$, we obtain a series $\psi:=\sum_{i=0}^\infty\nu^i\psi_i\in \mathcal C^\infty(Z)^\mathfrak g[[\nu]]$. It is clear that $S\psi=\varphi$. Auxiliary material ================== Two perturbation lemmata {#HPT} ------------------------ We consider (cochain) complexes in an additive $\mathbb K$-linear category $\mathscr C$ (e.g. the category of Fréchet spaces). A contraction* in $\mathscr C$ consists of the following data $$\begin{aligned} \label{contr} \xymatrix{(X,d_X)\ar@<-0.6ex>[r]_{i\:\:\:\:}^{}&(Y,d_Y),h_Y, \ar@<-0.6ex>[l]_{p\:\:\:\:}^{}}\end{aligned}$$ where $i$ and $p$ are chain maps between the chain complexes $(X,d_X)$ and $(Y,d_Y)$, $h_Y:Y\to Y[-1]$ is a morphism, and we have $pi={\operatorname{id}}_X$, $d_Yh_Y+h_Yd_Y={\operatorname{id}}_Y-ip$. The contraction is said to satisfy the side conditions (sc1–3), if moreover, $h_Y^2=0$ , $h_Yi=0$ and $ph_Y=0$ are true. It was observed in [@LambeSt], that in order to fulfill (sc2) and (sc3), one can replace $h_Y$ by $h'_Y:=(d_Yh_Y+h_Yd_Y) \:h_Y\:(d_Y h_Y+h_Y d_Y)$. If one wants to have in addition (sc1) to be satisfied, one may replace $h_Y'$ by $h_Y'':=h_Y'd_Yh'_Y$. Let $C:=Cone(p)$ be the mapping cone of $p$, i.e., $C=X[1]\oplus Y$ is the complex with differential $d_C(x,y):=(d_Xx+(-1)^{\|y\|}py,d_Y y)$. The homology of $C$ is trivial, because $h_C(x,y):=(0,h_Yy+(-1)^{\|x\|}ix)$ is a contracting homotopy, i.e., $d_Ch_C+h_Cd_C={\operatorname{id}}_C$, if (sc3) is true. In fact, we calculate $$\begin{aligned} (d_Ch_C+h_Cd_C)(x,y)&=&d_C\big(0,h_Yy+(-1)^{\|x\|}ix, d_Y y\big)+h_C\big(d_Xx+(-1)^{\|y\|}py,d_Yy\big)\\ &=&\big((-1)^{\|y\|+1}ph_Yy+pix,d_Yh_Yy+(-1)^{\|x\|}d_Yix\big)\\ &&\qquad\qquad\qquad\qquad\qquad\qquad+\big(0,h_Yd_Yy+(-1)^{\|x\|+1}i\:d_Xx+ipy\big)\\ &\stackrel{(sc3)}{=}&\big(pix,(h_Yd_Y+d_Yh_Y)y+ipy\big)=(x,y).\end{aligned}$$ Let us now assume that the objects $X$ and $Y$ carry descending Hausdorff filtrations and the structure maps are filtration preserving. Moreover, pretend that we have found a perturbation $D_Y=d_Y+t_Y$ of $d_Y$, i.e., $D_Y^2=0$ and $t_Y:Y\to Y[1]$, called the initiator, having the property that $t_Yh_Y+h_Yt_Y$ raises the filtration. Since, in general, $t_X:=pt_Yi$ does not need to be a perturbation of $d_X$, we impose that as an extra condition: we assume that $D_X=d_X+t_X$ is a differential. Setting $t_C:=(t_X,t_Y)$, we will get a perturbation $D_C:=d_C+t_C$ of $d_C$, if we have in addition $t_Xp=pt_Y$ (this will imply that $(d_X+t_X)^2=0$). Then an easy calculation yields that $H_C:=h_C(D_Ch_C+h_CD_C)^{-1}=h_C({\operatorname{id}}_C+t_Ch_C+h_Ct_C)^{-1}$ is well defined and satisfies $D_CH_C+H_CD_C={\operatorname{id}}_C$. In fact, we have $$\begin{aligned} D_CH_C+H_CD_C&=&\big(D_Ch_C+H_CD_C(D_Ch_C+h_CD_C)\big)(D_Ch_C+h_CD_C)^{-1}\\ &=&\big(D_Ch_C+H_C(D_Ch_C+h_CD_C)D_C\big)(D_Ch_C+h_CD_C)^{-1}\\ &=&\big(D_Ch_C+h_CD_C\big)(D_Ch_C+h_CD_C)^{-1}={\operatorname{id}}_C.\end{aligned}$$ Defining the morphism $I:X\to Y$, $H_C(x,0)=:(0,(-1)^{\|x\|}Ix)$ and the homotopy $H_Y:Y\to Y[-1]$, $H_C(0,y)=:(0,H_Yy)$ we obtain the following \[BPL1\]If the contraction (\[contr\]) satisfies (sc3) and $D_Y=d_Y+t_Y$ is a perturbation of $d_Y$ such that $t_Xp=pt_Y$, then $$\begin{aligned} \label{contrEins} \xymatrix{(X,D_X)\ar@<-0.6ex>[r]_{I\:\:\:\:}^{}&(Y,D_Y),H_Y \ar@<-0.6ex>[l]_{p\:\:\:\:}^{}}\end{aligned}$$ is a contraction fulfilling (sc3). Moreover, we have $H_Y= h_Y({\operatorname{id}}_Y+t_Yh_Y+h_Yt_Y)^{-1}$ and $Ix=ix-H_Y(t_Yix-it_Xx)$. If all side conditions are true for (\[contr\]), then they are for (\[contrEins\]), too. For all homogenuous $x\in X$, $y\in Y$ we have $$\begin{aligned} (x,y)&=&(D_CH_C+H_CD_C)(x,y)\\ &=&D_C\big(0,(-1)^{\|x\|}Ix+H_Yy\big)+H_C\big(D_Xx+(-1)^{\|y\|}py,D_Yy\big)\\ &=&\big(pIx+(-1)^{\|y\|+1}pH_Yy,(-1)^{\|x\|}D_YIx+D_YH_Yy\big)\\ &&\qquad\qquad\qquad\qquad+\big(0,(-1)^{\|x\|-1}ID_Xx+Ipy+H_YD_Yy\big),\end{aligned}$$ and we conclude that $pI={\operatorname{id}}_X$, $D_YI=ID_X$, $pH_Y=0$ and $D_YH_Y+H_YD_Y+Ip={\operatorname{id}}_Y$. Let us verify the formula for $H_Y$: $(0,H_Yy)=H_C(0,y)=h_C\sum_{i\ge 0}(-1)^i(t_Ch_C+h_Ct_C)^i(0,y) =\big(0,h_Y\sum_{i\ge 0}(-1)^i(t_Yh_Y+h_Yt_Y)^iy\big).$ Note that $(t_Ch_C+h_Ct_C)(x,0)=(-1)^{\|x\|}(0,t_Y ix-it_Xx)$. Using this result, it is straight forward to check: $(0,Ix)=(0,ix)-\big(0,h_Y\sum_{i\ge0}(-1)^i(t_Yh_Y+h_Yt_Y)^i(t_Y ix-it_Xx) =\big(0,ix-H_Y(t_Y ix-it_Xx)\big).$ Finally, let us address the side conditions. With the above formula for $H_Y$ the condition $i h_Y=0$ entails $iH_Y=0$. Note that $h_Y^2=0$ implies $h_Y(D_Yh_Y+h_YD_Y)^{-1}=(D_Yh_Y+h_YD_Y)^{-1}h_Y$, and we conlude that $H_Y^2=0$. \ \ Starting with the mapping cone $K=Cone(i)$, i.e., the complex $K=Y[1]\oplus X$ with the differential $d_K(y,x)=(d_Yy+(-1)^{\|x\|}ix,d_Xx)$, we may give a version of the above argument arriving at a contraction with all data perturbed except $i$. More precisely, we have a homotopy $h_K(y,x):=(h_Yy,(-1)^{\|y\|}py)$, for which $d_Kh_K+h_Kd_K={\operatorname{id}}_K$ follows from (sc2). In fact, we calculate $$\begin{aligned} (d_Kh_K+h_Kd_K)(y,x)&=&d_K\big(h_Y,(-1)^{\|y\|}py\big)+h_K\big(d_Yy+(-1)^{\|x\|}ix,d_X x\big)\\ &=&\big(d_Yh_Yy+ipy,(-1)^{\|y\|}d_Xpy)\big)\\ &&\qquad\qquad\qquad\qquad+\big(h_Yd_Y+(-1)^{\|x\|}h_Yix,(-1)^{\|y\|+1}pd_Yy+pix\big)\\ &\stackrel{(sc2)}{=}&\big(d_Yh_Yy+h_Yd_Y+ipy,pix\big)=(y,x).\end{aligned}$$ Mimicking the above argument for $C$, we get a differential $D_K:=d_K+t_K$ with $t_K:=(t_Y,t_X)$, if $t_Yi=it_X$ (this will imply $D_X^2=0$). Assuming (\[contr\]) to satisfy (sc2), $H_K:=h_K(D_Kh_K+h_KD_K)^{-1}$ will become a contracting homotopy $D_KH_K+H_KD_K={\operatorname{id}}_K$. Defining $P:Y\to X$ and $H'_Y:Y\to Y[-1]$ by $H_K(y,0)=H_K(y,x)=:(H'_Yy,(-1)^{\|y\|}Py)$ we get the following \[Perturbation Lemma – Version 2\] \[BPL2\]If the contraction (\[contr\]) satisfies (sc2) and $D_Y=d_Y+t_Y$ is a perturbation of $d_Y$ such that $t_Yi=it_X$, then $$\begin{aligned} \label{contrZwei} \xymatrix{(X,D_X)\ar@<-0.6ex>[r]_{i\:\:\:\:}^{}&(Y,D_Y),H'_Y \ar@<-0.6ex>[l]_{P\:\:\:\:}^{}}\end{aligned}$$ is a contraction fulfilling (sc2). Moreover, we have $H'_Y= h_Y({\operatorname{id}}_Y+t_Yh_Y+h_Yt_Y)^{-1}$ and $P=p({\operatorname{id}}+t_Yh_Y+h_Yt_Y)^{-1}$. If all side conditions are true for (\[contr\]), then they are for (\[contrZwei\]), too. For all homogenuous $x\in X$, $y\in Y$ we have $$\begin{aligned} (y,x)&=&(D_KH_K+H_KD_K)(y,x)\\ &=&D_K\big(H'_Yy,(-1)^{\|y\|}Py\big)+H_K\big(D_Y+(-1)^{\|x\|}ix,D_X x\big)\\ &=&\big(D_Y H'_Yy+iPy,(-1)^{\|y\|}D_XPy\big)\\ &&\qquad\qquad\qquad\qquad+\big(H'_YD_Yy+(-1)^{\|x\|}H'_Yix,(-1)^{\|y\|+1}PD_Yy+Pix\big),\end{aligned}$$ and we conclude that $D_YH'_Y+H'_YD_Y+iP={\operatorname{id}}_Y$, $H'_Yi=0$, $D_XP=PD_Y$ and $Pi={\operatorname{id}}_X$. Let us check the formulas for $H_Y'$ and $P$: $$\begin{aligned} (H_Y'y,(-1)^{\|y\|}Py)&=&H_K(y,0)=\big(h_K(\sum_{i\ge0}(-1)^i(t_Yh_Y+h_Yt_Y)^iy,0\big)\\ &=&\big(h_Y\sum_{i\ge0}(-1)^i(t_Yh_Y+h_Yt_Y)^iy,(-1)^{\|y\|}p\sum_{i\ge0}(-1)^i(t_Yh_Y+h_Yt_Y)^iy\big). \end{aligned}$$ Finally, let us address the side conditions. If $h_Y^2=0$ then $h_Y(1+t_Yh_Y+h_yt_Y)^{-1}=(1+t_Yh_Y+h_yt_Y)^{-1}h_Y$, which entails ${H'}^2_Y=0$. If in addition $ph_Y=0$, then we conclude that $PH'_Y=0$. Graded Lie-Rinehart pairs {#LieRinehart} ------------------------- The notion of a Lie-Rinehart pair is the algebraic counterpart of the notion of a Lie algebroid. It admits a more or less obvious translation to the graded situation. $\mathbbm Z_2$-graded Lie-Rinehart have been studied for example in [@Chemla]. In the following graded objects, morphisms etc. are understood to be in one of the abelian tensor categories ${{{\mathbbm K}\mathbf{-Mod}^{\mathbbm Z_2}}}$ and ${{{\mathbbm K}\mathbf{-Mod}^{\mathbbm Z}}}$, the category of $\mathbbm Z_2$-graded and $\mathbb Z$-graded vector spaces, respectively. Rinehart [@Rinehart] introduced the notion of an universal enveloping algebra of a Lie-Rinehart pair generalizing the universal enveloping algebra of a Lie algebra. In [@HuebPoiss] Hübschmann gave an alternative construction for it, which we will translate to the graded situation. A graded Lie Rinehart pair $(A,L)$ is a graded commutative $\mathbbm K$-algebra $A$ and a graded $\mathbbm K$-Lie algebra $L$ such that 1. $L$ is a graded left $A$-module $A\otimes L\to L$, $(a,X)\mapsto aX$, 2. $L$ acts on $A$ by graded left derivations $L\otimes A\to A$, $(X,a)\mapsto X(a)$, 3. the actions are compatible in the following sense $$\begin{aligned} aX(b)&=&(aX)(b),\\ {[X,aY]}&=&X(a)Y+(-1)^{\|a\|\|X\|}a[X,Y] \mbox{ for all homogeneous } X,Y\in L\mbox{ and }a,b\in A.\end{aligned}$$ Sometimes $L$ will also be called a $(\mathbbm K,A)$-Lie algebra. An $(A,L)$-module is a graded $\mathbbm K$-vector space $V$ which is at the same time an $A$-and an $L$-module such that the actions are compatible in the following sense $$\begin{aligned} aX(v)&=&(aX)(v),\\ X(av)&=&X(a)v+(-1)^{\|a\|\|X\|}aX(v) \mbox{ for all homogeneous }X\in L,\:a\in A\mbox{ and }v\in V.\end{aligned}$$ ##### Examples 1. If $A$ is a graded commutative $\mathbbm K$-algebra then $(A,\mathrm{Der}_{\mathbbm K}A)$ is a graded Lie-Rinehart pair. 2. If $A$ is a graded commutative algebra, $\mathfrak g$ a graded $\mathbbm K$-Lie algebra and $\rho:\mathfrak g\to \mathrm{Der}_\mathbbm K A$ a morphism of graded Lie algebras, then $A\otimes_\mathbbm K \mathfrak g$ is a graded Lie-Rinehart pair with bracket given by $$[a\otimes X,b\otimes Y]:=(-1)^{\|b\|\|X\|}ab\otimes [X,Y]+a\rho(X)b \otimes Y+(-1)^{(\|a\|+\|b\|+\|X\|)\|Y\|} (\rho(Y)a)b\otimes X.$$ 3. The above example can be generalized to the notion of a Lie algebroid, which we will use merely in the ordinary (even) manifold setup. By a Lie algebroid we mean the data $(E,[\:,\:],\rho)$, where $E\to M$ is a vector bundle over a manifold $M$, $[\:,\:]$ is a $\mathbbm K$-linear Lie bracket on the space of sections $\Gamma^\infty(M,E)$ of $E$ and $\rho:E\to TM$ is a vector bundle homomorphism such that $[a,fb]=f[a,b]+\rho(a)(f)\:b$ for all $a,b\in \Gamma^\infty(M,E)$ and $f\in\mathcal C^\infty(M)$. It is clear that $(\mathcal C^\infty(M),\Gamma^\infty(M,E))$ is a Lie-Rinehart pair. Associated to the graded Lie-Rinehart pair $(A,L)$ there is an universal enveloping algebra $(U(A,L),i_A,i_L)$, which is analog to the enveloping algebra of a Lie algebra and of the algebra of differential operators on a manifold, respectively. More precisely $U(A,L)$ is a graded $\mathbbm K$-algebra, $i_A:A \to U(A;L)$ is a morphism of graded $\mathbbm K$-algebras and $i_L:L\to U(A,L)$ is a morphism of graded Lie algebras having the following properties: $i_A(a)i_L(X)=i_L(aX)$ and $i_L(X)i_A(a)-(-1)^{\|X\|\|a\|}i_A(a)i_L(X)=i_A(X(a))$ for all homogeneous $a\in A$ and $X\in L$. $(U(A,L),i_A,i_L)$ is initial among the triples $(U,j_A,j_L)$ having these properties. In order to construct it, we follow the approach of Hübschmann [@HuebPoiss]. Let $U(\mathbbm K,L)$ be the universal enveloping algebra of the graded Lie algebra $L$, $i_{\mathbbm K}:\mathbbm K\to U(\mathbbm K,L)$ and $i_L:L\to U(\mathbbm K,L)$ the canonical embeddings and $\Delta:U(\mathbbm K,L)\to U(\mathbbm K,L)\otimes_{\mathbbm K} U(\mathbbm K,L)$ the standard comultiplication. Remember that $\Delta$ is the unique comultiplication such that $U(\mathbbm K,L)$ is a graded bialgebra and the image of $i_L$ is the space of primitives. As an intermediate step, let us define the algebra $$A\odot U(\mathbbm K,L)=(A\otimes_{\mathbbm K} U(\mathbbm K,L),\mu),$$ where the multiplication $\mu$ is given by $$(a\otimes u)(b\otimes v):=(-1)^{\|u\|\|b\|}ab\otimes uv+\sum (-1)^{\|u''\|\|b\|}au'(b)\otimes u''v$$ for homogeneous $a,b \in A$ and $u,v \in U(\mathbbm K,L)$. Here we used Sweedler’s notation $\Delta(u):=\sum u'\otimes u''$. One easily proves that $\mu$ is associative. Let $J$ be the right ideal generated by elements of the form $ab\otimes X-a\otimes b X$ for $a,b \in A$ and $X\in L$ (here we write $X$ and $bX$ for the respective images under $i_L$). As a consequence of the identity $$\begin{aligned} (c\otimes Y)(ab\otimes X-a\otimes b X)&=&(-1)^{\|Y\|(\|a\|+\|b\|)}(cab\otimes XY-c\otimes ab YX)\\ &&\qquad+(-1)^{\|Y\|\|a\|}(caY(b)\otimes X-c\otimes aY(b)X)\end{aligned}$$ for all homogeneous $a,b,c\in A$ and $X,Y\in L$, we have that $J$ is in fact a two sided ideal. The universal enveloping algebra is defined to be the quotient $$U(A,L):=A\odot U(\mathbbm K, L)/J,$$ and $i_A$, $i_L$ are the obvious morphisms. The universal enveloping algebra $U(A,L)$ is in a natural manner a filtered algebra $$U(A,L)=U_0(A,L)\supset U_1(A,L)\supset\dots U_{n-1}(A,L)\supset U_n(A,L)\supset\dots\quad.$$ Here $U_n(A,L)$ is the left $A$-module generated by at most $n$ products of the images of $L$ in $U(A,L)$. Clearly the induced left and right $A$-algebra structures on the associated graded algebra $\mathrm{gr}\,U(A,L)=\oplus_{n\ge 0}\mathrm{gr}_n \,U(A,L)=\oplus_{n\ge 0} U_n(A,L)/U_{n-1}(A,L)$ coincide (here we set $U_{-1}(A,L):=0$). Note that $\mathrm{gr}\,U(A,L)$ is a graded commutative algebra. The best understood Lie-Rinehart pairs are those, for which the $A$-module $L$ is projective. One reason for this is, that there is an analog of the Poincaré-Birkhoff-Witt theorem. The canonical $A$-module epimorphism $S_AL\to\mathrm{gr}\,U(A,L)$ is an isomorphism of graded commutative algebras. Adapt the proof of Rinehart [@Rinehart p.199–200] to the graded situation. \ A morphism of graded Lie-Rinehart pairs $(A,L)\to (A',L')$ is a morphism of graded commutative algebras $A\to A'$, $a\mapsto a'$ and a morphism of graded Lie algebras $L\to L'$, $X\mapsto X'$ such that $(aX)'=a'X'$ and $(X(a))'=X'(a')$ for all $a\in A$ and $X\in L$. A morphism of Lie Rinehart pairs extends uniquely to a ring homorphism of the universal envelopping algebras $U(A,L)\to U(A',L')$. There is also a natural notion of a module for a Lie-Rinehart pair $(A,L)$ (for details see [@Rinehart]). Equivalently, one may think of such a module as a module for the algebra $U(A,L)$. For example, the action of $L$ on $A$ extends natually to an $U(A,L)$-module structure on $A$. It has been shown by Rinehart, that if $L$ is a projective $A$-module, there is a Koszul resolution of the $U(A,L)$-module $A$, which generalizes the Koszul resolution of the ground field for Lie algebras. The space of chains of this complex is $K_\bullet(A,L)=U(A,L)\otimes_A \wedge ^\bullet_A L$. The differential $\partial:K_\bullet(A,L)\to K_{\bullet-1}(A,L)$ is given by the formula $$\begin{aligned} \partial(u\otimes X_1\wedge\cdots\wedge X_n)&=&\sum_{i=1}^n (-1)^{i+1}{\operatorname{sign}}(\tau_i,\|X\|)\:(u\: i_L(X_i))\otimes X_1\wedge \cdots \widehat{X_i}\cdots\wedge X_n\\ &+&\sum_{1\le i<j\le n} (-1)^{i+j}{\operatorname{sign}}(\tau_{i,j},\|X\|)\:u\otimes [X_i,X_j]\wedge X_1\wedge \cdots \widehat{X_i}\cdots\widehat{X_j}\cdots\wedge X_n.\end{aligned}$$ Here, ${\operatorname{sign}}(\tau_i,\|X\|)$ and ${\operatorname{sign}}(\tau_{i,j},\|X\|)$ are the Koszul signs of the permutations $\tau_i=(1\,2\dots i-1\,i)$ and $$\begin{aligned} \tau_{i,j}=\left( \begin{array}{cccccccccccc} 1&2&3&4&\dots& & & & & & &n \\ i&j&1&2&\dots&i-1&i+1&\dots &j-1&j+1&\dots&n \end{array} \right),\end{aligned}$$ respectively, for the multiindex $\|X\|=(\|X_1\|,\dots,\|X_n\|)$. Analog to one can show that $\partial$ is in fact well-defined and that $(K_\bullet(A,L),\partial)$ is a projective resolution of the $U(A,L)$-module $A$. We can therefore use this resolution to compute derived functors such as ${\operatorname{Ext}}_{U(A,L)}(A,M)$ for any $U(A,L)$-module $M$. More precisely, ${\operatorname{Ext}}_{U(A,L)}(A,M)$ is the cohomology of the complex $$\begin{aligned} \label{LieRinehartComplex} {\operatorname{Hom}}_A(\wedge_A^\bullet L,M)={\operatorname{Alt}}^{\bullet}_A(L,M),\end{aligned}$$ where the differential $d:={\operatorname{Alt}}^n(L,M)\to {\operatorname{Alt}}^{n+1}(L,M)$ is defined for homogeneous $f\in {\operatorname{Alt}}^n(L,M)$ by the formula $$\begin{aligned} (df)(X_1,\dots,X_{n+1})&=&\sum_{i=1}^{n+1} (-1)^{i+1+\|f\|\|X_i\|}{\operatorname{sign}}(\tau_i,\|X\|)\:X_i(f(X_1,\dots\widehat{X_i}\dots,X_{n+1}))\\ &&+\sum_{1\le i<j\le n+1}(-1)^{i+j}{\operatorname{sign}}(\tau_{i,j},\|X\|)\:f([X_i,X_j],X_1,\dots\widehat{X_i}\dots\widehat{X_j}\dots,X_{n+1}).\end{aligned}$$ In the case of a Lie algebroid ${\operatorname{Ext}}_{U(A,L)}(A,A)$ is just the ordinary Lie algebroid cohomology. In particular, for $A=\mathcal C^\infty(M)$ the ring of functions on a smooth manifold $M$ and $L=\Gamma^\infty(M,TM)$ this boils down to the ordinary de Rham cohomology. Also, in the $\mathbbm Z_2$-graded setting this construction reproduces the Chevalley-Eilenberg cochain complex of Lie super-algebras (see e.g. [@Scheunert] and references therein) and the de Rham complex for super-manifolds. In the general $\mathbbm Z$-graded case the complex (\[LieRinehartComplex\]) may become rather huge (note that we have to use ${\operatorname{Hom}}_A$ and not ${\operatorname{Hom}}_{{{A}\mathbf{-Mod}^{\mathbbm Z}}}$), and there are perhaps better ways to define Lie-Rinehart cohomology for $\mathbbm Z$-graded Lie-Rinehart pairs. The opposite of a $n$-Poisson algebra is $n$-Poisson {#oppPoiss} ---------------------------------------------------- If $(L,\cdot,[\:,\:])$ is a $n$-Poisson algebra then the opposite bracket $$\begin{aligned} [a,b]^{{\operatorname{opp}}}:=(-1)^{(\|a\|+n)(\|b\|+n)}[b,a]\end{aligned}$$ for homogeneous $a,b \in L$, is a Poisson bracket of degree $n$. We say that $(L,\cdot,[\:,\:]^{{\operatorname{opp}}})$ is the opposite of the $n$-Poisson algebra $(L,\cdot,[\:,\:])$. In order to prove the graded Leibniz rule note that $$\begin{aligned} [a,bc]^{{\operatorname{opp}}}&=&(-1)^{(\|b\|+\|c\|+n)(\|a\|+n)}[bc,a]\\ &=&(-1)^{\|a\|\|b\|+\|a\|\|c\|+n(\|a\|+\|b\|+\|c\|+1)}\left(b[c,a]+(-1)^{(\|a\|+n)}[b,a]c\right)\\ &=&(-1)^{\|a\|\|b\|+\|a\|\|c\|+n(\|a\|+\|b\|+\|c\|+1)}\Big(\underbrace{(-1)^{(\|a\|+n)(\|c\|+n)}}_{=(-1)^{\|a\|\|c\|+n(a\|+\|c\|+1)}} b[a,c]^{{\operatorname{opp}}}+\\ &&\qquad\qquad\qquad\qquad\qquad\qquad\qquad\underbrace{(-1)^{(\|a\|+n)\|c\|}(-1)^{(\|a\|+n)(\|b\|+n)}}_{=(-1)^{\|a\|\|b\|+\|a\|\|c\|+n(\|a\|+\|b\|+\|c\|+1)}}[a,b]^{{\operatorname{opp}}}c\Big)\\ &=&(-1)^{(\|a\|+n)\|b\|}b[a,c]^{{\operatorname{opp}}}+[a,b]^{{\operatorname{opp}}}c.\end{aligned}$$ Similarly we compute: $$\begin{aligned} [a,[b,c]^{{\operatorname{opp}}}]^{{\operatorname{opp}}}&=&(-1)^{(\|b\|+\|c\|)(\|a\|+n)}(-1)^{(\|c\|+n)(\|b\|+n)}[[c,b],a]\\ &=& (-1)^{\|a\|\|b\|+\|b\|\|c\|+\|a\|\|c\|+n}\left([c,[b,a]]+(-1)^{(\|a\|+n)(\|b\|+n)}[[c,a],b]\right)\\ &=& (-1)^{\|a\|\|b\|+\|b\|\|c\|+\|a\|\|c\|+n}\Big((-1)^{(\|a\|+\|b\|)(\|c\|+n)}(-1)^{(\|a\|+n)(\|b\|+n)}[[a,b]^{{\operatorname{opp}}},c]^{{\operatorname{opp}}}\\ &&+(-1)^{\|a\|\|b\|+n(\|a\|+\|b\|+1)}(-1)^{(\|a\|+\|c\|)(\|b\|+n)}(-1)^{(\|a\|+n)(\|c\|+n)}[b,[a,c]^{{\operatorname{opp}}}]^{{\operatorname{opp}}}\Big)\\ &=& (-1)^{\|a\|\|b\|+\|b\|\|c\|+\|a\|\|c\|+n}\Big((-1)^{\|a\|\|c\|+\|b\|\|c\|+n(\|a\|+\|b\|)}(-1)^{\|a\|\|b\|+n(\|a+\|b\|+1)}[[a,b]^{{\operatorname{opp}}},c]^{{\operatorname{opp}}}\\ &&+(-1)^{\|a\|\|b\|+n(\|a\|+\|b\|+1)}(-1)^{\|a\|\|b\|+\|b\|\|c\|+\|a\|\|c\|+n}[b,[a,c]^{{\operatorname{opp}}}]^{{\operatorname{opp}}}\Big)\\ &=&[[a,b]^{{\operatorname{opp}}},c]^{{\operatorname{opp}}}+(-1)^{(\|a\|+n)(\|b\|+n)}[b,[a,c]^{{\operatorname{opp}}}]^{{\operatorname{opp}}}.\end{aligned}$$ Proof of Theorem \[SchoutenNijenhuis\] {#Schoutenproof} -------------------------------------- First of all let us introduce some notation. Let $X=X_1X_2\dots X_r\in S_A(L[-1])$ be a monomial, such that $X_i\in L^{x_i}$ for $i=1,\dots,r$. We will call such monomials homogeneous. The degree $\|X\|$ of $X$ in $S_A(L[-1])$ is given by $x_1+\dots +x_r + r$. Moreover, we will use the shorthand $$\begin{aligned} X_{>i}&:=&X_{i+1}X_{i+2}\dots X_r,\qquad \|X_{>i}\|=x_{i+1}+\dots+ x_r+r-i\\ X^{<i}&:=&X_1X_2\dots X_{i-1},\qquad \|X^{<i}\|=x_1+\dots+x_{i-1}+i-1.\end{aligned}$$ It is not difficult to prove that, as a consequence of the Leibniz rule, the Schouten-Nijenhuis bracket with another such monomial $Y=Y_1Y_2\dots Y_s$ is given by: $$\begin{aligned} \label{Schoutenformel} [X,Y]=\sum_{i=1}^r\sum_{j=1}^s (-1)^{\|X_i\|\|X_{>i}\|+\|Y_j\|\|Y^{<j}\|}X^{<i}X_{>i}[X_i,Y_j]Y^{<j}Y_{>j}\end{aligned}$$ Moreover, it is easy to prove that the result of this formula does not depend of the choice of the decomposition. Note that the sign in this formula appears naturally as the Koszul sign of the permutation which moves $X_i$ to the right of $X$ and $Y_j$ to the left of $Y$. Conversely, the Leibniz rule follows effortlessly from equation (\[Schoutenformel\]). In fact, if $Z=Z_1Z_2\dots Z_t$ is a third homogeneous monomial we have $$\begin{aligned} [X,YZ]&=&[X,Y]Z+\sum_{i=1}^r\sum_{j=1}^t (-1)^{\|X_i\|\|X_{>i}\|+\|Z_j\|(\|Y\|+\|Z^{<j}\|)}X^{<i}X_{>i}[X_i,Z_j]YZ^{<j}Z_{>j}\\ &=&[X,Y]Z+(-1)^{\|Y\|(\|X\|-1)}Y[X,Z].\end{aligned}$$ In order to see that this bracket is graded antisymmetric note that $\|X^{<i}X_{>i}\|=\|X\|-\|X_i\|$ and that the degree of $[X_i,Y_j]=-(-1)^{(\|X_i\|-1)(\|Y_j\|-1)}[Y_j,X_i]$ is $\|X_i\|+\|Y_j\|-1$. Moreover, the Koszul sign of the operation which replaces $X^{<i}X_{>i}[X_i,Y_j]Y^{<j}Y_{>j}$ by $Y^{<j}Y_{>j}[X_i,Y_j]X^{<i}X_{>i}$ is given by $$\begin{aligned} &&(-1)^{(\|X\|-\|X_i\|)(\|X_i\|+\|Y_j\|-1)}(-1)^{(\|X\|-\|X_i\|)(\|Y\|-\|Y_j\|\|)}(-1)^{(\|X_i\|+\|Y_j\|-1)(\|Y\|-\|Y_j\|)}\\ &=&(-1)^{\|X\|\|X_i\|+\|X\|\|Y_j\|+\|X_i\|\|Y_j\|+\|X\|}(-1)^{\|X\|\|Y\|+\|X\|\|Y_j\|+\|X_i\|\|Y\|+\|X_i\|\|Y_j\|}(-1)^{\|X_i\|\|Y\|+\|Y\|\|Y_j\|+\|X_i\|\|Y_j\|+\|Y\|}\\ &=&(-1)^{\|X\|\|Y\|+\|X\|\|X_i\|+\|Y\|\|Y_j\|+\|X_i\|\|Y_j\|+\|X\|+\|Y\|}.\end{aligned}$$ It follows that $$\begin{aligned} [X,Y]&=&-\sum_{i=1}^r\sum_{j=1}^s (-1)^{\|X_i\|\|X_{>i}\|+\|Y_j\|\|Y^{<j}\|}(-1)^{\|X\|\|Y\|+\|X\|\|X_i\|+\|Y\|\|Y_j\|+\|X_i\|\|Y_j\|+\|X\|+\|Y\|}\\ &&\qquad \qquad \qquad \qquad \qquad \qquad (-1)^{(\|X_i\|-1)(\|Y_j\|-1)}\:Y^{<j}Y_{>j}[Y_j,X_i]X^{<i}X_{>i}\end{aligned}$$ and we have to analyse the sign in the above formula. To this end note that $\|X_{>i}\|=\|X\|-\|X^{<i}\|-\|X_i\|$ and $\|Y^{<j}\|=\|Y\|-\|Y_{>j}\|-\|Y_j\|$. Some bookkeeping yields that the sign works out correctly $$\begin{aligned} (-1)^{\|Y_j\|\|Y_{>j}\|+\|X_i\|\|X^{<i}\|}(-1)^{\|X\|\|Y\|+\|X\|+\|Y\|+1},\end{aligned}$$ and we have thus proved $[X,Y]=-(-1)^{(\|X\|-1)(\|Y\|-1)}[Y,X]$. In order to prove the Jacobi identity we introduce some further notation. For $1\le i<j\le r$ and a monomial $X=X_1X_2\dots X_r$ we introduce $$\begin{aligned} X_{>i}^{<j}:=X_{i+1}\dots X_{j-1}\qquad \|X_{>i}^{<j}\|=x_{i+1}+\dots+ x_{j-1}+j-i-1.\end{aligned}$$ Let $Y=Y_1Y_2\dots Y_s\in S^s_A(L[-1])$ and $Z=Z_1Z_2\dots Z_t\in S^t_A(L[-1])$ be two other homogeneous monomials. In order to proof $$\begin{aligned} \label{Jacobiidentitaet} [X,[Y,Z]]=[[X,Y],Z]+(-1)^{(\|X\|-1)(\|Y\|-1)}[Y,[X,Z]]\end{aligned}$$ we expand the three terms in the above equation according to equation (\[Schoutenformel\]), identify the various contributions and ‘mow down’ the signs. $$\begin{aligned} [X,[Y,Z]]&=&\sum_{i=1}^r\sum_{k=1}^s\sum_{j=1}^{k-1}\sum_{l=1}^t(-1)^{n_1} X^{<i}X_{>i}[X_i,Y_j]Y^{<j}Y_{>j}^{<k}Y_{>k}[Y_k,Z_l]Z^{<l}Z_{>l}\\ &&+\sum_{i=1}^r\sum_{k=1}^s\sum_{j=k+1}^{s}\sum_{l=1}^t(-1)^{n_2} X^{<i}X_{>i}[X_i,Y_j]Y^{<k}Y_{>k}^{<j}Y_{>j}[Y_k,Z_l]Z^{<l}Z_{>l}\\ &&+\sum_{i=1}^r\sum_{j=1}^s\sum_{l=1}^t (-1)^{n_3}X^{<i}X_{>i}[X_i,[Y_j,Z_l]]Y^{<j}Y_{>j} Z^{<l}Z_{>l}\\ &&+\sum_{i=1}^r \sum_{k=1}^s\sum_{l=1}^t\sum_{j=1}^{l-1}(-1)^{n_4}X^{<i}X_{>i}[X_i,Z_j]Y^{<k}Y_{>k}[Y_k,Z_l]Z^{<j}Z_{>j}^{<l}Z_{>l}\\ &&+\sum_{i=1}^r\sum_{k=1}^s\sum_{l=1}^t\sum_{j=l+1}^{t}(-1)^{n_5}X^{<i}X_{>i}[X_i,Z_j]Y^{<k}Y_{>k}[Y_k,Z_l]Z^{<l}Z_{>l}^{<j}Z_{>j}\\ &=:&A_1+A_2+A_3+A_4+A_5.\end{aligned}$$ The exponents $n_1,\dots,n_5$ in the above equation will be understood modulo $2$. They are given as follows $$\begin{aligned} n_1&=&\|X_i\|\|X_{>i}\|+\|Y_j\|\|Y^{<j}\|+\|Y_k\|\|Y_{>k}\|+\|Z_l\|\|Z^{<l}\|\\ n_2&=&\|X_i\|\|X_{>i}\|+\|Y_j\|(\|Y^{<j}\|-\|Y_k\|)+\|Y_k\|\|Y_{>k}\|+\|Z_l\|\|Z^{<l}\|\\ n_3&=&\|X_i\|\|X_{>i}\|+(\|Y\|-\|Y_j\|)(\|Y_j\|+\|Z_l\|-1)+\|Y_j\|\|Y_{>j}\|+\|Z_l\|\|Z^{<l}\|\\ n_4&=&\|X_i\|\|X_{>i}\|+\|Z_j\|(\|Z^{<j}\|+\|Y_k\|+\|Z_l\|-1+\|Y\|-\|Y_k\|) +\|Y_k\|\|Y_{>k}\|+\|Z_l\|\|Z^{<l}\|\\ &=&\|X_i\|\|X_{>i}\|+\|Z_j\|(\|Z^{<j}\|+\|Z_l\|+\|Y\|+1)+\|Y_k\|\|Y_{>k}\|+\|Z_l\|\|Z^{<l}\|\\ n_5&=&\|X_i\|\|X_{>i}\|+\|Z_j\|(\|Z^{<j}\|-\|Z_l\|+\|Z_l\|+\|Y\|-1)+\|Y_k\|\|Y_{>k}\|+\|Z_l\|\|Z^{<l}\|\\ &=&\|X_i\|\|X_{>i}\|+\|Z_j\|(\|Z^{<j}\|+\|Y\|+1)+\|Y_k\|\|Y_{>k}\|+\|Z_l\|\|Z^{<l}\|.\end{aligned}$$ Next, we have $$\begin{aligned} [[X,Y],Z]&=&\sum_{i=1}^r\sum_{k=1}^s\sum_{j=1}^{k-1}\sum_{l=1}^t(-1)^{m_1} X^{<i}X_{>i}[X_i,Y_j]Y^{<j}Y_{>j}^{<k}Y_{>k}[Y_k,Z_l]Z^{<l}Z_{>l}\\ &&+\sum_{i=1}^r\sum_{k=1}^s\sum_{j=k+1}^{s}\sum_{l=1}^t(-1)^{m_2} X^{<i}X_{>i}[X_i,Y_j]Y^{<k}Y_{>k}^{<j}Y_{>j}[Y_k,Z_l]Z^{<l}Z_{>l}\\ &&+\sum_{i=1}^r\sum_{j=1}^s\sum_{l=1}^t (-1)^{m_3}X^{<i}X_{>i}Y^{<j}Y_{>j}[[X_i,Y_j],Z_l] Z^{<l}Z_{>l}\\ &&+\sum_{i=1}^r \sum_{j=1}^s\sum_{l=1}^t\sum_{k=i+1}^{r}(-1)^{m_4}X^{<i}X_{>i}^{<k}X_{>k}[X_i,Y_j]Y^{<j}Y_{>j}[X_k,Z_l] Z^{<l}Z_{>l}\\ &&+\sum_{i=1}^r\sum_{j=1}^s\sum_{l=1}^t\sum_{k=1}^{i-1}(-1)^{m_5}X^{<k}X^{<i}_{>k} X_{>i}[X_i,Y_j]Y^{<j}Y_{>j}[X_k,Z_l]Z^{<l}Z_{>l}\\ &=:&B_1+B_2+B_3+B_4+B_5.\end{aligned}$$ The exponents $m_1,\dots,m_5$ in the above equation are given as follows $$\begin{aligned} m_1&=&\|X_i\|\|X_{>i}\|+\|Y_j\|\|Y^{<j}\|+\|Y_k\|\|Y_{>k}\|+\|Z_l\|\|Z^{<l}\|\\ m_2&=&\|X_i\|\|X_{>i}\|+\|Y_j\|\|Y^{<j}\|+\|Y_k\|(\|Y_{>k}\|-\|Y_j\|)+\|Z_l\|\|Z^{<l}\|\\ m_3&=&\|X_i\|\|X_{>i}\|+\|Y_j\|\|Y^{<j}\|+(\|Y\|-\|Y_j\|)(\|X_i\|+\|Y_j\|-1)+\|Z_l\|\|Z^{<l}\|\\ m_4&=&\|X_i\|\|X_{>i}\|+\|Y_j\|\|Y^{<j}\|+\|X_k\|(\|X_{>k}\|+\|Y\|-\|Y_j\|+(\|X_i\|+\|Y_j\|-1)) +\|Z_l\|\|Z^{<l}\|\\ &=&\|X_i\|\|X_{>i}\|+\|Y_j\|\|Y^{<j}\|+\|X_k\|(\|X_{>k}\| +\|Y\|+\|X_i\|+1)+\|Z_l\|\|Z^{<l}\|\\ m_5&=&\|X_i\|\|X_{>i}\|+\|Y_j\|\|Y^{<j}\|+\|X_k\|(\|X_{>k}\|-\|X_i\|+\|Y\|-\|Y_j\|+(\|X_i\|+\|Y_j\|-1)+\|Z_l\|\|Z^{<l}\|\\ &=&\|X_i\|\|X_{>i}\|+\|Y_j\|\|Y^{<j}\|+\|X_k\|(\|X_{>k}\|+\|Y\|+1)+\|Z_l\|\|Z^{<l}\|.\end{aligned}$$ Finally, we have $$\begin{aligned} [Y,[X,Z]]&=&\sum_{j=1}^s\sum_{k=1}^r\sum_{i=1}^{k-1}\sum_{l=1}^t(-1)^{k_1} Y^{<j}Y_{>j}[Y_j,X_i]X^{<i}X_{>i}^{<k}X_{>k}[X_k,Z_l]Z^{<l}Z_{>l}\\ &&+\sum_{j=1}^s\sum_{k=1}^r\sum_{i=k+1}^{s}\sum_{l=1}^t(-1)^{k_2} Y^{<j}Y_{>j}[Y_j,X_i]X^{<k}X_{>k}^{<i}X_{>i}[X_k,Z_l]Z^{<l}Z_{>l}\\ &&+\sum_{j=1}^s\sum_{i=1}^r\sum_{l=1}^t (-1)^{k_3}Y^{<j}Y_{>j}[Y_j,[X_i,Z_l]]X^{<i}X_{>i} Z^{<l}Z_{>l}\\ &&+\sum_{k=1}^s \sum_{i=1}^r\sum_{j=1}^t\sum_{l=1}^{j-1}(-1)^{k_4}Y^{<k}Y_{>k}[Y_k,Z_l]X^{<i}X_{>i}[X_i,Z_j]Z^{<l}Z_{>l}^{<j}Z_{>j}\\ &&+\sum_{i=1}^s\sum_{k=1}^r\sum_{j=1}^t\sum_{l=j+1}^{t}(-1)^{k_5}Y^{<k}Y_{>k}[Y_k,Z_l]X^{<i}X_{>i}[X_i,Z_j]Z^{<j}Z_{>j}^{<l}Z_{>l}\\ &=:&C_1+C_2+C_3+C_4+C_5.\end{aligned}$$ The exponents $k_1,\dots,k_5$ in the above equation are given as follows $$\begin{aligned} k_1&=&\|Y_j\|\|Y_{>j}\|+\|X_i\|\|X^{<i}\|+\|X_k\|\|X_{>k}\|+\|Z_l\|\|Z^{<l}\|\\ k_2&=&\|Y_j\|\|Y_{>j}\|+\|X_i\|(\|X^{<i}\|-\|X_k\|)+\|X_k\|\|X_{>k}\|+\|Z_l\|\|Z^{<l}\|\\ k_3&=&\|Y_j\|\|Y_{>j}\|+(\|X\|-\|X_i\|)(\|X_i\|+\|Z_l\|-1)+\|X_i\|\|X_{>i}\|+\|Z_l\|\|Z^{<l}\|\\ k_4&=&\|Y_k\|\|Y_{>k}\|+\|Z_l\|(\|Z^{<l}\|+\|X_i\|+\|Z_j\|-1+\|X\|-\|X_i\|) +\|X_i\|\|X_{>i}\|+\|Z_j\|\|Z^{<j}\|\\ &=&\|Y_k\|\|Y_{>k}\|+\|Z_l\|(\|Z^{<l}\|+\|Z_j\|+\|X\|+1)+\|X_i\|\|X_{>i}\|+\|Z_j\|\|Z^{<j}\|\\ k_5&=&\|Y_k\|\|Y_{>k}\|+\|Z_l\|(\|Z^{<l}\|-\|Z_j\|+\|Z_j\|+\|X_i\|-1+\|X\|-\|X_i\|)+\|X_i\|\|X_{>i}\|+\|Z_j\|\|Z^{<j}\|\\ &=&\|Y_k\|\|Y_{>k}\|+\|Z_l\|(\|Z^{<l}\|+\|X\|+1)+\|X_i\|\|X_{>i}\|+\|Z_j\|\|Z^{<j}\|.\end{aligned}$$ First of all we note that $$\begin{aligned} \label{A12id} A_1=B_1, \qquad A_2=B_2\end{aligned}$$ and we have to check some similar relations among the $A,B,C$’s. The only one which involves the Jacobi identity for $L$ is $$\begin{aligned} \label{Jac} A_3=B_3+(-1)^{(\|X\|-1)(\|Y\|-1)}C_3,\end{aligned}$$ which we would like to check right now. Let us compare $A_3$ and $B_3$: $$\begin{aligned} A_3&=&=\sum_{i=1}^r\sum_{j=1}^s\sum_{l=1}^t(-1)^{n_3} X^{<i}X_{>i}[X_i,[Y_j,Z_l]]Y^{<j}Y_{>j} Z^{<l}Z_{>l}\\ B_3&=&\sum_{i=1}^r\sum_{j=1}^s\sum_{l=1}^t (-1)^{m_3}X^{<i}X_{>i}Y^{<j}Y_{>j}[[X_i,Y_j],Z_l] Z^{<l}Z_{>l}\\ &=&\sum_{i=1}^r\sum_{j=1}^s\sum_{l=1}^t (-1)^{m_3+l_1}X^{<i}X_{>i}[[X_i,Y_j],Z_l]Y^{<j}Y_{>j}Z^{<l}Z_{>l},\end{aligned}$$ where $l_1$ is the Koszul sign of the permutation involved. Hence we need to check whether $$\begin{aligned} n_3&=&\|X_i\|\|X_{>i}\|+(\|Y\|-\|Y_j\|)(\|Y_j\|+\|Z_l\|-1)+\|Y_j\|\|Y_{>j}\|+\|Z_l\|\|Z^{<l}\|\\ &=&\|X_i\|\|X_{>i}\|+\|Y\|\|Y_j\|+\|Y\|\|Z_l\|+\|Y\|+\|Y_j\|\|Z_l\|+\|Y_j\|\|Y_{>j}\|+\|Z_l\|\|Z^{<l}\|\end{aligned}$$ coincides with $l_1+m_3$ $$\begin{aligned} l_1&=&(\|Y\|-\|Y_j\|)(\|X_i\|+\|Y_j\|+\|Z_l\|-2)\\ &=&\|X_i\|\|Y\|+\|Y\|\|Z_l\|+\|X_i\|\|Y_j\|+\|Y_j\|\|Z_l\|+\|Y_j\|\\ m_3&=&\|X_i\|\|X_{>i}\|+\|Y_j\|\|Y^{<j}\|+(\|Y\|-\|Y_j\|)(\|X_i\|+\|Y_j\|-1)+\|Z_l\|\|Z^{<l}\|\\ &=&\|X_i\|\|X_{>i}\|+\|Y_j\|\|Y^{<j}\|+\|X_i\|\|Y\|+\|X_i\|\|Y_j\|+\|Y\|\|Y_j\|+\|Y\|+\|Z_l\|\|Z^{<l}\|.\end{aligned}$$ By striking out all twice occuring terms, and using the fact that $\|Y_j\|\|Y^{<j}\|+\|Y_j\|\|Y_{>j}\|+\|Y_j\|=0\mod 2$, it follows that $n_3=l_1+m_3\mod 2$. On the other hand, we have $$\begin{aligned} C_3&=&\sum_{j=1}^s\sum_{i=1}^r\sum_{l=1}^t (-1)^{k_3}Y^{<j}Y_{>j}[Y_j,[X_i,Z_l]]X^{<i}X_{>i} Z^{<l}Z_{>l}\\ &=&\sum_{j=1}^s\sum_{i=1}^r\sum_{l=1}^t(-1)^{k_3+l_2}X^{<i}X_{>i}[Y_j,[X_i,Z_l]]Y^{<j}Y_{>j} Z^{<l}Z_{>l}.\end{aligned}$$ Again we have to check whether $k_3+l_2+\|X\|\|Y\|+\|X\|+\|Y\|+\|X_i\|\|Y_j\|+\|X_i\|+\|Y_j\|=n_3\mod 2$. $$\begin{aligned} l_2&=&(\|Y\|-\|Y_j\|+\|X\|-\|X_i\|)(\|X_i\|+\|Y_j\|+\|Z_l\|-2)+(\|Y\|-\|Y_j\|)(\|X\|-\|X_i\|)\\ &=&\|X\|\|Y\|+\|X_i\|\|Y_j\|+\|X\|\|X_i\|+\|Y\|\|Y_j\|+\|X_i\|+\|Y_j\|+(\|X\|+\|Y\|+\|X_i\|+\|Y_j\|)\|Z_l\|\\ k_3&=&\|Y_j\|\|Y_{>j}\|+(\|X\|-\|X_i\|)(\|X_i\|+\|Z_l\|-1)+\|X_i\|\|X_{>i}\|+\|Z_l\|\|Z^{<l}\|\\ &=&\|Y_j\|\|Y_{>j}\|+\|X\|\|X_i\|+\|X\|\|Z_l\|+\|X_i\|\|Z_l\|+\|X\|+\|X_i\|\|X_{>i}\|+\|Z_l\|\|Z^{<l}\|.\end{aligned}$$ The reader may convince himself that the sign works out correctly, and we have thus proved equation (\[Jac\]). The next identity we would like to prove is $$\begin{aligned} \label{A4id} A_4=(-1)^{(\|X\|-1)(\|Y\|-1)}C_5.\end{aligned}$$ To this end we need to compare $$\begin{aligned} A_4&=&\sum_{i=1}^r \sum_{k=1}^s\sum_{l=1}^t\sum_{j=1}^{l-1}(-1)^{n_4}X^{<i}X_{>i}[X_i,Z_j]Y^{<k}Y_{>k}[Y_k,Z_l]Z^{<j}Z_{>j}^{<l}Z_{>l}\qquad\mbox{with}\\ C_5&=&\sum_{i=1}^s\sum_{k=1}^r\sum_{j=1}^t\sum_{l=j+1}^{t}(-1)^{k_5}Y^{<k}Y_{>k}[Y_k,Z_l]X^{<i}X_{>i}[X_i,Z_j]Z^{<j}Z_{>j}^{<l}Z_{>l}\\ &=&\sum_{i=1}^r \sum_{k=1}^s\sum_{l=1}^t\sum_{j=1}^{l-1}(-1)^{k_5+l_3}X^{<i}X_{>i}[X_i,Z_j]Y^{<k}Y_{>k}[Y_k,Z_l]Z^{<j}Z_{>j}^{<l}Z_{>l}.\end{aligned}$$ The exponent $l_3$ works out as follows $$\begin{aligned} l_3&=&(\|X\|-\|X_i\|+\|X_i\|+\|Z_j\|-1)(\|Y\|-\|Y_k\|+\|Y_k\|+\|Z_l\|-1)\\ &=&(\|X\|+\|Z_j\|-1)(\|Y\|+\|Z_l\|-1)\\ &=&\|X\|\|Y\|+\|X\|\|Z_l\|+ \|Y\|\|Z_j\|+ \|Z_j\|\|Z_l\|+\|Z_j\|+\|Z_l\| +\|X\|+\|Y\|+1.\end{aligned}$$ The reader may convince himself that $n_4=k_5+l_3+(\|X\|-1)(\|Y\|-1)\mod 2$ $$\begin{aligned} n_4&=&\|X_i\|\|X_{>i}\|+\|Z_j\|(\|Z^{<j}\|+\|Z_l\|+\|Y\|+1)+\|Y_k\|\|Y_{>k}\|+\|Z_l\|\|Z^{<l}\|\\ k_5&=&\|Y_k\|\|Y_{>k}\|+\|Z_l\|(\|Z^{<l}\|+\|X\|+1)+\|X_i\|\|X_{>i}\|+\|Z_j\|\|Z^{<j}\|\end{aligned}$$ and we have proved identity (\[A4id\]). Next we claim $$\begin{aligned} \label{A5id} A_5=(-1)^{(\|X\|-1)(\|Y\|-1)}C_4,\end{aligned}$$ that is, we have to compare $$\begin{aligned} A_5&=&\sum_{i=1}^r\sum_{k=1}^s\sum_{l=1}^t\sum_{j=l+1}^{t}(-1)^{n_5}X^{<i}X_{>i}[X_i,Z_j]Y^{<k}Y_{>k}[Y_k,Z_l]Z^{<l}Z_{>l}^{<j}Z_{>j}\qquad\mbox{and}\\ C_4&=&\sum_{k=1}^s \sum_{i=1}^r\sum_{j=1}^t\sum_{l=1}^{j-1}(-1)^{k_4}Y^{<k}Y_{>k}[Y_k,Z_l]X^{<i}X_{>i}[X_i,Z_j]Z^{<l}Z_{>l}^{<j}Z_{>j}\\ &=&\sum_{i=1}^r\sum_{k=1}^s\sum_{l=1}^t\sum_{j=l+1}^{t}(-1)^{k_4+l_4}X^{<i}X_{>i}[X_i,Z_j]Y^{<k}Y_{>k}[Y_k,Z_l]Z^{<l}Z_{>l}^{<j}Z_{>j}.\end{aligned}$$ Here, $l_4$ coincides with $l_3$. The reader may convince himself that $n_5=k_4+l_3+(\|X\|-1)(\|Y\|-1)\mod 2$ usiing $$\begin{aligned} n_5&=&\|X_i\|\|X_{>i}\|+\|Z_j\|(\|Z^{<j}\|+\|Y\|+1)+\|Y_k\|\|Y_{>k}\|+\|Z_l\|\|Z^{<l}\|\\ k_4&=&\|Y_k\|\|Y_{>k}\|+\|Z_l\|(\|Z^{<l}\|+\|Z_j\|+\|X\|+1) +\|X_i\|\|X_{>i}\|+\|Z_j\|\|Z^{<j}\|.\end{aligned}$$ Now let us check $$\begin{aligned} \label{B4id} B_4=-(-1)^{(\|X\|-1)(\|Y\|-1)}C_1.\end{aligned}$$ We have to compare $$\begin{aligned} B_4&=&\sum_{i=1}^r \sum_{j=1}^s\sum_{l=1}^t\sum_{k=i+1}^{r}(-1)^{m_4}X^{<i}X_{>i}^{<k}X_{>k}[X_i,Y_j]Y^{<j}Y_{>j}[X_k,Z_l] Z^{<l}Z_{>l}\qquad\mbox{with}\\ C_1&=&\sum_{j=1}^s\sum_{k=1}^r\sum_{i=1}^{k-1}\sum_{l=1}^t(-1)^{k_1} Y^{<j}Y_{>j}[Y_j,X_i]X^{<i}X_{>i}^{<k}X_{>k}[X_k,Z_l]Z^{<l}Z_{>l}\\ &=&\sum_{i=1}^r \sum_{j=1}^s\sum_{l=1}^t\sum_{k=i+1}^{r}(-1)^{k_1+l_5}X^{<i}X_{>i}^{<k}X_{>k}[X_i,Y_j]Y^{<j}Y_{>j}[X_k,Z_l] Z^{<l}Z_{>l},\end{aligned}$$ where $$\begin{aligned} l_5&=&(\|X_i\|-1)(\|Y_j\|-1)+1+(\|X\|-\|X_i\|-\|X_k\|)(\|Y\|-\|Y_j\|+\|X_i\|+\|Y_j\|-1)\\ &&\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad+(\|Y\|-\|Y_j\|)(\|X_i\|+\|Y_j\|-1)\\ &=&\|X\|\|Y\|+\|X\|+\|Y\|+\|X\|\|X_i\|+\|Y\|\|Y_j\|+\|X_i\|\|X_k\|+\|X_k\|\|Y\|+\|X_k\|+\|X_i\|+\|Y_j\|\end{aligned}$$ Using the identity $\|X_i\|(\|X\|+\|X_{>i}\|+\|X^{<i}\|+1)=0=\|Y_j\|(\|Y\|+\|Y_{>j}\|+\|Y^{<j}\|+1)\mod 2$, the reader may check that $m_4=k_1+l_5+\|X\|\|Y\|+\|X\|+\|Y\|\mod 2$ using $$\begin{aligned} m_4&=&\|X_i\|\|X_{>i}\|+\|Y_j\|\|Y^{<j}\|+\|X_k\|(\|X_{>k}\|+\|Y\|+\|X_i\|+1) +\|Z_l\|\|Z^{<l}\|\\ k_1&=&\|Y_j\|\|Y_{>j}\|+\|X_i\|\|X^{<i}\|+\|X_k\|\|X_{>k}\|+\|Z_l\|\|Z^{<l}\|.\end{aligned}$$ and we have proved equation (\[B4id\]). Finally, let us check $$\begin{aligned} \label{B5id} B_5=-(-1)^{(\|X\|-1)(\|Y\|-1)}C_2.\end{aligned}$$ To this end we need to compare $$\begin{aligned} B_5&=&\sum_{i=1}^r\sum_{j=1}^s\sum_{l=1}^t\sum_{k=1}^{i-1}(-1)^{m_5}X^{<k}X^{<i}_{>k} X_{>i}[X_i,Y_j]Y^{<j}Y_{>j}[X_k,Z_l]Z^{<l}Z_{>l}\qquad\mbox{and}\\ C_2&=&\sum_{j=1}^s\sum_{k=1}^r\sum_{i=k+1}^{s}\sum_{l=1}^t(-1)^{k_2} Y^{<j}Y_{>j}[Y_j,X_i]X^{<k}X_{>k}^{<i}X_{>i}[X_k,Z_l]Z^{<l}Z_{>l}\\ &=&\sum_{i=1}^r\sum_{j=1}^s\sum_{l=1}^t\sum_{k=1}^{i-1}(-1)^{k_2+l_6}X^{<k}X^{<i}_{>k} X_{>i}[X_i,Y_j]Y^{<j}Y_{>j}[X_k,Z_l]Z^{<l}Z_{>l}.\end{aligned}$$ Again, $l_6$ coincides with $l_5$. The reader may check that $m_5=k_2+l_5+\|X\|\|Y\|+\|X\|+\|Y\|\mod 2$ using $$\begin{aligned} m_5&=&\|X_i\|\|X_{>i}\|+\|Y_j\|\|Y^{<j}\|+\|X_k\|(\|X_{>k}\|+\|Y\|+1)+\|Z_l\|\|Z^{<l}\|\\ k_2&=&\|Y_j\|\|Y_{>j}\|+\|X_i\|(\|X^{<i}\|-\|X_k\|)+\|X_k\|\|X_{>k}\|+\|Z_l\|\|Z^{<l}\|\end{aligned}$$ The Jacoibi identity (\[Jacobiidentitaet\]) now follows from the identities (\[A12id\])–(\[B5id\]). [^1]: The number of publications related to geometric quantization is presumably already of order $10^3$. [^2]: According to the citation index of the AMS there are over 500 publications related to deformation quantization (MSC 53D55) since 1998. [^3]: The guess $10^4$ is not too far-fetched. [^4]: Here, the dimension of a variety is the dimension of the smooth part. In general, we have $\dim_{\mathbbm R} (W\cap \mathbbm R^m)\le \dim_{\mathbbm C}(W)$. [^5]: The physicist’s denotation is in this case also mnemonically the simplest. [^6]: We have been hinted by L. Avramov and S. Iyengar that this also follows from the faithful flatness of $\mathcal C^\omega_x(M)\to\mathcal C^\infty(M)$ [@Matsumura Theorem 7.5]. [^7]: We have been hinted by L. Avramov and S. Iyengar that using the theorem of Vasconcelos [@Matsumura Theorem 19.9 and the remark that follows] we are done at this point. [^8]: Let us assume for convenience that it is finitely generated. [^9]: In this case all projective modules are free.	{ "pile_set_name": "ArXiv" }
--- author: - 'H. Abramowicz' - 'A. Abusleme' - 'K. Afanaciev' - 'Y. Benhammou' - 'O. Borysov' - 'M. Borysova' - 'I. Bozovic-Jelisavcic' - 'W. Daniluk' - 'D. Dannheim' - 'M. Demichev' - 'K. Elsener' - 'M. Firlej' - 'E. Firu' - 'T. Fiutowski' - 'V. Ghenescu' - 'M. Gostkin' - 'M. Hempel' - 'H. Henschel' - 'M. Idzik' - 'A. Ignatenko' - 'A. Ishikawa' - 'A. Joffe' - 'G. Kacarevic' - 'S. Kananov' - 'O. Karacheban' - 'W. Klempt' - 'S. Kotov' - 'J. Kotula' - 'U. Kruchonak' - 'Sz. Kulis' - 'W. Lange' - 'J. Leonard' - 'T. Lesiak' - 'A. Levy' - 'I. Levy' - 'L. Linssen' - 'W. Lohmann' - 'J. Moron' - 'A. Moszczynski' - 'A.T. Neagu' - 'B. Pawlik' - 'T. Preda' - 'A. Sailer' - 'B. Schumm' - 'S. Schuwalow' - 'E. Sicking' - 'K. Swientek' - 'B. Turbiarz' - 'N. Vukasinovic' - 'T. Wojton' - 'H. Yamamoto' - 'L. Zawiejski' - 'I.S. Zgura' - 'A. Zhemchugov' title: Performance and Molière radius measurements using a compact prototype of LumiCal in an electron test beam --- [example.eps]{} gsave newpath 20 20 moveto 20 220 lineto 220 220 lineto 220 20 lineto closepath 2 setlinewidth gsave .4 setgray fill grestore stroke grestore Introduction {#Introduction} ============ Forward calorimeters for future electron positron linear collider experiments have challenging requirements on a fast and high precision measurement of the luminosity [@FCAL_ILC], resulting in a stringent set of specifications for highly compact calorimeters. Two such calorimeters, LumiCal and BeamCal, are being considered for installation in the forward region of both International Linear Collider(ILC) [@ILC_TDR_v4_det; @ILC_TDR_v1_phys] detectors, ILD and SiD, and also in the Compact Linear Collider (CLIC) detector [@CLIC_UPDATE_YP]. The precise measurement of the integrated luminosity is provided by the LumiCal detector. BeamCal is designed for instant luminosity measurement and beam-tuning when included in a fast feedback system as well as for tagging beam particles scattered through low angles. Both detectors extend the capabilities of the experiments for physics studies in the high rapidity region. ![The very forward region of the ILD detector. LumiCal, BeamCal and LHCAL are carried by the support tube for the final focusing quadrupole and the beam-pipe. TPC, ECAL and HCAL are the Time Projection Chamber and the Electromagnetic and Hadron Calorimeters. []{data-label="fig_forward_ild"}](forward_region_new.png){width="\textwidth"} The layout of one arm of the forward region of the ILD detector is presented in Figure \[fig\_forward\_ild\]. LumiCal is positioned in a circular hole of the end-cap electromagnetic calorimeter ECAL. BeamCal is placed just in front of the final focus quadrupole. LumiCal is designed as a sampling calorimeter composed of 30 layers of 3.5 mm (1X$_0$) thick tungsten absorbers and silicon sensors placed in a one-millimeter gap between absorber plates. BeamCal has a similar design as LumiCal. For the current BeamCal baseline design, GaAs sensors are considered which can withstand higher radiation doses at room temperature. The similarity between LumiCal and BeamCal designs implies that the technology developed for one can be used also for the other. Luminosity in LumiCal is measured using Bhabha scattering, e$^{+}$e$^{-}\rightarrow$ e$^{+}$e$^{-}$($\gamma$), as a gauge process. The Bhabha scattering cross section can be precisely calculated in QED [@Bhabha_scatt] and the luminosity, $\sf{L}$, is obtained as $${ {\sf{L}} = \frac{{N}_{\mathrm{B}}}{\sigma_{\mathrm{B}}}, }\label{lumiDefEQ}$$ where $N_{\mathrm{B}}$ is the number of Bhabha events registered by LumiCal in a given range of polar angles ($\theta_{min}$, $\theta_{max}$) and $\sigma_{\mathrm{B}}$ is the integral of the differential cross section over the same range. This range defines the fiducial volume of the calorimeter. The fiducial volume for the LumiCal baseline design was studied in simulations [@FCAL_ILC] and found to be in the range from 41 to 67 mrad while the geometrical coverage of the LumiCal ranges from 31 to 77 mrad. The fiducial volume is reduced due to the lateral energy leakage which depends on the electromagnetic shower development in the transverse plane. The compact design of the LumiCal with small gaps between absorber plates allows the transverse size of the shower to be kept small and to achieve in a relatively small $\theta$ angle range a sufficiently large fiducial volume for a precise luminosity measurement. It also improves the efficiency to detect electrons on top of a widely-spread background originating from beamstrahlung and two-photon processes. In addition, the compact construction of LumiCal and BeamCal are essential to match the strict geometrical constraints imposed by the design of the detectors and accelerator needs near the interaction point. In an earlier test beam of a four-layer silicon-tungsten prototype of the LumiCal, an effective Molière radius[^1] of 24.0 $\pm$ 0.6 mm was measured [@LumiCal_multilayer_tb2014_epjc]. The reason for this large value was a large air gap between the silicon sensor plane and the absorber plates because space was needed for a 3.5 mm thick readout board. In order to get a smaller Molière radius, it was essential to design, build and use planes of sub-millimetre thickness to be inserted in a mechanical frame [@mech_frame] in one millimetre gaps between the tungsten absorber plates. This paper describes the design and construction of a compact LumiCal prototype calorimeter, hereafter referred to as calorimeter, and the results from test-beam measurements carried out at DESY, using an electron beam between 1 - 5 GeV energy. For the readout electronics, APV25 front-end boards [@APV_ieee; @APV_nima; @APV_Hybrid] were used. The effective Molière radius of this compact configuration was calculated in a similar way to that in Ref. [@LumiCal_multilayer_tb2014_epjc]. The energy dependence of the effective Molière radius in the energy range of 1 - 5 GeV is also measured. In addition, two sensor planes were put in front of the calorimeter to serve as tracker planes to distinguish between electrons and photons. The results of this latter study will be presented elsewhere. Thin detector plane construction {#LumiCal_Thin_subsection} ================================ The design of a LumiCal sensor was optimised in simulations to provide the required resolution of the polar angle reconstruction. A picture of a sensor is shown in Figure \[fig\_si\_sensor\]. The sensor is made of a 320 $\mu$m thick high resistivity n-type silicon wafer. It has the shape of a sector of a 30$^{\circ}$ angle, with inner and outer radii of the sensitive area of 80 mm and 195.2 mm, respectively. It comprises four sectors with 64 p-type pads of 1.8 mm pitch. ![Detector plane assembly. The thickness of adhesive layers (not shown) between components is within 10 - 15 $\mu$m. The total thickness is 650 $\mu$m.[]{data-label="fig_ThinLCAssembly"}](fig_si_sensor.pdf){width="\textwidth"} ![Detector plane assembly. The thickness of adhesive layers (not shown) between components is within 10 - 15 $\mu$m. The total thickness is 650 $\mu$m.[]{data-label="fig_ThinLCAssembly"}](lc_assembly_1.png){width="0.99\columnwidth"} The properties of the sensor were studied in the lab and beam tests. Results of beam tests and more details about the sensor can be found in Ref. [@TB2010_jinst; @LumiCal_multilayer_tb2014_epjc]. The first prototype of a LumiCal detector plane, which has been successfully used in a multi-layer configuration [@LumiCal_multilayer_tb2014_epjc], had a thickness of about 4 mm and only 32 pads were connected to the readout electronics. For the construction of a sub-millimetre detector plane we used the same silicon sensor. The bias voltage is supplied to the n-side of the sensor by a 70 $\mu$m flexible Kapton-copper foil, glued to the sensor with a conductive glue. The 256 pads of the sensor are connected to the front-end electronics using a fan-out made of 120 $\mu$m thick flexible Kapton foil with copper traces. The inner guard ring is grounded. Ultrasonic wire bonding was used to connect conductive traces on the fan-out to the sensor pads. A support structure, made of carbon fibre composite with a thickness of 100 $\mu$m in the sensor-gluing area, provides mechanical stability for the detector plane. Special fixtures were designed and produced to ensure the necessary thickness and uniformity of three glue layers between different components of the detector plane all over the area of the sensor. A sketch of the structure of the detector plane is shown in Figure \[fig\_ThinLCAssembly\] and a photo of a completed plane in Figure \[module\_photo\]. Since the multi-channel version of the dedicated front-end electronics is still under development, the APV25 front-end board [@APV_ieee; @APV_nima], used by the silicon strip detector of the CMS experiment, was chosen as a temporary solution. It has 128 channels, hence two boards read the whole sensor. The ultrasonic wire bonding proved to provide good electrical performance, but for a detector plane thinner than 1 mm, the wire loops, which are typically 100$-$200 $\mu$m high, cause a serious problem when the plane needs to be installed in a 1 mm gap between absorber plates. The parameters of the bonding machine were studied and tuned to make the loop as low as possible and technically acceptable. The sampling based measurements, which were done using a con-focal laser scanning microscope, show that the loop height is in the range from 50 $\mu$m to 100 $\mu$m. ![A thin detector plane. The black part is the carbon fibre support. The silicon sensor is covered by the Kapton fan-out which has two connectors for front-end boards.[]{data-label="module_photo"}](module_photo.png){width="50.00000%"} Beam Test Setup {#TB_setup} =============== The detector planes were installed in the 1-mm gap between the tungsten absorber layers. Each tungsten absorber layer is on average 3.5 mm thick and roughly one radiation length ($1~X_0$). As described in Figure \[tb2016\], the first calorimeter sensor layer was placed after 3 absorber layers, and the rest followed after each additional absorber layer. The last sensor layer was placed after 8 absorber layers with a total thickness of $7.7~X_0$, since, as noted in [@LumiCal_multilayer_tb2014_epjc], the absorber layers are not pure tungsten. The detector planes were tested in two beam test campaigns in 2015 and 2016 at the DESY-II Synchrotron using electrons with energies between 1 GeV and 5 GeV. The beam test aimed to study the performance of the compact calorimeter and to test the concept of tracking detectors in front of the calorimeter as a tool for electron and photon identification. The geometry of the setup is shown in Figure \[tb2016\]. ![Geometry of the beam test setup (not to scale). Sc1, Sc2 and Sc3 are scintillator counters; T1 and T2 the arms of three-pixel detector planes, Tg the copper target for bremsstrahlung photon production and LumiCal, the calorimeter prototype under test. Distances, rounded to integer numbers in centimetres, are shown in the upper part of the figure.[]{data-label="tb2016"}](tb2016_tab_dim.pdf){width="90.00000%"} The electron beam passed through a $5\times 5$ mm$^2$ square collimator that limits the beam spread along the test setup. The AIDA/EUDET beam telescope was placed upstream of the calorimeter. The telescope was split into two parts T1 and T2, each containing an arm with 3 layers of MIMOSA-26 pixel silicon detectors and 2 thin scintillator counters Sc1 and Sc2, for the trigger system. The telescope front arm was placed before the dipole magnet to record the incoming electrons. The rear arm was placed after the dipole magnet to record the electrons in the direction of the calorimeter, and to separate them from the photons generated in the copper target that was mounted just in front of the magnet. The calorimeter and tracker were assembled in a mechanical frame [@mech_frame] specially designed to provide high precision positioning of the sensor planes and absorber plates. The sensor planes are attached to the tungsten absorber plates by adhesive tape. The tungsten plates are glued to permaglass inserted into the comb slots of the mechanical structure. The assembly of the calorimeter is illustrated in Figure \[tb2016\_LumiCal\]. Two sub-millimetre planes, viewed separately in the upper part of Figure \[tb2016\_LumiCal\], are the tracker planes denoted as “Tracker” in Figure \[tb2016\]. They are installed in front of the calorimeter. ![Top view of the assembled calorimeter.[]{data-label="tb2016_LumiCal"}](tb2016_LumiCal.png){width="50.00000%"} The last module in the LumiCal stack shown in grey is assembled using the tape automatic bonding (TAB) technology [@TAB_Alice]. This sensor plane was not used in the present analysis. All detector planes, for both the calorimeter and the tracker, were powered with a reverse bias voltage of $120~V$. This bias voltage is about 2 to 3 times the depletion voltage [@TB2010_jinst], but well below the breaking voltage of these silicon sensors. Data Acquisition ================ A sketch of the Data Acquisition System (DAQ) is shown in Figure \[fig\_tb2016\_daq\]. It comprises two interdependent systems. The first one is the EUDAQ which controls the beam Telescope and the Trigger Logic Unit, TLU. The second, the calorimeter DAQ, is based on the Scalable Readout System (SRS) [@SRS], developed by the RD51 collaboration, and described below. A trigger signal is generated in the TLU, as a coincidence of signals from the scintillator counters Sc1 and Sc2, both consisting of two thin scintillators with attached photomultipliers. The TLU then sends the trigger signal to both the Telescope acquisition and to the SRS. In addition, a BUSY signal is provided by a NIM logic to prevent the TLU from sending more signals before the event acquisition ended. The SRS, with a front-end hybrid board [@APV_Hybrid] based on the APV25 front-end chip, is used for the readout. The APV25 front-end board has 128 readout channels, each consisting of a charge sensitive preamplifier and a shaper with a CR-RC filter producing a 50 ns shaped voltage pulse [@APV_ieee; @APV_nima]. The output of the shaper is sampled at 40 MHz and stored in an analog pipeline. During the beam test, the APV25 front-end boards are configured to operate in multi-mode, transmitting, upon receipt of a trigger from the TLU, 21 consecutive pipeline samples of each channel to the adapter board of the SRS through 3-m long HDMI cables. These samples are converted to 12-bit numbers in the SRS adapter board and transmitted to the data acquisition PC. Simulation results for the present configuration show that a single pad in a shower can be hit by 80 relativistic particles, hereafter referred to as MIPs \[fig\_si\_sensor\](see Figure \[fig\_pad\_energy\_depos\] in Section \[Calibration\_APV\_25\]). The usage of the APV25 front-end board, which has a dynamic range for energy depositions originating from up to 8 MIPs, is hence not appropriate to read out sensor pads inside an electromagnetic shower. In order to enable measurements of a wider range of deposited energies, a capacitive charge divider is connected to the input of the APV25 front-end board. The attenuation factor of the charge divider is optimised by using the results from MC simulation. However, small signals from pads with low energy depositions in the tails of the shower are then below the detection threshold. The simulation of the observed noise level and the geometry of the present calorimeter shows that an attenuation of the signal with a factor of 3.5 – 4.5, results in a 5 – 7% loss of the deposited energy, which can be corrected for as described in Section \[Calibration\_APV\_25\]. ![The Data Acquisition System.[]{data-label="fig_tb2016_daq"}](2016-daq_v2.pdf){width="80.00000%"} Signal Processing {#Signal_processing} ================= The APV25 front-end chip operating in the multi-mode provides readout of 21 consecutive pipeline samples. The baseline of the output for each channel is calculated as the average of these samples in a dedicated pedestal run with a random trigger without beam. The noise is estimated as the standard deviation of the samples in the pedestal run and is used for setting the threshold in data during the run with a beam. An example of the signal for a single channel, after baseline and common-mode noise subtraction, is shown in Figure \[fig\_raw\_signal\]. During data taking, the average of 21 samples of each channel is calculated and compared to the zero suppression (ZS) threshold. If it is below the ZS threshold, the data for the channel is not recorded. The threshold is set to 0.4 times the channel noise which results in a low enough threshold not to reject the signal from particles. Data is collected asynchronously, i.e. the readout electronics is not synchronised with the accelerator clock. As a consequence, most of the time the signal is not sampled exactly at its maximum. To determine the signal maximum, the samples are fitted with a CR-RC filter response function, as shown in Figure \[fig\_raw\_signal\], $${ S\left(t\right) = A\frac{t-t_{0}}{\tau} e^{-\frac{t-t_{0}}{\tau}} \Theta\left(t-t_{0}\right) , }\label{eq_signal_t0_tau}$$ where $t_{0}$ is the arrival time of the signal, $\tau$ = 50 ns is the peaking time of the APV25 front-end board and $A$ is the relative signal amplitude. The function $\Theta\left(t-t_{0}\right)$ is the Heaviside step function. The relatively low ZS threshold allows a significant amount of noise pulses to pass through and further signal selection criteria are applied in the analysis. First, an artificial neural network (ANN) is used to analyse the signal and classify the data based on its shape. The ANN is represented by multilayer perceptron model with 21 inputs fed from the APV25 samples and one hidden layer with 10 nodes. The training set for different signal amplitudes is generated using the function in eqn.(\[eq\_signal\_t0\_tau\]) with a Gaussian noise added to each sample. After signal preselection based on the ANN, the signal is fitted with eqn.(\[eq\_signal\_t0\_tau\]), where the amplitude, arrival time $t_{0}$ and peaking time $\tau$ are used as parameters. To further improve the purity of the signal, selection criteria are applied to the parameters $t_{0}$ and $\tau$. The efficiency of the selection is studied using external pulses, as described in Section \[Calibration\_APV\_25\]. ![Signal distribution in a single pad of the tracking layer. Green line – after zero suppression, blue line – after additional selection criteria, and red line – fit with a convolution of Landau and Gaussian distribution functions.[]{data-label="fig_tracker_en_distrib"}](raw_event_channel.pdf){width="\textwidth"} ![Signal distribution in a single pad of the tracking layer. Green line – after zero suppression, blue line – after additional selection criteria, and red line – fit with a convolution of Landau and Gaussian distribution functions.[]{data-label="fig_tracker_en_distrib"}](mip_pad_energy_distrib.pdf){width="\textwidth"} Figure \[fig\_tracker\_en\_distrib\] shows the distribution of the signal amplitudes produced by a 5 GeV electron beam and measured in a single channel of the tracking plane. The green line corresponds to the data which pass the ZS threshold. The blue line, which corresponds to the data after applying additional signal selection criteria, illustrates the effective noise suppression in the analysis. The most probable value (MPV) of the peak is estimated using a fit with a convolution of Landau and Gaussian distribution functions. The width $\sigma$ of the Gaussian distribution is considered as noise measurement. The MPV values of the amplitude distribution corresponding to 5 GeV electrons are shown in Figure \[fig\_mpv\_all\]. The higher values for small pad numbers reflect the geometry of the sensor where these pads have smaller area and smaller capacitance. The same effect is observed for the signal-to-noise ratio shown in Figure \[fig\_sn\_all\]. Since the beam profile has blurry edges, the statistical uncertainties increase for pads that correspond to the periphery of the beam. For most of the channels the signal-to-noise ratio is within a range from 7 to 10. The most probable value of the energy deposited by 5 GeV electrons is used to define the unit MIP for the energy deposition in the sensors. Based on MC simulations, a MIP corresponds to 88.5 keV. For the detector planes that are installed in the calorimeter, the capacitive charge divider is used. The signals from single particles are too small to be registered, and hence the signal-to-noise ratio cannot be measured. Taking into account the design of the charge divider and the noise measured in the pedestal run, as shown in Figure \[fig\_noise\_track\_calo\], the estimated value of signal-to-noise ratio is in the range of 2 – 3. For such a low ratio, the signal-shape analysis, using ANN and selection-criteria for the parameters retrieved from the fit, allows for the efficient identification of the signal with little contamination from noise. ![Signal to noise ratio for the pads of the tracking layers covered by the electron beam of 5 GeV.[]{data-label="fig_sn_all"}](mpv.pdf){width="\textwidth"} ![Signal to noise ratio for the pads of the tracking layers covered by the electron beam of 5 GeV.[]{data-label="fig_sn_all"}](sn.pdf){width="\textwidth"} ![Noise distribution in the channels of the tracking layers (blue) and the calorimeter layers (green).[]{data-label="fig_noise_track_calo"}](noise_tracker_calorimeter.pdf){width="48.00000%"} Calibration of the APV25 front-end board {#Calibration_APV_25} ======================================== The linearity of the APV25 front-end boards was studied with the bare chip [@APV_ieee; @APV_nima] and it was found to be very good for signals of up to 3 MIPs and remains better than 5% up to 5 MIPs. The relative response of the APV25 channels, equipped with a capacitive charge divider, is measured using a voltage pulse supplied to the channel input through a capacitor of 2 pF. The detector capacitance is simulated by a 7 pF capacitor connected in parallel to the channel input. About 10 randomly chosen channels for each APV25 front-end chip were measured and the average response curve was calculated for each APV25. ![Efficiency of signal identification as a function of the signal amplitude. Green triangles are measured for different channels, the red line is an average of the fit using eqn. (\[eq\_efficiency\]) to a large number of channels, and the shaded area corresponds to the spread of fits at small amplitudes.[]{data-label="fig_apv_efficiency"}](pad_energy2){width="\textwidth"} ![Efficiency of signal identification as a function of the signal amplitude. Green triangles are measured for different channels, the red line is an average of the fit using eqn. (\[eq\_efficiency\]) to a large number of channels, and the shaded area corresponds to the spread of fits at small amplitudes.[]{data-label="fig_apv_efficiency"}](efficiency_mip.pdf){width="\textwidth"} The APV25 front-end board with charge divider approaches saturation at about 1600 ADC counts. In this analysis, the maximum signal size is 1450 ADC counts, reasonably below the saturation. Figure \[fig\_pad\_energy\_depos\] shows the distribution of the deposited energy in a pad in the detector layer after 5 tungsten plates. The data were processed with the calibration obtained by interpolation between measured values. The sharp spikes are due to saturation which, after calibration, has slightly different thresholds for each APV25. The measured distribution of the deposited energy in a single pad is well reproduced by the simulations for signal amplitudes larger than 5 MIPs. However, smaller signals become masked by the noise. This loss of signals can also be seen in Figure \[fig\_pad\_energy\_depos\], where for small amplitudes the experimental distribution is below the MC expectation. In order to correct for this loss of signals, the efficiency $\epsilon$ of identifying the signal of a small amplitude is studied with the same setup using an external voltage pulse. We define the efficiency of signal identification as the ratio of the number of identified signals to the number of generated ones. This ratio depends on the signal-to-noise ratio and therefore is slightly different for different APV25 front-end chips, as shown in Figure \[fig\_apv\_efficiency\] where the results for channels of different APV25 front-end chips are presented. For each APV25 front-end chip, about 10 channels are measured. For signals larger than 10 MIPs, the efficiency is 100% in all channels. For a smaller number of MIPs, some channels give lower efficiencies. The measurements of the efficiency $\epsilon$ are fit by the following expression: $${ \epsilon = p_{0}\left(1+\mbox{\textit{erf}}\left( \frac{S-S_{0}}{p_{1}} \right) \right) }\label{eq_efficiency}$$ where is the error function, $S$ the signal amplitude and $p_{0}$, $p_{1}$ and $S_{0}$ are fit parameters. The red curve in Figure \[fig\_apv\_efficiency\] represents the average of the fit of a large number of channels and the shaded area the spread of the fit in these channels at low signal amplitudes. Since the noise level observed during lab calibration measurements and beam test are similar, the efficiency correction for small signal sizes is applied to the test-beam simulations using the results of the fit to eqn. (\[eq\_efficiency\]). Results ======= More than seven million events were collected in an electron beam from 1 GeV to 5 GeV energy, with 1 GeV steps, for different setup configurations to measure the precision of the shower position determination, the electromagnetic shower development in longitudinal and transverse directions and the effective Molière radius. ![Average deposited energy in the calorimeter, $E_{dep}$, as a function of beam energy before (red) and after applying the APV25 calibration and corrections to the leakage fraction estimated from the simulation (blue). The lines are straight line fits to the data. The lower part of the figure shows the ratio of the deposited energy to the straight line fit.[]{data-label="Avarage_E_dep_total_1_5gev"}](Edep_1-5_GeV.pdf){width="\textwidth"} ![Average deposited energy in the calorimeter, $E_{dep}$, as a function of beam energy before (red) and after applying the APV25 calibration and corrections to the leakage fraction estimated from the simulation (blue). The lines are straight line fits to the data. The lower part of the figure shows the ratio of the deposited energy to the straight line fit.[]{data-label="Avarage_E_dep_total_1_5gev"}](EdepEinc_corr_ratio.pdf){width="\textwidth"} Figure \[E\_dep\_total\_1\_5gev\] shows the distributions of the energy deposited in the sensors of the calorimeter for beam electrons of different energy. The average deposited energy as a function of the electron beam energy is presented in Figure \[Avarage\_E\_dep\_total\_1\_5gev\]. The measured raw values increase with increasing beam energy, with a tendency of a reduced slope at larger beam energies. After applying the APV25 calibration, as described in Section \[Calibration\_APV\_25\], and correcting for the energy leakage fraction, estimated from the simulation, the response becomes nearly linear. Reconstruction of the shower position ------------------------------------- For the reconstruction of the shower position, pads with deposited energy were combined into clusters. In the first step, the depositions in all pads at a given radial and azimuthal position are summed over all detector layers. The clustering algorithm used in this study builds a cluster including all nearest neighbour pads. The pad with radial number $n$ and sector number $k$ is assigned to a cluster if the cluster contains a pad with radial number $n'$ and sector numbers $k'$ such that both $\|n-n'\| \leq 1$ and $\|k-k'\| \leq 1$. If this holds, the cluster is considered as an electromagnetic shower. The shower position is determined using a weighted sum: $$\label{eq:cog} Y_c = \frac{\sum\limits_{m} Y_{m} w_m}{\sum\limits_{m} w_m},$$ where the index $m$ runs over all pads included in the shower. $Y_{m}$ is the position of the pad and $w_m$ is a weight, which in the simplest approach could be taken as the energy $E_m$ deposited in the pad. It has however been shown [@double_exp_cher; @double_exp_reco; @awes_position_ln] that this approach gives a biased estimate when the shower position is not in the centre of a pad. Several methods were developed to achieve more accurate position reconstruction, and the following choice of weights is found to be the most appropriate: $$\label{eq:log_weight} w_m = max\left\{0; W_0+\ln{\frac{E_m}{\sum\limits_{j} E_j}}\right\},$$ where $W_0$ is a free dimensionless parameter. The performance of the clustering algorithm is studied in a simulation and for the present configuration the best resolution for the radial coordinate of the shower is achieved with $W_0=3.4$. The resolution of the shower position reconstruction in the calorimeter is estimated using the tracker planes. Two detector planes are installed at distances of 86 mm and 63 mm in front of the first tungsten plate. Because of the relatively large pad size, about 95% of the reconstructed clusters in the tracking planes consist of one pad, hence charge sharing between pads cannot be used for the position reconstruction. The impact position of beam particles is set to the middle of the pad. Since the beam particle density is found to be almost constant, a uniform distribution of beam particles within the pitch of the sensor is given. Assuming that the uncertainty of the shower position reconstruction in the calorimeter has a Gaussian distribution, the distributions of the residuals between the particle position in the tracking plane and in the calorimeter is described by the convolution $$\label{eq_residuals_fit_function} f(x) = \frac{B}{p\sigma \sqrt{2\pi }} \int\limits_{x_{0}-\frac{p}{2}}^{x_{0}+\frac{p}{2}} e^{-\frac{(x-z)^2}{2\sigma ^2}} \mathrm{d}z,$$ where $\sigma$ is the position resolution in the calorimeter, $p$ the pitch of the tracking plane, $x_{0}$ accounts for relative displacement and $B$ provides the normalisation for a given number of events. Figure \[fig\_tracker\_lumical\_residuals\] shows the distribution of the residuals of the reconstructed radial position of the shower in the calorimeter and in the two planes of the tracker. To test the performance of the method the pitch of the sensor can be also considered as a fit parameter. In this case the values found from the fit are 1.86 mm and 1.71 mm for the first and second tracking planes, respectively. These numbers are within 5% equal to the sensor pitch of 1.8 mm. The resolution $\sigma$ of the shower position reconstruction, found from the fit when $p$ is fixed to the value of the sensor pitch, is (440 $\pm$ 20) $\mu$m, and the absolute values of relative displacements $x_{0}$ are less than 5 $\mu$m. The small distortion seen at the top part of the distribution for the second tracker plane in Figure \[fig\_tracker\_lumical\_residuals\] is explained by the small asymmetry of the beam profile and circular geometry of the sensor which, in combination, result in a decline from the uniform distribution of the position uncertainty in the tracking planes. ![Distribution of residuals of the radial position measurements in the tracking planes and the calorimeter, obtained with a 5 GeV electron beam. []{data-label="fig_tracker_lumical_residuals"}](tracker_lumical_residuals_5gev.pdf){width="60.00000%"} One dimensional transverse shower profile ----------------------------------------- The one dimensional profile of the deposited energy in the sensor layers for each event is obtained as the following sum: $$\label{eq_E_pad_layer} E^{det}_{nl} = \sum\limits_{k} \epsilon_{nkl} \ ,$$ where $\epsilon_{nkl}$ is the deposited energy measured in the sensor pad with radial number $n$, sector $k$ and layer $l$. The sector index $k$ runs over two central sectors of the sensor considered and the layer index $l$ corresponds to the 5 detector planes of the calorimeter. About 5% of randomly distributed channels in the calorimeter have a larger noise level corresponding to signal sizes of up to 40 MIPs. The influence of these channels, hereafter referred to as bad channels, on the shower development study is eliminated by calculating $\langle E^{det}_{nl} \rangle$ for all indexes $n$ and $l$ only from properly working channels. Since the particle position changes from event to event due to the transverse beam size within about 10 pads, for the estimation of the average value of $\langle E^{det}_{nl} \rangle$, the index $n$ in each event is set to n=0 for the pad which contains the centre of the shower. An example of the distributions of $E^{det}_{nl}$ for the shower core ($n=0$) and pads with $n=-2$, and $n=-5$ for the layer after seven tungsten plates are shown in Figure \[fig\_e\_dep\_layer\_pads\]. Data is well described by the simulation. ![Distributions of deposited energy $E^{det}_{nl}$ as defined in eqn. (\[eq\_E\_pad\_layer\]) summed over two sensor sectors L1 and R1 in a layer after seven tungsten plates for radial pads which correspond to shower core (a), two pads away from the core (b) and five pads away from the core (c). The distributions are obtained with a 5 GeV electron beam.[]{data-label="fig_e_dep_layer_pads"}](Sig_L6_P50.pdf "fig:"){width="\textwidth"} a) ![Distributions of deposited energy $E^{det}_{nl}$ as defined in eqn. (\[eq\_E\_pad\_layer\]) summed over two sensor sectors L1 and R1 in a layer after seven tungsten plates for radial pads which correspond to shower core (a), two pads away from the core (b) and five pads away from the core (c). The distributions are obtained with a 5 GeV electron beam.[]{data-label="fig_e_dep_layer_pads"}](Sig_L6_P48.pdf "fig:"){width="\textwidth"} b) ![Distributions of deposited energy $E^{det}_{nl}$ as defined in eqn. (\[eq\_E\_pad\_layer\]) summed over two sensor sectors L1 and R1 in a layer after seven tungsten plates for radial pads which correspond to shower core (a), two pads away from the core (b) and five pads away from the core (c). The distributions are obtained with a 5 GeV electron beam.[]{data-label="fig_e_dep_layer_pads"}](Sig_L6_P45.pdf "fig:"){width="\textwidth"} c) Longitudinal shower profile --------------------------- The average energy $\langle E^{layer}_{l} \rangle$ deposited in calorimeter layer $l$ is calculated as the following sum: $$\label{eq_E_longitudinal} \langle E^{layer}_{l} \rangle = \sum\limits_{n} \langle E^{det}_{nl} \rangle \ ,$$ where $n$ runs over the radial pads of the two central sectors of the sensor. About 5% of randomly distributed channels in the calorimeter have a larger noise level corresponding to signal sizes of up to 40 MIPs. A Monte Carlo simulation has been done to estimate the impact of bad channels on the longitudinal shower profile. The result is shown in Figure \[fig\_Longitudinal\_mca\]. The red distribution corresponds to a calorimeter without bad channels and the black one is obtained after dropping bad channels, introduced in the simulation in the same locations as observed in data. Both distributions agree very well within statistical uncertainties. ![Longitudinal shower profile, comparison between data and simulation. The distributions are obtained with a 5 GeV electron beam.[]{data-label="fig_Longitudinal_dmc"}](Lshower_MC.pdf){width="\textwidth"} ![Longitudinal shower profile, comparison between data and simulation. The distributions are obtained with a 5 GeV electron beam.[]{data-label="fig_Longitudinal_dmc"}](Lshower_DMC.pdf){width="\textwidth"} The development of the longitudinal shower profile is then measured using only events with properly working channels. In Figure \[fig\_Longitudinal\_dmc\] the deposited energy as a function of the layer $l$ is shown for data and Monte Carlo simulation. The maximum of the shower is reached in data at layer 7. Both distributions are, within statistical uncertainties, in reasonable agreement. The Molière Radius {#RM_Measurement} ------------------ The sensor in Figure \[fig\_si\_sensor\] has a relatively fine segmentation in the radial direction, with a pitch of 1.8 mm, but the size of the sectors is between 2 and 2.5 cm in the irradiated area. Such a geometry does not allow to uniformly sample the electromagnetic shower in the transverse plane and requires the development of a dedicated method to measure the effective Molière radius. Such a method was developed and presented in detail in Ref. [@LumiCal_multilayer_tb2014_epjc]. Since here the same sensors are used, this method will be applied. It is briefly described in the following with small modifications which are mainly motivated by the difference in the design of the detector plane. Denoting $F_{E}( {r})$ the density function of the average deposited energy in the transverse plane with respect to the shower axis, the energy in the area covered by a single detector pad can be expressed as the integral $$\label{eq_pad_energy_int_s} E_{n} = \int\limits_{S_{n}} \! F_{E} ( {r}) \, \mathrm{d}S,$$ where $S_{n}$ is the area which corresponds to the sensor pad $n$. The function $F_{E}( {r})$ is cylindrically symmetric with respect to the shower axis, and is expressed in cylindrical coordinates with the origin at the center of the shower. Hence it depends only on the radius $r$. Since, on average, 90% of the deposited energy lies inside a cylinder with a radius of one Molière radius $R_{\mathcal{M}}$, the following equation can be used for the Molière radius calculation: $$\label{eq_MR_1} 0.9 = \frac{\int\limits_{0}^{2\pi } d\varphi \int\limits_{0}^{R_\mathcal{M}} \! F_{E} ( r ) r \, \mathrm{d}r} {\int\limits_{0}^{2\pi } d\varphi \int\limits_{0}^{\infty} \! F_{E} ( r ) r \, \mathrm{d}r } .$$ The values of $E_{n}$ can be calculated using a parameterised trial functions $F_{E}( {r})$. Fitting this trial function to the average deposited energy measured in the corresponding pads, one can define their parameters and use them in eqn. (\[eq\_MR\_1\]) to obtain the Molière radius. In the previous paper [@LumiCal_multilayer_tb2014_epjc], the circular shape of the pads was approximated for simplicity by a straight strip. The effect of this approximation was studied in a simulation [@LumiCal_multilayer_tb2014_proc] and it was shown that the difference between values of $E_{n}$ calculated for pads of circular shape and for strip-like pads depends on the pad position with respect to the shower centre amounts to at most 2%. This difference was included in the systematic uncertainty. This effect was also diminished in the data analysis because the detector planes had limited number of pads connected to the readout and some values of $E_{n}$ could not be measured directly, but were recovered assuming the symmetry with respect to the shower core. In the present study, the numerical integration in eqn. (\[eq\_pad\_energy\_int\_s\]) is done using the correct geometry of the sensor pad. To this end it is convenient to use cylinder coordinates which are linked to the sensor geometry. Changing the coordinates to $\vec{r} = \vec{r'} - \vec{R'_{0}}$, where $\vec{R'_{0}}$ is the position of the shower axis in the sensor reference frame, the pad energy can be obtained by the integration: $$\label{eq_pad_energy_int_rphi} E_{n} = \int\limits_{\varphi'_{min}}^{\varphi'_{max}} \int\limits_{r'_{n}}^{r'_{n+1}} \! F_{E} (\| \vec{r'} - \vec{R'_{0}} \|) r' \, \mathrm{d}r' \mathrm{d}\varphi',$$ where $\varphi'_{min}$ and $\varphi'_{max}$ correspond to the sectors of the sensor and $r'_{n}$ to the radius of the sensor pad $n$. The integration over $\varphi'$ comprises the sectors L1 and R1 (see Figure \[fig\_si\_sensor\]) which corresponds to about 40 mm. Since the transverse size of the beam is $\sigma_{x,y}\approx$ 4.2 mm and the expected effective Molière radius is around 10 mm, the two sectors safely cover one effective Molière radius of the shower. The trial function used to describe the average transverse energy profile of the shower is a Gaussian for the core, dominated by the high energy component of the shower, and a form inspired by the Grindhammer-Peters parameterisation [@Grindhammer; @Grindhammer1] to account for the tails originating from the low energy photon halo, $$\label{eq_MR_FrFinel} F_{E} ( r ) = A_C e^{-(\frac{r}{R_C})^2} + A_T \frac{2r^{\alpha}R_T^2}{ (r^2 + R_T^2 )^2 } \ ,$$ where $A_C$, $R_C$, $A_T$, $R_T$ and $\alpha$ are parameters to be determined by fitting the function to the measured distribution. As can be seen from eqn.(\[eq\_pad\_energy\_int\_rphi\]), the energy $E_{n}$ deposited in the pad number $n$ depends on the shower position $\vec{R'_{0}}$ and pad position $\vec{r'_{n}}$. Since the beam transverse size is significantly smaller than the radius ${R'_{0}}$, the calculation of $E_{n}$ is done for a value of ${R'_{0}}$ which corresponds to the position of the maximum in the beam profile. This maximum is observed in a pad with $n$ = 45 and $r'_{n}$ = 161 mm. The effective Molière radius determination at 5 GeV --------------------------------------------------- The average profile of the electromagnetic transverse shower is determined by summing over all detector layers, $$\label{eq_E_pad} \langle E^{det}_{n} \rangle = \sum\limits_{l} \langle E^{det}_{nl} \rangle \ .$$ The measured averaged transverse energy values, $\langle E^{det}_{n} \rangle$, were fitted to the function in eqn.(\[eq\_MR\_FrFinel\]). Results for data and Monte Carlo simulation for electrons of 5 GeV energy, are shown in Figure \[MR\_5GeV\], where one sees the dependence of $\langle E^{det}_{n} \rangle$ on the distance from the shower core, $d_{core}$. The simulation agrees well with the data. ![The ratio of the integrals in eqn (\[eq\_MR\_1\]) using $F_E(r)$ obtained from the fit, as a function of the radius $R$ in units of the pad dimension (1.8 mm), for data (blue) and MC (red), for a 5 GeV electron beam. The insert shows an expanded view of the region $2 < R < 6$ pads.[]{data-label="fig_integral"}](MR_Data_5GeV_.pdf){width="\textwidth"} ![The ratio of the integrals in eqn (\[eq\_MR\_1\]) using $F_E(r)$ obtained from the fit, as a function of the radius $R$ in units of the pad dimension (1.8 mm), for data (blue) and MC (red), for a 5 GeV electron beam. The insert shows an expanded view of the region $2 < R < 6$ pads.[]{data-label="fig_integral"}](Integral_5GeV_.pdf){width="\textwidth"} The fitted function reproduces the experimental and the simulated transverse shower profile with an accuracy better than 5%. Figure \[fig\_integral\] shows the right part of eqn. (\[eq\_MR\_1\]) as a function of the radial integration limit ${R}$ for data and simulation with the horizontal line demonstrating a graphical solution for the effective Molière radius. The result is (8.1 $\pm$ 0.1 (stat) $\pm$ 0.3 (syst)) mm, a value well reproduced by the MC simulation (8.4 $\pm$ 0.1) mm. The result obtained here is much smaller than the one determined in the calorimeter prototype used during the 2014 test beam with larger gaps between the tungsten plates, which yielded (24.0 $\pm$ 1.6) mm [@LumiCal_multilayer_tb2014_epjc]. Energy dependence of the effective Molière radius ------------------------------------------------- The main analysis was performed for data taken at 5 GeV beam energy. In addition, data were taken for energies between 1 and 5 GeV. For the study of the energy dependence, about 50,000 events were used for each energy, and the measurement of the effective Molière radius was carried out as for the 5 GeV sample. An example of the average transverse shower profiles at 1, 3 and 5 GeV beam energy is shown in Figure \[fig\_MR\_Data135\]. The average deposited energies are lower at lower beam energies, and the distributions are wider, resulting in a larger value of the effective Molière radius. The data are again well described by the results of simulations. The effective Molière radius as a function of the incoming electron energy , $E_{inc}$, in the range of 1 - 5 GeV is shown in Figure \[fig\_MR\_Edep\]. It decreases with the electron energy as $E_{inc}^{(-0.15\pm 0.04)}$. The fit to the simulation yields an exponent (-0.11$\pm$0.01), in agreement with the data. ![The effective Molière radius as a function of the electron energy for data (blue) and simulation (red).[]{data-label="fig_MR_Edep"}](MR_Data_1_3_5GeV_v4.pdf){width="\textwidth"} ![The effective Molière radius as a function of the electron energy for data (blue) and simulation (red).[]{data-label="fig_MR_Edep"}](MR_Edep_.pdf){width="100.00000%"} In order to investigate the observed energy dependence of the effective Molière radius, a simulation of an “infinite” calorimeter was performed. In practice the simulated calorimeter consisted of 30 planes with transverse size of 40 $\times$ 40 cm$^2$. Absorbers, detector layers and gaps had the same composition and thickness as the ones of the tested prototype. Figure \[fig\_l\_shower\_MC\_1\_3\_5\_10\_GeV\] shows the normalised average longitudinal profile of the energy deposited in the detector layers for incident electrons of different energy. The depth of the calorimeter is sufficient to contain most of the shower even for 10 GeV electrons in which case the fraction of the energy deposited in the last sensor layer is below 0.3%, as can be seen in the insert in Figure \[fig\_l\_shower\_MC\_1\_3\_5\_10\_GeV\]. The detector layers from 3 to 7, as installed in the prototype (shaded area) probe different regions of the longitudinal shower profile for different energies. For 1 GeV electrons, the shower is measured almost symmetrically around its maximum, while for 5 GeV electrons the layers 3 – 7 cover mostly the left side from the maximum. Hence, the fraction of the energy recorded in these layers depends on the beam energy. In Figure \[fig\_e\_lfraction\_MC\_1\_3\_5\_10\_GeV\] the cumulative distribution of the fraction of the deposited energy is shown as a function of the number of layers. In layers 3 to 7, the fractions for 1 GeV, 3 GeV and 5 GeV electrons are 56%, 50% and 46%, respectively. This difference explains a small deviation from linearity in the observed prototype response as was shown in Figure \[Avarage\_E\_dep\_total\_1\_5gev\] with red line and squares. Those measurements corrected to represent equal fractions of beam energies are shown with blue triangles and they are in good agreement with linear fit. The measurement of the shower in fixed detector layer positions for different longitudinal shower profiles also influences the observed transverse shower size. As can be seen from Figure \[fig\_RMS\_MC\_1\_3\_5\_10\_GeV\], the RMS of the lateral projection of the deposited energy in each detector layer is expected to increase as a function of the sensor layer number, with a steeper slope for lower electron energies. The small increase of the RMS observed in the first and second layers are explained by the back-scattering of shower particles. According to the results in Figure \[fig\_RMS\_MC\_1\_3\_5\_10\_GeV\] it is expected that the effective Molière radius decreases with increasing beam energy for the beam test geometry. When the fraction of the sampled shower energy approaches unity for different electron energies, the Molière radii converge to the same value. This can be seen in Figure \[fig\_MR\_MC\_1\_3\_5\_10\_GeV\], where the calculated Molière radius is shown as a function of the number of detector layers included in the calculation. Thus, the observed dependence of the effective Molière radius in the prototype on the incident electron energy, as presented in Figure \[fig\_MR\_Edep\], is due to the limited number of detector layers installed near the shower maximum. The slightly higher values of the effective Molière radius observed in the simulated calorimeter originate from the fact that in the simulation the transverse size of the calorimeter was much larger than that of the prototype. The difference is well reproduced by the simulation. ![Cumulative distribution of the fraction of energy deposited in the detector layers as a function of the number of layers for different electron beam energies. The insert shows an expanded view of the region for planes 2 to 7.[]{data-label="fig_e_lfraction_MC_1_3_5_10_GeV"}](l_shower_1_3_5_10_gev_v1.pdf){width="\textwidth"} ![Cumulative distribution of the fraction of energy deposited in the detector layers as a function of the number of layers for different electron beam energies. The insert shows an expanded view of the region for planes 2 to 7.[]{data-label="fig_e_lfraction_MC_1_3_5_10_GeV"}](e_lfraction_1_3_5_10_gev_v1.pdf){width="\textwidth"} ![Molière radius, obtained in simulation, for different electron beam energies as a function of number of detector layers in the calorimeter. The insert shows an expanded view of the plane region 6 - 10.[]{data-label="fig_MR_MC_1_3_5_10_GeV"}](rms_lateral_profile_v1.pdf){width="\textwidth"} ![Molière radius, obtained in simulation, for different electron beam energies as a function of number of detector layers in the calorimeter. The insert shows an expanded view of the plane region 6 - 10.[]{data-label="fig_MR_MC_1_3_5_10_GeV"}](mc_rm_1_3_5_10_gev_v1.pdf){width="\textwidth"} Uncertainties ------------- The study of the systematic uncertainty of the measured average energy deposition in the transverse direction $\langle E^{det}_{nl} \rangle$ includes the following contributions: - uncertainty of the measured efficiency of the signal identification; - uncertainty of the particle impact position measurement and misalignment of detector planes; - uncertainty due to bad channels; - noise uncertainty; - calibration uncertainty. The uncertainty due to the efficiency of the signal reconstruction is evaluated by changing the efficiency according to high and low edges of the shaded area in Figure \[fig\_apv\_efficiency\]. The result for the effective Molière radius changes by $\pm$0.16 mm. The misalignment of the detector planes is estimated using occupancy plots for each layer. It is accounted for in the geometry of the simulation. The effect of misalignment on the effective Molière radius comes from the sum in eqn.(\[eq\_E\_pad\]) where the radial pad index $n$ denotes pads in different layers which are assumed to be aligned in the longitudinal direction. Due to misalignment, the average lateral deposited energy $\langle E^{det}_{n} \rangle$ for a given distance from the shower core, determined by the index $n$, gets contribution from pads which are at different distances from the shower core. A similar effect arises from the uncertainty of the particle impact position. This uncertainty is estimated by calculating the effective Molière radius from simulations with perfectly aligned sensors and sensors displaced within the estimated misalignment. The change of the effective Molière radius is found to be 0.08 mm. The influence of the bad channels, which are included into simulation, leads to a change of the effective Molière radius by 0.14 mm compared to the simulation where all channels work properly. The effect due to the usage of one single radius ${R'_{0}}$ in equation (\[eq\_pad\_energy\_int\_rphi\]) for the calculation of $E_{n}$ is estimated by selecting a narrow range of the particles impact position around the sensor pad with the radial index $n=45$. The relative changes of the effective Molière radius is within 0.13 mm. The contribution of the measured noise uncertainty was studied in the simulation and found to be significantly below 1%. A relative calibration uncertainty of 5% for each APV25 front-end board is assigned to each value $\langle E^{det}_{nl} \rangle$ in eqn. (\[eq\_E\_pad\]) and summed in quadrature to determine the uncertainty of $\langle E^{det}_{n} \rangle$. The calibration uncertainty is combined with the statistical one and used to produce 1000 transverse shower profiles where each $\langle E^{det}_{n} \rangle$ is randomly generated using a Gaussian distribution function with a mean value corresponding to the measured $\langle E^{det}_{n} \rangle$ and a $\sigma$ determined by the uncertainty. For each shower, the effective Molière radius is calculated and the RMS of their distribution is considered as a contribution to the statistical uncertainty of the effective Molière radius measurement. The contributions to the systematic uncertainty are considered to be independent. The total systematic uncertainty on the Molière radius measurement is obtained by adding all the contributions in quadrature. Summary and conclusions {#Summary_section} ======================= New sub-millimeter thickness detector layers for the luminosity calorimeter LumiCal have been designed and produced. Silicon sensors are read out using Kapton fan-outs with copper traces connected via wire bonding or TAB to the sensor pads. The eight assembled detector layers were installed in the 1 mm gap between the tungsten absorber plates and successfully operated during the 2016 beam-test campaign. Measurements of the shower position and the longitudinal and transverse shower shape are presented and compared to Monte Carlo simulations. The effective Molière radius of this compact calorimeter prototype was determined at 5 GeV to be (8.1 $\pm$ 0.1 (stat) $\pm$ 0.3 (syst)) mm, a value well reproduced by the MC simulation (8.4 $\pm$ 0.1) mm. Its energy dependence in the range 1 - 5 GeV was also studied. The observed slight decrease proportional to $E_{inc}^{(-0.15\pm 0.04)}$, can be explained by the limited number of detector planes used to probe the electromagnetic shower. These results demonstrate the feasibility of constructing a compact calorimeter consistent with the conceptual design, which is optimised for a high precision luminosity measurement in future e$^{+}$e$^{-}$ collider experiments. Acknowledgements {#acknowledgements .unnumbered} ================ This study was partly supported by the Israel Science Foundation (ISF), Israel German Foundation (GIF), the I-CORE program of the Israel Planning and Budgeting Committee, Israel Academy of Sciences and Humanities, by the National Commission for Scientific and Technological Research (CONICYT - Chile) under grant FONDECYT 1170345, by the Polish Ministry of Science and Higher Education under contract nrs 3585/H2020/2016/2 and 3501/H2020/2016/2, the Rumanian UEFISCDI agency under contracts PN-II-PT-PCCA-2013-4-0967 and PN-II-ID-PCE-2011-3-0978, by the Ministry of Education, Science and Technological Development of the Republic of Serbia within the project Ol171012, by the United States Department of Energy, grant DE-SC0010107, and by the European Union Horizon 2020 Research and Innovation programme under Grant Agreement no.654168 (AIDA-2020). The measurements leading to these results have been performed at the Test Beam Facility at DESY Hamburg (Germany), a member of the Helmholtz Association (HGF) [99]{} H. Abramowicz et al., Forward instrumentation for ILC detectors. JINST [5]{} (2010) P12002. T. Behnke et al., The International Linear Collider. Technical Design Report, Volume 4: Detectors, 2013. arXiv:1306.6329 \[physics.ins-det\]. H. Baer et al., The International Linear Collider. Technical Design Report, Volume 2: Physics. 2013. arXiv:1306.6352 \[hep-ph\]. The CLIC, CLICdp collaborations, Updated baseline for a staged Compact Linear Collider. CERN Yellow Report CERN-2016-004; arXiv:1608.07537 \[physics.acc-ph\]. D. Bardin et al., One-loop electroweak radiative corrections to polarized Bhabha scattering. arXiv:1801.00125 \[hep-ph\]. H. Abramowicz et al., Measurement of shower development and its Molière radius with a four-plane LumiCal test set-up. Eur. Phys. J. [C 78]{} (2018) 135 \[1705.03885\]. https://doi.org/10.1140/epjc/s10052-018-5611-9 F.-X. Nuiry, Collected documents on the FCAL-AIDA precision mechanical infrastructure and tungsten plates. https://edms.cern.ch/document/1475879/. M. Raymond et al., The APV25 0.25 $\mu$m CMOS readout chip for the CMS tracker. IEEE Nucl. Sci. Symp. Conf. Rec. 2 (2000) 9/113. M.J. French et al., Design and results from the APV25, a deep sub-micron CMOS front-end chip for the CMS tracker. Nucl. Instr. Meth. [A 466]{} (2001) 359. S. Martoiu, H. Muller, and J. Toledo. Front-end electronics for the Scalable Readout System of RD51. Proc. IEEE Nucl. Sci. Symp. Conf. Rec., Valencia, October 2011, pp 2036-2038 (2011), http://dx.doi.org/10.1109/NSSMIC.2011.6154414. H. Abramowicz et al., Performance of fully instrumented detector planes of the forward calorimeter of a Linear Collider detector. JINST [10]{} (2015) P05009. M. Bregant et al., Assembly and validation of the ALICE silicon microstrip detector. Nucl. Instr. Meth. [A 570]{} (2007) 312. S. Martoiu, H. Muller, A. Tarazona and J. Toledo. Development of the scalable readout system for micro-pattern gas detectors and other applications. Topical Workshop on Electronics for Particle Physics 2012, 17-21 September 2012, OXFORD, U.K. JINST [8]{} (2013) C03015. G. A. Akopdjanov et al., [Determination of Photon Coordinates in A Hodoscope Cherenkov Spectrometer.]{}, Nucl. Instr. Meth. [140]{} (1977) 441. L. Bugge, [On the Determination of Shower Central Positions from Lateral Samplings.]{}, Nucl. Instr. Meth. [A 242]{} (1986) 228. T.C. Awes, F.E. Obenshain, F. Plasil, S. Saini, S.P. Sorensen, G.R. Young. [A Simple Method of Shower Localization and Identification in Laterally Segmented Calorimeters.]{}, Nucl. Instr. Meth. [A 311]{} (1992) 130. O. Borysov for the FCAL collaboration. Beam Tests of a Multilayer LumiCal Prototype. Proceedings of the International Workshop on Future Linear Colliders - LCWS2016, Morioka, Japan, December 5-9, 2016. arXiv:1703.09955 \[physics.ins-det\] G. Grindhammer et al., In: Proceedings of the Workshop on Calorimetry for the Supercollider, Tuscaloosa, AL, March 13-17, 1989, edited by R. Donaldson and M.G.D. Gilchriese (World Sci- entific, Teaneck, NJ, p. 151) (1989) Grindhammer G., Rudowicz M., and Peters S. [The Parameterized Simulation of Electromagnetic Showers in Homogeneous and Sampling Calorimeters.]{}, hep-ex/0001020. [^1]: As we do not have a fully contained shower in the prototype of LumiCal used in the earlier and also this test beam, we measure an effective Molière radius.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The triple radio source detected in association with the luminous infrared source IRAS 16547-4247 has previously been studied with high angular resolution and high sensitivity with the Very Large Array (VLA) at 3.6-cm wavelength. In this paper, we present new 3.6 cm observations taken 2.68 years after the first epoch that allow a search for variability and proper motions, as well as the detection of additional faint sources in the region. We do not detect proper motions along the axis of the outflow in the outer lobes of this source at a 4-$\sigma$ upper limit of $\sim$160 km s$^{-1}$. This suggests that these lobes are probably working surfaces where the jet is interacting with a denser medium. However, the brightest components of the lobes show evidence of precession, at a rate of $0\rlap.^\circ08$ yr$^{-1}$ clockwise in the plane of the sky. It may be possible to understand the distribution of almost all the identified sources as the result of ejecta from a precessing jet. The core of the thermal jet shows significant variations in flux density and morphology. We compare this source with other jets in low and high mass young stars and suggest that the former can be understood as a scaled-up version of the latter.' address: - 'Centro de Radioastronomía y Astrofísica, UNAM, Apdo. Postal 3-72, Morelia, Michoacán, 58089 México' - 'Harvard-Smithsonian Center for Astrophysics, 60 Garden Street, Cambridge, MA 02138, USA' - 'Departamento de Astronomía, Universidad de Chile, Casilla 36-D, Santiago, Chile' - 'Australia Telescope National Facility, P.O. Box 76, Epping NSW 1710 Australia' - 'Departamento de Astronomía, Universidad de Chile, Casilla 36-D, Santiago, Chile' author: - 'Luis F. Rodríguez' - 'James M. Moran and Ramiro Franco-Hernández[^1]' - Guido Garay - 'Kate J. Brooks' - Diego Mardones title: 'The Collimated Jet Source in IRAS 16547-4247: Time Variation, Possible Precession, and Upper Limits to the Proper Motions Along the Jet Axis' --- To appear in The Astronomical Journal Introduction ============ A successful model of low-mass star formation, based on accretion via a circumstellar disk and a collimated outflow in the form of jets (Shu, Adams, & Lizano 1987), has been developed and found to be consistent with the observations. An important question related to star formation is whether or not this model can be scaled up for the case of high-mass protostars or if other physical processes (i.e., stellar merging; Bonnell, Bate, & Zinnecker 1998; Stahler, Palla, & Ho 2000; Bally & Zinnecker 2005) are important. A small number of B-type young stars have been found to be associated with collimated outflows and possibly even circumstellar disks (see Garay & Lizano 1999; Arce et al. 2007; Cesaroni et al. 2007). The source IRAS 16547-4247 is the best example of a highly-collimated outflow associated with an O-type protostar studied so far, and its study may reveal important information about the way high mass stars form. The systemic LSR velocity of the ambient molecular cloud where IRAS 16547-4247 is embedded is $-$30.6 km s$^{-1}$ (Garay et al. 2007). Adopting the galactic rotation model of Brand & Blitz (1993) and assuming that the one-dimensional rms velocity dispersion among molecular clouds is 7.8 km s$^{-1}$ (Stark & Brand 1989), we estimate a distance of 2.9$\pm$0.6 kpc for the source. IRAS 16547-4247 has a bolometric luminosity of 6.2$\times$10$^4$ $L_\odot$, equivalent to that of a single O8 zero-age main-sequence star, although it is probably a cluster for which the most massive star would have slightly lower luminosity. Garay et al. (2003) detected an embedded triple radio continuum source associated with the IRAS 16547-4247. The three radio components are aligned in a northwest-southeast direction, with the outer lobes symmetrically separated from the central source by an angular distance of $\sim10{''}$, equivalent to a physical separation in the plane of the sky of $\sim$0.14 pc. The positive spectral index of the central source is consistent with that expected for a radio thermal (free-free) jet (e.g., Anglada 1996; Rodríguez 1997), while the spectral index of the lobes suggests a mix of thermal and nonthermal emission. Forster & Caswell (1989) detected both mainline OH and H$_2$O masers at a position close to the central continuum source. The triple system is centered on the position of the IRAS source and is also coincident within measurement error with a 1.2 mm dust continuum and molecular line emission core whose mass is on the order of 10$^3$ $M_\odot$ (Garay et al. 2003). Brooks et al. (2003) reported a chain of $H_2$ 2.12 $\mu$m emission knots that trace a collimated flow extending over 1.5 pc that emanates from close to the central component of the triple radio source and has a position angle very similar to that defined by the outer lobes of the triple radio source. Most likely this extended component traces gas ejected in the past by the central component of the triple source. The molecular observations of Garay et al. (2007) revealed the presence of a collimated bipolar outflow with lobes $\sim$0.7 pc in extent and aligned with the thermal jet located at the center of the core. In a high angular resolution study made with the VLA and ATCA, Rodríguez et al. (2005) confirmed that the central object is a thermal radio jet, while the two outer lobes are most probably heavily obscured HH objects. The thermal radio jet was resolved angularly for the first time by these authors and found to align closely with the outer lobes. Several fainter sources detected in the region away from the main outflow axis were interpreted as most probably associated with other stars in a young cluster. Brooks et al. (2007) used ATCA observations to substantiate the jet nature of the central source and to detect emission at 88 GHz that most probably arises from dust associated with this source. In this paper, we present new sensitive high angular-resolution Very Large Array observations that provide new information on the characteristics of the radio triple source in IRAS 16547-4247 as well as other sources in the field. Observations ============ The new 3.6-cm wavelength observations were made using the Very Large Array (VLA) of the National Radio Astronomy Observatory (NRAO)[^2]. These VLA radio continuum observations were carried out in 2006 May 31 and June 08 at the frequency of 8.46 GHz. The array was in the BnA configuration and an effective bandwidth of 100 MHz was used. The absolute amplitude calibrator was 1331+305 (with an adopted flux density of 5.21 Jy), and a source model provided by NRAO was used for its calibration. The phase calibrator was 1626$-$298, with bootstrapped flux densities of 1.684$\pm$0.003 and 1.823$\pm$0.004 Jy for the first and second epochs, respectively. The phase center of the array was $\alpha=16^h58^m17{\rlap.}{^s}202$ and $\delta=-42^\circ52'09{\rlap.}{''}59$ (J2000.0). The data were edited and calibrated using the software package Astronomical Image Processing System (AIPS) of NRAO. To correct for amplitude and phase errors caused by the low elevation of the source, the data were self-calibrated in phase and amplitude. No significant variations were found between the two epochs of observations (separated by only eight days) and the final analysis was made from the result of concatenating all data. The average epoch for the two data sets is 2006.42, which we use in subsequent analysis. Clean maps were obtained using the task IMAGR of AIPS with the ROBUST parameter set to 0. The synthesized (FWHM) beam was $1\rlap.{''}18\times0\rlap.{''}60; PA = 8^\circ$. The noise level achieved in the final image was 23 $\mu$Jy beam$^{-1}$. The 3.6-cm wavelength archive data was taken on 2003 September 25 and 30 (epoch 2003.74) in very similar conditions to those of the new observations and the resulting beam was $1\rlap.{''}17\times0\rlap.{''}64; PA = 11^\circ$. This old data set is described in Rodríguez et al. (2005). These data were cross-calibrated in phase and amplitude using the 2006.42 epoch as the model, which serves to globally minimize the differences in positions and flux densities of the sources. The absolute flux density error is estimated to be 10%. Results and Discussion ====================== To allow a more reliable comparison between the two epochs, both images were convolved to the same angular resolution ($1\rlap.{''}20\times0\rlap.{''}65; PA = 9^\circ$). To improve the relative gain calibration between the two epochs we examined the relative strengths of the 13 sources we identified in the images. It is clear that the ratios of the flux densities of all the sources except S-1 and N-1 had the same constant value of 0.91$\pm$0.02. Hence, we adjusted the flux scale of the first epoch by this relative gain factor. Figure 1 shows the corrected ratio of the flux densities of all the sources. The reduced $\chi^2$ of the estimate of the gain constant is 0.76, which suggests that the flux density errors may be slightly overestimated by about 15 percent. Since the $\chi^2$ is close to unity we conclude that there is no evidence for variability among the 11 sources, whereas the changes in flux densities of S-1 and the jet are highly significant. Figure 2 shows contour images of the emission observed at 8.46 GHz in the two epochs, as well as the difference image (2006.42 - 2003.74). The positions, flux densities, and deconvolved angular sizes of the sources identified in Fig. 2 are given in Table 1. The observed parameters for each component were determined from a linearized least-squares fit to a Gaussian ellipsoid function using the task JMFIT of AIPS. In Table 2 we summarize the parameters of the proper motions of all sources in the field, derived from differences of the positions in the two epochs. Note that the position of the jet changes slightly between the epochs by about 6 mas in each coordinate ($\sim3\sigma$). We do not attach any significance to this shift. If we aligned the jet positions, the proper motions listed in Table 2 would change by about 2 mas yr$^{-1}$. This amount is insignificant and does not change any of our conclusions. We note that the faint features that appear marginally detected at the 4-$\sigma$ level in the images of the individual epochs (for example, some faint structures to the west of the central source in the 2006.42 image), do not appear in the difference image, since this is about $\sqrt{2}$ noisier than the individual images. As already discussed by Rodríguez et al. (2005), the lobes first observed by Garay et al. (2003) break into several components. A more careful examination of the 2003 map in conjunction with the 2006 map shows that there are two other components present in the northern lobe (components N-4 and N-5; see Table 1 and Figure 2), as well as two additional field sources (sources D and E, see Table 1). In this section we will discuss the sources individually when new information was found, and in the following section we focus on the search for variability and proper motions. Most of the interpretation is in the context of a jet and bipolar outflow. The position angle of the jet is $-16\pm1^{\circ}$, whereas the PA of the CO lobes is about $-6^{\circ}$ and the PA of the line joining N-1 and S-1 is $-16^{\circ}$. We adopt a nominal angle for the jet outflow of $-16^{\circ}$. The central jet source ---------------------- An important difference with respect to the analysis of Rodríguez et al. (2005) is that we recognize the presence of a compact source (source D) very close to the jet source (see Fig. 2), about $1\rlap.{''}2$ to its NW. To obtain the parameters of the jet and of source D separately we fitted simultaneously two Gaussian ellipsoids to this region of the image. The jet parameters for the two epochs discussed are given in Table 1. We note that the consideration of source D in the fitting procedure results in somewhat smaller deconvolved dimensions for the minor axis of the jet than obtained in Rodríguez et al. (2005). We assume that the opening angle of the thermal jet is the angle subtended by the deconvolved minor axis at a distance of one-half the deconvolved major axis (Eislöffel et al. 2000). Using the average of the deconvolved angular dimensions given in Table 1, we then estimate the opening angle of the thermal jet to be $\sim$15$^\circ$ (as opposed to the value of $\sim$25$^\circ$ derived by Rodríguez et al.), indicating significant collimation in this massive protostar. This result suggests that jets from high mass young stars can be as collimated as those found in lower mass objects, where HST studies indicate opening angles in the range of tens of degrees in scales of tens of AU from the star (Ray et al. 2007). On larger physical scales the optical jets from low mass young stars are known to show recollimation, resulting in opening angles of a few degrees (Ray et al. 2007). This recollimation is not evident in the thermal radio jets, which are usually detected only close to the star. The northern lobe ----------------- As noted before, we identify two additional components (N-4 and N-5) in this lobe, indicating an almost continuous sequence of knots between the central jet and the outermost N-1 component (see Fig. 2). The sequence of five knots, N-1 to N-5, shows a gentle monotonic curvature that may be indicative of precession (see Section 5). Four of the five knots are resolved angularly, and it is interesting to note that all four have position angles consistent with $\sim$$160^\circ \pm 10^\circ$. This result suggests that they are part of the northern outflow. This orientation does not seem to be valid for component S-1, as discussed in the next subsection. The N-1 and N-2 continuum components are the only ones in the whole region that appear to be associated with class I methanol masers (Voronkov et al. 2006). The southern lobe ----------------- Only the component S-1 is clearly resolved, but in contrast to the components of the northern lobe that have intrinsic position angles consistent with $\sim$$160^\circ \pm 10^\circ$, the southern lobe’s intrinsic position angle is $\sim$$40^\circ$. The misalignment between the intrinsic position angle of S-1 and the outflow axis, suggests that more than part of the main jet body, S-1 could be its working surface (e.g., Chakrabarti 1988) to the south. As we will see below, the time variability of this component could support this interpretation. On the other hand, the presence of an additional component (S-2) downstream suggests that our interpretation of S-1 as a working surface may be incorrect, or that S-2 is an independent source associated with a star. The fact that S-2 is unresolved is consistent with this interpretation. Source A -------- This source is clearly resolved. It appears to be part of a diffuse region of emission that connects with the brighter component of the northern lobe. Source A could actually be part of the outflow, either if the outflow is less collimated and the ionized gas preferentially highlights regions where the flow is interacting with dense ambient material or if source A is part of the northern flow being deflected to the east by interaction with a dense clump of gas. The second possibility is interesting and it may be related to the fact that components N1 and N2 are the only ones in the region showing methanol maser emission. It is possible that the methanol maser emission is being stimulated by the deflection of the jet, as it interacts with N1 and N2. Finally, the position of source A could be understood if the outflow has precession (see discussion in Section 5). High angular resolution observations of a tracer of dense molecular gas are needed to advance our understanding of this source. Source D -------- Source D is the faint source located $1\rlap.{''}2$ to the NW of the jet, and first identified as an independent source here. It is unresolved angularly and may probably trace an independent star. In the 2006.42 image, it shows a faint extension to the west. Source E -------- Source E is a barely detected source located about $10''$ to the east of the core of IRAS 16547-4247. It is located outside of the region shown in Figure 2, and is unresolved in angular size. A contour image of this source is shown in Figure 3. The a priori probability of finding a 3.6 cm background source with a flux density of $\sim$0.14 mJy (the average flux density of the two epochs) in a solid angle of $20'' \times 20''$ is only $\sim$0.003 (Windhorst et al. 1993). We conclude that this radio source most probably traces a young star embedded in this region. Search for variation and proper motions ======================================= Analysis of Figure 2 and Table 1 indicates that significant flux density or morphological variations are observed only in three sources: the jet source, N-1, and S-1. This is not unexpected, since these are the three brightest sources in the field and small variations in position or flux density are not evident in weaker sources where the signal-to-noise ratio is much smaller. In the case of N-1 the variation is only in position since the flux density remained constant and the flux density variations are evident only for the jet source and S-1. We will discuss our interpretation of the observed variabilities and proper motions in the following subsections. The central jet source and source S-1 ------------------------------------- The jet source seems to have increased its flux density by about 10% between the two epochs (see Table 1 and Figure 1). This difference is consistent with the fact that both thermal jets (Rodríguez et al. 2001; Galván-Madrid, Avila, & Rodríguez 2004) and HH knots (Rodríguez et al. 2000) can show small but statistically significant flux density variations on scales of years or even months. The difference image (Figure 4) suggests that the increase in emission comes from two discrete positions. One increase (of about 0.8 mJy) is unresolved and associated with the central source and we attribute it to an increase in mass loss at the core of the jet. The second increase (of about 0.4 mJy) comes from an unresolved component clearly displaced to the SE, and located at $\alpha(2000) = 16^h~ 58^m~ 17\rlap.^s2446 \pm 0\rlap.^s0033; \delta(2000) = -42^\circ~ 52'~ 08\rlap.^{''}354 \pm 0\rlap.^{''}067$, at $1\rlap.{''}3$ from the center of the jet. The OH maser emission detected by Caswell (1998) is close to the unresolved component associated with this second increase. We can think of three interpretations for this localized flux density increase. One is that we are observing a time-variable source powered by an independent star. There are two arguments against this interpretation: the first is that the sources in this region in general show little variation (see Table 1) and the second is that the source lies exactly in the path of the outflow, suggesting a relation with the jet. This latter fact suggests a second interpretation: we are seeing a discrete, new ejecta from the jet. This interpretation, on its turn, has several difficulties. The first is that, while a similar phenomenon was observed and monitored in the jet associated with HH 80-81 (Martí, Rodríguez, & Reipurth 1995; 1998), in this source the ejection was clearly bipolar, while in IRAS 16547-4247 we would have to consider a monopolar ejection. The second problem is related to the large velocity required for the ejection to move over $1\rlap.{''}3$ in 2.68 years or less, that at a distance of 2.9 kpc implies the unlikely velocity of $\sim10^4$ km s$^{-1}$ or more. Finally, a third possibility is that we are observing the brightening of the jet flow as it interacts with dense gas in its path. Observations of higher angular resolution are required to disentangle the nature of the variations seen in the jet source. Unfortunately, the present observations are the best that it can be done now and possibly for decades (until completion of the Square Kilometer Array). The source can be observed, but with lower angular resolution, with ATCA (as already done) but it cannot be observed from the latitude of the future e-MERLIN. Finally, the emission is thermal and not easily detectable with VLBI arrays. The source S-1 is the one with the largest flux density variation, with an increase of $\sim$40% (see Figure 1 and Table 1) between the two epochs. We believe that this large increase is related to our suggestion that S-1 may be a working surface of the jet, where kinetic energy is rapidly being dissipated and changes are expected. Lack of Proper Motions Along the Jet Axis ----------------------------------------- The proper motions of the prominent components N-1 and S-1 along the nominal jet axis of $-16^\circ$ are $-2 \pm 5$ and $-3 \pm 2$ mas yr$^{-1}$, respectively. These motions correspond to inward motions of 28 and 42 km s$^{-1}$, respectively, and are not statistically significant. The weighted average radial motion of all the components is $5.0 \pm 2.5$ mas yr$^{-1}$, corresponding to $70 \pm 40$ km s$^{-1}$. We adopt a conservative 4-$\sigma$ upper limit on the proper motion of 160 km s$^{-1}$. This upper limit is not very stringent, but certainly indicates that the IRAS 16547-4247 lobes are not moving as fast as the knots observed in the jets associated with the massive young stars HH 80-81 and Cep A HW2, where velocities in the plane of the sky of $\sim$500 km s$^{-1}$ have been reported (Martí, Rodríguez, & Reipurth 1998; Curiel et al. 2006). Proper Motions in the Direction Transverse to the Jet Axis ---------------------------------------------------------- The only components that show transverse motions to the nominal jet direction of $-16$ degrees above the 4-$\sigma$ level are N-1 and S-1 (see Table 2). These motions are $-13 \pm 3$ and $-16 \pm 2$ mas yr$^{-1}$, corresponding to clockwise transverse motions of $180 \pm 40$ and $220 \pm 30$ km s$^{-1}$, respectively. The discovery of transverse (clockwise) motion, with no radial motion, was unexpected. The total flux density of component N-1 hardly changed, and the difference images shown in Figures 2 and 5 clearly show the negative-positive residuals characteristic of a moving source with constant flux density. In contrast, the negative-positive signature of proper motion is not present in the case of S-1 (see Figures 2 and 6) because during the same period there was the strong brightening previously discussed that dominated the difference image. A Precession Model for IRAS 16547-4247 ====================================== We believe that it may be possible to understand the distribution of almost all the identified sources as well as the clockwise precession as the result of ejecta from a precessing jet of stellar origin. The gently curved and antisymmetric distribution of the radio components strongly suggests such a model. The basic idea is that the jet itself is not observed directly (because of the lack of radial motion in the outflow). Rather the observed sources are the result of the jet interacting with the ambient medium. We describe this model in simple empirical terms. We assume that the axis of symmetry is along the N-S direction and that the position angle (measured East of North) of the precession axis is given by $$\theta \sim\, \theta_m \,\sin(\omega t) - \theta _0~,~~~~~~~~ (1)$$ where $\theta_m$ is the amplitude of precession, $\omega$ is the precession rate and $\theta_0$ is the position angle at $t=0$. We assume that the time scale of the observable outflow phenomenon is much shorter than the precession period. Hence, over a short period the precession position angle changes linearly, $$\theta \simeq \theta_m \omega t - \theta_0 = \beta t - \theta_0~.~~~~~~~~~~~~~~(2)$$ The coordinates in the plane of the sky at the time of our observations, $t$, for the location of gas ejected from the star at time $t_e$ will be $$x(t) = r(t) \, \sin\,\theta_e~~~~~~~~~~~~~~~~~~(3)$$ $$y(t) = r(t)\cos \theta_e~~~~~~~~~~~~~~~~~~~(4)$$ where $\theta_e$ is the precession angle at the time of emission, and $r$ is the radial position of the ejecta, $r = v~(t-t_e)$. We assume that the ejection velocity, $v$, is constant. Since $t-t_e = r/v$ we obtain $$x(t) = r\,\sin\,(\frac{\beta r}{v} - \theta_0)~~~~~~~~~~~~~~~~~~~(5)$$ $$y(t) = r\,\cos\,(\frac{\beta r}{v} - \theta_0)~.~~~~~~~~~~~~~~~(6)$$ The prediction of this simple model is that $\theta$ is linearly related to $r$, that is $$\theta(r) = \beta~r/v - \theta_0. ~~~~~~~~~~~~~~~~~~~~(7)$$ Figure 7 shows the position angle of each component plotted versus its radial offset from the center of the jet component, which we identify as the origin. Source C deviates from the straight line fit to the data by about 19$^\circ$. We assume that it is not part of the jet, but a separate entity from the jet, perhaps evidence of another nearby star. The straight-line fit of equation 7 to the remaining 9 sources gives parameters, $\theta_0 = -43^\circ \pm 4^\circ$ (the current PA of the jet), and $\alpha = \beta/v = 2.3 \pm 0.4$ degrees/arcsecond. For a distance of 2.9 kpc the value of $\alpha$ in the source frame is 9.2 $\times 10^{-19}$ radians/cm or 2.9 radians/pc. Since the jet seems to be precessing linearly with time, we assume that the precession period is much greater than the time scale of the flow. Note that we cannot solve separately for the precession rate and the ejection velocity. The shape of the jet at the current epoch, as determined by equations 5 and 6, is shown in Figure 8. As noted before only two sources have significant proper motions, N-1 and S-1. Their motions are both in the transverse direction to the direction of the jet, whereas their motions along the jet are insignificant (see Table 2). Hence, we assume that the sources are caused by the interaction of the jet and clumps of ambient material as the jet sweeps across the clumps. In this model, the clumps will have apparent transverse velocities given by $$v_T = r \beta = r \alpha v~, ~~~~~~~~~~~~~~~~~~~~~(8)$$ which, with the observed value of $\alpha$, becomes $$v_T = 0.42 v (r/10")~. ~~~~~~~~~~~~~~~~~~~~~~~~~(9)$$ The proper motion measurements of all the sources are listed in Table 2. We believe that the apparent motions of source C could be due to the presence of an additional, time-variable component to the NW of the main component of source C (see Fig. 2). Source N-4 shows proper motions that are slightly below 4-$\sigma$ (see Table 2). However, these marginal proper motions are against the expected flow of the jet and, if real, are probably due to a “Christmas tree” effect more that to a real motion. The velocity vectors of S-1 and N-1 are nearly transverse to the jet direction, and agree in direction with the sense of precession of the jet (see Fig. 8). These velocities are 180 and 220 km s$^{-1}$, respectively. Since N-1 and S-1 both have about the same radial distance, we cannot confirm the expected linear trend in transverse velocity with radius predicted by eqn. 8. However, with these two data points we estimate from eqn. 9 that the ejection velocity is $490~\pm~80$ km s$^{-1}$. If we assume that we are indeed dealing with a working surface, the lack of proper motions ($\leq$160 km s$^{-1}$) allow us to estimate the ratio between the density of the medium and the density of the jet. Using the formulation of Raga, Rodríguez, & Cantó (1997), the ratio of ambient medium density, $\rho_a$, to jet density, $\rho_j$, is $${{\rho_a} \over {\rho_j}} = \biggl({{v_j} \over {v_{ws}}} - 1 \biggr)^2,~~~~~~~~~~(10)$$ where $v_j$ and $v_{ws}$ are the jet and working surface velocities, respectively. For a jet velocity of 490 km s$^{-1}$ and an upper limit of 160 km s$^{-1}$ for the velocity of the working surface, we obtain $\rho_a/\rho_j \geq$ 4. The precession rate of the jet is $\beta = \alpha v$. Hence with $\alpha = 2.3\pm0.4$ degree/arcsecond, and v = $490\pm80$ km s$^{-1}$, $\beta = 0.080\pm~0.02$ degrees/year. Note that the measurements of $\alpha$ and v are both 6-$\sigma$ results, but the measurement of $\beta$ is a 4-$\sigma$ result. The precession period of the jet is $T=2 \pi \theta_m/(\alpha v)$, which for a precession opening angle of 30$^\circ$, would be about 5500 years. Since the range in PA among the components is about 25$^{\circ}$ (see Fig. 7), the jet travel time to outer components is about 300 years. We might expect that the sources excited by the passing jet would persist after the jet moves on. However, there is no evidence (see Fig. 8) of structure in the transverse direction in any of the sources. This absence of “trails” suggests that the decay times for emission must be less than $\sim~100$ years, which would produce an extension of 1.5$''$ at the radius of sources N-1 and S-1. The recombination time for an H II region with electron density $n_e$ is estimated as $$\left(\frac{t}{\mbox{yr}}\right)=1.2\times10^5\left(\frac{n_e}{\mbox{cm}^{-3}}\right)^{-1}.~~~~~~(11)$$ The spectral index of N-1 indicates it is probably a thermal optically thin HII region (Rodriguez et al. 2005). Using the values for the flux density and size in Table 1 we estimate a emission measure of $1.8\times10^6$ cm$^{-6}$ pc$^{-1}$ and the electron density as $n_e=1.1\times10^4$ cm$^{-3}$. Using this value for the electron density in eqn. (10) we find a recombination time for the N-1 component to be about 11 years. This time is an order of magnitude smaller than that required for the jet to leave an observable transverse trail. For the S-1 source the emission may have a nonthermal component (Rodriguez et al 2005). The density of relativistic electrons, $n_{er}$, and magnetic field, $B$, can be estimated using eqns. (2) and (3) from Garay et al. (1996) $$\left(\frac{B}{\mbox{mG}}\right)=0.50 \left(\frac{S_{\nu}}{\mbox{mJy}}\right)^{2/7} \left(\frac{\theta_{s}}{\mbox{arcsec}}\right)^{-6/7} \left(\frac{\nu}{10~\mbox{GHz}}\right)^{1/7} \left(\frac{D}{\mbox{kpc}}\right)^{-2/7} \left(\log\frac{E_{max}}{E_{min}}\right)^{2/7},~~~~~~~~~~~~~~~~~~~~~~~~~(12)$$ $$\left(\frac{n_{er}E_{min}}{10^{-9}~\mbox{ergs cm}^{-3}}\right)=3.58 \left(\frac{S_{\nu}}{\mbox{mJy}}\right) \left(\frac{\theta_{s}}{\mbox{arcsec}}\right)^{-3} \left(\frac{\nu}{10~\mbox{GHz}}\right)^{1/2} \left(\frac{D}{\mbox{kpc}}\right)^{-1} \left(\frac{B}{\mbox{mG}}\right)^{2/7},~~~~~~~~~~~~~~~~~~~~~~~~~(13)$$ where $E_{max}$ and $E_{min}$ are the maximum and minimum energies of the relativistic electrons that we assume are $10^{11}$ and $10^{6}$ eV respectively. For the distance $D=2.9$ kpc and $\theta_s=0.76$ arcsec, we get $B=1.4$ mG and a density of relativistic electrons of $n_{er}=4.6\times10^{-3}$ cm$^{-3}$. The decay time is given by Krolick (1999) as $t_{d}=1.3\times10^{12}\nu^{-1/2}B^{-3/2}$ sec with $\nu$ in Hz and $B$ in G, or $t_{d}(yr) = 0.5B^{-3/2}$ for our frequency. Using the value derived above for $B$ we get $t_{d}=9\times10^4$ yr, which is much too long. A magnetic field of 30 mG would be required to achieve a synchrotron decay time of 100 years. Franco-Hernández et al. (2008, in preparation) have discovered two linear structures traced by water vapor masers on the mas scale. Their g2 group of maser has a PA of about $-30^\circ$, which is 13$^\circ$ away from the current PA of the proposed precessing source ($-43^\circ$). Another group of masers (g1) has a position angle of about 50$^\circ$. If these masers were in a disk, their pole PA would be $-40^\circ$, close to the current precession angle. Hence these masers may be associated with the precessing excitation source. The larger scale CO lobes are offset from the jet by about 15[”]{} along a PA of about $-6^{\circ}$. The line of sight velocity with respect to the ambient cloud is about $\pm$30 km s$^{-1}$. These lobes could also be associated with the precessing source. The expected PA for emission at 15[”]{} from eqn. (7) would be about $-8^{\circ}$, close to the observed value. If the molecular flow picks up the full jet velocity of 490 km s$^{-1}$, then the inclination would be about 90 – arctan (30/490), or about 86$^\circ$, very close to the value inferred from the analysis of the velocity structure of the CO lobes (Garay et al 2007). More realistically, the molecular flow velocity will be smaller, implying a smaller inclination angle. The precession model nicely accounts for the fact that the line from N-1 to S-1 does not intercept the known jet. However, if the angle of the driving precessing source is currently $-43^\circ$, it is significantly discrepant with the PA of $-16$ of the known radio jet. A possible explanation for this discrepancy might be that the central source is really a binary stellar system. Source I would be a non-precessing source associated with the known jet, which is related to the large scale CO outflow with about the same PA. Source II would be a precessing source which drives the continuum thermal sources and may be associated with the water masers and also the large scale CO outflow. Putting the IRAS 16547-4247 Jet in Context ========================================== How does the IRAS 16547-4247 jet compare with other jets found in regions of star formation? Anglada et al. (1992) have compared the centimeter radio luminosity (taken to be proportional to the flux density times the distance squared, $S_\nu d^2$) of thermal jets associated with low mass young stars with the momentum rate ($\dot P$) in the associated molecular outflow. For 16 sources studied, they conclude that these low mass objects can be fitted with a power law of the form $\dot P = 10^{-2.6} (S_\nu d^2)^{1.1}$, where $\dot P$ is in $M_\odot$ yr$^{-1}$ km s$^{-1}$, $S_\nu$ is in mJy and $d$ is in kpc. In Figure 9, we plot the data used by Anglada et al. (1992). They interpret the fit to be in agreement with a simple model in which the observed ionization is produced by shocks (Curiel, Cantó, & Rodríguez 1987; Curiel et al. 1989), where about 10% of the energy in the jet is thermalized. We have plotted in the same figure, the three best studied cases of thermal jets associated with massive young stars: IRAS 16547-4247, HH 80-81, and Cep A HW2, using the data listed in Table 3. Remarkably, these three data points fall reasonably well on the Anglada et al. correlation. We believe that this agreement suggests that the jets associated with massive young stars may be a scaled-up version of the phenomenon seen in low mass young stars, although a firmer conclusion requires the study of a larger sample of objects than now available. Note that the intrinsic radio luminosity of the IRAS 16547-4247 jet is about $2~\times~ 10^3$ larger than the typical radio luminosity of jets associated with low mass stars, and that it is the most luminous case known. Conclusions =========== Our main conclusions follow: 1\) We present new, sensitive 3.6-cm wavelength VLA observations of the multiple radio source associated with the luminous infrared source IRAS 16547-4247, the most massive example known of a thermal jet found in association with a forming star. The main purposes of these new observations were to search for variability and proper motions as well as to detect additional faint sources in the region. We detected four new components (N-4, N-5, D, and E) in the region; 2\) We do not detect proper motions along the axis of the flow in the outer lobes of this source at a 4-$\sigma$ upper limit of $\sim$160 km s$^{-1}$, suggesting that if these lobes are the working surfaces of the jets and the jet velocity is $\sim$490km s$^{-1}$, the ambient medium is at least four times as dense as the jet; 3\) The brightest components of the lobes, sources N-1 and S-1, show evidence of clockwise precession, at a rate of $0\rlap.^\circ08$ yr$^{-1}$ in the plane of the sky; 4\) A precessing model can account for the antisymmetric distribution of most of the sources in the field, as well as for the evidence of precession in sources N-1 and S-1; 5\) The thermal jet at the core of the region shows significant variations in flux density and morphology but our angular resolution is insufficient to reach a clear conclusion on what produces these changes; and 6\) The correlation found by Anglada et al. (1992) for outflows and jets in low mass stars extends to the handful of massive forming stars known. This result suggests that the jets associated with massive young stars are a scaled-up version of the phenomenon seen in low mass young stars, although a firmer conclusion requires the study of a larger sample of jets that now available. We thank an anonymous referee for valuable comments. LFR acknowledges support from grant CB0702172\_3 of COECyT, Michoacán, México. DM and GG acknowledge support from the Chilean [Centro de Astrofísica]{} FONDAP 15010003. References {#references .unnumbered} ========== Anglada, G., Rodríguez, L. F., Cantó, J., Estalella, R., & Torrelles, J. M. 1992, ApJ, 395, 494 Anglada, G. 1996, in ASP Conf. Ser. 93, Radio Emission from the Stars and the Sun, ed. A. R. Taylor & J. M. Paredes (San Francisco: ASP), 3 Arce, H. G., Shepherd, D., Gueth, F., Lee, C.-F., Bachiller, R., Rosen, A., & Beuther, H. 2007, in Protostars and Planets V, ed. B. Reipurth, D. Jewitt, & K. Keil (Tucson: Univ. Arizona Press), 245 Bally, J. & Zinnecker, H. 2005, AJ, 129, 2281 Bonnell, I. A., Bate, M. R., & Zinnecker, H. 1998, MNRAS, 298, 93 Brand, J. & Blitz, L. 1993, A&A, 275, 67 Brooks, K. J., Garay, G., Mardones, D., & Bronfman, L. 2003, ApJ, 594, L131 Brooks, K. J., Garay, G., Voronkov, M., & Rodríguez, L. F. 2007, ApJ, 669, 459 Caswell, J. L. 1998, MNRAS, 297, 215 Cesaroni, R., Galli, D., Lodato, G., Walmsley, C. M., & Zhang, Q. 2007, in Protostars and Planets V, ed. B. Reipurth, D. Jewitt, & K. Keil (Tucson: Univ. Arizona Press), 197 Chakrabarti, S. K. 1988, MNRAS, 235, 33 Curiel, S., Cantó, J., & Rodríguez, L. F. 1987, RevMexA&A, 14, 595 Curiel, S., Rodríguez, L. F., Bohigas, J., Roth, M., Cantó, J., & Torrelles, J. M. 1989, ApL&C, 27, 299 Curiel, S., et al. 2006, ApJ, 638, 878 Eislöffel, J., Mundt, R., Ray, T. P., & Rodríguez, L. F. 2000, in Protostars and Planets IV, ed. V. Mannings, A. P. Boss, & S. S. Russell (Tucson: Univ. Arizona Press), 815 Forster J. R., & Caswell J. L., 1989, A&A, 213, 339 Galván-Madrid, R., Avila, R., & Rodríguez, L. F. 2004, RevMexA&A, 40, 31 Garay, G., Ramirez, S., Rodríguez, L. F., Curiel, S., & Torrelles, J. M. 1996, ApJ, 459, 193 Garay, G., & Lizano, S. 1999, PASP, 111, 1049 Garay, G., Brooks, K. J., Mardones, D., & Norris, R. P. 2003, ApJ, 587, 739 Garay, G., Mardones, D., Bronfman, L., Brooks, K.J. Rodríguez, L.F., Güsten, R., Nyman, L-[Å]{}, Franco-Hernández, R., & Moran, J.M. 2007, A&A 463, 217 Krolik, J. H. 1999, Active galactic nuclei: from the central black hole to the galactic environment (Princeton, N.J.: Princeton University Press) Martí. J., Rodríguez, L. F., & Reipurth, B. 1993, ApJ, 416, 208 Martí. J., Rodríguez, L. F., & Reipurth, B. 1995, ApJ, 449, 184 Martí. J., Rodríguez, L. F., & Reipurth, B. 1998, ApJ, 502, 337 Narayanan, G., & Walker, C. F. 1996, ApJ, 466, 844 Raga, A. C., Rodríguez, L. F., & Cantó, J. 1997, RMxF, 43, 825 Ray, T., Dougados, C., Bacciotti, F., Eislöffel, J., & Chrysostomou, A. 2007, in Protostars and Planets V, ed. B. Reipurth, D. Jewitt, & K. Keil (Tucson: Univ. Arizona Press), 231 Rodríguez, L. F., Curiel, S., Moran, J. M., Mirabel, I. F., Roth, M. & Garay, G. 1989, ApJ, 346, L85 Rodríguez, L. F. 1997, in Herbig-Haro Flows and the Birth of Low Mass Stars, proceedings of IAU Symp. No. 182, eds. B. Reipurth & C. Bertout, p. 83 (Dordrecht: Kluwer) Rodríguez, L. F., Delgado-Arellano, V. G., Gómez, Y., Reipurth, B., Torrelles, J. M., Noriega-Crespo, A., Raga, A. C., & Cantó, J. 2000, AJ, 119, 882 Rodríguez, L. F., Torrelles, J. M., Anglada, G., & Martí, J. 2001, RevMexA&A, 37, 95 Rodríguez, L. F., Garay, G., Brooks, K. J., & Mardones, D. 2005, ApJ, 626, 953 Shu, F. H., Adams, F. C., & Lizano, S. 1987, ARA&A, 25, 23 Stahler, S. W., Palla, F., & Ho, P. T. P. in Protostars and Planets IV (eds. Mannings, V., Boss, A.P. & Russell, S. S.), 327 (University of Arizona Press, Tucson, 2000) Stark, A. A., & Brand, J. 1989, ApJ, 339, 763 Voronkov, M. A., Brooks, K. J., Sobolev, A. M., Ellingsen, S. P., Ostrovskii, A. B. & Caswell, J. L. 2006, MNRAS, 373, 411 Windhorst, R. A., Fomalont, E. B., Partridge, R. B., & Lowenthal, J. D. 1993, ApJ, 405, 498 Yamashita, T., Suzuki, H., Kaifu, N., Tamura, M., Mountain, C. M., & Moore, T. J. T. 1989, ApJ, 347, 894 ![Ratio of 2006 to 2003 flux densities versus the logarithm of the 2003 flux density, given in mJy. Only the jet and source S-1 show clear evidence of variability between the two epochs. \[fig1\]](f1.eps) ![VLA contour images at 8.46 GHz towards IRAS 16547-4247 for epochs 2003.74 (left) and 2006.42 (center), as well as the difference image (2006.42 - 2003.74). Contours are -25, -20, -15, -10, -8, -6, -5, -4, 4, 5, 6, 8, 10, 15, 20, 25, 30, 40, 60, 100, 140, 160, and 200 times 27 $\mu$Jy beam$^{-1}$ for the 2003.74 and 2006.42 images (the average value of the rms noises of the two images) and 38 $\mu$Jy beam$^{-1}$ for the difference image (the rms noise of this last image). The half power contour of the synthesized beams ($1\rlap.{''}20 \times 0\rlap.{''}65$; PA = $9^\circ$) is shown in the bottom left corner of each panel. The individual sources are identified in the 2003.74 image. \[fig2\]](f2.eps) ![VLA contour images at 8.46 GHz of the source E, made from the average of both epochs (2003.74 and 2006.42). Contours are -4, 4, 5, and 6 times 18 $\mu$Jy beam$^{-1}$, the rms noise of the image. The half power contour of the synthesized beam is as in Fig. 2. \[fig3\]](f3.eps) ![VLA contour images at 8.46 GHz towards the jet in IRAS 16547-4247 for epochs 2003.74 (left) and 2006.42 (center), as well as the difference image (2006.42 - 2003.74). Contours are -25, -20, -15, -10, -8, -6, -5, -4, 4, 5, 6, 8, 10, 15, 20, 25, 30, 40, 60, 100, 140, 160, and 200 times 27 $\mu$Jy beam$^{-1}$ for the 2003.74 and 2006.42 images and 38 $\mu$Jy beam$^{-1}$ for the difference image. The half power contour of the synthesized beams ($1\rlap.{''}20 \times 0\rlap.{''}65$; PA = $9^\circ$) is shown in the bottom right corner. The cross marks the peak position of the jet as determined from the average of the two positions given in Table 1. Individual sources are identified in the 2003.74 image. \[fig4\]](f4.eps) ![VLA contour images at 8.46 GHz towards the component N-1 in IRAS 16547-4247 for epochs 2003.74 (left) and 2006.42 (center), as well as the difference image (2006.42 - 2003.74). Contours are -25, -20, -15, -10, -8, -6, -5, -4, 4, 5, 6, 8, 10, 15, 20, 25, 30, 40, 60, 100, 140, 160, and 200 times 27 $\mu$Jy beam$^{-1}$ for the 2003.74 and 2006.42 images and 38 $\mu$Jy beam$^{-1}$ for the difference image. The half power contour of the synthesized beams ($1\rlap.{''}20 \times 0\rlap.{''}65$; PA = $9^\circ$) is shown in the bottom right corner. Individual sources are identified in the 2003.74 image. \[fig5\]](f5.eps) ![VLA contour images at 8.46 GHz towards the component S-1 in IRAS 16547-4247 for epochs 2003.74 (left) and 2006.42 (center), as well as the difference image (2006.42 - 2003.74). Contours are -25, -20, -15, -10, -8, -6, -5, -4, 4, 5, 6, 8, 10, 15, 20, 25, 30, 40, 60, 100, 140, 160, and 200 times 27 $\mu$Jy beam$^{-1}$ for the 2003.74 and 2006.42 images and 38 $\mu$Jy beam$^{-1}$ for the difference image. The half power contour of the synthesized beams ($1\rlap.{''}20 \times 0\rlap.{''}65$; PA = $9^\circ$) is shown in the bottom right corner. Individual sources are identified in the 2003.74 image. \[fig6\]](f6.eps) ![Position angle of the jet components as a function of radial offset from the jet center. The components south of the jet have been folded by 180$^\circ$ in this figure. The dashed line is the linear least squares fit to the components, with the exception of source C, that is taken to be an independent star. \[fig7\]](f7.eps) ![VLA contour image at 8.46 GHz towards IRAS 16547-4247 for epoch 2003.74. Contours and beam are as in Figure 2. The solid line indicates the position of the spiral model discussed in the text. The arrows indicate the proper motions of components N-1 and S-1 for a period of 300 years. \[fig8\]](f8.eps) ![Momentum rate in the molecular outflow, $\dot P$, versus the radio flux density times distance squared, $S_\nu d^2$. The solid dots are the 16 low mass young stars from the study of Anglada et al. (1992). We have included three high mass young stars (solid triangles, see also labels in figure) from this paper as well as from the references given in Table 3. The dashed line is the fit of Anglada et al. (1992) to the low mass young stars. The high mass young stars fall reasonably well on the correlation, suggesting a common nature for thermal jets associated with low and high mass young stars. \[fig9\]](f9.eps) [@lllll]{} Component & $\alpha$(2000)$^{\rm a}$ & $\delta$(2000)$^{\rm a}$ & S$_\nu^{\rm b}$ & (Epoch) & ($16^h~58^m$) & ($-42^\circ$) & (mJy) & Deconvolved Size$^{\rm c}$ A(2003) & 17$\rlap.^s$0597$\pm$0.0077 & 51$'$ 55$\rlap.{''}$039$\pm$0$\rlap.{''}$240 & 1.58$\pm$0.23 & 4.[”]{}2$\pm$0.[”]{}6$\times$1.[”]{}2$\pm$0.[”]{}2; 9$^\circ \pm$4$^\circ$\ A(2006) & 17$\rlap.^s$0641$\pm$0.0072 & 51$'$ 55$\rlap.{''}$212$\pm$0$\rlap.{''}$205 & 1.60$\pm$0.20 & 4.[”]{}1$\pm$0.[”]{}5$\times$1.[”]{}5$\pm$0.[”]{}2; 4$^\circ \pm$5$^\circ$\ N-1(2003) & 16$\rlap.^s$8165$\pm$0.0005 & 51$'$ 57$\rlap.{''}$048$\pm$0$\rlap.{''}$011 & $2.86\pm0.07$ & 1.[”]{}06$\pm$0.[”]{}04$\times$0.[”]{}31$\pm$0.[”]{}05; 163$^\circ \pm$3$^\circ$\ N-1(2006) & 16$\rlap.^s$8136$\pm$0.0004 & 51$'$ 57$\rlap.{''}$060$\pm$0$\rlap.{''}$009 & $2.85\pm0.06$ & 1.[”]{}15$\pm$0.[”]{}03$\times$0.[”]{}23$\pm$0.[”]{}08; 156$^\circ \pm$2$^\circ$\ N-2(2003) & 16$\rlap.^s$8706$\pm$0.0015 & 51$'$ 58$\rlap.{''}$339$\pm$0$\rlap.{''}$028 & $2.35\pm0.10$ & 1.[”]{}65$\pm$0.[”]{}09$\times$0.[”]{}80$\pm$0.[”]{}08; 162$^\circ \pm$5$^\circ$\ N-2(2006) & 16$\rlap.^s$8669$\pm$0.0011 & 51$'$ 58$\rlap.{''}$249$\pm$0$\rlap.{''}$019 & $2.24\pm0.07$ & 1.[”]{}47$\pm$0.[”]{}07$\times$0.[”]{}76$\pm$0.[”]{}07; 151$^\circ \pm$5$^\circ$\ N-3(2003) & 16$\rlap.^s$9142$\pm$0.0021 & 52$'$ 00$\rlap.{''}$776$\pm$0$\rlap.{''}$061 & $1.82\pm0.12$ & 2.[”]{}52$\pm$0.[”]{}17$\times$0.[”]{}62$\pm$0.[”]{}10; 168$^\circ \pm$3$^\circ$\ N-3(2006) & 16$\rlap.^s$9114$\pm$0.0022 & 52$'$ 00$\rlap.{''}$921$\pm$0$\rlap.{''}$052 & $1.76\pm0.10$ & 2.[”]{}52$\pm$0.[”]{}14$\times$0.[”]{}78$\pm$0.[”]{}09; 162$^\circ \pm$3$^\circ$\ E(2003) & 18$\rlap.^s$2348$\pm$0.0080 & 52$'$ 02$\rlap.{''}$183$\pm$0$\rlap.{''}$115 & 0.16$\pm$0.05 & $ \leq 1\rlap.{''}2^{\rm d}$\ E(2006) & 18$\rlap.^s$2128$\pm$0.0044 & 52$'$ 02$\rlap.{''}$319$\pm$0$\rlap.{''}$106 & 0.12$\pm$0.04 & $ \leq 1\rlap.{''}1^{\rm d}$\ N-4(2003) & 16$\rlap.^s$9645$\pm$0.0061 & 52$'$ 02$\rlap.{''}$650$\pm$0$\rlap.{''}$164 & $0.87\pm0.14$ & 2.[”]{}66$\pm$0.[”]{}43$\times$0.[”]{}88$\pm$0.[”]{}27; 171$^\circ \pm$9$^\circ$\ N-4(2006) & 16$\rlap.^s$9851$\pm$0.0033 & 52$'$ 03$\rlap.{''}$286$\pm$0$\rlap.{''}$092 & $0.86\pm0.09$ & 2.[”]{}40$\pm$0.[”]{}25$\times$0.[”]{}71$\pm$0.[”]{}16; 170$^\circ \pm$5$^\circ$\ N-5(2003) & 17$\rlap.^s$0764$\pm$0.0053 & 52$'$ 04$\rlap.{''}$881$\pm$0$\rlap.{''}$079 & $0.27\pm0.06$ & $\leq 1\rlap.{''}0^{\rm d}$\ N-5(2006) & 17$\rlap.^s$0715$\pm$0.0024 & 52$'$ 04$\rlap.{''}$737$\pm$0$\rlap.{''}$046 & $0.30\pm0.04$ & $\leq 0\rlap.{''}7^{\rm d}$\ D(2003) & 17$\rlap.^s$1265$\pm$0.0016 & 52$'$ 06$\rlap.{''}$840$\pm$0$\rlap.{''}$031 & 0.50$\pm$0.03 & $ \leq 0\rlap.{''}7^{\rm d}$\ D(2006) & 17$\rlap.^s$1211$\pm$0.0012 & 52$'$ 06$\rlap.{''}$833$\pm$0$\rlap.{''}$023 & 0.48$\pm$0.02 & $ \leq 0\rlap.{''}6^{\rm d}$\ Jet(2003) & 17$\rlap.^s$2097$\pm$0.0001 & 52$'$ 07$\rlap.{''}$142$\pm$0$\rlap.{''}$003 & 10.4$\pm$0.1 & 1.[”]{}21$\pm$0.[”]{}01$\times$0.[”]{}15$\pm$0.[”]{}03; 165$^\circ \pm$1$^\circ$\ Jet(2006) & 17$\rlap.^s$2093$\pm$0.0001 & 52$'$ 07$\rlap.{''}$150$\pm$0$\rlap.{''}$002 & 11.3$\pm$0.1 & 1.[”]{}20$\pm$0.[”]{}01$\times$0.[”]{}14$\pm$0.[”]{}02; 163$^\circ \pm$1$^\circ$\ B(2003) & 17$\rlap.^s$3946$\pm$0.0013 & 52$'$ 09$\rlap.{''}$675$\pm$0$\rlap.{''}$019 & 0.81$\pm$0.06 & $ \leq 0\rlap.{''}6^{\rm d}$\ B(2006) & 17$\rlap.^s$3904$\pm$0.0009 & 52$'$ 09$\rlap.{''}$732$\pm$0$\rlap.{''}$015 & 0.85$\pm$0.04 & $ \leq 0\rlap.{''}5^{\rm d}$\ C(2003) & 17$\rlap.^s$6549$\pm$0.0010 & 52$'$ 12$\rlap.{''}$150$\pm$0$\rlap.{''}$018 & 1.10$\pm$0.06 & $ \leq 0\rlap.{''}6^{\rm d}$\ C(2006) & 17$\rlap.^s$6603$\pm$0.0007 & 52$'$ 12$\rlap.{''}$221$\pm$0$\rlap.{''}$012 & 0.97$\pm$0.04 & $ \leq 0\rlap.{''}5^{\rm d}$\ S-1(2003) & 17$\rlap.^s$4519$\pm$0.0004 & 52$'$ 16$\rlap.{''}$301$\pm$0$\rlap.{''}$006 & $3.65\pm0.06$ & 0.[”]{}59$\pm$0.[”]{}04$\times$0.[”]{}40$\pm$0.[”]{}05; 51$^\circ \pm$9$^\circ$\ S-1(2006) & 17$\rlap.^s$4556$\pm$0.0002 & 52$'$ 16$\rlap.{''}$284$\pm$0$\rlap.{''}$003 & $4.97\pm0.05$ & 0.[”]{}66$\pm$0.[”]{}02$\times$0.[”]{}38$\pm$0.[”]{}02; 35$^\circ \pm$4$^\circ$\ S-2(2003) & 17$\rlap.^s$5481$\pm$0.0013 & 52$'$ 18$\rlap.{''}$113$\pm$0$\rlap.{''}$031 & 0.63$\pm$0.06 & $ \leq 0\rlap.{''}7^{\rm d}$\ S-2(2006) & 17$\rlap.^s$5473$\pm$0.0011 & 52$'$ 18$\rlap.{''}$210$\pm$0$\rlap.{''}$026 & 0.69$\pm$0.05 & $ \leq 0\rlap.{''}8^{\rm d}$\ $^{\rm a}$ Peak position. Right ascension ($\alpha$) given in hours, minutes, and seconds, and declination ($\delta$), given in degrees, arcmins, and arcsecs. The errors given are formal statistical errors. The absolute positional error of the images are estimated to be 0$\rlap .{''}$2 in right ascension and 0$\rlap .{''}$5 in declination. The sources are listed in order of decreasing declination. $^{\rm b}$ Total flux density. The errors given are formal statistical errors. The absolute flux density error is estimated to be on the order of 10%. $^{\rm c}$ Deconvolved dimensions of the source: FWHM major axis $\times$ FWHM minor axis; position angle of major axis. $^{\rm d}$ Unresolved [@lccccccccc]{} & D$^{\rm a}$ & PA(O)$^{\rm b}$ & PA(P)$^{\rm c}$ & $\mu(\alpha)^{\rm d}$ & $\mu (\delta)^{\rm e}$ & $\mu^{\rm f}$ & $PA(\mu)^{\rm g}$ & $\mu (R)^{\rm h}$ & $\mu (T)^{\rm i}$ Source & ([”]{}) & ($^\circ$) & ($^\circ$) & (mas yr$^{-1}$) & (mas yr$^{-1}$) & (mas yr$^{-1}$) & ($^\circ$) & (mas yr$^{-1}$) & (mas yr$^{-1}$) A & 12.215 & -7.76 & -15 & 18$\pm$43 & -65$\pm$118 & 67$\pm$114 & 164$\pm$97 & -67$\pm$114 & 0$\pm$53\ N-1 & 10.981 & -23.19 & -18 & -12$\pm$3 & -4$\pm$5 & 13$\pm$3 & -111$\pm$14 & -1$\pm$5 & -13$\pm$3\ N-2 & 9.560 & -22.96 & -21 & -15$\pm$8 & 34$\pm$13 & 37$\pm$12 & -24$\pm$19 & 36$\pm$12 & -5$\pm$8\ N-3 & 7.147 & -27.04 & -27 & -11$\pm$12 & -54$\pm$30 & 55$\pm$29 & -168$\pm$30 & -49$\pm$29 & -26$\pm$15\ E & 12.316 & 66.26 & -15 & -90$\pm$37 & -51$\pm$58 & 104$\pm$43 & -119$\pm$24 & -24$\pm$57 & -101$\pm$39\ N-4 & 5.239 & -30.98 & -31 & 85$\pm$28 & -237$\pm$70 & 252$\pm$67 & 160$\pm$15 & -251$\pm$68 & 16$\pm$34\ N-5 & 2.695 & -32.96 & -37 & -20$\pm$24 & 54$\pm$34 & 57$\pm$33 & -21$\pm$33 & 57$\pm$33 & -5$\pm$25\ D & 0.964 & -71.73 & -41 & -22$\pm$8 & 3$\pm$14 & 22$\pm$8 & -83$\pm$21 & 9$\pm$14 & -21$\pm$9\ Jet & 0.000 & 0.00 & -43 & -2$\pm$1 & -3$\pm$1 & 3$\pm$1 & -151$\pm$20 & -2$\pm$1 & -2$\pm$1\ B & 3.248 & 141.24 & 144 & -17$\pm$6 & -21$\pm$9 & 27$\pm$8 & -141$\pm$17 & 17$\pm$9 & 22$\pm$7\ C & 7.004 & 135.65 & 153 & 22$\pm$5 & -26$\pm$8 & 35$\pm$7 & 140$\pm$12 & 31$\pm$8 & -16$\pm$5\ S-1 & 9.538 & 163.79 & 159 & 15$\pm$2 & 6$\pm$3 & 16$\pm$2 & 67$\pm$7 & -3$\pm$2 & -16$\pm$2\ S-2 & 11.585 & 161.26 & 164 & -3$\pm$7 & -36$\pm$15 & 36$\pm$15 & -175$\pm$24 & 35$\pm$15 & 11$\pm$8\ $^{\rm a}$ Displacement with respect to the jet in arcsec. $^{\rm b}$ Observed position angle with respect to the jet in degrees. $^{\rm c}$ Predicted position angle in the precession model. $^{\rm d}$ Right Ascension proper motion in milliarcsec per year. (1 mas yr$^{-1}$ = 14 km s$^{-1}$ at a distance of 2.9 Kpc) $^{\rm e}$ Declination proper motion in milliarcsec per year. $^{\rm f}$ Total proper motion in milliarcsec per year. $^{\rm g}$ Position angle of the total proper motion in degrees. $^{\rm h}$ Radial proper motion with respect to an axis with PA = $-16^\circ$, in milliarcsec per year. A negative sign corresponds to motions approaching the jet. $^{\rm i}$ Transversal proper motion with respect to an axis with PA = $-16^\circ$, in milliarcsec per year. A positive sign corresponds to counterclockwise motion. $^{\rm j}$ Note that if the maps were aligned to make the position of the jet fixed, the proper motions of the other components would change by 2 and 3 mas yr$^{-1}$ in Right Ascension and Declination, respectively. [@lllll]{} & Distance & $S_\nu$(3.6 cm) & $\dot P$ &\ Source & (kpc) & (mJy) & ($M_\odot$ yr$^{-1}$ km s$^{-1}$) & References$^{\rm a}$\ IRAS 16547-4247 & 2.9 & 8.7 & $4.0\times10^{-1}$ & 1,2,3,4\ HH 80-81 & 1.7 & 3.0 & $5.3\times10^{-2}$ & 5,6\ Cep A HW2 & 0.725 & 6.9 & $5.4\times10^{-3}$ & 7,8\ $^{\rm a}$ 1) Garay et al. (2007); 2) Garay et al. (2003); 3) Rodríguez et al. (2005); 4) this paper; 5) Martí et al. (1993); 6) Yamashita et al. (1989); 7) Curiel et al. (2006); 8) Narayanan & Walker (1996). [^1]: Centro de Radioastronomía y Astrofísica, UNAM, Morelia 58089, México [^2]: NRAO is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc.	{ "pile_set_name": "ArXiv" }
--- abstract: \| We present observations of the molecular gas in the GN20 proto-cluster of galaxies at $z =4.05$ using the Expanded Very Large Array (EVLA). This group of galaxies is the ideal laboratory for studying the formation of massive galaxies via luminous, gas-rich starbursts within 1.6 Gyr of the Big Bang. We detect three galaxies in the proto-cluster in CO 2-1 emission, with gas masses (H$_2$) between $10^{10}$ and $10^{11} \times (\alpha/0.8)$ M$_\odot$. The emission from the brightest source, GN20, is resolved with a size $\sim 2"$, and has a clear north-south velocity gradient, possibly indicating ordered rotation. The gas mass in GN20 is comparable to the stellar mass ($1.3\times 10^{11} \times (\alpha/0.8)$ M$_\odot$ and $2.3\times 10^{11}$ M$_\odot$, respectively), and the sum of gas plus stellar mass is comparable to the dynamical mass of the system ($\sim 3.4\times 10^{11} [sin(i)/sin(45^o)]^{-2}$ M$_\odot$), within a 5kpc radius. There is also evidence for a tidal tail extending another $2"$ north of the galaxy with a narrow velocity dispersion. GN20 may be a massive, gas rich disk that is gravitationally disturbed, but not completely disrupted. There is one Lyman-break galaxy (BD29079) in the GN20 proto-cluster with an optical spectroscopic redshift within our search volume, and we set a 3$\sigma$ limit to the molecular gas mass of this galaxy of $1.1\times 10^{10} \times (\alpha/0.8)$ M$_\odot$. author: - 'C.L. Carilli, J. Hodge, F. Walter, D. Riechers, E. Daddi, H. Dannerbauer, G.E. Morrison' title: 'EVLA observations of a proto-cluster of molecular gas rich galaxies at $z = 4.05$' --- Introduction ============ Numerous lines of evidence support the hypothesis that massive elliptical galaxies form the majority of their stars quickly at early epochs, including: stellar population synthesis studies of nearby elliptical galaxies (Renzini 2006; Collins et al. 2009), a faster decline with cosmic time in the star formation rate per unit stellar mass with increasing galaxy mass (Moresco et al. 2010), and the direct observation of ‘red and dead’ ellipticals in clusters at high redshift (Kurk et al. 2009; Doherty et al. 2009; Andreon & Huertas-Company 2011). Plausible progenitors for these evolved galaxies at high redshift are even higher redshift submm galaxies (SMGs), corresponding to dusty, luminous starburst galaxies found in wide-field submm surveys (Blain et al. 2002). These hyper-luminous high-z galaxies may trace high over-densities, and are likely related to the formation of clusters and large ellipticals (Stevens et al. 2003; Aravena et al. 2010; although cf. Chapman et al. 2009; Williams et al. 2011). While the space density of SMGs peaks around $z \sim 2.3$ (Chapman et al. 2003), it is becoming clear that there is a substantial high-redshift tail of the SMG population, extending to $z > 4$ (Capak et al. 2011, Riechers et al. 2010; Wardlow et al. 2011). Daddi et al. (2009a) conclude, based on SMG space densities and duty cycles, that there are likely enough SMGs at $z > 3.5$ to account for the known populations of old massive galaxies at $z \sim 2$ to 3. A key question for the SMGs is: what drives the prolific star formation? Tacconi et al. (2006; 2008) argue, based on imaging of higher-order CO emission from a sample of $z \sim 2$ SMGs, that SMGs are predominantly nuclear starbursts, with median sizes $< 0.5"$ ($< 4$kpc), ‘representing extreme, short-lived, maximum star forming events in highly dissipative mergers of gas rich galaxies.’ This conclusion is supported by VLBI imaging of the star forming regions in two SMGs (Momjian et al. 2005; 2010). However, recent imaging of the lower order CO emission in a few SMGs (Ivison et al. 2010; Carilli et al. 2010; Riechers et al. 2011 in press), suggests that the lower-excitation molecular gas reservoirs can be significantly more extended. Daddi et al. (2009a) have identified a unique region in GOODS-N, containing the three galaxies: GN20, GN20.2a, and GN20.2b, at $z \sim 4.05 \pm 0.01$, within 20$''$ of each other (projected physical separation = 140 kpc). GN20 is among the brightest submm galaxies known, with a 350 GHz flux density of 23mJy. The galaxies GN20.2a and b are separated by only a few arcseconds, and hence are not spatially distinct in typical submm bolometer images. The two galaxies have a summed flux density of $S_{350\rm GHz} \sim 9.9$mJy. The implied IR luminosities (8 to 1000$\mu$m) are $L_{TIR} \ge 10^{13}$ L$_\odot$, with star formation rates are $> 1000$ M$_\odot$ yr$^{-1}$ (Daddi et al. 2009a). High order CO emission (4-3 and/or 6-5) has been detected from all three galaxies, although only marginally in the case of GN20.2b (Daddi et al. 2009a). The GN20 field also contains numerous Lyman Break Galaxies (LBGs) at similar redshift, and an overdensity of $z_{phot} > 3.5$ IRAC selected galaxies (Daddi et al. 2009a). Hence, this field offers an unprecedented opportunity to perform a definitive study of the gas distribution, kinematics, and physical conditions in a proto-cluster of galaxies within 1.6 Gyr of the Big Bang. In this paper we present a survey with the Expanded Very Large Array (EVLA; Perley et al. 2011), of the CO 2-1 emission from the GN20 proto-cluster. The low order transitions are critical for determining the total gas mass. We also search for CO emission from other galaxies in the cosmic volume surveyed. The GN20 proto-cluster: a massive galaxy formation laboratory ============================================================= The HST I-band image of the bright SMG GN20 shows diffuse emission about 1.5$"$ in extent (Daddi et al. 2009a), although offset from the radio and submm emission by $\sim 1"$, implying high obscuration of the most active star forming regions in the galaxy. GN20 is detected at 1.4 GHz, and resolved on a scale of $\sim 1.5"$, with a total flux density at $72\pm 13\mu$Jy (Morrison et al. 2010; Casey et al. 2009). High resolution imaging of the 850$\mu$m emission shows resolved structure, with a north-south extension possibly as large at 1.5$"$ (Younger et al. 2008; Iono et al. 2006). The 6.2$\mu$m PAH spectral feature has been detected in GN20 using Spitzer (Riechers et al. in prep). The integrated CO 1-0 and 2-1 emission from GN20 has been imaged at resolutions down to $0.2"$ with the VLA (Carilli et al. 2010). The molecular gas is extended on a scale of at least 1.5$"$, and at high resolution forms a partial ring, or disk. Unfortunately, the old VLA correlator provided no velocity information. The excitation of the integrated CO emission from GN20 is higher than the Milky Way, but lower than high redshift quasar hosts and the nuclear starburst regions of IR-luminous galaxies. The CO 4-3 line strength in GN20 is more than a factor of two lower than expected for thermal excitation (Carilli et al. 2010). GN20.2a shows a complex and extended morphology in the HST images (Daddi et al. 2009a), and it is a relatively strong radio source ($S_{1.4} = 181\mu$Jy), likely corresponding to a radio AGN. The SED in the optical through near-IR is consistent with a star forming galaxy. GN20.2a has been detected in CO 4-3 emission by Daddi et al. (2009a). Daddi et al. derive an IR luminosity (8 to 1000$\mu$m) of $L_{IR} = 1.6\times 10^{13}$ L$_\odot$. GN20.2b is faint and compact in the HST image. It is also detected in the radio continuum at $S_{1.4} = 32\mu$Jy, consistent with star formation. CO 4-3 emission is marginally detected with the Plateau de Bure Interferometer, and appears to be broad (700 km s$^{-1}$; Daddi et al. 2009a). Daddi et al. (2009a) find 15 B-band dropout galaxies in a 25$''$ radius centered on GN20, an overdensity of a factor of 6 compared to the full GOODS-N area, which is significant at the 7$\sigma$ level. A spike in the redshift distribution of galaxies at $z=4.06\pm0.02$ is observed in all of GOODS-N (13 spectroscopic redshifts in total at this redshift). Lastly, the SMG GN10 at $z = 4.04$ is located 9$'$ from GN20 (Daddi et al. 2009b; Dannerbauer et al. 2008). Therefore, it appears that the GN20 volume has a very significant overdensity, indicating a proto-cluster environment at $z\sim 4.05$. Daddi et al. (2009a) estimate a total mass for this structure of $\sim 10^{14}$ M$_\odot$. Observations ============ We observed the GN20 field with the Expanded Very Large Array in the D (1km) configuration in March and April, 2010. Observations were made of the CO 2-1 line, using a total bandwidth of 246 MHz and in each of two polarizations, centered at 45.655 GHz. The velocity coverage is 1600 km s$^{-1}$, including the CO 2-1 lines in the three galaxies. The total observing time was 28 hours, but only about 22 to 24 antennas were available at a time due to on-going EVLA work. Dynamic scheduling to ensure good weather is now standard at the EVLA. Fast switching phase calibration was employed (Carilli & Holdaway 1999) on timescales between two and three minutes using the VLA calibrator J1302+5748. Data were edited to remove time ranges of poor phase stability. The source 3C286 was used for flux density calibration. Standard EVLA data calibration and editing was performed using AIPS. After calibration and data editing, images were synthesized using the robust weighting scheme of the uv data with R=2 (Cornwell et al. 1999). The final resolution for the images was $1.9"$. Spectral line cubes were generated at 12MHz per channel (78 km s$^{-1}$). The rms noise per channel at 12MHz resolution is 0.11 mJy beam$^{-1}$. All images were corrected for the VLA primary beam response, which has a FWHM $\sim 1'$ at 45 GHz. The field pointing center was located $10"$ west of GN20. Results ======= Figure 1 shows the CO 2-1 emission integrated over the full frequency range covered by these observations for the GN20 group. The crosses mark the 1.4 GHz continuum positions of the three galaxies GN20, 20.2a,b (Morrison et al. 2010), and the optical position of one LBG with a spectoscopic redshift in the redshift range of the CO 2-1 observations (Section 5.2). We have added the previously published D array observations (Carilli et al. 2010) to improve signal-to-noise. Unfortunately, the previous data did not have the exact same velocity coverage, and hence the velocity-integrated flux densities are not accurate. Hence, this image acts as a finding chart for the CO emitting regions, but accurate line fluxes will be based on subsequent analysis of the new EVLA data itself. We detect CO 2-1 emission from GN20, GN20.2a, and GN20.2b. GN20 is extended on a scale $\sim 2"$ (see below). GN20.2a is not spatially resolved by these observations. GN20.2b may be extended, although the signal-to-noise of these data are not conclusive. Figure 2 shows spectra of the three sources. For GN20, we integrate over an area of $2"$. For the other two sources we only consider the spectra at the peak of the velocity integrated CO image. Results for Gaussian fits to the lines are given in Table 1. The lines are broad for GN20 and GN20.2a ($\sim 800$ km s$^{-1}$), although not outside the range of the zero intesity line widths seen in some nearby ULIRGs (Downes & Solomon 1998). However, we note athat, due to the bandwidth limitation for the early EVLA, the very lowest frequency (highest velocity) emission for GN20 may be truncated by these observations, although the higher order CO transitions show that we expect little emission beyond our lowest spectral channel (Daddi et al. 2009a). GN20.2b shows a narrower profile at the peak position. However, the detection in the velocity integrated image over the full frequency range (Figure 1), suggests that the emission may be extended both spatially and spectrally. A possibly broader line is also suggested for the higher order CO emission, but again, the high order CO detection is only marginal (Daddi et al 2009a). We have investigated this possibility with the current data, and conclude that more sensitive observations are required to better characterize the CO emission from GN20.2b, and we simply present the spectrum at the peak position herein. Figure 3 shows the velocity channel images for the GN20 CO 2-1 emission at 78 km s$^{-1}$ channel$^{-1}$ and $1.9"$ resolution. The emission clearly moves from the south to the north with increasing frequency. An interesting feature is seen in channel 6, corresponding to a frequency of 45.598 GHz (or $z =4.05588$). The emission in channel 6 appears to be substantially more extended than in other channels. Inspecting our bandpass calibration, we find no channel-dependent calibration errors that would lead to such a difference in morphology for this single channel. In channel 6 the emission extends well to the north, with a total of extent of $\sim 4"$. Admittedly, the signal-to-noise is not high, but this extended emission is suggestive of a gravitationally induced tidal feature. Such tidal tails in nearby galaxies often occur over a narrow velocity range (Hibbard & Mihos 1995). Figure 4a shows the CO 2-1 emission from GN20 integrated over the velocity range shown in the channel maps of Figure 3. The emission appears extended, in particular to the southwest. A formal Gaussian fit to the emission yields a total flux density of $0.69 \pm .07$ mJy, a peak surface brightness of $0.49\pm 0.03$ Jy beam$^{-1}$, and a deconvolved source size of $1.5" \times 0.9"$ with major axis position angle PA = 69$^o$. Figure 4b shows the iso-velocity contours (first moment) of the CO 2-1 emission, ie. the weighted mean velocity derived after first blanking the channel images below 3$\sigma$ per channel. There is a velocity gradient north-south, with a magnitude of $\pm 250$ km s$^{-1}$. Given the low spatial resolution of these images, this corresponds to a lower limit to the true projected rotational velocity. Analysis ======== Masses ------ We derive the mass in molecular gas (H$_2$) from the observed CO 2-1 luminosities using the standard relationships in Solomon & Vanden Bout (2005). We extrapolate from CO 2-1 to 1-0 luminosity assuming constant brightness temperature, which is certainly correct to within 10% for these galaxies (Carilli et al. 2010). We adopt a CO luminosity to H$_2$ conversion factor of $\alpha = 0.8$ M$_\odot$ K km s$^{-1}$ pc$^2$, appropriate for nearby nuclear starburst galaxies (Downes & Solomon 1998). The resulting values are listed in Table 1. Admittedly, the uncertainty on $\alpha$ is significant. A minimum gas mass can be derived assuming optically thin emission, and adopting a temperature and a CO to H$_2$ abundance ratio. Ivison et al. (2010) show that for reasonable assumptions (Galactic CO abundance in molecular clouds, and temperatures 40 K to 50 K), the lower limit to $\alpha \sim 0.65$ (see also Aalto et al. 1995). The gas mass in GN20 is $1.3\times 10^{11} \times (\alpha/0.8)$ M$_\odot$. Daddi et al. (2009a) derive a stellar mass of $2.3\times 10^{11}$ M$_\odot$ from IR through optical SED fitting. We can obtain a very rough estimate of the gravitational mass of GN20 from the CO velocity field, although there are substantial uncertainties due to the low spatial resolution of these observations. For the rotational velocity, we adopt the extremes of the velocity channels with significant emission in Figure 3, or $390/sin(i)$ km s$^{-1}$. For the radius, we adopt the value for the CO ring seen in the high resolution 2-1 observations presented in Carilli et al. (2010), $\sim 5$ kpc. The inclination angle remains uncertain. The gravitational mass inside this radius is $\sim 3.4\times 10^{11} [sin(i)/sin(45^o)]^{-2}$ M$_\odot$. This naively derived dynamical mass is comparable to the sum of the stellar and gas masses. Lyman-break galaxy molecular mass limits ---------------------------------------- There is one Lyman-break galaxy within our field with a spectroscopic redshift placing the CO 2-1 line within our band: J123711.48+622155.8 at $z = 4.058$ (BD29079; Daddi et al. 2009a; Shim et al. 2011). We find no CO emission from this galaxy, to a $1\sigma$ limit of 0.17 mJy beam$^{-1}$ at 78 km s$^{-1}$ channel$^{-1}$, after primary beam correction. Convolving to 312 km s$^{-1}$, we set a 3$\sigma$ limit to the CO 1-0 luminosity of $L'_{CO1-0} < 1.4\times 10^{10} \rm K~ km~ s^{-1}~ pc^2$, assuming constant brightness temperature from CO 2-1 to 1-0. The implied molecular gas mass of this galaxy is $< 1.1\times 10^{10} \times (\alpha/0.8)$ M$_\odot$. The stellar mass of this galaxy is $2.6\times 10^{10}$ M$_\odot$ (Daddi et al. 2009a), hence the gas-to-stellar mass ratio is $< 0.53 \times (\alpha/0.8)$. Daddi et al. (2009a) estimate a star formation rate of 150 M$_\odot$ year$^{-1}$ for BD29079, or an IR luminosity of $1.5\times 10^{12}$ L$_\odot$. Daddi et al. (2010) have considered the ’star formation law’ in high $z$ galaxies, ie. the relationship between IR luminosity and CO luminosity, and find for galaxies of this luminosity at $z \sim 2$: $L'_{CO1-0} \sim 0.02 L_{FIR}~\rm K~ km~ s^{-1}~ pc^2 $. Adopting this relation implies an expected CO luminosity of $L'_{CO1-0} \sim 3\times 10^{10}$ L$_\odot$. Hence, BD29079 is under-luminous in CO emission relative to the standard star formation law, although within the broad (factor three) scatter at any given luminosity. There are 11 other LBGs within our field, and with photometric redshifts that are consistent with the sampled range, although the $z_{phot}$ error bars are such that the galaxies could be outside the velocity range of the CO observations. We have looked for CO emission at all these positions, and find no significant detections to similar limits to those quoted above. We emphasize that there is a large uncertainty in the derived H$_2$ mass for the LBGs, due to the unknown conversion factor, $\alpha$. It is possible that metal poor star forming galaxies have a much larger value of $\alpha$ than ULIRGs or even the Milky Way due to CO dissociation by the more pervasive UV radiation field (Madden et al. 1997; Papadopoulos & Pelupessy 2010). Discussion ========== We have detected CO 2-1 emission from three galaxies in the GN20 proto-cluster at $z = 4.05$ using the EVLA. The molecular gas masses range from $1.9\times 10^{10} \times (\alpha/0.8)$ M$_\odot$ to $1.3\times 10^{11} \times (\alpha/0.8)$ M$_\odot$. Hence, we are observing a group of molecular gas rich galaxies undergoing extreme starburts within 1.6 Gyr of the Big Bang. GN20 presents a particularly interesting case, given its very high luminosity, and large spatial extent. We find that the sum of the stellar and gas mass in GN20 is comparable to the dynamical mass. While the assumptions involved in all three mass calculations are uncertain, it appears that the baryons likely dominate the mass content within 5 kpc radius of GN20. Moreover, their is little room for a CO luminosity to gas mass conversion factor substantially larger than 0.8 M$_\odot$ K km s$^{-1}$ pc$^2$. High resolution observations are required to confirm the rotational aspect of the emission, and to obtain a more accurate derivation of the dynamical mass. The morphology and velocity field for GN20 suggest a rotating disk, with a possible tidal tail extending 10kpc to the north in a single 78 km s$^{-1}$ channel. We can speculate that GN20 is a gas rich disk galaxy that has been gravitationally torqued by its neighboring galaxies. This gravitational disturbance has not distrupted the disk, but has greatly enhanced star formation in the disk, as well as generating the extended, tidal gas distribution. Lastly, we have searched for CO emission from LBGs in the GN20 protocluster. We set an upper limit of $1.1\times 10^{10} \times (\alpha/0.8)$ M$_\odot$ to the molecular gas mass in these galaxies, including one with an accurate spectroscopic redshift. Our limit is below the expected value based on the standard star formation law at this redshift, but within the broad scatter. We are obtaining high resolution CO imaging observations of the GN20 proto-cluster with the EVLA and the Plateau de Bure interferometer to perform a more detailed dynamical analysis, and study the spatial distribution of the CO excitation and gas-to-dust ratio of this forming cluster of massive galaxies at $z = 4.05$. We thank the referee for useful comments. ED acknowledges the funding support of ERC-StG-UPGAL-240039, and ANR-08-JCJC-0008. DR acknowledges support from from NASA through Hubble Fellowship grant HST-HF-51235.01 awarded by the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., for NASA, under contract NAS 5-26555. \[\] [\|c\|c\|c\|c\|c\|c\|c\|c\|]{} Source & Redshift & Peak & FWHM & I$_{CO}$ & Velocity$^a$ & L’CO & M(H$_2$)\ & & mJy & km s$^{-1}$ & Jy km s$^{-1}$ & km s$^{-1}$ & K km s$^{-1}$ pc$^2$ & M$_\odot$\ GN20 & 4.0554 & 1.21$\pm$0.11 & 679$\pm$88 & 0.87$\pm$0.088 & -11$\pm$32 & $1.6\times 10^{11}$ & $1.3\times10^{11}$\ GN20.2a & 4.0508 & 0.54$\pm$0.065 & 723$\pm$110 & 0.41$\pm$0.050 & -283$\pm$45 & $7.6\times 10^{10}$ & $6.1\times10^{10}$\ GN20.2b & 4.0567 & 0.42$\pm$0.083 & 290$\pm$70 & 0.13$\pm$0.025 & 67$\pm$29 & $2.4\times 10^{10}$ & $1.9\times10^{10}$\ $^a$Velocity relative to $z = 4.0556$. Aalto, S. et al. 1995, A& A, 300, 369 Andreon, S. & Huertas-Company, M. 2011, A& A, 526, 11 Aravena, M., Bertoldi, F., Carilli, C. et al. 2010, ApJL, 708, L36 Blain, A., Smail, I, Ivison, R., Kneib, J.-P., Frayer, D. 2002, PhR 369 111 Capak, P. et al. 2011, Nature, 470, 233 Carilli, C.L. & Holdaway, M. 1999, Radio Science, 34, 817 Casey, C.M., Chapman, S., Daddi, E. et al. 2009, MNRAS, in press Chapman, S., Blain, A., Ivison, R., Smail, I. 2003 Nature 422 695 Chapman, S., Blain, A., Ibata, R. et al. 2009, ApJ, 691, 560 Collins, D., Stott, J.P., Hilton, M. et al. 2009, Nature, 458, 603 Cornwell, T., Braun, R., Briggs, D. 1999, in [Synthesis Imaging in Radio Astronomy II]{}, eds. G. B. Taylor, C. L. Carilli, and R. A. Perley (ASP: San Francisco) 180, 151 Daddi, E., Dickinson, M., Chary, R. et al 2005, ApJ 631 L13 Daddi, E., Dannerbauer, H., Stern, D. et al. 2009a, ApJ, 694, 1517 Daddi, E., Dannerbauer, H., Krips, M. et al. 2009b, ApJ, 695, L176 Daddi et al. 2010, ApJ, 714, L118 Dannerbauer, H., Walter, F., Morrison, G. 2008, ApJ, 673, L127 Downes & Solomon 1998 ApJ, 507, 615 Doherty, M., Tanaka, M., de Breuck, C. et al. 2009, A& A, 509, 83 Hibbard, J. & Mihos, C. 1995, AJ, 110, 140 Iono, D., Peck, A., Pope, A. et al. 2006, ApJ, 640, L1 Ivison, R., Papadopoulos, P, Smail, I. et al. 2010, MNRAS, 412, 1913 Kurk, J., Cimatti, A., Zamorani, G. et al. 2009, A& A, 504, 331 Madden, S., Poglitsch, A., Geiss, N. et al. 1997, 483, 200 Momjian, E., Carilli, C. & Petric, A. 2005, AJ, 129, 1809 Momjian, E. et al. 2010, AJ, 139, 1622 Moresco, M. et al. 2010, A& A, 524, 67 Morrison, G., Owen, F., Dickinson, M., Ivison, R., Ibar, E. 2010, ApJ, 188, S178 Papadopoulos, P. & Pelupessy, F. 2010, ApJ, 717, 1037 Perley, R.A. et al. 2011, ApJL, this volume Renzini, A. 2006, ARAA, 44, 141 Riechers, D. et al. 2010, ApJL, 720, L131 Solomon & Vanden Bout 2005, ARAA, 43, 677 Stevens, J, Ivison, R., Dunlop, J. et al. 2003 Nature, 425, 264 Tacconi, L., Neri, R., Chapman, S. et al. 2006, ApJ, 640, 228 Tacconi, L., Genzel, R., Smail, I., et al. 2008, ApJ, 680, 246 Wardlow, J. et al. 2011, MNRAS, in press Williams, C.C., Giavaisco, M., Porciani, C. et al. 2011, ApJ, in press Younger, J., Fazio, G., Wilner, D. et al. 2008, ApJ, 688, 59	{ "pile_set_name": "ArXiv" }
--- abstract: 'We study the problem of anonymizing tables containing personal information before releasing them for public use. One of the formulations considered in this context is the $k$-anonymization problem: given a table, suppress a minimum number of cells so that in the transformed table, each row is identical to atleast $k-1$ other rows. The problem is known to be NP-hard and MAXSNP-hard; but in the known reductions, the number of columns in the constructed tables is arbitrarily large. However, in practical settings the number of columns is much smaller. So, we study the complexity of the practical setting in which the number of columns $m$ is small. We show that the problem is NP-hard, even when the number of columns $m$ is a constant ($m=3$). We also prove MAXSNP-hardness for this restricted version and derive that the problem cannot be approximated within a factor of $\frac{6238}{6237}$. Our reduction uses alphabets $\Sigma$ of arbitrarily large size. A natural question is whether the problem remains NP-hard when both $m$ and $\|\Sigma\|$ are small. We prove that the $k$-anonymization problem is in $P$ when both $m$ and $\|\Sigma\|$ are constants.' author: - 'Venkatesan T. Chakaravarthy, Vinayaka Pandit, Yogish Sabharwal' bibliography: - 'anon.bib' date: \| IBM Research - India, New Delhi and Bengaluru.\ [{vechakra, pvinayak, ysabharwal}@in.ibm.com]{} title: ' On the Complexity of the $k$-Anonymization Problem ' --- Introduction {#sec:intro} ============ Various organization such as hospitals and insurance companies collect massive amount of personal data. These need to be released publicly for the purpose of scientific data mining; for instance, data collected by hospitals could be mined to infer epidemics. However, a major risk in releasing personal data is that they can be used to infer sensitive information about individuals. A natural idea for protecting privacy is to remove obvious personal identifiers such as social security number, name and driving license number. However, Sweeney [@Swe02] showed that such a deidentified database can be joined with other publicly available databases (such as voter lists) to reidentify individuals. For instance, she showed that 87% of the population of the United States can be uniquely identified on the basis of gender, date of birth and zipcode. In the literature, such an identity leaking attribute combination is called a quasi-identifier. It is important to recognize quasi-identifiers and apply protective measures to eliminate the risk of identity disclosure via join attacks. Samaratti and Sweeyney [@Sam-Swe; @Swe02] introduced the notion of $k$-anonymity, which aims to preserve privacy either by suppressing or generalizing some of the sensitive data values. In this paper, we consider the basic $k$-anonymity problem with only suppression allowed. Suppose we have a table with $n$ rows and $m$ columns. In order to achieve anonymity, one is allowed to suppress the entries of the table so that in the modified table, every row is identical to at least $k-1$ other rows. The goal is to minimize the number of cells suppressed. This is called the [$k$-anonymization]{} problem. The motivation for the problem formulation are twofold: (i) any join attack would return groups of at least $k$ rows, thus preserving privacy with a parameter of $k$; (ii) lesser the number of entries suppressed, better is the value of the modified table for data mining. [Example: ]{}We now illustrate the problem definition with an example. An example input table and its anonymized output, for $k=2$, are shown in Figure \[fig:example\]. The number of rows is $n=4$ and number of columns is $m=3$. The suppressed cells are shown by “$$”. We see that in the anonymized output table, the first and the third rows are identical, and the second and the fourth rows are identical. Thus the table on the right is $2$-anonymized. The cost of this anonymization is $4$, since $4$ cells are suppressed. This is an optimal solution. [ccccc]{} $x$ $a$ $b$ ----- ----- ----- $z$ $c$ $d$ $y$ $a$ $b$ $z$ $c$ $e$ & & & & $$ $a$ $b$ ----- ----- ----- $z$ $c$ $$ $$ $a$ $b$ $z$ $c$ $$ \ &&&&\ Original table & & & & $2$-Anonymized table [Known and New Results:* ]{} Meyerson and Williams [@MW] proved the NP-hardness of the $k$-anonymization problem. Aggarwal et al. [@ICDT] improved the result by showing that the problem remains NP-hard even when the alphabet $\Sigma$ from which the symbols of the table are drawn is fixed to be ternary. Bonizzoni et al [@APX-hard] proved MAXSNP-hardness (and NP-hardness) even when the alphabet is binary. The value of the privacy parameter $k$ is a fixed constant in all the above results ($k=3$). On the algorithmic front, Meyerson and Williams gave a $O(k\log k)$-approximation algorithm. This was improved by Aggarwal et al. [@ICDT], who devised a $O(k)$-approximation algorithm. Park and Shim [@SIGMOD] presented an approximation algorithm with a ratio of $O(\log{k})$; however, we observe that the running time of their algorithm is exponential in the number of columns $m$ (but, polynomial in the number of rows $n$). We make the following observations regarding the previously known results. Firstly, the known NP-hardness reductions produce tables in which the number of columns is arbitrarily large. This is not satisfactory as the number of columns in practical settings is not large. Secondly, the algorithm of Park and Shim [@SIGMOD] is a polynomial time $O(\log k)$-approximation algorithm when the number of columns $m$ is small ($m=O(\log n)$). These observations raise a natural question: Does the $k$-anonymization problem remain NP-Complete even when the number of columns $m$ is small ($\log n$ or a constant)? We show that the $k$-anonymization problem remains NP-hard, even when the number columns $m$ is fixed to be a constant ($m=3$). In fact, we also show that the above restricted version is MAXSNP-hard, thus ruling out polynomial time approximation schemes. We also derive that the problem cannot be approximated within a factor of $\frac{6238}{6237}$. Even though our inapproximability bound is mild, it is the first explicit inapproximability bound proved for the $k$-anonymization problem. All our hardness results hold even when the privacy parameter $k$ is a constant ($k=7$). As we noted, the previous constructions ensured that the alphabet size is a fixed constant; but, in our constructions, the alphabet size is not a fixed constant, but it is arbitrarily large. However, this is not a serious issue; in most settings, tables have large number of unique entries (for example, a zipcode column takes a large number of distinct values). In the wake of previous results and our results mentioned above, a natural question is whether the problem is NP-hard when both the number of columns $m$ and the alphabet size $\|\Sigma\|$ are small. We show that the problem can be solved optimally in polynomial time when both $m$ and $\|\Sigma\|$ are fixed constants. Below, we summarize the new results proved in this paper. - The $k$-anonymization problem is NP-hard even when the number of columns $m$ and the privacy parameter $k$ are constants ($m=3$, $k=7$). We show that the restricted version is MAXSNP-hard and cannot be approximated within a factor of $\frac{6238}{6237}$. - When both the number of columns and the alphabet size are fixed constants, the problem can be solved optimally in polynomial time. The result is true, even when the privacy parameter $k$ is arbitrarily large. The known and new results put together provide a complete picture on the complexity of the problem for the four cases of the number of columns $m$ and the alphabet size $\Sigma$ being fixed constants or arbitrarily large. This is shown in Figure \[fig:summary\]. -- ------------------------------------------------------------------------------------------------------------------------------------------------- -- $\mathbf{\|\Sigma\|}$[=constant]{} & $\mathbf{\|\Sigma\|}$[=arbitrary]{}\ & &\ & &\ $\mathbf{m}$[=constant]{} & Solvable in P (even when $k$ is arbitrary) \[This paper\] & NP-hard (even when $k$ is a constant) \[This paper\]\ & &\ & &\ $\mathbf{m}$[=arbitrary]{} & NP-hard (even when $k$ is a constant) [@ICDT] & NP-hard (even when $k$ is a constant) [@ICDT]\ & &\ ******** -- ------------------------------------------------------------------------------------------------------------------------------------------------- -- Problem Definition ================== The input to the $k$-anonymization problem is a $n\times m$ table $T$ having $n$ rows and $m$ columns, with symbols of the table drawn from an alphabet $\Sigma$. The input also includes a [privacy parameter]{} $k$. A feasible solution $\sigma$ transforms the given table $T$ to a new table $T'$ by suppressing some of the cells of $T$; namely, it replaces some of the cells of $T$ with “$$”. In the transformed table $T'$, for any row $t$, there should exist $k-1$ other rows that are identical to $t$. The cost of the solution, denoted ${{\rm Cost}}(\sigma)$, is the number of suppressed cells. The goal is to find a solution having the minimum cost. Consider a solution $\sigma$. For a row $t$, we denote by ${{\rm Cost}}(t)$ the number of suppressed cells in $t$ and say that $t$ pays this cost. Thus, ${{\rm Cost}}(\sigma)$ is the sum of costs paid by all the rows. There is an equivalent way to view a solution in terms of partitioning the given table. Consider a subset of rows $S$. We say that a column is [good]{} with respect to $S$, if all the rows in $S$ take identical values on the column. A column is said to be [bad]{}, if it is not good; meaning, some two rows in $S$ have different values on the given column. Denote by $a(S)$ the number of bad columns in $S$. Then, our goal is to a find a partition of rows $\Pi={S_1, S_2, \ldots, S_{\ell}}$ such that each set $S_i$ is of size $\|S_i\|\geq k$. Each row $t$ in $S_i$ pays a cost of $a(S_i)$. The total cost of the solution is the sum of costs paid by all rows. Equivalently, the cost of the solution is given by $\sum_{i=1}^{\ell} \|S_i\|\cdot a(S_i)$. We shall interchangeably use either of the two descriptions in our discussions. Hardness Results with Three Columns =================================== In this section, we present results on the complexity of $k$-anonymization problem when both the number of columns and the privacy parameter are constants. NP-Hardness ----------- \[thm:AAA\] The $k$-anonymization problem is NP-hard even when the number of columns $m$ is 3 and the privacy parameter is fixed as $k=7$. [[Proof:* ]{}]{}We give a reduction from the vertex cover problem on 3-regular graphs, which is known to be NP-hard (see [@GJ]). Recall that a vertex cover of a graph refers to a subset of vertices such that each edge has at least one endpoint in the subset and that a graph is said to be 3-regular, if every vertex has degree exactly 3. Let $G=(V,E)$ be the input 3-regular graph having $r$ vertices. The alphabet of the output table is as follows: For each vertex $u \in V$, we add a symbol $u$. Next, we have additional symbols ‘$0$’ and ‘$Z$’. Further, we need a number of [special]{} symbols. A special symbol appears only once in the whole of the table. The exposition becomes somewhat clumsy, if we explicitly introduce these special symbols. Instead, we use the generic symbol ‘$?$’ to mean the special symbols. The symbol ‘$?$’ is not a single symbol, but a general placeholder to mean a special symbol. We maintain a running list of special symbols (say $s_1, s_2, \ldots$) and whenever a new row containing ‘$?$’ is added to the table, we actually get a new symbol from the list and replace ‘$?$’ by the new symbol. For instance, suppose $\langle ?, u, u\rangle$ and $\langle ?, v, ?\rangle$ are the first two rows added to the table. Then, the actual rows added are $\langle s_1, u, u\rangle$ and $\langle s_2, u, s_3\rangle$. With the above discussion in mind, notice that two different instances of ‘$?$’ do not match with each other. The output table $T$ is constructed as follows. 1. For each vertex $u\in V$, add the following 20 rows. These are said to be rows corresponding to $u$. 1. Add the following row six times: $\langle 0, u, u\rangle$. \[type:AAA\] 2. Add a row $\langle ?, u, u\rangle$. It is called the [critical]{} row of $u$ and it plays a vital role in the construction. \[type:BBB\] 3. Add seven rows: $\langle ?, u, ?\rangle$. \[type:CCC\] 4. Add the following row 3 times: $\langle 0, u, Z\rangle$. \[type:DDD\] 5. Add the following row 3 times: $\langle 0, Z, u\rangle$. \[type:EEE\] 2. For each edge $(x, y)$, add two rows $\langle 0, x, y\rangle$ and $\langle 0, y, x\rangle$. These are called [edge rows]{}. 3. Add the following two sets of [dummy rows]{}: 1. Add seven rows as below: $\langle 0, ?, Z\rangle$. \[type:DDD-A\] 2. Add seven rows as below: $\langle 0, Z, ?\rangle$. \[type:DDD-B\] This completes the construction of the table. The privacy parameter is set as $k=7$. Consider any $k$-anonymization solution to the constructed table. For any row of the table, we can derive a lowerbound on the cost paid by the row; we refer to the lowerbound as the [base cost]{}. The base costs are derived as follows, for the various types of rows. Consider any vertex $u\in V$. First consider rows of type \[type:AAA\]. These rows are of the form $\langle 0, u, u\rangle$ and there are exactly six of them. Since $k=7$, these rows must be participating in a cluster having a different row. Hence, each of these rows must pay a cost of at least 1. We set the base cost for each of these rows to be 1. Now, consider the critical row of type \[type:BBB\]. This row is of the form $\langle ?, u, u\rangle$ and it must pay a base cost of 1, since it involves a special symbol. The base cost of the critical row is deemed to be 1. A type \[type:CCC\] row (of the form $\langle ?, u, ?\rangle$) must pay cost of at least two, since it has two special symbols. The base cost of such a row is deemed to be 2. By similar arguments, we see that any other type of row must pay a base cost of 1. To summarize, every row of type \[type:CCC\] (of the form $\langle ?, u, ?\rangle$) pays a base cost of 2, whereas any row of any other type pays a base cost of 1. For each vertex $u$, the total base cost across the 20 rows can be calculated as follows: (i) The six (type \[type:AAA\]) rows of the form $\langle 0, u, u\rangle$ pay a cost of $6$ in total; (ii) The critical row (of type \[type:BBB\]) pays a cost of $1$; (iii) The seven (type \[type:CCC\]) rows of the form $\langle ?, u, ?\rangle$ pay a cost of $2$ each, totaling $14$; (iv) The three (type \[type:DDD\]) rows of the form $\langle 0, u, Z\rangle$ pay cost of $3$ in total; (v) The three (type \[type:EEE\]) rows of the form $\langle 0, Z, u \rangle$ pay cost of $3$ in total. Thus, the total base cost for each vertex $u$ is $27$. Then, each edge has a base cost of 2, coming from the two rows corresponding to it. The two blocks dummy rows (of type \[type:DDD-A\] and type \[type:DDD-B\]) contribute a base cost of $7$ each, summing up to $14$. Thus, the [aggregated base cost]{} is $ABC = 27r + 2\|E\|+14$. For any row, the difference between the actual cost paid and the base cost is denoted as [extra cost]{}. Similarly, the total extra cost is the sum of extra costs over all the rows. Notice that the cost of the solution is the sum of $ABC$ and the total extra cost. We claim that the given graph has a vertex cover of size $\leq t$, if and only if there exists a $k$-anonymization solution with an extra cost $\leq t$. It would follow that the graph has a vertex cover of size $\leq t$, if and only if there exists a $k$-anonymization solution of cost $\leq ABC+t$. This would prove the required NP-hardness. We next proceed to prove the above claim. We split the proof into two parts. First, we shall argue that if the given graph has a vertex cover of size $\leq t$, then there exists a $k$-anonymization solution with extra cost $\leq t$. Suppose $C$ is a vertex cover of size $\leq t$. We shall construct a $k$-anonymization solution $\sigma$ in which the critical rows corresponding to the vertices in the cover $C$ pay an extra cost of 1 and every other row pays no extra cost. For each edge $(x, y)$, if $x\in C$, then [attach]{} the edge to $x$, else attach it to $y$. (If both the endpoints of the edge are in the cover, the edge can be attached arbitrarily to any one of the two vertices). Without loss of generality, assume that each vertex in the cover has at least one edge attached to it. Otherwise, the vertex can be safely removed from $C$, yielding a smaller cover. -------------------- ---------------- Case: $u\not\in C$ Case: $u\in C$ -------------------- ---------------- Form a $k$-anonymization solution $\sigma$ as follows. See Figure \[fig:hard\] for an illustration. - Form two clusters combining the dummy rows. - Form a cluster by combining the seven (type \[type:DDD-A\]) dummy rows of the form $\langle 0, ?, Z\rangle$; call this cluster $D_1$. - Form a cluster by combining the seven (type \[type:DDD-B\]) dummy rows of the form $\langle 0, Z, ?\rangle$; call this cluster $D_2$. - Consider each vertex $u$ not in the cover $C$ (i.e., $u\not\in C$). - Form a cluster by adding the six (type \[type:AAA\]) rows of the form $\langle 0, u, u\rangle$ and the critical row $\langle ?, u, u\rangle$. Each row in the cluster pays a cost of 1, and hence the extra cost is 0 for all these rows. - Form a cluster by adding the seven (type \[type:CCC\]) rows of the form $\langle ?, u, ?\rangle$. Each row in the cluster pays a cost of 2, and hence the extra cost is 0 for all these rows. - Add the three (type \[type:DDD\]) rows of the form $\langle 0, u, Z\rangle$ to $D_1$. Add the three (type \[type:EEE\]) rows of the form $\langle 0, Z, u\rangle$ to $D_2$. Thus, each row in $D_1$ and $D_2$ pays a cost of 1, and hence their extra costs are 0. - Consider each vertex $u$ in the cover $C$ (i.e., $u\in C$). - Form a cluster $A_u$ by adding three of the (type \[type:AAA\]) rows of the form $\langle 0, u, u\rangle$ - Form a cluster $B_u$ by adding the remaining three (type \[type:AAA\]) rows of the form $\langle 0, u, u\rangle$ - Consider each edge attached to $u$, say $(u, x)$ for some $x\in V$. Add the edge row $\langle 0, u, x\rangle$ to $A_u$ and add the edge row $\langle 0, x, u\rangle$ to $B_u$. - Add the three (type \[type:DDD\]) rows of the form $\langle 0, u, Z\rangle$ to $A_u$ and add the three (type \[type:EEE\]) rows of the form $\langle 0, Z, u\rangle$ to $B_u$. Notice that both $A_u$ and $B_u$ have at least seven rows each, since each vertex has at least one edge attached to it. Every row in these two clusters pays a cost of $1$ (thus, the extra cost paid by these rows is 0). - Form a cluster by adding the seven (type \[type:CCC\]) rows of the form $\langle ?, u, ?\rangle$. Add the critical row $\langle ?, u, u\rangle$ to this cluster. Notice that the seven rows each pay a cost of $2$, and hence their extra cost is $0$. [The critical row pays a cost of $2$, and hence, its extra cost is $1$]{}. Observe that all the rows of the table have been assigned to some cluster and each cluster has size at least $7$. From the above discussion, we see that the only rows having non-zero extra cost are the critical rows corresponding to the vertices in the cover $C$ and they pay an extra cost of $1$ each. We conclude that the total extra cost is $\|C\|$. We have proved the following claim: [Claim 1: ]{} If the given graph has a vertex cover of size $\leq t$, then there exists a $k$-anonymization solution with extra cost $\leq t$. We next proceed to prove the reverse direction: if there exists a $k$-anonymization solution $\sigma$ of extra cost $\leq t$, then there exists a vertex cover of size $\leq t$. Consider such a solution $\sigma$. We first make the following claim. [Claim 2: ]{} Consider a vertex $u$. Suppose the critical row $\langle ?, u, u\rangle$ pays an extra cost of 0. Then, the only cluster in which it can participate is the one obtained by combining the critical row with the six (type \[type:AAA\]) rows of the form $\langle 0, u, u\rangle$.\ [[Proof: ]{}]{}Clearly, the critical row must pay a cost of $1$, since it has a special symbol. If it pays no extra cost, then the rows it is combined with should have the symbol ’$u$’ in their second and third columns. There are exactly six such rows available and these are the (type \[type:AAA\]) rows of the form $\langle 0, u, u\rangle$. [$\Box$]{} We say that a vertex is [perfect]{}, if all the 20 rows corresponding to it pay an extra cost of 0. A vertex is said to be [imperfect]{}, if at least one of its 20 rows pay an extra cost of at least 1. [Claim 3: ]{}Consider an edge $(x, y)$. If both $x$ and $y$ are perfect, then at least one of the two edge rows corresponding to the edge pays a cost of at least 2.\ [[Proof: ]{}]{}Consider the edge row $\langle 0, x, y\rangle$, corresponding to the given edge. Let $S$ be the cluster to which this row belongs. Since $k=7$, we have $\|S\|\geq 7$. Recall that a column is said to be good with respect to $S$, if the rows of $S$ have identical values on the column; a column is said to be bad with respect to $S$, otherwise. We shall argue that at least two of the three columns are bad with respect to $S$. Let us consider the three possible choices for two-column subsets out of the three columns. - Clearly, both the second and the third columns cannot be good with respect to $S$, since there are no other rows that contain $x$ in their second column and $y$ in their third column. - Next, we argue that both the first and the second column cannot be good with respect to $S$. Since, both $x$ and $y$ are perfect, their critical rows do not pay any extra cost. By Claim 2, the six (type \[type:AAA\]) rows of the form $\langle 0, x, x\rangle$ have gone to some cluster other than $S$. Similarly, the six (type \[type:AAA\]) rows of the form $\langle 0, y, y\rangle$ have also gone to some other cluster. These rows cannot be part of $S$. Now, since the graph is 3-regular, there are only two other edge rows that have ‘$0$’ in their first column and ‘$x$’ in their second column; these correspond to the two other edges incident on $x$. There are three other rows (corresponding to the vertex $x$ and of type \[type:DDD\]) that have ‘$0$’ in their first column and ‘$x$’ in their second column. Thus, totally there are only 5 other rows that that have the above property. Since $\|S\|\geq 7$, it follows that both the first and the second column cannot be good with respect to $S$. - A similar argument shows that there are only 5 other rows that have ‘$0$’ in their first column and ‘$y$’ in their third column. This means that both the first column and the third column cannot be good in $S$. We conclude at least two of the three columns are bad respect to $S$. Thus, the concerned edge row $\langle 0, x, y\rangle$ must pay a cost of at least 2. [$\Box$]{} Let $V'$ be the set of all imperfect vertices. Let $E'$ be the set edges whose both endpoints are perfect. Each imperfect vertex (by definition) contributes at least 1 to the extra cost. By Claim 3, each edge in $E'$ pays an extra cost of at least 1. Therefore, $$\mbox{total extra cost of $\sigma$} \geq \|V'\| + \|E'\|.$$ Construct a vertex cover $C$ as follows. Add every imperfect vertex to $C$. For each edge in $E'$, add one of its endpoints (arbitrarily) to $C$. Clearly, $C$ is a vertex cover. So, $$\|C\| \leq \|V'\|+\|E'\| \leq \mbox{extra cost of $\sigma$}$$ We have proved the following claim: [Claim 4: ]{} If there exists a $k$-anonymization solution $\sigma$ of extra cost $\leq t$, then there exists a vertex cover of size $\leq t$. [$\Box$]{} We observed that the cost of a $k$-anonymous solution is the sum of $ABC$ and the extra cost of the solution. Now, by combining Claim 1 and Claim 4, we get the following: there exists a vertex cover of size $\leq t$, if and only if there exists a $k$-anonymization solution of cost $\leq ABC+t$. This completes the NP-hardness proof. [$\Box$]{} It is easy to show that our reduction is an $L$-reduction (see [@Papa] for a discussion on $L$-reductions). As the vertex cover problem on 3-regular graphs is MAXSNP-hard [@Alimonti], it follows that, \[thm:BBB\] The $k$-anonymization problem is MAXSNP-hard, even when the number of columns in 3 and the privacy parameter is fixed as $k=7$. Moreover, [[Chlebík]{}]{} and [[Chlebíkov[á]{}]{}]{} [@Chlebik] showed that the vertex cover problem on 3-regular graphs cannot be approximated within a factor of $\frac{100}{99}$. Now, taking the parameters of the $L$-reduction of our construction, and based on the result of [[Chlebík]{}]{} and [[Chlebíkov[á]{}]{}]{}, we can show that, The $k$-anonymization problem cannot be approximated within a factor of $\frac{6238}{6237}$, even when the number of columns in 3 and the privacy parameter is fixed as $k=7$. MAXSNP-Hardness and Inapproximability Bound ------------------------------------------- We show that the $k$-anonymization problem is MAXSNP-hard even when the number of columns is 3; meaning it cannot be approximated within a factor of $1+\epsilon$, for some $\epsilon>0$. \[thm:BBB\] The $k$-anonymization problem is MAXSNP-hard, even when the number of columns in 3 and the privacy parameter is fixed as $k=7$. [[Proof: ]{}]{}The vertex cover problem on 3-regular graphs is known to be MAXSNP-hard [@Alimonti]. We shall argue that the reduction given in Theorem \[thm:AAA\] is an $L$-reduction (see [@Papa] for a discussion on $L$-reductions); it would follow that the $k$-anonymization problem is MAXSNP-hard. Recall that an $L$-reduction from a minimization problem $P_1$ to a minimization problem $P_2$ consists of two polynomial time algorithms $R$ and $S$. The algorithm $R$ takes an instance $I_1$ of the problem $P_1$ as input and outputs an instance $I_2$ of the problem $P_2$. The algorithm $S$ takes as input a solution $S_2$ of the instance $I_2$ and produces a solution $S_1$ for the instance $I_1$. The following two requirements are to be met, for some constants $\alpha$ and $\beta$: (i) ${{\rm OPT}}(I_2)\leq \alpha {{\rm OPT}}(I_1)$; (ii) ${{\rm Cost}}(S_1) - {{\rm OPT}}(I_1) \leq \beta ({{\rm Cost}}(S_2) - {{\rm OPT}}(I_2))$. Let $G=(V,E)$ be the input 3-regular graph over $r$ vertices. The reduction given in Theorem \[thm:AAA\] produces a table $T$; this constitutes the required algorithm $R$. Let $C^$ be the optimal vertex cover of $G$. In Theorem \[thm:AAA\], we showed (in Claim 1) that given a cover of size at most $t$, we can construct a $k$-anonymization solution $\sigma$ of cost at most $ABC+t$. It follows that the optimal $k$-anonymization solution satisfies ${{\rm Cost}}(\sigma^)\leq ABC+\|C^\|$. Recall that $ABC = 27r+2\|E\|+14$. Since the graph is 3-regular, number of edges is $\|E\|=3r/2$. Thus, $ABC=30r+14$. For sufficiently large $r$, we have $ABC\leq 31r$. Since $G$ is 3-regular, a vertex can cover at most three edges and so, $\|C^\|\geq r/2$. It follows that $ABC \leq 62r$ and that ${{\rm Cost}}(\sigma^)\leq 63\|C^\|$. Thus, we have $\alpha = 63$. In Theorem \[thm:AAA\], we showed how to construct a vertex cover $C$, from a given $k$-anonymization solution $\sigma$ such that $\|C\|$ is at most the extra cost of $\sigma$; this construction is our algorithm $S$. It follows that $\|C\|\leq {{\rm Cost}}(\sigma)-ABC$. We already saw that ${{\rm Cost}}(\sigma^) \leq ABC + \|C^\|$. Combining the two inequalities, we get that $\|C\|-\|C^\|\leq {{\rm Cost}}(\sigma)-{{\rm Cost}}(\sigma^)$. Thus, we have $\beta = 1$. We have provided an $L$-reduction with $\alpha = 63$ and $\beta=1$. [$\Box$]{} We next extend the Theorem \[thm:BBB\] to get a mild inapproximability bound. The $k$-anonymization problem cannot be approximated within a factor of $\frac{6238}{6237}$, even when the number of columns in 3 and the privacy parameter is fixed as $k=7$. [[Proof: ]{}]{}Suppose we have an algorithm for solving the $k$-anonymization problem with an approximation ratio of $a$. Using the $L$-reduction given in Theorem \[thm:BBB\], we can obtain an algorithm for the vertex cover problem on 3-regular graphs with an approximation ratio of $(1+\alpha\beta (a-1))$. [[Chlebík]{}]{} and [[Chlebíkov[á]{}]{}]{} [@Chlebik] showed that the vertex cover problem on 3-regular graphs cannot be approximated within a factor of $\frac{100}{99}$. It follows that the $k$-anonymization problem cannot be approximated within a factor of $\frac{6238}{6237}$. [$\Box$]{} Special case: $m$ and $\|\Sigma\|$ are constants ============================================== As our NP-hardness reduction utilizes alphabets of arbitrarily large size, a natural question is whether the problem remains NP-hard when both the number of columns and the alphabet size are fixed constants. Here, we show that this case can be solved optimally in polynomial time. In the problem definition, let $m$, the number of columns, to be a constant and let the size of $\Sigma$ be a constant $s$ and let $k$ be the privacy parameter. By a [row pattern]{}, we mean a vector over $m$ columns whose entries belong to $\Sigma$. Let ${{\cal R}}$ denote the set of all row patterns; $\|{{\cal R}}\| = \|\Sigma\|^{m}$ is a constant. By an [anonymization pattern]{}, we mean a vector over $m$ columns whose each entry is either a symbol from $\Sigma$ or the suppression symbol ‘$$’. Let ${{\cal P}}$ denote the set of all anonymization patterns; $\|{{\cal P}}\| = (\|\Sigma\|+1)^{m+1}$ is a constant . We say that a row pattern $t$ [matches]{} an anonymization pattern $p$, if $p$ and $t$ agree on all columns, except the columns suppressed in $p$. We use “$t\sim p$" as a shorthand to mean that $t$ matches $p$. Consider the optimal $k$-anonymization solution $\sigma^$. For each row pattern $t$, the solution $\sigma^$ chooses an anonymization pattern $p$ matching $t$ and applies $p$ to $t$. If a pattern $p\in {{\cal P}}$ is applied to a row pattern $t$, we say that $t$ is [attached]{} to $p$. The solution satisfies the property that, for each anonymization pattern $p \in {{\cal P}}$, number of row patterns attached to it is either zero or at least $k$. If no row pattern is attached to $p$, then we say that $p$ is [closed]{}; on the other hand, if at least $k$ row pattern are attached to $p$, we say that $p$ is [open]{}. Thus, the optimal solution $\sigma^$ opens up some subset of patterns from ${{\cal P}}$. Of course, we do not know which patterns are open and which are closed. But, we can guess the set of open patterns by iterating over all possible subsets of ${{\cal P}}$. For each subset $P\subseteq {{\cal P}}$, our goal is to compute the optimal solution whose set of open patterns is exactly equal to $P$. The number of such subsets is $2^{\|{{\cal P}}\|}$, which is a constant since $\|{{\cal P}}\|$ is a constant. Then, we take the minimum of the over these solutions. Consider a subset of patterns $P$. Our goal is to find the optimal solution in which the set of open patterns is exactly equal to $P$. Notice that there may not exist any feasible solution for the subset $P$; we also need to determine, if this is the case. This can be formulated as the following integer linear program. For each row pattern $t \in {{\cal R}}$, $s(t)$ denotes the number of copies (i.e., tuples) of the row pattern in the input table ($s(t) = 0$ if the row pattern $t$ does not occur in the table). For each pair $(p \in {{\cal P}},t\in {{\cal R}})$ such that the row pattern $t$ matches the pattern $p$, we introduce an integer variable $x_{p,t}$. This variable captures the number of copies of the row pattern $t$ attached to the anonymization pattern $p$. For a pattern $p$, let ${{\rm Cost}}(p)$ denote that number of suppressed cells in $p$; this is the cost each copy of a row pattern $t$ would pay, if $t$ is attached to $p$. $$\begin{aligned} \notag \min \sum_{(p,t):t\sim p} {{\rm Cost}}(p) x_{p,t} & &\\ \notag \mbox{subject to:} &\\ \sum_{t:t\sim p} x_{p,t} \geq k & & \mbox{for all } p\in P\\ \label{eqn:AAA} \sum_{p:t\sim p} x_{p,t} = s(t) & & \mbox{for all } t\in {{\cal R}}\\ \label{eqn:BBB} x_{p, t}\in \mathbb{N}_0 & & \mbox{for all } (p,t):t\sim p\end{aligned}$$ Note that this integer linear program has a constant number of variable as the number of $x_{p,t}$ variables is bounded by $\|{{\cal P}}\|\cdot\|{{\cal R}}\| \leq m^{2\|\Sigma\|+1}$. By the famous result of Lenstra [@Lenstra], an integer linear program on constant number of variables can be solved in polynomial time. We relax the integrality constraints to get the following linear program. The equality constraints (\[eqn:AAA\]) are split into two inequalities and relax the integrality constraints (\[eqn:BBB\]). [^1]. $$\begin{aligned} \notag \min \sum_{(p,t):t\sim p} {{\rm Cost}}(p) x_{p,t} & &\\ \notag \mbox{subject to:} & \\ \label{eqn:CCC} \sum_{t:t\sim p} x_{p,t} \geq k & & \mbox{for all } p\in P\\ \label{eqn:DDD} \sum_{p:t\sim p} x_{p,t} \geq 1 & & \mbox{for all } t\in T\\ \label{eqn:EEE} -\sum_{p:t\sim p} x_{p,t} \geq -1 & & \mbox{for all } t\in T\\ \label{eqn:FFF} x_{p, t}\geq 0 & & \mbox{for all } (p,t):t\sim p\end{aligned}$$ Let $A$ denote the constraint matrix where each row is a constraint and each column is a variable. Our main claim is that the matrix $A$ is totally unimodular (see [@opt-book] for a discussion on totally unimodular matrices). Recall that a square matrix is unimodular, if its determinant is -1, 0 or +1 and that a matrix is totally unimodular, if all its square submatrices are unimodular. It is well known that if the constraint matrix is totally unimodular and the right hand side of the constraints are all integral, then the optimal solution is integral. This means that in the optimal solution returned by the linear program, the values of all the variables $x_{p,t}$ will be either 0 or 1. Then, our constraints imply that for each row $t$, $x_{p,t}=1$ for exactly one pattern $p$ that $t$ matches to. It is now easy to convert the linear program solution into a $k$-anonymization solution of same cost. We shall show that $A^T$ (transpose of $A$) is totally unimodular. Taking transpose preserves total unimodularity, and thus we would get that $A$ is also totally unimodular. Let $X$ be a $a\times b$ matrix and $Y$ be a $a\times c$ matrix. Let $X\|\|Y$ denote the $a\times (b+c)$ matrix obtained by concatenating $X$ and $Y$ along the column-side (namely, the first $b$ columns of $X\|\|Y$ are taken from $X$ and the remaining $c$ columns are taken from $Y$). The matrix $A^T$ can be decomposed into four parts given as $A^T = A_1 \|\| A_2 \|\| \overline{A}_2 \|\| I$. The fours parts $A_1, A_2, \overline{A}_2$ and $I$, come from the four sets of constraints, \[eqn:CCC\], \[eqn:DDD\], \[eqn:EEE\] and \[eqn:FFF\], respectively. The matrix $A_1$ has a row corresponding to each variable $x_{p,t}$ and a column corresponding to each pattern $p'$, and a $1$ in the corresponding cell if $p=p'$ and $0$, otherwise. Similarly, the matrix $A_2$ has a row corresponding to each variable $x_{p,t}$ and a column corresponding to each row $t'$, and a $1$ in the corresponding cell if $t=t'$ and $0$, otherwise. The matrix $\overline{A}_2$ is given by $\overline{A}_2 = (-1)A_2$ and $I$ is an identity matrix. We can multiply each column of $A_2$ by $(-1)$; this may only change the sign of the determinants of the submatrices of $A$. Thus, it suffices to show that $A' = A_1 \|\| A_2 \|\| A_2 \|\| I$ is totally unimodular. Notice that $A_1 \|\| A_2$ is the incidence matrix of the bipartite graph $G$ defined below: the patterns in ${{\cal P}}$ form one side of the graph and the rows of the table $T$ form the other side; an edge is drawn between a pattern $p$ and a row $t$, if $t$ matches $p$. Now the claim that $A'$ is totally unimodular follows from the following three simple propositions. [@opt-book] The incidence matrix of any bipartite graph is totally unimodular. If $X\|\|Y$ is totally unimodular then $X\|\|Y\|\|Y$ is also totally unimodular. [[Proof: ]{}]{}Write $Z=X\|\|Y\|\|Y$. The matrix $Z$ has three parts to it. Consider any submatrix $S$ of $Z$. If $S$ includes only columns from the first two parts, then it is a submatrix of $X\|\|Y$, which is guaranteed to be unimodular. Now, suppose $S$ includes a column from the second part and the corresponding column from the third part. Then, these two columns can be canceled out yielding a submatrix with a column full of 0’s and hence, having a determinant of $0$. The remaining case is that $S$ does not include any pair of corresponding columns. In this case $S$ is also a submatrix of $X\|\|Y$ and hence, it is unimodular. [$\Box$]{} [@opt-book] If $X$ is totally unimodular then $X\|\|I$ is totally unimodular, where $I$ is an identity matrix. This approach, when applied to the practical case of $m=O(\log n)$ and $\|\Sigma\|$ being arbitrarily large, leads to a variant of facility location problem. The patterns ($n2^m=n^{O(1)}$ in number) can be viewed as facilities with a connection cost equal to the number of suppressed cells. The rows can be viewed as clients who can be serviced by any pattern that they match to. The goal is to open a subset of the facilities and attach the clients to the facilities such that every open facility has at least $k$ clients attached to it. Objective is to minimize the total connection cost of all the clients. Note that the distances here are non-metric. No approximation algorithms are known for this variant. Designing approximation algorithms for this facility location problem that can in turn yield approximation algorithm for the above case of anonymization problem would be interesting. Open Problems ============= For the general $k$-anonymization problem, the best known approximation algorithm, due to Aggarwal et al. [@ICDT], achieves a ratio of $O(k)$. Their algorithm is based on a natural graph theoretic framework. They showed that any poly-time algorithm that uses their framework cannot achieve a factor better than $O(k)$. Breaking the $O(k)$-approximation barrier seems to be a challenging open problem. Improving the $O(\log k)$ approximation ratio, due to Park and Shim [@SIGMOD], for the practical special case when $m=O(\log n)$ is an interesting open problem. For the case where $m$ is constant, a trivial constant factor approximation algorithm exists: suppressing all cells yields an $O(m)$ approximation ratio. However, it is challenging to design an algorithm that, for all constants $m$, guarantees a fixed constant approximation ratio (say, $2$); notice that such an algorithm is allowed to run in time $2^{2^m}$. Getting a hardness of approximation better than $\frac{6238}{6237}$ would be of interest. [^1]: We do not need to enforce the constraint “$x_{p,t}\leq 1$”, since this is automatically taken are by third set of constraints.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The Signorini problem for the Laplace operator is considered in a general polygonal domain. It is proved that the coincidence set consists of a finite number of boundary parts plus isolated points. The regularity of the solution is described. In particular, we show that the leading singularity is in general $r_i^{\pi/(2\alpha_i)}$ at transition points of Signorini to Dirichlet or Neumann conditions but $r_i^{\pi/\alpha_i}$ at kinks of the Signorini boundary, with $\alpha_i$ being the internal angle of the domain at these critical points.' author: - 'Thomas Apel[^1]' - 'Serge Nicaise[^2]' bibliography: - 'Signorini.bib' title: \| Regularity of the solution of the\ scalar Signorini problem in polygonal domains[^3] --- Signorini problem, coincidence set, regularity 35B65; 49N60 Introduction ============ In this paper we consider the Signorini problem $$\begin{aligned} \label{eq:bvp} -\Delta y &= 0 \quad\text{in }\Omega, \\ y &= 0 \quad\text{on }\Gamma_D,\\ \partial_n y&=0 \quad\text{on }\Gamma_N,\\ \partial_n y&=u \quad\text{on }\Gamma_U,\\ y\geq 0,\ \partial_n y\geq 0, \ y \partial_n y&=0 \quad\text{on }\Gamma_S, \label{eq:signorinibc}\end{aligned}$$ with a boundary datum $u\in L^2(\Gamma_U)$. We assume that the mutually disjoint, relatively open sets $\Gamma_D$, $\Gamma_N$, $\Gamma_U$, and $\Gamma_S$ satisfy $$\begin{aligned} \label{boundaryparts} \bar \Gamma_D\cup \bar \Gamma_N\cup \bar \Gamma_U\cup \bar \Gamma_S =\Gamma=\partial\Omega, \quad \bar\Gamma_S\cap\bar\Gamma_U=\emptyset, \quad \Gamma_D\ne \emptyset\end{aligned}$$ with $\Gamma$ being the boundary of the bounded polygonal domain $\Omega\subset\R^2$. The boundary parts $\Gamma_N$ and $\Gamma_U$ are distinguished because of the second assumption in . The condition $\Gamma_D\ne \emptyset$ is assumed to obtain a unique solution. The notation and our interest in the problem comes from an optimal control problem where $y$ is the state variable and $u$ is the control variable. Problem – is sometimes considered as the scalar version of the more important Signorini problem for the Lamé equations (“linear elasticity with unilateral boundary condition”) but it has its own application describing a steady-state fluid mechanics problem in media with a semi-permeable boundary, see [@GlowinskiLionsTremolieres81 Section 1.1.1]. Let $C=\{c_i\}_{i=1}^n$ be the set of critical boundary points, namely all points where the type of the boundary condition changes, that is $\Gamma\setminus(\Gamma_D\cup\Gamma_N\cup\Gamma_U\cup\Gamma_S)$, and all corners of the domain. Brézis [@Brezis:71] (see also [@Fichera:71] for the elasticity system) showed for the inhomogeneous equation in smooth domains with purely Signorini boundary condition that the solution is $H^2$-regular, Grisvard and Iooss [@Grisvard-Iooss:76] extended this result to the case of convex domains. Moussaoui and Khodja [@MoussaouiKhodja1992] showed $C^{1,\lambda}$-regularity away from $C$ for $\lambda<\frac12$, see also Theorem \[t:regy\]; they further discussed possible singular behavior near the critical points, see also Theorem \[t:reg\]. This last description suggests the $H^t$-regularity with $t\in (2,5/2)$ of the solution near $\Gamma_S$. Consequently some authors [@Belhachmi-Belgacem:03; @Drouet-Hild:15; @Wohlmuthetall:12] assume such a regularity without a complete proof, and use it for their numerical analysis of the problem. However, for the analysis of the behavior near the extremal points of $\Gamma_S$ and for sharper regularity results one needs that the coincidence set $$\begin{aligned} \label{def:GammaC} \Gamma_C=\{x\in\Gamma_S:y(x)=0\}\end{aligned}$$ consists of a finite number of connected boundary parts (“intervals”) plus isolated points. Otherwise the set of endpoints of the coincidence set (the set of points where the condition $u=0$ changes to $u>0$) could possess accumulation points while the analysis of the regularity near such points (or near corners of the domain) assumes the existence of a $\delta$-neighborhood where the type of the boundary condition does not change. As a consequence there are publications where the structure of the coincidence set is formulated as an assumption, see, e.g., [@ChristofHaubner18 Condition (A)]. One important result of our paper is the proof of this proposition in Section \[sec:coincidence\]. Such a result was previously obtained for the Signorini problem with the Lamé equations by Kinderlehrer in [@kinderlehrer:81Pisa; @kinderlehrer:81JMPA] under the assumptions - that the boundary of $\Omega$ is flat in a neighborhood of $\Gamma_S$, more precisely that $$\Gamma_S=(-c,c)\times \{0\}\subset \tilde \Gamma=(-\tilde c,\tilde c)\times \{0\} \subset \partial\Omega$$ for some positive constants $c$ and $\tilde c$ such that $c<\tilde c$, and - that the part $\tilde \Gamma\setminus\Gamma_S$ is included into $\Gamma_N$. While the transfer to the Laplace equation and to the case that $\tilde \Gamma\setminus\Gamma_C\subset\Gamma_D$ can be done with similar ideas, the avoidance of the the first assumption above is not straightforward. The main tool for our proof is a special conformal mapping which preserves the differential operator in and the normal derivative. It is not clear how to analyze other equations or a domain with curved boundary. For simplicity of presentation we assumed that the differential equation in and the gap function in are homogeneous. We admit that we cannot treat the general case but we discuss examples in Remark \[rem:f\]. With the knowledge of the structure of the coincidence set one can analyze the regularity of the solution, see, e.g., the already mentioned paper [@MoussaouiKhodja1992] by Moussaoui and Khodja for results in Hölder spaces. We discuss the regularity in Sobolev spaces in Section \[sec:reg\] where we use a form which is useful for our forthcoming numerical analysis. \[sec:coincidence\]The coincidence set ====================================== Problem – admits the following variational formulation. By introducing the convex set $$K=\{v\in H^1(\Omega): v=0 \text{ on } \Gamma_D \text{ and } v\geq 0 \text{ on } \Gamma_S\},$$ the function $y\in K$ satisfies the variational inequality $$\begin{aligned} \int_\Omega (\nabla y\cdot \nabla (v-y))\geq \int_{\Gamma_U} u (v-y)\quad \forall v\in K. \label{eq:VFsignorini}\end{aligned}$$ The solution of exists and is unique, see for instance [@KinderlehrerStampaccia Section II.2.1]. Let us start with a first regularity result of this solution. In particular it shows that the solution is continuous near the Signorini boundary such that the definition of $\Gamma_C$ in makes sense. To this end, introduce a domain $$\begin{aligned} \label{def:W} W\subset\mathbb{R}^2:\ \bar\Gamma_S\subset W,\ \bar\Gamma_U\cap\bar W=\emptyset,\end{aligned}$$ see the illustration in Figure \[fig:W\], and let ![\[fig:W\]Illustration of the domains $\Omega$, $W$ and $W_\delta$.](W) $$\begin{aligned} W_\delta:=(W\cap\Omega)\setminus\bigcup_{i=1}^n \bar B(c_i,\delta)\end{aligned}$$ where the set $C=\{c_i\}_{i=1}^n$ of centers of balls with radius $\delta>0$ is introduced in the introduction. \[t:regy\] The solution $y\in K$ of with $u\in H^{1/2}(\Gamma_U)$ satisfies $$\label{reg:locale} y\in H^2(W_\delta)\cap C^{1,\lambda}(\bar W_\delta) \quad\forall \lambda\in(0,\tfrac12)$$ for any $\delta>0$. We start with the proof of the property $$\begin{aligned} \label{eq:localreg} \forall x\in \bar W_\delta \quad \exists\, \varepsilon_x>0:\quad y\in H^2(W_\delta\cap B(x,\varepsilon_x)) \end{aligned}$$ using localization arguments. - If $x\in W_\delta$ we consider a ball $O_x=B(x,\varepsilon_x)$ with $\varepsilon_x<\mathrm{dist}(x,\partial\Omega)$. The solution is harmonic in $O_x$ and hence even real analytic in $O_x$ [@GLC Theorem 1.7.1]. - For $x\in\Gamma_D\cap\bar W_\delta$ or $x\in\Gamma_N\cap\bar W_\delta$ we consider a neighborhood $O_x=B(x,\varepsilon_x)\cap\Omega$ with $\varepsilon_x<\mathrm{dist}(x,C)$. Again, since the solution is harmonic in $O_x$ it is real analytic in $O_x$, [@GLC Theorem 2.7.1], i.e. near the smooth part of the Dirichlet or Neumann boundary. - For the remaining case $x\in \Gamma_S\cap \bar W_\delta$ we fix a rotationally symmetric cut-off function $\eta\in \mathcal{D}(\R^2)$ such that $\eta=1$ in a neighborhood of $x$ with a small support such that $\mathrm{supp}\,\eta\cap \Omega\subset W_{\delta/2}$, see the illustration in Figure \[fig:supp\]. ![\[fig:supp\]Illustration of $\mathrm{supp}\,\eta$ and $O_x$.](supp) Let now $O_x=B(x,\varepsilon_x)\cap\Omega$ with appropriately chosen $\varepsilon_x$ be a convex domain containing the support of $\eta$. Then $v=\eta w-\eta^2 y+y$ with arbitrary $$w\in K_x:=\{z\in H^1(O_x): z\geq 0\ \text{on } \partial O_x\}$$ satisfies $v=y=0$ on $\Gamma_D$ and $v=\eta w+(1-\eta^2)y\ge0$ on $\Gamma_S$ since all factors are greater than or equal to zero; hence $v\in K$. Inserting $v$ into gives $$\int_\Omega\nabla y\cdot\nabla(\eta(w-\eta y)) \ge \int_{\Gamma_U}u\eta(w-\eta y),$$ and with $\eta\equiv0$ in $\Omega\setminus O_x$ we get $$\begin{aligned} \label{eq:newstar} \int_{O_x}\nabla y\cdot\nabla(\eta(w-\eta y)) \ge0 \quad\forall x\in K_x. \end{aligned}$$ Since $\partial_n\eta=0$ on $\partial O_x$ we get $$\begin{aligned} \int_{O_x} &\nabla(\eta y)\cdot\nabla(w-\eta y) = \int_{O_x} (y\nabla\eta+\eta\nabla y)\cdot\nabla(w-\eta y) \\ &=-\int_{O_x} \nabla\cdot(y\nabla\eta)\cdot\nabla(w-\eta y) + \int_{O_x} \eta\nabla y\cdot\nabla(w-\eta y) \\ &=-\int_{O_x} \nabla\cdot(y\nabla\eta)(w-\eta y) + \int_{O_x} \nabla y\cdot\big(\nabla(\eta(w-\eta y))-\nabla\eta(w-\eta y)\big) \\ &\ge-\int_{O_x} \left(\nabla\cdot(y\nabla\eta)+\nabla y\cdot\nabla\eta\right)(w-\eta y) \end{aligned}$$ due to . Hence $\eta y\in K_x$ can be seen as the unique solution of $$\begin{aligned} \int_{O_x} (\nabla (\eta y)\cdot \nabla (w-\eta y)+\eta y(w-\eta y))\geq \int_{O_x} g_x (w-\eta y)\quad \forall w\in K_x \end{aligned}$$ with $g_x:=-\nabla\cdot(y\nabla\eta)-\nabla y\cdot\nabla\eta +\eta y \in L^2(O_x)$. Grisvard and Iooss showed that $\eta y\in H^2(O_x)$, see [@Grisvard-Iooss:76 Corollary 3.2]. Altogether the property is proved. The balls $B(x,\varepsilon_x)$ form an open covering of $\bar W_\delta$, hence there exists a finite covering, i.e., $$\begin{aligned} \exists\,x_j, \ j=1,\ldots,J: \quad \bar W_\delta\subset \bigcup_{j=1}^J B(x_j,\varepsilon_{x_j}). \end{aligned}$$ We conclude that $$y\in H^2(W_\delta)\subset W^{1,p}(W_\delta)\quad\forall p\in [1,\infty).$$ With the same procedure as above we can now prove that $y\in C^{1,\lambda}(W_\delta\cap B(x,\varepsilon_x))$ for all $x \in \bar W_{2\delta}$. The point is that now $g_x\in L^p(O_x)$ for all $p<\infty$ such that we can use a theorem from Khodja and Moussaoui, [@MoussaouiKhodja1992 Theorem 2] (see also [@Uraltseva:87]), to deduce that $\eta_x y$ belongs to $C^{1,\lambda}(O_x)$ for all $\lambda\in(0,\frac12)$. As above we conclude $y\in C^{1,\lambda}(W_{2\delta})$. Since $\delta>0$ was arbitrary we are done. The following lemma is inspired from [@kinderlehrer:81Pisa §6], see also [@MoussaouiKhodja1992 Lemma III.1.3]. \[l:zero\] Denote by $\partial_ty$ and $\partial_ny$ the tangential and normal derivatives along the boundary. Then the equality $$\begin{aligned} \partial_ty \, \partial_ny =0 \ \text{on } \Gamma_S\setminus C. \label{eq:prod=0} \end{aligned}$$ holds. \[rem:zero\] This result extends even to $\Gamma_D$ and $\Gamma_N$ since $\partial_ty=0$ on $\Gamma_D$ and $\partial_ny=0$ a.e. on $\Gamma_N$. Introduce the compact set $$\Gamma_{S,\varrho}:=\Gamma_S\setminus\bigcup_{i=1}^nB(c_i,\varrho)$$ for some $\varrho>0$. Then according to Theorem \[t:regy\], $\partial_n y$ is continuous on $\Gamma_{S,\varrho}$, hence we can introduce the sets $$\begin{aligned} \Gamma_{S,\varrho}^+&=\{x\in \Gamma_{S,\varrho}\ \| \ \partial_n y(x)>0\},& \Gamma_{S,\varrho}^0&=\{x\in \Gamma_{S,\varrho}\ \| \ \partial_n y(x)=0\}, \end{aligned}$$ and notice that $ \Gamma_{S,\varrho}=\Gamma_{S,\varrho}^+\cup \Gamma_{S,\varrho}^0. $ At this stage, we distinguish two cases: 1. If $x\in \Gamma_{S,\delta}^0$, we have $\partial_n y(x)=0$ and hence $$\begin{aligned} {\partial_t}y (x) \, \partial_ny(x) =0. \label{eq:prod=0ponctual} \end{aligned}$$ 2. In the other case, $x\in \Gamma_{S,\delta}^+$, we have $y(x)=0$ due to the Signorini conditions. Observe that the continuity of $\partial_n y$ implies that $\Gamma_{S,\varrho}^+$ is an open subset of $\Gamma_{S,\varrho}$. Hence, if $x\in \Gamma_{S,\delta}^+$, then $y=0$ holds in a neighborhood of $x$, and the tangential derivative is also zero in this neighborhood and consequently $\partial_ty (x)=0$, which shows that also holds in that case. We have just shown that is valid for all $x\in \Gamma_{S,\varrho}$ and letting $\varrho$ tend to zero we find . We prove now the main result of this section, namely the characterization of the coincidence set $\Gamma_C$, see . For that purpose, we adapt the method of Kinderlehrer in [@kinderlehrer:81Pisa §6] who treated the case of the elasticity system. \[thm:coincideset1\] Let $y\in K$ be the unique solution of , then the coincidence set $\Gamma_C$ is the union of a finite numbers of intervals and finitely many isolated points. We localize the problem by considering a finite covering of $\Gamma_S$. Introduce a finite number of open balls $B(c_i,\varrho_i)$, $i\in J$. The index set $J$ is chosen such that $J\supset \{i\in C : x_i\in \bar\Gamma_S\}$ and the radii $\varrho_i>0$ are chosen such that $\Gamma_S\subset\bigcup_{i\in J}B(c_i,\varrho_i)$ and $c_j\not\in B(c_i,\varrho_i)$ for $i\ne j$, see Figure \[fig:domain\]. Note that the index set may contain further points $c_i\in\Gamma_S\setminus C$. ![\[fig:domain\]Illustration of the domain, the boundary conditions and the covering.](domain) We consider now any ball $B(c_i,\varrho_i)$ and omit the index $i$ for better readability. Introduce a local polar coordinate system $(r,\theta)$ centered in $c$, such that $$O_\alpha:=B(c,\varrho)\cap\Omega = \{(r\cos\theta,r\sin\theta)\in\mathbb{R}^2: 0<r<\varrho,0<\theta<\alpha\}$$ where $\alpha$ is the angle of the domain at $c$. Consider now the situation where $$\Gamma_0:=\{(r\cos\theta,r\sin\theta)\in\mathbb{R}^2:0<r<\varrho,\theta=0\} \subset\Gamma_S.$$ The other leg $\Gamma_\alpha:=\{(r\cos\theta,r\sin\theta)\in\mathbb{R}^2: 0<r<\varrho, \theta=\alpha\}$ may be contained in $\Gamma_D$, $\Gamma_N$ or $\Gamma_S$ but not in $\Gamma_U$ because of $\bar\Gamma_S\cap\bar\Gamma_U=\emptyset$, see . Note that the situation where $\Gamma_\alpha\subset\Gamma_S$ and $\Gamma_0\not\subset\Gamma_S$ can be treated in a similar way. The function $y$ satisfies $$\begin{aligned} -\Delta y &=0 \quad\text{in } O_\alpha, \\ y\geq 0,\ \partial_n y\geq 0, \ y \partial_n y&=0 \quad\text{on }\Gamma_S\cap B(c,\varrho). \end{aligned}$$ Now regarding $y$ as a function of the complex variable $z=x_1+{\mathrm{i}\,}x_2$, we define the function $$w(z)=\partial_2 y (z)+{\mathrm{i}\,}\partial_1y(z),$$ defined in $O_\alpha$ (now considered as a subset of $\mathbb{C}$), see the illustration in Figure \[fig:trafo\]. ![\[fig:trafo\]Illustration of transformation.](trafo) As $y$ is harmonic in $O_\alpha$ and belongs to $C^2(O_\alpha)$, the function $w$ is analytic in $O_\alpha$, [@conway p. 41]. Furthermore, we introduce the biconformal mapping $$h: z\mapsto\hat z= z^{\pi/\alpha}, \quad O_\alpha\to O_{\pi,\alpha}:= \{z=r\mathrm{e}^{{\mathrm{i}\,}\theta}\in\mathbb{C}: 0<r<\varrho^{\pi/\alpha}, 0<\theta<\pi\},$$ and denote $\hat\Gamma_0:=h(\Gamma_0)$ and $\hat\Gamma_\pi:=h(\Gamma_\alpha)$. Note the simple rule $$h:r\mathrm{e}^{{\mathrm{i}\,}\theta}\mapsto\hat r\mathrm{e}^{{\mathrm{i}\,}\hat\theta}\quad\text{with } \hat r=r^{\pi/\alpha} \text{ and } \hat\theta=\frac{\theta\pi}{\alpha}.$$ Let us analyze now the function $$\hat y(\hat z):=y(z).$$ Since a conformal mapping preserves the Laplace operator (up to a factor) and since the normal derivative is up to a factor again the $\theta$-derivative we get $$\begin{aligned} -\Delta\hat y &=0 \quad\text{in }O_{\pi,\alpha}, \\ \hat y\ge0, \quad \hat\partial_n\hat y\ge0,\quad \hat y \hat\partial_n\hat y&=0 \quad\text{on }\hat\Gamma_0 , \end{aligned}$$ and the appropriate Dirichlet, Neumann or Signorini boundary condition on $\hat\Gamma_\pi$. Moreover, we can compute $$\begin{aligned} \partial_r y&= \partial_r\hat r\,\partial_{\hat r}\hat y= \tfrac\pi\alpha r^{\pi/\alpha-1} \partial_{\hat r}\hat y , & \partial_\theta y&= \partial_\theta\hat\theta\,\partial_{\hat\theta}\hat y= \tfrac\pi\alpha\partial_{\hat\theta}\hat y, \end{aligned}$$ such that $$\begin{aligned} \int_{O_{\pi,\alpha}} \|\hat\nabla\hat y\|^2 &= \int_{O_{\pi,\alpha}} \left( \|\partial_{\hat r}\hat y\|^2+ \|\hat r^{-1}\partial_{\hat\theta}\hat y\|^2 \right) \,\hat r\mathrm{d}\hat r\mathrm{d}\hat\theta \\ &= \int_{O_\alpha} \left( (\tfrac\pi\alpha)^{-2} r^{2(1-\pi/\alpha)} \|\partial_r y\|^2 + r^{-2\pi/\alpha}(\tfrac\pi\alpha)^{-2}\|\partial_\theta y\|^2 \right) \,r^{\pi/\alpha}\, \tfrac\pi\alpha r^{\pi/\alpha-1}\mathrm{d}r\, \tfrac\pi\alpha\mathrm{d}\theta\\ &= \int_{O_\alpha} \left( \|\partial_r y\|^2 + \|r^{-1}\partial_\theta y\|^2 \right) \,r\mathrm{d}r\mathrm{d}\theta = \int_{O_\alpha} \|\nabla y\|^2, \end{aligned}$$ i.e., for the function $$\hat w(\hat z):=\partial_1\hat y+{\mathrm{i}\,}\partial_2\hat y$$ the relation $$\begin{aligned} \|\hat w(\hat z)^2\|&=\|\hat w(\hat z)\|^2= \|\hat\nabla\hat y\|^2\in L^1(O_{\pi,\alpha}) \end{aligned}$$ holds. Lemma \[l:zero\] and Remark \[rem:zero\] imply that $$\Im \left(w(z)\right)^2=0 \hbox{ on } (-\hat\varrho,\hat\varrho)\setminus\{0\},$$ with $\hat\varrho=\varrho^{\pi/\alpha}$. Consequently on $U:=B(0,\hat\varrho)\setminus\{0\}$, we define the function $$F(z)= \begin{cases} w(z)^2 &\hbox{ if } \Im z\leq 0,\\[0.5ex] \overline{w(\bar z)^2} &\hbox{ if } \Im z> 0, \end{cases}$$ which is analytic in $U$ by the Schwarz reflection principle, see [@conway §IX.1]. Hence $F$ is meromorphic in $U$. As additionally $F$ belongs to $L^1(U)$ we conclude that $F$ admits the Laurent expansion $$F(z)=\frac{c}{z}+F_H(z),$$ with $c\in \mathbb{C}$ and a function $F_H$ which is an analytic in $B(0,\hat\varrho)$. (Terms $z^{-j}$ with $j>1$ are not in $L^1(U)$.) This implies that the function $$\Phi(z)=z F(z)$$ is holomorphic on $B(0,\hat\varrho)$. Therefore $\Phi$ has a finite number of zeroes on $\hat\Gamma:=\{z\in\bar U:\Im z=0\}$ if $\Phi$ is not identically equal to 0. Let us analyze two cases: 1. If $\Phi$ is identically equal to 0, then by we get $$w(z)^2=(\partial_2y)^2-(\partial_1 y)^2= 0 \hbox{ on } \hat\Gamma\setminus\{0\},$$ which again by implies $$\partial_2y=\partial_1 y=0 \hbox{ on } \hat\Gamma\setminus\{0\}.$$ Consequently $y$ is constant on $\hat\Gamma$, so either this constant is zero and $\hat\Gamma_C:=\{z\in\hat\Gamma:\hat y(z)=0\} = \hat\Gamma$, or this constant is different from zero and $\hat\Gamma_C= \emptyset$. 2. If $\Phi$ is not identically equal to 0, the sets $\{z\in \hat\Gamma: \Phi(z)>0\}$ and $\{z\in \hat\Gamma: \Phi(z)<0\}$ are unions of a finite number of intervals $I$. We are looking for the behavior of $y$ on any of these intervals $I$. Depending on the sign of $z\in\hat\Gamma$ we find that $\Phi(z)=zF(z)$ is positive or negative in $I$, hence $(\partial_2 y)^2-(\partial_1 y)^2$ does not change sign in $I$, and moreover $(\partial_2 y)^2>0$ or $(\partial_1 y)^2>0$ in $I$. If $(\partial_2 y(z))^2>0$ then we get by the Signorini condition that $y\equiv 0$ in $I$. If $(\partial_1 y(z))^2>0$ then the function $y$ is nowhere constant in $I$, hence $y$ has no or a finite number of zeros in $I$, and we get by the Signorini condition that $\partial_n y=0$ a.e. in $I$. In conclusion, in this case $\hat\Gamma_C$ is the union of a finite number of intervals plus eventually a finite number of points. Since the mapping $h$ is continuous this result is valid also for $\Gamma_C$. \[rem:f\] Let us finish this section with a discussion of our assumption that we assumed a homogeneous differential equation in and a homogeneous gap function in . - The assumption of a homogeneous differential equation in was made to simplify the discussion. For Lemma \[t:regy\] a right hand side $f\in L^\infty(\Omega)$ could be admitted. Recall also the introduction of the domain $W$ in . The whole analysis is untouched if the equation is homogeneous in a neighborhood of $\Gamma_S$ only since then the set $W$ could be defined accordingly. - In particular cases the solution of non-homogeneous problem could be homogenized. Assume that the differential equation in is replaced by $-\Delta y=f$ and the gap condition in is replaced by $y\ge\psi$. If $f$ and $\psi$ are such that there exists a function $y_{f,\psi}$ such that $$\begin{aligned} -\Delta y_{f,\psi} &= f \quad\text{in }\Omega, \\ y_{f,\psi} &= 0 \quad\text{on }\Gamma_D,\\ \partial_n y_{f,\psi}&=0 \quad\text{on }\Gamma_N\cup\Gamma_U,\\ y_{f,\psi}=\psi,\ \partial_n y_{f,\psi}&=0 \quad\text{on }\Gamma_S, \end{aligned}$$ then $y-y_{f,\psi}$ satisfies our assumptions. Of course this problem is overdetermined such that the existence of a solution cannot be expected for any $f$ and $\psi$. But examples can be constructed by choosing a function $$y_\in\{v\in H^2(\Omega):v=0 \text{ on }\Gamma_D, \partial_n v=0 \text{ on }\Gamma_N\cup\Gamma_U\cup\Gamma_S\} \supset H^2_0(\Omega)$$ and defining $f=-\Delta y_$ and $\psi=y_*\|_{\Gamma_S}$. Christof and Haubner investigated in [@ChristofHaubner18] a square domain and the case $\Gamma=\Gamma_S$. In the case of a homogeneous differential equation, Condition (A) in this paper is now proven in Theorem \[thm:coincideset1\], namely that the relative boundary of $\Gamma_C$ has one-dimensional Hausdorff measure zero and the relative interior of $\Gamma_C$ consists of at most finitely many connected components. \[sec:reg\]Regularity of the solution ===================================== We formulate now a regularity result in the spirit of Theorem 2.3 of the paper [@ChristofHaubner18] by Christof and Haubner where the regular part of the solution is considered in $W^{2,p}$, $p>2$, $p\not=4$. But we like to note that the regular part could also be smoother; the prize is that possibly more singular terms have to be included and the datum $u$ at the Neumann boundary must be sufficiently regular. \[t:reg\] Let $y$ be the solution of problem –. Recall the set $\{c_i\}_{i=1}^n$ of critical points and the interior angles $\alpha_i$. Recall also that there are points $\{c_i\}_{i=n+1}^m\subset\Gamma_S\setminus C$ of unknown location which are the endpoints of the intervals in the coincidence set and in that case, set $\alpha_i=\pi$. Furthermore denote by $(r_i,\theta_i)$ local polar coordinates at all these points. Let $p>2$, $p\not\in P$, where the finite set $P$ of exceptional values is a subset of the countable set $$\left\{ \frac{2}{2-\frac{k\pi}{2\alpha_i}}, \ k\in\mathbb{N},\ i=1,\ldots,m\right\}.$$ Assume that $u\in W^{1-1/p,p}(\Gamma_U)$ satisfies the compatibility condition $u(c_i)=0$ - if $c_i\in \bar\Gamma_D\cap\bar\Gamma_U$ and $\alpha_i=\frac12\pi$ or $\alpha_i=\frac32\pi$ or - if $c_i\in \bar\Gamma_N\cap\bar\Gamma_U$ and $\alpha_i=\pi$. Then there is a representation of $y$ $$\begin{aligned} y=y_R+\sum_{i=1}^n\sum_{j: 0<\lambda_{i,j}<2-2/p\atop \lambda_{i,j}\ne 1} d_{i,j} r_i^{\lambda_{i,j}} \Phi_{i,j}(\theta_i) + \sum_{i=n+1}^m d_i r_i^{3/2}\Phi_{i}(\theta_i) \end{aligned}$$ with $y_R\in W^{2,p}(\Omega)$, coefficients $d_{i,j}$ and $d_i$, smooth functions $\Phi_{i,j}$ and $\Phi_{i}$, and exponents $$\lambda_{i,j}= \begin{cases} j\pi/\alpha_i & \text{if D-D or N-N or U-U or U-N conditions near }c_i,\quad j\ge1, \\ (j-\frac12)\pi/\alpha_i & \text{if D-N or D-U conditions near }c_i,\quad j\ge1, \\ j\pi/(2\alpha_i) & \text{if S-S conditions near }c_i,\quad j\ge2, \\ j\pi/(2\alpha_i) & \text{if S-D or S-N conditions near }c_i,\quad j\ge1, \end{cases}$$ where D-N means that one boundary edge at $c_i$ is contained in $\Gamma_D$ and the other in $\Gamma_N$, and so on. The compatibility conditions could be omitted, but then a singularity of the form $r(\Phi_1(\theta)+\log r\Phi_2(\theta))$ with $\Phi_i$ smooth has to be added, see [@grisvard:85a p. 263]. Since we have a finite number of critical boundary points $c_i$ due to Theorem \[thm:coincideset1\] we can treat them separately and use classical theory as described for instance in [@grisvard:85a Corollary 4.4.4.14]. Let us discuss shortly the situation near the Signorini boundary. For $i\le n$ and $c_i\in\bar\Gamma_S\cap\bar\Gamma_D$ or $c_i\in\bar\Gamma_S\cap\bar\Gamma_N$ we do not know whether a Dirichlet or Neumann boundary condition occurs on $\Gamma_S$ near $c_i$. Therefore we consider the worst situation of mixed boundary conditions. In the remaining cases some singularities disappear at $c_i\in\bar\Gamma_S$: 1. For $i=n+1,\ldots,m$ the leading singularity is $r_i^{3/2}$ since the term $r_i^{1/2}$ is not in $H^2(\Omega)$, compare the result in Lemma \[t:reg\], and see the discussion in e.g. [@ChristofHaubner18; @MoussaouiKhodja1992]. 2. For $i\le n$ and $c_i\in\Gamma_S$ the worst situation could be mixed. Let us consider the case that Dirichlet conditions is valid for $\theta_i=0$. Then we have in the vicinity of $c_i$ $$\label{expwith small remainder} y=y_r+ dr_i^{\pi/(2\alpha_i)}\sin\left(\frac{\pi\theta_i}{2\alpha_i}\right).$$ We show now $y_r=o(r_i^{\pi/(2\alpha_i)})$ such that this term is neglectable sufficiently close to $c_i$. Indeed from [@grisvard:85a Corollary 4.4.4.14] near $c_i$, we have $$\label{fullexp} y_r=y_R+ \sum_{j\in \mathbb{N}\setminus\{0\}: 0<(j+1/2)\pi/\alpha_i <2-2/p\atop (j+1/2)\pi/\alpha_i \ne 1} d_{i,j}r_i^{(j+1/2)\pi/\alpha_i} \sin((j+1/2)\pi \theta_i/\alpha_i),$$ with $y_R\in W^{2,p}(\Omega\cap B(c_i, \rho))$ for $\rho$ small enough and $d_{i,j}\in \mathbb{R}$. Consequently, near $c_i$, $$\label{bcyR} y_{R}(r_i,0)=0,\quad \frac{\partial y_{R}}{\partial \theta_i}(r_i,\alpha_i)=0 \quad \forall r_i<\rho.$$ Notice that the Sobolev embedding theorem guarantees that $$\label{eq:star} y_R\in C^{1,\beta}(\bar \Omega\cap \bar B(c_i, \rho))\ \text{with }\beta=1-2/p.$$ We now notice that the second term in the sum in (if any) is trivially $o(r_i^{\pi/(2\alpha_i)})$, hence it remains to check the same behavior for $y_R$. We note that $\nabla y_R(c_i)=0$ except in the cases $\alpha_i=\frac12\pi$ or $\alpha_i=\frac32\pi$, and that $r_i^{(j+1/2)\pi/\alpha_i}$ is smooth when $\alpha_i=\frac12\pi$. For that purpose, we distinguish three cases. 1. If $\pi/(2\alpha_i)<1$, then by Taylor’s theorem (and since $ y_R(c_i)=0$), we have $$y_R(x)=\nabla y_R(c_i)\cdot (x-c_i)+o(r_i),$$ which yields $y_R(x)=O(r_i)=o(r_i^{\pi/(2\alpha_i)})$ as $\pi/(2\alpha_i)<1$. 2. If $\pi/(2\alpha_i)=1$, then from [@grisvard:85a Corollary 4.4.4.14], we directly have $$y=y_R\in W^{2,p}(\Omega\cap B(c_i, \rho)),$$ in other words the singular part is zero. 3. If $\pi/(2\alpha_i)>1$, then owing to and the regularity of $y_R$, we actually have $$\nabla y_R(c_i)=0,$$ hence by Taylor’s expansion with an integral remainder, we have $$y_R(x)=\int_0^1 \nabla y_R(c_i+t(x-c_i)) (x-c_i)\,dt, \forall \|x-c_i\|<\rho.$$ Therefore as $\|\nabla y_R(c_i+t(x-c_i))\| =\|\nabla y_R(c_i+t(x-c_i))-\nabla y_R(c_i)\|=O((tr_i)^{\beta})$ due to , one deduces that $$\| y_R(x)\|=O(r_i^{\beta+1})=o(r_i^{\pi/(2\alpha_i)}),$$ as $\pi/(2\alpha_i)<\beta+1$. (In the case $\pi/(2\alpha_i)>\beta+1=2-2/p$ the solution $y$ is $W^{2,p}$-regular in the vicinity of $c_i$. Equality is excluded by assumption.) Coming back to , for $\theta_i=0$ we get $\partial_ny=\partial_ny_r-d r_i^{\pi/(2\alpha_i)-1}$, hence $d\le0$ in order to satisfy the Signorini condition $\partial_ny\ge0$. For $\theta_i=\alpha_i$ we get $y=y_r+dr_i^{\pi/(2\alpha_i)}$, hence $d\ge0$ in order to satisfy the Signorini condition $y\ge0$. So we can have only $d=0$. Since all cases are treated the proof is complete. Let us shortly discuss the L-domain; that is a hexahedron with one interior angle $\alpha=\frac32\pi$ and all others being of size $\frac12\pi$. The leading singular term near the non-convex corner is of type $r^\lambda$ with $\lambda=\frac23$ if Signorini conditions are given at both legs of this angle, but with $\lambda=\frac13$ if a Signorini condition is given on one leg only, and a Dirichlet or Neumann condition at the other leg. These terms are in $H^s(\Omega)$ for $s<1+\lambda$ or in a suitable weighted Sobolev space. The set of exception values for $p$ is $P=\{3,6\}$. #### Acknowledgment The authors thank Constantin Christof for pointing to an incorrect argument in a previous version of the paper. The authors thank also Christof Haubner for preparing the illustrations. [^1]: `[email protected]`, Institut für Mathematik und Computergestützte Simulation, Universität der Bundeswehr München, D-85579 Neubiberg, Germany [^2]: `[email protected]`, Université Polytechnique Hauts-de-France, LAMAV, FR CNRS 2956, F-59313 - Valenciennes Cedex 9, France [^3]: Partially funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – Projektnummer 188264188/GRK1754.	{ "pile_set_name": "ArXiv" }
--- author: - 'Jessica Agarwal, Michael M[ü]{}ller, and Eberhard Gr[ü]{}n' bibliography: - '/usr4/users/jagarwal/Latex/refs.bib' title: 'Dust Environment Modelling of Comet 67P/Churyumov-Gerasimenko' --- Introduction {#sec:intro} ============ The inner comae of comets are the most dust rich environments in the solar system. Almost everything we see from a comet with the naked eye is dust. Both the coma and the tail are seen as sunlight scattered by [$\mu$m]{}-sized dust. Meteors, especially in meteor streams, are caused by mm- to cm-sized particles that originate to a large extent in comets. Fireballs are due to multi-tons boulders some of which are believed to stem from comets as well. Spacecraft missions to comets must take careful precautions to survive the hazards of the cometary dust cloud. Special dust shields have been designed to protect the spacecraft. Nevertheless, the dust environment can have detrimental effects on some aspects of the missions. The Giotto spacecraft flew by comet 1P/Halley at a distance of 600 km and a speed of 70 km/s. During this close fly-by Giotto was hit by mm-sized dust particles. The impacts caused nutation of the spacecraft spin, and the data transmission was disrupted. Some experiment sensors suffered damage during this fly-by. Similarly, the 350-kg impactor of the Deep Impact probe experienced attitude disturbances from dust grains before it hit the nucleus of comet 9P/Tempel 1. Owing to the low velocity of Rosetta relative to the comet, the consequences of dust impacts will be much less severe than for the fly-by missions. But detailed knowledge of the dust environment will be vital for the planning of spacecraft operations at the comet and is thus of crucial importance to optimise Rosetta’s scientific return. One of the problems in characterising the dust environment of a comet is that information on the nucleus, its dust and gas release is very limited. Before 1986, observations of cometary dust was the domain of astronomers. The method of Finson and Probstein [@finson-probstein1968a] was the first to be used to determine the size distribution of dust from observations of the tail at visible wavelengths. High resolution astronomical images of cometary comae revealed jets and other structures in the inner parts, some of which formed spirals which rotated like water from a lawn sprinkler indicating discrete dust emissions from localised active parts of the nucleus. A consequence of observing in visible light is that the results are biased to particle sizes in the range of 1 to 10 [$\mu$m]{}, because much smaller and much larger particles do not contribute significantly to the scattered light [@gruen-cometsI]. With the extension of the observable spectral range to infrared wavelengths, also the thermal emission of dust became accessible to astronomers. It revealed information on the abundance of larger grains and on the mineralogical composition of the dust, the latter from characteristic spectral features in the near and mid-infrared range. Breakthroughs in understanding cometary constituents came with the space missions to several comets: Giotto and two VeGa spacecraft to comet 1P/Halley in 1986, Deep Space 1 to comet 19P/Borelly in 1999, Stardust to comet 81P/Wild 2 in 2004, and Deep Impact to comet 9P/Tempel 1 in 2005. Water and CO were identified as the main species in the gas, and dust particles made of carbonaceous and silicate materials ranging from nanometre to millimetre sizes were detected. Active areas and corresponding dust jets were identified in spatially resolved images of some of the comets visited by spacecraft [@keller-delamere1987; @thomas-keller1987; @soderblom-boice2004; @sekanina-brownlee2004]. For 67P/C–G, however, such detail will only be observed when in 2014 the Rosetta spacecraft reaches the comet. Until then any information on the dust environment has to be derived from astronomical observations of the target comet or by assuming correspondence to other, better studied comets. The purpose of the present paper is twofold: on the one hand, to give an overview of the current knowledge of the dust environment of comet 67P/C–G, and on the other hand, to provide estimates of such quantities as the spatial density, flux, and speed of dust in the coma as functions of location and time. These values are meant to support the planning of measurements of the instruments on board Rosetta. Section \[sec:obs\] contains an overview of the available astronomical observations of comet 67P/C–G, and measurements of the albedo and temperature of dust from [67P/C-G]{} are presented. In addition, the phase function, size distribution and radiation pressure efficiency of cometary dust are defined and discussed. In Section \[sec:model\], several methods are presented to derive dust properties from modelling of images of the cometary tail or trail, and results for [67P/C-G]{} that were obtained by several authors are discussed. Finally, in Section \[sec:mmmodel\], the ESA Cometary Dust Environment Model [@mueller1997; @mueller1998; @muellerPhD; @landgraf-mueller1999] is briefly described and results from applying it to comet 67P/C–G are presented. Observations of [67P/C-G]{} Dust and Dust Properties {#sec:obs} ==================================================== In this section, we give an overview of the available observational data containing information on the dust of [67P/C-G]{}, and we introduce and discuss the major quantities that can be measured by means of such observations. Some of these quantities (the dust size distribution, the radiation pressure efficiency and the emission speeds) can – in the absence of in situ measurements – only be inferred through modelling of astronomical images, which is discussed in Section \[sec:model\]. Published astronomical data on the dust of [67P/C-G]{} include images of the dust coma, tail and trail in both visible and infrared light. They are available from 1982 onwards. Disambiguation: Dust tail, antitail, neckline, and trail {#subsec:dust_phenom} -------------------------------------------------------- In the following we give the definitions of some observational dust phenomena associated with comets as they are used in this paper: Outside the inner coma – in which the dust is accelerated by gas drag – dust dynamics is dominated by solar gravity and radiation pressure. Both forces follow a $1/r_{\rm h}^2\,$-law ($r_{\rm h}$ being the heliocentric distance) but act in opposite directions. Consequently, their ratio $\beta$ depends only on the material properties of the dust grains, such as size, composition, density, and shape. In general, radiation pressure is most efficient for particles of about the size of the dominant wavelength of the radiation. Such grains are driven away from the nucleus in the direction opposed to the Sun and trailing the nucleus, thus forming the comet’s [dust tail]{} before they disperse into interplanetary space. The term [antitail]{} refers to a part of the tail that seems to point toward the Sun instead of away from it. Often, this is a projection effect that occurs when the observer is in such a position that part of the normal tail appears to be on the Sun-facing side of the nucleus. Viewed in three dimensions, there is no difference between an ordinary tail and a projection antitail, but the dust seen in the antitail tends to be the larger and older component of the dust in the tail. A [neckline]{} [@kimura-liu1977; @richter-curdt1991; @mueller-green2001] consists of large particles emitted at a true anomaly of 180$^\circ$ before the observation: The orbital periods of large particles are similar to that of the parent comet. Their orbits are generally inclined with respect to the comet orbit, but the particles cross the orbital plane of the comet twice during each revolution around the Sun. One intersection point is the point of emission. The other lies on the line of nodes connecting the emission point and the Sun. The position of the second intersection point on the line of nodes depends on the emission velocity and $\beta$ of the particle. Large particles emitted at a given time cross the orbital plane of the comet almost simultaneously, but at different positions along the nodal line. To an observer in – or close to – the comet orbital plane, they appear as a bright line, the neckline. Necklines can appear both in the Sun- and the anti-Sun direction i.e. can contribute to the tail as well as to the antitail. In the case of comet C1995 O1 Hale-Bopp in early 1997 the neckline was mainly visible along the tail direction, but also gave rise to an antitail [@boehnhardt2003]. The [dust trail]{} of a comet consists of mm- to cm-sized particles that – because of low emission speeds and little sensitivity to radiation pressure – remain close to the comet orbit for many revolutions around the Sun and whose appearance reminds of an airplane contrail. Trails of eight short-period comets were first observed with IRAS in 1983 [@sykes-lebofsky1986; @sykes-hunten1986; @sykes-walker1992a], one of them being that of 67P/C–G. Morphology of Coma, Tail, Antitail and Trail {#subsec:obs_morph} -------------------------------------------- The coma of 67P/C–G showed azimuthal brightness variations during both the 1996/97 and the 2002/03 apparitions [@schleicher2006; @schulz-stuewe2004a; @schulz-stuewe2004b; @weiler-rauer2004a]. An example is displayed in Figure \[fig:dust\_jets\]. The bright regions have been interpreted as border lines of coma fans produced by active areas at different latitudes on the rotating nucleus. The 2003 fan pattern suggests the presence of 2 or 3 active regions on the nucleus. For the 2-active-region scenario, the rotation axis requires to be very much inclined to the orbital plane, while for the 3-active-region scenario a very wide range of rotation axis directions is possible (H. B[ö]{}hnhardt, private communication). The azimuthal direction of the bright features has been used to constrain the orientation of the rotation axis of the nucleus [@weiler-rauer2004a; @schleicher2006]. ![Left: Broadband R image of comet 67P/C–G on 3 May 2003 obtained with FORS1 at ESO/VLT/U1 [@schulz-stuewe2004a; @schulz-stuewe2004b]. Right: Same image, but structurally enhanced by subtracting the azimuthally averaged coma profile from the measured brightness distribution. Distinct features in the coma are clearly visible. The image is a reproduction of Figure 1 (lower left) in [@schulz-stuewe2004b], courtesy by R. Schulz. []{data-label="fig:dust_jets"}](cg_vlt.eps "fig:"){width=".5\textwidth"} ![Left: Broadband R image of comet 67P/C–G on 3 May 2003 obtained with FORS1 at ESO/VLT/U1 [@schulz-stuewe2004a; @schulz-stuewe2004b]. Right: Same image, but structurally enhanced by subtracting the azimuthally averaged coma profile from the measured brightness distribution. Distinct features in the coma are clearly visible. The image is a reproduction of Figure 1 (lower left) in [@schulz-stuewe2004b], courtesy by R. Schulz. []{data-label="fig:dust_jets"}](cg_vlt_proc.eps "fig:"){width=".5\textwidth"} The tail of 67P/C–G was characterised by a thin, bright feature close to the projected comet orbit and pointing away from the nucleus in the direction opposed to the motion of the comet. This feature was first observed shortly after perihelion in August 2002 and prevailed at least until April 2006 when the comet had already passed aphelion [@fulle-barbieri2004a; @moreno-lara2004; @agarwal-boehnhardt2007; @agarwalPhD; @ishiguro2008]. Different interpretations of this phenomenon are discussed in Section \[sec:model\] with the conclusion that it most probably was a very pronounced antitail due to the low inclination of the comet orbit with respect to the ecliptic. The dust trail of 67P/C–G was first observed with the Infrared Astronomical Satellite (IRAS) in 1983 [@sykes-lebofsky1986; @sykes-hunten1986; @sykes-walker1992a] with a reported length of 1.2$^\circ$ in mean anomaly and a width of 50000 km. In visible light, observations of the 67P/C–G trail were done in 2002/03 at heliocentric distances between 1.3 AU and 3.1 AU [@ishiguro2008] and in 2004 at $r_{\rm h} = 4.7 \,{\rm AU}$ (out-bound) [@agarwal-boehnhardt2007]. Infrared observations of the 67P/C–G trail between 2004 and 2006 were obtained with the MIPS instrument of NASA’s Spitzer Space Telescope at 24$\mu$m [@kelley-reach2008; @agarwalPhD]. ![Reproduction of Figure 1 in [@agarwal-boehnhardt2007]. The images show the dust trail and neckline of 67P/C–G in April 2004. The data were obtained with the Wide Field Imager (WFI) at the ESO/MPG 2.2m telescope in La Silla with a total exposure time of 7.5 hours and without filter. The size of the images is 35$^{\prime}$x4.7$^{\prime}$ each, corresponding to $1.1^\circ$ in mean anomaly parallel to the orbit. (a) Unfiltered image. (b) Same image, each pixel being replaced by the average over a neighbourhood of 200 pixels (140$^{\prime\prime}$) parallel and 10 pixels (7$^{\prime\prime}$) perpendicular to the trail axis after removal of the nucleus. The filtering window is indicated in the upper right corner of (b). A detailed discussion of the data acquisition, processing and interpretation is given in [@agarwal-boehnhardt2007] and modelling results are given in Section \[subsec:model\_trail\].[]{data-label="fig:trail"}](trail_orig.eps "fig:"){width="\textwidth"} ![Reproduction of Figure 1 in [@agarwal-boehnhardt2007]. The images show the dust trail and neckline of 67P/C–G in April 2004. The data were obtained with the Wide Field Imager (WFI) at the ESO/MPG 2.2m telescope in La Silla with a total exposure time of 7.5 hours and without filter. The size of the images is 35$^{\prime}$x4.7$^{\prime}$ each, corresponding to $1.1^\circ$ in mean anomaly parallel to the orbit. (a) Unfiltered image. (b) Same image, each pixel being replaced by the average over a neighbourhood of 200 pixels (140$^{\prime\prime}$) parallel and 10 pixels (7$^{\prime\prime}$) perpendicular to the trail axis after removal of the nucleus. The filtering window is indicated in the upper right corner of (b). A detailed discussion of the data acquisition, processing and interpretation is given in [@agarwal-boehnhardt2007] and modelling results are given in Section \[subsec:model\_trail\].[]{data-label="fig:trail"}](trail_bc200x10.eps "fig:"){width="\textwidth"} Albedo and Phase Function {#subsec:alb_phfn} ------------------------- The geometric albedo $p$ of an object is defined as the ratio of the intensities reflected backwards by the object and by a totally diffusely reflecting (i.e. Lambert scattering) disc of the same geometric cross section [@hanner-giese1981], while the Bond albedo $A_{\rm B}$ is the ratio of the total light reflected from a sphere to total light incident on it [@allen1973]. The phase angle $\alpha$ is the angle between the directions of the observer and of the incident radiation as seen from the scattering particle. The phase function $j(\alpha)$ describes the ratio of intensity scattered in the $\alpha$-direction to the intensity scattered at $\alpha = 0$, and the integral of the phase function over 4$\pi$ solid angle is called the [phase integral]{}, $q$. The Bond and geometric albedos are related by $A_{\rm B} = q\,p$. To derive the dust phase function from astronomical observations, an object must be observed at different phase angles, i.e. at different times. Since the total cross section of dust is not constant with time, an appropriate normalisation is required, for which two methods are used. One method employs gas production rates measured simultaneously with [$Af\!\rho$]{} for normalisation [@millis-ahearn1982; @meech-jewitt1987; @schleicher-millis1998], assuming that the dust-to-gas ratio as well as the dust size distribution and material properties remain constant over time. The other method, preferable but more laborious, normalises the scattered intensity to the simultaneously measured thermal infrared emission from the same volume [@tokunaga-golisch1986; @hanner-tedesco1985a; @hanner-newburn1989; @gehrz-ney1992; @ney1974; @ney-merrill1976; @ney1982]. The general shape of the phase function of cometary dust is characterised by a distinct forward and a gentle backscattering peak and is rather flat at medium phase angles [@kolokolova-hanner-cometsII-2004]. Divine [@divine1981] derived from data given in [@ney1974; @ney-merrill1976; @ney1982] the phase function shown in Figure \[fig:phase\_fns\] (solid line). The figure also shows the geometric phase function that describes the phase angle dependence of [$Af\!\rho$]{} and is discussed in Section \[subsec:obs\_afrhogas\]. Laboratory measurements and theoretical studies suggest that the dust albedo depends on particle size [@mcdonnell-lamy1991]. Earth based observations have so far not been suitable to investigate this dependence, because coma observations only provide data for the ensemble of particles of all sizes along a line of sight. The geometric albedo is derived from the simultaneous observation of the scattered visible and the thermally emitted infrared light, either directly at $\alpha = 0$, or at multiple phase angles and assuming a given phase function. The geometric albedo of dust in the coma of [67P/C-G]{} derived from the optical and infrared brightness was 0.04 at 1.25[$\mu$m]{} and 0.05 at 2.2[$\mu$m]{}[@hanner-tedesco1985a], which is in accordance with a large sample of comets [@divine-fechtig1986; @hanner-newburn1989]. From the low albedo it is inferred that there is no significant population of cold, bright (and possibly icy) grains that would contribute to the scattered light but not to the thermal emission [@hanner-tedesco1985a]. There is, however, some indication that the geometric albedo is higher for comets beyond $3 \,{\rm AU}$ [@hanner-newburn1989], which may point in the same direction. ![Solid line: phase function of an individual dust particle as given in [@divine1981], but here normalised to $j (\alpha\! = \!0) = 1$. Dashed line: geometric phase function $j_{\rm geo} (\alpha)$ [@muellerPhD] accounting for the anisotropy of an axis-symmetric coma with peak activity at the subsolar point.[]{data-label="fig:phase_fns"}](phase_fns.eps){width=".7\textwidth"} Afrho and Gas Production {#subsec:obs_afrhogas} ------------------------ The brightness of a cometary coma is proportional to the dust production rate. To infer the production rate from data obtained under different observational circumstances, the measured brightness must be corrected for all other parameters on which it depends. The quantity [$Af\!\rho$]{} was defined for this purpose [@ahearn-schleicher1984]. [$Af\!\rho$]{} stands for the product of albedo $A = p j(\alpha)$ (see Section \[subsec:alb\_phfn\]), filling factor $f$ of grains within the field of view, and the radius $\rho$ of the aperture at the comet. It is measured as follows: $$\label{eq:afr-def} Af\!\rho=4\frac{\Delta^2 (r_{\rm h} / 1 {\rm AU})^2} {I_{\rm sun}^{\rm (filter)}} \times \frac{I_{\rm dust}^{\rm (filter)}}{\rho},$$ where $r_{\rm h}$ and $\Delta$ are the heliocentric and geocentric distances of the comet during the observation, and $\rho$ is the radius of the circular aperture on which the coma intensity $I_{\rm dust}^{\rm (filter)}$ was measured using a given filter. $I_{\rm sun}^{\rm (filter)}$ is the intensity of the Sun at 1AU heliocentric distance seen through the same filter. Provided that the dust particles move away from the nucleus on straight trajectories and are not subject to processes altering their scattering behaviour, [$Af\!\rho$]{} is independent of the employed aperture radius, of the heliocentric and geocentric distances, and – to the extent that the dust can be considered as “grey” – of the spectral band in which the observation was carried out. For an isotropic coma and discrete dust sizes $s_j$, [$Af\!\rho$]{} is related to the production rates $Q_{{\rm d},j} \,(s_j)$ via the dust emission speeds $v_{{\rm d},j}$, the geometric albedo $p$, and the phase function $j(\alpha)$ [@muellerPhD]: $$\label{eq:afrnaper} Af\!\rho=2\pi \, p \, j(\alpha) \sum_{j} s_j^2 \,\frac{Q_{{\rm d},j}}{v_{{\rm d},j}}.$$ The relative magnitudes of the $Q_{{\rm d},j}$ rates are given by the size distribution. [$Af\!\rho$]{} depends on the phase angle of the observation, due both to the scattering properties of a single dust grain [@schleicher-millis1998] and – unless the coma is isotropic – to projection effects [@muellerPhD]. In a non-isotropic coma, the timescale on which the particles leave a given field of view depends on the angle between the main emission direction and the line of sight, with higher measured [$Af\!\rho$]{} for a line of sight parallel or close to the main emission direction. The phase-angle dependence of [$Af\!\rho$]{} is then better described by a [geometric]{} phase function $j_{\rm geo} (\alpha)$ [@muellerPhD] than by that of a single particle. The geometric phase function is characteristic of the specific pattern of emission of a given nucleus. In practice, [$Af\!\rho$]{} often depends on the aperture size despite its definition [@schleicher-millis1998; @schleicher2006 and references therein]. This implies that the brightness distribution in the coma deviates from the assumed 1/$\rho$-profile. Possible causes for this deviation include changes in the physical properties of the grains as they travel outward (e.g. loss of volatiles or fragmentation), the action of radiation pressure modifying the straight trajectories of small particles inside the field of view, or a long-lasting population of large particles [@schleicher-millis1998]. ![Top left: observed [$Af\!\rho$]{} values as function of heliocentric distance $r_{\rm h}$. Top right: observed [$Af\!\rho$]{} values corrected for phase angle through division by the geometric phase function (Figure \[fig:phase\_fns\]). Power-law fit for phase angle corrected [$Af\!\rho$]{}: $Af\!\rho\, (r_{\rm h}, \alpha\!\!=\!\!0$) $=2393\,\times (r_{\rm h} / 1 \, {\rm AU})^{-5.08}$ cm. Bottom: Measured H$_2$O production rates with corresponding power-law fit $Q_{\rm H_2O} \,(r_{\rm h}) = 3.4 \times\!10^{28}\,\times ( r_{\rm h}/ 1 \, {\rm AU})^{-5.71}$ molecules/s. []{data-label="fig:afr_h2o"}](afr_all_060624.eps "fig:"){width=".5\textwidth"} ![Top left: observed [$Af\!\rho$]{} values as function of heliocentric distance $r_{\rm h}$. Top right: observed [$Af\!\rho$]{} values corrected for phase angle through division by the geometric phase function (Figure \[fig:phase\_fns\]). Power-law fit for phase angle corrected [$Af\!\rho$]{}: $Af\!\rho\, (r_{\rm h}, \alpha\!\!=\!\!0$) $=2393\,\times (r_{\rm h} / 1 \, {\rm AU})^{-5.08}$ cm. Bottom: Measured H$_2$O production rates with corresponding power-law fit $Q_{\rm H_2O} \,(r_{\rm h}) = 3.4 \times\!10^{28}\,\times ( r_{\rm h}/ 1 \, {\rm AU})^{-5.71}$ molecules/s. []{data-label="fig:afr_h2o"}](afr_corr_all_060624.eps "fig:"){width=".5\textwidth"} ![Top left: observed [$Af\!\rho$]{} values as function of heliocentric distance $r_{\rm h}$. Top right: observed [$Af\!\rho$]{} values corrected for phase angle through division by the geometric phase function (Figure \[fig:phase\_fns\]). Power-law fit for phase angle corrected [$Af\!\rho$]{}: $Af\!\rho\, (r_{\rm h}, \alpha\!\!=\!\!0$) $=2393\,\times (r_{\rm h} / 1 \, {\rm AU})^{-5.08}$ cm. Bottom: Measured H$_2$O production rates with corresponding power-law fit $Q_{\rm H_2O} \,(r_{\rm h}) = 3.4 \times\!10^{28}\,\times ( r_{\rm h}/ 1 \, {\rm AU})^{-5.71}$ molecules/s. []{data-label="fig:afr_h2o"}](water_all_060624_fit.eps){width=".5\textwidth"} [$Af\!\rho$]{} has been measured for 67P/C–G during three perihelion passages (1982/83, 1996/97 and 2002/03) [@storrs-cochran1992a; @osip-schleicher1992a; @ahearn-millis1995a; @kidger2003; @lamy2003a; @weiler-rauer2004a; @schulz-stuewe2004a; @feldman-ahearn2004; @schleicher2006][^1]. Figure \[fig:afr\_h2o\](top left) shows all published [$Af\!\rho$]{}data as function of heliocentric distance. The same values after correction for the phase angle dependence using the geometric phase function (Figure \[fig:phase\_fns\]) are displayed in Figure \[fig:afr\_h2o\](top right) together with a power-law fit to the corrected values. Figure \[fig:afr\_h2o\](bottom) shows observed production rates of H$_2$O [@hanner-tedesco1985a; @feldman-ahearn2004; @maekinen2004; @crovisier-colom2002a] and a power-law fit to the data. The exponents of the derived power laws are untypically steep compared with other comets, and more data of high quality will be needed to confirm them. Both [$Af\!\rho$]{} and the H$_2$O production rate reach their maxima around 30 days after perihelion. No dust coma was detected beyond at least 4.9 AU [@tubiana-drahus2007_dps]. Dust Temperature {#subsec:obs_temp} ---------------- Assuming that a particle is characterised by the Bond albedo $A_{\rm B}$ at visible wavelengths and the emissivity $\epsilon$ in the infrared, its temperature $T$ at the heliocentric distance $r_{\rm h}$ (in AU) is given by the equilibrium between absorbed solar and emitted thermal radiation: $$\label{eq:gbeq} \frac{(1\!-\!A_{\rm B}) \, {I_{\odot}}}{r_{\rm h}^2} = 4\, \epsilon \sigma \,T^4,$$ where $\sigma$ is the Stefan-Boltzmann constant and ${I_{\odot}}$ = (1367 $\pm$ 2) W m$^{-2}$ the solar flux at 1AU [@cox2000]. The resulting equilibrium temperature is $$T (r_{\rm h}, A_{\rm B},\epsilon) = 278.8 \,{\rm K} \, \left(\frac{1-A_{\rm B}}{\epsilon}\right)^{\!\frac{1}{4}} \, \frac{1}{\sqrt{r_{\rm h}}}. \label{eq:T_eq}$$ A blackbody would be characterised by $A_{\rm B}$ = 0 and $\epsilon$ = 1. The temperatures of dust in the inner solar system are generally such that the main emission lies in the infrared. In practice, the temperature is derived from fitting a blackbody spectrum to measurements of the brightness at different infrared wavelengths. The coma of [67P/C-G]{} was monitored at multiple wavelengths in the range of 1 to 20[$\mu$m]{} between September 1982 (1.50 AU preperihelion) and March 1983 (1.88 AU postperihelion) [@hanner-tedesco1985a]. The derived temperatures of the dust in the coma were throughout higher than those of a theoretical blackbody at the same heliocentric distance, which is generally attributed either to the presence of submicron-sized particles (i.e. smaller than the dominant wavelength range of the thermal emission) [@hanner-tedesco1985a] or to very porous aggregates of small grains [@hanner-newburn1989; @kolokolova-hanner-cometsII-2004]. For a given heliocentric distance, the colour temperature was higher post- than pre-perihelion, suggesting a change in the particle properties or in the dominant size. Additionally, an 8- 13-[$\mu$m]{} spectrum was taken on 23 October 1983. It did not show a silicate feature, which is usually taken as an indication for the dominance of somewhat larger and more compact grains. Excess colour temperatures were also derived from IRAS observations of cometary trails at 12, 25, and 60[$\mu$m]{} [@sykes-walker1992a]. For [67P/C-G]{}, the derived temperature at 2.3 AU was approximately 14% above that of a blackbody at the same heliocentric distance, corresponding to $\epsilon/(1-A_{\rm B})$ = 0.6 $\pm$ 0.2. Figure \[fig:dust\_temp\] shows the temperature measurements discussed above and – for comparison – the equilibrium temperature of a blackbody as a function of heliocentric distance. The data from the coma [@hanner-tedesco1985a] and from the trail [@sykes-walker1992a] are remarkably consistent given that coma and trail are generally assumed to be dominated by different particle populations ([$\mu$m]{}- versus mm-cm sized). ![Colour temperatures of the dust from [67P/C-G]{} derived from multi-wavelength infrared observations. The symbols “+” and “x” refer to observations of the coma before and after perihelion, respectively [@hanner-tedesco1985a]. The asterisk indicates the temperature of the dust trail [@sykes-walker1992a], and the two dashed lines correspond – via Equation \[eq:T\_eq\] – to the upper and lower limit of $\epsilon/(1-A_{\rm B})$ derived from the same data. The solid line shows the temperature of a theoretical blackbody, described by Equation \[eq:T\_eq\] with $A_{\rm B}$ = 0 and $\epsilon$ = 1. []{data-label="fig:dust_temp"}](comparison_temperature_rh.eps){width=".7\textwidth"} Dust Size Distribution {#subsec:size_distr} ---------------------- The size distribution of cometary dust has been inferred from both astronomical images and in situ data. While the former yield a [size]{} distribution, the latter contain information on the [masses]{} of the particles. For optical images, the determined sizes scale directly with the particle albedo. The conversion from size to mass requires knowledge of the bulk density of the particles. A mass or size distribution can be specified in the form of either a [ differential]{} or a [cumulative]{} distribution. The cumulative mass distribution $F(m_0)$ gives information on the fraction $N$ of particles that have a mass greater than some mass $m_0$: $$N \, (m > m_0) = F(m_0).$$ The differential mass distribution $f(m)$ characterises the relative abundance $n$ of particles inside a mass interval $[m_1,m_2]$: $$n \, (m_1 \!<\! m \!<\! m_2) = \int\limits_{m_1}^{m_2} f(m) \: {{\rm d}}m = F(m_1) - F(m_2). \label{eq:nfrac_from_md}$$ If the mass of a particle can be converted to a size by a relation $s(m)$, a corresponding differential size distribution $g(s)$ exists: $$\int\limits_{m_1}^{m_2} f(m) \: {{\rm d}}m = \int\limits_{s(m_1\!)}^{s(m_2\!)} g\,(s) \: {{\rm d}}s.$$ It is generally assumed that – at least in intervals of the total mass range covered by cometary dust – the distributions can be approximated by power laws. In the literature, both the exponents $\gamma$ of the [cumulative mass distribution]{} $F(m) \sim m^{-\gamma}$ and $\alpha$ of the [differential size distribution]{} $g(s) \sim s^{\alpha}$ are commonly used. For a constant bulk density in the concerned size interval we have $m(s) \propto s^3$ and ${{\rm d}}m \propto s^2 {{\rm d}}s$. Hence the exponents $\gamma$ and $\alpha$ are related by $$\alpha = -3 \gamma -1. \label{eq:alpha2gamma}$$ The mass distribution at the nucleus is different from the one in the coma because of the size-dependence of the emission velocities. In general, large particles are more abundant in the coma than close to the surface because of their lower speeds. The relation between the size distributions in the coma and at the nucleus may be additionally complicated by fragmentation or evaporation of grains [@mcdonnell-alexander1987; @mcdonnell-lamy1991; @green-mcdonnell2004] and by an inhomogeneous distribution of surface activity [@fulle-colangeli1995]. In situ data on the dust mass distribution were obtained by the dust instruments on board the spacecraft VeGa 1 and 2 and Giotto at comet 1P/Halley in 1986 [@mcdonnell-alexander1987; @divine-newburn1987b; @mcdonnell-lamy1991; @fulle-colangeli1995], on board Giotto at comet 26P/Grigg-Skjellerup in 1992 [@mcdonnell-mcbride1993], and on board Stardust at comet 81P/Wild 2 in 2004 [@tuzzolino-economou2004; @green-mcdonnell2004]. The measured quantity is not the mass distribution of dust as released from the nucleus but the cumulative flux or fluence on the concerned instrument. The fluence is defined as the flux integrated over the spacecraft trajectory, it represents therefore an average mass distribution. The flux (or the fluence measured in only sections of the trajectory) showed significant variation with time during both the 1P/Halley and the 81P/Wild 2 fly-bys. Table \[tab:md\_lit\] lists mass- or size-distribution exponents given in the literature. Still no general agreement has been reached on the interpretation of the data with respect to the dust mass distribution at the nucleus, but the authors listed in Table \[tab:md\_lit\] do agree that the mass of dust in the coma is dominated by the largest emitted particles. The cumulative fluences registered at the various spacecraft show different exponents for large and small particles. The interpretations of this observation are not unanimous, either. \[tab:md\_lit\] Comet Instrument(s) $\tilde{\gamma}$ $\gamma$ $\alpha$ Ref. -------------------- ------------------ ------------------------- ---------------------------- --------------------------------- -------------------------------------------- ------------------------ ---------------------------- 1P/Halley DIDSY&PIA $10^{-13}$ $10^{-8}$ kg [1.02]{} $-$4.06 [@mcdonnell-alexander1987] $10^{-8}$ $10^{-6}$ kg [0.71]{} $-$3.13 DIDSY&PIA $10^{-19}$ $2\!\times\!10^{-14}$ kg [0.18]{} $0.18^\dagger$ $-$1.54 [@divine-newburn1987b] $2\!\times\!10^{-14}$ $10^{-6}$ kg [0.94]{} $1.11^\ast$ $-$4.32 VeGa 1 $10^{-19}$ $10^{-12}$ kg [0.26 ]{} $0.26^\dagger$ $-$1.78 $10^{-12}$ $10^{-9}$ kg [1.19 ]{} $1.36^\ast$ $-$5.07 VeGa 2 $10^{-19}$ $1.6\!\times\!10^{-13}$ kg [0.26 ]{} $0.26^\dagger$ $-$1.78 $1.6\!\times\!10^{-13}$ $10^{-9}$ kg [0.90 ]{} $1.07^\ast$ $-$4.2 DIDSY [$-$3.5]{} ${\mathbf \pm}$ [0.2]{} [@fulle-colangeli1995] \[4mm\] 26P/G–S DIDSY/GRE $10^{-9}$ $10^{-7}$ kg [0.27$^{+0.13}_{-0.20}$]{} $0.44^\ast$ $-$2.3 [@mcdonnell-mcbride1993] \[4mm\] 81P/Wild 2 DFMI $10^{-14}$ $10^{-9}$ kg [0.85]{} $\pm$ [0.05]{} $1.02^\ast$ $-$4.05 $\pm$ 0.15 [@green-mcdonnell2004] Stardust Samples $10^{-17}$ $10^{-3}$ kg [0.57 ]{} $0.74^\ast$ $-$3.21 [@hoerz-bastien2006] \[4mm\] : Exponents of the cumulative fluence on a spacecraft, $\tilde{\gamma}$, the cumulative mass distribution, $\gamma$, and the differential size distribution, $\alpha$, compiled from the literature. [Bold]{} values are taken directly from the publications, the remaining values were derived assuming the following relations. To translate $\tilde{\gamma}$ to $\gamma$, it is assumed that the speeds of small particles are limited by the speed of the gas and therefore independent of size. Hence $\gamma = \tilde{\gamma}$ for small particles (indicated by “$\dagger$”). For larger particles, the relation $v \propto m^{-1/6}$ is used, which was derived from a radially symmetric coma model [@divine-fechtig1986], resulting in $\gamma = \tilde{\gamma} + 1/6$ for large particles (indicated by “$\ast$”). The relation between the differential size distribution and the cumulative mass distribution is given by $\alpha = -3\gamma - 1$. Results from modelling of astronomical images are not included here. For [67P/C-G]{} a compilation of size distributions derived by various authors is shown in Figure \[fig:comparison\_model\](bottom). Radiation Pressure {#subsec:rad_press} ------------------ Both solar gravity and the radiation pressure force are inversely proportional to the heliocentric distance squared, and point radially away from the Sun. Radiation pressure can, therefore, be included in the equation of motion of a particle by introducing a modified potential substituting the gravitational constant, $G$, by $\tilde{G}$ = $G (1-\beta)$, where $\beta$ stands for the ratio of solar gravity to the radiation pressure force. It depends only on material properties of the dust grains, not on their distance from the Sun: $$\beta = \frac {3 \,L_\odot}{16 \,\pi c\, G M_\odot} \frac{Q_{\rm pr}}{\rho s}. \label{eq:beta}$$ $L_\odot$ and $M_\odot$ are the luminosity and mass of the Sun, and $c$ is the speed of light. The grain has the bulk density $\rho$ and the effective radius $s$, and it is characterised by the radiation pressure efficiency $Q_{\rm pr}$ [@burns-lamy1979; @divine-fechtig1986]. $Q_{\rm pr}$ represents the absorption and scattering properties of the grain averaged over the solar spectrum. For homogeneous spheres and some other simple shapes, $Q_{\rm pr}$ can be calculated in an exact way from the complex refractive index of the material [@mie1908; @bohren-huffman1983]. For more complicated structures, a variety of theoretical and experimental approaches exist to obtain $Q_{\rm pr}$ [@gustafson-dustbook-2001]. While for particle sizes on the order of 0.01 to 1[$\mu$m]{}, $\beta$ depends sensitively on material, shape, structure, surface properties, and size of the particles, it is approximately constant for much smaller particles, and proportional to 1/($\rho s$) for large ones, i.e. in the geometric optics regime. This implies that for sub-millimetre and larger particles, $Q_{\rm pr}$ is to first order independent of size [@burns-lamy1979; @bohren-huffman1983]. Results of Modelling the 67P Dust Tail and Trail {#sec:model} ================================================ In this Section, we give an overview of the efforts undertaken to derive properties of the [67P/C-G]{} dust from modelling astronomical observations of the thin, bright feature described in Section \[subsec:obs\_morph\], variantly referred to as neckline, antitail, or trail. The main quantities to be constrained by the models are the dust size distribution, the radiation pressure efficiency, and the emission speeds. The derived values will serve as input to the hydrodynamic coma model used in Section \[sec:mmmodel\] to predict the dust environment inside the coma. In the following, we summarise the results from neckline photometry applied to images taken in 2002/03 [@fulle-barbieri2004a; @moreno-lara2004], from analysis of the antitail as observed in May 2003 [@agarwal-mueller2007], and from simulations of the dust trail in visible and infrared wide-field observations between 2002 and 2006 [@agarwalPhD; @ishiguro2008; @kelley-reach2008]. The position of a dust particle in a cometary tail is a function of the radiation pressure coefficient $\beta$, the emission time, and the emission speed; and $\beta$ is closely related to the mass of a particle. Hence, the effect of solar radiation pressure is similar to that of a mass spectrometer. To understand the formation of dust tails, the concept of synchrones and syndynes was introduced by Bredikhin [@bredikhin1903]. Synchrones are the positions of particles of different $\beta$ emitted at a given time, while syndynes describe the positions of particles of fixed $\beta$ and varying emission time. Both terms refer to hypothetical particles emitted with zero velocity relative to the nucleus. Since, realistically, the initial velocity of a dust grain is different from zero, the resulting synchrone or syndyne will have a finite cross section with a radius proportional to the product of dust emission speed and time elapsed since emission. In particular, grains released at a given time with isotropic speed will form a spherical shell the centre of which moves along the appropriate syndyne. This description was used by Finson and Probstein [@finson-probstein1968a; @finson-probstein1968b] to derive properties of the dust size distribution from the brightness patterns observed in cometary dust tails. Their method is limited to recently emitted and small particles, because it neglects tidal effects and others (e.g. direction-dependent emission speed or production rate) that cause a dust shell to divert from the spherical shape [@kimura-liu1977; @fertig-schwehm1984; @mueller-green2001]. Various approaches to surmount these limitations are described in the following. Neckline Photometry {#subsec:model_fulle} ------------------- Fulle et al. [@fulle-barbieri2004a] analyse [67P/C-G]{} images obtained in March 2003 ([$r_{\rm h}$]{} = 2.6 AU) and before, when the comet was active and exhibited a significant coma. They employ both an analytical theory of neckline photometry and an inverse Monte Carlo model. The former is an analytical method to infer the emission speeds and abundance of particles as a function of the radiation pressure parameter $\beta$ [@fulle-sedmak1988]. It is applied to an image of the comet obtained with the Schmidt Telescope at the Th[ü]{}ringer Landessternwarte (TLS) in Tautenburg on 27-28 March 2003. The method relies on the assumption that the bright narrow feature in the tail is a neckline, and that all intensity observed in this feature is due to dust emitted within an interval of 10 hours at a true anomaly of $180^\circ$ before that of the observation, which corresponds to 5 May 2002 (105 days before perihelion) and . The authors find that – in this particular time interval – both the mass and the cross section of dust emitted by the comet were dominated by particles in the size range of $1.5$ to $10 \,{\rm mm}$, and that the exponent of the differential size distribution, $\alpha$, was between -3.5 and -3. From the fact that the described feature had been constantly observable since perihelion in August 2002, they infer that the characteristics of the dust emission as recorded in the neckline did not change between 3.6 and 1.7 AU before perihelion. The second method – the inverse Monte Carlo model – consists in a least-squares fit of simulated images to the measured surface intensity [@fulle1989a]. This method yields a set of time-dependent dust parameters (size distribution, emission velocity, size range, [$Af\!\rho$]{}, and dust production rate). The obtained solution is unique in the mathematical sense of a least-squares fit; its physical probability remains harder to evaluate. The method is applied to an image obtained with the Telescopio Nazionale Galileo (TNG) on La Palma during the same night as the TLS image. Quantitative interpretation of the model is achieved through matching of the modelled [$Af\!\rho$]{} to the value measured during a different observation close in time [@lamy2003a]. The authors [@fulle-barbieri2004a] conclude that the time dependence of all parameters must be asymmetric with respect to perihelion in order to match the observation. The size distribution exponent drops from $-3.5$ before to $-4.5$ after perihelion, the mass loss rate from $200 \,{\rm kg/s}$ to $10 \,{\rm kg/s}$. The emission speeds decrease by a factor of three during this time interval (from 3m/s to 1m/s for a particle of 1mm radius), and the dominant size by two orders of magnitude. Fulle et al. [@fulle-barbieri2004a] suggest the existence of two different source regions: one emitting large particles and only active before perihelion, and one emitting mostly small particles becoming active after perihelion. A decreasing production rate was also found by Moreno et al. [@moreno-lara2004] applying a similar method to the same image after calibration. Antitail Analysis in the Hypersonic Approximation {#subsec:model_hyperson} ------------------------------------------------- The dust size distribution has been evaluated by analysis of the brightness profile along the [67P/C-G]{} tail using the Finson-Probstein model in the [hypersonic approximation]{} [@agarwal-mueller2007], which was applied to an image (Figure \[fig:dust\_jets\]) obtained with ESO/VLT on 3 May 2003 [@schulz-stuewe2004a; @schulz-stuewe2004b]. At the time of the observation, [67P/C-G]{} was at a heliocentric distance of 2.9 AU postperihelion. A plot of synchrones and syndynes (Figure \[fig:synsyn\]) shows that all synchrones corresponding to ejection before 1.5 AU postperihelion (23 October 2002) appear projected in the direction towards the Sun, forming an antitail. They all are – in projection – inclined against the neckline by less than 0.8[$^{\circ}$]{}, which places doubts upon the assumption that all dust in the bright feature was emitted during an interval of merely 10 hours, on which the first model described in Section \[subsec:model\_fulle\] relies [@fulle-barbieri2004a]. ![ For 3 May 2003 synchrones are shown projected in the comet orbit plane (top) and image plane (bottom). The solid lines correspond to synchrones ejected in steps of $10^\circ$ in true anomaly starting $5^\circ$ in true anomaly before the observation time back to an ejection time that corresponds to a heliocentric distance of 3 AU before perihelion. Synchrones corresponding to ejection times after mid January 2003 appear in the direction opposed to the Sun. The synchrones of particles ejected before mid December 2003 appear in Sun direction. Only the synchrones of particles ejected in between (shown as long dashed) have significant angles with the Sun or anti-sun direction. Projected in the comet orbit plane (top panel) also syndynes are shown (dashed). The $\beta$-values corresponding to the syndynes are annotated for the larger particles. The $\beta$ values of the smaller particles are $1.5 \!\times\! 10^{-4}$, $3 \!\times\! 10^{-4}$, and $1 \!\times\! 10^{-3}$.[]{data-label="fig:synsyn"}](syncrodyn.eps "fig:"){width=".49\textwidth"} ![ For 3 May 2003 synchrones are shown projected in the comet orbit plane (top) and image plane (bottom). The solid lines correspond to synchrones ejected in steps of $10^\circ$ in true anomaly starting $5^\circ$ in true anomaly before the observation time back to an ejection time that corresponds to a heliocentric distance of 3 AU before perihelion. Synchrones corresponding to ejection times after mid January 2003 appear in the direction opposed to the Sun. The synchrones of particles ejected before mid December 2003 appear in Sun direction. Only the synchrones of particles ejected in between (shown as long dashed) have significant angles with the Sun or anti-sun direction. Projected in the comet orbit plane (top panel) also syndynes are shown (dashed). The $\beta$-values corresponding to the syndynes are annotated for the larger particles. The $\beta$ values of the smaller particles are $1.5 \!\times\! 10^{-4}$, $3 \!\times\! 10^{-4}$, and $1 \!\times\! 10^{-3}$.[]{data-label="fig:synsyn"}](syncrodyn2.eps "fig:"){width=".49\textwidth"} The hypersonic approximation holds when the extent of the dust shells along a synchrone is small, such that a unique relation between the position on the synchrone and $\beta$ can be assumed [@finson-probstein1968a]. It is then possible to infer the size distribution of the particles ejected from the nucleus from the brightness profile along the synchrone, if one of the following two conditions is met: (1) Either the antitail originates from the superposition of many synchrones that – in projection – have only a small offset from each other compared with the extension of the dust shells perpendicular to the synchrone. Or (2) the brightness along one direction in the image plane is dominated by one synchrone alone. A generalisation of the original formulation of the hypersonic approximation [@finson-probstein1968a] – which applied only to spherical shells – has been developed for tidally distorted elliptical shells, assuming radially symmetric particle emission [@agarwal-mueller2007]. The intensity along the antitail on 3 May 2003 could be fitted by a power law with an exponent of $-0.4$, which was translated to $\alpha = -4.1$ for the differential size distribution between at least 3 AU before and 1.5 AU after perihelion in 2002 [@agarwal-mueller2007]. To put this result in a wider context, we have compared the measured brightness distribution along the antitails of different comets rather than the derived size distributions, because we found that different authors use different models to derive the size distribution from the measured brightness exponent. We found that the brightness variation measured for the antitail of 67P/C–G is rather typical compared with the antitails of other comets, such that also the dust size distribution of 67P/C–G may be typical. Trail Analyses {#subsec:model_trail} -------------- Several wide field images of the dust along the orbit of [67P/C-G]{} were obtained between 2002 and 2006 in both optical and infrared light, and their analyses are summarised in the following. Three images – one optical and two infrared at 24[$\mu$m]{} – with fields of view of about half a degree were analysed using a generalised Finson-Probstein approach that takes into account the tidal deformation of dust shells on long time-scales [@agarwalPhD]. The optical image was obtained in April 2004 with the Wide Field Imager at the ESO/MPG 2.2m telescope on La Silla at [$r_{\rm h}$]{} = 4.7 AU (Figure \[fig:trail\]), and the infrared images were taken by the MIPS instrument on board the Spitzer Space Telescope of NASA in August 2005 and April 2006 ([$r_{\rm h}$]{}= 5.7 AU in both cases). All three images were thus taken when the comet was close to aphelion and not active, such that no particles larger than approximately 100[$\mu$m]{} are expected to be present in the fields of view. Simulated images were generated taking into account dust emitted during all seven perihelion passages since the last close encounter with Jupiter in 1959. The time-dependence of the dust production was modelled on the observed time-dependence of [$Af\!\rho$]{}. The emission speeds and surface activity were assumed to be isotropic, and the relative dependence of the emission speeds on size and heliocentric distance was obtained by help of the hydrodynamic coma model described in Section \[sec:mmmodel\] of this article, and from the observed time-dependence of the water production rate. Within this framework, the parameter values found most suitable to reproduce both the brightness profile along the trail and its width were the following: The emission speeds of a 1mm radius particle at perihelion ranged from 6m/s to 12m/s, decreasing by a factor of 10 until 3 AU. The size dependence of $\beta$ could be characterised by Equation \[eq:beta\] with $Q_{\rm pr}/\rho$ between 1 and 3 cm$^3$/g, which – assuming $Q_{\rm pr}$ = 1 for large particles – corresponds to bulk densities between 0.3 and 0.9 g/cm$^3$. The size distribution exponent for particles larger than 100[$\mu$m]{} was in the range of -4.1 to -3.9. In order to reproduce both the optical and the infrared images, a geometric albedo for visible light of $p$ = 4% was required if the dust was assumed to be emitting as a blackbody ($\epsilon$ = 1 in Equation \[eq:T\_eq\]), but $p$ = 10% if $\epsilon / (1-A)$ = 0.6 [@hanner-tedesco1985a; @sykes-walker1992a]. Lower limits for the production rates of particles with $s$ $>$ 100[$\mu$m]{}ranged from 100 kg/s at perihelion to 0.2 kg/s at 3 AU, corresponding to [$Af\!\rho$]{} values of 4 and 0.05m, respectively. This implies that particles larger than 100[$\mu$m]{} would have contributed at least 50% of the total [$Af\!\rho$]{} observed while the comet was in the inner solar system, which is difficult to reconcile with the size distribution exponent on the order of -4. Three optical images (R-band) were obtained between September 2002 and February 2003 with the 1.05-m Schmidt telescope of the Kiso Observatory at Nagano, Japan [@ishiguro2008]. The difference in position angle of freshly emitted dust and the trail (dust from previous apparitions) was 1[$^{\circ}$]{} or larger in these images, thus – in contrast to images taken at later dates – a clear separation of both dust populations is possible. Taking into account dust emitted after aphelion in 1986 and assuming $p$ = 4%, $\rho$ = 1 g/cm$^3$, and $Q_{\rm pr}$ = 1, the images were best reproduced by a model with cone-shaped emission with a half opening angle between 45[$^{\circ}$]{} and 90[$^{\circ}$]{}, a size distribution exponent of -3.5 with particles in the range between 6[$\mu$m]{} and 5mm, and dust production rates of about 200 kg/s at perihelion and 15 kg/s at 3 AU. The emission speeds were assumed to be proportional to $s^{-1/2}$ and $r_{\rm h}^{-1/2}$, absolute values for a 1mm particle ranging from 8 to 18 m/s at perihelion and from 5 to 12 m/s at 3 AU. Two further observations were analysed by help of a Monte Carlo model [@kelley-reach2008]: an optical image (Gunn r’ filter) obtained in June 2003 with the 5-m Hale telescope at Palomar Observatory, and an infrared (24[$\mu$m]{}) image made with Spitzer/MIPS in February 2004. The trail of particles from previous perihelion passages was not detected in the Palomar image, but was visible to Spitzer. The applied model includes the assumptions of $Q_{\rm pr}$ = 1, $\rho$ = 1 g/cm$^3$, dust production rates $Q_{\rm d} \propto r_{\rm h}^{-5.8}$, and emission speeds $v \propto \sqrt{\beta/r_{\rm h}}$. Particles larger than 0.5[$\mu$m]{} and emitted after March 1993 were included in the analysis. The images were best fitted by a cosine-shaped distribution of the surface activity, peaking at the subsolar point, and a differential size distribution exponent of -3.5. The emission speeds of a 1mm particle at the subsolar point varied between 10 m/s at perihelion and 7 m/s at 3 AU. Summary of Modelling Results {#subsec:model_summary} ---------------------------- Figure \[fig:comparison\_model\] summarises the dust production rates, emission speeds and size distributions derived by help of the various models described in Sections \[subsec:model\_fulle\] to \[subsec:model\_trail\]. We wish to stress that – with the exception of the TNG image obtained in March 2003 – each model has been applied to a different set of images, and each model was able to reproduce the images that were analysed by it. However, the ranges of parameter values derived from the different models are considerable, such that it is at present difficult to derive a consistent picture of the CG dust environment on the basis of these results. Future work on this matter should focus on finding a set of parameters that is able to reproduce all available observations of the CG tail and trail. ![Comparison of dust characteristics derived from the different models described in Section \[sec:model\]. Two lines of the same style represent upper and lower limits. Top left: dust production rates. Top right: emission speeds of a dust grain with 1mm radius. For anisotropic models, the peak emission speed (assumed above the subsolar point) is given. Bottom: Exponents of the differential size distribution.[]{data-label="fig:comparison_model"}](comparison_Qd.eps "fig:"){width=".5\textwidth"} ![Comparison of dust characteristics derived from the different models described in Section \[sec:model\]. Two lines of the same style represent upper and lower limits. Top left: dust production rates. Top right: emission speeds of a dust grain with 1mm radius. For anisotropic models, the peak emission speed (assumed above the subsolar point) is given. Bottom: Exponents of the differential size distribution.[]{data-label="fig:comparison_model"}](comparison_speeds.eps "fig:"){width=".5\textwidth"} ![Comparison of dust characteristics derived from the different models described in Section \[sec:model\]. Two lines of the same style represent upper and lower limits. Top left: dust production rates. Top right: emission speeds of a dust grain with 1mm radius. For anisotropic models, the peak emission speed (assumed above the subsolar point) is given. Bottom: Exponents of the differential size distribution.[]{data-label="fig:comparison_model"}](comparison_sd.eps){width=".5\textwidth"} Dust Environment Model and Predictions for 67P/C–G {#sec:mmmodel} ================================================== In this section, the state of dust in the coma of [67P/C-G]{} is predicted by means of a hydrodynamic model [@mueller1997; @mueller1998; @muellerPhD; @landgraf-mueller1999]. Three emission scenarios and three size distribution exponents are studied, and estimates of the fluxes on a spacecraft for several types of trajectory and the radiation environment in the coma are presented. Coma Model and Parameters {#subsec:mmmodel_param} ------------------------- The employed model can be considered as the simplest possible physically consistent model of the inner coma: The comet nucleus is assumed to be of spherical shape, and the activity distribution on its surface is axis-symmetric with respect to the comet-Sun line. The flows of gas and dust around the comet are calculated rendering at any position in the coma the gas density, velocity, and temperature, and the number density and velocity of dust particles of different mass classes. Each class represents particles having masses within one decade in the range of $10^{-20}$ to $10^{4} \,{\rm kg}$. All dust particles of one class are assigned the same representative mass and radius. It is assumed that the particles have spherical shapes and a constant bulk density of $1000 \, {\rm kg/m}^3$. The numerical method used to compute the dynamics of gas and dust takes advantage of the fact that the influence of the dust on the gas flow is of minor importance. In a first step, the gas flow is calculated without taking into account the presence of dust. In a second step, dust trajectories are integrated considering the gravitation of the nucleus and the gas drag force. The gas flow is calculated under the assumptions that the gas is in thermal equilibrium everywhere, and that the mean molecular mass is constant across the coma. It is given by the mean of the masses of the two most common molecules, H$_2$O and CO, weighted by their overall abundances in the coma. The production rates of these species are input parameters to the model. To first order, the CO-activity can be treated as independent of heliocentric distance, and the constant value of $10^{26}$ molecules/s is adopted for 67P/C–G. An upper limit for the CO-production of $10^{27}$ molecules/s at 3 AU has been inferred from radio observations [@bockelee-moreno2004]. The water activity is assumed to vary with heliocentric distance and time as indicated by observations (Figure \[fig:afr\_h2o\], bottom). The dust activity is scaled such that the observed [$Af\!\rho$]{} is reproduced for the employed size distribution and albedo. Since, at present, the size distribution estimates vary considerably (Section \[subsec:model\_summary\]), we consider in the following the same range of mass distributions as in earlier models for 46P/Wirtanen [@mueller1997]: two “extreme” and a “nominal” one. All are characterised by different exponents for light and heavy particles, and the analytical form used for the cumulative mass distribution is as follows [@divine-newburn1987b]: $$\label{eq:divnew_md} F(m) = \left( \frac{(1+x)^{b-1}}{x^b} \right)^{ac} {\rm with} \:\: x = \left( \frac{m}{m_{\rm t}} \right)^{1/c}.$$ This function is specified by the positive parameters $a$, $b$, $c$, and $m_{\rm t}$. The exponent $-\gamma$ of the cumulative mass distribution tends towards $-a$ for heavy particles ($m \ll m_{\rm t}$) and towards $-ab$ for light ones. $m_{\rm t}$ is the mass where the transition between the two exponents takes place, and $c$ determines the sharpness of the transition. The three mass distributions considered in this paper are different only by their slopes for heavy particles. For light particles we use $\gamma = ab = 0.26$ throughout (corresponding to $\alpha = -1.8$), for the transition mass $m_{\rm t} = 10^{-13} \,{\rm kg}$, and $c = 2$. The slopes for large particles are as follows: the (velocity-corrected) fit to the fluence measured on Giotto [@divine-newburn1987b] gave $\alpha = -4.3$ ($\gamma = 1.1$) which we use as one of the “extreme” distributions. From the same data, an exponent in the range of $\alpha \in [-3.7, 3.3]$ has been derived [@fulle-colangeli1995], wherefore we adopt $\alpha = -3.7$ ($\gamma = 0.9$) as the “nominal” and $\alpha = -3.3$ ($\gamma = 0.8$) as the other “extreme” distribution. It must, however, be emphasised that for $\alpha > -3.5$, the observed brightness is dominated by light scattered by large particles [@mueller1997]. Figure \[fig:md\_moments\] show different moments of the mass distribution. ![Different moments of the mass distributions studied in Section \[sec:mmmodel\]. Top left: number fraction. Top right: area fraction. Bottom: mass fraction. To convert mass to radius, a bulk density of $\rho$ = 1000 kg/m$^3$ was assumed.[]{data-label="fig:md_moments"}](nfrac.eps "fig:"){width=".5\textwidth"} ![Different moments of the mass distributions studied in Section \[sec:mmmodel\]. Top left: number fraction. Top right: area fraction. Bottom: mass fraction. To convert mass to radius, a bulk density of $\rho$ = 1000 kg/m$^3$ was assumed.[]{data-label="fig:md_moments"}](areafrac.eps "fig:"){width=".5\textwidth"} ![Different moments of the mass distributions studied in Section \[sec:mmmodel\]. Top left: number fraction. Top right: area fraction. Bottom: mass fraction. To convert mass to radius, a bulk density of $\rho$ = 1000 kg/m$^3$ was assumed.[]{data-label="fig:md_moments"}](massfrac.eps){width=".5\textwidth"} In the following, model predictions are given for several heliocentric distances: the perihelion distances in 2015 ($r_{\rm h} = 1.24 \, {\rm AU}$) and 2002 ($r_{\rm h} = 1.29 \, {\rm AU}$), and $1.87 \, {\rm AU}$, $2.5 \, {\rm AU}$, and $3 \, {\rm AU}$. The values used for the water production rate, for [$Af\!\rho$]{}, and for the phase angle $\alpha$ are listed in Table \[tab:prodrates\]. We use a geometric albedo of 4% and the phase function shown in Figure \[fig:phase\_fns\]. \[tab:prodrates\] [llll]{} $r_{\rm h}$ \[AU\] & ${\rm Q}_{{\rm H}_2{\rm O}}$ \[molecules/s\] & $Af\!\rho \, (\alpha) \, [{\rm cm}]$ & $\alpha [{\rm deg}]$\ $1.24$ & $1.3 \times 10^{28}$ & $400$ & 35\ $1.29$ & $1.0 \times 10^{28}$ & $300$ & 35\ $1.87$ & $1.2 \times 10^{27}$ & $65$ & 32\ $2.5$ & $1.8 \times 10^{26}$ & $20$ & 5\ $3.0$ & $6.4 \times 10^{25}$ & $5.8$ & 19\ While the production of CO is assumed to be isotropic, for the release of H$_2$O three different surface activity distributions are taken into account. The first represents a homogeneously composed surface such that the strength of gas production is determined by the amount of solar energy received by a given surface element ([insolation driven model]{}). In the second scenario, a [jet]{} is simulated describing the surface activity by a Gaussian profile centred at the subsolar point [@fulle-colangeli1995; @agarwal-mueller2004]. This model is expected to yield upper limits for local quantities inside a jet, while the insolation driven model gives estimates for quantities averaged over time and space. In addition, an even simpler [radially symmetric model]{} is used in which the water activity is distributed equally over the comet surface. The average radius of the nucleus of 67P/C–G as given in the literature is in the range of $1.7 \, .. \, 3.2 \,{\rm km}$ [@mueller1992; @lowry-fitzsimmons2003; @tancredi-fernandez2000; @tancredi-fernandez2006; @lamy-toth2007; @tubiana-drahus2007_dps; @kelley-reach2008]. The geometric albedo of the nucleus in R-band is in the range of $0.045$ to $0.06$, and the bulk density of the nucleus is estimated to $370 \,{\rm kg/m^3}$ [@lamy-toth2007]. The following results were obtained for a spherical nucleus having a radius of 2 km and a geometric albedo of 4%. Dust Terminal Speeds {#subsec:mmmodel_speeds} -------------------- The terminal speeds of dust as a function of particle size for the three different emission scenarios are shown in Figure \[fig:termvel\]. For comparison, the corresponding values for comet 46P/Wirtanen, the former Rosetta target, are shown as well. Note that all curves are at perihelion, hence the comets are compared at different heliocentric distances. While the terminal speeds for 46P/Wirtanen are given at 10 km from the nucleus centre, those for 67P/C–G are at 20 km, because the dust particles have not quite reached their terminal speeds at lower heights. In the radially symmetric models, the terminal speeds are higher in the case of 46P/Wirtanen than of 67P/C–G. The reason is that both comets have similar gas production rates although 46P/Wirtanen is smaller. Hence the gas directly over the surface is denser and the dust reaches higher velocities. The ratio between the speeds expected for 67P/C–G and 46P/Wirtanen in the radially symmetric model ranges from 0.8 for small particles to 0.5 for large ones. The speeds shown in Figure \[fig:termvel\] were computed for a dust bulk density of $\rho = 1000 \,{\rm kg/m}^3$. Scaling to other particle densities can to first order be done assuming that particles with the same cross section to mass ratio reach the same velocities. For the larger particles, a power law of the following form was fitted to the calculated speeds: $$\label{eq:termvel_fit} v\,(s) = v_0 \left(\! \frac{s}{s_0} \!\right)^{\!-d},$$ with $s_0 = 1 {\rm mm}$. The resulting parameters for the different scenarios are given in Table \[tab:termvel\_fit\]. For the fits, all dust classes with particle radii $s > 0.1 \, {\rm mm}$ have been used. The class with the largest liftable particles was not included because of the strong influence of the gravitation of the comet. For 46P/Wirtanen, the exponent $d$ is closer to the expected value of 0.5 than for 67P/C–G because of the latter’s larger nucleus and resulting stronger gravity. ![ Terminal speeds as functions of mass-to-cross-section ratio for various emission scenarios (r.s.: radially symmetric, i.d.: insolation driven). The corresponding particle radii for a bulk density of $\rho = 1000 \, {\rm kg/m^3}$ are indicated at the upper margin. Values for 67P/C–G are given at 20 km from the nucleus centre above the subsolar point and at perihelion ($r_{\rm h} = 1.29\,{\rm AU}$). For 46P/Wirtanen they are at 10 km from the nucleus centre and at the perihelion distance of $r_{\rm h} = 1.06\,{\rm AU}$.[]{data-label="fig:termvel"}](term_vel.eps){width=".7\textwidth"} \[tab:termvel\_fit\] [llll]{} Model & $r_{\rm h}$ \[AU\] & $v_0$ \[m/s\] & Exponent $d$\ Wirtanen r.s. & 1.06 & 7.0 & 0.51\ C–G r.s. & 1.29 & 3.9 & 0.53\ C–G i.d. & 1.29 & 8.0 & 0.53\ C–G jet & 1.29 & 22.2 & 0.52\ Dust Flux at 1.29 AU {#subsec:mmmodel_flux} -------------------- Here we give numerical values for the fluxes of gas and dust on a surface directed towards the nucleus at 20 km from the nucleus centre above the subsolar point for various mass distributions and emission scenarios. As the speeds of gas and dust outside the acceleration zone are almost constant, the fluxes there can be scaled by an inverse square law. Note that the values given for the gas are only valid as long as the gas is cold, i.e. far from the nucleus. Within a few comet radii from the surface, the velocity distribution of the gas should be taken into account. Consequently, there will also be a flux on surfaces which are directed away from the flux direction. The numerical values are given in the appendix of [@agarwal-mueller2007] and visualised in Figure \[fig:fluxes\]. The number and mass fluxes (upper panels Figure \[fig:fluxes\]) in are relevant for different sensors of the GIADA instrument. The lower left panel shows the surface coverage by dust which is important for estimating thermal effects on spacecraft and instruments. The lower right panel describes the momentum transfer onto the spacecraft by the dust. ![Fluxes over particle mass for various mass distributions and emission geometries. Top left: number flux, top right: mass flux, bottom left: area flux, bottom right: momentum flux.[]{data-label="fig:fluxes"}](nflux.eps "fig:"){width=".5\textwidth"} ![Fluxes over particle mass for various mass distributions and emission geometries. Top left: number flux, top right: mass flux, bottom left: area flux, bottom right: momentum flux.[]{data-label="fig:fluxes"}](mass_flux.eps "fig:"){width=".5\textwidth"} ![Fluxes over particle mass for various mass distributions and emission geometries. Top left: number flux, top right: mass flux, bottom left: area flux, bottom right: momentum flux.[]{data-label="fig:fluxes"}](area_flux.eps "fig:"){width=".5\textwidth"} ![Fluxes over particle mass for various mass distributions and emission geometries. Top left: number flux, top right: mass flux, bottom left: area flux, bottom right: momentum flux.[]{data-label="fig:fluxes"}](mom_flux.eps "fig:"){width=".5\textwidth"} Dust Collection along Sample Spacecraft Trajectories {#subsec:mmmodel_traj} ---------------------------------------------------- During most of the mission, regions of strong dust emission in the coma will be avoided, because impacts of dust particles constitute a hazard to the Rosetta spacecraft. This hazard includes damage of sensitive surfaces by cratering and erosion due to high velocity dust impacts, contamination of surfaces and modification of their optical and thermal properties due to sticking of dust, and even mechanical blockage of gears and hinges by larger grains. This hazard has been taken into account in the design of the Rosetta spacecraft and its instruments, e.g. by painting the spacecraft black so that sticking dust will not significantly change the thermal properties of the spacecraft, and by employing movable shutters in front of sensitive surfaces such as cameras. Nevertheless, in the vicinity of the comet, spacecraft operations will keep exposure to dust to a minimum [@koschny-dhiri2007]. In order to fulfil the dust collection requirements of the COSIMA and MIDAS instruments the spacecraft occasionally will have to pass through regions of high dust density. The design of these passages will depend on the collection requirements of the dust instruments. For simple dust collection trajectories, the surface coverage by dust (assuming all dust sticks to the surface) has been calculated for the different dust emission scenarios and mass distributions described in Section \[subsec:mmmodel\_param\]. The considered trajectory types are one revolution on a circular orbit, a parabolic orbit, and a straight line trajectory passing the nucleus at a speed v = 1 m/s. All trajectories are assumed to pass at closest approach ($r_{\rm CA}=4$km from the nucleus centre) over the subsolar point, i.e. the region of highest dust density in our model. The fraction of the surface covered by dust after one passage of the nucleus is given in Table \[tab:sc\_fluences\]. \[tab:sc\_fluences\] [@llllllll]{} Model & Orbit & Size &&& $r_{\rm h}$ &&\ && distr.&&& \[AU\] &&\ && exp. & 1.24 & 1.29 & 1.87 & 2.5 & 3.0\ radial & circle & -3.3 & 0.101 & 0.075 & 0.006 & 0.0008 & 0.0004\ sym. & & -3.7 & 0.272 & 0.204 & 0.015 & 0.0015 & 0.0005\ & & -4.3 & 0.539 & 0.404 & 0.030 & 0.0028 & 0.0009\ insol. & parabola & -3.3 & 0.189 & 0.142 & 0.017 & 0.0016 & 0.0005\ driven & & -3.7 & 0.606 & 0.455 & 0.035 & 0.0026 & 0.0007\ & & -4.3 & 1.213 & 0.909 & 0.062 & 0.0042 & 0.0010\ & straight & -3.3 & 0.148 & 0.111 & 0.013 & 0.0013 & 0.0004\ & line & -3.7 & 0.473 & 0.355 & 0.027 & 0.0021 & 0.0006\ & & -4.3 & 0.945 & 0.709 & 0.048 & 0.0034 & 0.0009\ jet & parabola & -3.3 & 0.917 & 0.688 & 0.106 & &\ & & -3.7 & 3.540 & 2.655 & 0.249 & &\ & & -4.3 & 6.896 & 5.172 & 0.421 & &\ & straight & -3.3 & 0.685 & 0.514 & 0.079 & &\ & line & -3.7 & 2.644 & 1.983 & 0.185 & &\ & & -4.3 & 5.153 & 3.865 & 0.313 & &\ For other closest approach distances the surface coverage levels scale with $r_{\rm CA}^{-0.5}$ for Keplerian orbits (circles and parabola) and with $r_{\rm CA}^{-1}$ and $v^{-1}$ for straight line passages. The values in Table \[tab:sc\_fluences\] refer to different dust production rates, i.e. different heliocentric distances. A surface coverage of one implies roughly a mono-layer of differently sized dust grains, most of which are grains in the 1 to 10 [$\mu$m]{} size range (cf. Figure \[fig:md\_moments\], top right). Since the dust instruments are sensitive to different sizes of dust grains it is useful to determine the dust coverage by particles above a certain minimum size. This is achieved by applying the scaling factors for different minimum dust sizes given in Table \[tab:areacum\]. \[tab:areacum\] [@lllll]{} $m_{\rm min}$ \[kg\]& $s_{\rm min}$ \[m\]& $\alpha = -3.3$ & $\alpha=-3.7$ & $\alpha=-4.3$\ $1.000\times 10^{-20}$ & $1.337\times 10^{-08}$ & $1.000\times 10^{+00}$ & $1.000\times 10^{+00}$ & $1.000\times 10^{+00}$\ $1.000\times 10^{-19}$ & $2.879\times 10^{-08}$ & $9.998\times 10^{-01}$ & $9.996\times 10^{-01}$ & $9.992\times 10^{-01}$\ $1.000\times 10^{-18}$ & $6.204\times 10^{-08}$ & $9.993\times 10^{-01}$ & $9.984\times 10^{-01}$ & $9.972\times 10^{-01}$\ $1.000\times 10^{-17}$ & $1.337\times 10^{-07}$ & $9.979\times 10^{-01}$ & $9.955\times 10^{-01}$ & $9.921\times 10^{-01}$\ $1.000\times 10^{-16}$ & $2.879\times 10^{-07}$ & $9.945\times 10^{-01}$ & $9.880\times 10^{-01}$ & $9.788\times 10^{-01}$\ $1.000\times 10^{-15}$ & $6.204\times 10^{-07}$ & $9.855\times 10^{-01}$ & $9.683\times 10^{-01}$ & $9.438\times 10^{-01}$\ $1.000\times 10^{-14}$ & $1.337\times 10^{-06}$ & $9.618\times 10^{-01}$ & $9.166\times 10^{-01}$ & $8.524\times 10^{-01}$\ $1.000\times 10^{-13}$ & $2.879\times 10^{-06}$ & $9.039\times 10^{-01}$ & $7.959\times 10^{-01}$ & $6.512\times 10^{-01}$\ $1.000\times 10^{-12}$ & $6.204\times 10^{-06}$ & $7.971\times 10^{-01}$ & $5.958\times 10^{-01}$ & $3.771\times 10^{-01}$\ $1.000\times 10^{-11}$ & $1.337\times 10^{-05}$ & $6.620\times 10^{-01}$ & $3.897\times 10^{-01}$ & $1.703\times 10^{-01}$\ $1.000\times 10^{-10}$ & $2.879\times 10^{-05}$ & $5.310\times 10^{-01}$ & $2.379\times 10^{-01}$ & $6.782\times 10^{-02}$\ $1.000\times 10^{-09}$ & $6.204\times 10^{-05}$ & $4.201\times 10^{-01}$ & $1.408\times 10^{-01}$ & $2.565\times 10^{-02}$\ $1.000\times 10^{-08}$ & $1.337\times 10^{-04}$ & $3.302\times 10^{-01}$ & $8.245\times 10^{-02}$ & $9.514\times 10^{-03}$\ $1.000\times 10^{-07}$ & $2.879\times 10^{-04}$ & $2.576\times 10^{-01}$ & $4.822\times 10^{-02}$ & $3.528\times 10^{-03}$\ $1.000\times 10^{-06}$ & $6.204\times 10^{-04}$ & $2.001\times 10^{-01}$ & $2.813\times 10^{-02}$ & $1.298\times 10^{-03}$\ $1.000\times 10^{-05}$ & $1.337\times 10^{-03}$ & $1.545\times 10^{-01}$ & $1.637\times 10^{-02}$ & $4.787\times 10^{-04}$\ $1.000\times 10^{-04}$ & $2.879\times 10^{-03}$ & $1.181\times 10^{-01}$ & $9.518\times 10^{-03}$ & $1.773\times 10^{-04}$\ $1.000\times 10^{-03}$ & $6.204\times 10^{-03}$ & $8.923\times 10^{-02}$ & $5.508\times 10^{-03}$ & $6.523\times 10^{-05}$\ $1.000\times 10^{-02}$ & $1.337\times 10^{-02}$ & $6.637\times 10^{-02}$ & $3.167\times 10^{-03}$ & $2.395\times 10^{-05}$\ $1.000\times 10^{-01}$ & $2.879\times 10^{-02}$ & $4.813\times 10^{-02}$ & $1.796\times 10^{-03}$ & $8.815\times 10^{-06}$\ $1.000\times 10^{+00}$ & $6.204\times 10^{-02}$ & $3.366\times 10^{-02}$ & $9.940\times 10^{-04}$ & $3.204\times 10^{-06}$\ $1.000\times 10^{+01}$ & $1.337\times 10^{-01}$ & $2.214\times 10^{-02}$ & $5.256\times 10^{-04}$ & $1.143\times 10^{-06}$\ $1.000\times 10^{+02}$ & $2.879\times 10^{-01}$ & $1.300\times 10^{-02}$ & $2.531\times 10^{-04}$ & $3.844\times 10^{-07}$\ $1.000\times 10^{+03}$ & $6.204\times 10^{-01}$ & $5.748\times 10^{-03}$ & $9.349\times 10^{-05}$ & $1.035\times 10^{-07}$\ Useful surface coverage levels of dust collectors range from $10^{-4}$ to $10^{-2}$, i.e. a $ 1 \,{\rm cm}^2$ collection surface will contain approximately 100 to $10^4$ particles of $10\, \mu{\rm m}$ in size. According to Table \[tab:areacum\] these particles constitute about 50% of the covered surface (for the size distribution with $\alpha = -3.7$). The other half of the covered surface consists of more finely dispersed smaller grains. The COSIMA instrument is sensitive to dust grains of 10[$\mu$m]{} and has 23 individual collectors. The MIDAS instrument is sensitive to submicron-sized dust grains and has 60 individual collectors. The exposure of these collectors to the cometary dust flux and the collection of sufficient numbers of dust particles will at the same time lead to significant dust coverage of all spacecraft surfaces facing the comet. During one dust collection passage of Rosetta at large heliocentric distances (2.5 and 3 AU) an individual dust collector may reach only a total dust coverage of approximately $2 \times 10^{-3}$ and $7 \times 10^{-4}$, respectively. Close to perihelion, however, the comet-facing side of the spacecraft will be almost completely covered by dust during a single dust collection passage. Radiation Environment {#subsec:mmmodel_rad} --------------------- In the following we predict the radiation levels received in the vicinity of the comet nucleus due to the presence of dust. It is beyond the scope of the present work to introduce a detailed wavelength-dependent model. Instead, values will only be given for the total amount of scattered visible light and thermally emitted radiation. The crucial parameters for these estimates are the albedo, phase function and temperature of the dust particles. We adopt the phase function derived for comet 1P/Halley [@divine1981] as shown in Figure \[fig:phase\_fns\], a dust geometric albedo of 4% [@hanner-tedesco1985a], and a temperature at perihelion of 285 K (see Figure \[fig:dust\_temp\]). A priori it is not clear that it is valid to assume that the temperature of a particle of a given size is independent of its position in the coma, because the gas molecules impacting on the particle surface give rise to a heat exchange between the gas and the dust phase. This effect was taken into account in early models [@probstein1968; @kitamura1986a], but was found to be of minor importance later [@divine-fechtig1986; @knollenbergPhD1994]. Neglecting the heat exchange, the temperature of a dust particle is given by the equilibrium between incident radiation and thermal emission of the particle. If the coma is optically thin, which is shown to be true for the dust coma of 67P/C–G below, the incident radiation is dominated by solar illumination. Hence it is justified to treat the particle temperature as independent of the position in the coma. Assuming furthermore that the physical properties of the particles on their way from the inner to the outer coma do not change significantly, the infrared spectrum of dust near the nucleus can be approximated by the spectrum observed from Earth. The thermal emission spectrum of cometary dust particles can only approximately be represented by the spectrum of a blackbody, because particles emit efficiently only at wavelengths smaller than their size. Consequently, the smallest (or very porous) particles have temperatures much higher than a blackbody at the given heliocentric distance, and it is not physically consistent to set the dust temperature to a constant value independent of particle size. However, for the purpose of estimating the amount of radiation received by an observer, only the properties of the dust particles as an ensemble are of importance. The relative contribution to the total cross section of particles of a given size in most of the coma volume is the same as in the ensemble seen by an Earth-based observer. This was found with the present model which consistently models the dynamics of particles of different sizes. To calculate the intensity received inside the coma it is therefore valid to assign every particle independently of its size the ensemble properties as measured in Earth-based observations, which is henceforth done for the optical and thermal characteristics of the particles. Compared with the results given in the previous subsections – which were strongly controlled by model parameters only very indirectly accessible to ground-based observations (e.g. particle size distribution, particle density) – the results presented here only depend on parameters readily measured from ground. For example the estimates of the scattered visual radiation are governed solely by the [$Af\!\rho$]{} value and the dust phase function. It may, at first glance, be surprising that the results are not influenced by the dust albedo. However, because in the present model the overall dust activity is adjusted in order to match an observed [$Af\!\rho$]{} value, the predictions of the scattered visual radiation inside the coma are merely a scaling of the Earth-based observations to a different geometry. To give an idea of the accuracy of the results obtained for the scattered radiation: The present model applied to comet 1P/Halley during the fly-by of Giotto reproduces the measurements of the Halley Optical Probe experiment [@levasseur-bertaux1986; @fulle-levasseur2000] by a factor of only 1.5. By contrast, the predictions of the optical thickness and of the thermal radiation by the dust depend on more parameters, introducing uncertainty: The estimated optical thickness is influenced by the dust albedo and by the – poorly constrained – extinction efficiency, while the thermal radiation relies on the dust temperature. Using an extinction efficiency of $q_{\rm ext}=1$ we find for the optical thickness at perihelion at the subsolar point values of $\tau=0.013$ in the insolation driven model and $\tau=0.09$ in the jet-model. In the insolation driven model the optical thickness of the coma along the line of sight from a point at the surface to the Sun is fairly constant over the comet surface and the results can well be approximated by the simplified model in [@mueller-green2002a]. In the jet model the optical thickness decreases towards the terminator by a factor of $5$. Since the extinction efficiency $q_{\rm ext}$ is little constrained and can assume values in the range $q_{\rm ext} \approx 1-2$ we estimate that at perihelion the optical thickness at the comet surface is typically in the range $\tau \approx 0.01-0.03$. Within jets it can reach values up to $\tau \approx 0.2$. Since the optical thickness is proportional to [$Af\!\rho$]{}, the results can easily be scaled to scenarios with different levels of cometary activity. The intensity received by an observer at 4 km (2 comet radii) from the nucleus centre over the subsolar point for different lines of sight is shown in Figure \[fig:rad\_los\]. For the insolation driven model, the observed visible intensity reaches maxima for lines of sight directed towards the Sun and the nucleus (phase angle $180^\circ$ and $0^\circ$, respectively). These peaks are not found in the infrared intensity received in the insolation driven model, which shows that they are due to the forward and backward scattering peaks of the dust phase function (Figure \[fig:phase\_fns\]). In the insolation driven model the intensities, both for visible and infrared radiation, reach maxima for the lines of sight that touch the nucleus tangentially (at about $150^\circ$). In the jet-model these peaks are not present because most of the dust is concentrated to a narrow region over the subsolar point. This also explains the maxima of the infrared intensity in the jet-model for the lines of sight in the Sun- and the nucleus-direction that are not observed in the insolation driven model. Correspondingly, the maxima for the visible radiation on these lines of sight are more pronounced in the jet-model than in the insolation driven model. The total intensity received by a surface depends on its orientation. To give an example, we consider a plane surface at 4 km over the subsolar point (see Figure \[fig:rad\_plane\]) taking into account also the radiation from the nucleus. The intensities emitted or scattered by the nucleus dominate over those from the dust; and, since the nucleus is a dark object, the infrared radiation dominates the total intensity. Note that the nucleus temperature in the model is calculated as that of a pure icy surface. Thus the given results for the infrared radiation received from the nucleus are lower limits. If the major part of the comet surface is inactive, the surface temperature and consequently the infrared flux will be much higher. ![Scattered sunlight and thermal infrared radiation at perihelion received from a line of sight by an observer above the subsolar point at 4 km from the nucleus centre. Only the contribution by the dust is considered, in contrast to radiation from the nucleus or direct sunlight. The nominal size distribution $\alpha = -3.7$ was used. Left: insolation driven model, right: jet model.[]{data-label="fig:rad_los"}](los_sd.eps "fig:"){width="50.00000%"} ![Scattered sunlight and thermal infrared radiation at perihelion received from a line of sight by an observer above the subsolar point at 4 km from the nucleus centre. Only the contribution by the dust is considered, in contrast to radiation from the nucleus or direct sunlight. The nominal size distribution $\alpha = -3.7$ was used. Left: insolation driven model, right: jet model.[]{data-label="fig:rad_los"}](los_jt.eps "fig:"){width="50.00000%"} ![Radiation at perihelion from all directions on a surface above the subsolar point at 4 km from the nucleus centre. Both curves with and without the contribution from the nucleus are shown, but no direct sunlight is considered. The nominal size distribution $\alpha = -3.7$ was used. Left: insolation driven model, right: jet model.[]{data-label="fig:rad_plane"}](plane_sd.eps "fig:"){width=".5\textwidth"} ![Radiation at perihelion from all directions on a surface above the subsolar point at 4 km from the nucleus centre. Both curves with and without the contribution from the nucleus are shown, but no direct sunlight is considered. The nominal size distribution $\alpha = -3.7$ was used. Left: insolation driven model, right: jet model.[]{data-label="fig:rad_plane"}](plane_jet.eps "fig:"){width=".5\textwidth"} Summary and Conclusions {#sec:summary} ======================= We have summarised the present knowledge on the dust emitted by Rosetta target comet 67P/Churyumov-Gerasimenko and discussed methods and results of modelling it. The observational data can be described as follows: The dust and gas production rates peaked about 30 days after perihelion passage during three past apparitions. The coma was characterised by azimuthal brightness variations probably resulting from active areas on the surface. Close to the comet orbit, a distinct line-shaped feature prevailed at least from shortly after perihelion 2002 until 2006, beyond aphelion. It is interpreted as a pronounced antitail due to the low inclination of the comet orbital plane towards the ecliptic. The colour temperatures in both the coma and the trail were higher than the blackbody equilibrium temperature at the concerned heliocentric distance. We have reviewed different approaches to derive parameters of the cometary dust production by modelling images of the tail and/or trail, and the results obtained by their application to observations of [67P/C-G]{}. We found that due to the large diversity of the derived parameter values it is at present not possible to formulate a consistent picture of the [67P/C-G]{} dust activity and its time evolution. We identified the need to find a common model that satisfies all available observations. Using the ESA Cometary Dust Environment Model we have predicted the terminal speeds of dust, the dust flux in the coma and along sample trajectories of the spacecraft, and the radiation flux in the coma. We listed results for three mass distributions and five heliocentric distances. We also considered different surface activity distributions, among which one implying a homogeneous surface of the nucleus and one assuming all dust to be emitted by a single active area. The former yielded estimates of quantities averaged over space and time, while the latter served to derive upper limits for local quantities. M. M. gratefully acknowledges funding by EDS Operations Services GmbH. We wish to thank R. Schulz et al. [@schulz-stuewe2004a; @schulz-stuewe2004b] for making their data available to us. [^1]: The data from these publications are summarised in the data base at . The site also includes data obtained by amateur astronomers.	{ "pile_set_name": "ArXiv" }
--- abstract: 'General-purpose, intelligent, learning agents cycle through sequences of observations, actions, and rewards that are complex, uncertain, unknown, and non-Markovian. On the other hand, reinforcement learning is well-developed for small finite state Markov decision processes (MDPs). Up to now, extracting the right state representations out of bare observations, that is, reducing the general agent setup to the MDP framework, is an art that involves significant effort by designers. The primary goal of this work is to automate the reduction process and thereby significantly expand the scope of many existing reinforcement learning algorithms and the agents that employ them. Before we can think of mechanizing this search for suitable MDPs, we need a formal objective criterion. The main contribution of this article is to develop such a criterion. I also integrate the various parts into one learning algorithm. Extensions to more realistic dynamic Bayesian networks are developed in Part II [@Hutter:09phidbnx]. The role of POMDPs is also considered there.' author: - \| [Marcus Hutter[^1]]{}\ RSISE$\,$@$\,$ANU and SML$\,$@$\,$NICTA\ Canberra, ACT, 0200, Australia\ `[email protected] www.hutter1.net` date: 9 June 2009 title: \| **** ------------------------------------------------------------------------ height5pt Feature Reinforcement Learning:\ Part I. Unstructured MDPs ------------------------------------------------------------------------ height2pt --- Reinforcement learning; Markov decision process; partial observability; feature learning; explore-exploit; information & complexity; rational agents. > “Approximations, after all, may be made in two places - in the construction of the model and in the solution of the associated equations. It is not at all clear which yields a more judicious approximation.” > > — [Richard Bellman (1961)]{} Introduction {#secIntro} ============ Artificial General Intelligence (AGI) is concerned with designing agents that perform well in a wide range of environments [@Goertzel:07; @Hutter:07iorx]. Among the well-established “narrow” Artificial Intelligence (AI) approaches [@Russell:03], arguably Reinforcement Learning (RL) [@Sutton:98] pursues most directly the same goal. RL considers the general agent-environment setup in which an agent interacts with an environment (acts and observes in cycles) and receives (occasional) rewards. The agent’s objective is to collect as much reward as possible. Most if not all AI problems can be formulated in this framework. Since the future is generally unknown and uncertain, the agent needs to learn a model of the environment based on past experience, which allows to predict future rewards and use this to maximize expected long-term reward. The simplest interesting environmental class consists of finite state fully observable Markov Decision Processes (MDPs) [@Puterman:94; @Sutton:98], which is reasonably well understood. Extensions to continuous states with (non)linear function approximation [@Sutton:98; @Gordon:99], partial observability (POMDP) [@Kaelbling:98; @Ross:08pomdp], structured MDPs (DBNs) [@Strehl:07], and others have been considered, but the algorithms are much more brittle. A way to tackle complex real-world problems is to reduce them to finite MDPs which we know how to deal with efficiently. This approach leaves a lot of work to the designer, namely to extract the right state representation (“features”) out of the bare observations in the initial (formal or informal) problem description. Even if [potentially]{} useful representations have been found, it is usually not clear which ones will turn out to be better, except in situations where we already know a perfect model. Think of a mobile robot equipped with a camera plunged into an unknown environment. While we can imagine which image features will potentially be useful, we cannot know in advance which ones will actually be useful. The primary goal of this paper is to develop and investigate a method that [automatically]{} selects those features that are necessary and sufficient for [reducing]{} a complex real-world problem to a computationally tractable MDP. Formally, we consider maps $\Phi$ from the past observation-reward-action history $h$ of the agent to an MDP state. Histories not worth being distinguished are mapped to the same state, i.e. $\Phi^{-1}$ induces a partition on the set of histories. We call this model $\Phi$MDP. A state may be simply an abstract label of the partition, but more often is itself a structured object like a discrete vector. Each vector component describes one feature of the history [@Hutter:09phidbn; @Hutter:09phidbnx]. For example, the state may be a 3-vector containing (shape,color,size) of the object a robot tracks. For this reason, we call the [reduction]{}, [Feature RL]{}, although in this Part I only the simpler unstructured case is considered. $\Phi$ maps the agent’s experience over time into a sequence of MDP states. Rather than informally constructing $\Phi$ by hand, our goal is to develop a formal objective criterion $\Cost(\Phi\|h)$ for [evaluating]{} different reductions $\Phi$. Obviously, at any point in time, if we want the criterion to be effective it can only depend on the agent’s past experience $h$ and possibly generic background knowledge. The “Cost” of $\Phi$ shall be small iff it leads to a “good” MDP representation. The establishment of such a criterion transforms the, in general, ill-defined RL problem to a formal optimization problem (minimizing Cost) for which efficient algorithms need to be developed. Another important question is which problems [can]{} profitably be reduced to MDPs [@Hutter:09phidbn; @Hutter:09phidbnx]. The real world does not conform itself to nice models: Reality is a non-ergodic partially observable uncertain unknown environment in which acquiring experience can be expensive. So we should exploit the data (past experience) at hand as well as possible, cannot generate virtual samples since the model is not given (need to be learned itself), and there is no reset-option. No criterion for this general setup exists. Of course, there is previous work which is in one or another way related to $\Phi$MDP. As partly detailed later, the suggested $\Phi$MDP model has interesting connections to many important ideas and approaches in RL and beyond: - $\Phi$MDP side-steps the open problem of learning POMDPs [@Kaelbling:98], - Unlike Bayesian RL algorithms [@Dearden:99; @Duff:02; @Poupart:06; @Ross:08bayes], $\Phi$MDP avoids learning a (complete stochastic) observation model, - $\Phi$MDP is a scaled-down practical instantiation of AIXI [@Hutter:04uaibook; @Hutter:07aixigentle], - $\Phi$MDP extends the idea of state-aggregation from planning (based on bi-simulation metrics [@Givan:03]) to RL (based on information), - $\Phi$MDP generalizes U-Tree [@McCallum:96] to arbitrary features, - $\Phi$MDP extends model selection criteria to general RL problems [@Gruenwald:07book], - $\Phi$MDP is an alternative to PSRs [@Singh:03] for which proper learning algorithms have yet to be developed, - $\Phi$MDP extends feature selection from supervised learning to RL [@Guyon:03]. Learning in agents via rewards is a much more demanding task than “classical” machine learning on independently and identically distributed (i.i.d.) data, largely due to the temporal credit assignment and exploration problem. Nevertheless, RL (and the closely related adaptive control theory in engineering) has been applied (often unrivaled) to a variety of real-world problems, occasionally with stunning success (Backgammon, Checkers, [@Sutton:98 Chp.11], helicopter control [@Ng:04]). $\Phi$MDP overcomes several of the limitations of the approaches in the items above and thus broadens the applicability of RL. $\Phi$MDP owes its general-purpose [learning]{} and [planning]{} ability to its [information]{} and [complexity]{} theoretical foundations. The implementation of $\Phi$MDP is based on (specialized and general) [search]{} and [optimization]{} algorithms used for finding good reductions $\Phi$. Given that $\Phi$MDP aims at general AI problems, one may wonder about the role of other aspects traditionally considered in AI [@Russell:03]: [knowledge representation]{} (KR) and [logic]{} may be useful for representing complex reductions $\Phi(h)$. Agent interface fields like [robotics]{}, computer [vision]{}, and natural [language]{} processing can speedup learning by pre&post-processing the raw observations and actions into more structured formats. These representational and interface aspects will only barely be discussed in this paper. The following diagram illustrates $\Phi$MDP in perspective. =2ex (32,20)(0,0) (16,17.5)[(10,3)(0,1)\[cc\][Universal AI]{}(0,-1)\[cc\][(AIXI)]{}]{} (16,11)[(14,2)(0,0)\[cc\][$\Phi$MDP / $\Phi$DBN ]{}]{} (5.5,6)[(7,2)(0,0)\[cc\][Information]{}]{} (12.5,6)[(7,2)(0,0)\[cc\][Learning]{}]{} (19.5,6)[(7,2)(0,0)\[cc\][Planning]{}]{} (26.5,6)[(7,2)(0,0)\[cc\][Complexity]{}]{} (0,2)[(32,2)\[cc\][Search – Optimization – Computation – Logic – KR]{}]{} (0,4)[(1,1)[2]{}]{}(3,7)[(1,1)[9]{}]{}(15,19)[(1,1)[1]{}]{} (32,4)[(-1,1)[2]{}]{}(29,7)[(-1,1)[9]{}]{}(17,19)[(-1,1)[1]{}]{} (8,7)[(2,3)[2]{}]{}(11,12)[(2,3)[2.667]{}]{} (24,7)[(-2,3)[2]{}]{}(21,12)[(-2,3)[2.667]{}]{} (13,7)[(1,4)[0.75]{}]{}(14,12)[(1,4)[1]{}]{} (19,7)[(-1,4)[0.75]{}]{}(18,12)[(-1,4)[1]{}]{} (5.5,4)[(0,1)[1]{}]{} (12.5,4)[(0,1)[1]{}]{} (19.5,4)[(0,1)[1]{}]{} (26.5,4)[(0,1)[1]{}]{} (0,0)[(16,2)\[lb\][Agents = Framework,]{}]{} (16,0)[(16,2)\[rb\][Interface = Robots,Vision,Language]{}]{} Section \[secPhiMDP\] formalizes our $\Phi$MDP setup, which consists of the agent model with a map $\Phi$ from observation-reward-action histories to MDP states. Section \[secCE\] develops our core $\Phi$ selection principle, which is illustrated in Section \[secTE\] on a tiny example. Section \[secCM\] discusses general search algorithms for finding (approximations of) the optimal $\Phi$, concretized for context tree MDPs. In Section \[secEE\] I find the optimal action for $\Phi$MDP, and present the overall algorithm. Section \[secICF\] improves the $\Phi$ selection criterion by “integrating” out the states. Section \[secDisc\] contains a brief discussion of $\Phi$MDP, including relations to prior work, incremental algorithms, and an outlook to more realistic [structured]{} MDPs (dynamic Bayesian networks, $\Phi$DBN) treated in Part II. Rather than leaving parts of $\Phi$MDP vague and unspecified, I decided to give at the very least a simplistic concrete algorithm for each building block, which may be assembled to one sound system on which one can build on. Throughout this article, $\log$ denotes the binary logarithm, $\epstr$ the empty string, and $\delta_{x,y}=\delta_{xy}=1$ if $x=y$ and $0$ else is the Kronecker symbol. I generally omit separating commas if no confusion arises, in particular in indices. For any $x$ of suitable type (string,vector,set), I define string $\v x = x_{1:l} = x_1...x_l$, sum $x_\p=\sum_j x_j$, union $x_=\bigcup_j x_j$, and vector $\v x_\vi=(x_1,...,x_l)$, where $j$ ranges over the full range $\{1,...,l\}$ and $l=\|{\rm x}\|$ is the length or dimension or size of $\rm x$. $\hat x$ denotes an estimate of $x$. $\P(\cdot)$ denotes a probability over states and rewards or parts thereof. I do not distinguish between random variables $X$ and realizations $x$, and abbreviation $\P(x):=\P[X=x]$ never leads to confusion. More specifically, $m\in\SetN$ denotes the number of states, $i\in\{1,...,m\}$ any state index, $n\in\SetN$ the current time, and $t\in\{1,...,n\}$ any time in history. Further, in order not to get distracted at several places I gloss over initial conditions or special cases where inessential. Also 0$$undefined=0$$infinity:=0. Feature Markov Decision Process ($\mathbf\Phi$MDP) {#secPhiMDP} ================================================== This section describes our formal setup. It consists of the agent-environment framework and maps $\Phi$ from observation-reward-action histories to MDP states. I call this arrangement “Feature MDP” or short $\Phi$MDP. I consider the standard agent-environment setup [@Russell:03] in which an [Agent]{} interacts with an [Environment]{} The agent can choose from actions $a\in\A$ (e.g. limb movements) and the environment provides (regular) observations $o\in\O$ (e.g.camera images) and real-valued rewards $r\in\R\subseteq\SetR$ to the agent. The reward may be very scarce, e.g. just $+1$ ($-1$) for winning (losing) a chess game, and 0 at all other times [@Hutter:04uaibook Sec.6.3]. This happens in cycles $t=1,2,3,...$: At time $t$, after observing $o_t$ and receiving reward $r_t$, the agent takes action $a_t$ based on history $h_t:=o_1 r_1 a_1...o_{t-1} r_{t-1} a_{t-1} o_t r_t$. Then the next cycle $t+1$ starts. The agent’s objective is to maximize his long-term reward. Without much loss of generality, I assume that $\R$ is finite. Finiteness of $\R$ is lifted in [@Hutter:09phidbn; @Hutter:09phidbnx]. I also assume that $\A$ is finite and small, which is restrictive. Part II deals with large state spaces, and large (structured) action spaces can be dealt with in a similar way. No assumptions are made on $\O$; it may be huge or even infinite. Indeed, $\Phi$MDP has been specifically designed to cope with huge observation spaces, e.g. camera images, which are mapped to a small space of relevant states. The agent and environment may be viewed as a pair or triple of interlocking functions of the history $\H:=(\O\times\A\times\R)^\times\O\times\R$: & & :Ø, o\_n r\_n = (h\_[n-1]{}a\_[n-1]{}),\ & & :, a\_n = (h\_n), =1.2ex (18,0)(0,-2) (3,3)[(6,2)\[cc\](0,0)\[cc\][Agent]{}]{} (15,3)[(6,2)\[cc\](0,0)\[cc\][Env()]{}]{} (3,2)[(0,-1)[2]{}]{} (3,0)[(1,0)[12]{}]{}(9,0.5)[(0,0)\[cb\][$\boldsymbol a$ction]{}]{} (15,0)[(0,1)[2]{}]{} (14,4)[(0,1)[2]{}]{} (14,6)[(-1,0)[10]{}]{}(9,6)[(0,0)\[ct\][$\boldsymbol r$eward]{}]{} (4,6)[(0,-1)[2]{}]{} (16,4)[(0,1)[4]{}]{} (16,8)[(-1,0)[14]{}]{}(9,8)[(0,0)\[ct\][$\boldsymbol o$bservation]{}]{} (2,8)[(0,-1)[4]{}]{} where $\leadsto$ indicates that mappings $\to$ might be stochastic. The goal of AI is to design agents that achieve high (expected) reward over the agent’s lifetime. For known (), finding the reward maximizing agent is a well-defined and formally solvable problem [@Hutter:04uaibook Chp.4], with computational efficiency being the “only” matter of concern. For most real-world AI problems () is at best partially known. For unknown (), the meaning of expected reward maximizing is even conceptually a challenge [@Hutter:04uaibook Chp.5]. Narrow AI considers the case where function () is either known (like planning in blocks world), or essentially known (like in chess, where one can safely model the opponent as a perfect minimax player), or () belongs to a relatively small class of environments (e.g.elevator or traffic control). The goal of AGI is to design agents that perform well in a large range of environments [@Hutter:07iorx], i.e. achieve high reward over their lifetime with as little as possible assumptions about Env(). A minimal necessary assumption is that the environment possesses [some]{} structure or pattern [@Wolpert:97]. From real-life experience (and from the examples below) we know that usually we do not need to know the complete history of events in order to determine (sufficiently well) what will happen next and to be able to perform well. Let $\Phi(h)$ be such a “useful” summary of history $h$. The following examples show that many problems can be reduced (approximately) to finite MDPs, thus showing that $\Phi$MDP can deal with a large variety of problems: In full-information [games]{} (like chess) with a static opponent, it is sufficient to know the current state of the game (board configuration) to play well (the history plays no role), hence $\Phi(h_t)=o_t$ is a sufficient summary (Markov condition). Classical [physics]{} is essentially predictable from the position and velocity of objects at a single time, or equivalently from the locations at two consecutive times, hence $\Phi(h_t)=o_{t-1}o_t$ is a sufficient summary (2nd order Markov). For [i.i.d. processes]{} of unknown probability (e.g. clinical trials $\simeq$ Bandits), the frequency of observations $\Phi(h_n)=(\sum_{t=1}^n\delta_{o_t o})_{o\in\O}$ is a sufficient statistic. In a [POMDP planning]{} problem, the so-called belief vector at time $t$ can be written down explicitly as some function of the complete history $h_t$ (by integrating out the hidden states). $\Phi(h_t)$ could be chosen as (a discretized version of) this belief vector, showing that $\Phi$MDP generalizes POMDPs. Obviously, the [identity]{} $\Phi(h)=h$ is always sufficient but not very useful, since () as a function of $\H$ is hard to impossible to “learn”. This suggests to look for $\Phi$ with small codomain, which allow to learn/estimate/approximate $\Env$ by $\widehat\Env$ such that $o_t r_t\approx\widehat\Env(\Phi(h_{t-1}))$ for $t=1...n$. Consider a robot equipped with a camera, i.e. $o$ is a pixel image. Computer vision algorithms usually extract a set of features from $o_{t-1}$ (or $h_{t-1}$), from low-level patterns to high-level objects with their spatial relation. Neither is it possible nor necessary to make a precise prediction of $o_t$ from summary $\Phi(h_{t-1})$. An approximate prediction must and will do. The difficulty is that the similarity measure “$\approx$” needs to be context dependent. Minor image nuances are irrelevant when driving a car, but when buying a painting it makes a huge difference in price whether it’s an original or a copy. Essentially only a bijection $\Phi$ would be able to extract [all potentially]{} interesting features, but such a $\Phi$ defeats its original purpose. It is of utmost importance to properly formalize the meaning of “$\approx$” in a general, domain-independent way. Let $s_t:=\Phi(h_t)$ summarize all relevant information in history $h_t$. I call $s$ a state or feature (vector) of $h$. “Relevant” means that the future is predictable from $s_t$ (and $a_t$) alone, and that the relevant future is coded in $s_{t+1}s_{t+2}...$. So we pass from the complete (and known) history $o_1 r_1 a_1...o_n r_n a_n$ to a “compressed” history $sra_{1:n}\equiv s_1 r_1 a_1...s_n r_n a_n$ and seek $\Phi$ such that $s_{t+1}$ is (approximately a stochastic) function of $s_t$ (and $a_t$). Since the goal of the agent is to maximize his rewards, the rewards $r_t$ are always relevant, so they (have to) stay untouched (this will become clearer below). The structure derived above is a classical Markov Decision Process (MDP), but the primary question I ask is not the usual one of finding the value function or best action or comparing different models of a given state sequence. I ask how well can the state-action-reward sequence generated by $\Phi$ be modeled as an MDP compared to other sequences resulting from different $\Phi$. A good $\Phi$ leads to a good model for predicting future rewards, which can be used to find good actions that maximize the agent’s expected long-term reward. $\mathbf\Phi$MDP Coding and Evaluation {#secCE} ====================================== I first review a few standard codes and model selection methods for i.i.d. sequences, subsequently adapt them to our situation, and show that they are suitable in our context. I state my Cost function for $\Phi$, and the $\Phi$ selection principle, and compare it to the Minimum Description Length (MDL) philosophy. Consider i.i.d. $x_1...x_n\in\X^n$ for finite $\X=\{1,...,m\}$. For known $\t_i=\P[x_t=i]$ we have $\P(x_{1:n}\|\v\t)=\t_{x_1}\cdot...\cdot\t_{x_n}$. It is well-known that there exists a code (e.g. arithmetic or Shannon-Fano) for $x_{1:n}$ of length $-\log\P(x_{1:n}\|\v\t)$, which is asymptotically optimal with probability one [@Barron:85 Thm.3.1]. This also easily follows from [@Cover:06 Thm.5.10.1]. [MDL/MML code [@Gruenwald:07book; @Wallace:05]:]{} For unknown $\v\t$ we may use a frequency estimate $\hat\t_i=n_i/n$, where $n_i=\|\{t\leq n:x_t=i\}\|$. Then it is easy to see that $-\log\P(x_{1:n}\|\v{\hat\t})=n\,H(\v{\hat\t})$, where H(v) := -\_[i=1]{}\^[m]{}\_i\_i v ($0\log 0:=0=:0\log{0\over 0}$). We also need to code $\v{\hat\t}$, or equivalently $(n_i)$, which naively needs $\log n$ bits for each $i$. In general, a sample size of $n$ allows estimating parameters only to accuracy $O(1/\sqrt{n})$, which is essentially equivalent to the fact that $\log\P(x_{1:n}\|\v{\hat\t}\pm O(1/\sqrt{n}))-\log\P(x_{1:n}\|\v{\hat\t}) = O(1)$. This shows that it is sufficient to code each $\hat\t_i$ to accuracy $O(1/\sqrt{n})$, which requires only $\fr12\log n+O(1)$ bits each. Hence, given $n$ and ignoring $O(1)$ terms, the overall code length (CL) of $x_{1:n}$ for unknown frequencies is \[iidCodeL\] (x\_[1:n]{}) (ň) := nH(ň/n) + 2n n>0 0 where $\v n=(n_1,...,n_m)$ and $n=n_+=n_1+...+n_m$. We have assumed that $n$ is given, hence only $m-1$ of the $n_i$ need to be coded, since the $m$th one can be reconstructed from them and $n$. The above is an exact code of $x_{1:n}$, which is optimal (within $+O(1)$) for all i.i.d. sources. This code may further be optimized by only coding $\hat\t_i$ for the $m'=\|\{i:n_i>0\}\|\leq m$ non-empty categories, resulting in a code of length \[iidCodeLp\] ’(ň) := nH(ň/n) + 2n + m, where the $m$ bits are needed to indicate which of the $\hat\t_i$ are coded. We refer to this improvement as [sparse]{} code. [Combinatorial code [@Li:08]:]{} A second way to code the data is to code $\v n$ exactly, and then, since there are $n!/n_1!...n_{m}!$ sequences $x_{1:n}$ with counts $\v n$, we can easily construct a code of length $\log(n!/n_1!...n_{m}!)$ given $\v n$ by enumeration, i.e. ”(ň) := (n!/n\_1!...n\_[m]{}!) + (m-1)n Within $\pm O(1)$ this code length also coincides with . [Incremental code [@Willems:97]:]{} A third way is to use a sequential estimate $\hat\t_i^{t+1}={t_i+\a\over t+m\a}$ based on [known]{} past counts $t_i=\|\{t'\leq t:x_{t'}=i\}\|$, where $\a>0$ is some regularizer. Then \[iidGamma\] ¶(x\_[1:n]{}) = \_[x\_1]{}\^1...\_[x\_n]{}\^n = C\_, C\_:=[(m)()\^[m]{}]{} where $\Gamma$ is the Gamma function. The logarithm of this expression again essentially reduces to (for any $\a>0$, typically $\fr12$ or 1), which can also be written as ”’(ň)=\_[i:n\_i>0]{}(n\_i)-(n)+O(1) n>0 [Bayesian code [@Schwarz:78; @MacKay:03]:]{} A fourth (the Bayesian) way is to assume a Dirichlet($\a$) prior over $\v\t$. The marginal distribution (evidence) is identical to and the Bayesian Information Criterion (BIC) approximation leads to code . [Conclusion:]{} All four methods lead to essentially the same code length. The references above contain rigorous derivations. In the following I will ignore the $O(1)$ terms and refer to simply as [the]{} code length. Note that $x_{1:n}$ is coded exactly (lossless). Similarly (see MDP below) sampling models more complex than i.i.d. may be considered, and the one that leads to the shortest code is selected as the best model [@Gruenwald:07book]. Recall that a sequence $sra_{1:n}$ is said to be sampled from an MDP $(\S,\A,T,R)$ iff the probability of $s_t$ only depends on $s_{t-1}$ and $a_{t-1}$; and $r_t$ only on $s_{t-1}$, $a_{t-1}$, and $s_t$. That is, ¶(s\_t\|h\_[t-1]{}a\_[t-1]{}) = ¶(s\_t\|s\_[t-1]{},a\_[t-1]{}) &=:& T\_[s\_[t-1]{}s\_t]{}\^[a\_[t-1]{}]{}\ ¶(r\_t\|h\_t) = ¶(r\_t\|s\_[t-1]{},a\_[t-1]{},s\_t) &=:& R\_[s\_[t-1]{}s\_t]{}\^[a\_[t-1]{}r\_t]{} In our case, we can identify the state-space $\S$ with the states $s_1,...,s_n$ “observed” so far. Hence $\S=\{s^1,...,s^m\}$ is finite and typically $m\ll n$, since states repeat. Let $s\stackrel{a}\to s'(r')$ be shorthand for “action $a$ in state $s$ resulted in state $s'$ (reward $r'$)”. Let $\T_{ss'}^{ar'}:=\{t\leq n:s_{t-1}=s, a_{t-1}=a, s_t=s', r_t=r'\}$ be the set of times $t-1$ at which $s\stackrel{a}\to s'r'$, and $n_{ss'}^{ar'}:=\|\T_{ss'}^{ar'}\|$ their number ($n_{\p\p}^{\p\p}=n$). For some fixed $s$ and $a$, consider the subsequence $s_{t_1}...s_{t_{n'}}$ of states reached from $s$ via $a$ ($s\stackrel{a}\to s_{t_i}$), i.e.$\{t_1,...,t_{n'}\}=\T_{s}^{a}$, where $n'=n_{s\p}^{a\p}$. By definition of an MDP, this sequence is i.i.d. with $s'$ occurring $n'_{s'}:=n_{ss'}^{a\p}$ times. By we can code this sequence in $\CL(\v n')$ bits. The whole sequence $s_{1:n}$ consists of $\|\S\times\A\|$ i.i.d. sequences, one for each $(s,a)\in\S\times\A$. We can join their codes and get a total code length \[sCode\] (s\_[1:n]{}\|a\_[1:n]{}) = \_[s,a]{} (ň\_[s]{}\^[a]{}) If instead of we use the improved sparse code , non-occurring transitions $s\stackrel{a}\to s'r'$ will contribute only one bit rather than $\fr12\log n$ bits to the code, so that large but sparse MDPs get penalized less. Similarly to the states we code the rewards. There are different “standard” reward models. I consider only the simplest case of a small discrete reward set $\R$ like $\{0,1\}$ or $\{-1,0,+1\}$ here and defer generalizations to $\SetR$ and a discussion of variants to the $\Phi$DBN model [@Hutter:09phidbn]. By the MDP assumption, for each $(s,a,s')$ triple, the rewards at times $\T_{ss'}^{a}$ are i.i.d. Hence they can be coded in \[rCode\] (r\_[1:n]{}\|s\_[1:n]{},a\_[1:n]{}) = \_[s,a,s’]{} (ň\_[ss’]{}\^[a]{}) bits. In order to increase the statistics it might be better to treat $r_t$ as a function of $s_t$ only. This is not restrictive, since dependence on $s_{t-1}$ and $a_{t-1}$ can be mimicked by coding aspects into an enlarged state space. Note that the code for $\v r$ depends on $\v s$. Indeed we may interpret the construction as follows: Ultimately we/the agent cares about the reward, so we want to measure how well we can predict the rewards, which we do with . But this code depends on $\v s$, so we need a code for $\v s$ too, which is . To see that we need both parts consider two extremes. A simplistic state transition model (small $\|\S\|$) results in a short code for $\v s$. For instance, for $\|\S\|=1$, nothing needs to be coded and is identically zero. But this obscures potential structure in the reward sequence, leading to a long code for $\v r$. On the other hand, the more detailed the state transition model (large $\|\S\|$) the easier it is to predict and hence compress $\v r$. But a large model is hard to learn, i.e. the code for $\v s$ will be large. For instance for $\Phi(h)=h$, no state repeats and the frequency-based coding breaks down. Let us define the [Cost]{} of $\Phi:\H\to\S$ on $h_n$ as the length of the $\Phi$MDP code for $\v s\v r$ given $\v a$ plus a complexity penalty $\CL(\Phi)$ for $\Phi$: \[costphi\] & & (\|h\_n) := (s\_[1:n]{}\|a\_[1:n]{}) + (r\_[1:n]{}\|s\_[1:n]{},a\_[1:n]{}) + (),\ & & s\_t=(h\_t) h\_t=ora\_[1:t-1]{}o\_t r\_t The discussion above suggests that the minimum of the joint code length and is attained for a $\Phi$ that keeps all and only relevant information for predicting rewards. Such a $\Phi$ may be regarded as best explaining the rewards. I added an additional complexity penalty $\CL(\Phi)$ for $\Phi$ such that from the set of $\Phi$ that minimize + (e.g. $\Phi$’s identical on $(\O\times\R\times\A)^n$ but different on longer histories) the simplest one is selected. The penalty is usually some code-length or log-index of $\Phi$. This conforms with Ockham’s razor and the MDL philosophy. So we are looking for a $\Phi$ of minimal cost: \[bestphi\] \^[best]{} := \_{ (\|h\_n) } If the minimization is restricted to some small class of reasonably simple $\Phi$, $\CL(\Phi)$ in may be dropped. The state sequence generated by $\Phi^{best}$ (or approximations thereof) will usually only be approximately MDP. While $\Cost(\Phi\|h)$ is an optimal code only for MDP sequences, it still yields good codes for approximate MDP sequences. Indeed, $\Phi^{best}$ balances closeness to MDP with simplicity. The primary purpose of the simplicity bias is [not]{} computational tractability, but generalization ability [@Legg:08; @Hutter:04uaibook]. In unsupervised learning (clustering and density estimation) and supervised learning (regression and classification), penalized maximum likelihood criteria [@Hastie:01 Chp.7] like BIC [@Schwarz:78], MDL [@Gruenwald:07book], and MML [@Wallace:05] have successfully been used for semi-parametric model selection. It is far from obvious how to apply them in RL. Indeed, our derived Cost function cannot be interpreted as a usual model+data code length. The problem is the following: Ultimately we do not care about the observations but the rewards. The rewards depend on the states, but the states are arbitrary in the sense that they are model-dependent functions of the bare data (observations). The existence of these unobserved states is what complicates matters, but their introduction is necessary in order to model the rewards. For instance, $\Phi$ is actually not needed for coding $\v r\v s\|\v a$, so from a strict coding/MDL perspective, $\CL(\Phi)$ in is redundant. Since $\v s$ is some “arbitrary” construct of $\Phi$, it is better to regard as a code of $\v r$ only. Since the agent chooses his actions, $\v a$ need not be coded, and $\v o$ is not coded, because they are only of indirect importance. The Cost() criterion is strongly motivated by the rigorous MDL principle, but invoked outside the usual induction/modeling/prediction context. A Tiny Example {#secTE} ============== The purpose of the tiny example in this section is to provide enough insight into how and why $\Phi$MDP works to convince the reader that our $\Phi$ selection principle is reasonable. I assume a simplified MDP model in which reward $r_t$ only depends on $s_t$, i.e.\[rCodeEx\] (r\_[1:n]{}\|s\_[1:n]{},a\_[1:n]{}) = \_[s’]{} (ň\_[s’]{}\^) This allows us to illustrate $\Phi$MDP on a tiny example. The same insight is gained using if an analogous larger example is considered. Furthermore I set $\CL(\Phi)\equiv 0$. Consider binary observation space $\O=\{0,1\}$, quaternary reward space $\R=\{0,1,2,3\}$, and a single action $\A=\{0\}$. Observations $o_t$ are independent fair coin flips, i.e. Bernoulli($\fr12$), and reward $r_t=2o_{t-1}+o_t$ a deterministic function of the two most recent observations. As features $\Phi$ I consider $\Phi_k:\H\to\O^k$ with $\Phi_k(h_t)=o_{t-k+1}...o_t$ for various $k=0,1,2,...$ which regard the last $k$ observations as “relevant”. Intuitively $\Phi_2$ is the best observation summary, which I confirm below. The state space $\S=\{0,1\}^k$ (for sufficiently large $n$). The $\Phi$MDPs for $k=0,1,2$ are as follows. =3ex (27,7)(0,0.3) (2,7)[(0,0)\[ct\][$\Phi_0$MDP]{}]{} (2,3)(2,3)[(0,0)\[cc\][$\epstr$]{}]{} (2,2)[(0,0)\[ct\][$r=0\|1\|2\|3$]{}]{} (2,5)[(1,1)\[t\]]{} (1.5,5)[(0,-1)[1.13]{}]{} (2.5,5)[(0,-1)[1.13]{}]{} (10,7)[(0,0)\[ct\][$\Phi_1$MDP]{}]{} (8,3)(8,3)[(0,0)\[cc\][0]{}]{} (8,2)[(0,0)\[ct\][$r=0\|2$]{}]{} (8,5)[(1,1)\[t\]]{} (7.5,5)[(0,-1)[1.13]{}]{} (8.5,5)[(0,-1)[1.13]{}]{} (10,3)[(1,0)[1]{}]{} (10,3)[(-1,0)[1]{}]{} (12,3)(12,3)[(0,0)\[cc\][1]{}]{} (12,2)[(0,0)\[ct\][$r=1\|3$]{}]{} (12,5)[(1,1)\[t\]]{} (11.5,5)[(0,-1)[1.13]{}]{} (12.5,5)[(0,-1)[1.13]{}]{} (22,7)[(0,0)\[ct\][$\Phi_2$MDP]{}]{} (20,5)(20,5)[(0,0)\[cc\][00]{}]{} (19,4)[(0,0)\[rt\][$r=0$]{}]{} (18,5)[(1,1)\[l\]]{} (18,4.5)[(1,0)[1.13]{}]{} (18,5.5)[(1,0)[1.13]{}]{} (24,1)(24,1)[(0,0)\[cc\][11]{}]{} (25,2)[(0,0)\[lb\][$r=3$]{}]{} (26,1)[(1,1)\[r\]]{} (26,1.5)[(-1,0)[1.13]{}]{} (26,0.5)[(-1,0)[1.13]{}]{} (24,5)(24,5)[(0,0)\[cc\][01]{}]{} (25,5)[(0,0)\[lc\][$\,r=1$]{}]{} (20,1)(20,1)[(0,0)\[cc\][10]{}]{} (19,1)[(0,0)\[rc\][$\,r=2\,$]{}]{} (21,5)[(1,0)[2]{}]{} (24,4)[(0,-1)[2]{}]{} (23,1)[(-1,0)[2]{}]{} (20,2)[(0,1)[2]{}]{} (22,3)[(1,1)[1.3]{}]{} (22,3)[(-1,-1)[1.3]{}]{} with all non-zero transition probabilities being 50% is an exact representation of our data source. The missing arrow (directions) are due to the fact that $s=o_{t-1}o_t$ can only lead to $s'=o'_t o'_{t+1}$ for which $o'_t=o_t$, denoted by $s=s'$ in the following. Note that $\Phi$MDP does not “know” this and has to learn the (non)zero transition probabilities. Each state has two successor states with equal probability, hence generates (see previous paragraph) a Bernoulli($\fr12$) state subsequence and a constant reward sequence, since the reward can be computed from the state = last two observations. Asymptotically, all four states occur equally often, hence the sequences have approximately the same length $n/4$. In general, if $\v s$ (and similarly $\v r$) consists of $x\in\SetN$ i.i.d. subsequences of equal length $n/x$ over $y\in\SetN$ symbols, the code length (and similarly ) is (š\|ǎ;x\_y) &=& ny + x[\|§\|-12]{}\ (ř\|š,ǎ;x\_y) &=& ny + x[\|\|-12]{} where the extra argument $x_y$ just indicates the sequence property. So for $\Phi_2$MDP we get (š\|ǎ;4\_2) = n+6n4 (ř\|š,ǎ;4\_1) = 6n4 The log-terms reflect the required memory to code the MDP structure and probabilities. Since each state has only 2 realized/possible successors, we need $n$ bits to code the state sequence. The reward is a deterministic function of the state, hence needs no memory to code given $\v s$. throws away all observations (left figure above), hence $\CL(\v s\|\v a;1_1)=0$. While the reward sequence is [not]{} i.i.d.(e.g. $r_{t+1}=3$ cannot follow $r_t=0$), $\Phi_0$MDP has no choice regarding them as i.i.d., resulting in $\CL(\v s\|\v a;1_4)=2n+\fr32\log n$. model is an interesting compromise (middle figure above). The state allows a partial prediction of the reward: State 0 allows rewards 0 and 2; state 1 allows rewards 1 and 3. Each of the two states creates a Bernoulli($\fr12$) state successor subsequence and a binary reward sequence, wrongly presumed to be Bernoulli($\fr12$). Hence $\CL(\v s\|\v a;2_2)=n+\log\fr n2$ and $\CL(\v r\|\v s,\v a;2_2)=n+3\log\fr n2$. The following table summarizes the results for general $k=0,1,2$ and beyond: [c\|\|c\|c\|c\|c\|c\|c\|c\|c\|c]{} k & §& \|§\| & n\_[ss’]{}\^[0]{} & n\_[s’]{}\^[r’]{} & n\_[s]{}\^[0]{}=n\_[s’]{}\^ & š+ř & (š\|ǎ) & (ř\|š,ǎ) & (\|h)\ 0 & {} & 1 & n & n/4 & n & 1\_1+1\_4 & 0 & 2n+32n & 2n+32n\ 1 & {0,1} & 2 & n/4 & n4\_[r’-s’=0\|1]{} & n/2 & 2\_2+2\_2 & n+n2 & n+3n2 & 2n+4n2\ 2 & {[00,0110,11]{}} & 4 & n8\_[s\,\s’]{} & n4\_[r’=s’]{} & n/4 & 4\_2+4\_1 & n+6n4 & 6n4 & n+12n4\ \^[\_]{}2 & {0,1}\^k & 2\^k & & n4\_[r’=s’]{} & n/2\^k & 2\^k\_2+2\^k\_2 & n+[2\^k-12\^[1-k]{}]{} & 32 2\^k & n+[2\^k+22\^[1-k]{}]{} The notation of the $\v s+\v r$ column follows the one used above in the text ($x_y$ for $\v s$ and $\v r$). $r'\widehat=s'$ means that $r'$ is the correct reward for state $s'$. The last column is the sum of the two preceding columns. The part linear in $n$ is the code length for the state/reward sequence. The part logarithmic in $n$ is the code length for the transition/reward probabilities of the MDP; each parameter needs $\fr12\log n$ bits. For large $n$, $\Phi_2$ results in the shortest code, as anticipated. The “approximate” model $\Phi_1$ is just not good enough to beat the vacuous model $\Phi_0$, but in more realistic examples some approximate model usually has the shortest code. In [@Hutter:09phidbn] I show on a more complex example how $\Phi^{best}$ will store long-term information in a POMDP environment. ($\mathbf\Phi$) Minimization {#secCM} ============================ So far I have reduced the reinforcement learning problem to a formal $\Phi$-optimization problem. This section briefly explains what we have gained by this reduction, and provide some general information about problem representations, stochastic search, and $\Phi$ neighborhoods. Finally I present a simplistic but concrete algorithm for searching context tree MDPs. I now discuss how to find good summaries $\Phi$. The introduced generic cost function $\Cost(\Phi\|h_n)$, based on only the known history $h_n$, makes this a well-defined task that is completely decoupled from the complex (ill-defined) reinforcement learning objective. This reduction should not be under-estimated. We can employ a wide range of optimizers and do not even have to worry about overfitting. The most challenging task is to come up with creative algorithms proposing $\Phi$’s. There are many optimization methods: Most of them are search-based: random, blind, informed, adaptive, local, global, population based, exhaustive, heuristic, and other search methods [@Aarts:97]. Most are or can be adapted to the structure of the objective function, here $\Cost(\cdot\|h_n)$. Some exploit the structure more directly (e.g. gradient methods for convex functions). Only in very simple cases can the minimum be found analytically (without search). Most search algorithms require the specification of a neighborhood relation or distance between candidate $\Phi$, which I define in the 2nd next paragraph. can be important: Since $\Phi$ is a discrete function, searching through (a large subset of) all computable functions, is a non-restrictive approach. Variants of Levin search [@Schmidhuber:04oops; @Hutter:04uaibook] and genetic programming [@Koza:92; @Banzhaff:98] and recurrent neural networks [@Pearlmutter:89; @Raedt:08] are the major approaches in this direction. A different representation is as follows: $\Phi$ effectively partitions the history space $\H$ and identifies each partition with a state. Conversely any partition of $\H$ can (up to a renaming of states) uniquely be characterized by a function $\Phi$. Formally, $\Phi$ induces a (finite) partition $\bigcup_s\{h':\Phi(h')=s\}$ of $\H$, where $s$ ranges over the codomain of $\Phi$. Conversely, any partition of $\H={\cal B}_1\dot\cup...\dot\cup{\cal B}_m$ induces a function $\Psi(h')=i$ iff $h'\in{\cal B}_i$, which is equivalent to $\Phi$ apart from an irrelevant permutation of the codomain (renaming of states). State aggregation methods have been suggested earlier for solving large-scale MDP planning problems by grouping (partitioning) similar states together, resulting in (much) smaller block MDPs [@Givan:03]. But the used bi-simulation metrics require knowledge of the MDP transition probabilities, while our Cost criterion does not. Decision trees/lists/grids/etc. are essentially space partitioners. The most powerful versions are rule-based, in which logical expressions recursively divide domain $\H$ into “true/false” regions [@Dzeroski:01; @Sanner:09]. A natural “minimal” change of a partition is to subdivide=split a partition or merge (two) partitions. Moving elements from one partition to another can be implemented as a split and merge operation. In our case this corresponds to splitting and merging states (state refinement and coarsening). Let $\Phi'$ split some state $s^a\in\S$ of $\Phi$ into $s^b,s^c\not\in\S$ ’(h) := { [(h) (h)s\^a s\^b s\^c (h)=s\^a]{} . where the histories mapped to state $s^a$ are distributed among $s^b$ and $s^c$ according to some splitting rule (e.g. randomly). The new state space is $\S'=\S\setminus\{s^a\}\cup\{s^b,s^c\}$. Similarly $\Phi'$ merges states $s^b,s^c\in\S$ into $s^a\not\in\S$ if ’(h) := { [(h) (h)s\^a s\^a (h)=s\^b s\^c]{} . where $\S'=\S\setminus\{s^b,s^c\}\cup\{s^s\}$. We can regard $\Phi'$ as being a neighbor of or similar to $\Phi$. Stochastic search is the method of choice for high-dimensional unstructured problems. Monte Carlo methods can actually be highly effective, despite their simplicity [@Liu:02; @Fishman:03]. The general idea is to randomly choose a neighbor $\Phi'$ of $\Phi$ and replace $\Phi$ by $\Phi'$ if it is better, i.e. has smaller Cost. Even if $\Cost(\Phi'\|h)>\Cost(\Phi\|h)$ we may keep $\Phi'$, but only with some (in the cost difference exponentially) small probability. Simulated annealing is a version which minimizes $\Cost(\Phi\|h)$. Apparently, $\Phi$ of small cost are (much) more likely to occur than high cost $\Phi$. The $\Phi_k$ in Section \[secTE\] depended on the last $k$ observations. Let us generalize this to a context dependent variable length: Consider a finite complete suffix free set of strings (= prefix tree of reversed strings) $\S\subset\O^$ as our state space (e.g.$\S=\{0,01,011,111\}$ for binary $\O$), and define $\Phi_\S(h_n):=s$ iff $o_{n-\|s\|+1:n}=s\in\S$, i.e. $s$ is the part of the history regarded as relevant. State splitting and merging works as follows: For binary $\O$, if history part $s\in\S$ of $h_n$ is deemed too short, we replace $s$ by $0s$ and $1s$ in $\S$, i.e. $\S'=\S\setminus\{s\}\cup\{0s,1s\}$. If histories $1s,0s\in\S$ are deemed too long, we replace them by $s$, i.e. $\S'=\S\setminus\{0s,1s\}\cup\{s\}$. Large $\O$ might be coded binary and then treated similarly. For small $\O$ we have the following simple $\Phi$-optimizer: [=0ex=0ex=0ex=0ex]{} [$\Phi$Improve($\Phi_\S,h_n$)]{} [=0ex=0ex=2ex=1ex=1ex]{} Randomly choose a state $s\in\S$; Let $p$ and $q$ be uniform random numbers in $[0,1]$; if $(p>1/2)$ then split $s$ i.e. $S'=S\setminus\{s\}\cup\{os:o\in\O\}$ else if $\{os':o\in\O\}\subseteq\S\quad$ ([$s'$ is $s$ without the first symbol]{}) then merge them, i.e. $S'=S\setminus\{os':o\in\O\}\cup\{s'\}$; if $(\Cost(\Phi_\S\|h_n)-\Cost(\Phi_{\S'}\|h_n) > \log(q))$ then $\S:=\S'$; [return ($\Phi_\S$); ]{} =1.3ex (12,6)(0,-10) (6,14)[(0,0)\[ct\][Example tree]{}]{} (2,11.3)[(0,0)\[cc\][$o_{n-2}$]{}]{} (6,11.3)[(0,0)\[cc\][$o_{n-1}$]{}]{} (10,11.3)[(0,0)\[cc\][$o_n$]{}]{} (10,1.5)[(0,0)\[rb\][$\S=$]{}]{} (12,-0.5)[(0,0)\[rb\][$\{0,01,011,111\}$]{}]{} (0,0)(4,2)[3]{}[ (4,4)[(-2,1)[4]{}]{} (4,4)[(-2,-1)[4]{}]{} (2,5)[(0,0)\[lb\][0]{}]{} (2,3)[(0,0)\[lt\][1]{}]{} (4,4) (0,6) (0,2) ]{} The idea of using suffix trees as state space is from [@McCallum:96] (see also [@Ring:94]). It might be interesting to compare the local split/merge criterion of [@McCallum:96] with our general global Cost criterion. On the other hand, due to their limitation, suffix trees are currently out of vogue. Exploration & Exploitation {#secEE} ========================== Having obtained a good estimate $\hat\Phi$ of $\Phi^{best}$ in the previous section, we can/must now determine a good action for our agent. For a finite MDP with known transition probabilities, finding the optimal action is routine. For estimated probabilities we run into the infamous exploration-exploitation problem, for which promising approximate solutions have recently been suggested [@Szita:08]. At the end of this section I present the overall algorithm for our $\Phi$MDP agent. For a known finite MDP $(\S,\A,T,R,\gamma)$, the maximal achievable (“optimal”) expected future discounted reward sum, called ($Q$) $V\!$alue (of action $a$) in state $s$, satisfies the following (Bellman) equations [@Sutton:98] \[BellmanEq\] Q\_s\^[\a]{} = \_[s’]{} T\_[ss’]{}\^a\[R\_[ss’]{}\^a + V\_[s’]{}\^\\] V\_s\^\=\_a Q\_s\^[\a]{} where $0<\g<1$ is a discount parameter, typically close to 1. See [@Hutter:04uaibook Sec.5.7] for proper choices. The equations can be solved by a simple (e.g. value or policy) iteration process or various other methods or in guaranteed polynomial time by dynamic programming [@Puterman:94]. The optimal next action is \[BellmanSol\] a\_n := \_a Q\_[s\_n]{}\^[\a]{} We can estimate the transition probability $T$ by \[hatT\] T\_[ss’]{}\^a := n\_[s]{}\^[a]{}>0 0 It is easy to see that the Shannon-Fano code of $s_{1:n}$ based on $\P_{\smash{\!\hat T}}(s_{1:n}\|a_{1:n})=\prod_{t=1}^n\hat T_{s_{t-1}s_t}^{a_{t-1}}$ plus the code of the (non-zero) transition probabilities $\hat T_{ss'}^a$ to relevant accuracy $O(1/\sqrt{n_{s\p}^{a\p}})$ has length , i.e.the frequency estimate is consistent with the attributed code length. The expected reward can be estimated as \[hatR\] R\_[ss’]{}\^a := \_[r’]{}R\_[ss’]{}\^[ar’]{} r’, R\_[ss’]{}\^[ar’]{} := [n\_[ss’]{}\^[ar’]{}n\_[ss’]{}\^[a]{}]{} Simply replacing $T$ and $R$ in and by their estimates and can lead to very poor behavior, since parts of the state space may never be explored, causing the estimates to stay poor. Estimate $\hat T$ improves with increasing $n_{s\p}^{a\p}$, which can (only) be ensured by trying all actions $a$ in all states $s$ sufficiently often. But the greedy policy above has no incentive to explore, which may cause the agent to perform very poorly: The agent stays with what he [believes]{} to be optimal without trying to solidify his belief. For instance, if treatment $A$ cured the first patient, and treatment $B$ killed the second, the greedy agent will stick to treatment $A$ and not explore the possibility that $B$ may just have failed due to bad luck. Trading off exploration versus exploitation optimally is computationally intractable [@Hutter:04uaibook; @Poupart:06; @Ross:08bayes] in all but extremely simple cases (e.g. Bandits [@Berry:85; @Kumar:86]). Recently, polynomially optimal algorithms (Rmax,E3,OIM) have been invented [@Kearns:98; @Brafman:02; @Szita:08]: An agent is more explorative if he expects a high reward in the unexplored regions. We can “deceive” the agent to believe this by adding another “absorbing” high-reward state $s^e$ to $\S$, not in the range of $\Phi(h)$, i.e. never observed. Henceforth, $\S$ denotes the extended state space. For instance $+$ in now includes $s^e$. We set \[extnR\] n\_[ss\^e]{}\^a=1,n\_[s\^e s]{}\^a=\_[s\^e s]{},R\_[ss\^e]{}\^a=R\_[max]{}\^e for all $s,a$, where exploration bonus $R_{max}^e$ is polynomially (in $(1-\g)^{-1}$ and $\|\S\times\A\|$) larger than $\max\R$ [@Szita:08]. Now compute the agent’s action by - but for the extended $\S$. The optimal policy $p^$ tries to find a chain of actions and states that likely leads to the high reward absorbing state $s^e$. Transition $\hat T_{ss^e}^a=1/n_{s\p}^a$ is only “large” for small $n_{s\p}^a$, hence $p^$ has a bias towards unexplored (state,action) regions. It can be shown that this algorithm makes only a polynomial number of sub-optimal actions. The overall algorithm for our $\Phi$MDP agent is as follows. [=0ex=0ex=0ex=0ex]{} [$\Phi$MDP-Agent($\A,\R$)]{} [=0ex=0ex=2ex=1ex=1ex]{} Initialize $\Phi\equiv\Phi'\equiv\epstr$; $\;\S=\{\epstr\}$; $\;h_0=a_0=r_0=\epstr$; for $n=1,2,3,...$ [=0ex=0ex=2ex=1ex=1ex]{} Choose e.g. $\g=1-1/n$; Set $R_{max}^e=$Polynomial$((1-\g)^{-1},\|\S\times\A\|)\cdot\max\R$; While waiting for $o_n$ and $r_n$ [=0ex=0ex=2ex=1ex=1ex]{} $\Phi':=\Phi$Improve($\Phi',h_{n-1}$); If $\Cost(\Phi'\|h_{n-1})<\Cost(\Phi\|h_{n-1})$ then $\Phi:=\Phi'$; Observe $o_n$ and $r_n$; $\;h_n:=h_{n-1}a_{n-1}r_{n-1}o_n r_n$; $s_n:=\Phi(h_n)$; $\;\S:=\S\cup\{s_n\}$; Compute action $a_n$ from Equations -; $\lfloor$ Output action $a_n$; Improved Cost Function {#secICF} ====================== As discussed, we ultimately only care about (modeling) the rewards, but this endeavor required introducing and coding states. The resulting Cost($\Phi\|h$) function is a code length of not only the rewards but also the “spurious” states. This likely leads to a too strong penalty of models $\Phi$ with large state spaces $\S$. The proper Bayesian formulation developed in this section allows to “integrate” out the states. This leads to a code for the rewards only, which better trades off accuracy of the reward model and state space size. For an MDP with transition and reward probabilities $T_{ss'}^a$ and $R_{ss'}^{ar'}$, the probabilities of the state and reward sequences are ¶\_[T]{}(s\_[1:n]{}\|a\_[1:n]{}) = \_[t=1]{}\^n T\_[s\_[t-1]{}s\_t]{}\^[a\_[t-1]{}]{}, ¶\_[R]{}(r\_[1:n]{}\|s\_[1:n]{}a\_[1:n]{}) = \_[t=1]{}\^n R\_[s\_[t-1]{}s\_t]{}\^[a\_[t-1]{}r\_t]{} The probability of $\v r\|\v a$ can be obtained by taking the product and marginalizing $\v s$: ¶\_[U]{}(r\_[1:n]{}\|a\_[1:n]{}) &=& \_[s\_[1:n]{}]{}¶\_[T]{}(s\_[1:n]{}\|a\_[1:n]{}) ¶\_[R]{}(r\_[1:n]{}\|s\_[1:n]{}a\_[1:n]{})\ &=& \_[s\_[1:n]{}]{}\_[t=1]{}\^n U\_[s\_[t-1]{}s\_t]{}\^[a\_[t-1]{}r\_t]{} = \_[s\_n]{}\[U\^[a\_0 r\_1]{}U\^[a\_[n-1]{}r\_n]{}\]\_[s\_0s\_n]{} where for each $a\in\A$ and $r'\in\R$, matrix $U^{ar'}\in\SetR^{m\times m}$ is defined as $[U^{ar'}]_{ss'}\equiv U_{ss'}^{ar'}:=T_{ss'}^a R_{ss'}^{ar'}$. The right $n$-fold matrix product can be evaluated in time $O(m^2 n)$. This shows that $\v r$ given $\v a$ and $U$ can be coded in $-\log\P_{\!U}$ bits. The unknown $U$ needs to be estimated, e.g. by the relative frequency $\hat U_{ss'}^{ar'}:=n_{ss'}^{ar'}/n_{s\p}^{a\p}$. Note that $\P_{\!U}$ completely ignores the observations $o_{1:n}$ and is essentially independent of $\Phi$. Map $\Phi$ and hence $o_{1:n}$ enter $\P_{\!\hat U}$ (only and crucially) via the estimate $\hat U$. The $M:=m(m-1)\|\A\|(\|\R\|-1)$ (independent) elements of $\hat U$ can be coded to sufficient accuracy in $\fr12 M\log n$ bits, and $\Phi$ will be coded in $\CL(\Phi)$ bits. Together this leads to a code for $\v r\|\v a$ of length \[ICost\] (\|h\_n) := - ¶\_[U]{}(r\_[1:n]{}\|a\_[1:n]{}) + 12 Mn + () In practice, $M$ can and should be chosen smaller like done in the original function, and/or by using the restrictive model for $R$, and/or by considering only non-zero frequencies . Analogous to we seek a $\Phi$ that minimizes ICost(). Since action evaluation is based on (discounted) reward sums, not individual rewards, one may think of marginalizing $\P_{\!U}(\v r\|\v a,\Phi)$ even further, or coding rewards only approximately. Unfortunately, the algorithms in Section \[secEE\] that learn, explore, and exploit MDPs require knowledge of the (exact) individual rewards, so this improvement is not feasible. Discussion {#secDisc} ========== This section summarizes $\Phi$MDP, relates it to previous work, and hints at more efficient incremental implementations and more realistic [structured]{} MDPs (dynamic Bayesian networks). Learning from rewards in general environments is an immensely complex problem. In this paper I have developed a generic reinforcement learning algorithm based on sound principles. The key idea was to reduce general learning problems to finite state MDPs for which efficient learning, exploration, and exploitation algorithms exist. For this purpose I have developed a formal criterion for evaluating and selecting good “feature” maps $\Phi$ from histories to states. One crucial property of $\Phi$MDP is that it neither requires nor learns a model of the complete observation space, but only for the reward-relevant observations as summarized in the states. The developed criterion has been inspired by MDL, which recommends to select the (coding) model that minimizes the length of a suitable code for the data at hand plus the complexity of the model itself. The novel and tricky part in $\Phi$MDP was to deal with the states, since they are not bare observations, but model-dependent processed data. An improved Bayesian criterion, which integrates out the states, has also been derived. Finally, I presented a complete feature reinforcement learning algorithm $\Phi$MDP-Agent(). The building blocks and computational flow are depicted in the following diagram: =2.7ex (20,14)(0,0) (0,0)[(20,2)\[cc\][Environment]{}]{} (3,5)[(6,2)(0,0)\[cc\][History $h$]{}]{} (3,9)[(6,2)(0,0)\[cc\][Feature Vec. $\v{\hat\Phi}$]{}]{} (5,13)[(6,2)(0,0)\[cb\][Transition Pr. $\hat T$]{} (0,0)\[ct\][Reward est. $\hat R\quad$]{}]{} (15,13)[(6,2)(0,0)\[cc\][$\hat T^e$, $\hat R^e$]{}]{} (17,9)[(6,2)(0,0)\[cc\][($\hat Q$) $\hat V\!$alue]{}]{} (17,5)[(6,2)(0,0)\[cc\][Best Policy $\hat p$]{}]{} (3,2)[(0,1)[2]{}(0,2)\[rc\][reward $r\;$]{}(0,2)\[lc\][$\;$observation $o$]{}]{} (3,6)[(0,1)[2]{}(0,2)\[lc\][$\;\;\Cost(\Phi\|h)$ minimization]{}]{} (3,10)[(1,1)[2]{}(2,2)\[lc\][$\!\!$frequency estimate]{}]{} (8,13)[(1,0)[4]{}(-4,0)\[cb\][exploration]{}(-4,0)\[ct\][bonus]{}]{} (15,12)[(1,-1)[2]{}(-1.2,-2)\[rc\][Bellman]{}]{} (17,8)[(0,-1)[2]{}(0,-2)\[rc\][implicit$\;$]{}]{} (17,4)[(0,-1)[2]{}(0,-2)\[rc\][action $a\;$]{}]{} As already indicated here and there, $\Phi$MDP can be regarded as extending the frontier of many previous important approaches to RL and beyond: [Partially Observable MDPs (POMDPs)]{} are a very important generalization of MDPs [@Kaelbling:98]. Nature is still assumed to be an MDP, but the states of nature are only partially observed via some non-injective or probabilistic function. Even for finite state space and known observation and transition functions, finding and even only approximating the optimal action is (harder than NP) hard [@Lusena:01; @Madani:03]. Lifting any of the assumptions causes conceptual problems, and when lifting more than one we enter scientific terra nullius. Assume a POMDP environment: POMDPs can formally (but not yet practically) be reduced to MDPs over so-called (continuous) belief states. Since $\Phi$MDP reduces every problem to an MDP, it is conceivable that it reduces the POMDP to (an approximation of) its belief MDP. This would be a profound relation between $\Phi$MDP and POMDP, likely leading to valuable insights into $\Phi$MDP and proper algorithms for learning POMDPs. It may also help us to restrict the space of potentially interesting features $\Phi$. [Predictive State Representations (PSRs)]{} are very interesting, but to this date in an even less developed stage [@Singh:03] than POMDPs. [Universal AI]{} [@Hutter:04uaibook] is able to optimally deal with arbitrary environments, but the resulting AIXI agent is computationally intractable [@Hutter:07aixigentle] and hard to approximate [@Pankov:08; @Hutter:06aixifoe]. [Bayesian RL]{} algorithms [@Dearden:99; @Duff:02; @Poupart:06; @Ross:08bayes] (see also [@Kumar:86 Chp.11]) can be regarded as implementations of the AI$\xi$ models [@Hutter:06aixifoe], which are down-scaled versions of AIXI, but the enormous computational demand still severely limits this approach. $\Phi$MDP essentially differs from “generative” Bayesian RL and AI$\xi$ in that it neither requires to specify nor to learn a (complete stochastic) observation model. It is a more “discriminative” approach [@Liang:08]. Since $\Phi$MDP “automatically” models only the relevant aspects of the environment, it should be computationally less demanding than full Bayesian RL. [State aggregation]{} methods have been suggested earlier for solving large-scale MDP planning problems by grouping (partitioning) similar states together, resulting in (much) smaller block MDPs [@Givan:03]. But the bi-simulation metrics used require knowledge of the MDP transition probabilities. $\Phi$MDP might be regarded as an approach that lifts this assumption. [Suffix trees]{} [@McCallum:96] are a simple class of features $\Phi$. $\Phi$MDP combined with a local search function that expands and deletes leaf nodes is closely related to McCallum’s U-Tree algorithm [@McCallum:96], with a related but likely different split&merge criterion. [Miscellaneous]{}: $\Phi$MDP also extends the theory of model selection (e.g. MDL [@Gruenwald:07book]) from passive to active learning. As discussed in Section \[secCM\], most search algorithms are local in the sense that they produce a chain of “slightly” modified candidate solutions, here $\Phi$. This suggests a potential speedup by computing quantities of interest incrementally, which becomes even more important in the $\Phi$DBN case [@Hutter:09phidbn; @Hutter:09phidbnx]. Computing $\Cost(\Phi)$ takes at most time $O(\|\S\|^2\|\A\|\|\R\|)$. If we split or merge two states, we can incrementally update the cost in time $O(\|\S\|\|\A\|\|\R\|)$, rather than computing it again from scratch. In practice, many transition $T_{ss'}^a$ don’t occur, and $\Cost(\Phi)$ can actually be computed much faster in time $O(\|\{n_{ss'}^{ar}>0\}\|)$, and incrementally even faster. Iteration algorithms for need an initial value for $V$ or $Q$. We can take the estimate $\hat V$ from a previous $\Phi$ as an initial value for the new $\Phi$. For a merge operation we can average the value of both states, for a split operation we could give them the same initial value. A significant further speedup can be obtained by using prioritized iteration algorithms that concentrate their time on badly estimated states, which are in our case (states close to) the new ones [@Sutton:98]. Similarly, results from cycle $n$ can be (re)used for the next cycle $n+1$. For instance, $\hat V$ can simply be reused as an initial value in the Bellman equations, and ICost$(\Phi)$ can be updated in time $O(\|\S\|^2)$ or even faster if $U$ is sparse. The use of “unstructured” MDPs, even our $\Phi$-optimal ones, is clearly limited to very simple tasks. Real world problems are structured and can often be represented by dynamic Bayesian networks (DBNs) with a reasonable number of nodes. Our $\Phi$ selection principle can be adapted from MDPs to the conceptually much more complex DBN case. The primary purpose of this Part I was to explain the key concepts on an as simple model as possible, namely unstructured finite MDPs, to set the stage for developing the more realistic $\Phi$DBN in Part II [@Hutter:09phidbnx]. The major open problems are to develop smart $\Phi$ generation and smart stochastic search algorithms for $\Phi^{best}$, and to determine whether minimizing is the right criterion. My thanks go to Pedro Ortega, Sergey Pankov, Scott Sanner, Jürgen Schmidhuber, and Hanna Suominen for feedback on earlier drafts. [ABCD]{}=0ex E. H. L. Aarts and J. K. Lenstra, editors. . Discrete Mathematics and Optimization. Wiley-Interscience, Chichester, England, 1997. A. R. Barron. . PhD thesis, Stanford University, 1985. D. A. Berry and B. Fristedt. . Chapman and Hall, London, 1985. W. Banzhaff, P. Nordin, E. Keller, and F.D. Francone. . Morgan-Kaufmann, San Francisco, CA, U.S.A., 1998. R. I. Brafman and M. Tennenholtz. R-max – a general polynomial time algorithm for near-optimal reinforcement learning. , 3:213–231, 2002. T. M. Cover and J. A. Thomas. . Wiley-Intersience, 2nd edition, 2006. S. Dzeroski, L. de Raedt, and K. Driessens. Relational reinforcement learning. , 43:7–52, 2001. R. Dearden, N. Friedman, and D. Andre. Model based [B]{}ayesian exploration. In [Proc. 15th Conference on Uncertainty in Artificial Intelligence ([UAI]{}-99)]{}, pages 150–159, 1999. M. Duff. . PhD thesis, Department of Computer Science, University of Massachusetts Amherst, 2002. G. Fishman. . Springer, 2003. R. Givan, T. Dean, and M. Greig. Equivalence notions and model minimization in [M]{}arkov decision processes. , 147(1–2):163–223, 2003. I. Guyon and A. Elisseeff, editors. , JMLR Special Issue, 2003. MIT Press. G. Gordon. . PhD thesis, School of Computer Science, Carnegie Mellon University, Pittsburgh, PA, 1999. B. Goertzel and C. Pennachin, editors. . Springer, 2007. P. D. Gr[ü]{}nwald. . The MIT Press, Cambridge, 2007. T. Hastie, R. Tibshirani, and J. H. Friedman. . Springer, 2001. M. Hutter. . Springer, Berlin, 2005. 300 pages, http://www.hutter1.net/ai/uaibook.htm. M. Hutter. Universal algorithmic intelligence: A mathematical top$\rightarrow$down approach. In [Artificial General Intelligence]{}, pages 227–290. Springer, Berlin, 2007. M. Hutter. Feature dynamic [B]{}ayesian networks. In [Proc. 2nd Conf. on Artificial General Intelligence ([AGI’09]{})]{}, volume 8, pages 67–73. Atlantis Press, 2009. M. Hutter. Feature [M]{}arkov decision processes. In [Proc. 2nd Conf. on Artificial General Intelligence ([AGI’09]{})]{}, volume 8, pages 61–66. Atlantis Press, 2009. M. Hutter. Feature reinforcement learning: Part [II]{}: Structured [MDP]{}s. , 2009. L. P. Kaelbling, M. L. Littman, and A. R. Cassandra. Planning and acting in partially observable stochastic domains. , 101:99–134, 1998. J. R. Koza. . The MIT Press, 1992. M. J. Kearns and S. Singh. Near-optimal reinforcement learning in polynomial time. In [Proc. 15th International Conf. on Machine Learning]{}, pages 260–268. Morgan Kaufmann, San Francisco, CA, 1998. P. R. Kumar and P. P. Varaiya. . Prentice Hall, Englewood Cliffs, NJ, 1986. S. Legg. . PhD thesis, IDSIA, Lugano, 2008. C. Lusena, J. Goldsmith, and M. Mundhenk. Nonapproximability results for partially observable [M]{}arkov decision processes. , 14:83–103, 2001. S. Legg and M. Hutter. Universal intelligence: A definition of machine intelligence. , 17(4):391–444, 2007. J. S. Liu. . Springer, 2002. P. Liang and M. Jordan. An asymptotic analysis of generative, discriminative, and pseudolikelihood estimators. In [Proc. 25th International Conf. on Machine Learning ([ICML’08]{})]{}, volume 307, pages 584–591. ACM, 2008. M. Li and P. M. B. Vitányi. . Springer, Berlin, 3rd edition, 2008. D. J. C. MacKay. . Cambridge University Press, Cambridge, MA, 2003. A. K. McCallum. . PhD thesis, Department of Computer Science, University of Rochester, 1996. O. Madani, S. Hanks, and A. Condon. On the undecidability of probabilistic planning and related stochastic optimization problems. , 147:5–34, 2003. A. Y. Ng, A. Coates, M. Diel, V. Ganapathi, J. Schulte, B. Tse, E. Berger, and E. Liang. Autonomous inverted helicopter flight via reinforcement learning. In [ISER]{}, volume 21 of [Springer Tracts in Advanced Robotics]{}, pages 363–372. Springer, 2004. S. Pankov. A computational approximation to the [AIXI]{} model. In [Proc. 1st Conference on Artificial General Intelligence]{}, volume 171, pages 256–267, 2008. B. A. Pearlmutter. Learning state space trajectories in recurrent neural networks. , 1(2):263–269, 1989. J. Poland and M. Hutter. Universal learning of repeated matrix games. In [Proc. 15th Annual Machine Learning Conf. of [B]{}elgium and [T]{}he [N]{}etherlands ([Benelearn’06]{})]{}, pages 7–14, Ghent, 2006. M. L. Puterman. . Wiley, New York, NY, 1994. P. Poupart, N. A. Vlassis, J. Hoey, and K. Regan. An analytic solution to discrete [B]{}ayesian reinforcement learning. In [Proc. 23rd International Conf. on Machine Learning ([ICML’06]{})]{}, volume 148, pages 697–704, Pittsburgh, PA, 2006. ACM. L. De Raedt, B. Hammer, P. Hitzler, and W. Maass, editors. , volume 08041 of [Dagstuhl Seminar Proceedings]{}. IBFI, Schloss Dagstuhl, Germany, 2008. M. Ring. . PhD thesis, University of Texas, Austin, 1994. S. J. Russell and P. Norvig. . Prentice-Hall, Englewood Cliffs, NJ, 2nd edition, 2003. S. Ross and J. Pineau. Model-based [B]{}ayesian reinforcement learning in large structured domains. In [Proc. 24th Conference in Uncertainty in Artificial Intelligence ([UAI’08]{})]{}, pages 476–483, Helsinki, 2008. AUAI Press. S. Ross, J. Pineau, S. Paquet, and B. Chaib-draa. Online planning algorithms for [POMDP]{}s. , 2008(32):663–704, 2008. R. S. Sutton and A. G. Barto. . MIT Press, Cambridge, MA, 1998. S. Sanner and C. Boutilier. Practical solution techniques for first-order [MDP]{}s. , 173(5–6):748–788, 2009. G. Schwarz. Estimating the dimension of a model. , 6(2):461–464, 1978. J. Schmidhuber. Optimal ordered problem solver. , 54(3):211–254, 2004. A. L. Strehl, C. Diuk, and M. L. Littman. Efficient structure learning in factored-state [MDP]{}s. In [Proc. 27th AAAI Conference on Artificial Intelligence]{}, pages 645–650, Vancouver, BC, 2007. AAAI Press. I. Szita and A. L[ö]{}rincz. The many faces of optimism: a unifying approach. In [Proc. 12th International Conference ([ICML]{} 2008)]{}, volume 307, Helsinki, Finland, June 2008. S. Singh, M. Littman, N. Jong, D. Pardoe, and P. Stone. Learning predictive state representations. In [Proc. 20th International Conference on Machine Learning ([ICML’03]{})]{}, pages 712–719, 2003. C. S. Wallace. . Springer, Berlin, 2005. D. H. Wolpert and W. G. Macready. No free lunch theorems for optimization. , 1(1):67–82, 1997. F. M. J. Willems, Y. M. Shtarkov, and T. J. Tjalkens. Reflections on the prize paper: The context-tree weighting method: Basic properties. , pages 20–27, 1997. List of Notation {#secApp} ================ = =\ $\O$ = finite or infinite set of possible observations\ $\A$ = (small) finite set of actions\ $\R$ = $\{0,1\}$ or $[0,R_{max}]$ or other set of rewards\ $n\in\SetN$ = current time\ $o_t r_t a_t$ = $ora_t\in\O\times\R\times\A$ = true observation, reward, action at time $t$\ [Internal structures for $\mathbf\Phi$MDP]{}\ $\log$ = binary logarithm\ $t\in\{1,...,n\}$ = any time\ $i\in\{1,...,m\}$ = any state index\ $\v x=x_{1:n}$ = $x_1...x_n$ (any $x$)\ $x_\p,x_,\v x_\vi$ = $\sum_j x_j$, $\bigcup_j x_j$, $(x_1,...,x_l)$ (any $x,j,l$)\ $\hat X$ = estimate of $X$ (any $X$)\ $\H$ = $(\O\times\R\times\A)^\times\O\times\R$ = possible histories\ $h_n$ = $ora_{1:n-1}o_n r_n$ = actual history at time $n$\ $\S$ = $\{s^1,...,s^m\}$ = internal finite state space (can vary with $n$)\ $\Phi:\H\to\S$ = state or feature summary of history\ $s_t$ = $\Phi(h_t)\in\S$ = realized state at time $t$\ $\P(\cdot)$ = probability over states and rewards or parts thereof\ $\CL(\cdot)$ = code length\ MDP = $(\S,\A,T,R)$ = Markov Decision Process\ $T_{ss'}^a$ = $\P(s_t=s'\|s_{t-1}=s,a_{t-1}=a)$ = transition matrix\ $s\stackrel{a}\to s'(r')$ = action $a$ in state $s$ resulted in state $s'$ (and reward $r'$)\ $\T_{ss'}^{ar'}$ = set of times $t\in\{1,...,n\}$ at which $s\stackrel{a}\to s'r'$\ $n_{ss'}^{ar'}$ = $\|\T_{ss'}^{ar'}\|$ = number of times $t\in\{1,...,n\}$ at which $s\stackrel{a}\to s'r'$\ $\Cost(\Phi\|h)$ = cost (evaluation function) of $\Phi$ based on history $h$\ ICost$(\Phi\|h)$ = improved cost function\ $Q_s^{a},V_s^*$ = optimal ($Q$) $V\!$alue (of action $a$) in state $s$\ $\g\in[0;1)$ = discount factor ($(1-\g)^{-1}$ is effective horizon)\ [^1]: A shorter version appeared in the proceedings of the AGI 2009 conference [@Hutter:09phimdp].	{ "pile_set_name": "ArXiv" }
--- abstract: 'We compute the non-linear memory contributions to the gravitational wave amplitudes for compact binaries in eccentric orbits at the third post-Newtonian order in general relativity. These contributions are hereditary in nature as they are sourced by gravitational waves emitted during the binary’s entire dynamical past. Combining these with already available instantaneous and tail contributions we get the complete 3PN accurate gravitational waveform.' author: - Michael Ebersold - Yannick Boetzel - Guillaume Faye - Chandra Kant Mishra - 'Bala R. Iyer' - Philippe Jetzer bibliography: - 'references-mem.bib' title: 'Gravitational wave amplitudes for compact binaries in eccentric orbits at the third post-Newtonian order: Memory contributions' --- Introduction {#sec: introduction} ============ The observation of the first gravitational wave (GW) signal by LIGO and Virgo opened up the new field of gravitational wave astronomy [@LIGO; @Virgo; @GEO600; @GW150914]. So far, ten confirmed binary black hole mergers and one binary neutron star coalescence have been reported [@GW151226; @GW170104; @GW170814; @GW170817; @GW170608; @GWTC-1]. Later this year KAGRA in Japan [@KAGRA] and in 2025 LIGO-India [@LIGO-India] are expected to join the global network of detectors, leading to improved parameter estimation and source localization. These ground-based detectors are sensitive to the decahertz-kilohertz frequency of the GW spectrum. In the future a space-based detector, LISA [@amaro-2017], will probe lower frequencies, around the millihertz range, and pulsar timing arrays (PTA) may measure ultra-low (nanohertz) frequency GWs [@IPTA-2013]. Currently – and this will most probably still hold in the future – compact binaries are the most important sources of observable GW signals. The events detected until now have all been found using circular waveform templates, this is because usually binaries have circularized due to GW emission [@peters-1963] when they enter the detection band of ground-based GW detectors. However, we know that binaries with substantial eccentricities exist. A simple example is provided by the Hulse-Taylor binary, with an eccentricity of $e \sim 0.6$ [@weisberg-2016]. Nonetheless, at the time this binary enters the detection band of ground-based GW detectors, it will have circularized to a negligible $e \sim 10^{-5}$ and not be distinguishable from a circular binary with current detector sensitivity [@sun-2015; @huerta-2018]. But especially in globular clusters and galactic nuclei, there are expected to be binaries with non-negligible eccentricity ($e>0.1$) emitting detectable GWs [@samsing-2018; @samsing-2014; @oleary-2007; @naoz-2013; @park-2017; @oleary-2009]. Hence, the detection of GWs from eccentric compact binaries could provide important information on compact object populations in globular clusters and galactic nuclei [@antonini-2016]. As soon as LISA will be operating, it will be able to observe compact binaries in our galaxy emitting GWs of much lower frequency. At this point they still are expected to have moderate eccentricities [@nelemans-2003; @sesana-2017]. On the other hand, LISA should be able to detect supermassive black hole binaries forming in the aftermath of galaxy mergers. Notably triple-induced coalescences are expected to have large eccentricities which remain significant until merger [@blaes-2002; @hoffman-2007; @amaro-2010; @bonetti-2018-1; @bonetti-2018-2]. The above anticipated prospects of future GW observations have motivated the development of eccentric waveform models. In the inspiralling phase one usually uses the post-Newtonian (PN) formalism to model the dynamics of the binary. This introduces three distinctive timescales. The first two, the orbital and the periastron precession timescales, are associated with the conservative dynamics and commonly described by the quasi-Keplerian parametrization [@damour-1985; @memmesheimer-2004]. The third timescale appears when the dissipative radiation-reaction effects are taken into account [@damour-2004; @koenigsdoerffer-2006]. Applying this description of the binary, several waveform models have been built [@yunes-2009; @cornish-2010; @key-2011; @huerta-2014; @huerta-2017; @huerta-2018; @gopakumar-2011; @tanay-2016; @hinder-2018; @cao-2017; @klein-2018]. In general, the far zone gravitational radiation field receives instantaneous and hereditary contributions. The instantaneous part is determined by the state of its source at a given retarded time while the hereditary part depends on the entire dynamical history of the source. The latter further contains, notably, tail and memory pieces. In this work, we concentrate on the memory contributions to the waveform from eccentric binaries. Normally we think of gravitational waves as oscillatory perturbations propagating on the background metric at the speed of light. However, all GW sources are subject to the so-called gravitational-wave memory effect, which manifests in a difference of the observed GW amplitudes at late and early times, $$\begin{aligned} \Delta h_{\textnormal{mem}}&= \lim_{t \rightarrow +\infty} h(t) - \lim_{t \rightarrow - \infty} h(t) \,.\end{aligned}$$ In an ideal, freely falling GW detector the GW memory causes a permanent displacement after the GW has passed. There are two main types of GW memory: The linear memory [@zeldovich-1974] originates from a net change in the time derivatives of the source multipole moments between early and late times, present mainly in unbound (e.g. hyperbolic binaries) systems. For bound systems the linear memory is negligible, as long as the components were formed, captured or underwent mass loss long before the GW driven regime. The non-linear memory, also called “Christodoulou memory" [@christodoulou-1991; @wiseman-1991; @thorne-1992; @blanchet-1992; @payne-1983], is a phenomenon directly related to the non-linearity of general relativity (GR). It arises from GWs sourced by previously emitted GWs. Since the non-linear memory is not produced directly by the source but rather by its radiation, it is present in all sources of GWs. From a more theoretical perspective, the memory effect and its variants can be interpreted in terms of conserved charges at null infinity and “soft theorems" [@strominger-2014; @pasterski-2016]. Several attempts are made to look for the memory effect, mostly with PTA’s, which would observe a sudden change in the pulse frequency of a pulsar [@seto-2009; @vanhaasteren-2010; @cordes-2012; @wang-2015]. On the other hand, ground-based detectors like LIGO, although not sensitive enough to the memory of a single event, could allow for a detection by the accumulation of several events. [@favata-2009-2; @lasky-2016; @mcneill-2017]. For circular binaries the non-linear, non-oscillatory memory contributions to the waveform have been computed at the 3PN order in [@favata-2009]. Regarding eccentric binaries, the leading order DC memory terms were obtained in [@favata-2011]. In this paper, we extend these computations to 3PN level by computing all terms coming from the memory contribution to the radiative mass multipoles. Note that this yields not only the “genuine” DC memory, but also oscillatory contributions. In the circular limit, the latter have been computed in [@blanchet-2008]. Due to complicated hereditary integrals, we calculate the memory contributions within a small eccentricity expansion. We present all our results in modified harmonic (MH) gauge in terms of the post-Newtonian parameter $x = ( G m \bar{\omega} / c^3 )^{2/3}$ and the eccentricity $e = {\bar{e}}_t$, with $\bar{\omega} = (1+\bar{k})\bar{n}$ being the orbital frequency and $\bar{n} = 2 \pi / P$ the mean motion. With the instantaneous contributions already available [@mishra-2015], and the tail and post-adiabatic contributions computed in a companion paper [@boetzel-2019] – hereafter called Paper I – this work aims to complete the knowledge of the 3PN waveform valid during the early inspiral of eccentric binary systems. This paper is structured as follows: In \[sec:prerequisites\] we discuss how the non-linear memory arises from the gravitational-wave energy flux and how it can be computed by integrating this flux over the binary’s past history. In \[sec:memcomp\] we explicitly evaluate the past-history integrals, which lead to two types of memory terms – DC memory and oscillatory memory – that are discussed separately in \[sec:DCmemory,sec:oscmemory\]. We next combine our results with the already available instantaneous and tail contributions and discuss the full 3PN waveform in \[sec:fullwaveform\]. In \[sec:summary\] we give a brief summary and conclude our work. Most expressions in this paper are presented only to leading order in eccentricity for convenience, though we provide the complete results to ${\mathcal{O}}(e^6)$ in a supplemental Mathematica notebook [@supplement]. Prerequisites {#sec:prerequisites} ============= Memory contribution to the mass multipole moments ------------------------------------------------- We state here briefly the essentials important for the memory calculation. The conventions and notations used are the same as outlined in Sec. II of Paper I. The gravitational waveform polarizations can be uniquely decomposed into the spherical harmonic modes, $h^{\ell m}$, via $$\begin{aligned} \label{eq:hlmmodes} h_+ - {\mathrm{i}}h_\times &= \sum_{\ell=2}^{\infty} \sum_{m=-\ell}^{\ell} h^{\ell m} \, Y^{\ell m}_{-2} (\Theta, \Phi) \,,\end{aligned}$$ where the basis is formed by the spin-weighted spherical harmonics $Y^{\ell m}_{-2} (\Theta, \Phi)$ and the amplitude modes $$\begin{aligned} \label{eq:hlmofu} h^{\ell m} &= -\frac{G}{\sqrt{2} R c^{\ell+2}} \left( U^{\ell m} - \frac{{\mathrm{i}}}{c} V^{\ell m} \right) \,,\end{aligned}$$ are given in terms of radiative mass and current multipoles, $U^{\ell m}$ and $V^{\ell m}$. These contain both instantaneous and hereditary parts. In the latter, we can further distinguish between tail and memory contributions (some of which may actually be tail induced) at the 3PN order, by schematically writing $$\begin{aligned} U^{\ell m} &= U^{\ell m}_{\textnormal{inst}}+ U^{\ell m}_{\textnormal{tail}}+ U^{\ell m}_{\textnormal{mem}}+ \delta U^{\ell m}\,,\\ V^{\ell m} &= V^{\ell m}_{\textnormal{inst}}+ V^{\ell m}_{\textnormal{tail}}+ \delta V^{\ell m} \,, \end{aligned}$$ where $\delta U^{\ell m}$ and $\delta V^{\ell m}$ represent possible higher-order hereditary terms. Note that there is no memory contribution to the radiative current-type moments [@blanchet-1992]. Now, employing the multipolar post-Minkowskian (MPM) post-Newtonian (PN) formalism, the radiative moments can be written in terms of the source moments. These relations can be found in Sec. III A of [@mishra-2015] for the instantaneous parts, which only require the knowledge of the source motion at a given moment in retarded time $T_R$, and in Sec. II B of Paper I for the hereditary parts, which involve integrals over the entire dynamical past of the source. It was shown in [@blanchet-1992; @favata-2009] that, since the memory is sourced by GWs, its contribution to the radiative mass moment can be alternatively written in terms of the gravitational-wave energy flux as $$\begin{aligned} \label{eq:Ulmem} U^{\ell m}_{\textnormal{mem}}&= \frac{32\pi}{c^{2-\ell}} \sqrt{\frac{(\ell -2)!}{2(\ell +2)!}} \int_{-\infty}^{T_R} {\mathrm{d}}t \int {\mathrm{d}}\Omega \frac{{\mathrm{d}}E^{\textnormal{GW}}}{{\mathrm{d}}t {\mathrm{d}}\Omega} \bar Y^{\ell m}(\Omega) \,.\end{aligned}$$ We will start from this equation to compute the memory contributions to the GW amplitude to the 3PN order. The GW energy flux is related to the GW stress-energy tensor [@thorne-1980] which can be expressed in terms of the polarization amplitudes as $$\begin{aligned} \frac{{\mathrm{d}}E^{\textnormal{GW}}}{{\mathrm{d}}t \, {\mathrm{d}}\Omega} &= c R^2 t_{00}^{\textnormal{GW}}= \frac{c^3 R^2}{16 \pi G} {\langle}\dot{h}_+^2 + \dot{h}_\times^2 {\rangle}\,,\end{aligned}$$ where the angle brackets denote an average over several wavelengths. Inserting the mode decomposition defined in \[eq:hlmmodes\] we find the GW energy flux in terms of the time derivatives of the $h^{\ell m}$ modes: $$\begin{aligned} \label{eq:gwflux} \frac{{\mathrm{d}}E^{\textnormal{GW}}}{{\mathrm{d}}t \, {\mathrm{d}}\Omega} = &\; \frac{c^3 R^2}{16 \pi G} \sum_{\ell'=2}^{\infty} \sum_{\ell''=2}^{\infty} \sum_{m'=-\ell'}^{\ell'} \sum_{m''=-\ell''}^{\ell''} {\langle}\dot{h}^{\ell' m'} \dot{\bar{h}}^{\ell'' m''} {\rangle}\no &\times Y^{\ell' m'}_{-2}(\theta,\phi) \bar Y^{\ell'' m''}_{-2}(\theta,\phi) \,.\end{aligned}$$ We insert this expression into \[eq:Ulmem\]. The time derivative of the memory contribution to the mass multipole moment, $U^{\ell m(1)}_{\textnormal{mem}}= {\mathrm{d}}U^{\ell m}_{\textnormal{mem}}/ {\mathrm{d}}T_R$, which is nothing but the memory contribution before integration over past history, may thus be expressed as $$\begin{aligned} \label{eq:Ulmmem} U^{\ell m(1)}_{\textnormal{mem}}=&\; \frac{c^{\ell+1} R^2}{G} \sqrt{\frac{2(\ell-2)!}{(\ell+2)!}} \sum_{\ell'=2}^{\infty} \sum_{\ell''=2}^{\infty} \sum_{m'=-\ell'}^{\ell'} \sum_{m''=-\ell''}^{\ell''} \no &\times G^{\ell \ell' \ell''}_{m m'm''} {\langle}\dot{h}^{\ell'm'} \dot{\bar h}^{\ell''m''}{\rangle}\,,\end{aligned}$$ where $G^{\ell \ell' \ell''}_{m m'm''}$ is the angular integral of a product of three spin-weighted spherical harmonics $$\begin{aligned} \label{eq:angint} G^{\ell \ell' \ell''}_{m m' m''} = \int {\mathrm{d}}\Omega \, \bar Y{^{\ell m}} \, Y^{\ell' m'}_{-2} \, \bar Y^{\ell'' m''}_{-2} \,.\end{aligned}$$ Reference [@NIST:DLMF] provides an explicit formula for this integral: $$\begin{aligned} \label{eq:G} G^{\ell \ell' \ell''}_{m m'm''} =&\; (-1)^{m + m'} \sqrt{\frac{(2\ell +1)(2\ell' +1)(2\ell'' +1)}{4\pi}} \no &\times \begin{pmatrix} \ell & \ell' & \ell'' \\ 0 & -2 & 2 \end{pmatrix} \begin{pmatrix} \ell & \ell' & \ell'' \\ -m & m' & -m'' \end{pmatrix} \,.\end{aligned}$$ The brackets denote the Wigner 3-$j$ symbols. Instantaneous and tail parts of the spherical harmonic modes {#sec:hlminsttail} ------------------------------------------------------------ Remembering that the dominant modes correspond to the quadrupolar case $\ell = 2$, with $h^{2m} = {\mathcal{O}}(c^{-4})$, we see from \[eq:Ulmmem\] that the memory integrands are of the order 2.5PN. However, as discussed below, the $U^{\ell 0 (1)}_{\textnormal{mem}}$ contain besides oscillatory complex exponentials also non-oscillatory terms. Due to the integration over the past history, their contributions at times $t\le T_R$ accumulate and enhance the result by a net factor $c^5$. It follows that the leading memory effect in the polarizations actually arises at the relative Newtonian order. Thus, \[eq:Ulmmem\] implies that, as an input for the computation of the 3PN accurate $U^{\ell m(1)}_{\textnormal{mem}}$, we need a priori all non-memory $h^{\ell m}$ modes to 3PN order. It is in fact not surprising that part of the waveform is required to calculate the full waveform since the non-linear memory originates from gravitational waves sourced by the energy flux of gravitational waves emitted in the past, as shown by \[eq:Ulmem\]. Note that the contribution from the memory to the memory itself turns out not to enter the waveform up to the 3PN order. In [@favata-2009] it was argued that for circular binaries these contributions would appear at the 5PN level, though for eccentric binaries we find, by explicit calculation, that these appear already at the 4PN order. This is due to additional oscillatory memory contributions that will be discussed in \[sec:oscmemory\] below. However, for the present work, these memory-of-memory terms can be safely ignored. The instantaneous parts of the 3PN accurate $h^{\ell m}$ modes describing inspiralling eccentric binaries have been computed in [@mishra-2015]. In Paper I the tail contributions have been derived, as well as the post-adiabatic corrections to the instantaneous contributions. Instantaneous mode amplitudes in [@mishra-2015] are written in terms of the post-Newtonian parameter $x$, the time eccentricity $e_t$ and parametrized by the eccentric anomaly $u$ and are valid for arbitrary eccentricities, while the tail contributions in Paper I are given in a small eccentricity expansion, parametrized by the mean anomaly $l$. The same will hold for the memory parts. Inverting the 3PN accurate Kepler equation by means of the solution developed in [@boetzel-2017], the instantaneous terms can be parametrized by the mean anomaly as well. Instead of restating the quasi-Keplerian parametrization and the phasing formalism describing the dynamics of the binary, we refer the reader to Sec. II C and D of Paper I where those aspects are summarized with the same conventions and notations. The $h^{\ell m}$ modes including instantaneous, tail and post-adiabatic contributions, are given in the following form $$\begin{aligned} \label{eq:hlm} h^{\ell m} &= \frac{8 Gm \nu}{c^2 R} x \sqrt{\frac{\pi}{5}} {\mathrm{e}}^{-{\mathrm{i}}m \psi} H^{\ell m} \,,\end{aligned}$$ where the $H^{\ell m}$ are written in terms of the adiabatic post-Newtonian parameter $x \equiv {\bar{x}}$, the time eccentricity $e \equiv {\bar{e}}_t$, and parametrized by the angles $\xi$ and $\psi$. See for instance, Eq. (76) of Paper I for the dominant mode ($h^{22}$) expression. The phase angles $\xi$ and $\psi$ arise naturally when applying a certain shift to the time coordinate aimed at eliminating the arbitrary constant $x_0$ appearing in both the instantaneous and tail parts [@blanchet-1996; @arun-2004], through the redefinitions $$\begin{aligned} \xi &= l - \frac{3 G M}{c^3} n \ln \left( \frac{x}{x_0'} \right) \label{eq:xi} \,,\\ \lambda_\xi &= \lambda - \frac{3 G M}{c^3} (1 + k) n \ln \left( \frac{x}{x_0'} \right) \,,\end{aligned}$$ where $M = m (1 - \nu x / 2)$ denotes the Arnowitt-Deser-Misner (ADM) mass, $m = m_1 + m_2$ the total mass, $\nu = m_1 m_2 / m^2$ the symmetric mass ratio and $x_0'$ is related to $x_0$ by $$\begin{aligned} \ln x_0' &= \frac{11}{18} - \frac{2}{3} {\gamma_\textnormal{E}}- \frac{4}{3} \ln 2 + \frac{2}{3} \ln x_0 \,,\end{aligned}$$ with ${\gamma_\textnormal{E}}$ being Euler’s constant. We refer to Appendix B of Paper I for the relations between the orbital elements ($l$, $\lambda$, $\phi$) and their redefined counterparts ($\xi$, $\lambda_\xi$, $\psi$). Computation of the non-linear memory {#sec:memcomp} ==================================== Memory contributions to the time derivative of the radiative moments {#sec:Umemdots} -------------------------------------------------------------------- The computation of the memory contributions to the radiative mass multipole using \[eq:Ulmmem\] involves products of the time derivatives of the $h^{\ell m}$ modes given in \[eq:hlm\]. These are obtained by expressing $\psi$ in terms of $\xi$ and $\lambda_\xi$ and applying the following time derivative operator, $$\begin{aligned} \frac{{\mathrm{d}}}{{\mathrm{d}}t} &= n \left[\frac{{\mathrm{d}}}{{\mathrm{d}}\xi} + (1+k) \frac{{\mathrm{d}}}{{\mathrm{d}}\lambda_\xi} \right] + \frac{{\mathrm{d}}x}{{\mathrm{d}}t} \frac{{\mathrm{d}}}{{\mathrm{d}}x} + \frac{{\mathrm{d}}e}{{\mathrm{d}}t} \frac{{\mathrm{d}}}{{\mathrm{d}}e} \,,\end{aligned}$$ where we have used the facts that ${\mathrm{d}}\xi/{\mathrm{d}}t = {\mathrm{d}}l/{\mathrm{d}}t = n$ and ${\mathrm{d}}\lambda_\xi/{\mathrm{d}}t = {\mathrm{d}}\lambda/{\mathrm{d}}t = (1+k)n$ to the required PN order. The secular time evolution of $x$ and $e$ is given, at leading order, by the formulas of Peters & Mathews [@peters-1963; @peters-1964] $$\begin{aligned} \frac{{\mathrm{d}}x}{{\mathrm{d}}t} =& \frac{c^3 \nu}{Gm} \frac{x^5}{(1-e^2)^{7/2}} \left(\frac{64}{5} + \frac{584}{15} e^2 + \frac{74}{15} e^4 \right) \,,\\ \frac{{\mathrm{d}}e}{{\mathrm{d}}t} =& -\frac{c^3 \nu}{Gm} \frac{e \, x^4}{(1-e^2)^{5/2}} \left( \frac{304}{15} + \frac{121}{15} e^2 \right) \,.\end{aligned}$$ Note that they cause a 2.5PN correction, thus the leading order is sufficient here. When computing the time derivatives of the amplitude modes $h^{\ell m}\sim x^{\ell/2}/c^2$, we have the following leading order PN scaling: $$\begin{aligned} \dot{h}^{\ell m} \sim (\omega/c^2)\, x^{\ell/2} \sim c\, x^{\ell/2+ 3/2}\,.\end{aligned}$$ As the dominant mode $\ell=2$ is of order $c\,x^{5/2}$, the knowledge of the waveform to 3PN order requires modes up to $\ell=8$. According to this argument, the sums in \[eq:Ulmmem\] consisting of products $\dot{h}^{\ell'm'} \dot{\bar h}^{\ell''m'}$ may be truncated at $\ell'=\ell''=8$. Moreover, the appearance of the 3-$j$ symbols in \[eq:G\] imply some selection rules: the three lower entries have to add up to zero, i.e., $m= m'-m''$. Since the mode products appearing in \[eq:Ulmmem\] scale like $$\begin{aligned} \dot{h}^{\ell'm'} \dot{\bar h}^{\ell''m''} \sim x^{n/2} \, {\mathrm{e}}^{-{\mathrm{i}}(m'-m'')\lambda_\xi}\,,\end{aligned}$$ for some integer $n$, only memory modes with $m=0$ will contain DC terms – as was previously found for circular orbits [@favata-2009]. The scaling being the same as in that case, we have to compute the $U^{\ell 0 (1)}_{\textnormal{mem}}$ up to $\ell=10$. On the other hand, a mode separation property holds for planar orbits [@kidder-2008; @faye-2012]: The $h^{\ell m}$ only depend on the mass (current) radiative moments if $\ell+m$ is even (odd). Thus, as there is no memory effect in the current radiative moment, neither is there when $\ell+m$ is odd. As an example, we show the leading order part of the $20$-mode up to ${\mathcal{O}}(e^2)$, which will represent the dominant memory contribution: $$\begin{aligned} U^{20 (1)}_{\textnormal{mem}}=&\; -\sqrt{\frac{\pi}{15}} \frac{c^5\nu^2}{G} x^5 \bigg(\frac{256}{7} +\frac{5008 e^2}{21}+\frac{768}{7} e\, {\mathrm{e}}^{-{\mathrm{i}}\xi} \no &+ \frac{768}{7} e\, {\mathrm{e}}^{{\mathrm{i}}\xi}+\frac{5176}{21} e^2\, {\mathrm{e}}^{-2 {\mathrm{i}}\xi}+\frac{5176}{21} e^2\, {\mathrm{e}}^{2 {\mathrm{i}}\xi}\bigg)\,.\end{aligned}$$ We observe two different type of terms – oscillatory terms proportional to ${\mathrm{e}}^{-{\mathrm{i}}m'\xi}$ and non-oscillatory ones, which give rise to the well-known leading order DC memory. As argued above, the $m \neq 0$ modes only contain oscillatory terms since they are proportional to ${\mathrm{e}}^{-{\mathrm{i}}m \lambda_\xi}$. For instance, the leading order of the $22$-mode explicitly is $$\begin{aligned} U^{22 (1)}_{\textnormal{mem}}= &-\sqrt{\frac{2\pi}{5}} \frac{c^5\nu^2}{G} x^5 \, {\mathrm{e}}^{-2 {\mathrm{i}}\lambda_\xi} \bigg( \frac{40}{7}e^2 -\frac{32}{21} e\, {\mathrm{e}}^{-{\mathrm{i}}\xi}\no &+\frac{32}{21} e\, {\mathrm{e}}^{{\mathrm{i}}\xi} -\frac{172}{21} e^2 {\mathrm{e}}^{-2 {\mathrm{i}}\xi} +\frac{52}{21} e^2 {\mathrm{e}}^{2 {\mathrm{i}}\xi}\bigg)\,.\end{aligned}$$ In \[eq:Ulmmem\], used for the calculation of the $U^{\ell m(1)}_{\textnormal{mem}}$, an average over several wavelengths denoted with angle brackets appears due to the way the gravitational-wave stress-energy tensor $t_{\mu\nu}^{\textnormal{GW}}$ is constructed [@isaacson-1968]. However, such an averaging procedure is unnecessary in the derivation of the non-linear memory valid at quadratic order in $G$ proposed in [@christodoulou-1991; @blanchet-1992]. The orbital average entering the calculation of the leading order eccentric memory performed in [@favata-2011] effectively removes the terms proportional to ${\mathrm{e}}^{{\mathrm{i}}m'\xi}$ in the $U^{\ell 0(1)}_{\textnormal{mem}}$ so that only the terms yielding the DC memory are left over, while the discarded pieces do not affect the amplitude of the DC memory. In the absence of orbital average, these pieces lead to small amplitude oscillatory contributions to the waveform, which we will call here oscillatory memory contributions. It would actually be delicate to introduce an orbital average in the $m \neq 0$ modes because these terms are not only oscillating on the orbital timescale but also on the much longer precession timescale. Based on the computation of the memory using directly the source moments, as outlined in \[sec:altmemory\], where such an average does not appear, we will ignore wavelength or orbital averages here. DC memory {#sec:DCmemory} --------- The next step consists in evaluating the hereditary time integral $$\begin{aligned} \label{eq:heredint} U^{\ell m}_{\textnormal{mem}}= \int_{-\infty}^{T_R} {\mathrm{d}}t \, U^{\ell m(1)}_{\textnormal{mem}}\,.\end{aligned}$$ To do so, we need a model for the secular evolution of the binary undergoing gravitational radiation-reaction forces. For a quasi-elliptical, inspiralling binary, the secular 3PN order evolution equations of the orbital elements has been obtained in [@arun-2008-1; @arun-2008-2; @arun-2009]. This model is an idealization since it assumes that the two components start at infinite separation and the orbital energy decreases solely due to the emission of gravitational waves. The explicit integrals appearing in \[eq:heredint\] are of two different types. The first one consists of a product of $x$ and $e$, each with some power $p$ and $q$, respectively: $$\begin{aligned} \label{eq:DCmemint} U^{\ell 0}_{\textnormal{DC}}&\sim \int_{-\infty}^{T_R} {\mathrm{d}}t\, x^p(t) \, e^q(t)\,.\end{aligned}$$ The leading Newtonian order corresponds to $p=5$. The possible values of the integer $q$ range from $0$ – the quasi-circular limit – to the order of eccentricity expansion. These integrals give the non-oscillatory contributions to the waveform, i.e. the DC memory. Note that as argued above, these terms are only present in the $m=0$ modes. The second type of integrals will lead to oscillatory terms appearing at 1.5PN, 2.5PN and 3PN order in the waveform. We will discuss these in \[sec:oscmemory\]. The strategy to evaluate the integral in \[eq:DCmemint\] is to express the PN-parameter $x$ in terms of the eccentricity $e$ and change the integration variable from time $t$ to $e$, so that the integral runs from some initial eccentricity $e_i$ at early times to $e(T_R)$ at current retarded time: $$\begin{aligned} \label{eq:DCint} U^{\ell 0}_{\textnormal{DC}}\sim \int_{e_i}^{e(T_R)} {\mathrm{d}}e\,\left(\frac{{\mathrm{d}}e}{{\mathrm{d}}t}\right)^{-1} x^p(e) \, e^q \,.\end{aligned}$$ The time evolution of $x$ and $e$ due to radiation reaction is stated to leading order in \[eq:peters-mathews\]. Here, we need the evolution equations up to 3PN order, they are provided in \[sec:evolveq\]. We form the ratio of the two equations, thereby canceling the time dependence, and expand the right-hand side in $x$ and $e$. This yields a differential equation with the following structure: $$\begin{aligned} \label{eq:dxde} \frac{{\mathrm{d}}x}{{\mathrm{d}}e} = &\; f_\mathrm{N}(e) \, x + f_1(e)\, x^2 + f_{1.5}(e) \,x^{5/2} \no &+f_2(e) \,x^3 + f_{2.5}(e) \,x^{7/2} + f_3(e, \ln x) \,x^4 \,.\end{aligned}$$ Here, the $f_i(e)$ terms represent the coefficients of $x^{i+1}$ in the expansion of ${\mathrm{d}}x/{\mathrm{d}}e$, with $f_{\mathrm{N}}=f_0$. To solve this differential equation, we search for the unknown function $x(e)$ in the form of a perturbative expansion, according to $$\begin{aligned} x = x_\mathrm{N} + \epsilon x_1 + \epsilon^{{3/2}} x_{1.5} + \epsilon^2 x_2 + \epsilon^{{5/2}} x_{2.5} + \epsilon^3 x_3\,,\end{aligned}$$ where $\epsilon$ is a formal parameter that allows one to keep track of the PN order. Inserting this expansion into \[eq:dxde\] and identifying the coefficients of $\epsilon^i$ on the left and the right-hand sides of the resulting equation, we find the set of differential relations satisfied by the post-Newtonian orders of $x$. This system can be straightforwardly solved by quadrature. Putting the pieces together yields the PN parameter $x$ as a function of eccentricity. At leading order in the PN and eccentricity expansion, we recover [@damour-2004] $$\begin{aligned} \label{eq:xofe} x(e) = x_0 \left(\frac{e_0}{e}\right)^{12/19}\,, \end{aligned}$$ where $x_0$ is the value of $x$ at some reference eccentricity $e_0$. The full 3PN result to leading order in eccentricity is provided in \[sec:evolveq\]. Note that for the expansion in eccentricity to be valid, the eccentricity has to be small at all times, hence $e_0$ has to be small as well. We are now in the position to insert the evolution equation for $e$ and the solution for $x(e)$ into \[eq:DCint\]. Expanding again in $x$ and $e$ yields elementary integrals to be calculated. We then re-express this result in terms of the time dependent quantities $x$ and $e$ by solving their relation (\[eq:xofe,eq:xofe3\]) for $x_0$ and reinsert the expression of this quantity in terms of $x$ and $e$ into the calculated memory terms. A last Taylor expansion then yields the DC memory pieces of the mass multipole moments. We present the memory contributions to the spherical harmonic modes in the following form: $$\begin{aligned} \label{eq:hlmmemory} h^{\ell m}_{\textnormal{mem}}=&\; - \frac{G}{\sqrt{2} c^{l+2} R} U^{\ell m}_{\textnormal{mem}}\no =&\; \frac{8 G m \nu}{c^2 R} x \sqrt{\frac{\pi}{5}} {\mathrm{e}}^{-{\mathrm{i}}m \psi} H^{\ell m}_{\textnormal{mem}}\,.\end{aligned}$$ With this convention, the memory pieces directly add to the waveform modes stated in \[eq:hlm\]. As the expressions are quite long, we present here only the $H^{20}_{\textnormal{DC}}$ mode to 3PN and leading order in eccentricity: \[eq:h20mem\] $$\begin{aligned} H^{20}_{\textnormal{DC}}=&\; -\frac{5}{14 \sqrt{6}} \left( H^{20}_{\textnormal{Newt}}+ x H^{20}_{\textnormal{{1}PN}} + x^{3/2} H^{20}_{\textnormal{{1.5}PN}} +x^{2} H^{20}_{\textnormal{{2}PN}} +x^{5/2} H^{20}_{\textnormal{{2.5}PN}} +x^{3} H^{20}_{\textnormal{{3}PN}}\right) \,,\\ H^{20}_{\textnormal{Newt}}=&\; 1 -\left( \frac{e}{e_i} \right)^{12/19} \,,\\ H^{20}_{\textnormal{{1}PN}} =&\; -\frac{4075}{4032} +\frac{67 \nu}{48} +\left(\frac{e}{e_i} \right)^{12/19} \left( -\frac{2833}{3192} +\frac{197 \nu}{114} \right) +\left( \frac{e}{e_i} \right)^{24/19} \left( \frac{145417}{76608}-\frac{2849 \nu}{912}\right) \,,\\ H^{20}_{\textnormal{{1.5}PN}} =&\; -\frac{377\pi}{228} \left(\frac{e}{e_i} \right)^{12/19} +\frac{377\pi}{228} \left( \frac{e}{e_i}\right)^{30/19} \,,\\ H^{20}_{\textnormal{{2}PN}} =&\; -\frac{151877213}{67060224} -\frac{123815 \nu}{44352} +\frac{205 \nu^2}{352} +\left(\frac{e}{e_i}\right)^{12/19} \left(\frac{358353209}{366799104} -\frac{738407 \nu}{727776} -\frac{20597 \nu^2}{17328}\right)\no &+\left(\frac{e}{e_i}\right)^{24/19} \left(\frac{411966361}{122266368} -\frac{825950 \nu}{68229} +\frac{561253 \nu^2}{51984}\right) \no &+\left(\frac{e}{e_i}\right)^{36/19} \biggl(-\frac{50392977379}{24208740864}+ \frac{764295307 \nu}{48033216} -\frac{11654209\nu^2}{1143648}\biggr) \,,\\ H^{20}_{\textnormal{{2.5}PN}} =&\; -\frac{253 \pi}{336} +\frac{253 \pi \nu}{84} +\left( \frac{e}{e_i} \right)^{12/19} \left( \frac{3763903 \pi}{7277760} +\frac{12788779 \pi \nu}{1819440}\right) +\left( \frac{e}{e_i} \right)^{24/19} \left(\frac{54822209 \pi}{8733312} -\frac{1074073 \pi \nu}{103968} \right) \no &+\left(\frac{e}{e_i}\right)^{30/19} \left(\frac{5340205 \pi}{1455552} -\frac{371345 \pi \nu}{51984} \right) +\left(\frac{e}{e_i}\right)^{42/19} \left(-\frac{424020733 \pi}{43666560} +\frac{27049187 \pi \nu}{3638880} \right) \,,\\ H^{20}_{\textnormal{{3}PN}} =&\; -\frac{4397711103307}{532580106240} +\left(\frac{700464542023}{13948526592} -\frac{205 \pi^2}{96}\right) \nu +\frac{69527951 \nu^2}{166053888} +\frac{1321981 \nu^3}{5930496} \no &+\left(\frac{e}{e_i}\right)^{12/19} \biggl[-\frac{4942027570449143}{96592876047360} -\frac{81025 \pi^2}{103968} +\frac{3317 {\gamma_\textnormal{E}}}{399} +\left( -\frac{10309531979}{7466981760} +\frac{3977 \pi^2}{3648}\right) \nu \no &+\frac{267351733 \nu^2}{82966464} +\frac{772583 \nu^3}{2222316} +\frac{12091 \ln 2}{5985} +\frac{78003 \ln 3}{5320} +\frac{3317 \ln x}{798}\biggr] +\frac{710645 \pi^2}{103968} \left( \frac{e}{e_i}\right)^{30/19} \no &+\left(\frac{e}{e_i}\right)^{24/19} \left(-\frac{31102835980319}{14049872879616} +\frac{279737759653 \nu}{167260391424} +\frac{26730466283 \nu^2}{1991195136} -\frac{397176241 \nu^3}{23704704}\right) \no &+\left(\frac{e}{e_i}\right)^{36/19} \left(-\frac{142763304914707}{25758100279296} +\frac{48901891428821 \nu}{919932152832} -\frac{400181473249 \nu^2}{3650524416} +\frac{2295879173 \nu^3}{43458624}\right) \no &+\left(\frac{e}{e_i}\right)^{48/19} \biggl[\frac{385621605844415513}{5740376633671680} -\frac{157405 \pi^2}{25992} -\frac{3317 {\gamma_\textnormal{E}}}{399} +\left( -\frac{49590995147570629}{478364719472640} +\frac{1271 \pi^2}{1216}\right) \nu \no &+\frac{3194536246463\nu^2}{34514049024} -\frac{1672948713 \nu^3}{45653504}-\frac{12091 \ln 2}{5985} -\frac{78003 \ln 3}{5320} -\frac{3317 \ln x}{798} -\frac{6634}{2527} \ln \left( \frac{e}{e_i} \right) \biggr] \,.\end{aligned}$$ All non-zero DC memory modes are presented to leading order in eccentricity in \[sec:DCmemlist\] and to ${\mathcal{O}}(e^6)$ in the supplemental material [@supplement]. An important check is to take the circular limit of our calculated memory modes and compare to the circular 3PN memory modes computed in [@favata-2009]. To illustrate this fact, we take the circular limit of the $20$-mode stated in \[eq:h20mem\] by setting $e=0$ and find $$\begin{aligned} H^{20}_{\textnormal{DC}}= &-\frac{5}{14 \sqrt{6}} \Bigg\{1+ x\left(-\frac{4075}{4032}+ \frac{67 \nu}{48}\right)+ x^2 \left(-\frac{151877213}{67060224}-\frac{123815 \nu}{44352}+\frac{205 \nu^2}{352}\right) +x^{5/2} \left(-\frac{253 \pi}{336}+\frac{253 \pi \nu}{84}\right) \no &+x^3 \left[-\frac{4397711103307}{532580106240} +\left(\frac{700464542023}{13948526592}-\frac{205 \pi^2}{96}\right) \nu +\frac{69527951 \nu^2}{166053888}+\frac{1321981 \nu^3}{5930496}\right] \Bigg\}\,,\end{aligned}$$ in perfect agreement with Eq. (4.3a) of [@favata-2009]. The higher DC modes up to $\ell = 10$ in the circular limit are consistent with Eq. (4.3) of [@favata-2009] as well. Moreover, we can check the leading eccentricity part at Newtonian order against Eq. (2.35) in [@favata-2011]. They are found to be equal. Note that at Newtonian order the computation of the DC memory is in principle possible for arbitrary eccentricities (see Eq. 2.34 in [@favata-2011]), however this becomes difficult at higher PN orders, especially when tail terms come into play. Oscillatory memory {#sec:oscmemory} ------------------ Before considering the oscillatory integrals, let us recall some properties of the non-linear memory. As mentioned at the beginning of \[sec:hlminsttail\], the memory contribution to the radiative mass multipole is formally of 2.5PN order. But due to the hereditary nature, the non-oscillatory terms are raised by 2.5PN orders to appear already at Newtonian level. From the oscillatory terms we cannot expect the same behavior, due to the fact that the oscillations in the remote past are effectively canceling each other out. Thus we expect only the recent past to be contributing. Examining the remaining oscillatory integrals, we notice that they are of the following form: $$\begin{aligned} \label{eq:oscmemint} U^{\ell m}_{\textnormal{osc}}&\sim \int_{-\infty}^{T_R} {\mathrm{d}}t\, x^p(t) \, e^q(t) \, {\mathrm{e}}^{{\mathrm{i}}(s \lambda_\xi + r \xi)} \,.\end{aligned}$$ Note that we have $s = -m$. Here we directly provide a formula to evaluate these integrals, its derivation is presented in \[sec:oscintegral\]. Using the fact that $\lambda_\xi = (1+k)\xi$ and $\xi = nt$ to the required PN order as well as the notion that the integral is essentially given by the contributions at recent time we find $$\begin{aligned} \label{eq:oscmem} U^{\ell m}_{\textnormal{osc}}\sim -\frac{{\mathrm{i}}}{n(r+s(1+k))} x^p\, e^q \, {\mathrm{e}}^{{\mathrm{i}}(s \lambda_\xi + r \xi)} \,,\end{aligned}$$ where the time dependence on $T_R$ is not written explicitly. Expanding the denominator, we have to distinguish between two different cases. The first applies if $r \neq -s$, we then find $$\begin{aligned} \label{eq:fastosc} U^{\ell m}_{\textnormal{osc}}\sim\; -\frac{{\mathrm{i}}}{r+s} x^{p-3/2}\, e^q \, {\mathrm{e}}^{{\mathrm{i}}(s \lambda_\xi + r \xi)}\,.\end{aligned}$$ Since $p=5$ at Newtonian order and the leading terms in the waveform are of order $x$, these integrals lead to 2.5PN contributions to the waveform. As we have expected, these kind of terms oscillating on the orbital timescale keep their formal PN order, we call them the fast oscillatory memory. On the other hand we find for $r = -s$: $$\begin{aligned} \label{eq:slowosc} U^{\ell m}_{\textnormal{osc}}\sim& -\frac{{\mathrm{i}}}{3s}\left[ x^{p-5/2} + x^{p-3/2}\left(-\frac{3}{2}+ \frac{7\nu}{3}\right) \right]\no &\times e^q \, {\mathrm{e}}^{{\mathrm{i}}s (\lambda_\xi - \xi)} + {\mathcal{O}}(e^{q+2})\,.\end{aligned}$$ This corresponds to terms that oscillate solely on the periastron precession timescale, we therefore call these terms the slow oscillatory memory. Because of the much slower oscillations they are enhanced by 1PN order corresponding to the PN order of precession and enter the waveform at 1.5PN. Note also that in \[eq:slowosc\] eccentricity corrections of ${\mathcal{O}}(e^{q+2})$ appear, whereas \[eq:fastosc\] would only be affected by eccentricity corrections starting at 3.5PN order. We provide the oscillatory memory contributions to the spherical harmonic modes in the same form as for the DC memory, according to \[eq:hlmmemory\]. Besides the DC memory contribution, the $20$-mode contains also fast oscillatory memory at 2.5PN: $$\begin{aligned} H^{20}_{\textnormal{osc}}=&\; {\mathrm{i}}\frac{16\sqrt{6}}{7} \nu\,e\, x^{5/2} \biggl( -{\mathrm{e}}^{-{\mathrm{i}}\xi} +{\mathrm{e}}^{{\mathrm{i}}\xi} \no &-\frac{647}{576} e \,{\mathrm{e}}^{-2 {\mathrm{i}}\xi} +\frac{647}{576} e\, {\mathrm{e}}^{2 {\mathrm{i}}\xi} \biggr)\,.\end{aligned}$$ Note that while the DC memory is purely real and therefore only affects the plus polarization (with the usual conventions on the polarization triad), the oscillatory contributions influence both polarizations. In the $m \neq 0$ modes, only oscillatory memory is present. For the dominant $22$-mode we find: $$\begin{aligned} H&^{22}_{\textnormal{osc}}= {\mathrm{i}}\, e^2 \nu \, {\mathrm{e}}^{2{\mathrm{i}}\xi} \biggl[ -\frac{13}{252} x^{3/2} \no &+ \left( \frac{697}{336} -\frac{865 \nu}{216} \right)x^{5/2} -\frac{29\pi}{126} x^{3} \biggr] \no &+ 21 {\mathrm{i}}x^{5/2} e \nu \left[ \frac{19}{6}e +\frac{4}{3} {\mathrm{e}}^{-{\mathrm{i}}\xi} - 4 {\mathrm{e}}^{{\mathrm{i}}\xi} +\frac{65}{24} e {\mathrm{e}}^{-2{\mathrm{i}}\xi}\right] \,.\end{aligned}$$ Here the slow oscillatory part in the first and second line is proportional to ${\mathrm{e}}^{2 {\mathrm{i}}\xi}$, as we factored out ${\mathrm{e}}^{-2{\mathrm{i}}\psi}$ according to \[eq:hlmmemory\]. Three different PN orders of slow oscillatory memory terms appear in this mode. The first one at 1.5PN arises from the leading order memory contribution to the radiative mass multipole at 2.5PN, so as expected it is enhanced by one post-Newtonian order. At 2.5PN, there is the 1PN correction to the first term as well as a part coming from the 1PN correction to the multipole. Finally, at 3PN there is a term originating from the 1.5PN correction to the memory part of the multipole – this corresponds to memory of the gravitational wave tail. The terms in the third line correspond to fast oscillatory memory entering at 2.5PN level. Full 3PN eccentric waveform {#sec:fullwaveform} =========================== In this section we summarize the results necessary for constructing the full waveform for eccentric binaries at third post-Newtonian order, including all instantaneous, hereditary and post-adiabatic contributions, as described in Sec. V of Paper I. Rather than listing the lengthy expressions, we give an overview at which PN order the individual terms enter the waveform and where they can be found. Explicit expressions for all spherical harmonic modes are given in a supplemental Mathematica notebook [@supplement]. We present the waveform in terms of the secular evolving PN parameter ${\bar{x}}$ and the time eccentricity ${\bar{e}}$, parametrized by the angles $\xi$ and $\psi$. We refer to Sec. V C of Paper I for their definition, and to Appendix B therein for various relations between the orbital elements ($l$, $\lambda$,$\phi$) and ($\xi$, $\lambda_\xi$, $\psi$). The secular evolution of the parameters ${\bar{x}}$ and ${\bar{e}}$ is given in \[sec:evolveq\]. The spherical harmonic modes describing the waveform are then written in the following form: $$\begin{aligned} \label{eq:hlmfull} h^{\ell m} = \frac{8 G m \nu}{c^2 R} {\bar{x}}\sqrt{\frac{\pi}{5}} {\mathrm{e}}^{-{\mathrm{i}}m \psi} H^{\ell m} \,.\end{aligned}$$ Modes with $m<0$ can be calculated from $$\begin{aligned} h^{\ell \, -m} = (-1)^\ell \, \bar h^{\ell m} \,.\end{aligned}$$ In general, the individual modes split in three types of contributions: $$\begin{aligned} H^{\ell m} = H^{\ell m}_{\textnormal{inst}}+ H^{\ell m}_{\textnormal{hered}}+ H^{\ell m}_{\textnormal{post-ad}}\,.\end{aligned}$$ The instantaneous terms depend only on the instantaneous state of the source at a given retarded time, with contributions at different orders relative to the leading Newtonian (${\textnormal{Newt}}$) order given as: $$\begin{aligned} H^{\ell m}_{\textnormal{inst}}=&\; (H^{\ell m}_{\textnormal{inst}})_{\textnormal{Newt}}+ (H^{\ell m}_{\textnormal{inst}})_{\textnormal{{1}PN}} + (H^{\ell m}_{\textnormal{inst}})_{\textnormal{{1.5}PN}} \nonumber\\ &+ (H^{\ell m}_{\textnormal{inst}})_{\textnormal{{2}PN}} + (H^{\ell m}_{\textnormal{inst}})_{\textnormal{{2.5}PN}} + (H^{\ell m}_{\textnormal{inst}})_{\textnormal{{3}PN}} \,.\end{aligned}$$ These are given in terms of $x$, $e$ and $u$ in Eqs. (5.09–5.11) and Eq. (A1) of [@mishra-2015]. The parametrization in terms of $u$ has to be transformed to $\xi$ using Eq. (B2b) in Paper I. The post-adiabatic contributions are introduced by radiation-reaction corrections to the quasi-Keplerian parametrization, at relative 2.5PN order: $$\begin{aligned} H^{\ell m}_{\textnormal{post-ad}}=&\; (H^{\ell m}_{\textnormal{post-ad}})_{\textnormal{{2.5}PN}} \,.\end{aligned}$$ They are given in Eqs. (66–67) of Paper I. The hereditary contributions on the other hand depend on the entire dynamical past of the binary system. They split further into tail and memory parts: $$\begin{aligned} H^{\ell m}_{\textnormal{hered}}= H^{\ell m}_{\textnormal{tail}}+ H^{\ell m}_{\textnormal{mem}}\,.\end{aligned}$$ For the tails we find contributions at different orders relative to the leading order as: $$\begin{aligned} H^{\ell m}_{\textnormal{tail}}=&\; (H^{\ell m}_{\textnormal{tail}})_{\textnormal{{1.5}PN}} + (H^{\ell m}_{\textnormal{tail}})_{\textnormal{{2.5}PN}} + (H^{\ell m}_{\textnormal{tail}})_{\textnormal{{3}PN}} \,.\end{aligned}$$ These are given in Eqs. (47–48) of Paper I. There is both DC memory and oscillatory memory: $$\begin{aligned} H^{\ell m}_{\textnormal{mem}}= H^{\ell 0}_{\textnormal{DC}}+ H^{\ell m}_{{\textnormal{osc}}} \,.\end{aligned}$$ DC memory enters the waveform in the $m=0$ modes at all relative orders $$\begin{aligned} H^{\ell 0}_{\textnormal{DC}}=&\; (H^{\ell 0}_{\textnormal{DC}})_{\textnormal{Newt}}+ (H^{\ell 0}_{\textnormal{DC}})_{\textnormal{{1}PN}} + (H^{\ell 0}_{\textnormal{DC}})_{\textnormal{{1.5}PN}} \nonumber\\ &+ (H^{\ell 0}_{\textnormal{DC}})_{\textnormal{{2}PN}} + (H^{\ell 0}_{\textnormal{DC}})_{\textnormal{{2.5}PN}} + (H^{\ell 0}_{\textnormal{DC}})_{\textnormal{{3}PN}} \,,\end{aligned}$$ while slow and fast oscillatory memory enter as $$\begin{aligned} H^{\ell m}_{\textnormal{osc}}=&\; (H^{\ell m}_{\textnormal{slow}\,{\textnormal{osc}}})_{\textnormal{{1.5}PN}} + (H^{\ell m}_{\textnormal{slow} \,{\textnormal{osc}}})_{\textnormal{{2.5}PN}} \nonumber\\ &+ (H^{\ell m}_{\textnormal{slow} \,{\textnormal{osc}}})_{\textnormal{{3}PN}} + (H^{\ell m}_{\textnormal{fast} \,{\textnormal{osc}}})_{\textnormal{{2.5}PN}} \,.\end{aligned}$$ Slow oscillatory is due to the double periodic nature of eccentric motion and is not present in quasi-circular binary systems. All memory modes have been computed in this paper and are listed in \[sec:DCmemlist,sec:oscmemlist\]. As an example, we present here the dominant $H^{22}$ mode including all contributions to ${\mathcal{O}}(e)$: \[eq:full\_22-mode\] $$\begin{aligned} H^{22}_{\textnormal{Newt}}=&\; 1 + {\bar{e}}\bigg( \frac{1}{4} {\mathrm{e}}^{-{\mathrm{i}}\xi} + \frac{5}{4} {\mathrm{e}}^{{\mathrm{i}}\xi} \bigg) \,,\\ H^{22}_{\textnormal{{1}PN}} =&\; {\bar{x}}\Bigg\{ -\frac{107}{42} + \frac{55 \nu}{42} + {\bar{e}}\bigg[ {\mathrm{e}}^{-{\mathrm{i}}\xi} \left( -\frac{257}{168} + \frac{169 \nu}{168} \right) + {\mathrm{e}}^{{\mathrm{i}}\xi} \left( -\frac{31}{24} + \frac{35 \nu}{24} \right) \bigg] \Bigg\} \,,\\ H^{22}_{\textnormal{{1.5}PN}} =&\; {\bar{x}}^{3/2} \Bigg\{ 2 \pi + {\bar{e}}\bigg[ {\mathrm{e}}^{-{\mathrm{i}}\xi} \left( \frac{11 \pi }{4} + \frac{27 {\mathrm{i}}}{2} \ln \left( \frac{3}{2} \right) \right) + {\mathrm{e}}^{{\mathrm{i}}\xi} \left(\frac{13 \pi }{4} + \frac{3 {\mathrm{i}}}{2} \ln(2) \right) \bigg] \Bigg\} \,,\\ H^{22}_{\textnormal{{2}PN}} =&\; {\bar{x}}^2 \Bigg\{ -\frac{2173}{1512} - \frac{1069 \nu}{216} + \frac{2047 \nu^2}{1512} + {\bar{e}}\bigg[ {\mathrm{e}}^{{\mathrm{i}}\xi} \left( -\frac{2155}{252} - \frac{1655 \nu}{672} + \frac{371 \nu^2}{288} \right) \no &+ {\mathrm{e}}^{-{\mathrm{i}}\xi} \left( -\frac{4271}{756} - \frac{35131 \nu}{6048} + \frac{421 \nu^2}{864} \right) \bigg] \Bigg\} \,,\\ H^{22}_{\textnormal{{2.5}PN}} =&\; {\bar{x}}^{5/2} \Bigg\{ -\frac{107 \pi}{21} + \left( -24 {\mathrm{i}}+ \frac{34 \pi}{21} \right) \nu \no &+ {\bar{e}}\bigg[ {\mathrm{e}}^{{\mathrm{i}}\xi} \bigg( -\frac{9 {\mathrm{i}}}{2} + \frac{229 \pi}{168} + \left( -\frac{14579 {\mathrm{i}}}{140} + \frac{61 \pi}{42} \right) \nu + \left( \frac{473 {\mathrm{i}}}{28} - \frac{3 {\mathrm{i}}\nu }{7} \right) \ln (2) \bigg) \no &+ {\mathrm{e}}^{-{\mathrm{i}}\xi} \bigg( -\frac{27 {\mathrm{i}}}{2} -\frac{1081 \pi}{168} + \left( -\frac{1291 {\mathrm{i}}}{180} + \frac{137 \pi}{42} \right) \nu + \left( \frac{27 {\mathrm{i}}}{4} + 9 {\mathrm{i}}\nu \right) \ln \left( \frac{3}{2} \right) \bigg) \bigg] \Bigg\} \,,\\ H^{22}_{\textnormal{{3}PN}} =&\; {\bar{x}}^3 \Bigg\{ \frac{27027409}{646800} + \frac{428 {\mathrm{i}}\pi}{105} + \frac{2 \pi^2}{3} - \frac{856 {\gamma_\textnormal{E}}}{105} + \left( -\frac{278185}{33264} + \frac{41 \pi^2}{96} \right) \nu - \frac{20261 \nu^2}{2772} + \frac{114635 \nu^3}{99792} \no &- \frac{1712 \ln(2)}{105} - \frac{428 \ln({\bar{x}})}{105} \no &+ {\bar{e}}\bigg[ {\mathrm{e}}^{-{\mathrm{i}}\xi} \bigg( \frac{219775769}{1663200} + \frac{749 {\mathrm{i}}\pi}{60} + \frac{49 \pi^2}{24} - \frac{749 {\gamma_\textnormal{E}}}{30} + \left( -\frac{121717}{20790} - \frac{41 \pi^2}{192}\right) \nu - \frac{86531 \nu^2}{8316} - \frac{33331 \nu^3}{399168} \no &+ \left( -\frac{2889}{70} + \frac{81 {\mathrm{i}}\pi}{2}\right) \ln \left( \frac{3}{2} \right) - \frac{81}{2} \ln^2 \left( \frac{3}{2} \right) - \frac{749 \ln(2)}{15} - \frac{749 \ln({\bar{x}})}{60} \bigg) \no &+ {\mathrm{e}}^{{\mathrm{i}}\xi} \bigg( \frac{55608313}{1058400} + \frac{3103 {\mathrm{i}}\pi}{420} + \frac{29 \pi^2}{24} - \frac{3103 {\gamma_\textnormal{E}}}{210} + \left( -\frac{199855}{3024} + \frac{41 \pi^2}{48} \right) \nu -\frac{9967 \nu^2}{1008} + \frac{35579 \nu^3}{36288} \no &+ \left( -\frac{6527}{210} + \frac{3 {\mathrm{i}}\pi}{2}\right) \ln(2) + \frac{3 \ln^2(2)}{2} - \frac{3103 \ln({\bar{x}})}{420} \bigg) \bigg] \Bigg\} \,.\end{aligned}$$ Note here the difference at 2.5PN order between \[eq:full\_22-mode\] and Eq. (76) of Paper I, due to additional memory terms not yet considered in Paper I. Complete expressions for all modes to ${\mathcal{O}}(e^6)$ are given in the supplemental material [@supplement]. By taking the quasi-circular limit of our modes as described in Sec. V E of Paper I, we can compare the instantaneous, tail and (fast) oscillatory memory contributions of our waveform modes with [@blanchet-2008] and the DC memory terms with [@favata-2009]. In all of them we find perfect agreement. Brief summary {#sec:summary} ============= In this paper we computed the memory contribution to the gravitational waveform from non-spinning compact binaries in eccentric orbits at the third post-Newtonian order. Our results complete the previous work on the instantaneous parts [@mishra-2015] and on the tail and post-adiabatic contributions [@boetzel-2019]. These waveforms form the basis for construction of increasingly accurate GW templates from binary systems in eccentric orbits. There are two fundamentally different types of memory. DC memory is a slowly increasing, non-oscillatory contribution to the gravitational wave amplitude, entering at Newtonian order, leading to a difference in the amplitude between early and late times. Oscillatory memory on the other hand enters at higher PN orders as a normal periodic contribution. Due to the double periodic nature of the eccentric motion, slow oscillatory memory contributions on the periastron precession timescale are enhanced by a factor of 1PN, and thus already enter the waveform at 1.5PN order. This is unlike the quasi-circular case, where oscillatory memory only enters at 2.5PN order. We thank Marc Favata for an early review and useful comments. We thank Maria Haney and Achamveedu Gopakumar for insightful discussions and comments. M. E. and Y. B. are supported by the Swiss National Science Foundation. Y. B. is supported by a Forschungskredit of the University of Zurich, grant no. FK-18-084. Computation of the memory via the radiative mass multipoles {#sec:altmemory} =========================================================== The computation of the non-linear memory in the paper has been done effectively via the GW energy flux with the formula given in \[eq:Ulmmem\]. An alternative way is to directly compute the required moments of the memory contribution to the radiative mass multipole. The leading order memory piece of the mass quadrupole moment contributes at 2.5PN, however, due to the hereditary integral the DC terms are raised by 2.5PN orders such that they contribute at leading order in the waveform polarization. Reference [@faye-2015] lists the memory contributions up to 3.5PN. From this we are able to compute the DC memory to 1PN accuracy. The hereditary integral enhances the slow oscillatory memory terms by 1PN, therefore from knowing the 3.5PN contribution to the mass moments, we find the leading order 2.5PN terms contributing at 1.5PN and 2.5 PN in the waveform, the 3PN terms appear at 2PN and 3PN and the 3.5PN terms at 2.5PN. However, what we miss are the 4PN terms that appear in the waveform at 3PN level. On the other hand, the fast oscillatory memory is not affected by the hereditary integral in its PN order, we recover it at 2.5PN and 3PN. The required memory contributions at 3.5PN to the radiative mass moments are: \[eq:Umem\] $$\begin{aligned} U_{ij}^{\textnormal{mem}}(T_R) =&\; \frac{G}{c^5} \int_{-\infty}^{T_R} {\mathrm{d}}\tau\, \left[ -\frac{2}{7} M^{(3)}_{a \langle i}\left(\tau \right) M^{(3)}_{j \rangle a}(\tau) \right] \no & +\frac{G}{c^7} \int_{ -\infty}^{T_R} {\mathrm{d}}\tau \biggl[-\frac{5}{756} M^{(4)}_{a b}(\tau) M^{(4)}_{i j a b}(\tau) -\frac{32}{63} S^{(3)}_{a \langle i}(\tau) S^{(3)}_{j \rangle a}(\tau) \no & + \varepsilon_{a b \langle i} \left(\frac{5}{42} S^{(4)}_{j \rangle bc}(\tau) M^{(3)}_{ac}(\tau) -\frac{20}{189} M^{(4)}_{j \rangle bc}(\tau) S^{(3)}_{ac}(\tau) \right) \biggr]\,,\\ U_{ijk}^\mathrm{mem}(T_R) =&\; \frac{G}{c^5} \int_{-\infty}^{T_R} {\mathrm{d}}\tau \left[-\frac{1}{3} M^{(3)}_{a \langle i}(\tau) M^{(4)}_{jk \rangle a}(\tau) -\frac{4}{5} \varepsilon_{ab \langle i} M^{(3)}_{j a}(\tau) S^{(3)}_{k \rangle b}(\tau)\right]\,,\\ U_{ijkl}^\mathrm{mem}(T_R) =&\; \frac{G}{c^3} \int_{-\infty}^{T_R} {\mathrm{d}}\tau \left[\frac{2}{5} M^{(3)}_{\langle ij}(\tau) M^{(3)}_{kl \rangle}(\tau)\right] \no &+\frac{G}{c^5} \int_{-\infty}^{T_R} {\mathrm{d}}\tau \biggl[\frac{12}{55} M^{(4)}_{a \langle i}(\tau) M^{(4)}_{jkl \rangle a}(\tau) -\frac{14}{99} M^{(4)}_{a \langle ij}(\tau) M^{(4)}_{k l\rangle a}(\tau) + \frac{32}{45} S^{(3)}_{\langle ij}(\tau) S^{(3)}_{k l\rangle}(\tau) \no & + \varepsilon_{a b \langle i} \left(-\frac{4}{5} M^{(3)}_{ja}(\tau) S^{(4)}_{kl \rangle b}(\tau) +\frac{32}{45} S^{(3)}_{ja}(\tau) M^{(4)}_{kl \rangle b}(\tau) \right) \biggr]\,,\\ U_{ijklm}^\mathrm{mem}(T_R) =&\; \frac{G}{c^3} \int_{-\infty}^{T_R} {\mathrm{d}}\tau \left[\frac{20}{21} M^{(3)}_{\langle ij}(\tau) M^{(4)}_{klm \rangle}(\tau)\right] \,,\\ U_{ijklmn}^\mathrm{mem}(T_R) =&\; \frac{G}{c^3} \int_{-\infty}^{T_R} {\mathrm{d}}\tau \left[\frac{5}{7} M^{(4)}_{\langle ijk}(\tau) M^{(4)}_{lmn \rangle}(\tau) -\frac{15}{14} M^{(3)}_{\langle ij}(\tau) M^{(4)}_{klmn \rangle}(\tau)\right]\,.\end{aligned}$$ Note that the STF projection $\langle \dots \rangle$ only applies to the free indices $ijk...$. The integrand in those equations consists of products of canonical mass and current moments, $M^{(n)}_{L}(\tau)$ and $S^{(n)}_{L}(\tau)$, and the superscript in brackets stands for the $n^{\mathrm{th}}$ derivative with respect to $\tau$. The canonical moments are related by a gauge transformation to the source moments $I_L$ and $J_L$ along with some more gauge moments that enter at 2.5PN in the $\delta I_L,\; \delta J_L$ terms, $$\begin{aligned} M_L &= I_L + G \delta I_L + {\mathcal{O}}(G^2) \,,\\ S_L &= J_L + G \delta J_L + {\mathcal{O}}(G^2) \,.\end{aligned}$$ For our purpose of calculating the memory contribution to next to leading order, we only need the 1PN part of the source moments. Here we list the relevant source moments at 1PN for two non-spinning compact objects in general orbits [@mishra-2015]. The source moments are written in terms of $x_i$ and $v_i$, which denote the binaries relative separation and relative velocity. Moreover, $r$ is the distance between the two objects, thus $r = \|\boldsymbol{x}\|$ and $\dot{r}$ is the radial velocity. For the mass quadrupole moment we have $$\begin{aligned} I_{ij} &= \nu m \left[ A_1 x_{\langle i} x_{j \rangle} + A_2 \frac{r \dot{r}}{c^2} x_{\langle i} x_{j \rangle} + A_3 \frac{r^2}{c^2} v_{\langle i} v_{j \rangle} \right]\,,\end{aligned}$$ where, $$\begin{aligned} A_1 &= 1 + \frac{1}{c^2} \left[v^2\left(\frac{29}{42} -\frac{29\nu}{14}\right) +\frac{G\,m}{r} \left( -\frac{5}{7} +\frac{8\nu}{7}\right)\right]\,,\\ A_2 &= -\frac{4}{7}+ \frac{12 \nu}{7}\,,\\ A_3 &= \frac{11}{21} - \frac{11 \nu}{7}\,.\end{aligned}$$ The 1PN mass octupole is $$\begin{aligned} I_{ijk} &= - \nu m \Delta \left[B_1 x_{\langle ijk \rangle} +B_2 \frac{r \dot{r}}{c^2} x_{\langle ij} v_{k \rangle} + B_3 \frac{r^2}{c^2} x_{\langle i} v_{jk \rangle} \right]\,,\end{aligned}$$ where, $$\begin{aligned} B_1 &= 1+ \frac{1}{c^2} \left[v^2\left(\frac{5}{6} -\frac{19\nu}{6}\right) +\frac{G\,m}{r} \left( -\frac{5}{6} +\frac{13\nu}{6}\right)\right]\,,\\ B_2 &= - (1-2\nu)\,,\\ B_3 &= 1 - 2 \nu\,,\end{aligned}$$ and $\Delta = (m_1 - m_2) / m$ is the mass difference ratio. Moreover, we need also the leading Newtonian order part of the mass hexadecapole, $$\begin{aligned} I_{ijkl} = \nu m\, x_{\langle ijkl \rangle} \left(1 - 3 \nu \right)\,.\end{aligned}$$ From the current source moments we need the quadrupole, which is $$\begin{aligned} J_{ij} &= - \nu m \Delta \left[C_1 \varepsilon_{ab \langle i} x_{j \rangle a} v_{b} + C_2 \frac{r \dot{r}}{c^2} \varepsilon_{ab \langle i} v_{j \rangle b} x_{a} \right]\,,\end{aligned}$$ where, $$\begin{aligned} C_1 &= 1 + \frac{1}{c^2}\left[v^2 \left(\frac{13}{28}-\frac{17\nu}{7}\right)+ \frac{G \, m}{r} \left( \frac{27}{14} + \frac{15\nu}{7}\right)\right] \,,\\ C_2 &= \frac{5}{28}(1-2\nu) \,,\end{aligned}$$ and at last the leading order of the current octupole, which is $$\begin{aligned} J_{ijk} &= \nu m \, \varepsilon_{ab \langle i} x_{jk \rangle a}v_{b} (1-3\nu) \,.\end{aligned}$$ Having the source moments in hand, thereby in our case also the canonical moments, we can calculate the products of time derivatives of the canonical moments occurring in the integrands of \[eq:Umem\]. Before treating the hereditary integral, we transform from the STF-moments $U_L^{\textnormal{mem}}$ computed here to the scalar version of the radiative mass moments using Eq. (4) of Paper I. These are the same moments as we find when computing the memory with \[eq:Ulmmem\]. The hereditary integral is evaluated in the same way as described in \[sec:DCmemory,sec:oscmemory\]. Using this method, we find the 1PN DC memory and the 1PN oscillatory memory. Be aware that the DC memory appears in the waveform at leading Newtonian order while the first slow oscillatory memory terms appear at 1.5PN and the fast oscillatory memory at 2.5PN. This method of computing the memory contribution serves as a check. We can compare the relative 1PN pieces of the DC and the oscillatory memory calculated before and here, they are found to be in perfect agreement. Radiation-reaction evolution equations {#sec:evolveq} ====================================== In this Appendix we provide the secular 3PN accurate evolution equations for $x$ and $e$ [@arun-2008-1; @arun-2008-2; @arun-2009] in modified harmonic (MH) gauge. The instantaneous terms are exact whereas the eccentricity enhancement functions appearing in the hereditary contributions are given in an eccentricity expansion. We begin by listing the pieces needed for the evolution of $x$: $$\begin{aligned} \frac{{\mathrm{d}}x}{{\mathrm{d}}t} = \frac{2 c^3 \nu x^5}{3 G m} \left(\mathcal{X}_{\textnormal{Newt}}+ x \mathcal{X}_{\textnormal{{1}PN}} + x^2 \mathcal{X}_{\textnormal{{2}PN}} + x^3\mathcal{X}_{\textnormal{{3}PN}} +\mathcal{X}_{\textnormal{hered}}\right)\,,\end{aligned}$$ where $$\begin{aligned} \mathcal{X}_{\textnormal{Newt}}=&\; \frac{1}{\left(1-e^2\right)^{7/2}} \left\{ \frac{96}{5}+\frac{292 e^2}{5} +\frac{37 e^4}{5}\right\}\,,\\ \mathcal{X}_{\textnormal{{1}PN}} =&\; \frac{1}{\left(1-e^2\right)^{9/2}}\left\{ -\frac{1486}{35} -\frac{264 \nu}{5} +e^2 \left(\frac{2193}{7}-570 \nu \right) +e^4 \left(\frac{12217}{20}-\frac{5061 \nu}{10}\right) +e^6 \left(\frac{11717}{280}-\frac{148 \nu}{5}\right) \right\}\,,\\ \mathcal{X}_{\textnormal{{2}PN}}= &\; \frac{1}{\left(1-e^2\right)^{11/2}}\biggl\{ -\frac{11257}{945} +\frac{15677 \nu}{105} +\frac{944 \nu^2}{15} + e^2 \left(-\frac{2960801}{945} -\frac{2781 \nu}{5} + \frac{182387 \nu^2}{90}\right) \no &+e^4 \left(-\frac{68647}{1260}-\frac{1150631 \nu}{140}+\frac{396443 \nu^2}{72}\right) +e^6 \left(\frac{925073}{336}-\frac{199939 \nu}{48} +\frac{192943 \nu^2}{90}\right) \no &+e^8 \left(\frac{391457}{3360}-\frac{6037 \nu}{56}+\frac{2923 \nu^2}{45}\right) +\sqrt{1-e^2} \biggl[ 48-\frac{96 \nu}{5} +e^2 \left(2134 -\frac{4268 \nu}{5}\right) \no &+e^4 \left(2193-\frac{4386 \nu}{5}\right) +e^6 \left(\frac{175}{2}-35 \nu \right)\biggr] \biggr\}\,,\\ \mathcal{X}_{\textnormal{{3}PN}} =&\; \frac{1}{\left(1 -e^2\right)^{13/2}} \biggl\{ \frac{614389219}{148500} +\left(-\frac{57265081}{11340}+\frac{369 \pi^2}{2}\right) \nu -\frac{16073 \nu^2}{140}- \frac{1121 \nu^3}{27}\no &+e^2 \left(\frac{19769277811}{693000} +\left(\frac{66358561}{3240} +\frac{42571 \pi^2}{80}\right) \nu -\frac{3161701 \nu^2}{840} - \frac{1287385 \nu^3}{324}\right)\no &+e^4 \left(-\frac{3983966927}{8316000} + \left(\frac{6451690597}{90720} -\frac{12403 \pi^2}{64}\right) \nu + \frac{34877019 \nu^2}{1120}- \frac{33769597 \nu^3}{1296}\right)\no &+e^6 \left(-\frac{4548320963}{5544000} +\left(-\frac{59823689}{4032} -\frac{242563 \pi^2}{640}\right) \nu +\frac{411401857 \nu^2}{6720} -\frac{3200965 \nu^3}{108}\right)\no &+e^8 \left(\frac{19593451667}{2464000} + \left(-\frac{6614711}{480} -\frac{12177 \pi^2}{640}\right) \nu +\frac{92762 \nu^2}{7} -\frac{982645 \nu^3}{162}\right) \no &+e^{10} \left(\frac{33332681}{197120} - \frac{1874543 \nu}{10080} +\frac{109733 \nu^2}{840} -\frac{8288 \nu^3}{81} \right) \no &+\sqrt{1-e^2} \biggl[\left( -\frac{1425319}{1125} +\left(\frac{9874}{105} -\frac{41 \pi^2}{10}\right) \nu + \frac{632 \nu^2}{5}\right) \no &+e^2 \left(\frac{933454}{375} + \left(-\frac{2257181}{63} +\frac{45961 \pi^2}{240}\right) \nu +\frac{125278 \nu^2}{15} \right) \no &+e^4 \left(\frac{840635951}{21000} + \left(-\frac{4927789}{60} +\frac{6191 \pi^2}{32}\right)\nu + \frac{317273 \nu^2}{15}\right)\no &+e^6 \left(\frac{702667207}{31500} + \left(-\frac{6830419}{252}+\frac{287 \pi^2}{960}\right) \nu +\frac{232177 \nu^2}{30}\right) + e^8 \left(\frac{56403}{112} -\frac{427733 \nu}{840}+\frac{4739 \nu^2}{30} \right)\biggr] \no &+\log \left[\frac{x \left(1+ \sqrt{1-e^2}\right)}{x_0 \left(2 \left(1 -e^2\right)\right)}\right] \left(\frac{54784}{175} +\frac{465664 e^2}{105} +\frac{4426376 e^4}{525} +\frac{ 1498856 e^6}{525} + \frac{31779 e^8}{350}\right)\biggr\}\,,\\ \mathcal{X}_{\textnormal{hered}}=&\; \frac{96}{5} \biggl\{4 \pi x^{3/2} \varphi (e) +\pi x^{5/2} \left[ -\frac{4159}{672} \psi_\omega (e) - \frac{189}{8} \nu \, \zeta_\omega (e) \right] \no &+x^3 \left[-\frac{116761}{3675} \kappa(e) + \left(\frac{16 \pi^2}{3} -\frac{1712 {\gamma_\textnormal{E}}}{105} - \frac{1712}{105} \log \left(\frac{4 x^{3/2}}{x_0}\right)\right) F(e)\right]\biggr\}\,.\end{aligned}$$ The helper functions appearing in the hereditary contribution are given by $$\begin{aligned} \psi_\omega(e) =&\; \frac{1344}{4159} \frac{1}{\left(1 -e^2\right)^{3/2}} \left[\sqrt{1-e^2} \left(1-5 e^2\right) \varphi (e) -4 \tilde{\varphi}(e)\right] + \frac{8191}{4159} \psi(e)\,, \\ \zeta_\omega(e) =&\; \frac{583}{567} \zeta(e) - \frac{16}{567} \varphi (e)\,.\end{aligned}$$ The various enhancement functions appearing in these equations are listed below. Next we state the evolution equation for the eccentricity. Note that we observed errors in the 2PN and 3PN order expressions in Eqs. (C10) and (C11) of [@arun-2009]. These are likely due to the fact that only the relation between $e^\mathrm{MH}$ and $e^\mathrm{ADM}$ was inserted, but one also has to transform ${\mathrm{d}}e^\mathrm{ADM}/{\mathrm{d}}t$ to ${\mathrm{d}}e^\mathrm{MH} / {\mathrm{d}}t$. $$\begin{aligned} \frac{{\mathrm{d}}e}{{\mathrm{d}}t} = - \frac{c^3 \nu e x^4}{G m} \left(\mathcal{E}_\mathrm{N} + x \mathcal{E}_{\mathrm{1PN}} + x^2 \mathcal{E}_{\mathrm{2PN}} + x^3\mathcal{E}_{\mathrm{3PN}} + \mathcal{E}_\mathrm{hered} \right)\,,\end{aligned}$$ where $$\begin{aligned} \mathcal{E}_{\textnormal{Newt}}=&\; \frac{1}{\left(1-e^2\right)^{5/2}} \left\{ \frac{304}{15}+\frac{121 e^2}{15} \right\}\,,\\ \mathcal{E}_{\textnormal{{1}PN}} =&\; \frac{1}{\left(1-e^2\right)^{7/2}}\left\{ -\frac{939}{35} -\frac{4084 \nu}{45} +e^2 \left(\frac{29917}{105} - \frac{7753}{30} \nu \right) +e^4 \left(\frac{13929}{280} - \frac{1664 \nu}{45}\right) \right\}\,,\\ \mathcal{E}_{\textnormal{{2}PN}} =&\; \frac{1}{\left(1 -e^2\right)^{9/2}} \biggl\{ -\frac{949877}{1890} +\frac{18763 \nu}{42} +\frac{752 \nu^2}{5} + e^2 \left(-\frac{3082783}{2520} -\frac{988423 \nu}{840} +\frac{64433 \nu^2}{40} \right)\no &+e^4 \left(\frac{23289859}{15120} - \frac{13018711 \nu}{5040} +\frac{127411 \nu^2}{90} \right) +e^6\left(\frac{420727}{3360} -\frac{362071\nu}{2520} +\frac{821\nu^2}{9}\right)\no &+\sqrt{1-e^2} \biggl[ \frac{1336}{3} - \frac{2672 \nu}{15} +e^2 \left(\frac{2321}{2} -\frac{2321 \nu}{5}\right) +e^4 \left(\frac{565}{6} - \frac{113 \nu}{3}\right)\biggr] \biggr\}\,, \\ \mathcal{E}_{\textnormal{{3}PN}} =&\; \frac{1}{\left(1 -e^2\right)^{11/2}} \biggl\{ \frac{54208557619}{6237000} +\left(\frac{50099023}{113400} + \frac{779 \pi^2}{10}\right) \nu -\frac{4088921 \nu^2}{2520} - \frac{61001 \nu^3}{486} \no &+e^2 \left(\frac{46226320013}{6237000} + \left(\frac{28141879}{900} -\frac{139031 \pi^2}{960} \right) \nu -\frac{21283907 \nu^2}{3024} - \frac{86910509 \nu^3}{19440} \right) \no &+e^4 \left(-\frac{116987170177}{16632000} + \left(\frac{11499615139}{907200} -\frac{271871 \pi^2}{1920}\right) \nu +\frac{61093675 \nu^2}{4032}- \frac{2223241 \nu^3}{180} \right) \no &+e^6 \left(\frac{5891934893}{1232000} + \left( -\frac{5028323}{560} -\frac{6519 \pi^2}{640}\right) \nu +\frac{24757667 \nu^2}{2520} - \frac{11792069 \nu^3}{2430}\right) \no &+e^8 \left( \frac{302322169}{1774080} -\frac{1921387 \nu}{10080} + \frac{41179 \nu^2}{216} -\frac{193396 \nu^3}{1215} \right) \no &+\sqrt{1-e^2} \biggl[-\frac{22713049}{15750} +\left(-\frac{5526991}{945} +\frac{8323 \pi^2}{180}\right) \nu +\frac{54332 \nu^2}{45} \no &+e^2 \left( \frac{89395687}{7875} +\left( -\frac{38295557}{1260} +\frac{94177 \pi^2}{960}\right) \nu +\frac{681989 \nu^2}{90} \right) \no &+e^4 \left(\frac{5321445613}{378000} +\left( -\frac{26478311}{1512} +\frac{2501 \pi^2}{2880} \right) \nu +\frac{225106 \nu^2}{45}\right) \no &+e^6 \left(\frac{186961}{336} -\frac{289691 \nu}{504} +\frac{3197 \nu^2}{18} \right)\biggr] +\frac{730168}{23625\left(1 +\sqrt{1 - e^2}\right)} \no &+\frac{304}{15} \left(\frac{82283}{1995} +\frac{297674}{1995} e^2 +\frac{1147147}{15960} e^4 +\frac{61311}{21280} e^6\right) \ln \left[\frac{x \left(1 +\sqrt{1-e^2}\right)}{2 x_0 \left(1-e^2\right)} \right] \biggr\}\,,\\ \mathcal{E}_{\textnormal{hered}}=&\; -\frac{32}{5} \biggl\{-\frac{985}{48} \pi x^{3/2} \varphi_e(e) +\pi x^{5/2} \biggl[ \frac{55691}{1344} \psi_e(e) + \frac{19067}{126} \nu \zeta_e(e) \biggr] \no &+x^3 \biggl[ \left(\frac{89789209}{352800} -\frac{87419 \ln 2}{630} +\frac{78003 \ln 3}{560}\right) \kappa_e(e) \no &-\frac{769}{96} \left(\frac{16 \pi^2}{3} -\frac{1712 {\gamma_\textnormal{E}}}{105} -\frac{1712}{105} \ln \left(\frac{4 x^{3/2}}{x_0}\right)\right) F_e(e)\biggr] \biggr\}\,.\end{aligned}$$ The additional functions in the hereditary contribution are: $$\begin{aligned} \varphi_e(e) =&\; \frac{192}{985} \frac{\sqrt{1-e^2}}{e^2} \left[\sqrt{1-e^2} \varphi(e) -\tilde{\varphi}(e)\right]\,,\\ \psi_e(e) =&\; \frac{18816}{55691} \frac{1}{e^2 \sqrt{1-e^2}} \left[\sqrt{1-e^2} \left(1 -\frac{11 e^2}{7}\right) \varphi (e) -\left(1-\frac{3}{7} e^2\right) \tilde{\varphi}(e) \right] \no &+\frac{16382}{55691} \frac{\sqrt{1-e^2}}{e^2} \left[\sqrt{1-e^2} \psi (e) -\tilde{\psi}(e)\right]\,,\\ \zeta_e (e) =&\; \frac{924}{19067} \frac{1}{e^2 \sqrt{1-e^2}} \left[-\left(1-e^2\right)^{3/2} \varphi (e) + \left(1-\frac{5}{11} e^2\right) \tilde{\varphi}(e)\right] + \frac{12243}{76268} \frac{\sqrt{1-e^2}}{e^2} \left[ \sqrt{1-e^2} \zeta (e) -\tilde{\zeta}(e) \right]\,,\\ \kappa_e (e) =&\; \frac{\sqrt{1-e^2}}{e^2} \left[ \sqrt{1-e^2} \kappa (e) -\tilde{\kappa}(e)\right] \left(\frac{769}{96}-\frac{3059665}{700566} \ln 2 +\frac{8190315}{1868176} \ln 3\right)^{-1}\,,\\ F_e (e) =&\; \frac{96}{769} \frac{\sqrt{1-e^2}}{e^2} \left[\sqrt{1-e^2} F(e) - \tilde{F}(e)\right]\,.\end{aligned}$$ The eccentricity enhancement functions arise from hereditary contributions to the energy flux (non-tilde) and the angular momentum flux (tilde). Most of them do not admit closed forms and have to be computed numerically or in a small eccentricity expansion. Here we list them in an eccentricity expansion to ${\mathcal{O}}(e^6)$: $$\begin{aligned} \varphi(e) =&\; 1+\frac{2335}{192} e^2 +\frac{42955}{768} e^4 +\frac{6204647}{36864} e^6\,,\\ \tilde{\varphi}(e) =&\; 1+\frac{209}{32} e^2+\frac{2415}{128} e^4 +\frac{730751}{18432} e^6\,,\\ \psi(e) =&\; 1-\frac{22988}{8191} e^2 -\frac{36508643}{524224} e^4 -\frac{1741390565}{4718016} e^6 \,,\\ \tilde{\psi}(e) =&\; 1-\frac{17416}{8191} e^2 -\frac{14199197}{524224} e^4 -\frac{467169215}{4718016} e^6 \,,\\ \kappa(e) =&\; 1 +e^2 \left(\frac{62}{3} -\frac{4613840}{350283} \ln 2 +\frac{24570945}{1868176} \ln 3 \right) + e^4 \left(\frac{9177}{64} +\frac{271636085}{1401132} \ln 2 -\frac{466847955}{7472704} \ln 3 \right) \no &+e^6\left( \frac{76615}{128} -\frac{4553279605}{2802264} \ln 2 +\frac{14144674005}{119563264} \ln 3 +\frac{914306640625}{1076069376} \ln 5 \right) \,,\\ \tilde{\kappa}(e) =&\; 1 +e^2 \left(\frac{389}{32} - \frac{2056005}{233522} \ln 2 +\frac{8190315}{934088}\ln 3\right) +e^4 \left(\frac{3577}{64} + \frac{50149295}{467044} \ln 2 -\frac{155615985}{3736352} \ln 3\right) \no &+e^6 \left( \frac{43049}{256} -\frac{12561332945}{16813584} \ln 2 +\frac{4709431125}{59781632} \ln 3 +\frac{182861328125}{538034688} \ln 5 \right) \,,\\ \zeta(e) =&\; 1 +\frac{1011565}{48972} e^2 +\frac{106573021}{783552} e^4 +\frac{456977827}{854784} e^6 \,,\\ \tilde{\zeta}(e) =&\; 1 +\frac{102371}{8162} e^2 +\frac{14250725}{261184} e^4 + \frac{722230667}{4701312} e^6\,,\\ F(e) =&\; 1+ \frac{62}{3} e^2 +\frac{9177}{64} e^4 +\frac{76615}{128} e^6 \,,\\ \tilde{F}(e) =&\; 1 +\frac{389}{32} e^2 +\frac{3577}{64} e^4 +\frac{43049}{256} e^6 \,.\end{aligned}$$ By dividing the evolution equations for $x$ and $e$ and expanding in these variables, we can find a solution for the evolution of $x$ in terms of $e$ at each order as described in \[sec:DCmemory\]. Here we provide $x(e)$ at 3PN and to leading order in eccentricity: $$\begin{aligned} \label{eq:xofe3} x(e) = x_{\textnormal{Newt}}+ x_{\textnormal{{1}PN}} + x_{\textnormal{{1.5}PN}} + x_{\textnormal{{2}PN}} + x_{\textnormal{{2.5}PN}} + x_{\textnormal{{3}PN}} \,,\end{aligned}$$ where $$\begin{aligned} x_{\textnormal{Newt}}=&\; x_0 \left[ \left(\frac{e_0}{e}\right)^{12/19} \right]\,, \\ x_{\textnormal{{1}PN}} =&\; x_0^2 \left[ \left(\frac{e_0}{e}\right)^{24/19} \left(-\frac{2833}{3192} +\frac{197 \nu}{114} \right) +\left(\frac{e_0}{e}\right)^{12/19} \left( \frac{2833}{3192} -\frac{197 \nu}{114} \right) \right]\,, \\ x_{\textnormal{{1.5}PN}} =&\; x_0^{5/2} \left[ \frac{377 \pi}{228} \left(\frac{e_0}{e}\right)^{12/19} -\frac{377\pi}{228} \left(\frac{e_0}{e}\right)^{30/19} \right]\,, \\ x_{\textnormal{{2}PN}} =&\; x_0^3 \biggl[ \left(\frac{e_0}{e}\right)^{12/19} \left(-\frac{358353209}{366799104} +\frac{738407 \nu}{727776}+\frac{20597 \nu^2}{17328} \right) +\left( \frac{e_0}{e} \right)^{24/19} \left( -\frac{8025889}{5094432} +\frac{558101 \nu}{90972} -\frac{38809 \nu^2}{6498} \right) \no &+\left(\frac{e_0}{e} \right)^{36/19} \left( \frac{936217217}{366799104} -\frac{578135\nu}{80864} +\frac{248681 \nu^2}{51984} \right) \biggr]\,, \\ x_{\textnormal{{2.5}PN}}=&\; x_0^{7/2} \biggl[ \left(\frac{e_0}{e}\right)^{12/19} \left( -\frac{3763903 \pi}{7277760} -\frac{12788779 \pi \nu}{1819440} \right) +\left( \frac{e_0}{e} \right)^{24/19} \left( -\frac{1068041 \pi}{363888} +\frac{74269 \pi \nu}{12996}\right)\no & +\left( \frac{e_0}{e} \right)^{30/19} \left( -\frac{5340205 \pi}{1455552} + \frac{371345 \pi \nu}{51984} \right) +\left( \frac{e_0}{e} \right)^{42/19} \left( \frac{12956437 \pi}{1819440} -\frac{2651489 \pi \nu}{454860} \right) \biggr] \,,\\ x_{\textnormal{{3}PN}} =&\; x_0^4 \Biggl\{ \left(\frac{e_0}{e}\right)^{12/19} \biggl[ \frac{4942027570449143}{96592876047360} +\frac{81025 \pi^2}{103968} -\frac{3317 {\gamma_\textnormal{E}}}{399} -\frac{12091 \ln 2}{5985} -\frac{78003 \ln 3}{5320} -\frac{3317 \ln x_0}{798}\no &+\left(\frac{10309531979}{7466981760} -\frac{3977 \pi^2}{3648}\right) \nu -\frac{267351733 \nu^2}{82966464} -\frac{772583 \nu^3}{2222316} \biggr] +\left( \frac{e_0}{e} \right)^{30/19} \left( -\frac{710645 \pi^2}{103968} \right) \no &+\left( \frac{e_0}{e} \right)^{24/19} \biggl( \frac{605942457431}{585411369984} -\frac{3267214507\nu}{2986792704} -\frac{543796927 \nu^2}{82966464} +\frac{27463573 \nu^3}{2963088} \biggr) \no &+\left( \frac{e_0}{e} \right)^{36/19} \biggl( \frac{2652303375761}{390274246656} -\frac{449767537459 \nu}{13938365952} +\frac{2754579983\nu^2}{55310976} -\frac{48990157 \nu^3}{1975392} \biggr) \no &+\left( \frac{e_0}{e} \right)^{48/19}\biggl[-\frac{1628129474693173}{27597964584960} +\frac{157405\pi^2}{25992} +\frac{3317 {\gamma_\textnormal{E}}}{399} +\frac{12091 \ln 2}{5985} +\frac{78003 \ln 3}{5320} +\frac{6634}{2527}\ln \left(\frac{e_0}{e}\right) \no &+\frac{3317 \ln x_0}{798} +\left(\frac{6686551181963}{209075489280} +\frac{3977 \pi^2}{3648}\right) \nu -\frac{6641442629\nu^2}{165932928} +\frac{282310639 \nu^3}{17778528} \biggr] \Biggr\}\,.\end{aligned}$$ Oscillatory memory integral {#sec:oscintegral} =========================== Here we derive the formula to evaluate the oscillatory memory integrals in \[eq:oscmem\]. For convenience we set $G=c=1$ in this Appendix. We define the integral which has to be computed as $$\begin{aligned} \label{eq:Jmem} J_{\textnormal{mem}}&= \int_{-\infty}^{T_R} {\mathrm{d}}t\, x^p(t) \, e^q(t) \, {\mathrm{e}}^{{\mathrm{i}}(s \lambda_\xi + r \xi)}\,.\end{aligned}$$ We follow the approach of [@arun-2004], where this integral is evaluated in the case of circular orbits $(q=0)$. The eccentric orbit is assumed to evolve only with the secular radiation-reaction equations given in \[eq:peters-mathews\] starting from $x=0$ and $e=1$ in the remote past. Every astrophysical process like capture or mass loss possibly happening to the binary is ignored. We start by restating the evolution equation for $x$ at leading order in $x$ and $e$, $$\begin{aligned} \label{eq:dxdt} \frac{{\mathrm{d}}x(t)}{{\mathrm{d}}t} = \frac{64 \nu x^5(t)}{5 m}\left[1+ \frac{157}{24} e^2(t)\right]\,,\end{aligned}$$ and integrate it over a time interval up to some coalescence time $T_C$, where the orbital frequency and therefore $x$ tends to infinity: $$\begin{aligned} \int_{t}^{T_C} {\mathrm{d}}t = \int_{x(t)}^{\infty} \frac{{\mathrm{d}}x(t)}{({\mathrm{d}}x/{\mathrm{d}}t)}\,.\end{aligned}$$ Thereby we find an explicit relation between orbital frequency (related to $x$) and time $t$: $$\begin{aligned} \label{eq:tofx} T_C - t = \frac{5m}{256 \nu}\frac{1}{x^4(t)} \left[1 -\frac{157}{43} e^2(t) \right]\,.\end{aligned}$$ We can now invert the $x(e)$-relation derived in \[eq:xofe\] to find $e$ as a function of $x$. Considering only the leading order, we find $$\begin{aligned} \label{eq:eofx} e(t) = e(T_R) \left(\frac{x(T_R)}{x(t)} \right)^{19/12}.\end{aligned}$$ Using \[eq:tofx,eq:eofx\] we get $x$ as an explicit function of $t$: $$\begin{aligned} \label{eq:xoft} x(t) = \frac{1}{4} \left(\frac{5 m}{\nu (T_C-t)}\right)^{1/4} \left[1 - \frac{157}{172} \,e^2(T_R) \left(\frac{T_C -t}{T_C -T_R}\right)^{19/24} \right]\,.\end{aligned}$$ A quick check reveals that this expression indeed solves the differential equation in \[eq:dxdt\]. Since the memory integral runs up to current time $T_R$, we introduce a new integration variable $y$ which is better suited to the integration limits we have: $$\begin{aligned} y = \frac{T_R - t}{T_C - T_R}\,.\end{aligned}$$ Next we express the time dependent quantities in the integral in terms of $y$ and their values at current time $T_R$. For $x$ we find $$\begin{aligned} \label{eq:xy} x(y) = x(T_R) (1+y)^{-1/4} \left[1 - \frac{157}{172} \,e^2(T_R) \left( (1+y)^{19/24} -1\right) \right]\,,\end{aligned}$$ and for the eccentricity $$\begin{aligned} \label{eq:ey} e(y) = e(T_R) \left(1+ y\right)^{19/48}\,.\end{aligned}$$ Note that while going back in time, with increasing $y$, we let the eccentricity only evolve until $e = 1$ is reached. Furthermore we need the redefined mean anomaly $\xi(t)$ in terms of $y$ and its value at current time. Because $\xi$ is defined in terms of $\dot{\xi} = n$, we have to calculate the integral $$\begin{aligned} \xi(t) = \xi(T_C) - \int^t_{T_C} {\mathrm{d}}t' \, n(t') = \xi(T_C) - \frac{1}{m} \int^t_{T_C} {\mathrm{d}}t' \, x^{3/2}(t') = \xi(T_C) - \frac{(T_C - T_R)}{m} \int_{-1}^{y} {\mathrm{d}}y' \, x^{3/2}(y') \,.\end{aligned}$$ We can now evaluate the latter integral by inserting the expression for $x(y)$ given in \[eq:xy\]. This leads to $$\begin{aligned} \xi(t) = \xi(T_C) - \frac{8(T_C - T_R) x^{3/2}(T_R)}{5m}(1+y)^{5/8}\left[1 - \frac{471}{11696} e^2(T_R) \left(15(1+y)^{19/24} -34\right)\right] \,,\end{aligned}$$ where $\xi(T_C)$ is the value of $\xi$ at the moment of coalescence. Thus at current time $T_R$ the mean anomaly is given by $$\begin{aligned} \label{eq:lTR} \xi(T_R) =&\; \xi(T_C) - \frac{8 (T_C - T_R) x^{3/2}(T_R)}{5 m} \left[1 + \frac{8949}{11696} e^2(T_R)\right]\,. \end{aligned}$$ Now we are able to express $\xi(t)$ in terms of $\xi(T_R)$ and $y$ $$\begin{aligned} \label{eq:xit} \xi(t) = \xi(T_R) - \frac{8 (T_C - T_R) x^{3/2}(T_R)}{5 m}\left[(1+y)^{5/8}-1\right]\left[1 - \frac{471}{11696} e^2(T_R) \frac{19 -34 (1+y)^{5/8} +15(1+y)^{17/12}}{(1+y)^{5/8} -1}\right] \,,\end{aligned}$$ where $x(T_R)$ and $e(T_R)$ stand for the respective current value of $x$ and $e$. At this point, we introduce a dimensionless “adiabatic parameter" $\chi(T_R)$, which is connected with the inspiral rate at current retarded time $T_R$. We define it as the ratio between current period and time left till coalescence, $$\begin{aligned} \chi(T_R) = \frac{1}{n(T_R)(T_C - T_R)}\,,\end{aligned}$$ where $n(T_R) = x^{3/2}(T_R) / m$ at leading order. Explicitly in terms of $x(T_R)$ and $e(T_R)$ it reads $$\begin{aligned} \chi(T_R) = \frac{256 \nu}{5} x^{5/2}(T_R) \left[1 +\frac{157}{43}e^2(T_R)\right] \,.\end{aligned}$$ Inserting $\chi(T_R)$ into \[eq:xit\], we find: $$\begin{aligned} \label{eq:xiy} \xi(t) = \xi(T_R) - \frac{8}{5 \chi(T_R)} \left[(1+y)^{5/8} -1\right]\left[1 - \frac{471}{11696} e^2(T_R) \frac{19 -34 (1+y)^{5/8} +15(1+y)^{17/12}}{(1+y)^{5/8} -1}\right] \,.\end{aligned}$$ Now we put \[eq:xiy,eq:xy,eq:ey\] into the oscillatory integral and write it as an integral over $y$: $$\begin{aligned} J_{\textnormal{mem}}=&\; (T_C-T_R)\int_{0}^{\infty} {\mathrm{d}}y \, x^p(y) \, e^q(y) \, {\mathrm{e}}^{{\mathrm{i}}(s\lambda_\xi(y) +r \xi(y))}\no =&\;(T_C-T_R)\, {\mathrm{e}}^{{\mathrm{i}}(r+s(1+k)) \xi(T_R)} \int_{0}^{\infty} {\mathrm{d}}y \, x^p(y) \, e^q(y) \exp\biggl\{ -\frac{8 {\mathrm{i}}(r+s(1+k))}{5 \chi(T_R)} \left[(1+y)^{5/8} -1\right]\no &\times \left[1 -\frac{471}{11696} e^2(T_R) \frac{19 -34 (1+y)^{5/8} + 15(1+y)^{17/12}}{(1+y)^{5/8} -1}\right] \biggr\} \,.\end{aligned}$$ Let us look at the form of this integral: $$\begin{aligned} \label{eq:Jstruc} J_{\textnormal{mem}}\sim \int_{0}^{\infty} {\mathrm{d}}y \, f(y) \exp \left[ \frac{{\mathrm{i}}}{\chi(T_R)}g(y) \right]\,.\end{aligned}$$ The strategy is to integrate by parts, therefore we need to know the following type of integral: $$\begin{aligned} \int {\mathrm{d}}y \, {\mathrm{e}}^{{\mathrm{i}}\sigma g(y)} = -\frac{{\mathrm{i}}}{\sigma g'(y)} {\mathrm{e}}^{{\mathrm{i}}\sigma g(y)} + {\mathcal{O}}(g'(y)^{-2}) \,.\end{aligned}$$ This is formula is valid as long as $g'(y)$ is sufficiently large. Integrating \[eq:Jstruc\] by parts we get $$\begin{aligned} \label{eq:Jpart} J_{\textnormal{mem}}\sim f(y) \left[ -\frac{{\mathrm{i}}\chi(T_R)}{g'(y)} \exp \left[ \frac{{\mathrm{i}}}{\chi(T_R)}g(y)\right] \right]_{0}^{\infty} + {\mathrm{i}}\chi(T_R) \int_{0}^{\infty} {\mathrm{d}}y \, \frac{f'(y)}{g'(y)} \exp \left[ \frac{{\mathrm{i}}}{\chi(T_R)}g(y) \right] \,.\end{aligned}$$ As $y$ approaches infinity in the remote past, we notice that $f(y) = x^p(y) e^q(y)$ goes towards zero. This is because at early times the frequency reaches zero and the eccentricity cannot grow past $e=1$ in our model. Evaluating the first term at $y=0$ we recover $x$ and $e$ at current time and the exponential factor is just $1$ since $g(0) = 0$. The derivative $g'(y)$ in the denominator evaluated at $y=0$ is effectively one multiplied by some constants. What remains in the first term of \[eq:Jpart\] is therefore of order $\chi(T_R)$. Looking at the second term we find the same form of integral as in \[eq:Jstruc\]. Successively integrating by parts would each time yield another factor of $\chi(T_R)$. Since this parameter is already of order 2.5PN, the higher order $\chi(T_R)$ contributions can be safely ignored. Including everything of order $\chi(T_R)$, we find the following formula $$\begin{aligned} J_{\textnormal{mem}}&= -(T_c-T_R) \,x^p\, e^q \, {\mathrm{e}}^{{\mathrm{i}}(s\lambda_\xi +r \xi)} \frac{{\mathrm{i}}\chi(T_R)}{(r +s(1+k))} \,,\no &= - \frac{{\mathrm{i}}}{n(r +s(1+k))} x^p\, e^q \, {\mathrm{e}}^{{\mathrm{i}}(s\lambda_\xi +r \xi)} \,,\end{aligned}$$ which allows us to compute the oscillatory hereditary integrals in \[sec:oscmemory\]. List of DC memory modes {#sec:DCmemlist} ======================= Here we list the 3PN accurate DC memory contributions to the $h^{\ell m}$ modes at leading order in eccentricity in the following form, $$\begin{aligned} h^{\ell m}_{\textnormal{DC}}= \frac{8 G m \nu}{c^2 R} x \sqrt{\frac{\pi}{5}} H^{\ell m}_{\textnormal{DC}}\,,\end{aligned}$$ where $H^{\ell m}_{\textnormal{DC}}$ is a function of $x$ and $e$. The non-zero modes read: H\^[20]{}\_[[DC]{.nodecor}]{}=& - ( H\^[20]{}\_[[Newt]{.nodecor}]{}+ x H\^[20]{}\_[[[1]{}PN]{.nodecor}]{} + x\^[3/2]{} H\^[20]{}\_[[[1.5]{}PN]{.nodecor}]{} +x\^[2]{} H\^[20]{}\_[[[2]{}PN]{.nodecor}]{} +x\^[5/2]{} H\^[20]{}\_[[[2.5]{}PN]{.nodecor}]{} +x\^[3]{} H\^[20]{}\_[[[3]{}PN]{.nodecor}]{}) ,\ H\^[20]{}\_[[Newt]{.nodecor}]{}=& 1 -( )\^[12/19]{} ,\ H\^[20]{}\_1 =& - + +( )\^[12/19]{} ( - + ) +( )\^[24/19]{} ( -) ,\ H\^[20]{}\_[[[1.5]{}PN]{.nodecor}]{} =& - ( )\^[12/19]{} + ( )\^[30/19]{} ,\ H\^[20]{}\_2 =& - - + +()\^[12/19]{} ( - -) &+()\^[24/19]{} ( - +) &+()\^[36/19]{} (-+ -) ,\ H\^[20]{}\_[[[2.5]{}PN]{.nodecor}]{} =& - + +( )\^[12/19]{} ( +) +( )\^[24/19]{} ( - ) &+()\^[30/19]{} ( - ) +()\^[42/19]{} (- + ) ,\ H\^[20]{}\_[[[3]{}PN]{.nodecor}]{} =& - +( -) + + &+()\^[12/19]{} + ( )\^[30/19]{} &+()\^[24/19]{} (- + + -) &+()\^[36/19]{} (- + - +) &+()\^[48/19]{} .&& H\^[40]{}\_[[DC]{.nodecor}]{}=& - ( H\^[40]{}\_[[Newt]{.nodecor}]{}+ x H\^[40]{}\_1 + x\^[3/2]{} H\^[40]{}\_[[[1.5]{}PN]{.nodecor}]{} +x\^[2]{} H\^[40]{}\_2 +x\^[5/2]{} H\^[40]{}\_pn[2.5]{} +x\^[3]{} H\^[40]{}\_3 ) ,\ H\^[40]{}\_[[Newt]{.nodecor}]{}=& 1 -( )\^[12/19]{},\ H\^[40]{}\_1 =& - + + ( )\^[12/19]{} (-+ ) + ( )\^[24/19]{} (- ) ,\ H\^[40]{}\_[[[1.5]{}PN]{.nodecor}]{} =& - ()\^[12/19]{} + ()\^[30/19]{},\ H\^[40]{}\_2 =& - + + ()\^[12/19]{} ( - - ) &+ ()\^[24/19]{} ( - + ) &+( )\^[36/19]{} ( - + -),\ H\^[40]{}\_[[[2.5]{}PN]{.nodecor}]{} =& - + +( )\^[12/19]{} ( +) &+()\^[24/19]{} ( - ) +()\^[30/19]{} ( -) &+()\^[42/19]{} ( - +),\ H\^[40]{}\_3 =& +( - - ) - + &+()\^[12/19]{} + ( )\^[30/19]{} &+( )\^[24/19]{} ( - + + -) &+( )\^[36/19]{} ( - + - + ) &+( )\^[48/19]{} ,&& H\^[60]{}\_[[DC]{.nodecor}]{}=& (x H\^[60]{}\_1 +x\^[2]{} H\^[60]{}\_2 +x\^[5/2]{} H\^[60]{}\_[[[2.5]{}PN]{.nodecor}]{} +x\^[3]{} H\^[60]{}\_3 ) ,\ H\^[60]{}\_1 =& 1 - ,\ H\^[60]{}\_2 =& - + - +()\^[24/19]{} (- + -) &+()\^[36/19]{} ( - + ) ,\ H\^[60]{}\_[[[2.5]{}PN]{.nodecor}]{} =& - +( )\^[24/19]{} (- + ) +()\^[42/19]{} ( - ),\ H\^[60]{}\_3 =& - + - &+()\^[24/19]{} ( - - + ) &+( )\^[36/19]{} ( - + -) &+()\^[48/19]{} (- + - + ). && H\^[80]{}\_[[DC]{.nodecor}]{}=& - (x\^[2]{} H\^[80]{}\_2 +x\^[3]{} H\^[80]{}\_3) ,\ H\^[80]{}\_2 =& 1- + +( )\^[36/19]{} ( -1 + - ),\ H\^[80]{}\_3 =& - + - + &+()\^[36/19]{} (-+ - + ) &+()\^[48/19]{} \[ - + - ).&& H\^[100]{}\_[[DC]{.nodecor}]{}=& x\^3 .&& List of oscillatory memory modes {#sec:oscmemlist} ================================ Here we list the non-zero oscillatory memory contributions to the $h^{\ell m}$ modes at 3PN order and to quadratic order in eccentricity in the following way, $$\begin{aligned} h^{\ell m}_{\textnormal{osc}}= \frac{8 G m \nu}{c^2 R} x \sqrt{\frac{\pi}{5}} {\mathrm{e}}^{-{\mathrm{i}}m \psi} H^{\ell m}_{\textnormal{osc}}\,,\end{aligned}$$ where, $H^{\ell m}_{\textnormal{osc}}$ is a function of $x$, $e$ and the modified mean anomaly $\xi$. To improve readability in the odd $m$ expressions we make the definition $\Delta = (m_1 - m_2)/m = \sqrt{1 - 4\nu}$: $$\begin{aligned} H^{20}_{\textnormal{osc}}=&\; \frac{ \sqrt{6}\, {\mathrm{i}}\,x^{5/2} e \nu}{7} \biggl[-16 {\mathrm{e}}^{-{\mathrm{i}}\xi} +16 {\mathrm{e}}^{{\mathrm{i}}\xi} -\frac{647}{36} e {\mathrm{e}}^{-2 {\mathrm{i}}\xi} +\frac{647}{36} e {\mathrm{e}}^{2 {\mathrm{i}}\xi} \biggr]\,,\\ H^{22}_{\textnormal{osc}}=&\; {\mathrm{i}}\, x^{3/2} \, e^2 \nu \, {\mathrm{e}}^{2{\mathrm{i}}\xi} \left[ -\frac{13}{252} + \left(\frac{697}{336} -\frac{865 \nu}{216} \right)x -\frac{29\pi}{126} x^{3/2}\right] \no &+ \frac{{\mathrm{i}}x^{5/2} e \nu}{21} \left[ \frac{19}{6}e +\frac{4}{3} {\mathrm{e}}^{-{\mathrm{i}}\xi} - 4 {\mathrm{e}}^{{\mathrm{i}}\xi} +\frac{65}{24} e {\mathrm{e}}^{-2{\mathrm{i}}\xi}\right] \,,\\ H^{31}_{\textnormal{osc}}=&\; \frac{\sqrt{14} x^2 \nu \Delta}{90}\biggl\{ \frac{44}{3} e^2 -\frac{44}{3} e\, {\mathrm{e}}^{{\mathrm{i}}\xi} -\frac{44}{3} e^2 {\mathrm{e}}^{2 {\mathrm{i}}\xi} +x \biggl[-\frac{121}{7} -\frac{43}{2} e {\mathrm{e}}^{-{\mathrm{i}}\xi } -\frac{2987}{84} e^2 {\mathrm{e}}^{-2 {\mathrm{i}}\xi}\no &+e {\mathrm{e}}^{{\mathrm{i}}\xi} \left(\frac{19801}{264} -\frac{2521 \nu}{231}\right) +e^2 {\mathrm{e}}^{2 {\mathrm{i}}\xi} \left(\frac{7957}{88} -\frac{827 \nu}{231}\right) +e^2 \left(- \frac{111821}{616} + \frac{827 \nu}{231} \right)\biggl]\biggr\}\,,\\ H^{33}_{\textnormal{osc}}=&\; \frac{x^3 \nu \Delta}{6 \sqrt{210}} \left[ \frac{22}{9} +19 e^2 +11 e\, {\mathrm{e}}^{-{\mathrm{i}}\xi} +\frac{1}{3} e\, {\mathrm{e}}^{{\mathrm{i}}\xi}+\frac{713}{30} e^2 {\mathrm{e}}^{-2 {\mathrm{i}}\xi} -\frac{119}{6} e^2 {\mathrm{e}}^{2 {\mathrm{i}}\xi}\right]\,,\\ H^{40}_{\textnormal{osc}}=&\; \frac{\sqrt{2} {\mathrm{i}}\,x^{5/2} e \nu}{210} \biggl[-8 {\mathrm{e}}^{-{\mathrm{i}}\xi} +8 {\mathrm{e}}^{{\mathrm{i}}\xi} -\frac{143}{16} e\, {\mathrm{e}}^{-2 {\mathrm{i}}\xi} +\frac{143}{16} e \,{\mathrm{e}}^{2 {\mathrm{i}}\xi}\biggr]\,,\\ H^{42}_{\textnormal{osc}}=&\; \frac{{\mathrm{i}}\,x^{3/2} e^2 \nu {\mathrm{e}}^{2{\mathrm{i}}\xi}}{216 \sqrt{5}} \left[-\frac{13}{14} +x \left(\frac{7943}{56} -\frac{25393 \nu}{66} \right) -\frac{29}{7} \pi x^{3/2}\right]\no &+\frac{{\mathrm{i}}\, x^{5/2} e \nu}{126 \sqrt{5}} \left[\frac{19}{12} e +\frac{2}{3} {\mathrm{e}}^{-{\mathrm{i}}\xi} -2 {\mathrm{e}}^{{\mathrm{i}}\xi} +\frac{65}{48} e \,{\mathrm{e}}^{-2 {\mathrm{i}}\xi}\right] \,,\\ H^{44}_{\textnormal{osc}}=&\; \frac{{\mathrm{i}}\, x^{5/2} \nu}{6 \sqrt{35}} \biggl[\frac{2}{3} +\frac{331 e^2}{240} + \frac{14}{15} e\, {\mathrm{e}}^{-{\mathrm{i}}\xi} +2 e\, {\mathrm{e}}^{{\mathrm{i}}\xi} +\frac{1037}{720} e^2 {\mathrm{e}}^{-2 {\mathrm{i}}\xi} +\frac{217}{48} e^2 {\mathrm{e}}^{2 {\mathrm{i}}\xi} \biggr]\,,\\ H^{51}_{\textnormal{osc}}=&\; \frac{x^2 \nu \Delta}{18 \sqrt{385}} \biggl\{ \frac{43 e^2}{12} - \frac{43}{12} e \,{\mathrm{e}}^{{\mathrm{i}}\xi} -\frac{43}{12} e^2 {\mathrm{e}}^{2 {\mathrm{i}}\xi} +x \biggl[ -\frac{26}{7} -\frac{251}{56} e \,{\mathrm{e}}^{-{\mathrm{i}}\xi} -\frac{1199}{168} e^2 {\mathrm{e}}^{-2 {\mathrm{i}}\xi} \no &+ e\, {\mathrm{e}}^{{\mathrm{i}}\xi} \left(\frac{8627}{156} -\frac{41807 \nu}{312}\right) +e^2 {\mathrm{e}}^{2 {\mathrm{i}}\xi} \left(\frac{785}{13} -\frac{5156 \nu}{39}\right) +e^2 \left(-\frac{8321}{104} +\frac{5156 \nu}{39}\right) \biggr] \biggr\} \,,\\ H^{53}_{\textnormal{osc}}=&\; \frac{x^3 \nu \Delta}{2\sqrt{330}} \biggl[-\frac{2}{189} +\frac{27 e^2}{112} + \frac{41}{336} e\, {\mathrm{e}}^{-{\mathrm{i}}\xi} -\frac{67}{504} e {\mathrm{e}}^{{\mathrm{i}}\xi} +\frac{1531 e^2 {\mathrm{e}}^{-2 {\mathrm{i}}\xi}}{5040} -\frac{47}{72} e^2 {\mathrm{e}}^{2 {\mathrm{i}}\xi} \biggr]\,,\\ H^{55}_{\textnormal{osc}}=&\; \frac{x^3 \nu \Delta}{14 \sqrt{66}} \biggl[\frac{18}{5} +\frac{8909 e^2}{720} + \frac{457}{72} e\, {\mathrm{e}}^{-{\mathrm{i}}\xi} +\frac{197}{16} e\, {\mathrm{e}}^{{\mathrm{i}}\xi} +\frac{787}{72} e^2 {\mathrm{e}}^{-2 {\mathrm{i}}\xi} +\frac{4369}{144} e^2 {\mathrm{e}}^{2 {\mathrm{i}}\xi}\biggr]\,,\\ H^{62}_{\textnormal{osc}}=&\; \frac{{\mathrm{i}}\,x^{5/2} e^2 \nu \, {\mathrm{e}}^{2 {\mathrm{i}}\xi}}{352 \sqrt{65}} \biggl[\frac{2783}{168} - 53 \nu \biggr]\,,\\ H^{71}_{\textnormal{osc}}=&\; \frac{5x^3 \nu e \Delta}{30888 \sqrt{2}} \biggl[ {\mathrm{e}}^{{\mathrm{i}}\xi} \left( \frac{5023}{168} - 97 \nu \right) +e\, {\mathrm{e}}^{2 {\mathrm{i}}\xi } \left(\frac{5023}{168} -97 \nu \right) +e \left(-\frac{5023}{168} +97 \nu \right)\biggr] \,.\end{aligned}$$	{ "pile_set_name": "ArXiv" }
--- author: - 'Alex Garivaltis[^1]' title: '[Cover’s Rebalancing Option with Discrete Hindsight Optimization]{}' --- JEL Classification: C44, D80, D81, G11 Introduction ============ The main alternative to the Markowitz (1952) mean-variance theory of portfolio selection was popularized by Kelly (1956) who sought to optimize a gambler’s asymptotic continuously-compounded capital growth rate in repeated bets on horse races in the presence of partial inside information. His reasoning is in fact applicable to all gambling, insurance, and investment problems. Rather than optimize the static reward per unit of risk, the Kelly Criterion (Poundstone 2010) is equivalent to the prescription that one should act each round so as to maximize the expected log of his capital. Breiman (1961) showed that the Kelly Criterion constitutes asymptotically dominant behavior: a Kelly gambler will almost surely beat any other gambler in the long run by an exponential factor, and he has the shortest expected hitting time for a distant wealth goal. With probability approaching 1 as time goes on, the Kelly gambler’s bankroll will (amusingly) overtake that of a mean-variance investor, who has a smooth ride but ultimately cannot “eat his Sharpe ratio.” The books by Cover and Thomas (2006) and Luenberger (1998) are excellent primers of the theory of asymptotic capital growth in discrete and continuous time, respectively. Thorp (cf. his 2017 biography) demonstrated the practical effectiveness of the Kelly Criterion when he used it to size his Blackjack bets in certain favorable situations that are identifiable via his trademark (1966) theory of card counting. In this connection, the correct behavior is to bet the fraction $b^:=p-q$ of your net worth on a given hand for which $p$ is the chance of winning and $q$ is the chance of losing. For growth opportunities in the stock market, the analog of Kelly’s fixed fraction betting scheme is a certain constant-rebalanced portfolio $b^$ that trades continuously so as to maintain a target growth-optimal fraction of wealth in each risk asset. For instance, rather than bet $b:=2\%$ of wealth on a (favorable) hand of Blackjack, one could bet $2\%$ of wealth (or even $b:=200\%$ of wealth) on the S&P $500$ index. In theory, if stock market returns are iid across (discrete) time then one can calculate the corresponding log-optimal portfolio directly from the return distribution. But in practice, equity investors must get along without complete knowledge of the return distribution. Thus, a real-world investor cannot measure the exact regret of his portfolio relative to the Kelly bet for the simple reason that he does not know the Kelly bet. The way out of this conundrum was discovered by information theorist Thomas Cover (1938-2012), who formulated the individual sequence approach to investment. For a given observed sequence of asset prices, one can look back and determine which constant-rebalanced asset allocation would have yielded the greatest final wealth for that particular sequence. By definition, a Kelly gambler (who knows the distribution of returns but not the individual sequence that will occur in the future) will achieve a final wealth that is no greater than that of the best constant-rebalanced portfolio determined in hindsight for the actual sequence of returns. Thus began Cover’s important universal portfolio theory that formulated various on-line investment schemes (1986, 1991, 1996, 1998) that guarantee to achieve a high percentage of the final wealth of the best constant unlevered rebalancing rule (of any kind) in hindsight. Of course, any such scheme would then also guarantee to achieve a high percentage of the Kelly final wealth in iid stock markets. Contribution ------------ One can consider Cover’s performance benchmark to be a financial derivative (“Cover’s rebalancing option”) whose final payoff is equal to the wealth that would have accrued to a $\$1$ deposit into the best rebalancing rule (or fixed-fraction betting scheme) determined in hindsight. Ordentlich and Cover (1998) began the work of pricing this option in the Black-Scholes (1973) market at time-$0$ for unlevered hindsight optimization over a single underlying risk asset. Garivaltis (2018) priced and replicated the rebalancing option at any time $t$ for levered hindsight optimization over an arbitrary number of correlated stocks in geometric Brownian motion. That paper obtained the elegant result that for completely relaxed (levered) hindsight optimization, the corresponding delta-hedging strategy simply looks back over the observed price history $[0,t]$, computes the best rebalancing rule in hindsight $b(S_t,t)$, and bets that fraction of wealth over the next differential time step $[t,t+dt]$. The present paper studies Cover’s rebalancing option with hindsight optimization over a discrete set $\mathbb{B}:=\{b_1,...,b_n\}$ of rebalancing rules. Apart from the scientific obligation to extend Ordentlich and Cover’s incisive (1998) chain of reasoning, our approach has some interesting advantages relative to hindsight-optimization over all possible rebalancing rules. In our world, the (delta-hedging) practitioner is now free to express any of his institutional constraints or beliefs about future returns through a judicious choice of the set $\mathbb{B}$. Our newly austere mode of hindsight optimization yields a rock-bottom option price and correspondingly better guarantees of relative performance at the end of the planning period, whose shortened length is now well within a human life span. Say, for robust betting on the S&P 500 index, the author himself is inclined to use $\mathbb{B}:=\{0,0.5,1,1.5,2\}$, which amounts to the following five (continuously-rebalanced) asset allocations: 1. $0\%$ stocks, $100\%$ cash 2. $50\%$ stocks, $50\%$ cash 3. $100\%$ stocks, $0\%$ cash 4. $150\%$ stocks, $-50\%$ cash (margin loans) 5. $200\%$ stocks, $-100\%$ cash (margin loans) In this example, the author would like to avoid paying the full Cost of Achieving the Best \[Rebalancing Rule\] in Hindsight that would correspond (Garivaltis 2018) to $\mathbb{B}:=\mathbb{R}$ or even $\mathbb{B}:=[0,2]$. The paper is organized as follows. Section \[notation\] explains our basic notation and terminology. Section \[two\] develops our main techniques in the context of hindsight optimization over a pair $b>c$ of rebalancing rules and a single underlying risk asset. We price and replicate both the horizon-$T$ and perpetual versions of the rebalancing option, and give performance simulations that illustrate the general behavior of the replicating strategy. Section \[several\] extends the methodology to general discrete sets of asset allocations. We show how the rebalancing option can be interpreted as a certain portfolio of Margrabe-Fischer (1978) exchange options, and derive the general replicating strategy, which is a time- and state-varying convex combination of the $b_i$. We close the paper by proving that American-style rebalancing options (with general exercise price $K$) are always “worth more alive than dead” in equilibrium. Definitions and Notation {#notation} ======================== We start in the Black-Scholes (1973) market with a single underlying stock whose price $S_t$ follows the geometric Brownian motion $$\frac{dS_t}{S_t}=\mu\,dt+\sigma dW_t,$$where $\mu$ is the drift, $\sigma$ is the volatility, and $W_t$ is a standard Brownian motion. There is a risk-free bond whose price $B_t:=e^{rt}$ follows $$\frac{dB_t}{B_t}=r\,dt.$$A constant rebalancing rule $b\in(-\infty,+\infty)$ is a fixed-fraction betting scheme that continuously maintains the fraction $b$ of wealth in the stock and the fraction $1-b$ of wealth in bonds. We let $V_t(b)$ denote the wealth at $t$ that accrues to a $\$1$ deposit into the rebalancing rule $b$. Thus, the trader holds $\Delta:=bV_t(b)/S_t$ shares of the stock at time $t$, and his remaining $(1-b)V_t$ dollars are invested in bonds. Maintenance of the target asset allocation generally requires continuous trading. If $0<b<1$, the trader must sell a precise number of shares on every uptick (more precisely, whenever $dS_t/S_t\geq r\,dt$) to restore the target allocation. Similarly, when the risk asset underperforms cash over $[t,t+dt]$ (i.e. when $dS_t/S_t\leq r\,dt$) the trader must buy additional shares to restore the balance. This amounts to a volatility harvesting scheme (cf. Luenberger 1998) that “lives off the fluctuations” of the underlying. For $b=1$ the trader just buys the stock and holds it; for $b>1$ he carries a margin (debit) balance of $(b-1)V_t(b)$ dollars at time $t$. A levered rebalancing rule $b>1$ must continuously maintain a fixed debt-to-assets ratio of $1-1/b$. Thus, when the stock rises (and debt is now a smaller percentage of assets) the trader will borrow against his new wealth to buy additional shares. Similarly, when the stock falls he must sell some shares to reduce the loan-to-value ratio. This “buy high, sell low” strategy is only appropriate for stocks with relatively high drift and low volatility. Finally, for low quality underlyings one can hold all cash ($b=0$) or a continuously-rebalanced short position $b<0$. We now imagine a trader who starts with $\$1$ and has two favored rebalancing rules $b>c$, who wants to perform well relative to the best of $\mathbb{B}:=\{b,c\}$ in hindsight. Accordingly, we create for him the financial derivative whose final payoff at $T$ is $$\boxed{V_T^:=\max\{V_T(b),V_T(c)\}}.$$ Ordentlich and Cover (1998) investigated the best unlevered rebalancing rule in hindsight, with payoff $V_T^:=\max\limits_{0\leq b\leq1}{V_T(b)}$. They found the time-0 price of this contingent claim to be $$C_0=1+\sigma\sqrt{\frac{T}{2\pi}}.$$The owner of this rebalancing option (cf. Garivaltis 2018) will compound his money at the same asymptotic rate as the best unlevered rebalancing rule in hindsight. Indeed, the final excess continuously-compounded growth rate of the best rebalancing rule in hindsight over that of the replicating strategy is $\log\big\{1+\sigma\sqrt{T/(2\pi)}\big\}/T$, which tends to 0 as $T\to\infty$. This growth rate spread obtains deterministically, regardless of the realized price path $(S_t)_{0\leq t\leq T}$. Garivaltis (2018) extended the Ordentlich-Cover (1998) analysis by computing the general time-$t$ price $C(S,t)$ of Cover’s rebalancing option for both levered and unlevered hindsight optimization. For levered hindsight optimization (with payoff $V_T^:=\max\limits_{b\in\mathbb{R}}{V_T(b)}$), Garivaltis (2018) found the general pricing formula $$C(S,t)=\sqrt{\frac{T}{t}}\exp\{rt+z_t^2/2\},$$ where $$z_t:=\frac{\log(S_t/S_0)-(r-\sigma^2/2)t}{\sigma\sqrt{t}}$$is an auxiliary variable that is distributed unit normal with respect to the equivalent martingale measure $\mathbb{Q}$. More generally, for a Black-Scholes market with $d$ correlated stocks in geometric Brownian motion, Garivaltis (2018) found that $$C(S,t)=\bigg(\frac{T}{t}\bigg)^{d/2}\cdot\exp\{rt+z_t'R^{-1}z_t/2\},$$where $R:=[\rho_{ij}]_{d\times d}$ is the correlation matrix of instantaneous returns, $$z_{it}:=\frac{\log(S_{it}/S_{i0})-(r-\sigma_i^2/2)t}{\sigma_i\sqrt{t}},$$are auxiliary variables, and $\sigma_i$ is the volatility of stock $i$. When we relax the hindsight optimization to include all levered rebalancing rules $b\in\mathbb{R}^d$, replication becomes especially simple. At time $t$, one just looks back at the observed price history $[0,t]$, finds the best ($d$-dimensional) rebalancing rule $b(S,t)$ in hindsight, and bets the fraction $b_i(S,t)$ of wealth on stock $i$ over $[t,t+dt]$. The relation $C(S,t;T)\propto T^{d/2}$ matches the model-independent $\mathcal{O}(T^{d/2})$ super-replicating price calculated by Cover $\&$ Company. In what follows, we work toward reducing the option price $\sqrt{T/t}\cdot\exp\{rt+z_t^2/2\}$ by replacing $\mathbb{B}=\mathbb{R}$ with $\mathbb{B}:=\{b,c\}$. In order to get the payoff $\max\{V_T(b),V_T(c)\}$ into a more practical form, we note that $V_t(b)$ is a geometric Brownian motion, since $$\label{eq:wealth} \frac{dV_t(b)}{V_t(b)}=b\frac{dS_t}{S_t}+(1-b)\frac{dB_t}{B_t}=[r+(\mu-r)b]dt+b\sigma dW_t.$$ Solving this stochastic differential equation, we obtain (cf. Wilmott 1998, 2001) $$V_t(b)=\exp\{[r+(\mu-r)b-\sigma^2b^2/2]t+b\sigma W_t\}.$$ In order to get the payoff in terms of the observable variable $S_t$ (rather than the Wiener process $W_t$), we start with the equation $$S_t=S_0\exp\{(\mu-\sigma^2/2)t+\sigma W_t\},$$ and solve for $\sigma W_t$ in terms of $S_t$. Substituting the resulting expression into \[eq:wealth\], we get $$V_t(b)=\exp\{(r-\sigma^2b^2/2)t+b[\log(S_t/S_0)-(r-\sigma^2/2)t]\}.$$ We thus have $$\label{fortune} \boxed{V_t(b)=\exp\{(r-\sigma^2b^2/2)t+b\sigma\sqrt{t}\cdot z_t\}},$$where $$\label{auxiliary} \boxed{z_t:=\frac{\log(S_t/S_0)-(r-\sigma^2/2)t}{\sigma\sqrt{t}}}$$is distributed unit normal with respect to the equivalent martingale measure $\mathbb{Q}$. Note that the drift $\mu$ (which is difficult to estimate) does not appear in this formula. The final wealth of the rebalancing rule $b$ is now expressed solely in terms of $z_t$, the risk-free rate $r$, the time $t$, and the volatility $\sigma$, which is easily estimated from high-frequency price data. The Best of Two Asset Allocations in Hindsight {#two} ============================================== Before we can price the rebalancing option with payoff $\max\{V_T(b),V_T(c)\}$, we must characterize the random outcomes under which $b$ will turn out to outperform $c$ over the interval $[0,t]$. Accordingly, we compare the exponents of $V_t(b)$ and $V_t(c)$ to obtain For two given rebalancing rules $b>c$, b outperforms $c$ over $[0,t]$ if and only if $$\boxed{z_t\geq\frac{b+c}{2}\sigma\sqrt{t}}.$$ The best rebalancing rule (of any kind) in hindsight over $[0,t]$, denoted $b(S,t)$, is $$\boxed{b(S,t):=\arg\max_{b\in\mathbb{R}}V_t(b)=\frac{z(S,t)}{\sigma\sqrt{t}}}.$$Given any closed set $\mathbb{B}$ of rebalancing rules, the best performer in hindsight is the $b\in\mathbb{B}$ that is nearest to $b(S,t)=z(S,t)/(\sigma\sqrt{t})$. We compute the abscissa of vertex of the parabola $b\mapsto\log\,V_t(b)$. This yields $$b(S,t)=\arg\max_{b\in\mathbb{R}}\log V_t(b)=\frac{-\sigma\sqrt{t}\cdot z_t}{2(-\sigma^2t/2)}=\frac{z_t}{\sigma\sqrt{t}}.$$Because the graph of a parabola is symmetric about its vertex, the $b\in\mathbb{B}$ that maximizes the height of this parabola is whichever element of $\mathbb{B}$ is nearest to the vertex $b(S,t)$. We proceed to compute the cost of achieving the best of two rebalancing rules in hindsight, by finding the expected present value of $\max\{V_T(b),V_T(c)\}$ at time-0 with respect to the equivalent martingale measure $\mathbb{Q}$. This cost is the sum of two integrals $I_1+I_2$, where $$I_1:=\frac{\exp(-\sigma^2b^2T/2)}{\sqrt{2\pi}}\int_{\frac{b+c}{2}\sigma\sqrt{T}}^\infty\exp(-z^2/2+b\sigma\sqrt{T}\cdot z)dz$$and $$I_2:=\frac{\exp(-\sigma^2c^2T/2)}{\sqrt{2\pi}}\int_{-\infty}^{\frac{b+c}{2}\sigma\sqrt{T}}\exp(-z^2/2+c\sigma\sqrt{T}\cdot z)dz.$$ In the sequel, we will often use the following general formula (i.e. the appendix to Reiner and Rubinstein 1992): $$\boxed{ \int_{A}^Be^{-\alpha y^2+\beta y}dy=\sqrt{\frac{\pi}{\alpha}}\exp\bigg(\frac{\beta^2}{4\alpha}\bigg)\bigg[N\bigg(B\sqrt{2\alpha}-\frac{\beta}{\sqrt{2\alpha}}\bigg)-N\bigg(A\sqrt{2\alpha}-\frac{\beta}{\sqrt{2\alpha}}\bigg)\bigg]},$$where $\alpha>0$ and $N(\bullet)$ is the cumulative normal distribution function. Simplifying the two integrals, we get $$I_1=I_2=N\bigg(\frac{b-c}{2}\sigma\sqrt{T}\bigg).$$ The time-$0$ cost of achieving the best of two rebalancing rules $\{b,c\}$ in hindsight is $$\boxed{C_0(\delta,\sigma,T)=2N\bigg(\frac{\delta}{2}\sigma\sqrt{T}\bigg)}.$$where $\delta:=\|b-c\|$ is the distance between the two rebalancing rules. The equilibrium price at $t=0$ of a perpetual option ($T:=\infty$) on the best of two rebalancing rules $\{b,c\}$ in hindsight is $C_0(\delta,\sigma,\infty)=\$2$. Note that the horizon-$T$ price is independent of the interest rate $r$, and it is translation invariant in the sense that it depends only on the distance $\delta=\|b-c\|$ between the two rebalancing rules. We always have $1\leq C_0(\delta,\sigma,T)\leq2$; besides the perpetual version of the option, the maximum $\$2$ price also obtains if $\sigma=\infty$ or $\delta=\infty$. The minimum $\$1$ price obtains if any of the parameters $\delta, \sigma, T$ tends to 0. Since the increasing function $N(\bullet)$ is concave over $[0,\infty)$, we see that the option price is increasing and concave separately in each of the parameters $\delta, \sigma, T$. Given two rebalancing rules $b>c$ with distance $\delta=\|b-c\|$, an initial $\$1$ deposit into the horizon-$T$ replicating strategy achieves at $T$ a compound growth-rate that is exactly $$\frac{100}{T}\log\bigg\{2N\bigg(\frac{\delta}{2}\sigma\sqrt{T}\bigg)\bigg\}$$percent lower than that of the best of $\{b,c\}$ in hindsight. A $\$1$ deposit into the corresponding horizon-free strategy (that replicates the perpetual version of the option) achieves a compound-growth rate at $T$ that is at most $100\log(2)/T$ percent lower than that of the best of $\{b,c\}$ in hindsight. The trader’s initial ($\$1$) deposit into the replicating strategy buys him $1/C_0$ units of the option at $t=0$. For the horizon-$T$ option, his wealth at expiration will be $\max\{V_T(b),V_T(c)\}/C_0$, and hence the excess continuously-compounded growth rate will be $$\frac{1}{T}\log[\max\{V_T(b),V_T(c)\}]-\frac{1}{T}\log[\max\{V_T(b),V_T(c)\}/C_0]=\frac{\log C_0(\delta,\sigma,T)}{T}.$$For the horizon-free option, the trader’s initial dollar buys him half a unit of the option at $t=0$. Thus, his wealth at $T$ will be at least half the exercise value of the option, which is $\max\{V_T(b),V_T(c)\}$. Hence, the excess continuously-compounded growth rate of the hindsight-optimized rule at $T$ is at most $$\frac{1}{T}\log[\max\{V_T(b),V_T(c)\}]-\frac{1}{T}\log[\max\{V_T(b),V_T(c)\}/2]=\frac{\log 2}{T}.$$ Consider the following robust scheme for $T:=25$ years of leveraged bets on the S&P 500 index. We put $b:=2$ and $c:=1$ (e.g. buy-and-hold), with $\sigma:=0.15$. We get $C_0=\$1.29$ and $\log(C_0)/T=1\%$, so the replicating strategy is guaranteed to achieve a final compound-growth rate that is $1\%$ lower than the best of $\{b,c\}$ in hindsight. If $b=2$ happens to outperform the index by more than $1\%$ per year, then the trader will beat the market over $t\in[0,25]$. If $b=2$ underperforms the index (or outperforms by less than $1\%$ a year), then the trader’s compound-growth rate will have lagged the market by at most $1\%$ a year. Note that the corresponding horizon-free strategy (that replicates the perpetual version of the option) can only guarantee to get within $\log(2)/T=2.8\%$ of the hindsight-optmized growth rate at $T=25$ We construct a robust $T:=25$ year scheme for long-run stock market investment that guarantees preservation of capital. We put $b:=1$ ($100\%$ stocks) and $c:=0$ (all cash). Assuming that $\sigma:=0.15$, the practitioner can rest easy, safe in the knowledge that his foray into risk assets will ultimately not cause him to lag the risk-free rate by more than $1\%$ a year. If $r>0.01$, then he is guaranteed not to lose money if he sticks to the Plan for $T=25$ years. At the same time, if stocks go through the roof, his strategy will earn the long-run market growth rate minus a $1\%$ “universality cost.” Would-be practitioners who enjoyed these example can use Figure \[fig:regret\] to inform the choice of horizon: it plots the excess continuously-compounded growth rate for different volatilities and maturities with $\delta:=1$. ![The excess percent growth rate of the best of two rebalancing rules over the replicating strategy, for different horizons and volatilities, with $\delta:=1$.[]{data-label="fig:regret"}](excess2.png){width="375px"} General Formulas for Pricing and Replication -------------------------------------------- Before we can put our on-line schemes for robust asset allocation into actual practice, we must derive general time-$t$ formulas for pricing and replication of the rebalancing option under discrete hindsight optimization. Thus, we proceed to extend the above integration technique to the general situation. To simplify the notation, we let $\tau:=T-t$ denote the remaining life of the option at time $t$. Inspired by Garivaltis (2018), we start with the decomposition $$\boxed{z_T=\sqrt{\frac{t}{T}}\cdot z_t+\sqrt{\frac{\tau}{T}}\cdot y},$$where $$\boxed{y:=\frac{\log(S_T/S_t)-(r-\sigma^2/2)\tau}{\sigma\sqrt{\tau}}}$$is distributed unit normal with respect to the equivalent martingale measure and the information available at $t$. Conditional on the values of time-$t$ variables, $b$ outperforms $c$ at $T$ if and only if $$\boxed{y\geq\frac{\frac{b+c}{2}\sigma T-\sqrt{t}\cdot z_t}{\sqrt{\tau}}}.$$ Thus, the general price $C(S,t)$ is equal to the sum of two integrals $I_1+I_2$, where $$I_1:=\frac{\exp(rt-\sigma^2b^2T/2+b\sigma\sqrt{t}\cdot z_t)}{\sqrt{2\pi}}\int_{[(b+c)\sigma T/2-\sqrt{t}\cdot z_t]/\sqrt{\tau}}^\infty\exp(-y^2/2+b\sigma\sqrt{\tau}\cdot y)dy$$and $$I_2:=\frac{\exp(rt-\sigma^2c^2T/2+c\sigma\sqrt{t}\cdot z_t)}{\sqrt{2\pi}}\int_{-\infty}^{[(b+c)\sigma T/2-\sqrt{t}\cdot z_t]/\sqrt{\tau}}\exp(-y^2/2+c\sigma\sqrt{\tau}\cdot y)dy.$$ These integrals simplify to $$I_1=N\bigg(\frac{[b-c]\sigma T/2+\sqrt{t}\cdot z_t-b\sigma t}{\sqrt{\tau}}\bigg)V_t(b)$$ and $$I_2=N\bigg(\frac{[b-c]\sigma T/2-\sqrt{t}\cdot z_t+c\sigma t}{\sqrt{\tau}}\bigg)V_t(c).$$ The general cost $C(S,t)$ of achieving the best of two rebalancing rules $b>c$ in hindsight is $$\label{eq:cost2} \boxed{C=N(d_1)V_t(b)+N(d_2)V_t(c)},$$where $$\label{d1} \boxed{d_1:=\frac{(b-c)\sigma T/2+\sqrt{t}\cdot z_t-b\sigma t}{\sqrt{\tau}}}$$ and $$\boxed{d_2:=(b-c)\sigma\sqrt{\tau}-d_1=\frac{(b-c)\sigma T/2-\sqrt{t}\cdot z_t+c\sigma t}{\sqrt{\tau}}}.$$ A perpetual option ($T:=\infty$) on the best of two rebalancing rules $b>c$ costs $C(S,t)=V_t(b)+V_t(c)$ in state $(S_t,t)$. To delta-hedge the perpetual option, one holds $$\boxed{\Delta=\frac{bV_t(b)+cV_t(c)}{S_t}}$$shares of the underlying in state $(S_t,t)$, and therefore bets the fraction $$\boxed{\hat{b}(S_t,t):=\frac{\Delta S}{C}=\frac{bV_t(b)+cV_t(c)}{V_t(b)+V_t(c)}}$$of wealth on the underlying at $t$. As $T\to\infty$, we see that $d_1,d_2\to+\infty$ and the option price converges to $V_t(b)+V_t(c)$. Next, one can verify by direct calculation from (\[fortune\]) and (\[auxiliary\]) that $$\frac{\partial V(b)}{\partial S}=\frac{\partial V(b)}{\partial z}\frac{\partial z}{\partial S}=\frac{bV_t(b)}{S_t}.$$Alternately, one can observe that the rebalancing rule $b$ keeps (by definition) $bV_t(b)$ dollars in the stock at time $t$, which amounts to $bV_t(b)/S_t$ shares. Thus, to replicate the sum $V_t(b)+V_t(c)$ we must own a total of $\Delta=bV_t(b)/S_t+cV_t(c)/S_t$ shares of the underlying. We should note that our general pricing formulas could have been obtained differently, by applying the theory of “exchange options” that was bequeathed to us in sumultaneous papers by Margrabe (1978) and Fischer (1978). Rather than the single underlying $S_t$, one could view the (perfectly correlated) geometric Brownian motions $U_1(t):=V_t(b)$ and $U_2(t):=V_t(c)$ as underlyings of a multi-asset option with payoff $$\max\{U_1,U_2\}=\max\{U_1-U_2,0\}+U_2.$$ This amounts to a $\$1$ deposit into the rebalancing rule $c$, plus the option to exchange the final wealth of $c$ for the final wealth of $b$ at $T$. Substituting the aggregate volatility $\sigma_a:=(b-c)\sigma$ into Margrabe’s Formula (cf. Zhang 1998) yields the same result $$\label{margrabe} \boxed{C(U_1,U_2,t)=N(d_1)U_1+N(\sigma_a\sqrt{\tau}-d_1)U_2},$$where $$\label{margrabe2} \boxed{d_1:=\frac{\log(U_1/U_2)}{\sigma_a\sqrt{\tau}}+\frac{\sigma_a\sqrt{\tau}}{2}},$$is in agreement with (\[d1\]). Figure \[fig:iv\] plots the price and intrinsic value of the rebalancing option for different values of $S$ under the parameters $r:=0.03, T:=10, S_0:=100, t:=5, \sigma:=0.7, b:=1.5,$ and $c:=0.5$. The horizon-$T$ replicating strategy for the best of two rebalancing rules $b>c$ in hindsight holds $$\boxed{\Delta=\frac{N(d_1)bV_t(b)+N(d_2)cV_t(c)}{S_t}}$$shares of the stock in state $(S_t,t)$, which amounts to betting the fraction $$\boxed{\hat{b}(S_t,t)=\frac{\Delta S}{C}=\frac{N(d_1)bV_t(b)+N(d_2)cV_t(c)}{N(d_1)V_t(b)+N(d_2)V_t(c)}}.$$ of wealth on the stock at $t$. Thus, the on-line fraction of wealth bet on the stock is a time- and state-varying convex combination of $b$ and $c$. First, we note the standard relations $\partial C/\partial U_1=N(d_1)$ and $\partial C/\partial U_2=N(d_2)$, which follow by direct calculation from (\[margrabe\]), (\[margrabe2\]), and the fact that $U_1\phi(d_1)=U_2\phi(d_2)$, where $\phi(\bullet)$ is the standard normal density function. Differentiating the option price, we get $$\frac{\partial C}{\partial S}=\frac{\partial C}{\partial U_1}\frac{\partial U_1}{\partial S}+\frac{\partial C}{\partial U_2}\frac{\partial U_2}{\partial S}=N(d_1)\frac{bU_1}{S}+N(d_2)\frac{cU_2}{S},$$which is the desired result. Thus, even if the best rebalancing rule (of any kind) in hindsight $b(S,t)=z(S,t)/(\sigma\sqrt{t})$ happens to lie between $b$ and $c$, the replicating strategy will not generally bet the hindsight-optimized fraction $\arg\max\limits_{b\in\mathbb{R}}V_t(b)$ of wealth on the stock. This phenomenon is illustrated in Figure \[fig:comparison\]. Instead, the relative weighting of $b$ and $c$ (which is initially $50/50$ at time-$0$) evolves with the observed performances $V_t(b), V_t(c)$ and the remaining life $\tau:=T-t$ of the option. For a fixed time $t$, if $U_1\to\infty$ or $U_2\to\infty$ then the on-line portfolio weight will converge to $b$ or $c$ accordingly. As $\tau\to 0$, $d_1$ tends to $\pm\infty$ and $d_2$ tends to $\mp\infty$ according as to whether $b$ or $c$ has outperformed over the known price history. Thus, small differences in the observed performances $V_t(b)$ and $V_t(c)$ get amplified in the on-line portfolio weight as $\tau\to0$. Figures \[fig:sim\] and \[fig:sim2\] simulate the performance of the replicating strategy for different parameter values over a $T:=30$ year horizon. ![The fraction of wealth bet by the replicating strategy for different stock prices, $r:=0.03, T:=10, S_0:=100, t:=5, \sigma:=0.7, b:=1.5, c:=0.5$.[]{data-label="fig:comparison"}](comparison.png){width="325px"} ![Option price and intrinsic value for different stock prices, $r:=0.03, T:=10, S_0:=100, t:=5, \sigma:=0.7, b:=1.5, c:=0.5$.[]{data-label="fig:iv"}](iv.png){width="300px"} ![Performance simulation over $T:=30$ years for the parameters $S_0:=1, b:=2, c:=0.5, r:=0.03, \sigma:=0.15, \nu:=0.1, \mu=\nu+\sigma^2/2.$[]{data-label="fig:sim"}](sim1.png){width="425px"} ![Performance simulation over $T:=30$ years for the parameters $S_0:=1, b:=2, c:=0.5, r:=0.03, \sigma:=0.7, \nu:=0.07, \mu=\nu+\sigma^2/2.$[]{data-label="fig:sim2"}](sim2.png){width="425px"} The General Discrete Set of Asset Allocations {#several} ============================================= We carry on with the general discrete set $\mathbb{B}:=\{b_1,...,b_n\}\subset\mathbb{R}$ of asset allocations, where the $b_i$ are arranged in increasing order: $b_1<b_2<\cdot\cdot\cdot<b_n$. Thus, we now deal with the payoff $\ V_t^:=\max\limits_{1\leq i\leq n}V_t(b_i)$. For notational convenience, we will also write $b_0:=-\infty$ and $b_{n+1}:=+\infty$. We let $\Delta b_i:=b_{i+1}-b_{i}$ for $0\leq i\leq n$, and thus $\Delta b_0=\Delta b_{n}=+\infty$. For a given rebalancing rule $b_i\,(1\leq i\leq n)$, the final payoff of the option is equal to $V_T(b_i)$ if and only if $$\frac{b_{i-1}+b_{i}}{2}\sigma\sqrt{T}\leq z_T\leq \frac{b_i+b_{i+1}}{2}\sigma\sqrt{T}.$$ Thus, conditional on the values of time-$t$ variables, $b_i$ will turn out to be the best performer over $[0,T]$ if and only if $$\frac{(b_{i-1}+b_{i})\sigma T/2-\sqrt{t}\cdot z_t}{\sqrt{\tau}}\leq y\leq \frac{(b_i+b_{i+1})\sigma T/2-\sqrt{t}\cdot z_t}{\sqrt{\tau}}.$$ For simplicity, we will write this interval as $y\in[A_{i-1},A_i]$. Thus $A_0=-\infty$ and $A_n=+\infty$. The expected present value of the final payoff with respect to $\mathbb{Q}$ and the information available at $t$ is equal a sum of integrals $I_1+\cdot\cdot\cdot+I_n$, where $$I_i:=\frac{\exp(rt-\sigma^2b_i^2T/2+b_i\sigma\sqrt{t}\cdot z_t)}{\sqrt{2\pi}}\int_{A_{i-1}}^{A_i}\exp(-y^2/2+b_i\sigma\sqrt{\tau}\cdot y)dy.$$ Evaluating these integrals, we obtain the general pricing formula $$\boxed{C(S,t)=\sum\limits_{i=1}^n\{N(A_i-\beta_i)-N(A_{i-1}-\beta_i)\}V_t(b_i)},$$where $$\boxed{A_i:=\frac{(b_i+b_{i+1})\sigma T/2-\sqrt{t}\cdot z_t}{\sqrt{\tau}}}.$$and $$\boxed{\beta_i:=b_i\sigma\sqrt{\tau}}.$$ Bearing in mind that $A_0=-\infty$ and $A_n=+\infty$, we can also write $$\begin{gathered} \label{eq:genprice} C(S,t)=N(A_1-\beta_1)V_t(b_1)+\sum\limits_{i=2}^{n-1}\{N(A_i-\beta_i)-N(A_{i-1}-\beta_i)\}V_t(b_i)\\ +N(\beta_n-A_{n-1})V_t(b_n).\\\end{gathered}$$The general option price could again have been obtained differently, by an interesting application of Margrabe’s theory of exchange options. Indeed, we could consider the wealth processes $(V_t(b_i))_{i=1}^n$ as separate underlyings $U_i:=V_t(b_i)$ of a multi-asset option whose final payoff is equal to $\max\{U_1,U_2,...,U_n\}$. First of all, we remark that at any given time the ordered sequence of numbers $U_1(t),...,U_n(t)$ is unimodal, or single-peaked. This happens because the $(\log U_i)_{i=1}^n$ trace out a sequence of heights on the parabola $b\mapsto\log V_t(b)$ as we move from left to right over the abscissae $b_1<b_2<\cdot\cdot\cdot<b_n$. The peak occurs for the index $$i^:=\arg\min\limits_{1\leq i\leq n}\|b_i-b(S_t,t)\|=\arg\max\limits_{1\leq i\leq n}V_t(b_i).$$Thus, $U_i$ is increasing in $i$ for $i<i^$ and decreasing in $i$ for $i\geq i^$. This unimodality in hand, we now have the identity $$\max\{U_1,...,U_n\}=U_1+(U_2-U_1)^++(U_3-U_2)^++\cdot\cdot\cdot+(U_n-U_{n-1})^+,$$where $x^+:=\max\{x,0\}$ denotes the positive part of $x$. Hence, the payoff $\max\limits_{1\leq i\leq n}U_i$ is equivalent to a portfolio consisting of one unit of $U_1$, plus an option to exchange $U_1$ for $U_2$, plus an option to exchange $U_2$ for $U_3$, $\cdot\cdot\cdot$ , plus an option to exchange $U_{n-1}$ for $U_n$. At expiration, the trader keeps exchanging $U_{i}$ for $U_{i+1}$ until the maximum $U_i^$ is reached. Applying the Margrabe Formula (cf. Zhang 1998) in conjunction with linear pricing, we find that the no-arbitrage price of this portfolio (consisting of a unit of $U_1$ plus $n-1$ exchange options) is $$\label{easy} \boxed{C(U_1,...,U_n,t)=U_1+\sum\limits_{i=1}^{n-1}\{N(d_{1i})U_{i+1}-N(d_{2i})U_i\}},$$where $$\boxed{d_{1i}:=\frac{\log(U_{i+1}/U_i)}{\sigma_{ai}\sqrt{\tau}}+\frac{\sigma_{ai}\sqrt{\tau}}{2}},$$$d_{2i}:=d_{1i}-\sigma_{ai}\sqrt{\tau}$, and $\sigma_{ai}:=\Delta b_i\sigma$ is the aggregate volatility in a two-asset market consisting of $U_i$ and $U_{i+1}$. Collecting terms, we get the linear combination $$\label{combo} \boxed{C(U_1,...,U_n,t)=N(-d_{21})U_1+\sum_{i=2}^{n-1}[N(d_{1,i-1})-N(d_{2i})]U_i+N(d_{1,n-1})U_n},$$which agrees with equation (\[eq:genprice\]) above. Figure \[payoff2\] plots the option price and intrinsic value for different stock prices under the parameters $r:=0.03, T:=10, S_0:=100, t:=5, \sigma:=0.7,$ and $\mathbb{B}:=\{0,0.5,1,1.5,2\}.$ In general there will be several implied volatilities $\sigma$ that could rationalize an observed price of the rebalancing option. Figure \[vol\] plots the option price for different volatilities under the parameters $r:=0.03, T:=10, S_0:=100, t:=5, S_t:=200,$ and $\mathbb{B}:=\{0,0.5,1.5\}$. ![Option price and intrinsic value for different stock prices, $r:=0.03, T:=10, S_0:=100, t:=5, \sigma:=0.7, \mathbb{B}:=\{0,0.5,1,1.5,2\} $.[]{data-label="payoff2"}](generaliv.png){width="300px"} ![Option prices for different volatilities, $r:=0.03, T:=10, S_0:=100, t:=5, S_t:=200, \mathbb{B}:=\{0,0.5,1.5\} $.[]{data-label="vol"}](vol.png){width="300px"} In specializing the pricing formula for $t:=0$ and simplifying (remembering that $V_0(b_i):=1$), we get For hindsight optimization over $n$ discrete rebalancing rules $b_1<\cdot\cdot\cdot<b_n$, the time-0 cost of achieving the best $b_i$ in hindsight is $$\boxed{C_0=2-n+2\sum_{i=1}^{n-1}N(\Delta b_i\sigma\sqrt{T}/2)},$$where $\Delta b_i:=b_{i+1}-b_i$. If $\delta:=\max\limits_{1\leq i\leq n-1}\Delta b_i$, then $$\boxed{C_0\leq 2-n+2(n-1)N(\delta\sigma\sqrt{T}/2)}.$$ A perpetual option on the best of any $n$ distinct rebalancing rules in hindsight is worth $C_0=n$ dollars at time-0. Thus, we see that the time-0 price of the general horizon-$T$ rebalancing option is independent of the interest rate, and it is increasing and concave separately in the parameters $\Delta b_i, \sigma, T$. We again observe that horizontal translations of the point set $\{b_1,...,b_n\}$ do not alter the option price. We always have the relation $1\leq C_0\leq n$; the maximum $n$ dollar price obtains if any of the parameters tends to infinity and the minimum $\$1$ price obtains if any of the parameters tends to zero. For a $T:=25$ year planning horizon, we cherry pick five favored asset allocations $\mathbb{B}:=\{0,0.5,1,1.5,2\}$. Assuming stock market volatility of $\sigma:=0.15$ going forward, we get $C_0=\$1.59$, and the excess growth rate of the hindsight-optimized asset allocation will be exactly $\log(C_0)/T=1.87\%$. Assuming that the risk-free rate is greater than $1.87\%$, the replicating strategy is guaranteed not to lose money if the practitioner sticks to the Plan for the next $T=25$ years. The horizon-$T$ replicating strategy for the best of the rebalancing rules $b_1<b_2<\cdot\cdot\cdot<b_n$ in hindsight holds $$\boxed{\Delta=N(-d_{21})\frac{b_1V_t(b_1)}{S_t}+\sum\limits_{i=2}^{n-1}[N(d_{1,i-1})-N(d_{2i})]\frac{b_iV_t(b_i)}{S_t}+N(d_{1,n-1})\frac{b_nV_t(b_n)}{S_t}}$$shares of the stock in state $(S_t,t)$, thereby betting the fraction $\hat{b}(S,t)=\Delta S/C$ of its bankroll on the stock. This amounts to a time- and state-varying convex combination of the $b_i$. As $\tau\to0$, the option price converges to $U_{i^}:=\max\limits_{1\leq i\leq n}U_i$ and the fraction of wealth bet by the replicating strategy converges to $\arg\max\limits_{b\in\mathbb{B}}V_t(b)$ if this set is a singleton; if $\arg\max\limits_{b\in\mathbb{B}}V_t(b)=U_{i^}=U_{i^+1}$ has two distinct points, then $\hat{b}$ converges to the midpoint $(b_{i^}+b_{i^+1})/2$ as $\tau\to0$. The horizon-free replicating strategy (corresponding to the perpetual version of the option) bets the performance-weighted average $$\hat{b}(S_t,t)=\frac{\sum\limits_{i=1}^nb_iV_t(b_i)}{\sum\limits_{i=1}^nV_t(b_i)}$$of the rebalancing rules $b_i$, which converges almost surely to $$\arg\max_{b\in\mathbb{B}}\bigg\{(\mu-r)b-\frac{\sigma^2b^2}{2}\bigg\}=\arg\min_{b\in\mathbb{B}}\bigg\|b-\frac{\mu-r}{\sigma^2}\bigg\|,$$i.e. it converges to whichever element of $\mathbb{B}$ is closest to the continuous time Kelly rule (cf. Luenberger 1998). Note that the pricing formula (\[combo\]) is a linearly homogeneous function of the underlyings $(U_1,...,U_n)$. By Euler’s theorem for homogeneous functions, we therefore have the relation $$C=\sum\limits_{i=1}^n\frac{\partial C}{\partial U_i}U_i.$$Accordingly, by direct calculation on (\[combo\]) one can (carefully) verify the partial derivatives $$\frac{\partial C}{\partial U_1}=N(-d_{21}),$$ $$\frac{\partial C}{\partial U_i}=N(d_{1,i-1})-N(d_{2i})\text{ for } 2\leq i\leq n-1,$$ $$\frac{\partial C}{\partial U_n}=N(d_{1,n-1}).$$To verify these partials easily, one needs the identity$$\frac{\phi(d_{2i})}{\phi(d_{1i})}=\frac{U_{i+1}}{U_i}\text{ for }1\leq i\leq n-1,$$where $\phi(\bullet)$ is the standard normal density function. Observe that $U_i$ generally appears in the terms of (\[combo\]) that correspond to the indices $i-1, i, \text{and } i+1$. $U_1$ appears in the first two terms and $U_n$ appears in the last two terms. This being done, the delta-hedging strategy now obtains from the chain rule $$\frac{\partial C}{\partial S}=\sum\limits_{i=1}^n\frac{\partial C}{\partial U_i}\frac{\partial U_i}{\partial S}$$in conjunction with the fact that $\partial U_i/\partial S=b_iV_t(b_i)/S$. To get the horizon-free result, we just observe that $d_{1i}\to+\infty$ and $d_{2i}\to-\infty$ as $T\to\infty$. Finally, consider what happens when $\tau\to0$. The numbers $d_{1i},d_{2i}$ will converge to the same limit $\pm\infty$ according as $U_{i+1}$ is greater or less than $U_i$. In the event that $U_{i+1}=U_i$ then $d_{1i}$ and $d_{2i}$ both converge to zero. The numbers $(U_i)_{i=1}^n$ will typically have a unique mode $U_{i^}$, e.g. $U_1<\cdot\cdot\cdot<U_{i^-1}<U_{i^}>U_{i^+1}>\cdot\cdot\cdot>U_n$. In this case, all coefficients in the linear combination (\[combo\]) converge to zero except the one corresponding to $i=i^$, which converges to $1$. If there are two modes $U_{i^}=U_{i^+1}$, then the corresponding coefficients in (\[combo\]) both converge to $1/2$, and the result follows. Note that for $n>2$, the initial weighting of the $b_i$ at time-$0$ is not uniform, even if the $b_i$ themselves are equally spaced. The endpoints $b_1$ and $b_n$ have initial weights $N(\Delta b_1\sigma\sqrt{T}/2)/C_0$ and $N(\Delta b_{n-1}\sigma\sqrt{T}/2)/C_0$, respectively, and the rest of the $b_i$ have initial weights $[N(\Delta b_{i-1}\sigma\sqrt{T}/2)-N(-\Delta b_{i}\sigma\sqrt{T}/2)]/C_0$ for $2\leq i\leq n-1$. If the $b_i$ are equally spaced, then each of the intermediate points ($2\leq i\leq n-1$) gets initial weight $[2N(\delta\sigma\sqrt{T}/2)-1]/C_0$, but the endpoints $b_1,b_n$ get the higher initial weight $N(\delta\sigma\sqrt{T}/2)/C_0$. For any closed set $\mathbb{B}\subseteq\mathbb{R}$ of rebalancing rules (finite or infinite), the American-style version of Cover’s rebalancing option (with exercise price $K$ and payoff $\max\{\max\limits_{b\in\mathbb{B}}V_t(b)-K,0\})$ will never be exercised early in equilibrium, and its price $C_a(S_t,t)$ equals the price $C_e(S_t,t)$ of the corresponding European-style option. For simplicity, let $V_t^:=\max\limits_{b\in\mathbb{B}}V_t(b)$ denote the hindsight-optimized wealth over $[0,t]$, and let $b^_t:=\arg\max\limits_{b\in\mathbb{B}}V_t(b)$ denote the best (allowable) rebalancing rule in hindsight over $[0,t]$. Consider, from the standpoint of time $t$, the following two trading strategies: - Invest $Ke^{-r\tau}$ dollars in the risk-free bond and buy $1$ unit of Cover’s (European-style) rebalancing option at a price of $C_e(S_t,t)$. - Invest $V_t^$ dollars into the rebalancing rule $b_t^$. That is, take the best rebalancing rule in hindsight over $[0,t]$, and adhere to that same (constant) continuously-rebalanced portfolio over $[t,T]$. Observe that Strategy 1 has a final payoff of $\max\{V_T^,K\}$ and Strategy 2 has a final payoff of $V_T(b_t^)$. Since the payoff at $T$ of Strategy 1 is guaranteed to be at least as great as that of Strategy 2, the initial investment of $Ke^{-r\tau}+C_e(S_t,t)$ dollars into Strategy 1 must be greater or equal to the investment $V_t^$ that is required for Strategy 2. Thus, we have the inequalities $$C_a(S_t,t)\geq C_e(S_t,t)\geq V_t^-Ke^{-r\tau}\geq V_t^-K.$$Hence, since the price of an American rebalancing option always exceeds the exercise value, the option “is worth more alive than dead” and will never be exercised in equilibrium. On account of the fact that early exercise rights are worthless anyhow, we must therefore have $C_a(S_t,t)=C_e(S_t,t)$. We remark that this is a general model-independent result that applies equally well to rebalancing rules $b\in\mathbb{B}\subseteq\mathbb{R}^d$ over arbitrary $d$-dimensional stock markets. The dominance argument only requires the market (and the set $\mathbb{B}$) to admit a well-defined hindsight-optimized rebalancing rule $b_t^:=\arg\max\limits_{b\in\mathbb{B}}V_t(b)$. For $\mathbb{B}:=\{1\}$ the best rebalancing rule in hindsight is just $b_t^=1$ and we get $V_t^=S_t$; this recovers the proof given by Merton (1973, 1990) of the no-exercise theorem for vanilla call options. The special cases $\mathbb{B}:=\mathbb{R}^d$ and $\mathbb{B}:=[0,1]$ were observed by Garivaltis (2018) for a continuous-time Black-Scholes market with $K:=0$. Conclusion ========== This paper studied Cover’s rebalancing option with discrete hindsight optimization. In the context of a single risk asset, a constant (perhaps levered) rebalancing rule is a simple trading strategy that continuously maintains some fixed fraction of wealth in the underlying asset. Cover’s discrete-time universal portfolio theory derives robust on-line trading strategies that are guaranteed to achieve an acceptable percentage of the final wealth of the best rebalancing rule (of any kind) in hindsight. Working in continuous time, we formulated the less aggressive benchmark of the best rebalancing rule in hindsight that hails from some finite set $\mathbb{B}:=\{b_1,...,b_n\}$. This approach allows the (delta-hedging) practitioner to cherry pick a small number of favored rebalancing rules that could embody institutional leverage constraints or the trader’s own speculative beliefs as to the future pattern of returns in the stock market. Accordingly, we priced and replicated the financial option whose final payoff is equal to the wealth $V_T^:=\max\limits_{1\leq i\leq n}V_T(b_i)$ that would have accrued to a $\$1$ deposit into the best $b_i$ in hindsight. We found that a perpetual option (with zero exercise price) on the best of $n$ distinct rebalancing rules costs $n$ dollars at $t=0$. The corresponding (horizon-free) replicating strategy amounts to depositing a dollar into each $b_i$ and “letting it ride.” If the option expires at some fixed date $T$ the price is lower; it is concavely increasing in $T$ and in the volatility $\sigma$ of the underlying risk asset. From the standpoint of $t=0$, the cost $C_0$ of achieving the best of the $b_i$ is translation invariant: it increases monotonically with the distances $\Delta b_i$ between adjacent rebalancing rules, but it does not otherwise depend on their precise location. In this connection, the replicating strategy amounts to a time- and state-varying convex combination of the $b_i$ that dynamically considers both the observed performances $V_t(b_i)$ and the remaining life $\tau:=T-t$ of the option. No-arbitrage considerations dictate that American-style rebalancing options (for general exercise price $K$) will never be exercised early in equilibrium. This model-independent result holds for arbitrary closed sets $\mathbb{B}\subseteq\mathbb{R}^d$ of rebalancing rules over any $d-$dimensional stock market that admits a well-defined best rebalancing rule in hindsight. Toward the end of the investment horizon (as it becomes more and more obvious which $b_i$ is likely to be the best in hindsight), even small differences in observed performance will cause the replicating strategy to dramatically over- or under-weight the various $b_i$. Any practitioner of the horizon-$T$ delta-hedging strategy is guaranteed to achieve at $T$ the deterministic* fraction $1/C_0$ of the final wealth of the best $b_i$ in hindsight. The excess compound-growth rate at $T$ of the best $b_i$ (over and above the trader) is $\log(C_0)/T$, which tends to $0$ as $T\to\infty$. The replicating strategy will asymptotically beat the underlying (i.e. an S&P 500 ETF) if any of the $b_i$ turns out to achieve a compound-growth rate that is higher than $b=1$. If there is no such $b_i\in\mathbb{B}$, but the trader had the good sense to put $1\in\mathbb{B}$, then the trader’s compound-annual growth rate will lag the underlying risk asset by at most $100\log(C_0)/T$ percent at $T$. If we have $0\in\mathbb{B}$, then the trader also guarantees that he will ultimately not lose money over $[0,T]$ if the condition $\log(C_0)/T<r$ is satisfied. Hence, our trading strategy is in a sense the most conservative attempt at detecting on-the-fly whether any of the rebalancing rules in some finite set is capable of beating the underlying over a given investment horizon. We have therefore obtained a universal continuous-time asset allocation scheme that is computationally pleasant as well as feasible for the human life span. The on-line behavior is Markovian in the sense that the relevant state vector is just $(S_t,t,(V_t(b_i))_{i=1}^n)$. The algorithm requires no prior knowledge of the (hard-to-estimate) drift parameter $\mu$ of the stock market. Apart from the finite-dimensional state vector, the trader’s behavior depends only on the known parameters $r, \sigma, T,$ and $\mathbb{B}$. In just 25 years, say, our method guarantees to achieve within $1.87\%$ of the compound-annual growth rate of whichever turns to be the most profitable asset allocation among $\mathbb{B}:=\{0,0.5,1,1.5,2\}$. [9]{} Black, F. and Scholes, M., 1973. The Pricing of Options and Corporate Liabilities. Journal of Political Economy, 81(3), pp.637-654. Breiman, L., 1961. Optimal Gambling Systems for Favorable Games. In Proceedings of the Fourth Berkeley Symposium on Mathematical Statistics and Probability, Volume 1: Contributions to the Theory of Statistics. The Regents of the University of California. Cover, T.M., 1991. Universal Portfolios. Mathematical Finance, 1(1), pp.1-29. Cover, T.M. and Gluss, D.H., 1986. Empirical Bayes Stock Market Portfolios. Advances in Applied Mathematics, 7(2), pp.170-181. Cover, T.M. and Ordentlich, E., 1996. Universal Portfolios with Side Information. IEEE Transactions on Information Theory, 42(2), pp.348-363. Cover, T.M. and Thomas, J.A., 2006. Elements of Information Theory. John Wiley & Sons. Fischer, S., 1978. Call Option Pricing When the Exercise Price is Uncertain, and the Valuation of Index Bonds. The Journal of Finance, 33(1), pp.169-176. Garivaltis, 2018. Exact Replication of the Best Rebalancing Rule in Hindsight. Working Paper. Kelly J.L., 1956. A New Interpretation of Information Rate. Bell System Technical Journal. Luenberger, D.G., 1998. Investment Science. Oxford University Press. MacLean, L.C., Thorp, E.O. and Ziemba, W.T., 2011. The Kelly Capital Growth Investment Criterion: Theory and Practice. World Scientific. Margrabe, W., 1978. The Value of an Option to Exchange one Asset for Another. The Journal of Finance, 33(1), pp.177-186. Markowitz, H., 1952. Portfolio Selection. The Journal of Finance, 7(1), pp.77-91. Merton, R.C., 1973. Theory of Rational Option Pricing. The Bell Journal of Economics and Management Science, pp.141-183. Merton, R.C., 1990. Continuous-Time Finance. Blackwell. Ordentlich, E. and Cover, T.M., 1998. The Cost of Achieving the Best Portfolio in Hindsight. Mathematics of Operations Research, 23(4), pp.960-982. Poundstone, W., 2010. Fortune’s Formula: The Untold Story of the Scientific Betting System That Beat the Casinos and Wall Street. Hill and Wang. Reiner, E. and Rubinstein, M., 1992. Exotic Options. Working Paper. Thorp, E.O., 1966. Beat the Dealer: a Winning Strategy for the Game of Twenty One. Random House. Thorp, E.O., 2017. A Man for All Markets. Random House. Wilmott, P., 1998. Derivatives: the Theory and Practice of Financial Engi- neering. John Wiley & Sons. Wilmott, P., 2001. Paul Wilmott Introduces Quantitative Finance. John Wiley & Sons. Zhang, P., 1998. Exotic Options: a Guide to Second Generation Options. World Scientific. [^1]: Assistant Professor of Economics, Northern Illinois University, 514 Zulauf Hall, DeKalb IL 60115. E-mail: [email protected]. ORCID: 0000-0003-0944-8517.	{ "pile_set_name": "ArXiv" }
--- author: - '[Ehsan Adeli$^{}$, Bojana Rosić$^{}$, Hermann G. Matthies$^{}$ and Sven Reinstädler$^{\dag}$]{}' date: July 2018 title: Effect of Load Path on Parameter Identification for Plasticity Models using Bayesian Methods --- Introduction {#sec:Introduction} ============ In order to predict the behavior of loaded metallic materials, constitutive models are applied, which present a mathematical frame for the description of elastic and inelastic deformation. The models by Miller, Krempl, Korhonen, Aubertin, Chan, and Bodner are well-known constitutive models for isotropic materials [@Miller; @Krempl; @Korhonen1; @Aubertin; @Chan]. In 1983, Chaboche [@Chaboche; @Chaboche1] put forward what has become known as the unified Chaboche viscoplasticity constitutive model, which has been widely accepted. All inelastic constitutive models contain parameters which have to be identified for a given material from experiments. In the literature only few investigations can be found dealing with identification problems using stochastic approaches. Klosowski and Mleczek have applied the least-squares method in the Marquardt-Levenberg variant to estimate the parameters of an inelastic model [@KLOSOWSKI]. Gong et al. have also used some modification of the least-squares method to identify the parameters [@Gong]. Harth and Lehn identified the model parameters of a model by employing some generated artificial data instead of experimental data using a stochastic technique [@Harth]. A similar study by Harth and Lehn has been done for other constitutive models like Lindholm and Chan [@Lindholm]. In this paper, a viscoplastic model of Chaboche is studied. The model contains five material parameters which have to be determined from experimental data. It should be noted that here virtual data are employed instead of real experimental data. A cyclic tension-compression test is applied in order to extract the virtual data. The model is described in Section 2, whereas Section 3 explains how to propagate the uncertainty in the model and how to perform the update. The probabilistic model is reformulated from the deterministic model, and once the forward model is provided, the model parameters are updated using a Bayesian approach. In Section 4 the desired parameters are identified from the measured data. In fact, the parameters which have been considered as uncertain parameters are updated and the uncertainties of the them are reduced while the random variables representing the uncertain parameters are updated during the process. The results are thoroughly studied and the identified parameters as well as the corresponding model responses are analyzed. Finally the prediction of the models is compared with the measured data for different applied load paths. It is also explained why different load paths cause different identification of model parameters. Model problem ============= The mathematical description of metals under cyclic loading beyond the yield limit that includes viscoplastic material behavior as well as the characterization of compulsory isotropic-kinematic hardening is here given in terms of a modified Chaboche model introduced in [@dinkler]. As we consider classical infinitesimal strains, we assume an additive strain decomposition. The material behavior is described for the elastic part by isotropic homogeneous elasticity, and for viscoplasticity the dissipation potential is given by $$\phi(\vek{\sigma}) = \frac{k}{n+1}\Biggl\langle\frac{\sigma_{ex}}{k}\Biggr\rangle^{n+1}, \label{eq:5}$$ with $\langle\cdot\rangle = \max(0, \cdot)$ and $k$ and $n$ as the material parameters. Here $\sigma_{ex}$ is the over-stress, defined via the equivalent stress ($\sigma_{eq}$) which reads $$\sigma_{eq}= \sqrt{\frac{3}{2} \text{tr} ((\vek{\sigma}-\vek{\chi})_D . ((\vek{\sigma}-\vek{\chi})_D)},$$ where $(\cdot)_D$ denotes the deviatoric part and $\vek{\chi}$ is the back-stress of kinematic hardening. The over-stress $\sigma_{ex}$ is given by $$\sigma_{ex} = \sigma_{eq}-\sigma_y -R= \sqrt{\frac{3}{2} \text{tr} ((\vek{\sigma}-\vek{\chi})_D . (\vek{\sigma}-\vek{\chi})} - \sigma_y - R,$$ where $\sigma_y$ is the yield stress and $R$ models the isotropic hardening which is introduced in the following. The partial derivative of the dissipation potential $\phi$ with respect to $\vek{\sigma}$ leads to the equation for the inelastic strain rate $$\dot{\vek{\epsilon}}_{vp} = \frac{\partial \phi}{\partial \vek{\sigma}} = \Biggl\langle\frac{\sigma_{ex}}{k}\Biggr\rangle^{n} \frac{\partial \sigma_{ex}}{\partial \vek{\sigma}}.$$ The viscoplastic model allows for isotropic and kinematic hardening, which is considered in order to describe different specifications. Assuming $R(t)$ and $\vek{\chi}(t)$ with $R(0) = 0$ and $\vek{\chi}(0) = 0$ to describe isotropic and kinematic hardening respectively, the evaluation equations for these two are $$\dot{R} = b_R(H_R-R)\dot{p}$$ and $$\dot{\vek{\chi}} = b_{\vek{\chi}}(\frac{2}{3} H_{\vek{\chi}}\frac{\partial \sigma_{eq}}{\partial \vek{\sigma}} - \vek{\chi} )\dot{p}$$ respectively. In the evaluation equations of the both hardening, $\dot{p}$ is the viscoplastic multiplier rate given as: $$\dot{p} = \Biggl\langle\frac{\sigma_{ex}}{k}\Biggr\rangle^{n},$$ which describes the rate of accumulated plastic strains. The parameter $b_R$ indicates the speed of stabilization, whereas the value of the parameter $H_R$ is an asymptotic value according to the evolution of the isotropic hardening. Similarly, the parameter $b_{\vek{\chi}}$ denotes the speed of saturation and the parameter $H_{\vek{\chi}}$ is the asymptotic value of the kinematic hardening variables. The complete model is stated in Table \[tab:1\]. Note that $\tnb{E}$ represents the elasticity tensor. [\|\| l\|\|]{} \[tab:1\] Strain\ $\vek{\epsilon}(t) = \vek{\epsilon}_e(t) + \vek{\epsilon}_{vp}(t)$ \ \ Hooke’s Law\ $\vek{\sigma}(t) = \tnb{E}: \vek{\epsilon}_e(t)$ \ \ Flow Rule\ $\dot{\vek{\epsilon}}_{vp}(t) = \langle\frac{\sigma_{eq}(t)-\sigma_y-R(t)}{k}\rangle^n \frac{\partial \sigma_{ex}}{\partial \vek{\sigma}}$ \ \ Hardening\ $\dot{R} = b_R(H_R-R)\dot{p}$ \ \ $\dot{\vek{\chi}} = b_{\vek{\chi}}(\frac{2}{3} H_{\vek{\chi}}\frac{\partial \sigma_{eq}}{\partial \vek{\sigma}} - \vek{\chi} )\dot{p}$ \ \ Initial Conditions\ $\vek{\epsilon}_{vp}(0)=0$, $R(0)=0$, $\vek{\chi}(0)=0$ \ \ Parameters\ $\sigma_y$ (Yield Stress) \ $k$, $n$ (Flow Rule) \ $b_R$, $H_R$, $b_{\vek{\chi}}$, $H_{\vek{\chi}}$ (Hardening) \ \[1ex\] By gathering all the desired material parameters to identify into the vector $\vek{q}=[\kappa~ G~ b_R~ b_{\vek{\chi}}~ \sigma_y]$, where $\kappa$ and $G$ are bulk and shear modulus, respectively, which determine the isotropic elasticity tensor, the goal is to estimate $\vek{q}$ given measurement displacement data, i.e. $$u=Y(\vek{q})+e \label{eq:pce_uf_yf},$$ in which $Y(\vek{q})$ represents the measurement operator and $e$ the measurement (also possibly the model) error. Being an ill-posed problem, the estimation of $\vek{q}$ given $u$ is not an easy task and usually requires regularization. This can be achieved either in a deterministic or a probabilistic setting. Here, the latter one is taken into consideration as further described in the text. Bayesian identification ======================= By using additional (prior) knowledge on the parameter set next to the observation data, the probabilistic approach regularizes the problem of estimating $\vek{q}$ with the help of Bayes’s theorem $$\label{eq:bayesrule} \pi_{q\|u}(\vek{q}\|u)\propto L(u\|\vek{q}) \pi_q(\vek{q}),$$ in which the likelihood $L(u\|\vek{q})$ describes how likely the measurement data are given prior knowledge $\pi_q(\vek{q})$. This in turn requires the reformulation of the deterministic model into a probabilistic one, and hence the propagation of material uncertainties through the model —the so-called forward problem— in order to obtain the likelihood [@Matthies0; @Matthies01]. The main difficulty in using Equation (\[eq:bayesrule\]) lies in the computation of the likelihood. Various numerical algorithms can be applied, the most popular example of which are the Markov chain Monte Carlo methods. Being constructed on the fundamentals of ergodic Markov theory, these methods are characterized by very slow convergence. To avoid this, an approximate method based on Kolmogorov’s definition of conditional expectation as already presented in [@Matthies2] is considered here. Let the material parameters $\vek{q}$ be modeled as random variables on a probability space $S:=L_2(\Omega,\mathcal{B},\mathbb{P})$. Here, $\Omega$ denotes the space of elementary events $\omega$, $\mathcal{B}$ is the $\sigma$-algebra and $\mathbb{P}$ stands for the probability measure. This alternative formulation of Bayes’s rule can be achieved by expressing the conditional probabilities in Equation (\[eq:bayesrule\]) in terms of conditional expectation. Following the mathematical derivation in [@Matthies2; @Matthies3; @Bojana; @Bojana1], this approach boils down to a quadratic minimization problem by considering the forecast random variable $q_f$ and the update of the forecast random variable $q_a$: $$\label{eq:minprob} \bar{q}^{\|z}=P_{\mathscr{Q}_{sn}} q_f=\underset{\eta\, \in \, \mathcal{Q}_{sn}}{\textrm{arg min }} \\|q_f-\eta\\|^2_{L_2} = \vek{\Xi} (u_f(\omega)),$$ where $P_{\mathcal{Q}_{sn}}$ is the orthogonal projection operator of $q_f$ onto the space of the new information $\mathcal{Q}_{sn}:=\mathscr{Q} \otimes S_n$ in which the space $S_n$ is the space of random variables generated by the measurement $u = Y(\vek{q}) + e$. Due to the Doob-Dynkin lemma, $\bar{q}^{\|z}$ is a function of the observation, where $u_f(\omega) = Y(q(\omega)) + e(\omega)$ is the forecast, and the assimilated value is $q_a(\omega) = q_f(\omega) + ( \vek{\Xi}(\hat{z}) - \vek{\Xi}(u_f(\omega)) )$. Constraining the space of all functions $\vek{\Xi}$ to the subspace of linear maps, the minimization problem in Equation (\[eq:minprob\]) leads to a unique solution $K$. Note that the projection is performed over a smaller space than $\mathcal{Q}_{sn}$. An implication of this is that available information is not completely used in the process of updating, introducing an approximation error. This gives an affine approximation of Equation (\[eq:minprob\]) $$\label{eq:gmarkovthm} q_a(\omega) = q_f(\omega) + K(\hat{z}-u_f(\omega)),$$ also known as a linear Bayesian posterior estimate or the so-called Gauss-Markov-Kalman filter (GMKF). Here, $q_f$ represents the prior random variable, $q_a$ is the posterior approximation, $u_f = Y(q_f(\omega)) + e(\omega)$ is the predicted measurement and $K$ represents the very well-known Kalman gain $$K:= C_{q_f u_f} \bigl(C_{u_f}+C_{\varepsilon}\bigr)^{-1}, \label{eq:kalman1}$$ which can be easily evaluated if the appropriate covariance matrices $C_{q_f u_f}$, $C_{u_f}$ and $C_{\varepsilon}$ are known. An advantage of Equation (\[eq:gmarkovthm\]) compared to Equation (\[eq:bayesrule\]) is that the inference in Equation (\[eq:gmarkovthm\]) is given in terms of random variables instead of conditional densities. Namely, $q_a(\omega)$, $q_f(\omega)$, and $u_f(\omega)$ denote the random variables used to model the posterior, prior, observation, and predicted observation, respectively. In this light the linear Bayesian procedure can be reduced to a simple algebraic method. Starting from the functional representation of the prior $$\hat{q}_f (\omega) = \sum_\alpha q_f^{(\alpha)} \psi_\alpha(\omega), \label{eq:q_f}$$ where $\psi_\alpha$ are multivariate Hermite polynomials, and by considering the proxy in Equation (\[eq:q\_f\]), one may discretize Equation (\[eq:gmarkovthm\]) as: $$\label{eq:lbunum} \vek{q}_a = \vek{q}_f +K\bigl(\hat{\vek{z}} - \vek{u}_f \bigr),$$ where $\vek{q}_a = [..., q^{(n)}_a,...]$, etc. are the PCE coefficients. As the measurement is a deterministic value, $\hat{z}= [\hat{z},0,...,0]$ has only a zero-th order tensor. The covariances for the Kalman gain Equation \[eq:kalman1\] are easily computed, e.g. $$C_{\hat{q}_f,\hat{u}_f} =\sum_{\alpha>0} \alpha ! \; q_f^{(\alpha)} (\vek{u}_f^{(\alpha)})^T.$$ Numerical results ================= The identification of the material constants in the Chaboche unified viscoplasticity model is a reverse process, here based on virtual data. In case of the Chaboche model the best way of parameter identification is using the results of the cyclic tests, since more information can be obtained from cyclic test rather than creep and relaxation tests, specifically information regarding hardening parameters. The aim of the parameter identification is to find a parameter vector $\vek{q}$ introduced in the previous section. The bulk modulus ($\kappa$), the shear modulus ($G$), the isotropic hardening coefficient ($b_{R}$), the kinematic hardening coefficient ($b_{\vek{\chi}}$) and the yield stress ($\sigma_y$) are considered as the uncertain parameters of the constitutive model. A preliminary study is on a regular cube, modeled with one 8 node element, completely restrained on the back face, and with normal traction on the opposite (front) face. Two cases are considered in order to compare the effect of applied force on identified parameters. For both cases the magnitude of the normal traction and a stress in the plane of the front face are plotted in Fig. \[fig:sigma-time\] and Fig. \[fig:sigma-time1\], respectively. Blue and red colors represent the stress value in normal and in plane directions, respectively. As it is seen, the magnitude of the applied force for the case 1 is constant all time but for the case 2 the magnitude of the applied force grows gradually by time.\ ![\[fig:sigma-time\] Decomposed applied force on desired node according to time- Case 1](pic/time-sigma1-sigma2){width="2.55in"} ![\[fig:sigma-time1\] Decomposed applied force on desired node according to time- Case 2](pic/time-sigma1-sigma2newf){width="2.55in"} Considering the parameters listed in Table \[tab:general11\], the related $\sigma$-$\epsilon$ hysteretic graph obtained for the applied force case 1 and 2 which can be seen in Fig. \[fig:sigma-epsilon\] and Fig. \[fig:sigma-epsilon1\], respectively. [l\[9]{}[c]{}r]{} & $\kappa$ & $G$ & $\sigma_y$ & $n$ & $k$ & $b_{R}$ & $H_{R}$ & $b_{\vek{\chi}}$ & $H_{\vek{\chi}}$ &\ & $1.66\mathrm{e}9$ & $7.69\mathrm{e}8$ &$1.7\mathrm{e}8$ & $1.0$ & $1.5\mathrm{e}8$ & $50$ & $0.5\mathrm{e}8$ & $50$ & $0.5\mathrm{e}8$ &\ \[tab:general11\] $\quad$ $\quad$ The displacements of one of the nodes on the front surface in normal and in plane directions are observed as the virtual data in this study. Applying the Gauss-Markov-Kalman filter with functional approximation as explained in the previous chapter and introducing measurement error in such a way that 15 percent of mean values are equal to the coefficient of variation for the related parameter, the probability density function (PDF) of prior and posterior of the identified parameters can be seen in Fig. \[fig:parameters\] and Fig. \[fig:parameters1\] for the first and second case, respectively. $\quad$ $\quad$ $\quad$ $\quad$ $\quad$ $\quad$ $\quad$ $\quad$ Summarizing the results, the true values $q_{\text{true}}$ and the mean and standard deviation of the estimated parameters, $\vek{q}^m_{\text{est}}$ and $\vek{q}^{\text{std}}_{\text{est}}$ respectively, for both cases are compared in Table \[tab:general41\]. [l\[6]{}[c]{}r ]{} Parameters & $\vek{q}_{\text{true}}$ & $\vek{q}^m_{\text{est-\text{1}}}$&$\vek{q}^{\text{std}}_{\text{est-\text{1}}}$ & $\vek{q}^m_{\text{est-\text{2}}}$&$\vek{q}^{\text{std}}_{\text{est-\text{2}}}$ &\ $\kappa$ & $1.66\mathrm{e}9$ & $1.66\mathrm{e}9$ & $1.13\mathrm{e}7$ & $1.66\mathrm{e}9$ & $2.59\mathrm{e}6$ &\ $G$ & $7.69\mathrm{e}8$ &$7.68\mathrm{e}8$ & $3.47\mathrm{e}6$ & $7.68\mathrm{e}8$ & $6.39\mathrm{e}5$ &\ $b_{R}$ & 50 & 52.36 & 3.71 & 50.27 & 0.29 &\ $b_{\vek{\chi}}$ & 50 & 52.04 & 3.01 & 50.19 & 0.53 &\ $\sigma_y$ & $1.7\mathrm{e}8$ & $1.69\mathrm{e}8$ & $1.35\mathrm{e}6$ & $1.69\mathrm{e}8$ & $1.52\mathrm{e}5$ &\ \[tab:general41\] Discussion of the results ------------------------- From the sharpness of the posterior PDF of $\kappa$, $G$ and $\sigma_{y}$, it can be concluded that enough information from virtual data is received and updating the parameters considering their uncertainty is done very properly for the both cases, as the standard deviation of the residual uncertainty is below $1\%$ of the mean. For the posterior PDF of $b_{R}$ and $b_{\vek{\chi}}$, it can be inferred that better updating is done for the second case compared to the first case. Not only are the more accurate estimations of the exact hardening parameters, $b_{R}$ and $b_{\vek{\chi}}$, predicted for the second case, but the uncertainty of the estimated hardening parameters are also reduced much more for the second case. One reason that can be mentioned is that the process is not always in the states that hardening equations are involved like the elastic states. Therefore less information from the whole simulation can be analyzed for estimating the hardening parameters and updating their parameters’ uncertainties. Fig. \[fig:VM\] and Fig. \[fig:VM1\] prove this fact that since more states are out of the von Mises yield criterion for the second case compared to the first case, in which the hardening equations are involved only in these states, the better identification can be done for the second case, where a gradually varying increasing applied force is considered, for hardening parameters in comparison with the first case where a constant magnitude applied force is employed. In fact, the cyclic applied force in the second case causes the more activation of the desired parameters in the studied set of equations comparing to the first case and accordingly a better determination of the parameters can be carried out for the second case. It should be pointed out that the von Mises yield criterion is illustrated by the green cylinder in the mentioned figures i.e. inside and outside of the cylinder respectively refer to the elasticity and plasticity states, respectively. Also the blue color represents the principal stresses in these figures. $\quad$ $\quad$ Summary ======= Using the Gauss-Markov-Kalman Filter method explained in Section 3 to identify the model parameters of the Choboche model indicates that it is possible to identify the model parameters by employing this method using functional approximation. The parameters are well estimated and the uncertainty of the parameters is reduced while the random variables of the parameters are updated during the process [@adeli]. The other conclusion that can be made is that the more information we receive, the better parameter identification we can do using the Gauss-Markov-Kalman filter method. This fact is observed by comparing the posterior probability density functions of hardening parameters, $b_{R}$ and $b_{\vek{\chi}}$ , for case 1 and 2. Therefore in terms of mechanical models, it should be always considered that the applied force should be applied in such a way that all time all equations should be involved. In other words, the applied load path should lead to activation of all uncertain parameters in the set of equations, as here a cyclic gradually varying increasing applied force leads to a better determination of the parameters. Otherwise only the parameters which are in the involved equations are updated.\ Acknowledgement*\ This work is partially supported by the DFG through GRK 2075. [10]{} A. Miller. An Inelastic Constitutive Model for Monotonic, Cyclic, and Creep Deformation: Part I–Equations Development and Analytical Procedures. J. Eng. Mater. Technol. 98(2), 97–105 (1976). E. Krempl, J. J. McMahon, and D. Yao. Viscoplasticity Based on Overstress with a Differential Growth Law for the Equilibrium Stress. Mechanics of Materials 5, 35–48 (1986). R. K. Korhonen, M. S. Laasanen, J. Toyras, R. Lappalainen, H. J. Helminen, and J. S. Jurvelin. Fibril reinforced poroelastic model predicts specifically mechanical behavior of normal, proteoglycan depleted and collagen degraded articular cartilage. J Biomech 36, 1373–1379. Michel Aubertin, Denis E. Gill, and Branko Ladanyi. A unified viscoplastic model for the inelastic flow of alkali halides. Mechanics of Materials 11, 63–82 (1991). K. S. Chan, S. R. Bodner, A.F. Fossum, and D.E. Munson. A constitutive model for inelastic flow and damage evolution in solids under triaxial compression. Mechanics of Materials 14, 1–14 (1992). J. L. Chaboche, and G. Rousselier. On the plastic and viscoplastic constitutive equations - part 1: rules developed with internal variable concept. J. Press. Vessel Technol., 105, 153–158 (1983). J. L. Chaboche, and G. Rousselier. On the plastic and viscoplastic constitutive equations - part 2: application of internal variable concepts to the 316 stainless steel. J. Press. Vessel Technol., 105, 159–164 (1983). P. Kłosowski, and A. Mleczek. Parameters’ Identification of Perzyna and Chaboche Viscoplastic Models for Aluminum Alloy at Temperature of $120^{\grad}C$. Engng. Trans. 62, 3, 291–305 (2014). Y. Gong, C. Hyde, W. Sun, and T. Hyde. Determination of material properties in the Chaboche unified viscoplasticity model. Journal of Materials Design and Applications, 224(1), 19–29 (2010). T. Harth, and Jürgen Lehn. Identification of Material Parameters for Inelastic Constitutive Models Using Stochastic Methods. GAMM-Mitt. 30, No. 2, 409–429 (2007). K. S. Chan, S. R. Bodner, and U. S. Lindholm. Phenomenological Modelling of Hardening and Thermal Recovery in Metals. Journal of Engineering Materials and Technology 110, 1–8 (1988). J. Velde. 3D Nonlocal Damage Modeling for Steel Structures under Earthquake Loading. Department of Architecture, Civil Engineering and Environmental Sciences University of Braunschweig - Institute of Technology (2010). H. G. Matthies. Stochastic finite elements: Computational Approaches to Stochastic Partial Differential Equations. Journal of Journal of Applied Mathematics and Mechanics, volume 88, 849–873 (2008). H. G. Matthies. Uncertainty Quantification with Stochastic Finite Elements. Encyclopedia of Computational Mechanics, edited by E. Stein, R. de Borst, T. R. J. Hughes, John Wiley and Sons, Chichester, (2007). H. G. Matthies, E. Zander, B. V. Rosić, and A. Litvinenko. Parameter Estimation via Conditional Expectation: A Bayesian Inversion. Journal of Advanced Modeling and Simulation in Engineering Sciences, 3–24 (2016). B. V. Rosić, A. Litvinenko, O. Pajonk, H. G. Matthies. Sampling-free linear Bayesian update of polynomial chaos representations. Journal of Computational Physics 231(17):5761–5787, (2012). H. G. Matthies, E. Zander, B. V. Rosić, A. Litvinenko, and Oliver Pajonk. Inverse Problems in a Bayesian Setting. Journal of Computational Methods for Solids and Fluids, volume 41, 245–286 (2016). B. V. Rosić and H. G. Matthies. Identification of Properties of Stochastic Elastoplastic Systems. Computational Methods in Stochastic Dynamics, 237–253, (2013). S. Ghahramani. Fundamentals of probability, with stochastic processes. 3rd ed. New Jersey, USA: Pearson, Prentice Hall (2005). E. Adeli, B. V. Rosić, H. G. Matthies and S. Reinst[ä]{}dler. Bayesian parameter identification in plasticity. XIV International Conference on Computational Plasticity. Fundamentals and Applications COMPLAS XIV E. Oñate, D.R.J. Owen, D. Peric and M. Chiumenti (Eds), 247–256, DOI: 10.13140/RG.2.2.22323.89124, (2017).	{ "pile_set_name": "ArXiv" }
--- author: - \| Cheng Li\ \ \ Xiaoxiao Guo\ \ \ Qiaozhu Mei\ \ \ bibliography: - 'attitude\_analysis\_bib.bib' title: Deep Memory Networks for Attitude Identification --- =10000 = 10000	{ "pile_set_name": "ArXiv" }
--- abstract: 'We report a new class of $SO(3,\mathbb{C})$ and diffeomorphism invariant formulations for general relativity with either a vanishing or a nonvanishing cosmological constant, which depends functionally on a $SO(3,\mathbb{C})$ gauge connection and a complex-valued 4-form via a holomorphic function of the trace of a symmetric $3\times3$ matrix that is constructed from these variables. We present two members of this class, one of which results from the implementation of a method for obtaining action principles belonging to the class. For the case of a nonvanishing cosmological constant, we solve for the complex-valued 4-form and get pure connection action principles. We perform the canonical analysis of the class. The analysis shows that only the Hamiltonian constraint is modified with respect to the Ashtekar formulation and that the members of the class have two physical degrees of freedom per space point.' author: - Diego González - Merced Montesinos title: Gauge connection formulations for general relativity --- Introduction ============ The gauge connection formulations for classical general relativity (GR) are relatively new ways of thought that open, in principle, new roads toward the quantization of the gravitational field [@quantum]. In these formulations, the gauge connection is the main structure used to describe the gravitational field, while the metric is a derived object. A remarkable feature that makes them attractive is that the field equations can be substantially simpler than the ones emerging directly from Einstein’s original formulation based on the metric field. Among these approaches we find the one due to Plebanski, who showed that GR can be expressed as a BF theory supplemented with a constraint on the B fields. This formulation involves a $SO(3,\mathbb{C})$ gauge connection, a $SO(3,\mathbb{C})$ 2-form, a $SO(3,\mathbb{C})$ scalar field, and a Lagrange multiplier [@plebanski]. Based on such a work, Capovilla, Dell, and Jacobson (CDJ) went one step further and presented a pure spin connection formulation for GR without a cosmological constant, where the variables involved are a gauge connection and a scalar density [@CJD1]. The generalization of this formulation to the case including a nonvanishing cosmological constant has shown to be a nontrivial task. Indeed, the same authors attempted to carry this out [@CJD2], however, erroneously [@erratum]. A correct action principle involving the cosmological constant was derived later by Capovilla and Jacobson [@CJ3] together with Peldan [@Peldan], using totally different methods. Unfortunately, this action principle has not been widely applied because of some technical aspects that prevent its handling. Recently, Krasnov has achieved a different action principle for GR that requires a nonvanishing cosmological constant and that depends functionally on a $SO(3)$ gauge connection only [@Kras1]. In this paper, we report a new class of gauge connection formulations for GR, in the Lorentzian signature case, which has as fundamental variables a $SO(3,\mathbb{C})$ gauge connection and a complex-valued 4-form. We explore the members of this class and find a new action principle for GR that works well with or without a cosmological constant. Furthermore, considering some caveats that we will clarify later on, a particular case of this member can be related to the action principle found in Ref. [@erratum]. We also develop a method for constructing the members of the class by integrating certain holomorphic functions. This method is illustrated with a new action principle for GR with a nonvanishing cosmological constant. We also derive pure connection action principles from the class with a nonvanishing cosmological constant, by eliminating the dependence of the action on the auxiliary 4-form. Finally, we develop the canonical analysis of the class and show that the Hamiltonian constraint is modified with respect to the Ashtekar formulation, whereas the Gauss and vector constraints remain unchanged. The class has two complex degrees of freedom per space point. The class of formulations for general relativity {#class} ================================================ Let us begin by fixing the notation and convention. The fundamental variables considered in this paper are a $SO(3,\mathbb{C})$ gauge connection $A^i$ with curvature $F^i=d A^i + \frac{1}{2} {\varepsilon}^i{}_{jk} A^j \wedge A^k$ and a nonvanishing complex-valued 4-form $\rho$. The indices $i,j,k=1,2,3$ are raised and lowered with the Kronecker delta $\delta_{ij}$ and ${\varepsilon}_{ijk}$ is the Levi-Cività symbol (${\varepsilon}_{123}=+1$). The wedge product of forms is denoted by $\wedge$. Now, let us define a complex $3\times3$ matrix $\Psi(A^i,\rho)$, which is a function of $A^i$ and $\rho$ via $$\begin{aligned} F^i\wedge F^j +2 \rho X^i{}_k X^{jk}=0, \label{FF-cond}\end{aligned}$$ where $X\equiv\Psi+(1/3)\Lambda I$. Here, $\Lambda$ is the usual cosmological constant and $I$ is the identity $3\times3$ matrix. From now on we restrict the analysis to configurations such that $X$ is symmetric and nonsingular, as is usual in the pure spin connection formulations for GR [@CJD2; @CJ3]. This restriction is needed in order to hold the equivalence between our formulation and Plebanki’s equations of motion for GR. Because $X$ is nonsingular, then the symmetric density matrix $\tilde{M}$ of weight 1 and defined by $\tilde{M}^{ij}d^4x=F^i\wedge F^j$ is also nonsingular on account of (\[FF-cond\]). Since $X$ is symmetric, then (\[FF-cond\]) becomes ${\tilde M}=-2 {\tilde \mu} X^2$ where $\rho = {\tilde \mu} d^4x$. The action principles considered in this paper depend on ${\rm Tr} \Psi \equiv \Psi^{ij} \delta_{ij}$; therefore, we need to compute $X$ from ${\tilde M}=-2 {\tilde \mu} X^2$, next compute $\Psi$, and finally compute ${\rm Tr} \Psi $. The solution for $X$ exists [@book], but is not unique generically [@book2]. Nevertheless, the action principles introduced below work well for any of these solutions. In the pure spin connection formulations for GR there is a debate about whether an additional criterion must be introduced in order to select just one or whether all the solutions must be allowed to describe GR [@CJ3; @Peldan]. Having defined $\Psi(A^i,\rho)$, we are ready to give the action principle $$\begin{aligned} S[A^i,\rho]=\int \rho f({\rm Tr}\Psi),\label{Action}\end{aligned}$$ where $f$ is a given holomorphic function that depends on ${\rm Tr}\Psi$ and that has the same dimensions of the cosmological constant. It is worth pointing out that $\Psi^{ij}$ is a tensor density of weight zero, and hence any function $f$ is also of weight zero and leaves the action (\[Action\]) correctly defined. Then it makes sense to consider an arbitrary holomorphic function $f$. However, an arbitrary choice of $f$ may lead to other theories of gravity different from GR. In this paper we emphasize that we only consider functions $f$ such that (\[Action\]) is an action for GR. The other possible cases are also interesting and will be treated elsewhere. We find that GR emerges from the class of formulations (\[Action\]) if $f$ is a holomorphic function in a domain $\Omega$ in $\mathbb{C}$ to which zero belongs, and if it satisfies the following properties: (i) The only zero of $f- \frac{1}{2} \left( {\rm Tr}\Psi + \Lambda \right) f'$, as a function of ${\rm Tr}\Psi$, is ${\rm Tr}\Psi=0$; (ii) $f'_0\equiv\left. f' \right\|_{{\rm Tr}\Psi=0} \neq 0$; where “$\prime$" denotes the derivative with respect to ${\rm Tr}\Psi$. The requirement of conditions (i) and (ii) will be evident below. Now we prove that given a function $f$ satisfying (i) and (ii), the class of formulations (\[Action\]) describes GR with and without a cosmological constant. To do this, it is useful to begin by determining the equations of motion of the action (\[Action\]), in the case in which $f$ is an arbitrary holomorphic function. These equations are given by $$\begin{aligned} \delta \rho: && f- \frac{1}{2} \left( {\rm Tr}\Psi + \Lambda \right) f'=0, \label{mot-rho}\\ \delta A^i:&& D\left[-\frac{f'}{2} (X^{-1})^i{}_j F^j \right] =0,\label{mot-con}\end{aligned}$$ where $D$ is the covariant derivative with respect to $A^i$. Here we have used our assumption that $X$ is symmetric and nonsingular. To compute the variations of ${\rm Tr}\Psi$ with respect to $\rho$ and $A^i$, we employ Eq. (\[FF-cond\]) to get $\rho\delta{\rm Tr}\Psi= -\frac{1}{2} {\rm Tr}X \delta\rho$ and $\rho\delta{\rm Tr}\Psi= -\frac{1}{2} (X^{-1})_{ij} F^i\wedge \delta F^j$, respectively. To simplify Eqs. (\[mot-rho\]) and (\[mot-con\]), we now consider that $f$ is endowed with the desired properties (i) and (ii). Since $f$ satisfies condition (i), equation of motion (\[mot-rho\]) implies $$\begin{aligned} {\rm Tr}\Psi=0. \label{cond-1}\end{aligned}$$ Notice that (\[cond-1\]) is equivalent to the constraint ${\rm Tr}X=\Lambda$, which also appears in the pure spin connection formulation of Ref. [@CJ3]. We remark that in our approach, as well as in Ref. [@CJ3], the appropriate value for ${\rm Tr}\Psi$ comes from the equation of motion for the auxiliary field $\rho$, and that a different value of ${\rm Tr}\Psi$ will lead to other theories of gravitation. This is one reason why the field $\rho$ is relevant. Moreover, note that $X^{ij}$ and $f'$ have weight zero, and hence the term in the argument of the covariant derivative of Eq. (\[mot-con\]) is actually a well-defined 2-form. Furthermore, this 2-form must be evaluated at ${\rm Tr}\Psi=0$ because of (\[cond-1\]), and, at this point, it does not vanish trivially because $f$ satisfies the property (ii). Therefore, the equation of motion (\[mot-con\]) acquires the form $$\begin{aligned} D\Sigma^i=0, \label{PlebConn}\end{aligned}$$ in which the 2-form $\Sigma^i$ is given by $$\begin{aligned} \Sigma^i= (X^{-1})^i{}_j F^j. \label{sigma}\end{aligned}$$ Note that $f'_0$ does not appear in (\[PlebConn\]), since it is a nonvanishing dimensionless constant. Now it is clear the role of conditions (i) and (ii). The remarkable fact is that Eqs. (\[cond-1\]) and (\[PlebConn\]) along with the definition (\[sigma\]) are Einstein’s equations for GR. Indeed, combining (\[FF-cond\]) with (\[sigma\]), we obtain the Plebanski constraints $$\begin{aligned} \Sigma^i \wedge \Sigma^j + 2 \rho \delta^{ij}=0. \label{PlebSigma}\end{aligned}$$ This means that the $\Sigma$’s are actually those of the Plebanski formulation for general relativity. Furthermore, notice that $\Sigma^i \wedge \Sigma_i\neq0$ since $ \rho\neq0$. Then, Eq. (\[PlebSigma\]) together with the reality conditions $$\begin{aligned} \Sigma^i \wedge {\overline \Sigma}^j = 0, \hspace{7mm} \Sigma^i \wedge \Sigma_i + {\overline \Sigma}^i \wedge {\overline \Sigma}_i =0,\end{aligned}$$ imply $\Sigma^i = {{\rm i }}\theta^0 \wedge \theta^i - \frac{1}{2} \varepsilon^i{}_{jk} \theta^j \wedge \theta^k$, where $\{ \theta^0 , \theta^i \}$ are four linearly independent real 1-forms and ${{\rm i }}$ is the imaginary unit. The overbar denotes the complex conjugate. We recall that the reality conditions do not come from the formulation (\[Action\]), but they are introduced by hand in order to relate the complex 2-forms $\Sigma^i$ with real 1-forms [@plebanski; @CDJ4]. On the other hand, Eq. (\[PlebConn\]) is a system of 12 linear equations for the 12 unknowns contained in the components of $A^i$. For Plebanski’s $\Sigma^i$ this system is nondegenerate and can be solved for $A^i=A^i(\Sigma)$ [@GMV]; however, further assumptions are needed to link $A^i$ with a spacetime connection because there are no spacetime geometrical structures involved in the class of gauge connection formulations (\[Action\]) nor in any other gauge connection formulation [@GMV]. According to Levi-Cività, a spacetime connection is uniquely defined by specifying its torsion and its action on a spacetime metric. Therefore, the first assumption consists in defining the spacetime metric, which is given by the Urbantke metric in terms of $\Sigma^i$ [@urbantke], namely, $(1/12) \varepsilon_{ijk} \Sigma^i{}_{MI} \Sigma^j{}_{JK} \Sigma^k{}_{LN} {\tilde \eta}^{IJKL}$, where $I,J,\ldots=0,1,2,3$ and ${\tilde \eta}^{IJKL}$ is a totally antisymmetric density of weight 1 (${\tilde \eta}^{0123}=1$). A direct calculation using the expression for $\Sigma^i$ in terms of the real 1-forms shows that this metric turns out to be the Minkowski metric. The second assumption is that the spacetime connection has no torsion. As a result, the connection $A^i$ is the self-dual part of the spin connection (i.e., the Levi-Cività spacetime connection) and Eq. (\[PlebConn\]) becomes the first Cartan’s structure equation with vanishing torsion. This in turn implies that $F^i$ is the self-dual part of the curvature of the spin connection. Next, $F^i$ can be solved from (\[sigma\]), $$\begin{aligned} F^i= \Psi^i{}_j \Sigma^j +\frac{1}{3} \Lambda \Sigma^i , \label{PlebCurv}\end{aligned}$$ which means that $F^i$ is self-dual as a 2-form. Finally, Eqs. (\[cond-1\]) and (\[PlebCurv\]) together with the relation between $F^i$ and curvature of the spin connection imply Einstein’s equations for GR. Indeed, Eqs. (\[cond-1\]), (\[PlebConn\]), (\[PlebSigma\]), and (\[PlebCurv\]) are exactly the set of equations of motion of the Plebanski formulation for complex GR with a cosmological constant. More details on the Plebanski formulation can be obtained in Refs. [@GMV; @Kras2]. In addition, the trace-free symmetric matrix $\Psi$ becomes the self-dual part of the Weyl tensor. Example: polynomial functions ============================= To illustrate how the class of formulations (\[Action\]) works, we present the simple, but rather significant, case of a quadratic function of the ${\rm Tr}\Psi$. To this end, let us begin by considering the function $$\begin{aligned} f =\alpha_1 (2 {\rm Tr}\Psi+\Lambda)+ \alpha_2 ({\rm Tr}\Psi+\Lambda)^2, \label{fquadratic}\end{aligned}$$ where $\alpha_1$ and $\alpha_2$ are arbitrary constants with appropriate units. Our goal is to demonstrate that, with a specific choice of $\alpha_1$ and $\alpha_2$, the function (\[fquadratic\]) is holomorphic on a certain region $\Omega$ containing zero and satisfying conditions (i) and (ii). The first requirement is satisfied because (\[fquadratic\]) is a polynomial function, and hence holomorphic in the whole complex plane. Conditions (i) and (ii) are satisfied if $\alpha_1\neq0$ and $\alpha_1+\alpha_2\Lambda\neq0$, respectively. Note that these conditions are fulfilled for $\Lambda=0$ as well as for $\Lambda\neq0$. Indeed if, for instance, $\Lambda=0$, then both conditions reduce to $\alpha_1\neq0$. In view of this, we can conclude that the new action principle (\[Action\]) with $f$ given by (\[fquadratic\]), where $\alpha_1\neq0$ and $\alpha_1+\alpha_2\Lambda\neq0$, has the equations of motion (\[cond-1\]) and (\[PlebConn\]), and therefore it describes GR with or without a cosmological constant $\Lambda$. For the special case of the choice $\alpha_2=0$, Eq. (\[fquadratic\]) is a linear function in ${\rm Tr}\Psi$ which, together with the action (\[Action\]), leads to $$\begin{aligned} S_{\rm{linear}}[A^i,\rho]=\alpha_1 \int \rho (2 {\rm Tr}X-\Lambda), \label{slinear}\end{aligned}$$ where we use $\Psi=X-(1/3)\Lambda$. This action, which also works well for both $\Lambda=0$ and $\Lambda\neq0$, corresponds to the simplest case of an action belonging to our class of formulations. Notice that (\[slinear\]) has the same form of the action principle (3) of Ref. [@erratum]. It is important to point out that the action principle (\[slinear\]) and the one of Ref. [@erratum] were obtained following different approaches. In fact, Eq. (3) in Ref. [@erratum] was obtained from the Plebanski action. In the case $\alpha_2\neq0$, the quadratic term in (\[fquadratic\]) remains and gives a different action principle with the same equations of motion of the case $\alpha_2=0$. Moreover, the action with the quadratic term has no analog with any other gauge connection formulation for GR. Note that $\alpha_2$ is an arbitrary parameter for $\Lambda=0$ and $\Lambda\neq0$ (except $\alpha_2\neq-\alpha_1/\Lambda$), and that appears in the action principle but not in the equations of motion. Then, in this sense, $\alpha_2$ resembles the Barbero-Immirzi parameter involved in the Holst action [@holst]. On the other hand, it can be shown that there are not polynomial functions on ${\rm Tr}\Psi$ with degree three or greater satisfying conditions (i) and (ii) either with $\Lambda=0$ or with $\Lambda\neq0$. Then, the quadratic function (\[fquadratic\]), under the mentioned conditions on $\alpha_1$ and $\alpha_2$, is the most general case for a viable polynomial function on ${\rm Tr}\Psi$ for GR. A method to obtain $\pmb{f({\rm Tr}\Psi)}$ ========================================== In addition to the quadratic function, there are many other functions satisfying conditions (i) and (ii). However, instead of giving another example, we shall briefly present a method to obtain such viable functions for GR, [i.e.]{}, a method to construct action principles belonging to our class of formulations (\[Action\]). Consider a holomorphic function $h({\rm Tr}\Psi)$ in the domain $\Omega$, which has only a zero of order $\kappa$ at ${\rm Tr}\Psi=0$. Furthermore, it is well known [@libro] that for such $h$, a holomophic function $q({\rm Tr}\Psi)$ exists in $\Omega$, which has no zeros and such that $h=({\rm Tr}\Psi)^\kappa q$. This suggests that it is convenient to construct $f$ from $h$ via the complex differential equation $$\begin{aligned} f- \frac{1}{2} \left( {\rm Tr}\Psi + \Lambda \right) f' = ({\rm Tr}\Psi)^\kappa q. \label{diffeq}\end{aligned}$$ The reason to consider this particular equation is that its solution, the function $f$, is holomorphic in $\Omega$ and satisfies condition (i). Indeed, the solution is found to be $$\begin{aligned} f= ({\rm Tr}\Psi+\Lambda)^2 (p+\alpha) , \label{sol-diff-eq}\end{aligned}$$ where the function $p({\rm Tr}\Psi)$ is a primitive of $-2({\rm Tr}\Psi)^\kappa q/({\rm Tr}\Psi + \Lambda)^3$ and $\alpha$ is an arbitrary constant. It can be verified that the term $ ({\rm Tr}\Psi+\Lambda)^2p$ is holomorphic at $-\Lambda$, despite the fact that $p$ is not. It can also be checked by direct calculation that (\[sol-diff-eq\]) satisfies condition (i). For instance, the quadratic function (\[fquadratic\]) arises from the simplest case of $h$, namely, $\kappa=1$ and $q=\alpha_1$ (=const). From this point of view, $\alpha_2=\alpha$ is actually an integration constant that is not necessarily equal to zero. Now, it is convenient to express condition (ii) in terms of $({\rm Tr}\Psi)^\kappa q$. Then, using (\[sol-diff-eq\]) to calculate $f'_0$, condition (ii) becomes $$\begin{aligned} \lim\limits_{{\rm Tr}\Psi \rightarrow 0} \left[ \frac{ ({\rm Tr}\Psi)^\kappa q }{ {\rm Tr}\Psi + \Lambda} - \left( {\rm Tr}\Psi + \Lambda \right) p \right] \neq \alpha\Lambda. \label{condii}\end{aligned}$$ In particular, if $\Lambda=0$, the only possible case satisfying condition (ii) is $\kappa=1$. In such a case, Eq. (\[condii\]) reduces to $q_0 \equiv q \|_{{\rm Tr}\Psi=0} \neq0$, which is satisfied by assumption. For $\Lambda\neq 0$, the first term on the left-hand side of (\[condii\]) vanishes, and therefore condition (ii) is equivalent to $p_0 \equiv p \|_{{\rm Tr}\Psi=0} \neq-\alpha$. In this case all possible values of $\kappa$ are allowed. At this point the reader may have noticed that there are many suitable functions for GR that can be found by following this approach. For example, consider the function $h=\beta {\rm Tr}\Psi \exp{( 2{\rm Tr}\Psi /\Lambda)}$ with $\beta$ a nonzero constant and $\Lambda\neq0$. This function is clearly holomorphic on $\mathbb{C}$ and has a zero at ${\rm Tr}\Psi=0$, as it is desirable. Thereby, Eq. (\[sol-diff-eq\]) constructed using $h$ directly satisfies condition (i). Indeed, Eq. (\[sol-diff-eq\]) gives the function $$\begin{aligned} f = - \Lambda \beta \exp(2 {\rm Tr}\Psi/\Lambda) + \alpha ({\rm Tr}\Psi+\Lambda)^2, \label{fexp}\end{aligned}$$ which satisfies condition (i) provided that $\beta\neq0$. It remains to test whether (\[fexp\]) satisfies condition (ii). To do this we can proceed in two different ways. One way is by verifying condition (ii) directly using (\[fexp\]). The second one is through the use of (\[condii\]) with $h$. From any of these ways it follows that condition (ii) is satisfied if $p_0=-\beta/\Lambda\neq-\alpha$. Then, the upshot of our example is a new action for GR with $\Lambda\neq0$, namely, the action (\[Action\]) with $f$ given by (\[fexp\]), provided that $\beta\neq0$ and $\beta\neq\alpha\Lambda$. To close the analysis of the method described above we can, alternatively, remark that the class of formulations (\[Action\]) with $f$ defined by the function (\[sol-diff-eq\]) describes GR provided that the holomorphic function $h({\rm Tr}\Psi)$ has only one zero at ${\rm Tr}\Psi=0$ and satisfies condition (\[condii\]). Pure connection formulation =========================== For the case $\Lambda\neq0$ of the class of formulations (\[Action\]) it is possible to eliminate the auxiliary field $\rho$, and then get formulations that depend on the gauge connection only. The aim of this section is to derive such pure connection formulations. We follow a procedure with the same logic as that of Refs. [@CJD2] and [@CJ3]. That is, the elimination of the field $\rho$ is achieved by using its equation of motion. Consider that $\Lambda\neq0$ and that $f$ satisfies conditions (i) and (ii). First, we want to express the equation of motion (\[cond-1\]) in terms of ${\tilde M}$, ${\tilde \mu}$, and $\Lambda$. To do this we make use of the fact that ${\tilde M}=-2 {\tilde \mu} X^2$ can be solved for $X$, and then we use $\Psi=X-(1/3)\Lambda I$ to get ${\rm Tr}\Psi$. Then the Eq. (\[cond-1\]) reads $$\begin{aligned} \pm \frac{1}{(-2{\tilde \mu})^{1/2} } {\rm Tr}{\tilde M}^{1/2}-\Lambda=0, \label{mov-rho}\end{aligned}$$ where $\tilde{M}^{1/2}$ is a symmetric square root of $\tilde{M}$, that is, $\tilde{M}=\tilde{M}^{1/2} \tilde{M}^{1/2}$ and $\tilde{M}^{1/2}=(\tilde{M}^{1/2})^T$. Recall that such $\tilde{M}^{1/2}$ exists since $\tilde{M}$ is symmetric and nonsingular [@book]. Now, it is straightforward to solve (\[mov-rho\]) for ${\tilde \mu}$, $$\begin{aligned} {\tilde \mu} = -\frac{1}{2 \Lambda^2} \left( {\rm Tr}{\tilde M}^{1/2} \right)^2. \label{mu}\end{aligned}$$ With this solution, it is time to get the pure connection actions. Substituting (\[cond-1\]) and (\[mu\]) into (\[Action\]), we obtain $$\begin{aligned} S[A^i]=-\frac{1}{4} \frac{f'_0}{\Lambda} \int \left( {\rm Tr}{\tilde M}^{1/2} \right)^2 d^4x,\label{Actionpure}\end{aligned}$$ where we have used $f_0\equiv\left. f \right\|_{{\rm Tr}\Psi=0}=\frac{1}{2}\Lambda f'_0$ that comes from the fact that $f$ satisfies condition (i). The outcome is that all the members with $\Lambda\neq0$ belonging to class (\[Action\]) reduce to the pure connection action principles (\[Actionpure\]), which share the same dependence on $A^i$, but the factor in front of the action has a dependence on $\Lambda$ through $f'_0/\Lambda$; thus there is a reminiscence of $f$ on this factor. This $\Lambda$ dependence is different from that of Ref. [@Kras1]. Recall that $f'_0$ does not vanish since $f$ satisfies condition (ii), and that in general it involves the cosmological constant. For instance, if we consider the functions (\[fquadratic\]) and (\[fexp\]) the factor is $f'_0=2(\alpha_1+\alpha_2\Lambda)$ and $f'_0=2(-\beta+\alpha \Lambda)$, respectively. Furthermore, we also remark that the procedure employed in the derivation of (\[Actionpure\]) is different from the one followed in Ref. [@Kras1]. An advantage of the class (\[Action\]) is that the resulting actions (\[Actionpure\]) have a very compact form, in contrast to the case of the pure connection formulation for GR with $\Lambda\neq0$ analyzed in Refs. [@CJ3] and [@Peldan]. Indeed, in Ref. [@Peldan] it is pointed out that the elimination of the auxiliary field from the formulation of the references aforementioned seems to lead to a “terrible expression” for the Lagrangian and that for such a reason that calculation was not completely done. For the treatment of the $\Lambda=0$ case, it is natural to try to apply the same strategy. However, this time the field $\rho$ cannot be solved from (\[cond-1\]), which is easy to see by setting $\Lambda=0$ in (\[mov-rho\]). Hence (\[Action\]), like the CDJ action [@CJD1], fails to reduce to pure connection action principles in this case. Canonical analysis ================== In Ref. [@Dadh] it was shown that the particular action (\[slinear\]) with $\Lambda\neq0$ leads to the usual constraints of GR in terms of the Ashtekar variables. In this section, we perform the canonical analysis of the class of formulations presented in Sec. \[class\]. By performing the spacetime decomposition we rewrite the class of actions (\[Action\]) as $$\begin{aligned} S=\int dt d^3x \tilde{\mathcal{L}}=\int dt d^3x {\tilde \mu} f({\rm Tr}\Psi),\label{Action31}\end{aligned}$$ and Eq. (\[FF-cond\]) as $$\begin{aligned} F^{(i}{}_{0a}F^{j)}{}_{bc}\tilde{\eta}^{abc} +2 {\tilde \mu} X^i{}_k X^{jk}=0, \label{FF-cond31}\end{aligned}$$ where $F^{i}{}_{0a}=\dot{A}^i{}_a-D_aA^i{}_0$ and $F^{i}{}_{ab}=\partial_a A^i{}_b-\partial_b A^i{}_a+\varepsilon^{i}{}_{jk}A^j{}_aA^k{}_b$. Here $D_a$ is the covariant derivative corresponding to the spatial connection $A^i{}_a$ and $a,b,c=1,2,3$ are spatial indices. We define the momenta conjugate to $A^i{}_a$ as $$\begin{aligned} \tilde{\pi}^a_i:=\frac{\partial \tilde{\mathcal{L}}}{\partial \dot{A}^i{}_a}=-\frac{f'}{4}(X^{-1})_{ij} \tilde{B}^{aj},\label{momentum}\end{aligned}$$ where $\tilde{B}^a_i=F_{ibc}\tilde{\eta}^{abc}$ is the corresponding “magnetic” field. We will restrict our analysis to nondegenerate magnetic fields, that is, $\det\tilde{B}:=(1/3!)\varepsilon^{ijk}\underaccent{\tilde}{\eta}{}_{abc} \tilde{B}^a_i \tilde{B}^b_j \tilde{B}^c_k\neq 0$. This restriction implies that $\det\tilde{\pi}:=(1/3!)\varepsilon^{ijk}\underaccent{\tilde}{\eta}{}_{abc} \tilde{\pi}^a_i \tilde{\pi}^b_j \tilde{\pi}^c_k=-(f'/4)^3 \det\tilde{B}/\det X\neq 0$ since $f'_0\neq 0$ and $\det X\neq 0$. The phase space variables are $A^i{}_a$ and $\tilde{\pi}^a_i$, whereas $A^i{}_0$ and $\tilde{\mu}$ are nondynamical variables, since they have vanishing conjugate momenta. The Gauss constraint follows from the fact that there are no time derivatives of the variables in the spatial projection of the equation of motion (\[mot-con\]). The constraint then reads $$\begin{aligned} \tilde{\mathcal{G}}_i\equiv D_a \tilde{\pi}^a_i \approx 0. \label{gauss}\end{aligned}$$ The vector constraint can be obtained from the requirement that $X_{ij}$ is symmetric. Indeed, this constraint turns out to be $$\begin{aligned} \tilde{\mathcal{V}}_b\equiv\tilde{\pi}^a_i F^i{}_{ab}=-\frac{f'}{4}(X^{-1})_{ij} F^i{}_{ab} \tilde{\eta}^{acd} F^j{}_{cd} = 0, \label{vector}\end{aligned}$$ where it can be checked that $F^i{}_{ab} \tilde{\eta}^{ade} F^j{}_{de}=-F^j{}_{ab} \tilde{\eta}^{ade} F^i{}_{de}$. Notice that the resulting Gauss and vector constraints of the class of formulations are still the same as those for the usual GR in terms of the Ashtekar variables. The Hamiltonian constraint is derived from the equation of motion (\[cond-1\]). Explicitly, using ${\rm Tr}\Psi={\rm Tr}X-\Lambda$ and multiplying by $\det\tilde{\pi}$ we can rewrite (\[cond-1\]) as $$\begin{aligned} \left({\rm Tr}X-\Lambda\right)\det\tilde{\pi} =0. \label{scalarTr}\end{aligned}$$ Using now the expression for the momenta $\tilde{\pi}^a_i$, we get $$\begin{aligned} X^{ij} \det\tilde{\pi}=-\frac{f'}{8}\varepsilon^{jkl}\underaccent{\tilde}{\eta}{}_{abc} \tilde{\pi}^a_k \tilde{\pi}^b_l \tilde{B}^{ci}. \label{Xmomenta}\end{aligned}$$ Taking the trace of (\[Xmomenta\]) and substituting it into (\[scalarTr\]) gives the Hamiltonian constraint $$\begin{aligned} \tilde{\tilde{\mathcal{H}}}&\equiv& \frac{f'}{8} \varepsilon^{ijk}\underaccent{\tilde}{\eta}{}_{abc} \tilde{\pi}^a_i \tilde{\pi}^b_j \tilde{B}^c_k + \frac{\Lambda}{6} \varepsilon^{ijk}\underaccent{\tilde}{\eta}{}_{abc} \tilde{\pi}^a_i \tilde{\pi}^b_j \tilde{\pi}^c_k \approx 0. \label{hamilconstraint}\end{aligned}$$ It is interesting to note the presence of function $f'$ in this constraint. The Hamiltonian constraint is then a modification of the familiar Ashtekar version of the Hamiltonian constraint. We recall that $f'$ is a dimensionless holomorphic function of ${\rm Tr}\Psi$ and that $f'_0\neq 0$ since $f$ enjoys the property (ii). For instance, the functions (\[fquadratic\]) and (\[fexp\]) lead to $f' =2\alpha_1+2\alpha_2({\rm Tr}\Psi+\Lambda) $ and $f' = - 2\beta \exp(2 {\rm Tr}\Psi/\Lambda) + 2 \alpha ({\rm Tr}\Psi+\Lambda)$, respectively. In particular, if $\alpha_2=0$ in (\[fquadratic\]), $f' =2\alpha_1=$ const and hence (\[hamilconstraint\]) reduces to the usual Hamiltonian constraint for GR, as was first found in Ref. [@Dadh]. It is worth mentioning that (\[hamilconstraint\]) is reminiscent of Krasnov’s modified Hamiltonian constraint [@Kras5], but is not the same. In Ref. [@Kras5] the cosmological constant that appears in the usual Hamiltonian constraint is replaced by an arbitrary function that has a dependence different from $f'$. Next, it remains only to consider the Poisson brackets among the constraints. Since the Gauss and vector constraints remain unchanged, the Poisson brackets that must be computed are those that involve the Hamiltonian constraint. Taking into account that $\tilde{\tilde{\mathcal{H}}}$ is gauge invariant, the Poisson bracket of $\tilde{\mathcal{G}}_i$ and $\tilde{\tilde{\mathcal{H}}}$ does not change. Similarly, the Poisson bracket between $ \tilde{\mathcal{V}}_a$ and $\tilde{\tilde{\mathcal{H}}}$ gives the known result. Therefore, we are bound to calculate only the Poisson bracket of the Hamiltonian constraint with itself. Let us introduce the smeared Hamiltonian constraint $$\begin{aligned} C_{\underaccent{\tilde}{N}}:=\int d^3x \underaccent{\tilde}{N} \tilde{\tilde{\mathcal{H}}},\label{smearedHamiltonian}\end{aligned}$$ where the test field $\underaccent{\tilde}{N}$ has weight $-1$. Then, the Poisson bracket of interest is $$\begin{aligned} \left\{ C_{\underaccent{\tilde}{N}{}_1},C_{\underaccent{\tilde}{N}{}_2} \right\}\!=\!\int\! d^3x \left( \frac{\delta C_{\underaccent{\tilde}{N}{}_1}}{\delta A^i{}_a } \frac{\delta C_{\underaccent{\tilde}{N}{}_2}}{\delta \tilde{\pi}^a_i }\! -\! \frac{\delta C_{\underaccent{\tilde}{N}{}_2}}{\delta A^i{}_a } \frac{\delta C_{\underaccent{\tilde}{N}{}_1}}{\delta \tilde{\pi}^a_i }\right). \label{poissonHamiltonian}\end{aligned}$$ For the purpose of computing this Poisson bracket, it is convenient to obtain the variation of $C_{\underaccent{\tilde}{N}}$ by using (\[Xmomenta\]) and $\delta f'=f'' \delta{\rm Tr}\Psi$. Then, $$\begin{aligned} \delta C_{\underaccent{\tilde}{N}}=\int d^3x \underaccent{\tilde}{N} \left[ \frac{1}{4} \gamma_1 \delta \left( \varepsilon^{ijk}\underaccent{\tilde}{\eta}{}_{abc} \tilde{\pi}^a_i \tilde{\pi}^b_j \tilde{B}^c_k \right) \right. \nonumber\\ + \left.\frac{1}{6} \gamma_2 \delta\left( \varepsilon^{ijk}\underaccent{\tilde}{\eta}{}_{abc} \tilde{\pi}^a_i \tilde{\pi}^b_j \tilde{\pi}^c_k \right) \right], \label{variationHamiltonian}\end{aligned}$$ where the functions $\gamma_1=\gamma_1({\rm Tr}\Psi)$ and $\gamma_2=\gamma_2({\rm Tr}\Psi)$ are given by $$\begin{aligned} \gamma_1&:=&\frac{1}{4} \frac{\left(f'\right)^2}{\left(f- \frac{1}{2} \left( {\rm Tr}\Psi + \Lambda \right) f'\right)'},\\ \gamma_2&:=&\Lambda+\frac{1}{2}\frac{\left( {\rm Tr}\Psi + \Lambda \right)^2f''}{\left(f- \frac{1}{2} \left( {\rm Tr}\Psi + \Lambda \right) f'\right)'}.\end{aligned}$$ Now having (\[variationHamiltonian\]), we get the required variations $$\begin{aligned} \frac{\delta C_{\underaccent{\tilde}{N}}}{\delta A^i{}_a }&=&D_b \left(\underaccent{\tilde}{N} \gamma_1 \varepsilon_i{}^{jk} \tilde{\pi}^a_j \tilde{\pi}^b_k \right), \label{variationA}\\ \frac{\delta C_{\underaccent{\tilde}{N}}}{\delta \tilde{\pi}^a_i } &=& \frac{1}{2} \underaccent{\tilde}{N} \varepsilon^{ijk}\underaccent{\tilde}{\eta}{}_{abc} \tilde{\pi}^b_j (\gamma_1 \tilde{B}^{c}_k +\gamma_2 \tilde{\pi}^{c}_k). \label{variationPi} \end{aligned}$$ By substituting (\[variationA\]) and (\[variationPi\]) into (\[poissonHamiltonian\]) the Poisson bracket can finally be written as $$\begin{aligned} \left\{ C_{\underaccent{\tilde}{N}{}_1},C_{\underaccent{\tilde}{N}{}_2} \right\}=\int d^3x \gamma_1^2 \underaccent{\tilde}{\underaccent{\tilde}{N}}{}_b \tilde{\tilde{Q}}^{ab} F^i{}_{ac} \tilde{\pi}^c_i, \label{poissonHamiltonian2} \end{aligned}$$ where $$\begin{aligned} \underaccent{\tilde}{\underaccent{\tilde}{N}}{}_b&:=&\partial_b(\underaccent{\tilde}{N}{}_1) \underaccent{\tilde}{N}{}_2 - \partial_b(\underaccent{\tilde}{N}{}_2) \underaccent{\tilde}{N}{}_1,\\ \tilde{\tilde{Q}}^{ab}&:=&\tilde{\pi}^a_i \tilde{\pi}^b_j \delta^{ij}.\end{aligned}$$ Therefore (\[poissonHamiltonian2\]) differs from the usual Poisson bracket only by the nontrivial function $\gamma_1$. For the functions (\[fquadratic\]) and (\[fexp\]), we obtain, respectively, $$\begin{aligned} \gamma_1=\frac{(\alpha_1+\alpha_2({\rm Tr}\Psi + \Lambda))^2}{\alpha_1}, \label{gamma1cuadratic}\end{aligned}$$ and $$\begin{aligned} \gamma_1=\frac{\Lambda \left( \beta \exp(2 {\rm Tr}\Psi/\Lambda) - \alpha ({\rm Tr}\Psi+\Lambda) \right)^2}{\beta (2{\rm Tr}\Psi+\Lambda) \exp(2 {\rm Tr}\Psi/\Lambda) }.\end{aligned}$$ Note that if $\alpha_2=0$ in (\[gamma1cuadratic\]), i.e., in the case of action principle (\[slinear\]), we get $\gamma_1=\alpha_1=$ const and hence (\[poissonHamiltonian2\]) is simply the usual Poisson bracket. Notice that $\tilde{\mathcal{G}}_i$, $\tilde{\mathcal{V}}_a$, and $\tilde{\tilde{\mathcal{H}}}$ form a set of $3+3+1=7$ first class constraints. Also we have $3\times3=9$ configuration variables $A^i{}_a$. Therefore, we are left with two complex degrees of freedom per space point, as it should be for complex GR. CONCLUSIONS =========== We conclude with some remarks. (a) In this paper we have introduced a new class of formulations for GR with either a vanishing or a nonvanishing cosmological constant that depends on a $SO(3,\mathbb{C})$ gauge connection and a complex-valued 4-form, via a holomorphic function of the trace of a symmetric $3\times3$ matrix $\Psi(A^i,\rho)$ that was constructed from these variables. (b) As a consequence of our class, we have achieved a very simple action principle for general relativity, given by (\[Action\]) with (\[fquadratic\]), that involves a quadratic function on the trace of $\Psi$. One advantage of this action principle is that it works well with and without a cosmological constant, a result that contrasts with the action principles of Refs. [@CJD1] and [@Kras1], which do not either support or necessarily require a nonvanishing cosmological constant. An interesting novelty of this action is that it involves an arbitrary parameter that resembles the Barbero-Immirzi parameter. Furthermore, in the particular case when this quadratic function is reduced to a linear one, the resulting action principle can be related to the one found in Ref. [@erratum]. (c) We have also developed a method for constructing action principles for general relativity belonging to the class of formulations by integrating holomorphic functions with certain desirable properties. (d) In particular, and as a straightforward implementation of the method, we have also reported a new action principle for general relativity with a nonzero cosmological constant, given by (\[Action\]) with (\[fexp\]), which involves an exponential function of the trace of $\Psi$. The action principles mentioned in (b) and (d) are new and had not been reported before. (e) We have obtained the pure connection action principles from the members of the class with a nonzero cosmological constant, by eliminating the field $\rho$, and also compared the resulting actions with that of Ref. [@Kras1]. (f) We have carried out the canonical analysis of the class of formulations. It was found that the only constraint that gets modified with respect to those of the Ashtekar formulation is the Hamiltonian constraint. The modification consists in promoting a constant factor to the nontrivial function $f'$. Therefore, each member of the class of formulations leads to a different Hamiltonian constraint, whereas the Gauss and vector constraints remain unchanged. The analysis shows that the members of the class have 2 complex degrees of freedom per space point, as it should be for complex formulations of GR. Furthermore, a natural way to generalize the class of formulations presented in this paper is in the spirit of Refs. [@bengtsson] and [@Kras3], i.e., by promoting the holomorphic function from a dependence on the trace of $\Psi$ to a dependence on the three independent scalar invariants of the matrix $\Psi$. Work in this direction is in progress. Another relevant aspect of future work is to linearize the class of formulations (\[Action\]), which can be done in the sense of Ref. [@Kras4]. ACKNOWLEDGMENTS {#acknowledgments .unnumbered} =============== We deeply thank Mariano Celada for useful discussions and help. We wish to thank Alejandro Perez and Riccardo Capovilla for useful comments. We also thank an anonymous referee for encouraging us to improve our manuscript. This work was supported in part by CONACyT, México, Grant No. 167477-F. [99]{} A. Ashtekar and J. Lewandowski, [[Classical Quantum Gravity ]{}[21]{}, R53 (2004)](http://iopscience.iop.org/0264-9381/21/15/R01); C. Rovelli, [Quantum Gravity]{} (Cambridge University Press, Cambridge, UK, 2004); T. Thiemann, [Modern Canonical Quantum General Relativity]{} (Cambridge University Press, Cambridge, UK, 2007). J. F. Plebański, [[J. Math. Phys. ]{}(N.Y.) [18]{}, 2511 (1977)](http://dx.doi.org/10.1063/1.523215). R. Capovilla, J. Dell, and T. Jacobson, [[Phys. Rev. Lett. ]{}[63]{}, 2325 (1989)](http://link.aps.org/doi/10.1103/PhysRevLett.63.2325). R. Capovilla, J. Dell, and T. Jacobson, [[Classical Quantum Gravity ]{}[8]{}, 59 (1991)](http://dx.doi.org/10.1088/0264-9381/8/1/010). R. Capovilla, J. Dell, and T. Jacobson, [[Classical Quantum Gravity ]{}[9]{}, 1839 (1992)](http://dx.doi.org/10.1088/0264-9381/9/7/015). R. Capovilla and T. Jacobson, [[Mod. Phys. Lett. A ]{}[07]{}, 1871 (1992)](http://dx.doi.org/10.1142/S0217732392001579). P. Peldan, [[Classical Quantum Gravity ]{}[8]{}, 1765 (1991)](http://iopscience.iop.org/0264-9381/8/10/005). K. Krasnov, [[Phys. Rev. Lett. ]{}[106]{}, 251103 (2011)](http://link.aps.org/doi/10.1103/PhysRevLett.106.251103). R. A. Horn and C. R. Johnson, [Topics in Matrix Analysis]{} (Cambridge University Press, Cambridge, UK, 1991), pp. 419 and 485. F. R. Gantmacher, [The Theory of Matrices]{} (Chelsea Publishing Company, New York, 1960), Vol. I, pp. $234-239$. R. Capovilla, T. Jacobson, and J. Dell, [[Classical Quantum Gravity ]{}[8]{}, 41 (1991)](http://dx.doi.org/10.1088/0264-9381/8/1/009). D. Gonzalez, M. Montesinos, and M. Velazquez, [arXiv:1204.4225](http://arxiv.org/abs/1204.4225). H. Urbantke, [[J. Math. Phys. ]{}(N.Y.) [25]{}, 2321 (1984)](http://dx.doi.org/10.1063/1.526402). K. Krasnov, [[Gen. Relativ. Gravit. ]{}[43]{}, 1 (2011)](http://dx.doi.org/10.1007/s10714-010-1061-x). S. Holst, [[Phys. Rev. D ]{}[53]{}, 5966 (1996)](http://link.aps.org/doi/10.1103/PhysRevD.53.5966). J. B. Conway, [Functions of One Complex Variable]{} (Springer-Verlag, New York, 1973). N. Dadhich, S. Kosht, and A. Kshirsagar, [[Classical Quantum Gravity ]{}[8]{}, L61 (1991)](http://stacks.iop.org/0264-9381/8/i=3/a=002). K. Krasnov, [[Phys. Rev. Lett. ]{}[100]{}, 081102 (2008)](http://link.aps.org/doi/10.1103/PhysRevLett.100.081102). I. Bengtsson, [[Phys. Lett. B ]{}[254]{}, 55 (1991)](http://dx.doi.org/10.1016/0370-2693(91)90395-7). K. Krasnov, [[Phys. Rev. D ]{}[81]{}, 084026 (2010)](http://dx.doi.org/10.1103/PhysRevD.81.084026). G. Delfino, K. Krasnov, and C. Scarinci, [arXiv:1205.7045](http://arxiv.org/abs/1205.7045v2).	{ "pile_set_name": "ArXiv" }
--- abstract: \| Results from spectroscopic observations of the Intermediate Polar (IP) EX Hya in quiescence during 1991 and 2001 are presented. Spin-modulated radial velocities consistent with an outer disc origin were detected for the first time in an IP. The spin pulsation was modulated with velocities near $\sim500-600$ [km s$^{-1}$]{}. These velocities are consistent with those of material circulating at the outer edge of the accretion disc, suggesting corotation of the accretion curtain with material near the Roche lobe radius. Furthermore, spin Doppler tomograms have revealed evidence of the accretion curtain emission extending from velocities of $\sim500$ [km s$^{-1}$]{} to $\sim1000$ [km s$^{-1}$]{}. These findings have confirmed the theoretical model predictions of King & Wynn (1999), Belle et al. (2002) and Norton et al. (2004) for EX Hya, which predict large accretion curtains that extend to a distance close to the Roche lobe radius in this system. Evidence for overflow stream of material falling onto the magnetosphere was observed, confirming the result of Belle et al. (2005) that disc overflow in EX Hya is present during quiescence as well as outburst. It appears that the [H$\beta$]{} and [H$\gamma$]{} spin [radial velocities]{} originated from the rotation of the funnel at the outer disc edge, while those of [H$\alpha$]{} were produced due to the flow of material along the field lines far from the [white dwarf]{} (narrow component) and close to the [white dwarf]{} (broad-base component), in agreement with the accretion curtain model. author: - \| N. Mhlahlo$^{1,2}$[^1], D.A.H. Buckley$^{2}$, V.S. Dhillon$^{3}$, S.B. Potter$^{2}$, B. Warner$^{1}$ and P.A. Woudt$^{1}$\ $^{1}$Astronomy Department, University of Cape Town, Rondebosch 7700, Cape Town, South Africa\ $^{2}$South African Astronomical Observatory, Observatory 7935, Cape Town, South Africa\ $^{3}$Physics and Astronomy Department, University of Sheffield, Sheffield, S3 7RH, UK title: Spectroscopic Observations of the Intermediate Polar EX Hydrae in Quiescence --- \[firstpage\] accretion discs, binary - stars: cataclysmic variables. Introduction ============ EX Hya is an Intermediate Polar (IP), a sub-class of magnetic Cataclysmic Variable Stars (mCVs) where a late-type main sequence star transfers material to the magnetic white dwarf star as the two stars orbit each other under the influence of their mutual gravitation. Unlike in Polars, another subclass of mCVs, where the white dwarf is in synchronous rotation with the binary rotation ($P_{spin}=P_{orb}$), the [white dwarf]{} in an IP is in asynchronous rotation with the orbital motion of the system. EX Hya, however, is nearer synchronism than the majority of IPs as it has a spin period ($\sim$67.03 min) which is about $2/3$ its orbital period (98.26 min) [@mum67; @hel87], and is one of only six out of thirty nine confirmed IPs with its orbital period below the 2-3 h CV period gap [@nor04]. It has an inclination $i=78^{\circ} \pm 1^{\circ}$. Recent studies have shown that EX Hya does not conform to the traditional IP model [@kin99; @wyn00; @bel02; @nor04; @bel05]. This system has a large $P_{spin}/P_{orb}$ ratio ($\sim0.68$) implying that it cannot be in the usual spin equilibrium rotation since most IPs have been shown to attain spin equilibrium near $P_{spin}/P_{orb} \sim0.1$ [@kin99; @wyn00]. This further implies that the corotation radius is far greater than the circularisation radius ($R_{co} \gg R_{cir}$) and that EX Hya cannot possess a Keplerian disc. Systems with Keplerian discs are expected to have $R_{co}<R_{cir}$ and thus a smaller $P_{spin}/P_{orb}$. These factors have prompted theorists to suggest that the spin equilibrium state in EX Hya is determined by $R_{co} \sim b$, where $b$ is the distance to the inner Lagrangian point, $L_{1}$ [@kin99; @wyn00; @nor04]. In this model the accretion curtains extend to near the $L_{1}$ point, and EX Hya resembles an asynchronous Polar where most of the material accretes via the stream [@kin99; @wyn00] and via both the ring of material near the Roche lobe radius of the [primary]{} and the stream [@nor04], depending on the orbital and spin periods of the system, and the magnetic field strength. In the later publication it was shown that material in EX Hya is fed from a ring of material at the outer edge of the Roche lobe, and that for the $P_{spin}/P_{orb}$ of EX Hya, this mode of accretion is preferred over stream-fed accretion. \[sec:obsred\] In this work we present spectroscopic data of EX Hya in quiescence obtained from the SAAO in 1991 (just before EX Hya went into outburst, and a day or two after outburst) and in 2001. Outburst data of 1991 will be discussed in a later publication. Observations And Data Reduction =================================== 1991 Observations ----------------- ---------- ----------------- -------- ------- 24-04-91 2448371.3884097 $3.28$ $90$ 25-04-91 2448372.3603850 $2.93$ $72$ 29-04-91 2448376.2402973 $7.00$ $100$ 24-03-01 2451993.5427099 $2.79$ $42$ 25-03-01 2451994.3581177 $3.77$ $56$ 25-03-01 2451994.5147196 $3.96$ $66$ 26-03-01 2451995.4360089 $1.94$ $48$ 26-03-01 2451995.5412087 $2.92$ $50$ ---------- ----------------- -------- ------- : Table of spectroscopic observations during quiescence in 1991 and 2001. The column Date denotes the date at the beginning of the observing night (before midnight), the column Time denotes the number of observing hours and Spectra the number of spectra obtained.[]{data-label="tabo:observ"} EX Hya was observed in April of 1991 by Buckley et al. (1991) using the SAAO 1.9-m telescope with the Reticon photon counting system (RPCS) detector on the Cassegrain spectrograph. A grating with a resolution of 1200 mm$^{-1}$ was used and a wavelength range of 4000 - 5080 Å was covered at a spectral resolution of $\Delta \lambda \sim$1.2 Å and at a time resolution of 100 - 120 s. The spectrograph slit width was 250 $\mu$m ($\sim$1.5 arcsecs). Wavelength calibration exposures were taken using a CuAr arc lamp. Three nights of observations (24, 25 and 29 April 1991) were covered in quiescence and, in total, 262 spectra were obtained. The observing log is given in Table \[tabo:observ\] together with the starting times of the observations. Following wavelength-calibration and sky-subtraction, the data were flux calibrated using the spectra of the standard star LTT3864. ![Radial-velocity Fourier amplitude spectra from the 1991 combined data are shown for $\alpha=3000$ km s$^{-1}$ and 900 km s$^{-1}$ for the [H$\beta$]{} line (top left panel and second left panel from the top) and for 3000 km s$^{-1}$ and 1200 km s$^{-1}$ for the [H$\gamma$]{} line (top right panel and second right panel from the top). $\Omega$ denotes the orbital frequency of the system, 2$\Omega$ its first harmonic and $\omega+\Omega$ the upper orbital side band where $\omega$ is the spin frequency. The data were prewhitened by the orbital frequency and are displayed in the third panels from the top. Window spectra are plotted below the amplitude spectra (bottom panels). []{data-label="q:amplspec1"}](hbgpspx4win_24591rvel1.ps){width="80mm"} 2001 Observations ----------------- The 2001 observations were obtained using the SITe CCD detector ($266\times1798$ pixels) on the Cassegrain spectrograph of the SAAO 1.9-m telescope. A grating with a resolution of 1200 mm$^{-1}$ was used over the wavelength range 4200 - 5100 Å on the nights of the 25th and 26th April. Another grating with a resolution of 1200 mm$^{-1}$ was used on the 24th, 25th and 26th April over the range 6300 - 7050 Å. The spectral resolution was $\sim$1.0 Å and a 1$\times2$ binning scheme was employed (i.e. binning by $2\times$ in the spatial direction). The exposure times during observations were 60 s. The observations covered the period 24 April - 26 April 2001 and, in total, 262 spectra were obtained (Table \[tabo:observ\]). The extraction and reduction of the data were performed the standard way using the Image Reduction and Analysis Facility (IRAF)[^2] package and the spectra were flux calibrated using observations of the standard star LTT3218. The Radial Velocities {#sec:radvels} ===================== It is widely accepted that in a canonical CV, the high velocity emission line wings are formed in the inner parts of the accretion disc orbiting close to the [white dwarf]{} and thus should reflect its orbital motion [@shaf83; @shaf84; @shaf85]. In IPs, however, high velocity emission line wings are formed in the gas streaming towards the [white dwarf]{} at high velocities [@hel87; @fer93]. ![Radial velocity amplitude spectra shown for the [H$\alpha$]{} line from 2001, for $\alpha=3500$ km s$^{-1}$ and 1200 km s$^{-1}$. The vertical dashed line shows the expected position of the orbital and spin period peaks. The data were prewhitened by $\Omega$ and are shown in the third panel from the top. A window spectrum is shown at the bottom.[]{data-label="q:amplspec3"}](hapspx4win_01rvelspin1.ps){width="60mm"} The radial velocities were determined by measuring the wings of the [H$\beta$]{} and [H$\gamma$]{} emission lines from the 1991 data (24 and 25/04/91 - the data obtained on the 29th was not added since EX Hya had not fully recovered from outburst); and the wings of [H$\alpha$]{}, [H$\beta$]{} and [H$\gamma$]{} emission lines from the 2001 data using the Gaussian Convolution Scheme (GCS, Schneider & Young 1980; Shafter & Szkody 1984; Shafter 1985). The GCS method convolves each spectrum with two identical Gaussian, one in the red wing and one in the blue wing. The separation between the two Gaussians is 2$\alpha$. Care was taken not to include regions far out in the wings where the continuum begins to dominate by choosing reasonable values of the width ($\sigma$) of the Gaussians and $\alpha$. Twelve standard Gaussian band-passes were used with $\alpha$ values ranging from 3500 to 100 km s$^{-1}$ and corresponding width values from 1200 to 100 km s$^{-1}$. Period Searches {#sec:qampspec} --------------- The [radial velocities]{} were Fourier transformed using the Discrete Fourier Transform (DFT) algorithm to search for any periods in the data [@dee75; @kur85]. The [H$\beta$]{} and [H$\gamma$]{} amplitude spectra from 1991 (24 and 25 April) are shown in Figure \[q:amplspec1\] and the [H$\alpha$]{} amplitude spectra from 2001 are shown in Figure \[q:amplspec3\]. ![image](trailer_orb_01avinvar2pixv.ps){width="175mm"} A prominent peak at a frequency corresponding to the 98-minute orbital frequency, $\Omega$, is observed in all the emission lines. Second in strength to the orbital frequency is the spin frequency, $\omega$, of the narrow s-wave component (NSC) ($\alpha=900, 1200$ [km s$^{-1}$]{}) (Figures \[q:amplspec1\] and \[q:amplspec3\]). The third panels from the top in Figures \[q:amplspec1\] and \[q:amplspec3\] show the data after prewhitening by $\Omega$. Power at $\omega$ is clearly present. The spin frequency was not detected at high values of $\alpha$ ($\alpha=3000,3500$ km s$^{-1}$) where it is most expected (since at these velocities the material is quite close to the [white dwarf]{} and its emission is expected to be modulated at the [white dwarf]{} spin period). It was also not present in [H$\beta$]{} and [H$\gamma$]{} [radial velocities]{} of 2001. The amplitude of $\omega$ relative to $\Omega$ was found to be $\sim$81% for [H$\beta$]{}, $\sim$64% for [H$\gamma$]{} and $\sim$18% for [H$\alpha$]{}. Orbital Variations of the Emission Lines {#sec:qew} ======================================== The data were phase-binned on the orbital ephemeris of Hellier & Sprouts (1992), $$T_{eclipse}=2437699.94179 + 0.068233846(4)E \label{eq:ephem},$$ where $E$ is the number of orbital cycles and T is the time of mid-eclipse. This ephemeris is defined by the zero point of mid-eclipse, where minimum intensity is at phase 0.0. This means that for the [radial velocities]{}, [maximum blueshift]{} is perpendicular to the line of centres, at phase 0.75, meaning that spectroscopic phase zero occurs at the blue-to-red crossing of the [emission line]{} [radial velocity]{} curve. 40 phase bins were used to produce the [radial velocities]{}. Orbital Tomograms and Trailed Spectra {#q:odopmap} ------------------------------------- ![2001 [H$\alpha$]{}, [H$\beta$]{}, [H$\gamma$]{} and [HeI $\lambda$4471]{} trailed spectra (top row of panels) and MEM orbital Doppler maps (second row of panels from the top) as well as the average-subtracted trailed spectra (third row of panels) are shown plotted on the same scale except for [H$\alpha$]{} panels. The HVC and the NSC are indicated. The fourth row shows the average-subtracted Doppler maps and the models plotted for $q=0.21$, $i=78^\circ$ and $M_{1}=0.50 M_{\odot}$. The bottom panels are the reconstruction of the average-subtracted data. The fourth and bottom panels are also plotted on the same scale except for [H$\alpha$]{} panels. The lookup table of this figure is such that the brightest emission features appear with decreasing intensity from black to light grey.[]{data-label="q:habegaidltrl"}](habgidltrl01-5xv.ps){width="85mm"} The [H$\alpha$]{}, [H$\beta$]{} and [H$\gamma$]{} Doppler tomograms were computed using the Back Projection Method (BPM) with the application of a filter, and the Maximum Entropy Method (MEM) [@mah88; @mar88; @hor91; @spr98]. The BPM Doppler maps are shown in Figure \[q:habegatrl\] and those constructed using MEM are shown in Figure \[q:habegaidltrl\]. Error in ephemerides (both orbital and spin) is sufficiently small to phase all of our data accurately on the orbital and spin cycles. A velocity amplitude of the primary, $K_{1}=74\pm2$ km s$^{-1}$ from our [radial velocity]{} measurements and that of the [secondary]{}, $K_{2}=360\pm35$ km s$^{-1}$, taken from Vande Putte et al. (2003) and Beuermann et al. (2003), were used to fix the positions of the Roche lobe and the stream trajectories on the tomograms. A [secondary]{} mass, $M_{2}= 0.10 \pm0.01$ M$_\odot$ for EX Hya was derived from the up-to-date secondary mass-period relation of Smith & Dhillon (1998). The mass of the primary, $M_{1}$, was then determined from the above values using $\frac{K_{1}}{K_{2}}=\frac{M_{2}}{M_{1}}$, and was found to be 0.50$\pm$0.05 M$_\odot$. The [H$\beta$]{}, [H$\gamma$]{} and [HeI $\lambda$4471]{} Doppler tomograms (Figure \[q:habegatrl\] and \[q:habegaidltrl\]) show strong emission at the bright spot, some at the Roche lobe and the stream, and some from the disc. Those of [H$\alpha$]{} also show strong bright spot emission but less or no emission from the stream. Disc emission is diminished in [H$\alpha$]{} when compared to other [emission lines]{}, especially at higher velocities. This is more obvious in the BPM tomogram. The bright spot emission falls near the region (-100, 350) km s$^{-1}$. Average-subtracted trailed spectra (Figure \[q:habegaidltrl\]) show the corresponding NSC. $\sim1\times10^{-11} {\rm ergs\hspace{0.1cm} cm^{-2} \hspace{0.1cm} s^{-1}}$ ($\sim$60-70%) of the original line fluxes is contained in the average-subtracted profiles of [H$\beta$]{} and [H$\gamma$]{}, and $\sim1\times10^{-12} {\rm ergs\hspace{0.1cm} cm^{-2}\hspace{0.1cm} s^{-1}}$ ($\sim$80-90%) is contained in [H$\alpha$]{} and [HeI $\lambda$4471]{}. This flux is mainly due to the bright spot and the stream. The trailed spectra have been repeated over 2 cycles for clarity. The trailed spectra of 2001 have revealed two interesting features. The first one is the asymmetry in the intensity of the s-wave (Figure \[q:habegaidltrl\]). In [H$\alpha$]{}, the red wing of the NSC is brighter at $\phi_{98} \sim 0.1-0.3$ and seems to reach maximum brightness near $\phi_{98} \sim 0.25$, whereas the blue wing is dimmer in the range $\phi_{98} \sim 0.7-0.9$ and seems to reach minimum brightness near $\phi_{98} \sim 0.75$. A similar effect is seen in the [H$\beta$]{} and the [H$\gamma$]{} lines, and to a lesser extent in [HeI $\lambda$4471]{}. The second feature is redshifted emission extending from the NSC to high velocities ($\sim1000$ [km s$^{-1}$]{}) at early binary phases ($\phi_{98} \sim 0.0-0.2$). The reconstructed trailed spectra suggest that this latter feature is another s-wave, which we shall refer to as the high velocity component (HVC), crossing the NSC near $\phi_{98} \sim0.2-0.3$. The Doppler tomograms show emission extending from the bright spot position, passing along the stream path, to the bottom-left quadrant at high velocities near 1000 [km s$^{-1}$]{} which is responsible for the HVC; most of this emission does not fall within the disc and gas stream velocities on the map, suggesting that there was small or no overlap of the stream component with the disc. The 1991 tomograms showed similar results. It is worth noting that most of the disc emission in 1991 came from the outer disc than in 2001. Even though the MEM and BPM pick out the same features, in the BP tomograms some features are more prominent than in the MEM tomograms while the reconstruction obtained using the MEM reproduces the observed data well. The advantage of BPM over MEM is that it is faster and it is easier to get a consistent set of maps of different emission lines (in terms of the apparent noise in the images). For this reason both methods have been used. Spin Variations of the Emission lines {#sec:svemq} ========================================= The Spin Radial Velocity Curve ------------------------------ The [radial velocities]{} were phase-folded using 30 bins on the quadratic spin ephemeris of Hellier & Sproats (1992), where spin maximum was defined as $\phi_{67}=0$. ![The [H$\beta$]{} (top panel), [H$\gamma$]{} (middle panel) and [H$\alpha$]{} (bottom panel) spin radial velocities of the narrow component from the 1991 combined data ([H$\beta$]{} and [H$\gamma$]{}) and 2001 data ([H$\alpha$]{}). The [radial velocities]{} were prewhitened by the orbital frequency and phase-folded on the spin frequency using 30 bins and are shown plotted as a function of the spin phase.[]{data-label="rv:hbg_245sppr"}](habg_24591-1t2wunbpr.eps){width="65mm"} Figure \[rv:hbg\_245sppr\] shows the variation of the [H$\beta$]{}, [H$\gamma$]{} and [H$\alpha$]{} narrow components with $\omega$. Maximum blueshift is seen at $\phi_{67}=0.79$ for [H$\beta$]{} and at $\phi_{67}=0.77$ for [H$\gamma$]{}. Whereas for [H$\alpha$]{}, maximum blueshift is seen at $\phi_{67}=0.90$. It should be noted that [H$\alpha$]{} and [H$\beta$]{} / [H$\gamma$]{} have not been observed simultaneously and so both data sets probably sample the spin phases at different orbital phases. Figure \[rv:2457ha-s\] shows the [H$\alpha$]{} narrow component ($\alpha=1200$ [km s$^{-1}$]{}) and the [broad-base component]{} ($\alpha=3500$ [km s$^{-1}$]{}) overplotted. The two components are in phase. The radial velocity variation with the spin period of the [H$\beta$]{} and [H$\gamma$]{} [broad-base component]{} could not be detected, possibly due to velocity cancellation We discuss this in Section \[sec:oqdic\]. ![The spin radial velocity curves of the [H$\alpha$]{} narrow (crosses) and broad (dots) components from 2001 (30 bins) plotted as a function of the spin phase. The solid line represents a fit to the data.[]{data-label="rv:2457ha-s"}](ha_01spinrvel3t51t2wf.ps){width="70mm"} Spin Tomograms and Trailed Spectra {#sec:spintomo} ---------------------------------- Spin tomograms of EX Hya were constructed by [@hel99] but revealed little information. Also, [@bel05] observed no coherent emission site/s on their tomograms folded on the spin phase. The [H$\beta$]{} and the [H$\gamma$]{} BPM and MEM spin tomograms from 2001, however, have revealed a coherent emission site between V$_{x}\sim$ 500 km s$^{-1}$ and $\sim$ 1000 km s$^{-1}$ which is evidence of emission from the accretion curtains (Figures \[q:spintrl\] and \[q:habgspintrlmem\]). But it is a well known fact that since the spin period is $\sim \frac{2}{3}$ of the orbital period in EX Hya, orbital cycle variations do not smear out when folded on the spin phase but repeat every 3 spin cycles [@hel87]. This is thought to be the origin of most of the structure in the [emission lines]{} at velocities $<1000$ [km s$^{-1}$]{} [@hel99]. To address this problem, phase-invariant subtraction is performed where emission that does not vary with the spin cycle is subtracted from the data. This is achieved by measuring minimum flux at each wavelength and subtracting this value. The results are shown in the second panels from the top of Figure \[q:habgspintrlmem\]. It should be mentioned though, that even subtracting the invariant part of the line profiles does not guarantee that the influence of the orbital period variations has been completely removed. ![The [H$\alpha$]{}, [H$\beta$]{}, [H$\gamma$]{} and HeI $\lambda$4471 trailed spectra from 2001 folded on the spin period are shown at the top panels and the average-subtracted spectra are shown at the second panels. Doppler maps constructed from the phase-invariant subtracted spectra are shown in the bottom panels. The Doppler maps were constructed using the BPM and are shown on the same velocity scale with the trailed spectra. The lookup table is as in Figure \[q:habegatrl\].[]{data-label="q:spintrl"}](trailer_spin_01avinvar4pxv.ps){width="80mm"} A spin-wave (to differentiate it from the s-wave which is normally caused by the bright spot) in the [H$\alpha$]{} trailed spectra (Figure \[q:habgspintrlmem\]) was detected from the data after the phase-invariant subtraction was performed. This is the first detection of modulation over the spin cycle in the optical emission line data of EX Hya. This spin-wave can be seen in the trailed spectra before (but hard to see) and after subtraction of the phase-invariant line profile. The narrow peak component is responsible for this spin-wave which is shown expanded in the second column of panels in Figure \[q:habgspintrlmem\] (the narrow peak component was selected by hand over a velocity range of $\pm500$ [km s$^{-1}$]{}). The spin-wave shows maximum blueshift near phase 1.0 and maximum redshift near phase 0.5, and has an amplitude of $\sim$ 500 km s$^{-1}$. The [H$\alpha$]{} tomogram shows corresponding emission near the “3 o’clock” position (blob of emission right at the edge of the map), around $\sim$ 500 km s$^{-1}$. DFTs show lower amplitude ($\sim40$ [km s$^{-1}$]{} for [H$\alpha$]{} and $\sim130-140$ [km s$^{-1}$]{} for [H$\beta$]{} and [H$\gamma$]{}) probably due to dilution by stationary material. Also, the [H$\alpha$]{} MEM tomogram shows stronger emission that peaks in a broad structure at lower velocities (V$_{x} \sim$ -200 km s$^{-1}$ – $\sim$ +200 km s$^{-1}$ – around the “5-6 o’clock” position). Circular motion gives rise to low or zero radial velocities when the motion is perpendicular to the line of sight, and the emission seen around the “5-6 o’clock” position could not be from such velocities since it shows [maximum blueshift]{} at $\phi_{67}\sim0.2-0.25$. Similar emission was observed in a Polar and was thought to be due to material that has just been decelerated after having attached to the magnetic field lines [@sch05] (we discuss an alternative explanation in Section \[sec:oqdic\]). The emission near the edge of the tomogram (at $\sim$ 500 km s$^{-1}$) shows [maximum blueshift]{} at $\phi_{67}\sim1.0$ and therefore cannot be due to motion perpendicular to the line of sight either. ![[H$\alpha$]{}, [H$\beta$]{} and [H$\gamma$]{} trailed spectra of 2001 folded on the spin period are shown in the top panels and the phase-invariant subtracted trailed spectra are shown in the second panels from the top. MEM spin Doppler tomograms constructed from the phase-invariant subtracted spectra are shown in the third panels with the reconstructed spectra in the bottom panels. The spin wave observed in the [H$\alpha$]{} phase-invariant subtracted trailed spectra, which was caused by the [H$\alpha$]{} narrow component, is shown expanded on a smaller velocity scale. The first column of panels are plotted between -1500 [km s$^{-1}$]{} and 1500 [km s$^{-1}$]{} and the last two column are plotted between -2000 [km s$^{-1}$]{} and +2000 [km s$^{-1}$]{}. The lookup table is as in Figure \[q:habegaidltrl\].[]{data-label="q:habgspintrlmem"}](habgidltrl01sp-4ixv.ps){width="80mm"} The [H$\alpha$]{} trailed spectra also show emission coupled to the spin-wave near $\phi_{67} \sim 0.1-0.6$ that extends to high velocities in the red, a similar situation to that seen the in orbital tomograms due to the HVC (Section \[q:odopmap\]). The corresponding emission in the [H$\alpha$]{} tomograms extending to higher velocities in the red spectral region is not clear. Both the [H$\beta$]{} and [H$\gamma$]{} phase-invariant subtracted trailed spectra show three weak-intensity spin-waves. The most clearly visible of the three is phased with maximum redshift near $\phi_{67} \sim 0.3-0.4$, with an estimated velocity amplitude of $\sim 900$ km s$^{-1}$ and corresponds to the emission near the “3 o’clock” position in the tomogram (Figure \[q:habgspintrlmem\]). The reconstructed trailed spectra reproduce the observed data. The [H$\beta$]{} and [H$\gamma$]{} trailed spectra in Figure \[q:spintrl\] seem to support these results. The emission observed near the “3 o’clock position” in the tomograms has also been seen in other IPs such as AO Psc and FO Aqr [@hel99] and was interpreted as emanating from the upper accretion curtain. The spin-waves are weak in intensity though, and more data are needed to support these results. Discussion of the Orbital and Spin Data {#sec:oqdic} ======================================= The generally accepted model of EX Hya has the material leaving the [secondary]{} star through the L$_{1}$ point, passing via a stream of material which orbits about the [white dwarf]{}, to form an accretion disc. The magnetic field lines of the [white dwarf]{} which form accretion curtains above and below the orbital plane channel the material from the disc, starting from the co-rotation radius ($R_{co}$) where the disc is truncated by the field lines, to the surface of the [white dwarf]{} [@hel87; @ros91]. King & Wynn (1999) challenged this model by arguing that systems with $P_{spin}/P_{orb} > 0.1$ cannot possess Keplerian discs since this implies $R_{co} \gg R_{cir}$. They showed that the spin equilibrium state in EX Hya is determined by $R_{co} \sim b$, where $b$ is the distance to the $L_{1}$ point. In this model the accretion curtains extend to near the $L_{1}$ point, and EX Hya resembles an asynchronous polar where most of the material accretes via the stream [@kin99; @wyn00]. Belle et al. (2002) revised the model of EX Hya after they showed that their EUV data support the model of King & Wynn (1999). Their revised model suggested that the magnetic field in EX Hya forms a large accretion curtain extending to the outer edge of the Roche lobe causing: - [part or all of the non-Keplerian disc (hereafter the ring of material or the ring) to rotate with the [white dwarf]{}]{}, - [an extended bulge (later, Belle et al. (2005) showed that there was Vertically Extended Material (VEM) obscuring the s-wave emission during $\phi_{98}=0.57-0.87$, and evidence for overflowing stream accretion in EX Hya)]{}, and - [the ring of material to feel magnetic force at the regions of the ring close to the poles, causing the ring material at these locations to be controlled by the magnetic field, forming two chunks along the accretion ring that rotate with the [white dwarf]{}.]{} Recently, Norton et al. (2004, 2004a) have shown that for systems with $P_{spin}/P_{orb} \sim0.72$, when the mass ratio is smaller at $q=0.2$, the material forms a ring near the edge of the [primary]{} Roche lobe, from where accretion curtains funnel down to the [white dwarf]{} surface, in agreement with King & Wynn (1999) and Belle et al. (2002). The material is fed from the ring (ring-fed accretion) and channeled along the magnetic field lines (when the angle between the [white dwarf]{} spin axis and magnetic dipole axis is small i.e $< 30$[$^{\circ}$]{}, which is true for EX Hya). The discussion by [@eis02] on the IR-UV flux distribution in EX Hya implies a disc (isobaric and isothermal) with an outer radius of 1.6$\times10^{10}$ cm and a thickness of 2$\times10^{8}$ cm, and an assumed central hole of 6$\times10^{9}$ cm, but [@eis02] suggested that the structure could also be a ring with a larger inner radius, in line with the suggestion of [@kin99; @bel02] and [@nor04]. They found that the disc component contains about 1/6 of the total flux which is a bit more than expected from gravitational energy release at the inner radius, R$_{in} > 6\times10^{9}$ cm. Our spectroscopic data support both the model of Belle et al. (2002) and Norton et al. (2004) in which material from a ring, circling the [white dwarf]{} and co-rotating with the magnetic field lines at the outer edge of the Roche lobe, is accreted by the [white dwarf]{}. The presence of the bright spot revealed by the trailed spectra, the DFTs of the [radial velocities]{} and the Doppler maps (Figures \[q:habegatrl\] and \[q:habegaidltrl\]) suggest the presence of a disc or ring of material extending to near the Roche lobe radius, around the [white dwarf]{}. When comparing the 1991 and 2001 tomograms for the [H$\beta$]{} they appear to be in the same state or similar, given the fact that they are ten years apart. It is reassuring that the fact that the two groups of lines have not been measured simultaneously is not a significant problem in the analysis. More importantly, a spin pulse modulated at velocities consistent with those of the material circulating at the outer edge of the disc ($\sim500-600$ [km s$^{-1}$]{}) (Figures \[q:amplspec1\] and \[q:amplspec3\]) was detected and provides evidence for co-rotation of the extended accretion curtains with the ring material. As discussed in Section \[sec:spintomo\], these low radial velocities mentioned above were not caused by motion perpendicular to the line of sight near the [white dwarf]{}, neither were they caused by velocity cancellation as will be shown later. A spin wave was detected in the spin-folded trailed spectra of [H$\alpha$]{} (Figure \[q:habgspintrlmem\]) with a velocity semi-amplitude of $\sim$ 500-600 km s$^{-1}$. The spin wave shows maximum blue-shift near phase $\phi_{67}\sim1.0$ (when the upper magnetic pole is pointed away from the observer) and maximum redshift near phase $\phi_{67}\sim0.5$. The [H$\alpha$]{} equivalent widths show maximum flux near $\phi_{67}\sim1.0$. This picture is consistent with the accretion curtain model of IPs and is possible if accretion occurs via a disc/ring. The spin tomograms (Figures \[q:spintrl\] and \[q:habgspintrlmem\]) show evidence of the accretion curtain emission extending from $\sim 500$ [km s$^{-1}$]{} to high velocities ($\sim 1000$ [km s$^{-1}$]{}), suggesting that material is channeled along the field lines from the outer ring. The [H$\alpha$]{} narrow and broad base components show similar phase variation, suggesting same position of maximum radial velocity as shown in Figures \[rv:2457ha-s\] and \[o:qexillus2e\] (line OA). This indicates that material is channeled from the ring (at low velocities) to high velocities along the field lines. A mass ratio of $q \sim$ 0.2 was measured from our data, and so the period ratio $P_{spin}/P_{orb}$ $\sim0.68$ is consistent with the ring accretion model of Norton et al. (2004). Decreased prominence of the [narrow s-wave component]{} around $\phi_{98}=0.57-0.87$ (Figure \[q:habegatrl\] and \[q:habegaidltrl\]) was observed and suggests the presence of VEM at the outer edge of the ring of material obscuring the emission at these phases. The presence of the overflow stream may be infered from this observation [@bel05]. But direct evidence comes from orbital Doppler tomograms which show an asymmetry in the emission, where more emission is observed from the [secondary]{} Roche lobe to the lower left quadrant than from the opposite side. Average-subtracted orbital tomograms show this emission at higher velocities ($\sim$900-1000 [km s$^{-1}$]{}) (Figures \[q:habegatrl\] and \[q:habegaidltrl\]), and it corresponds to the HVC observed in the trailed spectra, which is modulated with a velocity semi-amplitude of $\sim 1000$ [km s$^{-1}$]{}. This HVC is reminiscent of that detected by Rosen et al. (1987) in the trailed spectra of the AM Her system V834 Cen. Their HVC was blueshifted with a velocity of $\-900$ [km s$^{-1}$]{}and was said to be produced in the stream close to the [white dwarf]{}. The only difference is that there was no evidence of the HVC emission when it was expected to be seen redward of another component (medium-velocity component) in their data, whereas in EX Hya the evidence of the HVC emission is missing between $\phi_{98} \sim0.3-0.85$. The HVC emission is maximally blueshifted at $\phi_{98} \sim0.3-0.4$. This phasing is consistent with the expected phase of impact of a stream of material from the [secondary]{} with the disc or of the overflow stream material free-falling onto the magnetosphere of the [primary]{} [@hel89]. Support for overflow stream is also provided by spin tomograms where emission is observed on the upper accretion curtain with velocities consistent with stream velocities. This suggests that this emission site may also have resulted due to impact of overflow stream with the magnetosphere. The resulting emission is receding from the observer at [maximum redshift]{} near $\phi_{67} \sim 0.4$ (Figure \[q:habgspintrlmem\]), in agreement with the accretion curtain model. The model of [@kin99] is not fully supported by our observations since it predicts direct accretion via a stream. Our observations, however, fit the models of [@nor04] and [@bel02; @bel05]. ![A depiction of the regions where [H$\alpha$]{} was formed. Both the narrow and broad base components fall along the same radial direction, OA, resulting in similar phase variation.[]{data-label="o:qexillus2e"}](qexillus2e.eps){width="60mm"} There is evidence for strong [H$\alpha$]{} emission of the narrow s-wave component in the spin tomograms, centred around $\sim100$ [km s$^{-1}$]{} (Figure \[q:habgspintrlmem\]), that is not accounted for by these models. This emission shows [maximum blueshift]{} at phase $\phi_{67}\sim0.2$, suggesting that these are rotational velocities (or a combination of streaming and rotational velocities) of the antiphased motion of a source locked to the [white dwarf]{}. One possible explanation is that this emission comes from the opposite pole of the [white dwarf]{}, at a radial distance of 6$\times10^{9}$ cm ($\sim$8R$_{wd}$). [@sie89] found that the eclipsed optical source in EX Hya is centred at a radial distance of 1.5$\times10^{9}$ cm ($\sim$2R$_{wd}$), which is about four times closer to the [white dwarf]{} compared to our result. This could be the same emission region, but in our observations the emission is spread out, possibly due to the quality of the data, and this could account for the difference in the radial distance values quoted above. But we cannot imagine a geometry where such low rotational velocities can dominate over streaming velocities along the field lines near the [white dwarf]{}. We therefore suggest that this is evidence for material that is diverted out of the orbital plane. Since one of the assumptions of Doppler tomography is that everything lies on the plane, it is not possible to locate the exact position of this emission relative to the [white dwarf]{}. ![image](qexillus2d.eps){width="90mm"} ![image](qexillus2aip.eps){width="85mm"} White dwarf and [secondary]{} masses ------------------------------------ [@hel87] showed that maximum line widths of $\pm3500$ [km s$^{-1}$]{} constrain the mass of the [white dwarf]{}, and a free-fall velocity of this magnitude could be achieved for [white dwarf]{} masses greater than 0.48 M$_\odot$. We found M$_{1}=0.50\pm0.05 M_\odot$, in good agreement with the results obtained from recent studies by [@hoo04; @beu03] and [@put03]. For the [secondary]{}, we derived M$_{2}=0.10 \pm0.01$ M$_\odot$ from the secondary mass-period relation of Smith & Dhillon (1988), and this value agrees with that obtained by [@put03]. [@beu03] and [@hoo04] find lower values for M$_{2}$ consistent with 0.09 M$_\odot$. [@eis02] argues that for a [secondary]{} mass as low as 0.1 M$_\odot$ the secondary would have to be substantially expanded by $\sim$10%. The revised model of EX Hya --------------------------- We propose a model where one of the two chunks alluded to by Belle et al. (2002), which are formed by the magnetic pull along the accretion ring, co-rotates with the accretion curtains at the outer edge of the Roche lobe at $\sim$ 500-600 km s$^{-1}$, giving rise to the pulsation of emission at the spin period which we observe in our data, while the other is hidden by the accretion curtain below the ring of material. The resulting emission is maximally blueshifted near $\phi_{67} \sim0.8$ (Figure \[rv:hbg\_245sppr\]). In the accretion curtain model, at $\phi_{67} \sim0.5$ in the spin cycle, minimum flux (due to higher opacity) is observed when the upper accretion pole of the [white dwarf]{} is pointed towards the observer [@hel87], and so the phasing mentioned above is compatible with the motion of a rotating accretion funnel. This is illustrated in Figure \[o:qexillus2b\], where the position of the observer at pulse maximum is indicated, and the axis of the magnetic pole is shown. The disruption of the disc by the magnetic field at the outer disc is illustrated and part of the disc co-rotating with the magnetosphere is shown. At a corotation radius, $R_{c} \sim$ $b=a(0.500-0.227\log \frac{M_{2}}{M_{1}})$ ($\sim3\times10^{10}$ cm), the material is rotating at a velocity of $v^{2}=\frac{GM}{b}$ $\sim500$ [km s$^{-1}$]{}, in good agreement with the observations. Also, a rotation velocity of $\sim600$ [km s$^{-1}$]{} was measured from the spectra and the radial distance from the star to the ring of material was found to be $\sim3\times10^{10}$ cm, which is similar to ${b}$, for a [white dwarf]{} mass of 0.5 M$_{\odot}$ (Keplerian motion about the [white dwarf]{} had to be assumed in these calculations). At this radius, the accretion curtain is also rotating at a velocity of 2$\pi R_{co}$/P$_{spin}$ $\sim500$ [km s$^{-1}$]{}. $\sim6\times10^{-12}$ ergs cm$^{-2}$ s$^{-1}$ (64% - integrated over one spin cycle) of the original line fluxes that is contained in the average-subtracted profile of [H$\alpha$]{} shows [radial velocity]{} variations with the spin period. Assuming that [H$\beta$]{} and [H$\gamma$]{} also show a similar flux variation ([H$\beta$]{} and [H$\gamma$]{} spin tomograms also show a low-velocity s-wave but this result is not secured due to poor quality of data), the total line fluxes showing [radial velocity]{} variations with the spin period can be estimated to be $\sim2\times10^{-11}$ ergs cm$^{-2}$ s$^{-1}$ for the three [emission lines]{}. This is $\sim$2/10 of the total disc flux [@eis02], suggesting that only part of the ring corotates with the [white dwarf]{} while the rest of the material may be involved in a near Keplerian motion (this is a rough comparison since the flux is integrated over one spin cycle for [H$\alpha$]{}, [H$\beta$]{} and [H$\gamma$]{} in our data whereas [@eis02] derived their total flux values from one spectrum over the wavelength range $\lambda=912 - 24000$ Å). While some of the ring material co-rotates with the accretion curtains (i.e. remains in the disc rather than being immediately channeled along the field lines), some is channeled along the field lines at $\sim500$ [km s$^{-1}$]{} towards the [white dwarf]{}. There is also some material that overflows the ring and attaches onto the magnetic field lines. The overflow stream hits the magnetosphere, probably causing a second bright spot on the slowly rotating magnetosphere (Figure \[o:qexillus2b\]). The overflow stream is irradiated by the [white dwarf]{} in its inner regions close to the [white dwarf]{} (the regions facing the [white dwarf]{}). This results in the HVC emission being obscured at $\phi_{98} \sim0.4-0.9$, which are phases where the stream is viewed from behind-opposite the side facing the [white dwarf]{}, hiding the irradiated inner regions. HVC emission from the stream is blueshifted when that from the [narrow s-wave component]{} shows maximum redshift. Near $\phi_{98}\sim0.25$ the two s-waves intersect, explaining the asymmetry in the brightness of the s-wave seen near $\phi_{98}\sim0.25$ (Figure \[q:habegaidltrl\]). The overflow stream curls nearly behind the [white dwarf]{} and it is truncated by the field when the upper magnetic pole is facing the stream. Ferrario & Wickramasinghe (1993) and Ferrario, Wickramasinghe & King (1993) showed that in IPs the accretion curtain below the orbital plane can contribute in the radial velocities of a system if it can be seen either through the central hole of the truncated disc, or from below the disc, or both. This effect will result in velocity cancellation due to nearly equal quantities of material that are blueshifted and redshifted on the accretion curtains (Ferrario, Wickramasinghe & King 1993). In EX Hya where the inclination is high (78[$^{\circ}$]{}) and the disruption radius is large ($\sim$ 40 R$_{WD}$, for a [white dwarf]{} mass of 0.5 M$_{\odot}$) as proposed in Figure \[o:qexillus2b\], it is clear that we see spin-varying emission from two opposite magnetic poles, producing a fairly symmetric structure in the spin-folded line profiles [@hel87; @ros91]. If emission from these opposite poles is cancelling out then the sum will have a much lower velocity. This could explain the near zero and low amplitude of the [radial velocity]{} variation at the spin period of the [H$\beta$]{} and [H$\gamma$]{}, and [H$\alpha$]{} ($\leq$ 40 [km s$^{-1}$]{}) [broad-base component]{}, respectively (see also [@hel87] and Ferrario, Wickramasinghe & King (1993)). One could take this argument further by suggesting that the spin modulation we observe in our data at velocities near $\sim500$ [km s$^{-1}$]{} (Figures \[q:amplspec1\] and \[q:amplspec3\]) is just the slight asymmetries between the two poles. The resulting velocity could just be a measure of the degree to which the poles cancel their velocities near $\pm3500$ [km s$^{-1}$]{}(Coel Hellier; private communication). This, however, cannot be the case for [H$\beta$]{} and [H$\gamma$]{} since these two emission lines show motion that is consistent with that of a rotating object, suggesting that the line profiles are not dominated by the infall velocities at the two opposite accretion poles. If they were produced close to the [white dwarf]{} then maximum rotational velocity near $\pm3500$ [km s$^{-1}$]{} would be 2$\pi R/P_{spin} \sim$ 30 km s$^{-1}$, which is much smaller than $\sim500$ [km s$^{-1}$]{}. However, rotional velocities close to the ring are $\sim500$ [km s$^{-1}$]{}. For [H$\alpha$]{}, however, we observe [maximum blueshift]{} at $\phi_{67} \sim1.0$, and so velocity due to cancellation anywhere between 0 and $\pm3500$ [km s$^{-1}$]{}are expected, depending on how much the two poles cancel. If both accretion curtains are still visible and symmetric at large radii (which is possible as suggested by Ferrario, Wickramasinghe & King (1993) and our model), velocity cancellation will still result in smaller amplitudes than those of $\sim500$ [km s$^{-1}$]{}observed in our data. This would then count against the argument above. Furthermore, [H$\alpha$]{} orbital Doppler tomograms show strong emission at the bright spot. If our model is correct, the field lines should also attract this [H$\alpha$]{} dominated material, which is chanelled along the field lines, as already shown above. The velocity of this material due to streaming motion near the outer ring is less than that of the [H$\alpha$]{} broad-base component close to the [white dwarf]{}, as expected. A strong constraint on our model is that the disruption radius of EX Hya has been shown to be at 5-9$\times10^{9}$ cm [@hel87; @beu03] which implies a [white dwarf]{} magnetic moment of $\mu \sim 7 \times10^{31}$ G cm$^{3}$. For our model this would imply that the accretion curtains do not extend to near the Roche lobe radius. The theoretical analysis of [@kin99] and [@wyn00], however, has shown that equilibrium rotation is possible if the magnetic moment in EX Hya falls within the range of 10$^{33}\le\mu \le 10^{34}$ G cm$^{3}$. These are comparable to weakest field AM Hers below the period gap, and that EX Hya could possess such magnetic moments is supported to a certain extent by the average-subtracted trailed spectra of EX Hya that are reminiscent of [emission lines]{} seen in some Polars, e.g. V834 Cen (as discussed above), EF Eri [@cra81; @cow82], QS Tel [@rom03] and VV Pup [@dia94]. Furthermore, [@cum02] raised the possibility that the magnetic fields in IPs are buried by the material due to high accretion rates and so are not really as low as they appear. The ring structure in EX Hya could imply higher accretion rates in EX Hya than previously thought since the capacity of the ring of material to store matter may be low when compared to that of a classical disc, resulting in the accretion of more material than in a classical disc case. Summary ======= Optical observations of EX Hya and the analysis have suggested that large accretion curtains extending to a distance close to the L$_{1}$ point exist in this system. The DFTs and spin tomograms have for the first time provided evidence for corotation of the field lines with the ring material near the Roche lobe. Also, tomography and the phasing of the spin waves have suggested that feeding by the accretion curtains of the material from the ring (ring-fed accretion) takes place. These findings support the models of Belle et al. (2002) and Norton et al. (2004) for EX Hya and the simulations done by Norton et al. (2004a) which have shown that for systems with the parameters of EX Hya, the accreting material forms a ring at the outer edge of the [primary]{} Roche lobe, from where accretion curtains funnel down to the [white dwarf]{} surface. Evidence for stream overflow accretion has been observed. The HVC caused by the overflow stream disappeared at $\phi_{98}\sim0.4-0.9$ due to obscuration by the stream. Obscuration of the NSC at $\phi_{98}\sim0.57-0.87$ suggested the presence of the VEM which was irradiated by the [white dwarf]{} in its inner regions. The [H$\alpha$]{} [broad-base component]{} shows a radial velocity variation with the spin period whereas that of [H$\beta$]{} and [H$\gamma$]{} could not be detected. The low-amplitude velocity variations modulated at the spin period for [H$\alpha$]{} and for [H$\beta$]{} and [H$\gamma$]{} is explained in terms of velocity cancellation effects. We have provided an explanation for the asymmetry in the intensity of the narrow [s-wave component]{} seen in EX Hya trailed spectra in the optical. The narrow [s-wave component]{} and the HVC cross at $\phi_{98}\sim0.25$, resulting in the asymmetry in brightness that we observe at these phases. The spin-folded trailed spectra are not of good quality and more data are needed to confirm these results. Acknowledgments {#acknowledgments .unnumbered} =============== [NM would like to acknowledge financial support from the Sainsbury/Linsbury Fellowship Trust and the University of Cape Town. We would like to thank Kunegunda Belle, Coel Hellier and Andrew Norton for invaluable discussions and for their constructive comments. We acknowledge use of D. O’Donoghue’s and Tom Marsh’s programs Eagle and Molly, respectively.]{} [Asa87x]{} Belle, K., Howell S. B., Sirk, M., Huber, M.E., 2002, ApJ, 577, 359 Belle, K., Howell, S., Mukai, K., Szkody, P., Nishikida, K., Ciardi, D.R., Fried, R.E., Oliver, J.P., 2005, ApJ, 587, 373 Beuermann, K., Harrison, Th. E, McArthur, B. E., Benedict, G. F., Gansicke, B.T., 2003, A&A, ms3775 Buckley, D., Schwarzenberg-Czerny, A., 1991, Cataclysmic Variable and Related Physics, 2nd Technion Haifa Conference, Annals of the Israel Physical Society, 10 Cowley, A.P, Crampton, D., Hutchings, J.B., 1982, ApJ, 259, 370 Crampton, D., Hutchings, J.B., Cowley, A.P, 1981, ApJ, 243, 567 Cumming, A., 2002, MNRAS, 333, 589 Deeming, T. J., 1975, Ap&SS, 36, 137 Diaz, M. P., Steiner, J. E., 1994, A&A, 283, 508 Eisenbart, S., Beuermann, K., Reinsch, K., Gansicke B.T., 2002, A&A, 382, 984 Ferrario, L., Wickramasinghe, D., 1993, MNRAS, 265, 605 Ferrario, L., Wickramasinghe, D., King, A., 1993, MNRAS, 260, 149 Hellier, C., Mason, K., Rosen, R., 1987, MNRAS, 228, 463 Hellier, C., Mason, K., Smale, A. P., Corbet, R. H. D., O’Donogue, D., Barrett, P. E., Warner, B., 1989b, MNRAS, 238, 1107 Hellier, C., Sproats, L. N. 1992, IBVS, 3724 Hellier, C., 1999, ApJ, 519, 324 Hoogerwerf, R., Brickhouse, N. S., Mauche, C. W., 2004, ApJ, 610, 411 Horne, K., 1991, in: Fundamental Properties of Cataclysmic Variable Stars: 12th North American Workshop on Cataclysmic Variables and Low Mass X-ray Binaries, San Diego State University Publication, San Diego, ed. A.W. Shafter, 23 King, A., Wynn, G. A., 1999, MNRAS, 310, 203 Kurtz, D. W., 1985, MNRAS, 213, 773 Marsh, T. R., 1988, MNRAS, 231, 1117 Marsh, T. R., Horne, K., 1988, MNRAS, 235, 269 Mumford, G., 1967, ApJS, 15, 1 Norton, A., Wynn, A., Somerscales, C., 2004, ApJ, 614, 349 Norton, A., Somerscales, R. V., Parker, T. L., Wynn, A., West, R., 2004a, RevMexAA, 20, 138 Romero-Colmenero, E., Potter, S., Buckley, D., 2003, Astrotomography, 25th meeting of the IAU, Joint Discussion 9, 17 July 2003, Sydney, Australia Rosen, S.R., Mason, K.O., Cordova, F.A., 1987, MNRAS, 224, 987 Rosen, S.R., Mason, K.O., Mukai, K., Williams O.R, 1991, MNRAS, 249, 417 Schneider, D.P., Young, P., 1980, ApJ, 238, 946 Schwarz, R., Schwope, A.D., Staude, A., Remilland, R. A., 2005, A&A, 444, 213 Shafter, A.W. 1983, ApJ., 267, 222 Shafter, A.W., Szkody, P., 1984, ApJ., 276, 305 Shafter, A.W. 1985, Cataclysmic Variables and Low-Mass X-Ray Binaries, D.Q. Lamb and J. Patterson (eds), 355, 358 Siegel, N., Reinsch, K., Beuermann, K., van der Woerd, H., Wolff, E., 1989, A&A, 225, 97 Smith, D. A., Dhillon, V. S., 1998, MNRAS, 301, 767 Spruit, H. C., 1998, preprint(astro-ph/9806141) Vande Putte, D., Smith, R. C., Hawkins, N. A., Martin, J. S., 2003, MNRAS, 342, 151 Wynn, G. A., 2000, New Astr. Rev., 44, 75 [^1]: E-mail: [email protected] [^2]: IRAF is a software package for the reduction and analysis of astronomical data distributed by National Optical Astronomy Observatory (NOAO) which is operated by the Association of Universities for Research in Astronomy (AURA)	{ "pile_set_name": "ArXiv" }
--- abstract: 'X-ray polarimetry promises to give new information about high-energy astrophysical sources, such as binary black hole systems, micro-quasars, active galactic nuclei, and gamma-ray bursts. We designed, built and tested a hard X-ray polarimeter [X-Calibur]{} to be used in the focal plane of the InFOC$\mu$S grazing incidence hard X-ray telescope. X-Calibur combines a low-Z Compton scatterer with a CZT detector assembly to measure the polarization of $10-80 \, \rm{keV}$ X-rays making use of the fact that polarized photons Compton scatter preferentially perpendicular to the electric field orientation. X-Calibur achieves a high detection efficiency of order unity.' author: - 'M. Beilicke$^{}$, W.R. Binns, J. Buckley, R. Cowsik, P. Dowkontt, A. Garson, Q. Guo, M.H. Israel, K. Lee, H. Krawczynski' - 'M.G. Baring' - 'S. Barthelmy, T. Okajima, J. Schnittman, J. Tueller' - 'Y. Haba, H. Kunieda, H. Matsumoto, T. Miyazawa, K. Tamura' title: 'Design and Tests of the Hard X-ray Polarimeter X-Calibur' --- INTRODUCTION {#sec:intro} ============ [Motivation.]{} Spectro-polarimetric X-ray observations are capable of providing important information to study the non-thermal emission processes of various astrophysical sources – namely the fraction and orientation of linearly polarized photons [@Krawcz2011]. So far, only a few missions have successfully measured polarization in the soft (OSO-8 [@Weisskopf1978]) and hard (Integral [@Dean2008]) X-ray energy regime. The Crab nebula is the only source for which the polarization of the X-ray emission has been established with a high level of confidence [@Weisskopf1978; @Dean2008]. Integral observations of the X-ray binary CygnusX-1 indicate a high fraction of polarization in the hard X-ray/gamma-ray bands [@Laurent2011]. [Future missions.]{} There are currently no missions in orbit that are capable of making sensitive X-ray polarimetric observations. This will change by the launch of the satellite-borne [Gravity and Extreme Magnetism SMEX]{} (GEMS) mission [@GEMS] scheduled for 2014 ($2 - 10 \, \rm{keV}$). The [Soft Gamma-Ray Imager]{} on [ASTRO-H]{} [@Tajima2010] will also have capabilities of measuring polarization, but the results may be plagued by systematic uncertainties. The hard X-ray polarimeter [X-Calibur]{} discussed in this paper has the potential to cover the energy range above $10 \, \rm{keV}$, combining a high detection efficiency with a low level of background and well-controlled systematic errors. [Scientific potential.]{} Synchrotron emission results in linearly polarized photons (electric field oriented perpendicular to the magnetic field lines). The polarized synchrotron photons can be inverse-Compton (IC) scattered by relativistic electrons – weakening the fraction of polarization and imprinting a scattering angle dependence [@Poutanen1994] to the observed fraction of polarization. Other important mechanism for polarizing photons are Thomson scattering and curvature radiation. The scientific potentials of spectro-polarimetric hard X-ray observations are (see Krawczynski et al. (2011) and references therein [@Krawcz2011]): [1) Binary black hole (BH) systems.]{} Relativistic aberration and beaming, gravitational lensing, and gravitomagnetic frame-dragging will result in an energy-dependent fraction of X-ray polarization from a Newtonian accretion disk [@Connors1977] since photons with higher energies originate closer to the BH than the lower-energy photons. Schnittman and Krolik [@Schnittman2009; @Schnittman2010] calculate the expected polarization signature including the effects of deflection of photons emitted in the disk by the strong gravitational forces in the regions surrounding the black hole and of the re-scattering of these photons by the accretion disk. Spectro-polarimetric observations can constrain the mass and spin of the BH [@Schnittman2009], as well as the inclination of the inner accretion disk and the shape of the corona [@Schnittman2010]. [2) Pulsars and pulsar wind nebulae.]{} High-energy particles in pulsar magnetospheres are expected to emit synchrotron and/or curvature radiation which are difficult to distinguish from one another. However, since the orbital planes for accelerating charges that govern these two radiation processes are orthogonal to each other, their polarized emission will exhibit different behavior in position angle and polarization fraction as functions of energy and the phase of the pulsar [@Dean2008]. In magnetars, the highly-magnetized cousins of pulsars, polarization-dependent resonant Compton up-scattering is a leading candidate for generating the observed hard X-ray tails. In both these classes of neutron stars, phase-dependent spectro-polarimetry can probe the emission mechanism, and provide insights into the magnetospheric locale of the emission region. [3) Relativistic jets in active galactic nuclei (AGN).]{} Relativistic electrons in jets of AGN emit polarized synchrotron radiation at radio/optical wavelengths. The same electron population is believed to produce hard X-rays by inverse-Compton scattering of a photon field. Simultaneous measurements of the polarization angle and the fraction of polarization in the radio to hard X-ray band could help to study (i) synchrotron self-Compton models (X-ray polarization tracks polarization at radio/optical wavelengths in fraction and orientation angle) [@Poutanen1994] vs. (ii) external-Compton models for which the hard X-rays will have a relatively small ($<$10%) fraction of polarization [@McNamara2009]. Polarization also allows one to test the structure (i.e. helical) of the magnetic field of the jet. [4) Gamma-ray bursts (GRBs)]{}. GRBs are believed to be connected to hyper-nova explosions and the formation of relativistic jets. As in AGN jets, the structure and particle distribution responsible for GRBs can be revealed by X-ray polarization measurements. [5) Lorentz invariance.]{} X-ray polarimetric observations can be used to test theories violating Lorentz invariance [@Fan2007] with unprecedented accuracies by probing the helicity dependence of the speed of light. DESIGN OF X-CALIBUR {#sec:design} =================== Figure \[fig:Design\] illustrates the conceptual design of the X-Calibur polarimeter. A low-Z scintillator is used as Compton-scatterer. For sufficiently energetic photons, the Compton interaction produces a measurable scintillator signal which is read by a PMT. The scattered X-rays are photo-absorbed in surrounding rings of high-Z Cadmium Zinc Telluride (CZT) detectors. This combination of scatterer/absorber leads to a high fraction of unambiguously detected Compton events. Linearly polarized X-rays will preferably Compton-scatter perpendicular to their E field vector – resulting in a modulation of the azimuthal event distribution. ![\[fig:Design\] [Left:]{} Conceptual design of X-Calibur: Incoming X-rays are Compton-scattered in the scintillator rod (read by a PMT) and are subsequently photo-absorbed in one of the CZT detectors surrounding the rod. [Right:]{} X-Calibur design with read-out electronics, shielding and azimuthal rotation bearing.](Schematic_X-Calibur.eps "fig:"){width="55.00000%"} ![\[fig:Design\] [Left:]{} Conceptual design of X-Calibur: Incoming X-rays are Compton-scattered in the scintillator rod (read by a PMT) and are subsequently photo-absorbed in one of the CZT detectors surrounding the rod. [Right:]{} X-Calibur design with read-out electronics, shielding and azimuthal rotation bearing.](Drawing_X-CaliburAndShield.eps "fig:"){width="30.00000%"} Each CZT detector ($2 \times 2 \, \rm{cm}^{2}$) is contacted with a 64-pixel anode grid and a monolithic cathode facing the scintillator rod. Two detector thicknesses ($0.2 \, \rm{cm}$ and $0.5 \, \rm{cm}$) are being tested in the current setup. Each CZT detector is permanently bonded (anode side) to a ceramic chip carrier which is plugged into the readout board. Figure \[fig:Fotos\] (left) shows a single CZT detector unit as well as the readout electronics. Each CZT detector is read out by two digitizer boards (32 channel ASIC developed by G. De Geronimo (BNL) and E. Wulf (NRL) [@Wulf2007] and a 12-bit ADC). 16 digitizer boards (8 CZT detectors) are read out by one harvester board transmitting the data to a PC-104 computer. Four CZT detector units form a ring surrounding the scintillator slab. The scintillator EJ-200 is read by a Hamamatsu R7600U-200 PMT. The PMT trigger information allows to effectively select scintillator/CZT events from the data which represent likely Compton-scattering candidates. The polarimeter and the front-end readout electronics will be located inside an active CsI(Na) anti-coincidence shield with $5 \, \rm{cm}$ thickness and a passive lead shield/collimator at the top (Fig. \[fig:Design\]) to suppress charged and neutral particle backgrounds. ![\[fig:Fotos\] [Left:]{} Top: $2 \times 2 \times 0.2 \, \rm{cm}^{3}$ CZT detector (64 anode pixels). Bottom: $2 \times 2 \times 0.5 \, \rm{cm}^{3}$ detector bonded to a ceramic chip carrier, plugged into 2 ASIC readout boards. The high-voltage cable is glued to the detector cathode (red wire). [Right:]{} Laboratory version of X-Calibur. The scintillator (blueish glow) is surrounded by two detector rings – each consisting of four 64 pixel CZT detectors.](Combined_DetectorAndUnit.eps "fig:"){width="28.00000%"} ![\[fig:Fotos\] [Left:]{} Top: $2 \times 2 \times 0.2 \, \rm{cm}^{3}$ CZT detector (64 anode pixels). Bottom: $2 \times 2 \times 0.5 \, \rm{cm}^{3}$ detector bonded to a ceramic chip carrier, plugged into 2 ASIC readout boards. The high-voltage cable is glued to the detector cathode (red wire). [Right:]{} Laboratory version of X-Calibur. The scintillator (blueish glow) is surrounded by two detector rings – each consisting of four 64 pixel CZT detectors.](p1060223_zoom.eps "fig:"){width="19.00000%"} We plan to use the X-Calibur polarimeter in the focal plane of the InFOC$\mu$S experiment [@InFocus_FirstFlight]. A Wolter grazing incidence mirror focuses the X-rays on the X-Calibur polarimeter. In order to reduce the systematic uncertainties (including biases generated by the active shield, a possible pitch of the polarimeter with respect to the X-ray telescope, etc.), the polarimeter and the active shield will be rotated around the optical axis using a ring bearing (see Fig. \[fig:Design\]). Counter-rotating masses will be used to cancel the net-angular momentum of the polarimeter assembly during the flight. The advantages of the X-Calibur/InFOC$\mu$S design are (i) a high detection efficiency by using more than $80 \%$ of photons impinging on the polarimeter, (ii) low background due to the usage of a focusing optics instead of large detector volumes, and (iii) minimization and control of systematic effects. SIMULATIONS {#sec:simulations} =========== Simulations of the X-Calibur polarimeter were performed using the [Geant4*]{} package [@GEANT] with the Livermore low-energy electromagnetic model list. A balloon flight in the focal plane of the InFOC$\mu$S mirror assembly was assumed. We accounted for atmospheric absorption at a floating altitude of $130,000$ feet using the NIST XCOM attenuation coefficients [@NIST_XCOM] and an atmospheric depth of $2.9 \, \rm{g/cm}^{2}$ (observations performed at zenith). A Crab-like source was simulated. The X-Calibur modulation factor is $\mu = 0.52$ for a $100 \%$ polarized beam. The minimum detectable polarization [@Krawcz2011] in the $10-80 \, \rm{keV}$ energy range will be $4 \%$ assuming $5.6 \, \rm{hr}$ of on-source observations of a Crab-like source combined with a 1.4 hr background observation of an adjacent empty field. More details about the simulations can be found in Guo et al. (2010) [@Guo2010]. ![image](ComptonSpectrumAllRings.eps){width="50.00000%"} ![image](Phi_ExperimentalGeometry.eps){width="46.00000%"} FIRST MEASUREMENTS {#sec:FirstMeasurements} ================== Using funding from Washington University’s McDonnell Center for the Space Sciences, a flight-ready version of the X-Calibur polarimeter was assembled and tested in the laboratory. First measurements were performed with 3 detector rings installed ($0.5 \, \rm{cm}$ thickness). Before installation, IV-curves were taken for all detector pixels, followed by a calibration run using a Eu$^{152}$ source (lines at $39.9$, $45.7$, $121.8$ and $344.3 \, \rm{keV}$). After calibration, a collimated Eu$^{152}$ source was used to determine the azimuthal X-Calibur acceptance for an unpolarized beam. Another data run was taken without a source to determine the background induced by cosmic rays secondaries hitting the detector assembly. Only CZT events with a simultaneous ($30 \mu \rm{s}$) scintillator trigger are used. A polarized beam was generated by scattering a strong Cs$^{137}$ source (line at $662 \, \rm{keV}$) off a lead brick. A lead collimator allowed only X-rays with a scattering angle of $\sim 90 \deg$ to enter the polarimeter. The X-ray beam of the scattered photons has a mean energy of $288 \, \rm{keV}$ and was polarized to $\sim 55 \%$ (modulation factor of $\mu = 0.4$). The expected relative amplitude in the normalized $\Phi$ distribution is $0.55 \times 0.4 = 0.22$. Figure \[fig:Spectra\] (left) shows the raw spectrum of the first CZT polarimeter ring measured from (i) the polarized beam, (ii) a background spectrum measured without a source and (iii) the excess spectrum corresponding to the energy spectrum of the scattered/polarized beam. As expected, the excess spectrum drops off for energies higher than $288 \, \rm{keV}$ (vertical line) – the energy of the $90 \deg$-scattered Cs$^{137}$ photons entering the polarimeter. The continuum below this energy is the result of $288 \, \rm{keV}$ photons being Compton-scattered at different depths in the scintillator rod and therefore being reflected to the first CZT ring under different scattering angles and corresponding different Compton energy losses. The little bump in the spectrum around $288 \, \rm{keV}$ may originate from direct CZT hits without a Compton-scattering in the scintillator. Figure \[fig:Spectra\] (right) shows the azimuthal scattering distribution of the polarized and unpolarized beam for the second installed CZT detector ring. Only events with a simultaneous scintillator trigger and with a deposited CZT energy between $100-330 \, \rm{keV}$ are used (see spectrum in Fig. \[fig:Spectra\], left). The data of the polarized beam are corrected for the acceptance of the polarimeter (derived from the unpolarized X-ray beam). As expected for a polarized beam, a $180 \deg$ modulation can be seen with a maximum of azimuthal scattering angle perpendicular ($\Phi + 90 \deg$) to the plane of the $E$ field vector of the polarized beam (indicated by the gray arrows). A sine function was fit to the $\Phi$-distribution of the polarized beam resulting in a relative amplitude of $0.22$. The data are in excellent agreement with expectations. SUMMARY AND CONCLUSION {#sec:summary} ====================== We designed a hard X-ray polarimeter X-Calibur and studied its projected performance and sensitivity for a 1-day balloon flight with the InFOC$\mu$S X-ray telescope. X-Calibur combines a detection efficiency of close to $100\%$ with a high modulation factor of $\mu \approx 0.5$, as well as a good control over systematic effects. X-Calibur was successfully tested/calibrated in the laboratory with a polarized beam of $288 \, \rm{keV}$ photons. We applied for a 1-day X-Calibur/InFOC$\mu$S balloon flight. Our tentative observation program includes galactic sources (Crab, HerX-1, CygX-1, GRS1915, EXO0331) and one extragalactic source (Mrk421) for which sensitive polarization measurements would be carried through. We envision follow-up longer duration balloon flights (from the northern and southern hemisphere), possibly using a mirror with increased area. In the ideal case these flights would be performed while the GEMS mission is in orbit to achieve simultaneous coverage in the $0.5 - 80 \, \rm{keV}$ regime. Successful balloon flights would motivate a satellite-borne hard X-ray polarimetry mission. We are grateful for NASA funding from grant NNX10AJ56G and discretionary founding from the McDonnell Center for the Space Sciences to build the X-Calibur polarimeter. Q.Guo thanks the Chinese Scholarship Council for the financial support (NO.2009629064, stay at Washington University). [99]{} H. Krawczynski, et al., “Scientific prospects for hard X-ray polarimetry”, APh, 34 (2011) 550. M. Weisskopf, et al., “Precision measurement of the X-ray polarization of the Crab Nebula without pulsar contamination”, ApJ, 220 (1978) L117. A. J. Dean, et al., “Polarized Gamma-Ray Emission from the Crab”, Science, 321 (2008) 1183. P. Laurent, et al., “Polarized Gamma-Ray Emission from the Galactic Black Hole Cygnus X-1”, Science, 332 (2011) 438. http://www.cern.ch/accelconf/templates.html H. Tajima, et al., “Soft gamma-ray detector for the ASTRO-H Mission”, SPIE, 7732 (2010) 34. J. Poutanen, “Relativistic jets in blazars: Polarization of radiation”, ApJS, 92 (1994) 607. P. A. Connors, & R. F. Stark, “Observable gravitational effects on polarised radiation coming from near a black hole”, Nature, 269 (1977) 128. J. D. Schnittman, & J. H. Krolik, “X-ray Polarization from Accreting Black Holes: The Thermal State”, ApJ, 701 (2009) 1175. J. D. Schnittman, & J. H. Krolik, “X-ray Polarization from Accreting Black Holes: Coronal Emission”, ApJ, 712 (2010) 908. A. L. McNamara, et al., “X-ray polarization in relativistic jets”, MNRAS, 395 (2009) 1507. Y.-Z. Fan, et al., “Gamma-ray burst ultraviolet/optical afterglow polarimetry as a probe of quantum gravity”, MNRAS, 376 (2007) 1857. E. Wulf, et al., “Compton imager for detection of special nuclear material”, NIMA, 579 (2007) 371. Y. Ogasaka, et al., “First light of a hard-x-ray imaging experiment: the InFOCuS balloon flight”, SPIE, 5900 (2005) 217. http://geant4.cern.ch/ http://www.nist.gov/pml/data/xcom/index.cfm Q. Guo, et al., “Design of a Hard X-ray Polarimeter: X-Calibur”, arXiv 1101.0595 (2010).	{ "pile_set_name": "ArXiv" }
--- abstract: 'We present an ab initio quantum theory of the metal-insulator transition in Ti$_2$O$_3$. The recently developed cluster LDA+DMFT scheme is applied to describe the many-body features of this compound. The conventional single site DMFT cannot reproduce a low temperature insulating phase for any reasonable values of the Coulomb interaction. We show that the non-local Coulomb interactions and the strong chemical bonding within Ti-Ti pair is the origin of the small gap insulating ground state of Ti$_2$O$_3$.' author: - 'A.I. Poteryaev' - 'A.I. Lichtenstein' - 'G. Kotliar' title: \| Non-local Coulomb interactions and metal-insulator transition in Ti$_2$O$_3$:\ a cluster LDA+DMFT approach. --- The complicated electronic structure and the nature of the metal-insulator transition in Ti$_{2}$O$_{3}$ and V$_{2}$O$_{3}$ has been the object of intensive experimental and theoretical investigation over the past half century [@Imada]. Recent progress in high-energy photo-emission spectroscopy [@Spring8] and correlated electrons dynamical-mean field theory (DMFT) [@review] has shed new light on the first-order metal-insulator transition in V$_{2}$O$_{3}$. It has been shown that an realistic description of the metallic and insulating phases of V$_{2}$O$_{3}$ can be obtained from the combination of a band structure scheme with the local electron-electron interaction given from DMFT [@Held]. The correlation effects in Ti$_{2}$O$_{3}$ are less clear but angle resolved photo-emission experiment [@Smith] shows a strong reduction of the Ti 3$d$-bandwidth compared to band structure calculations. The important question is related to the mechanism of the small, about 0.1 eV, semiconductor band-gap formation. The generally accepted view is that the metal-insulator transition is related to the decrease of the $c/a$ ratio in rhombohedral Ti$_{2}$O$_{3}$ and the formation of a Ti-Ti pair along $z$-axis [@Goodenough]. Below the broad (almost 250 K in width) metal-insulator transition at around 470 K the Ti-Ti pair distance is seen to decrease without any change of the rhombohedral structure or the formation of long-range antiferromagnetic order [@mag]. This is in contrast to the case of V$_{2}$O$_{3}$ where the V-V pair distance increases within a monoclinic distortion in the antiferromagnetic phase [@Imada]. Ti$_{2}$O$_{3}$ has an $\alpha $-Al$_{2}$O$_{3}$ corundum structure (Fig. \[structure\]) in the metallic and insulating phases with two formula units per rhombohedral cell [@Abrahams_Rice]. Each Ti atom is surrounded by the octahedron of oxygens leading to the large $t_{2g}$-$e_{g}^{\sigma }$ splitting. The trigonal distortion gives an additional splitting of $t_{2g}$ bands into $e_{g}^{\pi }$-$a_{1g}$ states and $a_{1g}$ subbands of Ti-Ti pair form strong bonding-antibonding counterparts (Fig. \[structure\]). In principle, the large decrease of the Ti-Ti distance could split further an occupied single-degenerate $a_{1g}$ states from a double-degenerate $e_{g}^{\pi }$ states of $t_{2g}$ subband and form the insulating $d^{1}$ configuration of this Ti compound. Nevertheless, state of the art LDA calculations have shown that for reasonable Ti-Ti pair distances Ti$_{2}$O$_{3}$ will stay metallic [@Mattheiss]. In order to investigate the role of electron-electron interactions in the formation of this insulating low-temperature phase one needs an accurate estimation of the $a_{1g}$ and $e_{g}$ bandwidths in this complex structure [@Zeiger]. For example a simple free $[$Ti$_{2}$O$_{9}]^{12-}$ cluster mean-field investigation can easily produce a semiconducting gap due to drastic underestimation of the $a_{1g}$ and $e_{g}$ bandwidths [@Nakatsugawa]. On the other hand a more accurate band structure calculation within the unrestricted Hartree-Fock approximation results in large gap antiferromagnetic state [@Catti]. Thus it is crucial to use both the correct Green-function embedding of the Ti-Ti pairs as well as a more accurate treatment of the electron-electron interaction. The role of metal-metal pair formation and the “molecular” versus band pictures of the electronic structure have attracted much attention in these compounds [@pair]. The combination of a strong on-site Coulomb interaction and the large anisotropy between the hopping parameters in and perpendicular to the pair direction can favor a localized molecular-orbital picture of the insulating phase. However, realistic tight-binding calculations for V$_{2}$O$_{3}$ have shown the importance of long-range hopping parameters [@Elfimov]. It is also unclear how good an on-site approximation is for the electron-electron interaction. Since the pair forms a natural “molecular like” element in the corundum-type Ti$_{2}$O$_{3}$ structure it might be expected that non-local electron correlations are important in this system. Thus an approach which combines pair and beyond pair hopping with non-local electron interactions would be seem to be ideal for this problem. In this letter we apply for the first time a method, the cluster DMFT scheme [@cluster; @cdmft], which contains all the physics of correlated pairs in crystals to determine the origin of the insulating phase and the metal-insulator transition in Ti$_{2}$O$_{3}$. A numerically exact multi-orbital Quantum Monte-Carlo (QMC) scheme is deployed for the solution of the cluster DMFT problem and an accurate first principles tight-binding parametrization used for the one electronic structure. Our strategy here is to investigate the gap formation using single site [@anisimov] and cluster LDA+DMFT with only local correlations included. We then deploy the full non-local CDMFT and in this way are able to directly elucidate the impact non-local Coulomb interactions have on the physics. We show that the competition between strong bonding within the Ti-Ti pair and localization from correlation effects leads to the unique situation of the small semiconducting gap structure in Ti$_{2}$O$_{3}$ oxide and that non-local Coulomb correlations are of crucial importance for the physics of this small gap insulators. ![Left - rhombohedral unit cell of Ti$_{2}$O$_{3}$ corundum structure. Titanium ions are indicated by the red color, oxygens by green, and the pair of Ti atoms in $z$ direction by blue. Right - schematic representation of the $t_{2g}$ splitting in Ti$_{2}$O$_{3}$ (top part) and the intersite Coulomb interaction matrix (bottom part). $w$=0 in all our calculations.[]{data-label="structure"}](str1.ps){width=".5\textwidth"} ![Left - rhombohedral unit cell of Ti$_{2}$O$_{3}$ corundum structure. Titanium ions are indicated by the red color, oxygens by green, and the pair of Ti atoms in $z$ direction by blue. Right - schematic representation of the $t_{2g}$ splitting in Ti$_{2}$O$_{3}$ (top part) and the intersite Coulomb interaction matrix (bottom part). $w$=0 in all our calculations.[]{data-label="structure"}](scheme.ps){width=".7\textwidth"} [ $$\! \! \! \! \! \! \! \! \! \! \! \! \! \! \! U_{mm^{\prime}}^{ij} = \left( \begin{array}{ccc} \textcolor{blue}{V_{e_g}} & w & \textcolor{red}{W} \\ w & \textcolor{blue}{V_{e_g}} & \textcolor{red}{W} \\ \textcolor{red}{W} & \textcolor{red}{W} & \textcolor{green}{V_{a_{1g}}} \end{array} \right)$$ ]{} We start with the orthogonal LDA Hamiltonian $H_{mm^{\prime }}^{LDA}(\mathbf{k})$ in the massively downfolded N-th order muffin-tin orbital representation [@OKA] ($m$ corresponds to the 12 $t_{2g}$ orbitals of two Ti-Ti pairs in rhombohedral unit cell) and include different Coulomb interactions (see Fig. \[structure\]). DMFT results for the local and non-local Coulomb interactions are presented in Figs. \[dos\_ss\],\[dos\_2x2\]. The bare LDA density of states (DOS) is shown in Fig. \[dos\_2x2\] by the dashed lines for the low temperature structure (LTS, $\sim$300 K [@Abrahams_Rice]) and high temperature structural (HTS, $\sim$870 K [@Abrahams_Rice]) data on the upper and lower panels respectively. Both LTS and HTS electronic structures are metallic within the LDA scheme. The $a_{1g}$ subband (green dashed line in the Fig. \[dos\_2x2\]) has a strong bonding-antibonding splitting in contrast to the $e_{g}^{\pi}$ subbands (red dashed line). The bandwidth of the HTS is approximately 2.8 eV and smaller than the bandwidth of the LTS (3.2 eV) due to the reduction of the $t_{a_{1g},a_{1g}}$ hopping from -0.85 to -0.63 eV. The cluster DMFT maps the many-body crystal system onto an effective self-consistent multi-orbital quantum impurity-cluster problem [@cluster; @cdmft]. The corresponding Green-function matrix for the Ti-Ti cluster in the LDA+DMFT scheme is calculated via the BZ-integration $$\mathbf{G}(i\omega _{n})=\sum_{\mathbf{k}}[(i\omega _{n}+\mu )\mathbf{1}-% \mathbf{H}^{LDA}(\mathbf{k})-\mathbf{\Sigma }(\omega _{n})]^{-1}, \label{BZI}$$ where $\mu$ is the chemical potential defined self-consistently through the total number of electrons, $\omega _{n}=(2n+1)\pi T$ are the Matsubara frequencies for temperature $T\equiv \beta ^{-1}$ ($n=0,\pm 1,...$) and $\sigma$ is the spin index. The Hamiltonian and the self-energy matrix have the following super-matrix form corresponding to the symmetry of two Ti-Ti pairs in the unit cell $${\scriptstyle \left( \begin{array}{cccc} \mathbf{H}_{11}+\mathbf{\Sigma}_{11} & \mathbf{H}_{12}+\mathbf{\Sigma}_{12} & \mathbf{H}_{13} & \mathbf{H}_{14} \\ \mathbf{H}_{21}+\mathbf{\Sigma}_{21} & \mathbf{H}_{22}+\mathbf{\Sigma}_{11} & \mathbf{H}_{23} & \mathbf{H}_{24} \\ \mathbf{H}_{31} & \mathbf{H}_{32} & \mathbf{H}_{33}+\mathbf{\Sigma}_{11} & \mathbf{H}_{34}+\mathbf{\Sigma}_{12} \\ \mathbf{H}_{41} & \mathbf{H}_{42} & \mathbf{H}_{43}+\mathbf{\Sigma}_{21} & \mathbf{H}_{44}+\mathbf{\Sigma}_{11} \end{array} \right), }$$ where $\mathbf{H}_{ij}(\mathbf{k})$ and $\mathbf{\Sigma }_{ij}(\omega _{n})$ are 3$\times $3 matrices for the $t_{2g}$ states and $\mathbf{\Sigma }_{11}$ and $\mathbf{\Sigma }_{12}$ correspond to the intrasite and intersite contributions to the self-energy respectively. In the self-consistent cluster DMFT scheme the local Green-function (\[BZI\]) should coincide with the corresponding solution of the effective two-site quantum impurity problem [@review] $$\mathbf{G}_{\sigma }(\tau -\tau ^{\prime })=-\frac{1}{Z}\int D[\mathbf{c},% \mathbf{c}^{+}]e^{-S_{eff}}\mathbf{c}(\tau )\mathbf{c}^{+}(\tau ^{\prime }) \label{path}$$ here $\mathbf{c}(\tau )$=$[c_{im\sigma }(\tau )]$ is the super-vector of the Grassman variables, $Z$ is the partition function, $i$ runs over Ti-Ti pair and $m$ runs over $e_{g}^{\pi }$ or $a_{1g}$ orbitals. The effective cluster action $S_{eff}$ is defined in terms of so-called “bath” Green function matrix [@review] $\mathcal{G}_{\sigma}^{-1}(\omega _{n})=\mathbf{G}_{\sigma }^{-1}(\omega _{n})+ \mathbf{\Sigma }_{\sigma }(\omega _{n})$ which describes the energy, orbitals, spin and temperature dependent interactions of particular cluster with the rest of the crystal [$$\begin{aligned} S_{eff} & = & -\int_{0}^{\beta} d\tau \int_{0}^{\beta} d\tau^{\prime} Tr[\mathbf{c}^{+}(\tau)\mathcal{G}^{-1}(\tau,\tau^{\prime}) \mathbf{c}(\tau^{\prime})] + \\ & \frac{1}{2} & \sum_{im,jm^{\prime},\sigma} [ U_{mm^{\prime}}^{ij} n_{im}^{\sigma}n_{jm^{\prime}}^{-\sigma}+ (U_{mm^{\prime}}^{ij}-J_{mm^{\prime}}^{ij})n_{im}^{\sigma}n_{jm^{\prime}}^{\sigma}] \notag \end{aligned}$$ ]{} here $n_{im\sigma }$=$c_{im\sigma }^{+}c_{im\sigma }$. We have parameterized the screened local Coulomb and exchange matrices ($U_{mm^{\prime }}^{ii}$ and $J_{mm^{\prime }}^{ii}$) for the $t_{2g}$ electrons in terms of average Coulomb and exchange integrals [@Kotliar] and used a simple approximation to the intersite $U_{mm^{\prime }}^{ij}$ interactions as shown in the Fig. \[structure\]. The multi-band impurity QMC scheme [@Hirsch; @rozenberg] has been used for the numerically exact calculation of the cluster Green function (eq. \[path\]). The number of auxiliary Ising fields in the discrete Hirsh-Fye transformation were 15 and 58 for the local and non-local Coulomb interaction schemes respectively. For accurate QMC integration we used of the order of 10$^{6}$ sweeps, with 8000 $\mathbf{k}$-points for the BZ-integration. Within 15-20 DMFT iterations convergence in the self-energy was reached. The maximum entropy method [@MEM] has been used for analytical continuation of the diagonal part of the Green function matrix to the real energy axis. ![Density of states for the single site (upper panel) and cluster (lower panel) DMFT calculations with different values of the Coulomb repulsion $U$ and $J$=0.5 eV. Inset: $N(0)$ $versus$ Coulomb parameter. Filled dark green circles - DMFT results with $J$=0.5 eV, filled orange squares - DMFT with $J$=0 eV. Open blue circles and violet squares are CDMFT with $J$=0.5 and 0 eV respectively.[]{data-label="dos_ss"}](dos_ss.ps){width=".45\textwidth" height=".4\textwidth"} Firstly, in Fig.\[dos\_ss\] we show the total density of states for both conventional single site (DMFT) and the cluster (CDMFT) dynamical mean field theory where only local electron correlations have been included. The QMC simulation has been carried out for $\beta$=20 eV$^{-1}$ which corresponds to a temperature of $T$$\simeq$580 K which is on the border of the metal-insulator transition. In the upper panel of Fig. \[dos\_ss\] are shown the results of DMFT calculations with $U$=2,3,4 eV and exchange parameter $J$=0.5 eV. For all values of Coulomb interactions there is a peak below the Fermi level at around -0.5 eV, predominantly of $% a_{1g}$ character with in all cases the same intensity. Above the Fermi level there are two peaks. The first is at 0.5 eV and has $e_{g}$ character while the other peak is strongly dependent on the Coulomb parameter and can be associated with an upper Hubbard band. A lower Hubbard band can be seen at around -2 eV. We see that for all values of $U$ the the shape of pseudogap is unchanged and the system remains metallic. On the lower part of Fig. \[dos\_ss\] the results of the CDMFT calculation are shown for the same values of the Coulomb and exchange parameters. The general structure of the DOS is seen to be similar to the single site calculation, however one may note interesting differences. The lower $a_{1g}$ quasiparticle band is decreased in intensity and shifted towards the Fermi level from -0.6 eV to -0.3 eV on increasing $U$ from 2 to 4 eV. This has the result that for $U$=4 eV the pseudogap is now located directly at Fermi level, whereas for other $U$-values and for all DMFT results it lies on the slope of quasiparticle peak. Using the temperature DOS at the Fermi level, defined as $N(0) \equiv -ImG(\omega_0)/\pi$ with $\omega_0=\pi T$ we are able to estimate at what values of $U$ the system will become insulating. This is indicated in the inset in the upper panel of the Fig. \[dos\_ss\]. We see that for the single site calculations $N(0)$ depends weakly on $U$ and the system will remain metallic up to very large values, about 8 eV, of the Coulomb parameter. On the other hand for the cluster calculation $N(0)$ is seen to decrease strongly as a function of $U$ for both values of exchange parameter, and the critical value for an insulating solution is now lower at $U$$\sim$5-6 eV. As expected for the $d^1$ configuration the finite value of the exchange parameter effectively decreases the Coulomb interaction matrix. We see the single site results are in greater contradiction to the experiment as compared to LDA (see Fig. \[dos\_2x2\]): the local Coulomb interaction leads to the reduction of the bonding-antibonding splitting of the $a_{1g}$ subband and this acts to suppress gap formation. On the other hand in the cluster case a small semiconducting gap is developed for large $U$ due to dynamical antiferromagnetic correlation in the Ti-Ti pair. Nevertheless, using either the DMFT or CDMFT schemes with only local correlations there remains a dramatic absence of gap formation in Ti$_2$O$_3$. We now deploy the full non-local correlation in CDMFT to the effect of non-local correlations on low and high temperature electronic structure. We have used different values of the non-local Coulomb parameters and found that the most important correspond to non-diagonal interactions. For both structures we have chosen values of $U$=2 eV and $J$=0.5 eV which are close to those from constrained LDA estimations [@uvalue], while the off diagonal Coulomb parameter $W$ has been chosen at $W$=0.5 eV [@wvalue]. On the upper panel of Fig. \[dos\_2x2\] is shown the total and partial DOS for $\beta$=20 eV$^{-1}$. Shown also is the LDA result. One can see that for the reasonable parameters chosen we can reproduce the correct value of the semiconducting gap $\sim$0.1 eV while keeping the bonding-antibonding splitting on the LDA level. In the lower panel the high temperature metallic solution corresponding to $\beta$=10 eV$^{-1}$ is shown. Here we emphasize that the proper inclusion of the structural effect on the LDA level is important as evinced by the fact that for $\beta$=20 eV$^{-1}$ and high temperature hamiltonian we again obtain a metallic solution. The $e_g$ states are similar for both LTS and HTS calculations with a small shift of occupied part in LTS case. However, the difference between the LTS and HTS phases is more pronounced for the $a_{1g}$ states. The bonding-antibonding splitting in the LTS is about 2 eV whereas in the HTS case it is only 1.5 eV. The occupied $a_{1g}$ states in the LTS phase are shifted down opening the insulating gap. The important difference between the large $U$ and small $U$ plus non-local $W$ is the absence of well defined Hubbard bands. This absence makes possible a critical test of the theory proposed here, and thus it would be very interesting for photo-emission experiments to check the existence or not of a lower Hubbard band at around -2 eV. ![Partial and total CDMFT (solid line) compared to the LDA (dashed) DOS with $W$=0.5 eV and $V_{a_{1g}}$=$V_{e_{g}}$=0. Total DOS are shown by black, the $e_{g}$ states by red, and $a_{1g}$ states by green. On the upper panel the low temperature structure and $\beta$=20 eV$^{-1}$ are used. For the lower panel the high temperature structure and $\beta$=10 eV$^{-1}$ are used. The diagonal and the biggest $a_{1g}$ off-diagonal Green functions $G(\tau)$ are shown in the upper inset. In the lower inset the $Re \Sigma_{a_{1g},a_{1g}}$ with $W$=0.5 eV are shown by blue, $Re \Sigma^{\prime}_{a_{1g},a_{1g}}$ with $W$=0 eV by orange and $Im \Sigma_{a_{1g}}$ are shown by green[]{data-label="dos_2x2"}](dos_2x22.ps){width=".45\textwidth"} We have shown that the cluster LDA+DMFT calculation with a moderate Coulomb repulsion among the $a_{1g}$ orbitals is essential to produce the high temperature semimetallic state and the low temperature insulating state. To understand the role play of the intersite Coulomb interaction we focus on the the quantity $t_{a_{1g},a_{1g}} + Re \Sigma_{a_{1g},a_{1g}}(i \omega)$ which we can interpret as a frequency dependent “effective $a_{1g}-a_{1g}$ hopping” which describes the hopping matrix element in the titanium pair. We find that this quantity is surprisingly frequency dependent (see lower inset of Fig. \[dos\_2x2\]). We conclude that the main role of the intersite Coulomb interaction is dynamic (the Hartree contribution to this quantity is small) and results in the effective $a_{1g}-a_{1g}$ hopping that [increases]{} as the frequency decreases. This enhancement produces a strong level repulsion of the bonding antibonding $a_{1g}$ levels, lowering the $a_{1g}$ level relative to the $e_g$ level at the low frequency. This effect combined with a small narrowing of the $a_{1g}$ band opens the $e_g-a_{1g}$ band gap which results in the insulating state. We checked that this enhancement of the effective hopping as frequency is decreased is absent if we turned off the intersite Coulomb repulsion. This effect is the cluster DMFT analog of a mechanism first discussed in the context of the single impurity model by Haldane [@haldane]. He observed, that a Coulomb repulsion between the impurity level and additional conduction electron states or screening channels, [enhances]{} the hybridization (single impurity analog of the hopping matrix element) as we renormalize to low frequency. We would like to acknowledge O.K. Andersen, V.I. Anisimov, A. Georges and M.I. Katsnelson for useful discussions. This work was supported by the Netherlands Foundation for the Fundamental Study of Matter (FOM). GK was supported by the ONR, grant N000140210766. The authors are grateful to the Kavli Institute of Theoretical Physics, Santa Barbara, for hospitality during the initial stages of this work. This research was supported in part by the National Science Foundation under Grant No. PHY99-07949. M. Imada, A. Fujimori, Y. Tokura, Rev. Mod. Phys. [70]{}, 1039 (1998). S.-K. Mo, J.D. Denlinger, H.-D. Kim [et. al.]{} Phys. Rev. Lett. [90]{}, 186403 (2003). A. Georges, G. Kotliar, W. Krauth, M. Rozenberg, Rev. Mod. Phys. [68]{}, 13 (1996). K. Held, G. Keller, V. Eyert, D. Vollhardt, V.I. Anisimov, Phys. Rev. Lett. [86]{}, 5345 (2001). K.E. Smith and V.E. Henrich, Phys. Rev. B. [38]{}, 5965 (1988). J. B.Goodenough, Prog. Solid State Chem. [5]{}, 145 (1972). R.M. Moon, T. Riste, W.C. Koehler, S.C. Abrahams, J. Appl. Phys. [40]{}, 1445 (1969). S.C. Abrahams, Phys. Rev [130]{}, 2230 (1963). C.E. Rice, W.R. Robinson, Acta Cryst. B [33]{}, 1342 (1977). L.F. Mattheiss, J. Phys.: Condens. Matter [8]{}, 5987 (1996). H. J. Zeiger, Phys. Rev. B. [11]{}, 5132 (1975). H. Nakatsugawa and E. Iguchi, Phys. Rev. B. [56]{}, 12931 (1997). M. Catti, G. Sandrone, and R. Dovesi, Phys. Rev. B. [55]{}, 16122 (1997). R. Shiina, F. Mila, F.-C. Zhang, T. M. Rice, Phys. Rev. B. [63]{}, 144422 (2001). I.S. Elfimov, T. Saha-Dasgupta, M.A. Korotin, cond-mat/0303404 (2003). A. I. Lichtenstein and M. I. Katsnelson, Phys. Rev. B [62]{}, 9283(R) (2000). G. Kotliar, S.Y. Savrasov, G. Palsson, G. Biroli, Phys. Rev. Lett. [87]{}, 186401 (2001) . V.I. Anisimov, A.I. Poteryaev, M.A. Korotin, A.O. Anokhin, G. Kotliar, J. Phys.: Condens. Matter [9]{}, 7359 (1997). A.I. Lichtenstein and M.I. Katsnelson, Phys. Rev. B [57]{}, 6884 (1998). O.K. Andersen and T. Saha-Dasgupta, Phys. Rev B [62]{}, 16219 (2000). R. Frésard and G. Kotliar, Phys. Rev. B [56]{}, 12909 (1997). J.E. Hirsch and R.M. Fye, Phys. Rev. Lett. [56]{}, 2521 (1986). M.J. Rozenberg, Phys. Rev. B [55]{}, R4855, (1997). M. Jarrell and J.E. Gubernatis, Physics Reports [269]{}, 133 (1996). I. Solovyev, N. Hamada, K. Terakura, Phys. Rev. B [53]{}, 7158 (1996). G. Lucovsky, R.J. Sladek, J.W. Allen, Phys. Rev. B [16]{}, 5452 (1977). F.D.M. Haldane, Phys. Rev. Lett. [40]{}, 416 (1978) and Ph. D. Thesis University of Cambridge, 1977. See also Q. Si and G. Kotliar, Phys. Rev. Lett. [70]{}, 3143 (1993). Q. Si and G. Kotliar, Phys. Rev. B [48]{}, 13881 (1993).	{ "pile_set_name": "ArXiv" }
--- abstract: 'Let $M$ be a two dimensional complex manifold, $p \in M $ and a germ of holomorphic foliation of at $p$. Let $S\subset M$ be a germ of an irreducible, possibly singular, curve at $p$ in $M$ which is a separatrix for . We prove that if the Camacho-Sad-Suwa index ${\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},S,p)\not \in {\mbox{ $ \mathbb{ Q}$}}^+\cup \{0\} $ then there exists another separatrix for at $p$. A similar result is proved for the existence of parabolic curves for germs of holomorphic diffeomorphisms near a curve of fixed points.' address: 'Dipartimento di Matematica, Università di Pisa, via Buonarroti 2, 56127, Pisa.' author: - Francesco Degli Innocenti title: Holomorphic dynamics near germs of singular curves --- Introduction ============ Let $M$ be a two dimensional complex manifold and a germ of holomorphic foliation on $M$ near $p$. In local coordinates the foliation can be described by the vector field: $$A(x,y)\frac{\partial}{\partial x}+B(x,y)\frac{\partial}{\partial y},$$ with $A, B$ suitable holomorphic functions. A separatrix for is a non constant holomorphic solution of the system: $$\left \{ \begin{aligned} &\dot{x}=A(x,y)\\ &\dot{y}=B(x,y) \end{aligned}\hspace{2cm} x(0)=y(0)=0. \right.$$ Obviously the interesting case is when $(0,0)$ is a singularity for . In the singular case, in the well known paper [@CS], Camacho and Sad proved that there always exists (at least) one (possibly singular) irreducible separatrix - say $S$ - for at $(0,0)$. A natural question is whether the knowledge of this separatrix $S$ allows to infer the existence of another separatrix. There are essentially two types of results, one of local and the other of global flavour. The first kind of result is essentially a re-formulation of Camacho-Sad theorem (see the paper by J. Cano [@CA]) which says that if $S$ is non singular and $ {\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},S,p)\not \in {\mbox{ $ \mathbb{ Q}$}}^{+}\cup \{0\}$ (where ${\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},S,p)$ is the index introduced in [@CS]) then there exists another separatrix through $p$. The second type of result requires global conditions on $S$, like $S$ compact (but possibly singular), globally and locally irreducible and $S \cdot S < 0$ to provide the existence of another separatrix at some point of $S$ (see the paper by Sebastiani [@SB]). One aim of this paper is to prove a result of local nature when $S$ is possibly singular, using the index defined by Suwa [@SW]. We prove: \[teor:miomio\] Let be a complex two dimensional manifold, a holomorphic foliation on same open subset of $M$, $S \subset M $ a possibly singular curve locally irreducible at a point $p\in M $, such that it is a separatrix for at $ p $. If $ {\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},S,p)\not \in {\mbox{ $ \mathbb{ Q}$}}^+\cup\{0\} $ then there exists (at least) another separatrix for at $p $. Abate, Bracci and Tovena [@ABT], [@BR], [@BT] have recently shown how to translate results about foliations to holomorphic diffeomorphisms with curves of fixed points. The proof of Theorem \[teor:miomio\] respects their dictionary and so the results about the existence of separatrices for foliations can be translated in results about the existence of parabolic curves for diffeomorphisms. Using notations of [@ABT], [@BR] and [@BT] we obtain: \[teor:diff\] Let $M$ be a two dimensional complex manifold, a holomorphic map such that $\hbox{Fix}(f)=S$ with $S$ a locally irreducible, possibly singular curve at a point $p\in M$. Assume that $f$ is tangential on $S$ and ${\mbox{ \hbox{Ind}}}(f,S,p)\not \in {\mbox{ $ \mathbb{ Q}$}}^+\cup \{0\} $. Then there exists (at least) a parabolic curve for $f$ at $p$. Theorem \[teor:diff\] has been proved by Abate [@AB] in case $S$ is non singular and by Bracci in [@BR] in case $S$ is a generalized cusp, i.e. of the form $\{x^m=y^n\}$. I want to sincerely thank professor Filippo Bracci without whose help this work would not have came to be. Preliminary results =================== First of all we have to recall some basic notions about C.S.S. (Camacho-Sad-Suwa) index. This index was first introduced by Camacho and Sad in [@CS] for a complex one codimension singular foliation defined in a neighborhood of a non singular compact curve embedded in a two dimensional complex manifold. Later Suwa [@SW] generalized to a generic possibly singular compact invariant curve. The most interesting property of this is the following Index Theorem, that relates the dynamics of near a curve $S$ to the self intersection number of $S$. Let $S$ be a compact curve in a two dimensional complex manifold $M$ invariant by a possibly singular foliation ,then for every point $p\in S$ there exists a complex number ${\mbox{ \hbox{Ind}}}({\mbox{ $ \mathcal{F} $ }},S,p)\in {\mbox{ $ \mathbb{C} $ }}$ depending only on the local behaviour of and $S$ near $p$ such that: $$\sum_{p\in S}{\mbox{ \hbox{Ind}}}(\mathcal{F},S,p)= S \cdot S.$$ We now recall the behaviour of this index under blow-up. \[propos:blow-up\] Let $M $ be a two dimensional complex manifold, an holomorphic foliation, $S$ an -separatrix and $p\in S $ a singularity of $S$. We indicate by $ \pi: \tilde{M}\longrightarrow M $ the blow-up of $M$ in $p$, by $ \tilde{\mathcal{F}} $ the saturated foliation and by $ D :=\pi^{-1}(p) $ and $ \hat{S}:=\overline{\pi^{-1}(S\setminus \{p\})} $ respectively the exceptional divisor and the strict transform of $S$. Then $\hat{S} $ is an $\tilde{\mathcal{F}}$ separatrix. Moreover if $ \{\tilde{p}\}:=D\cap\hat{S} $ then $$\hbox{Ind}(\tilde{\mathcal{F}},\hat{S},\tilde{p})=\hbox{Ind}(\mathcal{F},S,p)-m^2$$ where $ m \geq 1 $ is the multiplicity of $S$ in $p$. Cano in [@CA] gives an algorithmic proof of Camacho-Sad result introducing a particular class of points that we will often use. Let $M$ be a two dimensional complex manifold, an holomorphic foliation and $S$ a local separatrix for . - We say that a point $p\in S$ is of type $(C_1)$ if $S$ is nonsingular at $p$ and $${\mbox{ \hbox{Ind}}}({\mbox{ $ \mathcal{F} $ }},S,p)\not \in {\mbox{ $ \mathbb{ Q}$}}^+ \cup \{0\}.$$ - We say that a point $p\in S$ is of type $(C_2)$ if $S$ has two nonsingular branches $S_0$, $S_1$ at $p$, intersecting trasversally at $p$, and there exists a real number $r>0$ such that $$\begin{aligned} &{\mbox{ \hbox{Ind}}}(\mathcal{F},S_0,p)\not \in {\mbox{ $ \mathbb{ Q}$}}_{\geq -\frac{1}{r}}=\{a\in {\mbox{ $ \mathbb{ Q}$}}: a\geq -\frac{1}{r}\}\\ &{\mbox{ \hbox{Ind}}}(\mathcal{F},S_1,p)\in {\mbox{ $ \mathbb{ Q}$}}_{\leq -r}=\{a\in {\mbox{ $ \mathbb{ Q}$}}: a\leq -r\}. \end{aligned}$$ According to Definition 7.6 of [@BR] and [@CA] we have: A point $p\in S \subset M$ where $S$ is an -invariant curve is said to be an appropriate singularity for if after a finite number of blow-ups there exists a $(C_1) $ or $(C_2)$ point on the total transform. The importance of this class of points is given by the following result: If $p\in S \subset M$ is an appropriate singularity for a foliation , then at least another separatrix trough $p$ for exists. Proof of the result =================== In order to get Theorem \[teor:miomio\] we will concentrate our attention on the particular class of points introduced in the previous section. The upshot is to show that under the hypotheses of Theorem \[teor:miomio\] the point $p$ is an appropriate singularity. We know that the resolution of curves singularities theorem [@LAU] ensures that after a finite number of blow-ups we have the geometric structure required for the existence of $(C_1)$ or $(C_2)$ points. To conclude we have to analize the C.S.S. index under this process. The behaviour of the index is strongly related to the evolution of the geometric structure under blow-up. We can divide the proof in two steps: 1. study of the geometric structure under the resolution of singularities, 2. study of the C.S.S. index under this process. Geometric structure under blow-up --------------------------------- In order to get step one we give the following definition: Let $M$ be a two dimensional complex manifold and $S_1,\cdots,S_n \subset M$ given curves. We say that a point $p$ is a double intersection point if $p$ belongs to exactly two distinct curves among $S_1,\cdots,S_n$. If instead $p$ belongs to exactly three of them it is called a triple intersection point. \[osserv:o1\] In the study of curve desingularization the set of curves we find is composed by the strict transform of the curve $S$ and the several exceptional divisors obtained by succesive blow-ups. Because of the structure of the blow-up process we can only have double and triple intersection points (see [@LAU]). A triple intersection point belongs to the strict transform of $S$ and to two exceptional divisors. To distinguish these two ${\mbox{ $ \mathbb{C} $ }}\mathbb{P}^1$ we will call old exceptional divisor the strict transform of a given exceptional divisor. Instead we will call new exceptional divisor the exceptional divisor produced by the last blow-up. Now we can describe the geometric evolution under blow-up. Note that the only intersection point that can be triple is the one made up by the strict transform of $S$. We will prove the following behaviour. \[propos:ordine\] Let $S$ be a singular curve and let $p$ be a singularity of $S$. The resolution process of $S$ in $p$ is related to the behavior of the multiplicity of $S$ in $p$ in the following way: - If we blow-up a singularity and the multiplicity does not reduce we have two cases: 1. if we are in a double intersection point at the next blow-up we find another double intersection, 2. if we are in a triple intersection point at the next blow-up we can find either a double intersection or a triple intersection point. More precisely we find a double intersection point if the tangent cone to the curve does not coincide with any exceptional divisor, while we find a triple intersection point if the tangent cone coincides with one of the two exceptional divisors and the new triple intersection point belongs to the strict transform of the old exceptional divisor. - If we blow-up a singularity and the multiplicity reduces we have two cases: 1. if we are in a double intersection point at the next blow-up we find a triple intersection point, 2. if we are in a triple intersection point at the next blow-up we find a triple intersection point that belongs to the strict transform of the new exceptional divisor. In the previous Proposition we have used inproperly the expression “the tangent cone coincides with one of the two exceptional divisors” to mean that the tangent cone of $S$ in $p$ coincides with the tangent space of $D$ in $p$. In order to get Proposition \[propos:ordine\] we will prove some elementary Lemmas. \[lemma:giallo\] Let $M$ be a two dimensional complex manifold, $S$ an analytic irreducible curve on $M$ and $p\in S$ a singularity of $S$. Blow-up $M$ in $p$ and let $\hat{S} $ be the strict transform of $S$, $D$ the exceptional divisor and $\hat{p}:=\hat{S} \cap D$. The multiplicity of $\hat{S}$ in $\hat{p}$ is strictly smaller than the multiplicity of $S$ in $p$ if and only if $D$ coincides with the tangent cone of $\hat{S}$ in $\hat{p}$. We can assume that $ p=(0,0) $ and $ S = \{ l (x,y)=0\} $ with $ l (x,y)= y^m + l_{m+1}(x,y) + \cdots .$ Blow-up in $p$ and using the chart such that the projection becomes $\pi(u,v)=(u,uv)$ we have: $ \hat{S}=\{\hat{l}(u,v)=0\} $, with $ l(u,v)=v^m + u l_{m+1}(1,v)+ \cdots = v^m + u q_{k-1} + \cdots $ and $ D = \{u=0\} . $ The multiplicity of $ \hat{S} $ in $ (0,0) $ is strictly less then $ m $ if and only if $ k<m $ and then if and only if the tangent cone is $ \{ u q_{k-1}(u,v)=0\} $ and so if and only if $ D $ is included in the tangent cone. Because $S$ is irreducible this can happen if and only if $ q_{k-1}(u,v)=u^{k-1}$, i.e. if and only if $ D $ is the tangent cone. \[lemma:viola\] Let $M$ be a two dimensional complex manifold, $S$ an analytic irreducible curve on $M$ and $p\in S$ a singularity of $S$. Blow-up $M$ in $p$ and let $\hat{S} $ be the strict transform of $S$, $D$ the exceptional divisor and $\hat{p}:=\hat{S} \cap D$. The exceptional divisor $D$ is the tangent cone of $\hat{S}$ in $\hat{p}$ if and only if blowing-up in $\hat{p}$ we get a triple intersection point. Let $\hat{D}$ be the strict tranform of $D$ and $D_1$ the new exceptional divisor. Now $\hat{D}$ intersects $D_1$ in the point corresponding to the tangent of $D$ in $p$, so $\hat{D}\cap \hat{\hat{S}} \neq \emptyset$ if and only if $D$ and $\hat{S}$ have the same tengent in $p$. So we get a triple intersection point if and only if the tangent cone of $\hat{S}$ coincides with $D$. Using the previous two Lemmas we obtain the following: Let $ S \subset M $ be an analytic irreducible curve of multiplicity $ m $ in the singular point $p$. Suppose that after a finite number of blows-up the strict transform of $S$, $ \tilde{S} $, intersects the exceptional divisor in a point $\tilde{p}$ and indicate with $D$ the irreducible component of the exceptional divisor conteining $\tilde{p}$, i.e. $\tilde{p}$ is a double intersection point. Blow-up in $\tilde{p}$ and let $ D_1 $ be the new exceptional divisor and $\hat{S} $ the strict transform of $ \tilde{S} $. If the multiplicity of $ \hat{S} $ in $ \hat{p}:=D_1\cap \hat{S} $ is equal to the multiplicity of $\tilde{S} $ in $\tilde{p}$ then at the following blow-up we find again a double intersection point. By Lemma \[lemma:viola\] we also get: \[lemma:verde\] Let $ S \subset M $ be an analytic irreducible curve of multiplicity $ m $ in the singular point $p$. Suppose that after a finite number of blows-up we have a triple intersection point. At the following blow-up we have two cases: 1. if the tangent cone in the singularity contains one of the two exceptional divisors then at the next blow-up we find agin a triple intersection point, 2. if the tangent cone in the singularity does not contain any of the two exceptional divisors then at the next blow-up we find a double intersection point. \[osserv:blu\] We observe that the demonstrative method used in Lemma \[lemma:verde\] does not give informations on which of the exceptional divisors goes to create the new triple intersection. To get this information we need some more calculations. Let $\hat{S}$ the strict transform of $S$ after some blow-ups and suppose to have a triple intersection point. We can assume that $p=(0,0) $ and $ \hat{S}=\{\hat{l}(u,v)=0\} $ with $\hat{l}(u,v)= v^{m} + u^{k_1}[q_{k_2-k_1}(u,v)+\cdots] ,$ and $ D_1=\{v=0\} $ , $ \hat{D}=\{u=0\}$ where $D_1$ is the new exceptional divisor and $\hat{D}$ is the old one (according to Remark \[osserv:o1\]). Let esaminate the various cases: 1. If $ m>k_2 $ then the tangent cone is $ \{u^{k_1}q_{k_2-k_1}(u,v)=0\} $ and by the irreducibility of $S$ is $ \{cu^{k_2}=0\} $ with $ c\neq 0 $ and so it contains an exceptional divisor, $ \hat{D} .$ Blow-up again $ (0,0) $ and using the chart by which the projection is $ \pi(x,y)=(xy,y)$ we have: $$\hat{l}(xy,y)= y^{m} + c x^{k_2}y^{k_2} + x^{k_1 }y^{k_2 +1}[ q_{k_2-k_1+1}+ \cdots ]$$ and because $ m>k_2 $ $$\hat{\hat{l}}(x,y)=y^{m_1-k_2}+ c x^{k_2} + x^{k_1} y[q_{k_2-k_1+1}(x,1)+\cdots ]$$ with $ D_2=\{y=0\}$ e $ \hat{\hat{D}}=\{x=0\} .$ So $ (0,0)$ is a triple intersection point made up by $D_2$, $\hat{\hat{S}}$, $\hat{\hat{D}}$. If instead we use the other chart we find only a double intersection points. 2. If $ m_1<k_2 $ we proceed in the same way obtaining a triple intersection point made by $D_2$, $\hat{D_1}$ and $\hat{\hat{S}}$. 3. If $ m_1=k_2 $ the tangent cone is given by $ \{v^{m_1}+u^{k_1}q_{k_2-k_1}(u,v)=0\} $ and by the irreducibility of the curve it is $ \{(v+c u)^{m_1}=0\} $ with $ c\neq 0 $ and it does not contain any exceptional divisor. So by Lemma \[lemma:verde\] at the next blow-up we find only double intersection points. C.S.S. index under blow-up -------------------------- Now we can proceed in order to get step two studying the behaviour of the index in a general resolution process via blow-up. The upshot is to prove that in the resolution process we necessarily find a $(C_1)$ or $(C_2)$ point ,i.e., $p$ is an appropriate singularity and then Theorem \[teor:miomio\] holds. The intent is to analyze the C.S.S. index in all possible geometric evolutions (see Proposition \[propos:ordine\]). In the analysis we will omit the case in which at some blow-up we find a dicritical point (see Definition 3.2 in [@BR]). In fact in this case the goal is obtained by Proposition 7.8 [@BR] and by the proper mapping theorem [@GR]. We will consider resolution processes only at a combinatoric level in a sense that will be specified later. Thanks to Proposition \[propos:ordine\] the structure of a resolution process of a singular point $p$ is completely described by the behaviour of the multiplicity of the strict transform at the intersection with the exceptional divisor. We can then consider a sequence of blow-ups only as a sequence of positive number (representing the evolution of the multiplicity) and forgetting any type of geometric obstruction. A process is an ordinate list of the form: $$P=\{(k, m), (\alpha_1, m_1), \cdots, (\alpha_n ,m_n)\}$$ where $k,\alpha_i, m_i \in \mathbb{N}$ and $ m> m_1\geq \cdots \geq m_n$. We associate to P, from a purely formal point of view, a blow-up sequence for a curve $S$ where the blows-up are made at the beginning at the point $p$ and then at the intersection point of the strict transform of the curve and the exceptional divisor. The blow-up sequence satisfies the following rules: - from the first to the $k-th$ blow-up we find only double intersection points and the curve multiplicity is constantly equal to $m$, - from the $(k+1)-th$ to the $(k+\alpha_1)-th$ blow-up we find a triple intersection point and the multiplicity of the strict transform of $S$ is constantly equal to $ m_1<m $, - - from the $(k+\alpha_1+\cdots +\alpha_{n-1}+1)-th$ to the $(k+\alpha_1 + \cdots \alpha_n)-th$ blow-up we find a triple intersection point and the multiplicity is constantly equal to $m_n \leq m_{n-1}$. At the end of $P$ the curve $S$ is not desingularized, in fact we have triple points and this type of point are not admitted in the desingularized curve. Now, according to Proposition \[propos:ordine\] we start to analyze all the possible cases. For notations we refer to [@BR] and [@BRU]. Case of double intersection --------------------------- It corresponds to a process $ P=\{(k, m)\}$, i.e. we start with multiplicity $m$ and we remain with this multiplicity for $k$ blows-up finding only double points. If we do not find $(C_1)$ or $(C_2)$ points in the total transform then (arguing as in Proposition 7.8(2) of [@BR]) at the $k-$th blow-up the indices are of type: $$\label{eq:doppia1} \begin{aligned} {\mbox{ \hbox{Ind}}}(\tilde{{\mbox{$\mathcal{F}$}}},D,q)&\in {\mbox{ $ \mathbb{ Q}$}}_{\leq-\frac{1}{k}}\\ {\mbox{ \hbox{Ind}}}(\tilde{{\mbox{$\mathcal{F}$}}},\hat{S},q)&\not\in {\mbox{ $ \mathbb{ Q}$}}_{\geq - k m^2}. \end{aligned}$$ where $q:= \hat{S}\cap D $. Case of triple intersection --------------------------- We consider now a slightly more complicated process,\ $P=\{(k,m),(1,m_1),\cdots,(1,m_n)\}$. Let us suppose not to find $(C_1)$ or $(C_2)$ points during $P$. We indicate at the last blow-up with $S$ the strict transform of the curve, the saturated foliation, $D_1$, $D_2$ the two exceptional divisors that intersect, with $S$, in the last triple intersection point $q$. \[propos:ciao\] In this situation at the last blow-up of $P$, if we have not found $(C_1)$ or $(C_2)$ points, we can find $ x,y\in\mathbb{N} $ and $ a,b \in\mathbb{N}\cup\{0\}$ such that the indices are: $$\label{eq:i2} \begin{aligned} &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},S,q)\not\in {\mbox{ $ \mathbb{ Q}$}}_{\geq-km^2- m_1^2 - \cdots - m_n^2}\\ &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},D_1,q)\in {\mbox{ $ \mathbb{ Q}$}}_{\leq -\frac{x}{y}}\\ &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},D_2,q)\in {\mbox{ $ \mathbb{ Q}$}}_{\leq -\frac{y k + a}{x k + b}}. \end{aligned}$$ At the k-th blow-up the indices are of type . Let blow-up again. As P describes we have a multiplicity decrease and we find a triple point on the total transform. Then if some point of the new exceptional divisor $D_1$ is of type $(C_1)$, $p$ is an appropriate singularity and we have the assertion. Otherwise ${\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},D_1,p)\in{\mbox{ $ \mathbb{ Q}$}}_{\geq 0}$ $ \forall p\in D_1\setminus \{q\}$ and then by Index Theorem: $${\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},D_1,q) \in {\mbox{ $ \mathbb{ Q}$}}_{\leq-1}.$$ Then by Proposition \[propos:blow-up\] and observing that $D$ has multiplicity one: $$\begin{aligned} &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},\hat{S},q) \not \in {\mbox{ $ \mathbb{ Q}$}}_{\geq -k m^2 - m_1^2}\\ &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},\hat{D_2},q)\in {\mbox{ $ \mathbb{ Q}$}}_{\leq-\frac{k+1}{k}}. \end{aligned}$$ Proceeding by induction on $n$ we can assume the assertion true for $n$ and we prove it for $n+1$. We have to analyze separately two different cases that can occur blowing-up: 1. the new triple point is made by $ \{\hat{S},\hat{D}_2,D\} ;$ 2. the new triple is made by $ \{\hat{S},\hat{D}_1,D\},$ where $D$ is the new exceptional divisor and $D_1$ and $D_2$ are the ones of the $n$ blow-up whose indices satisfy by inductive hypothesis. We consider only the case $(1)$ because the other is similar. By Proposition \[propos:blow-up\] the indices are of type: $$\begin{aligned} &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},\hat{S},q_1)={\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},S,p)-m_1^2- \cdots -m_n^2 - m_{n+1}^2\\ &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},\hat{D}_2,q_1)\in {\mbox{ $ \mathbb{ Q}$}}_{\leq -\frac{(x+y)k+(a+b)}{xk+b}}\\ &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},\hat{D}_1,q_0)\in {\mbox{ $ \mathbb{ Q}$}}_{\leq -\frac{x+y}{y}}, \end{aligned}$$ where $q_1$ is the new triple point and $q_0:=\hat{D_1} \cap D$. If there are not $ (C_1) $ points on $ D\setminus\{q_0,q_1\} $ then by Index Theorem $ q_0 $ is a $ (C_2) $ point or $ {\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},D,q_1)\in {\mbox{ $ \mathbb{ Q}$}}_{\leq -\frac{x}{x+y}} .$ In the last case the indices satisfy: $$\begin{aligned} &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},\hat{S},q_1)={\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},S,p)-m_1^2- \cdots -m_n^2 - m_{n+1}^2\\ &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},\hat{D}_2,q_1)\in {\mbox{ $ \mathbb{ Q}$}}_{\leq -\frac{(x+y)k+(a+b)}{xk+b}}\\ &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},D,q_1)\in {\mbox{ $ \mathbb{ Q}$}}_{\leq -\frac{x}{x+y}}. \end{aligned}$$ and then the assertion follows putting $ y'= x+y, x'=x, a'=a+b, b'=b.$ \[osserv:proc\] A general process can always be wrietten in the form $P=\{(k,m),(\alpha_1,m_1),\cdots ,(\alpha_n,m_n)\}$ with $m_i \neq m_j$ if $i\neq j$. The coefficients $(x,y,a,b)$ that occour in $P$, by Proposition \[propos:ciao\] depend only on the $ \alpha_i $ and to the order in which they appear but not to the multiplicities $m_i$ and the coefficient $k$. We propose now some simple properties of the index under a process that will be usefull later: \[lemma:l0\] In it follows that $ x a - y b = 1. $ We proceed by induction on the number of blows-up and argue as in the proof of Proposition \[propos:ciao\] With the same arguments we can also prove: \[lemma:l1\] Let consider a process $P=\{(k,m),(1,m_1),\cdots,(1,m_n)\}$ and indicate with $S, D_1, D_2$ the curves that create the triple intersection point. Then if $ (x,y,a,b) $ are the coefficients that appear in the indices we have, according to Remark \[osserv:o1\]: - if $ x > y $ then $ D_2 $ is the new exceptional divisor and $ D_1$ is the old one, - if $ x \leq y $ then $ D_1 $ is the new exceptional divisor and $ D_2 $ is the old one. Using Lemma \[lemma:l1\] and Remark \[osserv:blu\] we can easily prove: \[lemma:somma\] If we blow-up a triple intersection point and we have a multiplicity decrease then the coefficients $ (x',y',a',b')$ of the indices of the new triple are such that: - if $ x > y $ then $ x'=x, y'=x+y ,$ - if $ x \leq y $ then $ x'=x+y, y'=y .$ In the analysis of the $C.S.S.$ index in the triple intersection case the knowledge of the index is equivalent to the knowledge of the coefficients $(x,y,a,b)$. According to Remark \[osserv:blu\] the decrease or not of the multiplicity gives different coefficients. In the next subsections we are going to investigate these cases. To make clearer the possible evolutions of the coefficients we report below the coefficients $(x,y,a,b)$ that can appear in the first five blows-up in triple intersection. We indicate in black the coefficients related to a decrease of multiplicity and in grey the others. Transition from triple intersection with multiplicity lowering to triple with constant multiplicity --------------------------------------------------------------------------------------------------- We consider a process of type $P=\{(k,m),(\alpha_1, m_1),\cdots, (\alpha_{n-1},m_{n-1}),(\alpha_n, m_n)\}$ with $m_i\neq m_j $ if $i\neq j.$ We want to relate the coefficients of the last blow-up with the ones obtained at the first lowering of multiplicity $m_{n-1}\rightarrow m_n$, i.e., we want to relate the last indices of the process $\{(k,m),(\alpha_1, m_1),\cdots, (\alpha_{n-1},m_{n-1}),(1, m_n)\}$ to the last ones of $P$. \[section:TT\] Suppose to have indices of type: $$\label{eq:in} \begin{aligned} &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},S,p)\not \in {\mbox{ $ \mathbb{ Q}$}}_{\geq -k m^2- \alpha_1 m_1^2- \cdots -\alpha_{n-1}m_{n-1}^2-m_n^2}\\ &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},D_1,p)\in {\mbox{ $ \mathbb{ Q}$}}_{\leq{-\frac{x}{y}}}\\ &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},D_2,p)\in {\mbox{ $ \mathbb{ Q}$}}_{\leq-\frac{y k + a}{x k + b}} \end{aligned}$$ with $ m_i\neq m_j $ if $i\neq j $, i.e., $ n $ is the number of multiplicity lowerings. The indices at the end of the process $P$ are of type: - if $ x>y $ $$\label{eq:in1} \begin{aligned} &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},S,p)\not \in {\mbox{ $ \mathbb{ Q}$}}_{\geq -k m^2- \alpha_1 m_1^2- \cdots -\alpha_{n-1}m_{n-1}^2-\alpha_{n}m_n^2}\\ &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},D_1,p)\in {\mbox{ $ \mathbb{ Q}$}}_{\leq{-\frac{x+(\alpha_n-1)y}{y}}}\\ &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},D_2,p)\in {\mbox{ $ \mathbb{ Q}$}}_{\leq-\frac{y k + a}{(x+(\alpha_n-1)y) k +((\alpha_n-1)a+ b)}} \end{aligned}$$ - if $ x\leq y $ $$\label{eq:in2} \begin{aligned} &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},S,p)\not \in {\mbox{ $ \mathbb{ Q}$}}_{\geq -k m^2- \alpha_1 m_1^2- \cdots -\alpha_{n-1}m_{n-1}^2-\alpha_{n}m_n^2}\\ &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},D_1,p)\in {\mbox{ $ \mathbb{ Q}$}}_{\leq{-\frac{((\alpha_n-1)x+y)k + (a + (\alpha_n-1)b)}{x k+y}}}\\ &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},D_2,p)\in {\mbox{ $ \mathbb{ Q}$}}_{\leq-\frac{x}{(\alpha_n-1)x+y}}. \end{aligned}$$ By Proposition \[propos:ordine\] blowing-up with constant multiplicity we know that the new triple point is made up by the curve, the new exceptional divisor and the strict transform of the old one (see Remark \[osserv:o1\]). We have to analyze separately the case in which the old exceptional divisor is $D_1$ or $D_2.$ This distinction can be made in terms of $x>y$ or $x\leq y$ thanks to Lemma \[lemma:l1\]. Suppose, for instance, $ x>y $ in the indices , then we conclude that the old exceptional divisor is $D_1$. Now blowing-up again and using Proposition \[propos:blow-up\], the Index theorem and the assumption of non existence of $(C_1)$ or $(C_2)$ points we can prove the result for $\alpha_n=1,2$. Then proceeeding by induction and repeating the same argument for the case $x\geq y$ we have the assertion. Transition from triple to double intersection {#subsection:td} --------------------------------------------- Suppose that, after $ k $ blows-up in double intersection and a finite number of blows-up in triple intersection, we return to double intersection. Let consider the generic indices of the triple and we write the index along $S$ in the form: $${\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},S,p)\not \in {\mbox{ $ \mathbb{ Q}$}}_{\geq -k m^2-\alpha_1 m_1^2-\cdots -\alpha_n m_n^2},$$ with $m_i\neq m_j$ if $i\neq j$. Using Lemma \[lemma:l0\] we obtain that the indices in the double point we find are: $$\label{eq:doppia} \begin{aligned} &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},S,q)\not \in {\mbox{ $ \mathbb{ Q}$}}_{\geq-k m^2-\alpha_1 m_1^2-\cdots -\alpha_n m_n^2-m_n^2}\\ &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},D,q)\in {\mbox{ $ \mathbb{ Q}$}}_{\leq-\frac{1}{(x+y)^2 k + (x+y)(a+b)}}. \end{aligned}$$ Estimate of the term $ k m^2 $ {#subsection:fine1} ------------------------------ We estimate the term $-k m^2-\alpha_1 m_1^2-\cdots -\alpha_n m_n^2-m_n^2 $ showing that, if the curve is resolved, then $q$ is a point of type $(C_2)$; otherwise we obtain indices of the form and so we can utilize again the results found in the previous sections in order to get desingularization. In this subsection we estimate the term $k m^2 $. If we indicate with $ (x_i^j,y_i^j,a_i^j,b_i^j) $ the coefficients that appear in the indices of the triple intersection point at the $j-$th blow-up with multiplicity $ m_i $ then, if $ n \geq 2 $: $$\begin{aligned} \label{eq:st1} m&=& x_{n-1}^{\alpha_{n-1}} m_{n-1}+ y_{n-1}^{\alpha_{n-1}} m_n \mbox{ if } x_{n-1}^{\alpha_{n-1}}\geq y_{n-1}^{\alpha_{n-1}},\\\label{eq:st2} m&=& y_{n-1}^{\alpha_{n-1}} m_{n-1}+ x_{n-1}^{\alpha_{n-1}} m_n \mbox{ if } y_{n-1}^{\alpha_{n-1}}\geq x_{n-1}^{\alpha_{n-1}}.\end{aligned}$$ We proceed by induction on the number $ n $ of changes of multiplicity. For $n=2$ the indices are of the form: $$\begin{aligned} &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},S,p)\not\in {\mbox{ $ \mathbb{ Q}$}}_{\geq- k m^2-\alpha_1 m_1^2-\alpha_2m_2^2},\\ &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},D_1,p)\in{\mbox{ $ \mathbb{ Q}$}}_{\leq -\frac{\alpha_1\alpha_2+1}{\alpha_1}},\\ &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},D_2,p)\in{\mbox{ $ \mathbb{ Q}$}}_{\leq-\frac{\alpha_1 k + 1}{(\alpha_1\alpha_2+1)k + \alpha_2}}. \end{aligned}$$ The indices we find at the $\alpha_1-$th blow-up with multiplicity $ m_1 $ are: $$\label{eq:n} \begin{aligned} &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},S,p)\not\in {\mbox{ $ \mathbb{ Q}$}}_{\geq - k m^2-\alpha_1 m_1^2},\\ &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},D_2,p)\in{\mbox{ $ \mathbb{ Q}$}}_{\leq -\frac{\alpha_1k+1}{k}},\\ &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},D_1,p)\in{\mbox{ $ \mathbb{ Q}$}}_{\leq-\frac{1}{\alpha_1}}. \end{aligned}$$ Because we make $ \alpha_1 $ blows-up with multiplicity $ m_1 $ and because the curve is irreducible by Enriques-Chisini theorem ([@Bri] pag. 516) we have: $$m_2=m-\alpha_1 m_1$$ and then the assertion. We prove the inductive step. The index along $S$ is: $${\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},S,p)\not\in{\mbox{ $ \mathbb{ Q}$}}_{\geq -k m^2-\alpha_1 m_1^2-\cdots -\alpha_{n-1}m_{n-1}^2-\alpha_n m_n^2-\alpha_{n+1}m_{n+1}^2}$$ We consider the case $ x_{n-1}^{\alpha_{n-1}}\geq y_{n-1}^{\alpha_{n-1}}$ ( the other is similar ). Because we make $ \alpha_n $ blows-up with multiplicity $ m_n $ we have: $$m_{n+1}=m_{n-1}-\alpha_n m_n \mbox{ and then } m_{n-1}= \alpha_n m_n + m_{n+1}.$$ By inductive hypothesis and the above relation we find an expression of $m$ in terms of $m_n$ and $m_{n+1}$. Now we have to prove that this expression is the one of the statement. Using Lemma \[lemma:somma\] we have that $ x_n^1 \leq y_n^1 $ and for Proposition \[section:TT\] the indices at the $ \alpha_n-$th blow-up with multiplicity $ m_n $ are: $$\begin{aligned} &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},S,p)\not\in {\mbox{ $ \mathbb{ Q}$}}_{\geq -k m^2-\alpha_1 m_1^2- \cdots -\alpha_{n-1}m_{n-1}^2-\alpha_{n}m_n^2},\\ &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},D_2,p)\in {\mbox{ $ \mathbb{ Q}$}}_{\leq - \frac{(y_n^1+(\alpha_n-1)x_n^1)k + (a_n^1+(\alpha_n-1)b_n^1)} {x_n^1 k + b_n^1}},\\ &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},D_1,p)\in{\mbox{ $ \mathbb{ Q}$}}_{\leq -\frac{x_n^1}{(\alpha_n-1)x_n^1+y_n^1}}. \end{aligned}$$ Clearly $ x_n^{\alpha_n}\leq y_n^{\alpha_n}$ and so computing the expression $ y_n^{\alpha_n}m_n + x_n^{\alpha_n}m_{n+1},$ using the above form of the coefficients and Lemma \[lemma:somma\] we get the assertion. \[propos:c1\] When in the resolution process we return in double intersection the indices: $$\label{eq:i3} \begin{aligned} &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},S,p)\not \in {\mbox{ $ \mathbb{ Q}$}}_{\geq -k m^2-\alpha_1m_1^2-\cdots -\alpha_nm_n^2-m_n^2},\\ &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},D,q_0)\in {\mbox{ $ \mathbb{ Q}$}}_{\leq-\frac{1}{(x_n^{\alpha_n}+y_n^{\alpha_n})^2 k + (x_n^{\alpha_n}+y_n^{\alpha_n})(a_n^{\alpha_n}+b_n^{\alpha_n})}}, \end{aligned}$$ satisfy $$m \geq (x_n^{\alpha_n}+y_n^{\alpha_n})m_n .$$ It derives directly from the previous proposition and from Lemma \[lemma:somma\]. Estimate of the terms $k m^2+ \alpha_1 m_1^2\cdots + \alpha_n m_n^2+ m_n^2 $ {#subsection:fine2} ---------------------------------------------------------------------------- \[propos:restanti\] The indices at the return in double intersection , with $ n\geq 2$, satisfy: $$\alpha_1 m_1^2+\cdots \alpha_n m_n^2+ m_n^2 \geq (x_n^{\alpha_n}+y_n^{\alpha_n})(a_n^{\alpha_n}+b_n^{\alpha_n})m_n^2 .$$ Before proving this statement we consider the following one: \[propos:processi\] Let $ P=\{(k,m), (\alpha_1, m_1), \cdots, (\alpha_n, m_n)\} $ be a process and let indicate with $ (x,y,a,b) $ the coefficients of the indices that appear at the last blow-up described by $P$. We associate to $P$ the process $ \bar{P}=\{(k, m),(\alpha_2, m_2),\cdots,(\alpha_n, m_n)\} $ and we indicate with $ (\bar{x},\bar{y},\bar{a},\bar{b})$ the coefficients of the indices that appear at the last blow-up described by $\bar{P}$. Then: $$\begin{aligned} b&=\bar{y} \hspace{3cm} x=\alpha_1\bar{y}+\bar{a} \\ a&=\bar{x} \hspace{3cm} y=\alpha_1\bar{x}+\bar{b} \end{aligned}$$ We proceed by induction on the number $n$ of multiplicities decreases. By a direct calculation the Proposition is true for $n=2$. Suppose the assertion true for $ n $ and let prove it for $n+1$.\ Let consider the two processes $ P'=\{(k,m),(\alpha_1,m_1),\cdots,(\alpha_n, m_n),(\alpha_{n+1}, m_{n+1})\} $ and\ $ \bar{P}'=\{(k,m),(\alpha_2, m_2), \cdots, (\alpha_n,m_n),(\alpha_{n+1},m_{n+1})\} $ with respectively end coefficients $ (x',y',a',b') $ and $ (\bar{x}',\bar{y}',\bar{a}',\bar{b}'). $\ Let now construct the following two processes $ P=\{(k,m),(\alpha_1,m_1),\cdots,(\alpha_n, m_n)\},$\ $ \bar{P}=\{(k,m),(\alpha_2,m_2),\cdots,(\alpha_n, m_n)\} $ with end coefficients $ (x,y,a,b) $ and $(\bar{x},\bar{y},\bar{a},\bar{b}) .$ Starting by coefficients $ (x,y,a,b) $ we get $ (x',y',a',b') $ after one blow-up with multiplicity decrease and other $ \alpha_{n+1}-1 $ blows-up with constant multiplicity $m_{n+1}.$ By Propositions \[propos:ordine\] and \[section:TT\]: $$\begin{aligned} (x',y',a',b')&=(x,y+\alpha_n x,a+\alpha_n b,b) \mbox{ if } x>y,\\ (x',y',a',b')&=(x+\alpha_n y,y,a,\alpha_n a+b) \mbox{ if } x \leq y. \end{aligned}$$ Similarly we get: $$\begin{aligned} (\bar{x}',\bar{y}',\bar{a}',\bar{b}')&=(\bar{x},\bar{y}+\alpha_n \bar{x},\bar{a}+\alpha_n \bar{b},\bar{b}) \mbox{ if } \bar{x}>\bar{y},\\ (\bar{x}',\bar{y}',\bar{a}',\bar{b}')&=(\bar{x}+\alpha_n \bar{y},\bar{y},\bar{a},\alpha_n \bar{a}+\bar{b}) \mbox{ if } \bar{x}\leq\bar{y}. \end{aligned}$$ The processes $ P $ e $ P' $ differs only on one multiplicity decrease. Propositions \[propos:ordine\] and \[section:TT\] say that $ x $ and $ y $ relations invert only when a multiplicity decrease occurs. Then we can conclude that $ \bar{x}>\bar{y} $ if and only if $ x \leq y .$ If, for instance, $ x> y ,$ by inductive hypothesis: $$\begin{aligned} b'&=b=\bar{y}=\bar{y}',\\ a'&=a+\alpha_n b= \bar{x}+\alpha_n \bar{y}=\bar{x}',\\ x'&=x=\alpha_1 \bar{y}+\bar{a}=\alpha_1\bar{y}'+\bar{a}',\\ y'&=y+\alpha_n x=\alpha_1\bar{x}+\bar{b}+\alpha_n\bar{y}+\alpha_n\bar{a}=\alpha_1(\bar{x}+\alpha_n \bar{y})+ (\bar{b}+\alpha_n \bar{a})=\alpha_1\bar{x}'+\bar{b}'. \end{aligned}$$ and then the assertion. Now we can prove Proposition \[propos:restanti\]. Let proceed by induction on the number of changes of multiplicity. If $ n=2 $ the structure of the indices can be easily computed to obtain the assertion. Let prove the inductive step. Let $P=\{(k,m),(\alpha_1,m_1),\cdots,(\alpha_n,m_n),(\alpha_{n+1},m_{n+1})\}$ be a generic process. Thanks to the inductive step applied on the process $ \bar{P}=\{(k,m),(\alpha_2,m_2),\cdots,(\alpha_{n+1},m_{n+1})\} $ we have: $$\alpha_1m_1^2+\cdots\alpha_n m_n^2+ \alpha_{n+1} m_{n+1}^2 \geq \alpha_1 m_1^2 + (\bar{x}+\bar{y})(\bar{a}+\bar{b})m_{n+1}^2.$$ In order to estimate $ \alpha_1 m_1^2 $ we consider the process $ P'=\{(\alpha_1,m_1),(\alpha_2,m_2),\cdots,(\alpha_{n+1},m_{n+1})\} $ and thanks to Remark \[osserv:proc\] and Proposition \[propos:c1\] we have: $$m_1^2 \geq (\bar{x}+\bar{y})^2 m_{n+1}^2.$$ Then: $$\begin{aligned} \alpha_1m_1^2+\cdots\alpha_n m_n^2+ \alpha_{n+1} m_{n+1}^2 & \geq \alpha_1 (\bar{x}+\bar{y})^2m_{n+1}^2+ (\bar{x}+\bar{y})(\bar{a}+\bar{b})m_{n+1}^2\\ & =(\bar{x}+\bar{y})(\alpha_1 \bar{x}+\alpha_1 \bar{y} + \bar{a}+\bar{b})m_{n+1}^2. \end{aligned}$$ We conclude thanks to Proposition \[propos:processi\]. The estimate of $km^2$ and of the remaining terms are valid only if $n\geq 2$. The case $n=1$ can be easily proved using equation , Section \[subsection:td\] and observing that because of the $\alpha_1 + 1 $ blows-up $m \geq (\alpha_1 +1)m_1$. Proof of the Theorem -------------------- All the previous separate particular cases can now be glued together to get Theorem \[teor:miomio\]. We have observed that in the resolution process we can have only double or triple intersection points and so we studied the index in these cases. The triple point case presents two different subcases, linked to the multiplicity of the curve: it can decrease or not. This information is extremely useful for the study of the index evolution because it identifies the right exceptional divisor that will occur in the next triple point. Now we observe that if at the end of a process $P=\{(k,m),(\alpha_1, m_1),\cdots, (\alpha_n, m_n)\}$ we find a double point and the curve is desingularized we are in the geometric conditions of a $(C_2)$ point. The indices are the ones given by equation and by Propositions \[propos:c1\] and \[propos:restanti\] we can estimate them in such a way they became: $$\begin{aligned} &{\mbox{ \hbox{Ind}}}({\mbox{ $ \mathcal{F} $ }},S,p)\not \in {\mbox{ $ \mathbb{ Q}$}}_{\geq -[(x+y)k^2+(x+y)(a+b)]}\\ &{\mbox{ \hbox{Ind}}}({\mbox{ $ \mathcal{F} $ }},D,p)\in {\mbox{ $ \mathbb{ Q}$}}_{\leq \frac{1}{(x+y)^2k + (x+y)(a+b)}} \end{aligned}$$ and so $p$ is a $(C_2)$ point. Otherwise we are not in the right geometric conditions, i.e. the resolution is not just ended, but the same propositions gives indices: $$\begin{aligned} &{\mbox{ \hbox{Ind}}}({\mbox{ $ \mathcal{F} $ }},S,p)\not \in {\mbox{ $ \mathbb{ Q}$}}_{\geq -[(x+y)k^2+(x+y)(a+b)]m_n^2}\\ &{\mbox{ \hbox{Ind}}}({\mbox{ $ \mathcal{F} $ }},D,p)\in {\mbox{ $ \mathbb{ Q}$}}_{\leq \frac{1}{(x+y)^2k + (x+y)(a+b)}} \end{aligned}$$ and so we have indices exactly of the form of the ones associated to a process $P=\{h m^2\}$ and then we can apply all the previous argument to the new process which is starting. Such process terminates after a finite number of blows-up by theorem of resolution of singularities [@LAU]. Applications ============ The demonstrative method used to get Theorem \[teor:miomio\] allows us to generalize to the case in which we start with more than one separatrix: \[propos:mio2\] Let be a two dimensional complex manifold, an holomorphic foliation on and $ S_0,S_1,\cdots,S_n $ separatrices of passing through a point $ p\in M $. Let assume that $ S_1,\cdots,S_n $ are non singular and transverse each other and to $ S_0$. If, besides, the indices are of the sequent form: $$\begin{aligned} &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},S_0,p)\not\in {\mbox{ $ \mathbb{ Q}$}}_{\geq -m^2}\\ &{\mbox{ \hbox{Ind}}}({\mbox{$\mathcal{F}$}},S_i,p)\in{\mbox{ $ \mathbb{ Q}$}}_{\leq-(2n-1)} \quad \forall i\geq 1, \end{aligned}$$ then another separatrix through $p$ exists. We prove that $p$ is an appropriate singularity. We observe that after the first blow-up, if we have not finished, we have the same indices found in the study made to prove Theorem \[teor:miomio\] and so we conclude with the same argument. We show briefly that Theorem \[teor:miomio\] includes as particular cases the classical results in discrete and continuous dynamics. \[corol:mio1\] Let be a two dimensional complex manifold, an holomorphic foliation on and $ p\in M $ a singularity of . Then a separatrix of for $p$ exists. We blow-up $M$ in $p$. If the exceptional divisor is not a separatrix for the saturated foliation, Proposition $1$ in [@BRU]$($pag.15$)$ concludes. Otherwise using the index theorem $($see [@SW]$)$ and remembering that $D\cdot D =-1$, we obtain the existence of a singularity $ \tilde{p} $ of the saturated foliation $ \tilde{{\mbox{$\mathcal{F}$}}} $ such that $ {\mbox{ \hbox{Ind}}}(\tilde{{\mbox{$\mathcal{F}$}}},D,\tilde{p})\not \in {\mbox{ $ \mathbb{ Q}$}}^+\cup\{0\}$ and then by Theorem \[teor:miomio\] we have the existence of another separatrix for $ \tilde{p} $ that projects in a separatrix for in . With similar arguments we also have: Let $M $ be a two dimensional complex manifold, an holomorphic foliation on $M$. Let $ S\subset M $ be a compact, globally and locally irreducible curve with $S \cdot S<0$. If $S$ is a separatrix for then a point $ p\in S$ for which passes another separatrix exists. Analogously to what said for Theorem \[teor:diff\] we can obtain, in diffeomorphisms dynamics, a similar result to Proposition \[propos:mio2\] and find as particular cases results of Abate [@AB] and Bracci [@BR]. Let be a two dimensional complex manifold, $f:M \longrightarrow M$ an holomorphic map on $M$ with $ \hbox{Fix}(f)= S_0 \cup, S_1,\cdots,\cup S_n $ with $ S_0,\cdots,S_n $ analytic curves passing through the same point $p \in M$. Let suppose that $S_1,\cdots,S_n $ are non singular and transverse each other and to $ S_0$. If, besides, the indices are of the sequent form: $$\begin{aligned} &{\mbox{ \hbox{Ind}}}(f,S_0,p)\not\in {\mbox{ $ \mathbb{ Q}$}}_{\geq -m^2}\\ &{\mbox{ \hbox{Ind}}}(f,S_i,p)\in{\mbox{ $ \mathbb{ Q}$}}_{\leq-(2n-1)} \quad \forall i\geq 1, \end{aligned}$$ then a parabolic curve through $p$ exists. Let be a two dimensional complex manifold, $f:M \longrightarrow M$ an holomorphic map on and $ p\in M $ an isolated singularity of $f$ such that $df_p=Id$. Then a parabolic curve for $f$ through $p$ exists. Let $M $ be a two dimensional complex manifold, $ f:M \longrightarrow M$ an holomorphic map on $M$. Let $ S\subset M $ be a compact, globally and locally irreducible curve with $S \cdot S<0$. If $\hbox{f}=S$ and $f$ is non degenerate along $S$ is then a point $ p\in S$ for which passes a parabolic curve exists. [99]{} M. Abate, F. Bracci and F. Tovena, Index theorems for holomorphic self maps, to appear in Ann. of Math. M. Abate, The residual index and the dynamics of holomorphic maps tangent to the identity, Duke Math. J., (1) 107, 2001, 173-207. F. Bracci, The dynamics of holomorphic maps near curves of fixed points, Ann. Sc. Norm. Sup. Pisa Cl. Sci. (5) vol. II (2003), 493-520. F. Bracci and F. Tovena, Residual indices of holomorphic maps relative to singular curves of fixed points on surfaces, Math. Z., 242, 2002, 481-490. E.Brieskorn e H.Knorrer, Plane algebraic curves, Birkhauser, 1986. M. Brunella, Birational geometry of foliations, First Latin American Congress of Math., IMPA, Rio de Janeiro, Brazil 2000. C. Camacho and P. Sad Invariant varieties through singularities of holomorphic vector fields, Ann. of Math. (2) 115, 1982, 579-595. J. Cano, Construction of invariant curves for singular holomorfic vector fields, Proc. Amer. Math. Soc., 125(9) 1997,2649-2650. R. C. Gunning and H. Rossi, Analytic functions of several complex variables, Prentice-Hall,Inc., 1965. H. B. Laufer, Normal two dimensional singularities, Ann.Math.Stud.71, 1971. T. Suwa, Indices of holomorphic vector fields relative to invariant curves on surfaces, Proc. Amer. Math. Soc. 123, (1995), 2989-2997. J. F. Mattei and R. Moussu, Holonomie et intégrales premières, Ann. Scient. de l’Ecole Norm. Supérieure, 4 ème série-t.13, (1980), 469-523. M. Sebastiani, Sur l’existence de séparatices locales des feuilletages des surfaces, An. Acad. Bras.Ci., 69, 2, 1997, 159-162.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We study the coupled dynamics of normal and superfluid components of the superfluid $^4$He in the channel considering the counterflow turbulence with laminar normal component. In particular, we calculated profiles of the normal velocity, the mutual friction, the vortex line density and other flow properties and compared them to the case when the dynamic of the normal component is “frozen”. We have found that the coupling between the normal and superfluid components leads to flattening of the normal velocity profile, increasingly more pronounced with temperature, as the mutual friction, and therefore coupling, becomes stronger. The commonly measured flow properties also change when the coupling between two components is taken into account.' author: - 'D. Khomenko' - 'P. Mishra' - 'A. Pomyalov' date: 'Received: date / Accepted: date' title: 'Coupled dynamics for superfluid 4 in the channel. ' --- Introduction {#intro} ============ Superfluid $^4$He below transition temperature $T_\lambda\simeq 2.17\,$K may be viewed as a two-fluid system[@PNAS_intro; @VinenNiemela; @SS-2012] consisting of normal fluid with the very low kinematic viscosity $\nu{_{\text {n}}}(T)$ and inviscid superfluid component, that have their own densities, $\rho{_{\text {n}}}(T)$, $\rho{_{\text {s}}}(T)$, and velocity fields, ${{\bm{u}}}{_{\text {n}}}({{\bm{r}}},t)$, ${{\bm{u}}} {_{\text {s}}}({{\bm{r}}},t)$. Due to quantum mechanical restriction, the circulation around the superfluid vortices is equal to $\kappa= h /m\simeq 10^{-3}\,$cm$^2$/s, where $h$ is the Plank constant and $m$ denotes the mass of $^4$He atom. The singly quantized vortices usually arrange themselves in a tangle, referred to as quantum turbulence, that may be characterized by vortex line density (VLD) $\cal L$, i.e., total length of the quantized vortex line in a unit volume. The large scale motion of such a system may be described by Hall-Vinen-Bekarevich-Khalatnikov equations(HVBK)[@HV; @BK], where both components are considered as continuous fluids, coupled by a mutual friction force. The microscopic description of the vortex lines dynamics at scales that are still much larger than the vortex core size $a_0\approx 10^{-8}$ cm, was proposed by Schwarz[@Schwarz85; @Schwarz88] in the framework of Vortex Filament Method (VFM). In his approach, the influence of the normal component the dynamics on the quantum vortex lines is accounted through the mutual friction force with the predefined time-independent normal fluid velocity. It was soon realized that the back-reaction of the vortex tangle on the dynamics of the normal component may be important and several self-consistent methods, coupling Navier-Stokes equation for the normal component with Lagrangian description of the vortex lines dynamics, were proposed[@BarSam99; @Idowu00; @KBS_science00]. These methods are computationally challenging, given the wide range of scales involved in the problem. Such a self-consistent studies were limited to space-homogeneous flows with diluted vortex tangles. The VFM methods were extended to the wall-bounded flows[@Samuels92; @AaW94] and more recently included full Biot-Savart description of the tangle [@BL13; @BL15; @YT15; @Dima15]. Also here, a time-independent mean normal velocity profile, or a snapshot of the turbulent velocity field were imposed to generate the quantum vortex tangle. Depending on the way the quantum turbulence is generated in the channel or pipe, the normal and superfluid components flow in the same direction, or in opposite directions. In the latter case (the thermal counterflow), a relative, counterflow velocity $V{_{\text {ns}}}$ is established in the channel. At sufficiently large values of $V{_{\text {ns}}}$ the quantum vortex tangle is created. As $V{_{\text {ns}}}$ increases, the thermal counterflow passes several stages. At relatively low values of counterflow velocity, the normal components remains laminar, while a fully-developed quantum vortex tangle is formed, indicating superfluid turbulence. This regime was labeled at T1 state [@Tough_book]. At higher values of $V{_{\text {ns}}}$, the normal component also becomes turbulent, forming so-called T2 state. In this paper we address the back-reaction of the quantum vortex tangle on the normal velocity profile, restricting our study to T1 state of the counterflow in the channel. Here we describe the laminar normal velocity by its mean profile (neglecting normal velocity fluctuations), but allow its evolution, driven by the mutual friction force that couples dynamics of the mean normal velocity with the evolution of the vortex tangle. To be consistent with the nature of the mean profile, the mutual friction force, generated by the vortex tangle, is integrated over space and short intervals of time. Such a multi-scale, multi-time approach allows us to follow the coupled dynamics of two components and account for the influence of the vortex tangle on the normal component. We have found that initially parabolic mean normal velocity profile evolves to a flatter shape, with effect stronger for higher temperature and counterflow velocities. As a consequence, the VLD profile becomes more uniform in the channel core with the peaks pushed towards the wall. Both the global and microscopic properties of the flow change compared to the uncoupled dynamics, although this effect is significant only at high temperatures. [\|c\|c\|c\|c\|c\|c\|]{} T(K) & $\alpha$ & $\alpha'$ & $\rho_n$ (g/cm$^3$) & $\rho_n/\rho_s$ & $\nu=\eta/\rho_n$ (cm$^2$/sec)\ 1.45 & $6.1 \times 10^{-2}$ & $1.8 \times 10^{-2}$ & $1.3 \times 10^{-2}$ & $1.0 \times 10^{-1}$ & $1.0 \times 10^{-3}$\ 1.6 & $9.7 \times 10^{-2}$ & $1.6 \times 10^{-2}$ & $2.4 \times 10^{-2}$ & $1.9 \times 10^{-1}$ & $5.6 \times 10^{-4}$\ 1.9 & $2.1 \times 10^{-1}$ & $8.3 \times 10^{-2}$ & $6.1 \times 10^{-2}$ & $7.2 \times 10^{-1}$ & $2.2 \times 10^{-4}$\ \[TabPar\] Coupled system {#sec:1} ============== In the framework of two-fluid description of the superfluid $^4$He, dynamics of the normal component is given by HVBK equation[@HV; @BK; @BarSam99]: $$\label{eq:NS} \frac{\partial \bm u{_{\text {n}}}}{\partial t}+ (\bm u{_{\text {n}}} \cdot \nabla)\bm u{_{\text {n}}} = \frac{\nabla P}{\rho{_{\text {n}}}} + \frac{\bm F{_{\text {ns}}}}{\rho{_{\text {n}}}} +\nu {_{\text {n}}} \Delta \bm u{_{\text {n}}},$$ where mutual friction force $\bm F{_{\text {ns}}}$ couples the two components and the effective pressure $P$ is defined by: $\nabla P=-\rho{_{\text {n}}}/\rho ~\nabla p +\rho{_{\text {s}}} S \nabla T$ ($S$ is entropy and $\rho=\rho{_{\text {n}}}+\rho{_{\text {s}}}$ is density of Helium II). For the laminar flow in the channel, [Eq.(\[eq:NS\])]{} simplifies to an equation for the mean normal velocity: $$\label{VnEq} \frac{\partial V{_{\text {n}}}(y)}{\partial t}= \frac{d P}{d x} + \frac{{{\mathcal{F}}}{_{\text {ns}}}(y) }{\rho{_{\text {n}}}}+\nu{_{\text {n}}} \frac{\partial^2 V_n(y)}{\partial y^2} ,$$ Here we took into account that in the planar channel geometry the normal velocity field $V{_{\text {n}}}(y)$ depends on the wall-normal direction only and has only one non-zero (streamwise) component along channel. The mutual friction term ${{\mathcal{F}}}{_{\text {ns}}}(y)$ is a time average of a streamwise projection of the mutual friction force, $$\label{eq:Fns} \bm F{_{\text {ns}}}= \frac{\rho_s }{\Omega} \int_{{{\mathcal{C}}}'} (\alpha \bm s' \times [\bm s' \times \bm V{_{\text {ns}}}]+\alpha' \bm s' \times \bm V{_{\text {ns}}} )d \xi,$$ defined by dynamics of the quantum vortex tangle[@Schwarz88]. Here $\bm s$ is a radius vector to the points on the vortex line, $'$ denotes derivative along the vortex line and $\alpha, \alpha'$ are the mutual friction coefficients. The integral in [Eq.(\[eq:Fns\])]{} is taken along vortex lines ${{\mathcal{C}}}'$, residing inside suitably defined coarse-grained volume $\Omega$. The instantaneous counterflow velocity $\bm V{_{\text {ns}}}=\bm {V}{_{\text {n}}}-\bm {V}{_{\text {s}}}$ is defined by velocities of the normal and superfluid components. The superfluid velocity $\bm V{_{\text {s}}} = {V{_{\text {s}}} ^0}+ \bm {V}{_{\text {BS}}}(\bm s)$ contains contributions of the vortex tangle velocity $\bm {V}{_{\text {BS}}}(\bm s)$ in the Biot-Savart representation, integrated over entire vortex configuration ${{\mathcal{C}}}$ and the applied superfluid velocity ${V{_{\text {s}}} ^0}$, defined by the counterflow condition $\rho{_{\text {n}}} \langle \bm {V}{_{\text {n}}} \rangle_v +\rho{_{\text {s}}} \langle \bm {V}{_{\text {s}}}\rangle_v =0$. Here $\langle...\rangle_v$ stands for global volume averaging. ---------------------------------------- ---------------------------------------- -- ![image](Fig1a.pdf){width="42.00000%"} ![image](Fig1b.pdf){width="42.00000%"} ---------------------------------------- ---------------------------------------- -- ---------------------------------------- ---------------------------------------- ![image](Fig2a.pdf){width="42.00000%"} ![image](Fig2b.pdf){width="42.00000%"} ---------------------------------------- ---------------------------------------- #### Implementation Details The simulations were set up in a planar channel of the size $4h\times 2h\times2h, h=0.05$ cm. The vortex-lines dynamics was solved using VFM[@Schwarz88; @b:Kond-1; @Dima15] with 4th-order difference scheme for the derivatives ${{\bm{s}}}'$ and ${{\bm{s}}}''$ [@BB2011; @5point]. We used the periodic boundary condition in the streamwise($x$) and the spanwise($z$) directions, with slip conditions in the wall-normal $y$ direction, the line-resolution $\Delta \xi_0=0.0016$ cm and the timestep $\delta t{_{\text {s}}}$, defined by the stability condition of the 4th-order Runge-Kutta scheme. The dynamics of the mean normal velocity profile $V{_{\text {n}}} (y,t)$, [Eq.(\[VnEq\])]{}, was solved using 2nd-order finite-difference scheme for the viscous term, 2nd-order Adams-Bashforth method for the time marching and no-slip conditions on the solid walls. The effective pressure gradient $P{_{\text {x}}}=dP/dx$ was used as a free parameter of the system. To calculate the mutual friction term, we used two-stage approach. The term ${{\mathcal{F}}}{_{\text {ns}}}$ in [Eq.(\[VnEq\])]{} implies averaging over space and time, while the term $ F{_{\text {ns}}}$ is instantaneous and is averaged only over space. The parabolic normal velocity profile was initiated on a grid of a mesh size $\delta y=0.00625$ cm in the wall-normal direction. Using this profile, the superfluid system was propagated for $\Delta t=10 \delta t{_{\text {s}}}$. At each time step, the mutual friction force $ F{_{\text {ns}}}$ was integrated over thin slices of volume $4h\times\delta y\times 2h$ and then averaged during $\Delta t$. Resulting values of ${{\mathcal{F}}}{_{\text {ns}}}$ were assigned to the middle points of the slices in $y$-direction and linearly interpolated to the normal velocity grid points. With this mean mutual friction force, $V{_{\text {n}}}(y)$ was then propagated to $\Delta t$ using the same $\delta t{_{\text {s}}}$. The newly obtained values of $V{_{\text {n}}}(y)$ were interpolated to the positions of the line points for the next cycle of the superfluid dynamics, during which $V{_{\text {n}}}$ was considered constant. The resulting profiles of ${{\mathcal{L}}}(y)$, $V{_{\text {s,n}}}(y)$ and other related properties were obtained by integrating over the same slices as the mutual friction, assigned to the middle points of the slices and then averaged over more than 50 steady-state configurations. The simulations were carried out for $T=1.45, 1.6$ and $1.9$ K and a number of $P{_{\text {x}}}$ values. The material properties of the fluid components [@ExpData] are given in the Table \[TabPar\]. Notice that we didn’t use the modified mutual friction coefficients[@Idowu00], because we only consider the mean normal velocity field. Results and Discussion {#sec:res} ====================== The steady-state profiles of $V{_{\text {n}}}(y)$ for different $T$ and values of effective pressure are shown in Fig. \[fig:Vnprofiles\]. To compare the shapes of profiles we normalized them by $3/2~\langle V{_{\text {n}}}\rangle$. For the original parabolic profile this value corresponds to the centerline velocity. For modified profiles this relation no longer valid. The profiles of $V{_{\text {n}}}(y)$ become flatter in the center of the channel, increasingly so with increasing temperature and the applied pressure gradient. The resulting profile of counterflow velocity $V{_{\text {ns}}}(y)$ is shown by solid lines in Fig. \[fig:Lprofiles\], left panel, for $T=1.6$K and $1.9$K together with profiles, obtained using time-independent parabolic profile with similar $\langle V{_{\text {n}}}\rangle$(dashed lines). The flatter profiles of $V{_{\text {n}}}$ lead to a change in the profiles of $V{_{\text {ns}}}$, especially in the channel core. The vortex lines tend to concentrate in the regions with smaller $V{_{\text {n}}}$[@Dima15], such that the peaks in the VLD profiles are pushed further toward the walls. The statistical properties of the vortex tangle are commonly characterized by relation between mean VLD in the channel and $\langle V{_{\text {ns}}}\rangle$: $\langle \mathcal{ L} \rangle^{1/2}=\gamma(T) (\langle V_{ns}\rangle-v_0)$, where $\gamma(T)$ is a temperature-dependent coefficient and $v_0$ is virtual origin. This relation is valid only globally[@MechMom16]. We compare in Table \[tab:1\] the values of $\gamma$, obtained with different profiles. Notice that results for different profiles at temperatures $T=1.45$K and $1.6$K are close, while for higher $T=1.9$K, $\gamma{_{\text {c}}}$ is significantly larger than $\gamma{_{\text {p}}}$. The values of $\gamma{_{\text {p}}}$ are close to those of uniform $V{_{\text {n}}}$ for all temperatures. Notably, our results are higher than those, obtained in Refs.[@BL15; @YT15] for the channel counterflow with parabolic and Hagen-Poiseuille profiles. This discrepancy may stem from the fact that in both [@BL15; @YT15] the counterflow condition did not include contribution from the vortex tangle, which is not negligible for dense tangles. #### Model profile of $V{_{\text {n}}}$ To rationalize the observed modifications of the normal velocity profile, we notice that in steady state the mutual friction force is almost constant across the channel, except for the near-wall region, where it quickly falls to zero. Therefore, qualitatively, the mutual friction redefines the effective pressure gradient in the middle of the channel $P_x+\langle {{\mathcal{F}}}{_{\text {ns}}} \rangle/\rho{_{\text {n}}}$, while near the wall it remains $P_x$. Such a change leads to flattening of the normal velocity profile (as compared to classic parabolic profile). To find new, flattened, $ V{_{\text {n}}}$ we first model the shape of mutual friction force profile by a function: $$\label{Fmodel} {{{\mathcal{F}}}}{_{\text {ns}}}=-(1-{y^\dagger}^n) {{\mathcal{F}}}{_{\text {ns}}}(0) \, ,$$ where $n$ is an even integer and ${{\mathcal{F}}}{_{\text {ns}}}(0)$ is the mutual friction force in the middle of the channel. To account for the sharp transition from almost a constant value in the middle of the channel to zero at the wall, large values of $n$ are required. In this work we use $n=14$. Substituting into steady state equation for $V_n$ we get: $$\label{Vn2} \frac{\nu}{h^2} \frac{d^2 V{_{\text {n}}}}{d{y^\dagger}^2}=-P{_{\text {x}}}+\frac{ {{\mathcal{F}}}{_{\text {ns}}}(0)}{\rho_n}(1-{y^\dagger}^n)\, .$$ This equation is easily solved, giving: $$\begin{aligned} \label{Vn3} &&V{_{\text {n}}}(y^\dagger)=A\left(1-{y^\dagger}^2\right)\left[1+\frac{2C}{(n+2)(n+1)}\frac{1-{y^\dagger}^{n+2}}{1-{y^\dagger}^{2}}\right]\, ,\\\nonumber &&A=h^2P{_{\text {x,m}}}/(2\nu) \, , C= {{\mathcal{F}}}{_{\text {ns}}}(0) /(\rho_n P{_{\text {x,m}}})\, ,P{_{\text {x,m}}}=P{_{\text {x}}}- {{\mathcal{F}}}{_{\text {ns}}} (0)/\rho_n\, .\end{aligned}$$ Notice, that when mutual force is not taken into account $ {{\mathcal{F}}}{_{\text {ns}}} (0)=0$, the profile reduces to the classical parabolic profile. To proceed, we notice that $\langle V_n \rangle =2A/3 +2AC/[(n+1)(n+3)]$ and $\langle {{\mathcal{F}}}{_{\text {ns}}} \rangle=n{{\mathcal{F}}}{_{\text {ns}}} (0)/(n+1)$. Using counterflow condition and Gorter-Mellink relation between mean mutual friction force and $\langle V{_{\text {ns}}} \rangle$, we find $$\label{GM} \langle {{\mathcal{F}}}{_{\text {ns}}} \rangle =A{_{\text {GM}}} \rho{_{\text {s}}} \rho{_{\text {n}}} (1+\rho{_{\text {n}}}/\rho{_{\text {s}}})^3 \langle V{_{\text {n}}} \rangle^3\equiv\frac{\rho_n \nu}{h^2}\frac{\langle V_n \rangle^3}{V_0^2}\, .$$ The characteristic velocity $\displaystyle V_0=\left(A{_{\text {GM}}} \rho_s h^2(1+\rho{_{\text {n}}}/\rho{_{\text {s}}})^3 /\nu \right)^{-1/2}$ corresponds to a balance between mutual friction and viscous terms and $A{_{\text {GM}}}$ is the Gorter-Mellink constant. Combining all above we finally get: $$\begin{aligned} \label{Vn4} V{_{\text {n}}}(y^\dagger)&=&\frac{3}{2}\langle V{_{\text {n}}} \rangle \left[V_1 (1-{y^\dagger}^2)+V_2\left(1-{y^\dagger}^{n+2}\right) \right]\, ,$$ where $V_1=\Big(1-\frac{1}{n(n+3)}\Big[\frac{\langle V{_{\text {n}}}\rangle}{V_0}\Big]^2\Big)$ and $V_2=\frac{2}{3(n+2)n}\Big[\frac{\langle V{_{\text {n}}}\rangle}{V_0}\Big]^2$ are $T$-dependent expansion coefficients. The model profiles are compared with numerical results in Fig\[fig:Vnprofiles\], left panel. [\|c\|c\|c\|c\|c\|c\|c\|c\|]{} T(K) & $\gamma{_{\text {c}}}$ (s/cm$^2$)& $\gamma{_{\text {p}}}$(s/cm$^2$) &$\gamma{_{\text {p}}}$(s/cm$^2$) [@BL15]&$\gamma{_{\text {hp}}}$(s/cm$^2$)[@YT15]&$\gamma{_{\text {uni}}}$(s/cm$^2$)[@b:Kond-1]\ 1.3 & - & - &$67.9$&$31$&$72.1$\ 1.45 & 83 & 85 &$-$&$-$&$-$\ 1.6 & 114 & 113&$83.6$&$47$&$115.7$\ 1.9 & 165 & 144 &$105.7$&$103$&$148$\ Conclusions =========== We have studied back-reaction of the quantum vortex tangle on the mean laminar normal velocity profile in the channel counterflow. The initially parabolic $ V_n $ evolves to a flatter shape. The vortex tangle influences the mean normal velocity profile via mutual friction force. As a result, both the global properties of flow and the profiles of microscopic properties of the tangle, such as rms curvature, change compared to the uncoupled dynamics, with effect being stronger for high temperatures and counterflow velocities. At low and moderate temperatures the flow and tangle properties are close to those with the time-independent normal velocity profile. We propose a model of $ V{_{\text {n}}}(y)$, expressed via $\langle V{_{\text {n}}}\rangle$, that accounts for the effect of mutual friction. We acknowledge useful and encouraging discussions with V.L’vov and I. Procaccia. C. F. Barenghi, L. Skrbek and K. R. Sreenivasan, Proc. Natl. Acad. Sci. USA 111, 4647 (2014). W. F. Vinen and J. J. Niemela, J. Low Temp. Phys. [128]{}, 167 (2002). L. Skrbek and K.R. Sreenivasan, Phys. Fluids 24, 011301 (2012). R. J. Donnelly, Quantized Vortices in Hellium II, (Cambridge Univ. Press, 1991). Quantized Vortex Dynamics and Superfluid Turbulence, edited by C. F. Barenghi, R. J. Donnelly and W .F. Vinen, Lecture Notes in Physics 571, (Springer-Verlag, Berlin, 2001). H.E. Hall, W.F. Vinen, Proc. R. Soc. Lond. A Math Phys. Sci. 238, 215 (1956). I.L. Bekarevich, I.M. Khalatnikov, Sov. Phys. JETP 13, 643 (1961). C. Barenghi and David C. Samuels, Phys. Rev. B [60]{}, 1252), 1999.O.C. Idowu, D. Kivotides, C.F. Barenghi and D. C. Samuels, J. Low Temp. Phys., 120 (2000) 269.Kivotides D., Barenghi C. F. and Samuels D. C., Science, [290]{} (2000) 777. D. Kivotides, C. J. Vassilicos, D. C. Samuels and C. F. Barenghi, Europhys. Lett., [57]{}, 845 (2002) D. R. Poole, H. Scoffield, C. F. Barenghi, and D. C. Samuels, J. Low Temp. Phys.,[132]{}, 97 (2003) K.W. Schwarz, Phys. Rev. B 31, 5782 (1985). K.W. Schwarz, Phys. Rev. B 38, 2398 (1988). D.C. Samuels, Phys. Rev. B [46]{}, 11714 (1992). R. G. K. M. Aarts and A. T. A. M. de Waele, Phys. Rev. B 50, 10069 (1994). A. W. Baggaley and Laizet, Phys. Fluids 25, 115101 (2013). A. W. Baggaley, J. Laurie, J. Low Temp. Phys. 178, 35(2015) S. Yui and M.Tsubota,Phys. Rev. B 91, 184504 (2015). D. Khomenko, L. Kondaurova, V.S. L’vov, P. Mishra, A. Pomyalov and I.Procaccia, Phys. Rev. B [91]{}, 180504(R), 2015. J. T. Tough, in Progress in Low Temperature Physics, edited by D. F. Brewer (North-Holland, Amsterdam, 1982), Vol. VIII, Chap. 3. L. Galantucci, M. Sciacca, and C. F. Barenghi, Phys. Rev. B 92, 174530 (2015) A.W. Baggaley, C. Barenghi,Phys. Rev. B 83,134509(2011) L. Gamet, F. Ducros, F. Nicoud, and T. Poinsot, International Journal for Numerical Methods in Fluids 29, 2 (1999). L. Kondaurova, V.S. L’vov, A. Pomyalov and I. Procaccia, Phys. Rev. B 89, 014502 (2014). R. J. Donnelly and C. F. Barenghi, J. Phys. Chem. Ref. Data [27]{}, 6, (1998). D. Khomenko, V. S. L’vov, A. Pomyalov, and I. Procaccia, Phys. Rev. B, [93]{},134504(2016)	{ "pile_set_name": "ArXiv" }
--- abstract: 'In the timescape scenario cosmic acceleration is understand as an apparent effect, due to gravitational energy gradients that grow when spatial curvature gradients become significant with the nonlinear growth of cosmic structure. This affects the calibratation of local geometry to the solutions of the volume–average evolution equations corrected by backreaction. In this paper I discuss recent work on defining observational tests for average geometric quantities which can distinguish the timescape model from a cosmological constant or other models of dark energy.' author: - 'David L. Wiltshire' title: 'Gravitational energy as dark energy: Average observational quantities$^$' --- =cmr5=cmr7=cmmi7 \#1[\_]{} \#1[Phys. Rev. Lett. [\#1]{}]{} \#1[Phys. Rev. [\#1]{}]{} \#1[Astrophys. J. [\#1]{}]{} \#1[Astron. J. [\#1]{}]{} \#1[Mon. Not. R. Astr. Soc. [\#1]{}]{} \#1[Class. Quantum Grav. [\#1]{}]{} \#1[Gen. Relativ. Grav. [\#1]{}]{} \#1[\_]{} \#1 \#1[\_]{} =cmr7 \#1[\_]{} \#1 \#1 \#1\#2 \#1\#2\#3[[\#1\#3\#1\#2]{}]{} \#1\#2[[\#1\#1\#2]{}]{} \#1[[\|\#1]{}]{} [ address=[Department of Physics & Astronomy, University of Canterbury, Private Bag 4800, Christchurch 8140,\ New Zealand; and]{} ,altaddress=[International Center for Relativistic Astrophysics Network (ICRANet), P.le della Repubblica 10, Pescara 65121, Italy]{} ]{} Introduction ============ I will discuss some recent results on observational tests [@obs] of a model cosmology, which represents a new approach to understanding the phenomenology of dark energy as a consequence of the effect of the growth of inhomogeneous structures. The basic idea, outlined in a nontechnical manner in ref. [@dark07], is that as inhomogeneities grow one must consider not only their backreaction on average cosmic evolution – as discussed by other contributors to this volume – but also the variance in the geometry as it affects the calibration of clocks and rods of ideal observers. Dark energy is then effectively realised as a misidentification of gravitational energy gradients. Although the standard Lambda Cold Dark Matter () model provides a good fit to many tests, there are tensions between some tests, and also a number of puzzles and anomalies. Furthermore, at the present epoch the observed universe is only statistically homogeneous once one samples on scales of 150–300 Mpc. Below such scales it displays a web–like structure, dominated in volume by voids. Some 40%–50% of the volume of the present epoch universe is in voids with $\de\rho/\rho\goesas-1$ on scales of 30$h^{-1}$ Mpc [@HV], where $h$ is the dimensionless parameter related to the Hubble constant by $H\Z0=100h\kmsMpc$. Once one also accounts for numerous minivoids, and perhaps also a few larger voids, then it appears that the present epoch universe is void-dominated. Clusters of galaxies are spread in sheets that surround these voids, and in thin filaments that thread them. One particular consequence of a matter distribution that is only statistically homogeneous, rather than exactly homogeneous, is that when the Einstein equations are averaged they do not evolve as a smooth Friedmann–Lemaître–Robertson–Walker (FLRW) geometry. Instead the Friedmann equations are supplemented by additional backreaction terms [@buch00]. Whether or not one can fully explain the expansion history of the universe as a consequence of the growth of inhomogeneities and backreaction, without a fluid–like dark energy, is the subject of ongoing debate [@buch08]. Elsewhere in this volume, Peebles [@Peebles] provides some of the arguments that have been presented against backreaction. His line of reasoning is that of a plausibility argument: if we [assume]{} a FLRW geometry with small perturbations, and estimate the magnitude of the perturbations from the typical rotational and peculiar velocities of galaxies, then the corrections of inhomogeneities are consistently small. This would be a powerful argument, were it not for the fact that at the present epoch galaxies are not homogeneously distributed. The Hubble Deep Field reveals that galaxies were close to being homogeneous distributed at early epochs, but following the growth voids at redshifts $z\lsim1$ that is no longer the case today. Therefore galaxies cannot be consistently treated as randomly distributed gas particles on the 30$h^{-1}$ Mpc scales [@HV] that dominate present cosmic structure below the scale of statistical homogeneity. Over the past few years I have developed a new physical interpretation of cosmological solutions within the Buchert averaging scheme [@clocks; @sol; @equiv]. I start by noting that in the presence of strong spatial curvature gradients, not only should the average evolution equations be replaced by equations with terms involving backreaction, but the physical interpretation of average quantities must also account for the differences between the local geometry and the average geometry. In other words, geometric variance can be just as important as geometric averaging when it comes to the physical interpretation of the expansion history of the universe. I proceed from the fact that structure formation provides a natural division of scales in the observed universe. As observers in galaxies, we and the objects we observe in other galaxies are necessarily in bound structures, which formed from density perturbations that were greater than critical density. If we consider the evidence of the large scale structure surveys on the other hand, then the average location by volume in the present epoch universe is in a void, which is negatively curved. We can expect systematic differences in spatial curvature between the average mass environment, in bound structures, and the volume-average environment, in voids. Spatial curvature gradients will in general give rise to gravitational energy gradients. Physically this can be understood in terms of a relative deceleration of expanding regions of different densities. Those in the denser region decelerate more and age less. Since we are dealing with weak fields the relative deceleration of the background is small. Nonetheless even if the relative deceleration is typically of order $10^{-10}$ms$^{-2}$, cumulatively over the age of the universe it leads to significant clock rate variances [@equiv]. I proceed from an ansatz that the variance in gravitational energy is correlated with the average spatial curvature in such a way as to implicitly solve the Sandage–de Vaucouleurs paradox that a statistically quiet, broadly isotropic, Hubble flow is observed deep below the scale of statistical homogeneity. In particular, galaxy peculiar velocities have a small magnitude with respect to a local regional volume expansion. Expanding regions of different densities are patched together so that the regionally measured expansion remains uniform. Such regional expansion refers to the variation of the regional proper length, $\ell_r=\Vav^{1/3}$, with respect to proper time of isotropic observers (those who see an isotropic mean CMB). Although voids open up faster, so that their proper volume increases more quickly, on account of gravitational energy gradients the local clocks will also tick faster in a compensating manner. Details of the fitting of local observables to average quantities for solutions to the Buchert formalism are described in detail in refs.[@clocks; @sol]. Negatively curved voids, and spatially flat expanding wall regions within which galaxy clusters are located, are combined in a Buchert average (t)+(t)=1,where $\fw(t)=\fwi\aw^3/\ab^3$ is the [wall volume fraction]{} and $\fv(t)=\fvi\av^3/\ab^3$ is the [void volume fraction]{}, $\Vav=\Vav\ns i\ab^3$ being the present horizon volume, and $\fwi$, $\fvi$ and $\Vav\ns i$ initial values at last scattering. The time parameter, $t$, is the volume–average time parameter of the Buchert formalism, but does not coincide with that of local measurements in galaxies. In trying to fit a FLRW solution to the universe we attempt to match our local spatially flat wall geometry \^2=-\^2+\^2(). \[wgeom\]to the whole universe, when in reality the rods and clocks of ideal isotropic observers vary with gradients in spatial curvature and gravitational energy. By conformally matching radial null geodesics with those of the Buchert average solutions, the geometry (\[wgeom\]) may be extended to cosmological scales as the dressed geometry \^2=-\^2+\^2()\[dgeom\]where $a=\gb^{-1}\ab$, $\gb=\Deriv\dd\tw\ts$ is the relative lapse function between wall clocks and volume–average ones, $\dd\etb=\dd t/\ab= \dd\tw/ \ac$, and $\rw=\gb\fvf^{1/3}\fwi^{-1/3}\etw(\etb,\tw)$, where $\etw$ is given by integrating $\dd\etw=\fwi^{1/3}\dd\etb/[\gb\fvf^{1/3}]$ along null geodesics. In addition to the bare cosmological parameters which describe the Buchert equations, one obtains dressed parameters relative to the geometry (\[dgeom\]). For example, the dressed matter density parameter is $\Omega\Z M=\gb^3\OMM$, where $\OMM=8\pi G\rhb\Z{M0}\ab\Z0^3/(3\bH^2\ab^3)$ is the bare matter density parameter. The dressed parameters take numerical values close to the ones inferred in standard FLRW models. Apparent acceleration and Hubble flow variance ============================================== The gradient in gravitational energy and cumulative differences of clock rates between wall observers and volume average ones has important physical consequences. Using the exact solution obtained in ref. [@sol], one finds that a volume average observer would infer an effective deceleration parameter $\bq=-\ddot\ab/(\bH^2\ab)= 2\fvf^2/(2+\fv)^2$, which is always positive since there is no global acceleration. However, a wall observer infers a dressed deceleration parameter q=-[1H\^2 a]{}[\^2 a\^2]{}= [-(8\^3+39\^2-12-8)(4++4\^2)\^2]{}, \[qtrack\]where the dressed Hubble parameter is given by =\^[-1]{}=-=-\^[-1]{}. \[42\]At early times when $\fv\to0$ the dressed and bare deceleration parameter both take the Einstein–de Sitter value $q\simeq\bq\simeq\half$. However, unlike the bare parameter which monotonically decreases to zero, the dressed parameter becomes negative when $\fv\simeq0.59$ and $\bq\to0^-$ at late times. For the best-fit parameters [@LNW] the apparent acceleration begins at a redshift $z\simeq0.9$. Cosmic acceleration is thus revealed as an apparent effect which arises due to the cumulative clock rate variance of wall observers relative to volume–average observers. It becomes significant only when the voids begin to dominate the universe by volume. Since the epoch of onset of apparent acceleration is directly related to the void fraction, $\fv$, this solves one cosmic coincidence problem. In addition to apparent cosmic acceleration, a second important apparent effect will arise if one considers scales below that of statistical homogeneity. By any one set of clocks it will appear that voids expand faster than wall regions. Thus a wall observer will see galaxies on the far side of a dominant void of diameter $30h^{-1}$ Mpc recede at a rate greater than the dressed global average $\Hm$, while galaxies within an ideal wall will recede at a rate less than $\Hm$. Since the uniform bare rate $\bH$ would also be the local value within an ideal wall, eq.(\[42\]) gives a measure of the variance in the apparent Hubble flow. The best-fit parameters [@LNW] give a dressed Hubble constant $\Hm=61.7^{+1.2}_{-1.1}\kmsMpc$, and a bare Hubble constant $\Hb=48.2^{+2.0}_{-2.4}\kmsMpc$. The present epoch variance is 17–22%. Since voids dominate the universe by volume at the present epoch, any observer in a galaxy in a typical wall region will measure locally higher values of the Hubble constant, with peak values of order $72\kmsMpc$ at the $30h^{-1}$ Mpc scale of the dominant voids. Over larger distances, as the line of sight intersects more walls as well as voids, a radial spherically symmetric average will give an average Hubble constant whose value decreases from the maximum at the $30h^{-1}$ Mpc scale to the dressed global average value, as the scale of homogeneity is approached at roughly the baryon acoustic oscillation (BAO) scale of $110h^{-1}$Mpc. This predicted effect could account for the Hubble bubble [@JRK] and more detailed studies of the scale dependence of the local Hubble flow [@LS]. In fact, the variance of the local Hubble flow below the scale of homogeneity should correlate strongly to observed structures in a manner which has no equivalent prediction in FLRW models. Future observational tests ========================== There are two types of potential cosmological tests that can be developed; those relating to scales below that of statistical homogeneity as discussed above, and those that relate to averages on our past light cone on scales much greater than the scale of statistical homogeneity. The second class of tests includes equivalents to all the standard cosmological tests of the standard FLRW model with Newtonian perturbations. This second class of tests can be further divided into tests which just deal with the bulk cosmological averages (luminosity and angular diameter distances etc), and those that deal with the variance from the growth of structures (late epoch integrated Sachs–Wolfe effect, cosmic shear, redshift space distortions etc). Here I will concentrate solely on the simplest tests which are directly related to luminosity and angular diameter distance measures. In the cosmology we have an effective dressed luminosity distance =a0(1+z),where $a\Z0=\gc^{-1}\ab\Z0$, and =\^[1/3]{} \_\^[0]{}[()(1-())\^[1/3]{}()]{}. \[eq:dL\]We can also define an [effective angular diameter distance]{}, $\dA$, and an [effective comoving distance]{}, $D$, to a redshift $z$ in the standard fashion =[D1+z]{}=[(1+z)\^2]{}.\[dist\] A direct method of comparing the distance measures with those of homogeneous models with dark energy, is to observe that for a standard spatially flat cosmology with dark energy obeying an equation of state $P\Z D=w(z) \rh\Z D$, the quantity D=\_0\^z[z’]{}, \[rFLRW\]does not depend on the value of the Hubble constant, $\Hm$, but only directly on $\OmMn=1-\Omega\Z{D0}$. Since the best-fit values of $\Hm$ are potentially different for the different scenarios, a comparison of $\Hm D$ curves as a function of redshift for the model versus the model gives a good indication of where the largest differences can be expected, independently of the value of $\Hm$. Such a comparison is made in Fig. \[fig\_coD\]. [(a)]{}-15pt ![The effective comoving distance $\Hx D(z)$ is plotted for the best–fit (TS) model, with $\fvn=0.762$, (solid line); and for various spatially flat models (dashed lines). The parameters for the dashed lines are (i) $\OmMn=0.249$ (best–fit to WMAP5 only [@wmap5]); (ii) $\OmMn=0.279$ (joint best–fit to SneIa, BAO and WMAP5); (iii) $\OmMn=0.34$ (best–fit to Riess07 SneIa only [@Riess07]). Panel [(a)]{} shows the redshift range $z<6$, with an inset for $z<1.5$, which is the range tested by current SneIa data. Panel [(b)]{} shows the range $z<1100$ up to the surface of last scattering, tested by WMAP5.[]{data-label="fig_coD"}](rw6in15a.eps "fig:"){width="2.25in"} -15pt ![The effective comoving distance $\Hx D(z)$ is plotted for the best–fit (TS) model, with $\fvn=0.762$, (solid line); and for various spatially flat models (dashed lines). The parameters for the dashed lines are (i) $\OmMn=0.249$ (best–fit to WMAP5 only [@wmap5]); (ii) $\OmMn=0.279$ (joint best–fit to SneIa, BAO and WMAP5); (iii) $\OmMn=0.34$ (best–fit to Riess07 SneIa only [@Riess07]). Panel [(a)]{} shows the redshift range $z<6$, with an inset for $z<1.5$, which is the range tested by current SneIa data. Panel [(b)]{} shows the range $z<1100$ up to the surface of last scattering, tested by WMAP5.[]{data-label="fig_coD"}](rw1100a.eps "fig:"){width="2.in"} We see that as redshift increases the model interpolates between models with different values of $\OmMn$. For redshifts $z\lsim1.5$ $\Dfb$ is very close to $\Dlcdm$ for the parameter values $(\OmMn,\OmLn) =(0.34,0.66)$ (model (iii)) which best–fit the Riess07 supernovae (SneIa) data [@Riess07] only, by our own analysis. For very large redshifts that approach the surface of last scattering, $z\lsim1100$, on the other hand, $\Dfb$ very closely matches $\Dlcdm$ for the parameter values $(\OmMn,\OmLn) =(0.249,0.751)$ (model (i)) which best–fit WMAP5 only [@wmap5]. Over redshifts $2\lsim z\lsim10$, at which scales independent tests are conceivable, $\Dfb$ makes a transition over corresponding curves of $\Dlcdm$ with intermediate values of $(\OmMn,\OmLn)$. The $\Dlcdm$ curve for joint best-fit parameters to SneIa, BAO measurements and WMAP5 [@wmap5], $(\OmMn,\OmLn) =(0.279,0.721)$ is best–matched over the range $5\lsim z\lsim 6$, for example. The difference of $\Dfb$ from any single $\Dlcdm$ curve is perhaps most pronounced in the range $2\lsim z\lsim 6$, which may be an optimal regime to probe in future experiments. Gamma–ray bursters (GRBs) now probe distances to redshifts $z\lsim8.3$, and could be very useful. A considerable amount work of work has already been done on Hubble diagrams for GRBs. (See, e.g., [@GRB].) Much more work is needed to nail down systematic uncertainties, but GRBs may eventually provide a definitive test in future. An analysis of the model Hubble diagram using 69 GRBs has just been performed by Schaefer [@Schaefer], who finds that it fits the data better than the concordance model, but not yet by a huge margin. As more data is accumulated, it should become possible to distinguish the models if the issues with the standardization of GRBs can be ironed out. The effective “equation of state” --------------------------------- It should be noted that the shape of the $\Hm D$ curves depicted in Fig. \[fig\_coD\] represent the observable quantity one is actually measuring when some researchers loosely talk about “measuring the equation of state”. For spatially flat dark energy models, with $\Hm D$ given by (\[rFLRW\]), one finds that the function $w(z)$ appearing in the fluid equation of state $P\Z D=w(z)\rh\Z D$ is related to the first and second derivatives of (\[rFLRW\]) by w(z)=[23(1+z)D’\^[-1]{}D”+1(1+z)\^3\^2 D’\^2-1]{} \[eos\]where prime denotes a derivative with respect to $z$. Such a relation can be applied to observed distance measurements, regardless of whether the underlying cosmology has dark energy or not. Since it involves first and second derivatives of the observed quantities, it is actually much more difficult to determine observationally than directly fitting $\Hm D(z)$. The equivalent of the “equation of state”, $w(z)$, for the model is plotted in Fig. \[fig\_wz\]. The fact that $w(z)$ is undefined at a particular redshift and changes sign through $\pm\infty$ simply reflects the fact that in (\[eos\]) we are dividing by a quantity which goes to zero for the model, even though the underlying curve of Fig. \[fig\_coD\] is smooth. Since one is not dealing with a dark energy fluid in the present case, $w(z)$ simply has no physical meaning. Nonetheless, phenomenologically the results do agree with the usual inferences about $w(z)$ for fits of standard dark energy cosmologies to SneIa data. For the canonical model of Fig. \[fig\_wz\](a) one finds that the average value of $w(z)\simeq-1$ on the range $z\lsim0.7$, while the average value of $w(z)<-1$ if the range of redshifts is extended to higher values. The $w=-1$ “phantom divide” is crossed at $z\simeq0.46$ for $\fvn\simeq0.76$. One recent study [@ZZ] finds mild 95% evidence for an equation of state that crosses the phantom divide from $w>-1$ to $w<-1$ in the range $0.25<z<0.75$ in accord with the expectation. By contrast, another study [@SCHMPS] at redshifts $z<1$ draws different conclusions about dynamical dark energy, but for the given uncertainties in $w(z)$ the data is consistent with Fig. \[fig\_coD\](a) as well as with a cosmological constant [@obs]. The fact that $w(z)$ is a different sign to the dark energy case for $z>2$ is another way of viewing our statement above that the redshift range $2\lsim z\lsim6$ may be optimal for discriminating model differences. [(a)]{}-5pt ![The artificial equivalent of an equation of state constructed using the effective comoving distance (\[eos\]), plotted for the tracker solution with best–fit value $\fvn=0.762$, and two different values of $\OmMn$: [(a)]{} the canonical dressed value $\OmMn=\frn12(1-\fvn)(2+\fvn)=0.33$; [(b)]{} $\OmMn=0.279$.[]{data-label="fig_wz"}](wL.eps "fig:"){width="2.in"}-5pt ![The artificial equivalent of an equation of state constructed using the effective comoving distance (\[eos\]), plotted for the tracker solution with best–fit value $\fvn=0.762$, and two different values of $\OmMn$: [(a)]{} the canonical dressed value $\OmMn=\frn12(1-\fvn)(2+\fvn)=0.33$; [(b)]{} $\OmMn=0.279$.[]{data-label="fig_wz"}](wL2.eps "fig:"){width="2.in"} The $H(z)$ measure ------------------ Further observational diagnostics can be devised if the expansion rate $H(z)$ can be observationally determined as a function of redshift. Recently such a determination of $H(z)$ at $z=0.24$ and $z=0.43$ has been made using redshift space distortions of the BAO scale in the model [@GCH]. This technique is of course model dependent, and the Kaiser effect would have to be re-examined in the model before a direct comparison of observational results could be made. A model–independent measure of $H(z)$, the redshift time drift test, is discussed below. In Fig. \[fig\_HH0\] we compare $H(z)/\Hm$ for the model to spatially flat models with the same parameters chosen in Fig. \[fig\_coD\]. The most notable feature is that the slope of $H(z)/\Hm$ is less than in the case, as is to be expected for a model whose (dressed) deceleration parameter varies more slowly than for . ![The function $\Hm^{-1} H(z)$ for the model with $\fvn=0.762$ (solid line) is compared to $\Hm^{-1} H(z)$ for three spatially flat models with the same values of $(\OmMn,\OmLn)$ as in Fig. \[fig\_coD\] (dashed lines).[]{data-label="fig_HH0"}](HH0.eps){width="2.in"} The $Om(z)$ measure ------------------- Recently a number of authors [@GCC; @SSS; @ZC] have discussed various roughly equivalent diagnostics of dark energy. For example, Sahni, Shafieloo and Starobinsky [@SSS], have proposed a diagnostic function Om(z)=\^[-1]{}, \[dSSS\]on account of the fact that it is equal to the constant present epoch matter density parameter, $\OmMn$, at all redshifts for a spatially flat FLRW model with pressureless dust and a cosmological constant. However, it is not constant if the cosmological constant is replaced by other forms of dark energy. For general FLRW models, $H(z)=[D'(z)]^{-1}\sqrt{1+\Omkn\Hm^2 D^2(z)}$, which only involves a single derivatives of $D(z)$. Thus the diagnostic (\[dSSS\]) is easier to reconstruct observationally than the equation of state parameter, $w(z)$. [(a)]{}-5pt ![The dark energy diagnostic $Om(z)$ of Sahni, Shafieloo and Starobinsky [@SSS] plotted for the tracker solution with best–fit value $\fvn=0.762$ (solid line), and $1\si$ limits (dashed lines) from ref.[@LNW]: [(a)]{} for the redshift range $0<z<1.6$ as shown in ref. [@SSS2]; [(b)]{} for the redshift range $0<z<6$.[]{data-label="fig_Om"}](Om16a.eps "fig:"){width="2.2in"}-5pt ![The dark energy diagnostic $Om(z)$ of Sahni, Shafieloo and Starobinsky [@SSS] plotted for the tracker solution with best–fit value $\fvn=0.762$ (solid line), and $1\si$ limits (dashed lines) from ref.[@LNW]: [(a)]{} for the redshift range $0<z<1.6$ as shown in ref. [@SSS2]; [(b)]{} for the redshift range $0<z<6$.[]{data-label="fig_Om"}](Om6b.eps "fig:"){width="2.2in"} The quantity $Om(z)$ is readily calculated for the model, and the result is displayed in Fig. \[fig\_Om\]. What is striking about Fig. \[fig\_Om\], as compared to the curves for quintessence and phantom dark energy models as plotted in ref. [@SSS], is that the initial value Om(0)=23.H’\|\_0=[2(8\^3-3\^2+4)(2+)(4\^2++4)\^2]{} \[intc\]is substantially larger than in the spatially flat dark energy models. Furthermore, for the model $Om(z)$ does not asymptote to the dressed density parameter $\OmMn$ in any redshift range. For quintessence models $Om(z)>\OmMn$, while for phantom models $Om(z)<\OmMn$, and in both cases $Om(z)\to\OmMn$ as $z\to\infty$. In the model, $Om(z)>\OmMn\simeq0.33$ for $z\lsim1.7$, while $Om(z)<\OmMn$ for $z\gsim1.7$. It thus behaves more like a quintessence model for low $z$, in accordance with Fig. \[fig\_wz\]. However, the steeper slope and the different large $z$ behaviour mean the diagnostic is generally very different to that of typical dark energy models. For large $z$, $\OMMn<Om(\infty)<\OmMn$, if $\fvn>0.25$. Interestingly enough, a recent analysis of SneIa, BAO and CMB data [@SSS2] for dark energy models with two different empirical fitting functions for $w(z)$ gives an intercept $Om(0)$ which is larger than expected for typical quintessence or phantom energy models, and in the better fit of the two models the intercept (see Fig. 3 of ref. [@SSS2]) is close to the value expected for the model, which is tightly constrained to the range $0.638<Om(0)<0.646$ if $\fvn=0.76^{+0.12}_{-0.09}$. The Alcock–Paczyński test and baryon acoustic oscillations ---------------------------------------------------------- Some time ago Alcock and Paczyński devised a test [@AP] which relies on comparing the radial and transverse proper length scales of spherical standard volumes comoving with the Hubble flow. This test, which determines the function f=[1z]{}\|z\|=[HDz]{}, \[fAP\] was originally conceived to distinguish FLRW models with a cosmological constant from those without a $\Lambda$ term. The test is free from many evolutionary effects, but relies on one being able to remove systematic distortions due to peculiar velocities. Current detections of the BAO scale in galaxy clustering statistics [@bao; @Percival] can in fact be viewed as a variant of the Alcock–Paczyński test, as they make use of both the transverse and radial dilations of the fiducial comoving BAO scale to present a measure DV=\^[1/3]{}=Df\^[-1/3]{}. \[BAOr\] [(a)]{}-5pt ![[(a)]{} The Alcock–Paczyński test function $f\Ns{AP}=HD/z$; and [(b)]{} the BAO radial test function $\Hm D\Z V=\Hm Df\Ns{AP}^{-1/3}$. In each case the model with $\fvn=0.762$ (solid line) is compared to three spatially flat models with the same values of $(\OmMn,\OmLn)$ as in Fig. \[fig\_coD\] (dashed lines).[]{data-label="fig_AP"}](AP.eps "fig:"){width="1.8in"}-5pt ![[(a)]{} The Alcock–Paczyński test function $f\Ns{AP}=HD/z$; and [(b)]{} the BAO radial test function $\Hm D\Z V=\Hm Df\Ns{AP}^{-1/3}$. In each case the model with $\fvn=0.762$ (solid line) is compared to three spatially flat models with the same values of $(\OmMn,\OmLn)$ as in Fig. \[fig\_coD\] (dashed lines).[]{data-label="fig_AP"}](H0Dv.eps "fig:"){width="2.in"} In Fig. \[fig\_AP\] the Alcock–Paczyński test function (\[fAP\]) and BAO scale measure (\[BAOr\]) of the model are compared to those of the spatially flat model with different values of ($\OmLn$,$\OmLn$). Over the range of redshifts $z<1$ studied currently with galaxy clustering statistics, the $f\Ns{AP}$ curve distinguishes the model from the models much more strongly than the $D\Z V$ test function. In particular, the $f\Ns{AP}$ has a distinctly different shape to that of the model, being convex. The primary reason for use of the integral measure (\[BAOr\]) has been a lack of data. Future measurements with enough data to separate the radial and angular BAO scales are a potentially powerful way of distinguishing the model from . Recently Gaztañaga, Cabré and Hui [@GCH] have made the first efforts to separate the radial and angular BAO scales in different redshift slices. Although they have not yet published separate values for the radial and angular scales, their results are interesting when compared to the expectations of the model. Their study yields best-fit values of the present total matter and baryonic matter density parameters, $\OmMn$ and $\OmBn$, which are in tension with WMAP5 parameters fit to the model. In particular, the ratio of non-baryonic cold dark matter to baryonic matter has a best-fit value $\OmCn/\OmBn=(\OmMn-\OmBn)/ \OmBn$ of 3.7 in the $0.15<z<0.3$ sample, 2.6 in the $0.4<z<0.47$ sample, and 3.6 in the whole sample, as compared to the expected value of 6.1 from WMAP5. The analysis of the 3–point correlation function yields similar conclusions, with a best fit [@GCCCF] $\OmMn=0.28\pm0.05$, $\OmBn=0.079\pm0.025$. By comparison, the parameter fit to the model of ref. [@LNW] yields dressed parameters $\OmMn=0.33^{+0.11}_{-0.16}$, $\OmBn=0.080^{+0.021}_{-0.013}$, and a ratio $\OmCn/\OmBn=3.1^{+2.5}_{-2.4}$. Since other forms of dark energy are not generally expected to give rise to a renormalization of the ratio of non-baryonic to baryonic matter, this is encouraging for the model. Test of (in)homogeneity ----------------------- Recently Clarkson, Bassett and Lu [@CBL] have constructed what they call a “test of the Copernican principle” based on the observation that for homogeneous, isotropic models which obey the Friedmann equation, the present epoch curvature parameter, a constant, may be written as =[\[H(z)D’(z)\]\^2-1\^2]{}\[ctest1\] for all $z$, irrespective of the dark energy model or any other model parameters. Consequently, taking a further derivative, the quantity (z)1+H\^2(DD”-D’\^2)+HH’DD’\[ctest2\] must be zero for all redshifts for any FLRW geometry. A deviation of $\CC(z)$ from zero, or of (\[ctest1\]) from a constant value, would therefore mean that the assumption of homogeneity is violated. Although this only constitutes a test of the assumption of the Friedmann equation, i.e., of the Cosmological Principle rather than the broader Copernican Principle adopted in ref. [@clocks], the average inhomogeneity will give a clear and distinct prediction of a non-zero $\CC(z)$ for the model. [(a)]{} ![[Left panel:]{} The (in)homogeneity test function $\BB(z)=[HD']^2-1$ is plotted for the tracker solution with best–fit value $\fvn=0.762$ (solid line), and compared to the equivalent curves $\BB=\Omkn(\Hm D)^2$ for two different models with small curvature: $\OmMn=0.28$, $\OmLn=0.71$, $\Omkn=0.01$; [(b)]{} $\OmMn=0.28$, $\OmLn=0.73$, $\Omkn=-0.01$. The (in)homogeneity test function $\CC(z)$ is plotted for the $\fvn=0.762$ tracker solution.[]{data-label="fig_Bex"}](dok.eps "fig:"){width="2.2in"} ![[Left panel:]{} The (in)homogeneity test function $\BB(z)=[HD']^2-1$ is plotted for the tracker solution with best–fit value $\fvn=0.762$ (solid line), and compared to the equivalent curves $\BB=\Omkn(\Hm D)^2$ for two different models with small curvature: $\OmMn=0.28$, $\OmLn=0.71$, $\Omkn=0.01$; [(b)]{} $\OmMn=0.28$, $\OmLn=0.73$, $\Omkn=-0.01$. The (in)homogeneity test function $\CC(z)$ is plotted for the $\fvn=0.762$ tracker solution.[]{data-label="fig_Bex"}](Cex.eps "fig:"){width="2.in"} The functions (\[ctest1\]) and (\[ctest2\]) are computed in ref.[@obs]. Observationally it is more feasible to fit (\[ctest1\]) which involves one derivative less of redshift. In Fig. \[fig\_Bex\] we exhibit both $\CC(z)$, and also the function $\BB(z)=[HD']^2-1$ from the numerator of (\[ctest1\]) for the model, as compared to two models with a small amount of spatial curvature. A spatially flat FLRW model would have $\BB(z)\equiv0$. In other FLRW cases $\BB(z)$ is always a monotonic function whose sign is determined by that of $\Omkn$. An open $\Lambda=0$ universe with the same $\OmMn$ would have a monotonic function $\BB(z)$ very much greater than that of the model. Time drift of cosmological redshifts ------------------------------------ For the purpose of the $Om(z)$ and (in)homogeneity tests considered in the last section, $H(z)$ must be observationally determined, and this is difficult to achieve in a model independent way. There is one way of achieving this, however, namely by measuring the time variation of the redshifts of different sources over a sufficiently long time interval [@SML], as has been discussed recently by Uzan, Clarkson and Ellis [@UCE]. Although the measurement is extremely challenging, it may be feasible over a 20 year period by precision measurements of the Lyman-$\al$ forest in the redshift range $2<z<5$ with the next generation of Extremely Large Telescopes [@ELT]. In ref. [@obs] an analytic expression for $\Hm^{-1}\Deriv\dd\ta z$ is determined, the derivative being with respect to wall time for observers in galaxies. The resulting function is displayed in Fig. \[fig\_zdot\] for the best-fit model with $\fvn=0.762$, where it is compared to the equivalent function for three different spatially flat models. What is notable is that the curve for the model is considerably flatter than those of the models. This may be understood to arise from the fact that the magnitude of the apparent acceleration is considerably smaller in the model, as compared to the magnitude of the acceleration in models. For models in which there is no apparent acceleration whatsoever, one finds that $\Hm^{-1}\Deriv\dd\ta z$ is always negative. If there is cosmic acceleration, real or apparent, at late epochs then $\Hm^{-1}\Deriv\dd\ta z$ will become positive at low redshifts, though at a somewhat larger redshift than that at which acceleration is deemed to have begun. ![The function $\Hm^{-1}\Deriv\dd\ta z$ for the model with $\fvn=0.762$ (solid line) is compared to $\Hm^{-1}\Deriv\dd\ta z$ for three spatially flat models with the same values of $(\OmMn,\OmLn)$ as in Fig. \[fig\_coD\] (dashed lines).[]{data-label="fig_zdot"}](zdot.eps){width="2.2in"} Fig. \[fig\_zdot\] demonstrates that a very clear signal of differences in the redshift time drift between the model and models might be determined at low redshifts when $\Hm^{-1}\Deriv\dd\ta z$ should be positive. In particular, the magnitude of $\Hm^{-1}\Deriv\dd\ta z$ is considerably smaller for the model as compared to models. Observationally, however, it is expected that measurements will be best determined for sources in the Lyman $\al$ forest in the range, $2<z<5$. At such redshifts the magnitude of the drift is somewhat more pronounced in the case of the models. For a source at $z=4$, over a period of $\de\ta=10$ years we would have $\de z=-3.3\times10^{-10}$ for the model with $\fvn=0.762$ and $\Hm=61.7\kmsMpc$. By comparison, for a spatially flat model with $\Hm=70.5\kmsMpc$ a source at $z=4$ would over ten years give $\de z=-4.7\times10^{-10}$ for $(\OmMn,\OmLn)=(0.249,0.751)$, and $\de z=-7.0\times10^{-10}$ for $(\OmMn,\OmLn)=(0.279,0.721)$. Discussion ========== The tests outlined here demonstrate several lines of investigation to distinguish the model from models of homogeneous dark energy. The (in)homogeneity test of Bassett, Clarkson and Lu is definitive, since it tests the validity of the Friedmann equation directly. In performing these tests, however, one must be very careful to ensure that data has not been reduced with built–in assumptions that use the Friedmann equation. For example, current estimates of the BAO scale such as that of Percival [@Percival] do not determine $D\Z V$ directly from redshift and angular diameter measures, but first perform a Fourier space transformation to a power spectrum, assuming a FLRW cosmology. Redoing such analyses for the model may involve a recalibration of relevant transfer functions. In the case of supernovae, one must also take care as compilations such as the Union [@union] and Constitution [@Hicken] datasets use the SALT method to calibrate light curves. In this approach empirical light curve parameters and cosmological parameters – [assuming the Friedmann equation]{} – are simultaneously fit by analytic marginalisation before the raw apparent magnitudes are recalibrated. As Hicken discuss [@Hicken], a number of systematic discrepancies exist between data reduced by the SALT, SALT2, MLCS31 and MLCS17 techniques even within the model. In the case of the model, we find considerable differences between the different approaches [@SW]. In principle, at present there appear to be enough supernovae to decide between the and models on Bayesian evidence, but one is led to different conclusions depending on how the data is reduced. It is therefore important that the systematic issues are unravelled. The value of the dressed Hubble constant is also an observable quantity of considerable interest. A recent determination of $\Hm$ by Riess [@shoes] poses a challenge for the model. However, it is a feature of the model that a 17–22% variance in the apparent Hubble flow will exist on local scales below the scale of statistical homogeneity, and this may potentially complicate calibration of the cosmic distance ladder. Further quantification of the variance in the apparent Hubble flow in relationship to local cosmic structures would provide an interesting possibility for tests of the cosmology for which there are no counterparts in the standard cosmology. I thank Prof. Remo Ruffini and ICRANet for support and hospitality while the work of ref. [@obs] was undertaken. This work was also partly supported by the Marsden fund of the Royal Society of New Zealand. [9]{} D.L. Wiltshire, , 123512 (2009). D.L. Wiltshire, in [Dark Matter in Astroparticle and Particle Physics: Proc. of the 6th International Heidelberg Conference]{}, eds H.V. Klapdor–Kleingrothaus and G.F. Lewis, (World Scientific, Singapore, 2008) pp. 565-596 \[arXiv:0712.3984\]. F. Hoyle and M.S. Vogeley, , 641 (2002); , 751 (2004). T. Buchert, , 105 (2000); , 1381 (2001). T. Buchert, , 467 (2008). P.J.E. Peebles, in this volume; arXiv:0910.5142. D.L. Wiltshire, New J. Phys. [9]{}, 377 (2007). D.L. Wiltshire, , 251101 (2007). D.L. Wiltshire, , 084032 (2008). B.M. Leith, S.C.C. Ng and D.L. Wiltshire, , L91 (2008). S. Jha, A.G. Riess and R.P. Kirshner, , 122 (2007). N. Li and D.J. Schwarz, , 083531 (2008). E. Komatsu , Astrophys. J. Suppl. [180]{}, 330 (2009). A.G. Riess , , 98 (2007). B.E. Schaefer, , 16 (2007); N. Liang, W. K. Xiao, Y. Liu and S.N. Zhang, , 354 (2008). L. Amati, C. Guidorzi, F. Frontera, M. Della Valle, F. Finelli, R. Landi and E. Montanari, , 577 (2008); R. Tsutsui, T. Nakamura, D. Yonetoku, T. Murakami, Y. Kodama and K. Takahashi, JCAP [08]{} (2009) 015. B.E. Schaefer, in preparation. G.B. Zhao and X. Zhang, arXiv:0908.1568. P. Serra, A. Cooray, D.E. Holz, A. Melchiorri, S. Pandolfi and D. Sarkar, arXiv:0908.3186. E. Gaztañaga, A. Cabre and L. Hui, , 1663 (2009). J.A. Gu, C.W. Chen and P. Chen, New J. Phys. [11]{}, 073029 (2009). V. Sahni, A. Shafieloo and A. A. Starobinsky, , 103502 (2008). C. Zunckel and C. Clarkson, , 181301 (2008). A. Shafieloo, V. Sahni and A.A. Starobinsky, , 101301 (2009). C. Alcock and B. Paczyński, Nature [281]{}, 358 (1979). D.J. Eisenstein , (2005) 560; S. Cole , (2005) 505. W.J. Percival , , 1053 (2007); W.J. Percival , arXiv:0907.1660. E. Gaztañaga, A. Cabré, F. Castander, M. Crocce and P. Fosalba, , 801 (2009). C. Clarkson, B. Bassett and T.C. Lu, , 011301 (2008). A. Sandage, , 319 (1962); G.C. McVittie, , 334 (1962); A. Loeb, , L111 (1998). J.P. Uzan, C. Clarkson and G.F.R. Ellis, , 191303 (2008). P.S. Corasaniti, D. Huterer and A. Melchiorri, , 062001 (2007); J. Liske [et al.*]{}, , 1192 (2008). M. Kowalski , , 749 (2008). M. Hicken , , 1097 (2009). P.R. Smale and D.L. Wiltshire, in preparation. A.G. Riess , , 539 (2009).	{ "pile_set_name": "ArXiv" }
--- abstract: \| We study how the propagation speed of interfaces in the Allen-Cahn phase field model for phase transformations in solids consisting of the elasticity equations and the Allen-Cahn equation depends on two parameters of the model. The two parameters control the interface energy and the interface width but change also the interface speed. To this end we derive an asymptotic expansion of second order for the interface speed, called the kinetic relation, and prove that it is uniformly valid in both parameters. As a consequence we show that the model error is proportional to the interface width divided by the interface energy. We conclude that simulations of interfaces with low interface energy based on this model require a very small interface width, implying a large numerical effort. Effective simulations thus need adaptive mesh refinement or other advanced techniques. This version of the paper contains the proofs of Theorem 4.5 and Lemma 5.8, which are omitted in the version published in Continuum Mechanics and Thermodynamics. author: - \| Hans-Dieter Alber[^1]\ [Fachbereich Mathematik, Technische Universität Darmstadt]{}\ [Schlossgartenstr. 7, 64289 Darmstadt, Germany]{} title: 'Asymptotics and numerical efficiency of the Allen-Cahn model for phase interfaces with low energy in solids' --- Introduction {#S1} ============ In this paper we study how the propagation speed of interfaces in the Allen-Cahn phase field model for phase transformations in elastic solids depends on two parameters of the model. The model consists of the partial differential equations of linear elasticity coupled to the standard Allen-Cahn phase field equation. The two parameters, which we denote by $\mu$ and $\la$, control the interface energy and the interface width, but variation of these parameters also changes the interface speed, or more precisely, the form of the relation, which determines the interface speed as a function of the stress field and the curvature of the interface. In sharp interface models this relation is called kinetic relation. We use this notion also for phase field models. Our goal is therefore to determine the kinetic relation for the Allen-Cahn model and the dependence of it on the two model parameters. To this end we must derive an asymptotic expansion for the propagation speed of the interface and prove an error estimate for this asymptotic expansion, which holds uniformly in both parameters. Our results have consequences for the efficiency of the Allen-Cahn model in numerical simulations of interfaces with small interface energy. These consequences are also discussed. Let $\Om \subseteq \R^3$ be a bounded open set with a sufficiently smooth boundary $\pa\Om$. The points of $\Om$ represent the material points of a solid elastic body. The unknown functions in the model are the displacement $u(t,x) \in \R$ of the material point $x$ at time $t$, the Cauchy stress tensor $T(t,x)\in \ES^3$, where $\ES^3$ denotes the set of all symmetric $3\times 3$-matrices, and the order parameter $S(t,x)\in \R$. These unknowns must satisfy the model equations $$\begin{aligned} -\div_x\, T &=& {\sf b}, \label{E1.1} \\ T &=& D\big(\ve(\na_x u ) - \ov{\ve}S \big),\label{E1.2} \\ \pa_t S &=& - \frac{c}{(\mu\la)^{1/2}} \Big(\pa_S {\sf W} \big(\ve(\na_x u ), S \big) + \frac{1}{\mu^{1/2}} \hat{\psi}'(S) - \mu^{1/2} \la \Da_x S \Big) \label{E1.3} \end{aligned}$$ in the domain $[0,\infty)\ti \Om$. The boundary and initial conditions are $$\begin{aligned} {2} u(t,x) &= {\sf U}(t,x), &\qquad& (t,x) \in [0,\infty)\ti \pa\Om, \label{E1.4} \\ \pa_{n_{\pa\Om}} S(t,x) &= 0, && (t,x) \in [0,\infty) \ti \pa\Om, \label{E1.4a} \\ S(0,x) &= {\sf S}(x), && x \in \Om. \label{E1.5}\end{aligned}$$ Here ${\sf b}(t,x)\in \R^3$, ${\sf U}(t,x) \in \R^3$, ${\sf S}(t,x) \in \R$ denote given data, the volume force, boundary displacement and initial data. $\pa_{n_{\pa\Om}}$ denotes the derivative in direction of the unit normal vector $n_{\pa\Om}$ to the boundary. The deformation gradient $\na_x u(t,x)$ is the $3\ti 3$–matrix of first order partial derivatives of $u$ with respect to the components $x_k$ of $x$, and the strain tensor $$\ve(\na_x u) = \frac{1}{2}\big(\na_x u+(\na_x u)^T\big)$$ is the symmetric part of the deformation gradient, where $(\na_x u)^T$ denotes the transpose matrix. The elasticity tensor $D:\ES^3 \to \ES^3$ is a linear symmetric, positive definite mapping, $\ov{\ve}\in\ES^3$ is a given constant matrix, the transformation strain, and $\mu>0$ and $\lambda>0$ are parameters. The elastic energy is given by $$\label{E1.6} {\sf W}\big(\ve(\na_x u),S\big)= \frac{1}{2}\Big(D\big(\ve(\na_x u) - \ov{\ve} S \big)\Big):\big(\ve(\na_x u) - \ov{\ve} S\big),$$ with the matrix scalar product $A:B=\sum_{i,j}a_{ij}b_{ij}$. Using , we obtain for the derivative $$\label{E1.7} \pa_S {\sf W}(\ve,S)=-\ov{\ve}:D\big(\ve(\na_x u)-\ov{\ve}S\big)=-\ov{\ve}:T.$$ $c>0$ is a given constant and $\hat{\psi}:\R\to[0,\infty)$ is a double well potential satisfying $$\hat{\psi}(0) = \hat{\psi}(1)=0, \qquad \hat{\psi}(\zeta)>0 \mbox{ for } \zeta \not= 0,1.$$The precise assumptions on $\hat{\psi}$, which we need in our investigations, are stated in . This completes the formulation of the model. and are the equations of linear elasticity theory. This subsystem is coupled to the Allen-Cahn equation , which governs the evolution of the order parameter $S$. The system – satisfies the second law of thermodynamics. More precisely, the Clausius-Duhem inequality is satisfied with the free energy $$\label{E1.9} \psi_{\mu\la}^(\ve,S)={\sf W}(\ve,S)+ \frac{1}{\mu^{1/2}} \hat{\psi}(S) + \frac{\mu^{1/2}\lambda}{2} \|\na_x S\|^2.$$ From this expression we see that the parameter $\lambda$ determines the energy density of the phase interface. We assume that the parameters $\mu$ and $\la$ vary in intervals $(0,\mu_0]$ and $(0,\la_0]$, respectively, with $\mu_0 > 0$ and $\la_0 > 0$ chosen sufficiently small. The scaling $\frac{c}{(\mu\lambda)^{1/2}}$ on the right hand side of is necessary for otherwise the propagation speed of the diffuse interface would tend to zero for $\mu \ra 0$ or $\la \ra 0$. We give now a slightly sketchy overview of the main results in this article. The precise definitions and statements are given in Section \[S2\], and in particular in Section \[S2.4\]. We denote solutions of the Allen-Cahn equation by $(u_{\rm AC},T_{\rm AC},S_{\rm AC})$, to distinguish them from aproximate solutions, which we construct later. The values $S_{\rm AC}(t,x) \approx 0$ or $S_{\rm AC}(t,x) \approx 1$ indicate that at the point $x \in \Om$ at time $t$ the crystal structure of the material of the solid belongs to phase 1 or to phase 2, respectively. The set of all $x \in \Om$ with $0 < S_{\rm AC}(t,x) < 1$ is the region of the diffuse interface at time $t$. The level set $$\label{E1.1.9} \Gm_{\rm AC}(t) = \big\{ x \in \Om \bigm\| S_{\rm AC}(t,x) = \frac12 \big\}$$ belongs to this region. For $x \in \Gm_{\rm AC}(t)$ we denote by $s_{\rm AC}(t,x) \in \R$ the normal speed of this level set at $x$, and we call $s_{\rm AC}$ the speed of the diffuse interface. For this speed we derive an expression of the form $$\label{E1.1.10} s_{\rm AC}(t,x) = (s_{00} + \la^{1/2}s_{01}) + \mu^{1/2} ( s_{10} + \la^{1/2} s_{11} ) + \|\ln \mu\|^3 \mu\, s_\infty^{(\mu\la)}.$$ We call this expression the kinetic relation of the Allen-Cahn model. It is the central result of this paper. The remainder term $s_\infty^{(\mu\la)}$ depends on $\mu$ and $\la$, but in Section \[S6\] we prove that there exist numbers $\mu_0 > 0$ and $\la_0 > 0$ and a constant $C_{\cal E}$ such that the $L^2$–norm satisfies $$\label{E1.1.11} \\| s^{(\mu\la)}_\infty(t) \\|_{L^2_{\Gm_{\rm AC}(t)}} \leq C_{\cal E}$$ for all $0 < \mu \leq \mu_0$ and all $0 < \la \leq \la_0$, where $s^{(\mu\la)}_\infty(t)$ denotes the function $x \mapsto s^{(\mu\la)}_\infty(t,x):\Gm_{\rm AC}(t) \ra \R$. Therefore is an asymptotic expansion for the propagation speed of the diffuse interface, which is uniformly valid with respect to the parameters $\mu$ and $\la$. For sufficiently small $\mu$ the leading term $s_0(t,x) = s_{00} + \la^{1/2} s_{01}$ and the second term $\mu^{1/2}s_1(t,x) = \mu^{1/2} ( s_{10} + \la^{1/2} s_{11} )$ dominate over the remainder term $\|\ln \mu\|^3 \mu\, s_\infty^{(\mu\la)}$. We can therefore read off the behavior of the Allen-Cahn model with respect to the parameters $\mu$ and $\la$ from the first two terms in . The terms $s_{00} + \la^{1/2}s_{01}$ and $s_{10} + \la^{1/2}s_{11}$ are explicitly given in . We restrict ourselves here to state the form of the leading term. We assume that $0 \leq t_1 < t_2 < \infty$ are given times. We study the propagation of the interface for $t$ varying in the interval $[t_1,t_2]$. For $t$ from this interval the leading term is $$\label{E1.1.12} s_0(t,x) = s_{00} + \la^{1/2} s_{01} = \frac{c}{c_1} \Big( - \ov{\ve}:\langle \hat{T} \rangle(t,x) + \la^{1/2} c_1 \ka_\Gm(t,x) \Big),$$ where $c$ is the mobility constant from , $c_1 = \int_0^1 \sqrt{2\hat{\psi}(\vartheta)}\, d\vartheta$ is computed from the double well potential, $\ka_\Gm(t,x)$ is twice the mean curvature of the surface $\Gm_{\rm AC}(t)$ at the point $x$, and ${\hat{T}}$ is the stress field in the solution $(t,x) \mapsto \big({\hat{u}}(t,x),{\hat{T}}(t,x)\big)$ of the transmission problem $$\begin{aligned} -\div_x \hat{T} &=& {\sf b}, \label{E1.1.13} \\ \hat{T} &=& D\big( \ve(\na_x \hat{u}) - \ov{\ve} \hat{S}\big), \label{E1.1.14} \\ {[\hat{u}]} &=& 0, \label{E1.1.15} \\ {[\hat{T}]}n &=& 0,\label{E1.1.16} \\ {\hat{u}}(t)\rain{\pa\Om} &=& {\sf U}(t),\label{E1.1.17} \end{aligned}$$ Here $(t,x) \mapsto {\hat{S}}(t,x)$ is a function, which takes only the values $0$ or $1$ and jumps across the interface $$\Gm_{\rm AC} = \{ (t,x) \in [t_1,t_2] \ti \Om \mid x \in \Gm_{\rm AC}(t) \}.$$ $[{\hat{u}}]$ and $[{\hat{T}}]$ are the jumps of the functions ${\hat{u}}$ and ${\hat{T}}$ across $\Gm_{\rm AC}$, and ${\sf b}$, ${\sf U}$ are the volume force and boundary data from equations and . The equations , hold in the domain $\big([t_1,t_2] \ti \Om \big) \setminus \Gm_{\rm AC}$. For every fixed $t$ the problem – is an elliptic transmission problem for the function $x \mapsto \big({\hat{u}}(t,x),{\hat{T}}(t,x)\big)$ in the domain $\Om$. Therefore $t$ can be considered to be a parameter in this problem. Finally, the expression $$\langle {\hat{T}}\rangle = \frac12 \big({\hat{T}}^{(+)} + {\hat{T}}^{(-)}\big)$$ in equation denotes the mean value of the values ${\hat{T}}^{(+)}$ and ${\hat{T}}^{(-)}$ on both sides of the interface $\Gm_{\rm AC}$. The terms $s_{10}$ and $s_{11}$ are determined by a more complicated transmission problem, for which the transmission conditions are also posed on the interfae $\Gm_{\rm AC}$, and by a coupled system of ordinary differential equations for two functions $S_0$ and $S_1$, which are needed in the construction of an asymptotic solution of the Allen-Cahn model – . The coefficients of the second transmission problem and of the system of ordinary differential equations depend on the solution $({\hat{u}},{\hat{T}})$ of – . Both transmission problems together can thus be considered to be a larger transmission problem, which is recursively solvable. The derivation of the kinetic relation is based on the construction of an asymptotic solution $(u^{(\mu\la)},T^{(\mu\la)},S^{(\mu\la)})$ for the Allen-Cahn model. From this asymptotic solution it is seen that the width of the diffuse interface in the Allen-Cahn model is proportional to the parameter $$\label{E1.1.18} B = (\mu\la)^{1/2}.$$ We call $B$ the interface width parameter. As will be explained in Section \[S2.4\], the interface energy density is proportional to the parameter $$\label{E1.1.19} E = \la^{1/2}.$$ We call $E$ the interface energy parameter. The kinetic relation and the equation together have consequences for the efficiency of numerical simulations of interfaces with low interface energy density, which we sketch here. A precise discussion is given in Section \[S2.4\]. The explicit expressions in show that the second term $\mu^{1/2}(s_{10} + \la^{1/2} s_{11})$ in is of a very special form. We therefore argue that this term does not have a physical meaning, only the leading term $s_{00} + \la^{1/2} s_{01}$ is physically relevant. This means that in the term $${\cal E}^{(\mu\la)} = \mu^{1/2} ( s_{10} + \la^{1/2} s_{11} ) + \|\ln \mu\|^3 \mu\, s_\infty^{(\mu\la)}$$ is a mathematical error term, which in a precise numerical simulation of the evolution of the interface must be made small by choosing $\mu^{1/2}$ small enough. Therefore we call ${\cal E}^{(\mu\la)}$ the model error and $F = \mu^{1/2}$ the error parameter. By , the interface width is proportional to the error parameter $F$. This means that the interface width depends on the size of the model error. The smaller the model error is, which we want to allow, the smaller the interface width must be chosen. The total error in a numerical simulation consists of the model error and the numerical error. In order to make the numerical error small, the grid spacing must be chosen small enough to resolve the transition of the order parameter from $0$ to $1$ across the diffuse interface. When the interface width is small, we must therefore choose the grid spacing small, which means that the numerical effort is high. From and we see that for constant values of the error parameter $F$ the interface width is proportional to the interface energy density parameter $E$. Thus, when we want to precisely simulate an interface with small interface energy density, we must choose small values for $E$ and $F$, hence the interface width $B = EF$ becomes very small. As a consequence, also the grid spacing must be chosen very small, which means that numerical simulations of interfaces with low interface energy based on the Allen-Cahn model are not efficient. Of course, the efficiency can be improved by using adaptive mesh refinement and other advanced numerical techniques, but still it would be advantageous if such tools could be avoided. Often the Allen-Cahn model is formulated using the parameters $E$ and $F$ instead of $\mu$ and $\la$. It might therefore be helpful to shortly discuss the form, which our results take when this formulation is used. With these parameters the Allen-Cahn equation is $$\pa_t S = - \frac{c}{B} \Big(\pa_S {\sf W} \big(\ve(\na_x u ), S \big) + E \big(\frac{\hat{\psi}'(S)}{B} - B \Da_x S \big) \Big),$$ the free energy becomes $$\psi_{EB}^(\ve,S)={\sf W}(\ve,S)+ E \Big( \frac{ \hat{\psi}(S)}{B} + \frac{B}{2} \|\na_x S\|^2 \Big),$$ and the kinetic relation takes the form $$s_{\rm AC}(t,x) = (s_{00} + E s_{01}) + \frac{B}{E} ( s_{10} + E s_{11} ) + \Big\|2 \ln \Big(\frac{B}{E}\Big) \Big\|^3 \Big(\frac{B}{E}\Big)^2 s_\infty^{(BE)}.$$ From this equation we see that the model error $${\cal E}^{(BE)} = \frac{B}{E} ( s_{10} + E s_{11} ) + \Big\|2 \ln \Big(\frac{B}{E}\Big) \Big\|^3 \Big(\frac{B}{E}\Big)^2 s_\infty^{(BE)}$$ is governed by the ratio $\frac{B}{E}$. When one reduces the interface energy density $E$ and one wants to keep the model error constant, one must reduce the interface width $B$ by the same proportion. In a simulation of an interface without interface energy, the total model error is the sum ${\cal E}^{(BE)}_{\rm total} = E s_{01} + {\cal E}^{(BE)}$. To reduce the error in such a simulation we must reduce $E$ and the fraction $\frac{B}{E}$, hence $B$ must be reduced faster than $E$. The paper is organized as follows. The main results are presented in Section \[S2\], where we first state the transmission problems and the system of ordinary differential equations, whose solutions are needed to compute the coefficients $s_{00},\ldots,s_{11}$ in . These coefficients are explicitly given in . Moreover, in this theorem we also state properties of the asymptotic solution $(u^{(\mu\la)},T^{(\mu\la)},S^{(\mu\la)})$, which is constructed in later sections. In particular, we state the scaling law for the width of the diffuse interface. These properties are needed in Section \[S2.4\], where we precisely discuss the model error and the numerical efficiency. The estimate , which is the most important mathematical result of this paper, is stated in . Sections \[S3\] – \[S5\] contain the proof of . In Section \[S3\] we construct the approximate solution $(u^{(\mu\la)}, T^{(\mu\la)}, S^{(\mu\la)})$. That is, we state the inner and outer expansions which define the function $(u^{(\mu\la)}, T^{(\mu\la)}, S^{(\mu\la)})$. In these asymptotic expansions functions appear, which are obtained as solutions of systems of algebraic and differential equations. These systems are also stated in Section \[S3\]. The system for the outer expansion can be readily solved, and the solution of the system of ordinary differential equations for the inner expansion is more involved and is discussed in Section \[S4\]. In two equations of this system a linear differential operator appears with kernel different from $\{0\}$. In order that these differential equations be solvable the right hand sides must satisfy orthogonality conditions. The right hand sides contain the coefficients $s_{00}, \ldots,s_{11}$ of the kinetic relation . The orthogonality conditions dictate the form of these coefficients; therefrom the equation results. In Section \[S5\] we verify that $(u^{(\mu\la)}, T^{(\mu\la)}, S^{(\mu\la)})$ is really an asymptotic solution of the model equations – and prove the necessary estimates. In Section \[S6\] we prove the estimate . The proof uses the residue, with which the function $(u^{(\mu\la)}, T^{(\mu\la)}, S^{(\mu\la)})$ satisfies the equations – . The main difficulty in the proof is that though we want to prove that $s_\infty^{\mu\la}$ is bounded uniformly with respect to $\mu$ and $\la$, the residue term itself is not bounded for $\la \ra 0$, but instead behaves like $\la^{-1/2}$. In the bibliography of [@JElast2012] we gave many references to the literature on existence, uniqueness and asymptotics for models containing the Allen-Cahn and Cahn-Hilliard equations. We refer the reader to that bibliography and discuss here only some publications, which are of interest in the construction of asymptotic solutions. We believe that for the model – an asymptotic solution was constructed and used to identify the associated sharp interface problem for the first time in [@FG94], following earlier such investigations for other phase field models. For example, in [@Cagi86-1] these investigations were carried out for a model from solidification theory, which consists ot the Allen-Cahn equation coupled to the heat equation. The considerations in [@Cagi86-1; @FG94] are formal, since it is not shown that the asymptotic solution converges to an exact solution of the model equations for $\mu \ra 0$. Under the assumption that the associated sharp interface problems have smooth solutions, this was proved in [@MotSchatz95] for the Allen-Cahn equation, in [@ABCh94] for the Cahn-Hilliard equation, in [@CagiCh98] for the model from solidification theory and in [@AbelsSchau2014] for a model consisting of the Cahn-Hilliard equation coupled with the elasticity equations. The proofs use variants of a spectral estimate derived in [@Chen94]. For the model from solidification theory the associated sharp interface model is the Mullins-Sekerka model with surface tension. In [@CCO05] an asymptotic solution for the Cahn-Hilliard equation has recently been constructed with a method different from the one used in [@ABCh94], and which is similar to our method. In [@KarRap98] the numerical efficiency of simulations based on the phase field model consisting of the Allen-Cahn equation coupled to the heat equation is studied. It is shown that for suitably chosen coefficients of the model the second order term in an asymptotic expansion of the solution vanishes. By arguments similar to the ones we gave in the above discussion it is seen that this improves the numerical efficiency of the model. This result has been improved and generalized in [@Alm99; @ChenCagi06; @GarckeStin06]. A similar idea is also present in [@FifePen95]. Since the construction of asymptotic solutions is based on sharp interface problems, a rigorous analysis of these problems is of special interest. Of particular interest is the Hele-Shaw problem with surface tension, since this is the sharp interface problem associated with the Cahn-Hilliard equation. Existence, uniqueness, and regularity of classical solutions of this problem have been investigated in [@ChenEx96; @EschSim96; @EschSim97]. In [@EschSim98] it is shown that if the initial data are close to a sphere then a classical solution exists and converges to spheres. Existence of solutions to the Mullins-Sekerka problem mentioned above has been shown in [@Luckhaus90]. The model – describes the evolution of phase transitions in a solid when temperature effects are negligible. This model is the prototype of a large class of models obtained by extensions and generalizations of the model, which are used in the engineering sciences to simulate the behavior of complex and functional materials. From the very large literature in this field we cite here only [@Bhattacharya03; @SchradeMuellXuGross07; @XuSchradeMuellGrossRoedel10; @ZhangBhat2006; @ZuoGenKleinSteinXu14]. The kinetic relation {#S2} ==================== Notations {#S2.1} --------- For given fixed times $0 \leq t_1 < t_2 < \infty$ let $$Q = [t_1,t_2] \ti \Om \subseteq \R^4.$$ The construction of the asymptotic solution $(u^{(\mu\la)}, T^{(\mu\la)}, S^{(\mu\la)})$ is based on a surface $\Gm^{(\mu\la)}(t)$, which for $t_1 \leq t \leq t_2$ moves in $\Om$ and which will be the level set $$\Gm^{(\mu\la)}(t) = \big\{ x \in \Om \bigm\| S^{(\mu\la)}(t,x) = \frac12 \big\}.$$ We set $$\label{E2.1.1} \Gm^{(\mu\la)} = \{ (t,x) \in Q \mid x \in \Gm^{(\mu\la)}(t) \}.$$ To simplify the notation we often drop the index $\la$ or both indices $\mu$ and $\la$ and write $\Gm^{(\mu)}(t)$ and $\Gm^{(\mu)}$ or simply $\Gm(t)$ and $\Gm$. Similarly, we often write $(u^{(\mu)}, T^{(\mu)}, S^{(\mu)})$ for the asymptotic solution and use the same convention also in other notations. Both indices are specified if the dependence on $\la$ becomes important. The precise definition of the family $\{ \Gm^{(\mu\la)}(t) \}_{t_1 \leq t \leq t_2}$ is given in the nect section, and in Section \[S2.4\] we associate $\Gm^{(\mu\la)}(t)$ with the level set $\Gm^{(\mu\la)}_{\rm AC}(t) = \Gm_{\rm AC}(t)$ introduced in . To introduce notations we assume here that $\Gm$ is a known, orientable, three dimensional $C^k$–manifold with $k \geq 1$ sufficiently large embedded in $Q$ such that $\Gm(t)$ is a regular two dimensional surface in $\Om$ for every $t\in [t_1,t_2]$. Let $$\label{E2.1} n:\Gm \to \R^3$$ be a continuous vector field such that $n(t,x) \in \R^3$ is a unit normal vector to $\Gm(t)$ at $x \in \Gm(t)$, for every $t \in [t_1,t_2]$. For $\da > 0$ and $t \in [t_1,t_2]$ define the sets $$\label{E2.2} {\cal U}_\da(t) = \{x \in \Om \mid {\rm dist}(x,\Gm(t)) < \da \}\quad \mbox{and} \quad {\cal U}_\da = \{ (t,x) \in Q \mid x \in {\cal U}_\da(t) \}.$$ We assume that there is $\da > 0$ such that ${\cal U}_\da \subseteq Q$. Since $\Gm$ is a regular $C^1$–manifold in $Q$, then $\da$ can be chosen sufficiently small such that for all $t \in [t_1,t_2]$ the mapping $$\label{E2.3} (\eta,\xi) \mapsto x(t,\eta,\xi) = \eta + \xi n(t,\eta):\Gm(t) \ti (-\da,\da) \to {\cal U}_\da(t)$$ is bijective. We say that this mapping defines new coordinates $(\eta,\xi)$ in ${\cal U}_\da(t)$ and $(t,\eta,\xi)$ in ${\cal U}_\da$. If no confusion is püossible we switch freely between the coordinates $(t,x)$ and $(t,\eta,\xi)$. In particular, if $(t,x) \mapsto w(t,x)$ is a function defined on ${\cal U}_\da$ we write $w(t,\eta,\xi)$ for $w\big(t,x(t,\eta,\xi)\big)$, as usual. We use the standard convention and denote for a function $w$ defined on a subset $U$ of $Q$ by $w(t)$ the function $x \mapsto w(t,x)$, which is defined on the set $\{ x \mid (t,x) \in U \} \subseteq \R^3$. If $w$ is a function defined on ${\cal U}_\da(t) \setminus \Gm(t)$, we set for $\eta \in\Gm(t)$ $$\begin{aligned} w^{(\pm)} (\eta) &=& \lim_{\genfrac{}{}{0pt}{2}{\xi\to 0}{\xi > 0}} w \big( \eta \pm \xi n(t,\eta)\big), \\ (\pa_n^i w)^{(+)} (\eta) &=& \lim_{\genfrac{}{}{0pt}{2}{\xi\to 0}{\xi > 0} } \frac{\pa^i}{\pa\xi^i} w \big( \eta + \xi n(t,\eta)\big), \quad i \in \N, \\ (\pa_n^i w)^{(-)} (\eta) &=& \lim_{\genfrac{}{}{0pt}{2}{\xi\to 0}{\xi < 0} } \frac{\pa^i}{\pa\xi^i} w \big( \eta + \xi n(t,\eta)\big),\quad i \in \N, \\ {[w]} (\eta) &=& w^{(t)} (\eta) - w^{(-)}(\eta), \\[1ex] {[\pa_n^i w]}(\eta) &=& (\pa_n^i w)^{(+)}(\eta) - (\pa_n^i w)^{(-)}(\eta), \\ \lan w\ran (\eta) &=& \ha \big( w^{(+)} (\eta) + w^{(-)} (\eta) \big),\end{aligned}$$ provided that the one-sided limits in these equations exist. If $w$ is defined on ${\cal U}_\da \setminus \Gm$, we set $$w^{(\pm)} (t,\eta) = \big(w^{(\pm)} (t)\big)(\eta), \qquad (\pa_n^i w)^{(\pm)} (t,\eta) = \big( (\pa_n^i w)^{(\pm)} (t) \big)(\eta),$$ and define $[w](t,\eta)$, $\lan w\ran(t,\eta)$, $[\pa_n^i w](t,\eta)$ as above. Let $\tau_1(\eta)$, $\tau_2(\eta)\in\R^3$ be two orthogonal unit vectors to $\Gm(t)$ at $\eta \in\Gm(t)$. For functions $w:\Gm(t) \to\R$, $W:\Gm(t)\to\R^3$ we define the surface gradients by $$\begin{aligned} \label{E2.4} \na_\Gm w &=& (\pa_{\tau_1} w)\tau_1 + (\pa_{\tau_2} w)\tau_2, \\ \label{E2.5} \na_\Gm W &=& (\pa_{\tau_1} W)\otimes\tau_1 + (\pa_{\tau_2}W) \otimes \tau_2, \end{aligned}$$ where for vectors $c,d \in \R^3$ a $3\times 3$-matrix is defined by $$c\otimes d = (c_i d_j)_{i,j=1,2,3}.$$ With , we have for functions $w:{\cal U}_\da(t) \ra \R$ and $W:{\cal U}_\da(t) \ra \R^3$ at $\eta \in \Gm(t)$ the decompositions $$\begin{aligned} \na_x w &=& (\pa_n w) n + \na_\Gm w, \label{E2.6} \\ \na_x W &=& (\pa_n W)\otimes n + \na_\Gm W, \label{E2.7} \end{aligned}$$ where $n = n(t,\eta)$ is the unit normal vector to $\Gm(t)$. The normal speed of the family of surfaces $t \mapsto \Gm(t)$ is of fundamental importance in this paper. Therefore we give a precise definition. \[D2.1\] Let $m(t,\eta) = \big( m'(t,\eta),m''(t,\eta) \big) \in \R \ti \R^3$ be a normal vector to $\Gm$ at $(t,\eta) \in \Gm$. The normal speed of the family of surfaces $t \mapsto \Gm(t)$ at $\eta \in \Gm(t)$ is defined by $$\label{E2.normspeed1} s(t,\eta) = \frac{-m'(t,\eta)}{m''(t,\eta) \cdot n(t,\eta)},$$ with the unit normal vector $n(t,\eta) \in \R^3$ to $\Gm(t)$. Note that with this definition the speed is measured positive in the direction of the normal vector field $n$. Since $m''(t,\eta) \in \R^3$ is a normal vector to $\Gm(t)$, the denominator in is different from zero. If $\om = (\om',\om'') \in \R \ti \R^3$ is a tangential vector to $\Gm$ at $(t,\eta)$ with $\om' \neq 0$, then with the unit normal $n(t,\eta) \in \R^3$ to $\Gm(t)$ the vector $(-\om'' \cdot n,\om' n)$ is a normal vector to $\Gm$ at $(t,\eta)$, hence implies that the normal speed at $\eta \in \Gm(t)$ is given by $$\label{E2.normspeed2} s(t,\eta) = \frac{n \cdot \om'' }{\om' n\cdot n} = \frac{n \cdot \om'' }{\om'}.$$ For later use we prove the following \[L2.normspeed\] Let $x \in {\cal U}_\da(t_0)$ be a point having the representation $x = \eta + n(t,\eta) \xi$ in the $(\eta,\xi)$–coordinates, where $\eta = \eta(t,x) \in \Gm(t)$ and $\xi = \xi(t,x)$. Then the normal speed satisfies $$\label{E2.normspeed3} s(t_0,\eta) = n(t_0,\eta) \cdot \pa_t \eta(t_0,x) = - \pa_t\, \xi(t_0,x).$$ The tangential component of the vector $\pa_t \eta(t_0,x) \in \R^3$ to the surface $\Gm(t_0)$ is equal to $-\xi \pa_t n\big( t_0,\eta(t_0,x) \big)$. [Proof:]{} By definition of ${\cal U}_\da$, there is a neighborhood $U$ of $t_0$ in $[t_1,t_2]$ such that $\{x\} \ti U \subseteq {\cal U}_\da$, which implies that $x$ has the representation $$x = \eta(t,x) + \xi(t,x) n\big( t,\eta(t,x) \big).$$ for all $t \in U$. We differentiate this equation and obtain $$\label{E2.patetadeco} 0=\pa_t x=n \,\pa_t \xi + \xi\, \pa_t n + \pa_t \eta.$$ From $0=\pa_t 1 = \pa_t \|n\|^2 = 2n \cdot \pa_t n$ we see that $\pa_t n$ is tangential to $\Gm(t)$, hence implies that the tangential component of $\pa_t \eta$ is equal to $- \xi \pa_t n $. Multiplication of with $n$ yields $$\label{E2.normspeed4} \pa_t \xi= - n \cdot \pa_t \eta.$$ Since $\pa_t \big(t, \eta(t,x) \big) = \big(1, \pa_t \eta(t,x)\big)$ is a tangential vector to $\Gm$, it follows from that $s = \frac{n \cdot \pa_t \eta}{1} = n \cdot \pa_t \eta$, which together with implies . The evolution problem for the level set $\boldsymbol{\Gm^{(\mu)}}$ {#S2.2} ------------------------------------------------------------------ The level set $\Gm = \Gm^{(\mu\la)}$ of $S^{(\mu\la)}$ defined in is determined by an evolution problem for the family of surfaces $t \mapsto \Gm(t)$. To state this evolution problem let ${\cal N}$ be the operator, which assigns the normal speed to the family $t \mapsto \Gm(t)$, i.e. $$s(t,x) = {\cal N}(\Gm)(t,x),$$ with $s(t,x) = s^{(\mu)}(t,x)$ defined by . The evolution problem is given by $$\label{E2.evolution0} {\cal N}(\Gm)(t) = {\cal K}^{(\mu)}\big(\Gm (t)\big),\qquad t_1 \leq t \leq t_2\,,$$ where ${\cal K}^{(\mu)}$ is the non-local evolution operator, which has the form $$\label{E2.evolution1} {\cal K}^{(\mu)}\big(\Gm(t)\big)(x) = s_0 \big({\hat{T}},\ka_\Gm,\la^{1/2} \big)(t,x) + \mu^{1/2} s_1 \big({\hat{u}},{\hat{T}},{\check{T}},S_0,S_1,\la^{1/2} \big)(t,x),$$ for $x \in \Gm(t)$. Here $({\hat{u}},{\hat{T}},{\check{u}},{\check{T}},S_0,S_1)$ is the solution of a transmission-boundary value problem for a coupled system of elliptic partial differential equations and ordinary differential equations, which can be solved recursively. $\ka_\Gm(t,x)$ denotes twice the mean curvature of the surface $\Gm(t)$ at $x \in \Gm(t)$. With the principle curvatures $\kappa_1$, $\kappa_2$ of $\Gm(t)$ at $x \in \Gm(t)$ we thus have $$\kappa_\Gm (t,x)= \kappa_1(t,x) + \kappa_2(t,x).$$ The transmission condition is posed on $\Gm(t)$. Therefore the functions ${\hat{u}}$, ${\hat{T}}$, ${\check{u}}$, ${\check{T}}$ and $S_1$ depend on $\Gm(t)$. We first state and discuss the transmission-boundary value problem. The precise form of the functions $s_0$ and $s_1$ is given in following below. Let ${\hat{S}}:~Q \setminus \Gm \to \{ 0,1\}$ be a piecewise constant function, which only takes the values $0$ and $1$ with a jump across $\Gm$. The sets $$\begin{aligned} \gm &= \{(t,x) \in Q \setminus \Gm \mid {\hat{S}}(t,x) = 0 \},& \gm(t) &= \{ x \in \Om \setminus \Gm(t) \mid (t,x) \in \gm \}, \\ \gm' &= \{(t,x) \in Q \setminus \Gm \mid {\hat{S}}(t,x) = 1 \},& \gm'(t) &= \{ x \in \Om \setminus \Gm(t) \mid (t,x) \in \gm' \}\end{aligned}$$ yield partitions $Q = \gm \cup \Gm \cup \gm'$ and $\Om = \gm(t) \cup \Gm(t) \cup \gm'(t)$ of $Q$ and $\Om$, respectively. If $x$ belongs to $\gm(t)$ or $\gm'(t)$, then the crystal structure at the material point $x$ at time $t$ belongs to phase $1$ or phase $2$, respectively. We assume that the normal vector field $n$ given in is such that the vector $n(t,x)$ points into the set $\gm'(t)$ for every $x \in \Gm(t)$. The transmission-boundary value problem can be separated into two transmission-boundary value problems for the elasticity equations and a boundary value problem for a coupled system of two ordinary differential equations. To state the complete problem we fix $t \in [t_1,t_2]$ and assume that $\Gm(t)$ is known. In the first transmission-boundary problem the unknowns are the displacement $x \mapsto \hat{u}(t,x)\in\R^3$ and the stress tensor $x \mapsto \hat{T}(t,x)\in \ES^3$, which must satisfy the equations $$\begin{aligned} \label{E2.8} -\div_x \hat{T} &=& {\sf b}, \\ \label{E2.9} \hat{T} &=& D\big( \ve(\na_x \hat{u}) - \ov{\ve} \hat{S}\big), \\ \label{E2.10} {[\hat{u}]} &=& 0, \\ \label{E2.11} {[\hat{T}]}n &=& 0, \\ \label{E2.12} {\hat{u}}(t)\rain{\pa\Om} &=& {\sf U}(t), \end{aligned}$$ with ${\sf b}$ and ${\sf U}$ given in and . In the second transmission-boundary problem the unknowns are the displacement $x \mapsto {\check{u}}(t,x)\in \R^3$ and the stress tensor $x \mapsto {\check{T}}(t,x)\in \ES^3$, and the problem is $$\begin{aligned} \label{E2.13} -\div_x {\check{T}}&=& 0, \\ \label{E2.14} {\check{T}}&=& D\Big(\ve(\na_x {\check{u}}) - \ov{\ve}\, \frac{ \hat{T}:\ov{\ve} }{\hat{\psi}''({\hat{S}})} \Big), \\ \label{E2.15} [{\check{u}}] &=& 0, \\ \label{E2.16} {[{\check{T}}]} n &=& 0, \\ \label{E2.17} {\check{u}}(t)\rain{\pa\Om} &=& 0.\end{aligned}$$ The equations , and , must hold on the set $\Om \setminus \Gm(t)$, whereas the equations , and , are posed on $\Gm(t)$. In the boundary value problem for the ordinary differential equations the unknowns are $S_0:\R \to \R$, $S_1: \Gm \ti \R \to \R$ and $s_0: \Gm \ra \R$. We use the notations $S_1'(t,\eta,\zeta) = \pa_\zeta S_1(t,\eta,\zeta)$, $S_1''(t,\eta,\zeta) = \pa^2_\zeta S_1(t,\eta,\zeta)$. In this problem not only $t$, but also $\eta \in \Gm(t)$ is a parameter. For all $\zeta \in \R$ and all values of the parameter $\eta \in \Gm(t)$ the unknowns must satisfy the coupled ordinary differential equations $$\begin{aligned} \hat{\psi}'\big(S_0(\zeta)\big) - S_0''(\zeta) &=& 0, \label{E2.18} \\ \hat{\psi}'' \big(S_0(\zeta)\big) S_1(t,\eta,\zeta) - S_1''(t,\eta,\zeta) &=& F_1(t,\eta,\zeta), \label{E2.19} \end{aligned}$$ and the boundary conditions $$\begin{aligned} S_0(0)=\frac{1}{2},\quad \lim_{\zeta \to -\infty} S_0(\zeta) &= 0, \quad \lim_{\zeta \to \infty} S_0(\zeta) = 1, \label{E2.20} \\ \lim_{\zeta \to -\infty} S_1(t,\eta,\zeta) &= \frac{\overline{\ve}:\hat{T}^{(-)}(t,\eta)}{\hat{\psi}''(0)}, \label{E2.21} \\ \lim_{\zeta \to +\infty} S_1(t,\eta,\zeta) &= \frac{\overline{\ve}:\hat{T}^{(+)}(t,\eta)}{\hat{\psi}''(1)}, \label{E2.22} \\ S_1(t,\eta,0) &= 0, \label{E2.23} \end{aligned}$$ with the right hand side of given by $$\label{E2.24} \begin{split} F_1(t,\eta,\zeta)= \ov{\ve}: \Big(&[{\hat{T}}] (t,\eta) S_0(\zeta) + {\hat{T}}^{(-)}(t,\eta) \Big) \\ & + \Big(\frac{s_0(t,\eta)}{c} - \la^{1/2} \ka_\Gm (t,\eta) \Big) S'_0(\zeta), \end{split}$$ where the constant $c > 0$ is given in . The linear elliptic system , differs from the standard elasticity system only by the term $-D\ov{\ve} \hat{S}$. This term is known since $\Gm(t)$ is given. Under suitable regularity assumptions for the given functions ${\sf b}$ and ${\sf U}$ and very mild assumptions on the regularity of the interface $\Gm(t)$ the problem has a unique weak solution $({\hat{u}},{\hat{T}})$. This can be proved by standard methods from functional analysis. Of course, the regularity of the solution depends on the regularity of ${\sf b, \ U}$ and $\Gm(t)$. After insertion of the stress tensor ${\hat{T}}$ from this solution into , the equations – form a transmission-boundary value problem of the same type as – , with unique solution $({\check{u}},{\check{T}})$ determined by the same methods. We also insert ${\hat{T}}$ into , and , which determines the right hand side of the differential equation and the boundary conditions , posed at $\pm \infty$. The nonlinear differential equation has a unique solution $S_0$ satisfying the boundary conditions . By insertion of $S_0$ into and , equation becomes a linear differential equation for $S_1$, however with an additional unknown function $s_0$ in the right hand side. This function is constant with respect to $\zeta$. We sketch here the procedure used to determine $s_0$. This procedure is standard in investigations of the asymptotics of phase field models: The second order differential operator $\big(\hat{\psi}''(S_0) - \pa_\zeta^2\big)$ is selfadjoint in the Hilbert space $L^2(\R)$ with a one dimensional kernel spanned by the function $S_0'$. This is seen by differentiating the equation . From functional analysis we thus know that for $F_1 \in L^2(\R)$ the differential equation $\big(\hat{\psi}''(S_0) - \pa_\zeta^2\big)w = F_1$ has a solution $w \in L^2(\R)$ if and only if the orthogonality condition $$\label{E2.25} \int_{-\infty}^\infty F_1(t,\eta,\zeta) S_0'(\zeta)\,d\zeta = 0$$ holds. It turns out that though the function $F_1$ defined in does not in general belong to $L^2(\R)$ and the solution $S_1(t,\eta,\cdot)$ is not sought in $L^2(\R)$, which is seen from the boundary conditions , , the orthogonality condition is sufficient for the solution $S_1$ to exist. Comparison with shows that can be satisfied by choosing the constant $s_0(t,\eta)$ suitably. This defines the function $s_0:\Gm \ra \R$ uniquely. Since $F_1$ depends on ${\hat{T}}$, $\ka_\Gm$ and $\la^{1/2}$, it follows that also $s_0$ is a function of these variables: $$s_0(t,\eta) = s_0\big({\hat{T}},\ka_\Gm,\la^{1/2} \big)(t,x).$$ The explicit expression for $s_0$ obtained in this way is stated below in . In fact, $s_0({\hat{T}},\ka_\Gm,\la^{1/2})$ is the first term on the right hand side in the expression for ${\cal K}^{(\mu)}\big(\Gm(t) \big)$. The procedure sketched here is discussed precisely in Section \[S4.2\] when we determine the second term $s_1$ in , which is obtained from a similar, but more complicated boundary value problem. The asymptotic solution and the kinetic relation {#S2.3} ------------------------------------------------ To state the properties of the asymptotic solution and the kinetic relation in , we introduce some definitions. We need in our investigations that the second derivatives $\hat{\psi}''(0)$ and $\hat{\psi}''(1)$ of the double well potential at the minima $0$ and $1$ are positive, and we set $$a = \min\Big\{ \sqrt{\hat{\psi}''(0)},\sqrt{\hat{\psi}''(1)} \Big\}.$$ Depending on the parameters $\la$ and $\mu$, we partition $Q$ into the inner neighborhood $Q_{\rm inn}^{(\mu\la)}$ of $\Gm$, into the matching region $Q_{\rm match}^{(\mu\la)}$ and into the outer region $Q_{\rm out}^{(\mu\la)}$. These sets are defined by $$\label{E2.25a} \begin{split} Q_{\rm inn}^{(\mu\la)} &= \Big\{ (t,\eta,\xi) \in {\cal U}_\da \Bigm\| \|\xi\| < \frac32 \frac{ (\mu\la)^{1/2} \|\ln\mu\|}{a} \Big\}, \\ Q_{\rm match}^{(\mu\la)} &= \Big\{ (t,\eta,\xi) \in {\cal U}_\da \Bigm\| \frac32 \frac{ (\mu\la)^{1/2} \|\ln\mu\|}{a} \leq \|\xi\| \leq \frac{3 (\mu\la)^{1/2} \|\ln\mu\|}{a} \Big\}, \\[1ex] Q_{\rm out}^{(\mu\la)} &= Q \setminus \big( Q_{\rm inn}^{(\mu\la)} \cup Q_{\rm match}^{(\mu\la)} \big). \end{split}$$ We always assume that the parameters $\la$ and $\mu$ satisfy $0 < \la \leq \la_0$ and $0 < \mu \leq \mu_0$, where $\la_0$, $\mu_0$ are fixed constants satisfying $$\mu_0 \leq e^{-2}, \qquad \frac{3 (\mu_0\la_0)^{1/2} \|\ln \mu_0\|}{a} < \da.$$ The first condition is imposed for purely technical reasons and guarantees that the function $\mu \mapsto \mu^{1/2} \|\ln \mu\|$ is increasing, the second condition guarantees that $Q_{\rm inn}^{(\mu\la)},Q_{\rm match}^{(\mu\la)} \subset {\cal U}_\da$ and that $Q_{\rm out}^{(\mu\la)} \cap {\cal U}_\da $ is a nonempty, relatively open subset of ${\cal U}_\da$. By , the function $\hat{u}:Q \to \R^3$ is continuous at every point $(t,\eta) \in \Gamma$, but the first and higher derivatives of $\hat{u}$ in the direction of the normal vector $n(t,x)$ can jump across $\Gamma$. For these jumps we write $$\begin{aligned} u^\ast(t,\eta) &=& [\pa_n \hat{u}](t,\eta),\label{E2.26} \\ a^\ast (t,\eta) &=& [\pa_n^2 \hat{u}](t,\eta). \label{E2.27}\end{aligned}$$ We also set $$\label{E2.28} c_1 = \int_0^1 \sqrt{2\hat{\psi}(\vartheta)}\, d\vartheta.$$ \[T2.3\] Suppose that the double well potential $\hat{\psi} \in C^5(\R)$ satisfies $$\label{E2.29} \begin{split} \hat{\psi}(r)>0,\quad \mbox{for } 0 < r < 1, \\ \hat{\psi}(r)=\hat{\psi}'(r)= 0,\quad \mbox{for } r=0,1, \\ a = \min\Big\{ \sqrt{\hat{\psi}''(0)},\sqrt{\hat{\psi}''(1)} \Big\} > 0. \end{split}$$ Moreover, suppose that $\hat{\psi}$ satisfies the symmetry condition $$\label{E2.30} \hat{\psi}(\frac12 - \zeta) = \hat{\psi}(\frac12 + \zeta),\quad \zeta \in \R.$$ Assume that there is a solution $\Gm$ of the evolution problem , with $s_0 = s_0({\hat{T}},\ka_\Gm,\la^{1/2} ):\Gm \ra \R$ given by $$\label{E2.32} s_0 = \frac{c}{c_1} \Big( - \ov{\ve}:\langle \hat{T} \rangle + \la^{1/2} c_1 \ka_\Gm \Big),$$ and with $s_1 = s_1({\hat{u}},{\hat{T}},{\check{T}},S_0,S_1,\la^{1/2}):\Gm \ra \R$ defined by $$\label{E2.32s1} s_1 = s_{10} + \la^{1/2} s_{11} = s_{10}({\hat{T}},{\check{T}},S_0,S_1) + \la^{1/2} s_{11} ({\hat{u}},S_0),$$ where $$\begin{aligned} s_{10} &=& \frac{c}{c_1} \Biggl( - \ov{\ve}:\langle {\check{T}}\rangle + \ov{\ve}: [{\hat{T}}] \Big( \Big\langle \frac{\ov{\ve}:{\hat{T}}}{ \hat{\psi}''({\hat{S}})} \Big\rangle - \int_{-\infty}^\infty S_1 S'_0\, d\zeta \Big) \nn \\ && \hspace{8ex} \mbox{} + \frac{1}{c_1} \ov{\ve}: \langle {\hat{T}}\rangle \int_{-\infty}^\infty S_1' S_0' \,d\zeta + \frac12 \int_{-\infty}^\infty \hat{\psi}'''(S_0) S_1^2 S_0' \,d\zeta \Biggr), \label{E2.33} \\ s_{11} &=& - \frac{c}{c_1}\, \ov{\ve}:D \ve(a^* \otimes n + \na_\Gm u^) \int_{-\infty}^\infty S_0(\zeta) S_0(-\zeta)\, d\zeta. \label{E2.34}\end{aligned}$$ In and , we have $S_0 = S_0(\zeta)$ and $S_1 = S_1(t,\eta,\zeta)$, for all other functions the argument is $(t,\eta)$. The positive constant $c$ is defined in . The notations $[\cdot]$ and $\langle \cdot \rangle$ are introduced in Section \[S2.1\]. In particular, we have $$\Big\langle \frac{ \ov{\ve}:{\hat{T}}} { \hat{\psi}''({\hat{S}})} \Big\rangle = \frac12 \Big(\frac{\ov{\ve}:{\hat{T}}^{(+)}} {\hat{\psi}''(1)} + \frac{\ov{\ve}:{\hat{T}}^{(-)}} {\hat{\psi}''(0)} \Big).$$ With these functions the normal speed $s(t,\eta)$ of $\Gm(t)$ at $\eta \in \Gm(t)$ is thus given by $$\label{E2.31} s(t,\eta) = s_0 (t,\eta) + \mu^{1/2}s_1(t,\eta,\la^{1/2}) = s_0 (t,\eta) + \mu^{1/2} \big( s_{10}(t,\eta) + \la^{1/2} s_{11}(t,\eta) \big). $$ We assume moreover that the solution $\Gm$ is a $C^5$–manifold and that the functions ${\hat{u}}$ and ${\check{u}}$ defined by the evolution problem satisfy ${\hat{u}}\in C^4(\gm \cup \gm',\R^3)$, ${\check{u}}\in C^3(\gm \cup \gm',\R^3)$ and that ${\hat{u}}$ has $C^4$–extensions, ${\check{u}}$ has $C^3$–extensions from $\gm$ to $\gm \cup \Gm$ and from $\gm'$ to $\gm' \cup \Gm$. For the given right hand side of we assume that ${\sf b} \in C^1(\ov{Q})$. Under these assumptions there is an approximate solution $(u^{(\mu)},T^{(\mu)},S^{(\mu)})$ of the Allen-Cahn model – , for which $\Gm$ is the level set $$\label{E2.36} \Gm = \Big\{ (t,x) \in Q \bigm\| S^{(\mu)}(t,x) = \frac12 \Big\},$$ and which satisfies the equations $$\begin{gathered} -\div_x T^{(\mu)} = {\sf b} + f_1^{(\mu\la)}, \label{E2.36a} \\ T^{(\mu)} = D\big(\ve( \na_x u^{(\mu)}) - \ov{\ve}S^{(\mu)}\big), \label{E2.36b} \\ \pa_t S^{(\mu)} + \frac{c}{(\mu\la)^\frac12} \Big( \pa_S {\sf W} \big( \ve(\na_x u^{(\mu)}), S^{(\mu)} \big) + \frac{1}{\mu^\frac12} \hat{\psi}'(S^{(\mu)}) - \mu^\frac12\la\Da_x S^{(\mu)}\Big) = f_2^{(\mu\la)}, \label{E2.36c} \\ u^{(\mu)}(t,x) = {\sf U} (t, x), \qquad (t,x) \in [t_1,t_2]\ti \pa\Om, \label{E2.36d} \\ \pa_{n_{\pa\Om}} S^{(\mu)}(t,x) = f_3^{(\mu\la)}, \qquad (t,x) \in [t_1,t_2] \ti \pa\Om, \label{E2.36e}\end{gathered}$$ where to the right hand sides $f_1^{(\mu\la)}$,…, $f_3^{(\mu\la)}$ there exist nonnegative constants $K_1,\ldots, K_5$ such that for all $\mu \in (0,\mu_0]$ and all $\la \in (0,\la_0]$ $$\begin{aligned} \\| f_1^{(\mu\la)} \\|_{L^\infty(Q_{\rm inn}^{(\mu\la)} \cup Q_{\rm match}^{(\mu\la)})} &\leq& \|\ln \mu\|^2\, \Big( \frac{\mu}{\la} \Big)^\frac12 K_1\, , \label{E2.37a} \\ \\| f_1^{(\mu\la)} \\|_{L^\infty(Q_{\rm out}^{(\mu\la)})} &\leq& \mu^\frac32 K_2\,, \label{E2.37b} \\ \\| f_2^{(\mu\la)} \\|_{L^\infty(Q_{\rm inn}^{(\mu\la)} \cup Q_{\rm match}^{(\mu\la)})} &\leq& \|\ln \mu\|^2 \Big( \frac{\mu}{\la} \Big)^\frac12 K_3\,, \label{E2.38a} \\ \\| f_2^{(\mu\la)} \\|_{L^\infty(Q^{(\mu\la)}_{\rm out})} &\leq& \frac{\mu}{\la^{1/2}} K_4\,, \label{E2.38b} \\ \\| f_3^{(\mu\la)} \\|_{L^\infty(\pa\Om)} &\leq& \mu^\frac12 K_5\,, \label{E2.38c}\end{aligned}$$ In the neighborhood $Q_{\rm inn}^{(\mu\la)}$ of $\Gm$ the order parameter in the approximate solution is of the form $$\label{E2.39} S^{(\mu)}(t,x)=S_0\Big(\frac{\xi}{(\mu\lambda)^{1/2}}\Big)+ \mu^{1/2} S_1 \Big(t,\eta, \frac{\xi}{(\mu\lambda)^{1/2}} \Big) + \mu S_2 \Big(t,\eta, \frac{\xi}{(\mu\lambda)^{1/2}}\Big),$$ where the monotonically increasing transition profile $S_0: \R \to \R$ and the function $S_1: \Gm \ti \R \ra \R$ are given as solution of the coupled problem – , and where $S_2: \Gm \ti \R \ra \R$ satisfyies $S_2(t,\eta,0) = 0$ and $$\label{E2.40} \|S_2(t,\eta,\zeta)\| \leq C (1 + \|\zeta\|),\qquad \mbox{for } (t,\eta,\zeta) \in \Gm \ti \R,$$ with a constant $C$ independent of $(t,\eta,\zeta)$. We mention that the positive constant $c$ in does not play a major role in the analysis and could be replaced by $1$. We refrain from replacing it to show how $c$ appears in the kinetic relation. The [proof]{} of this theorem forms the content of Sections \[S3\] – \[S5\]. We remark that the symmetry assumption for the double well potential $\hat{\psi}$ serves to simplify the computations in the derivation of the asymptotic solution. Without this assumption the term $s_1$ in the kinetic relation would contain other terms in addition to the terms $s_{10}$ and $s_{11}$ given in an . The regularity properties of $\Gm$ and of ${\hat{u}}$, ${\check{u}}$ are of course not independent, since ${\hat{u}}$ and ${\check{u}}$ are solutions of the elliptic transmission problems – and – , respectively. Therefore the regularity theory of elliptic equations shows that ${\hat{u}}$ and ${\check{u}}$ automatically have the differentiability properties assumed in the theorem if the manifold $\Gm$ and the right hand side ${\sf b}$ are sufficiently smooth.\ Since by definition of $Q^{(\mu\la)}_{\rm inn}$ and $Q^{(\mu\la)}_{\rm match}$ in we have $${\rm meas}(Q^{(\mu\la)}_{\rm inn} \cup Q^{(\mu\la)}_{\rm match}) \leq C_3 (\mu\la)^{1/2} \|\ln \mu \|,$$ we immediately obtain from – the following \[C2.2\] There are constants $K_6$, $K_7$ such that for all $0 < \mu \leq \mu_0$ and all $0 < \la \leq \la_0$ $$\begin{aligned} \\| f_1^{(\mu\la)}\\|_{L^1(Q)} &\leq& \|\ln \mu\|^3 \mu K_6 \,, \label{E2.37} \\[1ex] \Big\\| f_2^{(\mu\la)} \\|_{L^1(Q)} &\leq& \frac{\|\ln \mu\|^3 \, \mu}{\la^{1/2}} K_7 \,. \label{E2.38}\end{aligned}$$ The leading term $s_0$ given in can be written in a more common and more general form. To give this form, we need a result on the jump of the Eshelby tensor. The Eshelby tensor to the solution $({\hat{u}},{\hat{T}})$ of the transmission problem – is defined by $$\label{E2.45} \hat{C} (\na_x \hat{u}, \hat{S})= \psi_\mu \big(\ve(\na_x \hat{u}), \hat{S}\big) I-(I+\na_x \hat{u})^T \hat{T},$$ where $I \in \ES^3$ is the unit matrix and where $$\label{E2.46} \psi_\mu (\ve,S)= {\sf W}(\ve, S)+ \frac{1}{\mu^{1/2}} \hat{\psi} (S)$$ is that part of the free energy $\psi^_\mu$ defined in without gradient term. The last term on the right hand side of is a matrix product. We use the standard convention to denote the matrix product of two matrices $A \in \R^{k \times m}$ and $B \in \R^{m \times \ell}$ by $A B \in \R^{k \ti \ell}$. \[L2.3\] Let $(\hat{u}, \hat{T})$ be the solution of the transmission problem – and let $n$ be a unit normal vector field to $\Gm(t)$. Then the jump $[\hat{C}]$ of the Eshelby tensor to $({\hat{u}},{\hat{T}})$ across $\Gamma$ satisfies $$\label{E2.44} n \cdot [\hat{C}]n = \frac{1}{\mu^{1/2}}\, [\hat{\psi}({\hat{S}})] - \ov\ve:\langle \hat{T} \rangle.$$ This result is known [@AK90]. In [@CMT2011] it is stated as equation (3.4) and proved on pages 154, 155. \[C2.4\] The leading term $s_0$ of the kinetic relation defined in satisfies $$\label{E2.55} s_0 = \frac{c}{c_1} \Big( n \cdot[\hat{C}]n + \lambda^{1/2} c_1 \kappa_\Gamma\Big).$$ This corollary follows immediately from , since by assumption we have $[\hat{\psi}(\hat{S})] = \hat{\psi}(1)- \hat{\psi}(0)=0$, which implies that $n \cdot[\hat{C}]n=-\ov\ve : \langle \hat{T}\rangle$. Consequences for numerical simulations {#S2.4} -------------------------------------- In this section we discuss the consequences of for numerical simulations of interfaces with small interface energy. In many functional materials the phase interfaces consist only of a few atomic layers. For interfaces with such small width mathematical models with sharp interface are appropriate. We therefore base the following considerations on the hypothesis that the propagation speed of the interface in the sharp interface model is a good approximation to the propagation speed of the interface in the real material. The model error of the Allen-Cahn model is then the difference of the propagation speed of the sharp interface and the propagation speed of the diffuse interface in the phase field model. The parameters $\mu$ and $\la$ in the Allen-Cahn model should be chosen such that this model error is small and such that numerical simulations based on the Allen-Cahn model are effective. To make this precise we must first determine the sharp interface model to be used. The model consists of the transmission problem – combined with a kinetic relation. To find this relation, one proceeds in the usual way and uses that by the second law of thermodynamics the Clausius-Duhem inequality $$\pa_t \psi_{\rm sharp} + \div_x\, q_{\rm sharp} \leq {\hat{u}}_t \cdot {\sf b}$$ must be satisfied to impose restrictions on the form of the kinetic relation. Here $\psi_{\rm sharp}$ denotes the free energy in the sharp interface problem and $q_{\rm sharp}$ is the flux of the free energy. We use the standard free energy and flux $$\begin{aligned} \psi_{\rm sharp}\big(\ve(\na_x \hat{u}), \hat{S}\big) &= {\sf W} \big(\ve(\na_x \hat{u}), \hat{S}\big) + \lambda^{1/2} c_1 \int_{\Gamma(t)}\,{\rm d} \sigma, \label{E2.56} \\ q_{\rm sharp} ( {\hat{T}}, {\hat{S}}) &= - {\hat{T}}\cdot {\hat{u}}_t \,. \nn \end{aligned}$$ The last term on the right hand side of is the interface energy, hence $\la^{1/2} c_1$ is the interface energy density. It is well known that if $\big({\hat{u}}(t),{\hat{T}}(t)\big)$ is a solution of the transmission problem – at time $t$ and if the interface $\Gm(t)$ in this problem moves with the given normal speed $s_{\rm sharp}(t,x)$ at $x \in \Gm(t)$, then the Clausius-Duhem inequality holds if and only if the inequality $$\label{E2.cdequiv} s_{\rm sharp}(t,x) \Big( n(t,x) \cdot [\hat{C}](t,x)n(t,x) + \la^{1/2} c_1 \ka_\Gm(t,x) \Big) \geq 0$$ is satisfied at every point $x \in \Gm(t)$. A proof of this well known result is given in [@Al00], however only for the case where $\la = 0$ in . The proof can be readily generalized to the case $\la > 0$. A simple linear kinetic relation, for which obviously holds, is $$\label{E2.kinrelsharp} s_{\rm sharp} = \frac{c}{c_1}\, \big( n \cdot [\hat{C}]n + \la^{1/2} c_1 \ka_\Gm \big).$$ The sharp interface problem thus consists of the transmission problem – combined with the kinetic relation . For this problem the Clausius-Duhem inequality is satisfied. We can now define the model error. To this end let $\big(u^{(\mu)}, T^{(\mu)}, S^{(\mu)} \big)$ be the asymptotic solution in the domain $Q = [t_1,t_2] \ti \Om$ constructed in , where by , the manifold $\Gm$ is the level set $\{ S^{(\mu)}=\frac12 \}$. Let $\hat{t}\in [t_1,t_2]$ be a fixed time and let$\big( u_{\rm AC}^{(\mu)}, T_{\rm AC}^{(\mu)}, S_{\rm AC}^{(\mu)} \big)$ be the exact solution of the Allen-Cahn model – in the domain $[\hat{t},t_2] \ti \Om$, which satisfies the boundary and initial conditions $$\begin{aligned} u^{(\mu)}_{\rm AC}(t,x) &=& {\sf U}(t,x), \qquad (t,x) \in [\hat{t},t_2] \ti \pa\Om, \label{E2.BC1AC} \\ \pa_{n_{\pa\Om}} S^{(\mu)}_{\rm AC}(t,x) &=& f_3^{(\mu\la)}(t,x), \qquad (t,x) \in [\hat{t},t_2] \ti \pa\Om, \label{E2.BC2AC} \\ S_{\rm AC}^{(\mu)}(\hat{t},x) &=& S^{(\mu)} (\hat{t},x), \quad x \in \Om, \label{E2.ICAC}\end{aligned}$$ where $f_3^{(\mu\la)}$ is the right hand side of . The level set of the order parameter $S_{\rm AC}^{(\mu)}$ is denoted by $$\label{E2.GmAC} \Gamma_{\rm AC} = \Big\{ (t,x) \in Q \Bigm\| S_{\rm AC}^{(\mu)}(t,x)=\frac{1}{2}\Big\}.$$ Let $\Gm_{\rm sharp} \subseteq Q$ be the sharp interface in the solution of the sharp interface problem – , , which satisfies the initial condition $$\label{E2.ICsharp} \Gm_{\rm sharp}(\hat{t}) = \Gm(\hat{t}).$$ The normal speeds of the different surfaces are $$s = s^{(\mu\la)} = \EN\big(\Gamma^{(\mu\la)}\big), \qquad s_{\rm AC} = s_{\rm AC}^{(\mu\la)}= \EN\big(\Gamma_{\rm AC}^{(\mu\la)}\big), \qquad s_{\rm sharp} = \EN\big(\Gm_{\rm sharp} \big),$$ where $\EN$ is the normal speed operator introduced at the beginning of Section \[S2.2\]. Of course, $s_{\rm sharp}$ is given by . Note that the functions $s^{(\mu\la)}(\hat{t})$, $s_{\rm AC}^{(\mu\la)}(\hat{t})$, $s_{\rm sharp}(\hat{t})$ are defined on the same set, since the initial condition and together imply $\Gm_{\rm AC}(\hat{t}) = \Gm(\hat{t}) = \Gm_{\rm sharp}(\hat{t})$. \[D2.7\] We call the function ${\cal E}= {\cal E}^{(\mu\la)}(\hat{t}):\Gm(\hat{t}) \ra \R$ defined by $$\label{E2.modelerror} {\cal E} = s_{\rm AC}(\hat{t}) - s_{\rm sharp}(\hat{t})$$ the model error of the Allen-Cahn model at time $\hat{t}$ to the parameters $\mu$ and $\la$. We can now discuss the choice of the parameters $\mu$ and $\la$. Since coincides with the leading term $s_0$ in the asymptotic expansion of the kinetic relation of the Allen-Cahn model, which is seen from , we have $$\label{E2.ssharp} s_{\rm sharp} = s_0.$$ Therefore yields $$\label{E2.decomperror} {\cal E} = s_{\rm AC} - s_{\rm sharp} = s_{\rm AC} - s_0 = ( s_{\rm AC} - s ) + ( s - s_0 ) = ( s_{\rm AC} - s ) + \mu^{1/2}(s_{10} + \la^{1/2} s_{11}) .$$ The difference $ s_{\rm AC} - s$ between the propagation speeds of the exact solution and the asymptotic solution tends to zero for $\mu \ra 0$ faster than the term $\mu^{1/2}s_{10}$, and the convergence is uniform with respect to $\la$. This is the basic result, which allows to discuss the optimal choice of $\mu$ and $\la$. The precise result is \[T2.8\] There is a constant $C_{\cal E} > 0$ such that for all $0 < \mu \leq \mu_0$ and all $0 < \la \leq \la_0$ we have the estimate $$\label{E2.assumpA} \\|s_{\rm AC}(\hat{t}) - s(\hat{t})\\|_{L^2(\Gm(\hat{t}))} \leq C_{\cal E} \| \ln \mu \|^3 \mu.$$ The [proof]{} of this theorem is given in Section \[S6\].\ and together yield $$\label{E2.esterror} \\| {\cal E}^{(\mu\la)} \\|_{L^2(\Gm(\hat{t}))} \leq C \mu^{1/2},$$ with a constant $C$, which can be chosen independently of $\la$. By this inequality, $\mu^{1/2}$ controls the model error. Therefore we write $F = \mu^{1/2}$ and call $F$ the error parameter. Moreover, since $\la^{1/2} c_1$ is the interface energy density, we call $E = \la^{1/2}$ the interface energy parameter. Also, since by the interface width is proportional to $(\mu\la)^{1/2}$, we call $B = (\mu\la)^{1/2}$ the interface width parameter. These three parameters and the propagation speed $s_{\rm AC}$ are connected by the fundamental relations $$\begin{gathered} B = EF, \label{E2.fundeqnAC1} \\ s_{\rm AC} = \frac{c}{c_1} n \cdot [\hat{C}] n + c \ka_\Gm E + {\cal E}[E,F], \label{E2.fundeqnAC2} \\ \\|{\cal E}[E,F]\\|_{L^2(\Gm(\hat{t}))} \leq C F, \label{E2.fundeqnAC3} \end{gathered}$$ where we use the notation ${\cal E}[E,F] = {\cal E}^{(\mu\la)}$. The first equation is an immediate consequence of the definition of the parameters, the second is obtained by insertion of into , and the last inequality is just a restatement of . Now assume that we want to use a phase field model to numerically simulate the propagation of a phase interface. In such a simulation the numerical effort is proportional to $h^{-p}$, where $h$ denotes the grid spacing and where the power $p > 1$ depends on whether we want to simulate a problem in $2$–d or in $3$–d and it depends on the numerical scheme we use. In order for the simulation to be precise, we must guarantee that the model error and the numerical error are small. To make the numerical error small, we must choose the grid spacing $h$ small enough to resolve the transition of the order parameter across the interface, which means that we must choose $h < B$, hence we have $h^{-p} > B^{-p}$. Therefore we see that the numerical effort of a simulation based on a phase field model is measured by the number $B^{-p}$. We call the number $$e_{\rm num} = B^{-p}$$ the parameter of numerical effort. For a simulation based on the Allen-Cahn model we see from that the numerical effort is $$e_{\rm num} = (EF)^{-p}.$$ Assume that the interface, which we want to simulate with the Allen-Cahn model, has very small interface energy density. Such interfaces are common in metallic or functional materials. For such materials the interface energy parameter $E$ is small. To make the model error small, we must also choose the error parameter $F$ small, which means that the numerical effort parameter $e_{\rm num} = (EF)^{-p}$ is very large as a product of two large numbers $E^{-p}$ and $F^{-p}$. To be more specific, we consider an interface without interface energy, which means that the free energy $\psi_{\rm sharp}$ does not contain the last term on the right hand side of . From we see that the propagation speed of the sharp interface with zero interface energy density is $$s_{\rm sharp} = \frac{c}{c_1} n \cdot [\hat{C}]n.$$ From this equation and from we see that in this case the total model error, which we denote by ${\cal E}_{\rm total}$, is $${\cal E}_{\rm total} = s_{\rm AC} - s_{\rm sharp} = c\ka_\Gm E + {\cal E}[E,F] .$$ This means that the term $c\ka_\Gm E$ is now part of the total model error. This term does not vanish identically, since we cannot set $\la = 0$ in the Allen-Cahn equation . Instead the values of $\la$ and of $E = \la^{1/2}$ must be positive. If we prescribe the $L^2$-norm ${\cal E}_{L^2} = \\|{\cal E}_{\rm total}\\|_{L^2(\Gm(\hat{t}))}$ of the total model error, we must therefore choose the parameters $E$ and $F$ such that $$\begin{aligned} c \\|\ka_\Gm\\|_{L^2(\Gm(\hat{t}))} E + \\|{\cal E}[E,F]\\|_{L^2(\Gm(\hat{t}))} &\leq& {\cal E}_{L^2}, \label{E2.opti1} \\ EF &\overset{!}{=}& \max, \label{E2.opti2}\end{aligned}$$ where the second condition is imposed by the requirement to make the numerical effort $e_{\rm num} = (EF)^{-p}$ as small as possible. To discuss this optimization problem, we assume first that the term $s_{10}$ in the asymptotic expansion of the kinetic relation of the Allen-Cahn model is not identically equal to zero. In this case we conclude from and by the inverse triangle inequality that for sufficiently small $\la^{1/2} = E$ and for sufficiently small $\mu^{1/2} = F$ $$\begin{gathered} \\|{\cal E}[E,F]\\|_{L^2(\Gm(\hat{t}))} = \\| \mu^{1/2} s_{10} + (\mu\la)^{1/2} s_{11} + (s_{\rm AC} -s) \\|_{L^2(\Gm(\hat{t}))} \\ \geq \mu^{1/2} \\|s_{10}\\|_{L^2(\Gm(\hat{t}))} - (\mu\la)^{1/2} \\|s_{11}\\|_{L^2(\Gm(\hat{t}))} - \\|s_{\rm AC} -s \\|_{L^2(\Gm(\hat{t}))} \\ \geq \mu^{1/2} \big( \\|s_{10} \\|_{L^2(\Gm(\hat{t}))} - \la^{1/2} \\|s_{11}\\|_{L^2(\Gm(\hat{t}))} - C_{\cal E} \|\ln \mu\|^3 \mu^{1/2} \big) \\ \geq \mu^{1/2} \Big( \\| s_{10} \\|_{L^2(\Gm(\hat{t}))} - \frac12 \\|s_{10} \\|_{L^2(\Gm(\hat{t}))} \Big) = \frac12 \\|s_{10}\\|_{L^2(\Gm(\hat{t}))} F. \end{gathered}$$ This inequality and imply that the solution $(E,F)$ of the optimization problem , satisfies $$F \leq \frac{2}{\\|s_{10}\\|_{L^2(\Gm(\hat{t}))}} \\|{\cal E}[E,F]\\|_{L^2(\Gm(\hat{t}))} \leq \frac{2}{\\|s_{10}\\|_{L^2(\Gm(\hat{t}))}}\, {\cal E}_{L^2} \qquad \mbox{and} \qquad E \leq \frac{1}{c \\|\ka_\Gm\\|_{L^2(\Gm(\hat{t}))}}\, {\cal E}_{L^2} .$$ From this result we obtain \[C2.5\] Let ${\cal E}_{\rm max}$ denote the total model error of the Allen-Cahn model in the simulation of an interface without interface energy. If the term $s_{10}$ in the asymptotic expansion of the kinetic relation of the Allen-Cahn model is not identically equal to zero, then the interface width $B$ satisfies $$\label{E2.withErr} B = EF \leq \frac{2}{c \\|s_{10}\\|_{L^2(\Gm(\hat{t}))}\\|\ka_\Gm\\|_{L^2(\Gm(\hat{t}))}}\, {\cal E}_{L^2}^2\,.$$ In a numerical simulation of an interface without interface energy the parameter of numerical effort satisfies $$\label{E2.numeff} e_{\rm num} \geq \left( \frac{c \\|s_{10}\\|_{L^2(\Gm(\hat{t}))} \\|\ka_\Gm\\|_{L^2(\Gm(\hat{t}))} }{ 2\,{\cal E}_{L^2}^2} \right)^p$$ with a power $p > 1$ depending on the space dimension and the numerical method used. The interface width thus decreases with the square of the model error. Since the time step in a simulation must be decreased when the grid spacing $h$ in $x$–direction is decreased, the number $p$ can be larger than $4$ in a three dimensional simulation. From we thus see that the numerical effort grows very rapidely when the required accuracy is increased. The Allen-Cahn model is therefore ineffective when used to accurately simulate interfaces with low interface energy. If the term $s_{10}$ vanishes identically, then the same considerations show that instead of and we would have $B = O({\cal E}_{L^2}^{3/2})$ and $e_{\rm num} \geq C {\cal E}_{L^2}^{-\frac32 p}$. The numerical effort would still grow fast when the required accuracy is increased, though less fast than for $s_{10} \neq 0$. However, a close investigation of the terms in the definition of $s_{10}$, which we do not present here, shows that only in very exceptional situations one can expect that $s_{10}$ vanishes identically. In we assumed that the mesh is globally refined. Of course, one can improve the effectivity of simulations by using local mesh refinement in the neighborhood of the interface. We do not discuss this question of numerical analysis here, but in fact shows that adaptive mesh refinement and other advanced numerical techniques are needed to make precise simulations of interfaces with small energy based on the Allen-Cahn model effective. #### Comparison to the hybrid phase field model With we can refine the comparison given in [@JElast2012] of the Allen-Cahn model and another phase field model, which we call the hybrid model. The hybrid model was introduced and discussed in [@Al00; @Al04; @AlPZ06; @AlPZ07; @JElast2012]. By formal construction of asymptotic solutions we showed in [@JElast2012] the following result:\ Let $B_{\rm AC}({\cal E}_{L^2})$ and $B_{\rm hyb}({\cal E}_{L^2})$ be the interface widths in the Allen-Cahn model and the hybrid model, respectively, which result when the model parameters are adjusted to model an interface without interface energy with the total model error ${\cal E}_{L^2}$. Then we have for ${\cal E}_{L^2} \ra 0$ that $$B_{\rm hyb}({\cal E}_{L^2}) = O({\cal E}_{L^2}), \qquad B_{AC}({\cal E}_{L^2}) = o(1)O({\cal E}_{L^2}) = o({\cal E}_{L^2}).$$ The Landau symbol $o(1)$ denotes terms, which tend to zero for ${\cal E}_{\rm max} \ra 0$. This result was obtained under the assumption that estimates corresponding to hold true for both models, without proving this assumption. To achieve a prescribed small value ${\cal E}_{L^2}$ of the total model error we must therefore choose the interface width in the Allen-Cahn model smaller than in the hybrid model. Consequently, the hybrid model is numerically more effective, but how much more depends on the rate of decay of the $o({\cal E}_{L^2})$ term in the result for the Allen-Cahn model. In [@JElast2012] we could not determine this decay rate, since the asymptotic solution constructed in [@JElast2012] for the Allen-Cahn model was only of first order. yields this decay rate. From the result for the hybrid model and from we thus obtain for the parameters $e_{\rm num}^{\rm hyb}$ and $e_{\rm num}^{\rm AC}$ of the hybrid model and the Allen-Cahn model, respectively, that $$e_{\rm num}^{\rm hyb} \leq C {\cal E}_{L^2}^{-p}, \qquad e_{\rm num}^{\rm AC} \geq C {\cal E}_{L^2}^{-2p},$$ which shows that when the prescribed error ${\cal E}_{L^2}$ is small, the hybrid model can be quite considerably more effectice in numerical simulations of interfaces with low interface energy or no interface energy than the Allen-Cahn model. The jump of solutions of the transmission problems {#S2.5} --------------------------------------------------- In this section we prove some results on the jumps of the solutions $({\hat{u}},{\hat{T}})$ and $({\check{u}},{\check{T}})$ of the transmission problems – and – , which we need in the following sections. We define a scalar product $\alpha:_D \beta$ on $\ES^3$ by $\alpha:_D \beta=\alpha:(D\beta)$, for $\alpha, \beta \in \ES^3$. For a unit vector $n \in \R^3$ let a linear subspace of $\ES^3$ be given by $$\label{E2.41} \ES_n^3= \Big\{\frac{1}{2} (\om \otimes n+n\otimes \om) \,\big\vert\, \om \in \R^3\Big\},$$ let $P_n:\ES^3 \to \ES^3$ be the projector onto $\ES_n^3$, which is orthogonal with respect to the scalar product $\alpha:_D \beta$ and let $Q_n=I-P_n$. \[L2.6\] Let $\om^* \in \R^3$ be a vector. This vector satisfies $\Big( D\big( \ve(\om^* \otimes n) - \ov{\ve} \big) \Big) n = 0$ if and only if $\ve(\om^* \otimes n) = P_n \ov{\ve}$ holds. This lemma is proved in [@CMT2011 Lemma 2.2]. \[L2.7\] Let $(\hat{u}, \hat{T})$ be a solution of the transmission problem – . Assume that ${\hat{u}}$ is continuous in $Q$ and that the limits $(\na_x {\hat{u}})^{(\pm)}$ exist and define continuous extensions of $\na_x {\hat{u}}$ from the set $\gm'$ to $\gm' \cup \Gm$ and from the set $\gm$ to $\gm \cup \Gm$, respectively. Then we have $$\begin{aligned} &[\ve(\na_x \hat{u})] =\ve(u^* \otimes n), && [\hat{T}]=D\big(\ve(u^\otimes n)- \ov{\ve}\big), \label{E2.42}\\ &[\ve(\na_x \hat{u})] = P_n \ov\ve, && [\hat{T}]=-D Q_n \ov\ve, \label{E2.43} \end{aligned}$$ [Proof:]{} Equation is proved in [@JElast2012 Lemma 2.2], is proved in [@CMT2011]. For completeness we give the short proofs here. Since by assumption ${\hat{u}}$ is continuous across $\Gm$ and since $\na_x {\hat{u}}$ has continuous extensions from both sides of $\Gm$ onto $\Gm$, the surface gradients $(\na_\Gm {\hat{u}})^{(+)}$ and $(\na_\Gm {\hat{u}})^{(-)}$ on both sides of $\Gm$ coincide, hence $[\na_\Gm {\hat{u}}] = 0$. Using the decomposition and the definition of $u^$ we therefore obtain $$\label{E2.47} [\na_x {\hat{u}}] = [(\pa_n {\hat{u}}) \otimes n + \na_\Gm {\hat{u}}] = [(\pa_n {\hat{u}}) \otimes n] + [\na_\Gm {\hat{u}}] = [\pa_n {\hat{u}}] \otimes n = u^* \otimes n.$$ Thus, by , $$D\big( \ve(u^* \otimes n) - \ov{\ve} \big) = D\big( [\ve(\na_x {\hat{u}})] - \ov{\ve} [{\hat{S}}] \big) = \big[ D\big( \ve (\na_x {\hat{u}}) - \ov{\ve} {\hat{S}}\big) \big] = [{\hat{T}}].$$ This proves . From and we infer that $$0 = [{\hat{T}}]n = \Big( D \big( \ve(u^* \otimes n) - \ov{\ve} \big) \Big) n,$$ so that $[\ve(\na_x {\hat{u}})] = \ve(u^* \otimes n) = P_n \ov{\ve}$, by . Therefore we find $$[{\hat{T}}] = D\big( [\ve(\na_x {\hat{u}})] - \ov{\ve} \big) = D(P_n \ov{\ve} - \ov{\ve}) = - D Q_n \ov{\ve},$$ which proves . \[L2.8\] Let $({\hat{u}},{\hat{T}})$ and $({\check{u}},{\check{T}})$ be solutions of the transmission problems – and – , respectively. Assume that ${\hat{u}}$ and ${\check{u}}$ are continuous in $Q$ and that the limits $(\na_x {\hat{u}})^{(\pm)}$ and $(\na_x {\check{u}})^{(\pm)}$ exist and define continuous extensions of $\na_x {\hat{u}}$ and of $\na_x {\check{u}}$ from the set $\gm'$ to $\gm' \cup \Gm$ and from the set $\gm$ to $\gm \cup \Gm$, respectively. Then we have $$\label{E3.jumpvu} [\pa_n {\check{u}}] = \big[ \frac{\ov{\ve}:{\hat{T}}}{\hat{\psi}''({\hat{S}})} \big] u^, \qquad [\na_x {\check{u}}] = \big[ \frac{\ov{\ve}:{\hat{T}}}{\hat{\psi}''({\hat{S}})} \Big] u^ \otimes n.$$ [Proof:]{} The decomposition yields $$\label{E3.jumppavu} [\na_x {\check{u}}] = [(\pa_n {\check{u}}) \otimes n + \na_\Gm {\check{u}}] = [(\pa_n {\check{u}}) \otimes n] + [\na_\Gm {\check{u}}] = [\pa_n {\check{u}}] \otimes n.$$ From this equation and from , we infer $$0 = [{\check{T}}]n = \Big( D \big( [\ve(\na_x {\check{u}})] - \big[ \frac{\ov{\ve}:{\hat{T}}}{\hat{\psi}''({\hat{S}})} \big] \ov{\ve} \big) \Big) n = \Big( D \big( \ve\big( [\pa_n {\check{u}}] \otimes n \big) - \big[ \frac{\ov{\ve}:{\hat{T}}}{\hat{\psi}''({\hat{S}})} \big] \ov{\ve} \big) \Big) n.$$ Thus, , the linearity of the projector $P_n$ and , imply $$\ve\big( [\pa_n {\check{u}}] \otimes n \big) = \big[ \frac{\ov{\ve}:{\hat{T}}} {\hat{\psi}''({\hat{S}})} \big] P_n \ov{\ve} = \big[ \frac{\ov{\ve}:{\hat{T}}} {\hat{\psi}''({\hat{S}})} \big] \ve(u^* \otimes n) = \ve\Big( \big[ \frac{\ov{\ve}:{\hat{T}}} {\hat{\psi}''({\hat{S}})} \big] u^* \otimes n \Big),$$ whence $$\Big( [\pa_n {\check{u}}] - \big[ \frac{\ov{\ve}:{\hat{T}}} {\hat{\psi}''({\hat{S}})} \big] u^* \Big) \otimes n + n \otimes \Big( [\pa_n {\check{u}}] - \big[ \frac{\ov{\ve}:{\hat{T}}} {\hat{\psi}''({\hat{S}})} \big] u^* \Big) = 0.$$ We multiply this equation from the right with $n$ and obtain $$\Big( [\pa_n {\check{u}}] - \big[ \frac{\ov{\ve}:{\hat{T}}} {\hat{\psi}''({\hat{S}})} \big] u^* \Big) + n \Big( [\pa_n {\check{u}}] - \big[ \frac{\ov{\ve}:{\hat{T}}} {\hat{\psi}''({\hat{S}})} \big] u^* \Big)\cdot n = 0,$$ which means that $[\pa_n {\check{u}}] - \big[ \frac{\ov{\ve}:{\hat{T}}} {\hat{\psi}''({\hat{S}})} \big] u^* $ is a multiple of $n$. Scalar multiplication of the last equation with $n$ yields $ \big( [\pa_n {\check{u}}] - \big[ \frac{\ov{\ve}:{\hat{T}}} {\hat{\psi}''({\hat{S}})} \big] u^* \big) \cdot n =0$, from which we now conclude that the first equation in holds. The second equation is obtained from . The asymptotic solution {#S3} ======================= This section forms the first part of the proof . We state in this section the form of the asymptotic solution $(u^{(\mu)},T^{(\mu)},S^{(\mu)})$. The proof continues in Section \[S4\], where we study properties of the functions $S_0,\ldots,S_2$ appearing in the asymptotic solution. In Section \[S5\] we use these properties to show that $(u^{(\mu)},T^{(\mu)},S^{(\mu)})$ is an asymptotic solution by verifying that the estimates – hold. This concludes the proof of . Notations {#S3.1} --------- Before we can start with the construction of the asymptotic solution we must introduce more definitions and notations. In particular, we must introduce parallel manifolds to the manifold $\Gm$ and we must extend the definition of the surface gradients for functions defined on $\Gm$, which are given in Section \[S2.1\], to functions defined on the parallel manifolds. These definitions and notations are needed throughout the remaining sections. Let $\da > 0$ be the number from . For $\xi$ satisfying $-\delta < \xi < \delta$ $$\Gm_\xi = \{ (t,\eta + n(t,\eta)\xi) \mid (t,\eta) \in \Gm \}$$ is a three dimensional parallel manifold of $\Gm$ embedded in ${\mathcal U}_\da$, and $$\Gm_\xi(t) = \{ x \in \Om \mid (t,x) \in \Gm_\xi \}$$ is a two-dimensional parallel surface of $\Gm(t)$ embedded in ${\mathcal U}_\da(t)$. Let $\tau_1,\ \tau_2 \in \R^3$ be two orthogonal unit vectors tangent to $\Gm_{\xi}(t)$ at $x \in \Gm_{\xi}(t)$. For functions $w:\Gm_\xi(t) \rightarrow \R$, $W: \Gm_\xi(t) \rightarrow \R^3$ and $\hat{W}: \Gm_\xi(t) \rightarrow \R^{3\ti 3}$ we define the surface gradient and the surface divergence on $\Gm_\xi(t)$ by $$\begin{aligned} \na_{\Gm_\xi} w &=& (\pa_{\tau_1} w)\tau_1 + (\pa_{\tau_2} w)\tau_2 , \label{Esurfgrad-intro} \\[1ex] \na_{\Gm_\xi} W &=& (\pa_{\tau_1}W)\otimes\tau_1 + (\pa_{\tau_2}W) \otimes \tau_2 , \label{Esurfgradmat-intro} \\ \div_{\Gm_\xi} W & = & \tau_1 \cdot \pa_{\tau_1} W + \tau_2 \cdot \pa_{\tau_2} W = \sum_{i=1}^2 \tau_i \cdot (\na_{\Gm_\xi} W)\tau_i, \label{Esurfdiv} \\ \div_{\Gm_\xi} \hat{W} & = & (\pa_{\tau_1} \hat{W}) \tau_1 + (\pa_{\tau_2} \hat{W}) \tau_2 . \label{Esurfdivmat} \end{aligned}$$ Clearly, with $\na_\Gm$ defined in and we have $\na_{\Gm_0}=\na_\Gm$. For brevity we write $\div_\Gm = \div_{\Gm_0}$. If $w$, $W$, $\hat{W}$ are defined on $\Gm_\xi$, we define $\na_{\Gm_\xi} w: \Gm_\xi \mapsto \R^3$, $\na_{\Gm_\xi} W: \Gm_\xi \mapsto \R^{3\times3}$, $\div_{\Gm_\xi} W: \Gm_\xi \mapsto \R$, $\div_{\Gm_\xi} \hat{W}: \Gm_\xi \mapsto \R^3$ by applying the operators $\na_{\Gm_\xi}$ and $\div_{\Gm_\xi}$ to the restrictions $w\rain{\Gm_\xi(t)}$, $W\rain{\Gm_\xi(t)}$, $\hat{W}\rain{\Gm_\xi(t)}$ for every $t$. With these definitions we have the splittings $$\begin{aligned} \na_x w(t,x) &=& \pa_\xi w(t,\eta,\xi)\, n(t,\eta) + \na_{\Gm_\xi} w(t,\eta,\xi), \label{Eskalgradsplit} \\ \na_x W(t,x) &=& \pa_\xi W(t,\eta,\xi)\otimes n(t,\eta) + \na_{\Gm_\xi} W(t,\eta,\xi), \label{Egradsplit} \\ \div_x \hat{W}(t,x) &=& \big(\pa_\xi \hat{W}(t,\eta,\xi)\big) n(t,\eta) + \div_{\Gm_\xi} \hat{W}(t,\eta,\xi), \label{Edivsplit} \end{aligned}$$ where, as usual, $W(t,\eta,\xi) = W\big(t,\eta + n(t,\eta) \xi\big)$. The operators $\na_\Gm$ and $\div_\Gm$ can be applied to functions defined on subsets of $\Gm$. In contrast, the operator $\na_\eta$ introduced next can be applied to functions defined on $\Gm \ti J$, where $J \subseteq \R$ is an interval. For $w:\Gm \ti J \ra \R$, $W:\Gm \ti J \ra \R^3$, $\hat{W}:\Gm \ti J \ra \R^{3\ti 3}$ consider the functions $\eta \mapsto w_{t,\xi}(\eta) = w(t,\eta,\xi)$, $\eta \mapsto W_{t,\xi}(\eta) = W(t,\eta,\xi)$, $\eta \mapsto \hat{W}_{t,\xi}(\eta) = \hat{W}(t,\eta,\xi)$, which are defined on $\Gm(t)$. To these functions the operators $\na_\Gm$ and $\div_{\Gm}$ can be applied. We set $$\begin{aligned} \na_\eta w(t,\eta,\xi) &=& \na_\Gm w_{t,\xi}(\eta) \in \R^3, \label{Enaeta} \\ \na_\eta W(t,\eta,\xi) &=& \na_\Gm W_{t,\xi}(\eta) \in \R^{3\ti 3}, \label{EnaetaW} \\ \div_\eta W(t,\eta,\xi) &=& \div_\Gm W_{t,\xi}(\eta) \in \R, \label{Ediveta} \\ \div_\eta \hat{W}(t,\eta,\xi) &=& \div_\Gm \hat{W}_{t,\xi}(\eta) \in \R^3. \label{EdivetahW} \end{aligned}$$ If $W$ is defined on ${\cal U}_\da$, then $(t,\eta,\xi) \ra W(t,\eta,\xi) = W(t,\eta + n(t,\eta) \xi)$ is defined on $\Gm\ti (-\da,\da)$. Consequently, the gradient $\na_\eta W$ is defined. The connection between $\na_\eta W$ and $\na_{\Gm_\xi} W = \na_{\Gm_\xi} W\rain{\Gm_\xi}$ is given by the chain rule, which yields $$\label{Eetatrans} \na_\eta W (t,\eta,\xi) = \big(\na_{\Gm_\xi} W(t,\eta + n(t,\eta) \xi)\big) \big( I + \xi \, \na_\eta n(t,\eta) \big).$$ In particular, we have $\na_\eta W(t,\eta,0) = \na_\Gm W(t,\eta)$. Similar formulas and relations hold for $\na_\eta w$, $\div_\eta W$, $\div_\eta \hat{W}$. If $W:{\cal U}_\da \ra \R^3$ is constant on all the lines normal to $\Gm(t)$, for all $t$, we have $W(t,\eta,\xi) = W(t,\eta)$. For such functions we sometimes interchangeably use the notations $\na_\eta W$ and $\na_\Gm W$. Similarly, we interchangeably use the notations $\na_\eta w$ and $\na_\Gm w$, $\div_\eta W$ and $\div_\Gm W$, $\div_\eta \hat{W}$ and $\div_\Gm \hat{W}$ if $w$ and $\hat{W}$ are independent of $\xi$. Note that by we have for $x \in \Gm_\xi(t)$ that $$\label{Esurfgradtrans} \na_{\Gm_\xi} W(t,x) = \big(\na_\eta W(t,\eta,\xi)\big) A(t,\eta,\xi),$$ where $A(t,\eta,\xi) \in \R^{3\ti3}$ is the inverse of the linear mapping $\big( I + \xi \na_\eta n(t,\eta)\big):\R^3 \ra \R^3 $. From the mean value theorem we obtain the expansion $$\label{EAexpansion} A(t,\eta,\xi) = I + \xi\, R_A(t,\eta,\xi)$$ where the remainder term $R_A(t,\eta,\xi) \in \R^{3\ti 3}$ is bounded when $(t,\eta,\xi)$ varies in $\Gm \ti (-\da,\da)$. Insertion into yields $$\label{Esurfgraddeco} \na_{\Gm_\xi} W(t,x) = \na_\eta W(t,\eta,\xi)\, \big(I + \xi R_A(t,\eta,\xi) \big).$$ For $w :{\cal U}_\da \to \R$ we consider $\na_{\Gm_\xi} w$ and $\na_\eta w$ to be column vectors. For such $w$ the equation corresponding to is $$\label{EnaGmxiw} \na_{\Gm_\xi} w(t,x) = A^T(t,\eta,\xi) \na_\eta w(t,\eta,\xi) = \big(I + \xi\, R_A^T(t,\eta,\xi) \big)\, \na_\eta w(t,\eta,\xi) .$$ Furthermore, , and together yield for $W:{\cal U}_\da \ra \R^3$ that $$\label{EdivGmxiW} \div_{\Gm_\xi} W = \sum_{i=1}^2 \tau_i \cdot \big((\na_\eta W)(I + \xi R_A)\tau_i\big) = \div_\eta W + \xi\, \div_{\Gm,\xi} W,$$ with the remainder term $$\label{Edivxi} \div_{\Gm,\xi} W(t,\eta,\xi) = \sum_{i=1}^2 \tau_i \cdot \big( ( \na_\eta W ) R_A \tau_i \big) = \sum_{i=1}^2 \tau_i \cdot \big( ( \na_\Gm W_{t,\xi} ) R_A \tau_i \big),$$ and this equation implies for $\hat{W}:{\cal U}_\da \ra \R^{3\ti3}$ that $$\label{Esurfdivmatdeco} \div_{\Gm_\xi} \hat{W}(t,\eta,\xi) = \div_\eta \hat{W} + \xi\, \div_{\Gm,\xi} \hat{W},$$ where $\div_{\Gm,\xi} \hat{W} = \sum_{i,j=1}^2 (\pa_{\tau_j} \hat{W}_{t,\xi})\, \tau_i (\tau_j \cdot R_A \tau_i)$. The terms $\div_{\Gm,\xi} W$ and $\div_{\Gm,\xi} \hat{W}$ are bounded when $(t,\eta,\xi)$ varies in $\Gm \ti (-\da,\da)$. For functions $w$ with values in $\R$ we define the second gradients $\na^2_{\Gm_\xi} w$, $\na^2_\eta w$ by applying the operators $\na_{\Gm_\xi}$, $\na_\eta$ to the vector functions $\na_{\Gm_\xi} w$, $\na_\eta w$. For $W$ with values in $\R^3$ we define second gradients $\na^2_{\Gm_\xi} W$, $\na^2_\eta W$ by applying these operators to the rows of $\na_{\Gm_\xi} W$, $\na_\eta W$. We remark that $$\Da_{\Gm_\xi} w = \div_{\Gm_\xi} \na_{\Gm\xi}\, w$$ is the surface Laplacian. \[D3.1\] Let $I \subseteq \R$ be an interval. For $k,m \in \N_0$ and $p = 1,3$ we define the space $$\begin{gathered} C^k \big( I, C^m (\Gm , \R^p)\big) \\ = \{ (t,\eta,\xi) \ra w(t,\eta,\xi): \Gm \ti I \ra \R^p \mid \pa^\ell_\xi \pa^i_t \na^j_\eta w \in C(\Gm \ti I), \ \ell \leq k, \ i+j \leq m \}. \end{gathered}$$ Construction of the asymptotic solution {#S3.2} --------------------------------------- We start with the construction of the asymptotic solution $(u^{(\mu)}, T^{(\mu)}, S^{(\mu)})$. We assume that the hypotheses of Theorem \[T2.3\] are satisfied. In particular, we assume that there is a sufficiently smooth solution $\Gm=\Gm^{(\mu)}$ of the evolution problem , , with $s_0 ({\hat{T}},\ka_\Gm,\la^{1/2})$, $s_1 ({\hat{u}},{\hat{T}},{\check{T}},S_0,S_1,\la^{1/2})$ defined in – . By this assumption, the function $(\hat{u}, \hat{T}, {\check{u}}, {\check{T}}, S_0, S_1)$ is known as a solution of the transmission-boundary value problem – . We use the notation $$\label{E3.1} 1^+(r)= \begin{cases} 1, & r >0\\ 0, & r \leq 0 \end{cases}, \qquad 1^-(r)=1-1^+(r), \qquad r^\pm = r\,1^\pm(r).$$ Let $\phi \in C_0^\infty((-2,2))$ be a function satisfying $0 \leq \phi(r) \leq 1$ for all $r \in \R$ and $\phi(r) = 1$ for $\|r\| \leq 1$. With the constant $a$ from we define a function $\phi_{\mu\la}:Q \ra [0,1]$ by $$\label{E3.2} \begin{split} \phi_{\mu \la} (t,x) &= \phi_{\mu\la}(t,\eta,\xi) = \phi \Big( \frac{2a \xi}{3 (\mu\la)^{1/2} \| \ln \mu\| } \Big), \quad \mbox{for } (t,x) \in {\cal U}_\da, \\ \phi_{\mu \la} (t,x) &= 0, \qquad \mbox{otherwise}. \end{split}$$ By , $\phi_{\mu\la}$ is equal to $1$ in $Q_{\rm inn}^{(\mu\la)}$, transits smoothly from $1$ to $0$ in $Q_{\rm match}^{(\mu\la)}$, and vanishes in $Q_{\rm out}^{(\mu\la)}$. With this function the asymptotic solution is defined by $$\begin{aligned} u^{(\mu)}(t,x) &= u_1^{(\mu)} (t,x)\, \phi_{\mu\la}(t,x) + u_2^{(\mu)}(t,x)\, \big( 1-\phi_{\mu\la} (t,x) \big) , \label{E3.3}\\ S^{(\mu)}(t,x) &= S_1^{(\mu)} (t,x)\, \phi_{\mu\la}(t,x) + S_2^{(\mu)}(t,x)\, \big( 1-\phi_{\mu\la} (t,x) \big) , \label{E3.4}\\ T^{(\mu)}(t,x) &= D \Big(\ve\big(\na_x u^{(\mu)}(t,x)\big)- \ov\ve S^{(\mu)}(t,x)\Big), \label{E3.5} \end{aligned}$$ where $u_1^{(\mu)}$, $S_1^{(\mu)}$ are components of the inner expansion $\big(u_1^{(\mu)}, T_1^{(\mu)}, S_1^{(\mu)}\big)$ defined in ${\cal U}_\da$, and $u_2^{(\mu)}$, $S_2^{(\mu)}$ are components of the outer expansion $\big(u_2^{(\mu)}, T_2^{(\mu)}, S_2^{(\mu)} \big)$ defined in $Q \setminus \Gamma$. The function $\big(u^{(\mu)},T^{(\mu)},S^{(\mu)} \big)$ is equal to the inner expansion $\big(u_1^{(\mu)},T_1^{(\mu)},S_1^{(\mu)} \big)$ in the region $Q_{\rm inn}^{(\mu\la)}$ and equal to the outer expansion $\big(u_2^{(\mu)},T_2^{(\mu)},S_2^{(\mu)} \big)$ in the region $Q_{\rm out}^{(\mu\la)}$. In the region $Q_{\rm match}^{(\mu\la)}$ both expansions are matched. #### The outer expansion The outer expansion is defined as follows. With the solutions $(\hat{u}, \hat{T})$ of the transmission problem – and $({\check{u}},{\check{T}})$ of the transmission problem – we set for $(t,x) \in Q \setminus \Gm$ $$\begin{aligned} u_2^{(\mu)} (t,x) &=& {\hat{u}}(t,x)+ \mu^{1/2} {\check{u}}(t,x)+ \mu\, \tilde{u}(t,x), \label{E3.6}\\[1ex] S_2^{(\mu)}(t,x) &=& \hat{S}(t,x)+ \mu^{1/2} \tilde{S}_1(t,x)+ \mu \tilde{S}_2(t,x) +\mu^{3/2} \tilde{S}_3(t,x), \label{E3.7}\\[1ex] T_2^{(\mu)}(t,x) &=& D \Big(\ve\big(\na_x u_2^{(\mu)}(t,x)\big)- \ov\ve S_2^{(\mu)}(t,x)\Big). \label{E3.8} \end{aligned}$$ The functions ${\tilde{u}}$, ${\tilde{S}}_1, \dots, {\tilde{S}}_3$ and another unknown function ${\tilde{T}}$ solve the system of algebraic and partial differential equations $$\begin{aligned} -\div_x {\tilde{T}}&=& 0, \label{E3.9} \\ {\tilde{T}}&=& D\big(\ve(\na_x {\tilde{u}}) - \ov{\ve} {\tilde{S}}_2 \big), \label{E3.10}\\ - {\hat{T}}:\ov{\ve} + \hat{\psi}'' ({\hat{S}}) {\tilde{S}}_1 &=& 0, \label{E3.11} \\ - {\check{T}}:\ov{\ve} + \hat{\psi}'' ({\hat{S}}) {\tilde{S}}_2 + \frac12 \hat{\psi}''' ({\hat{S}}) {\tilde{S}}_1^2 &=& 0, \label{E3.12} \\ - {\tilde{T}}:\ov{\ve} + \hat{\psi}'' ({\hat{S}}) {\tilde{S}}_3 + \hat{\psi}''' ({\hat{S}}) {\tilde{S}}_1 {\tilde{S}}_2 + \frac16 \hat{\psi}^{(IV)} ({\hat{S}}){\tilde{S}}_1^3 \quad \nn \\ \mbox{} + \frac{\la^{1/2}}{c} \pa_t {\tilde{S}}_1 - \la \Da_x {\tilde{S}}_1 &=& 0, \label{E3.13} \end{aligned}$$ in the set $Q \setminus \Gamma$. Moreover, ${\tilde{u}}$ satisfies the boundary conditions $$\begin{aligned} {\tilde{u}}(t,x) &=& 0, \qquad (t,x) \in [t_1,t_2] \ti \pa\Om, \label{E3.14} \\ {\tilde{u}}^{(-)}(t,\eta) &=& \la^{1/2} u^(t,\eta) \int_{-\infty}^0 S_1(t,\eta,\zeta) - \frac{\ov{\ve} : {\hat{T}}^{(-)}(t,\eta)} {\hat{\psi}''(1)}\, d\zeta, \quad (t,\eta) \in \Gm, \label{E3.15} \\ {\tilde{u}}^{(+)}(t,\eta) &=& \la^{1/2} u^(t,\eta) \int_0^\infty S_1(t,\eta,\zeta) - \frac{\ov{\ve} : {\hat{T}}^{(+)}(t,\eta)} {\hat{\psi}''(1)} \,d\zeta \nn \\ && \mbox{} + \la\, a^(t,\eta) \int_{-\infty}^\infty \Big( \int_{-\infty}^\zeta S_0(\vartheta) \, d\vartheta - \zeta^+ \Big) d\zeta, \quad (t,\eta) \in \Gm. \label{E3.16}\end{aligned}$$ Since by assumption $\Gm$, ${\hat{T}}$, ${\check{T}}$, $S_1$ are known from the evolution problem, this system can be solved recursively. To see this, note that yields $$\label{E3.17} \tilde{S}_1= \frac{\hat{T}:\ov{\ve}}{\hat{\psi}''(\hat{S})}.$$ We insert this equation into and solve this equation for ${\tilde{S}}_2$ to obtain $$\label{E3.18} {\tilde{S}}_2 = \frac{{\check{T}}: \ov{\ve}}{\hat{\psi}''({\hat{S}})} - \frac{\hat{\psi}'''({\hat{S}})}{2\hat{\psi}''({\hat{S}})} \Big( \frac{{\hat{T}}:\ov{\ve}}{\hat{\psi}''({\hat{S}})} \Big)^2.$$ Using this function in , we can determine ${\tilde{u}}$ and ${\tilde{T}}$ from the boundary value problem , , – . Finally, we can solve for $\tilde{S}_3$. #### The inner expansion The inner expansion $(u_1^{(\mu)},T_1^{(\mu)},S_1^{(\mu)})$ is essentially obtained by smoothing the jumps of the functions ${\hat{u}}$, ${\check{u}}$ and ${\hat{S}}$ from the evolution problem for the surface $\Gm$. Before we can define the inner expansion we must therefore study in the next two lemmas the jumps of ${\hat{u}}$ and ${\check{u}}$ across $\Gm$. Let $u^= [\pa_n \hat{u}]$ and $a^* = [\pa_n^2 \hat{u}]$ be the jumps of derivatives of ${\hat{u}}$ across $\Gm$. These functions are introduced in , . For $(t,x) = (t,x(t,\eta,\xi)) \in {\cal U}_\da$ we decompose ${\hat{u}}$ and ${\check{u}}$ in the form $$\begin{aligned} {\hat{u}}(t,x) &=& u^(t,\eta)\, \xi^+ + a^(t,\eta) \frac12(\xi^+)^2 + \hat{v}(t,x), \label{E3.19} \\ {\check{u}}(t,x) &=& u^(t,\eta) \Big(\frac{\ov{\ve} : \hat{T}^{(+)}(t,\eta)}{\hat{\psi}''(1)}\, \xi^+ + \frac{\ov{\ve} : \hat{T}^{(-)}(t,\eta)}{\hat{\psi}''(0)}\, \xi^- \Big) + {\check{v}}(t,x), \label{E3.20} \end{aligned}$$ where $\xi^+$, $\xi^-$ are defined in and where the remainder terms ${\hat{v}}$ and ${\check{v}}$ are defined by , . The decomposition is motivated by the fact that $$\label{E3.jumphv} [\pa_n^i \hat{v}] = 0, \qquad i = 0,1,2,$$ which follows immediately from , and , . The first two terms on the right hand side of thus serve to separate off the jumps of the first and second derivatives of ${\hat{u}}$ at $\Gm$. Similarly, the normal derivatives of first order of ${\check{v}}$ do not jump across $\Gm$. More precisely, we have the following result. \[L3.2\] Let $({\hat{u}},{\hat{T}})$ and $({\check{u}},{\check{T}})$ be solutions of the transmission problems – and – , respectively.\ (i) ${\check{v}}$ defined in satisfies $$\label{E3.jumpvv} [\pa_n^i {\check{v}}] = 0, \qquad i=0,1.$$ (ii) Assume that $\Gm$ is a $C^5$–manifold. Suppose that ${\hat{u}}\in C^4(\gm \cup \gm',\R^3)$, ${\check{u}}\in C^3(\gm \cup \gm',\R^3)$ and that ${\hat{u}}$ has $C^4$–extensions, ${\check{u}}$ has $C^3$–extensions from $\gm$ to $\gm \cup \Gm$ and from $\gm'$ to $\gm' \cup \Gm$. With the function spaces introduced in we then have $$\begin{aligned} {\hat{v}}&\in& C^2\big( (-\da,\da), C^2(\Gm)\big) \cap C^3\big( (-\da,0], C^1(\Gm)\big) \cap C^3\big( [0,\da), C^1(\Gm)\big), \label{E3.hvregu}\\ {\check{v}}&\in& C^1 \big( (-\da,\da), C^2(\Gm)\big) \cap C^3 \big( \Gm \ti (-\da,0] \big) \cap C^3 \big( \Gm \ti [0,\da) \big). \label{E3.vvregu}\end{aligned}$$ [Proof:]{} To prove note that by definition of $[\pa_n w]$ in Section \[S2.1\] and by definiton of $\xi^\pm$ in we have $$\Big[ \pa_n \Big(\frac{\ov{\ve} : \hat{T}^{(+)}}{\hat{\psi}''(1)}\, \xi^+ + \frac{\ov{\ve} : \hat{T}^{(-)}}{\hat{\psi}''(0)}\, \xi^- \Big) u^ \Big] = \Big( \frac{\ov{\ve} : \hat{T}^{(+)}}{\hat{\psi}''(1)} - \frac{\ov{\ve} : \hat{T}^{(-)}}{\hat{\psi}''(0)} \Big) u^* = \Big[ \frac{\ov{\ve} : \hat{T}}{\hat{\psi}''({\hat{S}}} \Big] u^.$$ From this equation, from the first equation in and from we obtain for $i=1$. For $i=0$ equation is an immediate consequence of and . This proves (i). \(ii) Since $\Gm$ is a $C^5$–manifold, the coordinate mapping $(t,\eta,\xi) \mapsto (t,x(t,\eta,\xi)) = \big(t,\eta + \xi\, n(t,\eta) \big)$ and the inverse mapping $(t,x) \mapsto \big(t,\eta(t,x),\xi(t,x)\big)$ are $C^4$. It follows from this differentiability property of the coordinate mapping and from our differentiability assumptions for ${\hat{u}}$ that $(t,\eta,\xi) \mapsto {\hat{u}}(t,\eta,\xi)$ is $C^4$ in $\Gm \ti (-\da,0]$ and in $\Gm \ti [0,\da)$, and that $(t,\eta) \mapsto u^(t,\eta) = n(t,\eta) \cdot [\na_x {\hat{u}}](t,\eta)$ belongs to $C^3 (\Gm)$ and $(t,\eta) \mapsto a^(t,\eta) = n(t,\eta) \cdot [\pa_n \na_x {\hat{u}}](t,\eta)$ belongs to $C^2 (\Gm)$. Since by we have $${\hat{v}}(t,\eta,\xi) = {\hat{u}}(t,\eta,\xi) - u^(t,\eta) \xi^+ - a^(t,\eta) \frac12(\xi^+)^2,$$ these properties imply that $${\hat{v}}\in \bigcap_{m=0}^1 \Big( C^{2+m}\big( (-\da,0],C^{2-m} (\Gm)\big) \cap C^{2+m}\big( [0,\da),C^{2-m}(\Gm)\big) \Big).$$ From this relation and from we conclude that holds. Relation is obtained in the same way, using instead of .\ For brevity in notation we define $$\begin{gathered} \hat{\si}(\xi) = \hat{\si}(t,\eta,\xi) = \ov{\ve}: D\ve\big(\na_x {\hat{v}}(t,x) \big), \qquad \hat{\si}'(\xi) = \pa_\xi \hat{\si}(t,\eta,\xi), \label{E3.21} \\ \check{\si}(\xi) = \check{\si}(t,\eta,\xi) = \ov{\ve}: D\ve\big(\na_x {\check{v}}(t,x) \big). \label{E3.22}\end{gathered}$$ Later we need the following result, which shows how ${\check{\sigma}}(t,\eta,0)$ can be computed from the limit values of ${\hat{T}}$ and ${\check{T}}$ at $\Gm$. \[L3.3\] The function ${\check{\sigma}}$ defined in satisfies $$\begin{aligned} \ov{\ve}:{\check{T}}^{(+)} &=& {\check{\sigma}}(0) + \ov{\ve}:[{\hat{T}}]\, \frac{\ov{\ve}:{\hat{T}}^{(+)}} {\hat{\psi}''(1)}, \label{E3.22a} \\ \ov{\ve}:{\check{T}}^{(-)} &=& {\check{\sigma}}(0) + \ov{\ve}:[{\hat{T}}]\, \frac{\ov{\ve}:{\hat{T}}^{(-)}} {\hat{\psi}''(0)}, \label{E3.22b} \\ {\check{\sigma}}(0) &=& \ov{\ve}: \langle {\check{T}}\rangle - \ov{\ve}:[{\hat{T}}] \Big\langle \frac{\ov{\ve}:{\hat{T}}} {\hat{\psi}''({\hat{S}})} \Big\rangle. \label{E3.22c}\end{aligned}$$ [Proof:]{} We apply the decomposition of the gradient to the function $W(t,\eta,\xi) = u^(t,\eta) \Big( \frac{\ov{\ve} : {\hat{T}}^{(+)}(t,\eta)}{\hat{\psi}''(1)} \xi^+ + \frac{\ov{\ve} : {\hat{T}}^{(-)}(t,\eta)}{\hat{\psi}''(0)} \xi^- \Big).$ This yields $$(\na_x W)^{(+)}(t,\eta) = \big((\pa_n W) \otimes n + \na_\Gm W\big)^{(+)} = u^* \otimes n\, \frac{\ov{\ve} : {\hat{T}}^{(+)}}{\hat{\psi}''(1)}.$$ thus implies $$(\na_x {\check{u}})^{(+)} = u^* \otimes n\, \frac{\ov{\ve} : {\hat{T}}^{(+)}}{\hat{\psi}''(1)} + \na_x {\check{v}}.$$ Insertion of these equations into yields $${\check{T}}^{(+)} = D\big( \ve(u^* \otimes n) - \ov{\ve} \big) \frac{\ov{\ve} : {\hat{T}}^{(+)}}{\hat{\psi}''(1)} + D\ve(\na_x {\check{v}}).$$ We take the scalar product with $\ov{\ve}$ on both sides of this equation and note and the definition of ${\check{\sigma}}$ in to obtain . Equation is obtained in the same way. To prove , we add and and solve the resulting equation for ${\check{\sigma}}(0)$. This proves the lemma. #### Definition of the inner expansion We can now construct the inner expansion $(u_1^{(\mu)},T_1^{(\mu)},S_1^{(\mu)})$. With the remainder terms ${\hat{v}}$, ${\check{v}}$ introduced in , we set in the neighborhood ${\cal U}_\da$ of $\Gm$ $$\begin{aligned} u^{(\mu)}_1(t,x) &=& (\mu\la)^{1/2} u_0 \big(t,\eta,\frac{\xi}{(\mu\la)^{1/2}} \big) + \mu \la^{1/2} u_1 \big(t,\eta,\frac{\xi}{(\mu\la)^{1/2}} \big) + \mu \la u_2 \big(t,\eta,\frac{\xi}{(\mu\la)^{1/2}} \big) \nn \\ && \mbox{} + \hat{v}(t,x) + \mu^{1/2}{\check{v}}(t,x), \label{E3.23} \\ S^{(\mu)}_1(t,x) &=& S_0\big(\frac{\xi}{(\mu\la)^{1/2}}\big) + \mu^{1/2} S_1\big(t,\eta,\frac{\xi}{(\mu\la)^{1/2}}\big) + \mu S_2\big(t,\eta,\frac{\xi}{(\mu\la)^{1/2}}\big), \label{E3.24} \\[1ex] T^{(\mu)}_1(t,x) &=& D\Big(\ve\big(\na_x u^{(\mu)}_1(t,x)\big) - \ov{\ve} S^{(\mu)}_1(t,x) \Big), \label{E3.25}\end{aligned}$$ where by assumption $S_0$, $S_1$ are known from the evolution problem for $\Gm$, and where the functions $u_0,\ldots ,u_2$ are defined by $$\begin{aligned} u_0(t,\eta,\zeta) &=& u^(t,\eta) \int_{-\infty}^\zeta S_0({\vartheta})\,d{\vartheta}, \label{E3.26} \\ u_1(t,\eta,\zeta) &=& u^(t,\eta) \int_0^\zeta S_1(t,\eta,{\vartheta}) \, d {\vartheta}, \label{E3.27} \\ u_2 (t,\eta,\zeta) &=& a^(t,\eta) \int_{-\infty}^\zeta \int_{-\infty}^{{\vartheta}} S_0({\vartheta}_1)\,d{\vartheta}_1 d{\vartheta}, \label{E3.28} $$ The function $S_2 = S_2(t,\eta,\zeta)$ together with another unknown function $s_1 = s_1(t,\eta)$ solve a boundary value problem. To state this boundary value problem let $\ka(t,\eta,\xi)$ denote twice the mean curvature of the surface $\Gm_\xi(t)$ at $\eta \in \Gm_\xi(t)$. With the notation introduced in Section \[S2.2\] we thus have $\ka(t,\eta,0) = \ka_\Gm(t,\eta)$. We write $\ka'(0) = \pa_\xi \ka(t,\eta,0)$. The boundary value problem for $S_2$ and $s_1$ consists of the ordinary differential equation $$\label{E3.29} \hat{\psi}'' \big(S_0(\zeta)\big) S_2(t,\eta,\zeta) - S_2''(t,\eta,\zeta) = F_2(t,\eta,\zeta),$$ with the right hand side given by $$\begin{aligned} F_2(t,\eta,\zeta) &= {\check{\sigma}}(0) + \ov{\ve}: [{\hat{T}}]\, S_1 - \frac{1}{c_1} \ov{\ve}:\langle {\hat{T}}\rangle S_1' - \frac12 \hat{\psi}'''(S_0) S_1^2 \nn \\ & \mbox{ } + \la^{1/2}\Big( \hat{\si}'(0) \zeta + \ov{\ve} : D\ve(a^ \otimes n + \na_\Gm u^) \int_{-\infty}^\zeta S_0(\vartheta)\, d\vartheta \Big) \nn \\ &\mbox{ } + \Big( \frac{s_1}{c} - \la \ka'(0) \zeta \Big) S_0'\,, \label{E3.30}\end{aligned}$$ and of boundary conditions. To formulate these boundary conditions, we choose $\vp \in C^\infty(\R,[0,1])$ such that $$\label{E3.31} \vp(\zeta) = \begin{cases} 0, & \zeta \leq 1, \\ 1, & \zeta \geq 2, \end{cases}$$ set $$\label{E3.32} \vp_+(\zeta) = \frac{\vp(\zeta)}{\hat{\psi}''(1)}, \qquad \vp_-(\zeta) = \frac{\vp(-\zeta)}{\hat{\psi}''(0)},$$ and define $$\begin{aligned} \rho_2(t,\eta,\zeta) = \vp_-(\zeta) \Biggl( \ov{\ve}:{\check{T}}^{(-)} - \frac{\hat{\psi}'''(0)}{2} & \Big( \frac{\ov{\ve} : {\hat{T}}^{(-)}}{\hat{\psi}''(0)} \Big)^2 + \la^{1/2}\, \hat{\si}'(0) \zeta \Biggr) \nn \\ \mbox{} + \vp_+(\zeta) \Biggl( \ov{\ve}:{\check{T}}^{(+)} - \frac{\hat{\psi}'''(1)}{2} & \Big( \frac{\ov{\ve} : {\hat{T}}^{(+)}}{\hat{\psi}''(1)} \Big)^2 + \la^{1/2} \hat{\si}'(0) \zeta \nn \\ \mbox{} + \la^{1/2}& \ov{\ve} : D\ve (a^ \otimes n + \na_\Gm u^) \zeta^+ \Biggr). \label{E3.33}\end{aligned}$$ With this function the boundary conditions are $$\begin{gathered} S_2(t,\eta,0) = 0, \label{E3.34} \\ \lim_{\zeta \ra \pm\infty} \big( S_2 (t,\eta,\zeta) - \rho_2(t,\eta,\zeta) \big) = 0. \label{E3.35} \end{gathered}$$ The function $s_1 = s_1(t,\eta)$ in is independent of $\zeta$. It is determined in Section \[S4.2\] by the procedure sketched at the end of Section \[S2.2\], which we apply of course to the boundary value problem , , , instead of the problem , – . The function $s_1$, whose explicit expression is given in – , forms the second term in the definition of the evolution operator ${\cal K}^{(\mu)}$. The functions $\boldsymbol{S_0,\ldots,S_2}$ from the inner expansion {#S4} ==================================================================== The functions $\tilde{S}_1,\ldots,\tilde{S}_3$ in the outer expansion can be determined explicitly from – , whereas the functions $S_0,\ldots,S_2$ in the inner expansion are determined as solutions of three coupled boundary value problems to linear and nonlinear ordinary differential equations. It is not obvious that these solutions exist and what properties they have. We study these solutions in this section. The function $\boldsymbol{S_0}$ {#S4.1} ------------------------------- The first boundary value problem determining $S_0$ is given by , , \[L4.1\] Assume that the double well potential $\hat{\psi}$ satisfies . Then $S_0$ is a solution of the boundary value problem , , if and only if $S_0$ satisfies the initial value problem $$\label{E4.1} S'_0(\zeta)= \sqrt{2 \hat{\psi}\big(S_0(\zeta)\big)}, \quad \zeta \in \R, \qquad S_0(0) = \frac{1}{2}.$$ [Proof:]{} Let $S_0$ be a solution of , . We multiply by $S'_0$ and obtain $$\frac{d}{d\zeta} \Big(\hat{\psi}(S_0)- \frac{1}{2}(S'_0)^2\Big) = 0,$$ or $$\label{E4.2} \hat{\psi}(S_0)- \frac{1}{2}(S'_0)^2 = C_1.$$ By we have $\lim_{\zeta \to \infty} S_0(\zeta)=1$. From and from we thus obtain that $\lim_{\zeta \to \infty}\big(S'_0(\zeta)\big)^2 = -2C_1$. Using again , we infer from this limit relation that $\lim_{\zeta \to \infty}S'_0(\zeta) = 0$, hence $C_1=0$. We solve for $S'_0$ and use that because of the boundary conditions the function $S_0$ must be increasing, hence $S'_0$ must be nonnegative. This shows that a solution of , must satisfy the initial value problem . To prove the converse we differentiate the differential equation in and obtain . We leave it to the reader to verify that the solution of satisfies the boundary conditions . \[T4.2\] Assume that $\hat{\psi} \in C^3([0,1], \R)$ has the properties . Then there is a unique solution $S_0 \in C^4(\R, (0,1))$ of the initial value problem . This solution is strictly increasing and satisfies and . Moreover, there are constants $K_1, \dots , K_3 >0$ such that for $a > 0$ defined in $$\begin{aligned} \label{E4.6} 0< S_0(\zeta) \leq K_1 e^{-a \|\zeta\|}, & \quad -\infty < \zeta \leq 0, \\ \label{E4.7} 1-K_2 e^{- a \zeta} \leq S_0 (\zeta) < 1, & \quad 0 \leq \zeta < \infty, \\ \label{E4.8} \|\pa^i S_0(\zeta)\| \leq K_3 e^{-a \|\zeta\|}, & \quad - \infty < \zeta < \infty,\ i=1,\ldots,4\,. \end{aligned}$$ This theorem follows immediately from the standard theory of ordinary differential equations, and we omit the proof. \[L4.3\] If $\hat{\psi}$ satisfies the symmetry condition , then the solution $S_0$ of satisfies for all $\zeta \in \R$ $$\begin{gathered} S_0(-\zeta)=1-S_0(\zeta), \qquad S'_0(\zeta)=S'_0(-\zeta), \label{E4.9a} \\ \int_{-\infty}^\zeta S_0(\vartheta)\, d \vartheta = \int_{-\infty}^{-\|\zeta\|} S_0(\vartheta)\, d\vartheta + \zeta^+, \label{E4.9b} \\ \Big\| \int_{-\infty}^\zeta S_0(\vartheta)\, d \vartheta - \zeta^+ \Big\| \leq \frac{K_1}{a} e^{-a \|\zeta\|}, \label{E4.9c} \\[1ex] \| \hat{\psi}''\big(S_0(\zeta)\big) - \hat{\psi}''\big({\hat{S}}(\zeta) \big)\| \leq K_4 e^{- a \|\zeta\|}. \label{E4.9d}\end{gathered}$$ [Proof:]{} If the symmetry condition holds and if $S_0$ is a solution of the initial value problem , then also $\zeta \mapsto \big(1-S_0(-\zeta)\big)$ is a solution. To see this, note that and imply $$\begin{aligned} \sqrt{2\hat{\psi}\big(1-S_0(-\zeta)\big)} &= \sqrt{2 \hat{\psi}\Big( \frac{1}{2}+\Big(\frac{1}{2}-S_0(-\zeta)\Big)\Big)} = \sqrt{2 \hat{\psi}\Big( \frac{1}{2}-\Big(\frac{1}{2}-S_0(-\zeta)\Big)\Big)} \\ &= \sqrt{2 \hat{\psi}\big(S_0(-\zeta)\big)} = (\pa_\zeta S_0 )(-\zeta) = \pa_\zeta \big(1-S_0(-\zeta)\big), \end{aligned}$$ whence $1-S_0(-\zeta)$ satisfies the differential equation in . Since we obviously have $1-S_0(0)=\frac{1}{2}$, we see that $1-S_0(-\zeta)$ is a solution of . Since the solution of this initial value problem is unique, we infer that $S_0(\zeta)=1-S_0(-\zeta)$ holds, which implies $S'_0(\zeta)=S'_0(-\zeta)$. To prove , note that implies for $\zeta > 0$ $$\begin{aligned} \int_{-\infty}^\zeta S_0(\vartheta)\, d \vartheta &= \int_{-\infty}^{-\zeta} S_0(\vartheta)\, d\vartheta + \int_{-\zeta}^\zeta S_0(\vartheta)\, d\vartheta \\ &= \int_{-\infty}^{-\zeta} S_0(\vartheta)\, d\vartheta + \int_0^\zeta S_0(\vartheta)+ S_0(-\vartheta)\, d\vartheta\\ &= \int_{-\infty}^{-\zeta} S_0(\vartheta)\, d\vartheta + \int_0^\zeta S_0(\vartheta)+ \big(1-S_0(\vartheta)\big)\,d\vartheta = \int_{-\infty}^{-\zeta} S_0(\vartheta)\, d\vartheta + \zeta. \end{aligned}$$ From this equation we immediately obtain . The inequality follows from and from , which yield $$\Big\| \int_{-\infty}^\zeta S_0(\vartheta)\, d \vartheta - \zeta^+ \Big\| = \Big\| \int_{-\infty}^{-\|\zeta\|} S_0(\vartheta)\, d\vartheta \Big\| \leq \int_{-\infty}^{-\|\zeta\|} K_1 e^{-a\|\vartheta\|} d\vartheta = \frac{K_1}{a} e^{-a \|\zeta\|}.$$ For the proof of note that ${\hat{S}}(\zeta) = 1$ for $\zeta > 0$. Consequently, the mean value theorem and together imply for $\zeta > 0$ that $$\| \hat{\psi}''\big(S_0(\zeta)\big) - \hat{\psi}''\big(\hat{S}(\zeta) \big)\| = \| \hat{\psi}''\big(S_0(\zeta)\big) - \hat{\psi}''(1)\| \leq \|\hat{\psi}'''(r^) (S_0(\zeta) - 1)\| \leq CK_2 e^{- a \|\zeta\|},$$ with a suitable number $r^$ between $S_0(\zeta)$ and $1$. For $\zeta < 0$ an analogous estimate is obtained using and noting that ${\hat{S}}(\zeta) = 0$ if $\zeta < 0$. The functions $\boldsymbol{S_1}$ and $\boldsymbol{S_2}$ {#S4.2} ------------------------------------------------------- The solutions $S_1$ and $S_2$ of the second and third boundary value problems are studied in this section. The second problem determining $S_1$ is given by the equations , – , the third problem, which determines $S_2$, consists of the equations , , , . The properties of $S_1$ and $S_2$, which we need in Section \[S5\], are summarized in the next two theorems. To state the first theorem we need the function $\rho_1:\Gm \ti \R \ra \R$, which is defined by $$\label{E4.10} \rho_1(t,\eta,\zeta) = \ov{\ve}:\hat{T}^{(-)}(t,\eta)\, \varphi_- (\zeta) + \ov{\ve}:\hat{T}^{(+)}(t,\eta)\, \varphi_+ (\zeta),$$ with $\varphi_\pm$ introduced in . \[T4.4\] Assume that $\hat{\psi}$ belongs to $C^5([0,1],\R)$ and satisfies the assumptions and the symmetry condition . Suppose that the function $s_0 = s_0(t,\eta)$ in is given by . Let $S_0$ be the solution of the boundary value problem , , which exists by . Then for every $(t, \eta) \in \Gamma$ there is a unique solution $\zeta \mapsto S_1 (t, \eta, \zeta):\R \ra \R$ of the boundary value problem , – . The function $S_1$ belongs to the space $C^2( \R, C^2(\Gm,\R))$. Moreover, there are constants $K_1,\ldots, K_3$ such that for the constant $a$ defined in and for all $(t,\eta,\zeta) \in \Gm\ti \R$ the estimates $$\begin{aligned} \label{E4.11} \\|D^{\alpha}_{(t, \eta)} S_1\\|_{L^{\infty}(\Gamma \times \R)} &\leq& K_1, \qquad \|\alpha\| \leq 2, \\ \label{E4.12} \big\| \pa^j_\zeta D^\al_{(t,\eta)} \big(S_1 (t, \eta, \zeta) - \rho_1(t,\eta,\zeta) \big) \big\| &\leq& K_2\, e^{-a \|\zeta\|}, \quad 0 \leq j,\|\al\| \leq 2, \\ \label{E4.14} \big\| \pa^j_\zeta D^\al_{(t,\eta)} S_1(t,\eta,\zeta) \big\| &\leq& K_3\, e^{-a \|\zeta\|}, \quad \|\al\| \leq 2,\ j=1,2, \end{aligned}$$ hold. We do not give the proof of this theorem, since it is almost the same as the proof of Theorem 1.2 in [@CMT2011]. Morover, it is obtained from the proof of the following theorem by simplification. The main difference between the two proofs is that the right hand side $F_1$ of the differential equation for $S_1$ is bounded, whereas the right hand side $F_2$ of the differential equation for $S_2$ grows linearly for $\zeta \ra \pm \infty$. \[T4.5\] Assume that $\hat{\psi}$ satisfies the assumptions given in the . Let $S_0$ be the solution of the boundary value problem , , and let $S_1$ be the solution of the boundary value problem , – . Suppose that the function $s_1 = s_1(t,\eta)$ in satisfies with $s_{10}$, $s_{11}$ given in , . \(i) Then for every $(t, \eta) \in \Gamma$ there is a unique solution $\zeta \mapsto S_2 (t, \eta, \zeta):\R \ra \R$ of the boundary value problem , , , . The function $S_2$ belongs to $C^2( \R, C^2(\Gm,\R))$, and there are constants $K_4,\ldots, K_6$ such that for the constant $a$ defined in and for all $(t,\eta,\zeta) \in \Gm\ti \R$ the estimates $$\begin{aligned} \label{E4.11a} \|\pa^j_\zeta D^{\alpha}_{(t, \eta, \zeta)} S_2(t,\eta,\zeta)\| &\leq& K_4 (1+\|\zeta\|)^{1-j}, \quad \|\alpha\| \leq 2,\ j=0,1, \\ \label{E4.12a} \big\| \pa_\zeta^j D^{\alpha}_{(t, \eta)} \big( S_2 (t, \eta, \zeta) - \rho_2(t,\eta,\zeta) \big) \big\| &\leq& K_5 (1 + \|\zeta\|)\, e^{-a \|\zeta\|}, \quad 0 \leq j,\|\al\| \leq 2, \\ \label{E4.13a} \big\| \pa^2_\zeta D^\al_{(t,\eta)} S_2(t,\eta,\zeta) \big\| &\leq& K_6 (1 + \|\zeta\|)\, e^{-a \|\zeta\|},\qquad \|\al\| \leq 2, \end{aligned}$$ hold, where $\rho_2$ is defined in . \(ii) $S_2$ is the only solution of the differential equation with $F_2$ given by , which satisfies and for which constants $C,\theta > 0$ exist such that $$\label{E4.14a} \| S_2(t,\eta,\zeta)\| \leq C e^{ (a - \theta) \|\zeta\| }, \qquad \zeta \in \R,$$ holds. Proof of Theorem 4.5 {#S4.3} -------------------- In this section we give the proof of , which is divided into five parts: #### (I) Reduction of the boundary value problem for $\boldsymbol{S_2}$ to a problem in $\boldsymbol{L^2}$. With $\rho_2$ defined in we make the ansatz $$\label{E4.16} S_2(t,\eta,\zeta) = w(t,\eta,\zeta) + \rho_2(t,\eta,\zeta).$$ Insertion of this ansatz into the equations and , shows that $S_2$ is a solution of the problem given by these equations if and only if $w$ solves the equations $$\begin{gathered} \hat{\psi}''\big( S_0(\zeta)\big) w(t,\eta,\zeta) - \pa^2_\zeta w(t,\eta,\zeta) = F_2(t,\eta,\zeta) + F_3(t,\eta,\zeta), \label{E4.17} \\ w(t,\eta,0) = 0, \label{E4.18} \\ \lim_{\zeta \ra \pm \infty} w(t,\eta,\zeta) = 0, \label{E4.19}\end{gathered}$$ where $F_2$ is given by and where $$\label{E4.20} F_3 = - (\hat{\psi}''(S_0) - \pa^2_\zeta ) \rho_2.$$ To get we used that $\vp_+(0) = \vp_-(0) = 0$, which by implies $\rho_2(t,\eta,0) = 0$. To show that the solution $S_2$ of the problem and , exists, it therefore suffices to prove that the reduced problem – has a solution. #### (II) Spectral theory For this proof note that $\hat{\psi}''(S_0)-\pa^2_\zeta$ is a linear self-adjoint differential operator in $L^2(\R)$. From the spectral theory of such operators we know that the continuous spectrum of $\hat{\psi}''(S_0)-\pa^2_\zeta$ is contained in the interval $[a_0,\infty)$, where $$a_0= \min\Big\{\lim\limits_{\zeta \to -\infty}\hat{\psi}''\big(S_0(\zeta)\big), \lim\limits_{\zeta \to \infty} \hat{\psi}''\big(S_0(\zeta)\big)\Big\},$$ and that the part of the spectrum in $(-\infty,a_0)$ is a pure point spectrum. yields $\lim\limits_{\zeta \to -\infty} \hat{\psi}'' \big(S_0(\zeta)\big) = \hat{\psi}''(0)$, $\lim\limits_{\zeta \to \infty} \hat{\psi}''\big(S_0(\zeta)\big)= \hat{\psi}''(1)$, hence the assumption implies $a_0=a^2 >0$. Therefore $0$ does not belong to the continuous spectrum. From the spectral theory we also know that for every $\om \in \C$, which is not in the continuous spectrum, the differential equation $\big(\hat{\psi}''(S_0)-\pa^2_\zeta\big) w -\om w=f$ has a solution $w \in L^2(\R)$, if and only if $f \in L^2(\R)$ is orthogonal to the kernel of the operator $\psi''(S_0)- \pa^2_\zeta -\om$. This implies in particular, that for every $(t,\eta)\in\Gamma$ the differential equation has a solution $w(t,\eta,\cdot) \in L^2(\R)$, if the right hand side $\zeta \mapsto f(\zeta)=F_2 (t,\eta,\zeta)+ F_3(t,\eta,\zeta)$ belongs to $L^2(\R)$ and is orthogonal to the kernel of the operator $\hat{\psi}''(S_0)-\pa^2_\zeta$. To show that the problem – has a solution, we therefore verify in the next two parts of the proof that $F_2+F_3$ satisfies these two conditions. #### (III) The asymptotic behavior of $\boldsymbol{F_2 + F_3}$ at infinity. We first show that the right hand side $F_2 + F_3$ of decays exponentially at $\pm \infty$, which implies that $F_2 + F_3 \in L^2(\R)$. To simplify the notation we define $$\ov{\vp}_+(\zeta) = \vp_+(\zeta) \hat{\psi}''\big (S_0(\zeta)\big), \qquad \ov{\vp}_-(\zeta) = \vp_-(\zeta) \hat{\psi}''\big (S_0(\zeta)\big), \label{E4.21}$$ with $\vp_+$, $\vp_-$ given in . For these functions we obtain from that $$\begin{aligned} &\|\ov{\vp}_- - \hat{\psi}''(0) \vp_- \| = \vp_-(\zeta)\, \|\hat{\psi}''(S_0) - \hat{\psi}''(0) \| \leq C K_4 e^{-a\|\zeta\|}, \label{E4.22} \\ &\|\ov{\vp}_+ - \hat{\psi}''(1) \vp_+\| = \vp_+(\zeta)\, \|\hat{\psi}''(S_0) - \hat{\psi}''(1)\| \leq C K_4 e^{-a\|\zeta\|}, \label{E4.23}\end{aligned}$$ for all $\zeta \in \R$. Since $ \hat{\psi}''(0) \vp_-(\zeta) = 1$ for $\zeta \leq -2$ and $\hat{\psi}''(1) \vp_+(\zeta) = 1$ for $\zeta \geq 2$, these estimates imply $$\begin{aligned} &\|1 - \ov{\vp}_-(\zeta) \| \leq C K_4 e^{-a \|\zeta\|}, & -\infty < \zeta \leq 0. \label{E4.24} \\ &\|1 - \ov{\vp}_+(\zeta) \| \leq C K_4 e^{-a \|\zeta\|}, & 0 \leq \zeta < \infty, \label{E4.25} \\ &\|1 - \ov{\vp}_-(\zeta) - \ov{\vp}_+(\zeta)\| \leq C K_4 e^{-a \|\zeta\|}, & \zeta \in \R. \label{E4.26}\end{aligned}$$ To get the last estimate we combined the first two estimates and noted that $\ov{\vp}_-(\zeta) = 0$ for $\zeta \geq -1$ and $\ov{\vp}_+(\zeta) = 0$ for $\zeta \leq 1$. Note that by , and the function $F_2 + F_3$ can be decomposed in the form $$\label{E4.27} F_2 + F_3 = F_2 - \hat{\psi}''(S_0) \rho_2 + \pa^2_\zeta \rho_2 = \sum_{j=1}^5 I_j ,$$ where $$\begin{aligned} I_1 &=& {\check{\sigma}}(0) + \ov{\ve}: [{\hat{T}}]\,S_1 - \ov{\ve}: {\check{T}}^{(-)}\, \ov{\vp}_- - \ov{\ve}: {\check{T}}^{(+)}\, \ov{\vp}_+ \,, \label{E4.28} \\[1ex] I_2 &=& -\frac{\hat{\psi}'''(S_0)}{2} S_1^2 + \frac{\hat{\psi}'''(0)}{2} \Big( \frac{ \ov{\ve}: {\hat{T}}^{(-)}} {\hat{\psi}''(0)} \Big)^2\, \ov{\vp}_- + \frac{\hat{\psi}'''(1)}{2} \Big( \frac{ \ov{\ve}: {\hat{T}}^{(+)}} {\hat{\psi}''(1)} \Big)^2\, \ov{\vp}_+ \,, \label{E4.29} \\[1ex] I_3 &=& \la^{1/2}\, \hat{\si}'(0) \zeta\, \big(1-\ov{\vp}_- - \ov{\vp}_+ \big), \label{E4.30} \\[1ex] I_4 &=& \la^{1/2}\,\ov{\ve}: D \ve(a^ \otimes n + \na_x u^) \Big( \int_{-\infty}^\zeta S_0(\vta) \, d\vta - \zeta^+ \ov{\vp}_+ \Big), \label{E4.31} \\[1ex] I_5 &=& - \frac{1}{c_1}\, \ov{\ve} : \langle {\hat{T}}\rangle S_1' + \Big( \frac{s_1} {c} - \la \ka'(0) \zeta \Big) S_0' + \pa^2_\zeta \rho_2\,. \label{E4.32}\end{aligned}$$ We show that everyone of these terms decays to zero for $\zeta \ra \pm \infty$. To verify this for the first term we insert and into , which results in $$I_1 = {\check{\sigma}}(0) + \ov{\ve}:[{\hat{T}}]\, S_1 - {\check{\sigma}}(0) ( \ov{\vp}_+ + \ov{\vp}_- ) - \ov{\ve}:[{\hat{T}}]\, \Big( \frac{\ov{\ve}:{\hat{T}}^{(-)} } {\hat{\psi}''(0)} \ov{\vp}_- + \frac{\ov{\ve}:{\hat{T}}^{(+)} } {\hat{\psi}''(1)} \ov{\vp}_+ \Big).$$ We introduce the terms $\hat{\psi}''(0) \vp_-$ and $\hat{\psi}''(1) \vp_+$ into this equation. Noting the definition of $\rho_1$ in , this leads to $$\begin{aligned} \|I_1\| & \leq \big\| {\check{\sigma}}(0)\, (1 - \ov{\vp}_+ + \ov{\vp}_- ) + \ov{\ve}:[{\hat{T}}]\, \big( S_1 - \rho_1 \big) \big\| \nn \\ & \mbox{} + \Big\| \ov{\ve}:[{\hat{T}}]\, \Big( \frac{\ov{\ve}:{\hat{T}}^{(-)} } {\hat{\psi}''(0)} (\ov{\vp}_- - \hat{\psi}''(0) \vp_- ) + \frac{\ov{\ve}:{\hat{T}}^{(+)} } {\hat{\psi}''(1)} (\ov{\vp}_+ - \hat{\psi}''(1) \vp_+ ) \Big) \Big\| \leq C e^{-a \|\zeta\|}, \label{E4.33}\end{aligned}$$ for all $\zeta \in \R$, where we applied the estimates , , and . Next we estimate $I_2$. By definition we have $\ov{\vp}_+(\zeta) = 0$ for $\zeta \leq 1$. From we thus have on the half axis $-\infty < \zeta \leq 0$ that $$\begin{aligned} I_2 &= -\frac{\hat{\psi}'''(S_0)}{2} S_1^2 + \frac{\hat{\psi}'''(0)}{2} \Big( \frac{ \ov{\ve}: {\hat{T}}^{(-)}} {\hat{\psi}''(0)} \Big)^2 \ov{\vp}_- \nn \\ &= \frac{\hat{\psi}'''(0) - \hat{\psi}'''(S_0)}{2} S_1^2 - \frac{\hat{\psi}'''(0)}{2} S_1^2 (1 - \ov{\vp}_-) - \frac{\hat{\psi}'''(0)}{2} \Big( S_1^2 - \Big( \frac{ \ov{\ve}: {\hat{T}}^{(-)}} {\hat{\psi}''(0)} \Big)^2 \Big) \ov{\vp}_- \nn \\ & = I_{21} + I_{22} + I_{23}. \label{E4.34}\end{aligned}$$ To estimate $I_{21}$ we apply the mean value theorem to $\hat{\psi}'''$ and use and , to estimate $I_{22}$ we use and . The result is $$\label{E4.35} \|I_{21} + I_{22}\| \leq C K_1 e^{-a\|\zeta\|}, \quad -\infty < \zeta \leq 0.$$ To estimate $I_{23}$ note that by , and we have for $-\infty < \zeta \leq -2$ that $$\rho_1(t,\eta,\zeta) = \vp_-(\zeta) \big( \ov{\ve}:\hat{T}^{(-)}(t,\eta) \big) = \frac{ \ov{\ve}: {\hat{T}}^{(-)}(t,\eta) } {\hat{\psi}''(0)}.$$ With this equation we infer from and with $\al,j=0$ that $$\label{E4.36} \|I_{23}\| = \Big\| \frac{\hat{\psi}'''(0)}{2} \Big( S_1- \frac{ \ov{\ve}: {\hat{T}}^{(-)}} {\hat{\psi}''(0)} \Big) \Big( S_1 + \frac{ \ov{\ve}: {\hat{T}}^{(-)}} {\hat{\psi}''(0)} \Big) \ov{\vp}_- \Big\| \leq C e^{-a\|\zeta\|}, \quad -\infty < \zeta \leq 0.$$ – together imply that $\|I_2(\zeta)\| \leq C e^{-a\|\zeta\|}$ for $-\infty < \zeta \leq 0$. On the half axis $0 \leq \zeta <\infty$ we estimate $I_2$ analogously. This proves that $$\label{E4.37} \|I_2\| \leq C e^{-a\|\zeta\|}, \qquad -\infty < \zeta < \infty.$$ The estimate for $I_3$ is obtained by application of to , which immediately yields $$\label{E4.38} \| I_3 \| \leq C (1+\|\zeta\|) e^{-a\|\zeta\|}, \qquad -\infty < \zeta < \infty.$$ To study the asymptotic behavior of $I_4$ note that and together imply $$\begin{gathered} \Big\|\int_{-\infty}^\zeta S_0(\vartheta)\, d\vartheta - \zeta^+ \ov{\varphi}_+(\zeta)\Big\| \leq \Big\|\int_{-\infty}^\zeta S_0(\vartheta)\, d\vartheta - \zeta^+ \Big\| + \zeta^+ \|1 - \ov{\varphi}_+(\zeta) \| \\ \leq \Big(\frac{1}{a} K_1+ \zeta^+ C K_4\Big) e^{-a\|\zeta\|}. \end{gathered}$$ Insertion of this inequality into results in $$\label{E4.39} \| I_4 \| \leq C (1+\|\zeta\|)\, e^{-a\|\zeta\|}.$$ It remains to investigate $I_5$. Note first that the third term on the right hand side of satisfies $$\label{E4.39a} \pa^2_\zeta \rho (t,\eta,\zeta) = 0, \qquad \mbox{for } \|\zeta\| \geq 2.$$ To show this it suffices to remark that the functions $\varphi_\pm$ are constant on the intervals $(-\infty,-2)$ and $(2,\infty)$, from which we see by inspection of that on these intervals the function $\zeta \mapsto \rho_2(t,\eta,\zeta)$ is a sum of constant and linear terms, whence follows. If we estimate the first term on the right hand side of by employing with $\al = 0$, $j=1$ and the second term by using , we obtain together with that $$\label{E4.40} \| I_5 \| \leq C(1 + \|\zeta\|) e^{-a\|\zeta\|}.$$ We combine , , – and to derive the estimate $$\label{E4.41} \| F_2 (t,\eta,\zeta) + F_3(t,\eta,\zeta) \| \leq C (1+\|\zeta\|) e^{-a\|\zeta\|}, \qquad \mbox{for all } \zeta \in \R,$$ which shows in particular that the right hand side of belongs to $L^2(\R)$. #### (IV) The orthogonality condition determining $\boldsymbol{s_1}$. Next we must show that the right hand side of is orthogonal to the kernel of $\hat{\psi}''(S_0)-\pa^2_\zeta$. This kernel is different from $\{0\}$, since $S'_0$ belongs to the kernel. This is immediately seen by differentiation of , which yields $$\label{E4.42} \hat{\psi}'' \big(S_0(\zeta)\big) S'_0(\vartheta)- \pa^2_\zeta S'_0(\zeta)=0.$$ Since by the function $S'_0$ is in the domain of definition of $\hat{\psi}''(S_0)-\pa^2_\zeta$, it belongs to the kernel of this operator. The theory of linear ordinary differential equations of second order implies now that the kernel is one-dimensional, hence every function from the kernel is a multiple of $S'_0$. Therefore the right hand side $F_2+F_3$ of is orthogonal to the kernel if it is orthogonal to $S'_0$. Note that the integrals $\int_{-\infty}^\infty F_2 S'_0 \,d\zeta$ and $\int_{-\infty}^\infty F_3 S'_0\,d\zeta$ both exist, since $F_2$ and $F_3$ grow at most linearly for $\zeta \to \pm \infty$, whereas by the function $S'_0$ decays exponentially at $\pm \infty$. Therefore we obtain from and by partial integration that $$\begin{gathered} \label{E4.43} \int_{-\infty}^\infty (F_2+F_3) S'_0 \, d\zeta = \int_{-\infty}^\infty F_2 S'_0 \, d\zeta - \int_{-\infty}^\infty \Big( \big(\hat{\psi}''(S_0)-\pa^2_\zeta\big) \rho_2 \Big) S'_0\, d\zeta \\ = \int_{-\infty}^\infty F_2 S'_0 \, d\zeta- \int_{-\infty}^\infty \rho_2 \big(\hat{\psi}''(S_0)-\pa^2_\zeta\big) S'_0 \, d\zeta = \int_{-\infty}^\infty F_2 S'_0 \, d\zeta. \end{gathered}$$ To study the last integral on the right hand side note that by the function $S'_0$ is even, which implies that $$\label{E4.44} \int_{-\infty}^\infty \hat{\sigma}' (t,\eta,0) \zeta\, S'_0(\zeta)\,d\zeta=0, \quad \mbox{and} \quad \int_{-\infty}^\infty \ka'(t,\eta,0) \zeta\, \big(S'_0(\zeta) \big)^2 d\zeta = 0.$$ Moreover, since by Lemma \[L4.1\] the function $S'_0$ satisfies , we obtain by substitution of $\vartheta=S_0(\zeta)$ $$\label{E4.45} \int_{-\infty}^\infty S'_0(\zeta) S'_0 (\zeta)\, d\zeta= \int_{-\infty}^\infty \sqrt{2 \hat{\psi}\big(S_0(\zeta)\big)}\, S'_0(\zeta)\, d\zeta = \int_0^1 \sqrt{2 \hat{\psi}(\vartheta)} \, d\vartheta=c_1,$$ where the last equality sign holds by definition of $c_1$ in . Finally, by partial integration, $$\begin{gathered} \label{E4.46} \int_{-\infty}^\infty \int_{-\infty}^\zeta S_0 (\vartheta) \, d\vartheta\, S'_0 (\zeta)\,d\zeta= \lim_{\zeta_1 \to\infty} \Big(\int_{-\infty}^{\zeta_1} S_0(\vartheta)\, d\vartheta\, S_0(\zeta_1)- \int_{-\infty}^{\zeta_1}S_0 (\zeta)^2\,\zeta\Big) \\ = \lim_{\zeta_1 \to\infty} \int_{-\infty}^{\zeta_1} S_0(\zeta) \big(S_0(\zeta_1)-S_0(\zeta)\big)\, d\zeta \\ = \int_{-\infty}^\infty S_0(\zeta) \big(1-S_0(\zeta)\big)\, d\zeta = \int_{-\infty}^\infty S_0(\zeta) S_0(-\zeta)\, d\zeta. \end{gathered}$$ In the second last step we used that $S_0$ is increasing. The equality sign thus follows from the theorem of Beppo Levi. The last equality sign is obtained from . The equations and – yield $$\begin{gathered} \int_{-\infty}^\infty F_2 (t,\eta,\zeta) S'_0(\zeta)\, d\zeta = \int_{-\infty}^\infty\big(\check{\sigma}(0)+ \ov\ve:[\hat{T}]S_1 \big) S'_0\, d\zeta - \frac{1}{c_1} \ov\ve:\langle \hat{T} \rangle \int_{-\infty}^\infty S'_1 S'_0 \, d\zeta \\ \mbox{}- \frac{1}{2} \int_{-\infty}^\infty \hat{\psi}''' (S_0) S_1^2 S'_0 \, d\zeta \mbox{}+ \lambda^{1/2}\, \ov\ve: D\ve (a^ \otimes n+ \na_\Gamma u^) \int_{-\infty}^\infty \int_{-\infty}^\zeta S_0 (\vartheta) \, d\vartheta\ S_0'(\zeta) \, d\zeta \\ \mbox{}+ \int_{-\infty}^\infty \Big(\la^{1/2} \hat{\si}'(0)\zeta + \big( \frac{s_1}{c} - \la \ka'(0)\zeta\big) S'_0 \Big) S'_0(\zeta) d\zeta \\ = {\check{\sigma}}(0) + \ov\ve:[{\hat{T}}] \int_{-\infty}^\infty S_1 S'_0\, d\zeta - \frac{1}{c_1} \ov\ve:\langle \hat{T} \rangle \int_{-\infty}^\infty S'_1 S'_0 \, d\zeta - \frac{1}{2} \int_{-\infty}^\infty \hat{\psi}''' (S_0) S_1^2 S'_0 \, d\zeta \\ \mbox{} + \la^{1/2}\, \ov\ve: D\ve (a^ \otimes n+ \na_\Gamma u^) \int_{-\infty}^\infty S_0(\zeta) S_0(-\zeta)\, d\zeta + c_1 \frac{s_1}{c}\,. \\ = - \frac{c_1}{c} s_{10} - \la^{1/2} \frac{c_1}{c} s_{11} + \frac{c_1} {c} s_1 = \frac{c_1}{c} (-s_{10} - \la^{1/2} s_{11} + s_1) . \end{gathered}$$ To get the second last equality sign we inserted for ${\check{\sigma}}(0)$ and used , . The right hand side of this equation vanishes if and only if $s_1$ satisfies . From we thus infer that $F_2+F_3$ is orthogonal to the kernel of $\hat{\psi}''(S_0)-\pa^2_\zeta$ if and only if holds. Consequently, from part (II) of the proof we conclude that the differential equation has a solution $w$ in $L^2(\R)$ if and only if $s_1$ satisfies with $s_{10}$ and $s_{11}$ given in , . In fact, there is exactly one such $w$, which also satisfies . To prove this assume that $\tilde{w} \in L^2(\R)$ is a special solution of . Then we obtain every solution contained in $L^2(\R)$ in the form $w= \tilde{w}+\beta S'_0$ with an arbitrary constant $\beta \in \R$. Since and yield $$S'_0(0)= \sqrt{2 \hat{\psi}\big(S_0(0)\big)}= \sqrt{2 \hat{\psi}\Big(\frac{1}{2}\Big)} >0,$$ we can choose $\beta$ such that $$w(t,\eta,0)= \tilde{w}(t,\eta,0)+ \beta S'_0(0)=0,$$ which is . This equation determines $\beta$ uniquely, hence $w$ is the unique solution of and in $L^2(\R)$. #### (V) Existence of the solution, estimates (4.18) – (4.20). We show next that this function $w$ satisfies . To this end we need the following \[L4.7\] Let $\hat{a}_- >0$, $\hat{a}_+ > 0$ and set $\hat{a} = \sqrt{\min\{\hat{a}_-, \hat{a}_+ \}}$. Let $g:\R \ra \R$ be a smooth function and let $f:\R \ra \R$ be a continuous function such that $$\begin{aligned} \| g(\zeta) - \hat{a}_-\| &\leq& C e^{- \hat{a} \|\zeta\|}, \hspace{11ex} \mbox{for } \zeta < 0, \label{E4.54} \\ \| g(\zeta) - \hat{a}_+\| &\leq& C e^{- \hat{a} \|\zeta\|}, \hspace{11ex} \mbox{for } \zeta > 0, \label{E4.55} \\ \| f (\zeta) \| &\leq& C(1+\|\zeta\|) e^{- \hat{a} \|\zeta\|}, \quad \mbox{for } \zeta \in \R. \label{E4.56}\end{aligned}$$ Let $\hat{w}$ be a solution of $$\label{E4.57} g(\zeta) \hat{w}(\zeta) - \pa^2 \hat{w}(\zeta) = f(\zeta), \qquad \zeta \in \R.$$ (i) Then $\hat{w}$ belongs to the space $C^2(\R)$. If $\hat{w} \in L^2(\R)$, then there is $C > 0$ such that $$\label{E4.58} \|\pa^j_\zeta \hat{w}(\zeta)\| \leq C(1 + \|\zeta\|) e^{-\hat{a} \|\zeta\|}, \qquad \mbox{for all }\zeta \in \R, \mbox{ for } j=0,1,2.$$ (ii) If there are $C,\theta > 0$ such that $$\label{E4.59} \|\hat{w}(\zeta)\| \leq C e^{ (\hat{a} - \theta) \|\zeta\|}$$ holds for all $\zeta \in \R$, then $\hat{w} \in L^2(\R)$. This is a standard result from the theory of ordinary differential equations, and we omit the proof. To show that $w$ satisfies , we apply this lemma with $\hat{a}_- = \hat{\psi}''(0)$, $\hat{a}_+ = \hat{\psi}''(1)$, $g(\zeta) = \hat{\psi}''\big(S_0(\zeta)\big)$ and $f(\zeta) = F_2(t,\eta,\zeta) + F_3(t,\eta,\zeta)$. Then we have $\hat{a} = \sqrt{\min\{\hat{\psi}''(0),\hat{\psi}''(1) \}} = a$, by , and from and we see that – hold for this choice of functions and constants. Moreover, with this choice of functions the differential equation is equal to . Since $w \in L^2(\R)$ is a solution of , we see that all assumptions for part (i) of are satisfied, hence holds for $w$, which means that $$\label{E4.60} \| \pa^j_\zeta w(\zeta)\| \leq C(1+ \|\zeta\|) e^{- a \|\zeta\| }, \qquad \zeta \in \R,\ j=0,1,2,$$ and this in particular implies that $w$ satisfies . We have now found a unique solution $w \in L^2(\R)$ of – . By part (I) of this proof this means that $S_2$ given by is a solution of the boundary value problem , , and . Since by we have $w = S_2 - \rho_2$, the inequality shows that $S_2$ satisfies for $\al = 0$. To verify that $S_2$ satisfies for $\al \neq 0$, it must first be shown that $S_2$ is two times continuously differentiable with respect to $(t,\eta)$. This follows if we can show that $w = S_2 - \rho_2$ is two times continuoulsly differentiable with respect to $(t,\eta)$, since by our regularity assumptions the function $\rho_2$ has this differentiability property. To prove this differentiability of $w$, we write , as a perturbation problem for the linear equation $$A w = f(t,\eta)$$ in $L^2(\R)$, where $A = (\hat{\psi}''(S_0) - \pa^2_\zeta)$ is the linear differential operator on the left hand side of and $f(t,\eta) = F_2(t,\eta,\cdot) + F_3(t,\eta,\cdot) \in L^2(\R)$ is the function on the right hand side of , which depends two times continuously differentiable on $(t,\eta)$ and satisfies the estimate for every $(t,\eta) \in \Gm$. Since $0$ is an eigenvalue of $A$ and since $f(t,\eta)$ is orthogonal to the kernel of $A$ for every $(t,\eta)$, this linear equation has infinitely many solutions and the solution set is affine. The condition defines a linear subspace, which is closed in the Sobolev space $H^1(\R)$, and which intersects the solution set in exactly one point $w$, which is the solution of , . To the problem set in this way we can apply the pertubation theory of linear operators. The theory yields that $w$ is two times continuously differentiable with respect to $(t,\eta)$. We avoid the details but refer to standard texts on the pertubation theory of linear operators, for example [@Kato]. With this knowledge we can derive the estimate for $\al \neq 0$ by applying the differential operator $D^\al_{(t,\eta)}$ with $1 \leq \|\al\| \leq 2$ to the differential equation and obtain $$\label{E4.61} \hat{\psi}'' (S_0) (D^\al_{(t,\eta)} w) - \pa^2_\zeta (D^\al_{(t,\eta)} w) = D^\al_{(t,\eta)} (F_2 + F_3).$$ This is a differential equation for the function $D^\al_{(t,\eta)} w$ with right hand side satisfying the estimate $$\label{E4.62} \|D^\al_{(t,\eta)} (F_2 + F_3) \| \leq C (1 + \|\zeta\|) e^{-a \|\zeta\|}.$$ The proof of this estimate proceeds in the same way as the proof of the corresponding estimate for $\al = 0$, which we gave in part (III). Essentially one has to replace the terms appearing in $F_2$ and $F_3$, which depend on $(t,\eta)$, by their derivatives. To avoid repetition of many technical details, we omit this proof. The differential equation has the same form as the differential equation . From we see that the assumption holds, hence we can apply (i) to this differential equation, from which we see that $w$ belongs to the space $C^2(\R,C^2(\Gm,\R))$ and that the inequality holds with $\hat{w}$ replaced by $D^\al_{(t,\eta)}w$, whence we have $$\| \pa^j_\zeta D^\al_{(t,\eta)}w(\zeta)\| \leq C(1+ \|\zeta\|) e^{- a \|\zeta\| }, \qquad \zeta \in \R,\ 0 \leq j \leq 2,\ 1 \leq \|\al\| \leq 2.$$ Since $w = S_2 - \rho_2$, this is inequality with $\al \neq 0$. Therefore we proved that holds for all $0\leq \|\al\| \leq 2$. The inequality is a consequence of and of , the inequality follows by combination of with the estimate $$\| \pa^j_\zeta D^\al_{(t,\eta)} \rho_2(t,\eta,\zeta) \| \leq C(1 + \|\zeta\|)^{1-j}, \qquad j = 0,1,$$ which is seen to hold by inspection of . We have now proved statement (i) of , and it remains to verify (ii). That is, we have to show that $S_2$ is the only solution of , , satisfying . Indeed, from it follows that $S_2$ satisfies . Assume that $S_2^$ is a second solution satisfying . Then $\hat{w} = S_2 - S_2^$ fulfills and the equation $$\hat{\psi}''(S_0) \hat{w} - \pa^2 \hat{w} = 0.$$ (ii) thus yields $\hat{w} \in L^2(\R)$. Consequently, by we have $S_2^ = w + \hat{w} + \rho_2$, where $w + \hat{w} \in L^2(\R)$ is a solution of , . At the end of part (IV) of this proof we showed that $w$ is the only solution of , in $L^2(\R)$, whence $\hat{w} = 0$, hence $S_2^* = S_2$. The proof of is complete. Proof of the estimates (\[E2.37a\]) – (\[E2.38b\]) in {#S5} ====================================================== The proof of – is straightforward: We insert the function $(u^{(\mu)},T^{(\mu)},S^{(\mu)})$ defined in Section \[S3\] into the model equations and and compute the residues. However, the necessary computations are long. Therefore we divide them into four parts: In Section \[S5.1\] we compute for the functions $T^{(\mu)}_1$ and $T^{(\mu)}_2$ from the inner and outer expansions of $T^{(\mu)}$ the residues $\div_x T^{(\mu)}_1 + {\sf b}$ and $\div_x T^{(\mu)}_2 + {\sf b}$ separately. Likewise, in Section \[S5.2\] we insert the inner expansion $(u_1^{(\mu)},T_1^{(\mu)},S_1^{(\mu)})$ and the outer expansion $(u_2^{(\mu)},T_2^{(\mu)},S_2^{(\mu)})$ into and compute and estimate the residues separately. With these residues we can prove – in the regions $Q^{(\mu\la)}_{\rm inn}$ and $Q^{(\mu\la)}_{\rm out}$, but in the matching region $Q^{(\mu\la)}_{\rm match}$ we need auxiliary estimates, which are stated in Section \[S5.3\]. All the estimates are put together in Section \[S5.4\] to complete the proof. Asymptotic expansion of $\boldsymbol{\div_x T^{(\mu)} + {\sf b}}$ {#S5.1} ------------------------------------------------------------------ \[L5.1\] Let $({\hat{u}},{\hat{T}})$ be the solution of the transmission problem – . With the splitting of ${\hat{u}}$ the stress tensor field ${\hat{T}}$ satisfies in the neighborhood ${\cal U}_\da$ of $\Gm$ $$\label{EhatT} \hat{T} = [{\hat{T}}] {\hat{S}}+ D\ve(\na_x {\hat{v}}) + D \ve \big( a^* \otimes n + \na_{\Gm_\xi} u^* \big) \xi^+ + D \ve (\na_{\Gm_\xi} a^) \frac12 (\xi^+)^2 .$$ With $\hat{\si}'$ defined in and with a remainder term $R_{\ov{\ve}:{\hat{T}}} \in L^{\infty}({\cal U}_\da)$ we have for $(t,\eta,\xi) \in {\cal U}_\da$ $$\begin{aligned} \ov{\ve}:{\hat{T}}(t,\eta,\xi) &= \ov{\ve}:{\hat{T}}^{(+)} (t,\eta) + \hat{\si}'(t,\eta,0)\, \xi \nn \\ \mbox{} + \ov{\ve}&:D\ve \big(a^(t,\eta) \otimes n(t,\eta) + \na_\Gm u^(t,\eta) \big) \, \xi + R_{\ov{\ve}:{\hat{T}}}(t,\eta,\xi)\, \xi^2, \quad \xi > 0, \label{E5.epsT+} \\ \ov{\ve}:{\hat{T}}(t,\eta,\xi) &= \ov{\ve}:{\hat{T}}^{(-)} (t,\eta) + \hat{\si}'(t,\eta,0)\, \xi + R_{\ov{\ve}:{\hat{T}}}(t,\eta,\xi)\, \xi^2, \qquad \xi < 0. \label{E5.epsT-}\end{aligned}$$ The functions $u^$ and $a^$ introduced in , and the normal vector $n$ satisfy on the interface $\Gm$ $$\begin{aligned} \Big( D \big( \ve(u^ \otimes n) - \ov\ve) \big) \Big) n &=& 0, \label{E5.Dn}\\ \big( D \ve (a^* \otimes n + \na_\Gm u^* ) \big) n + \div_\Gm D \ve(u^* \otimes n) &=& 0. \label{E5.an}\end{aligned}$$ [Proof:]{} With the splitting of the gradient operator we compute from that $$\begin{aligned} \na_x \hat{u}(t,x) &= \pa_\xi \hat{u} \otimes n + \na_{\Gamma_\xi} \hat{u} \\ &= \big(u^(t,\eta) 1^+(\xi)+ a^(t,\eta) \xi^+\big) \otimes n (t,\eta) + \na_{\Gamma_\xi} u^(t,\eta) \xi^+\\ &\phantom{=}+\na_{\Gamma_\xi} a^(t,\eta) \frac{1}{2} (\xi^+)^2+ \na_x \hat{v}(t,x). \end{aligned}$$ We insert this equation into , note that $\hat{S}(t,x)=1^+(\xi)$, and employ that by $$D \big(\ve(u^* \otimes n)- \ov\ve\big) \hat{S}=[\hat{T}] \hat{S}$$ to obtain . By we have $$\label{E5.24} \na_{\Gamma_\xi} u^(t,\eta) \xi^+= \na_\Gamma u^ (t,\eta) \xi^+ + (\na_\Gamma u^(t,\eta)) R_A(t,\eta,\xi)(\xi^+)^2,$$ where we used our convention to identify $\na_\Gamma u^$ and $\na_\eta u^$, since $u^$ does not depend on $\xi$. Noting the definition of $\hat{\sigma}$ in , we obtain from and from that $$\begin{aligned} \ov\ve : \hat{T}&(t,\eta,\xi) = \ov\ve : [\hat{T}](t,\eta) 1^+(\xi) + \hat{\sigma} (t,\eta,\xi) \nn \\ & + \ov\ve: D \ve \big(a^(t,\eta) \otimes n(t,\eta)+ \na_\Gamma u^(t,\eta)\big) \xi^+ \nn \\ & + \ov\ve: D \ve \big(\na_\Gamma u^(t,\eta) R_A (t,\eta,\xi) + \frac{1}{2} \na_{\Gamma_\xi} a^(t,\eta)\big) (\xi^+)^2. \label{E5.25}\end{aligned}$$ and together imply for $\xi<0$ that $$\label{E5.26} \hat{T} (t,\eta,\xi)=D \ve \big(\na_x \hat{v} (t,\eta,\xi)\big),$$ whence $$\ov\ve: \hat{T} (t,\eta,\xi)= \ov\ve: D \ve \big(\na_x \hat{v}(t,\eta,\xi)\big)= \hat{\sigma} (t,\eta,\xi), \quad \xi <0,$$ and therefore $$\label{E5.27} \ov\ve: \hat{T}^{(-)} (t,\eta)= \hat{\sigma} (t,\eta,0), \quad \ov\ve: \hat{T}^{(+)} (t,\eta)= \ov\ve: [\hat{T}] (t,\eta)+ \hat{\sigma}(t,\eta,0).$$ By Taylor’s formula we can express $\hat{\sigma}$ in the form $$\hat{\sigma}(t,\eta,\xi)= \hat{\sigma} (t,\eta,0)+ \pa_\xi \hat{\sigma}(t,\eta,0) \xi+ \pa^2_\xi \hat{\sigma}(t,\eta,\xi^) \xi^2.$$ We expand $\hat{\sigma}$ in with this formula and note the equations to obtain and . is an immediate consequence of and . To prove we apply to calculate from that $$\begin{aligned} 0 &=& \div_x \hat{T} + {\sf b} \nn \\[1ex] &=& \pa_\xi ( [{\hat{T}}] n {\hat{S}}) + \div_{\Gm_\xi} [{\hat{T}}] {\hat{S}}\nn \\ && \mbox{} + \pa_\xi\big( D \ve (a^ \otimes n + \na_{\Gm_\xi} u^* ) \xi^+ \big) n + \pa_\xi \big( \frac12 D \ve( \na_{\Gm_\xi} a^) (\xi^+)^2 \big)n \nn \\ &&\mbox{} + \div_{\Gm_\xi} D \ve \Big( \big(a^ \otimes n + \na_{\Gm_\xi} u^* \big) \xi^+ + \frac12 (\na_{\Gm_\xi} a^) (\xi^+)^2 \Big) \nn \\[1ex] &&\mbox{} + \div_x D\ve(\na_x {\hat{v}}) + {\sf b}. \label{EdivhatT}\end{aligned}$$ From this equation we obtain for $\xi < 0$ that $$\label{Eacdiv} \div_x D\ve(\na_x {\hat{v}}) + {\sf b} = \div_x {\hat{T}}+ {\sf b} = 0.$$ By assumption in , the function ${\sf b}$ is continuous at $\Gm$. Moreover, by and the differentiability properties of ${\hat{v}}$ required in the function ${\hat{v}}$ is two times continuously differentiable at $\Gm$. Therefore we infer from that $$\label{Ebcdiv} \big(\div_x D\ve(\na_x {\hat{v}}) + {\sf b} \big)^{(+)} = \big( \div_x D\ve(\na_x {\hat{v}}) + {\sf b} \big)^{(-)} = 0, \quad \mbox{on } \Gm.$$ With this equation and with $[{\hat{T}}]n = 0$, by , we conclude from that $$\begin{aligned} 0 = \lim_{\xi \ra 0+} (\div_x \hat{T} + {\sf b}) = \big( D \ve (a^ \otimes n + \na_\Gm u^* ) \big) n + \div_\Gm [{\hat{T}}] . \end{aligned}$$ From this relation and from $\div_\Gm [{\hat{T}}] = \div_\Gm D \ve(u^* \otimes n)$, which is a consequence of , we obtain .\ Next we study the stress field $T^{(\mu)}_1$ in the inner expansion. \[L5.2\] Let $u_1^{(\mu)}$, $S_1^{(\mu)}$, $T_1^{(\mu)}$ be given in – , let $u_0$, $u_1$, $u_2$ be defined in – , and let $R_A$ be the remainder term from . We set $\zeta= \frac{\xi}{(\mu\lambda)^{1/2}}$. Then we have for $(t, \eta, \xi)$ from the neighborhood $\cal{U}_\delta$ of $\Gamma$ $$\begin{gathered} \label{E5.30} T_1^{(\mu)}(t,\eta,\xi) = [\hat{T}] (S_0+ \mu^{1/2} S_1) +D \ve \big(\na_x (\hat{v}+ \mu^{1/2} \check{v})\big) \\ + (\mu\lambda)^{1/2} D \ve (a^* \otimes n + \na_\Gamma u^) \int_{-\infty}^\zeta S_0(\vartheta) \, d\vartheta + \mu R_{T_1} (\lambda,t,\eta,\xi, \zeta), \end{gathered}$$ where $$\label{E5.31} R_{T_1} (\lambda, t, \eta, \xi, \zeta) = D \Big(\ve \big(\na_{\Gamma_\xi} (\lambda^{1/2} u_1 + \lambda u_2) + \lambda\zeta (\na_\eta u_0) R_A\big)- \ov\ve S_2\Big).$$ The argument of $[\hat{T}]$, $u^$, $a^$, $n$ is $(t,\eta)$, the argument of $S_1$, $S_2$, $u_0$, $u_1$, $u_2$ is $(t,\eta, \frac{\xi}{(\mu\lambda)^{1/2}})$, the argument of $\na_x \hat{v}$, $\na_x \check{v}$, $R_A$ is $(t,\eta,\xi)$, and the argument of $S_0$ outside of the integral is $\frac{\xi}{(\mu\lambda)^{1/2}}$. Moreover, we have $$\begin{aligned} \div_x\, T_1^{(\mu)} + {\sf b} &= \div_x D\ve(\na_x \hat{v})+ {\sf b}+ \xi \,\div_{\Gamma,\xi} [{\hat{T}}] S_0 + \mu^{1/2} \div_{\Gamma_\xi} \big([{\hat{T}}] S_1\big) \nn \\ & + \mu^{1/2} \div_x D\ve(\na_x \check{v}) + (\mu\lambda)^{1/2} \div_{\Gamma_\xi} D\ve (a^ \otimes n+ \na_\Gm u^) \int_{-\infty}^\zeta S_0 (\vartheta)\, d\vartheta \nn \\ & + \mu \,\div_x R_{T_1}\,. \label{E5.32} \end{aligned}$$ With $\hat{\sigma}(0)$, $\hat{\sigma}'(0)$, $\check{\sigma}(0)$ defined in , we have $$\begin{gathered} \pa_S {\sf W}\big(\ve(\na_x u_1^{(\mu)}), S_1^{(\mu)} \big) \\ = -\ov\ve:T_1^{(\mu)} (t,\eta,\xi) = -\ov\ve: [\hat{T}] (S_0+ \mu^{1/2} S_1) - \hat{\sigma}(0) - (\mu\lambda)^{1/2} \hat{\sigma}'(0) \zeta - \mu^{1/2} \check{\sigma}(0) \\ - (\mu\lambda)^{1/2} \ov\ve :D \ve (a^ \otimes n + \na_\Gamma u^) \int_{-\infty}^\zeta S_0 (\vartheta) d \vartheta - \mu R_W (\lambda,t,\eta,\xi,\zeta), \label{E5.33}\end{gathered}$$ where $$\label{E5.34} R_W (\la,t,\eta,\xi,\zeta)= \ov\ve :R_{T_1} (\la,t,\eta,\xi,\zeta) + \la \pa_\xi^2 \hat{\si} (t,\eta, \hat{\xi}) \zeta^2+ \la^{1/2} \pa_\xi \check{\si} (t,\eta, \check{\xi}) \zeta,$$ with suitable $\hat{\xi}$, $\check{\xi}$ between $0$ and $\xi$ and with $R_{T_1}$ defined in . [Proof:]{} With the splitting of the gradient operator we obtain by definition of $u_1^{(\mu)}$ in , – that $$\begin{aligned} \label{E5.35} \na_x u_1^{(\mu)} &=& (\mu\la)^{1/2} \na_x u_0 + \mu\la^{1/2} \na_x u_1+ \mu\lambda \na_x u_2 + \na_x ({\hat{v}}+ \mu^{1/2} {\check{v}}) \nn \\ &=& (u^ \otimes n) S_0 +\mu^{1/2} u^* \otimes n S_1 +(\mu\lambda)^{1/2} a^* \otimes n \int_{-\infty}^\zeta S_0(\vartheta) d\vartheta \nn \\ && \mbox{} + (\mu\la)^{1/2} \na_{\Gm_\xi} u_0+ \mu\la^{1/2} \na_{\Gm_\xi} u_1 +\mu\la \na_{\Gm_\xi}u_2 + \na_x ({\hat{v}}+ \mu^{1/2} {\check{v}}). \end{aligned}$$ and together yield $$\label{E5.36} \na_{\Gamma_\xi} u_0 = \na_\eta u_0 (I+\xi R_A) = \na_\Gamma u^* \int_{-\infty}^\zeta S_0(\vartheta) d\vartheta +(\mu\lambda)^{1/2} \zeta\, (\na_\eta u_0) R_A.$$ We insert , and into . From the resulting equation we obtain and if we also note that by $$\begin{gathered} D\big(\ve(u^* \otimes n) (S_0+ \mu^{1/2} S_1) - \ov\ve (S_0+ \mu^{1/2} S_1)\big)\\ = D\big(\ve(u^* \otimes n)-\ov\ve\big) (S_0+ \mu^{1/2} S_1)= [\hat{T}] (S_0+ \mu^{1/2} S_1).\end{gathered}$$ To prove we employ the splitting of the divergence operator and to compute from $$\begin{aligned} \div_x\, T_1^{(\mu)}+ {\sf b} &= \pa_\xi ([\hat{T}] n S_0) + (\div_\Gm [\hat{T}]) S_0 + \xi \big(\div_{\Gm,\xi} [\hat{T}]\big) S_0 \nn \\[1ex] & + \mu^{1/2} \big( \pa_\xi ([\hat{T}] n S_1 ) + \div_{\Gamma_\xi} ([\hat{T}] S_1 )\big) \nn \\[1ex] & + \div_x D \ve (\na_x \hat{v}) +{\sf b} +\mu^{1/2} \div_x D\ve(\na_x \check{v}) + \big( D\ve (a^* \otimes n + \na_\Gamma u^)\big)n S_0 \nn \\ & + (\mu\lambda)^{1/2} \div_{\Gamma_\xi} D\ve (a^ \otimes n +\na_\Gamma u^) \int_{-\infty}^\zeta S_0(\vartheta) d \vartheta + \mu\, \div_x R_{T_1}. \label{E5.37} \end{aligned}$$ By and we have $$\big(D\ve (a^ \otimes n + \na_\Gamma u^)\big)n S_0 + \div_\Gamma [\hat{T}] S_0 = \Big( \big(D\ve(a^ \otimes n +\na_\Gamma u^)\big)n + \div_\Gm D\ve(u^ \otimes n)\Big) S_0=0.$$ With this equation and with we obtain from . , follow immediately from , which implies $\pa_S {\sf W}\big(\ve(\na_x u_1^{(\mu)}),S_1^{(\mu)} \big) = - \ov{\ve} : T_1^{(\mu)}$, and from , , using the Taylor expansions $$\begin{aligned} {\check{\sigma}}(\xi) &= {\check{\sigma}}(0)+ \pa_\xi {\check{\sigma}}(\check{\xi}) \xi = {\check{\sigma}}(0) + (\mu\la)^{1/2} \pa_\xi {\check{\sigma}}(\check{\xi}) \zeta, \\ \hat{\si} (\xi) &= \hat{\si}(0)+ \pa_\xi \hat{\si}(0)\xi+ \pa^2_\xi \hat{\si} (\hat{\xi}) \xi^2= \hat{\si}(0) + (\mu\la)^{1/2} \hat{\si}'(0)\zeta+ \mu\la \hat{\si}''(\hat{\xi}) \zeta^2. \end{aligned}$$ This completes the proof of . \[C5.3\] Let $Q_{\rm inn}^{(\mu\la)}$ and $Q_{\rm match}^{(\mu\la)}$ be defined in and let $T^{(\mu)}_1$ be given by . Then there is a constant $C$ such that for all $0<\mu \leq \mu_0$ and all $0 < \la \leq \la_0$ $$\label{E5.38} \\|\div_x\, T_1^{(\mu)}+{\sf b}\\|_{L^\infty (Q_{\rm inn}^{(\mu\la)} \cup Q_{\rm match}^{(\mu\la)})} \leq C \Big( \frac{\mu}{\la}\Big)^{1/2} \|\ln \mu\|^2.$$ [Proof:]{} We estimate the terms on the right hand side of . Note first that if $(t,\eta,\xi) \in Q_{\rm inn}^{(\mu\la)} \cup Q_{\rm match}^{(\mu\la)}$ and $\zeta= \frac{\xi}{(\mu\lambda)^{1/2}}$, then $$\label{E5.39} \|\xi\| \leq \frac{3}{a} (\mu\lambda)^{1/2} \|\ln\mu\|, \quad\|\zeta\| \leq \frac{3}{a} \|\ln \mu\|.$$ This follows from . With these inequalities the first two terms on the right hand side of can be estimated as follows: From the differentiability properties of ${\hat{v}}$ and ${\sf b}$, which in are assumed to hold, it follows by that $$\label{E5.regudivhu} \div_x D \ve(\na_x {\hat{v}}) + {\sf b} \in C({\cal U}_\da) \cap C^1\big( (-\da,0], C(\Gm)\big) \cap C^1\big( [0,\da), C(\Gm)\big).$$ Because of this differentiability property we can apply the mean value theorem to $\div_x D \ve(\na_x {\hat{v}}) + {\sf b}$, which together with and yields for all $(t,\eta,\xi) \in Q_{\rm inn}^{(\mu\la)} \cup Q_{\rm match}^{(\mu\la)}$ with $\xi \geq 0$ that $$\begin{gathered} \label{E5.41} \Big\| \big(\div_x D\ve(\na_x {\hat{v}}) + {\sf b}\big)(t,\eta,\xi) \Big\| = \Big\| \big(\div_x\, D\ve(\na_x \hat{v})+{\sf b}\big)^{(+)} + \pa_\xi \big(\div_x D\ve(\na_x {\hat{v}})+{\sf b} \big) (t,\eta,\xi^)\xi \Big\| \\ = \big\| \pa_\xi \big(\div_x D\ve(\na_x {\hat{v}})+{\sf b} \big) (t,\eta,\xi^) \big\| \xi \leq C_1 \xi \leq C_1 \frac{3}{a} (\mu\lambda)^{1/2} \|\ln \mu\|, \end{gathered}$$ with a suitable number $\xi^$ between $0$ and $\xi$. Since by the term $\div_x D\ve(\na_x {\hat{v}}) + {\sf b}$ vanishes for $\xi < 0$, the inequality holds for all $(t,\eta,\xi) \in Q_{\rm inn}^{(\mu\la)} \cup Q_{\rm match}^{(\mu\la)}$. To estimate the last term in note that , and yield $$\begin{aligned} 0 \leq \int_{-\infty}^\zeta S_0(\vartheta) d\vartheta &\leq \zeta^+ + C_2 \leq C_3 \frac{3}{a} \|\ln \mu\|, \label{E5.intS0}\\[1ex] 0 \leq \int_{-\infty}^\zeta \int_{-\infty}^\vartheta S_0(\vartheta_1) d\vartheta_1 d\vartheta &\leq \frac{1}{2}(\zeta^+)^2 + C_4 \leq C_5 \Big(\frac{3}{a} \|\ln\mu\| \Big)^2, \nn \\[1ex] \Big\|\int_0^\zeta S_1(t,\eta,\vartheta) d \vartheta\Big\| &\leq C_6\|\zeta\| \leq C_6 \frac{3}{a} \|\ln\mu\|. \nn\end{aligned}$$ Using these inequalities, the definitions of $u_0$, $u_1$, $u_2$ in – and the inequality we obtain from that $$\begin{aligned} \|R_{T_1}\| &\leq& \big(C_7 + \frac{C_8}{a^2} \big) \| \ln \mu\|^2, \label{E5.estRT1} \\ \|\mu\, \div_x R_{T_1}\| &\leq& \mu^{1/2} \la^{-1/2} \big(C_9 + \frac{C_{10}} {a^2} \big) \| \ln \mu\|^2. \label{E5.estdivRT1} \end{aligned}$$ is used later, is the desired estimate for the last term in . To estimate the other terms in we apply , , , and. Together with and we find for $(t,\eta,\xi)\in Q_{\rm inn}^{(\mu\la)} \cup Q_{\rm match}^{(\mu\la)}$ that $$\Big\|(\div_x T_1^{(\mu)} + {\sf b}) (t,\eta,\xi)\Big\| \leq \Big(\frac{C_{11}}{a^2}+ C_{12}\Big) \Big(\mu^{1/2} + (\mu \la)^{1/2} \|\ln\mu\| + \mu^{1/2}\la^{-1/2}) \|\ln\mu\|^2\Big),$$ which implies .\ In the next lemma we study the outer expansion $T_2^{(\mu)}$. \[L5.4\] Let $(\tilde{u}, \tilde{T})$ be the solution of the boundary value problem, which consists of the elliptic system , with $\tilde{S}_2$ given by , and of the boundary conditions – . Let $\tilde{S}_3$ be the solution of . Then $T_2^{(\mu)}$ defined in satisfies on $Q \setminus \Gamma$ $$\begin{aligned} T_2^{(\mu)} &=& \hat{T}+ \mu^{1/2} \check{T}+ \mu \tilde{T}- \mu^{3/2} D \ov\ve \tilde{S}_3, \label{E5.42} \\ \div_x T^{(\mu)}_2 +{\sf b} &=& -\mu^{3/2} \div_x(D \ov\ve \tilde{S}_3).\label{E5.43} \end{aligned}$$ [Proof:]{} Insertion of and into yields $$\begin{aligned} T_2^{(\mu)} = D \big(\ve (\na_x \hat{u}) - \ov\ve \hat{S}\big) &+ \mu^{1/2} D \big(\ve(\na_x \check{u})- \ov\ve \tilde{S}_1\big) \\ &+ \mu D \big(\ve(\na_x \tilde{u})- \ov\ve \tilde{S}_2\big)- \mu^{3/2} D \ov\ve \tilde{S}_3. \end{aligned}$$ Using , we see from this equation and from , , that holds. is an immediate consequence of and , , . Asymptotic expansion of $\boldsymbol{S_t + c({\sf W}_S + \hat{\psi} - \Da_x S)}$ {#S5.2} --------------------------------------------------------------------------------- In this section we compute the form of the residue $$\label{E5.44} (\mu\lambda)^{1/2} \pa_t S+ c\Big(\pa_S {\sf W}\big(\ve(\na_x u),S\big)+ \frac{1}{\mu^{1/2}} \hat{\psi}' (S)- \mu^{1/2}\lambda \Delta_x S\Big),$$ which is obtained when we either insert for $(u,S)$ the inner expansion $\big(u_1^{(\mu)},S_1^{(\mu)} \big)$ or the outer expansion $\big(u_2^{(\mu)},S_2^{(\mu)} \big)$ of the asymptotic solution $\big(u^{\mu)},S^{(\mu)}\big)$. For functions $(t,x) \mapsto w(t,x)$ defined in a neighborhood of $\Gm$ we write $w(t,\eta,\xi) = w(t,x)$ with $x = \eta + n(t,\eta)\xi$, as always. However, in the following computations this slight abuse of notation could lead to confusion when we consider derivatives with respect to $t$. To avoid this, we introduce the notations $$w_{\|t}(t,x) = w_{\|t}(t,\eta,\xi) = \pa_r w(r,\eta,\xi)\rain{r=t}\,, \quad (\pa_t w)(t,\eta,\xi) = \pa_t w(t,x).$$ As introduced previously, for $i=0,1,2$ we write $S_i'(t,\eta,\zeta) = \pa_\zeta S_i(t,\eta,\zeta)$ and $S_i''(t,\eta,\zeta) = \pa^2_\zeta S_i(t,\eta,\zeta)$. #### Inner expansion We first compute for $(u,S) = \big(u_1^{(\mu)},S_1^{(\mu)} \big)$. To this end we need \[L5.5\] Let $s^{(\mu)}(t,\eta)$ be the normal speed of the phase interface $\Gamma(t)$ at $\eta \in \Gamma(t)$, let $\na_\eta$ be the operator defined in , and let $w$ be a function defined in a neighborhood of $\Gm$. Then we have $$\pa_t w (t,x)= w_{\|t} (t,\eta,\xi) - \xi (\pa_t n)(t,\eta) \cdot \na_\eta w (t,\eta,\xi) - s^{(\mu)} (t,\eta) \pa_\xi w (t,\eta,\xi).$$ [Proof:]{} By definition, $\na_\eta w(t,\eta,\xi)$ is a tangential vector to $\Gm(t)$. thus yields $$\begin{gathered} \pa_t w (t,x)= \pa_t w(t,\eta,\xi) = w_{\|t} (t,\eta,\xi) + \pa_t \eta \cdot \na_\eta w (t,\eta,\xi) + \pa_\xi w (t,\eta,\xi) \pa_t \xi \\ = w_{\|t} (t,\eta,\xi) - \xi (\pa_t n)(t,\eta) \cdot \na_\eta w (t,\eta,\xi) - s^{(\mu)} (t,\eta) \pa_\xi w (t,\eta,\xi). \end{gathered}$$ This proves the lemma.\ We apply this lemma to the function $S_1^{(\mu)}$ defined in to obtain $$\pa_t S_1^{(\mu)} (t,x)= S_{1\|t}^{(\mu)} (t,\eta,\xi)- \xi (\pa_t n)(t,\eta) \cdot \na_\eta S_1^{(\mu)} (t,\eta,\xi)- s^{(\mu)} (t,\eta) \pa_\xi S_1^{(\mu)} (t,\eta,\xi).$$ From this equation and from the asymptotic expansion of $s^{(\mu)}$ we conclude for the first term in that $$\begin{aligned} (\mu\lambda)^{1/2}& \pa_t S_1^{(\mu)} (t,x) \nn \\ & = (\mu\lambda)^{1/2} \pa_t \Big(S_0 \big( \frac{\xi}{(\mu\lambda)^{1/2}}\big) + \mu^{1/2} S_1 \big(t,\eta,\frac{\xi}{(\mu\lambda)^{1/2}}\big) + \mu S_2 \big(t,\eta,\frac{\xi}{(\mu\lambda)^{1/2}}\big) \Big) \nn \\ &= -(s_0+\mu^{1/2} s_1) (S'_0 + \mu^{1/2} S'_1 + \mu S'_2) +\mu\la^{1/2} \tilde{R}_{S_t} \nn \\ &= -s_0 S_0' - \mu^{1/2} (s_1 S_0' + s_0 S_1') + \mu R_{S_t}, \label{E5.47} \end{aligned}$$ with $$\begin{aligned} \tilde{R}_{S_t}(\mu,\lambda,t,\eta,\xi) &=& (S_{1\|t}+ \mu^{1/2} S_{2\|t})- \xi (\pa_t n) \cdot \na_\eta (S_1+ \mu^{1/2} S_2), \nn \\ R_{S_t}(\mu,\lambda,t,\eta,\xi) &=& \la^{1/2} \tilde{R}_{S_t} - s_1 S_1' - (s_0 + \mu^{1/2} s_1)S_2'. \label{E5.48} \end{aligned}$$ For the third term in we get from Taylor’s formula and from $$\label{E5.49} \frac{1}{\mu^{1/2}} \hat{\psi}'(S^{(\mu)}_1) = \frac{1}{\mu^{1/2}} \hat{\psi}' (S_0) + \hat{\psi}'' (S_0) S_1 + \mu^{1/2} \big( \hat{\psi}''(S_0) S_2 + \frac12 \hat{\psi}'''(S_0) S_1^2 \big) + \mu R_{\hat{\psi}'}$$ where $$\label{E5.50} R_{\hat{\psi}'} = \frac12 \hat{\psi}'''(S_0)\big(2 S_1 S_2 + \mu^{1/2}S_2^2 \big) + \frac16 \hat{\psi}^{(IV)}\big(S_0 + \vartheta (\mu^{1/2} S_1 + \mu S_2)\big) (S_1 + \mu^{1/2} S_2)^3.$$ with suitable $0 < \vartheta < 1$. Observe next that $$\label{E5.51} \Delta_x S_1^{(\mu)} (t,x)= \pa_\xi^2 S_1^{(\mu)} (t,\eta,\xi)- \ka(t,\eta,\xi) \pa_\xi S^{(\mu)}_1(t,\eta,\xi) + \Delta_{\Gamma_\xi} S^{(\mu)}_1 (t,\eta,\xi),$$ where $\Delta_{\Gamma_\xi}= \div_{\Gm_\xi} \na_{\Gm_\xi}$ denotes the surface Laplacian and where $\ka(t,\eta,\xi)$ is twice the mean curvature of the surface $\Gm_\xi(t)$ at $\eta \in \Gm_\xi(t)$. With the notation $\ka'(t,\eta,0) = \pa_\xi \ka(t,\eta,0)$ we obtain from Taylor’s formula $$\begin{aligned} \kappa(t,\eta,\xi) &=& \kappa(t,\eta,0)+ \pa_\xi \kappa(t,\eta,0)\xi+ \frac{1}{2} \pa^2_\xi \kappa(t,\eta,\xi^)\xi^2 \nn \\ &=& \ka_\Gm(t,\eta)+ (\mu\la)^{1/2} \ka'(t,\eta,0) \zeta + \mu\la R_\ka(t,\eta,\xi) \zeta^2, \label{E5.52} \end{aligned}$$ where $R_\ka(t,\eta,\xi) = \frac{1}{2} \pa_\xi^2 \kappa(t,\eta,\xi^)$ is the remainder term, with suitable $\xi^$ between $0$ and $\xi$. We insert and into and obtain for the fourth term in that $$\begin{gathered} \label{E5.53} \mu^{1/2}\la \Da_x S^{(\mu)}_1 = \mu^{-1/2} S_0'' + S_1'' + \mu^{1/2} S_2'' \\ - \la^{1/2} \Big( \ka_\Gm + (\mu\la)^{1/2} \ka'\zeta + \mu \la R_\ka \zeta^2 \Big) (S_0' + \mu^{1/2} S_1') - \mu \la^{1/2} \ka(\xi) S_2' + \mu^{1/2}\la \Da_{\Gm_\xi} S^{(\mu)}_1 \\ = \mu^{-1/2} S_0'' + \big(S_1'' - \la^{1/2} \ka_\Gm S_0'\big) + \mu^{1/2} \Big( S_2'' - \la^{1/2} \ka_\Gm S_1' - \la \ka' \zeta S_0' \Big) + \mu R_\Da, \end{gathered}$$ where $$\label{E5.54} R_\Da = - \la \ka' \zeta S_1' - \la^{1/2} \ka(\xi) S_2' + \la \Da_{\Gm_\xi} (S_1 + \mu^{1/2} S_2) + \la^{3/2} R_\ka\zeta^2 (S_0' + \mu^{1/2} S_1').$$ From , , and we obtain $$\begin{aligned} (\mu\la)^{1/2} \pa_t S_1^{(\mu)} &+ c\Big(\pa_S {\sf W}\big(\ve(\na_x u_1^{(\mu)}),S_1^{(\mu)}\big)+ \frac{1}{\mu^{1/2}} \hat{\psi}' (S_1^{(\mu)})- \mu^{1/2}\la \Da_x S_1^{(\mu)}\Big) \nn \\ = \frac{c}{\mu^{1/2}}& \big(\hat{\psi}' (S_0) - S_0'' \big) + c\Big( \hat{\psi}'' (S_0) S_1 - S_1'' - \ov\ve: [\hat{T}] S_0 - \hat{\si}(0) + (\la^{1/2} \ka_\Gm - \frac{s_0}{c}) S_0' \Big) \nn \\ \mbox{} + c\mu^{1/2}& \Bigg(\hat{\psi}''(S_0) S_2 - S_2'' -\ov\ve : [\hat{T}] S_1 - \check{\si}(0) + \big(\la^{1/2} \ka_\Gm - \frac{s_0}{c} \big) S_1' \nn \\ & - \la^{1/2}\Big(\hat{\si}'(0) \zeta + \ov{\ve} : D \ve (a^* \otimes n + \na_\Gamma u^) \int_{-\infty}^\zeta S_0 (\vta) d \vta \Big) \nn \\ & + \frac12 \hat{\psi}'''(S_0) S_1^2 + \big( \la \ka' \zeta - \frac{s_1}{c} \big) S_0' \Bigg) + \mu R_{S_t + c(\ldots)}\,, \label{E5.55}\end{aligned}$$ where $$\label{E5.56} R_{S_t + c(\ldots)} = R_{S_t} + c(- R_W + R_{\hat{\psi}'} - R_\Da ),$$ with $R_{S_t}$, $R_W$, $R_{\hat{\psi}'}$ and $R_\Da$ given in , , and , respectively. \[C5.6\] Let $s_0$ be given by and assume that the functions $S_0$, $S_1$ and $S_2$ satisfy the ordinary differential equations , , , with $F_1$, $F_2$ given by , . Assume moreover that the conditions – and , hold. Then there is a constant $K$ such that the interior expansion $\big( u_1^{(\mu)},S_1^{(\mu)},T_1^{(\mu)} \big)$ defined in – satisfies for all $(t,\eta,\xi) \in Q_{\rm inn}^{(\mu\la)} \cup Q_{\rm match}^{(\mu\la)}$ and all $0 < \mu \leq \mu_0$, $0 < \la \leq \la_0$ the inequality $$\begin{gathered} \Big\| \pa_t S_1^{(\mu)} + \frac{c}{(\mu\la)^{1/2}} \Big(\pa_S {\sf W} \big(\ve(\na_x u_1^{(\mu)}),S_1^{(\mu)}\big)+ \frac{1}{\mu^{1/2}} \hat{\psi}' (S_1^{(\mu)})- \mu^{1/2}\la \Da_x S_1^{(\mu)}\Big) \Big\| \\ = \Big( \frac{\mu}{\la} \Big)^{1/2} \| R_{S_t} + c(- R_W + R_{\hat{\psi}'} - R_\Da ) \| \leq K \Big( \frac{\mu}{\la} \Big)^{1/2} \|\ln \mu\|^2. \label{E5.57} \end{gathered}$$ [Proof:]{} From we obtain $$\la^{1/2} \ka_\Gm - \frac{s_0}{c} = \frac{1}{c_1} \ov{\ve}:\langle {\hat{T}}\rangle,$$ and by we have $\ov{\ve}: {\hat{T}}^{(-)} = \hat{\si}(0)$. After insertion of these two equations into , the latter equation takes the form $$\begin{aligned} (\mu\la)^{1/2} \pa_t S_1^{(\mu)}& + c\Big(\pa_S {\sf W} \big(\ve(\na_x u_1^{(\mu)}),S_1^{(\mu)}\big) + \frac{1}{\mu^{1/2}} \hat{\psi}' (S_1^{(\mu)})- \mu^{1/2}\la \Da_x S_1^{(\mu)}\Big) \nn \\ = \frac{c}{\mu^{1/2}}& \Big(\hat{\psi}' (S_0) - S_0'' \Big) + c\Big( \hat{\psi}'' (S_0) S_1 - S_1'' - F_1 \Big) \mbox{} + c\mu^{1/2} \Big(\hat{\psi}''(S_0) S_2 - S_2'' - F_2 \Big) \nn \\ &+ \mu R_{S_t + c(\ldots)} = \mu \big(R_{S_t} + c(- R_W + R_{\hat{\psi}'} - R_\Da )\big). \label{E5.58}\end{aligned}$$ Here we also used , and . Noting the inequalities for $\xi$ and $\zeta$, the inequalities – , , , for $S_0$, $S_1$, $S_2$, and the inequality for the term $R_{T_1}$, which appears in $R_W$, we see by inspection of every term in , , and that the inequality $$\label{E5.59} \big\| R_{S_t} + c(- R_W + R_{\hat{\psi}'} - R_\Da )\big\| \leq K \| \ln \mu \|^2$$ holds. To obtain inequality we divide by $(\mu\la)^{1/2}$ and estimate the right hand side of the resulting equation using . #### Outer expansion Next we compute for $(u,S) = \big(u_2^{(\mu)},S_2^{(\mu)} \big)$. Note first that and yield $$\label{E5.60} \pa_S {\sf W}\big( \ve(\na_x u_2^{(\mu)}),S_2^{(\mu)} \big) = - \ov{\ve} : ({\hat{T}}+ \mu^{1/2} {\check{T}}+ \mu {\tilde{T}}) + \mu^{3/2} \ov{\ve} :D \ov{\ve} {\tilde{S}}_3\,.$$ Also, Taylor’s formula and yield for a suitable $0 < \vartheta(t,x) <1$ $$\begin{aligned} \hat{\psi}'(S_2^{(\mu)}) = \hat{\psi}'({\hat{S}})& + \hat{\psi}''({\hat{S}})\big(\mu^{1/2} {\tilde{S}}_1 + \mu {\tilde{S}}_2 + \mu^{3/2} {\tilde{S}}_3 \big) \nn \\ \mbox{} + \frac12& \hat{\psi}'''({\hat{S}}) \big(\mu^{1/2} {\tilde{S}}_1 + \mu {\tilde{S}}_2 + \mu^{3/2} {\tilde{S}}_3 \big)^2 + \frac16 \hat{\psi}^{(IV)}({\hat{S}}) \big(\mu^{1/2} {\tilde{S}}_1 + \mu {\tilde{S}}_2 + \mu^{3/2} {\tilde{S}}_3 \big)^3 \nn \\ &+ \frac1{24}\hat{\psi}^{(V)}\big({\hat{S}}+ \vartheta( \mu^{1/2} {\tilde{S}}_1 + \mu {\tilde{S}}_2 + \mu^{3/2} {\tilde{S}}_3) \big) \big(\mu^{1/2} {\tilde{S}}_1 + \mu {\tilde{S}}_2 + \mu^{3/2} {\tilde{S}}_3 \big)^4 \nn \\ = \mu^{1/2} \hat{\psi}''&({\hat{S}}) {\tilde{S}}_1 + \mu \Big(\hat{\psi}''({\hat{S}}) {\tilde{S}}_2 + \frac12 \hat{\psi}'''({\hat{S}}) {\tilde{S}}_1^2 \Big) \nn \\ &+ \mu^{3/2} \Big( \hat{\psi}''({\hat{S}}) {\tilde{S}}_3 + \hat{\psi}'''({\hat{S}}) {\tilde{S}}_1{\tilde{S}}_2 + \frac16 \hat{\psi}^{(IV)}({\hat{S}}) {\tilde{S}}_1^3 \Big) + \mu^2\, \ov{R}_{\hat{\psi}'}\,. \label{E5.61}\end{aligned}$$ Here we used that $\hat{\psi}'({\hat{S}}) = 0$, by . Equations and imply that in the domain $Q \setminus \Gm$ $$\begin{aligned} (\mu\la)^{1/2}& \pa_t S_2^{(\mu)} + c\Big(\pa_S {\sf W} \big(\ve(\na_x u_2^{(\mu)}),S_2^{(\mu)}\big) + \frac{1}{\mu^{1/2}} \hat{\psi}' (S_2^{(\mu)})- \mu^{1/2}\la \Da_x S_2^{(\mu)}\Big) \nn \\ &= c \Big( - \ov{\ve} : {\hat{T}}+ \hat{\psi}''({\hat{S}}) {\tilde{S}}_1 \Big) \nn \\ &+ c \mu^{1/2}\Big(- \ov{\ve} : {\check{T}}+ \hat{\psi}''({\hat{S}}) {\tilde{S}}_2 + \frac12 \hat{\psi}'''({\hat{S}}) {\tilde{S}}_1^2 \Big) \nn \\ &+ c\mu \Big(- \ov{\ve} : {\tilde{T}}+ \hat{\psi}''({\hat{S}}) {\tilde{S}}_3 + \hat{\psi}'''({\hat{S}}) {\tilde{S}}_1 {\tilde{S}}_2 + \frac16 \hat{\psi}^{(IV)}({\hat{S}}) {\tilde{S}}_1^3 - \la \Da_x {\tilde{S}}_1 + \frac{\la^{1/2}}{c} \pa_t {\tilde{S}}_1 \Big) \nn \\ & + \mu^{3/2}\, \ov{R}_{S_t + c(\ldots)}\,, \label{E5.62}\end{aligned}$$ where $$\label{E5.63} \ov{R}_{S_t + c(\ldots)} = c\,\big( \ov{\ve} :D \ov{\ve} {\tilde{S}}_3 + \ov{R}_{\hat{\psi}'} - \la \Da_x ( {\tilde{S}}_2 + \mu^{1/2} {\tilde{S}}_3 )\big) + \la^{1/2} \pa_t({\tilde{S}}_2 + \mu^{1/2} {\tilde{S}}_3).$$ Here we used that the function ${\hat{S}}$ has the constant values $0$ in $\gm$ and $1$ in $\gm'$. \[C5.7\] Assume that the functions ${\tilde{S}}_1$, ${\tilde{S}}_2$ and ${\tilde{S}}_3$ satisfy – . Then there is a constant $K$ such that for all $(t,x) \in Q\setminus \Gm$ and all $0 < \mu \leq \mu_0$, $0 < \la \leq \la_0$ $$\label{E5.64} \Big\| \pa_t S_2^{(\mu)} + \frac{c}{(\mu\la)^{1/2}} \Big(\pa_S {\sf W} \big(\ve(\na_x u_2^{(\mu)}),S_2^{(\mu)}\big) + \frac{1}{\mu^{1/2}} \hat{\psi}' (S_2^{(\mu)})- \mu^{1/2}\la \Da_x S_2^{(\mu)}\Big) \Big\| \leq K \frac{\mu}{\la^{1/2}}\,.$$ [Proof:]{} By – , the brackets on the right hand side of equation vanish. Therefore, if we divide the latter equation by $(\mu\la)^{1/2}$, we obtain $$\pa_t S_2^{(\mu)} + \frac{c}{(\mu\la)^{1/2}} \Big(\pa_S {\sf W} \big(\ve(\na_x u_2^{(\mu)}),S_2^{(\mu)}\big) + \frac{1}{\mu^{1/2}} \hat{\psi}' (S_2^{(\mu)})- \mu^{1/2}\la \Da_x S_2^{(\mu)}\Big) = \frac{\mu}{\la^{1/2}} \ov{R}_{S_t + c(\ldots)}.$$ ${\tilde{S}}_1$ and ${\tilde{S}}_2$ are given in , , and the function ${\tilde{S}}_3$ is obtained by solving for this function. From these equations we see by our general regularity assumptions that $\\| ({\tilde{S}}_1,{\tilde{S}}_2,{\tilde{S}}_3)\\|_{L^\infty(Q\setminus \Gm)} \leq K_1$, with the constant $K_1$ independent of $\mu$. Using this, we see by inspection of every term in that $\\| \ov{R}_{S_t + c(\ldots)}\\|_{L^\infty(Q\setminus \Gm)} \leq K$, with $K$ independent of $\mu$ and $\la$. This inequality and the equation above imply . Auxiliary estimates needed in the matching region {#S5.3} ------------------------------------------------- The following auxiliary estimates are needed to prove and in the matching region $Q_{\rm match}^{(\mu\la)}$. \[L5.6a\] The functions $u_2^{(\mu)}$, $T_2^{(\mu)}$, $S_2^{(\mu)}$ defined in – and $u_1^{(\mu)}$, $T_1^{(\mu)}$, $S_1^{(\mu)}$ defined in – satisfy $$\begin{aligned} \label{ESmu1Smu2} \\|S_1^{(\mu)} - S_2^{(\mu)}\\|_{L^\infty(Q_{\rm match}^{(\mu\la)} ) } &\leq& K \mu^{3/2} \|\ln \mu\|^2, \\ \label{EDxSmu1Smu2} \\| D^\alpha_{x}(S_1^{(\mu)} - S_2^{(\mu)})\\|_{L^\infty( Q_{\rm match}^{(\mu\la)}) } &\leq& K \la^{-\frac{\|\al\|}{2}} \mu^\frac{3-\|\al\|}{2},\quad 1 \leq \|\al\| \leq 2, \\[1ex] \label{EDtSmu1Smu2} \\| \pa_t(S_1^{(\mu)} - S_2^{(\mu)})\\|_{L^\infty(Q_{\rm match}^{(\mu\la)} )} &\leq& K \la^{-1/2} \mu, \\ \label{Eumu1umu2} \\|u_1^{(\mu)} - u_2^{(\mu)}\\|_{L^\infty(Q_{\rm match}^{(\mu\la)} )} &\leq& K \la^{1/2} \mu^{3/2} \|\ln \mu\| , \\ \label{Enaumu1umu2} \\| \na_x (u_1^{(\mu)} - u_2^{(\mu)}) \\|_{L^\infty( Q_{\rm match}^{(\mu\la)} )} &\leq& K \mu, \\ \label{ETmu1Tmu2} \\|T_1^{(\mu)} - T_2^{(\mu)}\\|_{L^\infty(Q_{\rm match}^{(\mu\la)} )} &\leq& K \mu, $$ for all $\mu \in (0,\mu_0]$, $\la \in (0,\la_0]$. Here $\al$ denotes a multi-index and $K$ denotes a positive constant, which does not necessarily have the same value in the six estimates. [Proof:]{} Since the proofs of these estimates are long and technical, we present here only the proofs of and . The proofs of the estimates , and run along the same lines. is an immediate consequence of the definitions , of $T_2^{(\mu)}$ and $T_1^{(\mu)}$, and of the estimates , . In this proof we mostly drop the arguments $t$ and $\eta$ to simplify the notation. As usual we write $\zeta = \frac{\xi}{(\mu\la)^{1/2}}$. We need that for $(t,\eta,\xi)\in Q_{\rm match}^{(\mu\la)}$ the inequalities $$\label{E5.8} \frac{3}{2} \frac{\| \ln \mu\|}{a} \leq \Big\|\frac{\xi}{(\mu\lambda)^{1/2}}\Big\| = \|\zeta\| \leq 3 \frac{\|\ln \mu\|}{a},$$ hold, by definition of $Q_{\rm match}^{(\mu\la)}$ in . We begin with the proof of . By definition of $S_1^{(\mu)}$ and $S_2^{(\mu)}$ in and we have $$\begin{aligned} \label{E5.9} \| S_1^{(\mu)}-S_2^{(\mu)}\| &=& \|S_0+ \mu^{1/2}S_1+\mu S_2- \hat{S} - \mu^{1/2} \tilde{S}_1- \mu \tilde{S}_2- \mu^{3/2} \tilde{S}_3\| \nn \\ &\leq& \|S_0- \hat{S}\|+ \mu^{1/2} \|S_1+ \mu^{1/2} S_2- \tilde{S}_1- \mu^{1/2} \tilde{S}_2\|+ \mu^{3/2} \|\tilde{S}_3\|. \end{aligned}$$ To estimate the first term on the right hand side note that since $\hat{S}(t,x)= \hat{S}(\xi)=1^+(\xi)$, relations , , and imply for $(t,\eta,\xi)\in Q_{\rm match}^{(\mu\la)}$ that $$\label{E5.10} \Big\|S_0 \Big(\frac{\xi}{(\mu\lambda)^{1/2}}\Big)- \hat{S}(\xi)\Big\| \leq K_1 e^{-a\|\zeta\|} \leq K_1 e^{- \frac{3}{2}\|\ln \mu\|} = K_1 \mu^{3/2}.$$ To estimate the second term on the right hand side of we introduce the notations $$\begin{aligned} \tilde{\rho}_1(\zeta) &= \begin{cases} \frac{\ov{\ve}:\hat{T}^{(-)}}{\hat{\psi}''(0)}, & \zeta < 0, \\ \frac{\ov{\ve}:\hat{T}^{(+)}}{\hat{\psi}''(1)}, & \zeta > 0, \end{cases} \label{E5.trho1} \\ \tilde{\rho}_2(\zeta) &= \begin{cases} \frac{1}{\hat{\psi}''(0)} \Big( \ov{\ve}:{\check{T}}^{(-)} - \frac{\hat{\psi}'''(0)}{2} \Big( \frac{\ov{\ve} : {\hat{T}}^{(-)}}{\hat{\psi}''(0)} \Big)^2 + \la^\frac12\, \hat{\si}'(0) \zeta \Big) , & \zeta < 0, \\ \frac{1}{\hat{\psi}''(1)} \Big( \ov{\ve}:{\check{T}}^{(+)} - \frac{\hat{\psi}'''(1)}{2} \Big( \frac{\ov{\ve} : {\hat{T}}^{(+)}}{\hat{\psi}''(1)} \Big)^2 + \la^\frac12 \hat{\si}'(0) \zeta \\ \hphantom{\frac{1}{\hat{\psi}''(0)} \Big( \ov{\ve}:{\check{T}}^{(+)} -} + \la^\frac12 \ov{\ve} : D\ve (a^ \otimes n + \na_\Gm u^) \zeta \Big), & \zeta > 0. \end{cases} \label{E5.trho2}\end{aligned}$$ By definition of the functions $\rho_1$, $\rho_2$ in and , and by definition of $\varphi$, $\varphi_\pm$ in and , we have $$\tilde{\rho}_i(\zeta) - \rho_i(\zeta) = (1 - \varphi(-\zeta) - \varphi(\zeta)) \tilde{\rho}_i (\zeta) = 0, \qquad \mbox{for } i=1,2 \mbox{ and } \|\zeta\| \geq 2.$$ We can therefore choose a suitable constant $\tilde{K}$ such that $\big\| \tilde{\rho}_i(\zeta) - \rho_i(\zeta) \big\| \leq \tilde{K} e^{-a \|\zeta\|}$ holds for $i=1,2$ and all $\zeta \in \R$. From this inequality and from the estimates , we conclude that $$\begin{aligned} \| S_1 (\zeta) - \tilde{\rho}_1\| &\leq \| S_1 (\zeta) - \rho_1(\zeta)\| + \|\rho_1(\zeta) - \tilde{\rho}_1\| \leq (K_2 + \tilde{K}) e^{-a \|\zeta\|}, \label{E5.S1rho1} \\ \| S_2 (\zeta) - \tilde{\rho}_2\| &\leq \| S_2 (\zeta) - \rho_2(\zeta)\| + \|\rho_2(\zeta) - \tilde{\rho}_2\| \leq \big( K_5 (1+\|\zeta\|) + \tilde{K} \big) e^{-a \|\zeta\|}, \label{E5.S2rho2} \end{aligned}$$ for $\zeta \in \R$. Now we proceed to estimate the second term on the right hand side of . We insert the functions $\tilde{\rho}_1$ and $\tilde{\rho}_2$ into this term, use the expressions for ${\tilde{S}}_1$ and ${\tilde{S}}_2$ given in , , and employ the triangle inequality to obtain $$\begin{gathered} \label{E5.11} \| S_1 + \mu^{1/2} S_2- \tilde{S}_1- \mu^{1/2} \tilde{S}_2\| \\ \leq \left\|\tilde{\rho}_1+ \mu^{1/2} \tilde{\rho}_2- \frac{\ov\ve: \hat{T}}{\hat{\psi}''(\hat{S})}- \mu^{1/2} \left(\frac{\ov\ve:\check{T}}{\hat{\psi}''(\hat{S})}- \frac{\hat{\psi}'''(\hat{S})}{2 \hat{\psi}''(\hat{S})} \Big(\frac{\ov\ve:\hat{T}}{\hat{\psi}''(\hat{S})}\Big)^2 \right)\right\| \\ + \|S_1-\tilde{\rho}_1\|+ \mu^{1/2} \|S_2-\tilde{\rho}_2\|= \|I_1\|+ \|I_2\|+ \|I_3\|. \end{gathered}$$ By , and we have for $(t,\eta,\xi)\in Q_{\rm match}^{(\mu\la)}$ $$\begin{aligned} \|I_2\|+ \|I_3\| &=& \big\| S_1 (t,\eta, \zeta ) - \tilde{\rho}_1 (t,\eta, \zeta )\big\| + \mu^{1/2} \big\|S_2 (t,\eta, \zeta) - \tilde{\rho}_2 (t,\eta, \zeta)\big\| \nn \\[1ex] &\leq& (K_2 + \tilde{K}) e^{-a\|\zeta\|} + \mu^{1/2} \big(K_5 (1 + \| \zeta\|) + \tilde{K} \big) e^{-a\|\zeta\|} \nn \\ &\leq& \left(C_1 + \mu^{1/2} C_2 \Big(1+3 \frac{\|\ln \mu\|}{a}\Big)\right) e^{- \frac{3}{2}\|\ln \mu\|} \leq C_3\, \mu^{3/2}. \label{E5.12}\end{aligned}$$ To find an estimate for $\|I_1\|$ note that the definitions of $\tilde{\rho}_1$, $\tilde{\rho}_2$ in , yield for $(t,\eta,\xi)\in Q_{\rm match}^{(\mu\la)}$ with $\xi > 0$ $$\begin{aligned} I_1 &=\tilde{\rho}_1 \Big(\frac{\xi}{(\mu\lambda)^{1/2}}\Big) + \mu^\frac12 \tilde{\rho}_2 \Big(\frac{\xi}{(\mu\lambda)^{1/2}}\Big) - \frac{\ov\ve: \hat{T}(\xi)}{\hat{\psi}''(1)} - \mu^\frac12 \left( \frac{\ov\ve: {\check{T}}(\xi)}{\hat{\psi}''(1)} - \frac{\hat{\psi}'''(1)}{2 \hat{\psi}''(1)} \Big( \frac{\ov\ve: \hat{T}(\xi)}{\hat{\psi}''(1)} \Big)^2\right) \nn \\ &= \frac{1}{\hat{\psi}''(1)} \Bigg( \ov\ve: \hat{T}^{(+)} - \ov\ve: \hat{T}(\xi) + \sigma'(0) \xi + \ov\ve: D \ve (a^ \otimes n + \na_\Gamma u^)\xi \label{E5.13} \\ &\phantom{= \frac{1}{\hat{\psi}''(1)}} + \mu^\frac12 \big(\ov\ve: \check{T}^{(+)} - \ov\ve: \check{T}(\xi)\big) - \mu^\frac12\, \frac{\hat{\psi}'''(1)}{2} \left( \Big( \frac{\ov\ve: \hat{T}^{(+)}}{\hat{\psi}''(1)}\Big)^2- \Big( \frac{\ov\ve: \hat{T}(\xi)}{\hat{\psi}''(1)}\Big)^2\right)\Bigg). \nn \end{aligned}$$ and together yield $$\begin{gathered} \label{E5.14} \|\ov\ve: \hat{T}^{(+)}+ \sigma'(0)\xi + \ov\ve:D \ve (a^ \otimes n + \na_\Gamma u^)\xi - \ov\ve: \hat{T}(\xi)\| \\ = \|R_{\ov{\ve}:{\hat{T}}}(\xi) \xi^2\| \leq C_4\, \mu\lambda \|\ln \mu\|^2. \end{gathered}$$ Since ${\check{T}}^{(+)}= {\check{T}}(0+)$, ${\hat{T}}^{(+)}= {\hat{T}}(0+)$, the mean value theorem and imply $$\begin{gathered} \label{E5.15} \mu^{1/2} \Big\| \ov\ve: \big( \check{T}^{(+)}- {\check{T}}(\xi)\big) - \frac{\hat{\psi}'''(1)}{2\big(\hat{\psi}''(1)\big)^2} \Big((\ov\ve: \hat{T}^{(+)})^2- \big(\ov\ve: \hat{T}(\xi)\big)^2\Big) \Big\| \\ = \mu^{1/2} \|R(\xi)\xi\| \leq C_5\, \mu \lambda^{1/2} \|\ln \mu\|,\end{gathered}$$ where the remainder term $R$ belongs to $L^\infty ({\cal U}_\da)$. Combination of – yields for $(t,\eta,\xi) \in Q^{(\mu\la)}_{\rm match}$ with $\xi > 0$ that $$\label{E5.16} \|I_1\| \leq \mu \big(C_4 \lambda \|\ln \mu\|^2 + C_5 \lambda^{1/2} \|\ln \mu\|\big) \leq C_6\, \lambda^{1/2} \|\ln \mu\|^2 \mu.$$ From the definitions of $\tilde{\rho}_1$, $\tilde{\rho}_2$ in and we see that for $(t,\eta,\xi) \in Q^{(\mu\la)}_{\rm match}$ with $\xi <0$ the term $I_1$ takes the form $$\begin{aligned} I_1 &= \frac{1}{\hat{\psi}''(0)} \Bigg( \ov\ve: {\hat{T}}^{(-)} - \ov\ve: {\hat{T}}(\xi) + \si'(0) \xi \\ & \phantom{= \frac{1}{\hat{\psi}''(1)}} +\mu^{1/2} \big(\ov\ve: {\check{T}}^{(-)}- \ov\ve: {\check{T}}(\xi)\big) - \mu^{1/2} \frac{\hat{\psi}'''(0)}{2} \left(\Big(\frac{\ov\ve : {\hat{T}}^{(-)}}{\hat{\psi}''(0)}\Big)^2- \Big(\frac{\ov\ve: {\check{T}}(\xi)}{\hat{\psi}''(0)}\Big)^2\right). \end{aligned}$$ Using instead of , we see as above that the estimate also holds in this case, whence the estimate is valid for all $(t,\eta,\xi) \in Q^{(\mu\la)}_{\rm match}$. To finish the proof of , we combine with , , and to obtain the estimate $$\|S_1^{(\mu)}- S_2^{(\mu)}\| \leq K_1 \mu^{3/2} + \mu^{1/2} \big( C_6\, \lambda^{1/2} \|\ln \mu\|^2 \mu + C_3\, \mu^{3/2} \big) + \mu^{3/2} \|{\tilde{S}}_3\|,$$ which implies . Next we prove . From and , we conclude for $(t,x) \in \U_\delta$ that $$\begin{aligned} u_2^{(\mu)}(t,x) &= \hat{u}(t,x)+ \mu^{1/2} \check{u}(t,x)+ \mu \tilde{u}(t,x)\\ \begin{split} = u^ \xi^+ +a^* \frac{1}{2}(\xi^+)^2+ \mu^{1/2} u^* \Big(\frac{\ov\ve : \hat{T}^{(+)}}{\hat{\psi}''(1)} \xi^+ + \frac{\ov\ve : \hat{T}^{(-)}}{\hat{\psi}''(0)} \xi^- \Big)\\ \mbox{} + \hat{v}(t,x)+ \mu^{1/2} \check{v}(t,x) + \mu \tilde{u}(t,x). \end{split}\end{aligned}$$ Combination of this equation with and insertion of – yields with $\zeta= \frac{\xi}{(\mu\lambda)^{1/2}}$ that $$\begin{aligned} u_1^{(\mu)} (t,x) &- u_2^{(\mu)}(t,x) \nn \\ = (&\mu\lambda)^{1/2} u_0 \Big(\frac{\xi}{(\mu\lambda)^{1/2}}\Big) + \mu\lambda^{1/2} u_1 \Big( \frac{\xi}{(\mu\lambda)^{1/2}}\Big) + \mu\lambda u_2 \Big(\frac{\xi}{(\mu\lambda)^{1/2}}\Big) \nn \\ & - \left(u^\xi^+ + a^ \frac{1}{2}(\xi^+)^2+ \mu^{1/2} u^* \Big(\frac{\ov\ve : \hat{T}^{(+)}}{\hat{\psi}''(1)} \xi^+ + \frac{\ov\ve: \hat{T}^{(-)}}{\hat{\psi}''(0)} \xi^-\Big) + \mu \tilde{u}(\xi)\right) \nn \\ =(&\mu\la)^{1/2} u^* \Big(\int_{-\infty}^\zeta S_0(\vartheta)\,{\rm d}\vartheta-\zeta^+\Big) + \mu\lambda\, a^* \int_{-\infty}^\zeta \Big(\int_{-\infty}^\vartheta S_0(\vartheta_1)\, d \vartheta_1 - \vartheta^+\Big)\,d \vartheta \nn \\ & \mbox{}+ \mu\lambda^{1/2} u^\int_0^\zeta \Big( S_1(\vartheta) - \frac{\ov\ve: \hat{T}^{(+)}}{\hat{\psi}''(1)} 1^+(\vartheta) - \frac{\ov\ve: \hat{T}^{(-)}}{\hat{\psi}''(0)} 1^-(\vartheta)\Big) d \vartheta - \mu \tilde{u}(\xi) \nn \\[1ex] = (&\mu\lambda)^{1/2} J_1(\zeta)+ \mu\big(J_2(\zeta)+J_3(\zeta) - \tilde{u}(\xi)\big). \label{E5.17}\end{aligned}$$ To estimate the right hand side we use the boundary condition for $\tilde{u}$, note that by the equation $\tilde{\rho}_1(\zeta) = \frac{\ov\ve: \hat{T}^{(+)}}{\hat{\psi}''(1)}$ holds for $\zeta > 0$, and employ the inequalities , to compute for $(t,\eta,\xi) \in Q^{(\mu\la)}_{\rm match}$ with $\xi >0$ $$\begin{aligned} \begin{split} \|J_2(\zeta)+ J_3(\zeta)- \tilde{u}^{(+)}\| =\Big\|J_2(\zeta)-\lambda a^ \int_{-\infty}^\infty \Big(\int_{-\infty}^\vartheta S_0(\vartheta_1)\, d \vartheta_1 - \vartheta^+\Big)\, d\vartheta \\ \mbox{}+ J_3(\zeta)-\lambda^{1/2} u^\int_0^\infty \Big(S_1(\vartheta)- \frac{\ov\ve: \hat{T}^{(+)}}{\hat{\psi}''(1)}\Big)\,d\vartheta\Big\| \end{split} \nn\\ &=\Big\|\lambda a^\int_\zeta^\infty \Big(\int_{-\infty}^\vartheta S_0(\vartheta_1)\vartheta_1-\vartheta^+\Big)\,d\vartheta + \lambda^{1/2} u^\int_\zeta^\infty \big(S_1(\vartheta)- \tilde{\rho}_1(\zeta) \big)\, d\vartheta\Big\| \nn\\[1ex] &\leq \lambda \|a^\| \int_\zeta^\infty \frac{K_1}{a} e^{-a\vartheta}\, d\vartheta +\lambda^{1/2} \|u^\| \int_\zeta^\infty (K_2 + \tilde{K}) e^{-a\vartheta}\, d\vartheta \nn \\[1ex] &= \Big(\lambda \|a^\| \frac{K_1}{a^2} +\lambda^{1/2} \|u^\| \frac{K_2 + \tilde{K}}{a}\Big) e^{-a\zeta} \leq C_1 \la^{1/2} \big(\\|a^\\|_{L^\infty(\Gamma)} + \\|u^\\|_{L^\infty(\Gamma)}\big) e^{-a \zeta}. \label{E5.18} \end{aligned}$$ The mean value theorem implies for $\xi >0$ $$\tilde{u}(t,x)= \tilde{u}(t,\eta,\xi)= \tilde{u}(t,\eta,0+)+R_{\tilde{u}}(t,\eta,\xi)\xi.$$ Since $\tilde{u}(t,\eta,0+)= \tilde{u}^{(+)}(t,\eta)$, we infer from this equation and from , for all $(t,\eta,\xi) \in Q_{\rm match}^{(\mu\la)}$ with $\xi>0$ that $$\begin{aligned} \|J_2 (\zeta)& + J_3(\zeta) - \tilde{u}(\xi)\| \leq \|J_2(\zeta)+J_3(\zeta) - \tilde{u}^{(+)}\| + \|\tilde{u}(\xi) - \tilde{u}^{(+)}\| \nn\\ &\leq C_1 \la^{1/2} \big(\\|a^\\|_{L^\infty(\Gamma)}+ \\|u^\\|_{L^\infty(\Gamma)}\big) e^{-a \zeta} + \\|R_{\tilde{u}}\\|_{L^\infty (\U_\delta)} \xi \nn\\[1ex] &\leq C_1 \la^{1/2}\big(\\|a^\\|_{L^\infty(\Gamma)}+ \\|u^\\|_{L^\infty(\Gamma)}\big) e^{- \frac{3}{2}\|\ln\mu\|} + \\|R_{\tilde{u}}\\|_{L^\infty (\U_\delta)} \frac{3}{a}(\mu\lambda)^{1/2} \|\ln\mu\| \nn\\ & \leq C_2 \la^{1/2} \mu^{1/2} \|\ln\mu\|. \label{E5.19}\end{aligned}$$ Using the boundary condition for $\tilde{u}$ instead of , we see by the analogous computation that also holds for $(t,\eta,\xi) \in Q_{\rm match}^{(\mu\la)}$ with $\xi<0$. Now use to estimate the second term on the right hand side of . The first term is estimated by . Because yields $e^{-a\|\zeta\|} \leq e^{- \frac32 \|\ln \mu\|} = \mu^{3/2}$, we obtain for $(t,\eta,\xi) \in Q_{\rm match}^{(\mu\la)}$ $$\begin{gathered} \|u_1^{(\mu)}(t,x)-u_2^{(\mu)}(t,x)\| \leq (\mu\lambda)^{1/2} \|u^\|\, \Big\| \int_{-\infty}^\zeta S_0(\vartheta)\, d\vartheta-\zeta^+\Big\|+ \mu \big\|J_2(\zeta)+ J_3(\zeta)- \tilde{u}(\xi)\big\| \\[1ex] \leq (\mu\lambda)^{1/2} \big( \max_\Gm \|u^\| \big)\, \frac{K_1}{a} e^{-a\|\zeta\|} + C_2 \la^{1/2} \mu^{3/2} \|\ln \mu\| \leq C_1 \lambda^{1/2} \mu^2+ C_2 \la^{1/2}\mu^{3/2} \|\ln\mu\|, \end{gathered}$$ which implies . End of the proof of {#S5.4} -------------------- To complete the proof of note first that , imply $$u^{(\mu)}\rain{\pa\Om} = u^{(\mu)}_2\rain{\pa\Om}, \qquad \pa_{n_{\pa\Om}} S^{(\mu)}\rain{\pa\Om)} = \pa_{n_{\pa\Om}} S^{(\mu)}_2\rain{\pa\Om},\qquad S^{(\mu)}\rain{Q_{\rm inn}^{(\mu\la)}} = S^{(\mu)}_1\rain{Q_{\rm inn}^{(\mu\la)}}.$$ Therefore follows from the definition of $S_1^{(\mu)}$ in , equation follows from , , , and is a consequence of the definition of $u^{(\mu)}_2\rain{\pa\Om}$ in and of , , . Moreover, the estimate for the right hand side $f_3^{(\mu\la)}$ of follows from the definition of $S^{(\mu)}_2\rain{\pa\Om}$ in and from $\pa_{n_{\pa\Om}} {\hat{S}}\rain{\pa\Om} = 0$. This last equation holds, since by assumption $\Gm(t) \subseteq \Om$, which implies that ${\hat{S}}(t)$ is identically equal to $0$ or $1$ in a neighborhood of $\pa\Om$. follows from the definition of $T^{(\mu)}$ in ; equation is an immediate consequence of . It remains to verify the estimates – for the right hand sides $f_1^{(\mu\la)}$, $f_2^{(\mu\la)}$ of the equations and . To this end we put together all the estimates derived in Sections \[S5.1\] – \[S5.3\]. We start with the proof of and . Equation yields $$\na_x u^{(\mu)} = \na_x u_1^{(\mu)}\phi_{\mu\la} + \na_x u_2^{(\mu)}(1-\phi_{\mu\la}) + \big(u_1^{(\mu)} - u_2^{(\mu)}\big) \otimes \na_x \phi_{\mu\la}.$$ We insert this equation into and use and , to obtain $$\label{ETmuexplicite} T^{(\mu)} = T_1^{(\mu)} \phi_{\mu\la} + T_2^{(\mu)} (1-\phi_{\mu\la}) + D\ve\big( (u_1^{(\mu)} - u_2^{(\mu)}) \otimes \na_x \phi_{\mu\la} \big).$$ The function $\phi_{\mu\la}$ defined in is independent of $(t,\eta)$. The decomposition of the gradient thus yields $$\label{Enaxphi} \na_x \phi_{\mu\la} = \frac{2a}{3(\la\mu)^{1/2} \|\ln \mu \|} \phi_{\mu\la}'n,$$ with the unit normal vector $n=n(t,\eta)$ to $\Gm_\xi(t)$ at $\eta \in \Gm_\xi(t)$. We write $$\phi_{\mu\la}' = \phi'\big(\frac{2a\xi}{3 (\mu\la)^{1/2}\|\ln \mu\|} \big), \qquad \phi_{\mu\la}'' = \phi''\big(\frac{2a\xi}{3 (\mu\la)^{1/2}\|\ln \mu\|} \big),$$ by a slight abuse of notation. With and we compute $$\begin{aligned} \div_x T^{(\mu)} + {\sf b} &= (\div_x T_1^{(\mu)}+{\sf b} ) \phi_{\mu\la} + ( \div_x T_2^{(\mu)} + {\sf b} ) (1-\phi_{\mu\la}) \nn \\[1ex] & + \Big( (T_1^{(\mu)} - T_2^{(\mu)}) n + \div_x D\ve\big( (u_1^{(\mu)} - u_2^{(\mu)}) \otimes n \big) \Big) \frac{2a}{3(\la\mu)^{1/2} \|\ln \mu \|} \phi_{\mu\la}' \nn \\ & + \Big(D\ve\big( (u_1^{(\mu)} - u_2^{(\mu)}) \otimes n \big)\Big) n\, \Big(\frac{2a} {3 (\la\mu)^{1/2} \|\ln \mu \|} \Big)^2 \phi_{\mu\la}''. \label{EdivTmu} \end{aligned}$$ Inequality is an immediate consequence of this equation and of , since $\phi_{\mu\la} = 0$ in $Q_{\rm out}^{(\mu\la)}$, and is obtained by estimating the right hand side of using the obvious inequality $$\label{Eelliest4} \| \div_x D\ve\big((u_1^{(\mu)} - u_2^{(\mu)}) \otimes n \big) \| \leq C \big( \|\na_x (u_1^{(\mu)} - u_2^{(\mu)})\| + \|u_1^{(\mu)} - u_2^{(\mu)}\| \big)$$ and the equation and inequalities , , – . We next proof and . The inequality follows immediately from , since $\phi_{\mu\la} = 0$ on $Q_{\rm out}^{(\mu\la)}$, which by and implies $(u^{(\mu)},S^{(\mu)}) = (u_2^{(\mu)}, S_2^{(\mu)})$ on $Q_{\rm out}^{(\mu\la)} \subseteq Q \setminus \Gm$. It remains to verify . Since ${\sf W}_S (\ve,S) = -\ov{\ve}:D(\ve - \ov{\ve} S)$, by , it follows from , , and that $$\begin{aligned} {\sf W}_S\big( \ve(\na_x u^{(\mu)}),&\, S^{(\mu)}\big) - {\sf W}_S \big( \ve(\na_x u_1^{(\mu)}),S_1^{(\mu)}\big) = -\ov{\ve}:(T^{(\mu)}-T_1^{(\mu)}) \\ & = -\ov{\ve} : (T_2^{(\mu)}-T_1^{(\mu)})(1-\phi_{\mu\la}) - \ov{\ve} : D\ve \big( (u_1^{(\mu)} - u_2^{(\mu)})\otimes \na_x \phi_{\mu\la}\big). \end{aligned}$$ The mean value theorem and imply $$\begin{aligned} \hat{\psi}' (S^{(\mu)}) - \hat{\psi}' (S_1^{(\mu)}) &= \hat{\psi}'' \Big( S_1^{(\mu)} + \vartheta (S_2^{(\mu)} - S_1^{(\mu)})(1 - \phi_{\mu\la}) \Big) (S_2^{(\mu)} - S_1^{(\mu)})(1 - \phi_{\mu\la}), \nn \\ \intertext{for a suitable $0<\vartheta (t,x)<1$, and } \Da_x S^{(\mu)} - \Da_x S_1^{(\mu)} &= \Da_x ( S_2^{(\mu)} - S_1^{(\mu)}) (1- \phi_{\mu\la}) + 2\na_x ( S_1^{(\mu)} - S_2^{(\mu)}) \cdot \na_x \phi_{\mu\la} \nn \\ & + ( S_1^{(\mu)} - S_2^{(\mu)}) \Da_x \phi_{\mu\la}. \label{E5.DeltaS-S1} \end{aligned}$$ The right hand sides of the last three equations vanish on the set $Q_{\rm inn}^{(\mu\la)}$, since $\phi_{\mu\la} = 1$ on this set. On the set $Q_{\rm match}^{(\mu\la)}$ the right hand sides of these equations can be estimated using , , , , . In the estimation of we also note that since $\phi_{\mu\la}$ is independent of $(t,\eta)$, analogous to the equation $$\Da_x \phi_{\mu\la} = - \ka \pa_\xi\phi_{\mu\la} + \pa_\xi^2 \phi_{\mu\la} = - \ka \frac{2a}{3(\la \mu)^{1/2}\|\ln \mu\|} \phi'_{\mu\la} + \Big( \frac{2a}{3 (\la \mu)^{1/2}\|\ln \mu\|}\Big)^2 \phi''_{\mu\la}$$ holds, with twice the mean curvature $\ka(t,\eta,\xi)$ of the surface $\Gm_\xi(t)$ at $\eta \in \Gm_\xi(t)$. Together we obtain that on $Q_{\rm inn}^{(\mu\la)} \cup Q_{\rm match}^{(\mu\la)}$ the inequality $$\begin{gathered} \Big \| \Big ({\sf W}_S \f \ve (\na_x u^{(\mu)}),S^{(\mu)}\g + \frac{1}{\mu^{1/2}} \hat{\psi}'(S^{(\mu)})-\mu^{1/2} \la \Da_x S^{(\mu)}\Big ) \\ - \Big ({\sf W}_S \f \ve(\na_x u_1^{(\mu)}),S_1^{(\mu)}\g + \frac{1}{\mu^{1/2}} \hat{\psi}'(S_1^{(\mu)})-\mu^{1/2} \la \Da_x S_1^{(\mu)}\Big ) \Big \| \\ \le K \big( \mu + \mu \|\ln \mu\|^2 + \mu^{1/2}\la \mu^{1/2} \la^{-1} \big) \le K \mu \| \ln \mu\|^2 \label{E5.68}\end{gathered}$$ holds. Similarly, implies $$\pa_t S^{(\mu)} - \pa_t S_1^{(\mu)} = \pa_t (S_2^{(\mu)} - S_1^{(\mu)})(1 - \phi_{\mu\la}) + (S_1^{(\mu)} - S_2^{(\mu)}) \pa_t \phi_{\mu\la}.$$ The right hand side of this equation vanishes on $Q_{\rm inn}^{(\mu\la)}$. To estimate the right hand side on the set $Q_{\rm match}^{(\mu\la)}$ we use the inequalities , and the equation $$\pa_t \phi_{\mu\la} = -\frac{2as^{(\mu)}}{3(\mu \la)^{1/2} \|\ln \mu\|} \phi'_{\mu\la},$$ which follows from and . The result is $$\label{E5.69} \|\pa_t S^{(\mu)} - \pa_t S_1^{(\mu)} \| \le K\la^{-1/2} \mu \|\ln \mu\|, \qquad \text{on } Q_{\rm inn}^{(\mu\la)} \cup Q_{\rm match}^{(\mu\la)}.$$ By combination of , and we see that the inequality $$\begin{gathered} \Big\| \pa_t S^{(\mu)} + \frac{c}{(\mu \la)^{1/2}} \Big( {\sf W}_S \big( \ve(\na_x u^{(\mu)}), S^{(\mu)}\big) + \frac{1}{\mu^{1/2}} \hat{\psi}' (S^{(\mu)}) - \mu^{1/2} \la \Da_x S^{(\mu)}\Big) \Big\| \\ \le \Big\| \pa_t S_1^{(\mu)} + \frac{c}{(\mu \la)^{1/2}} \Big( {\sf W}_S \big(\ve(\na_x u_1^{(\mu)}), S_1^{(\mu)} \big) + \frac{1}{\mu^{1/2}} \hat{\psi}' (S_1^{(\mu)}) - \mu^{1/2} \la \Da_x S_1^{(\mu)}\Big) \Big\| \\ + K \la^{-1/2} \mu \|\ln \mu\| + \frac{c}{(\mu\la)^{1/2}} K\mu \| \ln \mu\|^2 \\ \le K \Big(\frac{\mu}{\la}\Big)^{1/2} \|\ln \mu\|^2 + K \frac{\mu}{\la^{1/2}}\|\ln \mu\| + c K \Big( \frac{\mu}{\la}\Big)^{1/2} \| \ln \mu\|^2 \le K_1 \Big(\frac{\mu}{\la} \Big)^{1/2} \|\ln \mu\|^2 \end{gathered}$$ holds on the set $Q_{\rm inn}^{(\mu\la)} \cup Q_{\rm match}^{(\mu\la)}$. This proves and completes the proof of . Proof of {#S6} ========= This section contains the proof of . $(u,T,S)$ denotes the asymptotic solution constructed in and $(u_{\rm AC},T_{\rm AC},S_{\rm AC})$ denotes the exact solution of – , – . For the proof we need a lemma and a theorem, which we state first. \[L6.1\] For all $x \in \Gm(\hat{t})$ the propagation speeds $s_{\rm AC}$ and $s$ satisfy $$\label{E6.differences} s_{\rm AC}(\hat{t},x) - s(\hat{t},x) = \frac{1}{\|\na_x S(\hat{t},x)\|} \Big( f_2^{(\mu\la)}(\hat{t} ,x) - \frac{c}{(\mu\la)^{1/2}}\, \ov{\ve}: \big( T_{\rm AC} (\hat{t}, x) - T (\hat{t}, x) \big) \Big),$$ where $f_2^{(\mu\la)}$ is the right hand side of equation . [Proof:]{} Since the manifold $\Gm$ is a level set of $S$ and since by the manifold $\Gm_{\rm AC}$ is a level set of $S_{\rm AC}$, it follows that $(\pa_t S(\hat{t},x), \na_x S(\hat{t},x) )$ and $(\pa_t S_{\rm AC}(\hat{t},x), \na_x S_{\rm AC}(\hat{t},x) )$ are normal vectors to the respective manifolds at $(\hat{t},x)$. Moreover, implies that $\na_x S_{\rm AC}(\hat{t},x) = \na_x S(\hat{t},x)$. From we thus infer that $$\begin{gathered} s(\hat{t},x) = \frac{- \pa_t S^(\hat{t},x)} {\na_x S(\hat{t},x) \cdot n(\hat{t},x) } = \frac{- \pa_t S(\hat{t},x)} {\|\na_x S(\hat{t},x)\|}, \label{E6.2} \\ s_{\rm AC}(\hat{t},x) = \frac{- \pa_t S_{\rm AC}(\hat{t},x)} {\|\na_x S_{\rm AC}(\hat{t},x)\| } = \frac{- \pa_t S_{\rm AC}(\hat{t},x)} {\|\na_x S(\hat{t},x)\|}. \label{E6.3} \end{gathered}$$ For brevity we do not write the argument $(\hat{t},x)$ in the following computation. In we eliminate $\pa_t S$ with the help of , and in we replace $\pa_t S_{\rm AC}$ by the right hand side of . Together with another application of this results in $$\begin{gathered} s_{\rm AC} - s = \frac{c}{(\mu\la)^{1/2} \|\na_x S\|} \Big( \big( \pa_S {\sf W} \big( \ve(\na_x u_{\rm AC}), S_{\rm AC} \big) + \frac{1}{\mu^{1/2}} \hat{\psi}'(S) - \mu^{1/2}\la\Da_x S \big) \\ - \big( \pa_S {\sf W} \big( \ve(\na_x u), S \big) + \frac{1}{\mu^{1/2}} \hat{\psi}'(S) - \mu^{1/2} \la \Da_x S \big) \Big) + \frac{1}{ \|\na_x S\|} f_2^{(\mu\la)} \\ = \frac{1} { \|\na_x S\|} f_2^{(\mu\la)} - \frac{c}{(\mu\la)^{1/2} \|\na_x S\|} \Big(\ov{\ve}: T_{\rm AC} - \ov{\ve}: T \Big). \end{gathered}$$ which is . In the last step we used that by and we have $$\pa_S {\sf W} \big( \ve(\na_x u_{\rm AC}), S_{\rm AC} \big) = - \ov{\ve}: T_{\rm AC}, \quad \mbox{and} \quad \pa_S {\sf W} \big( \ve(\na_x u), S \big) = - \ov{\ve}: T.$$ The proof of is complete. \[T6.2\] Suppose that the order of differentiability of $\hat{\psi}$, $\Gm$, ${\hat{u}}$, ${\check{u}}$, ${\sf b}$, is higher by two than required in . Assume that the principal curvatures $\ka_1^{(\la\mu)}$, $\ka_2^{(\la\mu)}$ of the regular $C^1$–manifold $\Gm(\hat{t}) = \Gm^{(\mu\la)}(\hat{t})$ are bounded, uniformly with respect to $\mu \in (0,\mu_0]$ and $\la \in (0,\la_0]$, and that there is an open subset $\Om' \subset\subset \Om$ and $\delta > 0$ such that the neighborhood ${\cal U}^{(\mu\la)}_\delta(\hat{t})$ of $\Gm^{(\mu\la)}(\hat{t})$ defined in satisfies ${\cal U}^{(\mu\la)}_\delta(\hat{t}) \subseteq \Om'$. Then there is a constant $K_5$ such that for all $\mu \in (0,\mu_0]$ and all $\la \in (0,\la_0]$ $$\label{E6.differenceT} \\| T_{\rm AC} - T \\|_{L^2(\Gm(\hat{t}))} \leq K_5 \| \ln \mu\|^3 \mu.$$ We postpone the proof of this theorem and first finish the proof of .\ [End of the proof of* ]{} By and we have for $x = (\eta,0) \in \Gm(\hat{t})$ that $$\begin{gathered} \label{E6.naxlow} \na_x S(\hat{t},x) = n (\pa_n S)(\hat{t},x) + \na_\Gm S(\hat{t},x) \\ = \frac{1}{(\mu\la)^{1/2}} \Big( S_0'(\zeta) + \mu^{1/2} \pa_\zeta S_1(t,\eta,\zeta) + \mu \pa_\zeta S_2(t,\eta,\zeta) \Big)\rain{\zeta = 0}n(\hat{t},\eta) \\ + \mu^{1/2} \na_\Gm S_1(\hat{t},\eta,0) + \mu \na_\Gm S_2(\hat{t},\eta,0). \end{gathered}$$ implies $$S_0'(0) = \sqrt{2 \hat{\psi}(S_0(0))} = \sqrt{2 \hat{\psi}(1/2)} > 0,$$ whence, from for $\mu \in (0,\mu_0]$ and $\la \in (0,\la_0]$ with $\mu_0$ sufficiently small, $$\begin{aligned} \| \na_x S(\hat{t},x)\| &\geq \frac{1}{(\mu\la)^{1/2}} \Big(\sqrt{2 \hat{\psi}(1/2)} - \mu^{1/2} \|\pa_\zeta S_1(t,\eta,0)\| - \mu \|\pa_\zeta S_2(t,\eta,0)\| \Big) \nn \\ & - \mu^{1/2} \|\na_\Gm S_1(\hat{t},\eta,0)\| - \mu \|\na_\Gm S_2(\hat{t},\eta,0)\| \geq \frac{1}{2(\mu\la)^{1/2}} \sqrt{2 \hat{\psi}(1/2)}. \label{E6.naxlow2} \end{aligned}$$ Combination of with the inequalities , and yields $$\begin{gathered} \\|s_{\rm AC}(\hat{t}) - s(\hat{t})\\|_{L^2(\Gm(\hat{t}))} \\ \leq \frac{1}{\min_{\Gm(\hat{t})} \|\na_x S(\hat{t})\|} \Big( \\| f_2^{(\mu\la)}(\hat{t})\\|_{L^2(\Gm(\hat{t}))} + \frac{c\|\ov{\ve}\|}{(\mu\la)^{1/2}} \big\\| T_{\rm AC} (\hat{t})-T (\hat{t}) \big\\|_{L^2(\Gm(\hat{t}))} \Big) \\ \leq \frac{2(\mu\la)^{1/2}}{\sqrt{2\hat{\psi}(1/2)}} \Big( \|\ln \mu\|^2 \Big(\frac{\mu}{\la}\Big)^{1/2} K_3\, {\rm meas} (\Gm(\hat{t}))^{1/2} + \frac{c\|\ov{\ve}\|}{(\mu\la)^{1/2}} K_5 \|\ln \mu\|^3 \mu \Big) \\ \leq K_6 \|\ln \mu\|^2 \mu + K_7 \|\ln \mu \|^3 \mu. \end{gathered}$$ follows from this estimate. The proof of is complete.\ [Proof of :]{} Note that the function $\big( u_{\rm AC}(\hat{t}),T_{\rm AC}(\hat{t})\big)$ solves the equations , in $\Om$ with $S_{\rm AC}(\hat{t}) = S(\hat{t})$, by the initial condition . Moreover, holds for $u_{\rm AC}(\hat{t})$. From the equations , , we thus conclude that the difference $(u_{\rm AC} - u, T_{\rm AC} - T)$ satisfies $$\begin{aligned} - \div_x ( T_{\rm AC} - T) (\hat{t} ) &=& - f_1^{(\mu\la)}(\hat{t}), \label{E6.linelast1} \\[1ex] (T_{\rm AC} - T)(\hat{t}) &=& D \ve \big( \na_x (u_{\rm AC} - u) (\hat{t}) \big), \label{E6.linelast2} \\[1ex] (u_{\rm AC} - u) (\hat{t},x) &=& 0, \qquad x \in \pa \Om. \label{E6.linelast3}\end{aligned}$$ This is the Dirichlet boundary value problem for the elliptic system of elasticity theory in the domain $\Om$. It suggests itself to derive the inequality by using the $L^2$–regularity theory of elliptic systems, which allows to estimate the norm $\\|T_{\rm AC}-T\\|_{L^2(\Gm(\hat{t}))}$ by the $L^2$–norm of the right hand side $-f^{(\mu\la)}_1(\hat{t})$ of . To apply this theory directly we would need that the $L^2$–norm of $f^{(\mu\la)}_1(\hat{t})$ decays to zero for $\mu \ra 0$ uniformly with respect to $\la$. However, the relation ${\rm meas} \big(Q^{(\mu\la)}_{\rm inn} (\hat{t}) \cup Q^{(\mu\la)}_{\rm match} (\hat{t})\big) \leq C_1 (\mu\la)^{1/2} \|\ln \mu\|$, which follows from , and the estimates , yield $$\\| f^{(\mu\la)}_1 (\hat{t})\\|_{L^2(\Om)} \leq \|\ln \mu\|^{5/2} \frac{\mu^{3/4}}{\lambda^{1/4}} K_1 C_1^{1/2}.$$ The right hand side does not decay to zero for $\mu \ra \infty$ uniformly with respect to $\la$, but blows up for $\la \ra 0$. Therefore direct application of the $L^2$–regularity theory is not possible. Before giving the detailed proof of we sketch how to circumvent this difficulty. Set $$\label{E6.4n} {\sf A}(\mu)= \frac{3}{a} \mu^{1/2} \|\ln \mu\|,$$ where $a>0$ is the constant defined in . By we have $$\label{E6.3n} Q^{(\mu\lambda)}_{\rm inn} (\hat{t}) \cup Q^{(\mu\lambda)}_{\rm match} (\hat{t})= \left\{(\eta,\xi) \in \mathcal{U}_\delta (\hat{t}) \,\big\vert\, \|\xi\| \leq {\sf A}(\mu)\lambda^{1/2} \right\}.$$ Define $$\label{E6.1n} \begin{split} f_{11}^{(\mu\lambda)}(x) &= \begin{cases} f_1^{(\mu\lambda)}(\hat{t},x), & x \in Q^{(\mu\lambda)}_{\rm inn} (\hat{t}) \cup Q^{(\mu\lambda)}_{\rm match} (\hat{t})\\ 0, & x \in Q^{(\mu\lambda)}_{\rm out} (\hat{t}), \end{cases}\\ f_{12}^{(\mu\lambda)}(x) &= \begin{cases} 0, & x \in Q^{(\mu\lambda)}_{\rm inn} (\hat{t}) \cup Q^{(\mu\lambda)}_{\rm match} (\hat{t})\\ f_1^{(\mu\lambda)}(\hat{t},x), & x \in Q^{(\mu\lambda)}_{\rm out} (\hat{t}), \end{cases} \end{split}$$ hence $f_1^{(\mu\la)}(\hat{t})=f_{11}^{(\mu\la)}+ f_{12}^{(\mu\la)}$. For $x = x(\hat{t},\eta,\xi) \in {\cal U}_\da(\hat{t})$ we write as usual $f^{(\mu\la)}_{11}(x) = f^{(\mu\la)}_{11}(\eta,\xi)$. From and we obtain for $\eta \in \Gm(\hat{t})$ and $-{\sf A}(\mu)\la^{1/2} \leq \xi \leq {\sf A}(\mu) \la^{1/2}$ that $$\label{E6.2n} \|f_{11}^{(\mu\la)}(\eta,\xi)\| \leq \|\ln \mu\|^2 \left(\frac{\mu}{\lambda}\right)^{1/2} K_1.$$ Define $\da_^{(\mu\la)}:\Gm(\hat{t}) \to \R$ by $$\label{E6.5n} \da_^{(\mu\la)} (\eta) = \int_{-{\sf A}(\mu)}^{{\sf A}(\mu)} \la^{1/2} f_{11}^{(\mu\la)} \big(\eta,\la^{1/2} \zeta \big)\, d\zeta.$$ and together imply that $$\label{E6.6n} \|\da_^{(\mu\la)}(\eta)\| \leq 2 {\sf A}(\mu) \la^{1/2} \|\ln \mu\|^2 \left(\frac{\mu}{\la}\right)^{1/2} K_1 = \frac{6}{a} K_1 \mu \|\ln \mu\|^3,$$ for all $\eta \in \Gm(\hat{t})$. Examination of the boundary value problem – suggests that for $\mu$ fixed and $\la \ra 0$ the solution $(u_{\rm AC}-u, T_{\rm AC}-T)$ converges to the solution $(u_, T_):\Omega \to \R^3 \times {\cal S}^3$ of the transmission problem $$\begin{aligned} -\div_x T_ &=& 0, \label{E6.7n}\\ T_* &=& D \varepsilon (\na_x u_), \label{E6.8n}\\ {[T_]} n &=& \da^{(\mu)}_, \qquad \mbox{on } \Gm(\hat{t}), \label{E6.9n}\\ {[u_]} &=& 0, \qquad \mbox{on } \Gm(\hat{t}), \label{E6.10n}\\ u_* (x) &=& 0, \qquad x \in \pa \Om, \label{E6.11n}\end{aligned}$$ where $\da_^{(\mu)} (\eta)= \lim_{\la \to 0} \da_^{(\mu\la)}(\eta)$ for $\eta \in \Gm(\hat{t})$. If this limit exists, it follows from that $$\|\da_^{(\mu)} (\eta)\| \leq \frac{6}{a} K_1 \mu \|\ln \mu\|^3.$$ This implies that the solution $(u_, T_)$ will be bounded by $C \mu \|\ln \mu\|^3$ with a suitable constant $C$, and this limit behavior suggests that though the $L^2$–norm of $f_1^{(\mu\la)}(\hat{t})$ blows up for $\la \ra 0$, the solution $(u_{\rm AC}-u, T_{\rm AC}-T)(\hat{t})$ of – is bounded by $C \mu \|\ln \mu\|^3$ with $C$ independent of $\la$. The reason for the blow up of $\\| f_1^{(\mu\la)}(\hat{t})\\|_{L^2(\Om)}$ for $\la \ra 0$ is therefore not that the norm of the solution $(u_{\rm AC}-u, T_{\rm AC}-T)(\hat{t})$ would blow up, but that the solution looses regularity in a neighborhood of the surface $\Gm(\hat{t})$, which is shown by the equation for the limit solution. This equation implies that $T_$ does not belong to the Sobolev space $W^{1,2}(\Om)$. In the following proof we do not study the limit $(u_, T_)$. Instead, based on the idea of the behavior of the regularity of $(u_{\rm AC}-u, T_{\rm AC}-T)(\hat{t})$, we decompose this function in the form $$(u_{\rm AC}-u, T_{\rm AC}-T)(\hat{t}) = \big(u^{(\la)}, T^{(\la)} \big) + \big(u^{(\la)}_, T^{(\la)}_ \big),$$ where $(u^{(\la)}, T^{(\la)}) = (u^{(\la)}_\mu, T^{(\la)}_\mu)$ is bounded by $C\mu \|\ln \mu\|^3$, uniformly with respect to $\la$, and for $\la \ra 0$ has the same regularity behavior as $(u_{\rm AC}-u, T_{\rm AC}-T)$, but otherwise does not approximate $(u_{\rm AC}-u, T_{\rm AC}-T)$. The construction is such that the difference $(u^{(\la)}_, T^{(\la)}_) = (u^{(\la)}_{\mu}, T^{(\la)}_{\mu}) = (u_{\rm AC}-u, T_{\rm AC}-T) - (u^{(\la)}, T^{(\la)})$ does not loose its regularity for $\la \ra 0$. Hence, we can use the standard $L^2$–theory for elliptic equations to show that also $(u^{(\la)}_, T^{(\la)}_)$ is bounded by $C \mu \|\ln \mu\|^3$ independently of $\la$. To construct $(u^{(\la)}, T^{(\la)})$ let ${\cal U}_\da(\hat{t})$ be the neighborhood of $\Gm(\hat{t})$ defined in and let $\phi_* \in C_0^\infty \big((-\delta,\delta)\big)$ be a function satisfying $$\label{E6.12n} \phi_* (\xi)=1, \quad -\delta/2 \leq \xi \leq \delta/2.$$ We set $$\begin{aligned} u^{(\la)} (x) &=& \begin{cases} \displaystyle \la^{1/2} V \Big(\la,\eta,\frac{\xi}{\la^{1/2}}\Big) \phi_* (\xi), & x=x(\hat{t},\eta,\xi) \in {\cal U}_\da(\hat{t}),\\ 0, & x \in \Om \setminus {\cal U}_\da(\hat{t}), \end{cases}\label{E6.13n} \\ T^{(\la)} (x) &=& D \ve \big(\na_x u^{(\la)} (x)\big), \hspace{2cm} x \in \Omega, \label{E6.14n} \end{aligned}$$ where the function $\zeta \mapsto V(\lambda,\eta,\zeta):[-\frac{\delta}{\lambda^{1/2}}, \frac{\delta}{\lambda^{1/2}}] \to \R^3$ is constructed as follows: We use the notations $V'=\partial_\zeta V$, $V''=\partial^2_\zeta V$. In the interval $[-{\sf A}(\mu), {\sf A}(\mu)]$ the function $V$ is the solution of the boundary value problem $$\begin{aligned} \Big( D\ve \big( V''(\la,\eta,\zeta) \otimes n\big) \Big) n &=& \la^{1/2} f_{11}^{(\mu\la)} (\eta,\la^{1/2} \zeta ), \qquad -{\sf A}(\mu) \leq \zeta \leq {\sf A}(\mu), \label{E6.15n}\\ V \big(\la,\eta, \pm {\sf A}(\mu)\big) &=& 0, \label{E6.16n}\end{aligned}$$ where $n = n(\eta)$ is the unit normal vector to $\Gm(\hat{t})$ at $\eta \in \Gm(\hat{t})$. The equation is a second order linear system of ordinary differential equations for the three components of $V$, which can be written in the form $$\label{E6.16nn} B V'' = \lambda^{1/2} f_{11}^{(\mu\lambda)},$$ with a $3 \ti 3$–matrix $B = B(\eta)$ defined by the equation $$\label{E6.16nnn} B \om = \big( D \ve(\om \otimes n) \big)n ,$$ which must hold for all $\om \in \R^3$. The matrix $B$ is symmetric and positive definite uniformly with respect to $\eta$. To see this, note that since the elasticity tensor $D: {\cal S}^3 \to {\cal S}^3$ is a linear, symmetric, positive definite mapping, we compute for $\om_1, \om_2 \in \R^3$ $$\begin{gathered} (B \om_1) \cdot \om_2 = \Big(\big(D \ve(\om_1 \otimes n)\big) n\Big) \cdot \om_2\\ = (\om_2 \otimes n): D \ve(\om_1 \otimes n) = \ve (\om_2 \otimes n): D \ve (\om_1 \otimes n)\\ = \big(D\ve (\om_2 \otimes n)\big) : \ve (\om_1 \otimes n) = \Big(\big(D \ve (\om_2 \otimes n)\big)n\Big) \cdot \om_1 = (B \om_2) \cdot \om_1.\end{gathered}$$ This shows that $B$ is symmetric. For $\om \in \R^3$ we have with a suitable constant $C_0 > 0$, which only depends on $D$ but is independent of $\eta$, that $$(B \om) \cdot \om= \ve (\om \otimes n):D \ve (\om \otimes n) \geq C_0 \|\ve(\om \otimes n)\|^2 \geq \frac{C_0}{2} \|\om\|^2 ,$$ hence $B$ is positive definite uniformly with respect to $\eta \in \Gm(\hat{t})$. Therefore the boundary value problem , has a unique solution $V$ on $[-{\sf A}(\mu), {\sf A}(\mu)]$. To extend $\zeta \mapsto V(\lambda,\eta,\zeta)$ to all of $[- \frac{\delta}{\lambda^{1/2}}, \frac{\delta}{\lambda^{1/2}}]$, we continue $V$ to the intervals $\big(- \frac{\delta}{\lambda^{1/2}}, -{\sf A}(\mu)\big)$ and $\big({\sf A}(\mu), \frac{\delta}{\lambda^{1/2}}\big)$ by affine functions: $$\label{E6.17n} V(\lambda,\eta,\zeta)= \begin{cases} \big(\zeta+{\sf A}(\mu)\big) V'\big(\lambda,\eta,-{\sf A}(\mu)\big), & -\frac{\delta}{\lambda^{1/2}} \leq \zeta \leq -{\sf A}(\mu),\\ \big(\zeta-{\sf A}(\mu)\big) V'\big(\lambda,\eta,{\sf A}(\mu)\big), & {\sf A}(\mu) \leq \zeta \leq \frac{\delta}{\lambda^{1/2}}\,. \end{cases}$$ By this extension, $\zeta \mapsto V(\lambda,\eta,\zeta)$ is continuously differentiable at $\zeta= \pm {\sf A}(\mu)$. For $x = x(\hat{t},\eta,\xi) \in {\cal U}_\da$ we use the notation $$V(\la,x) = V\big(\la,\eta,\frac{\xi}{\la^{1/2}}\big).$$ In the remaining part of the proof of we need the following lemma, which we prove first. \[L6.3\] There are constants $C_1, \dots, C_4$ such that for all $\mu \in (0,\mu_0]$, $\la \in (0,\la_0]$, $(\eta,\zeta) \in \Gm(\hat{t}) \times \big(- \frac{\da}{\la^{1/2}}, \frac{\da}{\la^{1/2}}\big)$ and $x \in {\cal U}_\da(\hat{t})$ the estimates $$\begin{aligned} \|\na_\eta^j V'(\la,\eta,\zeta)\| &\leq& C_1 \|\ln \mu\|^3 \mu, \qquad j=0,1,2, \label{E6.18n} \\ \|\la^{1/2} \na_\eta^j V(\la,\eta,\zeta)\| &\leq& C_2 \|\ln \mu\|^3 \mu, \qquad j=0,1,2, \label{E6.19n} \\ \|\la^{1/2} \na_x V(\la,x) \| &\leq& C_3 \| \ln \mu \|^3 \mu, \label{E6.21n} \\ \|\la^{1/2} \pa_{x_k} \na_{\Gm_\xi} V(\la,x) \| &\leq& C_4 \| \ln \mu \|^3 \mu, \qquad k = 1,\ldots ,3, \label{E6.22n}\end{aligned}$$ hold. Moreover, there is a function $g^{(\mu\la)}:\Om \ra \R^3$ and a constant $C_5$ such that $T^{(\la)}$ defined in satisfies $$\label{E6.23n} \div_x T^{(\la)} = f^{(\mu\la)}_{11} + g^{(\mu\la)},$$ with $$\label{E6.24n} \| g^{(\mu\la)}(x)\| \leq C_5 \| \ln \mu \|^3 \mu,$$ for all $x \in \Om$ and all $\mu \in (0,\mu_0]$, $\la \in (0,\la_0]$. [Proof:]{} In the following computations we drop the arguments $\la$ and $\eta$. Integration of yields $$\begin{aligned} B V' (\zeta) &= \int_{-{\sf A}(\mu)}^\zeta \la^{1/2} f^{(\mu\la)}_{11} (\la^{1/2}\vta) d\vta + BV'(-{\sf A}(\mu)), \label{E6.25n}\\ BV (\zeta) &= \int_{-{\sf A}(\mu)}^\zeta \int_{-{\sf A}(\mu)}^{\vta_1} \la^{1/2} f^{(\mu\la)}_{11} (\la^{1/2}\vta) d\vta d\vta_1 + \big( \zeta + {\sf A}(\mu) \big) BV'(-{\sf A}(\mu)), \label{E6.26n}\end{aligned}$$ where we used the boundary condition to get the second equation. Since $V({\sf A}(\mu)) = 0$, the relations and together yield $$\begin{gathered} 2 {\sf A}(\mu) \| BV'(-{\sf A}(\mu))\| = \Big\| - \int_{-{\sf A}(\mu)}^{{\sf A}(\mu)} \int_{-{\sf A}(\mu)}^{\vartheta_1} \la^{1/2} f^{(\mu\la)}_{11} d\vartheta d\vartheta_1 \Big\| \\ \leq \int_{-{\sf A}(\mu)}^{{\sf A}(\mu)} \int_{-{\sf A}(\mu)}^{\vartheta_1} \|\ln \mu\|^2 \mu^{1/2} K_1 d\vartheta d\vartheta_1 = 2 {\sf A}(\mu)^2 \|\ln \mu\|^2 \mu^{1/2} K_1\,, \end{gathered}$$ hence, by , $$\| BV'(-{\sf A}(\mu))\| \leq {\sf A}(\mu) \|\ln \mu\|^2 \mu^{1/2} K_1 = \frac{3}{a} K_1 \|\ln \mu\|^3 \mu.$$ Since $B = B(\eta)$ is positive definite uniformly with respect to $\eta$, this inequality implies the estimate for $j=0$ and $-{\sf A}(\mu) \leq \zeta \leq {\sf A}(\mu)$. Since by definition we have $V'(\zeta) = V'(-{\sf A}(\mu))$ for $\zeta \leq -{\sf A}(\mu)$ and $V'(\zeta) = V'({\sf A}(\mu))$ for ${\sf A}(\mu) \leq \zeta$, the estimate with $j=0$ holds also for the values of $\zeta$ outside of the interval $[-{\sf A}(\mu),{\sf A}(\mu)]$. To prove for $j=0$ we use that $V(-{\sf A}(\mu))=0$. By integration we thus obtain from for $\zeta \in [\frac{-\da}{\la^{1/2}},\frac{\da} {\la^{1/2}}]$ that $$\|V(\zeta) \| = \Big\| \int_{-{\sf A}(\mu)}^\zeta V'(\vartheta) d\vartheta \Big\| \leq \|\zeta + {\sf A}(\mu)\| C_1 \|\ln \mu\|^3 \mu \leq \big( \la^{-1/2} \da + {\sf A}(\mu)\big) C_1 \|\ln \mu\|^3 \mu,$$ which implies for $j=0$. To verify and for $j=1,2$ we differentiate the differential equation and the boundary condition with respect to $\eta$. For $j=1$ we obtain the differential equation $$B(\eta) (\pa_{\eta_k} V)'' = \la^{1/2} \big(\pa_{\eta_k} f^{(\mu\la)}_{11} - \pa_{\eta_k} B(\eta) B(\eta)^{-1} f^{(\mu\la)}_{11} \big),$$ and a similar equation for $j=2$. We then use the estimate $$\|\na_\eta^j f^{(\mu\la)}_{11}(\eta,\xi)\| = \|\na_\eta^j f^{(\mu\la)}_1(\eta,\xi)\| \leq \|\ln \mu\|^2 \Big(\frac{\mu}{\la}\Big)^{1/2} K, \qquad j =1,2.$$ This estimate is obtained by differentiation with respect to $\eta$ of the asymptotic expansions in Section \[S5.1\] leading to . Under the regularity assumptions in these derivatives exist. With this estimate we can employ the same arguments as above for the case $j=0$ to derive and for $j=1,2$. To prove we use the decomposition of the gradient and to compute $$\na_x V(\la,x) = \pa_\xi V\big(\la,\eta, \frac{\xi}{\la^{1/2}} \big) \otimes n + \na_{\Gm_\xi} V\big(\la,\eta, \frac{\xi}{\la^{1/2}} \big) = \la^{-\frac12} V'\otimes n + (\na_\eta V) A(\hat{t},\eta,\xi).$$ The right hand side is estimated by and to obtain . The estimate is obtained from and by similar decompositions. To prove , note that by , we have $T^{(\la)} = D \ve (\na_x u^{(\la)}) = D \ve \big(\na_x (\la^{1/2} V \phi_* )\big)$. Using and we therefore obtain by a similar computation as in that $$\begin{aligned} \div_x T^{(\la)} \nn &=& \div_x D\ve \Big(\na_x \big( \la^{1/2} V(\la,\eta,\frac{\xi}{\la^{1/2}}) \phi_(\xi)\big) \Big) \nn\\ &=& \Big( \la^{-1/2} \big( D\ve (V'' \otimes n)\big) n + \div_{\Gm_\xi} D\ve (V' \otimes n ) \Big) \phi_ \nn\\ && \mbox{} + \la^{1/2} \Big( D\ve ( \pa_\xi \na_{\Gm_\xi} V)\big)n + \div_{\Gm_\xi} D\ve (\na_{\Gm_\xi} V) \Big) \phi_* \nn\\ && \mbox{} + \Big( \big( D\ve( \la^{1/2} \na_x V)\big) n + \div_x D\ve(\la^{1/2} V \otimes n) \Big) \phi_' \nn\\ &&\mbox{} + \Big( D\ve(\la^{1/2} V \otimes n) \Big)n\, \phi_'' \nn\\ &=& f^{(\mu\la)}_{11} + g^{(\mu\la)}. \label{E6.28n}\end{aligned}$$ In the last step we used the differential equation and noted that for $\xi \in ([-\da,-{\sf A}(\mu)\la^{1/2}] \cup [{\sf A}(\mu)\la^{1/2},\da])$ we have $V''\big(\la,\eta,\frac{\xi}{\la^{1/2}}\big) = 0$, by definition of $V$ for such values of $\xi$ in . We also used that $\phi_(\xi) = 1$ for $\xi \in [-{\sf A}(\mu)\la^{1/2},{\sf A}(\mu)\la^{1/2}]$, which follows from and , since we have chosen $\mu_0$ and $\la_0$ small enough such that ${\sf A}(\mu)\la^{1/2} < \da/2$ for all $0 < \mu\leq \mu_0$ and $0< \la \leq \la_0$. The function $g^{(\mu\la)}$ is the sum of terms number $2$ to $7$ in the middle expression of equation . If we examine everyone of these six terms and apply – and also note that the functions $\phi_$, $\phi_'$ and $\phi_''$ are bounded independently of $\mu$ and $\la$ and vanish outside of ${\cal U}_\da(\hat{t})$, which follows from $\phi_* \in C_0^\infty((-\da, \da))$, we see that holds for $g^{(\mu\la)}$. This completes the proof of .\ To conclude the proof of let $( u^{(\la)}_,T^{(\la)}_)$ be the solution of the boundary value problem $$\begin{aligned} -\div_x T^{(\la)}_* &=& g^{(\mu\la)} - f^{(\mu\la)}_{12}, \label{E6.30n}\\ T^{(\la)}_* &=& D\ve(\na_x u^{(\la)}_), \label{E6.31n}\\ u^{(\la)}_(x) &=& 0, \qquad x \in \pa \Om. \label{E6.32n} \end{aligned}$$ From these equations and from , , we see that the function $( u^{(\la)} + u^{(\la)}_,T^{(\la)} + T^{(\la)}_)$ satisfies $$\begin{aligned} -\div_x (T^{(\la)} + T^{(\la)}_) &=& - f^{(\mu\la)}_{11} - g^{(\mu\la)} + g^{(\mu\la)} - f^{(\mu\la)}_{12} = - f^{(\mu\la)}_1(\hat{t}), \\ (T^{(\la)} + T^{(\la)}_) &=& D\ve\big(\na_x (u^{(\la)} + u^{(\la)}_) \big), \\ (u^{(\la)} + u^{(\la)}_)(x) &=& 0, \qquad x \in \pa \Om,\end{aligned}$$ hence $( u^{(\la)} + u^{(\la)}_,T^{(\la)} + T^{(\la)}_)$ is equal to the unique solution of the boundary value problem – , which means that $( u^{(\la)} + u^{(\la)}_,T^{(\la)} + T^{(\la)}_) = (u_{\rm AC} - u,T_{\rm AC} - u)(\hat{t})$. Consequently, we have $$\label{E6.35n} \\| T_{\rm AC} - T\\|_{L^2(\Gm(\hat{t}))} \leq \\| T^{(\la)} \\|_{L^2(\Gm(\hat{t}))} + \\| T^{(\la)}_\\|_{L^2(\Gm(\hat{t}))}.$$ To estimate $\\| T^{(\la)}_\\|_{L^2(\Gm(\hat{t}))}$ we can use the theory of interior regularity for the elliptic boundary value problem – . By this theory there is a constant $C$ such that $\\| u^{(\la)}_\\|_{W^{2,2}(\Om')} \leq C \\| g^{(\mu\la)} - f^{(\mu\la)}_{12} \\|_{L^2(\Om)}$, where $\Om'$ is the subdomain of $\Om$ introduced in , hence by the Sobolev embedding theorem and by , $$\label{E6.regularityest} \\|T^{(\la)}_\\|_{L^2(\Gm(\hat{t}))} \leq C_1 \\|T^{(\la)}_\\|_{W^{1,2}(\Om')} \leq C_2 \\| g^{(\mu\la)} - f^{(\mu\la)}_{12} \\|_{L^2(\Om)},$$ where by our assumptions on $\Gm^{(\mu\la)}(\hat{t})$ in the constants $C_1,C_2$ can be chosen independently of $\mu$ and $\la$. By definition of $f^{(\mu\la)}_{12}$ in and by we have $\|f^{(\mu\la)}_{12}(x)\| \leq \mu^{3/2} K_2$ for all $x \in \Om$. From this inequality, from and from we conclude that $$\begin{gathered} \label{E6.36n} \\|T^{(\la)}_\\|_{L^2(\Gm(\hat{t}))} \leq C_2 \Big(\int_\Om \big(\|g^{(\mu\la)}(x)\| + \|f^{(\mu\la)}_{12}(x)\|\big)^2 dx\Big)^{1/2} \\ \leq C_2 (C_5 \| \ln \mu \|^3 \mu + \mu^{3/2} K_2) \Big(\int_\Om dx\Big)^{1/2} \leq K \|\ln \mu\|^3 \mu. \end{gathered}$$ From , and from the inequalities , we infer that $$\begin{gathered} \| T^{(\la)}(x)\| \leq C \| \na_x u^{(\la)}(x)\| = C \|\na_x \big(\la^{1/2} V(\la,x) \phi_(\xi) \big)\| \\ = C \| \big( \la^{1/2} \na_x V(\la,x)\big) \phi_(\xi) + \la^{1/2} V(\la,x) \otimes \big(n \phi_'(\xi)\big) \| \leq K' \|\ln \mu\|^3 \mu, \end{gathered}$$ whence $$\\| T^{(\la)} \\|_{L^2(\Gm(\hat{t}))} \leq K'' \|\ln \mu\|^3 \mu.$$ Combination of this inequality with and yields . The proof of is complete. [99]{} H. Abels and S. Schaubeck, Sharp interface limit for the Cahn-Larché system, [Asymptotic Analysis]{} [91]{} No.3–4 (2015), 283–340. R. Abeyaratne and J.K. Knowles, On the driving traction acting on a surface of strain discontinuity in a continuum. [J. Mech. Phys. Solids]{} [38,3]{} (1990), 345-360. H.-D. Alber, Evolving microstructure and homogenization, [Continum. Mech. Thermodyn.]{} [12]{} (2000), 235-286. H.-D. Alber, Evolution of phase interfaces by configurational forces: A phase field model. Thermodynamische Materialtheorien, Oberwolfach 12.2.2004 – 18.12.2004. [Oberwolfach Reports]{} [ 1,4]{} (2004), 2981–2985 H.-D. Alber and Peicheng Zhu, Solutions to a Model with nonuniformly parabolic terms for phase evolution driven by configurational forces, [SIAM J. Appl. Math.]{} [66,2]{} (2006), 680–699. H.-D. Alber and Peicheng Zhu, Evolution of phase boundaries by configurational forces. [Arch. Rational Mech. Anal.]{} [185,2]{} (2007), 235 – 286. H.-D. Alber and Peicheng Zhu, Interface motion by interface diffusion driven by bulk energy: Justification of a diffusive interface model, [Continuum Mech. Thermodyn.]{} [ 23]{} No.2, (2011), 139 – 176. H.-D. Alber and Peicheng Zhu, Comparison of a rapidely converging phase field model for interfaces in solids with the Allen-Cahn model. [J. Elast.]{} [111]{} No.2, (2013), 153 – 221. R.F. Almgren, Second-order phase field asymptotics for unequal conductivities, [SIAM J. Appl. Math.]{} [59]{} No.6 (1999), 2086-2107 N.D. Alikakos, P.W. Bates and Xinfu Chen, Convergence of the Cahn-Hilliard equation to the Hele-Shaw model, [ Arch. Rational Mech. Anal.]{} [128]{} (1994), 165–205. K. Bhattachary. [Microstructure of martensite]{}. Oxford University Press, Oxford, 2003. G. Buratti, Y. Huo and I. Müller, Eshelby tensor as a tensor of free enthalpy, [J. Elast.]{} 72 (2003), 3142. G. Caginalp, The role of microscopic anisotropy in the macroscopic behavior of a phase boundary, [Annals. of Physics]{} [172]{} (1986), 136–155. G. Caginalp and X. Chen. Convergence of the phase field model to its sharp interface limits, [Euro J. Appl. Math.]{} [9]{} (1998), 417–445 & 795–801. E. A. Carlen, M. C. Carvalho and E. Orlandi. Approximate solutions of the [C]{}ahn-[H]{}illiard equation via corrections to the [M]{}ullins-[S]{}ekerka motion. [ Arch. Ration. Mech. Anal.]{} [178]{} No.1 (2005), 1–55. Xinfu Chen, Spectrums for the Allen-Cahn, Cahn-Hilliard, and phase-field equations for generic interface, [ Comm. Partial Diff. Eqns.]{} [19]{} (1994), 1371-1395. Xinfu Chen, G. Caginalp and Ch. Eck, A rapidely converging phase field model, [Discrete Contin. Dyn. Syst.]{} [15]{}, No. 4 (2006), 1017-1034. Xinfu Chen, J. Hong, and F. Yi, Existence, uniqueness, and regularity of classical solutions of the Mullins-Sekerka problem. [Comm. Partial Differential Equations]{}, [21]{} (1996), 1705–1727. J. Escher and G. Simonett, On Hele-Shaw models with surface tension, [Math. Res. Lett.]{} [3]{} No. 4 (1996), 467–474. J. Escher and G. Simonett, Classical solutions for Hele-Shaw models with surface tension, [Adv. Differ. Equ.]{} [2]{} No.4 (1997), 619–642. J. Escher and G. Simonett, A center manifold analysis for the Mullins-Sekerka model. [J. Differential Equations]{}, [143]{} No.2 (1998), 267–292. P. Fife and O. Penrose, Interfacial dynamics for thermodynamically consistent phase field models with nonconserved order parameter, [Electron. J. Diff. Eq.]{} [1]{} (1995), 1–49. E. Fried and M. Gurtin, Dynamic solid-solid transitions with phase characterized by an order parameter, [Phys. D.]{} [72]{} (1994), 287–308. H. Garcke and B. Stinner, Second order phase field asymptotics for multi-component systems, [Interfaces Free Bound.]{} [ 8]{} No. 2 (2006), 131–157. A. Karma and W.-J. Rappel, Quantitative phase-field modeling of dendritic growth in two and three dimensions, [Phys. Rev. E]{} [57]{} (1998), 4323–4349. T. Kato, Perturbation theory for linear operators, Springer (1966) S. Luckhaus, Solutions for the two-phase Stefan problem with the Gibbs-Thomson law for the melting temperature, [ Eur. J. Appl. Math.]{} [1]{} No.2, (1990), 101–111. P. de Mottoni and M. Schatzman, Geometrical evolution of developed interfaces, [Trans. Amer. Math. Soc.]{} [347]{} (1995), 1533–1589. D. Schrade, R. Mueller, B.X. Xu and D. Gross, Domain evolution in ferroelectric materials: A continuum phase field model and finite element implementation, [ Comput. Methods Appl. Engrg.]{} [196]{} (2007), 4365–4374. Bai-Xiang Xu, D. Schrade, R. Müller, D. Gross, T. Granzow and J. Rödel, Phase field simulation and experimental investigation of the electro-mechanical behavior of ferroelectrics, [Z. Angw. Mat. Mech.]{} [90]{} No. 7–8 (2010), 623–632. W. Zhang and K. Bhattacharya, A computational model of ferroelectric domains, Part I: model formulation and domain switching, [Acta Materialia]{} [53]{} (2005), 182–198. Yinan Zuo, Y.A. Genenko, A. Klein, P. Stein and Baixiang Xu, Domain wall stability in ferroelectrics with space charges, [J. Appl. Phys.]{} [115*]{}, 084110 (2014). [^1]: [email protected]	{ "pile_set_name": "ArXiv" }
--- abstract: 'In this contribution, a comprehensive review of the main aspects of high ${\rm p_{\rm T}}$ jet physics in Run II at the Tevatron is presented. Recent measurements on inclusive jet production are discussed using different jet algorithms and covering a wide region of jet transverse momentum and jet rapidity. Several measurements, sensitive to a proper description of soft gluon radiation and the underlying event in hadron collisions, are shown. Finally, high ${\rm p_{\rm T}}$ prompt photon measurements and studies on the production of electroweak bosons in association with jets in the final state are discussed.' author: - 'M. Martínez-Pérez' title: 'High ${\rm p_{\rm T}}$ Jet Physics' --- INCLUSIVE JET PRODUCTION ======================== The measurement of the inclusive jet cross section in $p\overline{p}$ collisions at $\sqrt{s} = 1.96 \ \rm TeV$ constitutes a stringent test of perturbative QCD (pQCD) predictions over almost nine orders of magnitude. The increased center-of-mass energy and integrated luminosity in Run II at the Tevatron, compared to Run I, allows to extend the measured jet cross section to jets with transverse momentum, ${\rm p_{\rm T}^{\rm jet}}$, above $650 \ {\rm GeV/c}$, and to search for signals of quark compositeness down to $\sim 10^{-19} \rm m$. The pQCD calculations are written as matrix elements, describing the hard interaction between partons, convoluted with parton density functions (PDFs) in the proton and antiproton that require input from the experiments. Inclusive jet cross section measurements from Run I at the Tevatron [@d0runI; @runIjet], performed in different jet rapidity regions, have been used to partially constrain the gluon distribution in the proton. The pQCD predictions are affected by the still limited knowledge on the gluon PDF, which translates into a big uncertainty on the theoretical cross sections at high ${\rm p_{\rm T}^{\rm jet}}$. The hadronic final states in hadron-hadron collisions are characterized by the presence of soft contributions (the so-called [[underlying event]{}]{}) from initial-state gluon radiation and multiple parton interactions between remnants, in addition to the jets of hadrons originated by the hard interaction. A proper comparison with pQCD predictions at the parton level requires an adequate modeling of these soft contributions which become important at low ${\rm p_{\rm T}^{\rm jet}}$. In Run II, both CDF and D0 experiments explore new jet algorithms following the theoretical work that indicates that the cone-based jet algorithm employed in Run I is not infrared safe and compromises a future meaningful comparison with pQCD calculations at NNLO. The CDF experiment recently published results [@ktprl; @runIIjet] on inclusive jet production using the ${\rm k_{\rm T}}$ [@ktalgo; @soper] and midpoint [@midpoint] algorithms for jets with ${\rm p_{\rm T}^{\rm jet}}> 54 \ \rm GeV/c$ and rapidity in the region $0.1 < \| {\rm y^{\rm jet}}\| < 0.7$, which are well described by NLO pQCD [@jetrad] predictions [^1]. CDF has presented new measurements [@cdfwww] of the inclusive jet production cross section as a function of ${\rm p_{\rm T}^{\rm jet}}$ in five different jet rapidity regions up to $\| {\rm y^{\rm jet}}\| = 2.1$, based on $1.0 \ \rm fb^{-1}$ of CDF Run II data. ![ Measured inclusive differential jet cross sections, using the ${\rm k_{\rm T}}$ algorithm with $\rm D=0.7$, (black dots) as a function of ${\rm p_{\rm T}^{\rm jet}}$ for jets with ${\rm p_{\rm T}^{\rm jet}}> 54$ GeV/c in different $\|{\rm y^{\rm jet}}\|$ regions compared to NLO pQCD predictions (open circles). The shaded bands show the total systematic uncertainty on the measurements. The dashed lines indicate the PDF uncertainty on the theoretical predictions. For presentation, each measurement is scaled by a given factor.[]{data-label="xs"}](xs.eps){width="8"} Figure \[xs\] shows the measured cross sections using the ${\rm k_{\rm T}}$ algorithm, $$\rm k_{\rm T,i} = p_{T,i}^2 \ \ \ \rm{;} \ \ \rm k_{\rm T, (i,j)} = \rm min(p_{T,i}^2,p_{T,j}^2) \cdot {\Delta R_{i,j}^2}/{D^2},$$ with $\rm D=0.7$, as a function of ${\rm p_{\rm T}^{\rm jet}}$ in five different $\| {\rm y^{\rm jet}}\|$ regions compared to NLO pQCD predictions where, for presentation, each measurement has been scaled by a given factor. The measured cross sections decrease by more than seven to eight orders of magnitude as ${\rm p_{\rm T}^{\rm jet}}$ increases. Figure \[ratio\] shows the ratios data/theory as a function of ${\rm p_{\rm T}^{\rm jet}}$ in the five different $\|{\rm y^{\rm jet}}\|$ regions. Good agreement is observed in the whole range in ${\rm p_{\rm T}^{\rm jet}}$ and ${\rm y^{\rm jet}}$ between the measured cross sections and the theoretical predictions. In particular, no significant deviation from the pQCD prediction is observed for central jets at high ${\rm p_{\rm T}^{\rm jet}}$. ![Ratio Data/Theory as a function of ${\rm p_{\rm T}^{\rm jet}}$ in different $\|{\rm y^{\rm jet}}\|$ regions. The error bars (shaded bands) show the total statistical (systematic) uncertainty on the data. A $5.8 \%$ uncertainty on the luminosity is not included. The dashed lines indicate the PDF uncertainty on the theoretical predictions. The dotted lines present the ratios of MRST2004 and CTEQ6.1M predictions. The dotted-dashed lines show the ratios of predictions with $2 \mu_0$ and $ \mu_0$.[]{data-label="ratio"}](ratio.eps){width="8.5"} In the most forward region, the uncertainty on the measured cross section at high ${\rm p_{\rm T}^{\rm jet}}$, compared to that on the theoretical prediction, indicates that the data will contribute to a better understanding of the gluon PDF. In the region $0.1 < \|{\rm y^{\rm jet}}\| < 0.7$, the analysis is repeated using different values for $\rm D$ in the ${\rm k_{\rm T}}$ algorithm: $\rm D=0.5$ and $\rm D=1.0$ (see Figure \[xsd\]). In both cases, good agreement is observed between the measured cross sections and the NLO pQCD predictions in the whole range in ${\rm p_{\rm T}^{\rm jet}}$. As $\rm D$ decreases, the measurement is less sensitive to contributions from multiple proton-antiproton interactions, and the presence and proper modeling of the underlying event. These measurements support the validity of the ${\rm k_{\rm T}}$ algorithm to search for jets in hadron collisions that will be further explored at the LHC energies. ![Measured inclusive differential jet cross sections (black dots) as a function of ${\rm p_{\rm T}^{\rm jet}}$ for jets with ${\rm p_{\rm T}^{\rm jet}}> 54$ GeV/c and $0.1 < \|{\rm y^{\rm jet}}\| < 0.7$ using $\rm D=0.5$ (left) and $\rm D=1.0$ (right), compared to NLO pQCD predictions. The shaded bands show the total systematic uncertainty on the measurements. A $5.8 \%$ uncertainty on the luminosity is not included. The dashed lines indicate the PDF uncertainty on the theoretical predictions. (bottom) Magnitude of the parton-to-hadron corrections, $\rm C_{\rm HAD}({\rm p_{\rm T}^{\rm jet}})$, used to correct the NLO pQCD predictions for $\rm D=0.5$ (left) and $\rm D=1.0$ (right). The shaded bands indicate the quoted Monte Carlo modeling uncertainty.[]{data-label="xsd"}](xsd.eps){width="8.5"} Figure \[d0QCD1\] shows the measured inclusive jet cross section by D0 [@d0www] based on the first 800 ${\rm pb}^{-1}$ of Run II data. The new midpoint jet algorithm has been used with a cone size R=0.7. The measurements have been performed for central jets in two different ${\rm y^{\rm jet}}$ bins. ![Measured inclusive jet cross section as a function of ${\rm p_{\rm T}^{\rm jet}}$ in different $\|{\rm y^{\rm jet}}\|$ ranges (dots), compared to pQCD NLO predictions [@d0nlo] (full lines). For presentation one of the measurements is scaled by a given factor.[]{data-label="d0QCD1"}](d0QCD1.eps){width="7"} Figure \[d0QCD2\] presents the ratio data vs NLO pQCD predictions [@d0nlo] as a function of ${\rm p_{\rm T}^{\rm jet}}$ for jets in the region $\|{\rm y^{\rm jet}}\|< 0.4$. In order to eliminate the uncertainty on the integrated luminosity the ratio has been normalized to unity at ${\rm p_{\rm T}^{\rm jet}}= 100$ GeV/c. The data is in good agreement with the pQCD NLO predictions. ![Ratio data/theory as a function of ${\rm p_{\rm T}^{\rm jet}}$ for jets in the region $\|{\rm y^{\rm jet}}\|<0.4$. The band indicates the uncertainty on the data and the dashed lines the uncertainty on the NLO prediction due to PDFs. The data is scaled to the theoretical prediction at ${\rm p_{\rm T}^{\rm jet}}= 100$ GeV/c.[]{data-label="d0QCD2"}](d0QCD2.eps){width="7"} UNDERLYING EVENT STUDIES ======================== As mentioned in the previous section, the hadronic final states at the Tevatron are characterized by the presence of soft underlying emissions, usually denoted as [[underlying event]{}]{}, in addition to highly energetic jets coming from the hard interaction. The underlying event contains contributions from initial- and final-state soft gluon radiation, secondary semi-hard partonic interactions and interactions between the proton and anti-proton remnants that cannot be described by perturbation theory. These processes must be approximately modeled using Monte Carlo programs tuned to describe the data. The jet energies measured in the detector contain an underlying event contribution that has to be subtracted in order to compare the measurements to pQCD predictions. Hence, a proper understanding of this underlying event contribution is crucial to reach the desired precision in the measured jet cross sections. In the analysis presented here, the underlying event in dijet events has been studied by looking at regions well separated from the leading jets, where the underlying event contribution is expected to dominate the observed hadronic activity. Jets have been reconstructed using tracks with $p_T^{\rm track} > 0.5$ GeV and $\|\eta^{\rm track}\|<1$ and a cone algorithm with R=0.7. ![Scheme of the different $\phi$ regions defined around the leading jet.[]{data-label="rf1"}](RF1.eps){width="3.5"} ![Measured average track density in the transverse region as a function of the $E_T^{\rm jet}$ of the leading jet. The measurements are compared to different Monte Carlo models.[]{data-label="rf2"}](RF3.eps){width="8"} Figure \[rf2\] shows the average track density in the transverse region as a function of $E_T^{\rm jet}$ of the leading jet for the dijet inclusive sample and for events where the leading jets are forced to be back-to-back in $\phi$, in order to further reduce extra hard-gluon radiation. The observed plateau indicates that the underlying event activity is, to a large extend, independent from the hard interaction. The measurements have been compared to the predictions from PYTHIA [@pythia] and HERWIG [@herwig] Monte Carlo programs including leading-order QCD matrix elements plus initial and final parton showers. The PYTHIA samples have been created using a special tuned set of parameters, denoted as PYTHIA-Tune A [@tunea], which includes an enhanced contribution from initial-state soft gluon radiation and a tuned set of parameter to control secondary parton interactions. It was determined as a result of similar studies of the underlying event performed using CDF Run I data [@underlying]. PYTHIA-Tune A describes the hadronic activity in transverse region while HERWIG underestimates the radiation at low $E_T^{\rm jet}$. JET SHAPES ========== The internal structure of jets is dominated by multi-gluon emissions from the primary final-state parton. It is sensitive to the relative quark- and gluon-jet fraction and receives contributions from soft-gluon initial-state radiation and beam remnant-remnant interactions. The study of jet shapes at the Tevatron provides a stringent test of QCD predictions and tests the validity of the models for parton cascades and soft-gluon emissions in hadron-hadron collisions. ![Integrated jet shapes definition[]{data-label="psi"}](psi.eps){width="3.5"} The CDF experiment has presented results [@shapes] on jet shapes for central jets with transverse momentum in the region $37 < {\rm p_{\rm T}^{\rm jet}}< 380$ GeV, where jets are searched for using the midpoint[^2] algorithm and a cone size $R=0.7$. The integrated jet shape, $\Psi(r)$, is defined as the average fraction of the jet transverse momentum that lies inside a cone of radius $r$ concentric to the jet cone: $$\Psi(r) = \frac{1}{\rm N_{jet}} \sum_{\rm jets} \frac{P_T(0,r) }{P_T(0,R)}, \ \ \ \ 0 \leq r \leq R$$ where $\rm N_{\rm jet}$ denotes the number of jets. The measured jet shapes have been compared to the predictions from PYTHIA-Tune A and HERWIG Monte Carlo programs. ![The measured integrated jet shape, $\Psi(r/R)$, in inclusive jet production for jets with $0.1 < \|{\rm y^{\rm jet}}\| < 0.7$ and $37 \ {\rm GeV/c} < {\rm p_{\rm T}^{\rm jet}}< 45 \ {\rm GeV/c}$. The predictions of PYTHIA-Tune A (solid lines), PYTHIA (dashed-dotted lines), PYTHIA-(no MPI) (dotted lines) and HERWIG (dashed lines) are shown for comparison.[]{data-label="sh1"}](sh1.eps){width="9.0"} In addition, two different PYTHIA samples have been used with default parameters and with and without the contribution from multiple parton interactions (MPI) between proton and antiproton remnants, the latter denoted as PYTHIA-(no MPI), to illustrate the importance of a proper modeling of soft-gluon radiation in describing the measured jet shapes. Figure \[sh1\] presents the measured integrated jet shapes, $\Psi(r/R)$, for jets with $37 < {\rm p_{\rm T}^{\rm jet}}< 45$ GeV, compared to HERWIG, PYTHIA-Tune A, PYTHIA and PYTHIA-(no MPI) predictions. ![The measured $1 - \Psi(0.3/R)$ as a function of ${\rm p_{\rm T}^{\rm jet}}$ for jets with $0.1 < \|{\rm y^{\rm jet}}\| < 0.7$ and $37 \ {\rm GeV/c} < {\rm p_{\rm T}^{\rm jet}}< 380 \ {\rm GeV/c}$. Error bars indicate the statistical and systematic uncertainties added in quadrature. The predictions of PYTHIA-Tune A (solid line) , PYTHIA (dashed-dotted line), PYTHIA-(no MPI) (dotted line) and HERWIG (dashed line) are shown for comparison.[]{data-label="sh2"}](sh2.eps){width="9.0"} Figure \[sh2\] shows, for a fixed radius $r_0 = 0.3$, the average fraction of the jet transverse momentum outside $r=r_0$, $1-\Psi(r_0/R)$, as a function of ${\rm p_{\rm T}^{\rm jet}}$ where the points are located at the weighted mean in each ${\rm p_{\rm T}^{\rm jet}}$ range. The measurements show that the fraction of jet transverse momentum at a given fixed $r_0/R$ increases ($1-\Psi(r_0/R)$ decreases) with ${\rm p_{\rm T}^{\rm jet}}$, indicating that the jets become narrower as ${\rm p_{\rm T}^{\rm jet}}$ increases. PYTHIA with default parameters produces jets systematically narrower than the data in the whole region in ${\rm p_{\rm T}^{\rm jet}}$. The contribution from secondary parton interactions between remnants to the predicted jet shapes (shown by the difference between PYTHIA and PYTHIA-(no MPI) predictions) is relatively small and decreases as ${\rm p_{\rm T}^{\rm jet}}$ increases. PYTHIA-Tune A predictions describe all of the data well. HERWIG predictions describe the measured jet shapes well for ${\rm p_{\rm T}^{\rm jet}}> 55$ GeV but produces jets that are too narrow at lower ${\rm p_{\rm T}^{\rm jet}}$. DIJET AZIMUTHAL DECORRELATIONS ============================== The D0 experiment has employed the dijet sample to study azimuthal decorrelations, $\Delta \phi_{\rm dijet}$, between the two leading jets [@dphi]. The normalized cross section, $$\frac{1}{\sigma_{\rm dijet}} \frac{d \sigma}{d \Delta \phi_{\rm dijet}},$$ is sensitive to the spectrum of the gluon radiation in the event. The measurements has been performed in different regions of the leading jet ${\rm p_{\rm T}^{\rm jet}}$ starting at ${\rm p_{\rm T}^{\rm jet}}> 75$ GeV, where the second jet is required to have at least ${\rm p_{\rm T}^{\rm jet}}> 40$ GeV. Figure \[dph1\] shows the measured cross section compared to LO and NLO predictions [@dphinlo]. The LO predictions, with at most three partons in the final state, is limited to $\Delta \phi_{\rm dijet} > 2 \pi/3$, for which the three partons define a [[Mercedes-star]{}]{} topology. It presents a prominent peak at $\Delta \phi_{\rm dijet} = \phi$ corresponding to the soft limit for which the third parton is collinear to the direction of the two leading partons. The NLO predictions, with four partons in the final state, describes the measured $\Delta \phi_{\rm dijet}$ distribution except at very high and very low values of $\Delta \phi_{\rm dijet}$ where additional soft contributions, corresponding to a resummed calculation, are necessary. A reasonable approximation to such calculations is provided by parton shower Monte Carlo programs. ![ Measured azimuthal decorrelations in dijet production for central jets compared to pQCD predictions in different regions of ${\rm p_{\rm T}^{\rm jet}}$ of the leading jet.[]{data-label="dph1"}](dph1.eps){width="6"} Figure \[dph2\] present the measured cross section compared to PYTHIA and HERWIG predictions in different regions of ${\rm p_{\rm T}^{\rm jet}}$. PYTHIA with default parameters underestimates the gluon radiation at large angles. Different tunes of PYTHIA predictions are possible, which include an enhanced contribution from initial-state soft gluon radiation, to properly describe the azimuthal distribution. HERWIG also describes the data although tends to produce less radiation than PYTHIA close to the direction of the leading jets. This measurements clearly show that angular correlations between jets can be employed to tune Monte Carlo predictions of soft gluon radiation in the final state. ![ Measured azimuthal decorrelations in dijet production for central jets compared to PYTHIA and HERWIG predictions in different regions of leading ${\rm p_{\rm T}^{\rm jet}}$. The band covers PYTHIA predictions with different amount of initial-state soft-gluon radiation.[]{data-label="dph2"}](dph3.eps){width="6"} DIRECT PHOTON PRODUCTION ======================== The measurement of the inclusive photon production as a function of photon transverse momentum, ${\rm p_{\rm T}^{\gamma}}$, at the Tevatron constitutes a precise test of pQCD predictions since the energy scale of the electromagnetic calorimeters is well understood by the experiments. The measured cross section is partially dominated by contributions from quark-gluon scattering and therefore provides a powerful constrain of the gluon PDF at high-x. However, it is a rather difficult measurement where a good understanding of QCD backgrounds from $\pi^0$ and $\eta$ decays into photons is necessary. The D0 collaboration has presented results [@photon] on inclusive photon production, based on 326 $\rm pb^{-1}$ of Run II data, in the region $\|\eta^{\gamma}\| < 0.9$ and ${\rm p_{\rm T}^{\gamma}}< 300$ GeV/c (see Figure \[gamma1\]). The measurements are compared to NLO pQCD predictions [@gnlo]. ![Measured inclusive prompt photon cross section as a function of ${\rm p_{\rm T}^{\gamma}}$ (dots) compared to pQCD NLO prediction (solid line).[]{data-label="gamma1"}](gamma1.eps){width="8"} Figure \[gamma2\] shows the ratio data/NLO as a function of ${\rm p_{\rm T}^{\gamma}}$. The measured cross section is well described by the theoretical prediction, where the latter presents uncertainties at the level of about $10 \%$. Future measurements based on few $\rm fb^{-1}$ of data will provide valuable information about the proton structure as well as imposes strong constrains to the presence of new physics with very energetic photons in the final state. ![Ratio data/theory as a function of ${\rm p_{\rm T}^{\gamma}}$. The error bars show the uncertainty on the measurements. The band denotes the PDF uncertainty on the theoretical prediction and the dashed lines indicate the uncertainty from the variation of the renormalization/factorization scales in the pQCD NLO calculation.[]{data-label="gamma2"}](gamma2.eps){width="8"} BOSON PLUS JETS PRODUCTION ========================== The study of the production of electroweak bosons in association with jets of hadrons in the final state constitutes a fundamental item in the high-${\rm p_{\rm T}}$ physics program at the Tevatron. These events are main backgrounds to many interesting physics processes like, for example, top production, and the search for the SM Higgs and supersymmetry. Therefore, during the last few years a significant effort is being made to develop and validate the necessary Monte Carlo tools to describe these complicated multijet final states. As a result, a number of leading-order Monte Carlo programs are available that describe boson+jet production processes up to large parton multiplicities [@vecbos; @alpgen; @madgraph], and NLO pQCD parton-level predictions are also available for a limited number of processes (up to boson+2jets production) [@mcfm]. The interface between parton level calculations and parton showers, necessary to describe the complexity of the observed hadronic final states, requires the introduction of different prescriptions to resolve resulting double counts across processes with different parton multiplicities. The theoretical prescriptions employed require validation using data. For this purpose, both CDF and D0 collaborations have performed a careful set of measurements on Boson+jets production. Figure \[cdfW1\] shows the measured ${\rm p_{\rm T}^{\rm jet}}$ distribution for the ${\rm n}^{th}$ jet in inclusive $\rm W+n_{\rm jet}$ production by CDF, based on 320 $\rm pb^{-1}$ of Run II data. This observable is particularly sensitive to the details on the implementation of the parton shower interface in the Monte Carlo models. The measurements are compared to leading-order Monte Carlo [@alpgen] predictions interfaced with PYTHIA parton shower, and normalized to the data. Similarly, Figure \[cdfW2\] shows the measured distance, ($\eta - \phi$) space, between the two leading jets in inclusive $\rm W + 2 \rm jets$ production. The Monte Carlo model provides a reasonable description of the shape of the measured distributions. ![Measured inclusive cross section as a function of ${\rm E_{\rm T}^{\rm jet}}$ for the n${}^{th}$ jet in inclusive $\rm W+n_{\rm jet}$ production. The measurements are compared to different leading-order Monte Carlo predictions normalized to the data.[]{data-label="cdfW1"}](cdfW1.eps){width="8"} ![Measured cross section as a function of $\Delta R_{\rm jet1-jet2}$ in inclusive $\rm W+ \geq 2$ jets production compared to leading-order Monte Carlo predictions. The Monte Carlo distribution is normalized to the data. []{data-label="cdfW2"}](cdfW2.eps){width="8.0"} The D0 collaboration has performed a detailed measurement on inclusive $\rm Z +n_{\rm jet}$ production [@d0Zs]. Figure \[d0Z1\] presents the measured cross section, normalized to the total Drell-Yan cross section, as a function of the inclusive jet multiplicity. The measurements are compared to PYTHIA, Matrix Elements+Parton Shower (ME+PS) [@madgraph], and parton-level NLO calculations [@mcfm]. As expected, PYTHIA, that only includes matrix elements for two-to-two processes, only provides a reasonable description for one and two inclusive jet production[^3], while underestimates the production of events with large jet multiplicities. The prediction from ME+PS describes the observed normalized yields as well as the shape of the $\rm n^{\rm th}$ jet ${\rm p_{\rm T}^{\rm jet}}$ distributions (see Figure \[d0Z2\]). Finally, pQCD NLO predictions for one and two inclusive jet production provide a reasonable description of the data. ![Measured inclusive cross section as a function of jet multiplicity for $Z+N_{\rm jet}$ production. The measurements are normalized to the total Drell-Yan production cross section. The data is compared to different leading-order hadron level Monte Carlo and NLO pQCD parton level predictions. ME+PS predictions are normalized to the measured inclusive $Z+1{\rm jet}$ production yield.[]{data-label="d0Z1"}](d0Z1.eps){width="8"} ![Measured ${\rm E_{\rm T}^{\rm jet}}$ distribution for the n${}^{th}$ jet in $\rm W+N_{\rm jet}$ production. The measurements are compared to different leading-order Monte Carlo predictions.[]{data-label="d0Z2"}](d0Z2.eps){width="8"} The CDF collaboration has presented first measurements of jet shapes and energy flows in inclusive Z+jet production, which are necessary to validate the modeling for underlying event and soft gluon radiation implemented in the Monte Carlo generators. Figure \[cdfZ1\] shows the measured momentum flow (projected to the transverse plane) as a function of the azimuthal distance with respect to the Z direction ($\phi = 0$), where only the central region of the calorimeter ($\|\eta\|<0.7$) is considered. In the region $\|\Delta \phi\| = \phi$ the distribution shows the presence of the leading jet. The region $\|\Delta \phi\| \sim 1$ is particularly sensitive to a proper description of the underlying event. PYTHIA-Tune A provides a good description of the measured transverse momentum flow while ALPGEN+HERWIG would require additional underlying event activity. ![Measured mometum flow (transverse plane) and a function of $\Delta \phi$ in inclusive Z+jet production, where the Z direction defined $\phi = 0$. The measurements are compared with PYTHIA and ALPGEN+HERWIG.[]{data-label="cdfZ1"}](cdfZ1.eps){width="7"} Similar conclusions can be drawn from Figure \[cdfZ2\], where the measured jet shape in inclusive Z+jet production is presented. PYTHIA Tune A provides the best description of the data, and the difference between models can be attributed to differences on the underlying event implementation. ![Measured integrated jet shape in inclusive Z+jet production for jets in the region ${\rm p_{\rm T}^{\rm jet}}> 25$ GeV/c and $\|{\rm y^{\rm jet}}\|<2.1$, compared to different leading-order Monte Carlo predictions.[]{data-label="cdfZ2"}](cdfZ2.eps){width="7"} FINAL NOTES =========== The Tevatron is delivering luminosity according to expectations and each experiment plans to collect more than $\rm 4 \ fb^{-1}$ of data. During the next few year the jet physics program at the Tevatron will translate into a number of precise measurements that will test the SM and provide very valuable information on how to model QCD processes at the LHC. I would like to thank the organizers for their kind invitation to the conference. [9]{} B. Abbott [et al.]{} (DØ Collaboration), Phys. Rev. Let t. [82]{}, 2451 (1999).\ T. Affolder [et al.]{} (CDF Collaboration), Phys. Rev. D [f 64]{}, 032001 (2001). \[Erratum-ibid. D [65]{}, 039903 (2002)\]. A. Abulencia [et al.]{} (CDF Collaboration), Phys. Rev. Lett. [96]{}, 122001 (2006). A. Abulencia [et al.]{}, hep-ex/0512020. S. Catani [et al.]{}, Nucl. Phys. B [406]{}, 187 (1993). S.D. Ellis and D.E. Soper, Phys. Rev. D [48]{}, 3160 (1993). G. C. Blazey, et al., hep-ex/0005012 (2000).\ S.D. Ellis, et al., hep-ph/0111434 (2001). W.T. Giele, E.W.N. Glover and David A. Kosower, Nucl. Phys. B [403]{}, 633 (1993). V.M. Abazov [et al.]{} (DO Collaboration), Phys. Lett. B [525]{}, 211 (2002). see http://www-cdf.fnal.gov see http://www-d0.fnal.gov N. Kidonakis and J.F. Owens, Phys. Rev. D 63, 054019 (2001), hep-ph/0007268 T. Sjöstrand [et al.]{}, Comp. Phys. Comm. [135]{}, 238 (2001). G. Corcella [et al.]{}, JHEP [0101]{}, 010 (2001). Monte Carlo samples are generated using the following tuned parameters in [pyth ia]{}: PARP(67) = 4.0, MSTP(82) = 4, PARP(82) = 2.0, PARP(84) = 0.4, PARP(85) = 0.9, PARP(86) = 0.95, PARP(89) = 1800.0, PAR P(90) = 0.25. T. Affolder [et al.]{} (CDF Collaboration), Phys. Rev. D [65]{}, 092002 (2002). D. Acosta [et al.]{} (CDF Collaboration), Phys. Rev. D [71]{}, 112002 (2005). V.M. Abazov [et al.]{} (D0 Collaboration), Phys. Rev. Lett. [94]{}, 221801 (2005). Z. Nagy, Phys. Rev. Lett. [88]{}, 122003 (2002). V.M. Abazov [et al.]{} (D0 Collaboration), Phys. Lett. B [639]{} 151 (2006). T. Binoth et al., Eur. Phys. J. C[16]{}, 311 (2000).\ S. Catani et al., JHEP [05]{}, 028 (2002). F.A. Berends [et al.]{}, Nucl Phys B [357]{} 32 (1991). M.L. Mangano [et al.]{}, JHEP [0307]{} 001 (2003). F. Maltoni and T. Stelzer, JHEP [0302]{} 027 (2003). J. Campbell, R.K. Ellis and D. Rainwater, Phys. Rev. D [68]{} 094021 (2003). V.M. Abazov [et al.]{} (D0 Collaboration), FERMILAB-PUB-06-283-E. [^1]: Previous measurements using the ${\rm k_{\rm T}}$ algorithm at the Tevatron [@d0kt] observed a marginal agreement with NLO pQCD at low ${\rm p_{\rm T}^{\rm jet}}$ but this discrepancy is removed after non-perturbative corrections are included. [^2]: A $75 \%$ merging fraction has been used instead of the default $50 \%$. [^3]: The hardness of the first gluon radiation in a parton showers approximately follows that of matrix elements.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Given a set of $n$ unit squares in the plane, the goal is to rank them in space in such a way that only few squares see each other vertically. We prove that ranking the squares according to the lexicographic order of their centers results in at most $3n-7$ pairwise visibilities for $n\geq 4$. We also show that this bound is best possible, by exhibiting a set of $n$ squares with at least $3n-7$ pairwise visibilities under any ranking.' author: - 'Bernd Gärtner[^1]' bibliography: - 'biblio.bib' date: 'August 4, 2008' title: Ranking Unit Squares with Few Visibilities --- Problem statement ================= The unit square with center $c\in{\mathds{R}}^2$ is the set $$S(c) = \{p\in{\mathds{R}}^2\mid \\|p-c\\|_{\infty}\leq 1\}.$$ Given a set ${\cal S}$ of $n$ unit squares (simply called squares in the sequel), a ranking of ${\cal S}$ is a sequence $\rho=(S_1,S_2,\ldots,S_n)$ such that ${\cal S}=\{S_1,S_2,\ldots,S_n\}$. For $i<k$, $S_i$ sees $S_k$ under $\rho$ if there exists a point $p\in{\mathds{R}}^2$ such that $$\begin{array}{rcll} p&\in& S_i\cap S_k,\\ p&\notin& S_j, & i<j<k. \end{array}$$ The graph on ${\cal S}$ formed by all pairs $\{S_i,S_k\}$ such that $S_i$ sees $S_k$ under $\rho$ is called the visibility graph of $\rho$ and will be denoted by $G(\rho)$. The goal is to find a ranking $\rho$ such that $G(\rho)$ has as few edges as possible. We do not know how to find the best ranking for a given set ${\cal S}$, but we prove that there is always a ranking $\rho$ under which $G(\rho)$ has no more than $3n-7$ edges. For some sets ${\cal S}$, this is the best bound that can be achieved. This research is motivated by similar questions for intervals in ${\mathds{R}}^1$ [@intervals]. Main result =========== Given a ranking $\rho=(S_1,S_2,\ldots,S_n)$, the center of square $S_i$ will be denoted by $c_i=(x_i,y_i), i=1,2,\ldots,n$. We will repeatedly use the following simple fact. \[obs:help\] Let $\rho=(S_1,S_2,\ldots,S_n)$ be a ranking, and suppose that there are centers $c_i,c_j,c_k$ with $i<j<k$, such that $c_j$ is contained in the (axis-parallel) rectangle spanned by $c_i$ and $c_k$. Then $\{S_i,S_k\}$ is not an edge of $G(\rho)$. [[Proof.]{} ]{} Center $c_j$ being contained in the rectangle spanned by $c_i$ and $c_k$ is easily seen to be equivalent to $S_j\supseteq S_i\cap S_k$. It follows that $S_i$ does not see $S_k$ under $\rho$. The size (number of edges) of the visibility graph may be $\Theta(n^2)$ for a “bad” ranking, see Figure \[fig:n2\] (left). The right part of the figure depicts the lexicographic ranking. According to the next lemma, this ranking always results in $O(n)$ visibilities. ![Upward view on a ranked set of squares. Left: All squares in the lower half see all squares in the upper half. Right: the lexicographic ranking incurs only linearly many visibilities.\[fig:n2\]](n2.eps){width="12cm"} \[lem:main\] Let $\rho_{{\rm lex}}=(S_1,S_2,\ldots,S_n)$ be the lexicographic ranking, meaning that $i<j$ if and only if the center of $S_i$ is lexicographically smaller than the center of $S_j$. Then $G(\rho_{{\rm lex}})$ is a planar graph. More precisely, the straight-line embedding of $G(\rho_{{\rm lex}})$ obtained by mapping each square $S_i$ to its center $c_i\in{\mathds{R}}^2$ is a plane graph. [[Proof.]{} ]{}We show that if two segments $\overline{c_ic_j}$ and $\overline{c_kc_{\ell}}$ cross, then $G(\rho_{{\rm lex}})$ contains at most one of the edges $\{S_i,S_j\}$ and $\{S_k,S_{\ell}\}$. Since deletion of other squares can only add visibilities between the four squares involved in the crossing, we may w.l.o.g.assume that ${\cal S}=\{S_1,S_2,S_3,S_4\}$. By lexicographic order of the centers, there are only two cases. Case (a): Segments $\overline{c_1c_3}$ and $\overline{c_2c_4}$ cross. If $y_2$ is between $y_1$ and $y_3$, we get that $c_2$ is in the rectangle spanned by $c_1$ and $c_3$, so $\{S_1,S_3\}$ is not an edge of $G(\rho_{{\rm lex}})$ by Observation \[obs:help\]. If $y_2$ is not between $y_1$ and $y_3$, we either have $y_2<\min(y_1,y_3)$ and thus $y_4>y_3$ (otherwise, there would be no crossing), or $y_2>\max(y_1,y_3)$ and thus $y_4 < y_3$. In both cases, $y_3$ is between $y_2$ and $y_4$ which means that $c_3$ is in the rectangle spanned by $c_2$ and $c_4$. This in turn shows that $\{S_2,S_4\}$ is not an edge of $G(\rho_{{\rm lex}})$. Case (b): Segments $\overline{c_1c_4}$ and $\overline{c_2c_3}$ cross. An easy case occurs if one of $c_2$ and $c_3$ is in the rectangle spanned by $c_1$ and $c_4$, since Observation \[obs:help\] then implies that $\{S_1,S_4\}$ is not an edge. Otherwise, we have $$\label{eq:visa} y_2<\min(y_1,y_4)\leq y_4$$ and thus $$\label{eq:visb} y_3>\max(y_1,y_4)\geq y_1$$ (because of the crossing), or $y_3<\min(y_1,y_4)$ and $y_2>\max(y_1,y_4)$. Let us only treat the first case; the second one is symmetric under exchange of indices $2$ and $3$. We now show that every point in $S_1\cap S_4$ is in $S_2\cup S_3$, given that $S_2\cap S_3\neq\emptyset$. Therefore, if $\{S_2,S_3\}$ is an edge of $G(\rho_{{\rm lex}})$, then $\{S_1,S_4\}$ is not an edge. Let $p=(x,y)$ be any point in $S_1\cap S_4$. From $x_4-1\leq x \leq x_1+1$ and lexicographic order it follows that $$x_2-1, x_3-1 \leq x \leq x_2+1, x_3+1.\label{eq:vis1}$$ Using $y_4-1 \leq y \leq y_1+1$ together with (\[eq:visa\]) and (\[eq:visb\]), we also conclude that $$y_2-1 \leq y \leq y_3+1.\label{eq:vis2}$$ If $y$ in addition satisfies $y\leq y_2+1$, (\[eq:vis1\]) and (\[eq:vis2\]) imply that $p\in S_2$. But if $y>y_2+1$, we can use the assumption $S_2\cap S_3\neq\emptyset$ to conclude that $y_3-y_2\leq 2$ and hence $y>y_3-1$. With (\[eq:vis1\]) and (\[eq:vis2\]), we then get $p\in S_3$. Now we are ready to prove our main theorem. \[thm:main\] If $n\geq 4$, $G(\rho_{{\rm lex}})$ has at most $3n-7$ edges. [[Proof.]{} ]{}An upper bound of $3n-6$ already follows from Lemma \[lem:main\]. In order for this bound to be tight, the outer face of $G(\rho_{{\rm lex}})$’s straight-line embedding would have to be a triangle $\Delta$ spanned by three centers $c_i, c_j,c_k$, $i<j<k$, and with all other centers inside $\Delta$. Indeed, for $n=3$ this is possible, but for $n\geq 4$, we get a contradiction: Let $c=(x,y)$ be any center distinct from $c_i,c_j,c_k$. From $c\in\Delta$, it easily follows that $c_i< c <c_k$ (comparison being lexicographically). If $y$ is between $y_i$ and $y_k$, $c$ is in the rectangle spanned by $c_i$ and $c_k$, so $\{S_i,S_k\}$ can’t be an edge of $G(\rho_{{\rm lex}})$ by Observation \[obs:help\]. If $y$ is not between $y_i$ and $y_k$, then $y$ must be between $y_i$ and $y_j$, and between $y_j$ and $y_k$. Depending on whether $c_i < c < c_j$ or $c_j < c < c_k$, we get that either $\{S_i,S_j\}$ or $\{S_j,S_k\}$ is not an edge of $G(\rho_{{\rm lex}})$. Lower Bound =========== The bound of $3n-7$ derived in Theorem \[thm:main\] is best possible in the worst case, not only under the lexicographic ranking, but under every ranking. For a proof by picture see Figure \[fig:best\]. ![A set of $n$ unit squares such that the visibility graph $G(\rho)$ has at least $3n-7$ edges for every ranking $\rho$: a square from the middle bunch of $n-2$ squares always sees the next square above it in the bunch (these are $n-3$ visibilities), and it sees (or is seen by) both of the two special squares ($2n-4$ visibilities).\[fig:best\]](best.eps){width="5cm"} Acknowledgment {#acknowledgment .unnumbered} ============== The problem of ranking unit squares was given to me as an undergraduate student by Emo Welzl in early 1989. After some initial programming tasks, this was the first actual research assignment during my time as “Forschungstutor” (research assistant) with Emo. My initial manuscript from 1989, typeset in bumpy LaTeX and containing manually drawn figures, is lost. But throughout the almost twenty years that have passed since then, I could never forget this first (simple) result of mine. While writing it up now, many memories of all these years have come back to me with surprising strength. I thank Emo for bringing the problem to my attention, and for so much more. [^1]: Institute of Theoretical Computer Science, ETH Z[ü]{}rich, CAB G32.2, CH-8092 Z[ü]{}rich, Switzerland. Email: [email protected]	{ "pile_set_name": "ArXiv" }
--- abstract: '= 11pt = 0.65in = 0.65in' author: - 'Sean N. Raymond' - 'Eiichiro Kokubo' - 'Alessandro Morbidelli' - 'Ryuji Morishima' - 'Kevin J. Walsh' title: - 'TERRESTRIAL PLANET FORMATION AT HOME AND ABROAD' - Key inner Solar System Constraints --- INTRODUCTION ================ The term “terrestrial planet” evokes landscapes of a rocky planet like Earth or Mars but given recent discoveries it has become somewhat ambiguous. Does a $5 \mearth$ Super Earth count as a terrestrial planet? What about the Mars-sized moon of a giant planet? These objects are terrestrial planet-sized but their compositions and corresponding landscapes probably differ significantly from our terrestrial planets’. In addition, while Earth is thought to have formed via successive collisions of planetesimals and planetary embryos, the other objects may have formed via different mechanisms. For instance, under some conditions a $10 \mearth$ or larger body can form by accreting only planetesimals, or even only cm-sized pebbles. In the context of the classical stages of accretion this might be considered a “giant embryo” rather than a planet (see §7.1). What criteria should be used to classify a planet as terrestrial? A bulk density higher than a few $g\,cm^{-3}$ probably indicates a rock-dominated planet, but densities of low-mass exoplanets are extremely challenging to pin down [see @marcy13]. A planet with a bulk density of $0.5-2 g\,cm^{-3}$ could either be rocky with a small H-rich envelope or an ocean planet [@fortney07; @valencia07; @adams08]. Bulk densities larger than $3 g\,cm^{-3}$ have been measured for planets as massive as $10-20 \mearth$, although higher-density planets are generally smaller [@weiss13]. Planets with radii $R \lesssim 1.5-2 \rearth$ or masses $M \lesssim 5-10 \mearth$ are likely to preferentially have densities of $3 g\,cm^{-3}$ or larger and thus be rocky [@weiss13b; @lopez13]. In this review we address the formation of planets in orbit around stars that are between roughly a lunar mass ($\sim 0.01 \mearth$) and ten Earth masses. Although the compositions of planets in this mass range certainly vary substantially, these planets are capable of having solid surfaces, whether they are covered by thick atmospheres or not. These planets are also below the expected threshold for giant planet cores [e.g. @lissauer07]. We refer to these as terrestrial planets. We start our discussion of terrestrial planet formation when planetesimals have already formed; for a discussion of planetesimal formation please see the chapter by Johansen et al. Our understanding of terrestrial planet formation has undergone a dramatic improvement in recent years. This was driven mainly by two factors: increased computational power and observations of extra-solar planets. Computing power is the currency of numerical simulations, which continually increase in resolution and have become more and more complex and realistic. At the same time, dramatic advances in exoplanetary science have encouraged many talented young scientists to join the ranks of the planet formation community. This manpower and computing power provided a timely kick in the proverbial butt. Despite the encouraging prognosis, planet formation models lag behind observations. Half of all Sun-like stars are orbited by close-in “super Earths”, yet we do not know how they form. There exist ideas as to why Mercury is so much smaller than Earth and Venus but they remain speculative and narrow. Only recently was a cohesive theory presented to explain why Mars is smaller than Earth, and more work is needed to confirm or refute it. We first present the observational constraints in the Solar System and extra-solar planetary systems in §2. Next, we review the dynamics of accretion of planetary embryos from planetesimals in §3, and of terrestrial planets from embryos in §4, including a discussion of the importance of a range of parameters. In §5 we apply accretion models to extra-solar planets and in §6 to the Solar System. We discuss different modes of accretion and current limitations in §7 and summarize in §8. OBSERVATIONAL CONSTRAINTS ============================= Given the explosion of new discoveries in extra-solar planets and our detailed knowledge of the Solar System, there are ample observations with which to constrain accretion models. Given the relatively low resolution of numerical simulations, accretion models generally attempt to reproduce large-scale constraints such as planetary mass-and orbital distributions rather than smaller-scale ones like the exact characteristics of each planet. We now summarize the key constraints for the Solar System and exoplanets. 2.1 The Solar System [The masses and orbits of the terrestrial planets.]{} There exist metrics to quantify different aspects of a planetary system and to compare it with the Solar System. The angular momentum deficit $AMD$ [@laskar97] measures the difference in orbital angular momentum between the planets’ orbits and the same planets on circular, coplanar orbits. The $AMD$ is generally used in its normalized form: $$AMD = \frac{\sum_{j} m_j \sqrt{a_j} \left(1 - cos(i_j) \sqrt{1-e_j^2}\right)} {\sum_j m_j \sqrt{a_j}},$$ where $a_j$, $e_j$, $i_j$, and $m_j$ are planet $j$’s semimajor axis, eccentricity, inclination and mass. The $AMD$ of the Solar System’s terrestrial planets is 0.0018. The radial mass concentration $RMC$ [defined as $S_c$ by @chambers01] measures the degree to which a system’s mass is concentrated in a small region: $$RMC = max \left(\frac{\sum m_j}{\sum m_j [log_{10}(a/a_j)]^2} \right).$$ Here, the function in brackets is calculated for $a$ across the planetary system, and the $RMC$ is the maximum of that function. For a single-planet system the $RMC$ is infinite. The $RMC$ is higher for systems in which the total mass is packed in smaller and smaller radial zones. The $RMC$ is thus smaller for a system with equal-mass planets than a system in which a subset of planets dominate the mass. The $RMC$ of the Solar System’s terrestrial planets is 89.9. [The geochemically-determined accretion histories of Earth and Mars.]{} Radiogenic elements with half-lives of a few to 100 Myr can offer concrete constraints on the accretion of the terrestrial planets. Of particular interest is the $^{182}$Hf-$^{182}$W system, which has a half life of 9 Myr. Hf is lithophile (“rock-loving") and W is siderophile (“iron-loving”). The amount of W in a planet’s mantle relative to Hf depends on the timing of core formation [@nimmo06]. Early core formation (also called “core closure”) would strand still-active Hf and later its product W in the mantle, while late core formation would cause all W to be sequestered in the core and leave behind a W-poor mantle. Studies of the Hf-W system have concluded that the last core formation event on Earth happened roughly 30-100 Myr after the start of planet formation [@kleine02; @yin02; @kleine09; @konig11]. Similar studies on martian meteorites show that Mars’ accretion finished far earlier, within 5 Myr [@nimmo07; @dauphas11]. The highly-siderophile element (HSE) contents of the terrestrial planets’ mantles also provide constraints on the total amount of mass accreted by a planet after core closure [@drake02]. This phase of accretion is called the [late veneer]{} [@kimura74]. Several unsolved problems exist regarding the late veneer, notably the very high Earth/Moon HSE abundance ratio [@day07; @walker09], which has been proposed to be the result of either a top-heavy [@bottke10; @raymond13] or bottom-heavy [@schlichting12] distribution of planetesimal masses. [The large-scale structure of the asteroid belt.]{} Reproducing the asteroid belt is not the main objective of formation models. But any successful accretion model must be consistent with the asteroid belt’s observed structure, and that structure can offer valuable information about planet formation. Populations of small bodies can be thought of as the “blood spatter on the wall” that helps detectives solve the “crime”, figuratively speaking of course. The asteroid belt’s total mass is just $5 \times10^{-4} \mearth$, about four percent of a lunar mass. This is 3-4 orders of magnitude smaller than the mass contained within the belt for any disk surface density profile with a smooth radial slope. In addition, the inner belt is dominated by more volatile-poor bodies such as E-types and S-types whereas the outer belt contains more volatile-rich bodies such as C-types and D-types [@gradie82; @demeo13]. There are no large gaps in the distribution of asteroids – apart from the Kirkwood gaps associated with strong resonances with Jupiter – and this indicates that no large ($\gtrsim 0.05 \mearth$) embryos were stranded in the belt after accretion, even if the embryos could have been removed during the late heavy bombardment [@raymond09c]. [The existence and abundance of volatile species – especially water – on Earth.]{} Although it contains just 0.05-0.1% water by mass [@lecuyer98; @marty12], Earth is the wettest terrestrial planet. It is as wet as ordinary chondrite meteorites, thought to represent the S-type asteroids that dominate the inner main belt, and wetter than enstatite chondrites that represent E-types interior to the main belt [see, for example, figure 5 from @morby12b]. We think that this means that the rocky building blocks in the inner Solar System were dry. In addition, heating mechanisms such as collisional heating and radiogenic heating from $^{26}$Al may have dehydrated fast-forming planetesimals [e.g. @grimm93]. The source of Earth’s water therefore requires an explanation. The isotopic composition of Earth’s water constrains its origins. The D/H ratio of water on Earth is a good match to carbonaceous chondrite meteorites thought to originate in the outer asteroid belt [@marty06]. The D/H of most observed comets is $2\times$ higher – although one comet was recently measured to have the same D/H as Earth [@hartogh11] – and that of the Sun (and presumably the gaseous component of the protoplanetary disk) is $6\times$ smaller [@geiss98]. It is interesting to note that, while the D/H of Earth’s water can be matched with a weighted mixture of material with Solar and cometary D/H, that same combination does not match the $^{15}$N/$^{14}$N isotopic ratio [@marty06]. Carbonaceous chondrites, on the other hand, match both measured ratios. ----------------------------------- ------------------------------------- Angular momentum deficit $AMD$ 0.0018 Radial Mass Concentration $RMC$ 89.9 Mars’ accretion timescale$^1$ 3-5 Myr Earth’s accretion timescale$^2$ $\sim 50$ Myr Earth’s late veneer$^3$ $(2.5-10) \times 10^{-3} \mearth$ Total mass in asteroid belt $5 \times 10^{-4} \mearth$ Earth’s water content by mass$^4$ $5\times 10^{-4} - 3\times 10^{-3}$ ----------------------------------- ------------------------------------- : $^1$[@dauphas11]. $^2$[@kleine09; @konig11]. $^3$[@day07; @walker09], see also [@bottke10; @schlichting12; @raymond13]. $^4$[@lecuyer98; @marty12] The bulk compositions of the planets are another constraint. For example, the core/mantle (iron/silicate) mass ratideo of the terrestrial planets ranges from 0.4 (Mars) to 2.1 (Mercury). The bulk compositions of the terrestrial planets depend on several factors in addition to orbital dynamics and accretion: the initial compositional gradients of embryos and planetesimals, evolving condensation fronts, and the compositional evolution of bodies due to collisions and evaporation. Current models for the bulk composition of terrestrial planets piggyback on dynamical simulations such as the ones discussed in sections 4-6 below [e.g. @bond10; @carterbond12; @elser12]. These represent a promising avenue for future work. 2.2 Extrasolar Planetary Systems [The abundance and large-scale characteristics of “hot Super Earths"]{}. These are the terrestrial exoplanets whose origin we want to understand. Radial velocity and transit surveys have shown that roughly 30-50% of main sequence stars host at least one planet with $M_p \lesssim 10 \mearth$ with orbital period $P \lesssim 85-100$ days [@mayor11; @howard10; @howard12; @fressin13]. Hot super Earths are preferentially found in multiple systems [e.g. @udry07; @lissauer11]. These systems are in compact orbital configurations that are similar to the Solar System’s terrestrial planets’ as measured by the orbital period ratios of adjacent planets. The orbital spacing of adjacent Kepler planet candidates is also consistent with that of the Solar System’s planets when measured in mutual Hill radii [@fang13]. Figure \[fig:kepler\_ss\] shows eight systems each containing 4-5 presumably terrestrial exoplanets discovered by the Kepler mission. The largest planet in each system is less than 1.5 Earth radii, and in one system the largest planet is actually smaller than Earth (KOI-2169). The Solar System is included for scale, with the orbit of each terrestrial planet shrunk by a factor of ten (but with their actual masses). Given that the x axis is on a log scale, the spacing between planets is representative of the ratio between their orbital periods (for scale, the Earth/Venus period ratio is about 1.6). Given the uncertainties in the orbits of extra-solar planets and observational biases that hamper the detection of low-mass, long-period planets we do not generally apply the $AMD$ and $RMC$ metrics to these systems. Rather, the main constraints come from the systems’ orbital spacing, masses and mass ratios. [The existence of giant planets on exotic orbits.]{} Simulations have shown in planetary systems with giant planets the giants play a key role in shaping the accretion of terrestrial planets [e.g., @chambers02; @levison03; @raymond04]. Giant exoplanets have been discovered on diverse orbits that indicate rich dynamical histories. Gas giants exist on orbits with eccentricities as high as 0.9. It is thought that these planets formed in systems with multiple gas giants that underwent strong dynamical instabilities that ejected one or more planets and left behind surviving planets on eccentric orbits [@chatterjee08; @juric08; @raymond10]. Hot Jupiters – gas giants very close to their host stars – are thought to have either undergone extensive inward gas-driven migration [@lin96] or been re-circularized by star-planet tidal interactions from very eccentric orbits produced by planet-planet scattering [@nagasawa08; @beauge12] or other mechanisms [e.g. @fabrycky07; @naoz11 see chapter by Davies et al]. There also exist gas giants on nearly-circular Jupiter-like orbits [e.g. @wright08]. However, from the current discoveries systems of gas giants like the Solar System’s – with giant planets confined beyond 5 AU on low-eccentricity orbits – appear to be the exception rather than the rule. Of course, many planetary systems do not host currently-detected giant planets. Radial velocity surveys show that at least 14% of Sun-like stars have gas giants with orbits shorter than 1000 days [@mayor09], and projections to somewhat larger radii predict that $\sim20\%$ have gas giants within 10 AU [@cumming08]. Although they are limited by small number statistics, the statistics of high-magnification (planetary) microlensing events suggest that 50% or more of stars have gas giants on wide orbits [@gould10]. In addition, the statistics of short-duration microlensing events suggests that there exists a very abundant population of gas giants on orbits that are separated from their stars; these could either be gas giants on orbits larger than $\sim 10$ AU or free-floating planets [@sumi11]. [The planet-metallicity correlation.]{} Gas giants – at least those easily detectable with current techniques – are observed to be far more abundant around stars with high metallicities [@gonzalez97; @santos01; @laws03; @fischer05]. However, this correlation does not hold for low-mass planets, which appear to be able to form around stars with a wide range of metallicities [@ghezzi10; @buchhave12; @mann12]. It is interesting to note that there is no observed trend between stellar metallicity and the presence of debris disks [@greaves06; @moromartin07], although disks do appear to dissipate faster in low-metallicity environments [@yasui09]. The planet-metallicity correlation in itself does strongly constraint the planet formation models we discuss here. What is important is that the formation of systems of hot Super Earths does not appear to depend on the stellar metallicity, i.e. the solids-to-gas ratio in the disk. Additional constraints on the initial conditions of planet formation come from observations of protoplanetary disks around other stars [@williams11]. These observations measure the approximate masses and radial surface densities of planet-forming disks, mainly in their outer parts. They show that protoplanetary disks tend to have masses on the order of $10^{-3}$-$10^{-1}$ times the stellar mass [e.g. @scholz06; @andrews07a; @eisner08; @eisner12], with typical radial surface density slopes of $\Sigma \propto r^{-(0.5-1)}$ in their outer parts [@mundy00; @looney03; @andrews07b]. In addition, statistics of the disk fraction in clusters with different ages show that the gaseous component of disks dissipate within a few Myr [@haisch01; @hillenbrand08; @fedele10]. It is also interesting to note that disks appear to dissipate more slowly around low-mass stars than Solar-mass stars [@pascucci09]. FROM PLANETESIMALS TO PLANETARY EMBRYOS =========================================== In this section we summarize the dynamics of accretion of planetary embryos. We first present the standard model of runaway and oligarchic growth from planetesimals (§3.1). We next present a newer model based on the accretion of small pebbles (§3.2). 3.1 Runaway and Oligarchic Growth [Growth Modes]{} There are two growth modes: “orderly” and “runaway”. In orderly growth, all planetesimals grow at the same rate, so the mass ratios between planetesimals tend to unity. During runaway growth, on the other hand, larger planetesimals grow faster than smaller ones and mass ratios increase monotonically. Consider the evolution of the mass ratio between two planetesimals with masses $M_1$ and $M_2$, assuming $M_1>M_2$. The time derivative of the mass ratio is given by $$\frac{\mathrm{d}}{\mathrm{d}t}\left(\frac{M_1}{M_2}\right) = \frac{M_1}{M_2} \left( \frac{1}{M_1}\frac{\mathrm{d}M_1}{\mathrm{d}t} - \frac{1}{M_2}\frac{\mathrm{d}M_2}{\mathrm{d}t} \right).$$ It is the relative growth rate $(1/M)\mathrm{d}M/\mathrm{d}t$ that determines the growth mode. If the relative growth rate decreases with $M$, $\mathrm{d}(M_1/M_2)/\mathrm{d}t$ is negative then the mass ratio tends to be unity. This corresponds to orderly growth. If the relative growth rate increases with $M$, $\mathrm{d}(M_1/M_2)/\mathrm{d}t$ is positive and the mass ratio increases, leading to runaway growth. The growth rate of a planetesimal with mass $M$ and radius $R$ that is accreting field planetesimals with mass $m$ ($M>m$) can be written as $$\frac{\mathrm{d}M}{\mathrm{d}t} \simeq n_m\pi R^2 \left(1+\frac{v_{\rm esc}^2}{v_{\rm rel}^2}\right) v_{\rm rel}m, \label{eq:growth_rate}$$ where $n_m$ is the number density of field planetesimals, and $v_{\rm rel}$ and $v_{\rm esc}$ are the relative velocity between the test and the field planetesimals and the escape velocity from the surface of the test planetesimal, respectively [e.g., @kokubo96]. The term $v_{\rm esc}^2/v_{\rm rel}^2$ indicates the enhancement of collisional cross-section by gravitational focusing. [Runaway Growth of Planetesimals]{} The first dramatic stage of accretion through which a population of planetesimals passes is runaway growth [@greenberg78; @wetherill89; @kokubo96]. During planetesimal accretion gravitational focusing is efficient because the velocity dispersion of planetesimals is kept smaller than the escape velocity due to gas drag. In this case Eq.\[eq:growth\_rate\] reduces to $$\label{eq:rdmdt} \frac{\mathrm{d}M}{\mathrm{d}t} \propto \Sigma_\mathrm{dust} M^{4/3}v^{-2},$$ where $\Sigma_\mathrm{dust}$ and $v$ are the surface density and velocity dispersion of planetesimals and we used $n_m\propto \Sigma_\mathrm{dust} v^{-1}$, $v_{\rm esc} \propto M^{1/3}$, $R \propto M^{1/3}$, and $v_{\rm rel}\simeq v$. During the early stages of accretion, $\Sigma_\mathrm{dust}$ and $v$ barely depend on $M$, in other words, the reaction of growth on $\Sigma_\mathrm{dust}$ and $v$ can be neglected since the mass in small planetesimals dominate the system. In this case we have $$\frac{1}{M}\frac{\mathrm{d}M}{\mathrm{d}t} \propto M^{1/3},$$ which leads to runaway growth. During runaway growth, the eccentricities and inclinations of the largest bodies are kept small by dynamical friction from smaller bodies [@wetherill89; @ida92]. Dynamical friction is an equipartitioning of energy that maintains lower random velocities – and therefore lower-eccentricity and lower-inclination orbits – for the largest bodies. The mass distribution relaxes to a distribution that is well approximated by a power-law distribution. Among the large bodies that form in simulations of runaway growth, the mass follows a distribution $\mathrm{d}n_{\rm c}/\mathrm{d}m \propto m^{y}$, where $y \simeq -2.5$. This index can be derived analytically as a stationary distribution [@makino98]. The power index smaller than -2 is characteristic of runaway growth, as most of the system mass is contained in small bodies. We also note that runaway growth does not necessarily mean that the growth time decreases with mass, but rather that the mass ratio of any two bodies increases with time. [Oligarchic Growth of Planetary Embryos]{} During the late stages of runaway growth, embryos grow while interacting with one another. The dynamics of the system become dominated by a relatively small number – a few tens to a few hundred – oligarchs [@kokubo98; @kokubo00; @thommes03]. Oligarchic growth is the result of the self-limiting nature of runaway growth and orbital repulsion of planetary embryos. The formation of similar-sized planetary embryos is due to a slow-down of runaway growth [@lissauer87; @ida93; @ormel10b]. When the mass of a planetary embryo $M$ exceeds about 100 times that of the average planetesimal, the embryo increases the random velocity of neighboring planetesimals to be $v \propto M^{1/3}$ [but note that this depends on the planetesimal size; @ida93; @rafikov04; @chambers06]. The relative growth rate (from Eq.\[eq:rdmdt\]) becomes $$\frac{1}{M}\frac{\mathrm{d}M}{\mathrm{d}t} \propto \Sigma_\mathrm{dust} M^{-1/3}.$$ $\Sigma_\mathrm{dust}$ decreases through accretion of planetesimals by the embryo as $M$ increases [@lissauer87]. The relative growth rate is a decreasing function of $M$, which changes the growth mode to orderly. Neighboring embryos grow while maintaining similar masses. During this stage, the mass ratio of an embryo to its neighboring planetesimals increases because for the planetesimals with mass $m$, $(1/m)\mathrm{d}m/\mathrm{d}t \propto \Sigma_\mathrm{dust} m^{1/3}M^{-2/3}$, such that $$\frac{(1/M)\mathrm{d}M/\mathrm{d}t}{(1/m)\mathrm{d}m/\mathrm{d}t} \propto \left(\frac{M}{m}\right)^{1/3}.$$ The relative growth rate of the embryo is by a factor of $(M/m)^{1/3}$ larger than the planetesimals’. A bi-modal embryo-planetesimal system is formed. While the planetary embryos grow, a process called orbital repulsion keeps their orbital separations at roughly 10 mutual Hill radii $R_{H,m}$, where $R_{H,m} = 1/2 \left(a_1+a_2\right) \left[(M_1+M_2)/ (3 M_\star)\right]^{1/3}$; here subscripts 1 and 2 refer to adjacent embryos. Orbital repulsion is a coupling effect of gravitational scattering between planetary embryos that increases their orbital separation and eccentricities and dynamical friction from small planetesimals that decreases the eccentricities [@kokubo95]. Essentially, if two embryos come too close to each other their eccentricities are increased by gravitational perturbations. Dynamical friction from the planetesimals re-circularizes their orbits at a wider separation. ![Oligarchic growth of planetary embryos. Snapshots of the planetesimal system on the $a$-$e$ plane are shown for $t=0$, $10^5$, $2\times 10^5$, and $4\times 10^5$ years. The circles represent planetesimals with radii proportional to their true values. The initial planetesimal system consists of 10000 equal-mass ($m = 2.5 \times 10^{-4} \mearth$) bodies. In this simulation, a 6-fold increase in the planetesimal radius was used to accelerate accretion. In $4\times 10^5$ years, the number of bodies decreases to 333. From [@kokubo02]. []{data-label="fig:a-e-t_oligarchic"}](fig2.eps){width="45.00000%"} An example of oligarchic growth is shown in Figure \[fig:a-e-t\_oligarchic\] [@kokubo02]. About 10 embryos form with masses comparable to Mars’ ($M \approx 0.1 \mearth$) on nearly circular non-inclined orbits with characteristic orbital separations of $10 R_{H,m}$ . At large $a$ the planetary embryos are still growing at the end of the simulation. Although oligarchic growth describes the accretion of embryos from planetesimals, it implies giant collisions between embryos that happen relatively early and are followed by a phase of planetesimal accretion. Consider the last pairwise accretion of a system of oligarchs on their way to becoming planetary embryos. The oligarchs have masses $M_{olig}$ and are spaced by $N$ mutual Hill radii $R_{H,m}$, where $N \approx 10$ is the rough stability limit for such a system. The final system of embryos will likewise be separated by $N \, R_{H,m}$, but with larger masses $M_{emb}$. The embryos grow by accreting material within an annulus defined by the inter-embryo separation. Assuming pairwise collisions between equal-mass oligarchs to form a system of equal-mass embryos, the following simple relation should hold: $N R_{H,m} (M) = 2 \,N \,R_{H,m}(M_{emb})$. Given that $R_{H,m}(M) \sim (2M)^{1/3}$, this implies that $M_{emb} = 8 M_{olig}$. After the collision between a pair of oligarchs, each embryo must therefore accrete the remaining three quarters of its mass from planetesimals. We can estimate the dynamical properties of a system of embryos formed by oligarchic growth. We introduce a protoplanetary disk with surface density of dust and gas $\Sigma_\mathrm{dust}$ and $\Sigma_\mathrm{gas}$ defined as: $$\begin{aligned} \Sigma_\mathrm{dust} = f_\mathrm{ice}\Sigma_1{\left(\frac{a}{1\,{\rm AU}}\right)^{-x}} \ {\rm gcm}^{-2} \nonumber \\ \Sigma_\mathrm{gas} = f_\mathrm{gas}\Sigma_1{\left(\frac{a}{1\,{\rm AU}}\right)^{-x}} \ {\rm gcm}^{-2},\end{aligned}$$ where $\Sigma_1$ is simply a reference surface density in solids at 1 AU and $x$ is the radial exponent. $f_\mathrm{ice}$ and $f_\mathrm{gas}$ are factors that enhance the surface density of ice and gas with respect to dust. In practice $f_\mathrm{ice}$ is generally taken to be 2-4 [see @kokubo02; @lodders03] and $f_\mathrm{gas}\approx 100$. Given an orbital separation $b$ of embryos, the isolation (final) mass of a planetary embryo at orbital radius $a$ is estimated as [@kokubo02]: $$\begin{aligned} \label{eq:m_iso} M_\mathrm{iso} & \simeq 2\pi a b \Sigma_\mathrm{dust} = 0.16 {\left(\frac{b}{10 r_\mathrm{H}}\right)^{3/2}} {\left(\frac{f_\mathrm{ice}\Sigma_1}{10}\right)^{3/2}}\nonumber \\ & {\left(\frac{a}{1\,{\rm AU}}\right)^{(3/2)(2-x)}} {\left(\frac{M_\star}{M_\odot}\right)^{-1/2}} M_\oplus, \end{aligned}$$ where $M_{\star}$ is the stellar mass. The time evolution of an oligarchic body is [@thommes03; @chambers06]: $$M(t)= M_{\rm iso}\tanh^3\left(\frac{t}{\tau_{\rm grow}} \right). \label{eq:me2}$$ The growth timescale $\tau_{\rm grow}$ is estimated as $$\begin{aligned} \tau_{\rm grow} &=& 1.1 \times 10^{6} f_{\rm ice}^{-1/2} \left(\frac{f_{\rm gas}}{240}\right)^{-2/5} \left(\frac{\Sigma_1}{10}\right)^{-9/10} \nonumber \\ & & \left(\frac{b}{10 r_{\rm H}}\right)^{1/10} \left(\frac{a}{\rm 1 \hspace{0.3em} AU}\right)^{8/5+9x/10} \left(\frac{M_{\star}}{M_{\odot}}\right)^{-8/15} \nonumber \\ & & \left(\frac{\rho_{\rm p}}{2 \hspace{0.3em} {\rm gcm}^{-3}}\right)^{11/15} \left(\frac{r_{\rm p}}{100 \hspace{0.3em} {\rm km}}\right)^{2/5} \hspace{0.3em} {\rm yr},\end{aligned}$$ where $r_{\rm p}$ and $\rho_{\rm p}$ are the physical radius and internal density of planetesimals. Eq. (\[eq:me2\]) indicates that the embryo gains 44$\%$, 90$\%$, and 99$\%$ of its final mass during 1$\tau_{\rm grow}$, 2$\tau_{\rm grow}$, and 3$\tau_{\rm grow}$. For the standard disk model defined above, $M_\mathrm{iso} \sim 0.1 \mearth$ in the terrestrial planet region. This suggests that if they formed by oligarchic growth, Mercury and Mars may simply represent leftover planetary embryos. A short growth timescale ($\tau_{\rm grow} < 2$ Myr) of Mars estimated by the Hf-W chronology [@dauphas11] would suggest that Mars accreted from a massive disk of small planetesimals [@kobayashi13; @morishima13]. Alternately, accretion of larger planetesimals might have been truncated as proposed by the Grand Tack model (see §6.3). Unlike Mars and Mercury, further accretion of planetary embryos is necessary to complete Venus and Earth. This next, final stage is called late-stage accretion (see Section 4). 3.2 Embryo formation by pebble accretion [@lambrechts12], hereafter LJ12, proposed a new model of growth for planetary embryos and giant planet cores. They argued that if the disk’s mass is dominated by pebbles of a few decimeters in size, the largest planetesimals accrete pebbles very efficiently and can rapidly grow to several Earth masses [see also @johansen10; @ormel10; @murrayclay11]. This model builds on a recent planetesimal formation model in which large planetesimals (with sizes from $\sim 100$ up to $\sim$1,000km) form by the collapse of a self-gravitating clump of pebbles, concentrated to high densities by disk turbulence and the streaming instability [@youdin05; @johansen06; @johansen07; @johansen09 see also chapter by Johansen et al]. The pebble accretion scenario essentially describes how large planetesimals continue to accrete. There is observational evidence for the existence of pebble-sized objects in protoplanetary disks [@wilner05; @rodmann06; @lommen07; @perez12], although their abundance relative to larger objects (planetesimals) is unconstrained. Pebbles are strongly coupled with the gas so they encounter the already-formed planetesimals with a velocity $\Delta v$ that is equal to the difference between the Keplerian velocity and the orbital velocity of the gas, which is slightly sub-Keplerian due to the outward pressure gradient. LJ12 define the planetesimal [Bondi radius]{} as the distance at which the planetesimal exerts a deflection of one radian on a particle approaching with a velocity $\Delta v$: $$R_B={{GM}\over{\Delta v^2}} \label{Bondi}$$ where $G$ is the gravitational constant and $M$ is the planetesimal mass (the deflection is larger if the particle passes closer than $R_B$). LJ12 showed that all pebbles with a stopping time $t_f$ smaller than the Bondi time $t_B=R_B/\Delta v$ that pass within a distance $R=(t_f/t_B)^{1/2} R_B$ spiral down towards the planetesimal and are accreted by it. Thus, the growth rate of the planetesimal is: $$\mathrm{d}M/\mathrm{d}t=\pi\rho R^2\Delta v \label{Bondi-accrete}$$ where $\rho$ is the volume density of the pebbles in the disk. Because $R\propto M$, the accretion rate $\mathrm{d}M/\mathrm{d}t\propto M^2$. Thus, pebble accretion is at the start a super-runaway process that is faster than the runaway accretion scenario (see Sec 3.1) in which $\mathrm{d}M/\mathrm{d}t\propto M^{4/3}$. According to LJ12, this implies that in practice, only planetesimals more massive than $\sim 10^{-4}\mearth$ (comparable to Ceres’ mass) undergo significant pebble accretion and can become embryos/cores. The super-runaway phase cannot last very long. When the Bondi radius exceeds the scale height of the pebble layer, the accretion rate becomes $$\mathrm{d}M/\mathrm{d}t=2R\Sigma \Delta v$$ where $\Sigma$ is the surface density of the pebbles. This rate is proportional to $M$, at the boundary between runaway and orderly (oligarchic) growth. ![Trajectories of particles in the vicinity of a growing embryo. The black curves represent particles strongly coupled to the gas and the gray curves particles that are weakly coupled, as measured by the ratio of the stopping time $t_f$ to the Bondi time $t_B$. The orbits of weakly-coupled particles are deflected by the embryo’s gravity, but the strongly coupled particles spiral inward and are quickly accreted onto the embryo. From [@lambrechts12].[]{data-label="fig:pebble"}](fig3.eps){width="45.00000%"} Moreover, when the Bondi radius exceeds the Hill radius $R_H = a \left[M/(3 M_\star)\right]^{1/3}$, the accretion rate becomes $$\mathrm{d}M/\mathrm{d}t=2R_H \Sigma v_H \label{Hill-accretion}$$ where $v_H$ is the Hill velocity (i.e. the difference in Keplerian velocities between two circular orbits separated by $R_H$). Here $\mathrm{d}M/\mathrm{d}t \propto M^{2/3}$ and pebble accretion enters an oligarchic regime. For a given surface density of solids $\Sigma$ the growth of an embryo is much faster if the solids are pebble-sized than planetesimal sized. This is the main advantage of the pebble-accretion model. However, pebble accretion ends when the gas disappears from the protoplanetary disk, whereas runaway/oligarchic accretion of planetesimals can continue. Also, the ratio between $\Sigma_{\rm planetesimals}/\Sigma_{\rm pebbles}$ remains to be quantified, and ultimately it is this ratio that determines which accretion mechanism is dominant. An important problem in Solar System formation is that the planetary embryos in the inner solar system are thought to have grown only up to at most a Mars-mass, whereas in the outer solar system some of them reached many Earth masses, enough to capture a primitive atmosphere and become giant planets. The difference between these masses can probably be better understood in the framework of the pebble-accretion model than in the planetesimal-accretion model. The dichotomy in embryo mass in the inner/outer Solar System might have been caused by radial drift of pebbles. We consider a disk with a “pressure bump” [@johansen09] at a given radius $R_{\rm bump}$. At this location the gas’ azimuthal velocity $v_\theta$ is larger than the Kepler velocity $v_K$. Pebbles cannot drift from beyond $R_{\rm bump}$to within $R_{\rm bump}$ because they are too strongly coupled to the gas. Embryos growing interior to $R_{\rm bump}$ are thus “starved” in the sense that they can only accrete pebbles within $R_{\rm bump}$, and are not in contact with the presumably much larger pebble reservoir beyond $R_{\rm bump}$. Of course, embryos growing exterior to $R_{\rm bump}$ would not be starved and could grow much faster and achieve much larger masses within the gaseous disk’s lifetime. On the contrary, the planetesimal accretion model does not seem to present a sharp radial boundary for slow/fast accretion and so it is harder to understand the dichotomy of embryo masses in that framework. [@ida08] argued that a pressure bump could be located at the snow line. If this is true, then we can speculate that giant planet cores should form in the icy part of the disk and sub-Mars-mass planetary embryos in the rocky part of the disk. This seems to be consistent with the structure of the Solar System. FROM PLANETARY EMBRYOS TO TERRESTRIAL PLANETS ================================================= The final accumulation of terrestrial planets – sometimes called late-stage accretion – is a chaotic phase characterized by giant embryo-embryo collisions. It is during this phase that the main characteristics of the planetary system are set: the planets’ masses and orbital architecture, the planets’ feeding zones and thus their bulk compositions, and their spin rates and obliquities [although their spins may be altered by other processes on long timescales – see e.g., @correia09]. Whether embryos form by accreting planetesimals or pebbles, the late evolution of a system of embryos is likely in the oligarchic regime. The transition from oligarchic growth to late-stage accretion happens when there is insufficient damping of random velocities by gas drag and dynamical friction from planetesimals [@kenyon06]. The timescale of the orbital instability of an embryo system has been numerically calculated by $N$-body simulations to be $$\log t_\mathrm{inst} \simeq c_1 \left(\frac{b_\mathrm{ini}}{r_\mathrm{H}}\right) + c_2,$$ where $b_\mathrm{ini}$ is the initial orbital separation of adjacent embryos and $c_1$ and $c_2$ are functions of the initial $\langle e^2\rangle^{1/2}$ and $\langle i^2\rangle^{1/2}$ of the system [@chambers96; @yoshinaga99]. The most important quantity in determining the outcome of accretion is the level of eccentricity excitation of the embryos. This is determined by a number of parameters including forcing from any giant planets that exist in the system [@chambers02; @levison03; @raymond04]. Although giant planets are far larger than terrestrials, they are thought to form far faster and strongly influence late-stage terrestrial accretion. The lifetimes of gaseous protoplanetary disks are just a few Myr [@haisch01] whereas geochemical constraints indicate that Earth took 50-100 Myr to complete its formation [@touboul07; @kleine09; @konig11]. The dynamics described in this section are assumed to occur in a gas-free environment (we consider the effects of gas in other sections). We first describe the dynamics of accretion and radial mixing (§4.1), then the effect of accretion on the final planets’ spins (§4.2) and the effect of embryo and disk parameters on accretion (§4.3). We explain the consequences of taking into account imperfect accretion (§4.4) and the effect of giant planets on terrestrial accretion (§4.5). 4.1 Timescales and Radial Mixing Figure \[fig:sim0\] shows the evolution of a simulation of late-stage accretion from [@raymond06b] that included a single Jupiter-mass giant planet on a circular orbit at 5.5 AU. The population of embryos is excited from the inside-out by mutual scattering among bodies and from the outside-in by secular and resonant excitation by the giant planet. Accretion is faster closer-in and proceeds as a wave sweeping outward in time. At 10 Myr the disk inside 1 AU is dominated by 4 large embryos with masses close to Earth’s. The population of close-in (red) planetesimals has been strongly depleted, mainly by accretion but also by some scattering to larger orbital radii. Over the rest of the simulation the wave of accretion sweeps outward across the entire system. Small bodies are scattered onto highly-eccentric orbits and either collide with growing embryos or venture too close to the giant planet and are ejected from the system. Embryos maintain modest eccentricities by dynamical friction from the planetesimals. Nonetheless, strong embryo-embryo gravitational scattering events spread out the planets and lead to giant impacts such as the one thought to be responsible for creating Earth’s Moon [@cuk12; @canup12]. After 200 Myr three terrestrial planets remain in the system with masses of 1.54, 2.04, and $0.95 \mearth$ (inner to outer). Although modestly more massive, the orbits of the two inner planets are decent analogs for Earth and Venus. The outer planet does a poor job of reproducing Mars: it is nine times too massive and too far from the star. This underscores the [small Mars]{} problem: simulations that do not invoke strong external excitation of the embryo swarm systematically produce Mars analogs that are far too massive [@wetherill91; @raymond09c]. We will return to this problem in §6. A large reservoir of water-rich material is delivered to the terrestrial planets in the simulation from Fig. \[fig:sim0\]. By 10 Myr four large embryos have formed inside 1 AU but they remain dry because to this point their feeding zones have been restricted to the inner planetary system. Over the following 20 Myr planetesimals and embryos from the outer planetary system are scattered inward by repeated gravitational encounters with growing embryos. These bodies sometimes collide with the growing terrestrial planets. This effectively widens the feeding zones of the terrestrial planets to include objects that condensed at different temperatures and therefore have different initial compositions [see also @bond10; @carterbond12; @elser12]. The compositions of the terrestrial planets become mixtures of the compositions of their constituent embryos and planetesimals. The planets’ feeding zones represent those constituents. When objects from past 2.5 AU are accreted, water-rich material is delivered to the planet in the form of hydrated embryos and planetesimals. In the simulations, from 30-200 Myr the terrestrial planets accrete objects from a wide range of initial locations and are delivered more water. Given that the water delivered to the planets in this simulation originated in the region between 2.5 and 4 AU, its composition should be represented by carbonaceous chondrites, which provide a very good match to Earth’s water [@morby00; @marty06]. The planets are delivered a volume of water that may be too large. For example, the Earth analog’s final water content by mass was $8 \times 10^{-3}$, roughly 8-20 times the actual value. However, water loss during giant impacts was not taken into account in the simulation [see, e.g., @genda05]. 4.2 Planetary spins Giant impacts impart large amounts of spin angular momentum on the terrestrial planets [e.g., @safronov69; @lissauer91; @dones93]. The last few giant impacts tend to dominate the spin angular momentum [@agnor99; @kokubo07; @kokubo10]. Using a “realistic” accretion condition of planetary embryos [@genda12 see §4.4)], [@kokubo10] found that the spin angular velocity of accreted terrestrial planets follows a Gaussian distribution with a nearly mass-independent average value of about 70% of the critical angular velocity for rotational breakup $$\omega_{\rm cr} = \left(\frac{GM}{R^3}\right)^{1/2},$$ where $M$ and $R$ are the mass and radius of a planet. This appears to be a natural outcome of embryo-embryo impacts at speeds slightly larger than escape velocity. At later times, during the late veneer phase, the terrestrial planets’ spins are further affected by impacts with planetesimals [@raymond13]. The obliquity of accreted planets ranges from 0$^\circ$ to 180$^\circ$ and follows an isotropic distribution [@agnor99; @kokubo07; @kokubo10]. Both prograde and retrograde spins are equally probable. The isotropic distribution of $\varepsilon$ is a natural outcome of giant impacts. During the giant impact stage, the thickness of a planetary embryo system is $\sim a\langle i^2\rangle^{1/2} \sim 10r_{\rm H}$, far larger than the radius $R$ of planetary embryos $R\sim 10^{-2}r_{\rm H}$, where $a$, $i$, and $r_\mathrm{H}$ are the semimajor axis, inclination and Hill radius of planetary embryos. Thus, collisions are fully three-dimensional and isotropic, which leads to isotropic spin angular momentum. This result clearly shows that prograde spin with small obliquity, which is common to the terrestrial planets in the solar system except for Venus, is not a common feature for planets assembled by giant impacts. Note that the initial obliquity of a planet determined by giant impacts can be modified substantially by stellar tide if the planet is close to the star and by satellite tide if the planet has a large satellite. 4.3 Effect of disk and embryo parameters The properties of a system of terrestrial planets are shaped in large part by the total mass and mass distribution within the disk, and the physical and orbital properties of planetary embryos and planetesimals within the disk. However, while certain parameters have a strong impact on the outcome, others have little to no effect. [@kokubo06] performed a suite of simulations of accretion of populations of planetary embryos to test the importance of the embryo density, mass, spacing and number. They found that the bulk density of the embryos had little to no effect on the accretion within the range that they tested, $\rho = 3.0-5.5 {\rm g \, cm^{-3}}$. One can imagine that the dynamics could be affected for extremely high values of $\rho$, if the escape speed from embryos were to approach a significant fraction of the escape speed from the planetary system [@goldreich04]. In practice this is unlikely to occur in the terrestrial planet forming region because it would require unphysically-large densities. The initial spacing likewise had no meaningful impact on the outcome, at least when planetary embryos were spaced by 6-12 mutual Hill radii [@kokubo06]. Likewise, for a fixed total mass in embryos, the embryo mass was not important. The total mass in embryos does affect the outcome. A more massive disk of embryos and planetesimals produces fewer, more massive planets than a less massive disk [@kokubo06; @raymond07b]. Embryos’ eccentricities are excited more strongly in massive disks by encounters with massive embryos. With larger mean eccentricities, the planets’ feeding zones are wider than if the embryos’ eccentricities were small, simply because any given embryos crosses a wider range of orbital radii. The scaling between the mean accreted planet mass and the disk mass is therefore slightly steeper than linear: the mean planet mass $M_p$ scales with the local surface density $\Sigma_0$ as $M_p \propto \Sigma_0^{1.1}$ [@kokubo06]. It is interesting to note that this scaling is somewhat shallower than the $\Sigma_0^{1.5}$ scaling of embryo mass with the disk mass [@kokubo00]. Accretion also proceeds faster in high-mass disks, as the timescale for interaction drops. Terrestrial planets that grow from a disk of planetesimals and planetary embryos retain a memory of the surface density profile of their parent disk. In addition, the dynamics is influenced by which part of the disk contains the most mass. In disks with steep density profiles – i.e., if the surface density scales with orbital radius as $\Sigma \propto r^{-x}$, disks with large values of $x$ – more mass is concentrated in the inner parts of the disk, where the accretion times are faster and protoplanets are dry. Compared with disks with shallower density profiles (with small $x$), in disks with steep profiles the terrestrial planets tend to be more massive, form more quickly, form closer-in, and contain less water [@raymond05b; @kokubo06]. 4.4 Effect of imperfect accretion As planetesimalsÕ eccentricities are excited by growing embryos, they undergo considerable collisional grinding. Collisional disruption can be divided into two types: catastrophic disruption due to high-energy impacts and cratering due to low-energy impacts. [@kobayashi10b] found that cratering collisions are much more effective in collisional grinding than collisions causing catastrophic disruption, simply because the former impacts occur much more frequently than the latter ones. Small fragments are easily accreted by embryos in the presence of nebular gas [@wetherill93], although they rapidly drift inward due to strong gas drag, leading to small embryo masses [@chambers08; @kobayashi10]. Giant impacts between planetary embryos often do not result in net accretion. Rather, there exists a diversity of collisional outcomes. These include near-perfect merging at low impact speeds and near head-on configurations, partial accretion at somewhat higher impact speeds and angles, “hit and run” collisions at near-grazing angles, and even net erosion for high-speed, near head-on collisions [@agnor04; @asphaug06; @asphaug10]. Two recent studies used large suites of SPH simulations to map out the conditions required for accretion in the parameter space of large impacts [@genda12; @leinhardt12]. However, most $N$-body simulations of terrestrial planet formation to date have assumed perfect accretion in which all collisions lead to accretion. About half of the embryo-embryo impacts in a typical simulation of late-stage accretion do not lead to net growth [@agnor04; @kokubo10]. Rather, the outcomes are dominated by partially accreting collision, hit-and-run impacts, and graze-and-merge events in which two embryos dissipate sufficient energy during a grazing impact to become gravitationally bound and collide [@leinhardt12]. Taking into account only the accretion condition for embryo-embryo impacts, the final number, mass, orbital elements, and even growth timescale of planets are barely affected [@kokubo10; @alexander98]. This is because even though collisions do not lead to accretion, the colliding bodies stay on the colliding orbits after the collision and thus the system is unstable and the next collision occurs shortly. However, by allowing non-accretionary impacts to both erode the target embryo and to produce debris particles, [@chambers13] found that fragmentation does have a noted effect on accretion. The final stages of accretion are lengthened by the sweep up of collisional fragments. The planets that formed in simulations with fragmentation had smaller masses and smaller eccentricities than their counterparts in simulations without fragmentation. Imperfect accretion also affects the planets’ spin rates. [@kokubo10] found that the spin angular momentum of accreted planets was 30% smaller than in simulations with perfect accretion. This is because grazing collisions that have high angular momentum are likely to result in a hit-and-run, while nearly head-on collisions that have small angular momentum lead to accretion. The production of unbound collisional fragments with high angular momentum could further reduce the spin angular velocity. The effect of non-accretionary impacts on the planetary spins has yet to be carefully studied. A final consequence of fragmentation is on the core mass fraction. Giant impacts lead to an increase in the core mass fraction because the mantle is preferentially lost during imperfect merging events [@benz07; @stewart12; @genda12]. However, the sweep-up of these collisional fragments on 100 Myr timescales re-balances the composition of planets to roughly the initial embryo composition [@chambers13]. We speculate that a net increase in core mass fraction should be retained if the rocky fragments are allowed to collisionally evolve and lose mass. 4.5 Effect of outer giant planets We now consider the effect of giant planets on terrestrial accretion. We restrict ourselves to systems with giant planets similar to our own Jupiter and Saturn. That is, systems with non-migrating giant planets on stable orbits exterior to the terrestrial planet-forming region. In §5.2 we will consider the effects of giant planet migration and planet-planet scattering. The most important effect of giant planets on terrestrial accretion is the excitation of the eccentricities of planetary embryos. This generally occurs by the giant planet-embryo gravitational eccentricity forcing followed by the transmission of that forcing by embryo-embryo or embryo-planetesimal forcing. The giant planet forcing typically occurs via mean motion or secular resonances, or secular dynamical forcing. Giant planet-embryo excitation is particularly sensitive to the giant planets’ orbital architecture [@chambers02; @levison03; @raymond06a]. Figure \[fig:testp\] shows the eccentricities of test particles excited for 1 Myr by two different configurations of Jupiter and Saturn [@raymond09c], both of which are consistent with the present-day Solar System (see §6). The spikes in eccentricity seen in Fig. \[fig:testp\] come from specific resonances: in the [JSRES]{} configuration (for “Jupiter and Saturn in RESonance”), the $\nu_5$ secular resonance at 1.3 AU and the 2:1 mean motion resonance with Jupiter at 3.4 AU; and in the [EEJS]{} configuration (for “Extra-Eccentric Jupiter and Saturn”) the $\nu_5$ and $\nu_6$ secular resonances at 0.7 and 2.1 AU, and a hint of the 2:1 mean motion resonance with Jupiter at 3.3 AU. The “background” level of excitation seen in Fig. \[fig:testp\] comes from secular forcing, following a smooth function of the orbital radius. The eccentricity excitation of terrestrial embryos is significant even for modest values of the giant planets’ eccentricity. In Fig. \[fig:testp\], Jupiter and Saturn have eccentricities of 0.01-0.02 in the [JSRES]{} configuration and of 0.1 in the [EEJS]{} configuration. The test particles in the [JSRES]{} system are barely excited by the giant planets interior to 3 AU; the magnitude of the spike at 1.3 AU is far smaller than the secular forcing anywhere in the [EEJS]{} simulation. Note also that this figure represents just the first link in the chain. The eccentricities imparted to embryos are systematically transmitted to the entire embryo swarm, and it is the mean eccentricity of the embryo swarm that dictates the outcome of accretion. In a population of embryos with near-circular orbits, the communication zone – the radial distance across which a given embryo has gravitational contact with its neighbors – is very narrow. Embryos grow by collisions with their immediate neighbors. The planets that form are thus limited in mass by the mass in their immediate vicinity. In contrast, in a population of embryos with significant eccentricities, the communication zone of embryos is wider. Each embryo’s orbit crosses the orbits of multiple other bodies and, by secular forcing, gravitationally affects the orbits of even more. This of course does not imply any imminent collisions, but it does mean that the planets that form will sample a wider radial range of the disk than in the case of very low embryo eccentricities. This naturally produces a smaller number of more massive planets. Given that collisions preferentially occur at pericenter, the terrestrial planets that form tend to also be located closer-in when the mean embryo eccentricity is larger [@levison03]. In systems with one or more giant planets on orbits exterior to the terrestrial planet-forming region, the amplitude of excitation of the eccentricities of terrestrial embryos is larger when the giant planets’ orbits are eccentric or closer-in. The timescale for excitation is shorter when the giant planets are more massive. Thus, the strongest perturbations come from massive eccentric gas giants. Simulations have indeed shown that systems with massive or eccentric outer gas giants systematically produce fewer, more massive terrestrial planets [@chambers02; @levison03; @raymond04]. However, the efficiency of terrestrial accretion is smaller in the presence of a massive or eccentric gas giant because a fraction of embryos and planetesimals are excited onto orbits that are unstable and are thus removed from the system. The most common mechanism for the removal of such bodies is by having their eccentricities increased to the point where their orbits cross those of a giant planet, then being ejected entirely from the system into interstellar space. The strong outside-in perturbations produced by massive or eccentric outer gas giants also act to accelerate terrestrial planet formation. This happens for two reasons. First, when embryos have significant mean eccentricities the typical time between encounters decreases, as long as eccentricities are more strongly perturbed than inclinations. Second, accretion is slower in the outer parts of planetary systems because of the longer orbital and encounter timescales, and it is these slow-growing regions that are most efficiently cleared by the giant planets’ perturbations. Given their outside-in influence, outer gas giants also play a key role in water delivery to terrestrial planets. It should be noted up front that the gas giants’ role in water delivery is purely detrimental, at least in the context of outer giant planets on static orbits. Stimulating the eccentricities of water-rich embryos at a few AU can in theory cause some embryos to be scattered inward and deliver water to the terrestrial planets. In practice, a much larger fraction of bodies is scattered outward, encounters the giant planets and is ejected from the system than is scattered inward to deliver water [@raymond06b]. Finally, simulations with setups similar to the one from Fig. \[fig:sim0\] confirm that the presence of one or more giant planets strongly anti-correlates with the water content of the terrestrial planets in those systems [@chambers02; @raymond04; @raymond06b; @raymond07a; @raymond09c; @obrien06]. There is a critical orbital radius beyond which a giant planet must lie for terrestrial planets to accrete and survive in a star’s liquid water habitable zone [@raymond06a]. This limit is eccentricity dependent: a zero-eccentricity (single) giant planet must lie beyond 2.5 AU to allow a terrestrial planet to form between 0.8 and 1.5 AU whereas a giant planet with an eccentricity of 0.3 must lie beyond 4.2 AU. For water to be delivered to the terrestrial planets from a presumed source at 2-4 AU (as in Fig. \[fig:sim0\]) the giant planet must be farther still [@raymond06a]. TERRESTRIAL ACCRETION IN EXTRA-SOLAR PLANETARY SYSTEMS ========================================================== Extra-solar planetary systems do not typically look like the Solar System. To extrapolate to extra-solar planetary systems is therefore not trivial. Additional mechanisms must be taken into account, in particular orbital migration both of planetary embryos (Type 1 migration) and of gas giant planets (Type 2 migration) and dynamical instabilities in systems of multiple gas giant planets. There exists ample evidence that accretion does indeed occur around other stars. Not only has an abundance of low-mass planets been detected [@mayor11; @batalha13], but the dust produced during terrestrial planet formation [@kenyon04] has also been detected [e.g. @meyer08; @lisse08], including the potential debris from giant embryo-embryo impacts [@lisse09]. In this section we first address the issue of the formation of hot Super Earths. Then we discuss how the dynamics shaping the known systems of giant planets may have sculpted unseen terrestrial planets in those systems. 5.1 Hot Super Earths Hot Super Earths are extremely common. Roughly one third to one half of Sun-like (FGK) stars host at least one planet with a mass less than $10 \mearth$ and a period of less than $50-100$ days [@howard10; @mayor11]. The frequency of Hot Super Earths is at least as high around M stars as around FGK stars and possibly higher [@howard12; @bonfils13; @fressin13]. Hot Super Earths are typically found in systems of many planets on compact but non-resonant orbits [e.g. @udry07; @lovis11; @lissauer11]. Several mechanisms have been proposed to explain the origin of Hot Super Earths [see @raymond08a]: 1) In situ accretion from massive disks of planetary embryos and planetesimals; 2) Accretion during inward type 1 migration of planetary embryos; 3) Shepherding in interior mean motion resonances with inward-migrating gas giant planets; 4) Shepherding by inward-migrating secular resonances driven by dissipation of the gaseous disk; 5) Circularization of planets on highly-eccentric orbits by star-planet tidal interactions; 6) Photo-evaporation of close-in gas giant planets. Theoretical and observational constraints effectively rule out mechanisms 3-6. The shepherding of embryos by migrating resonances (mechanisms 3 and 4) can robustly transport material inward [@zhou05; @fogg05; @fogg07; @raymond06c; @mandell07; @gaidos07]. An embryo that finds itself in resonance with a migrating giant planet will have its eccentricity simultaneously excited by the giant planet and damped by tidal interactions with the gaseous disk [@tanaka04; @cresswell07]. As the tidal damping process is non-conservative, the embryo’s orbit loses energy and shrinks, removing the embryo from the resonance. The migrating resonance catches up to the embryo and the process repeats itself, moving the embryo inward, potentially across large distances. This mechanism is powered by the migration of a strong resonance. This requires a connection between Hot Super Earths and giant planets. If a giant planet migrated inward, and the shepherd was a mean motion resonance (likely the 3:2, 2:1 or 3:1 resonance) then hot Super Earths should be found just interior to close-in giant planets, which is not observed. If a strong secular resonance migrated inward then at least one giant planet on an eccentric orbit must exist exterior to the hot Super Earth, and there should only be a small number of Hot Super Earths. This is also not observed. Tidal circularization of highly-eccentric Hot Super Earths (mechanism 5) is physically possible but requires extreme conditions [@raymond08a]. Star-planet tidal friction of planets on short-pericenter orbits can rapidly dissipate energy, thereby shrinking and re-circularizing the planets’ orbits. This process has been proposed to explain the origin of hot Jupiters [@ford06; @fabrycky07; @beauge12], and the same mechanism could operate for low-mass planets. Very close pericenter passages – within 0.02 AU – are required for significant radial migration [@raymond08a]. Although such orbits are plausible, another implication of the model is that, given their large prior eccentricities, hot Super Earths should be found in single systems with no other planets nearby. This is not observed. The atmospheres of very close-in giant planets can be removed by photo-evaporation from the host star [mechanism 6; @lammer03; @baraffe04; @baraffe06; @yelle04; @erkaev07; @hubbard07a; @raymond08a; @murrayclay09; @lopez13]. The process is driven by UV heating from the central star. Mass loss is most efficient for planets with low surface gravities extremely close to UV-bright stars. Within $\sim 0.02$ AU, planets as large as Saturn can be photo-evaporated down to their cores on Gyr timescales. Since both the photoevaporation rate and the rate of tidal evolution depend on the planet mass, a very close-in rocky planet like [Corot-7b]{} [@leger09] could have started as a Saturn-mass planet on a much wider orbit [@jackson10]. Although photo-evaporation may cause mass loss in some very close-in planets, it cannot explain the systems of hot Super Earths. [@hubbard07b] showed that the mass distributions of very highly-irradiated planets within 0.07 AU was statistically indistinguishable from the mass distribution of planets at larger distances. In addition, given the very strong radial dependence of photo-evaporative mass loss, the mechanism is likely to produce systems with a single hot Super Earth as the closest-in planet rather than multiple systems of hot Super Earths. Given the current constraints from Kepler and radial velocity observations, mechanisms 1 and 2 – in situ accretion and type 1 migration – are the leading candidates to explain the formation of the observed Hot Super Earths. Of course, we cannot rule out additional mechanisms that have yet to come to light. For systems of hot Super Earths to have accreted in situ from massive populations of planetesimals and planetary embryos, their protoplanetary disks must have been very massive [@raymond08a; @hansen12; @hansen13; @chiang13; @raymond14]. The observed systems of hot Super Earths often contain $20-40 \mearth$ in planets within a fraction of an AU of the star [@batalha13]. Let us put this in the context of simplified power-law disks: $$\Sigma = \Sigma_0 \left(\frac{r}{1 {\rm AU}}\right)^{-x}.$$ The minimum-mass Solar Nebula (MMSN) model [@weidenschilling77b; @hayashi85] has $x = 3/2$, although modified versions have $x = 1/2$ [@davis05] and $x \approx 2$ [@desch07]. [@chiang13] created a minimum-mass [extrasolar]{} nebula using the Kepler sample of hot Super Earths and found a best fit for $x = 1.6-1.7$ with a mass normalization roughly ten times higher than the MMSN. However, [@raymond14] showed that minimum-mass disks based on Kepler multiple-planet systems actually cover a broad range in surface density slopes and are inconsistent with a universal underlying disk profile. Only steep power-law disks allow for a significant amount of mass inside 1 AU. Consider a disk with a mass of $0.05 M_{\odot}$ extending from zero to 50 AU with an assumed dust-to-gas ratio of 1%. This disk contains a total of $150 \mearth$ in solids. If the disk follows an $r^{-1/2}$ profile (i.e., with $x = 1/2$) then it only contains $0.4 \mearth$ in solids inside 1 AU. If the disk has $x=1$ then it contains $3 \mearth$ inside 1 AU. If the disk has $x=1.5-1.7$ then it contains $21-46 \mearth$ inside 1 AU. Sub-mm observations of cold dust in the outer parts of nearby protoplanetary disks generally find values of $x$ between $1/2$ and 1 [@mundy00; @looney03; @andrews07b]. However, the inner parts of disks have yet to be adequately measured. The dynamics of in situ accretion of hot Super Earths would presumably be similar to the well-studied dynamics of accretion presented in sections 3 and 4. Accretion would proceed faster than at 1 AU due to the shorter relevant timescales, but would consist of embryo-embryo and embryo-planetesimal impacts [@raymond08a]. However, even if Super Earths accrete modest gaseous envelopes from the disk, these envelopes are expected be lost during the dispersal of the protoplanetary disk under most conditions [@ikoma12]. This loss process is most efficient at high temperatures, making it hard to explain the large radii of some detected Super Earths. Nonetheless, Super Earths that form by in situ accretion appear to match several other features of the observed population, including their low mutual inclination orbits and the distributions of eccentricity and orbital spacing [@hansen13]. Alternately, the formation of hot Super Earths may involve long-range orbital migration [@terquem07]. Once they reach $\sim 0.1 \mearth$, embryos are susceptible to type 1 migration [@goldreich80; @ward86]. Type 1 migration may be directed inward or outward depending on the local disk properties and the planet mass [@paardekooper11; @masset10; @kretke12]. In most disks outward migration is only possible for embryos larger than a few Earth masses. All embryos therefore migrate inward when they are small. If they grow quickly enough during the migration then in some regions they can activate their corotation torque and migrate outward. A population of inward-migrating embryos naturally forms a resonant chain. Migration is stopped at the inner edge of the disk [@masset06] and the resonant chain piles up against the edge [@ogihara09]. If the resonant chain gets too long, cumulative perturbations from the embryos act to destabilize the chain, leading to accretionary collisions and a new shorter resonant chain [@morby08; @cresswell08]. This process can continue throughout the lifetime of the gaseous disk and include multiple generations of inward-migrating embryos or populations of embryos. Figure \[fig:type1\] shows the formation of a system of hot Super Earths by type 1 migration from [@cossou13b]. In this simulation $60 \mearth$ in embryos with masses of $0.1-2\mearth$ started from 2-15 AU. The embryos accreted as they migrated inward in successive waves. One embryo (shown in red in Fig. \[fig:type1\]) grew large enough to trigger outward migration and stabilized at a zero-torque zone in the outer disk, presumably to become giant planet core. The system of hot Super Earths that formed is similar in mass and spacing to the Kepler-11 system [@lissauer11]. The four outer super Earths are in a resonant chain but the inner one was pushed interior to the inner edge of the gas disk and removed from resonance. It was proposed by [@raymond08a] that transit measurements of hot Super Earths could differentiate between the in situ accretion and type 1 migration models. They argued that planets formed in situ should be naked high-density rocks whereas migrated planets are more likely to be dominated by low-density material such as ice. It has been claimed that planets that accrete in situ can have thick gaseous envelopes and thus inflated radii [@hansen12; @chiang13]. However, detailed atmospheric calculations by [@ikoma12] suggest that it is likely that low-mass planets generally lose their atmospheres during disk dispersal. This is a key point. If these planets can indeed retain thick atmospheres then simple measurements of the bulk density of Super Earths wold not provide a mechanism for differentiation between the models. However, if hot Super Earths cannot retain thick atmospheres after forming in situ, then low density planets must have formed at larger orbital distances and migrated inward. It is possible that migration and in situ accretion both operate to reproduce the observed hot Super Earths. The main shortcoming of in situ accretion model is that the requisite inner disk masses are extremely large and do not fit the surface density profiles measured in the outskirts of protoplanetary disks. Type 1 migration of planetary embryos provides a natural way to concentrate solids in the inner parts of protoplanetary disks. One can envision a scenario that proceeds as follows. Embryos start to accrete locally throughout the disk. Any embryo that grows larger than roughly a Mars mass type 1 migrates inward. Most embryos migrate all the way to the inner edge of the disk, or at least to the pileup of embryos bordering on the inner edge. There are frequent close encounters and impacts between embryos. The embryos form long resonant chains that are successively broken by perturbations from other embryos or by stochastic forcing from disk turbulence [@terquem07; @pierens11]. As the disk dissipates the resonant chain can be broken, leading to a last phase of collisions that effectively mimics the in situ accretion model. There remains sufficient gas and collisional debris to damp the inclinations of the surviving Super Earths to values small enough to be consistent with observations. However, that it is possible that many Super Earths actually remain in resonant orbits but with period ratios altered by tidal dissipation [@batygin13]. 5.2 Sculpting by giant planets: type 2 migration and dynamical instabilities The orbital distribution of giant exoplanets is thought to have been sculpted by two dynamical processes: type 2 migration and planet-planet scattering [@moorhead05; @armitage07]. These processes each involve long-range radial shifts in giant planets’ orbits and have strong consequences for terrestrial planet formation in those systems. In fact, each of these processes has been proposed to explain the origin of hot Jupiters [@lin96; @nagasawa08], so differences in the populations of terrestrial planets, once observed, could help resolve the question of the origin of hot Jupiters. Only a fraction of planetary systems contain giant planets. About 14% of Sun-like stars host a gas giant with period shorter than 1000 days [@mayor11], although the fraction of stars with more distant giant planets could be significantly higher [@gould10]. When a giant planet becomes massive enough to open a gap in the protoplanetary disk, its orbital evolution becomes linked to the radial viscous evolution of the gas. This is called Type 2 migration [@lin86; @ward97]. As a giant planet migrates inward it encounters other small bodies in various stages of accretion. Given the strong damping of eccentricities by the gaseous disk, a significant fraction of the material interior to the giant planet’s initial orbit is shepherded inward by strong resonances as explained in §5.1 [@zhou05; @fogg05; @fogg07; @fogg09; @raymond06c; @mandell07]. Indeed, the simulation from the left panel of Figure \[fig:aet\_giants\] formed two hot Super Earth planets, one just interior to the 2:1 and 3:1 resonance. The orbits of the two planets became destabilized after several Myr, collided and fused into a single $4 \mearth$ hot Super Earth. There also exists a population of very close-in planetesimals in the simulation from Fig. \[fig:aet\_giants\]; these were produced by the same shepherding mechanism as the hot Super Earths but, because the dissipative forces from gas drag were so much stronger for these objects than the damping due to disk-planet tidal interactions felt by the embryos [@adachi76; @ida08b], they were shepherded by a much higher-order resonance, here the 8:1. Planetesimals or embryos that come too close to the migrating giant are scattered outward onto eccentric orbits. These orbits are slowly re-circularized by gas drag and dynamical friction. On 10-100 Myr or longer timescales a second generation of terrestrial planets can form from this scattered material [@raymond06c; @mandell07]. The building blocks of this new generation of planets are significantly different than the original ones. This new distribution is comprised of two components: bodies that originated across the inner planetary system that were scattered outward by the migrating gas giant, and bodies that originated exterior to the gas giant. When taking into account the original location of these protoplanets, the effective feeding zone of the new terrestrial planets essentially spans the entire planetary system. This new generation of terrestrial planets therefore inevitably contains material that condensed at a wide range of orbital distances. Their volatile contents are huge. Indeed, the water content of the $3 \mearth$ planet that formed at 0.9 AU (in the shaded habitable zone) in Fig. \[fig:aet\_giants\] is roughly 10% by mass. Even if 90% of the water were lost during accretion, that still corresponds to ten times Earth’s water content (by mass), meaning that this planet is likely to be covered in global oceans. The simulation from Fig. \[fig:aet\_giants\] showed the simple case of a single giant planet on a low-eccentricity ($e \approx 0.05$) migrating through a disk of growing planetesimals and embryos. Migration would be more destructive to planet formation under certain circumstances. For example, if migration occurs very late in the evolution of the disk then less gas remains to damp the eccentricities of scattered bodies. This is probably more of an issue for the formation of hot Super Earths than for scattered embryos: since the viscous timescale is shorter closer-in, much of the inner disk may in fact drain onto the star during type 2 migration [@thommes08] and reduce the efficiency of the shepherding mechanism. In addition, multiple giant planets may often migrate inward together. In that case the giant planets’ eccentricities would likely be excited to modest values, and any embryo scattered outward would likely encounter another giant planet, increasing the probability of very strong scattering events onto unbound orbits. Although type 2 migration certainly does not provide a comfortable environment for terrestrial accretion, planet-planet scattering is far more disruptive. The broad eccentricity distribution of observed giant exoplanets is naturally reproduced if at least 75% of the observed planets are the survivors of violent dynamical instabilities [@chatterjee08; @juric08; @raymond10]. It is thought that giant planets form in multiple systems on near-circular orbits but in time, perturbations destabilize these systems and lead to a phase of close gravitational encounters. Repeated planet-planet scattering usually leads to the ejection of one or more giant planets [@rasio96; @weidenschilling96; see chapter by Davies et al]. The large eccentricities of the observed planets are essentially the scars of past instabilities. Instabilities are also destructive for terrestrial planets or their building blocks. The timing of instabilities is poorly-constrained, although it is thought that many instabilities may be triggered by either migration in systems of multiple gas giants [@adams03; @moorhead05] or by the removal of damping during the dissipation of the gaseous disk [@moeckel08; @matsumura10; @moeckel12]. On the other hand, systems of planets on more widely-spaced orbits or systems with wide binary companions may naturally experience instabilities on Gyr timescales [@marzari02; @kaib13]. Although early instabilities may allow for additional sources of damping via gas drag from remaining gas and dynamical friction from abundant planetesimals, in practice the timing of the instability makes little difference for the survival of terrestrial bodies [@raymond12]. Instabilities between Jupiter-sized planets typically only last for $\sim10^5$ years. When a giant planet is scattered onto a highly-eccentric orbit, even if it only lasts for a relatively short time, very strong secular forcing can drive the orbits of inner terrestrial bodies to very high eccentricities. The outcome of the perturbation is extremely sensitive to the proximity of the giant planet to the terrestrial planet zone: giant planets whose pericenter distances come within a given separation act so strongly that terrestrial planets or embryos are driven entirely into the central star [@veras05; @veras06; @raymond11; @raymond12]. The giant planet instabilities that are the least disruptive to the terrestrial planets are those that are very short in duration, that are confined to the outer parts of the planetary system, or that result in a collision between giant planets. The right panel of Figure \[fig:aet\_giants\] shows a simulation in which all terrestrial bodies were removed from the system by an instability between three $\sim$Jupiter-mass giant planets that occurred after 42 Myr. During the first 42 Myr of the simulation, accretion in the inner disk proceeded in the same manner as in Fig. \[fig:sim0\]. Once the instability was triggered after 42.8 Myr, the inner disk of planets – including two planets that had grown to nearly an Earth mass – were driven into the central star. The entire outer disk of planetesimals was ejected by repeated giant planet-planetesimal scattering over the next few Myr [@raymond12]. Instabilities systematically perturb both the terrestrial planet-forming region and outer disks of planetesimals. The dynamics of gas giant planets thus creates a natural correlation between terrestrial planets and outer planetesimal disks. On Gyr timescales planetesimal disks collisionally grind down and produce cold dust that is observable at wavelengths as debris disks [@wyatt08; @krivov10]. On dynamical grounds, [@raymond11; @raymond12] predicted a correlation between debris disks and systems of low-mass planets, as each of these forms naturally in dynamically calm environments, i.e. in systems with giant planets on stable orbits or in systems with no gas giants. FORMATION OF THE SOLAR SYSTEM’S TERRESTRIAL PLANETS ======================================================= A longstanding goal of planet formation studies has been to reproduce the Solar System using numerical simulations. Although that goal has not yet been achieved, substantial progress has been made. Jupiter and Saturn are key players in this story. Their large masses help shape the final stages of terrestrial accretion (§4.5). However, there exist few constraints on their orbits during late-stage terrestrial accretion, and these are model-dependent. The Nice model [e.g., @tsiganis05; @morby07] proposes that the [Late Heavy Bombardment]{} (LHB) – a spike in the impact rate on multiple Solar System bodies that lasted from roughly 400 until 700 Myr after the start of planet formation [@tera74; @cohen00; @chapman07] – was triggered by an instability in the giant planets’ orbits. The instability was triggered by gravitational interactions between the giant planets and a disk of planetesimals exterior to the planets’ orbits comprising perhaps $30-50\mearth$. Before the Nice model instability, the giant planets’ orbits would have been in a more compact configuration, with Jupiter and Saturn interior to the 2:1 resonance and perhaps lodged in 3:2 resonance. Although there is no direct constraint, hydrodynamical simulations indicate that the gas giants’ eccentricities were likely lower than their current values, probably around 0.01-0.02 [@morby07]. An alternate but still self-consistent assumption is that the gas giants were already at their current orbital radii during terrestrial accretion. In that case, Jupiter and Saturn must have had slightly higher eccentricities than their current ones because scattering of embryos during accretion tends to modestly decrease eccentricities [e.g. @chambers02]. In this scenario, an alternate explanation for the LHB is needed. In this section we first consider “classical” models that assume that the orbits of the giant planets were stationary (§6.1). Based on the above arguments we consider two reasonable cases. In the first case, Jupiter and Saturn were trapped in 3:2 mean motion resonance at 5.4 and 7.2 AU with low eccentricities ($e_{giants} \approx 0.01-0.02$). In the second, Jupiter and Saturn were at their current orbital radii but with higher eccentricities ($e_{giants} = 0.07-0.1$). Of course, Jupiter and Saturn’s orbits need not have been stationary at this time. It is well-known that giant planets’ orbits can migrate long distances, inward or outward, driven by exchanges with the gaseous protoplanetary disk [e.g. @lin86; @veras04] or a disk of planetesimals [e.g. @fernandez84; @murray98]. Although the last phases of accretion are constrained by Hf-W measurements of Earth samples to occur after the dissipation of the typical gas disk, giant planet migration at early times can sculpt the population of embryos and thus affect the “initial conditions” for late-stage growth. [l\|ccccccl]{} Resonant Jup, Sat & $\checkmark$ & $\times$ &$\times$ & $\checkmark $ & $\times$ &$\checkmark$ & Consistent with Nice model\ Eccentric Jup, Sat & $\sim$ & $\sim$ &$\checkmark$ & $\checkmark $ & $\checkmark$ &$\times$ & Not consistent with Nice model\ Grand Tack & $\checkmark$ &$\checkmark$ &$\checkmark$ & $\sim$ & $\checkmark$ &$\checkmark$ & Requires tack at 1.5 AU\ Planetesimal-driven & $\checkmark$ & $\times$ & $\times$ & $\checkmark$ & $\times$ &$\checkmark$ & Requires other source of LHB\ migration\ -0.5cm While the Nice model relies on a delayed planetesimal-driven instability, earlier planetesimal-driven migration of the giant planets has recently been invoked [@agnor12]. In §6.2 we consider the effect of this migration, which must occur on a timescale shorter than Mars’ measured few Myr accretion time [@dauphas11] to have an effect. Finally, in §6.3 we describe a new model called the [Grand Tack]{} [@walsh11] that invokes early gas-driven migration of Jupiter and Saturn. It is possible that disks are not radially smooth, or at least that planetesimals do not form in a radially-uniform way [e.g. @johansen07; @chambers10]. [@jin08] proposed that a discontinuity in viscosity regimes at $\sim$2 AU could decrease the local surface density and thus form a small Mars. However, the dip produced is too narrow to cut off Mars’ accretion [@raymond09c]. It has also been known for decades that an embryo distribution with an abrupt radial edge naturally forms large planets within the disk but small planets beyond the edge [@wetherill78]. This “edge effect” can explain the large Earth/Mars mass ratio (see below). Table 2 summarizes the ability of various models to reproduce the observational constraints discussed in §2. 6.1 Classical models with stationary gas giants Fig. \[fig:testp\] shows how the giant planets excite the eccentricities of test particles for each assumption [@raymond09c]. In the left panel (labeled JSRES for “Jupiter and Saturn in RESonance”) the giant planets are in a low-eccentricity compact configuration consistent with the Nice model whereas in the right panel (labeled EEJS for “Extra-Eccentric Jupiter and Saturn”) the giant planets have significant eccentricities and are located at their current orbital radii. The much stronger eccentricity excitation imparted by eccentric gas giants and the presence of strong resonances such as the $\nu_6$ resonance seen at 2.1 AU in the right panel of Fig. \[fig:testp\] have a direct influence on terrestrial planet formation. Simulations with Jupiter and Saturn on circular orbits reproduce several aspects of the terrestrial planets [@wetherill78; @wetherill96; @wetherill85; @chambers98; @morby00; @chambers01; @raymond04; @raymond06b; @raymond07a; @raymond09c; @obrien06; @morishima10]. Simulations typically form about the right number (3-5) of terrestrial planets with masses comparable to their actual masses. Earth analogs tend to complete their accretion on 50-100 Myr timescales, consistent with geochemical constraints. Simulations include late giant impacts between embryos with similar characteristics to the one that is thought to have formed the Moon [@cuk12; @canup12]. Embryos originating at 2.5-4 AU, presumed to be represented by carbonaceous chondrites and therefore to be volatile-rich, naturally deliver water to Earth during accretion (see Fig. \[fig:sim0\]). There are three problems. First and most importantly, simulations with Jupiter and Saturn on circular orbits are unable to form good Mars analogs. Rather, planets at Mars’ orbital distance are an order of magnitude too massive, a situation called the [small Mars problem]{} [@wetherill91; @raymond09c]. Second, the terrestrial planet systems that form tend to be far too spread out radially. Their radial mass concentration $RMC$ (see Eq. 2) are far smaller than the Solar System’s value of 89.9 (see Table 1). Third, large ($\sim$Mars-sized) embryos are often stranded in the asteroid belt. All three of these problems are related: the large $RMC$ in these systems is a consequence of too much mass existing beyond 1 AU. This mass is in the form of large Mars analogs and embryos in the asteroid belt. Simulations starting with Jupiter and Saturn at their current orbital radii but with larger initial eccentricities ($e = 0.07-0.1$) reproduce many of the same terrestrial planet constraints [@raymond09c; @morishima10]. Simulations tend to again form the same number of terrestrial planets with masses comparable to the actual planets’. Moon-forming impacts also occur. Beyond this the accreted planets contrast with those that accrete in simulations with circular gas giants. With eccentric Jupiter and Saturn, the terrestrial planets accrete faster, in modest agreement with Earth’s geochemical constraints. The delivery of water to Earth is much less efficient. But Mars analogs form with about the right mass! In these simulations, a strong secular resonance with Saturn – the $\nu_6$ at 2.1 AU – acts to clear out the material in the inner asteroid belt and in Mars’ vicinity. The resonance is so strong that bodies that are injected into it are driven to very high eccentricities and collide with the Sun within a few Myr [@gladman97]. Any embryo from the inner planetary system that is scattered out near the $\nu_6$ is quickly removed from the system. The Mars region is quickly drained and a small Mars forms. The $\nu_6$ acts as a firm outer edge such that the terrestrial planet systems form in more compact configurations, with $RMC$ values that approach the Solar System’s (but still remain roughly a factor of two too small; see Fig.\[fig:amdrmc\]). The $AMD$ of the terrestrial planets are systematically higher than the Solar System value because the planetesimals that could provide damping at late times are too efficiently depleted. The terrestrial planet forming region is effectively cut off from the asteroid belt by the resonance, and water delivery is inefficient. If the gravitational potential from the dissipating gas disk is accounted for, the $\nu_5$ and $\nu_6$ resonances sweep inward and can perhaps shepherd water-rich embryos in to Earth’s feeding zone by the same mechanism presented in Sec 5.2 [@thommes08; @morishima10]. However, hydrodynamical simulations suggest that Jupiter and Saturn’s eccentricities are unlikely to remain high enough during the gaseous disk phase for this to occur [e.g. @morby07; @pierens11]. The early orbits of Jupiter and Saturn sculpt dramatically different terrestrial planet systems. Systems with gas giants on circular orbits form Mars analogs that are far too large and strand embryos in the asteroid belt. Systems with gas giants on eccentric orbits do not deliver water to Earth and have eccentricities that are too large. To date, no other configuration of Jupiter and Saturn with static orbits has been shown to satisfy all constraints simultaneously. ![Orbital statistics of the terrestrial planet systems formed in different models. The configuration of each system is represented by its angular momentum deficit and radial mass concentration values; see section 2.1 for the definition of these terms. The simulations with eccentric and resonant gas giants are from [@raymond09c], those including planetesimal-driven migration of the gas giants are from [@lykawka13], and the Grand Tack simulations are from [@obrien13]. []{data-label="fig:amdrmc"}](fig8.eps){width="45.00000%"} To quantify the failings of the classical model, Figure \[fig:amdrmc\] shows the angular momentum deficit $AMD$ and radial mass concentration $RMC$ statistics for simulated terrestrial planets under the two assumptions considered here. The accreted planets are far too radially spread out (have small $RMC$ values). In many cases their orbits are also too excited, with larger $AMD$ values than the actual terrestrial planets’. 6.2 Accretion with planetesimal-driven migration of Jupiter and Saturn If Jupiter and Saturn formed in a more compact orbital configuration, then the migration to their current configuration may have perturbed the terrestrial planets, or even the building blocks of the terrestrial planets if their formation was not complete. [@brasser09; @brasser13] and [@agnor12] simulated the influence of planetesimal-driven migration of the giant planets on the terrestrial planets assuming that the migration occurred late, after the terrestrial planets were fully-formed. They found that if Jupiter and Saturn migrated with eccentricities comparable to their present-day values, a smooth migration with an exponential timescale characteristic of planetesimal-driven migration ($\tau \sim$ 5-10 Myr) would have perturbed the eccentricities of the terrestrial planets to values far in excess of the observed ones. To resolve this issue, [@brasser09; @brasser13] suggested a Òjumping JupiterÓ in which encounters between an ice giant and Jupiter caused Jupiter and Saturn’s orbits to spread much faster than if migration were driven solely by encounters with planetesimals [see also @morby10]. On the other hand, [@agnor12] suggested that the bulk of any giant planet migration occurred during accretion of terrestrial planets. Whenever the migration occurred, the degree of eccentricity excitation of Jupiter and Saturn is constrained by the dynamics of resonance crossing. Jupiter and Saturn are naturally excited to $e_{giants} \sim 0.05$ but cannot reach the higher eccentricities invoked by the eccentric Jupiter and Saturn model described above [@tsiganis05]. Given that the eccentricity excitation is the key difference between this model and those with stationary giant planets discussed above, the only free parameter is the timing of the eccentricity excitation. Two recent papers simulated the effect of planetesimal-driven migration of Jupiter and Saturn’s orbits on terrestrial planet formation [@walsh11b; @lykawka13]. In both studies terrestrial planets accrete from a disk of material which stretches from $\sim$0.5 AU to 4.0 AU. In [@walsh11b], Jupiter and Saturn are initially at 5.4 and 8.7 AU respectively (slightly outside the 2:1 mean motion resonance), with eccentricities comparable to the current ones, and migrate to 5.2 and 9.4 AU with an e-folding time of 5 Myr. In their simulations Mars is typically far too massive and the distribution of surviving planetesimals in the asteroid belt is inconsistent with the observed distribution. [@lykawka13] performed similar simulations but included the 2:1 resonance crossing of Jupiter and Saturn, which provides a sharp increase in the giant planets’ eccentricities and thus in the perturbations felt by the terrestrial planets. They tested the timing of the giant planets’ 2:1 resonance crossing between 1 and 50 Myr. They found the expected strong excitation in the asteroid belt once the giant planets’ eccentricities increased, but the perturbations were too small to produce a small Mars. Although they produced four Mars analogs in their simulations, they remained significantly more massive than the real Mars, accreted on far longer timescales than the geochemically-constrained one, and stranded large embryos in the asteroid belt. Their $AMD$ and $RMC$ values remain incompatible with the real Solar System (Fig. \[fig:amdrmc\]). If another mechanism is invoked to explain the late heavy bombardment, planetesimal-driven migration of Jupiter and Saturn is plausible. However, it does not appear likely to have occurred as it is incapable of solving the Mars problem. 6.3 The Grand Tack model Prior to 2009, several studies of terrestrial accretion had demonstrated an edge effect in terrestrial accretion. A distribution of embryos with an abrupt edge naturally produces a large mass gradient between the massive planets that formed within the disk and the smaller planets that were scattered beyond the disk’s edge [@wetherill78; @wetherill91; @chambers98; @agnor99; @chambers01; @kominami04]. These studies had outer edges at 1.5-2 AU and generally considered their initial conditions a deficiency imposed by limited computational resources. [@hansen09] turned the tables by proposing that, rather than a deficiency, initial conditions with edges might actually represent the true initial state of the disk. Indeed, [@morishima08] and [@hansen09] showed that most observed constraints could be reproduced by a disk of embryos spanning only from 0.7 to 1 AU. Earth and Venus are massive because they formed within the annulus whereas Mars and Mercury’s small masses are explained as edge effects, embryos that were scattered exterior and interior, respectively, to the annulus at early times, stranding and starving them. Mars analogs consistently accrete on the short observed timescale. The main unanswered question in these studies was the origin of the edges of the annulus. [@walsh11] presented a mechanism to produce the outer edge of the disk by invoking migration of the giant planets to dramatically sculpt the distribution of solid material in the inner Solar System. Given that gas giant planets must form in the presence of gaseous disks and that these disks invariably drive radial migration [@ward97], it is natural to presume that Jupiter and Saturn must have migrated to some extent. A Jupiter-mass planet naturally carves an annular gap in the gaseous disk and migrates inward on the local viscous timescale [@lin86]. In contrast, a Saturn-mass planet migrates much more quickly because of a strong gravitational feedback during disk clearing [@masset03]. Assuming that Jupiter underwent rapid gas accretion before Saturn, hydrodynamical simulations show that Jupiter would have migrated inward relatively slowly. When Saturn underwent rapid gas accreted it migrated inward quickly, caught up to Jupiter and became trapped in 2:3 resonance. At this point the direction of migration was reversed and Jupiter “tacked”, that is it changed its direction of migration [@masset01; @morby07; @pierens08; @pierens11; @dangelo12]. The outward migration of the two gas giants slowed as the gaseous disk dissipated, stranding Jupiter and Saturn on resonant orbits. This naturally produces the initial conditions for a recently revised version of the Nice model [@morby07; @levison11], with Jupiter at 5.4 AU and Saturn at 7.2 AU. This model is called the [Grand Tack]{}. One cannot know the precise migration history of the gas giants [a priori]{} given uncertainties in disk properties and evolution. [@walsh11] anchored Jupiter’s migration reversal point at 1.5 AU because this truncates the inner disk of embryos and planetesimals at 1.0 AU, creating an outer edge at the same location as invoked by [@hansen09]. Jupiter’s formation zone was assumed to be $\sim3-5$ AU [although a range of values was tested by @walsh11], in the vicinity of the snow line [e.g. @sasselov00; @kornet04; @martin12], presumably a favorable location for giant planet formation. The Grand Tack model also proposes that the compositional gradient seen in the asteroid belt can be explained by the planetesimals’ formation zones. Volatile-poor bodies (“S-class") are primarily located in the inner belt and volatile-rich bodies (“C-class") primarily in the outer belt [@gradie82; @demeo13]. The Grand Tack scenario presumes that S-class bodies formed interior to Jupiter’s initial orbit and that C-class bodies formed exterior. The evolution of the Grand Tack is illustrated in Figure \[fig:aet\_GT\] [@walsh11]. Jupiter and Saturn’s inward migration scattered S-class planetesimals from the inner disk, with $\sim$10% ending on eccentric orbits beyond the giant planets. Meanwhile a large fraction of planetesimals and embryos were shepherded inward by the same mechanism discussed in §5.2 onto orbits inside 1 AU. Following Jupiter’s “tack” the outward-migrating gas giants first encountered the scattered S-class planetesimals, about 1% of which were scattered inward onto stable orbits in the asteroid belt. The giant planets then encountered the disk of C-class planetesimals that originated beyond Jupiter’s orbit. Again, a small fraction ($\sim1\%$) were scattered inward and trapped in the asteroid belt. The final position of a scattered body depends on the orbital radius of the scattering body, in this case Jupiter. Jupiter was closer in when it scattered the S-class planetesimals and farther out when it scattered the C-class planetesimals. The S-class bodies were therefore preferentially implanted in the inner part of the asteroid belt and the C-class bodies preferentially in the outer part of the belt, as in the present-day belt [@gradie82; @demeo13]. The total mass of the asteroid population is set by the need to have $\sim 2 \mearth$ of material remaining in the inner truncated disk of embryos and planetesimals (to form the planets). This requirement for the planets sets the total mass in S-class bodies implanted into the asteroid belt as they originate from the same inner disk. The current ratio of S-class to C-class asteroids sets the mass in outer disk planetesimals. The Grand Tack model reproduces many aspects of the terrestrial planets. Planets that accrete from a truncated disk have similar properties to those in [@hansen09] and [@morishima08]. Earth/Mars mass ratios are close matches to the actual planets, and Mars’ accretion timescale is a good match to Hf/W constraints. Figure \[fig:amdrmc\] shows that the angular momentum deficit $AMD$ is systematically lower than in simulations of the classical model (§6.1) and the radial mass concentration $RMC$ is systematically higher [@walsh11; @obrien13]. In contrast with other models, the Grand Tack simulations provide a reasonable match to the inner Solar System. The Grand Tack delivers water-rich material to the terrestrial planets by a novel mechanism. As Jupiter and Saturn migrate outward, they scatter about 1% of the C-class asteroids that they encountered onto stable orbits in the asteroid belt. And for every implanted C-type asteroid, 10-20 C-class bodies are scattered onto [unstable]{} orbits that cross the orbits of the terrestrial planets. These scattered C-class planetesimals accrete with the growing terrestrial planets and naturally deliver water. The amount of water-rich material accreted by Earth is less than in classical simulations with stationary giant planets like the one presented in Fig. \[fig:sim0\], but is still significantly larger than the Earth’s current water budget [@obrien13]. The chemical signature of the delivered water is the same as C-type asteroids (and therefore carbonaceous chondrites), and thus provides a match to the signature of Earth’s water [@marty06]. Thus, in the Grand Tack model Earth was delivered water not by C-type asteroids but by the same parent population as for C-type asteroids. There remain some issues with the Grand Tack model. The accretion timescales are much faster for all of the planets than what was typically found in previous models. This is a consequence of the removal of embryos beyond 1 AU, where growth timescales are long. In simulations Mars analogs typically form in less than 10 Myr [@obrien13]. Earth analogs form in 10-20 Myr, with giant embryo-embryo impacts occurring after 20 Myr in only a modest fraction ($\sim20\%$) of simulations. This is roughly a factor of two faster than the Hf-W constraints [@touboul07; @kleine09; @konig11]. However, new simulations show that the accretion timescale of the terrestrial planets can be lengthened to match observations simply by increasing the total embryo-to-planetesimal mass ratio in the annulus, which is itself an unconstrained parameter [@jacobson13]. An open question related to the origin of Mercury’s small mass is the origin of the [inner]{} edge of the annulus proposed by [@hansen09]. One possibility is that, as embryos grow larger from the inside-out, they also become subject to type 1 migration from the inside-out [@mcneil05; @daisaka06; @ida08]. For embryo-mass objects migration is directed inward [@paardekooper11], so as each embryo forms it migrates inward. If, by some process, inward-migrating planets are removed from the system (presumably by colliding with the star), then an inner edge in the distribution of [surviving]{} embryos could correspond to the outermost orbital radius at which an embryo formed and was destroyed. Another possibility is that planetesimals could only form in a narrow annulus. If a pressure bump [@johansen09] were located in that region it could act to concentrate small particles [@haghighipour03; @youdin04] and efficiently form planetesimals (see chapter by Johansen et al.). DISCUSSION ============== 7.1 Terrestrial planets vs. giant embryos We think that Earth formed via successive collisions between planetesimals and planetary embryos, including a protracted stage of giant impacts between embryos. But does the formation of most terrestrial planets follow the same blueprint as Earth? The alternative is that terrestrial exoplanets are essentially giant planetary embryos. They form from planetesimals or pebbles and do not undergo a phase of giant impacts after the dissipation of the gaseous disk. This is a wholly reasonable possibility. Imagine a disk that only forms planetesimals in a few preferred locations, perhaps at pressure bumps. The planetesimals in each location could be efficiently swept up into a single large embryo, perhaps by the largest planetesimal undergoing a rapid burst of super-runaway pebble accretion. Isolated giant embryos would evolve with no direct contact with other embryos. Only if several embryos formed and migrated toward a common location would embryo-embryo interactions become important, and collisions would only occur if a critical number of embryos was present [the critical number is about 5; @morby08; @pierens13]. Terrestrial planets and giant embryos should differ in terms of their accretion timescales, their atmospheres, and perhaps their geological evolution. The timescale for the completion of Earth’s accretion is at least ten times longer than the typical gas disk lifetime (see §2). Giant embryos must form within the lifetime of the gaseous disk, while the mechanisms to efficiently concentrate are active. How would Earth be different if it had accreted ten times faster? The additional heat of formation and from trapped short-lived radionuclides could act to rapidly devolatilize the giant embryo’s interior. However, giant embryos may be able to gravitationally capture thick envelopes of gas from the disk, at least at cooler locations within the disk [@ikoma12]. The fate of giant embryos’ volatiles remain unstudied. Nonetheless, given that only a very small amount of H and He are needed to significantly inflate a planet’s radius [@fortney07], giant embryos would likely have low bulk densities. Many low-density planets have indeed been discovered [@marcy13; @weiss13], although we stress that this does not indicate that these are giant embryos. How could we tell observationally whether late phases of giant impacts are common? Perhaps the simplest approach would be to search for signatures of such impacts around stars that no longer harbor gaseous disks. The evolution of warm dust, detected as excess emission at mid-infrared wavelengths, has recently been measured to decline on 100 Myr timescales [@meyer08; @carpenter09; @melis10]. This dust is thought to trace the terrestrial planet-forming region [@kenyon04] and indicates the presence of planetesimals or other large dust-producing bodies in that region. In some cases the signature of specific minerals in the dust can indicate that it originated in a larger body. In fact, the signature of a giant impact was reported by [@lisse09] around the $\sim$12 Myr-old A star HD 172555. Given the 1-10 Myr interval between giant impacts in accretion simulations and the short lifetime of dust produced [@kenyon05; @melis12], a direct measure of the frequency of systems in which giant impacts occur will require a large sample of young stars surveyed at mid-infrared wavelengths [e.g., @kennedy12]. 7.2 Limitations of the simulations Despite marked advances in the last few years, simulations of terrestrial planet formation remain both computationally and physically limited. Even the best numerical integrators [@chambers99; @duncan98; @stadel01] can follow the orbits of at most a few thousand particles at $\sim 1$ AU for the $>$100 Myr timescales of terrestrial planet formation. There are 3-4 orders of magnitude in uncertainty in the sizes of initial planetesimals, and a corresponding 9-12 orders of magnitude uncertainty in the initial number of planetesimals. It is clear that current simulations cannot fully simulate the conditions of planet formation except in very constrained settings [e.g. @barnes09]. Simulations thus resort to including planetesimals that are far more massive than they should be. There exist several processes thought to be important in planet formation that have yet to be adequately modeled. For example, the full evolution of a planetesimal swarm including growth, dynamical excitation, and collisional grinding has yet to be fully simulated [but see @bromley11; @levison12]. In addition to the numerical and computational challenges, this task is complicated by the fact that the initial distribution and sizes of planetesimals, pebbles and dust remain at best modestly-constrained by models and observations (see chapters by Johansen et al. and Testi et al). Likewise, the masses, structure and evolution of the dominant, gaseous components of protoplanetary disks is an issue of ongoing study. SUMMARY =========== This chapter has flown over a broad swath of the landscape of terrestrial planet formation. We now summarize the take home messages. 1\. The term “terrestrial planet” is well-defined in the confines of our Solar System but not in extra-solar planetary systems (§1). 2\. There exist ample observed and measured constraints on terrestrial planet formation models in the Solar System and in extra-solar planetary systems (§2). 3\. There exist two models for the growth of planetary embryos (§3). Oligarchic growth proposes that embryos grow from swarms of planetesimals. Pebble accretion proposes that they grow directly from cm-sized pebbles. 4\. Starting from systems of embryos and planetesimals, the main factors determining the outcome of terrestrial accretion have been determined by simulations (§4). The most important one is the level of eccentricity excitation of the embryo swarm – by both gravitational self-stirring and perturbations from giant planets – as this determines the number, masses and feeding zones of terrestrial planets. 5\. The observed systems of hot Super Earths probably formed by either in-situ accretion from massive disks or inward migration of embryos driven by interactions with the gaseous disk (§5.1). The key debate in differentiating between the models is whether rocky planets that accrete in situ could retain thick gaseous envelopes. 6\. The dynamical histories of giant exoplanets are thought to include gas-driven migration and planet-planet scattering. An inward-migrating gas giant forms low-mass planets interior to strong resonances and stimulates the formation of outer volatile-rich planets. Dynamical instabilities among giant planets can destroy terrestrial planets or their building blocks, and this naturally produces a correlation between debris disks and terrestrial planets (§5.2). 7\. Historical simulations of terrestrial planet formation could not reproduce Mars’ small mass (§6.1). This is called the [small Mars problem]{}. Simulations can reproduce Mars’ small mass by invoking large initial eccentricities of Jupiter and Saturn at their current orbital radii. Invoking early planetesimal-driven migration of Jupiter and Saturn does not produce a small Mars (§6.2). 8\. The [Grand Tack]{} model proposes that Jupiter migrated inward to 1.5 AU then back outward due to disk torques before and after Saturn’s formation (§6.3). The inner disk was truncated at 1 AU, producing a large Earth/Mars mass ratio. Water was delivered to the terrestrial planets in the form of C-class bodies scattered inward during the gas giants’ outward migration. Finally, it remains unclear whether most systems of terrestrial planets undergo phases of giant collisions between embryos during their formation (§7.1). Acknowledgments. We are grateful to a long list of colleagues whose work contributed to this chapter. S.N.R. thanks NASA’s Astrobiology Institute for their funding via the University of Washington, University of Colorado, and Virtual Planetary Laboratory lead teams. S.N.R. and A.M. thank the CNRS’s Programme Nationale de Planetologie. KJW acknowledges support from NLSI CLOE. [273]{} \[1\][\#1]{} I. et al. (1976) Progress of Theoretical Physics, 56, 1756. E. R. et al. (2008) *, 673, 1160. F. C. and [Laughlin]{} G. (2003) Icarus, 163, 290. C. and [Asphaug]{} E. (2004) , 613, L157. C. B. and [Lin]{} D. N. C. (2012) , 745, 143. C. B. et al. (1999) , 142, 219. S. G. and [Agnor]{} C. B. (1998) , 132, 113. S. M. and [Williams]{} J. P. (2007) , 671, 1800. S. M. and [Williams]{} J. P. (2007) , 659, 705. P. J. (2007) , 665, 1381. E. (2010) Chemie der Erde / Geochemistry, 70, 199. E. et al. (2006) , 439, 155. I. et al. (2004) , 419, L13. I. et al. (2006) , 450, 1221. R. et al. (2009) , 203, 626. N. M. et al. (2013) , 204, 24. K. and [Morbidelli]{} A. (2013) , 145, 1. C. and [Nesvorn[ý]{}]{} D. (2012) , 751, 119. W. et al. (2007) , 132, 189. J. C. et al. (2010) , 205, 321. X. et al. (2013) , 549, A109. W. F. et al. (2010) Science, 330, 1527. R. et al. (2009) , 507, 1053. R. et al. (2013) . B. C. and [Kenyon]{} S. J. (2011) , 731, 101. L. A. et al. (2012) , 486, 375. R. M. (2012) Science, 338, 1052. J. M. et al. (2009) , 181, 197. J. C. et al. (2012) , 760, 44. J. (2006) Icarus, 180, 496. J. (2008) , 198, 256. J. E. (1999) , 304, 793. J. E. (2001) Icarus, 152, 205. J. E. (2010) , 208, 505. J. E. (2013) , 224, 43. J. E. and [Cassen]{} P. (2002) Meteoritics and Planetary Science, 37, 1523. J. E. and [Wetherill]{} G. W. (1998) , 136, 304. J. E. et al. (1996) Icarus, 119, 261. C. R. et al. (2007) , 189, 233. S. et al. (2008) , 686, 580. E. and [Laughlin]{} G. (2013) , 431, 3444. B. A. et al. (2000) Science, 290, 1754. A. C. M. and [Laskar]{} J. (2009) , 201, 1. C. et al. (2014) A&A, in press, arXiv:1407.6011. P. and [Nelson]{} R. P. (2008) , 482, 677. P. et al. (2007) , 473, 329. M. and [Stewart]{} S. T. (2012) Science, 338, 1047. A. et al. (2008) , 120, 531. J. K. et al. (2006) , 185, 492. G. and [Marzari]{} F. (2012) , 757, 50. N. and [Pourmand]{} A. (2011) , 473, 489. S. S. (2005) , 627, L153. J. M. D. et al. (2007) Science, 315, 217. F. E. and [Carry]{} B. (2013) , 226, 723. S. J. (2007) , 671, 878. L. and [Tremaine]{} S. (1993) , 103, 67. M. J. and [Righter]{} K. (2002) , 416, 39. M. J. et al. (1998) , 116, 2067. J. A. (2012) , 755, 23. J. A. et al. (2008) , 683, 304. S. et al. (2012) , 221, 859. N. V. et al. (2007) , 472, 329. D. and [Tremaine]{} S. (2007) , 669, 1298. J. and [Margot]{} J.-L. (2013) , 767, 115. D. et al. (2010) , 510, A72. J. A. and [Ip]{} W. (1984) Icarus, 58, 109. D. A. and [Valenti]{} J. (2005) , 622, 1102. M. J. and [Nelson]{} R. P. (2005) , 441, 791. M. J. and [Nelson]{} R. P. (2007) , 461, 1195. M. J. and [Nelson]{} R. P. (2009) , 498, 575. E. B. and [Rasio]{} F. A. (2006) , 638, L45. J. J. et al. (2007) , 659, 1661. F. et al. (2013) , 766, 81. E. et al. (2007) Science, 318, 210. J. and [Gloeckler]{} G. (1998) , 84, 239. H. and [Abe]{} Y. (2005) , 433, 842. H. et al. (2012) , 744, 137. L. et al. (2010) , 720, 1290. B. J. et al. (1997) Science, 277, 197. P. and [Tremaine]{} S. (1980) , 241, 425. P. et al. (2004) , 614, 497. G. (1997) , 285, 403. A. et al. (2010) , 720, 1073. J. and [Tedesco]{} E. (1982) Science, 216, 1405. J. S. et al. (2006) , 366, 283. R. et al. (1978) , 35, 1. R. E. and [McSween]{} H. Y. (1993) Science, 259, 653. N. and [Boss]{} A. P. (2003) , 583, 996. Jr. K. E. et al. (2001) , 553, L153. B. M. S. (2009) , 703, 1131. B. M. S. and [Murray]{} N. (2012) , 751, 158. B. M. S. and [Murray]{} N. (2013) , 775, 53. P. et al. (2011) , 478, 218. C. et al. (1985) in: Protostars and Planets II, (edited by D. C. [Black]{} and M. S. [Matthews]{}), pp. 1100–1153. L. A. et al. (2008) , 677, 630. A. W. et al. (2010) Science, 330, 653. A. W. et al. (2012) , 201, 15. W. B. et al. (2007) , 658, L59. W. B. et al. (2007) , 187, 358. S. and [Lin]{} D. N. C. (2008) , 673, 487. S. and [Makino]{} J. (1992) , 96, 107. S. and [Makino]{} J. (1993) , 106, 210. S. et al. (2008) , 686, 1292. M. and [Hori]{} Y. (2012) , 753, 66. B. et al. (2010) , 407, 910. S. et al. (2013) in: Protostars and Planets VI, Heidelberg, July 15-20, 2013. Poster \#2H024, p. 24. L. et al. (2008) , 674, L105. A. and [Lacerda]{} P. (2010) , 404, 475. A. et al. (2006) , 636, 1121. A. et al. (2007) , 448, 1022. A. et al. (2009) , 697, 1269. M. and [Tremaine]{} S. (2008) , 686, 603. N. A. et al. (2013) , 493, 381. G. M. and [Wyatt]{} M. C. (2012) , 426, 91. S. J. and [Bromley]{} B. C. (2004) , 602, L133. S. J. and [Bromley]{} B. C. (2005) , 130, 269. S. J. and [Bromley]{} B. C. (2006) , 131, 1837. K. et al. (1974) , 38, 683. T. et al. (2002) , 418, 952. T. et al. (2009) , 73, 5150. H. and [Dauphas]{} N. (2013) , 225, 122. H. and [Tanaka]{} H. (2010) , 206, 735. H. and [Tanaka]{} H. (2010) , 206, 735. E. and [Genda]{} H. (2010) , 714, L21. E. and [Ida]{} S. (1995) , 114, 247. E. and [Ida]{} S. (1996) , 123, 180. E. and [Ida]{} S. (1998) , 131, 171. E. and [Ida]{} S. (2000) , 143, 15. E. and [Ida]{} S. (2002) , 581, 666. E. and [Ida]{} S. (2007) , 671, 2082. E. et al. (2006) , 642, 1131. J. and [Ida]{} S. (2004) , 167, 231. S. et al. (2011) , 75, 2119. K. et al. (2004) , 417, 151. K. A. and [Lin]{} D. N. C. (2012) , 755, 74. A. V. (2010) Research in Astronomy and Astrophysics, 10, 383. M. and [Johansen]{} A. (2012) , 544, A32. H. et al. (2003) , 598, L121. J. (1997) , 317, L75. C. et al. (2003) , 125, 2664. C. et al. (1998) Chem. Geol., 145, 249. A. et al. (2009) , 506, 287. Z. M. and [Stewart]{} S. T. (2012) , 745, 79. H. F. and [Agnor]{} C. (2003) , 125, 2692. H. F. et al. (2011) , 142, 152. H. F. et al. (2012) , 144, 119. D. N. C. and [Papaloizou]{} J. (1986) , 309, 846. D. N. C. et al. (1996) , 380, 606. J. J. (1987) , 69, 249. J. J. and [Kary]{} D. M. (1991) , 94, 126. J. J. and [Stevenson]{} D. J. (2007) Protostars and Planets V, pp. 591–606. J. J. et al. (2011) , 470, 53. C. M. et al. (2008) , 673, 1106. C. M. et al. (2009) , 701, 2019. K. (2003) , 591, 1220. D. et al. (2007) , 462, 211. L. W. et al. (2003) , 592, 255. E. D. and [Fortney]{} J. J. (2014) , 792, 1. C. et al. (2011) , 528, A112. P. S. and [Ito]{} T. (2013) , 773, 65. J. et al. (1998) , 3, 411. A. M. et al. (2007) , 660, 823. A. W. et al. (2012) , 753, 90. G. W. et al. (2013) Submitted. R. G. and [Livio]{} M. (2012) , 425, L6. B. (2012) Earth and Planetary Science Letters, 313, 56. B. and [Yokochi]{} R. (2006) Rev. Mineral Geophys., 62, 421. F. and [Weidenschilling]{} S. J. (2002) Icarus, 156, 570. F. and [Snellgrove]{} M. (2001) , 320, L55. F. S. and [Casoli]{} J. (2010) , 723, 1393. F. S. and [Papaloizou]{} J. C. B. (2003) , 588, 494. F. S. et al. (2006) , 642, 478. S. et al. (2010) , 714, 194. M. et al. (2009) , 507, 487. M. et al. (2011) arXiv:1109.2497. D. et al. (2005) , 130, 2884. C. et al. (2010) , 717, L57. C. et al. (2012) , 487, 74. M. R. et al. (2008) , 673, L181. N. and [Armitage]{} P. J. (2012) , 419, 366. N. et al. (2008) , 688, 1361. A. V. and [Adams]{} F. C. (2005) , 178, 517. A. et al. (2000) Meteoritics and Planetary Science, 35, 1309. A. et al. (2007) , 134, 1790. A. et al. (2008) , 478, 929. A. et al. (2010) , 140, 1391. A. et al. (2012) Annual Review of Earth and Planetary Sciences, 40, 251. R. et al. (2008) , 685, 1247. R. et al. (2010) Icarus, 207, 517. R. et al. (2013) Earth and Planetary Science Letters, 366, 6. A. et al. (2007) , 658, 1312. L. G. et al. (2000) Protostars and Planets IV, pp. 355–+. N. et al. (1998) Science, 279, 69. R. et al. (2011) in: Extreme Solar Systems, p. 804. R. A. et al. (2009) , 693, 23. M. et al. (2008) , 678, 498. S. et al. (2011) , 473, 187. F. and [Agnor]{} C. B. (2006) Earth and Planetary Science Letters, 243, 26. F. and [Kleine]{} T. (2007) , 191, 497. D. P. et al. (2006) , 184, 39. D. P. et al. (2014) , 239, 74. M. and [Ida]{} S. (2009) , 699, 824. C. W. and [Klahr]{} H. H. (2010) , 520, A43. C. W. et al. (2010) , 714, L103. S.-J. et al. (2010) , 401, 1950. I. et al. (2009) , 696, 143. L. M. et al. (2012) , 760, L17. A. and [Nelson]{} R. P. (2008) , 482, 333. A. and [Raymond]{} S. N. (2011) , 533, A131. A. et al. (2013) , 558, A105. R. R. (2004) , 128, 1348. F. A. and [Ford]{} E. B. (1996) Science, 274, 954. S. N. (2006) , 643, L131. S. N. and [Cossou]{} C. (2014) , 440, L11. S. N. et al. (2004) Icarus, 168, 1. S. N. et al. (2005) , 632, 670. S. N. et al. (2006) Science, 313, 1413. S. N. et al. (2006) , 183, 265. S. N. et al. (2007) , 669, 606. S. N. et al. (2007) Astrobiology, 7, 66. S. N. et al. (2008) , 384, 663. S. N. et al. (2009) Icarus, 203, 644. S. N. et al. (2010) , 711, 772. S. N. et al. (2011) , 530, A62. S. N. et al. (2012) , 541, A11. S. N. et al. (2013) , 226, 671. J. et al. (2006) , 446, 211. V. S. (1969) [Evoliutsiia doplanetnogo oblaka.]{}* N. C. et al. (2001) *, 373, 1019. D. D. and [Lecar]{} M. (2000) , 528, 995. H. E. et al. (2012) , 752, 8. A. et al. (2006) , 645, 1498. J. G. (2001) [Cosmological N-body simulations and their analysis]{}, Ph.D. thesis, University of Washington. S. T. and [Leinhardt]{} Z. M. (2012) , 751, 32. T. et al. (2011) , 473, 349. H. and [Ward]{} W. R. (2004) , 602, 388. F. et al. (1974) Earth and Planetary Science Letters, 22, 1. C. and [Papaloizou]{} J. C. B. (2007) , 654, 1110. E. W. et al. (2003) Icarus, 161, 431. E. W. et al. (2008) Science, 321, 814. M. et al. (2007) , 450, 1206. K. et al. (2005) , 435, 459. S. et al. (2007) , 469, L43. D. et al. (2007) , 665, 1413. D. and [Armitage]{} P. J. (2004) , 347, 613. D. and [Armitage]{} P. J. (2005) , 620, L111. D. and [Armitage]{} P. J. (2006) , 645, 1509. R. J. (2009) Chemie der Erde / Geochemistry, 69, 101. K. J. and [Morbidelli]{} A. (2011) , 526, A126. K. J. et al. (2011) , 475, 206. W. R. (1986) , 67, 164. W. R. (1997) , 126, 261. S. J. (1977) , 51, 153. S. J. and [Marzari]{} F. (1996) , 384, 619. L. M. and [Marcy]{} G. W. (2014) , 783, L6. L. M. et al. (2013) , 768, 14. G. W. (1978) in: IAU Colloq. 52: Protostars and Planets, (edited by T. [Gehrels]{}), pp. 565–598. G. W. (1985) Science, 228, 877. G. W. (1991) in: Lunar and Planetary Institute Science Conference Abstracts, vol. 22 of Lunar and Planetary Inst. Technical Report, p. 1495. G. W. (1996) , 119, 219. G. W. and [Stewart]{} G. R. (1989) , 77, 330. G. W. and [Stewart]{} G. R. (1993) , 106, 190. J. P. and [Cieza]{} L. A. (2011) , 49, 67. D. J. et al. (2005) , 626, L109. J. T. et al. (2008) , 683, L63. M. C. (2008) , 46, 339. C. et al. (2009) , 705, 54. R. V. (2004) , 170, 167. Q. et al. (2002) , 418, 949. K. et al. (1999) , 139, 328. A. N. and [Chiang]{} E. I. (2004) , 601, 1109. A. N. and [Goodman]{} J. (2005) , 620, 459. J.-L. et al. (2005) , 631*, L85.	{ "pile_set_name": "ArXiv" }
--- abstract: 'A reliable single photon source is a prerequisite for linear optical quantum computation and for secure quantum key distribution. A criterion yielding a conclusive test of the single photon character of a given source, attainable with realistic detectors, is therefore highly desirable. In the context of heralded single photon sources, such a criterion should be sensitive to the effects of higher photon number contributions, and to vacuum introduced through optical losses, which tend to degrade source performance. In this paper we present, theoretically and experimentally, a criterion meeting the above requirements.' author: - 'Alfred B. U’Ren$^{1,2}$' - Christine Silberhorn$^1$ - 'Jonathan L. Ball$^1$' - Konrad Banaszek$^1$ - 'Ian A. Walmsley$^1$' title: 'Characterization of the non-classical nature of conditionally prepared single photons' --- High fidelity single photon sources are an essential ingredient for quantum-enhanced technologies including linear optical quantum computation (LOQC) and secure quantum key distribution. Thus, the endeavor to generate single photons in controlled, well-defined spatio-temporal modes is an active area of research. Current single photon source candidates can be classified into two categories: deterministic sources producing single photons on demand at predefined trigger times and heralded single photon sources relying on the spontaneous emission of distinguishable photon pairs in conjunction with conditional preparation. While the emission times for conditional single photon sources cannot be controlled beyond the restriction of emission time slots through a pulsed pump, it has been shown that waveguided PDC can yield heralded single photons in well defined modes together with high collection efficiencies[@uren04]. Conditional state preparation has been utilized in various physical systems including atomic cascades [@grangier86], ensembles of cold atoms [@chou04] and in parametric downconversion (PDC). In the case of PDC, conditional preparation was first reported by Mandel et al.[@hong86] and since then has been optimized to generate approximately true $n=1$ Fock states [@uren04; @rarity87; @kwiat91; @lvovsky01; @alibart04; @pittman04]. In order to assess the performance of heralded single photon sources a criterion that takes into account the detrimental contributions of higher photon numbers and optical losses is needed. In addition, such a criterion should take into full consideration limitations of existing photodetectors such as the binary behavior of avalanche photodiodes operated in the Geiger mode where a single click signifies the detection of one or more photons. In this paper we derive such a criterion and show that our previously reported waveguided PDC source[@uren04] represents a high fidelity source of heralded single photons. A standard approach used to determine whether a light source exhibits classical or quantum photon statistics is the measurement of a $g^{(2)}(\tau)$ second-order intensity autocorrelation function in a Hanbury-Brown Twiss geometry. The semi-classical theory of photodetection predicts, firstly, that $g^{(2)}(0) \geq g^{(2)}(\tau)$ for all time delays $\tau$, and, secondly, that $g^{(2)}(0) \geq 1$. The observation of photon anti-bunching, i.e. $g^{(2)}(0) \leq g^{(2)}(\tau)$, has been utilized, for example, to verify the non-classical character of deterministic single photon sources implemented by strongly coupled atom cavity systems [@mcKeever04]. For PDC sources, the probability of generating simultaneously two photon pairs at a given instant of time is of the same order as the probability of generating two independent pairs separated by the interval $\tau$. This obliterates the effect of antibunching, unless we employ selective heralding that identifies specifically a single-pair component. For PDC sources the non-classical character of the generated radiation is usually tested by violating the lower bound on the second-order intensity autocorrelation function $g^{(2)}(0) \geq 1$. The value of $g^{(2)}(0)$ constitutes a figure of merit which determines the degree to which higher photon number contributions degrade the single photon character [@alibart04]. Based on a classical wave description and intensity measurements, Grangier et al. derived from the Cauchy Schwarz inequality a similar “anti-correlation” criterion for characterizing conditionally prepared single photons by coincidence detection rates [@grangier86]. For the experimental configuration shown in Fig. \[Fi:BBineqSchematic\] an anti-correlation parameter: $$\alpha=\frac{R_1 R_{123} }{R_{12} R_{13}}$$ can be defined, which indicates non-classical photon statistics for $\alpha <1$, where $R_i$ represents the singles count rates at detector $i$, and $R_{ij}$, $R_{ijk}$ the double and triple coincidences for the respective detectors $i,j,k$. A variant of the $g^{(2)}(0)$ measurement specifically designed to study conditional single photon sources independently from losses, which has been pioneered by Clauser [@Clauser74], has recently been implemented for single photons generated from an ensemble of cold atoms [@chou04]. In the above works, the theoretical modeling of experimental data was carried out in terms of intensity correlation functions. In a typical experiment, however, the count rates are directly related to light intensities only under certain auxiliary assumptions. The reason for this is that standard photodetectors sensitive to single photons, such as avalanche photodiodes operated in the Geiger regime, do not resolve multiphoton absorption events and yield only a binary response telling us whether at least one photon was present in the detected mode or none at all. With such detectors, the light intensity can be read out from the count rates only in the limit of weak fields, where the probability of detecting a single photon is proportional to the intensity. In a general case, the probability of obtaining a click is a nonlinear function of the incident intensity. This aspect is particularly important in schemes utilizing ultrashort pulses, where the incoming light energy is concentrated in sub-picosecond time intervals that cannot be resolved even by the fastest photodetectors. It is therefore interesting to go beyond the basic intensity correlation theory and examine whether count statistics collected with binary non-photon-resolving detectors can serve as a test of source non-classicality. We will demonstrate in the following that this is indeed the case. Furthermore, the non-classicality criterion based on measuring $g^{(2)}(0)$ relies on a coincidence basis measurement so that losses can be neglected. However, for applications such as cascaded logic gates in LOQC[@knill01] and loophole free tests of Bell inequalities[@kwiat95] post-selection is not desirable, as it leads to vacuum contamination. The latter diminishes the usability of the single photon states: heralding no longer necessarily corresponds to the successful generation of a single-photon or LOQC gate operation. In this paper we derive a criterion designed to test the non-classical nature of conditionally prepared single photon states. Our criterion takes into account both, the non-linearity of the detectors and the fidelity of the generated single-photon state, which measures the probability that a single-photon is actually present when it is heralded. The criterion can be tested in a standard setup in which the signal field is subdivided into two submodes, each monitored by a non-photon number resolving detector. Consider a source emitting two light beams whose intensities, integrated over the detector active area, are $W_A$ and $W_B$. In the semiclassical theory of photodetection we will treat these intensities as positive-definite stochastic variables described by a joint probability distribution ${\cal P}(W_A;W_B)$. Beam $B$ is divided by a beam splitter with power reflection and transmission coefficients $r$ and $t$. Finally, the resulting beams are detected by three photodetectors. We will assume that the probability of obtaining a click on the $i$th detector illuminated by intensity $W$ is given by $p_i(W)$, bounded between $0$ and $1$. We furthermore assume $p_i(W)$ to be a monotonic increasing function of its argument $W$. Under these assumptions it is easy to show that the following inequality is satisfied for an arbitrary pair of arguments $W_B$ and $W'_B$: $$[p_2(rW_B) - p_2(rW'_B)][p_3(tW_B) - p_3(tW_B')] \ge 0$$ Indeed, the sign of both the factors in square brackets is always the same, depending on the sign of the difference $W_B-W'_B$; their product is therefore never negative. Let us now multiply both sides of the above inequality by the factor ${\cal P}(W_A;W_B){\cal P}(W_A';W_B')p_1(W_A)p_1(W'_A)$ which is likewise nonegative, and perform a double integral $\int_0^\infty dW_A dW_B \int_0^\infty dW_A' dW_B'$. This yields the inequality: $$\label{R1R123-R12E13}B= R_1 R_{123} - R_{12} R_{13} \ge 0$$ where the single, double, and triple count rates are given by averages $\langle \ldots \rangle = \int_0^\infty dW_A dW_B {\cal P}(W_A;W_B) \ldots$ defined with respect to the probability distribution ${\cal P}(W_A;W_B)$: $$\begin{aligned} R_1 & = & \langle p_1(W_A) \rangle \\ R_{12} & = & \langle p_1(W_A) p_2(rW_B) \rangle \\ R_{13} & = & \langle p_1(W_A) p_3(tW_B) \rangle \\ R_{123} & = & \langle p_1(W_A) p_2(rW_B) p_3(tW_B) \rangle\end{aligned}$$ It is seen that the inequality derived in Eq. (\[R1R123-R12E13\]) which can be transformed into: $$\frac{R_1 R_{123}}{R_{12} R_{13}} \ge 1$$ has formally the same structure as the condition derived by Grangier [et al.]{}[@grangier86]. However, the meaning of the count rates is different, as we have incorporated the binary response of realistic detectors. It is noteworthy that this inequality has been derived with a very general model of a detector, assuming essentially only a monotonic response with increasing light intensity. Our experimental apparatus is similar to that reported in Ref. [@uren04]. PDC is generated by a KTP nonlinear waveguide pumped by femtosecond pulses from a modelocked, frequency doubled 87MHz repetition rate Ti:sapphire laser. In contrast to that reported in Ref. [@uren04], the approach here is to record time-resolved detection information for the three spatial modes involved with respect to the Ti:sapphire pulse train as detected by a fast photodiode. We thus obtain a reference clock signal with respect to which post-detection event selection can be performed in order to implement temporal gating. The latter is important for the suppression of uncorrelated background photons, the presence of which can lead to heralded vacuum (rather than a true single photon). Through this approach, we are able to freely specify the time-gating characteristics; arbitrarily complicated logic can be performed without added experimental hardware. Drawbacks include the lack of real-time data processing as well as the deadtime in the region of $\mu$s between subsequent triggers exhibited by the digital oscilloscope (LeCroy WavePro 7100) used for data acquisition. In our setup, source brightness information is obtained via a separate NIM electronics-based measurement. For a given trigger event, three numbers are recorded: the time difference between the electronic pulse positive edge corresponding to the trigger and to the two signal modes $t_{S1}$, $t_{S2}$, as well as the trigger-clock reference time difference $t_{CLK}$. Time-gating involves discarding trigger events outside a certain range of $t_{CLK}$ values, while coincidence events with $t_{S1}$ and $t_{S2}$ outside a $1.1$ns wide coincidence window are regarded as accidental and ignored. We collected $75000$ trigger events and measured pre-time gating detection efficiencies (defined as the rate of coincidences normalized by singles) for each of the two signal channels of $14.4\%$ and $13.7\%$. Fig. \[Fig:GrangierData\] shows the post-processed data using a scanned temporal band-pass filter with $300$ps width (selected to approximately match the measured APD jitter). Fig. \[Fig:GrangierData\](A)\[(B)\] shows the time-resolved signal$_1$-trigger \[signal$_2$-trigger\] coincidence count rate, compared to the time-resolved trigger singles count rate. Fig. \[Fig:GrangierData\](C)\[(D)\] shows the resulting time-gated detection efficiency for the signal$_1$ \[signal$_2$\] channel, showing maximum values of $\sim17.4\%$ \[$\sim17.0\%$\]. Fig. \[Fig:GrangierData\](E) shows time-resolved triple coincidences, for identical coincidence windows as used in computing double coincidences. Thus, our time-gating procedure filters the PDC flux so that for the pump-power used the generated light is described essentially by a superposition of vacuum with single photon pairs, showing nearly vanishing multiple pair generation. Fig. \[Fig:GrangierData\](F) shows the time-resolved inequality parameter \[see Eq. \[R1R123-R12E13\]\] resulting from the count rates presented above. As a numerical example, at the peak of the triples counts, we obtain the following time-gated counting rates for $75000$ trigger events: $R_{123}=2$, $R_{12}=5329$, $R_{13}=5067$ and $R_1=30629$, yielding an inequality parameter value of $B=-0.029 \pm .001$. For comparison, our results correspond to a value of the anti-correlation parameter of $\alpha= (2.3 \pm 1.6)\times 10^{-3}$. Fig. \[Fig:GrangierData\] indicates an overall signal transmission \[defined as the sum of the two individual efficiencies $(R_{12}+R_{13})/R_1$\] of $\sim$34.5$\%$. The main contribution to losses is the non-unit quantum efficiency of the single photon detectors. The overall detection efficiency is also degraded due to imperfect optical transmission and remaining unsuppressed uncorrelated photons. From the above count rates, we can also calculate $g^{(2)}(0)=2 p_{(2)}/p_{(1)}^2$ in terms of the probability of observing a single photon in the signal arm $p_{(1)}=(R_{12}+R_{13})/R_1$ and the probability of observing two photons in the signal arm $p_{(2)}=R_{123}/R_1$. We thus obtain $g^{(2)}(0)=(1.1\pm 0.8)\times 10^{-3}$, amongst the lowest reported for a single photon source. Ignoring the spectral and transverse momentum degrees of freedom, the signal and idler photon-number distribution in a realistic PDC source is expressed as: $$\|\Psi\rangle=\sqrt{1-\|\lambda\|^2}\sum\limits_{n=0}^\infty \lambda^n\|n\rangle_s\|n\rangle_i$$ where $n$ represents the photon number describing each of the signal and idler modes and $\lambda$ represents the parametric gain. PDC experiments often operate in a regime where $\lambda$ is small enough that the probability of multiple pair generation becomes negligible. For larger values of $\lambda$ (accessed for example by a higher pump power or higher non-linearity), however, the higher order terms (e.g. $\|2\rangle_s\|2\rangle_i$, $\|3\rangle_s\|3\rangle_i$...) become important. While these higher photon number terms are desirable for conditional preparation via photon number resolving detection, in the context of the present work, where the detectors used are not photon-number resolving and where the emphasis is on high-fidelity preparation of single photons, multiple pair generation must be avoided. As discussed earlier,in order to characterize a source of conditionally prepared single photons based on PDC, besides the parametric gain $\lambda$, optical losses must be taken into account. Losses in the signal arm imply that a trigger detection event can incorrectly indicate the existence of a signal photon, while in reality vacuum is present. Fig. \[Fig:BBtheo\] shows the expected inequality behavior based on a quantum mechanical calculation in which it is assumed that the detection probability is given by the expectation value of the operator $1-\exp(-\eta \hat{W})$ (where $\hat{W}$ is the time-integrated incoming intensity operator and $\eta$ is the corresponding overall transmission including all optical and detection losses). Fig. \[Fig:BBtheo\](A) shows the calculated inequality parameter $B$ for PDC light as a function of the overall signal optical transmission $\eta_s=(R_{12}+R_{13})/R_1$ for a fixed value of the parametric gain $\lambda$. Fig. \[Fig:BBtheo\](B) shows the inequality coefficient as a function of the parametric gain $\lambda$ for different levels of optical loss. Note that a strong violation of the inequality is only observed in the low parametric gain limit coupled with low losses. Note further that the minimum value of $B$, corresponding to the strongest violation and which is only reached in the ideal lossless case, is $-0.25$. In an experimental realization, while accessing very low values of $\lambda$ is straightforward e.g. by using a low pump power, attaining a sufficiently low level of loss to yield a nearly ideal violation is challenging. An analysis of expected detection rates, under the assumption that all uncorrelated photons in the trigger arm are suppressed, yields the parametric gain $\lambda$ in terms of experimentally measurable quantities: $$\lambda^2=\frac{R_2+R_3}{\eta_s R_{rep} (1+f)}$$ where $R_{rep}$ is the pump repetition rate and $f$ is the uncorrelated photon intensity normalized by that of PDC. We estimate that in our experiment $f$ is constrained by: $0<f\lesssim 2$. Our experimental values of $R_2+R_3\approx 70000 s^{-1}$, $R_{rep}=87\times10^6 s^{-1}$ and $\eta_s=0.345$ thus yield: $0.016<\lambda<0.047$. The experimentally observed violation \[see Fig. \[Fig:GrangierData\](F)\] is in good agreement with the theory curves in Fig. \[Fig:BBtheo\]. The black squares in Fig. \[Fig:BBtheo\](A) and (B) depict the observed violation as compared with the theoretical curves, where the uncertainty is smaller than the square dimensions. The plot in Fig. \[Fig:BBtheo\](A) assumes a fixed value of the parametric gain $\lambda$ (with different curves shown for a choice of $\lambda$ values). The signal arm transmission is obtained as the sum of the two individual signal detection efficiencies \[see Fig. \[Fig:GrangierData\](A) and (B)\]. In summary, we have derived a criterion which allows a conclusive test of the single photon character of conditionally prepared single photon states. We have shown that the inequality in Eq. \[R1R123-R12E13\] is fulfilled by all classical light sources, as well as by states generated by PDC exhibiting higher photon numbers through a large parametric gain. On the contrary, a strong violation of the inequality is observed only for states that constitute a good approximation to a conditionally prepared single photon. Our criterion is realistic enough to include binary non-photon number resolving photon counting detectors while it is sensitive to the degradation observed in the prepared state caused by a vacuum component due to losses, crucial for assessing heralded single photon source performance. Through the application of our criterion it is shown that our waveguided PDC source[@uren04] constitutes a high-fidelity conditional single photon source. Our derived inequality yields a new figure of merit quantifying the overall performance of conditional single photon sources taking into full consideration experimental imperfections. [99]{} A.B. U’Ren, Ch. Silberhorn, K. Banaszek, I.A. Walmsley, Phys. Rev. Lett. 93, 093601 (2004) P. Grangier, G. Roger, A. Aspect, Europhys. Lett. 1, 173 (1986) C.W.Chou, S.V. Polyakov, A. Kuzmich, H.J. Kimble, Phys. Rev. Lett. 92, 213601 (2004) C. K. Hong and L. Mandel,Phys. Rev. Lett. 56, 58 (1986) J. McKeever et al., Science 303, 1992 (2004); M. Hennrich, T. Legero, A. Kuhn, and G. Rempe, quant-ph/0406034 (2004) J. G. Rarity, P. R. Tapster, and E. Jakeman, Opt. Comm. [62]{}, 201 (1987). P.G. Kwiat, R.Y. Chiao, Phys. Rev. Lett. 66, 588 (1991) A. I. Lvovsky et al., Phys. Rev. Lett. [87]{}, 050402 (2001). O. Alibart, S. Tanzilli, D. B. Ostrowsky, P. Baldi, quant-ph/0405075 T. B. Pittman, C. C. Jacobs, J. D. Franson, quant-ph/0408093 J. F. Clauser, Phys. Rev. D 9, 853 (1974) E. Knill, R. LaFlamme and G.J. Milburn, Nature 409, 46 (2001); T.C. Ralph, A.G. White, W.J. Munro and G.J. Milburn, Phys. Rev. A 65, 012314 (2001); T.B. Pittman, M.J. Fitch, B.C. Jacobs and J.D. Franson, quant-ph/0303095 (2003); M. Fiorentino and F.N.C. Wong, Phys. Rev. Lett. 93, 070502 (2004). P. G. Kwiat et al. Phys Rev. Lett. 75, 4337 (1995); G. Weihs et al. Phys. Rev. Lett. 81, 5039 (1998). M.G. Roelofs, A. Suna, W. Bindloss and J.D. Bierlein, J. Appl. Phys. 76, 4999 (1994).	{ "pile_set_name": "ArXiv" }
--- abstract: \| This work presents efficient solution techniques for radiative transfer in the smoothed particle hydrodynamics discretization. Two choices that impact efficiency are how the material and radiation energy are coupled, which determines the number of iterations needed to converge the emission source, and how the radiation diffusion equation is solved, which must be done in each iteration. The coupled material and radiation energy equations are solved using an inexact Newton iteration scheme based on nonlinear elimination, which reduces the number of Newton iterations needed to converge within each time step. During each Newton iteration, the radiation diffusion equation is solved using Krylov iterative methods with a multigrid preconditioner, which abstracts and optimizes much of the communication when running in parallel. The code is verified for an infinite medium problem, a one-dimensional Marshak wave, and a two and three-dimensional manufactured problem, and exhibits first-order convergence in time and second-order convergence in space. For these problems, the number of iterations needed to converge the inexact Newton scheme and the diffusion equation are independent of the number of spatial points and the number of processors.\ author: - \| Brody R. Bassett$^{a,\star}$, J. Michael Owen$^{a}$, Thomas A. Brunner$^{a}$\ \ [<[email protected]>, <[email protected]>, <[email protected]>]{}\ \ [$^{a}$Lawrence Livermore National Laboratory]{}\ [7000 East Avenue, Livermore, CA, 94550]{}\ \ [$^{\star}$Corresponding author]{} bibliography: - 'sph\_rad\_hydro\_refs.bib' date: $ $ title: \| Efficient smoothed particle radiation hydrodynamics I:\ Thermal radiative transfer --- Keywords: radiation hydrodynamics, smoothed particle hydrodynamics, radiative transfer, meshless method Introduction ============ Smoothed particle hydrodynamics (SPH) is a meshless approach to solving the hydrodynamics equations, in which the fluid is separated into discrete masses that are used as interpolation points (for an overview, see Refs. [@monaghan2005smoothed; @liu2010smoothed]). SPH has several good properties, such as automatic conservation of mass and enforced conservation of energy and momentum, flexible point topology, and Galilean invariance. Like mesh-based Lagrangian codes, the resolution of the problem follows the mass, but unlike mesh-based codes, SPH does not have issues with mesh tangling. Drawbacks of SPH include a lack of zeroth-order consistency (the interpolant cannot generally reproduce a constant exactly) [@morris1996study], the use of a relatively low-order, unlimited artificial viscosity for shock hydrodynamics [@monaghan1983shock], and additional expense that comes from computing connectivity at each time step. There are extensions to SPH that correct some of these issues, such as moving least squares particle hydrodynamics [@dilts1999moving; @dilts2000moving] and conservative reproducing kernel smoothed particle hydrodynamics [@frontiere2017crksph]. The thermal radiative transfer equations have been solved using a smoothed particle hydrodynamics discretization previously. The most popular method is implicit, two-temperature, flux-limited diffusion, first implemented in Refs. [@whitehouse2004smoothed; @whitehouse2005faster]. These papers assume an ideal gas equation of state. In the latter paper, the system is reduced to solving a quartic equation per point within a Gauss-Seidel iteration scheme. There are other implementations of flux-limited diffusion, most of which have been applied to astrophysical problems [@viau2006implicit; @mayer2007fragmentation]. The form of the second derivative used most often for the radiation diffusion second derivative was first applied to heat diffusion [@brookshaw1985method]. The optically-thin variable Eddington factor equations with Eddington factors determined by source information (as opposed to by a full transport calculation) have been applied to cosmological simulation [@petkova2009implementation]. Similar methods that have been studied include ray-tracing [@altay2008sphray], Monte Carlo [@nayakshin2009dynamic], and neutrino flux-limited diffusion [@fryer2006snsph]. Thermal radiative transfer couples the transport of photons with the hydrodynamic state. Common coupling techniques for the radiation and material equations include simple convergence of the residuals of both equations (including Newton-Krylov methods), using a single Newton iteration, lagging the nonlinear terms in the equation, predictor-corrector schemes, and linearization of the nonlinear emission source [@lowrie2004comparison; @mousseau2000physics]. The method used in this paper is based on nonlinear elimination methods, as described in Ref. [@lanzkron1996analysis] and applied to radiation diffusion in Ref. [@brunner2019nonlinear]. The radiation diffusion and transport equations have also been solved using meshless methods other than SPH, including for coupled radiative transport and conductive heat transfer [@sadat2006use; @liu2007meshless; @kindelan2010application], neutron transport [@bassett2019meshless], and neutron diffusion [@rokrok2012element; @tanbay2013numerical]. Many of these discretizations involve either relatively flat meshless functions (which increases accuracy but makes the system ill-conditioned) or integration of the meshless functions. The advantage of using the SPH discretization directly is that the same functions that are used for hydrodynamics can be reused for the radiation, without recomputing the topology. The goal of this research is to make SPH radiation diffusion more efficient. This is done by improving the solution of the diffusion problem using fast and accurate preconditioners and applying material-radiation coupling methods that speed up the convergence of the radiation emission and absorption. With appropriate time step constraints, the solver is stable and exhibits first-order convergence in time and second-order convergence in space to analytic solutions and manufactured problems. The methods used here scale well with the number of points and under domain decomposition due to the use of fast multigrid solvers [@falgout2002hypre], which also makes the method simple to implement on distributed architectures. The remainder of this paper is structured as follows. In Sec. \[sec:theory\], the SPH thermal radiative transfer equations are derived and discretized in time and space. In Sec. \[sec:methodology\], the implementation of the equations is discussed. Finally, in Sec. \[sec:results\], results for problems in one, two and three dimensions with known solutions are presented to verify the accuracy of the code. In Paper II, the full SPH radiation hydrodynamics equations are derived with appropriate radiation-material coupling terms, and results for full radiation hydrodynamics are presented. Theory\[sec:theory\] ==================== The smoothed particle hydrodynamics method is used to discretize the thermal radiative transfer equations in space. The discretization in time reduces a coupled set of equations for the material and radiation energies to separate updates for the material and radiation energies in an iterative process. The resulting equations conserve energy. The thermal radiative transfer equations ---------------------------------------- The thermal radiative transfer equations are $$\begin{gathered} \rho\frac{\partial e}{\partial t}=-c\sigma_{a}B+c\sigma_{a}E+Q_{e},\label{eq:material-energy}\\ \frac{\partial E}{\partial t}=-\partial_{\alpha}F_{\alpha}-c\sigma_{a}E+c\sigma_{a}B+Q_{E},\label{eq:zeroth-moment}\\ \frac{1}{c}\frac{\partial F_{\alpha}}{\partial t}=-c\partial_{\beta}P_{\alpha\beta}-\sigma_{t}F_{\alpha},\label{eq:first-moment}\end{gathered}$$ with the summation coefficients $\alpha$ and $\beta$ and the variables [00.00.0000]{} time, position, velocity, mass density, specific material energy, [0.7]{} radiation energy density, radiation flux, radiation pressure, [0.7]{} material temperature, integrated photon emission function, speed of light in a vacuum, [0.7]{} total opacity, absorption opacity, black-body constant, nonhomogeneous material energy source, nonhomogeneous radiation energy source. The radiation transport equation has been integrated over all energy frequencies (the grey approximation) and integrated over angle to produce the first two angular moments. For the derivation in Paper I, the material is assumed to be stationary. The emission term is defined as $$B=aT^{4},$$ where the black body constant $a$ is defined in terms of the Stefan-Boltzmann constant $\sigma_{SB}$ and the speed of light $c$ as $a=4\sigma_{SB}/c$. Assuming the specific intensity is linearly anisotropic in angle, the pressure term becomes $$P_{\alpha\beta}=\frac{1}{3}I_{\alpha\beta}E,$$ where $I$ is the tensor identity. Assuming that the time derivatives are neglected, the first angular moment of the transport equation [\[]{}Eq. (\[eq:zeroth-moment\])[\]]{} can be solved for the radiation flux to get Fick’s law, $$F_{\alpha}=-\frac{c}{3\sigma_{t}}\partial_{\alpha}E.\label{eq:ficks}$$ Replacing the flux term in the zeroth angular moment equation [\[]{}Eq. (\[eq:first-moment\])[\]]{} with Fick’s law and replacing the pressure term by the linearly anisotropic value results in the diffusion equation, $$\frac{\partial E}{\partial t}=\partial_{\alpha}\frac{c}{3\sigma_{t}}\partial_{\alpha}E-c\sigma_{a}E+c\sigma_{a}B+Q_{E}.$$ In the original radiation equations [\[]{}Eqs. (\[eq:zeroth-moment\]) and (\[eq:first-moment\])[\]]{}, if the pressure were not isotropic, the radiation would propagate correctly at the speed of light. However, when the time derivative on the first moment equation is dropped, this adds an error that allows the radiation to propagate faster than the speed of light, $\left\|\bm{F}\right\|>cE$. This effect can be prevented by applying a flux limiter $\lambda$ to the diffusion equation, $$\frac{\partial E}{\partial t}=\partial_{\alpha}\frac{c\lambda}{\sigma_{t}}\partial_{\alpha}E-c\sigma_{a}E+c\sigma_{a}B+Q_{E}.\label{eq:diffusion}$$ With a constant $\lambda=1/3$, the original equation is recovered. The flux limiter is usually written in terms of the constant $R$, $$R=\frac{\sqrt{\left(\partial_{\alpha}E\right)\left(\partial_{\alpha}E\right)}}{\sigma_{t}E},$$ with examples such as the Levermore flux limiter [@levermore1984relating], $$\lambda=\frac{2+R}{6+3R+R^{2}},$$ and the Larsen flux limiter [@morel2000diffusion], $$\lambda=\frac{1}{\left(3^{k}+R^{k}\right)^{1/k}},$$ where $k$ is a chosen constant. Both of these have the correct limits for diffusive regions $\left(R\to0\right)$, which is the diffusion equation, and optically thin regions $\left(R\to\infty\right)$, which describes radiation propagating at the speed of light [@mihalas2013foundations], $$F_{\alpha}\to\begin{cases} -\frac{c}{3\sigma_{t}}\partial_{\alpha}E, & R\to0,\\ -\frac{\partial_{\alpha}E}{\sqrt{\left(\partial_{\beta}E\right)\left(\partial_{\beta}E\right)}}cE, & R\to\infty. \end{cases}$$ The Larsen flux limiter with $k=2$ is used in most of the results in this paper. Time discretization ------------------- The coupled material and radiation energy equations [\[]{}Eqs. (\[eq:material-energy\]) and (\[eq:diffusion\])[\]]{} are discretized fully implicitly in time, using the nonlinear elimination methodology described in Ref. [@brunner2019nonlinear]. The derivation begins with the coupled material and radiation diffusion equations discretized using backward Euler, $$\bm{G}=\left[\begin{array}{c} m\left(e^{n},E^{n}\right)\\ r\left(e^{n},E^{n}\right) \end{array}\right]=\left[\begin{array}{c} \dfrac{\rho}{\Delta t}\left(e^{n}-e^{n-1}\right)+c\sigma_{a}B^{n}-c\sigma_{a}E^{n}-Q_{e}^{n}\\ \dfrac{1}{\Delta t}\left(E^{n}-E^{n-1}\right)-\partial_{\alpha}\dfrac{c\lambda}{\sigma_{t}}\partial_{\alpha}E^{n}+c\sigma_{a}E^{n}-c\sigma_{a}B^{n}-Q_{E}^{n} \end{array}\right]=\left[\begin{array}{c} 0\\ 0 \end{array}\right],\label{eq:ne-starting-point}$$ where $\bm{U}=\left[\begin{array}{cc} e & E\end{array}\right]^{\top}$. Newton’s method updates the state according to \[eq:regular-newton\] $$\begin{gathered} \bm{J}^{\ell}\bm{\delta}^{\ell}=-\bm{G}^{\ell},\\ \bm{U}^{\ell+1}=\bm{U}^{\ell}+\bm{\delta}^{\ell},\label{eq:newton-update}\end{gathered}$$ where $\bm{J}$ is the Jacobian with respect to $e$ and $E$, $$\bm{J}=\left[\begin{array}{cc} J_{m,e} & J_{m,E}\\ J_{r,e} & J_{r,E} \end{array}\right]=\left[\begin{array}{cc} \dfrac{\rho}{\Delta t}+\dfrac{c}{c_{v}}\sigma_{a}\partial_{T}B & -c\sigma_{a}\\ -\dfrac{c}{c_{v}}\sigma_{a}\partial_{T}B & \dfrac{1}{\Delta t}-\partial_{\alpha}\frac{c\lambda}{\sigma_{t}}\partial_{\alpha}+c\sigma_{a} \end{array}\right],$$ and $\bm{\delta}=\left[\delta_{e},\delta_{E}\right]^{\top}$ is the change in the state variables and $c_{v}=\frac{\partial e}{\partial T}$ is the specific heat capacity. To simplify the Newton iteration, the Schur compliment of the Jacobian is taken, resulting in the Jacobian system $$\left[\begin{array}{cc} J_{m,e} & J_{m,E}\\ 0 & J_{r,E}-J_{r,e}J_{m,e}^{-1}J_{m,E} \end{array}\right]\left[\begin{array}{c} \delta_{e}\\ \delta_{E} \end{array}\right]=\left[\begin{array}{c} -F_{e}\\ -F_{E}+J_{r,e}J_{m,e}^{-1}F_{e} \end{array}\right]\label{eq:newton-schur}$$ that can be solved first for $\delta_{E}$ and then $\delta_{e}$. This system of equations can be first solved for $\delta_{E}$ and then $\delta_{e}$. The solution in Eq. (\[eq:newton-schur\]) combined with the update in Eq. (\[eq:newton-update\]) is full Newton iteration without approximation. Noting that $\partial_{T}B=4aT^{3}$, the term $J_{r,e}J_{m,e}^{-1}$ can be written in terms of the Fleck factor commonly used in implicit Monte Carlo [@fleckjr1971implicit], $$f=\left(1+c\sigma_{a}\Delta t\frac{4aT^{3}}{\rho c_{v}}\right)^{-1},$$ which is used to forward-predict the emission source, as $$J_{r,e}J_{e,e}^{-1}=f-1.$$ Finally, the temperature is held constant in $f$ during each time step, which results in an inexact Newton iteration scheme, $$\dfrac{1}{\Delta t}E^{n,\ell+1}-\partial_{\alpha}\dfrac{c\lambda}{\sigma_{t}}\partial_{\alpha}E^{n,\ell+1}+c\sigma_{a}fE^{n,\ell+1}=\dfrac{1}{\Delta t}E^{n-1}+c\sigma_{a}B^{n,\ell}-\left(1-f\right)c\sigma_{a}E^{n,\ell}+Q_{E}^{n},$$ with the time step index $n$ and the iteration index $\ell$. When $E^{n}$ has converged, the terms involving $f$ in the diffusion equation cancel out and the original, time-discretized diffusion equation is recovered. Equations (\[eq:regular-newton\]) represent the standard Newton solution method for the thermal radiative transfer equations. Here, a nonlinear acceleration technique from Ref. [@brunner2019nonlinear] is additionally applied. The following two functions are helpful in describing the iterative process of solving the equations, \[eq:ne-equations\] $$\begin{gathered} m\left(e^{\star},E\right)=\dfrac{\rho}{\Delta t}\left(e^{\star}-e^{n-1}\right)+c\sigma_{a}B^{\star}-c\sigma_{a}E-Q_{e}^{n}=0,\\ r^{\dagger}\left(e,E,E^{\star}\right)=\dfrac{1}{\Delta t}E^{\star}-\partial_{\alpha}\dfrac{c\lambda}{\sigma_{t}}\partial_{\alpha}E^{\star}+c\sigma_{a}fE^{\star}-\dfrac{1}{\Delta t}E^{n-1}-c\sigma_{a}B+\left(1-f\right)c\sigma_{a}E-Q_{E}^{n}=0\end{gathered}$$ where $e^{\star}$ and $E^{\star}$ represent the values to be solved for. Note the use of $r^{\dagger}$ to distinguish the diffusion equation from the starting equation $r$. The solution of $r^{\dagger}\left(e,E,E^{\star}\right)=0$ represents solving a spatially-coupled linear equation, as discussed in Sec. \[subsec:mat-rad-solve\]. Note that the value of the material energy in these equations is entirely dependent on the radiation energy, which means that a solution of $r^{\dagger}$ represents an inexact Newton iteration of the full material and radiation energy system. After a solution of $r^{\dagger}$ for the radiation energy, the material energy is updated by solving $m\left(e^{},E\right)=0$, which represents a nonlinear solve using Newton’s method, $$e^{\star}\leftarrow e^{\star}-\left(J_{m,e}^{\star}\right)^{-1}\left[\dfrac{\rho}{\Delta t}\left(e^{\star}-e^{n-1}\right)+c\sigma_{a}B^{\star}-c\sigma_{a}E-Q_{e}^{n}\right],\label{eq:full-material-solve}$$ where $\leftarrow$ represents an update to the value of $e^{\star}$. This update is performed until $e^{\star}$ is converged. This differs from standard Newton iteration, where the material energy update would be done simultaneously with the radiation energy update and without solving $m\left(e^{},E\right)=0$ until $e^{*}$ is converged. The iterative solution procedure for these two equations is described in Alg. \[alg:mat-rad-solve\]. The outer iterations, or the inexact Newton iteration of the material and radiation energy system, continue until the convergence criteria are met, \[eq:mat-rad-convergence\] $$\begin{gathered} \left\|\frac{e^{n,\ell+1}-e^{n,\ell}}{e^{n,\ell+1}}\right\|<\epsilon_{e},\\ \left\|\frac{E^{n,\ell+1}-E^{n,\ell}}{E^{n,\ell+1}}\right\|<\epsilon_{E}.\end{gathered}$$ The inner iterations, or the iterations required to converge $m$ and $r$ within an outer iteration, use similar equations for determining convergence. As the absorption term appears with both $E^{n,\ell}$ and $E^{n,\ell+1}$, convergence is required for energy conservation to be preserved (see Sec. \[subsec:conservation\]). Introduction to smoothed particle hydrodynamics\[subsec:sph-intro\] ------------------------------------------------------------------- Smoothed particle hydrodynamics involves interpolation of fields using kernels, which are functions centered at interpolation points. The initial assumption needed is that SPH kernels approximate delta functions, or $W\left(x-x',h\right)\to\delta\left(x-x'\right)$ as $h\to0$, where $h$ is the smoothing parameter that determines the width of the kernels $$\begin{aligned} g\left(x\right) & =\int_{V}\delta\left(x-x'\right)g\left(x'\right)dV'\\ & \approx\int_{V}W\left(x-x',h\right)g\left(x'\right)dV'.\end{aligned}$$ This also means that the integral of the kernel should be equal to one, $$\int_{V}W\left(x-x',h\right)dV'=1.\label{eq:kernel-normalization}$$ In practice, the $h$ can depend on both $x$ and $x'$. The second assumption is that a set of kernels in space can be used to form a quadrature, with abscissas at the kernel centers $\bm{x}_{i}$ and weights equal to the kernel volume $V_{i}$. With these two assumptions, the kernels can be used to interpolate a function, $$\begin{aligned} \left\langle g\left(x\right)\right\rangle & =\int_{V}W\left(x-x',h\right)g\left(x'\right)dV'\nonumber \\ & \approx\sum_{j}V_{j}W\left(x-x_{j},h\right)g_{j}.\label{eq:sph-interpolant}\end{aligned}$$ Here $g_{j}$ is the function evaluated at point $j$, $g_{j}=g\left(x_{j}\right)$, and $\left\langle \cdot\right\rangle $ indicates an interpolated quantity. The interpolant is used to calculate derivatives of the field. This interpolant does not have the Kronecker delta property, $$\left\langle g\left(x\right)\right\rangle _{i}=\sum_{j}V_{j}W\left(x_{i}-x_{j},h\right)g_{j}\neq g_{i}.$$ To approximate a derivative in SPH, another three properties are needed from the kernel: that the kernel does not intersect a boundary, that the derivative of the kernel is antisymmetric in $\bm{x}$ and $\bm{x}'$, and that the integral of the derivative of the kernel is zero. With these approximations, the derivative of a function can be approximated as $$\begin{aligned} \left\langle \partial_{x,\alpha}g\left(x\right)\right\rangle & =\int_{V}W\left(x-x',h\right)\partial_{x',\alpha}g\left(x'\right)dV'\nonumber \\ & =-\int_{V}\partial_{x',\alpha}W\left(x-x',h\right)g\left(x'\right)dV'\nonumber \\ & =\partial_{x,\alpha}\int_{V}W\left(x-x',h\right)g\left(x'\right)dV'-g\left(x\right)\partial_{x,\alpha}\int_{V}W\left(x-x',h\right)dV'\nonumber \\ & \approx\sum_{j}V_{j}\left(g_{j}-g\left(x\right)\right)\partial_{x,\alpha}W\left(x-x_{j},h\right).\label{eq:sph-derivative}\end{aligned}$$ Note that because of the added $g\left(x\right)$ term (which equals zero in integral form due to the derivative), this approximation of the derivative goes to zero for a constant function. For a diffusion-like problem, a second-order series expansion for $g\left(x\right)$ about $x'$ and $g\left(x'\right)$ about $x$, $$\partial_{x',\alpha}g\left(x'\right)\approx\frac{x_{\alpha}-x_{\alpha}'}{\left(x_{\beta}-x_{\beta}'\right)\left(x_{\beta}-x_{\beta}'\right)}\left[g\left(x'\right)-g\left(x\right)\right],$$ can be used to produce a commonly-used second-order approximation to the second derivative [@monaghan2005smoothed], $$\begin{aligned} \left\langle \partial_{x,\alpha}\left[k\left(x\right)\partial_{x,\alpha}g\left(x\right)\right]\right\rangle & =\int_{V}W\left(x-x',h\right)\partial_{x',\alpha}\left[k\left(x'\right)\partial_{x',\alpha}g\left(x'\right)\right]dV'\nonumber \\ & =-\int_{V}k\left(x'\right)\partial_{x',\alpha}W\left(x-x',h\right)\partial_{x',\alpha}g\left(x'\right)dV'\nonumber \\ & =\partial_{x,\alpha}\int_{V}k\left(x'\right)W\left(x-x',h\right)\partial_{x',\alpha}g\left(x'\right)dV'+k\left(x\right)\partial_{x,\alpha}g\left(x\right)\partial_{x,\alpha}\int_{V}W\left(x-x',h\right)dV'\nonumber \\ & =\int_{V}\left[k\left(x\right)\partial_{x,\alpha}g\left(x\right)+k\left(x'\right)\partial_{x',\alpha}g\left(x'\right)\right]\partial_{x,\alpha}W\left(x-x',h\right)dV'\nonumber \\ & \approx\int_{V}\left[k\left(x\right)+k\left(x'\right)\right]\left[g\left(x\right)-g\left(x'\right)\right]\frac{x_{\alpha}-x_{\alpha}'}{\left(x_{\beta}-x_{\beta}'\right)\left(x_{\beta}-x_{\beta}'\right)}\partial_{x,\alpha}W\left(x-x',h\right)dV'\nonumber \\ & \approx\sum_{j}V_{j}\left(k\left(x\right)+k_{j}\right)\left(g\left(x\right)-g_{j}\right)\frac{x_{\alpha}-x_{\alpha,j}}{\left(x_{\beta}-x_{\beta,j}\right)\left(x_{\beta}-x_{\beta,j}\right)}\partial_{x,\alpha}W\left(x-x_{j},h\right).\label{eq:sph-second-derivative}\end{aligned}$$ Like the first derivative approximation, the second derivative is zero for a constant function because the $g$ terms cancel. When discretizing an equation in SPH, the equation is first multiplied by the kernel $W\left(x-x',h\right)$ and integrated, similar to the finite element method. The distinction is that the interpolant is only used to calculate derivatives. The terms without derivatives are approximated using the definition of the interpolant, $$\left\langle g\left(x\right)\right\rangle =\int_{V}W\left(x-x',h\right)g\left(x'\right)dV'\approx g\left(x\right).\label{eq:interpolant-approx}$$ The terms with derivatives are approximated as derivatives of the interpolant, as in Eq. \[eq:sph-derivative\]. The final step in an SPH discretization is to add constraints to match the free variables in the equations. This can be done by multiplying the equations by a delta function $\delta\left(x-x_{i}\right)$ and integrating, or equivalently, evaluating the functions at $x=x_{i}$. This results in the fully-discrete approximations to Eqs. (\[eq:sph-interpolant\]), (\[eq:sph-derivative\]), and (\[eq:sph-second-derivative\]), $$\begin{gathered} \left\langle g\left(x\right)\right\rangle _{i}\approx\sum_{j}V_{j}W_{ij}g_{j},\\ \left\langle \partial_{x,\alpha}g\left(x\right)\right\rangle _{i}\approx\sum_{j}V_{j}\left(g_{j}-g_{i}\right)\partial_{i,\alpha}W_{ij},\\ \left\langle \partial_{x,\alpha}\left[k\left(x\right)\partial_{x,\alpha}g\left(x\right)\right]\right\rangle _{i}\approx\sum_{j}V_{j}\left(k_{i}+k_{j}\right)\left(g_{i}-g_{j}\right)\frac{x_{ij,\alpha}}{x_{ij,\beta}x_{ij,\beta}}\partial_{i,\alpha}W_{ij},\end{gathered}$$ where $$\begin{gathered} x_{ij,\alpha}=x_{i,\alpha}-x_{j,\alpha},\\ W_{ij}=W\left(x_{ij},h\right),\\ \partial_{i,\alpha}W_{ij}=\partial_{x_{i},\alpha}W\left(x_{ij},h\right).\end{gathered}$$ The smoothing length $h$ connects the analytic kernel in Eq. (\[eq:wendlandc4\]) with the SPH kernels, $$W\left(x-x',h\right)=\psi\left(\frac{\sqrt{\left(x_{\alpha}-x_{\alpha}^{\prime}\right)\left(x_{\alpha}-x_{\alpha}^{\prime}\right)}}{h}\right).$$ In practice, the smoothing length is not constant and instead has discrete values at the SPH nodes. To retain symmetry in the derivatives [\[]{}such that $\partial_{i,\alpha}W\left(x_{i}-x_{j},h\right)=\partial_{j,\alpha}W\left(x_{j}-x_{i},h\right)$[\]]{}, the kernel values are taken to be an average of the evaluations using these two smoothing lengths, $$\begin{gathered} W_{ij}=\frac{1}{2}\left[W\left(x_{ij},h_{i}\right)+W\left(x_{ij},h_{j}\right)\right]\\ \partial_{i,\alpha}W_{ij}=\frac{1}{2}\left[\partial_{i,\alpha}W\left(x_{ij},h_{i}\right)+\partial_{i,\alpha}W\left(x_{ij},h_{j}\right)\right].\end{gathered}$$ The Wendland functions [@wendland1995piecewise] are one example of a kernel that can be used in SPH. The Wendland C4 kernel is used here, $$W_{\text{wendland}}\left(x_{ij},h\right)=\begin{cases} k\left(1-r\right)^{5}\left(8r^{2}+5r+1\right), & r\equiv x_{ij}/h\leq1,\\ 0, & \text{otherwise}, \end{cases}\label{eq:wendlandc4}$$ with a dimensionally-dependent normalization constant $k$ that is chosen such that Eq. (\[eq:kernel-normalization\]) is satisfied. Spatial discretization ---------------------- To perform the spatial discretization, the material and radiation energy equations [\[]{}Eqs. (\[eq:ne-equations\])[\]]{} are multiplied by $W\left(x_{i}-x',h\right)$ and integrated, $$\begin{gathered} \left\langle \frac{\rho}{\Delta t}\left(e^{n,\ell+1}-e^{n-1}\right)+c\sigma_{a}B^{n,\ell+1}-c\sigma_{a}E^{n,\ell+1}-Q_{e}^{n}\right\rangle _{i}=0,\\ \left\langle \frac{1}{\Delta t}E^{n,\ell+1}-\partial_{\alpha}\frac{c\lambda}{\sigma_{t}}\partial_{\alpha}E^{n,\ell+1}+c\sigma_{a}fE^{n,\ell+1}\right\rangle _{i}=\left\langle \frac{1}{\Delta t}E^{n-1}+c\sigma_{a}B^{n,\ell}-\left(1-f\right)c\sigma_{a}E^{n,\ell}+Q_{E}^{n}\right\rangle _{i},\end{gathered}$$ where $\left\langle \cdot\right\rangle $ is the interpolant notation as defined in Sec. \[subsec:sph-intro\]. The terms not involving derivatives can be approximated using Eq. (\[eq:interpolant-approx\]), e.g. $$\begin{gathered} \left\langle \frac{\rho}{\Delta t}e^{n-1}\right\rangle \approx\frac{\rho_{i}}{\Delta t}e_{i}^{n-1},\\ \left\langle c\sigma_{a}fE^{n,\ell+1}\right\rangle _{i}\approx c\sigma_{a,i}f_{i}E_{i}^{n,\ell+1}.\end{gathered}$$ The second derivative in the radiation energy equation can be approximated using Eq. (\[eq:sph-second-derivative\]), $$\left\langle -\partial_{\alpha}D\left(x\right)\partial_{\alpha}E\left(x\right)\right\rangle \approx-\sum_{j}V_{j}\left(D_{i}+D_{j}\right)\left(E_{i}-E_{j}\right)\frac{x_{ij,\alpha}}{x_{ij,\beta}x_{ij,\beta}}\partial_{i,\alpha}W_{ij},$$ with the diffusion coefficient $$D_{i}=\frac{c\lambda_{i}}{\sigma_{t,i}}.$$ With these approximations, the material and radiation equations are \[eq:fully-disc-mat-rad\] $$\frac{\rho_{i}}{\Delta t}\left(e_{i}^{n,\ell+1}-e_{i}^{n-1}\right)+c\sigma_{a,i}B_{i}^{n,\ell+1}-c\sigma_{a,i}E_{i}^{n,\ell+1}-Q_{e,i}^{n}=0,$$ $$\begin{aligned} \frac{1}{\Delta t}E_{i}^{n,\ell+1}-\sum_{j}V_{j}\left(D_{i}+D_{j}\right)\left(E_{i}^{n,\ell+1}-E_{j}^{n,\ell+1}\right)\frac{x_{ij,\alpha}}{x_{ij,\beta}x_{ij,\beta}}\partial_{i,\alpha}W_{ij}+c\sigma_{a,i}f_{i}E_{i}^{n,\ell+1}\nonumber \\ \qquad=\frac{1}{\Delta t}E_{i}^{n-1}+c\sigma_{a,i}B_{i}^{n,\ell}-\left(1-f_{i}\right)c\sigma_{a,i}E_{i}^{n,\ell}+Q_{E,i}^{n}.\end{aligned}$$ The material energy equation is independent for each point $i$, while the radiation energy equation for point $i$ linearly couples to the radiation energy from nearby nodes. For more information on how these equations are solved, see Sec. \[subsec:mat-rad-solve\]. Conservation\[subsec:conservation\] ----------------------------------- The coupled material and radiation equations [\[]{}Eqs. (\[eq:fully-disc-mat-rad\])[\]]{} are conservative in energy to the tolerance of the nonlinear iteration process once converged. The quantity that is conserved is the total energy over all points, which is the material energy plus the radiation energy, $$\sum_{i}\left(m_{i}e_{i}+V_{i}E_{i}\right)=\text{const}.$$ At convergence, $e^{n,\ell+1}\approx e^{n,\ell}$ and $E^{n,\ell+1}\approx E^{n,\ell}$, which results in the simplified equations $$\begin{gathered} \frac{\rho_{i}}{\Delta t}\left(e_{i}^{n}-e_{i}^{n-1}\right)+c\sigma_{a,i}B_{i}^{n}-c\sigma_{a,i}E_{i}^{n}=Q_{e,i}^{n},\\ \frac{1}{\Delta t}\left(E_{i}^{n}-E_{i}^{n-1}\right)-\sum_{j}V_{j}\left(D_{i}+D_{j}\right)\left(E_{i}^{n}-E_{j}^{n}\right)\frac{x_{ij,\alpha}}{x_{ij,\beta}x_{ij,\beta}}\partial_{i,\alpha}W_{ij}+c\sigma_{a,i}E_{i}^{n}-c\sigma_{a,i}B_{i}^{n}=Q_{E,i}^{n}.\end{gathered}$$ Adding the two equations, multiplying by $V_{i}$, and summing over $i$ results in $$\sum_{i}\left[m_{i}\left(e_{i}^{n}-e_{i}^{n-1}\right)+V_{i}\left(E_{i}^{n}-E_{i}^{n-1}\right)\right]=\Delta t\sum_{i}\left(Q_{e,i}^{n}+Q_{E,i}^{n}\right).$$ Because the diffusion spatial derivative is antisymmetric about $i$ and $j$, it disappears under summation over both $i$ and $j$. This final equation says that any gains or losses in the total energy are as a result of specified sources. At equilibrium and absence external sources, both the material and the radiation equations reduce to an equilibrium absorption-emission rate at every point $x_{i}$, which is $E_{i}=B_{i}$. Methodology\[sec:methodology\] ============================== The thermal radiative transfer methods described here are implemented within the open-source SPH code Spheral, described at <https://wci.llnl.gov/simulation/computer-codes/spheral> and publicly available at <https://github.com/jmikeowen/spheral>, although the radiative transfer methods are not available in the open source version. For access to the full code, including thermal radiative transfer and input scripts for the problems presented in this paper, please contact the authors. Spheral is written in C++ with a Python interface, which allows for simple addition of new physics in either programming language. The opacities, flux limiters, and equations of state can be chosen arbitrarily for subsets of the SPH points as needed, allowing problems with any number of distinct materials to be studied. In the succeeding examples Spheral’s default method [@Owen2010] for computing the smoothing scale per point ($h_{i}$) is used with the C4 Wendland kernel [\[]{}Eq. (\[eq:wendlandc4\])[\]]{}. The target radius of support is 4 points, i.e., the local smoothing scale for each point is chosen to be 4 times the local particle spacing. Material and radiation energy solve\[subsec:mat-rad-solve\] ----------------------------------------------------------- The iterative process in Alg. \[alg:mat-rad-solve\] [@brunner2019nonlinear] can be written in terms of operator matrices as $$\left[\begin{array}{c} e^{\ell+1}\\ E^{\ell+1} \end{array}\right]=\left[\begin{array}{cc} \mathbb{M}^{-1} & 0\\ 0 & \mathcal{I} \end{array}\right]\left[\begin{array}{cc} \mathcal{R} & \mathcal{A}\\ 0 & \mathcal{I} \end{array}\right]\left[\begin{array}{cc} \mathcal{I} & 0\\ 0 & \mathcal{D}^{-1} \end{array}\right]\left[\begin{array}{cc} \mathcal{I} & 0\\ \mathbb{N} & \mathcal{S} \end{array}\right]\left[\begin{array}{c} e^{\ell}\\ E^{\ell} \end{array}\right],\label{eq:matrix-iteration}$$ with the operators defined by [00.00.0000]{} [0.8]{} Solve the nonlinear material equation, given a source $P_{i}$, $$\dfrac{\rho_{i}}{\Delta t}\left(e_{i}-e_{i}^{n-1}\right)+c\sigma_{a,i}aT_{i}^{4}=P_{i},\label{eq:material-energy-solve}$$ for each $e_{i}$, Solve the diffusion equation, given a source $P_{i}$, $$\left(\frac{1}{\Delta t}+c\sigma_{a,i}f_{i}\right)E_{i}-\sum_{j}V_{j}\left(D_{i}+D_{j}\right)\left(E_{i}-E_{j}\right)\frac{x_{ij,\alpha}}{x_{ij,\beta}x_{ij,\beta}}\partial_{i,\alpha}W_{ij}=P_{i},\label{eq:diffusion-solve}$$ for all $E_{i}$, Calculate the radiation diffusion source, $\dfrac{1}{\Delta t}E_{i}^{n-1}-\left(1-f_{i}\right)c\sigma_{a,i}E_{i}+Q_{E,i}$, Calculate the emission of radiation from the material, $c\sigma_{a,i}aT_{i}^{4}$, Calculate the radiation energy absorption, $c\sigma_{a,i}E_{i}$, Set the material energy source, $Q_{e,i}^{n}$, and the identity operator $\mathcal{I}$. This simplifies the addition of new physics, reduces code duplication, and allows testing of each operator independently. Note that $\mathbb{M}^{-1}$ and $\mathbb{N}$ are nonlinear operators, and should not be mistaken for matrices. The operators $\mathcal{D}^{-1}$, $\mathcal{S},$and $\mathcal{A}$ are linear, but in practice, they are not explicitly formed into matrices. The application of the combined operator in Eq. (\[eq:matrix-iteration\]) is equivalent to a single outer iteration in Alg. \[alg:mat-rad-solve\]. The opacities and other material data ($\sigma_{a}$, $\sigma_{s}$, $D$, $f$, $\lambda$, and $c_{v}$) are calculated at the start of the time step and held constant within the time step. The material energy solve in Eq. (\[eq:material-energy-solve\]) is done independently for each point using Newton’s method, as shown in Eq. (\[eq:full-material-solve\]). Defining the function and Jacobian for the material energy solve as \[eq:material-energy-solve-operator\] $$\begin{gathered} u\left(e_{i}\right)=\frac{\rho}{\Delta t}\left(e_{i}-e^{n-1}\right)+c\sigma_{a}aT^{4}-P,\\ J_{m,e}\left(e\right)=\frac{\rho}{\Delta t}+\frac{4c\sigma_{a}aT^{3}}{c_{v}},\end{gathered}$$ the material energy is solved iteratively as $$e_{i}^{k+1}=e_{i}^{k}-J_{m,e}^{-1}\left(e_{i}^{k}\right)u\left(e_{i}^{k}\right)\label{eq:material-energy-update}$$ until converged, where $k$ is the Newton iteration index. As the call to an external equation of state can be expensive, the temperature is updated for all points simultaneously, followed by an update of the material energy for each point independently. The iterative process in Eq. (\[eq:material-energy-update\]) proceeds until convergence each time the $\mathcal{M}^{-1}$ operator is called. The radiation energy solve in Eq. (\[eq:diffusion-solve\]) is done using the Hypre BoomerAMG preconditioner with GMRES [@falgout2002hypre]. Because the material properties (opacities, specific heats, hydrodynamic variables) are held constant within a time step, the preconditioner can be initialized once at the start of each time step and then reused each time the $\mathcal{D}^{-1}$ operator is called within the time step. The radiation energy $E^{\ell}$ is used as an initial guess for the GMRES solver when calculating $E^{\ell+1}$, which reduces the number of GMRES iterations as the solution nears convergence, as is discussed in Sec. \[subsec:manufactured-solution\]. Time step choice\[subsec:time-step\] ------------------------------------ Due to the implicit solve, the time step is updated not based on the propagation speed of the radiation, but instead based on the observed change in radiation energy over a time step. This is a heuristic for the stability and accuracy of one Newton step. The idea is to limit the fractional change in the material and radiation energies to some finite value and increase or decrease the time step according to the change from the previous time step. For example, for the material energy, the fractional change in energy is calculated as $$\eta_{e}=\max_{i}\left(\frac{\left\|e_{i}^{n}-e_{i}^{n-1}\right\|}{e_{i}^{n}+\eta_{e}^{\text{target}}\bar{e}^{n}}\right),$$ where $\eta_{e}^{\text{target}}$ is the target fractional change ($\eta_{e}^{\text{target}}=0.05$ for the problems in this paper) and $\bar{e}^{n}$ is the volume-averaged energy, $$\bar{e}^{n}=\frac{\sum_{i}e_{i}^{n}V_{i}}{\sum_{i}V_{i}}.$$ The $\bar{e}^{n}$ in the denominator of the fractional change prevents points with very small radiation energy from dominating the time step when changing a small magnitude relative to the other points in the problem. Given the fractional change over the previous time step, a proposed time step is calculated as $$\Delta t_{e}^{n+1}=\Delta t^{n}\left(\frac{\eta_{e}^{\text{target}}}{\eta_{e}}\right)^{1/2}.$$ The same is process is done for the radiation energy $E$, and the time step is chosen to be the more restrictive of these, $\Delta t^{n+1}=\min\left(\Delta t_{e}^{n+1},\Delta t_{E}^{n+1}\right)$, within the minimum and maximum time steps specified by the user. For more information on time stepping strategies for thermal radiative transfer with diffusion, including the one used here, see Ref. [@rider1999time]. Results\[sec:results\] ====================== The results include three problems. The first is an infinite medium equilibrium test, which is designed to show that the material and radiation energies come to equilibrium at the correct rate, testing the time discretization, emission, and absorption. The second is the Su-Olson Marshak wave, which simulates the diffusion of a planar radiation source into a vacuum, additionally testing the diffusion rate of the radiation. The final problem is a manufactured solution that extends the results to two and three dimensions. The tolerances used for each of these problems are listed in Table \[tab:tolerances\], with the inner tolerances referring to the convergence metric for ending the Newton solve for the material energy and the GMRES solve for the radiation energy and the outer tolerances used for ending the nonlinear elimination iteration process (see Sec. \[subsec:mat-rad-solve\]). Infinite medium equilibrium test\[subsec:infinite-medium\] ---------------------------------------------------------- The first problem involves a single material in an infinite medium in which the material and radiation energies are initially out of equilibrium. The units of length, time, temperature, and mass are chosen such that the absorption opacity $\sigma_{a}$, speed of light $c$, black body constant $a$, and ratio of proton mass to the Boltzmann constant, $m_{p}/k_{B}$, are all one. The equation of state is chosen to be an ideal gas, which in these units is defined as $$T=\left(\gamma-1\right)\mu e=\alpha e,$$ where $\mu$ is the molecular mass and $\gamma$ is the heat capacity ratio. The problem can be written in the simplified unit system as a system of two ordinary differential equations, $$\begin{gathered} \rho\frac{\partial e}{\partial t}=E-\left(\alpha e\right)^{4},\\ \frac{\partial E}{\partial t}=-E+\left(\alpha e\right)^{4}.\end{gathered}$$ These equations are solved using a fifth-order Radau IIA integrator [@hairer1999stiff] for comparison to the results given by the code. The problem is run for two different cases, both of which have $\alpha=1$ and a density of $\rho=1$, and in each case, the simulation proceeds until $t_{end}=10$, which allows the material and radiation energies to reach equilibrium. The first case starts with hot material and cold radiation, while the second starts with cold material and hot radiation, as shown in Table \[tab:infinite-initial\]. The time step is allowed to increase by up to an order of magnitude per time step, up to the limit set by the changing material and radiation energies, with a hard cap of $\Delta t_{max}=0.1$ (see Sec. \[subsec:time-step\]). The first and second cases take, respectively, 779 and 155 time steps to reach the goal time. The results are shown in Fig. \[fig:infinite\]. The reference and numeric solutions agree well for both problems. The $L_{1}$ relative error for this problem is calculated with an integral in time, $$L_{1}\text{ error}=\frac{\int_{0}^{t_{end}}\left\|T_{\text{ref}}-T_{\text{num}}\right\|dt}{\int_{0}^{t_{end}}T_{\text{ref}}dt},\label{eq:infinite-error}$$ using Simpson’s rule evaluated at the time steps. The error between the reference and numeric solutions for the first case (with hot material) is $3.32\times10^{-4}$ for the material temperature and $2.55\times10^{-4}$ for the radiation temperature. The error calculated in the same way for the second case (with hot radiation) is $1.22\times10^{-3}$ for the material temperature and $6.21\times10^{-4}$ for the radiation temperature. For a fixed time step $\Delta t$, the error decreases linearly with $\Delta t$ as expected, as shown in Fig. \[fig:infinite-convergence\]. The relative error between the initial and final energies ranges from $1.11\times10^{-15}$ (hot radiation) and $4.44\times10^{-16}$ (hot material) for $\Delta t=0.1$ to $4.22\times10^{-13}$ (hot radiation) and $5.53\times10^{-13}$ (hot material) for $\Delta t=0.0001$, which is of the correct order given the outer tolerance of $10^{-12}$ and inner tolerance of $10^{-14}$ (Table \[tab:tolerances\]). Su-Olson Marshak wave --------------------- The second problem is a Marshak planar wave as described in Ref. [@su1997analytical]. The problem consists of a reflecting plane at $x=0.0$ and a radiation source from $0\leq x\leq x_{0}$ that is turned on for $0\leq\tau\leq\tau_{0}$, where $x$ is a scaled distance and $\tau$ is a scaled time ($z$ is used in the paper for physical distance). The problem is started with a material and radiation energy of $10^{-5}$ so the heat capacity (an analytic $c_{v}\propto T^{3}$) is nonzero. The units for the problem are chosen such that $\sigma_{t}=1$, $\epsilon c=1$, and $a=1$, which makes the scaled units referenced in the paper equal to the physical units ($x=z$ and $\tau=t$). The solution to the problem is semianalytic, and involves evaluating integrals to high precision over $\eta\in\left[0,1\right]$ for each spatial point at each time. For these results, due to the complicated structure of the solution near $\eta=1$, a quadrature is first generated in $\xi\in\left[-1,1\right]$ and then transformed into $\eta$ via the relation $$\eta=\frac{1}{\ell}\ln\left(\frac{1}{2}\left(1-\xi+e^{\ell}\left(1+\xi\right)\right)\right),$$ in which the transformation becomes linear as $\ell\to0$ and concentrates more points near $\eta=1$ as $\ell$ increases. A 5000-point Gauss-Legendre quadrature with $\ell=3$ works well for the specific problem considered below. The numerical result at the end time is compared to this semianalytic result using an $L_{1}$ error, $$L_{1}\text{ error}=\frac{\sum_{i}\left\|e_{i}-e_{i}^{\text{semianalytic}}\right\|}{\sum_{i}e_{i}^{\text{semianalytic}}},\label{eq:marshak-error}$$ which differs from the standard approach of comparison to specific $x$ and $\tau$ values tabulated in the paper. The parameters used for these results include an absorption fraction of $c_{a}=0.5$, a time scaling factor of $\epsilon=1.0$, a radiation source extent of $x_{0}=0.5$, and a scaled time at which the source is turned off of $\tau_{0}=10.0$. The problem is run until $\tau=100.0$. The spatial extent of the problem is $0\leq x\leq100$. The time evolution of the semianalytic and numeric solutions is shown in Fig. \[fig:marshak-solution\]. The spatial convergence results are shown in Fig. \[fig:marshak\] for three different time steps. The code converges to the semianalytic result with second-order accuracy spatially as the point spacing decreases and with first-order accuracy temporally as the time step decreases, as we expect based on SPH spatial interpolation and backward Euler time integration. Manufactured solution for material-radiation coupling\[subsec:manufactured-solution\] ------------------------------------------------------------------------------------- This problem tests the material-radiation coupling using the method of manufactured solutions (MMS). To apply MMS to Eqs. (\[eq:fully-disc-mat-rad\]), solutions for $e$ and $E$ are chosen for the equations, \[eq:manufactured-solutions\] $$\begin{gathered} e_{m}=e_{0}\left[1.2+\prod_{\alpha=1}^{\text{dim}}\cos\left(2\pi\frac{x_{\alpha}-v_{m}t}{d_{m}}\right)\right],\\ E_{m}=E_{0}\left[1.2+\prod_{\alpha=1}^{\text{dim}}\cos\left(2\pi\frac{x_{\alpha}+v_{m}t}{d_{m}}-\omega_{m}\right)\right],\end{gathered}$$ with $e_{0}=10^{13}\text{ ergs\ensuremath{\cdot}g}^{-1}$, $E_{0}=7a\left(e_{0}/c_{v}\right)^{4}\text{ ergs\ensuremath{\cdot}cm}^{-3}$, $v_{m}=5\times10^{9}\text{ cm\ensuremath{\cdot}s}^{-1}$, $d_{m}=5\text{ cm}$, and $\omega_{m}=1/16$. The opacities are set to constant values of $\sigma_{a}=0.05\text{ cm}^{-1}$ and $\sigma_{s}=0.95\text{ cm}^{-1}$. These solutions are inserted into the non-discretized equations [\[]{}Eqs. (\[eq:material-energy\]) and (\[eq:diffusion\])[\]]{} to calculate sources, $Q_{e}$ and $Q_{E}$. The solutions are chosen such that the magnitude of the individual terms in Eqs. (\[eq:material-energy\]) and (\[eq:diffusion\]) is approximately equal, with the opposite time movement of the waves $\pm v$ allowing each individual piece of physics to dominate at certain times. The spatial extent of the problem is set to one full wavelength, or $d_{m}$ in each dimension (to allow for periodic boundary conditions), and the problem is run for one cycle, or until $t_{end}=d_{m}/v_{m}$, using a time step of $\Delta t=t_{end}/1000$. At $t_{end}$, the numerical solution is compared to the analytic solution in Eqs. (\[eq:manufactured-solutions\]) to calculate the $L_{1}$ relative error, $$L_{1}\text{ error}=\frac{\sum_{i}\left\|e_{i}-e_{m,i}\right\|}{\sum_{i}e_{m,i}}\label{eq:manufactured-error}$$ (with a similar equation for $E$). The solution for the problem with $128^{2}$ points is shown in Fig. \[fig:manufactured-solution-2d\]. At the end time, the solution should be equal to the initial condition. The relative error between the solution after one cycle (at $10^{-9}$ s) and the manufactured solution for the specific thermal energy and radiation energy density is shown in Fig. \[fig:manufactured-error\]. The relative error is highest where the solution is lowest, with maxima of 0.006 for the specific material energy and 0.008 for the radiation energy. For much of the domain, the relative error is below 0.001 for both the material and radiation energy. At the end time, the relative difference between the starting and ending energy of the system is $1.67\times10^{-12}$, which indicates good energy conservation. The $L_{1}$ error for the manufactured problem in 1D for several time step values is shown in Fig. \[fig:manufactured-convergence-1d\]. The numeric solution converges to the manufactured solution with second-order accuracy in space and first-order accuracy in time (again as expected). For a time step of $2\times10^{-13}$, the error in the radiation energy indicates second-order convergence for all values of point spacing. The error in the material energy, which only has spatial coupling through the radiation, is much more dependent on the time step for this problem than the radiation energy. As in the previous problems, the error decreases linearly with the time step, when not limited by the spatial error. Based on the results in 1D, a time step of $4\times10^{-13}$ is chosen for the comparison of 1D, 2D, and 3D results up to $128^{d}$ points (for the dimension $d$). The convergence results for 1D, 2D, and 3D are shown in Fig. \[fig:manufactured-convergence\]. The 2D and 3D results have similar errors compared to the manufactured solution for similar spacing of points, although not identical due to differences in the interpolation functions for 2D and 3D, and also show second-order convergence when not limited by the time step. While the number of processors used for each case was not chosen to show scaling with a constant number of points per processor or a constant number of total points, the performance results in Table \[tab:manufactured-timing\] show some general scaling trends. In 2D, the small number of points per processor means that communication costs dominate, and adding more points per each processor helps reduce that cost. In 3D, the cases with 910 points per processor show around a 74 percent efficiency when the number of points and processors are quadrupled. With many more points per processor than the other cases, the 2,097,152-point problem on 576 processors runs with a similar efficiency to the 262,144-point problem in 288 processors. The number of outer iterations per time step is governed for this problem by the tolerance. For the tolerance given in Table \[tab:tolerances\], the solver needed two outer iterations to converge (and a third to check for convergence) in each time step. The number of GMRES iterations per outer iteration is around 3-4 on average. At the start of a time step, when the guess for the material and radiation energies is based on the values from the previous time step, the number is generally higher, around 6-8. On the last outer iteration, the number of GMRES solves is 1, indicating that the guess from the previous iteration satisfies the diffusion equation. The solution to the problem is the same at the start and end times, so despite adding and subtracting energy in certain regions of the problem, the initial and final energies should be equal. The relative error between the initial and final energy in 2D ranges from $6.35\times10^{-14}$ for the 256 points to $1.69\times10^{-12}$ for 16,384 points. For 3D, the relative error ranges from $1.33\times10^{-13}$ for 4,096 points to $4.53\times10^{-12}$ for 2,097,152 points. In both cases, the relative error is on the order of the tolerance for the radiation solve of $10^{-12}$ and much lower than the tolerance of the coupled radiation and material energy solve of $10^{-8}$. Conclusions and future work\[sec:conclusions\] ============================================== The thermal radiative transfer equations with grey diffusion is discretized fully implicitly in time and solved using an efficient nonlinear elimination method that leads to fast convergence of the emission source. A standard SPH diffusion derivative is applied to the modified equations to form a fully discretized set of equations that are conservative in energy. The diffusion term is solved using optimized linear solvers, which permits good parallel efficiency. The code is verified by comparison to three test problems with known solutions. The code gets the correct infinite medium behavior for either a hot material emitting radiation or a cold material absorbing radiation, with the expected first-order convergence in time. The results for a one-dimensional semianalytic Marshak wave problem are consistent with second-order spatial convergence. Finally, the code is run in 2D and 3D for a manufactured problem with sinusoidal radiation and material energy solutions traveling in opposite directions, which likewise exhibits second-order convergence in space and first-order convergence in time. The diffusion discretization for SPH is stable and performs well, but lacks zeroth-order consistency. Other methods, such as reproducing kernel particle methods [@liu1995reproducing] and moving least squares particle hydrodynamics [@dilts1999moving; @dilts2000moving], do not lack the zeroth-order consistency and perform better near boundaries. Conservative reproducing kernel smoothed particle hydrodynamics is implemented in the SPH code used for these results [@frontiere2017crksph], and a similar diffusion discretization would make the hydrodynamics and radiation discretizations consistent. Other discretizations would also make the application of vacuum or incoming radiation boundary conditions more feasible. Acknowledgements {#acknowledgements .unnumbered} ================ This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344. This document was prepared as an account of work sponsored by an agency of the United States government. Neither the United States government nor Lawrence Livermore National Security, LLC, nor any of their employees makes any warranty, expressed or implied, or assumes any legal liability or responsibility for the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately owned rights. Reference herein to any specific commercial product, process, or service by trade name, trademark, manufacturer, or otherwise does not necessarily constitute or imply its endorsement, recommendation, or favoring by the United States government or Lawrence Livermore National Security, LLC. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States government or Lawrence Livermore National Security, LLC, and shall not be used for advertising or product endorsement purposes. ---------------------- ------------ ------------ ------------ ------------ Material Radiation Material Radiation Infinite medium $10^{-14}$ $10^{-14}$ $10^{-12}$ $10^{-12}$ Marshak wave $10^{-12}$ $10^{-12}$ $10^{-8}$ $10^{-8}$ Manufactured problem $10^{-12}$ $10^{-12}$ $10^{-8}$ $10^{-8}$ ---------------------- ------------ ------------ ------------ ------------ : Convergence tolerances for the results. Then inner tolerance is for the independent radiation and material energy solves, while the outer tolerance is for the combined solve.[]{data-label="tab:tolerances"} Case $e_{0}$ $E_{0}$ $T_{e,0}$ $T_{r,0}$ $\Delta t_{init}$ --------------- ----------- ------------ ----------- ----------- ------------------- Hot material $1.0$ $10^{-16}$ $1.0$ $10^{-4}$ $10^{-20}$ Hot radiation $10^{-4}$ $1.0$ $10^{-4}$ $1.0$ $10^{-7}$ : Initial conditions for infinite medium problem for two cases, one starting with a high thermal temperature and low radiation temperature and one with the temperatures reversed. See Sec. \[subsec:infinite-medium\] for units.[]{data-label="tab:infinite-initial"} ![Convergence of the numerical solution for the infinite medium problem to the reference solution with decreasing time step. The error is calculated using Eq. (\[eq:infinite-error\]).[]{data-label="fig:infinite-convergence"}](figs/infinite_error){width="47.00000%"} Dimension Points Procs Points/proc GMRES/Outer Outer/Step Wall time (s) Time$\times$Proc/Point ----------- ----------- ------- ------------- ------------- ------------ --------------- ------------------------ 256 36 7 3.66 3 59 8.29 1,024 36 28 4.66 3 79 2.78 4,096 36 113 3.33 3 261 2.29 16,384 36 455 3.66 3 817 1.79 4096 36 113 3.66 3 815 7.21 32,768 36 910 4.66 3 5,610 6.20 262,144 288 910 3.33 3 7,660 8.41 2,097,152 576 3,640 3.66 3 31,975 8.78 : Timing for the manufactured problem in 3D.[]{data-label="tab:manufactured-timing"}	{ "pile_set_name": "ArXiv" }
--- abstract: 'Recently solar flares exhibiting a double J-shaped ribbons in the lower solar atmosphere have been paid increasing attention in the context of extending the two-dimensional standard flare model to three dimensions, as motivated by the spatial correlation between photospheric current channels and flare ribbons. Here we study the electric currents through the photospheric area swept by flare ribbons (termed synthesized ribbon area or SRA), with a sample of 71 two-ribbon flares, of which 36 are J-shaped. Electric currents flowing through one ribbon are highly correlated with those through the other, therefore belonging to the same current system. The non-neutrality factor of this current system is independent of the flare magnitude, implying that both direct and return currents participate in flares. J-shaped flares are distinct from non-J-shaped flares in the following aspects: 1) Electric current densities within J-shaped SRA are significantly smaller than those within non-J-shaped SRA, but J-shaped SRA and its associated magnetic flux are also significantly larger. 2) Electric currents through SRA are positively correlated with the flare magnitude, but J-shaped flares show stronger correlation than non-J-shaped flares. 3) The majority (75%) of J-shaped flares are eruptive, while the majority (86%) of non-J-shaped flares are confined; accordingly, hosting active regions of J-shaped flares are more likely to be sigmoidal than non-J-shaped flares. Thus, J-shaped flares constitute a distinct subset of two-ribbon flares, probably the representative of eruptive ones. Further, we found that combining SRA and its associated magnetic flux has the potential to differentiate eruptive from confined flares.' author: - Yuwei He - Rui Liu - Lijuan Liu - Jun Chen - Wensi Wang - Yuming Wang title: 'Electric Currents through J-shaped and Non-J-shaped Flare Ribbons' --- Introduction ============ Solar flares are among the most energetic phenomena in our solar system. The flare emission spans the whole range of electromagnetic spectrum, often associated with particle acceleration and coronal mass ejections. The physical mechanism of solar flares is not only important for forecasting the space weather at Earth, but for understanding similar physical processes in stellar flares as well as on planetary magnetospheres [e.g., @Zhong2020] and active galactic nuclei . Although the complexity and diversity of flare phenomena makes it impossible to build a ‘universal’ flare model that is capable of explaining all observational aspects in all events, the standard or CSHKP flare model [@KoppRA1976; @Sturrock1966; @Carmichael1964; @Hirayama1974] is successful in explaining the major characteristics of two-ribbon eruptive flares, and has formed the basis for our understanding of solar flares for decades . Solar eruptive phenomena draw energy from coronal magnetic fields [@Forbes2000]. The magnetic field $\mathbf{B}$ can be always decomposed into a current-free, potential component $\mathbf{B}_p$ and a current-carrying, non-potential component $\mathbf{B}_c$, so that the magnetic energy $E_m$ in a volume $V$ can be written as [@Sakurai1981] $$E_m=\int_V \frac{B^2}{8\pi}\,dV=\frac{1}{8\pi}\int_VB_p^2\,dV+\frac{1}{2c}\int_V \mathbf{A}_c\cdot\mathbf{J}\,dV,$$ where $\mathbf{B}_c=\nabla\times\mathbf{A}_c$ and $\mathbf{J}=\frac{c}{4\pi}\nabla\times\mathbf{B}_c$, as $\nabla\times\mathbf{B}_p=0$. The first term is the energy of the potential field, which is produced by current sources located within the interior of the Sun. The energy powering solar eruptions can only be contained in the second term that is associated with electric currents in the solar corona. The most rapid release of this ‘free’ magnetic energy is manifested as flares. Because of the huge magnetic Reynolds number in the corona, electric currents must be concentrated into small regions such as narrow current sheets, across which magnetic connectivity changes rapidly, generally known as quasi-separatrix layers [QSLs; @Demoulin2006]. QSLs are preferential sites for magnetic reconnection, which drives the processes of plasma heating and particle acceleration in flares. Due to the requirement of the electric current continuity, concentrations of electric currents in the photosphere are believed to be the imprints of coronal current sheets . During the flare, the release of free magnetic energy stored in the corona can be attributed to a geometric reconfiguration of the current paths, with the magnitudes of currents at the footpoints of the current system being fixed to a large extent [@Melrose2017]. During some flares, however, horizontal electric currents tend to concentrate at lower altitudes around the polarity inversion line (PIL) than before the eruption [e.g., @Sun2012; @Liu2012; @Liu2014], while vertical electric currents tend to increase in localized ribbons [e.g., @Janvier2014; @Janvier2016; @Sharykin2020]. From the observational perspective, only the vertical component of the electric current density $J_z$ can be derived from photospheric vector magnetograms, which are normally limited to a single height. With the measurements of photospheric transverse magnetic fields becoming more and more reliable, it has been well known that there exists a close spatial relationship between the vertical electric currents at the photosphere and the deposition sites of flare energy at the chromosphere as represented by bright H$\alpha$ kernels [e.g., @Moreton1968; @Yuanzhang1987; @Romanov1990; @Leka1993; @deLaBeaujardiere1993; @Sharykin2014], or UV/EUV emission [e.g., @Janvier2014; @Janvier2016; @Sharykin2020], or hard X-ray emission [e.g., @deLaBeaujardiere1993; @Canfield1992; @lijing1997; @Musset2015; @Sharykin2020]. However, these flare kernels are often not exactly co-spatial with the locations of strongest electric current densities, but adjacent to the current channels [e.g., @Romanov1990; @Leka1993; @deLaBeaujardiere1993; @Canfield1992; @lijing1997], which might result from magnetic reconnection in a quadrupolar magnetic field [@Aschwanden1999]. Thus, the distribution of pre-flare electric current densities in the photosphere provides important clues to the coronal currents accessible to flares. More recently, aided by nonlinear force-free field or MHD modeling of the coronal magnetic field, it has been demonstrated that H$\alpha$ and UV flare ribbons often coincide with the footprints of QSLs [e.g., @Liu2014; @Liu2016SR; @Janvier2014; @Janvier2016; @Liu2018; @Su2018; @Jiang2018]. In particular, the footprints of the QSLs wrapping around a magnetic flux rope, the core structure of coronal mass ejections [@Vourlidas2013; @Georgoulis2019], correspond to a pair of J-shaped ribbons of high electric current densities; the hooks of J-shaped ribbons outline the edge of the rope legs [@Janvier2014; @Janvier2016; @Wang2017]. Motivated by these observational and modeling results, it has been proposed that the two-dimensional standard model can be extended to three dimensions to address the shape, location, and dynamics of flares with a double J-shaped ribbons [@Aulanier2013; @Aulanier2012; @Janvier2013; @Janvier2015]. However, not all two-ribbon flares are J-shaped; in fact, many typical two-ribbon flares [e.g., Figure 1 in @Qiu2010] do not exhibit hooks at the ends of flare ribbons. Hence, it is obscure whether a canonical J-shaped flare, which breaks the translational symmetry along ribbons but introduce a 2-fold rotational symmetry (i.e., the ribbon morphology does not change by a rotation of $180^\circ$), can represent classic two-ribbon flares in the general 3D context. In this paper, we tackle on this question by investigating the distribution of pre-eruption photospheric electric currents associated with flare ribbons. Since the two ribbons often move away from each other during the impulsive phase of flares, we consider the photospheric electric currents through the total area swept by flare ribbons, termed synthesized ribbon area (SRA) hereafter, instead of the area covered by flare ribbons at any time instant. In the rest of the paper, we present the methods in §\[sec:method\] and the statistical analysis in §\[sec:statistics\], and make concluding remarks in §\[sec:conclusion\]. Data and Methods {#sec:method} ================ Selection of Events ------------------- We selected two-ribbon flares of GOES-class M1.0 and above from the database provided by @Kazachenko2017. The database covers all flare ribbon events between 2010 April and 2016 April recorded by the Solar Dynamics Observatory [SDO; @Pesnell2012], corresponding to GOES-class C1.0 and above within 45$^\circ$ from the central meridian. By visually examining flare ribbons observed by the ultraviolet 1600 [Å]{} passband of the Atmospheric Imaging Assembly [AIA; @Lemen2012] onboard SDO, we categorized the selected events into J-shaped flares (Figure \[ffg:J-ribbon\]a), if at least one of the two ribbons exhibits a hooked shape at the ribbon end, and non-J-shaped flares (Figure \[ffg:non-J-ribbon\]a), if neither ribbon has a pronounced hook. It is interesting that J-shaped flares do not necessarily occur in a sigmoidal active region (indicated by an asterisk in Tables \[tab:Non-J-shaped flares\] and \[tab:J-shaped flares\]), which exhibits an overall S shape with two opposing bundles of coronal loops; but non-J-shaped flares can sometimes be hosted by a sigmoidal active region. In addition, we excluded two types of events: 1) those with material ejection visible near flare ribbons in AIA 1600 [Å]{} images; and 2) those with complex or remote ribbons beyond two major ribbons in AIA 1600 [Å]{}. Both scenarios could compromise a clear identification of the two major ribbons. In total, we have 36 J-shaped and 35 non-J-shaped flares. Calculation of Electric Current Densities ----------------------------------------- We calculated the distribution of current density $J_z$ in active regions of interest by Amepére’s law, $$\mu_0J_z= \frac{\partial B_y}{\partial x} - \frac{\partial B_x}{\partial y}.$$ $B_{x,y}$ is obtained by the Helioseismic and Magnetic Imager [HMI; @Scherrer2012] onboard SDO. For each flare, we worked with a vector magnetogram taken immediately before the flare onset. These vector magnetograms are disambiguated and de-projected to the heliographic coordinates with a Lambert (cylindrical equal area; CEA) projection method, also known as the Space-Weather HMI Active Region Patches (SHARP) data. Computation of electric currents based on the above equation is subject to several uncertainties that are difficult to quantify, e.g., the 180$^\circ$ ambiguity in the horizontal field direction, and these uncertainties can be further amplified by partial derivatives. Considering the uncertainty in measuring the transverse component of photospheric magnetic field [$\sim\,$100 G; @Hoeksema2014] and SHARP’s pixel size (0.36 Mm), we estimated that the uncertainty of $J_z$ is about 22 mA m$^{-2}$. As a comparison, the average and standard deviation of $\|J_z\|$ in a box area in the quiet region is listed in the column ‘noise’ in Tables \[tab:Non-J-shaped flares\] and \[tab:J-shaped flares\], and the box typically occupies the lower left corner (1/12 in length and 1/8 in width) of the SHARP map. One can see that in all cases the value of 22 mA m$^{-2}$, which significantly exceeds the average $\|J_z\|$ in the quiet region ($\sim\,$10 mA m$^{-2}$), is a reliable representative of the noise level. Hence we considered only $J_z$ above 22 mA m$^{-2}$ in the statistical analysis below (§\[sec:statistics\]). We have also carried out the same analysis without considering this threshold value, but got similar results and reached the same conclusions. Identification of Flare Ribbons ------------------------------- We used AIA 1600 [Å]{} images to identify flare ribbons. The AIA 1600 [Å]{} passbands provides full-disk images of the lower atmosphere at a temporal cadence of 24 s and a spatial resolution of $1''.2$. To match the map of electric current density with that of flare ribbons, we remapped 1600 [Å]{} images from the CCD coordinates to the CEA coordinates of the corresponding SHARP data. We detected flare ribbons by setting a threshold of about 3$\sim$5 times the average brightness of an AIA 1600 [Å]{} image immediately before the onset of each flare under investigation. We fine-tuned the exact threshold value case by case so that flare ribbons are visually captured as accurately as possible while as many as bright plages are excluded. This threshold value is varied by 15% to be taken as the upper and lower threshold. We flagged the pixels with brightness above the upper/lower threshold in each 1600 [Å]{} image, and collected all the flagged pixels during the interval from the GOES flare start to end time to construct ‘synthesized’ flare ribbons. Meanwhile we removed isolated pixels flagged outside the main ribbon area, and then constructed a binary mask for the synthesized flare ribbons by setting the flagged pixels to be unity and the rest to be zero. Applying the upper and lower threshold yields two slightly different ribbon masks and different ribbon areas. The average area is taken as the synthesized ribbon area (SRA). In the following, when calculating an electric-current parameter, we also apply the two ribbon masks that are associated with the upper and lower threshold, respectively, and then take the average to be the final result and half of the range to be the uncertainty. Calculation of Electric Current Parameters ------------------------------------------ Multiplying the map of pre-flare $J_z$ by the synthesized flare-ribbon mask, we obtained the distribution of $J_z$ within SRA (Figures \[ffg:J-ribbon\]c and \[ffg:non-J-ribbon\]c) as well as the mean ($\overline{\|J_z\|}$) and median ($\widetilde{\|J_z\|}$) of the absolute values of $J_z$. The direct current density $J_\mathrm{DC}$ is selected from $J_z$ whose sign is consistent with the dominant current through the ribbon, as opposed to the return current density $J_\mathrm{RC}$. Note that for simplicity, here the signs of direct and return currents are not defined according to the sign of magnetic helicity of the host active region as in MHD simulations . In observation, it is more difficult to determine the sign of helicity than that of dominant current. Summing up $J_\mathrm{DC}$ and $J_\mathrm{RC}$ on a flare ribbon, we obtained the direct current $I_\mathrm{DC}$ and return current $I_\mathrm{RC}$, respectively. The net current $I_\mathrm{net}$ is the algebraic sum of $J_z$ over SRA, as opposed to the unsigned current $I_\mathrm{uns}$, which is the sum of $\|J_z\|$. The above quantities are listed in Tables \[tab:Non-J-shaped flares\] and \[tab:J-shaped flares\]. We adopted the superscript ‘+’ or ‘-’ to indicate that a parameter is given for a ribbon located on the positive or negative polarity side of the PIL. When the superscript is dropped, electric current densities are sampled through both ribbons. The above defined parameters of electric current densities (Fig. \[fig:ribbon\_vs\_ribbon\_J\]) and currents (Fig. \[fig:ribbon\_vs\_ribbon\_I\]) flowing through the SRA of positive polarity are highly correlated with that of negative polarity. Hence it is reasonable to assume that they belong to the same current system, i.e., the electric currents flow out of one ribbon and into the other. Thus, following @Georgoulis2012 we defined a non-neutrality factor $I_\mathrm{nn}$, $$\label{eq:Inn} I_\mathrm{nn}=\frac{1}{2}\left(\frac{\|I_\mathrm{net}^+\|}{\|I_\mathrm{DC}^+\|}+\frac{\|I_\mathrm{net}^-\|}{\|I_\mathrm{DC}^-\|}\right),$$ so that $I_\mathrm{nn}=0$ if electric currents through the SRA are balanced separately on either side of the PIL, i.e., neutralized, and that $I_\mathrm{nn}=1$ if electric currents through the SRA are unneutralized. We noticed that [@LiuY2017] used $\|I_\mathrm{DC}/I_\mathrm{RC}\|$ to measure the degree of current neutralization in active regions, but this parameter is less tractable than $I_\mathrm{nn}$ when $I_\mathrm{RC}$ is small. Statistics {#sec:statistics} ========== . To evaluate whether J-shaped flares are distinct from non-J-shaped flares in terms of electric currents through SRA, we performed Welch’s $t$-test to determine whether two samples of different sizes and variances belong to the same population, by comparing their sample means. The test defines a statistic $t$ as follows, $$t=\frac{\overline{X}_1-\overline{X}_2}{\sqrt{\frac{s_1^2}{N_1}+\frac{s_2^2}{N_2}}},$$ where $\overline{X}_1$, $s_1^2$, and $N_1$ are the 1st sample mean, sample variance, and sample size, respectively. The null hypothesis that two populations have equal means can be rejected if the resultant $p$-value is small enough. As a rule of thumb, one may deem that the two populations have significantly different means if $p\le 0.01$, as there is less than a 1% probability that the null hypothesis is correct. We found that the distributions of most current-density parameters are different for the two flare categories, which is manifested by the statistical significance in Welch’s $t$-test. Specifically, both median and mean $\|J_z\|$ ($\widetilde{\|J_z\|}$ and $\overline{\|J_z\|}$, respectively) through J-shaped ribbons are smaller than those through non-J-shaped ribbons (Figure \[fig:student\_test\](a & b)). A similar result is found for the mean and median current densities for DC (Figure \[fig:student\_test\](c & d)) and RC (not shown). But we found no statistically significant differences between J-shaped and non-J-shaped flares in electric currents through SRA, no matter if it is the direct, return, unsigned, or net current, or the dominance of direct current as measured by $I_\mathrm{DC}^+/I_\mathrm{RC}^+$ and $I_\mathrm{DC}^-/I_\mathrm{RC}^-$, or the current imbalance as measured by $I_\mathrm{DC}^+/I_\mathrm{DC}^-$ and $I_\mathrm{RC}^+/I_\mathrm{RC}^-$. Note all the ratios are taken in their absolute values. However, J-shaped flares have significantly larger SRA (Figure \[fig:student\_test\]e) as well as larger unsigned magnetic flux through it $(\|\Phi^+\| + \|\Phi^-\|)/2$ (averaged over two ribbons; Figure \[fig:student\_test\]f) than non-J-shaped flares. We further examined the correlations between the flare magnitude as measured by the GOES class and the above parameters. None of the current density parameters, $\|J_z\|$, $\|J_\mathrm{RC}\|$, or $\|J_\mathrm{DC}\|$ within SRA, shows significant correlation with the flare magnitude. Averaged over two ribbons, the mean net current $(\|I_\mathrm{net}^+\|+\|I_\mathrm{net}^-\|)/2$, mean unsigned current $(\|I_\mathrm{uns}^+\|+\|I_\mathrm{uns}^-\|)/2$, and mean direct current $(\|I_\mathrm{DC}^+\|+\|I_\mathrm{DC}^-\|)/2$ are all positively correlated with the flare magnitude; but the correlation coefficient for J-shaped flares is consistently larger than non-J-shaped flare (Figure \[fig:I\_vs\_goes\](a–c)). The flare magnitude is also positively correlated with both SRA (Figure \[fig:I\_vs\_goes\]e) and magnetic flux through SRA $(\|\Phi^+\| + \|\Phi^-\|)/2$ (Figure \[fig:I\_vs\_goes\]f), which is consistent with @Kazachenko2017 and previous studies. On the other hand, it is surprising that the non-neutrality factor $I_\mathrm{nn}$ is almost independent of the flare magnitude (Figure \[fig:I\_vs\_goes\]d). Discussion and Conclusion {#sec:conclusion} ========================= . To summarize, we confirmed that electric currents flowing through one ribbon are highly correlated with those through the other, hence belonging to the same current system pertinent to the two-ribbon flare. Importantly we found that J-shaped flares are distinct from non-J-shaped flares in a few aspects as follows. First, $\|J_z\|$, $\|J_\mathrm{DC}\|$, and $\|J_\mathrm{RC}\|$ within the SRA of J-shaped flares are significantly smaller than those of non-J-shaped flares, but the SRA of J-shaped flares as well as the associated magnetic flux is significantly larger than that of non-J-shaped flares. In other words, smaller current densities in J-shaped flare ribbons are compensated by larger ribbon area, which explains why J-shaped flares are similar to non-J-shaped ones in terms of electric currents through SRA. Second, the magnitudes of direct, net, and unsigned currents through SRA are all positively correlated with the flare magnitude, but the correlation coefficient for J-shaped flares is generally larger than that for non-J-shaped flares. Third, the majority (27 of 36; 75%) of J-shaped flares are eruptive, while the majority (30 of 35; 86%) of non-J-shaped flares are confined; among the 28 hosting active region for the J-shaped flares, 12 are sigmoidal regions [^1], yet among the 21 hosting active regions for the non-J-shaped flares, only 3 are sigmoidal. These differences highlight J-shaped flares as a distinct subset of two-ribbon flares, probably the representative of eruptive ones. These differences also imply a different magnetic configuration in J-shaped from non-J-shaped flares, which will be further explored in future work and should be taken into account in developing a three-dimensional model for two-ribbon flares as a whole. Furthermore, the flare magnitude is independent of non-neutrality factor $I_\mathrm{nn}$. In other words, it does not matter whether the electric currents flowing through flare ribbons are neutralized ($I_\mathrm{nn}=0$), i.e., balanced separately on each ribbon, or, unneutralized ($I_\mathrm{nn}=1$), i.e., the return current is negligible relative to the direct current through each ribbon. This may imply that both direct and return current are involved in flares. With a weak positive correlation, stronger direct current is associated with $I_\mathrm{nn}$ being closer to unity (left panel of Figure \[fig:eruptive-confined\]), but neither direct current nor $I_\mathrm{nn}$ is able to differentiate eruptive from confined flares. We further checked other parameters and any two of their combinations. The only pair that stands out is the SRA and the associated magnetic flux (right panel of Figure \[fig:eruptive-confined\]). Most of the flares are eruptive, if SRA exceeds $10^{15}$ m$^2$ and average unsigned magnetic flux through each ribbon exceeds $4\times10^{21}$ Mx. In contrast, if both parameters are below the above-mentioned threshold values, most of the flares are confined. Using $\|I_\mathrm{DC}/I_\mathrm{RC}\|$ to measure the degree of current neutralization, [@LiuY2017] found in a sample of four active regions that the two CME-producing regions are strongly nonneutralized ($\|I_\mathrm{DC}/I_\mathrm{RC}\|\approx2$); the other two regions that did not produce CMEs are almost perfectly neutralized, but can still produce strong flares. They hence speculated that the presence or absence of a double-J-pattern of direct currents around the PIL may indicate whether or not an active region will produce CMEs. More recent studies on larger samples found that CME/flare-productive active regions are more likely to be non-neutralized than quiet regions . In stead of examining the whole active region, our statistical study focuses on electric currents flowing though flare ribbons, in other words, the currents accessible to flares. This study shows that $I_\mathrm{nn}$ is similar to $\|I_\mathrm{DC}/I_\mathrm{RC}\|$ in its poor correlation with the flare magnitude, consistent with @LiuY2017, and that J-shaped flares are more likely to be eruptive than non-J-shaped ones. Moreover, we found that combining the ribbon area and its associated magnetic flux has the potential to differentiate eruptive from confined flares. Further study is needed to clarify the connection between electric currents in an active region and its productivity in flares and CMEs. The authors thank the SDO consortium for the excellent data. This work was supported by NSFC (Grant No. 41761134088, 41774150, and 11925302), CAS Key Research Program (Grant No. KZZD-EW-01-4), the fundamental research funds for the central universities, and the Strategic Priority Program of the Chinese Academy of Sciences (Grant No. XDB41000000) natexlab\#1[\#1]{} \[1\][[\#1](#1)]{} \[1\][doi: [](http://doi.org/#1)]{} \[1\][[](http://ascl.net/#1)]{} \[1\][[](https://arxiv.org/abs/#1)]{} , M. J., [Kosugi]{}, T., [Hanaoka]{}, Y., [Nishio]{}, M., & [Melrose]{}, D. B. 1999, , 526, 1026, , G., [D[é]{}moulin]{}, P., [Schrijver]{}, C. J., [et al.]{} 2013, , 549, A66, , G., [Janvier]{}, M., & [Schmieder]{}, B. 2012, , 543, A110, , E. A., & [Sun]{}, X. 2020, , 893, 123, , D., & [Cappi]{}, M. 2019, , 628, A5, , R. C., [Hudson]{}, H. S., [Leka]{}, K. D., [et al.]{} 1992, , 44, L111 , H. 1964, [A Process for Flares]{}, Vol. 50, 451 , J. F., [Canfield]{}, R. C., & [Leka]{}, K. D. 1993, , 411, 378, , P. 2006, Advances in Space Research, 37, 1269, , G. D., & [Pevtsov]{}, A. A. 2018, in Electric Currents in Geospace and Beyond, ed. A. [Keiling]{}, O. [Marghitu]{}, & M. [Wheatland]{}, Vol. 235, 43–65, , T. G. 2000, , 105, 23153, , M. K., [Nindos]{}, A., & [Zhang]{}, H. 2019, Philosophical Transactions of the Royal Society of London Series A, 377, 20180094, , M. K., [Titov]{}, V. S., & [Miki[ć]{}]{}, Z. 2012, , 761, 61, , T. 1974, , 34, 323, , J. T., [Liu]{}, Y., [Hayashi]{}, K., [et al.]{} 2014, , 289, 3483, , M., [Aulanier]{}, G., [Bommier]{}, V., [et al.]{} 2014, , 788, 60, , M., [Aulanier]{}, G., & [D[é]{}moulin]{}, P. 2015, , 290, 3425, , M., [Aulanier]{}, G., [Pariat]{}, E., & [D[é]{}moulin]{}, P. 2013, , 555, A77, , M., [Savcheva]{}, A., [Pariat]{}, E., [et al.]{} 2016, , 591, A141, , C., [Zou]{}, P., [Feng]{}, X., [et al.]{} 2018, , 869, 13, , M. D., [Lynch]{}, B. J., [Welsch]{}, B. T., & [Sun]{}, X. 2017, , 845, 49, , R. A., & [Pneuman]{}, G. W. 1976, , 50, 85, , K. D., [Canfield]{}, R. C., [McClymont]{}, A. N., [et al.]{} 1993, , 411, 370, , J. R., [Title]{}, A. M., [Akin]{}, D. J., [et al.]{} 2012, , 275, 17, , J., [Metcalf]{}, T. R., [Canfield]{}, R. C., [W[ü]{}lser]{}, J.-P., & [Kosugi]{}, T. 1997, , 482, 490, , Y., & [Gaizauskas]{}, V. 1987, , 109, 81 , C., [Deng]{}, N., [Liu]{}, R., [et al.]{} 2012, , 745, L4, , R., [Chen]{}, J., & [Wang]{}, Y. 2018, Science China Physics, Mechanics, and Astronomy, 61, 69611, , R., [Chen]{}, J., [Wang]{}, Y., & [Liu]{}, K. 2016, Scientific Reports, 6, 34021, , R., [Titov]{}, V. S., [Gou]{}, T., [et al.]{} 2014, , 790, 8, , Y., [Sun]{}, X., [T[ö]{}r[ö]{}k]{}, T., [Titov]{}, V. S., & [Leake]{}, J. E. 2017, , 846, L6, , D. B. 2017, Journal of Geophysical Research (Space Physics), 122, 7963, , G. E., & [Severny]{}, A. B. 1968, , 3, 282, , S., [Vilmer]{}, N., & [Bommier]{}, V. 2015, , 580, A106, , W. D., [Thompson]{}, B. J., & [Chamberlin]{}, P. C. 2012, , 275, 3, , E. R., & [Forbes]{}, T. G. 2002, , 10, 313, , J., [Liu]{}, W., [Hill]{}, N., & [Kazachenko]{}, M. 2010, , 725, 319, , V. A., & [Tsap]{}, T. T. 1990, , 34, 656 , T. 1981, , 69, 343, , P. H., [Schou]{}, J., [Bush]{}, R. I., [et al.]{} 2012, , 275, 207, , B., & [Aulanier]{}, G. 2018, in Electric Currents in Geospace and Beyond, ed. A. [Keiling]{}, O. [Marghitu]{}, & M. [Wheatland]{}, Vol. 235, 391–406, , I. N., & [Kosovichev]{}, A. G. 2014, , 788, L18, , I. N., [Zimovets]{}, I. V., & [Myshyakov]{}, I. I. 2020, , 893, 159, , P. A. 1966, , 211, 695, , Y., [Liu]{}, R., [Li]{}, S., [et al.]{} 2018, , 855, 77, , X., [Hoeksema]{}, J. T., [Liu]{}, Y., [et al.]{} 2012, , 748, 77, , P. 2019, , 486, 4936, , A., [Lynch]{}, B. J., [Howard]{}, R. A., & [Li]{}, Y. 2013, , 284, 179, , W., [Liu]{}, R., [Wang]{}, Y., [et al.]{} 2017, Nature Communications, 8, 1330, , J., [Wei]{}, Y., [Lee]{}, L. C., [et al.]{} 2020, , 893, L18, [cccccccccccccc]{} 1 & 2011-03-09T14:02 & N9W6 & 11166 & C & 28.98 & 30.03 & 25.84 & 26.76 & 3.10E+11 & -1.59E+11 & -4.13E+11 & 2.18E+11 & 6.67 $\pm$ 5.29\ 2 & 2011-04-22T04:57 & S18E43 & 11195 & C & 27.25 & 26.45 & 25.06 & 24.68 & 2.99E+11 & -1.68E+11 & -4.04E+11 & 2.98E+11 & 7.33 $\pm$ 5.48\ 3 & 2011-11-06T06:35 & N21E31 & 11339 & C & 40.02 & 35.04 & 30.42 & 29.68 & -2.23E+12 & 1.61E+12 & 1.83E+12 & -8.33E+11 & 7.87 $\pm$ 6.00\ 4 & 2011-12-26T02:27 & S21W33 & 11387 & E & 36.96 & 35.01 & 32.33 & 29.37 & -1.33E+12 & 7.42E+11 & 1.25E+12 & -4.77E+11 & 10.12 $\pm$ 7.99\ 5 & 2011-12-31T16:26 & S26E42 & 11389 & C & 35.29 & 40.32 & 29.53 & 30.80 & -8.65E+11 & 6.69E+11 & 1.80E+12 & -5.04E+11 & 8.44 $\pm$ 6.66\ 6 & 2012-03-06T00:28 & N16E41 & 11429\* & C & 38.32 & 45.62 & 32.45 & 36.34 & -1.70E+12 & 1.55E+12 & 4.61E+12 & -1.47E+12 & 8.99 $\pm$ 6.97\ 7 & 2012-03-06T01:44 & N17E41 & 11429\* & C & 41.11 & 44.64 & 34.29 & 34.23 & -2.96E+12 & 1.69E+12 & 5.26E+12 & -1.77E+12 & 9.15 $\pm$ 7.05\ 8 & 2012-03-06T22:53 & N17E35 & 11429\* & C & 40.85 & 50.77 & 34.16 & 39.45 & -3.68E+12 & 1.31E+12 & 6.65E+12 & -1.58E+12 & 8.81 $\pm$ 6.74\ 9 & 2012-03-06T12:41 & N18E36 & 11429\* & C & 34.96 & 46.73 & 30.29 & 37.24 & -3.16E+12 & 1.51E+12 & 5.78E+12 & -1.39E+12 & 10.54 $\pm$ 7.88\ 10 & 2012-06-06T20:06 & S19W5 & 11494 & E & 30.80 & 33.40 & 27.45 & 28.30 & -6.31E+11 & 5.56E+11 & 9.60E+11 & -4.01E+11 & 9.05 $\pm$ 6.87\ 11 & 2012-06-13T13:17 & S16E18 & 11504\* & E & 34.83 & 37.93 & 28.59 & 31.25 & 1.28E+12 & -9.24E+11 & -3.54E+12 & 1.72E+12 & 7.01 $\pm$ 5.54\ 12 & 2012-07-14T04:58 & S22W30 & 11521 & C & 33.40 & 31.86 & 28.00 & 29.97 & 7.47E+11 & -3.67E+11 & -1.07E+12 & 2.50E+11 & 7.56 $\pm$ 5.98\ 13 & 2012-11-27T21:26 & S14W41 & 11620 & C & 45.00 & 35.05 & 33.97 & 29.88 & 3.03E+12 & -1.14E+12 & -1.34E+12 & 8.72E+11 & 9.04 $\pm$ 7.07\ 14 & 2013-08-12T10:41 & S17E19 & 11817\* & E & 36.53 & 50.98 & 31.57 & 37.60 & -1.61E+12 & 3.51E+11 & 3.05E+12 & -6.00E+11 & 9.71 $\pm$ 7.59\ 15 & 2013-10-22T00:22 & N6E17 & 11875 & C & 37.33 & 35.00 & 31.24 & 30.13 & 1.26E+12 & -1.24E+12 & -1.03E+12 & 7.26E+11 & 9.67 $\pm$ 7.61\ 16 & 2013-10-24T00:08 & N8W11 & 11875 & C & 39.12 & 40.47 & 32.38 & 31.89 & 2.58E+12 & -1.73E+12 & -3.00E+12 & 2.05E+12 & 8.56 $\pm$ 6.41\ 17 & 2013-10-24T10:09 & N6W14 & 11875 & C & 36.56 & 40.75 & 31.68 & 31.37 & 1.93E+12 & -1.27E+12 & -2.34E+12 & 1.45E+12 & 6.13 $\pm$ 4.83\ 18 & 2013-11-16T04:53 & S19W29 & 11900 & C & 39.82 & 46.78 & 32.97 & 34.14 & -1.76E+12 & 1.78E+11 & 1.87E+12 & -2.20E+11 & 7.80 $\pm$ 6.28\ 19 & 2014-02-01T01:25 & S11E26 & 11967 & C & 33.61 & 54.77 & 28.59 & 40.56 & -1.83E+12 & 7.78E+11 & 3.17E+12 & -1.63E+12 & 7.07 $\pm$ 5.66\ 20 & 2014-02-02T22:04 & S13E5 & 11967 & C & 41.83 & 42.75 & 35.92 & 34.25 & -3.87E+12 & 1.69E+12 & 3.57E+12 & -1.53E+12 & 7.64 $\pm$ 6.62\ 21 & 2014-02-04T01:23 & S13W14 & 11967 & C & 48.31 & 61.78 & 38.07 & 44.68 & -4.42E+12 & 8.43E+11 & 4.73E+12 & -5.35E+11 & 9.49 $\pm$ 7.40\ 22 & 2014-02-04T09:49 & S13W12 & 11967 & C & 50.76 & 54.57 & 40.46 & 40.12 & -2.92E+12 & 8.55E+11 & 3.40E+12 & -5.79E+11 & 8.70 $\pm$ 6.70\ 23 & 2014-02-11T16:51 & S13E12 & 11974 & E & 36.11 & 38.15 & 31.12 & 30.31 & 2.31E+12 & -1.38E+12 & -2.30E+12 & 8.90E+11 & 7.65 $\pm$ 6.29\ 24 & 2014-02-13T08:12 & S12W13 & 11974 & C & 46.40 & 41.96 & 34.43 & 33.43 & 3.31E+12 & -2.71E+12 & -2.66E+12 & 2.56E+12 & 9.04 $\pm$ 7.14\ 25 & 2014-10-20T09:11 & S13E43 & 12192 & C & 39.38 & 44.86 & 32.87 & 33.12 & -2.60E+12 & 1.88E+12 & 5.16E+12 & -3.20E+12 & 8.33 $\pm$ 6.40\ 26 & 2014-10-20T20:04 & S13E43 & 12192 & C & 31.65 & 36.85 & 28.04 & 32.86 & -7.64E+11 & 6.68E+11 & -6.55E+11 & 5.36E+11 & 8.32 $\pm$ 6.49\ 27 & 2014-11-07T10:22 & N15E43 & 12205 & C & 46.59 & 37.01 & 36.24 & 30.78 & 2.50E+12 & -7.27E+11 & -1.50E+12 & 7.26E+11 & 11.14 $\pm$ 8.36\ 28 & 2014-12-01T06:41 & S21E17 & 12222 & C & 28.58 & 29.31 & 27.19 & 26.59 & -5.97E+11 & 2.68E+11 & 6.73E+11 & -2.98E+11 & 10.23 $\pm$ 7.77\ 29 & 2014-12-04T18:25 & S20W32 & 12222 & C & 36.07 & 33.92 & 29.82 & 28.11 & 3.12E+12 & -3.08E+12 & 2.19E+12 & -1.74E+12 & 9.68 $\pm$ 7.57\ 30 & 2014-12-04T19:41 & S20W32 & 12222 & C & 32.81 & 25.76 & 30.01 & 24.70 & -4.89E+11 & 1.54E+11 & -1.79E+11 & 6.00E+10 & 8.46 $\pm$ 6.69\ 31 & 2014-12-05T12:25 & S19W37 & 12222 & C & 31.72 & 29.36 & 27.67 & 26.11 & 1.21E+12 & -8.91E+11 & 8.47E+11 & -4.81E+11 & 9.03 $\pm$ 6.99\ 32 & 2014-12-19T09:44 & S19W27 & 12242 & C & 29.66 & 28.30 & 27.58 & 25.75 & 5.17E+11 & -3.67E+11 & 4.75E+11 & -2.80E+11 & 7.85 $\pm$ 6.18\ 33 & 2014-12-21T07:32 & S19W44 & 12242 & C & 38.49 & 34.61 & 32.19 & 29.55 & 2.21E+12 & -1.10E+12 & -1.29E+12 & 1.17E+12 & 8.10 $\pm$ 6.38\ 34 & 2015-01-26T16:53 & S10E25 & 12268 & C & 37.74 & 39.66 & 30.47 & 32.43 & -1.00E+12 & 5.68E+11 & 1.86E+12 & -1.09E+12 & 9.62 $\pm$ 7.40\ 35 & 2015-06-20T06:48 & N13E27 & 12371 & C & 42.39 & 44.21 & 30.17 & 33.78 & -1.28E+12 & 2.86E+11 & 8.87E+11 & -6.07E+11 & 10.73 $\pm$ 8.13 [cccccccccccccc]{} 1 & 2010-08-07T18:24 & N11E34 & 11093\* & E & 26.16 & 25.77 & 24.83 & 24.92 & -6.74E+11 & 5.81E+11 & 6.05E+11 & -4.20E+11 & 8.46 $\pm$ 6.46\ 2 & 2011-07-27T16:07 & N20E37 & 11260 & C & 44.39 & 33.95 & 32.26 & 29.15 & 3.07E+12 & -2.53E+12 & 2.39E+12 & -1.63E+12 & 11.73 $\pm$ 8.64\ 3 & 2011-08-02T06:19 & N16W8 & 11261 & E & 45.68 & 42.81 & 32.82 & 33.81 & 4.28E+12 & -8.23E+11 & -4.54E+12 & 1.94E+12 & 6.62 $\pm$ 5.50\ 4 & 2011-08-03T13:48 & N16W30 & 11261 & E & 38.88 & 35.78 & 30.92 & 29.52 & 3.48E+12 & -9.35E+11 & -2.70E+12 & 1.23E+12 & 5.98 $\pm$ 4.62\ 5 & 2011-11-09T13:35 & N25E34 & 11342 & E & 26.93 & 26.66 & 25.72 & 25.49 & -5.50E+12 & 5.44E+12 & -3.48E+12 & 3.14E+12 & 10.38 $\pm$ 7.74\ 6 & 2012-03-10T17:44 & N18W18 & 11429 & E & 30.79 & 30.18 & 27.17 & 27.30 & -3.06E+12 & 1.63E+12 & 3.87E+12 & -1.86E+12 & 8.61 $\pm$ 6.74\ 7 & 2012-03-14T15:21 & N14E5 & 11432 & E & 28.82 & 30.13 & 26.83 & 26.87 & 8.59E+11 & -4.22E+11 & -6.12E+11 & 4.76E+11 & 7.28 $\pm$ 5.78\ 8 & 2012-04-27T08:24 & N11W30 & 11466\* & C & 27.90 & 29.52 & 25.89 & 26.53 & -5.11E+11 & 4.04E+11 & 8.45E+11 & -5.17E+11 & 9.22 $\pm$ 7.24\ 9 & 2012-07-10T05:14 & S17E33 & 11520 & C & 34.37 & 49.04 & 29.38 & 41.64 & 9.58E+11 & -3.38E+11 & -2.33E+12 & 4.16E+11 & 7.23 $\pm$ 5.92\ 10 & 2012-07-10T06:27 & S17E30 & 11520 & C & 34.11 & 39.05 & 30.13 & 34.74 & 1.00E+12 & -4.01E+11 & -1.57E+12 & 1.46E+11 & 8.77 $\pm$ 6.73\ 11 & 2013-04-11T07:16 & N9E12 & 11719\* & E & 27.15 & 27.10 & 25.99 & 25.87 & -2.60E+12 & 2.40E+12 & 1.02E+12 & -9.50E+11 & 10.35 $\pm$ 7.73\ 12 & 2013-04-12T20:38 & N22W42 & 11718 & C & 30.32 & 34.79 & 26.69 & 30.13 & 4.71E+11 & -3.74E+11 & -1.33E+12 & 1.11E+12 & 9.05 $\pm$ 7.43\ 13 & 2013-08-17T19:33 & S7W29 & 11818\* & E & 39.30 & 29.16 & 31.43 & 26.34 & 1.66E+12 & -6.56E+11 & -1.61E+12 & 1.38E+12 & 10.27 $\pm$ 8.23\ 14 & 2013-10-13T00:43 & S22E17 & 11865 & E & 38.67 & 34.96 & 29.42 & 31.25 & 2.11E+12 & -3.79E+11 & -1.78E+12 & 5.49E+11 & 10.03 $\pm$ 7.69\ 15 & 2013-10-15T08:38 & S22W13 & 11865\* & E & 62.40 & 42.13 & 36.87 & 33.01 & 2.58E+12 & -2.69E+11 & -1.91E+12 & 2.92E+11 & 8.32 $\pm$ 6.40\ 16 & 2014-01-07T18:32 & S14W7 & 11944 & E & 29.16 & 27.48 & 26.07 & 25.68 & 1.36E+12 & -1.19E+12 & -2.73E+12 & 2.51E+12 & 8.31 $\pm$ 6.53\ 17 & 2014-01-31T15:42 & N9E36 & 11968 & E & 27.95 & 28.00 & 26.60 & 26.49 & 2.57E+12 & -2.39E+12 & -2.14E+12 & 1.90E+12 & 12.56 $\pm$ 9.45\ 18 & 2014-02-01T07:23 & S11E23 & 11967 & C & 35.37 & 47.78 & 29.21 & 36.78 & -1.50E+12 & 6.58E+11 & 3.67E+12 & -1.43E+12 & 8.18 $\pm$ 7.08\ 19 & 2014-03-20T03:56 & S14E35 & 12010 & E & 27.96 & 31.87 & 26.38 & 27.21 & -4.49E+11 & 3.12E+11 & -5.50E+11 & 3.94E+11 & 7.82 $\pm$ 6.27\ 20 & 2014-06-16T00:01 & S22E7 & 12087 & E & 27.65 & 28.35 & 25.93 & 25.38 & -4.70E+11 & 4.35E+11 & -5.11E+11 & 4.37E+11 & 8.48 $\pm$ 6.69\ 21 & 2014-08-01T18:13 & S9E12 & 12127 & E & 29.22 & 28.47 & 26.63 & 25.53 & -1.02E+12 & 6.58E+11 & 7.45E+11 & -4.24E+11 & 6.95 $\pm$ 5.37\ 22 & 2014-08-25T15:11 & N5W36 & 12146 & E & 31.37 & 45.25 & 28.01 & 35.50 & -1.46E+12 & 4.83E+11 & 2.07E+12 & -5.76E+11 & 6.75 $\pm$ 5.35\ 23 & 2014-08-25T20:21 & N9W38 & 12146 & E & 30.86 & 39.91 & 28.80 & 30.56 & -1.81E+12 & 5.60E+11 & 2.27E+12 & -3.58E+11 & 7.52 $\pm$ 6.02\ 24 & 2014-09-09T00:29 & N12E29 & 12158\* & E & 34.83 & 34.11 & 30.14 & 29.05 & -2.66E+12 & 1.50E+12 & 2.50E+12 & -1.29E+12 & 8.63 $\pm$ 6.59\ 25 & 2014-09-28T02:58 & S13W23 & 12173\* & E & 33.31 & 29.67 & 28.33 & 26.77 & -1.58E+12 & 1.15E+12 & 9.70E+11 & -4.65E+11 & 8.12 $\pm$ 6.23\ 26 & 2014-10-24T21:41 & S16W21 & 12192\* & C & 33.19 & 30.54 & 27.29 & 26.88 & -6.07E+12 & 4.73E+12 & 4.52E+12 & -3.36E+12 & 8.21 $\pm$ 6.35\ 27 & 2014-11-07T17:26 & N15E33 & 12205 & E & 46.31 & 35.80 & 30.97 & 29.95 & 7.78E+12 & -1.60E+12 & -5.02E+12 & 1.92E+12 & 8.19 $\pm$ 6.51\ 28 & 2014-12-17T19:01 & S10E24 & 12241 & C & 37.50 & 37.16 & 30.06 & 31.31 & 1.26E+12 & -2.71E+11 & -1.66E+12 & 8.66E+11 & 9.81 $\pm$ 7.56\ 29 & 2014-12-18T21:58 & S15E8 & 12241 & E & 37.02 & 36.65 & 31.02 & 32.10 & 3.35E+12 & -5.08E+11 & -4.39E+12 & 1.04E+12 & 9.69 $\pm$ 7.57\ 30 & 2014-12-21T12;17 & S11W21 & 12241 & E & 32.45 & 27.36 & 26.76 & 25.72 & 6.35E+11 & -2.60E+11 & -5.53E+11 & 2.85E+11 & 8.16 $\pm$ 6.36\ 31 & 2015-03-12T12:14 & S16E6 & 12297 & C & 35.59 & 41.53 & 29.54 & 28.62 & 9.92E+11 & -2.77E+11 & -1.07E+12 & 4.60E+11 & 7.54 $\pm$ 5.98\ 32 & 2015-06-21T02:36 & N13E14 & 12371\* & E & 37.24 & 34.42 & 29.87 & 28.73 & -1.70E+12 & 4.72E+11 & 1.36E+12 & -9.95E+11 & 9.52 $\pm$ 7.25\ 33 & 2015-06-21T01:42 & N12E13 & 12371\* & E & 38.77 & 33.72 & 29.92 & 30.27 & -2.37E+12 & 7.99E+11 & 2.19E+12 & -1.08E+12 & 8.21 $\pm$ 6.31\ 34 & 2015-06-22T18:23 & N12W8 & 12371\* & E & 30.88 & 30.92 & 28.17 & 27.18 & -3.01E+12 & 6.49E+11 & 2.97E+12 & -1.22E+12 & 7.62 $\pm$ 6.02\ 35 & 2015-11-04T13:52 & N6E3 & 12443 & E & 29.62 & 28.18 & 26.75 & 26.16 & 1.72E+12 & -1.22E+12 & -1.44E+12 & 1.43E+12 & 6.83 $\pm$ 5.56\ 36 & 2015-11-09T13:12 & S11E41 & 12449\* & E & 27.87 & 28.42 & 26.24 & 26.44 & -1.55E+12 & 1.48E+12 & 2.74E+12 & -1.94E+12 & 9.81 $\pm$ 7.42 [^1]: A special case is AR 11865, which is not yet sigmoidal at the time of Event \#14 of J-shaped flare, but becomes sigmoidal at the time of Event \#15.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Learning Automata (LA) are considered as one of the most powerful tools in the field of reinforcement learning. The family of estimator algorithms is proposed to improve the convergence rate of LA and has made great achievements. However, the estimators perform poorly on estimating the reward probabilities of actions in the initial stage of the learning process of LA. In this situation, a lot of rewards would be added to the probabilities of non-optimal actions. Thus, a large number of extra iterations are needed to compensate for these wrong rewards. In order to improve the speed of convergence, we propose a new P-model absorbing learning automaton by utilizing a double competitive strategy which is designed for updating the action probability vector. In this way, the wrong rewards can be corrected instantly. Hence, the proposed Double Competitive Algorithm overcomes the drawbacks of existing estimator algorithms. A refined analysis is presented to show the $\epsilon-optimality$ of the proposed scheme. The extensive experimental results in benchmark environments demonstrate that our proposed learning automata perform more efficiently than the most classic LA $SE_{RI}$ and the current fastest LA $DGCPA^{}$.' author: - Chong Di$^1$ - 'Shenghong Li$^{1}$' - Xudie Ren$^1$ - Yinghua Ma$^1$ - Bo Zhang$^2$ bibliography: - 'template.bib' date: 'Received: date / Accepted: date' title: A double competitive strategy based learning automata algorithm --- [example.eps]{} gsave newpath 20 20 moveto 20 220 lineto 220 220 lineto 220 20 lineto closepath 2 setlinewidth gsave .4 setgray fill grestore stroke grestore Introduction {#intro} ============ Learning automaton(LA) is one of reinforcement approaches. It is a decision maker which acn choose the optimal action and update its strategy through interacting with the random environment [@narendra2012learning]. As one of the most powerful tools in adaptive learning system, LA has had a myriad of applications [@oommen1996graph][@nicopolitidis2002using][@esnaashari2010data][@wang2014learning][@zhao2015cellular][@jiang2014new]. As illustrated in Fig.\[fig:1\], the process of learning is based on a learning loop involving two entities: the random environment(RE) and the LA. In this process, the LA continuously interacts with the RE to get the feedback to its various actions. According to the responses from to the various actions the environment, LA will update the probability vector with a certain method. Finally, the LA attempts to learn the optimal action by interacting with the RE through sufficient iterations. ![Learning automata that interact with a random environment [@narendra2012learning][]{data-label="fig:1"}](fig_1.eps) The first study concerning LA models dates back to the studies by $Tsetlin$ [@tsetlin1973automaton] which investigated deterministic LA in detail. $Varshavskii$ and $Vorontsova$ [@varshavskii1963behavior] introduced the stochastic variable structure versions of LA. Since then LA has been extensively researched to develop various kinds of algorithms based on deterministic LA \[6\] and stochastic LA \[7\]. A comprehensive overview of these researches has been summarized by $Thathachar$ [@thathachar2002varieties]. In general, the rate of convergence is one of the vital considerations of learning algorithms. Therefore, $Thathachar$ and $Sastry$ designed a new class of learning automata, called estimator algorithms [@thathachar1985new][@thathachar1986estimator]. The estimator algorithms have faster rate of convergence than all previous ones. These algorithms, not only maintain and update the probabilities vector of actions like the previously, but also keep estimating the reward probabilities for each action with using a reward-estimator vector to update the action probability vector. In this strategy, even when an action is rewarded, it is possible that the probability of choosing another action is increased [@papadimitriou2004new]. Compared with the traditional learning algorithms, the estimator algorithms have been demonstrated to be more efficient. However, the performance of early estimator algorithms is strictly dependent on the reliability of the estimator’s contents and an unreliable estimator may cause a significant decrease of the accuracy and the speed of convergence [@papadimitriou2004new]. In this situation, $Papadimitriou$, $Sklira$ and $Pomportsis$ [@papadimitriou2004new] designed a stochastic estimator reward-inaction learning automaton ($SE_{RI}$) which is based on the use of a stochastic estimator. As its much faster speed of convergence and much higher accuracy in choosing the correct action than other estimator algorithms, $SE_{RI}$ is is widely accepted as the most classic LA model by now. Due to the superiority of the estimator algorithms, there are many novel estimators [@ge2015novel][@jiang2011new][@jiang2016new] are proposed in recent years. In 2005, Hao Ge [@ge2015novel] proposed a deterministic estimator based LA (Discretized Generalized Confidence Pursuit Algorithm, $DGCPA$) of which the estimate of each action is the upper bound of a confidence interval and extended the algorithm to stochastic estimator schemes. The improved stochastic estimator based LA $DGCPA^{}$ is the current fastest LA model. Although the family of estimator learning automata has achieved great improvements in the field of LA, there are still some drawbacks. Because of the fundamental defect, the value of an estimator could not always be strictly unmistakable. Especially in the initial stage of the learning process of LA, the estimator may perform poorly on estimating the reward probabilities of each action. In this situation, a lot of reward would be added to the probabilities of non-optimal actions. Thus, a large number of extra iterations are needed to compensate for these wrong reward. In this paper, in order to overcome the drawbacks of estimator algorithms, a novel method based on a double competitive strategy to update the action probability vector is introduced. The proposed Double Competitive Algorithm($DCA$) learning automata use a stochastic estimator as same as $SE_{RI}$. The first competitive strategy of $DCA$ is that only the action which has the highest current stochastic estimate of reward probability gets the opportunity to increase its probability. And its second strategy is that whenever the ‘optimal’ action which has the highest current stochastic estimate of reward probability changes, the probability of new ‘optimal’ action gets a huge increase while the probability of original ‘optimal’ action decreases a lot. Accordingly, the wrong rewards could be corrected instantly. Consequently, the $DCA$ learning automata converge rapidly and accurately. The key contributions of this paper are summarized as follows. - We propose a new algorithm, referred to as Double Competitive Algorithm ($DCA$) and prove that the proposed scheme is $\epsilon-optimal$ in all random stationary environments. - The proposed $DCA$ is compared with the most classic LA $SE_{RI}$ and the fastest LA $DGCPA^{}$ in various stationary P-model random environments. The results indicate that the proposed $DCA$ is more efficient. The paper is organized as follows. In section 2, we introduce the general idea of LA and the estimator algorithms. The $DCA$ scheme is presented in section3. In section 4, we prove that the proposed scheme is $\epsilon-optimal$. Extensive simulation results are presented to describe the superiority of the proposed $DCA$ model over the most classic LA $SE_{RI}$ and the fastest LA $DGCPA^{}$ in Section 5. We conclude the paper in the last section. Learning Automata and Estimator Algorithms {#sec:1} ========================================== LA and stochastic environment ----------------------------- A LA is defined by a quintuple $ < A,B,Q,F( \cdot , \cdot ),G( \cdot ) > $, where: $\bullet$ $A = \{ {\alpha _1},{\alpha _2}, \cdots ,{\alpha _r}\} $ is the set of outputs or actions, and ${\alpha _t}$ is the action chosen by the automata at any time instant $t$. $\bullet$ $B = \{ {\beta _1},{\beta _2}, \cdots ,{\beta _m}\} $ is the set of inputs to the automata, and ${\beta _t}$ is the input at any time instant $t$. The set $t$ could be finite or infinite. In this paper, we consider the case when $B = \{0,1\} $, where $\beta = 0$ represents the events that the LA has been penalized, and $\beta = 1$ represents the events that the LA has been rewarded. $\bullet$ $Q = \{ {q_1},{q_2}, \cdots ,{q_s}\} $ is the set of finite states, and ${q_t}$ is the state of the automata at any time instant $t$. $\bullet$ $F( \cdot , \cdot ):Q \times B \to Q$ is a mapping in terms of the state and input at any time instant $t$, such that, $q(t + 1) = F(q(t),\beta (t))$. $\bullet$ $G( \cdot )$ is a mapping $G:Q \to A$, and is called the output function which determines the output of the automata depending on the state ${q_t}$, such that, $\alpha (t) = G(q(t))$. The random environment interacted with LA is defined as $ < A,B,C > $, where $A$ and $B$ has been defined above. $C = \{ {c_1},{c_2}, \cdots ,{c_r}\} $ is the set of reward probability, and ${c_i}$ corresponds to an input action ${\alpha _t}$. Estimator Algorithms -------------------- For the purpose of improving the convergence rate of LA, $Thathachar$ and $Sastry$ designed a new-class of algorithms, called estimator algorithms \[9\]\[10\]. These algorithms keep running estimates for each action using a reward-estimate vector and then use the estimate to update probabilities. According to the contents of estimators, the estimator algorithms could be divided into two classes, deterministic estimator algorithm and stochastic estimator algorithm. The class of deterministic estimator is the majority of estimator algorithms, such as $D{P_{ri}}$ [@oommen1990discretized] and $DGPA$ [@agache2002generalized]. In these algorithms, the deterministic estimate vector $D'(t) = [d_i'(t), \cdots ,d_r'(t)]$ can be computed using the following formula which yields the maximum-likelihood estimate [@sastry1985systems][@thathachar1979discretized] $$d_i'(t) = \frac{{{W_i}(t)}}{{{Z_i}(t)}},\forall i = 1,2, \cdots ,r$$ Where $W_i(t)$ is the number of times the action $\alpha_i$ has been rewarded until the current time $t$, and $Z_i(t)$ is the number of times the action $\alpha_i$ has been selected until the current time $t$. $Papadimitrtriou$ [@vasilakos1995new] introduced a new type of estimator called “stochastic estimator”. The non-stationary environments indicate that the reward probability $c_i$ will vary with time instant $t$ which means the optimal action may change from time to time. In [@vasilakos1995new], the author added an zero mean normally distributed random number to each action’s estimate probability. $Papadimitrtriou$ also extended the use of stochastic estimator to stationary environment [@papadimitriou2004new]. The implementation of the stochastic estimator in [@papadimitriou2004new] is to impose a random perturbation to the deterministic estimate, such that $${u_i}(t) = d_i'(t) + R_i^t$$ where $u_i(t)$ is the stochastic estimate of reward probability of action $\alpha_i$ at time $t$, $d_i'(t)$ is the deterministic estimate of reward probability of action $\alpha_i$ at time $t$, and $R_i^t$ is a random number which is uniformly distributed in an interval. The length of the interval depends on a design parameter $\gamma$ and the number of times that action $\alpha_i$ has been selected up to time instant $t$. Double Competitive Algorithm {#sec:2} ============================ It is clear that, in estimator learning automata fields, the most important part is to estimate the reward probabilities of each possible action accurately. However, Because of the fundamental defect, the value of an estimator could not always be strictly unmistakable. Especially in the initial stage of the learning process of LA, the estimator may perform poorly on estimating the reward probabilities of each action. Thus, a lot of rewards would be added to the probabilities of non-optimal actions. As a result, a large number of extra iterations are needed to compensate for these wrong rewards. The proposed double competitive algorithm ($DCA$) is a learning automaton which updates the action probability with a double competitive strategy to update the action probability. The first competitive strategy of $DCA$ is that only the action which has the highest current stochastic estimate of the reward probability gets the opportunity to increase its probability. And the second competitive strategy is that whenever the ‘optimal’ action which has the highest current stochastic estimate of reward probability changes, the probability of new ‘optimal’ action gets a huge increase while the probability of the original ‘optimal’ action decreases a lot. With the unique two competitive strategies, the wrong rewards could be corrected instantly. Clearly, the ‘optimal’ action would be constantly changing as the estimator is not reliable enough in the early stages of learning, resulting in the probability of each action fluctuates continually. But eventually, when the estimator is fully reliable, as the action which has the highest current stochastic estimate of reward probability tends to be invariable, the LA will converge rapidly. Besides, since the dramatic changes of the probabilities of any possible action during the learning process, the actions whose probabilities used to be relatively small get more opportunities to be selected. Then their deterministic estimates would be further updated. Therefore, during the learning process, the estimate of each non-optimal action gets more opportunities to be updated. According to the Law of Large Numbers, the precision of the stochastic estimator would be higher. So the stochastic estimator in $DCA$ scheme would be more reliable than that in $SE_{RI}$ scheme. The procedure of $DCA$ is briefly introduced below. The $DCA$ scheme* Algorithm $DCA$ Parameters $n,\gamma$ resolution parameters $\mu $ attenuation factor ${W_i}(t)$ the number of events that $i$th action has been rewarded up to time instant $t$, for $1 \le i \le r$ ${Z_i}(t)$ the number of events that the $i$th action has been selected up to time instant $t$, for $1 \le i \le r$ $\Delta = 1/r/n$ smallest step size ${m_0}$ the action that has the highest stochastic estimate of reward probability at the last time $t-1$. Method Initialize $\mu $=0.1 Initialize ${p_i}(t) = 1/r$ for $1 \le i \le r$ Initialize ${W_i}(t)$ and ${Z_i}(t)$ by selecting each action a number of times Initialize ${m_0} = $ a random integer within $[1,r]$ Repeat Step 1: At time $t$, choose an action $\alpha (t) = \alpha $, according to the probability distribution $P(t)$. Step 2: Receive a feedback $\beta (t) \in \{ 0,1\} $ from stochastic environment. Step 3: Set $W{}_i(t) = {W_i}(t - 1) + \beta (t)$, ${Z_i}(t) = {Z_i}(t - 1) + 1$. Step 4: Compute the deterministic estimate $d_i'(t)$, by setting $d_i'(t) = \frac{{{W_i}(t)}}{{{Z_i}(t)}}$. Step 5: If $\beta (t) = 0$, go to Step 9. Step 6: Compute stochastic estimates ${u_i}(t) = d_i'(t) + {R_i}(t)$, where ${R_i}(t)$ is a random number uniformly distributed within $( - \frac{\gamma }{{{Z_i}(t)}},\frac{\gamma }{{{Z_i}(t)}})$. Step 7: Select the action ${\alpha _m}$ that has the highest stochastic estimate of reward probability, where ${\alpha _m} = \max \{ {u_i}(t)\} $. Step 8: Update the probability vector $P(t)$ according to the following equations: $$\begin{array}{l} {p_i}(t) = \max \{ {p_i}(t) - \Delta \} ,\forall i \ne m;\\ {p_m}(t) = 1 - \sum\limits_{i \ne m} {{p_i}(t)} . \end{array}$$ Step 9: Compute stochastic estimates $u_i'(t)$ in the same way with Step 6 and select the action ${\alpha _m'}$ like Step 7 where ${\alpha _m'} = \max \{ u_i'(t)\} $. Step 10: if $m' = {m_0}$, go to Step 12. Step 11: Update the probability vector $P(t)$ as follows $$\begin{array}{l} {p_{m'}}(t) = {p_{m'}}(t) + (1 - \mu ){p_{{m_0}}}(t);\\ {p_{{m_0}}}(t) = \mu {p_{{m_0}}}(t);\\ {m_0} = m'. \end{array}$$ Step 12: Update the probability vector $P(t)$ at time $t+1$. $${p_i}(t + 1) = {p_i}(t),1 \le i \le r$$ End Repeat\ End Algorithm $DCA$\ Note that the double competitive strategy is reflected in the twice updating probability procedures. Step 7 and Step 8 are the implementation of the first competitive strategy, only the action which has the highest current stochastic estimate of the reward probability gets the opportunity to increase its probability and in order to satisfy $\sum_{i = 1}^r {p_i(t) = 1} $, all the probabilities of the others decrease. The second competitive strategy is summarized in Step 10 and Step 11 where whenever the ‘optimal’ action which has the highest current stochastic estimate of reward probability changes ($m' \ne {m_0}$), the probability of the original ‘optimal’ action ${\alpha _{{m_0}}}$ is reduced by $90\%$ (determined by the attenuation factor $\mu$), and then the new ‘optimal’ action ${\alpha _{m'}}$ will get an additional reward which equals to the reduced probability of action ${\alpha _{{m_0}}}$. Proof of $\epsilon-optimality$ {#sec:3} ============================== Whether the given algorithm is $\epsilon-optimality$ is an important standard in LA contexts. Thus, we will show that the proposed $DCA$ scheme is $\epsilon-optimal$ in every stationary environment.\ Definition 1 $\epsilon-optimality$ Given any arbitrarily small $\varepsilon > 0$ and $\delta > 0$, there exists a ${n_0} < \infty $ (that depends on $\varepsilon$ and $\delta $) and a ${t_0} < \infty $ such that for all resolution parameter $n \ge {n_0}$ and all time $t \ge {t_0}$ :$\Pr \{ \|p{}_m(t) - 1\| < \varepsilon \} > 1 - \delta$.\ To prove the $\epsilon-optimality$ of $DCA$ scheme, the following two theorems would be used.\ Theorem 1:Suppose there exists an index $m$ and a time instant ${t_0} < \infty $ such that ${u_m}(t) > {u_j}(t)$ for all $j$ with $j \ne m$ and for all time $t \ge {t_0}$. Then there exists an integer $n_0$ such that for all resolution parameters $n > {n_0}$, ${p_m}(t) \to 1$ with probability one as $t \to \infty $.\ Proof: Since we have supposed that ${u_m}(t) > {u_j}(t)$ for all $j$ with $j \ne m$ and for all time $t \ge t_0$, the action $\alpha_m$ which has the highest stochastic estimate of reward probability will not change, there is no difference between the proposed $DCA$ and $SE_{ri}$ scheme. The scheme has been introduced and proved in \[11\].\ Theorem 2:For action $\alpha_i$， assume ${p_i}(0) \ne 0$, for any given constants $\delta > 0$ and ${\rm M} > 0$, there exists ${n_0} < \infty $ and ${t_0} < \infty $ such that for all resolution parameters $n \ge {n_0}$ and all time $t \ge {t_0}$: $Pr [{\alpha _i}{\rm{\ is \ selected\ more\ than\ M\ times\ at\ time\ t]}} \ge {\rm{1 - }}\delta $. Proof:Define the random variable ${\alpha _i}$ as the number of times the action ${\alpha _i}$ is chosen up to time instant $t$. And then we should prove that $$\Pr \{ G_i^t > M\} \ge 1 - \delta .$$ Equivalent to prove that $$\Pr \{ G_i^t \le M\} \le \delta .$$ It is clear that the events $\{ G_i^t = k\} $ and $\{ G_i^t = j\} $ are mutually exclusive for any $j \ne k$.Then \[4\] is equivalent to $$\sum\limits_{k = 1}^M {\Pr \{ G_i^t = k\} } \le \delta .$$ Now, consider an extreme situation in the proposed $DCA$ learning automata. If the random initialization ${m_0} = i$ and the action ${\alpha _i}$ does not have the highest stochastic estimate of reward probability in the first iteration, then the probability of action ${\alpha _i}$ gets a ninety percent decay. And worse, the $i$th action would not get any reward in the subsequent iterations which means the stochastic estimate of reward probability of action ${\alpha _i}$ meets that ${u_i}(t) < {u_m}(t)$ at all time instant $t$. Thus, during any of the first $t$ iterations, the largest decrease for any action is $(1-\mu)p{}_i(0) + t\Delta $. So it is clear that: $$\Pr \{ \alpha {}_i{\rm{\ is\ not\ chosen\} }} \le {\rm{(1}} - \mu{p_i}(0) + t\Delta ).$$ The probability that action ${\alpha_i}$ is chosen up to $M$ times among $t$ iterations has the following upper bound. $$\Pr \{ G_i^t \le M{\rm{\} }} \le \sum\limits_{k = 1}^M {C(t,k){{(1)}^k}(} {\rm{1}} - \mu{p_i}(0) + t\Delta {)^{t - k}}.$$ It is clear that a sum of $M$ terms would less than $\delta $ if each element of the sum less than $\delta /M$. And when $k = m$, $C(t,m) \le {t^m}$. Thus we should prove that: $$M{t^m}{{\rm{(1}} - \mu{p_i}(0) + t\Delta )^{t - m}} \le \delta .$$ Observe the inequality, it is necessary to make sure that ${\rm{(1}} - \mu{p_i}(0) + t\Delta )$ is strictly is less than unity when $t$ increases. Thus $\Delta > \frac{{\mu{p_i}(0)}}{t}$ with $\Delta = \frac{1}{{rn}}$, such that $\Delta > \frac{{\mu {p_i}(0)}}{t}$. Let, $$n = \frac{{2t}}{{\mu rp{}_i(0)}}.$$ Now, we should prove that $$\Pr \{ G_i^t \le M\} \le M{t^m}{\Psi ^{t - m}}.$$ Where $$\Psi = 1 - \frac{\mu }{2}{{\rm{p}}_i}{\rm{(0) and }}0 < \Psi < 1.$$ Then we calculate that $$\mathop {\lim }\limits_{t \to \infty } M{t^m}{\Psi ^{t - m}} = M{\lim _{t \to \propto }}\frac{{{t^m}}}{{{{(1/\Psi )}^{t - m}}}}.$$ Using l’Hopital’s rule $m$ times, we could get the following equation: $$M\mathop {\lim }\limits_{t \to \infty } \frac{{{t^m}}}{{{{(1/\Psi )}^{t - m}}}} = M\mathop {\lim }\limits_{t \to \infty } \frac{{m!}}{{(\ln {{(1/\Psi )}^m}{{(1/\Psi )}^{t - m)}}}} = 0.$$ Thus, $M\mathop {\lim }\limits_{t \to \infty } \frac{{{t^m}}}{{{{(1/\Psi )}^{t - m}}}}$ has a limit of zero as $t$ tends towards infinity with $n = \frac{{2t}}{{\mu rp{}_i(0)}}$. In this case, for every action ${\alpha _i}$, there is a ${t_i}$, and for all $t > t(i)$, $M{t^m}{{\rm{(1}} - \mu{p_i}(0) + t\Delta )^{t - m}}$ is less than $\delta$. And, it’s clear that (8) is monotonically decreasing as $n$ increases. Let $n = \frac{{2t}}{{\mu rp{}_i(0)}}$. Hence, (8) is satisfied for all $n > {n_i}$ and $t > {t_i}$.Furthermore, for any $t > {t_i}$, we have $$\Pr \{ G_i^t \ge M\} \ge \Pr \{ G_i^{{t_i}} \ge M\} .$$ Thus, we could get that $$G_i^{{t_i}} \ge M \Rightarrow G_i^t \ge M.$$ Hence, for any action ${\alpha _i}$, $$\Pr \{ G_i^t \le M\} \le \delta {\rm{ whenever }}t > {t_i}{\rm{ and }}n > {n_i}.$$ Now, we could repeat this argument for all the actions. Define ${n_0}$ and ${t_0}$ as follows: $$\begin{array}{l} {n_0} = {\max _{1 \le i \le r}}\{ {n_i}\} ;\\ {t_0} = {\max _{1 \le i \le r}}\{ {t_i}\} . \end{array}$$ Thus, for each action, $\Pr \{ G_i^t < M\} < \delta$ is satisfied for all $t > {t_0}$ and $n > {n_0}$, and the theorem is proved. Now we are ready to prove that $DCA$ scheme is $\epsilon-optimal$. According to the Definition 1, we should prove the following theorem. Theorem 3:The $DCA$ is $\epsilon-optimal$ in every random environment. Given any $\varepsilon > 0$ and $\delta > 0$, there exists a ${n_0} < \infty $(that depends on $\varepsilon $ and $\delta $)and ${t_0} < \infty $ such that for all $n > {n_0}$ and $t > {t_0}$: $\Pr \{ \|{P_m}(t) - 1\| < \varepsilon \} > 1 - \delta $. Proof: The only difference between the proposed $DCA$ scheme and $SE_{ri}$ scheme is the method to update the probabilities. Since we have shown that the theorem 1 and theorem 2 work well in $DCA$, we can prove the $\epsilon-optimality$ of $DCA$ in the same method with $SE_{ri}$ which has been introduced in detail in [@papadimitriou2004new].\ Simulation results {#sec:4} ================== In the following, the proposed $DCA$ scheme is compared with the most classic LA $SE_{RI}$ and $DGCPA^{}$ which is considered as the current fastest LA. All of the schemes have been proved to be $\epsilon-optimal$. Within the context of LA, the speed of convergence is compared by the iterations needed to converge under the five benchmark environments given in [@papadimitriou2004new]. The actions’ reward probabilities for each environment are as follows: - $E_1$: D={0.65,0.50,0.45,0.40,0.35,0.30,0.25,0.20,0.15,0.10}. - $E_2$: D={0.60,0.50,0.45,0.40,0.35,0.30,0.25,0.20,0.15,0.10}. - $E_3$: D={0.55,0.50,0.45,0.40,0.35,0.30,0.25,0.20,0.15,0.10}. - $E_4$: D={0.70,0.50,0.30,0.20,0.40,0.50,0.40,0.30,0.50,0.20}. - $E_5$: D={0.10,0.45,0.84,0.76,0.20,0.40,0.60,0.70,0.50,0.30}. In all the simulations performed, we have the same setting as [@papadimitriou2004new]. The computation of an algorithm is considered to have converged if the probability of choosing an action is greater than or equal to a threshold $T$($0 < T < 1$). The automaton is considered to have converged correctly when it converges to the action that has the highest reward probability. Before comparing the performance of different learning automata, a large number of evaluation tests were carried out to determine the ‘best’ parameters for each scheme. The values of ‘best’ parameters are considered to be the best if they yield the fastest convergence and the automaton converges to the correct action in a sequence of experiments. The values of $DCA$ and $SE_{RI}$ are taken to the same as those used in [@papadimitriou2004new]. Hence, $T = 0.999$ and $NE = 750$. As long as we have determined the ‘best’ parameters, each algorithm was executed 250,000 times for each environment by using the ‘best’ parameters. Before the simulation, to initialize the estimator vector, all the actions were sampled 10 times, and these extra 100 iterations are included in the iteration counts. Before comparing the overall simulation results, a single ordinary experiment would be executed to show the difference between $DCA$ and $SE_{ri}$ during the convergence process. The curves that represent the probability of the optimal action as a function of time are presented in Fig \[fig:2\]. ![image](fig_2.eps){width="12cm"} ![image](fig_3.eps){width="12cm"} The results presented in Fig \[fig:2\] indicate that the probability of the optimal action changes dramatically in the initial stage of the $DCA$ learning process as we have explained earlier. With the number of iterations increasing, the stochastic estimator becomes more and more reliable. When the estimator is sufficiently reliable, the learning automaton converges rapidly. On the other hand, during the convergence process of $SE_{RI}$ scheme, once the probability of the optimal action decreases, a lot of extra iterations are needed to compensate for the lost probability.\ Besides, as presented in Fig \[fig:3\], the non-optimal actions in $DCA$ scheme have more chances to be selected than in $SE_{RI}$ scheme. Thus, during the learning process, the estimate of each action gets more opportunities to be updated and the precision of the stochastic estimator would be higher. So the time $t$ when the stochastic estimator is reliable enough in $DCA$ scheme would be earlier than that in $SE_{ri}$ scheme. Thus, with the benefits that have been explained above, the overall simulation results are presented as follows. [llllll]{} & $E_1$ & $E_2$ & $E_3$ & $E_4$ & $E_5$\ $DCA$ & 0.998 & 0.997 & 0.996 & 0.999 & 0.998\ $SE_{RI}$ & 0.997 & 0.996 & 0.995 & 0.998 & 0.997\ $DGCPA^{}$ & 0.997 & 0.996 & 0.995 & 0.998 & 0.997\ [lllllll]{} Environment & $DCA$ & & $SE_{RI}$ & & $DGCPA$ &\ & Parameter & Iterations & Parameter & Iterations & Parameter & Iterations\ $E_1$ & $n=13 ,\gamma=6 $ & 377 & $n=16 ,\gamma=8 $ & 426 & $n=3 ,\gamma=5 $ & 351\ $E_2$ & $n=23 ,\gamma=8 $ & 664 & $n=32 ,\gamma=12$ & 834 & $n=6 ,\gamma=9 $ & 678\ $E_3$ & $n=43 ,\gamma=16 $ & 2134 & $n=105 ,\gamma=25 $ & 2540 & $n=19 ,\gamma=20 $ & 2032\ $E_4$ & $n=12 ,\gamma=5 $ & 299 & $n=13 ,\gamma=6 $ & 325 & $n=2 ,\gamma=4 $ & 298\ $E_5$ & $n=40 ,\gamma=7 $ & 633 & $n=33 ,\gamma=12 $ & 729 & $n=5 ,\gamma=7 $ & 598\ [llllll]{} Environment & $DCA$ & & $SE_{RI}$ & & Improvement\ & Parameter & Iterations & Parameter & Iterations\ $E_1$ & $n=10 ,\gamma=6 $ & 338 & $n=16 ,\gamma=8 $ & 426 & 20.66%\ $E_2$ & $n=18 ,\gamma=8 $ & 633 & $n=32 ,\gamma=12 $ & 834 & 24.10%\ $E_3$ & $n=30 ,\gamma=16 $ & 1990 & $n=105 ,\gamma=25 $ & 2540 & 21.65%\ $E_4$ & $n=9 ,\gamma=5 $ & 282 & $n=13 ,\gamma=6 $ & 325 & 13.23%\ $E_5$ & $n= 28,\gamma=7 $ & 582 & $n=33 ,\gamma=12 $ & 729 & 20.16%\ [lllllll]{} Environment & $DCA$ & & & $DGCPA^{}$\ & Parameter & Iterations & Time(ms) & Parameter & Iterations & Time(ms)\ $E_1$ & $n=10 ,\gamma=6 $ & 338 & 0.162 & $n=16 ,\gamma=8 $ & 426 & 3.423\ $E_2$ & $n=18 ,\gamma=8 $ & 633 & 0.339 & $n=32 ,\gamma=12 $ & 834 & 7.417\ $E_3$ & $n=30 ,\gamma=16 $ & 1990 & 1.167 & $n=105 ,\gamma=25 $ & 2540 & 26.577\ $E_4$ & $n=9 ,\gamma=5 $ & 282 & 0.126 & $n=13 ,\gamma=6 $ & 325 & 2.744\ $E_5$ & $n= 28,\gamma=7 $ & 582 & 0.351 & $n=33 ,\gamma=12 $ & 729 & 9.252\ The accuracies (number of correct convergences/number of experiments) of $DCA$, $SE_{RI}$ and $DGCPA^{}$ in environment $E_1$ to $E_5$ when using the ‘best’ learning parameters are presented in Table \[tab:1\]. The results show that $DCA$ always has a better accuracy than the other two algorithms. The average numbers of iterations required for convergence are summarized in Table \[tab:2\], which demonstrate that the $DCA$ scheme converges with a faster speed than $SE_{RI}$ and with a little lower speed than the current fastest LA $DGCPA^{}$ with higher accuracy. In order to ensure that the performance comparison between $DCA$, $SE_{RI}$ and $DGCPA^{}$ is fair, let us verify the number of iterations required to achieve the same accuracy, a series of experiments have been carried out. The results are shown in Table \[tab:3\] and Table \[tab:4\]. On the one hand, compared with the most classic LA model $SE_{RI}$, the proposed scheme $DCA$ achieves a great improvement in the speed of convergence in all benchmark environments. For example, in environment $E_2$, The $DCA$ converges in 633 iterations, while the $SE_{RI}$ requires 834 iterations. Thus, an improvement of $24.10\%$ in comparison with $SE_{RI}$ is obtained. On the other hand, as indicated in Table \[tab:4\], the current fastest LA model $DGCPA^{}$ performs less competitively than the proposed $DCA$ scheme. The superiority of $DCA$ is not only reflected in the fewer number of iterations for convergence, but also established in the time efficiency. Because of the complexity of $DGCPA^{}$ model when computing the confidence interval, the time required for convergence increases rapidly. Thus, the superiority of the proposed $DCA$ scheme is clear. In summary, the $DCA$ scheme using a double competitive strategy is more efficient than $SE_{RI}$ and $DGCPA^{}$. It overcomes the drawbacks of estimator algorithms and provides a novel idea to make breakthroughs in LA fields. Conclusions {#sec:5} =========== In this paper, a novel P-model absorbing learning automaton is introduced. With the use of double competitive strategy, the proposed $DCA$ scheme overcomes the drawbacks of existing estimator algorithms. The benefits of proposed scheme are analysed and it is proved to be $\epsilon-optimal$ in every stationary random environment. Extensive simulations have been performed in five benchmark environments, and the results indicate that the proposed $DCA$ scheme converges faster and performs more efficiently than the most classic LA $SE_{RI}$ and the current fastest LA $DGCPA^{}$. Since the reliability of an estimator is the key to guarantee the convergence of LA, the future work will focus on studying how to make the estimator being reliable enough as soon as possible. This research work is funded by the National Key Research and Development Project of China (2016YFB0801003), Science and Technology Project of State Grid Corporation of China ( SGCC)，Key Laboratory for Shanghai Integrated Information Security Management Technology Research.	{ "pile_set_name": "ArXiv" }
--- author: - 'F. Dulieu ' - 'L. Amiaud ' - 'E. Congiu' - 'J-H. Fillion ' - 'E. Matar ' - 'A. Momeni' - 'V. Pirronello' - 'J. L. Lemaire ' date: 'Received XXX; accepted YYY' title: Experimental evidence for water formation on interstellar dust grains by hydrogen and oxygen atoms --- [The synthesis of water is one necessary step in the origin and development of life. It is believed that pristine water is formed and grows on the surface of icy dust grains in dark interstellar clouds. Until now, there has been no experimental evidence whether this scenario is feasible or not on an astrophysically relevant template and by hydrogen and oxygen atom reactions.]{} [We present here the first experimental evidence of water synthesis by such a process on a realistic grain surface analogue in dense clouds, i.e., amorphous water ice.]{} [Atomic beams of oxygen and deuterium are aimed at a porous water ice substrate (H$_{2}$O) held at 10 K. Products are analyzed by the temperature-programmed desorption technique]{} [We observe production of HDO and D$_{2}$O, indicating that water is formed under conditions of the dense interstellar medium from hydrogen and oxygen atoms. This experiment opens up the field of a little explored complex chemistry that could occur on dust grains, believed to be the site where key processes lead to the molecular diversity and complexity observed in the Universe.]{} Introduction ============ Water, the spring of life [@brack], is the most abundant molecule in biological systems, and it is almost certainly of extraterrestrial origin. Water has been detected, in gaseous or solid form, in numerous astrophysical environments such as planets, comets, interstellar clouds and star forming regions where strong maser emission can be also observed [@ehrenfreund; @dartois]. Amorphous water ice was directly detected in dark interstellar clouds through infra-red absorption [@leger]. During the formation of stars deep inside molecular clouds, gas and dust become part of the infalling material feeding the central object. Part of this gas and dust grains, covered with icy mantles (mainly composed of water), ends up in the rotating disks surrounding young stars and forms the basic material from which icy planetesimals and later planets, together with comets in the external regions, are formed [@vanD]. While the means of delivery of water to Earth remain a subject of debate [@morbidelli], the synthesis of water in the Universe is a fundamental link in establishing our origins. Water molecule formation in the gas phase is not efficient enough to reproduce the observed abundances in dark clouds, especially in its solid form [@pariseb; @ceccarelli]. Therefore water ice must form directly on the cold interstellar grains and not as a condensate after being formed in the gas phase. A complete review of the processes involved both in the gas and solid phase has been recently published [@tielensb]. It was suggested many years ago that interstellar dust grains act as catalysts [@oort; @vandeH]. Starting from simple atoms or molecules such as H, O, C, N, CO, grains are believed to be chemical nanofactories on which more complex molecules are synthesized leading eventually to prebiotic species produced concurrently by surface reactions and by UV photons and cosmic rays irradiation, as already shown long ago [@hagen; @pirronelloa]. The most volatile species may be released in the gas phase upon formation [@garrod], while the refractory ones remain on the grain surface, building up a so called “dirty icy mantle”, and at least partially may be sputtered by the heavy component of cosmic rays [@johnson]. Such mantles, having a typical thickness of a hundred monolayers, are mainly composed of water, the most abundant solid phase species in the Universe. Under dark cloud conditions, except for the very first monolayer that has to grow on bare silicate or carbonaceous grains [@papoular], most water molecules should be subsequently synthesized on a surface mainly composed of water. Chemical models including water formation on grain surfaces were proposed years ago by [@tielensa]. They suggested that H$_{2}$O formation would be initiated by H-atoms reacting with O, O$_{2}$ and O$_{3}$, although the O$_{3}$+H pathway was considered the most effective and O$_{2}$ would play more a catalytic role. Recent Monte Carlo simulations [@cuppen] show that while the main route to water formation on cosmic dust grains in diffuse and translucent clouds is the reaction H + OH, in dense clouds the principal source of H$_{2}$O is the reaction between H$_{2}$ and OH. This study also emphasizes the non-negligible contribution from the H + H$_{2}$O$_{2}$ reaction (H$_{2}$O$_{2}$ being a product of the H + O$_{2}$ pathway) and the unusual high abundance of reactants such as OH and O$_{3}$. Interestingly, another code by @parisea proposed a water formation scheme where O$_{3}$ molecules react with H- or D-atoms to form OH or OD, and subsequently the reaction H$_{2}$ + OH/OD leads to H$_{2}$O/HDO. It should be noted that this scheme was in part also constrained by the observed abundances of deuterated species. In previous laboratory works, [@hiraoka] succeeded in producing water molecules from the reaction of H- and O-atoms initially trapped in a N$_{2}$O matrix. Very recently [@miyauchi] investigated the reaction between cold H-atoms and an O$_{2}$ ice at 10 K and demonstrated the production of H$_{2}$O$_{2}$ and H$_{2}$O molecules and estimated the efficiency of the reactions. [@ioppolo] did a similar experiment but with varying O$_{2}$ substrate temperatures. They confirmed the production of H$_{2}$O$_{2}$ and H$_{2}$O, made an estimate of the reactions efficiency and also drew conclusions upon the temperature dependence of the amount of species produced. These two experiments dealt with the H$_{2}$O production pathway in which O$_2$ is the species consumed to produce water as shown in preliminary experiments by our group [@momeni; @matar] In the present study, the formation of water is studied for the first time using hydrogen and oxygen atoms interacting on the surface of an amorphous solid water (ASW) ice film, hence under conditions that are much more relevant to the interstellar medium. The aim of this first attempt to synthesize water under conditions close to those encountered in dense clouds was to investigate how water formation continues on the icy surfaces of cosmic grains and to give an estimate of the efficiency of the chemical path(s) actually active. Experimental procedures ======================= Experiments were performed with the FORMOLISM set-up [@amiaud]. In brief, a copper sample surface whose temperature can be controlled by computer in the 8-800 K range is maintained under UHV conditions; on it an amorphous solid water ice substrate, on which water formation is studied, is prepared in two steps. First, a 100-layer film of non-porous ASW ice is first deposited at 120 K [@kimmel], then an overlayer of 10 layers of porous ASW ice is grown at 10 K. The underlying non-porous water film isolates the ice layer from the copper substrate [@engquist]. This double ASW ice film is annealed to 90 K prior to each experimental run in order to avoid any further collapse of the pores between 10 and 80 K in the subsequent temperature-programmed desorption (TPD) experiments (see below). In fact, the species D$_{2}$, O, O$_{2}$ used in this experiment will be thoroughly evaporated by 90 K and thus before any rearrangement in the porous structure of the ice template. The annealed water ice substrate is still porous [@kimmel] and mimics an amorphous ice processed by UV and cosmic rays [@palumbo] that is thought to constitute the icy mantle of interstellar grains. By using an architecture with two separate channels [@pirronellob], each consisting of a triply differentially pumped beam line, two atomic beams of O and D are aimed at the ASW ice substrate held at 10 K. Atoms are produced by dissociation of O${_{2}}$ and D${_{2}}$ in microwave discharges. The dissociation efficiency of the oxygen beam is typically 40%, meaning that for 100 O${_{2}}$ molecules that initially feed the discharge, 60 O${_{2}}$ molecules and 80 O atoms will finally reach the cold target. Trace quantities of residual gases (i.e., CO, N$_2$, CO$_{2}$ and H$_{2}$O) are present in the beam, although neither O$_3$ nor deuterated compounds were detected. The D beam has a 60% dissociation efficiency, and no UV photons from the D$_{2}$-discharge plasma can reach the water ice sample. The purity of the beams was checked with a rotatable quadrupole mass spectrometer (QMS) that intercepted the incoming beam. The O- and D-beams have a flux of $10^{12}$ atoms/cm$^2$/s and $5\times 10^{12}$ atoms/cm$^2$/s, respectively. After concurrent or sequential injection of the beams, a TPD is performed at a heating rate of 20 K/minute from 10 K to 200 K and simultaneously the desorbing species are monitored with the QMS. The signals of mass 19 (HDO) and mass 20 (D$_2$O) presented here are corrected by subtracting the contribution of the H$_2$O water substrate, which naturally includes some isotopes. Results and Discussion ====================== Several experiments have been performed. In the first one D-atoms were sent onto the water surface to confirm that hydrogen atoms do not react with water molecules in the substrate to produce deuterated water molecules (see the solid thick line in Fig. \[Fig1\]), as already mentioned by [@nagaoka]. In addition, we determined that D$_2$ molecules do not react with O-atoms nor with O$_2$ molecules residing on the ASW ice surface. This, that may seem of secondary importance, is on the contrary a decisive result that proves that the water formation process requires hydrogen in atomic form. ![TPD profiles of D$_2^{16}$O (thin line) and HD$^{16}$O (circles) after irradiation of the water substrate (H$_2^{16}$O) with D atoms, $^{16}$O atoms and $^{16}$O${_{2}}$ molecules. Thick line: TPD profile of HD$^{16}$O, after 30 minutes of D irradiation of the water ice substrate. The water ice substrate is held at 10 K during all exposures. Traces are vertically shifted for clarity. []{data-label="Fig1"}](Figure1.EPS){width="9cm"} Fig. \[Fig1\] shows the mass spectrum recorded during the TPD performed after the irradiation of the water substrate with D-atoms and $^{16}$O-atoms (and $^{16}$O${_{2}}$ molecules). Water formation clearly occurs as is testified by the presence of D$_2^{16}$O and HD$^{16}$O peaks. These findings are not surprising although one could expect that D$_2$O should be mainly formed, because D is used as precursor. But it is also known that during the heating, isotope exchange between water molecules will occur [@smith97]. For a very thin layer of deuterated water, a complete isotope exchange with the underlying H$_2$O substrate molecules is expected around 150 K. In such a context, even if D$_2$O is formed at 10 K, most of the deuterated water will desorb as HD$^{16}$O, as is observed in our experiments. Due to the mass spectrometric detection method of water molecules in the gas phase one could wonder whether molecules synthesized by the reaction of D-atoms and O-atoms and O$_2$ were formed during the warm up of the whole ice substrate, a fact that would render the result not relevant to interstellar space. However, this is certainly not the case because reactions involving D-atoms must proceed at temperature below 20 K. At temperatures higher than that value the residence time of H-/D-atoms becomes exceedingly small for reactions to occur. In order to confirm unequivocally that water was formed starting from the reactants deposited from the gas phase and provide quantitative results, we repeated the experiment presented above with the oxygen beam line fed with $^{18}$O$_2$ molecules. To be sure that the concentration on the substrate of D-atoms on the surface was always much higher than O-atoms and O$_2$ molecules, we opened the oxygen beam by time-slices of 20 s every 2 minutes while the D-beam was constantly running and irradiating the target held at 10 K. Thus, the relative concentration on the surface consists of $\ge 10$ D-atoms per O-nucleus (in atomic or molecular form). [@matar] have shown that D reacts very efficiently with O$_2$, and that O$_2$ disappears from the surface in a one D for one O$_2$ fashion. In such a scenario, due to the very low mobility of O-atoms [@tielensa] and O$_2$ molecules at the irradiation temperature and in the temperature range in which H- and D-atoms remain adsorbed on the substrate, it is reasonable to assume that O and O$_2$ surface concentration reflects the proportions in which they were produced in the beam. O-atoms have therefore a very low probability to form O$_2$ and to encounter O$_2$ molecules to form O$_3$, especially because they react with D with a higher probability. Fig. \[fig2\] shows a TPD mass spectrum obtained after a 10-minute dose (30 time-windows of 20 s every 2 minutes) of $^{18}$O-atoms and simultaneous D-atom irradiation of the ASW ice substrate at 10 K. The two peaks at mass 21 and 22 can only be assigned to HD$^{18}$O and D$_2^{18}$O and this clearly shows again that formation of water molecules has occurred on the amorphous H$_2$O ice substrate, namely, a realistic analogue of cosmic dust surfaces in dense clouds. ![ HD$^{18}$O (open triangles) and D${_{2}}$$^{18}$O (solid line) desorption profiles after ten minutes of simultaneous irradiation of the amorphous H$_2$O ice substrate with $^{18}$O- and D-atoms. []{data-label="fig2"}](Figure2.EPS){width="9cm"} In Table \[Table:1\] are shown the relative abundances (in terms of uncorrected ion counts) of the species detected during the TPD, and, in addition, the total number of oxygen nuclei ($^{18}$O atoms and twice the number of $^{18}$O$_2$ molecules) that have been exposed to the substrate during the irradiation phase. We have not performed here any correction (except the correction from the H$_2$O substrate). -------------- ---------------- ---------------- --------------- -------------- ------------ --------- Total $^{18}$O Mass 19 Mass 20 Mass 21 Mass 22 Mass 23 Area (Cts.K) 15000$\pm$1000 30000$\pm$2000 10000$\pm$500 1250$\pm$150 250$\pm$70 20 Model 15000 31802 8076 1434 188 0 -------------- ---------------- ---------------- --------------- -------------- ------------ --------- As a preliminary remark, the signal at mass 23 is expected to be 0, and it serves as a good indicator of the level of noise. To evaluate the conversion efficiency of O-atoms into water molecules in our experiment, we may evaluate the ratio $\alpha$ between the sum of mass 20 ($^{18}$OD and H$_2^{18}$O ), mass 21 (HD$^{18}$O), mass 22 (D$_2^{18}$O) and the total amount of $^{18}$O nuclei: $$\alpha =\frac{\mathrm{M}20 + \mathrm{M}21+ \mathrm{M}22}{\mathrm{total} \, ^{18}\mathrm{O}} = 0.76$$ This ratio represents an upper limit to the formation of water molecules from $^{18}$O assuming that the contribution of D$_2^{16}$O to the signal at mass 20 is negligible, as we have actually seen above (Fig. \[Fig1\]). $\alpha$, however, is not 1. The missing $^{18}$O nuclei are likely to have partially desorbed during the reaction or ended up in the form $^{18}$OH radicals or even as H$_2^{18}$O$_2$ (and its deuterated isotopologues). The formation of water molecules through the D + O$_2$ pathway has already been studied and confirmed experimentally by [@miyauchi; @ioppolo; @matar]. In the present experiment, we can deduce a lower limit for the formation efficiency for the D+O pathway. Actually 60% of the available $^{18}$O on the surface is contained in molecules. We can then see that, even if we assume that the molecular pathway has a 100% efficiency, it is less than $\alpha$ and a lower limit for water formation via the D + O pathway is about 0.16 ($0.76-0.60$). Therefore, if we define the formation efficiency via atoms as the ratio between the fraction of water molecules that certainly formed via the D + O route (0.16) and the fraction of O-atoms available (0.4), we obtain 0.4 as a lower limit for the formation efficiency. On the other hand we can also deduce an upper limit. If we assume that O and O$_2$ react with equal probability with D, we can estimate that 31% ($0.76 \times 0.4$) of the water we have formed is the product of the D + O pathway and the efficiency of water formation via atoms is equal to $\alpha$. Considering both boundary conditions, we conclude that a reasonable estimate for the efficiency of water formation via atoms is about $0.5$. In order to know if we can go deeper in the understanding of the different pathways of water formation, we have made a very simple statistical model which is able to reproduce correctly the data (see Table \[Table:1\]). The model considers only statistical equilibrium. For sake of clarity we give here a simple example. Suppose there is an initial population of 90 H$_2$O and 10 OD , the final equilibrium obtained by randomly mixing the isotopes is OD = 0.5, OH = 9.5 , H$_2$O = 81.225, HDO = 8.55, D$_2$0 = 0.225. We have also included oxygen scrambling and calculated the statistical weight of all the final compounds. Finally, we sum the final products by masses taking into account QMS fractionation and are able to compare with the experimental results. The question is the following: is it possible that the OH + H reaction has a activation barrier as previously suggested in gas phase. In this case, $^{18}$OD could be mixed with H$_2$O at higher temperature during the isotope exchange? We first assume that the OD + D reaction is actually negligible compared to the O$_2$ + D reaction, which leads to formation of OD and D$_2$O in equal amounts. Therefore, at the end of the D irradiation the ratio $a=$ $^{18}$OD /D$_2^{18}$O is equal to (40+30 =70)/30 because $^{18}$O atoms will form $^{18}$OD and $^{18}$O$_2$ molecules will form an equal amount of $^{18}$OD and D$_2$$^{18}$O (via $^{18}$O$_2$D$_2$) . We assume only one adjustable parameter, the ratio $b$=$^{16}$O/$^{18}$O of atoms that will exchange during the TPD and statistically equilibrate their populations via isotope exchanges. We find a best fit (Table \[Table:1\]) for the only free parmeter $b=15$. This means that the overlayer of products ($^{18}$OD and D$_2^{18}$O) is mixed with 15 layers of the H$_2$O substrate and this is reasonably consistent with previous results. [@smith00] found an efficient mixing of about 50 ML of both isotopes, although they used a lower heating ramp and a different amorphous ice. If one believes in our simple statistical model, it is also possible to explain our data starting with an high population of OD, and therefore that an activation barrier for the OD + D reaction exists at 10 K. Indeed, the isotope exchange blurs the chemical pathways, and the data provided here give not enough constraints to make a firm conclusion about which pathway is more efficient. None the less, on average, we have estimated that water formation via the D + O pathway is highly efficient (about 50%) and certainly is responsible for producing a significant fraction of water molecules. More laboratory investigations should be performed and we actually plan to do that by combining TPD and infra-red spectroscopy. The relevance of our results to astrochemistry is clear and gives strong experimental support to simulations of the formation and time scale of growth of water ice mantles in dense clouds. For the first time it is demonstrated on experimental grounds that water molecules can form on an amorphous water ice substrate under interstellar conditions (i.e., through surface reactions between atomic hydrogen and atomic and molecular oxygen on ASW ice), allowing the growth of icy mantles that are observed in dense clouds. Also, in our experiments, as occurs on interstellar grains, the formation of water ice via reactions between hydrogen and oxygen atoms suffers from the competition between the formation of molecular hydrogen and molecular oxygen and probably this competition might reflect in the depth profile inside the ice layer and in the time evolution of the mantle itself. Several issues still need to be investigated: the efficiency at various coverages and with an abundance ratio between hydrogen and oxygen atoms closer to the interstellar one; the interaction of the atoms that land from the gas phase and those belonging to the surface; other chemical pathways that contribute to the formation of water and so on. Besides all these investigations, of course, still to prove is how the first monolayer of water molecules may form efficiently on an amorphous silicate and/or and amorphous carbon layer. If it can not form efficiently enough, it will be compulsory to accept that the very first monolayer has to be built by water molecules formed by gas-phase reactions and then accreted on the bare refractory grain surface; a fact of major interest inside dense clouds near the threshold of observability of icy mantles. These and other issues will be addressed in forthcoming papers. Conclusions =========== We presented the first laboratory attempt to reproduce the formation of water molecules on of realistic space analogue of grain surfaces in dense molecular clouds. By exposing O- and D-atoms to an amorphous water ice substrate held at 10 K deuterated water molecules were formed with a high efficiency ($\sim$ 50 %). We acknowledge the support of the French National PCMI program funded by the CNRS, as well as the strong financial support from the Conseil Régional d’Ile de France through a SESAME program (No. E1315) and from the Conseil Général du Val d’Oise. We also thank E. herbst, S. Leach, H. J. Fraser and P. Cernicharo for their comments on an early version of this work. Amiaud, L. et al. 2006, J. Chem. Phys., 124, 094702 Borget, F., Chiavassa, T., Allouche, A., Aycard, J. P. 2001, J. Phys. Chem. B, 105, 449 Brack, A. 2002, in Astrobiology: The quest for the conditions of life, ed. G. Horneck, & C. Baumstark-Khan (Physics and astronomy online library, Springer, Berlin) 79 Caselli, P., Stantcheva, T., Shalabiea, O. M., Shematovich, V. I., & Herbst, E. 2002, Planetary and Space Science, 50, 1257 Ceccarelli, C., Caselli, P., Herbst, E., Tielens, A. G. G. M., & Caux, E. 2007, in Protostars and Planets, ed. V. B. Reipurth, D. Jewitt, & K. Keil (U. Arizona Press) 47 Cuppen, H.M., & Herbst, E. 2007 ApJ, 668, 294 Dartois, E. 2005, Space Science Reviews, 119, 293 Ehrenfreund, P. et al. 2003, Planet. Space Sci., 51, 473 Engquist, I., Lundström, I., & Lieberg, B. 1995, J. Phys. Chem., 99, 12257 Garrod, R. T., Wakelam, V., & Herbst, E. 2007, , 467, 1103 Hagen, W., Allamandola, L. J., & Greenberg, J. M. 1979, Astrophysics and Space Science, 65, 215 Hiraoka, K., Miyagoshi, T., Takayama, T., Yamamoto, K., Kihara, Y. 1998, ApJ, 498, 710 Ioppolo, S., Cuppen, H. M., Romanzin, C., van Dishoeck, E. F., & Linnartz, H. 2008, ArXiv e-prints, 807, arXiv:0807.0129 Johnson, R. E., Pirronello, V.,Donn, B. & Sundqvist, B. 1991, ApJ, 379, L75 Kimmel, G. A., Stevenson, K. P., Dohnalek, Z., Smith, R. S., & Kay, B. D. 2001, J. Chem. Phys., 114, 5284 Leger, A. et al. 1979, A&A 79, 256 Matar, E., Congiu, E., Dulieu, F., Momeni, A., & Lemaire, J. L. 2008, A&A Lett., 492, L17-L20 Meyer, D. M., Jura, M., & Cardelli, J. A. 1998, ApJ, 493, 222 Miyauchi, N., Hidaka, H., Chigai, T., Nagaoka, A., Watanabe, N., & Kouchi, A. 2008, Chemical Physics Letters, 456, 27 Momeni, A. ,Web pubication of New Astronomical Challenges In Surface Science. 2007, Heriot-Watt University, Edinburgh, http://www.chem.ucl.ac.uk/astrosurf/challenge07.html Morbidelli, A. et al. 2000, Meteoritics & Planetary Science, 35, 1309 Nagaoka, A., Watanabe, N., & Kouchi, A. 2005, ApJ, 624, L29 Oort, J. H., & van de Hulst, H. C. 1946, BAN, 10, 1870 Palumbo, M. E. 2006, A&A, 453, 903 Papoular, R. 2005, Mon. Not. R. Astron. Soc., 362, 489 Parise, B. 2004, PhD Thesis, CESR (Toulouse, France) Parise, B., Ceccarelli, C., & Maret, S. 2005, A&A, 441, 171 Pirronello, V., Brown, W. L., Lanzerotti, L. J., Marcantonio, K. J., & Simmons, E. H., 1982, ApJ, 262, 636 Pirronello, V., Liu, C., Shen, L., & Vidali, G. 1997, ApJ 475, L69 Smith, S. R., Huang, C., & Kay , B. D., 1997, J. Phys. Chem. B 101, 6123-6126 Smith, S. R., Dohalek, Z., Kimmel, G. A. & Kay , B. D., 2000, Chemical Physics 258, 291 Tielens, A. G. G. M., & Hagen, W. 1982, A&A 114, 245 Tielens, A. G. G. M. 2005, The Physics and Chemistry of the Interstellar Medium (Cambridge University Press) van de Hulst, H. C. 1946, Recherches Astronomiques de l’Observatoire d’Utrecht, 11, 2 van Dishoeck, E. F. 2004, Annu. Rev. Astron. Astrophys., 42, 119 Whittet, D.C.B., Gerakines, P.A., Hough, J.H., & Shenoy, S.S. 2001, ApJ, 547, 872	{ "pile_set_name": "ArXiv" }
--- abstract: \| Motivated by the fact that in several cases a matching in a graph is stable if and only if it is produced by a greedy algorithm, we study the problem of computing a maximum weight greedy matching on weighted graphs, termed [<span style="font-variant:small-caps;">GreedyMatching</span>]{}. In wide contrast to the maximum weight matching problem, for which many efficient algorithms are known, we prove that <span style="font-variant:small-caps;">GreedyMatching</span> is strongly NP-hard and APX-complete, and thus it does not admit a PTAS unless P=NP, even on graphs with maximum degree at most 3 and with at most three different integer edge weights. Furthermore we prove that [<span style="font-variant:small-caps;">GreedyMatching</span>]{}is strongly NP-hard if the input graph is in addition bipartite. Moreover we consider two natural parameters of the problem, for which we establish a sharp threshold behavior between NP-hardness and tractability. On the positive side, we present a randomized approximation algorithm ([<span style="font-variant:small-caps;">Rgma</span>]{}) for [<span style="font-variant:small-caps;">GreedyMatching</span>]{}on a special class of weighted graphs, called bush graphs. We highlight an unexpected connection between [<span style="font-variant:small-caps;">Rgma</span>]{}and the approximation of maximum cardinality matching in unweighted graphs via randomized greedy algorithms. We show that, if the approximation ratio of <span style="font-variant:small-caps;">Rgma</span> is $\rho$, then for every $\epsilon >0$ the randomized <span style="font-variant:small-caps;">Mrg</span> algorithm of [@ADFS95] gives a $(\rho-\epsilon)$-approximation for the maximum cardinality matching. We conjecture that a tight bound for $\rho $ is $\frac{2}{3}$; we prove our conjecture true for two subclasses of bush graphs. Proving a tight bound for the approximation ratio of <span style="font-variant:small-caps;">Mrg</span> on unweighted graphs (and thus also proving a tight value for $\rho $) is a long-standing open problem [@PS12]. This unexpected relation of our <span style="font-variant:small-caps;">Rgma</span> algorithm with the <span style="font-variant:small-caps;">Mrg</span> algorithm may provide new insights for solving this problem. Keywords: Greedy weighted matching, maximum cardinality matching, NP-hard, approximation, randomized algorithm. author: - 'Argyrios Deligkas[^1]' - 'George B. Mertzios[^2]' - 'Paul G. Spirakis[^3]' title: The Complexity of Weighted Greedy Matching --- Introduction ============ The matching problem is one of the fundamental and most well studied problems in combinatorial optimization. Several different versions of the matching problem have been studied over the years: matchings on weighted or unweighted bipartite [@FL10] and general [@Edmonds] graphs, popular matchings [@AIKM], stable matchings [@R84], and greedy matchings [@HK78], to name but a few. In this article we investigate the problem of computing and approximating a maximum weight greedy matching on edge-weighted graphs; i.e. the matching with the maximum weight when every edge that is added to the matching is chosen greedily from the set of available edges with the largest weight. Although various polynomial time algorithms are known for the maximum weight matching problem, these algorithms are not always very fast in practice, see [@Duan14; @San09] and references therein for a comprehensive list of known results. One way to deal with this inefficiency is to turn into approximation algorithms; a recent major result in this direction was given by Duan and Pettie [@Duan14] who provided an algorithm for computing an $(1-{\epsilon\xspace})$-approximation of the maximum weight matching in $O(m {\epsilon\xspace}^{-1} \log{\epsilon\xspace}^{-1})$ time, where $m$ is the number of edges. However, there are cases where it makes sense to search for algorithms that are fast in practice and easy to implement, such as greedy algorithms. To make best use of such algorithms, their algorithmic performance needs to be properly understood. There are cases where a greedy approach for computing a weighted matching is preferred or even required, as the classical notion of a maximum weight matching does not necessarily fit the underlying problem. Consider for example the case where the vertices of a graph represent players and the edge weights represent the “happiness” that the corresponding players get from this match. It is possible that two players, which are not matched to each other in a maximum weighted matching, can still coordinate and match together instead of staying in the current matching, thus becoming both individually “happier”. This is the so-called stable matching problem which has received a lot of attention the previous years due to the plethora of its applications in real life problems, including kidney exchange [@R04] and matching medical students to hospitals [@RP99]. In many applications of the stable matching problem, such as in kidney exchange, there are only a few feasible values of “happiness”. Thus in the underlying graph there are only few discrete edge weights, while many edges share the same weight. In such cases it is not clear how to compute a stable matching, as ties among edges with the same weight must be resolved. It turns out that in many cases a matching is stable in the Nash equilibrium sense if and only if it can produced by a greedy algorithm [@ADN13]. In the graph-matching game described above, a Nash equilibrium is a matching [$\mathcal{M}$]{}in which no two vertices $u,v$ can become individually happier by replacing their currently matched edges in [$\mathcal{M}$]{}with the edge $(u,v)$. Thus, a maximum weight greedy matching is an equilibrium (i.e. stable matching) that maximizes the social welfare, that is, the cumulative “happiness” of all the players. A natural algorithmic question is whether a maximum weight greedy matching can be efficiently computed. Although greedy algorithms for matching problems have been studied extensively in the past [@PS12; @ADFS95; @HK78; @DF91; @MP97; @DFP93], to the best of our knowledge not much is known about the problem of computing a maximum weight greedy matching. Related work ------------ The scenarios of matching problems where the vertices of the graph correspond to players can vary from matching employees and employers [@J79], to matching kidney donors and recipients [@R04; @ABS07]. Anshelevich, Das, and Naamad [@ADN13] and Anshelevich, Bhardwaj, and Hoefer [@ABH13] studied the price of anarchy and stability of stable matchings on weighted graphs. Furthermore, the authors in [@ADN13] provided algorithms that compute almost stable matchings. Our work is closely related to [@ADN13], although their techniques cannot be applied to our problem since we study only matchings that are greedy, whereas almost stable matchings are not. Greedy matchings have been studied extensively over the years. The classical result by Korte and Hausmann [@HK78] states that an arbitrary greedy matching is a $\frac{1}{2}$-approximation of the maximum cardinality matching, i.e. every greedy matching on unweighted graphs picks at least half of maximum number of edges that any matching can pick. For edge-weighted graphs, Avis [@Avis83] showed that every algorithm that greedily picks edges with the maximum currently available weight is a $\frac{1}{2}$-approximation of the maximum weight matching. Hence, every weighted greedy matching is also a $\frac{1}{2}$-approximation for the maximum weight greedy matching problem. Several authors studied randomized greedy algorithms for the maximum cardinality matching problem. The currently best randomized algorithm, known as [<span style="font-variant:small-caps;">Mrg</span>]{} [@ADFS95], picks the next edge to add to the matching by first selecting a random unmatched vertex $V$ of the graph and then a random unmatched neighbor of $v$. Aronson, Dyer, Frieze and Suen [@ADFS95] showed that [<span style="font-variant:small-caps;">Mrg</span>]{}breaks the $\frac{1}{2}$-barrier and that it achieves a $\frac{1}{2}+ 1/400,000$-approximation guarantee on every graph. Recently, Poloczek and Szegedy [@PS12] provided a different analysis for [<span style="font-variant:small-caps;">Mrg</span>]{}and shown that it achieves an approximation guarantee of at least $\frac{1}{2}+ \frac{1}{256}$. However, as experiments suggest, the approximation guarantee of [<span style="font-variant:small-caps;">Mrg</span>]{}can be as large as $\frac{2}{3}$ [@PS12]. Our contribution ---------------- In this paper we study the computational complexity of computing and approximating a maximum weight greedy matching in a given edge-weighted graph, i.e. a greedy matching with the greatest weight among all greedy matchings. This problem is termed <span style="font-variant:small-caps;">GreedyMatching</span>. In wide contrast to the maximum weight matching, for which many efficient algorithms are known (see [@Duan14] and the references therein), we prove that <span style="font-variant:small-caps;">GreedyMatching</span> is strongly NP-hard by a reduction from a special case of MAX2SAT. Our reduction also implies hardness of approximation; we prove that <span style="font-variant:small-caps;">GreedyMatching</span> is APX-complete, and thus it does not admit a PTAS unless P=NP. These hardness results hold even for input graphs with maximum degree at most 3 and with at most three different integer edge weights, namely with weights in the set $\{1,2,4\}$. Furthermore, by using a technique of Papadimitriou and Yannakakis [@PY91], we extend the NP-hardness proof to the interesting case where the input graph is in addition bipartite. Next, we study the decision variations <span style="font-variant:small-caps;">GreedyVertex</span> and <span style="font-variant:small-caps;">GreedyEdge</span> of the problem, where we now ask whether there exists a greedy matching in which a specific vertex $u$ or a specific edge $(u,v)$ is matched. These are both natural questions, as the designer of the stable matching might want to ensure that a specific player or a specific pair of players is matched in the solution. We prove that both <span style="font-variant:small-caps;">GreedyVertex</span> and <span style="font-variant:small-caps;">GreedyEdge</span> are also strongly <span style="font-variant:small-caps;">NP</span>-hard. As <span style="font-variant:small-caps;">GreedyMatching</span> turns out to be computationally hard, it makes sense to investigate how the complexity is affected by appropriately restricting the input. In this line of research we consider two natural parameters of the problem, for which we establish a sharp threshold behavior. As the first parameter we consider the minimum ratio $\lambda _{0}$ of any two consecutive weights. Assume that the graph has $\ell $ different edge weights $w_{1}>w_{2}>\ldots >w_{\ell }$; we define for every $i\in \lbrack \ell -1]$ the ratio $\lambda _{i}=\frac{w_{i}}{w_{i+1}}$ and the minimum ratio $\lambda _{0}=\min_{i\in \lbrack \ell -1]}\lambda _{i}$. We prove that, if $\lambda _{0}\geq 2$ then <span style="font-variant:small-caps;">GreedyMatching</span> can be solved in polynomial time, while for any constant $\lambda _{0}<2$ <span style="font-variant:small-caps;">GreedyMatching</span> is strongly NP-hard and APX-complete, even on graphs with maximum degree at most 3 and with at most three different edge weights. As the second parameter we consider the maximum edge cardinality $\mu $ of the connected components of $G(w_{i})$, among all different weights $w_{i}$, where $G(w_{i})$ is the subgraph of $G$ spanned by the edges of weight $w_{i}$. Although at first sight this parameter may seem unnatural, it resembles the number of times that the greedy algorithm has to break ties. At the stage where we have to choose among all available edges of weight $w_{i}$, it suffices to consider each connected component of the available edges of $G(w_{i})$ separately from the other components. In particular, although the weight of the final greedy matching may highly depend on the order of the chosen edges within a connected component, it is independent of the ordering that the various different connected components are processed. Thus $\mu $ is a reasonable parameter for <span style="font-variant:small-caps;">GreedyMatching</span>. In the case $\mu =1$ there exists a unique greedy matching for $G$ which can be clearly computed in polynomial time. We prove that <span style="font-variant:small-caps;">GreedyMatching</span> is strongly NP-hard and APX-complete for $\mu \geq 2$, even on graphs with maximum degree at most 3 and with at most five different edge weights. On the positive side, we consider a special class of weighted graphs, called bush graphs, where all edges of the same weight in $G$ form a star (bush). We present a randomized approximation algorithm ([<span style="font-variant:small-caps;">Rgma</span>]{}) for [<span style="font-variant:small-caps;">GreedyMatching</span>]{}on bush graphs and we highlight an unexpected connection between [<span style="font-variant:small-caps;">Rgma</span>]{}and the randomized [<span style="font-variant:small-caps;">Mrg</span>]{}algorithm for greedily approximating the maximum cardinality matching on unweighted graphs. In particular we show that, if the approximation ratio of <span style="font-variant:small-caps;">Rgma</span> for <span style="font-variant:small-caps;">GreedyMatching</span> on bush graphs is $\rho$, then for every $\epsilon >0$ <span style="font-variant:small-caps;">Mrg</span> [@ADFS95] is a $(\rho-\epsilon )$-approximation algorithm for the maximum cardinality matching. We conjecture that a tight bound for $\rho $ is $\frac{2}{3}$; among our results we prove our conjecture true for two subclasses of bush graphs. Proving a tight bound for the approximation ratio of <span style="font-variant:small-caps;">Mrg</span> on unweighted graphs (and thus also proving a tight value for $\rho $) is a long-standing open problem [@PS12; @ADFS95; @DF91]. This unexpected relation of our <span style="font-variant:small-caps;">Rgma</span> algorithm with the <span style="font-variant:small-caps;">Mrg</span> algorithm may provide new insights for solving this problem. Preliminaries {#sec:pre} ============= Every graph considered in this paper is undirected. For any graph $G = (V, E)$ we use $G + u$ to denote the graph $G' = (V', E')$ where $V' = V \cup \{u\}$ and $E'$ is consisted by the set $E$ and all the edges the vertex $u$ belongs to. Similarly $G - V'$ denotes the induced graph of $G$ defined by $V \setminus V'$, where $V'\subseteq V$. We study graphs $G = (V, E)$ with positive edge weights, i.e. each edge $e=(u,v) \in E$ has a weight $w(e) = w(u,v) > 0$. The degree of a vertex $u$ is the number of its adjacent vertices in $G$. We use $G(w_i)$ to denote the subgraph of $G$ spanned by the edges of weight $w_i$. A matching ${\ensuremath{\mathcal{M}}\xspace}\subseteq E$ is a set of edges such that no pair of them are adjacent. The weight of a matching [$\mathcal{M}$]{}is the sum of the weights of the edges in [$\mathcal{M}$]{}, formally $w({\ensuremath{\mathcal{M}}\xspace}) = \sum_{e \in {\ensuremath{\mathcal{M}}\xspace}}w(e)$. A greedy matching is a maximal matching constructed by the Greedy Matching Procedure. Input: Graph $G = (V, E)$, with $w_1 > w_2 > \ldots > w_{\ell}$ edge weight values\ Output: Greedy matching [$\mathcal{M}$]{} 1. ${\ensuremath{\mathcal{M}}\xspace}\leftarrow \emptyset$ 2. for $i = 1 \ldots \ell$ do 3. while there is an $e \in E$ such that $w(e) = w_i$ do 4. Pick an edge $e^* \in E$ with $w(e^) = w_i$ and add it to [$\mathcal{M}$]{}; \[step4\] 5. Remove all edges adjacent to $e^$ from $E$; Notice that in Step \[step4\] the edge that is added to the matching [$\mathcal{M}$]{}is not specified explicitly. The rule that specifies which edge is chosen in Step \[step4\] can be deterministic or randomized, resulting to a specific greedy matching algorithm. We will use ${\textsc{Opt}\xspace}(G)$ to denote the optimum of [<span style="font-variant:small-caps;">GreedyMatching</span>]{}with input $G$, i.e. the weight of the maximum greedy matching of $G$. <span style="font-variant:small-caps;">Instance</span>: Graph $G=(V,E)$ with positive edge weights.\ <span style="font-variant:small-caps;">Task</span>: Compute a maximum weight greedy matching [$\mathcal{M}$]{}for $G$. Furthermore, we study another two related problems, where we ask whether there is a greedy matching that matches a specific vertex or a specific edge. <span style="font-variant:small-caps;">Instance</span>: Graph $G=(V,E)$ with positive edge weights and a vertex ${v\in V}$.\ <span style="font-variant:small-caps;">Question</span>: Is there a greedy matching [$\mathcal{M}$]{}such that $(v,u) \in {\ensuremath{\mathcal{M}}\xspace}$, for some ${u \in V}$? <span style="font-variant:small-caps;">Instance</span>: Graph $G=(V,E)$ with positive edge weights and an edge ${(u,v) \in E}$.\ <span style="font-variant:small-caps;">Question</span>: Is there a greedy matching [$\mathcal{M}$]{}such that ${(v,u) \in {\ensuremath{\mathcal{M}}\xspace}}$? Computational hardness of [<span style="font-variant:small-caps;">GreedyMatching</span>]{}\[sec:hardness\] ========================================================================================================== In this section we study the complexity of computing a maximum weight greedy matching. In Section \[APX-subsec\] we prove that [<span style="font-variant:small-caps;">GreedyMatching</span>]{}is strongly NP-hard and APX-complete, even on graphs with maximum degree at most 3 and with at most three different integer weight values. By slightly modifying our reduction of Section \[APX-subsec\], we first prove in Section [hardness-bipartite-subsec]{} that [<span style="font-variant:small-caps;">GreedyMatching</span>]{}remains strongly NP-hard also when the graph is in addition bipartite, and we then prove in Section \[hardness-additional-subsec\] that also the two decision problem variations [<span style="font-variant:small-caps;">GreedyVertex</span>]{}and [<span style="font-variant:small-caps;">GreedyEdge</span>]{}are also strongly NP-hard. Our hardness reductions are from the MAX2SAT(3) problem [@Ausiello1999; @RRR98], which is the special case of MAX-SAT where in the input CNF formula $\phi $ every clause has at most 2 literals and every variable appears in at most 3 clauses; we call such a formula $\phi $ a 2SAT(3) formula. Note that the decision version of [<span style="font-variant:small-caps;">GreedyMatching</span>]{}, where we ask whether there exists a greedy matching with weight at least $B$, belongs to the class NP. Indeed we are able to verify in polynomial time whether a given matching ${\ensuremath{\mathcal{M}}\xspace}$ is maximal, greedy and has weight at least $B$. The maximality and the weight of the matching ${\ensuremath{\mathcal{M}}\xspace}$ can be computed and checked in linear time. To check whether ${\ensuremath{\mathcal{M}}\xspace}$ is greedy, we first check whether the largest edge weight in ${\ensuremath{\mathcal{M}}\xspace}$ equals the largest edge weight in $G$. In this case we remove from $G$ all vertices incident to the highest weight edges of ${\ensuremath{\mathcal{M}}\xspace}$ and we apply recursively the same process in the resulting induced subgraph. Then ${\ensuremath{\mathcal{M}}\xspace}$ is greedy if and only if we end up with a graph with no edges. Overview of the reduction.\[overview-subsec\] --------------------------------------------- Given a 2SAT(3) formula $\phi $ with $m$ clauses and $n$ variables $x_{1},\ldots ,x_{n}$ we construct an undirected graph $G$ with $10n+m$ vertices and $9n+2m$ edges. Then we prove that there exists a truth assignment that satisfies at least $k$ clauses of $\phi $ if and only if there exists a greedy matching [$\mathcal{M}$]{}in $G$ with weight at least $14n+k$. Without loss of generality we make the following assumptions on $\phi $. Firstly, if a variable occurs only with positive (resp. only with negative) literals, then we trivially set it true (resp. false) and remove the associated clauses. Furthermore, without loss of generality, if a variable $x_{i}$ appears three times in $\phi $, we assume that it appears once as a positive literal $x_{i}$ and two times as a negative literal $\overline{x_{i}}$; otherwise we rename the negation with a new variable. Similarly, if $x_{i}$ appears two times in $\phi $, then it appears once as a positive literal $x_{i}$ and once as a negative literal $\overline{x_{i}}$. For each variable $x_{i}$ we create a subgraph ${\ensuremath{\mathcal{G}}\xspace}_{x_{i}}$ and for each clause $C_{j}$ we create one vertex $v_{j}$. The vertices created from the clauses will be called $v$-vertices. Each subgraph ${\ensuremath{\mathcal{G}}\xspace}_{x_{i}}$ is a path with 10 vertices, where three of them are distinguished; the vertices $\alpha _{x_{i}},\beta _{x_{i}}$ and $\gamma _{x_{i}}$. Each distinguished vertex can be connected with at most one $v$-vertex that represents a clause. Furthermore, every $v$-vertex is connected with at most two vertices from the subgraphs ${\ensuremath{\mathcal{G}}\xspace}_{x_{i}}$; one distinguished vertex from each of the subgraphs ${\ensuremath{\mathcal{G}}\xspace}_{x_{i}}$ that correspond to the variables of the clause. The edge weights in the subgraphs ${\ensuremath{\mathcal{G}}\xspace}_{x_{i}}$ are not smaller than the weights of the edges connecting the $v$-vertices with the distinguished vertices of the subgraphs ${\ensuremath{\mathcal{G}}\xspace}_{x_{i}}$. The construction\[construction-subsec\] --------------------------------------- The gadget ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ that we create for variable $x_i$ is illustrated in Figure \[fig:gx\]; the distinguished vertices of ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ are $\alpha_{x_i}$, $\beta_{x_i}$ and $\gamma_{x_i}$. The vertex $\alpha_{x_i}$ corresponds to the positive literal of the variable and vertices $\beta_{x_i}$ and $\gamma_{x_i}$ correspond to the negative literal $\overline{x_i}$. \[fig:gx\] \(01) at (0,1) [$\beta_{x_i}$]{}; (10) at (1,0) \[label=below:$p_{x_i}$\]; (01) – node\[weight\] [$1$]{} (10); (21) at (2,1) \[label=above:$q_{x_i}$\]; (21) – node\[weight\] [$3$]{} (10); (30) at (3,0) \[label=below:$r_{x_i}$\]; (21) – node\[weight\] [$4$]{} (30); (41) at (4,1) [$\alpha_{x_i}$]{}; (41) – node\[weight\] [$4$]{} (30); (51) at (6,1) [$\gamma_{x_i}$]{}; (41) – node\[weight\] [$4$]{} (51); (70) at (7,0) \[label=below:$y_{x_i}$\]; (51) – node\[weight\] [$4$]{} (70); (81) at (8,1) \[label=above:$z_{x_i}$\]; (81) – node\[weight\] [$4$]{} (70); (90) at (9,0) \[label=below:$s_{x_i}$\]; (81) – node\[weight\] [$3$]{} (90); (101) at (10,1) \[label=above:$t_{x_i}$\]; (101) – node\[weight\] [$1$]{} (90); The vertex $v_{j}$ associated to clause $C_{j}$, where $j\in \lbrack m]$, is made adjacent to the vertices that correspond to the literals associated with that clause. For example, if $C_{j}=(x_{1}\vee \overline{x_{2}})$ we will connect the vertex $v_{j}$ with one of the vertices $\alpha _{x_{1}},\beta _{x_{1}},\gamma _{x_{1}}$ and with one of the vertices $\alpha _{x_{2}},\beta _{x_{2}},\gamma _{x_{2}}$. In order to make these connections in a consistent way, we first fix an arbitrary ordering over the clauses. If the variable $x_{i}$ occurs as a positive literal in the clause $C_{j}$, then we add the edge $(v_{j},\alpha _{x_{i}})$ of weight 3. Next, if $C_{j}$ is the first clause that the variable $x_{i}$ occurs with a negative literal (in the fixed ordering of the clauses), then we add the edge $(v_{j},\beta _{x_{i}})$ of weight 1. Finally, if the clause $C_{j}$ is the second clause that the variable $x_{i}$ occurs as a negative literal, then we add the edge $(v_{j},\gamma _{x_{i}})$ of weight 3. That is, if a variable $x_{i}$ appears only two times in $\phi $, then only the two distinguished vertices $\alpha _{x_{i}}$ and $\beta _{x_{i}}$ of ${\ensuremath{\mathcal{G}}\xspace}_{x_{i}}$ are adjacent to a $v$-vertex. This completes the construction of the graph $G$. Note that, by the construction of $G$, in any maximum greedy matching of $G$, there are exactly four alternative ways to match the edges of each of the subgraphs ${\ensuremath{\mathcal{G}}\xspace}_{x_{i}}$, as illustrated in Fig. \[fig:man\]-\[fig:bad2\]. APX-completeness\[APX-subsec\] ------------------------------ In order to prove that [<span style="font-variant:small-caps;">GreedyMatching</span>]{}is APX-complete, first we prove in the next lemma that given an assignment that satisfies at least $k$ clauses we can construct a greedy matching with weight at least $14n+k$. The intuition for this lemma is as follows. Starting with a given satisfying truth assignment $\tau$ for the input formula $\phi$, we first construct the matching ${\ensuremath{\mathcal{M}^-}\xspace}$ in every ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ (cf. Figure \[fig:man\]), and thus the $\beta$-vertices are initially free to be matched to $v$-vertices. Then, if a variable $x_i$ is true in $\tau$, we change the matching of ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ from ${\ensuremath{\mathcal{M}^-}\xspace}$ to ${\ensuremath{\mathcal{M}^+}\xspace}$ (cf. Figure \[fig:map\]), such that only the $\alpha$-vertex (and not the $\beta$ and $\gamma$-vertices) of ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ is free to be matched to a $v$-vertex. On the other hand, if the variable $x_i$ is false in $\tau$, then we either keep the matching ${\ensuremath{\mathcal{M}^-}\xspace}$ in ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$, or we replace ${\ensuremath{\mathcal{M}^-}\xspace}$ with ${\ensuremath{\mathcal{M}^{--}}\xspace}$ in ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ (cf. Figure \[fig:mann\]). Note that in ${\ensuremath{\mathcal{M}^-}\xspace}$ only $\beta_{x_i}$ is free to be matched, while in ${\ensuremath{\mathcal{M}^{--}}\xspace}$ both $\beta_{x_i}$ and $\gamma_{x_i}$ are free to be matched with a $v$-vertex; in both cases the $\alpha$-vertex of ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ is “blocked” from being matched to a $v$-vertex. Then, using the fact that $\tau$ satisfies at least $k$ clauses of $\phi$, we can construct a matching of $G$ where $k$ $v$-vertices are matched and the total weight of this matching is at least ${14n+k}$. \[lem:one\] If there is an assignment that satisfies at least $k$ clauses then, there is a greedy matching with weight at least $14n+k$. Given an assignment that satisfies $k$ clauses we will construct a greedy matching of $G$ with weight $14n+k$ by making use of the three matchings ${\ensuremath{\mathcal{M}^-}\xspace}$, ${\ensuremath{\mathcal{M}^{--}}\xspace}$, and ${\ensuremath{\mathcal{M}^+}\xspace}$ of ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$, as illustrated in Figures \[fig:man\]-\[fig:map\]. All these three matchings of ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ are greedy; furthermore note that there also exists a fourth greedy matching [$\mathcal{M}^b$]{}of ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ (see Fig. \[fig:bad2\]) which will not be used in the proof of the lemma. \(01) at (0,1) [$\beta_{x_i}$]{}; (10) at (1,0) \[label=below:$p$\]; (01) – node\[weight\] [$1$]{} (10); (21) at (2,1) \[label=above:$q$\]; (21) – node\[weight\] [$3$]{} (10); (30) at (3,0) \[label=below:$r$\]; (21) – node\[weight\] [$4$]{} (30); (41) at (4,1) [$\alpha_{x_i}$]{}; (41) – node\[weight\] [$4$]{} (30); (51) at (6,1) [$\gamma_{x_i}$]{}; (41) – node\[weight\] [$4$]{} (51); (70) at (7,0) \[label=below:$y$\]; (51) – node\[weight\] [$4$]{} (70); (81) at (8,1) \[label=above:$z$\]; (81) – node\[weight\] [$4$]{} (70); (90) at (9,0) \[label=below:$s$\]; (81) – node\[weight\] [$3$]{} (90); (101) at (10,1)\[label=above:$t$\] ; (101) – node\[weight\] [$1$]{} (90); \(01) at (0,1) [$\beta_{x_i}$]{}; (10) at (1,0) \[label=below:$p$\]; (01) – node\[weight\] [$1$]{} (10); (21) at (2,1) \[label=above:$q$\]; (21) – node\[weight\] [$3$]{} (10); (30) at (3,0) \[label=below:$r$\]; (21) – node\[weight\] [$4$]{} (30); (41) at (4,1) [$\alpha_{x_i}$]{}; (41) – node\[weight\] [$4$]{} (30); (51) at (6,1) [$\gamma_{x_i}$]{}; (41) – node\[weight\] [$4$]{} (51); (70) at (7,0) \[label=below:$y$\]; (51) – node\[weight\] [$4$]{} (70); (81) at (8,1) \[label=above:$z$\]; (81) – node\[weight\] [$4$]{} (70); (90) at (9,0) \[label=below:$s$\] ; (81) – node\[weight\] [$3$]{} (90); (101) at (10,1) \[label=above:$t$\]; (101) – node\[weight\] [$1$]{} (90); \(01) at (0,1) [$\beta_{x_i}$]{}; (10) at (1,0) \[label=below:$p$\]; (01) – node\[weight\] [$1$]{} (10); (21) at (2,1) \[label=above:$q$\]; (21) – node\[weight\] [$3$]{} (10); (30) at (3,0) \[label=below:$r$\]; (21) – node\[weight\] [$4$]{} (30); (41) at (4,1) [$\alpha_{x_i}$]{}; (41) – node\[weight\] [$4$]{} (30); (51) at (6,1) [$\gamma_{x_i}$]{}; (41) – node\[weight\] [$4$]{} (51); (70) at (7,0) \[label=below:$y$\]; (51) – node\[weight\] [$4$]{} (70); (81) at (8,1) \[label=above:$z$\]; (81) – node\[weight\] [$4$]{} (70); (90) at (9,0) \[label=below:$s$\]; (81) – node\[weight\] [$3$]{} (90); (101) at (10,1) \[label=above:$t$\]; (101) – node\[weight\] [$1$]{} (90); We construct the greedy matching of $G$ with weight $14n+k$ as follows. Firstly, we set all the matchings for the subgraphs ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ to be the greedy matching ${\ensuremath{\mathcal{M}^-}\xspace}$, thus incurring a total weight of $14n$ from the currently matched edges. Then we process sequentially each clause $C_j$ of the formula $\phi$. If a clause $C_j$ is satisfied in the given truth assignment by at least one positive literal, then we choose one of these literals arbitrarily, say $x_i$, and we change the matching of ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ to ${\ensuremath{\mathcal{M}^+}\xspace}$; furthermore we match the edge $(v_{j},\alpha_{x_i})$ which has weight 3. In this case we replaced the matched edges $(p_{x_i},q_{x_i})$ and $(r_{x_i},\alpha_{x_i})$ of ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ with total weight 7 by the matched edges $(\beta_{x_i},p_{x_i})$, $(q_{x_i},r_{x_i})$, and $(v_j,\alpha_{x_i})$ with total weight 8, i.e. we increased the weight of the matching by 1. Assume that a clause $C_j$ is satisfied in the given assignment only by negative literals. If at least one of these literals of $C_j$ corresponds to a $\beta$-vertex, then we match the edge $(v_{j},\beta_{x_i})$ of weight 1. Thus in this case we also increase the total weight of the matched edges by 1. Finally, if all of these literals of $C_j$ correspond to $\gamma$-vertices, then we choose one of them arbitrarily, say $\overline{x_i}$, and we change the matching of ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ to ${\ensuremath{\mathcal{M}^{--}}\xspace}$; furthermore we match the edge $(v_{j},\gamma_{x_i})$ of weight 3. In this case we replaced the matched edges $(\gamma_{x_i},y_{x_i})$ and $(z_{x_i},s_{x_i})$ of ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ with total weight 7 by the matched edges $(v_j,\gamma_{x_i})$, $(y_{x_i},z_{x_i})$, and $(s_{x_i},t_{x_i})$ with total weight 8, i.e. we increased the weight of the matching by 1. This completes the required matching $\mathcal{M}$ of the graph $G$. Since we started with a matching of total weight $14n$ and we added weight 1 for each of the $k$ satisfied clauses in $\phi$, note that the total weight of $\mathcal{M}$ is $14n+k$. In this matching $\mathcal{M}$, each of the induced subgraphs ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ of $G$ is greedily matched. Furthermore all the remaining edges of $G$ are edges that join a $v$-vertex with a distinguished vertex $\alpha_{x_i}$ (resp. $\beta_{x_i}$, $\gamma_{x_i}$). Note that the weight of each of these edges is smaller than or equal to the weight of the edges adjacent to $\alpha_{x_i}$ (resp. $\beta_{x_i}$, $\gamma_{x_i}$) within the subgraph ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$. Thus, the matching $\mathcal{M}$ of $G$ can be constructed greedily. Moreover, since $\mathcal{M}$ can be potentially further extended greedily to a matching with larger weight, it follows that the maximum greedy matching of $G$ is at least $14n+k$. Next we prove in Lemma \[lem:at-least\] that, if there is a greedy matching with weight $14n+k$, then there is an assignment that satisfies at least $k$ clauses. In order to prove Lemma \[lem:at-least\], first we prove in Lemma \[lem:one-of-ag\] a crucial property of the constructed graph $G$, namely that in any greedy matching at most one of the vertices $\alpha_{x_i}$ and $\gamma_{x_i}$ can be matched with a $v$-vertex. \[lem:one-of-ag\] Let $\mathcal{M}$ be an arbitrary greedy matching of $G$ and let $i\in \{1,2,\ldots,n\}$. Then, in the subgraph ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$, at most one of the vertices $\alpha_{x_i}$ and $\gamma_{x_i}$ can be matched with a $v$-vertex. The proof is done by contradiction. Assume otherwise that both $\alpha_{x_i}$ and $\gamma_{x_i}$ are matched with some $v$-vertices in $\mathcal{M}$. Note that both these edges that connect the vertices $\alpha_{x_i}$ and $\gamma_{x_i}$ with the corresponding $v$-vertices have weight 3. Furthermore, none of the edges $(\alpha_{x_i},r)$, $(\gamma_{x_i},y)$, and $(\alpha_{x_i},\gamma_{x_i})$ belong to $\mathcal{M}$. Thus, since the weight of the edge $(\alpha_{x_i},\gamma_{x_i})\notin \mathcal{M}$ is 4, it follows $\mathcal{M}$ is not greedy, which is a contradiction. That is, if both edges $(\alpha_{x_i},r)$ and $(\gamma_{x_i},y)$ of the subgraph ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ are not matched within $\mathcal{M}$, then $(\alpha_{x_i},\gamma_{x_i})\in \mathcal{M}$, as it is illustrated in the “bad” matching $\mathcal{M}^{b}$ of Fig. \[fig:bad2\]. \(01) at (0,1) [$\beta_{x_i}$]{}; (10) at (1,0) \[label=below:$p$\]; (01) – node\[weight\] [$1$]{} (10); (21) at (2,1) \[label=above:$q$\]; (21) – node\[weight\] [$3$]{} (10); (30) at (3,0) \[label=below:$r$\]; (21) – node\[weight\] [$4$]{} (30); (41) at (4,1) [$\alpha_{x_i}$]{}; (41) – node\[weight\] [$4$]{} (30); (51) at (6,1) [$\gamma_{x_i}$]{}; (41) – node\[weight\] [$4$]{} (51); (70) at (7,0) \[label=below:$y$\]; (51) – node\[weight\] [$4$]{} (70); (81) at (8,1) \[label=above:$z$\]; (81) – node\[weight\] [$4$]{} (70); (90) at (9,0) \[label=below:$s$\]; (81) – node\[weight\] [$3$]{} (90); (101) at (10,1) \[label=above:$t$\] ; (101) – node\[weight\] [$1$]{} (90); We are now ready to prove Lemma \[lem:at-least\]. \[lem:at-least\] If there is a greedy matching with weight at least $14n+k$ in $G$, then there exists an assignment that satisfies at least $k$ clauses of the formula $\phi$. Let [$\mathcal{M}$]{}be a maximum weight greedy matching of $G$ and assume that [$\mathcal{M}$]{}has weight at least $14n+k$. First we show that we can assume without loss of generality that, for every $i\in [n]$, the edges of the induced subgraph ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ are matched in [$\mathcal{M}$]{} according to one of the four matchings ${\ensuremath{\mathcal{M}^-}\xspace}$, ${\ensuremath{\mathcal{M}^{--}}\xspace}$, ${\ensuremath{\mathcal{M}^+}\xspace}$, and [$\mathcal{M}^b$]{}(see Figures \[fig:man\]-\[fig:bad2\]). Assume that the edge $(\gamma_{x_i},y_{x_i})$ is matched in [$\mathcal{M}$]{}. Then clearly the edge $(z_{x_i},s_{x_i})$ is also matched in [$\mathcal{M}$]{}, since this is the only valid greedy option for the right part of ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$. Assume that the vertex $\gamma_{x_i}$ is matched in [$\mathcal{M}$]{}with a vertex different than $y_{x_i}$. Then similarly the edges $(y_{x_i},z_{x_i})$ and $(s_{x_i},t_{x_i})$ are matched in [$\mathcal{M}$]{}. On the other hand, assume that the edge $(\alpha_{x_i},r_{x_i})$ is matched in [$\mathcal{M}$]{}. Then the edge $(p_{x_i},q_{x_i})$ is also matched in [$\mathcal{M}$]{}due to the assumption that [$\mathcal{M}$]{}is greedy. Finally assume that the vertex $\alpha_{x_i}$ is matched in [$\mathcal{M}$]{}with a vertex different than $r_{x_i}$. Then, since [$\mathcal{M}$]{}is greedy, the edge $(q_{x_i},r_{x_i})$ is matched in [$\mathcal{M}$]{}. Furthermore, since [$\mathcal{M}$]{}has the greatest weight among the greedy matchings of $G$ by assumption, the vertex $\beta_{x_i}$ is matched in [$\mathcal{M}$]{}either with vertex $p_{x_i}$ or with its adjacent $v$-vertex. If $\beta_{x_i}$ is matched with the $v$-vertex, then we replace this matched edge in [$\mathcal{M}$]{}by the matched edge $(\beta_{x_i},p_{x_i})$ (if the vertex $p_{x_i}$ is unmatched) and we get an other greedy matching with the same weight. Therefore, we can assume without loss of generality that, for every $i\in [n]$, the edges of the induced subgraph ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ are matched in [$\mathcal{M}$]{} according to one of the four matchings ${\ensuremath{\mathcal{M}^-}\xspace}$, ${\ensuremath{\mathcal{M}^{--}}\xspace}$, ${\ensuremath{\mathcal{M}^+}\xspace}$, and [$\mathcal{M}^b$]{}, see Figures \[fig:man\]-\[fig:bad2\]. We construct from [$\mathcal{M}$]{}a truth assignment that satisfies at least $k$ clauses of the formula $\phi$, as follows. If the edges of the induced subgraph ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ are matched in [$\mathcal{M}$]{}according to one of the matchings ${\ensuremath{\mathcal{M}^-}\xspace}$ or ${\ensuremath{\mathcal{M}^{--}}\xspace}$, then we set the value of $x_i$ to false. If the edges of ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ are matched according to ${\ensuremath{\mathcal{M}^+}\xspace}$, then we set the value of $x_i$ to true. Otherwise, if the edges of ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ are matched according to [$\mathcal{M}^b$]{}, then set the truth value of $x_i$ arbitrarily. Let now $i \in [n]$. Since the edges of the induced subgraph ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ are matched according to one of the matchings ${\ensuremath{\mathcal{M}^-}\xspace}$, ${\ensuremath{\mathcal{M}^{--}}\xspace}$, ${\ensuremath{\mathcal{M}^+}\xspace}$, and [$\mathcal{M}^b$]{}, as we proved above, it follows that the vertices $\alpha_{x_i}$ and $\beta_{x_i}$ are not simultaneously matched with their associated $v$-vertices in [$\mathcal{M}$]{}. Furthermore, Lemma \[lem:one-of-ag\] implies that at most one of the vertices $\alpha_{x_i}$ and $\gamma_{x_i}$ can be matched with their associated $v$-vertices in [$\mathcal{M}$]{}. Therefore the constructed truth assignment is valid. Let $i \in [n]$. If ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ is matched according to [$\mathcal{M}^b$]{}in [$\mathcal{M}$]{}, then ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ clearly contributes weight 14 to the total weight of [$\mathcal{M}$]{}. Assume that ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ is matched according to ${\ensuremath{\mathcal{M}^+}\xspace}$ in [$\mathcal{M}$]{}, i.e. assume that $\alpha_{x_i}$ is matched with a $v$-vertex. Then $\gamma_{x_i}$ is matched with $y_{x_i}$, and thus the right part of ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ (i.e. the part between vertices $\gamma_{x_i}$ and $t_{x_i}$) contributes weight 7 to the total weight of [$\mathcal{M}$]{}. Furthermore the left part of ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ (i.e. the part between vertices $\beta_{x_i}$ and $\alpha_{x_i}$) contributes weight 1+4+3=8 to the total weight of [$\mathcal{M}$]{}. That is, ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ contributes weight 15 to the total weight of [$\mathcal{M}$]{}. Assume now that ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ is matched according to ${\ensuremath{\mathcal{M}^-}\xspace}$ or ${\ensuremath{\mathcal{M}^{--}}\xspace}$ in [$\mathcal{M}$]{}. If $\beta_{x_i}$ is matched with $p_{x_i}$ in [$\mathcal{M}$]{}, then the left part of ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$contributes weight 3+4=7 to the total weight of [$\mathcal{M}$]{}. Otherwise, if $\beta_{x_i}$ is matched with its adjacent $v$-vertex in [$\mathcal{M}$]{}, then the left part of ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ contributes weight 1+3+4=8 to the total weight of [$\mathcal{M}$]{}. Similarly, if $\gamma_{x_i}$ is matched with $y_{x_i}$ in [$\mathcal{M}$]{}, then the right part of ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ contributes weight 4+3=7 to the total weight of [$\mathcal{M}$]{}. Otherwise, if $\gamma_{x_i}$ is matched with its adjacent $v$-vertex in [$\mathcal{M}$]{}, then the right part of ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ contributes weight 3+4+1=8 to the total weight of [$\mathcal{M}$]{}. Summarizing, if $0 \leq \ell \leq 2$ of the vertices $\{\alpha_{x_i}, \beta_{x_i}, \gamma_{x_i}\}$ are matched with $v$-vertices in [$\mathcal{M}$]{}, then ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ contributes weight $14+\ell$ to the total weight of [$\mathcal{M}$]{}. Therefore, since $G$ has $n$ induced subgraphs and the weight of [$\mathcal{M}$]{}is at least $14n+k$, it follows that $\ell\geq k$ $v$-vertices are matched in [$\mathcal{M}$]{}. For every vertex $v_j$ that is matched with a vertex $\alpha_{x_i}$ in [$\mathcal{M}$]{}, the clause $C_j$ contains the literal $x_i$ and the variable $x_i$ is set to true by the construction of the truth assignment. Thus $C_j$ is satisfied. Similarly, for every vertex $v_j$ that is matched with a vertex $\beta_{x_i}$ or $\gamma_{x_i}$ in [$\mathcal{M}$]{}, the clause $C_j$ contains the literal $\overline{x_i}$ and the variable $x_i$ is set to false by the construction of the truth assignment. Thus $C_j$ is again satisfied. Therefore, since $\ell\geq k$ $v$-vertices are matched in [$\mathcal{M}$]{}, it follows that there are $\ell\geq k$ satisfied clauses of $\phi$ in the constructed assignment. In the following theorem we conclude with the main result of this section. \[thm:inai\] [<span style="font-variant:small-caps;">GreedyMatching</span>]{}is strongly NP-hard and APX-complete. In particular, unless P=NP, <span style="font-variant:small-caps;">GreedyMatching</span>admits no PTAS, even on graphs with maximum degree at most 3 and with at most three different integer weight values. It follows by Lemmas \[lem:one\] and \[lem:at-least\] that there is a greedy matching [$\mathcal{M}$]{}in $G$ with weight at least $14n+k$ if and only if there is a truth assignment that satisfies at least $k$ clauses in the 2SAT(3) formula $\phi $. Thus it follows that <span style="font-variant:small-caps;">GreedyMatching</span> is NP-hard, since MAX2SAT(3) is also NP-hard [@Ausiello1999; @RRR98]. Furthermore, since the graph $G$ has three different weight values (namely 1, 3, and 4), it follows that <span style="font-variant:small-caps;">GreedyMatching</span> is strongly NP-hard. Denote by OPT$_{\text{Max2SAT(3)}}(\phi )$ the greatest number of clauses that can be simultaneously satisfied by a truth assignment of $\phi $. Furthermore denote by OPT$_{\text{Greedy}}(G)$ the maximum weight of a greedy matching of the graph $G$ that is constructed from $\phi $ by our reduction. Recall by construction that $G$ has 3 different integer weights and the maximum degree is 3. Then Lemma \[lem:one\] implies that OPT$_{\text{Greedy}}(G)\geq 14n+$OPT$_{\text{Max2SAT(3)}}(\phi )$. Note that a random truth assignment satisfies each clause of $\phi $ with probability $\frac{7}{8}$, and thus there exists a truth assignment that satisfies at least $\frac{7}{8}m$ clauses of $\phi $, where $m$ is the number of clauses, and thus OPT$_{\text{Max2SAT(3)}}(\phi )\geq \frac{7}{8}m$. Since every clause has at most 2 literals in $\phi $, it follows that $m\geq \frac{n}{2}$, and thus OPT$_{\text{Max2SAT(3)}}(\phi )\geq \frac{7}{8}m\geq \frac{7}{16}n $. Assume that there is a PTAS for computing OPT$_{\text{Greedy}}(G)$. Then for every $\varepsilon >0$ we can compute in polynomial time a greedy matching $\mathcal{M}$ of $G$ such that $\|\mathcal{M}\|\geq (1-\varepsilon )\cdot $OPT$_{\text{Greedy}}(G)$. Given such a matching $\mathcal{M}$ we can compute by Lemma \[lem:at-least\] a truth assignment $\tau $ of $\phi $ such that $\|\tau (\phi )\|\geq \|\mathcal{M}\|-14n$. Therefore:$$\begin{aligned} \|\tau (\phi )\| &\geq &(1-\varepsilon )\cdot OPT_{\text{Greedy}}(G)-14n \\ &\geq &(1-\varepsilon )\cdot (14n+OPT_{\text{Max2SAT(3)}}(\phi ))-14n \\ &\geq &(1-\varepsilon )\cdot OPT_{\text{Max2SAT(3)}}(\phi )-14\varepsilon \cdot \frac{16}{7}OPT_{\text{Max2SAT(3)}}(\phi ) \\ &\geq &(1-33\varepsilon )\cdot OPT_{\text{Max2SAT(3)}}(\phi )\end{aligned}$$ That is, assuming a PTAS for computing OPT$_{\text{Greedy}}(G)$, we obtain a PTAS for computing OPT$_{\text{Max2SAT(3)}}(\phi )$. This is a contradiction by [@Ausiello1999], unless P=NP. This proves that <span style="font-variant:small-caps;">GreedyMatching</span>is APX-hard. Furthermore [<span style="font-variant:small-caps;">GreedyMatching</span>]{}clearly belongs to the class APX, as any greedy matching algorithm achieves an $\frac{1}{2}$-approximation for [<span style="font-variant:small-caps;">GreedyMatching</span>]{}, and thus [<span style="font-variant:small-caps;">GreedyMatching</span>]{}is APX-complete. Hardness of [<span style="font-variant:small-caps;">GreedyMatching</span>]{}in Bipartite graphs \[hardness-bipartite-subsec\] ----------------------------------------------------------------------------------------------------------------------------- The graph $G$ that we constructed from $\phi$ (see Section [construction-subsec]{}) is not necessarily bipartite, as it may contain an odd-length cycle. More specifically, it is possible that the following cycle of length 9 exists: $$v \rightarrow \beta_{x_i} \rightarrow p \rightarrow q \rightarrow r \rightarrow \alpha_{x_i} \rightarrow v^{\prime }\rightarrow \gamma_{x_j} \rightarrow \alpha_{x_j} \rightarrow v.$$ However, as we prove in this section, [<span style="font-variant:small-caps;">GreedyMatching</span>]{}remains strongly NP-hard also when the graph is in addition bipartite. To prove this (cf. Theorem \[thm:bip-hard\]), we slightly modify our reduction of Section \[construction-subsec\] and the proofs of Section [APX-subsec]{}, as follows. We start with a 2-CNF formula $\phi $, where every variable appears in an arbitrary number of clauses. We may assume without loss of generality that every variable appears in $\phi $ at least three times; otherwise we may add dummy copies of existing clauses. Then we create from $\phi $ an equivalent 2-CNF formula $\phi ^{\prime }$ using a technique of Papadimitriou and Yannakakis [@PY91]. More specifically, for every variable $x_{i}$ that appears $l\geq 3$ times in $\phi $, we replace $x_{i}$ by $l$ new variables $x_{i_{1}},\ldots ,x_{i_{l}}$, one for every clause of $\phi _{1}$ in which $x_{i}$ initially appeared. Furthermore we add the $l$ extra clauses $(\overline{x_{i_{1}}}\vee x_{i_{2}})$, $(\overline{x_{i_{2}}}\vee x_{i_{3}})$, $\ldots $, $(\overline{x_{i_{l}}}\vee x_{i_{1}})$. Denote by $\phi ^{\prime }$ the resulting 2-CNF formula after performing these operations for every $i=1,2,\ldots ,n$. Note that in $\phi ^{\prime }$ a variable $x_{i_{j}}$ occurs exactly in three clauses: two times as $\overline{x_{i_{j}}}$ and one as $x_{i_{j}}$ if $x_{i_{j}}$ was negative in $\phi $, or two times as $x_{i_{j}}$ and one as $\overline{x_{i_{j}}}$ if $x_{i_{j}}$ was positive in $\phi $. Furthermore, each variable $x_{i_{j}}$ occurs in one old clause from $\phi $ and in two new clauses in $\phi ^{\prime }$. We will use again the gadgets ${\ensuremath{\mathcal{G}}\xspace}_{x_{i_{j}}}$ for each variable $x_{i_{j}}$ with a small modification. If the variable $x_{i_{j}}$ occurs two times in $\phi ^{\prime }$ as $\overline{x_{i_{j}}}$, then the vertices $\beta _{x_{i_{j}}}$ and $\gamma _{x_{i_{j}}}$ of $\mathcal{G}\xspace_{x_{i_{j}}}$ will correspond to the negative assignment of $x_{i_{j}} $. Otherwise, if $x_{i_{j}}$ occurs two times in $\phi ^{\prime }$ as a positive literal, then the vertices $\beta _{x_{i_{j}}}$ and $\gamma _{x_{i_{j}}}$ of ${\ensuremath{\mathcal{G}}\xspace}_{x_{i_{j}}}$ will correspond to the positive assignment of $x_{i_{j}}$. Again we will create one vertex $v_{k}$ for every clause $C_{k}$ of $\phi ^{\prime }$. If the vertex $v_{k}$ corresponds to an old clause (i.e. from the initial formula $\phi $) then we connect it to the $\gamma $-vertices of the subgraphs ${\ensuremath{\mathcal{G}}\xspace}_{x_{i_{j}}}$ that correspond to these literals. If $v_{k}$ corresponds to a new clause in $\phi ^{\prime }$ then this clause is of the form $(\overline{x_{i_{j}}}\vee x_{i_{j+1}})$. In this case we connect the corresponding $v$-vertex with the vertex $\beta _{x_{i_{j}}}$, if the variable $x_{i_{j}}$ occurs two times as a negative literal in $\phi ^{\prime }$, or with the vertex $\alpha _{x_{i_{j}}}$, if $x_{i_{j}}$ occurs two times as a positive literal in $\phi ^{\prime }$. Similarly, the $v$-vertex is connected with the vertex $\alpha _{x_{i_{j+1}}}$, if the variable $x_{i_{j+1}}$ occurs two times as a negative literal in $\phi ^{\prime }$, or with $\beta _{x_{i_{j+1}}}$, if $x_{i_{j+1}}$ occurs two times as a positive literal in $\phi ^{\prime }$. The weights of these edges will be the same as before, i.e. each edge between a $v$-vertex and a $\beta $-vertex has weight 1 and between a $v$-vertex and an $\alpha $-vertex or a $\gamma $-vertex has weight 3. In order to prove that the constructed graph is bipartite, it is sufficient to prove that there is no cycle with odd length. Let $A$ be the set of all $\alpha $-vertices and of all $\beta $-vertices and let $\Gamma $ be the set of all $\gamma $-vertices. First note that any cycle in the graph $G$ must contain at least two vertices from $A\cup \Gamma $. Furthermore note that, by the above construction, every path that connects two different vertices of the set $A$, without touching any vertex of the set $\Gamma $, has even length. Similarly, every path that connects two different vertices of the set $\Gamma $, without touching any vertex of the set $A$, has also even length. Thus every cycle in $G$ that does not contain any vertex from $\Gamma $ (resp. from $A$) has even length. Consider now a cycle in $G$ that contains vertices from both sets $A$ and $\Gamma $. Then, if we traverse this cycle in any direction, we will encounter the same number of transition edges from set $A$ to set $\Gamma $ and from the set $\Gamma $ to the set $A$. Therefore the length of the cycle is even, and thus $G$ is bipartite. Thus, using the same argumentation as in Lemmas \[lem:one\] and [lem:at-least]{}, we obtain the following theorem. \[thm:bip-hard\] [<span style="font-variant:small-caps;">GreedyMatching</span>]{}is strongly NP-hard, even on bipartite graphs with maximum degree at most 3 and with at most three different integer weight values. Hardness results for [<span style="font-variant:small-caps;">GreedyVertex</span>]{}and <span style="font-variant:small-caps;">GreedyEdge</span>problems\[hardness-additional-subsec\] ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Having established the hardness results for [<span style="font-variant:small-caps;">GreedyMatching</span>]{}in Sections \[APX-subsec\] and \[hardness-bipartite-subsec\], we now prove that also the decision problems [<span style="font-variant:small-caps;">GreedyVertex</span>]{}and [<span style="font-variant:small-caps;">GreedyEdge</span>]{}are also strongly NP-hard. \[thm:twohard\] The decision problems [<span style="font-variant:small-caps;">GreedyVertex</span>]{}and [<span style="font-variant:small-caps;">GreedyEdge</span>]{}are strongly NP-hard, even on graphs with at most five different edge weights. For the proof we amend the construction of Section \[construction-subsec\] and the proofs of Section \[APX-subsec\]. Instead of reducing from the MAX2SAT(3) problem, we provide a reduction from the decision problem 3SAT(3). In this problem we are given a formula $\phi$, in which every clause has at most 3 literals and every variable appears in at most 3 clauses, and the question is whether there exists a truth assignment that satisfies all clauses of $\phi$. This problem is NP-hard [@Ausiello1999]. Let $\phi $ be a 3SAT(3) formula with $n$ variables and $m$ clauses. We construct from $\phi$ a weighted graph $G$ in the same way as in Section \[construction-subsec\], with the only difference that now every $v$-vertex is connected to at most three (instead of at most two) distinguished vertices from the subgraphs ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$. By following exactly the same proofs of Lemmas \[lem:one\] and \[lem:at-least\], we can prove that this graph $G$ has a greedy matching with weight at least $14n+k$ if and only if there exists a truth assignment that satisfies at least $k$ clauses of the 3SAT(3) formula $\phi$. Now we augment this graph $G$ to a new graph $G^{\prime }$ by adding two new vertices $u$ and $u^{\ast }$. Vertex $u$ is adjacent in $G^{\prime }$ to all the $v$-vertices of $G$ with edges of weight $\frac{1}{2}$, while vertex $u^{\ast }$ is adjacent in $G^{\prime }$ only to vertex $u$ with an edge of weight $\frac{1}{4}$. Note that $G^{\prime }$ has five different edge weights. Let $\mathcal{M}^{\prime }$ be a greedy matching in $G^{\prime }$ and let $\mathcal{M}$ be the restriction of $\mathcal{M}^{\prime }$ on the edges of the graph $G$. Since every edge of $G$ has larger weight than every edge that is adjacent to $u$ or to $u^$ in $G^{\prime }$, it follows that $\mathcal{M}$ is also a greedy matching of $G$. Assume that each of the $m$ $v $-vertices is matched in $\mathcal{M}$ with a vertex from the subgraphs ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$. Then clearly $(u,u^) \in \mathcal{M}^{\prime }$. Conversely, assume that $(u,u^) \in \mathcal{M}^{\prime }$. If there exists at least one vertex $v_j$ that is not matched in $\mathcal{M}$ with any vertex from a subgraph ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$, then the edge $(v_j, u)$ with weight $\frac{1}{2}$ will be available to be matched in $\mathcal{M}^{\prime }$, and thus the edge $(u, u^)$ with weight $\frac{1}{4}$ will not belong to $\mathcal{M}^{\prime }$, a contradiction. Therefore, $(u,u^) \in \mathcal{M}^{\prime }$ if and only if each of the $m$ $v$-vertices is matched in $\mathcal{M}$ a vertex from the subgraphs ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$. Furthermore, it follows by the proofs of Lemmas \[lem:one\] and \[lem:at-least\] that each of the $m$ $v$-vertices is matched in $\mathcal{M}$ with a vertex from the subgraphs ${\ensuremath{\mathcal{G}}\xspace}_{x_i}$ if and only if the weight of the greedy matching $\mathcal{M}$ of $G$ is at least $14n+m$, or equivalently, if and only if there exists a truth assignment that satisfies all $m$ clauses of the formula $\phi$. Summarizing, there exists a greedy matching $\mathcal{M}^{\prime }$ of the graph $G^{\prime }$, in which the given edge $(u,u^)$ (resp. the given vertex $u^$) is matched, if and only if the formula $\phi$ is satisfiable. Thus, since 3SAT(3) is NP-hard, it follows that both decision problems [<span style="font-variant:small-caps;">GreedyVertex</span>]{}and [<span style="font-variant:small-caps;">GreedyEdge</span>]{} are strongly NP-hard, even on graphs with at most five different edge weights. Further natural parameters of <span style="font-variant:small-caps;">GreedyMatching</span> ========================================================================================== In this section we investigate the influence of two further natural parameters to the computational complexity of [<span style="font-variant:small-caps;">GreedyMatching</span>]{}, other than the parameters maximum degree and number of different edge weights that we considered in Section \[sec:hardness\]. As the first parameter we consider in Section \[ratio-subsec\] the minimum ratio $\lambda_{0}$ between two consecutive weight values, and as the second parameter we consider in Section \[sec:param2\] the maximum cardinality $\mu$ of the connected components of $G(w_i)$, over all possible weight values $w_i$. We prove that [<span style="font-variant:small-caps;">GreedyMatching</span>]{}has a sharp threshold* behavior with respect to each of these parameters $\lambda_{0}$ and $\mu$. Minimum ratio of consecutive weights\[ratio-subsec\] ---------------------------------------------------- Here we consider the parameter $\lambda_0 = \min_{i} \lambda_i$, where $\lambda_i = \frac{w_i}{w_{i+1}}>1$ is the ratio between the $i$th pair of consecutive edge weights. First we prove that, if $\lambda_0 \geq 2$, then there exists at least one maximum weight matching of $G$ that is an optimum solution for [<span style="font-variant:small-caps;">GreedyMatching</span>]{}on $G$, obtaining the next theorem. \[thm:rg2\] [<span style="font-variant:small-caps;">GreedyMatching</span>]{}can be computed in polynomial time if $\lambda_0 \geq 2$. Let [$\mathcal{M}$]{}be a maximum weight matching for $G$. Note that [$\mathcal{M}$]{}can be computed in polynomial time [@Duan14]. Assume that [$\mathcal{M}$]{}is not a greedy matching. We will construct from [$\mathcal{M}$]{}a greedy matching of $G$ which has the same weight as [$\mathcal{M}$]{}, as follows. Since [$\mathcal{M}$]{}is not greedy, there must exist at least one edge $e \notin {\ensuremath{\mathcal{M}}\xspace}$ and two incident edges $e^{\prime }, e^{\prime \prime }\in {\ensuremath{\mathcal{M}}\xspace}$, where each of the weights $w(e^{\prime })$ and $w(e^{\prime \prime })$ of the edges $e^{\prime }$ and $e^{\prime \prime }$, respectively, is smaller than the weight $w(e)$ of the edge $e$. Since $w(e^{\prime }),w(e^{\prime \prime })<w(e)$ and $\lambda_0 \geq 2$, it follows that $w(e^{\prime }),w(e^{\prime \prime }) \leq \frac{w(e)}{2}$, and thus $w(e) \geq w(e^{\prime }) + w(e^{\prime \prime })$. On the other hand $w(e) \leq w(e^{\prime }) + w(e^{\prime \prime })$, since [$\mathcal{M}$]{}is a maximum weight matching by assumption. Therefore $w(e) = w(e^{\prime }) + w(e^{\prime \prime })$, and thus we can replace in [$\mathcal{M}$]{}the edges $e^{\prime },e^{\prime \prime }$ with the edge $e$ without changing the weight of [$\mathcal{M}$]{}. We call all such edges $e\notin {\ensuremath{\mathcal{M}}\xspace}$ “problematic”. Among all problematic edges pick one edge $e$ with the maximum weight and replace its incident matched edges $e^{\prime },e^{\prime \prime }$ with the edge $e$ in [$\mathcal{M}$]{}. We repeat this procedure until no problematic edge is left, and thus we obtain a greedy matching $\mathcal{M}^{\prime }$ with equal weight as [$\mathcal{M}$]{}. At each iteration the choice of the maximum weight problematic edge ensures that no new problematic edges are created. We perform at most $\|\mathcal{M} \xspace\|/2 = \|E\|/4$ iterations, and thus $\mathcal{M}^{\prime }$ is computed in polynomial time. Recall that in the proof of the Theorem \[thm:inai\] the weight values 1, 3, and 4 were used, thus the [<span style="font-variant:small-caps;">GreedyMatching</span>]{}is hard for $\lambda_0 \leq 4/3$. In the next theorem we amplify this result by showing that [<span style="font-variant:small-caps;">GreedyMatching</span>]{}is NP-hard for any constant $\lambda_0 <2$. That is, complexity of [<span style="font-variant:small-caps;">GreedyMatching</span>]{}has a threshold behavior at the parameter value $\lambda_0 = 2$. \[thm:parlb\] [<span style="font-variant:small-caps;">GreedyMatching</span>]{}is strongly NP-hard and APX-complete for any constant $\lambda_0 < 2$, even on graphs with maximum degree at most 3 and with at most three different integer weight values. For the proof we amend the weight values in the construction of Section \[construction-subsec\] and the proofs of Section \[APX-subsec\]. More specifically, in the construction of the graph $G$ from the formula $\phi$ in Section \[construction-subsec\], we replace each edge of weight 4 with an edge of weight $2x$, and each edge of weight 3 with an edge of weight $x+1$, where $x>1$ is an arbitrary integer. In particular, the results of Sections \[construction-subsec\] and \[APX-subsec\] are given for the value $x=2$. By the proofs of Lemmas \[lem:one\] and \[lem:at-least\] (adapted for these new weights) it follows that there exists a truth assignment that satisfies at least $k$ clauses of the 2SAT(3) formula $\phi$ if and only if there is a greedy matching with weight at least $(6x+2)n+k$ in the constructed graph $G$. Similarly to Sections \[construction-subsec\] and \[APX-subsec\], this graph $G$ maximum maximum degree 3 and three different integer weight values. Furthermore, $\lambda _{0}=\frac{2x}{x+1}$ can go arbitrarily close to 2 as $x$ increases. The statement of the theorem follows exactly by the proof of Theorem \[thm:inai\], adapted for these new weights of the edges of $G$. Maximum edge cardinality of a connected component in $G(w_{i})$\[sec:param2\] ----------------------------------------------------------------------------- Another parameter that we can consider is the maximum edge cardinality $\mu$ of the connected components of $G(w_{i})$, among all different weights $w_{i}$. Since $\mu =1$ implies that there is a unique greedy matching for $G$ which can be clearly computed in polynomial time, we consider the case $\mu \geq 2$. In the original construction of Section \[construction-subsec\], in every gadget $\mathcal{G}_{x_{i}}$ there is a path with five edges where each edge has weight 4. Thus $\mu =5$ in the graph $G$ of Section \[construction-subsec\]. To prove our hardness result for $\mu =2$ in Theorem [thm-maximum-component-parameter]{}, we modify the gadgets $\mathcal{G}_{x_{i}}$ as illustrated in Figure \[gadget-cardinality-parameter-fig\]. \[fig:modgx\] \(01) at (0,1) [$\beta_{x_i}$]{}; (10) at (1,0) \[label=below:$p_{x_i}$\]; (01) – node\[weight\] [$2$]{} (10); (21) at (2,1) \[label=above:$q_{x_i}$\]; (21) – node\[weight\] [$4$]{} (10); (30) at (3,0) \[label=below:$r_{x_i}$\]; (21) – node\[weight\] [$5$]{} (30); (41) at (4,1) [$\alpha_{x_i}$]{}; (41) – node\[weight\] [$5$]{} (30); (51) at (6,1) [$\gamma_{x_i}$]{}; (41) – node\[weight\] [$4$]{} (51); (70) at (7,0) \[label=below:$y_{x_i}$\]; (51) – node\[weight\] [$5$]{} (70); (81) at (8,1) \[label=above:$z_{x_i}$\]; (81) – node\[weight\] [$5$]{} (70); (90) at (9,0) \[label=below:$s_{x_i}$\]; (81) – node\[weight\] [$4$]{} (90); (101) at (10,1) \[label=above:$t_{x_i}$\]; (101) – node\[weight\] [$2$]{} (90); Notice that in every subgraph $\mathcal{G}_{x_{i}}$ (see Figure [gadget-cardinality-parameter-fig]{}) the connected components of each weight have edge cardinality at most 2. Furthermore, the weight of the edge between a $v$-vertex and a $\beta$-vertex has weight 1, while the edges between a $v$-vertex and an $\alpha$-vertex or a $\gamma$-vertex have weight 3. Thus these edges do not belong to any connected component with edges from $\mathcal{G}_{x_{i}}$. However, each $v$-vertex is connected with at most two distinguished vertices in different gadgets $\mathcal{G}_{x_{i}}$ and $\mathcal{G}_{x_{i^{\prime }}}$. Therefore $\mu =2$ in the graph $G$ of this modified construction. Considering these updated gadgets $\mathcal{G}_{x_{i}} $ and using the same argumentation as in Lemmas \[lem:one\] and [lem:at-least]{}, we obtain that there is a greedy matching with weight at least $18n+k$ in the constructed graph $G$ if and only if there is a truth assignment that satisfies at least $k$ clauses from the original 2SAT(3) formula $\phi $, which implies the next theorem. \[thm-maximum-component-parameter\] [<span style="font-variant:small-caps;">GreedyMatching</span>]{}is strongly NP-hard and APX-complete for ${\mu \geq 2}$, even on graphs with maximum degree at most 3 and with at most five different integer weight values. A randomized approximation algorithm\[sec:approx\] ================================================== In this section we provide a randomized approximation algorithm ([<span style="font-variant:small-caps;">Rgma</span>]{}) for [<span style="font-variant:small-caps;">GreedyMatching</span>]{}with approximation ratio $\frac{2}{3}$ on two special classes of graphs (cf. Section \[two-bush-subsec\]). Furthermore we highlight an unexpected relation between [<span style="font-variant:small-caps;">Rgma</span>]{}and the randomized [<span style="font-variant:small-caps;">Mrg</span>]{}algorithm for greedily approximating the maximum cardinality matching (cf. Section \[bush-cardinality\]), the exact approximation ratio of which is a long-standing open problem [@PS12; @ADFS95; @DF91]. Before we present our randomized algorithm [<span style="font-variant:small-caps;">Rgma</span>]{}, we first introduce the following class of weighted graphs, called bush graphs. \[def:bushg\] An edge-weighted graph $G = (V, E)$ with $\ell$ edge weight values $w_1 > w_2 > \ldots > w_{\ell}$ is a bush graph if, for every $i\in\{1,2,\ldots,\ell\}$, the edges of $G(w_{i})$ form a star, which we call the $i$-th bush of $G$. Input: Bush Graph $G$ with edge weight values $w_1 > w_2 > \ldots > w_{\ell}$.\ Output: A greedy matching [$\mathcal{M}_{RG}$]{}. 1. ${\ensuremath{\mathcal{M}_{RG}}\xspace}\leftarrow \emptyset$ 2. for $i = 1 \ldots \ell$ do 3. if $G_i \neq \emptyset$ 4. Select uniformly at random an edge $e_i \in G_i$ and add $e_i$ to [$\mathcal{M}_{RG}$]{} 5. Remove from $G$ the endpoints of $e_i$ and all edges of $G_{i}$ Bush graphs and the maximum cardinality matching\[bush-cardinality\] -------------------------------------------------------------------- In this section we present the connection of the problem [<span style="font-variant:small-caps;">GreedyMatching</span>]{}on (weighted) bush graphs to the problem of approximating the maximum cardinality matching in unweighted graphs via randomized greedy algorithms, cf. Theorem \[thm:cardappx\]. Notice that we cannot directly apply the [<span style="font-variant:small-caps;">Rgma</span>]{}algorithm on unweighted graphs, since the algorithm has to consider the different bushes in a specific total order which is imposed by the order of the weights. Thus, in order to approximate a maximum cardinality matching in a given unweighted graph $G$ using the [<span style="font-variant:small-caps;">Rgma</span>]{}algorithm, we first appropriately convert $G$ to a (weighted) bush graph $G^$ using the next Bush Decomposition algorithm, and then we apply [<span style="font-variant:small-caps;">Rgma</span>]{}on $G^$. Input: Unweighted graph $G=(V,E)$ and ${\epsilon\xspace}\ll \frac{1}{\|V\|^3}$.\ Output: A (weighted) bush graph $G^$. 1. Set $k \leftarrow 0$ 2. while* $E \neq \emptyset$ do 3. Chose a random vertex $u \in V$ \[gws3\] 4. For every $v' \in S := \{v' \in V: (u, v') \in E\}$ set $w(u, v') = 1 - k\cdot {\epsilon\xspace}$ 5. Remove the edges of $S$ from $E$ 6. $k \leftarrow k+1$ Any unweighted graph $G = (V,E)$ can be considered as a weighted graph with edge weights $w(u, v) = 1$ for every edge $(u, v) \in E$, and thus in this case ${\textsc{Opt}\xspace}(G)$ coincides with the maximum cardinality matching in $G$. In the next lemma we relate ${\textsc{Opt}\xspace}(G^)$ with ${\textsc{Opt}\xspace}(G)$. \[lem:cardopt\] ${\textsc{Opt}\xspace}(G) \geq {\textsc{Opt}\xspace}(G^) \geq {\textsc{Opt}\xspace}(G) - \frac{1}{n}$. Assume that ${\textsc{Opt}\xspace}(G^) \geq {\textsc{Opt}\xspace}(G)$. Then, since the weight of each edge of $G$ is 1 and the weight of each edge of $G^$ is by construction smaller than 1, it follows that ${\textsc{Opt}\xspace}(G^)$ has strictly more edges than ${\textsc{Opt}\xspace}(G)$. This is a contradiction, since ${\textsc{Opt}\xspace}(G)$ is a maximum cardinality matching of $G$. Therefore ${\textsc{Opt}\xspace}(G) \geq {\textsc{Opt}\xspace}(G^)$. To prove that ${\textsc{Opt}\xspace}(G^) \geq {\textsc{Opt}\xspace}(G) - \frac{1}{n}$, we construct from a maximum cardinality matching [$\mathcal{M}$]{}of $G$ a maximum weight greedy matching ${\ensuremath{\mathcal{M}}\xspace}^$ for $G^$ with the same cardinality as [$\mathcal{M}$]{}, i.e. $\|{\ensuremath{\mathcal{M}}\xspace}^\|=\|{\ensuremath{\mathcal{M}}\xspace}\|$, as follows. Starting from ${\ensuremath{\mathcal{M}}\xspace}$, we sequentially visit all centers $x_1, x_2, \ldots$ of the bushes in the weighted graph $G^$, in a decreasing order of their edge weights. Whenever a center $x_{i}$ of a bush in $G^$ is unmatched in ${\ensuremath{\mathcal{M}}\xspace}$, then all its neighbors must be matched. If one of these neighbors of $x_i$ is matched in the current matching with an edge that is lighter than the edges of the bush of $x_i$, then we swap one of these edges with an edge incident to $x_{i}$. That is, the only case where $x_i$ stays unmatched is when all neighbors of $x_i$ are matched with edges of larger weight in the current matching. In this case there exists a maximum cardinality matching for $G$ such that the vertex $x_i$ is unmatched. At the end we obtain a matching ${\ensuremath{\mathcal{M}}\xspace}^$ with the same cardinality as the initial matching ${\ensuremath{\mathcal{M}}\xspace}$, but now ${\ensuremath{\mathcal{M}}\xspace}^$ is a greedy matching for $G^$. Thus, since $\|{\ensuremath{\mathcal{M}}\xspace}\|=\|{\ensuremath{\mathcal{M}}\xspace}^\|$ and the weight of ${\ensuremath{\mathcal{M}}\xspace}^$ is $w({\ensuremath{\mathcal{M}}\xspace}^)\leq OPT(G^)\leq OPT(G)=\|{\ensuremath{\mathcal{M}}\xspace}\|$, it follows that $\|{\ensuremath{\mathcal{M}}\xspace}\|-w({\ensuremath{\mathcal{M}}\xspace}^)$ is less than or equal to the sum of the weight differences that have been introduced by “Bush Decomposition”, i.e. $\|{\ensuremath{\mathcal{M}}\xspace}\|-w({\ensuremath{\mathcal{M}}\xspace}^)\leq\frac{1}{n}$, and thus $OPT(G)-OPT(G^)\leq \frac{1}{n}$. With Lemma \[lem:cardopt\] in hand the next theorem follows: \[thm:cardappx\] Let $\rho$ be the approximation guarantee of [<span style="font-variant:small-caps;">Rgma</span>]{}algorithm on every bush graph. Then, for every ${\epsilon\xspace}<1$, [<span style="font-variant:small-caps;">Rgma</span>]{}computes a $(\rho-{\epsilon\xspace})$-approximation of the maximum cardinality matching for unweighted graphs. We conjecture that a tight bound for $\rho$ is $\frac{2}{3}$; in Section \[two-bush-subsec\] we prove our conjecture true for two subclasses of bush graphs. Note that, although vertex $u$ in Step \[gws3\] of the Bush decomposition is selected at random, we do not use anywhere this fact in the proof of Lemma \[lem:cardopt\]. In particular, both Lemma \[lem:cardopt\] and Theorem \[thm:cardappx\] hold even when the choice of $u$ in Step \[gws3\] is arbitrary. Randomized greedy matching in subclasses of bush graphs[two-bush-subsec]{} -------------------------------------------------------------------------- In this section we prove that our <span style="font-variant:small-caps;">Rgma</span>algorithm achieves an approximation ratio $\rho=\frac{2}{3}$ in two special classes of bush graphs, cf. Theorems \[thm:app2bush\] and \[thm:appbush\]. Before we prove these two theorems we first need to prove the following three lemmas which will be useful for our analysis. \[claim0-lem\] Let $v$ be the center of the bush with the largest weight in the graph $G+v$. If the edge $(u,v)$ belongs to a maximum greedy matching of $G+v$ then ${\textsc{Opt}\xspace}(G+v)=w(u,v)+{\textsc{Opt}\xspace}(G-u)$. Let $\mathcal{M}$ be a maximum greedy matching of $G+v$ that contains the edge $(u,v)$. Then $\mathcal{M}\setminus \{(u,v)\}$ is a greedy matching of the graph $G-u$, and thus its weight is at most ${\textsc{Opt}\xspace}(G-u)$. That is, ${\textsc{Opt}\xspace}(G+v)-w(u,v)\leq {\textsc{Opt}\xspace}(G-u)$, and thus ${\textsc{Opt}\xspace}(G+v)\leq w(u,v)+{\textsc{Opt}\xspace}(G-u)$. Conversely, let now $\mathcal{M}^{\prime }$ be a maximum greedy matching of $G-u$. Since $v$ is the center of the bush with the largest weight in $G+v$, it follows that every edge of $G-u$ has weight less than $w(u,v)$. Therefore $\mathcal{M}^{\prime }\cup \{(u,v)\}$ is a greedy matching of $G+v$, and thus its weight is at most ${\textsc{Opt}\xspace}(G+v)$. That is, ${\textsc{Opt}\xspace}(G-u)+w(u,v)\leq {\textsc{Opt}\xspace}(G+v)$. \[claim1-1-lem\]Let $w_{0}$ be the largest edge weight $G$ and let $u$ be a vertex $G$. Then ${\textsc{Opt}\xspace}(G-u)\geq {\textsc{Opt}\xspace}(G)-w_{0}$. Let $\mathcal{M}$ be a maximum greedy matching of $G$. If $u$ is not matched in $\mathcal{M}$ then ${\textsc{Opt}\xspace}(G-u)={\textsc{Opt}\xspace}(G)$, which satisfies the statement of the lemma. Suppose now that $u$ is matched in $\mathcal{M}$ and let $(u,v)\in \mathcal{M}$. We will modify the matching $\mathcal{M}$ of $G$ to a matching $\mathcal{M}^{\prime }$ of $G-u$ as follows. First remove the edge $(u,v)$ from $\mathcal{M}$ and let $\mathcal{M}_{0}=\mathcal{M}\setminus \{(u,v)\}$. If $\mathcal{M}_{0}$ is a greedy matching of $G-u$ then define $\mathcal{M}^{\prime }=\mathcal{M}_{0}$; note that in this case $w(\mathcal{M}^{\prime })=w(\mathcal{M})-w(u,v)\geq w(\mathcal{M})-w_{0}$. Otherwise, if $\mathcal{M}_{0}$ is not greedy, $v$ must have either (a) a neighbor $v^{\prime }$ such that $v^{\prime }$ is unmatched in $\mathcal{M}_{0}$ or (b) a neighbor $v_{1}$ that is matched in $\mathcal{M}_{0}$ with an edge $(v_{1},v_{2})$, where $w(v_{1},v_{2})<w(v,v_{1})$. If $v$ has both such neighbors $v^{\prime }$ and $v_{1}$, we choose to consider only case (a) if $w(v,v^{\prime })>w(v,v_{1})$, or only case (b) if $w(v,v_{1})>w(v,v^{\prime })$, breaking ties arbitrarily if $w(v,v^{\prime })=w(v,v_{1})$. In case (a) we define $\mathcal{M}^{\prime }=\mathcal{M}_{0}\cup \{(v,v^{\prime })\}$; then $\mathcal{M}^{\prime }$ is greedy and $$\begin{aligned} w(\mathcal{M}^{\prime }) &=&w(\mathcal{M})-w(u,v)+w(v,v^{\prime }) \\ &\geq &w(\mathcal{M})-w(u,v) \\ &\geq &w(\mathcal{M})-w_{0}.\end{aligned}$$ In case (b) we define $\mathcal{M}^{\ast }=\mathcal{M}_{0}\cup \{(v,v_{1})\}\setminus \{(v_{1},v_{2})\}$. In this case $$\begin{aligned} w(\mathcal{M}^{\ast }) &=&w(\mathcal{M})-w(u,v)+\left( w(v,v_{1})-w(v_{1},v_{2})\right) \\ &>&w(\mathcal{M})-w(u,v) \\ &\geq &w(\mathcal{M})-w_{0}.\end{aligned}$$If $\mathcal{M}^{\ast }$ is greedy then we define $\mathcal{M}^{\prime }=\mathcal{M}^{\ast }$. Otherwise, if $\mathcal{M}^{\ast }$ is not greedy, $v_{2}$ must have (similarly to the above) either (a) a neighbor $v_{2}^{\prime }$ such that $v_{2}^{\prime }$ is unmatched in $\mathcal{M}^{\ast } $ or (b) a neighbor $v_{3}$ that is matched in $\mathcal{M}^{\ast }$ with an edge $(v_{3},v_{4})$, where $w(v_{3},v_{4})<w(v_{2},v_{3})$. We continue to update the matching $\mathcal{M}^{\ast }$ as above, until we reach a matching $\mathcal{M}^{\prime }$ of $G-u$ such that $\mathcal{M}^{\prime }$ is greedy and $w(\mathcal{M}^{\prime })\geq w(\mathcal{M})-w_{0}$. This completes the proof of the lemma, since $w(\mathcal{M}^{\prime })\leq {\textsc{Opt}\xspace}(G-u)$ and $w(\mathcal{M})={\textsc{Opt}\xspace}(G)$. In the next theorem we prove that [<span style="font-variant:small-caps;">Rgma</span>]{}achieves an approximation ratio of $\frac{2}{3}$ when applied to a bush graph with only two different edge weights. Using this theorem as the induction basis, we then prove in Theorem \[thm:app2bush\] that [<span style="font-variant:small-caps;">Rgma</span>]{}achieves an approximation ratio of $\frac{2}{3}$ also when applied to a bush graph in which every bush has at most two edges. \[thm:app2bush\] [<span style="font-variant:small-caps;">Rgma</span>]{}is a $\frac{2}{3}$-approximation when applied on bush graphs with only two weights and with an arbitrary number of edges per bush. Let $g_1$ and $g_2$ be the two bushes and let $x_2$ be the center of the bush $g_2$. We have to consider the following three cases: $x_2$ is the center of the bush $g_1$ too, $x_2$ does not belong to the bush $g_1$, $x_2$ is a leaf of the bush $g_1$. If $x_2$ is the center of the bush $g_1$ too, then any greedy matching is consisted by only one edge from $g_1$, thus [<span style="font-variant:small-caps;">Rgma</span>]{}always finds an optimal greedy matching. When $x_2$ does not belong to the bush $g_1$ we can partition the edges of $g_2$ in three sets: - $g_{21}$: edges of $g_2$ that are incident to the center of the bush $g_1$. - $g_{22}$: edges of $g_2$ that are not incident to leaves of $g_1$. - $g_{23}$: edges of $g_2$ that are incident to a leaf of $g_1$. Without loss of generality we may assume that $g_{21} = \emptyset$, since none of these edges can be chosen in any greedy matching. Furthermore, if $g_{22} \neq \emptyset$, then [<span style="font-variant:small-caps;">Rgma</span>]{}will always choose one edge from the $g_2$, thus it will return an optimal greedy matching. Finally, if $\|g_{23}\|>1$, then for every choice of edge from $g_1$, there will be at least one edge in $g_2$ that can be chosen in the greedy matching. Hence [<span style="font-variant:small-caps;">Rgma</span>]{}will construct an optimal greedy matching. Finally suppose that $e' \in g_{23}$ be the unique edge $g_{23}$ has. Then the probability that $e'$ is deleted in the first iteration of [<span style="font-variant:small-caps;">Rgma</span>]{}is $\frac{1}{\|g_1\|}$, thus [<span style="font-variant:small-caps;">Rgma</span>]{}will return a matching with expected weight $w_1 + \frac{\|g_1\| - 1}{\|g_1\|}w_2 \geq w_1 + \frac{1}{2}w_2$. Hence the approximation guarantee of [<span style="font-variant:small-caps;">Rgma</span>]{}in this case is $\frac{w_1 + w_2/2}{w_1 + w_2} > \frac{3}{4}$. Finally, we consider the case where $x_2$ is a leaf of the bush $g_1$. Then we may partition the edges of $g_2$ into the following two sets: - $g_{24}$: edges of $g_2$ that are adjacent only to one leaf of $g_1$. - $g_{25}$: edges of $g_2$ that are adjacent to two leaves of $g_1$. If $g_{24} \neq \emptyset$, then with probability $\frac{\|g_1\|-1}{\|g_1\|}$ the algorithm returns a matching with weight $w_1 + w_2$. Thus, if this is the case the approximation guarantee of the algorithm is at least $\frac{3}{4}$. Hence, without loss of generality we may assume that $g_{24} = \emptyset$. If $\|g_{25}\| > 1$, then the only case the algorithm [<span style="font-variant:small-caps;">Rgma</span>]{}returns a matching with weight $w_1$ is when the edge $(x_1, x_2)$ is chosen; in every other case there exists at least one in $g_2$ that it is not deleted. This happens with probability $\frac{1}{\|g_1\|}$. Notice that by our assumption it must be true that $\|g_1\| \geq \|g_2\| > 1$. Thus, the expected weight of the matching returned by the algorithm is $\frac{\|g_1\|-1}{\|g_1\|}(w_1 + w_2) + w_1/\|g_1\| \geq \frac{1}{2}(w_1 + w_2) + w_1/2$. Hence the approximation guarantee of the [<span style="font-variant:small-caps;">Rgma</span>]{} is better than $\frac{2}{3}$. The remaining case is when there is a unique edge $(x_2, v) \in g_{25}$. Then the [<span style="font-variant:small-caps;">Rgma</span>]{}returns a matching of weight $w_1$ when at least one of the edges $x_2$ and $v$ is chosen in the first iteration. This happens with probability $\frac{2}{\|g_1\|}$. Hence the expected weight of the matching returned by the algorithm is $\frac{\|g_1\|-2}{\|g_1\|}(w_1 + w_2) + 2w_1/\|g_1\| \geq \frac{1}{3}(w_1 + w_2) + w_1$. Hence, in this case the approximation guarantee of the algorithm is at least $\frac{2}{3}$. We are now ready to prove the main theorem of this section. \[thm:appbush\] <span style="font-variant:small-caps;">Rgma</span>is a $\frac{2}{3}$-approximation when applied on bush graphs where each bush has at most two edges. We will prove the claim by induction on the number of the different weight values $w_{1},\ldots ,w_{\lambda }$ the bush graph $G$ has. We will use ${\textsc{Av}\xspace}\xspace(G)$ to denote the expected weight of the greedy matching produced by <span style="font-variant:small-caps;">Rgma</span>on the bush graph $G$. We know from Theorem \[thm:app2bush\] that the claim holds when there are only two weight values in $G$. Assume that for any bush graph $G$ with $i\geq 2$ different weight values such that every bush of $G$ has at most two edges, it holds that ${\textsc{Av}\xspace}\xspace(G)\geq \frac{2}{3}{\textsc{Opt}\xspace}(G)$. We will prove that the claim holds also for bush graphs with $i+1$ different weight values. Let $x_{0}$ be the center of the bush with the largest weight $w_{0}$ and let $\alpha $ and $\beta $ be the leaves of this bush. Without loss of generality we can assume that all incident edges to $x_{0}$ have weight $w_{0}$, since every other incident edge of $x_{0}$ with weight $w_{i}<w_{0}$ would never be selected by the greedy algorithm. Assume that the edge $(x_{0},\alpha )$ belongs to the optimal greedy matching of $G+x_{0}$. Thus, ${\textsc{Opt}\xspace}(G+x_{0})=w_{0}+{\textsc{Opt}\xspace}(G-\alpha )$ by Lemma \[claim0-lem\]. Furthermore ${\textsc{Opt}\xspace}(G-\beta )\geq {\textsc{Opt}\xspace}(G)-w_{0}$ by Lemma \[claim1-1-lem\]. Hence we get $$\begin{aligned} {\textsc{Av}\xspace}\xspace(G+x_{0})& =w_{0}+\frac{1}{2}({\textsc{Av}\xspace}\xspace(G-\alpha )+{\textsc{Av}\xspace}\xspace(G-\beta )) \\ & \geq w_{0}+\frac{1}{2}\cdot \frac{2}{3}({\textsc{Opt}\xspace}(G-\alpha )+{\textsc{Opt}\xspace}(G-\beta )) \\ & =\frac{1}{3}({\textsc{Opt}\xspace}(G+x_{0})+{\textsc{Opt}\xspace}(G-\beta )+2w_{0}).\end{aligned}$$Now note that ${\textsc{Opt}\xspace}(G-\beta )\geq {\textsc{Opt}\xspace}(G)-w_{0}$ and ${\textsc{Opt}\xspace}(G)\geq {\textsc{Opt}\xspace}(G+x_{0})-w_{0}$ by Lemma \[claim1-1-lem\]. Therefore ${\textsc{Opt}\xspace}(G-\beta)+2w_{0}\geq {\textsc{Opt}\xspace}(G)+w_{0}\geq {\textsc{Opt}\xspace}(G+x_{0})$, and thus $${\textsc{Av}\xspace}\xspace(G+x_{0})\geq \frac{1}{3}({\textsc{Opt}\xspace}(G+x_{0})+{\textsc{Opt}\xspace}(G+x_{0}))=\frac{2}{3}{\textsc{Opt}\xspace}(G+x_{0}).$$ Conclusions =========== Several interesting open questions stem from our paper. Probably the most important one is to derive tight approximation guarantees $\rho$ for the maximum weight greedy matching problem, even for bush graphs. We conjecture that $\rho=\frac{2}{3}$; an affirmative answer to our conjecture would imply that the algorithm [<span style="font-variant:small-caps;">Mrg</span>]{}for maximum cardinality matching in unweighted graphs has an approximation ratio of almost $\frac{2}{3}$, thus solving a longstanding open problem [@PS12; @ADFS95; @DF91]. We believe that our approach might provide novel ways of better analysis of the [<span style="font-variant:small-caps;">Mrg</span>]{}algorithm. As we proved, [<span style="font-variant:small-caps;">GreedyMatching</span>]{}is NP-hard even on graphs of maximum degree three with at most three different weight values on their edges. It remains open whether [<span style="font-variant:small-caps;">GreedyMatching</span>]{}can be solved efficiently when there are only two weight values on the edges of the input graph. [10]{} D. Abraham, A. Blum, and T. Sandholm. Clearing algorithms for barter exchange markets: enabling nationwide kidney exchanges. In [Proceedings 8th [ACM]{} Conference on Electronic Commerce ([EC]{})]{}, pages 295–304, 2007. D. Abraham, R. Irving, T. Kavitha, and K. Mehlhorn. Popular matchings. , 37(4):1030–1045, 2007. E. Anshelevich, O. Bhardwaj, and M. Hoefer. Friendship and stable matching. In [Algorithms - [ESA]{} 2013 - 21st Annual European Symposium, Sophia Antipolis, France, September 2-4, 2013. Proceedings]{}, pages 49–60, 2013. E. Anshelevich, S. Das, and Y. Naamad. Anarchy, stability, and utopia: creating better matchings. , 26(1):120–140, 2013. J. Aronson, M. Dyer, A. Frieze, and S. Suen. Randomized greedy matching [II]{}. , 6(1):55–73, 1995. G. Ausiello, M. Protasi, A. Marchetti-Spaccamela, G. Gambosi, P. Crescenzi, and V. Kann. . Springer-Verlag, 1999. D. Avis. A survey of heuristics for the weighted matching problem. , 13(4):475–493, 1983. R. Duan and S. Pettie. Linear-time approximation for maximum weight matching. , 61(1):1–23, 2014. M. Dyer and A. Frieze. Randomized greedy matching. , 2(1):29–45, 1991. M. Dyer, A. Frieze, and B. Pittel. The average performance of the greedy matching algorithm. , 3(2):526–552, 05 1993. J. Edmonds. , chapter Paths, Trees, and Flowers, pages 361–379. Birkh[ä]{}user Boston, Boston, MA, 1987. D. R. Ford and D. R. Fulkerson. . Princeton University Press, 2010. D. Hausmann and B. Korte. K-greedy algorithms for independence systems. , 22(1):219–228, 1978. B. Jovanovic. Job matching and the theory of turnover. , 87(5):972–990, 1979. Z. Miller and D. Pritikin. On randomized greedy matchings. , 10(3):353–383, 1997. C. Papadimitriou and M. Yannakakis. Optimization, approximation, and complexity classes. , 43(3):425–440, 1991. M. Poloczek and M. Szegedy. Randomized greedy algorithms for the maximum matching problem with new analysis. In [Proceedings of the 53rd Annual [IEEE]{} Symposium on Foundations of Computer Science ([FOCS]{})]{}, pages 708–717, 2012. V. Raman, B. Ravikumar, and S. S. Rao. A simplified [NP]{}-complete [MAXSAT]{} problem. , 65(1):1–6, 1998. A. Roth. The evolution of the labor market for medical interns and residents: [A]{} case study in game theory. , 92(6):991–1016, 1984. A. Roth and E. Peranson. . Technical report, NBER Working Papers 6963, 1999. A. Roth, T. Sönmez, and M. Ünver. Kidney exchange. , 119(2):457–488, 2004. P. Sankowski. Maximum weight bipartite matching in matrix multiplication time. , 410(44):4480 – 4488, 2009. [^1]: Department of Computer Science, University of Liverpool, UK. Email: `[email protected]` [^2]: School of Engineering and Computing Sciences, Durham University, UK. Email: `[email protected]` [^3]: Department of Computer Science, University of Liverpool, UK, and Computer Technology Institute (CTI), Greece. Email: `[email protected]`	{ "pile_set_name": "ArXiv" }
--- abstract: \| Let $X$ be a $n\times p$ matrix and $l_1$ the largest eigenvalue of the covariance matrix $X^{}X$. The “null case" where $X_{i,j}\sim {\cal N}(0,1)$ is of particular interest for principal component analysis. For this model, when $n, p{\rightarrow}\infty$ and $n/p {\rightarrow}\gamma \in \mathbb{R}_+^$, it was shown in @imj that $l_1$, properly centered and scaled, converges to the Tracy-Widom law. We show that with the same centering and scaling, the result is true even when $p/n$ or $n/p{\rightarrow}\infty$, therefore extending the previous result to $\gamma \in \overline{\mathbb{R}}_+$. The derivation uses ideas and techniques quite similar to the ones presented in @imj. Following @sosh, we also show that the same is true for the joint distribution of the $k$ largest eigenvalues, where $k$ is a fixed integer. Numerical experiments illustrate the fact that the Tracy-Widom approximation is reasonable even when one of the dimension is small. author: - \| Noureddine El Karoui[^1]\ Department of Statistics,\ Stanford University bibliography: - 'research.bib' title: 'On the largest eigenvalue of Wishart matrices with identity covariance when $n$, $p$ and $p/n {\rightarrow}\infty$' --- Introduction ============ Large scale principal component analysis (PCA) - concerning an $n\times p$ matrix $X$ where $n$ and $p$ are both large - is nowadays a widely used tools in many fields, such as image analysis, signal processing, functional data analysis and quantitative finance. Several examples come to mind, including Eigenfaces, subspace filtering, or @lalouxetal where PCA (as well as some random matrix theory) is used to try to improve on the naive solution to Markovitz’s portfolio optimization problem. Important progress has been made recently in our understanding of the statistical properties of PCA in such settings. Emblematic of this is work of @imj, which explains the properties of the square of the largest singular value of a random matrix $X$ under the “null model" where its entries are iid ${\cal N}(0,1)$. Specifically, if we denote the sample eigenvalues of $X'X$ by $l_1\geq \ldots \geq l_p$, call $$\begin{aligned} n_1&=\max{(n,p)}-1\;, \; \; \; \; p_1=\min{(n,p)} \;,\\ \mu_{np}&=(\sqrt{n_1}+\sqrt{p_1})^2 \; ,\\ \sigma_{np}&=(\sqrt{n_1}+\sqrt{p_1})\left(\frac{1}{\sqrt{n_1}}+\frac{1}{\sqrt{p_1}}\right)^{1/3} \; ,\end{aligned}$$ and $W_1$ the Tracy-Widom distribution (see A0), it was shown in @imj that \[ThJ\] If $n,p {\rightarrow}\infty$ and $n/p{\rightarrow}\gamma \in (0,\infty)$, $$\frac{l_1-\mu_{np}}{\sigma_{np}} \overset{\cal L}{\rightarrow}W_1 \;.$$ Building on @imj and using properties of determinantal point processes, @sosh showed that the same result holds for the $k$ largest eigenvalues, where $k$ is a fixed integer: their joint distribution converges to their Tracy-Widom counterpart. This is a very interesting development because the classical theory (e.g @anderson) was developed under the assumption that $p$ was fixed and $n$ grew to $\infty$, whereas in modern day applications both $p$ and $n$ are large. However, Johnstone’s assumption $n/p{\rightarrow}\gamma$ imposes a limit on the validity of his result which one would like to remove. In an actual data analysis, with given $p$ and $n$, $n={\mathrm{o}}(p)$ and $n\asymp p$ could be equally plausible. Furthermore, a specific $X$ of size $n\times p$ could arise in many triangular arrays settings, where we have $X_j$ of size $n_j\times p_j$, and the limitation $n_j/p_j{\rightarrow}\gamma$ finite might only hold in some triangular situations and not in others. Accordingly in this paper we weaken the assumption that $n/p {\rightarrow}\gamma$ finite and show that \[ThJextended\] If $n,p {\rightarrow}\infty$ and $n/p{\rightarrow}\infty$, $$\frac{l_1-\mu_{np}}{\sigma_{np}} \overset{\cal L}{\rightarrow}W_1 \;.$$ Moreover, with the same centering and scaling, the joint distribution of the $k$ largest eigenvalues converges in law to its Tracy-Widom counterpart. Dually, the same result holds if $n/p{\rightarrow}0$. Let us note that the remark we made about centering and scaling sequences after Theorem \[ThJ\] is still valid in this context. There is clearly a mathematical motivation for dealing with this problem: the result completes the picture about the properties of $l_1$ with large $p$ and $n$ and, in a sense, closes Theorem \[ThJ\]. But is it interesting from a statistical standpoint? The situation $p\gg n$ is indeed a fairly common one in modern statistics. Microarray data are a prototypical example: currently they usually have $p$ of the order of a few thousands and $n$ of the order of a few tens. One encounters $p\gg n$ or $n\gg p$ in many other instances: data collection mechanisms are now effective enough so as to, for example, collect and retain thousands of piece of information for millions of customers (transactional data), or millions of pieces of information for thousands of stocks (tick-by-tick data in Finance). Analyzing these very high dimensional datasets raises new challenges and is at the center of recent statistical work, both applied and theoretical. Microarray analysis in particular is a very active field, and has contributed a flurry of activity in non classical situations (very high dimensional data), raising theoretical questions and sometimes revisiting classical techniques or results. As illustrated for instance in @wrr, PCA or PCA-related methods are used for various tasks in the microarray context, from traditional dimensionality reduction procedures to gene grouping. Having a good understanding of the behavior of the singular values of gaussian “white noise" matrices could provide valuable insights for these applications. Recent work of @bl03 about the properties of naive Bayes and Fisher’s linear discriminant function when $p\gg n$ illustrates the impetus these dimensionality assumptions are also gaining in theoretical studies. Our work is part of the larger effort to investigate the properties of high dimensional data structures. Here it is done in a simple, “null" situation. We now present a few numerical experiments we realized to assess how big (or small) $n$ or $p$ should be for Theorems \[ThJ\] and \[ThJextended\] to be practically useful. Numerical experiments --------------------- @imj showed empirically that in that situation the Tracy-Widom approximation was reasonably satisfying, even for small matrices. Similarly, to try to assess its accuracy in our setup, we ran the following experiments in `Matlab`: we picked $n$ and $p$ and generated $10,000$ $n\times p$ matrices $X$ with entries iid ${\cal N}(0,1)$. Then we used standard routines (`normest` in `Matlab`) to compute their spectral norms and squared them to obtain a dataset of $l_1$-s. Following [@imj2], we adjust centering and scaling to $$\begin{aligned} \tilde{\mu}_{np}&=\sqrt{n-1/2}+\sqrt{p-1/2}\; ,\\ \tilde{\sigma}_{np}&=(\sqrt{n-1/2}+\sqrt{p-1/2})\left(\frac{1}{\sqrt{n-1/2}}+\frac{1}{\sqrt{p-1/2}}\right)^{1/3} \;.\end{aligned}$$ This leads to a very significant improvement in the quality of the Tracy-Widom approximation for our simulations. Simple manipulations (explained in section 2.2) show that we have some freedom in choosing the centering and scaling: if we replace $n$ by $n+a$ and $p$ by $p+b$ (where $a$ and $b$ are fixed real numbers) in the definitions of $\mu_{np}$ and $\sigma_{np}$, Theorem \[ThJ\] and Theorem \[ThJextended\] still hold. The particular choice used here is motivated by a careful theoretical analysis of the entries of $K_N$ mentioned in section 2.2. Table 1 summarizes the “quantile" properties of the empirical distributions we obtained and compare them to the Tracy-Widom reference. We used the same reference points as @imj. TW Quantiles TW 10$\times$1000 10$\times$ 4000 10$\times$ 10000 100$\times$4000 30$\times$5000 -------------- ----- ---------------- ----------------- ------------------ ----------------- ---------------- -3.9 .01 0.009 0.010 0.015 0.012 0.013 -3.18 .05 0.047 0.050 0.060 0.053 0.055 -2.78 .10 0.102 0.107 0.112 0.103 0.105 -1.91 .30 0.303 0.308 0.316 0.304 0.303 -1.27 .50 0.506 0.506 0.522 0.508 0.503 -0.59 .70 0.705 0.704 0.723 0.706 0.702 0.45 0.9 0.904 0.904 0.913 0.901 0.904 0.98 .95 0.953 0.951 0.958 0.951 0.953 2.02 .99 0.992 0.990 0.992 0.991 0.991 : Quality of the Tracy-Widom Approximation for some large matrices: the leftmost columns displays certain quantiles of the Tracy-Widom distribution. The second column gives the corresponding value of its cdf. Other columns give the value of the empirical distribution functions obtained from simulations at these quantiles. $\tilde{\mu}_{np}$ and $\tilde{\sigma}_{np}$ are the centering and scaling sequences. \ TW Quantiles TW 50$\times$5000 50$\times$20000 50$\times$50000 5$\times$200 5$\times$2000 5$\times$20000 -------------- ----- ---------------- ----------------- ----------------- -------------- --------------- ---------------- -3.9 .01 0.010 0.017 0.021 0.008 0.014 0.018 -3.18 .05 0.053 0.067 0.079 0.047 0.057 0.069 -2.78 .10 0.104 0.125 0.139 0.094 0.110 0.120 -1.91 .30 0.309 0.331 0.345 0.293 0.314 0.320 -1.27 .50 0.502 0.522 0.538 0.500 0.506 0.519 -0.59 .70 0.705 0.718 0.727 0.714 0.712 0.710 0.45 .90 0.899 0.905 0.911 0.911 0.906 0.907 0.98 .95 0.949 0.955 0.957 0.959 0.951 0.954 2.02 .99 0.991 0.992 0.992 0.994 0.992 0.992 : Quality of the Tracy-Widom Approximation for some large matrices: the leftmost columns displays certain quantiles of the Tracy-Widom distribution. The second column gives the corresponding value of its cdf. Other columns give the value of the empirical distribution functions obtained from simulations at these quantiles. $\tilde{\mu}_{np}$ and $\tilde{\sigma}_{np}$ are the centering and scaling sequences. We picked the dimensions according to two criteria: $100\times4000$, $30\times5000$, and $50\times5000$ were chosen to investigate “representative" microarray situations. We chose the other to have a range of ratios and estimate how valuable the Tracy-Widom approximation would be in situations that could be considered classical, i.e one small dimension (less than 10) and one large (several hundreds to several thousands). For the sake of completeness, we redid the simulations presented in @imj and present in Table 2 the results obtained with $\tilde{\mu}_{np}$ and $\tilde{\sigma}_{np}$ as centering and scaling. TW Quantiles TW 5$\times$5 10$\times$ 10 100$\times$ 100 5$\times$20 10$\times$ 40 100$\times$400 -------------- ------ ------------ --------------- ----------------- ------------- --------------- ---------------- -3.9 .01 0 0.002 0.008 0.001 0.004 0.008 -3.18 .05 0.003 0.018 0.043 0.019 0.032 0.044 -2.78 0.10 0.022 0.054 0.090 0.056 0.077 0.095 -1.91 .30 0.217 0.257 0.295 0.262 0.279 0.294 -1.27 .50 0.464 0.486 0.497 0.490 0.494 0.489 -.59 .70 0.702 0.703 0.700 0.702 0.707 0.702 0.45 .90 0.903 0.903 0.901 0.905 0.906 0.899 0.98 .95 0.949 0.950 0.950 0.952 0.953 0.949 2.02 .99 0.988 0.990 0.991 0.989 0.990 0.990 : Quality of the Tracy-Widom Approximation (Continued): the columns have the same meaning as in Table 1. The ratio $p/n$ is smaller than in Table 1 and the matrices are not as big, but the Tracy-Widom approximation is already acceptable for the upper quantiles. $\tilde{\mu}_{np}$ and $\tilde{\sigma}_{np}$ are the centering and scaling sequences. We see that the fit is good to very good for the upper quantiles ($.9$ and beyond) across the range of dimensions we investigated. The practical interest of this remark is clear: these are the quantiles one would naturally use in a testing problem. We note that it appears empirically that the problem gets harder when the ratio $r$ of the larger dimension to the smaller one ($p_1$ in our notation) gets bigger: the larger $r$, the larger $p_1$ should be for the approximation to be acceptable. Conclusions and Organization ---------------------------- From a technical standpoint, the method developed in @imj proves to be versatile, and, at least conceptually, relatively easy to adapt to the case where $n/p{\rightarrow}\infty$. Nevertheless, substantial technical work is needed to obtain Theorem \[ThJextended\]. Using the elementary fact (see e.g theorem 7.3.7 in @hj) that the largest eigenvalue of $X^{}X$ is the same as the largest eigenvalue of $XX^{}$, it will be sufficient to give the proof in the case $n/p{\rightarrow}\infty$. From a practical point of view, we show that the Tracy-Widom limit law does not depend of how the sequence $(n,p)$ is embedded. As long as both dimensions go to infinity, the properly re-centered and re-scaled largest eigenvalue converges weakly to this law.\ We can compare this with the “classical" situation where $p$ is held fixed, in which case the limiting joint distribution is known, too (see e.g @anderson, corollary 13.3.2). In this case, the centering is done around $n$ and the scaling is $\sqrt{n}$; elementary computations show that $(l_1-\mu_{np})/\sigma_{np}$ also has a non-degenerate limiting distribution (possibly changing with each $p$). Nevertheless, even with the classical centering, it is hard to evaluate the marginals in this context and the results are therefore difficult to use in practice. Our simulation results show that the Tracy-Widom approximation is reasonably good (for the upper quantiles) even when $p$ or $n$ are small. As remarked by @imj, Proposition 1.2, this implies that when doing PCA, one could develop (conservative) tests based on the Tracy-Widom distribution that could serve as alternatives to the scree plot or the Wachter plot.\ The paper is organized as follows: after presenting (Section 2) the main elements of the proof of Theorem \[ThJ\], we describe (Section 3) the strategy that will lead to the proof of Theorem \[ThJextended\]. We prove the two crucial points needed in Section 4. To make the paper self-contained, we give some background information about different aspects of the problem in the appendices. Several technical issues are also treated there in order to avoid obscuring the proof of the main result. Outline of Johnstone’s proof ============================ Before describing the backbone of the proof presented in @imj, we need to introduce a few notational conventions. In what follows, we will use $N$ instead of $p$ to be consistent with the literature. We also denote by ${\mathbf{AB}}$ (for “asymptotic behavior") the situation where $n,N, \text{ and } n/N {\rightarrow}\infty$. We will frequently index functions that depend on both $N$ and $n$ with only $N$. The reason for this is that it will allow us to keep the notations relatively light, and that we think of $n$ as being a function of $N$. Notations like $\textbf{E}_N$ and $\textbf{P}_N$ will denote expectation and probability under the measure induced by the matrices (of size $n(N)\times N$) we are working with.\ Finally, it is technically simpler to work with a matrix $X$ whose entries are standard complex Gaussians (i.e the real and imaginary parts are independent, and they are both ${\cal N}(0,1/2)$), rather than with entries that are ${\cal N}(0,1)$. When we mention the complex case, we refer to this situation. We now give a quick overview of the important points around which the proof of Theorem \[ThJ\] was articulated. At the core of several random matrix theory results lie the fact that the joint distribution of the eigenvalues of the random matrices of interest is known and can be represented as the Fredholm determinant of a certain operator (or a totally explicit function of it). Building on this, if we introduce a number $b$ that is $1$ in the real case and $2$ in the complex one, it turns out that one has the representation formula $$\label{freddet} {\mathbf{E}_{N}\left(\prod_{i=1}^N (1+f(l_i))\right)}=\left[\det({\mathrm{Id}}+S_Nf)\right]^{b/2} \; ,$$ where $S_N$ is an explicit kernel, depending of course upon the kind of matrices in which one is interested. Here, $f$ treated as an operator means multiplication by this function. It is clear that if $\chi_t=-\mathbf{1}\{x:x\geq t\}$, we have $$\mathbf{P}_N(l_1\leq t)=\left[\det({\mathrm{Id}}+S_N\chi_t)\right]^{b/2}\;.$$ The interested reader can find background information on this in @mehta, chapters 5 and 6, @tw98 or @deift, chapter 5, which in turn (p.109) points to @simonreed, section 17, vol 4, for background on operator determinants. We stress the fact that all these formulas are finite dimensional. From the last display, the strategy to show convergence in law in either Theorem \[ThJ\] or \[ThJextended\] is clear: fix $s_0$, show that under the relevant assumptions, $\mathbf{P}(l_{1,N}\leq s_0){\rightarrow}W_1(s_0)$, and use the fact that $W_1$ is continuous to conclude. Complex case ------------ We just saw that to find the asymptotic behavior of $l_1$ is equivalent to showing the convergence of the determinant of a certain operator. This task can be reduced to showing convergence in trace class norm of this operator (see @simonreed for background on this, e.g, Lemma XIII.17.4 (p.323)). Through work from @widom99, @imj exhibits an integral representation formula for his operator, and the original problem is essentially transformed into showing that certain integrals have a predetermined limit. In somewhat more detail, if we call $\alpha=n-N$, and $L_k^{\alpha}$ the $k$-th Laguerre polynomial associated with $\alpha$ (as in @szego, p.100), let $$\phi_k(x)=\sqrt{\frac{k!}{(k+\alpha)!}}x^{\alpha/2}{\mathrm{e}}^{-x/2}L_k^{\alpha}(x) \; ,$$ $\xi_k(x)=\phi_k(x)/x$, $a_N=\sqrt{Nn}$, and finally $$\left\{ \begin{array}{cll} \phi(x)&=(-1)^N\sqrt{\frac{a_N}{2}} (\sqrt{n}\xi_N(x)-\sqrt{N}\xi_{N-1}(x)) \; ,\\ \psi(x)&=(-1)^N\sqrt{\frac{a_N}{2}}(\sqrt{N}\xi_N(x)-\sqrt{n}\xi_{N-1}(x)) \; . \end{array} \right.$$ We note two things: first, there is a slight abuse of notation since $\phi$ and $\psi$ obviously depend on $n$ and $N$, but as in @imj, we choose to not carry these indices in the interest of readability. Also, $\phi$ and $\psi$ admit more “compact" representations, in terms of a single Laguerre polynomial, with a modified $\alpha$, or another degree. These are easy to derive using @szego, p.102, for instance. Nevertheless we choose to work (except in A7) with the previous representations because of the symmetries they present. The kernel $S_N$ mentioned in (\[freddet\]) has the representation (@imj, equation (3.6)) $$S_N(x,y)=\int_{0}^{\infty} \phi(x+z)\psi(y+z)+ \psi(x+z)\phi(y+z) dz \; .$$ Now let $\bar{S}$ be the Airy operator. Its kernel is $$\bar{S}(x,y)=\frac{{\mathrm{Ai}}(x){\mathrm{Ai}}'(y)-{\mathrm{Ai}}(y){\mathrm{Ai}}'(x)}{x-y}=\int_0^{\infty}{\mathrm{Ai}}(x+u){\mathrm{Ai}}(y+u) du \; ,$$ where Ai denotes the Airy function. It was shown in @tw94 that, viewing $\bar{S}$ as an operator on $L^{2}[s,\infty)$, one had $$\det({\mathrm{Id}}-\bar{S})=W_2(s) \; ,$$ where $W_2$ is the Tracy-Widom law “emerging" in the complex case (see A0). So the complex analog of theorem \[ThJ\] follows from the fact that, after defining $S_{\tau}(x,y)={\sigma_N}S_N(\mu_N+{\sigma_N}x,\mu_N+{\sigma_N}y)$, Johnstone managed to show, for all $s$, that $$\det({\mathrm{Id}}-S_{\tau}) {\rightarrow}\det({\mathrm{Id}}-\bar{S}) \; .$$ To do this, he introduced $\phi_{\tau}(s)=\sigma_N \phi(\mu_N+s\sigma_N)$, and similarly $\psi_{\tau}$. Note that we have $$S_{\tau}(x,y)=\int_{0}^{\infty} \phi_{\tau}(x+z)\psi_{\tau}(y+z)+ \psi_{\tau}(x+z)\phi_{\tau}(y+z) dz \; .$$ Since what we are interested in is really $S_{\tau} \chi_s$, for some fixed $s$, we will view $S_{\tau}$ as an operator acting on $L^2[s,\infty)$ in what follows.\ So the problem becomes to show that, as $n,N {\rightarrow}\infty$ $$\label{whattoshow1} \phi_{\tau}(s),\psi_{\tau}(s){\rightarrow}\frac{1}{\sqrt{2}}\mathrm{Ai}(s) \; ,$$ and that $\forall s_0 \in \mathbb{R}$, there exists $N_0(s_0)$ such that if $N>N_0$, we have on $[s_0,\infty)$, $$\label{whattoshow2} \phi_{\tau}(s),\psi_{\tau}(s)=O({\mathrm{e}}^{-s/2}) \; .$$ Once this is shown (we give more details on this later), we can show that $S_{\tau} {\rightarrow}\bar{S}$ in the trace class norm of operators on $L^2[s,\infty)$. A classical way to do it is described in the remark at the end of section 3 of @imj, which bounds the trace class norm of the difference of $S_{\tau}-\bar{S}$ in terms of the Hilbert-Schmidt norm of operators whose kernels are related to $\phi_{\tau}, \psi_{\tau}$ and ${\mathrm{Ai}}$. This leads to the conclusion that $$\det({\mathrm{Id}}-S_{\tau}) {\rightarrow}\det({\mathrm{Id}}-\bar{S}) \;,$$ since $\det$ is continuous with respect to trace class norm. Therefore, the largest eigenvalue of $X^X$ has the behavior it was claimed it has. Real Case --------- In the real case, using arguments from @tw96 and @widom99, @imj gets a representation similar to (\[freddet\]), this time involving an operator with kernel a $2\times 2$ matrix (instead of scalar in the complex case). He is then able to relate it to the complex case problem - the matrix operator determinant can be computed as the product of two scalar operator determinants - and shows that the “reduced" variable he works with ought to have the same limit as it had in the Gaussian Orthogonal Ensemble case, which was studied in depth by Tracy and Widom. For the sake of completeness, we recall that in this situation $\alpha=n-1-N$ and $$\mathbf{P}_N(l_1\leq t)=\sqrt{\det({\mathrm{Id}}+K_N\chi_t)} \;.$$ $K_N$ has the representation (in the $N$ even case) $$K_N= \begin{pmatrix} S_N+ \psi\otimes \epsilon \phi & S_N D-\psi\otimes\phi \\ {\epsilon}S_N -{\epsilon}+{\epsilon}\psi \otimes {\epsilon}\phi & S_N + {\epsilon}\phi \otimes \psi \end{pmatrix} \; ,$$ where $D$ is the differential operator, ${\epsilon}$ is convolution with the kernel ${\epsilon}(x-y)$, and ${\epsilon}(x)=\text{sgn}(x)/2$. We note the slight change in $\alpha$ and replace $n$ by $n-1$ when we need to use the results or formulas derived in the complex case (for instance, the $S_N$ we just mentioned is $S_{n-1,N}$, and not $S_{n,N}$). We refer the reader to @ggk for a complement of information on operator determinants and to the end of section VIII in @tw96 for details on the technical problems that $K_N$ poses. From a purely technical standpoint, one critical issue is to evaluate the large $n,N$ limit of $c_{\phi}=\int_{0}^{\infty}\phi(x)dx/2$. If one can show that it is $1/\sqrt{2}$ when $N{\rightarrow}\infty$ through even values, then Johnstone’s considerations hold true all the way and we have the same conclusion as in Theorem \[ThJ\].\ We note that using the interlacing properties of the singular values (as mentioned for instance in @sosh, Remark 5; see also @hj, theorem 7.3.9), as well as the estimates of the difference (resp. ratio) between two consecutive terms of the centering (resp. scaling) sequence, the $N$ odd case follows immediately from the $N$ even case. To be more precise, we use the fact that $$\frac{\mu_{n,N}-\mu_{n,N-1}}{{\sigma_N}}={\mathrm{O}}(N^{-1/3}) \; \; {\rightarrow}0 \text{ as } N {\rightarrow}\infty$$ to check that the $N$ even terms lower and upper bounding the $N$ odd probability have the same limit. Note that the same relationship holds for $\mu_{n+a,N+b}$ and $\mu_{n,N}$, if $a$ and $b$ are fixed real numbers. Therefore, after doing the proof with centering sequence $\mu_{n+3/2,N+1/2}$ (which is technically simpler), we will be able to conclude that the theorem holds true for $\mu_{n,N}$. Last, to be able to use @sosh, Lemma 2, which gives the result we wish for the joint distribution of the $k$-largest eigenvalues, we will need to verify that the entries of the $2\times 2$ operator converge pointwise, and are bounded above in an exponential way. This is what is done in the proof of Lemma 1 of @sosh, and we will show in A8* that the arguments given there can be extended to handle our situation. Further Remarks and Agenda ========================== Most of the work in @imj is done in closed form, and in the finite dimensional case. That has two advantages from our standpoint: as the limiting behavior is only investigated in the last “step", most of the arguments given there carry through for our problem, and the method certainly does. Therefore, our contribution is mostly technical; it follows very closely the ideas of @imj, providing solutions to technical problems appearing in the case we consider. Only at a few points could we not use the approach developed in @imj. This led us to an analysis of the complex case that is slightly different from the original one, but the core reasons for which the result holds are the same. In what follows, we first focus on showing that (\[whattoshow1\]) and (\[whattoshow2\]) hold true when $n,N$ and their ratio tend to infinity. This takes care of the complex case. We then turn to the problem of the asymptotic behavior of $c_{\phi}$, and the technical points we have to verify for @sosh results to hold. The following remarks outline the differences between the analysis we present here and the one done in @imj. Remarks on adaptation of the original proof ------------------------------------------- ### Complex case {#errorcontrolproblem} To show that (\[whattoshow1\]) and (\[whattoshow2\]) held true, @imj essentially reduced his problem to studying the solution of a “perturbed" Airy equation and used tools from @olver to carefully study it. One point that was used repeatedly was that the turning points of the equation were bounded away from one another when $n,N$ were large. This is not true anymore in the case we consider, and we show how to get around this difficulty. So we do not work with a perturbed Airy equation anymore, but rather with Whittaker functions, which have a close relationship to Laguerre polynomials, and their expansion in terms of parabolic cylinder functions (see A9 for some background information on special functions). In @olver80, the case we are interested in was studied in detail, giving us most of the tools we need to show (\[whattoshow2\]). Using @olver75, we reinterpret the [parabolic cylinder functions]{} results in terms of Airy functions and derive the elements we need to complete the proof of (\[whattoshow1\]) and (\[whattoshow2\]). The reason for which we could not exactly follow the “original" method is related to the error control function called ${\cal V}(\zeta)$ in @imj. This function depends upon the parameter $\omega=2\lambda/\kappa$, which in the case $n/N{\rightarrow}\gamma \in \mathbb{R}$ is bounded away from 2. This essentially allows a uniform control over ${\cal V}$, and it is possible to show that this error control function is bounded as a function of $N$. Since the control is actually something like $\exp(\lambda_0 {\cal V}/\kappa)-1$, it tends to zero as $N{\rightarrow}\infty$. This gave @imj a way to get part of (\[whattoshow2\]).\ In our case, it seems that ${\cal V}$ would tend to $\infty$, at a rate that is nevertheless ${\mathrm{o}}(\kappa)$. As it seems easier and more promising to use @olver80 than to derive the growth of ${\cal V}$, we choose this approach. Nevertheless, this is the only (but crucial) technicality (in the complex case) that did not carry through by the method described in @imj under ${\mathbf{AB}}$. ### Real Case For the $c_{\phi}$ problem, we provide a closed form expression at given $n,N$ and show that in the limit is the “right" one as long as $n$ and $N$ tend to $\infty$.. This does not use the saddlepoint method, but relies on the availability of a generating function formula for Laguerre polynomials. The proof is done in A7.\ A simple modification to @imj would give the same result: in the display preceding (6.13) there, we could write $$h(t)=\sum_{k=0}^{\infty}c_k t^k = 2^{\alpha/2}\Gamma(\alpha/2)(1+t)(1-t^2)^{-(\alpha/2+1)}$$ and expand $(1-t^2)^{-(\alpha/2+1)}$. Multiplying by $1+t$ has a very simple effect on the series, and so $c_k$ is known explicitly. In A8, we show how to check that the conditions required for Soshnikov’s results to hold are indeed met. They are straightforward consequences of the analysis we will carry below. Since the real case is derived from the complex one after analyzing a few technical points, we verify these in the appendices and present here the study of the complex case. We now turn to the main problem we solve in this note: showing (\[whattoshow1\]) and (\[whattoshow2\]) under our set of assumptions. Complex case: study of asymptotics ================================== In this section, we work on the problem of showing pointwise convergence and uniform boundedness, setting the problem in a way similar to section 5 of @imj. We recall his notations, slightly modified to avoid confusions: $N_+=N+1/2$, $n_+=n+1/2$, $z=\mu_N+\sigma_N s$, with $\mu_N=(\sqrt{(N+\alpha)_+}+\sqrt{N_+})^{1/2}$ and $\sigma_N=(\sqrt{(N+\alpha)_+}+\sqrt{N_+})(1/\sqrt{N_+}+1/\sqrt{(N+\alpha)_+})^{1/3}$. For reasons that will be transparent later on, our aim is to show that $$\label{pointwisecv} F_N(z)=(-1)^N \sigma_N^{-1/2}\sqrt{N!/n!}\,z^{(\alpha+1)/2}{\mathrm{e}}^{-z/2}L_N^{\alpha_N}(z) {\rightarrow}{\mathrm{Ai}}(s), \; \forall s \in \mathbb{R} \; ,$$ and $$\label{uniformcv} F_N(z)={\mathrm{O}}({\mathrm{e}}^{-s})\; \text{uniformly in } \; [s_0,\infty), s_0 \in \mathbb{R} \; .$$ The scaling is slightly different from the original proof: $N^{-1/6}$ has been replaced by ${\sigma_N}^{-1/2}$. As in @imj, we focus on $w_N(z)=z^{(\alpha+1)/2}{\mathrm{e}}^{-z/2}L_N^{\alpha}(z)$, which satisfies $$\frac{d^2w}{dz^2}=\left(\frac{1}{4}-\frac{\kappa}{z}+\frac{\lambda^2-1/4}{z^2}\right)w \; ,$$ where $\kappa=N+(\alpha+1)/2$ and $\lambda=\alpha/2$. Remark that under ${\mathbf{AB}}{\overset{\text{def}}{\Leftrightarrow}}n,N,n/N {\rightarrow}\infty$, $\kappa \sim \lambda$. Our strategy is to reformulate the problem in terms of so-called Whittaker functions, denoted $W_{k,m}$, and to use the extensive available studies of these functions to show (\[whattoshow1\]) and (\[whattoshow2\]). @temme, formula (3.1) p.117 shows that $$w_N(z)=\frac{(-1)^N}{N!}W_{\kappa,\lambda}(z) \; .$$ From now on, we will closely follow @olver80. Let us remark that $$F_N(z)={\sigma_N}^{-1/2}\frac{1}{\sqrt{n!N!}}{W_{\kappa,\lambda}}(z) \;.$$ We fix $s_0 \in \mathbb{R}$, and we work only with $z=\mu_N+{\sigma_N}s$, where $s\geq s_0$. #### Preliminaries Following @olver80, we introduce $l=\kappa/\lambda$, $\beta=\sqrt{2(l-1)}$, and the turning points $x_1=2l-2\sqrt{l^2-1}$, $x_2=2l+2\sqrt{l^2-1}$, after the rescaling $x=z/\lambda$. We remark that the two turning points coalesce at 2 under the hypothesis ${\mathbf{AB}}$. In the new variable $x$, we have $$\frac{d^2W}{dx^2}=\left(\lambda^2 g(x) - \frac{1}{4x^2}\right)W \;,$$ where $g(x)=\frac{(x-x_1)(x-x_2)}{4x^2}$. Using the ideas explained in @imj, we shall be - eventually - interested in the asymptotics for $z=\mu_N+\sigma_N s$, or $x=z/\lambda=x_2+\sigma_N s/\lambda$ of $F_N(z)$. Let us now define an auxiliary variable $\upsilon$ by $$\begin{aligned} \int_{\beta}^{\upsilon}(\tau^2-\beta^2)^{1/2}d\tau &= \int_{x_2}^x g^{1/2}(t)dt \hspace{1cm} \text{if } x_2 \leq x < \infty \;, \\ \int_{-\beta}^{\upsilon}(\beta^2-\tau^2)^{1/2}d\tau &= \int_{x_1}^x (-g)^{1/2}(t)dt \hspace{1cm} \text{if } x_1 \leq x \leq x_2 \;.\end{aligned}$$ We limit $x$ to this range because of the technically important following point: ${\sigma_N}/\lambda$ tends to zero faster than $x_2-x_1$ does, and so, when $s$ is bounded below, $x$ will stay in the range $(x_1,\infty)$ for all $N$ greater than a certain $N_0$. This is shown in A2, along with the closely related fact that we can focus on $\upsilon\geq 0$. Our analysis is based on section 3 of @olver80, where he builds on @olver75, in which he expands Whittaker functions in terms of parabolic cylinder functions. The condition ${\upsilon}\geq 0$ is critical, since Olver’s expansions depend on the sign of $\upsilon$. Therefore, A2 entitles us to focus on only one specific form of these. From (3.10) p.219 in @olver80, one has $$\begin{aligned} {W_{\kappa,\lambda}}(\lambda x)= (2\lambda)^{1/4}&\{\lambda(2+\beta^2/2)/e\}^{\lambda(1+\beta^2/4)} &\times\left(\frac{{\upsilon}^2-\beta^2}{x^2-4lx+4}\right)^{1/4}x^{1/2} \{U(-\frac{1}{2}\lambda\beta^2,{\upsilon}\sqrt{2\lambda})+{\epsilon}_1(\lambda^2,\beta^2,{\upsilon})\} \; ,\end{aligned}$$ where, if ${\mathbf{E}}$ and ${\mathbf{M}}$ are the weight and modulus functions associated with $U$ in @olver75 (p.156), we have, according to @olver80 (3.11) p.219, $$\label{error} {\epsilon}_1(\lambda^2,\beta^2,{\upsilon})={\mathbf{E}}^{-1}(-\frac{1}{2}\lambda\beta^2,{\upsilon}\sqrt{2\lambda}) {\mathbf{M}}(-\frac{1}{2}\lambda\beta^2,{\upsilon}\sqrt{2\lambda})\;\mathrm{O}(\lambda^{-2/3})$$ uniformly with respect to $\beta\in [0,B]$ and ${\upsilon}\in [0,\infty)$, $B$ being an arbitrary positive constant. We recall that the main relationship between $U$, ${\mathbf{E}}$ and ${\mathbf{M}}$ : for $b \leq 0$ and $x\geq 0$, $\|U(b,x)\|\leq {\mathbf{E}}^{-1}(b,x){\mathbf{M}}(b,x)$. We now show that we have uniform boundedness on $[s_0,\infty)$. The pointwise convergence result will be a straightforward consequence of the arguments we need to develop to solve this first problem. Uniform Boundedness ------------------- Following up on the previous displays, if $n,N$ are large enough so that ${\upsilon}\geq 0$, we have $$\begin{gathered} \label{Wbound} \left\|{W_{\kappa,\lambda}}(\lambda x)\right\|\leq (2\lambda)^{1/4}\{\lambda(2+\beta^2/2)/e\}^{\lambda(1+\beta^2/4)} \times\left(\frac{{\upsilon}^2-\beta^2}{x^2-4lx+4}\right)^{1/4}x^{1/2}{\mathbf{M}}{\mathbf{E}^{-1}}(1+\mathrm{O}(\lambda^{-2/3})) \;,\end{gathered}$$ where we omitted the argument $(-\frac{1}{2}\lambda\beta^2,{\upsilon}\sqrt{2\lambda})$ for readability purposes. Our plan is now to transform this upper bound into a somewhat similar one, involving the modulus and weight function associated with the Airy function, which have the advantage of having only one parameter and known asymptotics. To carry out this program, we need to split the investigation into two parts: first $s\geq 0$ or ${\upsilon}\geq \beta$. This will allow us to find an $s_1 \geq 0$ such that $F_N(z)={\mathrm{O}}({\mathrm{e}}^{-s})$ on $[2s_1,\infty)$. In the second part, we will just have to consider the case $s\in [s_0,2s_1]$, and show that $F_N$ is merely uniformly bounded on this interval. ### Case $\mathbf{ s\geq 0} $ In order to use the results linking parabolic cylinder functions and the Airy function (proved in @olver59 and cited in @olver75), let us define yet another auxiliary variable, $\eta$, by $$\frac{2}{3}\eta^{3/2}\beta^2=\int_{x_2}^xg^{1/2}(t)dt \;.$$ Then, if we call ${{\cal E}}$ and ${{\cal M}}$ the weight and modulus functions associated with the Airy function, we have, as shown in A3: $$\begin{aligned} {\mathbf{E}^{-1}}(-\frac{1}{2}\lambda\beta^2,{\upsilon}\sqrt{2\lambda}) &\leq {{\cal E}^{-1}}(\lambda^{2/3}\beta^{4/3}\eta)(1+{\mathrm{O}}((\lambda\beta^2)^{-1})) \; ,\\ {\mathbf{M}}(-\frac{1}{2}\lambda\beta^2,{\upsilon}\sqrt{2\lambda})&\leq \frac{\sqrt{2}\pi^{1/4}\left(\Gamma((1+\lambda\beta^2)/2)\right)^{1/2} \beta^{1/2}}{(\lambda\beta^2)^{1/12}}\\ &\times\left(\frac{\eta}{{\upsilon}^2-\beta^2}\right)^{1/4}{{\cal M}}(\lambda^{2/3}\beta^{4/3}\eta) \left(1+{\mathrm{O}}((\lambda\beta^2)^{-1})\right) \; .\end{aligned}$$ Whence, if we call $\theta\triangleq \lambda^{2/3}\beta^{4/3}\eta$, $$\|F_N(\lambda x)\|\leq K_{n,N}x^{1/2}(\eta/(x^2-4lx+4))^{1/4}{{\cal E}^{-1}}(\theta){{\cal M}}(\theta) \left(1+{\mathrm{O}}((\lambda\beta^2)^{-1}\vee \lambda^{-2/3})\right) \; .$$ In A4, we show that $K_{n,N}\sim 2^{2/3}(N/n)^{1/4}$ under ${\mathbf{AB}}$. From now on, $\Delta$ will denote a generic constant; its value may change from display to display. As long as $x\geq x_2$, or $s\geq 0$, we have $$\|F_N(\lambda x)\|\leq \Delta (N/n)^{1/4} x^{1/2}(\eta/(x^2-4lx+4))^{1/4}{{\cal E}^{-1}}(\theta){{\cal M}}(\theta)) \; .$$ Now using the fact that (see @olver, chap. 11) $x^{1/4}{{\cal M}}(x)\leq \Delta$, ${{\cal E}^{-1}}(x)\leq \Delta \exp(-2x^{3/2}/3)$ for $x\geq 0$ and $\lambda\beta^2=2N+1$, we get the new inequality $$\|F_N(\lambda x)\|\leq \Delta \left(\frac{N}{n}\right)^{1/4}N^{-1/6}\left(\frac{x^2}{x^2-4lx+4}\right)^{1/4}\exp(-(2\theta^{3/2})/3) \; .$$ In A5.1, we show that there exists $s_1$ such that if $s\geq 2 s_1$, $(2\theta^{3/2})/3 \geq s$. Also, as shown in A6.1, if $s\geq 0$, $g$ is positive and increasing in $x$ (or, equivalently, in $s$). Since the rational function of $x$ appearing in the previous display is just $(4g(x))^{-1/4}$, we can bound it by its value at $x(2s_1)$ on $[2s_1,\infty)$. In A6.2, we show that, at $s$ fixed, under ${\mathbf{AB}}$, we have $4g(x)\sim \beta{\sigma_N}s/\lambda$, and using the equivalents mentioned in A1, we have ${\sigma_N}\beta/\lambda\sim 4 N^{1/3}/n$, from which we conclude that $$\left(\frac{N}{n}\right)^{1/4}N^{-1/6}(4g(2s_1))^{-1/4}\sim N^{1/12}n^{-1/4}(8s_1N^{1/3}/n)^{-1/4}\sim (8s_1)^{-1/4} \; .$$ Therefore, if $N$ is large enough, $$\forall s\in[2s_1,+\infty)\;\;\|F_N(\lambda x)\| \leq \Delta \exp(-s)$$ ### Case $s\in[s_0,2s_1]$ {#casesinterval} Our aim now is just to show that $F_N$ as a function of $s$ is bounded on this interval; from this we shall immediately have that $F_N={\mathrm{O}}(\exp(-s))$ on this interval, and we will have a proof of (\[uniformcv\]).\ This part is comparatively simpler: we use equation (\[Wbound\]), in which we have ${\mathbf{E}^{-1}}\leq 1$, by definition (@olver75, p.156, (5.22)). Now using the display between (6.12) and (6.13) p.159 of the same article, we have for ${\lambda\beta^2}\geq 1$ and ${\upsilon}\geq 0$, $$\frac{{\mathbf{M}}(-\lambda\beta^2/2,{\upsilon}\sqrt{2\lambda})}{\left(\Gamma((1+\lambda\beta^2)/2)\right)^{1/2}}\leq \frac{\Delta\beta^{1/2}}{(\lambda\beta^2)^{1/12}}\left(\frac{\eta}{{\upsilon}^2-\beta^2}\right)^{1/4} \; .$$ Hence, $$\|F_N(\lambda x)\| \leq K_{n,N}\Delta\left(\frac{\eta}{x^2-4lx+4}\right)^{1/4}x^{1/2} \; .$$ However on this interval, $x{\rightarrow}2$, by A5.2 $\eta=(\lambda\beta^2)^{-2/3}s+{\mathrm{o}}((\lambda\beta^2)^{-2/3})$, and by A6.2 $(x^2-4lx+4)=4s{\sigma_N}\beta(1+{\mathrm{o}}(1))/\lambda$. Therefore, $$\frac{\eta}{x^2-4lx+4}\sim 2^{-2/3}2^{-4} n/N$$ on the whole interval, and, because of the asymptotic estimate of $K_{n,N}$ given in A4, $F_N$ is bounded uniformly in $N$ on the interval $[s_0,2s_1]$. We can thus conclude that $$\forall s_0, \; \; \exists N_0(s_0) \; N>N_0(s_0) ,\; \; \; F_N(s)={\mathrm{O}}_{s_0}({\mathrm{e}}^{-s}) \; \; \text{ on } [s_0,\infty) \; . $$ Pointwise convergence --------------------- Having studied in detail the uniform boundedness of $F_N$ makes the pointwise convergence problem easier. First, since we bounded above $F_N$ in terms of ${\mathbf{M}}$ and ${\mathbf{E}^{-1}}$, equation (\[error\]) shows that ${\epsilon}_1={\mathrm{O}}(\lambda^{-2/3}{\mathrm{e}}^{-s})$ on $[s_0,\infty)$. So for fixed $s$, it tends to zero as $N$ gets large. The pointwise limit of $F_N$ will be the pointwise limit of the parabolic cylinder function part of the expansion. We call this part $\wp F_N$, for “principal part". Using the relationship between $U$ and ${\mathrm{Ai}}$ that we mention in A3, we have, with $\theta=(\lambda\beta^2)^{2/3}\eta$, $$\wp F_N(\lambda x)=K_{n,N}x^{1/2}\left(\frac{\eta}{x^2-4lx+4}\right)^{1/4}({\mathrm{Ai}}(\theta)+{{\cal E}^{-1}}(\theta){{\cal M}}(\theta) \, {\mathrm{O}}((\lambda\beta^2)^{-1})) \; .$$ Since $x{\rightarrow}2$, $K_{n,N}\sim 2^{2/3}(N/n)^{1/4}$ and given the estimate we just mentioned for the ratio $\eta/(x^2-4lx+4)$, we have $$K_{n,N}\, x^{1/2}\left(\frac{\eta}{x^2-4lx+4}\right)^{1/4} \sim 1 \; .$$ In other respects, we show in A5.2 that $\theta{\rightarrow}s$ under ${\mathbf{AB}}$. Finally, ${{\cal E}^{-1}}$ and ${{\cal M}}$ are bounded on $\mathbb{R}$, as shown in 11.2 (pp.394-397) of @olver. Hence ${{\cal E}^{-1}}(\theta){{\cal M}}(\theta) \, (\lambda\beta^2)^{-1} {\rightarrow}0 $ under ${\mathbf{AB}}$, and we can conclude that $\wp F_N(\lambda x){\rightarrow}{\mathrm{Ai}}(s)$; combining all the elements gives $$\forall s \in \mathbb{R}, \; \; \; \;F_N(\lambda x){\rightarrow}{\mathrm{Ai}}(s) \; \text{ under }{\mathbf{AB}}\; .$$ Asymptotics for $\mathbf{\phi_{\tau}} \text{ and } \mathbf{\psi_{\tau}}$ ------------------------------------------------------------------------ So far we have shown that $F_N(z)=(-1)^N {\sigma_N}^{-1/2}\sqrt{z}\phi_N(z){\rightarrow}{\mathrm{Ai}}(s)$, and that ${\mathrm{e}}^s F_N$ was bounded when $N>N_0$ and $s \geq s_0$.\ Our aim is to show (\[whattoshow1\]) and (\[whattoshow2\]). Let us write, as in @imj, $$\phi_{\tau}=\phi_{I,N}+\phi_{I\!I,N} \; ,$$ where $$\phi_{I,N}(z)=(-1)^N{\sigma_N}\sqrt{a_N n}\phi_N(z)/(\sqrt{2} z)= F_N(z)d_N (z/\mu_N)^{-3/2}\;.$$ #### Study of $\phi_{I,N}$ In the previous display, we have $d_N=({\sigma_N}/\mu_N)^{3/2}\sqrt{a_N n/2}$. As ${\sigma_N}\sim n^{1/2}N^{-1/6}$ and $\mu_N\sim n$, $({\sigma_N}/\mu_N)^{3/2}\sim n^{-3/4}N^{-1/4}$. Since $a_N=\sqrt{Nn}$, $a_Nn=n^{3/2}N^{1/2}$, and therefore $d_N{\rightarrow}1/\sqrt{2}$. But when $s$ is fixed, $z/\mu_N {\rightarrow}1$, so it follows that $$\text{Under }{\mathbf{AB}},\hspace{.5cm} \phi_{I,N}(\mu_N+{\sigma_N}s){\rightarrow}\frac{{\mathrm{Ai}}(s)}{\sqrt{2}} \;.$$ To bound $\phi_{I,N}$ for $N>N_0$ and $s\geq s_0$, we use, as in @imj, the uniform bound for $F_N$ and $(z/\mu_N)^{-3/2}\leq \exp(-3{\sigma_N}s/(2\mu_N))$, if $s\geq 0$. If $s\leq 0$, we have $(z/\mu_N)^{-3/2}\leq (1+s_0{\sigma_N}/\mu_N)^{-3/2}$, and since this converges to 1 under ${\mathbf{AB}}$, it is bounded if $N$ is large enough. So we have shown that, $$\phi_{I,N}(\mu_N+s{\sigma_N}) \left\{ \begin{array}{clc} {\rightarrow}& 2^{-1/2}{\mathrm{Ai}}(s)\;, \; \; N{\rightarrow}\infty \;, \\ \leq & M {\mathrm{e}}^{-s} \; \; \text{ on } [s_0,\infty) \text{ if } N>N_0(s_0) \; . \end{array} \right.$$ #### Study of $\phi_{I\!I,N}$ We use once again the same approach as in @imj. We have $$\phi_{I\!I,N}=u_N v_{N-1} \phi_{I,N-1} \; ,$$ where $u_N=({\sigma_N}/\sigma_{N-1})\sqrt{a_N/a_{N-1}}$ and $v_N=(N/n)^{1/2}$, and $n_{N-1}$ appearing in $\sigma_{N-1}$ is $n_N-1$ (for $\phi_{N-1}$ is defined in terms of $L_{N-1}^{\alpha_N}$ and we should therefore have the same $\alpha=n-N=(n-1)-(N-1)$). Remark that under ${\mathbf{AB}}$, $v_N {\rightarrow}0 $ and $u_N {\rightarrow}1$.\ Define $s'$ by $\mu_N+{\sigma_N}s=\mu_{N-1}+\sigma_{N-1}s'$. From $$s'=\frac{\mu_N-\mu_{N-1}}{\sigma_{N-1}}+\frac{{\sigma_N}}{\sigma_{N-1}}s \; ,$$ we deduce that $s'\geq s/2 $ on $[0,\infty)$, if $N$ is large enough: as a matter of fact, under ${\mathbf{AB}}$, $\mu_N -\mu_{N-1 }={\mathrm{O}}(\sqrt{n/N})$, $\sigma_{N} \sim n^{1/2}N^{-1/6}$, and ${\sigma_N}/\sigma_{N-1}{\rightarrow}1$, so it is larger than $1/2$ when $N$ is large enough. To summarize, we just showed that $$\phi_{I\!I,N}(\mu_N+s{\sigma_N})\leq M v_N {\mathrm{e}}^{-s/2} \; \; \text{for } s\in [0,\infty) \; ,$$ by applying the bound we got for $\phi_{I,N}$ to $\phi_{I,N-1}$ and $s'$ as the dummy variable. Here, we are implicitly using the fact that since $n/N{\rightarrow}\infty$, $(n-1)/(N-1)$ does too, and we can apply all the results we derived before. On the other hand, when $s \in[s_0,0]$, we can use the fact that $(\mu_N-\mu_{N-1})\geq 0$ and ${\sigma_N}/\sigma_{N-1} \leq 2$ to show that $s'\geq 2 s $ and hence $$\phi_{I\!I,N}(\mu_N+s{\sigma_N})\leq M v_N {\mathrm{e}}^{-2s}\leq M' v_N {\mathrm{e}}^{-s/2} \; \; \text{for } s\in [s_0,0] \; .$$ The conclusion is therefore that $$\phi_{I\!I,N}(\mu_N+s{\sigma_N}) \left\{\begin{array}{clc} {\rightarrow}& 0 \;, \; \; N{\rightarrow}\infty ,\\ \leq & \Delta {\mathrm{e}}^{-s/2} \; \; \text{ on } [s_0,\infty)\;, \text{ if } N>N_0(s_0) \;. \end{array} \right.$$ Hence we have shown that (\[whattoshow1\]) and (\[whattoshow2\]) held for $\phi_{\tau}$. The analysis for $\psi_{\tau}$ is similar. Appendices ========== This section is devoted to giving background information needed to understand the problem and make the paper relatively self-contained. We also establish many of the properties needed in the course of the proofs of equations (\[whattoshow1\]) and (\[whattoshow2\]) here.\ Before we start, let us mention a notation issue: $\alpha$ changes value depending on whether we treat the complex case or the real one. For the complex case $\alpha+N=n$, whereas for the real one $\alpha+N=n-1$. We frequently replace $\alpha+N$ by $n$ in what follows; this is because the proof of equations (\[whattoshow1\]) and (\[whattoshow2\]) is done in the complex case and applies to the real one by just changing $n$ into $n-1$ everywhere. When dealing with problems which are real case specific, we keep the notation $N+\alpha$. The definition of ${\mu_N}$ and ${\sigma_N}$ are also given in terms of $N+\alpha$ to highlight the adjustments needed when dealing with the real or the complex case. A0: Tracy-Widom distributions {#a0-tracy-widom-distributions .unnumbered} ----------------------------- We recall here the definition of the Tracy-Widom distributions. We split the description according to whether the entries of the matrix we are considering are real or complex.\ We first need to introduce the function $q$, defined as $$\left\{ \begin{array}{l} q''(x)=xq(x)+2q^3(x) \;,\\ q(x)\sim {\mathrm{Ai}}(x) \; \; \; \text{ as } x{\rightarrow}\infty \;. \end{array} \right.$$ $\bullet$ Complex Case The Tracy-Widom distribution appearing in the complex case, $W_2$, has cumulative distribution function $F_2$ given by $$F_2(s)=\exp\left(-\int_s^{\infty}(x-s)q^2(x) dx\right)\;.$$ The joint distribution is slightly more involved to define. Following @sosh, we do it through its $k$-point correlation functions, using its determinantal point process character (see e.g @soshDet).\ Let us first call $\bar{S}$ be the Airy operator. Its kernel is $$\bar{S}(x,y)=\frac{{\mathrm{Ai}}(x){\mathrm{Ai}}'(y)-{\mathrm{Ai}}(y){\mathrm{Ai}}'(x)}{x-y}=\int_0^{\infty}{\mathrm{Ai}}(x+u){\mathrm{Ai}}(y+u) du \;.$$ In the complex case, the $k$-point correlation functions have the property that $$\rho_k(x_1,\ldots,x_k)=\det_{1\leq i,j\leq k} \bar{S}(x_i,x_j) \;.$$ $\bullet$ Real Case The real counterpart of $W_2$, which is called $W_1$, has cdf $F_1$ with $$F_1(s)=\exp\left(-\frac{1}{2}\int_s^{\infty}q(x)+(x-s)q^2(x) dx\right) \;.$$ The $k$-point correlation functions satisfy $$\rho_k(x_1,\ldots,x_k)=\left(\det_{1\leq i,j\leq k} K(x_i,x_j)\right)^{1/2} \;,$$ where the $2\times 2$ matrix kernel of $K$ has entries (see @sosh, eq (2.18) to (2.21)) $$\begin{aligned} K_{1,1}(x,y)&=\bar{S}(x,y)+\frac{1}{2}{\mathrm{Ai}}(x)\int_{-\infty}^y {\mathrm{Ai}}(u) du \; ,\\ K_{2,2}(x,y)&=K_{1,1}(y,x)\; ,\\ K_{1,2}(x,y)&=-\frac{1}{2}{\mathrm{Ai}}(x){\mathrm{Ai}}(y)-\frac{\partial}{\partial y}\bar{S}(x,y)\;,\\ K_{2,1}(x,y)&=-\int_0^{\infty}dt\left(\int_{x+t}^{\infty}{\mathrm{Ai}}(v) dv \right){\mathrm{Ai}}(y+t) -{\epsilon}(x-y)+\frac{1}{2}\int_y^x{\mathrm{Ai}}(u)du+\frac{1}{2}\int_x^{\infty}{\mathrm{Ai}}(u)du\int_{-\infty}^y{\mathrm{Ai}}(v)dv \; .\end{aligned}$$ A1: Asymptotic behavior of some simple functions {#a1-asymptotic-behavior-of-some-simple-functions .unnumbered} ------------------------------------------------ In this appendix, we present some basic facts and identities that we used throughout the proof.\ We will make repeated use of the following observations: since ${\sigma_N}=(\sqrt{(N+\alpha)_+}+\sqrt{N_+})(1/\sqrt{N_+}+1/\sqrt{(N+\alpha)_+})^{1/3} $ and $ \lambda =\alpha/2$, under AB we have $$\begin{aligned} {\sigma_N}&\sim n^{1/2}N^{-1/6} \;,\\ \lambda&\sim n/2 \; .\end{aligned}$$ We also use several times the following identities: \[factequal\] With $\lambda=\alpha/2$, $\kappa=N+(\alpha+1)/2$, and $l=\kappa/\lambda$, $\beta=\sqrt{2(l-1)}$, we have $$\begin{aligned} \lambda\beta^2&=(2N+1) \;,\\ \beta &\sim 2\sqrt{N/n} \;.\end{aligned}$$ The first remark is simple algebra, and the second one comes from $\beta^2=2(l-1)=2(2N+1)/\alpha\sim 4N/n$ under ${\mathbf{AB}}$. We have the estimates: \[factequiv\] $x_2-x_1 \sim 8 \sqrt{N/n}$ and $\sigma_N/\lambda \sim 2 n^{-1/2}N^{-1/6}$ . The second one is obvious; the first one comes from the fact that $x_2-x_1 = 2\sqrt{2}\beta (l+1)^{1/2}$ as $x_{2,1}=2l \pm 2\sqrt{l^2-1}$. Using Fact \[factequal\] immediately gives the claimed result. Finally, we have the following estimates \[factfunctions\] $\beta{\sigma_N}/\lambda \sim 4 N^{1/3}/n$ and ${\sigma_N}^3/(\lambda\beta^2)\sim(n/N)^{3/2}/2$ . The result directly follows from the aforementioned estimates. A2: Working with ${\upsilon}\geq 0$ {#a2-working-with-upsilongeq-0 .unnumbered} ----------------------------------- Here we assume that $s\in[s_0,\infty)$. We also assume that $s<0$, for otherwise we can work with ${\upsilon}\geq \beta>0$. From A1, we have $\|x-x_2\|=\|s\|{\sigma_N}/\lambda \leq \|s_0\|{\sigma_N}/\lambda \ll x_2-x_1 $ by Fact \[factequiv\]. Now ${\upsilon}=0$ corresponds to $x_0\leq \bar{x}=(x_1+x_2)/2$: as a matter of fact, since $(x_2-x)(x_1-x)$ is symmetric around $\bar{x}$ and $1/x$ is obviously larger on $[x_1,\bar{x}]$ than it is on $[\bar{x},x_2]$, we have $$\int_{-\beta}^{{\upsilon}_{\bar{x}}} (\beta^2 - \tau^2) d\tau = \int_{x_1}^{\bar{x}}(-g(t))^{1/2}dt \geq \int_{\bar{x}}^{x_2}(-g(t))^{1/2}dt \; .$$ By symmetry, we also get $$\begin{aligned} \int_{-\beta}^{0} (\beta^2 - \tau^2) d\tau = \int_{0}^{\beta} (\beta^2 - \tau^2) d\tau & = \frac{1}{2} \int_{x_1}^{x_2}(-g(t))^{1/2}dt \\ &\leq \int_{x_1}^{\bar{x}}(-g(t))^{1/2}dt = \int_{-\beta}^{{\upsilon}_{\bar{x}}} (\beta^2 - \tau^2) d\tau \; ,\end{aligned}$$ and therefore, ${\upsilon}_{\bar{x}}>0$.\ However $\bar{x}$ is always smaller than $x(s_0)$ if $N$ is large enough. So we can limit our investigations to the case ${\upsilon}\geq 0$. A3: Relationship between ${\mathbf{E}^{-1}}$, ${{\cal E}^{-1}}$, ${\mathbf{M}}$ and ${{\cal M}}$ {#a3-relationship-between-mathbfe-1-cal-e-1-mathbfm-and-cal-m .unnumbered} ------------------------------------------------------------------------------------------------ We claim that if $s\geq 0$, and we define $\theta=(\lambda\beta^2)^{2/3}\eta$, the following inequalities hold true: $$\begin{aligned} {\mathbf{E}^{-1}}(-{\lambda\beta^2}/2,{\upsilon}\sqrt{2\lambda})&\leq {{\cal E}^{-1}}(\theta)(1+{\mathrm{O}}(({\lambda\beta^2})^{-1})) \; ,\\ {\mathbf{M}}(-{\lambda\beta^2}/2,{\upsilon}\sqrt{2\lambda})&\leq \frac{(4\pi)^{1/4}}{({\lambda\beta^2})^{1/12}}[\Gamma((1+\lambda\beta^2)/2)]^{1/2}\beta^{1/2}\left(\frac{\eta}{{\upsilon}^2-\beta^2}\right)^{1/4}\\ & \times{{\cal M}}(\theta)(1+{\mathrm{O}}(({\lambda\beta^2})^{-1})) \;.\end{aligned}$$ For the sake of simplicity we call $\Xi$ the part that precedes the sign “$\times$" in the last inequality.\ According to @olver75, equations (5.12) and (5.13), we have $$\begin{aligned} U(-{\lambda\beta^2}/2,{\upsilon}\sqrt{2\lambda})&=\Xi\left\{{\mathrm{Ai}}(\theta)+{{\cal M}}(\theta){{\cal E}^{-1}}(\theta)\,{\mathrm{O}}(({\lambda\beta^2})^{-1})\right\} \; ,\\ \bar{U}(-{\lambda\beta^2}/2,{\upsilon}\sqrt{2\lambda})&=\Xi\left\{{\mathrm{Bi}}(\theta)+{{\cal M}}(\theta){{\cal E}^{-1}}(\theta)\,{\mathrm{O}}(({\lambda\beta^2})^{-1})\right\} \;.\end{aligned}$$ We have, if $s\geq 0$, $x\geq x_2$, so $$\label{vardef} 2/3\beta^2\eta^{3/2}=\int_{x_2}^x g^{1/2}(t)dt=\int_{\beta}^{{\upsilon}}(\tau^2-\beta^2)^{1/2}d\tau \;.$$ For the Airy function, the weight and modulus functions had different definition depending on whether the argument was bigger than the largest root, $c$, of ${\mathrm{Ai}}(z)={\mathrm{Bi}}(z)$ or not. Likewise, the definition of ${\mathbf{E}^{-1}}$ and ${\mathbf{M}}$ depends on the position of the argument with respect to the largest root of the equation $\bar{U}(b,x)=U(b,x)$, which is called $\rho(b)$ in @olver75. #### Where do the auxiliary variables lie when $\mathbf{s\geq 0}$? We claim that the answer is that $\theta \geq 0 > c$, and ${\upsilon}\sqrt{2\lambda}\geq \rho(-{\lambda\beta^2}/2)$.\ The first part of equation (\[vardef\]) implies that $\eta\geq 0$, so $\theta \geq c$, as $c<0$. This means that we can use the definition ${{\cal M}}^2=2{\mathrm{Ai}}{\mathrm{Bi}}$ and ${{\cal E}^{-1}}{{\cal M}}=2^{1/2}{\mathrm{Ai}}$. The second part implies that ${\upsilon}\geq \beta$; therefore, $2\lambda{\upsilon}^2\geq 2{\lambda\beta^2}\geq \rho(-{\lambda\beta^2}/2)^2$, since by @olver75, equation (5.21), $\rho(b)\leq 2(-b)^{1/2}$ when $b{\rightarrow}-\infty$. This means that we have similar relationships between ${\mathbf{E}^{-1}}$, ${\mathbf{M}}$, $U$, and $\bar{U}$, to the one we had in the Airy case, $\bar{U}$ playing the role of ${\mathrm{Bi}}$, and $U$ playing the role of ${\mathrm{Ai}}$. #### Consequences of their positions The interesting consequence of the previous remarks is that we can write, if $N$ is large enough, for all $s\geq 0$ $$\mathbf{E}^{-2}(-{\lambda\beta^2}/2,{\upsilon}\sqrt{2\lambda})=\frac{U}{\bar{U}}=\frac{{{\cal M}}(\theta){{\cal E}^{-1}}(\theta)}{{{\cal M}}(\theta){\cal E}(\theta)} \frac{2^{-1/2}+{\mathrm{O}}(({\lambda\beta^2})^{-1})}{2^{-1/2}+{\mathrm{O}}(({\lambda\beta^2})^{-1})} \;.$$ In other words, we just proved that $\exists N_0$ such that $N>N_0$ implies, $\forall \, s\geq 0$ $${\mathbf{E}^{-1}}(-{\lambda\beta^2}/2,{\upsilon}\sqrt{2\lambda}) \leq {{\cal E}^{-1}}(\theta)(1+{\mathrm{O}}(({\lambda\beta^2})^{-1}) \;.$$ By the same arguments, we derive that $${\mathbf{M}}(-{\lambda\beta^2}/2,{\upsilon}\sqrt{2\lambda})\leq \Xi {{\cal M}}(\theta)(1+{\mathrm{O}}(({\lambda\beta^2})^{-1})\;.$$ A4: Asymptotic behavior of $K_{n,N}$ {#a4-asymptotic-behavior-of-k_nn .unnumbered} ------------------------------------ The aim here is to show that $$K_{n,N}\sim 2^{2/3} (N/n)^{1/4} \;.$$ $K_{n,N}$ has the following expression: $$K_{n,N}=\frac{(2\lambda)^{1/4}\{\lambda(2+1/2\beta^2)/e\}^{\lambda(1+\beta^2/4)}\sqrt{2}\pi^{1/4}[\Gamma((1+\lambda\beta^2)/2)]^{1/2}\beta^{1/2}}{(\lambda\beta^2)^{1/12}\sqrt{n!N!{\sigma_N}}} \; .$$ Since $\lambda\beta^2=(2N+1)$, $\Gamma((1+\lambda\beta^2)/2)=\Gamma(N+1)=N!$ .\ In other respects, let $A_n=\{\lambda(2+1/2\beta^2)/e\}^{\lambda(1+\beta^2/4)}/\sqrt{n!}$ . Note that $2\lambda+\lambda\beta^2/2=n-N+(2N+1)/2=n+1/2=n_+$. So $A_n=(n_+/e)^{n_+/2}/\sqrt{n!}$. Using Stirling’s formula, we get that $A_n\sim (n_+/n)^{n/2}(n_+/n)^{1/4}(2\pi {\mathrm{e}})^{-1/4} \sim (2\pi)^{-1/4}$.\ Now rewriting $$K_{n,N}=\frac{A_n(\lambda\beta^2)^{1/4}(8\pi)^{1/4}}{(\lambda\beta^2)^{1/12}\sqrt{{\sigma_N}}} \;,$$ we get that $K_{n,N}\sim 2^{2/3} (N/n)^{1/4}$, from using $A_n(8\pi)^{1/4}\sim \sqrt{2}$ and the second estimates of Fact \[factfunctions\] in A1. A5: Asymptotic properties of $\mathbf{\eta}$ {#a5-asymptotic-properties-of-mathbfeta .unnumbered} -------------------------------------------- This appendix is divided into two parts. We first show that there exists $s_1$ such that, if $s\geq 2s_1$, $$\begin{gathered} \frac{2}{3}{\lambda\beta^2}\eta^{3/2} \geq s \;.\tag{P1}\end{gathered}$$ Then we shall show: $$\begin{gathered} \text{uniformly in }s \in [a,b], \; \; \;(2N+1)^{2/3}\eta=s+{\mathrm{o}}(1) \;.\tag{P2}\end{gathered}$$ ### A5.1: Proof of P1 {#a5.1-proof-of-p1 .unnumbered} This is the argument that was used in A8 of @imj. We repeat it for the sake of completeness.\ Let us first suppose that $s$ is given. Since $g(x)=(x-x_1)(x-x_2)/(4x^2)$, we have $${\sigma_N}^2g(x) = s\frac{\sigma_N^3}{\lambda}\frac{(x_2-x_1)+s{\sigma_N}/\lambda}{4(x_2+s{\sigma_N}/\lambda)^2} \sim s\frac{{\sigma_N}^32\sqrt{2}\beta(l+1)^{1/2}}{16 \lambda}\sim s\frac{\beta{\sigma_N}^3}{4\lambda}\;,$$ the first equivalent coming from the fact that when $s$ is fixed, $x_2-x_1 \gg s{\sigma_N}/\lambda$, and $x_2{\rightarrow}2$. The second is just $l{\rightarrow}1$ under ${\mathbf{AB}}$. Now using the first point of Fact \[factfunctions\] in A1, together with ${\sigma_N}^2\sim n N^{-1/3}$, we get that $(\beta{\sigma_N}^3)/(4\lambda){\rightarrow}1$. So at $s$ fixed, $${\sigma_N}^2 g(x){\rightarrow}s \;.$$ Having this information let us now pick $s_1=8$. If $N$ is large enough, we have ${\sigma_N}^2 g(x(s_1))\geq s_1/2=4$. For all (fixed) $N$ $g$ is an increasing function of $s$. Therefore for the same $N$ we will have $$\forall \, s\geq s_1 \; {\sigma_N}^2 g(x) \geq {\sigma_N}^2 g(x(s_1))\geq s_1/2=4 \; ,$$ and hence, since $s\geq s_1\geq 0$, $g$ is positive and we have $g^{1/2}(s)\geq 2/{\sigma_N}$. Therefore, $$\frac{2}{3}{\lambda\beta^2}\eta^{2/3}=\lambda \int_{x_2}^xg^{1/2}(t)dt \geq \int_{x(s_1)}^xg^{1/2}(t)dt \geq \frac{2\lambda}{{\sigma_N}}\frac{{\sigma_N}}{\lambda}(s-s_1)=2(s-s_1)\;.$$ Consequently, if $s\geq 2s_1$, we have (P1). ### A5.2: Proof of P2 {#a5.2-proof-of-p2 .unnumbered} Without loss of generality, we can suppose that $a$ and $b$ have the same sign, and $a\geq 0$. (If it is not the case, we can split $[a,b]=[a,0]\bigcup[0,b]$, apply the reasoning on each of these, and get the claimed result for the original interval.)\ The idea is that on $[a,b]$, we have $$\frac{(x-x_2)(x_2+b{\sigma_N}/\lambda -x_1)}{4(x_2+a{\sigma_N}/\lambda)^2} \geq g(x)\geq \frac{(x-x_2)(x_2-x_1+a{\sigma_N}/\lambda)}{4(x_2+b{\sigma_N}/\lambda)^2} \; .$$ Now on both sides, the terms which are not $(x-x_2)$ are $(x_2-x_1)(1+{\mathrm{o}}(1))=4\beta(1+{\mathrm{o}}(1))$, again because ${\sigma_N}/\lambda \ll \beta$. So if we integrate the square root of the previous inequality between $x_2$ and $x(s)$, we get $$\begin{aligned} 2/3(s{\sigma_N}/\lambda)^{3/2}2\sqrt{\beta}(1+{\mathrm{o}}(1))/4 \geq 2/3 \eta^{3/2}\beta^2 \geq 2/3(s{\sigma_N}/\lambda)^{3/2}2\sqrt{\beta}(1+{\mathrm{o}}(1))/4 \;,\end{aligned}$$ or $$\frac{1}{2}s^{3/2}({\sigma_N}^3\beta/\lambda)^{1/2}(1+{\mathrm{o}}(1)) \geq \eta^{3/2}\lambda\beta^2 \geq \frac{1}{2}s^{3/2}({\sigma_N}^3\beta/\lambda)^{1/2}(1+{\mathrm{o}}(1)) \;.$$ The conclusion follows from A1, Fact \[factfunctions\], whose first point, along with the estimate of ${\sigma_N}$ mentioned there, shows that ${\sigma_N}^3\beta/\lambda\sim 4$. We note that (P2) also gives us pointwise convergence of $({\lambda\beta^2})^{2/3}\eta$ to $s$. A6: Properties of $\mathbf{g}$ {#a6-properties-of-mathbfg .unnumbered} ------------------------------ We first show that $g$ is increasing - at $N$ fixed - as a function of $s$, if $s\geq 0$. Then we give an estimate of $4x^2g(x)$ as $N{\rightarrow}\infty$ and $s\in[a,b]$. ### A6.1: $\mathbf{g}$ is increasing on $s\geq 0$ {#a6.1-mathbfg-is-increasing-on-sgeq-0 .unnumbered} Since $g(t)=(t-x_2)(t-x_1)/(4t^2)=(t^2-4lt+4)/(4t^2)$, we have $$g'(t)=\frac{l}{t^2}-\frac{2}{t^3}=\frac{lt-2}{t^3} \;.$$ Now $lx_2=2l^2+2l\sqrt{l^2-1}\geq 2$, since $l=1+(2N+1)/\alpha\geq 1$. But $lx\geq lx_2$ when $s\geq 0$, and the assertion is proved. ### A6.2: On the asymptotic behavior of $\mathbf{4x^2g(x)}$ for $\mathbf{s\in [a,b]}$ {#a6.2-on-the-asymptotic-behavior-of-mathbf4x2gx-for-mathbfsin-ab .unnumbered} This estimate is motivated by the fact that in the course of the proof of the main result, we have to deal with an expression of the form $$\frac{\eta}{x^2-4lx+4} \;.$$ We already studied in detail $\eta$ as a function of $s$ and $N$. We now focus on $x^2-4lx+4$. Recalling that $x^2-4lx+4=(x-x_2)(x-x_1)$ and $x=x_2+s{\sigma_N}/\lambda$, we have $$x^2-4lx+4=s\frac{{\sigma_N}}{\lambda}(x_2-x_1+s\frac{{\sigma_N}}{\lambda})=s\frac{{\sigma_N}}{\lambda}(x_2-x_1+{\mathrm{o}}(\beta)) \;,$$ because the first estimate in Fact \[factfunctions\] shows that ${\sigma_N}/\lambda = {\mathrm{o}}(\beta)$, and since $s \in [a,b]$, the previous statement holds true uniformly on this interval. Now $x_2-x_1\sim 4\beta$ under ${\mathbf{AB}}$, and therefore, uniformly on $[a,b]$, $$x^2-4lx+4=s\frac{{\sigma_N}}{\lambda}4\beta(1+{\mathrm{o}}(1)) \; ,$$ as was claimed in \[casesinterval\]. Also, since $x=x_2+s{\sigma_N}/\lambda$, and $x_2=2+(2l+2)^{1/2}\beta+\beta^2$, $$4g(x)=s\frac{{\sigma_N}}{\lambda}\beta(1+{\mathrm{o}}(1)) \;.$$ A7: Limit of $c_{\phi}$ {#a7-limit-of-c_phi .unnumbered} ----------------------- Recall that under the notation of @imj, $$\sqrt{2} c_{\phi}=\frac{1}{2}\sqrt{a_N}\left(\sqrt{N+\alpha}\int \xi_N - \sqrt{N} \int \xi_{N-1}\right)$$ where $\xi_k(x)=x^{\alpha/2-1}e^{(-x/2)}L_k^{(\alpha)}(x)\sqrt{\frac{k!}{(k+\alpha)!}}$. We are interested in\ $$\begin{aligned} v_{k,\alpha}&=\sqrt{k+\alpha}\int \xi_k - \sqrt{k} \int \xi_{k-1} \\ &=\sqrt{\frac{k!}{(k+\alpha-1)!}}\int_0^{\infty}x^{\alpha/2-1}e^{(-x/2)} \left(L^{\alpha}_k(x)-L^{\alpha}_{k-1}(x)\right) dx \\ \text{(by \citet{szego} 5.1.13 p.102)} &= \sqrt{\frac{k!}{(k+\alpha-1)!}}\int_0^{\infty}x^{\alpha/2-1}e^{(-x/2)} L^{\alpha-1}_k(x) dx \\ &=\sqrt{\frac{k!}{(k+\alpha-1)!}} I_{k,\alpha} \;.\end{aligned}$$ Now using @szego 5.1.9 p.101, $$\sum_{k=0}^{\infty}w^k L_k^{\alpha-1}(x)=(1-w)^{-\alpha} \exp\left(-\frac{xw}{1-w}\right) \;.$$ So if $F(\alpha)=\int_0^{\infty} \left(\sum_{k=0}^{\infty}w^k L_k^{\alpha-1}(x)\right)x^{\alpha/2-1}e^{(-x/2)} dx$, we have: $$\begin{aligned} F(\alpha)&=(1-w)^{-\alpha}\int_0^{\infty} x^{\alpha/2-1}e^{(-x/2)} {\mathrm{e}}^{-xw/(1-w)}dx\\ &=(1-w)^{-\alpha}\Gamma(\alpha/2)\left(\frac{2(1-w)}{1+w}\right)^{\alpha/2}\\ &=2^{\alpha/2}(1-w^2)^{-\alpha/2}\Gamma(\alpha/2) \;.\end{aligned}$$ Now if $x\geq 0,\hspace{.3cm} \|L_n^{\alpha-1}(x)\| \leq L_n^{\alpha-1}(-x)$, by 5.1.6 in @szego, and hence $$\left\|\sum_{k=0}^{+\infty}w^k L_k^{\alpha-1}(x)\right\|\leq \sum_{k=0}^{+\infty}\|w\|^k L_k^{\alpha-1}(-x)=(1-\|w\|)^{-\alpha}\exp\left(\frac{x\|w\|}{1-\|w\|}\right) \;.$$ Therefore, as long as $w \in(-1/3,1/3)$, we can switch orders of summation, and get $$\sum_{k=0}^{\infty}w^k I_{k,\alpha}=2^{\alpha/2}(1-w^2)^{-\alpha/2}\Gamma(\alpha/2) \;.$$ But $(1-w)^{-\alpha/2}\Gamma(\alpha/2)=\sum_{k=0}^{+\infty}\frac{\Gamma(\alpha/2+k)}{k!}w^k$, since the right-hand side converges without any difficulty on $(-1/3,1/3)$, and hence $$\sum_{k=0}^{\infty}w^kI_{k,\alpha}=\sum_{m=0}^{\infty}\frac{2^{\alpha/2}\Gamma(\alpha/2+m)}{m!}w^{2m}\;.$$ So we have $$\forall k\in 2\mathbb{N}, \hspace{.5cm} I_{k,\alpha}=\frac{2^{\alpha/2}\Gamma((\alpha+k)/2)}{(k/2)!} \;.$$ Now $v_{k,\alpha}=\sqrt{\frac{k!}{(k+\alpha-1)!}}I_{k,\alpha}=2^{\alpha/2}\frac{\Gamma((\alpha+k)/2)}{\sqrt{(k+\alpha-1)!}}\frac{\sqrt{k!}}{(k/2)!}$. Since $\Gamma(z)\sim (z/e)^z \sqrt{2\pi/z}$, we have $$\begin{aligned} \frac{\Gamma((\alpha+k)/2)}{\sqrt{\Gamma(k+\alpha)}}&\sim 2^{-(\alpha+k)/2}(\alpha+k)^{-1/4}(2\pi)^{1/4} \sqrt{2} \;, \\ \frac{\sqrt{k!}}{(k/2)!}&\sim 2^{k/2}(\pi k)^{-1/4}2^{1/4} \;,\end{aligned}$$ which in turn leads to $$\begin{aligned} v_{k,\alpha}&\sim 2^{\alpha/2}(k(\alpha+k))^{-1/4}2^{-(\alpha+k)/2}2^{k/2} \sqrt{2} \sqrt{2}\\ &\sim 2 (k(\alpha+k))^{-1/4} \\ &\sim 2/\sqrt{a_k} \;.\end{aligned}$$ Hence, as $N$ is even, $\sqrt{2}c_{\phi}=v_{N,\alpha} \sqrt{a_N}/2 {\rightarrow}1 $. A8: On @sosh Lemma 1 {#a8-on-lemma-1 .unnumbered} -------------------- For @sosh Lemma 1 to hold true in our case, we have to check two things. First that not only does ${\sigma_N}\phi({\mu_N}+{\sigma_N}s){\rightarrow}{\mathrm{Ai}}(s)/\sqrt{2}$, but also that this is true for the derivative: $$\label{eq:sosh1} {\sigma_N}^2 \phi'({\mu_N}+{\sigma_N}s) {\rightarrow}\frac{1}{\sqrt{2}} {\mathrm{Ai}}'(s) \; .\tag{S1}$$ We also have to verify that ${\sigma_N}^2 \phi'({\mu_N}+{\sigma_N}s)$ is bounded above by $\Delta(s_0) \exp(-\Delta s)$ on $[s_0,\infty)$, where $\Delta$ is a positive constant. We need to verify this for $\psi$ as well, but the techniques are similar, so we will verify it only for $\phi$. The second point that we need to check is that $$\label{eq:sosh2} \int_0^{\infty}\left(\int_0^{z}\phi(u) du \; \psi(y+z)\right) dz {\rightarrow}0 \; \text { as } N {\rightarrow}\infty \; .\tag{S2}$$ ### A8.1: Proof of (\[eq:sosh1\]) {#a8.1-proof-of-eqsosh1 .unnumbered} It is easy to see that all we need to work on are the properties of $g_N(s)=F_N({\mu_N}+{\sigma_N}s)$; if we can show that ${\sigma_N}F_N'({\mu_N}+{\sigma_N}s) {\rightarrow}{\mathrm{Ai}}'(s)$, and that it is bounded by $\Delta(s_0) {\mathrm{e}}^{-\Delta s}$ on $[s_0,\infty)$, we will be done.\ We have very easily that $$-{\sigma_N}F_N'({\mu_N}+{\sigma_N}s_1)=\int_{s_1}^{\infty} {\sigma_N}^2 \left. \frac{d^2 F_N}{ds^2} \right\|_{{\mu_N}+{\sigma_N}u} du \; .$$ So the strategy is clear: we want to show that the integrand in the right-hand side is bounded by an integrable function and that it converges pointwise to ${\mathrm{Ai}}''(u)=u{\mathrm{Ai}}(u)$.\ However, $$\left. \frac{d^2 F_N(x)}{dx^2} \right\|_{{\mu_N}+{\sigma_N}u} = \left[ \frac{1}{4}-\frac{{\kappa_N}}{{\mu_N}+{\sigma_N}u}+\frac{\lambda^2-1/4}{({\mu_N}+{\sigma_N}u)^2}\right] F_N({\mu_N}+{\sigma_N}u) \;,$$ and since we already know that $F_N({\mu_N}+{\sigma_N}s){\rightarrow}{\mathrm{Ai}}(s)$, we first need to check that, pointwise, $${\sigma_N}^2 \left[ \frac{1}{4}-\frac{{\kappa_N}}{{\mu_N}+{\sigma_N}s}+\frac{\lambda^2-1/4}{({\mu_N}+{\sigma_N}s)^2}\right] {\rightarrow}s \;.$$ In turn, this reduces to showing that $$\begin{gathered} {\sigma_N}^2 \left[ \frac{1}{4}-\frac{{\kappa_N}}{{\mu_N}}+\frac{\lambda^2-1/4}{{\mu_N}^2}\right] {\rightarrow}0 \;, \text{ and }\\ \frac{{\sigma_N}^3}{{\mu_N}}\left[\frac{{\kappa_N}}{{\mu_N}}-2 \frac{\lambda^2-1/4}{{\mu_N}^2}\right] {\rightarrow}1 \;.\end{gathered}$$ The first result comes from the remarkable equality ${\kappa_N}/{\mu_N}-\lambda^2/{\mu_N}^2=1/4$, which follows from the fact that if we call $x=\sqrt{N_+/(N+\alpha)_+}$, we have ${\kappa_N}/{\mu_N}=.5-x/(1+x)^2$ and $\lambda^2/{\mu_N}^2=.25-x/(1+x)^2$. Using these estimates, we see that ${\kappa_N}/{\mu_N}-2(\lambda/{\mu_N})^2 =x/(1+x)^2 \sim \sqrt{N_+/n_+}$, from which we conclude that the second result holds.\ Note that if we changed the centering and scaling (replacing $n$ by $\tilde{n}=n+\alpha$ and $N$ by $\tilde{N}=N+\beta$), by studying the first expression in this case as a “perturbation" of the study we just did, and using the fact that $\mu_{\tilde{N}}-{\mu_N}={\mathrm{O}}(\sqrt{n/N})$, one could show that the first expression is then ${\mathrm{O}}(N^{-1/3})$, and so the result would hold. We also have corresponding results for the second expression. This shows that we have some freedom in the centering and scaling we pick. It is also needed to show that $${\sigma_N}^2 \phi'({\mu_N}+{\sigma_N}s){\rightarrow}{\mathrm{Ai}}(s) \;,$$ since in our splitting of $\phi$, the second part $\phi_{II,N}$ corresponds to parameters $(n-1,N-1)$, but is centered and scaled using ${\mu_N}$ and ${\sigma_N}$, defined with $(n,N)$. To show that the sequence of functions we are interested in is bounded above by an integrable function, we split $[s_0,\infty)$ into $[s_0,\sqrt{n}]$ and $[\sqrt{n},\infty)$. On the first interval, we can apply the previous results since ${\sigma_N}s/{\mu_N}$ is small compared to 1. So in particular the whole integrand will be smaller that $\Delta(s_0) (1+\|s\|)^2 \exp(-s/2)$, after taking into account the properties of $F_N$. On the other hand, on $[\sqrt{n},\infty)$, ${\sigma_N}^2 \leq s^2$, and the denominators involving $s$ are bigger than ${\mu_N}$ and ${\mu_N}^2$ respectively, which gives immediately that the integrand is less than $\Delta(s_0) s^2 \exp(-s/2)$. From this we conclude that the integrand is less than $\Delta(s_0) \exp(-s/4)$, for instance, and that therefore the derivative we are interested in is too. It then follows easily that (\[eq:sosh1\]) is true, and we also showed that the left-hand side of (\[eq:sosh1\]) is dominated on $[s_0,\infty)$ and for $N>N_0(s_0)$ by $\Delta(s_0){\mathrm{e}}^{-s/4}$. ### A8.2: Proof of (\[eq:sosh2\]) {#a8.2-proof-of-eqsosh2 .unnumbered} The approach laid out in @sosh p.1044 works after some modifications. We first write $$\int_0^{\infty}\left(\int_0^{z}\phi(u) du \; \psi(y+z)\right) dz = \int_0^{n^{5/8}}\left(\int_0^{z}\phi(u) du \; \psi(y+z)\right) dz+ \int_{n^{5/8}}^{\infty}\left(\int_0^{z}\phi(u) du \; \psi(y+z)\right) dz \;.$$ Then we can check, via a third order asymptotic development in $x$ of the right-hand side of equation (2.10) in @olver80, that equation (2.18) therein is still true in our case, since, with his notations, $x_N\leq n^{-3/8}$. Therefore, the analysis carried out after equation (3.21) of the same reference applies, and after integration of the expansion following (3.22) adapted to our situation, we can show that $$\int_0^{n^{5/8}} \phi(u) du = {\mathrm{O}}(n^{-n/16})$$ With this estimate and this splitting of $[0,\infty)$, the rest of Soshnikov’s argument holds true and therefore (\[eq:sosh2\]) can be verified. A9: A quick look at special functions {#a9-a-quick-look-at-special-functions .unnumbered} ------------------------------------- In this note, we mentioned three types of special functions, Airy, Whittaker, and parabolic cylinder functions. We recall their definition in this appendix, as well as the main ideas behind some of the transformations Olver used. To justify their introduction, let us say that they play a special role because it is possible, in the setting we were in, to write the functions we studied as a perturbation of the differential equations these functions satisfy. ### A9.1: Airy function {#a9.1-airy-function .unnumbered} Let us consider the following second order differential equation: $$\frac{d^2w}{dx^2}=xw \;.$$ #### General remark: Recessive solutions Since these functions are used to get asymptotic expansions, it makes sense to define the independent solutions with respect to their behavior at $+\infty$. Usually, independent solutions $w_1$ and $w_2$ are sought, so that $w_2={\mathrm{o}}(w_1)$ at a particular point of the (extended) real line. In our cases, it will be $\infty$. $w_2$ is called the recessive solution. That leaves the problem underdetermined, but with this in mind, one can then give enough constraints so the problem is fully determined, and solve in terms of recessive and dominant solutions. For a more precise definition of recessivity, see @olver, p.155. In the case of the Airy function, we have for example: (from @olver, 11.1, p.392) $$\begin{aligned} {\mathrm{Ai}}(x)&= \frac{1}{\pi}\int_0^{\infty}\cos(t^3/3+xt)dt \\ {\mathrm{Bi}}(x)&= \frac{1}{\pi}\int_0^{\infty}\{\exp(-t^3/3+xt)+\sin(t^3/3+xt)\}dt\end{aligned}$$ ### A9.2: Whittaker functions {#a9.2-whittaker-functions .unnumbered} These are solution of the following differential equation $$\label{eq:whittaker} \frac{d^2 W}{dx^2}=\left(\frac{1}{4}-\frac{\kappa}{z}+\frac{\lambda^2 -1/4}{z^2}\right)W \;.$$ ${W_{\kappa,\lambda}}$, the recessive solution at $\infty$, is obtained by requiring $${W_{\kappa,\lambda}}(x) \sim e^{-x/2}x^{\kappa} \; \text{as } x{\rightarrow}\infty \;.$$ The other solution is $M_{\kappa,\lambda}$, which is required to satisfy $$M_{\kappa,\lambda}(x) \sim x^{\lambda+1/2} \; \text{as } x{\rightarrow}0^+ \;.$$ For more detail on these, see @olver, p.260, or @olver80. ### A9.3: Parabolic cylinder functions {#a9.3-parabolic-cylinder-functions .unnumbered} According to @olver59, equation (2.9) p.133, [parabolic cylinder functions]{} satisfy (in the case we are interested in) $$\frac{d^2 W}{dx^2}=\left(\frac{1}{4}x^2+a\right)W \;.$$ $U(a,x)$ is chosen to satisfy $$U(a,x) \sim x^{-a-1/2}{\mathrm{e}}^{-x^2/4} \; \text{as } x{\rightarrow}+\infty \;.$$ On the other hand, $\bar{U}$ satisfies $$\bar{U}(a,x) \sim (2/\pi)^{1/2} \Gamma(1/2-a)x^{a-1/2}{\mathrm{e}}^{x^2/4} \; \text{as } x{\rightarrow}+\infty\;.$$ $\bar{U}$’s definition is actually fairly complicated, and can be found in @olver59, equation (2.12) or in @olver75, section 5.1. ### A9.4: On the usage of these functions {#a9.4-on-the-usage-of-these-functions .unnumbered} As we mentioned earlier, these functions play a central role because it is relatively easy to transform the equations in which we are interested into one of the three mentioned above, or a perturbation of it. Then a range of techniques are available to study the effect of the perturbation, and one can sometimes, and obviously in the case we examine, get asymptotic expansions in terms of the “non-perturbed" solutions. Since these functions are quite well known, information can be gathered about the function of original interest this way.\ For example, in @imj, section 5, after the scaling $\xi=x/\kappa$, the Whittaker equation (\[eq:whittaker\]) becomes $$\frac{d^2W}{d\xi^2}=\left(\kappa^2\frac{(\xi-\xi_1)(\xi-\xi_2)}{4\xi^2}-\frac{1}{4\xi^2}\right)W \;.$$ Using the Liouville-Green transformation $\zeta(d\zeta/d\xi)^2=(\xi-\xi_1)(\xi-\xi_2)/(4\xi^2)$, with $w=(d\zeta/d\xi)^{-1/2} W$, one has $$\frac{d^2w}{d\zeta^2}=\{\kappa^2 \zeta + \psi(\zeta)\} w \;.$$ This is a perturbation of the (scaled) Airy equation, for ${\mathrm{Ai}}(\kappa^{2/3} \zeta)$ and ${\mathrm{Bi}}(\kappa^{2/3} \zeta)$ are solutions of $d^2w/d\zeta^2=\kappa^2 \zeta w$.\ $w$ is not $W$, but it can be related to it, and it is through this mean that Johnstone did his original analysis. As $\psi$ is a relatively involved function of $\xi$ and $\zeta$, we do not explicit it, but just mention that the understanding of $\psi$ is key to getting the uniform bound (\[whattoshow2\]). For more on this, see @imj or @olver, theorem 11.3.1 p.399. The problem we encountered (and mentioned in \[errorcontrolproblem\]) about the error control function is exactly here: we could not get enough information about the behavior of $\psi$ under ${\mathbf{AB}}$, so we slightly changed approach and turned to other studies. In @olver80, Olver starts with equation (\[eq:whittaker\]), where the dummy variable was $z$. Writing $x=z/\lambda$ and $l=\kappa/\lambda$, he gets $$\frac{d^2 W}{dx^2}=\left(\lambda^2g(x)-\frac{1}{4x^2}\right)W \;.$$ As he aims to expand the solution in terms of parabolic cylinder functions, he changes variables another time, by writing $$W=\left(\frac{dx}{d\zeta}\right)^{1/2} w \; \; \;, \; \left(\frac{d\zeta}{dx}\right)^2=\frac{x^2-4lx+4}{4x^2(\zeta^2-\beta^2)} \;,$$ with $\beta=\{2(l-1)\}^{1/2}$. Hence, he gets $$\frac{d^2w}{d\zeta^2}=\{\kappa^2 (\zeta^2-\beta^2) + \psi(\kappa,\beta,\zeta)\} w \;,$$ with $\psi(\kappa,\beta,\zeta)=-\dot{x}^2/(4x^2)+\dot{x}^{1/2}d^2(\dot{x}^{-1/2})/d\zeta^2$. His @olver75 is a study of this type of equations, and in particular of the control of the deviation of the solution of the previous equation to the corresponding parabolic cylinder function. In @olver80, he studies very explicitly the abstract estimate he gets in @olver75 in the case of Whittaker functions. We use this repeatedly in our study, as it is essential to get the crucial property (\[whattoshow2\]). [^1]: Acknowledgements: The author is grateful to Pr. Iain Johnstone for many discussions, key references, and guidance and to Pr. David Donoho for his helpful comments and support. Supported in part by NSF -0140698 and -008584 (ITR). AMS 2000 SC: Primary 62E20, Secondary 62H25. Key words and Phrases : Principal Component Analysis, largest singular value, Tracy-Widom distribution, Fredholm determinant, Random Matrix Theory, Wishart Matrices. Contact : `[email protected]`	{ "pile_set_name": "ArXiv" }
--- abstract: 'We prove quantitative equidistribution results for actions of Abelian subgroups of the $2g+1$ dimensional Heisenberg group acting on compact $2g+1$-dimensional homogeneous nilmanifolds. The results are based on the study of the $C^\infty$-cohomology of the action of such groups, on tame estimates of the associated cohomological equations and on a renormalisation method initially applied by Forni to surface flows and by Forni and the second author to other parabolic flows. As an application we obtain bounds for finite Theta sums defined by real quadratic forms in $g$ variables, generalizing the classical results of Hardy and Littlewood [@MR1555099; @MR1555214] and the optimal result of Fiedler, Jurkat and Körner [@MR0563894] to higher dimension.' address: - 'Centro de Matem'' atica, Universidade do Minho, Campus de Gualtar, 4710-057 Braga, PORTUGAL.' - 'Unit'' e Mixte de Recherche CNRS 8524, Unit'' e de Formation et Recherche de Math'' ematiques, Universit'' e de Lille 1, F59655 Villeneuve d’Asq CEDEX, FRANCE. ' author: - Salvatore Cosentino - Livio Flaminio bibliography: - 'heisenberg.bib' title: 'Equidistribution for higher-rank Abelian actions on Heisenberg nilmanifolds' ---	{ "pile_set_name": "ArXiv" }
--- abstract: 'We have reanalyzed SUMER observations of a parcel of coronal gas using new collisional ionization equilibrium (CIE) calculations. These improved CIE fractional abundances were calculated using state-of-the-art electron-ion recombination data for K-shell, L-shell, Na-like, and Mg-like ions of all elements from H through Zn and, additionally, Al- through Ar-like ions of Fe. They also incorporate the latest recommended electron impact ionization data for all ions of H through Zn. Improved CIE calculations based on these recombination and ionization data are presented here. We have also developed a new systematic method for determining the average emission measure ($EM$) and electron temperature ($T_e$) of an isothermal plasma. With our new CIE data and our new approach for determining average $EM$ and $T_e$, we have reanalyzed SUMER observations of the solar corona. We have compared our results with those of previous studies and found some significant differences for the derived $EM$ and $T_e$. We have also calculated the enhancement of coronal elemental abundances compared to their photospheric abundances, using the SUMER observations themselves to determine the abundance enhancement factor for each of the emitting elements. Our observationally derived first ionization potential (FIP) factors are in reasonable agreement with the theoretical model of @Lami08a.' author: - 'P. Bryans, E. Landi, and D. W. Savin' title: 'A new approach to analyzing solar coronal spectra and updated collisional ionization equilibrium calculations. II. Updated ionization rate coefficients ' --- Introduction {#sec:intro} ============ Investigating the dynamics of the solar corona is crucial if one is to understand fundamental solar and heliospheric physics. The corona also greatly influences the Sun-Earth interaction, as it is from here that the solar wind originates. Explosive events in the corona can deposit up to $2 \times 10^{16}$ g of ionized particles into the solar wind [@Hund93a]. These can have a profound effect on the Earth’s magnetosphere and ionosphere. Hence the investigation of the corona is of obvious importance. Over the years there has been a significant amount of research invested in developing our understanding of the corona (reviewed by Aschwanden 2004 and Foukal 2004). However, gaps remain in our understanding of some of the most fundamental processes taking place in the corona. For example, the so-called coronal heating problem remains unsolved [@Gudi05a; @Klim06a] and we are still unable to explain the onset processes that cause solar flares and coronal mass ejections [@Forb00a; @Prie02a]. One of the most powerful tools for understanding the properties of the solar corona is spectroscopy [@Tand88a; @Fouk04a]. Analyzing the spectral emission of the corona can give the temperature and density of the plasma, as well as information on the complex plasma structures common in this region of the Sun’s atmosphere. One common approach to this end is to calculate the emission measure ($EM$) of the gas [e.g., @Raym81a]. The $EM$ technique is particularly useful for studying the properties of the upper solar atmosphere. In this region, conditions are such that the plasma can often be described as low-density and in steady-state and the emitting region as constant in density and temperature. These relatively simple conditions allow one to neglect density effects and to assume all emission is from an isothermal plasma. For example, @Land02a compared off-disk spectral observations of the solar corona with predictions from the CHIANTI version 3 atomic database [@Dere97a; @Dere01a]. @Land02a calculated the $EM$ of the plasma based on the observed intensities using the atomic data assembled together in CHIANTI. From this, they also infer the electron temperature ($T_e$) of the emitting plasma. However, the power of this spectroscopic diagnostic can be limited by our understanding of the underlying atomic physics that produce the observed spectrum. Reliable $EM$ calculations require accurate fractional abundances for the ionization stages of the elements present in the plasma. For a plasma in collisional ionization equilibrium (CIE; sometimes also called coronal equilibrium), the atomic data needed for such a spectral analysis includes rate coefficients for electron-ion recombination and electron-impact ionization. These data directly affect the calculated ionic fractional abundances of the gas. The fractional abundances, in turn, are used to determine the $EM$. Hence the reliability of the CIE calculations is critical. The recommended CIE calculations at the time of the work by @Land02a were those of @Mazz98a. Recently, however, state-of-the-art electron-ion recombination data have been published for K-shell, L-shell, and Na-like ions of all elements from H through Zn [@Badn03a; @Badn06a; @Badn06b; @Badn06c; @Gu03a; @Gu03b; @Gu04a]. Based on these new recombination data, a significant update of the recommended CIE fractional abundances was published recently by @Brya06a [Paper I in this series]. Since then additional recombination data have been published for Mg-like ions of H through Zn [@Altu07a] and Al- through Ar-like ions of Fe [@Badn06d; @Badn06e]. Electron impact ionization (EII) data have also been updated recently by @Suno06a, @Dere07a, and @Matt07a. Of these three, the recommended EII data of @Dere07a, which we adopt, provide the only complete available set of rate coefficients for all ions of H through Zn. Here we have updated the results of @Brya06a using these new recombination and ionization data. One of the motivations behind this paper is to investigate the effects of the recent improvements in CIE calculations on solar observations. Since the @Land02a paper there have been other improved atomic data (e.g., the improvement of the model for N-like ions). These have been made available in a more recent CHIANTI release—version 5.2 [@Land06a]. It is this version we use here. We also investigate here the observed relative elemental abundances and the first ionization potential (FIP) effect. The FIP effect is the discrepancy between the coronal and photospheric elemental abundances, possibly explained by the pondermotive force induced by the propagation of Alfvén waves through the chromosphere [@Lami04a; @Lami08a]. Elements with a FIP of below $\sim 10$ eV appear to have a coronal abundance that is enhanced by a factor of a few over their photospheric abundance [see, e.g., the review by @Feld00a]. Often, the FIP effect is accounted for by multiplying the abundance of the low-FIP elements by a single scaling factor [such as 3.5 as was done in @Land02a]. In the present work, we investigate the reliability of this approach by quantifying the FIP effect based on the observations themselves. We determine the $EM$ from the high-FIP element Ar and then scale the elemental abundances of the moderate- and low-FIP elements so that their derived $EM$s match that of Ar. We compare our derived abundances with those of a previous analysis of the same observation [@Feld98a] as well as with theoretical predictions [@Lami08a]. An important aspect of the present paper is the development of a sound mathematical method of determining the average $EM$ and $T_e$ of an isothermal plasma. Previous studies have done this in a less rigorous manner. @Land02a, for example, evaluate plots of $EM$ versus $T_e$ curves and give a “by eye” estimate of the average value of the $EM$ and $T_e$ and their associated errors. This method allows human bias to become important when deciding which curve crossings to include in the selection. In addition, it is unclear to what this “average” actually corresponds mathematically. The fact that the analysis is performed on graphs with logarithmic axes suggests that by-eye average is closer to the geometric mean than the arithmetic mean. Finally, no account is taken of the reliability of the atomic data used to calculate fractional abundances. @Brya06a showed that CIE results are unreliable at temperatures where the ionic fractional abundances are less than 1%. Previous studies have failed to account for this when using the CIE data in the $EM$ analysis. Taking the above four paragraphs into account, we have reanalyzed the observations of @Land02a. The rest of this paper is organized as follows: In Sec. \[sec:observations\] we give a description of the observing sequence, the observed lines and their categorizations by @Land02a. Section \[sec:em\] defines the $EM$ and explains the method we use to determine the plasma temperature from the observed line intensities. In Sec. \[sec:cie\] we review the recent developments in the understanding of dielectronic and radiative recombination and electron impact ionization, and the subsequent improvement in CIE calculations. We also present updated tables of these CIE calculations, which supersede those of @Brya06a. In Sec. \[sec:method\] we describe our new approach for determining the $EM$ and temperature of an isothermal plasma based on the observed spectral line intensities. Section \[sec:fip\] discusses our method of determining the elemental abundance enhancement factors due to the FIP effect. In Sec. \[sec:groups\] we present the results of our $EM$ calculations for each of the categorizations introduced by @Land02a. Section \[sec:discuss\] discusses the consequences of these results, in particular highlighting discrepancies between the results of this paper and those of @Land02a. In Sec. \[sec:future\] we propose future observations needed to address some of the remaining issues raised by our results here. Concluding remarks are given in Sec. \[sec:conclusion\]. Observations {#sec:observations} ============ The spectrum analyzed by @Land02a, and revisited here, was detected using the Solar Ultraviolet Measurements of Emitted Radiation spectrometer [SUMER; @Wilh95a] onboard the Solar and Heliospheric Observatory (SOHO). The observation spans over 5 hours, from 21:16 UT on 1996 November 21 to 02:28 UT November 22, and was collected in 61 spectral sections. The observing slit imaged at a height $h$ of $1.03R_\odot\lesssim h \lesssim 1.3R_\odot$ above the western limb. The resulting spectrum covers the entire SUMER spectral range of 660–1500 Å. @Land02a give a full description of the observation sequence and data reduction. Table \[tab:linelist\] lists the coronal lines identified in the spectrum and their corresponding transitions [reproduced from @Land02a]. Known typos in the line assignment labels of @Land02a have been corrected; these do not affect their reported results. Landi et al. estimate uncertainties on the extracted line intensities of 25–30%. Twelve of the emission lines observed in this run are omitted from the table here due to their being blended with other emission lines or having uncertain intensities. The remaining spectral lines are split into three distinct groups, labeled in the first column of Table \[tab:linelist\] as: - Forbidden transitions within the ground configuration. - Non–N-like transitions. - N-like transitions. - Transitions between the ground and the first excited configuration: - Allowed $2s-2p$ transitions in the Li-like isoelectronic sequence and allowed $3s-3p$ transitions in the Na-like isoelectronic sequence. - Intercombination transitions in the Be-, B-, C-, and Mg-like isoelectronic sequences. - Transitions between the first and second excited configuration. Within each group and subgroup we have derived the average $T_e$ and $EM$. Categorizing the transitions in this way helps us to better identify any trends in the $EM$ with respect to the transition type. Group I transitions have been further divided into non–N-like and N-like transitions. This separation was originally proposed by @Land02a due to the poor agreement they found for the $T_e$ derived within each of these transition types. This is discussed further in Secs. \[sec:fip\] and \[sec:discuss\]. The subdivision of transition Group II is to allow us to investigate a longstanding discrepancy between $EM$s derived using Li- and Na-like ions and those derived using other isoelectronic sequences [e.g., @Dupr72a; @Feld98a; @Land02a]. We also discuss this further in Secs. \[sec:fip\] and \[sec:discuss\]. Method of Calculating Temperature and Emission Measure {#sec:em} ====================================================== The intensity of an observed spectral line due to a transition from level $j$ to level $i$ in element X of ionization state $+m$ can be written as $$I_{ji} = \frac{1}{4\pi d^2}\int_{V}G_{ji}(T_e,n_e)n_e^2\,{\rm d}V$$ where $n_e$ is the electron density, $V$ is the emitting volume along the line of sight, and $d$ is the distance to the source. $G_{ji}(T_e,n_e)$ is the contribution function, which is defined as $$\label{eqn:cont func} G_{ji}(T_e,n_e)=\frac{n_j({\rm X}^{+m})}{n({\rm X}^{+m})} \frac{n({\rm X}^{+m})}{n({\rm X})} \frac{n({\rm X})}{n({\rm H})} \frac{n({\rm H})}{n_e} \frac{A_{ji}}{n_e}$$ where $n_j({\rm X}^{+m})/n({\rm X}^{+m})$ is the population of upper level $j$ relative to all levels in X$^{+m}$, $n({\rm X}^{+m})/n({\rm X})$ is the fractional abundance of ionization stage $+m$ relative to the sum of all ionization stages of X, $n({\rm X})/n({\rm H})$ is the abundance of element X relative to hydrogen, and $n({\rm H})/n_e$ is the abundance of hydrogen relative to the electron density. $A_{ji}$ is the spontaneous emission coefficient for the transition. For the observation analyzed here, the emitting plasma was found to be isothermal by @Feld99a and @Land02a. For the moment we assume this to be correct but we revisit the validity of the isothermal assumption in Sec. \[sec:discuss groups\]. One can also make the assumption that the region emitting the observed line intensities is at a constant density. While the line-of-sight of the observation covers plasma where densities vary by orders of magnitude, the emission is dominated by a region with a small range of densities around the peak density. Only those emission lines that have a strong density sensitivity in this range will be affected by the density gradient [@Lang90a]. @Feld99a inferred a density of $1.8\times10^8$ cm$^{-3}$ for this observation. A density dependent study of the 74 lines observed here is beyond the scope of our paper. Here we use the inferred density of @Feld99a in our analysis. If we now assume that all the emission comes from the same parcel of gas of nearly constant temperature, $T_c$, and density, we can approximate $$I_{ji}=\frac{G_{ji}(T_c,n_e)}{4\pi d^2} EM$$ where the emission measure $EM$ is defined as $$EM=\int n_e^2\,{\rm d}V$$ and can be evaluated from the observed line intensity as $$\label{eqn:em} EM=4\pi d^2\frac{I_{ji}}{G_{ji}(T_c,n_e)} .$$ This has the same value for all transitions if the constant temperature and density assumption is correct, which we label $EM_c$. Thus, from the observed line intensities, $I_{ji}$, and using accurate data for $G_{ji}(T_e,n_e)$, one can calculate the emission measure and electron temperature of the emitting region. This is done by plotting the $EM$ against $T_e$. The resulting curves for each observed line should intersect at a common point yielding $[T_c,EM_c]$. But this depends on the assumption of constant temperature and density being correct and on the accuracy of the underlying atomic data. Here, one of the issues we are investigating is the effect on solar coronal observations of the newly calculated fractional abundances $$f^m=\frac{n({\rm X}^{+m})}{n({\rm X})}.$$ The units used throughout this paper for $EM$ and $T_e$ are cm$^{-3}$ and K, respectively. For ease of reading, we typically drop these units below. Improved collisional ionization equilibrium (CIE) calculations {#sec:cie} ============================================================== The plasma conditions of the solar upper atmosphere are often described as being optically-thin, low-density, dust-free, and in steady-state or quasi-steady-state. Under these conditions the effects of any radiation field can be ignored, three-body collisions are unimportant, and the ionization balance of the gas is time-independent. This is commonly called CIE or coronal equilibrium. These conditions are not always the case in the solar upper atmosphere in the event of impulsive heating events but, given the inactivity and low density of the plasma analyzed here, they sufficiently describe the observed conditions. For a thorough discussion of plasma conditions where one must treat the time scales and density effects more carefully, we direct the reader to @Summ06a. In CIE, recombination is due primarily to dielectronic recombination (DR) and radiative recombination (RR). At the temperature of peak formation in CIE, DR dominates over RR for most ions. Ionization is primarily a result of electron impact ionization (EII). At temperatures low enough for both atoms and ions to exist, charge transfer (CT) can be both an important recombination and ionization process [@Arna85a; @King96a]. CT is not expected to be important at solar coronal temperatures and is not included in the work of @Mazz98a, @Brya06a, or the present paper. Considering all the ions and levels that need to be taken into account, it is clear that vast quantities of data are needed. Generating them to the accuracy required pushes atomic theoretical and experimental methods to the edge of what is currently achievable and often beyond. For this reason, the CIE data used by the solar physics and astrophysics communities have gone through numerous updates over the years as more reliable atomic data have become available. Recombination rate coefficients {#sec:recom} ------------------------------- The DR and RR rate coefficients used to determine the CIE fractional abundances utilized by @Land02a were those recommended by @Mazz98a. However, there has been a significant improvement in the recombination rate coefficients since then. @Badn03a and @Badn06a [@Badn06b; @Badn06c] have calculated DR and RR rate coefficients for all ionization stages from bare through Na-like for all elements from H through Zn and @Gu03a [@Gu03b; @Gu04a] for a subset of these elements. The methods of Badnell and Gu are of comparable sophistication and their DR results for a given ion agree with one another typically to better than 35% at the electron temperatures where the CIE fractional abundance of that ion is $\ge1\%$. The RR rate coefficients are in even better agreement, typically within 10% over this temperature range. These differences for the DR and RR rate coefficients do not appear to be systematic in any way [@Brya06a]. For both DR and RR outside this temperature range, agreement between these two state-of-the-art theories can become significantly worse. The DR calculations have also been compared to experimental measurements, where they exist, and found to be in agreement to within 35% in the temperature range where the ion forms in CIE. For a fuller discussion of the agreement between recent theories and the agreement between theory and experiment, we direct the reader to @Brya06a. Electron impact ionization rate coefficients {#sec:eii} -------------------------------------------- There have also been recent attempts to improve the state of the EII rate coefficients used in CIE calculations. The most complete of these studies is that of @Dere07a, who produced recommended rate coefficients for all ionization stages of the elements H through Zn. These data are based on a combination of laboratory experiments and theoretical calculations. In addition, there have been works by @Suno06a and @Matt07a that also address the issue of updating the EII database. These works are less complete than that of @Dere07a. @Suno06a provides EII cross sections for all ionization stages of C. @Matt07a provides EII cross sections for all ionization stages of H through O plus Ne and a selection of other ions up to Ge. Between these recent compilations there remain sizable differences in the EII rate coefficients for certain elements, often in the temperature range where an ion forms in CIE. For the ions important to the present work, differences between recent recommended rate coefficients of up to 50% are seen. Larger differences, of up to a factor of $\sim 4$, are found for other ions not observed in this SUMER observation. In short, we do not see the uniform agreement between recommended sets of EII data as we do for the state-of-the-art DR and RR calculations. Despite these outstanding issues regarding the accuracy of the various EII databases, we have used the compilation of @Dere07a to calculate fractional CIE abundances. Of the recent EII compilations, the Dere database offers the most complete selection of rate coefficients. However, given the large differences between the @Dere07a and @Matt07a results, we believe that further analysis of the EII database is required to resolve these differences. Updated CIE calculations {#sec:cie update} ------------------------ The new recombination data of @Badn03a and @Badn06a [@Badn06b; @Badn06c] motivated @Brya06a to calculate new CIE fractional abundances. Their results show large differences from the @Mazz98a data for certain elements. Here we revise the work of @Brya06a to include these newly recommended EII rate coefficients for all elements from H through Zn and some further updates to the DR and RR rate coefficients for selected ions. We calculate CIE fractional abundances using the EII data of @Dere07a for all ions of the elements H through Zn. We also include some corrections for Ca-like ions (K. P. Dere 2007, private communication). The DR and RR rate coefficients used here are those of @Brya06a but updated to include recent corrections to the fitting of some of the rate coefficients [@Badn06b]. We also include recent DR work for Mg-like ions of H through Zn and for Al- through Ar-like ions of Fe [@Altu07a; @Badn06b; @Badn06d; @Badn06e]. The DR and RR data for all other ions are those of @Mazz98a. Here we provide electronic tables of the CIE fractional abundances for all elements from H through Zn calculated using these data (Tables \[tab:H\]–\[tab:Zn\]). These tabulations are provided for a $T_e$ range of $10^4$–$10^9$ K. For ease of comparison with previous CIE fractional abundance calculations we present figures showing the present results along with those of @Mazz98a in Figs. \[fig:H Mazz\]–\[fig:Zn Mazz\], and the present results along with those of @Brya06a in Figs. \[fig:H Bryans\]–\[fig:Zn Bryans\]. A new approach to derive average emission measures and temperatures {#sec:method} =================================================================== Using the method described in Sec. \[sec:em\], the assumption of constant temperature and density, and our updated CIE results, we can calculate the $EM$ curve for each of the observed spectral lines listed in Table \[tab:linelist\]. Due to oversimplifications of the plasma model, uncertainties in the observations, and errors in the atomic data, there is no common intersection of all $EM$ curves at a single $[T_c,EM_c]$. So one must calculate the most likely $EM$ and $T_e$ of the plasma based on the range of values where the $EM$ curves cross one another. To determine these values we have developed a mathematically more rigorous approach than has been used in the past for isothermal plasmas. Here we use the emission lines from Si to illustrate this new method. We calculate the $EM$ curves using a constant electron density of $1.8\times 10^8$ cm$^{-3}$ as was reported by @Feld99a for the same source region. Step 1 of our approach is to take the mean of all crossing points of the $EM$ curves for a given group of lines. This can be seen in the left panel of Fig. \[fig:si\]. In this panel we have marked with an asterisk every crossing point of the $EM$ curves shown. The $EM$ vs. $T_e$ curves vary more slowly in log-log space than in linear space. Also, because of the shape of the curves, any outlying crossings are far more likely to occur at a higher $EM$ than at a lower $EM$. Thus, those crossings that fall far from the preponderance skew the average always towards higher values of the $EM$. To avoid giving undue weight to these points we calculate the mean in log space, where $\langle \log_{10} EM\rangle \le \log_{10}\langle EM\rangle$. This is equivalent to taking the geometric mean rather than the more common arithmetic mean. The log of the geometric mean $EM$ is given by $$\langle\log_{10} EM\rangle=\log_{10}\left( \prod_{i=1}^{n} EM_i \right) ^{1/n} = \frac{1}{n}\sum_{i=1}^{n} \log_{10} EM_i$$ and its standard deviation by $$\delta\langle \log_{10}EM\rangle=\sqrt{\frac{\sum_{i=1}^{n} (\log_{10} EM_i -\langle\log_{10} EM\rangle)^2}{n}}$$ where $n$ is the number of crossing points over which the mean is being taken and $EM_i$ is the value of $EM$ at each of these crossings. By a similar argument the mean and standard deviation of $T_e$ are calculated in the same way. From here on, unless otherwise stated, when we discuss the mean and standard deviation of the $EM$ and $T_e$ we are referring to the geometric mean and geometric standard deviation. In Fig. \[fig:si\] the mean $\log_{10}EM$ and mean $\log_{10}T_e$ are shown as dashed lines and the standard deviations by dotted lines. Step 2 eliminates the less physically probable crossings when two $EM$ curves cross one another more than once. For any two curves we select only the crossing point that is closest, in the $\log EM$-$\log T_e$ plane, to the mean calculated values of the $EM$ and $T_e$ from Step 1. In Step 2 we also exclude some additional unphysical crossing points. In all cases where there are multiple emission lines from a single ion, the $EM$ curves are nearly parallel. Often, these curves nearly overlap with one another and can cross in one or more places. We attribute the crossings of these lines to errors in the effective line emission rate coefficients and/or issues with the observed line intensities. For this reason, we exclude these crossings from our calculation. For Si emission lines, such crossings are seen for Si [viii]{}, [x]{}, and [xi]{} (Fig. \[fig:si\]). Using this reduced set of crossings we recalculate the mean and standard deviation of the $EM$ and temperature. This plot is shown in the middle panel of Fig. \[fig:si\]. In Step 3 we further reduce the dataset by considering only $EM$ curves in the temperature range where $f^m\ge0.01$. The reliability of all published CIE calculations is uncertain below this fractional abundance. @Brya06a compared the results of CIE calculations using 2 different compilations of state-of-the-art DR and RR datasets. Agreement at peak abundance was found to be within 10% and within 50% when going to temperatures where the fractional abundance is 0.01. Outside this temperature range, for values of $f^m<0.01$, the reliability of the CIE calculations grows significantly worse. In the right panel of Fig. \[fig:si\] we show the same $EM$ curves as in the middle panel but only for the temperature range where $f^m\ge0.01$. It is this mean $EM$ and $T_e$ after Step 3 that we consider the most likely $EM$ and $T_e$ for a given set of emission lines. Henceforth, when discussing the results after all three steps of our analysis, we refer to the $EM$ and $T_e$ as coming from the Geometric mean Emission Measure (GEM) method. Coronal abundance enhancement factors {#sec:fip} ===================================== Our first step in determining the coronal abundance of the observed elements is to assume that the high-FIP elements Ne and Ar have the same abundance in the corona as they do in the photosphere. This follows the approach taken by @Feld98a. Using the photospheric abundances of Feldman & Laming (2000; see Table \[tab:abundances\]) we have calculated the geometric mean $EM$ from emission lines of Ne and Ar using the GEM method outlined in Sec. \[sec:method\], giving $\langle\log_{10} EM_{\rm high-FIP}\rangle$. An objective of the present paper is to investigate the apparent abundance discrepancy of Li- and Na-like ions. Previous studies, such as those of @Dupr72a, @Feld98a, and @Land02a, have found the abundance of these ions to be greater than those of ions in other isoelectronic sequences. In order that this discrepancy does not affect our calculation of the FIP factors of each element we do not include any Li- or Na-like lines in the calculations of the FIP factors detailed below. @Land02a also reported a difference in $T_e$ derived from N-like and non–N-like ions within Group I. However, unlike the Li- and Na-like abundance discrepancy, this has not been reported in the literature previously. If we adopt the uniform FIP factor of 3.5 used by @Land02a for all high-FIP N-like and non–N-like ions and implement our GEM method we find no discrepancy in the $T_e$ derived from N-like and non–N-like ions. This is discussed in more detail in Sec. \[sec:discuss\]. For these reasons, in this section we include both N-like and non–N-like ions in our analysis. The results of the GEM analysis of the high-FIP elements can be seen in the upper left panel of Fig. \[fig:all elem\]. The exclusion of the Li- and Na-like ions results in only a single crossing remaining after the 3 steps—due to 2 emission lines from Ar [xi]{} and Ar [xii]{}. We use the value of the $EM$ at this point as our reference value. Restricting the temperature range to that where the fractional abundance of an ion is greater than 1% limits us to this single crossing since the 2 Ne [vii]{} $EM$ curves are below this limit at the $T_e$ values where they cross the $EM$ curves of Ar [xi]{} and Ar [xii]{}. It is not ideal that we are left with only a single crossing but we believe this represents an improvement over the work of @Feld98a. There they used only a single line, whereas here we use two. Additionally, the line they selected was Li-like O [vi]{}. As we have discussed above, and will also discuss in Sec. \[sec:discuss\], there are several reasons to treat this line with suspicion. It is also worth noting that the crossing of the Ar lines results in $\log_{10} T_e=6.24$. This is $\sim 0.1$ in the dex higher than the temperature derived from the other emission lines (see later in this section and Sec. \[sec:groups\]). However, in the absence of additional non–Li- and Na-like emission from other high-FIP elements, we consider normalizing to this crossing of Ar lines to be the best approach to analyzing this particular observation. We next separate all other emission lines by the element responsible for the emission and, again using the GEM method, calculate the mean $EM$ of each of the low- and moderate-FIP elements individually using the photospheric elemental abundances as our starting value, giving $\langle\log_{10} EM_{\rm X}\rangle$ for each element X. For each of these low- to moderate-FIP elements, we determine an “enhancement factor” $f_{\rm X}$ for the elemental abundance that will result in the same derived $EM$ as found for the high-FIP element Ar. From Eqns. \[eqn:cont func\] and \[eqn:em\] we see that the elemental abundance, $$f(X)=\frac{n({\rm X})}{n({\rm H})},$$ is inversely proportional to the $EM$ of the emitting plasma so the $f_{\rm X}$ values can be calculated as $$\label{eqn:fip} \log_{10}f_{\rm X} = \langle\log_{10}EM_{\rm X}\rangle -\langle\log_{10}EM_{\rm high-FIP}\rangle$$ where $$f_{\rm X} = \frac{f(X)_{\rm corona}}{f(X)_{\rm photosphere}}.$$ For the emission from the elements Mg, Al, Si, S, and Fe we show the $EM$ as a function of $T_e$ in Fig. \[fig:all elem\] where we have used our derived coronal elemental abundances. These are subject to the 3 steps of the GEM method but in this case we only show the last step. The derived elemental abundances are given in Table \[tab:abundances\]. From Eq. \[eqn:fip\], we estimate the absolute error in $\log_{10}f_{\rm X}$ as the quadrature sum of the standard deviations of the $EM$ from the high-FIP elements and the $EM$ from the individual element X, i.e., $$\delta\langle\log_{10}f_{\rm X}\rangle = \sqrt{\delta\langle\log_{10}EM_{\rm X}\rangle^2 + \delta\langle\log_{10}EM_{\rm high-FIP}\rangle^2}$$ However, since only a single crossing of Ar lines is used to determine the high-FIP $EM$, there is no error associated with $\log_{10}EM_{\rm high-FIP}$ and the error in $\log_{10}f_{\rm X}$ reduces to $\delta\langle\log_{10}EM_{\rm X}\rangle$ and is thus probably an underestimate. Given our derived errors in $\langle\log_{10}EM\rangle$ presented in Sec. \[sec:groups\] we estimate $\langle\log_{10}EM_{\rm high-FIP}\rangle$ is good to $\sim\pm 0.3$ in the dex. However, due to insufficient data, we do not attempt to assign an error to $\delta\langle\log_{10}EM_{\rm high-FIP}\rangle$. Instead, we leave the errors in the FIP factors as they are, but note that they are likely underestimates. With the Li- and Na-like lines omitted, we are left with only 2 emission lines from Na and Ca and a single emission line from K in this observation. Both Na emission lines are from the same charge state and their associated $EM$ curves are therefore almost parallel. The same is true for the 2 emission lines from Ca. We thus have no crossing points of the $EM$ curves over the $T_e$ range considered for Na, K, and Ca. Also, as discussed in Sec. \[sec:discuss\], we believe the emission from Li-like N [v]{} and O [vi]{} ions to be from a cooler region of plasma so we do not determine an average $EM$ from the curves crossings of these elements. To determine the FIP factors of Na, K and Ca we use their $EM$ values at the $T_e$ determined from those emission lines for which we have already calculated FIP factors. This mean $T_e$ determination is shown in Fig. \[fig:te\] where emission lines of Ne, Mg, Al, Si, S, Ar, and Fe have been considered (excluding Li- and Na-like ions). This gives a value of $\log_{10} T_e=6.13\pm 0.06$, at which value we calculate the $EM$ of each of the Na, K and Ca lines. (For Na and Ca, for which we have 2 emission lines, we take the average value of the 2 $EM$ values at $\log_{10} T_e=6.13$.) We then determine a FIP factor for each of these elements that will give the same $EM$ as for the high-FIP element Ar. We estimate the absolute error in $\log_{10}f_{\rm X}$ of these 3 elements by calculating their $EM$ at the values of the extremes of the errors associated with $T_e$, i.e., the $EM$ at $\log_{10} T_e=6.07$ and $\log_{10} T_e=6.19$. Table \[tab:abundances\] lists the enhancement factors, which are often called FIP factors or FIP biases, for all of the elements present in the observation. We also give the resulting coronal elemental abundances. The FIP factors are also shown in Fig. \[fig:fip\] alongside the results of @Feld98a. Note that Figs. \[fig:si\]–\[fig:fip\] show the results when using the CIE fractional abundances of the present paper. We have repeated the analysis using the @Mazz98a CIE fractional abundances. We do not show figures of these results, but in Table \[tab:abundances\] we list the FIP factors and coronal abundances determined when using these older CIE data. Analysis by groups {#sec:groups} ================== Using our derived coronal abundances we calculate the $EM$ and $T_e$ of each of the line categorizations given in Sec. \[sec:observations\]. Figures \[fig:1a\]–\[fig:3\] show the GEM approach as applied to each of these groups. We also give the results in Table \[tab:averages\] listing the geometric mean and standard deviation of the $EM$ and $T_e$ after each step of the GEM method. For the Group I and II categorizations we show their individual subcategorizations as well as the groups as a whole. In the case of Group IIb, the emission lines have been further subdivided by separating out the N [v]{} and O [vi]{} lines. This is because the $EM$ curves from these lines do not match well with the others in this group. We elaborate on the possible reasons for this in Sec. \[sec:discuss\]. When we consider Group II as a whole, these lines are also excluded. In addition to the division by groups we calculate the mean $EM$ and $T_e$ from every emission line (but again excluding the Li-like N [v]{} and O [vi]{} lines). This is done both including and excluding Li- and Na-like ions with the results shown in Figs. \[fig:all\] and \[fig:allx\], respectively, and listed in Table \[tab:averages\]. It should be noted that this is not simply the sum of all the crossings from the individual groups. It also includes crossings between lines from different groups and results in a total of 1428 and 872 crossings (including and excluding Li- and Na-like lines, respectively). The results of our analysis, as given in Table \[tab:averages\], are shown in graphical form in Fig. \[fig:em figs\] for the variation of $\log_{10}EM$ versus group and Fig. \[fig:temp figs\] for the variation of $\log_{10}T_e$ with group. The numbers in the data points in these figures are the number of crossings that were used to determine the average value and the errors shown are $\pm \delta\langle\log_{10}EM\rangle$ of the mean. The average and standard deviation of $\log_{10}EM$ and $\log_{10}T_e$ as determined from every emission line are shown for comparison as dashed and dotted lines, respectively. We show these values with and without Li- and Na-like ions included in the $EM$ calculation (i.e., Figs. \[fig:all\] and \[fig:allx\], respectively). The thick lines are the average and standard deviations when Li- and Na-like ions are included (excluding N [v]{} and O [vi]{}) and the thin are when they are excluded. Discussion {#sec:discuss} ========== Updated CIE fractional abundances {#sec:discuss cie} --------------------------------- One of the aims of the present paper is to investigate the effect of our new CIE fractional abundances on the $EM$ analysis. But first we look at how the updated recombination and ionization data impact the fractional abundances themselves. Perhaps the most widely used recommended CIE fractional abundances are those of @Mazz98a. Comparison of the current CIE fractional abundances with these are shown in Figs. \[fig:H Mazz\]–\[fig:Zn Mazz\]. We also compare with the recently recommended CIE fractional abundances of @Brya06a in Figs. \[fig:H Bryans\]–\[fig:Zn Bryans\]. A comparison of the works of @Mazz98a and @Brya06a was discussed in @Brya06a, showing the effects of the new DR and RR data on the @Mazz98a results. We compare the current CIE results with those of @Mazz98a for temperatures where $f^m \ge 0.01$. As discussed in @Brya06a and in Sec. \[sec:method\], the reliability of the atomic data is uncertain below this abundance. Differences between the current CIE results and those of @Mazz98a are large for all elements other than H, He, and Li. Factors of typically at least 2 difference in abundance are found for at least one ionization stage of each of these elements. Differences are often much larger. We draw particular attention to the extremely large differences in abundance and peak formation temperature of Sc, Ti, V, Cr, Mn, Co, Ni, Cu, and Zn in the $T_e$ range of $10^4$–$10^6$ K. Differences for these elements can be up to a factor of 30. Such variation between the current results and those of @Mazz98a is a result of the new recombination and ionization rate coefficients being used here. We also compare our present results with the more recent recommended CIE fractional abundances of @Brya06a. The DR and RR rate coefficients used in this work are largely the same as those used by Bryans et al. with the exception of the Mg-like ions of H through Zn, the Al- through Ar-like ions of Fe, and some corrections to the fitting of other ions. The most significant changes in atomic data bewteen this work and @Brya06a is the introduction of the @Dere07a EII rate coefficients. As expected, differences between the present results and those of @Brya06a are not as large as those found between the present results and those of @Mazz98a. However, large differences do remain. The differences highlighted above for Sc, Ti, V, Cr, Mn, Co, Ni, Cu, and Zn in the $T_e$ range of $10^4$–$10^6$ K are also present in the comparison with Bryans et al. For other elements, abundance differences of a factor a few are not uncommon. We attribute all these differences primarily to the EII rate coefficients. In Secs. \[sec:discuss fip obs\] and \[sec:discuss groups\] we discuss the impact of these updated CIE calculations on the analysis of the present SUMER observation. However, only a selection of the ions discussed in this section are present in the SUMER observation. We recommend that the CIE fractional abundances provided here be used in all future analysis of astrophysical spectra until the next revision of the CIE fractional abundances is published. Comparison with FIP factor observations {#sec:discuss fip obs} --------------------------------------- For this same SUMER observation, FIP factors were also determined by @Feld98a. The present results are shown in comparison to those of Feldman et al. in Fig. \[fig:fip\]. They recommend a FIP factor of 1 for the high-FIP elements, a factor of 4 for the low-FIP elements, and a factor of somewhere between 1 and 2 for S. These results are the basis of the approach taken by @Land02a who assumed a FIP factor of unity for the moderate- and high-FIP elements, S, O, N, Ar, and Ne, and a uniform factor of 3.5 for all of the low-FIP elements, K, Na, Al, Ca, Mg, Fe, and Si. We believe our present results are more robust than those of @Feld98a. Firstly, our reference $EM$ value is taken from the crossing of 2 Ar $EM$ curves whereas @Feld98a use the emission from a single Li-like O [vi]{} line as their reference value. This O [vi]{} line had an order of magnitude more counts than any other line in the dataset used by Feldman et al. and thus seems a natural reference emission line. However, given the apparent systematic abundance discrepancy of Li-like ions (which the authors acknowledge), the O [vi]{} line may not be the most reliable to use as an $EM$ reference value. Furthermore, in determining the FIP factor for each element we generally use more emission lines than @Feld98a. For most elements we have multiple emission lines, ranging from 3 lines for Fe to as many as 18 for Si. The only exceptions are the elements K, Na, and Ca as have already been discussed in Sec. \[sec:fip\]. For K we only have 1 emission line and for Na and Ca we have 2. @Feld98a, however, use only 1 or 2 emission lines to determine the FIP factors for each of the elements they consider. An additional source of unreliability in the @Feld98a results lies in the method they use to estimate the plasma temperature. They use the crossing points of curves of FIP factors vs. $T_e$ from different elements. They estimate $\log_{10}T_e=6.13$ (the same value at which we ultimately arrive) but only calculate these FIP factor vs. $T_e$ curves on a temperature grid of 0.1 in the log. From their figures it is reasonable to conclude that any value in the range of $\log_{10}T_e=6.1$ to 6.2 would fit their data points. In which case, their reported FIP factors could range from $\sim 1.5$ to 11. However, @Feld00a estimate the error in these FIP factors to be of the order of 25% which seems to be a significant underestimate. Of the low-FIP elements, we find rough agreement between our results and those of Feldman et al. in the sense that the abundance of the low-FIP elements is enhanced over the high-FIP elements, though one should note that Feldman et al. did not ascribe errors to their results. The largest differences between our results and those of Feldman et al. occur for Na and Ca, where we find differences of a factor of 2.5 and 1.5 respectively. However, our results for these elements should be considered with some care since they are not determined from an average of crossing points but from the $EM$ at a given $T_e$. The error bars on our results for Na and Ca are also relatively large and the Feldman et al. results lie within these errors. Our result for K (a FIP factor of 1.75) is in disagreement with the Feldman et al. conclusion that the low-FIP elements are best fitted with an enhancement factor of 4. However, it should be noted that Feldman et al. did not calculate the FIP factor for K itself and that our analysis uses only one line of K. Finally, we compare the FIP factor results of our GEM method when we utilize the CIE fractional abundances of the present paper and those of @Mazz98a. These results are given in Table \[tab:abundances\]. We find that the effect of our new CIE fractional abundances is largest for K, Ca, and Fe. In the case of K and Fe the differences in FIP factor are not within the estimated errors using our new CIE results. Naturally, these differences are also seen in the log of the inferred coronal abundances. Comparison with FIP factor model {#sec:discuss fip model} -------------------------------- The FIP effect model of @Lami04a [@Lami08a] allows an opportunity to quantitatively compare our derived coronal elemental abundances with those of theory. The Laming model builds on that of @Schw99a by explaining the FIP effect in terms of Alfvén waves in the chromosphere. These Alfvén waves drive a pondermotive force on their reflection or transmission at the chromosphere-corona boundary which results in the elemental fractionation. The extent of the FIP effect on each species is dependent on the upward energy flux of the Alfvén waves. @Lami08a gives results for a number of wave energy fluxes and we compare these results with ours for wave energy fluxes of 2, 8 and 32 in units of $10^6$ ergs cm$^{-2}$ s$^{-1}$. We show these comparisons in Fig. \[fig:fip model\]. Our results suggest that upward wave energy fluxes in this range best describe the solar conditions at the time of this particular SUMER observation. Our data generally fit the model well, with the exception of K. However, as has already been discussed in Sec. \[sec:fip\], our result for K should be considered less reliable than the other elements since we were limited to only a single K emission line in the SUMER observation. It should also be noted that the low-FIP results of the present work were calculated relative to a high-FIP enhancement of 1, while in the Laming model the high-FIP elements do show a slight abundance variation dependent on their FIP value. If we were to normalize to the Ar FIP factor of the @Lami08a model this would introduce a shift in our FIP factors of somewhere between a factor of 0.88 to a factor of 1.77. Groups {#sec:discuss groups} ------ We have used the same group splitting as that used by @Land02a and thus can compare directly with their results. Table \[tab:Landi\] shows their results for the mean and “error” of $\log_{10}EM$ and $\log_{10}T_e$ for the various groups. However, unlike the present work, Landi et al. quotes the mean and the error as judged by eye as opposed to our GEM method. As our results demonstrate, they have considerably underestimated the uncertainty of their results. The results of @Land02a suggest a difference in the temperature derived from the subsets of Group I, with $\log_{10} T_e=6.13\pm 0.01$ and $6.17\pm 0.01$ for Groups Ia and Ib, respectively. We do not see this difference in our analysis. In the present work, Groups Ia and Ib give $\log_{10} T_e=6.16\pm 0.07$ and $6.16\pm 0.05$, respectively. Within our error bars, we see no distinction between the N-like and non–N-like ions in this group. Using the same uniform low-FIP factor of 3.5 used by @Land02a and the GEM method, the distinction remains unobserved as we find values of $\log_{10} T_e=6.16\pm 0.05$ and $6.17\pm 0.03$ for Groups Ia and Ib, respectively. The temperatures derived from Groups IIa, IIb, and III agree reasonably well with those of @Land02a. We note that Landi et al. excluded the N [v]{} and O [vi]{} lines from their calculation of the Group IIa lines. When comparing with their results we also exclude these lines. Figure \[fig:2a\] shows the $EM$ curves for the lines in Group IIa. The largest discrepancies from the other lines in this group can be seen to come from the 2 pairs of N [v]{} and O [vi]{} lines. It is interesting to note that these lines are the lowest in temperature of peak formation of all the ions considered here (see Figs. \[fig:H Bryans\]–\[fig:Zn Bryans\] and Tables \[tab:H\]–\[tab:Zn\]) and as a result the majority of the crossings from these lines are excluded in the right panel of Fig. \[fig:2a\] when we ignore fractional abundances below 0.01. This perhaps goes some way to highlight the need for care when using fractional abundances of such low values. Given the disagreement with the other lines in Group IIa, and the lower formation temperature of N [v]{} and O [vi]{} compared to the other ions in the group, it is possible that the emission lines from these two ions originate from a different region of plasma. Thus, we have excluded the O [vi]{} and N [v]{} lines and recalculated the $EM$ curves. Figure \[fig:2ax\] shows this reduced set of $EM$ curves. We have also done the same for the O and N lines on their own in Fig. \[fig:2a\_on\]. A much lower average temperature of $\log_{10} T_e=5.44$ is derived from these curves. We do not give an estimated error on this value since the final result comes from a single crossing point and has no standard deviation. Given that the N [v]{} and O [vi]{} ions have lower formation temperatures than the other ions of this observation, this suggests that the source of emission from these ions is from a different region of plasma with a lower temperature than that emitting the lines from other elements. One of the questions that this paper seeks to address is the apparent discrepancy between the $EM$s derived from Li- and Na-like and that of all other ions. This has been identified previously [e.g., @Dupr72a; @Feld98a]. All Li- and Na-like lines in this observation come from transitions between the ground and first excited configuration, i.e., our Group II. So we first investigate the difference between Li- and Na-like ions and all other ions within this group. Our results for the Li- and Na-like ions are shown in Fig. \[fig:2ax\] (excluding N [v]{} and O [vi]{}) and can be compared to the $EM$ from the other Group II ions shown in Fig. \[fig:2b\]. Results are also given in Table \[tab:averages\]. A comparison of our present results and those of @Land02a can be seen in Table \[tab:Landi\]. The combination of using the most up-to-date atomic data, an improved method of arriving at the most likely $EM$ and $T_e$, and a reanalysis of the FIP factors, has led to us finding no sign of any discrepancy between the emission of Li- and Na-like ions and all the other Group II ions. In the present case, the difference between the $EM$ from Groups IIa(excluding N [v]{} and O [vi]{} lines) and IIb is within the error bars on the $EM$ of Group IIa. In addition to the comparison of $EM$ within Group II, we also compare the $EM$ from the Li- and Na-like ions (Group IIa$^1$) with the $EM$ derived from every other ion in the observation (i.e., those from Groups I, IIb, and III). The comparison between the $EM$s from the Li- and Na-like ions and all other ions is shown in Fig. \[fig:em figs\]. One can see that the $EM$ from Li- and Na-like lines alone (point IIa$^1$) overlaps, within the errors, with the average determined excluding Li- and Na-like ions (thin dashed line). Thus, we find no statistically meaningful difference in the $EM$ derived from Li- and Na-like ions and that from every other ion in the observation. Given that we find fairly good agreement in $EM$ and $T_e$ between each of the group categories, our best estimate of the $EM$ and $T_e$ of the emitting plasma is found by applying our analysis method to every emission line (excluding the discrepant N [v]{} and O [vi]{} lines). These results are shown in Fig. \[fig:all\] and give $\log_{10} EM=42.98\pm 0.29$ and $\log_{10} T_e=6.12\pm 0.07$. If, in addition to excluding the N [v]{} and O [vi]{} lines, we also exclude the Li- and Na-like lines then the calculated values become $\log_{10} EM=43.02\pm 0.29$ and $\log_{10} T_e=6.13\pm 0.06$ (see Fig. \[fig:allx\]). Landi et al. estimate $\log_{10} EM=43.20\pm 0.15$ and $\log_{10} T_e=6.13$ (no error given) for the plasma by combining results from Groups I and IIb. Our results agree, within our errors, with those of @Land02a. Our results have larger errors, which we believe to be more realistic due to our more rigorous method of calculating the mean and standard deviation of $EM$ and $T_e$. To investigate the effects of the updated CIE data on these results, we compare the $EM$ and $T_e$ derived for each group when utilizing the recommended CIE fractional abundances of the present paper and those of @Mazz98a. These results are given in Table \[tab:Landi\]. While differences are found, they are all within the errors. It is interesting to note that the large differences found in the FIP factors do not translate into differences on the same scale for the derived $EM$ and $T_e$. Nonetheless, this would not necessarily be the case when applied to other observations, so we recommend the future use of the CIE fractional abundances presented here. Other Issues ------------ Despite an overall general agreement in the $EM$ and $T_e$ of each of the groups, there are a number of indications that the observed emission does not come from an isothermal plasma. We have already discussed the possibility that the Li-like N [v]{} and O [vi]{} lines come from a cooler region of gas. Even when these lines are removed from the Group IIa categorization, there remains a large scatter in the crossing points of the emission from Li- and Na-like ions (Fig. \[fig:2ax\]) suggesting the isothermal assumption is not entirely accurate. There is also some evidence of a low temperature component from Groups I and IIb (Figs. \[fig:1\] and \[fig:2b\], respectively). It is also possible that the relatively large errors in $EM$ and $T_e$ are suggestive of a non-isothermal plasma. To determine whether the crossings that fall away from the average are indeed a product of the non-isothermal nature of the plasma, and not some error in the atomic data, one would have to perform a differential emission measure (DEM) analysis, which is beyond the scope of this paper. A further possible source of error in our analysis is that the ionization balance was calculated using the zero-density approximation. This issue has been raised by @Feld00a in reference to Fe$^{8+}$ emission. They claim that over half of the population can be in metastable levels at coronal densities, but there are no emission lines from Fe$^{8+}$ in the observation analyzed in the present paper. The sensitivity of emission from Li-like ions has been investigated by @Doyl05a. These authors found that the contribution function of emission from Li-like lines only becomes significantly affected on reaching densities $\ge 10^{11}$ cm$^{-3}$, orders of magnitude higher than the density of $1.8\times 10^8$ cm$^{-3}$ inferred by @Feld99a for the observation analyzed here. We thus expect the zero-density approximation to be valid in the present case, but a full density-dependent analysis of every emission line in the observation would be required before one could answer this issue with complete certainty. Again, such a study is beyond the scope of this paper. Proposals for future observations {#sec:future} ================================= Our work shows that SUMER observations can go a long way towards constraining FIP models such as those of @Lami04a [@Lami08a]. Even better constraints can be achieved through the simultaneous observation of lines from a number of additional charge states. More lines from high-FIP elements such as N, O, Ne, and Ar are required to better determine the $EM$ for these high-FIP elements, which can then be used to normalize the low-FIP elements. For N, O, and Ne, emission lines from H- and He-like stages need to be observed to avoid using Li-like ions. These charge states are predicted to be abundant for these elements at coronal temperatures. This may require simultaneous observations using separate, cross-calibrated spectrometers. For Ar, emission lines from the Ne-, F-, O-, and N-like ions would lie in the $6.0\le\log T_e\le 6.2$ range typical of the corona. Additional line observations from elements such as Na, K, and Ca, for which we have few lines in the present observation, are also needed to better constrain their FIP factors. Summary {#sec:conclusion} ======= This work has reanalyzed data from a SUMER coronal observation in an attempt to improve upon previous methods of analysis. We have given a brief review of and implemented state-of-the-art electron-ion recombination and ionization data. We have updated the CIE results of @Brya06a by using recently published DR data for Mg-like ions of the elements from H through Zn and for Al- through Ar-like Fe ions, and have updated the EII data to those of @Dere07a for all ions of H through Zn. We have also set out a new, mathematically rigorous, approach for determining the $EM$ and $T_e$ of an emitting plasma within the isothermal approximation. Using these new CIE data and our approach for determining the $EM$, we calculated the FIP factors of the observed elements. Our assessment is generally in reasonable agreement with a previous study of the FIP factors [@Feld98a]. Also, we are in reasonable agreement with the FIP-effect model of @Lami08a using an Alfvén wave energy flux in the range $\sim 2$–$32\times10^6$ erg cm$^{-2}$ s$^{-1}$. However, our results differ from those of @Land02a in certain respects. The difference between the temperature derived using lines from non–N- and N-like ions is not evident when we apply out analysis technique. Also, the previously reported discrepancy between the $EM$ derived from Li- and Na-like lines and the $EM$ from all other lines (Groups I, IIb, and III) is not supported by our results, rather the two agree at the $1\sigma$ level. Our best estimate of the $EM$ and $T_e$ of the emitting plasma of this observation is $\log_{10} EM=42.98\pm 0.29$ and $\log_{10} T_e=6.12\pm 0.07$ when we include all lines except those from N [v]{} and O [vi]{}, and $\log_{10} EM=43.02\pm 0.29$ and $\log_{10} T_e=6.13\pm 0.06$ when we additionally exclude all Li- and Na-like lines. There remains variation in the crossing points of the $EM$ vs. $T_e$ curves that are suggestive of errors in the atomic data, the observations, or the solar physics model used. However, from the results of the present work it is not possible to say where the source of these errors lie. Further improvements to the atomic database and new insight into the physical conditions of the upper solar atmosphere are needed before these questions can be answered. Given the evidence for regions of differing temperature, a DEM analysis might go some way to resolving the discrepancies found in the present paper. We thank H. Bruhns, H. Kreckel, J. M. Laming, and M. Lestinsky for stimulating discussions. We also thank K. P. Dere for providing corrected versions of his EII rate coefficients and for discussions thereon. CHIANTI is a collaborative project involving the NRL (USA), RAL (UK), MSSL (UK), the Universities of Florence (Italy) and Cambridge (UK), and George Mason University (USA). P.B. and D.W.S. were supported in part by the NASA Solar and Heliospheric Physics Supporting Research and Technology program and the NASA Astronomy and Physics Research and Analysis Program. E.L. was supported by NASA grants NNG06EA14I and NNH06CD24 as well as other NASA grants. [JUNK]{} Altun, Z., Yumak, A., Yavuz, I., Badnell, N. R., Loch, S. D., and Pindzola, M. S. 2007, 474, 1051 Arnaud, M., & Rothenflug, R. 1985, , 60, 425 Aschwanden, M. J. 2004, Physics of the Solar Corona (Chichester: Praxis Publishing) Badnell, N. R., et al. 2003, , 406, 1151 Badnell, N. R. 2006a, , 167, 334 Badnell, N. R. 2006b, http://amdpp.phys.strath.ac.uk/tamoc/DR/ Badnell, N. R. 2006c, http://amdpp.phys.strath.ac.uk/tamoc/RR/ Badnell, N. R. 2006d, J. Phys. B, 39, 4825 Badnell, N. R. 2006e, , 651, L73 Bryans, P., Badnell, N. R., Gorczyca, T. W., Laming, J. M., Witthumsiri, W., & Savin, D. W. 2006, , 167, 343 Dere, K. P., Landi, E., Mason, H. E., Monsignori Fossi, B. C., & Young, P. R. 1997, , 125, 149 Dere, K. P., Landi, E., Del Zanna, G., & Young, P. R. 2001, , 134, 331 Dere, K. P. 2007, , 466, 771 Doyle, J. G., Summers, H. P., & Bryans, P. 2005, , 430, L29 Dragoset, R. A., Musgrove, A., Clark, C. W., & Martin, W. C. 2001, http://physics.nist.gov/PhysRefData/PerTable/index.html Dupree, A. K. 1972, , 178, 527 Feldman, U., & Laming, J. M. 2000, Phys. Scr., 61, 222 Feldman, U., & Widing, K. G. 2003, Space Sci. Rev., 107, 665 Feldman, U., Schühle, U., Widing, K. G., & Laming, J. M. 1998, , 505, 999 Feldman, U., Doschek, G. A., Schühle, U., & Wilhelm, K. 1999, , 518, 500 Forbes, T. G. 2000, Phil. Trans. Roy. Soc. A, 358, 711 Foukal, P. V. 2004, Solar Astrophysics (2nd, Revised Edition; Weinheim: Wiley-VCH) Gu, M. F. 2003a, , 589, 1085 Gu, M. F. 2003b, , 590, 1131 Gu, M. F. 2004, , 153, 389 Gudiksen, B. V., & Norlund, Å 2005 , 618, 1020 Hundhausen, A. J. 1993, , 98, 13177 Kingdon, J. B., & Ferland, G. J. 1996, , 106, 205 Klimchuk, J. A. 2006, Sol. Phys., 234, 41 Laming, J. M. 2004, , 614, 1063 Laming, J. M. 2008, , submitted Landi, E., Landini, M., Dere, K. P., Young, P. R., & Mason, H. E. 1999, , 135, 339 Landi, E., Feldman, U., & Dere, K. P. 2002, , 139, 281 Landi, E., et al. 2006, , 162, 261 Lang, J., Mason, H. E., & McWhirter, R. W. P. 1990, Sol. Phys., 129, 31 Lanzafame, A. C., Brooks, D. H., Lang, J., Summers, H. P., Thomas, R. J., & Thompson, A. M. 2002, , 384, 242 Mandrini, C. H., Démoulin, P., & Klimchuk, J. A. 2000, , 530, 999 Mattioli, M., et al. 2007, J. Phys. B, 40, 3569 Mazzotta, P., Mazzitelli, G., Colafrancesco, S., & Vittorio, N. 1998, , 133, 403 Priest, E. R. & Forbes, T. G. 2002, , 10, 313 Raymond, J. & Doyle, J. G. 1981, , 247, 686 Schwadron, N. A., Fisk, L. A., & Zurbuchen, T. H. 1999, , 521, 859 Summers, H. P., et al. 2006, Plasma Phys. Contr. Fusion, 48, 263 Suno, H., & Kato, T. 2006, ADNDT, 92, 407 Tandberg-Hanssen, E. & Emslie, A. G. 1988, The Physics of Solar Flares (Cambridge: Cambridge University Press) Wilhelm, K., et al. 1995, Sol. Phys., 170, 75 Zirker, J. B. 1993, Sol. Phys., 148, 43 [cccccc]{} IIa & N [v]{} &Li & 1238.82 & $2s\:^2S_{1/2}-2p\:^2P_{3/2}$ &2.520\ IIa & N [v]{} &Li & 1242.80 & $2s\:^2S_{1/2}-2p\:^2P_{1/2}$ &1.420\ IIa & O [vi]{} &Li & 1031.91 & $2s\:^2S_{1/2}-2p\:^2P_{3/2}$ &63.000\ IIa & O [vi]{} &Li & 1037.62 & $2s\:^2S_{1/2}-2p\:^2P_{1/2}$ &28.500\ IIa & Ne [viii]{} &Li & 770.41 & $2s\:^2S_{1/2}-2p\:^2P_{3/2}$ &30.700\ IIa & Ne [viii]{} &Li & 780.32 & $2s\:^2S_{1/2}-2p\:^2P_{1/2}$ &14.200\ IIb & Ne [vii]{} &Be & 895.17 & $2s^2\:^1S_{0}-2s2p\:^3P_{1}$ &0.132\ III & Ne [vii]{} &Be & 973.33 & $2s2p\:^1P_{1}-2p^2\:^1D_{2}$ &0.070\ IIa & Na [ix]{} &Li & 681.72 & $2s\:^2S_{1/2}-2p\:^2P_{3/2}$ &4.515\ IIa & Na [ix]{} &Li & 694.13 & $2s\:^2S_{1/2}-2p\:^2P_{1/2}$ &2.600\ IIb & Na [viii]{} &Be & 789.78 & $2s^2\:^1S_{0}-2s2p\:^3P_{1}$ &0.074\ III & Na [viii]{} &Be & 847.91 & $2s2p\:^1P_{1}-2p^2\:^1D_{2}$ &0.058\ IIa & Mg [x ]{} &Li & 609.79 & $2s\:^2S_{1/2}-2p\:^2P_{3/2}$ &153.000\ IIa & Mg [x ]{} &Li & 624.94 & $2s\:^2S_{1/2}-2p\:^2P_{1/2}$ &91.700\ IIb & Mg [ix]{} &Be & 693.98 & $2s^2\:^1S_{0}-2s2p\:^3P_{2}$ &0.898\ IIb & Mg [ix]{} &Be & 706.06 & $2s^2\:^1S_{0}-2s2p\:^3P_{1}$ &8.160\ III & Mg [ix ]{} &Be & 749.55 & $2s2p\:^1P_{1}-2p^2\:^1D_{2}$ &1.490\ IIb & Mg [viii ]{} &B &762.66 & $2s^22p\:^2P_{1/2}-2s2p^2\:^4P_{3/2}$ &0.047\ IIb & Mg [viii ]{} &B &769.38 & $2s^22p\:^2P_{1/2}-2s2p^2\:^4P_{1/2}$ &0.152\ IIb & Mg [viii ]{} &B &772.28 & $2s^22p\:^2P_{3/2}-2s2p^2\:^4P_{5/2}$ &0.670\ IIb & Mg [viii ]{} &B &782.36 & $2s^22p\:^2P_{3/2}-2s2p^2\:^4P_{3/2}$ &0.357\ IIb & Mg [viii]{} &B &789.43 & $2s^22p\:^2P_{3/2}-2s2p^2\:^4P_{1/2}$ &0.099\ IIb & Mg [vii]{} &C &868.11 & $2s^22p^2\:^3P_{2}-2s2p^3\:^5S_{2}$ &0.048\ IIa & Al [xi ]{} &Li & 549.98 & $2s\:^2S_{1/2}-2p\:^2P_{3/2}$ &7.820\ IIa & Al [xi ]{} &Li & 568.18 & $2s\:^2S_{1/2}-2p\:^2P_{1/2}$ &5.050\ IIb & Al [x ]{} &Be & 637.76 & $2s^2\:^1S_{0}-2s2p\:^3P_{1}$ &2.070\ III & Al [x ]{} &Be & 670.01 & $2s2p\:^1P_{1}-2p^2\:^1D_{2}$ &0.265\ IIb & Al [ix ]{} &B &688.25 & $2s^22p\:^2P_{1/2}-2s2p^2\:^4P_{1/2}$ &0.076\ IIb & Al [ix ]{} &B &691.54 & $2s^22p\:^2P_{3/2}-2s2p^2\:^4P_{5/2}$ &0.441\ IIb & Al [ix ]{} &B &703.65 & $2s^22p\:^2P_{3/2}-2s2p^2\:^4P_{3/2}$ &0.205\ IIb & Al [ix ]{} &B &712.23 & $2s^22p\:^2P_{3/2}-2s2p^2\:^4P_{1/2}$ &0.058\ IIb & Al [viii ]{} &C &756.70 & $2s^22p^2\:^3P_{1}-2s2p^3\:^5S_{2}$ &0.036\ IIb & Al [viii]{} &C &772.54 & $2s^22p^2\:^3P_{2}-2s2p^3\:^5S_{2}$ &0.055\ Ib &Al [vii ]{} &N &1053.84 & $2s^22p^3\:^4S_{3/2}-2s^22p^3\:^2P_{3/2}$ &0.017\ Ia &Al [viii ]{} &C &1057.85 & $2s^22p^2\:^3P_{1}-2s^22p^2\:^1S_{0}$ &0.036\ IIa & Si [xii ]{} &Li & 499.40 & $2s\:^2S_{1/2}-2p\:^2P_{3/2}$ &19.200\ IIa & Si [xii ]{} &Li & 520.67 & $2s\:^2S_{1/2}-2p\:^2P_{1/2}$ &9.160\ III & Si [x ]{} &B &551.18 & $2s2p^2\:^2P_{3/2}-2p^3\:^2D_{5/2}$ &0.200\ IIb & Si [xi ]{} &Be & 564.02 & $2s^2\:^1S_{0}-2s2p\:^3P_{2}$ &1.110\ IIb & Si [xi ]{} &Be & 580.91 & $2s^2\:^1S_{0}-2s2p\:^3P_{1}$ &16.000\ III & Si [xi ]{} &Be & 604.15 & $2s2p\:^1P_{1}-2p^2\:^1D_{2}$ &1.840\ IIb & Si [x ]{} &B &611.60 & $2s^22p\:^2P_{1/2}-2s2p^2\:^4P_{3/2}$ &0.553\ IIb & Si [x ]{} &B &624.70 & $2s^22p\:^2P_{3/2}-2s2p^2\:^4P_{5/2}$ &6.970\ IIb & Si [x ]{} &B &638.94 & $2s^22p\:^2P_{3/2}-2s2p^2\:^4P_{3/2}$ &5.210\ IIb & Si [x ]{} &B &649.19 & $2s^22p\:^2P_{3/2}-2s2p^2\:^4P_{1/2}$ &1.510\ IIb & Si [ix ]{} &C &676.50 & $2s^22p^2\:^3P_{1}-2s2p^3\:^5S_{2}$ &1.560\ IIb & Si [ix ]{} &C &694.70 & $2s^22p^2\:^3P_{2}-2s2p^3\:^5S_{2}$ &3.550\ Ib &Si [viii ]{} &N &944.38 & $2s^22p^3\:^4S_{3/2}-2s^22p^3\:^2P_{3/2}$ &2.030\ Ib &Si [viii ]{} &N &949.22 & $2s^22p^3\:^4S_{3/2}-2s^22p^3\:^2P_{1/2}$ &0.870\ Ia &Si [ix ]{} &C &950.14 & $2s^22p^2\:^3P_{1}-2s^22p^2\:^1S_{0}$ &1.920\ Ia &Si [vii ]{} &O &1049.22 & $2s^22p^4\:^3P_{1}-2s^22p^4\:^1S_{0}$ &0.045\ Ib &Si [viii ]{} &N &1440.49 & $2s^22p^3\:^4S_{3/2}-2s^22p^3\:^2D_{5/2}$ &0.176\ Ib &Si [viii ]{} &N &1445.76 & $2s^22p^3\:^4S_{3/2}-2s^22p^3\:^2D_{3/2}$ &2.090\ IIb & S [xi ]{} &C &552.12 & $2s^22p^2\:^3P_{1}-2s2p^3\:^5S_{2}$ &0.190\ IIb & S [xi ]{} &C &574.89 & $2s^22p^2\:^3P_{2}-2s2p^3\:^5S_{2}$ &0.470\ Ib &S [x ]{} &N &776.25 & $2s^22p^3\:^4S_{3/2}-2s^22p^3\:^2P_{3/2}$ &1.910\ Ia &S [xi ]{} &C &782.96 & $2s^22p^2\:^3P_{1}-2s^22p^2\:^1S_{0}$ &0.4005\ Ib &S [x ]{} &N &787.56 & $2s^22p^3\:^4S_{3/2}-2s^22p^3\:^2P_{1/2}$ &0.946\ Ia &S [ix ]{} &O &871.73 & $2s^22p^4\:^3P_{1}-2s^22p^4\:^1S_{0}$ &0.440\ Ib &S [x ]{} &N &1196.26 & $2s^22p^3\:^4S_{3/2}-2s^22p^3\:^2D_{5/2}$ &1.590\ Ib &S [x ]{} &N &1212.93 & $2s^22p^3\:^4S_{3/2}-2s^22p^3\:^2D_{3/2}$ &3.410\ IIa & Ar [viii]{} &Na & 700.25 & $3s\:^2S_{1/2}-3p\:^2P_{3/2}$ &0.635\ IIa & Ar [viii]{} &Na & 713.81 & $3s\:^2S_{1/2}-3p\:^2P_{1/2}$ &0.310\ Ib &Ar [xii ]{} &N &1018.89 & $2s^22p^3\:^4S_{3/2}-2s^22p^3\:^2D_{5/2}$ &0.274\ Ib &Ar [xii ]{} &N &1054.57 & $2s^22p^3\:^4S_{3/2}-2s^22p^3\:^2D_{3/2}$ &0.056\ Ia &Ar [xi ]{} &O &1392.11 & $2s^22p^4\:^3P_{2}-2s^22p^4\:^1D_{2}$ &0.134\ IIa & K [ix ]{} &Na & 636.29 & $3s\:^2S_{1/2}-3p\:^2P_{1/2}$ &0.133\ IIa & Ca [x ]{} &Na & 557.76 & $3s\:^2S_{1/2}-3p\:^2P_{3/2}$ &8.000\ IIa & Ca [x ]{} &Na & 574.00 & $3s\:^2S_{1/2}-3p\:^2P_{1/2}$ &4.880\ IIb & Ca [ix ]{} &Mg & 691.41 & $3s^2\:^1S_{0}-3s3p\:^3P_{1}$ &0.109\ III & Ca [ix ]{} &Mg & 821.23 & $3s3p\:^1P_{1}-3p^2\:^1D_{2}$ &0.048\ Ia &Fe [xii ]{} &P &1242.00 & $3s^23p^3\:^4S_{3/2}-3s^23p^3\:^2P_{3/2}$ &11.840\ Ia &Fe [xii ]{} &P &1349.36 & $3s^23p^3\:^4S_{3/2}-3s^23p^3\:^2P_{1/2}$ &5.353\ Ia &Fe [xi]{} &S &1467.06 & $3s^23p^4\:^3P_{1}-3s^23p^4\:^1S_{0}$ &3.390 [ccc]{} 4.00 & 0.001 & 2.751\ 4.10 & 0.026 & 1.240\ 4.20 & 0.318 & 0.285\ 4.30 & 1.089 & 0.037\ 4.40 & 1.895 & 0.006\ 4.50 & 2.584 & 0.001\ 4.60 & 3.162 & 0.000\ 4.70 & 3.649 & 0.000\ 4.80 & 4.064 & 0.000\ 4.90 & 4.420 & 0.000\ 5.00 & 4.730 & 0.000\ 5.10 & 5.002 & 0.000\ 5.20 & 5.244 & 0.000\ 5.30 & 5.461 & 0.000\ 5.40 & 5.658 & 0.000\ 5.50 & 5.838 & 0.000\ 5.60 & 6.005 & 0.000\ 5.70 & 6.160 & 0.000\ 5.80 & 6.306 & 0.000\ 5.90 & 6.445 & 0.000\ 6.00 & 6.577 & 0.000\ 6.10 & 6.703 & 0.000\ 6.20 & 6.824 & 0.000\ 6.30 & 6.942 & 0.000\ 6.40 & 7.056 & 0.000\ 6.50 & 7.168 & 0.000\ 6.60 & 7.277 & 0.000\ 6.70 & 7.384 & 0.000\ 6.80 & 7.490 & 0.000\ 6.90 & 7.594 & 0.000\ 7.00 & 7.699 & 0.000\ 7.10 & 7.803 & 0.000\ 7.20 & 7.906 & 0.000\ 7.30 & 8.010 & 0.000\ 7.40 & 8.113 & 0.000\ 7.50 & 8.216 & 0.000\ 7.60 & 8.318 & 0.000\ 7.70 & 8.421 & 0.000\ 7.80 & 8.524 & 0.000\ 7.90 & 8.627 & 0.000\ 8.00 & 8.731 & 0.000\ 8.10 & 8.835 & 0.000\ 8.20 & 8.939 & 0.000\ 8.30 & 9.043 & 0.000\ 8.40 & 9.148 & 0.000\ 8.50 & 9.253 & 0.000\ 8.60 & 9.358 & 0.000\ 8.70 & 9.463 & 0.000\ 8.80 & 9.569 & 0.000\ 8.90 & 9.675 & 0.000\ 9.00 & 9.780 & 0.000\ [cccc]{} 4.00 & 0.000 & 8.813 & 15.000\ 4.10 & 0.000 & 6.135 & 15.000\ 4.20 & 0.000 & 3.980 & 15.000\ 4.30 & 0.003 & 2.240 & 13.153\ 4.40 & 0.061 & 0.885 & 8.981\ 4.50 & 0.496 & 0.167 & 5.896\ 4.60 & 1.295 & 0.023 & 3.845\ 4.70 & 2.043 & 0.006 & 2.288\ 4.80 & 2.650 & 0.040 & 1.071\ 4.90 & 3.265 & 0.295 & 0.307\ 5.00 & 4.052 & 0.884 & 0.061\ 5.10 & 4.827 & 1.527 & 0.013\ 5.20 & 5.511 & 2.092 & 0.004\ 5.30 & 6.116 & 2.571 & 0.001\ 5.40 & 6.658 & 2.979 & 0.000\ 5.50 & 7.150 & 3.329 & 0.000\ 5.60 & 7.603 & 3.633 & 0.000\ 5.70 & 8.022 & 3.899 & 0.000\ 5.80 & 8.413 & 4.135 & 0.000\ 5.90 & 8.779 & 4.346 & 0.000\ 6.00 & 9.125 & 4.538 & 0.000\ 6.10 & 9.452 & 4.713 & 0.000\ 6.20 & 9.764 & 4.874 & 0.000\ 6.30 & 10.061 & 5.023 & 0.000\ 6.40 & 10.346 & 5.164 & 0.000\ 6.50 & 10.621 & 5.296 & 0.000\ 6.60 & 10.886 & 5.422 & 0.000\ 6.70 & 11.143 & 5.542 & 0.000\ 6.80 & 11.392 & 5.658 & 0.000\ 6.90 & 11.636 & 5.770 & 0.000\ 7.00 & 11.874 & 5.880 & 0.000\ 7.10 & 12.109 & 5.988 & 0.000\ 7.20 & 12.341 & 6.095 & 0.000\ 7.30 & 12.570 & 6.201 & 0.000\ 7.40 & 12.796 & 6.306 & 0.000\ 7.50 & 13.020 & 6.410 & 0.000\ 7.60 & 13.242 & 6.514 & 0.000\ 7.70 & 13.462 & 6.617 & 0.000\ 7.80 & 13.681 & 6.721 & 0.000\ 7.90 & 13.899 & 6.824 & 0.000\ 8.00 & 14.115 & 6.927 & 0.000\ 8.10 & 14.332 & 7.030 & 0.000\ 8.20 & 14.547 & 7.133 & 0.000\ 8.30 & 14.762 & 7.237 & 0.000\ 8.40 & 14.977 & 7.340 & 0.000\ 8.50 & 15.000 & 7.445 & 0.000\ 8.60 & 15.000 & 7.549 & 0.000\ 8.70 & 15.000 & 7.653 & 0.000\ 8.80 & 15.000 & 7.758 & 0.000\ 8.90 & 15.000 & 7.863 & 0.000\ 9.00 & 15.000 & 7.968 & 0.000\ [ccccc]{} 4.00 & 2.408 & 0.002 & 15.000 & 15.000\ 4.10 & 3.095 & 0.000 & 15.000 & 15.000\ 4.20 & 3.667 & 0.000 & 15.000 & 15.000\ 4.30 & 4.147 & 0.000 & 14.841 & 15.000\ 4.40 & 4.553 & 0.000 & 12.035 & 15.000\ 4.50 & 4.897 & 0.000 & 9.228 & 15.000\ 4.60 & 5.192 & 0.000 & 6.625 & 15.000\ 4.70 & 5.445 & 0.000 & 4.531 & 14.616\ 4.80 & 5.663 & 0.001 & 2.841 & 10.276\ 4.90 & 5.861 & 0.014 & 1.486 & 6.790\ 5.00 & 6.141 & 0.157 & 0.517 & 4.106\ 5.10 & 6.694 & 0.644 & 0.115 & 2.317\ 5.20 & 7.294 & 1.253 & 0.060 & 1.137\ 5.30 & 7.857 & 1.868 & 0.234 & 0.395\ 5.40 & 8.542 & 2.609 & 0.691 & 0.100\ 5.50 & 9.263 & 3.367 & 1.237 & 0.026\ 5.60 & 9.936 & 4.050 & 1.735 & 0.008\ 5.70 & 10.558 & 4.654 & 2.163 & 0.003\ 5.80 & 11.139 & 5.195 & 2.531 & 0.001\ 5.90 & 11.689 & 5.685 & 2.849 & 0.001\ 6.00 & 12.210 & 6.132 & 3.127 & 0.000\ 6.10 & 12.706 & 6.545 & 3.371 & 0.000\ 6.20 & 13.179 & 6.929 & 3.589 & 0.000\ 6.30 & 13.631 & 7.287 & 3.784 & 0.000\ 6.40 & 14.066 & 7.624 & 3.960 & 0.000\ 6.50 & 14.486 & 7.944 & 4.122 & 0.000\ 6.60 & 14.894 & 8.248 & 4.272 & 0.000\ 6.70 & 15.000 & 8.538 & 4.410 & 0.000\ 6.80 & 15.000 & 8.815 & 4.539 & 0.000\ 6.90 & 15.000 & 9.081 & 4.659 & 0.000\ 7.00 & 15.000 & 9.338 & 4.773 & 0.000\ 7.10 & 15.000 & 9.587 & 4.883 & 0.000\ 7.20 & 15.000 & 9.832 & 4.990 & 0.000\ 7.30 & 15.000 & 10.072 & 5.097 & 0.000\ 7.40 & 15.000 & 10.308 & 5.202 & 0.000\ 7.50 & 15.000 & 10.541 & 5.306 & 0.000\ 7.60 & 15.000 & 10.771 & 5.409 & 0.000\ 7.70 & 15.000 & 10.997 & 5.512 & 0.000\ 7.80 & 15.000 & 11.221 & 5.613 & 0.000\ 7.90 & 15.000 & 11.443 & 5.714 & 0.000\ 8.00 & 15.000 & 11.664 & 5.815 & 0.000\ 8.10 & 15.000 & 11.883 & 5.916 & 0.000\ 8.20 & 15.000 & 12.101 & 6.017 & 0.000\ 8.30 & 15.000 & 12.319 & 6.119 & 0.000\ 8.40 & 15.000 & 12.536 & 6.221 & 0.000\ 8.50 & 15.000 & 12.752 & 6.324 & 0.000\ 8.60 & 15.000 & 12.968 & 6.427 & 0.000\ 8.70 & 15.000 & 13.184 & 6.530 & 0.000\ 8.80 & 15.000 & 13.399 & 6.634 & 0.000\ 8.90 & 15.000 & 13.615 & 6.738 & 0.000\ 9.00 & 15.000 & 13.830 & 6.843 & 0.000\ [cccccc]{} 4.00 & 0.021 & 1.335 & 6.887 & 15.000 & 15.000\ 4.10 & 0.114 & 0.637 & 4.177 & 15.000 & 15.000\ 4.20 & 0.379 & 0.240 & 2.156 & 15.000 & 15.000\ 4.30 & 0.865 & 0.161 & 0.762 & 15.000 & 15.000\ 4.40 & 1.797 & 0.603 & 0.134 & 15.000 & 15.000\ 4.50 & 2.987 & 1.363 & 0.020 & 15.000 & 15.000\ 4.60 & 4.072 & 2.065 & 0.004 & 15.000 & 15.000\ 4.70 & 5.007 & 2.656 & 0.001 & 12.940 & 15.000\ 4.80 & 5.812 & 3.149 & 0.000 & 10.076 & 15.000\ 4.90 & 6.513 & 3.564 & 0.000 & 7.419 & 15.000\ 5.00 & 7.127 & 3.914 & 0.000 & 5.280 & 14.466\ 5.10 & 7.671 & 4.212 & 0.000 & 3.552 & 10.362\ 5.20 & 8.149 & 4.461 & 0.003 & 2.155 & 7.053\ 5.30 & 8.577 & 4.673 & 0.040 & 1.058 & 4.411\ 5.40 & 9.064 & 4.957 & 0.248 & 0.365 & 2.467\ 5.50 & 9.683 & 5.384 & 0.688 & 0.134 & 1.217\ 5.60 & 10.367 & 5.887 & 1.239 & 0.220 & 0.469\ 5.70 & 11.187 & 6.535 & 1.927 & 0.578 & 0.140\ 5.80 & 12.066 & 7.250 & 2.658 & 1.048 & 0.042\ 5.90 & 12.909 & 7.938 & 3.334 & 1.496 & 0.014\ 6.00 & 13.698 & 8.576 & 3.936 & 1.888 & 0.006\ 6.10 & 14.437 & 9.170 & 4.473 & 2.226 & 0.003\ 6.20 & 15.000 & 9.726 & 4.956 & 2.519 & 0.001\ 6.30 & 15.000 & 10.249 & 5.394 & 2.775 & 0.001\ 6.40 & 15.000 & 10.745 & 5.795 & 3.001 & 0.000\ 6.50 & 15.000 & 11.215 & 6.164 & 3.202 & 0.000\ 6.60 & 15.000 & 11.664 & 6.506 & 3.385 & 0.000\ 6.70 & 15.000 & 12.093 & 6.827 & 3.551 & 0.000\ 6.80 & 15.000 & 12.505 & 7.128 & 3.703 & 0.000\ 6.90 & 15.000 & 12.902 & 7.412 & 3.846 & 0.000\ 7.00 & 15.000 & 13.287 & 7.684 & 3.979 & 0.000\ 7.10 & 15.000 & 13.660 & 7.944 & 4.105 & 0.000\ 7.20 & 15.000 & 14.024 & 8.194 & 4.225 & 0.000\ 7.30 & 15.000 & 14.380 & 8.436 & 4.339 & 0.000\ 7.40 & 15.000 & 14.730 & 8.673 & 4.449 & 0.000\ 7.50 & 15.000 & 15.000 & 8.904 & 4.556 & 0.000\ 7.60 & 15.000 & 15.000 & 9.132 & 4.660 & 0.000\ 7.70 & 15.000 & 15.000 & 9.357 & 4.764 & 0.000\ 7.80 & 15.000 & 15.000 & 9.580 & 4.867 & 0.000\ 7.90 & 15.000 & 15.000 & 9.800 & 4.970 & 0.000\ 8.00 & 15.000 & 15.000 & 10.018 & 5.072 & 0.000\ 8.10 & 15.000 & 15.000 & 10.235 & 5.174 & 0.000\ 8.20 & 15.000 & 15.000 & 10.451 & 5.275 & 0.000\ 8.30 & 15.000 & 15.000 & 10.666 & 5.377 & 0.000\ 8.40 & 15.000 & 15.000 & 10.881 & 5.478 & 0.000\ 8.50 & 15.000 & 15.000 & 11.095 & 5.579 & 0.000\ 8.60 & 15.000 & 15.000 & 11.310 & 5.681 & 0.000\ 8.70 & 15.000 & 15.000 & 11.524 & 5.783 & 0.000\ 8.80 & 15.000 & 15.000 & 11.739 & 5.886 & 0.000\ 8.90 & 15.000 & 15.000 & 11.953 & 5.989 & 0.000\ 9.00 & 15.000 & 15.000 & 12.168 & 6.092 & 0.000\ [ccccccc]{} 4.00 & 0.382 & 0.233 & 9.957 & 15.000 & 15.000 & 15.000\ 4.10 & 1.029 & 0.043 & 7.367 & 15.000 & 15.000 & 15.000\ 4.20 & 1.516 & 0.013 & 5.445 & 14.843 & 15.000 & 15.000\ 4.30 & 1.822 & 0.007 & 3.913 & 10.708 & 15.000 & 15.000\ 4.40 & 2.052 & 0.005 & 2.668 & 7.368 & 15.000 & 15.000\ 4.50 & 2.265 & 0.012 & 1.653 & 4.664 & 15.000 & 15.000\ 4.60 & 2.518 & 0.068 & 0.859 & 2.503 & 15.000 & 15.000\ 4.70 & 2.964 & 0.318 & 0.397 & 0.930 & 15.000 & 15.000\ 4.80 & 3.929 & 1.088 & 0.565 & 0.190 & 15.000 & 15.000\ 4.90 & 5.229 & 2.193 & 1.154 & 0.035 & 13.989 & 15.000\ 5.00 & 6.458 & 3.228 & 1.743 & 0.008 & 11.241 & 15.000\ 5.10 & 7.546 & 4.124 & 2.249 & 0.002 & 8.543 & 15.000\ 5.20 & 8.508 & 4.897 & 2.677 & 0.001 & 6.275 & 15.000\ 5.30 & 9.365 & 5.568 & 3.037 & 0.000 & 4.446 & 11.514\ 5.40 & 10.127 & 6.148 & 3.337 & 0.001 & 2.965 & 8.145\ 5.50 & 10.791 & 6.635 & 3.566 & 0.008 & 1.772 & 5.428\ 5.60 & 11.371 & 7.043 & 3.736 & 0.063 & 0.869 & 3.289\ 5.70 & 11.980 & 7.486 & 3.957 & 0.273 & 0.347 & 1.761\ 5.80 & 12.702 & 8.046 & 4.310 & 0.662 & 0.206 & 0.796\ 5.90 & 13.574 & 8.762 & 4.831 & 1.216 & 0.375 & 0.286\ 6.00 & 14.572 & 9.609 & 5.496 & 1.889 & 0.744 & 0.093\ 6.10 & 15.000 & 10.469 & 6.186 & 2.559 & 1.154 & 0.033\ 6.20 & 15.000 & 11.288 & 6.842 & 3.170 & 1.531 & 0.013\ 6.30 & 15.000 & 12.056 & 7.456 & 3.718 & 1.861 & 0.006\ 6.40 & 15.000 & 12.780 & 8.033 & 4.212 & 2.150 & 0.003\ 6.50 & 15.000 & 13.465 & 8.575 & 4.659 & 2.403 & 0.002\ 6.60 & 15.000 & 14.116 & 9.088 & 5.067 & 2.627 & 0.001\ 6.70 & 15.000 & 14.735 & 9.574 & 5.442 & 2.827 & 0.001\ 6.80 & 15.000 & 15.000 & 10.038 & 5.790 & 3.008 & 0.000\ 6.90 & 15.000 & 15.000 & 10.481 & 6.114 & 3.173 & 0.000\ 7.00 & 15.000 & 15.000 & 10.905 & 6.418 & 3.325 & 0.000\ 7.10 & 15.000 & 15.000 & 11.312 & 6.705 & 3.466 & 0.000\ 7.20 & 15.000 & 15.000 & 11.706 & 6.978 & 3.599 & 0.000\ 7.30 & 15.000 & 15.000 & 12.088 & 7.239 & 3.725 & 0.000\ 7.40 & 15.000 & 15.000 & 12.459 & 7.490 & 3.844 & 0.000\ 7.50 & 15.000 & 15.000 & 12.822 & 7.733 & 3.958 & 0.000\ 7.60 & 15.000 & 15.000 & 13.178 & 7.970 & 4.068 & 0.000\ 7.70 & 15.000 & 15.000 & 13.529 & 8.201 & 4.174 & 0.000\ 7.80 & 15.000 & 15.000 & 13.874 & 8.429 & 4.279 & 0.000\ 7.90 & 15.000 & 15.000 & 14.216 & 8.654 & 4.382 & 0.000\ 8.00 & 15.000 & 15.000 & 14.554 & 8.877 & 4.486 & 0.000\ 8.10 & 15.000 & 15.000 & 14.889 & 9.097 & 4.588 & 0.000\ 8.20 & 15.000 & 15.000 & 15.000 & 9.315 & 4.691 & 0.000\ 8.30 & 15.000 & 15.000 & 15.000 & 9.532 & 4.792 & 0.000\ 8.40 & 15.000 & 15.000 & 15.000 & 9.747 & 4.894 & 0.000\ 8.50 & 15.000 & 15.000 & 15.000 & 9.962 & 4.995 & 0.000\ 8.60 & 15.000 & 15.000 & 15.000 & 10.176 & 5.096 & 0.000\ 8.70 & 15.000 & 15.000 & 15.000 & 10.390 & 5.198 & 0.000\ 8.80 & 15.000 & 15.000 & 15.000 & 10.604 & 5.300 & 0.000\ 8.90 & 15.000 & 15.000 & 15.000 & 10.819 & 5.402 & 0.000\ 9.00 & 15.000 & 15.000 & 15.000 & 11.033 & 5.504 & 0.000\ [cccccccc]{} 4.00 & 0.014 & 1.487 & 10.947 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 0.203 & 0.428 & 7.258 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 0.795 & 0.076 & 4.809 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 1.369 & 0.019 & 3.125 & 13.250 & 15.000 & 15.000 & 15.000\ 4.40 & 1.768 & 0.013 & 1.918 & 9.615 & 15.000 & 15.000 & 15.000\ 4.50 & 2.099 & 0.043 & 1.068 & 6.842 & 14.727 & 15.000 & 15.000\ 4.60 & 2.476 & 0.164 & 0.508 & 4.733 & 10.382 & 15.000 & 15.000\ 4.70 & 2.970 & 0.423 & 0.207 & 3.170 & 7.020 & 15.000 & 15.000\ 4.80 & 3.546 & 0.778 & 0.085 & 2.010 & 4.404 & 15.000 & 15.000\ 4.90 & 4.158 & 1.177 & 0.068 & 1.131 & 2.346 & 15.000 & 15.000\ 5.00 & 4.905 & 1.717 & 0.241 & 0.583 & 0.836 & 15.000 & 15.000\ 5.10 & 6.158 & 2.769 & 0.965 & 0.696 & 0.162 & 13.909 & 15.000\ 5.20 & 7.692 & 4.106 & 2.006 & 1.214 & 0.032 & 10.883 & 15.000\ 5.30 & 9.120 & 5.342 & 2.971 & 1.730 & 0.009 & 8.174 & 15.000\ 5.40 & 10.396 & 6.429 & 3.808 & 2.174 & 0.003 & 6.010 & 14.630\ 5.50 & 11.535 & 7.383 & 4.530 & 2.550 & 0.001 & 4.269 & 10.752\ 5.60 & 12.549 & 8.217 & 5.147 & 2.859 & 0.001 & 2.864 & 7.625\ 5.70 & 13.434 & 8.927 & 5.655 & 3.089 & 0.008 & 1.739 & 5.108\ 5.80 & 14.211 & 9.533 & 6.070 & 3.254 & 0.060 & 0.895 & 3.135\ 5.90 & 14.984 & 10.139 & 6.495 & 3.453 & 0.238 & 0.395 & 1.715\ 6.00 & 15.000 & 10.838 & 7.020 & 3.773 & 0.578 & 0.238 & 0.803\ 6.10 & 15.000 & 11.677 & 7.693 & 4.260 & 1.081 & 0.369 & 0.310\ 6.20 & 15.000 & 12.629 & 8.485 & 4.882 & 1.699 & 0.688 & 0.111\ 6.30 & 15.000 & 13.591 & 9.294 & 5.534 & 2.325 & 1.055 & 0.042\ 6.40 & 15.000 & 14.512 & 10.067 & 6.162 & 2.904 & 1.401 & 0.018\ 6.50 & 15.000 & 15.000 & 10.796 & 6.753 & 3.428 & 1.707 & 0.009\ 6.60 & 15.000 & 15.000 & 11.486 & 7.309 & 3.902 & 1.977 & 0.005\ 6.70 & 15.000 & 15.000 & 12.141 & 7.835 & 4.336 & 2.215 & 0.003\ 6.80 & 15.000 & 15.000 & 12.768 & 8.335 & 4.734 & 2.427 & 0.002\ 6.90 & 15.000 & 15.000 & 13.370 & 8.812 & 5.102 & 2.617 & 0.001\ 7.00 & 15.000 & 15.000 & 13.951 & 9.268 & 5.446 & 2.791 & 0.001\ 7.10 & 15.000 & 15.000 & 14.512 & 9.707 & 5.770 & 2.950 & 0.000\ 7.20 & 15.000 & 15.000 & 15.000 & 10.131 & 6.076 & 3.099 & 0.000\ 7.30 & 15.000 & 15.000 & 15.000 & 10.541 & 6.367 & 3.239 & 0.000\ 7.40 & 15.000 & 15.000 & 15.000 & 10.940 & 6.647 & 3.373 & 0.000\ 7.50 & 15.000 & 15.000 & 15.000 & 11.330 & 6.918 & 3.502 & 0.000\ 7.60 & 15.000 & 15.000 & 15.000 & 11.712 & 7.182 & 3.629 & 0.000\ 7.70 & 15.000 & 15.000 & 15.000 & 12.089 & 7.440 & 3.753 & 0.000\ 7.80 & 15.000 & 15.000 & 15.000 & 12.462 & 7.695 & 3.878 & 0.000\ 7.90 & 15.000 & 15.000 & 15.000 & 12.832 & 7.947 & 4.002 & 0.000\ 8.00 & 15.000 & 15.000 & 15.000 & 13.200 & 8.199 & 4.128 & 0.000\ 8.10 & 15.000 & 15.000 & 15.000 & 13.568 & 8.450 & 4.257 & 0.000\ 8.20 & 15.000 & 15.000 & 15.000 & 13.935 & 8.702 & 4.388 & 0.000\ 8.30 & 15.000 & 15.000 & 15.000 & 14.304 & 8.956 & 4.523 & 0.000\ 8.40 & 15.000 & 15.000 & 15.000 & 14.673 & 9.211 & 4.661 & 0.000\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 9.470 & 4.804 & 0.000\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 9.731 & 4.951 & 0.000\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 9.996 & 5.102 & 0.000\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 10.265 & 5.258 & 0.000\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 10.537 & 5.419 & 0.000\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 10.813 & 5.584 & 0.000\ [ccccccccc]{} 4.00 & 0.000 & 3.390 & 14.613 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 0.008 & 1.746 & 10.223 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 0.140 & 0.559 & 6.507 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 0.680 & 0.102 & 4.089 & 13.747 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 1.290 & 0.024 & 2.554 & 9.663 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 1.758 & 0.023 & 1.476 & 6.596 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 2.180 & 0.095 & 0.722 & 4.316 & 12.687 & 15.000 & 15.000 & 15.000\ 4.70 & 2.690 & 0.318 & 0.288 & 2.706 & 9.065 & 15.000 & 15.000 & 15.000\ 4.80 & 3.304 & 0.681 & 0.116 & 1.605 & 6.337 & 11.980 & 15.000 & 15.000\ 4.90 & 3.974 & 1.121 & 0.107 & 0.844 & 4.249 & 8.163 & 15.000 & 15.000\ 5.00 & 4.725 & 1.655 & 0.259 & 0.372 & 2.685 & 5.202 & 15.000 & 15.000\ 5.10 & 5.580 & 2.302 & 0.571 & 0.156 & 1.565 & 2.948 & 15.000 & 15.000\ 5.20 & 6.535 & 3.054 & 1.023 & 0.157 & 0.811 & 1.269 & 15.000 & 15.000\ 5.30 & 7.836 & 4.156 & 1.854 & 0.595 & 0.612 & 0.312 & 12.862 & 15.000\ 5.40 & 9.546 & 5.672 & 3.118 & 1.513 & 0.988 & 0.063 & 9.758 & 15.000\ 5.50 & 11.219 & 7.154 & 4.366 & 2.452 & 1.462 & 0.017 & 7.312 & 15.000\ 5.60 & 12.731 & 8.480 & 5.471 & 3.279 & 1.886 & 0.006 & 5.370 & 12.795\ 5.70 & 14.087 & 9.655 & 6.436 & 3.990 & 2.244 & 0.003 & 3.809 & 9.381\ 5.80 & 15.000 & 10.686 & 7.268 & 4.587 & 2.529 & 0.003 & 2.550 & 6.627\ 5.90 & 15.000 & 11.569 & 7.960 & 5.062 & 2.725 & 0.013 & 1.553 & 4.418\ 6.00 & 15.000 & 12.343 & 8.552 & 5.450 & 2.864 & 0.072 & 0.825 & 2.705\ 6.10 & 15.000 & 13.118 & 9.150 & 5.858 & 3.047 & 0.243 & 0.404 & 1.476\ 6.20 & 15.000 & 13.984 & 9.848 & 6.376 & 3.364 & 0.566 & 0.281 & 0.689\ 6.30 & 15.000 & 14.988 & 10.690 & 7.049 & 3.855 & 1.052 & 0.419 & 0.275\ 6.40 & 15.000 & 15.000 & 11.624 & 7.823 & 4.464 & 1.635 & 0.712 & 0.106\ 6.50 & 15.000 & 15.000 & 12.558 & 8.605 & 5.095 & 2.220 & 1.040 & 0.044\ 6.60 & 15.000 & 15.000 & 13.453 & 9.354 & 5.702 & 2.762 & 1.349 & 0.021\ 6.70 & 15.000 & 15.000 & 14.303 & 10.064 & 6.277 & 3.257 & 1.627 & 0.011\ 6.80 & 15.000 & 15.000 & 15.000 & 10.737 & 6.819 & 3.707 & 1.873 & 0.006\ 6.90 & 15.000 & 15.000 & 15.000 & 11.379 & 7.334 & 4.121 & 2.092 & 0.004\ 7.00 & 15.000 & 15.000 & 15.000 & 11.994 & 7.823 & 4.502 & 2.288 & 0.002\ 7.10 & 15.000 & 15.000 & 15.000 & 12.586 & 8.290 & 4.856 & 2.467 & 0.001\ 7.20 & 15.000 & 15.000 & 15.000 & 13.157 & 8.739 & 5.188 & 2.630 & 0.001\ 7.30 & 15.000 & 15.000 & 15.000 & 13.711 & 9.170 & 5.501 & 2.782 & 0.001\ 7.40 & 15.000 & 15.000 & 15.000 & 14.248 & 9.587 & 5.798 & 2.925 & 0.001\ 7.50 & 15.000 & 15.000 & 15.000 & 14.773 & 9.991 & 6.082 & 3.061 & 0.000\ 7.60 & 15.000 & 15.000 & 15.000 & 15.000 & 10.385 & 6.356 & 3.191 & 0.000\ 7.70 & 15.000 & 15.000 & 15.000 & 15.000 & 10.771 & 6.622 & 3.318 & 0.000\ 7.80 & 15.000 & 15.000 & 15.000 & 15.000 & 11.150 & 6.882 & 3.443 & 0.000\ 7.90 & 15.000 & 15.000 & 15.000 & 15.000 & 11.524 & 7.137 & 3.568 & 0.000\ 8.00 & 15.000 & 15.000 & 15.000 & 15.000 & 11.895 & 7.390 & 3.692 & 0.000\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 12.264 & 7.641 & 3.818 & 0.000\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 12.631 & 7.892 & 3.945 & 0.000\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 12.998 & 8.143 & 4.076 & 0.000\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 13.366 & 8.395 & 4.209 & 0.000\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 13.735 & 8.649 & 4.346 & 0.000\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 14.106 & 8.906 & 4.487 & 0.000\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 14.479 & 9.166 & 4.633 & 0.000\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 14.855 & 9.429 & 4.783 & 0.000\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 9.696 & 4.937 & 0.000\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 9.966 & 5.096 & 0.000\ [cccccccccc]{} 4.00 & 0.000 & 2.959 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 0.017 & 1.409 & 12.192 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 0.244 & 0.367 & 8.276 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 0.910 & 0.057 & 5.558 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 1.510 & 0.014 & 3.616 & 11.944 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 1.890 & 0.009 & 2.183 & 8.206 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 2.188 & 0.033 & 1.173 & 5.441 & 13.325 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 2.550 & 0.158 & 0.521 & 3.452 & 9.310 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 3.070 & 0.453 & 0.195 & 2.077 & 6.346 & 14.286 & 15.000 & 15.000 & 15.000\ 4.90 & 3.714 & 0.876 & 0.100 & 1.135 & 4.149 & 10.184 & 15.000 & 15.000 & 15.000\ 5.00 & 4.459 & 1.404 & 0.183 & 0.521 & 2.527 & 7.014 & 12.023 & 15.000 & 15.000\ 5.10 & 5.341 & 2.069 & 0.460 & 0.218 & 1.402 & 4.623 & 8.080 & 15.000 & 15.000\ 5.20 & 6.346 & 2.862 & 0.907 & 0.178 & 0.680 & 2.860 & 5.061 & 15.000 & 15.000\ 5.30 & 7.468 & 3.774 & 1.507 & 0.362 & 0.294 & 1.611 & 2.790 & 15.000 & 15.000\ 5.40 & 8.727 & 4.828 & 2.278 & 0.773 & 0.219 & 0.816 & 1.160 & 14.593 & 15.000\ 5.50 & 10.345 & 6.245 & 3.434 & 1.612 & 0.640 & 0.632 & 0.290 & 10.978 & 15.000\ 5.60 & 12.288 & 7.993 & 4.940 & 2.834 & 1.497 & 0.976 & 0.065 & 8.208 & 15.000\ 5.70 & 14.153 & 9.668 & 6.389 & 4.025 & 2.365 & 1.404 & 0.019 & 6.116 & 14.110\ 5.80 & 15.000 & 11.174 & 7.681 & 5.080 & 3.130 & 1.788 & 0.007 & 4.454 & 10.530\ 5.90 & 15.000 & 12.511 & 8.816 & 5.995 & 3.782 & 2.106 & 0.004 & 3.117 & 7.646\ 6.00 & 15.000 & 13.682 & 9.793 & 6.765 & 4.310 & 2.340 & 0.006 & 2.047 & 5.323\ 6.10 & 15.000 & 14.696 & 10.622 & 7.399 & 4.720 & 2.489 & 0.029 & 1.221 & 3.479\ 6.20 & 15.000 & 15.000 & 11.377 & 7.969 & 5.081 & 2.619 & 0.115 & 0.655 & 2.080\ 6.30 & 15.000 & 15.000 & 12.164 & 8.578 & 5.496 & 2.830 & 0.315 & 0.361 & 1.100\ 6.40 & 15.000 & 15.000 & 13.061 & 9.307 & 6.041 & 3.197 & 0.669 & 0.329 & 0.500\ 6.50 & 15.000 & 15.000 & 14.082 & 10.165 & 6.728 & 3.724 & 1.168 & 0.510 & 0.206\ 6.60 & 15.000 & 15.000 & 15.000 & 11.076 & 7.478 & 4.329 & 1.724 & 0.791 & 0.087\ 6.70 & 15.000 & 15.000 & 15.000 & 11.972 & 8.220 & 4.936 & 2.265 & 1.086 & 0.040\ 6.80 & 15.000 & 15.000 & 15.000 & 12.828 & 8.932 & 5.518 & 2.767 & 1.360 & 0.020\ 6.90 & 15.000 & 15.000 & 15.000 & 13.646 & 9.609 & 6.070 & 3.226 & 1.607 & 0.011\ 7.00 & 15.000 & 15.000 & 15.000 & 14.427 & 10.256 & 6.593 & 3.647 & 1.828 & 0.007\ 7.10 & 15.000 & 15.000 & 15.000 & 15.000 & 10.875 & 7.090 & 4.035 & 2.027 & 0.004\ 7.20 & 15.000 & 15.000 & 15.000 & 15.000 & 11.471 & 7.564 & 4.395 & 2.208 & 0.003\ 7.30 & 15.000 & 15.000 & 15.000 & 15.000 & 12.047 & 8.017 & 4.732 & 2.373 & 0.002\ 7.40 & 15.000 & 15.000 & 15.000 & 15.000 & 12.603 & 8.453 & 5.049 & 2.527 & 0.001\ 7.50 & 15.000 & 15.000 & 15.000 & 15.000 & 13.144 & 8.873 & 5.349 & 2.671 & 0.001\ 7.60 & 15.000 & 15.000 & 15.000 & 15.000 & 13.671 & 9.279 & 5.635 & 2.807 & 0.001\ 7.70 & 15.000 & 15.000 & 15.000 & 15.000 & 14.186 & 9.675 & 5.910 & 2.939 & 0.001\ 7.80 & 15.000 & 15.000 & 15.000 & 15.000 & 14.691 & 10.062 & 6.177 & 3.066 & 0.000\ 7.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 10.442 & 6.437 & 3.191 & 0.000\ 8.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 10.816 & 6.692 & 3.316 & 0.000\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.187 & 6.945 & 3.440 & 0.000\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.555 & 7.195 & 3.565 & 0.000\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.922 & 7.445 & 3.693 & 0.000\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.288 & 7.695 & 3.822 & 0.000\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.655 & 7.947 & 3.955 & 0.000\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.023 & 8.200 & 4.092 & 0.000\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.392 & 8.455 & 4.232 & 0.000\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.764 & 8.714 & 4.377 & 0.000\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.139 & 8.975 & 4.526 & 0.000\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.517 & 9.241 & 4.680 & 0.000\ [ccccccccccc]{} 4.00 & 0.000 & 4.765 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 0.001 & 2.820 & 13.944 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 0.024 & 1.268 & 9.524 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 0.315 & 0.288 & 6.129 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 1.052 & 0.040 & 3.951 & 13.757 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 1.673 & 0.011 & 2.429 & 9.568 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 2.046 & 0.024 & 1.354 & 6.422 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 2.364 & 0.112 & 0.653 & 4.134 & 10.748 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 2.791 & 0.341 & 0.268 & 2.527 & 7.363 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 3.344 & 0.701 & 0.118 & 1.414 & 4.858 & 10.657 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 4.009 & 1.169 & 0.146 & 0.665 & 3.001 & 7.297 & 14.277 & 15.000 & 15.000 & 15.000\ 5.10 & 4.834 & 1.793 & 0.380 & 0.262 & 1.700 & 4.786 & 10.056 & 15.000 & 15.000 & 15.000\ 5.20 & 5.816 & 2.569 & 0.806 & 0.155 & 0.855 & 2.953 & 6.831 & 10.942 & 15.000 & 15.000\ 5.30 & 6.927 & 3.474 & 1.391 & 0.288 & 0.375 & 1.659 & 4.394 & 7.175 & 15.000 & 15.000\ 5.40 & 8.170 & 4.512 & 2.135 & 0.642 & 0.215 & 0.818 & 2.613 & 4.313 & 15.000 & 15.000\ 5.50 & 9.544 & 5.684 & 3.034 & 1.201 & 0.337 & 0.368 & 1.383 & 2.201 & 15.000 & 15.000\ 5.60 & 11.126 & 7.067 & 4.163 & 2.028 & 0.789 & 0.332 & 0.697 & 0.791 & 12.007 & 15.000\ 5.70 & 13.118 & 8.867 & 5.725 & 3.318 & 1.751 & 0.875 & 0.694 & 0.190 & 8.800 & 15.000\ 5.80 & 15.000 & 10.817 & 7.452 & 4.796 & 2.938 & 1.698 & 1.053 & 0.050 & 6.567 & 14.817\ 5.90 & 15.000 & 12.620 & 9.045 & 6.159 & 4.040 & 2.479 & 1.435 & 0.018 & 4.846 & 11.168\ 6.00 & 15.000 & 14.234 & 10.460 & 7.359 & 5.001 & 3.154 & 1.764 & 0.008 & 3.472 & 8.239\ 6.10 & 15.000 & 15.000 & 11.696 & 8.393 & 5.814 & 3.710 & 2.017 & 0.006 & 2.371 & 5.879\ 6.20 & 15.000 & 15.000 & 12.762 & 9.268 & 6.484 & 4.146 & 2.184 & 0.017 & 1.507 & 3.992\ 6.30 & 15.000 & 15.000 & 13.712 & 10.036 & 7.058 & 4.505 & 2.309 & 0.066 & 0.877 & 2.524\ 6.40 & 15.000 & 15.000 & 14.639 & 10.790 & 7.629 & 4.877 & 2.476 & 0.198 & 0.486 & 1.444\ 6.50 & 15.000 & 15.000 & 15.000 & 11.615 & 8.280 & 5.345 & 2.764 & 0.459 & 0.336 & 0.724\ 6.60 & 15.000 & 15.000 & 15.000 & 12.561 & 9.059 & 5.952 & 3.214 & 0.870 & 0.411 & 0.322\ 6.70 & 15.000 & 15.000 & 15.000 & 13.587 & 9.926 & 6.658 & 3.777 & 1.379 & 0.631 & 0.140\ 6.80 & 15.000 & 15.000 & 15.000 & 14.619 & 10.806 & 7.385 & 4.370 & 1.903 & 0.899 & 0.065\ 6.90 & 15.000 & 15.000 & 15.000 & 15.000 & 11.660 & 8.092 & 4.951 & 2.400 & 1.162 & 0.033\ 7.00 & 15.000 & 15.000 & 15.000 & 15.000 & 12.479 & 8.770 & 5.505 & 2.860 & 1.404 & 0.018\ 7.10 & 15.000 & 15.000 & 15.000 & 15.000 & 13.264 & 9.419 & 6.030 & 3.283 & 1.623 & 0.011\ 7.20 & 15.000 & 15.000 & 15.000 & 15.000 & 14.018 & 10.041 & 6.530 & 3.673 & 1.821 & 0.007\ 7.30 & 15.000 & 15.000 & 15.000 & 15.000 & 14.745 & 10.638 & 7.006 & 4.034 & 2.001 & 0.004\ 7.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.215 & 7.461 & 4.372 & 2.166 & 0.003\ 7.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.774 & 7.898 & 4.690 & 2.319 & 0.002\ 7.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.315 & 8.318 & 4.991 & 2.463 & 0.002\ 7.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.843 & 8.726 & 5.277 & 2.600 & 0.001\ 7.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.358 & 9.121 & 5.552 & 2.731 & 0.001\ 7.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.863 & 9.508 & 5.819 & 2.859 & 0.001\ 8.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.360 & 9.887 & 6.078 & 2.984 & 0.000\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.851 & 10.261 & 6.333 & 3.108 & 0.000\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 10.631 & 6.585 & 3.233 & 0.000\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 10.998 & 6.835 & 3.358 & 0.000\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.364 & 7.084 & 3.485 & 0.000\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.729 & 7.333 & 3.615 & 0.000\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.095 & 7.583 & 3.748 & 0.000\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.461 & 7.836 & 3.884 & 0.000\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.830 & 8.090 & 4.024 & 0.000\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.201 & 8.348 & 4.169 & 0.000\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.575 & 8.609 & 4.318 & 0.000\ [cccccccccccc]{} 4.00 & 0.000 & 7.324 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 0.000 & 4.891 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 0.001 & 2.918 & 12.845 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 0.021 & 1.330 & 8.442 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 0.303 & 0.299 & 5.146 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 1.090 & 0.037 & 3.079 & 10.508 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 1.800 & 0.016 & 1.686 & 6.961 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 2.261 & 0.081 & 0.784 & 4.411 & 11.839 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 2.705 & 0.296 & 0.310 & 2.689 & 8.104 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 3.249 & 0.652 & 0.126 & 1.538 & 5.395 & 11.656 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 3.888 & 1.106 & 0.126 & 0.760 & 3.391 & 8.010 & 14.847 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 4.668 & 1.695 & 0.319 & 0.310 & 1.955 & 5.264 & 10.426 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 5.607 & 2.435 & 0.705 & 0.156 & 0.996 & 3.246 & 7.056 & 12.781 & 15.000 & 15.000 & 15.000\ 5.30 & 6.690 & 3.314 & 1.263 & 0.254 & 0.429 & 1.815 & 4.523 & 8.788 & 13.291 & 15.000 & 15.000\ 5.40 & 7.927 & 4.345 & 1.999 & 0.595 & 0.218 & 0.891 & 2.695 & 5.766 & 8.907 & 15.000 & 15.000\ 5.50 & 9.306 & 5.520 & 2.899 & 1.151 & 0.311 & 0.394 & 1.447 & 3.538 & 5.573 & 15.000 & 15.000\ 5.60 & 10.819 & 6.830 & 3.953 & 1.902 & 0.668 & 0.257 & 0.683 & 1.965 & 3.098 & 15.000 & 15.000\ 5.70 & 12.486 & 8.300 & 5.182 & 2.858 & 1.285 & 0.456 & 0.353 & 0.963 & 1.356 & 12.638 & 15.000\ 5.80 & 14.471 & 10.094 & 6.748 & 4.176 & 2.305 & 1.118 & 0.564 & 0.614 & 0.396 & 9.086 & 15.000\ 5.90 & 15.000 & 12.195 & 8.634 & 5.831 & 3.694 & 2.195 & 1.252 & 0.829 & 0.103 & 6.710 & 14.965\ 6.00 & 15.000 & 14.232 & 10.465 & 7.448 & 5.068 & 3.292 & 2.010 & 1.184 & 0.034 & 4.963 & 11.325\ 6.10 & 15.000 & 15.000 & 12.114 & 8.895 & 6.292 & 4.267 & 2.686 & 1.511 & 0.015 & 3.589 & 8.423\ 6.20 & 15.000 & 15.000 & 13.570 & 10.159 & 7.349 & 5.096 & 3.248 & 1.770 & 0.009 & 2.493 & 6.089\ 6.30 & 15.000 & 15.000 & 14.844 & 11.250 & 8.246 & 5.782 & 3.694 & 1.949 & 0.015 & 1.630 & 4.220\ 6.40 & 15.000 & 15.000 & 15.000 & 12.213 & 9.025 & 6.365 & 4.057 & 2.080 & 0.052 & 0.987 & 2.754\ 6.50 & 15.000 & 15.000 & 15.000 & 13.130 & 9.768 & 6.923 & 4.415 & 2.234 & 0.155 & 0.565 & 1.654\ 6.60 & 15.000 & 15.000 & 15.000 & 14.083 & 10.555 & 7.535 & 4.842 & 2.484 & 0.365 & 0.361 & 0.888\ 6.70 & 15.000 & 15.000 & 15.000 & 15.000 & 11.442 & 8.255 & 5.390 & 2.876 & 0.709 & 0.369 & 0.426\ 6.80 & 15.000 & 15.000 & 15.000 & 15.000 & 12.418 & 9.071 & 6.045 & 3.390 & 1.161 & 0.534 & 0.195\ 6.90 & 15.000 & 15.000 & 15.000 & 15.000 & 13.419 & 9.920 & 6.741 & 3.953 & 1.650 & 0.770 & 0.093\ 7.00 & 15.000 & 15.000 & 15.000 & 15.000 & 14.401 & 10.756 & 7.430 & 4.515 & 2.125 & 1.017 & 0.048\ 7.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.563 & 8.095 & 5.056 & 2.569 & 1.250 & 0.026\ 7.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.339 & 8.734 & 5.572 & 2.980 & 1.463 & 0.016\ 7.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.086 & 9.348 & 6.063 & 3.360 & 1.657 & 0.010\ 7.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.807 & 9.939 & 6.532 & 3.714 & 1.835 & 0.006\ 7.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.506 & 10.512 & 6.982 & 4.046 & 1.998 & 0.004\ 7.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.067 & 7.415 & 4.360 & 2.150 & 0.003\ 7.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.608 & 7.835 & 4.659 & 2.293 & 0.002\ 7.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.137 & 8.244 & 4.947 & 2.429 & 0.002\ 7.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.657 & 8.643 & 5.226 & 2.560 & 0.001\ 8.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.168 & 9.036 & 5.498 & 2.687 & 0.001\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.674 & 9.424 & 5.767 & 2.812 & 0.001\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.177 & 9.809 & 6.034 & 2.936 & 0.001\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.677 & 10.194 & 6.300 & 3.061 & 0.000\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 10.579 & 6.569 & 3.186 & 0.000\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 10.967 & 6.840 & 3.313 & 0.000\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.358 & 7.115 & 3.443 & 0.000\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.754 & 7.396 & 3.576 & 0.000\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.155 & 7.684 & 3.713 & 0.000\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.563 & 7.979 & 3.853 & 0.000\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.979 & 8.281 & 3.998 & 0.000\ [ccccccccccccc]{} 4.00 & 3.201 & 0.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 3.873 & 0.000 & 14.221 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 4.435 & 0.000 & 11.417 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 4.908 & 0.000 & 8.613 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 5.310 & 0.000 & 6.012 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 5.652 & 0.000 & 3.918 & 12.315 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 5.931 & 0.002 & 2.243 & 8.194 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 6.126 & 0.046 & 1.000 & 5.061 & 13.104 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 6.326 & 0.278 & 0.326 & 2.966 & 8.923 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 6.599 & 0.691 & 0.111 & 1.670 & 5.999 & 13.108 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 6.932 & 1.165 & 0.108 & 0.819 & 3.862 & 9.126 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 7.417 & 1.752 & 0.298 & 0.324 & 2.329 & 6.130 & 11.815 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 8.082 & 2.476 & 0.675 & 0.135 & 1.286 & 3.912 & 8.144 & 14.045 & 15.000 & 15.000 & 15.000 & 15.000\ 5.30 & 8.883 & 3.298 & 1.185 & 0.167 & 0.606 & 2.277 & 5.328 & 9.781 & 15.000 & 15.000 & 15.000 & 15.000\ 5.40 & 9.852 & 4.256 & 1.854 & 0.427 & 0.266 & 1.155 & 3.239 & 6.508 & 11.002 & 15.000 & 15.000 & 15.000\ 5.50 & 11.015 & 5.382 & 2.713 & 0.927 & 0.260 & 0.504 & 1.789 & 4.083 & 7.360 & 10.682 & 15.000 & 15.000\ 5.60 & 12.361 & 6.671 & 3.750 & 1.646 & 0.546 & 0.255 & 0.878 & 2.363 & 4.642 & 6.862 & 15.000 & 15.000\ 5.70 & 13.864 & 8.100 & 4.941 & 2.550 & 1.079 & 0.337 & 0.409 & 1.216 & 2.675 & 3.995 & 15.000 & 15.000\ 5.80 & 15.000 & 9.682 & 6.297 & 3.644 & 1.848 & 0.721 & 0.329 & 0.568 & 1.348 & 1.930 & 13.055 & 15.000\ 5.90 & 15.000 & 11.538 & 7.936 & 5.043 & 2.960 & 1.498 & 0.710 & 0.465 & 0.682 & 0.655 & 9.254 & 15.000\ 6.00 & 15.000 & 13.777 & 9.970 & 6.852 & 4.512 & 2.755 & 1.624 & 0.964 & 0.708 & 0.174 & 6.743 & 14.810\ 6.10 & 15.000 & 15.000 & 12.047 & 8.720 & 6.144 & 4.123 & 2.689 & 1.670 & 1.012 & 0.056 & 4.988 & 11.231\ 6.20 & 15.000 & 15.000 & 13.945 & 10.421 & 7.630 & 5.368 & 3.664 & 2.330 & 1.325 & 0.023 & 3.635 & 8.405\ 6.30 & 15.000 & 15.000 & 15.000 & 11.923 & 8.932 & 6.448 & 4.500 & 2.887 & 1.580 & 0.013 & 2.561 & 6.137\ 6.40 & 15.000 & 15.000 & 15.000 & 13.232 & 10.054 & 7.365 & 5.194 & 3.331 & 1.761 & 0.016 & 1.715 & 4.319\ 6.50 & 15.000 & 15.000 & 15.000 & 14.394 & 11.039 & 8.158 & 5.781 & 3.694 & 1.895 & 0.045 & 1.076 & 2.884\ 6.60 & 15.000 & 15.000 & 15.000 & 15.000 & 11.960 & 8.898 & 6.330 & 4.041 & 2.043 & 0.128 & 0.638 & 1.790\ 6.70 & 15.000 & 15.000 & 15.000 & 15.000 & 12.897 & 9.664 & 6.915 & 4.441 & 2.271 & 0.303 & 0.399 & 1.006\ 6.80 & 15.000 & 15.000 & 15.000 & 15.000 & 13.910 & 10.515 & 7.595 & 4.951 & 2.627 & 0.602 & 0.358 & 0.509\ 6.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.455 & 8.373 & 5.570 & 3.106 & 1.013 & 0.479 & 0.244\ 7.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.432 & 9.195 & 6.242 & 3.647 & 1.473 & 0.684 & 0.120\ 7.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.399 & 10.014 & 6.919 & 4.197 & 1.931 & 0.911 & 0.063\ 7.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.338 & 10.809 & 7.577 & 4.730 & 2.364 & 1.132 & 0.035\ 7.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.576 & 8.211 & 5.240 & 2.767 & 1.338 & 0.021\ 7.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.316 & 8.822 & 5.728 & 3.142 & 1.527 & 0.013\ 7.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.032 & 9.410 & 6.193 & 3.491 & 1.700 & 0.009\ 7.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.727 & 9.980 & 6.640 & 3.819 & 1.861 & 0.006\ 7.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.404 & 10.533 & 7.071 & 4.130 & 2.010 & 0.004\ 7.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.072 & 7.488 & 4.427 & 2.151 & 0.003\ 7.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.599 & 7.894 & 4.712 & 2.286 & 0.002\ 8.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.116 & 8.291 & 4.989 & 2.416 & 0.002\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.625 & 8.682 & 5.259 & 2.543 & 0.001\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.129 & 9.068 & 5.527 & 2.667 & 0.001\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.630 & 9.451 & 5.792 & 2.792 & 0.001\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.130 & 9.834 & 6.058 & 2.916 & 0.001\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.630 & 10.219 & 6.325 & 3.042 & 0.000\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 10.605 & 6.596 & 3.169 & 0.000\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 10.995 & 6.871 & 3.300 & 0.000\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.391 & 7.152 & 3.433 & 0.000\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.792 & 7.440 & 3.571 & 0.000\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.200 & 7.735 & 3.712 & 0.000\ [cccccccccccccc]{} 4.00 & 0.272 & 0.333 & 4.014 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 0.616 & 0.125 & 2.118 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 1.093 & 0.129 & 0.754 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 1.991 & 0.614 & 0.127 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 3.156 & 1.413 & 0.017 & 12.784 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 4.215 & 2.146 & 0.003 & 9.795 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 5.125 & 2.761 & 0.001 & 7.040 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 5.904 & 3.273 & 0.000 & 4.844 & 13.296 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 6.567 & 3.689 & 0.000 & 3.124 & 9.250 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 7.097 & 3.992 & 0.006 & 1.836 & 6.171 & 13.637 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 7.488 & 4.171 & 0.054 & 0.930 & 3.891 & 9.485 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 7.864 & 4.350 & 0.236 & 0.383 & 2.268 & 6.370 & 12.364 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 8.367 & 4.668 & 0.603 & 0.164 & 1.185 & 4.080 & 8.509 & 14.650 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.30 & 9.034 & 5.161 & 1.133 & 0.207 & 0.520 & 2.424 & 5.592 & 10.209 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.40 & 9.892 & 5.852 & 1.832 & 0.489 & 0.214 & 1.298 & 3.440 & 6.820 & 11.419 & 15.000 & 15.000 & 15.000 & 15.000\ 5.50 & 10.919 & 6.721 & 2.677 & 0.970 & 0.201 & 0.604 & 1.907 & 4.276 & 7.679 & 12.197 & 15.000 & 15.000 & 15.000\ 5.60 & 12.125 & 7.777 & 3.681 & 1.652 & 0.461 & 0.296 & 0.914 & 2.454 & 4.872 & 8.185 & 11.548 & 15.000 & 15.000\ 5.70 & 13.533 & 9.039 & 4.868 & 2.548 & 0.990 & 0.351 & 0.405 & 1.264 & 2.866 & 5.193 & 7.480 & 15.000 & 15.000\ 5.80 & 15.000 & 10.487 & 6.220 & 3.633 & 1.754 & 0.711 & 0.299 & 0.594 & 1.517 & 3.036 & 4.444 & 15.000 & 15.000\ 5.90 & 15.000 & 12.108 & 7.732 & 4.898 & 2.730 & 1.342 & 0.537 & 0.360 & 0.714 & 1.569 & 2.255 & 12.993 & 15.000\ 6.00 & 15.000 & 13.985 & 9.487 & 6.423 & 3.994 & 2.302 & 1.158 & 0.583 & 0.456 & 0.761 & 0.851 & 9.171 & 15.000\ 6.10 & 15.000 & 15.000 & 11.615 & 8.334 & 5.665 & 3.703 & 2.262 & 1.343 & 0.806 & 0.651 & 0.244 & 6.620 & 14.358\ 6.20 & 15.000 & 15.000 & 13.847 & 10.360 & 7.468 & 5.262 & 3.556 & 2.335 & 1.445 & 0.899 & 0.079 & 4.887 & 10.894\ 6.30 & 15.000 & 15.000 & 15.000 & 12.238 & 9.138 & 6.710 & 4.763 & 3.273 & 2.073 & 1.191 & 0.033 & 3.575 & 8.182\ 6.40 & 15.000 & 15.000 & 15.000 & 13.913 & 10.618 & 7.983 & 5.816 & 4.083 & 2.609 & 1.434 & 0.019 & 2.541 & 6.013\ 6.50 & 15.000 & 15.000 & 15.000 & 15.000 & 11.912 & 9.085 & 6.715 & 4.760 & 3.041 & 1.611 & 0.020 & 1.728 & 4.274\ 6.60 & 15.000 & 15.000 & 15.000 & 15.000 & 13.064 & 10.057 & 7.495 & 5.335 & 3.397 & 1.746 & 0.044 & 1.110 & 2.898\ 6.70 & 15.000 & 15.000 & 15.000 & 15.000 & 14.144 & 10.965 & 8.224 & 5.873 & 3.737 & 1.894 & 0.118 & 0.677 & 1.838\ 6.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.884 & 8.973 & 6.442 & 4.124 & 2.113 & 0.274 & 0.427 & 1.067\ 6.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.871 & 9.799 & 7.098 & 4.612 & 2.450 & 0.544 & 0.360 & 0.562\ 7.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.938 & 10.711 & 7.848 & 5.205 & 2.904 & 0.921 & 0.450 & 0.281\ 7.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.667 & 8.649 & 5.858 & 3.425 & 1.355 & 0.629 & 0.142\ 7.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.620 & 9.453 & 6.521 & 3.960 & 1.794 & 0.838 & 0.076\ 7.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.549 & 10.239 & 7.171 & 4.484 & 2.214 & 1.046 & 0.044\ 7.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.448 & 10.999 & 7.798 & 4.987 & 2.607 & 1.242 & 0.027\ 7.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.734 & 8.404 & 5.468 & 2.973 & 1.424 & 0.017\ 7.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.445 & 8.988 & 5.928 & 3.317 & 1.592 & 0.011\ 7.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.136 & 9.554 & 6.371 & 3.639 & 1.749 & 0.008\ 7.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.809 & 10.103 & 6.797 & 3.946 & 1.896 & 0.006\ 7.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.465 & 10.639 & 7.211 & 4.238 & 2.035 & 0.004\ 8.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.162 & 7.613 & 4.520 & 2.168 & 0.003\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.676 & 8.007 & 4.794 & 2.296 & 0.002\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.183 & 8.395 & 5.063 & 2.422 & 0.002\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.685 & 8.779 & 5.328 & 2.547 & 0.001\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.184 & 9.160 & 5.592 & 2.670 & 0.001\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.682 & 9.542 & 5.857 & 2.795 & 0.001\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.181 & 9.925 & 6.124 & 2.921 & 0.001\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.682 & 10.311 & 6.395 & 3.049 & 0.000\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 10.701 & 6.670 & 3.180 & 0.000\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.096 & 6.952 & 3.315 & 0.000\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.497 & 7.240 & 3.453 & 0.000\ [ccccccccccccccc]{} 4.00 & 1.422 & 0.017 & 6.249 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 1.765 & 0.008 & 4.422 & 12.816 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 1.969 & 0.005 & 3.074 & 9.001 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 2.145 & 0.007 & 2.030 & 5.972 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 2.338 & 0.030 & 1.206 & 3.545 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 2.616 & 0.137 & 0.610 & 1.653 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 3.214 & 0.562 & 0.446 & 0.434 & 12.598 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 4.382 & 1.558 & 0.940 & 0.067 & 9.420 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 5.639 & 2.643 & 1.592 & 0.012 & 6.780 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 6.760 & 3.592 & 2.165 & 0.003 & 4.742 & 12.007 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 7.731 & 4.390 & 2.631 & 0.001 & 3.178 & 8.413 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 8.539 & 5.028 & 2.973 & 0.005 & 1.984 & 5.652 & 11.991 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 9.193 & 5.515 & 3.193 & 0.038 & 1.085 & 3.533 & 8.253 & 14.722 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.30 & 9.815 & 5.972 & 3.407 & 0.180 & 0.484 & 1.966 & 5.378 & 10.220 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.40 & 10.583 & 6.577 & 3.789 & 0.531 & 0.231 & 0.932 & 3.279 & 6.809 & 11.616 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.50 & 11.567 & 7.401 & 4.405 & 1.112 & 0.311 & 0.377 & 1.849 & 4.313 & 7.867 & 12.533 & 15.000 & 15.000 & 15.000 & 15.000\ 5.60 & 12.736 & 8.410 & 5.220 & 1.871 & 0.648 & 0.191 & 0.943 & 2.540 & 5.069 & 8.540 & 12.979 & 15.000 & 15.000 & 15.000\ 5.70 & 14.048 & 9.564 & 6.191 & 2.761 & 1.175 & 0.283 & 0.441 & 1.331 & 3.025 & 5.511 & 8.778 & 12.068 & 15.000 & 15.000\ 5.80 & 15.000 & 10.884 & 7.337 & 3.806 & 1.901 & 0.640 & 0.306 & 0.620 & 1.629 & 3.303 & 5.610 & 7.867 & 15.000 & 15.000\ 5.90 & 15.000 & 12.385 & 8.671 & 5.022 & 2.832 & 1.252 & 0.506 & 0.345 & 0.792 & 1.792 & 3.314 & 4.726 & 15.000 & 15.000\ 6.00 & 15.000 & 14.065 & 10.192 & 6.413 & 3.965 & 2.103 & 1.005 & 0.445 & 0.426 & 0.863 & 1.741 & 2.459 & 12.648 & 15.000\ 6.10 & 15.000 & 15.000 & 11.965 & 8.049 & 5.362 & 3.247 & 1.843 & 0.940 & 0.528 & 0.491 & 0.837 & 0.980 & 8.888 & 15.000\ 6.20 & 15.000 & 15.000 & 14.117 & 10.058 & 7.151 & 4.803 & 3.127 & 1.925 & 1.174 & 0.733 & 0.634 & 0.298 & 6.371 & 13.683\ 6.30 & 15.000 & 15.000 & 15.000 & 12.215 & 9.102 & 6.539 & 4.618 & 3.148 & 2.100 & 1.310 & 0.833 & 0.099 & 4.691 & 10.381\ 6.40 & 15.000 & 15.000 & 15.000 & 14.240 & 10.933 & 8.168 & 6.023 & 4.311 & 2.998 & 1.903 & 1.099 & 0.043 & 3.438 & 7.815\ 6.50 & 15.000 & 15.000 & 15.000 & 15.000 & 12.572 & 9.618 & 7.266 & 5.332 & 3.779 & 2.414 & 1.326 & 0.024 & 2.456 & 5.767\ 6.60 & 15.000 & 15.000 & 15.000 & 15.000 & 14.028 & 10.892 & 8.348 & 6.208 & 4.435 & 2.830 & 1.495 & 0.024 & 1.687 & 4.126\ 6.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.035 & 9.309 & 6.977 & 5.001 & 3.180 & 1.631 & 0.048 & 1.101 & 2.825\ 6.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.110 & 10.213 & 7.700 & 5.534 & 3.518 & 1.782 & 0.118 & 0.686 & 1.818\ 6.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.186 & 11.126 & 8.442 & 6.097 & 3.902 & 2.002 & 0.266 & 0.439 & 1.077\ 7.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.101 & 9.254 & 6.740 & 4.380 & 2.331 & 0.519 & 0.363 & 0.586\ 7.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.150 & 10.147 & 7.472 & 4.957 & 2.770 & 0.874 & 0.435 & 0.304\ 7.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.235 & 11.084 & 8.254 & 5.593 & 3.274 & 1.285 & 0.594 & 0.159\ 7.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.021 & 9.043 & 6.242 & 3.795 & 1.705 & 0.786 & 0.088\ 7.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.938 & 9.817 & 6.881 & 4.307 & 2.110 & 0.982 & 0.052\ 7.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.828 & 10.568 & 7.500 & 4.800 & 2.491 & 1.169 & 0.032\ 7.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.690 & 11.295 & 8.099 & 5.273 & 2.848 & 1.343 & 0.021\ 7.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.000 & 8.677 & 5.727 & 3.183 & 1.506 & 0.014\ 7.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.684 & 9.237 & 6.163 & 3.499 & 1.658 & 0.010\ 7.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.351 & 9.781 & 6.584 & 3.800 & 1.802 & 0.007\ 8.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.003 & 10.312 & 6.993 & 4.088 & 1.938 & 0.005\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.642 & 10.831 & 7.391 & 4.366 & 2.069 & 0.004\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.342 & 7.781 & 4.637 & 2.197 & 0.003\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.845 & 8.166 & 4.903 & 2.322 & 0.002\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.345 & 8.548 & 5.167 & 2.446 & 0.002\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.842 & 8.927 & 5.430 & 2.570 & 0.001\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.338 & 9.308 & 5.694 & 2.695 & 0.001\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.836 & 9.690 & 5.961 & 2.821 & 0.001\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.336 & 10.075 & 6.232 & 2.950 & 0.000\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.840 & 10.465 & 6.508 & 3.082 & 0.000\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 10.861 & 6.791 & 3.218 & 0.000\ [cccccccccccccccc]{} 4.00 & 0.638 & 0.113 & 4.786 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 1.435 & 0.017 & 2.955 & 13.766 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 2.109 & 0.014 & 1.627 & 9.882 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 2.645 & 0.089 & 0.739 & 6.960 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 3.180 & 0.323 & 0.281 & 4.929 & 11.328 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 3.749 & 0.670 & 0.105 & 3.499 & 7.915 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 4.299 & 1.030 & 0.044 & 2.420 & 5.235 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 4.820 & 1.372 & 0.032 & 1.567 & 3.088 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 5.359 & 1.737 & 0.086 & 0.920 & 1.390 & 12.195 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 6.196 & 2.402 & 0.473 & 0.718 & 0.330 & 8.432 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 7.543 & 3.578 & 1.400 & 1.142 & 0.052 & 5.929 & 14.241 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 8.904 & 4.771 & 2.366 & 1.675 & 0.011 & 4.161 & 10.346 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 10.094 & 5.794 & 3.183 & 2.114 & 0.004 & 2.814 & 7.331 & 14.010 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.30 & 11.088 & 6.621 & 3.820 & 2.418 & 0.009 & 1.766 & 4.957 & 9.988 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.40 & 11.935 & 7.303 & 4.318 & 2.617 & 0.051 & 0.968 & 3.091 & 6.783 & 11.810 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.50 & 12.788 & 7.992 & 4.825 & 2.854 & 0.208 & 0.445 & 1.701 & 4.301 & 7.990 & 12.847 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.60 & 13.794 & 8.834 & 5.494 & 3.277 & 0.565 & 0.248 & 0.799 & 2.502 & 5.112 & 8.718 & 13.344 & 15.000 & 15.000 & 15.000 & 15.000\ 5.70 & 15.000 & 9.883 & 6.378 & 3.935 & 1.145 & 0.370 & 0.345 & 1.315 & 3.056 & 5.644 & 9.099 & 13.327 & 15.000 & 15.000 & 15.000\ 5.80 & 15.000 & 11.116 & 7.453 & 4.802 & 1.917 & 0.755 & 0.256 & 0.628 & 1.669 & 3.429 & 5.921 & 9.032 & 12.178 & 15.000 & 15.000\ 5.90 & 15.000 & 12.512 & 8.697 & 5.855 & 2.857 & 1.361 & 0.466 & 0.347 & 0.821 & 1.907 & 3.604 & 5.805 & 7.971 & 15.000 & 15.000\ 6.00 & 15.000 & 14.077 & 10.112 & 7.092 & 3.969 & 2.178 & 0.945 & 0.417 & 0.424 & 0.958 & 1.997 & 3.455 & 4.820 & 15.000 & 15.000\ 6.10 & 15.000 & 15.000 & 11.709 & 8.522 & 5.266 & 3.210 & 1.680 & 0.805 & 0.419 & 0.496 & 0.984 & 1.835 & 2.539 & 12.072 & 15.000\ 6.20 & 15.000 & 15.000 & 13.557 & 10.213 & 6.818 & 4.520 & 2.725 & 1.547 & 0.822 & 0.512 & 0.537 & 0.885 & 1.042 & 8.450 & 15.000\ 6.30 & 15.000 & 15.000 & 15.000 & 12.284 & 8.748 & 6.226 & 4.189 & 2.742 & 1.718 & 1.076 & 0.705 & 0.632 & 0.333 & 6.028 & 12.853\ 6.40 & 15.000 & 15.000 & 15.000 & 14.525 & 10.847 & 8.116 & 5.856 & 4.166 & 2.873 & 1.938 & 1.225 & 0.793 & 0.116 & 4.428 & 9.749\ 6.50 & 15.000 & 15.000 & 15.000 & 15.000 & 12.832 & 9.904 & 7.436 & 5.524 & 3.986 & 2.789 & 1.776 & 1.033 & 0.051 & 3.247 & 7.349\ 6.60 & 15.000 & 15.000 & 15.000 & 15.000 & 14.628 & 11.514 & 8.851 & 6.732 & 4.968 & 3.534 & 2.255 & 1.242 & 0.031 & 2.327 & 5.437\ 6.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.953 & 10.104 & 7.793 & 5.819 & 4.168 & 2.650 & 1.401 & 0.030 & 1.608 & 3.907\ 6.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.266 & 11.239 & 8.748 & 6.576 & 4.724 & 2.993 & 1.538 & 0.055 & 1.061 & 2.691\ 6.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.318 & 9.655 & 7.297 & 5.256 & 3.332 & 1.696 & 0.126 & 0.671 & 1.747\ 7.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.400 & 10.574 & 8.040 & 5.822 & 3.719 & 1.922 & 0.272 & 0.439 & 1.051\ 7.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.534 & 11.552 & 8.850 & 6.464 & 4.197 & 2.251 & 0.517 & 0.366 & 0.585\ 7.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.595 & 9.734 & 7.188 & 4.766 & 2.682 & 0.857 & 0.428 & 0.313\ 7.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.672 & 10.657 & 7.958 & 5.390 & 3.172 & 1.249 & 0.573 & 0.170\ 7.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.744 & 11.582 & 8.735 & 6.027 & 3.679 & 1.650 & 0.752 & 0.097\ 7.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.488 & 9.499 & 6.655 & 4.178 & 2.039 & 0.935 & 0.058\ 7.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.369 & 10.241 & 7.265 & 4.661 & 2.407 & 1.113 & 0.037\ 7.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.222 & 10.960 & 7.855 & 5.125 & 2.754 & 1.280 & 0.024\ 7.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.658 & 8.425 & 5.571 & 3.080 & 1.437 & 0.017\ 7.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.337 & 8.979 & 6.000 & 3.389 & 1.584 & 0.012\ 8.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.998 & 9.517 & 6.415 & 3.683 & 1.724 & 0.008\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.645 & 10.042 & 6.818 & 3.966 & 1.858 & 0.006\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.279 & 10.557 & 7.211 & 4.240 & 1.987 & 0.005\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.904 & 11.064 & 7.598 & 4.508 & 2.114 & 0.003\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.564 & 7.979 & 4.772 & 2.238 & 0.003\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.061 & 8.358 & 5.034 & 2.362 & 0.002\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.556 & 8.737 & 5.296 & 2.486 & 0.001\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.051 & 9.116 & 5.560 & 2.611 & 0.001\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.548 & 9.497 & 5.827 & 2.738 & 0.001\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.047 & 9.883 & 6.099 & 2.868 & 0.001\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.551 & 10.273 & 6.376 & 3.001 & 0.000\ [ccccccccccccccccc]{} 4.00 & 0.089 & 0.733 & 6.302 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 0.670 & 0.105 & 3.573 & 13.045 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 1.541 & 0.017 & 2.014 & 8.887 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 2.213 & 0.040 & 1.091 & 5.922 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 2.647 & 0.157 & 0.522 & 3.792 & 12.363 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 3.050 & 0.409 & 0.219 & 2.322 & 8.920 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 3.524 & 0.767 & 0.107 & 1.329 & 6.353 & 12.260 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 4.082 & 1.207 & 0.140 & 0.671 & 4.421 & 8.516 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 4.749 & 1.748 & 0.325 & 0.294 & 3.001 & 5.632 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 5.502 & 2.363 & 0.625 & 0.126 & 1.970 & 3.415 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 6.290 & 3.007 & 0.987 & 0.089 & 1.212 & 1.693 & 10.639 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 7.250 & 3.817 & 1.540 & 0.297 & 0.811 & 0.505 & 7.118 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 8.686 & 5.099 & 2.590 & 1.042 & 1.037 & 0.089 & 4.881 & 11.666 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.30 & 10.218 & 6.475 & 3.751 & 1.932 & 1.478 & 0.020 & 3.387 & 8.429 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.40 & 11.568 & 7.668 & 4.743 & 2.678 & 1.829 & 0.010 & 2.248 & 5.897 & 11.078 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.50 & 12.724 & 8.667 & 5.553 & 3.262 & 2.062 & 0.024 & 1.361 & 3.889 & 7.716 & 12.646 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.60 & 13.795 & 9.580 & 6.287 & 3.787 & 2.269 & 0.098 & 0.715 & 2.336 & 5.060 & 8.698 & 13.408 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.70 & 14.925 & 10.553 & 7.090 & 4.393 & 2.586 & 0.310 & 0.348 & 1.236 & 3.061 & 5.655 & 9.157 & 13.562 & 15.000 & 15.000 & 15.000 & 15.000\ 5.80 & 15.000 & 11.680 & 8.053 & 5.172 & 3.102 & 0.717 & 0.283 & 0.577 & 1.673 & 3.426 & 5.946 & 9.233 & 13.213 & 15.000 & 15.000 & 15.000\ 5.90 & 15.000 & 12.993 & 9.207 & 6.151 & 3.841 & 1.336 & 0.509 & 0.318 & 0.823 & 1.901 & 3.626 & 5.994 & 8.919 & 11.863 & 15.000 & 15.000\ 6.00 & 15.000 & 14.487 & 10.548 & 7.326 & 4.795 & 2.161 & 0.997 & 0.401 & 0.425 & 0.958 & 2.038 & 3.649 & 5.716 & 7.743 & 15.000 & 15.000\ 6.10 & 15.000 & 15.000 & 12.069 & 8.688 & 5.956 & 3.184 & 1.723 & 0.782 & 0.406 & 0.491 & 1.045 & 2.029 & 3.399 & 4.673 & 15.000 & 15.000\ 6.20 & 15.000 & 15.000 & 13.776 & 10.243 & 7.324 & 4.409 & 2.680 & 1.439 & 0.722 & 0.430 & 0.548 & 1.010 & 1.807 & 2.460 & 11.275 & 15.000\ 6.30 & 15.000 & 15.000 & 15.000 & 12.058 & 8.964 & 5.903 & 3.930 & 2.420 & 1.406 & 0.788 & 0.537 & 0.559 & 0.882 & 1.019 & 7.873 & 15.000\ 6.40 & 15.000 & 15.000 & 15.000 & 14.237 & 10.980 & 7.770 & 5.572 & 3.818 & 2.539 & 1.634 & 1.063 & 0.709 & 0.632 & 0.339 & 5.611 & 11.918\ 6.50 & 15.000 & 15.000 & 15.000 & 15.000 & 13.152 & 9.793 & 7.385 & 5.408 & 3.889 & 2.726 & 1.872 & 1.192 & 0.772 & 0.124 & 4.119 & 9.043\ 6.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.706 & 9.101 & 6.916 & 5.177 & 3.780 & 2.673 & 1.705 & 0.989 & 0.058 & 3.021 & 6.823\ 6.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.439 & 10.647 & 8.269 & 6.327 & 4.713 & 3.375 & 2.153 & 1.178 & 0.037 & 2.169 & 5.057\ 6.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.037 & 9.477 & 7.344 & 5.529 & 3.980 & 2.529 & 1.328 & 0.037 & 1.505 & 3.642\ 6.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.316 & 10.583 & 8.271 & 6.267 & 4.522 & 2.864 & 1.466 & 0.065 & 1.000 & 2.517\ 7.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.543 & 11.645 & 9.162 & 6.980 & 5.051 & 3.206 & 1.632 & 0.140 & 0.642 & 1.643\ 7.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.717 & 10.072 & 7.722 & 5.619 & 3.601 & 1.868 & 0.288 & 0.431 & 0.998\ 7.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.841 & 11.040 & 8.529 & 6.261 & 4.082 & 2.203 & 0.532 & 0.368 & 0.568\ 7.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.068 & 9.403 & 6.978 & 4.646 & 2.628 & 0.860 & 0.428 & 0.313\ 7.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.125 & 10.311 & 7.735 & 5.259 & 3.107 & 1.236 & 0.563 & 0.175\ 7.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.176 & 11.220 & 8.498 & 5.883 & 3.600 & 1.620 & 0.729 & 0.103\ 7.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.111 & 9.248 & 6.499 & 4.087 & 1.993 & 0.902 & 0.063\ 7.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.978 & 9.979 & 7.098 & 4.559 & 2.348 & 1.070 & 0.041\ 7.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.820 & 10.688 & 7.678 & 5.013 & 2.683 & 1.230 & 0.027\ 7.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.639 & 11.376 & 8.240 & 5.450 & 3.000 & 1.381 & 0.019\ 8.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.046 & 8.786 & 5.871 & 3.301 & 1.524 & 0.013\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.700 & 9.317 & 6.280 & 3.589 & 1.660 & 0.010\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.340 & 9.837 & 6.677 & 3.867 & 1.791 & 0.007\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.969 & 10.346 & 7.066 & 4.137 & 1.919 & 0.005\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.589 & 10.849 & 7.449 & 4.401 & 2.044 & 0.004\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.346 & 7.828 & 4.663 & 2.167 & 0.003\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.840 & 8.205 & 4.924 & 2.291 & 0.002\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.333 & 8.581 & 5.185 & 2.415 & 0.002\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.827 & 8.960 & 5.449 & 2.541 & 0.001\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.323 & 9.342 & 5.717 & 2.669 & 0.001\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.823 & 9.727 & 5.990 & 2.800 & 0.001\ [cccccccccccccccccc]{} 4.00 & 0.181 & 0.468 & 8.954 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 0.949 & 0.052 & 5.894 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 1.850 & 0.006 & 3.766 & 11.913 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 2.521 & 0.004 & 2.163 & 7.967 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 2.899 & 0.044 & 1.023 & 5.016 & 12.472 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 3.228 & 0.232 & 0.384 & 3.068 & 8.554 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 3.671 & 0.580 & 0.141 & 1.884 & 5.855 & 13.801 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 4.176 & 0.983 & 0.087 & 1.108 & 3.903 & 10.045 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 4.737 & 1.427 & 0.154 & 0.588 & 2.449 & 7.133 & 12.084 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 5.394 & 1.954 & 0.355 & 0.295 & 1.398 & 4.894 & 8.280 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 6.174 & 2.596 & 0.706 & 0.221 & 0.700 & 3.219 & 5.337 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 7.091 & 3.362 & 1.211 & 0.356 & 0.313 & 2.021 & 3.109 & 12.438 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 8.119 & 4.233 & 1.844 & 0.660 & 0.175 & 1.203 & 1.451 & 8.425 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.30 & 9.391 & 5.344 & 2.737 & 1.256 & 0.391 & 0.842 & 0.406 & 5.534 & 12.461 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.40 & 11.075 & 6.865 & 4.055 & 2.305 & 1.107 & 1.062 & 0.081 & 3.758 & 8.942 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.50 & 12.749 & 8.377 & 5.375 & 3.377 & 1.885 & 1.406 & 0.025 & 2.551 & 6.335 & 11.502 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.60 & 14.231 & 9.696 & 6.512 & 4.284 & 2.527 & 1.662 & 0.022 & 1.628 & 4.285 & 8.111 & 12.906 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.70 & 15.000 & 10.884 & 7.525 & 5.081 & 3.081 & 1.871 & 0.062 & 0.935 & 2.686 & 5.425 & 8.971 & 13.453 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.80 & 15.000 & 12.059 & 8.532 & 5.885 & 3.661 & 2.140 & 0.198 & 0.485 & 1.513 & 3.374 & 5.914 & 9.252 & 13.395 & 15.000 & 15.000 & 15.000 & 15.000\ 5.90 & 15.000 & 13.327 & 9.637 & 6.798 & 4.365 & 2.564 & 0.496 & 0.309 & 0.757 & 1.912 & 3.646 & 6.057 & 9.141 & 12.821 & 15.000 & 15.000 & 15.000\ 6.00 & 15.000 & 14.747 & 10.899 & 7.879 & 5.250 & 3.197 & 0.998 & 0.419 & 0.394 & 0.980 & 2.070 & 3.731 & 5.947 & 8.645 & 11.353 & 15.000 & 15.000\ 6.10 & 15.000 & 15.000 & 12.340 & 9.146 & 6.332 & 4.052 & 1.716 & 0.800 & 0.381 & 0.504 & 1.073 & 2.127 & 3.627 & 5.529 & 7.385 & 15.000 & 15.000\ 6.20 & 15.000 & 15.000 & 13.960 & 10.602 & 7.613 & 5.125 & 2.647 & 1.435 & 0.678 & 0.415 & 0.554 & 1.112 & 2.020 & 3.273 & 4.429 & 14.702 & 15.000\ 6.30 & 15.000 & 15.000 & 15.000 & 12.258 & 9.103 & 6.425 & 3.801 & 2.322 & 1.269 & 0.676 & 0.451 & 0.594 & 1.009 & 1.729 & 2.306 & 10.351 & 15.000\ 6.40 & 15.000 & 15.000 & 15.000 & 14.194 & 10.881 & 8.026 & 5.253 & 3.530 & 2.213 & 1.329 & 0.789 & 0.578 & 0.575 & 0.852 & 0.946 & 7.199 & 14.509\ 6.50 & 15.000 & 15.000 & 15.000 & 15.000 & 13.026 & 10.006 & 7.083 & 5.134 & 3.576 & 2.433 & 1.613 & 1.094 & 0.733 & 0.633 & 0.325 & 5.130 & 10.884\ 6.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.117 & 9.045 & 6.885 & 5.106 & 3.728 & 2.654 & 1.861 & 1.188 & 0.763 & 0.127 & 3.762 & 8.255\ 6.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.115 & 10.895 & 8.538 & 6.553 & 4.959 & 3.654 & 2.614 & 1.665 & 0.958 & 0.064 & 2.754 & 6.222\ 6.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.576 & 10.032 & 7.856 & 6.060 & 4.543 & 3.276 & 2.082 & 1.129 & 0.043 & 1.973 & 4.604\ 6.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.107 & 11.387 & 9.030 & 7.046 & 5.331 & 3.855 & 2.441 & 1.273 & 0.046 & 1.366 & 3.308\ 7.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.650 & 10.120 & 7.958 & 6.058 & 4.387 & 2.774 & 1.416 & 0.080 & 0.908 & 2.279\ 7.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.873 & 11.179 & 8.849 & 6.772 & 4.919 & 3.125 & 1.597 & 0.165 & 0.589 & 1.484\ 7.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.258 & 9.766 & 7.523 & 5.497 & 3.534 & 1.851 & 0.326 & 0.411 & 0.904\ 7.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.386 & 10.740 & 8.337 & 6.147 & 4.027 & 2.199 & 0.576 & 0.370 & 0.521\ 7.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.559 & 11.765 & 9.209 & 6.862 & 4.593 & 2.627 & 0.902 & 0.438 & 0.295\ 7.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.809 & 10.107 & 7.609 & 5.198 & 3.098 & 1.264 & 0.569 & 0.171\ 7.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.847 & 11.002 & 8.360 & 5.811 & 3.580 & 1.633 & 0.727 & 0.103\ 7.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.862 & 11.880 & 9.097 & 6.415 & 4.054 & 1.992 & 0.891 & 0.065\ 7.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.734 & 9.815 & 7.003 & 4.515 & 2.334 & 1.051 & 0.043\ 7.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.566 & 10.514 & 7.574 & 4.960 & 2.659 & 1.204 & 0.029\ 8.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.377 & 11.195 & 8.129 & 5.389 & 2.969 & 1.350 & 0.020\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.860 & 8.670 & 5.806 & 3.265 & 1.491 & 0.015\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.510 & 9.198 & 6.212 & 3.551 & 1.626 & 0.011\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.149 & 9.717 & 6.609 & 3.829 & 1.757 & 0.008\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.779 & 10.228 & 7.000 & 4.102 & 1.886 & 0.006\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.402 & 10.734 & 7.387 & 4.372 & 2.014 & 0.004\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.238 & 7.773 & 4.641 & 2.142 & 0.003\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.740 & 8.159 & 4.911 & 2.271 & 0.002\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.244 & 8.546 & 5.185 & 2.401 & 0.002\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.750 & 8.938 & 5.462 & 2.534 & 0.001\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.260 & 9.334 & 5.745 & 2.670 & 0.001\ [ccccccccccccccccccc]{} 4.00 & 0.003 & 2.230 & 9.867 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 0.074 & 0.808 & 5.856 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 0.606 & 0.124 & 3.103 & 11.838 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 1.505 & 0.031 & 1.414 & 7.445 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 2.455 & 0.184 & 0.466 & 4.346 & 12.168 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 3.415 & 0.572 & 0.138 & 2.415 & 8.341 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 4.178 & 1.021 & 0.067 & 1.318 & 5.854 & 12.960 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 4.794 & 1.489 & 0.114 & 0.701 & 4.134 & 9.499 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 5.407 & 2.003 & 0.268 & 0.348 & 2.879 & 6.869 & 13.750 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 6.065 & 2.562 & 0.511 & 0.169 & 1.952 & 4.842 & 10.148 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 6.755 & 3.143 & 0.809 & 0.103 & 1.257 & 3.255 & 7.282 & 11.099 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 7.478 & 3.747 & 1.156 & 0.130 & 0.749 & 2.012 & 4.988 & 7.506 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 8.282 & 4.420 & 1.591 & 0.278 & 0.437 & 1.088 & 3.195 & 4.658 & 14.087 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.30 & 9.243 & 5.238 & 2.188 & 0.616 & 0.374 & 0.506 & 1.885 & 2.488 & 9.589 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.40 & 10.433 & 6.280 & 3.021 & 1.212 & 0.614 & 0.301 & 1.063 & 0.968 & 6.227 & 12.996 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.50 & 12.049 & 7.744 & 4.286 & 2.256 & 1.340 & 0.641 & 0.872 & 0.232 & 4.034 & 9.118 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.60 & 13.909 & 9.451 & 5.804 & 3.567 & 2.363 & 1.323 & 1.088 & 0.063 & 2.712 & 6.440 & 11.397 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.70 & 15.000 & 11.040 & 7.212 & 4.779 & 3.311 & 1.965 & 1.318 & 0.035 & 1.770 & 4.407 & 8.085 & 12.659 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.80 & 15.000 & 12.500 & 8.498 & 5.878 & 4.166 & 2.543 & 1.529 & 0.055 & 1.070 & 2.833 & 5.475 & 8.858 & 13.063 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.90 & 15.000 & 13.912 & 9.741 & 6.944 & 5.003 & 3.127 & 1.788 & 0.151 & 0.594 & 1.660 & 3.472 & 5.902 & 9.027 & 12.868 & 15.000 & 15.000 & 15.000 & 15.000\ 6.00 & 15.000 & 15.000 & 11.030 & 8.062 & 5.908 & 3.798 & 2.171 & 0.383 & 0.362 & 0.871 & 2.019 & 3.690 & 5.944 & 8.794 & 12.137 & 15.000 & 15.000 & 15.000\ 6.10 & 15.000 & 15.000 & 12.438 & 9.306 & 6.951 & 4.624 & 2.743 & 0.807 & 0.399 & 0.455 & 1.069 & 2.133 & 3.686 & 5.724 & 8.167 & 10.620 & 15.000 & 15.000\ 6.20 & 15.000 & 15.000 & 14.002 & 10.713 & 8.168 & 5.640 & 3.533 & 1.449 & 0.709 & 0.385 & 0.561 & 1.133 & 2.118 & 3.488 & 5.201 & 6.870 & 15.000 & 15.000\ 6.30 & 15.000 & 15.000 & 15.000 & 12.296 & 9.573 & 6.857 & 4.546 & 2.313 & 1.279 & 0.628 & 0.436 & 0.599 & 1.117 & 1.933 & 3.052 & 4.074 & 13.393 & 15.000\ 6.40 & 15.000 & 15.000 & 15.000 & 14.086 & 11.193 & 8.300 & 5.804 & 3.421 & 2.122 & 1.183 & 0.673 & 0.485 & 0.612 & 0.965 & 1.592 & 2.077 & 9.377 & 15.000\ 6.50 & 15.000 & 15.000 & 15.000 & 15.000 & 13.122 & 10.063 & 7.395 & 4.860 & 3.319 & 2.121 & 1.330 & 0.837 & 0.626 & 0.585 & 0.801 & 0.839 & 6.511 & 13.153\ 6.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.181 & 9.354 & 6.667 & 4.901 & 3.467 & 2.422 & 1.657 & 1.150 & 0.770 & 0.635 & 0.300 & 4.657 & 9.890\ 6.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.384 & 11.404 & 8.567 & 6.591 & 4.940 & 3.664 & 2.652 & 1.880 & 1.203 & 0.764 & 0.126 & 3.419 & 7.509\ 6.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.339 & 10.351 & 8.179 & 6.326 & 4.836 & 3.599 & 2.587 & 1.647 & 0.942 & 0.067 & 2.502 & 5.661\ 6.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.979 & 9.621 & 7.578 & 5.890 & 4.445 & 3.213 & 2.038 & 1.100 & 0.048 & 1.792 & 4.188\ 7.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.473 & 10.940 & 8.718 & 6.844 & 5.205 & 3.769 & 2.383 & 1.239 & 0.056 & 1.242 & 3.008\ 7.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.881 & 12.182 & 9.790 & 7.740 & 5.918 & 4.291 & 2.714 & 1.388 & 0.097 & 0.829 & 2.071\ 7.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.398 & 10.844 & 8.626 & 6.630 & 4.825 & 3.072 & 1.580 & 0.191 & 0.546 & 1.349\ 7.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.631 & 11.922 & 9.544 & 7.384 & 5.409 & 3.492 & 1.848 & 0.362 & 0.396 & 0.828\ 7.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.048 & 10.517 & 8.200 & 6.063 & 3.992 & 2.205 & 0.616 & 0.372 & 0.486\ 7.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.211 & 11.532 & 9.064 & 6.773 & 4.556 & 2.630 & 0.935 & 0.444 & 0.283\ 7.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.562 & 9.949 & 7.509 & 5.151 & 3.092 & 1.285 & 0.571 & 0.168\ 7.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.582 & 10.828 & 8.245 & 5.751 & 3.561 & 1.639 & 0.720 & 0.104\ 7.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.581 & 11.691 & 8.969 & 6.343 & 4.023 & 1.983 & 0.875 & 0.067\ 7.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.532 & 9.675 & 6.919 & 4.472 & 2.312 & 1.028 & 0.045\ 8.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.352 & 10.364 & 7.481 & 4.907 & 2.626 & 1.175 & 0.031\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.152 & 11.035 & 8.027 & 5.328 & 2.927 & 1.316 & 0.022\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.934 & 11.691 & 8.560 & 5.737 & 3.216 & 1.452 & 0.016\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.334 & 9.082 & 6.137 & 3.496 & 1.585 & 0.012\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.967 & 9.595 & 6.529 & 3.770 & 1.714 & 0.009\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.592 & 10.102 & 6.917 & 4.040 & 1.842 & 0.006\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.212 & 10.605 & 7.302 & 4.308 & 1.970 & 0.005\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.828 & 11.106 & 7.686 & 4.577 & 2.098 & 0.003\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.608 & 8.071 & 4.847 & 2.227 & 0.003\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.111 & 8.459 & 5.122 & 2.359 & 0.002\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.618 & 8.851 & 5.401 & 2.493 & 0.001\ [cccccccccccccccccccc]{} 4.00 & 0.000 & 3.459 & 13.305 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 0.009 & 1.693 & 8.681 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 0.197 & 0.439 & 5.025 & 14.395 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 0.984 & 0.049 & 2.700 & 9.243 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 1.939 & 0.039 & 1.133 & 5.405 & 14.578 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 3.053 & 0.365 & 0.247 & 2.759 & 9.717 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 4.288 & 1.005 & 0.065 & 1.400 & 6.732 & 14.248 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 5.172 & 1.527 & 0.109 & 0.718 & 4.780 & 10.658 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 5.816 & 1.997 & 0.265 & 0.350 & 3.385 & 7.943 & 14.178 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 6.411 & 2.495 & 0.510 & 0.165 & 2.356 & 5.832 & 10.556 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 7.018 & 3.013 & 0.804 & 0.089 & 1.576 & 4.155 & 7.674 & 13.293 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 7.650 & 3.546 & 1.128 & 0.087 & 0.978 & 2.808 & 5.351 & 9.650 & 13.685 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 8.338 & 4.123 & 1.507 & 0.173 & 0.554 & 1.751 & 3.494 & 6.709 & 9.446 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.30 & 9.137 & 4.797 & 1.991 & 0.390 & 0.328 & 0.984 & 2.062 & 4.381 & 6.065 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.40 & 10.108 & 5.634 & 2.642 & 0.796 & 0.343 & 0.534 & 1.051 & 2.622 & 3.445 & 10.313 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.50 & 11.324 & 6.709 & 3.539 & 1.464 & 0.664 & 0.447 & 0.487 & 1.424 & 1.550 & 6.640 & 13.187 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.60 & 12.932 & 8.172 & 4.830 & 2.541 & 1.429 & 0.851 & 0.477 & 0.866 & 0.457 & 4.144 & 9.059 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.70 & 14.946 & 10.037 & 6.528 & 4.036 & 2.640 & 1.741 & 1.002 & 0.910 & 0.122 & 2.700 & 6.308 & 10.985 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.80 & 15.000 & 11.901 & 8.226 & 5.541 & 3.885 & 2.693 & 1.629 & 1.108 & 0.056 & 1.760 & 4.323 & 7.792 & 11.975 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.90 & 15.000 & 13.647 & 9.808 & 6.940 & 5.042 & 3.583 & 2.226 & 1.323 & 0.065 & 1.082 & 2.814 & 5.310 & 8.397 & 12.285 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.00 & 15.000 & 15.000 & 11.321 & 8.278 & 6.156 & 4.450 & 2.824 & 1.583 & 0.147 & 0.621 & 1.694 & 3.412 & 5.625 & 8.512 & 12.021 & 15.000 & 15.000 & 15.000 & 15.000\ 6.10 & 15.000 & 15.000 & 12.839 & 9.628 & 7.295 & 5.361 & 3.490 & 1.950 & 0.350 & 0.383 & 0.928 & 2.025 & 3.543 & 5.625 & 8.217 & 11.232 & 14.299 & 15.000 & 15.000\ 6.20 & 15.000 & 15.000 & 14.434 & 11.063 & 8.532 & 6.385 & 4.288 & 2.484 & 0.727 & 0.391 & 0.503 & 1.100 & 2.061 & 3.495 & 5.335 & 7.530 & 9.722 & 15.000 & 15.000\ 6.30 & 15.000 & 15.000 & 15.000 & 12.637 & 9.920 & 7.574 & 5.268 & 3.230 & 1.319 & 0.663 & 0.410 & 0.594 & 1.100 & 2.007 & 3.228 & 4.757 & 6.230 & 15.000 & 15.000\ 6.40 & 15.000 & 15.000 & 15.000 & 14.385 & 11.491 & 8.958 & 6.459 & 4.210 & 2.145 & 1.206 & 0.637 & 0.469 & 0.594 & 1.063 & 1.771 & 2.755 & 3.634 & 12.027 & 15.000\ 6.50 & 15.000 & 15.000 & 15.000 & 15.000 & 13.284 & 10.575 & 7.895 & 5.454 & 3.236 & 2.041 & 1.191 & 0.718 & 0.513 & 0.612 & 0.888 & 1.418 & 1.802 & 8.376 & 15.000\ 6.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.525 & 9.675 & 7.057 & 4.686 & 3.256 & 2.153 & 1.407 & 0.913 & 0.691 & 0.597 & 0.742 & 0.714 & 5.820 & 11.816\ 6.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.790 & 11.780 & 8.996 & 6.473 & 4.824 & 3.490 & 2.497 & 1.744 & 1.238 & 0.823 & 0.637 & 0.268 & 4.187 & 8.915\ 6.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.919 & 10.978 & 8.304 & 6.452 & 4.904 & 3.685 & 2.697 & 1.935 & 1.238 & 0.764 & 0.122 & 3.080 & 6.778\ 6.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.838 & 10.014 & 7.970 & 6.223 & 4.796 & 3.592 & 2.597 & 1.650 & 0.923 & 0.071 & 2.253 & 5.109\ 7.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.553 & 11.580 & 9.355 & 7.422 & 5.799 & 4.395 & 3.187 & 2.016 & 1.068 & 0.056 & 1.612 & 3.775\ 7.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.033 & 10.637 & 8.527 & 6.721 & 5.130 & 3.724 & 2.350 & 1.207 & 0.068 & 1.116 & 2.705\ 7.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.419 & 11.860 & 9.582 & 7.602 & 5.834 & 4.242 & 2.684 & 1.367 & 0.118 & 0.747 & 1.858\ 7.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.069 & 10.631 & 8.485 & 6.550 & 4.782 & 3.053 & 1.578 & 0.224 & 0.502 & 1.210\ 7.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.301 & 11.709 & 9.406 & 7.311 & 5.377 & 3.489 & 1.865 & 0.406 & 0.382 & 0.747\ 7.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.831 & 10.377 & 8.129 & 6.035 & 3.997 & 2.231 & 0.665 & 0.377 & 0.446\ 7.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.982 & 11.381 & 8.987 & 6.741 & 4.559 & 2.657 & 0.979 & 0.454 & 0.266\ 7.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.395 & 9.858 & 7.466 & 5.146 & 3.109 & 1.316 & 0.578 & 0.163\ 7.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.396 & 10.722 & 8.189 & 5.734 & 3.566 & 1.656 & 0.720 & 0.104\ 7.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.378 & 11.569 & 8.899 & 6.313 & 4.016 & 1.986 & 0.868 & 0.069\ 8.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.396 & 9.592 & 6.878 & 4.454 & 2.303 & 1.013 & 0.047\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.203 & 10.269 & 7.428 & 4.878 & 2.607 & 1.155 & 0.033\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.992 & 10.930 & 7.965 & 5.290 & 2.899 & 1.291 & 0.023\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.764 & 11.577 & 8.490 & 5.692 & 3.182 & 1.424 & 0.017\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.213 & 9.006 & 6.086 & 3.457 & 1.554 & 0.012\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.840 & 9.514 & 6.474 & 3.727 & 1.682 & 0.009\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.460 & 10.016 & 6.858 & 3.995 & 1.809 & 0.007\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.076 & 10.516 & 7.240 & 4.262 & 1.936 & 0.005\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.690 & 11.016 & 7.623 & 4.530 & 2.065 & 0.004\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.516 & 8.008 & 4.801 & 2.195 & 0.003\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.019 & 8.397 & 5.077 & 2.328 & 0.002\ [ccccccccccc]{} 4.00 & 3.425 & 0.000 & 11.456 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 3.999 & 0.000 & 8.520 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 4.478 & 0.000 & 5.799 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 4.879 & 0.000 & 3.614 & 11.687 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 5.224 & 0.006 & 1.861 & 7.434 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 5.639 & 0.134 & 0.577 & 4.155 & 10.438 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 6.457 & 0.733 & 0.092 & 2.176 & 6.539 & 14.837 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 7.060 & 1.376 & 0.051 & 1.169 & 4.218 & 10.723 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 7.095 & 1.813 & 0.130 & 0.618 & 2.707 & 7.794 & 14.459 & 15.000 & 15.000 & 15.000\ 4.90 & 7.068 & 2.224 & 0.312 & 0.316 & 1.640 & 5.570 & 10.795 & 15.000 & 15.000 & 15.000\ 5.00 & 7.217 & 2.714 & 0.610 & 0.204 & 0.895 & 3.868 & 7.912 & 13.068 & 15.000 & 15.000\ 5.10 & 7.587 & 3.324 & 1.039 & 0.274 & 0.428 & 2.599 & 5.667 & 9.571 & 15.000 & 15.000\ 5.20 & 8.154 & 4.051 & 1.589 & 0.502 & 0.195 & 1.687 & 3.941 & 6.827 & 11.237 & 15.000\ 5.30 & 8.856 & 4.848 & 2.213 & 0.834 & 0.123 & 1.036 & 2.604 & 4.651 & 7.971 & 10.769\ 5.40 & 9.664 & 5.702 & 2.895 & 1.249 & 0.181 & 0.592 & 1.576 & 2.926 & 5.344 & 7.112\ 5.50 & 10.615 & 6.658 & 3.682 & 1.788 & 0.401 & 0.375 & 0.856 & 1.618 & 3.283 & 4.209\ 5.60 & 11.796 & 7.815 & 4.673 & 2.547 & 0.870 & 0.458 & 0.501 & 0.762 & 1.786 & 2.036\ 5.70 & 13.365 & 9.339 & 6.033 & 3.689 & 1.749 & 0.991 & 0.650 & 0.476 & 0.943 & 0.683\ 5.80 & 15.000 & 11.346 & 7.878 & 5.329 & 3.144 & 2.076 & 1.390 & 0.834 & 0.810 & 0.189\ 5.90 & 15.000 & 13.452 & 9.825 & 7.081 & 4.669 & 3.318 & 2.321 & 1.424 & 0.964 & 0.079\ 6.00 & 15.000 & 15.000 & 11.663 & 8.734 & 6.109 & 4.498 & 3.218 & 2.013 & 1.167 & 0.070\ 6.10 & 15.000 & 15.000 & 13.402 & 10.296 & 7.472 & 5.620 & 4.078 & 2.594 & 1.412 & 0.127\ 6.20 & 15.000 & 15.000 & 15.000 & 11.830 & 8.816 & 6.740 & 4.957 & 3.220 & 1.743 & 0.287\ 6.30 & 15.000 & 15.000 & 15.000 & 13.416 & 10.224 & 7.939 & 5.932 & 3.963 & 2.229 & 0.609\ 6.40 & 15.000 & 15.000 & 15.000 & 15.000 & 11.772 & 9.291 & 7.075 & 4.894 & 2.931 & 1.152\ 6.50 & 15.000 & 15.000 & 15.000 & 15.000 & 13.512 & 10.846 & 8.435 & 6.058 & 3.891 & 1.955\ 6.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.658 & 10.064 & 7.505 & 5.151 & 3.060\ 6.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.824 & 12.056 & 9.327 & 6.799 & 4.555\ 6.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.310 & 11.421 & 8.730 & 6.333\ 6.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.503 & 10.658 & 8.109\ 7.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.462 & 9.763\ 7.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.137 & 11.290\ 7.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.722\ 7.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.101\ 7.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.00 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.30 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.40 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.50 &10.821 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.60 & 6.966 & 13.130 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.70 & 4.287 & 8.918 & 14.684 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.80 & 2.749 & 6.155 & 10.516 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.90 & 1.820 & 4.253 & 7.482 & 11.366 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.00 & 1.167 & 2.830 & 5.156 & 8.023 & 11.557 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.10 & 0.710 & 1.764 & 3.373 & 5.431 & 8.047 & 11.190 & 14.806 & 15.000 & 15.000 & 15.000\ 6.20 & 0.447 & 1.009 & 2.044 & 3.459 & 5.340 & 7.646 & 10.342 & 13.069 & 15.000 & 15.000\ 6.30 & 0.409 & 0.562 & 1.131 & 2.027 & 3.316 & 4.936 & 6.887 & 8.819 & 15.000 & 15.000\ 6.40 & 0.638 & 0.441 & 0.620 & 1.088 & 1.891 & 2.947 & 4.292 & 5.568 & 15.000 & 15.000\ 6.50 & 1.159 & 0.654 & 0.498 & 0.603 & 1.000 & 1.585 & 2.431 & 3.164 & 10.675 & 15.000\ 6.60 & 2.008 & 1.225 & 0.776 & 0.566 & 0.614 & 0.802 & 1.230 & 1.509 & 7.390 & 14.178\ 6.70 & 3.267 & 2.232 & 1.519 & 1.029 & 0.773 & 0.619 & 0.689 & 0.591 & 5.156 & 10.543\ 6.80 & 4.827 & 3.557 & 2.605 & 1.864 & 1.336 & 0.884 & 0.644 & 0.237 & 3.735 & 7.988\ 6.90 & 6.399 & 4.912 & 3.740 & 2.770 & 1.997 & 1.279 & 0.770 & 0.117 & 2.753 & 6.080\ 7.00 & 7.861 & 6.171 & 4.795 & 3.614 & 2.617 & 1.663 & 0.916 & 0.073 & 2.012 & 4.580\ 7.10 & 9.207 & 7.325 & 5.758 & 4.381 & 3.177 & 2.010 & 1.054 & 0.063 & 1.437 & 3.378\ 7.20 &10.467 & 8.404 & 6.657 & 5.097 & 3.699 & 2.339 & 1.197 & 0.083 & 0.994 & 2.414\ 7.30 &11.684 & 9.448 & 7.531 & 5.796 & 4.217 & 2.679 & 1.369 & 0.144 & 0.669 & 1.653\ 7.40 &12.897 & 10.497 & 8.417 & 6.518 & 4.766 & 3.064 & 1.597 & 0.264 & 0.462 & 1.076\ 7.50 &14.135 & 11.578 & 9.343 & 7.286 & 5.372 & 3.514 & 1.900 & 0.458 & 0.372 & 0.670\ 7.60 &15.000 & 12.698 & 10.313 & 8.105 & 6.035 & 4.029 & 2.275 & 0.720 & 0.384 & 0.408\ 7.70 &15.000 & 13.837 & 11.308 & 8.956 & 6.735 & 4.588 & 2.698 & 1.028 & 0.466 & 0.250\ 7.80 &15.000 & 14.975 & 12.306 & 9.813 & 7.449 & 5.165 & 3.141 & 1.354 & 0.586 & 0.157\ 7.90 &15.000 & 15.000 & 13.291 & 10.662 & 8.158 & 5.741 & 3.587 & 1.680 & 0.723 & 0.103\ 8.00 &15.000 & 15.000 & 14.256 & 11.494 & 8.855 & 6.308 & 4.024 & 1.998 & 0.864 & 0.069\ 8.10 &15.000 & 15.000 & 15.000 & 12.307 & 9.536 & 6.861 & 4.451 & 2.304 & 1.003 & 0.048\ 8.20 &15.000 & 15.000 & 15.000 & 13.101 & 10.201 & 7.401 & 4.866 & 2.598 & 1.139 & 0.034\ 8.30 &15.000 & 15.000 & 15.000 & 13.879 & 10.852 & 7.929 & 5.270 & 2.882 & 1.272 & 0.024\ 8.40 &15.000 & 15.000 & 15.000 & 14.641 & 11.491 & 8.446 & 5.665 & 3.159 & 1.402 & 0.018\ 8.50 &15.000 & 15.000 & 15.000 & 15.000 & 12.119 & 8.955 & 6.053 & 3.429 & 1.529 & 0.013\ 8.60 &15.000 & 15.000 & 15.000 & 15.000 & 12.741 & 9.458 & 6.437 & 3.697 & 1.656 & 0.010\ 8.70 &15.000 & 15.000 & 15.000 & 15.000 & 13.356 & 9.957 & 6.818 & 3.962 & 1.782 & 0.007\ 8.80 &15.000 & 15.000 & 15.000 & 15.000 & 13.969 & 10.455 & 7.199 & 4.229 & 1.910 & 0.005\ 8.90 &15.000 & 15.000 & 15.000 & 15.000 & 14.582 & 10.953 & 7.582 & 4.498 & 2.039 & 0.004\ 9.00 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 11.453 & 7.967 & 4.770 & 2.170 & 0.003\ [cccccccccccc]{} 4.00 & 1.131 & 0.040 & 1.884 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 1.734 & 0.129 & 0.624 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 2.747 & 0.697 & 0.098 & 12.421 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 3.945 & 1.505 & 0.014 & 9.346 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 5.015 & 2.230 & 0.003 & 6.564 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 5.931 & 2.835 & 0.001 & 4.342 & 11.865 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 6.613 & 3.234 & 0.001 & 2.634 & 7.897 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 6.908 & 3.269 & 0.013 & 1.534 & 5.230 & 10.863 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 7.041 & 3.158 & 0.058 & 0.909 & 3.559 & 7.682 & 14.963 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 7.243 & 3.131 & 0.158 & 0.523 & 2.382 & 5.401 & 11.120 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 7.571 & 3.242 & 0.335 & 0.296 & 1.509 & 3.655 & 8.098 & 13.758 & 15.000 & 15.000 & 15.000\ 5.10 & 8.034 & 3.498 & 0.603 & 0.212 & 0.882 & 2.306 & 5.698 & 10.120 & 15.000 & 15.000 & 15.000\ 5.20 & 8.647 & 3.913 & 0.988 & 0.283 & 0.486 & 1.304 & 3.822 & 7.219 & 11.313 & 15.000 & 15.000\ 5.30 & 9.445 & 4.521 & 1.535 & 0.546 & 0.342 & 0.642 & 2.426 & 4.966 & 8.027 & 12.408 & 15.000\ 5.40 & 10.440 & 5.331 & 2.260 & 1.010 & 0.447 & 0.297 & 1.460 & 3.278 & 5.488 & 8.800 & 11.544\ 5.50 & 11.582 & 6.290 & 3.116 & 1.625 & 0.740 & 0.197 & 0.828 & 2.032 & 3.533 & 5.960 & 7.713\ 5.60 & 12.841 & 7.369 & 4.081 & 2.364 & 1.188 & 0.295 & 0.465 & 1.143 & 2.048 & 3.728 & 4.671\ 5.70 & 14.268 & 8.623 & 5.211 & 3.281 & 1.837 & 0.631 & 0.398 & 0.618 & 1.013 & 2.054 & 2.354\ 5.80 & 15.000 & 10.216 & 6.673 & 4.541 & 2.851 & 1.360 & 0.770 & 0.587 & 0.539 & 1.021 & 0.847\ 5.90 & 15.000 & 12.331 & 8.654 & 6.329 & 4.410 & 2.658 & 1.747 & 1.206 & 0.767 & 0.756 & 0.245\ 6.00 & 15.000 & 14.618 & 10.804 & 8.297 & 6.163 & 4.170 & 2.967 & 2.102 & 1.315 & 0.869 & 0.102\ 6.10 & 15.000 & 15.000 & 12.852 & 10.169 & 7.833 & 5.616 & 4.146 & 2.986 & 1.885 & 1.062 & 0.083\ 6.20 & 15.000 & 15.000 & 14.785 & 11.935 & 9.408 & 6.981 & 5.264 & 3.834 & 2.451 & 1.301 & 0.134\ 6.30 & 15.000 & 15.000 & 15.000 & 13.651 & 10.943 & 8.319 & 6.373 & 4.694 & 3.057 & 1.624 & 0.283\ 6.40 & 15.000 & 15.000 & 15.000 & 15.000 & 12.523 & 9.712 & 7.553 & 5.645 & 3.776 & 2.096 & 0.590\ 6.50 & 15.000 & 15.000 & 15.000 & 15.000 & 14.233 & 11.246 & 8.889 & 6.767 & 4.688 & 2.789 & 1.123\ 6.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.987 & 10.442 & 8.122 & 5.849 & 3.753 & 1.931\ 6.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.286 & 9.781 & 7.327 & 5.051 & 3.073\ 6.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.496 & 11.816 & 9.194 & 6.750 & 4.618\ 6.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.041 & 11.260 & 8.657 & 6.373\ 7.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.274 & 10.520 & 8.084\ 7.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.260 & 9.675\ 7.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.887 & 11.154\ 7.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.556\ 7.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.921\ 7.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.00 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.30 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.40 &11.544 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.50 & 7.713 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.60 & 4.671 & 10.899 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.70 & 2.354 & 7.006 & 12.708 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.80 & 0.847 & 4.254 & 8.535 & 13.807 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.90 & 0.245 & 2.674 & 5.826 & 9.807 & 14.462 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.00 & 0.102 & 1.769 & 4.029 & 6.975 & 10.477 & 14.681 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.10 & 0.083 & 1.151 & 2.707 & 4.828 & 7.412 & 10.588 & 14.322 & 15.000 & 15.000 & 15.000 & 15.000\ 6.20 & 0.134 & 0.719 & 1.714 & 3.181 & 5.035 & 7.386 & 10.187 & 13.397 & 15.000 & 15.000 & 15.000\ 6.30 & 0.283 & 0.466 & 1.002 & 1.943 & 3.215 & 4.902 & 6.942 & 9.324 & 11.723 & 15.000 & 15.000\ 6.40 & 0.590 & 0.427 & 0.577 & 1.086 & 1.882 & 3.029 & 4.445 & 6.158 & 7.837 & 15.000 & 15.000\ 6.50 & 1.123 & 0.655 & 0.470 & 0.614 & 1.011 & 1.711 & 2.612 & 3.775 & 4.860 & 13.344 & 15.000\ 6.60 & 1.931 & 1.188 & 0.705 & 0.533 & 0.588 & 0.910 & 1.378 & 2.084 & 2.673 & 9.357 & 15.000\ 6.70 & 3.073 & 2.080 & 1.329 & 0.876 & 0.632 & 0.626 & 0.723 & 1.042 & 1.217 & 6.449 & 12.516\ 6.80 & 4.618 & 3.393 & 2.397 & 1.690 & 1.179 & 0.884 & 0.660 & 0.645 & 0.478 & 4.534 & 9.355\ 6.90 & 6.373 & 4.934 & 3.709 & 2.771 & 2.018 & 1.463 & 0.957 & 0.652 & 0.207 & 3.311 & 7.118\ 7.00 & 8.084 & 6.444 & 5.007 & 3.856 & 2.881 & 2.091 & 1.334 & 0.776 & 0.112 & 2.445 & 5.423\ 7.10 & 9.675 & 7.846 & 6.210 & 4.861 & 3.681 & 2.674 & 1.694 & 0.913 & 0.076 & 1.784 & 4.080\ 7.20 &11.154 & 9.147 & 7.322 & 5.788 & 4.417 & 3.210 & 2.027 & 1.048 & 0.073 & 1.270 & 3.001\ 7.30 &12.556 & 10.381 & 8.378 & 6.667 & 5.117 & 3.721 & 2.354 & 1.198 & 0.101 & 0.877 & 2.136\ 7.40 &13.921 & 11.586 & 9.413 & 7.535 & 5.815 & 4.243 & 2.704 & 1.385 & 0.175 & 0.595 & 1.457\ 7.50 &15.000 & 12.796 & 10.462 & 8.425 & 6.543 & 4.803 & 3.104 & 1.633 & 0.311 & 0.426 & 0.949\ 7.60 &15.000 & 14.034 & 11.545 & 9.354 & 7.318 & 5.418 & 3.568 & 1.952 & 0.517 & 0.365 & 0.596\ 7.70 &15.000 & 15.000 & 12.659 & 10.321 & 8.136 & 6.083 & 4.088 & 2.334 & 0.783 & 0.394 & 0.370\ 7.80 &15.000 & 15.000 & 13.785 & 11.306 & 8.978 & 6.778 & 4.643 & 2.753 & 1.085 & 0.480 & 0.233\ 7.90 &15.000 & 15.000 & 14.905 & 12.288 & 9.821 & 7.480 & 5.210 & 3.187 & 1.399 & 0.598 & 0.150\ 8.00 &15.000 & 15.000 & 15.000 & 13.256 & 10.655 & 8.175 & 5.774 & 3.621 & 1.713 & 0.730 & 0.100\ 8.10 &15.000 & 15.000 & 15.000 & 14.204 & 11.471 & 8.858 & 6.328 & 4.048 & 2.019 & 0.865 & 0.069\ 8.20 &15.000 & 15.000 & 15.000 & 15.000 & 12.271 & 9.527 & 6.870 & 4.464 & 2.314 & 0.999 & 0.048\ 8.30 &15.000 & 15.000 & 15.000 & 15.000 & 13.053 & 10.181 & 7.400 & 4.869 & 2.600 & 1.131 & 0.035\ 8.40 &15.000 & 15.000 & 15.000 & 15.000 & 13.819 & 10.823 & 7.919 & 5.266 & 2.877 & 1.260 & 0.025\ 8.50 &15.000 & 15.000 & 15.000 & 15.000 & 14.572 & 11.453 & 8.429 & 5.655 & 3.148 & 1.387 & 0.019\ 8.60 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.075 & 8.932 & 6.038 & 3.415 & 1.513 & 0.014\ 8.70 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.691 & 9.431 & 6.419 & 3.680 & 1.639 & 0.010\ 8.80 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.303 & 9.927 & 6.798 & 3.945 & 1.766 & 0.008\ 8.90 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.914 & 10.423 & 7.178 & 4.212 & 1.893 & 0.006\ 9.00 &15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.525 & 10.921 & 7.561 & 4.482 & 2.023 & 0.004\ [cccccccccccc]{} 4.00 & 1.265 & 0.028 & 2.130 & 10.843 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 1.650 & 0.050 & 1.060 & 7.050 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 2.069 & 0.144 & 0.563 & 4.354 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 2.530 & 0.307 & 0.302 & 2.313 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 3.097 & 0.599 & 0.231 & 0.797 & 12.234 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 4.172 & 1.417 & 0.721 & 0.113 & 8.814 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 5.562 & 2.567 & 1.566 & 0.013 & 6.185 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 6.727 & 3.505 & 2.217 & 0.003 & 4.198 & 10.998 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 7.469 & 4.033 & 2.472 & 0.002 & 2.791 & 8.006 & 14.555 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 7.958 & 4.318 & 2.498 & 0.007 & 1.867 & 5.835 & 10.816 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 8.406 & 4.571 & 2.505 & 0.029 & 1.210 & 4.166 & 7.942 & 13.790 & 15.000 & 15.000 & 15.000\ 5.10 & 8.895 & 4.874 & 2.575 & 0.093 & 0.726 & 2.845 & 5.647 & 10.198 & 15.000 & 15.000 & 15.000\ 5.20 & 9.471 & 5.272 & 2.751 & 0.232 & 0.402 & 1.821 & 3.816 & 7.302 & 11.914 & 15.000 & 15.000\ 5.30 & 10.157 & 5.787 & 3.054 & 0.474 & 0.241 & 1.069 & 2.384 & 4.988 & 8.537 & 12.653 & 15.000\ 5.40 & 10.978 & 6.442 & 3.507 & 0.850 & 0.261 & 0.583 & 1.317 & 3.184 & 5.845 & 8.939 & 13.185\ 5.50 & 11.991 & 7.295 & 4.166 & 1.417 & 0.510 & 0.394 & 0.628 & 1.869 & 3.786 & 6.037 & 9.247\ 5.60 & 13.229 & 8.378 & 5.062 & 2.210 & 1.014 & 0.514 & 0.310 & 1.018 & 2.303 & 3.852 & 6.200\ 5.70 & 14.652 & 9.649 & 6.155 & 3.187 & 1.727 & 0.885 & 0.296 & 0.542 & 1.286 & 2.242 & 3.861\ 5.80 & 15.000 & 11.107 & 7.440 & 4.348 & 2.643 & 1.495 & 0.563 & 0.403 & 0.678 & 1.125 & 2.114\ 5.90 & 15.000 & 12.894 & 9.062 & 5.838 & 3.906 & 2.479 & 1.238 & 0.718 & 0.580 & 0.583 & 1.019\ 6.00 & 15.000 & 15.000 & 11.204 & 7.848 & 5.703 & 4.021 & 2.500 & 1.656 & 1.149 & 0.761 & 0.711\ 6.10 & 15.000 & 15.000 & 13.516 & 10.031 & 7.685 & 5.769 & 3.990 & 2.851 & 2.013 & 1.275 & 0.802\ 6.20 & 15.000 & 15.000 & 15.000 & 12.108 & 9.572 & 7.437 & 5.420 & 4.011 & 2.872 & 1.821 & 0.986\ 6.30 & 15.000 & 15.000 & 15.000 & 14.063 & 11.348 & 9.008 & 6.771 & 5.112 & 3.698 & 2.367 & 1.221\ 6.40 & 15.000 & 15.000 & 15.000 & 15.000 & 13.071 & 10.539 & 8.096 & 6.207 & 4.540 & 2.958 & 1.546\ 6.50 & 15.000 & 15.000 & 15.000 & 15.000 & 14.831 & 12.117 & 9.481 & 7.377 & 5.479 & 3.670 & 2.025\ 6.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.833 & 11.016 & 8.712 & 6.599 & 4.584 & 2.732\ 6.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.774 & 10.282 & 7.971 & 5.766 & 3.726\ 6.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.849 & 12.180 & 9.683 & 7.302 & 5.089\ 6.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.420 & 11.749 & 9.202 & 6.827\ 7.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.918 & 11.215 & 8.686\ 7.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.149 & 10.472\ 7.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.963 & 12.143\ 7.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.718\ 7.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.50 & 11.859 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.60 & 7.871 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.70 & 4.764 & 10.536 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.80 & 2.420 & 6.733 & 11.937 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.90 & 0.894 & 4.057 & 7.959 & 12.744 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.00 & 0.274 & 2.540 & 5.407 & 9.008 & 13.222 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.10 & 0.121 & 1.687 & 3.740 & 6.399 & 9.566 & 13.314 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.20 & 0.097 & 1.110 & 2.520 & 4.433 & 6.768 & 9.590 & 12.896 & 15.000 & 15.000 & 15.000 & 15.000\ 6.30 & 0.147 & 0.707 & 1.601 & 2.922 & 4.593 & 6.674 & 9.138 & 11.956 & 14.832 & 15.000 & 15.000\ 6.40 & 0.297 & 0.476 & 0.943 & 1.783 & 2.920 & 4.402 & 6.180 & 8.259 & 10.347 & 15.000 & 15.000\ 6.50 & 0.610 & 0.457 & 0.563 & 1.002 & 1.697 & 2.689 & 3.902 & 5.379 & 6.818 & 15.000 & 15.000\ 6.60 & 1.155 & 0.706 & 0.498 & 0.594 & 0.916 & 1.498 & 2.240 & 3.221 & 4.121 & 11.637 & 15.000\ 6.70 & 1.989 & 1.275 & 0.787 & 0.583 & 0.581 & 0.811 & 1.156 & 1.721 & 2.171 & 8.091 & 14.859\ 6.80 & 3.195 & 2.236 & 1.496 & 1.025 & 0.737 & 0.659 & 0.662 & 0.872 & 0.947 & 5.577 & 10.982\ 6.90 & 4.778 & 3.594 & 2.623 & 1.910 & 1.365 & 1.013 & 0.717 & 0.619 & 0.384 & 3.970 & 8.268\ 7.00 & 6.483 & 5.089 & 3.906 & 2.972 & 2.194 & 1.594 & 1.033 & 0.667 & 0.182 & 2.921 & 6.316\ 7.10 & 8.118 & 6.527 & 5.147 & 4.008 & 3.015 & 2.190 & 1.392 & 0.789 & 0.107 & 2.158 & 4.813\ 7.20 & 9.640 & 7.864 & 6.298 & 4.969 & 3.778 & 2.743 & 1.733 & 0.920 & 0.080 & 1.571 & 3.615\ 7.30 & 11.067 & 9.117 & 7.376 & 5.867 & 4.490 & 3.262 & 2.058 & 1.057 & 0.084 & 1.115 & 2.650\ 7.40 & 12.434 & 10.319 & 8.412 & 6.733 & 5.181 & 3.770 & 2.389 & 1.216 & 0.124 & 0.769 & 1.877\ 7.50 & 13.778 & 11.506 & 9.440 & 7.599 & 5.880 & 4.297 & 2.751 & 1.420 & 0.213 & 0.531 & 1.276\ 7.60 & 15.000 & 12.706 & 10.488 & 8.492 & 6.615 & 4.868 & 3.166 & 1.687 & 0.365 & 0.398 & 0.832\ 7.70 & 15.000 & 13.932 & 11.568 & 9.424 & 7.395 & 5.491 & 3.641 & 2.020 & 0.582 & 0.364 & 0.529\ 7.80 & 15.000 & 15.000 & 12.672 & 10.385 & 8.209 & 6.156 & 4.164 & 2.405 & 0.850 & 0.407 & 0.335\ 7.90 & 15.000 & 15.000 & 13.782 & 11.356 & 9.040 & 6.843 & 4.713 & 2.821 & 1.145 & 0.497 & 0.216\ 8.00 & 15.000 & 15.000 & 14.883 & 12.322 & 9.869 & 7.533 & 5.270 & 3.246 & 1.450 & 0.614 & 0.142\ 8.10 & 15.000 & 15.000 & 15.000 & 13.273 & 10.687 & 8.215 & 5.822 & 3.669 & 1.753 & 0.741 & 0.097\ 8.20 & 15.000 & 15.000 & 15.000 & 14.206 & 11.490 & 8.886 & 6.365 & 4.085 & 2.048 & 0.871 & 0.067\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 12.276 & 9.543 & 6.897 & 4.492 & 2.335 & 1.001 & 0.048\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 13.047 & 10.188 & 7.419 & 4.890 & 2.613 & 1.130 & 0.035\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 13.803 & 10.821 & 7.930 & 5.280 & 2.885 & 1.257 & 0.025\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 14.548 & 11.445 & 8.435 & 5.665 & 3.153 & 1.383 & 0.019\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.062 & 8.934 & 6.045 & 3.418 & 1.509 & 0.014\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.674 & 9.430 & 6.424 & 3.682 & 1.635 & 0.010\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.285 & 9.926 & 6.803 & 3.948 & 1.763 & 0.008\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.895 & 10.422 & 7.185 & 4.217 & 1.893 & 0.006\ [ccccccccccccc]{} 4.00 & 2.503 & 0.002 & 3.397 & 13.345 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 2.796 & 0.007 & 1.859 & 8.946 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 2.907 & 0.098 & 0.698 & 5.553 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 3.342 & 0.530 & 0.152 & 3.457 & 10.413 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 3.880 & 1.042 & 0.044 & 2.205 & 6.796 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 4.223 & 1.331 & 0.044 & 1.296 & 3.988 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 4.554 & 1.587 & 0.137 & 0.640 & 1.805 & 12.271 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 5.233 & 2.173 & 0.578 & 0.452 & 0.425 & 8.688 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 6.452 & 3.286 & 1.539 & 0.874 & 0.077 & 6.597 & 14.611 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 7.505 & 4.223 & 2.316 & 1.181 & 0.032 & 5.141 & 11.388 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 8.383 & 4.978 & 2.904 & 1.355 & 0.020 & 3.975 & 8.788 & 14.142 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 9.196 & 5.664 & 3.418 & 1.500 & 0.015 & 3.017 & 6.658 & 10.755 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 9.978 & 6.314 & 3.896 & 1.645 & 0.013 & 2.225 & 4.900 & 7.976 & 12.432 & 15.000 & 15.000 & 15.000\ 5.30 & 10.743 & 6.944 & 4.352 & 1.800 & 0.019 & 1.570 & 3.444 & 5.673 & 9.082 & 13.619 & 15.000 & 15.000\ 5.40 & 11.507 & 7.572 & 4.809 & 1.981 & 0.049 & 1.040 & 2.247 & 3.764 & 6.302 & 9.795 & 13.791 & 15.000\ 5.50 & 12.314 & 8.244 & 5.311 & 2.229 & 0.144 & 0.656 & 1.304 & 2.216 & 4.023 & 6.647 & 9.654 & 13.659\ 5.60 & 13.274 & 9.070 & 5.969 & 2.650 & 0.410 & 0.506 & 0.683 & 1.076 & 2.266 & 4.159 & 6.349 & 9.370\ 5.70 & 14.553 & 10.216 & 6.951 & 3.411 & 1.014 & 0.745 & 0.522 & 0.464 & 1.127 & 2.403 & 3.909 & 6.106\ 5.80 & 15.000 & 11.689 & 8.264 & 4.516 & 1.961 & 1.368 & 0.800 & 0.345 & 0.554 & 1.301 & 2.227 & 3.719\ 5.90 & 15.000 & 13.388 & 9.807 & 5.863 & 3.150 & 2.265 & 1.395 & 0.588 & 0.400 & 0.688 & 1.115 & 1.993\ 6.00 & 15.000 & 15.000 & 11.680 & 7.550 & 4.681 & 3.532 & 2.392 & 1.270 & 0.729 & 0.613 & 0.603 & 0.942\ 6.10 & 15.000 & 15.000 & 14.047 & 9.740 & 6.717 & 5.324 & 3.941 & 2.532 & 1.675 & 1.199 & 0.802 & 0.672\ 6.20 & 15.000 & 15.000 & 15.000 & 12.055 & 8.880 & 7.262 & 5.654 & 3.984 & 2.840 & 2.038 & 1.297 & 0.768\ 6.30 & 15.000 & 15.000 & 15.000 & 14.250 & 10.925 & 9.097 & 7.281 & 5.370 & 3.962 & 2.867 & 1.820 & 0.952\ 6.40 & 15.000 & 15.000 & 15.000 & 15.000 & 12.849 & 10.826 & 8.814 & 6.680 & 5.029 & 3.666 & 2.347 & 1.192\ 6.50 & 15.000 & 15.000 & 15.000 & 15.000 & 14.719 & 12.512 & 10.315 & 7.973 & 6.098 & 4.489 & 2.927 & 1.525\ 6.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.246 & 11.874 & 9.338 & 7.254 & 5.419 & 3.639 & 2.021\ 6.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.590 & 10.871 & 8.591 & 6.548 & 4.569 & 2.757\ 6.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.658 & 10.194 & 7.958 & 5.796 & 3.806\ 6.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.800 & 12.164 & 9.746 & 7.415 & 5.259\ 7.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.422 & 11.834 & 9.342 & 7.030\ 7.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.950 & 11.306 & 8.844\ 7.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.175 & 10.570\ 7.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.938 & 12.192\ 7.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.734\ 7.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.60 & 11.788 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.70 & 7.638 & 14.578 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.80 & 4.538 & 9.791 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.90 & 2.271 & 6.193 & 10.907 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.00 & 0.830 & 3.713 & 7.240 & 11.509 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.10 & 0.270 & 2.345 & 4.938 & 8.148 & 11.943 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.20 & 0.130 & 1.572 & 3.431 & 5.804 & 8.655 & 12.012 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.30 & 0.110 & 1.044 & 2.320 & 4.026 & 6.127 & 8.650 & 11.562 & 14.835 & 15.000 & 15.000 & 15.000\ 6.40 & 0.165 & 0.676 & 1.478 & 2.648 & 4.147 & 6.001 & 8.157 & 10.619 & 13.123 & 15.000 & 15.000\ 6.50 & 0.324 & 0.474 & 0.879 & 1.608 & 2.616 & 3.926 & 5.462 & 7.266 & 9.064 & 15.000 & 15.000\ 6.60 & 0.653 & 0.487 & 0.550 & 0.907 & 1.503 & 2.363 & 3.387 & 4.651 & 5.865 & 14.241 & 15.000\ 6.70 & 1.227 & 0.777 & 0.540 & 0.575 & 0.819 & 1.297 & 1.892 & 2.705 & 3.432 & 10.063 & 15.000\ 6.80 & 2.117 & 1.410 & 0.904 & 0.655 & 0.591 & 0.738 & 0.965 & 1.397 & 1.717 & 6.938 & 12.943\ 6.90 & 3.413 & 2.469 & 1.720 & 1.215 & 0.876 & 0.732 & 0.640 & 0.743 & 0.725 & 4.804 & 9.606\ 7.00 & 5.030 & 3.866 & 2.894 & 2.157 & 1.569 & 1.163 & 0.791 & 0.607 & 0.313 & 3.466 & 7.287\ 7.10 & 6.691 & 5.323 & 4.145 & 3.194 & 2.379 & 1.735 & 1.114 & 0.679 & 0.162 & 2.564 & 5.584\ 7.20 & 8.266 & 6.706 & 5.336 & 4.187 & 3.162 & 2.300 & 1.456 & 0.799 & 0.104 & 1.893 & 4.253\ 7.30 & 9.740 & 7.999 & 6.448 & 5.114 & 3.894 & 2.829 & 1.783 & 0.928 & 0.086 & 1.373 & 3.185\ 7.40 & 11.135 & 9.225 & 7.503 & 5.994 & 4.589 & 3.336 & 2.105 & 1.072 & 0.100 & 0.970 & 2.323\ 7.50 & 12.487 & 10.415 & 8.530 & 6.855 & 5.277 & 3.846 & 2.444 & 1.247 & 0.153 & 0.671 & 1.637\ 7.60 & 13.825 & 11.600 & 9.559 & 7.725 & 5.983 & 4.384 & 2.821 & 1.471 & 0.259 & 0.474 & 1.109\ 7.70 & 15.000 & 12.802 & 10.611 & 8.626 & 6.726 & 4.967 & 3.253 & 1.757 & 0.427 & 0.377 & 0.726\ 7.80 & 15.000 & 14.026 & 11.691 & 9.560 & 7.510 & 5.598 & 3.739 & 2.103 & 0.653 & 0.367 & 0.468\ 7.90 & 15.000 & 15.000 & 12.788 & 10.515 & 8.320 & 6.262 & 4.264 & 2.491 & 0.922 & 0.422 & 0.302\ 8.00 & 15.000 & 15.000 & 13.885 & 11.475 & 9.140 & 6.940 & 4.807 & 2.901 & 1.211 & 0.516 & 0.199\ 8.10 & 15.000 & 15.000 & 14.969 & 12.426 & 9.955 & 7.618 & 5.353 & 3.318 & 1.506 & 0.631 & 0.134\ 8.20 & 15.000 & 15.000 & 15.000 & 13.362 & 10.759 & 8.288 & 5.894 & 3.731 & 1.798 & 0.755 & 0.093\ 8.30 & 15.000 & 15.000 & 15.000 & 14.280 & 11.547 & 8.947 & 6.426 & 4.137 & 2.084 & 0.881 & 0.065\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 12.321 & 9.593 & 6.948 & 4.536 & 2.363 & 1.008 & 0.047\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 13.081 & 10.228 & 7.461 & 4.926 & 2.636 & 1.135 & 0.034\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 13.828 & 10.854 & 7.967 & 5.311 & 2.904 & 1.260 & 0.025\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 14.566 & 11.472 & 8.466 & 5.692 & 3.169 & 1.386 & 0.019\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.085 & 8.962 & 6.070 & 3.433 & 1.512 & 0.014\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.695 & 9.457 & 6.449 & 3.698 & 1.640 & 0.010\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.304 & 9.952 & 6.829 & 3.966 & 1.770 & 0.007\ [ccccccccccccc]{} 4.00 & 2.694 & 0.001 & 3.460 & 14.920 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 3.381 & 0.007 & 1.820 & 10.534 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 3.666 & 0.117 & 0.626 & 6.793 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 4.056 & 0.553 & 0.143 & 4.255 & 11.807 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 4.432 & 0.972 & 0.051 & 2.508 & 7.532 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 4.725 & 1.267 & 0.052 & 1.227 & 4.220 & 11.611 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 5.127 & 1.635 & 0.194 & 0.491 & 1.845 & 6.896 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 5.860 & 2.305 & 0.658 & 0.392 & 0.432 & 3.626 & 14.037 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 7.158 & 3.517 & 1.672 & 0.990 & 0.064 & 1.877 & 9.549 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 8.333 & 4.590 & 2.549 & 1.515 & 0.063 & 0.993 & 6.611 & 13.925 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 9.297 & 5.440 & 3.204 & 1.855 & 0.161 & 0.530 & 4.711 & 10.298 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 10.229 & 6.249 & 3.820 & 2.183 & 0.331 & 0.279 & 3.411 & 7.702 & 13.261 & 15.000 & 15.000 & 15.000\ 5.20 & 11.174 & 7.065 & 4.445 & 2.542 & 0.550 & 0.148 & 2.452 & 5.711 & 9.991 & 15.000 & 15.000 & 15.000\ 5.30 & 12.123 & 7.883 & 5.073 & 2.923 & 0.795 & 0.087 & 1.707 & 4.122 & 7.354 & 11.984 & 15.000 & 15.000\ 5.40 & 13.074 & 8.700 & 5.705 & 3.324 & 1.059 & 0.078 & 1.122 & 2.834 & 5.196 & 8.756 & 13.213 & 15.000\ 5.50 & 14.050 & 9.541 & 6.364 & 3.766 & 1.364 & 0.134 & 0.685 & 1.803 & 3.435 & 6.109 & 9.540 & 13.182\ 5.60 & 15.000 & 10.458 & 7.104 & 4.301 & 1.758 & 0.300 & 0.421 & 1.032 & 2.045 & 3.980 & 6.555 & 9.277\ 5.70 & 15.000 & 11.530 & 8.004 & 5.006 & 2.323 & 0.650 & 0.393 & 0.565 & 1.047 & 2.358 & 4.214 & 6.178\ 5.80 & 15.000 & 12.845 & 9.152 & 5.969 & 3.143 & 1.267 & 0.674 & 0.460 & 0.483 & 1.261 & 2.506 & 3.835\ 5.90 & 15.000 & 14.416 & 10.561 & 7.201 & 4.231 & 2.162 & 1.265 & 0.710 & 0.330 & 0.647 & 1.365 & 2.152\ 6.00 & 15.000 & 15.000 & 12.219 & 8.687 & 5.575 & 3.319 & 2.146 & 1.284 & 0.548 & 0.459 & 0.718 & 1.033\ 6.10 & 15.000 & 15.000 & 14.301 & 10.605 & 7.349 & 4.914 & 3.488 & 2.347 & 1.293 & 0.842 & 0.693 & 0.594\ 6.20 & 15.000 & 15.000 & 15.000 & 13.012 & 9.614 & 7.005 & 5.344 & 3.948 & 2.606 & 1.829 & 1.315 & 0.848\ 6.30 & 15.000 & 15.000 & 15.000 & 15.000 & 11.928 & 9.153 & 7.273 & 5.641 & 4.035 & 2.963 & 2.119 & 1.328\ 6.40 & 15.000 & 15.000 & 15.000 & 15.000 & 14.109 & 11.171 & 9.088 & 7.235 & 5.388 & 4.045 & 2.904 & 1.825\ 6.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.076 & 10.800 & 8.742 & 6.671 & 5.080 & 3.669 & 2.336\ 6.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.938 & 12.481 & 10.230 & 7.951 & 6.131 & 4.474 & 2.913\ 6.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.227 & 11.794 & 9.321 & 7.289 & 5.406 & 3.639\ 6.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.537 & 10.881 & 8.652 & 6.561 & 4.606\ 6.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.739 & 10.325 & 8.041 & 5.911\ 7.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.967 & 12.380 & 9.915 & 7.622\ 7.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.632 & 11.998 & 9.550\ 7.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.054 & 11.457\ 7.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.266\ 7.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.981\ 7.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.60 & 13.026 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.70 & 8.996 & 11.185 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.80 & 5.865 & 7.228 & 13.481 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.90 & 3.502 & 4.212 & 8.940 & 14.534 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.00 & 1.788 & 2.017 & 5.551 & 9.792 & 14.787 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.10 & 0.833 & 0.710 & 3.316 & 6.484 & 10.289 & 14.728 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.20 & 0.646 & 0.252 & 2.135 & 4.463 & 7.321 & 10.701 & 14.598 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.30 & 0.754 & 0.136 & 1.444 & 3.111 & 5.220 & 7.756 & 10.727 & 14.097 & 15.000 & 15.000 & 15.000 & 15.000\ 6.40 & 0.940 & 0.124 & 0.963 & 2.101 & 3.612 & 5.475 & 7.701 & 10.243 & 13.100 & 15.000 & 15.000 & 15.000\ 6.50 & 1.187 & 0.189 & 0.633 & 1.335 & 2.361 & 3.678 & 5.304 & 7.169 & 9.306 & 11.467 & 15.000 & 15.000\ 6.60 & 1.537 & 0.368 & 0.469 & 0.801 & 1.422 & 2.289 & 3.421 & 4.729 & 6.278 & 7.807 & 15.000 & 15.000\ 6.70 & 2.063 & 0.728 & 0.527 & 0.537 & 0.811 & 1.297 & 2.015 & 2.859 & 3.921 & 4.925 & 12.288 & 15.000\ 6.80 & 2.848 & 1.352 & 0.877 & 0.602 & 0.573 & 0.728 & 1.092 & 1.543 & 2.195 & 2.761 & 8.588 & 15.000\ 6.90 & 3.985 & 2.330 & 1.607 & 1.075 & 0.775 & 0.638 & 0.692 & 0.804 & 1.106 & 1.304 & 5.890 & 11.207\ 7.00 & 5.538 & 3.726 & 2.774 & 2.007 & 1.463 & 1.064 & 0.842 & 0.657 & 0.655 & 0.550 & 4.129 & 8.385\ 7.10 & 7.315 & 5.349 & 4.184 & 3.201 & 2.433 & 1.796 & 1.325 & 0.878 & 0.609 & 0.257 & 3.018 & 6.408\ 7.20 & 9.076 & 6.958 & 5.594 & 4.411 & 3.437 & 2.581 & 1.883 & 1.201 & 0.697 & 0.146 & 2.241 & 4.919\ 7.30 & 10.744 & 8.476 & 6.925 & 5.555 & 4.388 & 3.330 & 2.422 & 1.529 & 0.817 & 0.102 & 1.652 & 3.741\ 7.40 & 12.321 & 9.906 & 8.179 & 6.632 & 5.285 & 4.037 & 2.935 & 1.849 & 0.950 & 0.093 & 1.193 & 2.791\ 7.50 & 13.833 & 11.273 & 9.381 & 7.665 & 6.147 & 4.721 & 3.436 & 2.173 & 1.103 & 0.119 & 0.840 & 2.025\ 7.60 & 15.000 & 12.609 & 10.559 & 8.684 & 7.004 & 5.408 & 3.951 & 2.523 & 1.295 & 0.188 & 0.586 & 1.420\ 7.70 & 15.000 & 13.941 & 11.741 & 9.713 & 7.877 & 6.121 & 4.500 & 2.916 & 1.539 & 0.312 & 0.429 & 0.961\ 7.80 & 15.000 & 15.000 & 12.940 & 10.765 & 8.781 & 6.871 & 5.094 & 3.363 & 1.842 & 0.495 & 0.364 & 0.634\ 7.90 & 15.000 & 15.000 & 14.157 & 11.840 & 9.712 & 7.655 & 5.728 & 3.855 & 2.197 & 0.729 & 0.375 & 0.414\ 8.00 & 15.000 & 15.000 & 15.000 & 12.925 & 10.658 & 8.458 & 6.388 & 4.378 & 2.585 & 0.996 & 0.440 & 0.273\ 8.10 & 15.000 & 15.000 & 15.000 & 14.005 & 11.603 & 9.266 & 7.057 & 4.913 & 2.989 & 1.279 & 0.537 & 0.183\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 12.538 & 10.067 & 7.723 & 5.449 & 3.397 & 1.565 & 0.650 & 0.126\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 13.457 & 10.856 & 8.380 & 5.980 & 3.800 & 1.848 & 0.771 & 0.088\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 14.360 & 11.632 & 9.028 & 6.502 & 4.198 & 2.126 & 0.894 & 0.063\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.393 & 9.664 & 7.015 & 4.589 & 2.398 & 1.019 & 0.046\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.143 & 10.291 & 7.521 & 4.974 & 2.666 & 1.144 & 0.033\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.883 & 10.909 & 8.021 & 5.354 & 2.931 & 1.269 & 0.025\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.615 & 11.523 & 8.517 & 5.732 & 3.195 & 1.395 & 0.018\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.133 & 9.011 & 6.110 & 3.459 & 1.523 & 0.013\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.741 & 9.505 & 6.489 & 3.726 & 1.652 & 0.010\ [cccccccccccccc]{} 4.00 & 1.948 & 0.005 & 4.512 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 2.814 & 0.002 & 2.700 & 11.913 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 3.542 & 0.018 & 1.385 & 7.922 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 4.088 & 0.111 & 0.646 & 5.032 & 14.278 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 4.318 & 0.300 & 0.303 & 2.963 & 9.446 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 4.460 & 0.564 & 0.158 & 1.489 & 5.770 & 13.316 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 4.739 & 0.938 & 0.186 & 0.634 & 3.207 & 8.447 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 5.195 & 1.436 & 0.376 & 0.284 & 1.648 & 5.203 & 11.647 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 5.801 & 2.035 & 0.695 & 0.210 & 0.767 & 3.173 & 7.556 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 6.584 & 2.776 & 1.176 & 0.353 & 0.324 & 1.884 & 4.692 & 14.254 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 7.520 & 3.645 & 1.800 & 0.668 & 0.169 & 1.049 & 2.794 & 10.300 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 8.578 & 4.617 & 2.537 & 1.118 & 0.221 & 0.532 & 1.588 & 7.423 & 14.003 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 9.772 & 5.711 & 3.404 & 1.715 & 0.472 & 0.300 & 0.847 & 5.320 & 10.333 & 15.000 & 15.000 & 15.000 & 15.000\ 5.30 & 11.073 & 6.904 & 4.381 & 2.440 & 0.890 & 0.302 & 0.434 & 3.793 & 7.529 & 11.753 & 15.000 & 15.000 & 15.000\ 5.40 & 12.412 & 8.128 & 5.400 & 3.222 & 1.397 & 0.446 & 0.223 & 2.670 & 5.369 & 8.560 & 12.434 & 15.000 & 15.000\ 5.50 & 13.736 & 9.334 & 6.408 & 4.006 & 1.929 & 0.659 & 0.123 & 1.825 & 3.690 & 6.022 & 8.941 & 13.168 & 15.000\ 5.60 & 15.000 & 10.515 & 7.390 & 4.776 & 2.467 & 0.911 & 0.094 & 1.187 & 2.387 & 3.999 & 6.125 & 9.366 & 12.762\ 5.70 & 15.000 & 11.696 & 8.374 & 5.559 & 3.034 & 1.220 & 0.148 & 0.735 & 1.410 & 2.410 & 3.873 & 6.291 & 8.832\ 5.80 & 15.000 & 12.962 & 9.451 & 6.444 & 3.716 & 1.664 & 0.355 & 0.519 & 0.769 & 1.246 & 2.149 & 3.876 & 5.704\ 5.90 & 15.000 & 14.468 & 10.776 & 7.587 & 4.666 & 2.395 & 0.863 & 0.663 & 0.560 & 0.584 & 1.008 & 2.149 & 3.373\ 6.00 & 15.000 & 15.000 & 12.449 & 9.084 & 5.980 & 3.503 & 1.760 & 1.242 & 0.835 & 0.461 & 0.471 & 1.107 & 1.810\ 6.10 & 15.000 & 15.000 & 14.418 & 10.884 & 7.605 & 4.934 & 2.990 & 2.188 & 1.511 & 0.785 & 0.430 & 0.626 & 0.873\ 6.20 & 15.000 & 15.000 & 15.000 & 13.099 & 9.652 & 6.796 & 4.660 & 3.601 & 2.677 & 1.637 & 0.955 & 0.762 & 0.607\ 6.30 & 15.000 & 15.000 & 15.000 & 15.000 & 12.060 & 9.029 & 6.708 & 5.412 & 4.261 & 2.936 & 1.959 & 1.418 & 0.909\ 6.40 & 15.000 & 15.000 & 15.000 & 15.000 & 14.446 & 11.246 & 8.747 & 7.232 & 5.869 & 4.284 & 3.038 & 2.185 & 1.365\ 6.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.332 & 10.663 & 8.943 & 7.380 & 5.555 & 4.062 & 2.928 & 1.834\ 6.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.484 & 10.571 & 8.819 & 6.773 & 5.051 & 3.663 & 2.326\ 6.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.281 & 12.187 & 10.255 & 8.003 & 6.070 & 4.450 & 2.895\ 6.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.891 & 11.788 & 9.345 & 7.214 & 5.380 & 3.629\ 6.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.544 & 10.921 & 8.606 & 6.575 & 4.643\ 7.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.861 & 10.372 & 8.156 & 6.056\ 7.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.495 & 10.107 & 7.848\ 7.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.718 & 12.167 & 9.757\ 7.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.165 & 11.608\ 7.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.367\ 7.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.70 & 12.294 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.80 & 8.285 & 10.222 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.90 & 5.200 & 6.382 & 11.932 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.00 & 2.983 & 3.587 & 7.785 & 12.827 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.10 & 1.478 & 1.660 & 4.804 & 8.646 & 13.087 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.20 & 0.727 & 0.594 & 2.921 & 5.819 & 9.204 & 13.075 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.30 & 0.623 & 0.240 & 1.923 & 4.079 & 6.625 & 9.570 & 13.016 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.40 & 0.741 & 0.144 & 1.308 & 2.873 & 4.752 & 6.957 & 9.581 & 12.501 & 15.000 & 15.000 & 15.000 & 15.000\ 6.50 & 0.931 & 0.142 & 0.874 & 1.955 & 3.296 & 4.906 & 6.867 & 9.054 & 11.535 & 14.066 & 15.000 & 15.000\ 6.60 & 1.186 & 0.218 & 0.581 & 1.257 & 2.153 & 3.275 & 4.694 & 6.281 & 8.122 & 9.971 & 15.000 & 15.000\ 6.70 & 1.549 & 0.412 & 0.451 & 0.777 & 1.296 & 2.008 & 2.977 & 4.067 & 5.381 & 6.665 & 14.783 & 15.000\ 6.80 & 2.099 & 0.796 & 0.546 & 0.566 & 0.758 & 1.117 & 1.704 & 2.375 & 3.250 & 4.061 & 10.514 & 15.000\ 6.90 & 2.943 & 1.479 & 0.967 & 0.712 & 0.614 & 0.665 & 0.921 & 1.233 & 1.733 & 2.151 & 7.257 & 13.101\ 7.00 & 4.197 & 2.575 & 1.823 & 1.319 & 0.960 & 0.735 & 0.699 & 0.699 & 0.877 & 0.965 & 4.980 & 9.679\ 7.10 & 5.837 & 4.059 & 3.085 & 2.354 & 1.757 & 1.283 & 0.985 & 0.710 & 0.605 & 0.423 & 3.551 & 7.317\ 7.20 & 7.599 & 5.667 & 4.487 & 3.545 & 2.730 & 2.028 & 1.493 & 0.972 & 0.620 & 0.217 & 2.624 & 5.624\ 7.30 & 9.308 & 7.225 & 5.851 & 4.713 & 3.695 & 2.784 & 2.031 & 1.290 & 0.718 & 0.134 & 1.951 & 4.320\ 7.40 & 10.928 & 8.695 & 7.140 & 5.817 & 4.610 & 3.503 & 2.548 & 1.608 & 0.840 & 0.102 & 1.433 & 3.277\ 7.50 & 12.471 & 10.091 & 8.365 & 6.868 & 5.481 & 4.191 & 3.048 & 1.925 & 0.979 & 0.105 & 1.030 & 2.434\ 7.60 & 13.963 & 11.439 & 9.552 & 7.889 & 6.333 & 4.869 & 3.549 & 2.256 & 1.146 & 0.145 & 0.725 & 1.756\ 7.70 & 15.000 & 12.767 & 10.726 & 8.905 & 7.187 & 5.560 & 4.072 & 2.619 & 1.356 & 0.231 & 0.514 & 1.226\ 7.80 & 15.000 & 14.096 & 11.908 & 9.935 & 8.064 & 6.280 & 4.632 & 3.029 & 1.620 & 0.372 & 0.395 & 0.831\ 7.90 & 15.000 & 15.000 & 13.107 & 10.988 & 8.969 & 7.035 & 5.235 & 3.488 & 1.938 & 0.568 & 0.359 & 0.552\ 8.00 & 15.000 & 15.000 & 14.318 & 12.058 & 9.896 & 7.818 & 5.872 & 3.985 & 2.299 & 0.807 & 0.388 & 0.367\ 8.10 & 15.000 & 15.000 & 15.000 & 13.131 & 10.831 & 8.613 & 6.526 & 4.504 & 2.686 & 1.072 & 0.461 & 0.246\ 8.20 & 15.000 & 15.000 & 15.000 & 14.196 & 11.762 & 9.409 & 7.185 & 5.031 & 3.084 & 1.348 & 0.560 & 0.168\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 12.682 & 10.197 & 7.840 & 5.558 & 3.483 & 1.627 & 0.672 & 0.117\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 13.587 & 10.973 & 8.487 & 6.078 & 3.878 & 1.902 & 0.791 & 0.083\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 14.476 & 11.736 & 9.124 & 6.592 & 4.268 & 2.174 & 0.913 & 0.060\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.488 & 9.752 & 7.098 & 4.652 & 2.441 & 1.036 & 0.044\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.229 & 10.371 & 7.598 & 5.033 & 2.706 & 1.160 & 0.032\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.963 & 10.985 & 8.093 & 5.411 & 2.969 & 1.286 & 0.024\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.690 & 11.595 & 8.587 & 5.787 & 3.233 & 1.413 & 0.017\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.203 & 9.080 & 6.165 & 3.498 & 1.542 & 0.013\ [cccccccccccccc]{} 4.00 & 1.767 & 0.008 & 4.065 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 2.611 & 0.003 & 2.314 & 12.339 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 3.110 & 0.052 & 0.953 & 8.212 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 3.591 & 0.415 & 0.211 & 5.152 & 14.884 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 4.200 & 0.987 & 0.048 & 3.176 & 10.057 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 4.607 & 1.345 & 0.026 & 1.876 & 6.516 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 4.934 & 1.597 & 0.051 & 1.071 & 4.079 & 10.443 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 5.321 & 1.887 & 0.148 & 0.565 & 2.470 & 6.819 & 13.200 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 5.841 & 2.293 & 0.364 & 0.282 & 1.389 & 4.247 & 8.631 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 6.534 & 2.857 & 0.736 & 0.217 & 0.689 & 2.516 & 5.445 & 10.337 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 7.415 & 3.601 & 1.283 & 0.373 & 0.316 & 1.393 & 3.343 & 6.684 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 8.471 & 4.512 & 1.996 & 0.732 & 0.224 & 0.706 & 1.951 & 4.231 & 11.427 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 9.686 & 5.576 & 2.860 & 1.276 & 0.377 & 0.364 & 1.045 & 2.575 & 8.123 & 14.353 & 15.000 & 15.000 & 15.000\ 5.30 & 11.052 & 6.790 & 3.874 & 1.999 & 0.754 & 0.317 & 0.525 & 1.473 & 5.679 & 10.391 & 15.000 & 15.000 & 15.000\ 5.40 & 12.548 & 8.130 & 5.018 & 2.875 & 1.322 & 0.517 & 0.317 & 0.784 & 3.898 & 7.380 & 11.483 & 15.000 & 15.000\ 5.50 & 14.127 & 9.553 & 6.247 & 3.856 & 2.025 & 0.894 & 0.338 & 0.396 & 2.625 & 5.118 & 8.211 & 11.920 & 15.000\ 5.60 & 15.000 & 10.999 & 7.505 & 4.883 & 2.798 & 1.373 & 0.503 & 0.206 & 1.719 & 3.429 & 5.681 & 8.484 & 12.419\ 5.70 & 15.000 & 12.433 & 8.756 & 5.918 & 3.598 & 1.908 & 0.754 & 0.142 & 1.077 & 2.171 & 3.716 & 5.766 & 8.767\ 5.80 & 15.000 & 13.874 & 10.021 & 6.978 & 4.440 & 2.504 & 1.092 & 0.198 & 0.656 & 1.261 & 2.207 & 3.627 & 5.849\ 5.90 & 15.000 & 15.000 & 11.381 & 8.144 & 5.401 & 3.238 & 1.588 & 0.435 & 0.494 & 0.697 & 1.129 & 2.016 & 3.584\ 6.00 & 15.000 & 15.000 & 12.976 & 9.554 & 6.617 & 4.241 & 2.369 & 0.976 & 0.691 & 0.553 & 0.540 & 0.971 & 1.980\ 6.10 & 15.000 & 15.000 & 14.877 & 11.280 & 8.159 & 5.580 & 3.501 & 1.883 & 1.295 & 0.858 & 0.455 & 0.487 & 1.015\ 6.20 & 15.000 & 15.000 & 15.000 & 13.307 & 10.010 & 7.240 & 4.963 & 3.134 & 2.273 & 1.569 & 0.819 & 0.499 & 0.603\ 6.30 & 15.000 & 15.000 & 15.000 & 15.000 & 12.308 & 9.357 & 6.892 & 4.861 & 3.752 & 2.803 & 1.743 & 1.105 & 0.836\ 6.40 & 15.000 & 15.000 & 15.000 & 15.000 & 14.870 & 11.745 & 9.101 & 6.879 & 5.540 & 4.363 & 3.022 & 2.097 & 1.494\ 6.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.064 & 11.248 & 8.844 & 7.292 & 5.902 & 4.304 & 3.117 & 2.214\ 6.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.272 & 10.694 & 8.941 & 7.351 & 5.518 & 4.089 & 2.916\ 6.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.471 & 10.529 & 8.749 & 6.698 & 5.046 & 3.629\ 6.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.254 & 12.133 & 10.174 & 7.921 & 6.062 & 4.422\ 6.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.867 & 11.737 & 9.295 & 7.245 & 5.398\ 7.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.580 & 10.962 & 8.731 & 6.693\ 7.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.024 & 10.623 & 8.407\ 7.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.781 & 10.397\ 7.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.954 & 12.411\ 7.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.344\ 7.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.70 & 12.102 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.80 & 8.346 & 11.474 & 14.014 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.90 & 5.376 & 7.676 & 9.365 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.00 & 3.170 & 4.756 & 5.772 & 10.716 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.10 & 1.681 & 2.649 & 3.167 & 6.915 & 11.360 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.20 & 0.813 & 1.253 & 1.408 & 4.224 & 7.614 & 11.506 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.30 & 0.646 & 0.644 & 0.522 & 2.609 & 5.168 & 8.133 & 11.544 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.40 & 0.958 & 0.592 & 0.240 & 1.745 & 3.648 & 5.876 & 8.465 & 11.443 & 14.765 & 15.000 & 15.000 & 15.000 & 15.000\ 6.50 & 1.377 & 0.714 & 0.161 & 1.189 & 2.561 & 4.200 & 6.130 & 8.387 & 10.913 & 13.744 & 15.000 & 15.000 & 15.000\ 6.60 & 1.817 & 0.906 & 0.168 & 0.794 & 1.727 & 2.882 & 4.277 & 5.950 & 7.825 & 9.963 & 12.136 & 15.000 & 15.000\ 6.70 & 2.298 & 1.175 & 0.264 & 0.541 & 1.100 & 1.850 & 2.800 & 3.991 & 5.330 & 6.900 & 8.465 & 15.000 & 15.000\ 6.80 & 2.883 & 1.572 & 0.494 & 0.463 & 0.696 & 1.097 & 1.669 & 2.454 & 3.345 & 4.442 & 5.502 & 12.597 & 15.000\ 6.90 & 3.671 & 2.188 & 0.947 & 0.640 & 0.583 & 0.678 & 0.921 & 1.358 & 1.867 & 2.565 & 3.202 & 8.842 & 15.000\ 7.00 & 4.791 & 3.148 & 1.748 & 1.190 & 0.872 & 0.693 & 0.645 & 0.775 & 0.954 & 1.309 & 1.593 & 6.054 & 11.205\ 7.10 & 6.341 & 4.546 & 2.991 & 2.201 & 1.645 & 1.218 & 0.908 & 0.766 & 0.655 & 0.712 & 0.701 & 4.205 & 8.353\ 7.20 & 8.176 & 6.234 & 4.526 & 3.522 & 2.746 & 2.092 & 1.545 & 1.156 & 0.789 & 0.585 & 0.330 & 3.055 & 6.383\ 7.30 & 10.040 & 7.955 & 6.097 & 4.894 & 3.914 & 3.050 & 2.285 & 1.670 & 1.073 & 0.639 & 0.185 & 2.274 & 4.926\ 7.40 & 11.830 & 9.605 & 7.599 & 6.209 & 5.039 & 3.979 & 3.012 & 2.189 & 1.385 & 0.746 & 0.124 & 1.690 & 3.782\ 7.50 & 13.535 & 11.173 & 9.021 & 7.455 & 6.106 & 4.863 & 3.706 & 2.691 & 1.697 & 0.872 & 0.106 & 1.236 & 2.860\ 7.60 & 15.000 & 12.678 & 10.384 & 8.652 & 7.133 & 5.715 & 4.381 & 3.184 & 2.015 & 1.021 & 0.121 & 0.885 & 2.113\ 7.70 & 15.000 & 14.145 & 11.711 & 9.822 & 8.141 & 6.559 & 5.056 & 3.688 & 2.355 & 1.203 & 0.177 & 0.626 & 1.517\ 7.80 & 15.000 & 15.000 & 13.027 & 10.989 & 9.153 & 7.414 & 5.751 & 4.221 & 2.733 & 1.432 & 0.281 & 0.456 & 1.056\ 7.90 & 15.000 & 15.000 & 14.348 & 12.166 & 10.183 & 8.293 & 6.478 & 4.792 & 3.158 & 1.713 & 0.440 & 0.371 & 0.718\ 8.00 & 15.000 & 15.000 & 15.000 & 13.359 & 11.232 & 9.197 & 7.235 & 5.402 & 3.625 & 2.043 & 0.647 & 0.361 & 0.483\ 8.10 & 15.000 & 15.000 & 15.000 & 14.557 & 12.291 & 10.117 & 8.014 & 6.037 & 4.124 & 2.407 & 0.889 & 0.404 & 0.325\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 13.349 & 11.039 & 8.799 & 6.684 & 4.638 & 2.791 & 1.150 & 0.484 & 0.222\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 14.397 & 11.955 & 9.582 & 7.333 & 5.157 & 3.182 & 1.420 & 0.584 & 0.155\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.859 & 10.357 & 7.977 & 5.674 & 3.573 & 1.691 & 0.695 & 0.109\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.749 & 11.120 & 8.613 & 6.186 & 3.961 & 1.960 & 0.813 & 0.078\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.626 & 11.873 & 9.241 & 6.691 & 4.344 & 2.226 & 0.933 & 0.057\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.615 & 9.861 & 7.190 & 4.724 & 2.490 & 1.056 & 0.042\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.349 & 10.474 & 7.686 & 5.101 & 2.753 & 1.180 & 0.030\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.076 & 11.084 & 8.178 & 5.477 & 3.015 & 1.307 & 0.022\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.800 & 11.691 & 8.670 & 5.854 & 3.280 & 1.435 & 0.016\ [ccccccccccccccc]{} 4.00 & 0.901 & 0.058 & 4.016 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 1.416 & 0.020 & 2.172 & 10.829 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 1.932 & 0.085 & 0.781 & 6.792 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 2.825 & 0.587 & 0.131 & 4.017 & 14.447 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 3.877 & 1.291 & 0.026 & 2.200 & 9.715 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 4.692 & 1.794 & 0.070 & 0.879 & 5.950 & 14.727 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 5.580 & 2.399 & 0.412 & 0.216 & 3.337 & 9.594 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 6.502 & 3.061 & 0.881 & 0.072 & 1.699 & 5.924 & 13.063 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 7.319 & 3.639 & 1.291 & 0.120 & 0.723 & 3.326 & 8.332 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 8.194 & 4.290 & 1.787 & 0.347 & 0.290 & 1.671 & 5.119 & 10.228 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 9.169 & 5.056 & 2.405 & 0.726 & 0.197 & 0.764 & 3.114 & 6.561 & 13.158 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 10.239 & 5.927 & 3.132 & 1.227 & 0.319 & 0.349 & 1.880 & 4.149 & 9.077 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 11.367 & 6.865 & 3.926 & 1.804 & 0.575 & 0.200 & 1.075 & 2.526 & 6.144 & 13.002 & 15.000 & 15.000 & 15.000 & 15.000\ 5.30 & 12.564 & 7.879 & 4.795 & 2.464 & 0.953 & 0.239 & 0.570 & 1.405 & 4.000 & 9.299 & 15.000 & 15.000 & 15.000 & 15.000\ 5.40 & 13.890 & 9.028 & 5.795 & 3.263 & 1.500 & 0.489 & 0.357 & 0.696 & 2.489 & 6.546 & 11.387 & 15.000 & 15.000 & 15.000\ 5.50 & 15.000 & 10.326 & 6.942 & 4.214 & 2.222 & 0.945 & 0.412 & 0.336 & 1.490 & 4.547 & 8.251 & 12.743 & 15.000 & 15.000\ 5.60 & 15.000 & 11.714 & 8.176 & 5.259 & 3.053 & 1.534 & 0.650 & 0.218 & 0.852 & 3.093 & 5.872 & 9.319 & 13.388 & 15.000\ 5.70 & 15.000 & 13.149 & 9.456 & 6.355 & 3.949 & 2.205 & 1.008 & 0.266 & 0.464 & 2.029 & 4.047 & 6.635 & 9.722 & 13.346\ 5.80 & 15.000 & 14.619 & 10.768 & 7.490 & 4.892 & 2.938 & 1.459 & 0.440 & 0.265 & 1.260 & 2.644 & 4.517 & 6.790 & 9.531\ 5.90 & 15.000 & 15.000 & 12.120 & 8.670 & 5.889 & 3.736 & 1.999 & 0.729 & 0.228 & 0.736 & 1.583 & 2.854 & 4.445 & 6.447\ 6.00 & 15.000 & 15.000 & 13.563 & 9.946 & 6.988 & 4.646 & 2.672 & 1.171 & 0.377 & 0.465 & 0.851 & 1.607 & 2.618 & 3.994\ 6.10 & 15.000 & 15.000 & 15.000 & 11.421 & 8.292 & 5.768 & 3.575 & 1.859 & 0.800 & 0.517 & 0.502 & 0.814 & 1.324 & 2.161\ 6.20 & 15.000 & 15.000 & 15.000 & 13.230 & 9.935 & 7.236 & 4.841 & 2.921 & 1.618 & 1.005 & 0.637 & 0.561 & 0.632 & 0.998\ 6.30 & 15.000 & 15.000 & 15.000 & 15.000 & 12.089 & 9.222 & 6.638 & 4.525 & 2.995 & 2.087 & 1.403 & 0.983 & 0.664 & 0.617\ 6.40 & 15.000 & 15.000 & 15.000 & 15.000 & 14.785 & 11.755 & 8.994 & 6.698 & 4.955 & 3.777 & 2.809 & 2.081 & 1.414 & 1.001\ 6.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.395 & 11.469 & 8.996 & 7.052 & 5.628 & 4.400 & 3.394 & 2.413 & 1.678\ 6.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.811 & 11.169 & 9.037 & 7.385 & 5.918 & 4.657 & 3.394 & 2.372\ 6.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.210 & 10.899 & 9.036 & 7.348 & 5.854 & 4.334 & 3.055\ 6.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.696 & 10.636 & 8.743 & 7.033 & 5.279 & 3.770\ 6.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.523 & 12.278 & 10.193 & 8.282 & 6.313 & 4.595\ 7.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.093 & 11.828 & 9.730 & 7.562 & 5.652\ 7.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.803 & 11.529 & 9.177 & 7.088\ 7.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.702 & 11.177 & 8.920\ 7.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.329 & 10.912\ 7.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.874\ 7.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.753\ 7.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.80 & 12.579 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.90 & 8.713 & 11.489 & 13.683 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.00 & 5.597 & 7.595 & 9.022 & 14.711 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.10 & 3.191 & 4.516 & 5.355 & 9.701 & 14.791 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.20 & 1.528 & 2.276 & 2.684 & 5.984 & 9.901 & 14.395 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.30 & 0.710 & 0.974 & 1.058 & 3.542 & 6.535 & 9.992 & 13.851 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.40 & 0.714 & 0.582 & 0.408 & 2.247 & 4.511 & 7.146 & 10.104 & 13.495 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.50 & 1.061 & 0.608 & 0.214 & 1.530 & 3.209 & 5.189 & 7.426 & 10.029 & 12.901 & 15.000 & 15.000 & 15.000 & 15.000\ 6.60 & 1.470 & 0.751 & 0.163 & 1.043 & 2.242 & 3.689 & 5.343 & 7.308 & 9.477 & 11.930 & 14.428 & 15.000 & 15.000\ 6.70 & 1.903 & 0.961 & 0.193 & 0.699 & 1.494 & 2.499 & 3.672 & 5.114 & 6.705 & 8.543 & 10.391 & 15.000 & 15.000\ 6.80 & 2.397 & 1.259 & 0.320 & 0.500 & 0.945 & 1.574 & 2.341 & 3.346 & 4.457 & 5.788 & 7.095 & 14.852 & 15.000\ 6.90 & 3.023 & 1.707 & 0.604 & 0.492 & 0.630 & 0.932 & 1.350 & 1.981 & 2.687 & 3.591 & 4.448 & 10.635 & 15.000\ 7.00 & 3.896 & 2.417 & 1.153 & 0.779 & 0.642 & 0.653 & 0.765 & 1.071 & 1.428 & 1.968 & 2.447 & 7.365 & 12.979\ 7.10 & 5.162 & 3.527 & 2.107 & 1.491 & 1.105 & 0.854 & 0.693 & 0.713 & 0.765 & 0.990 & 1.152 & 5.041 & 9.578\ 7.20 & 6.834 & 5.050 & 3.476 & 2.638 & 2.023 & 1.535 & 1.126 & 0.888 & 0.671 & 0.621 & 0.521 & 3.572 & 7.230\ 7.30 & 8.674 & 6.746 & 5.019 & 3.975 & 3.149 & 2.442 & 1.807 & 1.335 & 0.878 & 0.585 & 0.269 & 2.635 & 5.571\ 7.40 & 10.489 & 8.419 & 6.543 & 5.306 & 4.283 & 3.373 & 2.531 & 1.844 & 1.170 & 0.661 & 0.164 & 1.968 & 4.308\ 7.50 & 12.225 & 10.017 & 7.995 & 6.576 & 5.369 & 4.269 & 3.233 & 2.346 & 1.476 & 0.774 & 0.120 & 1.459 & 3.301\ 7.60 & 13.889 & 11.544 & 9.378 & 7.789 & 6.407 & 5.129 & 3.910 & 2.838 & 1.787 & 0.909 & 0.114 & 1.062 & 2.486\ 7.70 & 15.000 & 13.021 & 10.714 & 8.964 & 7.417 & 5.969 & 4.579 & 3.331 & 2.112 & 1.071 & 0.144 & 0.759 & 1.827\ 7.80 & 15.000 & 14.471 & 12.026 & 10.124 & 8.418 & 6.810 & 5.256 & 3.843 & 2.465 & 1.272 & 0.217 & 0.543 & 1.305\ 7.90 & 15.000 & 15.000 & 13.333 & 11.285 & 9.428 & 7.666 & 5.957 & 4.386 & 2.859 & 1.520 & 0.339 & 0.412 & 0.909\ 8.00 & 15.000 & 15.000 & 14.645 & 12.458 & 10.456 & 8.546 & 6.689 & 4.967 & 3.296 & 1.817 & 0.512 & 0.359 & 0.621\ 8.10 & 15.000 & 15.000 & 15.000 & 13.641 & 11.498 & 9.447 & 7.446 & 5.579 & 3.769 & 2.155 & 0.727 & 0.369 & 0.423\ 8.20 & 15.000 & 15.000 & 15.000 & 14.825 & 12.545 & 10.356 & 8.218 & 6.211 & 4.267 & 2.520 & 0.970 & 0.424 & 0.289\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 13.588 & 11.265 & 8.992 & 6.850 & 4.775 & 2.900 & 1.228 & 0.509 & 0.201\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 14.620 & 12.167 & 9.763 & 7.489 & 5.286 & 3.284 & 1.492 & 0.610 & 0.141\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.057 & 10.525 & 8.123 & 5.795 & 3.668 & 1.757 & 0.721 & 0.101\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.934 & 11.277 & 8.749 & 6.298 & 4.049 & 2.021 & 0.838 & 0.073\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.798 & 12.019 & 9.369 & 6.797 & 4.427 & 2.283 & 0.958 & 0.053\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.753 & 9.982 & 7.291 & 4.803 & 2.544 & 1.081 & 0.039\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.481 & 10.591 & 7.783 & 5.177 & 2.806 & 1.206 & 0.029\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.205 & 11.198 & 8.274 & 5.553 & 3.070 & 1.334 & 0.021\ [ccccccccccccccc]{} 4.00 & 0.710 & 0.094 & 3.841 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 1.222 & 0.033 & 1.903 & 11.225 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 1.786 & 0.129 & 0.620 & 7.184 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 2.575 & 0.511 & 0.162 & 4.407 & 13.899 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 3.360 & 0.934 & 0.056 & 2.437 & 9.135 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 4.112 & 1.357 & 0.070 & 0.978 & 5.423 & 14.976 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 5.084 & 2.028 & 0.389 & 0.236 & 2.859 & 9.674 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 6.196 & 2.864 & 0.910 & 0.085 & 1.265 & 5.886 & 13.098 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 7.354 & 3.767 & 1.526 & 0.251 & 0.389 & 3.278 & 8.237 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 8.622 & 4.801 & 2.295 & 0.664 & 0.119 & 1.733 & 4.920 & 10.389 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 9.862 & 5.827 & 3.069 & 1.129 & 0.106 & 0.852 & 2.737 & 6.434 & 11.503 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 11.124 & 6.889 & 3.893 & 1.673 & 0.268 & 0.398 & 1.408 & 3.824 & 7.216 & 12.581 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 12.484 & 8.062 & 4.838 & 2.358 & 0.625 & 0.269 & 0.670 & 2.208 & 4.378 & 8.327 & 15.000 & 15.000 & 15.000 & 15.000\ 5.30 & 13.958 & 9.359 & 5.915 & 3.191 & 1.166 & 0.401 & 0.327 & 1.224 & 2.549 & 5.452 & 11.857 & 15.000 & 15.000 & 15.000\ 5.40 & 15.000 & 10.765 & 7.107 & 4.152 & 1.861 & 0.740 & 0.269 & 0.653 & 1.370 & 3.457 & 8.416 & 13.943 & 15.000 & 15.000\ 5.50 & 15.000 & 12.299 & 8.434 & 5.259 & 2.723 & 1.284 & 0.467 & 0.422 & 0.662 & 2.086 & 5.860 & 10.087 & 14.831 & 15.000\ 5.60 & 15.000 & 13.968 & 9.902 & 6.515 & 3.751 & 2.024 & 0.899 & 0.481 & 0.327 & 1.199 & 3.998 & 7.162 & 10.779 & 15.000\ 5.70 & 15.000 & 15.000 & 11.456 & 7.864 & 4.886 & 2.894 & 1.491 & 0.743 & 0.251 & 0.653 & 2.645 & 4.936 & 7.625 & 11.161\ 5.80 & 15.000 & 15.000 & 13.051 & 9.262 & 6.080 & 3.843 & 2.185 & 1.140 & 0.349 & 0.346 & 1.665 & 3.235 & 5.154 & 7.847\ 5.90 & 15.000 & 15.000 & 14.677 & 10.696 & 7.319 & 4.854 & 2.959 & 1.642 & 0.584 & 0.223 & 0.980 & 1.949 & 3.223 & 5.212\ 6.00 & 15.000 & 15.000 & 15.000 & 12.203 & 8.641 & 5.960 & 3.843 & 2.275 & 0.972 & 0.292 & 0.572 & 1.037 & 1.765 & 3.158\ 6.10 & 15.000 & 15.000 & 15.000 & 13.904 & 10.164 & 7.278 & 4.952 & 3.151 & 1.620 & 0.652 & 0.522 & 0.555 & 0.817 & 1.699\ 6.20 & 15.000 & 15.000 & 15.000 & 15.000 & 11.988 & 8.907 & 6.384 & 4.362 & 2.618 & 1.385 & 0.900 & 0.559 & 0.416 & 0.851\ 6.30 & 15.000 & 15.000 & 15.000 & 15.000 & 14.084 & 10.818 & 8.106 & 5.876 & 3.930 & 2.454 & 1.654 & 0.983 & 0.485 & 0.524\ 6.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.135 & 10.242 & 7.815 & 5.677 & 3.974 & 2.891 & 1.926 & 1.112 & 0.797\ 6.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.794 & 10.179 & 7.858 & 5.942 & 4.603 & 3.371 & 2.272 & 1.639\ 6.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.558 & 10.062 & 7.945 & 6.372 & 4.896 & 3.537 & 2.616\ 6.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.804 & 12.140 & 9.835 & 8.045 & 6.343 & 4.747 & 3.564\ 6.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.115 & 11.631 & 9.639 & 7.728 & 5.912 & 4.491\ 6.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.401 & 11.220 & 9.114 & 7.095 & 5.454\ 7.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.892 & 10.603 & 8.395 & 6.552\ 7.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.806 & 12.346 & 9.961 & 7.929\ 7.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.474 & 11.923 & 9.714\ 7.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.169 & 11.795\ 7.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.922\ 7.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.80 & 11.152 & 14.747 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.90 & 7.711 & 10.445 & 13.727 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.00 & 4.980 & 6.987 & 9.424 & 11.284 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.10 & 2.941 & 4.324 & 6.030 & 7.215 & 12.209 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.20 & 1.589 & 2.431 & 3.510 & 4.190 & 8.011 & 12.479 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.30 & 0.820 & 1.188 & 1.739 & 2.045 & 4.950 & 8.389 & 12.278 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.40 & 0.700 & 0.659 & 0.775 & 0.790 & 2.975 & 5.602 & 8.592 & 11.964 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.50 & 1.193 & 0.800 & 0.560 & 0.336 & 1.944 & 3.926 & 6.201 & 8.780 & 11.694 & 14.924 & 15.000 & 15.000 & 15.000 & 15.000\ 6.60 & 1.859 & 1.162 & 0.631 & 0.199 & 1.333 & 2.791 & 4.489 & 6.430 & 8.657 & 11.128 & 13.878 & 15.000 & 15.000 & 15.000\ 6.70 & 2.528 & 1.567 & 0.794 & 0.173 & 0.905 & 1.926 & 3.150 & 4.570 & 6.236 & 8.086 & 10.177 & 12.312 & 15.000 & 15.000\ 6.80 & 3.202 & 2.010 & 1.027 & 0.230 & 0.614 & 1.262 & 2.086 & 3.071 & 4.271 & 5.608 & 7.156 & 8.714 & 15.000 & 15.000\ 6.90 & 3.936 & 2.537 & 1.368 & 0.402 & 0.478 & 0.801 & 1.279 & 1.893 & 2.699 & 3.603 & 4.697 & 5.775 & 12.523 & 15.000\ 7.00 & 4.824 & 3.236 & 1.896 & 0.768 & 0.568 & 0.601 & 0.774 & 1.064 & 1.529 & 2.064 & 2.773 & 3.448 & 8.833 & 14.920\ 7.10 & 6.008 & 4.245 & 2.745 & 1.459 & 1.007 & 0.779 & 0.680 & 0.682 & 0.848 & 1.062 & 1.440 & 1.776 & 6.055 & 10.989\ 7.20 & 7.612 & 5.686 & 4.034 & 2.592 & 1.909 & 1.443 & 1.097 & 0.841 & 0.738 & 0.670 & 0.759 & 0.812 & 4.193 & 8.188\ 7.30 & 9.524 & 7.442 & 5.643 & 4.048 & 3.151 & 2.466 & 1.894 & 1.403 & 1.058 & 0.739 & 0.574 & 0.392 & 3.039 & 6.264\ 7.40 & 11.491 & 9.260 & 7.317 & 5.572 & 4.477 & 3.588 & 2.808 & 2.100 & 1.533 & 0.989 & 0.601 & 0.220 & 2.266 & 4.856\ 7.50 & 13.398 & 11.023 & 8.940 & 7.047 & 5.765 & 4.687 & 3.712 & 2.804 & 2.032 & 1.284 & 0.695 & 0.146 & 1.694 & 3.756\ 7.60 & 15.000 & 12.711 & 10.491 & 8.453 & 6.996 & 5.739 & 4.581 & 3.485 & 2.523 & 1.589 & 0.818 & 0.118 & 1.251 & 2.870\ 7.70 & 15.000 & 14.338 & 11.982 & 9.803 & 8.181 & 6.754 & 5.423 & 4.152 & 3.010 & 1.903 & 0.964 & 0.127 & 0.907 & 2.150\ 7.80 & 15.000 & 15.000 & 13.436 & 11.117 & 9.339 & 7.750 & 6.256 & 4.818 & 3.507 & 2.238 & 1.142 & 0.173 & 0.650 & 1.572\ 7.90 & 15.000 & 15.000 & 14.870 & 12.415 & 10.489 & 8.747 & 7.096 & 5.499 & 4.028 & 2.606 & 1.362 & 0.265 & 0.476 & 1.120\ 8.00 & 15.000 & 15.000 & 15.000 & 13.712 & 11.645 & 9.755 & 7.955 & 6.206 & 4.583 & 3.014 & 1.628 & 0.405 & 0.381 & 0.781\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 12.811 & 10.779 & 8.835 & 6.941 & 5.170 & 3.462 & 1.938 & 0.591 & 0.355 & 0.538\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 13.983 & 11.812 & 9.729 & 7.695 & 5.783 & 3.939 & 2.282 & 0.811 & 0.382 & 0.371\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.846 & 10.628 & 8.459 & 6.410 & 4.433 & 2.647 & 1.054 & 0.447 & 0.258\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.873 & 11.524 & 9.223 & 7.041 & 4.936 & 3.022 & 1.309 & 0.536 & 0.181\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.889 & 12.411 & 9.981 & 7.670 & 5.438 & 3.400 & 1.567 & 0.639 & 0.129\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.287 & 10.732 & 8.294 & 5.939 & 3.777 & 1.827 & 0.750 & 0.093\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.152 & 11.473 & 8.912 & 6.435 & 4.153 & 2.087 & 0.867 & 0.068\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.207 & 9.525 & 6.928 & 4.527 & 2.347 & 0.987 & 0.049\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.934 & 10.133 & 7.418 & 4.900 & 2.607 & 1.111 & 0.036\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.657 & 10.738 & 7.907 & 5.274 & 2.869 & 1.237 & 0.027\ [cccccccccccccccc]{} 4.00 & 1.265 & 0.024 & 4.955 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 2.154 & 0.004 & 2.936 & 13.025 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 2.875 & 0.021 & 1.343 & 8.551 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 3.552 & 0.289 & 0.314 & 5.102 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 4.404 & 1.022 & 0.044 & 2.885 & 10.117 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 5.028 & 1.630 & 0.033 & 1.310 & 6.154 & 14.805 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 5.543 & 2.129 & 0.267 & 0.345 & 3.274 & 9.302 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 6.208 & 2.753 & 0.760 & 0.099 & 1.542 & 5.469 & 13.084 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 6.815 & 3.295 & 1.192 & 0.164 & 0.604 & 2.920 & 8.226 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 7.502 & 3.895 & 1.678 & 0.414 & 0.250 & 1.496 & 5.147 & 10.782 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 8.290 & 4.581 & 2.240 & 0.774 & 0.187 & 0.755 & 3.281 & 7.386 & 12.913 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 9.195 & 5.371 & 2.898 & 1.238 & 0.301 & 0.364 & 2.059 & 5.057 & 9.044 & 14.315 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 10.212 & 6.264 & 3.649 & 1.799 & 0.552 & 0.192 & 1.222 & 3.353 & 6.215 & 9.840 & 15.000 & 15.000 & 15.000 & 15.000\ 5.30 & 11.329 & 7.253 & 4.489 & 2.448 & 0.918 & 0.187 & 0.663 & 2.082 & 4.070 & 6.466 & 10.375 & 15.000 & 15.000 & 15.000\ 5.40 & 12.564 & 8.355 & 5.435 & 3.205 & 1.412 & 0.350 & 0.353 & 1.175 & 2.452 & 3.981 & 6.890 & 12.708 & 15.000 & 15.000\ 5.50 & 13.955 & 9.611 & 6.531 & 4.110 & 2.070 & 0.709 & 0.303 & 0.616 & 1.302 & 2.201 & 4.312 & 8.874 & 13.947 & 15.000\ 5.60 & 15.000 & 11.104 & 7.859 & 5.249 & 2.976 & 1.340 & 0.573 & 0.446 & 0.633 & 1.038 & 2.494 & 6.024 & 9.971 & 14.310\ 5.70 & 15.000 & 12.903 & 9.490 & 6.691 & 4.195 & 2.305 & 1.215 & 0.704 & 0.462 & 0.458 & 1.365 & 4.039 & 7.056 & 10.347\ 5.80 & 15.000 & 14.897 & 11.316 & 8.328 & 5.619 & 3.492 & 2.108 & 1.258 & 0.642 & 0.286 & 0.724 & 2.683 & 4.925 & 7.351\ 5.90 & 15.000 & 15.000 & 13.205 & 10.031 & 7.116 & 4.766 & 3.114 & 1.958 & 1.012 & 0.345 & 0.377 & 1.732 & 3.322 & 5.029\ 6.00 & 15.000 & 15.000 & 15.000 & 11.752 & 8.638 & 6.077 & 4.175 & 2.742 & 1.499 & 0.554 & 0.227 & 1.067 & 2.102 & 3.205\ 6.10 & 15.000 & 15.000 & 15.000 & 13.501 & 10.194 & 7.433 & 5.299 & 3.611 & 2.098 & 0.898 & 0.252 & 0.646 & 1.204 & 1.794\ 6.20 & 15.000 & 15.000 & 15.000 & 15.000 & 11.893 & 8.941 & 6.588 & 4.664 & 2.903 & 1.468 & 0.532 & 0.536 & 0.679 & 0.828\ 6.30 & 15.000 & 15.000 & 15.000 & 15.000 & 13.902 & 10.766 & 8.208 & 6.063 & 4.071 & 2.417 & 1.216 & 0.875 & 0.652 & 0.416\ 6.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.962 & 10.209 & 7.857 & 5.649 & 3.788 & 2.343 & 1.693 & 1.144 & 0.567\ 6.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.734 & 10.187 & 7.776 & 5.719 & 4.047 & 3.118 & 2.275 & 1.393\ 6.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.796 & 10.193 & 7.949 & 6.065 & 4.881 & 3.771 & 2.613\ 6.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.533 & 10.112 & 8.029 & 6.611 & 5.255 & 3.846\ 6.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.739 & 12.149 & 9.879 & 8.244 & 6.662 & 5.022\ 6.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.102 & 11.656 & 9.819 & 8.028 & 6.176\ 7.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.443 & 11.417 & 9.431 & 7.381\ 7.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.164 & 10.996 & 8.763\ 7.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.887 & 10.483\ 7.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.580\ 7.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.838\ 7.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.70 & 14.776 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.80 & 10.802 & 14.657 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.90 & 7.666 & 10.636 & 13.840 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.00 & 5.158 & 7.386 & 9.794 & 12.676 & 14.991 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.10 & 3.163 & 4.760 & 6.490 & 8.584 & 10.139 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.20 & 1.692 & 2.742 & 3.886 & 5.303 & 6.277 & 10.652 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.30 & 0.837 & 1.408 & 2.041 & 2.884 & 3.426 & 6.778 & 10.667 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.40 & 0.595 & 0.743 & 0.933 & 1.300 & 1.513 & 4.062 & 7.059 & 10.437 & 14.209 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.50 & 1.072 & 0.844 & 0.651 & 0.630 & 0.579 & 2.490 & 4.780 & 7.375 & 10.285 & 13.565 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.60 & 1.979 & 1.417 & 0.894 & 0.558 & 0.282 & 1.675 & 3.395 & 5.362 & 7.579 & 10.111 & 12.890 & 15.000 & 15.000 & 15.000 & 15.000\ 6.70 & 2.930 & 2.070 & 1.263 & 0.666 & 0.191 & 1.150 & 2.399 & 3.854 & 5.508 & 7.432 & 9.543 & 11.904 & 14.328 & 15.000 & 15.000\ 6.80 & 3.852 & 2.724 & 1.669 & 0.850 & 0.192 & 0.780 & 1.630 & 2.656 & 3.843 & 5.266 & 6.826 & 8.605 & 10.415 & 15.000 & 15.000\ 6.90 & 4.774 & 3.404 & 2.130 & 1.116 & 0.284 & 0.546 & 1.052 & 1.712 & 2.503 & 3.504 & 4.604 & 5.899 & 7.199 & 14.523 & 15.000\ 7.00 & 5.768 & 4.178 & 2.706 & 1.515 & 0.517 & 0.489 & 0.691 & 1.030 & 1.480 & 2.118 & 2.827 & 3.715 & 4.587 & 10.451 & 15.000\ 7.10 & 6.953 & 5.162 & 3.509 & 2.153 & 0.995 & 0.704 & 0.633 & 0.688 & 0.836 & 1.158 & 1.530 & 2.069 & 2.581 & 7.261 & 12.601\ 7.20 & 8.491 & 6.512 & 4.691 & 3.180 & 1.864 & 1.333 & 1.015 & 0.813 & 0.692 & 0.733 & 0.810 & 1.046 & 1.257 & 4.975 & 9.312\ 7.30 & 10.418 & 8.265 & 6.284 & 4.624 & 3.153 & 2.401 & 1.856 & 1.420 & 1.055 & 0.844 & 0.659 & 0.629 & 0.589 & 3.523 & 7.040\ 7.40 & 12.515 & 10.198 & 8.066 & 6.260 & 4.638 & 3.680 & 2.926 & 2.274 & 1.686 & 1.245 & 0.826 & 0.562 & 0.310 & 2.601 & 5.444\ 7.50 & 14.591 & 12.120 & 9.841 & 7.894 & 6.123 & 4.973 & 4.024 & 3.172 & 2.378 & 1.725 & 1.096 & 0.621 & 0.190 & 1.951 & 4.234\ 7.60 & 15.000 & 13.972 & 11.551 & 9.465 & 7.548 & 6.218 & 5.086 & 4.047 & 3.061 & 2.212 & 1.390 & 0.726 & 0.137 & 1.457 & 3.271\ 7.70 & 15.000 & 15.000 & 13.195 & 10.972 & 8.912 & 7.413 & 6.109 & 4.893 & 3.727 & 2.694 & 1.694 & 0.857 & 0.125 & 1.071 & 2.491\ 7.80 & 15.000 & 15.000 & 14.790 & 12.432 & 10.231 & 8.573 & 7.104 & 5.721 & 4.385 & 3.180 & 2.014 & 1.016 & 0.148 & 0.775 & 1.858\ 7.90 & 15.000 & 15.000 & 15.000 & 13.864 & 11.525 & 9.716 & 8.090 & 6.548 & 5.050 & 3.682 & 2.359 & 1.210 & 0.211 & 0.561 & 1.352\ 8.00 & 15.000 & 15.000 & 15.000 & 15.000 & 12.809 & 10.857 & 9.081 & 7.387 & 5.736 & 4.211 & 2.740 & 1.447 & 0.321 & 0.425 & 0.962\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 14.095 & 12.005 & 10.085 & 8.245 & 6.447 & 4.774 & 3.160 & 1.728 & 0.476 & 0.362 & 0.674\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.162 & 11.102 & 9.121 & 7.181 & 5.365 & 3.614 & 2.047 & 0.672 & 0.359 & 0.469\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.320 & 12.124 & 10.007 & 7.930 & 5.976 & 4.092 & 2.394 & 0.896 & 0.401 & 0.327\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.145 & 10.895 & 8.685 & 6.597 & 4.582 & 2.757 & 1.138 & 0.473 & 0.230\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.157 & 11.777 & 9.438 & 7.219 & 5.078 & 3.127 & 1.388 & 0.565 & 0.163\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.651 & 10.185 & 7.840 & 5.574 & 3.500 & 1.643 & 0.670 & 0.117\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.515 & 10.925 & 8.455 & 6.068 & 3.873 & 1.900 & 0.782 & 0.085\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.370 & 11.658 & 9.066 & 6.559 & 4.245 & 2.157 & 0.899 & 0.062\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.385 & 9.673 & 7.047 & 4.616 & 2.416 & 1.021 & 0.045\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.107 & 10.278 & 7.535 & 4.988 & 2.677 & 1.146 & 0.033\ [cccccccccccccccc]{} 4.00 & 1.049 & 0.041 & 5.484 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 1.939 & 0.005 & 3.230 & 13.805 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 2.710 & 0.016 & 1.456 & 9.302 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 3.623 & 0.302 & 0.301 & 5.616 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 5.041 & 1.196 & 0.029 & 3.308 & 11.055 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 6.412 & 2.128 & 0.013 & 1.651 & 6.910 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 7.717 & 3.061 & 0.173 & 0.485 & 3.744 & 10.228 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 9.260 & 4.287 & 0.805 & 0.084 & 1.734 & 6.017 & 12.885 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 10.437 & 5.191 & 1.489 & 0.164 & 0.551 & 3.072 & 7.663 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 11.109 & 5.625 & 2.099 & 0.587 & 0.163 & 1.339 & 4.127 & 9.899 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 11.575 & 5.882 & 2.583 & 1.005 & 0.221 & 0.540 & 2.053 & 6.066 & 11.630 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 12.095 & 6.217 & 3.082 & 1.437 & 0.461 & 0.282 & 1.024 & 3.854 & 7.885 & 13.391 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 12.769 & 6.726 & 3.688 & 1.957 & 0.820 & 0.283 & 0.505 & 2.489 & 5.434 & 9.469 & 14.540 & 15.000 & 15.000 & 15.000 & 15.000\ 5.30 & 13.611 & 7.417 & 4.420 & 2.584 & 1.288 & 0.447 & 0.251 & 1.567 & 3.672 & 6.640 & 10.128 & 15.000 & 15.000 & 15.000 & 15.000\ 5.40 & 14.591 & 8.260 & 5.261 & 3.303 & 1.843 & 0.727 & 0.170 & 0.930 & 2.351 & 4.487 & 6.799 & 10.567 & 15.000 & 15.000 & 15.000\ 5.50 & 15.000 & 9.244 & 6.209 & 4.117 & 2.488 & 1.115 & 0.239 & 0.523 & 1.371 & 2.831 & 4.313 & 7.143 & 12.494 & 15.000 & 15.000\ 5.60 & 15.000 & 10.406 & 7.308 & 5.071 & 3.270 & 1.656 & 0.490 & 0.360 & 0.721 & 1.614 & 2.494 & 4.571 & 8.767 & 13.387 & 15.000\ 5.70 & 15.000 & 11.812 & 8.632 & 6.242 & 4.265 & 2.421 & 0.990 & 0.492 & 0.432 & 0.842 & 1.250 & 2.704 & 5.950 & 9.540 & 13.434\ 5.80 & 15.000 & 13.531 & 10.254 & 7.706 & 5.549 & 3.486 & 1.807 & 0.980 & 0.551 & 0.540 & 0.557 & 1.485 & 3.940 & 6.677 & 9.615\ 5.90 & 15.000 & 15.000 & 12.133 & 9.421 & 7.084 & 4.809 & 2.899 & 1.772 & 1.014 & 0.633 & 0.313 & 0.787 & 2.579 & 4.602 & 6.749\ 6.00 & 15.000 & 15.000 & 14.130 & 11.253 & 8.734 & 6.254 & 4.126 & 2.722 & 1.669 & 0.957 & 0.338 & 0.416 & 1.645 & 3.064 & 4.552\ 6.10 & 15.000 & 15.000 & 15.000 & 13.128 & 10.428 & 7.750 & 5.414 & 3.754 & 2.432 & 1.420 & 0.531 & 0.259 & 1.005 & 1.907 & 2.838\ 6.20 & 15.000 & 15.000 & 15.000 & 15.000 & 12.179 & 9.308 & 6.774 & 4.876 & 3.305 & 2.019 & 0.883 & 0.296 & 0.622 & 1.079 & 1.533\ 6.30 & 15.000 & 15.000 & 15.000 & 15.000 & 14.111 & 11.053 & 8.329 & 6.207 & 4.406 & 2.866 & 1.502 & 0.629 & 0.586 & 0.652 & 0.694\ 6.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.145 & 10.239 & 7.907 & 5.889 & 4.115 & 2.535 & 1.400 & 1.030 & 0.749 & 0.430\ 6.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.597 & 10.066 & 7.845 & 5.851 & 4.068 & 2.692 & 2.027 & 1.436 & 0.794\ 6.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.728 & 10.316 & 8.115 & 6.140 & 4.539 & 3.606 & 2.734 & 1.803\ 6.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.883 & 10.488 & 8.330 & 6.519 & 5.341 & 4.213 & 3.018\ 6.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.748 & 10.416 & 8.409 & 7.005 & 5.640 & 4.205\ 6.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.893 & 12.394 & 10.203 & 8.588 & 7.006 & 5.350\ 7.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.323 & 11.957 & 10.148 & 8.363 & 6.503\ 7.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.773 & 11.780 & 9.807 & 7.757\ 7.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.638 & 11.489 & 9.262\ 7.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.549 & 11.158\ 7.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.342\ 7.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.80 & 13.582 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.90 & 9.827 & 13.263 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.00 & 6.886 & 9.515 & 12.355 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.10 & 4.544 & 6.491 & 8.596 & 11.082 & 13.023 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.20 & 2.698 & 4.059 & 5.533 & 7.289 & 8.572 & 13.497 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.30 & 1.386 & 2.236 & 3.161 & 4.298 & 5.087 & 8.885 & 13.270 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.40 & 0.704 & 1.104 & 1.551 & 2.173 & 2.590 & 5.505 & 8.908 & 12.692 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.50 & 0.696 & 0.698 & 0.733 & 0.933 & 1.058 & 3.275 & 5.899 & 8.830 & 12.101 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.60 & 1.372 & 1.022 & 0.701 & 0.559 & 0.442 & 2.094 & 4.094 & 6.343 & 8.861 & 11.677 & 14.773 & 15.000 & 15.000 & 15.000 & 15.000\ 6.70 & 2.289 & 1.625 & 0.999 & 0.576 & 0.247 & 1.432 & 2.920 & 4.615 & 6.523 & 8.686 & 11.065 & 13.695 & 15.000 & 15.000 & 15.000\ 6.80 & 3.206 & 2.261 & 1.372 & 0.713 & 0.192 & 0.979 & 2.039 & 3.274 & 4.680 & 6.307 & 8.097 & 10.106 & 12.172 & 15.000 & 15.000\ 6.90 & 4.106 & 2.909 & 1.789 & 0.924 & 0.224 & 0.666 & 1.358 & 2.204 & 3.187 & 4.366 & 5.667 & 7.162 & 8.685 & 15.000 & 15.000\ 7.00 & 5.035 & 3.609 & 2.283 & 1.234 & 0.364 & 0.501 & 0.871 & 1.378 & 1.999 & 2.797 & 3.684 & 4.748 & 5.817 & 12.169 & 15.000\ 7.10 & 6.083 & 4.448 & 2.935 & 1.716 & 0.683 & 0.545 & 0.629 & 0.837 & 1.141 & 1.608 & 2.140 & 2.835 & 3.523 & 8.611 & 14.364\ 7.20 & 7.397 & 5.570 & 3.884 & 2.505 & 1.312 & 0.924 & 0.750 & 0.691 & 0.713 & 0.887 & 1.110 & 1.487 & 1.855 & 5.916 & 10.601\ 7.30 & 9.115 & 7.108 & 5.259 & 3.729 & 2.379 & 1.762 & 1.353 & 1.051 & 0.819 & 0.731 & 0.680 & 0.779 & 0.878 & 4.103 & 7.916\ 7.40 & 11.131 & 8.957 & 6.954 & 5.276 & 3.773 & 2.944 & 2.319 & 1.794 & 1.331 & 1.005 & 0.711 & 0.565 & 0.438 & 2.978 & 6.075\ 7.50 & 13.208 & 10.876 & 8.724 & 6.903 & 5.249 & 4.223 & 3.397 & 2.666 & 1.991 & 1.447 & 0.934 & 0.571 & 0.252 & 2.227 & 4.732\ 7.60 & 15.000 & 12.747 & 10.452 & 8.490 & 6.689 & 5.478 & 4.465 & 3.541 & 2.669 & 1.924 & 1.212 & 0.653 & 0.167 & 1.674 & 3.685\ 7.70 & 15.000 & 14.549 & 12.114 & 10.015 & 8.068 & 6.685 & 5.495 & 4.391 & 3.334 & 2.401 & 1.506 & 0.769 & 0.133 & 1.246 & 2.841\ 7.80 & 15.000 & 15.000 & 13.720 & 11.485 & 9.397 & 7.851 & 6.493 & 5.218 & 3.988 & 2.879 & 1.814 & 0.912 & 0.136 & 0.913 & 2.154\ 7.90 & 15.000 & 15.000 & 15.000 & 12.918 & 10.691 & 8.991 & 7.474 & 6.037 & 4.642 & 3.366 & 2.142 & 1.085 & 0.176 & 0.662 & 1.599\ 8.00 & 15.000 & 15.000 & 15.000 & 14.331 & 11.968 & 10.122 & 8.452 & 6.860 & 5.309 & 3.876 & 2.500 & 1.296 & 0.257 & 0.489 & 1.160\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 13.240 & 11.255 & 9.439 & 7.699 & 5.999 & 4.414 & 2.894 & 1.550 & 0.384 & 0.389 & 0.826\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 14.514 & 12.396 & 10.438 & 8.556 & 6.713 & 4.983 & 3.324 & 1.844 & 0.553 & 0.353 & 0.582\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.541 & 11.447 & 9.427 & 7.445 & 5.576 & 3.782 & 2.170 & 0.756 & 0.369 & 0.409\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.685 & 12.457 & 10.303 & 8.187 & 6.183 & 4.258 & 2.518 & 0.982 & 0.423 & 0.288\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.463 & 11.179 & 8.932 & 6.797 & 4.745 & 2.878 & 1.223 & 0.502 & 0.205\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.461 & 12.049 & 9.675 & 7.412 & 5.234 & 3.244 & 1.471 & 0.597 & 0.147\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.911 & 10.412 & 8.024 & 5.724 & 3.612 & 1.723 & 0.703 & 0.106\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.764 & 11.143 & 8.633 & 6.212 & 3.981 & 1.977 & 0.816 & 0.077\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.609 & 11.868 & 9.238 & 6.698 & 4.350 & 2.233 & 0.935 & 0.057\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.589 & 9.841 & 7.184 & 4.721 & 2.492 & 1.058 & 0.041\ [cccccccccccccccccccccccccccccccc]{} 4.00 & 0.475 & 0.177 & 4.862 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 1.414 & 0.018 & 2.719 & 13.737 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 2.328 & 0.035 & 1.135 & 9.224 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 3.441 & 0.411 & 0.214 & 5.589 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 4.918 & 1.277 & 0.024 & 3.220 & 11.485 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 6.286 & 2.135 & 0.019 & 1.462 & 7.154 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 7.693 & 3.113 & 0.287 & 0.316 & 3.940 & 10.630 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 9.510 & 4.567 & 1.153 & 0.036 & 1.992 & 6.432 & 13.876 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 11.306 & 6.052 & 2.148 & 0.099 & 0.706 & 3.335 & 8.424 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 13.265 & 7.744 & 3.428 & 0.617 & 0.149 & 1.320 & 4.516 & 9.289 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 15.000 & 9.532 & 4.868 & 1.456 & 0.295 & 0.348 & 2.045 & 4.951 & 10.587 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 15.000 & 10.712 & 5.750 & 2.116 & 0.760 & 0.178 & 0.820 & 2.372 & 6.265 & 11.537 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 15.000 & 11.111 & 5.893 & 2.526 & 1.159 & 0.347 & 0.392 & 1.142 & 3.881 & 7.705 & 12.895 & 15.000 & 15.000 & 15.000 & 15.000\ 5.30 & 15.000 & 11.446 & 6.006 & 2.957 & 1.578 & 0.644 & 0.322 & 0.575 & 2.509 & 5.316 & 9.147 & 13.597 & 15.000 & 15.000 & 15.000\ 5.40 & 15.000 & 11.963 & 6.328 & 3.512 & 2.092 & 1.042 & 0.430 & 0.297 & 1.601 & 3.622 & 6.462 & 9.468 & 14.150 & 15.000 & 15.000\ 5.50 & 15.000 & 12.670 & 6.863 & 4.196 & 2.706 & 1.530 & 0.658 & 0.195 & 0.973 & 2.352 & 4.413 & 6.361 & 9.927 & 15.000 & 15.000\ 5.60 & 15.000 & 13.556 & 7.596 & 5.007 & 3.425 & 2.111 & 0.996 & 0.238 & 0.565 & 1.404 & 2.826 & 4.033 & 6.729 & 11.603 & 15.000\ 5.70 & 15.000 & 14.646 & 8.546 & 5.983 & 4.289 & 2.827 & 1.481 & 0.453 & 0.385 & 0.762 & 1.644 & 2.313 & 4.303 & 8.124 & 12.353\ 5.80 & 15.000 & 15.000 & 9.775 & 7.196 & 5.376 & 3.757 & 2.190 & 0.912 & 0.491 & 0.468 & 0.885 & 1.130 & 2.529 & 5.479 & 8.762\ 5.90 & 15.000 & 15.000 & 11.367 & 8.741 & 6.782 & 4.999 & 3.219 & 1.705 & 0.966 & 0.590 & 0.600 & 0.492 & 1.385 & 3.608 & 6.103\ 6.00 & 15.000 & 15.000 & 13.259 & 10.562 & 8.455 & 6.503 & 4.517 & 2.780 & 1.750 & 1.058 & 0.706 & 0.291 & 0.744 & 2.355 & 4.187\ 6.10 & 15.000 & 15.000 & 15.000 & 12.504 & 10.243 & 8.119 & 5.932 & 3.982 & 2.684 & 1.707 & 1.029 & 0.340 & 0.405 & 1.493 & 2.762\ 6.20 & 15.000 & 15.000 & 15.000 & 14.504 & 12.085 & 9.786 & 7.403 & 5.250 & 3.703 & 2.464 & 1.490 & 0.552 & 0.271 & 0.908 & 1.693\ 6.30 & 15.000 & 15.000 & 15.000 & 15.000 & 14.010 & 11.537 & 8.963 & 6.615 & 4.835 & 3.353 & 2.108 & 0.941 & 0.346 & 0.591 & 0.954\ 6.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.519 & 10.759 & 8.223 & 6.225 & 4.517 & 3.020 & 1.638 & 0.759 & 0.656 & 0.649\ 6.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.952 & 10.235 & 8.031 & 6.111 & 4.380 & 2.796 & 1.656 & 1.242 & 0.905\ 6.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.780 & 10.380 & 8.260 & 6.312 & 4.535 & 3.154 & 2.459 & 1.825\ 6.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.091 & 10.782 & 8.629 & 6.668 & 5.064 & 4.112 & 3.208\ 6.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.272 & 10.926 & 8.790 & 6.978 & 5.791 & 4.640\ 6.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.106 & 10.803 & 8.795 & 7.390 & 6.012\ 7.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.733 & 10.543 & 8.935 & 7.346\ 7.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.656 & 12.293 & 10.496 & 8.712\ 7.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.170 & 12.195 & 10.228\ 7.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.200 & 12.061\ 7.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.257\ 7.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 7.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 4.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.10 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.80 & 12.236 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 5.90 & 8.707 & 12.264 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.00 & 6.069 & 8.816 & 11.864 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.10 & 4.039 & 6.105 & 8.415 & 10.899 & 13.790 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.20 & 2.456 & 3.940 & 5.621 & 7.426 & 9.535 & 11.144 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.30 & 1.276 & 2.254 & 3.389 & 4.605 & 6.048 & 7.099 & 11.358 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.40 & 0.585 & 1.118 & 1.774 & 2.479 & 3.365 & 3.996 & 7.290 & 11.101 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.50 & 0.500 & 0.639 & 0.872 & 1.134 & 1.564 & 1.870 & 4.403 & 7.361 & 10.637 & 14.286 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.60 & 1.115 & 0.902 & 0.763 & 0.644 & 0.704 & 0.746 & 2.665 & 4.943 & 7.478 & 10.310 & 13.448 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000\ 6.70 & 2.223 & 1.695 & 1.225 & 0.781 & 0.539 & 0.355 & 1.768 & 3.493 & 5.431 & 7.602 & 10.035 & 12.695 & 15.000 & 15.000 & 15.000 & 15.000\ 6.80 & 3.403 & 2.593 & 1.828 & 1.103 & 0.612 & 0.226 & 1.213 & 2.478 & 3.923 & 5.554 & 7.410 & 9.441 & 11.690 & 14.018 & 15.000 & 15.000\ 6.90 & 4.545 & 3.478 & 2.450 & 1.482 & 0.776 & 0.205 & 0.825 & 1.699 & 2.729 & 3.908 & 5.286 & 6.796 & 8.500 & 10.254 & 15.000 & 15.000\ 7.00 & 5.667 & 4.368 & 3.102 & 1.918 & 1.023 & 0.277 & 0.574 & 1.110 & 1.782 & 2.577 & 3.550 & 4.622 & 5.870 & 7.145 & 14.001 & 15.000\ 7.10 & 6.837 & 5.324 & 3.841 & 2.464 & 1.396 & 0.482 & 0.493 & 0.728 & 1.087 & 1.548 & 2.172 & 2.870 & 3.731 & 4.601 & 10.109 & 15.000\ 7.20 & 8.171 & 6.461 & 4.779 & 3.224 & 1.995 & 0.919 & 0.670 & 0.637 & 0.716 & 0.882 & 1.199 & 1.573 & 2.100 & 2.629 & 7.043 & 12.081\ 7.30 & 9.833 & 7.940 & 6.074 & 4.352 & 2.971 & 1.736 & 1.250 & 0.973 & 0.799 & 0.702 & 0.745 & 0.834 & 1.070 & 1.312 & 4.835 & 8.947\ 7.40 & 11.870 & 9.805 & 7.767 & 5.888 & 4.360 & 2.970 & 2.265 & 1.764 & 1.360 & 1.023 & 0.820 & 0.655 & 0.635 & 0.634 & 3.429 & 6.782\ 7.50 & 14.073 & 11.846 & 9.647 & 7.617 & 5.947 & 4.404 & 3.497 & 2.789 & 2.172 & 1.616 & 1.188 & 0.795 & 0.549 & 0.340 & 2.537 & 5.264\ 7.60 & 15.000 & 13.880 & 11.528 & 9.353 & 7.543 & 5.851 & 4.755 & 3.854 & 3.041 & 2.281 & 1.646 & 1.047 & 0.598 & 0.210 & 1.911 & 4.117\ 7.70 & 15.000 & 15.000 & 13.351 & 11.034 & 9.088 & 7.250 & 5.977 & 4.895 & 3.896 & 2.948 & 2.121 & 1.333 & 0.700 & 0.150 & 1.436 & 3.204\ 7.80 & 15.000 & 15.000 & 15.000 & 12.656 & 10.575 & 8.594 & 7.156 & 5.902 & 4.729 & 3.603 & 2.596 & 1.633 & 0.831 & 0.133 & 1.066 & 2.463\ 7.90 & 15.000 & 15.000 & 15.000 & 14.232 & 12.018 & 9.897 & 8.302 & 6.885 & 5.548 & 4.254 & 3.077 & 1.950 & 0.990 & 0.151 & 0.781 & 1.860\ 8.00 & 15.000 & 15.000 & 15.000 & 15.000 & 13.431 & 11.174 & 9.430 & 7.858 & 6.364 & 4.911 & 3.573 & 2.292 & 1.182 & 0.208 & 0.573 & 1.376\ 8.10 & 15.000 & 15.000 & 15.000 & 15.000 & 14.830 & 12.439 & 10.554 & 8.834 & 7.189 & 5.585 & 4.094 & 2.666 & 1.413 & 0.307 & 0.437 & 0.998\ 8.20 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.702 & 11.682 & 9.819 & 8.030 & 6.280 & 4.643 & 3.074 & 1.683 & 0.449 & 0.367 & 0.712\ 8.30 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.965 & 12.816 & 10.815 & 8.886 & 6.997 & 5.218 & 3.513 & 1.989 & 0.629 & 0.355 & 0.505\ 8.40 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.952 & 11.815 & 9.752 & 7.727 & 5.813 & 3.975 & 2.322 & 0.837 & 0.386 & 0.358\ 8.50 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 12.816 & 10.621 & 8.465 & 6.418 & 4.452 & 2.672 & 1.064 & 0.449 & 0.255\ 8.60 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.811 & 11.488 & 9.204 & 7.028 & 4.936 & 3.031 & 1.304 & 0.533 & 0.183\ 8.70 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.797 & 12.349 & 9.939 & 7.637 & 5.422 & 3.395 & 1.550 & 0.632 & 0.132\ 8.80 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 13.202 & 10.669 & 8.245 & 5.908 & 3.762 & 1.800 & 0.740 & 0.096\ 8.90 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.047 & 11.394 & 8.849 & 6.394 & 4.130 & 2.053 & 0.855 & 0.070\ 9.00 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 15.000 & 14.886 & 12.116 & 9.452 & 6.880 & 4.499 & 2.310 & 0.976 & 0.051\ [cccccccc]{} N & 14.53 & -4.08 & 1.00 & 1.00 && -4.08 & -4.08\ O & 13.62 & -3.17 & 1.00 & 1.00 && -3.17 & -3.17\ Ne & 21.56 & -3.89 & 1.00 & 1.00 && -3.89 & -3.89\ Na & 5.14 & -5.67 & $7.75^{+12.98}_{-4.97}$ & $7.84^{+19.98}_{-6.40}$ && $-4.78^{+0.42}_{-0.45}$ & $-4.78^{+0.55}_{-0.73}$\ Mg & 7.65 & -4.42 & $2.78^{+2.28}_{-1.25}$ & $2.91^{+0.14}_{-0.13}$ && $-3.98\pm 0.26$ & $-3.96\pm 0.02$\ Al & 5.99 & -5.51 & $3.58^{+1.72}_{-1.16}$ & $4.51^{+1.57}_{-1.17}$ && $-4.96\pm 0.17$ & $-4.86\pm 0.13$\ Si & 8.15 & -4.44 & $4.94^{+2.90}_{-1.82}$ & $5.18^{+3.41}_{-2.06}$ && $-3.75\pm 0.20$ & $-3.73\pm 0.22$\ S & 10.36 & -4.67 & $2.19^{+0.21}_{-0.19}$ & $1.78^{+0.31}_{-0.33}$ && $-4.33\pm 0.04$ & $-4.42\pm 0.09$\ Ar & 15.76 & -5.41 & 1.00 & 1.00 && -5.41 & -5.41\ K & 4.34 & -6.87 & $1.75^{+0.44}_{-0.59}$ & $3.84^{+2.99}_{-1.30}$ && $-6.63\pm 0.35$ & $-6.29^{+0.25}_{-0.18}$\ Ca & 6.11 & -5.65 & $3.46^{+4.29}_{-1.93}$ & $6.67^{+11.70}_{-4.25}$ && $-5.11^{+0.18}_{-0.10}$ & $-4.83\pm 0.44$\ Fe & 7.90 & -4.50 & $6.98^{+1.42}_{-1.17}$ & $5.06^{+0.62}_{-0.55}$ && $-3.66\pm 0.08$ & $-3.80\pm 0.05$ [ccccccc]{} \[fig:1a\] & Ia & 1 & 6.16 & 0.11 & 43.24 & 0.66\ \[fig:1a\] & Ia & 2 & 6.17 & 0.07 & 43.10 & 0.45\ \[fig:1a\] & Ia & 3 & 6.16 & 0.07 & 43.02 & 0.15\ \ \[fig:1b\] & Ib & 1 & 6.15 & 0.10 & 43.03 & 0.62\ \[fig:1b\] & Ib & 2 & 6.16 & 0.05 & 42.97 & 0.27\ \[fig:1b\] & Ib & 3 & 6.16 & 0.05 & 42.94 & 0.27\ \ \[fig:1\] & I & 1 & 6.16 & 0.09 & 43.09 & 0.51\ \[fig:1\] & I & 2 & 6.16 & 0.07 & 43.05 & 0.37\ \[fig:1\] & I & 3 & 6.17 & 0.05 & 43.00 & 0.20\ \ \[fig:2a\] & IIa & 1 & 6.04 & 0.26 & 42.99 & 0.74\ \[fig:2a\] & IIa & 2 & 6.02 & 0.22 & 42.92 & 0.53\ \[fig:2a\] & IIa & 3 & 5.98 & 0.19 & 42.62 & 0.36\ \ \[fig:2ax\] & IIa & 1 & 6.13 & 0.20 & 42.97 & 0.56\ \[fig:2ax\] & IIa & 2 & 6.11 & 0.19 & 42.95 & 0.56\ \[fig:2ax\] & IIa & 3 & 6.06 & 0.12 & 42.73 & 0.30\ \ \[fig:2a\_on\] & IIa & 1 & 5.79 & 0.30 & 42.50 & 0.91\ \[fig:2a\_on\] & IIa & 2 & 5.84 & 0.27 & 42.70 & 0.84\ \[fig:2a\_on\] & IIa & 3 & 5.44 & 0.00 & 41.44 & 0.00\ \ \[fig:2b\] & IIb & 1 & 6.14 & 0.12 & 43.25 & 0.75\ \[fig:2b\] & IIb & 2 & 6.11 & 0.06 & 43.05 & 0.32\ \[fig:2b\] & IIb & 3 & 6.11 & 0.04 & 43.00 & 0.21\ \ \[fig:2\] & II & 1 & 6.12 & 0.15 & 43.09 & 0.70\ \[fig:2\] & II & 2 & 6.11 & 0.09 & 42.98 & 0.39\ \[fig:2\] & II & 3 & 6.10 & 0.06 & 42.91 & 0.26\ \ \[fig:3\] & III & 1 & 6.23 & 0.21 & 43.66 & 1.14\ \[fig:3\] & III & 2 & 6.17 & 0.10 & 43.40 & 0.59\ \[fig:3\] & III & 3 & 6.12 & 0.01 & 43.07 & 0.03\ \ \[fig:all\] & All Lines & 1 & 6.16 & 0.16 & 43.26 & 0.83\ \[fig:all\] & All Lines & 2 & 6.14 & 0.10 & 43.08 & 0.46\ \[fig:all\] & All Lines & 3 & 6.12 & 0.07 & 42.98 & 0.29\ \ \[fig:allx\] & All Lines & 1 & 6.17 & 0.15 & 43.40 & 0.90\ \[fig:allx\] & All Lines & 2 & 6.14 & 0.08 & 43.15 & 0.46\ \[fig:allx\] & All Lines & 3 & 6.13 & 0.06 & 43.02 & 0.29 [ccccccccc]{} \[fig:1a\] & Ia & $6.16\pm 0.07$ & $6.13\pm 0.06$ & $6.13\pm 0.01$ & & $43.02\pm 0.15$ & $43.09\pm 0.13$ & $43.15\pm 0.10$\ \[fig:1b\] & Ib & $6.16\pm 0.05$ & $6.17\pm 0.03$ & $6.17\pm 0.01$ & & $42.94\pm 0.27$ & $42.92\pm 0.16$ & $43.15\pm 0.05$\ \[fig:2ax\] & IIa & $6.06\pm 0.12$ & $6.03\pm 0.12$ & $6.11\pm 0.02$ & & $42.73\pm 0.30$ & $42.63\pm 0.26$ & $42.90\pm 0.15$\ \[fig:2b\] & IIb & $6.11\pm 0.04$ & $6.10\pm 0.04$& $6.13\pm 0.01$ & & $43.00\pm 0.21$ & $42.99\pm 0.22$ & $43.30\pm 0.15$\ \[fig:3\] & III & $6.12\pm 0.01$ & $6.11\pm 0.01$& $6.13\pm 0.01$ & & $43.07\pm 0.03$ & $43.05\pm 0.04$ & $43.45\pm 0.10$ ![Ionization fractional abundance versus electron temperature for all ionization stages of H. The upper graph shows our results ([solid curves]{}) and the abundances calculated by Mazzotta et al. (1998; [dashed curves]{}). The lower graph shows the ratio of the calculated abundances. Comparison is made only for fractional abundances greater than $10^{-2}$. We label our results as “New” and those of @Mazz98a as “Old”.[]{data-label="fig:H Mazz"}](f1.eps) ![Same as Fig. \[fig:H Mazz\] but for He.[]{data-label="fig:He Mazz"}](f2.eps) ![Same as Fig. \[fig:H Mazz\] but for Li.[]{data-label="fig:Li Mazz"}](f3.eps) ![Same as Fig. \[fig:H Mazz\] but for Be.[]{data-label="fig:Be Mazz"}](f4.eps) ![Same as Fig. \[fig:H Mazz\] but for B.[]{data-label="fig:B Mazz"}](f5.eps) ![Same as Fig. \[fig:H Mazz\] but for C.[]{data-label="fig:C Mazz"}](f6.eps) ![Same as Fig. \[fig:H Mazz\] but for N.[]{data-label="fig:N Mazz"}](f7.eps) ![Same as Fig. \[fig:H Mazz\] but for O.[]{data-label="fig:O Mazz"}](f8.eps) ![Same as Fig. \[fig:H Mazz\] but for F.[]{data-label="fig:F Mazz"}](f9.eps) ![Same as Fig. \[fig:H Mazz\] but for Ne.[]{data-label="fig:Ne Mazz"}](f10.eps) ![Same as Fig. \[fig:H Mazz\] but for Na.[]{data-label="fig:Na Mazz"}](f11.eps) ![Same as Fig. \[fig:H Mazz\] but for Mg.[]{data-label="fig:Mg Mazz"}](f12.eps) ![Same as Fig. \[fig:H Mazz\] but for Al.[]{data-label="fig:Al Mazz"}](f13.eps) ![Same as Fig. \[fig:H Mazz\] but for Si.[]{data-label="fig:Si Mazz"}](f14.eps) ![Same as Fig. \[fig:H Mazz\] but for P.[]{data-label="fig:P Mazz"}](f15.eps) ![Same as Fig. \[fig:H Mazz\] but for S.[]{data-label="fig:S Mazz"}](f16.eps) ![Same as Fig. \[fig:H Mazz\] but for Cl.[]{data-label="fig:Cl Mazz"}](f17.eps) ![Same as Fig. \[fig:H Mazz\] but for Ar.[]{data-label="fig:Ar Mazz"}](f18.eps) ![Same as Fig. \[fig:H Mazz\] but for K.[]{data-label="fig:K Mazz"}](f19.eps) ![Same as Fig. \[fig:H Mazz\] but for Ca.[]{data-label="fig:Ca Mazz"}](f20.eps) ![Same as Fig. \[fig:H Mazz\] but for Sc.[]{data-label="fig:Sc Mazz"}](f21.eps) ![Same as Fig. \[fig:H Mazz\] but for Ti.[]{data-label="fig:Ti Mazz"}](f22.eps) ![Same as Fig. \[fig:H Mazz\] but for V.[]{data-label="fig:V Mazz"}](f23.eps) ![Same as Fig. \[fig:H Mazz\] but for Cr.[]{data-label="fig:Cr Mazz"}](f24.eps) ![Same as Fig. \[fig:H Mazz\] but for Mn.[]{data-label="fig:Mn Mazz"}](f25.eps) ![Same as Fig. \[fig:H Mazz\] but for Fe.[]{data-label="fig:Fe Mazz"}](f26.eps) ![Same as Fig. \[fig:H Mazz\] but for Co.[]{data-label="fig:Co Mazz"}](f27.eps) ![Same as Fig. \[fig:H Mazz\] but for Ni.[]{data-label="fig:Ni Mazz"}](f28.eps) ![Same as Fig. \[fig:H Mazz\] but for Cu.[]{data-label="fig:Cu Mazz"}](f29.eps) ![Same as Fig. \[fig:H Mazz\] but for Zn.[]{data-label="fig:Zn Mazz"}](f30.eps) ![Ionization fractional abundance versus electron temperature for all ionization stages of H. The upper graph shows our results ([solid curves]{}) and the abundances calculated by Bryans et al. (2006; [dashed curves]{}). The lower graph shows the ratio of the calculated abundances. Comparison is made only for fractional abundances greater than $10^{-2}$. We label our results as “New” and those of @Brya06a as “Old”.[]{data-label="fig:H Bryans"}](f31.eps) ![Same as Fig. \[fig:H Bryans\] but for He.[]{data-label="fig:He Bryans"}](f32.eps) ![Same as Fig. \[fig:H Bryans\] but for Li.[]{data-label="fig:Li Bryans"}](f33.eps) ![Same as Fig. \[fig:H Bryans\] but for Be.[]{data-label="fig:Be Bryans"}](f34.eps) ![Same as Fig. \[fig:H Bryans\] but for B.[]{data-label="fig:B Bryans"}](f35.eps) ![Same as Fig. \[fig:H Bryans\] but for C.[]{data-label="fig:C Bryans"}](f36.eps) ![Same as Fig. \[fig:H Bryans\] but for N.[]{data-label="fig:N Bryans"}](f37.eps) ![Same as Fig. \[fig:H Bryans\] but for O.[]{data-label="fig:O Bryans"}](f38.eps) ![Same as Fig. \[fig:H Bryans\] but for F.[]{data-label="fig:F Bryans"}](f39.eps) ![Same as Fig. \[fig:H Bryans\] but for Ne.[]{data-label="fig:Ne Bryans"}](f40.eps) ![Same as Fig. \[fig:H Bryans\] but for Na.[]{data-label="fig:Na Bryans"}](f41.eps) ![Same as Fig. \[fig:H Bryans\] but for Mg.[]{data-label="fig:Mg Bryans"}](f42.eps) ![Same as Fig. \[fig:H Bryans\] but for Al.[]{data-label="fig:Al Bryans"}](f43.eps) ![Same as Fig. \[fig:H Bryans\] but for Si.[]{data-label="fig:Si Bryans"}](f44.eps) ![Same as Fig. \[fig:H Bryans\] but for P.[]{data-label="fig:P Bryans"}](f45.eps) ![Same as Fig. \[fig:H Bryans\] but for S.[]{data-label="fig:S Bryans"}](f46.eps) ![Same as Fig. \[fig:H Bryans\] but for Cl.[]{data-label="fig:Cl Bryans"}](f47.eps) ![Same as Fig. \[fig:H Bryans\] but for Ar.[]{data-label="fig:Ar Bryans"}](f48.eps) ![Same as Fig. \[fig:H Bryans\] but for K.[]{data-label="fig:K Bryans"}](f49.eps) ![Same as Fig. \[fig:H Bryans\] but for Ca.[]{data-label="fig:Ca Bryans"}](f50.eps) ![Same as Fig. \[fig:H Bryans\] but for Sc.[]{data-label="fig:Sc Bryans"}](f51.eps) ![Same as Fig. \[fig:H Bryans\] but for Ti.[]{data-label="fig:Ti Bryans"}](f52.eps) ![Same as Fig. \[fig:H Bryans\] but for V.[]{data-label="fig:V Bryans"}](f53.eps) ![Same as Fig. \[fig:H Bryans\] but for Cr.[]{data-label="fig:Cr Bryans"}](f54.eps) ![Same as Fig. \[fig:H Bryans\] but for Mn.[]{data-label="fig:Mn Bryans"}](f55.eps) ![Same as Fig. \[fig:H Bryans\] but for Fe.[]{data-label="fig:Fe Bryans"}](f56.eps) ![Same as Fig. \[fig:H Bryans\] but for Co.[]{data-label="fig:Co Bryans"}](f57.eps) ![Same as Fig. \[fig:H Bryans\] but for Ni.[]{data-label="fig:Ni Bryans"}](f58.eps) ![Same as Fig. \[fig:H Bryans\] but for Cu.[]{data-label="fig:Cu Bryans"}](f59.eps) ![Same as Fig. \[fig:H Bryans\] but for Zn.[]{data-label="fig:Zn Bryans"}](f60.eps) ![$EM$ versus $T_e$ curves of all emission lines observed here from Si. The dashed lines indicate the mean $\log_{10}EM$ and $\log_{10}T_e$ and the dotted lines show the standard deviations of these values. Asterisks indicate where the curves cross. The left panel shows the results after Step 1 of the analysis, the middle panel shows the results after Step 2, and the right panel shows the results after Step 3. See Sec. \[sec:method\] for a description of each step.[]{data-label="fig:si"}](f61.eps) ![$EM$ versus $T_e$ curves of all the emission lines from each of the low- and moderate-FIP elements using the GEM method described in Sec. \[sec:method\]. Na, K and Ca are excluded as in this SUMER dataset there are not enough observed emission lines from these elements to determine a mean $EM$. The upper left panel shows the high-FIP elements Ne and Ar.[]{data-label="fig:all elem"}](f62.eps) ![$EM$ versus $T_e$ curves of all the emission lines from each of the elements Ne, Mg, Al, Si, S, Ar, and Fe (excluding Li- and Na-like ions). The $T_e$ derived from these elements were used to determine the FIP factors of Na, K and Ca. Asterisks indicate where the curves cross. The three panels show the three steps of the GEM method as in Fig. \[fig:si\].[]{data-label="fig:te"}](f63.eps) ![Coronal abundance enhancement factor (i.e., FIP factor) used for each of the elements versus their first ionization potential. Open circles indicate the high-FIP elements O, N, Ne and Ar where no enhancement was assumed. The elements in squares indicate those which had their enhancement factor determined by matching their mean $EM$ with that of the high-FIP element Ar. Elements marked with diamonds are those that did not have enough crossings to determine their $EM$ in this way (see Sec. \[sec:fip\] for further details). The dotted line is purely to guide the eye. The solid circles are the results of @Feld98a for N, O, Ne, Na, Mg, Si, S, Ca, and Fe. Feldman et al. scaled their results to O and assumed a FIP factor of 1 for this element. We also set the O FIP factor to 1, so their and our FIP factors for O lie directly on top of one another in this plot. []{data-label="fig:fip"}](f64.eps) ![$EM$ versus $T_e$ curves for the emission lines of Group Ia using our inferred coronal abundances. Asterisks indicate where the curves cross. The three panels show the three steps of the GEM method as in Fig. \[fig:si\].[]{data-label="fig:1a"}](f65.eps) ![Same as Fig. \[fig:1a\] but for Group Ib.[]{data-label="fig:1b"}](f66.eps) ![Same as Fig. \[fig:1a\] but for Group I as a whole.[]{data-label="fig:1"}](f67.eps) ![Same as Fig. \[fig:1a\] but for Group IIa.[]{data-label="fig:2a"}](f68.eps) ![Same as Fig. \[fig:2a\] but excluding emission lines from N [v]{} and O [vi]{}.[]{data-label="fig:2ax"}](f69.eps) ![Same as Fig. \[fig:2a\] but showing only emission lines from N [v]{} and O [vi]{}.[]{data-label="fig:2a_on"}](f70.eps) ![Same as Fig. \[fig:1a\] but for Group IIb.[]{data-label="fig:2b"}](f71.eps) ![Same as Fig. \[fig:1a\] but for Group II as a whole, excluding emission lines from N [v]{} and O [vi]{}.[]{data-label="fig:2"}](f72.eps) ![Same as Fig. \[fig:1a\] but for Group III.[]{data-label="fig:3"}](f73.eps) ![Same as Fig. \[fig:1a\] but for all emission lines in the observation except those from N [v]{} and O [vi]{}. Due to the large number of crossings we exclude the asterisks for clarity.[]{data-label="fig:all"}](f74.eps) ![Same as Fig. \[fig:all\] but excluding emission lines from all Li- and Na-like ions.[]{data-label="fig:allx"}](f75.eps) ![Mean $\log_{10}EM$ values for each of the groups using the GEM method (as listed in Table \[tab:averages\]). The numbers in the data points represent the number of $EM$ curve crossings that were used to derive the mean $EM$. The error bars on the points are $\pm \delta\langle\log_{10}EM\rangle$. Group IIa$^1$ excludes the O [vi]{} and N [v]{} lines. The dashed and dotted lines indicate the mean and standard deviation, respectively, when every emission line, except N [v]{} and O [vi]{}, is considered. The thick lines include emission lines from Li- and Na-like ions (Fig. \[fig:all\]) and the thin lines exclude emission lines from these ions (Fig. \[fig:allx\]). []{data-label="fig:em figs"}](f76.eps) ![Same as Fig. \[fig:em figs\] but for $\log_{10}T_e$. The upper values of the standard deviation for the thick and thin lines lie on top of one another.[]{data-label="fig:temp figs"}](f77.eps) ![Coronal abundance enhancement factor (i.e., FIP factor) used for each of the elements versus their first ionization potential. Open symbols represent the present results; refer to Fig. \[fig:fip\] for details. The solid circles are the results of the model of @Lami08a for for upward Alfvén wave energy fluxes of 2, 8, and 32 ([solid]{}, [dashed]{}, and [dot-dashed lines]{}, respectively) in units of $10^6$ ergs cm$^{-2}$ s$^{-1}$. Lines have been drawn between points only to guide the eye.[]{data-label="fig:fip model"}](f78.eps)	{ "pile_set_name": "ArXiv" }
--- abstract: 'In this paper we study properties of effective impedance of finite electrical networks and calculate the effective impedance of a finite ladder network over an ordered field. Moreover, we consider two particular examples of infinite ladder networks (Feynman’s network or $LC$-network and $CL$-network, both with zero on infinity) as networks over the ordered Levi-Civita field $\mathcal R$. We show, that effective impedances of finite $LC$-networks converge to the limit in order topology of $\mathcal R$, but the effective impedances of finite $CL$-networks do not converge in the same topology.' address: 'Anna Muranova: IRTG 2235, University Bielefeld, Postfach 10 01 31, 33501 Bielefeld, Germany' author: - Anna Muranova title: Effective impedance over ordered fields --- [Mathematics Subject Classification 2010:]{}[ 05C22, 34B45, 05C25, 39A12, 12J15.]{} Introduction ============ It is known that electrical networks with resistances are related to weighted graphs (see e.g. [@DS], [@LPW], [@Soardi]). Moreover, it is shown in [@Grimmett] and [@LPW], that effective resistance for finite networks satisfies the basic physical properties (e.g. parallel and series laws). In [@DS] and [@Grimmett] the notion of effective resistance for infinite network is introduced. An effective resistance is tightly related to random walk and Dirichlet problem on graphs, which are described in many papers and books (e. g. [@Barlow], [@Grigoryan], [@Woess],[@Soardi]). In [@Muranova] a finite electrical network with alternating current and passive elements is considered as a generalization of electrical network with resistances. It is shown there, that such a network is related to weighted graphs over non-Archimedean ordered field of rational functions $\Bbb R(\lambda)$. The generalization of effective resistance for this case is called effective impedance. The inverse of effective impedance is called effective admittance. The most known in physics infinite network with passive elements is Feynman’s ladder network ($LC$-network, see [@Feynman]). In [@Yoon] the effective impedances of $LC$-network and $CL$-network are considered as limits of complex-valued effective impedances of corresponding sequences of finite networks. The present paper consists of two parts. In the first part we describe some properties of electrical network over an ordered field. The main result of this part is Theorem \[SM\], which gives the mathematical description of well-known in physics star-mesh transform. The mathematical conceptions of parallel and series laws, as well as $\Delta-Y$ and $Y-\Delta$ transform, follow from Theorem \[SM\] as corollaries. In the second part of the present paper we discuss the question whether one can generalize the notion of effective resistance for infinite networks with zero on infinity (see e.g. [@Grimmett]) for the case of non-Archimedean weighted graphs. The main theorem of this section is Theorem \[Thmmonotonicity\]. It shows, that a sequence of effective admittances of finite networks, exhausted a given infinite network, decreases. Unfortunately, it does not give a convergence over non-Archimedean field. As examples, we consider $LC$-network and $CL$-network (with zero on infinity) as electrical networks over ordered Levi-Civita field $\mathcal R$, which contains a subfield isomorphic to $\Bbb R(\lambda)$ (see [@Berz], [@Hall]). Firstly, we present the general calculation of admittance of a finite ladder network (see Figure \[figFin\]) over an ordered field. Then the closeness of the Levi-Civita field in order topology ([@Berz], [@Shams]) gives us an opportunity to arise the question whether effective admittance of infinite network could be defined as limit of effective admittances of corresponding finite electrical networks in this case. We show, that in case of the $LC$-network the sequence of effective admittances of finite networks converge in ordered topology of the Levi-Civita field (Theorem \[ThmLC\]). Moreover, we show, that admittances of finite $CL$-network do not converge in the same topology (Example \[ExCL\]). This shows, that in general it is not possible to generalize the notion of effective resistance for infinite networks for the case of non-Archimedean weights. Properties of effective impedance of the finite network over ordered field ========================================================================== A network over an ordered field $(K,\succ)$ is a structure $$\Gamma=(V,\rho,a_0, B),$$ where - $(V,E)$ is a locally finite connected graph ($\|V\|\ge 2$), - $\rho:E\rightarrow K$ is a positive function called admittance, - $a_0\in V$ is a fixed vertex, - $B=\{a_1,\dots,a_{\|B\|}\}\in V\setminus\{a_0\}$ is a fixed non-empty subset of vertices. Let us denote by $B_0=B\cup\{a_0\}$ the set of boundary vertices and by $z=\frac{1}{\rho}$ the positive function of impedance. The network is called finite if $\|V\|<\infty$. Otherwise, it is called infinite. Note that we can consider $\rho$ as a function from $V\times V$ to $K$ by setting $\rho_{xy}=0$, if $xy$ is not an edge. Then the weight $\rho_{xy}$ gives rise to a function on vertices as follows: $$\rho(x)=\sum_y \rho_{xy},$$ where the notation $\sum\limits_y$ means $\sum\limits_{y\in V}$. Then $\rho(x)$ is called the weight of a vertex $x$. We have $0<\rho(x)<\infty$ for any vertex $x$ of a network . Let us consider the following Dirichlet problem on the given finite network $\Gamma=(V,\rho,a_0, B)$: $$\label{dirpr} \begin{cases} \Delta_\rho v(x)=0 \mbox { on } V\setminus B_0, \\ v(x)= 0 \mbox { on } B, \\ v(a_0)=1, \end{cases}$$ where $\Delta_\rho v(x)=\sum_y(v(y)-v(x))\rho_{xy}.$ The physical meaning of Dirichlet problem is the following: if we take $K=\Bbb R(\lambda)$ and admittance of each edge in the form $$\rho_{xy}=\frac{\lambda}{ L_{xy}\lambda^2+R_{xy}\lambda+{D_{xy}}}, \mbox{ }L_{xy},R_{xy},D_{xy}\ge0 \mbox{ and } L_{xy}+R_{xy}+D_{xy}\ne 0,$$ then the real voltage at the vertex $x$ at time $t$ will be equal to $\Re(v(x)e^{i\omega t})$, assuming that we keep potential $1$ at the vertex $a_0$, ground all the vertices from $B$ and apply alternating current of frequency $\omega=-i\lambda$ to the network (see [@Muranova]). Note that if $\|V\|=n$, then the Dirichlet problem is a $n\times n$ system of linear equations over the field $K$. It can be also written in a matrix form (note, that here we already have substituted $v(a_0)=1, v(a_i)=0, i=\overline{1,\|B\|}$ in the first $(n-\|B\|-1)$ equations): $$\label{dirprM} \begin{cases} A {\hat v}={b}, \\ v(a_0)=1, \\ v(a_i)=0, i=\overline{1,k}, \end{cases}$$ where $k=\|B\|$, $A$ is a symmetric matrix ($A=A^T$), ${\hat v},{b}$ are vector-columns of length $(n-k-1)$: $$A= \begin{bmatrix} \sum_{x\sim x_1}\rho_{xx_1}&-\rho_{x_1x_2}&\dots&-\rho_{x_1x_{n-k-1}}\\ -\rho_{x_1x_2}&\sum_{x\sim x_2}\rho_{xx_2}&\dots&-\rho_{x_2x_{n-k-1}}\\ \dots\\ -\rho_{x_1x_{n-k-1}}&-\rho_{x_2x_{n-k-1}}&\dots&\sum_{x\sim x_{n-k-1}}\rho_{xx_{n-k-1}}\\ \end{bmatrix},$$ $$b=(\rho_{a_0x_1},\rho_{a_0x_2},\dots,\rho_{a_0x_{n-k-1}})^T,$$ $$\hat v=(v(x_1),v(x_2),\dots,v(x_{n-k-1}))^T.$$ In [@Muranova] it is proved, that the Dirichlet problem has a unique solution for any finite network over an ordered field. We define effective impedance of the finite network $\Gamma$ as $$Z_{eff}(\Gamma)=\frac{1}{\sum_{x:x\sim a_0}(1-v(x))\rho_{xa_0}},$$ where $v$ is the solution of the Dirichlet problem . The effective admittance is defined by $$\mathcal P_{eff}{(\Gamma)}=\sum_{x:x\sim a_0}(1-v(x))\rho_{xa_0}.$$ For the solution $v$ of the Dirichlet problem we have $$\label{differentEqforP} \mathcal P_{eff}{(\Gamma)}=\sum_{i=1}^{\|B\|}\sum_{x:x\sim a_i}v(x)\rho_{xa_i}=\sum_{i=1}^{\|B\|}\Delta_\rho v(a_i)=-\Delta_\rho v(a_0)=\frac{1}{2}\sum_{\substack{x\sim y\\x,y\in V}}(\nabla_{xy}v)^2\rho_{xy},$$ where $\nabla_{xy}v=v(y)-v(x)$. The proof of this result follows the same outline as the proof of the similar result in [@Muranova]. [(Star-mesh transform)]{}\[SM\] Let $\Gamma = (V,\rho,a_0,B)$ be a finite network, $\|V\|=n$, $B_0=B\cup\{a_0\}$, and $x_1,\dots ,x_m\in V$, $3 \le m \le n$, are such that 1. $x_1\not \in B_0$, 2. $y\not\sim x_1$ for all $y \in V\setminus\{x_2,\dots ,x_m\}$, If one removes the vertex $x_1$, edges $(x_1,x_i), i=\overline{2,m}$ and change the admittances of the edges $(x_i,x_j), i,j=\overline{2,m}, i\ne j$ as follows: $$\label{rhoSM} \rho'_{x_ix_j}=\rho_{x_ix_j}+\frac{\rho_{x_1x_i}\rho_{x_1x_j}}{\rho(x_1)},$$ not changing the other admittances, then for the new network the solution of the Dirichlet problem for all the vertices will be the same as the solution of the Dirichlet problem on the original network at corresponding vertices. \(0) [$x_1$]{}; \(1) \[left of=0\] [$x_9$]{}; (2) \[above left of=0\] [$x_2$]{}; (3) \[above of=0\] [$x_3$]{}; (4) \[above right of=0\] [$x_4$]{}; (5) \[right of=0\] [$x_5$]{}; (6) \[below right of=0\] [$x_6$]{}; (7) \[below of=0\] [$x_7$]{}; (8) \[below left of=0\] [$x_8$]{}; \(0) edge node \[bend right\] (1) (0) edge node \[bend right\] (2) (0) edge node \[bend right\] (3) (0) edge node \[bend right\] (4) (0) edge node \[bend right\] (5) (0) edge node \[bend right\] (6) (0) edge node \[bend right\] (7) (0) edge node \[bend right\] (8) (1) edge node \[bend right\] (2) (2) edge node \[bend right\] (3) (3) edge node \[bend right\] (4) (4) edge node \[bend right\] (5) (5) edge node \[bend right\] (6) (6) edge node \[bend right\] (7) (7) edge node \[bend right\] (8) (1) edge node \[bend right\] (8); (10)\[right of=5\] [$x_9$]{}; (00)\[right of=10\] ; (20) \[above left of=00\] [$x_2$]{}; (30) \[above of=00\] [$x_3$]{}; (40) \[above right of=00\] [$x_4$]{}; (50) \[right of=00\] [$x_5$]{}; (60) \[below right of=00\] [$x_6$]{}; (70) \[below of=00\] [$x_7$]{}; (80) \[below left of=00\] [$x_8$]{}; \(10) edge node \[bend right\] (20) (10) edge node \[bend right\] (30) (10) edge node \[bend right\] (40) (10) edge node \[bend right\] (50) (10) edge node \[bend right\] (60) (10) edge node \[bend right\] (70) (10) edge node \[bend right\] (80) \(20) edge node \[bend right\] (30) (20) edge node \[bend right\] (40) (20) edge node \[bend right\] (50) (20) edge node \[bend right\] (60) (20) edge node \[bend right\] (70) (20) edge node \[bend right\] (80) \(30) edge node \[bend right\] (40) (30) edge node \[bend right\] (50) (30) edge node \[bend right\] (60) (30) edge node \[bend right\] (70) (30) edge node \[bend right\] (80) \(40) edge node \[bend right\] (50) (40) edge node \[bend right\] (60) (40) edge node \[bend right\] (70) (40) edge node \[bend right\] (80) \(50) edge node \[bend right\] (60) (50) edge node \[bend right\] (70) (50) edge node \[bend right\] (80) \(60) edge node \[bend right\] (70) (60) edge node \[bend right\] (80) (70) edge node \[bend right\] (80); Let us consider the Dirichlet problem for the network $\Gamma$ in a matrix form (\[dirprM\]). Obviously, it is enough to solve the matrix equation $A \hat v=b$. Without loss of generality we can assume that $x_1,\dots,x_l\not\in B_0$, where $l=m-\|\{x_1,\dots,x_m\}\cap B_0\|$. Writing equations for $x_1,\dots,x_l$ as the first ones and denoting $k=\|B\|$, we have $$A= \begin{bmatrix} \rho(x_1)&-\rho_{x_1x_2}&\dots&-\rho_{x_1x_l}&0&\dots&0\\ -\rho_{x_1x_2}&\rho(x_2)&\dots&-\rho_{x_2x_l}&-\rho_{x_2x_{m+1}}&\dots&-\rho_{x_2x_{n-k-1}}\\ \dots&\dots&\dots&\dots&\dots&\dots&\dots\\ -\rho_{x_1x_l}&-\rho_{x_2x_l}&\dots&\rho(x_l)&-\rho_{x_lx_{m+1}}&\dots&-\rho_{x_lx_{n-k-1}}\\ 0&-\rho_{x_2x_{m+1}}&\dots&-\rho_{x_lx_{m+1}}&\rho(x_{m+1})&\dots&-\rho_{x_{m+1}x_{n-k-1}}\\ \dots&\dots&\dots&\dots&\dots&\dots&\dots\\ 0&-\rho_{x_2x_{n-k-1}}&\dots&-\rho_{x_lx_{n-k-1}}&-\rho_{x_{m+1}x_{n-k-1}}&\dots&\rho(x_{n-k-1})\\ \end{bmatrix},$$ since $y\not\sim x_1$ for all $y \in V\setminus\{x_2,\dots ,x_m\}$, and $$b=(\rho_{a_0x_1},\rho_{a_0x_2},\dots,\rho_{a_0x_l},\rho_{a_0x_{m+1}},\dots,\rho_{a_0x_{n-k-1}})^T,$$ Now it is easy to calculate, that the star-mesh transform is just applying of Gaussian elimination method for the first row. Indeed, applying Gaussian elimination method for the first row of the augmented matrix $\bar A=[A\|b]$ we obtain: $$\left[ \begin{array}{ccccccc\|c} 1&-\frac{\rho_{x_1x_2}}{\rho(x_1)}&\dots&-\frac{\rho_{x_1x_l}}{\rho(x_1)}&0&\dots&0&\frac{\rho_{a_0x_1}}{\rho(x_1)}\\ 0&\rho^(x_2)&\dots&-\rho'_{x_2x_l}&-\rho_{x_2x_{m+1}}&\dots&-\rho_{x_2x_{n-k-1}}&\rho^_{a_0x_2}\\ \dots&\dots&\dots&\dots&\dots&\dots&\dots&\dots\\ 0&-\rho'_{x_2x_l}&\dots&\rho^(x_l)&-\rho_{x_lx_{m+1}}&\dots&-\rho_{x_lx_{n-k-1}}&\rho^_{a_0x_l}\\ 0&-\rho_{x_2x_{m+1}}&\dots&-\rho_{x_lx_{m+1}}&\rho(x_{m+1})&\dots&-\rho_{x_{m+1}x_{n-k-1}}&\rho_{a_0x_{m+1}}\\ \dots&\dots&\dots&\dots&\dots&\dots&\dots&\dots\\ 0&-\rho_{x_2x_{n-k-1}}&\dots&-\rho_{x_lx_{n-k-1}}&-\rho_{x_{m+1}x_{n-k-1}}&\dots&\rho(x_{n-k-1})&\rho_{a_0x_{n-k-1}}\\ \end{array} \right]$$ since $\rho(x_1)\ne 0$, where $$\rho^(x_i)=\rho(x_i)-\frac{\rho_{x_1x_i}^2}{\rho(x_1)} \mbox{ and } \rho^_{a_0x_i}=\rho_{a_0x_i}+\frac{\rho_{x_1x_i}\rho_{a_0x_1}}{\rho(x_1)} \mbox{ for all } i=\overline{2,l}.$$ Note, that for all $i=\overline{2,l}$ $$\begin{split} \rho'(x_i)=&\rho(x_i) -\rho_{x_1x_i}-\sum_{\substack{j=2\\j\ne i}}^m \rho_{x_ix_j}+\sum_{\substack{j=2\\j\ne i}}^m \rho'_{x_ix_j}\\ =&\rho(x_i) -\rho_{x_1x_i}-\sum_{\substack{j=2\\j\ne i}}^m \rho_{x_ix_j}+\sum_{\substack{j=2\\j\ne i}}^m \left(\rho_{x_ix_j}+\frac{\rho_{x_1x_i}\rho_{x_1x_j}}{\rho(x_1)}\right)\\ =&\rho(x_i) -\rho_{x_1x_i}+\sum_{{\substack{j=2\\j\ne i}}}^m \frac{\rho_{x_1x_i}\rho_{x_1x_j}}{\rho(x_1)}\\ =&\rho(x_i) -\rho_{x_1x_i}+\frac{\rho_{x_1x_i}}{\rho(x_1)}\sum_{{\substack{j=2\\j\ne i}}}^m\rho_{x_1x_j} \\ =&\rho(x_i) -\rho_{x_1x_i}+\frac{\rho_{x_1x_i}}{\rho(x_1)}\sum_{j=2}^m\rho_{x_1x_j}-\frac{\rho_{x_1x_i}}{\rho(x_1)}\rho_{x_1x_i}\\ =&\rho(x_i) -\rho_{x_1x_i}+\frac{\rho_{x_1x_i}}{\rho(x_1)}\rho(x_1)-\frac{\rho_{x_1x_i}^2}{\rho(x_1)}\\ =&\rho(x_i)-\frac{\rho_{x_1x_i}^2}{\rho(x_1)}=\rho^(x_i) \end{split}$$ and $$\begin{split} \rho^_{a_0x_i}=\rho_{a_0x_i}+\frac{\rho_{x_1x_i}\rho_{a_0x_1}}{\rho(x_1)}= \begin{cases} \rho'_{a_0x_i}, & \text{if}\ a_0 \in\{x_2,\dots,x_m\}\\ \rho_{a_0x_i}, & \text{otherwise, since}\ \rho_{a_0x_1}=0. \end{cases} \end{split}$$ Hence, $\bar A=$ $$\left[ \begin{array}{ccccccc\|c} 1&-\frac{\rho_{x_1x_2}}{\rho(x_1)}&\dots&-\frac{\rho_{x_1x_l}}{\rho(x_1)}&0&\dots&0&\frac{\rho_{a_0x_1}}{\rho(x_1)}\\ 0&\rho'(x_2)&\dots&-\rho'_{x_2x_l}&-\rho_{x_2x_{m+1}}&\dots&-\rho_{x_2x_{n-k-1}}&\rho'_{a_0x_2}\\ \dots&\dots&\dots&\dots&\dots&\dots&\dots&\dots\\ 0&-\rho'_{x_2x_l}&\dots&\rho'(x_l)&-\rho_{x_lx_{m+1}}&\dots&-\rho_{x_lx_{n-k-1}}&\rho'_{a_0x_l}\\ 0&-\rho_{x_2x_{m+1}}&\dots&-\rho_{x_lx_{m+1}}&\rho(x_{m+1})&\dots&-\rho_{x_{m+1}x_{n-k-1}}&\rho_{a_0x_{m+1}}\\ \dots&\dots&\dots&\dots&\dots&\dots&\dots&\dots\\ 0&-\rho_{x_2x_{n-k-1}}&\dots&-\rho_{x_lx_{n-k-1}}&-\rho_{x_{m+1}x_{n-k-1}}&\dots&\rho(x_{n-k-1})&\rho_{a_0x_{n-k-1}}\\ \end{array} \right],$$ Therefore, we can eliminate the variable $v(x_1)$ from the Dirichlet problem, changing admittances as in the statement of the theorem. \[CorSM\] Under the star-mesh transform of the network the effective impedance and effective admittance do not change. In the proof we will use the notations from the proof of the Theorem \[SM\]. The case $\{x_1,\dots, x_m\}\cap B_0=\emptyset$ is trivial. The cases, when $\{x_1,\dots, x_m\}\cap B=\emptyset$ or $\{x_1,\dots, x_m\}\cap \{a_0\}=\emptyset$ are obvious, due to .\ Otherwise, we can assume, without loss of generality, that $$x_{m-j}=a_j,j=\overline{0,\|\{x_1,\dots, x_m\}\cap B_0\|},$$ in particular, $x_m=a_0$. Then, if we denote the new network by $\Gamma'$ we have by $$\begin{aligned} \mathcal P_{eff}(\Gamma)=&-\Delta_\rho v(a_0)=\sum_{x\ne a_0}(1-v(x))\rho_{x a_0}\\ =&(1-v(x_1))\rho_{x_1 a_0}+\sum_{i=2}^{m-1} (1-v(x_i))\rho_{x_ia_0}+\sum_{x\not\in\{x_1,\dots,x_m\}}(1-v(x))\rho_{xa_0}\\ =&\mathcal P_{eff}(\Gamma')-\sum_{i=2}^{m-1} (1-v(x_i))\rho'_{x_ia_0}+(1-v(x_1))\rho_{x_1 a_0}+\sum_{i=2}^{m-1} (1-v(x_i))\rho_{x_ia_0}\\ =&\mathcal P_{eff}(\Gamma')-\sum_{i=2}^{m-1} (1-v(x_i))\frac{\rho_{x_1a_0}\rho_{x_1x_i}}{\rho(x_1)}+(1-v(x_1))\rho_{x_1 a_0}\\ =&\mathcal P_{eff}(\Gamma')-\sum_{i=2}^{m-1} (1-v(x_i))\frac{\rho_{x_1a_0}\rho_{x_1x_i}}{\rho(x_1)}+\left(1-\sum_{i=2}^{m-1}v(x_i)\frac{\rho_{x_1x_i}}{\rho(x_1)}-\frac{\rho_{x_1a_0}}{\rho(x_1)}\right)\rho_{x_1 a_0}\\ =&\mathcal P_{eff}(\Gamma')-\rho_{x_1a_0}\sum_{i=2}^{m-1} (1-v(x_i))\frac{\rho_{x_1x_i}}{\rho(x_1)}+\left(1-\sum_{i=2}^{m-1}v(x_i)\frac{\rho_{x_1x_i}}{\rho(x_1)}-\frac{\rho_{x_1a_0}}{\rho(x_1)}\right)\rho_{x_1 a_0}\\ =&\mathcal P_{eff}(\Gamma')-\rho_{x_1a_0}\left(\sum_{i=2}^{m-1} \frac{\rho_{x_1x_i}}{\rho(x_1)}-\sum_{i=2}^{m-1} v(x_i)\frac{\rho_{x_1x_i}}{\rho(x_1)}\right)+\left(1-\sum_{i=2}^{m-1}v(x_i)\frac{\rho_{x_1x_i}}{\rho(x_1)}-\frac{\rho_{x_1a_0}}{\rho(x_1)}\right)\rho_{x_1 a_0}\\ =&\mathcal P_{eff}(\Gamma')-\rho_{x_1a_0}\left(1-\frac{\rho_{x_1x_m}}{\rho(x_1)}-\sum_{i=2}^{m-1} v(x_i)\frac{\rho_{x_1x_i}}{\rho(x_1)}\right)+\left(1-\sum_{i=2}^{m-1}v(x_i)\frac{\rho_{x_1x_i}}{\rho(x_1)}-\frac{\rho_{x_1a_0}}{\rho(x_1)}\right)\rho_{x_1 a_0}\\ =&\mathcal P_{eff}(\Gamma')\end{aligned}$$ since $$\rho(x_1)=\sum_{i=2}^m\rho_{x_1x_i}\mbox{ and }v(x_1)=\sum_{i=2}^{l}v(x_i)\frac{\rho_{x_1x_i}}{\rho(x_1)}+\frac{\rho_{x_1a_0}}{\rho(x_1)}=\sum_{i=2}^{m-1}v(x_i)\frac{\rho_{x_1x_i}}{\rho(x_1)}+\frac{\rho_{x_1a_0}}{\rho(x_1)}$$ (see the first line of $\bar A$ and note that $v(x_j)=0$ for all $j=\overline{j+1,m-1}$ and $a_0=x_m$). Series law and $Y-\Delta$ transform are just particular cases of star-mesh transform. Since multigraphs are not allowed in this paper, we will use a modification of parallel law and call it parallel-series law. [(Series law)]{}\[SeriesLaw\] Let $\Gamma = (V,\rho,a_0,B)$ be a finite network, $B_0=B\cup\{a_0\}$. Let $a,b,c\in V$ are such, that 1. $b\not\in B_0$, 2. $a\not\sim c$, $a\sim b$, $b\sim c$, 3. $b\not\sim x$ for all $x\not \in \{a,c\}$. If one removes the vertex $b$, edges $(a,b),(b,c)$ and add the edge $(a,c)$ with the addmittance $$\rho'_{ac}=\frac{\rho_{ab}\rho_{ac}}{\rho_{ab}+\rho_{ac}},$$ not changing other admittances, then for the new network the solution of the Dirichlet problem for all the vertices will be the same as the solution of the Dirichlet problem on the original network at corresponding vertices. The effective impedance (admittance) of the new network coincides with the effective impedance (addmittance) of the original one. The corresponding equation for impedances is then $$z'_{ac}={z_{ab}+z_{ac}},$$ which corresponds to the well-known physical series law. \(1) [$a$]{}; (2) \[right of=1\] [$b$]{}; (3) \[right of=2\] [$c$]{}; \(1) edge node \[bend right\] [$z_{ab}$]{} (2) (2) edge node \[bend right\] [$z_{bc}$]{} (3); \(10) \[right of=3\][$a$]{}; (20) \[right of=10\]\[node distance=4.5cm\] [$c$]{}; \(10) edge node \[bend right\][$z'_{ac}=z_{ab}+z_{bc}$]{} (20); Apply Theorem \[SM\] and Corollary \[CorSM\] ($x_1=b$) for the case $m=3$ and $\rho_{ac}=0$. [(Parallel-series law)]{}\[SPLaw\] Let $\Gamma = (V,\rho,a_0,B )$ be a finite network, $B_0=B\cup\{a_0\}$.\ Let $a,b,c\in V$ are such, that 1. $b\not \in B_0$, 2. $a\sim b,b\sim c, a\sim c$, 3. $b\not \sim x$ for all $x\not \in \{a,c\}$. Then if one removes the vertex $b$, edges $(a,b),(b,c)$ and add the edge $(a,c)$ with the admittance $$\rho'_{ac}=\frac{\rho_{ab}\rho_{bc}}{\rho_{ab}+\rho_{bc}}+\rho_{ac},$$ not changing other admittances, then for the new network the solution of the Dirichlet problem for all the vertices will be the same as the solution of the Dirichlet problem on the original network for corresponding vertices. The effective impedance (admittance) of the new network coincides with the effective impedance (admittance) of the original one. The corresponding equation for impedances is then $$\frac{1}{z'_{ac}}=\frac{1}{z_{ab}+z_{bc}}+\frac{1}{z_{ac}}$$ which corresponds to the application of the physical series law and then the physical parallel law. \(1) [$a$]{}; (2) \[above right of=1\] [$b$]{}; (3) \[below right of=2\] [$c$]{}; \(1) edge node \[bend right\] [$z_{ab}$]{} (2) (2) edge node \[bend right\] [$z_{bc}$]{} (3) (1) edge node \[bend right\] [$z_{ac}$]{} (3); \(10) \[right of=3\] [$a$]{}; (20) \[right of=10\] [$c$]{}; \(10) edge node \[bend right\] [$z'_{ac}$]{} (20); Apply Theorem \[SM\] and Corollary \[CorSM\] ($x_1=b$) for the case $m=3$. [($Y-\Delta$ transform)]{}\[YDLaw\] Let $\Gamma = (V,\rho,a_0,B)$ be a finite network, $B_0=B\cup\{a_0\}$. Let $a,b,c,d\in V$ are such, that 1. $d\not \in B_0$, 2. $d \sim a,d\sim b,d\sim c$, 3. $d\not \sim x$ for all $x\not \in \{a,b,c\}$. If one removes the vertex $d$, edges $(d,a),(d,b),(d,c)$ and set $$\label{YD} \begin{split} {\rho'_{ab}}=&\frac{\rho_{da}\rho_{db}}{\rho_{da}+\rho_{db}+\rho_{dc}}+\rho_{ab},\\ {\rho'_{bc}}=&\frac{\rho_{db}\rho_{dc}}{\rho_{da}+\rho_{db}+\rho_{dc}}+\rho_{bc}, \\ {\rho'_{ac}}=&\frac{\rho_{da}\rho_{dc}}{\rho_{da}+\rho_{db}+\rho_{dc}}+\rho_{ac}, \end{split}$$ not changing other admittances, then for the new network the solution of the Dirichlet problem for all the vertices will be the same as the solution of the Dirichlet problem on the original network for the corresponding vertices. The effective impedance (admittance) of the new network coincides with the effective impedance (admittance) of the original one. The corresponding equalities for the impedances are $$\begin{split} {z'_{ab}}=&{\frac{z_{dc}}{z_{da}z_{db}+z_{db}z_{dc}+z_{da}z_{dc}}+\frac{1}{z_{ab}}},\\ {z'_{bc}}=&{\frac{z_{da}}{z_{da}z_{db}+z_{db}z_{dc}+z_{da}z_{dc}}+\frac{1}{z_{bc}}}, \\ {z'_{ac}}=&{\frac{z_{db}}{z_{da}z_{db}+z_{db}z_{dc}+z_{da}z_{dc}}+\frac{1}{z_{ac}}}. \end{split}$$ From the physical point of view, if $\rho_{ab},\rho_{bc},\rho_{ac}$ are all equal to zero, then it is just $Y-\Delta$ transform, otherwise, it is $Y-\Delta$ transform and the parallel law. at (-0.4,-3.3) (1) [$a$]{} ; at (2,0) (2)[$b$]{}; at (4.4,-3.3) (3) [$c$]{}; at (2,-2) (4) [$d$]{}; \(4) edge node \[bend right\] [$z_{da}$]{} (1) (4) edge node \[bend right\] [$z_{db}$]{} (2) (3) edge node \[bend right\] [$z_{dc}$]{} (4) (1) edge node \[bend right\] [$z_{ab}$]{} (2) (2) edge node \[bend right\] [$z_{bc}$]{} (3) (3) edge node \[bend right\] [$z_{ac}$]{} (1); at (6.6,-3.3) (1) [$a$]{} ; at (9,0) (2)[$b$]{}; at (11.4,-3.3) (3) [$c$]{}; \(1) edge node \[bend right\] [$z'_{ab}$]{} (2) (2) edge node \[bend right\] [$z'_{bc}$]{} (3) (3) edge node \[bend right\] [$z'_{ac}$]{} (1); Theorem \[SM\] and Corollary \[CorSM\] ($x_1=d$) for the case $m=4$. The $Y-\Delta$ transform is invertible. In general, it is not the case for star-mesh transform. [($\Delta-Y$ transform)]{}\[DYLaw\] Let $\Gamma' = (V',\rho',a_0,B)$ be a finite network and let $a,b,c\in V$ are such, that $a \sim b,b \sim c$, and $a\sim c$. If one add a vertex $d$ and edges $(d,a),(d,b),(d,c)$ setting $$\begin{aligned} \rho_{da}&=\frac{\rho'_{ac}\rho'_{bc}+\rho'_{ac}\rho'_{ab}+\rho'_{ab}\rho'_{bc}}{\rho'_{bc}}, \\ \rho_{db}&=\frac{\rho'_{ac}\rho'_{bc}+\rho'_{ac}\rho'_{ab}+\rho'_{ab}\rho'_{bc}}{\rho'_{ac}}, \\ \rho_{dc}&=\frac{\rho'_{ac}\rho'_{bc}+\rho'_{ac}\rho'_{ab}+\rho'_{ab}\rho'_{bc}}{\rho'_{ab}},\end{aligned}$$ and remove the edges $(a,b),(b,c),(a,c)$ not changing other admittances, then for the new network $$\Gamma=(V\cup\{d\},\rho,a_0,B),$$ the solution of the Dirichlet problem for any vertex in $V$ will be the same as the solution of the Dirichlet problem on the original network for the corresponding vertex. Moreover, the effective impedance and effective admittance do not change under this transform. The corresponding equalities for the impedances are $$\begin{aligned} \label{zad} z_{da}&=\frac{z'_{ab}z'_{ac}}{z'_{ab}+z'_{bc}+z'_{ac}}, \\ z_{db}&=\frac{z'_{ab}z'_{bc}}{z'_{ab}+z'_{bc}+z'_{ac}}, \\ z_{dc}&=\frac{z'_{bc}z'_{ac}}{z'_{ab}+z'_{bc}+z'_{ac}}.\end{aligned}$$ at (-0.4,-3.3) (1) [$a$]{} ; at (2,0) (2)[$b$]{}; at (4.4,-3.3) (3) [$c$]{}; \(1) edge node \[bend right\] [$z'_{ab}$]{} (2) (2) edge node \[bend right\] [$z'_{bc}$]{} (3) (3) edge node \[bend right\] [$z'_{ac}$]{} (1); at (6.6,-3.3) (1) [$a$]{} ; at (9,0) (2)[$b$]{}; at (11.4,-3.3) (3) [$c$]{}; at (9,-2) (4)[$d$]{}; \(4) edge node \[bend right\] [$z_{da}$]{} (1) (4) edge node \[bend right\] [$z_{db}$]{} (2) (3) edge node \[bend right\] [$z_{dc}$]{} (4); To prove the theorem it is enough to express $\rho_{da}$, $\rho_{db}$ and $\rho_{dc}$ from , assuming $\rho_{ab}=0$, $\rho_{bc}=0$, and $\rho_{ac}=0$. Summing up the inverses of all three equations one obtains $$\frac{1}{\rho'_{ab}}+\frac{1}{\rho'_{bc}}+\frac{1}{\rho'_{ac}}=\frac{(\rho_{ab}+\rho_{bc}+\rho_{ac})^2}{\rho_{da}\rho_{db}\rho_{dc}}$$ Since both sides are strictly positive, the last equation is equivalent to $$\label{eq1} \frac{\rho'_{ab}\rho'_{bc}\rho'_{ac}}{\rho'_{ab}\rho'_{bc}+\rho'_{bc}\rho'_{ac}+\rho'_{ab}\rho'_{ac}}=\frac{\rho_{da}\rho_{db}\rho_{dc}}{(\rho_{ab}+\rho_{bc}+\rho_{ac})^2}$$ Multiplying the both sides of the last equation by $$\frac{1}{\rho'_{ab}\rho'_{ac}}=\frac{(\rho_{ab}+\rho_{bc}+\rho_{ac})^2}{\rho_{da}^2\rho_{db}\rho_{dc}},$$ which follows from , we get $$\frac{\rho'_{bc}}{\rho'_{ab}\rho'_{bc}+\rho'_{bc}\rho'_{ac}+\rho'_{ab}\rho'_{ac}}=\frac{1}{\rho_{da}}.$$ Then the equation for $\rho_{da}$ follows. To obtain the equations for $\rho_{db}$ and $\rho_{dc}$ one should multiply by $\frac{1}{\rho'_{ab}\rho'_{bc}}$ and $\frac{1}{\rho'_{ac}\rho'_{bc}}$ respectively.\ The fact that effective impedance and effective admittance do not change follows from Theorem \[YDLaw\]. All described in this section transforms preserve the positivity of admittances and impedances on the edges. Effective impedance of infinite networks over an ordered field ============================================================== Infinite networks with zero potential on infinity ------------------------------------------------- Let $\Gamma=(V,\rho,a_0,B)$ be an infinite network over an ordered field $(K,\succ)$. Let us consider the sequence of finite graphs $(V_n, \left.\rho\right\|_{V_n})$, where $V_n=\{x\in V\mid\operatorname{dist}(a_0,x)\le n\}$, $n\in\Bbb N$. We denote by $$\partial V_n=\{x\in V\mid\operatorname{dist}(a_0,x)= n\}$$ the boundary of the graph $(V_n,\left.\rho\right\|_{V_n})$. Note that $V_{n+1}= \partial V_{n+1}\cup V_n$. Let us denote $B_n=B\cap V_n$. Then $$\Gamma_n=(V_n,\left.\rho\right\|_{V_n},a_0,B_n\cup \partial V_n),n\in\Bbb N$$ is a sequence of finite networks exhausted the infinite network $\Gamma$. This is an analogue to the approach to infinite networks in [@Grimmett]. Let us consider the Dirichlet problem on each $\Gamma_n$: $$\label{DirprFin} \begin{cases} \sum_{y: y\sim x}(v^{(n)}(y)-v^{(n)}(x))\rho_{xy}=0 \mbox { on } V_n\setminus (\partial V_n\cup B_n\cup\{a_0\}), \\ v^{(n)}(x)=0\mbox { on } \partial V_n\cup B_n. \\ v^{(n)}(a_0)=1, \end{cases}$$ \[Thmmonotonicity\] $$\label{monotonicity} \mathcal P_{eff}(\Gamma_{n+1})\preceq \mathcal P_{eff}(\Gamma_n).$$ By Dirichlet/Thomson’s principle [@Muranova] we have $$\label{DTfineq} \mathcal P_{eff}(\Gamma_{n+1})\preceq \frac{1}{2}\sum_{x,y\in V_{n+1}}(\nabla_{xy} f)^2\rho_{xy}$$ for any $f:V_{n+1}\rightarrow K $ such that $f(a_0)=1, \left.f\right\|_{ \partial V_{n+1}\cup B_{n+1}}\equiv 0$.\ Since $(V_{n+1}\setminus \partial V_{n+1})=V_n$ and $B_{n+1}\cap V_n=B_n$, the inequality is true for $$f(x)= \begin{cases} v^{(n)}(x), \mbox{ if } x\in V_n,\\ 0, \mbox{ if } x\in \partial V_{n+1},\\ \end{cases}$$ where $v^{(n)}$ is the solution of for $\Gamma_n$. Then $$\begin{aligned} \frac{1}{2}\sum_{x,y\in V_{n+1}}(\nabla_{xy} f)^2\rho_{xy}=&\frac{1}{2}\sum_{x,y\in V_{n}}(\nabla_{xy} f)^2\rho_{xy}+\frac{1}{2}\sum_{x,y\in \partial V_{n+1}}(\nabla_{xy} f)^2\rho_{xy}\\ =&\mathcal P_{eff}(\Gamma_n)+0.\end{aligned}$$ The last equality, together with , gives us $\eqref{monotonicity}$. Even in a Cauchy complete non-Archimedean ordered field inequalities for all $n\in\Bbb N$ do not imply, that the sequence $\{\mathcal P_{eff}(\Gamma_n)\}_{n=1}^\infty$ converges. Obviously, if the sequence of effective admittances of finite networks converges, then the corresponding sequence of the effective impedances also has a limit (finite or infinite). If for given infinite network $\Gamma$ the limit of effective admittances (impedances) of exhausted finite networks exists in $K$, we call it effective admittance (impedance) of the network $\Gamma$ with zero potential at infinity and denote it by $\mathcal P_{eff}(\Gamma)$ ($Z_{eff}(\Gamma)$). Examples: ladder networks over Levi-Civita field ------------------------------------------------ In this subsection we will investigate the behavior of the sequence $\{\mathcal P_{eff}(\Gamma^{\alpha\beta}_n)\}_{n=1}^\infty$ of effective admittances of finite networks exhausted the ladder network ($\alpha\beta$-network) at the Figure \[fig\] ($\alpha,\beta \in K$, $\alpha,\beta\succ 0$). More precisely, $\alpha\beta$-network is a network $\Gamma^{\alpha\beta}=\{V,\rho, a_0,B\}$, where - $V=\{a_0,a_1,a_2,\dots,x_1,x_2,\dots\}$, - $\rho_{a_0x_1}=\alpha$, $\rho_{x_ix_{i+1}}=\alpha, \rho_{a_ix_i}=\beta, i\in \Bbb N$, and $\rho_{xy}=0,\mbox {otherwise}$, - $B=\{a_0,a_1,a_2,\dots\}$. This network is similar to Feynman’s ladder network and $CL$-network (see [@Feynman], [@Yoon]), but has zero potential at infinity. Therefore, for any ordered field $K$ the Theorem \[Thmmonotonicity\] is true for this network. We will show (Theorem \[ThmLC\] and Example \[ExCL\]) that whether $\{\mathcal P_{eff}(\Gamma_n)\}_{n=1}^\infty$ converges in Cauchy completness of $K$ depends on $\alpha$ and $\beta$. \(0) \[draw=none\]; (1) \[right of=0\] [$a_1$]{}; (2) \[right of=1\] [$a_2$]{}; (3) \[right of=2\] [$a_k$]{}; (4) \[right of=3\] [$a_{n}$]{}; \(5) \[above of=0\][$a_0$]{}; (6) \[right of=5\] [$x_1$]{}; (7) \[right of=6\] [$x_2$]{}; (8) \[right of=7\] [$x_k$]{}; (9) \[right of=8\] [$x_{n}$]{}; (10) \[right of=9\] \[draw=none\]; \(5) edge node \[bend right\] [$\alpha$]{} (6) (6) edge node \[bend right\] [$\alpha$]{} (7) (9) edge node \[bend right\] [$\alpha$]{} (10); (7) to (8); (8) to (9); \(1) edge node \[bend right\] [$\beta$]{} (6) (2) edge node \[bend right\] [$\beta$]{} (7) (3) edge node \[bend right\] [$\beta$]{} (8) (4) edge node \[bend right\] [$\beta$]{} (9); ### Finite ladder network over ordered field Let $\Gamma_{n}$ be a sequence of finite networks exhausted an $\alpha\beta$-network (see Figure \[figFin\]). \(0) \[draw=none\]; (1) \[right of=0\] [$a_1$]{}; (2) \[right of=1\] [$a_2$]{}; (3) \[right of=2\] [$a_k$]{}; (4) \[right of=3\] [$a_{n-1}$]{}; \(5) \[above of=0\][$a_0$]{}; (6) \[right of=5\] [$x_1$]{}; (7) \[right of=6\] [$x_2$]{}; (8) \[right of=7\] [$x_k$]{}; (9) \[right of=8\] [$x_{n-1}$]{}; (10) \[right of=9\] [$x_{n}$]{}; \(5) edge node \[bend right\] [$\alpha$]{} (6) (6) edge node \[bend right\] [$\alpha$]{} (7) (9) edge node \[bend right\] [$\alpha$]{} (10); (7) to (8); (8) to (9); \(1) edge node \[bend right\] [$\beta$]{} (6) (2) edge node \[bend right\] [$\beta$]{} (7) (3) edge node \[bend right\] [$\beta$]{} (8) (4) edge node \[bend right\] [$\beta$]{} (9); The Dirichlet problem for this network is the following $$\label{dirprF} \begin{cases} \alpha v(a_0)+\beta v(a_1)+\alpha v(x_2)-(2\alpha+\beta)v(x_1)=0, \\ \alpha v(x_{i-1})+\beta v(a_i)+\alpha v(x_{i+1})-(2\alpha+\beta)v(x_i)=0 \mbox { for } i=\overline{2,n-1}, \\ v(x_n)=0, \\ v(a_i)= 0 \mbox { for } i=\overline{1,n-1}, \\ v(a_0)=1, \end{cases}$$ Using the second line in we obtain the following recurrence relation for $v(x_i), i=\overline{2,n-1}$ $$\label{recuk} v(x_{i+1})-\left(2+\frac{\beta}{\alpha}\right)v(x_i)+v(x_{i-1})=0$$ since $v(a_i)=0$ for $i=\overline{2,n-1}$. The characteristic polynomial of is $$\label{psi} \psi^2-\left(2+\frac{\beta}{\alpha}\right)\psi+1=0.$$ Its roots are $$\psi_{1,2}=1+\frac{\beta}{2\alpha}\pm \xi,$$ where $$\label{xi} \xi =\sqrt {\frac{\beta}{\alpha}+\left(\frac{\beta}{2\alpha}\right)^2}.$$ Note that $\xi$ should not necessary belong to $K$. It is known, that any ordered field posses a real-closed (or maximal ordered) extension $\overline {K}$. Then in $\overline {K}$ exists exactly one positive square root of ${\frac{\beta}{\alpha}+\left(\frac{\beta}{2\alpha}\right)^2}$ ([@Bourbaki]). Therefore, we fix the extension $\overline {K}$, denote the positive square root by $\xi$, and make all the further calculations in $\overline {K}$. The solution of the recurrence equation is $$\label{ukeq} v(x_i)=c_1\psi_1^i+c_2\psi_2^i,$$ where $c_1,c_2\in \overline{K}$ are arbitrary constants. We use first and third equations in as boundary conditions for this recurrence equation. Substituting in the boundary conditions we obtain the following equations for the constants: $$\label{const} \begin{cases} 1+c_1\psi_1^2+c_2\psi_2^2-\left(2+\frac{\beta}{\alpha}\right)(c_1\psi_1+c_2\psi_2)=0,\\ c_1\psi_1^{n}+c_2\psi_2^{n}=0, \end{cases}$$ which, by is equivalent to $$\label{const} \begin{cases} c_1+c_2=1,\\ c_1\psi_1^{n}+c_2\psi_2^{n}=0. \end{cases}$$ Therefore, $$\label{const} \begin{cases} c_1=\frac{1}{1-\psi_1^{2n}},\\ c_2=\frac{1}{1-\psi_2^{2n}}=\frac{-\psi_1^{2n}}{1-\psi_1^{2n}}, \end{cases}$$ since $\psi_1\psi_2=1$ by . Now we can calculate the effective admittance of $\Gamma_n$: $$\label{PeffF} \mathcal P_{eff}\left(\Gamma_n\right)={\alpha\left(1-v(x_1)\right)}=\alpha\left(1-c_1 \psi_1-c_2 \psi_2\right)=\frac{\alpha\left(\left(1+\frac{\beta}{2\alpha}+ \xi\right)^{2n-1}+1\right)\left(\frac{\beta}{2\alpha}+\xi\right)}{\left(1+\frac{\beta}{2\alpha}+ \xi\right)^{2n}-1}.$$ Since $\mathcal P_{eff}\left(\Gamma_n\right)$ is an element of $K$ as the solution of the Dirichlet problem over $K$, it can be written without $\xi$: $$\begin{aligned} \mathcal P_{eff}\left(\Gamma_n\right)=&\alpha\left(1-c_1 \psi_1-c_2 \psi_2\right)=\alpha\left(1-\frac{\psi_1}{1-\psi_1^{2n}}-\frac{\psi_2}{1-\psi_2^{2n}}\right)\\ =&\alpha\left(1-\frac{\psi_1\left(1-\psi_2^{2n}\right)+\psi_2\left(1-\psi_1^{2n}\right)}{\left(1-\psi_1^{2n}\right)\left(1-\psi_2^{2n}\right)}\right)=\alpha\left(1-\frac{\psi_1+\psi_2-\left(\psi_2^{2n-1}+\psi_1^{2n-1}\right)}{2-\left(\psi_1^{2n}+\psi_2^{2n}\right)}\right)\\ =&\alpha\left(1-\frac{2+\frac{\beta}{\alpha}-2\sum_{k=0}^{n-1}{{2n-1}\choose{2k}}\left(1+\frac{\beta}{2\alpha}\right)^{2n-2k-1}\left(\frac{\beta}{\alpha}+\left(\frac{\beta}{2\alpha}\right)^2\right)^{k}}{2-2\sum_{k=0}^{n}{{2n}\choose{2k}}\left(1+\frac{\beta}{2\alpha}\right)^{2n-2k}\left(\frac{\beta}{\alpha}+\left(\frac{\beta}{2\alpha}\right)^2\right)^{k}}\right),\end{aligned}$$ where in the last line we have used binomial expansion. ### Infinite ladder networks over Levi-Civita field $\mathcal R$ We will consider two examples of $\alpha\beta$-network over Levi-Civita field $\mathcal R$. Firstly, let us describe the Levi-Civita field $\mathcal R$. We take the definition of $\mathcal R$ and theorems about its properties from [@Berz] and [@Shams]. \[LCfield\] A subset $M$ of the rational numbers $\Bbb Q$ is called left-finite if for every $r\in \Bbb Q$ there are only finitely elements of $M$ that are smaller than $r$. Then the Levi-Civita field is the set of all real valued functions on $\Bbb Q$ with left-finite support with the following operations: - addition is defined component-wise $$(\alpha+\beta)(q)=\alpha(q)+\beta(q),$$ - multiplication is defined as follows $$(\alpha\cdot \beta)(q)=\sum_{\substack{q_\alpha,q_\beta\in \Bbb Q,\\ q_\alpha+q_\beta=q}} \alpha(q_\alpha)\cdot \beta(q_\beta).$$ It is proved in [@Berz], that $\mathcal R$ is an ordered field with a set of positive elements $$\mathcal R^{+}=\{\alpha\in \mathcal R\|\alpha(\min \{q\in \Bbb Q\mid \alpha(q)\ne 0\})>0\}.$$ We denote by $\tau$ the following element in $\mathcal R$: $$\tau(q)= \begin{cases} 1, \mbox{ if } q=1,\\ 0, \mbox { otherwise}, \end{cases}$$ which plays role of infinitesimal in Levi-Civita field. Therefore, the Levi-Civita field is non-Archimedean. By [@Berz] we can write any $\alpha\in \mathcal R$ as $$\label{LCseries} \alpha=\sum_{i=1}^\infty \alpha(q_i)\tau^{q_i},$$ since $\alpha_n=\sum_{i=1}^n\alpha (q_i)\tau^{q_i}$ converges strongly to the limit $\alpha$ in the order topology. The set of all polynomials over real numbers $\Bbb R[\tau]=\{a_n \tau^n+\dots a_1 \tau+a_0\mid a_i\in \Bbb R, n\in \Bbb N\}$ is a subring of Levi-Civita field $\mathcal R$ due to . Therefore, since $\mathcal R$ is a field, the field of rational functions over real numbers $$\Bbb R(\tau)=\left\{\frac{\sum_{i=k}^n a_i \tau^i}{\sum_{i=l}^m b_i \tau^i}=\frac{a_k\tau^k+a_{k+1}\tau^{k+1}+\cdots+a_n\tau^n}{b_l\tau^l+b_{l+1}\tau^{l+1}+\cdots+b_m\tau^m}\mid a_i,b_i\in\Bbb R, n,m,k,l\in\Bbb N_0\right\}$$ is isomorphic to a subfield of $\mathcal R$. Let us find the element in $\mathcal R$ which corresponds to the rational function $\frac{1}{\tau^2-4\tau+3}$, i. e. we should find the sequences $\{q_i\}\in \Bbb Q$ and $\{\alpha(q_i)\}\in\Bbb R$ such that $$\left(3-4\tau +\tau^2\right)\left(\sum_{q_i}\alpha(q_i)\tau^{q_i} \right)=1.$$ Comparing the coefficients at powers of $\tau$ at right hand side and left hand side, starting from the lowest power, one obtains $$q_1=0, \alpha(q_1)=\frac{1}{3},$$ $$q_2=1, \alpha(q_2)=\frac{4}{9},$$ and the recurrence relation $$q_i=q_{i-1}+1, 3 \alpha(q_i)-4 \alpha(q_{i-1})+\alpha (q_{i-2})=0 \mbox{ for } i>2.$$ Therefore, solving the recurrence relation for $\alpha(q_i)$ we obtain $$\alpha(q_i)=-\frac{1}{2\cdot 3^i}+\frac{1}{2}$$ and $$\frac{1}{3-4\tau +\tau^2}=\sum_{i\in \Bbb N}\left(-\frac{1}{2\cdot 3^i}+\frac{1}{2}\right)\tau^{i-1}.$$ Note, that the corresponding order in the field $\Bbb R(\tau)$ is the following: $$\label{OrdR(tau)} \frac{a_k\tau^k+a_{k+1}\tau^{k+1}+\cdots+a_n\tau^n}{b_l\tau^l+b_{l+1}\tau^{l+1}+\cdots+b_m\tau^m}\succ 0\mbox{ if } \frac{a_k}{b_l}>0.$$ Therefore, we can consider Levi-Civita field $\mathcal R$ as an ordered extension of the ordered field $\Bbb R(\tau)$ with the positiveness defined as . To consider $\mathcal R$ as an ordered extension of the ordered field $\Bbb R(\tau)$ with the positiveness defined in [@Muranova] we make a substitution $$\tau=\frac{1}{\lambda}.$$ Consequently, we can consider electrical networks over field of rational numbers [@Muranova], as networks over Levi-Civita field and investigate the behavior of the sequence of effective admittances of finite electrical networks. By [@Berz] the Levi-Civita field is Cauchy complete in order topology and real-closed. From the physical point of view we have the following impedances of passive elements - $(Li\omega)^{-1}=L^{-1}\tau$, $L>0$ for the coil, - $(Ci\omega)=C\tau^{-1}$, $C>0$ for the capacitor, - $R>0$ for the resistor, where $\omega$ is a frequency of the alternating current (see [@Muranova]). Let us consider the Feynman’s infinite ladder $LC$-network, assuming that it has zero potential at infinity. It is an $\alpha\beta$-network with $\alpha=L^{-1}\tau$, $\beta={C}\tau^{-1}$, where $L, C>0$, $\alpha,\beta\in \mathcal R$. \[ThmLC\] For the Feynman’s ladder $LC$-network ($\alpha=L^{-1}\tau$, $\beta={C}\tau^{-1}$, where $L, C>0$) with zero potential at infinity $$\label{ThmF} \mathcal P_{eff}\left(\Gamma_n\right)\rightarrow \frac{\beta}{\frac{\beta}{2\alpha}+\xi} \mbox{ as } n\rightarrow\infty$$ in the order topology of Levi-Civita field $\mathcal R$, where $\Gamma_n$ is the sequence of the exhausted finite networks For the Feynman’s ladder $LC$-network $$\frac{\beta}{\frac{\beta}{2\alpha}+\xi}=\left(\frac{C}{2\tau}-\frac{\tau}{L}\sqrt{\frac{CL}{\tau^{2}}+\left(\frac{CL}{2 \tau^{2}}\right)^2}\right)$$ and the motivation for this quantity was Feynman’s impedance for infinite ladder $LC$-network (see [@Feynman 22-13]). Firstly, we should write $\xi$ as an element of Levi-Civita field, i. e. as power series . $$\begin{aligned} \xi=&\sqrt{\frac{CL}{\tau^{2}}+\left(\frac{CL}{2 \tau^{2}}\right)^2}=\frac{CL}{2 \tau^{2}}\sqrt{\frac{4 \tau^2}{CL}+1}=\frac{CL}{2 \tau^{2}}\sum_{k=0}^{\infty}\binom{\frac{1}{2}}{k}\left(\frac{4 \tau^{2}}{CL}\right)^k\\ =&\frac{CL}{2\tau^2}\cdot 1+\frac{CL}{2\tau^2}\cdot \frac{1}{2}\cdot\frac{4 \tau^2}{CL}- \frac{1}{8}\cdot\frac{CL}{2\tau^2}\cdot \left(\frac{4 \tau^2}{CL}\right)^2+o\left(\tau^2\right)\\ =&\frac{CL}{2}\tau^{-2}+1-\tau^2+o\left(\tau^2\right).\end{aligned}$$ Note that here and further $o\left(\tau^m\right)$, where $m\in \Bbb Z$ means $\sum_{k=m+1}^{\infty}a_k \tau^k, a_k\in \Bbb R$. Let us calculate the difference $\mathcal P_{eff}\left(\Gamma_n\right)-\frac{\beta}{\frac{\beta}{2\alpha}+ \xi}$. $$\begin{aligned} \begin{split} &\mathcal P_{eff}\left(\Gamma_n\right)-\frac{\beta}{\frac{\beta}{2\alpha}+ \xi}=\frac{\alpha\left(\frac{\beta}{2\alpha}+ \xi\right)\left(\left(1+\frac{\beta}{2\alpha}+\xi\right)^{2n-1}+1\right)}{\left(1+\frac{\beta}{2\alpha}+\xi\right)^{2n}-1}-\frac{\beta}{\frac{\beta}{2\alpha}+ \xi}\\ =&\frac{\alpha\left(\frac{\beta}{2\alpha}+ \xi\right)^2\left(\left(1+\frac{\beta}{2\alpha}+\xi\right)^{2n-1}+1\right)-\beta\left(\left(1+\frac{\beta}{2\alpha}+\xi\right)^{2n}-1\right)}{\left(\left(1+\frac{\beta}{2\alpha}+\xi\right)^{2n}-1\right)\left(\frac{\beta}{2\alpha}+ \xi\right)} \end{split}\end{aligned}$$ The nominator $A$ of the last expression is $$\begin{aligned} \begin{split} A=&\alpha\left(\left(1+\frac{\beta}{2\alpha}+ \xi\right)^{2n-1}+1\right)\left(\frac{\beta}{2\alpha}+\xi\right)^2-\beta\left(\left(1+\frac{\beta}{2\alpha}+ \xi\right)^{2n}-1\right) \\ =&\alpha\left(D+1\right)\left(\frac{\beta}{2\alpha}+\xi\right)^2-\beta\left(\left(1+\frac{\beta}{2\alpha}+ \xi\right)D-1\right), \end{split}\end{aligned}$$ where $D=\left(1+\frac{\beta}{2\alpha}+ \xi\right)^{2n-1}$. Since $$\left(\frac{\beta}{2\alpha}+\xi\right)^2=\frac{\beta^2}{4\alpha^2}+\frac{\beta}{\alpha}\xi+\xi^2=\frac{\beta^2}{2\alpha^2}+\frac{\beta}{\alpha}+\frac{\beta}{\alpha}\xi=\frac{\beta}{\alpha}\left(\frac{\beta}{2\alpha}+1+\xi\right)$$ and $$\xi^2=\frac{\beta}{\alpha}+\frac{\beta^2}{4\alpha^2},$$ we have $$\begin{aligned} \begin{split} A=&\alpha\left(D+1\right)\frac{\beta}{\alpha}\left(\frac{\beta}{2\alpha}+1+\xi\right)-\beta\left(\left(1+\frac{\beta}{2\alpha}+ \xi\right)D-1\right)\\ =&D\cdot 0+2\beta+\frac{\beta^2}{2\alpha}+ \beta\xi=2\alpha\left( \frac{\beta}{\alpha}+\frac{\beta^2}{4\alpha^2}+ \frac{\beta}{2\alpha}\xi\right)\\ =&2\alpha\left( \xi^2+ \frac{\beta}{2\alpha}\xi\right)=2\alpha\xi\left( \xi+ \frac{\beta}{2\alpha}\right). \end{split}\end{aligned}$$ Therefore, $$\label{diff} \mathcal P_{eff}\left(\Gamma_n\right)-\frac{\beta}{\frac{\beta}{2\alpha}+ \xi}=\frac{2\alpha\xi}{\left(1+\frac{\beta}{2\alpha}+ \xi\right)^{2n}-1}.$$ The right hand side of the last expression is, obviously, positive in $(\mathcal R, \succ)$, therefore $$\begin{aligned} \left\|\mathcal P_{eff}\left(\Gamma_n\right)-\frac{\beta}{\frac{\beta}{2\alpha}+ \xi}\right\|=&\frac{2\alpha\xi}{\left(\left(1+\frac{\beta}{2\alpha}+ \xi\right)^{2n}-1\right)}\\ =&\frac{2\tau\xi}{L\left(\left(1+\frac{C L}{2\tau^2}+ \xi\right)^{2n}-1\right)}\\ =&\frac{CL\tau^{-1}+2 \tau -2\tau^3+o\left(\tau^3\right)}{L\left(\left(CL\tau^{-2}+2-\frac{1}{CL}\tau^2+o\left(\tau^2\right)\right)^{2n}-1\right)}\\ =&\frac{CL\tau^{-1}+2 \tau-\frac{2}{CL}\tau^3+o\left(\tau^3\right)}{L\left(CL\tau^{-2}\right)^{2n}+o\left(\tau^{-4n-2}\right)}\\ =&\left(C\tau^{-1}+o\left(\tau^{-1}\right)\right)\left(\frac{1}{\left(CL\right)^{2n}}\tau^{4n}+o\left(\tau^{4n+2}\right)\right)\\ =&\frac{C}{\left(CL\right)^{2n}}\tau^{4n-1}+o\left(\tau^{4n-1}\right)\rightarrow 0,\end{aligned}$$ when $n\rightarrow \infty$. \[absuf\] From the proof one can see that is true for $\alpha\beta$-network whenever for any $\gamma\in \mathcal R$ exists $N_0\in \Bbb N$ such that $n>N_0$ implies $\left(\frac{\beta}{\alpha}\right)^n\succ \gamma$. \[ExCL\] For the $CL$-network ($\alpha={C}\tau^{-1}$, $\beta=L^{-1}\tau$, $L,C>0$) effective admittances of the exhausted finite networks do not converge in the Levi-Civita field $\mathcal R$. In this case $\xi=\frac{\tau}{\sqrt{CL}}+\left(\frac{\tau}{2\sqrt{CL}}\right)^3+o(\tau^3)$. Let us prove, that $\{\mathcal P (\Gamma_n)\}_{n=1}^\infty$ is not a Cauchy sequence in $\mathcal R$. Indeed $$\begin{aligned} \mathcal P &(\Gamma_{n+1})-\mathcal P (\Gamma_n)\\ =&\frac{\alpha\left(\frac{\beta}{2\alpha}+ \xi\right)\left(\left(1+\frac{\beta}{2\alpha}+\xi\right)^{2n+1}+1\right)}{\left(1+\frac{\beta}{2\alpha}+\xi\right)^{2n+2}-1} -\frac{\alpha\left(\frac{\beta}{2\alpha}+ \xi\right)\left(\left(1+\frac{\beta}{2\alpha}+\xi\right)^{2n-1}+1\right)}{\left(1+\frac{\beta}{2\alpha}+\xi\right)^{2n}-1}\\ =&\alpha\left(\frac{\beta}{2\alpha}+ \xi\right)\left(\frac{\left(1+\frac{\beta}{2\alpha}+\xi\right)^{2n+1}+1}{\left(1+\frac{\beta}{2\alpha}+\xi\right)^{2n+2}-1} -\frac{\left(1+\frac{\beta}{2\alpha}+\xi\right)^{2n-1}+1}{\left(1+\frac{\beta}{2\alpha}+\xi\right)^{2n}-1}\right)\end{aligned}$$ Since $$\psi_1=1+\frac{\beta}{2\alpha}+\xi=1+\frac{\tau}{\sqrt{CL}}+o\left(\tau^{1}\right),$$ we can rewrite $$\begin{aligned} \mathcal P &(\Gamma_{n+1})-\mathcal P (\Gamma_n)=\alpha\left(\frac{\beta}{2\alpha}+ \xi\right)\left(\frac{\psi_1^{2n+1}+1}{\psi_1^{2n+2}-1}-\frac{\psi_1^{2n-1}+1}{\psi_1^{2n}-1}\right)\\ =&\alpha\left(\frac{\beta}{2\alpha}+ \xi\right)\frac{\left(\psi_1^{2n+1}+1\right)\left(\psi_1^{2n}-1\right)-\left(\psi_1^{2n-1}+1\right)\left(\psi_1^{2n+2}-1\right)}{\left(\psi_1^{2n+2}-1\right)\left(\psi_1^{2n}-1\right)}\\ =&C\tau^{-1}\left(\frac{\tau}{\sqrt{CL}}+o\left(\tau^1\right)\right)\frac{\left(\psi_1^{2n+1}+1\right)\left(\psi_1^{2n}-1\right)-\left(\psi_1^{2n-1}+1\right)\left(\psi_1^{2n+2}-1\right)}{\left(\psi_1^{2n+2}-1\right)\left(\psi_1^{2n}-1\right)}\end{aligned}$$ Substituting $$\begin{aligned} &\left(\psi_1^{2n+1}+1\right)\left(\psi_1^{2n}-1\right)-\left(\psi_1^{2n-1}+1\right)\left(\psi_1^{2n+2}-1\right)\\ =&\left(2+\frac{\tau}{\sqrt{CL}}(2n+1)+o\left(\tau^1\right)\right)\left(\frac{\tau}{\sqrt{CL}}(2n)+o\left(\tau^1\right)\right)\\ -&\left(2+\frac{\tau}{\sqrt{CL}}(2n-1)+o\left(\tau^1\right)\right)\left(\frac{\tau}{\sqrt{CL}}(2n+2)+o\left(\tau^1\right)\right)\\ =&-4\frac{\tau}{\sqrt{CL}}+o\left(\tau^1\right)\end{aligned}$$ and $$\begin{aligned} &\left(\psi_1^{2n+2}-1\right)\left(\psi_1^{2n}-1\right)=\left(\frac{\tau}{\sqrt{CL}}(2n+2)+o\left(\tau^1\right)\right)\left(\frac{\tau}{\sqrt{CL}}(2n)+o\left(\tau^1\right)\right)\\ =&\frac{\tau^2}{CL}(4n^2+4n)+o\left(\tau^2\right)\end{aligned}$$ we obtain $$\begin{aligned} \mathcal P &(\Gamma_{n+1})-\mathcal P (\Gamma_n)\\ &=\left(\frac{C}{\sqrt{CL}}+0(\tau^0)\right)\left(-4\frac{\tau}{\sqrt{CL}}+o\left(\tau^1\right)\right)\left(\frac{CL}{4n^2+4n}\tau^{-2}+o\left(\tau^{-2}\right)\right)\\ &=\frac{-4C}{4n^2+n}\tau^{-1}+o\left(\tau^{-1}\right)\succ \tau \mbox{ for any } n\in \Bbb N \mbox { and for any } L,C>0.\end{aligned}$$ Therefore, $\{\mathcal P (\Gamma_n)\}_{n=1}^\infty$ is not a Cauchy sequence in $\mathcal R$. Therefore, the following question: > Under what conditions the effective admittance of infinite network over non-Archimedean field could be defined? remains open. Note, that Remark \[absuf\] gives some sufficient condition for $\alpha\beta$-network. Acknowledgement {#acknowledgement .unnumbered} =============== The author thanks her scientific advisor, Professor Alexander Grigor’yan, for fruitful discussions on the topic. [99]{} Martin T. Barlow. Random Walks and Heat Kernels on Graphs. London Mathematical Society, Lecture Note Series: 438. Cambridge University Press, 2017. Martin Berz, Christian Bischof, George Corliss, Andreas Griewank. Computational Differentiation: Techniques, Applications, and Tools. Chapter 2: Calculus and Numerics on Levi-Civita Fields. eds., SIAM, 1996. N. Bourbaki. Elements of Mathematica. Algebra II, Chapters 4-7. Springer-Verlag Berlin Heidelberg, 2003. P.G. Doyle, J.L. Snell. Random walks and electric networks. Carus Mathematical Monographs 22, Mathematical Association of America. Washington, DC, 1984. Richard P. Feynman, Robert B. Leighton, Matthew Sands. The Feynman lectures on physics, Volume 1: Mainly mechanics, radiation, and heat. Addison-Wesley publishing company. Reading, Massachusetts, Fourth printing – 1966. Richard P. Feynman, Robert B. Leighton, Matthew Sands. The Feynman lectures on physics, Volume 2: Mainly Electromagnetism and Matter. Addison-Wesley publishing company. Reading, Massachusetts, Fourth printing – 1966. A. Grigor’yan. Introduction to Analysis on Graphs. AMS University Lecture Series, Volume: 71. Providence, Rhode Island, 2018. G. Grimmett. Probability on Graphs: Random Processes on Graphs and Lattices. Cambridge University Press. New York, 2010. J. F. Hall, T. D. Todorov. Ordered Fields, the Purge of Infinitesimals from Mathematics and the Rigorousness of Infinitesimal Calculus. [Bulgarian Journal of Physics]{},2015. Vol. 42, n.2, 99–127. David A. Levin, Yuval Peres, Elizabeth L. Wilmer. Markov Chains and Mixing Times. AMS University Lecture Series. Providence, Rhode Island, 2009. Anna Muranova. On the notion of effective impedance. arXiv e-prints, page arXiv:1905.02047, May 2019. K. Shamseddine. New Elements of Analysis on the Levi-Civita Field. PhD thesis, Michigan State University, East Lansing, Michigan, USA, 1999. Paolo M. Soardi. Potential Theory on Infinite Networks. Springer-Verlag, Berlin Heidelberg, 1994. Wolfgang Woess. Random Walks on Infinite Graphs and Groups. Cambridge Tracts in Mathematics: 138. Cambridge University Press, 2000. Sung Hyun Yoon. Ladder-type circuits revisited. [European journal of physics]{}, 2007. Vol. 22, n. 22, 277–288.	{ "pile_set_name": "ArXiv" }
--- bibliography: - 'my.bib' title: A Massively Parallel Associative Memory Based on Sparse Neural Networks ---	{ "pile_set_name": "ArXiv" }
--- abstract: \| We demonstrate a construction of error-correcting codes from graphs by means of $k$-resolving sets, and present a decoding algorithm which makes use of covering designs. Along the way, we determine the $k$-metric dimension of grid graphs (i.e. Cartesian products of paths).\ [MSC 2010:]{} 05C12, 94B25 (primary); 05B40, 94B35 (secondary).\ [Keywords:]{} error-correcting code; $k$-resolving set; $k$-metric dimension; covering design; uncovering; grid graph. author: - 'Robert F. Bailey[^1]' - 'Ismael G. Yero[^2]' title: 'Error-correcting codes from $k$-resolving sets' --- Introduction ============ Error-correcting codes ---------------------- Error-correcting codes are applied to the accurate transmission and storage of data. When information is received by a target, or read from a storage medium, errors may be introduced—for example, due to signal noise, or the medium being damaged—so to alleviate this problem, redundancy is introduced in order that the intended message can still be understood. For an introduction to coding theory, see [@Pless]. Formally, an [error-correcting code]{} (or simply a [code]{}) is a collection $\mathcal{C}$ of vectors, called [codewords]{}, of given length $\ell$ over a fixed alphabet. The [Hamming distance]{} between two codewords ${\mathbf{x}}=(x_1,\ldots,x_\ell)$, ${\mathbf{y}}=(y_1,\ldots,y_\ell)$ is the number of positions where they differ, i.e. $\|\{i\, : \, x_i\neq y_i \}\|$. The [minimum distance]{} of $\mathcal{C}$ is the least Hamming distance between any two distinct codewords; if the minimum distance is $D$, then the [correction capability]{} of $\mathcal{C}$ is $r=\left\lfloor (D-1)/2 \right\rfloor$. Suppose that a codeword ${\mathbf{x}}$ is transmitted via a noisy channel which causes errors to appear, i.e. some symbols are replaced with others. If there are $r$ errors or fewer, the received word has a unique nearest neighbour in $\mathcal{C}$, which is necessarily the transmitted word ${\mathbf{x}}$. For this to be useful in practice, an efficient decoding algorithm is needed to determine the nearest neighbour. Traditionally, the most familiar error-correcting codes are [linear codes]{} (i.e. subspaces of vector spaces over finite fields) [@Pless], where the alphabet size is small (such as binary codes, which have an alphabet of size 2). Other classes of codes include [permutation codes]{} [@Cameron2010], where each codeword is a permutation of $n$ symbols, so the length and alphabet size are both equal to $n$; codes with larger alphabet sizes have been the subject of more recent attention, in part because of applications such as powerline communications [@Chu] and flash memory devices [@TamoSchwartz10]. $k$-resolving sets ------------------ We consider finite, simple, connected, undirected graphs. The [distance]{} between two vertices $u$ and $v$ of a graph $G$ is the length of a shortest path between $u$ and $v$, and we denote this by ${\mathrm{d}}_G(u,v)$. In recent years, much attention has been paid to the [metric dimension]{} of graphs: this is the smallest size of a subset of vertices (called a [resolving set]{}) with the property that the list of distances from any vertex to those in the set uniquely identifies that vertex, and is denoted by $\dim(G)$. These concepts were introduced to graph theory in the 1970s by Harary and Melter [@Harary76] and, independently, Slater [@Slater75]; however, in the context of arbitrary metric spaces, the concept dates back at least as far as the 1950s [@Blumenthal53]. Various applications have been suggested for resolving sets and metric dimension of graphs, including combinatorial optimization [@SeboTannier04], pharmaceutical chemistry [@Chartrand00], robot navigation [@Khuller96] and sonar [@Slater75]. For more information, see [@bsmd; @Chartrand00]. The following definition is a natural generalization of the notion of resolving sets. Let $G=(V,E)$ be a graph. An ordered set of vertices $(v_1,\ldots,v_{\ell})$ is a [$k$-resolving set]{} for $G$ if, for any distinct vertices $u,w\in V$, the lists of distances $({\mathrm{d}}_G(u,v_1),\ldots,{\mathrm{d}}_G(u,v_{\ell}))$ and $({\mathrm{d}}_G(w,v_1),\ldots,{\mathrm{d}}_G(w,v_{\ell}))$ differ in at least $k$ positions. In the case $k=1$, we have the usual notion of a resolving set for $G$. For $k>2$ it is not necessarily the case that an arbitrary graph $G$ has a $k$-resolving set; for example, a complete graph $K_n$ with $n\geq 3$ has a $2$-resolving set but not a $3$-resolving set. If $G$ has a $k$-resolving set, we denote the least size of a $k$-resolving set by $\dim_k(G)$, the [$k$-metric dimension]{} of $G$. A $k$-resolving set of size $\dim_k(G)$ is called a [$k$-metric basis]{} for $G$. If $k$ is the largest integer for which $G$ has a $k$-resolving set, then we say that $G$ is a [$k$-metric dimensional]{} graph. The notion of $k$-resolving sets in graphs was introduced in [@yero1] and further studied in [@yero3; @yero4; @yero2]; it was then extended to more general metric spaces in [@BeardonRV]. Codes from $k$-resolving sets ============================= The two themes of this paper are tied together by the following definition. \[defn:kRS\] Let $G$ be a graph with $n$ vertices and diameter $d$, and let $S=\{v_1,v_2,\ldots,v_{\ell}\}$ be a $k$-resolving set for $G$ of size $\ell$. Then the set $$\mathcal{C}(G,S) = \{ ({\mathrm{d}}_G(u,v_1),{\mathrm{d}}_G(u,v_2),\ldots,{\mathrm{d}}_G(u,v_{\ell})) \, : \, u\in V \}$$ is called a [$(G,k)$-code]{}. It follows from the definition that $\mathcal{C}(G,S)$ is an error-correcting code of length $\ell$, size $n$ and minimum Hamming distance at least $k$, over the alphabet $\{0,\ldots,d\}$, which can correct $r=\lfloor (k-1)/2 \rfloor$ errors. In order for $r$ to be non-zero, we require that $k\geq 3$ (otherwise the code has no practical purpose). For $\mathcal{C}(G,S)$ to be used for error correction, we need a decoding algorithm. Let $G_j(u)$ denote the subset of vertices of $G$ at distance $j$ from $u$. Now suppose that $u\in V$ and ${\mathbf{u}}=({\mathrm{d}}_G(u,v_1),{\mathrm{d}}_G(u,v_2),\ldots,{\mathrm{d}}_G(u,v_{\ell})) \in \mathcal{C}(G,S)$; suppose that we transmit ${\mathbf{u}}$ and receive the word ${\mathbf{x}}=(x_1,x_2,\ldots,x_{\ell})$, which is assumed to have at most $r$ errors. \[lemma:uniqueness\] Let $u$, ${\mathbf{u}}$ and ${\mathbf{x}}$ be as above, and suppose that $I$ is an $(\ell-r)$-subset of $\{1,\ldots,\ell\}$. - If the received word ${\mathbf{x}}$ contains no errors in the positions indexed by $I$, then $$\bigcap_{i\in I} G_{x_i}(v_i) = \{u\}.$$ - If the received word ${\mathbf{x}}$ does contain an error in a position in $I$, then $$\bigcap_{i\in I} G_{x_i}(v_i) = \varnothing.$$ If there are no errors in the positions indexed by $I$, then both ${\mathbf{x}}$ and ${\mathbf{u}}$ contain the same entries in those positions, and thus $u$ is the unique vertex at those distances from the corresponding entries in the $k$-resolving set. If, however, there are errors in those positions, then no such vertex can exist, and thus the intersection is empty. The goal, therefore, when decoding a $(G,k)$-code is to find (as quickly as possible) an $(\ell-r)$-subset of positions for which the intersection $\displaystyle{\bigcap_{i\in I} G_{x_i}(v_i)}$ is non-empty. Successively enumerating the $(\ell-r)$-subsets of $\{1,\ldots,\ell\}$ will achieve this, but will be slow in practice. However, if we assume that there are at most $r'<r$ errors (say $r'=2$ or $r'=3$), we can make use of the following idea. \[defn:uncovering\] Let $\nu$, $\kappa$, $\tau$ be integers such that $\nu\geq \kappa \geq \tau \geq 0$. A [$(\nu,\nu-\kappa,\tau)$-uncovering]{} is a collection $\mathcal{U}$ of $(\nu-\kappa)$-subsets of $\{1,\ldots,\nu\}$ with the property that any $\tau$-subset of $\{1,\ldots,\nu\}$ is disjoint from at least one member of $\mathcal{U}$. If we take the complements of each $(\nu-\kappa)$-subset in $\mathcal{U}$, we obtain a [$(\nu,\kappa,\tau)$-covering design]{}, which are much more widespread in the literature: see the survey by Mills and Mullin [@MillsMullin92] for details of these. Uncoverings were introduced by the first author in [@btubb; @ecpg] where they were applied to decoding permutation codes; the same concept was also devised under the name [antiblocking system]{} by Kroll and Vincenti [@Kroll08] for a decoding algorithm for linear codes. A further application to network reliability was given in [@ubbnet]. The best known bound on the minimum size of coverings (and thus uncoverings also) is known as the [Schönheim bound]{}, proved in [@Schonheim]. It states that for given $\nu$, $\kappa$ and $\tau$, the least size of a $(\nu,\kappa,\tau)$-covering design is $$L(\nu,\kappa,\tau) = \left\lceil \frac{\nu}{\kappa} \left\lceil \frac{\nu-1}{\kappa-1} \left\lceil \cdots \left\lceil \frac{\nu-\tau+1}{\kappa-\tau+1} \right\rceil \cdots \right\rceil \right\rceil \right\rceil.$$ Covering designs meeting this bound are known (or known asymptotically) in many cases: see [@handbook; @MillsMullin92] for tables of results. The database of best-known covering designs [@lajolla] is useful for finding uncoverings with small parameters. \[example:cov-unc\] The following is an $(8,5,2)$-covering design: $$\begin{array}{ccccc} 4 & 5 & 6 & 7 & 8 \\ 1 & 2 & 3 & 7 & 8 \\ 1 & 4 & 5 & 6 & 8 \\ 2 & 3 & 4 & 5 & 6 \end{array}$$ By taking the complements of each block, we obtain an $(8,3,2)$-uncovering: $$\begin{array}{ccc} 1 & 2 & 3 \\ 4 & 5 & 6 \\ 2 & 3 & 7 \\ 1 & 7 & 8 \end{array}$$ It can easily be seen that any pair chosen from $\{1,\ldots,8\}$ is disjoint from at least one row. A decoding algorithm -------------------- Suppose we have a $(G,k)$-code of length $\ell$, and we wish to correct $r'$ errors. If $\mathcal{U}$ is an $(\ell, \ell-r, r')$-uncovering, we proceed as follows: for a received word ${\mathbf{x}}$, we consider each $I\in\mathcal{U}$ and obtain $\displaystyle{\bigcap_{i\in I} G_{x_i}(v_i)}$. By Lemma \[lemma:uniqueness\], this intersection will either be empty or contain the vertex $u$ corresponding to the transmitted word ${\mathbf{u}}$. If $\mathbf{u}$ contains at most $r'$ errors, by the definition of uncovering we know that there exists an $I\in\mathcal{U}$ disjoint from the error positions; consequently, we are guaranteed to be able to find the transmitted word. To compute $\displaystyle{\bigcap_{i\in I} G_{x_i}(v_i)}$, consider the matrix $M$ whose rows are indexed by $V$ and whose columns are indexed by $S$, and where the entries are $M_{uv}={\mathrm{d}}_G(u,v)$ (so the rows of $M$ are precisely the codewords). For a given $I\subseteq S$, examine the rows of the submatrix to find a row which agrees with ${\mathbf{x}}$ in those positions; if such a row exists, by Lemma \[lemma:uniqueness\] it must be unique and correspond to the vertex $u$. Complexity ---------- The matrix $M$ is a submatrix of the distance matrix of $G$, which can be computed in $O(\|V\|^3)$ time (for instance, by the Floyd–Warshall algorithm; see [@intro-alg [§]{}25.2]); however, this need only be done once, prior to the implementation of the code. For a given instance of the decoding problem, where the input is a received word ${\mathbf{x}}$ and the output the transmitted word ${\mathbf{u}}$, the matrix $M$ must be examined at most $\|\mathcal{U}\|$ times, and at most $\|I\|\cdot\|V\|$ steps are required each time. Thus the overall complexity of the decoding algorithm is $O(\|\mathcal{U}\|\cdot\|I\|\cdot\|V\|)$. Some covering designs ===================== Consider a $(\nu,\kappa,\tau)$-covering design where $\nu$ is given by a linear function in $\kappa$. Then for fixed $\tau$ there exists a threshold value $\kappa_0$ such that, beyond this threshold, the Schönheim bound $L(\nu,\kappa,\tau)$ remains constant. For example, if $\nu=2\kappa+3$ we see that for all $\kappa\geq 9$, we have $L(2\kappa+3,\kappa,2)=7$, while for all $\kappa\geq 21$, we have $L(2\kappa+3,\kappa,3)=15$. Unfortunately, examples of covering designs which actually achieve this lower bound are rare. However, while it would be desirable to have optimal coverings, from the perspective of complexity a family of coverings of constant size (for a given value of $\tau$) is an acceptable solution. So the following construction (suggested by F. Petrov[^3]) is very useful. \[prop:petrov\] Let $\nu=a\kappa+b$ (where $a\neq 0$ and $b$ are fixed constants), and let $\tau$ be a fixed constant. Then, provided $\kappa$ is sufficiently large, there is a $(\nu,\kappa,\tau)$-covering design of size bounded by a constant dependent only on $a$ and $\tau$. Let $m=\lfloor \kappa/\tau \rfloor$ and $s=\lceil \nu/m \rceil$. Form a partition $\Pi$ of the set of $\nu=a\kappa+b$ points into $s$ subsets, including as many as possible of size $m$ and (unless $\kappa$ divides $\tau$) one of smaller size. By the division algorithm, $\kappa=m\tau+\rho$, where $0 \leq \rho \leq \tau-1$. Then we have $$s= \left\lceil \frac{\nu}{m} \right\rceil = \left\lceil \frac{a\kappa+b}{m} \right\rceil = \left\lceil \frac{a(m\tau+\rho)+b}{m} \right\rceil = a\tau + \left\lceil \frac{a\rho+b}{m} \right\rceil = a\tau + \left\lceil \frac{\tau(a\rho+b)}{\kappa-\rho} \right\rceil .$$ The quantity $\displaystyle \frac{\tau(a\rho+b)}{\kappa-\rho}$ is zero if and only if both $b=0$ and $\rho=0$; furthermore, since $a$, $b$ and $\tau$ are constants and $\rho<\tau$, we have that$$\lim_{k\to\infty} \frac{\tau(a\rho+b)}{\kappa-\rho} = 0.$$ Thus, by the definition of a limit, there exists some $\kappa_1$ such that for all $\kappa\geq\kappa_1$, $\displaystyle \frac{\tau(a\rho+b)}{\kappa-\rho} \leq 1$, and therefore $s\in \{ a\tau, a\tau+1 \}$. We form a covering design as follows: for any combination of $\tau$ of the sets in $\Pi$, take any $\kappa$-subset of points which contains their union. Then this collection of $\kappa$-subsets forms the blocks of an $(a\kappa+b,\kappa,\tau)$-covering design: the number of blocks is at most $\binom{a\tau+1}{\tau}$; this depends only on the constants $a$ and $\tau$. \[example:petrov\] We will use Proposition \[prop:petrov\] to construct a $(23,10,2)$-covering design. The Schönheim bound gives $L(23,10,2)=7$, while the best-known covering has size $8$ (see [@lajolla]); we can obtain a covering of size 10 as follows. First, we partition the set $\{1,\ldots,23\}$ into $\lceil 23/ (10/2) \rceil =5$ subsets, four of which have size $10/2=5$ and the remaining set has size $3$: $$1\ 2\ 3\ 4\ 5 \mid 6\ 7\ 8\ 9\ 10 \mid 11\ 12\ 13\ 14\ 15 \mid 16\ 17\ 18\ 19\ 20 \mid 21\ 22\ 23$$ Then we form the blocks of our covering by taking the unions of any pair of these subsets, adding extra points arbitrarily if required. The blocks are the rows of the array below (where $\ast$ indicates symbols which can be replaced arbitrarily): $$\begin{array}{cccccccccc} 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 \\ 1 & 2 & 3 & 4 & 5 & 11 & 12 & 13 & 14 & 15 \\ 1 & 2 & 3 & 4 & 5 & 16 & 17 & 18 & 19 & 20 \\ 1 & 2 & 3 & 4 & 5 & 21 & 22 & 23 & \ast & \ast \\ 6 & 7 & 8 & 9 & 10 & 11 & 12 & 13 & 14 & 15 \\ 6 & 7 & 8 & 9 & 10 & 16 & 17 & 18 & 19 & 20 \\ 6 & 7 & 8 & 9 & 10 & 21 & 22 & 23 & \ast & \ast \\ 11 & 12 & 13 & 14 & 15 & 16 & 17 & 18 & 19 & 20 \\ 11 & 12 & 13 & 14 & 15 & 21 & 22 & 23 & \ast & \ast \\ 16 & 17 & 18 & 19 & 20 & 21 & 22 & 23 & \ast & \ast \end{array}$$ In fact, for any sufficiently large value of $\kappa$, Proposition \[prop:petrov\] will yield a $(2\kappa+3,\kappa,2)$-covering with $\binom{5}{2}=10$ blocks. Codes from paths and cycles =========================== In this section, we show how two straightforward classes of graphs—namely paths and cycles—may be used to obtain families of $(G,k)$-codes to which our decoding algorithm can be applied. While the parameters for codes obtained from paths and cycles are very similar, the key distinction is that the alphabet size is smaller for codes from cycles than for codes from paths, on account of the diameter of a cycle $C_n$ being (approximately) half that of a path $P_n$. Paths ----- Let $P_n$ denote a path on $n$ vertices, which has a $k$-resolving set for $k\leq n-1$, and for $k\geq 3$ has $\dim_k(P_n)=k+1$; see [@yero1] for details. (We remark that $\dim_k(P_n)=k$ for $k=1$ and $k=2$, but this is of no interest from the perspective of error-correction.) A $(P_n,k)$-code will therefore have $n$ codewords over an alphabet of size $\operatorname{diam}(P_n)=n-1$, of length $\ell=\dim_k(P_n)=k+1$, and with minimum distance at least $k$, so can correct $r=\lfloor (k-1)/2 \rfloor$ errors. We note that in the extreme case where $k=n-1$, the $k$-metric dimension is $k+1=n$, and thus every vertex is required in a $k$-resolving set. This also means that the alphabet size, length of the codewords and number of codewords are all equal, making these codes comparable to Latin squares as permutation codes (but with minimum distance one less). Consider a path $P_5$ on $5$ vertices, and let $k=4$. All $5$ vertices are needed in a $4$-resolving set, and the code obtained is as follows: $$\begin{array}{c\|ccccc} \textnormal{Vertex} & \multicolumn{5}{\|c}{\textnormal{Codeword}} \\ \hline 1 & 0 & 1 & 2 & 3 & 4 \\ 2 & 1 & 0 & 1 & 2 & 3 \\ 3 & 2 & 1 & 0 & 1 & 2 \\ 4 & 3 & 2 & 1 & 0 & 1 \\ 5 & 4 & 3 & 2 & 1 & 0 \end{array}$$ This code has $5$ codewords over the alphabet $\{0,1,2,3,4\}$, and has minimum distance $4$, so can correct $\lfloor (4-1)/2 \rfloor =1$ error. In order to decode a $(P_n,k)$-code, we will require an uncovering with parameters $(k+1, \ell-\lfloor (k-1)/2 \rfloor, r')$, or equivalently a $(k+1, \lfloor (k-1)/2 \rfloor, r')$-covering design. By letting $m=\lfloor (k-1)/2 \rfloor$, this is either a $(2m+2,m,r')$-covering when $k$ is odd, or a $(2m+3,m,r')$-covering when $k$ is even. In either case, for a given value of $r'$, we can obtain appropriate covering designs from Proposition \[prop:petrov\]. For example, for any path $P_n$ with $n\geq 23$ may use the uncovering arising from Example \[example:petrov\] to correct two errors. Cycles ------ For a cycle $C_n$ on $n$ vertices, we must consider the cases where $n$ is odd or even separately; details of $k$-resolvability of cycles were given in [@BeardonRV]. ### Odd cycles When $n$ is odd, $C_n$ has a $k$-resolving set for $k\leq n-1$ with $\dim_k(C_n)=k+1$ for all $k$. A $(C_n,k)$-code will therefore have $n$ codewords of length $\ell=\dim_k(C_n)=k+1$, and with minimum distance $k$, so can correct $r=\lfloor (k-1)/2 \rfloor$ errors; all of these parameters are identical to those for a path $P_n$, but this time the codewords are over an alphabet of size $\operatorname{diam}(C_n)=(n-1)/2$. As the alphabet size does not affect the uncovering needed for decoding, any uncovering for a $(P_n,k)$-code may also be used for a $(C_n,k)$-code when $n$ is odd. Consider a cycle $C_5$ on $5$ vertices (labelled $0,\ldots,4$), and let $k=4$. All $5$ vertices are needed in a $4$-resolving set, and the code obtained is as follows: $$\begin{array}{c\|ccccc} \textnormal{Vertex} & \multicolumn{5}{\|c}{\textnormal{Codeword}} \\ \hline 0 & 0 & 1 & 2 & 2 & 1 \\ 1 & 1 & 0 & 1 & 2 & 2 \\ 2 & 2 & 1 & 0 & 1 & 2 \\ 3 & 2 & 2 & 1 & 0 & 1 \\ 4 & 1 & 2 & 2 & 1 & 0 \end{array}$$ This has $5$ codewords over the alphabet $\{0,1,2\}$, and has minimum distance $4$, so can therefore correct $\lfloor (4-1)/2 \rfloor =1$ error. We note that, because the underlying graph is a cycle, any codeword may be obtained from another by a cyclic permutation. ### Even cycles When $n$ is even, a little more care is required. Letting $n=2q$, we have that $C_n$ has a $k$-resolving set for $k\leq n-2=2q-2$, and $\dim_k(C_n)=k+1$ for $k\leq q-1$, or $\dim_k(C_n)=k+2$ for $q\leq k \leq 2q-2=n-2$. Thus for $k\leq q-1$, the parameters of a $(C_n,k)$-code when $n$ is even (as well as those of the uncovering needed for decoding) are the same as those from an odd cycle (although with an alphabet of size $q=n/2$). For $q\leq k \leq 2q-2$, they are slightly different: we have $n$ codewords of length $k+2$ and minimum distance $k$. For decoding, we will need a $(k+2, k+2-\lfloor (k-1)/2 \rfloor, r')$-uncovering, or equivalently a $(k+2,\lfloor (k-1)/2 \rfloor, r')$-covering design. Letting $m=\lfloor (k-1)/2 \rfloor$, we require either a $(2m+3,m,r')$-covering or a $(2m+4,m,r')$-covering if $k$ is odd or even respectively; Proposition \[prop:petrov\] is applicable in either case. Consider a cycle $C_6$ on $6$ vertices (labelled $0,\ldots,5$), and let $k=4$. All $6$ vertices are needed in a $4$-resolving set, and the code obtained is as follows: $$\begin{array}{c\|cccccc} \textnormal{Vertex} & \multicolumn{6}{\|c}{\textnormal{Codeword}} \\ \hline 0 & 0 & 1 & 2 & 3 & 2 & 1 \\ 1 & 1 & 0 & 1 & 2 & 3 & 2 \\ 2 & 2 & 1 & 0 & 1 & 2 & 3 \\ 3 & 3 & 2 & 1 & 0 & 1 & 2 \\ 4 & 2 & 3 & 2 & 1 & 0 & 1 \\ 5 & 1 & 2 & 3 & 2 & 1 & 0 \end{array}$$ This code has $6$ codewords over the alphabet $\{0,1,2,3\}$, and has minimum distance $4$, so can correct $\lfloor (4-1)/2 \rfloor =1$ error. Codes from grid graphs ====================== In this section, we will use the family of grid graphs $P_s{\,\Box\,}P_t$, i.e. the Cartesian product of the paths $P_s$ and $P_t$, to obtain $(P_s{\,\Box\,}P_t,k)$-codes. As we will show, for any grid graph $P_s{\,\Box\,}P_t$ of order $st$ and any $k\in \{1,\ldots,s+t-2\}$, we have that $\dim_k(P_s{\,\Box\,}P_t)=2k$. This goes on to provide an interesting infinite family of examples. The $k$-metric dimension of grid graphs --------------------------------------- Suppose that $G$ is the grid graph $P_s{\,\Box\,}P_t$, that $s\ge t$, and $U=\{u_1,u_2,\ldots,u_s\}$ and $V=\{v_1,v_2,\ldots,v_t\}$ are the vertex sets of $P_s$ and $P_t$, respectively. It is clear that, if $G$ is a $k$-metric dimensional graph, then for every positive integer $k'\le k$, $G$ also has a $k'$-metric basis. Next we present a characterization of $k$-metric dimensional graphs, obtained in [@yero1], which will be useful in our work. To do so, we need some additional terminology. Given two vertices $x,y\in V(G)$, we say that the set of [distinctive vertices]{} of $x,y$ is $${\cal D}(x,y)=\{z\in V(G): d_{G}(x,z)\ne d_{G}(y,z)\}.$$ \[theokmetric\] A connected graph $G$ is $k$-metric dimensional if and only if $k=\displaystyle\min_{x,y\in V(G)}\vert {\cal D}(x,y)\vert .$ To compute the $k$-metric dimension of a grid graph, we need first to determine for which values of $k$ there exists a $k$-metric basis. This is answered by our next result. For any $s,t\ge 2$, the graph $G=P_s{\,\Box\,}P_t$ is $(s+t-2)$-metric dimensional. First, we consider the vertices $(u_2,v_1)$ and $(u_1,v_2)$. Notice that $${\cal D}_G((u_2,v_1),(u_1,v_2))= (U\times \{v_1\})\cup (\{u_1\}\times V)-\{(u_1,v_1)\}.$$ Thus, $\|{\cal D}_G((u_2,v_1),(u_1,v_2))\|=s+t-2$ and $G$ is $k$-metric dimensional for some $k\le s+t-2$. On the other hand, let $(u_i,v_j)$ and $(u_g,v_h)$ be two distinct vertices of $G$. We consider the following cases.\ Case 1: $i=g$. Hence $j\ne h$ and it follows that $U\times\{v_j,v_h\}\subseteq {\cal D}_G((u_i,v_j),(u_g,v_h))$. Also, $$\|(\{u_i\}\times V)\cap {\cal D}_G((u_i,v_j),(u_g,v_h))\|\ge t-1.$$ Thus we have $$\begin{aligned} \|{\cal D}_G((u_i,v_j),(u_g,v_h))\|&\ge \|U\times\{v_j,v_h\}\|+\|(\{u_i\}\times V)\cap {\cal D}_G((u_i,v_j),(u_g,v_h))\|-2\\ &\ge 2s+t-3\\ &\ge s+t-1.\end{aligned}$$ Case 2: $j=h$. Analogous to Case 1 above, we obtain that $\|{\cal D}_G((u_i,v_j),(u_g,v_h))\|\ge s+t-1.$\ Case 3: $i\ne g$ and $j\ne h$. We may assume that $i<g$. Hence we have one of the following situations. - If $j<h$, then we notice that at most two vertices of the set $(\{u_1,\ldots,u_{g}\}\times \{v_j\})\cup (\{u_{i},\ldots,u_s\}\times \{v_h\})\cup (\{u_i\}\times \{v_1,\ldots,v_{h}\})\cup (\{u_g\}\times \{v_{j},\ldots,v_t\})$ do not belong to the set ${\cal D}_G((u_i,v_j),(u_g,v_h))$ (see Figure \[fig-Case-1\] for an example with $i=3$, $j=2$, $g=9$ and $h=6$). in [0,2,4,6,8,10,12,14,16,18,20,22]{} [ in [0,2,4,6,8,10,12,14]{} [ at (,) \[draw, shape=circle,scale=.5\] ; ]{} ]{} in [2,4,6,8,10,12,14]{} [ at (16,) \[draw, fill=black, shape=circle,scale=.5\] ; ]{} in [0,2,4,6,8,10]{} [ at (4,) \[draw, fill=black, shape=circle,scale=.5\] ; ]{} in [0,2,4,6,8,10,12,14,16]{} [ at (,2) \[draw, shape=circle,fill=black, scale=.5\] ; ]{} in [4,6,8,10,12,14,16,18,20,22]{} [ at (,10) \[draw, shape=circle,fill=black, scale=.5\] ; ]{} in [1,2,3,4,5,6]{} [ at (+-2,-2) [$u_{\x}$]{}; ]{} in [7,8,9,10,11,12]{} [ at (+-2,-2) [$u_{\x}$]{}; ]{} in [1,2,3,4,5,6,7,8]{} [ at (-2,+-2) [$v_{\y}$]{}; ]{} at (14,2) \[rectangle,fill=gray, scale=.85\] ; at (6,10) \[rectangle, fill=gray,scale=.85\] ; at (4,2) \[rectangle, fill=black,scale=.85\] ; at (16,10) \[rectangle, fill=black,scale=.85\] ; Consequently, $$\begin{aligned} \|{\cal D}_G((u_i,v_j),(u_g,v_h))\|&\ge \|\{u_1,\ldots,u_{g}\}\times \{v_j\}\|+\|\{u_{i},\ldots,u_s\}\times \{v_h\}\|+\\ &\hspace{0.5cm}+\|\{u_i\}\times \{v_1,\ldots,v_{h}\}\|+\|\{u_g\}\times \{v_{j},\ldots,v_t\}\|-6\\ & \ge g+s-i+1+h+t-j+1-6\\ &= s+t+g+h-i-j-4\\ &\ge s+t+i+j-i-j-2 \quad \mbox{(since $i<g$ and $j<h$)} \\ &=s+t-2.\end{aligned}$$ - If $j>h$, then a similar procedure yields $\|{\cal D}_G((u_i,v_j),(u_g,v_h))\|\ge s+t-2$. As a consequence, we obtain that $G$ is $(s+t-2)$-metric dimensional. Having established that $G=P_s{\,\Box\,}P_t$ is $(s+t-2)$-metric dimensional, next we obtain its $k$-metric dimension for every $k\in \{1,\ldots,s+t-2\}$. As an observation regarding the set of distinctive vertices of a pair of vertices $x,y$, we notice that if $S$ is a $k$-resolving set for a graph $G$, then $\|{\cal D}(x,y)\cap S\|\ge k$. This simple fact will be used frequently in our next proof. For any grid graph $G=P_s{\,\Box\,}P_t$ and every $k\in \{1,\ldots,s+t-2\}$, $$\dim_k(G)=2k.$$ If $k=1$, then it is already known (see [@Chartrand00]) that $\dim_1(G)=\dim(G)=2$, so from now on we consider only $k\ge 2$. Without loss of generality, we suppose that $s\ge t$. Let $(u_i,v_j)$ and $(u_g,v_h)$ be two distinct vertices of $G$. Next we consider the following three cases according to the value of $k$.\ Case 1: $k\le s$. Let $S=\{u_1,\dots,u_k\}\times \{v_1,v_t\}$. We consider the following subcases.\ Subcase 1.1: $i=g$. Hence $j\ne h$, and for any vertex $(u_p,v_q)\in S$ it follows $d_{G}((u_i,v_j),(u_p,v_q))\ne d_{G}((u_g,v_h),(u_p,v_q))$. Thus, $(u_i,v_j)$ and $(u_g,v_h)$ are distinguished by $2k$ vertices of $S$.\ Subcase 1.2: $i\ne g$. Without loss of generality we assume that $i<g$. Moreover, we consider $j\le h$. First notice that ${\cal D}_G((u_i,v_j),(u_g,v_h))\supseteq \{u_1,\dots,u_i\}\times \{v_1\}$. If $i\ge k$, then clearly $\|{\cal D}_G((u_i,v_j),(u_g,v_h))\cap S\|\ge k$. Hence, we may assume $i<k$. If $g\ge k$, then there is at most one vertex $(u_f,v_1)\in \{u_{i+1},\dots,u_k\}\times \{v_1\}$ such that $d_{G}((u_i,v_j),(u_f,v_1))= d_{G}((u_g,v_h),(u_f,v_1))$, which leads to $$\|{\cal D}_G((u_i,v_j),(u_g,v_h))\cap (\{u_{1},\dots,u_k\}\times \{v_1\})\|\ge k-1.$$ On the other hand, since $i<k$, there is at least one vertex $(u_d,v_t)\in \{u_1,\dots,u_k\}\times \{v_t\}$ such that $d_{G}((u_i,v_j),(u_d,v_t))\ne d_{G}((u_g,v_h),(u_d,v_t))$. As a consequence, $$\begin{aligned} \|{\cal D}_G((u_i,v_j),(u_g,v_h))\cap S\|& = \|{\cal D}_G((u_i,v_j),(u_g,v_h))\cap (\{u_{1},\dots,u_k\}\times \{v_1\})\| \\ & \hspace{0.5cm} + \|{\cal D}_G((u_i,v_j),(u_g,v_h))\cap (\{u_{1},\dots,u_k\}\times \{v_t\})\| \\ & \ge k.\end{aligned}$$ We now consider $g < k$. Thus, ${\cal D}_G((u_i,v_j),(u_g,v_h))\supseteq \{u_g,\dots,u_k\}\times \{v_t\}$ and at most one vertex in $\{u_1,\dots,u_g\}\times \{v_1\}$ does not belong to ${\cal D}_G((u_i,v_j),(u_g,v_h))$. Moreover, ${\cal D}_G((u_i,v_j),(u_g,v_h))$ contains at least one vertex in $\{u_i,\dots,u_g\}\times \{v_t\}$. As a consequence, $$\begin{aligned} \|{\cal D}_G((u_i,v_j),(u_g,v_h))\cap S\|& \ge \|{\cal D}_G((u_i,v_j),(u_g,v_h))\cap (\{u_1,\dots,u_g\}\times \{v_1\})\| \\ & \hspace{0.5cm} + \|{\cal D}_G((u_i,v_j),(u_g,v_h))\cap (\{u_i,\dots,u_g\}\times \{v_t\}\| \\ & \hspace{1cm} + \|\{u_g,\dots,u_k\}\times \{v_t\}\| \\ & \ge g-1+1 + k-g+1 \\ & > k.\end{aligned}$$ See Figure \[fig-Case-1-2\] for examples of the situations above considering $k=7$. (-1.1,-1.1) rectangle (11.1,1.1); (-1.1, 12.9) rectangle (7.1,15.1); (8.9, 12.9) rectangle (13.1,15.1); (-0.8,-0.8) rectangle (2.8,0.8); (3.2,13.2) rectangle (12.8,14.8); (-0.8,-0.8) rectangle (12.8,0.8); in [0,2,4,6,8,10,12,14,16,18,20,22]{} [ in [0,2,4,6,8,10,12,14]{} [ at (,) \[draw, shape=circle,scale=.5\] ; ]{} ]{} in [0,2,4,6,8,10,12]{} [ at (,0) \[draw, shape=circle,fill=gray, scale=.6\] ; at (,14) \[draw, shape=circle,fill=gray, scale=.6\] ; ]{} at (2,4) \[rectangle,draw, fill=gray, scale=1\] ; at (14,4) \[rectangle,draw, pattern=north west lines, scale=1\] ; at (18,6) \[rectangle,draw, pattern=north west lines, scale=1\] ; at (6,10) \[rectangle,draw, pattern=north east lines, scale=1\] ; at (2,6) \[rectangle,draw, pattern=north east lines, scale=1\] ; at (18,8) \[rectangle,draw,fill=gray, scale=1\] ; in [1,2,3,4,5,6]{} [ at (+-2,-2) [$u_{\x}$]{}; ]{} in [7,8,9,10,11,12]{} [ at (+-2,-2) [$u_{\x}$]{}; ]{} in [1,2,3,4,5,6,7,8]{} [ at (-2,+-2) [$v_{\y}$]{}; ]{} Finally, if $j>h$, then a similar procedure gives an analogous result and we observe that $S$ is a $k$-resolving set for $G$.\ Case 2: $s<k\le s+t-2$. Note that $s\ge 3$, since the case $s=2$ would lead to the graph $C_4$, which is $2$-metric dimensional. Assume that $k=s+\alpha$ for some integer $\alpha\ge 1$. Let $S=(U\times\{v_1,v_t\})\cup (\{u_1,u_s\}\times\{v_2,\ldots,v_{\alpha+1}\})$. According to Case 1, the set $S_1= U\times\{v_1,v_t\}\subset S$ is a $s$-resolving set of $G$, which means that any pair of vertices of $G$ is recognized by at least $s$ vertices of $S_1$. On the other hand, also by Case 1 (in a similar version), the set $S_2=\{u_1,u_s\}\times\{v_2,\ldots,v_{\alpha+1}\}\subset S$ is an $(\alpha+1)$-resolving set of $G$, which similarly means any pair of vertices of $G$ is recognized by at least $\alpha+1$ vertices of $S_2$. Since only two vertices, i.e. $(u_1,v_1)$ and $(u_s,v_1)$, belong to both sets $S_1$ and $S_2$, it must happen that at least $s+\alpha+1-2$ vertices of $S$ recognize each pair of vertices of $G$. Suppose that there is a pair of vertices $(u_i,v_j)$ and $(u_g,v_h)$ such that they are distinguished by exactly $s+\alpha-1$ vertices of $S$. Thus, they are distinguished by $(u_1,v_1)$ and $(u_s,v_1)$, by $s-2$ vertices in $S_1\setminus \{(u_1,v_1),(u_s,v_1)\}$ and by $\alpha-1$ vertices in $S_2\setminus \{(u_1,v_1),(u_s,v_1)\}$. However, we can now notice that if $(u_i,v_j)$ and $(u_g,v_h)$ are distinguished by $(u_1,v_1)$ and $(u_s,v_1)$, then $(u_i,v_j)$ and $(u_g,v_h)$ are not distinguished by at most one vertex in $U\times\{v_1\}$ and by at most one vertex in $U\times\{v_t\}$, that is, $2s-2>s$ (since $s\ge 3$) vertices of $S$, which is a contradiction. Thus, each pair of vertices of $G$ is distinguished by at least $s+\alpha=k$ vertices of $S$. Therefore, $S$ is a $k$-resolving set of $G$. As a consequence of both cases, we obtain that $\dim_k(G)\le 2k$. We now want to show that no smaller $k$-resolving set can exist. Let $S'$ be a basis for $G$. We consider the vertices $(u_1,v_2)$, $(u_2,v_1)$, $(u_s,v_{t-1})$ and $(u_{s-1},v_t)$. Since $$\|A\|=\|{\cal D}_G((u_1,v_2),(u_2,v_1))\cap S'\|=\|(U\times \{v_1\})\cup (\{u_1\}\times V)-\{(u_1,v_1)\}\cap S'\|\ge k$$ and $$\|B\|=\|{\cal D}_G((u_s,v_{t-1}),(u_{s-1},v_t))\cap S'\|=\|(U\times \{v_t\})\cup (\{u_s\}\times V)-\{(u_s,v_t)\}\cap S'\|\ge k$$ we have that $$\|S'\|\ge \|A\|+\|B\|-2\ge 2k-2.$$ Now suppose $\|S'\|=2k-2$. Thus, it follows $\|A\|=k$, $\|B\|=k$, $(u_s,v_1),(u_1,v_t)\in S$ and $S=A\cup B$. So $\|(\{u_s\}\times \{v_2,\ldots,v_{t-1}\})\cup (\{u_2,\ldots,u_{s-1}\}\times \{v_t\})\cap S'\|=k-1$ and $\|(\{u_1\}\times \{v_2,\ldots,v_{t-1}\})\cup (\{u_2,\ldots,u_{s-1}\}\times \{v_1\})\cap S'\|=k-1$. We now consider the vertices $(u_1,v_1)$, $(u_2,v_2)$, $(u_{s-1},v_{t-1})$ and $(u_s,v_t)$. Hence, if we denote $Q={\cal D}_G((u_1,v_1),(u_2,v_2))$, then $$\begin{aligned} \|Q\cap S'\|&\le \|((\{u_s\}\times \{v_2,\ldots,v_{t-1}\})\cup (\{u_2,\ldots,u_{s-1}\}\times \{v_t\}))\cap S'\|\\ &= k-1\end{aligned}$$ and $$\begin{aligned} \|Q\cap S'\|&\le \|(\{u_1\}\times \{v_2,\ldots,v_{t-1}\})\cup (\{u_2,\ldots,u_{s-1}\}\times \{v_1\})\cap S'\|\\ &= k-1,\end{aligned}$$ which is a contradiction. So $\|S'\|\ge 2k-1$. If we suppose that $\|S'\|=2k-1$, then an analogous procedure to the one above gives a contradiction again. Therefore, we have that $\|S'\|\ge 2k$ and the proof is complete. Decoding -------- For the grid graphs $G=P_s {\,\Box\,}P_t$ described above, and for any $k\leq s+t-2$, we obtain $(G,k)$-codes with $n=st$ codewords of length $\ell=2k$ over an alphabet of size $s+t+1$. Such a $(G,k)$-code has correction capability $r=\left\lfloor (k-1)/2 \right\rfloor$. To decode $r'<r$ errors using the algorithm described above will require a $(2k,2k-r,r')$-uncovering, or equivalently a $(2k,r,r')$-covering design. Depending on the parity of $k$, we therefore require either: (i) if $k=2m$ is even, a $(4m,m-1,r')$-covering design; or (ii) if $k=2m+1$ is odd, a $(4m+2,m,r')$-covering design. Proposition \[prop:petrov\] provides the coverings (and thus uncoverings) we require. Given that these uncoverings have constant size, the complexity of the decoding algorithm in this case is $O(kn)$. Conclusion ========== The main achievement of this paper was to obtain a new application of $k$-resolving sets in graphs to coding theory, by obtaining a new method of constructing error-correcting codes. We considered three families of graphs, namely paths, cycles and grid graphs; naturally, there are many more graph families which could be investigated with this application in mind. We conclude by mentioning that, while $1$-resolving sets have many applications, for the application of $k$-resolving sets to error-correcting codes we require that $k\geq 3$; however, $2$-resolving sets could potentially be applied to the detection of errors. [99]{} R. F. Bailey, Uncoverings-by-bases for base-transitive permutation groups, [Des. Codes Cryptogr.]{} [41]{} (2006), 153–176. R. F. Bailey, Error-correcting codes from permutation groups, [Discrete Math.]{} [309]{} (2009), 4253–4265. R. F. Bailey and P. J. Cameron, Base size, metric dimension and other invariants of groups and graphs, [Bull. London Math. Soc.]{} [43]{} (2011), 209–242. R. F. Bailey and B. Stevens, Uncoverings on graphs and network reliability, [Australas. J. Combin.]{} [50]{} (2011), 219–231. A. F. Beardon and J. A. Rodríguez-Velázquez, On the $k$-metric dimension of metric spaces, preprint; `arXiv:1603.04049`. L. M. Blumenthal, [Theory and Applications of Distance Geometry]{}, Clarendon Press, Oxford, 1953. P. J. Cameron, Permutation codes, [European J. Combin.]{} [31]{} (2010), 482–490. G. Chartrand, L. Eroh, M. A. Johnson and O. R. Oellermann, Resolvability in graphs and the metric dimension of a graph, [Discrete Appl. Math.]{} [105]{} (2000), 99–113. W. Chu, C. J. Colbourn and P. Dukes, Constructions for permutation codes in powerline communications, [Des. Codes Cryptogr.]{} [32]{} (2004), 51–64. C. J. Colbourn and J. H. Dinitz (editors), [Handbook of Combinatorial Designs]{} (second edition), CRC Press, Boca Raton, 2007. T. H. Cormen, C. E. Leiserson, R. L. Rivest and C. Stein, [Introduction to Algorithms]{} (second edition), MIT Press, Cambridge/McGraw-Hill, Boston, 2001. A. Estrada-Moreno, J. A. Rodríguez-Velázquez and I. G. Yero, The $k$-metric dimension of a graph, [Appl. Math. Inf. Sci.]{} [9]{} (2015), 2829–2840. A. Estrada-Moreno, I. G. Yero and J. A. Rodríguez-Velázquez, The $k$-metric dimension of corona product graphs, [Bull. Malaysian Math. Sci. Soc.]{} [39:1]{} (2016), 135–156. A. Estrada-Moreno, I. G. Yero and J. A. Rodríguez-Velázquez, The $k$-metric dimension of the lexicographic product of graphs, [Discrete Math.]{} [339]{} (2016), 1924–1934. D. M. Gordon, La Jolla Covering Repository, <http://www.ccrwest.org/cover.html>. F. Harary and R. A. Melter, On the metric dimension of a graph, [Ars Combin.]{}, [2]{} (1976), 191–195. S. Khuller, B. Raghavachari and A. Rosenfeld, Landmarks in graphs, [Discrete Appl. Math.]{} [70]{} (1996), 217–229. H.-J. Kroll and R. Vincenti, Antiblocking systems and PD-sets, [Discrete Math.]{} [308]{} (2008), 401–407. W. H. Mills and R. C. Mullin, Coverings and packings, in Contemporary Design Theory: A collection of surveys, (eds J. H. Dinitz and D. R. Stinson), John Wiley & Sons, New York, 1992. V. S. Pless, [Introduction to the Theory of Error-Correcting Codes]{} (third edition), John Wiley & Sons, New York, 1998. J. Schönheim, On coverings, [Pacific J. Math]{} [14]{} (1964), 1405–1411. A. Sebő and E. Tannier, On metric generators of graphs, [Math. Oper. Res.]{} [29]{} (2004), 383–393. P. J. Slater, Leaves of trees, [Congr. Numer.]{} [14]{} (1975), 549–568. I. Tamo and M. Schwartz, Correcting limited-magnitude errors in the rank-modulation scheme, [IEEE Trans. Inform. Theory]{} [56]{} (2010), 2551–2560. I. G. Yero, A. Estrada-Moreno and J. A. Rodríguez-Velázquez, Computing the $k$-metric dimension of graphs, [Applied Mathematics and Computation]{} [300]{} (2017), 60–69. [^1]: School of Science & Environment (Mathematics), Grenfell Campus, Memorial University of Newfoundland, Corner Brook, NL A2H 6P9, Canada. Email: `[email protected]` [^2]: Departamento de Matemáticas, Escuela Politécnica Superior de Algeciras, Universidad de Cádiz, Av. Ramón Puyol s/n, 11202 Algeciras, Spain. Email: `[email protected]` [^3]: Personal communication via `mathoverflow.net`, February 2016.	{ "pile_set_name": "ArXiv" }
Ł ¶ =21 cm =0 cm =16 cm =0.2 in =0.2 in V. Aldaya$^{2,3}$, M. Calixto$^2$ and J. Guerrero$^{2,4}$ September 20, 1995 [Abstract]{} The problems arising when quantizing systems with periodic boundary conditions are analysed, in an algebraic (group-) quantization scheme, and the “failure" of the Ehrenfest theorem is clarified in terms of the already defined notion of [good]{} (and [bad]{}) operators. The analysis of “constrained" Heisenberg-Weyl groups according to this quantization scheme reveals the possibility for new quantum (fractional) numbers extending those allowed for Chern classes in traditional Geometric Quantization. This study is illustrated with the examples of the free particle on the circumference and the charged particle in a homogeneous magnetic field on the torus, both examples featuring “anomalous" operators, non-equivalent quantization and the latter, fractional quantum numbers. These provide the rationale behind flux quantization in superconducting rings and Fractional Quantum Hall Effect, respectively. Introduction ============ The need for a consistent quantization scheme which is truly suitable for systems wearing a non-trivial topology is increasing daily. Configuration spaces with non-trivial topology appear in as diverse cases as Gauge Theories, Quantum Gravity, and the more palpable ones of the superconducting ring and the Quantum Hall effect, where the measuring tools change the topology of the system in a non-trivial way. The most common problem which appears when the configuration-space manifold possesses a non-trivial topology is the failure of the Ehrenfest theorem for certain operators, a problem usually referred to as an anomaly. In the sequel, we shall add the qualifier [topologic]{} to distinguish these from others directly attached to the Lie algebra of the quantum operators and characterized, roughly speaking, by the appearance of a term in a quantum commutator not present at the classical, Poisson-algebra level. We call them [algebraic]{} anomalies and refer the reader to [@Schrodinger] for a detailed analysis. The failure of the Ehrenfest theorem for a given operator is primarily related to the non-globality of the corresponding classical function, such as is the case of the local co-ordinate on a one-dimensional closed submanifold. Geometric Quantization was intended to go further than ordinary canonical quantization does, allowing for the quantization of arbitrary symplectic manifolds. Unfortunately, Geometric Quantization only partially accomplished this task, one of the reasons being the difficulty in (or, even more, the impossibility of) finding a polarization suitable enough to quantize a given set of classical functions [@Woodhouse; @Isham], or quantizing a set of operators in a way that would preserve a given polarization. A quantization procedure based on a group structure, Group Approach to Quantization (GAQ) [@GAQ; @Schrodinger], improves the standard Geometric Quantization approach in that it provides two sets of mutually commuting operators, namely, the left- and right-invariant vector fields. This enables us to impose the polarization conditions by means of the left-invariant vector fields, say, while the right-invariant ones will be the quantum operators, which automatically preserve the polarization. The quantization group is endowed with a $U(1)$-principal bundle structure so that generators fall into two classes according to whether or not they give rise to a term proportional to the vertical generator on the r.h.s. of a commutator. Generators which do not reproduce any $U(1)$-term close a [horizontal]{} subalgebra, the characteristic subalgebra, of non-dynamical generators, which should be included in the polarization subalgebra. The principal drawback of GAQ is the need for a group symmetry associated with the system to be quantized, and the apparent restriction in the number of functions which can be quantized. However, this last limitation is slighter than it might seem, since Canonical Quantization on a particular phase space does not quantize the entire set of functions on phase space, but rather, a restricted Poisson subalgebra. Even more, it could well happen in some cases that a more standard quantization provides only quantum operators corresponding to a finite-dimensional Lie algebra. This is the case, for instance, of the symplectic manifold $S^2$, where the quantum operators are only those of $su(2)+R$ [@Gotay]. Moreover, most of the interesting systems in Physics possess a symmetry group large enough to achieve a proper quantization. To be more precise, not only the right invariant vector fields preserve the polarization, but rather the entire right enveloping algebra preserve the structure of the Hilbert space. This means that any element in the right enveloping algebra can be realized as a quantum operator, although the relation between the quantum algebra and the standard Poisson algebra on the co-adjoint orbits of the group is no longer an isomorphism; GQA provides a quantum theory rather than the quantization of a classical theory. A reformulation of GAQ was proposed a few years ago [@Alfonso], the Algebraic Quantization on a Group (AQG) \[some of the basic ideas in [@Alfonso] have also appeared in the context of quantum systems with non-trivial topology [@Landsman] and in Quantum Gravity ([@Ashtekar] and references therein)\] , which generalizes GAQ in two respects. Firstly, finite transformations generalize the infinitesimal ones throughout the method; that is, any concept or condition relative to Lie subalgebras is generalized by its counterpart in terms of Lie subgroups, thus allowing discrete transformations to enter the theory. Needless to say, infinitesimal objects are employed whenever possible. Secondly, it generalizes the $U(1)$ phase invariance in Quantum Mechanics (the structure group of the principal bundle fibration of the quantum symmetry) incorporating other symmetries, eventually interpreted as constraints. The new structure group $T$, which must include the traditional $U(1)$, may also contain discrete symmetries especially suitable to simulate manifold surgery as, for instance, toral compactification, by means of periodic boundary conditions. From now on we shall call “compactified" (cylindrical or toral) Heisenberg-Weyl (H-W) group a H-W group where the structure group is $T$ rather than $U(1)$, this subgroup $T$ being the factor subgroup leading to a compactified classical (the cylinder or the torus) phase space by the quotient $\Gtm/T$. The generalization of the $U(1)$-equivariance to $T$-equivariance condition on the wave functions gives rise to two new, closely related features: a) the existence of non-equivalent quantizations associated with non-equivalent representations of the larger structural subgroup $T$, and b) the notion of [good]{} operators, constituting the subgroup of transformations compatible with the $T$-equivariance condition, in a sense to be specified later (see [@Alfonso]). Furthermore, those operators not preserving the $T$-equivariance condition, the [bad]{} operators, may be seen as quantization-changing transformations, and exhibit topologic anomalies. Like in the $T=U(1)$ case, all the elements of the right enveloping algebra [compatible with the $T$-equivariance condition]{}, for arbitrary $T$, can be realized as [good]{} quantum operators. It should be noted that, as mentioned above, AGQ is formulated in terms of finite objects. This means that some algebraic indices must replace the well-know Chern class $[\omega]$ of the symplectic form in Geometric Quantization. In fact, the indices characterizing the (not necessarily central) extension by $T$ of the “classical" group $G$ generalize the Chern class, providing also [fractional]{} values. This is precisely the case of the motion of a charged particle on a torus in the presence of a homogeneous magnetic field, closely related to the (Fractional) Quantum Hall Effect. The appearance of fractional quantum numbers generalizing the integer Chern classes reveals, once again, that the procedure of taking constraints and that of quantizing, depending at least on the specific methods employed, may not commute. This paper is organized as follows. Section 2 illustrates the way in which AQG operates with the help of the examples of the Heisenberg-Weyl group in 1D with constraints mimicking the compactification of the coordinate $x$ (Sec 2.1) and that of the compactification of both $x$ and $p$ (Sec. 2.2). In the latter case, generalizing the quantization of a compact phase space [à la Dirac]{}, a not necessarily integer quantization condition is obtained which generalizes that of Geometric Quantization, i.e. the condition $[\omega]\in Z$ (Chern class), and, associated with it, vector-valued wave functions. In solving this problem, real (versus holomorphic) polarizations have been employed, leading to a generalized $kq$-representation. This technique simplifies the treatment and is much more intuitive, even though the configuration-space wave functions contain delta functions. The results obtained in Sec. 2 are applied to the quantization of the free particle on the circumference (Sec. 3, where the failure of Ehrenfest theorem is analysed), directly related to flux quantization in superconducting rings, and to the quantization of a charged particle on a torus in the presence of an homogeneous transverse magnetic field (Sec. 4), providing the rationale behind Integer and Fractional Quantum Hall Effect. Algebraic Quantization of “compactified" Heisenberg -Weyl groups: Fractional quantum numbers ============================================================================================ In this section, we shall explain the AQG formalism over the example of the Heisenberg-Weyl group with one co-ordinate “compactified", i.e. with constraints associated with the compactification of one co-ordinate, and with one co-ordinate and its canonically conjugate momentum “compactified". We nevertheless recommend the reading of the Ref. [@Alfonso]. Although explicit calculations are given for the Heisenberg-Weyl group with only one co-ordinate-momentum pair, the results can be generalized, immediately, to any finite number of them. Cylindrical Heisenberg-Weyl group --------------------------------- Let us firstly proceed with the case of the Heisenberg-Weyl group with only one of the coordinates “compactified", i.e. with structure group $T$ such that the quotient $\Gtm/T$ leads to the cylinder as the symplectic manifold. The starting point in AQG is a Lie group which is a right-principal bundle with structure group $T$. $T$ is itself a principal bundle with $U(1)$ as structure group. In our case is the ordinary Heisenberg-Weyl group in 1D (throughout the paper, 1D means one coordinate-momentum pair, $x$ and $p$), and $T=U(1)\times\{e_k,\,k\in Z\}$, where $\{e_k,\,k\in Z\}$ is the subgroup of of finite translations in the coordinate $x$ by an amount of $kL$, $L$ being the spatial period. Note that $T$ is isomorphic to $U(1)\times Z$, so that its fibration is trivial. The group law $g''=g'g$ for is: x” & = & x’+ x\ p” & = & p’ + p \[H-WcGLaw\]\ ” & = & ’\^[\[(1+)x’p+xp’\]]{} where the first two lines correspond to the group law of $G$, and the third to that of $U(1)$. The real parameter $\lambda$ has been introduced to account for a complete class of central extensions differing in a coboundary \[coboundaries have the form $\xi(g',g)=\eta(g'g)-\eta(g')-\eta(g)$, where $\eta:G\rightarrow R$ is called the generating function of the coboundary\] generated by the function $\eta(x,p)=\lambda xp$ (In particular, for $\lambda=-\medio$ we have Bargmann’s cocycle). From this group law we can read immediately the right and left translations, $R_{g'}g=g'g=L_gg'$. In particular, the left- and right-invariant vector fields (generating the finite translations) become: [ \^[L]{}\_[x]{} ]{} & = & [ ]{} + p\ [ \^[L]{}\_[p]{} ]{}& = & [ ]{} + x\ [ \^[L]{}\_ ]{} & = & i[ ]{} [ \^[R]{}\_[x]{} ]{}& = & [ ]{} + p\ [ \^[R]{}\_[p]{} ]{}& = & [ ]{} + x\ [ \^[R]{}\_ ]{} & = & i[ ]{} \[XLRH-Wc\] The quantization 1-form (the left-invariant 1-form associated with the parameter $\z$) can also be obtained: = - pdx - (1+)xdp + \[1formC\] Since we are not considering time evolution, the quantization 1-form has no characteristic subalgebra, there exists no discrete characteristic subgroup $\GC$, and any combination of the two generators ${ {\tilde{X}}^{L}_{x} }$ and ${ {\tilde{X}}^{L}_{p} }$ constitutes a first-order full polarization (with the time evolution added, as in the free particle in 1D, things are a bit more complicated, see Sec. 3). Two polarizations are singled out, ${\cal P}_p = <{ {\tilde{X}}^{L}_{x} }>$ and ${\cal P}_x = <{ {\tilde{X}}^{L}_{p} }>$, or their finite (versus infinitesimal) counterparts $\GP{}_p=\{\hbox{Space translations}\}$ and $\GP{}_x=\{\hbox{Boosts transformations}\}$, leading to momentum and configuration space representations, respectively. It should be borne in mind that the polarization conditions are needed to reduce the group representation which otherwise would provide only the Bohr-Sommerfeld quantization. These polarization conditions read, in general, ${ {\tilde{X}}^{L}_{} }\Psi=0,\,\forall { {\tilde{X}}^{L}_{} }\in {\cal P}$ or $\Psi(g\GP)=\Psi(g)$ in finite terms. The $T$-function condition generalizes ordinary phase invariance ($U(1)$-equivariance) in Quantum Mechanics, which is written $\Psi(\zg)=\rho(\z)\Psi(g)$, where $\rho(\z)$ is the natural representation of $U(1)$ on the complex numbers, $\rho(\z)=\z$. The generalization to a bigger group $T$ involves the use of a general representation ${\cal D}$ of $T$ (or, to be precise, of $T_B\equiv U(1)\cup T_p$, where $T_p$ is a maximal polarization subgroup of $T$; see [@Alfonso]) on a complex vector space $E$, where the wave functions themselves take their values. In the formalism of AQG, the representation of $T_B$ is constructed from the very representation of , i.e. the vector space $E$ on which the constrained functions are evaluated is made out of the unconstrained wave functions by properly choosing their arguments. This is the reason why the group $T_B$ is interpreted as constraints: the representation of $T_B$ is not an abstract representation, but rather built with the same functions of the representation of . The $T_B$-function condition then reads $\Psi(g_{T_B}g)={\cal D}(g_{T_B})\Psi(g), \forall g_{T_B}\in T_B, \forall g\in \Gtm$. In the present case $T_B=T$ and ${\cal D}(\z,e_k)=\z D(e_k)$, where $D(e_k)$ is a representation of $\{e_k,\,k\in Z\}$ ($\approx Z$) in the complex numbers, and there is an infinity of non-equivalent irreducible representations, of the form $D^\ep(e_k)=\e^{\idhbar \ep kL}$, with $\ep\in [0,\frac{2\pi\hbar}{L})$ (the first Brillouin zone, in Solid State nomenclature). Therefore, [there is a non-equivalent quantization associated with each choice of non-equivalent representation of $T$, parameterized by $\ep$]{}. The $T$-function condition for the wave function implies the restriction: \^[(1+)kLp]{}\^(x+kL,p,)=\^[kL]{} \^(x,p,) \[H-WcConstraint\] Note that the constrained wave functions can be identified with the space of sections of a $U(1)$-bundle on the cylinder, with connection given by (\[1formC\]). We now impose the polarization conditions in order to reduce the representation. Firstly, we shall consider the momentum space representation, where the polarization conditions (either in finite or infinitesimal form) lead to the following form of the wave functions: \^(x,p,) = \^[-xp]{} \^(p) \[WFH-WcMom\] where the fact that $\Psi(\z g)=\z\Psi(g)$ (by the $T$-function property) has been used. Both conditions (\[H-WcConstraint\]) and (\[WFH-WcMom\]) together imply for the wave function $\Phi^\ep(p)$ a form like: \^(p)=\_[kZ]{} \_k \^\_k(p) \[WFH-WcMom2\] where $\phi^\ep_k(p)\equiv\delta(p-\ep-\frac{2\pi\hbar}{L}k)$, i.e. the wave function is peaked at the values of the momentum $p_k^\ep=\ep+\frac{2\pi\hbar}{L}k$, $k\in Z$. The Hilbert space ${\cal H}^\ep(\Gtm)$ is made from the wave functions defined by (\[WFH-WcMom\]) and (\[WFH-WcMom2\]). The quantum operators, defined as $\P\equiv -i\hbar{ {\tilde{X}}^{R}_{x} }\,\,$ and $\X\equiv i\hbar{ {\tilde{X}}^{R}_{p} }$, act on the wave functions as: ¶\^& = & p \^\ \^& = & \^[-xp]{} \^\[H-WcOMom\] One of the main consequences of having generalized the structure group in AQG is the classification of the operators (actually left translations) as good and bad operators according to whether or not they are compatible with the $T$-function condition. More precisely, the subgroup of good operators, $\GH$, is characterized by the condition (see [@Alfonso]) Ad(), g\_TT \[Good\] In the present case, and due to the discrete character of the “physical" momenta, the position operator $\X$ is expected to be problematic, since the subgroup of good transformations compatible with (\[H-WcConstraint\]) and (\[WFH-WcMom\]) is the subgroup of , in which the continuous variable $p$ is substituted by the discrete variable $p_k\equiv p_k^0=\frac{2\pi\hbar}{L}k$, $k\in Z$, as can be deduced from $Ad(\Gtm)\left[\en,g\right]=(0,0,\e^{\idhbar nLp})\subset\GP\,\, \forall n\in Z$. Therefore, the good operators are $\P$ and the finite boosts transformations by the amount of $p_k$. Position is not a good operator in the sense that it does not preserve the structure of the wave functions, i.e. it does not leave the Hilbert space (for fixed $\ep$) ${\cal H}^\ep(\Gtm)$ stable. This fact will be further discussed in Sec. 2.1.1. With regard to the configuration space representation given by the polarization ${\cal P}_x$ or the polarization subgroup $\GP{}_x$, the solutions to this polarization are: (x,p,) = \^[-(1+) px]{} (x) \[WFH-WcConf\] Applying the condition of $T$-function (\[H-WcConstraint\]) to this wave functions in configuration space, we obtain: \^[(1+)kLp]{} \^[-(1+)(x+kL)p]{}\^(x+kL) = \^[kL]{}\^[-(1+)xp]{}\^(x) $\forall k\in Z$, where the quasi-periodicity condition for $\Phi^\ep(x)$ immediately follows: \^(x+L)=\^[L]{}\^(x) \[H-WcPeriod\] It should be stressed that this result is independent of the chosen cocycle, since it does not depend on $\lambda$, as expected. The quantum operators are: ¶\^& = & \^[-(1+) px]{} \^\ \^& = & x \^\[H-WcOConf\] Again, the position operator $\X$ is not a good operator, for the same reason as in the momentum-space case, and the subgroup of (left) transformations leaving the structure of the wave functions (\[WFH-WcConf\]) and (\[H-WcPeriod\]) stable is the same $\GH$ as before, containing only $\P$ and the finite boosts in $p_k$, $k\in Z$. Therefore, the standard position has no meaning for any (Galilean) system with the circumference as configuration space (see Sec. 3). ### Is there any good position-like operator? The position operator $\X$ is not a good operator because the variable $x$ is not periodic: if $\phi(x)$ is a quasi-periodic function, $x\phi(x)$ is no longer quasi-periodic. However, the function $\eta=\e^{i\frac{2\pi}{L}x}$ is periodic, so that we could define the operator $\hat{\eta}\equiv\e^{i\frac{2\pi}{L}\X}$, and verify that $\hat{\eta}\Psi^\ep=\e^{i\frac{2\pi}{L}x}\Psi^\ep$ satisfies the same quasi-periodicity condition as $\Psi^\ep$. We can then say that $\hat{\eta}$ is a good operator. The reason why $\hat{\eta}$ is a good operator is precisely that it generates a good finite boost. We know that the only good boosts are indexed by $p_k=\frac{2\pi\hbar}{L}k$, i.e. \^(p\_k\g)=\^[k[ \^[R]{}\_[p]{} ]{}]{}\^(g)= (\^[i]{})\^[-k]{}\^(g)=\^[-k]{}\^(g) This means that $\hat{\eta}^k,\,k\in Z$ are the only good position operators. The finite operator $\hat{\eta}$ is obviously not Hermitian; rather, it is unitary as it should be. However $\hat{\eta}$ can be written as $\hat{\eta}=\cos(\frac{2\pi}{L}\X) + i\sin(\frac{2\pi}{L}\X)$, the good operators $\cos(\frac{2\pi}{L}\X)$ and $\sin(\frac{2\pi}{L}\X)$ being Hermitian. These are good operators, given that they are periodic functions of the operator $\X$. Since the set of functions $\{\e^{i\frac{2\pi}{L}mx},\,m\in Z\}$ constitutes a basis for the periodic functions of $x$ in the interval $[0,L]$, any operator which is a periodic function of the position operator $\X$ is a good operator. In any case, we might wonder about the finite boosts transformations for $\tilde{p}\neq p_k$, i.e. about transformations of the form $\Phi'(x)=\e^{\idhbar \tilde{p}\X}\Phi^\ep(x)=\e^{\idhbar \tilde{p}x} \Phi^\ep(x)$. This new function verifies the boundary conditions $\Phi'(x+L)=\e^{\idhbar \tilde{p}(x+L)}\Phi^\ep(x+L)= \e^{\idhbar (\ep+\tilde{p})L}\Phi'(x)$, and therefore belongs to the Hilbert space ${\cal H}^{\ep+\tilde{p}}(\Gtm)$. In fact, since the representations parameterized by $\ep$ and $\ep+\frac{2\pi\hbar}{L}k, \,k\in Z$ are equivalent, the transformed wave functions lie in the representation $(\ep+\tilde{p}) \hbox{ mod } \frac{2\pi\hbar}{L}$. Of course, if $\tilde{p}=p_k$ for some $k$, the transformed wave function lies in the same Hilbert space as before and we recover the result that the finite boosts in $p_k$ are good operators. In particular, $\Phi^\ep(x)=\e^{\idhbar \ep\X}\Phi^0(x)=\e^{\idhbar \ep x}\Phi^0(x) \equiv \e^{\idhbar \ep x}\Phi(x)$, with $\Phi(x)$ satisfying (the usual) periodic boundary conditions. This means that all the Hilbert spaces ${\cal H}^{\ep}(\Gtm)$, although yielding non-equivalent representations, are related to each other by means of finite boosts transformations, which are unitary transformations considered in the union of all these Hilbert spaces $\cup_{\ep\in[0,2\pi\hbar/L)}{\cal H}^{\ep}(\Gtm)$. We could say that ${\cal H}^{\ep}(\Gtm)$ for a fixed $\ep$ is too small for the boosts operator to live in. The momentum operator, however, preserves (and is Hermitian in) each one of these Hilbert spaces, but it is not Hermitian in the union of all of them. It is worth mentioning that the set of operators $\P$, $\hat{\eta}$ and $\hat{\eta}^{\dag}$ close a Lie algebra under ordinary commutation which is isomorphic to the non-extended harmonic oscillator algebra. The operators $\hat{\eta}$ and $\hat{\eta}^{\dag}$ act as ladder operators on the eigenfunctions of $\P$ (this fact has been used in [@O-K] to study Quantum Mechanics on the circumference). Toral Heisenberg-Weyl group --------------------------- Let us now proceed with the case of the Heisenberg-Weyl group with both the coordinate and the momentum “compactified", i.e. with a structure group $T$ such that $\Gtm/T$ leads to the torus as the symplectic manifold. We shall parameterize the plane with coordinates $(x_1,x_2)$ because in physical applications the coordinates play the double rôle of coordinate and momentum (see Sec. 4.1). We also apply AQG to this system. Here again, is the Heisenberg-Weyl group in 1D (now parameterized by $\x=(x_1,x_2)$ and $\z$). Given $L_1$ and $L_2$, we introduce the lattice points $\L_{\k}\equiv (k_1 L_1,k_2 L_2)$, $k_1$ and $k_2$ being integers (thus defining a rectangular torus if we would take quotient by them). The structure group $T$ will be a principal bundle with base $\left\{e_{\k},\,\k\in Z\times Z \right\}$ and fiber $U(1)$, where $\left\{e_{\k},\,\k\in Z\times Z \right\}\subset \Gtm$ is the set of finite translations in the coordinates $\x$ by an amount of $\L_{\k}$. The fibration of $T$ is non-trivial in general. The group law for now reads: ” & = & ’+\ ” & = & ’\^[m]{} \[H-WtGLaw\] where a new numerical constant $\w$ (with dimensions of $T^{-1}$), besides the mass $m$, which was implicit in the momentum $p=mv$, has been introduced to accommodate the dimensions in the exponent above. The left and right invariant vector fields can be obtained: [ \^[L]{}\_[x\_1]{} ]{} & = & [ ]{} + mx\_2\ [ \^[L]{}\_[x\_2]{} ]{}& = & [ ]{} + mx\_1\ [ \^[L]{}\_ ]{} & = & i[ ]{} [ \^[R]{}\_[x\_1]{} ]{}& = & [ ]{} + mx\_2\ [ \^[R]{}\_[x\_2]{} ]{}& = & [ ]{} + mx\_1\ [ \^[R]{}\_ ]{} & = & i[ ]{} , \[XLRH-Wt\] and the quantization 1-form is: = - mx\_2dx\_1 - (1+)mx\_1dx\_2 + \[1formT\] As before, the quantization 1-form has no characteristic module, and any combination of the two generators ${ {\tilde{X}}^{L}_{x_1} }$ and ${ {\tilde{X}}^{L}_{x_2} }$ constitutes a first-order full polarization. These can be written as ${\cal P}_{\n} = < \n\cdot{ {\tilde{X}}^{L}_{\x} }>$, where $\n=(n_1,n_2)$ is an arbitrary unit vector. The choice of an $\n$ corresponds to the selection of a particular direction in the plane. \[All directions are indistinguishable, but on the mimicked torus, there are geodesics (directions) which close, as happens with the lines $x_2=0$ and $x_1=0$, and others which are open and fill the torus densely. It can be easily checked that the condition for a geodesic with direction given by $\n$ to close is either that $\frac{n_2}{n_1}=\frac{k_{02}}{k_{01}} \frac{L_2}{L_1}\,,k_{01},k_{02}\in Z$ or that $\n=(1,0)$ or $\n=(0,1)$, i.e. $\n$ is of the form $\n=\L_{\kO}/\|\L_{\kO}\|$, with $\kO\in Z\times Z$. Also, for a geodesic and its orthogonal one to close, it is necessary and sufficient that $\frac{L_2^2}{L_1^2}$ be a rational, except for the case $\n=(1,0)$ and $\n=(0,1)$, which are always orthogonal and closed. This condition is similar to the condition of commensurability of the frequencies for a Lissajoux figure to be closed\]. The polarization condition ${\cal P}_{\n}$ leads to the following wave functions: =\^[-m]{} (y\_2) \[H-WtWF\] where $y_1\equiv \n\cdot\x\,,\,\, y_2\equiv \n\cdot\J\cdot\x$, and $(\J)_{ij}=\epsilon_{ij}\,$, with $\epsilon_{12}=1$. The action of the right operators on these wave functions is: [ \^[R]{}\_ ]{}&=& my\_2\ [ \^[R]{}\_ ]{}&=& \^[-m]{}\ & & (y\_2) \[H-WtOp\] Before imposing the constraints, we have to determine the structure of the group $T$. It must be done by means of finite transformations, since it is basically a discrete group (times $U(1)$). We then compute the group commutator of two elements of $\left\{e_{\k}\right\}_{\k\in Z\times Z}$, with the result $[e_{\kp},e_{\k}]=(0,0,\e^{\idhbar m\w L_1L_2(k_1'k_2-k_2'k_1)})$. Two cases have to be considered: - $\e^{\idhbar m\w L_1L_2(k_1'k_2-k_2'k_1)}=1 \,\forall\, \k,\,\kp \in Z\times Z \;\Rightarrow \; [e_{\kp},e_{\k}]={\bf1}_{\Gtm}$ , - $\exists\,\k$ and $\kp$ / $\e^{\idhbar m\w L_1L_2(k_1'k_2-k_2'k_1)}\neq 1\; \Rightarrow \; {\bf 1}_{\Gtm}\neq [e_{\kp},e_{\k}]\in U(1)$ , For the case [i)]{}, $T$ is the direct product $T=\left\{e_{\k},\,\k\in Z\times Z\right\}\times U(1)$ and the whole group $T$ can be imposed as constraints. For the case [ii)]{}, when $\e^{\idhbar m\w L_1L_2(k_1'k_2-k_2'k_1)}\neq 1$ for some values of $\k$ and $\kp$ (an infinite discrete set of values, in fact), there are two possibilities, depending on whether $\frac{m\w L_1L_2}{2\pi\hbar}$ is rational or irrational. In neither case can we impose the entire group $T$ as a constraint group and we have to choose a polarization subgroup $T_p$ of $T$ (see [@Alfonso]). ### Integer quantum numbers For the condition [i)]{} to hold, it is necessary that =nN \[H-WtQuant\] , which implies a [quantization]{} of the “frequency" $\w$. As we shall see in Sec. 4, this condition will imply the quantization of the magnetic flux through the torus surface. This quantization condition is of the same nature as that of the Dirac monopole case. Concerning this case, AQG simply reproduces the quantization condition of the standard Geometric Quantization: the symplectic form must be of integer class, defining the Chern class of the quantum manifold. The rest of the procedure follows the same lines as in the case of the cylindrical H-W group: the condition of $T$-function is $\Psi(g_{T}g)={\cal D}(g_T)\Psi(g)$, with ${\cal D}(e_{\k},\z)=\z D(e_{\k})$, where $D(e_{\k})$ is a representation of the group $\left\{e_{\k},\,\k\in Z\times Z\right\}\approx Z\times Z$ on the complex numbers. For the moment, we shall use the trivial representation $D^0(e_{\k})=1$, and later the rest of non-equivalent representations (leading to non-equivalent quantization of ) will be computed with the help of the bad operators, as was shown in Sec. 2.1.1 for the cylindrical H-W group. The $T$-function condition then reads: \^[m]{} \^[0]{}(+Ł\_,)= \^[0]{}(,) \[H-WtConstraint\] $\forall \k\in Z\times Z$. Note that the space of constrained wave functions can be identified with the space of sections of a $U(1)$-bundle on the torus, with Chern class $n$ and connection given by (\[1formT\]). Applying this constraint to the polarized wave functions (\[H-WtWF\]) the following restriction is obtained: \^[m{ y\_2 + (1+2)n\_1n\_2y\_1+ \[(1+)n\_2\^2-n\_1\^2\](Ł\_)} (Ł\_)]{}& &\ \^[-m(Ł\_)]{} \^[0]{}(y\_2+Ł\_) &=& \^[0]{}(y\_2) $\forall \k\in Z\times Z$. This restriction has important consequences: $a)$ the possible polarizations are only those given by $\n=(1,0)$ and $\n=(0,1)$; $b)$ the wave function is peaked at certain equally spaced values of $y_2$; and $c)$ the parameter $\lambda$ is also quantized. From these facts it can also be deduced that the dimension of the representation is $n$, i.e. the representations of the toral Heisenberg-Weyl group are finite-dimensional, having dimension $n$, where $n$ is given by (\[H-WtQuant\]). Explicitly, the “allowed" values for the coordinates are $x_2=\frac{k}{n}L_2,\, k\in Z$ for $\n=(1,0)$ and $x_1=\frac{k}{n}L_1,\, k\in Z$ for $\n=(0,1)$. The wave functions then turn out to be, respectively: \^[0]{}(x\_2)=\_[kZ]{}a\_k (x\_2-L\_2) && =(1,0) \[WFt1\]\ \^[0]{}(x\_1)=\_[kZ]{}b\_k (x\_1-L\_1) && =(0,1) \[WFt2\] The coefficients $a_k$ and $b_k$ are not completely arbitrary; due to the $T$-function condition, which now reads $\Phi^0(x_2+k_2L_2)=\Phi^0(x_2)$ (for $\n=(1,0)$) and $\Phi^0(x_1+k_1L_1)=\Phi^0(x_1)$ (for $\n=(0,1)$) $\forall \k\in Z\times Z$, they satisfy $a_{k+n}=a_k$, and $b_{k+n}=b_k,\,\forall k\in Z$. Then, there are only $n$ independent coefficients, so that the dimension of the representation is $n$. The allowed values for $\lambda$ are given by $\lambda=\frac{k}{n}\,,\, k\in Z$, i.e. the possible (equivalent) cocycles, or that which is the same, the possible coboundaries are quantized. This fact can be easily understood in terms of the generating function of the coboundary parameterized by $\lambda$, which has the form $\lambda x_1x_2$, or better, $\e^{\idhbar m\w\lambda x_1x_2}$. For this function to be quasi-periodic, i.e. $\e^{\idhbar m\w\lambda (x_1+k_1L_1)(x_2+k_2L_2)}= \e^{i\ep\cdot\L_{\k}}\e^{\idhbar m\w\lambda x_1x_2},\forall\k\in Z\times Z$, the quantization condition for $\lambda$ is necessary, besides the quantization condition for $x_1$ and $x_2$. Let us focus on the case $\n=(1,0)$ for concreteness (the case $\n=(0,1)$ is completely analogous and in fact equivalent). Using the expression (\[WFt1\]) and the fact that $a_{k+n}=a_k, \forall k\in Z$, the summatory can be regrouped, and we arrive at a rather compact form for the wave functions: \^0(x\_2)&=&\_[kZ]{}a\_k (x\_2-L\_2) =\_[k=0]{}\^[n-1]{}\_[k\_2Z]{}a\_[k+nk\_2]{} (x\_2-L\_2)\ &=&\_[k=0]{}\^[n-1]{} a\_k \_[k\_2Z]{} (x\_2-L\_2)\ &=&\_[k=0]{}\^[n-1]{} a\_k \^0\_k(x\_2) where ($x_2^{(k)}\equiv x_2-\frac{k}{n}L_2$) \^0\_k(x\_2)\_[k\_2Z]{}(x\_2\^[(k)]{}-k\_2L\_2) = \_[qZ]{} \^[i2q x\_2\^[(k)]{}/L\_2]{} Therefore, the dimension of the Hilbert space ${\cal H}^0(\Gtm)$ is $n$, since it is spanned by the functions $\Lambda^0_k(x_2),\,k=0,1,...,n-1$. Next, we determine the subgroup $\GH$ of good transformations (those preserving the structure of the wave function). As in the case of the cylindrical H-W group, it is deduced from $Ad(\Gtm)\left[e_{\kp},g\right]= \left(0,0,\e^{\idhbar m\w(k_1L_1x_2 - k_2L_2x_1)}\right)\subset\GP\,\, \forall \k\in Z\times Z$, which implies $\frac{m\w}{\hbar}(k_1L_1x_2-k_2L_2x_1)=2\pi k, \,k\in Z$, and, together with the quantization condition (\[H-WtQuant\]) for $\w$, leads to $\x=\frac{1}{n}\L_{\k}$. Therefore, the subgroup $\GH$ of good transformations is the subgroup of in which the parameters $\x$ are restricted to be $\x=\frac{1}{n}\L_{\k}$, although only a finite number of them corresponding to $\{\x=(\frac{k_1}{n}L_1,\frac{k_2}{n}L_2),\, k_1,k_2=0,1,...,n-1\}\,$ are actually different, due to the $T$-function condition. Consequently, no infinitesimal transformation (apart from that of $U(1)$) preserves the structure of the wave function. If we introduce the (finite) operators $\hat{\eta}_i\equiv \e^{L_i{ {\tilde{X}}^{R}_{x_i} }},\, i=1,2$, in a similar way as in Sec. 2.1.1 (although here they represent finite translations), we can write the elements of $T$ as $e_{\k}\equiv (\hat{\eta}_1)^{k_1}(\hat{\eta}_2)^{k_2}$, and the subgroup of good operators is: = {(\_1)\^(\_2)\^, k\_1,k\_2Z,U(1) } \[IGood\] As in Sec. 2.1.1, the set of bad operators can be interpreted as quantization-changing operators, sweeping the space of all non-equivalent quantizations. As was proven there, the action of a bad operator takes the wave function out of our Hilbert space ${\cal H}^0(\Gtm)$ and puts it into a different Hilbert space ${\cal H}^{\av}(\Gtm)$ corresponding to a non-equivalent representation of $T_B$ ($=T$) parameterized by $\av$. Thus, we define the new functions (we restrict ourselves to the $\Phi(x_2)$ part of the wave function): \^(x\_2)&& \^[\_1[ \^[R]{}\_[x\_1]{} ]{}+\_2[ \^[R]{}\_[x\_2]{} ]{}]{}\^0(x\_2) = \^[i2n ]{}\^0(x\_2+\_2)\ &=& \_[k=0]{}\^[n-1]{}a\_k \^[i2n ]{} \^0\_k(x\_2+\_2) = \_[k=0]{}\^[n-1]{}a\_k \^\_k(x\_2) \[IWF1\] where \^\_k(x\_2)\^[i2n ]{}\^0\_k(x\_2+\_2) = \^[i2n ]{} \_[qZ]{} \^[i2q(x\_2\^[(k)]{}+\_2)/L\_2]{} \[IWF2\] and the values of $\av$ are different from $\frac{1}{n}\L_{\k}$ (good transformations). To determine the non-equivalent quantizations (i.e. the minimum range of values of the parameters $\a_1$ and $\a_2$ that sweeps the whole set of non-equivalent quantizations) we let the transformations of $T$ act on these new functions and then we determine the quasi-periodicity conditions: (\_1)\^[k\_1]{} \^(x\_2)&=&\^[-i2nk\_1]{} \^(x\_2) \[IQuasiP1\]\ (\_2)\^[k\_2]{} \^(x\_2)&=&\^[i2nk\_2]{} \^(x\_2) \[IQuasiP2\] from which it can be deduced that $\a_1\in [0,\frac{L_1}{n})$ and $\a_2\in [0,\frac{L_2}{n})$. This range of values is associated with the first Brillouin zone of the reciprocal lattice, as can be checked if we define the parameters $\epv\equiv m\w \J\cdot\av$. It is easy to verify that the wave functions $\{\Lambda^{\av}_k(x_2),\,k=0,1,...,n-1\}$ constitute the carrier space (they span ${\cal H}^{\av}(\Gtm)$) for unitary irreducible representations (parameterized by $\av$) of the subgroup of good operators. Under these operators the wave functions transform as: (\_1)\^[k\_1/n]{}\^\_k(x\_2)&=& \^[i2(-)k\_1]{}\^\_k(x\_2)\ (\_2)\^[k\_2/n]{}\^\_k(x\_2)&=& \^[i2k\_2]{}\^\_[k-k\_2[mod]{}n ]{}(x\_2) In a recent paper, [@GotayToro], it is shown that, for the case $n=1$, the symplectic manifold defined by the torus can be fully quantized, i.e. the entire Poisson algebra on the torus can be irreducibily represented by self-adjoint operators acting on a Hilbert space. Here, the same result is obtained for arbitrary integer $n$. Even more, more operators than those associated with classical functions (those of $T$) can be irreducibily represented, namely $(\hat{\eta}_i)^{\frac{k_i}{n}}, k_i\in Z, i=1,2$. To be precise, the $n^{th}$’s roots of the classical functions can be quantized according to our scheme. This is possible thanks to the fact that the representation defined by the equations above is a vector representation, i.e. wave functions are really sections of an associated vector bundle of dimension $n$ over the torus. As in Sec. 2.1.1, we could consider the union of all the Hilbert spaces $\cup_{\av}{\cal H}^{\av}(\Gtm)$. In this Hilbert space, the bad operators ${ {\tilde{X}}^{R}_{\x} }$ are Hermitian and act irreducibily, carrying a unitary irreducible representation of the toral H-W group, turning out to be a generalization, for arbitrary integer $n$, of that called $kq$-representation in Solid State Physics [@Zak], where only the case $n=1$ is considered. [Summarizing]{} the integer case, there is a continuum of non-equivalent quantizations, corresponding to non-equivalent representations of $T$ parameterized by $\av$, giving rise to different quasi-periodic boundary conditions. The value $\av=0$, corresponding to the trivial representation $D^0(e_{\k})=1$ of $\left\{e_{\k};\,\k\in Z\times Z\right\}$, reproduces the standard periodic boundary conditions. The wave functions are (\[IWF1\]-\[IWF2\]) with quasi-periodicity conditions given by (\[IQuasiP1\]-\[IQuasiP2\]) and the subgroup of good operators is (\[IGood\]). It should be stressed the difference between the two representations here obtained. On the one hand, for a fixed $\av$, the Hilbert space ${\cal H}^{\av}(\Gtm)$ carries an irreducible representation of the subgroup of good operators. In this representation the operators ${ {\tilde{X}}^{R}_{\x} }$ do not preserve the Hilbert space; they are bad operators. On the other hand, the union of all the Hilbert spaces $\cup_{\av}{\cal H}^{\av}(\Gtm)$ carries an irreducible representation of the entire toral H-W group, in such a way that the operators ${ {\tilde{X}}^{R}_{\x} }$ are Hermitian and the good operators act in a diagonal form. A brief comment is now in order. Let us consider the discrete (infinite) subgroup generated by $\left\{(\hat{\eta}_1)^{\frac{k_1}{n}}(\hat{\eta}_2)^{\frac{k_2}{n}},\;\; k_1,k_2\in Z \right\}$, which constitutes a principal fibre bundle with base $Z\times Z$ and fibre $Z_n\subset U(1)$. The group algebra of this discrete group can be proven to be (in a suitable basis) an infinite-dimensional trigonometric algebra [@Fairlie]. Since $n$ is an integer, this discrete group has a centre, which can be removed by means of the $T$-function condition. The quotient group is the finite group generated by $\left\{(\hat{\eta}_1)^{\frac{k_1}{n}}(\hat{\eta}_2)^{\frac{k_2}{n}},\right.$ $\;\;$$\left. k_1,k_2=0,...,n-1 \right\}$. This finite group (which can be seen as a finite version of the Heisenberg-Weyl group) constitutes a principal fibre bundle with base $Z_n \times Z_n$ and fibre $Z_n\subset U(1)$, and admits a simple matrix representation given in Ref. [@Weyl] (see also [@Fairlie] and [@Floratos]). The corresponding group algebra is the algebra of $SU(n)\times U(1)$ for $n$ odd or $U(n/2)$ for $n$ even, in a trigonometric basis [@Fairlie]. By means of this representation, the limit $n\rightarrow \infty$ (the “classical" limit) is particularly simple, leading to the algebra of infinitesimal area-preserving diffeomorphisms of a 2D-surface (the torus, in this case). This algebra, referred to as $\w_\infty$ in the literature, is the classical version of a variety of infinite-dimensional algebras called collectively $W_\infty$, of increasing interest nowadays (see [@Shen] for a review). In this sense, the subgroup of good operators $\GH$ can be seen as the quantum version of the area-preserving diffeomorphisms of the torus, thus constituting a realization of the $W_\infty$ algebras on the torus. ### Fractional quantum numbers We now consider the rational case, in which $\frac{m\w L_1L_2}{2\pi\hbar}=\frac{n}{r}$. In this case $T$ has a non-trivial characteristic subgroup, i.e. there are non-trivial elements commuting with the whole group $T$. This is $\GC=\{ r\L_{\k}, \k\in Z^2\}$, and the polarization subgroup, which must contain $\GC$, is $T_p=\GC\cup \{k\L_{\k_p}, k\in Z\}$, where $\k_p$ is a vector the components of which are either relative prime integers, (1,0) or (0,1). This condition is required for maximality of the polarization subgroup, and therefore for the irreducibility of the representation of $T$. The $T$-function condition now reads $\Psi(g_{T_B}g)= {\cal D}(g_{T_B})\Psi(g)$, where $T_B\equiv T_p\cup U(1)$ is the maximal subgroup of compatible constraints that can be applied to the wave function, and ${\cal D}(g_{T_B})$ is a representation of $T_B$ on the complex numbers. For the moment, we shall use the representation ${\cal D}^0(e_{T_p},\z)=\z$, which is trivial for the elements in $T_p$. Later, the non-equivalent representations of $T_B$ will be straightforwardly computed, as in Sec 2.2.1. The $T_B$-function condition on the polarized wave functions (\[H-WtWF\]) is then: \^[m{ y\_2 + (1+2)n\_1n\_2y\_1+ \[(1+)n\_2\^2-n\_1\^2\](Ł\_[r+k\_p]{}) }(Ł\_[r+k\_p]{})]{}& &\ \^[-m(Ł\_[r+k\_p]{})]{} \^0(y\_2+(Ł\_[r+k\_p]{})) &=& \^0(y\_2) $\forall k\in Z$, and $\forall \k\in Z\times Z$. As in the integer case, the only polarization vectors $\n$ consistent with these restrictions are $\n=(1,0)$ and $\n=(0,1)$, and the same for $\k_p$, for which the only possible values are $\k_p=(1,0)$ and $\k_p=(0,1)$. Let us fix the polarization to $\n=(1,0)$ for concreteness (the case $\n=(0,1)$ is completely analogous and in fact leads to an equivalent representation). The two different choices of $\k_p$, perpendicular and parallel to $\n$, lead to [non-equivalent]{} representations \[this is a general feature in AQG: for a given polarization in , different choices of polarization subgroups $T_p$ in $T$ can lead to non-equivalent quantizations, even though the polarization subgroups were equivalent from the point of view of the subgroup $T$ itself (see [@Alfonso])\], both with dimension $n$ and with $\lambda$ restricted to be $\lambda=k/n,\,k\in Z$: [ ]{} i.e. $\k_p=(0,1)$, then the wave function is peaked at the values $y_2=x_2=\frac{k}{n}L_2,\, k\in Z$, satisfies $\Phi^0_\perp(x_2+k_2L_2)=\Phi^0_\perp(x_2)$, and has the form \^0\_(x\_2)=\_[k=0]{}\^[n-1]{}a\_k\_k\^0(x\_2) where $\Lambda_k^0(x_2)$ is defined as in Sec 2.2.1, and the subgroup of good transformations is $\GH=\{\frac{r}{n}\L_{\k},\,\k\in Z\times Z\} \cup\{\frac{k}{n}L_2,\,k\in Z\}$, although only a finite subgroup of them are distinct: \^={(\_1)\^[r]{}, (\_2)\^, k\_1,k\_2=0,...,n-1 } \[FGoodperp\] i.e. $\k_p=(1,0)$, then the wave function is peaked at the values $y_2=x_2=k\frac{r}{n}L_2,\, k=0,1,...n-1$, satisfies $\Phi^0_\parallel(x_2+rk_2L_2)=\Phi^0_\parallel(x_2)$, and has the form \^0\_(x\_2)=\_[k=0]{}\^[n-1]{}a\_k\^[r,0]{}\_k(x\_2) where $\Lambda^{r,0}_k(x_2)\equiv \frac{1}{rL_2}\sum_{q\in Z} \e^{i2\pi q x_2^{r,(k)}/(rL_2)}$, with $x_2^{r,(k)}\equiv x_2-\frac{k}{n}rL_2$, and the subgroup of good transformations is $\GH=\{\frac{r}{n}\L_{\k},\, \k\in Z\times Z\}\cup\{\frac{k}{n}L_1,\,k\in Z\}$. Again, only a finite subgroup of them are distinct: \^={(\_1)\^, (\_2)\^[r]{}, k\_1,k\_2=0,...,n-1 } \[FGoodparallel\] As in Sec. 2.1.1, we can compute the non-equivalent representations by applying the whole set of bad operators to the wave functions. We proceed as in the integer case (Sec 2.2.1) and obtain: [ ]{} the wave functions have the form \^[\_p]{}\_(x\_2)=\_[k=0]{}\^[n-1]{}a\_k\_k\^[\_p]{}(x\_2) \[FWFperp\] with $\Lambda_k^{\av_p}(x_2)\equiv \e^{i2\pi\frac{n}{r} \frac{\a_{p1}}{L_1}\frac{x_2}{L_2}}\Lambda_k^{0}(x_2)$. They satisfy (\_1)\^[rk\_1]{}\^[\_p]{}\_(x\_2)&=& \^[-i2nk\_1]{}\^[\_p]{}\_(x\_2) \[FQuasiPperp1\]\ (\_2)\^[k\_2]{}\^[\_p]{}\_(x\_2)&=& \^[i2nk\_2]{}\^[\_p]{}\_(x\_2) \[FQuasiPperp2\] with $\a_{p1}\in [0,r\frac{L_1}{n}),\,\a_{p2}\in [0,\frac{L_2}{n})$. the wave functions have the form \^[\_p]{}\_(x\_2)=\_[k=0]{}\^[n-1]{}a\_k\^[r,\_p]{}\_k(x\_2) \[FWFparallel\] with $\Lambda_k^{r,\bv_p}(x_2)\equiv \e^{i2\pi\frac{n}{r} \frac{\b_{p1}}{L_1}\frac{x_2}{L_2}}\Lambda_k^{r,0}(x_2)$. They satisfy (\_1)\^[k\_1]{}\^[\_p]{}\_(x\_2)&=& \^[-i2nk\_1]{}\^[\_p]{}\_(x\_2) \[FQuasiPparallel1\]\ (\_2)\^[rk\_2]{}\^[\_p]{}\_(x\_2)&=& \^[i2nk\_2]{}\^[\_p]{}\_(x\_2) \[FQuasiPparallel2\] with $\b_{p1}\in [0,\frac{L_1}{n}),\,\b_{p2}\in [0,r\frac{L_2}{n})$. It is easy to verify that the wave functions $\{\Lambda^{\av_p}_k(x_2),\,k=0,1,...,n-1\}$ and $\{\Lambda^{r,\bv_p}_k(x_2),\,k=0,1,...,n-1\}$ constitute the carrier spaces for unitary irreducible representations (parameterized by $\av$ and $\bv$, respectively) of the subgroup of good operators. Under these operators they transform as: [ ]{} (\_1)\^[rk\_1/n]{}\^[\_p]{}\_k(x\_2)&=& \^[i2(-)]{}\^[\_p]{}\_k(x\_2)\ (\_2)\^[k\_2/n]{}\^[\_p]{}\_k(x\_2)&=& \^[i2k\_2]{}\^[\_p]{}\_[k-k\_2[mod]{}n ]{}(x\_2) (\_1)\^[k\_1/n]{}\^[r,\_p]{}\_k(x\_2)&=& \^[i2(-)]{}\^[r,\_p]{}\_k(x\_2)\ (\_2)\^[rk\_2/n]{}\^[r,\_p]{}\_k(x\_2)&=& \^[i2k\_2]{}\^[r,\_p]{}\_[k-k\_2[mod]{}n ]{}(x\_2) It should be noted that although $\frac{m\w L_1L_2}{2\pi\hbar}=\frac{n}{r}$, the dimension of the representations is $n$, and $\lambda=k/n,\,k\in Z$, as in the integer case (even more, in the case $\L_{\k_p}\perp \n$ the wave functions coincide); the difference is found in the subgroups of good operators, which, although isomorphic, differ in the specific values of the transformations. This representation can be reinterpreted as mimicking a torus $r$ times greater in one direction (determined by the orthogonal vector to $\k_p$), i.e., the area of the effective torus is $rL_1L_2$, and therefore $\frac{m\w (r L_1L_2)}{2\pi\hbar}=n$. Thus, the same results as in the integer case now apply, although changing $L_2$ by $rL_2$ if $\k_p=(1,0)$ or $L_1$ by $rL_1$ if $\k_p=(0,1)$. [Summarizing]{} the fractional case, there are two continua of non-equivalent quantizations, according to the choices $\L_{\k_p}\perp \n$ and $\L_{\k_p}\parallel \n$, parameterized by $\av_p$ and $\bv_p$, respectively. The wave functions are given by (\[FWFperp\]) and (\[FWFparallel\]), satisfying quasi-periodicity conditions given by (\[FQuasiPperp1\]-\[FQuasiPperp2\]) and (\[FQuasiPparallel1\]-\[FQuasiPparallel2\]), respectively. The subgroups of good operators are given by (\[FGoodperp\]) and (\[FGoodparallel\]), respectively. If we act on the wave functions with the bad operators of $T$ (i.e. those operators of $T$ which are not in $T_B$) the resulting wave functions lie in a different Hilbert space belonging to a different quantization. However, as these operators are finite and their $r^{th}$ power are good operators, these new wave functions transform among each other under the action of the subgroup $T_{\rm bad}$, defined as the set of bad operators of $T$ and the identity. Therefore, constructing the vector space spanned by these $r$ functions ($T_{\rm bad}$ has $r$ elements), we obtain an $r$-dimensional, unitary irreducible representation of the group $T$ as a whole, including the bad operators. Explicitly: [ ]{} we define \^[\_p]{}\_[k,j]{}(x\_2)(\_1)\^[j]{}\^[\_p]{}\_k(x\_2)= \^[i2j]{}\^[\_p]{}\_k(x\_2) for $j=0,1,...,r-1$, where they satisfy: (\_1)\^[j’]{}\^[\_p]{}\_[k,j]{}(x\_2)= \^[-i2n(j+j’[div]{} r)]{} \^[\_p]{}\_[k,j+j’[mod]{} r]{}(x\_2) for $j,j'=0,1,...,r-1$. we define \^[r,\_p]{}\_[k,j]{}(x\_2)(\_2)\^[j]{}\^[r,\_p]{}\_k(x\_2)= \^[r,\_p]{}\_k(x\_2+jL\_2)\ for $j=0,1,...,r-1$, satisfying: (\_2 )\^[j’]{}\^[r,\_p]{}\_[k,j]{}(x\_2)= \^[i2n(j+j’ [div]{} r)]{} \^[r,\_p]{}\_[k,j+j’[mod]{} r]{}(x\_2) for $j,j'=0,1,...,r-1$. This construction can be viewed as the $r$-dimensional vector bundle associated with the principal bundle , which has structure group $T$. The $r$-component wave functions are sections of this associated vector bundle. As stated before, AQG generalizes Geometric Quantization in some respects, in particular in that which concerns (topologic) quantum numbers. The fractional value $\frac{m\w L_1L_2}{2\pi\hbar}=\frac{n}{r}$ generalizes the integer class of the standard symplectic form (the Chern class of the line bundle). The geometric quantization of a symplectic manifold with “fractional class" $\frac{n}{r}$ would have led to r-valued wave functions (as opposed to single-valued). Eventually, this trouble could have been circumvented by replacing the usual line bundle by a complex vector bundle $E$ of rank $r$ and Chern class $n$, as constructed before. The comments at the end of Sec. 2.2.2 concernig the generalized $kq$-representation can be translated to the $r$-bundle structure associated with the fractional case. ### Irrational case Finally, and for the sake of thoroughness, let us briefly comment on the case in which $\rho\equiv\frac{m\w L_1L_2}{2\pi\hbar}$ is an irrational number. In this case the characteristic group is trivial, and $T_B=T_p\cup U(1)$, with $T_p=\{k\L_{\k_p}, k\in Z\}$ only. As before, it can be proven that the only possible polarization vectors are $\n=(1,0)$ and $\n=(0,1)$. Moreover, the only consistent choice of polarizations $T_p$ in $T$ are also $\k_p=(1,0)$ and $\k_p=(0,1)$. No restriction for $\lambda$ appears in this case, and the structure of $T_B$-function condition closely resembles that of the case of the cylindrical H-W group: the wave functions are either peaked at an infinite series of equally spaced values of $y_2$ if $\k_p\parallel \n$ (as in the momentum space representation in the cylindrical H-W group), or quasi-periodic if $\k_p\perp \n$ (as in the configuration space representation in the cylindrical H-W group). In both cases the non-equivalent representations are labelled by $\ep\in[0,\frac{2\pi\hbar}{\|\L_{\k_p}\|})$. The representations are therefore infinite dimensional, and the subgroup of good operators is given by $\GH=\{\frac{1}{\rho}\L_{\k}, \k\in Z\times Z\}\cup \{\alpha\L_{\k_p}, \alpha\in R\}$. Consequently, besides the discrete transformations in $\x=\frac{1}{\rho}\L_{\k}$, the infinitesimal operator $\L_{\k_p}\cdot{ {\tilde{X}}^{R}_{\x} }$ is also a good operator, that is, arbitrary translations in the direction of $\L_{\k_p}$ are good transformations. Note that, $\rho$ being an irrational number, $\frac{1}{\rho}\L_{\k}$ never reaches a point of the lattice defined by $\L_{\k}$, although it fills the corresponding torus densely when varying $\k\in Z\times Z$. Therefore, in this case, the subgroups $x_1=0$ and $x_2=0$ (the classical circumferences), are represented faithfully, as in the case of the cylindrical H-W group, but the rest of the group is not faithfully represented, nor are even the points of the lattice (the group $T$). In particular, for the infinite-order operators $\hat{\eta}_1,\hat{\eta}_2$, defined as in Sec 2.1.1 for the directions $x_1,x_2$, only one is a good operator (the one in the direction of $\L_{\k_p}$), the other being a bad operator. Consequently, we cannot represent the toral H-W group faithfully for irrational values of $\rho$. Free Galilean particle on the circumference =========================================== Let us apply the results obtained in the last section to the simple example of the free particle moving on the circumference. We can study this problem easily by simply adding the temporal evolution to the results obtained in Sec. 2.1 (for the group law, vector fields, polarizations, Schrödinger equation, etc., see [@Position] and references therein), without affecting the main conclusions of that section. The main new features are the introduction of a new operator $\E$ associated with the temporal evolution and the fact that, by using the Schrödinger equation, this operator can be written in terms of the momentum operator as $\frac{1}{2m}\P^2$. Since $\P$ is a good operator, $\E$ proves also to be a good operator. A common set of eigenfunctions is given by \^\_n(x,t)&=&\^[-(+n)\^2 t]{} \^[(+n)x]{}\ \^\_n&=& (+n)\^2\^\_n\ ¶\^\_n&=& (+n)\^\_n where $n\in Z$. Note that for $\ep=0$ the states $n$ and $-n$ have the same energy, which means that all the energy eigenstates except for the vacuum are degenerate. For any other value of $\ep$, the states $n$ and $-(n+2\ep\frac{L}{2\pi\hbar})$ have the same energy, but $-(n+2\ep\frac{L}{2\pi\hbar})$ is an eigenstate only if $2\ep\frac{L}{2\pi\hbar}\in Z$, i.e. $\ep\frac{L}{2\pi\hbar}$ is integer or half-integer. This means that, in general, there is no degeneracy for any value of $\ep$ except for the integer values, in which case all the eigenstates are doubly degenerate except for the vacuum, and half-integers, for which all the eigenstates, including the vacuum, are doubly degenerate. The phenomenon of degenerate ground state in this simple model parallels $\theta$-vacuum phenomenon in Yang-Mills field theories [@Asorey]. The feature of non-equivalent quantizations can be reproduced (in an equivalent way, indeed) by the introduction of an extra coboundary in (\[H-WcGLaw\]) (more precisely, in its counterpart when the temporal evolution is added; see [@Position]) generated by the function $\ep x$, i.e. a multiplicative factor of the form $\e^{\idhbar \ep \frac{p'}{m}t}$ in the $\z\in U(1)$ composition law \[We remind that $x''=x'+x+\frac{p'}{m}t$ is the composition law for $x$ when the temporal evolution is added\]. In the case of the free Galilean particle on the real line, the only consequence of this term is the appearance of a total derivative in the quantization 1-form $\Theta$ (or, what is the same, in the Lagrangian), leading thus to equivalent (classical and quantum) theories, as expected from the fact that $\ep \frac{p'}{m}t$ is a coboundary. The situation is quite different when the particle is on the circumference: the generating function $\ep x$, or better $\e^{\idhbar\ep x}$, is not single-valuate on the circumference unless $\ep=\frac{2\pi\hbar}{L}k,\,k\in Z$. As a consequence, two cocycles differing in a coboundary generated by $\ep x$ (and therefore leading to equivalent theories on the real line) lead to non-equivalent theories on the circumference if $\ep\neq\frac{2\pi\hbar}{L}k,\,k\in Z$. This process of creation of non-trivial cohomology closely resembles the appearance of cohomology under the process of group contraction, as in the case of the Poincaré group, in which a certain class of coboundaries (generated by a linear function in time) become true cocycles in the $c\rightarrow\infty$ limit since their generating function goes to infinity in this limit. Another interesting way of interpreting the feature of non-equivalent quantizations parameterized by $\ep$, at least in the case of charged particles, is as an Aharonov-Bohm-like effect. The different quantizations can be carried out physically by producing (externally, with the help of a solenoid) a magnetic flux $\Phi$ through the circumference, in a way that the particle does not feel the magnetic field, but rather the vector potential only. Under these circumstances, the effect of the vector potential is the same as that of a boost, leading to non-equivalent quantizations depending on the flux through the circumference, in such a way that $\ep=e\Phi/c$. An interesting physical application is that of a superconducting ring threaded by a magnetic flux, where by Meissner effect the magnetic flux is pulled out of the interior region of the superconducting ring, and therefore the magnetic field is effectively zero and only the vector potential is relevant (Aharonov-Bohm effect). If the flux is \[in this case the effective electric charge is $e^=2e$ because electrons form Cooper pairs\] $ k\Phi_0,\,k\in Z$, where $\Phi_0\equiv hc/e^$ is the quantum unit of flux, there is no net current in the superconducting ring, but for any other value of the flux there is a net current which has the form given in Figure 1. =9.0cm Note that for half-integer values of $k$, the net current has no definite sign, as a consequence, precisely, of the double degeneracy of states, in such a way that states with opposite signs of velocity have the same energy and therefore there is no energy cost to pass from one to the other. Failure of the Ehrenfest theorem -------------------------------- As mentioned in the introduction, the most common problem appearing in systems with topologically non-trivial configuration space is the failure of Ehrenfest theorem for certain operators (“anomalous" operators) [@Esteve]. Ehrenfest theorem asserts that the expectation values of quantum operators follow classical equations of motion: <> = \[Ehrenfest\] In the framework of AQG, this is a natural consequence of the appearance of bad operators, those which do not preserve the Hilbert space of wave functions that verify the $T$-function condition. In Ref. [@Esteve] it is claimed that when the operator $\hat{A}$ does not keep invariant the domain of $\hat{H}$, then an extra term appears in the r.h.s. of (\[Ehrenfest\]), which is interpreted as an anomaly. In the language of AQG, we would say that $\hat{A}$ is a bad operator, so that neither the left-hand side nor the right-hand side of (\[Ehrenfest\]) would make sense, since the operator $\hat{A}$ takes the wave function off the Hilbert space where the Hamiltonian $\hat{H}$ is self-adjoint (of course $\hat{H}$ is a good operator; otherwise the temporal evolution would take the physical states off the Hilbert space, and the system would have no physical meaning). The appearance of the “anomalous" term violating the Ehrenfest theorem is a consequence of this fact. Returning to the free Galilean particle on the circumference, the Ehrenfest theorem will fail for the position operator, which is a bad operator and therefore Eq. (\[Ehrenfest\]) makes no sense in this regard. In conclusion, whenever there are bad operators in the theory, the Ehrenfest theorem will fail for each of these operators and, in general, any expectation value involving these operators will be ill-defined, giving extra terms that can eventually be interpreted as topologic anomalies. Charged particle in a homogeneous magnetic field on the torus ============================================================= Now, we shall consider the most interesting problem of a charged particle moving on a torus in the presence of a homogeneous magnetic field. This problem is related to the Schwinger model [@Manton], and has important applications in the Quantum Hall effect [@Hall1; @Hall2; @Thouless]. The magnetic field is perpendicular to the torus surface, and the total flux is quantized (as we shall see), much in the same manner the Dirac monopole charge is quantized [@Wu-Yang]. The actual connection of this system with Quantum Hall Effect is based on the fact that the wave function of the complete system factorizes in a relative-coordinate dependent term (which includes interactions) and a centre of mass dependent term, which behaves essentially as a particle in a transverse homogeneous magnetic field, and on the effective topology of the experimental device in the latter system; the topology of the (semiconductor) strip along with the current and voltage leads is that of a punctured torus [@Thouless]. Firstly, we shall study the planar case, i.e. the charged particle on the plane, to clarify the meaning of the different magnitudes appearing in the problem, and to obtain a proper parameterization of the system. Charged particle in a homogeneous magnetic field ------------------------------------------------ The movement of a charged particle in a homogeneous magnetic field can be factorized into a 2-dimensional problem (on the plane normal to the magnetic field) times a free movement in the direction of the magnetic field. Thus, we restrict ourselves to a 2-dimensional system characterized by a non-zero commutator between the translation generators, $[{ {\tilde{X}}^{L}_{x^1} },{ {\tilde{X}}^{L}_{x^2} }]=im\wc/\hbar$, where $\wc$ is the cyclotron frequency, $\omega_c=\frac{q{\cal H}}{mc}$, ${\cal H}$ the magnetic field strength and $q$ the particle electric charge [@Landau; @Cohen-Tannoudji]. We have to build up a group law for this system, which must be a deformation of the Galilean group law (in two dimensions) due to the non-zero commutator between the translation generators. In fact, the Galilean group does not admit any central extension giving rise to $[{ {\tilde{X}}^{L}_{x^1} },{ {\tilde{X}}^{L}_{x^2} }]=im\wc/\hbar$, and a deformation of the non-extended algebra is required: $\left[X^L_{t},X^L_{\x}\right] = \Bm \J\cdot X^L_{\x}$. We then arrive at the following Lie algebra as the quantum symmetry for our system: &=& [ \^[L]{}\_ ]{}\ &=&- [ \^[L]{}\_ ]{}\ &=&\ &=& \_[ij]{} A group law for this centrally extended Lie algebra becomes: t” &=& t’ + t\ ”&=& + ’ + (+ )’\ ”&=& ’+ \[BGLaw\]\ ”&=& ’\^[(g’,g)]{} where the cocycle is given by: (g’,g) = {m’- ’+ ’+ ’(- )} \[Bcocycle\] The $2\times 2$ matrices are given by $\M \equiv \cos\Bmt\, \I - \sin\Bmt\, \J\,$, $\N\equiv \sin\Bmt\, \I + \cos\Bmt\, \J$, $\M$ and $\N$ being orthogonal, and $\J_{ij}\equiv\epsilon_{ij}$, $\epsilon_{12}=1$. We have not taken into account the rotations, since they do not play any dynamical rôle, although they are of interest in that, when considered on the torus, they represent a very simple example of a local (in the strict mathematical sense) symmetry of the equation of motion which cannot be realized globally. The left and right invariants vector fields are easily deduced from the group law: [ \^[L]{}\_[t]{} ]{} & = & [ ]{} + [ ]{} -[ ]{}\ [ \^[L]{}\_ ]{} & = & [ ]{} - \[XLB\]\ [ \^[L]{}\_ ]{} & = & [ ]{} +\ [ \^[L]{}\_ ]{} & = & i[ ]{}\ & &\ & &\ [ \^[R]{}\_[t]{} ]{} & = & [ ]{}\ [ \^[R]{}\_ ]{} & = & [ ]{} + \[XRB\]\ [ \^[R]{}\_ ]{} & = & [ ]{} + (-)[ ]{} -\ [ \^[R]{}\_ ]{} & = & i[ ]{} and from (\[XLB\]) the quantization 1-form is computed: = - dt + the characteristic module of which is $\Gthetam=<{ {\tilde{X}}^{L}_{t} }>$. From this, the classical equations of motion are written: &=&\ &=&+ where $\vec{P}$ and $\ro$ are arbitrary constant vectors, parameterizing the (classical) solution manifold. With the aid of the constant $\wc$, we may introduce $\R\equiv\frac{1}{m\wc}\J\cdot\vec{P}$, so that the second line of the equation above reads $\x=\Mi\cdot\ro + \R$, i.e. the classical trajectories are circumferences centred at $\R$, with radius $\|\ro\|$. The Noether invariants, in terms of the constants $\ro$ and $\R$, are: [ i\_[[ \^[R]{}\_[t]{} ]{}]{}]{} &=& \^2H\ [ i\_[[ \^[R]{}\_ ]{}]{}]{} &=& m\ [ i\_[[ \^[R]{}\_ ]{}]{}]{} &=& -(+ ) where $H$ is the classical energy of the system. It should be noted that the energy depends only on the radius $\|\ro\|$ of the circumference, and not on the position $\R$ of its centre, as corresponds to a system with translational invariance \[The system possesses translational invariance in the more conventional sense (the magnetic field is homogeneous) although the translation generator ${ {\tilde{X}}^{R}_{\x} }$ does not commute with the Hamiltonian ${ {\tilde{X}}^{R}_{t} }$. In fact, as we shall see later, there exists a translation generator in the Lie algebra (the magnetic translations) which commutes with the Hamiltonian generator\]. To obtain the representation in configuration space, we need to impose polarization conditions similar to those of the Galilean case [@Position]: \^[HO]{}=<[ \^[L]{}\_ ]{},[ \^[L]{}\_[t]{} ]{}-([ \^[L]{}\_ ]{})\^2 > Solving the polarization equations we obtain for the wave functions the general form: = \^[-]{}(,t) \[BWF\] where $\Phi(\x,t)$ satisfies the Schrödinger equation i[ ]{}= {- \^2 + i + \^2} \[BSEQN\] The quantum operators are: & = & i[ ]{}= \^[- ]{}(,t)\ & = & \^[- ]{} (,t) \[BOConf\]\ & = & \^[- ]{} (,t) Instead of proceeding further and solving the Schrödinger equation explicitly, we shall perform a change of variables which will clarify the meaning of the different magnitudes entering the theory and which will make facilitate the accomplishment of AQG in the next subsection. If we define $\r\equiv\Mi\cdot\ro$, we can easily rewrite the group law (\[BGLaw\]) and (\[Bcocycle\]) in terms of $\r$ and $\R$: t” &=& t’ + t\ ” &=& + ’\ ” &=& ’ +\ ” &=& ’\^[m]{} where we have added the coboundary generated by $-m\wc(\medio+\lambda)R_1R_2$ to accommodate the cocycle, in its $\R$-dependent term, to the expression of Sec. 2.2 (except for a global minus sign). From this group law we can compute again the left- and right-invariant vector fields: [ \^[L]{}\_[t]{} ]{} & = & [ ]{} -[ ]{}\ [ \^[L]{}\_ ]{} & = & [ ]{} -\ [ \^[L]{}\_[R\_1]{} ]{} & = & [ ]{} - mR\_2\ [ \^[L]{}\_[R\_2]{} ]{} & = & [ ]{} - mR\_1\ [ \^[L]{}\_ ]{} & = & i[ ]{} [ \^[R]{}\_[t]{} ]{} & = & [ ]{}\ [ \^[R]{}\_ ]{} & = & [ ]{} + \[XRBn\]\ [ \^[R]{}\_[R\_1]{} ]{} & = & [ ]{} - mR\_2\ [ \^[R]{}\_[R\_2]{} ]{} & = & [ ]{} - mR\_1\ [ \^[R]{}\_ ]{} & = & i[ ]{} \[XLRBn\] and the commutation relations are now: &=& [ \^[L]{}\_ ]{}\ &=& 0\ &=& \_[ij]{}\[CommB\]\ &=& - \_[ij]{}\ &=& 0 A glance at the algebra (\[CommB\]) reveals that it is the central extension of the direct sum of the harmonic oscillator algebra and the Heisenberg algebra. Consequently, the wave function factorizes into a harmonic oscillator wave function (depending on $t$ and $\r$) times a function of $\R$, and the energy spectrum coincides with that of the harmonic oscillator, the degeneracy being infinite due to the Heisenberg-Weyl symmetry, which in the plane has only infinite-dimensional unitary irreducible representations. We are interested in a configuration-space representation, so that a second-order polarization is needed. This is found to be: \^[HO]{}=<[ \^[L]{}\_[p]{} ]{},[ \^[L]{}\_[t]{} ]{}-([ \^[L]{}\_ ]{})\^2, [ \^[L]{}\_ ]{},[ \^[L]{}\_ ]{} > where $\n$ and $\np$ are arbitrary unit vectors. They can be chosen to be $(1,0)$ or $(0,1)$, for instance. Imposing these polarization conditions to the wave functions, we obtain the general form: = \^[-m]{} (y\_2) \^[\_2\_1]{} (\_2,t)\[BWFr-R\] where $y_1\equiv \n\cdot\R,\, y_2\equiv \n\cdot\J\cdot\R,\, \kappa_1\equiv\np\cdot\r,\,\kappa_2\equiv\np\cdot\J\cdot\r,\,\,\Phi(y_2)$ is an arbitrary function and $\Theta(\kappa_2,t)$ satisfies the Schrödinger equation: i[ ]{}(\_2,t)=(\_2,t) This is nothing other than the Schrödinger equation for the harmonic oscillator, so that the solutions are (\_2,t)=\_n A\_n \^[-i(n+)t]{} \^[-\_2\^2]{}H\_n(\_2) where $H_n$ are the Hermite polynomials. Since the wave functions factorize, the operators ${ {\tilde{X}}^{R}_{\R} }$ will act only on the $\R$-dependent part of it, having the same expressions as in (\[H-WtOp\]) (changing there $\x$ to $\R$), and the operators ${ {\tilde{X}}^{R}_{\x} }$ will act only on the $(\r,t)$-dependent part, with the expressions: [ \^[R]{}\_[\_1]{} ]{}[ \^[R]{}\_ ]{} (\_2,t)&=& (\_2,t)\ [ \^[R]{}\_[\_2]{} ]{}[ \^[R]{}\_ ]{} (\_2,t)&=& (\_2,t) once the (irrelevant) phase factors have been factorized out. Using the dual transformation to the one taking $(\x,\p)$ to $(\r,\R)$, we obtain the expression of the operators $\Xv$ and $\Pv$ in terms of ${ {\tilde{X}}^{R}_{\r} }$ and ${ {\tilde{X}}^{R}_{\R} }$: [ \^[R]{}\_ ]{} &=& [ \^[R]{}\_ ]{}\ -[ \^[R]{}\_ ]{} &=& ([ \^[R]{}\_ ]{}-[ \^[R]{}\_ ]{}) In addition, by $\Tv$ we denote the operator $-i\hbar { {\tilde{X}}^{R}_{\R} }=\Pv-m\wc\J\cdot\Xv$. It can be easily deduced that $\Pv$ has the physical meaning of a linear momentum (mass times velocity), which we shall simply call momentum, while $\Tv$ is a momentum commuting with the Hamiltonian, generally called magnetic translations, and this is associated with the coordinate $\R$ of the centre of the circumferences. We can still define another momentum in the theory, the canonical momentum, as $\Piv\equiv -\frac{i\hbar}{2}\left({ {\tilde{X}}^{R}_{\r} }+{ {\tilde{X}}^{R}_{\R} }\right)$, which has the particularity that its components mutually commute, and, as can be easily checked, is a proper translation generator: it is written (for $t=0$) as $\vec{\nabla}_{\x}$ when acting on $\Phi(x,t)$ in (\[BWF\]) at $t=0$. Its explicit expression and that of $\Tv$ on $\Phi(x,t)$ are: & = & \^[- ]{} (,t)\ & = & \^[- ]{} (,t) The rôle of the different momenta can be clarified by introducing the vector potential operator in the usual form, $\Av\equiv-\frac{m\wc}{2}\J\cdot\Xv = \frac{i\hbar}{2}\left({ {\tilde{X}}^{R}_{\r} }-{ {\tilde{X}}^{R}_{\R} }\right)$. Then, the canonical momentum is rewritten $\Piv=\Pv+\Av$, and $\Tv=\Pv+2\Av=\Piv+\Av$. Then it is easy to verify that $\E=\frac{1}{2m}\Pv^2= \frac{1}{2m}\left(\Piv-\Av\right)^2= \frac{1}{2m}\left(\Tv-2\Av\right)^2$. Charged particle in a homogeneous magnetic field in the plane with periodic boundary conditions ----------------------------------------------------------------------------------------------- Before imposing the periodic boundary conditions which define the torus, as in Sec. 2.2, we must determine how these boundary conditions affect each of the coordinates. Clearly, $\x$ will be affected by the boundary conditions, but it is not clear what happens to $\p$. Let us return to $\r$ and $\R$ coordinates, where $\R$ is the (absolute) position of the centre of the circumference (the classical trajectory) and $\r$ is the (relative) position of the particle with respect to the centre of the circumference, i.e. $\r=\x-\R$. Therefore, $\R$ will be subject to periodic boundary conditions (the same that $\x$) while $\r$ will not, being a relative coordinate (since the classical energy $H$ is a function of $\ro{}^2=\r{\:}^2$, periodic boundary conditions for $\r$ would imply an upper bound to the energy spectrum, and even more, a periodic energy spectrum). This makes $\r$ and $\R$ coordinates more appropriate to describe the system with periodic boundary conditions. Now we are ready to apply the results of Sec. 2.2, having reduced the problem, roughly speaking, to the study of an harmonic oscillator times a Heisenberg-Weyl group on the torus, the latter being parameterized by $\R$. Regarding the H-W subgroup, we can apply the results of Sec. 2.2. We also consider the two cases [i)]{} and [ii)]{}, corresponding to $T$ being a trivial or non-trivial principal fibre bundle, respectively. Let us consider first the case [i)]{} (Sec. 2.2.1), which is the more conventional one. The actual condition to be satisfied is =nZ \[BQuant\] which implies, as already anticipated, a quantization of the magnetic flux through the torus surface, in the same manner as in the Dirac monopole case. If this flux were produced by a monopole charge, the quantization of the magnetic charge would follow. This kind of quantization condition guarantees, for instance, that the Wilson loop variables in gauge theories are single-valued [@Manton]. The wave functions turn out to be (\[BWFr-R\]), where $\Phi(y_2)$ is subject to exactly the same restrictions as in Sec. 2.2.1, thus leading to the expression (for $\n=(1,0)$): \^(R\_2)=\_[k=0]{}\^[n-1]{}a\_k \^\_k(R\_2) where $\av$ is defined as before. The wave function is therefore peaked at $R_2=\a_2+\frac{k}{n}L_2,\, k\in Z$ ($R_1=\a_2+\frac{k}{n}L_1,\, k\in Z$ for $\n=(0,1)$). The subgroup $\GH$ of good transformations (the ones that preserve the structure of the wave function) is the subgroup of with the parameters $\R$ restricted to be $\R=\frac{1}{n}\L_{\k}$. The quantum operators $\E$ and ${ {\tilde{X}}^{R}_{\r} }$ are good operators (since the harmonic oscillator part is not subject to constraints), while the operator ${ {\tilde{X}}^{R}_{\R} }$ is a bad operator. If we analyse these results in terms of the operators $\Xv,\, \Pv,\, \Tv\,$ and $\Piv$ by means of the expressions given in Sec. 4.1, we conclude that the operator $\Pv$ is a good operator, while $\Xv,\, \Tv\,$ and $\Piv$ are not. Consequently, the momentum (or velocity) of the particle is a measurable quantity, but the position, the canonical momentum and the magnetic translations are not observables. The vector potential operator is of course also a bad operator. For all the bad operators, their finite expressions (counterparts of $\hat{\eta}$ of Sec 2.1.1) can be nevertheless constructed, since all these expressions are good operators. For the case $ii)$, only the fractional case $a)$ is physically meaningful. Now the wave function is defined on a torus $r$ times greater in one direction [@Niu], or, what is the same, it is a vector-valued (with $r$ components) function; or, in FQHE terminology, the centre of mass function corresponding to the vacuum is degenerate, $r$ being the degeneration. In this case the Hall conductivity is associated with the quotient $\frac{n}{r}$ of the Chern class of the associated determinant bundle by the rank of the vector bundle [@determinante]. This result lends support to the idea that Fractional Quantum Hall Effect is always associated with multiple-valued wave functions, i.e. degenerate vacua. Acknowledgement {#acknowledgement .unnumbered} =============== We wish to thank Mark Gotay for valuable discussions. [99]{} Aldaya, V., Azcárraga, J.A.: J. Math. Phys. [23]{}, 1297 (1982) Aldaya, V., Navarro-Salas, J., Ramirez, A.: Comm. Math. Phys. [121]{}, 541 (1989) Aldaya, V., Navarro-Salas, J., Bisquert, J., Loll, R.: J. Math. Phys. [33]{}, 3087 (1992) Aldaya, V., Bisquert, J., Guerrero, J., Navarro-Salas, J.: J. Phys. [A26]{}, 5375 (1993) Ashtekar, A.: [ Mathematical problems of non-perturbative Quantum General Relativity]{}, Lectures delivered at the 1992 Les Houches summer school on Gravitation and Cosmology (1992) Asorey, M., Esteve, J.G., Pacheco, A.F.: Phys. Rev. [D27]{}, 1852 (1983) Cohen-Tannoudji, C., Diu, B., Laloë, F.: [Mécanique quantique]{}, Tome I, Hermann (1977) Esteve, J.G.: Phys. Rev. [D34]{}, 674 (1986) Fairlie, D.B., Fletcher, P., Zachos, C.K.: J. Math. Phys. [31]{}, 1088 (1990) Floratos, E.G.: Phys. Lett. [B 228]{}, 335 (1989) Gotay, M.J., Grundling, H., Hurst, C.A.: [A Groenewold-Van Hove Theorem for $S^2$]{} (dg-ga/9502008) (to appear in Transactions of the AMS) (1995) Gotay, M.J., [On a Full Quantization of the Torus]{}, dg-ga/9507005 (1995) Isham, C.J., Linden, N.: Class. Quant. Grav. [5]{}, 71 (1988) Klitzing, K. v., Dorda, G., Pepper, M.: Phys. Rev. Lett. [45]{}, 494 (1980); Landau, L.D., Lifshitz, E.M.: [Quantum Mechanics, Nonrelativistic Theory]{}, Pergamon Press, Oxford (1965) Landsman, N.P.: Let. Math. Phys. [20]{}, 11 (1990) Laughlin, R.B.: Phys. Rev. Lett. [54]{}, 1395 (1983); Manton, N.S.: Ann. of Phys. [159]{}, 220 (1985) Niu, Q., Thouless, D.J., Wu, Y.S.: Phys. Rev. [B31]{}, 3372 (1985) Ohnuki, Y., Kitakado, S.: J. Math. Phys. [34]{}, 2827 (1993) Shen, X.: Int. J. Mod. Phys. [A 7]{}, 6953 (1992) Thouless, D.J.: J. Math. Phys. [35]{}, 5362 (1994) Varnhagen, R.: [Topology and Fractional Quantum Hall Effect]{}, preprint BONN-Th-94-22, hep-th/9411040 (1994). Weyl, H.: [The Theory of Groups and Quantum Mechanics]{}, Dover, New York (1931) Woodhouse, N.: [Geometric Quantization]{}, Oxford University Press, Oxford (1980) Wu, T.T., Yang, C.N.: Phys. Rev. [D 12]{}, 3845 (1975) Zak, J.: Phys. Rev. [168*]{}, 686 (1968)	{ "pile_set_name": "ArXiv" }
--- abstract: 'We study operation of a new device, the superconducting differential double contour interferometer (DDCI), in application for the ultra sensitive detection of magnetic flux and for digital read out of the state of the superconducting flux qubit. DDCI consists of two superconducting contours weakly coupled by Josephson Junctions. In such a device a change of the critical current and the voltage happens in a step-like manner when the angular momentum quantum number changes in one of the two contours. The DDCI may outperform traditional Superconducting Quantum Interference Devices when the change of the quantum number occurs in a narrow magnetic field region near the half of the flux quantum due to thermal fluctuations, quantum fluctuations, or the switching a loop segment in the normal state for a while by short pulse of an external current.' author: - 'V.L. Gurtovoi$^{1,2}$, V.N. Antonov$^{2,3}$, A.V. Nikulov$^{1}$, R. Shaikhaidarov$^{3}$, and V.A. Tulin$^{1}$' title: A development of superconducting differential double contour interferometer --- It is about 50 years since first measurements of Superconducting Quantum Interference Device (SQUID) [@SQUID1964], which is one of the most sensitive detectors of magnetic field [@SQUID1977]. The classical dc SQUID consists of two Josephson junctions mounted in a superconducting loop with a square $S$. The critical current $I _{c}$ of the two junctions oscillates as a function of the external magnetic flux $\Phi = BS$ threading the loop, with the period equal the flux quantum $\Phi _{0} = 2\pi \hbar /2e \approx 20.7 \ Oe \ \mu m^{2}$ [@SQUID1977]. The voltage $V$ across the junctions at constant bias current $I$ is also periodic in $\Phi $. The classical dc SQUID is used as the detector of magnetic flux $\Phi $ or magnetic field $B = \Phi /S$ due to the periodic dependence $V(\Phi )$ at $I = const $. Therefore the maximum value $(\partial V/\partial \Phi ) _{I}$ is important parameter for the dc SQUID sensitivity [@SQUID1977]. The amplitude of the voltage oscillations $\Delta V = R _{d}\Delta I _{c}$ cannot exceed the value $R _{d}I _{c} < \Delta /e$, where $ R _{d}$ is the dynamical resistance of the Josephson junctions, $\Delta $ is the energy gap of the superconductor and $e$ is the electron charge [@Barone]. According to the relation $I _{c} = 2 I _{c,j}\|\cos \pi \Phi \|$ [@Barone] valid in the case of week screening $\beta = 2L I _{c,j}/\Phi _{0} \ll 1$ [@SQUID1977] the critical current of the dc SQUID changes in the interval $\Delta \Phi = \Phi _{0}/2$. Consequently the maximum value $(\partial V/\partial \Phi ) _{I}$ of the classical dc SQUID cannot exceed $2\Delta /e\Phi _{0}$. The real value $(V/\Phi ) _{I} \approx 2 \ \mu V/\Phi _{0}$ [@SQUID1977] of a typical dc SQUID is substantially smaller than the maximum value. In this work we explore an alternative device for the measurement of the weak magnetic filed, a differential double contour interferometer (DDCI). The idea of this new device arose thanks to the experiment made in [@Zhilyaev2000] and its explanation [@Zhilyaev2001]. Higher sensitivity of DDCI compared to the conventional SQUID is provided by strong discreteness of the energy spectrum of the continuous superconducting loop [@PLA2012QF]. According to the conventional theory [@Tink75] the total energy of the persistent current $$I_{p} = \frac{n\Phi_{0} - \Phi }{L_{k}} \eqno{(1)}$$ in a loop with small cross section $s \ll \lambda _{L}^{2}(T)$ is determined mainly by the kinetic energy [@QuSMF2016]: $$E _{t} \approx E _{k} = \frac{L_{k}I_{p}^{2}}{2}= \frac{\Phi _{0}^{2}}{2L_{k}}(n - \frac{\Phi }{\Phi _{0}})^{2} \eqno {(2)}$$ Here $L _{k} = ml/ q^{2}n_{s}s = (\lambda _{L}^{2}/s) \mu _{0}l \approx (\lambda _{L}^{2}/s)L$ is the kinetic inductance of the loop of side $l$; $L \approx \mu _{0}l$ is the magnetic inductance; $s$ is the cross section of superconducting wires; $n _{s}$ is the density of the Cooper pairs; $\lambda _{L} = (m/\mu _{0}q^{2}n_{s})^{0.5} $ is the London penetration depth. Two permitted states $n$ and $n+1$ have equal energy at $\Phi = (n+0.5) \Phi _{0} $ according to (2) and thus equal probability $P \propto \exp (-E _{k}/k _{B}T) $. The probability of other permitted states is negligible and $P(n+1) \approx 1 - P(n)$ at $\Phi \approx (n+0.5) \Phi _{0} $ when $\Phi _{0}^{2}/2L_{k} = I _{p,A}\Phi _{0} \gg k _{B}T$. The probability of the $n$ state may be described with the relation $$P(n) \approx \frac{1}{1+\exp \epsilon 2\frac{\delta \Phi}{\Phi _{0}}} \eqno {(3)}$$ at the magnetic flux $\Phi = (n+0.5) \Phi _{0} + \delta \Phi $, when $\delta \Phi \ll \Phi _{0}$. Here $ I _{p,A} = \Phi _{0}/2L_{k}$ is the persistent current (1) at $\|n - \Phi /\Phi _{0}\| = 1/2$ and $\epsilon = I _{p,A}\Phi _{0}/k _{B}T $. The probability (3) changes from $P(n) \approx 1$ to $P(n) \approx 0$ in a region of the magnetic flux from $\delta \Phi \approx -\Phi _{0}/2\epsilon $ to $\delta \Phi \approx \Phi _{0}/2\epsilon $. This region may be very narrow due to a big value $\epsilon = I _{p,A}\Phi _{0}/k _{B}T \gg 1$ equal, for example $\epsilon \approx 1500$ at the temperature of measurement $T \approx 1 \ K$ and a typical value $I _{p,A} = 10 \ \mu A$ measured, for example in [@PCJETP07]. Measurements [@Tanaka04] of flux qubit (superconducting loop with three Josephson junctions) corroborate the change of the probability $P(n) $ in a narrow region of magnetic flux predicted by the relation (3). For example the probability changes from $P(n) \approx 1$ to $P(n) \approx 0$ in a region from $\delta \Phi \approx -0.002\Phi _{0}$ to $\delta \Phi \approx 0.002\Phi _{0}$ at the temperature of measurement $T \approx 0.1 \ K$, see Fig.5 of [@Tanaka04], and the value $I _{p,A} < 0.6 \mu A$ of the flux qubit measured in [@Tanaka04]. ![\[fig:epsart\] The work principle of the superconducting differential double contour interferometer (DDCI). A bias current $I$ flows from the upper loop to the lower loop through two Josephson junctions $J _{a}$ and $J _{b}$ (labeled with white circles). The maximum value of the superconducting current through these Josephson junctions should depend on parity of the sum $n _{u} + n _{d}$ of the quantum numbers of the upper loop $n _{u}$ and the lower loop $n _{d}$ according to (6). Therefore the voltage measured on DDCI average in the time $\overline{V} = \Theta ^{-1}\int _{\Theta} dt V(t) \approx V _{min}P(n _{u} =0) + V _{max}P(n _{u} =1)$ should change from the minimum value $\overline{V} \approx V _{min}$ to the maximum value $\overline{V} \approx V_{max}$ in the narrow interval $\Delta \Phi \approx \Phi _{0}/\epsilon \ll \Phi _{0}/2$ at $\epsilon \gg 1$ when $n _{d} = 0$ and the quantum number of the upper loop takes two values $n _{u} = 0$ and $n _{u} = 1$ due to the switching of a segment of this loop in normal state by the short pulses of an external current $I_{sw}$ with the amplitude exceeding the critical current $I_{c}$ of the superconducting strip.](Fig1) The quantum number of the continuous superconducting loop, in contrast to the flux qubit [@Tanaka04], can change only if it or its segment is switched in normal state for a while [@PRB2014C]. The numerous measurements of the critical current [@PCJETP07; @APL2016; @JETP07J] and other parameters [@PCScien07; @PLA2012] testify that the loop comes back in the superconducting state with the quantum number $n$ corresponding to the minimal kinetic energy (2) at the $(n-0.5)\Phi _{0} < \Phi < (n+0.5) \Phi _{0}$ with the predominant probability $P(n) \approx 1$.The probability $P(n)$ of the $n$ state should change in a narrow interval of magnetic flux $\Delta \Phi \approx \Phi _{0}/\epsilon $ near $\Phi = (n+0.5) \Phi _{0} $ when a segment of the continuous superconducting loop is switched in the normal state by the short pulses of an external current $I_{sw}$ with a frequency $f$. Such switching may be provided with the help of the additional current leads shown on Fig.1. The quantum number $\overline{n} = \Theta ^{-1}\int _{\Theta} dt n(t)$ average in a time $\Theta \gg 1/f$ equal $\overline{n} \approx nP(n) + (n+1)P(n+1)$ should change in this narrow interval $\Delta \Phi \approx \Phi _{0}/\epsilon \ll \Phi _{0}/2$ from $n$ to $n+1$. We propose to use the differential double contour interferometer, shown on Fig.1, in order to transform the $\overline{n}$ variation in the variation of the voltage average in the time $\overline{V} = \Theta ^{-1}\int _{\Theta} dt V(t)$ in the same narrow interval $\Delta \Phi \ll \Phi _{0}/2$. It may be possible due to the jump of the critical current of the DDCI with the change of the quantum number of one of its contours from $n$ to $n+1$ predicted in [@NANO2010]. The superconducting current between points $l _{u}$ and $r _{d}$ of DDCI, Fig.1, equals the sum $$I _{s} = I _{a}\sin \varphi _{a} + I _{b}\sin \varphi _{b} \eqno{(4)}$$ of the currents $I _{a}\sin \varphi _{a}$ and $I _{b}\sin \varphi _{b}$ through the Josephson junctions $J _{a}$ and $J _{b}$. Here $I _{a}$ and $I _{b}$ are the critical currents of the Josephson junctions; $\varphi _{a}$ and $\varphi _{b}$ are the phase differences between the up $a _{u}$, $b _{u}$ and down $a _{d}$, $b _{d}$ boundaries of the Josephson junctions. The relation $$\oint_{l}dl \bigtriangledown \varphi = 2\pi n \eqno{(5)}$$ must be valid for the both contours $l _{u} - a _{u}- r_{u} - b_{u}- l _{u}$ and $l _{d} - a _{d}- r_{d} - b_{d}- l _{d}$ due to the requirement that the complex wave function must be single-valued in any point of the circumference $l $ of each contour $\Psi = \|\Psi \|e^{i\varphi } = \|\Psi \|e^{i(\varphi + n2\pi )}$. The relation (5) should be valid also for the contours $l _{u} - a _{u} - a _{d} - r_{d} - b_{d}- b_{u}- l _{u}$ when the current through the Josephson junctions does not exceed the critical current $I _{a}$ and $I _{b}$. Then, the superconducting current between $l _{u}$ and $r _{d}$ is equal to: $$I _{s} = I _{a}\sin \varphi _{a} + I _{b}\sin (\varphi _{a}+ \pi (n _{u} + n _{d})) \eqno{(6)}$$ The critical current $I _{c,DD}$ of the DDCI, i.e. the maximum value of the superconducting current (6), depends only on parity of the quantum number sum, and it does not explicitly depend on the magnetic flux, which makes the system to be an ideal detector of the quantum states. The critical current has only two values at $I _{a}= I _{b} = I _{cj}$ according to (6): $I _{c,DD} = 2 I _{c,j}$ when the sum $n _{u} + n _{d}$ of the quantum numbers of the upper loop $n _{u}$ and the bottom loop $n _{d}$ is even and $I _{c} = 0$ when it is odd. Therefore, for example at $\Phi = 0.5 \Phi _{0} + \delta \Phi $, the maximum value of the superconducting current (6) and the voltage measured on the DDCI at a bias current $I$ may change by jump when the quantum number of the bottom loop should be equal permanently zero $n _{d} = 0$ whereas the quantum number the upper loop may change from $n _{u} =0$ to $n _{u} =1$ (or from $n _{u} =1$ to $n _{u} = 0$) after successive switching of the loop segment in the normal state by the short pulse of the external current $I_{sw}$, Fig.1. The probability of the minimum value of the voltage $V _{min}$ (the maximum value of the critical current $I _{c,DD} = 2 I _{c,j}$ ) should be equal the probability (3) of the state $n _{u} =0$ because $n _{u} + n _{d} = 0 + 0 = 0$ is even. Therefore the voltage average in the time $\overline{V} = \Theta ^{-1}\int _{\Theta} dt V(t) \approx V _{min}P(n _{u} =0) + V _{max}P(n _{u} =1)$ should change from the minimum value $\overline{V} \approx V _{min}$ to the maximum value $\overline{V} \approx V_{max}$ in the narrow interval $\Delta \Phi \approx \Phi _{0}/\epsilon \ll \Phi _{0}/2$ at $\epsilon \gg 1$. The DDCI may be more sensitive than the conventional dc SQUID due to this sharpness of the voltage $\overline{V}$ change at its most sensitive point $\Phi \approx 0.5 \Phi _{0} $ (or $\Phi \approx (n+0.5) \Phi _{0} $). One may expect to increase the flux sensitivities in $\epsilon /2$ times if the jump $V_{max} - V _{min}$ is not less than the voltage variation of the conventional dc SQUID. In order to ensure that the change of the quantum number $n _{u}$ results to the voltage jump and to measure its value we used a real structures shown on Fig.2. The structure was fabricated by e-beam lithography to form suspended resist mask and double angle shadow evaporation of aluminum ($d \approx 30 \ nm$ and $d \approx 35 \ nm$) with intermediate first aluminum layer oxidation. This technology allows to make two independent superconducting square contours weakly connected by two Josephson junctions in the two points $J _{a}$ and $J _{b}$, Fig. 1. The contours are shifted relative to one another and there are two extra Josephson junctions $J _{t1}$ and $J _{t2}$ because of this technology. The structures with square side of the loops $a \approx 4 \ \mu m$ and $a \approx 20 \ \mu m$ were made and investigated. The width of the loops was $w \approx 400 \ nm$ and their cross section $s = dw \approx 12000 \ nm^{2}$ corresponds to $\lambda _{L}^{2}(T) = \lambda _{L}^{2}(0)(1 - T/ T _{c})^{-1}$ of aluminium with its London penetration depth $\lambda _{L}(0) \approx 50 \ nm$ at $T \approx 0.8T _{c}$. The critical temperature of aluminium loops was $T _{c} \approx 1.3 \ K$. ![\[fig:epsart\] The real structure used for the observation of the voltage jumps with the change of the quantum numbers of its two loops weakly connected by two Josephson junctions $J _{a}$ and $J _{b}$. The two aluminium loops are shifted with respect to each other because of the simple technology of two-angle evaporation used in our work. The structure has two unnecessary Josephson junctions $J _{l}$ and $J _{r}$ because of this technology. Photo of 4 $\mu m $ aluminium loops is shown.](Fig2) The structure, shown on Fig.2, without the additional current leads $cl$, Fig.1, cannot be used as a sensitive magnetometer because the quantum number $n _{u}$ of the continuous superconducting loop may not meet the minimum kinetic energy (2) when no segment is switched in the normal state. We use this structure only for the observation of the voltage jump due to the change of the quantum number. This change takes place when the persistent current (1) increases in the loop without the current leads with the change of the magnetic flux $\Phi = BS + LI_{p}$ up to a critical value [@Geim2003; @Moler2007; @Nat2016]. The pair velocity $v = (2\pi \hbar/ml)(n - \Phi /\Phi _{0})$ can increase with the $\Phi $ variation up to the depairing velocity $v _{c} = \hbar /m \surd 3 \xi (T)$ [@Tink75] but cannot exceed this value $\|v\| \leq v _{c}$. Therefore the quantum number should change at $\|n - \Phi /\Phi _{0}\| \leq l/2\pi \surd 3 \xi (T)$ [@Geim2003; @Moler2007; @Nat2016]. The magnetic flux at which the quantum number $n$ changes may be not equal $\Phi = (n+0.5)\Phi _{0}$ and may be different in ascending and descending magnetic field. The jumps of the critical current (6) and the voltage at a non-zero bias current $I \neq 0$ can be observed only if the quantum numbers of the upper loop $n _{u}$ and the lower loop $n _{d}$ change at different $\Phi $ values. The voltage may jump up (when the sum $n _{u} + n _{d}$ becomes an odd number) and down (when the sum $n _{u} + n _{d}$ becomes again an even number) in this case with the period corresponding to the flux quantum inside each loop. In the experiments we indeed observe such digital type oscillations of voltage, Fig. 3. The observed periodicity leaves no doubt that the voltage jumps up and down are a consequence of the change of $n _{u}$ and $n _{d}$. The period of the digital type oscillations in magnetic field $B _{0} = \Phi _{0}/S$ corresponds to the flux quantum $\Phi _{0} \approx 20.7 \ Oe \ \mu m^{2}$ inside the loop both with square side $a \approx 20 \ \mu m$ and with $a \approx 4 \ \mu m$: the period $B _{0} \approx 0.053 \ Oe $ observed in the first case corresponds to the area $S = \Phi _{0}/B _{0} \approx 390 \ \mu m^{2} \approx a^{2} $ and the period $B _{0} \approx 1.2 \ Oe $ observed in the second case corresponds to the area $S = \Phi _{0}/B _{0} \approx 17 \ \mu m^{2} \approx a^{2}$. We observed more than 110 voltage jump up and down at measurement of the structure with square side $a \approx 20 \ \mu m$ in the magnetic field from $B = -3 \ Oe$ to $B = 3 \ Oe$. ![\[fig:epsart\] Voltage jumps due to changes of quantum numbers in magnetic field at the bias currents through the structure $I \approx 20 \ nA $ observed at the temperature $T \approx 1.1 \ K $ at measurement of the structures with square side of the loops $a \approx 20 \ \mu m$. A part of magnetic dependence measured from $B = 3 \ Oe$ to $B = -3 \ Oe$ is shown. The voltage jumps up and down with the period $B _{0} \approx 0.053 \ Oe $ and the amplitude of the jumps is modulated with the period $B _{m} \approx 0.8 \ Oe $ in the whole region from $B = 3 \ Oe$ to $B = -3 \ Oe$, see Fig.8 of Supporting Information.](Fig3) The observation of these jumps has the critical importance for the opportunity to use the DDCI for high sensitive detection of magnetic flux. The conventional theory [@Tink75] and numerous experiments [@PCJETP07; @Tanaka04; @APL2016; @JETP07J; @PCScien07; @PLA2012] testify to the sharpness of the change of the probability $P(n)$ of the $n$ state of superconducting loop with enough big value of the persistent current $I _{p,A}$. The theory predicts the jump of the critical current with the $n$ change at $\Phi = (n+0.5) \Phi _{0}$ not only of the double contour interferometer but also, for example, of a superconducting ring with asymmetric link-up of current leads. A simple magnetometer based on the latter prediction was proposed in [@Letter2014]. But measurements of aluminium ring with asymmetric link-up of current leads have revealed that a smooth change of its critical current is observed at $\Phi \approx (n+0.5) \Phi _{0}$ instead of the jump which must be observed due to the change of the quantum number from $n$ to $n+1$ [@PLA2017]. Our observations of the voltage jumps up to $V _{max} - V _{min} \approx 20 \ \mu V$, Fig.3, and higher mean that the derivative $(\partial \overline{V}/\partial \Phi _{e}) _{I}$ can reach high values when the voltage $\overline{V}$ average in the time $\Theta \gg 1/f$ changes from $\overline{V} \approx V _{min}$ to $\overline{V} \approx V_{max}$ in the narrow interval $\Delta \Phi \approx \Phi _{0}/\epsilon \ll \Phi _{0}/2$: for example $(\partial \overline{V}/\partial \Phi _{e}) _{I} \approx (V _{max} - V _{min})/\Delta \Phi \approx 20 \ mV/\Phi _{0}$ at $V _{max} - V _{min} \approx 20 \ \mu V$ and the real value $\epsilon = I _{p,A}\Phi _{0}/k _{B}T \approx 1000$. ![\[fig:epsart\] Current-voltage characteristics of the DDCI structures with square side of the loops $a \approx 4 \ \mu m$ measured at the temperature $T \approx 0.44 \ K $ and different magnetic field $B \approx 9.2 \ Oe$ and $B \approx 10.2 \ Oe$. Two extra steps correspond to the two additional Josephson junctions of the real DDCI at $ l _{u}$ and $r _{d}$.](Fig4) The experimental corroboration of the voltage jumps, Fig.3, is the central objective of our measurements of the real double contour interferometer shown on Fig.2. Other peculiarities observed at these measurements are considered partly in the Supporting Information. Here we say only about the periodical modulation of the jump amplitude, Fig.3, and some peculiarities of the current-voltage characteristics of the real structures. We assume that the modulation of the jump $V _{max} - V _{min}$ with the period of $B _{m} \approx 0.8 \ Oe $ may be connected with the shift $\approx 0.5 \ \mu m $ of the loops relative to one another in the measured structure, Fig.2, see the Supporting Information. Three steps at the current-voltage characteristics shown in Fig. 4 are observed because of the necessary Josephson junctions $J _{a}$, $J _{b}$ and two unnecessary Josephson junctions $J _{l}$, $J _{r}$ created because of our technology. Two current-voltage characteristics, shown in Fig. 4, measured at different value of magnetic field correspond to even and odd sum $n _{u} + n _{d}$. In summary we propose a new type of magnetometer, the DDCI, and present experimental evidence of its opportunity. The DDCI can reach sensitivities better than $20 \ mV/\Phi _{0}$, which exceeds that of dc SQUID by more than one order of magnitude. The effect is due to the strong discreteness of the energy spectrum of the continuous superconducting loop. The advantage of DDCI lies also in circuitry, as one does not need to couple the device to a flux transformer for the measurement of small magnetic fields. The flux transformer is used for the measurement of tiny magnetic fields $B = \Phi /S$ since the area $S$ of the dc SQUID cannot be too large because of the strong screening $\Phi _{I} = LI _{c,j}> \Phi _{0}/2$ and $\beta = 2L I _{c,j}/\Phi _{0} > 1$ in the loop with a high magnetic inductance $L \approx \mu _{0}l$ [@SQUID1977]. In the DDCI the magnetic flux induced by the persistent current does not depend on the loop size $l$ since $\Phi _{I} = LI _{p} = (L/L _{k})(n\Phi_{0} - \Phi) \approx (s/\lambda _{L}^{2}(T))(n\Phi_{0} - \Phi)$ [@QuSMF2016]. This work has been supported by the Russian Science Foundation, Grant No. 16-12-00070. Supporting Information =========================== Strong discreteness of the energy spectrum of continuous superconducting loop ----------------------------------------------------------------------------- Niels Bohr postulated as far back as 1913 a strong discreteness of the energy spectrum of atoms. According to Bohr’s condition $pr = n\hbar $, that the angular momentum $pr$ is an integer multiple of Planck’s constant $\hbar = h/2\pi $, the energy discreteness $E_{n+1} - E_{n} = p_{n+1}^{2}/2m - p_{n}^{2}/2m = (2n+1)\hbar ^{2}/2mr^{2}$ decreases with the increase of the orbit radius $r$. The energy spectrum of atom is strongly discrete due to small radius of electron orbits: the energy difference between adjacent permitted states $\Delta E \approx \hbar ^{2}/2mr^{2} \approx 2 \ 10^{-18} \ J$ for the Bohr radius $r_{B} \approx 0.05 \ nm = 5 \ 10^{-11} \ m$ corresponds to the temperature $T = \Delta E/k _{B} \approx 100000 \ K$. The discreteness decreases down to $\Delta E \approx \hbar ^{2}/2mr^{2} \approx 2 \ 10^{-18} \ J$ corresponding to $T = \Delta E/k _{B} \approx 0.001 \ K$ in a mesoscopic ring with a radius $r \approx 500 \ nm = 5 \ 10^{-7} \ m$. Therefore the quantum phenomena connected with the discrete spectrum, such as the persistent current of electrons, can be observed in nano-rings with a real radius $r > 300 \ nm$ only at very low temperature [@PCScien09; @PCPRL09]. The energy difference of superconducting loop $\Delta E \approx N _{s}\hbar ^{2}/2mr^{2} $ is much larger due to the impossibility for all $N _{s}$ Cooper pairs in the loop to change their quantum state $n$ individually [@FPP2009]. This impossibility of individual motion of quantum particle was postulated first by Lev Landau as far back as 1941 [@Landau41] for the description of superfluidity of $^{4}He$ liquid. All $N _{s}$ pairs in the loop with the volume $V = ls$, the length $l$ and the section area $s$ are described with the wave function $\Psi = \|\Psi \|\exp{i\varphi }$, according to the Ginzburg-Landau theory [@GL1950]: $\|\Psi \|^{2} = n_{s}$ is the density of Cooper pairs and $\int_{V}dV \|\Psi \|^{2} = \oint_{l}dl s\|\Psi \|^{2} = N_{s} \gg 1$ is the total number of Cooper pairs in the loop. This number exceeds 100000 in a typical superconducting loop. Therefore its spectrum of the permitted states is strongly discrete: the energy difference between adjacent permitted states of a ring with a radius $r \approx 500 \ nm $ exceeds the value $\Delta E \approx N _{s}\hbar ^{2}/2mr^{2} \approx 2 \ 10^{-21} \ J$ corresponding to the temperature $T = \Delta E/k _{B} \approx 100 \ K$ at $N _{s} > 10^{5}$. The discreteness increases with the increase of all three sizes of the ring $\Delta E \approx N _{s}\hbar ^{2}/2mr^{2} \approx n_{s}s2\pi r(\hbar ^{2}/2mr^{2}) \propto (s/r)$ due to the increase of the number of Cooper pairs $N _{s}$. The quantization of angular momentum postulated by Bohr may be deduced from the requirement that the complex wave function must be single-valued in any point $l$ of the loop $\Psi = \|\Psi \|e^{i\varphi } = \|\Psi \|e^{i(\varphi + n2\pi )} $. The relation $$\oint_{l}dl \bigtriangledown \varphi = 2\pi n \eqno{(1)}$$ must be valid for any contour $l$ along which the wave function $\Psi = \|\Psi \|e^{i\varphi }$ is defined, according to this requirement. The angular momentum of each Cooper pair has a discrete value $n\hbar $ and the total angular momentum of all $N _{s}$ pairs equals $M _{p} = \oint _{l} dl sr\Psi ^{}\hat{p} \Psi = \oint _{l} dl sr\Psi ^{} (-i\hbar \nabla )\Psi = sr\|\Psi \|^{2}\hbar \oint _{l} dl \nabla \varphi = N _{s}\hbar n $. ![\[fig:epsart\] The picture on the left: The virgin magnetization $\Delta M \propto L_{f}I _{p}$ of aluminum ring with the radius $r \approx 1 \ \mu m$ increases up to $\|v\| \approx v _{c} $ with the magnetic flux $\Phi \approx BS$. The quantum number changes from $n = 0$ to $n = 1$ $\|\Phi /\Phi _{0}\| \approx 1.7 \approx r/\surd 3 \xi (T)$, $n = 1$ to $n = 2$ $\|\Phi /\Phi _{0}\| \approx 1.7 + 1 = 2.7$, $n = 2$ to $n = 3$ $\|\Phi /\Phi _{0}\| \approx 1.7 + 2 = 3.7$. The drawing at the upper right: The velocity of Cooper pairs changes by jump at a definite magnetic flux $\Phi = (n+0.5) \Phi _{0}$ due to the switching of ring segment in the normal state for a while by an external current, a noise or thermal fluctuations, or due to quantum fluctuations. The two states $n$ and $n+1$ have the opposite velocity $v$ and equal kinetic energy $E _{k} \propto v^{2}$. The picture below on the right: Measurements of the critical current of symmetric aluminum ring with the radius $r \approx 2 \ \mu m$ [@JETP07J] corroborate the predominant probability of the $n$ state in the interval $(n-0.5) \Phi _{0} < \Phi < (n+0.5) \Phi _{0}$.](Fig5) According to the canonical definition the gradient operator $\hat{p} = -i\hbar \nabla $ corresponds to the canonical momentum $p = mv + qA$ of a particle with a mass $m$ and a charge $q$ both with $A \neq 0$ and without $A = 0$ magnetic field [@LandauL]. The velocity operator $\hat{v} = (\hat{p} - qA)/m = (-i\hbar \nabla - qA)/m$ [@FeynmanL], in contrast to the momentum operator, depends on the magnetic vector potential $A$. Therefore the velocity of Cooper pair with $q = 2e$ $$\oint_{l}dlv = \frac{2\pi \hbar }{m}(n-\frac{\Phi }{\Phi_{0}}) \eqno{(2)}$$ cannot be equal zero when the magnetic flux $\Phi = \oint_{l}dl A$ inside the loop $l$ is not divisible by the flux quantum $\Phi _{0} = 2\pi \hbar /q = \pi \hbar /e \approx 20.7 \ Oe \ \mu m^{2}$. The velocity $ldv/dt = -(q/m)d\Phi /dt$ and the persistent current $L _{k}dI _{p}/dt = L _{k}sqn_{s}dv/dt = -d\Phi /dt $ change with the magnetic flux $\Phi $ in accordance with the Newton’s second law $mdv/dt = qE$, where $E = -\nabla V - dA/dt$ is the electric field, $L _{k} = ml/ q^{2}n_{s}s = (\lambda _{L}^{2}/s) \mu _{0}l$ is the kinetic inductance of the loop with the length $l$, the cross-sectional area $s = wd$ and the density of Cooper pairs $n_{s}$ and $\lambda _{L} = (m/\mu _{0}q^{2}n_{s})^{0.5} $ is the London penetration depth. The magnetic inductance $L_{f}$ of superconducting loop with a small cross-sectional area $s \ll \lambda _{L}^{2}$ is much lower than the kinetic inductance $L_{f} \approx \mu _{0}l \ll L _{k} = (\lambda _{L}^{2}/s) \mu _{0}l $ at $s \ll \lambda _{L}^{2}$. The magnetic flux $\Delta \Psi _{I} = L_{f}I _{p}$ induced by the persistent current $I _{p}$ and the energy of the magnetic field $L_{f}I _{p}^{2}/2 \ll L_{k}I _{p}^{2}/2$ are small in this case of weak screening. ![\[fig:epsart\] The scheme of an ideal DDCI. The loops located one above the other and therefore the boundaries of the Josephson junctions $J _{a}$ and $J _{b}$ coincide with the points $a _{u}$, $a _{d}$ and $b _{u}$, $b _{d}$. The critical current of such DDCI should not depend directly on the magnetic flux inside the loops.](Fig6) The pair velocity may increase up to the depairing velocity $\|v\| \leq v _{c} = \hbar /m \surd 3 \xi (T)$ in narrow loop segments [@Tink75] without a change of the quantum number $n$, Fig.5. The quantum number can change when a density of Cooper pairs diminishes for a while from $n_{s} = 2n_{s0}/3$ [@Tink75] to $n_{s} = 0$ at $\|v\| \geq v _{c} $ in a segment, where $n_{s0}$ is the density at $\|v\| = 0$. The quantum number of homogeneous ring changes at $\|n - \Phi /\Phi _{0}\| \approx r/\surd 3 \xi (T)$, Fig.5, when the velocity increases only because of magnetic flux $\Phi \approx BS$ change in accordance with (2). The jump of the velocity (2) and the persistent current, observed in this case [@Geim2003; @Moler2007; @Nat2016], cannot be used for a magnetometer designing because of the problem with a hysteretic field characteristic. The quantum number can change from $n$ to $n+1$ at a definite magnetic flux $\Phi = (n+0.5) \Phi _{0}$, Fig.5, due to the switching of a segment in the normal state for a while by an external current at $T < T _{c}$ [@PRB2014C], by thermal fluctuations at $T \approx T _{c}$ [@JLTP1998] or quantum tunneling [@Tanaka2002]. ![\[fig:epsart\] The scheme of the DDCI with the loops shifted relative to each other. The boundaries of the Josephson junctions $J _{a}$ and $J _{b}$ do not coincide with the points $a _{u}$, $a _{d}$ and $b _{u}$, $b _{d}$. Therefore the critical current may depend directly on the magnetic flux inside the loops.](Fig7) The ring is switched in the normal state for a while by the external current in the process of measuring the critical current [@PRB2014C]. The measurements [@PCJETP07] corroborate that the quantum number of the ring corresponds to the minimal kinetic energy $\propto v^{2}$ with the predominant probability after its coming back in superconducting state. The oscillations of the critical current of a symmetric superconducting ring, Fig.5, observed for example in [@JETP07J], are similar to the one of a conventional dc SQUID, i.e. superconducting loop with two Josephson junctions [@Barone]. But there is a principal difference between superconducting loops with and without the Josephson junction. The phase change at a complete turn along the loop with Josephson junctions depends on the sum of the phase differences on the Josephson junctions and the magnetic field inside the loop: $2\pi n = \sum _{i}\varphi _{i} - 2\pi \Phi /\Phi _{0}$ [@Barone]. The persistent current in the loop with one Josephson junction, i.e. in an rf SQUID, $I _{p} = I _{c,j}\sin \varphi = I _{c,j}\sin (2\pi n + 2\pi \Phi /\Phi _{0}) = I _{c,j}\sin (2\pi \Phi /\Phi _{0})$ should not depend on the quantum number $n$ because of the mathematical equality $\sin (\varphi + 2\pi n) \equiv \sin (\varphi )$. Therefore no jump connected with the $n$ change should be observed in this case. The jump of the persistent current should be observed in the loop without the Josephson junction. This jump should not result to the jump of the critical current of a symmetric ring. The jump of the critical current must be observed at measurements of asymmetric rings. But this jump is not observed [@PCJETP07; @PLA2017] for some strange reason. Superconducting current through the differential double contour interferometer ------------------------------------------------------------------------------ The phases $\varphi $ of the wave functions $\Psi = \|\Psi \|\exp{i\varphi }$ describing superconducting state of the two loops of the differential double contour interferometer are connected due to the two Josephson junctions $J _{a}$ and $J _{b}$, Fig.6. According to the Josephson relation $I _{s} = I _{c,j}\sin \varphi $ the superconducting currents depend of the phase difference $\varphi _{a}$ and $\varphi _{b}$ between the boundaries of the Josephson junctions $J _{a}$ $\varphi _{a} = \varphi _{a,d} - \varphi _{a,u}$ and $J _{b}$ $\varphi _{b} = \varphi _{b,d} - \varphi _{b,u}$, where $\varphi _{a,d}$, $\varphi _{a,u}$, $\varphi _{b,d}$ and $\varphi _{b,u}$ are the phase of the wave functions at the points $a _{d}$, $a _{u}$, $b _{d}$ and $b _{u}$. The total superconducting current through the two Josephson junctions $$I _{s} = I _{a}\sin \varphi _{a} + I _{b}\sin \varphi _{b} \eqno{(3)}$$ should depend on the phases $\varphi _{a,d}$, $\varphi _{a,u}$, $\varphi _{b,d}$ and $\varphi _{b,u}$. The phase change from $a _{d}$ to $b _{d}$ and from $b _{d}$ to $a _{d}$, clockwise for example, should be equal $2\pi n _{d}$ and the phase change from $a _{u}$ to $b _{u}$ and from $b _{u}$ to $a _{u}$ should be equal $2\pi n _{u}$ according to (1). The phase changes from $a _{d}$ to $b _{d}$ and from $b _{d}$ to $a _{d}$ are equal when the velocity of Cooper pairs in the ring halves is the same. The velocity should be the same in a homogeneous loop with the persistent current exceeding strongly the superconducting current (3) through the two Josephson junctions $I _{p} \gg I _{s}$. In this case the phase differences of the Josephson junctions $J _{a}$ and $J _{b}$ are connected with the quantum numbers of the loops by the simple relation $\varphi _{a} - \varphi _{b} = (\varphi _{a,d} - \varphi _{a,u}) - (\varphi _{b,d} - \varphi _{b,u}) = \pi (n _{u} + n _{d})$ due to the equality $\varphi _{a,d} - \varphi _{b,d} = \pi n _{d}$ and $\varphi _{b,d} - \varphi _{a,d} = \pi n _{d}$. According to this relation $\varphi _{b} = \varphi _{a} - \pi (n _{u} + n _{d})$ and the superconducting current (3) depends on the single phase difference $\varphi _{a}$ and two quantum numbers $I _{s} = (I _{a} + I _{b})\sin \varphi _{a} = 2I _{cj}\sin \varphi _{a}$ at $I _{a}= I _{b} = I _{cj}$ when the sum $n _{u} + n _{d}$ is even and $I _{s} = (I _{a} - I _{b})\sin \varphi _{a} = 0$ at $I _{a}= I _{b}$ when $n _{u} + n _{d}$ is odd. ![\[fig:epsart\] Voltage oscillations in the DDCI with the side of the loops $a \approx 20 \ \mu m$ measured in the opposite directions of magnetic field sweep, from $B \approx +3 \ Oe$ to $ B \approx -3 \ Oe$ and $B \approx -3 \ Oe$ to $ B \approx +3 \ Oe$.](Fig8) The differential double contour interferometer with shifted loops ----------------------------------------------------------------- The phase differences between the boundaries of the Josephson junctions are connected with the quantum numbers of superconducting loops by the simple relation $\varphi _{a} - \varphi _{b} = \pi (n _{u} - n _{d})$ due to the overlap of the points $a _{d}$, $J _{a}$, $a _{u}$ and $b _{d}$, $J _{b}$, $b _{u}$ of the ideal structure shown on Fig.6. These points are shifted with respect to each other, Fig.7, in the real structure because of the simple shadow evaporation technique used in our work for the DDCI fabrication. The upper loop is shifted relatively the bottom one on $a _{sh} \approx 0.6 \ \mu m$ because of this technique. The path from the upper (bottom) boundary of the Josephson junction $J _{a}$ to the upper (bottom) boundary of the $J _{b}$ through $l _{u}$ ($r _{d}$) is not equal the path through $r _{u}$ ($l _{d}$), Fig.7. The first path is longer than the second one on $4a _{sh}\surd{2}$. The path along the contour $l _{d} - a _{d} - J _{a} - a _{u} - r _{u} - b _{u} - J _{b} - b _{d} - l _{d}$ increases on the value $4a _{di} = 4a _{sh}\surd 2$, Fig.7. Therefore the superconducting current through the real DDCI used in our work for the experimental investigations $$I _{s} = I _{a}\sin \varphi _{a} + I _{b}\sin (\varphi _{a}+ \pi (1+\frac{a _{di}}{a})(n _{u} + n _{d}) + 2\pi \frac{a _{di}aB}{\Phi _{0}}) \eqno{(4)}$$ depends not only on the phase difference $\varphi _{a}$ and the quantum numbers $n _{u} - n _{d}$ but also on the magnetic field value $B$. The term $2\pi a _{di}aB/\Phi _{0}$ in the relation can explain the modulation of the jump amplitude with the period $B _{m} \approx 0.8 \ Oe$ observed at measurements of the DDCI with the side of the loops $a \approx 20 \ \mu m$, Fig.8. What is the maximal magnetic field in which the voltage jumps may be observed? ------------------------------------------------------------------------------ The jumps of the critical current of the DDCI and the voltage jumps at a biased current can be observed until their loops are in superconducting state. Magnetic field depresses superconductivity in a strip with a finite width $w$. The value of the depression depends of the temperature and the stripe width. For example the critical current of the aluminum strip with $w \approx 0.6 \ \mu m$ decreases in two time at the magnetic field $B \approx 20 \ Oe$ and in four time at $B \approx 30 \ Oe$ at the temperature $T \approx 0.986T _{c}$, see Fig.9 of [@PCJETP07]. More than 1000 jumps may be observed in the interval $30 \ Oe < B < 30 \ Oe$ at the measurement of the DDCI with the side of the loops $a \approx 20 \ \mu m$ with the period of jumps $B _{0} \approx 0.053 \ Oe$. We made the measurement in the interval $3 \ Oe < B < 3 \ Oe$ and observed more than 100 jumps, Fig.8. The depression of superconductivity by magnetic field decreases with the temperature $T$ and width $w$ decrease. The voltage jumps may be observed in the interval of magnetic field much wider than $30 \ Oe < B < 30 \ Oe$ when $T < 0.986T _{c}$ and $w < 0.6 \ \mu m$. The number of the jumps may be much more than 1000. We demonstrate on Fig.8 the hysteretic field characteristic of our DDCI. The magnetic field at which the quantum numbers of the loops change depends on the direction of sweep because the loop is switched in normal state for a while by the persistent current $I _{p}$ at $\|n - \Phi /\Phi _{0}\| \approx r/\surd 3 \xi (T)$ rather than by an external current $I _{ext}$, a noise or thermal fluctuations. The quantum number of the superconducting loop can change both on 1 and 2, 3… in this case [@Geim2003; @Moler2007]. It is observed on Fig.8 that the period of the voltage jumps $B _{0} \approx 0.11 \ Oe$ in the beginning of the sweep and $B _{0} \approx 0.053 \ Oe$ in the continuation of the sweep. The period $B _{0} \approx 0.053 \ Oe$ corresponds to the flux quantum $B _{0}S = \Phi _{0} \approx 20.7 \ Oe \ \mu m^{2}$ in the loop with the area $S = \Phi _{0}/B _{0} \approx 390 \ \mu m^{2} \approx (19.8 \ \mu m)^{2}$. Consequently, the observations shown on Fig.8 mean that the quantum numbers of the DDCI loops change on 2 in the beginning of the sweep and 1 in the continuation of the sweep. It is not clear why this difference is observed. [99]{} R. C. Jaklevic, J. Lambe, A. H. Silver, and J. E. Mercereau, Quantum Interference Effects in Josephson Tunneling. Phys. Rev. Lett. [12]{}, 159 (1964) B.B. Schwartz and S. Foner, [Superconductor Applications: SQUIDs and Machines]{}. Plenum Press, New York, 1977 A. Barone and G. Paterno, [Physics and Applications of the Josephson Effect]{}. Wiley-Interscience, New-York, 1982. I. N. Zhilyaev, S. G. Boronin, K. Fossheim. Step-like oscillations in the resistance of two weakly linked aluminum rings. [Physica C]{}, [332]{}, 422 (2000). I. N. Zhilyaev, S. G.Boronin, A. V. Nikulov and K. Fossheim, States in the Structure of Weakly Connected Superconducting Rings. [Quantum Computers and Computing]{}, [2]{}, 49 (2001). A.V. Nikulov, The Meissner effect puzzle and the quantum force in superconductor. [Phys.Lett. A]{}, [376]{}, 3392 (2012). M.Tinkham, [Introduction to Superconductivity.]{} McGraw-Hill Book Company (1975). A.V. Nikulov, Could ordinary quantum mechanics be just fine for all practical purposes? [Quantum Studies: Mathematics and Foundations]{}, [3]{}, 41 (2016) V. L. Gurtovoi, S. V. Dubonos, A. V. Nikulov, N. N. Osipov, and V. A. Tulin, Dependence of the magnitude and direction of the persistent current on the magnetic flux in superconducting rings. [JETP]{} [105]{}, 1157 (2007). H. Tanaka, S. Saito, H. Nakano, K. Semba, M. Ueda, H. Takayanagi, Single-Shot Readout of Macroscopic Quantum Superposition State in a Superconducting Flux Qubit. e-print arXiv:cond-mat/0407299 (2004) V.L. Gurtovoi and A. V. Nikulov, Comment on “Vortices induced in a superconducting loop by asymmetric kinetic inductance and their detection in transport measurements”. [Phys. Rev. B]{} [90]{}, 056501 (2014). V. L. Gurtovoi, M. Exarchos, V. N. Antonov, A. V. Nikulov and V. A. Tulin, Multiple superconducting ring ratchets for ultrasensitive detection of non-equilibrium noises. [Appl. Phys. Lett.]{} [109]{}, 032602 (2016) V.L. Gurtovoi, S.V. Dubonos, S.V. Karpi, A.V. Nikulov, and V.A. Tulin, Contradiction between the Results of Observations of Resistance and Critical Current Quantum Oscillations in Asymmetric Superconducting Rings. [JETP]{} [105]{}, 262 (2007). N.C. Koshnick, H. Bluhm, M. E. Huber, and K.A. Moler, Fluctuation Superconductivity in Mesoscopic Aluminum Rings. [Science]{} [318]{}, 1440 (2007). A.A. Burlakov, V.L. Gurtovoi, A.I. Ilin, A.V. Nikulov, and V.A. Tulin, Possibility of persistent voltage observation in a system of asymmetric superconducting rings. [Phys. Lett. A]{} [376]{}, 2325 (2012). A.V. Nikulov, Discrete spectrum of measured parameters of a superconductor nanostructure. in Proceedings of [18th International Symposium “NANOSTRUCTURES: Physics and Technology”]{} St Petersburg: Ioffe Institute, p. 367 (2010); e-print arXiv: 1006.5332 D. Y. Vodolazov, F. M. Peeters, S. V. Dubonos, A. K. Geim, Multiple flux jumps and irreversible behavior of thin Al superconducting rings. [Phys. Rev. B]{} [67]{}, 054506 (2003). Hendrik Bluhm, Nicholas C. Koshnick, Martin E. Huber, Kathryn A. Moler, Multiple fluxoid transitions in mesoscopic superconducting rings. e-print arXiv: 0709.1175 (2007) Petkovich, I.; Lollo, A.; Glazman, L.; Harris, J. Deterministic phase slips in mesoscopic superconducting rings. [Nat.Comm.]{} [7]{}, 13551 (2016). A. A. Burlakov, V. L. Gurtovoi, A. I. Il’in, A. V. Nikulov, V. A. Tulin, Superconducting Quantum Interference Device without Josephson Junctions. [JETP Letters]{}, [99]{}, 169 (2014). A.A.Burlakov, A.V.Chernykh, V.L.Gurtovoi, A.I.Ilin, G.M.Mikhailov, A.V.Nikulov, V.A.Tulin, Quantum periodicity in the critical current of superconducting rings with asymmetric link-up of current leads. [Phys.Lett. A]{} [381]{}, 2432 (2017). A.C. Bleszynski-Jayich, W. E. Shanks, B. Peaudecerf, E. Ginossar, F. von Oppen, L. Glazman, J. G. E. Harris, Persistent Currents in Normal Metal Rings. [Science]{} [326]{}, 272 (2009). H. Bluhm, N.C. Koshnick, Ju.A. Bert, M.E. Huber, K.A. Moler, Persistent Currents in Normal Metal Rings. [Phys. Rev. Lett.]{} [102]{}, 136802 (2009) A. V. Nikulov, Bohm’s quantum potential and quantum force in superconductor. in AIP Conference Proceedings, Vol. 1101 [“Foundations of Probability and Physics-5”]{} p. 134 (2009); arXiv: 0812.4118, (2008). L.D. Landau, Theory of superfluidity of helium II. [Zh.Eksp.Teor.Fiz.]{} [11]{}, 592 (1941). V.L.Ginzburg and L.D. Landau, On theory of superconductivity. [Zh.Eksp.Teor.Fiz.]{} [20]{}, 1064 (1950). L.D. Landau, and E.M. Lifshitz, [Quantum Mechanics: Non-Relativistic Theory]{}. Volume 3, Third Edition, Elsevier Science, Oxford, 1977. R.P. Feynman, R.B. Leighton, M. Sands, [The Feynman Lectures on Physics]{}, Addison-Wesley Publishing Company, Reading, Massachusetts, 1963. A.V. Nikulov and I.N. Zhilyaev, The Little-Parks Effect in an Inhomogeneous Superconducting Ring. [J. Low Temp.Phys.]{} [112]{}, 227 (1998). H. Tanaka, Y. Sekine, S. Saito, H. Takayanagi, DC-SQUID readout for qubit. [Physica C]{} [368]{}, 300 (2002).	{ "pile_set_name": "ArXiv" }
--- author: - 'M. K. Volkov' - 'A. A. Pivovarov' - 'A. A Osipov' date: '30 January 2018 / Revised version: date' title: '$\tau \to f_1 (1285)\pi^{-} \nu_{\tau}$ decay in the extended Nambu – Jona-Lasinio model' --- [leer.eps]{} gsave 72 31 moveto 72 342 lineto 601 342 lineto 601 31 lineto 72 31 lineto showpage grestore Introduction {#intro} ============ The recent measurements of the branching fractions of three-prong $\tau$ decay modes made by BABAR Collaboration [@Lees12] contain important new results on the decay $\tau \to f_1(1285) \pi^{-} \nu_{\tau}$ providing confidence that the precise information on the corresponding invariant mass distributions will also soon become available. Together with the large data sets obtained on this mode in the past by BABAR [@Aubert08; @Aubert05] and CLEO [@Bergfeld97] Collaborations this calls for improved understanding of the theoretical description of the process. The main purpose of our paper is to make a step in this direction. The $\tau \to f_1\,\pi^{-} \nu_{\tau}$ decay is driven by the hadronization of the QCD axial-vector currents involved. The details of this mechanism are not yet clearly understood due to poor knowledge of the QCD dynamics at low-energies. Indeed, the invariant mass of the $(f_1, \pi )$-system belongs to the interval $m_{f_1}+m_\pi\leq \sqrt{s}\leq m_\tau$, so it is too low to apply the QCD perturbation theory, but it is too large that the original chiral perturbation theory (ChPT) of QCD would be applicable [@Weinberg79; @Gasser84; @Gasser85a; @Gasser85b]. In addition, an order of magnitude of energies involved seems to indicate that not only the ground axial-vector $a_1(1260)$ state contributes to the pertinent hadronic axial-vector current. It is not excluded that the first radial excitation, $a_1(1640)$, of the ground resonance state may be also important. We have found that both resonances affect the form of the spectral function, giving an interesting interference picture. Presently there is no clear understanding of the nature of the $a_1(1260)$ and $f_1(1285)$ mesons. In our work we hold the view that these resonances are the standard quark-antiquark bound states. Our reasoning is based on the large-$N_c$ expansion of QCD [@Hooft74; @Witten79], indicating that at $N_c=\infty$ mesons are pure $q\bar q$ states. Phenomenologically these states can be described by the local effective meson Lagrangians [@Schwinger67; @Wess67; @Weinberg68; @Gasiorowicz69]. These Lagrangians are not known from first principles, however, they basically can be constructed on the chiral symmetry grounds. Comprehensive reviews of such attempts can be found in [@Meissner88; @Bando88]. The general problem of including these states in ChPT has been addressed in [@Ecker89; @Pich89]. Notice, that there is a different interpretation of $a_1(1260)$ and $f_1(1285)$. Assuming that these resonances not belong to the large-$N_c$ ground state of QCD (i.e. the $qq\bar q\bar q$ states), one can generate them in ChPT by implementing unitarity in coupled-channels (see, for instance, [@Lutz04; @Roca05; @Zhou14; @Xie15] and references therein). This approach can provide an alternative platform for studying of the $\tau\to f_1\pi^-\nu_\tau$ decay. Over last years, not much has been done with respect to the theoretical study of this particular mode of the $\tau$-decay. One can indicate just a few attempts. This is an approach based on a meson dominance model considered in [@Calderon13], which leads to a strong disagreement with the data of BABAR Collaboration, namely the calculated branching ratio Br$(\tau \to f_1\, \pi^{-} \nu_{\tau})=1.3\times 10^{-4}$ is three times less of the experimental value Br$(\tau \to f_1\, \pi^{-} \nu_{\tau})=(3.9\pm 0.5)\times 10^{-4}$ reported by the Particle Data Group [@Patrignani16]. More encouraging but relatively old result, Br$(\tau \to f_1\, \pi^{-} \nu_{\tau})=2.91\times 10^{-4}$, gives the model [@Li97] based on the hypothesis of $a_1(1260)$-meson dominance. Both ideas are naturally realized under the framework of the Nambu – Jona-Lasinio (NJL) model [@Eguchi76; @Volkov82; @Ebert83; @Volkov84; @Volkov86; @Ebert86; @Vogl91; @Klevansky92; @Volkov93; @Bijnens93; @Hatsuda94; @Volkov94; @Osipov96]. The first attempt to apply this model to describe the hadron part of the $\tau \to f_1 (1285) \pi^{-} \nu_{\tau}$ decay is presented in [@Vishneva14]. Though an analysis made there allows to reproduce the experimental value of branching ratio, the picture presented cannot be considered as a fully satisfactory description. Here we improve it on the following aspects. First, as opposed to [@Vishneva14] we demonstrate that the pseudoscalar channel does not give contribution to the decay $\tau \to f_1 \pi^{-} \nu_{\tau}$. The latter is a direct consequence of the anomaly structure of the corresponding quark triangle diagram. Indeed, for the pseudoscalar-exchange channel the hadron axial-vector current, $J_\mu^A$, is proportional to a gradient of the pion field: $J_\mu^A\propto\partial_\mu\pi$. If the pion turns into the $f_1\pi$ couple, the corresponding $f_1\pi\pi$ vertex vanishes because it has an anomaly structure $e_{\mu\nu\alpha\beta}\epsilon^\mu_{f_1} p_1^\nu p_2^\alpha p_3^\beta$, where the antisymmetric Levi-Civita symbol is contracted with a polarization vector of the axial-vector field, $\epsilon^\mu_{f_1}$, and with three momenta $p_1, p_2, p_3$ of particles involved. However, due to the conservation of the total momenta, there are only two independent vectors. The linear dependence between $p_1, p_2$ and $p_3$ makes the product to be zero. On the other hand, if pion turns into $a_1$ (the $\pi-a_1$ transition is described by the Lagrangian density $\propto \partial_\nu \vec\pi\vec a_1^\nu$ [@Osipov17je; @Osipov17ap; @Morais17]), the amplitude of the $a_1\to f_1\pi$ transition, which is not zero by itself $e_{\mu\nu\alpha\beta}\epsilon^\mu_{f_1} \epsilon^\nu_{a_1} p_1^\alpha p_2^\beta\neq 0$, will be changed to $\epsilon^\nu_{a_1}\to p^\nu_{\pi}$ and this vanishes the product for the reasons just mentioned above. Second, we show that it is necessary to take into account the first radially-excited state of the $a_1(1260)$-meson, that is the $a_1(1640)$ resonance. This state has not been considered in [@Vishneva14]. The evidence of this hadron resonance has been recently approved by the new data of COMPASS collaboration [@Wallner17]. To take the $a_1(1640)$ into account, we carry out the calculations in the framework of the extended NJL model [@VolkovW97; @Volkov97; @VolkovE97; @VolkovYu00; @Volkov06; @Volkov16; @Volkov17]. The model allows one to describe both the ground and the first radially-excited meson states in accord with the chiral symmetry requirements. And finally, we study the influence of the $\pi -a_1$ transitions on the amplitude and demonstrate their significance. Notice the essential difference between present calculations of $\pi -a_1$ effects, and the previous ones, presented in [@Vishneva14]. In our work we take into account the $\pi -a_1$ transition on the external pion line, that has not been done in [@Vishneva14]. On the contrary, we show that the $\pi -a_1$ transition on the virtual axial-vector line, considered in that paper, does not contribute. Thus, we properly account for $\pi -a_1$ mixing effects in the $\tau \to f_1 \pi^{-} \nu_{\tau}$ decay amplitude. As a result we obtain a reasonable theoretical description of the branching ratio of the $\tau \to f_1\, \pi^{-} \nu_{\tau}$ decay. We believe that as soon as new experimental data on the spectral functions of this process will be available a more detailed test of the extended NJL model will be possible. The material of this paper is distributed as follows. In Section 2 we establish some convenient notations and review the basic properties of the extended NJL model including its Lagrangian, which is the basis of all our calculations. In Section 3 we give a derivation of the $\tau \to f_1 \pi^{-} \nu_{\tau}$ decay amplitude which does not take into account the $\pi -a_1$ mixing effects. To trace the numerical effect coming out of $\pi-a_1$ transitions, we intentionally delayed this material up to the Section 4, where we also calculate the spectral distribution of $f_1\pi$ pair and find the two particle decay width of $a_1(1640)\to f_1(1285)\pi$. In Section 5 we summarize our results and make conclusions. In the Appendix we present the integral form of the amplitude describing the $\tau \to f_1 \pi^{-} \nu_{\tau}$ decay and collect some general expressions for the quark-loop-integrals considered. The quark-meson Lagrangian of the extended NJL model {#sec:1} ==================================================== Let us review the main ingredients of the model which we apply to study the $\tau \to f_1\, \pi^{-} \nu_{\tau}$ decay. Its dynamics is described by the quark Lagrangian density ${\cal L}(x)$ with the effective $U(2)_L\times U(2)_R$ chiral symmetric four-quark interactions $$\begin{aligned} \label{qint} {\cal L}(x)&=&\bar q(x) \left(i\hat\partial-m^0 \right)q(x)+{\cal L}_{int}(x), \nonumber \\ {\cal L}_{int}(x)&=&\frac{G_S}{2}\sum_{i=1}^{2}\sum_{a=0}^{3}\left\{[j^S_{a(i)}(x)]^2+[j^P_{a(i)}(x)]^2\right\} \nonumber \\ &-&\frac{G_V}{2}\sum_{i=1}^{2}\sum_{a=0}^{3}\left\{[j^V_{a(i)}(x)]^2+[j^A_{a(i)}(x)]^2\right\},\end{aligned}$$ where $q=(u,d)$ are the up and down current quark fields of mass $m^0=(m^0_u, m^0_d)$; $G_S$ is a coupling determining the strength of the four-quark interactions of scalar and pseudoscalar types, $G_V$ is a coupling of vector and axial-vector interactions. The general form of quark currents is $$j^{S,P,V,A}_{a(i)}(x)=\!\!\int\! d^4x_1 d^4 x_2 \bar q(x_1) F^{S,P,V,A}_{a(i)}(x; x_1,x_2)q(x_2).$$ At each value of index $i$ the sum over $a$ in (\[qint\]) is invariant with respect to chiral transformations. Thus, there are four independent $U(2)_L\times U(2)_R$ invariant interactions. The local interactions with $i=1$ represent the conventional NJL type model (see, for instance, [@Volkov86; @Bijnens93]) which describes the physics of ground-state mesons with quantum numbers $J^{PC}=0^{++}, 0^{-+}, 1^{--}$ and $1^{++}$. The covariant form factors of local interactions are given by $$\begin{aligned} &&F^{S,P,V,A}_{a(1)}(x; x_1,x_2)=F^{S,P,V,A}_{a} \delta(x-x_1)\delta(x-x_2), \nonumber \\ &&F^{S,P,V,A}_{a} =\tau_a\, (1,i\gamma_5,\gamma_\mu, \gamma_\mu\gamma_5 ),\end{aligned}$$ where the flavour matrices $\tau_a=(1, \vec\tau)$, with the standard notation of the isospin Pauli matrices $\vec\tau$. The non-local part of the Lagrangian density $i=2$ represents the first radial excitations of ground-states. The corresponding covariant non-local form factors have been constructed in [@VolkovW97] with the use of the following transformation $$\begin{aligned} F^{S,\ldots}_{a(2)}(x; x_1,x_2)&=&\!\int\!\frac{d^4k}{(2\pi )^4}\frac{d^4p}{(2\pi )^4} \exp i\left[ p\left(x-\frac{x_1+x_2}{2}\right) \right.\nonumber \\ &-& \left. k\left(x_1-x_2\right)\right] F^{S,\ldots}_{a(2)}(k_\perp ),\end{aligned}$$ where $k_\mu$ is a relative momentum of the quark-antiquark pair, and $p_\mu$ is a 4-momentum of their center of mass reference frame, i.e. a meson momentum. The total momentum $p_\mu$ of the composite hadron provides a naturally preferred direction which forms the basis for a covariant three dimensional support to the interaction kernel $F^{S,\ldots}_{a(2)}(k_\perp )$ which is implemented here in accord with the concept of bilocal fields. The dependence of the interaction kernel on the transverse part of the quark momentum $k$, i.e. $k_\perp =k-(kp)p/p^2$ is a consequence of a subsidiary condition [@Markov40; @Yukawa50], which as it was shown in [@Lukierski77] is equivalent to a ’gauge principle’ and expresses the redundance of the longitudinal component of the relative momentum $k$ for the physical interaction between the quark-antiquark constituents. The covariant kernels of non-local interactions are $$\label{ff2} F^{S,P,V,A}_{a(2)}(k_\perp^2)=F^{S,P,V,A}_{a} c^{S,P,V,A} f (k_\perp^2 )\theta (\Lambda_3-\|k_\perp \|).$$ The coefficients $c^{S,P,V,A}$ renormalize the couplings of four-quark interactions $G_S$ and $G_V$ increasing a strength of these forces for the non-local interactions. They are fixed from the empirical values of meson masses. The step function $\theta (\Lambda_3-\|k_\perp \|)$, where $\Lambda_3$ is a covariant cutoff, restricts the integration over relative momentum of a bound quark-antiquark pair to the size of the bag. The function $$\label{fk} f (k_\perp^2) =1+d \|k_\perp \|^2, \quad \|k_\perp \|=\sqrt{-k_\perp^2},$$ being a Lorentz scalar, can be calculated in any convenient reference frame. In the following we use the instantaneous rest frame of the meson. In this case $k_\perp =(0, \vec k)$, and $f (k_\perp^2)= 1+d \vec k^2\equiv f(\vec k^2)$. The slope parameter $d$ is fixed by the requirement that the numerical values of the quark condensate and constituent quark masses are not changed due to an inclusion of the radially excited states. Equivalently, one can require that the single excited quark-antiquark states averaged over vacuum are vanishing (an absence of the vacuum tadpoles). It gives $d = -1.784 \,\textrm{GeV}^{-2}$. There is a simple argument in favour of this requirement: the non-local bound states do not survive in the large $N_c$ limit, therefore they cannot affect the main characteristics of the QCD ground state. The form factor $f (k_\perp^2)$ has for $d\leq \Lambda_3^{-2}$ the form of an excited-state wave function, with a node in the interval $0\leq \|k_\perp \|\leq \Lambda_3$. In (\[fk\]) we consider only the first two terms in a series of polynomials in $\|k_\perp \|^2$; inclusion of higher excited states would require polynomials of higher degree. The Lagrangian density (\[qint\]) has to be bosonized. The boson variables can be introduced in the two stages. On the first stage, the four quark interactions can be equivalently rewritten as a Yukawa type quark-antiquark-meson interactions. In this form the model Lagrangian has a structure of a linear sigma model. On the second stage, one should integrate out the quark fields completely. Exactly this way the Lagrangian of the extended NJL model has been worked out in [@Volkov97; @VolkovYu00]. Its numerous applications were reviewed in [@Volkov17]. Let us stress the main features of such calculations: (a) The model reveals the mechanism of spontaneous chiral symmetry breaking. Starting from some critical value of coupling $G_S\geq G_{crit}$ the Wigner-Weyl ground state of the system is changed to the Nambu-Goldstone phase. The transition is described by the gap equation. In particular, the current quark mass $m^0$ is replaced on the constituent quark mass $m$. It is assumed that the non-local sector of the model does not contribute to $m$, and does not affect the value of the quark condensate. This assumption help us to fix a slope parameter $d$; (b) Several mixing effects take place at the level of free Lagrangian. First, as a consequence of the phase transition, the mixing between $J^{P}=0^-$ and $J^P=1^+$ states, the so-called $\pi -a_1$ transitions, occurs. Second, there are mixings between ground and excited states with the same quantum numbers; (c) The effective meson vertices and corresponding coupling constants follow from the one-quark-loop calculations. It means that one should separate divergences in a regularized form, and renormalize the meson fields. All these effects are taken into account in our calculations. The description of collective bound states can be facilitated if we introduce, as it was discussed above, the set of bosonic variables and pass to the semi-bosonized effective meson action which is responsible for the $\tau \to f_1\, \pi^{-} \nu_{\tau}$ decay, namely, the part which describes both the interactions of the ground pseudoscalar $\pi^{\pm}$, axial-vector $a_1^{\pm}(1260)$, $f_1(1285)$ mesons and their first radially excited states with light constituent quarks. In momentum representation the action takes the form $$\begin{aligned} \label{Lagrangiane} S &=&\!\int\!\frac{d^4p}{(2\pi )^4}\bar q(p) \left(\hat p-m \right)q(p)+ \Delta S_{mass}+ \Delta S_{int}, \nonumber \\ \Delta S_{int} &=&\! \int\!\frac{d^4p}{(2\pi )^4}\!\int\!\frac{d^4k}{(2\pi )^4}\, \bar{q}\left(k+\frac{p}{2}\right) \nonumber \\ &\times& \left[ \frac{1}{2} \gamma^{\mu}\gamma^{5}\left(A_{f_{1}}f_{1\mu}(p) + B_{f_{1}}f^{'}_{1\mu}(p)\right) \right.\nonumber \\ &+& \frac{1}{2} \gamma^{\mu}\gamma^{5} \vec\tau \left(A_{a_{1}}\vec a_{1\mu}(p) + B_{a_{1}}\vec a'_{1\mu}(p)\right) \nonumber \\ &+& \left. \gamma^{5} \vec\tau \left(A_{\pi}\vec \pi (p) + B_{\pi}\vec \pi'\right)\right]q\left(k-\frac{p}{2}\right) .\end{aligned}$$ We shall not discuss here the mass part of the action $\Delta S_{mass}$ (see, for instance, [@Volkov17] or references therein). Our notations are as follows: $f_{1\mu}$, $\vec a_{1\mu}$ and $\vec \pi $ are the fields corresponding to the axial-vector and pseudoscalar mesons, the related excited states are marked with a prime. The constituent quark fields, $q$, have equal masses $m_{u} = m_{d} = 280$ MeV [@VolkovYu00; @Volkov17]. The summation over colour index is assumed. The couplings of the ground state meson $M=(\pi, a_1, f_1)$ with quarks, $A_M$, and the corresponding couplings of its radially excited state $M'=(\pi', a_1', f_1')$, $B_M$, can be written in a form $$\begin{aligned} \label{coefficients} &&\!\!\!\! A_M = A_M^0\left[g_M \sin\theta_{M}^+ +g'_M f(k_\perp^2)\sin\theta_M^ -\right]\! , \nonumber \\ &&\!\!\!\! B_M =-A_M^0\left[g_M \cos\theta_{M}^+ +g'_{M} f (k_\perp^2)\cos\theta_M^ -\right].\end{aligned}$$ These expressions are obtained as a result of several procedures. The renormalization factors $g_M$ and $g_M'$ eliminate the divergent parts of the amplitudes describing the one-quark-loop self-energy transitions $M\to M$ and $M'\to M'$ correspondingly. In the NJL model they are fully determined by the requirement that such transitions generate a free Lagrangian of meson fields. Their specific values are expressed through the divergent quark-loop integrals $$I_{2, n} = -i\frac{N_{c}}{(2\pi)^{4}}\int\frac{f^{n}(k_\perp^2)}{(m^{2} - k^2)^{2}}\theta(\Lambda_{3}^2 - k_\perp^2) \mathrm{d}^{4}k,$$ which are regularized by imposing the three-dimensional cutoff on $\|k_\perp\|\leq \Lambda_3 = 1.03$ GeV (in accord with eq.(\[ff2\])), after carrying out the $k_0$ integration in the self-energy diagrams [@Volkov97; @Volkov17]. In particular, we have $$\begin{aligned} &&g_{f_1} = g_{a_{1}}=\sqrt{\frac{3}{2I_{2,0}}}, \quad g_{\pi}=\sqrt{\frac{Z_\pi}{4I_{2,0}}}, \nonumber\\ &&g'_{f_1} = g'_{a_1} = \sqrt{\frac{3}{2I_{2,2}}}, \quad g'_{\pi}=\sqrt{\frac{1}{4I_{2,2}}}.\end{aligned}$$ Note that $Z_{\pi}$ is the factor induced by a diagonalization of the free Lagrangian for ground state mesons (the $\pi -a_{1}$ transitions). To avoid this mixing the unphysical axial-vector fields $\vec a_{1\mu}(x)$ should be redefined $$\label{pa-trans} \vec a_{1\mu}(x)=\vec a_{1\mu}^{phys}(x)+\sqrt{\frac{2Z_\pi}{3}}\kappa m\partial_\mu\vec\pi (x),$$ where $$\begin{aligned} Z_{\pi} = \left(1 - \frac{6m^{2}}{M^{2}_{a_{1}}}\right)^{-1} \approx 1.45. \end{aligned}$$ A dimensional parameter $\kappa$ is fixed by requiring that the free meson Lagrangian does not contain the unphysical $\vec\pi-\vec a_{1\mu}$ transitions; it gives $\kappa =3/M_{a_1}^2$, where $M_{a_1}$ is a mass of the $a_1(1260)$ meson, $M_{a_1}=1230\pm 40\,\mbox{MeV}$. The replacement (\[pa-trans\]) does not affect the Green function of the axial-vector field, but it affects the kinetic term of a pion free Lagrangian. Consequently, the pion field wave-function is additionally renormalized by the factor $\sqrt{Z_\pi}$ in (\[pa-trans\]). There are other mixings between $J^P=1^+$ and $J^P=0^-$ states. The $\pi' -a_1'$ mixing does not contribute to the $\tau\to f_1\pi\nu_\tau$ decay. In the following we neglect the $a_1'\to\pi$ mixing due to a heavy mass of the $a_1'$ state which is associated with $a_1(1640)$ meson. Eq. (\[coefficients\]) includes the angles, $\theta_{M}^{0}$ and $\theta_{M}$ [@Volkov97; @Volkov17]. These parameters appear due to $M\to M'$ mixing ($\theta_{M}^{0}$), and as a result of diagonalization, which aimed to avoid such a mixing ($\theta_{M}$). We arranged them in the following combinations: $A_M^0=1/\sin(2\theta_{M}^{0})$, $\theta_M^\pm = \theta_{M} \pm \theta_{M}^{0}$. The numerical values of mixing angles follow from the meson mass formulae. They are $$\begin{aligned} \label{angles} && \theta_{f_{1}} = \theta_{a_{1}} = 81.8^{\circ}, \quad \theta_{\pi} = 59.48^{\circ}, \nonumber\\ && \theta_{f_{1}}^{0} = \theta_{a_{1}}^{0} = 61.5^{\circ}, \quad \theta_{\pi}^{0} = 59.12^{\circ}. \end{aligned}$$ Notice, that all parameters of the model are determined from the empirical data, which are not related with the characteristics of the processes considered in this work. Therefore, our estimations can be used to test the predictive power of the extended NJL model. The decay amplitude without $\pi -a_{1}$ transitions {#sec:2} ==================================================== The decay amplitude of the process $\tau \to f_1(1285)\, \pi^- \nu_\tau$ is described by two types of diagrams which are shown in Fig.\[Contact\], and Fig.\[Intermediate\] (the $\pi -a_1$ transitions are neglected there). The first diagram describes the so-called direct contribution to the amplitude. This means that $W$-boson decays directly to the final products of the reaction, $f_1\pi$ - pair, i.e. without a resonance exchange. The latter is taken into account by the second diagram. Notice, that hadrons are alway interact through the one-quark-loop in accord with our action (\[Lagrangiane\]). In addition, we would like to point out that these loop integrals are expanded in momenta of external fields, and only the divergent parts are kept (in the case of anomalies, which lead to the finite result we take the first term of such expansion). This approximation is qualitatively justified by the $1/N_c$ expansion which states that meson physics in the large $N_c$ limit is described by the local vertices [@Hooft74; @Witten79]. It is also well known from the sigma model that divergent parts of radiative corrections have a strictly chiral-symmetric structure [@Eguchi78]. The amplitude corresponding to the graphs shown in Fig.\[Contact\] and Fig.\[Intermediate\] is presented in an Appendix. Here we show the result obtained after derivative expansion of quark vertices, as it was discussed above. Note also that from (\[angles\]) it follows that the angle $\theta_\pi^-$ is small, and $\theta_\pi\simeq \theta^0_\pi$. Neglecting $\theta_\pi^-$, one concludes from (\[coefficients\]) that $A_{\pi} \simeq g_{\pi}$. As a result, an amplitude takes the form $$\begin{aligned} \label{M-without} \mathcal{M} & = & 4mG_{F}V_{ud}\,g_\pi l^\mu e_{\mu\nu\alpha\beta}\epsilon^{\nu}(p_{f_1})p_{f_1}^\alpha p_{\pi}^\beta\left\{ I_{3}^{A_{f_{1}}} \right. \nonumber \\ &+& \frac{C_{a_1}}{g_{a_{1}}} I_{3}^{A_{f_{1}}A_{a_{1}}} \frac{s - 6m^2}{M_{a_{1}}^2 - s - i \sqrt{s}\Gamma_{a_{1}}} \nonumber\\ & + & \left. \frac{C_{a'_1}}{g_{a_{1}}} I_{3}^{A_{f_{1}}B_{a_{1}}} \frac{s - 6m^2}{M_{a'_{1}}^2 - s - i \sqrt{s}\Gamma_{a'_{1}}} \right\}\!.\end{aligned}$$ Here $G_F = 1.1663787(6) \times 10^{-11}$ MeV$^{-2}$ is the Fermi coupling constant; $V_{ud} = 0.97417\pm 0.00021$ is the Cabbibo-Kobayashi-Maskawa quark-mixing matrix element; $l_{\mu}=\bar\nu_\tau(Q')\gamma_\mu(1-\gamma_5)\tau (Q)$ is a lepton current, where $Q$ and $Q'$ are momenta of the tau-lepton and neutrino; $\epsilon_{\nu}(p_{f_{1}})$ is a polarization vector of the $f_1$ meson with the momentum $p_{f_1}$; $s = (p_{f_1} + p_{\pi})^2$ is a square of the invariant mass of the $f_1\pi$-pair, the masses and widths of resonances are $M_{a_{1}} = 1230$ MeV, $M_{a'_{1}} = 1640$ MeV, $\Gamma_{a_{1}} \approx 400$ MeV, $\Gamma_{a'_{1}} = 254$ MeV [@Patrignani16]. The factors $C_{a_1}$ and $C_{a'_1}$ are the remaining of functions $A_{a_1}$ and $B_{a_1}$ (see eq.(\[coefficients\])) after integration over an internal momentum in the $W-a_1$ and $W-a_1'$ quark-loops (see Fig.\[Intermediate\]) correspondingly. $$\begin{aligned} C_{a_1} & = & A^0_{a_1} \left( \sin\theta_{a_1}^+ +R_V\sin\theta_{a_1}^- \right), \nonumber \\ C_{a'_1} & = & -A^0_{a_1} \left(\cos\theta_{a_1}^+ +R_V\cos\theta_{a_1}^-\right), \nonumber \\ R_{V} & = & \frac{I_{2,1}}{\sqrt{I_{2,0}I_{2,2}}}\, .\end{aligned}$$ The integrals $I_3^{A_{f_1}\dots}$ come out from the evaluation of the one-loop-quark triangle diagrams of Fig.\[Contact\] and Fig.\[Intermediate\]. They correspond to the case $n=3$ of the general expression (in the next section the case $n=4$ will be also required) $$\label{In} I_{n}^{A_{M}B_{M}\dots} = \frac{-iN_c}{(2\pi)^4}\!\int\! \frac{A_M (k_\perp^2) B_M (k_\perp^2)\dots}{(m^2 - k^2)^n}\theta(\Lambda_3^2 - k_\perp^2) \mathrm{d}^{4}k,$$ where $A_{M}, B_{M}$ are given in (\[coefficients\]). The amplitude (\[M-without\]) has a conventional form $\mathcal{M}=G_{F}V_{ud}l^\mu H_\mu$, where we have found the hadron current $H_\mu$ with the help of the extended NJL model. The square of this amplitude has a simple form $$\|\mathcal{M}\|^2=4G_{F}^2V_{ud}^2\left[2(QH)(Q'H)-H^2 (QQ')\right].$$ Thus one can easily find the decay width of the process $$\Gamma = \frac{1}{32M_\tau^3(2\pi )^3}\!\! \int\limits_{(M_{f_1}+M_\pi )^2}^{M_\tau^2}\!\!\!\!\! ds \int\limits_{t_-}^{t_+} dt \,\| {\cal M} \|^2$$ by performing the numerical integration over kinematical variables $t=(Q-p_{f_1})^2$ and $s=(p_{f_1}+p_\pi )^2$. Here a boundary of the physical region at fixed value of $s$ belongs to the interval $t_-\leq t\leq t_+$, where $$\begin{aligned} &&t_\pm = \frac{1}{2} \left( t_0\pm\sqrt{D}\right), \nonumber \\ && t_0= M_\tau^2+M_{f_1}^2+M_\pi^2 -s -\frac{M_\tau^2}{s}\left(M_{f_1}^2-M_\pi^2\right), \nonumber \\ &&\sqrt{D}=\frac{1}{s}\left(M_\tau^2-s\right)\sqrt{\lambda (s,M_{f_1}^2,M_\pi^2 )}, \nonumber \\ &&\lambda (x, y, z )=[ x - (\surd y-\surd z )^2] [ x - (\surd y+ \surd z )^2 ]. \end{aligned}$$ Finally we arrive at the following result for the branching ratio Br$(\tau\to f_1\pi^-\nu_\tau )=6.04 \times 10^{-4}$. It can be schematically presented in terms of individual contributions as follows $$\begin{aligned} &&10^{4}\times \mbox{Br}(\tau\to f_1\pi^-\nu_\tau)= 6.04 \nonumber\\ &&= \|c\|^2+\|a_1\|^2+\|a'_1\|^2+2\mbox{Re}(a_1 c^+a'_1 c^+a'_1a_1^) \nonumber\\ &&= 2.25+8.43+0.78-7.34+0.53+1.39.\end{aligned}$$ where $c, a_1, a'_1$ represent the contributions from the contact (direct) term, and from $a_1$ and $a'_1$ exchanges correspondingly. One can see that if one neglects the $\pi -a_1$ transitions (as we did here) the result overestimates the experimental value Br$(\tau \to f_1\, \pi^{-} \nu_{\tau})=(3.9\pm 0.5) \times 10^{-4}$. The effect of $\pi -a_{1}$ transitions ====================================== The replacement (\[pa-trans\]) in the quark-meson Lagrangian originates new vertices of the axial-vector type with the gradient of the pion field. This can be equivalently considered as a creation of the final pion via old $a_1$ meson field ($\pi -a_1$ transitions). To take them into account one has to append the diagrams shown in Fig. \[pi-a1-contact\] and Fig. \[pi-a1-intermediate\]. The corresponding additional contribution to the amplitude is given in the Appendix (see eq.(\[add-pi-a1\])). It can be simplified after the following observation. Let us consider the trace of the quark triangles corresponding to these diagrams $$\mbox{tr}[\gamma^{\nu}\gamma^{5}(\hat{k} + \hat{p}_{f_{1}} + m)\gamma^{\mu}\gamma^{5}(\hat{k} - \hat{p}_{\pi} + m) \gamma^{\lambda}\gamma^{5}(\hat{k} + m)].$$ It can be written as a sum $$\begin{aligned} \label{trace} && \mbox{tr}\,[\gamma^{\nu}(\hat{k} + \hat{p}_{f_{1}} + m)\gamma^{\mu}\gamma^{5}(\hat{k} - \hat{p}_{\pi} + m)\gamma^{\lambda}(\hat{k} + m) \nonumber\\ && - 2m \gamma^{\nu}(\hat{k} + \hat{p}_{f_{1}} + m)\gamma^{\mu}\gamma^{5}(\hat{k} - \hat{p}_{\pi} + m)\gamma^{\lambda}].\end{aligned}$$ It is not difficult to see that the first term coincides with a trace from the triangle of the process $f_{1}(1285) \to \rho \gamma$ considered in our previous work [@Osipov17]. After integration over loop-momentum $k$ in the corresponding quark triangle expression (see, for instance, eq.(\[afa\])) one finds that it contributes to the amplitude as $$\begin{aligned} &&e^{\mu\nu\lambda}_{\, . \, . \ . \, \alpha} \, p^\alpha_{\pi} \left(2 p_{f_{1}}^{2} + p_\pi p_{f_{1}}\right) - e^{\mu\nu\lambda}_{\, . \, . \ . \, \alpha}\, p^\alpha_{f_{1}} \! \left( p_\pi p_{f_{1}} \right) \nonumber \\ && + e^{\mu\nu}_{\, . \, .\, \alpha\beta}\, p^\alpha_{\pi} p^\beta_{f_{1}} p_{f_{1}}^{\lambda} - e^{\lambda\mu}_{\, . \, .\, \alpha\beta}\, p^\alpha_{f_1} p^\beta_{\pi} p_{\pi}^{\nu}.\end{aligned}$$ Due to $\pi -a_1$ transitions this result is multiplied by the 4-momentum $p_{\pi}^{\lambda}$. That vanishes it. The second term in eq. (\[trace\]) is easily calculated, giving $-8im^2e^{\mu\nu\lambda\alpha} (2k+p_{f_1}-p_\pi )_\alpha$. The result of its integration over $k$ depends on the structure of the integrand. In the case considered, we have three similar structures which differ only by the functions $A_M$ and $B_M$ at the vertices of anomalous triangle diagrams. These alternatives are absorbed by the corresponding coefficient $I_{4}^{A_{f_1}A_{\alpha_1}\dots}$ in the following common structure $$\begin{aligned} 8 m^{4} I_{4}^{A_{f_1}A_{\alpha_1}\dots} e^{\mu\nu\lambda}_{\, . \, . \ . \, \alpha}\, p^\alpha_{f_{1}},\end{aligned}$$ where the integrals $I_{4}^{A_{f_1}A_{\alpha_1}\dots}$ are given by (\[In\]). Taking into account these remarks, one obtains the corrections induced by the $\pi -a_1$ transitions to the total amplitude of the $\tau \to f_1 \pi^{-} \nu_{\tau}$ decay. As a result the total amplitude is $$\begin{aligned} \label{Mtot} && \mathcal{M}_{tot} = G_{F}V_{ud}l^\mu 4mg_\pi \left\{ \left[I_{3}^{A_{f_{1}}} - \frac{C_{a_1}6m^4}{g_{a_1}M_{a_1}^2} I_4^{A_{f_{1}}A_{a_{1}}} \right] \right. \nonumber\\ &&+ \frac{C_{a_1}}{g_{a_1}} \left[I_{3}^{A_{f_{1}}A_{a_{1}}} - \frac{C_{a_1}6m^4}{g_{a_1}M_{a_1}^2} I_{4}^{A_{f_{1}}A_{a_{1}}A_{a_{1}}}\right] \nonumber \\ &&\times \frac{s - 6m^2}{M_{a_{1}}^2 - s - i \sqrt{s}\Gamma_{a_{1}}} \nonumber\\ &&+ \frac{C_{a'_1}}{g_{a_{1}}} \left[I_{3}^{A_{f_{1}}B_{a_{1}}} - \frac{C_{a_1}6m^4}{g_{a_1}M_{a_1}^2} I_{4}^{A_{f_{1}}B_{a_{1}}A_{a_{1}}}\right] \nonumber\\ &&\times\left.\frac{s - 6m^2}{M_{a'_{1}}^2 - s - i \sqrt{s}\Gamma_{a'_{1}}} \right\} e_{\mu\nu\alpha\beta} \epsilon^{\nu}(p_{f_{1}}) p^\alpha_{f_{1} }p^\beta_{\pi}.\end{aligned}$$ Here, in the first square brackets, the contact contribution is presented. The second and third square brackets contain the contributions of $a_1(1260)$ and $a_1(1640)$ resonances correspondingly. The numerical integration over the $ f_1\pi\nu_\tau$ three-body phase space with the amplitude (\[Mtot\]) gives the branching ratio Br$(\tau\to f_1\pi^-\nu_\tau )=3.98 \times 10^{-4}$. The result can be schematically presented in terms of the individual contributions as follows $$\begin{aligned} &&10^{4}\times \mbox{Br}(\tau\to f_1\pi^-\nu_\tau)= 3.98 \nonumber\\ &&= \|c\|^2+\|a_1\|^2+\|a'_1\|^2+2\mbox{Re}(a_1 c^+a'_1 c^+a'_1a_1^) \nonumber\\ &&=1.62+5.92+0.45-5.22+0.33 +0.88,\end{aligned}$$ where $c, a_1, a'_1$ represent the contributions from the contact term, and from $a_1$ and $a'_1$ exchanges correspondingly. One can see that if one takes into account the $\pi -a_1$ transitions the branching ratio is in a good agreement with the experimental value Br$(\tau \to f_1\, \pi^{-} \nu_{\tau})=(3.9\pm 0.5) \times 10^{-4}$. The differential decay distribution, $d\Gamma /d \sqrt s$, is shown in Fig. \[distribution\]. It reaches its maximum value near the $a_1(1640)$ resonance mass shell. Therefore, it appears worth while to estimate the decay width of the state, that dominates the spectral function. The amplitude of the $a_1(1640)\to f_1(1285)\pi$ decay is given by $$\begin{aligned} \mathcal{M} & = & 4img_\pi e_{\mu\nu\alpha\beta }\, \epsilon^{\mu}(p_{a_{1}})\epsilon^{\nu}(p_{f_{1}}) p^\alpha_{f_{1}}p^\beta_{\pi} \nonumber \\ &\times& \left\{I_{3}^{A_{f_{1}}B_{a_{1}}} - \frac{C_{a_1}6m^4}{g_{a_{1} M_{a_1}^2}} I_{4}^{A_{f_{1}}B_{a_{1}}A_{a_{1}}} \right\}. \end{aligned}$$ It follows then that $$\Gamma \left[a_1(1640)\to f_1(1285)\pi \right]=14.1\,\mbox{MeV}.$$ The future measurements will show how good is the extended NJL model in its predictions here. Probably, the tau decay mode studied in this work can serve us with a detailed information on the nature of $a_1(1640)$ state. Conclusions =========== The main purpose of our calculations was to apply the extended NJL model to the decay $\tau\to f_1(1285)\,\pi^- \nu_\tau$. Presently available phenomenological data on this mode give a rare opportunity to test the model. As a result we have found that hadronic part of the amplitude is sensitive to the four types of different contributions which are equally important. These are the contact interaction, the exchange by the $a_1(1260)$ meson, the exchange by the first radially excited state, $a_1(1640)$, and the pion creation by the intermediate $a_1(1260)$ meson (the $\pi a_1$-transitions). Indeed, the contact term alone gives the value Br$\ =2.25\times 10^{-4}$. The $a_1(1260)$ ground state exchange alone gives a larger number Br$\ =8.43\times 10^{-4}$. The sum of these two contributions is Br$\ =3.34\times 10^{-4}$. The radially excited state increase this value up to Br$\ =6.04\times 10^{-4}$. And finally taking into account the $\pi -a_1$ transitions we come to the final number Br$\ =3.98\times 10^{-4}$ which agrees with presently known empirical values. Our result shows that both $a_1(1260)$ and $a_1(1640)$ exchanges are very important for the description of data, and it indicates that many ideas about description of the first radial exited meson states in the NJL model seems to be correct. Let us stress that we did not introduce any new parameters to get the consistent result. All model parameters were fixed with the use of other input data, which are not related with the process $\tau\to f_1(1285)\,\pi^- \nu_\tau$. Note, that this is not the only case where the model predictions correspond to the empirical values. We refer to the recent review [@Volkov17], which contains many other examples. Those include the meson production processes in $e^+e^-$ collisions and tau decay modes. Of some relevance may be the fact that our result does not leave the place for the contribution of a new axial-vector resonance $a_1(1420)$ observed recently by the COMPASS collaboration [@Adolph15]. Our study shows that even if there is a contribution due to $a_1(1420)$ exchange to the process discussed here, this contribution is negligible. It indicates that most probably $a_1(1420)$ is not the $q\bar q$ state. Some reasonings in favour of multi-quark structure of this state are recently given in [@Ivanov17]. Acknowledgments {#acknowledgments .unnumbered} =============== We would like to thank A. B. Arbuzov for his interest to this work and useful discussions. Appendix {#appendix .unnumbered} ======== Here we present the general expressions for the amplitude of $\tau\to f_1(1285)\pi^-\nu_\tau$ decay shown in Figs.\[Contact\]-\[pi-a1-intermediate\]. In the text we make derivative expansions of these expressions to obtain effective meson vertices in the long wavelength approximation. We consider the process $$\tau (Q)\to \nu_\tau (Q') f_1(p_{f_1}) \pi (p_\pi )$$ with the quantities in the parentheses denoting the four moments of the particles. The amplitude, corresponding to diagrams of Fig.\[Contact\]-\[Intermediate\], has the following form $$\mathcal{M} = G_{F} V_{ud} l_{\mu} \mathcal{H}^\mu ,$$ where a hadron current $\mathcal{H}^\mu$ is $$\begin{aligned} \mathcal{H}^\mu & = & \frac{g_{\pi}}{2} \epsilon_{\nu}(p_{f_{1}}) \left\{ I_{Wf_{1}\pi 1}^{\mu\nu} + I_{Wf_{1}\pi 2}^{\mu\nu} \right. \nonumber\\ & + & \frac{I_{Wa_{1}}^{\mu\lambda}\left(I_{a_{1}f_{1}\pi1}^{\lambda\nu} + I_{a_{1}f_{1}\pi2}^{\lambda\nu}\right)}{M_{a_{1}}^2 - s - i \sqrt{s}\Gamma_{a_{1}}} \nonumber \\ & + & \left. \frac{I_{Wa'_{1}}^{\mu\lambda}\left(I_{a'_{1}f_{1}\pi1}^{\lambda\nu} + I_{a'_{1}f_{1}\pi2}^{\lambda\nu}\right) }{M_{a'_{1}}^2 - s - i \sqrt{s}\Gamma_{a'_{1}}} \right\}.\end{aligned}$$ The first two integrals $I_{W f_1\pi 1}^{\mu\nu}$ and $I_{Wf_{1}\pi 2}^{\mu\nu}$ describe the direct contribution from the transition $W^\mu\to f_1^\nu\pi$ generated by the triangle quark loop (see Fig.\[Contact\]). In accord with two different directions for the loop momenta we specify them by indices 1 (clockwise) and 2 (counter-clockwise). Their expressions are $$\begin{aligned} && I_{Wf_{1}\pi1}^{\mu\nu} = N_c \int \frac{\mathrm{d}^4k}{(2\pi)^4} A_{f_1}(k_\perp^2) \\ &&\times\frac{ \mbox{tr}[\gamma^\nu\gamma^5(\hat{k} + \hat{p}_{f_{1}} + m)\gamma^{\mu}\gamma^{5}(\hat{k} - \hat{p}_{\pi} + m)\gamma^{5}(\hat{k} + m)]}{[(k + p_{f_{1}})^2 - m^2][(k - p_{\pi})^2 - m^2](k^2 - m^2)}, \nonumber \end{aligned}$$ $$\begin{aligned} &&I_{Wf_{1}\pi2}^{\mu\nu} = N_c \int \frac{\mathrm{d}^4k}{(2\pi)^4} A_{f_{1}} (k_\perp^2) \\ &&\times\frac{\mbox{tr}[\gamma^{5}(\hat{k} + \hat{p}_{\pi} + m)\gamma^{\mu}\gamma^{5}(\hat{k} - \hat{p}_{f_{1}} + m)\gamma^{\nu}\gamma^{5}(\hat{k} + m)]}{[(k - p_{f_{1}})^2 - m^2][(k + p_{\pi})^2 - m^2](k^2 - m^2)} . \nonumber\end{aligned}$$ The other two integrals $I_{Wa_1}^{\mu\lambda}$ and $I_{Wa'_1}^{\mu\lambda}$ describe the $W^\mu\to a_1^\lambda$ and $W^\mu\to {a'_1}^\lambda$ transitions correspondingly (see Fig.\[Intermediate\]). The first of them is equal to $$\begin{aligned} \label{IWa} &&I_{Wa_{1}}^{\mu\lambda} = N_c \int \frac{\mathrm{d}^4k}{(2\pi)^4}A_{a_{1}} (k_\perp^2) \nonumber \\ &&\times\frac{ \mbox{tr} [\gamma^{\lambda}\gamma^{5}(\hat{k} + \frac{\hat{q}}{2} + m)\gamma^{\mu}\gamma^5(\hat{k} - \frac{\hat{q}}{2} + m)]}{[(k + \frac{q}{2})^2 - m^2][(k - \frac{q}{2})^2 - m^2]}, \end{aligned}$$ with $q=Q-Q'=p_{f_1}+p_\pi$, and $q^2=s$. The second one can be obtained from (\[IWa\]) by the replacement $A_{a_{1}} (k_\perp^2) \to B_{a_{1}} (k_\perp^2)$. The anomalous quark triangle integrals $I_{a_{1}f_{1}\pi1}^{\lambda\nu}$ and $I_{a_{1}f_{1}\pi2}^{\lambda\nu}$ of Fig.\[Intermediate\] differ from other similar pair $I_{a'_{1}f_{1}\pi1}^{\lambda\nu}$ and $I_{a'_{1}f_{1}\pi2}^{\lambda\nu}$ by the replacement $A_{a_{1}}(k_\perp^2)\to B_{a_{1}}(k_\perp^2)$. Therefore, we give here only the first two expressions $$\begin{aligned} && I_{a_{1}f_{1}\pi1}^{\lambda\nu} = N_c \int \frac{\mathrm{d}^4k}{(2\pi)^4}A_{a_{1}} (k_\perp^2) A_{f_{1}}(k_\perp^2) \\ &&\times\frac{\mbox{tr} [\gamma^{\nu}\gamma^{5}(\hat{k} + \hat{p}_{f_{1}} + m)\gamma^{\lambda}\gamma^{5}(\hat{k} - \hat{p}_{\pi} + m)\gamma^{5}(\hat{k} + m)]}{[(k + p_{f_{1}})^2 - m^2][(k - p_{\pi})^2 - m^2](k^2 - m^2)},\nonumber\end{aligned}$$ $$\begin{aligned} &&I_{a_{1}f_{1}\pi2}^{\lambda\nu} = N_c \int \frac{\mathrm{d}^4k}{(2\pi)^4} A_{a_{1}} (k_\perp^2) A_{f_{1}}(k_\perp^2) \\ &&\times\frac{ \mbox{tr}[\gamma^{5}(\hat{k} + \hat{p}_{\pi} + m)\gamma^{\lambda}\gamma^{5}(\hat{k} - \hat{p}_{f_{1}} + m)\gamma^{\nu}\gamma^{5}(\hat{k} + m)]}{[(k - p_{f_{1}})^2 - m^2][(k + p_{\pi})^2 - m^2](k^2 - m^2)}. \nonumber \end{aligned}$$ Consider now the diagrams shown in Fig.\[pi-a1-contact\]-\[pi-a1-intermediate\]. They give an additional contribution to the axial current $\mathcal{H}^\mu$ by taking into account the $\pi -a_1$ transitions $$\begin{aligned} \label{add-pi-a1} \Delta\mathcal{H}^\mu &=& \frac{g_{\pi}}{4} \epsilon_{\nu}(p_{f_{1}})\frac{I_{a_{1}\pi}^{\lambda}}{M_{a_1}^2}\left\{ I_{Wf_{1}a_{1}1}^{\mu\nu\lambda} + I_{Wf_{1}a_{1}2}^{\mu\nu\lambda} \right. \nonumber\\ &+& \frac{I_{Wa_{1}}^{\mu\delta}\left(I_{a_{1}f_{1}a_{1}1}^{\delta\nu\lambda} + I_{a_{1}f_{1}a_{1}2}^{\delta\nu\lambda}\right)}{M_{a_{1}}^2 - s - i \sqrt{s}\Gamma_{a_{1}}} \nonumber\\ &+& \left. \frac{I_{Wa'_{1}}^{\mu\delta} \left(I_{a'_{1}f_{1}a_{1}1}^{\delta\nu\lambda} + I_{a'_{1}f_{1}a_{1}2}^{\delta\nu\lambda}\right)}{M_{a'_{1}}^2 - s - i \sqrt{s}\Gamma_{a'_{1}}} \right\}\end{aligned}$$ Here, the integrals with two Lorentz indices have been already defined in (\[IWa\]). The one-index integral describes the quark loop corresponding to the $\pi -a_1$ transition, i.e. $$\begin{aligned} && I_{a_{1}\pi}^{\lambda} = N_c \int \frac{\mathrm{d}^4k}{(2\pi)^4} A_{a_{1}} (k_\perp^2) \nonumber \\ &&\times \frac{ \mbox{tr} [\gamma^{5}(\hat{k} + \frac{\hat{p}_{\pi}}{2} + m)\gamma^{\lambda}\gamma^{5}(\hat{k} - \frac{\hat{p}_{\pi}}{2} + m)]}{[(k + \frac{p_{\pi}}{2})^2 - m^2][(k - \frac{p_{\pi}}{2})^2 - m^2]} . \end{aligned}$$ The integral $I_{Wf_{1}a_{1}2}^{\mu\nu\lambda}$ can be obtained from the integral $$\begin{aligned} && I_{Wf_{1}a_{1}1}^{\mu\nu\lambda} = N_c \int \frac{\mathrm{d}^4k}{(2\pi)^4} A_{f_{1}}(k_\perp^2)A_{a_{1}} (k_\perp^2) \\ && \frac{ \mbox{tr}[\gamma^{\nu}\gamma^{5}(\hat{k} + \hat{p}_{f_{1}} + m)\gamma^{\mu}\gamma^{5}(\hat{k} - \hat{p}_{\pi} + m)\gamma^{\lambda}\gamma^{5}(\hat{k} + m)]}{(k^2 - m^2)[(k + p_{f_{1}})^2 - m^2][(k - p_{\pi})^2 - m^2]} \nonumber\end{aligned}$$ by replacements $\nu\leftrightarrow \lambda$ and $p_{f_1} \leftrightarrow p_\pi$. Both describe the direct transition $W^\mu\to f_1^\nu a_1^\lambda$. The last four integrals can be obtained, for instance, from the integral $$\begin{aligned} \label{afa} && I_{a_{1}f_{1}a_{1}1}^{\delta\nu\lambda} = N_c \int \frac{\mathrm{d}^4k}{(2\pi)^4} A_{a_{1}} (k_\perp^2) A_{a_{1}}(k_\perp^2) A_{f_{1}}(k_\perp^2) \\ && \frac{\mbox{tr}[\gamma^{\nu}\gamma^{5}(\hat{k} + \hat{p}_{f_{1}} + m)\gamma^{\delta}\gamma^{5}(\hat{k} - \hat{p}_{\pi} + m)\gamma^{\lambda}\gamma^{5}(\hat{k} + m)]}{(k^2 - m^2)[(k + p_{f_{1}})^2 - m^2][(k - p_{\pi})^2 - m^2]}. \nonumber \end{aligned}$$ In this case $I_{a_{1}f_{1}a_{1}2}^{\delta\nu\lambda}$ is given by the replacement $\nu\leftrightarrow \lambda$ and $p_{f_1} \leftrightarrow p_\pi$; the integral $I_{a'_{1}f_{1}a_{1}1}^{\delta\nu\lambda}$ follows from (\[afa\]) by the replacement of one of the two functions $A_{a_{1}} (k_\perp^2) $ by $B_{a_{1}} (k_\perp^2) $; and finally, the integral $I_{a'_{1}f_{1}a_{1}2}^{\delta\nu\lambda}$ can be obtained from (\[afa\]) by the replacement of one of the two functions $A_{a_{1}} (k_\perp^2) $ by $B_{a_{1}} (k_\perp^2) $ together with $\nu\leftrightarrow \lambda$ and $p_{f_1} \leftrightarrow p_\pi$. J. P. Lees [et al.]{} (BABAR Collaboration), Phys. Rev. D [86]{} (2012) 092010. B. Aubert [et al.]{} (BABAR Collaboration), Phys. Rev. D [77]{} (2008) 112002. B. Aubert [et al.]{} (BABAR Collaboration), Phys. Rev. D [72]{} (2005) 072001. T. Bergfeld [et al.]{} (CLEO Collaboration) Phys. Rev. Lett. [79]{} (1997) 2406. S. Weinberg, Physica A [96]{} (1979) 327. J. Gasser, H. Leutwyler, Ann. Physics. [158]{} (1984) 142. J. Gasser, H. Leutwyler, Nucl. Phys. B [250]{} (1985) 465. J. Gasser, H. Leutwyler, Nucl. Phys. B [250]{} (1985) 517. G. ’t Hooft, Nucl. Phys. B [72]{} (1974) 461. E. Witten, Nucl. Phys. B [160]{} (1979) 57. J. Schwinger, Phys. Lett. B [24]{} (1967) 473. J. Wess, B. Zumino, Phys. Rev. [163]{} (1967) 1727. S. Weinberg, Phys. Rev. [166]{} (1968) 1568. S. Gasiorowicz, D. A. Geffen, Rev. Mod. Phys. [41]{} (1969) 531. U.-G. Meißner, Phys. Rep. [161]{} (1988) 213. M. Bando, T. Kugo, K. Yamawaki, Phys. Rep. [164]{} (1988) 217. G. Ecker, J. Gasser, A. Pich, E. de Rafael, Nucl. Phys. B [231]{} (1989) 311. G. Ecker, J. Gasser, H. Leutwyler, A. Pich, E. de Rafael, Phys. Lett. B [223]{} (1989) 425. M. F. M. Lutz, E. E. Kolomeitsev, Nucl. Phys. A [730]{} (2004) 392. L. Roca, E. Oset and J. Singh, Phys. Rev. D [72]{} (2005) 014002. Y. Zhou, X. L. Ren, H. X. Chen, L. S. Geng, Phys. Rev. D [90]{} (2014) 014020. J. J. Xie, Phys. Rev. C [92]{} (2015) 065203. G. Calderon, J. H. Munoz and C. E. Vera, Phys. Rev. D [87]{} (2013) 114011. C. Patrignani [et al.]{} \[Particle Data Group\], Chin. Phys. C [40]{} (2016) 100001. B. A. Li, Phys. Rev. D [55]{} (1997) 1436. T. Eguchi, Phys. Rev. D [14]{} (1976) 2755. M. K. Volkov,D. Ebert, Sov. J. Nucl. Phys. [36]{} (1982) 736 \[Yad. Fiz. [36]{} (1982) 1265\]. D. Ebert, M. K. Volkov, Z. Phys. C [16]{} (1983) 205. M. K. Volkov, Ann. Physics [157]{} (1984) 282. M. K. Volkov, Sov. J. Part. Nucl. [17]{} (1986) 186 \[Fiz. Elem. Chast. Atom. Yadra [17]{} (1986) 433\]. D. Ebert, H. Reinhardt, Nucl. Phys. B [271]{} (1986) 188. U. Vogl, W. Weise, Prog. Part. Nucl. Phys. [27]{} (1991) 195. S. P. Klevansky, Rev. Mod. Phys. [64]{} (1992) 649. M. K. Volkov, Phys. Part. Nucl. [24]{} (1993) 35 \[Fiz. Elem. Chast. Atom. Yadra [24]{} (1993) 81\]. J. Bijnens, C. Bruno, E. de Rafael, Nucl. Phys. B [390]{} (1993) 501. T. Hatsuda, T. Kunihiro, Phys. Rep. [247]{} (1994) 221. D. Ebert, H. Reinhardt, M. K. Volkov, Prog. Part. Nucl. Phys. [33]{} (1994) 1. V. Bernard, A. H. Blin, B. Hiller, Y. P. Ivanov, A. A. Osipov, Ulf-G. Meißner, Ann. Physics. [249]{} (1996) 499. A. V. Vishneva, M. K. Volkov, D. G. Kostunin, Eur. Phys. J. A [50]{} (2014) 137. M. K. Volkov, A. A. Osipov, JETP Letters [105]{} (2017) 215 \[Pis’ma ZETF [105]{} (2017), 205\]. A. A. Osipov, M. K. Volkov, Ann. Phys. [382]{} (2017) 50. J. Morais, B. Hiller, A. A. Osipov, Phys. Lett. B [773]{} (2017) 277. S. Wallner, arXiv:1711.09782 \[hep-ex\]. M. K. Volkov, C. Weiss, Phys. Rev. D [56]{} (1997) 221. M. K. Volkov, Phys. Atom. Nucl. [60]{} (1997) 1920 \[Yad. Fiz. [60]{} (1997) 2094\]. M. K. Volkov, D. Ebert, M. Nagy, Int. J. Mod. Phys. A [13]{} (1998) 5443. M. K. Volkov, V. L. Yudichev, Phys. Part. Nucl. [31]{} (2000) 282 \[Fiz. Elem. Chast. Atom. Yadra [31]{} (2000) 576\]. M. K. Volkov and A. E. Radzhabov, Phys. Usp. [49]{} (2006) 551 \[Uspekhi Fizicheskikh Nauk [176]{} (2006) 569\]. M. K. Volkov, A. B. Arbuzov, Phys. Part. Nucl. [47]{} (2016) 489 \[Fiz. Elem. Chast. Atom. Yadra [47]{} (2016) 489\]. M. K. Volkov, A. B. Arbuzov, Phys. Usp. [60]{} (2017) 643 \[Uspekhi Fizicheskikh Nauk [187]{} (2017) 689\]. M. A. Markov, Sov. J. Phys. [3]{} (1940) 452. H. Yukawa, Phys. Rev. [77]{} (1950) 219. J. Lukierski, M. Oziewicz, Phys. Lett. B [69]{} 339 (1977). T. Eguchi, Phys. Rev. D [17]{} (1978) 611. A. A. Osipov, A. A. Pivovarov and M. K. Volkov, Phys. Rev. D [96]{}, (2017) 054012. C. Adolph [et al.]{} (COMPASS Collaboration), Phys. Rev. Lett. [115]{} (2015) 082001. T. Gutsche, M. A. Ivanov, J. G. Körner, V. E. Lyubovitskij, K. Xu, Phys. Rev. D [96]{}, (2017) 114004.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Within the neurosciences, to observe variability across time in the dynamics of an underlying brain process is neither new nor unexpected. Wavelets are essential in analyzing brain signals because, even within a single trial, brain signals exhibit nonstationary behaviour. However, neurological signals generated within an experiment may also potentially exhibit evolution across trials (replicates). As neurologists consider localised spectra of brain signals to be most informative, here we develop a novel wavelet-based tool capable to formally represent process nonstationarities across both time and replicate dimensions. Specifically, we propose the Replicate Locally Stationary Wavelet (RLSW) process, that captures the potential nonstationary behaviour within and across trials. Estimation using wavelets gives a natural desired time- and replicate-localisation of the process dynamics. We develop the associated spectral estimation framework and establish its asymptotic properties. By means of thorough simulation studies, we demonstrate the theoretical estimator properties hold in practice. A real data investigation into the evolutionary dynamics of the hippocampus and nucleus accumbens during an associative learning experiment, demonstrate the applicability of our proposed methodology, as well as the new insights it provides.' author: - 'Jonathan Embleton[^1]' - 'Marina I. Knight' - Hernando Ombao title: Multiscale modelling of replicated nonstationary time series --- [1]{} [Keywords:]{} replicate time series; cross-trial dependence; neuroscience; wavelet-based spectra Introduction {#sec:intro} ============ In an experimental setting consisting of repeated trials, inference is typically carried out on the average dynamics of the underlying process over all trials. However, a recent study on neurological signals [@fo:medbp] suggests that this approach is naive due to its failure to account for the possibility of a change in the process dynamics over the course of the experiment. Their data example focusses on the hippocampus (Hc) and the nucleus accumbens (NAc), both known to play important roles in cognitive processing as they are individually associated with memory recall and the processing of reward, respectively. Recordings of electrical activity (at approximately 1000Hz) using local field potentials (LFPs) were obtained from the Hc and NAc of an awake behaving macaque during an associative learning experiment. For each trial, the macaque was presented with one of four pictures and was then tasked with associating this picture with one of the four doors appearing on the screen. Upon making a correct association, the macaque was rewarded with a small quantity of juice. Plots of the LFPs obtained from trials in which a correct association was made are shown in Figures \[fig:rb7\_Ct\_Hc\] and \[fig:rb7\_Ct\_NAc\]. Variability in neuronal activity within both brain regions has also been observed over the trials of a learning experiment in other recent studies, including [@segcin:dynamics; @Gorrombao:monkeydepend; @abela:2015]. Such traits present the challenge of modelling time series that display potential nonstationary behaviour not only across time, but also across trials. In the specific context of brain signals, the usefulness of time-scale spectral decompositions that are typical of wavelet constructions has already been established in the literature [@sand:2010; @park:mvlsw]. Our aim is therefore to develop a [wavelet]{}-based model that, in the spectral domain, captures in a scale-dependent manner not only the evolutionary dynamics of the underlying brain process across time (within each trial) but also across the trials of the experiment. An important assumption typically undertaken is that of independent/ uncorrelated trials. However, some studies document evidence of correlation across trials in an experiment [@arieli:1996; @huk:2018]. Thus, one important feature of our model is that it accounts for the dependence between trials by means of a coherence quantity that acts as a measure for cross-trial dependence. Time series data encountered in practice are often of a nonstationary nature and much recent research has been concerned with developing statistical models that capture this behaviour within the series. [@priest:1965] was first to study in detail nonstationary processes with time varying spectral characteristics. Further models using Fourier waveforms and adaptations of, have been developed by [@dahl:1997] who introduced the methodology of locally stationary processes which allowed for asymptotics to be established, and [@omb:2002] whom utilise the SLEX (smoothed localised complex exponentials) library of waveforms which give localisation in both time and frequency. Similar representations have been developed using wavelets. Approaches that use wavelet thresholding for smoothing the spectra of locally stationary time series have been considered by [@vSsch:1996] and [@neumvs:1997]. The locally stationary wavelet framework was developed by [@nvsk:2000]. Their approach used a set of discrete non-decimated wavelets to replace the Fourier exponentials used in [@dahl:1997], and this in turn offered localisation in both time and scale. Extensions to the multivariate setting appear in the work of [@omb:2005], [@sand:2010] and [@park:mvlsw], where a structure for coherence is also embedded in the model. What is not well accounted for in the literature, is the statistical modelling of second order nonstationarity across a collection of (constant mean) time series arising from the same experiment. In earlier work for stationary time series, [@digwas:1997] accounted for the stochastic variation that arises across replicated time series via a subject-specific realisation of a random process. They proposed a generalised linear mixed-effects model to estimate population characteristics where the subject-specific random term in the model captures the cross-subject variation. Extension to the nonstationary case has been addressed by [@QGL:2009] whom, through the locally stationary processes framework, developed a time-frequency functional model, where the time-varying log-spectra determine the evolution of the stochastic variation. More recently, [@fo:medbp] proposed a model that captures the evolution in the spectral characteristics of dynamic brain processes across a collection of trials in an associative learning experiment. Their methodology, developed for multivariate analysis, captures process evolution within a trial and across trials through a Fourier spectrum. The authors refer to trials as [replicates]{}, a term we will also borrow in our nomenclature. Our approach acknowledges the work of [@fo:medbp], who developed their time–replicate model using Fourier waveforms, but the limitation of their work is the assumption of uncorrelated replicates. To the best of our knowledge, our proposed methodology is the first to model nonstationary stochastic variation across replicate time series in the [wavelet domain]{}. Major advantages of our model are that (i) it offers the superior time-localisation typical of wavelet constructions, and (ii) it takes into account the correlation of brain signals across trials. We propose to model the replicate process within a locally stationary wavelet process paradigm that builds upon the framework introduced by [@nvsk:2000] for a single process (here, replicate or trial). Two new models are developed, first under the constraint of uncorrelated replicates, and then relaxing this assumption and allowing for replicate correlation. This amounts to developing novel evolutionary wavelet quantities and associated estimation theory that encompass variation [both]{} across time and replicate. To obtain well-behaved, consistent spectral estimates we propose to perform local smoothing of the raw wavelet peridograms across replicates, as opposed to employing the smoothing over time typically undertaken in the locally stationary process context. Replicate-coherence estimation theory is also treated and is shown to provide useful information about the process replicate evolution. The article proceeds as follows. Section \[sec:model\] introduces our proposed model as well as its associated estimation theory, both under the assumption of replicate uncorrelation. Section \[sec:sims\] details simulation studies that illustrate the behaviour of the proposed methodology. Section \[sec:modelext\] describes the new model that allows for correlation between replicate processes. Corresponding estimation theory is developed in Section \[sec:modelext\]. Section \[sec:simsext\] illustrates through simulation the advantage of our proposed work, both for within- and across- replicate behaviour characterisation. Section \[sec:rda\] details an application of the proposed methodologies to a real data study within neuroscience, and Section \[sec:concl\] concludes the paper. Theoretical model {#sec:model} ================= Brief introduction of locally stationary wavelet processes {#brief-introduction-of-locally-stationary-wavelet-processes .unnumbered} ---------------------------------------------------------- Before we describe the framework for the replicate locally stationary wavelet model, we recall some of the defining features of the locally stationary wavelet (LSW) framework of [@nvsk:2000]. The LSW model provides a time-scale representation of nonstationary time series with time-varying second order structure, where the building blocks are the discrete non-decimated wavelets (see [@vid:smbw] or [@nas:wms] for an extensive introduction to wavelets). For $T=2^{J(T)}$, a sequence of stochastic processes $\{X_{t;T}\}_{t=0,...,T-1}$ is a LSW process if it admits the representation $$\label{eq:lsw} X_{t;T} = \sum_{j=1}^{\infty}\sum_{k\in\mathbb{Z}} \omega_{j,k;T}\psi_{j,k}(t)\xi_{j,k},$$ where for scale $j$ and location $k$, $\omega_{j,k;T}$ is the amplitude corresponding to the discrete non-decimated wavelet $\psi_{j,k}(t)$ and $\{\xi_{j,k}\}$ are a set of orthonormal random variables. Modelling under the concept of local stationarity means that the variation of the amplitudes $\{\omega_{j,k;T}\}_k$, happens slowly over time and this is controlled by a smoothly varying continuous Lipschitz function $W_{j}(k/T)$, that can be thought of as a scale ($j$) and time ($k$) dependent transfer function [@frynas:HFisz]. [@nvsk:2000] further propose the evolutionary wavelet spectrum (EWS) as a means to quantify the contribution to the overall process variance at a scale $j$ and rescaled time $z=k/T$ and define this as $S_{j}(z) = \|W_{j}(z)\|^2$. The raw wavelet periodogram is used for estimation of the EWS and is defined as $I_{j,k;T}=\|d_{j,k;T}\|^2$ where $d_{j,k;T}$ denotes the process wavelet coefficient at scale $j$ and location $k$ associated to a discrete non-decimated family of wavelets as defined in [@nvsk:2000]. The original LSW model does not capture the dynamics of time series data recorded for several trials over the course of an entire experiment. This setting presents additional challenges, notably the fact that these signals behave in a way that is nonstationary at multiple scales, (i) within the signal in each trial, and (ii) across trials over the course of the entire experiment. Replicate Locally Stationary Wavelet (RLSW) process --------------------------------------------------- \[def:rlsw\] We define a sequence of stochastic processes $\{X_{t;T}^{r;R}\}$, with time $t=0, \ldots, T-1$ where $T=2^{J(T)}$ and replicate $r=0, \ldots, R-1$ where $R=2^{J(R)}$ to be a replicate locally stationary wavelet process if it admits the following representation $$\label{eq:rlsw} X_{t;T}^{r;R} = \sum_{j=1}^{\infty}\sum_{k\in\mathbb{Z}} \omega_{j,k;T}^{r;R}\psi_{j,k}(t)\xi_{j,k}^{r},$$ where within each replicate $r$, for a scale $j$ and time $k$, $\omega_{j,k;T}^{r;R}$ are the amplitudes for the non-decimated wavelets $\psi_{j,k}(t)$ and $\{\xi_{j,k}^{r}\}$ are a set of orthonormal random variables with properties as detailed below. Letting $\nu = r/R$ denote rescaled replicate and $z=k/T$ denote rescaled within-trial time, the quantities in possess the following properties: 1. For all $j$, $k$ and $r$, ${\mathbf{E}}[\xi_{j,k}^{r}] = 0 \ (\Rightarrow {\mathbf{E}}[X_{t;T}^{r;R}] = 0)$. 2. ${\mathbf{E}}[\xi_{j,k}^{r} \xi_{j',k'}^{r'}] = \delta_{j,j'}\delta_{k,k'}\delta_{r,r'} \ (= \text{cov}(\xi_{j,k}^{r}, \xi_{j',k'}^{r'}))$. This amounts to assuming uncorrelated replicates. 3. For each scale $j \geq 1$, there exists a Lipschitz continuous transfer function in both rescaled time ($z$) and rescaled replicate ($\nu$), denoted by ${\widetilde{W}}_{j}(z,\nu)$ with the following properties 1. [ $$\begin{aligned} \label{eq:unifW} \sum_{j=1}^{\infty} \left\|{\widetilde{W}}_{j}(z,\nu)\right\|^{2} < \infty \ \text{uniformly in } z \in (0,1), \nu \in (0,1).\end{aligned}$$ ]{} 2. Let $L_j^\nu$ denote the bounded Lipschitz constant corresponding to the time dimension at a particular (rescaled) replicate ($\nu$) and scale $j$. Similarly, denote by $N_j^z$ the bounded Lipschitz constant corresponding to the replicate dimension at a particular (rescaled) time ($z$) and scale $j$. Denote $L_j=\sup_{\substack{\nu\in (0,1)}} L_j^\nu$ and $N_j=\sup_{\substack{z\in (0,1)}} N_j^z$, and assume they are uniformly bounded in $j$. Further assume that $$\label{eq:condL} \sum_{j=1}^{\infty} 2^{j}L_{j} < \infty \text{ and } \sum_{j=1}^{\infty} 2^{j} N_j< \infty.$$ 3. \[enum:seq1\] There exist sequences of bounded replicate-specific constants $\{C_{j}^{r}\}_r$ and location-specific constants $\{D_{j}^{k}\}_k$, such that for each $T$ and $R$ respectively, the amplitudes are forced to vary slowly, in the sense that $$\begin{aligned} \sup_{\substack{k=0: T-1}} \left\| \omega_{j,k;T}^{r;R} - {\widetilde{W}}_{j}\left(\frac{k}{T},\frac{r}{R}\right) \right\| &\leq \frac{C_{j}^{r}}{T}, \quad \forall j, r,\label{eq:supWr}\\ \sup_{\substack{r=0: R-1}}\left\| \omega_{j,k;T}^{r;R} - {\widetilde{W}}_{j}\left(\frac{k}{T},\frac{r}{R}\right)\right\| &\leq \frac{D_{j}^{k}}{R}, \quad \forall j, k. \label{eq:supWt}\end{aligned}$$ Denote $C_j=\sup_r C_j^r$ and $D_j=\sup_k D_j^k$ and assume the sequences $\{C_{j}\}$, $\{D_{j}\}$ fulfill $\sum_{j=1}^{\infty} 2^{j}C_{j} < \infty$ and $\sum_{j=1}^{\infty} 2^{j}D_{j} < \infty$. [Remark (rescaled time and replicate evolution).]{} Within each scale $j$, the transfer function ${\widetilde{W}}_{j}(z,\nu)$ controls the evolution of the amplitudes, forcing them to vary slowly over [both]{} rescaled time ($z$) and replicate ($\nu$) dimensions. The evolution of the amplitudes over time within each replicate happens in a smooth manner. The evolution across replicates is such that while the spectral properties of different replicates may also be different, however across neighbouring replicates there is a larger degree of commonality. Nevertheless, further apart replicates may display different traits. Such a meta-process evolution appears later in Figure \[fig:wavplots\_SIM3\] (Section \[sec:sims\]). Replicate evolutionary wavelet spectrum --------------------------------------- As is common in spectral domain analysis (both Fourier and wavelet-based), we do not work directly with the time- and replicate-specific multiscale transfer functions $\{{\widetilde{W}}_{j}(\cdotp,\cdotp)\}_{j}$, but instead we define a transformed version that quantifies the contribution to the process variance attributed across scales ($j$) to each time [and]{} replicate. As noted, the current LSW model and its related quantities are not capable of capturing the multiscale evolution of brain signals. Here, we develop a novel evolutionary wavelet spectrum corresponding to a RLSW process, a quantity that for simplicity we refer to as the [replicate evolutionary wavelet spectrum]{}. The replicate evolutionary wavelet spectrum (REWS) at scale $j$, rescaled replicate $\nu$, rescaled within-trial time $z$ is given by $$\label{eq:rews} S_{j}(z,\nu) = \left\|{\widetilde{W}}_{j}(z,\nu)\right\|^2 = \lim_{\substack{T\to\infty \\ R\to\infty}}\left(\left\|\omega_{j,\lfloor zT\rfloor;T}^{\lfloor \nu R\rfloor;R}\right\|^2\right),$$ where $\lfloor zT\rfloor$ and $\lfloor \nu R \rfloor$ denote the largest integer less than or equal to $zT$ and $\nu R $, respectively. Note that from equations and we directly obtain that for each $T$ and $R$ we have $$\begin{aligned} \label{eq:approx} \sup_{\substack{r=0: R-1}}\sup_{\substack{k=0: T-1}}\left\|\omega_{j,k;T}^{r;R} - {\widetilde{W}}_{j}\left(\frac{k}{T},\frac{r}{R}\right)\right\| &= {\mathcal{O}}(C_{j}{T}^{-1}) + {\mathcal{O}}(D_{j}{R}^{-1}),\end{aligned}$$ hence the right-hand equality in equation . [Remark (RLSW versus LSW processes).]{} An innovation of the proposed RLSW model is to impose within each scale not only a smooth spectral behaviour across each (replicate) time series, but also to constrain the ‘meta’-spectral evolution across replicates to happen in a smooth manner, as detailed by the conditions in Definition \[def:rlsw\]. Note that a replicate locally stationary wavelet (RLSW) process is thus [not]{} to be understood only as a collection of locally stationary wavelet (LSW) processes that happen to be observed across several replicates, as this would limit its capacity to represent multiscale behaviour across trials. [Remark (bounded variation jumps).]{} Our theoretical development could of course be extended to encompass bounded variation jumps, but this is outside the scope of this work. Nevertheless, we show through simulation that such behaviour is well handled by the proposed methodology. We define the replicate local autocovariance (RLACV) for the replicated process $\{X_{t;T}^{r;R}\}$ for some rescaled time $z$ and rescaled replicate $\nu$ to be given by $$c(z,\nu; \uptau) = \sum_{j=1}^\infty S_{j}(z, \nu)\Psi_{j}(\uptau),$$ where $\uptau$ is an integer time-lag, $z\in (0,1)$, $\nu \in (0,1)$ and $\Psi_{j}(\uptau)=\sum_{k\in\Z}\psi_{j,k}(0)\psi_{j,k}(\uptau)$ denotes the scale $j$ autocorrelation wavelet. Note that $\|c(z,\nu; \uptau)\|<\infty$ follows directly from the uniform bounds in $\uptau$ and $\nu$ for both the limiting amplitudes and autocorrelation wavelets (see equation ). The local autocovariance defined above can be shown to be an approximation of the process autocovariance corresponding to a particular (rescaled) replicate, as follows. \[prop:autocov\] For a RLSW process $\{X_{t;T}^{r;R}\}$ with properties as in Definition \[def:rlsw\],\ $\left\| {\mathrm{cov}}(X_{\lfloor zT\rfloor;T}^{\lfloor \nu R\rfloor;T}, X_{\lfloor zT \rfloor +\uptau;T}^{\lfloor \nu R\rfloor;T} )- c(z,\nu; \uptau) \right\|={\mathcal{O}}(T^{-1})+{\mathcal{O}}(R^{-1})$, uniformly in $\uptau$ at (rescaled) time $z$ and replicate $\nu$. The proof appears in Section \[supp:proofs:autocov\] of the Supplementary Material and uses the approximation properties in the definition of the RLSW process. Spectral estimation {#sec:specestRLSW} ------------------- We start our proposed estimation procedure for the REWS (and thus also for the RLACV), by first computing the raw wavelet periodogram and exploring its asymptotic properties as an estimator for the true, unknown REWS. We note that the theoretical results in this section are derived under the assumption of Gaussianity. \[def:wp1\] We define the raw wavelet periodogram of a RLSW process $\{X_{t;T}^{r;R}\}$ as $$\label{eq:wp1} I_{j,k;T}^{r;R} = \left\|d_{j,k;T}^{ r;R}\right\|^{2},$$ where for scale $j$, replicate $r$ and location $k$, $d_{j,k;T}^{r;R}=\sum_{t=0}^{T-1} X_{t;T}^{r;R}\psi_{j,k}(t)$ are the process empirical wavelet coefficients constructed using a family of discrete non-decimated wavelets, $\{\psi_{j,k}(t)\}_{j,k}$. We note here that unlike the Fourier periodogram, the wavelet-based raw periodogram is typically not an unbiased estimator of the wavelet spectrum, and this will also turn out to be the case here. For reasons that will become obvious next, we also define a transformed spectral quantity $\beta_{j}(z,\nu)=\sum_{l=1}^\infty A_{j,l} S_{l} (z,\nu)$, where $A_{j,l}= \langle \Psi_{j},\Psi_{l} \rangle = \sum _{\uptau\in\Z}\Psi_{j}(\uptau)\Psi_{l}(\uptau)$ is the inner product matrix of the autocorrelation wavelets. The invertibility of the matrix $A$ and boundedness of its inverse norm [@nvsk:2000] ensure that finding a well-behaved estimator of the REWS is equivalent to finding a well-behaved estimator for the spectral quantity $\beta$. Hence we next focus on estimating $\beta$. \[prop:EI\] For a RLSW process $\{X_{t;T}^{r;R}\}$ with properties as in Definition \[def:rlsw\], its associated raw wavelet periodogram in equation has the following asymptotic properties: Expectation $$\label{eq:EI} {\mathbf{E}}\left[I_{j,\lfloor zT\rfloor;T}^{\lfloor \nu R \rfloor;R}\right] = \beta_{j}(z,\nu) + \textcolor{black}{{\mathcal{O}}(2^{j}T^{-1}) + {\mathcal{O}}(R^{-1})},$$ Variance $$\label{eq:varI} \text{var}\left(I_{j,\lfloor zT\rfloor;T}^{\lfloor \nu R \rfloor;R}\right) = 2\beta_{j}^2(z,\nu) + \textcolor{black}{{\mathcal{O}}(2^{2j}T^{-1}) + {\mathcal{O}}(2^{j}R^{-1})}.$$ Appendix \[app:proofs:rawi\] contains the proof. From Proposition \[prop:EI\], we see that the raw periodogram is asymptotically unbiased for $\beta$, but inconsistent due its asymptotically non-vanishing variance. Thus we next propose to smooth the raw periodogram in order to obtain consistency (and then we can correct for bias to obtain an asymptotically unbiased estimator for $S$). We define a replicate-smoothed estimator for the quantity $\beta_{j}(\frac{k}{T},\frac{r}{R})$ to be $$\label{eq:defItilde} {\tilde{I}}_{j,k;T}^{r;R} = (2M + 1)^{-1} \sum_{s=-M}^{M}I_{j,k;T}^{r+s;R},$$ where $(2M+1)$ is the length of the smoothing window and $M$ is an integer such that as $T, R \to \infty$, we have that $M \to \infty$ and $M/R \to 0$. [Remark (replicate smoothing).]{} Unlike for the usual locally stationary processes where the periodogram is smoothed over time in order to achieve consistency, here we propose a smoothing procedure that operates over [replicates]{} by locally averaging the spectral estimates across a window of neighbouring replicates. This approach is indeed theoretically justified by the assumption of spectral smoothness across the replicate-dimension. In practice, these assumptions will have to be verified in order to determine some empirically guided choice of $M$, as seen in the simulation study Section \[sec:sims\]. \[prop:ESI\] For a RLSW process $\{X_{t;T}^{r;R}\}$ with properties as in Definition \[def:rlsw\], the replicate-smoothed wavelet periodogram in equation has the following asymptotic properties: Expectation $$\label{eq:ESI} {\mathbf{E}}\left[{\tilde{I}}_{j,\lfloor zT\rfloor;T}^{\lfloor \nu R \rfloor;R}\right] = \beta_{j}(z,\nu) +\textcolor{black}{{\mathcal{O}}(MR^{-1}) + {\mathcal{O}}(2^{j}T^{-1})}, $$ Variance $$\label{eq:varSI} \text{var}\left({\tilde{I}}_{j,\lfloor zT\rfloor;T}^{\lfloor \nu R \rfloor;R}\right) = \textcolor{black}{{\mathcal{O}}(2^{2j}M^{-1})+{\mathcal{O}}(2^jR^{-1})+{\mathcal{O}}(MR^{-2})}.$$ Appendix \[app:proofs:itilde\] contains the proof which manipulates the amplitude properties across replicates as opposed to those across time. Note that as $T$, $R$ and $M \to \infty$ and using the condition $M/R \to 0$, the bias of the smoothed periodogram becomes asymptotically negligible, while its variance tends to zero for any fixed fine enough scale $j$ (with $2^j=o\left(\mbox{min}\left\{T,R,(2M+1)^{1/2}\right\}\right)$). The usual bias–variance trade-off here is manifest through the increase of $M$ resulting in a decrease of the variance at the price of an increase in the bias. As the replicate-smoothed periodogram proposed above is an asymptotically unbiased and consistent estimator for the true $\beta$, the relationship between the true spectral quantities $\beta$ and $S$ suggests a natural way of constructing a well-behaved spectral estimator for the unknown spectrum $S$. We thus propose to estimate the unknown REWS by means of $$\label{eq:Sestr} \hat{S}_{j}(z,\nu) = \sum_{l=1}^{J}A_{j,l}^{-1}{\tilde{I}}_{l,\lfloor zT \rfloor;T}^{\lfloor \nu R \rfloor;R},$$ where $A_{j,l}^{-1}$ is the $(j,l)$ entry of the inverse of the inner product matrix $A$ of the autocorrelation wavelets and $J=\lfloor\alpha J(T)\rfloor$ with $\alpha\in(0,1)$. \[prop:CS\] For a RLSW process $\{X_{t;T}^{r;R}\}$ with properties as in Definition \[def:rlsw\], the spectral estimator proposed in equation is asymptotically unbiased and consistent for $S_{j}(z,\nu)$ for each fixed scale $j$, rescaled time and replicate $z$ and $\nu$, respectively, provided that $M/R \to 0$ as $T$, $R$ and $M \to \infty$. Appendix \[app:proofs:corrper\] contains the proof which hinges on the properties of the replicate-smoothed periodogram shown in Proposition \[prop:ESI\] above. [Remark (replicate and time smoothing).]{} The results in Proposition \[prop:ESI\] highlight the small sample dependence of the bias and variance of the smoothed periodogram on the number of replicates $R$, on the time series length $T$ and on the smoothing window $(2M+1)$, as well as well as on the ratio of (replicate) smoothing window to the total number of replicates. While still having a bias–variance tradeoff, the variance can be further improved by additionally smoothing across the time-dimension. Specifically, using a time-smoothing window of length $(2M_T+1)$ such that $M_T \to \infty$ and $M_T/T \to 0$ (the reader may also refer to [@park:mvlsw]) and chosen as usual under LSW modelling (see e.g. [@nas:station]), and preserving the previous notation of $(2M+1)$ for the replicate-smoothing window, we define the replicate- and time-smoothed periodogram $$\label{eq:defItildetilde} \tilde{{\tilde{I}}}_{j,k;T}^{r;R} = (2M + 1)^{-1} (2M_T + 1)^{-1} \sum_{s=-M}^{M} \sum_{t=-M_T}^{M_T} I_{j,k+t;T}^{r+s;R},$$ to act as an estimator for the transformed spectral quantity $\beta_{j}(\frac{k}{T},\frac{r}{R})$. Below we show that this estimator has desirable asymptotic properties, leading to faster convergence than its counterpart involving only replicate-smoothing. \[prop:ESIts\] For a RLSW process as in Definition \[def:rlsw\] and satisfying the additional assumption of autocovariance summability, $\sup_{z,\nu}\sum_{n\in\Z}\left\|c(z,\nu;n\right\|)={\mathcal{O}}(1)$, the smoothed time- and replicate-specific wavelet periodogram defined in equation has the following asymptotic properties: Expectation $$\label{eq:ESIts} {\mathbf{E}}\left[\tilde{{\tilde{I}}}_{j,\lfloor zT\rfloor;T}^{\lfloor \nu R \rfloor;R}\right] = \beta_{j}(z,\nu)+ {\mathcal{O}}(M_T T^{-1}) +\textcolor{black}{{\mathcal{O}}(MR^{-1}) + {\mathcal{O}}(2^{j}T^{-1})}, $$ Variance $$\label{eq:varSIts} \text{var}\left(\tilde{{\tilde{I}}}_{j,k;T}^{r;R}\right) ={\mathcal{O}}(2^{2j}(M_TM)^{-1})+{\mathcal{O}}(2^{j}(MR)^{-1}).$$ Appendix \[app:proofs:itildetilde\] contains the proof which makes use of the smoothing in both directions. The replicate- and time- smoothed periodogram can then be used to further build a well-behaved estimator of the unknown REWS by means of $$\label{eq:Sestrt} \hat{\hat{S}}_{j}(z,\nu) = \sum_{l=1}^{J}A_{j,l}^{-1}\tilde{{\tilde{I}}}_{l,\lfloor zT \rfloor;T}^{\lfloor \nu R \rfloor;R}.$$ It is straightforward to show that this is also asymptotically unbiased and consistent for $S_j(z,\nu)$, in the same manner as in the proof of Proposition \[prop:CS\]. Simulation Study {#sec:sims} ================ Here we aim to assess the behaviour of our proposed RLSW methodology as well as compare it to a classical approach involving the LSW methodology [@nvsk:2000]. Specifically, we evaluate (i) the classical approach where one would independently estimate the spectrum for each replicate using a localised time smoother and then average over all replicates (‘LSW’), (ii) our proposed methodology involving localised smoothing over replicates (‘RLSW$_1$’), and (iii) our proposed methodology involving localised smoothing over time and replicates (‘RLSW$_2$’). In order to match the current practice for LSW estimation, e.g. [@nvsk:2000; @park:mvlsw], we have set $J=J(T)$ (corresponding to $\alpha=1$), although in a bivariate spectral estimation context [@sand:2010] set a similar measure to $\alpha=0.7$ and remark on its improved results when compared to $\alpha=1$. We carry out simulations over $N=100$ runs and explore performance across a range of time series lengths $T$ from $128$ to $1024$, number of replicates $R$ from $64$ to $512$ and smoothing windows $(2M + 1)$ from $9$ to $25$. We report the mean squared errors (MSE) and squared bias results. Overall, based on our findings, we recommend the use of the method involving both time and replicate smoothing (RLSW$_2$) with a window length choice of $(2M+1)=0.15 R$ as a rule of thumb. [Illustrative example]{}. We choose to present here the behaviour of our proposed methodology on a process with a challenging spectral structure, as shown in Figure \[fig:wavplots\_SIM3\] and mathematically defined in Section \[supp:extradetail\] of the Supplementary Material. Further simulation findings are detailed in Section \[supp:furthersims\] of the Supplementary Material, including MSE and squared bias tables. For $R = 256$ replicates each of length $T = 256$, the process places spectral content at level $j = 5$, manifest through a decreasing amplitude of the cosine across the last 192 replicates, and at level $j = 6$, where the periodicity of the cosine increases across the first 128 replicates. A concatenated realisation of this process is shown in Figure \[fig:reptsBY7\_SIM3proc\]. Note however that this is an abuse of representation, since each replicate is a time series of its own, and the sole purpose of this visualisation is to highlight the evolution of the meta-process. Furthermore this process departs somewhat from the requirement that the amplitudes evolve slowly over both rescaled time $(z)$ and replicate $(\nu)$ dimensions, however we show that despite this the methodology still performs well. Spectral estimates have been computed using discrete non-decimated wavelets built by means of Daubechies Least Asymmetric family with 6 vanishing moments (see [@daub:ten] for an understanding of Daubechies compactly supported wavelets). For the RLSW method, local averaging involved windows of 9 replicates corresponding to $M = 4$ and we note that numerical MSE results in Appendix \[app:tables:sims\] highlight that we chose to visually present here some of our least performant results. The LSW and RLSW($_1$) spectral estimates appear in Figure \[fig:coef\_SIM3proc\_R256\_T256\], along with the truth. From the figures we get a visual clarification that the RLSW model is doing a good job at capturing the evolving characteristics of the spectra across replicates and the leakage across the neighbouring levels $j = 5$ and 6 is minor. They also highlight that when neglecting the possibility of evolutionary behaviour over replicates, when it is in fact present as seen for levels $5$ and $6$ in the top row plots of the true spectrum, the LSW model struggles to reflect this and either under or over-estimates, as seen in the middle row plots. The bottom row plots show that the RLSW$_1$ estimates do indeed pick up the evolution over replicates. Figures \[fig:lev2\_SIM3proc\_R256\_T256\] and \[fig:lev3\_SIM3proc\_R256\_T256\] further support the evolutionary behaviour of the spectral quantities over time and replicates in levels 5 and 6, respectively. Histograms of the MSEs over the 100 simulations are shown in Section \[supp:extradetail\] (Supplementary Material) and highlight not only how the increase in $M$ improves performance but also how the increase in $R$ and $T$ reduces the MSEs, thus demonstrating the expected asymptotic behaviour of our smoothed estimator. To numerically strengthen our visual inference, we examine the MSEs and squared bias results in Table \[tab:msesim3\] (Appendix \[app:tables:sims\]). The best results in terms of lowest MSEs are obtained by RLSW$_2$. Specifically, RLSW$_2$ does incur a somewhat higher bias than RLSW$_1$, but both our methods have a substantially lower bias than the LSW. The higher bias of the LSW estimates is unsurprisingly resulting from averaging over all the replicates and thus failing to account for the evolutionary behaviour through replicates. The benefit of taking a local smoothing approach over both time and replicates is that it always results in spectral estimates with lower bias and MSE when compared to LSW, although it is worth pointing out that taking a local smoothing approach over replicates only, while yielding lower bias, might increase the MSE for inappropriately small windows. Asymptotically, the MSEs associated to our methods decay much faster than for the LSW and as we increase the local averaging window length ($2M + 1$), the performance of our the RLSW methodology improves. A replicate window length choice of $(2M+1)$ equal to about 15% of $R$ appears to work well across all our investigations. RLSW model embedding replicate coherence {#sec:modelext} ======================================== So far, a serious limitation is the assumption that the replicate time series are uncorrelated. We now develop the theory to allow for cross-trial dependence by means of a coherence structure along the replicate dimension. This is a major innovation of this work, as to the best of our knowledge this is the first paper that accounts for correlation across trials. \[def:rlswc\] Let the sequence of stochastic processes $\{X_{t;T}^{r;R}\}_{t,r}$ be a replicate locally stationary wavelet process as defined in Definition \[def:rlsw\] with the following amendments to the properties: 2. (replacing property 2) Additional to $\{\xi_{j,k}^{r}\}_{j,k}$ being orthonormal within replicate $r$, we have ${\mathbf{E}}[\xi_{j,k}^{r} \xi_{j',k'}^{r'}] = \delta_{j,j'}\delta_{k,k'}\rho_{j,k;T}^{r,r';R}$, where $\{\rho_{j,k;T}^{r,r';R}\}_k$ determine the innovation dependence structure between replicates $r$ and $r'$, at each scale $j$. Note that the within-replicate orthonormality induces $\left\|\rho_{j,k;T}^{r,r';R}\right\| \leq1$ for all $j$, $k$ and $r$, $r'$, with equality when $r=r'$. 3. (replacing property 3(a)) $$\begin{aligned} \label{eq:unifWc} \sum_{j=1}^{\infty} \left\|{\widetilde{W}}_{j}(z,\nu){\widetilde{W}}_{j}(z,\nu^\prime)\right\| < \infty \ \text{uniformly in } z \in (0,1), \nu , \nu^\prime \in (0,1).\end{aligned}$$ 4. (additional property) For each scale $j \geq 1$, there exists a Lipschitz continuous function in rescaled time ($z$) and rescaled replicate arguments $\nu$ and $\nu'$, denoted by $\rho_{j}(z, \nu, \nu')$, which constrains the covariance structure and fulfills the assumptions below, as follows 1. Let $Q_j^{\nu,\nu'}$ denote the bounded Lipschitz constant corresponding to the time dimension at particular (rescaled) replicates $\nu$ and $\nu'$, at scale $j$. Similarly, denote by $P_j^z$ the bounded Lipschitz constant corresponding to the replicate dimension at a particular (rescaled) time ($z$), at scale $j$. Denote $Q_j=\sup_{\substack{\nu,\nu'\in (0,1)}} Q_j^{\nu,\nu'}$, $P_j=\sup_{\substack{z\in (0,1)}} P_j^z$ and assume they are uniformly bounded in $j$. Further assume that $$\label{eq:condLrho} \sum_{j=1}^{\infty} 2^{j}Q_{j} < \infty \text{ and } \sum_{j=1}^{\infty} 2^{j} P_j< \infty.$$ 2. \[enum:seq2\] There exist sequences of bounded replicate-specific constants $\{\tilde{C}_{j}^{r,r'}\}_{r,r'}$ and location-specific constants $\{\tilde{D}_{j}^{k}\}_k$, such that for each $T$ and $R$ respectively, the covariances are forced to vary slowly, in the sense that $$\begin{aligned} \sup_{\substack{k=0: T-1}} \left\| \rho_{j,k;T}^{r,r';R} - \rho_{j}\left(\frac{k}{T},\frac{r}{R},\frac{r'}{R}\right) \right\| &\leq \frac{\tilde{C}_{j}^{r,r'}}{T}, \quad \forall j, r, r'\label{eq:supcohr}\\ \sup_{\substack{r,r'=0: R-1}}\left\| \rho_{j,k;T}^{r,r';R} - \rho_{j}\left(\frac{k}{T},\frac{r}{R},\frac{r'}{R}\right)\right\| &\leq \frac{\tilde{D}_{j}^{k}}{R}, \quad \forall j, k. \label{eq:supcoht}\end{aligned}$$ Denote $\tilde{C}_j=\sup_{r,r'} \tilde{C}_j^{r,r'}$ and $\tilde{D}_j=\sup_k \tilde{D}_j^k$ and assume the sequences $\{\tilde{C}_{j}\}$, $\{\tilde{C}_{j}\}$ fulfill $\sum_{j=1}^{\infty} 2^{j}\tilde{C}_{j} < \infty$ and $\sum_{j=1}^{\infty} 2^{j}\tilde{D}_{j} < \infty$. [Remark (rescaled replicate dependence).]{} Note that from equations and we directly obtain for each $T$ and $R$ $$\begin{aligned} \label{eq:approxc} \sup_{\substack{r,r'=0: R-1}}\sup_{\substack{k=0: T-1}}\left\|\rho_{j,k;T}^{r,r';R} - \rho_{j}\left(\frac{k}{T},\frac{r}{R},\frac{r'}{R}\right) \right\| &= {\mathcal{O}}(\tilde{C}_{j}{T}^{-1}) + {\mathcal{O}}(\tilde{D}_{j}{R}^{-1}).\end{aligned}$$ Hence for some rescaled time $z$ and rescaled replicates $\nu$ and $\nu'$, we have in the limit $$\rho_{j}(z, \nu,\nu') = \lim_{\substack{T\to\infty \\ R\to\infty}}\left(\rho_{j,\lfloor zT\rfloor;T}^{\lfloor \nu R\rfloor, \lfloor \nu' R\rfloor;R}\right).$$ For a scale $j$ and time location $k$, the quantity $\rho_{j}(\frac{k}{T}, \frac{r}{R},\frac{r'}{R})$ thus gives a measure of the dependence between replicates $r$ and $r'$. Evolutionary wavelet replicate–(cross-)spectrum and coherence ------------------------------------------------------------- We next define a measure for the scale, time and cross-replicate contribution to the overall process variance. The [evolutionary wavelet replicate–cross-spectrum]{} defined at scale $j$, rescaled time $z$ within rescaled replicates $\nu$ and $\nu'$ is given by $$S_{j}(z,\nu,\nu') = {\widetilde{W}}_{j}(z,\nu){\widetilde{W}}_{j}(z,\nu')\rho_{j}(z, \nu, \nu').$$ In order to simplify the terminology, we also refer to this quantity as the replicate evolutionary wavelet cross-spectrum (REWCS). Note in the above definition that the replicate–cross-spectrum corresponding to any rescaled replicate $\nu$, is nothing else but the spectrum corresponding to that replicate, i.e. $S_{j}(z,\nu,\nu) =S_{j}(z,\nu)$. Alternatively, $\|\tilde{W}_{j}(z,\nu)\|=\left(S_{j}(z,\nu,\nu)\right)^{1/2}$. For a replicate locally stationary wavelet process $\{X_{t;T}^{r;R}\}$, we define the replicate local cross-covariance (RLCCV) at rescaled time $z\in (0,1)$ within rescaled replicates $\nu$ and $\nu'$, both in $(0,1)$, to be given by $$c(z,\nu, \nu'; \uptau) = \sum_{j=1}^{\infty}S_{j}(z, \nu, \nu')\Psi_{j}(\uptau),$$ where $\uptau \in \mathbb{Z}$ is an integer time-lag and we recall that $\Psi_{j}(\uptau)=\sum_{k\in\Z}\psi_{j,k}(0)\psi_{j,k}(\uptau)$ is the scale $j$ autocorrelation wavelet. Akin to established time series literature, we can next rephrase the dependence measure as $$\label{eq:lsrc} \rho_{j}(z, \nu, \nu') = \frac{S_{j}(z,\nu,\nu')}{\left\{S_{j}(z,\nu)S_{j}(z,\nu')\right\}^{1/2}},$$ and we shall refer to it as the locally stationary replicate–coherence, with values ranging from $-1$, indicating an absolute negative correlation, to $1$ indicating an absolute positive correlation. Note that $\|c(z,\nu, \nu'; \uptau)\|<\infty$ follows directly from the coherence range between $-1$ and $1$, and from the uniform bounds in lag ($\uptau$) and rescaled-replicate time ($\nu$, $\nu'$) for both the limiting amplitudes and the autocorrelation wavelets (see equation ). Unsurprisingly, the local autocovariance can be shown to be an approximation of the process autocovariance, as follows. \[prop:autocovc\] For a RLSW process $\{X_{t;T}^{r;R}\}$ with properties as in Definition \[def:rlswc\],\ $\left\| {\mathrm{cov}}(X_{\lfloor zT\rfloor;T}^{\lfloor \nu R\rfloor;T}, X_{\lfloor zT \rfloor +\uptau ;T}^{\lfloor \nu' R\rfloor;T} )- c(z,\nu,\nu'; \uptau) \right\|={\mathcal{O}}(T^{-1})+{\mathcal{O}}(R^{-1})$, uniformly in $\uptau$ at (rescaled) time $z$ and replicates $\nu$, $\nu'$. The proof appears in Section \[supp:proofs:autocovc\] of the Supplementary Material. Estimation theory {#sec:modelext_est} ----------------- Recall, in the absence of cross-trial dependence, the raw wavelet periodogram is given by $I_{j,k;T}^{r;R} = \|d_{j,k;T}^{ r;R}\|^{2}$. We next introduce a cross-replicate version. For a scale $j$ and location $k$, we define the raw wavelet cross-periodogram between replicates $r$ and $r'$ of a RLSW process to be $$\label{eq:rawIc} I_{j,k;T}^{(r,r');R} = d_{j,k;T}^{ r;R}d_{j,k;T}^{ r';R}.$$ The theoretical results below are derived under the Gaussianity assumption. \[prop:EIc\] For a RLSW process $\{X_{t;T}^{r;R}\}$ as in Definition \[def:rlswc\], the wavelet cross-periodogram has the following asymptotic properties: Expectation $${\mathbf{E}}\left[I_{j,\lfloor kT \rfloor;T}^{(\lfloor \nu R \rfloor,\lfloor \nu' R \rfloor);R}\right] = \sum_{l=1}^{\infty}A_{j,l} S_{l}\left(z,\nu,\nu'\right) + {\mathcal{O}}(2^jT^{-1}) + {\mathcal{O}}(R^{-1}),$$ Variance $$\begin{aligned} \nonumber \text{var}\left(I_{j,\lfloor kT \rfloor;T}^{\lfloor \nu R \rfloor,\lfloor \nu' R \rfloor;R}\right) &= \left(\sum_{l=1}^{\infty}A_{j,l} S_{l}\left(z,\nu\right)\right) \left(\sum_{l=1}^{\infty}A_{j,l} S_{l}\left(z,\nu'\right)\right) \nonumber \\ & \qquad + \left(\sum_{l=1}^{\infty}A_{j,l} S_{l}\left(z,\nu,\nu'\right)\right)^{2} + {\mathcal{O}}(2^{2j}T^{-1}) + {\mathcal{O}}(2^{j}R^{-1}).\nonumber\end{aligned}$$ The proofs follow similar steps to their non-coherence counterpart in Appendix \[app:proofs:rawi\] and are thus omitted here. [Remark (replicate smoothing).]{} In line with the spectral ‘similarity’ amongst neighbouring replicates, we proceed by smoothing the cross-periodograms across replicates for consistency and then correcting them for bias. We define a smoothed estimator across the replicate dimension to be $$\label{eq:reppersmooth} {\tilde{I}}_{j,k;T}^{(r,r');R} = (2M + 1)^{-1} \sum_{s=-M}^{M}I_{j,k;T}^{(r+s,r'+s);R},$$ where $(2M+1)$ is the length of the smoothing window and $M$ is an integer such that as $T, R \to \infty$, we have that $M \to \infty$ and $M/R \to 0$. \[prop:ESIc\] Under the properties of Definition \[def:rlswc\] and the additional assumption\ $\sup_{z,\nu\in (0,1)}\sum_{\eta\in\Z} \left\| c(z,\nu,\nu+\frac{\eta}{R};\uptau) \right\|={\mathcal{O}}(1)$ for any time lag $\uptau$, the [replicate-smoothed]{} wavelet cross-periodogram defined above has the following asymptotic properties: Expectation $$\label{eq:ESIc} {\mathbf{E}}\left[{\tilde{I}}_{j,\lfloor kT \rfloor;T}^{(\lfloor \nu R \rfloor,\lfloor \nu' R \rfloor);R}\right] = \sum_{l=1}^{\infty}A_{j,l} S_{l}\left(z,\nu,\nu'\right)+ {\mathcal{O}}(MR^{-1}) + {\mathcal{O}}(2^{j}T^{-1}),$$ Variance $$\nonumber \text{var}\left({\tilde{I}}_{j,\lfloor kT \rfloor;T}^{(\lfloor \nu R \rfloor,\lfloor \nu' R \rfloor);R}\right)={\mathcal{O}}(2^{2j}M^{-1})+ {\mathcal{O}}(2^j R^{-1})+ {\mathcal{O}}(M^2R^{-2}).$$ Appendix \[app:proofs:itildec\] contains the proof. The bias of the replicate-smoothed wavelet cross-periodogram becomes asymptotically negligible and $\text{var}\left({\tilde{I}}_{j,k;T}^{(r,r');R}\right) \to 0$ for any scale $j$ with $2^j=o\left( \mbox{min}\{ T, R, (2M+1)^{1/2}\}\right)$ as $T$, $R$, $M \to \infty$ and $M/R \to 0$. Then correcting for the bias will yield a desirable spectral estimator, as follows. \[prop:CSC\] The following is an asymptotically unbiased and consistent replicate-smoothed estimator for the REWCS $$\label{eq:corrspecc} \hat{S}_{j}(z,\nu, \nu') = \sum_{l=1}^{J}A_{j,l}^{-1}{\tilde{I}}_{l,\lfloor zT \rfloor;T}^{(\lfloor \nu R \rfloor,\lfloor \nu' R \rfloor);R},$$ where $A_{j,l}^{-1}$ is the $(j,l)$ entry of the inverse of the inner product matrix $A$ of the autocorrelation wavelets and $J=\lfloor \alpha J(T) \rfloor$ with $\alpha \in (0,1)$. The proof follows the same steps as for Proposition \[prop:CS\]. This paves the way towards proposing the replicate–coherence estimator $$\label{eq:rhoc} \hat{\rho}_{j}(z,\nu, \nu') = \frac{\hat{S}_{j}(z,\nu,\nu')}{\left\{\hat{S}_{j}(z,\nu)\hat{S}_{j}(z,\nu')\right\}^{1/2}},$$ where the involved spectral quantities are consistently estimated as proposed in Proposition \[prop:CSC\] (see also equation ) and the use of the same smoothing windows guarantees that the values of the resulting coherence estimator are indeed quantities between $-1$ and $1$ (a proof appears in Section \[supp:proofs:rhoclim\] of the Supplementary Material). The following proposition shows that the step of examining $\hat{\rho}_{j}(z,\nu, \nu')$ is theoretically justified. \[prop:rhoc\] Under the assumptions of Proposition \[prop:ESIc\], the coherence estimator in is asymptotically consistent for the true coherence $\rho_{j}(z,\nu, \nu')$. See Appendix \[app:proofs:rhoc\]. Coherence illustration via simulation {#sec:simsext} ===================================== We shall now investigate through simulation the performance of our proposed methodology for coherence estimation. We use the mean squared error (MSE) and squared bias of the estimates $\hat \rho$, averaged over all time-scale points and replicates, as detailed in Appendix \[app:tables:simsext\]. We display the behaviour of our estimators on a simulated example below, and provide a further simulation study in Section \[supp:furthersimscoh\] of the Supplementary Material. [Illustrative example]{}. We simulate a replicate locally stationary wavelet process with $R=256$ replicates that feature dependence, measured at $T=512=2^{9}$ time points. The locally stationary wavelet autospectra are defined by a sine wave whose periodicity and magnitude evolve slowly over the replicates in such a way that the spectral characteristics of neighbouring replicates do not look too dissimilar whilst there is a noticeable difference between replicates further apart (for their mathematical expression, see Simulation 1, Section \[supp:furthersims\] of the Supplementary Material). Here we have $J(T)=9$ (in short, $J$) and the spectral characteristics are placed in level $j=J(T)-4=5$. In addition to the autospectral characteristics, we also define a challenging cross-replicate spectral structure by means of defining their (true) coherence at each level $j$ and location $k$. Specifically, we set the coherence to be zero over the last $256$ locations, yielding an $R \times R$ coherence identity matrix. For level $j=5$ and time $k = 1,\ldots, 256$ we define the non-zero replicate coherence matrices as follows: the first $128$ replicates have a strong positive coherence (0.99) with one another, however this coherence becomes negative (-0.71) with the last $128$ replicates. A (weaker) positive coherence (0.5) also exists between the last $128$ replicates. The expressions of the non-zero coherence matrices can be found in Section \[supp:furthersims\] of the Supplementary Material and the illustrative true coherence structures for replicates 50 (top row) and 200 (bottom row) can be visualised in Figure \[fig:coh\_sim2\] (left panels). Coherence estimates obtained using the methodology proposed in Section \[sec:modelext\_est\] are represented in Figure \[fig:coh\_sim2\] (right panels) for replicates 50 (top row) and 200 (bottom row). Non-decimated discrete wavelets built using Daubechies least asymmetric family with 10 vanishing moments and local averaging over a window of 9 replicates were employed. In terms of correctly estimating the coherence structure switch over times and replicates, as well as identifying the positive or negative character of the coherence, the proposed estimation procedure does a good job. We do however note that the estimated coherence intensity does exhibit some bias, which we may attribute to the smoothing performed in order to address the practical computation considerations (see the remark in Appendix \[app:tables:simsext\]). Nevertheless, we could argue that the model does give a good indication for the degree of the positiveness of the coherence, approximately $0.99$ and $0.5$ for replicates 50 and 200 respectively (right panels of Figure \[fig:coh\_sim2\]). Tables \[tab:rhomsesim2\] (Appendix \[app:tables:simsext\]) and \[tab:rhomsesim1\] (Section \[supp:furthersimscoh\] in the Supplementary Material) illustrate MSE results for two smoothing approaches: the first involves smoothing only over a window of replicates; the second involves local averaging through time and through replicates. For both simulations, the results paint the same picture. As we increase the replicate smoothing window (such that $M/R \to 0$) the performance of our models improves in terms of MSEs, and the double smoothing over time and replicates further reduces the MSEs. The price to pay for double smoothing as usual is a slightly higher bias (than when using averaging over replicates only). In order to ensure that our spectral estimates are positive, our correction procedure uses the correction matrix $A^{-1}$ truncated at zero. Inevitably this introduces bias, evident through the increasing MSEs as $R$ and $T$ increase. In their work on bivariate coherence estimation, [@sand:2010] reported better results when additionally employing smoothing over scales. We conjecture that this is also applicable for our work, but leave the further numerical treatment for future research. Analysis of Macaque Local Field Potentials {#sec:rda} ========================================== We perform our analysis on the dataset of local field potentials (LFPs) from the hippocampus (Hc) and nucleus accumbens (NAc) of a macaque over the course of an associative learning experiment. Due to their roles in the consolidation of memory information and the processing of rewarding stimuli, the Hc and NAc have been studied in relation to learning tasks for monkeys, rats and humans [@wirth:2003; @abela:2015; @segcin:dynamics]. Our RLSW-based spectral analysis offers not only confirmations to the results of previous studies, but also provides additional insights to the understanding of the dynamics of the LFPs through capturing the evolutionary characteristics of brain processes [within]{} and [across]{} the trials of the experiment, in a [scale-dependent]{} manner, through the use of the wavelet transform. One notable benefit of our model in contrast to previous Fourier-based methodology [@fo:medbp], is its superior time-localisation, as we will see next. Each trial consists of $T = 2048$ time points, corresponding to approximately 2 seconds of data. The design splits each trial into four time blocks of 512 milliseconds each and it is in the final time block that the macaque was tasked with associating one of four doors (appearing on a screen) with the picture visual presented in the second time block. The macaque had to learn the associations through repeated trials and for each correct association made the macaque was given a juice reward. Further details are given in Section \[supp:exper\] of the Supplementary Material. The data has been grouped into sets of ‘correct’ and ‘incorrect’ responses, in order to investigate how the contributions of the Hc and NAc to the learning process differ between groups [@Gorrombao:monkeydepend]. The groups containing the correct and incorrect responses consist of 241 and 264 trials, respectively. We opt to carry out the analysis on $R = 256$ replicates (trials). To obtain a dyadic number of replicates necessary for estimation (here, 256) for the correct response group, we mirror the last 15 trials. This is for computational purposes only, and we naturally discard the corresponding estimates from our discussions and plots. To ensure comparability across trials, each trial is standardised to have mean zero and unit variance. Plots for the correct responses appeared in Figures \[fig:rb7\_Ct\_Hc\] and \[fig:rb7\_Ct\_NAc\] (Section \[sec:intro\]) and for the incorrect responses in Figures \[fig:rb7\_It\_Hc\] and \[fig:rb7\_It\_NAc\] of the Supplementary Material Section \[supp:exper\], which also details the implemented methodology. Results ------- Hippocampus. The proposed spectral estimates for the correct and incorrect sets of trials appear in Figures \[fig:coef\_CtHc\_M10D10\] (below) and \[fig:coef\_ItHc\_M10D10\] (Section \[supp:exper\], Supplementary Material), respectively. The bottom row plots show the RLSW($_2$)-spectral estimates averaged across 30 replicates to illustrate the process behaviour in the beginning, middle and end of the experiment. These demonstrate (i) the sequenced activation of within-trial time blocks and (ii) the evolutionary behaviour of the wavelet spectrum along the course of the experiment. For both correct and incorrect datasets, the ‘activity’ is primarily captured within the coarser levels of the wavelet periodograms, approximately corresponding to frequencies 2-8Hz. Of these, the theta band frequencies 4-8Hz are typical of slow activity, known for their association to hippocampal activity in mammals and to promote memory [@buzs06]. [@fo:medbp] report the low frequency range 1-12Hz to account for most variability in the Hc data. Our analysis offers a finer characterisation that does not fully support activation of low delta waves (under 2Hz), known to be typical of deep sleep, and shows weak alpha band (8-12Hz) and low beta band (12-16Hz) alertness at certain time blocks within each trial. Due to its construction, the RLSW model has been able to capture the process evolutionary behaviour across trials, which naturally cannot be achieved by LSW alone. The LSW-based estimates do capture some of the time-dependencies but remained constant across replicates. The RLSW estimates have the capacity to highlight the individual time blocks as they activate through the course of the experiment, an insight invisible to LSW and weakly represented in the Fourier approach of [@fo:medbp]. For the correct Hc trials, the RLSW models capture the bulk of ‘activity’ in levels 3 and 4. In early replicates this is fairly even through time, while for middle replicates the bursts of activity shift centrally within time, thus coinciding with the second block of the macaque being shown the visual stimulus and possibly with the expectation of the picture to continue being shown. For the final trials, the activity is clearly localised around time-point 500 (corresponding to the visual exposure) and towards the final quarter of time (corresponding to the time when the macaque made the correct association). When compared to a Fourier approach, our wavelet-based analysis thus brings to the fore novel information that links the experimental time blocks to Hc activation. Specifically, as the correct trials progress, the activity in the Hc is evident at the visual cue time and also at the selection task time, thus suggesting learning of the picture associations. Although we cannot compare the correct and incorrect trials like-for-like, we are still able to see evidence of evolutionary behaviour across the incorrect trials. As the experiment progresses, there is evidence of less spectral activity in the incorrect trials, with a brief Hc activation in the visual exposure block for the middling trials, and a burst of Hc activity localised in the last time block, when the task is carried out, for the end trials. The spectrum suggests that whereas the Hc displays prolonged activity in the second time block for the correct trials (corresponding to the picture being presented), this feature is not as marked in the incorrect trials and thus the macaque is not making the association between the picture presented and the selection task. Scientific literature has shown [@segcin:dynamics] that during a learning experiment, activity in the Hc decreases as associations/rules are learned but would spike upon the application of an association/rule. The capacity of our model to extract time-localised information thus highlights novel traits that suggest that the macaque in this experiment has not yet fully learned the associations, but evidence of learning is indeed present. Figure \[fig:Hc\_avlevplots\] further illustrates spectral evolution through time is captured by both the LSW and RLSW models. As the experiment progresses, the RLSW model identifies that the activity in the Hc increases across the correct trials with higher activity along the middle and end replicates, and (within-replicate) time dependent peaks gradually spanning the course of the experiment. The final trials display more Hc activity towards the task time than starting trials, again indicating a learning process. In contrast, the incorrect trials display a much less structured behaviour, with time-dependent activity distributed more evenly across the replicates. Nucleus accumbens. The resulting (R)LSW spectral estimates appear in Figures \[fig:coef\_CtNAc\_M10D6\] (below) and \[fig:coef\_ItNAc\_M10D6\] (Section \[supp:exper\], Supplementary Material) for the correct and incorrect sets of trials, respectively. These plots are to be understood in the same manner as those for the hippocampus. [@fo:medbp] find that the bulk of variability in the NAc is accounted for by (high) beta band frequencies (20-30Hz), while we place this in the wider range of beta band waves 16-30Hz, associated to focussed activity. Our analysis also offers evidence for low gamma frequency waves (31-60Hz), typical of working memory activation [@iacc:2016]. Additional to Fourier analysis, the RLSW model also shows that nonstationarity across time is clearly present, as well as some spectral evolution across the replicates. Although not as obvious as for the Hc, for the beginning and middle replicates of the correct group, NAc activity is manifest towards the trial start and end, while for the final replicates activity is captured in the final quarter of time. A similar pattern of behaviour is displayed by the incorrect group of trials. Also note that the NAc activity decreases in intensity from the beginning to end replicates for both groups. The NAc is part of the ventral striatum and plays a role in the processing of rewarding stimuli. The activity seen in the final 512 milliseconds can be attributed to the macaque expecting and receiving the juice reward in the correct trials, or expectation of reward in the incorrect trials. The impact of reward expectation [@schultz:rewardexp; @Hollerman:1998; @mulder:2005] could also explain the activity we see at the trial start for the beginning replicates and its observed periodicity across the experiment. Upon receiving no reward in an incorrect trial, the NAc activity decreases and with it the reward expectation falls for the next trial. The opposite holds for a correct trial when the reward is given. Our analysis reflects the results of other studies on learning experiments [@Hollerman:1998; @fo:medbp] that highlight that the activity in the ventral striatum decreases as the stimuli are learned. Plots for the average RLSW spectral estimates of the NAc in the finer levels 6 and 7 appear in Figure \[fig:NAc\_avlevplots\]. Our previous comments are also reflected in these plots. Evolution in the spectra across time is captured by both models, with the NAc activity displaying periodic patterns, while the RLSW-based estimation highlights a decrease in NAc activity along the trials of the experiment. Remarks. Our analysis demonstrates how the simplifying assumption of trials that are identical realisations of the same process, leading one to draw conclusions solely based on averaging across all replicates, could cause an important understanding in the process evolution through the experiment to be missed. Our proposed RLSW methodology has captured the spectral [time-]{} and [replicate-]{}evolutionary behaviour, thus yielding new scale-based results and refining the findings of [@fo:medbp] in the Fourier domain. Nevertheless, these results are still underpinned by the assumption of uncorrelated trials. We next explore whether this assumption is tenable. Allowing for correlation between trials --------------------------------------- We relax the assumption of uncorrelated trials and allow for (potential) dependencies across trials. Hence we now estimate the locally stationary replicate-coherence, a quantity defined in . The aim of our analysis is to investigate whether dependencies across trials exist. For the Hc data, we were unable to capture any substantial coherence structure over the trials (correct and incorrect groups). However, our analysis did find evidence of a moderate dependence across neighbouring trials in the NAc at the beta band frequencies (16-30Hz), known to be responsible for brain activity related to reward feedback mechanisms. The estimated NAc replicate-coherence (absolute value) is shown in Figure \[fig:NAc\_cohplots\] at level 6 for trials 20, 100 and 200 in the incorrect group and for correct trial 200, depicting typical behaviour. Some burst areas are present, indicating a moderate neighbouring replicate coherence, with most meaningful values either side of 0.4 and a few above 0.5. For the beginning and middling incorrect trials, this is apparent in the time periods leading up to and inclusive of the trial task phase, upon which the macaque would receive a juice reward if the task was done correctly. Remarks. Our analysis provides novel evidence in the temporal and scale (frequency)-dimensions that mild to moderate dependence is exhibited in both the correct and incorrect trials. This is primarily evident in the final correct trials, potentially as the manifest result of learning, and at the onset of the incorrect trials as the likely result to the expectation of reward. Within the neuroscience literature (e.g. [@Gorrombao:monkeydepend]), dependence between brain regions is also of interest, with coherence measures setup between channels of interest (here, Hc and NAc). Such measures have not formed the scope of our work here, but in relating our replicate-coherence results to the reported evolutionary coherence between the Hc and NAc, the dependence we observed at the beginning of NAc correct trial 200 (approximately rescaled replicate 0.8) is reminiscent of the dependence between Hc and NAc captured at rescaled replicate-time 0.8 by [@fo:medbp]. Concluding remarks {#sec:concl} ================== Our work proposed a novel wavelet-based methodology that successfully captures nonstationary process characteristics across a replicated time series, often encountered in experimental studies. Its desirable properties were evidenced by simulation studies and through a real data application from the neurosciences. However, the methodology itself is not restricted to use within this field, and the authors envisage its utility in other experimental areas where wavelet spectral analysis has proved to be ideally suited, e.g. circadian biology [@hargreaves19:wavelet]. This work demonstrates the dangers of approaching replicated time series as identical process realisations and the misleading results this can yield when studying the process dynamics along the course of the experiment. Our work proposes statistical models and associated estimation theory for processes that embed not only uncorrelated replicates, but also allow for dependence between replicates. A next natural step would be to extend the RLSW methodology to a multivariate setting ([@sand:2010; @park:mvlsw]) and to additionally define and investigate the variate replicate coherence. [99]{} Supporting evidence for simulation studies {#app:tables} ========================================== MSE and squared bias tables for Section \[sec:sims\] {#app:tables:sims} ---------------------------------------------------- As means to quantify the performance of the models, we employ the [mean squared error]{} (MSE) and [squared bias]{}, calculated as the average over all time-scale points and replicates as follows $$\begin{aligned} MSE(\hat{S}) &= {(R \, J \, T)}^{-1} \sum_{r, j, k}\left[ \frac{1}{N}\sum_{n=1}^{N} \left( \hat{S}_{j}^{(n)}\left(\frac{k}{T},\frac{r}{R}\right) - S_{j}\left(\frac{k}{T},\frac{r}{R}\right) \right)^{2}\right],\\ Bias^2(\hat{S}) &= {(R \, J \, T)}^{-1} \sum_{r, j, k}\left[ \frac{1}{N}\sum_{n=1}^{N} \hat{S}_{j}^{(n)}\left(\frac{k}{T},\frac{r}{R}\right) - S_{j}\left(\frac{k}{T},\frac{r}{R}\right) \right]^{2}.\end{aligned}$$ [Remarks on estimates at the boundaries]{}. We do not assess edges that involve local averaging over the first and last $(M-1)$ replicates. This has also been accounted for when calculating the MSE and squared bias. As a result, the reported measures whose values correspond to modelling via LSW will appear to change (in a very minor way), when in fact they should be the same for all choices of $M$ at fixed $R$ and $T$. [@ccccccccc@]{} & &\ & & &\ R & T & M & mse & bias[2]{} & mse & bias[2]{} & mse & bias[2]{}\ 256 & 128 & 4 & 17.55 & 17.27 & 22.31 & 2.53 & 11.52 & 3.88\ & & 12 & 16.74 & 16.46 & 9.45 & 2.40 & 6.54 & 3.82\ & 256 & 4 & 14.21 & 13.95 & 19.62 & 1.04 & 8.46 & 1.29\ & & 12 & 13.48 & 13.22 & 7.58 & 0.92 & 3.83 & 1.23\ & 512 & 4 & 12.25 & 12.01 & 17.59 & 0.53 & 7.15 & 0.52\ & & 12 & 11.59 & 11.35 & 6.51 & 0.42 & 2.85 & 0.48\ [@ccccccccc@]{} & &\ & & &\ R & T & M & mse & bias[2]{} & mse & bias[2]{} & mse & bias[2]{}\ 512 & 256 & 4 & 14.29 & 14.14 & 19.68 & 1.05 & 8.49 & 1.28\ & & 12 & 13.89 & 13.76 & 7.62 & 0.92 & 3.81 & 1.22\ & 512 & 4 & 12.30 & 12.18 & 17.61 & 0.53 & 7.15 & 0.52\ & & 12 & 11.96 & 11.84 & 6.54 & 0.42 & 2.85 & 0.48\ & 1024 & 4 & 10.92 & 10.81 & 15.83 & 0.32 & 6.29 & 0.25\ & & 12 & 10.61 & 10.50 & 5.78 & 0.22 & 2.38 & 0.21\ MSE and squared bias tables for Section \[sec:simsext\] {#app:tables:simsext} ------------------------------------------------------- As numerical tools that quantify the performance of an estimate $\hat{\rho}_{j}(z,\nu)$ across all time points $z=k/T$ with $k=0, \ldots, T-1$, replicates $\nu=r/R$ with $r=0, \ldots, R-1$ and scales $j=1, \ldots, J$, we will use the mean squared error and squared bias, defined in this context as $$\nonumber MSE(\hat{\rho})= ( R^{} \, J \, T )^{-1} \sum_{r, r^{}, j, k}\left[ \frac{1}{N}\sum_{n=1}^{N} \left( \hat{\rho}_{j}^{(n)}\left(\frac{k}{T},\frac{r}{R},\frac{r^{}}{R}\right) - \rho_{j}\left(\frac{k}{T},\frac{r}{R},\frac{r^{}}{R}\right) \right)^{2}\right],$$ $$\nonumber Bias^2(\hat{\rho}) = (R^{} \, J \, T)^{-1} \sum_{r, r^{}, j, k}\left( \frac{1}{N}\sum_{n=1}^{N} \hat{\rho}_{j}^{(n)}\left(\frac{k}{T},\frac{r}{R},\frac{r^{}}{R}\right) - \rho_{j}\left(\frac{k}{T},\frac{r}{R},\frac{r^{}}{R}\right) \right)^{2},$$ where due to the symmetry of the coherence matrix, we have used $R^{} = R(R+1)/2$ and $r^{} = r' \geq r$, and $N$ denotes the number of simulation runs. As in Section \[sec:sims\], we also adopt here $J=J(T)$. [Remarks on implementation]{}. From the theoretical model construction, the autospectra are positive quantities. However, our spectral estimates may take values that are negative or close to zero after correction, and this in turn can cause problems when normalising for coherence estimation. In order to bypass this issue, we choose to correct our raw wavelet periodogram estimates before smoothing. The theoretical properties of the coherence estimator show that using replicate-smoothing does yield an estimator with good properties, albeit its rate of convergence is heavily dependent on the smoothing window width $(2M+1)$. A local averaging window over time for smoothing each replicate before applying smoothing over replicates could also be employed, just as proposed for spectral estimation in Section \[sec:specestRLSW\]. A possible further step, to be done after smoothing through time, is to smooth over scales as proposed by [@sand:2010] for estimating the linear dependence between bivariate LSW time series. However, we do not pursue this approach here. [cccccccc]{} R & T & M & & mse[1]{} & bias[2]{}[1]{} & mse[2]{} & bias[2]{}[2]{}\ 128 & 256 & 7 & & 18.10 & 9.02 & 15.50 & 10.71\ & & 12 & & 14.53 & 8.22 & 13.05 & 9.64\ 128 & 512 & 7 & & 18.91 & 10.56 & 16.72 & 12.37\ & & 12 & & 15.10 & 9.40 & 13.95 & 10.92\ 256 & 512 & 7 & & 20.82 & 11.58 & 18.40 & 13.60\ & & 12 & & 17.78 & 10.93 & 16.43 & 12.81\ Proofs of results on the asymptotic behaviour of proposed estimators under the uncorrelated replicates assumption {#app:proofs} ================================================================================================================= In this section, we give details of the proofs in Section \[sec:model\] using the notation described therein. For ease we recall here that the auto- and cross-correlation wavelets are defined for $\uptau\in\Z$ as $\Psi_j(\uptau)=\sum_{k\in \Z}\psi_{j,k}(0)\psi_{j,k}(\uptau)$ and $\Psi_{j,l}(\uptau)=\sum_{k\in \Z}\psi_{j,k}(0)\psi_{l,k}(\uptau)$ respectively, where in general $\psi_{j,k}(\uptau)=\psi_{j,k-\uptau}$ are compactly supported discrete wavelets as defined in [@nvsk:2000]. Note that from their construction, $\Psi_j(-\uptau)=\Psi_j(\uptau)$ and $\Psi_{j,l}(-\uptau)=\Psi_{l,j}(\uptau)$, and both have compact support of order $2^j$, $2^j+2^l$ respectively [@nvsk:2000; @sand:2010]. In what follows, wherever the summation domain is not specified, it is to be understood as $\Z$ for time indices (e.g. $k, \, n, \, \uptau$) and as $\Z^\star_+$ (strictly positive integers) for scale indices (e.g. $j, \, j', \, l$). We also recall the autocorrelation wavelet inner product matrices, defined as A\_[j,l]{}&=&\_ \_j()\_l()=\_ \| \_[j,l]{}()\|\^2,\ A\_[j,l]{}\^&=&\_[n]{} \_[j]{}(n) \_[l]{}(n+)=\_[n]{}\_[j,l]{}(n)\_[j,l]{}(n+) , with properties $A_{j,l}\geq 0$, $A_{j,l}=A_{l,j}$, $\sum_j 2^{-j} A_{j,l}=1$ [@fryz03:forecasting]. In the proofs that follow, we make use of the results in the following lemmas, whose proofs appear in Section \[supp:proofs:sec3\] of the Supplementary Material. \[lem:ASord\] Under the assumptions of Definition \[def:rlsw\], we have for every rescaled time and replicate, $z, \, \nu$ respectively, and scale $j$, $\sum_{l=1}^{\infty}A_{j,l} S_{l}\left(z,\nu\right)={\mathcal{O}}(2^{j})$. \[lem:ASordtau\] Under the assumptions of Definition \[def:rlsw\], we have for every rescaled time and replicate, $z, \, \nu$ respectively, lag $\uptau$ and scale $j$, $\sum_{l=1}^{\infty}A_{l,j}^{\uptau} S_{l}\left(z,\nu\right)={\mathcal{O}}(2^{j})$. \[lem:ASordjj\] Under the assumptions of Definition \[def:rlsw\], we have for every rescaled time and replicate, $z, \, \nu$ respectively, lag $\uptau$ and scales $j, j'$, $$\sum_{l=1}^{\infty} \left\|\sum _{n\in \Z} \Psi_{j,j'}(n) \Psi_{l}(n+\uptau) S_{l}\left(z,\nu\right)\right\|={\mathcal{O}}(2^{(j+j')/2}).$$ Proof of Proposition \[prop:EI\] {#app:proofs:rawi} --------------------------------- ### Proof of Proposition \[prop:EI\] (Expectation) {#proof-of-proposition-propei-expectation .unnumbered} For a fuller understanding of the process behaviour over replicates, we start by defining the cross-replicate-periodogram as $I_{j,k;T}^{(r,r');R} = d_{j,k;T}^{ r;R} d_{j,k;T}^{ r';R}$. Taking the expectation, it follows that $$\begin{aligned} {\mathbf{E}}\left[I_{j,k;T}^{(r,r');R}\right] &= {\mathbf{E}}\left[d_{j,k;T}^{ r;R} d_{j,k;T}^{ r';R}\right] \nonumber \\ &= {\mathbf{E}}\left[\left\{ \sum_{t} X_{t;T}^{r;R}\psi_{j,k}(t) \right\} \left\{ \sum_{t'} X_{t';T}^{r;R}\psi_{j,k}(t') \right\}\right] \nonumber \\ &= {\mathbf{E}}\left[\left\{ \sum_{t} \sum_{l=1}^{\infty}\sum_{m\in\mathbb{Z}} \omega_{l,m;T}^{r;R}\psi_{l,m}(t)\xi_{l,m}^{r}\psi_{j,k}(t) \right\}\right. \nonumber \\ & \quad\qquad\qquad \times \left.\left\{ \sum_{t'} \sum_{l'=1}^{\infty}\sum_{m'\in\mathbb{Z}} \omega_{l',m';T}^{r;R}\psi_{l',m'}(t')\xi_{l',m'}^{r}\psi_{j,k}(t') \right\}\right] \nonumber \nonumber \\ &= \sum_{t}\sum_{t'} \sum_{l=1}^{\infty}\sum_{l'=1}^{\infty}\sum_{m\in\mathbb{Z}}\sum_{m'\in\mathbb{Z}}\omega_{l,m;T}^{r;R}\omega_{l',m';T}^{r';R}\psi_{l,m}(t)\psi_{l',m'}(t')\psi_{j,k}(t)\psi_{j,k}(t'){\mathbf{E}}\left[\xi_{l,m}^{r}\xi_{l',m'}^{r'}\right] \nonumber \\ &= \sum_{l=1}^{\infty}\sum_{m\in\mathbb{Z}} (\omega_{l,m;T}^{r;R})^{2} \sum_{t}\psi_{l,m}(t)\psi_{j,k}(t)\sum_{t'}\psi_{l,m}(t')\psi_{j,k}(t') \nonumber \\ &= {\mathbf{E}}\left[I_{j,k;T}^{r;R}\right],\nonumber\end{aligned}$$ since for the orthogonal increments, ${\mathbf{E}}\left[\xi_{l,m}^{r}\xi_{l',m'}^{r'}\right] = \delta_{l,l'}\delta_{m,m'}\delta_{r,r'}$ and so $l=l'$, $m=m'$ and $r=r'$. Using the properties in equation and the definition of the cross-correlation wavelet functions ($\Psi_{j,l}$), we obtain by replacing $n: =m-k$ &=& \_[l=1]{}\^\_[n]{} (\_[l,n+k;T]{}\^[r;R]{})\^[2]{} \_[t]{}\_[l,n+k]{}(t)\_[j,k]{}(t)\_[t’]{}\_[l,n+k]{}(t’)\_[j,k]{}(t’) ,\ &=& \_[l=1]{}\^\_[n]{} (S\_[l]{}(,) + (D\_[l]{}R\^[-1]{})+(C\_[l]{}T\^[-1]{})) { \_[t]{}\_[l,n+k]{}(t)\_[j,k]{}(t) }\^2,\ &=& \_[l=1]{}\^\_[n]{} (S\_[l]{}(,) + (L\_[l]{} \|n\| T\^[-1]{})+(D\_[l]{}R\^[-1]{})+(C\_[l]{}T\^[-1]{})) (\_[j,l]{}(n))\^2, where in the last equality we used the Lipschitz continuity of the spectrum in the (rescaled) time argument and a substitution $n:= -n$. Let us now bound the order terms, as follows. Since the number of terms in the wavelet cross-correlation is finite and bounded as a function of $n$ [@nvsk:2000] and the compact support of $\Psi_{j,l}^2$ is bounded by $K(2^{j}+2^{l})$ for some constant $K$ [@sand:2010], we have \_[l]{}\_[n]{} L\_[l]{} \|n\| (T\^[-1]{})\_[j,l]{}\^2(n) && (T\^[-1]{}) \_[l]{} (2\^[j]{}+2\^[l]{}) L\_l A\_[j,l]{},\ && (T\^[-1]{} 2\^j) \_[l]{} L\_l A\_[j,l]{} + (T\^[-1]{}) \_[l]{} 2\^[l]{}L\_l A\_[j,l]{} ,\ &=& (2\^j T\^[-1]{}), where we used $\sum_{l} L_l A_{j,l}=\sum_{l} 2^l L_l 2^{-l}A_{j,l}\leq \sum_{l^\prime} 2^{l^\prime} L_{l^\prime} \sum_{l} 2^{-l} A_{j,l} < \infty$ since $\sum_{l} 2^{l}L_l<\infty$ and $\sum_{l} 2^{-l} A_{j,l}=1$ [@fryz03:forecasting], and T\^[-1]{}\_[l]{} 2\^l L\_l A\_[j,l]{}&=&2\^j T\^[-1]{} \_[l]{} 2\^l L\_l 2\^[-j]{}A\_[j,l]{},\ && 2\^j T\^[-1]{} \_[l]{} ( 2\^[l]{} L\_[l]{} \_[j\^]{} 2\^[-j\^]{} A\_[j\^,l]{} ),\ &=& (2\^j T\^[-1]{}), again as $\sum_{l} 2^{l}L_l<\infty$ and $\sum_{j} 2^{-j} A_{j,l}=1$. Using the definition of the $A$ matrix, $A_{j,l}=\sum_{n} \Psi_{j,l}^2(n)$, in the next order term, we obtain \_[l]{}\_[n]{} D\_[l]{} (R\^[-1]{})\_[j,l]{}\^2(n) &=& (R\^[-1]{}) \_[l]{} 2\^l D\_l 2\^[-l]{} A\_[j,l]{},\ && (R\^[-1]{}) \_[l\^]{} 2\^[l\^]{}D\_[l\^]{} \_[l]{} 2\^[-l]{}A\_[j,l]{},\ &=& (R\^[-1]{}) where we used the condition $\sum_{l} 2^l D_l<\infty$ and $\sum_{l} 2^{-l} A_{j,l}=1$ [@fryz03:forecasting]. Using the same set of arguments and the condition $\sum_{l} 2^l C_l<\infty$, we also have $\sum_{l}\sum_{n} C_{l} {\mathcal{O}}(T^{-1})\Psi_{j,l}^2(n)={\mathcal{O}}(T^{-1})$. Hence, retaining the maximum order terms, we obtain &=& \_[l=1]{}\^\_[n]{} S\_[l]{}(,) (\_[j,l]{}(n))\^2 + (2\^j T\^[-1]{})+ (R\^[-1]{}),\ &=& \_[l=1]{}\^ A\_[j,l]{}S\_[l]{}(,) + (2\^j T\^[-1]{})+ (R\^[-1]{}). ### Proof of Proposition \[prop:EI\] (Variance) {#proof-of-proposition-propei-variance .unnumbered} For ease of notation let $X_{t;T}^{r;R} = X_{t}^{r}$ and $d_{j,k;T}^{ r;R}=d_{j,k}^{ r}$. $$\begin{aligned} \label{eq:EII1} {\mathbf{E}}\left[I_{j,k;T}^{r;R}I_{j',k';T}^{r';R}\right] &= {\mathbf{E}}\left[d_{j,k}^{ r} d_{j,k}^{ r}d_{j',k'}^{ r'} d_{j',k'}^{ r'}\right] \nonumber \\ &= {\mathbf{E}}\left[ \sum_{t} X_{t}^{r}\psi_{j,k}(t) \sum_{t'} X_{t'}^{r}\psi_{j,k}(t')\right. \nonumber \\ & \qquad\qquad\qquad\qquad \times \left.\sum_{h} X_{h}^{r'}\psi_{j',k'}(h)\sum_{h'} X_{h'}^{r'}\psi_{j',k'}(h')\right] \nonumber \\ &= \sum_{t}\sum_{t'}\sum_{h}\sum_{h'}\psi_{j,k}(t)\psi_{j,k}(t')\psi_{j',k'}(h)\psi_{j',k'}(h'){\mathbf{E}}\left[ X_{t;T}^{r;R}X_{t';T}^{r;R} X_{h;T}^{r';R}X_{h';T}^{r';R}\right].\end{aligned}$$ Using the result of [@isserlis:1918] for , we can write that $$\label{eq:iss} {\mathbf{E}}\left[ X_{t}^{r}X_{t'}^{r} X_{h}^{r'}X_{h'}^{r'}\right] = {\mathbf{E}}\Big[ X_{t}^{r}X_{t'}^{r}\Big]{\mathbf{E}}\left[ X_{h}^{r'}X_{h'}^{r'}\right] + {\mathbf{E}}\left[ X_{t}^{r}X_{h}^{r'}\right]{\mathbf{E}}\left[ X_{t'}^{r}X_{h'}^{r'}\right] + {\mathbf{E}}\left[ X_{t}^{r}X_{h'}^{r'}\right]{\mathbf{E}}\left[ X_{t}^{r}X_{h}^{r'}\right],$$ and substituting into re-write \[eq:EII\] = + + . Let us now consider $$\begin{aligned} \label{eq:EXX} {\mathbf{E}}\Big[ X_{t}^{r}X_{t'}^{r}\Big] &= {\mathbf{E}}\left[\left\{ \sum_{l=1}^{\infty}\sum_{m\in\mathbb{Z}} \omega_{l,m;T}^{r;R}\psi_{l,m}(t)\xi_{l,m}^{r} \right\}\left\{ \sum_{l'=1}^{\infty}\sum_{m'\in\mathbb{Z}} \omega_{l',m';T}^{r;R}\psi_{l',m'}(t')\xi_{l',m'}^{r}\right\}\right] \nonumber \\ &= \sum_{l=1}^{\infty}\sum_{m\in\mathbb{Z}} \omega_{l,m;T}^{r;R}\psi_{l,m}(t)\sum_{l'=1}^{\infty}\sum_{m'\in\mathbb{Z}} \omega_{l',m';T}^{r;R}\psi_{l',m'}(t'){\mathbf{E}}\left[\xi_{l,m}^{r}\xi_{l',m'}^{r}\right] \nonumber\\ &= \sum_{l=1}^{\infty}\sum_{m\in\mathbb{Z}} (\omega_{l,m;T}^{r;R})^{2}\psi_{l,m}(t)\psi_{l,m}(t'),\end{aligned}$$ since for the orthogonal increments, ${\mathbf{E}}\left[\xi_{l,m}^{r}\xi_{l',m'}^{r}\right] = \delta_{l,l'}\delta_{m,m'}$ and so is non-zero only when $l=l'$ and $m=m'$. Then for the first term in equation , we have $$\label{eq:EEXX} {\mathbf{E}}\Big[ X_{t}^{r}X_{t'}^{r}\Big]{\mathbf{E}}\left[ X_{h}^{r'}X_{h'}^{r'}\right] = \sum_{l=1}^{\infty}\sum_{m\in\mathbb{Z}} (\omega_{l,m;T}^{r;R})^{2}\psi_{l,m}(t)\psi_{l,m}(t')\sum_{e=1}^{\infty}\sum_{f\in\mathbb{Z}} (\omega_{e,f;T}^{r';R})^{2}\psi_{e,f}(h)\psi_{e,f}(h').$$ Thus the first term in equation , $\alpha$, is $$\begin{aligned} \alpha &= \sum_{t}\sum_{t'}\sum_{h}\sum_{h'}\psi_{j,k}(t)\psi_{j,k}(t')\psi_{j',k'}(h)\psi_{j',k'}(h'){\mathbf{E}}\Big[ X_{t}^{r}X_{t'}^{r}\Big]{\mathbf{E}}\left[ X_{h}^{r'}X_{h'}^{r'}\right] \\ &= \sum_{l=1}^{\infty}\sum_{m\in\mathbb{Z}} (\omega_{l,m;T}^{r;R})^{2}\sum_{t}\psi_{l,m}(t)\psi_{j,k}(t)\sum_{t'}\psi_{l,m}(t')\psi_{j,k}(t') \\ & \qquad\qquad \times \sum_{e=1}^{\infty}\sum_{f\in\mathbb{Z}} (\omega_{e,f;T}^{r';R})^{2}\sum_{h}\psi_{e,f}(h)\psi_{j',k'}(h)\sum_{h'}\psi_{e,f}(h')\psi_{j',k'}(h') \\ &= \left(\sum_{l=1}^{\infty}A_{j,l} S_{l}\left(\frac{k}{T},\frac{r}{R}\right) + {\mathcal{O}}(2^{j}T^{-1}) + {\mathcal{O}}(R^{-1})\right) \\ & \qquad\qquad \times \left(\sum_{e=1}^{\infty}A_{j',e} S_{e}\left(\frac{k'}{T},\frac{r'}{R}\right) + {\mathcal{O}}(2^{j'}T^{-1}) + {\mathcal{O}}(R^{-1})\right),\end{aligned}$$ where we used the asymptotic expectation result obtained above (see equation ). Using the result of Lemma \[lem:ASord\], we obtain $$\begin{aligned} \alpha&=\left( \sum_{l=1}^{\infty}A_{j,l} S_{l}\left(\frac{k}{T},\frac{r}{R}\right)\right) \left(\sum_{e=1}^{\infty}A_{j',e} S_{e}\left(\frac{k'}{T},\frac{r'}{R}\right)\right)+ {\mathcal{O}}(2^{j+j'}T^{-1})+{\mathcal{O}}(2^{\mbox{max}\{j,j'\}}R^{-1}),\end{aligned}$$ where we used ${\mathcal{O}}(2^{\mbox{max}\{j,j'\}}R^{-1})={\mathcal{O}}(2^{j}R^{-1})+{\mathcal{O}}(2^{j'}R^{-1})$. In the same manner as for and , we can show that $$\begin{aligned} \label{eq:EEXX2} {\mathbf{E}}\left[ X_{t}^{r}X_{h}^{r'}\right]{\mathbf{E}}\left[ X_{t'}^{r}X_{h'}^{r'}\right] &= \sum_{l=1}^{\infty}\sum_{m\in\mathbb{Z}} (\omega_{l,m;T}^{r;R})^{2}\psi_{l,m}(t)\psi_{l,m}(h)\sum_{l'=1}^{\infty}\sum_{m'\in\mathbb{Z}} (\omega_{l',m';T}^{r;R})^{2}\psi_{l',m'}(t')\psi_{l',m'}(h'), \nonumber \\ {\mathbf{E}}\left[ X_{t}^{r}X_{h'}^{r'}\right]{\mathbf{E}}\left[ X_{t'}^{r}X_{h}^{r'}\right] &= \sum_{l=1}^{\infty}\sum_{m\in\mathbb{Z}} (\omega_{l,m;T}^{r;R})^{2}\psi_{l,m}(t)\psi_{l,m}(h')\sum_{l'=1}^{\infty}\sum_{m'\in\mathbb{Z}} (\omega_{l',m';T}^{r;R})^{2}\psi_{l',m'}(t')\psi_{l',m'}(h),\end{aligned}$$ noting that the above expressions are non-zero only when $r=r'$ since ${\mathbf{E}}\left[\xi_{j,k}^{r}\xi_{j',k'}^{r'}\right] = \delta_{j,j'}\delta_{k,k'}\delta_{r,r'}$. Then the second term in , $\beta$, is $$\begin{aligned} \beta &= \sum_{t}\sum_{t'}\sum_{h}\sum_{h'}\psi_{j,k}(t)\psi_{j,k}(t')\psi_{j',k'}(h)\psi_{j',k'}(h'){\mathbf{E}}\left[ X_{t}^{r}X_{h}^{r'}\right]{\mathbf{E}}\left[ X_{t'}^{r}X_{h'}^{r'}\right] \\ &= \sum_{l=1}^{\infty}\sum_{m\in\mathbb{Z}} (\omega_{l,m;T}^{r;R})^{2}\sum_{t}\psi_{l,m}(t)\psi_{j,k}(t)\sum_{h}\psi_{l,m}(h)\psi_{j',k'}(h)\\ & \qquad\qquad\qquad\qquad\qquad \times \sum_{l'=1}^{\infty}\sum_{m'\in\mathbb{Z}}(\omega_{l',m';T}^{r;R})^{2}\sum_{t'}\psi_{l',m'}(t')\psi_{j,k}(t') \sum_{h'}\psi_{l',m'}(h')\psi_{j',k'}(h')\\ &= \left(\sum_{l=1}^{\infty}\sum_{m\in\mathbb{Z}} (\omega_{l,m;T}^{r;R})^{2}\sum_{t}\psi_{l,m}(t)\psi_{j,k}(t)\sum_{h}\psi_{l,m}(h)\psi_{j',k'}(h)\right)^{2}.\end{aligned}$$ Making the substitution $m=n+k$, and noticing from the above that the term $\beta \geq 0$, it follows that $$\begin{aligned} \beta^{\frac{1}{2}} &= \left\| \sum_{l=1}^{\infty}\sum_{n\in\mathbb{Z}} (\omega_{l,n+k;T}^{r;R})^{2} \sum_{t}\psi_{l,n+k-t}\psi_{j,k-t}\sum_{h}\psi_{l,n+k-h}\psi_{j',k'-h} \right\|,\\ &= \left\| \sum_{l=1}^{\infty}\sum_{n\in\mathbb{Z}} \left\{S_{l}\left(\frac{n+k}{T},\frac{r}{R}\right)+{\mathcal{O}}(D_{l}R^{-1})+ {\mathcal{O}}(C_{l}T^{-1}) \right\} \sum_{t}\psi_{l,n-t}\psi_{j,-t}\sum_{h}\psi_{l,n+k-h}\psi_{j',k'-h}\right\|\\ &= \left\| \sum_{l=1}^{\infty}\sum_{n\in\mathbb{Z}} \left\{S_{l}\left(\frac{k}{T},\frac{r}{R}\right)+{\mathcal{O}}(D_{l}R^{-1})+ {\mathcal{O}}(C_{l}T^{-1})+ {\mathcal{O}}(\|n\|L_{l}T^{-1}) \right\}\right. \\ &\quad \left. \times \sum_{t}\psi_{l,n-t}\psi_{j,-t}\sum_{h}\psi_{l,n+k-h}\psi_{j',k'-h}\right\|,\end{aligned}$$ where we have used the spectrum definition and the Lipschitz continuity of $S_{j}(\cdotp,r/R)$ in the rescaled time argument. Using a substitution $u=h-t$, the sums of wavelet products above can be manipulated as follows, $$\begin{aligned} \sum_{n\in\mathbb{Z}}\sum_{t}\psi_{l,n-t}\psi_{j,-t} \sum_{h}\psi_{l,n+k-h}\psi_{j',k'-h} &= \sum_{n\in\mathbb{Z}}\sum_{t}\psi_{l,n-t}\psi_{j,-t} \sum_{u}\psi_{l,n+k-u-t}\psi_{j',k'-u-t} \nonumber \\ &= \sum_{u}\sum_{t}\psi_{j,-t}\psi_{j',k'-u-t} \sum_{n\in\mathbb{Z}}\psi_{l,n-t}\psi_{l,n+k-u-t} \nonumber\\ &= \sum_{u}\Psi_{j,j'}(u-k')\Psi_{l}(u-k), \nonumber \\ &= \sum_{u}\Psi_{j,j'}(u)\Psi_{l}(u-k+k'), \nonumber $$ where for the last equality we used a substitution $u:=u-k'$. Equivalently, using the first result in the proof of Lemma \[lem:ASordjj\], the above could have been directly written as $\sum_{u}\Psi_{j,l}(u)\Psi_{j',l}(u-k+k')$. Hence \^ = \| \_[l=1]{}\^\_[u]{} \_[j,j’]{}(u)\_[l]{}(u-k+k’) S\_[l]{}(,) +I+II+III \|. The term $I$ can be bounded as follows \|I\|&=&\|\_l (D\_[l]{}R\^[-1]{})\_[u]{}\_[j,l]{}(u)\_[j’,l]{}(u-k+k’)\|,\ && (R\^[-1]{})\_[l]{} ( D\_l \_u \| \_[j,l]{}(u)\_[j’,l]{}(u-k+k’) \| ),\ && (R\^[-1]{})\_[l]{} D\_l (\_u \| \_[j,l]{}(u)\|\^2)\^[1/2]{} (\_u \| \_[j’,l]{}(u-k+k’)\|\^2)\^[1/2]{},\ &=& (R\^[-1]{})\_[l]{} D\_l (A\_[j,l]{})\^[1/2]{} (A\_[j’,l]{})\^[1/2]{}, A\_[j,l]{},\ &=& (R\^[-1]{})\_[l]{} (D\_l A\_[j,l]{})\^[1/2]{} (D\_l A\_[j’,l]{})\^[1/2]{},\ && (R\^[-1]{})(\_[l]{} D\_l A\_[j,l]{})\^[1/2]{} (\_[l]{} D\_l A\_[j’,l]{})\^[1/2]{},\ && (R\^[-1]{})(1), \_[l]{} D\_l A\_[j,l]{}=(1),\ &=& (R\^[-1]{}). Similarly, $\|II\| =\left\| \sum_l {\mathcal{O}}(C_{l}T^{-1})\sum_{u}\Psi_{j,l}(u)\Psi_{j',l}(u-k+k')\right\|={\mathcal{O}}(T^{-1})$. We bound the term $III$ by noting that the function $\Psi_{j,l}\Psi_{j',l}$ is compactly supported, with the support bounded by $K'\left( 2^l +2^{\mbox{min}\{j,j'\}}\right)$, hence \|III\|&& \_l (L\_[l]{}T\^[-1]{})( 2\^l +2\^[{j,j’}]{})\_[u]{} \|\_[j,l]{}(u)\_[j’,l]{}(u-k+k’)\|,\ && (T\^[-1]{})\_l L\_[l]{}( 2\^l +2\^[{j,j’}]{}) (A\_[j,l]{})\^[1/2]{} (A\_[j’,l]{})\^[1/2]{},\ &=& (T\^[-1]{})\_l L\_[l]{}2\^l (A\_[j,l]{})\^[1/2]{} (A\_[j’,l]{})\^[1/2]{}\ &+& (T\^[-1]{})\_l L\_[l]{}2\^[{j,j’}]{} (A\_[j,l]{})\^[1/2]{} (A\_[j’,l]{})\^[1/2]{},\ && (T\^[-1]{})2\^[(j+j’)/2]{}\_l L\_[l]{}2\^l (2\^[-j/2]{}A\_[j,l]{}\^[1/2]{}) (2\^[-j’/2]{}A\_[j’,l]{}\^[1/2]{})\ &+& (T\^[-1]{})2\^[(j+j’)/2]{}\_l L\_[l]{} (A\_[j,l]{})\^[1/2]{} (A\_[j’,l]{})\^[1/2]{}, 2\^[{j,j’}]{} 2\^[(j+j’)/2]{}. The term $\sum_l L_{l}2^l (2^{-j/2}A_{j,l}^{1/2}) (2^{-j'/2}A_{j',l}^{1/2})\leq \left(\sum_l L_{l}2^l 2^{-j}A_{j,l}\right)^{1/2}\left(\sum_l L_{l}2^l 2^{-j'}A_{j',l}\right)^{1/2}$ from the Cauchy-Schwarz inequality, and using $\sum_{j} 2^{-j}A_{j,l}=1$ and $\sum_{l} 2^l L_l<\infty$, we obtain its ${\mathcal{O}}(1)$ bound. The term $\sum_l L_{l} (A_{j,l})^{1/2} (A_{j',l})^{1/2} \leq \left(\sum_l L_{l}A_{j,l}\right)^{1/2} \left(\sum_l L_{l}A_{j',l}\right)^{1/2}={\mathcal{O}}(1)$ since we already established that $\sum_l L_{l}A_{j,l}<\infty$ in the expectation part of the proof. Thus term $III$ is bounded by ${\mathcal{O}}(2^{(j+j')/2}T^{-1})$. We therefore have \^ = \| \_[l=1]{}\^\_[u]{} \_[j,j’]{}(u)\_[l]{}(u-k+k’) S\_[l]{}(,) \| +(2\^[(j+j’)/2]{}T\^[-1]{})+(R\^[-1]{}), and using the result in Lemma \[lem:ASordjj\], we obtain = (\_[l=1]{}\^\_[u]{} \_[j,j’]{}(u)\_[l]{}(u-k+k’) S\_[l]{}(,))\^2 +(2\^[(j+j’)/2+(j+j’)/2]{}T\^[-1]{})+ (2\^[(j+j’)/2]{}R\^[-1]{}), where we retained the highest order terms only. The third term in , $\gamma$, can be establish using precisely the same arguments. Using all previous results, when $r \neq r'$ $$\begin{aligned} {\mathbf{E}}\left[I_{j,k;T}^{r;R}I_{j',k';T}^{r';R}\right] &= \alpha + \beta + \gamma = \alpha \nonumber\\ &= \sum_{l=1}^{\infty}A_{j,l} S_{l}\left(\frac{k}{T},\frac{r}{R}\right)\sum_{e=1}^{\infty}A_{j',e} S_{e}\left(\frac{k'}{T},\frac{r'}{R}\right)+ {\mathcal{O}}(2^{j+j'}T^{-1}) \nonumber \\ & \quad +{\mathcal{O}}(2^{\text{max}\{j,j'\}} R^{-1}),\end{aligned}$$ as the $\beta$ and $\gamma$ terms are both $0$ due to . When $r = r'$, we obtain the following $$\begin{aligned} \label{eq:covIdt} {\mathbf{E}}\left[I_{j,k;T}^{r;R}I_{j',k';T}^{r;R}\right] &= \alpha + \beta + \gamma \nonumber\\ &= \sum_{l=1}^{\infty}A_{j,l} S_{l}\left(\frac{k}{T},\frac{r}{R}\right)\sum_{e=1}^{\infty}A_{j',e} S_{e}\left(\frac{k'}{T},\frac{r}{R}\right)+ {\mathcal{O}}(2^{j+j'}T^{-1})+{\mathcal{O}}(2^{\text{max}(j,j')} R^{-1}) \nonumber \\ & \quad+ 2\left(\sum_{l=1}^{\infty}\sum_{u\in\Z}\Psi_{j,j'}(u)\Psi_{l}(u+k-k')S_{l}\left(\frac{k}{T},\frac{r}{R}\right) \right)^2 \nonumber \\ & \quad+ {\mathcal{O}}(2^{j+j'}T^{-1}) +{\mathcal{O}}(2^{(j+j')/2}R^{-1}).$$ Thus for $r = r', j = j'$ and $k = k'$, the variance of $I_{j,k;T}^{r;R}$ can be written as $$\begin{aligned} \text{var}\left(I_{j,k;T}^{r;R}\right) &= {\mathbf{E}}\left[I_{j,k;T}^{r;R}I_{j,k;T}^{r;R}\right] - {\mathbf{E}}\left[I_{j,k;T}^{r;R}\right]{\mathbf{E}}\left[I_{j,k;T}^{r;R}\right] \\ &= 3\left(\sum_{l=1}^{\infty}A_{j,l} S_{l}\left(\frac{k}{T},\frac{r}{R}\right)\right)^{2} + {\mathcal{O}}(2^{2j}T^{-1})+{\mathcal{O}}(2^{j}R^{-1}) \\ & \qquad\qquad\qquad\qquad\qquad - \left(\sum_{l=1}^{\infty}A_{j,l} S_{l}\left(\frac{k}{T},\frac{r}{R}\right) + {\mathcal{O}}(2^{j}T^{-1}) + {\mathcal{O}}(R^{-1})\right)^{2} \\ &= 2\left(\sum_{l=1}^{\infty}A_{j,l} S_{l}\left(\frac{k}{T},\frac{r}{R}\right)\right)^{2} + {\mathcal{O}}(2^{2j}T^{-1}) + {\mathcal{O}}(2^{j}R^{-1}),\end{aligned}$$ where for the last equality we used the result in Lemma \[lem:ASord\] and we have only kept the highest order terms. Proof of Proposition \[prop:ESI\] {#app:proofs:itilde} ---------------------------------- ### Proof of Proposition \[prop:ESI\] (Expectation) {#proof-of-proposition-propesi-expectation .unnumbered} From the definition of the replicate-smoothed periodogram in , we have $$\nonumber {\mathbf{E}}\left[{\tilde{I}}_{j,k;T}^{r;R}\right] = \frac{1}{2M+1}\sum_{s=-M}^{M}{\mathbf{E}}\left[I_{j,k;T}^{r+s;R}\right]$$ and substituting the asymptotic result for the expectation ${\mathbf{E}}\left[I_{j,k;T}^{r+s;R}\right]$ (see equation ), we further obtain $$\nonumber {\mathbf{E}}\left[{\tilde{I}}_{j,k;T}^{r;R}\right] = (2M+1)^{-1}\sum_{s=-M}^{M}\left\{\sum_{l=1}^{\infty}A_{j,l} S_{l}\left(\frac{k}{T},\frac{r+s}{R}\right) + {\mathcal{O}}(2^jT^{-1}) + {\mathcal{O}}(R^{-1})\right\}.$$ Using the Lipschitz continuity of the spectrum $S_{j}(k/T,\cdotp)$ in replicate-dimension, we obtain $$\begin{aligned} {\mathbf{E}}\left[{\tilde{I}}_{j,k;T}^{r;R}\right] &= (2M+1)^{-1}\sum_{s=-M}^{M}\left\{\sum_{l=1}^{\infty}A_{j,l} \left(S_{l}\left(\frac{k}{T},\frac{r}{R}\right)+ {\mathcal{O}}(\|s\|N_{l}R^{-1})\right) + {\mathcal{O}}(2^{j}T^{-1}) + {\mathcal{O}}(R^{-1})\right\}\\ &= (2M+1)^{-1}\sum_{s=-M}^{M}\left\{\sum_{l=1}^{\infty}A_{j,l} S_{l}\left(\frac{k}{T},\frac{r}{R}\right)+ {\mathcal{O}}(\|s\|R^{-1})\sum_l N_l A_{j,l}+ {\mathcal{O}}(2^{j}T^{-1}) + {\mathcal{O}}(R^{-1})\right\} \\ &= \sum_{l=1}^{\infty}A_{j,l} S_{l}\left(\frac{k}{T},\frac{r}{R}\right)+ {\mathcal{O}}(MR^{-1}) + {\mathcal{O}}(2^j T^{-1}),\end{aligned}$$ where we have used that $\sum_l N_l A_{j,l} < \infty$ as $\sum_{l}2^{l}N_{l} < \infty$ and $\sum_l 2^{-l} A_{j,l}=1$.\ ### Proof of Proposition \[prop:ESI\] (Variance) {#proof-of-proposition-propesi-variance .unnumbered} Let us take $$\begin{aligned} \label{eq:CSII} \text{var}\left({\tilde{I}}_{j,k;T}^{r;R}\right)&= \text{var}\left( (2M+1)^{-1} \sum_{s=-M}^{M} I_{j,k;T}^{r+s;R} \right) \nonumber\\ & = (2M+1)^{-2} \left( {\mathbf{E}}\left[\sum_{s=-M}^{M} I_{j,k;T}^{r+s;R}\sum_{s'=-M}^{M} I_{j,k;T}^{r+s';R} \right] - {\mathbf{E}}\left[\sum_{s=-M}^{M} I_{j,k;T}^{r+s;R}\right]{\mathbf{E}}\left[\sum_{s'=-M}^{M} I_{j,k;T}^{r+s';R}\right] \right) \nonumber\\ &= (2M+1)^{-2}\sum_{s=-M}^{M}\sum_{s'=-M}^{M} {\mathrm{cov}}\left(I_{j,k;T}^{r+s;R},I_{j,k;T}^{r+s';R}\right).\end{aligned}$$ As the replicates are uncorrelated, we have $s = s'$ in and using the variance result for the raw periodogram (Proposition \[prop:EI\]) we obtain $$\begin{aligned} \text{var}\left({\tilde{I}}_{j,k;T}^{r;R}\right) &=(2M+1)^{-2}\sum_{s=-M}^{M}\text{var}\left(I_{j,k;T}^{r+s;R}\right) \nonumber \\ &= (2M+1)^{-2}\sum_{s=-M}^{M}\left\{2\left(\sum_{l=1}^{\infty}A_{j,l} S_{l}\left(\frac{k}{T},\frac{r+s}{R}\right)\right)^{2} + {\mathcal{O}}(2^{2j}T^{-1}) + {\mathcal{O}}(2^{j}R^{-1})\right\} \nonumber\\ &= (2M+1)^{-2}\sum_{s=-M}^{M}\left\{2\left(\sum_{l=1}^{\infty}A_{j,l} S_{l}\left(\frac{k}{T},\frac{r}{R}\right)+ {\mathcal{O}}(\|s\|R^{-1})\sum_l N_l A_{j,l} \right)^{2}\right\} \\ &\quad +(2M+1)^{-2}\sum_{s=-M}^{M}\left\{{\mathcal{O}}(2^{2j}T^{-1}) + {\mathcal{O}}(2^{j}R^{-1})\right\} \nonumber\\ &={\mathcal{O}}(M^{-1})\left(\sum_{l=1}^{\infty}A_{j,l}S_{l}\left(\frac{k}{T},\frac{r}{R}\right)\right)^{2}+ (2M+1)^{-2}\sum_{s=-M}^{M}\left\{ {\mathcal{O}}(2^j\|s\|R^{-1})+ {\mathcal{O}}(\|s\|^2R^{-2})\right\} \\ &\quad +{\mathcal{O}}(2^{2j}(MT)^{-1}) + {\mathcal{O}}(2^{j}(MR)^{-1}), $$ from Lemma \[lem:ASord\] and as $\sum_l N_l A_{j,l}<\infty$. Retaining the largest order terms, it then follows that $$\begin{aligned} \text{var}\left({\tilde{I}}_{j,k;T}^{r;R}\right) &={\mathcal{O}}(2^{2j}M^{-1})+{\mathcal{O}}(2^jR^{-1})+{\mathcal{O}}(MR^{-2})+{\mathcal{O}}(2^{2j}(MT)^{-1}) + {\mathcal{O}}(2^{j}(MR)^{-1}),\\ &={\mathcal{O}}(2^{2j}M^{-1})+{\mathcal{O}}(2^jR^{-1})+{\mathcal{O}}(MR^{-2}).\end{aligned}$$ The expectation and variance results show that for fixed coarse enough scales $j$ (to guard against asymptotic bias and non-vanishing variance), the proposed replicate-smoothed periodogram is an asymptotically consistent estimator for the spectral quantity $\beta$, as it is asymptotically unbiased and its variance converges to zero as $T \to \infty, R \to \infty$, $M \to \infty$ and $M/R \to 0$.\ Proof of Proposition \[prop:CS\] {#app:proofs:corrper} --------------------------------- As $M, T \rightarrow \infty$, for each $j$, $z$ and $\nu$, the consistency result $\hat{S}_{j}(z,\nu)\stackrel{P}{\rightarrow} {S}_{j}(z,\nu)$ follows from the consistency results ${\tilde{I}}_{l,\lfloor zT \rfloor;T}^{\lfloor \nu R \rfloor;R} \stackrel{P}{\rightarrow} \beta_{l}(z,\nu)$ for all fine enough scales $l$ (as shown in Proposition \[prop:ESI\]) and then using the continuous mapping theorem [@billingsley1999convergence] for the continuous function $g(x_1, \ldots, x_J)= \sum_{l=1}^J A^{-1}_{j,l} x_l$ that defines their linear combination with coefficients given by the matrix $A^{-1}$ entries. Additionally, using the properties of the matrix $A$, we obtain the estimator asymptotic unbiasedness from the linearity of the expectation operator and from the asymptotic unbiasedness of the corrected periodogram, as follows (\_[j]{}(z,))&=&(\_[l=1]{}\^J A\^[-1]{}\_[j,l]{} \_[l,zT ;T]{}\^[R ;R]{}),\ &=& \_[l=1]{}\^J A\^[-1]{}\_[j,l]{} (\_[l,zT ;T]{}\^[R ;R]{}),\ &=&\_[l=1]{}\^J A\^[-1]{}\_[j,l]{} (\_[l]{}(z,) + (2\^l T\^[-1]{})+(MR\^[-1]{})),\ &=& \_[l=1]{}\^J A\^[-1]{}\_[j,l]{}\_[l’]{} A\_[l,l’]{}S\_[l’]{}(z,)+ \_[l=1]{}\^J A\^[-1]{}\_[j,l]{}( (2\^l T\^[-1]{})+(MR\^[-1]{}) ),\ &=& \_[l’]{} (\_[l=1]{}\^J A\^[-1]{}\_[j,l]{}A\_[l,l’]{})S\_[l’]{}(z,)+ \_[l=1]{}\^J A\^[-1]{}\_[j,l]{}( (2\^l T\^[-1]{})+(MR\^[-1]{}) ),\ &=& \_[l’]{} ( A\^[-1]{}A)\_[j,l’]{}S\_[l’]{}(z,) + \_[l=1]{}\^J A\^[-1]{}\_[j,l]{}( (2\^l T\^[-1]{})+(MR\^[-1]{}) ),\ &=& S\_[j]{}(z,)+(T\^[-1]{})+(MR\^[-1]{}), where we used the boundedness of $A^{-1}_{j,l}$ and $\sum_{l=1}^J 2^l ={\mathcal{O}}(T^\alpha)$. In fact, it can be shown that for Haar wavelets, the above approximation rate is ${\mathcal{O}}(T^{-1/2})+{\mathcal{O}}(MR^{-1})$ since $A^{-1}_{j,l} ={\mathcal{O}}(2^{-(j+l)/2})$ [@nvsk:2000]. Proof of Proposition \[prop:ESIts\] {#app:proofs:itildetilde} ------------------------------------ ### Proof of Proposition \[prop:ESIts\] (Expectation) {#proof-of-proposition-propesits-expectation .unnumbered} From the definition of the time- and replicate-smoothed periodogram in , we have $$\nonumber {\mathbf{E}}\left[\tilde{{\tilde{I}}}_{j,k;T}^{r;R}\right] = \frac{1}{2M+1}\frac{1}{2M_T+1}\sum_{s=-M}^{M}\sum_{t=-M_T}^{M_T}{\mathbf{E}}\left[I_{j,k+t;T}^{r+s;R}\right]$$ and substituting the asymptotic result for the expectation ${\mathbf{E}}\left[I_{j,k+t;T}^{r+s;R}\right]$ (see for instance equation ), we further obtain $${\mathbf{E}}\left[\tilde{{\tilde{I}}}_{j,k;T}^{r;R}\right] = (2M+1)^{-1}(2M_T+1)^{-1}\sum_{s=-M}^{M} \sum_{t=-M_T}^{M_T}\left\{\sum_{l=1}^{\infty}A_{j,l} S_{l}\left(\frac{k+t}{T},\frac{r+s}{R}\right) + {\mathcal{O}}(2^jT^{-1}) + {\mathcal{O}}(R^{-1})\right\}.$$ Now using the Lipschitz continuity of the spectrum $S_{l}(\cdotp,\cdotp)$ in both time and replicate dimensions, we have $$\begin{aligned} {\mathbf{E}}\left[\tilde{{\tilde{I}}}_{j,k;T}^{r;R}\right] &= (2M+1)^{-1}(2M_T+1)^{-1}\sum_{s=-M}^{M}\sum_{t=-M_T}^{M_T} \left\{\sum_{l=1}^{\infty}A_{j,l} S_{l}\left(\frac{k}{T},\frac{r}{R}\right)\right\}\\ &+(2M+1)^{-1}(2M_T+1)^{-1}\sum_{s=-M}^{M}\sum_{t=-M_T}^{M_T}\sum_{l=1}^{\infty}A_{j,l} \left({\mathcal{O}}(\|t\|L_{l}T^{-1})+{\mathcal{O}}(\|s\|N_{l}R^{-1})\right)\\ &+ {\mathcal{O}}(2^{j}T^{-1}) + {\mathcal{O}}(R^{-1}),\\ &= (2M+1)^{-1}(2M_T+1)^{-1}\sum_{s=-M}^{M}\sum_{t=-M_T}^{M_T} \sum_{l=1}^{\infty}A_{j,l} S_{l}\left(\frac{k}{T},\frac{r}{R}\right)\\ &+ (2M+1)^{-1}(2M_T+1)^{-1}\sum_{s=-M}^{M}\sum_{t=-M_T}^{M_T} \left\{ {\mathcal{O}}(\|t\|T^{-1})\sum_l L_l A_{j,l} + {\mathcal{O}}(\|s\|R^{-1})\sum_l N_l A_{j,l} \right\} \\ &+ {\mathcal{O}}(2^{j}T^{-1}) + {\mathcal{O}}(R^{-1}),\\ &= \sum_{l=1}^{\infty}A_{j,l} S_{l}\left(\frac{k}{T},\frac{r}{R}\right)+ {\mathcal{O}}(M_T T^{-1})+ {\mathcal{O}}(MR^{-1}) + {\mathcal{O}}(2^j T^{-1}),\end{aligned}$$ where we have used that $\sum_l L_l A_{j,l} < \infty$ and $\sum_l N_l A_{j,l} < \infty$ as previously shown, and as usual we retained the highest order terms. ### Proof of Proposition \[prop:ESIts\] (Variance) {#proof-of-proposition-propesits-variance .unnumbered} Under the assumption of summable autocovariance, $$\sup_{z,\nu}\sum_{n\in\Z}\left\|c\left(z,\nu;n \right)\right\| ={\mathcal{O}}(1),$$ the result in Lemma \[lem:ASordtau\] further implies $$\label{eq:Aeta} \sum_{\uptau\in \Z}\left\|\sum_{l=1}^{\infty}A_{l,j}^{\uptau} S_{l}\left(z,\nu\right)\right\|={\mathcal{O}}(2^{j}).$$ This can be easily seen by taking $$\begin{aligned} \sum_{\uptau}\left\|\sum_{l=1}^{\infty}A_{l,j}^{\uptau} S_{l}\left(z,\nu\right)\right\| & = \sum_{\uptau} \left\|\sum_{l=1}^{\infty}\sum _{n}\Psi_{l}(n) \Psi_{j}(n+\uptau) S_{l}\left(z,\nu\right)\right\| \mbox{ from the definition of the matrix }A^\uptau\\ &= \sum_{\uptau} \left\|\sum _{n}\left(\sum_{l=1}^{\infty} S_{l}\left(z,\nu\right)\Psi_{l}(n)\right)\Psi_{j}(n+\uptau)\right\| \\ &= \sum_{\uptau} \left\|\sum _{n}c\left(z,\nu;n \right)\Psi_{j}(n+\uptau)\right\| \mbox { from the local autocovariance definition}\\ & \leq \sum _{n} \left( \sum_{\uptau} \left\|\Psi_{j}(n+\uptau)\right\| \right) \left\|c\left(z,\nu;n \right)\right\| \mbox { using the triangle inequality}\\ &= {\mathcal{O}}(2^{j}) \sum_n \left\|c\left(z,\nu;n \right)\right\| = {\mathcal{O}}(2^{j}),\end{aligned}$$ where we used $\sup_{z,\nu}\sum_n \left\|c\left(z,\nu;n \right)\right\|={\mathcal{O}}(1)$ and that $\sum _{\uptau}\left\|\Psi_{j}(n+\uptau)\right\| = {\mathcal{O}}(2^{j})$. In the same vein, we show that $$\label{eq:Aetal}\sum_{\uptau\in\Z} \left\|\sum_{l=1}^\infty L_{l} A_{j,l}^{\uptau} \right\|= {\mathcal{O}}(2^j),$$ and similarly for terms involving $N_l$, since $$\begin{aligned} \sum_{\uptau}\left\|\sum_{l=1}^{\infty}L_l A_{l,j}^{\uptau} \right\| & = \sum_{\uptau} \left\|\sum_{l=1}^{\infty} L_l \sum _{n}\Psi_{j}(n) \Psi_{l}(n+\uptau) \right\| \mbox{ from the definition of the matrix }A^\uptau\\ &\leq \sum_{l=1}^{\infty} L_l \sum_{\uptau} \sum _{n} \left\| \Psi_{j}(n)\Psi_{l}(n+\uptau)\right\| \mbox { using the triangle inequality} \\ &= \sum_{l=1}^{\infty} L_l \sum _{n} \left(\left\| \Psi_{j}(n)\right\| \sum_{\uptau}\left\| \Psi_{l}(n+\uptau)\right\| \right)\\ & \leq K \sum_{l=1}^{\infty} 2^l L_l \sum _{n} \left\| \Psi_{j}(n)\right\| \\ &= {\mathcal{O}}(2^{j}),\end{aligned}$$ where we used (in order) that $\sum _{\uptau}\left\|\Psi_{l}(n+\uptau)\right\| = {\mathcal{O}}(2^{l})$, $\sum _{n}\left\|\Psi_{j}(n)\right\| = {\mathcal{O}}(2^{j})$ and $\sum_l 2^l L_l <\infty$. We also note here that $\left\| \sum_{l=1}^\infty 2^{-l} A_{j,l}^{\uptau} \right\|= {\mathcal{O}}(1)$, which can be immediately obtained by taking $$\begin{aligned} \left\| \sum_l 2^{-l} A_{j,l}^{\uptau} \right\|&= \left\|\sum_{l} 2^{-l} \sum _{n}\Psi_{j}(n) \Psi_{l}(n+\uptau) \right\|,\\ &= \left\|\sum _{n}\Psi_{j}(n) \sum_{l} 2^{-l} \Psi_{l}(n+\uptau) \right\|,\\ &= \left\|\sum _{n}\Psi_{j}(n) \delta_{0,n+\tau} \right\|, \\ &=\left\| \Psi_{j}(-\uptau) \right\|= {\mathcal{O}}(1).\end{aligned}$$ where we used $\sum_{l} 2^{-l} \Psi_{l}(n+\uptau)=\delta_{0,n+\tau}$ as shown in [@fryz03:forecasting]. Now let us take $$\begin{aligned} \label{eq:CSIIts} \text{var}\left(\tilde{{\tilde{I}}}_{j,k;T}^{r;R}\right)&= \text{var}\left( (2M_T+1)^{-1}\sum_{t=-M_T}^{M_T} {\tilde{I}}_{j,k+t;T}^{r;R} \right), \nonumber\\ &= (2M_T+1)^{-2} \sum_{t=-M_T}^{M_T} \sum_{\eta=-M_T-t}^{M_T-t} {\mathrm{cov}}\left({\tilde{I}}_{j,k+t;T}^{r;R}, {\tilde{I}}_{j,k+t+\eta;T}^{r;R}\right), \nonumber\\ &= (2M_T+1)^{-2} (2M+1)^{-2} \sum_{t=-M_T}^{M_T} \sum_{\eta=-M_T-t}^{M_T-t} \sum_{s=-M}^{M} {\mathrm{cov}}\left(I_{j,k+t;T}^{r+s;R}, I_{j,k+t+\eta;T}^{r+s;R}\right),\end{aligned}$$ where we have used the definition of the replicate-smoothed periodogram in and the assumption of uncorrelated replicate series. Using the covariance definition and the result in equation of the variance proof for the raw periodogram (Appendix \[app:proofs:rawi\] for Proposition \[prop:EI\]), as well as the expectation result for the raw periodogram (Proposition \[prop:EI\]), we can re-write equation above as follows $$\begin{aligned} \text{var}\left(\tilde{{\tilde{I}}}_{j,k;T}^{r;R}\right) &= (2M_T+1)^{-2} (2M+1)^{-2} \sum_{t=-M_T}^{M_T} \sum_{\eta=-M_T-t}^{M_T-t} \sum_{s=-M}^{M} \left\{2\left(\sum_{l=1}^{\infty}A_{j,l}^{\eta} S_{l}\left(\frac{k+t}{T},\frac{r+s}{R}\right)\right)^{2}\right\} \\ &+ (2M_T+1)^{-2} (2M+1)^{-2} \sum_{t=-M_T}^{M_T} \sum_{\eta=-M_T-t}^{M_T-t} \sum_{s=-M}^{M} \left\{ {\mathcal{O}}(2^{2j}T^{-1}) + {\mathcal{O}}(2^{j}R^{-1})\right\}.\end{aligned}$$ From the Lipschitz continuity of $S_{l}(\cdotp,\cdotp)$, we have $$\begin{aligned} \sum_{l=1}^{\infty}A_{j,l}^{\eta} S_{l}\left(\frac{k+t}{T},\frac{r+s}{R}\right)&= \sum_{l=1}^{\infty}A_{j,l}^{\eta} S_{l}\left(\frac{k}{T},\frac{r}{R}\right)+ \left({\mathcal{O}}(\|t\|T^{-1})\sum_l L_l A_{j,l}^{\eta}+ {\mathcal{O}}(\|s\|R^{-1})\sum_l N_l A_{j,l}^{\eta}\right),\\ &= \sum_{l=1}^{\infty}A_{j,l}^{\eta} S_{l}\left(\frac{k}{T},\frac{r}{R}\right)+ \left(\mbox{term }I \right), $$ where the notation ‘term $I$’ above is used for brevity. Replacing this into the variance formula above, we obtain $$\begin{aligned} \text{var}\left(\tilde{{\tilde{I}}}_{j,k;T}^{r;R}\right) &\leq (2M_T+1)^{-2} (2M+1)^{-2} \sum_{t=-M_T}^{M_T} \sum_{s=-M}^{M} \sum_{\eta\in\Z} 2\left(\sum_{l=1}^{J}A_{j,l}^{\eta} S_{l}\left(\frac{k}{T},\frac{r}{R}\right)\right)^2\\ &+ (2M_T+1)^{-2} (2M+1)^{-2} \sum_{t=-M_T}^{M_T} \sum_{s=-M}^{M} \sum_{\eta\in\Z} 2\left(\mbox{term }I\right)^2 \\ &+ (2M_T+1)^{-2} (2M+1)^{-2} \sum_{t=-M_T}^{M_T}\sum_{s=-M}^{M} \sum_{\eta\in\Z} 4\left(\sum_{l=1}^{J}A_{j,l}^{\eta}S_{l}\left(\frac{k}{T},\frac{r}{R}\right)\right) \times \left(\mbox{term }I\right) \\ &+(2M_T+1)^{-2} (2M+1)^{-2} \sum_{t=-M_T}^{M_T} \sum_{\eta=-M_T-t}^{M_T-t} \sum_{s=-M}^{M}\left\{{\mathcal{O}}(2^{2j}T^{-1}) + {\mathcal{O}}(2^{j}R^{-1})\right\}.\end{aligned}$$ As $\sum_{\eta} \left(\sum_{l}A_{j,l}^{\eta} S_{l}\left(\frac{k}{T},\frac{r}{R}\right)\right)^2 \leq \left(\sum_{\eta} \left\|\sum_{l}A_{j,l}^{\eta} S_{l}\left(\frac{k}{T},\frac{r}{R}\right)\right\| \right)^2$, from we obtain\ $\sum_{\eta} \left(\sum_{l}A_{j,l}^{\eta} S_{l}\left(\frac{k}{T},\frac{r}{R}\right)\right)^2={\mathcal{O}}(2^{2j})$, hence the first term in the variance sum above is ${\mathcal{O}}(2^{2j}(M_TM)^{-1})$. To determine the order of the second term in the variance sum, as above we first observe that $\sum_\eta \left(\sum_l L_l A_{j,l}^{\eta} \right)={\mathcal{O}}(2^{2j})$ from equation , and similarly for the terms involving $N_l$ (and cross-terms), which leads to the second term in the variance sum above to be rephrased as $$(2M_T+1)^{-2} (2M+1)^{-2} {\mathcal{O}}(2^{2j}) \sum_{t=-M_T}^{M_T} \sum_{s=-M}^{M} \left\{ {\mathcal{O}}(\|t\|^2T^{-2})+{\mathcal{O}}(\|s\|^2R^{-2})+{\mathcal{O}}(\|t\|\|s\|(TR)^{-1})\right\}.$$ The third (cross-)term of the variance sum can be obtained from $$\sum_\eta \left\|\sum_{l}A_{j,l}^{\eta} S_{l}\left(\frac{k}{T},\frac{r}{R}\right)\right\| \left\|\sum_l L_l A_{j,l}^{\eta} \right\|={\mathcal{O}}(2^{2j}),$$ using similar arguments as above and results and ; the same holds for terms involving $N_l$. Hence the third term can be re-expressed as $$(2M_T+1)^{-2} (2M+1)^{-2} {\mathcal{O}}(2^{2j}) \sum_{t=-M_T}^{M_T} \sum_{s=-M}^{M} \left\{{\mathcal{O}}(\|t\|T^{-1})+ {\mathcal{O}}(\|s\|R^{-1})\right\}.$$ Now replacing all order terms in the variance formula above leads to $$\begin{aligned} \text{var}\left(\tilde{{\tilde{I}}}_{j,k;T}^{r;R}\right) &= {\mathcal{O}}(2^{2j}(M_TM)^{-1})\\ &+ {\mathcal{O}}(2^{2j}M_T T^{-2}M^{-1})+{\mathcal{O}}(2^{2j}(M_T)^{-1}M R^{-2})+{\mathcal{O}}(2^{2j}(TR)^{-1})\\ &+ {\mathcal{O}}(2^{2j}(MT)^{-1})+{\mathcal{O}}(2^{2j}(M_TR)^{-1})\\ &+ {\mathcal{O}}(2^{2j}(MT)^{-1}) + {\mathcal{O}}(2^{j}(MR)^{-1}).\end{aligned}$$ Retaining the largest order terms, it then follows that $$\begin{aligned} \text{var}\left(\tilde{{\tilde{I}}}_{j,k;T}^{r;R}\right) &={\mathcal{O}}(2^{2j}(M_TM)^{-1})+{\mathcal{O}}(2^{j}(MR)^{-1}),\end{aligned}$$ which completes the proof. Proofs of results on the asymptotic behaviour of proposed estimators embedding replicate coherence {#app:proofsc} ================================================================================================== In this section, we give details of the proofs in Section \[sec:modelext\], using the notation described therein. In the proofs that follow, we make use of the following results. \[lem:lipz\] Under the assumptions of Definition \[def:rlswc\], we have a sequence $\{B_j\}$ of uniformly bounded Lipschitz constants in $j$ with $\sum_j2^jB_j<\infty$ such that $$\begin{aligned} & \left\|{\widetilde{W}}_{j}\left(\frac{k+n}{T},\frac{r}{R}\right){\widetilde{W}}_{j}\left(\frac{k+n}{T},\frac{r'}{R}\right)\rho_{j}\left(\frac{k+n}{T},\frac{r}{R},\frac{r'}{R}\right) \right. \nonumber \\ & \qquad \qquad \left. - {\widetilde{W}}_{j}\left(\frac{k}{T},\frac{r}{R}\right){\widetilde{W}}_{j}\left(\frac{k}{T},\frac{r'}{R}\right)\rho_{j}\left(\frac{k}{T},\frac{r}{R},\frac{r'}{R}\right)\right\| \leq \|n\| B_{j}T^{-1}, \label{eq:crosslipz}\end{aligned}$$ for any replicates $r$, $r'$ and times $k$, $n$. Note that the above result means that the replicate cross-spectrum, i.e. $S_{j}(\cdotp,\nu,\nu')$, is Lipschitz continuous in the rescaled time argument for any rescaled replicates $\nu$, $\nu'$. The proof can be seen in Section \[supp:proofs:sec5\] of the Supplementary Material. \[lem:lipr\] Under the assumptions of Definition \[def:rlswc\], we have a sequence $\{B_j^\prime\}$ of uniformly bounded Lipschitz constants in $j$ with $\sum_j2^jB^\prime_j<\infty$ such that $$\begin{aligned} & \left\|{\widetilde{W}}_{j}\left(\frac{k}{T},\frac{r+s}{R}\right){\widetilde{W}}_{j}\left(\frac{k}{T},\frac{r'+s}{R}\right)\rho_{j}\left(\frac{k}{T},\frac{r+s}{R},\frac{r'+s}{R}\right) \right. \nonumber\\ & \qquad \qquad \left. - {\widetilde{W}}_{j}\left(\frac{k}{T},\frac{r}{R}\right){\widetilde{W}}_{j}\left(\frac{k}{T},\frac{r'}{R}\right)\rho_{j}\left(\frac{k}{T},\frac{r}{R},\frac{r'}{R}\right)\right\| \leq \|s\| B^\prime_{j}R^{-1}, $$ for any times $k$, $n$ and replicates $r$, $r'$. Note that the above result effectively states that the replicate cross-spectrum, i.e. $S_{j}(z,\nu+\cdotp,\nu'+ \cdotp)$, is Lipschitz continuous in the rescaled replicate arguments. The proof can be seen in Section \[supp:proofs:sec5\] of the Supplementary Material. \[lem:ASordc\] Under the assumptions of Definition \[def:rlswc\], we have $\sum_{l=1}^{J}A_{j,l} S_{l}\left(z,\nu,\nu'\right)={\mathcal{O}}(2^{j})$. The proof follows the same steps as for Lemma \[lem:ASord\]. \[lem:ASordjjc\] Under the assumptions of Definition \[def:rlswc\], we have $$\sum_{l=1}^{J} \sum _{n}\Psi_{l}(n+\uptau) \Psi_{j,j'}(n) S_{l}\left(z,\nu,\nu'\right)={\mathcal{O}}(2^{\mbox{max}\{j,j'\}}).$$ The proof follows the same steps as for Lemma \[lem:ASordtau\]. Proof of Proposition \[prop:ESIc\] {#app:proofs:itildec} ---------------------------------- ### Proof of Proposition \[prop:ESIc\] (Expectation) {#proof-of-proposition-propesic-expectation .unnumbered} The proof follows similar steps to the proof of its non-coherence counterpart in Appendix \[app:proofs:itilde\]. ### Proof of Proposition \[prop:ESIc\] (Variance) {#proof-of-proposition-propesic-variance .unnumbered} $$\begin{aligned} \label{eq:CSIIc} \text{var}\left({\tilde{I}}_{j,k;T}^{(r,r');R}\right)&= \text{var}\left( (2M+1)^{-1} \sum_{s=-M}^{M} I_{j,k;T}^{(r+s,r'+s);R} \right) \nonumber\\ &= (2M+1)^{-2}\sum_{s=-M}^{M}\sum_{s'=-M}^{M} {\mathrm{cov}}\left(I_{j,k;T}^{(r+s,r'+s);R},I_{j,k;T}^{(r+s',r'+s');R}\right) \nonumber\\ &= (2M+1)^{-2}\sum_{s=-M}^{M}\sum_{\eta=-M-s}^{M-s} {\mathrm{cov}}\left(I_{j,k;T}^{(r+s,r'+s);R},I_{j,k;T}^{(r+s+\eta,r'+s+\eta);R}\right)\end{aligned}$$ where we have let $\eta=s'-s$. Let us now take $$\begin{aligned} \label{eq:covtau} {\mathrm{cov}}\left(I_{j,k;T}^{(r+s,r'+s);R},I_{j,k;T}^{(r+s+\eta,r'+s+\eta);R}\right) &= {\mathbf{E}}\left[d_{j,k}^{ r+s} d_{j,k}^{ r'+s}d_{j,k}^{ r+s+\eta} d_{j,k}^{ r'+s+\eta}\right] \nonumber \\ & \qquad - {\mathbf{E}}\left[d_{j,k}^{ r+s} d_{j,k}^{ r'+s}\right]{\mathbf{E}}\left[d_{j,k}^{ r+s+\eta} d_{j,k}^{ r'+s+\eta}\right] \nonumber \\ &= {\mathbf{E}}\left[d_{j,k}^{ r+s} d_{j,k}^{ r+s+\eta}\right]{\mathbf{E}}\left[d_{j,k}^{ r'+s} d_{j,k}^{ r'+s+\eta}\right] \nonumber \\ & \qquad + {\mathbf{E}}\left[d_{j,k}^{ r+s} d_{j,k}^{ r'+s+\eta}\right]{\mathbf{E}}\left[d_{j,k}^{ r'+s} d_{j,k}^{ r+s+\eta}\right],\end{aligned}$$ where we have use Isserlis’ theorem [@isserlis:1918] in the last equality. Using the expectation of the wavelet cross-periodogram in Proposition \[prop:EIc\] and the result in Lemma \[lem:lipr\], we can rewrite the terms in equation as follows. $$\begin{aligned} {\mathbf{E}}\left[d_{j,k}^{ r+s} d_{j,k}^{ r+s+\eta}\right] &= {\mathbf{E}}\left[I_{j,k;T}^{(r+s,r+s+\eta);R}\right] \\ &= \sum_{l=1}^{J}A_{j,l} S_{l}\left(\frac{k}{T},\frac{r+s}{R},\frac{r+s+\eta}{R}\right) + {\mathcal{O}}(2^j T^{-1}) + {\mathcal{O}}(R^{-1}), \\ &= \sum_{l=1}^{J}\left\{A_{j,l} S_{l}\left(\frac{k}{T},\frac{r}{R},\frac{r+\eta}{R}\right) + {\mathcal{O}}(\|s\| B^\prime_l R^{-1})\right\} + {\mathcal{O}}(2^j T^{-1}) + {\mathcal{O}}(R^{-1}), \\ &= \sum_{l=1}^{J} A_{j,l} S_{l}\left(\frac{k}{T},\frac{r}{R},\frac{r+\eta}{R}\right)+ {\mathcal{O}}(2^j T^{-1}) + {\mathcal{O}}(\|s\| R^{-1}),\end{aligned}$$ where we have also used the Lipschitz constants’ property $\sum_l 2^l B^\prime_l <\infty$ that yields $\sum_l B^\prime_l A_{j,l}<\infty$. Using the same steps, we also have $$\begin{aligned} {\mathbf{E}}\left[d_{j,k}^{ r'+s} d_{j,k}^{ r'+s+\eta}\right] &= {\mathbf{E}}\left[I_{j,k;T}^{(r'+s,r'+s+\eta);R}\right] \\ &= \sum_{l=1}^{J} A_{j,l} S_{l}\left(\frac{k}{T},\frac{r'}{R},\frac{r'+\eta}{R}\right)+ {\mathcal{O}}(2^j T^{-1}) + {\mathcal{O}}(\|s\| R^{-1}).\end{aligned}$$ Similarly, from the expectation of the wavelet cross-periodogram in Proposition \[prop:EIc\] and the cross-spectrum Lipschitz continuity in replicate time, it can be shown that $$\begin{aligned} {\mathbf{E}}\left[d_{j,k}^{ r+s} d_{j,k}^{ r'+s+\uptau}\right] &= {\mathbf{E}}\left[I_{j,k;T}^{(r+s,r'+s+\uptau);R}\right] \\ &= \sum_{l=1}^{J} A_{j,l} S_{l}\left(\frac{k}{T},\frac{r+s}{R},\frac{r'+s+\eta}{R}\right)+ {\mathcal{O}}(2^j T^{-1}) + {\mathcal{O}}(R^{-1}),\\ &= \sum_{l=1}^{J} A_{j,l} S_{l}\left(\frac{k}{T},\frac{r+s}{R},\frac{r+s+\eta}{R}\right)+ {\mathcal{O}}(2^j T^{-1}) + {\mathcal{O}}(\|r-r'\| R^{-1}),\end{aligned}$$ and using the same arguments as above, one obtains $$\begin{aligned} {\mathbf{E}}\left[d_{j,k}^{ r+s} d_{j,k}^{ r'+s+\eta}\right] &= \sum_{l=1}^{J} A_{j,l} S_{l}\left(\frac{k}{T},\frac{r}{R},\frac{r+\eta}{R}\right)+ {\mathcal{O}}(2^j T^{-1}) + {\mathcal{O}}(\|s\| R^{-1}),\end{aligned}$$ under the condition that the replicates $r, r'$ are such that $\|r-r'\|<\infty$. Similarly, $$\begin{aligned} {\mathbf{E}}\left[d_{j,k}^{ r'+s} d_{j,k}^{ r+s+\eta}\right] &= \sum_{l=1}^{J} A_{j,l} S_{l}\left(\frac{k}{T},\frac{r'}{R},\frac{r'+\eta}{R}\right)+ {\mathcal{O}}(2^j T^{-1}) + {\mathcal{O}}(\|s\| R^{-1}).\end{aligned}$$ Thus we can write the covariance in equation as follows $$\begin{aligned} {\mathrm{cov}}&\left(I_{j,k;T}^{(r+s,r'+s);R},I_{j,k;T}^{(r+s+\eta,r'+s+\eta);R}\right) \\ &= {\mathbf{E}}\left[d_{j,k}^{ r+s} d_{j,k}^{ r+s+\eta}\right]{\mathbf{E}}\left[d_{j,k}^{ r'+s} d_{j,k}^{ r'+s+\eta}\right] + {\mathbf{E}}\left[d_{j,k}^{ r+s} d_{j,k}^{ r'+s+\eta}\right]{\mathbf{E}}\left[d_{j,k}^{ r'+s} d_{j,k}^{ r+s+\eta}\right] \\ &= 2\left(\sum_{l=1}^{J}A_{j,l}S_{l}\left(\frac{k}{T},\frac{r}{R},\frac{r+\eta}{R}\right) + {\mathcal{O}}(2^j T^{-1}) + {\mathcal{O}}(\|s\|R^{-1}) \right) \\ & \qquad \times \left( \sum_{l=1}^{J}A_{j,l}S_{l}\left(\frac{k}{T},\frac{r'}{R},\frac{r'+\eta}{R}\right) + {\mathcal{O}}(2^j T^{-1}) + {\mathcal{O}}(\|s\|R^{-1}) \right).\end{aligned}$$ Using this expression in equation , we obtain the variance of the replicate-smoothed wavelet cross-periodogram to be $$\begin{aligned} \text{var}\left({\tilde{I}}_{j,k;T}^{(r,r');R}\right)&= {\mathcal{O}}(M^{-2})\sum_{s=-M}^{M}\sum_{\eta=-M-s}^{M-s} \left[ \sum_{l=1}^{J}A_{j,l}S_{l}\left(\frac{k}{T},\frac{r}{R},\frac{r+\eta}{R}\right) \sum_{l=1}^{J}A_{j,l}S_{l}\left(\frac{k}{T},\frac{r'}{R},\frac{r'+\eta}{R}\right) \right. \\ & \qquad \qquad \qquad \quad \left. + \left( \sum_{l=1}^{J}A_{j,l}S_{l}\left(\frac{k}{T},\frac{r}{R},\frac{r+\eta}{R}\right) {\mathcal{O}}(2^j T^{-1}) \right)\right. \\ & \qquad \qquad \qquad \quad \left. + \left( \sum_{l=1}^{J}A_{j,l}S_{l}\left(\frac{k}{T},\frac{r}{R},\frac{r+\eta}{R}\right) {\mathcal{O}}(\|s\|R^{-1}) \right) \right. \\ & \qquad \qquad \qquad \quad \left. + \left( \sum_{l=1}^{J}A_{j,l}S_{l}\left(\frac{k}{T},\frac{r'}{R},\frac{r'+\eta}{R}\right) {\mathcal{O}}(2^j T^{-1}) \right)\right. \\ & \qquad \qquad \qquad \quad \left. + \left( \sum_{l=1}^{J}A_{j,l}S_{l}\left(\frac{k}{T},\frac{r'}{R},\frac{r'+\eta}{R}\right) {\mathcal{O}}(\|s\|R^{-1}) \right) \right. \\ & \qquad \qquad \qquad \quad \left. + {\mathcal{O}}(\|s\|^2R^{-2})+ {\mathcal{O}}(2^j \|s\| (TR)^{-1})+{\mathcal{O}}(2^{2j} T^{-2}) \vphantom{\sum_{l=1}^{J}S_{l}\left(\frac{k}{T},\frac{r}{R}\right)} \right]. $$ Under the assumption $\sup_{z,\nu,\uptau}\sum_{\eta\in\Z} \left\| c(z,\nu,\nu+\frac{\eta}{R};\uptau) \right\|={\mathcal{O}}(1)$, we obtain for time $k$, replicate $r$ and replicate-lag $\eta$ that $\sum_{\eta\in\Z} \left\| \sum_{l}A_{j,l}S_{l}\left(\frac{k}{T},\frac{r}{R},\frac{r+\eta}{R}\right) \right\| ={\mathcal{O}}(2^j)$, since using the definition of the $A$ matrix and of the local cross-covariance we have $$\begin{aligned} \sum_{\eta\in\Z} \left\| \sum_{l}A_{j,l}S_{l}\left(\frac{k}{T},\frac{r}{R},\frac{r+\eta}{R}\right) \right\| &=\sum_{\eta\in\Z} \left\| \sum_{\uptau\in\Z} \left( \sum_l S_{l}\left(\frac{k}{T},\frac{r}{R},\frac{r+\eta}{R}\right) \Psi_{l}(\uptau) \right) \Psi_{j}(\uptau) \right\|,\\ &= \sum_{\eta} \left\| \sum_{\uptau} c\left(\frac{k}{T},\frac{r}{R},\frac{r+\eta}{R}; \uptau\right)\Psi_{j}(\uptau) \right\|, \\ &\leq \sum_{\uptau} \left( \left\|\Psi_{j}(\uptau)\right\| \sum_{\eta} \left\|c\left(\frac{k}{T},\frac{r}{R},\frac{r+\eta}{R}; \uptau\right) \right\| \right), \\ &={\mathcal{O}}(1) \sum_{\uptau}\left\|\Psi_{j}(\uptau)\right\|= {\mathcal{O}}(2^j),\end{aligned}$$ where we used the triangle inequality and the autocorrelation wavelet property $\sum_{\uptau}\left\|\Psi_{j}(\uptau)\right\|= {\mathcal{O}}(2^j)$. Using Lemma \[lem:ASordc\] and the property above, we readily obtain that the first term in the variance equation is of order ${\mathcal{O}}(2^{2j}M^{-1})$; the second and fourth terms, are both of order ${\mathcal{O}}(2^{2j}(MT)^{-1})$; and the third and fifth terms are both of order ${\mathcal{O}}(2^j R^{-1})$. Now considering the final order terms, $${\mathcal{O}}(M^{-2})\sum_{s=-M}^{M}\sum_{\eta=-M-s}^{M-s} \left[{\mathcal{O}}(\|s\|^2R^{-2})+ {\mathcal{O}}(2^j \|s\| (TR)^{-1})+{\mathcal{O}}(2^{2j} T^{-2})\right],$$ we have $${\mathcal{O}}(M^2R^{-2})+{\mathcal{O}}(2^j M(TR)^{-1})+ {\mathcal{O}}(2^{2j}T^{-2}).$$ Putting these results together in the variance equation, we obtain $$\begin{aligned} \text{var}\left({\tilde{I}}_{j,k;T}^{(r,r');R}\right)&= {\mathcal{O}}(2^{2j}M^{-1})+{\mathcal{O}}(2^{2j}(MT)^{-1})+{\mathcal{O}}(2^j R^{-1})\\ & \qquad + {\mathcal{O}}(M^2R^{-2})+{\mathcal{O}}(2^j M(TR)^{-1})+ {\mathcal{O}}(2^{2j}T^{-2}),\\ &= {\mathcal{O}}(2^{2j}M^{-1})+ {\mathcal{O}}(2^j R^{-1})+ {\mathcal{O}}(M^2R^{-2}).\end{aligned}$$ Proof of Proposition \[prop:rhoc\] {#app:proofs:rhoc} ---------------------------------- It follows directly from the continuous mapping theorem [@billingsley1999convergence] and the consistency results for the corrected replicate-smoothed periodogram in , i.e. $\hat{S}_{j}(z,\nu, \nu') \stackrel{P}{\to} {S}_{j}(z,\nu, \nu')$ as $M, T, R \to \infty$ and $M/R \to 0$. [1]{} \ [Jonathan Embleton$^{1}$, Marina I. Knight$^1$, and Hernando Ombao$^2$\ ]{} [${}^1$Department of Mathematics, University of York, UK\ ${}^2$King Abdullah University of Science and Technology (KAUST), Saudi Arabia\ ]{} [\ ]{} Experimental data description and overview of implemented methodology {#supp:exper} ===================================================================== Each trial (replicate) consists of $T = 2048$ time points, corresponding to approximately 2 seconds of data. The design of the experiment splits each trial into four time blocks of 512 milliseconds each, as follows. For the first block the macaque fixated on a screen; a picture (one of four) was then presented on the screen for the next time block; this was followed by an empty screen for the next interval; for the last 512 milliseconds the macaque was presented with a picture of four doors, one of which associated with the picture visual from the second time block. The macaque’s task was to select the correct door using a joystick. Correct and incorrect choices were signified via a visual cue and a juice reward was given each time a correct selection was made. The macaque had to learn the associations through repeated trials. The data has been grouped into sets of ‘correct’ and ‘incorrect’ responses, in order to investigate how the contributions of the Hc and NAc to the learning process differ between groups [@Gorrombao:monkeydepend]. The plots of the incorrect responses appear below, as well as their corresponding proposed spectral estimates. For both the correct and incorrect sets of the hippocampus (Hc) trial data, we compute the wavelet periodograms using non-decimated discrete wavelets built by means of Daubechies Least Asymmetric wavelet family with 10 vanishing moments. Similarly, for both sets of the nucleus accumbens (NAc) trial data we again choose Daubechies Least Asymmetric wavelet family, but however we now opt for a coarser choice of wavelet with 6 vanishing moments to reflect the behaviour of the signal. In accordance with our simulation study findings, to obtain an asymptotically unbiased and consistent estimator for the replicate evolutionary wavelet spectrum (REWS), we smooth the wavelet periodograms using a local averaging window over 21 replicates ($M = 10$ neighbouring replicates) and then correct for bias. For completeness, we also run the analysis to include a time-smoothing step before locally averaging across replicates, as this was shown to lead to better performance (see the simulation study in Section \[sec:sims\]). For comparison, we additionally report the LSW estimator embedding averaging over all replicates. Note that averaging over all replicates here refers to the averaging over the first 241 and 256 correct and incorrect response trials, respectively. To explore the consistency of the results, the analysis has also been repeated using wavelets with different vanishing moments and varying smoothing windows across the replicates, which yielded extremely similar results to those reported here. Further simulation evidence for Section \[sec:sims\] {#supp:extra} ==================================================== Supporting evidence for the simulation study of Section \[sec:sims\] {#supp:extradetail} -------------------------------------------------------------------- For $R = 256$ replicates and $T=256=2^8$, we generated a RLSW process with spectral structure at scales 5 and 6, defined as $$S_{j}(z,\nu) = \begin{cases} 4(1-\nu)\cos^{2}\left(\pi z \right), & \text{for } j=J(T)-3, z\in(65/256,1), \nu \in(0,1) \\ 4\cos^{2}\left(2\pi z + 5\nu \right), & \text{for } j=J(T)-2, z\in(0,128/256), \nu \in(0,1) \\ 0, & \text{otherwise}. \end{cases} \label{eq:specsim3}$$ The histogram in Figure \[fig:hist\_SIM3proc\] illustrates that the expected (theoretical) asymptotic behaviour of the proposed estimator holds in practice. Further simulation studies {#supp:furthersims} -------------------------- [Simulation 1]{}. We simulate a RLSW process consisting of $R = 128$ replicates, each of length $T=256=2^8$ and whose REWS, illustrated in Figure \[fig:wavplots\_usual\], evolves slowly over both rescaled time and replicates, as follows $$S_{j}(z, \nu) = \begin{cases} 4\nu \sin^{2}\left(2\pi z(1 + 2\nu) \right), & \text{for } j=J(T)-4, z\in(0,1), \nu \in(0,1) \\ 0, & \text{otherwise}, \end{cases} \label{eq:specsim1}$$ recalling that $z = k/T$ and $\nu = r/R$ for $k=0,\ldots,T-1$ and $r=0,\ldots,R-1$. The spectral characteristics thus appear at scale $j = 8-4=4$. The periodicity and magnitude of the sine wave evolve slowly over the replicates in such a way that the spectral characteristics of neighbouring replicates do not look too dissimilar whilst there is a noticeable difference between replicates further apart. One concatenated realisation of the meta-process with the specified spectral structure in , viewed as a series of length $RT$, can be seen in Figure \[fig:reptsBY7\_usualproc\]. We display in Figure \[fig:coef\_usual\_128\_256\] the true spectra and the average spectral estimates for replicates 20, 64 and 108. The non-decimated wavelet transform was computed using discrete wavelets built by means of Daubechies Least Asymmetric family with 10 vanishing moments and the local averaging for our RLSW$_1$ method was carried out using $M=4$, corresponding to a window of $9$ replicates (numerical MSE results in Table \[tab:msesim1\] highlight that we chose to visually present some of our least performant results). Figure \[fig:coef\_usual\_128\_256\] displays the danger of neglecting the possibility of an existing evolutionary behaviour over replicates (see e.g. level $4$ in the top row plots of the true spectrum), conducive to either under or over-estimation (see the middle row plots). The bottom row plots show that the RLSW$_1$ estimates do reflect the evolution over replicates. To further support this, Figure \[fig:lev4\_usual\_128\_256\] takes a closer look at the evolutionary behaviour of the spectral quantities over the time and replicates in level 4. The MSE and squared bias results in Table \[tab:msesim1\], highlight that for this example, which adheres well to the RLSW assumptions, the MSEs for the LSW model are higher than those computed for the RLSW model. The benefit of taking a local smoothing approach over replicates results in spectral estimates with lower bias and MSE, although it is worth pointing out that taking a local smoothing approach over both time and replicates, while yielding lower MSEs, does increase the bias. Just as for the simulation in the main body of the paper, notice the RLSW methodology performance improves with the replicate local averaging window length increase ($2M + 1$). Figure \[fig:hist\_usualproc\] provides a visualisation on how the RLSW model performed over the 100 simulations via histograms of the simulation-specific MSE. The histograms highlight not only how the increase in $M$ improves performance but also how the increase in $R$ and $T$ reduces the MSEs, thus demonstrating the expected asymptotic behaviour of our smoothed estimator $\hat{S}_{j}\left(z,\nu\right)$. [@ccccccccc@]{} & &\ & & &\ R & T & M & mse & bias[2]{} & mse & bias[2]{} & mse & bias[2]{}\ 64 & 128 & 4 & 13.21 & 12.68 & 12.92 & 7.91 & 11.59 & 7.97\ & & 7 & 12.01 & 11.47 & 10.74 & 7.81 & 9.99 & 7.87\ & & 10 & 11.01 & 10.47 & 9.88 & 7.81 & 9.36 & 7.87\ & & 12 & 10.48 & 9.94 & 9.51 & 7.79 & 9.08 & 7.84\ & 256 & 4 & 11.10 & 10.60 & 7.64 & 2.79 & 6.32 & 2.89\ & & 7 & 10.06 & 9.56 & 6.54 & 3.71 & 5.82 & 3.82\ & & 10 & 9.21 & 8.71 & 6.83 & 4.84 & 6.35 & 4.94\ & & 12 & 8.77 & 8.27 & 7.09 & 5.43 & 6.70 & 5.53\ 128 & 128 & 4 & 13.67 & 13.40 & 13.11 & 7.92 & 11.70 & 7.99\ & & 7 & 13.02 & 12.76 & 10.81 & 7.76 & 10.01 & 7.83\ & & 10 & 12.40 & 12.14 & 9.80 & 7.67 & 9.26 & 7.75\ & & 12 & 12.00 & 11.74 & 9.40 & 7.64 & 8.97 & 7.72\ & 256 & 4 & 11.49 & 11.25 & 7.34 & 2.41 & 6.01 & 2.52\ & & 7 & 10.95 & 10.70 & 5.47 & 2.58 & 4.74 & 2.70\ & & 10 & 10.41 & 10.17 & 4.94 & 2.91 & 4.47 & 3.03\ & & 12 & 10.07 & 9.82 & 4.90 & 3.20 & 4.52 & 3.33\ [@ccccccccc@]{} & &\ & & &\ R & T & M & mse & bias[2]{} & mse & bias[2]{} & mse & bias[2]{}\ 256 & 512 & 4 & 10.16 & 10.03 & 5.66 & 0.39 & 4.13 & 0.40\ & & 7 & 9.91 & 9.78 & 3.54 & 0.40 & 2.64 & 0.43\ & & 10 & 9.67 & 9.54 & 2.67 & 0.46 & 2.05 & 0.49\ & & 12 & 9.51 & 9.38 & 2.36 & 0.52 & 1.86 & 0.55\ & 1024 & 4 & 9.11 & 8.98 & 5.28 & 0.12 & 3.77 & 0.11\ & & 7 & 8.88 & 8.75 & 3.17 & 0.12 & 2.28 & 0.12\ & & 10 & 8.66 & 8.53 & 2.30 & 0.15 & 1.68 & 0.15\ & & 12 & 8.52 & 8.39 & 1.99 & 0.19 & 1.47 & 0.20\ 512 & 512 & 4 & 10.25 & 10.18 & 5.72 & 0.37 & 4.17 & 0.39\ & & 7 & 10.12 & 10.05 & 3.55 & 0.36 & 2.64 & 0.38\ & & 10 & 9.99 & 9.93 & 2.62 & 0.36 & 1.99 & 0.39\ & & 12 & 9.91 & 9.84 & 2.26 & 0.37 & 1.73 & 0.39\ & 1024 & 4 & 9.18 & 9.12 & 5.29 & 0.11 & 3.77 & 0.10\ & & 7 & 9.06 & 9.00 & 3.19 & 0.09 & 2.28 & 0.09\ & & 10 & 8.95 & 8.89 & 2.29 & 0.09 & 1.65 & 0.09\ & & 12 & 8.87 & 8.81 & 1.93 & 0.09 & 1.40 & 0.09\ [Simulation 2]{}. We consider a RLSW process consisting of $R = 128$ replicates, each of length $T= 256 = 2^{8}$, and driven by the following REWS $$S_{j}(z,\nu) = \begin{cases} \sin^{2}\left(2\pi z + 10\nu \right), & \text{for } j=J(T)-1, z\in(0,1), \nu \in(0,1) \\ 0, & \text{otherwise}. \end{cases} \label{eq:specsim2}$$ Our spectra in this example are characterised by a squared sine wave in the finest level that experiences a shift from replicate to replicate. This behaviour is displayed in Figure \[fig:wavplots\_shift\] and then a concatenated realisation of the meta-process appears in Figure \[fig:reptsBY7\_shiftproc\]. Note that visually the meta-process behaviour does not offer any indication of transitioning through the replicates, despite this actually happening. We obtain estimates for both the LSW model averaged over all replicates and for the RLSW model which adopts the local averaging procedure (over replicate, and over time and replicate dimensions). On visually examining the concatenated process in Figure \[fig:reptsBY7\_shiftproc\], one may not question the existence of evolutionary behaviour across the replicates. Inspecting our RLSW estimates of the spectral characteristics of this process tells us otherwise. Figure \[fig:coef\_shiftproc\_R128\_T256\] highlights that the RLSW($_1$) method (bottom row) manages to capture the evolution of the spectra in the finest level. On the other hand, the LSW method (middle row) fails to capture this behaviour, and this is further demonstrated in Figure \[fig:lev7\_shiftproc\_R128\_T256\] which shows the spectral estimates across rescaled time and replicate in the finest level.\ Mean squared errors and squared bias results for Simulation 2 are given in Table \[tab:msesim2\]. When comparing models, we notice that for low $R$, performing local averaging over the replicates only (RLSW$_1$) appears to yield low bias estimates that nevertheless have poorer MSE results than the blanket LSW involving local time smoothing and then averaging over all replicates (despite its high bias). The small squared bias for the RLSW estimates implies that much of the MSE can be attributed to the variance. A possible explanation for this is the spectral leakage across neighbouring scales, a known artefact in the locally stationary spectral estimation context: the narrower choices of window to smooth over replicates in the RLSW model were not sufficient enough to remove the effects of the ‘leaked’ characteristics. These simulations provide a stronger highlight of the impact of the ratio $M/R$ when choosing the replicate window width $(2M+1)$, with the guideline of a window width of 15% of $R$ again appearing to yield competitive results. As $R$ and $T$ increase, and our choice of $M$ becomes larger with larger $R$ (increasing the smoothing window), the MSEs improve and RLSW$_2$ estimation performs dramatically better than LSW. This is again in line with our asymptotic results. Also, we note here that while the spectra in Simulation 2 does not behave in quite as slowly evolving manner across replicates as in Definition \[def:rlsw\], nevertheless the RLSW methodology still performs well at estimating the spectra and capturing the evolutionary behaviour across replicates. As with Simulation 1, we also provide histograms for the MSEs over 100 simulations involving $R=128$ replicates of length $T=256$. Figure \[fig:hist\_shiftproc\] highlights how increasing the smoothing window improves the performance of our RLSW method and leads to better estimates. In this setting, Table \[tab:msesim2\] shows how the LSW model performs better in comparison to RLSW($_1$) when $M=4$ but as $M$ increased, the RLSW($_1$) model soon outperformed LSW. This can also be visualised nicely through the histograms in Figure \[fig:hist\_shiftproc\_B\], where we can see the switch in performance of the models and the MSE improvement with the increase of $M$. [@ccccccccc@]{} & &\ & & &\ R & T & M & mse & bias[2]{} & mse & bias[2]{} & mse & bias[2]{}\ 64 & 128 & 4 & 19.37 & 18.09 & 33.91 & 1.97 & 10.69 & 1.85\ & & 7 & 19.49 & 18.21 & 28.04 & 8.85 & 14.10 & 8.79\ & & 10 & 19.39 & 18.11 & 33.34 & 19.57 & 23.24 & 19.44\ & & 12 & 19.20 & 17.92 & 36.49 & 24.87 & 27.88 & 24.69\ & 256 & 4 & 16.87 & 15.81 & 29.63 & 1.66 & 9.10 & 1.47\ & & 7 & 16.97 & 15.91 & 24.48 & 7.71 & 12.17 & 7.60\ & & 10 & 16.89 & 15.83 & 29.11 & 17.13 & 20.29 & 17.03\ & & 12 & 16.74 & 15.68 & 31.86 & 21.79 & 24.43 & 21.69\ 128 & 128 & 4 & 18.55 & 17.94 & 32.12 & 0.51 & 9.10 & 0.37\ & & 7 & 18.65 & 18.04 & 20.06 & 1.09 & 6.32 & 1.06\ & & 10 & 18.74 & 18.13 & 16.51 & 2.97 & 6.75 & 2.99\ & & 12 & 18.78 & 18.17 & 16.50 & 5.12 & 8.31 & 5.15\ & 256 & 4 & 16.23 & 15.69 & 28.27 & 0.39 & 7.85 & 0.21\ & & 7 & 16.32 & 15.77 & 17.61 & 0.90 & 5.39 & 0.80\ & & 10 & 16.39 & 15.84 & 14.48 & 2.54 & 5.77 & 2.48\ & & 12 & 16.43 & 15.88 & 14.46 & 4.42 & 7.14 & 4.38\ [@ccccccccc@]{} & &\ & & &\ R & T & M & mse & bias[2]{} & mse & bias[2]{} & mse & bias[2]{}\ 256 & 512 & 4 & 14.13 & 13.89 & 25.00 & 0.25 & 6.85 & 0.08\ & & 7 & 14.16 & 13.93 & 15.04 & 0.19 & 4.15 & 0.09\ & & 10 & 14.20 & 13.97 & 10.87 & 0.27 & 3.10 & 0.20\ & & 12 & 14.23 & 13.99 & 9.30 & 0.40 & 2.78 & 0.35\ & 1024 & 4 & 12.72 & 12.50 & 22.58 & 0.23 & 6.15 & 0.07\ & & 7 & 12.75 & 12.54 & 13.58 & 0.17 & 3.73 & 0.08\ & & 10 & 12.78 & 12.57 & 9.82 & 0.24 & 2.79 & 0.17\ & & 12 & 12.81 & 12.59 & 8.41 & 0.37 & 2.50 & 0.31\ 512 & 512 & 4 & 13.99 & 13.87 & 25.06 & 0.26 & 6.84 & 0.08\ & & 7 & 14.01 & 13.89 & 15.03 & 0.16 & 4.10 & 0.05\ & & 10 & 14.02 & 13.90 & 10.74 & 0.12 & 2.94 & 0.05\ & & 12 & 14.04 & 13.91 & 9.03 & 0.12 & 2.49 & 0.05\ & 1024 & 4 & 12.59 & 12.48 & 22.58 & 0.23 & 6.14 & 0.06\ & & 7 & 12.60 & 12.50 & 13.55 & 0.14 & 3.69 & 0.04\ & & 10 & 12.62 & 12.51 & 9.69 & 0.11 & 2.64 & 0.04\ & & 12 & 12.63 & 12.52 & 8.15 & 0.10 & 2.23 & 0.04\ Further simulation evidence for Section \[sec:simsext\] {#supp:furthersimscoh} ======================================================= For level $j=5$ and time $k = 1,\ldots, 256$ we define the non-zero replicate coherence matrices of the illustrative example in Section \[sec:simsext\], as follows $$\begin{aligned} \left(\rho_{j}(\frac{k}{T},\frac{r}{R},\frac{r'}{R})\right)_{r,r'} &= \begin{bmatrix} \rho_{5,k}^{1,1} & \cdots & \rho_{5,k}^{1,128} & \rho_{5,k}^{1,129} & \cdots & \rho_{5,k}^{1,R} \\ \vdots & \ddots & \vdots & \vdots & \ddots & \vdots \\ \rho_{5,k}^{128,1} & \cdots & \rho_{5,k}^{128,128} & \vdots & \cdots & \rho_{5,k}^{128,R} \\ \rho_{5,k}^{129,1} & \cdots & \cdots & \rho_{5,k}^{129,129} & \cdots & \rho_{5,k}^{129,R} \\ \vdots & \ddots & \vdots & \vdots & \ddots & \vdots \\ \rho_{5,k}^{R,1} & \cdots & \rho_{5,k}^{R,128} & \rho_{5,k}^{R,129} & \cdots & \rho_{5,k}^{R,R} \end{bmatrix} \\ &= \begin{bmatrix} 1 & 0.99 & \cdots & 0.99 & -0.71 & \cdots & \cdots & -0.71 \\ 0.99 & \ddots & \ddots & \vdots & \vdots & \ddots & \ddots & \vdots \\ \vdots & \ddots & \ddots & 0.99 & \vdots & \ddots & \ddots & \vdots \\ 0.99 & \cdots & 0. 99 & 1 & -0.71 & \cdots & \cdots & -0.71 \\ -0.71 & \cdots & \cdots & -0.71 & 1 & 0.5 & \cdots & 0.5\\ \vdots & \ddots & \ddots & \vdots & 0.5 & \ddots & \ddots & \vdots \\ \vdots & \ddots & \ddots & \vdots & \vdots & \ddots & \ddots & 0.5 \\ -0.71 & \cdots & \cdots & -0.71 & 0.5 & \cdots & 0.5 & 1 \end{bmatrix}\end{aligned}$$ where we have (abusively) let $\rho_{j}(\frac{k}{T}, \frac{r}{R},\frac{r'}{R}) = \rho_{j,k}^{r,r'}$ to ease notation. Further simulation study {#further-simulation-study .unnumbered} ------------------------ We simulate a replicate locally stationary wavelet process with $R=256$ replicates that feature dependence, measured at $T=512=2^{9}$ time points. The within-replicate locally stationary wavelet (auto)spectra are as defined in Simulation 1 of Section \[sec:sims\]. Here we have $J(T)=9$ (in short, $J$) and the spectral characteristics are placed in level $j=J(T)-4=5$. In addition to the autospectral characteristics, we also define the cross-replicate spectral structure by means of defining their (true) coherence at each level $j$ and location $k$. For level $j = 5$, we choose a coherence of $0.7$ between all replicates over the first $256$ locations and zero (no) coherence over the last $256$ locations. All other levels have no coherence between replicates. Visual representations of this dependence structure appear in Figure \[fig:coh\_sim1\] (left panels). For $j=5$ and $k=1, \ldots,256$, the non-zero coherence matrices are defined as follows $$\nonumber \left(\rho_{j}(\frac{k}{T},\frac{r}{R},\frac{r'}{R})\right)_{r,r'} = \begin{bmatrix} \rho_{5,k}^{1,1} & \rho_{5,k}^{1,2} & \cdots & \rho_{5,k}^{1,R} \\ \rho_{5,k}^{2,1} & \ddots & \ddots & \vdots \\ \vdots & \ddots & \ddots & \rho_{5,k}^{R-1,R} \\ \rho_{5,k}^{R,1} & \cdots & \rho_{5,k}^{R,R-1} & \rho_{5,k}^{R,R} \end{bmatrix} = \begin{bmatrix} 1 & 0.7 & \cdots & 0.7 \\ 0.7 & \ddots & \ddots & \vdots \\ \vdots & \ddots & \ddots & 0.7 \\ 0.7 & \cdots & 0.7 & 1 \end{bmatrix}.$$ Using the spectral estimation methodology proposed in Section \[sec:modelext\_est\], we obtain coherence estimates visually represented in Figure \[fig:coh\_sim1\] (right panels) for replicate 50 (top row) and replicate 200 (bottom row). The non-decimated wavelet transform was computed using discrete wavelets built by means of Daubechies least asymmetric family with 10 vanishing moments and a value $M=4$ was chosen for the local replicate-smoothing. It is apparent that the coherence structure is being picked up by the modelling framework in Section \[sec:modelext\] in terms of the locations and the positiveness of the defined true coherence. We do however note that the intensity of the estimated coherence is not quite as strong as the true coherence, just as for the simulation in Section \[sec:simsext\]. The numerical MSE and squared bias results reported in Table \[tab:rhomsesim1\] highlight that our correction procedure that aims to ensure positive spectral estimates has the undesired effect of increasing the MSEs and introducing bias as $R$ and $T$ increase, just as reported in a bivariate coherence estimation framework by [@sand:2010]. [cccccccc]{} R & T & M & & mse[1]{} & bias[2]{}[1]{} & mse[2]{} & bias[2]{}[2]{}\ 128 & 256 & 7 & & 17.29 & 7.41 & 15.16 & 9.66\ & & 12 & & 15.06 & 7.42 & 14.04 & 9.66\ 128 & 512 & 7 & & 17.64 & 8.11 & 16.01 & 10.66\ & & 12 & & 15.29 & 8.01 & 14.73 & 10.52\ 256 & 512 & 7 & & 18.06 & 8.27 & 16.34 & 10.85\ & & 12 & & 15.85 & 8.18 & 15.18 & 10.75\ Further proofs for the material of Appendix \[app:proofs\] {#supp:proofs:sec3} ========================================================== Under the assumptions of Definition \[def:rlsw\], we have for every rescaled time and replicate, $z, \, \nu$ respectively, and scale $j$, $\sum_{l=1}^{\infty}A_{j,l} S_{l}\left(z,\nu\right)={\mathcal{O}}(2^{j})$. $$\begin{aligned} \left\|\sum_l A_{j,l} S_{l}\left(z,\nu\right)\right\| &= \left\|\sum_{l}\sum _{\uptau}\Psi_{j}(\uptau)\Psi_{l}(\uptau) S_{l}\left(z,\nu\right)\right\| \mbox{ from the definition of the matrix }A \\ &= \left\|\sum _{\uptau}\left(\sum_{l} S_{l}\left(z,\nu\right)\Psi_{l}(\uptau)\right)\Psi_{j}(\uptau)\right\|\\ &= \left\|\sum _{\uptau}c\left(z,\nu;\uptau \right)\Psi_{j}(\uptau)\right\| \mbox { from the local autocovariance definition}\\ & \leq \sum _{\uptau}\left\|c\left(z,\nu;\uptau \right)\right\| \left\|\Psi_{j}(\uptau)\right\| \mbox { using the triangle inequality}\\ &= {\mathcal{O}}(2^{j}),\end{aligned}$$ where we used $\left\|c\left(z,\nu;\uptau \right)\right\| <\infty$ for all $\nu, \uptau$ and $\sum _{\uptau}\left\|\Psi_{j}(\uptau)\right\| = {\mathcal{O}}(2^{j})$ [@nvsk:2000]. Under the assumptions of Definition \[def:rlsw\], we have for every rescaled time and replicate, $z, \, \nu$ respectively, lag $\uptau$ and scale $j$, $\sum_{l=1}^{\infty}A_{l,j}^{\uptau} S_{l}\left(z,\nu\right)={\mathcal{O}}(2^{j})$. $$\begin{aligned} \left\|\sum_{l}A_{l,j}^{\uptau} S_{l}\left(z,\nu\right)\right\|&= \left\|\sum_{l}\sum _{n}\Psi_{l}(n) \Psi_{j}(n+\uptau) S_{l}\left(z,\nu\right)\right\| \mbox{ from the definition of the matrix }A^\uptau \\ &= \left\|\sum _{n}\left(\sum_{l} S_{l}\left(z,\nu\right)\Psi_{l}(n)\right)\Psi_{j}(n+\uptau)\right\| \\ &= \left\|\sum _{n}c\left(z,\nu;n \right)\Psi_{j}(n+\uptau)\right\| \mbox { from the local autocovariance definition}\\ & \leq \sum _{n}\left\|c\left(z,\nu;n \right)\right\| \left\|\Psi_{j}(n+\uptau)\right\| \mbox { using the triangle inequality}\\ &= {\mathcal{O}}(2^{j}),\end{aligned}$$ where we used $\left\|c\left(z,\nu;n \right)\right\| <\infty$ for all $\nu, n$ and $\sum _{n}\left\|\Psi_{j}(n+\uptau)\right\| = \sum _{n}\left\|\Psi_{j}(n)\right\|= {\mathcal{O}}(2^{j})$. Under the assumptions of Definition \[def:rlsw\], we have for every rescaled time and replicate, $z, \, \nu$ respectively, lag $\uptau$ and scales $j, j'$, $$\sum_{l=1}^{\infty} \left\|\sum _{n\in \Z} \Psi_{j,j'}(n) \Psi_{l}(n+\uptau) S_{l}\left(z,\nu\right)\right\|={\mathcal{O}}(2^{(j+j')/2}).$$ Let us first show that for any $\uptau$ and scales $j,\, j',\, l$, we have $$\sum _{n\in\Z}\Psi_{j,j'}(n)\Psi_{l}(n+\uptau)= \sum _{u\in\Z}\Psi_{j,l}(u)\Psi_{j',l}(u+\uptau).$$ This can be seen by re-expressing the right hand of the equality above by means of the definition of the cross-correlation wavelets, hence obtaining $$\begin{aligned} \sum _{u}\Psi_{j,l}(u)\Psi_{j',l}(u+\uptau) &= \sum_{u} \left( \sum_{p} \psi_{j,p}\psi_{l,p-u}\right)\left( \sum_{p'} \psi_{j',p'}\psi_{l,p'-u-\uptau}\right),\\ &= \sum_{p}\sum_{p'} \psi_{j,p} \psi_{j',p'} \left( \sum_u \psi_{l,p-u} \psi_{l,p'-u-\uptau}\right),\\ &= \sum_{p}\sum_{p'} \psi_{j,p} \psi_{j',p'} \Psi_l(p-p'+\uptau),\\ &= \sum_{n} \left(\sum_{p'} \psi_{j,n+p'} \psi_{j',p'} \right)\Psi_l(n+\uptau), \mbox{ where }n:=p-p',\\ &= \sum_{n} \Psi_{j,j'}(n)\Psi_l(n+\uptau).\end{aligned}$$ Using the above equality, we use the fact that the spectrum is positive and now take the triangle inequality $$\begin{aligned} \sum_{l} S_{l}\left(z,\nu\right) \left\| \sum _{n}\Psi_{j,l}(n) \Psi_{j',l}(n+\uptau)\right\| & \leq \sum_{l} S_{l}\left(z,\nu\right) \sum _{n} \left\| \Psi_{j,l}(n) \Psi_{j',l}(n+\uptau)\right\|, \\ &\leq \sum_{l} S_{l}\left(z,\nu\right) \left( \sum _{n} \Psi_{j,l}^2(n)\right)^{1/2} \left(\sum _{n} \Psi_{j',l}^2(n+\uptau)\right)^{1/2}, \\ &= \sum_{l} S_{l}\left(z,\nu\right) A_{j,l}^{1/2}A_{j',l}^{1/2}, \mbox{ from the definition of matrix }A\\ &\leq \left( \sum_{l} A_{j,l}S_{l}\left(z,\nu\right) \right) ^{1/2} \left(\sum_{l'} A_{j',l'}S_{l'}\left(z,\nu\right) \right) ^{1/2},\\ &= {\mathcal{O}}(2^{(j+j')/2}),\end{aligned}$$ where for the last two inequalities above we used the Cauchy-Schwarz inequality and for the last equality we used the result in Lemma \[lem:ASord\]. Proof of Proposition \[prop:autocov\] {#supp:proofs:autocov} ====================================== As the RLSW process is defined to have zero-mean, we have ${\mathrm{cov}}\left(X_{\lfloor zT\rfloor;T}^{\lfloor \nu R\rfloor;T}, X_{\lfloor zT \rfloor+\uptau;T}^{\lfloor \nu R\rfloor;T} \right)= {\mathbf{E}}\left(X_{\lfloor zT\rfloor;T}^{\lfloor \nu R\rfloor;T} X_{\lfloor zT \rfloor +\uptau ;T}^{\lfloor \nu R\rfloor;T} \right)$ and using the RLSW process definition in equation , we obtain (X\_[zT;T]{}\^[R;R]{}, X\_[zT +;T]{}\^[R;R]{} )&=& \_[j=1]{}\^\_[k]{} ( \_[j,k;T]{}\^[R;R]{} )\^2 \_[j,k]{}(zT)\_[j,k]{}(zT+), k:=n+zT,\ &=&\_[j=1]{}\^\_[n]{} (\_[j,n+zT;T]{}\^[R;R]{})\^2 \_[j,n+zT]{}(zT)\_[j,n+zT]{}(zT+),\ &=& \_[j=1]{}\^\_[n]{} (\_[j,n+zT;T]{}\^[R;R]{})\^2 \_[j,n]{}(0)\_[j,n]{}(). Denoting $A=\left\| {\mathrm{cov}}\left(X_{\lfloor zT\rfloor;T}^{\lfloor \nu R\rfloor;R}, X_{\lfloor zT \rfloor +\uptau ;T}^{\lfloor \nu R\rfloor;R} \right)-c(z,\nu;\uptau) \right\|$ and using the local autocovariance definition as well as the amplitude approximations in equation , we obtain A =\| \_[j=1]{}\^\_[n]{} (\_[j,n+zT;T]{}\^[R;R]{})\^2 \_[j,n]{}(0)\_[j,n]{}() - c(z,; )\|, a quantity bounded by\ $\left\| \sum_{j}\sum_{n} \left(S_{j}(z+\frac{n}{T},\nu)+ (D_j R^{-1})+(C_j T^{-1})\right) \psi_{j,n}(0)\psi_{j,n}(\uptau)- \sum_{j}S_j(z,\nu)\Psi_j(\uptau)\right\|$. Using the spectrum Lipschitz continuity in time, we further bound $A$ by\ $\left\| \sum_{j}\sum_{n} \left(S_{j}(z,\nu)+ (D_j R^{-1})+(C_j T^{-1})+ (L_{j}\|n\|T^{-1})\right) \psi_{j,n}(0)\psi_{j,n}(\uptau)- \sum_{j}S_j(z,\nu)\Psi_j(\uptau)\right\|$. Hence A && \| \_[j]{}\_[n]{} ((D\_j R\^[-1]{})+(C\_j T\^[-1]{})+ (L\_[j]{}\|n\|T\^[-1]{})) \_[j,n]{}(0)\_[j,n]{}()\| \_[j]{}()=\_[k]{}\_[j,k]{}(0)\_[j,k]{}(),\ && \| \_[j]{}\_[n]{} (D\_j R\^[-1]{}) \_[j,n]{}(0)\_[j,n]{}()\|\ &+& \| \_[j]{}\_[n]{} (C\_j T\^[-1]{} \_[j,n]{}(0)\_[j,n]{}()\|\ &+& \| \_[j]{}\_[n]{} (L\_[j]{}\|n\|T\^[-1]{})\_[j,n]{}(0)\_[j,n]{}()\|,\ &=& (R\^[-1]{})+ (T\^[-1]{}), where the last equality follows as the terms $\sum_{j}\sum_{n} (D_j R^{-1}) \psi_{j,n}(0)\psi_{j,n}(\uptau)=R^{-1} \sum_{j} D_j \Psi_{j}(\uptau)= {\mathcal{O}}(R^{-1})$ since $\Psi_{j}(\uptau)={\mathcal{O}}(1)$ [@nvsk:2000] and we assumed $\sum_{j} D_j <\infty$, and similarly $\sum_{j}\sum_{n} (C_j T^{-1}) \psi_{j,n}(0)\psi_{j,n}(\uptau)={\mathcal{O}}(T^{-1})$ using $\sum_{j} C_j <\infty$; the last term\ $\left\|\sum_{j}\sum_{n} (L_{j}\|n\|T^{-1})\psi_{j,n}(0)\psi_{j,n}(\uptau)\right\|\leq {\mathcal{O}}(T^{-1}) \sum_{j} 2^j L_{j}$, since there are at most order $2^j$ non-zero terms in the wavelet product, $\Psi_{j}(\uptau)={\mathcal{O}}(1)$ and we assumed $\sum_{j} 2^j L_{j}<\infty$ (see also [@nvsk:2000]). Further proofs for the material of Appendix \[app:proofsc\] {#supp:proofs:sec5} =========================================================== Under the assumptions of Definition \[def:rlswc\], we have a sequence $\{B_j\}$ of uniformly bounded Lipschitz constants in $j$ with $\sum_j2^jB_j<\infty$ such that $$\begin{aligned} & \left\|{\widetilde{W}}_{j}\left(\frac{k+n}{T},\frac{r}{R}\right){\widetilde{W}}_{j}\left(\frac{k+n}{T},\frac{r'}{R}\right)\rho_{j}\left(\frac{k+n}{T},\frac{r}{R},\frac{r'}{R}\right) \right. \nonumber \\ & \qquad \qquad \left. - {\widetilde{W}}_{j}\left(\frac{k}{T},\frac{r}{R}\right){\widetilde{W}}_{j}\left(\frac{k}{T},\frac{r'}{R}\right)\rho_{j}\left(\frac{k}{T},\frac{r}{R},\frac{r'}{R}\right)\right\| \leq \|n\| B_{j}T^{-1}, \label{eq:crosslipz}\end{aligned}$$ for any replicates $r$, $r'$ and times $k$, $n$. Note that the above result means that the replicate cross-spectrum, i.e. $S_{j}(\cdotp,\nu,\nu')$, is Lipschitz continuous in the rescaled time argument for any rescaled replicates $\nu$, $\nu'$. Under the assumptions of Definition \[def:rlswc\], the functions ${\widetilde{W}}_{j}(\cdotp,\nu)$, ${\widetilde{W}}_{j}(\cdotp,\nu')$ and $\rho_{j}(\cdotp, \nu, \nu')$ are Lipschitz continuous in rescaled time, with Lipschitz constants $L_{j}^{\nu}$, $L_{j}^{\nu'}$ and $Q_{j}^{\nu,\nu'}$ respectively. Equivalently, this can be written as $\left\|{\widetilde{W}}_{j}((k+n)/T,\nu)- {\widetilde{W}}_{j}(k/T,\nu)\right\| \leq \|n\| L_{j}^{\nu}T^{-1}$,\ $\left\|{\widetilde{W}}_{j}((k+n)/T,\nu')- {\widetilde{W}}_{j}(k/T,\nu')\right\| \leq \|n\| L_{j}^{\nu'}T^{-1}$ and $\left\|\rho_{j}((k+n)/T,\nu,\nu')- \rho_{j}(k/T,\nu,\nu')\right\| \leq \|n\| Q_{j}^{\nu,\nu'}T^{-1}$. From standard Lipschitz function theory, the product of the Lipschitz continuous functions defined on a compact interval is also Lipschitz continuous with Lipschitz constant the maximum of the individual constants. Hence working across replicates and denoting $B_{j}=\sup_{\nu,\nu'}\mbox{max}\{L_{j}^{\nu},L_{j}^{\nu'},Q_{j}^{(\nu,\nu')}\}$, we readily obtain that $S_{j}(\cdotp,\nu,\nu')={\widetilde{W}}_{j}(\cdotp,\nu){\widetilde{W}}_{j}(\cdotp,\nu')\rho_{j}(\cdotp, \nu, \nu')$ is Lipschitz continuous (the result in equation ) with constants $\{B_j\}$. The properties of the $\{B_j\}$ follow immediately from the similar properties of the individual Lipschitz constants $L_j=\sup_{\nu}L_j^\nu$ and $Q_j=\sup_{\nu,\nu'}Q_j^{\nu,\nu'}$. Under the assumptions of Definition \[def:rlswc\], we have a sequence $\{B_j^\prime\}$ of uniformly bounded Lipschitz constants in $j$ with $\sum_j2^jB^\prime_j<\infty$ such that $$\begin{aligned} & \left\|{\widetilde{W}}_{j}\left(\frac{k}{T},\frac{r+s}{R}\right){\widetilde{W}}_{j}\left(\frac{k}{T},\frac{r'+s}{R}\right)\rho_{j}\left(\frac{k}{T},\frac{r+s}{R},\frac{r'+s}{R}\right) \right. \nonumber\\ & \qquad \qquad \left. - {\widetilde{W}}_{j}\left(\frac{k}{T},\frac{r}{R}\right){\widetilde{W}}_{j}\left(\frac{k}{T},\frac{r'}{R}\right)\rho_{j}\left(\frac{k}{T},\frac{r}{R},\frac{r'}{R}\right)\right\| \leq \|s\| B^\prime_{j}R^{-1}, \label{eq:crosslipr}\end{aligned}$$ for any times $k$, $n$ and replicates $r$, $r'$. Note that the above result effectively states that the replicate cross-spectrum, i.e. $S_{j}(z,\nu+\cdotp,\nu'+ \cdotp)$, is Lipschitz continuous in the rescaled replicate arguments. Under the assumptions of Definition \[def:rlswc\], for any rescaled time $z$ and replicates respectively $\nu'$, $\nu$, the limiting coherence functions $\rho_{j}(z, \cdotp, \nu')$ and $\rho_{j}(z, \nu, \cdotp)$ are Lipschitz continuous with Lipschitz constants $\{P_j^z\}$ and are defined on a compact interval. Then it immediately follows that there exists a bounded constant $C$ such that $\left\|\rho_{j}(z, \nu+a, \nu'+a) - \rho_{j}(z, \nu, \nu') \right\| \leq C \, P_{j}^{z} \|a\|$ for any $a$ such that $\nu+a, \nu'+a \in (0,1)$. Specifically, the Lipschitz continuity assumption in rescaled replicate-dimension implies that for each $z$ we have $\left\|{\widetilde{W}}_{j}(z,(r+s)/R)- {\widetilde{W}}_{j}(z,r/R)\right\| \leq \|s\| N_{j}^{z}R^{-1}$, $\left\|{\widetilde{W}}_{j}(z,(r'+s)/R)- {\widetilde{W}}_{j}(z,r'/R)\right\| \leq \|s\| N_{j}^{z}R^{-1}$ and\ $\left\|\rho_{j}(z,(r+s)/R,(r'+s)/R)- \rho_{j}(z,r/R,r'/R)\right\| \leq \|s\| P_{j}^{z}R^{-1}$ at any replicates $r$, $r'$ and $s$. Then, from standard Lipschitz function theory it follows that their product $S_{j}(z, \nu+\cdotp, \nu'+\cdotp)={\widetilde{W}}_{j}(z,\nu+\cdotp){\widetilde{W}}_{j}(z,\nu'+\cdotp)\rho_{j}(z, \nu+\cdotp, \nu'+\cdotp)$ is also Lipschitz continuous (the desired result in equation ) with constants $B^\prime_{j}=\sup_{z}\mbox{max}\{N_{j}^{z},P_{j}^{z}\}$. The properties of the $\{B^\prime_j\}$ follow immediately from the similar properties of the individual Lipschitz constants $N_j=\sup_{z}N_j^z$ and $P_j=\sup_{z}P_j^z$. Proof of Proposition \[prop:autocovc\] {#supp:proofs:autocovc} ====================================== The proof follows immediately from the amplitude and coherence approximation properties in Definition \[def:rlswc\] of the RLSW process, which yield \|\_[j,zT;T]{}\^[R;R]{} \_[j,zT;T]{}\^[’ R;R]{} \_[j,zT;T]{}\^[R, ’ R;R]{}- S\_[j]{}(z,,’)\|=(C\_j\^T\^[-1]{})+ (D\_j \^R\^[-1]{}), where $C_j^\prime$ and $D_j^\prime$ can be shown to also fulfill equations of the type in Definition \[def:rlswc\], \[enum:seq1\] and \[enum:seq2\]. As the limiting amplitudes (${\widetilde{W}}_{j}(\cdotp,\nu)$, ${\widetilde{W}}_{j}(\cdotp,\nu')$) and coherence ($\rho_{j}(\cdotp, \nu, \nu')$) are Lipschitz continuous functions in rescaled time for rescaled replicates $\nu$, $\nu'$ and are defined on a compact interval, it follows from standard Lipschitz function theory that their product ($S_{j}(\cdotp,\nu,\nu')$) is also Lipschitz continuous in rescaled time, as shown in Lemma \[lem:lipz\]. The desired conclusion then follows using the same type of arguments as employed in Section \[supp:proofs:autocov\]. Proof of estimated coherence limits {#supp:proofs:rhoclim} =================================== Let us first show that $\left\|{\tilde{I}}_{l,k;T}^{(r,r');R}\right\| \leq \sqrt{\left\|{\tilde{I}}_{l,k;T}^{(r,r);R}\right\|} \sqrt{\left\|{\tilde{I}}_{l,k;T}^{(r',r');R}\right\|}$ for any scale $l$, time $k$ and replicates $r,\, r'$. Equivalently, we want to show that $\left({\tilde{I}}_{l,k;T}^{(r,r');R}\right)^2 \leq \left({\tilde{I}}_{l,k;T}^{(r,r);R}\right) \left({\tilde{I}}_{l,k;T}^{(r',r');R}\right)$. Recalling that we use the same smoothing window and using the definitions of the smoothed periodogram in and of the raw periodogram in , the above is equivalent to having to show that ( \_s d\_[l,k;T]{}\^[r+s;R]{}d\_[l,k;T]{}\^[r’+s;R]{} )\^2( \_s (d\_[l,k;T]{}\^[r+s;R]{})\^2 ) ( \_[s’]{} (d\_[l,k;T]{}\^[r’+s’;R]{})\^2 ), which indeed follows from the Cauchy-Schwarz inequality. Recalling that the components in the estimated replicate–coherence equation are of the form $\hat{S}_{j,k;T}^{(r,r');R} = \sum_l A^{-1}_{j,l} {\tilde{I}}_{l,k;T}^{(r,r');R}$ (as in equation ), next let us take \| \_l A\^[-1]{}\_[j,l]{} \_[l,k;T]{}\^[(r,r’);R]{} \|\^2 && ( \_l \|A\^[-1]{}\_[j,l]{}\| \|\_[l,k;T]{}\^[(r,r’);R]{}\| )\^2,\ && ( \_l ()\^2 )\^2,\ &=& ( \_l () () )\^2,\ && ( \_l A\^[-1]{}\_[j,l]{} \_[l,k;T]{}\^[(r,r);R]{} ) ( \_[l’]{} A\^[-1]{}\_[j,l’]{} \_[l’,k;T]{}\^[(r’,r’);R]{} ), where the last inequality resulted from the Cauchy-Schwarz inequality. Hence the values of the squared estimated replicate-coherence have the property $\left\| \hat{\rho}_{j,k;T}^{(r,r');R} \right\|\leq 1$. [99]{} [^1]: Corresponding author: [email protected]	{ "pile_set_name": "ArXiv" }
--- abstract: \| We investigate the integer solutions of Diophantine equations related to perfect numbers. These solutions generalize the example, found by Descartes in 1638, of an odd, “spoof” perfect factorization $3^2\cdot 7^2\cdot 11^2\cdot 13^2\cdot 22021^1$. More recently, Voight found the spoof perfect factorization $3^4\cdot 7^2\cdot 11^2\cdot 19^2\cdot(-127)^1$. No other examples appear in the literature. We compute all nontrivial, odd, primitive spoof perfect factorizations with fewer than seven bases—there are twenty-one in total. We show that the structure of odd, spoof perfect factorizations is extremely rich, and there are multiple infinite families of them. This implies that certain approaches to the odd perfect number problem that use only the multiplicative nature of the sum-of-divisors function are unworkable. On the other hand, we prove that there are only finitely many nontrivial, odd, primitive spoof perfect factorizations with a fixed number of bases. author: - BYU Computational Number Theory Group title: 'Odd, spoof perfect factorizations' --- [^1] Introduction ============ Let $\sigma$ denote the sum-of-divisors function. A positive integer $n$ is said to be perfect if $\sigma(n)=2n$; in other words, the sum of the proper divisors of $n$ equals $n$. At present there are fifty-one perfect numbers known, all of the form $2^{p-1}(2^p-1)$ where both $p$ and $2^{p}-1$ are prime. The smallest example is when $p=2$, and the current largest example is when $p=82589933$. Euclid, in his Elements, defined the perfect numbers and proved that if $2^p-1$ is prime (which necessitates $p$ being prime), then $2^{p-1}(2^p-1)$ is perfect. Euler proved, conversely, that every even perfect number is of this form. Two of the oldest open problems in mathematics are whether there are infinitely many perfect numbers, and whether any of them are odd. This paper focuses on the second problem. Many of the methods employed in the study of odd perfect numbers (hereafter denoted OPNs) apply to a much broader class of structures. To motivate this generalization, we consider the (nonprime) factorization $n=3^2\cdot 7^2\cdot 11^2\cdot 13^2\cdot 22021^1$ discovered by Descartes. Recall that $\sigma$ is multiplicative, and for any prime $p$ and any positive integer $a$ we have $\sigma(p^a)=1+p+\cdots+p^a$. Thus, we might (falsely) compute $$\begin{aligned} \sigma(n) & = & \sigma(3^2\cdot 7^2\cdot 11^2\cdot 13^2\cdot 22021^1) \\ & = & (1+3+3^2)(1+7+7^2)(1+11+11^2)(1+13+13^2)(1+22021)=2n.\end{aligned}$$ The problem is that $22021=19^2\cdot 61$, so $22021$ is not prime and the second equality above is false. However, the given factorization of $n$ satisfies the condition for an OPN if we pretend that $22021$ is prime and apply the usual rules for $\sigma$. This motivates the following definitions. Let $n\in {\ensuremath \mathbb{Z}}$. We call an expression of the form $n=\prod_{i=1}^{k}x_i^{a_i}$, where each $x_i\in {\ensuremath \mathbb{Z}}$ and each $a_i\in {\ensuremath \mathbb{Z}}_{\geq 1}$, a factorization of $n$. We call each $x_i$ a base of the factorization, and each $a_i$ is the corresponding exponent of the $i$th base. If, moreover, each $x_i\in {\ensuremath \mathbb{Z}}_{\geq 1}$, then the factorization is positive. A factorization is odd when $n$ is odd; otherwise it is even. We define a function $\tilde{\sigma}$ on the collection of ordered pairs describing such a factorization by the rule $$\tilde{\sigma}\Big(\{(x_i,a_i):1\leq i\leq k\}\Big):=\prod_{i=1}^{k}\left(\sum_{j=0}^{a_i}x_i^{j} \right).$$ As an abuse of notation, we will write $\tilde{\sigma}(\prod_{i=1}^{k}x_i^{a_i})$ instead of $\tilde{\sigma}(\{(x_i,a_i):1\leq i\leq k\})$. We see that $\tilde{\sigma}(\prod_{i=1}^{k}x_i^{a_i})$ agrees with $\sigma(n)$ when the bases are distinct, positive primes. A factorization as above is spoof perfect if $\tilde{\sigma}(\prod_{i=1}^{k}x_i^{a_i})=2\prod_{i=1}^{k}x_i^{a_i}$. (Note that the prime factorizations of actual perfect numbers are being treated as spoof perfect factorizations. We use the word “spoof” rather than something like “generalized” for historical reasons.) Descartes’s example is a positive, odd, spoof perfect factorization. Banks, Güloğlu, Nevans, and Saidak [@Banks] searched for other positive, odd, spoof perfect factorizations of a form similar to Descartes’s example, and Dittmer [@Dittmer] searched for an even more general class of positive, odd, spoof perfect factorizations. Neither was successful in finding any additional examples. Voight in [@Voight] found the odd, spoof perfect factorization $3^4\cdot 7^2\cdot 11^2\cdot 19^2\cdot (-127)^{1}$, which is not positive. Previous to this paper, Voight’s and Descartes’s examples were the only odd, spoof perfect factorizations to have been published. \[SpoofEvenPerfectFirstForm\] For each positive integer $a\in {\ensuremath \mathbb{Z}}_{\geq 1}$, the factorization $2^{a-1}\cdot(2^a-1)^1$ is even and spoof perfect. Dittmer found other infinite families of even, spoof perfect factorizations in [@Dittmer]. These examples suggest a natural dichotomy between the spoof perfect factorizations that are odd and those that are even. We are able to prove that this dichotomy holds quite generally. Our main result appears as Theorem \[MainTheorem\], which says that after restricting ourselves to what we will call nontrivial, primitive spoof perfect factorizations, there are only finitely many odd such factorizations with a fixed number of bases. We find all twenty-one of them that have less than seven bases. These examples not only complement the examples of Descartes and Voight, but they provide natural barriers to certain proof strategies in trying to show that OPNs cannot exist. An outline of the paper follows. In Section \[Section:Trivialities\] we characterize the spoof perfect factorizations with a single base, and also those with $0$ or $-1$ as a base. The latter factorizations are called trivial, and throughout the rest of the paper only nontrivial factorizations are considered. In Section \[Section:AbundDef\] we apply the classical notions of abundant and deficient numbers to spoof perfect factorizations. Using these concepts, we characterize the spoof perfect factorizations with two bases. Next, in Section \[Section:2adic\] we describe two additional tools that have been used in the study of OPNs, and apply them to this broader context of spoof perfect factorizations. With these tools in hand, we give a (possibly complete) list of spoof perfect factorizations with three bases in Section \[Section:3base\]. Finally, in Section \[Section:OddSpoofsLots\], we characterize the odd, primitive spoof perfect factorizations with six or fewer bases, as well as prove our main theorem. Along the way we discover a rich structure inherent in the odd, spoof perfect factorizations, including the fact that there are multiple infinite (nontrivial) families of such factorizations. These infinite families necessarily have an increasing number of bases. We discuss one final class of examples in Section \[Section:EndingQuestions\], where we also pose some additional open problems. Preliminaries and triviality {#Section:Trivialities} ============================ One of the primary goals of this paper is to characterize many of the spoof perfect factorizations, in the hope that this information will be useful in the study of actual perfect numbers. The spoof perfect factorizations with one base are easy to characterize. The only spoof perfect factorization with one base is $1^1$. By inspection, $\tilde{\sigma}(1^1)=2\cdot 1^1$. Suppose we have a general solution $\tilde{\sigma}(x^a)=2x^a$. We then have $$1+x+x^2+\cdots + x^{a-1}=x^a.$$ Clearly $x\neq 0$. If $x\neq \pm 1$, then looking at this equation modulo $x$ we see there are no solutions. When $x=-1$, if $a$ is even the equation becomes $0=1$, while if $a$ is odd the equation becomes $1=-1$. Thus, the only possible solution is when $x=1$, which implies $a=1$. Before moving on to more bases, we first handle some trivialities. Given a factorization $\prod_{i=1}^{k}x_i^{a_i}$, we have $\tilde{\sigma}(\prod_{i=1}^{k}x_i^{a_i})=0$ if and only if at least one of the bases is $-1$ and the corresponding exponent is odd. If $\tilde{\sigma}(x^a)=0$, then $x\neq 1$ and the geometric sum formula yields $$0=\sum_{j=0}^{a}x^j=\frac{x^{a+1}-1}{x-1}.$$ Hence $x^{a+1}=1$, and thus $x=-1$ and $a$ is odd. The converse is easy to verify. \[Cor:Trivial01\] A factorization where one of the bases equals $0$ is spoof perfect if and only if at least one other factor has base $-1$ with an odd exponent. We note that if $a$ is even, then $\tilde{\sigma}((-1)^a)=1$. Thus, if we have a spoof perfect factorization, we can obtain a new spoof perfect factorization by adjoining (or, when possible, removing) a factor whose base is $-1$ and whose exponent is even. In light of Corollary \[Cor:Trivial01\] and this fact, we say that a factorization is trivial if some base is $0$ or $-1$, otherwise it is nontrivial. For the remainder of the paper we will implicity assume all factorizations are nontrivial. Abundant and deficient factorizations {#Section:AbundDef} ===================================== Recall that the function $\sigma_{-1}(n)=\sum_{d\|n}\frac{1}{d}$ is, just like $\sigma$, multiplicative. In analogy to $\tilde{\sigma}$, we define (for nontrivial factorizations) $$\tilde{\sigma}_{-1}\Big(\{(x_i,a_i):1\leq i\leq k\}\Big)=\tilde{\sigma}_{-1}\left(\prod_{i=1}^{k}x_i^{a_i}\right) :=\prod_{i=1}^{k}\left(\sum_{j=0}^{a_i}\frac{1}{x_i^{j}} \right).$$ Note that a factorization $\prod_{i=1}^{k}x_i^{a_i}$ is spoof perfect if and only if $\tilde{\sigma}_{-1}(\prod_{i=1}^{k}x_i^{a_i})=2$. There are a few benefits to working with $\tilde{\sigma}_{-1}$ instead of $\tilde{\sigma}$. First, when $x\in {\ensuremath \mathbb{Z}}\setminus \{-1,0,1\}$ we may set $\tilde{\sigma}_{-1}(x^{\infty}):=\frac{x}{x-1}$ because $$\lim_{a\to\infty}\tilde{\sigma}_{-1}(x^a)=\sum_{j=0}^{\infty}\frac{1}{x^j}=\frac{x}{x-1},$$ thus allowing us to consider what happens to exponents “at infinity.” When $x=1$ we put $\tilde{\sigma}_{-1}(1^{\infty})=\infty$, and we continue to use the formula $x/(x-1)$ by treating $1/0$ as $\infty$. Note that if we were to allow infinite exponents, there would be another spoof perfect factorization with one base, namely $2^{\infty}$. More generally, note that since $\tilde{\sigma}_{-1}(x^{\infty})=\tilde{\sigma}_{-1}((x-1)^1)$ when $x\in{\ensuremath \mathbb{Z}}\setminus\{-1,0,1\}$, we can always replace infinite exponents with finite exponents in factorizations. Thus, in our results characterizing spoof perfect factorizations with a given number of bases we will only list those factorizations involving finite exponents, leaving the reader to figure out the possible factorizations with infinite exponents, if they so desire. Generalizing the literature, let us say that a factorization is deficient if $\tilde{\sigma}_{-1}$ applied to the factorization yields an output that is less than $2$, and abundant when the output is greater than $2$. A quick calculation shows that an abundant factorization remains abundant if we adjoin an additional factor with positive base. More generally, we have the following growth conditions. \[Prop:AbundDefic\] [(1)]{} If $x\in {\ensuremath \mathbb{Z}}_{\geq 1}$ and $a,b\in {\ensuremath \mathbb{Z}}_{\geq 1}\cup \{\infty\}$ with $a<b$, then $$1<\frac{x+1}{x}\leq \tilde{\sigma}_{-1}(x^a)<\tilde{\sigma}_{-1}(x^b)\leq \frac{x}{x-1}.$$ Thus, for a fixed positive base, $\tilde{\sigma}_{-1}$ is strictly increasing as the exponent increases. Moreover, the base determines the interval $\left[\frac{x+1}{x}, \frac{x}{x-1}\right]$ in which $\tilde{\sigma}_{-1}$ takes values. [(2)]{} If $1\leq x<y$ are integers, and $a,b\in {\ensuremath \mathbb{Z}}_{\geq 1}\cup\{\infty\}$, then $$\tilde{\sigma}_{-1}(x^a)\geq \tilde{\sigma}_{-1}(y^b)$$ with equality only if $y=x+1$, $a=1$, and $b=\infty$. Thus, on positive bases the function $\tilde{\sigma}_{-1}$ decreases as bases increase. [(3)]{} If $x\in {\ensuremath \mathbb{Z}}_{<-1}$, then $$\frac{1}{2}\leq \frac{x+1}{x}=\tilde{\sigma}_{-1}(x^1) < \tilde{\sigma}_{-1}(x^3)<\ldots <\frac{x}{x-1} <\ldots <\tilde{\sigma}_{-1}(x^4) < \tilde{\sigma}_{-1}(x^2)<1.$$ Thus, the values of $\tilde{\sigma}_{-1}(x^a)$ oscillate around the limiting value $\tilde{\sigma}_{-1}(x^{\infty})=\frac{x}{x-1}$ when the base is negative. [(4)]{} If $y<x<-1$ are integers, and $a,b\in {\ensuremath \mathbb{Z}}_{\geq 1}\cup \{\infty\}$, then $$\tilde{\sigma}_{-1}(x^a)< \tilde{\sigma}_{-1}(y^b)$$ apart from four cases. For $x\in {\ensuremath \mathbb{Z}}_{<-1}$ and $n\in {\ensuremath \mathbb{Z}}_{\geq 1}$, those four cases are given by $$\begin{array}{rclrcl} \tilde{\sigma}_{-1}(x^{2n}) & > & \tilde{\sigma}_{-1}((x-1)^1), \qquad & \tilde{\sigma}_{-1}(x^{\infty}) & = & \tilde{\sigma}_{-1}((x-1)^1),\\ \tilde{\sigma}_{-1}((-2)^{2}) & > & \tilde{\sigma}_{-1}((-3)^{2n-1}), \qquad & \tilde{\sigma}_{-1}((-2)^2) & = & \tilde{\sigma}_{-1}((-4)^1). \end{array}$$ All of these are straightforward computations, left to the reader. We will now demonstrate the usefulness of these notions by characterizing the perfect factorizations with two bases. \[Prop:TwoBase\] If $x^a\cdot y^b$ is a nontrivial, spoof perfect factorization with $x\leq y$, then it is of the form $2^a\cdot (2^{a+1}-1)^{1}$ or it is one of the two sporadic solutions $(-2)^{1}\cdot 1^3$ and $(-3)^{1}\cdot 1^2$. If $x$ and $y$ are both negative then by part (3) of Proposition \[Prop:AbundDefic\], $\tilde{\sigma}_{-1}(x^a\cdot y^b)<1$, so the factorization isn’t spoof perfect. Next consider the case when $x$ and $y$ are both positive. If $x=1$, then part (1) of Proposition \[Prop:AbundDefic\] yields $$\tilde{\sigma}_{-1}(x^a\cdot y^b)>\tilde{\sigma}_{-1}(1^1)=2$$ so the factorization is abundant. If $x=2$, then $$\tilde{\sigma}_{-1}(y^b)=\frac{2}{\tilde{\sigma}_{-1}(x^a)}=\tilde{\sigma}_{-1}((2^{a+1}-1)^1).$$ By part (2) of Proposition \[Prop:AbundDefic\] we must have $y=2^{a+1}-1$, and then by part (1) we get $b=1$. This gives us the infinite family of solutions given in the statement of the proposition. Next, if $x\geq 3$ and $y\geq 4$, then we compute $$\tilde{\sigma}_{-1}(x^a\cdot y^b)<\tilde{\sigma}_{-1}(3^\infty\cdot 4^{\infty})=\frac{3}{2}\cdot\frac{4}{3}=2,$$ so the factorization is deficient. Thus the only other possible case with both $x$ and $y$ positive is $x=y=3$. If $a=1$ or $b=1$, then the factorization is not perfect due to $2$-adic considerations. (We look at general $p$-adic restrictions in the next section.) Thus $a,b\geq 2$. We then compute $$\tilde{\sigma}_{-1}(x^a\cdot y^b)\geq \tilde{\sigma}_{-1}(3^2\cdot 3^2)=\frac{169}{81}>2,$$ so the factorization is abundant. Finally, consider the case when $x\leq -2$ and $y$ is positive. By part (3) of Proposition \[Prop:AbundDefic\], $\tilde{\sigma}_{-1}(x^a)<1$. If $y\geq 2$, then $\tilde{\sigma}_{-1}(y^b)\leq 2$ and so $\tilde{\sigma}_{-1}(x^a\cdot y^b)<2$, hence the factorization is deficient. So we reduce to the case when $y=1$. If $b=1$, the factorization is deficient. If $b\geq 4$, then we find $$\tilde{\sigma}_{-1}(x^a\cdot y^b)\geq \frac{1}{2}\tilde{\sigma}_{-1}(1^4)=\frac{5}{2}>2,$$ so the factorization is abundant. It is straightforward to check, using part (4) of Proposition \[Prop:AbundDefic\], that the remaining two cases give the sporadic solutions. 2-adic valuation considerations {#Section:2adic} =============================== Given that we want to solve $\tilde{\sigma}_{-1}(\prod_{i=1}^{k}x_i^{a_i})=2$, it is natural to look at the $2$-adic conditions necessary for this equality to hold. These conditions are especially nice in the case where all the bases are odd. \[Proposition:Eulerian\] In any odd, spoof perfect factorization, exactly one of the exponents in the factorization is odd. If $x$ is the unique base with an odd exponent $a$, in such a factorization, then $$x\equiv a\equiv 1\pmod{4}.$$ Let $v_2$ denote the $2$-adic valuation on ${\ensuremath \mathbb{Q}}^{\times}$. If $x$ is any odd base, then $v_2(\tilde{\sigma}_{-1}(x^a))\geq 0$. The valuation is strictly positive if and only if $a$ is odd. As we want the $2$-adic valuation on the entire factorization to equal $1$, we see that exactly one base has an odd exponent. A direct computation through the cases mod $4$ shows that $v_2(\tilde{\sigma}_{-1}(x^a))=1$ if and only if $x\equiv a\equiv 1\pmod{4}$. This proposition restricts the class of potential spoof perfect factorizations, as it rules out many possible exponents on bases. Euler was the first to prove and use this restriction, in the case of actual OPNs. Thus, in any odd, spoof perfect factorization we will call the unique base with an odd exponent the Eulerian base. There is another result in the literature that applies to spoof perfect factorizations as well as actual perfect numbers; however, it applies only for positive, odd factorizations. \[Prop:PaceBound\] If $n=\prod_{i=1}^{k}x_i^{a_i}$ is any positive, odd, spoof perfect factorization, then $n<2^{2^{2k}}$. Follow the argument in [@NielsenBestBound], but now for spoof perfect factorizations. There are finitely many positive, odd, spoof perfect factorizations with a fixed number of bases. While the previous proposition gives an effective bound on the size of any positive, odd, spoof perfect factorization in terms of the number of factors, it is often possible to further restrict cases by applying abundance and deficiency arguments. Using such ideas, Dittmer [@Dittmer] showed that other than the spoof perfect factorization $1^1$ and Descartes’s example, there are no other positive, odd, spoof perfect factorizations with fewer than seven bases. In the remainder of this paper, we explore what happens if we weaken the positivity and parity conditions. Surprisingly, many (but not all) of the finiteness conditions disappear, and a much richer structural pattern emerges. Three bases—more infinite families and complicated behavior {#Section:3base} =========================================================== If we remove the restriction that bases are odd, there are many more spoof perfect factorizations. We list below some spoof perfect factorizations having exactly three bases. (Exactly one of the given factorizations is odd.) We found fifteen sporadic solutions. $$\begin{array}{lll} \bullet\ (-10)^1\cdot (-3)^3\cdot 1^2\qquad & \bullet\ (-5)^1\cdot (-2)^3\cdot 1^3\qquad & \bullet\ (-2)^2\cdot 1^1\cdot 3^1\\ \bullet\ (-9)^1\cdot (-4)^1\cdot 1^2 & \bullet\ (-4)^1\cdot (-3)^1\cdot 1^3 & \bullet\ 3^1\cdot 4^1\cdot 5^1\\ \bullet\ (-7)^1\cdot (-3)^2\cdot 1^2 & \bullet\ (-3)^1\cdot (-2)^1\cdot 1^5 & \bullet\ 3^1\cdot 4^2\cdot 7^1\\ \bullet\ (-6)^1\cdot (-5)^1\cdot 1^2 & \bullet\ (-3)^1\cdot (-2)^2\cdot 1^3 & \bullet\ 3^3\cdot 4^2\cdot 35^1\\ \bullet\ (-5)^1\cdot (-2)^1\cdot 1^4 & \bullet\ (-2)^1\cdot (-2)^1\cdot 1^7 & \bullet\ 3^3\cdot 5^1\cdot 8^1 \end{array}$$ Letting $n\in {\ensuremath \mathbb{Z}}_{>0}$ serve as an index, there are also six infinite families. $$\begin{array}{ll} \bullet\ (-2^{2n+2}+2^{n+2}-2)^1\cdot 2^n\cdot (2^{n+1}-1)^3 \qquad & \bullet\ (-2)^{2n-1}\cdot 1^2\cdot (2^{2n}-1)^1\\ \bullet\ (-2^{2n+2}+2^{n+1}-1)^1\cdot 2^n\cdot (2^{n+1}-1)^2 & \bullet\ (-n-1)^1\cdot 1^1\cdot n^1\\ \bullet\ (-2^{2n+1}-1)^1\cdot(-2)^{2n}\cdot 1^2 & \bullet\ 3^1\cdot 3^n\cdot (3^{n+1}-1)^1 \end{array}$$ Further, letting $m\in {\ensuremath \mathbb{Z}}_{>0}$ serve as a second index, there are three doubly-indexed infinite families. $$\begin{array}{l} \bullet\ 2^n\cdot (2^{n+1})^{m}\cdot (2^{(n+1)(m+1)}-1)^1\\ \bullet\ \left(-(2^{n+1}-1)\left(\frac{2^{n+1}}{m}-1\right)\right)^1\cdot 2^n\cdot (2^{n+1}-1-m)^1, \text{ where }\, m\|(2^{2n+2}-2^{n+1})\, \text{ and }\,m<2^{n+1}-2\\ \bullet\ 2^n\cdot(2^{n+1}+m)^1\cdot \left((2^{n+1}-1)\left(\frac{2^{n+1}}{m+1}+1\right)\right)^1,\text{ where }\, m\|(2^{2n+2}-2^{n+1})\,\text{ and }\,m\leq 2^{n+1}-2 \end{array}$$ These examples can be found employing abundance and deficiency computations and a case-by-case analysis. Many cases can be eliminated by considering $p$-adic information, applying the following two useful results. In the following, $\Phi_n$ denotes the $n$th cyclotomic polynomial, $v_q(x)$ denotes the $q$-adic valuation of $x$ for an arbitrary prime $q$, and $o_q(x)$ denotes the order of $x$ modulo $q$. Let $q$ be a prime and $n\geq 1$ an integer. Setting $k:=v_q(n)$, write $n=q^{k}m$. The equation $$\Phi_n(x)\equiv 0\pmod q$$ is solvable for some $x\in {\ensuremath \mathbb{Z}}$ if and only if $q\equiv 1 \pmod{m}$, and the solutions are exactly those integers $x$ with $o_q(x)=m$. Moreover, if $x$ is such a solution, then $$v_q(\Phi_n(x))=\begin{cases} v_q(x^n-1) & \text{ if }k=0,\\ 1 & \text{ if } k\geq 2, \text{ or } k=1 \text{ and } n>2,\\ v_2(x+1) & \text{ if }k=1 \text{ and }n=2 \text{ \textup{(}so $q=2$\textup{)}.} \end{cases}$$ \[Cor:CyclotomicAdic\] If $q\geq 3$ is a prime, $x\in {\ensuremath \mathbb{Z}}\setminus \{0\}$, and $a\in {\ensuremath \mathbb{Z}}_{>0}$, then $$v_q(\tilde{\sigma}_{-1}(x^a)) = \begin{cases} -a v_q(x) & \text{ if }q\|x,\\ v_q(x^{o_q(x)}-1)+v_q(a+1) & \text{ if }o_q(x)\|(a+1) \text{ and }o_q(x)\neq 1,\\ v_q(a+1) & \text{ if }o_q(x)=1,\\ 0 & \text{ otherwise}. \end{cases}$$ A similar statement holds when $q=2$. Write $$\tilde{\sigma}_{-1}(x^a)=\frac{x^{a+1}-1}{x^a(x-1)}=\frac{\prod_{n\|(a+1),\, n>1}\Phi_n(x)}{x^a}$$ and use the previous proposition. Here is one example of how these results can eliminate candidate spoof perfect factorizations. Consider a (supposed) spoof perfect factorization of the form $(-3)^a\cdot (-2)^b\cdot 1^3$ with $a\geq 2$ and $b\geq 4$. Since it is spoof perfect we must have $$v_3(\tilde{\sigma}_{-1}((-2)^b))=-v_3(\tilde{\sigma}_{-1}((-3)^a\cdot 1^3))=a\geq 2.$$ Corollary \[Cor:CyclotomicAdic\] then implies that $3^2\|(b+1)$. Now, since $o_{19}(-2)=9$ and $9\|b+1$, by the same corollary this forces $v_{19}(\tilde{\sigma}_{-1}((-2)^b))\geq 1$, so the $19$-adic valuation of the factorization is positive, yielding the needed contradiction. We conjecture that our twenty-four bullet points above provide a complete list of the nontrivial, spoof perfect factorizations with three bases. However, we are currently unable to rule out two additional cases. First, writing the factorization as $x^a\cdot y^b\cdot z^c$ with $x\leq y\leq z$, we can’t rule out the possibility that $x$ is negative, $y=2<z<2^{b+1}-1$, and one of $a$ or $c$ is not $1$. The other case is similar, and occurs when $x=2$, $2^{a+1}<y\leq z$, and one of $b$ or $c$ is not $1$. Infinitely many nontrivial, odd, spoof perfect factorizations {#Section:OddSpoofsLots} ============================================================= The three nontrivial, odd, spoof perfect factorizations with three or fewer bases are as follows. $$\begin{array}{l} 1^1\\ 1^2\cdot (-3)^1\\ 1^2\cdot (-3)^2\cdot (-7)^1. \end{array}$$ There is a pattern to these factorizations that continues indefinitely. We can increase the exponent on the Eulerian base by one and then adjoin a new negative base, according to the formula $$\label{Eq:ImportantTrick} \arraycolsep=1.4pt\def\arraystretch{2.8} \begin{array}{rcl} \tilde{\sigma}_{-1}(x^a) & = & \displaystyle{ \frac{x^a+x^{a-1}+\cdots+1}{x^a}= \frac{x^{a+1}+x^a+\cdots+1}{x^{a+1}}\cdot \frac{x^{a+1}+x^{a}+\cdots+x}{x^{a+1}+x^{a}+\cdots+1}}\\ & = & \displaystyle{\tilde{\sigma}_{-1}\left(x^{a+1}\cdot \left(-\sum_{j=0}^{a+1}x^j\right)^1\right)}. \end{array}$$ Notice that the new base $-\sum_{j=0}^{a+1}x^j$ is odd exactly when $a$ is odd, so this process gives a new odd, spoof perfect factorization that changes the Eulerian base. Thus, the three factorizations above are part of an infinite family that continues $$\begin{array}{l} 1^2\cdot (-3)^2\cdot (-7)^2\cdot (-43)^1 \\ 1^2\cdot (-3)^2\cdot (-7)^2\cdot (-43)^2\cdot (-1807)^1 \end{array}$$ and so forth. The two spoof perfect factorizations found by Descartes and Voight, respectively, do not lie in the infinite family generated by applying this process to $1^1$, but each generates its own infinite family of odd, spoof perfect factorizations. The “base expansion trick” encapsulated in (\[Eq:ImportantTrick\]) can be applied to any of the factors in an even spoof perfect factorization to give another such factorization. Thus, in particular, each even infinite family with three bases gives rise to three new infinite families with four bases. There are other (nontrivial) ways to get new spoof perfect factorizations from old ones. Noting that $\tilde{\sigma}_{-1}((-3)^2\cdot 7^2\cdot 7^2\cdot (-19)^2)=1$, we can adjoin this product to any given spoof perfect factorization and obtain a new spoof perfect factorization with exactly four more bases. To avoid this type of extension, we say that a spoof perfect factorization $\prod_{i=1}^{k}x_i^{a_i}$ is primitive if for each proper subset $S\subsetneq \{1,2,\ldots, k\}$ the factorization $\prod_{i\in S}x_i^{a_i}$ is not spoof perfect. Requiring primitivity, and bounding the number of bases, strongly limits the number of spoof perfect factorizations, at least among odd factorizations. \[MainTheorem\] For each integer $k\geq 1$, there are finitely many nontrivial, odd, primitive spoof perfect factorizations with $k$ bases. We follow the ideas of [@Dittmer] and describe a finite process to construct all possible odd, primitive spoof perfect factorizations with at most $k$ bases. Hereafter, $k$ is fixed and all spoof perfect factorizations we will consider have $\leq k$ bases and are odd. First, define a partial factorization to be an ordered set of triples of the form $(x_i,b_i,c_i)$ where: - $x_i\neq -1$ is an odd integer, which we think of as one of the bases in the factorization, - $b_i$ is a positive integer, which we think of as a lower bound on the exponent of the base $x_i$, and - $c_i$ equals either $b_i$ or $\infty$; in either case we think of $c_i$ as an upper bound on the exponent. Thus, for instance, the set $\{(3,2,\infty),(5,4,4),(-3,8,\infty)\}$ is a partial factorization which tells us that one of the bases is $3$ and the corresponding exponent is at least $2$, another base is $5$ and the corresponding exponent is exactly $4$, and a third base is $-3$ with corresponding exponent at least $8$. Our ultimate goal is to recursively build up all complete factorizations from partial factorizations, for odd, primitive spoof perfect factorizations. Next, we define a strict partial ordering on partial factorizations, which will give us a notion of improving a partial factorization. We put $\{(x_i,b_i,c_i)\}_{i=1}^{m}< \{(y_i,d_i,e_i)\}_{i=1}^{n}$ exactly when the following three conditions hold: - We have $m\leq n$. This guarantees that the factorization cannot become shorter. - For each $i\leq m$, we have $x_i=y_i$, $b_i\leq d_i$, and $c_i\geq e_i$. Thus, the bases we have already chosen do not change, and the ranges on the previously chosen exponents can only tighten up. - If $m=n$, then there exists some $i\leq m$ with $c_i=\infty$ and $d_i\neq \infty$. This is to prevent the possibility of endlessly tightening the range of an exponent, such as $$\{(3,2,\infty)\}\to \{(3,4,\infty)\}\to\{3,6,\infty)\}\to \cdots.$$ When improving a partial factorization we see that the new partial factorization always either contains an additional base, or the range on one of the exponents that was previously infinite will now be limited to a single positive integer. Thus, any chain of improvements will have length at most $2k$, as we can only add at most $k$ bases, and can fix each exponent exactly once. Hence, we just need a way to limit choices on exponents and limit choices of bases to be adjoined. Let $S=\{(x_i,b_i,c_i)\}_{i=1}^{m}$ be a partial factorization. If $P=\prod_{i=1}^k x_i^{a_i}$ is any odd, primitive spoof perfect factorization that is compatible with $S$ (meaning $S\leq \{(x_i,a_i,a_i)\}_{i=1}^{k}$, and in particular $m\leq k$), then Proposition \[Prop:AbundDefic\] gives upper and lower bounds on $\tilde{\sigma}_{-1}(\prod_{i=1}^{m}x_i^{a_i})$ using only the information in $S$. Namely, the lower bound is $$L(S):=\tilde{\sigma}_{-1}\left(\prod_{i=1}^{m}x_i^{b_i'}\right)$$ where $$b_i'=\begin{cases} b_i & \text{if $x_i>0$, or $b_i$ is odd, or $b_i=c_i$},\\ b_{i}+1 & \text{otherwise}. \end{cases}$$ The upper bound $U(S)$ is defined similarly. There are three cases we need to consider. [Case 1:]{} $L(S)>2$. In this case we say that $S$ is abundant, since any factorization compatible with $S$ with the same number of factors is abundant. To correct this defect, at least one of the remaining bases in $P$ must be negative. Let $y$ be the largest (under the usual ordering on ${\ensuremath \mathbb{Z}}$) remaining negative base in $P$. By Proposition \[Prop:AbundDefic\], parts (3) and (4), we have $$2=\tilde{\sigma}_{-1}(P)\geq L(S)\prod_{i=m+1}^{k}\tilde{\sigma}_{-1}(y^1)=L(S)\left(1+\frac{1}{y}\right)^{k-m}$$ and solving the inequality for $y$ we reach $$\label{Eq:Limit1} 0>y\geq \frac{1}{\left(\frac{2}{L(S)}\right)^{1/(k-m)}-1}.$$ In particular, we see that the next base in an extension of $S$ is forced to belong to a finite interval. Putting it another way, if we extend $S$ to a new partial factorization with exactly one new base, we may do so in only a finite number of ways, as limited by (\[Eq:Limit1\]). [Case 2:]{} $U(S)<2$. In this case we say that the partial factorization is deficient. To correct this defect, at least one of the remaining bases in $P$ must be positive. Let $y$ be the smallest remaining positive base in $P$. Applying Proposition \[Prop:AbundDefic\], parts (1) and (2), we have $$2=\tilde{\sigma}_{-1}(P)\leq U(S)\prod_{i=m+1}^k \tilde{\sigma}_{-1}(y^{\infty})=U(S)\left(\frac{y}{y-1}\right)^{k-m}$$ and solving for $y$ we reach $$0<y\leq \frac{1}{1-\left(\frac{U(S)}{2}\right)^{1/(k-m)}}.$$ Once again, this limits the next base to a finite interval. [Case 3:]{} $L(S)\leq 2\leq U(S)$. If $L(S)=U(S)=2$, we have a spoof perfect factorization, and any further extension will not be primitive. Thus, we may reduce to considering the case when $L(S)<U(S)$, and in particular at least one of the elements of $S$ has $\infty$ as a third coordinate (else, from the definitions we gave for $L(S)$ and $U(S)$, their values would match). Let $S_b$ be the partial factorization obtained from $S$ by replacing any triple $(x_i,b_i,c_i)$ where $c_i=\infty$ with the new triple $(x_i,b,b)$. Note that $S<S_b$ as long as $$b\geq \max_{\{1\leq i\leq m\, :\, c_i=\infty\}}b_i.$$ Further, $$\lim_{b\to \infty}L(S_b)=\lim_{b\to \infty}U(S_b);$$ call this common value $n'$. We now appeal to Proposition \[Proposition:Eulerian\], and note that $\tilde{\sigma}_{-1}(x^{\infty})=\frac{x}{x-1}$ has negative $2$-adic valuation for any odd $x$. This implies that $n'\neq 2$. Thus, there exists some (computable) positive integer $b$ such that $S<S_b$ and either $L(S_b)>2$ or $U(S_b)<2$. Now, if $P$ is any odd, primitive spoof perfect factorization compatible with $S$, then one of two situations holds. Either - $P$ is also compatible with $S'$, where $S'$ is obtained from $S$ by replacing one of the triples $(x_i,b_i,\infty)$ with $(x_i,b_i',b_i')$, for one of the finitely many integers $b_i'\in [b_i,b]$, or - $P$ is compatible with $S_b$, so we may reduce to either Case 1 or Case 2 (after replacing $S$ with $S_b$) and adjoin one of finitely many new bases. In every case, there are only finitely many improvements. Further, as mentioned previously, the longest possible chain of improvements is $2k$, so there are only finitely many partial factorizations compatible with odd, spoof perfect factorizations. The process described in the proof of Theorem \[MainTheorem\] can be turned into a program for finding all odd, spoof perfect factorizations with a given number of bases. We implemented such a program in <span style="font-variant:small-caps;">Mathematica</span>, also including a subroutine incorporating the $2$-adic valuation conditions expressed in Proposition \[Proposition:Eulerian\]. The $k=5$ case terminated after about 30 minutes, yielding a complete list of the corresponding odd, primitive spoof perfect factorizations. The $k=6$ case took considerably longer. Over the space of two years we distributed the computation over multiple processes (ranging from between 20 to 80 processors at any given time), totaling over 30 processor years. The corresponding computation for positive factorizations, as done in [@Dittmer], takes around 13.3 hours on a single processor. A large majority of the extra time used in our computation was spent considering the cases where the first three bases are $3$, $5$, and $15$. It may be possible to eliminate some of these cases without the need for a brute-force search. We found a total of twenty-one nontrivial, odd, primitive spoof perfect factorizations with six or fewer bases. Ten of these factorizations are listed below. The other eleven can be constructed from those ten by (repeatedly) applying the formula (\[Eq:ImportantTrick\]) to Eulerian factors, thus increasing the number of bases by one. - $1^1$ - $1^2\cdot (-3)^2\cdot (-5)^2\cdot 49^1$ - $1^2\cdot (-3)^2\cdot (-3)^2\cdot 7^2\cdot (-19)^1$ - $3^2\cdot 7^2\cdot 7^2\cdot 13^1\cdot (-19)^2$ - $3^2\cdot 7^2\cdot 11^2\cdot 13^2\cdot 22021^1$ - $3^4\cdot 7^2\cdot 11^2\cdot 19^2\cdot (-127)^1$ - $1^2\cdot (-3)^2\cdot (-3)^2\cdot 7^4\cdot (-17)^2\cdot 36413^1$ - $1^2\cdot (-3)^2\cdot (-5)^2\cdot 7^2\cdot (-7)^2\cdot (-2451)^1$ - $3^4\cdot 7^2\cdot 7^2\cdot (-19)^1\cdot 11^2\cdot (-19)^2$ - $3^4\cdot 7^2\cdot 7^2\cdot (-19)^2\cdot 25^2\cdot (-3751)^1$ While each of these factorizations is primitive, and none arises from any other by repeated use of the base expansion trick in (\[Eq:ImportantTrick\]), it is still sometimes possible to generate one from another. To see this, first apply the base expansion trick to the Eulerian factor of $1^2\cdot (-3)^2\cdot (-5)^2\cdot 49^1$ to get the new “derived” spoof perfect factorization $1^2\cdot (-3)^2\cdot (-5)^2\cdot 49^2\cdot (-2451)^1$. For any base $x\neq 0$, it happens that $$\label{Eq:SquareSquare} \tilde{\sigma}_{-1}((x^2)^2)=\tilde{\sigma}_{-1}(x^2\cdot (-x)^2),$$ so we can replace $49^2$ with $7^2\cdot (-7)^2$, and thus the eighth listed spoof perfect factorization arises in a natural way from the second. We can us the same squared-square trick given in (\[Eq:SquareSquare\]) to get another primitive spoof perfect factorization as follows. Starting with $1^1$ as a seed, applying (\[Eq:ImportantTrick\]) to the Eulerian factor three times in succession we have the spoof perfect factorization $1^2\cdot (-3)^2\cdot (-7)^2\cdot (-43)^1$. Multiplying by $(-3)^2\cdot 7^2\cdot 7^2\cdot (-19)^2$, and then replacing $(-7)^2\cdot 7^2$ with $49^2$, we have the seven-base, primitive spoof perfect factorization $$1^2\cdot (-3)^2\cdot (-3)^2\cdot 7^2\cdot (-19)^2\cdot (-43)^1\cdot 49^2.$$ Doing a partial search with $k=7$, our program found an additional seven-base, odd, spoof perfect factorization: $$1^2\cdot (-5)^2\cdot (-5)^2\cdot (-9)^2\cdot 7^2\cdot (-9)^2\cdot (-101251)^1.$$ More infinite families of odd, spoof perfect factorizations {#Section:EndingQuestions} =========================================================== There are three main properties that can prevent an odd, spoof perfect factorization from corresponding to an actual OPN. They are: - The bases are allowed to be non-primes. - The bases are allowed to share prime factors. - The bases are allowed to be negative (and possibly zero if one allows trivial factorizations). Thus, if any result on OPNs similarly does not use primality, relative primality, or positivity of the bases, then that result necessarily applies to the corresponding spoof perfect factorizations as well. Consequently, Voight’s spoof perfect factorization, which has prime bases that are pairwise relatively prime, shows us that any purported proof of the nonexistence of OPNs must necessarily use the positivity of the bases in the factorization. Similarly, Descartes’s example shows that one must necessarily use the primality of the bases. We were unable to find any example with only positive, prime bases (allowing for repetitions of bases). This work raises the question of whether or not there is additional structure inherent in all spoof perfect factorizations, which in turn would limit the structure of OPNs. For instance, for each of the spoof perfect factorizations listed above, the Eulerian exponent is always $1$. Is this true in general? It turns out that, no, the Eulerian exponent can be arbitrarily large for odd, spoof perfect factorizations. For instance, consider the following 74-base factorization, where we use $[x^a]^b$ to mean that the factor $x^a$ is being repeated $b$ times: $$[(-619)^2]^4\cdot [(-31)^2]^7\cdot [(-19)^2]^2\cdot [(-11)^2]^6\cdot [(-7)^4]^{14}\cdot [7^2]^8\cdot 11^2\cdot [37^2]^6 \cdot [67^2]^5\cdot [163^2]^4 \cdot [191^2]^7\cdot [211^2]^2\cdot [2223^2]^8.$$ If $P$ is this factorization, then $\tilde{\sigma}_{-1}(P)=1/3$. Therefore, in the odd, spoof perfect factorization $1^5\cdot P$, the Eulerian exponent is $5$. The odd, spoof perfect factorization $1^{17}\cdot[P]^2$, where we have repeated the factors in $P$ twice, has an even larger Eulerian exponent. Repeating this process, we can make the Eulerian exponent arbitrarily large. Other questions that we are currently unable to answer include: - Are there only finitely many nontrivial, odd, (not necessarily primitive) spoof perfect factorizations with a given number of bases? If so, can Proposition \[Prop:PaceBound\] be modified to give an effective upper bound on the number of such factorizations? - Are there infinitely many nontrivial, odd, primitive spoof perfect factorizations that do not use $1$ as a base, and that do not arise from simpler factorizations using (\[Eq:ImportantTrick\])? If so, can we additionally guarantee that the bases are pairwise relatively prime? Acknowledgements {#acknowledgements .unnumbered} ================ The project was sponsored by the National Security Agency under Grant Number H98230-16-1-0048. This work was partially supported by a grant from the Simons Foundation (\#281876). [1]{} William D. Banks, Ahmet M. Güloğlu, C. Wesley Nevans, and Filip Saidak, Descartes numbers, Anatomy of integers, CRM Proc. Lecture Notes, vol. 46, Amer. Math. Soc., Providence, RI, 2008, pp. 167–173. [MR ]{}[2437973]{} Samuel J. Dittmer, Spoof odd perfect numbers, Math. Comp. 83 (2014), no. 289, 2575–2582. [MR ]{}[3223347]{} Trygve Nagell, Introduction to number theory, Second edition, Chelsea Publishing Co., New York, 1964. [MR ]{}[0174513]{} Pace P. Nielsen, Odd perfect numbers, [D]{}iophantine equations, and upper bounds, Math. Comp. 84 (2015), no. 295, 2549–2567. [MR ]{}[3356038]{} John Voight, On the nonexistence of odd perfect numbers, M[ASS]{} selecta, Amer. Math. Soc., Providence, RI, 2003, pp. 293–300. [MR ]{}[2027187]{} Here is a full list of the authors in the BYU Computational Number Theory Group for “Odd, spoof perfect factorizations”: - Nickolas Andersen - Spencer Durham - Michael J. Griffin - Jonathan Hales - Paul Jenkins - Ryan Keck - Hankun Ko - Grant Molnar - Kyle Niendorf - Eric Moss - Pace P. Nielsen - Vandy Tombs - Merrill Warnick - Dongsheng Wu Correspondence should be directed to Pace P. Nielsen.\ Department of Mathematics, Brigham Young University, Provo, UT 84602, USA\ Email address: [email protected] [^1]: A full list of authors may be found at: <https://math.byu.edu/~pace/spoof_authors.pdf>	{ "pile_set_name": "ArXiv" }
--- abstract: 'We have developed a compact partially ferromagnetic electromagnet to produce a Ioffe-Pritchard trap for neutral atoms. Our structure permits strong magnetic confinement with low power consumption. Compared to the previous iron-core electromagnet [@old; @electromagnet], it allows for easy compensation of remnant fields and very high stability, along with cost-effective realization and compactness. We describe and characterize our apparatus and demonstrate trapping and cooling of $^{87}$Rb atoms to quantum degeneracy. Pure Bose-Einstein condensates containing $\mathrm{10^6}$ atoms are routinely realized on a half-minute cycle. In addition we test the stability of the magnetic trap by producing atom lasers.' author: - \| M. Fauquembergue, J-F. Riou, W. Guerin, S. Rangwala[^1], F. Moron, A. Villing, Y. Le Coq, P. Bouyer and\ A. Aspect - 'M. Lécrivain' title: Partially ferromagnetic electromagnet for trapping and cooling neutral atoms to quantum degeneracy --- Introduction ============ The realization of Bose-Einstein condensates (BECs) of a dilute gas of trapped atoms [@firstBEC] has opened the way to great improvements in the domain of atom optics and atom interferometry. The macroscopic population of atoms in a single quantum state is the matter-wave analog to a laser in optics [@Mewes97; @Anderson98; @Hagley99; @Bloch99] for which the generalization of its use has revolutionized interferometry technology [@Cho85; @Ste95; @Stedman97]. It is therefore expected that the use of Bose-Einstein condensed atoms will advance the field of atom optics and non linear atom optics [@nonlinearatopt]. In particular, BECs will bring about an unprecedented level of accuracy in atom interferometry [@Bouyer97; @Gupta02; @andrews]. In the several years since their first observation, BECs have been studied extensively [@ReviewBEC]. In most cases the condensate properties themselves are the focus of the investigations. Relatively little work has been done so far using a condensate as a tool to explore questions in other fields [@Gupta02; @hmlecoq; @casimircornell]. Therefore, developing a system that could be used for these purposes is worthwhile, but needs a special focus in the conception of the key elements. One of the main challenges when designing such a BEC apparatus is to produce in a stable and reproducible way Bose-Einstein condensates with a large number of atoms. For that, one needs to optically collect many atoms and have a long lifetime for the atoms in the trap where the atoms will be cooled down to a few hundred nanokelvins. To reach these ultra low temperatures and enter the quantum degeneracy regime, evaporative cooling of trapped atoms [@Hess] remains at this time the most efficient method. This technique relies on a sufficiently high binary collision rate in order to allow fast thermalization of the external degrees of freedom of the atoms [@Walraven]. Thus, besides requiring an ultrahigh vacuum and a sophisticated timing system, the major challenge lies in designing an adequate and easy-to-use non-dissipative trap with high confinement, either using magnetic or optical fields, which forms the core element of each BEC experiment. Potential applications, such as inertial sensors in space [@hmlecoq], mandate the highest gradients for the lowest power. Several groups around the world have unveiled condensate-producing technologies that may in the future prove to fullfill the requirements. One such system involves “macroscopic" magnets, either permanent or electrically controlled [@lewandowski; @QUIC; @old; @electromagnet; @hulet]. A magnetic configuration commonly used is the Ioffe-Pritchard trap configuration [@IoffePritchard] because it avoids trap losses due to nonadiabatic spin flips: a quadrupolar linear trap providing the radial confinement is combined with a dipolar trap providing the axial confinement. To increase the density of trapped atoms and ensure a high collision rate, tight confinement is necessary. To accomplish those requirements, Ioffe traps typically dissipate kilowatts of power necessitating considerable cooling and stabilization as well as causing electronic switching problems. Different structures of magnetic traps have been widely used and a few, such as the QUIC traps [@QUIC], the ferromagnetic traps [@old; @electromagnet], the permament magnets [@hulet] or a combination [@lewandowski] have very low power consumption. Another system called “atom chip" involves a surface magneto-optical trap (MOT) and a magnetic trap based on a wafer with lithographically patterned wires [@Puce]. This technology is compact, generates condensates with unprecedented rapidity, and holds promise for eventually being simpler and more robust than traditional condensate machines. Nevertheless it requires more expertise to fabricate [@rugositechip] and might not allow for high fluxes because of its small size. The latest approach is the all optical method [@OpticalBEC]. By eliminating the magnetic trap altogether, this method indeed may eventually become the simplest route to BEC, but the need for high power lasers makes it still difficult with the present technology. Our system is a hybrid magnetic trap that falls in the first category. It uses strong iron core magnets to produce radial confining fields and low power electromagnetic coils to produce axial confinement and a bias field. The ferromagnets guide and concentrate the magnetic flux where it is needed, i.e. on the atoms, and thus can be used to create a compact and light aparatus that consumes very little power. They are used above saturation threshold and thus, do not need to be actively temperature controlled like permanent magnets [@hulet] and only need moderate cooling compared to electromagnetic coils producing the same field. The tight confinement from the ferromagnets insures that we have the necessary collision rate to evaporatively cool the atoms. The longitudinal bias field, the only parameter for which stability is critical, is produced by servo-controlled electromagnetic coils, which are water cooled. In this paper, we report on the development of this new hybrid electromagnet. In the first section we explain the main advantages of ferromagnetic trapping. In the second and third section we describe this electromagnet and present the way it is used to cool $^{87}$Rb atoms. In the last section we present stability measurements made using atom lasers. 1. Why ferromagnetic trapping? ============================== Ferromagnetic structures [@old; @electromagnet], which have been developed [@patent] in our group, make possible high gradient fields with reasonable power consumption and large condensates. Ferromagnetism [@Ashcroft] consists in exciting a magnetic material with coils. This excitation orientates the microscopic magnetic moments, so that field in the material is enhanced. Due to the high permittivity of ferromagnetic materials, the ferromagnetic core guides the magnetic flux. To understand this effect, let us consider the magnetic circuit represented in figure \[Culasse\]. The two tips are separated by a gap $e$ of a few centimeters. The ferromagnetic structure has a total length $l$ and a section $S\propto r^2$. The whole structure is excited by a coil of $N$ loops driven by a current $I$, leading to an excitation $NI$. From Ampere’s theorem [@Jackson], we can introduce the reluctance $R_{\rm iron}$ : $$R_{\rm iron}=\oint_{\Gamma_{\rm i}} \frac{dl}{\mu_{r}\mu_{0}}\simeq\frac{l}{\mu_{r}\mu_{0}} \label{eq:1}$$ inside the iron core and $R_{\rm gap}$ : $$R_{\rm gap}=\oint_{\Gamma_{\rm g}} \frac{dl}{\mu_{0}}\simeq\frac{e}{\mu_{0}} \label{eq:1bis}$$ in the gap between the tips. A simple relation between the excitation $NI$ and the magnetic flux $BS$ can be written : $$BS=\frac{NI}{R_{\rm iron}+R_{\rm gap}} \label{eq:2}$$ Since $\mu_{r}$ is very large ($\mu_{r}> 10^{4}$) for ferromagnetic materials, only the gap contribution is important ($\mu_{r}\gg l/e$) [@foot1]. A more complete calculation [@DesThes] shows that the field created in the gap is similar to that created with two coils of excitation $NI/2$ placed close to the tips as represented in figure \[Culasse\]. Thus, guiding of the magnetic field created by arbitrary large coils far away from the rather small trapping volume is achieved. This is only true if a yoke links a north pole to a south pole. If not, no guiding occurs and the field in the gap is significantly reduced. Ferromagnetic materials are also subjected to hysteresis cycles generating remnant fields which must be compensated to offer an appropriate magnetic environment. This is somewhat difficult to accomplish as long as some coupling exists between the dipole and the quadrupole causing the remnant field to have a complex geometry. In this case, several additional coils are required to provide the right compensating field as previously described in [@old; @electromagnet]. To avoid these couplings, we use a partially-ferromagnetic electromagnet where iron cores are only used where strong fields (gradients) are needed (i.e. the quadrupole part of the Ioffe-Pritchard configuration). With this configuration, we can independantly act on the bias and the gradients, and the remnant fields retain a simple geometry. 2. Description and characterization of the electromagnet ======================================================== The configuration of the electromagnet is shown in Fig. \[FigElectromagnet\] and \[FigPolesAndFields\]. The quadrupole field is produced by two ferromagnetic pieces, each one composing two poles in a way very similar to figure \[Culasse\]. Each ferromagnetic piece is excited by a coil of 5 cm diameter made with hollow copper wire (5 mm diameter) with 10 windings. The hollow wire allows the exciting coils to be water cooled. Poles are 2 cm wide and are bevelled mainly in order to better guide the magnetic field but also to leave some space for the vertical magneto-optical trap beams. The dipolar coils are 100 turn coils of copper wire excited by equal current and separated each other by 3 cm. They are conical to facilitate the path of the magneto-optical trap beams and positioned on a macor and brass support, which is also water cooled. Another pair of electromagnetic coils called “anti-dipole coils" is used to make the bias adjustable. These coils are larger than the dipole coils, separated from each other by 10 cm and placed at the support termination. The dipole and antidipole coils are mounted in series, but with the currents flowing in opposite directions, so as to cancel out noise in the magnetic field due to current fluctuations. We can set the value for the bias field by controling the number of windings of the anti-dipole coils [@foot2]. Static performances of the ferromagnetic structure The maximum achievable quadrupole gradient $b'$ is 830 G/cm, obtained for a saturating current of 60 A (see Fig. \[FigQuadGrad\]). We limit the current in the dipole coils to 15 A, beyond which heating becomes too important. With this value, a bias $B_0$ and a curvature $b''$ of respectively 54 G and 150 G/cm$^2$ are obtained for the dipole and 49 G and 68 G/cm$^2$ for the anti-dipole (see Fig. \[FigDipPlusAntiDip\]). The total bias can thus be varied from 54 to 5 Gauss. The modulus of the magnetic field created by the electromagnet is then given by: $$B=\sqrt{\left(B_0+\frac{b''x^2}{2}\right)^2+\left(b'^2-\frac{b''B_0}{2}\right)\left(y^2+z^2\right)} \label{chpmag}$$ Dynamical properties The dynamical characteristics of the electromagnet are also of a great interest since it is necessary to be able to switch on and off the magnetic field in times shorter than the oscillation period of the trap. There are two important reasons for this condition. The first one is that we wish to transfer the atoms from the magneto-optical trap to the magnetic trap without losing in density. The second reason is due to the way we image the atomic cloud: it is released from the trap and expands ballistically before being illuminated by a near-resonant probe beam (absorption imaging [@WKetterle]). From the images, we can deduce temperature, density and the number of trapped atoms. To avoid eddy currents in the ferromagnetic structure, which would limit switching times, we use a laminated material. It is worth noting that the shape of the present quadrupole is simple enough to be easily produced by assembling inexpensive laser-cut sheets of ferromagnetic material. It consists of 100 $\rm{\mu m}$ thick iron-silicium plates, each isolated from the other. To control the switching times, we use a capacitor $C$ in series with the coils of inductance $L$. To turn on the magnetic field, the capacitors ($C_{\rm charge}$) are first loaded with a high voltage supply $U$. Then the energy $1/2\,C_{\rm charge}U^2$ accumulated in the capacitors is transferred to the coils in a time equal to $2\pi\sqrt{LC_{\rm charge}}/4$, i.e. a quarter of a period of the oscillating circuit. Then, the power supply takes over. The same idea applies to the procedure to switch off the magnetic field, but now, the switch off current will charge a capacitor called $C_{\rm discharge}$ which limits overvoltage in the circuit at switch off. For our setup, the switching time is $150~\rm{\mu s}$. The electronic circuit used for the quadrupole is presented in Fig. \[FigCircQuad\]. For the dipolar coils (Fig. \[FigCircDip\]), the switching procedure is the same, but now we also ramp the current in “anti-dipole" coils by using MOSFET so as to precisely adjust the bias at a low value $B_0$ without decreasing the curvature $b''$. Cancelling of the remnant magnetic field To generate the compensating field in order to cancel remnant fields, two thin coils (1 mm diameter wire) have been rolled on the two quadrupole exciting coils. They have thus the same configuration and the compensation field automatically possesses the right geometry to cancel the remnant field. The procedure to find the compensation current is quite simple. We adjust the current after having cooled atoms into a magneto-optical trap [@Raab] and then cut the MOT magnetic field. If the compensating current is too high or too low, the remnant field provokes a rotation of the atomic cloud, a situation which can be compared to the mechanical Hanle effect [@Hanle]. At the right value, there is no rotation anymore, which can be seen by observing fluorescence along the x axis. We should stress the point that the remnant fields are so easy to compensate only because we limited the ferromagnetic structure to the quadrupolar plane so as to suppress coupling between quadrupole and dipole. 3. Atoms trapped in the electromagnet ===================================== In order to trap the atoms, the electromagnet is placed around a 1 cm square section parallelepipedic glass vacuum cell (see Fig. \[FigBobines\]). The atoms are first loaded into a magneto-optical trap whose magnetic field is produced by two thin printed circuits (400 $\mu$m thick) positioned along the $y$ direction. On each one are imprinted 6.5 loops which are 1 mm wide and 35 micrometers thick, such that the full resistance is 1 $\Omega$. A current of 2 A in those circuits generates a spherical quadrupole field with an axial gradient of 15 Gauss/cm. A slave laser, injected by a grating stabilized diode laser, produces the six independent circularly polarized trapping beams with a diameter of one centimeter. The MOT sequence ends with an optical molasses [@DalCoh] phase during which the spherical quadrupole field is switched off and the laser frequency is detuned. It is important that there is no residual magnetic field -bias or gradient- during the molasses phase so as not to disturb the sub-doppler cooling mechanism. To ensure this requirement, three pairs of coils (“Molasses coils" in Fig. \[FigBobines\]) are placed along the three axes to compensate any residual fields (terrestrial field, ionic pump field...) additional to remnants fields. We obtain $6\times 10^8$ atoms at $50~\mu$K after the molasses. After the molasses phase, the atoms are optically transferred into the $F=1$ hyperfine state, in which only the $m_F=-1$ sublevel experiences an attractive potential which can be harmonically approximated around the minimum of the magnetic field with the following trapping frequencies: $$\begin{aligned} \omega_{\rm ax} & = & \sqrt{\frac{g_F\,m_F\,\mu_B}{m} b''}\\ \omega_{\rm rad} & = & \sqrt{\frac{g_F\,m_F\,\mu_B}{m}\left(\frac{b'^2}{B_0}-\frac{b''}{2}\right)}\end{aligned}$$ where $g_F=-1/2$ is the Landé factor and $\mu_B$ the Bohr magneton. The values of the trap parameters are presented in table \[TableCaract\]. [8.5cm]{}[l@ \[5]{}[c @]{}]{} Trap & $B_0{\rm(G)}$ & $b''{\rm(G/cm^2)}$ & $b'{(\rm G/cm)}$ & $\omega_{\rm rad}{\rm (Hz)}$ & $\omega_{\rm ax}{\rm (Hz)}$\ NCIPT & 54 & 150 & 200 & 2 $\pi \times$ 15 & 2 $\pi \times$ 11\ CIPT & 5 & 82 & 830 & 2 $\pi \times$ 330 & 2 $\pi \times$ 8 The magnetic Ioffe-Pritchard trap (IPT) is switched on by applying a current of 15 A in the dipole coils and 5 A in the quadrupole exciting coils. These values are selected so that the size of the two traps -MOT and IPT- are well matched. Moreover, the centers of the two traps have to overlap, otherwise the atoms will acquire a larger energy during the transfer. To adjust the relative position of the trap centers, we have added two pairs of coils in a Helmholtz configuration in series with the MOT’s coil. They act on the field in the $x$ and $z$ directions. Variable resistances are added in parallel to the three pairs of coils (including the MOT coils) as shown in Fig. \[FigPositioning\], allowing for tuning of the current in the positioning coils. The MOT, which will be centered around the zero of the magnetic field, can now be placed appropriately using the positioning coils system described above. It is to be noted that the right values for each resistance must be found following an iterative method. By moving the center of the MOT, as explained above, the magnetic environment is slightly modified and the molasses coils need to be fine tuned. This in turn displaces the MOT and we have to center the cloud again. After having transfered the atoms into the IPT, current in the quadrupole is increased to 60 A and the current in the anti-dipole coils is slowly ramped up. At this point, the quadrupole field gradient reaches 830 G/cm and the bias is reduced to 5 Gauss. This provides a tight trap for which the collision rate is around 60 s$^{-1}$, more than one order of magnitude larger than before compression. The compression of the trap is adiabatically done in one second in order to minimize the loss of phase space density. Compression time is thus chosen larger than the oscillation time and the rethermalization time. Current values at each stage are summarized in Fig. \[FigCycleMag\]. At this stage, we obtain typically $2\times 10^8$ atoms at $250~\mu$K i.e. a phase space density of $\mathrm{5 \times 10^{-7}}$. Finally, evaporative cooling of the atoms is performed by radio-frequency (rf) induced spin flips. The rf magnetic field is produced by the MOT coils (Fig. \[FigPositioning\]), in which the rf current now flows in the same direction for both coils such that the field is nearly homogeneous on the atoms. It is applied perpendicular to the IPT magnetic field at the center of the trap. Then, the radiofrequency is swept from 70 MHz to a final value of around 4.2 MHz. Fig. \[FigTransitionImg\] shows images of cooled atomic clouds after 25 ms of ballistic expansion for different final values of the ramp. The sudden appearence of the BEC can be seen in the density profile measured by absorption imaging along the x axis, as the distribution becomes bimodal. The distribution of thermal atoms is Gaussian and is represented by a dashed line in the profiles, whereas the distribution of condensed atoms is an inverted parabola (in the Thomas-Fermi approximation)[@Stringari]. For decreasing temperatures (lower final rf frequency), the fraction of thermal atoms decreases and at sufficiently low temperatures, we obtain a pure condensate containing $10^6$ atoms. 4. Stability ============ One efficient way to check the stability of the bias created by the electromagnet is to produce atom lasers by radio-frequency outcoupling [@Bloch99](Fig. \[atomlas\]) and mesure the spectral width of the condensate. The principle of weak radio-frequency outcoupled atom lasers has been developped in [@paplaser]. Main features are summarized in figure \[atomlasprinciple\]. The condensate is displaced from the center of the magnetic field due to gravity. Thus, it crosses magnetic equipotentials in the vertical direction only. By applying radio-frequency one flips the spin of atoms from the trapped state $\|F=1,m_F=-1>$ to the untrapped one $\|F=1,m_F=0>$. The wave falls under the effect of gravity into the continuum of Airy functions [@Landau]. This gives rise to an atom laser since the source (the condensate) is fully coherent [@cohbec]. The radio-frequency value $\nu_{RF}$ is directly related to the difference in energy between the two involved states: $$h\nu_{RF}=E_{\rm BEC}-E_{\rm free}(\vec{r})=E_{\rm bias}+\mu-E_{\rm free}(\vec{r})$$ where $E_{ \rm free}(\vec{r})$ stands for the potential energy of outcoupled atoms and depends on the extrating point $\vec{r}$, $\mu$ and $E_{\rm bias}$, respectively the chemical potential and the magnetic potential energy of the BEC. For a given radiofrequency, fluctuations of the coupling zone are only related to bias fluctuations through $E_{\rm bias}$. To quantitavely give an estimation on the stability of the experiment, we have to determine the spectral width of the condensate. For a given weak radio-frequency power, we have measured the number of extracted atoms as a function of the value of the rf knife (experimental points in figure \[width\]). If we assume that the number $N_e$ of extracted atoms is proportional to the number of condensed atoms on the surface $\Gamma(\delta\nu)$ defined by the rf value $\nu_{RF}$ (i.e. one neglects the effect of interactions on the coupling rate [@paplaser]). Taking a Thomas-Fermi profile for the condensate leads to: $$N_e(\delta\nu)\propto \int_{\Gamma(\delta\nu) } \,d \vec{x} \,{\rm max}\left(1-\sum_{i=x,y,z}\left(\frac{x_i}{R_i}\right)^2;0\right)$$ where $\delta\nu=\nu_{RF}-\nu_0$ is the difference between outcoupling radio-frequency and its value in the center of the BEC, $x_i$ and $R_i$ being respectively the coordinate and the Thomas-Fermi radius of the dimension $i$. The surface $\Gamma(\delta\nu)$ corresponding to the radiofrequency knife resonance condition is defined by: $$\omega_{\rm ax}^2 x^2 +\omega_{\rm rad}^2\left(y^2+z^2\right) + 2gy=\frac{2h}{m}\delta\nu$$ If we assume that fluctuations are gaussian with r.m.s value $\sigma$, the final output flux is the convolution between this noise and $N_e$: $$F(\delta \nu)=\int du \,N_e(u) \,e^{-(u-\delta\nu)^2/2\sigma^2}$$ We have fitted the experimental data with $F$ as shown in figure \[width\], and we obtain $\sigma=$1.5 kHz. This broadening could come from bias variations during one scan. In order to look more precisely for the different timescales involved in these fluctuations, we have fixed the rf outcoupler at a detuning of $\mathrm{+5\, kHz}$ and we have looked at the long term drift and fluctuations of the number of outcoupled atoms. Once we have deduced the effects of atom number fluctuations in the original BEC ($\mathrm{8\,\%}$), we have measured a drift of $\mathrm{500 \,Hz/hour} \,(\mathrm{0.7 \,mG/hour})$ and r.m.s. fluctuations of $\mathrm{700\, Hz}$ which correspond to shot to shot bias fluctuations of $\mathrm{1\, mG}$. Nevertheless, it is likely that the rapid fluctuations are also of the order of $\mathrm{1\,mG}$, which is sufficient to produce stable quasi-continuous atom lasers as one can see on figure \[atomlas\]. Conclusion ========== In this paper, we have presented the realization of a new generation hybrid electromagnet designed to produce large Bose-Einstein condensates containing one million Rubidium atoms. Due to the compact ferromagnetic design (20 cm diameter) one achieves gradients as large as 830 G/cm which are sufficient to compress the atomic cloud for an efficient runaway evaporative cooling. The total power consumption is less than 800 W which is easy to dissipate. Due to the lamellar structure of the ferromagnetic core and the driving electronic circuit, eddy currents are suppressed and the switching time is reduced to 150 $\mu$s. The high stability of the bias (fluctuations measured to be of the order of $\mathrm{1 \, mG}$) enables us to produce stable radio-frequency outcoupled quasi-continuous atom lasers. Y. Le Coq acknowledges support from the CNES post-doctoral fellowship program. This work is part of the CNES supported ICE project (DA No. 10030054) with initial support from Laboratoire National d’Essai, Délégation Générale de l’Armement (Contract No. 99-34-050), the European Union (Cold Quantum Gas network) and INTAS (Contract No. 211-855). We would like to thank John Gaebler for his careful reading of this manuscript. [llml]{} B.Desruelle [[et al]{}]{}, Eur. Phys. J. D [1]{}, 255 (1998) M.H. Anderson [[et al]{}]{}, Science [269]{}, 198 (1995); C.C. Bradley [[et al]{}]{}, Phys. Rev. Lett. [75]{}, 1687 (1995); K.B. Davis [[et al]{}]{}, [[ibid.]{}]{} [75]{}, 3969 (1995); C.C. Bradley [[et al]{}]{}, [[ibid.]{}]{} [78]{}, 985 (1997) M.-O. Mewes [[et al]{}]{}, Phys. Rev. Lett. [78]{}, 582 (1997) B.P. Anderson and M.A. Kasevich , Science [282]{}, 1686 (1998) E.W. Hagley [[et al]{}]{}, Science [283]{}, 1706 (1999) I.Bloch [[et al]{}]{}, Phys. Rev. Lett. [82]{}, 3008 (1999) W. Chow [[et al]{}]{}, Rev. Mod. Phys. [57]{}, 61 (1985) G.E. Stedman [[et al]{}]{}, Phys. Rev. A [51]{}, 4944 (1995) G.E. Stedman, Rep. Prog. Phys. [60]{}, 615 (1997) L. Deng [[et al]{}]{}, Nature [398]{}, 218 (1999) T.L. Gustavson [[et al]{}]{}, Phys. Rev. Lett. [78]{}, 2046 (1997) S. Gupta [[et al]{}]{}, Phys. Rev. Lett. [89]{}, 140401 (2002) M. Andrews [[et al]{}]{}, Science [275]{}, 637 (1997) E.A. Cornell and C.E. Wienman, Rev. Mod. Phys. [74]{}, 875 (2002); W. Ketterle, Rev. Mod. Phys. [74]{}, 1131 (2002) Y. Le Coq [[et al]{}]{}, cond-mat/0501520 (2005) J.M. McGuirck [[et al]{}]{}, cond-mat/0403254 (2004) K. B. Davis [[et al]{}]{}, Phys. Rev. Lett. [74]{}, 5202 (1995) O.J. Luiten [[et al]{}]{}, Phys. Rev. A [53]{}, 381 (1996) H.J. Lewandowski [[et al]{}]{}, Jour. of Low Temp. Phys. [132]{}, 309 (2003) T.Esslinger [[et al]{}]{}, Phys. Rev. A [58]{}, R2664 (1998) J.J. Tollet [[et al]{}]{}, Phys. Rev. A [51]{}, R22 (1995) D.E. Pritchard, Phys. Rev. Lett. [51]{}, 1336 (1983) Hänsel [[et al]{}]{}, Nature [413]{}, 498 (2001) J. Estève [[et al]{}]{}, Phys. Rev. A [70]{}, 043629 (2004) M.D. Barett [[et al]{}]{}, Phys. Rev. Lett. [87]{}, 010404 (2001) P. Bouyer, A. Aspect [[et al]{}]{}, [[Dispositif électromagnétique pour la production d’atomes neutres froids]{}]{}, Brevet No. 00 02704 (France) (2000) N.W. Ashcroft, N.D. Mermin, [[Solid State Physics]{}]{}, Brooks Cole (1976) J. Jackson, [[Classical Electrodynamics]{}]{}, Wiley, New York (1962) The case of very small gap where $R_{iron}\geq R_{gap}$ was studied in [@Vuletic96]. In this case, the ferromagnetic materials amplify the magnetic field in the gap. V. Vuletic [[et al]{}]{}, Europhys. Lett. [36]{}, 349 (1996) B. Desruelle [Evaporation par radio-fréquence et condensation de Bose-Einstein d’un gaz d’alcalins en régime de champ fort* ]{} (1999) Thesis, Université Paris XI It is worth noting that having an adjustable bias is not simple in the case of an entire ferromagnetic structure : the anti-dipole field would couple into the ferromagnetic structure, which would lead to a desexcitation of the material and would reduce the longitudinal curvature too much. W. Ketterle [[et al]{}]{}, [Bose-Einstein condensation in atomic gases, Proceedings of the international school of Physics Enrico Fermi, course CXL]{} (1999) E.L. Raab [[et al]{}]{}, Phys. Rev. Lett. [59]{}, 2631 (1987) R. Kaiser [[et al]{}]{}, Z. Phys. [D 18]{}, 17 (1991) J. Dalibard and C. Cohen-Tannoudji, J. Opt. Soc. Am. B [6]{}, 2058 (1989) F. Dalfovo [[et al]{}]{}, Rev. Mod. Phys. [71]{}, 463 (1999) F. Gerbier [[et al]{}]{}, Phys. Rev. Lett. [86]{}, 4729 (2001) L.Landau, E. Lifchitz, [[Cours de Physique théorique, Mécanique quantique]{}]{}, Editions Mir, Moscou (1975) I. Bloch [[et al]{}]{}, Nature (London) [403]{}, 166 (2000) [^1]: Present address: Raman Research Institute, C.V. Raman Avenue, Sadashivanagar, BANGALORE 560080 (INDIA)	{ "pile_set_name": "ArXiv" }
--- abstract: 'We present a measurement protocol for a flux qubit coupled to a dc-Superconducting QUantum Interference Device (SQUID), representative of any two-state system with a controllable coupling to an harmonic oscillator quadrature, which consists of two steps. First, the qubit state is imprinted onto the SQUID via a very short and strong interaction. We show that at the end of this step the qubit dephases completely, although the perturbation of the measured qubit observable during this step is weak. In the second step, information about the qubit is extracted by measuring the SQUID. This step can have arbitrarily long duration, since it no longer induces qubit errors.' author: - 'I. Serban' - 'B.L.T. Plourde' - 'F.K. Wilhelm' date: title: 'Quantum nondemolition-like, fast measurement scheme for a superconducting qubit' --- Introduction ============ The quantum measurement postulate is often viewed as the most intriguing assumption of quantum physics. Much of it has been demystified by the study of the physics of quantum measurements. The dynamics of the measurement process can be described by a coupled many-body Hamiltonian, consisting of the system to be measured and the detector with a heat bath component [@Braginsky95; @Caves80]. Thus, the measurement process can be investigated using the established tools of quantum mechanics of open systems [@Nato06II; @Weiss99; @Keil01; @Lindblad76]. Most interest has been focused on the physics of weak measurements, where the system-observer coupling can be treated within perturbation theory. Famously, this research has shown that only a certain class of measurements satisfy von Neumann’s quantum measurement postulate [@Cohen92; @Neumann55] and indeed project the system wavefunction onto an eigenstate of the measured observable. Measurements of this type are termed quantum nondemolition (QND) measurements. Within the weak measurement paradigm, the QND regime is achieved when the measured observable is a constant of the free motion and commutes with the system-detector coupling Hamiltonian. Weak QND measurements have been investigated in various systems, ranging from spins to oscillators and even photons [@Bulaevskii04; @Averin02; @Jordan05; @Jordan06; @Milburn83; @Sanders89; @Brune90; @Boulant07; @boissonneault08]. The dynamics of the weak measurement process has practical relevance in the context of quantum computing. Specifically, superconducting qubits have been proposed as building blocks of a scalable quantum computer [@Makhlin01; @Devoret04; @Nato06I; @Clarke08], and a fast measurement with a high resolution and visibility is important for readout and also for error correction. There are a variety of different measurement techniques used in superconducting qubits. Weak measurements can be performed using single-electron transistors [@Makhlin01]. A different approach is the switching measurement, where the detector switches out of a metastable state depending on the state of the qubit [@Science00; @EPJB03; @Vion02; @Martinis02; @steffen:050502]. Such switching measurements have been a quite successful readout scheme for many superconducting qubit experiments to date. However, the dissipative nature of the switching process imposes limitations on the measurement speed and perturbs the qubit state. A QND measurement could be achieved by using a pointer system, and measuring one of its observables influenced by the state of the qubit [@PRBR033]. Recent developments of such detection schemes, using an oscillator as the pointer, have led to vast improvements [@Lupascu04; @Lupascu07; @Lee05; @Blais04; @Schuster05; @Wallraff05; @Metcalfe07] over previous measurement protocols. It has previously been shown [@Lougovski06; @Bastin06; @Santos07; @Storcz06] that infinitesimally short interaction between a qubit and an oscillator is sufficient to imprint information about the state of the oscillator onto the qubit. The similar idea of using a short interaction to transfer information about the qubit into the oscillator has been used [@measurement07] in a dispersive readout scheme. In this case, after a short interaction, the state of the oscillator contains information about the qubit which can be extracted by further measuring one of its observables, for example, momentum. However, this scheme did not take possible bit flip errors into account. These errors may occur in the short yet finite time when the qubit is in contact with its environment. Thus, the full power of a quasi-instantaneous measurement has not yet been explored. In this paper we describe the effect of an ideally extremely short and arbitrarily strong interaction of a qubit with its environment (consisting of a weakly damped harmonic oscillator). We investigate the back-action on the qubit when the measured observable does not commute with the Hamiltonian describing the interaction with the environment, and study how close this result approximates the QND measurement. We study a setup consisting of a flux qubit inductively coupled to a dc-SQUID magnetometer. The flux qubit consists of a superconducting loop with three Josephson junctions [@Mooij99; @Orlando99]. For flux bias near odd half-integer multiple of $h/2e$, the qubit is represented by two circulating current states with opposite directions. During the entire measurement process the SQUID is coupled to measurement circuitry, with associated dissipative elements. However, it never switches out of the zero dc-voltage state. The qubit-SQUID interaction of arbitrary strength is turned on only for a short time by applying a very short bias current pulse to the latter. During this time, information about the qubit is imprinted onto the SQUID and can later be extracted from it during the post-interaction phase by monitoring voltage oscillations across the device. When the current pulse is switched off, the qubit-SQUID interaction ideally vanishes and the environment no longer perturbs the qubit. Thus, one can afford a long time to measure the SQUID and determine the state of the qubit. In section \[method\], following Ref. [@Tian02], we model the qubit-SQUID system by a two-level system linearly coupled to a dissipative oscillator. We describe the evolution of this system by means of a master equation in the Born-Markov approximation [@Blum96], valid for the underdamped SQUID. In section \[results\] we discuss the qubit-oscillator evolution during both interaction and post-interaction phases. We study the qubit dephasing and relaxation during the interaction phase. We show that, at the end of this phase, the qubit appears completely dephased. In other words, the qubit has been measured and its information has been transferred in the form of a classical probability to the oscillator. During the same time interval, we find that qubit relaxation has remained negligible. For the post-interaction phase we describe the evolution of the oscillator under the influence of the environment, starting from the state prepared by the interaction with the qubit. Technically, extracting the qubit information amounts to measuring the amplitude of the ringdown of the oscillator momentum. In section \[implementation\] we discuss some of the details involved with implementing this measurement scheme. Model and method {#method} ================ We study a flux qubit inductively coupled to a dc-SQUID, with one possible setup shown schematically in Fig. \[circuit\] (a). We describe a more detailed setup for implementing this scheme in section \[implementation\]. ![(Color online). (a) Simplified circuit consisting of a flux qubit inductively coupled to a SQUID with two identical junctions and shunt capacitance $C_S$. The SQUID is driven by an bias step-like dc pulse $I_B(t)$ and the voltage drop $V(t)$ is measured by a device with internal resistance $R$. (b) Illustration of the measurement scheme: coupling ($t=0$) and decoupling ($t=\tau$) of the qubit and the SQUID (oscillator) and the evolution of a point of mass in the transition of potential from one harmonic oscillator to a superposition of two displaced oscillators and back. The dashed (red) and the continuous (green) lines correspond to the different states of the qubit.[]{data-label="circuit"}](circuit.pdf){width="\columnwidth"} The SQUID is characterized by a two-dimensional washboard potential for the two independent phases corresponding to the two junctions [@Lefevre92]. Their sum couples to bias current driven through the SQUID, while the difference of phases couples to the magnetic flux applied to the SQUID. The small oscillations in these two directions can have vastly different characteristic frequencies. In particular, a small geometric inductance and a low critical current can make the flux mode frequency large while a shunt capacitor can lower the bias current mode frequency substantially. In the limit of very different frequencies, one can approximate the SQUID dynamics as that of a one-dimensional oscillator in the bias current direction, with the position of the oscillator minimum dependent on both $I_B$ and the total flux coupled to the SQUID which, for example, could vary depending on the state of the qubit. The setup of Fig. \[circuit\] (a) can be described by the effective Hamiltonian [@Tian02] $$\begin{aligned} \hat{H}&=&\hat{H}_S+\hat{H}_I+\hat{H}_B,\\ \hat{H}_S&=&\hbar w\hat{\sigma}_z+\hbar\delta\hat{\sigma}_x+\hbar\Omega(\hat{a}^\dagger \hat{a}+1/2)\nonumber\\ &+&\hbar(\Theta(t)-\Theta(t-\tau))(\hat{a}+\hat{a}^\dagger)(\hat{\sigma}_z\gamma+K),\nonumber\end{aligned}$$ where $\hat{H}_S$ is the Hamiltonian for the qubit-SQUID oscillator system, $w$ is the qubit energy and $\delta$ the tunneling matrix element, $\hat{H}_B$ is the Hamiltonian for the dissipative environment of the measurement circuitry, $\hat{H}_I$ describes the interaction between the SQUID oscillator and the environment, and $\Theta$ is the Heaviside step function. We note that for a continuous shape of the current pulse similar results are expected, as long as the switching is not adiabatic. Here the SQUID is described, in the lowest-order approximation, by a harmonic oscillator with frequency $\Omega$, i.e. the plasma frequency of the bias current degree of freedom. This frequency also depends on the applied bias current, as shown in Appendix \[parameters\]. This dependence leads to a enhanced ring-down frequency after the pulse is switched off. This change in the SQUID plasma frequency does not, in the first approximation, depend on the qubit state, therefore it will not qualitatively affect this method of discrimination. For the following derivation we assume the SQUID plasma frequency constant (the value during the bias pulse), noting that the ring-down oscillations occurring in the post-interaction phase have in practice a somewhat higher frequency, but otherwise unchanged behavior. The dispersive, next-to-leading order component of the qubit-oscillator coupling [@measurement07] becomes significant in the absence of a linear component for very weak bias pulse, which is not the limit we investigate here. In the following, the effects of the linear component are investigated. We focus on the regime where the qubit-SQUID interaction displaces the state by more than its zero-point fluctuation but does not yet explore the classical nonlinearity. The first consequence of the nonlinear component may be to add more phase shift to the ringdown oscillations. In the measurement protocol proposed here we assume a symmetric SQUID. The qubit-oscillator coupling strength is tuned by the bias current $I_B$ [@Bertet05d]. When $I_B=0$, the qubit and the SQUID are decoupled. By using a fast current pulse, the qubit-oscillator interaction of arbitrary strength $\gamma$ is turned on only for the short time $\tau$ allowing information about the qubit to be imprinted onto the oscillator. During this time, the SQUID oscillator is displaced according to the qubit state. After the coupling is switched off, the SQUID oscillator phase particle returns to the original position after undergoing ring-down oscillations that decay with a damping determined by the SQUID measurement circuitry. The parameter $K$ describes the strength of the bias current kick induced in the oscillator, caused by the abrupt shift in the minimum of the SQUID potential energy from the bias current pulse, in the absence of a qubit. For the expressions of the parameters $\gamma$ and $K$ and their explicit dependence on $I_B$ see Appendix \[parameters\]. During the entire measurement process the oscillator is coupled via a linear Hamiltonian $\hat{H}_I$ $$\begin{aligned} \hat{H}_I&=&\sum_i\frac{\hbar\lambda_i(\hat{a}\hat{b}^{\dagger}_i+\hat{a}^{\dagger}\hat{b}_i)}{\sqrt{2m\Omega}},\end{aligned}$$ to a dissipative environment described by a bath of harmonic oscillators $$\begin{aligned} \hat{H}_B&=&\sum_i\hbar\omega_i\left(\hat{b}^{\dagger}_i\hat{b}_i+\frac{1}{2}\right),\end{aligned}$$ with Ohmic spectral density $J(\omega)=\sum_i\lambda_i^2\hbar\delta(\omega-\omega_i)=m\hbar\kappa\omega\Theta(\omega-\omega_c)/\pi$ [@Ingold98]. Here $[\kappa]=s^{-1}$ is the photon loss rate. The cut-off frequency $\omega_c$ is physically motivated by the high-frequency filter introduced by the capacitors. This environment represents the dissipative element contained in any measuring device. We now describe the dynamics of the qubit and SQUID oscillator during the various phases of our measurement scheme. The interaction phase --------------------- At $t=0$, before the bias current is rapidly pulsed on and the qubit and SQUID interact strongly, we assume the factorized initial state $\hat\rho(0)=\hat\rho_S(0)\otimes\hat\rho_B(0)$. The oscillator interaction with the bath is supposed to be weak, and assuming a Markovian environment, we obtain the standard master equation for the qubit-oscillator reduced density matrix $\hat\rho_S(t)={\rm Tr}_B\left\{\hat\rho(t)\right\}$ in the Born approximation $$\begin{aligned} \dot{\hat{\rho}}_S(t)&=&\frac{1}{\mathbbm{i} \hbar}\left[\hat{H}_S,\hat{\rho}_S(t)\right] \label{me_bm1}\\ &-&\frac{1}{\hbar^2}\int_0^t dt'{\rm Tr}_B \!\left[\hat{H}_{I},\left[\hat{H}_{I}(t,t'),\hat{\rho}_S(t)\otimes\hat{\rho}_B(0)\right]\right],\nonumber \end{aligned}$$ where $$\hat{H}_{I}(t,t')= \hat{U}_{t'}^{t}\hat{H}_{I}\hat{U}_{t}^{t'},\:\:\hat{U}_t^{t'}=\mathcal{T}\exp\left(\int_t^{t'}\!\!\!\!d\tau\frac{\hat{H}_S+\hat{H}_B}{\mathbbm{i} \hbar}\right),$$ and $\mathcal{T}$ is the time-ordering operator. This approach is valid at finite temperatures $k_BT \gg \hbar \kappa$ and times $t\gg1/\omega_c$ [@Alicki06; @Nato06II], which is the limit we will discuss henceforth. In the qubit $\hat{\sigma}_z$ eigen-basis the density matrix and the qubit-oscillator Hamiltonian read $$\begin{aligned} \hat{\rho}_S & = & \left(\begin{matrix} \hat{\rho}_{\uparrow\uparrow}&\hat{\rho}_{\uparrow\downarrow} \\ \hat{\rho}_{\downarrow\uparrow}&\hat{\rho}_{\downarrow\downarrow}\end{matrix}\right),\label{matrix}\\ \hat{H}_{S\downarrow\uparrow}&=&\hat{H}_{S\uparrow\downarrow}=\hbar\delta,\:\:\:\:r_{\sigma}=\langle\sigma\|\hat{\sigma}_z\|\sigma\rangle,\:\:\sigma\in\{\uparrow,\downarrow\},\\ \hat{H}_{S\sigma\sigma}&=&\hbar (r_{\sigma} w+\Omega(\hat{a}^\dagger \hat{a}+1/2)\\ &+&(r_{\sigma}\gamma+K)(\hat{a}+\hat{a}^\dagger)).\nonumber\end{aligned}$$ In the following, we assume that the environment acts on each matrix element of (\[matrix\]) in the same way. This is a valid assumption in the case of very weak damping and $\delta/ w\ll1$ for an Ohmic bath. Within this assumption we obtain $$\begin{aligned} \dot{\hat{\rho}}_{\sigma\sigma}&=&\frac{1}{\mathbbm{i} \hbar}[\hat{H}_{\sigma\sigma},\hat{\rho}_{\sigma\sigma }]-\mathbbm{i}\delta r_\sigma (\hat{\rho}_{\downarrow\uparrow}-\hat{\rho}_{\uparrow\downarrow})+\hat{\mathcal{L}}\hat{\rho}_{\sigma\sigma},\\ \dot{\hat{\rho}}_{\uparrow\downarrow}&=&\frac{1}{\mathbbm{i} \hbar}(\hat{H}_{\uparrow\uparrow}\hat{\rho}_{\uparrow\downarrow}-\hat{\rho}_{\uparrow\downarrow}\hat{H}_{\downarrow\downarrow})+ \mathbbm{i} \delta(\hat{\rho}_{\uparrow\uparrow}-\hat{\rho}_{\downarrow\downarrow})+\hat{\mathcal{L}}\hat{\rho}_{\uparrow\uparrow}\nonumber\\ \dot{\hat{\rho}}_{\downarrow\uparrow}&=&\frac{1}{\mathbbm{i} \hbar}(\hat{H}_{\downarrow\downarrow}\hat{\rho}_{\downarrow\uparrow}-\hat{\rho}_{\downarrow\uparrow}\hat{H}_{\uparrow\uparrow})- \mathbbm{i} \delta(\hat{\rho}_{\uparrow\uparrow}-\hat{\rho}_{\downarrow\downarrow})+\hat{\mathcal{L}}\hat{\rho}_{\downarrow\uparrow}\nonumber,\end{aligned}$$ where $$\begin{aligned} \hat{\mathcal{L}}\hat{\rho}_{\sigma\sigma'}&=&-\kappa(\hat{a}^\dagger \hat{a}\hat{\rho}_{\sigma\sigma'}+\hat{\rho}_{\sigma\sigma'} \hat{a}^\dagger \hat{a}-2\hat{a}\hat{\rho}_{\sigma\sigma'} \hat{a}^\dagger)\\ &-&2\kappa n(\hat{a}^\dagger \hat{a}\hat{\rho}_{\sigma\sigma'}+\hat{\rho}_{\sigma\sigma'} \hat{a}\hat{a}^\dagger-\hat{a}\hat{\rho}_{\sigma\sigma'} \hat{a}^\dagger-\hat{a}^\dagger\hat{\rho}_{\sigma\sigma'} \hat{a}).\nonumber\end{aligned}$$ At $t=0$ we assume a factorized initial state for the qubit-oscillator reduced density matrix $$\hat{\rho}_S(0)=\hat{\rho}_{\rm q}(0)\otimes\hat{\rho}_{\rm HO}(0),$$ and use the Wigner representation of the oscillator density matrix in phase-space [@Cahill69] $$\begin{aligned} \hat{\rho}_{\rm HO}(0)&=&\frac{1}{\pi}\int\!d^2\!\alpha\:\chi_0(\alpha)\hat{D}(-\alpha),\\ \hat{D}(-\alpha)&=&\exp\left(-\alpha\hat{a}^{\dagger}+\alpha^\hat{a}\right),\end{aligned}$$ where $\chi_0$ is the Fourier transform of the Wigner function. We assume the oscillator to be initially in a thermal state $$\begin{aligned} \chi_0(\alpha)&=&\frac{1}{4\pi}\exp\left(-\frac{\eta}{2}\|\alpha\|^2\right),\:\:\eta=1+2n(\Omega),\end{aligned}$$ where $n(\Omega)$ is the Bose function at bath temperature $T$. The qubit is assumed to be initially in the pure state $\|\Psi\rangle=q_{\uparrow}\|\uparrow\rangle+q_{\downarrow}\mathbbm{e}^{\mathbbm i\phi}\|\downarrow\rangle$ such that $$\begin{aligned} \hat{\rho}_{\rm q}(0)&=&\left(\begin{matrix}q_{\uparrow}^2&q_{\uparrow}q_{\downarrow}\mathbbm{e}^{-\mathbbm{i}\phi}\\ q_{\uparrow}q_{\downarrow}\mathbbm{e}^{\mathbbm{i}\phi}& q_{\downarrow}^2\end{matrix}\right).\end{aligned}$$ For the corresponding Wigner characteristic functions we obtain the following coupled partial differential equations: $$\begin{aligned} \dot{\chi}_{\sigma\sigma}&=&(\mathbbm{i}(r_{\sigma}\gamma+K)(\alpha+\alpha^)+\mathbbm{i} \Omega(\alpha\partial_{\alpha}-\alpha^\partial_{\alpha^})\nonumber\\ &+&\mathcal{D})\chi_{\sigma\sigma}-r_{\sigma}\mathbbm{i} {\delta}(\chi_{\downarrow\uparrow}-\chi_{\uparrow\downarrow})),\label{ugly_eqs}\\ \dot{\chi}_{\uparrow\downarrow}&=&(2\mathbbm{i} \gamma(\partial_{\alpha^}-\partial_{\alpha})+\mathbbm{i} \Omega(\alpha\partial_{\alpha}-\alpha^\partial_{\alpha^})-2\mathbbm{i} w\nonumber\\ &+&\mathbbm{i}K(\alpha+\alpha^)+\mathcal{D})\chi_{\uparrow\downarrow}-\mathbbm{i} {\delta}(\chi_{\downarrow\downarrow}-\chi_{\uparrow\uparrow})),\nonumber\\ \dot{\chi}_{\downarrow\uparrow}&=&(-2\mathbbm{i} \gamma(\partial_{\alpha^}-\partial_{\alpha})+\mathbbm{i} \Omega(\alpha\partial_{\alpha}-\alpha^\partial_{\alpha^})+2\mathbbm{i} w \nonumber\\ &+&\mathbbm{i}K(\alpha+\alpha^)+\mathcal{D})\chi_{\downarrow\uparrow}+\mathbbm{i} {\delta}(\chi_{\downarrow\downarrow}-\chi_{\uparrow\uparrow})),\nonumber\end{aligned}$$ where the differential operator $\mathcal{D}$ is given by $$\begin{aligned} \mathcal{D}&=&-\kappa(\alpha\partial_{\alpha}+\alpha^\partial_{\alpha^})-\eta\kappa \|\alpha \|^2.\end{aligned}$$ To solve these equations, we approximate the inhomogeneous parts, in the limit of short time $\tau$ and weak tunneling $\delta$, by $$\begin{aligned} \chi_{\sigma\sigma'}(t)\simeq\chi_{\sigma\sigma'}(0)+t\dot{\chi}_{\sigma\sigma'}(0),\:\:\:\sigma,\sigma'\in\{\uparrow,\downarrow\} .\label{important}\end{aligned}$$ For details on the solution see Appendix \[solution\_chi\]. The post-interaction phase -------------------------- The state prepared by the interaction with the qubit at $t=\tau$, as the bias current pulse ends, is described by $$\begin{aligned} \hat\rho(\tau)&=&\sum_{\sigma,\sigma'\in\{\uparrow,\downarrow\}}\|\sigma\rangle\langle\sigma'\|\hat\rho_{\sigma\sigma'}(\tau)\otimes\hat\rho_B(0).\end{aligned}$$ Since the system Hamiltonian no longer contains any qubit-oscillator interaction, we can write the time evolution of this density matrix as follows $$\begin{aligned} \hat\rho(t)&=&\!\!\!\!\sum_{\sigma,\sigma'\in\{\uparrow,\downarrow\}}\!\!\!\! \hat U_{\rm q}(t)\|\sigma\rangle\langle\sigma'\|\hat U^{\dagger}_{\rm q}(t)\nonumber\\ &\cdot&\hat U_{{\rm HO}-B}(t)\hat\rho_{\sigma\sigma'}(\tau)\otimes\hat\rho_B(0)\hat U^{\dagger}_{{\rm HO}-B}(t),\label{ref6a}\end{aligned}$$ where the evolution operators are given by $$\begin{aligned} \hat U_{\rm q}(t)&=&\exp(-\mathbbm{i}(t-\tau)(\delta\hat\sigma_x+ w \hat\sigma_z)),\\ \hat U_{{\rm HO}-B}&=&\mathcal{T}\exp\left(\int_\tau^t\! dt'\:\frac{\hat H_B+\hat H_I+\hbar\Omega\hat a^\dagger\hat a)}{\mathbbm{i}\hbar}\right).\end{aligned}$$ In the reduced density matrix $$\begin{aligned} \hat \rho_S(t)&=&{\rm Tr}_B\hat\rho(t)=\!\!\!\!\sum_{\sigma,\sigma'\in\{\uparrow,\downarrow\}}\!\!\!\! \hat U_{\rm q}(t)\|\sigma\rangle\langle\sigma'\|\hat U^{\dagger}_{\rm q}(t)\\ &\cdot&{\rm Tr}_B\left\{\hat U_{{\rm HO}-B}(t)\hat\rho_{\sigma\sigma'}(\tau)\otimes\hat\rho_B(0)\hat U^{\dagger}_{{\rm HO}-B}(t)\right\},\nonumber\end{aligned}$$ we can treat the time evolution of the oscillator components by means of a master equation in the Born-Markov approximation and, in a similar manner to Eq. (\[me\_bm1\]), we obtain $$\begin{aligned} \dot{\hat{\rho}}_{\sigma\sigma'}(t)&=&-\mathbbm{i}\Omega[\hat a^\dagger \hat a,\hat{\rho}_{\sigma\sigma'}(t)]\label{ref6b}\\ &-&\frac{1}{\hbar^2}\int_0^\infty \!\!\!\!dt'{\rm Tr}_B \!\left[\hat{H}_{I},[\hat{H}_{I}(t,t'),\hat{\rho}_{\sigma\sigma'}(t)\otimes\hat{\rho}_B(0)]\right].\nonumber \end{aligned}$$ Using the Wigner representation $$\begin{aligned} \hat{\rho}_{\sigma\sigma'}(t)&=&\frac{1}{\pi}\int\!d^2\alpha\:\tilde\chi_{\sigma\sigma'}(\alpha,t)\hat{D}(-\alpha),\end{aligned}$$ we obtain the differential equation $$\begin{aligned} \dot{\tilde{\chi}}_{\sigma\sigma'}(\alpha, t)&=&(\mathbbm{i} \Omega(\alpha\partial_{\alpha}-\alpha^\partial_{\alpha^})+\mathcal{D})\tilde{\chi}_{\sigma\sigma'}(\alpha,t),\end{aligned}$$ with the initial condition prepared at the end of the interaction phase $$\begin{aligned} \tilde\chi_{\sigma\sigma'}(\alpha,\tau)=\chi_{\sigma\sigma'}(\alpha,\tau),\end{aligned}$$ and the analytic solution $$\begin{aligned} \tilde{\chi}_{\sigma\sigma'}(\alpha,t)&=&\tilde{\chi}_{\sigma\sigma'}(\alpha \mathbbm{e}^{-(t-\tau)(\kappa-\mathbbm{i} \Omega)},\tau)\nonumber\\ &\times&\exp\left(\frac{\eta}{2}\|\alpha\|^2(\mathbbm{e}^{-2(t-\tau)\kappa}-1)\right).\end{aligned}$$ The reduced density matrix in the post-interaction phase is given by $$\begin{aligned} \hat\rho_S(t)&=&\!\!\!\!\!\sum_{s,s'\in\{\uparrow,\downarrow\}}\!\!\!\! \|s\rangle\langle s'\|\frac{1}{\pi}\int d^2\alpha \chi_{ss'}(\alpha,t)\hat{D}(-\alpha),\end{aligned}$$ where $$\begin{aligned} \chi_{ss'}(\alpha,t)\!\!&=&\!\!\!\!\!\!\!\!\!\!\sum_{\sigma,\sigma'\in\{\uparrow,\downarrow\}}\!\!\!\!\!\!\langle s\|\hat U_{\rm q}(t)\|\sigma\rangle\langle \sigma'\|\hat U_{\rm q}^\dagger(t)\|s'\rangle\tilde\chi_{\sigma\sigma'}(\alpha,t)\label{rel1}.\end{aligned}$$ In the the post-interaction phase, the qubit and the oscillator are decoupled. The trace of the oscillator-bath part in Eq. (\[ref6a\]) is time independent, as one can see after a circular permutation of the involved operators. One finds that the qubit time evolution is given only by the unitary $\hat U_q$, and thus is independent of the oscillator. Physically, this means that in the post-interaction phase no further information about the qubit can be transferred to the oscillator-bath system, and thus the qubit suffers no further decoherence. Results ======= In this section we analyze the qubit decoherence and the evolution of its detector, the dissipative oscillator, during the entire measurement process. Qubit decoherence ----------------- During the interaction phase, $t\in(0,\tau)$, the qubit is in contact with an environment represented by the dissipative oscillator, and thus subject to decoherence. The qubit can be prepared in a well defined state by thermal relaxation or (if the temperature is too high) by measurement post-selection and conditional rotation by microwave pulses. We analyze the qubit relaxation described by $$\langle\hat{\sigma}_z\rangle(t)=4\pi(\chi_{\uparrow\uparrow}(0,t)-\chi_{\downarrow\downarrow}(0,t))$$ and from Eq. (\[chi\_d\]) we obtain the analytic result $$\begin{aligned} \langle\hat{\sigma}_z\rangle(t)&=&(q_{\uparrow}^2-q_{\downarrow}^2)(1- 2t^2\delta^2)\\ &+&4q_{\uparrow}q_{\downarrow}t\delta(t w \cos(\phi)+\sin(\phi))\nonumber. \end{aligned}$$ We observe that the above expression is identical with the expansion up to the second order in time of $\langle\hat{\sigma}_z\rangle(t)$ when the qubit evolves under the free Hamiltonian $\hat H_{\rm q}$ only. Thus, the evolution of $\langle\hat{\sigma}_z\rangle(t)$ in this short time expansion is indistinguishable from the free evolution of the unperturbed qubit. This can be understood as follows: the observable $\hat \sigma_z$ commutes with the environment coupling, but is not an integral of the free motion, as required for a QND measurement [@Braginsky95]. Thus, the perturbation of the measured observable comes only from the free evolution of the system. One can restrict this perturbation by reducing the time $\tau$ when it takes place. Fig. \[decoherence\] (a) shows the evolution of $\langle\hat{\sigma}_z\rangle(t)$ for a set of parameters closely related to a feasible experiment, see also Appendix \[parameters\]. The initial state chosen for panel (a) was $\|\uparrow\rangle$. ![ (a) Evolution of $\langle\hat \sigma_z\rangle$ with qubit initially in $\|\uparrow\rangle$ state. (b) Dephasing from the $1/\sqrt{2}(\|\uparrow\rangle+\|\downarrow\rangle)$ state for the time $\tau$ that the qubit is in contact with the oscillator. For both plots, the following parameters were used: $\Omega/(2\pi)=0.97$ GHz, $\kappa/\Omega=10^{-2}$, $ w =\Omega$, $\Omega\tau=1.83$, $\delta\tau=0.015$, $\gamma\tau=3$, $T=30$ mK. The assumed values of the circuit parameters are given in Appendix \[parameters\].[]{data-label="decoherence"}](decoherence.pdf){width="\columnwidth"} Furthermore, we analyze the qubit coherence $\langle\hat{\sigma}_x\rangle$ which is given by $$\begin{aligned} \langle\hat{\sigma}_x\rangle(t)&=&8\pi{\rm Re}\chi_{\uparrow\downarrow}(0,t),\end{aligned}$$ and can be evaluated from Eqs. (\[chi\_off1\],\[chi\_off2\]), where $\chi_{\uparrow\downarrow}^{\rm inh}(0,t)$ can be integrated numerically. We observe that, if the interaction time $\tau$ is long enough to allow the oscillator a full period evolution, one finds a revival in the qubit coherence at the end of this period. As the oscillator returns to (almost) its initial state, the information about the qubit is “erased" from the oscillator, as the oscillator states corresponding to $\|\uparrow\rangle$ and $\|\downarrow\rangle$ are no longer discernible. The height of the coherence revival peaks at $\Omega t=2\pi n$ decays in time as the information about the coupled qubit-oscillator system flows (irreversibly in this case) into the environment. The qubit dephasing for the same parameters of Appendix \[parameters\] is shown in Fig. \[decoherence\] (b). The appropriate initial state for this study is the equal superposition $(1/\sqrt{2})(\|\uparrow\rangle+\|\downarrow\rangle)$. We observe that the qubit appears completely dephased after the strong interaction with the damped oscillator, such that only a classical probability is imprinted onto the latter. In Fig. \[decoherence\] (a) we observe that the relaxation from the excited qubit state is very weak during the interaction time, as $\langle\hat\sigma_z\rangle$ differs at most by $10^{-3}$ from the initial value of 1. This combination of low coherence (b), indicating the fact that the information about the qubit has been imprinted onto the oscillator, and very low relaxation (a) demonstrates that the first step of the measurement protocol produces a good starting point for the second one, the oscillator readout. The negligible relaxation brings the qubit close to QND dynamics. We observe that the qubit coherence time is essentially dominated by the coupling between the qubit and its complex environment $\gamma^{-1}$ such that it is desirable to achieve $\gamma\tau\gg 1$. The relaxation of the qubit has been described in the first order in time, and essential to the almost-QND result is that $\tau\delta\ll1$. We note that the implied condition $\gamma\ll\delta$ contradicts none of our approximations, and can also be realized in experiment. Detector dynamics ----------------- In this section we study the evolution of the damped oscillator, which represents the detector. To achieve the strong qubit-oscillator coupling during the short interaction phase required to imprint the qubit state onto the oscillator, one needs a bias current pulse that approaches the critical current for the SQUID. Nonetheless, it is important that the SQUID does not switch out to the running state during the bias current pulse. For the parameters given in Appendix \[parameters\], we can evaluate the SQUID escape rate [@Martinis87] from the zero-voltage state during the bias current pulse in the regime of quantum assisted thermal activation $(k_BT\apprge\hbar\Omega)$ $$\begin{aligned} \Gamma_{\rm sw}&=&\frac{\sinh\left(\frac{\hbar\Omega}{2k_BT}\right)}{\sin\left(\frac{\hbar\Omega}{2k_BT}\right)}\frac{\Omega}{2\pi}\exp\left(\frac{-\Delta U}{k_BT}\right),\end{aligned}$$ where $\Delta U$ is the potential barrier. We obtain, for the worst case, $\Gamma_{\rm sw}\approx 3.6\cdot 10^{7}s^{-1}$ such that the escape time is much larger than the duration of the bias current pulse. The output of the detector is the time dependent voltage across the SQUID, which is proportional to the momentum of the oscillator. The probability distribution of momentum is given by $$\begin{aligned} P(p,\tau,t)&=&\mu\langle\delta(\hat{p}-p)\rangle\label{prob_dist}\\ &=& 2 \int d \alpha_x\sum_{\sigma\in\{\uparrow,\downarrow\}}\chi_{\sigma\sigma}(\alpha_x,t)\exp\left(\frac{\mathbbm{i} p \alpha_x}{\mu}\right),\nonumber\\ \mu&=&\sqrt{\frac{m\Omega\hbar}{2}},\:\:\alpha=\alpha_x+\mathbbm{i}\alpha_y, \end{aligned}$$ where, in the post-interaction phase $(t>\tau)$, $\chi_{\sigma\sigma}(\alpha_x,t)$ also depends on $\tau$ via its initial condition. The expectation values for the ${\rm n}^{th}$ moment of the oscillator momentum and position are then $$\begin{aligned} \langle\hat{p}^n\rangle(t)&=&\frac{4\pi\mu^n}{\mathbbm{i}^n}(-1)^n(\partial_{\alpha_x})^n\!\!\!\!\sum_{\sigma\in\{\uparrow,\downarrow\}}\!\!\!\chi_{\sigma\sigma}(\alpha_x,t)\|_{\alpha_x=0}\label{momentum},\\ \langle\hat{x}^n\rangle(t)&=&\!\!\left(\!\!\sqrt{\frac{\hbar}{2m\Omega}}\!\right)^n\!\!\frac{4\pi}{\mathbbm{i}^n}(\partial_{\alpha_y})^n\!\!\!\!\sum_{\sigma\in\{\uparrow,\downarrow\}}\!\!\!\chi_{\sigma\sigma}(\mathbbm{i}\alpha_y,t)\|_{\alpha_y=0}.\nonumber\end{aligned}$$ Furthermore, in the post-interaction phase we have, from Eq. (\[rel1\]), $$\begin{aligned} \!\!\!\sum_{\sigma\in\{\uparrow,\downarrow\}}\!\!\!\!\chi_{\sigma\sigma}&=&\!\!\!\! \!\!\!\!\sum_{s,s',\sigma\in\{\uparrow,\downarrow\}}\langle\sigma\|\hat U_{\rm q}(t)\|s\rangle\langle s'\|\hat U^\dagger_{\rm q}(t)\|\sigma\rangle\tilde\chi_{ss'}\nonumber\\ &=&\sum_{s,s'}\langle s'\|\hat U^\dagger_{\rm q}(t)\hat U_{\rm q}(t)\|s\rangle\tilde\chi_{ss'}=\sum_{s}\tilde\chi_{ss},\end{aligned}$$ which shows, as expected, that no measurement of the oscillator can provide information about the post-interaction evolution of the qubit, provided this evolution is unitary (i.e. the qubit is not being measured by something else). For the evaluation of both Eqs. (\[prob\_dist\], \[momentum\]) the $s$-integration in $\chi_{\sigma\sigma}^{\rm inh}$, Eq. (\[chi\_diag\]) should be evaluated last. Thus, one obtains an analytic (but rather long) expression for the expectation value of momentum, while for the probability density a numerical $s$-integration is required. Nevertheless, the components originating in $\chi_{\sigma\sigma}^{\rm hom}$ turn out to be dominant, and we give their analytic expressions in the following: $$\begin{aligned} &&\langle\hat p\rangle(\tau,t)=\langle\hat p\rangle_{\rm hom}(\tau,t) +\langle\hat p\rangle_{\rm inh}(\tau,t),\label{mom_expect}\\ &&\langle\hat p\rangle_{\rm hom}=\left(K+\gamma q_{\uparrow}^2-\gamma q_{\downarrow}^2\right) \mu\mathbbm{e}^{-(t-\tau)\kappa} \nonumber\\ &\cdot&\left(\mathbbm{e}^{-(t-\tau)\mathbbm{i} \Omega}\frac{1-\mathbbm{e}^{-\tau (\kappa +\mathbbm{i} \Omega)}}{-\kappa -\mathbbm{i} \Omega } +\mathbbm{e}^{(t-\tau)\mathbbm{i} \Omega}\frac{1-\mathbbm{e}^{-\tau(\kappa-\mathbbm{i} \Omega)}}{-\kappa +\mathbbm{i}\Omega }\right)\nonumber,\end{aligned}$$ The explicit form of the probability distribution of momentum, Eq.(\[prob\_dist\]), is given by $$\begin{aligned} P(p,\tau,t)&=&P_{\rm hom}(p,\tau,t)+P_{\rm inh}(p,\tau,t),\label{prob_dist_hom}\end{aligned}$$ where $$\begin{aligned} P_{\rm hom}(p,\tau,t)= \sum_\sigma\frac{\|\langle\sigma\|\Psi\rangle\|^2}{\sqrt{2\pi\eta}} \exp\left(\!\!\frac{\mathbbm{i}p}{\sqrt{2\eta}\mu}-\mathbbm{i}\frac{K+r_\sigma \gamma}{\sqrt{2\eta}} \mathbbm{e}^{(\tau-t)\kappa} \right. &&\nonumber\\ \cdot\!\!\left.\left(\mathbbm{e}^{-(t-\tau)\mathbbm{i} \Omega}\frac{1-\mathbbm{e}^{-\tau (\kappa +\mathbbm{i} \Omega)}}{-\kappa -\mathbbm{i} \Omega } +\mathbbm{e}^{(t-\tau)\mathbbm{i} \Omega}\frac{1-\mathbbm{e}^{-\tau(\kappa-\mathbbm{i} \Omega)}}{-\kappa +\mathbbm{i}\Omega }\right) \!\!\right)^2.&&\end{aligned}$$ The results above refer to the post-interaction phase $t>\tau$. For the interaction phase, $t\in(0,\tau)$, the probability distribution of momentum is given by $P(p,t,t)$ in Eq. (\[prob\_dist\_hom\]) and the expectation value of momentum by $\langle \hat p\rangle(t,t)$ in Eq. (\[mom\_expect\]), i.e. by replacing $\tau$ by $t$. The expectation value of momentum $\langle \hat p\rangle(\tau,t)$ in the post-interaction phase contains information about the qubit initial state. We observe that the momentum oscillations corresponding to the two different initial qubit states $\|\uparrow\rangle$ and $\|\downarrow\rangle$ for $t>\tau$ are in phase. Disregarding the inhomogeneous contributions, which are relatively small in the limit of small $\tau\delta$, the envelope of the homogeneous part is given by $$\begin{aligned} \mathcal{A}(q_{\uparrow},q_{\downarrow})&=&\frac{2\mathbbm{e}^{-(t-\tau) \kappa }(K+\gamma q_{\uparrow}^2-\gamma q_{\downarrow}^2) \mu }{\sqrt{\kappa ^2+\Omega ^2}}\nonumber\\ &\cdot&\sqrt{-2 \mathbbm{e}^{-\kappa \tau } \cos (\tau \Omega )+\mathbbm{e}^{-2 \kappa \tau }+1}.\label{amplitude} \end{aligned}$$ Fig. \[spiral\] illustrates the phase-space trajectories of the oscillator corresponding to the qubit being in either the $\|\uparrow\rangle$ or $\|\downarrow\rangle$ state. During the interaction phase the system moves away from the origin. After switching off the interaction, the trajectories spiral back towards the origin, without crossing. For $K=0$ the trajectories are symmetric with respect to the origin, while $K\not=0$ introduces an asymmetry. We note that the artificial situation $K=0$ includes only the bare oscillator response for the different qubit states. This situation has been introduced in order to more easily illustrate the difference between the two oscillations. ![(Color online). Phase space representation of the oscillator trajectories $(\langle \hat x\rangle(t),\langle \hat p\rangle(t),t)$ corresponding to the two qubit states $\|\downarrow\rangle$ (dashed, red) and $\|\uparrow\rangle$ (continuous, green) for the parameters given in Appendix \[parameters\], an oscillator quality factor of $10$, with $K=0$ (a) and $K\not =0$ (b). Projections on the $(x,p)$, $(x,t)$ and $(p,t)$ planes are included. Both trajectories start at the origin and move away from it under the influence of the interaction with the qubit. At the point marked with $\bullet$ the interaction is switched off, and the system evolves freely spiraling around the origin. The trajectories circle around each other without crossing. []{data-label="spiral"}](phase_sp.pdf){width="48.00000%"} ![(Color online) Probability distribution of output voltage (density plot, dark color indicates high and white low density) and expectation value of momentum for the two qubit states $\|\downarrow\rangle$ (dashed, red) and $\|\uparrow\rangle$ (continuous, green). Here $\Omega/(2\pi)=0.97$ GHz, $\Omega/\kappa=20$, $ w =\Omega$, $\Omega\tau=1.83$, $\delta\tau=0.015$, $\gamma\tau=3$, $T=30$ mK. The assumed values of the circuit parameters are given in Appendix \[parameters\].[]{data-label="detection"}](voltage.pdf){width="\columnwidth"} Fig. \[detection\] shows the output of the detector for the two qubit states $\|\downarrow\rangle$ and $\|\uparrow\rangle$. The standard condition for the possibility of single-shot readout, i. e., the maximal separation of the two peaks corresponding to different qubit states in the probability distribution Eq. (\[prob\_dist\]) should be larger than the peak width, is given by $$\varepsilon\approx\frac{\|\mathcal{A}(1,0)-\mathcal{A}(0,1)\|}{3\mu\sqrt{\eta}}>1,\label{single_shot}$$ where the envelope (\[amplitude\]) has been evaluated at $t=\tau$. We note that $q_{\uparrow}$ and $q_{\downarrow}$ are continuous variables with values between $0$ and $1$ and the condition presented above takes into account the extremal case of the difference between the states $\|\uparrow\rangle$ and $\|\downarrow\rangle$. The result is independent of $K$. For the parameters of Fig. \[detection\] we have $\varepsilon\approx2.5$. Practical implementation {#implementation} ======================== A possible measurement protocol involves discriminating the amplitudes of the ringdown oscillations corresponding to different qubit states. As demonstrated by Eq. (\[amplitude\]), the amplitude difference is independent of $K$. This discrimination could be performed more accurately with an interferometric technique, where ringdown oscillations from a second, reference SQUID oscillator that is not coupled to the qubit are combined with those from the original SQUID oscillator.The reference SQUID is biased such that it undergoes ringdown oscillations with the same phase and amplitude as those of the measurement SQUID oscillator for one of the two qubit states. In this case, the resultant signal after the subtraction would be exactly zero for perfect cancellation when the qubit state causes the two SQUID oscillators to have identical ringdown signals. A residual ringdown oscillation would be produced for the other qubit state. This scheme requires that the two SQUIDs receive an identical kick and begin their ringdown oscillations at the same time. This can be achieved by splitting the bias current pulse signal along two separate lines, one going to each SQUID, as shown in Fig. \[implement\], where the layout is such that the reference SQUID has a vanishing coupling to the qubit. ![Circuit diagram for SQUID oscillator and qubit, along with reference SQUID oscillator, dual-input gradiometer microstrip amplifier and a cryogenic High Electron Mobility Transistor (HEMT). Dashed boxes indicate different chips and/or different temperatures.[]{data-label="implement"}](implement.pdf){width="\columnwidth"} Fig. \[detection2\] shows the total signal, i.e. the difference of the ringdown oscillations from the measurement and reference SQUIDs for the two qubit states. We have considered the case where the total flux bias for the reference SQUID is equal to the total flux bias for the measurement SQUID in the case where the qubit state is $\|\uparrow\rangle$. In this case the difference signal is smeared around $0$ for the qubit in state $\|\uparrow\rangle$. If the qubit is in the $\|\downarrow\rangle$ state, the output signal oscillates with an amplitude is given by the difference between the two ringdown oscillations in Fig. \[detection\]. ![(Color online) Probability distribution of output voltage (density plot, dark color indicates high and white low density) and expectation value of momentum for the two qubit states $\|\downarrow\rangle$ (dashed, red) and $\|\uparrow\rangle$ (continuous, green). Here the contribution of the reference SQUID has been introduced. Parameters: $\Omega/(2\pi)=0.97$ GHz, $\Omega/\kappa=20$, $ w =\Omega$, $\Omega\tau=1.83$, $\delta\tau=0.015$, $\gamma\tau=3$, $T=30$ mK. The assumed values of the circuit parameters are given in Appendix \[parameters\].\[detection2\]](voltage2.pdf){width="\columnwidth"} The subtraction of the two ringdown signals can be achieved by using a microstrip SQUID amplifier arranged as a gradiometer with two separate microstrip inputs with their senses indicated in Fig. \[implement\] [@Mueck08]. The microstrip SQUID amplifier consists of a dc SQUID with a multi-turn superconducting input coil above a conventional SQUID washer, where the signal is connected between one side of the input coil and ground and the other end of the input coil is left open. Input signals near the stripline resonance frequency, related to the total length of the input coil, typically of the order of $1$ GHz, couple strongly to the SQUID loop and the SQUID produces an output signal with a gain of $\sim 10-20$ dB [@Mueck03]. A gradiometer microstrip SQUID amplifier for amplifying the difference between two separate signals near the stripline resonance can be produced as a straightforward extension from previous microstrip SQUID layouts by using a SQUID geometry with two loops and a separate stripline coil coupled to each of the loops, with one signal input connected to each stripline [@Mueck08]. With no crosstalk between the two inputs, the circulating currents in the two loops of the SQUID amplifier cancel out when the input signals are identical, resulting in a vanishing output signal. Thus, with the arrangement in Fig. \[implement\], the microstrip SQUID amplifier produces the difference between the two oscillator ringdowns. Of course, in any practical gradiometer, there will be non-zero crosstalk, where a signal at one input induces circulating currents in the other loop of the SQUID amplifier. However, for reasonable layouts of the device, this crosstalk could be kept at the $1$% level, thus setting a limit on the fidelity of the subtraction [@Mueck08]. Based on the calculated difference signals for the ringdown oscillations in the two qubit states from Fig. \[detection2\], one must be able to discriminate the oscillations for the $\|\downarrow\rangle$ qubit state from the non-oscillatory signal for the $\|\uparrow\rangle$ state. Thus one needs to resolve a $\sim 1$ GHz signal with an amplitude of $\sim 0.5$ $\mu$V in a $\sim 100$ MHz bandwidth, i.e. before the ringdown is completed. Microstrip SQUID amplifiers operated at $20$ mK have achieved noise temperatures as low as $\sim 50$ mK [@Mueck01]. If we assume a conservative noise temperature estimate of $200$ mK for our gradiometer microstrip SQUID amplifier, this would correspond to a noise of $250$ nV in the $100$ MHz bandwidth referred back to the SQUID oscillators. Thus, it should be possible to discriminate between the two possible output signals corresponding to the two qubit states in a single shot. In the non-ideal case, the noise of the reference SQUID increases the broadening of the curves in Fig. \[detection2\] such that the single shot condition (\[single\_shot\]) must accommodate another width $\eta$. Still, at the parameters used in Fig. \[detection2\], this condition will still hold. Conclusion ========== We have demonstrated that a non-QND Hamiltonian can induce a close to QND backaction on the qubit, despite arbitrarily strong interaction with the environment, provided that the interaction time is very short, i.e. the measurement is quasi-instantaneous. The relaxation of the qubit has been described in the first order in time and, essential to the almost-QND results presented above, is that $\tau\delta\ll1$. We observe that the measurement time, i.e., the time needed to reduce the qubit density matrix to a classical mixture is essentially dominated by the coupling between the qubit and its complex environment $\gamma^{-1}$ such that it is desirable to achieve $\gamma\tau\gg 1$. The readout time for the oscillator is restricted only by the ring-down of the two possible oscillations of momentum, i.e. $\kappa^{-1}$. The amplitude of these oscillations is proportional to $\gamma$, which again stresses the usefulness of a strong qubit-oscillator coupling. If the two peaks in $P(p,\tau,t)$ become separated by significantly more than their widths, single shot measurement may become possible. The method presented above has the advantage of a very short interaction between the qubit and its environment, compared to e.g. the dispersive readout of Ref. [@Lupascu04], and results in a QND-type of readout, without the requirement of strong, continuous AC driving of e.g. Ref. [@Lupascu07] which may induce spurious qubit relaxation. As a figure of merit we consider the QND fidelity in Ref. [@ralph06]. For the parameters used in Fig. \[decoherence\] and an initial qubit state $\Psi=1/\sqrt{2}(\|\uparrow\rangle+\|\downarrow\rangle)$, our scheme achieves at the end of the post interaction phase a QND fidelity of $99.92\%$. Furthermore, if the aim is to apply the idea of a short interaction with an intermediate system, dispersive measurement, with all its potential advantages, may be difficult due to the continuous driving which implies continuous interaction between the qubit and its environment. Acknowledgment ============== We acknowledge useful discussions with Michael Mück. This work was supported by DFG through SFB 631, by NSERC discovery grants, by QuanumWorks and by EU through EuroSQIP. Solution for the Wigner characteristic functions {#solution_chi} ================================================ In this section we solve Eqs. (\[ugly\_eqs\]) using the approximation (\[important\]). ### The diagonal density matrix elements We solve the diagonal equations needed for evaluation of expectation values such as $\langle \hat p\rangle(t)$, which characterize the output of the detector: $$\begin{aligned} \dot{\chi}_{\sigma\sigma}&=&(\mathbbm{i}(r_{\sigma}\gamma+K)(\alpha+\alpha^) +\mathbbm{i}\Omega(\alpha\partial_{\alpha}-\alpha^\partial_{\alpha^})\nonumber\\ &+&\mathcal{D})\chi_{\sigma\sigma}-r_\sigma\mathbbm{i} {\delta}\chi_0(\alpha)F(\alpha,t),\label{eq:1}\end{aligned}$$ where $$\begin{aligned} F(\alpha,t)&=&2q_{\uparrow}q_{\downarrow}\sin(\phi)(\mathbbm{i}-K(\alpha+\alpha^)t)-2\mathbbm{i}(q_{\uparrow}^2-q_{\downarrow}^2)\delta t\nonumber\\ &-&2\mathbbm{i}q_{\uparrow}q_{\downarrow}\cos(\phi)t(\eta\gamma(\alpha^-\alpha)-2 w ).\end{aligned}$$ We perform a variable transformation in order to remove the first order derivatives in Eq. (\[eq:1\]) $$\begin{aligned} \alpha&=&z \mathbbm{e}^{s(\kappa-\mathbbm{i} \Omega)}, \:\:\: \alpha^=z^* \mathbbm{e}^{s(\kappa+\mathbbm{i} \Omega)}, \:\:\: t=s,\end{aligned}$$ and obtain $$\begin{aligned} \partial_s{\chi}_{\sigma\sigma}&=&(\mathbbm{i}(r_{\sigma} \gamma+K)\mathbbm{e}^{s\kappa}(z \mathbbm{e}^{-s\mathbbm{i} \Omega}+z^* \mathbbm{e}^{s\mathbbm{i} \Omega})\nonumber\\ &-&\eta\kappa \|z \|^2\mathbbm{e}^{2s\kappa})\chi_{\sigma\sigma}\nonumber\\ &-&r_\sigma\mathbbm{i} {\delta}\chi_0(z\mathbbm{e}^{s(\kappa-\mathbbm{i} \Omega)})F(z\mathbbm{e}^{s(\kappa-\mathbbm{i} \Omega)},s),\end{aligned}$$ which can be solved analytically, and transformed back to the initial variables $\alpha, t$. The solution reads $$\chi_{\sigma\sigma}(\alpha,t)=\frac{\|\langle\sigma\|\Psi\rangle\|^2}{4\pi}\chi_{\sigma\sigma}^{\rm hom}(\alpha,t)-\frac{\mathbbm{i} r_{\sigma} \delta}{4\pi}\chi_{\sigma\sigma}^{\rm inh}(\alpha,t),\label{chi_d}$$ where $$\begin{aligned} &\chi_{\sigma\sigma}^{\rm hom}&(\alpha,t)= \exp\left(-\frac{\|\alpha\|^2\eta}{2} +\mathbbm{i}(r_{\sigma}\gamma+K) \right.\\ &\cdot&\!\!\!\!\!\!\left. \left(\frac{\alpha(1-\mathbbm{e}^{-t(\kappa-\mathbbm{i} \Omega)})}{\kappa-\mathbbm{i}\Omega} +\frac{\alpha^(1-\mathbbm{e}^{-t(\kappa+\mathbbm{i} \Omega)})}{\kappa+\mathbbm{i}\Omega }\right)\right),\nonumber\end{aligned}$$ and $$\chi_{\sigma\sigma}^{\rm inh}(\alpha,t)=\int_0^t \!\!ds\chi_{\sigma\sigma}^{\rm hom}(\alpha,s)F\left(\alpha\mathbbm{e}^{-s(\kappa-\mathbbm{i}\Omega)},t-s\right)\label{chi_diag}.$$ ### The off-diagonal density matrix elements The method and approximations of the previous section can be used to solve the off-diagonal equations. From this solution we intend to extract information about the qubit coherence $\langle \hat \sigma_x\rangle(t)$. We start with $$\begin{aligned} \dot{\chi}_{\uparrow\downarrow}&=&(2\mathbbm{i} \gamma(\partial_{\alpha^}-\partial_\alpha)+\mathbbm{i} \Omega(\alpha\partial_{\alpha}-\alpha^\partial_{\alpha^})-2\mathbbm{i} w \nonumber\\ &+&\mathbbm{i}K(\alpha+\alpha^)+\mathcal{D})\chi_{\sigma\sigma}-\mathbbm{i} {\delta}\chi_0(\alpha)G(\alpha,t),\end{aligned}$$ where $$\begin{aligned} G(\alpha,t)&=&q_{\downarrow}^2-q_{\uparrow}^2-t\mathbbm{i}(\gamma-K(q_{\downarrow}^2-q_{\uparrow}^2))(\alpha+\alpha^)\nonumber\\ &-&4t \delta q_{\uparrow}q_{\downarrow}\sin(\phi).\end{aligned}$$ The variable transformation in this case originates from $$\begin{aligned} \partial_s\alpha&=&(-\mathbbm{i}\Omega+\kappa)\alpha+2\mathbbm{i}\gamma\nonumber,\\ \partial_s\alpha^&=&(\mathbbm{i}\Omega+\kappa)\alpha^-2\mathbbm{i}\gamma\nonumber,\end{aligned}$$ and reads $$\begin{aligned} \alpha&=&\frac{2\mathbbm{i}\gamma}{\kappa-\mathbbm{i}\Omega}\left(\mathbbm{e}^{s(\kappa-\mathbbm{i}\Omega)}-1\right)+z\mathbbm{e}^{s(\kappa-\mathbbm{i}\Omega)}\nonumber,\\ \alpha^&=&-\frac{2\mathbbm{i}\gamma}{\kappa-\mathbbm{i}\Omega}\left(\mathbbm{e}^{s(\kappa+\mathbbm{i}\Omega)}-1\right)+z^\mathbbm{e}^{s(\kappa+\mathbbm{i}\Omega)},\nonumber\\ t&=&s.\end{aligned}$$ We obtain $$\begin{aligned} \partial_s{\chi}_{\uparrow\downarrow}&=&(-2\mathbbm{i} w -\eta\kappa \alpha(z,s)\alpha^(z^,s)\nonumber\\ &+&\mathbbm{i}{K}(\alpha(z,s)+\alpha^(z^,s)))\chi_{\uparrow\downarrow}\nonumber\\ &-&\mathbbm{i} {\delta}\chi_0(\alpha(z,s))G(\alpha(z,s),s),\end{aligned}$$ which can be solved analytically, and transformed back to $\alpha, t$. The solution reads $$\chi_{\uparrow\downarrow}(\alpha,t)=\frac{q_{\uparrow}q_{\downarrow}\mathbbm{e}^{-\mathbbm{i}\phi}}{4\pi}\chi_{\uparrow\downarrow}^{\rm hom}(\alpha,t) -\frac{\mathbbm{i} \delta}{4\pi}\chi_{\uparrow\downarrow}^{\rm inh}(\alpha,t),$$ where $$\begin{aligned} &\chi_{\uparrow\downarrow}^{\rm hom}&\!\!\!\!\!(\alpha,t)= \exp\left(-\frac{\|\alpha\|^2}{2}\eta -2 \mathbbm{i} t w -\frac{4 t \gamma (\gamma \eta \kappa -\mathbbm{i} K \Omega )}{\kappa^2+\Omega ^2}\right. \nonumber\\ &+&\!\!\!\!\!\frac{4 \gamma(\gamma\eta(\kappa^2-\Omega ^2)-2 \mathbbm{i} K \kappa \Omega)}{(\kappa^2+\Omega ^2)^2}\label{chi_off1}\\ &+&\!\!\!\!\!\frac{K +\gamma \eta }{\kappa+\mathbbm{i} \Omega }\left(\mathbbm{i}(1-\mathbbm{e}^{-t (\kappa +\mathbbm{i} \Omega )}) \alpha^* -\frac{2 \mathbbm{e}^{-t (\kappa +\mathbbm{i} \Omega )} \gamma}{\kappa+\mathbbm{i}\Omega}\right)\nonumber\\ &+&\!\!\!\!\!\left. \frac{K -\gamma \eta}{\kappa-\mathbbm{i} \Omega}\left(\mathbbm{i} (1-\mathbbm{e}^{-t(\kappa-\mathbbm{i}\Omega)})\alpha +\frac{2 \mathbbm{e}^{-t(\kappa-\mathbbm{i}\Omega)} \gamma }{\kappa-\mathbbm{i}\Omega}\right)\right), \nonumber\end{aligned}$$ and $$\begin{aligned} &\chi_{\uparrow\downarrow}^{\rm inh}&\!\!\!\!\!(\alpha,t)= \int_0^t ds \chi_{\uparrow\downarrow}^{\rm hom}(\alpha,s)\label{chi_off2}\\ &&G\left(\mathbbm{e}^{-s(\kappa-\mathbbm{i} \Omega) } \alpha +\frac{2\left(1-\mathbbm{e}^{-s(\kappa-\mathbbm{i} \Omega)}\right) \gamma }{\mathbbm{i}\kappa +\Omega },t-s\right)\nonumber.\end{aligned}$$ From the density matrix calculated above we can extract information about the qubit relaxation and dephasing during the short interaction with the dissipative oscillator. Conversion to circuit parameters {#parameters} ================================ In the following we give a recipe [@Tian02] to obtain the parameters entering the calculation of this paper from the circuit components $$\begin{aligned} \Omega=\sqrt{\frac{2\pi I_c^{\rm eff}}{C_S\Phi_0}}\left(1-\left(\frac{I_B}{I_{\rm c}^{\rm eff}}\right)^2\right)^{\frac{1}{4}}, && m=\left(\frac{\Phi_0}{2\pi}\right)^2C_S, \\ \gamma =-\frac{M_{qS}I_{\rm q}I_B\tan\phi_m^0}{4\mu},&& \kappa=\frac{1}{2RC_S},\\ \tan\phi_m^0=\frac{I_B}{\sqrt{{I_{\rm c}^{\rm eff}}^2-I_B^2}},&&K=\frac{I_B}{2e}\sqrt{\frac{\hbar}{2 m \Omega}},\end{aligned}$$ where $\Phi_0=h/2e$ is the magnetic flux quantum for a superconductor, $M_{qS}$ is the qubit-SQUID mutual inductance, $I_{\rm c}^{\rm eff}$ is the effective critical current of the SQUID at the particular flux bias, $I_B$ is the amplitude of the dc bias pulse applied to the SQUID, $C_S$ the SQUID shunt capacitance, $R$ the internal resistance of the measurement circuitry, and $I_{\rm q}$ is the circulating current of the localized states of the qubit. The expression for $K$ is derived in the limit of a small geometric inductance, low critical current and large shunt capacitor where one can approximate the SQUID dynamics as that of a single Josephson junction with a variable critical current. The momentum of the oscillator $p$ and the voltage across the SQUID are related by $$\begin{aligned} V&=&\frac{ep}{C_S\hbar},\end{aligned}$$ where $e$ is the electron charge. The parameters used to generate Figs. \[decoherence\], \[spiral\], \[detection\], \[detection2\] are $$\begin{aligned} I_{\rm c}^{\rm eff}=0.5\cdot 10^{-6} {\rm A},&& I_B=0.87 I_{\rm c}^{\rm eff},\\ C_S=2\cdot 10^{-11}{\rm F},&& M_{qS}=100\cdot 10^{-12} {\rm H},\\ I_{\rm q}=438\cdot 10^{-9}{\rm A}, &&\tau=0.3\cdot 10^{-9}{\rm s},\\ \delta/(2\pi)&=&0.8\cdot 10^{7}{\rm Hz}.\end{aligned}$$ [10]{} V. Braginsky, F. Y. Khalili, and K. Thorne, [Quantum Measurement]{}(Cambridge University Press, Cambridge, 1995). C. M. Caves, K. S. Thorne, R. W. P. Drever, V. D. Sandberg, and M. Zimmermann, Rev. Mod. Phys. [52]{}, 341 (1980). F. Wilhelm, U. Hartmann, M. Storcz, and M. Geller, in [Quantum computing with superconductors II: Decoherence]{}, [NATO Science Series II: Mathematics, Physics and Chemistry]{}, edited by M. E. Flatte and I. Tifrea (Springer, Dordrecht, 2007). U. Weiss, [Quantum Dissipative Systems]{}, No. 10 in [Series in modern condensed matter physics]{}, 2 ed. (World Scientific, Singapore, 1999). M. Keil and H. Schoeller, Phys. Rev. B [63]{}, 180302 (2001). G. Lindblad, Commun. Math. Phys. [48]{}, 119 (1976). C. Cohen-Tannoudji, B. Diu, and F. Laloë, [Quantum Mechanics]{} (Wiley Interscience, Weinheim, 1992). J. von Neumann, [Mathematical Foundations of Quantum Mechanics]{} (Princeton University Press, ADDRESS, 1955). L. Bulaevskii, M. Hru ška, A. Shnirman, D. Smith, and Y. Makhlin, Phys. Rev. Lett. [92]{}, 177001 (2004). D. V. Averin, Phys. Rev. Lett. [88]{}, 207901 (2002). A. N. Jordan and M. Buttiker, Phys. Rev. B [71]{}, 125333 (2005). A. N. Jordan and A. N. Korotkov, Phys. Rev. B [74]{}, 085307 (2006). G. J. Milburn and D. F. Walls, Phys. Rev. A [28]{}, 2065 (1983). B. C. Sanders and G. J. Milburn, Phys. Rev. A [39]{}, 694 (1989). M. Brune, S. Haroche, V. Lefevre, J. M. Raimond, and N. Zagury, Phys. Rev. Lett. [65]{}, 976 (1990). N. Boulant, G. Ithier, P. Meeson, F. Nguyen, D. Vion, D. Esteve, I. Siddiqi, R. Vijay, C. Rigetti, F. Pierre, and M. Devoret, Phys. Rev. B [76]{}, 014525 (2007). M. Boissonneault, J. Gambetta, and A. Blais, Phys. Rev. A [77]{}, 012112 (2008). Y. Makhlin, G. Schön, and A. Shnirman, Rev. Mod. Phys. [73]{}, 357 (2001). M. Devoret, A. Wallraff, and J. Martinis, cond-mat/0411174 (unpublished). M. Geller, E. Pritchett, A. Sornborger, and F. Wilhelm, in [Quantum computing with superconductors I: Architectures]{}, [NATO Science Series II: Mathematics, Physics and Chemistry]{}, edited by M. E. Flatte and I. Tifrea (Springer, Dordrecht, 2007). J. Clarke, F.K. Wilhelm, Nature [453]{}, 1031 (2008). C. van der Wal, A. ter Haar, F. Wilhelm, R. Schouten, C. Harmans, T. Orlando, S. Lloyd, and J. Mooij, Science [290]{}, 773 (2000). C. van der Wal, F. Wilhelm, C. Harmans, and J. Mooij, Eur. Phys. J. B [31]{}, 111 (2003). D. Vion, A. Aassime, A. Cottet, P. Joyez, H. Pothier, C. Urbina, D. Esteve, and M. Devoret, Science [296]{}, 886 (2002). J. M. Martinis, S. Nam, J. Aumentado, and C. Urbina, Phys. Rev. Lett. [89]{}, 117901 (2002). M. Steffen, M. Ansmann, R. McDermott, N. Katz, R. C. Bialczak, E. Lucero, M. Neeley, E. M. Weig, A. N. Cleland, and J. M. Martinis, Phys. Rev. Lett. [97]{}, 050502 (2006). F. K. Wilhelm, Phys. Rev. B [68]{}, 060503 (2003). A. Lupascu, C. Verwijs, R. N. Schouten, C. Harmans, and J. E. Mooij, Phys. Rev. Lett. [93]{}, 177006 (2004). A. Lupascu, S. Saito, T. Picot, P. C. de Groot, C. J. P. M. Harmans, and J. E. Mooij, Nat. Phys. [3]{}, 119 (2007). J. Lee, W. Oliver, T. Orlando, and K. Berggren, IEEE Trans. Appl. Superc. [ 15]{}, 841 (2005). A. Blais, R.-S. Huang, A. Wallraff, S. Girvin, and R. Schoelkopf, Phys. Rev. A [69]{}, 062320 (2004). D. Schuster, A. Wallraff, A. Blais, L. Frunzio, R. Huang, J. Majer, S. Girvin, and R. Schoelkopf, Phys. Rev. Lett. [94]{}, 123602 (2005). A. Wallraff, D. I. Schuster, A. Blais, L. Frunzio, J. Majer, M. H. Devoret, S. M. Girvin, and R. J. Schoelkopf, Phys. Rev. Lett. [95]{}, 060501 (2005). M. Metcalfe, E. Boaknin, V. Manucharyan, R. Vijay, I. Siddiqi, C. Rigetti, L. Frunzio, R. J. Schoelkopf, and M. H. Devoret, Phys. Rev. B [76]{}, 174516 (2007). P. Lougovski, H. Walther, and E. Solano, Eur. Phys. J. D [38]{}, 423 (2006). T. Bastin, J. Zanthier, and E. Solano, J. Phys. B. [39]{}, 685 (2006). M. F. Santos, G. Giedke, and E. Solano, Phys. Rev. Lett. [98]{}, 020401 (2007). M. J. Storcz, M. Mariantoni, H. Christ, A. Emmert, A. Marx, W. D. Oliver, R. Gross, F. K. Wilhelm, and E. Solano, cond-mat/0612226 (unpublished). I. Serban, E. Solano, and F. Wilhelm, Phys. Rev. B [76]{}, 104510 (2007). J. Mooij, T. Orlando, L. Levitov, L. Tian, C. van der Wal, and S. Lloyd, Science [285]{}, 1036 (1999). T. P. Orlando, J. E. Mooij, L. Tian, C. van der Wal, L. S. Levitov, S. Lloyd, and J. J. Mazo, Phys. Rev. B [60]{}, 15398 (1999). L. Tian, S. Lloyd, and T. Orlando, Phys. Rev. B [65]{}, 144516 (2002). K. Blum, [Density Matrix Theory and Applications]{} (Plenum, New York, 1996). V. Lefevre-Seguin, E. Turlot, C. Urbina, D. Esteve, and M. H. Devoret, Phys. Rev. B [46]{}, 5507 (1992). P. Bertet, I. Chiorescu, G. Burkard, K. Semba, C. J. P. M. Harmans, D. P. DiVincenzo, and J. E. Mooij, Phys. Rev. Lett. [95]{}, 257002 (2005). G. Ingold, in [Quantum transport and dissipation]{} (Wiley-VCH, Weinheim, 1998), Chap. Dissipative quantum systems. R. Alicki, D. Lidar, and P. Zanardi, Phys. Rev. A [73]{}, 052311 (2006). K. Cahill and R. Glauber, Phys. Rev. [177]{}, 1882 (1969). J. M. Martinis, M. H. Devoret, and J. Clarke, Phys. Rev. B [35]{}, 4682 (1987). M. Mück, Private communication, 2008. M. Muck, C. Welzel, and J. Clarke, Appl. Phys. Lett. [82]{}, 3266 (2003). M. Mück, J. Kycia, and J. Clarke, Appl. Phys. Lett. [78]{}, 967 (2001). T. C. Ralph, S. D. Bartlett, J. L. O’Brien, G. J. Pryde, H. M. Wiseman, Phys. Rev. A [73]{}, 012113 (2006).	{ "pile_set_name": "ArXiv" }
--- abstract: 'It is shown here that a strengthening of Wallach’s Unentangled Gleason Theorem can be obtained by applying results of the present authors on generalised Gleason theorems for quantum multi-measures arising from investigations of quantum decoherence functionals.' address: - 'Physics Division, Starlab nv/sa, Boulevard Saint Michel 47, B-1040 Brussels, Belgium' - 'Analysis and Combinatorics Research Centre, Mathematics, University of Reading, Reading RG6 6AX, England' author: - 'Oliver Rudolph [^1]' - 'J.D. Maitland Wright [^2]' title: On unentangled Gleason theorems for quantum information theory --- 17.7cm -0.4cm -0.8cm -0.2cm \[section\] \[principle\][Theorem]{} \[principle\][Proposition]{} \[principle\][Lemma]{} \[principle\][Corollary]{} \[principle\][Definition]{} \[principle\][Example]{} \[principle\][Remark]{} Introduction ============ In an interesting recent paper Wallach [@Wallach] obtained an unentangled Gleason theorem. His work was motivated by fundamental problems in quantum information theory, in particular: to what extent do local operations and measurements on multipartite quantum systems suffice to guarantee the validity of a theorem of Gleason-type and thus a Born-type rule for probabilities? His positive result is formulated in terms of partially defined frame functions, defined only on the unentangled states of a finite product of finite dimensional Hilbert spaces. We shall show that Wallach’s theorem, and also its generalisation to infinite dimensions, can readily be derived from results which were obtained by us in our investigations of generalised Gleason theorems for quantum bi-measures and multi-measures [@Wright95; @RudolphW97; @Wright98; @RudolphW98; @RudolphW99]. The physical motivation for our earlier work arose from the so-called histories approach to quantum mechanics [@Isham94; @IshamLS94]. Our more general approach relies on the generalised Gleason theorem obtained by Bunce and one of us [@BunceW92a; @BunceW92b; @BunceW92c]. Preliminaries ============= Throughout this note ${\cal H}$ is a Hilbert space, ${\cal S}({\cal H})$ is the set of unit vectors in ${\cal H}$, and the sets of projections, compact operators or bounded operators are denoted by ${\cal P}({\cal H}), {\cal K}({\cal H}),$ or ${\cal B}({\cal H})$ respectively.\ A quantum measure for ${\cal H}$ is a map $m: {\cal P}({\cal H}) \to {\Bbb C}$ such that $m(p + q) = m(p) + m(q)$ whenever $p$ and $q$ are orthogonal. If $m$ takes only positive values and $m(1) = 1$, then $m$ is a quantum probability measure. If, whenever $\{p_i \}_{i \in I}$ is a family of mutually orthogonal projections, $\sum_i m(p_i)$ is absolutely convergent and $m(\sum_i p_i) = \sum_i m(p_i)$, then $m$ is said to be completely additive.\ The essential content of Gleason’s original theorem [@Dvurecenskij93] is that if $m$ is a positive, completely additive quantum measure on ${\cal P}({\cal H})$, then it has a unique extension to a positive normal functional $\phi_m$ on ${\cal B}({\cal H})$, whenever the Hilbert space ${\cal H}$ is not of dimension 2. It then follows from routine functional analysis that there exists a unique positive, self-adjoint trace class operator $T$ on ${\cal H}$ such that $\phi_m(x) = {\mathrm{Tr}}(Tx)$ for each $x \in {\cal B}({\cal H})$, i.e., $m(p) = {\mathrm{Tr}}(Tp)$ for each $p \in {\cal P}({\cal H})$. As a tool to help him prove his theorem, Gleason introduced the notion of a frame function. A (positive) frame function for ${\cal H}$ is a function $f: {\cal S}({\cal H}) \to {\mathbb{R}}^+$ such that there exists a real number $w$ (the weight of $f$) such that, for any orthonormal basis of ${\cal H}$, $\{ x_i \}_{i \in I}$, $\sum_i f(x_i) = w$. There is a bijective correspondence between frame functions for $\mathcal{H}$ and positive, completely additive quantum measures on ${\cal P}({\cal H})$, see [@Dvurecenskij93]. Unentangled frame functions and quantum multi-measures ====================================================== Let ${\cal H}_1, \cdots, {\cal H}_n$ be Hilbert spaces. An unentangled element of ${\cal H}_1 \otimes \cdots \otimes {\cal H}_n$ is a vector which can be expressed in the form $x_1 \otimes \cdots \otimes x_n$. (Unentangled elements are sometimes referred to as simple tensors.) Let $\Sigma({\cal H}_1, \cdots, {\cal H}_n)$ be the set of all unentangled vectors of norm 1 in ${\cal H}_1 \otimes \cdots \otimes {\cal H}_n$. Every element in $\Sigma({\cal H}_1, \cdots, {\cal H}_n)$ can be expressed as a tensor product of unit vectors in ${\cal H}_1, \cdots, {\cal H}_n$ respectively. Following Wallach [@Wallach], an unentangled frame function for ${\cal H}_1, \cdots, {\cal H}_n$ is a function $f: \Sigma({\cal H}_1, \cdots, {\cal H}_n) \to \mathbb{R}^+$ such that, for some positive real number $w$ (the weight of $f$) whenever $\{ \xi_i \}_{i \in I}$ is an orthonormal basis of ${\cal H}_1 \otimes \cdots \otimes {\cal H}_n$ with each $\xi_i \in \Sigma({\cal H}_1, \cdots, {\cal H}_n)$, then $\sum_i f(\xi_i) = w$. The physical idea behind this definition is that the elements of $\Sigma({\cal H}_1, \cdots, {\cal H}_n)$ represent the outcomes of elementary local operations or measurements.\ It turns out that unentangled frame functions have natural links with quantum multi-measures. For the purposes of this note we define a (positive) quantum multi-measure for ${\cal H}_1, \cdots, {\cal H}_n$ to be a function $m: {\cal P}({\cal H}_1) \times \cdots \times {\cal P}({\cal H}_n) \to {\Bbb R}^+$, such that $m$ is completely orthoadditive in each variable separately, see [@RudolphW98]. (Our results in [@RudolphW98] apply to more general, vector valued quantum multi-measures.) When $n = 2$, a multi-measure is called a bi-measure. These arise naturally in the study of quantum decoherence functionals [@IshamLS94; @Wright95; @RudolphW97; @Wright98; @RudolphW98; @RudolphW99]. \[l31\] Let ${\cal H}_1, \cdots, {\cal H}_n$ be Hilbert spaces, none of which is of dimension 2. Let $m$ be a (positive) quantum multi-measure for ${\cal H}_1, \cdots, {\cal H}_n$. Then there exists a unique bounded, multi-linear map $M : {\cal B}({\cal H}_1) \times \cdots \times {\cal B}({\cal H}_n) \to {\Bbb C}$, such that $$M(p_1,p_2, \cdots, p_n) = m(p_1,p_2, \cdots, p_n) \mbox{ for each } p_j \in {\cal B}({\cal H}_j).$$ Furthermore, given $r$, with $1 \leq r \leq n$ and assuming $n \geq 2$, for each positive $x_j \in {\cal B}({\cal H}_j)$, with $1 \leq j \leq n$ and $j \neq r$, the map $y \mapsto M(x_1, \cdots, x_{r-1},y,x_{r+1}, \cdots, x_n)$ is a positive normal functional on ${\cal B}({\cal H}_r)$. Proof: The existence and uniqueness of $M$ is a consequence of results obtained in [@RudolphW98]. Whenever $x$ is a positive operator in ${\mathcal{B}}({\mathcal{H}})$, there exists a sequence of commuting projections $\{ p_j \}_{j = 1, 2, \cdots}$ such that $x = \Vert x \Vert \sum_j \frac{1}{2^j} p_j$ (for a proof see, e.g., [@Pedersen79] page 27). This observation, together with the positivity of $m$, shows that if $x_r$ is positive for $r = 1,2, \cdots, n$, then $M(x_1, \cdots,x_n) \geq 0$. It now follows from the results of [@RudolphW98] that given $r$, with $1 \leq r \leq n$ and assuming $n \geq 2$, for each positive $x_j \in {\mathcal{B}}({\mathcal{H}}_j)$, with $1 \leq j \leq n$ and $j \neq r$, the map $y \mapsto M(x_1, \cdots, x_{r-1},y,x_{r+1}, \cdots, x_n)$ is a positive normal functional on ${\mathcal{B}}({\mathcal{H}}_r)$. $\Box$\ Let us recall that the algebraic tensor product ${\cal B}({\cal H}_1) \otimes_{\rm alg} \cdots \otimes_{\rm alg} {\cal B}({\cal H}_n)$ may be identified with the linear span of $\{ x_1 \otimes x_2 \otimes \cdots \otimes x_n: x_j \in {\cal B}({\cal H}_j) \}$ in (the von Neumann tensor product) ${\cal B}({\cal H}_1 \otimes \cdots \otimes {\cal H}_n) = {\cal B}({\cal H}_1) \otimes \cdots \otimes {\mathcal{B}}({\cal H}_n)$. Let $m$ and $M$ be as in Lemma \[l31\], then by the basic property of the algebraic tensor product, there exists a unique linear functional ${\mathfrak{M}}$ on ${\cal B}({\cal H}_1) \otimes_{\rm alg} \cdots \otimes_{\rm alg} {\cal B}({\cal H}_n)$ such that\ ${\mathfrak{M}}(x_1 \otimes x_2 \otimes \cdots \otimes x_n) = M(x_1,x_2, \cdots, x_n)$. Let ${\cal H}_1, \cdots, {\cal H}_n$ be finite dimensional Hilbert spaces, none of which has dimension 2. Let $m$ be a positive quantum multi-measure for ${\cal H}_1, \cdots, {\cal H}_n$. Then there exists an unentangled frame function $f$ for ${\cal H}_1, \cdots, {\cal H}_n$ such that, whenever $\nu_1 \otimes \nu_2 \otimes \cdots \otimes \nu_n$ is in $\Sigma({\cal H}_1, \cdots, {\cal H}_n)$ and $p_j$ is the projection of ${\cal H}_j$ onto the one-dimensional subspace generated by $\nu_j$, $$f(\nu_1 \otimes \nu_2 \otimes \cdots \otimes \nu_n) = m(p_1,p_2, \cdots, p_n).$$ Proof: Fix a unit vector $\nu_j$ in ${\cal H}_j$ for $j = 1,2, \cdots, n$. Then the projection from ${\cal H}_1 \otimes \cdots \otimes {\cal H}_n$ onto the subspace spanned by $\nu_1 \otimes \nu_2 \otimes \cdots \otimes \nu_n$ can be identified with the projection $p_1 \otimes p_2 \otimes \cdots \otimes p_n$ in ${\cal B}({\cal H}_1 \otimes \cdots \otimes {\cal H}_n) = {\cal B}({\cal H}_1) \otimes \cdots \otimes {\cal B}({\cal H}_n)$. Define $f(\nu_1 \otimes \nu_2 \otimes \cdots \otimes \nu_n)$ to be ${\mathfrak{M}}(p_1 \otimes p_2 \otimes \cdots \otimes p_n) = M(p_1,p_2, \cdots,p_n) = m(p_1,p_2, \cdots,p_n)$. $\Box$\ The following technical lemma allows us to associate a canonical multi-measure with each unentangled frame function. \[l33\] Let ${\cal H}_1, \cdots, {\cal H}_n$ be Hilbert spaces of arbitrary dimension and let $f: \Sigma({\cal H}_1, \cdots, {\cal H}_n) \to \mathbb{R}^+$ be an unentangled frame function. Then there is a (positive, completely additive) quantum multi-measure $m$ for ${\cal H}_1, \cdots, {\cal H}_n$ such that whenever $\nu_1 \otimes \nu_2 \otimes \cdots \otimes \nu_n$ is in $\Sigma({\cal H}_1, \cdots, {\cal H}_n)$ and $p_j$ is the projection of ${\cal H}_j$ onto the one-dimensional subspace generated by $\nu_j$, $$m(p_1,p_2, \cdots, p_n) = f(\nu_1 \otimes \nu_2 \otimes \cdots \otimes \nu_n).$$ Proof: To simplify our notation we shall prove this for $n = 2$, but the method is perfectly general. Let $e_1$ and $e_2$ be projections in ${\mathcal{P}}({\mathcal{H}}_1)$ and ${\mathcal{P}}({\mathcal{H}}_2)$, respectively. Let $E_1$ and $E_2$ be the subspaces of ${\mathcal{H}}_1$ and ${\mathcal{H}}_2$ which are the respective ranges of $e_1$ and $e_2$. Let $\{ \xi_j \}_{j \in J}$ and $\{ \psi_i \}_{i \in I}$ be orthonormal bases of $E_1$ and $E_2$, respectively. We wish to define $m(e_1,e_2)$ to be $$\sum_{j \in J} \sum_{i \in I} f(\xi_j \otimes \psi_i).$$ The only difficulty here is that we do not know that this number is independent of the choice of orthonormal bases for $E_1$ and $E_2$, respectively. To establish this we argue as follows. Let $w$ be the weight of $f$. Let $\{ \xi_j \}_{j \in J^\perp}$ and $\{ \psi_i \}_{i \in I^\perp}$ be orthonormal bases for $E_1^\perp$ and $E_2^\perp$, respectively. Then $\{ \xi_j \otimes \psi_i \}_{j \in J \cup J^\perp, i \in I \cup I^\perp}$ is an orthonormal basis for ${\mathcal{H}}_1 \otimes {\mathcal{H}}_2$. So $$\sum_{(j,i) \in J \times I} f(\xi_j \otimes \psi_i) + \sum_{(j,i) \in J^\perp \times (I \cup I^\perp)} f(\xi_j \otimes \psi_i) + \sum_{(j,i) \in (J \cup J^\perp) \times I^\perp} f(\xi_j \otimes \psi_i) = w.$$ Let $\{ \xi_j' \}_{j \in J}$ and $\{ \psi_i' \}_{i \in I}$ be orthonormal bases of $E_1$ and $E_2$, respectively. Then $$\sum_{(j,i) \in J \times I} f(\xi_j' \otimes \psi_i') + \sum_{(j,i) \in J^\perp \times (I \cup I^\perp)} f(\xi_j \otimes \psi_i) + \sum_{(j,i) \in (J \cup J^\perp) \times I^\perp} f(\xi_j \otimes \psi_i) = w.$$ Hence $$\begin{aligned} \sum_{(j,i) \in J \times I} f(\xi_j' \otimes \psi_i') & = & w - \sum_{(j,i) \in J^\perp \times (I \cup I^\perp)} f(\xi_j \otimes \psi_i) - \sum_{(j,i) \in (J \cup J^\perp) \times I^\perp} f(\xi_j \otimes \psi_i) \\ & = & \sum_{(j,i) \in J \times I} f(\xi_j \otimes \psi_i). \end{aligned}$$ So $m$ is well-defined. It is straightforward to verify that $m$ has all the required properties. $\Box$\ Remark: In the above argument we made essential use of the property that $f$ is an unentangled frame function. Suppose that we only knew that $f$ satisfied the weaker property: for some positive real number $w$ whenever $\{ \xi_i \}_{i \in I}$ is a product orthonormal basis of ${\cal H}_1 \otimes \cdots \otimes {\cal H}_n$, then $\sum_i f(\xi_i) = w$. Then the proof of the preceding lemma would break down. This throws fresh light on the counterexample constructed in Proposition 5 in [@Wallach].\ In our investigations on quantum decoherence functionals we were led to obtain results on generalised quantum bi-measures and multi-measures [@Wright95; @RudolphW97; @Wright98; @RudolphW98; @RudolphW99]. The statement of the next theorem is Wallach’s Theorem 1 [@Wallach]. Our proof shows that Wallach’s Theorem is a natural consequence of our earlier results on quantum multi-measures. \[w\] Let ${\cal H}_1, \cdots, {\cal H}_n$ be finite dimensional Hilbert spaces, each of dimension at least 3. Let $f: \Sigma({\cal H}_1, \cdots, {\cal H}_n) \to \mathbb{R}^+$ be an unentangled frame function. Then there exists a self-adjoint operator $T$ in ${\cal B}({\cal H}_1 \otimes \cdots \otimes {\cal H}_n)$ such that whenever $\nu_1 \otimes \nu_2 \otimes \cdots \otimes \nu_n$ is in $\Sigma({\cal H}_1, \cdots, {\cal H}_n)$ and $p_j$ is the projection of ${\cal H}_j$ onto the one-dimensional subspace generated by $\nu_j$, $$f(\nu_1 \otimes \nu_2 \otimes \cdots \otimes \nu_n) = {\mathrm{Tr}}((p_1 \otimes p_2 \otimes \cdots \otimes p_n)T).$$ Proof: Since each of the Hilbert spaces ${\cal H}_1, \cdots, {\cal H}_n$ is finite dimensional, ${\cal B}({\cal H}_1) \otimes \cdots \otimes {\cal B}( {\cal H}_n) = {\cal B}({\cal H}_1) \otimes_{\mathrm{alg}} \cdots \otimes_{\mathrm{alg}} {\cal B}({\cal H}_n)$. Let $m$ be the quantum multi-measure constructed from $f$ as in Lemma \[l33\]. Let $\mathfrak{M}$ be the linear functional on ${\cal B}({\cal H}_1) \otimes_{\mathrm{alg}} \cdots \otimes_{\mathrm{alg}} {\cal B}({\cal H}_n) = {\cal B}({\cal H}_1) \otimes \cdots \otimes {\cal B}({\cal H}_n) = {\cal B}({\cal H}_1 \otimes \cdots \otimes {\cal H}_n)$ such that ${\mathfrak{M}}(q_1 \otimes q_2 \otimes \cdots \otimes q_n) = M(q_1,q_2, \cdots, q_n) = m(q_1,q_2, \cdots, q_n)$ for each $q_j \in {\cal P}({\cal H}_j)$. Since $\mathfrak{M}$ is a linear functional on a finite dimensional space, it is bounded. Hence there is a unique bounded operator $T$ in ${\cal B}({\cal H}_1 \otimes \cdots \otimes {\cal H}_n)$ such that ${\mathfrak{M}}(x) = {\mathrm{Tr}}(xT)$ for all $x$. Thus $$\label{beepbeep} f(\nu_1 \otimes \nu_2 \otimes \cdots \otimes \nu_n) = {\mathfrak{M}}(p_1 \otimes p_2 \otimes \cdots \otimes p_n) = {\mathrm{Tr}}((p_1 \otimes p_2 \otimes \cdots \otimes p_n)T).$$ On taking complex conjugates of the Equation (\[beepbeep\]) we find that $T$ may be replaced by $T^$. So in (\[beepbeep\]) we may replace $T$ by $\frac{1}{2} (T+T^)$. Hence we may suppose in (\[beepbeep\]) that $T$ is self-adjoint. $\Box$\ The work of [@RudolphW98; @RudolphW99] shows that Wallach’s Theorem can be generalised to the situation where the Hilbert spaces are not required to be finite dimensional provided an appropriate boundedness condition is imposed. More precisely: Let ${\cal H}_1, \cdots, {\cal H}_n$ be Hilbert spaces, each of dimension at least 3.\ Let $f: \Sigma({\cal H}_1, \cdots, {\cal H}_n) \to \mathbb{R}^+$ be an unentangled frame function. Let ${\mathfrak{M}}$ be the associated linear functional on ${\cal B}({\cal H}_1) \otimes_{\mathrm{alg}} \cdots \otimes_{\mathrm{alg}} {\cal B}({\cal H}_n)$. If the restriction of $\mathfrak{M}$ to ${\cal K}({\cal H}_1) \otimes_{\mathrm{alg}} \cdots \otimes_{\mathrm{alg}} {\mathcal{K}}({\cal H}_n)$ is bounded, then there exists a unique bounded self-adjoint, trace class operator $T$ in ${\cal B}({\cal H}_1 \otimes \cdots \otimes {\cal H}_n)$ such that whenever $\nu_1 \otimes \nu_2 \otimes \cdots \otimes \nu_n$ is in $\Sigma({\cal H}_1, \cdots, {\cal H}_n)$ and $p_j$ is the projection of ${\cal H}_j$ onto the one-dimensional subspace generated by $\nu_j$, $$f(\nu_1 \otimes \nu_2 \otimes \cdots \otimes \nu_n) = {\mathrm{Tr}}((p_1 \otimes p_2 \otimes \cdots \otimes p_n)T).$$ \[t35\] Proof: Let ${\mathfrak{M}}_0$ be the restriction of ${\mathfrak{M}}$ to . By hypothesis ${\mathfrak{M}}_0$ is bounded and so has a unique bounded extension, also denoted by ${\mathfrak{M}}_0$, to . By standard functional analysis, there exists a trace class operator $T$ such that ${\mathfrak{M}}_0(z) = {\mathrm{Tr}}(zT)$ for each $z$ in . Since each one-dimensional projection in ${\mathcal{B}}({\mathcal{H}}_j)$ is in ${\mathcal{K}}({\mathcal{H}}_j)$, $$m(p_1,p_2, \cdots, p_n) = M(p_1,p_2, \cdots,p_n) = {\mathfrak{M}}_0(p_1 \otimes p_2 \otimes \cdots \otimes p_n) = {\mathrm{Tr}}((p_1 \otimes p_2 \otimes \cdots \otimes p_n)T).$$ So $$f(\nu_1 \otimes \nu_2 \otimes \cdots \otimes \nu_n) = {\mathrm{Tr}}((p_1 \otimes p_2 \otimes \cdots \otimes p_n)T) = \langle T (\nu_1 \otimes \nu_2 \otimes \cdots \otimes \nu_n), \nu_1 \otimes \nu_2 \otimes \cdots \otimes \nu_n \rangle,$$ where $\langle \cdot, \cdot \rangle$ denotes the inner product on ${\cal H}_1 \otimes \cdots \otimes {\cal H}_n.$ It now follows from Lemma 5.9 \[6\] that $T$ is unique. But, arguing as in the proof of Proposition \[w\], we can replace $T$ by $\frac{1}{2}(T+T^)$. So, by uniqueness, $T$ is self-adjoint. $\Box$\ Remark: When the spaces ${\cal H}_1, \cdots, {\cal H}_n$ are finite dimensional, then the boundedness condition of Theorem \[t35\] is automatically satisfied. So Wallach’s Theorem is a corollary of Theorem \[t35\] which, in turn, follows from the work of [@RudolphW98; @RudolphW99]. Moreover, it can be shown along the lines of [@RudolphW97] that, for $n = 2$, there exists a self-adjoint operator $T$, not necessarily of trace class, on ${\mathcal{H}} = {\mathcal{H}}_1 \otimes {\mathcal{H}}_2$, such that $${\mathfrak{M}}(p) = {\mathrm{Tr}}(Tp)$$ for all finite rank projections $p$ in ${\cal P}({\cal H})$ if, and only if, ${\mathfrak{M}}$ is bounded on one dimensional projection operators whose ranges are generated by vectors in the algebraic tensor product ${\mathcal{H}}_1 \otimes_{{\mathrm{alg}}} {\mathcal{H}}_2$. [99]{} , An unentangled Gleason theorem, preprint, quant-ph/0002058. J.D.M. Wright, The structure of decoherence functionals for von Neumann quantum histories,* J. Math. Phys. 36 (1995), 5409-5413. , On tracial operator representations of quantum decoherence functionals, J. Math. Phys. [38]{} (1997), 5643-5652. J.D.M. Wright, Decoherence functionals for von Neumann quantum histories: boundedness and countable additivity, Comm. Math. Phys. 191 (1998), 493-500. , The multi-form generalized Gleason theorem, Comm. Math. Phys. [198]{} (1998), 705-709. , Homogeneous decoherence functionals in standard and history quantum mechanics, Comm. Math. Phys. [204]{} (1999), 249-267. , Quantum temporal logic and decoherence functionals in the histories approach to generalized quantum theory, J. Math. Phys. 35 (1994), 2157-2185. , The classification of decoherence functionals: An analogue of Gleason’s theorem, J. Math. Phys. 35 (1994), 6360-6370. L.J. Bunce and J.D.M. Wright, The Mackey-Gleason problem, Bull. Amer. Math. Soc. [26]{} (1992), 288-293. L.J. Bunce and J.D.M. Wright, Complex measures on projections in von Neumann algebras, J. London Math. Soc. 46 (1992), 269-279. L.J. Bunce and J.D.M. Wright, The Mackey-Gleason problem for vector measures on projections on von Neumann algebras, J. London Math. Soc. [49]{} (1994), 133-149. A. Dvurečenskij, Gleason’s theorem and its applications, (Kluwer, Dordrecht, 1993). G.K. Pedersen, $C^$-algebras and their automorphism groups*, (Academic, London, 1979). [^1]: email: [email protected] [^2]: email: [email protected]	{ "pile_set_name": "ArXiv" }
--- abstract: 'Using the non-linear optimal velocity (OV) models as an example, we show that there exists an emergent intrinsic scale that characterises the interaction strength between multiple clusters appearing in the solutions of such models. The interaction characterises the dynamics of the localised quasisoliton structures given by the time derivative of the headways, and the intrinsic scale is analogous to the “charge" of the quasisolitons, leading to non-trivial cluster statistics from the random perturbations to the initial steady states of uniform headways. The cluster statistics depend both on the quasisoliton charge and the density of the traffic. The intrinsic scale is also related to an emergent quantity that gives the extremum headways in the cluster formation, as well as the coexistence curve separating the absolute stable phase from the metastable phase. The relationship is qualitatively universal for general optimal velocity models.' author: - 'Bo Yang$^1$, Xihua Xu$^{2,3}$, John Z.F. Pang$^{1}$, and Christopher Monterola$^1$' date: - - title: Cluster Statistics and Quasisoliton Dynamics in Microscopic Optimal Velocity Models --- Modelling traffic flow, especially in an attempt to understand the occurrence of the traffic jams[@treiberbook; @yukawa; @yukawa1; @rehborn; @rehborn1; @kerner1], has been a fascinating subject leading to interesting development in many related fields. Several common approaches in modelling the evolution of the traffic flow include the microscopic car-following model[@bando; @stepan; @werner], cellular automata[@schall; @nagel] and the macroscopic hydrodynamic model[@seibold; @kim]; more thorough reviews can be found in [@helbing; @dogbe; @as]. Most microscopic models involve anisotropic nearest neighbour interactions. One elegant class of models is the optimal velocity (OV) models, with an explicit optimal velocity function dependent on the relative distance between the car and the next one ahead, or the headway[@bando]. Extensions of such models include additional force terms so that the acceleration or deceleration of the cars leaving/entering jammed region is not too large[@Helbing_PRE98; @JiangR_PRE01; @PengGH_PhyA13; @GLW_PhyA08]. Other more realistic microscopic models include the intelligent driver models (IDM)[@helbingidm], Shamoto’s models[@shamoto] and various types of the sophisticated three-phase traffic models[@k1; @k2; @k3]. The relationships between these models are also explored in[@boyang]. Multiple preceding cars and even following cars are included to better model the driver decision-making process[@xue; @hasebe], and non-linear velocity difference effects are studied in[@xihua]. Controversies still remain on what aspects of real traffic dynamics can be captured by simple models like the optimal velocity model[@kerner2; @schreck; @helbing1]. Such models assume the existence of a fundamental diagram, thus all steady states have a unique relationship between the flow and density. This is in contrast with the fundamental assumptions of the three-phase traffic theories[@kernerbook], that a multitude of steady states with non-unqiue flow-density relationship exists in the “synchronised phase". While one does not expect such simple traffic models to capture all the empirical features of the congested traffic flow, these models offer a physically intuitive way to understand the formation of jams from the non-linear interactions between the system components, which are useful in designing intelligent mass transport systems[@hasebe] made of, for example, sensor equipped driverless cars. In addition, they have the potential to characterise a wide range of physical phenomena including the complex spatiotemporal traffic patterns, dynamics of (quasi-)one dimensional granular flows and the clustering of dissipative “granular gases"[@zanetti]. It is thus of great theoretical interest to study the universal behaviours of these models especially in the non-linear regime. In this paper, we do not concern ourselves with the capabilities of the models in capturing the empirical features of the traffic flow. Instead, we study the formal non-linear dynamics of the OV model class, especially focusing on the multi-cluster solutions. Using the original OV model as an example for its simplicity, we show that by properly non-dimensionalizing the model, the emergent symmetry of the cluster formation is rendered explicit, and the extremum headway of the clusters is an emergent quantity which gives the coexistence curve separating the absolutely stable and metastable phase of the model. Our numerical calculation shows that the probability distribution of cluster numbers depends both on an intrinsic scale of the model and the density of the traffic lane. This can be explained by the dynamics of the “quasisolitons" in the domain of headway velocity, which will be explained in details later. The strength of attraction between quasisolitons of opposite charges depends both on the intrinsic scale and the distance between them. The intrinsic scale is thus analogous to the charge of the quasisolitons. A general car-following model can be written as $$\begin{aligned} \label{general} \tau\dot v_n=-v_n+V\left(h_{n-i},\dot h_{n-i},\cdots,h_n,\dot h_n,\cdots,h_{n+j},\dot h_{n+j}\right)\end{aligned}$$ where the dot represents time derivative and $n\in \mathbb Z^+$ is the index of the cars; $v_n$ is the velocity of the $n^{\text{th}}$ car; $h_{n+i}$ is the distance between the $n^{\text{th}}$ car and the $(i+1)^{\text{th}}$ car in front of it, while $h_{n-i}$ is the distance between the $n^{\text{th}}$ car and the $i^{\text{th}}$ car behind it, which by convention is negative. The first viscosity term on the right models the increasing tendency for the driver to decelerate when the car travels faster, and $\tau$ is the reaction time for the driver to maintain the optimal velocity given by the second term on the right. In this work the higher derivatives are suppressed as we assume the reaction time is small. The periodic boundary conditions gives $v_{N+n}=v_n$, where $N$ is the total number of cars. For physically relevant cases the optimal velocity is non-linear: it is generally assumed that $V$ is monotonically increasing for all its arguments, and it is bounded from above and below by the maximum and minimum acceleration of the car. We will now proceed with the simplest case of the OV model, where the optimal velocity function only depends on a single headway and is given by $$\begin{aligned} V\left(h_n\right)=V_1+V_2\tanh\left(s_n\right), s_n=C_1(h_n-l)-C_2\label{vc}\end{aligned}$$ The physical significance of different parameters in Eq.(\[vc\]) can be found in[@bando; @Helbing_PRE98]. We can now rewrite Eq.(\[general\]) as $$\begin{aligned} \label{ovmm} \ddot s_n+\kappa_1\dot s_n=\kappa_2\left(\tanh s_{n+1}-\tanh s_n\right)\end{aligned}$$ where $\kappa_1=\tau^{-1},\kappa_2=\tau^{-1}C_1V_2$. By rescaling the time variable $t\rightarrow \kappa_2t/\kappa_1$, the only dimensionless parameter in Eq.(\[ovmm\]) is $\kappa=\kappa_1^2/\kappa_2$. This equation in general describes an array of particles moving in a viscous media with anisotropic non-linear nearest neighbour interaction. We will now focus on Eq.(\[ovmm\]), where $s_n$ is dimensionless. The change of variable in Eq.(\[vc\]) not only tells us seemingly different driving behaviors are actually equivalent within the model; it also makes the symmetry of ODE’s in Eq.(\[ovmm\]) explicit. While the physical headway $h_n$ has to be positive, there is no such constraint on $s_n$; one should note the average of $s_n$ over all cars is inversely proportional to the linear car density of the lane with a shift, according to Eq.(\[vc\]). Thus Eq.(\[vc\]-\[ovmm\]) completely define the physical model at hand, and mathematically Eq.(\[ovmm\]) alone is sufficient. We will first discuss the properties of the individual clusters appearing in the solutions of Eq.(\[ovmm\]). While many of these properties are known, here we derive them in the most general way. We also present the relations of the coexistence curve with the emergent extremal headways from the non-linear dynamics, which are not reported before and useful for numerical analysis. Linear analysis leads to a stable phase of $s_n=s_0$ against small perturbation above the spinodal line (or the neutral stability line) given by $$\begin{aligned} \label{linear} 2\text{sech}^2 s_0=\kappa.\end{aligned}$$ In the regime $\|s_0\|>s_{c1}=\|\text{sech}^{-1}\sqrt{\kappa/2}\|$, a small perturbation to a uniform headway $s_0$ with $s_n(t\rightarrow 0)=s_0+\delta s_n$ leads to $s_n(t\rightarrow\infty)=s_0$. Here we take $\sum_n\delta s_n=0$ for technical convenience. Thus a random small initial variation of the positions of the cars in a single lane would not lead to the development of clusters, or traffic jams, in this regime. Note Eq.(\[linear\]) is only exact in the limit when the perturbation goes to zero; close to the spinodal line, the uniform headway configuration is metastable, a large enough perturbation will also lead to the formation of clusters[@kerner]. We now show that the coexistence curve that separates the metastable phase and the absolutely stable phase can be numerically read off from the cluster formation alone. Firstly, in the regime $\|s_0\|<s_{c1}$, it is well known that small perturbations will grow in time with the formation of clusters, as shown in Fig.(\[singlepeak\]), where a random initial condition settles into a configuration with the majority number of cars having two extremum headways given by $\pm s_{c2}$. As smaller $s_n$ implies higher physical car density, cars with headway $-s_{c2}$ form clusters or jams of very high density with minimal velocity, while cars with headway $s_{c2}$ moves with very high velocity, forming anti-clusters. Interestingly like $s_{c1}$, the numerical value of $s_{c2}$ only depends on $\kappa$ but not on $s_0$, even for $s_0$ in the metastable regime. 1= ![The plot of the headway as the function of the car index, when a jam or a cluster is formed. This cluster configuration evolves from a random initial headway distribution, as shown in the top inset. The bottom inset is the spinodal curve (the solid line without circles, plotted from Eq.(\[linear\])), and the coexistence curve from numerical calculations (the solid line fitting the solid circles). The solid circles are numerically observed extremum headways at different $\kappa$.[]{data-label="singlepeak"}](single_peak.pdf "fig:"){height="8cm"} Secondly the number of cars involved in the “kink" or “anti-kinks" are independent of $s_0$ and the total number of cars $N$. A “kink" is the “go front“, or the transition region from a cluster with $s_n\sim-s_{c2}$ to an anti-cluster with $s_n\sim s_{c2}$, while an ”anti-kink" is the “stop front", or the transition region from an anti-cluster to a cluster. Thus for large $N$ we can ignore cars in the “(anti-)kink". Since the sum of the headways over all vehicles is conserved during the time evolution, the number of cars in the cluster is given by $$\begin{aligned} \label{jnumber} N_j=\frac{N}{2}\frac{s_{c2}-s_0}{s_{c2}}\end{aligned}$$ Clearly for $s_0\ge s_{c2}$, no clusters can be formed, given random initial perturbations of any magnitude. Similarly, no anti-clusters can exist for $s_0<-s_{c2}$. We thus identify $s_{c2}$ as the coexistence curve [@kerner; @yu; @naka] and plot it together with $s_{c1}$ in Fig.(\[singlepeak\]). The numerically calculated coexistence curve and the spinodal line coincides at the critical neutral stability point located at $s_0=0, \kappa=2$, agreeing with previous analysis[@nagatani; @xue]. Note that $s_n$ can be negative, and the physical car density is calculated from Eq.(\[vc\]). There is also a duality between $s_0\leftrightarrow -s_0$, where clusters at $s_0$ corresponds to anti-clusters at $-s_0$, and all behaviors at $s_0$ are identical to those at $-s_0$. Progresses have been made in treating non-linear ODE describing car-following models analytically[@nakashini; @yokokawa1; @hayakawa; @nagatani]; For Eq.(\[ovmm\]) it is generally accepted that one can do a controled expansion near the critical neutral stability point and close to the neutral stability line; the former leads to modified KdV equations plus correction terms, that gives the approximate “(anti-)kink" solutions; the latter reduces the original model to KdV equations plus corrections that give rise to soliton solutions[@kurtze]. However, away from the neutral stability line, it is clear from numerical calculation that if one makes the car index continuous, the transition between the two extremum headways is discontinuous and analytically intractable. One can, however, show that the “kink" and “anti-kink" of a single cluster move at the same velocity, by taking $s=\sum_{n=i}^js_n$. For the “kink", the $i^{\text{th}}$ car is located in the cluster, while the $j^{\text{th}}$ car is located in the anti-cluster. From Eq.(\[ovmm\]) we have $$\begin{aligned} \label{calcv} \ddot s+\kappa_1\dot s=2\kappa_2\tanh s_{c2}\end{aligned}$$ The relevant set of solutions is $s=\left(2\kappa_2\tanh (s_{c2})/\kappa_1\right)t+C$, where $C$ is an unimportant constant of integration. This gives the velocity of the “kink" as the number of cars per unit time as follows $$\begin{aligned} \label{kinkv} v_{k}=\frac{\kappa_2}{\kappa_1}\frac{\tanh s_{c2}}{s_{c2}}\end{aligned}$$ The velocity of the “anti-kink" is calculated similarly, thus $v_k$ gives the velocity of the cluster, which again is independent of the car density of the traffic lane. Here we make the assumption that for cars far away from the “(anti-)kink", their headway takes the value of $\pm s_{c2}$. More importantly, if we concatenate two clusters together, as long as the assumption holds (e.g. when the two clusters are far away), they will move at the same velocity and will never merge. We will proceed to study the dynamics of the multi-cluster solutions. One would naively expect that a random initial state like the inset of Fig.(\[singlepeak\]) should lead to a random number of clusters[@zhang], at least in the limit of large $N$, subjecting to the constraint of Eq.(\[jnumber\]). However, our numerical results show that the probability distribution of the number of clusters is not random; it strongly depends on the initial headway $s_0$ and $\kappa$. We first calculate the probability distribution by fixing the strength of the initial random perturbation and $\kappa$ in Eq.(\[ovmm\]), and only vary the initial headway $s_0$. For each value of $s_0$, sufficiently large number of random initial states are generated until the probability for each number of clusters converges. The probability distribution is plotted in Fig.(\[probability\]), which is one of the main results of this work. 1= ![(Color online) The probability of having one to six clusters in a single traffic lane, plotted as the function of the initial average headway $s_0$. The probability is calculated with three hundred cars and random initial headway perturbation. Inset: The probability of having only one cluster, as the function of the initial average headway. The probability is calculated for three hundred cars to seven hundred cars, showing some numerical evidence that in the limit of large number of cars, the probability curve converges to a well-defined limit. The probability is calculated at $t=30000 s$.[]{data-label="probability"}](probability1.pdf "fig:"){height="7.5cm"} A few comments are in order here. In Fig.(\[probability\]) we take $\kappa=1$ and only plot the part where $s_0$ is negative, because the probability distribution is identical for $s_0$ and $-s_0$. For $\|s_0\|>0.87$ we can see the final state is dominated by one cluster, and this is true even for an infinitely long traffic lane as $N\rightarrow\infty$; in this case, most probably one very large cluster develops, instead of several clusters with smaller lengths. As $\|s_0\|$ decreases, the probability of having more than one cluster increases, and for $\|s_0\|<0.82$, it is almost impossible to have just one cluster. As $\|s_0\|$ further decreases towards zero, the average number of clusters most probably will tend to infinity. This cannot be observed numerically for a finite number $N$, since at $s_0=0$ the total number of cars in the clusters is $\sim N/2$(see Eq.(\[jnumber\])). For a physical traffic lane, from Eq.(\[vc\]) the maximum number of jams will occur at car density $\sim \left(C_2/C_1+l\right)^{-1}$. Increasing or decreasing from that car density reduces the number of jams. This phenomenon of large number of “phantom jams" occuring at some intermediate density could be used to empirically check the validity of the OV model. 1= ![The quasisoliton structure of $ds_n/dt$ as a function of the car index. The fewer the cars involved in the quasisoliton, the smaller the width of the quasisoliton, which depends only on $\kappa$ and not on the initial headway $s_0$. By convention a kink gives a positively charged quasisoliton as shown in this figure. An anti-kink gives a negatively charged quasisoliton. The top inset shows the dependence of the annihilation time $t_a$ on the number of cars between the quasisolitons of opposite charges, the exponential fit is numerically perfect. The intrinsic scale as a function of $s_{c2}$ is shown in the bottom inset.[]{data-label="interaction"}](bump.pdf "fig:"){width="9cm"} To understand the probability distribution of the number of clusters, we characterize quantitatively the strength of interaction between two clusters by the time it takes for them to merge. It is useful to plot $ds_n/dt$ instead of $s_n$ as a function of the car index $n$. The “kinks" and “anti-kinks" lead to exponentially localized “quasisolitons" of opposite charges (see Fig.(\[interaction\])), which closely resemble the “autosolitons" in dissipative non-linear systems[@book]. When quasisolitons of opposite charges annihilate each other, two (anti-)clusters merge into one. We numerically observe that the time needed for annihilation, $t_a$, increases exponentially with the number cars $n$ between the peaks of these two quasisolitons, giving the relationship $$\begin{aligned} \label{ta} t_a\sim e^{n/n_0}\end{aligned}$$ While Eq.(\[kinkv\]) dictates that kinks and anti-kinks travel at the same velocity, implying they would never merge, one should note the velocity is calculated from the extremal headways $s_{c2}$. The extremal headways are only attainable infinitely far away from the kink (or the anti-kink). Thus in principle, when multiple kinks and anti-kinks coexist in the same solution, they only move at the same velocity when they are infinitely apart. For finite separations, Eq.(\[kinkv\]) is only an approximation, thus leading to the annihilations between the kink and anti-kink pair. One thus note that when $\|s_0\|$ increases, the cluster (for $s_0>0$) or the anti-cluster (for $s_0<0$) region gets narrower(see Eq.(\[jnumber\])), leading to higher probability of short distances between the quasisolitons. Thus the probability of having multiple (anti-) clusters is suppressed, as shown in Fig.(\[probability\]). The intrinsic “scale" $n_0$ in Eq.(\[ta\]) depends on $s_{c2}$ or $\kappa$, which is also plotted in Fig.(\[interaction\]). This is analogous to the interaction and collapsing of kinks and anti-kinks in the Ginzburg-Landau theory[@rougemont], though here the total number of cars in the cluster has to satisfy Eq.(\[jnumber\]), so that at least one cluster will remain for a finite system with periodic boundary condition. Thus the greater the intrinsic scale, the stronger the interactions between quasisolitons, so this scale can be used to quantify the absolute value of the quasisoliton charge. The interaction leads to merging of clusters, reducing the probability of having a large number of clusters in the traffic lane. While the magnitude of the charge does not depend on $s_0$, Fig.(\[probability\]) will look qualitatively the same if the x-axis is replaced with increasing $s_{c2}$. The dependence of average number of clusters as a function of $s_0$ and $s_{c2}$ are plotted separately in Fig.(\[average\]), numerically supporting the above explanation. For any finite number of cars, all clusters will eventually merge in the limit of very long time; thus the statements here are only rigorous in the limit that the number of cars $N\rightarrow \infty$. However because of the exponential dependence of the annihilation time on the number of cars between quasisolitons of opposite charges, the statements here are true for all practical purposes when the number of cars is reasonably large (even for computer simulation because of the finite numerical resolutions). 1= ![The average number of clusters of a single lane traffic as a function of the intial headway (the top plot, while keeping the perturbation strength and $\kappa$ fixed), and as a function of $s_{c2}$ (the bottom plot, while keeping the perturbation strength and $s_0$ fixed).[]{data-label="average"}](average1.pdf "fig:"){width="8cm"} We would also like to make a cautionary note here that both the cluster statistics in Fig.(\[probability\]) and the average number of clusters in Fig.(\[average\]) depend on the number of vehicles $N$ and the time of simulation $t$. In principle, however, those two quantities are only well-defined in the limit of both $N$ and $t$ going to infinity. The finite scaling of the OV model is unfortunately very expensive numerically. On the other hand, the exponential dependence of the annihilation time on the number of cars between quasisolitons of opposite charges implies the cluster statistics and the average number of clusters converge very fast when $N$ increases (see also the inset of Fig.(\[probability\])). One should also note that formally, the emergent quantities discussed in this paper are only well-defined when we take the limit of $N\rightarrow\infty$ first, followed by the limit of $t\rightarrow\infty$. In conclusion, we have investigated the OV model in the non-linear regime, where the metastable phase is delineated by the critical average initial headway $s_{c1}$ and $s_{c2}$. The behavior of the traffic jam evolution seems to be completely determined by the charge of, and the distance between, quasisolitons of opposite signs. This leads to non-trivial statistics of multiple clusters that depends both on $s_0$ and $s_{c2}$. This property is not only present in the OV model shown in details here. We have done extensive (but not necessarily thorough) numerical calculations for various extended OV models, which suggests that all features discussed above are qualitatively the same. A comprehensive and quantitative study of extended OV models will be presented elsewhere. Apart from its theoretical interest, we believe such studies are useful in designing and optimizing autonomous intelligent transport systems, where multiple clusters lead to undesirable wear-and-tear and need to be suppressed. It would also be interesting to see how the cluster statistics could be modified for more complicated traffic lanes with road works[@werner]. Given the universality of our results, it is also important to check the cluster statistics against the empirical data when modeling of real traffic dynamics is concerned, so as to understand what aspect of the real traffic complexity can really be captured by the General Motors model classes[@schreck]. We would like to thank Prof. Weizhu Bao and Prof. Ren Weiqing from National University of Singapore for useful comments. This research was partially supported by Singapore A$^{\star}$STAR SERC “Complex Systems" Research Programme grant 1224504056. The numerical calculations in this work is supported by ACRC of A$^{\star}$STAR. [99]{} M. Treiber and A. Kesting, Traffic Flow Dynamics, Springer-Verlag Berlin Heidelberg 2013, and the references therein. Y. Sugiyama, M. Fukui, M. Kikuchi, K. Hasebe, A. Nakayama, K. Nishinari, S. Tadaki and S. Yukawa, New. J. Phys. [10]{}, 033001 (2008). A. Nakayama, M. Fukui, M. Kikuchi, K. Hasebe, K. Nishinari, Y. Sugiyama, S. Tadaki and S. Yukawa, New. J. Phys. [11]{}, 083025 (2009). B.S. Kerner and H. Rehborn, Phys. Rev. E. [53]{}, R1297 (1996). B.S. Kerner and H. Rehborn, Phys. Rev. Lett. [79]{}, 4030 (1997). B.S. Kerner, Phys. Rev. Lett. [81]{}, 3797 (1998). G. Orosz, R.E. Wilson, R. Szalai and G. Stepan, Phys. Rev. E. [80]{}, 046205 (2009). M. Bando, K. Hasebe, A. Nakayama, A. Shibata, and Y. Sugiyama, Phys. Rev. E. [51]{}, 1035 (1995). T. Seidel, I. Gasser and B. Werner, SIAM J. Appl. Dyn. Syst. [8]{}, 1305 (2009). D.M. Miedema, A.S. de Wijn and P. Schall, Phys. Rev. E. [89]{}, 062812 (2014). K. Nagel and M. Schreckenberg, J. Phys. I. France [2]{}, 2221 (1992). H.K. Lee, H.-W. Lee and D. Kim, Phys. Rev. E. [69]{}, 016118 (2004). M.R. Flynn, A.R. Kasimov, J.-C. Nave, R.R. Rosales and B. Seibold, Phys. Rev. E. [79]{}, 056113 (2009). N. Bellomo and C. Dogbe, Soc. Ind. App. Math. [53]{}, 409 (2011). D. Helbing, Rev. Mod. Phys. [73]{}, 1067 (2001). D. Chowdhury, L. Santen and A. Schadschneider, Phys. Rep. [329]{}, 199 (2000). D. Helbing and B. Tilch, Phys. Rev. E. [58]{}, 133 (1998). R. Jiang, Q. Wu and Z. Zhu, Phys. Rev. E. [64]{}, 017101 (2001). G. Peng and R. Cheng, Phys. A. [392]{}, 3563 (2013). H. Gong, H. Liu and B. Wang, Phys. A. [387]{}, 2595 (2008). M. Treiber, A. Hennecke and D. Helbing, Phys. Rev. E. [62]{} 1805 (2000). D. Shamoto, A. Tomoeda, R. Nishi and K. Nishinari, Phys. Rev. E. [83]{}, 046105 (2011). B.S. Kerner and S.L. Klenov, J. Phys. A: Math. Gen. [35]{} L31-L43 (2002). B.S. Kerner and S.L. Klenov, J. Phys. A: Math. Gen. [39]{} 1775 (2006). B.S. Kerner and S.L. Klenov, Phys. Rev. E. [68]{}, 036130 (2003). B. Yang and C. Monterola, Phys. Rev. E, [92]{}, 042802(2015). H.X. Ge, S.Q. Dai, L.Y. Dong and Y. Xue, Phys. Rev. E. [70]{}, 066134 (2004). A. Nakayama, Y. Sugiyama and K. Hasebe, Phys. Rev. E. [65]{}, 016112 (2001). X. Xu, J. Pang and C. Monterola, Physica A, [436]{}, 565(2015). B.S. Kerner, Physica A, [392]{}, 5261(2013). B.S. Kerner, S.L. Klenov and M. Schreckenberg, Phys. Rev. E. [89]{}, 052807(2014). M. Schonhof and D. Helbing, Transport. Res. B. [43]{}, 784 (2009). M. Treiber, A. Kesting and D. Helbing, Transport. Res. B. [44]{}, 983 (2010). B.S. Kerner, Introduction to Modern Traffic Flow Theory and Control: The Long Road to Three-phase Traffic Theory, Springer-Verlag Berlin Heidelberg 2009, and the references therein. I. Goldhirsch and G. Zanetti, Phys. Rev. Lett. [70]{}, 1619 (1993). S.R. Waitukaitis, H.F. Grutjen, J.R. Royer and H.M. Jaeger, Phys. Rev. E. [83]{}, 051302 (2011). A. Kudrolli, M. Wolpert and J.P. Gollub, Phys. Rev. Lett. [78]{}, 1383 (1997). B.S. Kerner and P. Konhauser, Phys. Rev. E. [50]{}, 54 (1994). W. Shi, N-G Chen and Y. Xue, Commun. Theor. Phys. [48]{}, 1088 (2007). H. Hayakawa and K. Nakanishi, Phys. Rev. E. [57]{}, 3839 (1998). H. Hayakawa and K. Nakanishi, Phys. Rev. E. [57]{}, 3839 (1998). T. Nagatani, Phys. Rev. E. [61]{}, 3564 (2000). Y. Igarashi, K. Itoh and K. Nakanishi, J. Phys. Soc. Jpn. [68]{}, 791 (1999). Y. Igarashi, K. Itoh, K. Nakanishi, K. Ogura and K. Yokokawa, Phys. Rev. Lett. [83]{}, 718 (1999). W.L. Jin and H.M. Zhang, Transport. Res B. [37]{}, 207 (2003). B. S. Kerner and V. V. Osipov, Autosolitons. A New Approach to Problems of Self-Organization and Turbulence, Kluwer Academic Publishers (1995). J.-P. Eckmann and J. Rougemont, Commun. Math. Phys. [199]{}, 441 (1998). J. Rougemont, Nonlinearity. [12]{}, 539 (1999). H.C. Rosu, O. Cornejo-Perez and P. Ojeda-May, Phys. Rev. E. [85]{}, 037102 (2012). D.A. Kurtze, Phys. Rev. E. [88]{}, 032804 (2013).	{ "pile_set_name": "ArXiv" }
--- abstract: 'We consider conditions which force a well-quasi-ordered poset (wqo) to be better-quasi-ordered (bqo). In particular we obtain that if a poset $P$ is wqo and the set $S_{\omega}(P)$ of strictly increasing sequences of elements of $P$ is bqo under domination, then $P$ is bqo. As a consequence, we get the same conclusion if $S_{\omega} (P)$ is replaced by $\mathcal J^1(P)$, the collection of non-principal ideals of $P$, or by $AM(P)$, the collection of maximal antichains of $P$ ordered by domination. It then follows that an interval order which is wqo is in fact bqo.' author: - \| Maurice Pouzet [^1]\ PCS, Université Claude-Bernard Lyon1, Domaine de Gerland -bât.\ Recherche \[B\],\ 50 avenue Tony-Garnier, F$69365$ Lyon cedex 07, France\ [[email protected] ]{} - \| Norbert Sauer [^2]\ Department of Mathematics and Statistics, The University of Calgary,\ Calgary, T2N1N4, Alberta, Canada\ [[email protected]]{} title: 'From well-quasi-ordered sets to better-quasi-ordered sets ' --- [*Key words: poset, ideal, antichain, critical pair, interval-order, barrier, well-quasi-ordered set, better-quasi-ordered set.* ]{} Introduction and presentation of the results ============================================ How to read this paper ---------------------- Section \[index\] contains a collection of definitions, notations and basic facts. The specialist reader should be able to read the paper with only occasional use of Section \[index\] to check up on some notation. Section \[index\] provides readers which are not very familiar with the topic of the paper with some background, definitions and simple derivations from those definitions. Such readers will have to peruse Section \[index\] frequently. The paper is organized as follows. Section \[secbarriers\] provides the basics behind the notion of bqo posets and develops the technical tools we need to work with barriers and concludes with the proof of a result about $\alpha$-bqo’s from which Theorem \[nonprincipal\] follows. We present some topological properties of ideals in Section \[sec:ideals\] and discuss minimal type posets in Section \[sec:min\]. The proof of Theorem \[criticbqo\] is contained in Section \[sec:pred\]. In Section \[maximal\] we present constructions involving maximal antichains of prescribed size. Background ---------- Since their introduction by G.Higman [@higm], well-quasi-ordered, (wqo), sets have played an important role in several areas of mathematics: algebra (embeddability of free algebras in skew-fields, elimination orderings) set theory and logic (comparison of chains, termination of rewriting systems, decision problems) analysis (asymptotic computations, symbolic dynamic ). A recent example is given by the Robertson-Seymour Theorem [@Rob-Sey] asserting that the collection of finite graphs is well-quasi-ordered by the minor relation. In this paper we deal with the stronger notion of better-quasi-ordered, (bqo), posets. Bqo posets where introduced by C. St. J. A. Nash-Williams, see [@nwbqots], to prove that the class of infinite trees is wqo under topological embedding. Better-quasi-orders enjoy several properties of well-quasi-orders. For example, finite posets are bqo. Well ordered chains are bqo, finite unions and finite products of bqo posets are bqo. The property of being bqo is preserved under restrictions and epimorphic images. Still there is a substantial difference: Well-quasi-ordered posets are not preserved under the infinitary construction described in the next paragraph, but better-quasi-ordered posets are . A basic result due to G.Higman, see [@higm], asserts that a poset $P$ is wqo if and only if ${\bf I}(P)$, the set of initial segments of $P$, is well-founded. On the other hand, Rado [@rado] has produced an example of a well-founded partial order $P$ for which ${\bf I}(P)$ is well-founded and contains infinite antichains. The idea behind the bqo notion is to forbid this situation: ${\bf I}(P)$ and all its iterates, ${\bf I}({\bf I}(\cdots ({\bf I}(P)\cdots ))$ up to the ordinal $\omega_1$, have to be well-founded and hence wqo. This idea is quite natural but not workable. (Proving that a two element set satisfies this property is far from being an easy task). The working definition, based upon the notion of [barrier]{}, invented by C. St. J. A. Nash-Williams, is quite involved, see [@nwwqoft] and [@nwbqots]. Even using this working condition, it is not so easy to see wether a wqo is a bqo or not. We aim to arrive at a better understanding of bqo posets and consider two special problems to see if indeed we obtained such a better understanding. We solved the first problem, to characterize bqo interval orders, completely, see Theorem \[thm:int\]. The second was Bonnet’s problem, see Problem \[prob:B\]. We related the property of a poset to be bqo to the bqo of various posets associated to a given poset, in particular the poset of the maximal antichains under the domination order. We think that those results stand on their own but unfortunately don’t seem to be strong enough to solve Bonnet’s problem. The results ----------- Let $P$ be a poset. For $X,Y\subseteq P$ let $X\leq_{dom}Y$ if for every $x\in X$ there is a $y\in Y$ with $x\leq y$. The quasi-order $\leq_{dome}$ is the [domination]{} order on $P$ and $S_\omega(P)$ is the set of strictly increasing $\omega$-sequences of $P$. We will prove, see Theorem \[nonprincipal2\] and the paragraph before Theorem \[nonprincipal2\]: \[nonprincipal\] If $P$ is wqo, and $(S_{\omega}(P);\leq_{dom})$ is bqo then $P$ is bqo. Let $C\in S_\omega(P)$. Then $\downarrow\hskip-2pt C$ is an ideal of $P$. On the other hand if $I$ is an ideal with denumerable cofinality then $I=\downarrow\hskip-2pt C$ for some $C\in S_\omega(P)$. Let $\mathcal J^{\neg \downarrow\hskip -2pt }(P)$ be the set of non principal ideals. Since ideals with denumerable cofinality are non-principal, we obtain from Theorem \[nonprincipal\] and the property of bqo to be preserved under restrictions that: \[cornonprincipal\] If $P$ is wqo and $\mathcal J^{\neg \downarrow\hskip -2pt }(P)$ is bqo then $P$ is bqo. The poset $(S_{\omega}(P);\leq_{dom})$ is often more simple than the poset $P$. So for example if $P$ is finite $S_\omega(P)=\emptyset$. It follows trivially from Definition \[defin:bqo\] that the empty poset is bqo and hence from Theorem \[nonprincipal\] that finite posets are bqo. A result which is of course well known. Also: \[cor2\]If $P$ is wqo and $\mathcal J^{\neg \downarrow\hskip -2pt }(P)$ is finite then $P$ is bqo. Corollary \[cornonprincipal\] was conjectured by the first author in his thesis[@pouzettr] and a proof of Corollary \[cor2\] given there. The proof is given in [@fraissetr] Chapter $7$, subsections $7.7.7$ and $7.7.8$. pp $217-219$. The above considerations suggest that $S_\omega(P)$ corresponds to some sort of derivative. As already observed, the elements of $S_\omega(P)$ generate the non-principal ideals of $P$ with denumerable cofinality. The set of ideals $\mathcal{J}(P)$ of $P$ form the base set of a Cantor space $C$, see Section \[sec:ideals\]. It follows that $P$ is finite if and only if the first Cantor-Bendixson derivative of $C$ is finite and that $\mathcal{J}(P)$ is finite if and only if the second Cantor-Bendixson derivative of $C$ is finite. The space $C$ contains just one limit if and only if $\mathcal{J}(P)$ is a singleton space. Such posets are called minimal type posets. Minimal type posets occur naturally in symbolic dynamics. See section \[sec:min\] for details. If $P$ is an interval order then $\mathcal J^{\neg \downarrow\hskip -2pt }(P)$ is a chain. Let $I,J\in \mathcal J^{\neg \downarrow\hskip -2pt }(P)$. If $I\setminus J\not = \emptyset$ and $J\setminus I \not = \emptyset$ pick $x \in I \setminus J$ and $y\in J\setminus I$. Since $I$ is not a principal ideal then $x$ is not a maximal element in $I$, so we may pick $x'\in I$ such that $ x<x'$. For the same reason, we may pick $y'\in J$ such that $y<y'$. Clearly, the poset induced on $\{x,x',y,y'\}$ is a $\underline 2\oplus \underline 2$. But then $P$ is not an interval order. Corollary \[cornonprincipal\] has an other immediate consequence: \[chain\] If $P$ is wqo and $\mathcal J^{\neg \downarrow\hskip -2pt } (P)$ is a chain then $P$ is bqo. Indeed, if $P$ is wqo then ${\bf I}(P)$ is well-founded. In particular $\mathcal J^{\neg \downarrow\hskip -2pt }(P)$ is well-founded. If $\mathcal J^{\neg \downarrow\hskip -2pt }(P)$ is a chain, this is a well-ordered chain, hence a bqo. From Corollary \[cornonprincipal\] $P$ is bqo. From Corollary \[chain\], this gives: \[thm:int\] An interval order is bqo iff it is wqo. The following Theorem is an immediate consequence of Theorem \[thmsucc\]. (See Section \[index\] for a definition of the notions used in Theorem \[criticbqo\].) \[criticbqo\] Let $P$ be a poset. If $P$ has no infinite antichain, then the following properties are equivalent: 1. $P$ is bqo. 2. $(P;\leq_{succ})$ is bqo. 3. $(P;\leq_{pred})$ is bqo. 4. $(P;\leq_{crit})$ is bqo. 5. $AM(P)$ is bqo.\[item:5\] As indicated earlier part of the motivation for this research was an intriguing problem due to Bonnet, see [@bonnet]. \[prob:B\] Is every wqo poset a countable union of bqo posets? Item [\[item:5\]]{} of Theorem \[criticbqo\] may suggest to attack Bonnet’s problem using the antichains of the poset. Note that if a poset $P$ is wqo but not bqo, then it contains antichains of arbitrarily large finite size. Indeed, if the size of antichains of a poset $P$ is bounded by some integer, say $m$, then from Dilworth’s theorem, $P$ is the union of at most $m$ chains. If $P$ is well founded, these chains are well ordered chains, hence are bqo, and $P$ is bqo as a finite union of bqo’s. For each integer $m$, let $AM_{m}(P)$ be the collection of maximal antichains having size $m$ and $\bigcup {AM_{m}(P)}$ be the union of these maximal antichains. Trivially, $AM(P)$ is the union of the sets $ {AM_{m}(P)}$ for $m\in {\mathbb{N}}$ if and only if $P$ contains no infinite antichain. It its tempting to use this decomposition to attack Bonnet’s problem. That is, consider the union $\bigcup {AM_{m}(P)}=P$ of a well-quasi-ordered poset $P$. This does not work in general: there are wqo posets $P$ for which $\bigcup AM_2(P)$ is not bqo ( Lemmas \[construct1\] and \[counterexample\]). Still, this works for wqo posets $P$ for which $\bigcup {AM_{m}(P)}$ is bqo for every $m$ and $AM(P)$ is not bqo (these $P$ are not bqo). Rado’s poset provides an example, see Lemma \[AM(R)\]. Looking at the relationship between $AM_m(P)$ and $\bigcup {AM_m(P)}$, we prove, see Theorem \[thm:222\]: Let $P$ be a poset with no infinite antichain, then $AM_2(P)$ is bqo if and only if $\bigcup {AM_2(P)}$ is bqo. This does not extend: it follows from Corollary \[construct2\] that there exists a wqo poset $P$ for which $AM_{3}(P)$ is bqo but $\bigcup {AM_{3}(P)}$ is not bqo. Barriers and better-quasi-orders {#secbarriers} ================================ Basics ------ We use Nash-William’s notion of bqo, see [@nwwqoft], and refer to Milner’s exposition of bqo theory, see [@milnerwqobqo]. See Section \[index\] for the basic definitions. The following result due to F.Galvin(1968) extends the partition theorem of F.P.Ramsey. [@galvin] For every subset $B$ of $[{\mathbb{N}}]^{<\omega}$ there is an infinite subset $X$ of ${\mathbb{N}}$ such that either $[X]^{<\omega}\cap B=\emptyset $ or $[X]^{<\omega}\cap B$ is a block. Trivially, every block contains a thin block, the set $\min_{\leq _{in}}(B)$ of $\leq_{in}$ minimal elements of the block. Moreover, if $B$ is a block, resp.a thin block, and $X$ is an infinite subset of $\bigcup B$ then $B_{\restriction X}$ is a block, resp a thin block. The theorem of Galvin implies the following result of Nash-Williams, see [@milnerwqobqo]. \[partnash\] 1. Every block contains a barrier. 2. For every partition of a barrier into finitely many parts, one contains a barrier. The partial order $(B,\leq_{lex})$ is the lexicographic sum of the partial orders $(B_{(i)},\leq_{lex})$: $$(B,\leq_{lex})=\sum_{i\in {\mathbb{N}}}(B_{(i)}, \leq_{lex}).$$ Let $\mathrm{T}(B)$ be the tree $\mathrm{T}(B):=\bigl(\{t : \exists s\in B(t\leq_{in}s)\}, \leq_{in}\bigr)$ with root $\emptyset$ and $\mathrm{T}^d(B)$ the dual order of $\mathrm{T}(B)$. If $\mathrm{T}(B)$ does not contain an infinite chain then $\mathrm{T}^{d}(B)$ is well founded and the height function satisfies $$\begin{aligned} & h\bigl(\emptyset , \mathrm{T}^{d}(B)\bigr)= \\ & \sup\{h\bigl((a), \mathrm{T}^{d}(B)\bigr) +1: (a)\in \mathrm{T}(B)\} = \sup\{h\bigl(\emptyset, \mathrm{T}^d(_{(a)}B)\bigr) +1: (a)\in \mathrm{T}(B) \}. \end{aligned}$$ Induction on the height gives then that $\mathrm{T}(B)$ is well ordered under the lexicographic order. The order type of $\mathrm{T}$ being at most $\omega^{\alpha}$ where $\alpha:=h(\emptyset , \mathrm{T}^{d}(B))$. From this fact, we deduce: [@pouz72] Every thin block, and in particular every barrier, is well ordered under the lexicographic order. This allows to associate with every barrier its order-type. We note that $\omega$ is the least possible order-type. An ordinal $\gamma$ is the order-type of a barrier if and only if $\gamma=\omega^{\alpha}\cdot n$ where $n<\omega$ and $n=1$ if $\alpha <\omega$ [@assous74]. Every barrier contains a barrier whose order-type is an indecomposable ordinal. \[defin:bqo\] A map $f$ from a barrier $B$ into a poset $P$ is [good]{} if there are $s,t\in B$ with $s\lhd t$ and $f(s) \leq f(t)$. Otherwise $f$ is [bad]{}.\ Let $\alpha$ be a denumerable ordinal. A poset $P$ is [$\alpha$-better-quasi-ordered]{} if every map $f: B\rightarrow P$, where $B$ is a barrier of order type at most $\alpha$, is good.\ A poset $P$ is [better-quasi-ordered]{} if it is $\alpha$-better-quasi-ordered for every denumerable ordinal $\alpha$. It is known and easy to see that a poset $P$ is $\omega$-better-quasi-ordered if and only if it is well-quasi-ordered. Remember that we abbreviate better-quasi-order by bqo. Since every barrier contains a barrier with indecomposable order type, only barriers with indecomposable order type need to be taken into account in the definition of bqo. In particular, we only need to consider $\alpha$-bqo for indecomposable ordinals $\alpha$. Note that for indecomposable ordinals $\alpha$ the notion of $\alpha$-bqo leeds to different objets [@marcone1994] . We will need the following results of Nash-Williams (for proofs in the context of $\alpha$-bqo, see [@milnerwqobqo] or [@pouzet92]): \[higbqo\] Let $P$ and $Q$ be partial orders, then: 1. Finite partial orders and well ordered chains are bqo. 2. If $P,Q$ are $\alpha$-bqo then the direct sum $P\oplus Q$ and the direct product $P\times Q$ are $\alpha$-bqo. 3. If $P$ is $\alpha$-bqo and $f:P\longrightarrow Q$ is order-preserving then $f(P)$ is $\alpha$- bqo. 4. If $P$ embeds into $Q$ and $Q$ is $\alpha$-bqo then $P$ is $\alpha$-bqo. 5. If $C\subseteq \mathfrak {P}(P)$ is $\alpha$-bqo then the set of finite unions of members of $C$ is $\alpha$-bqo. It follows from Item $e$ that if $P$ is $\alpha$-bqo then $I_{<\omega}(P)$ is $\alpha$-bqo which in turn implies that if $P$ is $\alpha$-bqo then $AM(P)$ is $\alpha$-bqo. (If $P$ is $\alpha$-bqo then it is $\omega$-bqo and hence well-quasi-ordered and hence does not contain infinite antichains. Then $AM(P)$ embeds into $A(P)$ which in turn embeds into $I_{<\omega}(P)$.) It follows from Item $d$ that if $P$ is bqo then every restriction of $P$ to a subset of its elements is also bqo. Barrier constructions --------------------- Let $B$ be a subset of $[{\mathbb{N}}]^{<\omega}$. See Section \[index\] for notation. If $B$ is a block then $B^2$ is a block and if $B$ is a thin block then $B^2$ is a thin block. Moreover, if $B$ is a thin block, and $u\in B^2$, then there is a unique pair $s,t\in B$ such that $s\lhd t$ and $u= s\cup t$. If $B$ is a block, then $\bigcup {_{}B}=\bigcup{_{}B}= \bigcup {B}\setminus \min(\bigcup B)$ and $_{}B$ is a block. Moreover, if $B$ is well ordered under the lexicographic order then $_{}B$ is well ordered too and if the type of $B$ is an indecomposable ordinal $\omega^{\gamma}$ then the type of $_{}B$ is at most $\omega^{\gamma}$. If $C$ is a block and $B:= C^2$ then $_{}B= C\setminus C_{(a)}$, where $a$ is the least element of $\bigcup C$. The following Lemma is well known and follows easily from the definition. If $B$ is a barrier, then $B^{2}$ is a barrier and if $B$ has type $\alpha$ then $B^2$ has type $\alpha\cdot\omega$. A generalization, Lemma \[lem:Marcone\] below, was given by A.Marcone. We recall his construction and result, see [@marcone2001] Lemma 8 pp. 343. Let $B$ be a subset of $[{\mathbb{N}}]^{<\omega}$. Then $B^{\circ}$ is the set of all elements $s\in B$ with the property that for all $i\in \bigcup{B}$ with $i<s(0)$ there is an element $t\in B$ with $(i)\cdot {_\ast s}\leq_{in} t$. In other words $s\in B^{\circ}$ if $(i)\cdot {_\ast s}\in \mathrm{T}(B)$ for all $i\in \bigcup{B}$ with $i<s(0)$. Let $B^\prime:=\{{_\ast s} : s\in B^\circ\}\setminus \{\emptyset\}$. \[lem:Marcone\] Let $B$ be a thin block of type larger than $\omega$, then: 1. $B'$ is a thin block. 2. For every $u\in (B')^2$ there is some $s\in B$ such that $s\leq_{in} u$. 3. If the type of $B$ is at most $\omega^{\gamma}$ then $B'$ contains a barrier of type at most $\omega^{\gamma}$ if $\gamma$ is a limit ordinal and at most $\omega^{\gamma-1}$ otherwise. We may note that for every $s\in B$ such that $s\leq_{in} u:= s'\cup t'$ we have $s'\leq_{in}s$. Indeed, otherwise $s\leq_{in }s'$, but $s':=_{}s''$ for some $s''\in B$, hence $s\subseteq s''$ contradicting the fact that $B$ is a barrier. A barrier $B$ is [end-closed]{} if $$\label{eqendclosed} s\cdot(a)\in B \; \text{ for every }s\in B_{} \text{ and } a\in \bigcup B \; \text{ with } a>\lambda (s).$$ For example, $[{\mathbb{N}}]^{n}$ is end-closed for every $n$, $n\geq 1$, as well as the barrier $B:= \{s\in {\mathbb{N}}^{<\omega}: l(s)=s(0)+2\}$. If $B, C$ are two barriers with the same domain, the set $BC:= \{s\cdot t: s\in C, t\in B, \lambda(s)<t(0)\}$ is a barrier, the [product* ]{} of $B$ and $C$, see [@pouz72]. Its order-type is $\omega^{\gamma+\beta}$ if $\omega^{\gamma}$ and $\omega^{\beta}$ are the order-types of $B$ and $C$ respectively. For example, the product $[\bigcup B]^1B$ is end-closed. Provided that $B$ has type $\omega^{\beta}$, it has type $\omega^{1+\beta}$. The converse holds, namely: \[fact0\] The set $D\subseteq {\mathbb{N}}^{<\omega}$ is an end-closed barrier of type larger than $\omega$ if and only if $D_{}$ is a barrier and $D:= [\bigcup {D_{}}]^1{D_{}}$. Every barrier $B$ contains an end-closed subbarrier $B^\prime$. Induction on the order-type $\beta$ of $B$. If $\beta:=\omega$ then $B=[\bigcup B]^{1}$ and we may set $B':= B$. Suppose $\beta>\omega$ and every barrier of type smaller than $\beta$ contains an end-closed subbarrier. The set $S(B):= \{i \in \bigcup B: (i) \in B\}$ is an initial segment of $\bigcup B$. (Indeed, let $i\in S(B)$ and $j<i$ with $j\in \bigcup B$. Select $X\in [\bigcup B]^{\omega}$ such that $(j,i)\leq_{in}X$. Since $B$ is a barrier, $X$ has an initial segment $s\in B$. Since $B$ is an antichain w.r.t. inclusion $i\not\in s$, hence $s=(j)$.) The type of $B$ is larger than $\omega$, hence $S(B)\not=\bigcup{B}$. Set $i_{0}:=\min( \bigcup {B}\setminus S(B))$. The set $_{(i_{0})}B$ is a barrier because $i_{0}\not \in S(B)$. Hence induction applies providing some $X_{0} \subseteq \bigcup {B}\setminus (S(B)\cup \{i_{0}\})$ such that $\{s: (i_0)\cdot s \in B\cap [X_{0}]^{<\omega} \}$ is an end-closed barrier of domain $X_{0}$. It follows that $\{i_0\}\cup X_0\subseteq \bigcup{B}$. Starting with $(x_0,X_0)$ we construct a sequence $( i_{n}, X_{n})_{n<\omega}$ such that for every $n<\omega$: 1. $\{s: (i_n)\cdot s \in B\cap [X_{n}]^{<\omega} \}$ is an end-closed barrier of domain $X_{n}$. 2. $\{i_{n+1}\}\cup X_{n+1}\subseteq X_{n}$. Let $n<\omega$. If $(i_m, X_{m})_{m<n}$ is defined for all $m<n$ replace $B$ by $B\cap [X_{n-1}]^{<\omega}$ in the construction of $x_0$ and $X_0$ to obtain $i_{n}$ and $X_{n}$. Then for $X:= \{i_{n}: n<\omega\}$ set $B':= B\cap [X]^{<\omega}$. On the comparison of blocks ----------------------------- Let $B,B'$ be two subsets of $[{\mathbb{N}}]^{<\omega}$. We write $B'\leq_{in} B$ if : $$\label{eqbarrier} \mbox{For every} \; s'\in B'\; \mbox { there is some}\; s\in B\; \mbox { such that } \; s'\leq_{in} s.$$ This is the quasi-order of domination associated with the order $\leq_{in}$ on $[{\mathbb{N}}]^{<\omega}$. \[fact1\] Let $B$, $B'$ be two thin blocks. If $B'\leq_{in} B$ then for all $s,t\in B$ and $s^\prime, t^\prime\in B^\prime$: 1. $\bigcup B'\subseteq \bigcup B$. 2. If $s\leq_{in} s'$ then $s=s'$, hence $\leq_{in}$ is a partial order on thin blocks. 3. If $\bigcup B=\bigcup B'$ then for every $s\in B$ there is some $s'\in B'$ such that $s'\leq_{in} s$. 4. If $s'\lhd t' $ then $s\lhd t$ for some $s,t\in B$ with $s'\leq_{in} s$ and $t'\leq_{in} t$. 5. The set $B'':= B'\cup D$ with $D:= \{s\in B: \forall s'\in B'\, (s'\not \leq_{in} s)\}$ is a thin block and $\bigcup B''= \bigcup B$ and $B''\leq_{in}B$. $(a), (b), (c)$ follow from the definitions. $(d)$. Let $s', t'\in B'$ with $s'\lhd t' $ and let $t'':= s'\cup t'$. Since $\bigcup{B'}\subseteq \bigcup B$, $t''\subseteq \bigcup B$. Let $X\in [\bigcup B] ^{\omega}$ such that $t''\leq_{in} X$. There are $s, t\in B$ such that $s\leq _{in}X$ and $t\leq_{in}{_}X$. We have $s\lhd t$. It follows from $(b)$ that $s'\leq_{in} s$ and $t'\leq_{in} t$. $(e)$ $\bigcup B''= \bigcup B'\cup \bigcup D$ and $\bigcup B'\subseteq \bigcup B$ imply $\bigcup B''\subseteq \bigcup B$. For the converse, let $x\in \bigcup B\setminus \bigcup B'$. Since $B$ is a block there is some $s\in B$ having $x$ as first element. Clearly $s\in D$, hence $x\in D$, proving $\bigcup B''= \bigcup B$. From the definition, $B''$ is an antichain. Now, let $X\subseteq B$. We prove that some initial segment $s''$ belongs to $B''$. Since $B$ is a block, some initial segment $s$ of $X$ belongs to $B$. If $s\in D$ set $s'':= s$. Otherwise some initial segment $s'$ of $s$ is in $B'$. Set $s'':=s'$. Let $f: B\rightarrow P$ and $f': B'\rightarrow P$ be two maps. Set $f'\leq_{in} f$ if $B'\leq_{in} B$ and $f'(s')=f(s)$ for every $s' \in B'$, $s\in B$ with $s'\leq_{in} s$. Let $\mathcal{H}_X(P)$ be the set of maps $f: B\rightarrow P$ for which $B$ is a thin block with domain $X$. \[fact2\] Let $f: B\rightarrow P$ and $f': B'\rightarrow P$ with $f'\leq_{in}f$. If $B' $ and $B$ are thin blocks then $B'$ extends to a thin block $B''$ and $f'$ to a map $f''$ such that $\bigcup B''=\bigcup B$ and $f''\leq _{in} f$. Applying $(e)$ of Fact \[fact1\], set $B'':= B'\cup D$ and define $f''$ by setting $f''(s''):=f(s)$ if $s''\in D$ and $f''(s''):=f'(s'')$ if $s''\in B''$. Then $f''\leq_{in} f$. \[fact3\] Let $P$ be a poset and $X\in {\mathbb{N}}^\omega$, then: 1. The relation $\leq _{in}$ is an order on the collection of maps $f$ whose domain is a thin block and whose range is $P$. 2. Every $\leq_{in}$-chain has an infimum on the set $\mathcal{H}_X(P)$. 3. An element $f$ is minimal in $\mathcal{H}_X(P)$ if and only if every $f'$ with $f'\leq_{in} f$ is the restriction of $f$ to a sub-block of the domain of $f$. 4. If $f$ is minimal in $\mathcal{H}_X(P)$ and $f'\leq_{in} f$ has domain $C$ then $f^\prime$ is minimal in $\mathcal{H}_{\bigcup{C}}(P)$. 5. Let $B$ be a thin block and $f:B\to P$. If $f$ is bad and $f'\leq_{in} f$ then $f'$ is bad. $(a)$ Obvious. $(b)$ Let $\mathcal D:= \{f_{\alpha}: B_{\alpha} \rightarrow P, \alpha\}$ be a chain of maps. Let $\mathcal C:= \{dom (f): f\in \mathcal D\}$. Then $\{s \in \bigcup \mathcal C: s'\in \bigcup \mathcal C \Rightarrow s' \not <s\}$ is a thin block and the infimum of $\mathcal{C}$. For $s\in D$, let $f'(s)$ be the common value of all maps $f_{\alpha}$. This map is the infimum of $\mathcal D$. $(c)$ Apply Fact \[fact2\]. $(d)$ Follows from $(c)$. $(e)$ Apply $(d)$ of Fact \[fact1\]. \[minimalbad\] Let $f$ be a map from a thin block $B$ into $P$ and let $\mathcal F:= \{f'\in \mathcal{H}_{\bigcup B} (P) : f'\leq_{in} f\}$. Then there is a minimal $f'\in \mathcal{F}$ such that $f'\leq_{in} f$. Follows from Fact \[fact3\] $(b)$ using Zorn’s Lemma. Let $s,t\in [{\mathbb{N}}]^{s}$. Set $s\leq_{end}t$ if $\lambda(s)\leq \lambda(t)$ and $s_{}=t_{}$. \[key\]Let $f:B\rightarrow P$ a bad map. If $P$ is wqo and $f$ is minimal then there is an end-closed barrier $B'\subseteq B$ such that: $$\label{eqendclosed2} s<_{end}t \mbox {\; in\; } B' \Rightarrow f(s)<f(t) \mbox{\; in\; } P.$$ Let $B_1$ be a an end-closed subbarrier of $B$ and let $C_1:= \{s\cdot(b) : s\in B_1, b\in \bigcup B_1, b>\lambda(s)\}$. Divide $C_1$ into three parts $D_{i}$, $i<3$, with $D_{i}:= \{ s'\cdot (ab)\in C_1 : f(s'\cdot (a))\rho_{i}f(s'\cdot(b))\}$ where $\rho_{0}$ is the equality relation, $\rho_{1}$ is the strict order $<$ and $\rho_{3}$ is $\not \leq $ the negation of the order relation on $P$. Since $C_1$ is a barrier, Nash-Williams ’s partition theorem (Theorem \[partnash\] $(b)$) asserts that one of these parts contains a barrier $D$. Let $X$ be an infinite subset of $\bigcup C_1$ such that $D=C_1\cap [X]^{<\omega}$. The inclusion $D\subseteq D_{2}$ is impossible. Otherwise, let $s\in B_{1}$ such that $s\leq_{in} X$, set $Y:= X\setminus s_{}$ and set $g(a):= f(s_{}\cdot (a))$ for $a\in Y$. Then $g$ is a bad map from $X$ into $P$. This contradicts the fact that $P$ is wqo. The inclusion $D\subseteq D_{0}$ is also impossible. Otherwise, set $B':= \{s': s'\cdot (a)\in B_1\cap [X]^{<\omega} $ for some $a\}$. For $s'\in B'$, set $f'(s'):=f(s'\cdot(a))$ where $a\in X$. In this case $f'(s')$ is well-defined. Since $P$ is wqo and $f$ is bad, the order type of $B_{1}$ is at least $\omega^2$, hence $B'$ is a barrier. The map $f'$ satisfies $f'\leq_{in}f$. According to Fact \[fact3\] $(c)$, the minimality of $f$ implies that $f'$ is the restriction of $f$ to $B'$. Since $B'$ is not included into $B$ this is it not the case. A contradiction. Thus we have $D\subseteq D_{1}$. Set $B':= B_1\cap [X]^{<\omega}$. Then (\[eqendclosed2\]) holds. An application to $S_{\omega}(P)$ --------------------------------- We deduce Theorem \[nonprincipal\] from the equivalence $(i)\Longleftrightarrow (ii)$ in the following result. Without clause $(ii)$, the result is due to A.Marcone [@marcone2001]. Without Marcone’s result our proof only shows that under clause $(ii)$ $P$ is $\alpha$-bqo. This suffices to prove Theorem \[nonprincipal\] but the result below is more precise. \[nonprincipal2\] Let $\alpha$ be a denumerable ordinal and $P$ be a poset. Then the following properties are equivalent: 1. $P$ is $\alpha \omega$- bqo; 2. $P$ is $\omega$-bqo and $S_{\omega} (P)$ is $\alpha$-bqo. 3. $\mathfrak{P}_{\leq \omega}(P)$ is $\alpha$-bqo 4. $\mathfrak{P}(P)$ is $\alpha$-bqo $(i)\Rightarrow (iv)$. Let $B$ be a barrier with order type at most $\alpha$ and $f: B\rightarrow \mathfrak {P}(P)$. If $f$ is bad, let $f': B'\rightarrow P$ where $B':= B^2$ and $f'(s\cup t)\in f(s)\setminus {\downarrow\hskip -2pt \hskip -2pt f(t)}$. (See Equation \[downarrow\].) This map $f'$ is bad and the order type of $B'$ is at most $\alpha\omega$. $(iv)\Rightarrow (iii)$ Trivial. $(iii)\Rightarrow (ii)$ $P$ and $S_{\omega} (P)$ identify to subsets of $\mathfrak{P}_{\leq \omega}(P)$, hence are $\alpha$-bqo. $(ii)\Rightarrow (i)$ Induction on $\alpha$. Suppose that $P$ is not $\alpha\omega$-bqo. Let $\beta$ be the smallest ordinal such that $P$ is not $\beta $-bqo. Then $\beta\leq \alpha \omega$. [Case 1]{}. $\beta= \alpha'\omega$. According to Marcone [@marcone2001] the implication $(iii)\Rightarrow (i)$ holds for all denumerable ordinals, hence there is a bad map $f':B'\rightarrow \mathfrak{P}_{\leq \omega}(P)$ for which $B'$ is a barrier of type at most $\alpha'$. Let $X\in \mathfrak{P}_{\leq \omega}(P)$. Since $P$ is wqo, $\downarrow\hskip -2pt \hskip -2pt X$ is a finite union of ideals according to a theorem of Erdös and Tarski (1943), see [@fraissetr]. Hence there are a finite antichain $A_X$ and a finite set $B_X$ of strictly increasing sequences such that $\downarrow\hskip -2pt \hskip -2pt X=\downarrow\hskip -2pt \hskip -2pt A_X \cup \downarrow\hskip -2pt \hskip -2pt B_X$. Let $g: B'\rightarrow \mathfrak {P}_{<\omega}(P)\times \mathfrak{P}_{<\omega}(S_{\omega}(P))$ defined by $g(s')= (A_f'(s'), B'_f(s'))$. This map is bad. Hence, from $(b)$ of Lemma \[higbqo\] there is a bad map from a subbarrier $B''$ of $B'$ into $P$ or into $S_{\omega}(P)$. The latter case is impossible since $S_{\omega}(P)$ is $\alpha$-bqo and so is the former case according to the the induction hypothesis. [Case 2]{}. Case 1 does not hold, that is $\beta= \omega^{\gamma}$ where ${\gamma}$ is a limit ordinal it follows that $\beta\leq \alpha$. Let $f: B\rightarrow P$ be a bad map where $B$ is a barrier of type $\beta$. According to Lemma \[minimalbad\] there is a minimal $f': B'\rightarrow P$ with $\bigcup B'=\bigcup B$ and $f'\leq_{end} f$ and according to Fact \[fact3\] $(e)$ the map $f'$ is bad. Since $P$ is wqo, Lemma \[key\] applies. Thus $B'$ contains a subbarrier $B''$ on which $s\leq_{end }t$ implies $f'(s)< f'(t)$. Let $F: B''_{} \rightarrow P$ be given by $F(s'):= \downarrow\hskip -2pt \hskip -2pt \{f'(t)\in P: t\in B''$ and $s'\leq _{in} t \}$. [Claim 1]{} $F(s')$ is a finite union of non-principal ideals of $P$. Since $P$ is wqo, every initial segment is a finite union of ideals. Hence in order to show that $F(s')$ is a finite union of non-principal ideals it suffices to show that it contains no maximal element. Let $x \in F(s')$. Let $t \in B''$ such that $s'\leq_{in} t$, $f''(t)=x$. Let $u\in B''$ such that $t<_{end}u$. Then $s'\leq_{in}u$ hence $f''(u)\in F(s')$. From Lemma \[key\] $f''(t)<f(u)$, proving our claim. [Claim 2* ]{} $F$ is good. Indeed, since $S_{\omega}(P)$ is $\alpha$-bqo, it follows from $(e)$ of Lemma \[higbqo\] that the collection of finite unions of its members is $\alpha$-bqo. Hence $f'$ is good. Indeed, since $F$ is good, there are $s', t'\in B''_{}$ such that $s' \lhd t'$ and $F(s')\subseteq F(t')$. Let $a:=t'(l(s')-1)$ then $s'\leq_{in} s:= s'.(a)\in B''$ then $f'(s'.(a))\in F(s')$. Since $F(s')\subseteq F(t')$ there is some $t\in B''$ such that $t'\leq_{in }t$ and $f'(s)\leq f'(t)$. Because $s\lhd t$ the map $f'$ is good. This contradicts the hypothesis that $f'$ is bad and finishes the proof of the theorem. The set of ideals of a well-quasi-ordered-poset {#sec:ideals} =============================================== In this section, we illustrate the relevance of the notion of ideal w.r.t. well-quasi-ordering. Define a topology on $\mathfrak{P}(P)$. A basis of open sets consists of subsets of the form $O(F,G):=\{X\in \mathfrak{P}(P): F\subseteq X \text{ and } G\cap X=\emptyset \}$, where $F, G$ are finite subsets of $P$. The topological closure of $down(P)$ in $\mathfrak{P}(P)$ is a Stone space which is homeomorphic to the Stone space of $Tailalg (P)$, the Boolean algebra generated by $up(P)$ . With the order of inclusion added the closure of $down(P)$, $\bigcup{down(P)}$, is isomorphic to the Priestley space of $Taillat (P)$ [@bekkali-pouzet-zhani]. Note that $\bold{I}_{ <\omega}(P)$ is the set of compact elements of $\bold{I}(P)$, hence $\mathcal{J}(\bold{I}_{ <\omega}(P))\cong \bold{I}(P)$. We also note that $\mathcal{J}(P)$ is the set of join-irreducible elements of $\bold{I}(P)$. We have \[emptyset\] $\emptyset \not \in \bigcup {down(P)}\Longleftrightarrow P\in \bold{F}_{<\omega}(P)$. \[ideals1\] $down(P)\subseteq \mathcal {J}(P)\subseteq \bigcup {down (P)}\setminus\{\emptyset\}$. In particular, the topological closures in $\mathfrak{P}(P)$ of $down(P)$ and $\mathcal {J}(P)$ are the same. A poset $P$ is [up-closed]{} if every intersection of two members of $up(P)$ is a finite union (possibly empty) of members of $up(P)$. \[ideals\] The following properties for a poset $P$ are equivalent: 1. $\mathcal {J}(P)\cup \{\emptyset \}$ is closed for the product topology; 2. $\mathcal {J}(P)=\bigcup {down(P)}\setminus \{\emptyset\}$; 3. $P$ is up-closed; 4. $\bold {F}_{<\omega}(P)$ is a meet-semi-lattice; 5. $Taillat(P)=\bold{F}_{<\omega}(P)\cup \{P\}$. Let us recall that a topological space $X$ is [scattered]{} if every non-empty subset $Y$ of $X$ contains an isolated point with respect to the topology induced on $Y$. We have: Let $P$ be a poset. If $P$ is well-quasi-ordered then ${\mathcal J}(P)$ is a compact scattered space whose set of isolated points coincides with $down(P)$. [Claim 1]{} $\mathcal {J}(P)=\bigcup {down(P)}$. Indeed, since $P$ is wqo it is up-closed. Hence, from Proposition \[ideals\], $\mathcal {J}(P)=\bigcup {down(P)}\setminus \{\emptyset\}$. Again, since $P$ is wqo, $P\in \bold{F}_{<\omega}(P)$. Hence, from Lemma \[emptyset\], $\emptyset \not \in \bigcup {down(P)}$ proving that $\mathcal {J}(P)=\bigcup {down(P)}$, as claimed. [Claim 2]{} As a subspace of the Cantor space $\mathfrak {P}(P)$, $\bold I(P)$ are compact and scattered. $\bold I(P)$ is closed. To see that it is scattered, let $X$ be a non-empty subset of $\bold {I}(P)$. Since $P$ is wqo, $\bold {I}(P)$ is well-founded. Select a minimal element $I$ in $X$. Let $G:= \min(P\setminus I)$ and $O(G, \emptyset):= \{I'\in \bold {I}(P): G\cap I'=\emptyset \} $. Since $P$ is wqo, $G$ is finite, hence $O(\emptyset, G)$ is a clopen subset of $\mathfrak {P}(P)$. Since $O(\emptyset, G)\cap X= I$, $I$ is isolated in $X$. [Claim 3]{} Let $J\in \mathcal J(P)$, then $J$ is isolated in $\mathcal J(P)$ if and only $J$ is principal. Suppose that $J$ is isolated. Then there is a clopen set of the form $O(F,G)$ such that $O(F,G)\cap \mathcal J(P)=\{J\}$. Since $J$ is up-directed, there is some $z$ in $J$ which majorises $F$. Clearly, $\downarrow\hskip -2pt z \in O(F,G)\cap \mathcal J(P)$, hence $J=\downarrow\hskip -2pt z$, proving that $J$ is principal. Conversely, let $z\in P$. Let $G:=\min( P \setminus \downarrow\hskip -2pt z)$. Since $P$ is wqo, $G$ is finite. Hence $O(\{z\}, G)$ is a clopen set. It contains only $\downarrow\hskip -2pt z$, proving that $\downarrow\hskip -2pt z$ is isolated. From this result, $\mathcal J^{\neg\downarrow\hskip -2pt }:= \mathcal J(P)\setminus down(P)$, the set of non-principal ideals of $P$, coincides with $ \mathcal J^1 {(P)}$, the first derivative of $\mathcal J(P)$ in the Cantor-Bendixson reduction procedure. Our main result establishes a link between the bqo characters of $\mathcal J(P)$ and $ \mathcal J^1 {(P)}$. This suggests to look at the other derivatives. Minimal type posets {#sec:min} ==================== Well-quasi-ordered posets with just one non-principal ideal are easy to describe. Each is a finite unions of ideals, all but one being finite. The infinite one, called a [minimal type]{} poset, can be characterized in several ways: Let $P$ be an infinite poset. Then, the following properties are equivalent: 1. $P$ is wqo and all ideals distinct from $P$ are principal; 2. $P$ has no infinite antichain and all ideals distinct from $P$ are finite; 3. Every proper initial segment of $P$ is finite. 4. Every linear extension of $P$ has order type $\omega$. 5. $P$ is level-finite, of height $\omega$, and for each $n<\omega$ there is $m<\omega$ such that each element of height at most $n$ is below every element of height at least $m$. 6. $P$ embeds none of the following posets: an infinite antichain; a chain of order type $\omega ^{dual} $; a chain of order type $\omega +1$; the direct sum $\omega\oplus 1$ of a chain of order type $\omega$ and a one element chain. An easy way of obtaining posets with minimal type is given by the following corollary Let $n$ be an integer and $P$ be a poset. The order on $P$ is the intersection of $n$ linear orders of order type $\omega$ if and only if $P$ is the intersection of $n$ linear orders and $P$ has minimal type. Minimal type posets occur quite naturally in symbolic dynamic. Indeed, let $S: A^{\omega}\rightarrow A^{\omega}$ be the shift operator on the set $A^{\omega}$ of infinite sequences $s:= (s_{n})_{n<\omega}$ of members of a finite set $A$ (that is $S(s):= (s_{n+1})_{n<\omega}$). A subset $F$ of $A^{\omega}$ is [invariant]{} if $S(F)\subseteq F$. As it is well-known, every compact (non-empty) invariant subset contains a minimal one. To a compact invariant subset $F$ we may associate the set $\mathcal A(F)$ of finite sequences $s:= (s_{0}, \dots, s_{n-1})$ such that $s$ is an initial segment of some member of $F$. Looking as these sequences as words, we may order $\mathcal A(F)$ by the factor ordering: a sequence $s$ being a [factor]{} of a sequence $t$ if $s$ can be obtained from $t$ by deleting an initial segment and an end segment of $ t$. We have then $\mathcal A(F)$ has minimal type if and only if $F$ is a minimal compact invariant subset. Maximal antichains, “pred” and “succ” {#sec:pred} ===================================== Let $P$ be a poset. We consider both $A(P)$ and $AM(P)$ to be ordered by domination. The main result of this section will be that if $AM(P)$ $\alpha$-bqo then $P$ is $\alpha$-bqo. Our first aim is to prove that if $AM(P)$ is well founded then $P$ is well founded. To this end we will associate with every element $x\in P$ an antichain $\varphi(x)$ and investigate the connection between $x$ and $\varphi(x)$. First the following: \[lem:1\] Let $P$ be a poset and $X\subseteq A(P)$. Then the following properties are equivalent: 1. $X$ is the minimum of $AM(P)$. 2. $X$ is a minimal element of $AM(P)$. 3. $P=\uparrow\hskip-2pt\hskip -2pt X$. Implications $(iii)\Rightarrow (i)\Rightarrow (ii)$ are obvious. $(ii)\Rightarrow (iii)$: Suppose for a contradiction that $P\setminus \uparrow\hskip-2pt X\not=\emptyset$. For every element $z\in P\setminus \uparrow\hskip-2pt X$ there is an element $x\in X$ with $z<x$. Let $Y$ be a maximal antichain of the set $P\setminus \uparrow\hskip-2pt X$ and $Z=X\setminus \uparrow\hskip-2pt Y$. Then $Z\cup Y$ is a maximal antichain which is strictly dominated by $X$. Let $P$ be a poset. Let $\mathcal{P}$ be the order that is the set of pairs $\mathcal{P}:=\{(x,y) : x\leq y\}$. For $S\subseteq P$ let $\min S:=\{x\in S : y\in S \text{ and } y\leq x \text{ implies } y=x\}$ and for $x\in P$ let $\Phi(x):=\{z\in P : z\not< x\}$ and let $\varphi (x):= \min\Phi(x)$. Note that $x\in \varphi(x)$. \[lem:2\] $x<y \Longleftrightarrow \varphi(x)<\varphi (y) \text{ and }x\not\in \varphi(y)$ for all $x,y\in P$. Suppose $x<y$, then $\Phi(y) \subset \Phi(x)$. Hence $\varphi (y):=\min \Phi(y)\leq_{dom} \min\Phi(x)=:\varphi(x) $. ($\leq_{dom}$ is the domination order.) Then $\varphi(x)<\varphi(y)$ because $\varphi(x)\ni x \notin \varphi(y)$. Conversely, suppose $\varphi(x)<\varphi (y)$ and $x\not \in \varphi (y)$. If $x\not < y$, then the definition of $\varphi (y)$ insures that $x'\leq x$ for some $x'\in \varphi(y)$. Since $\varphi(x)<\varphi (y)$, we have $x=x'$ proving $x\in \varphi(y)$, a contradiction. Note that if $\varphi(x)$ is a maximal antichain for every $x\in P$ and $AM(P)$ is well founded then we obtain, using Lemma \[lem:2\], that $P$ is well founded. Actually we will show below that if $AM(P)$ is well-founded then $\varphi(x)$ is a maximal antichain. Let $x\in P$ and the set $\mathcal F\subseteq \mathfrak{P}(P)$. Then $\mathcal F(x):= \{F\in \mathcal F: x\in F\}$ and $Inc_{P}(x):= \{y\in P: x \text{ and } y \text{ are incomparable}\}$. Note that $AM(Inc_{P}(x))\cong AM(P)(x)$. ($\cong$ is the order isomorphism between he maximal antichains of $Inc_{P}(x)$ and the maximal antichains of $AM(P)(x)$ both ordered under domination.) This implies that if $AM(P)$ is well-founded then $AM(\Phi(x))$ is well-founded. Let $S$ be a minimal element of $AM(\Phi(x))$. It follows then from Lemma \[lem:1\] that $S$ is the minimum of $AM(\Phi(x))$ and $\uparrow\hskip-2pt S=\Phi(x)$. Hence $S=\phi(x)$. That is: $\varphi(x)$ is the least maximal antichain of $P$ containing $x$. Also, if $P$ is well-founded then $\Phi(x)$ is well-founded, hence $\uparrow\hskip-2pt \varphi(x)=\Phi(x)$ which in turned implies using Lemma \[lem:1\] that $\varphi(x)$ is the least maximal antichain of $P$ containing $x$. Hence we established the following Lemma: \[lem:3\] If $AM(P)$ is well-founded or if $P$ is well founded then for all $x,y\in P$: 1. $\uparrow\hskip-2pt \varphi(x)=\Phi(x)$. 2. $\varphi(x)$ is the minimum of all maximal antichains of $P$ containing $x$. 3. $x<y \Longleftrightarrow \varphi(x)<\varphi (y)$ and $x\not\in \varphi(y)$. 4. $P$ is well founded. Associated with the quasi order $(P;\leq_{pred})$ is the equivalence relation $\equiv$ equal to the set $\{(x,y) : x\leq_{pred} y \text{ and } y\leq_{pred}x\}$. Let $(P;\leq_{pred})/\equiv $ be the quotient equipped with the order induced by $pred(P)$. Let $\pi$ be the canonical map of $(P;\leq_{pred})$ to $(P;\leq_{pred})/\equiv $. For every subset $S$ of $P$ let $\overline{\pi}(S):=\{\downarrow\hskip-2pt p(s); s\in S\}$. \[lem:AMembed\] [ ]{} 1. The function $\overline{\pi}$ induces an embedding of $AM(P)$ into $\mathbf{I}\bigl((P;\leq_{pred})/\equiv\bigr) $ and if $P$ is well founded then $(P;\leq_{pred})/\equiv $ embeds into $AM(P)$. 2. $\overline{\pi}$ induces an embedding of $\mathcal J^{\neg \downarrow\hskip -2pt } (P)$ into $\mathcal J^{\neg \downarrow\hskip -2pt }((P;\leq_{pred})/\equiv)$. 3. If $P$ has no infinite antichain then this embedding is surjective, hence $\mathcal J^{\neg \downarrow\hskip -2pt } (P)\cong \mathcal J^{\neg \downarrow\hskip -2pt }((P;\leq_{pred})/\equiv )$. Let $A,B\in AM(P)$ with $\overline{\pi}(A)\subseteq \overline{\pi}(B)$. For every $a\in A$ there is an element $b\in B$ so that $a$ and $b$ are related under $\leq$. Assume for a contradiction that $b<a$. Then $\pi(b)<\pi(a)$. Because $\overline{\pi}(A)\subseteq \overline{\pi}(B)$ there is a $c\in B$ with $\pi(a)<\pi(c)$. This implies, because $b<a$, that $b<c$ a contradiction. Hence $\overline{\pi}(A)\subseteq \overline{\pi}(B)$ implies that $A$ is less than or equal to $B$ in the domination order which in turn implies that if $\overline{\pi}(A)= \overline{\pi}(B)$ then $A=B$. If $A$ is less than or equal to $B$ in the domination order then $\overline{\pi}(A)\subseteq \overline{\pi}(B)$ and hence we conclude that $\overline{\pi}$ is an embedding of $AM(P)$ into $\mathbf{I}\bigl((P;\leq_{pred})/\equiv\bigr) $. We have: $x\leq_{pred (P)}y$ if and only if $\Phi(x)\supseteq \Phi(y)$ if and only if $\varphi(x)\leq \varphi (y)$ in $AM(P)$. Hence $x\leq_{pred} y \text{ and } y\leq_{pred}x$, that is $x\equiv y$, if and only if $\varphi(x)=\varphi(y)$. This establishes item 1. In order to establish item 2 let $J,J^\prime\in \mathcal J^{\neg \downarrow\hskip -2pt } (P)$. If $J\subseteq J'$ then clearly $\overline{\pi}(J)\subseteq \overline{\pi}(J')$. The functions $\pi$ and $\overline{\pi}$ are order-preserving. Hence $J\in \mathcal J(P)$ implies $\overline{\pi}(J)\in \mathcal J(Q)$. Since $\pi$ is strictly increasing $\overline{\pi}(J)\in \mathcal J^{\neg \downarrow }(Q)$ if and only if $J\in \mathcal J^{\neg \downarrow }(P)$. Suppose $J\not \subseteq J'$. Let $x\in J\setminus J'$. Since $J$ is not principal, there is some $x'\in J$ such that $x<x'$. Since $J'$ is an intial segment, $x'\not \in J'$. Assume for a contradiction that $\pi(x') \in \overline{\pi}(J')$. Then there is an $x''\in J^\prime$ such that $\pi(x')\leq \pi(x'')$ hence $x'\leq_{pred} x''$. Therefore $x<x''$ follows from $x<x'$. Since $J'$ is an initial segment, we have $x\in J'$ contradicting the choice of $x$. This proves that $\pi(J)\not \subseteq \overline{\pi}(J')$. Hence $\overline{\pi}$ is an embedding. Item 3: Let $K\in \mathcal J^{\neg \downarrow\hskip -2pt } (Q)$. Let $J:= \{ x\in P: \pi(x)\in K\}$. The set $J$ is an initial segment of $P$ since $\pi$ is order preserving. The set $J$ is a finite union of ideals since $P$ has no infinite antichain; see Fact \[fact:union\]. Let $J:=J_{1}\cup \cdots \cup J_{k}$. We have $K= \overline{\pi}(J)= \overline{\pi}(J_1) \cup \cdots \cup \overline{\pi}(J_{k})$. From the fact that $K$ is an ideal it follows that $K= \overline{\pi}(J_i)$ for some $i$. Since $K$ is not principal, $J_i$ cannot be principal. We derive Theorem \[criticbqo\] from the following result: \[thmsucc\] Let $P$ be a poset with no infinite antichain and $\alpha$ be a countable ordinal . The following properties are equivalent: 1. $P$ is $\alpha$-bqo. 2. $(P;\leq_{succ})$ is $\alpha$-bqo. 3. $(P;\leq_{pred})$ is $\alpha$-bqo. 4. $(P; \leq_{crit})$ $\alpha$-bqo. 5. $AM(P)$ is $\alpha$-bqo. Implications $(i)\Longrightarrow (iv)\Longrightarrow (iii)$ follow from the sequence of inclusions $(\leq)\, \, \subseteq\, \, (\leq_{crit})\, \subseteq (\leq_{pred})$ and the implication $(iv)\Longrightarrow (ii)$ follows from the inclusion $( \leq_{crit})\, \subseteq\, \, ( \leq_{succ(P)})$. $(iii)\Longleftrightarrow (v)$ Suppose $AM(P)$ is $\alpha$-bqo then $P$ is well-founded according to Lemma \[lem:3\] and hence according to item 1 of Theorem \[lem:AMembed\], $(P;\leq_{pred})/\equiv$ embeds into $AM(P)$ implying that $(P;\leq_{pred})$ is $\alpha$-bqo. Conversely, suppose $(P;\leq_{pred})$ is $\alpha$-bqo. Then ${\bf I}_{<\omega}(P;\leq_{pred})/\equiv)$ is $\alpha$-bqo according to item 3 of Theorem \[lem:AMembed\]. From item 1 of Theorem \[lem:AMembed\], the poset $AM(P)$ embeds into ${\bf I}_{<\omega}{<\omega}(P;\leq_{pred})/\equiv)$ and hence is $\alpha$-bqo. We prove implications $(ii)\Longrightarrow (i)$ and $(iii)\Longrightarrow (i)$ in a similar way as we have proven the implication $(ii)\Longrightarrow (i)$ in Theorem \[nonprincipal2\]. Induction on $\alpha$. Let $Q$ be equal to $(P;\leq_{succ})$ or equal to $ (P;\leq_{pred})$. Suppose that $Q$ is $\alpha$-bqo. Since $(\leq)\, \subseteq\, (\leq_{ succ}\, \cap\, \leq_{pred})$, the partial order $P$ is well-founded and since it has no infinite antichain it is wqo. If $P$ is not $\alpha$-bqo there is a barrier $B$ of type at most $\alpha$ and a bad map $f :B\rightarrow P$. From Lemma \[minimalbad\] there is a minimal $f': B'\rightarrow P$ such that $\bigcup B'=\bigcup B$ and $f'\leq_{end} f$. According to Fact \[fact3\] $(e)$ the map $f'$ is bad. Since $P$ is wqo, Lemma \[key\] applies. Thus $B'$ contains a subbarrier $B''$ on which $$\label{eq1} s<_{end }t \Longrightarrow f'(s)< f'(t).$$ Suppose $Q:=(P;\leq_{succ})$. Since $(P;\leq_{pred})$ is $\alpha$-bqo, $f'$ cannot be bad thus there are $s, t \in B''$ such that $s\lhd t$ and $f'(s)\leq_{succ(P)} f'(t)$. Pick $t'\in B''$ such that $t<_{end}t'$. From (\[eq1\]) we have $f'(t)< f'(t')$ . According to the definition of $(P;\leq_{succ})$, we have $f'(s)\leq f'(t')$. Since $s\lhd t'$ it follows that $f'$ is good for $P$. A contradiction. Suppose $Q:= (P;\leq_{pred})$. For $s\in B''$ set $s^{+}:= s_{}.(a)$ where $a$ is the successor of $\lambda(s)$ in $\bigcup{ B''}$. Set $f'^{+}(s):= f'(s^{+})$. Since $(P;\leq_{pred})$ is $\alpha$-bqo there are $s$ and $t$ such that $s\lhd t$ and $f'^{+}(s)\leq_{pred}f'^{+}(t)$, that is $f'(s^{+})\leq_{pred} f'(t^{+})$. Since $s<_{end}s^{+}$ we have $f'(s)<f'(s^{+})$. According to the definition of $(P;\leq_{pred})$, this gives $f'(s) < f'(t^{+})$. Since $s\lhd t^{+}$, the function $f'$ cannot be bad. A contradiction. Maximal antichains with a prescribed size {#maximal} ========================================= Two element maximal antichains ------------------------------ Let $P$ be a poset. The structure $(P(2);\leq)$ is defined on $P(2):=P\times 2$ so that: $$(x,i)\leq (y,j) \text{ if } \begin{cases} i=j & \text{ and $x\leq y$, or } \\ i=0 \text{ and } j=1 & \text{ and there exist incomparable elements }\\ &\text{ $x^\prime, y^\prime\in P$ with $x\leq x^\prime$ and $y^\prime\leq y$ }. \end{cases}$$ It is easy to see that $P(2)$ is a poset. \[construct1\]Every poset $P$ embeds into the poset $AM_2(P(2))$. [Claim 1.]{} If $y\leq x$ then $(x,0)$ and $(y,1)$ are incomparable in $P(2)$. The converse holds if for every $x <y$ there are two incomparable elements $x', y'$ such that $x\leq x'$ and $y'\leq y$. If $(x,0)$ and $(y,1)$ are comparable then necessarily $(x,0)<(y,1)$. In this case there are two incomparable elements $x',y'$ such that $x\leq x'$ and $y'\leq y$. But if $y\leq x$, we get $y'\leq x'$, a contradiction. Conversely, suppose that $(x,0)$ and $(y,1)$ are incomparable. Then clearly, $x$ and $y$ are comparable. Necessarily, $x\leq y$. Otherwise $x<y$. But, from the condition stated, we have $(x,0)\leq (y,1)$, a contradiction. For $x\in V$, set $X_{x}:=\{(x, 0),(x,1)\}$. [Claim 2.]{} $X_{x}\in AM_2(P(2))$. The set $X_x$ is an antichain according to Claim 1. Moreover, every element $(x', i')$ different from $(x,0)$ and $(x,1)$ is comparable to one of these two elements. Indeed, if $x'$ is comparable to $x$, then $(x',i')$ is comparable to $(x,i')$. If $x'$ is incomparable to $x$ then $(x',i')$ is comparable to $(x, \neg i')$ where $\neg i' \not =i'$. This proves that $X_x$ is maximal. [Claim 3. ]{} The map $x \rightarrow X_{x}$ is an embedding of $P$ into $AM_{2}(P(2))$. That is: $$x\leq y \Longleftrightarrow X_{x}\leq X_{y}.$$ Suppose $x\leq y$. Then we have $(x,0)\leq (y, 0)$ and $(x,1)\leq (y,1)$ proving $X_{x}\leq X_{y}$. Conversely, suppose $X_{x}\leq X_{y}$, that is $(x,0)\leq(y,i)$ and $(x,1)\leq (y, j)$ for some $i,j\in \{0,1\}$. Due to our ordering, we have $j=1$, hence $x\leq y$ as required. With this construction, a poset $P$ which is not $\alpha$-bqo but is $\beta$-bqo for every $\beta<\alpha$ leads to a poset $Q$ having the same property and for which neither $AM_2(Q)$ nor $\bigcup {AM_2(P)}$ is $\alpha$-bqo. The simplest example of this situation is given below. \[thm:222\] Let $P$ be a poset with no infinite antichain, and $\alpha$ be a denumerable ordinal, then $AM_2(P)$ is $\alpha$-bqo if and only if $\bigcup {AM_2(P)}$ is $\alpha $-bqo. If $Q:=\overline {AM_2(P)}$ is $\alpha $-bqo then $A(Q)$ is $\alpha$-bqo. In particular, $AM_2(Q)$ is $\alpha $-bqo. This set is simply $AM_2(P)$ and the conclusion follows. For the converse, we prove a bit more. Let $\mathcal P$ be a subset of $[P]^{2}$. We quasi-order $\mathcal P$ as follows: $X\leq Y$ if for every $ x\in X$ there is some $y\in Y$ such that $x\leq y$ and for every $y\in Y$ there is some $x\in X$ such that $x\leq y$. Let $T$ be a subset of $P':= \cup \mathcal P$. [Claim]{} If $\mathcal P$ is $\alpha$-bqo then $T$ is $\alpha$-bqo. Let $f: B\rightarrow T$ be a map from a barrier $B$ of type at most $\alpha$ into $T$. For each $s\in B$, select a map $F(s): 2:=\{0,1\}\rightarrow P$ such that $f(s)\in rg(F(s))\in \mathcal P$. For $s\in B$, set $p(s):=i$ if $ F(s)(i)=f(s)$ and for $(s,t)\in B\times B$, set $\rho_{(s,t)}:=\{(i,j)\in 2\times 2: F(s)(i)\leq F(t)(j)\}$. Note that since an order is transitive, for $s,t, u \in B$ the composition of relations satisfies $$\label{eqcomp} \rho_{(t ,u)}\circ \rho_{(s ,t)} \subseteq \rho_{(s ,u)}$$ [Subclaim 1]{} We may suppose that: 1. $p(s)=i_0$ for all $s\in B$ and some $i_0\in 2$; 2. $\rho_{(s,t)}= \rho$ for all pairs $(s,t)\in B\times B$ such that $s\lhd t$ and some $\rho\subseteq 2\times 2$; 3. for every $i\in 2$ there are some $j,j'\in 2$ such that $(i,j), (j',i)\in \rho$ [Proof of Subclaim 1]{} Since the map $p$ takes only two values, we get from the partition theorem of Nash-Williams $p$ is constant on a subbarrier of $B$. With no loss of generality, we may suppose that this barrier is $B$ proving that $1$ holds. Similarly, the map which associate $\rho_{(s,t)}$ to each element $s\cup t\in B^2$ takes only finitely many values hence, by the same token, this map is constant on a subbarrier $C$ of $B^2$. Necessarily $C= B^{2}\cap [X]^{<\omega}$for some $X\subseteq \overline B$. For $B':=B\cap [X]^{<\omega}$ the condition stated in $2$ holds. We may suppose $B'= B$. Finally, since $\mathcal P$ is $\alpha$-bqo, the map which associates $rg(F(s)$ to $s\in B$ cannot be bad. According to the partition theorem of Nash-Williams this map is perfect on a subbarrier . With no loss of generality, we may suppose this subbarrier equals to $B$. From this $3$ follows. We will prove that $\rho$ is reflexive. With conditions $1$ and $2$ it follows that $f$ is perfect, proving our claim (indeed, let $s \lhd t$. From $1$, $f(s)=F(s)(i_0)$ and $f(t)=F(t)(i_0)$, from $2$ $\rho_{(s,t)}=\rho$. The reflexivity of $\rho$ insures that $(i_0,i_0)\in \rho$ that is $(i_0,i_0)\in\rho_{(s,t)}$ which amounts to $F(s)(i_0)\leq F(t)(i_0)$. This yields $f(s)\leq f(t)$ as required). If $\rho$ is not reflexive, then it follows from condition $3$ that $\{(0,1), (1,0)\}\subseteq \rho$. From now, on we will suppose this later condition fulfilled. We say that two elements $s_0, s_1 \in B$ are [intertwined]{} and we set $s_0\lhd_{\frac{1}{2}} s_1 $ if there is an infinite sequence $X:= a_0< \dots a_n <\dots $ of elements of $\overline B$ such that $s_0<_{init} X_{even}$ and $s_1<_{init} X_{odd}$, where $X_{even}:= a_0< \dots a_{2n} <\dots $ and $X_{odd}:= a_1< \dots a_{2n+1}<\dots$. We set $B^{(\frac{1}{2})}:= \{(s_0,s_1): s_0\lhd_{\frac{1}{2}} s_1 \}$ and $B^{\frac{1}{2}}:= \{s_0\cup s_1: (s_0,s_1)\in B^{(\frac{1}{2})} \}$ where $s_0\cup s_1$ denotes the sequence $w$ whose range is the union of the ranges of $s_0$ and $s_1$. We note that 1. if $w\in B^{\frac{1}{2}}$ then the pair $(s_0,s_1)\in B^{(\frac{1}{2})}$ such that $w= s_0\cup s_1$ is unique; 2. $B^{\frac{1}{2}}$ is a thin block; 3. if $X:= a_0< \dots a_n <\dots $ is an infinite sequence of elements of $\overline B$, $Y:= _{}X$, $s_0, s_1, s_2 \in B$ such that $s_0<_{init} X_{even}$, $s_1<_{init} X_{odd}$, $s_2<_{init}Y_{odd}$ then $s_0\lhd_{\frac{1}{2}} s_1 \lhd_{\frac{1}{2}} s_2$ and $s_0\lhd s_2$. Let $w:= s_0\cup s_1, w':= s'_0\cup s'1\in B^{\frac{1}{2}}$. We say that $w$ and $w'$ are equivalent if there is a map $g$ from $\cup \{ rg(F(s_i)): i<2\}$ onto $\cup \{rg(F(s'_i)): i<2\}$ such that 1. $g\circ F(s_i)=F(s'_i)$ for $i<m$; 2. $\rho_{(s_i,s_j)}=\rho_{(s'_i,s'_j)}$ for all $i, j<2$; As one can check easily, this is an equivalence relation on $B^{\frac{1}{2}}$. Furthermore, the number of equivalence classes is finite (one can code each equivalence classe by a relational structure on a set of at most $4$ elements, this structure been made of four binary relations and four unary relations). Since $B^{\frac{1}{2}}$ is a block, it follows from the partition theorem of Nash-Williams that one class contains a barrier. Let $C$ be such a barrier, $X\subseteq \overline B$ such that $C:= B^{\frac{1}{2}}\cap [X]^{<\omega}$ and let $B':=B\cap [X]^{<\omega}$. For $ s_0, s_1\in B'$ such that $s_0\lhd_{\frac{1}{2}} s_1 $, $\rho_{ s_0, s_1}$ and $\rho_{ s_1, s_0}$ are constant; let $\rho_{\frac{1}{2}}$ and $\rho'_{\frac{1}{2}}$ their common value. For the proof of the next subclaims, we select $s_0, s_1, s_2 \in B'$ such that $s_0\lhd_{\frac{1}{2}} s_1 \lhd_{\frac{1}{2}} s_2$ and $s_0\lhd s_2$; according to condition 3 above this is possible. [Subclaim 2]{} $\rho_{\frac{1}{2}}\circ \rho_{\frac{1}{2}}\subseteq \rho$ [Proof of Subclaim 2]{} We have $\rho_{\frac{1}{2}}= \rho_{s_0,s_1}= \rho_{s_1,s_2}$ and $\rho= \rho_{s_0,s_2}$. The claimed inclusion follows from the composition of relations. [Subclaim 3]{} $\rho'_{\frac{1}{2}}=\emptyset$ [Proof of Subclaim 3]{} Suppose the contrary; let $(i,j)\in \rho'_{\frac{1}{2}}$. Case 1. $i=j$. Let $k\not = i$. We have $(k,i) \in \rho= \rho_{s_0,s_2}$ and $(i,i)\in\rho'_{\frac{1}{2}}= \rho_{s_2,s_1}= \rho_{s_1,s_0}$. By composing these relations, we get with \[eqcomp\] $(k,i)\in \rho_{s_0,s_0}$ contradicting the fact that $rg(F(s_0))$ is an antichain. Case 2. $i\not= j$. then from $(i,j)\in\rho'_{\frac{1}{2}}= \rho_{s_2,s_1}= \rho_{s_1,s_0}$ and $(j,i)\in \rho= \rho_{s_0,s_2}$ we get, by composing these relations, $(i,j)\in \rho_{s_1,s_1}$ contradicting the fact that $rg(F(s_1))$ is an antichain and proving Subclaim 3. [Subclaim 4]{} $\rho_{\frac{1}{2}}$ satisfies condition $3$ of Subclaim $1$. [Proof of Subclaim 4]{} Since $rg(F(s_0))$ and $rg(F(s_1))$ are two maximal antichains, each element of one is comparable to some element of the other. Since $\rho_{s_1, s_0}=\rho'_{\frac {1}{2}}=\emptyset$, $rg(F(s_0))\leq rg(F(s_1))$ and the result follows. . Now, if $\rho_{\frac{1}{2}}$ is reflexive, it follows from Subclaim 2 that $\rho$ is reflexive and our claim is proved. If $\rho_{\frac{1}{2}}$ is not reflexive then from Subclaim 4 it follows that $\{(0,1), (1,0)\}\subseteq \rho_{\frac{1}{2}}$. With Subclaim 4 this yields $(0,0), (1,1)\in \rho$ that is $\rho$ is reflexive and the proof of our claim is complete. Rado’s poset ------------- Let $V:=\{ (m,n) \in {\mathbb{N}}^2: m<n\}$. We denote by $\leq_R$ the following relation on $V$: $$\label{eqrado}(m,n)\leq_R (m',n')\; \mbox {if either} \; m=m'\; \mbox {and}\; n \leq n' \; \mbox {or}\; n<m'$$ This relation is an order. We denote by $R$ the resulting poset. This poset, discovered by R. Rado [@rado], is at the root of the discovery of bqo’s. R. Rado observed that $R$ is wqo but $I(R)$ is not wqo and has shown that a poset $P$ is $\omega ^2$-bqo if and only if $I(P)$ is wqo. R. Laver [@laver] has shown that a poset $P$ which is wqo, and not $\omega^2$-bqo contains a copy of $R$. Applying the construction given in Lemma \[construct1\] we have: \[counterexample\] The poset $AM_{2}(R(2))$ is wqo but not $\omega^{2}$-bqo. As a union of two wqo posets, $R(2)$ is wqo. Hence $AM(R(2))$ is wqo for the domination order. In particular $AM_2(R(2))$ is wqo. Since $AM_2(R(2))$ embeds $R$, it cannot be $\omega^{2}$-bqo. \[AM(R)\] $\bigcup {AM_{m}(R)}$ is bqo for every integer $m$ and $R$ embeds into $AM(R)$. [ $\, $]{}\ a) $\bigcup {AM_{m}(P)}$ is bqo. Let $m$, $m<\omega$.Then $\bigcup {AM_m(P)}\subseteq \{(i,j): i<m, i<j<\omega\}$. Indeed, let $A\in AM_{m}(P)$ then for each $i, i<m$ there is some $(i,j)\in A$ with $i<j$ (otherwise, add to $A$ an element $(i,j)$ with $j$ large enought). Consequently $\bigcup {AM_m(P)}$ is bqo. b\) $R$ embeds into $AM(R)$. Since $R$ is not $\omega^2$ bqo, $AM(R)$ is not $\omega^2$ bqo (Theorem \[thmsucc\] ). Hence from Laver’s result mentioned above, the poset $R$ embeds into $AM(R)$. For the sake of simplicity we give a direct proof. Set $X_{(0,1)}:=\{(0,1),(1,2)\}$, $X_{(0,n)}:=\{(m,n): m<n\}$ for $n\geq 2$, $X_{(m,n)}:=\{(m',m): m'<m\}\cup \{(m,n)\}$ for $m\geq 1$. One has to check successively that: [Claim 1.]{} $X_{(m,n)}$ is the least antichain in $AM(R)$ which contains $(m,n )$. [Claim 2.]{} $(m,n)\leq (m',n')\Rightarrow X_{(m,n)}\leq X_{(m',n')}$. [Claim 3]{} $m,m'\geq 1$ and $X_{(m,n)}\leq X_{(m',n')}$ imply $(m,n)\leq (m',n')$. Three element maximal antichains -------------------------------- \[construct2\] Let $P:=(V;\leq)$ be a poset. Let $L:=(V;\sqsubseteq)$ be a linear extension of $P$ with the property that if $x<y$ then there is a $z$ with $x\sqsubset z\sqsubset y$ and $z$ is incomparably in $P$ to both $x$ and $y$. Then there is a poset $Q$ which is a union of a copy of $P$ and two copies of $L$ for which $\bigcup {AM_3(Q)}=Q$ and $AM_3(Q)$ is isomorphic to $L$. On $V\times 3$ define the following strict order relation $<_Q$: $$(x,i)<_Q(y,j) \text{ if } \begin{cases} i=j=1 & \text{ and $x < y$, or} \\ 1\not=i \text{ or } j\not=1 & \text{ and $i\leq j$ and $x\sqsubset y$. } \end{cases}$$ Let $Q=(V\times 3;\leq_Q)$ be the resulting poset by adding the identity relation to $<_Q$. The order induced by $\leq_Q$ on $V\times \{i\}$ coincides with the order $\leq$ on $V$ if $i=1$, whereas it coincides with $\sqsubseteq$ if $i\not = 1$. Let $A:=\{(x_0,i_0), (x_1,i_1), (x_2,i_2), \dots, (x_{n-1},i_{n-1})\}$ be a finite antichain of $Q$ with $x_0\sqsubseteq x_1\sqsubseteq x_2\sqsubseteq x_3\sqsubseteq \dots\sqsubseteq x_n$. If $i_j\not=2$ for any $j\in n$ then $\{(x_0,2)\}\cup A$ is an antichain of $Q$. If $i_j\not=0$ for any $j\in n$ then $\{(x_{n-1},0)\}\cup A$ is an antichain of $Q$. It follows that every element of $AM_3(Q)$ is of the form $\{(x_0,0), (x_1,1), (x_2,2)\}$ with $x_0\sqsupseteq x_1\sqsupseteq x_2$. Let $A:=\{(x_0,0), (x_1,1), (x_2,2)\}\in AM_3(Q)$. Assume for a contradiction that $x_0\not= x_1$. Because $A$ is a maximal antichain it follows that $x_1<x_0$. According to the assumptions of the Lemma, there exists an element $y\in V$ which is not related to $x_1$ and $x_0$ and with $x_1\sqsubset y \sqsubset x_0$. Then $\{(x_0,0), (y,1), (x_1,1), (x_2,2)\}$ is an antichain. In a similar way we obtain that $x_2=x_1$. It follows that $AM_3(Q)=\{\{y\}\times 3 : y\in V\}$. We conclude that $AM_3(Q)$ is isomorphic to $L$ and $\bigcup AM_3(Q)$=Q. \[construct2\] There exists a poset $Q$ for which $AM_3(Q)$ is bqo but $\bigcup AM_3(Q)$ is not bqo. A poset $P$ which is wqo and not bqo but satisfies the conditions of Lemma \[construct2\] leads to a poset $Q$ which is wqo, not bqo, and for which $AM_3(Q)$ is bqo but $\bigcup {AM_3(Q)}$ is not bqo. One may take for $P$ Rado’s example and for a linear extension $L$ the lexicographic order according to the second difference. Index, notation, basic definitions and facts {#index} ============================================ Let $P$ denote a partially ordered set. $\underline{\text{ \large Poset, qoset, chain, well-founded, wqo, well-quasi-ordered:}}$ If $(P;\leq)$ is a partially ordered set, a [poset]{}, we will often just write $P$ for $(P;\leq)$. We write $a\leq b$ for $(a,b)\in\, \leq$. A [qoset]{} is a quasi ordered set and a linearly ordered poset is a [chain]{}. A qoset $P$ is [well-founded]{} if it contains no infinite descending chain $$\cdots < x_n < \cdots < x_0$$ and if in addition, $P$ contains no infinite antichain then it is [well-quasi-ordered]{} (wqo). If $P$ is a chain and well-quasi-ordered then it is [well-ordered]{}. $\underline{\text{ \large Initial segment, principal, ${\mathbf I}(P)$, $\bold{I}_{<\omega}(P)$, $\downarrow\hskip -2pt X$: }}$ A subset $I$ of $P$ is an [initial segment]{} (or is [closed downward]{}) if $x\leq y$ and $y\in I$ imply $x\in I$. We denote by ${\bf I}(P)$ the set of initial segments of $P$ ordered by inclusion. Let $X$ be a subset of $P$, then: $$\begin{aligned} \label{downarrow} \downarrow\hskip -2pt X:= \{y\in P: y\leq x \text{ for some } x\in X\}.\end{aligned}$$ We say that $\downarrow\hskip-2pt X$ is generated by $X$. If $X$ contains only one element $x$, we write $\downarrow\hskip -2pt x$ instead of $\downarrow\hskip -2pt \{x\}$. An initial segment generated by a singleton is [principal]{} and it is [finitely generated]{} if it is generated by a finite subset of $P$. We denote by $\bold{I}_{<\omega}(P)$ the set of finitely generated initial segments. $\underline{\text{ \large $up(P)$, $down(P)$}}$ We set $up(P):=\{\uparrow\hskip-2pt\hskip -2pt x: x\in P \}$ and $down(P):=\{\downarrow\hskip -2pt x: x\in P\}$. $\underline{\text{ \large $\leq_{dom}$, domination relation: }}$ A subset $X$ of $P$ is being [dominated]{} by the subset $Y$ of $P$, $X\leq_{dom}Y$, if for every $x\in X$ there is a $y\in Y$ such that $x\leq y$. The domination relation is a quasi-order on the power-set $\mathfrak{P}(P)$. The resulting ordered set is isomorphic to ${\bf I} (P)$, ordered by inclusion, via the map which associates with $X\in \mathfrak{P}(P)$ the initial segment $\downarrow\hskip-2pt X$. $\underline{\text{ \large $S_{\omega}(P)$, strictly increasing sequence: }}$ A sequence $(x_{n})_{n<\omega}$ of elements of $P$ is [strictly increasing]{} if $$x_{0}<x_{1}\cdots <x_{n}<x_{n+1}<\cdots$$ We denote by $S_{\omega}(P)$ the set of strictly increasing sequences of elements of $P$. For $(x_{n})_{n<\omega} \in S_{\omega}(P)$ and $(y_{n})_{n<\omega}\in S_{\omega}(P)$ we set $(x_{n})_{n<\omega} \leq (y_{n})_{n<\omega}$ if for every $n<\omega$ there is some $m<\omega$ such that $x_{n}\leq y_{m}$. This defines a quasi-order on $S_{\omega}(P)$. If we identify each $(x_{n})_{n<\omega}\in S_{\omega}(P)$ with the subset $\{x_{n}: n<\omega\}$ of $P$, this quasi-order is induced by the domination relation on subsets. $\underline{\text{ \large $\mathcal{J}(P)$, $\mathcal J^{\neg \downarrow\hskip -2pt }(P)$, ideal, non principal ideal:}}$ An [ideal]{} of $P$ is a non empty initial segment $I$ which is up-directed, that is every pair $x, y\in I$ has an upper bound $z\in I$. Its [cofinality]{}, $cf(I)$, is the least cardinal $\kappa$ such that there is some set $X$ of size $\kappa$ such that $I= \downarrow\hskip -2pt X$. The cofinality of $I$ is either $1$, in which case it has a largest element and is said to be [principal]{}, or is infinite. We denote by $\mathcal J^{\neg \downarrow\hskip -2pt }(P)$ the set of non principal ideals of $P$. Note the following fact, which goes back to Erdös-Tarski (1943)(see [@fraissetr]): \[fact:union\] A poset $P$ has no infinite antichain if and only if every initial segment of $P$ is a finite union of ideals. $\underline{\text{ \large $P^{dual} $, $\bold {F}(P)$, $\bold{F}_{<\omega}(P)$, $\mathcal {F}(P)$, filter:}}$ The [dual]{} of $P$ is the poset obtained from $P$ by reversing the order; we denote it by $P^{dual} $. A subset which is respectively an initial segment, a finitely generated initial segment or an ideal of $P^{dual} $ will be called a [final segment]{}, a [finitely generated final segment]{} or a [filter]{} of $P$. We denote by $\bold {F}(P)$, $\bold{F}_{<\omega}(P)$, and $\mathcal {F}(P)$ respectively, the collection of initial segments, finitely generated initial segments, and ideals of $P$ ordered by inclusion. $\underline{\text{ \large $\mathfrak{P}(P)$, $A(P)$, $AM(P)$, $AM_n(P)$:}}$ $\mathfrak{P}(P)$ denotes the set of subsets of $P$ and $A(P)$ is the collection of antichains of $P$, $AM(P)$ the collection of maximal antichains of $P$ and $AM_n(P)$ is the collection of $n$-element maximal antichains of $P$. The quasi-order of domination defined on $\mathfrak{P}(P)$ induces an ordering on the set $A(P)$ of antichains of $P$. The sets $A(P)$ and $A(P^{dual})$ are equal. Hence, we may order $A(P)$ by the domination order of $P^{dual}$. In general, these two orders are distinct. But, they coincide on $AM(P)$. $\underline{\text{ \large $\leq_{succ}$, $\leq_{pred}$, $\leq_{crit}$:}}$ Let $P$ be a poset. We write $a\leq_{pred(P)}b$ if $x<a$ implies $x<b$ for every $x\in P$. We write $a\leq_{succ(P)}b$ if $b<y$ implies $a<y$ for every $y\in P \}$. We write $a\leq_{prec}b$, or $a\leq_{succ}b$, if $P$ is understood. We denote by $(P;\leq_{prec})$ and $(P;\leq_{succ})$ the corresponding quasi-ordered sets. We write $a\leq_{crit(P)}b$ if $a\leq_{pred(P)}b$ and $a\leq_{succ(P)}b$. If in addition $a$ and $b$ are incomparable, the pair $(a,b)$ [critical]{}. $\underline{\text{ \large Interval order: }}$ The poset $P$ is an [interval-order]{} if $P$ is isomorphic to a subset $\mathcal J$ of the set $Int (C)$ of non-empty intervals of some chain $C$. The intervals are ordered as follows: for every $I, J\in Int(C)$, $I<J$ if $x<y$ for every $ x\in I$, every $y\in J$. Interval orders have neat characterizations in different ways: maximal antichains, associated preorders or obstructions, see [@fishburn], [@wiener]. We recall this important characterization: The following properties are equivalent: (i) $P$ is an interval order. (ii) $Pred({ P})$ is total qoset. (iii) $Succ( P)$ is a total qoset. (iv) $P$ does not contain a subset isomorphic to $\underline{2}\oplus \underline{2}$, the direct sum of two copies of the two-element chain. (v) $AM(P)$ is a chain. $\underline{\text{ \large ${\mathbb{N}}$, $[X]^{<\omega}$, $l(s)$, $\lambda (s)$, $_{}s$, $s_{}$, $s\cdot t$, $s\leq_{in} t$, $s\leq_{lex} t$:}}$ The set of non-negative integers is denoted by ${\mathbb{N}}$, the set of $n$-element subsets of $X\subseteq {\mathbb{N}}$ by $[X]^n$ and the set of finite subsets of $X\subseteq {\mathbb{N}}$ by $[X]^{<\omega}$. We identify each member $s$ of $[{\mathbb{N}}]^{<\omega}$ with a strictly increasing sequence, namely the list of its elements written in an increasing order, eg $\{3,4,8\}$. Let $s\in [{\mathbb{N}}]^{<\omega}$; the [length]{} of $s$, $l(s)$, is the number of its elements. For $m:= l(s)\not = 0$, we write $s:=\{s(0), \dots, s( m-1)\}$ with $s(0)<s(1)<\cdots<s(m-1)$. The smallest element of $s$ is $s(0)$ the largest, denoted by $\lambda (s)$, is $s(m-1)$. We denote by $_{}s$ the sequence obtained from $s$ by deleting its first element and by $s_{}$ the sequence obtained by deleting the last element. (With the convention that $_{}\emptyset=\emptyset_{}=\emptyset$.) We denote by $(a)$ the one element sequence with entry $a$. Let $s,t\in [{\mathbb{N}}]^{<\omega}$. If $\lambda(s) <t(0)$ then $s\cdot t$ is the concatenation of $s$ and $t$. We denote by $s\leq_{in} t$ the fact that $s$ is an initial segment of $t$ and by $s\leq_{lex} t$ the fact that $s$ is smaller than $t$ in the lexicographic order. (For example $\{3,5, 8,9\}\leq_{lex} \{3,5, 9, 15\}$.) If there exists an $r\in [{\mathbb{N}}]^{<\omega}$ with $s<_{in} r$ and $t=_{}r$ then $s\lhd t$. For example $\{ i\}\lhd\{j\}$ if and only if $i<j$, ($r=\{i,j\}$); also $\{i,j\}\lhd\{i',j'\}$ if and only if $j=i'$ and then $r=\{i, i^\prime, j^\prime\}$. $\underline{\text{ \large $\bigcup B$, $B_{\restriction X}$, block, thin block, barrier }}$ Let $B\subseteq [{\mathbb{N}}]^{<\omega}$ and $X\subseteq {\mathbb{N}}$. Then $\bigcup B:= \cup B$ and $B_{\restriction X}:= B\cap [X]^{<\omega}$. The set $B$ is a [block ]{}[^3] if: 1. $B$ is infinite. 2. For every infinite subset $X\subset \bigcup B$ there is some $s\in B\setminus \{\emptyset\}$ such that $s\leq_{in} X$. If $B$ is a block and an antichain for the order $\leq_{in}$ then $B$ is a [thin block]{}, whereas $B$ is a [ barrier ]{} if it is a block and an antichain for the inclusion order. A typical barrier is the set $[{\mathbb{N}}]^{n}$ of $n$-element subsets of ${\mathbb{N}}$. Trivially, every block contains a thin block, the set $\min_{\leq _{in}}(B)$ of $\leq_{in}$ minimal elements of the block. Moreover, if $B$ is a block, resp.a thin block, and $X$ is an infinite subset of $\bigcup B$ then $B_{\restriction X}$ is a block, resp a thin block. $\underline{\text{ \large $_{}B$, $B_{s}$, $_{s}B$, $B_{}$, $B^2$, $B^{\circ}$, $B\leq_{in} B^\prime$:}}$ Let $B, B^\prime\subseteq [{\mathbb{N}}]^{\omega}$. $_{}B:=\{_{}s: s\in B\}$. For $s\in [ {\mathbb{N}}]^{<\omega}$ $B_{s}:= \{t \in B: s\leq_{in} t\}$ $_{s}B:= \{t\in [{\mathbb{N}}]^{<\omega}: s\cdot t\in B\}=\{r\setminus s : r\in B_s\}$. Note that if $B$ is a thin block and $_{s}B$ is non-empty then it is a thin-block. $B_{}:= \{s_{}: s\in B\}$. $B^2:= \{ u:=s\cup t: s, t \in B$ and $ s\lhd t\}$. (This despite the possible confusion with the cartesian square of $B$.) $B^{\circ}$ is the set of all elements $s\in B$ with the property that for all $i\in \bigcup{B}$ with $i<s(0)$ there is an element $t\in B$ with $(i)\cdot {_\ast s}\leq_{in} t$. In other words $s\in B^{\circ}$ if $(i)\cdot {_\ast s}\in \mathrm{T}(B)$ for all $i\in \bigcup{B}$ with $i<s(0)$. Let $B^\prime:=\{{_\ast s} : s\in B^\circ\}\setminus \{\emptyset\}$. $B'\leq_{in} B$ if for every $s'\in B'$ there is some $s\in B$ such that $ s'\leq_{in} s$. This is the quasi-order of domination associated with the order $\leq_{in}$ on $[{\mathbb{N}}]^{<\omega}$ $\underline{\text{ \large bqo, good, bad }}$ A map $f$ from a barrier $B$ into a poset $P$ is [good]{} if there are $s,t\in B$ with $s\lhd t$ and $f(s) \leq f(t)$. Otherwise $f$ is [bad]{}.\ Let $\alpha$ be a denumerable ordinal. A poset $P$ is [$\alpha$-better-quasi-ordered]{} if every map $f: B\rightarrow P$, where $B$ is a barrier of order type at most $\alpha$, is good.\ A poset $P$ is [better-quasi-ordered]{} if it is $\alpha$-better-quasi-ordered for every denumerable ordinal $\alpha$. [999]{} M. Assous, Caractérisation du type d’ordre des barrières de Nash-Williams. Publ. Dép. Math. (Lyon) 11 (1974), no. 4, 89–106. G.Behrendt, Maximal antichains in partially ordered sets. Eleventh British Combinatorial Conference (London, 1987). Ars Combin. 25 (1988), C, 149–157. M.Bekkali, M.Pouzet, D.Zhani, Incidence structures and Stone-Priestley duality, 2004. R.Bonnet, Private communication, december 2004. R. Fra[ï]{}ss[é]{}, . North-Holland Publishing Co., Amsterdam, 2000. P.C. Fischburn, Interval Orders and Interval graphs, Wiley, 1985. F. Galvin, A generalization of RamseyÕs theorem. Notices Amer. Math. Soc. vol.15 p.548 (1969) Abstract 68 T 368. G. Higman, Ordering by divisibility in abstract algebras, Proc. London. Math. Soc. 2 (3), (1952), 326-336. J. B. Kruskal, Well-quasi-ordering, the [T]{}ree [T]{}heorem, and [V]{}azsonyi’s conjecture. , 95 (1960)210–225. J. B. Kruskal, The theory of well-quasi-ordering : a frequently discovered concept. , 13:297–305, 1972. R. Laver, Well-quasi-orderings and sets of finite sequences. Math. Proc. Cambridge Philos. Soc. 79 (1976), no. 1, 1–10. A.Marcone, Foundations of bqo theory Trans. Amer. Math. Soc., 345 (1994), 641-660 A.Marcone, Fine analysis of the quasi-orderings on the power set, Order, 18, (2001)339-347. E. C. Milner, Basic wqo- and bqo-theory, In [Graphs and order (Banff, Alta., 1984)]{}, pages 487–502. Reidel, Dordrecht, 1985. C. St. J. A. Nash-Williams, On well-quasi-ordering finite trees. , 59:833–835, 1963. C. St. J. A. Nash-Williams, On better-quasi-ordering transfinite sequences. , 64:273–290, 1968. M.Pouzet, Sur les prémeilleurordres, Ann. Inst. Fourier(Grenoble) 22(1972)1-20. M. Pouzet, Sur la théorie des relations. Thèse d’état, Université Claude-Bernard, Lyon 1, 1978. M.Pouzet, Graphs and posets with no infinite independent set. Finite and infinite combinatorics in sets and logic (Banff, AB, 1991), 313–335, NATO Adv. Sci. Inst. Ser. C Math. Phys. Sci., 411, Kluwer Acad. Publ., Dordrecht, 1993. F. P. Ramsey, On a problem of formal logic, Proc. London Math. Soc., 30, (1930), 264-286. R.Rado, Partial well-ordering of a set of vectors, Mathematika, 1 (1954), 89-95. N.Robertson, P. Seymour, Graph minors—a survey. Surveys in combinatorics 1985 (Glasgow, 1985), 153–171, London Math. Soc. Lecture Note Ser., 103, Cambridge Univ. Press, Cambridge, 1985. M. Wiener, A contribution to the theory of relative position, Proc. Camb. Philos. Soc., 17, 1914, p. 441-449. [^1]: Supported by Intas [^2]: Supported by NSERC of Canada Grant \# 691325 [^3]: We stick to the definition of Nash-Williams, 1968 [@nwbqots]; in some papers, a block is what we call a thin block	{ "pile_set_name": "ArXiv" }
--- abstract: 'We study the role of coded side information in single-server Private Information Retrieval (PIR). An instance of the single-server PIR problem includes a server that stores a database of $K$ independently and uniformly distributed messages, and a user who wants to retrieve one of these messages from the server. We consider settings in which the user initially has access to a coded side information which includes a linear combination of a subset of $M$ messages in the database. We assume that the identities of the $M$ messages that form the support set of the coded side information as well as the coding coefficients are initially unknown to the server. We consider two different models, depending on whether the support set of the coded side information includes the requested message or not. We also consider the following two privacy requirements: (i) the identities of both the demand and the support set of the coded side information need to be protected, or (ii) only the identity of the demand needs to be protected. For each model and for each of the privacy requirements, we consider the problem of designing a protocol for generating the user’s query and the server’s answer that enables the user to decode the message they need while satisfying the privacy requirement. We characterize the (scalar-linear) capacity of each setting, defined as the ratio of the number of information bits in a message to the minimum number of information bits downloaded from the server over all (scalar-linear) protocols that satisfy the privacy condition. Our converse proofs rely on new information-theoretic arguments—tailored to the setting of single-server PIR and different from the commonly-used techniques in multi-server PIR settings. We also present novel capacity-achieving scalar-linear protocols for each of the settings being considered.' author: - 'Anoosheh Heidarzadeh, Fatemeh Kazemi, and Alex Sprintson[^1][^2][^3]' bibliography: - 'PIR\_salim.bib' - 'pir\_bib.bib' - 'coding1.bib' - 'coding2.bib' title: 'The Role of Coded Side Information in Single-Server Private Information Retrieval' --- Private information retrieval, information-theoretic privacy, single server, coded side information Introduction ============ In the Private Information Retrieval (PIR) problem, there is a user that wishes to privately download a single or multiple messages belonging to a database stored on a single or multiple (non-colluding or colluding) servers. There are two different types of PIR in the literature: computational and information-theoretic. In the computational PIR (see, e.g., [@KO1997]), the identity of the requested message(s) must be protected from the server(s), assuming that the server(s) is computationally bounded. Aside from the computational PIR is the information-theoretic PIR, introduced by Chor et al. in [@Chor:PIR1995], where no such assumption is made on the computational power of the server(s), and the identity of the requested message(s) need to be protected in an information-theoretic sense. The drawback of this strong requirement is that in the single-server case, the user must download the entire database from the server [@Chor:PIR1995]. This has led to an extensive body of work on multi-server information-theoretic PIR (see, e.g., [@beimel2002breaking; @gasarch2004survey; @yekhanin2010private; @Sun2017; @JafarPIR3new; @BU2018; @shah2014one; @CHY2015; @tajeddine2016private; @extended; @BU18; @fazeli2015pir; @blackburn2016pir; @freij2016private]). Initiated by the work of Kadhe et al. in [@Kadhe2017] and [@KGHERS2017], the information-theoretic PIR was recently extended to the settings wherein the user has a random subset of messages in the database as side information, and the identities of side information messages are unknown to the server(s) [@Kadhe2017; @KGHERS2017; @HKGRS:2018; @KHSO2019; @LG:2018; @KKHS32019; @Chen2017side; @Maddah2018; @HKRS2019]. (Some other types of side information, not closely related to our work, were also studied, see, e.g.,[@Tandon2017; @Wei2017CacheAidedPI; @Wei2017FundamentalLO; @WU2018].) Three different notions of privacy were considered: (i) $({W},{S})$-privacy, where both the identities of the requested messages (denoted by the index set $W$) and the identities of the side information messages (denoted by the index set $S$) must be protected [@Kadhe2017; @KGHERS2017; @Chen2017side; @Maddah2018; @HKGRS:2018; @KHSO2019]; (ii) joint ${W}$-privacy, where only the identities of the requested messages (and not necessarily the identities of the side information messages) must be protected [@Kadhe2017; @KGHERS2017; @HKGRS:2018; @KHSO2019; @KKHS32019; @LG:2018]; and (iii) individual ${W}$-privacy, where the identity of each requested message must be protected individually (but not necessarily jointly) [@HKRS2019]. In single-message PIR, where the user wants to retrieve one message only, the notions of joint and individual ${W}$-privacy, referred to as ${W}$-privacy for brevity, are equivalent. The differences between these two notions of ${W}$-privacy in multi-message PIR were studied in [@HKRS2019]. [\|c\|c\|c\|c\|c\|]{} Privacy Condition & &\ Model & ------------------ ${W}\not\in {S}$ (PIR-PCSI–I) ------------------ & --------------- ${W}\in {S}$ (PIR-PCSI–II) --------------- & ------------------ ${W}\not\in {S}$ (PIR-CSI–I) ------------------ & -------------- ${W}\in {S}$ (PIR-CSI–II) -------------- \ Parameters & $1\leq M\leq K-1$ & $2\leq M\leq K$ & $1\leq M\leq K-1$ & $2\leq M\leq K$\ Capacity & & --------------------------------------------- $(K-M+1)^{-1}$ for $M>\frac{K+1}{2}$ (Theorem \[thm:PIRPCSI-II\]) Open for $M\leq \frac{K+1}{2}$ (Lower bound in Theorem \[thm:PIRPCSI-II\]) --------------------------------------------- & &\ --------------- Scalar-Linear Capacity --------------- & & ------------------------------ $(K-M+1)^{-1}$ (Theorem \[thm:PIRPCSI-II\]) ------------------------------ & &\ --------------- Achievability Scheme --------------- & ------------- Specialized GRS Code ------------- & ---------------------- Modified Specialized GRS Code ---------------------- & -------------------- Modified Partition-and-Code -------------------- & -------------------- Randomized Selection-and-Code -------------------- \ In this work, we focus on single-message single-server information-theoretic PIR in the presence of a coded side information. We initiated this study in [@HKS2018] and [@HKS2019] for the cases in which ${W}$-privacy and $({W},{S})$-privacy are required, respectively, where $W$ denotes the index of the requested message, and $S$ denotes the index set of the messages in the support set of the coded side information. We have recently extended these works to the multi-server setting in [@KKHS12019] and [@KKHS22019]. In this problem, there is a single server storing a database of $K$ independently and uniformly distributed messages; and there is a user who is interested in retrieving a single message from the server. The user initially knows a linear coded combination of a subset of $M$ messages in the database, where the identities of the messages in the support set of the user’s coded side information as well as their coding coefficients are initially unknown to the server. This setting can be motivated by several practical scenarios. For instance, the user may have obtained a coded side information via overhearing in a wireless network; or on-the-fly recording of a random linear combination of messages being broadcast by an information source; or from a trusted agent, e.g., an entity who makes profit by offering privacy to users, with limited knowledge about the database; or from the information which is locally stored, e.g., using an erasure code, in the user’s cache of limited size. Recently, inspired by [@Kadhe2017], a group of researchers from Google in [@PPY2018] used the idea of a coded side information in a new single-server PIR scheme, which leverages both the information-theoretic and computational PIR. The problem is to design a protocol for generating the user’s query and the server’s answer which satisfy one of the following two privacy conditions: $({W},{S})$-privacy, i.e., the privacy of both the requested message and the messages in the support set of the coded side information must be preserved, or ${W}$-privacy, i.e., only the privacy of the requested message needs to be protected. We refer to this problem as PIR with Private Coded Side Information (PIR-PCSI) or PIR with Coded Side Information (PIR-CSI) when $({W},{S})$-privacy or ${W}$-privacy is required, respectively. Depending on whether the support set of the user’s coded side information includes the user’s demand or not, we consider two different models for each of the PIR-PCSI and PIR-CSI problems. In the first model, referred to as Model I, the demand does not belong to the support set of the coded side information, whereas in the second model, referred to as Model II, the demand belongs to the support set of the coded side information. We refer to the PIR-PCSI (or PIR-CSI) problem under Model I and Model II as PIR-PCSI–I (or PIR-CSI–I) and PIR-PCSI–II (or PIR-CSI–II), respectively. For each of these settings, we define the capacity as the ratio of the number of information bits in a message to the minimum number of information bits downloaded from the server over all protocols that satisfy the privacy condition. We similarly define the scalar-linear capacity of each setting, except when the minimum is taken over all scalar-linear protocols—the protocols in which the server’s answer contains only scalar-linear combinations of the messages in the database (i.e., linear combinations with scalar coding coefficients), that satisfy the privacy condition. In this work, our goal is to characterize the capacity and the scalar-linear capacity of each of the PIR-PCSI and PIR-CSI settings, and design a capacity-achieving protocol for each of these settings. The main contributions of this work are as follows. The results are also summarized in Table \[tab:1\]. For the PIR-PCSI–I setting, we prove that the capacity and the scalar-linear capacity are both given by $(K-M)^{-1}$ for any ${1\leq M\leq K-1}$. This is interesting because, as shown in [@Kadhe2017 Theorem 2], the capacity of PIR with $M$ randomly chosen messages as side information is equal to $(K-M)^{-1}$ when $({W},{S})$-privacy is required. This shows that for achieving $({W},{S})$-privacy, even one random linear coded combination of a random subset of $M$ messages is as efficient as $M$ randomly chosen messages separately, as side information. For the PIR-PCSI–II setting, we prove that the scalar-linear capacity for any value of ${2\leq M\leq K}$ and the capacity for any value of ${\frac{K+1}{2}<M\leq K}$ are given by $(K-M+1)^{-1}$, whereas the capacity for any value of $2\leq M\leq \frac{K+1}{2}$ remains open. This shows that when the user knows only one random linear coded combination whose support set consists of the requested message along with $M-1$ other randomly chosen messages, achieving $({W},{S})$-privacy is no more costly than that when the user knows $M-1$ randomly chosen (uncoded) messages, different from the requested message. For the PIR-CSI–I setting, we prove that the capacity and the scalar-linear capacity are given by $\lceil\frac{K}{M+1}\rceil^{-1}$ for any $0\leq M< K$. Interestingly, this is the same as the capacity of PIR with $M$ randomly chosen messages as side information [@Kadhe2017 Theorem 1]. For the PIR-CSI–II setting, we prove that the capacity and the scalar-linear capacity are equal to $1$ for $M=2$ and $M=K$, and are equal to $\frac{1}{2}$ for any ${3\leq M\leq K-1}$. This result is particularly interesting because, unlike the previous settings, the gap between the capacity and the trivial capacity upper bound $1$ is a constant, regardless of the size of support set of the side information ($M$). The converse proofs are based on new information-theoretic arguments. These arguments are tailored to the setting of single-server PIR and are different from the proof techniques being commonly used in the multi-server PIR settings. In particular, the main ingredients in the proofs are a necessary condition for $({W},{S})$-privacy and a necessary condition for ${W}$-privacy, which reveal the combinatorial nature of the problem of single-server PIR in the presence of (uncoded or coded) side information. In addition, our converse proofs for the PIR-PCSI–I and PIR-CSI–I settings serve as alternative information-theoretic proofs for the results in [@Kadhe2017] which were proven using index coding arguments. The achievability proofs are based on novel scalar-linear PIR-PCSI and PIR-CSI protocols. In particular, the proposed PIR-PCSI–I and PIR-PCSI–II protocols, termed the Specialized GRS Code protocol and the Modified Specialized GRS Code protocol, rely on the Generalized Reed-Solomon (GRS) codes that contain a specific codeword, depending on the index of the requested message as well as the indices of the messages in the support set of the coded side information and their coding coefficients. The proposed protocol for the PIR-CSI–I setting, termed the Modified Partition-and-Code (MPC) protocol, is inspired by our recently proposed Partition-and-Code with Interference Alignment protocol in [@HS2019] for single-server private computation with uncoded side information. The MPC protocol also generalizes the Partition-and-Code protocol of [@Kadhe2017] for single-server PIR with uncoded side information. It is noteworthy that we originally introduced a different PIR-CSI–I protocol in [@HKS2018], termed Randomized Partitioning (RP) protocol, which is also capacity-achieving. For the PIR-CSI–II setting, we propose a protocol, termed the Randomized Selection-and-Code protocol, which is based on the idea of randomizing the structure of the user’s query and the server’s answer (instead of always using a fixed structure for query/answer). We introduced this idea in [@HKS2018] for the first time, and Tian et al., concurrently and independently, used a similar idea in [@TSC2018] for multi-server PIR without side information. Problem Setup and Formulation {#sec:SN} ============================= Basic Notation -------------- Throughout this paper, we denote random variables and their realizations by bold-face and regular letters, respectively. The functions $\mathbb{P}(\cdot)$, $\mathbb{P}(\cdot\|\cdot)$, $H(\cdot)$, $H(\cdot\|\cdot)$, and $I(\cdot;\cdot \|\cdot)$ denote probability, conditional probability, (Shannon) entropy, conditional entropy, and conditional mutual information, respectively. Let $\mathbb{F}_q$ be a finite field for a prime power $q$, and let $\mathbb{F}^{\times}_q\triangleq \mathbb{F}_q\setminus \{0\}$ be the multiplicative group of $\mathbb{F}_q$. Let $\mathbb{F}_{q^l}$ be an extension field of $\mathbb{F}_q$ for an integer $l\geq 1$, and let $L\triangleq l\log_2 q$. The parameters $q$ and $l$ are referred to as the base-field size and the field-extension degree, respectively. Let $K\geq 1$ and $1\leq M\leq K$ be two integers. Let $\mathcal{K} \triangleq \{1,\dots,K\}$. We denote by $\mathcal{S}$ the set of all $M$-subsets (i.e., all subsets of size $M$) of $\mathcal{K}$, and denote by $\mathcal{C}$ the set of all ordered multisets of size $M$ (i.e., all length-$M$ sequences) with elements from $\mathbb{F}_q^{\times}$. Note that $\|\mathcal{S}\| = \binom{K}{M}$ and $\|\mathcal{C}\| = (q-1)^{M}$. Setup and Assumptions --------------------- There is a server that stores a set of $K$ messages ${X}_1,\dots,{X}_K$, denoted by $X_{\mathcal{K}}\triangleq \{X_1,\dots,X_K\}$, where ${\mathbf{X}}_i$’s are independently and uniformly distributed over $\mathbb{F}_{q^l}$, i.e., ${H({\mathbf{X}}_i) = L}$ for $i\in \mathcal{K}$ and $H({\mathbf{X}}_{\mathcal{K}}) = KL$, where ${\mathbf{X}}_{\mathcal{K}}\triangleq \{{\mathbf{X}}_1,\dots,{\mathbf{X}}_K\}$. There is a user who wants to retrieve a message ${X}_{{W}}$ for some ${W}\in \mathcal{K}$ from the server, and knows a linear combination ${Y}^{[{S},{C}]}\triangleq \sum_{i\in {S}} {c}_i {X}_i$ on the messages $X_{{S}}\triangleq \{X_i: i\in S\}$, for some ${S} \triangleq \{i_1,\dots,i_M\}\in \mathcal{S}$ and ${{C} \triangleq \{{c}_{i_1},\dots,{c}_{i_M}\} \in \mathcal{C}}$. We refer to ${X}_{{W}}$ as the demand, ${W}$ as the demand index, $X_{{S}}$ as the side information support set, ${S}$ as the side information support index set, $M$ as the side information support size, and ${Y}^{[{S},{C}]}$ as the (coded) side information. We assume that ${\mathbf{S}}$ and ${\mathbf{C}}$ are uniformly distributed over $\mathcal{S}$ and $\mathcal{C}$, respectively. Also, we consider two different models for the conditional distribution of ${\mathbf{W}}$ given ${\mathbf{S}}={S}$: ### Model I {#modeli .unnumbered} ${\mathbf{W}}$ is uniformly distributed over $\mathcal{K}\setminus{S}$, $$\mathbb{P}({\mathbf{W}}={W}\|{\mathbf{S}}={S}) = \begin{cases} \frac{1}{K-M}, & {W}\in \mathcal{K}\setminus{S},\\ 0, & \text{otherwise}; \end{cases}$$ ### Model II {#modelii .unnumbered} ${\mathbf{W}}$ is uniformly distributed over ${S}$, $$\mathbb{P}({\mathbf{W}}={W}\|{\mathbf{S}}={S}) = \begin{cases} \frac{1}{M}, & {W}\in {S},\\ 0, & \text{otherwise}. \end{cases}$$ For both Models I and II, ${\mathbf{W}}$ is distributed uniformly over $\mathcal{K}$. Let $\mathds{1}_{\{\mathbf{W}\in \mathbf{S}\}}$ be an indicator random variable such that that $\mathds{1}_{\{\mathbf{W}\in \mathbf{S}\}}=0$ if $\mathbf{W}\not\in \mathbf{S}$, and $\mathds{1}_{\{\mathbf{W}\in \mathbf{S}\}}=1$ otherwise. Note that $\mathds{1}_{\{\mathbf{W}\in \mathbf{S}\}} = 0$ in Model I, and $\mathds{1}_{\{\mathbf{W}\in \mathbf{S}\}} = 1$ in Model II. We assume that the server knows the underlying model (i.e., whether ${\mathbf{W}}\not\in {\mathbf{S}}$ or ${\mathbf{W}}\in {\mathbf{S}}$), the side information support size $M$, the distributions of ${\mathbf{S}}$ and ${\mathbf{C}}$, and the conditional distribution of ${\mathbf{W}}$ given ${\mathbf{S}}$, in advance; whereas the realizations ${W},{S},{C}$ are unknown to the server in advance. Privacy and Recoverability Conditions ------------------------------------- For any given ${W}, {S}, {C}$, in order to retrieve ${X}_{{W}}$, the user sends to the server a query ${Q}^{[{W},{S},{C}]}$, which is a (potentially stochastic) function of ${W},{S},{C}$.[^4] For simplifying the notation, we denote ${\mathbf{Q}}^{[{\mathbf{W}},{\mathbf{S}},{\mathbf{C}}]}$ by ${\mathbf{Q}}$. The query must satisfy one of the following two privacy conditions: - both the user’s demand index and side information support index set must be protected from the server; - only the user’s demand index (and not necessarily the side information support index set) must be protected from the server. The condition (i) is referred to as the $({W},{S})$-privacy condition, and the condition (ii) is referred to as the ${W}$-privacy condition. (Note that $({W},{S})$-privacy is a stronger condition than ${W}$-privacy.) The $({W},{S})$-privacy condition implies that $({\mathbf{W}},{\mathbf{S}})$ and $\mathbf{Q}$ must be conditionally independent given $\mathds{1}_{\{{\mathbf{W}}\in{\mathbf{S}}\}}$, $$I({\mathbf{W}},{\mathbf{S}};{\mathbf{Q}}\|\mathds{1}_{\{{\mathbf{W}}\in{\mathbf{S}}\}}) = 0.$$ The ${W}$-privacy condition implies that ${\mathbf{W}}$ and $\mathbf{Q}$ must be conditionally independent given $\mathds{1}_{\{{\mathbf{W}}\in{\mathbf{S}}\}}$, $$I({\mathbf{W}};{\mathbf{Q}}\|\mathds{1}_{\{{\mathbf{W}}\in{\mathbf{S}}\}}) = 0.$$ Equivalently, for a given $\theta\in \{0,1\}$, when $({W},{S})$-privacy is required, it must hold that $$\begin{aligned} & \mathbb{P}({\mathbf{W}}= {W}^{},{\mathbf{S}}={S}^{}\| {\mathbf{Q}}= {Q}^{[{W},{S},{C}]},\mathds{1}_{\{{\mathbf{W}}\in {\mathbf{S}}\}}=\theta) \\ & = \mathbb{P}({\mathbf{W}}= {W}^{},{\mathbf{S}}={S}^{}\|\mathds{1}_{\{{\mathbf{W}}\in {\mathbf{S}}\}}=\theta)\end{aligned}$$ for all ${W}^{}\in \mathcal{K}$ and ${S}^{}\in \mathcal{S}$, and when ${W}$-privacy is required, it must hold that $$\begin{aligned} &\mathbb{P}({{\mathbf{W}}={W}^{}}\|{{\mathbf{Q}} = {Q}^{[{W},{S},{C}]}},{\mathds{1}_{\{{\mathbf{W}}\in {\mathbf{S}}\}}=\theta}) \\ & = {\mathbb{P}({\mathbf{W}}={W}^{}\|{\mathds{1}_{\{{\mathbf{W}}\in {\mathbf{S}}\}}=\theta})}\end{aligned}$$ for all ${W}^{}\in \mathcal{K}$.[^5] Upon receiving ${Q}^{[{W},{S},{C}]}$, the server sends to the user an answer ${A}^{[{W},{S},{C}]}$, which is a (deterministic) function of the query ${Q}^{[{W},{S},{C}]}$, the indicator variable $\mathds{1}_{\{{W}\in {S}\}}$, and the messages in $X_{\mathcal{K}}$. For simplifying the notation, we denote ${\mathbf{A}}^{[{\mathbf{W}},{\mathbf{S}},{\mathbf{C}}]}$ by ${\mathbf{A}}$. Note that $({\mathbf{W}},{\mathbf{S}},{\mathbf{C}}) \leftrightarrow ({\mathbf{Q}},\mathds{1}_{\{{\mathbf{W}}\in {\mathbf{S}}\}},{\mathbf{X}}_{\mathcal{K}}) \leftrightarrow {\mathbf{A}}$ forms a Markov chain, and $H({\mathbf{A}}\| {\mathbf{Q}},\mathds{1}_{\{{\mathbf{W}}\in {\mathbf{S}}\}}, {\mathbf{X}}_{\mathcal{K}},{\mathbf{W}},{\mathbf{S}},{\mathbf{C}}) = 0$. The answer ${A}^{[{W},{S},{C}]}$ along with ${Q}^{[{W},{S},{C}]}, \mathds{1}_{\{{W}\in {S}\}}, Y^{[{S},{C}]}$, and ${W}, {S}, {C}$ must enable the user to retrieve the demand ${X}_{{W}}$. That is, it must hold that $$H({\mathbf{X}}_{{\mathbf{W}}}\| {\mathbf{A}}, {\mathbf{Q}}, \mathds{1}_{\{{\mathbf{W}}\in {\mathbf{S}}\}},{\mathbf{Y}}^{[{\mathbf{S}},{\mathbf{C}}]}, {\mathbf{W}},{\mathbf{S}},{\mathbf{C}})=0.$$ We refer to this condition as the recoverability condition. PIR-PCSI and PIR-CSI Problems ----------------------------- For each type of privacy and for each model, the problem is to design a protocol for generating a query ${Q}^{[{W},{S},{C}]}$ (and the corresponding answer ${A}^{[{W},{S},{C}]}$, given ${Q}^{[{W},{S},{C}]}$, $\mathds{1}_{\{{W}\in {S}\}}$, and ${X}_{\mathcal{K}}$) for any given ${W},{S},{C}$, such that both the privacy and recoverability conditions are satisfied. Note that the protocol is assumed to be known at the server. When $({W},{S})$-privacy is required, we refer to this problem as Private Information Retrieval (PIR) with Private Coded Side Information* (PIR-PCSI), and when ${W}$-privacy is required we refer to this problem as PIR with Coded Side Information (PIR-CSI). The PIR-PCSI problem under Model I (or Model II) is referred to as the PIR-PCSI–I (or PIR-PCSI–II) setting; and the PIR-CSI problem under Model I (or Model II) is referred to as the PIR-CSI–I (or PIR-CSI–II) setting. A protocol for generating query/answer for the PIR-PCSI–I (or PIR-PCSI–II) setting is referred to as a PIR-PCSI–I (or PIR-PCSI–II) protocol. A PIR-CSI–I (or PIR-CSI–II) protocol is defined similarly. Capacity and Scalar-Linear Capacity ----------------------------------- The rate of a PIR-PCSI–I (or PIR-PCSI–II) protocol is defined as the ratio of the entropy of a message, i.e., $L$, to the conditional entropy of ${\mathbf{A}}^{[{\mathbf{W}},{\mathbf{S}},{\mathbf{C}}]}$ given that $\mathds{1}_{\{{\mathbf{W}}\in {\mathbf{S}}\}}=0$ (or $\mathds{1}_{\{{\mathbf{W}}\in {\mathbf{S}}\}}=1$). The rate of a PIR-CSI–I (or PIR-CSI–II) protocol is defined similarly. The capacity of PIR-PCSI–I (or PIR-PCSI–II) setting is defined as the supremum of rates over all PIR-PCSI–I (or PIR-PCSI–II) protocols and over all base-field sizes $q$ and all field-extension degrees $l$; and the capacity of PIR-CSI–I (or PIR-CSI–II) setting is defined similarly. The scalar-linear capacity of PIR-PCSI–I (or PIR-PCSI–II) setting is defined as the supremum of rates over all scalar-linear PIR-PCSI–I (or PIR-PCSI–II) protocols (i.e., the protocols in which the answer of the server consists only of the scalar-linear combinations of the messages in $X_{\mathcal{K}}$) and over all $q$ and $l$. The scalar-linear capacity of PIR-CSI–I (or PIR-CSI–II) setting is defined similarly.[^6] Problem Statement ----------------- In this work, our goal is to derive upper bounds on the capacity and the scalar-linear capacity of the PIR-PCSI–I, PIR-PCSI–II, PIR-CSI–I, and PIR-CSI–II settings, and to design protocols that achieve the corresponding upper-bounds. Main Results ============ We present our main results in this section. The results for the PIR-PCSI–I and PIR-PCSI–II settings are summarized in Section \[subsec:PCSIMainResults\], and the results for the PIR-CSI–I and PIR-CSI–II settings are summarized in Section \[subsec:CSIMainResults\] The following two lemmas give a necessary condition for $({W},{S})$-privacy and ${W}$-privacy, respectively. These simple but powerful lemmas are the key components in the converse proofs of our main results. \[prop:1\] For $({W},{S})$-privacy, for a given ${\theta\in \{0,1\}}$, for any ${W}^{}\in\mathcal{K}$ and ${S}^{}\in\mathcal{S}$ with ${\mathds{1}_{\{{W}^{}\in {S}^{}\}}=\theta}$, there must exist ${{C}^{}\in\mathcal{C}}$ such that $$H({\mathbf{X}}_{{W}^{}}\| {\mathbf{A}}, {\mathbf{Q}}, \mathds{1}_{\{{\mathbf{W}}\in {\mathbf{S}}\}}=\theta, {\mathbf{Y}}^{[{S}^{},{C}^{}]}) = 0.$$ [[[* Proof.* ]{}]{}]{}The proof is by the way of contradiction. For a given $\theta\in \{0,1\}$, consider an arbitrary ${W}^{}\in \mathcal{K}$ and an arbitrary ${S}^{}\in \mathcal{S}$ such that $\mathds{1}_{\{{W}^{}\in {S}^{}\}}=\theta$. Suppose that there does not exist any ${C}^{}\in \mathcal{C}$ such that ${H({\mathbf{X}}_{{W}^{}}\| {\mathbf{A}}, {\mathbf{Q}}, \mathds{1}_{\{{\mathbf{W}}\in {\mathbf{S}}\}}=\theta, {\mathbf{Y}}^{[{S}^{},{C}^{}]}) = 0}$. If ${W}^{}$ and ${S}^{}$ are respectively the user’s demand index and side information support index set, no matter what the user’s side information $Y^{[{S}^{},\cdot]}$ is, the user cannot recover $X_{{W}^{}}$ given the answer, query, and the side information $Y^{[{S}^{},\cdot]}$. This violates the recoverability condition. Thus, ${W}^{}$ and ${S}^{}$ cannot be the user’s demand index and side information support index set, respectively. This obviously violates the $({W},{S})$-privacy condition, because given the query, every ${W}^{}\in \mathcal{K}$ and every ${S}^{}\in \mathcal{S}$ such that $\mathds{1}_{\{{W}^{}\in {S}^{}\}}=\theta$ must be equally likely to be the user’s demand index and side information support index set, respectively. \[prop:2\] For ${W}$-privacy, for a given ${\theta\in \{0,1\}}$, for any ${{W}^{}\in\mathcal{K}}$, there must exist ${S}^{}\in\mathcal{S}$ with ${\mathds{1}_{\{{W}^{}\in {S}^{}\}}=\theta}$ and ${{C}^{}\in\mathcal{C}}$ such that $$H({\mathbf{X}}_{{W}^{}}\| {\mathbf{A}}, {\mathbf{Q}}, \mathds{1}_{\{{\mathbf{W}}\in {\mathbf{S}}\}}=\theta, {\mathbf{Y}}^{[{S}^{},{C}^{}]}) = 0.$$ [[[ Proof.* ]{}]{}]{}The proof is similar to the proof of Lemma \[prop:1\]—except that the ${W}$-privacy condition is used instead of the $({W},{S})$-privacy condition. The proof is omitted for brevity. PIR-PCSI {#subsec:PCSIMainResults} -------- In this section, we present our main results for the PIR-PCSI–I and PIR-PCSI–II settings. The capacity and the scalar-linear capacity of the PIR-PCSI–I setting (for all ${1\leq M\leq K-1}$) are characterized in Theorem \[thm:PIRPCSI-I\], and the capacity (for all $\frac{K+1}{2}<M\leq K$) and the scalar-linear capacity (for all $2\leq M\leq K$) of the PIR-PCSI–II setting are characterized in Theorem \[thm:PIRPCSI-II\]. For any $2\leq M\leq \frac{K+1}{2}$, the capacity of the PIR-PCSI–II setting, which we conjecture to be the same as the scalar-linear capacity, remains open. The proofs are given in Sections \[sec:PIRPCSI-I\] and \[sec:PIRPCSI-II\]. \[thm:PIRPCSI-I\] For the PIR-PCSI–I setting with $K$ messages and side information support size $M$, the capacity and the scalar-linear capacity are given by $(K-M)^{-1}$ for all $1\leq M\leq K-1$. The converse follows directly from the result of [@Kadhe2017 Theorem 2], which was proven using an index coding argument, for single-server single-message PIR with (uncoded) side information when $({W},{S})$-privacy is required. In this work, we provide an alternative proof of converse by upper bounding the rate of any PIR-PCSI–I protocol using the information-theoretic arguments (see Section \[subsec:ConvThm1\]). The key component of the proof is the necessary condition for $({W},{S})$-privacy, stated in Lemma \[prop:1\]. The achievability proof relies on a new scalar-linear PIR-PCSI–I protocol, termed the Specialized GRS Code protocol, which achieves the rate $(K-M)^{-1}$ (see Section \[subsec:AchThm1\]). This protocol is based on the Generalized Reed-Solomon (GRS) codes that contain a specific codeword depending on ${W}, {S},{C}$. \[rem:1\] As shown in [@Kadhe2017], when there is a single server storing $K$ independent and identically distributed messages, and there is a user that knows $M$ randomly chosen (uncoded) messages as their side information and demands a single message not in their side information, in order to guarantee $({W},{S})$-privacy, the minimum download cost is $(K-M)L$, where $L$ is the entropy of a message. Surprisingly, this result matches the result of Theorem \[thm:PIRPCSI-I\]. This shows that, when compared to having $M$ random messages separately as side information, for achieving $({W},{S})$-privacy there will be no additional loss in capacity even if only one* random linear coded combination of $M$ random messages is known by the user.* \[thm:PIRPCSI-II\] For the PIR-PCSI–II setting with $K$ messages and side information support size $M$, the capacity is given by ${(K-M+1)^{-1}}$ for all $\frac{K+1}{2}<M\leq K$, and it is lower bounded by ${(K-M+1)^{-1}}$ for all $2\leq M\leq \frac{K+1}{2}$. Moreover, the scalar-linear capacity is given by $(K-M+1)^{-1}$ for all ${2\leq M\leq K}$. The converse proof for the scalar-linear case is based on a mix of algebraic and information-theoretic arguments (see Section \[subsec:ConvThm2\]), and the proof of converse for the general case relies on different information-theoretic arguments. The main ingredient of the proofs is the result of Lemma \[prop:1\]. The proof of achievability is based on a novel scalar-linear protocol, referred to as the Modified Specialized GRS Code protocol—a modified version of the Specialized GRS Code protocol, which achieves the rate ${(K-M+1)^{-1}}$ (see Section \[subsec:AchThm2\]). \[rem:2\] Interestingly, comparing the results of [@Kadhe2017 Theorem 2] and Theorem \[thm:PIRPCSI-II\], one can see that when the side information is composed of $M-1$ randomly chosen messages (different from the requested message), $({W},{S})$-privacy cannot be achieved more efficiently than the case in which the side information is only one* random linear coded combination of $M$ randomly chosen messages including the demand.* PIR-CSI {#subsec:CSIMainResults} ------- In this section, we present our main results for the PIR-CSI–I and PIR-CSI–II settings. The capacity and the scalar-linear capacity of the PIR-CSI–I setting (for all ${1\leq M\leq K-1}$) and the capacity and the scalar-linear capacity of the PIR-CSI–II setting (for all ${2\leq M\leq K}$) are characterized in Theorems \[thm:PIRCSI-I\] and \[thm:PIRCSI-II\], respectively. The proofs are given in Sections \[sec:PIRCSI-I\] and \[sec:PIRCSI-II\]. \[thm:PIRCSI-I\] For the PIR-CSI–I setting with $K$ messages and side information support size $M$, the capacity and the scalar-linear capacity are given by ${\lceil \frac{K}{M+1} \rceil}^{-1}$ for all $1\leq M\leq K-1$. The proof consists of two parts. In the first part, using information-theoretic arguments, we give an upper bound on the rate of any PIR-CSI–I protocol (see Section \[subsec:ConvThm3\]). The proofs rely primarily on the necessary condition for ${W}$-privacy, stated in Lemma \[prop:2\]. In the second part, we construct a new scalar-linear PIR-CSI–I protocol, termed the Modified Partition-and-Code (MPC) protocol, which achieves this rate upper-bound (see Section \[subsec:AchThm3\]). The proposed protocol is inspired by our recently proposed Partition-and-Code with Interference Alignment protocol in [@HS2019] for single-server private computation with uncoded side information. \[rem:3\] Interestingly, the capacity of PIR with (uncoded) side information [@Kadhe2017] is also equal to $\lceil\frac{K}{M+1}\rceil^{-1}$ where $M$ is the number of (uncoded) messages known to the user in advance as side information. This shows that there will be no loss in capacity, when compared to the case that the user knows $M$ randomly chosen messages separately, even if the user knows only one* random linear coded combination of $M$ randomly chosen messages.* \[rem:4\] When $({W},{S})$-privacy is required, the result of Theorem \[thm:PIRPCSI-I\] shows that the capacity of single-server PIR with a coded side information with support size $M$ that does not include the demand is equal to $(K-M)^{-1}$. Note that $\lceil\frac{K}{M+1}\rceil< K-M$ for all ${1\leq M\leq K-2}$. This implies that the capacity of the PIR-CSI–I setting is strictly greater than that of the PIR-PCSI–I setting for any ${1\leq M\leq K-2}$. This is expected because ${W}$-privacy is a weaker notion of privacy when compared to $({W},{S})$-privacy. However, for the extremal case of $M=K-1$, as can be seen $({W},{S})$-privacy comes at no extra cost compared to ${W}$-privacy. \[thm:PIRCSI-II\] For the PIR-CSI–II setting with $K$ messages and side information support size $M$, the capacity and the scalar-linear capacity are equal to $1$ for $M=2,K$, and $1/2$ for all $3\leq M\leq K-1$. For each range of values of $M$, the proof consists of two parts. In the first part, we use information-theoretic arguments—based on the result of Lemma \[prop:2\], so as to upper bound the rate of any PIR-CSI–II protocol (see Section \[subsec:ConvThm4\]). In the second part, we construct novel scalar-linear [PIR-CSI–II]{} protocols, collectively termed the Randomized Selection-and-Code (RSC) protocols, for different ranges of values of $M$. The proposed protocols rely on probabilistic techniques, and achieve the corresponding rate upper-bounds (see Section \[subsec:AchThm4\]). \[rem:5\] Interestingly, Theorem \[thm:PIRCSI-II\] shows that when ${W}$-privacy is required, no matter what the size of support set of the side information is, the user can privately retrieve any message belonging to the support set of their coded side information, with a download cost at most twice the cost of downloading the message directly—which obviously does not preserve the privacy of the requested message. \[rem:6\] As shown in Theorem \[thm:PIRPCSI-II\], when $({W},{S})$-privacy is required, the (scalar-linear) capacity of single-server PIR with a coded side information whose support set includes the requested message is equal to $(K-M+1)^{-1}$, where $M$ is the side information support size. The result of Theorem \[thm:PIRCSI-II\] matches this result for the cases of $M=K$ and $M=K-1$, and hence, $({W},{S})$-privacy and ${W}$-privacy are attainable at the same cost in these cases; whereas for the other cases of $M$, achieving $({W},{S})$-privacy is much more costly than achieving ${W}$-privacy. Proof of Theorem \[thm:PIRPCSI-I\] {#sec:PIRPCSI-I} ================================== Converse {#subsec:ConvThm1} -------- As shown in [@Kadhe2017] using an index-coding argument, when $({W},{S})$-privacy is required, the capacity of PIR with $M$ uncoded messages as side information is given by ${(K-M)^{-1}}$. Obviously, the capacity of the PIR-PCSI–I setting is upper bounded by this quantity. This proves the converse for Theorem \[thm:PIRPCSI-I\]. In this section, we present an alternative information-theoretic proof for the general case, which also proves the converse for the scalar-linear case. \[lem:Conv1\] For any $1\leq M\leq K-1$, the (scalar-linear) capacity of the PIR-PCSI–I setting is upper bounded by ${(K-M)^{-1}}$. [[[* Proof.* ]{}]{}]{}In the following, all entropies are conditional on the event $\mathds{1}_{\{{\mathbf{W}}\in {\mathbf{S}}\}}=0$, and we remove this event from the conditions everywhere, for the ease of notation. We need to show that ${H({\mathbf{A}})\geq (K-M)L}$. Take arbitrary ${W},{S},{C}$ (and ${\mathbf{Y}}\triangleq {\mathbf{Y}}^{[{S},{C}]}$) such that ${W}\not\in {S}$. Then, we have $$\begin{aligned} H({\mathbf{A}}) &\geq H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}}) \label{eq:PCSIlineI1}\\ & = H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}})+H({\mathbf{X}}_{{W}}\|{\mathbf{A}},{\mathbf{Q}},{\mathbf{Y}}) \label{eq:PCSIlineI2}\\ &=H({\mathbf{A}},{\mathbf{X}}_{{W}}\|{\mathbf{Q}},{\mathbf{Y}})\label{eq:PCSIlineI3}\\ &=H({\mathbf{X}}_{{W}}\|{\mathbf{Q}},{\mathbf{Y}}) + H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}}) \label{eq:PCSIlineI4}\\ &=H({\mathbf{X}}_{{W}})+H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}})\label{eq:PCSIlineI5}\end{aligned}$$ where follows since conditioning does not increase the entropy; holds because $H({\mathbf{X}}_{{W}}\|{\mathbf{A}},{\mathbf{Q}},{\mathbf{Y}})=0$ (by the recoverability condition); and follow from the chain rule of entropy; and follows from $H({\mathbf{X}}_{{W}}\|{\mathbf{Q}},{\mathbf{Y}}) = H({\mathbf{X}}_{{W}})$ since ${\mathbf{X}}_{{W}}$ is independent of $({\mathbf{Q}},{\mathbf{Y}})$ (noting that ${W}\not\in {S}$). If ${W}\cup {S} = \mathcal{K}$ (i.e., $M=K-1$), then $H({\mathbf{A}})\geq H({\mathbf{X}}_{{W}})=L$ (by using the first term in ), as was to be shown. If ${W}\cup {S} \neq \mathcal{K}$, we proceed by lower bounding the second term in , $H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}})$. By Lemma \[prop:1\], for each ${i\in \mathcal{K}\setminus ({W}\cup {S})}$, there exists ${C}_{i}\in \mathcal{C}$ (and ${\mathbf{Y}}_i\triangleq {\mathbf{Y}}^{[{S},{C}_i]}$) such that $H({\mathbf{X}}_{i}\|{\mathbf{A}},{\mathbf{Q}},{\mathbf{Y}}_{i})=0$. Let $I$ be a maximal subset of ${\mathcal{K}\setminus ({W}\cup {S})}$ such that ${\mathbf{Y}}$ and ${\mathbf{Y}}_{I}\triangleq \{{\mathbf{Y}}_{i}\}_{i\in I}$ are linearly independent. (Note that ${\|I\|\leq \|{S}\|-1=M-1}$.) Let ${\mathbf{X}}_{I}\triangleq \{{\mathbf{X}}_i\}_{i\in I}$. Then, we have $$\begin{aligned} H(\mathbf{A}\|\mathbf{Q},\mathbf{Y},\mathbf{X}_{{W}}) & \geq H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_I) \nonumber\\ &\geq H(\mathbf{A}\|\mathbf{Q},\mathbf{Y},\mathbf{X}_{{W}},\mathbf{Y}_I) \nonumber \\ &\quad + H(\mathbf{X}_I\|\mathbf{A},\mathbf{Q},\mathbf{Y},\mathbf{X}_{{W}},\mathbf{Y}_I)\label{eq:PCSIlineI6}\\ &= H(\mathbf{A},\mathbf{X}_I\|\mathbf{Q},\mathbf{Y},\mathbf{X}_{{W}},\mathbf{Y}_I)\nonumber\\ & = H(\mathbf{X}_I\|\mathbf{Q},\mathbf{Y},\mathbf{X}_{{W}},\mathbf{Y}_I)\nonumber\\ & \quad + H(\mathbf{A}\|\mathbf{Q},\mathbf{Y},\mathbf{X}_{{W}},\mathbf{Y}_I,\mathbf{X}_I)\nonumber\\ & = H(\mathbf{X}_I) \nonumber \\ & \quad +H(\mathbf{A}\|\mathbf{Q},\mathbf{Y},\mathbf{X}_{{W}},\mathbf{Y}_I,\mathbf{X}_I)\label{eq:PCSIlineI7} \end{aligned}$$ where holds because $H(\mathbf{X}_{i}\|\mathbf{A},\mathbf{Q},\mathbf{Y}_{i})=0$ for all $i\in I$ (by assumption); and holds since $\mathbf{X}_I$ is independent of $(\mathbf{Q},\mathbf{Y},\mathbf{X}_{{W}},\mathbf{Y}_I)$ by construction (noting that $I$ and ${{W}\cup {S}}$ are disjoint). The first term in , $H({\mathbf{X}}_{I})$, is lower bounded by $\|I\|L\geq 0$. Thus, in order to further lower bound $H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}})$, we need to lower bound the second term in , $H(\mathbf{A}\|\mathbf{Q},\mathbf{Y},\mathbf{X}_{{W}},\mathbf{Y}_I,\mathbf{X}_I)$. By the maximality of $I$, for each $j\in J\triangleq {\mathcal{K}\setminus ({W}\cup {S}\cup I)}$, there exists ${C}_j\in \mathcal{C}$ (and $\mathbf{Y}_j\triangleq \mathbf{Y}^{[{S},{C}_j]}$, which is linearly dependent on $\mathbf{Y}$ and $\mathbf{Y}_I$) such that $H(\mathbf{X}_j\|\mathbf{A},\mathbf{Q},\mathbf{Y}_j) = 0$, and as a consequence, $H(\mathbf{X}_j\|\mathbf{A},\mathbf{Q},\mathbf{Y},\mathbf{Y}_I) = 0$. (Note that $\|J\|={K-M-1-\|I\|}$.) Let $\mathbf{X}_J\triangleq \{X_j\}_{j\in J}$. Then, we can write $$\begin{aligned} & H(\mathbf{A}\|\mathbf{Q},\mathbf{Y},\mathbf{X}_{{W}},\mathbf{Y}_I,\mathbf{X}_I)\nonumber\\ & \quad = H(\mathbf{A}\|\mathbf{Q},\mathbf{Y},\mathbf{X}_{{W}},\mathbf{Y}_I,\mathbf{X}_I)\nonumber\\ & \quad\quad + H(\mathbf{X}_{J}\|\mathbf{A},\mathbf{Q},\mathbf{Y},\mathbf{X}_{{W}},\mathbf{Y}_{I},\mathbf{X}_I)\label{eq:PCSIlineI8}\\ & \quad = H({\mathbf{A}},{\mathbf{X}}_{J}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_{I},{\mathbf{X}}_I)\nonumber\\ & \quad = H(\mathbf{X}_J\|\mathbf{Q},\mathbf{Y},\mathbf{X}_{{W}},\mathbf{Y}_I,\mathbf{X}_I)\nonumber\\ & \quad \quad + H(\mathbf{A}\|\mathbf{Q},\mathbf{Y},\mathbf{X}_{{W}},\mathbf{Y}_I,\mathbf{X}_I,\mathbf{X}_J)\nonumber\\ & \quad \geq H(\mathbf{X}_{J})\label{eq:PCSIlineI9}\end{aligned}$$ where holds since $H(\mathbf{X}_{j}\|\mathbf{A},\mathbf{Q},\mathbf{Y},\mathbf{Y}_{I})=0$ for all $j\in J$ (by assumption); and holds because $\mathbf{X}_{J}$ and $(\mathbf{Q},\mathbf{Y},\mathbf{X}_{{W}},\mathbf{Y}_I,\mathbf{X}_I)$ are independent by construction (noting that $J$ and ${{W}\cup {S}\cup I}$ are disjoint). Putting , , and together, $H(\mathbf{A})\geq H(\mathbf{X}_{{W}})+H(\mathbf{X}_I)+H(\mathbf{X}_{J}) = L+\|I\|L+ (K-M-1-\|I\|)L= (K-M)L$, as was to be shown. Achievability {#subsec:AchThm1} ------------- In this section, we propose a scalar-linear PIR-PCSI–I protocol that achieves the rate $(K-M)^{-1}$. The proposed protocol requires a base-field size $q\geq K$ (and arbitrary field-extension degree $l\geq 1$) where the messages $X_i$’s are elements from $\mathbb{F}_{q^{l}}$. It is noteworthy that the rate $(K-M)^{-1}$ is not necessarily achievable for ${q<K}$, and for the special case of scalar-linear schemes, the achievability of this rate is conditional upon the existence of a $(K,K-M)$ maximum distance seperable (MDS) code over $\mathbb{F}_q$ that has a codeword with support ${W}\cup {S}$ such that the $j$th code symbol is non-zero for $j={W}$ and it is equal to ${c}_j$ for each $j\in {S}$ where ${c}_j$ is the coefficient of the message ${X}_j$ in the coded side information ${Y}^{[{S},{C}]}$. Specialized GRS Code Protocol: This protocol consists of three steps as follows: Step 1: First, the user arbitrarily chooses $K$ distinct elements $\omega_1,\dots,\omega_K$ from $\mathbb{F}_q$, and constructs a polynomial $${p(x)\triangleq \prod_{i\in \mathcal{K}\setminus ({W}\cup {S})} (x-\omega_i)}.$$ Then, the user constructs $K-M$ (ordered) sets ${Q}_1,\dots,{Q}_{K-M}$, each of size $K$, defined as $${Q}_i=\{v_1\omega_1^{i-1},\dots,v_K\omega_K^{i-1}\},$$ where the parameters $v_j$’s are chosen as follows. For each $j\in {S}$, $v_j=\frac{c_j}{p(\omega_j)}$ where $c_j$ is the coefficient of $X_j$ in $Y^{[{S},{C}]}$; and for each $j\not\in {S}$, $v_j$ is chosen at random from $\mathbb{F}_q^{\times}$. The user then sends to the server the query ${Q}^{[{W},{S},{C}]} = \{{Q}_{1},\dots,{Q}_{K-M}\}$. Note that the $j$th element in the set ${Q}_i$ can be thought of as the entry $(i,j)$ of a $(K-M)\times K$ matrix $G\triangleq {[g_1^{{T}},\dots,g_{K-M}^{{T}}]}^{{T}}$, which generates a $(K,K-M)$ GRS code with distinct parameters ${\omega_1,\dots,\omega_{K}}$ and non-zero multipliers $v_1,\dots,v_K$ [@Roth:06]. This construction ensures that such a GRS code has a specific codeword, namely $\sum_{i=1}^{K-M} p_{i} g_{i}$ where $p_i$ is the coefficient of $x^{i-1}$ in the expansion of the polynomial $p(x)= \sum_{i=1}^{K-M} p_{i}x^{i-1}$, with support ${W}\cup {S}$ such that the $j$th code symbol is non-zero for $j={W}$, and it is equal to $c_j$ for each $j\in {S}$. This observation is the chief idea in the proof of the recoverability condition for the proposed protocol. Step 2: By using ${Q}_i$’s, the server computes $A_i$’s, defined as $A_{i} = \sum_{j=1}^{K} v_j\omega_j^{i-1} X_{j}$, and it sends the answer ${A}^{[{W},{S},{C}]}=\{A_{1},\dots,A_{K-M}\}$ to the user. Note that $A_i$’s are the parity check equations of a $(K,M)$ GRS code which is the dual code of the GRS code generated by the matrix $G$ defined earlier. *Step 3:* Upon receiving the answer, the user retrieves $X_{{W}}$ by subtracting off the contribution of the side information $Y^{[{S},{C}]}$ from $\sum_{i=1}^{K-M} p_{i} A_{i} = c_{{W}} X_{{W}}+\sum_{i\in {S}} c_{i}X_{i}$. \[ex:PCSI-I\] Consider a scenario where the server has ${K=4}$ messages $X_1,\dots,X_{4}\in \mathbb{F}_{5}$, and the user demands the message $X_1$ and has a coded side information $Y = X_2+X_3$ with support size $M=2$. For this example, ${W} = 1$, ${S}= \{2,3\}$, and ${{C} =\{c_2,c_3\}= \{1,1\}}$. First, the user chooses ${K=4}$ distinct elements $\omega_1,\dots,\omega_4$ from $\mathbb{F}_5$, say ${(\omega_1,\omega_2,\omega_3,\omega_4)=(0,1,2,3)}$. Then, the user constructs the polynomial $$p(x) = \prod_{i\not\in {W}\cup {S}} (x-\omega_i)={x-\omega_4}= x+2.$$ Note that $p(x) = p_1+p_2x=2+x$. The user then computes $v_j$ for $j\in {S}$, i.e., $v_2$ and $v_3$, by setting ${v_2=\frac{c_2}{p(\omega_2)}=2}$ and ${v_3=\frac{c_3}{p(\omega_3)}=4}$, and chooses $v_j$ for ${j\not\in {S}}$, i.e., $v_1$ and $v_4$, at random (from $\mathbb{F}^{\times}_5$). Suppose that the user chooses $v_1=1$ and $v_4=2$. Then, the user constructs ${K-M=2}$ (ordered) sets ${Q}_1 = \{v_1,\dots,v_4\}=\{1,2,4,2\}$ and ${Q}_2 = \{v_1\omega_1,\dots,v_4\omega_4\}=\{0,2,3,1\}$. The user then sends the query $Q = \{Q_1,Q_2\}$ to the server. The server computes $A_1 = \sum_{j=1}^{4}v_j X_j = X_1+2X_2+4X_3+2X_4$ and $A_2 = \sum_{j=1}^{4}v_j\omega_j X_j = 2X_2+3X_3+X_4$, and sends the answer $A = \{A_1,A_2\}$ back to the user. Then, the user computes $\sum_{i=1}^{2} p_{i}A_i = 2A_1+A_2 = 2X_1+X_2+X_3$, and recovers $X_1$ by subtracting off $Y = X_2+X_3$. For this example, the rate of the proposed protocol is $1/2$. Note that the server knows the protocol, including the parameters $\omega_1,\dots,\omega_4$, and can compute the multipliers $v_1,\dots,v_4$, given the query. Since the side information coefficients $c_2$ and $c_3$ are uniformly distributed, the server finds each of the polynomials $x-\omega_1=x$, $x-\omega_2=4+x$, $x-\omega_3=3+x$, and $x-\omega_4=2+x$ equally likely to be the polynomial $p(x) = p_1+p_2x$, constructed in Step 1 of the protocol. Since the server knows that by the protocol the user requires the linear combination $p_1A_1+p_2A_2$ to recover the demand, from the server’s perspective, each of the linear combinations $Z_1 = A_2$, $Z_2 = 4A_1+A_2$, $Z_3 = 3A_1+A_2$, $Z_4 = 2A_1+A_2$, i.e., $Z_1 = 2X_2+3X_3+X_4$, $Z_2 = 4X_1+4X_3+4X_4$, $Z_3 = 3X_1+3X_2+2X_4$, $Z_4 = 2X_1+X_2+X_3$, are equally likely to be the linear combination required by the user. Note, also, that, for each candidate demand index (e.g., $\{1\}$) and each candidate side information support index set (e.g., $\{2,3\}$), there exists exactly one of the linear combinations $Z_1,\dots,Z_4$ (e.g., $Z_4$) from which the candidate demand (e.g., $X_{1}$) can be recovered, given some linear combination (e.g., $X_2+X_3$) of the messages in the candidate side information support set (e.g., $X_2,X_3$). By these arguments, the server finds every index $i\in \{1,\dots,5\}$ and every pair of indices $\{i_1,i_2\}$ such that $i\not\in \{i_1,i_2\}$ equally likely to be the user’s demand index and side information support index set, respectively. This confirms that the proposed protocol achieves $({W},{S})$-privacy in this example. \[lem:Ach1\] The Specialized GRS Code protocol is a scalar-linear PIR-PCSI–I protocol, and achieves the rate $(K-M)^{-1}$. [[[* Proof.* ]{}]{}]{}See Appendix \[subsec:PLAch1\]. Proof of Theorem \[thm:PIRPCSI-II\] {#sec:PIRPCSI-II} =================================== Converse {#subsec:ConvThm2} -------- First, we prove the converse for the scalar-linear case of Theorem \[thm:PIRPCSI-II\] for all $2\leq M\leq K$. The proof is based on a combination of algebraic and information-theoretic arguments. \[lem:Conv2\] For any $2\leq M\leq K$, the scalar-linear capacity of the PIR-PCSI–II setting is upper bounded by ${(K-M+1)^{-1}}$. [[[* Proof.* ]{}]{}]{}In the following, all the entropies are conditional on the event $\mathds{1}_{\{{\mathbf{W}}\in {\mathbf{S}}\}}=1$, and for simplifying the notation, we remove this event from the conditions. We need to show that $H(\mathbf{A})\geq {(K-M+1)L}$. Let $I$ be the set of all $i\in \mathcal{K}$ such that $H(\mathbf{X}_i\|\mathbf{A},\mathbf{Q})=0$. (Note that $0\leq \|I\|\leq K$). Let $\mathbf{X}_{I}\triangleq \{\mathbf{X}_i\}_{i\in I}$. By assumption, $\mathbf{X}_I$ and $\mathbf{Q}$ are independent and $H(\mathbf{X}_I\|\mathbf{A},\mathbf{Q})=0$. Then, we can write $$\begin{aligned} H(\mathbf{A}) & \geq H(\mathbf{A}\|\mathbf{Q})\nonumber \\ & = H(\mathbf{A}\|\mathbf{Q})+H(\mathbf{X}_I\|\mathbf{A},\mathbf{Q})\nonumber\\ & = H(\mathbf{A},\mathbf{X}_I\|\mathbf{Q})\nonumber \\ & = H(\mathbf{X}_I\|\mathbf{Q})+H(\mathbf{A}\|\mathbf{Q},\mathbf{X}_I)\nonumber\\ & = H(\mathbf{X}_I)+H(\mathbf{A}\|\mathbf{Q},\mathbf{X}_I).\label{eq:PCSIlineII10}\end{aligned}$$ If ${\|I\|\geq K-M+1}$, the first term in , $H({\mathbf{X}}_I)$, is lower bounded by $(K-M+1)L$, and hence, $H(\mathbf{A})\geq {(K-M+1)L}$, as was to be shown. If $0\leq \|I\|\leq K-M$, the second term in , $H(\mathbf{A}\|\mathbf{Q},\mathbf{X}_I)$, can be further lower bounded as follows. Assume, w.l.o.g., that $I=\{1,\dots,\|I\|\}$. (Note that $I = \emptyset$ for $\|I\|=0$.) Let $J\triangleq \{1,\dots,K-M-\|I\|+1\}$, and let ${S}_j\triangleq {\{\|I\|+1,\|I\|+j+1,\dots,\|I\|+j+M-1\}}$ for $j\in J$. By Lemma \[prop:1\], for each $j\in J$, there exists ${C}_j\in \mathcal{C}$ (and $\mathbf{Y}_j\triangleq \mathbf{Y}^{[{S}_j,{C}_j]}$) such that $H(\mathbf{X}_{\|I\|+1}\|\mathbf{A},\mathbf{Q},\mathbf{Y}_j)=0$. Let $\mathbf{Z}_j\triangleq \mathbf{Y}_j - c_j\mathbf{X}_{\|I\|+1}$ where $c_j$ is the coefficient of $\mathbf{X}_{\|I\|+1}$ in $\mathbf{Y}_j$. For any scalar-linear protocol where the answer consists only of scalar-linear combinations of messages in $X_{\mathcal{K}}$, it is easy to see that for each $j\in J$, (i) $H(\mathbf{Z}_j\|\mathbf{A},\mathbf{Q})=0$, or (ii) ${H(\mathbf{Z}_j+c\mathbf{X}_{\|I\|+1}\|\mathbf{A},\mathbf{Q})=0}$ for some $c\in \mathbb{F}^{\times}_q\setminus \{c_j\}$. (Otherwise, the server learns that ${W}$ and ${S}$ cannot be $\|I\|+1$ and ${S}_j$, respectively. This obviously violates the $({W},{S})$-privacy condition.) In either case (i) or (ii), one can see that $H(\mathbf{Z}_j\|\mathbf{A},\mathbf{Q},\mathbf{X}_{\|I\|+1})=0$. (Note that this observation, which is the key in the proof of Lemma \[lem:Conv2\], holds for all scalar-linear schemes, but not necessarily for all vector-linear or non-linear schemes in general. This implies the need for a different proof technique for the general schemes, and an example of such a technique is used in the proof of Lemma \[lem:GConv2\].) Let $\mathbf{Z}_{J}\triangleq \{\mathbf{Z}_j\}_{j\in J}$. Then, we have $$\begin{aligned} H(\mathbf{A}\|\mathbf{Q},\mathbf{X}_I) & \geq H(\mathbf{A}\|\mathbf{Q},\mathbf{X}_I,\mathbf{X}_{\|I\|+1})\nonumber \\ & = H(\mathbf{A}\|\mathbf{Q},\mathbf{X}_I,\mathbf{X}_{\|I\|+1})\nonumber\\ & \quad + H(\mathbf{Z}_J\|\mathbf{A},\mathbf{Q},\mathbf{X}_I,\mathbf{X}_{\|I\|+1})\label{eq:PCSIlineII11}\\ & = H(\mathbf{A},\mathbf{Z}_J\|\mathbf{Q},\mathbf{X}_I,\mathbf{X}_{\|I\|+1})\nonumber\\ & = H(\mathbf{Z}_J\|\mathbf{Q},\mathbf{X}_I,\mathbf{X}_{\|I\|+1}) \nonumber \\ &\quad + H(\mathbf{A}\|\mathbf{Q},\mathbf{X}_I,\mathbf{X}_{\|I\|+1},\mathbf{Z}_J)\nonumber\\ & \geq H(\mathbf{Z}_J) \label{eq:PCSIlineII12}\end{aligned}$$ where holds since $H(\mathbf{Z}_j\|\mathbf{A},\mathbf{Q},\mathbf{X}_{\|I\|+1})=0$ for all $j\in J$ (by assumption); and follows because $\mathbf{Z}_J$ is independent of $(\mathbf{Q},\mathbf{X}_I,\mathbf{X}_{\|I\|+1})$ by construction, noting that $\mathbf{Z}_J$, $\mathbf{X}_I$, and $\mathbf{X}_{\|I\|+1}$ are linearly independent. By the linear independence of $\mathbf{Z}_j$’s for all $j\in J$, it follows that $H(\mathbf{Z}_J) = {(K-M-\|I\|+1)L}$. By and , we get $H(\mathbf{A})\geq H({\mathbf{X}}_I) +H({\mathbf{Z}}_J) = {\|I\|L}+{(K-M-\|I\|+1)L} ={(K-M+1)L}$, as was to be shown. Next, we give an information-theoretic proof of converse for the general case of Theorem \[thm:PIRPCSI-II\] for all ${\frac{K+1}{2}< M\leq K}$. For any $2\leq M\leq \frac{K+1}{2}$, the converse proof remains open. \[lem:GConv2\] For any $\frac{K+1}{2}<M\leq K$, the capacity of the PIR-PCSI–II setting is upper bounded by ${(K-M+1)^{-1}}$. [[[* Proof.* ]{}]{}]{}Similar to the proof of Lemma \[lem:Conv2\], for the ease of notation in the following we remove the event $\mathds{1}_{\{{\mathbf{W}}\in {\mathbf{S}}\}}=1$ from the conditions of all the entropies. We need to show that $H({\mathbf{A}})\geq {(K-M+1)L}$. Let $J\triangleq \{1,\dots,K-M+1\}$ and ${S}_j\triangleq \{j,\dots,{j+M-1}\}$ for $j\in J$. By Lemma \[prop:1\], for each $j\in J$, there exists ${C}_j\in \mathcal{C}$ (and ${\mathbf{Y}}_j\triangleq {\mathbf{Y}}^{[{S}_j,{C}_j]}$) such that $H({\mathbf{X}}_{j}\|{\mathbf{A}},{\mathbf{Q}},{\mathbf{Y}}_j)=0$. Let ${\mathbf{X}}_J\triangleq \{{\mathbf{X}}_j\}_{j\in J}$. (Note that $\|J\|={K-M+1}<M$ when $M>\frac{K+1}{2}$). Then, we can write $$\begin{aligned} \label{eq:PCSIlineII13} H({\mathbf{X}}_J,{\mathbf{Y}}_J\|{\mathbf{Q}}) & = H({\mathbf{X}}_J,{\mathbf{Y}}_J) \\ \label{eq:PCSIlineII14} & = 2(K-M+1)L, \end{aligned}$$ where holds since ${\mathbf{Q}}$ is independent of $({\mathbf{X}}_J,{\mathbf{Y}}_J)$ (by assumption); and follows because ${\mathbf{X}}_J$ and ${\mathbf{Y}}_J$ are independent by construction. (Note that ${\mathbf{X}}_J$ and ${\mathbf{Y}}_J$ are not necessarily independent for $\|J\| = K-M+1\geq M$, and a different technique—which remains open, is required for the proof of converse when $2\leq M\leq \frac{K+1}{2}$.) Moreover, we have $$\begin{aligned} H({\mathbf{X}}_J,{\mathbf{Y}}_J\|{\mathbf{A}},{\mathbf{Q}}) & \leq \sum_{j\in J}H({\mathbf{X}}_j,{\mathbf{Y}}_j\|{\mathbf{A}},{\mathbf{Q}}) \label{eq:PCSIlineII15}\\ & = \sum_{j\in J} H({\mathbf{Y}}_j\|{\mathbf{A}},{\mathbf{Q}}) \nonumber \\ & \quad + \sum_{j\in J} H({\mathbf{X}}_j\|{\mathbf{A}},{\mathbf{Q}},{\mathbf{Y}}_j)\nonumber\\ & = \sum_{j\in J} H({\mathbf{Y}}_j\|{\mathbf{A}},{\mathbf{Q}})\label{eq:PCSIlineII16}\\ & \leq \sum_{j\in J} H({\mathbf{Y}}_j) \nonumber\\ & = (K-M+1)L, \label{eq:PCSIlineII17} \end{aligned}$$ where follows from the chain rule of entropy; holds because $H({\mathbf{X}}_{j}\|{\mathbf{A}},{\mathbf{Q}},{\mathbf{Y}}_j)=0$ for $j\in J$ (by assumption); and holds because ${\mathbf{Y}}_j$’s for all $j\in J$ are independent by construction, and ${\mathbf{Y}}_j$ for each $j\in J$ is a scalar-linear combination of ${\mathbf{X}}_j,{\mathbf{X}}_{j+1},\dots,{\mathbf{X}}_{j+M-1}$. Using and , we can bound $H({\mathbf{X}}_J,{\mathbf{Y}}_J,{\mathbf{A}}\|{\mathbf{Q}})$ from below and above. On the one hand, we have $$\begin{aligned} \nonumber H({\mathbf{X}}_J,{\mathbf{Y}}_J,{\mathbf{A}}\|{\mathbf{Q}}) & \geq H({\mathbf{X}}_J,{\mathbf{Y}}_J\|{\mathbf{Q}}) \nonumber \\ & = 2(K-M+1)L, \label{eq:PCSIlineII18}\end{aligned}$$ where follows from . On the other hand, we have $$\begin{aligned} H({\mathbf{X}}_J,{\mathbf{Y}}_J,{\mathbf{A}}\|{\mathbf{Q}}) & = H({\mathbf{A}}\|{\mathbf{Q}}) \nonumber\\ & \quad +H({\mathbf{X}}_J,{\mathbf{Y}}_J\|{\mathbf{A}},{\mathbf{Q}}) \nonumber\\ & \leq H({\mathbf{A}}\|{\mathbf{Q}}) \nonumber \\ & \quad +(K-M+1)L, \label{eq:PCSIlineII19} \end{aligned}$$ where follows from . Now, combining and , we have $H({\mathbf{A}}\|{\mathbf{Q}})\geq (K-M+1)L$, and as a consequence, $H({\mathbf{A}})\geq H({\mathbf{A}}\|{\mathbf{Q}})\geq (K-M+1)L$, as was to be shown. Achievability {#subsec:AchThm2} ------------- In this section, we propose a scalar-linear PIR-PCSI–II protocol, termed the Modified Specialized GRS Code protocol, that achieves the rate $(K-M+1)^{-1}$. For this protocol, the requirements for the parameters $q$ and $l$ are the same as those for the Specialized GRS Code protocol. Modified Specialized GRS Code Protocol: This protocol consists of three steps, where the steps 2-3 are the same as Steps 2-3 in the Specialized GRS Code protocol (Section \[subsec:AchThm1\]), when the parameter $M$ is replaced with $M-1$ everywhere. The step 1 of the proposed protocol is as follows: Step 1: For $K$ arbitrarily chosen distinct elements $\omega_1,\dots,\omega_K$ from $\mathbb{F}_q$, the user constructs a polynomial $${p(x) = \sum_{i=1}^{K-M+1} p_i x^{i-1} \triangleq \prod_{i\in\mathcal{K}\setminus {S}} (x-\omega_i)},$$ and constructs $K-M+1$ (ordered) sets $Q_1,\dots,Q_{K-M+1}$, each of size $K$, defined as $$Q_i=\{v_1\omega_1^{i-1},\dots,v_K\omega_K^{i-1}\},$$ where $v_j$’s are chosen as follows. For each $j\in {S}\setminus {W}$, $v_j=\frac{c_j}{p(\omega_j)}$ where $c_j$ is the coefficient of $X_j$ in $Y^{[{S},{C}]}$; ${v_{{W}}=\frac{c}{p(\omega_{{W}})}}$ for a randomly chosen element $c$ from ${\mathbb{F}^{\times}_q\setminus \{c_{{W}}\}}$ where $c_{{W}}$ is the coefficient of $X_{{W}}$ in $Y^{[{S},{C}]}$; and for each $j\not\in {S}$, $v_j$ is chosen at random from $\mathbb{F}_q^{\times}$. The user then sends to the server the query ${Q}^{[{W},{S},{C}]} = \{{Q}_{1},\dots,{Q}_{K-M+1}\}$. Consider a scenario where the server has ${K=4}$ messages $X_1,\dots,X_{4}\in \mathbb{F}_{5}$, and the user demands the message $X_1$ and has a coded side information $Y = X_1+X_2$ with support size $M=2$. For this example, ${W} = 1$, ${S}= \{1,2\}$, and ${{C} =\{c_1,c_2\}= \{1,1\}}$. First, the user chooses ${K=4}$ distinct elements $\omega_1,\dots,\omega_4$ from $\mathbb{F}_5$, say ${(\omega_1,\omega_2,\omega_3,\omega_4)=(0,1,2,3)}$. Then, the user constructs the polynomial $$p(x) = \prod_{i\not\in {S}} (x-\omega_i)=(x-\omega_3)(x-\omega_4)= (x+3)(x+2).$$ Note that $p(x) = p_1+p_2x+p_3x^2=1+x^2$. The user then computes $v_j$ for $j\in {S}\setminus {W}$, i.e., $v_2$, by setting ${v_2=\frac{c_2}{p(\omega_2)}=3}$; computes $v_{{W}}$, i.e., $v_1$, for a randomly chosen element $c$, say $c = 4$, from $\mathbb{F}^{\times}_5\setminus \{c_1 = 1\}$ by setting $v_1 = \frac{c}{p(\omega_1)} = 4$; and chooses $v_j$ for ${j\not\in {S}}$, i.e., $v_3$ and $v_4$, at random (from $\mathbb{F}^{\times}_5$). Suppose that the user chooses $v_3=1$ and $v_4=3$. Then, the user constructs ${K-M+1=3}$ (ordered) sets ${Q}_1 = \{v_1,\dots,v_4\}=\{4,3,1,3\}$, ${Q}_2 = \{v_1\omega_1,\dots,v_4\omega_4\}=\{0,3,2,4\}$, and ${Q}_3 = \{v_1\omega^2_1,\dots,v_4\omega^2_4\}=\{0,3,4,2\}$. The user then sends the query $Q = \{Q_1,Q_2,Q_3\}$ to the server. The server computes $A_1 = \sum_{j=1}^{4}v_j X_j = 4X_1+3X_2+X_3+3X_4$, $A_2 = \sum_{j=1}^{4}v_j\omega_j X_j = 3X_2+2X_3+4X_4$, and $A_3 = \sum_{j=1}^{4}v_j\omega^2_j X_j = 3X_2+4X_3+2X_4$, and sends the answer $A = \{A_1,A_2,A_3\}$ back to the user. Then, the user computes $\sum_{i=1}^{3} p_{i}A_i = A_1+A_3 = 4X_1+X_2$, and recovers $X_1$ by subtracting off $Y = X_1+X_2$. For this example, the rate of the proposed protocol is $1/3$. The proof of $({W},{S})$-privacy for the proposed protocol in this example is similar to the proof of $({W},{S})$-privacy for the Specialized GRS Code protocol in Example \[ex:PCSI-I\]. \[lem:Ach2\] The Modified Specialized GRS Code protocol is a scalar-linear PIR-PCSI–II protocol, and achieves the rate $(K-M+1)^{-1}$. [[[* Proof.* ]{}]{}]{}See Appendix \[subsec:PLAch2\]. Proof of Theorem \[thm:PIRCSI-I\] {#sec:PIRCSI-I} ================================= Converse {#subsec:ConvThm3} -------- The capacity of the PIR-CSI–I setting is naturally upper bounded by the capacity of PIR with uncoded side information [@Kadhe2017] where $M$ uncoded messages are available at the user as side information. As shown in [@Kadhe2017], the capacity of this problem is equal to $\lceil \frac{K}{M+1} \rceil^{-1}$, and the proof of this result relies on an index coding argument. In this section, we present an alternative converse proof for the case of general PIR-CSI–I protocols, by using information-theoretic arguments. Obviously, this proof also serves for the special case of scalar-linear PIR-CSI–I protocols. \[lem:Converse1\] For any $1\leq M\leq K-1$, the (scalar-linear) capacity of the PIR-CSI–I setting is upper bounded by ${\lceil \frac{K}{M+1}\rceil}^{-1}$. [[[* Proof.* ]{}]{}]{} In the following, all entropies are conditional on the event $\mathds{1}_{\{{\mathbf{W}}\in {\mathbf{S}}\}}=0$, and this event is removed from the conditions for the ease of notation. We need to show that $H({\mathbf{A}})\geq \lceil\frac{K}{M+1}\rceil L$. Take arbitrary ${W},{S},{C}$ (and ${\mathbf{Y}}\triangleq {\mathbf{Y}}^{[{S},{C}]}$) such that ${W}\not\in {S}$. Similar to the proof of Lemma \[lem:Conv1\], it can be shown that $$\label{eq:CSIlineI20} H({\mathbf{A}})\geq H({\mathbf{X}}_{{W}})+H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}}).$$ There are two cases: (i) $W \cup S = \mathcal{K}$, and (ii) $W \cup S\neq \mathcal{K}$. In the case (i), $M = K-1$, and so, $\lceil \frac{K}{M+1}\rceil L = L$. Since $H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}})\geq 0$, then $H({\mathbf{A}})\geq H({\mathbf{X}}_{{W}}) = L$ (by ), as was to be shown. In the case (ii), we proceed by lower bounding $H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}})$ as follows. We arbitrarily choose a message, say ${\mathbf{X}}_{{W}_1}$, for some ${{W}_1 \not\in {W} \cup {S}}$. By Lemma \[prop:2\], there exist ${S}_1\in \mathcal{S}$ with ${W}_1\not\in {S}_1$ and ${C}_1\in \mathcal{C}$ (and ${\mathbf{Y}}_1 = {\mathbf{Y}}^{[{S}_1,{C}_1]}$) such that $H({\mathbf{X}}_{{W}_1}\|{\mathbf{A}},{\mathbf{Q}},{\mathbf{Y}}_1) = 0$. Since conditioning does not increase the entropy, then $H({\mathbf{X}}_{{W}_1}\|{\mathbf{A}},{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1) = 0$. Thus, $$\begin{aligned} H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}}) &\geq H\big({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1\big)\nonumber\\ & = H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1) \nonumber\\ & \quad +H({\mathbf{X}}_{{W}_1}\|{\mathbf{A}},{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1)\nonumber \\ &= H\big({\mathbf{A}},{\mathbf{X}}_{{W}_1}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1\big)\nonumber \\ &=H({\mathbf{X}}_{{W}_1}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1)\nonumber \\ & \quad +H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1,{\mathbf{X}}_{{W}_1})\nonumber\\ & = H({\mathbf{X}}_{{W}_1})\nonumber \\ & \quad +H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1,{\mathbf{X}}_{{W}_1})\label{eq:CSIlineI21}\end{aligned}$$ where holds because ${\mathbf{X}}_{{W}_1}$ and $({\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1)$ are independent (noting that ${{W}_1\not\in {W}\cup {S}\cup {S}_1}$), and hence, $H({\mathbf{X}}_{{W}_1}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1) = H({\mathbf{X}}_{{W}_1})$. Let $n \triangleq \lceil\frac{K}{M+1}\rceil$. Similarly as above, it can be shown that for all ${1\leq i\leq n-1}$ there exist ${W}_1,\dots,{W}_{i}\in \mathcal{K}$ and ${{S}_1,\dots,{S}_{i}\in \mathcal{S}}$ with ${W}_j\not\in {S}_j$ for all $1\leq j\leq i$ and ${{W}_i\not\in \cup_{j=1}^{i-1} ({W}_{j}\cup {S}_{j}) \cup ({W}\cup {S})}$, and ${C}_{1},\dots,{C}_{i}\in \mathcal{C}$ (and ${\mathbf{Y}}_1 = {\mathbf{Y}}^{[{S}_1,{C}_1]},\dots,{\mathbf{Y}}_{i} = {\mathbf{Y}}^{[{S}_i,{C}_i]}$), such that $$H({\mathbf{X}}_{{W}_{i}}\|{\mathbf{A}},{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1,{\mathbf{X}}_{{W}_1},\dots,{\mathbf{Y}}_{i-1},{\mathbf{X}}_{{W}_{i-1}},{\mathbf{Y}}_{i})=0.$$ Note that by construction, $$\left\|\cup_{j=1}^{i-1} ({W}_j \cup {S}_j) \cup ({W}\cup {S})\right\|\leq (M+1)i$$ for all $1\leq i\leq n-1$. Repeating an argument similar to the one being used for lower bounding $H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}})$ as in , it can be shown that $$\begin{aligned} & H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1,{\mathbf{X}}_{{W}_1},\dots,{\mathbf{Y}}_{i-1},{\mathbf{X}}_{{W}_{i-1}})\nonumber \\ & \quad \geq H({\mathbf{X}}_{{W}_{i}}) +H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1,{\mathbf{X}}_{{W}_1},\dots,{\mathbf{Y}}_{i},{\mathbf{X}}_{{W}_{i}})\end{aligned}$$ for all $1\leq i\leq n-1$. Combining these lower bounds for all $1\leq i\leq n-1$, we have $$\begin{aligned} H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}}) &\geq \sum_{i=1}^{n-1} H({\mathbf{X}}_{{W}_i})\nonumber \\ & = (n-1)L. \label{eq:CSIlineI22}\end{aligned}$$ Putting and together, we get $H({\mathbf{A}})\geq nL = \lceil\frac{K}{M+1}\rceil L$. Achievability {#subsec:AchThm3} ------------- In this section, we propose a scalar-linear PIR-CSI–I protocol for arbitrary ${1\leq M\leq K-1}$. This protocol, termed Modified Partition-and-Code (MPC), is inspired by our recently proposed Partition-and-Code with Interference Alignment protocol in [@HS2019] for private computation with uncoded side information. The MPC protocol does not make any assumption on the base-field size $q$ and the field-extension degree $l$, and is applicable for arbitrary $q\geq 2$ and $l\geq 1$. It should be noted that the Partition-and-Code protocol of [@Kadhe2017] is only applicable to the PIR-CSI–I setting when $M+1$ divides $K$. Otherwise, when $M+1$ is not a divisor of $K$, the Partition-and-Code protocol will generate one part of size less than $M+1$. This immediately results in a violation of the ${W}$-privacy condition. This is because the user’s demand cannot be any of the messages pertaining to this part, noting that (i) the support set of the user’s side information has size $M$, and (ii) all messages in the user’s side information support set need to be combined with the user’s demand. Modified Partition-and-Code (MPC) Protocol: This protocol consists of three steps as follows: Step 1: Let $n\triangleq\lceil \frac{K}{M+1}\rceil$. For $1\leq i\leq n-1$, we define $I_i \triangleq \{{(i-1)(M+1)+1},\dots,i(M+1)\}$, and $I_n\triangleq \{(n-1)(M+1)+1,\dots,K,1,\dots,n(M+1)-K\}$. (Note that $I_n = \{(n-1)(M+1)+1,\dots,K\}$ when $M+1$ divides $K$.) First, the user constructs a random permutation $\pi$ on $\mathcal{K} = \{1,\dots,K\}$ as follows. The user randomly chooses an index $j^{}$ from $\mathcal{K}$, and assigns the demand index ${W}$ to $\pi(j^{})$, i.e., $\pi(j^{}) = {W}$. Let $i^{}\triangleq \lceil \frac{j^{}}{M+1}\rceil$ be the smallest index $i\in \{1,\dots,n\}$ such that $j^{}\in I_i$. Then, the user randomly assigns the side information support indices in ${S}$ to $\{\pi(j): j\in I_{i^{}}\setminus \{j^{}\}\}$ and randomly assigns the (not-yet-assigned) indices in $\mathcal{K}\setminus ({W}\cup {S})$ to $\{\pi(j): j\in \mathcal{K}\setminus I_{i^{}}\}$. Next, the user constructs $n$ (ordered) sets $U_1,\dots,U_n$, each of size $M+1$, defined as $U_i = \{\pi(j): j\in I_i\}$; and constructs an (ordered) multiset $V$, defined as $V = \{c_{\pi(j)}: j\in I_{i^{}}\}$ where $c_{\pi(j)}$ for $j\in I_{i^{}}\setminus \{j^{}\}$ is the coefficient of message $X_{\pi(j)}$ in the side information $Y^{[{S},{C}]}$, and $c_{\pi(j^{})} = c_{{W}}$ is a randomly chosen element from $\mathbb{F}^{\times}_q$. The user then constructs $Q_i = (U_i,V)$ for each ${1\leq i\leq n}$, and sends to the server the query $Q^{[{W},{S},{C}]} = \{Q_{1},\dots,Q_{n}\}$. Step 2:* By using $Q_i=(U_i,V)$’s, the server computes $A_i$’s, defined as $A_{i} = \sum_{j=1}^{M+1} c_{i_j} X_{i_j}$ where $U_{i} = \{i_1,\dots,i_{M+1}\}$ and $V = \{c_{i_1},\dots,c_{i_{M+1}}\}$, and sends back to the user the answer $A^{[{W},{S},{C}]}=\{A_{1},\dots,A_{n}\}$. *Step 3:* Upon receiving the answer from the server, the user retrieves $X_{{W}}$ by subtracting off the contribution of the side information $Y^{[{S},{C}]}$ from $A_{i^{}}=c_{{W}}X_{{W}}+\sum_{i\in {S}} c_{i}X_{i}$. Consider a scenario where the server has ${K=5}$ messages $X_1,\dots,X_{5}\in \mathbb{F}_{3}$, and the user demands the message $X_1$ and has a coded side information $Y = X_2+2X_3$ with support size $M=2$. For this example, ${W} = 1$, ${S}= \{2,3\}$, and ${{C} =\{c_2,c_3\}= \{1,2\}}$. The parameters of the MPC protocol for this example are as follows: $n = \lceil\frac{K}{M+1}\rceil=2$, $I_1 = \{1,2,3\}$, and $I_2 = \{4,5,1\}$. First, the user constructs a permutation $\pi$ of $\{1,\dots,5\}$ as follows. The user randomly chooses an index $j^{}$ from $\{1,\dots,5\}$, say $4$, and assigns the index ${W} = 1$ to $\pi(j^{}) = \pi(4)$, i.e., $\pi(4) = 1$. Note that, in this case, $i^{}\triangleq \lceil \frac{j^{}}{M+1}\rceil = 2$, and $I_{i^{}} = I_2 = \{4,5,1\}$. The user then randomly assigns the indices in ${S}$, i.e., $2$ and $3$, to $\{\pi(j): {j\in I_{i^{}}\setminus \{j^{}}\}\} = \{\pi(5),\pi(1)\}$, say $\pi(5)=3$ and $\pi(1)=2$; and randomly assigns the (not-yet-assigned) indices $4$ and $5$ to $\{\pi(j): j\in \{1,\dots,5\}\setminus I_{i^{}}\} = \{\pi(2),\pi(3)\}$, say $\pi(2)=4$ and $\pi(3)=5$. Thus, the permutation $\pi$ maps $\{1,2,3,4,5\}$ to $\{2,4,5,1,3\}$. Next, the user constructs $n=2$ (ordered) sets $U_1,U_2$, each of size $M+1=3$, defined as $U_1 = \{\pi(j): j\in I_1\} = \{2,4,5\}$ and $U_2 = \{\pi(j): j\in I_2\} = \{1,3,2\}$; and constructs an (ordered) multiset $V$, defined as $V= \{c_{\pi(j)}: j\in I_{2}\} = \{c_1,c_3,c_2\}$ where $c_2=1$ and $c_3=2$ are the coefficients of $X_2$ and $X_3$ in the side information $Y$, and $c_1$ is a randomly chosen element from $\mathbb{F}^{\times}_3 = \{1,2\}$, say $c_1=2$. Thus, $V = \{2,2,1\}$. The user constructs $Q_1 = (U_1,V) = (\{2,4,5\},\{2,2,1\})$ and $Q_2 = (U_2,V) = (\{1,3,2\},\{2,2,1\})$, and sends the query $Q = \{Q_1,Q_2\}$ to the server. The server then computes $A_1 = 2X_2+2X_4+X_5$ and $A_2 = 2X_1+2X_3+X_2$, and sends the answer $A = \{A_1,A_2\}$ back to the user. Then, the user subtracts off the contribution of $Y = X_2+2X_3$ from $A_{i^{}}=A_2 = 2X_2+X_2+2X_3$, and recovers $X_1$. For this example, the rate of the MPC protocol is $1/2$. Note that the rate of the Specialized GRS Code protocol—which achieves $({W},{S})$-privacy and hence ${W}$-privacy, for the scenario of this example is $(K-M)^{-1} = 1/3$. From the perspective of the server, who knows the model and the parameters as well as the protocol, the messages $X_1,\dots,X_5$ are equally likely to be the user’s demand. This is because, given the query, for each candidate demand, the server finds a unique potential side information. In particular, by the protocol, there must exist a linear combination $A_i$ in the answer $A = \{A_1,\dots,A_n\}$ (i.e., $\{A_1,A_2\}$ in this example) which is a function of the demand and the side information, and not a function of any other message. For example, given that the candidate demand is $X_1$, the server finds $X_2+2X_3$ as the only potential side information, noting that only $A_2 = 2X_1+X_2+2X_3$ is a linear combination of $X_1$ and $M=2$ other messages (i.e., $X_2$ and $X_3$). As an another example, consider the message $X_2$. Given that the candidate demand is $X_2$, there exist two linear combinations $A_1$ and $A_2$, each of which is a function of $X_2$ and $M=2$ other messages. However, by the protocol, among all linear combinations $A_i$ that are functions of the candidate demand and $M$ other messages, only the linear combination $A_i$ with the smallest index $i$ is a function of the demand and the side information. Thus, for the candidate demand $X_2$, the server finds $2X_4+X_5$ as the only potential side information, noting that among $A_1$ and $A_2$—which are both functions of $X_2$ and $M=2$ other messages, the linear combination $A_1 = 2X_2+2X_4+X_5$ has the smallest index. Similarly, for each of the other candidate demands $X_3,X_4,X_5$, the server finds a unique potential side information. Moreover, the side information support index set is uniformly distributed and the demand index is conditionally distributed uniformly given the side information support index set. Putting these arguments together, one can see that given the query each message is equally likely to be the user’s demand. This confirms that the MPC protocol satisfies the ${W}$-privacy condition for this example. It is worth noting that the existence of a unique potential side information for each candidate demand, which ensures ${W}$-privacy, results in the violation of the $({W},{S})$-privacy condition. For instance, in this example, given the query, for the candidate demand index $1$ the only potential side information support index set is $\{2,3\}$; and no other pair of indices in $\{2,\dots,5\}$ can be a potential side information support index set for the candidate demand index $1$. \[lem:Ach3\] The Modified Partition-and-Code (MPC) protocol is a scalar-linear PIR-CSI–I protocol, and achieves the rate ${\lceil \frac{K}{M+1} \rceil}^{-1}$. [[[* Proof.* ]{}]{}]{}See Appendix \[subsec:PLAch3\]. Proof of Theorem \[thm:PIRCSI-II\] {#sec:PIRCSI-II} ================================== Converse {#subsec:ConvThm4} -------- In this section, we give an information-theoretic proof of converse for the case of general PIR-CSI–II protocols, which also serves as a converse proof for the special case of scalar-linear PIR-CSI–II protocols. \[lem:Converse2\] For $M=2$ and $M=K$, the (scalar-linear) capacity of the PIR-CSI–II setting is upper bounded by $1$, and for any $3\leq M\leq K-1$, the (scalar-linear) capacity of the PIR-CSI–II setting is upper bounded by $1/2$. [[[* Proof.* ]{}]{}]{}In the following, all entropies are conditional on the event $\mathds{1}_{\{{\mathbf{W}}\in {\mathbf{S}}\}}=1$, and for simplifying the notation, we remove this event from the conditions everywhere. Take arbitrary ${W},{S},{C}$ (and ${\mathbf{Y}}\triangleq {\mathbf{Y}}^{[{S},{C}]}$) such that ${W}\in {S}$. For the cases of $M=2$ and $M=K$, it suffices to show that $H({\mathbf{A}})\geq L$. Note that $H({\mathbf{A}})\geq H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}})=H({\mathbf{A}},{\mathbf{X}}_{{W}}\|{\mathbf{Q}},{\mathbf{Y}})$, where the equality follows from the recoverability condition, and $H({\mathbf{A}},{\mathbf{X}}_{{W}}\|{\mathbf{Q}},{\mathbf{Y}})= H({\mathbf{X}}_{{W}}\|{\mathbf{Q}},{\mathbf{Y}})+H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}})\geq H({\mathbf{X}}_{{W}})$, where the inequality follows from the independence of ${\mathbf{X}}_{{W}}$ and $({\mathbf{Q}},{\mathbf{Y}})$ by assumption. Putting these arguments together, $H({\mathbf{A}})\geq H({\mathbf{X}}_{{W}})=L$. For the cases of $3\leq M \leq K-1$, we need to show that $H({\mathbf{A}})\geq 2L$. By the above arguments, we have $$\label{eq:CSIlineII23} H({\mathbf{A}}) \geq H({\mathbf{X}}_{{W}})+H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}}).$$ Consider an arbitrary index ${W}_1\in {S}$. By the result of Lemma \[prop:2\], there exist ${S}_1\in \mathcal{S}$ with ${W}_1\in {S}_1$ and ${C}_1\in \mathcal{C}$ (and ${\mathbf{Y}}_1 = {\mathbf{Y}}^{[{S}_1,{C}_1]}$) such that $H({\mathbf{X}}_{{W}_1}\|{\mathbf{A}},{\mathbf{Q}},{\mathbf{Y}}_1) = 0$. Since conditioning does not increase the entropy, then $H({\mathbf{X}}_{{W}_1}\|{\mathbf{A}},{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1) = 0$. Then, we can write $$\begin{aligned} H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}}) & \geq H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1) \nonumber \\ & = H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1) \nonumber \\ & \quad + H({\mathbf{X}}_{{W}_1}\|{\mathbf{A}},{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1) \nonumber\\ &=H({\mathbf{A}},{\mathbf{X}}_{{W}_1}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1) \nonumber\\ &=H({\mathbf{X}}_{{\mathbf{W}}_1}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1)\nonumber \\&\quad + H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1,{\mathbf{X}}_{{W}_1})\nonumber\\ & \geq H({\mathbf{X}}_{{\mathbf{W}}_1}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1). \label{eq:CSIlineII24}\end{aligned}$$ Noting that ${\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1,{\mathbf{X}}_{{W}_1}$ are linear functions of the messages in ${\mathbf{X}}_{\mathcal{K}}$, and ${\mathbf{Q}}$ is independent of ${\mathbf{X}}_{\mathcal{K}}$, there are two possible cases: (i) $H({\mathbf{X}}_{{W}_1}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1)=H({\mathbf{X}}_{{W}_1})$, i.e., ${\mathbf{X}}_{{W}_1}$ is independent of $({\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1)$, or (ii) ${H({\mathbf{X}}_{{W}_1}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1)=0}$, i.e., ${\mathbf{X}}_{{W}_1}$ can be recovered from ${\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1$. In the case (i), $H({\mathbf{X}}_{{W}_1}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1)=H({\mathbf{X}}_{{W}_1})$ by assumption. Rewriting , we have $$\begin{aligned} H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}}) &\geq H({\mathbf{X}}_{{W}_1}).\label{eq:CSIlineII25} \end{aligned}$$ By and , $H({\mathbf{A}})\geq H({\mathbf{X}}_{{W}})+H({\mathbf{X}}_{{W}_1})=2L$. In the case (ii), by assumption, $H({\mathbf{X}}_{{W}_1}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1)=0$. Again, by the linearity of ${\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_1,{\mathbf{X}}_{{W}_1}$, it must hold that ${\mathbf{Y}}=c_{{W}} {\mathbf{X}}_{{W}}+c_{{W}_1} {\mathbf{X}}_{{W}_1}+{\mathbf{Z}}$ and ${\mathbf{Y}}_1=c'_{{W}} {\mathbf{X}}_{{W}} + c'_{{W}_1} {\mathbf{X}}_{{W}_1}+c' {\mathbf{Z}}$ for some $c'_{{W}},c'_{{W}_1},c'\in \mathbb{F}^{\times}_q$, where ${\mathbf{Z}} = \sum_{i\in {S}\setminus \{{W},{W}_1\}} c_{i} {\mathbf{X}}_i$. Unlike the previous case, this time we turn to an arbitrary index ${W}_2\not\in {S}$. Again, by the result of Lemma \[prop:2\], there exist ${S}_2\in \mathcal{S}$ with ${W}_2\in {S}_2$ and ${C}_2\in \mathcal{C}$ (and ${\mathbf{Y}}_2 = {\mathbf{Y}}^{[{S}_2,{C}_2]}$) such that $H({\mathbf{X}}_{{W}_2}\|{\mathbf{A}},{\mathbf{Q}},{\mathbf{Y}}_2) = 0$. Similar to , it can be shown that $$\begin{aligned} H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}}) & \geq H({\mathbf{X}}_{{W}_2}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_2)\nonumber \\&\quad + H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_2,{\mathbf{X}}_{{W}_2}). \label{eq:CSIlineII26}\end{aligned}$$ If ${\mathbf{X}}_{{W}_2}$ is independent of $({\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_2)$, similarly as in the case (i) we can show that $H({\mathbf{A}})\geq H({\mathbf{X}}_{{W}})+H({\mathbf{X}}_{{W}_2})=2L$. If ${\mathbf{X}}_{{W}_2}$ is recoverable from $({\mathbf{Q}},{\mathbf{Y}},{\mathbf{X}}_{{W}},{\mathbf{Y}}_2)$, it must hold that ${\mathbf{Y}}_2=c''_{{W}_2} {\mathbf{X}}_{{W}_2}+c'' (c_{{W}_1} {\mathbf{X}}_{{W}_1}+{\mathbf{Z}})$ for some $c''_{{W}_2},c''\in \mathbb{F}^{\times}_q$. Note that ${\mathbf{X}}_{{W}_2}$ is independent of $({\mathbf{Q}},{\mathbf{Y}}_1,{\mathbf{X}}_{{W}_1},{\mathbf{Y}}_2)$ since by construction, ${\mathbf{X}}_{{W}_2}$ cannot be recovered from $c'_{{W}}{\mathbf{X}}_{{W}}+c'{\mathbf{Z}}$ and $c''_{{W}_2}{\mathbf{X}}_{{W}_2}+c''{\mathbf{Z}}$, or in turn, from ${\mathbf{Y}}_1$ and ${\mathbf{Y}}_2$ given ${\mathbf{X}}_{{W}_1}$. Also, ${\mathbf{X}}_{{W}_1}$ is independent of $({\mathbf{Q}},{\mathbf{Y}}_1,{\mathbf{Y}}_2)$ since, again by construction, ${\mathbf{X}}_{{W}_1}$ cannot be recovered from ${\mathbf{Y}}_1$ and ${\mathbf{Y}}_2$. Thus, we can write $$\begin{aligned} H({\mathbf{A}}) &\geq H({\mathbf{A}}\|{\mathbf{Q}},{\mathbf{Y}}_1,{\mathbf{Y}}_2)\nonumber \\ &=H({\mathbf{A}},{\mathbf{X}}_{{W}_1},{\mathbf{X}}_{{W}_2}\|{\mathbf{Q}},{\mathbf{Y}}_1,{\mathbf{Y}}_2) \label{eq:CSIlineII27}\\ & \geq H({\mathbf{X}}_{{W}_1}\|{\mathbf{Q}},{\mathbf{Y}}_1,{\mathbf{Y}}_2)\nonumber\\ &\quad +H({\mathbf{X}}_{{W}_2}\|{\mathbf{Q}},{\mathbf{X}}_{{W}_1},{\mathbf{Y}}_1,{\mathbf{Y}}_2)\nonumber\\ & = H({\mathbf{X}}_{{W}_1})+H({\mathbf{X}}_{{W}_2})\label{eq:CSIlineII28}\end{aligned}$$ where holds because $H({\mathbf{X}}_{{W}_1}\|{\mathbf{A}},{\mathbf{Q}},{\mathbf{Y}}_1,{\mathbf{Y}}_2)=0$ and $H({\mathbf{X}}_{{W}_2}\|{\mathbf{A}},{\mathbf{Q}},{\mathbf{Y}}_1,{\mathbf{X}}_{{W}_1},{\mathbf{Y}}_2)=0$, noting that by assumption, ${H({\mathbf{X}}_{{W}_1}\|{\mathbf{A}},{\mathbf{Q}},{\mathbf{Y}}_1)=0}$ and $H({\mathbf{X}}_{{W}_2}\|{\mathbf{A}},{\mathbf{Q}},{\mathbf{Y}}_2)=0$; and holds since as was shown earlier, ${\mathbf{X}}_{{W}_1}$ and ${\mathbf{X}}_{{W}_2}$ are independent of $({\mathbf{Q}},{\mathbf{Y}}_1,{\mathbf{Y}}_2)$ and $({\mathbf{Q}},{\mathbf{Y}}_1,{\mathbf{X}}_{{W}_1},{\mathbf{Y}}_2)$, respectively. By , we get $H({\mathbf{A}})\geq 2L$. Achievability {#subsec:AchThm4} ------------- In this section, we propose a scalar-linear PIR-CSI–II protocol for each of the following cases: (Case 1) $M=2$; (Case 2) $3\leq M\leq \frac{K}{2}+1$; (Case 3) $\frac{K+1}{2}\leq M\leq K-1$; and (Case 4) $M=K$. (Note that Cases 2 and 3 are overlapping at $M=\frac{K}{2}+1$ or $M=\frac{K+1}{2}$ when $K$ is even or odd, respectively. In these scenarios, either of the proposed protocols for Cases 2 and 3 applies.) It should be noted that the proposed protocols for Cases 1 and 2 are applicable for any base-field size $q\geq 2$ and any field-extension degree $l\geq 1$; whereas the proposed protocols for Cases 3 and 4 are applicable for any $q\geq 3$ and any $l\geq 1$. The proposed protocols are based on the idea of randomizing the structure of query/answer, and are referred to as the Randomized Selection-and-Code (RSC) protocols. In particular, in these protocols, for any given instance of the problem, there exist multiple different query/answer structures, each of which satisfies the recoverability condition; and one of these structures will be chosen at random according to a probability distribution, which is carefully designed to ensure the ${W}$-privacy condition. For example, consider a scenario of Case 1 in which the server stores $X_1,X_2,X_3,\dots,X_K$, and the user’s demand and side information are $X_1$ and $X_1+X_2$, respectively. The RSC protocol for Case 1 has two different (query/answer) structures: (i) the user queries $X_1$, which is the user’s demand, and the server sends $X_1$ back to the user; or (ii) the user queries $X_2$, which is the other message in the user’s side information, and the server sends back $X_2$ to the user. (Note that neither of these structures depend on the other messages $X_3,\dots,X_K$.) The RSC protocol for Case 1 randomly generates one of the two structures (i) and (ii), according to a probability distribution—specified shortly in the description of the protocol, designed in order to guarantee ${W}$-privacy (i.e., given the query, each message in $X_1,\dots,X_K$ is equally likely to be the user’s demand.) Note that using either of the two structures (i) and (ii), the user can recover $X_1$. The RSC protocols for Cases 2-4 use a similar idea. For the ease of exposition, w.l.o.g., we assume that ${{W}=\{1\}}$, ${{S} = \{1,\dots,M\}}$, and ${{C} = \{c_{1},\dots,c_{M}\}}$. Randomized Selection-and-Code (RSC) Protocols: The RSC protocol for each case consists of three steps, where the steps 2-3 are the same as Steps 2-3 in the MPC protocol (Section \[subsec:AchThm3\]). The step 1 of the RSC protocols are as follows: Case 1: The user randomly selects the index ${W}$ (i.e., $1$) with probability $\frac{1}{K}$, or the other index in ${S}$ (i.e., $2$) with probability $\frac{K-1}{K}$, and constructs two sets $U = \{i\}$ and $V=\{1\}$, where $i$ is the selected index by the user. The user then constructs $Q = (U,V)$, and sends the query $Q^{[{W},{S},{C}]} = Q$ to the server. Consider a scenario where the server has ${K=6}$ messages $X_1,\dots,X_{6}\in \mathbb{F}_{3}$, and the user demands the message $X_1$ and has a coded side information $Y = 2X_1+X_2$ with support size $M=2$. For this example, ${W} = 1$, ${S}= \{1,2\}$, and ${{C} =\{c_1,c_2\}= \{2,1\}}$. The user randomly selects an index $i$ from ${S} = \{1,2\}$, where the probability of selecting the index $i=1$ is $\frac{1}{K} = \frac{1}{6}$, and the probability of selecting the index $i=2$ is $\frac{K-1}{K} = \frac{5}{6}$. Suppose that the user selects the index $i=2$. Then, the user requests the server for the message $X_i = X_2$, and the server responds by sending $X_2$ to the user. Subtracting off $X_2$ from $Y=2X_1+X_2$, the user then recovers $X_1$. For this example, the rate of the RSC protocol is $1$. Note that the Modified Specialized GRS Code protocol—which yields $({W},{S})$-privacy and hence ${W}$-privacy, achieves the rate ${(K-M+1)^{-1}} = 1/5$ for the scenario of this example. From the server’s perspective, the probability that the message $X_2$ is the user’s demand is $\frac{1}{6}$, and the probability that one of the messages $X_1,X_3,\dots,X_6$ is the user’s demand is $\frac{5}{6}$. Since these messages are equally likely to be the demand, the probability of any of them to be the user’s demand is $\frac{5}{6}\times \frac{1}{5} = \frac{1}{6}$. This guarantees the ${W}$-privacy. Now, suppose that the user selects $i=1$. In this case, the user requests their demand $X_1$ from the server, and the server responds by sending $X_1$ back to the user. Again, from the perspective of the server, the probability that the message $X_1$ is the user’s demand is $\frac{1}{6}$, and the probability of any of the messages $X_2,\dots,X_6$ to be the user’s demand is $\frac{5}{6}\times\frac{1}{5} = \frac{1}{6}$. This again ensures the ${W}$-privacy. Case 2: The user constructs two (ordered) sets $U_1,U_2$, each of size $M-1$, with elements from the indices in $\mathcal{K}$, and an (ordered) multiset $V$ of size $M-1$ with elements from $\mathbb{F}^{\times}_q$. The constructions of $U_1,U_2,V$ are as follows. First, the user chooses an integer $r\in \{M-2,M-1\}$ by sampling from a probability distribution given by $$\mathbb{P}({\mathbf{r}}=r) = \begin{cases} \frac{2M-2}{K}, & r = M-2,\\ 1-\frac{2M-2}{K}, & r = M-1. \end{cases}$$ If $r=M-1$ is chosen, the user randomly selects $M-1$ indices from ${\mathcal{K}\setminus {S}}$; otherwise, if $r=M-2$ is chosen, the user selects the index ${W}$ along with $M-2$ randomly chosen indices from $\mathcal{K}\setminus {S}$. Denote by $\{i_1,\dots,i_{M-1}\}$ the (ordered) set of the $M-1$ selected indices (in increasing order). Then, the user constructs $U_1 = \{2,\dots,M\}$ (i.e., the set of elements in ${S}\setminus W$ in increasing order) and $U_2=\{i_1,\dots,i_{M-1}\}$. Next, the user constructs the (ordered) multiset $V=\{c_{2},\dots,c_{M}\}$ (i.e., the sequence of elements in ${C}$ excluding the element $c_{{W}}$). The user then constructs $Q_{i} = (U_{i},V)$ for each $i\in \{1,2\}$, and for a randomly chosen permutation $\sigma: \{1,2\}\mapsto \{1,2\}$, sends the query $Q^{[{W},{S},{C}]} = \{Q_{\sigma(1)},Q_{\sigma(2)}\}$ to the server. Consider a scenario where the server has ${K=6}$ messages $X_1,\dots,X_{6}\in \mathbb{F}_{3}$, and the user demands the message $X_1$ and has a coded side information $Y = 2X_1+X_2+2X_3$ with support size $M=3$. For this example, ${W} = 1$, ${S}= \{1,2,3\}$, and ${{C} =\{c_1,c_2,c_3\}= \{2,1,2\}}$. First, the user randomly chooses an integer $r\in \{M-2 = 1,M-1 = 2\}$, where the probability of choosing $r=1$ is $\frac{2}{3}$, and the probability of choosing $r=2$ is $\frac{1}{3}$. Suppose that the user chooses $r=1$. The user then selects the index ${W}=1$ along with $r=1$ randomly chosen index from $\{1,\dots,6\}\setminus\{1,2,3\} = \{4,5,6\}$, say the index $4$. Then, the user constructs two (ordered) sets $U_1=\{2,3\}$ and $U_2 = \{1,4\}$, and the (ordered) multiset $V=\{c_2,c_3\}=\{1,2\}$. Then, the user constructs $Q_1 = (U_1,V) = (\{2,3\},\{1,2\})$ and $Q_2 = (U_2,V) = (\{1,4\},\{1,2\})$. For a randomly chosen permutation $\sigma$ on $\{1,2\}$, say $\sigma(1)=2$ and ${\sigma(2)=1}$, the user constructs the query $Q = \{Q_{\sigma(1)},Q_{\sigma(2)}\} = \{Q_2,Q_1\}$, and sends it to the server. The server computes $A_{i} = \sum_{j=1}^{M-1} c_{i_j} X_{i_j}$ for each $i\in \{1,2\}$ where $Q_i = (\{i_1,i_2\},\{c_{i_1},c_{i_2}\})$. For this example, $A_1 = X_2+2X_3$ and $A_2=X_1+2X_4$. Then, the server sends the answer $A=\{A_{\sigma(1)},A_{\sigma(2)}\}=\{A_2,A_1\}$ back to the user. Subtracting off $A_1$ from $Y=2X_1+X_2+2X_3$, the user recovers $X_1$. For this example, the rate of the RSC protocol is $1/2$; whereas the rate of the Modified Specialized GRS Code protocol for the scenario of this example is $(K-M+1)^{-1} = 1/4$. From the server’s perspective, $U_1=\{2,3\}$ and $U_2=\{1,4\}$ are equally likely to be the index set of the user’s side information support set (excluding the demand index). Let us refer to the event that $X_2$ and $X_3$ (or $X_1$ and $X_4$) are the two messages in the user’s side information support set as E1 (or E2). Then, E1 (or E2) has probability $\frac{1}{2}$. Note also that, given E1 (or E2), $X_2$ and $X_3$ (or $X_1$ and $X_4$) have zero probability to be the user’s demand. Given E1, (i) with probability $\frac{1}{3}$, the user’s demand is neither $X_1$ nor $X_4$, or (ii) with probability $\frac{2}{3}$, the user’s demand is either $X_1$ or $X_4$. Given E1-(i), $X_5$ and $X_6$ are equally likely to be the user’s demand. That is, given E1, $X_5$ (or $X_6$) is the user’s demand with probability $\frac{1}{3}\times \frac{1}{2} = \frac{1}{6}$. Given E1-(ii), $X_1$ and $X_4$ are equally likely to be the user’s demand. Then, given E1, $X_1$ (or $X_4$) is the user’s demand with probability $\frac{2}{3}\times\frac{1}{2}=\frac{1}{3}$. Given E2, (i) with probability $\frac{1}{3}$, the user’s demand is neither $X_2$ nor $X_3$, or (ii) with probability $\frac{2}{3}$, the user’s demand is either $X_2$ or $X_3$. Given E2-(i), either of $X_5$ and $X_6$ is the user’s demand with probability $\frac{1}{2}$. Then, given E2, $X_5$ (or $X_6$) is the user’s demand with probability $\frac{1}{3}\times \frac{1}{2} = \frac{1}{6}$. Given E2-(ii), either of $X_2$ and $X_3$ is the user’s demand with probability $\frac{1}{2}$. Then, given E1, $X_2$ (or $X_3$) is the user’s demand with probability $\frac{2}{3}\times\frac{1}{2}=\frac{1}{3}$. From the above arguments, it is easy to see that given the query, each message $X_i$ is equally likely to be the user’s demand, and hence the ${W}$-privacy condition is satisfied. For example, $X_1$ has probability $\frac{1}{3}$ (or $0$) to be the user’s demand given E1 (or E2). Since E1 and E2 each have probability $\frac{1}{2}$, the probability of $X_1$ to be the user’s demand is $\frac{1}{2}\times\frac{1}{3}+\frac{1}{2}\times 0 = \frac{1}{6}$. As an another example, consider $X_5$. Given either of E1 or E2, $X_5$ has probability $\frac{1}{6}$ to be the user’s demand. Thus, the probability of $X_5$ to be the user’s demand is $\frac{1}{2}\times\frac{1}{6}+\frac{1}{2}\times \frac{1}{6} = \frac{1}{6}$. Case 3: The user constructs two (ordered) sets $U_1,U_2$, each of size $M$, with elements from the indices in $\mathcal{K}$, and an (ordered) multiset $V$ of size $M$ with elements from $\mathbb{F}^{\times}_q$. The constructions of $U_1,U_2,V$ are as follows. The user chooses an integer ${s\in\{2M-K-1,2M-K\}}$ by sampling from a probability distribution given by $$\mathbb{P}({\mathbf{s}}=s) = \begin{cases} 1-\frac{2K-2M}{K}, & s = 2M-K-1,\\ \frac{2K-2M}{K}, & s = 2M-K. \end{cases}$$ If $s=2M-K$ is chosen, the user randomly selects $2M-K$ indices from ${S}\setminus {W}$; otherwise, if $s=2M-K-1$ is chosen, the user selects the index ${W}$ together with $2M-K-1$ randomly chosen indices from ${S}\setminus {W}$. Denote by $\{i_1,\dots,i_{M}\}$ the (ordered) set of the $2M-K$ selected indices and the $K-M$ indices in $\mathcal{K}\setminus {S}$ (in increasing order). Then, the user constructs ${U_1 = \{1,\dots,M\}}$ (i.e., the set of elements in ${S}$ in increasing order) and $U_2=\{i_1,\dots,i_M\}$. Next, the user constructs the (ordered) multiset $V=\{c,c_{2},\dots,c_{M}\}$ (i.e., the sequence of the elements in ${C}$, except when the element $c_{{W}}$ is replaced by the element $c$) where $c$ is randomly chosen from ${\mathbb{F}^{\times}_q\setminus \{c_{1}\}}$ (i.e., $\mathbb{F}^{\times}_q\setminus \{c_{{W}}\}$). The user then constructs $Q_i = (U_i,V)$ for each $i\in \{1,2\}$, and for a randomly chosen permutation $\sigma: \{1,2\}\mapsto \{1,2\}$, sends the query $Q^{[{W},{S},{C}]} = \{Q_{\sigma(1)},Q_{\sigma(2)}\}$ to the server. Case 4: The user creates two (ordered) sets $U = \{1,\dots,K\}$ and $V=\{c,c_{2},\dots,c_{K}\}$ (i.e., the sequence of elements in ${C}$, except when the element $c_{{W}}$ is replaced by the element $c$) where $c$ is randomly chosen from $\mathbb{F}^{\times}_q\setminus \{c_{1}\}$ (i.e., $\mathbb{F}^{\times}_q\setminus \{c_{{W}}\}$). The user then constructs $Q = (U,V)$, and sends the query $Q^{[{W},{S},{C}]} = Q$ to the server. \[lem:Ach4\] The Randomized Selection-and-Code (RSC) protocols for $M=2$, $3\leq M\leq \frac{K}{2}+1$, ${\frac{K+1}{2}\leq M\leq K-1}$, and $M=K$ are scalar-linear PIR-CSI–II protocols, and achieve the rates $1$, $1/2$, $1/2$, and $1$, respectively. [[[* Proof.* ]{}]{}]{}See Appendix \[subsec:PLAch4\]. Conclusion and Future Work ========================== In this work, we studied the fundamental limits of single-message single-server information-theoretic PIR in the presence of a coded side information. Considering two different types of privacy, namely $({W},{S})$-privacy and ${W}$-privacy, we characterized the capacity and the scalar-linear capacity of the problem under two different models depending on whether the support set of the user’s coded side information includes the requested message or not. In addition, for each problem setting we proposed a novel scalar-linear scheme that achieves the capacity. One natural question that remains open is that how much the capacity will increase if we relax the assumption that the server knows the considered model, i.e., whether the side information is a function of the demand or not. Our preliminary results, beyond the scope of this work and hence not presented here, suggest that in an asymptotic regime (when the number of messages in the database grows unbounded), the capacity remains the same even if the server is aware of whether the side information depends on the demand or not. A detailed study of this observation remains open. Another direction for future work is to characterize the capacity of the single-server PIR when the user has multiple coded side information and/or wants multiple messages from the server. Our initial attempts at studying these settings suggest that there is a close relation between these problems and the problem of single-server private computation with coded side information, which is the focus of an ongoing work. Last but not least, the extensions of this work for multi-server setting were also recently studied in [@KKHS12019] and [@KKHS22019] for the cases in which ${W}$-privacy and $({W},{S})$-privacy are required, respectively. Some achievability schemes based on those in this work were proposed; notwithstanding, the capacity of these settings are still open in general. Proofs of Lemmas \[lem:Ach1\] and \[lem:Ach2\] {#sec:A1} ============================================== Proof of Lemma \[lem:Ach1\] {#subsec:PLAch1} --------------------------- Since the matrix $G$, defined in Step 1 of the Specialized GRS Code protocol, generates a $(K,K-M)$ GRS code which is an MDS code, the rows of $G$ are linearly independent. Accordingly, ${\mathbf{A}}_1,\dots,{\mathbf{A}}_{K-M}$, defined in Step 2, are linearly independent combinations of the messages in ${\mathbf{X}}_{\mathcal{K}}$, which are themselves independently and uniformly distributed over $\mathbb{F}_{q^{l}}$. This implies that ${\mathbf{A}}_1,\dots,{\mathbf{A}}_{K-M}$ are independently and uniformly distributed over $\mathbb{F}_{q^{l}}$. Since $H(\mathbf{X}_j)=L$ for all $j\in \mathcal{K}$, then $H(\mathbf{A}_i)=L$ for all $i\in \{1,\dots,K-M\}$. Thus, for all ${W}\in \mathcal{K},{S}\in \mathcal{S},{C}\in \mathcal{C}$ such that ${W}\not\in{S}$, we have $H(\mathbf{A}^{[{W},{S},{C}]}) = H(\mathbf{A}_1,\dots,\mathbf{A}_{K-M})=\sum_{i=1}^{K-M} H(\mathbf{A}_i)=(K-M)L$. (Note that $H(\mathbf{A}^{[{W},{S},{C}]})=(K-M)L$ does not depend on the realizations ${W},{S},{C}$.) Given that ${{\mathbf{W}}\not\in {\mathbf{S}}}$, ${\mathbf{W}}$ and ${\mathbf{S}}$ are jointly distributed uniformly, and ${\mathbf{C}}$ is distributed uniformly (and independently from $({\mathbf{W}},{\mathbf{S}})$). Thus, $H(\mathbf{A}^{[{\mathbf{W}},{\mathbf{S}},{\mathbf{C}}]}\|{\mathbf{W}}\not\in {\mathbf{S}})=H(\mathbf{A}^{[{W},{S},{C}]}) = (K-M)L$, implying that the rate of the Specialized GRS Code protocol is equal to ${L/((K-M)L)} = (K-M)^{-1}$. The scalar-linearity of ${\mathbf{A}}_i$’s in the messages ${\mathbf{X}}_{j}$’s confirms that the Specialized GRS Code protocol is scalar-linear. From the construction, it should also be obvious that the recoverability condition is satisfied. The proof of $({W},{S})$-privacy relies on two facts: (i) the $(K,K-M)$ GRS code, generated by the matrix $G$, is an MDS code, and hence the minimum (Hamming) weight of a codeword is $K-(K-M)+1 = M+1$; and (ii) there exist the same number of minimum-weight codewords for any support of size ${M+1}$ [@Roth:06]. The rest of the proof is as follows. From (i) and (ii), for any ${W}^{}\in \mathcal{K}, {S}^{}\in \mathcal{S}$ such that ${{W}^{}\not\in {S}^{}}$, the dual code, whose parity check matrix is $G$, contains the same number of parity check equations with support ${W}^{}\cup {S}^{}$ (i.e., the messages $\{X_i\}_{i\in {W}^{}\cup {S}^{}}$ have non-zero coefficients and the rest of the messages all have zero coefficients). For given ${W}^{},{S}^{}$, consider an arbitrary such parity check equation $Z = c_{{W}^{}}X_{{W}^{}}+\sum_{i\in {S}^{}} c_i X_i$ where $c_{i}\in \mathbb{F}^{\times}_q$ for all $i\in {W}^{}\cup {S}^{}$. The candidate demand $X_{{W}^{}}$ can be recovered from $Z$, only given a potential side information $\sum_{i\in {S}^{}} c (c_i X_i)$ for arbitrary $c\in \mathbb{F}^{\times}_q$. Noting that $\|\mathbb{F}^{\times}_q\| = q-1$, for any given parity check equation $Z$ with support ${W}^{}\cup {S}^{}$, there exist only $q-1$ potential side information, namely $\{c(Z-\sum_{i\in {S}^{}} c_i X_i): c\in \mathbb{F}^{\times}_q\}$, from each of which the candidate demand $X_{{W}^{*}}$ can be recovered. This proves the $({W},{S})$-privacy of the Specialized GRS Code protocol. Proof of Lemma \[lem:Ach2\] {#subsec:PLAch2} --------------------------- The proof, omitted to avoid repetition, follows from the same lines as in the proof of Lemma \[lem:Ach1\] (Appendix \[subsec:PLAch1\]). Proofs of Lemmas \[lem:Ach3\] and \[lem:Ach4\] {#sec:A2} ============================================== Proof of Lemma \[lem:Ach3\] {#subsec:PLAch3} --------------------------- By the construction of the Modified Partition-and-Code (MPC) protocol (see Steps 1-2), ${\mathbf{A}}_1,\dots,{\mathbf{A}}_n$ are linearly independent combinations of the messages in ${\mathbf{X}}_{\mathcal{K}}$. Using a similar argument as the one in the proof of Lemma \[lem:Ach1\] (Appendix \[subsec:PLAch1\]), it can be shown that $H({\mathbf{A}}^{[{W},{S},{C}]}) = H({\mathbf{A}}_1,\dots,{\mathbf{A}}_n) = nL$ for all ${W}\in \mathcal{K}, {S}\in \mathcal{S},{C}\in \mathcal{C}$ such that ${W}\not\in {S}$, and $H({\mathbf{A}}^{[{\mathbf{W}},{\mathbf{S}},{\mathbf{C}}]}\|{\mathbf{W}}\not\in {\mathbf{S}})= H({\mathbf{A}}^{[{W},{S},{C}]}) = nL$. This implies that the rate of the MPC protocol is equal to $L/nL = \lceil\frac{K}{M+1}\rceil^{-1}$. The scalar-linearity of the MPC protocol should be obvious from the construction. The recoverability condition is also obviously satisfied (see Step 3). To prove that the MPC protocol satisfies the ${W}$-privacy condition, we need to show that for any query $Q$ generated by the protocol, $${\mathbb{P}({\mathbf{W}}={W}\|\mathbf{Q}=Q,{\mathbf{W}}\not\in {\mathbf{S}})=\mathbb{P}({\mathbf{W}}={W}\|{\mathbf{W}}\not\in {\mathbf{S}})}$$ for all ${{W}\in \mathcal{K}}$, or in turn, $\mathbb{P}({\mathbf{W}}={W}\|\mathbf{Q}=Q,{\mathbf{W}}\not\in {\mathbf{S}})$ does not depend on ${W}$. (Note that by construction, ${\mathbf{Q}}$ is independent of the messages in ${\mathbf{X}}_{\mathcal{K}}$.) By Step 1 of the protocol, for any given ${W}\in \mathcal{K}$, there exist a unique ${S}_{{W}}\in \mathcal{S}$ (with ${W}\not\in {S}_{{W}}$) and a unique ${C}_{{W}}\in \mathcal{C}$ such that the triple $({W},{S}_{{W}},{C}_{{W}})$ complies with the query $Q$, i.e., given that $X_{{W}}$ and $Y^{[{S}_{{W}},{C}_{{W}}]}$ are the user’s demand and side information, respectively, the protocol could potentially generate the query $Q$. Then, we have $$\begin{aligned} &\mathbb{P}({\mathbf{W}}={W}\|\mathbf{Q}=Q, {\mathbf{W}}\not\in {\mathbf{S}}) \\ & = {\mathbb{P}({\mathbf{W}}={W},{\mathbf{S}}={S}_{{W}},{\mathbf{C}}={C}_{{W}}\|\mathbf{Q}=Q,{\mathbf{W}}\not\in {\mathbf{S}})}.\end{aligned}$$ Since the conditional distribution of $({\mathbf{W}},{\mathbf{S}},{\mathbf{C}})$ given ${\mathbf{W}}\not\in {\mathbf{S}}$ is uniform, by applying the Bayes’ rule one can see that ${\mathbb{P}({\mathbf{W}}={W},{\mathbf{S}}={S}_{{W}},{\mathbf{C}}={C}_{{W}}\|\mathbf{Q}=Q,{\mathbf{W}}\not\in {\mathbf{S}})}$ does not depend on ${W}$ so long as ${\mathbb{P}(\mathbf{Q}=Q\|{\mathbf{W}}={W},{\mathbf{S}}={S}_{{W}},{\mathbf{C}}={C}_{{W}})}$ does not depend on ${W}$. By the design of the protocol, $$\begin{aligned} & \mathbb{P}(\mathbf{Q}=Q\|{\mathbf{W}}={W},{\mathbf{S}}={S}_{{W}},{\mathbf{C}}={C}_{{W}}) \\ & =\frac{1}{K!}\binom{K-1}{M}(q-1)^{-1} \end{aligned}$$ for all ${W}\in \mathcal{K}$, and hence $\mathbb{P}({\mathbf{W}}={W}\|\mathbf{Q} = Q,{\mathbf{W}}\not\in {\mathbf{S}})$ does not depend on ${W}$. Proof of Lemma \[lem:Ach4\] {#subsec:PLAch4} --------------------------- The proofs for the rates of the RSC protocols follow the same line as in the proof of the rate of the MPC protocol in Lemma \[lem:Ach3\] (Appendix \[subsec:PLAch3\]), and hence omitted. From the construction, it should also be obvious that the RSC protocols are scalar-linear. Moreover, it should not be hard to see from the description of these protocols that the recoverability condition is satisfied. To prove that the RSC protocols satisfy the ${W}$-privacy condition, we need to show that $${\mathbb{P}({\mathbf{W}}={W}\|{\mathbf{Q}}=Q,{\mathbf{W}}\in {\mathbf{S}})}={\mathbb{P}({\mathbf{W}}={W}\|{\mathbf{W}}\in {\mathbf{S}})}$$ for all ${W}\in \mathcal{K}$. Alternatively, by the Bayes’ rule, it suffices to show that ${\mathbb{P}({\mathbf{Q}}=Q\|{\mathbf{W}}={W},{\mathbf{W}}\in{\mathbf{S}})}$ does not depend on ${W}$. Recall that $Q = (U,V)$ for Cases 1 and 4, and $Q = \{Q_1,Q_2\} = \{(U_1,V),(U_2,V)\}$ for Cases 2 and 3. For simplifying the notation, let us denote $\{U_1,U_2\}$ by $U$ for Cases 2 and 3. By the construction of the RSC protocols and the model assumptions, given ${\mathbf{W}}\in {\mathbf{S}}$, the following two observations hold: (i) ${\mathbf{U}}$ and ${\mathbf{V}}$ are conditionally independent given ${\mathbf{W}}$, and (ii) ${\mathbf{V}}$ and ${\mathbf{W}}$ are independent. The observation (i) should be obvious, and the observation (ii) holds because ${\mathbf{V}}$ is uniformly distributed over all possible choices of $V$ for each case. (For example, for Case 1, ${\mathbf{V}}=\{1\}$; and for Case 2, ${\mathbf{V}}=\{{\mathbf{c}}_i: i\in {\mathbf{S}}\setminus {\mathbf{W}}\}$—where ${\mathbf{c}}_i$’s are uniformly distributed over $\mathbb{F}^{\times}_q$, is uniformly distributed over all ordered multisets of size $M-1$ with elements from $\mathbb{F}^{\times}_q$.) Using (i) and (ii), $$\begin{aligned} & \mathbb{P}({\mathbf{Q}}=Q\|{\mathbf{W}}={W},{\mathbf{W}}\in{\mathbf{S}}) \\ & = \mathbb{P}({\mathbf{U}}=U, {\mathbf{V}}=V\|{\mathbf{W}}={W},{\mathbf{W}}\in{\mathbf{S}})\\ & = \mathbb{P}({\mathbf{V}}=V\|{\mathbf{W}}\in{\mathbf{S}})\times \mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W},{\mathbf{W}}\in{\mathbf{S}}).\end{aligned}$$ Since $\mathbb{P}({\mathbf{V}}=V\|{\mathbf{W}}\in{\mathbf{S}})$ does not depend on ${W}$, instead of showing that ${\mathbb{P}({\mathbf{Q}}=Q\|{\mathbf{W}}={W},{\mathbf{W}}\in{\mathbf{S}})}$ is not a function of ${W}$, it suffices to show that ${\mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W},{\mathbf{W}}\in{\mathbf{S}})}$ does not depend on ${W}$. In the following, we prove this claim for the RSC protocol for each case separately. With a slight abuse of notation, hereafter for the ease of exposition, we treat the ordered sets $U_1,U_2$ as (unordered) sets.\ ### Case 1 {#case1 .unnumbered} For an arbitrary $i\in \mathcal{K}$, consider $U=\{i\}$. Take an arbitrary ${W}\in \mathcal{K}$. There are two cases as follows: (i) ${W}=i$, and (ii) ${W}\neq i$. In the case (i), we have $$\begin{aligned} & \mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}}) \nonumber\\ & =\sum_{j\in\mathcal{K}\setminus {W}}\mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W}, {\mathbf{S}}=\{{W},j\})\nonumber\\ & \quad\quad\quad\hspace{0.275cm} \times \mathbb{P}({\mathbf{S}}=\{{W},j\}\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}}).\label{eq:CSIlineII29}\end{aligned}$$ By the model assumption, we have $$\label{eq:CSIlineII30} \mathbb{P}({\mathbf{S}}=\{{W},j\}\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}})=\frac{1}{K-1}$$ for all $j\in \mathcal{K}\setminus {W}$. Moreover, given that ${\mathbf{W}}={W}$ and ${{\mathbf{S}}=\{{W},j\}}$, the protocol constructs $U = \{{W}\}$ with probability $\frac{1}{K}$. This implies that $$\label{eq:CSIlineII31} \mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W}, {\mathbf{S}}=\{{W},j\})=\frac{1}{K}$$ for all $j\in \mathcal{K}\setminus {W}$. Substituting and into , $$\label{eq:CSIlineII32} \mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}}) = \frac{1}{K}.$$ In the case (ii), we have $$\begin{aligned} & \mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}}) \nonumber\\ & =\mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W}, {\mathbf{S}}=\{{W},i\})\nonumber\\ &\quad\times \mathbb{P}({\mathbf{S}}=\{{W},i\}\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}})\nonumber\\ & = \frac{1}{K}, \label{eq:CSIlineII33}\end{aligned}$$ noting that by the model assumption, $$\mathbb{P}({\mathbf{S}}=\{{W},i\}\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}})=\frac{1}{K-1},$$ and by the design of the protocol, $${\mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W}, {\mathbf{S}}=\{{W},i\})}=\frac{K-1}{K}.$$ From and , we can conclude that ${\mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}})}$ does not depend on ${W}$.\ ### Case 2 {#case2 .unnumbered} Consider an arbitrary $U=\{U_1,U_2\}$. (Recall that $\|U_1\|=\|U_2\|=M-1$.) Take an arbitrary ${W}\in \mathcal{K}$. There are two cases as follows: (i) ${W}\in U_1\cup U_2$, and (ii) ${W}\not\in U_1\cup U_2$. In the case (i), w.l.o.g., assume that ${W}\in U_1$. Note that ${\mathbf{W}}={W}$ and ${W}\in U_1$ together imply that ${\mathbf{S}} = {W}\cup U_2$ (by the design of the protocol). Then, we have $$\begin{aligned} & \mathbb{P}({\mathbf{U}} =U\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}}) \nonumber\\ & =\mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W}, {\mathbf{S}}={W}\cup U_2)\nonumber\\ &\quad \times \mathbb{P}({\mathbf{S}}={W}\cup U_2\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}}).\label{eq:CSIlineII34}\end{aligned}$$ By the model assumption, we have $$\label{eq:CSIlineII35} \mathbb{P}({\mathbf{S}}={W}\cup U_2\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}})=\binom{K-1}{M-1}^{-1}.$$ Moreover, given that ${\mathbf{W}}={W}$ and ${\mathbf{S}} = {W}\cup U_2$, the protocol constructs $U_1$ with probability $(\frac{2M-2}{K})\times\binom{K-M}{M-2}^{-1}$, noting that ${W}\in U_1$. (The protocol selects the demand index ${W}$ to be one of the elements in $U_1$ with probability $\frac{2M-2}{K}$, and selects the set of other $M-2$ elements in $U_1$ from the set of $K-M$ indices in $\mathcal{K}\setminus {S}$ with probability $\binom{K-M}{M-2}^{-1}$.) This implies that $$\begin{aligned} & \mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W},{\mathbf{S}}={W}\cup U_2) \nonumber\\ & =2\left(\frac{M-1}{K}\right) \binom{K-M}{M-2}^{-1}.\label{eq:CSIlineII36} \end{aligned}$$ Substituting and into , $$\begin{aligned} & \mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}}) \nonumber\\ & = 2\left(\frac{M-1}{K}\right) \binom{K-M}{M-2}^{-1}\binom{K-1}{M-1}^{-1}.\label{eq:CSIlineII37} \end{aligned}$$ In the case (ii), we have $$\begin{aligned} & \mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}}) \nonumber \\ & =\mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W}, {\mathbf{S}}={W}\cup U_1)\nonumber\\ & \quad\quad\times \mathbb{P}({\mathbf{S}}={W}\cup U_1\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}})\nonumber\\ & \quad+\mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W}, {\mathbf{S}}={W}\cup U_2)\nonumber\\ & \quad\quad\times \mathbb{P}({\mathbf{S}}={W}\cup U_2\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}})\nonumber\\ & = 2\left(1-\frac{2M-2}{K}\right)\binom{K-M}{M-1}^{-1}\binom{K-1}{M-1}^{-1},\label{eq:CSIlineII38}\end{aligned}$$ noting that $$\begin{aligned} &\mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W},{\mathbf{S}}={W}\cup U_1)\\ &=\mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W},{\mathbf{S}}={W}\cup U_2)\\ & =\left(1-\frac{2M-2}{K}\right)\binom{K-M}{M-1}^{-1}, \end{aligned}$$ and $$\begin{aligned} & \mathbb{P}({\mathbf{S}}={W}\cup U_1\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}})\\ & =\mathbb{P}({\mathbf{S}}={W}\cup U_2\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}})\\ & = \binom{K-1}{M-1}^{-1} .\end{aligned}$$ Now, it is easy to verify that $$\begin{aligned} & \left(\frac{M-1}{K}\right)\binom{K-M}{M-2}^{-1} \\ & = \left(1-\frac{2M-2}{K}\right)\binom{K-M}{M-1}^{-1}. \end{aligned}$$ This shows that and are equal, completing the proof that $\mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}})$ does not depend on ${W}$.\ ### Case 3 {#case3 .unnumbered} Consider an arbitrary query $U=\{U_1,U_2\}$. (Recall that $\|U_1\|=\|U_2\|=M$.) Take an arbitrary ${W}\in \mathcal{K}$. There are two cases as follows: (i) ${W}\in U_1\cap U_2$, and (ii) ${W}\not\in U_1\cap U_2$. In the case (i), we have $$\begin{aligned} & \mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}})\nonumber \\ & =\mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W}, {\mathbf{S}}= U_1)\nonumber\\ &\quad\quad\times \mathbb{P}({\mathbf{S}}=U_1\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}})\nonumber\\ &\quad+\mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W}, {\mathbf{S}}=U_2)\nonumber\\ &\quad\quad\times \mathbb{P}({\mathbf{S}}=U_2\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}})\nonumber\\ & = 2\left(\frac{2M-K}{K}\right)\binom{M-1}{2M-K-1}^{-1}\binom{K-1}{M-1}^{-1}, \label{eq:CSIlineII39}\end{aligned}$$ noting that $$\begin{aligned} &\mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W},{\mathbf{S}}=U_1)\\ & =\mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W},{\mathbf{S}}=U_2)\\ & =\left(\frac{2M-K}{K}\right) \binom{M-1}{2M-K-1}^{-1}, \end{aligned}$$ and $$\begin{aligned} & \mathbb{P}({\mathbf{S}}=U_1\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}})\\ & =\mathbb{P}({\mathbf{S}}=U_2\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}})\\ & = \binom{K-1}{M-1}^{-1}. \end{aligned}$$ In the case (ii), w.l.o.g., assume that ${W}\in U_1$. Then, we have $$\begin{aligned} & \mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}}) \nonumber \\ &=\mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W}, {\mathbf{S}}=U_1)\nonumber\\ &\quad\times \mathbb{P}({\mathbf{S}}=U_1\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}})\nonumber\\ & = 2\left(\frac{K-M}{K}\right) \binom{M-1}{2M-K}^{-1}\binom{K-1}{M-1}^{-1},\label{eq:CSIlineII40}\end{aligned}$$ noting that $$\begin{aligned} & \mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W},{\mathbf{S}}=U_1) \\ & =2\left(\frac{K-M}{K}\right) \binom{M-1}{2M-K}^{-1}, \end{aligned}$$ and $$\begin{aligned} \mathbb{P}({\mathbf{S}}=U_1\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}}) = \binom{K-1}{M-1}^{-1}. \end{aligned}$$ It is easy to verify that $$\begin{aligned} & \left(\frac{2M-K}{K}\right)\binom{M-1}{2M-K-1}^{-1} \\ & = \left(\frac{K-M}{K}\right)\binom{M-1}{2M-K}^{-1}. \end{aligned}$$ This shows that and are equal, completing the proof that $\mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}})$ does not depend on ${W}$.\ ### Case 4 {#case4 .unnumbered} By the protocol, we have $U = \mathcal{K}$, and hence ${\mathbb{P}({\mathbf{U}}=U\|{\mathbf{W}}={W},{\mathbf{W}}\in {\mathbf{S}})=1}$ for all ${W}$. [^1]: This work was presented in part at the 2018 IEEE Information Theory Workshop, Guangzhou, China, November 2018, and the 2019 IEEE International Symposium on Information Theory, Paris, France, July 2019. [^2]: The authors are with the Department of Electrical and Computer Engineering, Texas A&M University, College Station, TX 77843 USA (E-mail: {anoosheh, fatemeh.kazemi, spalex}@tamu.edu). [^3]: This material is based upon work supported by the National Science Foundation under Grants No. 1718658 and 1642983. [^4]: In general, the query may also depend on the content of the side information—notwithstanding, in this work we focus on queries that are “universal” in the sense that any such query achieves privacy for all realizations of the messages. [^5]: The mutual information based definitions of the $(W,S)$-privacy and $W$-privacy conditions will be used in the converse proofs, whereas their probability based counterparts will be used in the achievability proofs. [^6]: Although our definitions of capacity and scalar-linear capacity are independent of the base-field size $q$ and the field-extension degree $l$, these quantities may depend on $q$ and $l$ in general. In this work, we show that the capacity and the scalar-linear capacity of the PIR-PCSI settings are achievable so long as $q\geq K$ and $l\geq 1$; and depending on the parameters $K,M$ and the model (I or II), the capacity and the scalar-linear capacity of the PIR-CSI settings are achievable so long as $q\geq 2$ or $q\geq 3$ and $l\geq 1$.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We experimentally investigate transport properties of a single junction between a superconductor and the edge of a two-dimensional electron system in a narrow $In_{0.75}Ga_{0.25}As$ quantum well with strong Rashba-type spin-orbit coupling. We experimentally demonstrate suppression of Andreev reflection within a narrow energy range within the superconducting gap. This suppression is shown to be independent of the superconductor material and the interface disorder. The effect is only present at lowest temperatures and magnetic fields, far from the corresponding critical values of a superconducting transition. We connect the observed suppression with a spin-Hall effect due to a strong spin-orbit coupling on the normal side of the junction.' author: - 'A. Kononov' - 'N. Titova' - 'G. Biasiol' - 'L. Sorba' - 'E.V. Deviatov' title: Suppression of Andreev reflection at energies well below the superconducting gap --- Introduction ============ Recent interest to transport investigations of the interface between a superconductor (S) and a low-dimensional semiconductor structure (N) is mostly stimulated by the search for Majorana fermions [@Wilczek]. The essential condition of this search [@reviews] is the realization of the topological superconductivity near the SN interface. This regime is characterized [@Fu; @Sau1; @Potter; @alicea] by spectrum modification in a low dimensional structure, caused by cooperation of a spin-orbit (SO) coupling, Zeeman energy $E_Z$, and the energy gap $\Delta_{ind}$ induced due to a proximity with s-wave superconductor. Because of the spectrum modification, it seems to be quite reasonable [@Nakosai] to study charge transport through a single SN interface, where the normal side is the edge of a two-dimensional electron gas (2DEG) with strong Rashba-type spin-orbit coupling. By contrast to Ref. , where a good correspondence with standard BTK theory [@tinkham] has been observed, we concentrate on a low-temperature limit, to extract features, specific to the strong SO coupling. The SO coupling induces an energy splitting $\Delta_{SO}$ which lifts the spin degeneracy, but the energy splitting does not break the time reversal symmetry unlike an exchange splitting in ferromagnet. Even in the simplest case of a two-dimensional metal, the physics is expected to be quite sophisticated [@inoue]. On the other hand, the problem is even more complicated because of a spin-Hall effect [@sinova] in a 2DEG with Rashba SO coupling. ![(Color online) Image of the central part of the sample with electrical connections depicted. The 100 $\mu$m wide corner-shape mesa has a number of leads to Ni-Au Ohmic contacts (contacts are not in the image area). Several superconducting (Nb or NbN) stripes (yellow, denoted by numbers) are placed to overlap the mesa step at two perpendicular mesa edges. In every overlap region, a junction is formed between the superconducting film and the 2DEG edge. The width of each junction is equal to 20 $\mu$m. The junctions are separated by 2$\mu$m distance. We study electron transport across one particular S-2DEG junction in a three-point configuration: the corresponding superconducting electrode is grounded (no. 6 in the figure), others are disconnected; a current is applied between it and one of the Ni-Au Ohmic contacts; another Ni-Au contact traces the 2DEG potential. \[sample\]](nbinas1.eps){width="0.9\columnwidth"} Here, we experimentally investigate transport properties of a single junction between a superconductor and the edge of a two-dimensional electron system in a narrow $In_{0.75}Ga_{0.25}As$ quantum well with strong Rashba-type spin-orbit coupling. We experimentally demonstrate suppression of Andreev reflection in a small energy range within the superconducting gap. This suppression is shown to be independent of the superconductor material and the interface disorder. The effect is only present at lowest temperatures and magnetic fields, far from the corresponding critical values of a superconducting transition. We connect the observed suppression with a spin-Hall effect due to a strong spin-orbit coupling on the normal side of the junction. Samples and technique ===================== Our samples are grown by solid source molecular beam epitaxy on semi-insulating GaAs substrates. The active layer is composed of a 20-nm thick $In_{0.75}Ga_{0.25}As$ quantum well sandwiched between a lower 50-nm thick and an upper 120-nm thick $In_{0.75}Al_{0.25}As$ barriers. Details on the growth parameters can be found elsewhere [@biasiol05]. A 2DEG, confined in a narrow asymmetric $In_{0.75}Ga_{0.25}As$ quantum well, is characterized [@holmes; @inas] by strong Rashba-type SO coupling with Rashba constant $\alpha$ about $1-2 \times 10^{-11}$ eVm. The 2DEG mobility at 4K is $5 \cdot 10^{5} $cm$^{2}$/Vs and the carrier density is $4.1 \cdot 10^{11} $cm$^{-2}$, as obtained from standard magnetoresistance measurements. ![(Color online) Differential resistance $dV/dI$ of a single Nb-2DEG junction as a function of the dc voltage drop $V$ across the junction. The curves are denoted by the junction numbers, see Fig. \[sample\]. In this case of Nb electrodes, the limit of intermediate ($Z\approx 1.4$) scattering at the Nb-2DEG interface is realized, see Ref. . Each curve demonstrates a well developed $dV/dI$ resistance peak at low bias. The peak width $\approx 0.2$ mV (denoted by dashed lines) is the same for different Nb-2DEG junctions, so it is independent of the disorder at the interface. The inset demonstrates the enlarged region of normal resistance at $eV>\Delta_{Nb}$ for two perpendicular mesa edges. Black arrows indicate the expected from $T_c$ superconducting gap in niobium $\Delta_{Nb}\approx1.3$ meV. \[IVnb\]](nbinas2.eps){width="\columnwidth"} The sample is presented in Fig. \[sample\]. The 200 nm height mesa step is formed by wet chemical etching. Two perpendicular mesa edges are oriented along $[011]$ and $[01\overline{1}]$ crystallographic directions. In our $In_{0.75}Ga_{0.25}As$ structure a high quality edge contact to a 2DEG can be realized by evaporation of a metal over the mesa step, without annealing procedure [@biasiol08]. We thermally evaporate 10 nm Ni and 100 nm Au to obtain (normal) Ohmic contacts. Any of them is characterized by a strictly linear $I-V$ dependence at low temperatures with $\approx 500\Omega$ resistance. In addition, we use dc sputtering to place 100 nm thick niobium Nb (or 50 nm thick niobium nitride, NbN) superconductor stripes to overlap with mesa edges, see Fig. \[sample\]. The stripes are formed by lift-off technique, the surface is mildly cleaned in Ar plasma before dc sputtering. As it is usual for the superconductor thin film sputtering, a critical temperature $T_c$ can be seriously affected by the substrate. To avoid the 2DEG degradation, the sputtering is performed at room temperature. The Nb film is more robust: the superconducting gap $\Delta_{Nb}\approx 1.3$ meV obtained in Fig. \[IVnb\] is in a good correspondence with the expected $T_c\approx 9$ K. Room temperature sputtered NbN films are known to be subjected to a fragmentation onto normal and superconducting regions under a mechanical stress, which should be important at the mesa step [@nbn]. We measure $T_c$ to be about 11 K for a test NbN film at a flat surface (compare with 15 K bulk value), but Fig. \[IVnbn\] demonstrates even smaller values at the mesa step. A junction is formed between the superconducting (Nb or NbN) electrode and the 2DEG at the mesa edge. We study electron transport across one particular S-2DEG junction in a three-point configuration: a current is applied between one of the Ni-Au Ohmic contacts and a superconducting electrode which is grounded (contact 6 in Fig. \[sample\]) while another Ni-Au contact measures the 2DEG potential. To obtain $dV/dI(V)$ characteristics, we sweep the dc current through the interface from -5 $\mu$A to +5 $\mu$A. This dc current is modulated by a low (0.85 nA) ac (110 Hz) component. We measure both the dc ($V$) and ac ($\sim dV/dI$) components of the 2DEG potential by using a dc voltmeter and a lock-in amplifier, respectively. We have checked, that the lock-in signal is independent of the modulation frequency in the range 50 Hz – 300 Hz. This range is defined by applied ac filters. As usual in a three-point configuration, both $V$ and $dV/dI$ are determined by the ground contact, i.e. the measured $dV/dI(V)$ curves reflect the behavior of the S-2DEG interface. We verify this statement for our setup by demonstrating that the measured $V$ and $dV/dI$ are independent of a particular choice of current/voltage Ohmic contacts in Fig. \[sample\] for a fixed ground. The setup is also verified by obtaining a strictly linear $dV/dI(V)$ curve if the superconductivity is suppressed by high magnetic field or temperature. The measurements were performed at a temperature of 30 mK. Similar results were obtained from different samples of every type (Nb or NbN) in several cooling cycles. Experimental results ==================== The examples of $dV/dI$ characteristics are presented in Figs. \[IVnb\],\[IVnbn\] for two different superconductor materials, Nb and NbN. For both cases, the differential resistance $dV/dI$ is still finite within the superconducting gap $\Delta_{Nb}$ or $\Delta_{NbN}$ and strongly non-linear. We check, that the superconductivity, and, therefore, any $dV/dI(V)$ nonlinearity can be suppressed by the temperature or the magnetic field. The most important, we observe a well developed $dV/dI$ resistance peak at low biases, which can not be predicted by standard theories. The main $dV/dI$ behavior can be well understood within the framework of standard BTK theory [@tinkham]. Finite resistance is only allowed because of Andreev reflection, since a single-particle tunneling is prohibited within the superconducting gap. Andreev reflection is essentially affected by a single-particle scattering at the interface, which can be characterized by a scattering potential strength $Z$. We can define the regime of Andreev reflection by estimating $Z$ through a single-particle transmission $T$ of the interface ($T=1/(1+Z^2)$, see Ref. ). From the junction normal resistance and the junction width we obtain $T\approx 1$ for the NbN case and only $T\approx 0.3$ for the Nb junction, which corresponds to $Z_{NbN}\approx 0$ and $Z_{Nb}\approx 1.4$ respectively. The main behavior of the $dV/dI$ characteristics in Figs. \[IVnb\],\[IVnbn\] is in a good agreement with this $Z$ estimation. In the case of Nb electrode (see Fig. \[IVnb\]), the scattering is dominant at the interface ($Z\approx 1.4$), so the differential resistance $dV/dI$ at $eV<\Delta_{Nb}$ is increased with respect to the normal value. [@tinkham] This increase depends on a junction, because the Andreev process requires two-particle co-tunneling and therefore is extremelly sensitive to the disorder. At higher voltages $eV>\Delta_{Nb}$, $dV/dI$ approaches the normal resistance value, which is approximately the same for the junctions placed at the same mesa edge, cp. junctions 5,6 and 4,7 in Figs. \[sample\], \[IVnb\]. On the other hand, the normal resistance values are clearly different for two perpendicular $[011]$ and $[01\overline{1}]$ mesa edges, which reflects intrinsic in-plane mobility anisotropy of 2DEG in $In_{0.75}Ga_{0.25}As$ structure [@biasiol08]. ![(Color online) Differential resistance $dV/dI$ of a single NbN-2DEG junction as a function of the dc current $I$ through the junction. The curves are denoted by the junction numbers, see Fig. \[sample\]. In this case of NbN electrodes, the limit of low ($Z\approx 0$) scattering at the NbN-2DEG interface is realized, see Ref. . In this case, because of low 2$\mu$m junctions’ spacing, NbN-2DEG-NbN two-point resistance (dashed curve) is of standard Josephson behavior. Each particular NbN-2DEG junction demonstrates a clear visible $dV/dI$ resistance peak at low bias, similarly to Nb-2DEG case. By contrast, two-point NbN-2DEG-NbN resistance has no peculiarities at low currents. \[IVnbn\]](nbinas3.eps){width="0.85\columnwidth"} In the case of NbN electrode (see Fig. \[IVnbn\]), the limit of low scattering ($Z\approx 0$) at the S-2DEG interface is realized, so the differential resistance $dV/dI$ is diminished [@tinkham] within the superconducting gap $\Delta_{NbN}$. While the curves in Fig. \[IVnbn\] are qualitatively similar for two neighbor junctions, they demonstrate different normal resistance values, different drop in $dV/dI$ within the gap, and even different $\Delta_{NbN}$ values. This indicates the presence of normal regions of NbN at the mesa step. [@batov] In this case, the measured voltage $V$ is not directly a voltage drop at the S-2DEG interface. For this reason, we draw $dV/dI$ in Fig. \[IVnbn\] as a function of a current $I$ through the junction which is unambiguously defined in our setup. We can be sure, however, that the qualitative $dV/dI$ behavior reflects the S-2DEG interface properties. For example, because of low 2$\mu$m junctions’ spacing and transparent interfaces, NbN-2DEG-NbN two-point resistance (dashed curve in Fig. \[IVnbn\]) is of standard Josephson behavior. The most unexpected our experimental finding is a well developed $dV/dI$ resistance peak at low biases, which is present in both (Nb and NbN) scattering regimes. The resistance peak indicates, that the Andreev reflection is (partially) suppressed within a narrow energy region around zero bias. Fig. \[IVnb\] demonstrates that the $dV/dI$ peak width $\approx 0.2$ mV is universal for different Nb-2DEG junctions. Similarly to Nb-2DEG case, every NbN-2DEG junction demonstrates a clear visible $dV/dI$ resistance peak of a roughly constant width. We should conclude, that the central $dV/dI$ peak is independent of the disorder at the particular interface. It is important that two-point NbN-2DEG-NbN Josephson curve has no peculiarities at low currents. We find that the central $dV/dI$ peak is extremely sensitive to the magnetic field and the temperature, see Fig. \[IVev\]. Since the resistance peak is mostly prominent in Fig. \[IVnb\] for Nb electrodes, we only demonstrate this case below. The central $dV/dI$ peak disappears completely above $B=1.5$ T, see Fig. \[IVev\] (a). By contrast, the $dV/dI(V)$ curve is clearly non-linear below the upper critical field $B_c=3.8$ T of a niobium film in a present magnetic field configuration. We use the in-plane oriented field to avoid orbital effects in the 2DEG. Fig. \[IVev\] (b) demonstrates typical evolution of the $dV/dI(V)$ curve with temperature. The $dV/dI$ resistance peak is gradually diminishing with the temperature increase and disappears completely at 0.88 K. By contrast, the curve itself is practically insensitive to the temperature below 0.88 K, as it can be expected because of much higher $T_c=9$ K in niobium. The characteristic values of the $dV/dI$ peak suppression (1.5 T and 0.88 K) are also the same for the NbN-2DEG junction. Similarly to the Nb case, $dV/dI(V)$ curves are not affected below this values except for the central peak region. ![(Color online) The central $dV/dI$ peak suppression by in-plane magnetic field (a) or temperature (b) in the case of a Nb-2DEG junction. The central $dV/dI$ peak disappears completely in 1.5 T, much before the upper critical field $B_c=3.8$ T in the present in-plane field configuration. The $dV/dI$ resistance peak also disappears completely at 0.88 K, while the $dV/dI(V)$ curve is insensitive to the temperature up to this value. The curves in (a) panel are shifted vertically for clarity. \[IVev\]](nbinas4.eps){width="\columnwidth"} Discussion ========== Let us highlight the most important experimental result: a strong increase of the resistance at low bias indicates a suppression of the Andreev reflection within a narrow energy range. This suppression is very unusual, by contrast to widely observed Andreev reflection enhancement in different systems, see, e.g., Ref. . Since the central $dV/dI$ peak is present for two different superconductor materials with clearly different S-2DEG interface properties, the disorder at the interface cannot be responsible for the quite universal $dV/dI$ behavior at low bias. We have to connect the observed $dV/dI$ peak with some 2DEG-related effect, which is substantiated by the flat NbN-2DEG-NbN Josephson curve at low currents. The only relevant [@coulomb] energy scale in this case is the Rashba SO splitting [@holmes] $\Delta_{SO}=\hbar^2 k_{SO}^2/2m \sim 0.1$ meV in the 2DEG spectrum, where $k_{SO}$ is defined as $\alpha m/\hbar^2$. In principle, strong Rashba-type SO coupling affects the Andreev reflection as predicted by calculation in Ref. . However, the effect is quite small, and Ref. does not predict any low energy physics [@takane]. A clear physical explanation of our results can be connected with a spin-Hall effect [@sinova], which has been directly demonstrated in our 2DEG by using ferromagnetic side contacts [@feinas]. In the case of a strong Rashba-type SO coupling, the electric current within the 2DEG plane is expected to cause a non-zero, out-of-plane spin polarization (spin accumulation) $S_z$ near the sample edges even in zero magnetic field [@halperin04]. For a current, flowing through a S-2DEG junction, out-of-plane spin polarization $S_z$ is accumulated around the junction corners. Andreev reflection is suppressed for transport from these regions to the Nb (NbN) contact, because of broken usual requirements [@tinkham] for spin projections and energies of two electrons to be combined in a pair. The junction width is effectively diminished, which gives rise to the increased differential resistance $dV/dI$ around zero bias as shown in Figs. \[IVnb\] and \[IVnbn\]. In picture the central $dV/dI$ peak width is determined by the chemical potential difference for spin-up and spin-down electrons and obviously independent of the disorder at the interface, as we observe in Fig. \[IVnb\]. On the other hand, flat two-point NbN-2DEG-NbN curve in Fig. \[IVnbn\] supports the proposed explanation, since a zero-resistant Josephson state is insensitive to the contact effective width. This conclusion is also supported by the strong temperature dependence at $T<<T_c$ in Fig. \[IVev\] (b): if the temperature exceeds the value of the spin-orbit splitting $\Delta_{SO} \approx 0.1$ meV, all the effects of the spin polarization disappear in the 2DEG, and the interface resistance is diminished, as we do observe in Fig. \[IVev\] (b) at $T\sim 1$ K. An in-plane magnetic field has a similar effect: in our $In_{0.75}Ga_{0.25}As$ quantum well the Zeeman splitting exceeds [@holmes] $\Delta_{SO}$ at $B\approx 1.5$ T, so the spin-Hall effect disappears, and the interface resistance is diminished, as can be seen in Fig. \[IVev\] (a). Conclusion ========== As a result, we experimentally investigate transport properties of a single junction between a superconductor and the edge of a two-dimensional electron system in a narrow $In_{0.75}Ga_{0.25}As$ quantum well with strong Rashba-type spin-orbit coupling. We experimentally demonstrate suppression of Andreev reflection within a narrow energy range within the superconducting gap. This suppression is shown to be independent of the superconductor material and the interface disorder. The effect is only present at lowest temperatures and magnetic fields, far from the corresponding critical values of a superconducting transition. We connect the observed suppression with a spin-Hall effect due to a strong spin-orbit coupling on the normal side of the junction. We wish to thank Ya. Fominov, A.M. Bobkov, and I.V. Bobkova for fruitful discussions. We gratefully acknowledge financial support by RFBR and RAS. N.Titova acknowledges financial support from the Ministry of Education and Science of the Russian Federationunder Contract No. 14.B25.31.0007. [99]{} F. Wilczek, Nature Phys. 5, 614–618 (2009). For recent reviews, see C. W. J. Beenakker, Annu. Rev. Con. Mat. Phys. 4, 113 (2013) and J. Alicea, Rep. Prog. Phys. 75, 076501 (2012). L. Fu and C. L. Kane, Phys. Rev. Lett. 100, 96407 (2008). J. D. Sau, R. M. Lutchyn, S. Tewari, S. Das Sarma, Phys. Rev. Lett. 104, 040502 (2010). A. C. Potter, P. A. Lee, Phys. Rev. Lett. 105, 227003 (2010). J. Alicea, Phys. Rev. B 81, 125318 (2010). S. Nakosai, Y. Tanaka, and N. Nagaosa, Phys. Rev. Lett. 108, 147003 (2012) I. E. Batov, Th. Schäpers, A. A. Golubov, and A. V. Ustinov, J. Appl. Phys. 96, 3366 (2004). G.E. Blonder, M. Tinkham, T.M. Klapwijk, Physical Review B, 25, 4515, (1982); M. Tinkham, Introduction to Superconductivity (2d ed., McGraw–Hill, New York, 1996). T. Yokoyama, Y. Tanaka, and J. Inoue, Phys. Rev. B 74, 035318 (2006). J. Sinova, D. Culcer, Q. Niu, N. A. Sinitsyn, T. Jungwirth, and A. H. MacDonald, Phys. Rev. Lett. 92, 126603 (2004). F. Capotondi, G. Biasiol, D. Ercolani, V. Grillo, E. Carlino, F. Romanato, and L. Sorba, Thin Solid Films 484, 400 (2005). S.N. Holmes, P.J. Simmonds, H.E. Beere, F. Sfigakis, I. Farrer, D.A. Ritchie, and M. Pepper, J. Phys.: Condens. Matter, 20, 472207 (2008). A. Kononov, G. Biasiol, L. Sorba, and E. V. Deviatov Phys. Rev. B 86, 125304 (2012). Y. M. Shy, L. E. Toth, and R. Somasundaram, Journal of Applied Physics 44, 5539 (1973); Zhen Wang, Akira Kawakami, Yoshinori Uzawa, and Bokuji Komiyama, Journal of Applied Physics 79, 7837 (1996); D. D. Bacon, A. T. English, S. Nakahara, F. G. Peters, H. Schreiber, W. R. Sinclair, and R. B. van Dover, Journal of Applied Physics 54, 6509 (1983). D. Ercolani, G. Biasiol, E. Cancellieri, M. Rosini, C. Jacoboni, F. Carillo, S. Heun, L. Sorba, and F. Nolting, Phys. Rev. B 77, 235307 (2008). F. Giazotto, P. Pingue, and F. Beltram Modern Physics Letters B, 17, 955 (2003); F. Carillo, D. Born, V. Pellegrini, F. Tafuri, G. Biasiol, L. Sorba, and F. Beltram, Phys. Rev. B 78, 052506 (2008). For example, Coulomb blockade effects are negligible for the 20-$\mu$m wide junction. If we even suppose that current flows within the much more narrow region, e.g. because of some defect, we have to suppose that this defect is exactly the same for all the junctions for different samples. This seems to be impossible. Y. Takane and R. Ando, J. Phys. Soc. Jpn. 83, 014706 (2014). A. Kononov, S.V. Egorov, G. Biasiol, L. Sorba, E.V. Deviatov, Phys. Rev. B 89, 075312 (2014). E. G. Mishchenko, A.V. Shytov, and B. I. Halperin, Phys. Rev. Lett. 93, 226602 (2004).	{ "pile_set_name": "ArXiv" }
--- abstract: 'We show that the displacement and translation distance of non-elementary random walks on isometry groups of hyperbolic spaces satisfy large deviation principles with the same rate function $I$. Roughly, this means that there exists function $I(t)$ which accurately predicts the exponential decay rate of the probability that the translation distance of a random product of length $n$ is $tn$, and similarly for the displacement. This settles a special case of a conjecture concerning the large deviation principle for the spectral radius of random matrix products. In a second part, we give a characterization of the effective support of the rate function only in terms of the deterministic notion of joint stable length. Finally, as a by-product of our techniques, we deduce some further deterministic results on the asymptotics of a bounded set of isometries. Some of the results in this paper were obtained simultaneously and independently by Boulanger–Mathieu as we discuss in the introduction.' address: - 'Institut für Mathematik, Universität Zürich, 190, Winterthurerstrasse, 8057 Zürich, Switzerland' - 'Departement Mathematik, ETH Zürich, 101, Rämistrasse, 8092 Zürich, Switzerland' author: - Cagri Sert - Alessandro Sisto bibliography: - 'biblio.bib' title: 'Large deviation principles for non-elementary random walks on hyperbolic spaces' --- Introduction ============ Let $(X,d)$ be a (Gromov) hyperbolic metric space, e.g. a CAT$(-1)$-space, and let ${\operatorname{Isom}}(X)$ be the group of isometries of $X$. Let $\mu$ be a probability measure on ${\operatorname{Isom}}(X)$ and $Y_1,Y_2,\cdots$ be a sequence of independent random variables with distribution $\mu$. Given $x \in X$ and for $n \in \mathbb{N}$, denoting by $L_n$ the random product $Y_n \cdots Y_1$, we obtain a natural random process on $X$ given by $L_n x$. (We remark that whereas we state our probabilistic results for the left random walk $L_n=Y_n \ldots Y_1$, the same results hold for the right random walk $R_n=Y_1 \ldots Y_n$, since $L_n$ and $R_n$ have the same distributions given by the convolution $\mu^{\ast n}$.) Fixing a basepoint $x \in X$ and for $g \in {\operatorname{Isom}}(X)$, denote by $\|g\|_{x}$ the distance $d(gx,x)$. The goal of this paper is to study the probabilistic properties of $\|L_n\|_{x}$ as well as those of the translation distance $\tau(L_n):=\inf_{y \in X}d(L_n y,y)$ from a large deviations perspective, that is, studying the probability that average $\frac{1}{n}\tau(L_n)$ takes some specific value far from its expectation (see Definition \[defn.LDP\]). The process $L_n x$ can be thought of as a random walk on the metric space $(X,d)$; indeed, in the special case where $(X,d)$ is the Cayley graph of a hyperbolic group $G$ with respect to a generating set $S$ and $\mu$ is a probability measure supported on $S$, the process $L_n x$ is precisely a random walk on the Cayley graph starting from $x \in G$. The study of random walks on non-commutative groups and the relations of their properties with the geometric and algebraic structure of groups have a long history, we refer the readers to the books [@woess; @pete] for a general overview. Regarding the asymptotic properties of $\|L_n\|_{x}$, despite the non-commutativity of the underlying group and mere subadditivity of the displacement functional $\|.\|_{x}$, under quite general assumptions, the random variables $\|L_n\|_{x}$ tend to behave as a sum of independent real random variables. Under some more restrictive assumptions, the same is true for the translation distance $\tau(L_n)$ which do not enjoy subadditivity either. It is a general fact that, with a finite first moment assumption on $\mu$, the sequence $\|L_n\|_{x}$ satisfies a law of large numbers, namely $\frac{1}{n} \|L_n\|_{x}$ converges almost surely to a limit $\ell_\mu \in \mathbb{R}$, called the drift or rate of escape of the random walk. This follows from Kingman’s subadditive ergodic theorem. In special cases of the general setting that we will consider in this article, other limit theorems for $\|L_n\|_{x}$, such as central or local limit theorem and large deviation estimates were studied by several authors, establishing a relatively satisfactory understanding of the random walks on groups with hyperbolic features (see below). Our goal here is to prove the analogue of the classical result of Cramér’s on large deviation principles, for the random variables $\|L_n\|_{x}$ and $\tau(L_n)$. We now describe our setting. Throughout the article, $(X,d)$ is a separable geodesic metric space that is Gromov hyperbolic (see [@CDP Ch. 1]). The group ${\operatorname{Isom}}(X)$ is endowed with the topology of pointwise convergence and the corresponding Borel $\sigma$-algebra. For the probability measure $\mu$, we shall always assume that there exists a closed, first countable and separable subgroup $\mathcal{H}<{\operatorname{Isom}}(X)$ such that $\mu(\mathcal{H})=1$. Such a probability measure on ${\operatorname{Isom}}(X)$ will be called admissible. We note that if $\mu(G)=1$ for some countable group $G$, then $\mu$ is admissible and the reader is invited to think of this case in the sequel. We shall say that a subset $S<{\operatorname{Isom}}(X)$ is non-elementary if the smallest group containing $S$ contains at least two hyperbolic elements with disjoint fixed point sets on the Gromov boundary $\partial X$ of $X$. Note that the group ${\operatorname{Isom}}(X)$ acts on the Gromov boundary $\partial X$ and an element $g \in {\operatorname{Isom}}(X)$ is said to be hyperbolic if it has two fixed points on $\partial X$. Accordingly, a probability measure $\mu$ on ${\operatorname{Isom}}(X)$ is called non-elementary if its support is (see §\[subsec.weak.LDP\]). These will be the standing assumptions on the distribution $\mu$ of random products $Y_i$ considered in this article. This setting comprises a broad set of actions of non-amenable groups, for some particular cases, see [@maher-tiozzo page 1]. Some of the first instances where the negative curvature was exploited in the setting of random walks on groups to deduce some limit theorems are in the works of Ledrappier [@ledrappier.free] (central limit theorem), Gerl [@gerl.uber] and Sawyer [@sawyer] (local limit theorems) on free groups (although, earlier works studying various probabilistic phenomenon in relation to negative or non-positive curvature exist: e.g. Dynkin–Malyutov [@dynkin-malyutov], Tutubalin [@tutubalin], Cartier [@cartier] and others [@derriennic; @bougerol.symmetric; @furstenberg.poisson]). Ledrappier’s central limit theorem was later extended, in increasing order of generality, by Björklund [@bjorklund], Benoist–Quint [@benoist-quint.hyperbolic] and Mathieu–Sisto [@mathieu-sisto] who proved it for acylindrical actions. The local limit theorem for free groups was later extended by Lalley and Gouëzel [@lalley.local.free; @gouezel-lalley; @gouezel.local] for symmetric random walks on hyperbolic groups. The study of large deviations on hyperbolic-like groups, compared to aforementioned limit theorems, is rather incomplete. By Kingman’s theorem, the probability of large deviations off the mean converges to zero, i.e. for every $\varepsilon>0$, $\mathbb{P}(\|\frac{1}{n}\|L_n\|_{x}-\ell_\mu\|>\varepsilon) \to 0$ as $n \to \infty$. For free groups, using the spectral gap result of Ledrappier [@ledrappier.free], one can deduce that the decay rate is exponential for every $\varepsilon>0$, following the strategy of Le Page [@lepage] (see also [@bougerol-lacroix]). As Gouëzel remarks [@gouezel.gap Page 4], this analytic approach carries over to random walks on hyperbolic groups and can be used to prove an exponential decay result in that setting. These parallel the corresponding large deviation results of Le Page [@lepage] for random matrix products which turned out to be a key ingredient in relatively recent works on dynamics on homogeneous spaces [@BFLM; @BQ2]. The relation with random matrix products is more than a mere analogy, since rank-one simple linear groups over local fields act isometrically on their symmetric spaces or the associated Bruhat–Tits buildings, which are hyperbolic metric spaces. In Corollary \[corol.conjecture\], we will exploit this relation to settle a case of a conjecture appearing in [@sert.LDP] for the spectral radius of random matrix products. We note that the aforementioned exponential decay result pertains rather, but not exclusively, to the non-amenable setting. Indeed, for example, for symmetric random walks on countable amenable groups, i.e. random walks driven by a probability measure $\mu$ satisfying $\mu(g)=\mu(g^{-1})$ for every $g\in {\operatorname{Isom}}(X)$, by a classical result of Kesten [@kesten.symmetric], the decay rate of $\mu^{\ast 2n}(g)$ is never exponential (see [@berg-christensen] for the locally compact case). Finally, we mention the recent work of Corso [@corso] where a large deviation principle was shown for word-lengths of random walks on free products. Main result on displacement --------------------------- To describe our results on large deviations, we first recall a definition. Let $Y$ be a topological space and let $\mathcal{F}$ be a $\sigma$-algebra on $Y$. \[defn.LDP\] A sequence $Z_{n}$ of $Y$-valued random variables is said to satisfy a large deviation principle (LDP), if there exists a lower-semicontinuous function (called the rate function) $I:Y \longrightarrow [0, \infty]$ such that for every measurable subset $R$ of $Y$, we have $$\underset{\alpha \in {\operatorname{int}}(R)}{-\inf I(\alpha)} {\leqslant}\underset{n \rightarrow \infty}{\liminf} \frac{1}{n}\log \mathbb{P}(Z_{n} \in R) {\leqslant}\underset{n \rightarrow \infty}{\limsup} \frac{1}{n}\log \mathbb{P}(Z_{n} \in R) {\leqslant}\underset{\alpha \in \overline{R}}{-\inf I(\alpha)}$$ where, ${\operatorname{int}}(R)$ denotes the interior and $\overline{R}$ the closure of $R$. Very roughly, the definition says that the probability that $Z_n$ (to be thought of as $\frac{1}{n}\|L_n\|_{x}$ or $\frac{1}{n}\tau(L_n)$ in our case) takes value approximately $\alpha$ is equivalent to $e^{-I(\alpha)n}$. With this definition, Cramér’s theorem says that for a sequence of real valued independent and identically distributed (i.i.d.) random variables $(Y_i)$ with finite exponential moment, the sequence of averages $\frac{1}{n}\sum_{i=1}^{n}Y_{i}$ satisfies a LDP with a proper convex rate function $I$, given by the convex conjugate (Legendre transform) of the Laplace transform of $Y_{i}$’s. A probability measure $\mu$ on ${\operatorname{Isom}}(X)$ is said to have a finite exponential moment if for some $\alpha>0$ and $x \in X$, we have $\int \exp(\alpha \|g\|_{x}) d\mu(g)<\infty$. This condition is obviously independent of the choice of $x$. We can now state our first result. \[thm.LDP.norm\] Let $(X,d)$ be a separable geodesic hyperbolic metric space and $\mu$ a non-elementary and admissible probability measure on ${\operatorname{Isom}}(X)$. Suppose that $\mu$ has a finite exponential moment. Then, for every $x \in X$, the sequence $\frac{1}{n}\|L_n\|_{x}$ of random variables satisfies a large deviation principle with a proper convex rate function $I:[0,\infty) \to [0,\infty]$. \[rk.after.thm.LDP.norm\] 1. We also prove the existence of a weak large deviation principle without any moment assumption (see Theorem \[thm.LDP.precise\]).\ 2. Under a stronger moment condition, by exploiting the convexity of $I$, we identify the rate function $I$ with the convex conjugate of a limit log-Laplace transform of the random variables $\frac{1}{n} \|L_n\|_x$. It is easy to see that the rate function $I$ given by Theorem \[thm.LDP.norm\] does not depend on the choice of the base point $x \in X$. Below, we list some more remarks on this result: \[rk.affter.thm.LDP.norm\] 1. By convexity of $I$, the set $D_I=\{\alpha \in [0,\infty) \, \| \, I(\alpha)<\infty \}$, called the effective support of $I$, is an interval and $I$ is continuous on $D_I$. By Theorem \[thm.LDP.norm\], this in turn implies that for every subset $J$ of $D_I$ satisfying $\overline{{\operatorname{int}}(J)}=\overline{J}$ (e.g. any interval with non-empty interior), the limit $\lim_{n \to \infty} \frac{1}{n} \log \mathbb{P}(\frac{1}{n}\|L_n\|_x \in J)$ exists and is equal to $\max_{\alpha \in J}-I(\alpha)$ (see Theorem \[thm.support\] for more on $D_I$).\ 2. It is not hard to see that under mild assumptions on the space $(X,d)$ and the action (see §\[subsec.doubling\]), for any symmetric finitely supported probability measure $\mu$, we have $I(0)=-\log (r_\mu)$, where $r_\mu$ is the spectral radius of the random walk, i.e. the spectral radius of the self-adjoint operator associated to $\mu$ through the regular representation on $\ell^2(G_\mu)$, which, by Kesten’s formula, is given by $\lim_{n \to \infty} \mu^{\ast 2n}(e)^{\frac{1}{2n}}$. Here $G_\mu$ denotes the group generated by $\mu$.\ 3. It is clear that the drift $\ell_\mu$ is a zero of the rate function $I$. The aforementioned property of exponential decay of probabilities of large deviations off the drift, which can be deduced from a spectral gap result [@gouezel.gap page 4] in the particular case of random walks on hyperbolic groups, is equivalent to $\ell_\mu$ being the unique zero of $I$. We were informed that an ongoing work of Boulanger–Mathieu proves this decay result in our setting (see Remark \[rk.comment\]). We do not deal with this point in this article. It is not very difficult to pinpoint the explicit expression of the rate function for the standard random walk on the free group $F_q$ of rank $q {\geqslant}1$. It is given by the following $$I(\alpha)= \begin{cases} \frac{1+\alpha}{2}\log(1+\alpha) + \frac{1-\alpha}{2} \log(1-\alpha) +\log (q) - \frac{1+\alpha}{2} \log(2q-1) \qquad &\alpha \in [0,1]\\ \infty &\text{otherwise} \end{cases}$$ We remark that, among others, this function satisfies the following properties:\ 1) it is analytic and strictly convex on its effective support $D_{I}:=\{x \in [0,\infty) \, \|\, I(x)<\infty\}$,\ 2) $I(0)=-\log \frac{\sqrt{2q-1}}{q}$ where $\frac{\sqrt{2q-1}}{q}$ is the spectral radius of the standard random walk on $F_q$ calculated by Kesten [@kesten.symmetric],\ 3) the drift $\frac{q-1}{q}$ is the unique zero of $I$,\ 4) if $\Lambda(\lambda)$ denotes the Legendre transform of $I$ given by $\Lambda(\lambda)=\sup_{\alpha \in \mathbb{R}} (\lambda \alpha-I(\alpha))$, then $\Lambda''(0)-(\frac{q-1}{q})^2$ is the asymptotic variance appearing in the central limit theorem for the standard random walk on the free group (this fact can be deduced either directly or as in [@bougerol-lacroix Lemma 5.2]).\ Whereas finding an explicit expression for the rate function $I$ does not seem to be feasible in general, pinning down some of its general properties, paralleling the above ones, is a more tractable challenge. For example, as mentioned previously, the properties 2) and 3) above are known to hold under general assumptions and the properties 1) and 4) naturally suggest the corresponding open problems. We mention only a few of them:\ Questions: Is the rate function appearing in Theorem \[thm.LDP.norm\] strictly convex? Analytic? Do these properties depend on generating set or probability measure? Regarding the proof of Theorem \[thm.LDP.norm\], we use a general criterion provided by Theorem \[thm.LDP.criterion\] implementing it in a similar way as in [@sert.LDP]. This requires being able to find, in quantitative and uniform fashion, local perturbations of the random walk such that the displacement functionals $\|.\|_{x}$ of the corresponding random walk on $X$ will have an almost-additive property. Such perturbations are found thanks to the use the negatively curved geometric structure in a form which manifests itself in the unfolding lemma (Lemma \[lemma.alignment\]). In our setting, this lemma yields the analogous tools provided by the results of Abels–Margulis–Soifer [@AMS] and Benoist [@benoist.proper] on the proximality and almost-norm-multiplicativity of linear transformations. Main result on translation distance ----------------------------------- Another important notion to measure the size of an isometry $g$ of $(X,d)$ is its translation length $\tau(g)=\inf_{x \in X} \|g\|_x$. It has the advantage of being more intrinsic, in particular it does not depend on the choice of a basepoint $x \in X$ and it is a conjugacy invariant. On the other hand, it displays a more chaotic behaviour compared to $\|g\|_x$. For example, the useful subadditivity property $\|gh\|_x {\leqslant}\|g\|_x + \|h\|_x$ fails for $\tau(.)$. The following theorem proves the existence of a LDP for the sequence $\frac{1}{n}\tau(L_n)$. Moreover, we note that, under a moment assumption, it can be shown that for a non-elementary random walk, on a set of large probability, the behaviour of $\frac{1}{n}\tau(L_n)$ is similar to that of $\frac{1}{n} \|L_n\|_{x}$ (for example, they converge almost surely to the same constant $\ell_\mu$). The following result also says that, at an exponential scale, the probabilistic behaviours of these two different notions of size, coincide not only for events of large probability but also for the rare (exponentially small) ones: \[Thm.LDP.translation\] Let $(X,d)$ be a separable geodesic hyperbolic metric space and $\mu$ be a boundedly supported, non-elementary and admissible probability measure on the group of isometries of $(X,d)$. The sequence $\frac{1}{n} \tau(L_n)$ of random variables satisfies a large deviation principle with the same rate function $I:[0,\infty) \to [0,\infty]$ given by Theorem \[thm.LDP.norm\]. Here, a probability measure $\mu$ on ${\operatorname{Isom}}(X)$ is said to be bounded if $\sup_{s \in {\operatorname{supp}}\mu} \|s\|_x $ is finite for some $x$ (equivalently for all $x \in X$). A common and sometimes more convenient way to express a notion of translation length of $g \in {\operatorname{Isom}}(X)$ is given by that of asymptotic translation length or stable length defined as $\ell(g)= \lim_{n \to \infty}\frac{1}{n}\|g^n\|_{x}$. The limit exists by subbadditivity and does not depend on $x$. The difference $\|\ell(.)-\tau(.)\|$ is uniformly bounded on ${\operatorname{Isom}}(X,d)$ (see [@CDP Ch.10, Prop. 6.4]). Consequently, the previous theorem applies equally to the random variables $\frac{1}{n} \ell(L_n)$ with the same conclusion. Consequences for rank-one linear groups --------------------------------------- Let us explain a consequence of this result that confirms a conjecture stated in [@sert.LDP] in a particular case. A simple linear algebraic group $H$ of rank $1$ over a local field $k$ (e.g. ${\operatorname{SL}}_2(\mathbb{R})$ or ${\operatorname{SL}}_2(\mathbb{Q}_p)$), has a natural, up to finite index, faithful action by isometries on its symmetric space or the associated Bruhat–Tits tree $(X,d)$. The metric space $(X,d)$ is a Gromov hyperbolic space and for some finite dimensional representation of $H$ on a $k$-vector space $V$ endowed with a Hermitian (if $k$ is archimedean) or ultrametric (if $k$ is non-archimedean) norm, and for $x \in X$ and $h \in H$, the displacement functional $\|h\|_x$ is given by the logarithm of the associated operator norm $\|\|.\|\|$ (see e.g. [@BQ.book Chapter 6,8] and [@quint.cones §6]). Moreover, the translation distance $\tau(h)$ corresponds to the logarithm of the spectral radius $\rho(h)$ of $h$, defined by the spectral radius formula $\rho(h)=\lim_{n \to \infty} \|\|h^n\|\|^{\frac{1}{n}}$. In this case, under the assumptions of Theorem \[thm.LDP.norm\], the existence of a convex rate function for $\frac{1}{n}\|L_n\|_x$ follows from the main result of [@sert.LDP] (as well as, from Theorem \[thm.LDP.norm\]). It was conjectured [@sert.LDP Conjecture 6.2] (see also [@breuillard-sert §5.15]) that, if the support of the probability measure $\mu$ on $H$ generates a Zariski-dense semigroup (equivalently, if $\mu$ is non-elementary), then the sequence $\frac{1}{n}\log \rho(L_n)$ satisfies a LDP and the rate function coincides with the rate function of the sequence $\frac{1}{n}\log \|\|L_n\|\|$. Now one sees that for a probability measure $\mu$ with a compact support, this conjecture follows from Theorem \[Thm.LDP.translation\]: \[corol.conjecture\] Let $H$ be a simple linear algebraic group of rank one over a local field $k$ endowed with an absolute value $\|.\|$. Let $\mu$ be a compactly supported probability measure on $H$ whose support generates a Zariski dense semigroup in $H$. Let $\|\|.\|\|$ be an operator norm on a finite dimensional representation $V$ of $H$ as above and $I:[0,\infty) \to [0,\infty]$ be the rate function of the LDP of $\frac{1}{n}\log \|\|L_n\|\|$. Then, the sequence $\frac{1}{n}\log \rho(L_n)$ of random variables satisfies a LDP with rate function $I$. On the proof of Theorem \[Thm.LDP.translation\] ----------------------------------------------- Theorem \[Thm.LDP.translation\] cannot be proven with methods similar to those that work for Theorem \[thm.LDP.norm\], due to the more chaotic behavior of $\tau(.)$. We use genuinely different ideas for its proof, in particular cyclic permutations of trajectories. In fact, the proof of Theorem \[Thm.LDP.translation\] uses Theorem \[thm.LDP.norm\], and consists of several parts. In a first part, using the unfolding lemma, we show that, given a prescribed speed $\alpha {\geqslant}0$, for large $n$, the event $\frac{1}{n}\tau(L_n) \sim \alpha$ is, at an exponential scale of the probability, more likely than $\frac{1}{n}\|L_n\|_x \sim \alpha$. In the second part, for all prescribed speeds $\alpha>0$, we show that, up to a linear loss (i.e. up to multiplying one of the probabilities by $n$), the converse inequality holds for probabilities of these events, which implies the desired inequality of exponential rates. This step relies on Lemma \[lemma.getting.norm.close.to.tau\] that uses an argument that finds, among the cyclic permutations of a given trajectory, a word whose displacement $\|.\|_x$ is uniformly close to the translation distance, which is invariant by cyclic permutation. Finally, the control of exponential decay rates for sublinear speed, requires an upgrade of the previous combinatorial lemma in order to carry out a contradiction argument using Theorem \[thm.LDP.norm\]. Properties of the rate function ------------------------------- A natural question motivated by the previous results concerns the understanding of the effective support $D_I=\{\alpha \in [0,\infty) \, \| \, I(\alpha)<\infty \}$ of the rate function $I$. By convexity of the rate function, the effective support $D_I$ is an interval in $[0,\infty)$. The following result gives a geometric characterization of this interval only in terms of the support of the probability measure $\mu$ and it relates the effective support with the recently introduced notion of asymptotic joint displacement of a bounded set of isometries of a metric space. To state this result, we need some terminology. A set $S$ of isometries of a metric space $(X,d)$ is said to be non-arithmetic if there exist $n \in \mathbb{N}$ and $g,g' \in S^n$ such that $\ell(g)\neq \ell(g')$. As in [@benoist-quint.hyperbolic], we shall also call a probability measure non-arithmetic if its support is. A set $S$ is said to be bounded if the quantity $L(S,x):=\sup_{s \in S} \|s\|_x $, called the maximal displacement of $S$ at $x \in X$, is bounded for some $x$ (equivalently for all $x \in X$). The joint minimal displacement of $S$ is defined to be $L(S):=\inf_{x \in X} L(S,x)$. We shall denote the limit $\lim_{n \to \infty}\frac{1}{n}L(S^n)$, which exists by subadditivity, by $\ell(S)$; it is called the asymptotic joint displacement [@breuillard-fujiwara] or joint stable length [@oregon.reyes:properties] of $S$. Similarly, let $\ell_{sub}(S)$ denote the limit $\lim_{n \to \infty} \inf_{g \in S^n} \frac{1}{n}\|g\|_x$ that we call lower asymptotic joint displacement of $S$. The previous limit exists by subadditivity and does not depend on $x$. \[thm.support\] Let $(X,d)$ be a separable geodesic hyperbolic metric space and $\mu$ be a non-elementary and admissible probability measure on the group of isometries of $(X,d)$. Let $I$ be the rate function given by Theorem \[thm.LDP.norm\] (see also Theorem \[thm.LDP.precise\]).\ 1) The probability measure $\mu$ is non-arithmetic if and only if the effective support $D_I$ of $I$ is an interval with non-empty interior.\ 2) If the support $S$ of $\mu$ is a bounded set, we have $(\ell_{sub}(S),\ell(S)) \subseteq D_I \subseteq \overline{D}_I \subseteq [\ell_{sub}(S),\ell(S)]$.\ 3) If $S$ is finite, then $D_I=[\ell_{sub}(S),\ell(S)]$. Regarding the second statement in the previous theorem, in §\[subsec.support\], we provide examples of probability measures $\mu$ with bounded support $S$ for which the rate function $I$ explodes on the boundary points of $D_I$. The notion of asymptotic joint displacement is analogous to the classical notion of joint spectral radius from linear algebra. In this geometric setting, it was recently studied by Oregón-Reyes [@oregon.reyes:properties] and Breuillard–Fujiwara [@breuillard-fujiwara] who proved the geometric analogues of some of the main results on joint spectral radius. The previous result parallels [@sert.LDP Theorem 1.7] where the effective support of the rate function of the norms of random matrix products was related to joint spectral radii. Deterministic results --------------------- Here we record some deterministic (as opposed to probabilistic) consequences of our results and the ingredients we develop to prove them. ### Hausdorff convergence The following is a consequence of the combination of Theorems \[thm.LDP.norm\], \[Thm.LDP.translation\] and \[thm.support\]: \[prop.joint.spectrum\] Given a bounded non-elementary subset $S$ of ${\operatorname{Isom}}(X)$, the sequences of subsets $\frac{1}{n}\|S^n\|_x$ and $\frac{1}{n}\tau(S^n)$ of $\mathbb{R}$ converge to $[\ell_{sub}(S),\ell(S)]$ with respect to the Hausdorff metric. In other words, the sequences $\frac{1}{n}\|S^n\|_x$ and $\frac{1}{n}\tau(S^n)$ become more and more dense in the interval $[\ell_{sub}(S),\ell(S)]$ as $n$ grows. In fact, Theorems \[thm.LDP.norm\] and \[Thm.LDP.translation\] can be seen as quantitative refinements of this convergence. This result parallels the convergence result proven in [@breuillard-sert Theorem 1.3] for the vectors of singular values and moduli of eigenvalues of powers of a set of matrices. Indeed, for the particular case of rank-one symmetric spaces of non-compact type, this convergence follows from [@breuillard-sert]. The interval $[\ell_{sub}(S),\ell(S)]$ corresponds to what is called the joint spectrum of $S$ in that article. In passing, we also note that a direct deterministic consequence of Theorem \[thm.support\] (in fact, rather, of its proof, see Remark \[rk.arithmetic\]) regarding the asymptotic joint displacements/joint stable length is the following: Given a non-elementary and bounded subset $S$ of the isometry group of a hyperbolic space $(X,d)$, the set $S$ is non-arithmetic if and only if $\ell_{sub}(S) \neq \ell(S)$. To describe our next result, let $\mu$ be a probability measure of bounded support on ${\operatorname{Isom}}(X,d)$ for some metric space $(X,d)$. Without any appeal to the previous theorem, it is obvious that the drift $\lambda(\mu)$ of a random walk belongs to the interval $[0,\ell(S)]$. In the setting of random walks on acylindrical hyperbolic groups, previous results of Maher–Tiozzo (who else to cite here) ... establishes that, in fact, $\lambda(\mu) \neq 0$. Together with the previous results on limit theorems for random walks on acylindrically hyperbolic groups (namely [@benoist-quint.hyperbolic; @mathieu-sisto]), the geometric techniques that we develop to prove our results on large deviations allow us to prove that the drift cannot attain the other bound either: \[thm.drift\] Let $(X,d)$ be a geodesic and hyperbolic metric space and $\mu$ be a probability measure on the group of isometries of $(X,d)$ whose support generates an acylindrically hyperbolic group (plus exponential moment to be able to use Mathieu-Sisto). Let $I$ be the rate function given by Theorem \[thm.LDP.norm\]. Then, $\lambda(\mu) \in {\operatorname{int}}{D}_I$, where ${\operatorname{int}}{D}_I$ is the interior of the interval $D_I \subseteq [0,\infty]$. [[Ideally, this result would read: let $\mu$ be a non-arithmetic and non-elementary probability measure with finite first moment (or second..) then $\lambda(\mu)$ lies in the interior of $D_I$.]{}]{} [[Mathieu-Sisto says that deviation inequalities give lower bound for the variance. Adrien and Pierre are getting deviation inequalities in this more general setting, so maybe they can see that there is some sort of positive variance: I remind that we don’t even need that the variance is positive.. for this result we just need that if the drift of $\mu$ supported on $S$ is $\lambda$, then $S^n$ contains elements whose $\|.\|_x$ is (as $n$ gets larger) arbitrarily far from $n \lambda$ in both directions (i.e. larger and smaller).]{}]{} In this result, ### Berger–Wang equality We finally mention a deterministic consequence of the geometrical ingredients that we develop to prove Theorem \[thm.LDP.norm\] on large deviations. For a subset $S$ of ${\operatorname{Isom}}(X,d)$, we denote by $\ell_{\max}(S)$ the maximal stable length $\max_{g \in S} \ell(g)$ and by $\ell_{\infty}(S)=\sup_{k \in \mathbb{N}}\frac{1}{k}\ell_{\max}(S^k)$. Clearly $\ell_{\max}(S) {\leqslant}\ell_{\infty}(S) {\leqslant}\ell(S)$ (see also [@breuillard-fujiwara Lemma 1.1 ]). Using the unfolding lemma, we can easily deduce \[prop.berger-wang\] For a bounded non-elementary subset $S$ of ${\operatorname{Isom}}(X,d)$, we have $\ell_\infty(S)=\ell(S)$. This result is a geometric analogue of a result due to Berger–Wang which tells that for a bounded set $S$ of matrices in ${\operatorname{Mat}}(d,\mathbb{C})$, one has $\sup_{k \in \mathbb{N}} \sup_{g \in S^k}\frac{1}{k} \log \rho(g)=\lim_{n \to \infty} \sup_{g \in S^n} \frac{1}{n} \log \|\|g\|\|$, where $\rho(.)$ denotes the spectral radius and $\|\|.\|\|$ is any operator norm on the algebra ${\operatorname{Mat}}(d,\mathbb{C})$. We note that a stronger (without the non-elementary assumption) form of Proposition \[prop.berger-wang\] as well as more general Bochi-type inequalities were recently proven by Oregón-Reyes [@oregon.reyes:properties] and by Breuillard–Fujiwara [@breuillard-fujiwara]. We note that using the Berger–Wang inequality, one can show that the joint stable length/asymptotic joint displacement $\ell(.)$ is continuous on the set of compact subsets of ${\operatorname{Isom}}(X)$ endowed with the Vietoris topology (see [@oregon.reyes:properties]). On the other hand, in the particular case of $X=\mathbb{H}^2$, in [@bochi-morris Corollary 1.9] Bochi–Morris proves also the continuity of the lower asymptotic joint displacement $\ell_{sub}(.)$. This leaves us wonder whether the interval $[\ell_{sub}(.),\ell(.)]$ is continuous in general (see also [@oregon.reyes:properties §6]). Outline ------- The paper is organized as follows. In Section \[sec.LDP.displacement\], we introduce some preliminaries of large deviation theory and we prove Theorem \[thm.LDP.precise\] which is a more general version of Theorem \[thm.LDP.norm\]. In Section \[sec.tau\], after introducing the necessary geometric ingredients, we prove Theorem \[Thm.LDP.translation\] in three subsection corresponding to three parts of the proof. Finally, in Section \[sec.support\], we prove Theorem \[thm.support\] as well as Propositions \[prop.joint.spectrum\] and \[prop.berger-wang\]. \[rk.comment\] Before the completion of this article, we were informed that Adrien Boulanger and Pierre Mathieu were also studying the large deviations phenomena in our setting. In particular, for countably supported probability measures, Theorem \[thm.LDP.norm\] also follows from their work. Our proofs for that result are somewhat similar. Boulanger–Mathieu then concentrate on the positivity properties of the rate function and they show that in our setting, for countable groups, the rate function has a unique zero. Unlike the analytic approach mentioned above to tackle such problems, their work, together with an observation of Hamana [@hamana], makes use of various nice elaborations of deviation inequalities developed in [@mathieu-sisto], thereby exploiting directly the coarse negative curvature to tackle this problem. However, they do not treat large deviations of translation distance (Theorem \[Thm.LDP.translation\]) and the characterization of the support of the rate function (Theorem \[thm.support\]). Our works can be seen to be complementary to each other. Acknowledgements {#acknowledgements .unnumbered} ---------------- We thank Pierre Mathieu for remarks on a draft version of this article. C.S. is supported by SNF grant 182089. Large deviation principle for displacement functionals {#sec.LDP.displacement} ====================================================== In §\[subsec.weak.LDP\], we introduce the notion of weak LDP and state the more precise version of Theorem \[thm.LDP.norm\]. We prove the existence of weak LDP for displacement functionals in §\[subsec.weak.LDP\] and in §\[subsec.convexity\], we show the convexity of the rate function. Finally, in §\[subsec.proof.of.A\], we complete the proof of Theorem \[thm.LDP.precise\], and hence that of Theorem \[thm.LDP.norm\]. Existence of Weak LDP with a convex rate function {#subsec.weak.LDP} ------------------------------------------------- Recall that by definition of an admissible probability measure $\mu$, there exists a closed, first countable and separable group $\mathcal{H}$ such that $\mu(\mathcal{H})=1$. By Birkhoff-Kakutani theorem, such a group is metrizable and by [@parthasarathy Theorem 2.1], there exists a unique minimal closed subset $S$ of $\mathcal{H}$ with $\mu(S)=1$. The subset $S$ is called the support of $\mu$. It has the property that for every $g \in S$ and any neighborhood $U$ of $g$, $\mu(U)>0$. The subgroup of $\mathcal{H}$ generated by $S$ will be denoted by $G$. To avoid any measurability issue, we complete the Borel $\sigma$-algebra of $\mathcal{H}$ with respect to the convolutions $\mu^{\ast n}$, which are defined as push-forwards of the product measures $\mu^{\otimes n}$ on $G^n$ to $G$ by the multiplication map. Finally, given such a probability measure $\mu$ on ${\operatorname{Isom}}(X)$, we fix some probability space $(\Omega,\mathfrak{F},\mathbb{P})$ on which the independent random walk increments $Y_i$ for $i {\geqslant}1$ are defined and have distribution $\mu$.\ In Definition \[defn.LDP\], a LDP with a rate function $I$ for a sequence of random variables $Z_{n}$ (in our case, to be thought of as $\frac{1}{n} \|L_n\|_x$ or $\frac{1}{n}\tau(L_n)$) with values in a topological space $Y$ (in our case $Y=[0,\infty)$), can be reformulated as saying\ 1. (Upper bound) For any closed set $F \subset Y$, $\underset{n \rightarrow \infty}{\limsup} \frac{1}{n}\log \mathbb{P}(Z_{n} \in F) {\leqslant}\underset{\alpha \in F}{-\inf I(\alpha)}$.\ 2. (Lower bound) For any open set $O \subset Y$, $\underset{n \rightarrow \infty}{\liminf} \frac{1}{n}\log \mathbb{P}(Z_{n} \in O) {\geqslant}\underset{\alpha \in O}{-\inf I(\alpha)}$. The definition of a weak LDP is a slight weakening of the upper bound in the previous reformulation: A sequence of $Y$-valued random variables $Z_{n}$ is said to satisfy a weak LDP with a rate function $I:Y \to [0,\infty]$ if the upper bound 1. (above) holds for all compact sets and the lower bound 2. also holds, for all open sets in $Y$. We note in passing that if $Y$ is locally compact or a polish space and a sequence of random variables $Z_{n}$ on $Y$ satisfies a weak LDP with a rate function $I$, then $I$ is unique. Recall that the limit Laplace transform of the sequence $(\|L_n\|_x)_{n}$ is the function $\Lambda: \mathbb{R} \to \overline{\mathbb{R}}$ defined as $$\Lambda(\lambda) = \limsup_{n \rightarrow \infty}\frac{1}{n}\log \mathbb{E}[e^{\lambda \|L_n\|_x}]$$ Nice analytic properties (e.g. differentiability, steepness) of this function have direct implications for the LDP (see e.g. Gärtner-Ellis theorem [@dembo-zeitouni §4]). Finally, we say that the probability measure $\mu$ has a strong exponential moment if $ \int \exp(\alpha \|g\|_x) d\mu(g)<\infty$ for every $\alpha>0$. We can now state the more precise version of Theorem \[thm.LDP.norm\]. \[thm.LDP.precise\] Let $(X,d)$ be a separable geodesic hyperbolic metric space and $\mu$ be a non-elementary and admissible probability measure on ${\operatorname{Isom}}(X)$. Then, for every $x \in X$, the sequence $\frac{1}{n}\|L_n\|_x$ of random variables satisfies a weak large deviation principle with a convex rate function $I:[0,\infty) \to [0,\infty]$. If $\mu$ has a finite exponential moment, the (full) LDP holds and the rate function $I$ is proper. If, furthermore, $\mu$ has strong exponential moment, then $I$ is given by the Legendre transform of $\Lambda$, i.e. $I(\alpha)= \sup_{\lambda \in \mathbb{R}}(\alpha \lambda -\Lambda(\lambda))$. The rest of this section is devoted to the proof of this theorem. We fix a basepoint $x \in X$ for the sequel. Existence of weak LDP for displacement functionals {#subsec.existence} -------------------------------------------------- The aim of this part is to prove the existence of weak LDP for displacement functionals without any moment assumption on the probability measure. This is done in the following Proposition \[prop.weak.ldp\]. Under the additional exponential moment assumption, the existence of a proper rate function as claimed in Theorem \[thm.LDP.precise\] will follow by standard techniques in large deviation theory; this will be shown in §\[subsec.proof.of.A\]. \[prop.weak.ldp\] Keep the assumptions of Theorem \[thm.LDP.precise\]. Then, the sequence $\frac{1}{n}\|L_n\|_x$ of random variables satisfies a large deviation principle with a rate function $I:[0,\infty) \to [0,\infty]$. The following technical result contains the crucial stability property that will be key in the proof of Theorem \[thm.LDP.norm\] together with the Lipschitz property of the displacement functionals. It says that if the averaged displacement functionals visit a certain neighborhood of a speed $\alpha$ with an exponential decay rate $\beta$ along some subsequence of times, then they will visit a slightly larger neighborhood of $\alpha$ with at most a slightly larger decay rate along periodic times. To ease the notation, we denote by $B(\alpha,r)$ the intersection of the interval $[\alpha-r,\alpha+r]$ with $[0,\infty)$. \[lemma.stability\] Let $\alpha \in [0,\infty)$. For every $\varepsilon,\delta>0$ and neighborhood $B(\alpha,r)$ of $\alpha$ in $[0,\infty)$, there exists $m_0 \in \mathbb{N}$ such that for every $k {\geqslant}1$, we have $$\label{eq.lemma.stability} - \frac{1}{km_0} \log \mathbb{P} \left(\frac{1}{km_0}\|L_{km_0}\|_x \in B(\alpha,r+\varepsilon)\right) -\delta {\leqslant}-\limsup_{n \to \infty} \frac{1}{n} \log \mathbb{P}\left(\frac{1}{n}\|L_n\|_x \in B(\alpha,r)\right)$$ For the proof of this lemma, we shall require the following key geometric ingredient. The reader interested in discrete supports only does not need to worry about neighborhoods, and just take $V_f=\{f\}$. \[lemma.alignment\] Given a non-elementary set $S$, there exist $k_S\in \mathbb N$, a constant $L>0$, a finite subset $\mathcal F\subseteq S^{k_S}$ and bounded neighborhoods $\{V_f\}_{f\in\mathcal F}$ of the elements of $\mathcal F$ so that the following holds. For every $n \in \mathbb{N}$ and elements $g_1,\ldots, g_n \in {\operatorname{Isom}}(X)$ with $\min_{i=1,\ldots,n} \|g_i\|_x {\geqslant}L$, there exist elements $f_i\in \mathcal F$ with the property that $$\tau(f'_ng_n f'_{n-1} g_{n-1} \ldots f'_1 g_1) {\geqslant}\sum_{i=1}^n \|g_i\|_x.$$ for every $f'_i \in V_{f_i}$. Regarding the neighborhoods, recall that we call a subset of $G$ bounded if it moves a given point of $X$ a bounded amount. This lemma parallels an observation of Benoist [@benoist.proper Proposition 6.4] concerning the coarse multiplicativity of spectral radius and norms of perturbations of products matrices based on Abels–Margulis–Soifer’s [@AMS] main result. In our geometric setting, for statements of somewhat similar flavor, see [@gouezel.local Lemma 2.4], [@mathieu-sisto Lemma 9.5] and [@delzant-steenbock Lemma 6.3]. The bound on the translation length will follow from the following well-known fact, several variations of which have appeared in the literature, see e.g. [@gromov:hyperbolic 7.2.C], [@ghys-delaharpe Théorème 5.16],[@MS:hyp-emb Proposition 14]. There exists $C=C(\delta)>0$ so that, given $h_i\in G$, $i=1,\dots, n$, satisfying $$\|h_i\|_x{\geqslant}(h_{i-1}^{-1}x\|h_ix)_x+ (h^{-1}_ix\|h_ix)_x +C,$$ where we set $h_0=h_n$ and $h_{n+1}=h_1$, we have $$\tau(h_1\dots h_n){\geqslant}\sum_{i=1}^n \|h_i\|_x - \sum_{i=1}^n (h_i^{-1}x\|h_{i+1}x)_x- nC.$$ It is possible to give a more constructive argument for the proof of the lemma, but we chose a probabilistic argument instead. First, notice that there is a finite subset $S'$ of $S$ so that there are independent loxodromic elements in the semigroup generated by $S'$. We consider the uniform probability distribution $\nu$ on $S'$, and the random walk driven by $\nu$/the convolution powers of $\nu$. Also, we denote the Gromov product at $z$ by $(\cdot\|\cdot)_z$. [Claim.]{} There exists $k$ so that for each $g\in G$ the following hold: - $\nu^{\ast k}(\{f: \|f\|_x{\geqslant}10C+1\})> .9,$ - $\nu^{\ast k}(\{f: (fx\|gx)_x,(f^{-1}x\| gx)_x {\leqslant}\|f\|_x/2 -C-1\})>.9,$ This is a combination of the facts that - the random walk driven by $\nu$ makes linear progress in $X$, meaning that with probability going to $1$ as $k$ goes to infinity it has a linear lower bound on $\|\cdot\|_x$ ([@maher-tiozzo Theorem 1.2]), and - the exponential decay of shadows estimate given by [@maher:exponential Lemma 2.10] (which, as observed [@maher-tiozzo Page 220], holds in our setting as well). We fix $k_S=k$ as in the claim and set $\mathcal F=(S')^k$. We can choose neighborhoods $V_f$, for $f\in \mathcal F$, so that $d(fx,f'x){\leqslant}1$ and $d(f^{-1}x,f'^{-1}x){\leqslant}1$ for all $f'\in F$. In particular, if $\|f\|_x{\geqslant}10C+1$ then $\|f'\|_x{\geqslant}10C+1$ for all $f\in V_f$, and similarly Gromov products vary by at most 1 when replacing $fx$ with $f'x$ or $f^{-1}x$ with $f'^{-1}x$. Finally, we choose $L=10k_S+10C$. In view of the claim, for each $i$ there is $f_i\in \mathcal F$ so that for any $f'_i\in V_{f_i}$ we have $$(f_i^{'-1}x\|g_ix)_x{\leqslant}\min\{\|g_i\|_x/2,\|f'_i\|_x/2\}-C,$$ and $$(g_{i+1}^{-1}x\|f'_ix)_x {\leqslant}\min\{\|g_{i+1}\|_x/2,\|f'_i\|_x/2\}-C,$$ where we set $g_{n+1}=g_1$. The required bound on $\tau(f'_n g_n \dots f'_1 g_1)$ now follows from the estimate given at the beginning of the proof (notice that the positive contributions of the $\|f'_i\|_x$ outweigh the negative contributions of the Gromov products and the “$-nC$”). When $\alpha=0$, follows by subadditivity of the displacement functional and i.i.d. property of the increments $Y_i$’s. So, we can suppose that $\alpha > 0$. It is also easy to see that it suffices to prove the statement for every $r>0$ small enough, so we suppose $0<r<\alpha$ and fix $\varepsilon>0$ with $\alpha>r+\varepsilon$. Consider the finite set $\mathcal{F}$, the neighborhoods $(V_f)_{f \in \mathcal{F}}$, and the constants $k_S,L>0$ as given by Lemma \[lemma.alignment\]. Let $\sup_{f' \in \tilde{V}} \|f'\|_x:=M_F \in \mathbb{R}$, where $\tilde{V}=\cup_{f \in \mathcal{F}} V_f$. By the property of $k_S \in \mathbb{N}$, the finite set $\mathcal{F}$ is in the support of $\mu^{\ast k_S}$ and we have $\mu^{\ast k_S}(V_f) >0$ for every $f \in \mathcal{F}$; set $\alpha_F:=\min_{f \in \mathcal{F}} \mu^{\ast k_S}(V_f)$. We have $\alpha_F>0$. For any $k \in \mathbb{N}$ and $k$-tuple $(g_k,\ldots,g_1) \in G^k$ with $\|g_i\|_x {\geqslant}L$ as in Lemma \[lemma.alignment\], using that lemma, choose elements $f(g_i,g_{i-1}) \in \mathcal{F}$ so that we have $$\label{eq.alignment} \|g_k f'(g_k,g_{k-1})g_{k-1}\ldots g_2 f'(g_2,g_1)g_1\|_x {\geqslant}\sum_{i=1}^k \|g_i\|_x$$ for every $f'(g_i,g_{i-1}) \in V_{f(g_i,g_{i-1})}$. Setting $\beta:= - \limsup_{n \to \infty} \frac{1}{n} \log \mathbb{P}(\frac{1}{n}\|L_n\|_x \in B(\alpha,r))$, fix $n_0 \in \mathbb{N}$ large enough so that $$\label{eq.conditions} \begin{aligned} \text{(1)} & \; \mathbb{P}\left(\frac{1}{n_0}\|L_{n_0}\|_x \in B(\alpha,r)\right) {\geqslant}e^{-n_0 (\beta +\delta/2)}, \hspace{5cm} {}\\ \text{(2)} & \; n_0(\alpha-r) {\geqslant}L, \hfill {}\\ \text{(3)} & \; n_0 {\geqslant}\frac{1}{\varepsilon}\max\{k_S\alpha_F,M_F\}, \hfill {}\\ \text{(4)} & \; n_0 {\geqslant}-\frac{2}{\delta} \log \alpha_F. \end{aligned}$$ Denote by $E_0:=\{g \in S^{n_0} \, \| \, \|g\|_x \in B(n_0\alpha,n_0r)\}$. Note that the item $(1)$ above is equivalent to $\mathbb{P}(L_{n_0} \in E_0){\geqslant}e^{-n_0 (\beta +\delta/2)}$. To make the notation lighter, for $n {\geqslant}m$, denote $L_m^n=Y_n \ldots Y_m$. Now for every $k {\geqslant}1$ and for $1 {\leqslant}i {\leqslant}k$, we set $$\begin{aligned} \Omega_k: & =\left \{\frac{1}{kn_0 +(k-1)k_S}\|L_{kn_0 +(k-1)k_S}\|_x \in B(\alpha,r+\varepsilon) \right \},\\ \hat{E}_i: & =\{L^{(i+1)n_0 +ik_S}_{i(n_0+k_S)} \in E_0\},\\ \hat{F}_i: & =\{L^{i(n_0+k_S)}_{in_0+(i-1)k_S+1} \in V_{f}, \; \text{where}, \; f=f(L^{(i+1)n_0 +ik_S}_{i(n_0+k_S)},L^{in_0 +(i-1)k_S}_{(i-1)(n_0+k_S)})\}. \end{aligned}$$ We described these events because when all events $\hat{E}_i$ and $\hat{F}_i$ occur, then Lemma \[lemma.alignment\] applies. The following inclusion of events holds true in view of subadditivity of displacement functionals, and the choice of $n_0 \in \mathbb{N}$ (namely items (2) and (3) above): $$\Omega_k \supseteq \bigcap_{i=0}^{k-1} \hat{E}_i \cap \bigcap_{i=1}^{k-1} \hat{F}_i$$ As a consequence denoting by $\mathcal{F}_{\hat{E}} $ the $\sigma$-algebra generated by the random variables used in the events $\hat{E}_i$’s, we obtain $\mathbb{P}$-a.s.$$\label{eq.conditional} \mathbb{P}(\Omega_k) {\geqslant}\mathbb{E}[\mathbb{E}[ \Pi_{i=0}^{k-1} 1_{\hat{E}_i} \Pi_{i=1}^{k-1} 1_{\hat{F}_i} \, \| \, \mathcal{F}_{\hat{E}}]] =\mathbb{E}[\Pi_{i=0}^{k-1} 1_{\hat{E}_i} \mathbb{E}[ \Pi_{i=1}^{k-1} 1_{\hat{F}_i} \, \| \, \mathcal{F}_{\hat{E}}]] $$ On the other hand, by definition of the events $\hat{F}_i$’s and $\alpha_F>0$, we readily have $\mathbb{P}$-a.s. $\mathbb{E}[ \Pi_{i=1}^{k-1} 1_{\hat{F}_i} \, \| \, \mathcal{F}_{\hat{E}}] {\geqslant}\alpha_F^{k-1}$. Using this lower bound and the fact that $Y_i$’s are i.i.d. we deduce from that we have $$\mathbb{P}(\Omega_k) {\geqslant}\mathbb{P}(\hat{E}_0)^k \alpha_F^{k}$$ Now using items $(1)$ and $(4)$ above, we obtain that $$\label{eq.last.in.lemma} \frac{1}{kn_0 +(k-1)k_S}\log \mathbb{P}(\Omega_k) {\geqslant}-\beta - \delta$$ The statement follows by setting $m_0=n_0+k_S$. The following lemma crucially relies on the Lipschitz property of displacement functionals and it says that the exponential rate of decay of probabilities of events concerning the asymptotic behaviour of displacement functionals along periodic times matches the decay rate considered along all times. \[lemma.fill.the.gaps\] For every $m {\geqslant}1$, $\alpha,r {\geqslant}0$, $\varepsilon>0$, we have $$-\liminf_{k \to \infty} \frac{1}{mk} \log \mathbb{P}\left(\frac{1}{mk} \|L_{mk}\|_x \in B(\alpha,r)\right) {\geqslant}- \liminf_{n \to \infty} \frac{1}{n} \log \mathbb{P}\left(\frac{1}{n}\|L_n\|_x \in B(\alpha,r+\varepsilon)\right).$$ Fix a a subset $K$ of $G$ such that for every $i=1,\ldots,m$, we have $\mu^{\ast i}(K) {\geqslant}\frac{1}{2}$ and set $\max_{g \in K} \|g\|_x=:C_K \in \mathbb{R}$. This is possible since the sequence of increasing closed sets $\{g \in G \, \| \, \|g\|_x {\leqslant}n \}$ covers $G$. Setting $\beta:= -\liminf_{k \to \infty}\frac{1}{mk} \log \mathbb{P}(\frac{1}{mk}\|L_{mk}\|_x \in B(\alpha,r))$, for $\delta>0$, choose $k_0$ large enough so that for every $k {\geqslant}k_0$, we have $-\frac{1}{mk}\log \mathbb{P}(\frac{1}{mk}\|L_{mk}\|_x \in B(\alpha,r)) {\geqslant}\beta - \delta$. The following inclusion of events follows from the Lipschitz property of displacement functionals and the choice of $k_0$: for every $n {\geqslant}mk_0$, denoting by $k_n$ the integer satisfying $m(k_n+1) > n {\geqslant}mk_n$, we have $$\left\{\frac{1}{n} \|L_n\|_x \in B(\alpha,r+\varepsilon)\right\} \supseteq \left\{\frac{1}{mk_n}\|L_{mk_n}\|_x \in B(\alpha,r)\right\} \cap \left\{Y_n \ldots Y_{mk_n+1} \in K\right\}.$$ Evaluating the probabilities and using the independence of $Y_i$’s, we get $$\label{eq.gaps} \mathbb{P}\left(\frac{1}{n} \|L_n\|_x \in B(\alpha,r+\varepsilon)\right) {\geqslant}\mathbb{P}\left(\frac{1}{mk_n}\|L_{mk_n}\|_x \in B(\alpha,r)\right) \mathbb{P}(Y_n \ldots Y_{mk_n+1} \in K)$$ Since $Y_i$’s are identically distributed, we have $\mathbb{P}(Y_n \ldots Y_{mk_n+1} \in K) {\geqslant}\frac{1}{2}$. Taking the logarithm, dividing my $n$ and passing to $\liminf$ on both sides of , the claim follows since $\delta>0$ is arbitrary. In the proof of Theorem \[thm.LDP.norm\], we shall make use of the following general criterion (see also the subsequent remark) for the existence of a LDP: \[thm.LDP.criterion\] Let $Y$ be a topological space endowed with its Borel $\sigma$-algebra $\beta_{Y}$, and $Z_{n}$ be a sequence of $Y$-valued random variables. Denote by $\mu_{n}$ the distribution of $Z_{n}$. Let $\mathcal{A}$ be a base of open sets for the topology of $Y$. For each $\alpha \in Y$, define: $$I_{li}(\alpha):= \underset{\underset{\alpha \in A}{A \in \mathcal{A}}} {\sup} - \underset{n \rightarrow \infty}{\liminf} \frac{1}{n}\log \mu_{n}(A) \quad \text{and} \quad I_{ls}(\alpha):= \underset{\underset{\alpha \in A}{A \in \mathcal{A}}}{\sup} - \underset{n \rightarrow \infty} {\limsup} \frac{1}{n}\log \mu_{n}(A)$$ Suppose that for all $\alpha \in Y$, we have $I_{li}(\alpha)=I_{ls}(\alpha)$. Then, the sequence $Z_{n}$ satisfies a LDP with rate function $I$ given by $I(\alpha):=I_{li}(\alpha)=I_{ls}(\alpha)$. \[rk.ldp.implies\] In our setting where $Y=[0,\infty)$, the hypothesis of the previous theorem is actually equivalent to the existence of a weak LDP (see [@dembo-zeitouni]). In particular, one readily sees that a rate function of a LDP is lower semi-continuous. We are now ready to elucidate To prove the statement, taking $Y=[0,\infty)$, $Z_n$ to be the random variables $\frac{1}{n} \|L_n\|_x$ in Theorem \[thm.LDP.criterion\], we need (see Remark \[rk.ldp.implies\]) to show that $I_{li}(\alpha)=I_{ls}(\alpha)$ for every $\alpha \in [0,\infty)$. Suppose this is not the case for some $\alpha \in [0,\infty)$. This means that there exists $r_1>0$ such that for every $r_1>r_0>0$, we have $$\beta_1:=-\liminf_{n \to \infty} \frac{1}{n} \log \mathbb{P}\left(\frac{1}{n} \|L_n\|_x \in B(\alpha,r_1)\right)>-\limsup_{n \to \infty} \frac{1}{n} \log \mathbb{P}\left(\frac{1}{n}\|L_n\|_x \in B(\alpha,r_0)\right)=:\beta_2.$$ Fix $r_0 \in (0,r_1)$. Applying Lemma \[lemma.stability\] with $r=r_0$, some $\varepsilon \in (0,\frac{r_1-r_0}{2})$ and $\delta \in (0, \beta_1-\beta_2)$, we deduce that for some $m_0 \in \mathbb{N}$ and for every $k {\geqslant}1$, we have $$\beta_2>-\frac{1}{km_0} \log \mathbb{P}\left(\frac{1}{km_0} \|L_{km_0}\|_x \in B(\alpha,r_0+\varepsilon)\right) -\delta.$$ Now applying Lemma \[lemma.fill.the.gaps\] with $m=m_0$ and $r=r_0+\varepsilon$, since $r_0+2\varepsilon<r_1$, we deduce that $$\beta_2 {\geqslant}-\liminf_{n \to \infty} \frac{1}{n} \log \mathbb{P}\left(\frac{1}{n}\|L_n\|_x \in B(\alpha,r_{0}+2 \varepsilon)\right)-\delta {\geqslant}\beta_1-\delta,$$ in contradiction with the choice of $\delta$. This proves the proposition. Convexity of the rate function {#subsec.convexity} ------------------------------ Here we prove \[prop.convexity\] Keep the assumptions of Theorem \[thm.LDP.precise\]. The rate function $I:[0,\infty) \to [0,\infty]$ given by Proposition \[prop.weak.ldp\] is convex. Recall that it readily follows from the expression of the rate function given by Theorem \[thm.LDP.criterion\] that $I$ is lower semi-continuous (see Remark \[rk.ldp.implies\]). The previous result shows in particular that $I$ is continuous in the interior of its effective support. The proof uses the expression of the rate function $I$ as in Theorem \[thm.LDP.criterion\] (see also Remark \[rk.ldp.implies\]) and the Unfolding Lemma (Lemma \[lemma.alignment\]) in a crucial way. Since the rate function $I:[0,\infty) \to [0,\infty]$ is lower semi-continuous, to show that it is convex, it suffices to show that for every $\alpha_1,\alpha_2 \in [0,\infty)$ and $\varepsilon,\delta>0$, we have $$\label{eq.to.convex} I\left(\frac{\alpha_1+\alpha_2}{2}\right)-\delta{\leqslant}\left(\frac{1}{2}+\varepsilon\right)(I(\alpha_1)+I(\alpha_2)).$$ So, let $\alpha_2 >\alpha_1 \in [0,\infty)$, $\delta>0$ be given. We will suppose that $\alpha_1 \neq 0$; this case can be treated similarly as below. Fix some $0<r< \min\{\alpha_1,\alpha_2\}$ and some positive constant $\varepsilon < \min\{ \frac{r}{\alpha_2-\alpha_1}, 1\}$. Let $\mathcal{F}$ be the finite set, $(V_f)_{f \in \mathcal{F}}$ be the neighborhoods and $k_S, L>0$ be the constants given by Lemma \[lemma.alignment\]. Set $\max_{f' \in \hat{V}} \|f'\|_x:=M_F \in \mathbb{R}$, where $\hat{V}=\cup_{f \in \mathcal{F}} V_f$. By Lemma \[lemma.alignment\], we have $\mu^{\ast k_S}(V_f) >0$ for every $f \in \mathcal{F}$; set $\alpha_F:=\min_{f \in \mathcal{F}} \mu^{\ast k_S}(V_f)$. We have $\alpha_F>0$. Given any $k$-tuple $(g_k,\ldots,g_1) \in G^k$ with $\|g_i\|_x {\geqslant}L$ as in Lemma \[lemma.alignment\], using that lemma, choose elements $f(g_i,g_{i-1}) \in F$ so that holds. In view of Remark \[rk.ldp.implies\], by Proposition \[prop.weak.ldp\] and Theorem \[thm.LDP.criterion\], there exists $n_0 \in \mathbb{N}$ large enough so that for $j=1,2$ $$\label{eq.n0.so.that} e^{-n_0(I(\alpha_j)+\delta/2)} {\leqslant}\mathbb{P}\left(\frac{1}{n_0}\|L_{n_0}\|_x \in B(\alpha_j,r)\right),$$ and 1. $n_0(\alpha_1-r){\geqslant}L$, 2. $n_0 {\geqslant}2\frac{1}{r}\max\{M_F, k_S(\alpha_1+\alpha_2)\}$, 3. $n_0 {\geqslant}-\frac{2}{\delta}\log \alpha_F$. For $j=1,2$, let us denote by $E_j$ the subset of $G$ given by $\{g \in S^{n_0} \, \| \, \frac{1}{n_0} \|g\|_x \in B(\alpha_j,r) \}$. For $i$ an integer, denote $j(i)=1$ if $i$ is odd and $j(i)=2$ if $i$ is even. For $n {\geqslant}m$, using the same notation $L_m^n=Y_n \ldots Y_m$, as before, consider the following events for every $k {\geqslant}1$ and $1 {\leqslant}i {\leqslant}k$: $$\begin{aligned} \Omega_k:&=\left \{\frac{1}{kn_0 +(k-1)k_S}\|L_{kn_0 +(k-1)k_S}\|_x \in B\left(\frac{\alpha_1+\alpha_2}{2},2r\right) \right \},\\ \hat{E}_{i}:&=\{L^{(i+1)n_0 +ik_S}_{i(n_0+k_S)} \in E_{j(i)}\},\\ \hat{F}_i:&=\{L^{i(n_0+k_S)}_{in_0+(i-1)k_S+1} \in V_f \; \text{where} \, f=f(L^{(i+1)n_0 +ik_S}_{i(n_0+k_S)},L^{in_0 +(i-1)k_S}_{(i-1)(n_0+k_S)})\}. \end{aligned}$$ Using subadditivity of displacement functionals, , it is not hard to verify that the following inclusion of events holds true for every $k {\geqslant}1$, thanks to the choice of $n_0$ (i.e. satisfying (1) and (2) above): $$\label{eq.inclusion.events.convexity} \Omega_k \supseteq \bigcap_{i=1}^{k-1} \hat{F}_i \cap \bigcap_{i=1}^k \hat{E}_{i}.$$ Now let us denote by $\mathcal{F}_{\hat{E}} $ the $\sigma$-algebra generated by the random variables $Y_i$’s for $i=1,\ldots,kn_0+(k-1)k_S$ except those $i \in \cup_{t=1}^{k-1} \{t(n_0+k_S),\ldots,tn_0+(t-1)k_S +1\}$. Using independence of $Y_i$’s, evaluating the probabilities on both sides of we obtain $\mathbb{P}$-a.s. $$\label{eq.convexity.1} \mathbb{P}(\Omega_k) {\geqslant}\prod_{i=1}^k \mathbb{E}[ 1_{\hat{E}_{i}} \mathbb{E}[\Pi_{i=1}^{k-1} 1_{\hat{F}_i} \, \| \, \mathcal{F}_{\hat{E}}]].$$ As in the proof of Proposition \[prop.weak.ldp\], using the $\mathbb{P}$-a.s. bound $\mathbb{E}[ \Pi_{i=1}^{k-1} 1_{\hat{F}_i} \, \| \, \mathcal{F}_{\hat{E}}] {\geqslant}\alpha_F^{k-1}$, reads $$\label{eq.convexity.2} \mathbb{P}(\Omega_k) {\geqslant}\alpha_F^{k-1} \prod_{i=1}^k \mathbb{P}( \hat{E}_{i})$$ Using the fact that $Y_i$’s are identically distributed, $\eqref{eq.n0.so.that}$ implies that for every $i$, we have $\mathbb{P}(\hat{E}_{i})=\mathbb{P}(L_{n_0} \in E_{j(i)}) {\geqslant}e^{-n_0(I(\alpha_{j(i)})+\delta/2)}$, so that reads $$\mathbb{P}(\Omega_k) {\geqslant}\alpha_F^{k-1} e^{-n_0\|j^{-1}(1)\|(I(\alpha_{1})+\delta/2)}e^{-n_0\|j^{-1}(2)\|(I(\alpha_{2})+\delta/2)}.$$ Note that $\|j^{-1}(i)\|{\leqslant}k/2+1$. By (3) in the choice of $n_0$ above, we deduce that for every $k$ large enough, we have $$\frac{1}{kn_0 +(k-1)k_S}\log \mathbb{P}(\Omega_k) {\geqslant}-\delta-\frac{1}{2}(I(\alpha_1)+I(\alpha_2))-\varepsilon(I(\alpha_1)+I(\alpha_2)).$$ Now applying Lemma \[lemma.fill.the.gaps\] with $m=n_0+k_S$, $\alpha=\frac{\alpha_1+\alpha_2}{2}$ and $2r$, and using the expression of $I$ given by Theorem \[thm.LDP.criterion\], we deduce $$I\left(\frac{\alpha_1+\alpha_2}{2}\right)-\delta{\leqslant}\left(\frac{1}{2}+\varepsilon\right)(I(\alpha_1)+I(\alpha_2)).$$ This shows and proves the proposition. For $j=1,2$, let us denote by $E_j$ the subset of $G$ given by $\{g \in S^{n_0} \, \| \, \frac{1}{n_0} \|g\|_x \in B(\alpha_j,r) \}$. For $n {\geqslant}m$, using the same notation $L_m^n=Y_n \ldots Y_m$, as before, consider the following events for every $k {\geqslant}1$ and $1 {\leqslant}i {\leqslant}k$: $$\begin{aligned} \Omega_k:&=\left \{\frac{1}{kn_0 +(k-1)k_S}\|L_{kn_0 +(k-1)k_S}\|_x \in B\left(\frac{\alpha_1+\alpha_2}{2},2r\right) \right \},\\ \hat{E}_{i,j}:&=\{L^{(i+1)n_0 +ik_S}_{i(n_0+k_S)} \in E_j\} \quad \text{for} \, j=1,2,\\ \hat{F}_i:&=\{L^{i(n_0+k_S)}_{in_0+(i-1)k_S+1} \in V_f \; \text{where} \, f=f(L^{(i+1)n_0 +ik_S}_{i(n_0+k_S)},L^{in_0 +(i-1)k_S}_{(i-1)(n_0+k_S)})\}. \end{aligned}$$ Using subadditivity of displacement functionals, , it is not hard to verify that the following inclusion of events holds true for every $k {\geqslant}1$, thanks to the choice of $n_0$ (i.e. satisfying (1) and (2) above): $$\label{eq.inclusion.events.convexity} \Omega_k \supseteq \bigcap_{i=1}^{k-1} \hat{F}_i \cap \left(\bigcup_{\underset{\|\|f^{-1}(1)\|-\|f^{-1}(2)\|\|<k\varepsilon}{f \in \{1,2\}^{\{1,\ldots,k\}}}}\left(\cap_{i=1}^k \hat{E}_{i,f(i)} \right)\right),$$ where for $j=1,2$, $\|f^{-1}(j)\|$ denotes the cardinality of $i$’s in $\{1,\ldots,k\}$ satisfying $f(i)=j$. Now let us denote by $\mathcal{F}_{\hat{E}} $ the $\sigma$-algebra generated by the random variables $Y_i$’s for $i=1,\ldots,kn_0+(k-1)k_S$ except those $i \in \cup_{t=1}^{k-1} \{t(n_0+k_S),\ldots,tn_0+(t-1)k_S +1\}$. Since the events $\cap_{i=1}^k \hat{E}_{i,f(i)}$ corresponding to different $f \in \{1,2\}^{\{1,\ldots,k\}}$ are disjoint, evaluating the probabilities on both sides of , using independence of $Y_i$’s, we obtain $\mathbb{P}$-a.s. $$\label{eq.convexity.1} \mathbb{P}(\Omega_k) {\geqslant}\sum_{\underset{\|\|f^{-1}(1)\|-\|f^{-1}(2)\|\|<k\varepsilon}{f \in \{1,2\}^{\{1,\ldots,k\}}}} \prod_{i=1}^k \mathbb{E}[ 1_{\hat{E}_{i,f(i)}} \mathbb{E}[\Pi_{i=1}^{k-1} 1_{\hat{F}_i} \, \| \, \mathcal{F}_{\hat{E}}]].$$ As in the proof of Proposition \[prop.weak.ldp\], using the $\mathbb{P}$-a.s. bound $\mathbb{E}[ \Pi_{i=1}^{k-1} 1_{\hat{F}_i} \, \| \, \mathcal{F}_{\hat{E}}] {\geqslant}\alpha_F^{k-1}$, reads $$\label{eq.convexity.2} \mathbb{P}(\Omega_k) {\geqslant}\alpha_F^{k-1} \sum_{\underset{\|\|f^{-1}(1)\|-\|f^{-1}(2)\|\|<k\varepsilon}{f \in \{1,2\}^{\{1,\ldots,k\}}}} \prod_{i=1}^k \mathbb{P}( \hat{E}_{i,f(i)})$$ Using the fact that $Y_i$’s are identically distributed, $\eqref{eq.n0.so.that}$ implies that for every $f$ and $i$, we have $\mathbb{P}(\hat{E}_{i,f(i)})=\mathbb{P}(L_{n_0} \in E_{f(i)}) {\geqslant}e^{-n_0(I(\alpha_{f(i)})+\delta/2)}$, so that reads $$\mathbb{P}(\Omega_k) {\geqslant}\alpha_F^{k-1} \sum_{\underset{\|\|f^{-1}(1)\|-\|f^{-1}(2)\|\|<k\varepsilon}{f \in \{1,2\}^{\{1,\ldots,k\}}}} e^{-n_0\|f^{-1}(1)\|(I(\alpha_{1})+\delta/2)}e^{-n_0\|f^{-1}(2)\|(I(\alpha_{2})+\delta/2)}$$ Using the crude lower bound (enough for our purposes) given by a single summand in the sum on the right-hand-side of the previous equation, by (3) in the choice of $n_0$ above, we deduce that for every $k {\geqslant}1$, we have $$\frac{1}{kn_0 +(k-1)k_S}\log \mathbb{P}(\Omega_k) {\geqslant}-\delta-\frac{1}{2}(I(\alpha_1)+I(\alpha_2))-\varepsilon(I(\alpha_1)+I(\alpha_2)).$$ Now applying Lemma \[lemma.fill.the.gaps\] with $m=n_0+k_S$, $\alpha=\frac{\alpha_1+\alpha_2}{2}$ and $2r$, and using the expression of $I$ given by Theorem \[thm.LDP.criterion\], we deduce $$I\left(\frac{\alpha_1+\alpha_2}{2}\right)-\delta{\leqslant}\left(\frac{1}{2}+\varepsilon\right)(I(\alpha_1)+I(\alpha_2)).$$ This shows and proves the proposition. Proof of Theorem \[thm.LDP.precise\] {#subsec.proof.of.A} ------------------------------------ The rest of this section is devoted to completing the proof of Theorem \[thm.LDP.precise\]. It remains to show that the (full) LDP holds under a finite exponential moment condition and that we can give an alternative expression for the rate function under a strong exponential moment condition. These will follow using rather standard techniques in large deviation theory. We spell out the details for reader’s convenience. ### Existence of (full) LDP under exponential moment condition {#subsec.upgrade.to.full.LDP} The following classical notion of large deviations theory allows one to formulate a sufficient condition (see Lemma \[corolLDP\]) to strengthen a weak LDP to a LDP with proper rate function: \[defexptight\] A sequence of random variables $Z_{n}$ on a topological space $Y$ is said to be exponentially tight, if for all $\alpha \in \mathbb{R}$, there exists a compact set $K_{\alpha} \subset Y $ such that $\underset{n \rightarrow \infty}{\limsup} \frac{1}{n} \log \mathbb{P}(Z_{n} \in K_{\alpha}^{c})<-\alpha$. With this definition, we have \[corolLDP\] If an exponentially tight sequence of random variables on $X$ satisfies a weak LDP with a rate function $I$, then it satisfies a (full) LDP with a proper rate function $I$. Therefore, by Proposition \[prop.weak.ldp\], to prove the full LDP and properness of $I$ in Theorem \[thm.LDP.precise\], we only need to show that a finite exponential moment condition on $\mu$ implies that the sequence $\frac{1}{n}\|L_n\|_x$ of random variables is exponentially tight. \[lemma.expmoment.implies.exptight\] If $\mu$ has a finite exponential moment, then the sequence random variables $\frac{1}{n}\|L_n\|_x$ is exponentially tight. It suffices to show that $$\lim_{t \rightarrow \infty} \limsup_{n \rightarrow \infty} \frac{1}{n} \log \mathbb{P}\left(\frac{1}{n}\|L_n\|_x{\geqslant}t\right)=-\infty$$ By Chebyshev inequality, for every $s {\geqslant}0$, we have $$\mathbb{P}( \|L_n\|_x {\geqslant}tn) {\leqslant}\mathbb{E}[e^{s \|L_n\|_x}] e^{-stn}$$ In this inequality, taking $\log$, dividing by $n$ and specializing to some $s_{0} \in \mathbb{R}$ such that $c {\geqslant}s_{0}>0$, we get $$\frac{1}{n} \log \mathbb{P}(\|L_n\|_x {\geqslant}tn) {\leqslant}-(s_{0}t - \frac{1}{n}\log \mathbb{E}[e^{s_{0} \|L_n\|_x}])$$ On the other hand, it follows by the independence of random walk increments and the subadditivity of the displacement functional $\|.\|_x$ that for all $n {\geqslant}1$, we have $\frac{1}{n} \log \mathbb{E}[e^{s_{0} \|L_n\|_x}] {\leqslant}\log \mathbb{E}[e^{s_{0} \|Y_{1}\|_x}]$. Therefore, we obtain $$\limsup_{n \rightarrow \infty} \frac{1}{n} \log \mathbb{P}\left(\frac{1}{n}\|L_n\|_x{\geqslant}t\right) {\leqslant}-(s_{0}t-\mathbb{E}[e^{s_{0}\|Y_{1}\|_x}])$$ Since $\mathbb{E}[e^{s_{0}\|Y_{1}\|_x}]$ is finite by the exponential moment condition, the result follows by taking limit in both sides as $t$ goes to $+\infty$. ### Identification of the rate function {#subsec.identification} Using Fenchel–Moreau duality and Varadhan’s integral lemma, we now give an alternative expression for the rate function $I$ as the Legendre transform of a limit Laplace transform of the distributions of $\|L_n\|_x$. We now complete the It is easy to see that if $\mu$ has a strong exponential moment, then we have $\Lambda(\lambda)<\infty$ for all $\lambda \in \mathbb{R}$. Hence, it follows from Varadhan’s integral lemma (see [@dembo-zeitouni] section 4.3) that in fact for all $\lambda \in \mathbb{R}$, one has $$\Lambda(\lambda)=\lim_{n} \frac{1}{n} \log \mathbb{E}[e^{\lambda \|L_n\|_x}] =\underset{ \alpha \in \mathbb{R}}{\sup}(\lambda \alpha-I(\alpha))$$ where $I$ is the proper rate function of the LDP. Now, for a function $f$ on $\mathbb{R}$, denote its convex conjugate (Legendre tranform) by $f^{\ast}(.)$, where $f^{\ast}(\lambda):= \sup_{\alpha \in \mathbb{R}}(\lambda \alpha -f( \alpha)) $. The above conclusion of Varadhan’s integral lemma hence reads as $\Lambda(\lambda)=I^{\ast}(\lambda)$. Now, since $I$ is a convex rate function, Fenchel-Moreau duality tells us that $I(\alpha)=I^{\ast \ast}(\alpha)=\Lambda^{\ast}(\alpha)$, identifying $I(\alpha)$ with $\Lambda^{\ast}(\alpha)$ and completing the proof. The value of the rate function at zero {#subsec.doubling} -------------------------------------- ${}$ To end this section, here we briefly spell out the details of 3. of Remark \[rk.affter.thm.LDP.norm\]. We shall say that a proper metric space $(X,d)$ has at most exponential doubling property if there exists a constant $\alpha>1$ such that for every $R>0$ large enough, any ball of radius $2R$ can be covered with at most $\alpha^R$ balls of radius $R$. This condition is clearly satisfied for every finite dimensional Riemannian manifold with pinched negative curvature. Let $(X,d)$ be a hyperbolic space with at most exponential doubling property. Let $\mu$ be a finitely supported symmetric non-elementary probability measure on ${\operatorname{Isom}}(X)$ and suppose that the group $G$ generated by the support of $\mu$ acts properly discontinuously on $X$. Let $I$ be the rate function given by Theorem \[thm.LDP.norm\]. Then, we have $I(0)=-\log r_\mu$, where $r_\mu$ is the spectral radius of the random walk. Note since $\mu$ is non-elementary, its support generates a non-amenable countable subgroup of ${\operatorname{Isom}}(X)$ and by Kesten’s criterion [@kesten.banach.mean], we have $I(0)>0$. By applying at most exponential doubling property, we find that for some $R>0$ and $x \in X$, we have $\|\{g \in G \,\|\, gx \in B(x, nR) \}\|=O(\alpha^{nR})$. Denote by $\lambda(\mu)$ the self-adjoint operator on $\ell^2(G)$ defined by $\lambda(\mu)f(x)=\sum_\gamma f(g^{-1}x) \mu(g)$. We have $\lambda(\mu^{\ast n})=\lambda(\mu)^n$ and by Cauchy-Schwarz inequality, $\mu^{\ast 2n }(g)=<\lambda(\mu)^{2n}\delta_g,\delta_e>{\leqslant}\|\|\lambda(\mu)\delta_g\|\|.\|\|\lambda(\mu) \delta_2\|\|=\mu^{\ast 2n}(e)$. Now using this and Kesten’s formula, we deduce that for every $\varepsilon>0$, $ \log r_\mu {\leqslant}\limsup_{n \to \infty} \frac{1}{2n} \log \mathbb{P}(\|L_{2n}\|_{x} {\leqslant}2nR \varepsilon) {\leqslant}\log r_\mu + R \varepsilon$, which shows that for the rate function $I$ given by Theorem \[thm.LDP.norm\], we have $I(0)=-\log r_\mu$. It is easy to see that this argument also gives a lower bound to the drift of the random walk. Such relations between drift, spectral radius and entropy of the random walk, and growth rate of the group are well-known [@avez; @guivarch.fundamental]. Large deviation principle for translation distance {#sec.tau} ================================================== This section is devoted to the proof of Theorem \[Thm.LDP.translation\]. The proof will be carried out in three parts consecutively in §\[subsec.translation.distance.as.likely\],§\[subsec.translation.functional.as.likely\], and §\[subsec.deal.with.zero\]. In the first part (§\[subsec.ingredients.tau\]), we start with developing the necessary geometric ingredients. Geometric ingredients {#subsec.ingredients.tau} --------------------- Let $(X,d)$ be a hyperbolic space, with a fixed basepoint $x\in X$, and let $S$ be an arbitrary bounded subset of ${\operatorname{Isom}}(X)$. Some of the results in this subsection might be known to experts, but we provide detailed proofs for completeness. Here, the key results that will be used for the probabilistic arguments are Lemmas \[lemma.getting.norm.close.to.tau\] and \[lemma.moving.tau\]. We start with some preliminary results to prepare the proofs of those lemmas. For convenience, in the proofs below we will assume that $d(x,sx){\leqslant}1$ for each $s\in S$. This can be achieved by rescaling $X$, and it is readily seen that all the statements hold for $X$ if and only if they hold for a rescaling of $X$, up to changing the constants. \[lem:cant\_avoid\_geod\] For every $\epsilon>0$ there exist $D_0,N{\geqslant}1$ so that the following holds. Let $s_1,\dots,s_n\in S$, for some $n{\geqslant}N$, and let $g_i=s_1\dots s_i$. Let $\gamma$ be a subpath of length ${\geqslant}\epsilon n$ of a geodesic from $x$ to $g_nx$. Then there exists $i$ with $d(g_ix,\gamma){\leqslant}D_0$. This can be deduced from [@HS:coarse Claim 2 within Lemma 2.6], which in our setting says the following. There exist $\epsilon_0>0$ and $D'>0$ (independent of $x$ and $g_n$) so that, given disjoint balls $B_1,\dots,B_k$ of radius $D{\geqslant}D'$ centered on $\gamma$, any path $\alpha$ from $x$ to $g_nx$ that avoids all $B_i$ satisfies $l(\alpha){\geqslant}k(1+\epsilon_0)^D$. Choose $D_0{\geqslant}D'+1$ so that $(1+\epsilon_0)^{D_0-1}>3D_0/\epsilon$. Also, we let $N{\geqslant}6D_0/\epsilon$, and check that these choices work. In the setting of the statement, suppose by contradiction that we have $d(g_ix,\gamma)> D_0$ for all $i$. Then we can find at least $k{\geqslant}\epsilon n/(2D_0)-1{\geqslant}\epsilon n/(3D_0)$ disjoint balls $B_i$ of radius $D_0-1$ centered on $\gamma$ so that the path $\alpha$ in $X$ obtained concatenating geodesics from $s_ix$ to $s_{i+1}x$ avoids all $B_i$. The length of $\alpha$ is at most $n$, so we obtain: $$n{\geqslant}\frac{\epsilon n}{3D_0} (1+\epsilon_0)^{D_0-1}>n,$$ a contradiction. Let $\delta{\geqslant}1$ be a hyperbolicity constant for $X$. For $g\in G$, define $$Min(g)=\{z\in X: d(z,gz){\leqslant}\tau(g)+4\delta\}.$$ Also, for $z\in Z$, denote by $\pi^g(z)$ a point in $Min(g)$ so that $d(z,\pi^g(z)){\leqslant}d(z,Min(g))+1$. (That is, $\pi^g$ is coarsely the closest-point projection to $Min(g)$.) We can and will assume that $g\pi^g(z)=\pi^g(gz)$ holds for all $g$ and $z$. It is known that $Min(g)$ is quasiconvex (see e.g. [@delzant-gromov Proposition 2.3.3] and [@coulon Proposition 2.28]), but we will only need the following special case of quasiconvexity, which has a very short proof: \[lem:Min\_qconv\] If $y\in Min(g)$, then any point on any geodesic from $y$ to $gy$ is also contained in $Min(g)$. First, observe that given $y\in X$ and a geodesic $[y,gy]$, any $z\in [y,gy]$ has $$d(z,gz){\leqslant}d(z,gy)+d(gy,gz)=d(z,gy)+d(y,z)=d(y,gy).$$ The desired statement easily follows. We now show that geodesics from $z$ to $gz$ pass close to the projection points of the endpoints onto $Min(g)$. \[lem:far\_from\_min\] There exists $D_1{\geqslant}0$ so that the following holds. For every $g\in G$ and $z\in X$, we have that any geodesic $\gamma$ from $z$ to $gz$ passes $D_1$-close to $\pi^g(z)$ and $\pi^g(gz)$. Moreover, we have $$d(z,gz){\geqslant}2d(z,\pi^g(z))+\tau(g)-D_1.$$ Consider any geodesic $\gamma$ from $z$ to $gz$. We will show that $\gamma$ passes $(4\delta+2)$-close to $\pi^g(z)$, the argument for $\pi^g(gz)$ being similar. We will use $2\delta$-thinness of a quadrangle with vertices $z,\pi^g(z),\pi^g(gz),gz$. Suppose by contradiction that $\gamma$ does not pass $(4\delta+2)$-close to $\pi^g(z)$. Consider the point $z'$ on a geodesic from $z$ to $\pi^g(z)$ at distance $2\delta+2$ from $\pi^g(z)$. We observe that $z'$ cannot be $2\delta$-close to any geodesic $[\pi^g(z),\pi^g(gz)]$, for otherwise there would be a point $q$ on said geodesic, whence on $Min(g)$ by Lemma \[lem:Min\_qconv\], which satisfies $d(z,q)<d(z,\pi^g(z))-1$, contradicting the defining property of $\pi^g$. Also, $z'$ cannot be $2\delta$-close to $\gamma$ by hypothesis, so $z'$ is $2\delta$-close to $g[z,\pi^g(z)]$. But then it must be $4\delta$-close to the point on that geodesic at distance $4\delta+2$ from $\pi^g(gz)$, this point being $gz'$. We just showed $d(z',gz'){\leqslant}4\delta$, which implies $z'\in Min(g)$. But $d(z,z')<d(z,\pi^g(z))-1$, contradicting the defining property of $\pi^g(z)$. Now, the fact that $\gamma$ passes $(4\delta+2)$-close to $\pi^g(z)$ and $\pi^g(gz)$ implies the following inequality: $$d(z,gz){\geqslant}d(z,\pi^g(z))+d(\pi^g(z),\pi^g(gz))+d(\pi^g(gz),gz)-4(4\delta+2).$$ The first and third terms on the right-hand side are both equal to $d(z,\pi^g(z))$, while the second term is at least $\tau(g)$. Therefore, we can conclude by setting $D_1=4(4\delta+2)$. We are now ready to prove the following key lemma which will be instrumental in showing that for a fixed speed $\alpha>0$, up to multiplying its probability by the linear factor $n$, the event $\|L_n\|_x \sim \alpha n$ is at least as likely as the event $\tau(L_n) \sim \alpha n$. \[lemma.getting.norm.close.to.tau\] For each $\epsilon>0$ there exist $D,N{\geqslant}1$ so that the following holds. Let $s_1,\dots,s_n\in S$, for some $n{\geqslant}N$, and let $g_i=s_1\dots s_i$ and $r_i=s_{i+1}\dots s_n$. Suppose that $\tau(g_n){\geqslant}\epsilon n$. Then there exists $i$ so that $\|\,\|r_ig_i\|_x -\tau(g_n)\,\|{\leqslant}D$. Assuming that $\epsilon n$ is sufficiently large compared to the constant $D_1$ of Lemma \[lem:far\_from\_min\], we see that any geodesic from $x$ to $g_nx$ contains a subgeodesic $\gamma$ of length ${\geqslant}\epsilon n/2$ with endpoints $D_1$-close to $Min(g_n)$. In view of Lemma \[lem:Min\_qconv\] and $2\delta$-thinness of quadrangles, any point $p$ on $\gamma$ is $(D_1+2\delta)$-close to $Min(g_n)$, and in particular $p$ satisfies $$d(p,g_np){\leqslant}\tau(g_n)+4\delta+2(D_1+2\delta).$$ In view of Lemma \[lem:cant\_avoid\_geod\] (applied with $\epsilon/2$ replacing $\epsilon$, and denoting $D_0$ the corresponding constant), we have $d(g_ix,p){\leqslant}D_0$ for some $p\in\gamma$, provided that $n$ is large enough. We now have: $$\|r_ig_i\|_x=d(x,r_ig_ix)=d(g_ix,g_ng_ix){\leqslant}d(p,g_np)+2D_0{\leqslant}\tau(g_n)+2(D_1+D_0)+8\delta.$$ On the other hand, $\|r_ig_i\|_x{\geqslant}\tau(r_ig_i)=\tau(g_n)$, where the equality holds since $r_ig_i$ and $g_n$ are conjugate. This concludes the proof. The following useful lemma will be used to show that the exponential rate of decay of the event that $\|L_n\|$ stays sublinear is comparable to the event that $\tau(L_n)$ stays sublinear. \[lemma.moving.tau\] For each $\epsilon>0$ there exists $N{\geqslant}1$ so that the following holds. Let $s_1,\dots,s_n\in S$, for some $n{\geqslant}N$, and let $g_i=s_1\dots s_i$ and $r_i=s_{i+1}\dots s_n$. Then for every $r\in [\tau(g_n),\|g_n\|_x]$ there exists $i$ so that $\|\,\|r_ig_i\|_x - r\,\|{\leqslant}\epsilon n$. First, notice that for $r{\geqslant}\|g_n\|_x-\epsilon n$ we can just choose $i=n$, so in the arguments below we assume $r{\leqslant}\|g_n\|_x-\epsilon n$. Set $d=(r-\tau(g_n))/2$, and assume that $n$ is larger than the $N$ from Lemma \[lem:cant\_avoid\_geod\] with $\epsilon/4$ replacing $\epsilon$. We impose further constraints on $n$ later. [Claim:]{} If $n$ is sufficiently large, then we can find a subgeodesic $\gamma$ of length $\epsilon n/4$ of a geodesic $\gamma'$ from $x$ to $g_nx$ so that any $p\in\gamma$ has - $d(p,q){\leqslant}D_1+\delta$ for some $q$ on a geodesic from $x$ to $\pi^{g_n}(x)$, - $\|d(p,\pi^{g_n}(x))-d\|{\leqslant}\epsilon n/3$. We let $p_0$ be the point along $\gamma'$ so that $$d(x,p_0)=d(x,\pi^{g_n}(x))-d-\epsilon n/4=\hat d,$$ and we let $\gamma$ be the subgeodesic of $\gamma'$ of length $\epsilon n/4$ with starting point $p_0$. We now check that, for $n$ large enough, this is all well-defined, and that $\gamma$ has the required property. Let us make the preliminary observation that $$\|g_n\|_x{\leqslant}d(x,\pi^{g_n}(x))+d(\pi^{g_n}(x),\pi^{g_n}(g_nx))+ d(\pi^{g_n}(g_nx),g_nx) {\leqslant}2d(x,\pi^{g_n}(x))+\tau(g_n)+4\delta.$$ Observe now that we have $$d{\leqslant}(\|g_n\|_x-\epsilon n-\tau(g_n))/2{\leqslant}d(x,\pi^{g_n}(x))+4\delta-\epsilon n/2,$$ implying $\hat d{\geqslant}d-4\delta +\epsilon n/4$, which is a positive quantity if $n$ is sufficiently large. Also, again by Lemma \[lem:far\_from\_min\], there exists $p'\in \gamma'$ so that $d(p',\pi^{g_n}(x)){\leqslant}D_1$; denote by $\gamma''$ the initial subgeodesic of $\gamma'$ with terminal point $p'$. Notice that for $n$ large enough we have $$d(x,p'){\geqslant}d(x,\pi^{g_n}(x))-D_1= d+\hat d +\epsilon n/4-D_1 {\geqslant}\hat d +\epsilon n/4.$$ The inequalities we just showed imply that $p_0$ and $\gamma$ are well-defined and, furthermore, that $\gamma$ is a subgeodesic of $\gamma''$. Considering a triangle with vertices $x,p', \pi^{g_n}(x)$ and containing $\gamma''$, we see that any point on $\gamma''$, whence any point on $\gamma$, is $(D_1+\delta)$-close to a point on a geodesic from $x$ to $\pi^{g_n}(x)$. In particular, for any $p\in \gamma$ we have $$\|d(p,\pi^{g_n}(x))+d(x,p) -d(x,\pi^{g_n}(x)\|{\leqslant}2D_1+2\delta,$$ and hence $$\|d(p,\pi^{g_n}(x)) - d\|{\leqslant}\|d(x,\pi^{g_n}(x))-d(x,p)- d\| +2D_1+2\delta {\leqslant}$$ $$\|d(x,\pi^{g_n}(x))-d(x,p_0)- d- d(p_0,p)\| + 2D_1+2\delta=\|\epsilon n/4 - d(p_0,p)\| + 2D_1+2\delta {\leqslant}\epsilon n/4 +2D_1+2\delta.$$ Provided that $n$ is large enough, this concludes the proof of the claim. By Lemma \[lem:cant\_avoid\_geod\], there exists $i$ so that we have $d(g_i,p){\leqslant}D_0$ for some $p\in\gamma$, and by the claim we have $d(p,q){\leqslant}D_1+\delta$, whence $d(g_ix,q){\leqslant}D_0+D_1+\delta$, for some $q$ on a geodesic from $x$ to $\pi^{g_n}(x)$. Notice that we can assume that $\pi^{g_n}(q)=\pi^{g_n}(x)$ since $$d(q,\pi^{g_n}(x))=d(z,\pi^{g_n}(x))-d(q,x){\leqslant}d(x,Min(g_n))+1-d(q,x){\leqslant}d(q,Min(g_n))+1.$$ In particular, in view of Lemma \[lem:far\_from\_min\], we have $$d(q,g_nq){\geqslant}2d(q,\pi^{g_n}(x))+\tau(g_n)-D_1.$$ We can now compute $$\|r_ig_i\|_x=d(x,r_ig_ix)=d(g_ix,g_ng_ix) {\geqslant}d(q,g_nq)- 2(D_0+D_1+\delta){\geqslant}2d(q,\pi^{g_n}(x))+\tau(g_n)- 3(D_0+D_1+\delta).$$ Hence, $$\|r_ig_i\|_x{\geqslant}2d +\tau(g_n) - 2\epsilon n/3- 5(D_0+D_1+\delta)= r - 2\epsilon n/3 -5(D_0+D_1+\delta).$$ For $n$ sufficiently large, this last quantity is ${\geqslant}r-\epsilon n$. On the other hand, we also have $$\|r_ig_i\|_x= d(g_ix,g_ng_ix){\leqslant}d(p,\pi^{g_n}(x))+d(\pi^{g_n}(x),\pi^{g_n}(g_nx)) +d(g_np,\pi^{g_n}(g_nx))+2D_0{\leqslant}$$ $$2d+2\epsilon n/3 +2D_1+\tau(g_n)+4\delta+2D_0{\leqslant}r+2\epsilon n/3+2D_0+2D_1+4\delta.$$ For $n$ sufficiently large, this last quantity is ${\leqslant}r+\epsilon n$, and we are done. We now proceed with the proof of Theorem \[Thm.LDP.translation\]. Following Theorem \[thm.LDP.criterion\], for $\alpha\in [0,\infty)$, we set $$J_{li}(\alpha):=\sup_{r \to 0} -\liminf_{n \to \infty}\frac{1}{n} \log \mathbb{P}\left(\frac{1}{n}\tau(L_n)\in B(\alpha,r)\right)$$ and $$J_{ls}(\alpha):=\sup_{r \to 0} -\limsup_{n \to \infty}\frac{1}{n} \log \mathbb{P}\left(\frac{1}{n}\tau(L_n)\in B(\alpha,r)\right).$$ Let $I:[0,\infty) \to [0,\infty]$ be the convex rate function given by Theorem \[thm.LDP.precise\]. In §\[subsec.translation.distance.as.likely\], we show that for every $\alpha {\geqslant}0$, $I(\alpha) {\geqslant}J_{li}(\alpha)$. In §\[subsec.translation.functional.as.likely\], we show that for every $\alpha>0$, $J_{ls}(\alpha) {\geqslant}I(\alpha)$. And in §\[subsec.deal.with.zero\], we show that $I(0)=J_{ls}(0)=J_{li}(0)$ and this concludes the proof of Theorem \[Thm.LDP.translation\] in view of Theorem \[thm.LDP.criterion\]. One direction: translation distance at least as likely as displacement functional (at the exponential scale) {#subsec.translation.distance.as.likely} ------------------------------------------------------------------------------------------------------------ The following is a straightforward consequence of Lemma \[lemma.alignment\]: \[lemma.proximal\] There exist a constant $k_S \in \mathbb{N}$, a finite set $\mathcal{F} \subset S^{k_S}$, for every $f \in \mathcal{F}$ a neighborhood $V_f$, and a constant $L>0$ such that for every $g \in {\operatorname{Isom}}(X)$, there exists $f \in \mathcal{F}$ with $\tau(f'g) {\geqslant}\|g\|_x -L$ for every $f' \in V_f$. Note that here, the constant $L$ accounts for the lack of hypothesis on the size of $\|g\|_x$ as in Lemma \[lemma.alignment\]. Indeed, one can take $L$ to be the same constant given by that lemma. This geometric ingredient is sufficient to show that the exponential rates of decay of probabilities of rare events for the lengths of translation distances of random walks are bounded above by those of displacement functionals: Let $\alpha \in [0,\infty)$ and $\delta>0$ be given. As before, set $M_F:=\max_{f' \in \hat{V}}\|f'\|_x$, where $\hat{V}=\cup_{f \in \mathcal{F}}V_f$ and $\alpha_F:=\min_{f \in F} \mu^{\ast k_S}(V_f)>0$. By definition of the function $J_{li}:[0,\infty) \to [0,\infty]$, for every small enough $r_0>0$, there exists an increasing sequence $n_\ell$ of positive integers such that for every $\ell {\geqslant}1$, we have $$\label{eq.start.tau} \mathbb{P}\left(\frac{1}{n_\ell}\tau(L_{n_\ell}) \in B(\alpha,r_0)\right)=\mathbb{P}\left(\frac{1}{n_\ell}\tau(R_{n_\ell}) \in B(\alpha,r_0)\right) {\leqslant}e^{-n_\ell(J_{li}(\alpha)-\delta)},$$ where the first equality follows since the left random walk $L_n$ and the right random walk $R_n$ have the same distribution. On the other hand, for $n {\geqslant}m$ noting $R_m^n=Y_{m+1} \ldots Y_n$, an application of Theorem \[thm.LDP.precise\] and using fact that $Y_i$’s are i.i.d, yield that for every $\ell$ large enough, we have $$\label{eq.lower.bd.tau.1} \mathbb{P}(\frac{1}{n_\ell - k_S}\|R^{n_\ell}_{k_S}\|_x \in B(\alpha,r_0/2)) {\geqslant}e^{-(n_\ell-k_S)(I(\alpha)+\delta)}.$$ Using Lemma \[lemma.proximal\], for every $g \in G$, let us fix an element $f(g)\in \mathcal{F}$ such that $\tau(f'(g)g) {\geqslant}\|g\|_x -L$ for every $f'(g) \in V_{f(g)}$. Now, taking $n_\ell {\geqslant}\frac{1}{r_0} \max\{4 \alpha k_S,M_F,L\}$, it follows by this lemma that we have the following inclusion of events $$\label{eq.inclusion.tau.1} \{R_{k_S}=f(R_{k_S}^{n_\ell})\} \cap \{\frac{1}{n_\ell-k_S} \|R^{n_\ell}_{k_S}\|_x \in B(\alpha,r_0/2)\} \subseteq \{\frac{1}{n_\ell}\tau(R_{n_\ell})\in B(\alpha,r_0)\}$$ Denote the events $\{\frac{1}{n_\ell}\tau(R_{n_\ell})\in B(\alpha,r_0)\}$, $\{\frac{1}{n_\ell-k_S} \|R^{n_\ell}_{k_S}\|_x \in B(\alpha,r_0/2)\}$ and $\{R_{k_S} \in V_f \; \text{where} \; f=f(R_{k_S}^{n_\ell})\}$, respectively, by $\Omega_\ell$, $E_\ell$ and $F_\ell$. Furthermore, for every $\ell {\geqslant}1$, denote by $\mathcal{E}_\ell$ the $\sigma$-algebra generated by $Y_{k_S+1},\ldots,Y_{n_\ell}$. Taking expectation of the inequality $1_{F_\ell}1_{E_\ell} {\leqslant}1_{\Omega_\ell}$ given by , we have $\mathbb{P}$-a.s.$$\mathbb{P}(\Omega_\ell) {\geqslant}\mathbb{E}[1_{F_\ell}1_{E_\ell}]=\mathbb{E}[1_{E_\ell} \mathbb{E}[ 1_{F_\ell} \, \| \, \mathcal{E}_\ell]]{\geqslant}\mathbb{P}(E_\ell) \alpha_F,$$ where the last inequality follows since by construction $\mathbb{P}$-a.s. $\mathbb{E}[1_{F_\ell} \, \|\, \mathcal{E}_\ell ] {\geqslant}\alpha_F$. Combining the previous equation with , we deduce $\mathbb{P}(\Omega_\ell) {\geqslant}e^{(-n_\ell-k_S)(I(\alpha)+\delta)} \alpha_F$. This yields that for every $n_\ell$ large enough, $$\mathbb{P}(\frac{1}{n_\ell}\tau(R_{n_\ell}) \in B(\alpha,r_0)) {\geqslant}e^{-n_\ell ((1+\delta)I(\alpha)+\delta)}$$ Taking logarithm, dividing by $n_\ell$, we deduce from that we have $$(1+\delta)I(\alpha)+2\delta {\geqslant}J_{li}(\alpha),$$ which shows that $I(\alpha) {\geqslant}J_{li}(\alpha)$ since $\delta>0$ is arbitrary. The other direction: displacement functional at least as likely as translation distance (at the exponential scale) {#subsec.translation.functional.as.likely} ------------------------------------------------------------------------------------------------------------------ Here we show that for every $\alpha>0$, $J_{ls}(\alpha) {\geqslant}I(\alpha)$. Let $\alpha \in (0,\infty)$ be given. Fix some $\delta>0$. Using Theorem \[thm.LDP.precise\], there exists $r_0>0$ such for every positive $r < \min\{r_0,\alpha\}$, there exists $n_0 \in \mathbb{N}$ such that 1. $\frac{\log n_0}{n_0} {\leqslant}\frac{\delta}{3}$, 2. $n_0 {\geqslant}\max\{N,\frac{DD}{r}\}$ where $N$ and $D$ are as given by Lemma \[lemma.getting.norm.close.to.tau\] with $\epsilon=\frac{1}{2}(\alpha-r)$. 3. $ \mathbb{P}(\frac{1}{n_0}\tau(L_{n_0}) \in B(\alpha,r)) {\geqslant}e^{-n_0(J_{ls}(\alpha)+\delta/3)}. $ 4. $\mathbb{P}(\frac{1}{n_0}\|L_{n_0}\|_x \in B(\alpha,2r)) {\leqslant}e^{-n_0 (I(\alpha)-\delta/3)}$. Let $E_0$ be the subset of the Cartesian product $S^{n_0}$ given by $$E_0=\{(s_1,\ldots,s_{n_0}) \, \| \, \frac{1}{n_0}\tau(s_1\ldots s_{n_0}) \in B(\alpha,r)\}.$$ By choice of $\epsilon=\frac{1}{2}(\alpha-r)$ and $n_0$ with (2) above, applying Lemma \[lemma.getting.norm.close.to.tau\], we get that for some $D>0$ and for every $(s_1,\ldots,s_{n_0}) \in E_0$, there exists an integer $i \in [1,n_0]$ such that we have $$\label{eq.cycling} \|\|s_{i+1}\ldots s_{n_0}s_1\ldots s_i\|_x -\tau(s_1\ldots s_n)\| {\leqslant}D$$ Recall that the asymptotic translation distance $\ell(g)$ of an element $g \in {\operatorname{Isom}}(X)$ is continuous [@oregon.reyes:properties Theorem 1.9] and the difference $\|\ell(g)-\tau(g)\|$ is uniformly bounded (depending only on the hyperbolicity constant of $X$) [@CDP Ch.10, Prop. 6.4]. Using this fact and the continuity of $\|.\|_x$, for each $(s_1,\ldots,s_{n_0})\in E_0$, up to increasing $D$ by a bounded amount (depending only on the hyperbolicity constant of $X$), we can fix such a choice of $i \in [1,n_0]$ (as in ) so that the corresponding function $p:\hat{E}_0 \to [1,n_0]$ is measurable. Now for every integer $k \in [1,n_0]$, let $$E_{0,k}=\{(s_1,\ldots,s_{n_0}) \in E_0 \; \| \; p(s_1,\ldots,s_{n_0})=k\}.$$ By construction, the $E_{0,k}$ give a partition of $E_0$, so we have $$\sum_{k=1}^{n_0} \mathbb{P}((Y_1,\ldots,Y_{n_0}) \in E_{0,k}) {\geqslant}\mathbb{P}\left(\frac{1}{n_0}\tau(R_{n_0}) \in B(\alpha,r)\right) {\geqslant}e^{-n_0 (J_{ls}(\alpha)+\delta/3)}.$$ Consequently, there exists some integer $k_0 \in [1,n_0]$ such that we have $$\label{eq.Ek0.lower.bound} \mathbb{P}((Y_1,\ldots,Y_{n_0}) \in E_{0,k_0}) {\geqslant}\frac{1}{n_0}e^{-n_0 (J_{ls}(\alpha)+2\delta/3)} {\geqslant}e^{-n_0(J_{ls}(\alpha)+\delta)},$$ where the last inequality follows by (1) in the choice of $n_0$. On the other hand, by , for every $(s_1,\ldots,s_{n_0}) \in E_{0,k_0}$, we have $$\|\|s_{k_0+1}\ldots s_{n_0}s_1\ldots s_{k_0}\|_x-\tau(s_1\ldots s_{n_0})\| {\leqslant}D,$$ so that we have the inclusion of events $$\label{eq.inclusion.tau.norm} \{\|Y_1\ldots Y_{n_0}\|_x \in B(\alpha n_0, r n_0 +D)\} \supseteq \{(Y_{k_0+1},\ldots,Y_{n_0},Y_1,\ldots,Y_{k_0}) \in E_{0,k_0}\}.$$ Moreover, using the fact that $Y_i$’s are i.i.d, we also have $$\label{eq.probas.equal.tau} \mathbb{P}((Y_{k_0+1},\ldots,Y_{n_0},Y_1,\ldots,Y_{k_0}) \in E_{0,k_0})=\mathbb{P}((Y_1,\ldots,Y_{n_0}) \in E_{0,k_0}).$$ Combining , and , and since $D {\leqslant}rn_0$, we deduce $$\mathbb{P}\left(\frac{1}{n_0}\|R_{n_0}\|_x \in B(\alpha,2r)\right) {\geqslant}e^{-n_0(J_{ls}(\alpha)+2\delta/3)}.$$ Now taking logarithm, dividing by $n_0$, plugging in (4) above, we deduce $$I(\alpha) {\leqslant}J_{ls}(\alpha)+\delta.$$ Since $\delta>0$ is arbitrary, we are done with showing $J_{ls}(\alpha) {\geqslant}I(\alpha)$. Dealing with the decay rates at zero {#subsec.deal.with.zero} ------------------------------------ We recall that up to now, we showed that $I(0) {\geqslant}J_{li}(0)$ and for every $\alpha>0$, $I(\alpha)=J_{li}(\alpha)=J_{ls}(\alpha)$. It only remains to show $J_{ls}(0) {\geqslant}I(0)$. To do this, we now suppose that $J_{ls}(0)<I(0)$ and we shall deduce a contradiction using our large deviation Theorem \[thm.LDP.precise\] for the displacement functionals and geometric Lemma \[lemma.moving.tau\]. Since the function $J_{ls}$ is non-negative, the function $I:[0,\infty) \to [0,\infty]$ attains zero at some $\alpha_0$ strictly positive. Since $I$ is lower semi-continuous at $0$ and convex on $[0,\infty)$, it follows that there exists $\epsilon_0>0$ and $\delta>0$ such that for every $\alpha \in [0,5\epsilon_0]$, we have $I(\alpha)> J_{ls}(0)+5\delta$. Up to reducing $\epsilon_0>0$, it follows by Theorem \[thm.LDP.precise\] and the definition of $J_{ls}$ that there exists a sequence $n_\ell$ of positive integers such that $$\mathbb{P}(\tau(R_{n_\ell}){\leqslant}n_\ell \epsilon_0) {\geqslant}e^{-n_\ell (I(0)-3\delta)}$$ and $$\label{eq.at.zero.1} \mathbb{P}(\|R_{n_\ell}\|_x {\leqslant}4n_\ell \epsilon_0) {\leqslant}e^{-n_\ell (I(0)+\delta)}$$ It follows that for every $\ell {\geqslant}1$ large enough, we have $$\label{eq.at.zero.2} \mathbb{P}(\{\tau(R_{n_\ell}) {\leqslant}n_\ell \epsilon_0\} \cap \{\|R_{n_\ell}\| {\geqslant}4 n_\ell \epsilon_0\}) {\geqslant}e^{-n_\ell (I(0)-2\delta)}.$$ Denote by $E_\ell$ the subset of the Cartesian product $S^{n_\ell}$ given by $$\{(s_1,\ldots,s_{n_\ell}) \in S^{n_\ell} \, \| \, \tau(s_1\ldots s_{n_\ell}) {\leqslant}n_\ell \epsilon_0 \; \text{and}\; \|s_1\ldots s_{n_\ell}\|_x {\geqslant}4n_\ell \epsilon_0\},$$ so that by , for every $\ell {\geqslant}1$ large enough, we have $$\label{eq.at.zero.3} \mathbb{P}((Y_1,\ldots,Y_{n_\ell})\in E_{\ell}) {\geqslant}e^{-n_\ell (I(0)-2\delta)}.$$ By Lemma \[lemma.moving.tau\], there exists $N \in \mathbb{N}$ such that for every $\ell$ large enough (so that, we also have $n_\ell {\geqslant}N$) and for every $(s_1,\ldots,s_{n_\ell}) \in E_{n_\ell}$, there exists an integer $i \in [1,\ldots,n_\ell]$ such that $$\label{eq.at.zero.4} \|s_{i+1}\ldots s_{n_\ell} s_1 \ldots s_i\|_x \in [2\epsilon_0 n_\ell , 3 \epsilon_0 n_\ell].$$ For every large enough $\ell$ and $(s_1,\ldots,s_{n_\ell}) \in E_{n_\ell}$, fix such a choice of $i$ and denote the corresponding mapping by $p$, i.e. $p(s_1,\ldots,s_{n_\ell})=i$ (as in §\[subsec.translation.functional.as.likely\], we can make sure that $p$ is measurable). Furthermore, for every large enough $\ell$ and $i=1,\ldots,n_\ell$, set $$E_{n_\ell,i}=\{(s_1,\ldots,s_{n_\ell}) \in E_{n_\ell} \, \| \, p(s_1,\ldots,s_{n_\ell})=i\},$$ so that $(E_{n_\ell,i})_{i=1,\ldots,n_\ell}$ gives a partition of $E_{n_\ell}$. As in , it follows that for every $\ell$ large enough, there exists an integer $k_\ell \in [1,n_\ell]$ such that $$\label{eq.at.zero.5} \mathbb{P}((Y_1,\ldots,Y_{n_\ell}) \in E_{n_\ell,k_\ell}) {\geqslant}\frac{1}{n_\ell}e^{-n_\ell (I(0)-2\delta)} {\geqslant}e^{-n_\ell(I(0)-\delta)}.$$ On the other hand, by , we have the following inclusion of events $$\label{eq.at.zero.6} \{\|Y_{k_\ell+1}\ldots Y_{n_\ell}Y_1 \ldots Y_{k_\ell}\|_x \in [2n_\ell \epsilon_0,3n_\ell\epsilon_0]\} \supseteq \{(Y_1,\ldots,Y_{n_\ell}) \in E_{n_\ell,k_\ell}\}.$$ Evaluating probabilities on both sides of , plugging in and using the fact that $Y_i$’s are i.i.d, we get $$\mathbb{P}\left(\frac{1}{n_\ell}\|R_{n_\ell}\|_x \in [2 \epsilon_0,3\epsilon_0]\right)=\mathbb{P}\left(\left\{\frac{1}{n_\ell}\|Y_{k_\ell+1}\ldots Y_{n_\ell}Y_1 \ldots Y_{k_\ell}\|_x \in [2 \epsilon_0,3\epsilon_0]\right\}\right) {\geqslant}e^{-n_\ell(I(0)-\delta)},$$ contradicting and showing that $J_{ls}(0) {\geqslant}I(0)$. As indicated earlier, we have thus shown that $\frac{1}{n} \tau(L_n)$ satisfies a weak LDP with the same rate function $I:[0,\infty) \to [0,\infty]$ given by Theorem \[thm.LDP.precise\]. To see that the full LDP holds under the exponential moment assumption, one just observes that since for every $g \in G$, $\tau(g) {\leqslant}\|g\|_x$, exponential tightness of $\frac{1}{n}\|L_n\|$ (Lemma \[lemma.expmoment.implies.exptight\]) implies that of $\frac{1}{n}\tau(L_n)$. This finishes the proof of Theorem \[Thm.LDP.translation\]. Support of the rate function {#sec.support} ============================ This section is devoted to the proofs of Theorem \[thm.support\] and Propositions \[prop.joint.spectrum\], \[prop.berger-wang\]. Support of the rate function {#subsec.support} ---------------------------- 1\) Suppose that the interval given by effective support of $I$ has non-empty interior. Taking $\alpha<\beta$ inside the interior of $D_I$, we see that for some $\delta<\frac{1}{2}(\beta-\alpha)$ and for every $t \in \mathbb{N}$ large enough, we have $\mathbb{P}(\frac{1}{t}\|g\|_x {\geqslant}\beta-\delta) {\geqslant}e^{-t I(\beta)}>0$ and $\mathbb{P}(\frac{1}{t}\|g\|_x {\leqslant}\alpha+\delta) {\geqslant}e^{-t I(\alpha)}>0$. So, for every $t$ large enough, there exist elements $g_1(t),g_2(t)$ in $S^t$ satisfying $\|g_1(t)\|_x {\leqslant}t(\alpha+\delta)$ and $\|g_2(t)\|_x {\geqslant}t(\beta-\delta)$. Applying Lemma \[lemma.proximal\] to $g_2(t)$, we obtain an element $g_2'(t) \in S^{k_S+t}$ with $\tau(g_2'(t)) {\geqslant}t(\beta-\alpha)-L$. Similarly, multiplying $g_1(t)$ with the $k_S^{th}$-power of a fixed element $h$ in $S$, we obtain $hg_1(t) \in S^{k_S+t}$ with $\tau(hg_1(t)){\leqslant}\|hg_1(t)\|_x {\leqslant}t(\alpha+\delta)$. Since $2\delta< \beta-\alpha$ and the difference $\|\tau(.)-\ell(.)\|$ is uniformly bounded, non-arithmeticity follows provided that $t$ is large enough. \[rk.arithmetic\] The same proof above also shows that for a bounded set $S$, if $\ell_{sub}(S)\neq \ell(S)$ then $S$ is non-arithmetic. The converse implication being obvious, it follows that $S$ is non-arithmetic if and only if $\ell_{sub}(S) \neq \ell(S)$. For the other direction, since $\mu$ is non-elementary and non-arithmetic, it is easy to see that there exists $n_0 \in \mathbb{N}$ such that $S^{n_0}$ contains two hyperbolic elements $h,h'$ with disjoint sets of fixed points on $\partial X$ as well as two elements $g,g'$ with distinct asymptotic translation distances $\ell(g) \neq \ell(g')$. Since the action of $G$ on $\partial X$ as well as the function $\ell$ is continuous (for the latter see [@oregon.reyes:properties Theorem 1.9]), we can choose some disjoint neighborhoods $U_h,U_h',U_g,U_g'$ such that the corresponding properties (i.e. disjoint sets of fixed points and distinct stable lengths) holds for every pair of elements in these neighborhoods. Let $\nu$ be the probability measure given by the restriction of $\mu^{\ast n_0}$ to $U_h \cup U_{h'} \cup U_g \cup U_{g'}$. It follows that $\nu$ is a non-elementary probability measure and satisfies $\nu {\leqslant}c \mu^{\ast n_0}$ where $\frac{1}{c}=\mu^{\ast n_0}(U_h \cup U_{h'} \cup U_g \cup U_{g'})>0$. On the other hand, it is readily observed (by definitions) that, if we denote by $I_{\mu}$ and $I_{\nu}$, the rate functions given by Theorem \[thm.LDP.norm\] of the random walks, respectively, driven by $\mu$ and $\nu$, we have $I_{\nu}(\alpha) {\geqslant}n_0 I_\mu(\frac{\alpha}{n_0}) - \log c$ for every $\alpha {\geqslant}0$. Now, as $\nu$ has bounded support, the claim will follow from 2. applied to $D_{I_\nu}$, since for the support $S_\nu$ of $\nu$, we clearly have $\ell_{sub}(S_\nu)< \ell(S_\nu)$.\ 2) The fact that $\overline{D}_I \subseteq [\ell_{sub}(S),\ell(S)]$ is readily seen from the definitions of $\ell_{sub}(S)$ and $\ell(S)$. So we only need to prove $D_I \supseteq (\ell_{sub}(S),\ell(S))$. There is nothing to prove if $\ell_{sub}(S)=\ell(S)$, so suppose $\ell(S)>\ell_{sub}(S)$. First, suppose that $\lambda(\mu) \neq \ell_{sub}(S)$ and let $\alpha \in (\ell_{sub}(S),\ell(S))$ be such that $\alpha {\leqslant}\lambda(\mu)$. Then, by definition of $\ell_{sub}(S)$, there exists $n_0 \in \mathbb{N}$ and $g \in S^{n_0}$ such that $\frac{1}{n_0}\|g\|_x<\alpha$. Let $U_g$ be a neighborhood of $g$ such that $\mu^{\ast n_0}(U)>0$ and $\frac{1}{n_0}\|g'\|_x<\alpha$ for every $g' \in U$. Using Theorem \[thm.LDP.norm\], we can now write, $$\begin{aligned} I(\alpha)&=-\lim_{n \to \infty} \frac{1}{n} \log \mathbb{P}(\frac{1}{n}\|L_n\|_x <\alpha){\leqslant}-\liminf_{t \to \infty} \frac{1}{tn_0} \log \mathbb{P}(\frac{1}{tn_0}\|L_{tn_0}\|_x<\alpha)\\ & {\leqslant}- \liminf_{t \to \infty} \frac{1}{tn_0} \log \mathbb{P}(L_{tn_0} \in U^t) {\leqslant}-\lim_{t \to \infty} \frac{1}{tn_0} \log \mathbb{P}(L_{n_0}\in U)^t \\ &=-\frac{1}{n_0}\log \mu^{\ast n_0}(U)<\infty, \end{aligned}$$ where in the first equality we used the convexity of $I$ together with the fact that $\alpha {\leqslant}\lambda(\mu)$, in the second inequality we used the subadditivity of the displacement functional $\|.\|_x$ and in the third inequality we used the fact that the random walk increments are i.i.d. The proof of the other case is very similar to the proof of Lemma \[lemma.stability\]; to avoid repetition, we only indicate how to adapt that proof. So, suppose that $\lambda(\mu) \neq \ell(S)$ and let $\alpha \in (\ell_{sub}(S),\ell(S))$ be such that $\alpha {\geqslant}\lambda(\mu)$. Fix $\eta>r>0$ and $\varepsilon>0$ with $r+\varepsilon<\eta$ and $\alpha+2\eta {\leqslant}\ell(S)$. By definition of $\ell(S)$, for every $n\in \mathbb{N}$, there exists $g_n \in S^n$ such that $\frac{1}{n}\|g_n\|_x {\geqslant}\alpha+2\eta$. By continuity, for every $n \in \mathbb{N}$ there exists a neighborhood $U_n$ of $g_n$ such that for every $g'_n \in U_n$, we have $\frac{1}{n}\|g_n'\|_x {\geqslant}\alpha +\eta$ and $\mu^{\ast n}(U_n)>0$. Now replace $\alpha$ in the proof of Lemma \[lemma.stability\] by $\alpha+\eta$, fix $\delta>0$, then choose $n_0 \in \mathbb{N}$ satisfying (2),(3) and (4) of . Let $\beta>$ be a constant so that we have $-\frac{1}{n_0} \log \mathbb{P}(\frac{1}{n_0}\|L_{n_0}\|_x \in B(\alpha+\eta,r)) {\leqslant}\beta+\delta/2$. Note that the choice of $\beta$ to be finite is guaranteed by the fact that $\mu^{\ast n_0}(U_{n_0})>0$. Now the rest of the proof of Lemma \[lemma.stability\] (from onward) goes precisely the same and yield , i.e. for every $k {\geqslant}1$, $$\begin{aligned} -I(\alpha) & {\geqslant}\limsup_{n \to \infty} \frac{1}{n} \log \mathbb{P}\left(\frac{1}{n}\|L_n\|_x {\geqslant}\alpha\right)\\ & {\geqslant}\frac{1}{kn_0 +(k-1)k_S} \log \mathbb{P}\left(\frac{1}{kn_0 +(k-1)k_S}\|L_{kn_0 +(k-1)k_S}\|_x \in B(\alpha,r+\varepsilon)\right) \\ &{\geqslant}-\beta -\delta. \end{aligned}$$ where we used Theorem \[thm.LDP.norm\] in the first inequality. This shows $I(\alpha)<\infty$ and completes the proof.\ 3) If $\ell_{sub}(S)=\ell(S)$, then $\lambda(\mu)=\ell(S)$ and $I(\ell(S))=0$ so that by 2), $D_I=\{\ell(S)\}$. Therefore we suppose that $\ell_{sub}(S)<\ell(S)$ and we only need to show that on $(\ell_{sub}(S),\ell(S))$, the rate function $I$ is bounded above by $-\min_{g\in S}\log \mu(g)<\infty$; lower semi-continuity of $I$ then entails that $I$ is bounded above by $-\min_{g\in S}\log \mu(g)$ on $[\ell_{sub}(S),\ell(S)]$ proving the claim. So let $\alpha \in (\ell_{sub}(S),\ell(S))$. By 2), $\alpha \in D_{I}$ and by Theorem \[thm.LDP.norm\], for any $\delta>0$, for every small enough $r>0$ and large enough $n \in \mathbb{N}$, we have $$\label{eq.supp2} 0<e^{-n(I(\alpha)+\delta)}{\leqslant}\mathbb{P}(\frac{1}{n}\|L_n\|_x \in B(\alpha,r)) {\leqslant}e^{-n(I(\alpha)-\delta)}.$$ It follows that for every such $n \in \mathbb{N}$, there exists $g_n \in {\operatorname{supp}}(\mu^{\ast n})$ with $\frac{1}{n}\|g_n\|_x \in B(\alpha,r)$. By the i.i.d. property of random walk increments, writing $g_n$ as a product $s_n\ldots s_1$ with $s_i \in {\operatorname{supp}}(\mu)$, it follows that $\mathbb{P}(L_n=g_n) {\geqslant}(\min_{s \in S}\mu(s))^n$. Plugging this in , since $\delta>0$ is arbitrary, we deduce that $I(\alpha){\leqslant}-\min_{s \in S} \log \mu(s)$, as required. We now construct some examples illustrating in the setting of Theorem \[thm.support\] that when the support of the probability measure is not finite, the rate function of LDP can explode on $\ell_{sub}(S)$ or $\ell(S)$ or both. In fact, by considering the action of ${\operatorname{SL}}_2(\mathbb{R})$ on the Poincaré disc $\mathbb{D}$ and using the relation $\frac{1}{2}\log\|\|g\|\|=\|g\|_o$ where $g \in G$, $o$ denotes the origin in $\mathbb{D}$ and $\|\|.\|\|$ is the operator norm induced by the Euclidean norm on $\mathbb{R}^2$, [@sert.LDP Example 5.5] already provides an example of a rate function that explodes on $\ell(S)$. Below, we shall give more examples where rate function explodes on any subset of $\{\ell_{sub}(S),\ell(S)\}$. ${}$\ Consider $G={\operatorname{SL}}_2(\mathbb{R})$ acting isometrically on the Poincaré disc $\mathbb{D}$ endowed with the usual hyperbolic metric $d$. Let $c_a$ and $c_b$ be two geodesics in $\mathbb{D}$ that are of distance $d=1$ to the origin and denoting their endpoints on $\partial \mathbb{D}$, respectively, by $\{x_a^+,x_a^-\}$ and $\{x_b^+,x_b^-\}$, suppose that these are ordered as $(x_b^-,x_b^+,x_a^+,x_a^-)$. For $k {\geqslant}1$, $t \in \{a,b\}$, $t_k$ be hyperbolic elements of $G$ with translation axis $c_t$, attracting/repelling fixed points $x_t^+$ and $x_t^-$ and translation distance $10-\frac{1}{k}$ for $t=b$ and $\frac{1}{k}$ for $t=a$. Let $S=\{a_k,b_k \, \|\, k {\geqslant}1\}$ and let $Y_i$’s be the coordinate functions on $S^{\mathbb{N}}$. It is easy to see that the semigroup $\Gamma$ generated by $S$ consists of hyperbolic elements whose translation axes is contained in the connected region bounded by $c_a$ and $c_b$. It follows, e.g. by [@breuillard-sert Lemma 6.3], that denoting by $D>0$ twice the distance between $c_a$ and $c_b$, for any $g,h \in \Gamma$, we have $$\label{eq.ex.1} \tau(g)+\tau(h) {\leqslant}\tau(gh) {\leqslant}\tau(g)+\tau(h)+D.$$ Now for $t \in \{a,b\}$, $n {\geqslant}k {\geqslant}1$ and $(s_n,\ldots,s_1) \in S^n$, denote by $\hat{t}_{n,k}$ the number of $t_{i}$’s in $(s_n,\ldots,s_1)$ with $i {\geqslant}k$. It is readily seen by that we have the following inclusion of events for every $n {\geqslant}1$ and $1 {\leqslant}k {\leqslant}n$: $$\label{eq.ex.inc1} \left\{\frac{1}{n}\tau(Y_n\ldots Y_1)>10-\frac{1}{2k}\right\} \subset \left\{\hat{b}_{n,k} {\geqslant}\frac{n}{2}\right\}$$ and $$\label{eq.ex.inc2} \left\{\frac{1}{n}\tau(Y_n\ldots Y_1)<\frac{1}{2k}\right\} \subset \left\{\hat{a}_{n,k} {\geqslant}\frac{n}{2}\right\}.$$ Notice also that by elementary plane hyperbolic geometry, for every $g \in \Gamma$, we have $$\label{eq.ex.2} 0 {\leqslant}\|g\|_0-\tau(g) {\leqslant}D +4.$$ For any probability measure $\mu$ on $S$, the random variables $\frac{1}{n}\|L_n\|_o$ satisfies a LDP with some rate function $I$; this follows from Theorem \[thm.LDP.norm\] if the support of $\mu$ contains $a_i$’s and $b_i$’s (so that $\mu$ is non-elementary) and from classical theorem of Cramér (which can be readily deduced from Theorem \[thm.LDP.criterion\]) if the support contains only $a_i$’s or $b_i$’s. The inequality entails by Theorem \[thm.LDP.criterion\] (or using [@dembo-zeitouni Theorem 4.2.13]) that $\frac{1}{n} \tau(L_n)$ satisfies a LDP with rate function $I$ too. Now let $\alpha_k>0$ be such that $\sum_{k {\geqslant}1} \alpha_k=\frac{1}{2}$ and consider $\mu_1=\frac{1}{2}\delta_{a_1}+\sum_{k {\geqslant}1} \alpha_k \delta_{b_k}$ supported on $S_1$ and $\mu_2=\frac{1}{2}\delta_{b_1}+\sum_{k {\geqslant}1} \alpha_k \delta_{a_k}$ supported on $S_2$ and $\mu=\frac{1}{2}(\mu_1+\mu_2)$ supported on $S$. Denote by $I_1$, $I_2$ and $I$ the rate functions of the LDP of $\frac{1}{n}\|L_n\|_o$ when $S^\mathbb{N}$ is endowed with, respectively, $\mu_1^{\otimes \mathbb{N}},\mu_2^{\otimes \mathbb{N}}$ and $\mu^{\otimes \mathbb{N}}$. Using Theorem \[thm.support\] and Stirling’s formula, it is not hard to deduce from that $\ell(S_1) \notin D_{I_1}$, and from that $\ell_{sub}(S_2) \notin D_{I_2}$, and finally, that we have $D_I=(\ell_{sub}(S),\ell(S))$. Moreover, using , one sees that $D_{I_1}=[\ell_{sub}(S_1),\ell(S_1))$ and $D_{I_2}=(\ell_{sub}(S_2),\ell(S_2)]$. Proofs of deterministic consequences ------------------------------------ ### Hausdorff convergence The unfolding Lemma \[lemma.alignment\] and Lemmas \[lemma.getting.norm.close.to.tau\] and \[lemma.moving.tau\] can be used to give a direct proof of Proposition \[prop.joint.spectrum\]. Here, we give a short proof based on our large deviations results. It follows from the definitions that for every $\varepsilon>0$, there exists $N \in \mathbb{N}$ such that for every $n {\geqslant}N$ and $g\in S^n$, $\ell_{sub}(S) {\leqslant}\frac{1}{n} \|g\|_x {\leqslant}\ell(S)+\varepsilon$. This already implies the statement if $\ell_{sub}(S)=\ell(S)$, so let $\ell_{sub}(S)<\ell(S)$. Then, for every $\varepsilon \in (0,(\ell(S)-\ell_{sub}(S))/2) $, the set $S$ contains a finite subset $S'$ such that $\ell(S'){\geqslant}\ell(S)-\varepsilon$ and $\ell_{sub}(S'){\leqslant}\ell_{sub}(S)-\varepsilon$. This follows from the definitions of $\ell(S)$ and $\ell_{sub}(S)$. Therefore, in view of the previous paragraph, we can suppose that $S$ is finite. Now endow $S$ with the uniform probability measure and consider the corresponding random walk on ${\operatorname{Isom}}(X)$. Now given an interval $J$ of non-empty interior in $[\ell_{sub}(S),\ell(S)]$, by Theorem \[thm.LDP.norm\] and 3. of Theorem \[thm.support\], $-\liminf_{n \to \infty} \frac{1}{n} \log \mathbb{P}(\frac{1}{n}\|L_n\|_x \in J)>\infty$, which in particular says that for every $n \in \mathbb{N}$, large enough $J \cap \frac{1}{n}\|L_n\|_x \neq \emptyset$. Together with the first paragraph above, this shows the Hausdorff convergence of $\frac{1}{n}\|S^n\|_x$. The convergence of $\frac{1}{n}\tau(S^n)$ is deduced similarly using Theorem \[Thm.LDP.translation\]. ### Berger–Wang equality for isometries of hyperbolic spaces Assume for a contradiction that we have $\ell_\infty(S)<\ell(S)$. It follows that there exists $\delta>0$ such that for every $n {\geqslant}1$, there exists $g_n \in S^n$ with $\frac{1}{n}\|g_n\|_x {\geqslant}\ell_\infty(S)+\delta$. By Lemma \[lemma.alignment\] and the fact that $\|\tau(.)-\ell(.)\|$ is uniformly bounded, we deduce that for every $n$ large enough, there exists $f_n \in \mathcal{F} \subset S^{k_S}$ such that $\frac{1}{n}\ell(f_ng_n) {\geqslant}\ell_\infty(S)+\delta$. Since $f_n g_n \in S^{k_S+n}$, for every $n> \frac{2k_S \ell_\infty(S)}{\delta}+k_S$, we get $$\frac{\ell(f_ng_n)}{k_S+n} {\geqslant}\ell_\infty(S)+\frac{\delta}{2}$$ yielding a contradiction since $\ell(g^n)=n\ell(g)$ for every $n \in \mathbb{N}$ and $g \in {\operatorname{Isom}}(X)$.	{ "pile_set_name": "ArXiv" }
mpi-pks/9907008 Geometrical theory of diffraction and spectral statistics Martin Sieber[^1] Max-Planck-Institut für Physik komplexer Systeme, Nöthnitzer Str. 38, 01187 Dresden, Germany Abstract We investigate the influence of diffraction on the statistics of energy levels in quantum systems with a chaotic classical limit. By applying the geometrical theory of diffraction we show that diffraction on singularities of the potential can lead to modifications in semiclassical approximations for spectral statistics that persist in the semiclassical limit $\hbar \rightarrow 0$. This result is obtained by deriving a classical sum rule for trajectories that connect two points in coordinate space. PACS numbers:\ 03.65.Sq Semiclassical theories and applications.\ 05.45.Mt Semiclassical chaos (“quantum chaos”). Introduction ============ Correlations in the spectra of quantum systems with a chaotic classical limit are expected to follow random matrix theory. This is the content of the random matrix hypothesis [@BGS84; @Boh91], which has found confirmation by numerical investigations in many systems. Theoretical support for this conjecture has been obtained by semiclassical approximations based on periodic orbits [@Ber85; @BK96] as well as field theoretical methods [@AAA95]. In particular, it has been shown by the semiclassical method that the leading asymptotic behaviour of two-point correlation functions for long-range correlations agrees with results of random matrix theory. This result has been obtained by using mean properties of periodic orbits as expressed by the sum rule of Hannay and Ozorio de Almeida [@HO84]. It is expected that a proof of the random matrix hypothesis would be possible if certain correlations between periodic orbits were known [@ADDKKSS93]. The semiclassical analysis is based on the trace formula for the density of states [@Gut90]. This is a leading order approximation, as $\hbar \rightarrow 0$, in terms of the periodic orbits of the corresponding classical system. For most systems this approximation is not exact, and there are corrections in higher order of $\hbar$. A particular important correction occurs if the potential has discontinuities or singularities which lead to wave diffraction. This is the case in many standard examples of chaotic systems. It leads to additional contributions to the trace formula in terms of so-called creeping orbits or in terms of trajectories that are closed or connect two points in coordinate space [@VWR94; @PSc95; @BW96; @PSSU96]. In the present article we investigate systems in which the diffraction occurs at point-like objects, for example at a delta-like singularity of the potential, at a magnetic flux line in a two-dimensional system, or at a corner in a billiard system. For these systems we examine whether diffractive corrections to the trace formula can have an influence on spectral statistics that persists in the semiclassical limit $\hbar \rightarrow 0$. The analysis is done by deriving a classical sum rule for the orbits that arise in the geometrical theory of diffraction. With this input the diffractive corrections to the diagonal approximation for the spectral form factor are calculated, and it is shown that these corrections in general do not vanish as $\hbar \rightarrow 0$. Sum rules for transient orbits ============================== Semiclassical approximations for the Green function $G({\boldsymbol q}_b,{\boldsymbol q}_a,E)$ of a quantum system involve classical trajectories that go from ${\boldsymbol q}_a$ to ${\boldsymbol q}_b$ at energy $E$. In general there is an infinite number of these trajectories. Moreover, in a chaotic system the number of trajectories that connect the two points in a time less than $T$ increases exponentially with $T$. In this section we use the ergodic property of chaotic systems in order to obtain a sum rule for orbits connecting two points in coordinate space. We call these trajectories transient orbits. Consider a particle with energy $E$ that starts at time $t=0$ at a point ${\boldsymbol q}_a$ in coordinate space. The classical probability density for the particle to be found at time $t=T$ at a point ${\boldsymbol q}_b$ is given by $$\label{prop} P({\boldsymbol q}_b;T,{\boldsymbol q}_a,E) = \frac{\int \! {\text{d}}^f p_a \; \delta(E - H({\boldsymbol q}_a,{\boldsymbol p}_a)) \; \delta({\boldsymbol q}(T) - {\boldsymbol q}_b)} {\int \! {\text{d}}^f q_b \int \! {\text{d}}^f p_a \; \delta(E - H({\boldsymbol q}_a,{\boldsymbol p}_a)) \; \delta({\boldsymbol q}(T) - {\boldsymbol q}_b)} \; ,$$ where the function ${\boldsymbol q}(t)$ denotes the position of the particle with initial position ${\boldsymbol q}_a$ and momentum ${\boldsymbol p}_a$ at $t=0$, and $f$ is the number of degrees of freedom of the system. The integration in the numerator extends over all initial momenta corresponding to energy $E$, and the denominator gives the normalisation constant. The probability density is only different from zero if there are classical trajectories that connect the points ${\boldsymbol q}_a$ and ${\boldsymbol q}_b$ in time $T$. Let us assume that the particle moves in the field of a vector and a scalar potential and that the Hamiltonian is given by $$H({\boldsymbol q},{\boldsymbol p}) = \frac{1}{2m} \left({\boldsymbol p}- \frac{e}{c} \operatorname{A}({\boldsymbol q}) \right)^2 + V({\boldsymbol q}) \; .$$ In the following we express the transition probability density for this Hamiltonian in terms of classical trajectories. The numerator in (\[prop\]) is evaluated by introducing a local coordinate system in the vicinity of a trajectory where one coordinate is along the trajectory and the others are perpendicular to it, $$\int \! {\text{d}}^f p_a \; \delta(E - H({\boldsymbol q}_a,{\boldsymbol p}_a)) \; \delta({\boldsymbol q}(T) - {\boldsymbol q}_b) = \sum_\gamma \frac{1}{v_a \, v_b} \, \left\| {\det}' \left( \frac{\partial {\boldsymbol q}_b}{\partial {\boldsymbol p}_a} \right)_\gamma \right\|^{-1} \, \delta(T_\gamma - T) \; .$$ The sum is over all trajectories from ${\boldsymbol q}_a$ to ${\boldsymbol q}_b$ at energy $E$, $v_a$ and $v_b$ are the velocities at ${\boldsymbol q}_a$ and ${\boldsymbol q}_b$, respectively, $T_\gamma$ is the time along the trajectory $\gamma$, and the prime at the determinant denotes that the matrix of derivatives involves only the coordinates orthogonal to the trajectory. The denominator is evaluated by changing the integration over momentum into one over velocity and evaluating it in hyperspherical coordinates. The remaining integral over coordinates gives one, and one obtains $$\int \! {\text{d}}^f q_b \int \! {\text{d}}^f p_a \; \delta(E - H({\boldsymbol q}_a,{\boldsymbol p}_a)) \; \delta({\boldsymbol q}(T) - {\boldsymbol q}_b) = \frac{(m v_a)^{f-1}}{v_a} \, {\cal S}^{(f)} \; ,$$ where ${\cal S}^{(f)} = 2 \, \pi^{f/2} / \Gamma(f/2)$ is the surface area of an $f$-dimensional hypersphere with unit radius. Altogether the result is $$\label{traj} P({\boldsymbol q}_b;T,{\boldsymbol q}_a,E) = \sum_\gamma \frac{1}{v_b \, (m v_a)^{f-1} \, {\cal S}^{(f)}} \, \left\| {\det}' \left( \frac{\partial {\boldsymbol q}_b}{\partial {\boldsymbol p}_a} \right)_\gamma \right\|^{-1} \, \delta(T_\gamma - T) \; .$$ In an ergodic system the property that typical trajectories fill out phase space uniformly can be used to describe the probability density for long times $T$. In order to apply ergodicity we smooth the singular function $P$ over some range of the final coordinate $$\label{smooth} \langle P({\boldsymbol q}_b;T,{\boldsymbol q}_a,E) \rangle_{\varepsilon} \equiv \int \! {\text{d}}^f q_b' \; P({\boldsymbol q}_b';T,{\boldsymbol q}_a,E) \; \delta_\varepsilon ({\boldsymbol q}_b -{\boldsymbol q}_b') \; ,$$ where $\delta_\varepsilon$ is a smoothed delta-function whose width is parameterised by $\varepsilon$, and which is normalised to one $$\int_{V({\boldsymbol q}) < E} \! {\text{d}}^f q \: \delta_\varepsilon ({\boldsymbol q}- {\boldsymbol q}_b) = 1 \; .$$ It is assumed that the width $\varepsilon$ is classically small so that the potentials do not change appreciably within this region. The integral in (\[smooth\]) results in the replacement of the second delta function in the numerator of (\[prop\]) by $\delta_\varepsilon$. In order to describe the properties of the probability density $P$ we employ the property of an ergodic system that the time-average of a function along a trajectory is equal to its phase space average for almost all trajectories $$\label{ergod} \lim_{T \rightarrow \infty} \frac{1}{T} \; \int_{T_0}^{T} \! {\text{d}}t \; f({\boldsymbol q}(t),{\boldsymbol p}(t)) = \frac{\int \! {\text{d}}^f q \, {\text{d}}^f p \; \delta(E - H({\boldsymbol q},{\boldsymbol p})) \; f({\boldsymbol q},{\boldsymbol p})}{ \int \! {\text{d}}^f q \, {\text{d}}^f p \; \delta(E - H({\boldsymbol q},{\boldsymbol p}))} \; ,$$ and we choose for $f$ the smoothed delta-function $\delta_\varepsilon$. The phase space average over this function is evaluated similarly as before and yields $$\label{aver} \frac{\int \! {\text{d}}^f q \, {\text{d}}^f p \; \delta(E - H({\boldsymbol q},{\boldsymbol p})) \; \delta_\varepsilon({\boldsymbol q}- {\boldsymbol q}_b)}{ \int \! {\text{d}}^f q \, {\text{d}}^f p \; \delta(E - H({\boldsymbol q},{\boldsymbol p}))} = \frac{(m v_b)^{f-1} \, {\cal S}^{(f)}}{v_b \, \Sigma(E)} + {\cal O}(\varepsilon) \; ,$$ where $\Sigma(E)$ is the volume of the energy surface in phase space. After integrating (\[ergod\]) with (\[aver\]) over the initial momenta and dividing by the normalisation constant one obtains the following result for the probability density $$\label{sumprop} \lim_{T \rightarrow \infty} \frac{1}{T} \; \int_{T_0}^{T} \! {\text{d}}t \; \langle P({\boldsymbol q}_b;t,{\boldsymbol q}_a,E) \rangle_{\varepsilon} = \frac{(m v_b)^{f-1} \, {\cal S}^{(f)}}{v_b \, \Sigma(E)} + {\cal O}(\varepsilon) \; .$$ This property of $P$ has the following interpretation. When the first term on the right-hand side is multiplied by a volume element ${\text{d}}^f q$ it is the ratio between the volume of the part of the energy surface in phase space corresponding the element ${\text{d}}^f q$ around ${\boldsymbol q}_b$ and the volume of the total energy surface. Thus the probability of a particle to be found in a neighbourhood of ${\boldsymbol q}_b$ is equal to the relative volume of the energy surface of this neighbourhood. Equation (\[sumprop\]) gives an expression for the probability density $P$ that is smoothed over an $\varepsilon$-neighbourhood of the final point in an ergodic system. Using (\[traj\]), it is a sum rule for classical trajectories that start at ${\boldsymbol q}_a$ and visit the $\varepsilon$-neighbourhood of ${\boldsymbol q}_b$. In order to obtain a sum rule for trajectories that hit ${\boldsymbol q}_b$ one has to take the limit $\varepsilon \rightarrow 0$. Here one faces the problem that one would like to interchange the two limits $\varepsilon \rightarrow 0$ and $T \rightarrow \infty$, a problem that occurs also in the derivation of sum rules for periodic orbits [@HO84]. One can argue that the two limits can be interchanged if the classical trajectories do not have conjugate points. Then the terms $[{\det}' (\partial {\boldsymbol q}_b/\partial{\boldsymbol p}_a)]^{-1}$ decrease exponentially with the transition time $T_\gamma$. This means that the contribution of a single trajectory to the probability density is exponentially small for long times. The integral over the probability density $\int^T_{T_0} {\text{d}}t \, P({\boldsymbol q}_b;t,{\boldsymbol q}_a,E)$ is a discontinuous function of the final coordinate ${\boldsymbol q}_b$, but its variation inside the $\varepsilon$-environment of ${\boldsymbol q}_b$ becomes very small for long times $T$, and the two limits $\varepsilon \rightarrow 0$ and $T \rightarrow \infty$ can be interchanged. This argumentation does not apply, however, if there are caustics of the classical motion inside the $\varepsilon$-neighbourhood. On a caustic the term $[{\det}' (\partial {\boldsymbol q}_b/\partial{\boldsymbol p}_a)]^{-1}$ is divergent. This divergence is integrable so that the smoothed version of the sum rule (\[sumprop\]) is still valid. It might be possible that in the generic situation where the point ${\boldsymbol q}_b$ itself does not lie on a caustic, the two limits can still be interchanged. We point out, however, that for applications in semiclassical approximations quantum mechanics provides a natural smoothing of the probability density, so that only a smoothed version of the sum rule is required. In cases where the two limits can be interchanged we obtain the following sum rule for trajectories from ${\boldsymbol q}_a$ to ${\boldsymbol q}_b$ at energy $E$ $$\lim_{T \rightarrow \infty} \frac{1}{T} \sum_{T_\gamma < T} \frac{1}{v_b \, (m v_a)^{f-1} \, {\cal S}^{(f)}} \, \left\| {\det}' \left( \frac{\partial {\boldsymbol q}_b}{\partial {\boldsymbol p}_a} \right)_\gamma \right\|^{-1} = \frac{(m v_b)^{f-1} \, {\cal S}^{(f)}}{v_b \, \Sigma(E)} \; ,$$ and the differentiated version is $$\label{sumt} \sum_\gamma \frac{1}{v_b \, (m v_a)^{f-1} \, {\cal S}^{(f)}} \, \left\| {\det}' \left( \frac{\partial {\boldsymbol q}_b}{\partial {\boldsymbol p}_a} \right)_\gamma \right\|^{-1} \, \delta(T- T_\gamma) \approx \frac{(m v_b)^{f-1} \, {\cal S}^{(f)}}{v_b \, \Sigma(E)} \, , \qquad T \rightarrow \infty \; ,$$ where it is implied that the left-hand side has to be smoothed over some small time-interval $\Delta T$ in order to obtain a smooth function. In the following we give several variants of the sum rule that can be obtained in an analogous way. - For chaotic area-preserving maps on an $(2f)$-dimensional unit torus, $({\boldsymbol q}_{n+1},{\boldsymbol p}_{n+1}) = h({\boldsymbol q}_n,{\boldsymbol p}_n)$, the corresponding result is $$\label{summap} \sum_{({\boldsymbol q}_n,{\boldsymbol q}_0)=({\boldsymbol q}_b,{\boldsymbol q}_a)} \left\| \det \left( \frac{ \partial {\boldsymbol q}_n}{\partial {\boldsymbol p}_0} \right) \right\|^{-1} \approx 1 \; , \qquad n \rightarrow \infty \; ,$$ where the sum extends over all points $(q_0,p_0)$ for which $q_0=q_a$ and $q_n=q_b$. In appendix \[cats\] it is shown that this sum rule is exact for cat maps for all $n>0$. - For billiard systems the sum rule (\[sumt\]) can be expressed in a geometrical form. In two dimensions the result is $$\label{sumbill} \frac{1}{2 \pi} \sum_\gamma \frac{1}{\|M_{12}\|} \delta(L - L_\gamma) \approx \frac{1}{A} \; , \qquad L \rightarrow \infty \; ,$$ where $A$ is the area of the billiard, $L_\gamma$ is the geometrical length of a trajectory and $M_{12} = p_a \, (\partial q_b^\perp / \partial p_a^\perp)$ is an element of the stability matrix, scaled by $p_a$ in order to be energy independent. Eq. (\[sumbill\]) expresses that a particle is equally likely to be found anywhere in the billiard if it travels a sufficiently long distance $L$. Consequently, the probability density that it is found at some point ${\boldsymbol q}_b$ is equal to $A^{-1}$. - For applications in the next section one requires sum rules for a subset of transient orbits for which the angular orientations of the initial and the final velocities, ${\boldsymbol \phi}_a$ and ${\boldsymbol \phi}_b$, are fixed, $$\label{sumpart} \sum_{{\boldsymbol \phi}_a,{\boldsymbol \phi}_b \, \text{fixed}} \left\| {\det}' \left( \frac{\partial {\boldsymbol q}_b}{\partial {\boldsymbol p}_a} \right)_\gamma \right\|^{-1} \, \delta(T- T_\gamma) \approx \frac{(m v_a)^{f-1} \, (m v_b)^{f-1} }{\Sigma(E)} \, {\text{d}}\Omega_a \, {\text{d}}\Omega_b \; , \qquad T \rightarrow \infty \; .$$ Here the sum extends over all transient trajectories for which the initial and final angular directions lie in solid angle elements ${\text{d}}\Omega_a$ and ${\text{d}}\Omega_b$, respectively, around the directions ${\boldsymbol \phi}_a$ and ${\boldsymbol \phi}_b$. - In integrable systems a similar sum rule can be obtained, if the ergodic average is performed over an invariant torus instead of the energy surface. $$\begin{aligned} \label{sumint} & \sum_{{\boldsymbol \phi}_a \, \text{fixed}} \frac{1}{v_b \, (m v_a)^{f-1} \, {\cal S}^{(f)}} \, \left\| {\det}' \left( \frac{\partial {\boldsymbol q}_b}{\partial {\boldsymbol p}_a} \right)_\gamma \right\|^{-1} \, \delta(T- T_\gamma) \notag \\ \approx & \frac{1}{(2 \pi)^f \, {\cal S}^{(f)}} \, \sum_{{\boldsymbol I}({\boldsymbol q}_b,{\boldsymbol p}_b) = {\boldsymbol I}_a} \left\| \det \left( \frac{\partial {\boldsymbol I}({\boldsymbol q}_b,{\boldsymbol p}_b)}{\partial {\boldsymbol p}_b} \right) \right\|^{-1} \, {\text{d}}\Omega_a \, , \quad T \rightarrow \infty \; ,\end{aligned}$$ where ${\boldsymbol I}$ are the action variables of the system. This equation has the following interpretation. The left-hand side is a sum over all trajectories for which the initial angular direction lies in a solid angle element ${\text{d}}\Omega_a$ around the direction ${\boldsymbol \phi}_a$. This direction ${\boldsymbol \phi}_a$, the energy $E$ and the initial position ${\boldsymbol q}_a$ determine the action variable ${\boldsymbol I}_a={\boldsymbol I}({\boldsymbol q}_a,{\boldsymbol p}_a)$ and thus the torus on which the motion occurs. The right-hand side is a summation over all final momenta ${\boldsymbol p}_b$ for which the points $({\boldsymbol q}_b,{\boldsymbol p}_b)$ lie on this torus. In order to obtain a sum rule for all transient trajectories one has to integrate (\[sumint\]) over all initial directions. In contrast to the chaotic case this is not always possible, since the integral can be divergent due to caustics. In cases where it is possible one has $$\begin{aligned} \label{sumint2} & \sum_{\gamma} \frac{1}{v_b \, (m v_a)^{f-1} \, {\cal S}^{(f)}} \, \left\| {\det}' \left( \frac{\partial {\boldsymbol q}_b}{\partial {\boldsymbol p}_a} \right)_\gamma \right\|^{-1} \, \delta(T- T_\gamma) \notag \\ \approx & \frac{1}{(2 \pi)^f \, {\cal S}^{(f)}} \, \int \! {\text{d}}\Omega_a \, \sum_{{\boldsymbol I}({\boldsymbol q}_b,{\boldsymbol p}_b) = {\boldsymbol I}_a} \left\| \det \left( \frac{\partial {\boldsymbol I}({\boldsymbol q}_b,{\boldsymbol p}_b)}{\partial {\boldsymbol p}_b} \right) \right\|^{-1} \, , \quad T \rightarrow \infty \; .\end{aligned}$$ This sum rule can be verified, for example, for a $f$-dimensional rectangular billiard with side lengths $a_i$, $i=1,\dots,f$. There the number of transient trajectories with length in an interval ${\text{d}}L$ around $L$ is given asymptotically for large $L$ by ${\cal S}^{(f)} \, L^{f-1} \, {\text{d}}L/V$ where $V=\prod_{i=1}^n a_i$ is the volume of the billiard, and $L=v_a \, T$. The absolute value of the determinant on the left-hand-side of (\[sumint2\]) is $(L/(m v_a))^{f-1}$, the actions are $I_i=a_i \, \|p_i\|/\pi$, and the sum on the right-hand-side is $(2 \pi)^f/V$. Altogether one obtains on both sides $V^{-1}$. Incidentally the sum rule for chaotic billiard systems yields the same result in this case. The reason is that for rectangular billiards the determinant on the right-hand side of (\[sumint2\]) is the same for all tori. The sum rule (\[sumt\]) contains implicit information about the number of transient orbits. Let us assume that due to the exponential sensitivity of trajectories on initial conditions one has $\langle \|{\det}' (\partial {\boldsymbol q}_b / \partial {\boldsymbol p}_a)\|^{-1} \rangle \sim c' \, \exp(-h T)$ where $c'$ and $h$ are constants, and the averaging is performed over trajectories with $T_\gamma \approx T$. With this assumption one obtains the following asymptotic law for the number of transient orbits, $$\label{number} \rho(T) = \frac{{\text{d}}{\cal N}(T)}{{\text{d}}T} \sim c \, \exp(h T) \; , \qquad T \rightarrow \infty \; ,$$ where ${\cal N}(T)$ is the number of transient trajectories with $T_\gamma < T$, $\rho(T)$ the density of these orbits, and $c$ is a constant that is determined by $c'$. For Riemannian manifolds with no conjugate points one can show that $\lim_{T \rightarrow \infty} T^{-1} \log {\cal N}(T) = h$ is the topological entropy of the system [@BP97]. Eq. (\[number\]) with topological entropy $h$ would imply that the number of transient orbits grows by an order of $T$ stronger than the number of periodic orbits for which $\rho_{po}(T) \sim \exp(h T)/T$. One arrives at a similar conclusion for the number of transient orbits if one considers systems in which a code for these orbits exists [@BW96; @RVW96; @BD97]. For cat maps the law (\[number\]) is proved in appendix \[cats\]. Influence of diffractive orbits on spectral statistics {#secstat} ====================================================== Diffraction of quantum wave functions on singularities of the potential leads to corrections in semiclassical expansions in terms of classical trajectories. In approximations for the density of states there are additional terms besides periodic orbits. These are expressed in terms of diffractive orbits. We concentrate on cases where the source of diffraction is point-like for which some examples are given below. In these cases the diffractive orbits are trajectories that start from and return to the source of diffraction $n$ times, where $n$ is an arbitrary positive integer. They are composed of a sequence of arbitrary $n$ transient orbits of the last section for which the initial and the final points of the trajectory are identical and are located at the source of diffraction. Alternatively, they can be considered as closed trajectories that are scattered $n$ times on the source of diffraction. Within the geometrical theory of diffraction (GTD) the contribution of all diffractive orbits with $n$ scattering events to the density of states is given by [@VWR94; @PSc95; @BW96] $$\begin{aligned} \label{gtd} d^{(n)}(E) & = \frac{1}{\pi \, n} \left( \frac{\hbar^2}{2 m} \right)^n \frac{{\text{d}}}{{\text{d}}E} {\operatorname{Im}}\left[ \sum_{\gamma_1} \dots \sum_{\gamma_n} {\cal G}_{\gamma_1}(E) {\cal D}({\boldsymbol \phi}_{a,\gamma_{2}},{\boldsymbol \phi}_{b,\gamma_1}) {\cal G}_{\gamma_2}(E) {\cal D}({\boldsymbol \phi}_{a,\gamma_{3}},{\boldsymbol \phi}_{b,\gamma_2}) \right. \notag \\ & \left. \phantom{\sum_{\gamma_n}} \cdots {\cal G}_{\gamma_n}(E) {\cal D}({\boldsymbol \phi}_{a,\gamma_{1}},{\boldsymbol \phi}_{b,\gamma_n}) \right] \; ,\end{aligned}$$ where $$\label{green} {\cal G}_{\gamma_i}(E) = \frac{1}{i \hbar (2 \pi i \hbar)^{(f-1)/2} } \, \sqrt{\frac{1}{v_a \, v_b} \, \left\| {\det}' \left( \frac{\partial {\boldsymbol q}_b}{\partial {\boldsymbol p}_a} \right)_{\gamma_i} \right\|^{-1}} \, \exp \left\{ \frac{i}{\hbar} S_{\gamma_i}(E) - i \frac{\pi}{2} \nu_{\gamma_i} \right\} \; .$$ The expression (\[green\]) is the semiclassical contribution of a transient orbit $\gamma_i$ to the Green function, and $S_{\gamma_i}$ is its action and $\nu_{\gamma_i}$ the number of conjugate points along it. Furthermore, $v_a$ and $v_b$ are the initial and final velocities. They are identical since the initial and final point of the trajectory are identical, and we write them in the following without subscript. A diffractive orbit is composed of $n$ transient orbits, and it is labelled by a set of $n$ indices $\gamma_1, \dots , \gamma_n$. It is customary to consider cyclic permutations of these indices to correspond to the same diffractive orbit. The contribution of a diffractive orbit to the density of states in (\[gtd\]) can be interpreted in the following way. In the GTD approximation the diffraction is treated as a local process that occurs at the source of diffraction. The diffraction coefficient ${\cal D}({\boldsymbol \phi}_{a,\gamma_{i+1}},{\boldsymbol \phi}_{b,\gamma_i})$ contains the amplitude and phase for the scattering from the incoming direction of the transient orbits $\gamma_i$ into the outgoing direction of the transient orbit $\gamma_{i+1}$. In some sense, the diffractive trajectories can be considered as generalised periodic orbits which are closed in momentum space by the scattering at the singularity of the potential. Some examples for the diffraction coefficients are - Diffraction on a corner with angle $\theta$ in a two-dimensional billiard system [@Kel62; @Jam76]: $${\cal D}(\phi_a,\phi_b) = \frac{2}{N} \, \sin\frac{\pi}{N} \left[ \left(\cos\frac{\pi}{N} -\cos\frac{\phi_a+\phi_b}{N} \right)^{-1} - \left(\cos\frac{\pi}{N} -\cos\frac{\phi_a-\phi_b}{N} \right)^{-1} \right] \; ,$$ where $N=\theta/\pi$, and $\phi_a$ and $\phi_b$ are the angles of the outgoing and incoming trajectory, measured with respect to one side of the corner such that $\phi_a,\,\phi_b \in [0,\theta]$. - Diffraction on a flux line in a two-dimensional system [@Sie99a]: $${\cal D}(\phi_a,\phi_b) = \frac{2 \sin(\alpha \pi)}{ \cos \left(\frac{\phi_a - \phi_b}{2} \right) } \exp\left\{ i \frac{\phi_a - \phi_b}{2} \right\} \; ,$$ where $\alpha$ is the flux parameter and $\phi_a$ and $\phi_b$ are the directions of the outgoing and incoming trajectory. - Diffraction on a delta-like singularity of the potential in two dimensions [@AGHH88; @ES96]: $${\cal D} = \frac{2 \pi}{i \frac{\pi}{2} - \gamma - \log\left(\frac{k a}{2}\right)} \; ,$$ where $k=\sqrt{2 m E}/\hbar$ and $a$ is a parameter characterising the strength of the potential, and $\gamma$ is Euler’s constant. - Diffraction on a delta-like singularity of the potential in three dimensions [@AGHH88; @ES96]: $${\cal D} = \frac{4 \pi a }{1 + i k a} \; .$$ In the first two examples the diffraction coefficient is energy independent but it depends on the incoming and outgoing directions of the trajectories. There are directions in which the diffraction coefficient diverges. There the GTD approximation breaks down and has to be replaced by a uniform approximation [@SPS97; @Sie99a]. In the last two cases, the diffraction coefficient has no angular dependence, but depends on energy. It corresponds to pure $s$-wave scattering, and the GTD approximation is valid for all angular directions. In the following we consider the influence of diffractive orbits on semiclassical approximations for the spectral form factor which is defined by $$K(\tau) = \int_{-\infty}^{\infty} \! \frac{{\text{d}}\eta}{\bar{d}(E)} \; \left\langle d_{\text{osc}}\left( E + \frac{\eta}{2} \right) d_{\text{osc}}\left( E - \frac{\eta}{2} \right) \right\rangle_E \; \exp\left( 2 \pi i \eta \tau \bar{d}(E) \right) \; ,$$ where $\langle \dots \rangle_E$ denotes an average over an energy interval that is small in comparison with $E$ but contains many energy levels. If the oscillatory part of the density of states is semiclassically approximated by classical trajectories in the form $$\label{density} d_{\text{osc}}(E) \approx \sum_\gamma A_\gamma \exp\left( \frac{i}{\hbar} S_\gamma(E) \right) \; ,$$ then the spectral form factor is expressed by a double sum over trajectories. We consider here only the diagonal approximation to this double sum that describes the form factor for small values of $\tau$. In leading semiclassical order $$K_d(\tau) = \frac{2 \pi \hbar}{\bar{d}(E)} \sum_\gamma \, g_\gamma \, \|\bar{A}_\gamma\|^2 \, \delta(T-T_\gamma) \; ,$$ where $T=2 \pi \hbar \bar{d}(E) \tau$ and $g_\gamma$ is the number of terms in the sum for which the actions $S_\gamma(E)$ are identical. $\bar{A}_\gamma$ is the average over all $g_\gamma$ amplitudes of the orbits which have the same action as the orbit $\gamma$ in case that these amplitudes are different. Let us first consider the case of single-diffractive orbits $n=1$. Then $$A_\gamma = \frac{T_\gamma \, {\cal D}({\boldsymbol \phi}_{a,\gamma},{\boldsymbol \phi}_{b,\gamma}) }{4 \pi m v (2 \pi \hbar)^{(f-1)/2}} \sqrt{\left\| {\det}' \left( \frac{\partial {\boldsymbol q}_b}{\partial {\boldsymbol p}_a} \right)_\gamma \right\|^{-1}} \; ,$$ and one obtains for the contribution of diffractive orbits with one scattering event to the diagonal approximation of the form factor $$K^{(1)}_d(\tau) = \frac{2 \pi \hbar (2/\beta)}{\bar{d}(E)} \sum_{\gamma} \frac{T_\gamma^2 \, \|{\cal D}({\boldsymbol \phi}_{a,\gamma},{\boldsymbol \phi}_{b,\gamma})\|^2 }{(4 \pi m v)^2 \, (2 \pi \hbar)^{f-1}} \, \left\| {\det}' \left( \frac{\partial {\boldsymbol q}_b}{\partial {\boldsymbol p}_a} \right)_\gamma \right\|^{-1} \, \delta(T - T_{\gamma}) \; .$$ The degeneracy of the actions is in this case $g=2/\beta$ (except for an exponentially small fraction of the orbits for large $T$) where $\beta$ is an integer denoting the symmetry class of the system. $\beta=1$ for systems in which the only symmetry is an anti-unitary symmetry, and $\beta=2$ for systems without any symmetry. Applying the sum rule (\[sumpart\]) and the leading asymptotic approximation for the mean density of states $\bar{d}(E) \sim \Sigma(E)/(2 \pi \hbar)^f$ one obtains $${\text{d}}^2 K^{(1)}_d(\tau) = \frac{1}{8 \beta \pi^2} \, \left( \frac{m v}{2 \pi \hbar} \right)^{2 f-4} \, \|{\cal D}({\boldsymbol \phi}_a,{\boldsymbol \phi}_b)\|^2 \, \tau^2 \, {\text{d}}\Omega_a {\text{d}}\Omega_b$$ for the partial contribution to $K^{(1)}_d(\tau)$ from diffractive orbits with fixed initial and final directions. For the examples given above this result has the following implications. For the corner diffraction and diffraction on a flux line the prefactor of $\tau^2$ is independent of $\hbar$. It implies that diffraction has an influence on statistical properties of energy levels in the semiclassical limit $\hbar \rightarrow 0$. Although the contributions of diffractive orbits to the density of states are by an order $\sqrt{\hbar}$ smaller than those of periodic orbits, they still give a finite contribution to the diagonal approximation of the spectral form factor as $\hbar \rightarrow 0$. The complete contribution to $K_d^{(1)}$ that is obtained by integrating over all angular directions ${\boldsymbol \phi}_a$ and ${\boldsymbol \phi}_b$ requires the use of uniform approximations, since the GTD approximation yields a divergent result. However, since all contributions to $K_d^{(1)}$ are positive there is a non-vanishing contribution of order $\tau^2$ to the diagonal approximation of the form factor in the semiclassical limit. For the examples of diffraction on delta-singularities the diffraction is isotropic and the integration over the angular directions can be performed as is done below. Here the prefactor of $\tau^2$ is energy dependent, and there is a significant difference between the two- and three-dimensional result. In two dimensions the prefactor vanishes in the semiclassical limit, although very slowly like $(\log \hbar)^{-2}$, whereas in three dimensions the prefactor approaches a constant. The contributions of diffractive orbits with $n$ scattering events to $K_d(\tau)$ are obtained in a similar way. They are composed of $n$ transient orbits and it is sufficient to consider only cases in which all these $n$ orbits are different, since the relative number of the other cases is exponentially suppressed for large times $T$. The degeneracy of the actions of diffractive orbits is then $g=(2/\beta)^n \, n!$. The index $\gamma$ in (\[density\]) is now a multiple index $\gamma = (\gamma_1,\dots,\gamma_n)$ and $$\bar{A}_\gamma = \frac{T_\gamma}{2 \pi \hbar n} \, \left( \frac{\hbar}{2 m v (2 \pi \hbar)^{(f-1)/2}} \right)^n \overline{\cal D}^{(n)}({\boldsymbol \phi}_{a,\gamma_1},{\boldsymbol \phi}_{b,\gamma_1}, \dots,{\boldsymbol \phi}_{a,\gamma_n},{\boldsymbol \phi}_{b,\gamma_n}) \prod_{i=1}^n \sqrt{\left\| {\det}' \left( \frac{\partial {\boldsymbol q}_b}{\partial {\boldsymbol p}_a} \right)_{\gamma_i} \right\|^{-1}} \; ,$$ where $T_\gamma = \sum_{i=1}^n T_{\gamma_i}$, and $$\overline{\cal D}^{(n)}({\boldsymbol \phi}_{a,\gamma_1},{\boldsymbol \phi}_{b,\gamma_1}, \dots,{\boldsymbol \phi}_{a,\gamma_n},{\boldsymbol \phi}_{b,\gamma_n}) = \left\langle \prod_{i=1}^n {\cal D}({\boldsymbol \phi}_{a,\gamma_{i+1}},{\boldsymbol \phi}_{b,\gamma_i}) \right\rangle$$ is the average over all $n!$ permutations of the $\gamma_i$. For simplicity of notation we abbreviate the argument of $\overline{\cal D}^{(n)}$ by an index $\gamma$. Using the relation $$\delta(T - T_\gamma) = \int_0^\infty \! {\text{d}}T_1 \cdots {\text{d}}T_n \; \left[ \prod_{i=1}^n \delta(T_i - T_{\gamma_i}) \right] \, \delta( T - \sum_{i=1}^{n} T_i) \; ,$$ one can write the contribution of $n$-fold diffractive orbits to the diagonal approximation of the form factor in the form $$\begin{aligned} K_d^{(n)}(\tau) & = \frac{T^2 (2/\beta)^n n!}{2 \pi \hbar n^2 \bar{d}(E)} \, \int_0^\infty \! {\text{d}}T_1 \cdots {\text{d}}T_n \; \delta( T - \sum_{i=1}^{n} T_i) \, \times \notag \\ & \qquad \sum_\gamma \|\overline{\cal D}_\gamma^{(n)}\|^2 \prod_{i=1}^n \left[ \frac{1}{(4 \pi m v_a)^2 (2 \pi \hbar)^{f-3}} \, \left\| {\det}' \left( \frac{\partial {\boldsymbol q}_b}{\partial {\boldsymbol p}_a} \right)_{\gamma_i} \right\|^{-1} \, \delta(T_i - T_{\gamma_i}) \right] \; .\end{aligned}$$ After using sum rule (\[sumpart\]) $n$ times one is left with the integrals over the $T_i$ which yield $$\int_0^\infty \! {\text{d}}T_1 \cdots {\text{d}}T_n \; \delta( T - \sum_{i=1}^{n} T_i) = \frac{T^{n-1}}{(n-1)!} \; ,$$ which can be shown by induction, and the final result is $${\text{d}}^{2 n} K^{(n)}_d(\tau) = \frac{\tau^{n+1}}{n} \, \left( \frac{1}{8 \beta \pi^2} \, \left( \frac{m v}{2 \pi \hbar} \right)^{2 f-4} \, \right)^n \, \|\overline{\cal D}^{(n)}({\boldsymbol \phi}_{a,1},\dots,{\boldsymbol \phi}_{b,n})\|^2 \, {\text{d}}\Omega_{a,1} \cdots {\text{d}}\Omega_{b,n} \; .$$ The interpretation is similar as before. The $n$-fold diffractive orbits contribute to $K_d(\tau)$ in order $\tau^{n+1}$. For corner diffraction and diffraction on a flux line the prefactor of $\tau^{n+1}$ is independent of $\hbar$ and persists in the semiclassical limit. The remarkable point is that the order in $\hbar$ of the contributions of these orbits to the density of states can be arbitrarily large, their amplitude is by an order $\hbar^{n/2}$ smaller than those of periodic orbits, but they still give a finite contribution to the form factor in the semiclassical limit! For the diffraction on delta-singularities of the potential, the prefactor of $\tau^{n+1}$ again vanishes slowly in the two-dimensional case and approaches a constant in three dimensions as $\hbar \rightarrow 0$. For the isotropic case where the diffraction coefficient is angular independent, the integration over the angular directions can be performed, and the contributions of all diffractive orbits can be summed. Adding also the contribution of periodic orbits, the complete expression for the diagonal form factor is $$K_d(\tau) = \frac{2}{\beta} \tau + \sum_{n=1}^\infty \frac{\tau^{n+1}}{n} {\cal C}^n = \frac{2}{\beta} \tau - \tau \, \log (1-{\cal C} \tau) \; , \qquad \qquad \tau < {\cal C}^{-1} \; ,$$ where $${\cal C} = \frac{\|{\cal D}\|^2 \, [{\cal S}^{(f)}]^2}{8 \beta \pi^2} \, \left( \frac{m v}{2 \pi \hbar} \right)^{2 f-4} \, .$$ For an $s$-wave scatterer in a three-dimensional billiard system the constant approaches the value ${\cal C}=8$ in the limit $\hbar \rightarrow 0$. The results for the contributions of diffractive orbits to the spectral form factor can also be applied to a rectangular billiard with an $s$-wave scatterer, since the same sum-rule applies in this case. The form factor for this system has been studied in [@DV98]. Conclusions =========== The main result of this article is that diffraction can have an influence on spectral statistics in the semiclassical limit. The semiclassical treatment of diffraction leads to corrections to the diagonal approximation for the spectral form factor $K(\tau)$ that do not vanish in the semiclassical limit $\hbar \rightarrow 0$. This is in contrast, for example, to corrections due to bouncing ball orbits, which vanish if the semiclassical limit is performed while keeping the argument of $K(\tau)$ fixed. Although the corrections to the form factor do not agree with those expected from random matrix theory, the present results do not show that diffraction leads to a deviation from random matrix statistics, since off-diagonal contributions to the form factor have been neglected. They show, however, that a semiclassical proof of the random matrix hypothesis cannot be based on periodic orbits alone if diffraction occurs. Consider, for example, chaotic billiard systems with flux lines. These systems are considered to be standard examples in which spectral correlations are expected to follow the statistics of the Gaussian unitary ensemble (GUE). The diagonal approximation for the form factor in terms of periodic orbits already yields the correct linear GUE form factor up to the Heisenberg time $\tau=1$ [@Ber85]. Moreover, it can be shown that in certain ensemble averages the off-diagonal contributions of periodic orbits vanish [@FK98]. The present results show, however, that diffraction on the flux line leads to additional semiclassical contributions to the diagonal form factor in all higher orders of $\tau$. If the spectral statistics follow indeed random matrix theory, then these additional contributions have to be cancelled by off-diagonal terms involving diffractive orbits. Such a cancellation would require specific correlations between classical trajectories that start from and return to one point in coordinate space, or between these trajectories and periodic orbits, analogous to the action correlations of periodic orbits [@ADDKKSS93]. For another example let us assume that the spectral correlations in a particular three-dimensional chaotic system are described correctly by random matrix theory in the semiclassical limit. Then it is generally expected that this property is not changed, if an $s$-wave scatterer is added to the system. As is shown in the present article, however, the diagonal approximation for $K(\tau)$ is modified also in this case. This would imply a deviation from random matrix results, if this is not corrected by the off-diagonal terms. Nevertheless, it shows that a single $s$-wave scatterer [can]{} lead to deviations from random matrix statistics. The analysis in this article is based on the geometrical theory of diffraction. For corner diffraction or diffraction on a flux line this is not sufficient for calculating the complete diagonal approximation of the form factor. It would require the application of uniform approximations. Since the semiclassical weight of diffractive orbits is larger in the uniform regime, one can expect an even stronger total influence of diffraction on the spectral form factor in these cases. The main remaining question is, whether deviations from random matrix statistics can be observed in these systems. [Acknowledgement]{} It is a pleasure to thank R. Artuso, E. Bogomolny, S. Fishman, E. Gutkin, J. Keating, and U. Smilansky for informative and helpful discussions on this subject. Cat maps {#cats} ======== In this appendix we give a simple example for the sum rule for transient orbits. Cat maps are linear, area-preserving, hyperbolic maps on the unit 2-torus $$\begin{pmatrix} q_{n+1} \\ p_{n+1} \end{pmatrix} = M \, \begin{pmatrix} q_n \\ p_n \end{pmatrix} \; \mod 1 \; ,$$ where $\det M = 1$ due to area preservation, $\|{\operatorname{Tr}}M\| > 2$ due to hyperbolicity, and the elements $M_{ij}$ of the matrix $M$ are integers for continuity. The sum rule (\[summap\]) has the form $$\label{sumcat} \sum_{(q_n,q_0)=(q_b,q_a)} \frac{1}{\|(M^n)_{12}\|} = 1 \; ,$$ where the sum extends over all points $(q_0,p_0)$ for which $q_0=q_a$ and $q_n=q_b$. The matrix element $(M^n)_{12}$ is the same for all points and can be taken in front of the sum. The number of points over which the sum extends is given by the number of solutions of the equation $$q_b = (M^n)_{11} \, q_a + (M^n)_{12} \, p_a \; \mod 1 \; .$$ Since $p_a$ varies in the interval $[0,1)$ there are exactly $\|(M^n)_{12}\|$ solutions of this equation. This shows that the sum rule (\[sumcat\]) is exact for all $n>0$. Let $\|{\operatorname{Tr}}M\| = 2 \cosh(\lambda)$ where $\lambda>0$ is the Lyapunov exponent of the map, and $\sigma = \operatorname{sign}({\operatorname{Tr}}M)$. Then $$M^n = \frac{1}{u_1 s_2 - u_2 s_1} \, \begin{pmatrix} u_1 & s_1 \\ u_2 & s_2 \end{pmatrix} \begin{pmatrix} \sigma^n e^{n \lambda} & 0 \\ 0 & \sigma^n e^{-n \lambda} \end{pmatrix} \begin{pmatrix} s_2 & -s_1 \\ -u_2 & u_1 \end{pmatrix} \; ,$$ where $(u_1,u_2)$ and $(s_1,s_2)$ are the components of the unstable and stable eigenvectors of $M$, respectively. One obtains $(M^n)_{12} = - \sigma^n u_1 s_1 (e^{n \lambda} + e^{-n \lambda})/(u_1 s_2 - u_2 s_1)$ and it follows that the number of points contributing to the sum rule is given by $$\rho(n) \sim \frac{\|u_1 \, s_1\|}{\|u_1 s_2 - u_2 s_1\|} \, e^{n \lambda} \; , \qquad n \rightarrow \infty \, .$$ This is of the form (\[number\]) and it shows, moreover, that the constant $c$ in (\[number\]) is not universal. For cat maps it is determined by the directions of the eigenvectors of $M$. [10]{} O. Bohigas, M. J. Giannoni and C. Schmit: [Characterization of Chaotic Quantum Spectra and Universality of Level Fluctuation Laws]{}, Phys. Rev. Lett. [52]{} (1984) 1–4. O. Bohigas: [Random Matrix Theories and Chaotic Dynamics]{}, in: [Les Houches 1989 Session [LII]{} on [Chaos and Quantum Physics]{}]{} (Eds. M. J. Giannoni, A. Voros and J. Zinn-Justin), 87–199. North-Holland, Amsterdam, (1991). M. V. Berry: [Semiclassical Theory of Spectral Rigidity]{}, Proc. R. Soc. Lond. A [400]{} (1985) 229–251. E. Bogomolny and J. Keating: [Gutzwiller’s Trace Formula and Spectral Statistics: Beyond the Diagonal Approximation]{}, Phys. Rev. Lett. [77]{} (1996) 1472–1475. O. Agam, B. L. Altshuler and A. V. Andreev: [Spectral Statistics: From Disordered to Chaotic Systems]{}, Phys. Rev. Lett. [75]{} (1995) 4389–4392. J. H. Hannay and A. M. [Ozorio de Almeida]{}: [Periodic Orbits and a Correlation Function for the Semiclassical Density of States]{}, J. Phys. A [17]{} (1984) 3429–3440. N. Argaman, F. Dittes, E. Doron, J. Keating, A. Kitaev, M. Sieber and U. Smilansky: [Correlations in the Actions of Periodic Orbits Derived from Quantum Chaos]{}, Phys. Rev. Lett. [71]{} (1993) 4326–4329. M. C. Gutzwiller: [Chaos in Classical and Quantum Mechanics]{}, Springer, New York, (1990). G. Vattay, A. Wirzba and P. E. Rosenqvist: [Periodic Orbit Theory of Diffraction]{}, Phys. Rev. Lett. [73]{} (1994) 2304–2307. N. Pavloff and C. Schmit: [Diffractive Orbits in Quantum Billiards]{}, Phys. Rev. Lett. [75]{} (1995) 61–64, Erratum in Phys. Rev. Lett. [ 75]{} (1995) 3779. H. Bruus and N. D. Whelan: [Edge Diffraction, Trace Formulae and the Cardioid Billiard]{}, Nonlinearity [9]{} (1996) 1023–1047. H. Primack, H. Schanz, U. Smilansky and I. Ussishkin: [Penumbra Diffraction in the Quantization of Dispersing Billiards]{}, Phys. Rev. Lett. [76]{} (1996) 1615–1618. K. Burns and G. P. Paternain: [Counting Geodesics on a [R]{}iemannian Manifold and Topological Entropy of Geodesic Flows]{}, Ergod. Theor. Dyn. Syst. [17]{} (1997) 1043–1059. P. E. Rosenqvist, G. Vattay and A. Wirzba: [Application of the Diffraction Trace Formula to the Three-Disk Scattering System]{}, J. Stat. Phys. [83]{} (1996) 243–257. A. Bäcker and H. R. Dullin: [Symbolic Dynamics and Periodic Orbits for the Cardioid Billiard]{}, J. Phys. A [30]{} (1997) 1991–2020. J. B. Keller: [Geometrical Theory of Diffraction]{}, J. Opt. Soc. Am. [ 52]{} (1962) 116–130. G. L. James: [Geometrical Theory of Diffraction for Electromagnetic Waves]{}, Peregrinus, Stevenage, England, (1976). M. Sieber: [Semiclassical Treatment of Diffraction in Billiard Systems with a Flux Line]{}, to appear in Phys. Rev. E. S. Albeverio, F. Gesztesy, R. H[ø]{}egh-Krohn and H. Holden: [Solvable Models in Quantum Mechanics]{}, Springer, New York, (1988). P. Exner and P. Šeba: [Point Interactions in Two and Three Dimensions as Models of Small Scatterers]{}, Phys. Lett. A [222]{} (1996) 1–4. M. Sieber, N. Pavloff and C. Schmit: [Uniform Approximation for Diffractive Contributions to the Trace Formula in Billiard Systems]{}, Phys. Rev. E [ 55]{} (1997) 2279–2299. P. Dahlqvist and G. Vattay: [Periodic Orbit Quantization of the Sinai Billiard in the Small Scatterer Limit]{}, J. Phys. A [31]{} (1998) 6333–6345. S. Fishman and J. P. Keating: [The Diagonal Approximation for Non-Time-Reversal-Symmetric Systems]{}, J. Phys. A [31]{} (1998) L313–L319. [^1]: E-mail: [email protected]	{ "pile_set_name": "ArXiv" }
--- abstract: 'This paper provides a critical discussion of the observational evidence for winds in our own Galaxy, in nearby star-forming and active galaxies, and in the high-redshift universe. The implications of galactic winds on the formation and evolution of galaxies and the intergalactic medium are briefly discussed. A number of observational challenges are mentioned to inspire future research directions.' author: - Sylvain Veilleux title: \| Multiwavelength Observations of Galactic Winds:\ Near and Far --- \#1[[\#1]{}]{} \#1[[\#1]{}]{} = \#1 1.25in .125in .25in Introduction ============ Active galactic nuclei (AGN) and nuclear starbursts may severely disrupt the gas phase of galaxies through deposition of a large amount of mechanical energy in the centers of galaxies. As a result, a large-scale galactic wind that encompasses much of the central regions of these galaxies may be created (e.g., Chevalier & Clegg 1985; Schiano 1985). Depending upon the extent of the gaseous halo and its density and upon the wind’s mechanical luminosity and duration, the wind may ultimately blow out through the halo and into the intergalactic medium. The effects of these winds may be far-reaching. Bregman (1978) has suggested that the Hubble sequence can be understood in terms of a galaxy’s greater ability to sustain winds with increasing bulge-to-disk ratio. Galactic winds may affect the thermal and chemical evolution of galaxies and the intergalactic medium by depositing large quantities of hot, metal-enriched material on the outskirts of galaxies and beyond. This widespread circulation of matter and energy between the disks and halos of galaxies may be responsible for the mass-metallicity relation between galaxies. This paper reviews the observed properties (§2) and impact (§3) of starburst- and AGN-driven winds in both the local and distant universe; the discussion on starburst-driven winds is largely borrowed from Veilleux (2003), while new elements on AGN outflows are also included. The last section (§4) discusses future avenues of research. The theory and numerical modelling of galactic winds are not discussed here due to space limitations (see, e.g., Veilleux et al. 2002; Strickland 2002; Heckman 2002; Veilleux 2003). Collimated jet outflows and unresolved nuclear winds in AGNs are also beyond the scope of this paper; recent reviews of these topics include Zensus (1997), Veilleux et al. (2002), and Crenshaw, Kraemer, & George (2003). Observed Properties of Galactic Winds ===================================== AGN- and starburst-driven winds are common among local galaxies. Galaxies with global star formation rates per unit area $\Sigma_* \equiv SFR / \pi R_{\rm opt}^2 \ga 0.1$ M$_\odot$ yr$^{-1}$ kpc$^{-2}$, where $R_{\rm opt}$ is the optical radius, often show signs of large-scale winds. This general rule-of-thumb also appears to apply to ultra/luminous infrared galaxies (see §2.3) and distant Lyman break galaxies (see §2.4). “Quiescent” galaxies with lower star formation rates per unit area often show thick ionized disks, but no galactic-scale outflow (e.g., Miller & Veilleux 2003a, 2003b). This rule-of-thumb is conservative since a number of known starburst-driven wind galaxies, including our own Galaxy (§2.1) and several dwarf galaxies, have $\Sigma_* <<$ 0.1 M$_\odot$ yr$^{-1}$ kpc$^{-2}$ (e.g., Hunter & Gallagher 1990, 1997; Meurer et al. 1992; Marlowe et al. 1995; Kunth et al. 1998; Martin 1998, 1999). The production of detectable winds depends not only on the characteristics of the energy source (AGN vs starburst, power, age), but also on the detailed properties of the ISM in the host galaxies (e.g., see the theoretical blowout criterion of MacLow & McCray 1988). The winds in active and star-forming galaxies in the local universe show a very broad range of properties, with opening angles of $\sim$ 0.1 – 0.5 $\times$ (4$\pi$ sr), radii ranging from $<$ 1 kpc to several 10s of kpc, outflow velocities of a few 10s of km s$^{-1}$ to more than 1000 km s$^{-1}$ (with clear evidence for a positive correlation with the temperature of the gas phase), total (kinetic and thermal) outflow energies of $\sim$ 10$^{53}$ – 10$^{57}$ ergs, and mass outflow rates ranging from $<$ 1 M$_\odot$ yr$^{-1}$ to $>$ 100 M$_\odot$ yr$^{-1}$. In AGNs, the mass outflow rates on kpc scale are larger than the mass accretion rates needed to power the central nucleus. In starbursts, the mass outflow rates scale roughly with the star formation rates (see §2.3 below). In the remainder of this section, we repeat the discussion of Veilleux (2003) on a few well-studied cases of galactic winds in the local universe and summarize the evidence for winds in luminous and ultraluminous infrared galaxies at low and moderate redshifts as well as in distant Lyman break galaxies. The Milky Way ------------- By far the closest case of a large-scale outflow is the wind in our own Galaxy. Evidence for a dusty bipolar wind extending $\sim$ 350 pc ($\sim$ 1$^\circ$) above and below the disk of our Galaxy has recently been reported by Bland-Hawthorn & Cohen (2003) based on data from the Midcourse Space Experiment (MSX). The position of the warm dust structure coincides closely with the well-known Galactic Center Lobe detected at radio wavelengths (e.g., Sofue 2000 and references therein). Simple arguments suggest that the energy requirement for this structure is of order $\sim$ 10$^{55}$ ergs with a dynamical time scale of $\sim$ 1 Myr. Bland-Hawthorn & Cohen (2003) also argue that the North Polar Spur, a thermal X-ray/radio loop that extends from the Galactic plane to $b = +80^\circ$ (e.g., Sofue 2000), can naturally be explained as an open-ended bipolar wind, when viewed in projection in the near field. This structure extends on a scale of 10 – 20 kpc and implies an energy requirement of $\sim$ (1 – 30) $\times$ 10$^{55}$ ergs and a dynamical timescale of $\sim$ 15 Myr, i.e. considerably longer than that of the smaller structure seen in the MSX maps. If confirmed, this may indicate that the Milky Way Galaxy has gone through multiple galactic wind episodes. Bland-Hawthorn & Cohen (2003) point out that the North Polar Spur would escape detection in external galaxies; it is therefore possible that the number of galaxies with large-scale winds has been (severely?) underestimated. Nearby Starburst Galaxies ------------------------- Two classic examples of starburst-driven outflows are described in this section to illustrate the wide variety of processes taking place in these objects. 0.1in [M 82.]{} This archetype starburst galaxy hosts arguably the best studied galactic wind. Some of the strongest evidence for the wind is found at optical wavelengths, where long-slit and Fabry-Perot spectroscopy of the warm ionized filaments above and below the disk shows line splittings of up to $\sim$ 250 km s$^{-1}$, corresponding to deprojected velocities of order 525 – 655 km s$^{-1}$ (e.g., McKeith et al. 1995; Shopbell & Bland-Hawthorn 1998). Combining these velocities with estimates for the ionized masses of the outflowing filamentary complex, the kinetic energy involved in the warm ionized outflow is of order $\sim$ 2 $\times$ 10$^{55}$ ergs or $\sim$ 1% of the total mechanical energy input from the starburst. The ionized filaments are found to lie on the surface of cones with relatively narrow opening angles ($\sim$ 5 – 25$^\circ$) slightly tilted ($\sim$ 5 – 15$^\circ$) with respect to the spin axis of the galaxy. Deep narrow-band images of M82 have shown that the outflow extends out to at least 12 kpc on one side (e.g., Devine & Bally 1999), coincident with X-ray emitting material seen by $ROSAT$ (Lehnert, Heckman, & Weaver 1999) and $XMM$-Newton (Stevens, Read, & Bravo-Guerrero 2003). The wind fluid in this object has apparently been detected by both $CXO$ (Griffiths et al. 2000) and $XMM$-Newton (Stevens et al. 2003). The well-known H I complex around this system (e.g., Yun et al. 1994) may be taking part, and perhaps even focussing, the outflow on scales of a few kpc (Stevens et al. 2003). Recently published high-quality CO maps of this object now indicate that some of the molecular material in this system is also involved in the large-scale outflow (Walter, Weiss, & Scoville 2002; see also Garcia-Burillo et al. 2001). The outflow velocities derived from the CO data ($\sim$ 100 km s$^{-1}$ on average) are considerably lower than the velocities of the warm ionized gas, but the mass involved in the molecular outflow is substantially larger ($\sim$ 3 x 10$^8$ M$_\odot$), implying a kinetic energy ($\sim$ 3 $\times$ 10$^{55}$ ergs) that is comparable if not larger than that involved in the warm ionized filaments. The molecular gas is clearly a very important dynamical component of this outflow. 0.1in [NGC 3079. ]{} An outstanding example of starburst-driven superbubble is present in the edge-on disk galaxy, NGC 3079. High-resolution $HST$ H$\alpha$ maps of this object show that the bubble is made of four separate bundles of ionized filaments (Cecil et al. 2001). The two-dimensional velocity field of the ionized bubble material derived from Fabry-Perot data (Veilleux et al. 1994) indicates that the ionized bubble material is entrained in a mushroom vortex above the disk with velocities of up to $\sim$ 1500 km s$^{-1}$ (Cecil et al. 2001). A recently published X-ray map obtained with the $CXO$ (Cecil, Bland-Hawthorn, & Veilleux 2002) reveals excellent spatial correlation between the hot X-ray emitting gas and the warm optical line-emitting material of the bubble, suggesting that the X-rays are being emitted either as upstream, standoff bow shocks or by cooling at cloud/wind conductive interfaces. This good spatial correlation between the hot and warm gas phases appears to be common in galactic winds (Strickland et al. 2000, 2002; Veilleux et al. 2003, and references therein). The total energy involved in the outflow of NGC 3079 appears to be slightly smaller than that in M 82, although it is a lower limit since the total extent of the X-ray emitting material beyond the nuclear bubble of NGC 3079 is not well constrained (Cecil et al. 2002). Contrary to M 82, the hot wind fluid that drives the outflow in NGC 3079 has not yet been detected, and evidence for entrained molecular gas is sparse and controversial (e.g., Irwin & Sofue 1992; Baan & Irwin 1995; Israel et al. 1998; but see Koda et al. 2002). Luminous and Ultraluminous Infrared Galaxies. --------------------------------------------- Given that the far-infrared energy output of a (dusty) galaxy is a direct measure of its star formation rate, it is not surprising [a posteriori]{} to find evidence for large-scale galactic winds in several starburst-dominated luminous and ultraluminous infrared galaxies (LIRGs and ULIRGs; e.g., Heckman et al. 1990; Veilleux et al. 1995). Systematic searches for winds have been carried out in recent years in these objects to look for the unambiguous wind signature of blueshifted absorbing material in front of the continuum source (Heckman et al. 2000; Rupke et al. 2002). The feature of choice to search for outflowing neutral material in galaxies of moderate redshifts ($z \la$ 0.6) is the Na ID interstellar absorption doublet at 5890, 5896 Å. The wind detection frequency derived from a set of 44 starburst-dominated LIRGs and ULIRGs is high, of order $\sim$ 70 – 80% (Rupke et al. 2002, 2003 in prep.). The outflow velocities reach values in excess of 1700 km s$^{-1}$ (even more extreme velocities are found in some AGN-dominated ULIRGs; e.g., Mrk 231; Rupke et al. 2002). A simple model of a mass-conserving free wind (details of the model are given in Rupke et al. 2002) is used to infer mass outflow rates in the range $\dot{M}_{\mathrm{tot}}$(H)$\;= {\mathrm few} - 120\;$ for galaxies hosting a wind. These values of $\dot{M}_{\mathrm{tot}}$, normalized to the corresponding global star formation rates inferred from infrared luminosities, are in the range $\eta \equiv \dot{M}_{\mathrm{tot}} / \mathrm{SFR} = 0.01 - 1$. The parameter $\eta$, often called the “mass entrainment efficiency” or “reheating efficiency” shows no dependence on the mass of the host (parameterized by host galaxy kinematics and absolute $R$- and $K^{\prime}$-band magnitudes), but there is a possible tendency for $\eta$ to decrease with increasing infrared luminosities (i.e. star formation rates). The large molecular gas content in ULIRGs may impede the formation of large-scale winds and reduce $\eta$ in these objects. A lower thermalization efficiency (i.e. higher radiative efficiency) in these dense gas-rich systems may also help explain the lower $\eta$ (Rupke et al. 2003, in prep.). Lyman Break Galaxies -------------------- Evidence for galactic winds has now been found in a number of z $\sim$ 3 – 5 galaxies, including an important fraction of Lyman break galaxies (LBGs; e.g., Franx et al. 1997; Pettini et al. 2000, 2002; Frye, Broadhurst, & Benitez 2002; Dawson et al. 2002; Ajiki et al. 2002; Adelberger et al. 2003; Shapley et al. 2003). The best studied wind at high redshift is that of the gravitationally lensed LBG MS 1512-cB58 (Pettini et al. 2000, 2002). An outflow velocity of $\sim$ 255 km s$^{-1}$ is derived in this object, based on the positions of the low-ionization absorption lines relative to the rest-frame optical emission lines (Ly$\alpha$ is to be avoided for this purpose since resonant scattering and selective dust absorption of the Ly$\alpha$ photons may severely distort the profile of this line; e.g., Tenorio-Tagle et al. 1999). The mass-conserving free wind model of Rupke et al. (2002) applied to MS 1512-cB58 (for consistency) results in a mass outflow rate of $\sim$ 20 $M_\odot$ yr$^{-1}$, equivalent to about 50% the star formation rate of this galaxy based on the dust-corrected UV continuum level. Similar outflow velocities are derived in other LBGs (Pettini et al. 2001). The possibly strong impact of these LBG winds on the environment at high $z$ is discussed in the next section (§3.2). Impact of Galactic Winds on the Environment =========================================== Due to space limitations, it is not possible to discuss in detail the profound influence of galactic winds on galaxy formation and evolution and on the properties of the intergalactic medium. This section reviews a few key results on the heating and enrichment of the ISM and IGM, and describe new optical constraints on the size of the zone of influence of galactic winds. Heating and Enrichment of the ISM and IGM ----------------------------------------- [Hot Metal-Enriched Gas in Starburst-Driven Winds.]{} Nuclear starbursts inject both mechanical energy and metals in the centers of galaxies. This hot, chemically-enriched material is eventually deposited on the outskirts of the host galaxies, and contributes to the heating and metal enrichment of galaxy halos and the IGM. Surprisingly little evidence exists for the presence of this enriched wind fluid. This is due to the fact that the wind fluid is tenuous and hot and therefore very hard to detect in the X-rays. The current best evidence for the existence of the wind fluid is found in M 82 (Griffiths et al. 2000; Stevens et al. 2003), NGC 1569 (Martin, Kobulnicky, & Heckman 2002), and possibly the Milky Way (e.g., Koyama et al. 1989; Yamauchi et al. 1990). The ratio of alpha elements to iron appears to be slightly super-solar in the winds of both NGC 1569 and M 82, as expected if the stellar ejecta from SNe II are providing some, but not all of the wind fluid. 0.1in [Selective Loss of Metals in Starburst-Driven Winds. ]{} The outflow velocities in starburst-dominated LIRGs and ULIRGs do not appear to be correlated with the rotation velocity (or equivalently, the escape velocity) of the host galaxy, implying selective loss of metal-enriched gas from shallower potentials (Heckman et al. 2000; Rupke et al. 2002). If confirmed over a broader range of galaxy masses (e.g., Martin 1999; but see Martin 2003 and Rupke et al. 2003, in prep.), this result may help explain the mass-metallicity relation and radial metallicity gradients in elliptical galaxies and galaxy bulges and disks (e.g., Bender, Burstein, & Faber 1993; Franx & Illingworth 1990; Carollo & Danziger 1994; Zaritsky et al. 1994; Trager et al. 1998). The ejected gas may also contribute to the heating and chemical enrichment of the ICM in galaxy clusters (e.g., Dupke & Arnaud 2001; Finoguenov et al. 2002, and references therein). 0.1in [Heating by AGN-Driven Outflows. ]{} The large “cavities” in the X-ray surface brightness of several cooling flow clusters with radio-loud cD galaxies (e.g., Böhringer et al. 1993; Fabian et al. 2000; McNamara et al. 2000, 2001; David et al. 2001; Heinz et al. 2002) point to the direct influence of AGN-driven outflows on the ICM. The hot/relativistic buoyant gas injected into the ICM by the AGN reduces and perhaps even quenches the mass accretion rates associated with the cooling flows, possibly through thermal conduction or “effervescent” heating (e.g., Quilis, Bower, & Balogh 2001; Churazov et al. 2002; Ruszkowski & Begelman 2003; see Kim & Narayan 2003, however). 0.1in [Dust Outflows.]{} Galactic winds also act as conveyor belts for the dust in the hosts. The evidence for a large-scale dusty outflow in our own Galaxy has already been mentioned in §2.1 (Bland-Hawthorn & Cohen 2003). Far-infrared maps of external galaxies with known galactic winds show extended dust emission along the galaxy minor axis, suggestive of dust entrainment in the outflow (e.g., Hughes, Gear, & Robson 1994; Alton et al. 1998, 1999; Radovich, Kahanpää, & Lemke 2001). Direct evidence is also found at optical wavelengths in the form of elevated dust filaments in a few galaxies (e.g., NGC 1808, Phillips 1993; NGC 3079, Cecil et al. 2001). A strong correlation between color excesses, $E(B - V)$, and the equivalent widths of the blueshifted low-ionization lines in star-forming galaxies at low (e.g., Armus, Heckman, & Miley 1989; Veilleux et al. 1995; Heckman et al. 2000; Rupke et al. 2003) and moderate-to-high redshifts (e.g., Rupke et al. 2003; Shapley et al. 2003) provides additional support for the prevalence of dust outflows. Assuming a Galactic dust-to-gas ratio, Heckman et al. (2000) estimate that the dust outflow rate is about 1% of the total mass outflow rate in LIRGs. Dust ejected from galaxies may help feed the reservoir of intergalactic dust (e.g., Coma cluster; Stickel et al. 1998; note, however, that tidal and ram-pressure stripping may be more efficient than winds at carrying dust into the ICM; see contribution by Stickel at this conference). Zone of Influence of Winds -------------------------- The impact of galactic winds on the host galaxies and the environment depends sensitively on the size of the “zone of influence” of these winds, i.e. the region affected either directly (e.g., heating, metals) or indirectly (e.g., ionizing radiation) by these winds. But the true extent of galactic winds is often difficult to determine in practice due to the steeply declining density profile of both the wind material and the host ISM. The zone of influence of galactic winds is therefore often estimated using indirect means which rely on a number of assumptions. A popular method is to use the measured velocity of the outflow and compare it with the local escape velocity derived from some model for the gravitational potential of the host galaxy. If the measured outflow velocity exceeds the predicted escape velocity [and]{} if the halo drag is negligible, then the outflowing material is presumed to escape the host galaxy and be deposited in the IGM on scales $\ga$ 50 – 100 kpc (see, e.g., Rupke et al. 2002 for an application of this method). Another method is to rely on the expected terminal velocity of an adiabatic wind at the measured X-ray temperature $T_X$ \[$v_X \sim (5KT_X/\mu)^{0.5}$, where $\mu$ is the mean mass per particle\] to provide a lower limit to the velocity of the wind fluid (this is a lower limit because it only takes into account the thermal energy of this gas and neglects any bulk motion; e.g., Chevalier & Clegg 1985; Martin 1999; Heckman et al. 2000). Both of these methods make the important assumption that halo drag is negligible. Silich & Tenorio-Tagle (2001) have argued that halo drag may severely limit the extent of the wind and the escape fraction. Drag by a dense halo or a complex of tidal debris may be particularly important in ULIRGs if they are created by galaxy interactions (e.g., Veilleux, Kim, & Sanders 2002b). The large uncertainties on these indirect estimates of the zone of influence of galactic winds emphasize the need for more direct measurements; these are discussed next. 0.1in [Deep Multiwavelength Maps of Local Galaxies.]{} The fundamental limitation in directly measuring the zone of influence of winds is the sensitivity of the instruments. Fortunately, $CXO$ and $XMM$-Newton now provide powerful tools to better constrain the extent of the hot medium (e.g., M 82, Stevens et al. 2003; NGC 3079, Cecil et al. 2002; NGC 6240, Komossa et al. 2003; Veilleux et al. 2003; NGC 1511, Dahlem et 2003). The reader should refer to the contribution of M. Ehle at this conference for a summary of recent X-ray results (see also Strickland et al. 2003 and references therein). Technological advancements have also allowed to detect galactic winds on very large scales at radio wavelengths (e.g., Irwin & Saikia 2003). A discussion of these results is beyond the scope of this short review. The present discussion focusses on optical constraints derived from the detection of warm ionized gas on the outskirts of wind hosts. Progress in this area of research has been possible thanks to advances in the fabrication of low-order Fabry-Perot etalons which are used as tunable filters to provide monochromatic images over a large fraction of the field of view of the imager. The central wavelength (3500 Å – 1.0 $\mu$m) is tuned to the emission-line feature of interest and the bandwidth (10 – 100 Å) is chosen to minimize the sky background. The data acquisition methods used to reach very faint flux limits are discussed in Veilleux (2003) and references therein, and are not repeated here. The Taurus Tunable Filter (TTF; Bland-Hawthorn & Jones 1998; Bland-Hawthorn & Kedziora-Chudczer 2003) has been used on the AAT and WHT to produce emission-line images of several “quiescent” disk galaxies (Miller & Veilleux 2003a) and a few starburst galaxies (Veilleux et al. 2003) down to unprecedented emission-line flux levels. Gaseous complexes or filaments larger than $\sim$ 20 kpc have been discovered or confirmed in a number of wind hosts (e.g., NGC 1482 and NGC 6240; the presence of warm ionized gas at $\sim$ 12 kpc from the center of M 82 was discussed in §2.2). Multi-line imaging and long-slit spectroscopy of the gas found on large scale reveal line ratios which are generally not H II region-like. Shocks often contribute significantly to the ionization of the outflowing gas on the outskirts of starburst galaxies. As expected from shock models (e.g., Dopita & Sutherland 1995), the importance of shocks over photoionization by OB stars appears to scale with the velocity of the outflowing gas (e.g., NGC 1482, or ESO484-G036 versus NGC 1705; NGC 3079 is an extreme example of a shock-excited wind nebula; Veilleux et al. 1994), although other factors like the starburst age, star formation rate, and the dynamical state of the outflowing structure (e.g., pre- or post-blowout) must also be important in determining the excitation properties of the gas at these large radii (e.g., Shopbell & Bland-Hawthorn 1998 and Veilleux & Rupke 2002). In the cases of AGN-driven winds, the hard radiation from the central source sometimes produces highly ionized winds and/or large-scale ionization cones (e.g., NGC 1068, NGC 1365 and NGC 4388; Veilleux et al. 2003). 0.1in [Lyman Break Galaxies.]{} Large absorption-line data sets collected on high-$z$ galaxies provide new constraints on the zone of influence of winds in the early universe. Adelberger et al. (2003) have recently presented tantalizing evidence for a deficit of neutral hydrogen clouds within a comoving radius of $\sim$ 0.5 $h^{-1}$ Mpc from $z \sim 3$ LBGs. The uncertainties are large and the results are significant at less than the $\sim$ 2$\sigma$ level. Adelberger et al. (2003) argue that this deficit, if real, is unlikely to be due solely to the ionizing radiation from LBGs (e.g., Steidel et al. 2001; Giallongo et al. 2002). They favor a scenario in which the winds in LBGs directly influence the surrounding IGM. They also argue that the excess of absorption-line systems with large CIV column densities near LBGs is evidence for chemical enrichment of the IGM by the LBG winds. Future Avenues of Research ========================== Although great strides have been made over the last decade in understanding the physics and impact of galactic winds in the local and distant universe, much work remains to be done to be able to quantify the overall role of these winds on the formation and evolution of galaxy-sized structures. Absorption-line studies of bright background galaxies (e.g., high-$z$ quasars, LBGs) have proven to be a very powerful tool to constrain the zone of influence of galactic winds at large redshifts. The next generation of instruments on $HST$ will provide the capabilities to extend the sample to a larger set of wind galaxies. $CXO$ and $XMM$-Newton will continue their harvest of high-quality data on the hot medium in galactic winds, and within five years a new generation of radio telescopes (e.g., $EVLA$, $SKA$, $CARMA$, $ALMA$) will probe the hot relativistic component of galactic winds better than ever before and provide the sensitivity to better quantify the role of the molecular gas in the dynamics of local winds. This component may be particularly important in determining the overall thermalization efficiency of galactic winds, or the percentage of the mechanical energy from the starburst or AGN that goes into heating the gas and driving the outflow; this quantity is currently very poorly constrained. The advent of tunable filters on 8-meter class telescopes \[e.g., OSIRIS on the GranTeCan (Cepa et al. 2000) and the Maryland-Magellan Tunable Filter on the Baade 6.5-m telescope\] should improve the sensitivity of optical wind surveys at least tenfold. Measurements with this second generation of tunable filters will provide direct quantitative constraints on the gaseous cross-section of active and star-forming galaxies, and the importance of mass exchange between galaxies and their environment. These powerful instruments will be ideally suited to search for galaxies with starburst-driven winds, exploiting the contrast in the excitation properties of the wind component and the star-forming disk (Veilleux & Rupke 2002). Special thanks to P.-A. Duc for organizing an excellent conference. Some of the results presented in this paper are part of a long-term effort involving many collaborators, including J. Bland-Hawthorn, G. Cecil, P. L. Shopbell, and R. B. Tully and Maryland graduate students S. T. Miller and D. S. Rupke. This article was written while the author was on sabbatical at the California Institute of Technology and the Observatories of the Carnegie Institution of Washington; the author thanks both of these institutions for their hospitality. The author acknowledges partial support of this research by a Cottrell Scholarship awarded by the Research Corporation, NASA/LTSA grant NAG 56547, and NSF/CAREER grant AST-9874973. Adelberger, K. L., et al. 2003, , 584, 45 Ajiki, M., et al. 2002, , 576, L25 Alton, P. B., Davies, J. I., & Bianchi, S. 1999, , 343, 51 Alton, P. B., et al. 1998, , 507, L125 Armus, L., Heckman, T., & Miley, G. 1989, , 347, 727 Baan, W. A., & Irwin, J. A. 1995, , 446, 602 Bender, R., Burstein, D., & Faber, S. M. 1993, , 411, 153 Bland-Hawthorn, J., & Cohen, M. 2003, , 582, 246 Bland-Hawthorn, J., & Jones, D. H. 1998, PASA, 15, 44 Bland-Hawthorn, J. & Kedziora-Chudczer, L. 2003, PASA, 20, 242 Böhringer, H., et al. 1993, , 318, L25 Carollo, C. M., & Danziger, I. J. 1994, , 270, 523 Cecil, G., Bland-Hawthorn, J., & Veilleux, S. 2002, , 576, 745 Cecil, G., et al. 2001, , 555, 338 Cepa, J., et al. 1990, in Optical and IR Telescope Instrumentation and Detectors, eds. M. Iye and A. F. Moorwood, Proc. SPIE, 4008, 623 Chevalier, R. A., & Clegg, A. W. 1985, Nature, 317, 44 Churazov, E., et al. 2002, , 332, 729 Crenshaw, D. M., Kraemer, S. B., & George, I. M 2003, , 41, 117 Dahlem, M., et al. 2003, , 403, 547 David, L. P., et al. 2001, , 557, 546 Dawson, S., et al. 2002, , 570, 92 D’Ercole, A., & Brighenti, F., 1999, , 309, 941 Devine, D., & Bally, J. 1999, , 510, 197 Dopita, M. A., & Sutherland, R. S. 1995, , 455, 468 Dupke, R. A., & Arnaud, K. A. 2001, , 548, 141 Fabian, A. C., et al. 2000, , 318, L65 Finoguenov, A., et al. 2002, , 381, 21 Franx, M., & Illingworth, G. 1990, , 359, L41 Franx, M., et al. 1997, , 486, L75 Frye, B., Broadhurst, T., & Benitez, N. 2002, , 568, 558 Garcia-Burillo, S., et al. 2001, , 563, L27 Giallongo, E., et al. 2002, , 568, L9 Gonzalez Delgado, R. M., et al. 1998, , 495, 698 Griffiths, R. E., et al. 2000, Science, 290, 1325 Heckman, T. M. 2002, in ASP Conf. Ser. 254, Extragalactic Gas at Low Redshift, eds. J. Mulchaey and J. Stocke (San Francisco: ASP), 292 Heckman, T. M., Armus, L., & Miley, G. K. 1990, , 74, 833 Heckman, T. M., et al. 2000, , 129, 493 —–. 2001, , 554, 1021 Heinz, S., Choi, Y. Y., Reynolds, C. S., & Begelman, M. C. 2002, , 569, L79 Hughes, D. H., Gear, W. K., & Robson, E. I. 1994, , 270, 641 Hunter, D. A., & Gallagher, J. S. III 1990, , 362, 480 —–. 1997, , 475, 65 Irwin, J. A., & Saikia, D. J. 2003, , in press (astro-ph/0309373) Irwin, J. A., & Sofue, Y. 1992, , 396, L75 Israel, F. P., et al. 1998, , 336, 433 Kim, W.-T., & Narayan, R. 2003, ApJ, in press (astro-ph/0308376) Koda, J., et al. 2002, , 573, 105 Komossa, St., et al. 2003, , 582, L15 Koyama, K., et al. 1989, Nature, 339, 603 Kunth, D., et al. 1998, , 334, 11 Lehnert, M. D., & Heckman, T. M. 1995, , 97, 89 —–. 1996, , 462, 651 Lehnert, M. D., Heckman, T. M., & Weaver, K. A. 1999, , 523, 575 MacLow, M.-M., & McCray, R. 1988, , 324, 776 Marlowe, A. T., et al. 1995, , 438, 285 Martin, C. L. 1998, , 506, 222 —–. 1999, , 513, 156 —–. 2003, in The Neutral ISM in Starburst Galaxies, ASP Conf. Series, eds. S. Aalto, S. Hüttemeister, and A. Pedlar, in press. Martin, C. L., Kobulnicky, H. A., & Heckman, T. M. 2002, , 574, 663 McKeith, C. D., et al. 1995, , 293, 703 McNamara, B. R., et al. 2000, , 534, 135 McNamara, B. R., et al. 2001, , 562, L149 Meurer, G. R., et al. 1992, , 103, 60 Miller, S. T., & Veilleux, S. 2003a, ApJS, 148, 000 —–. 2003b, ApJ, 592, 79 Pettini, M., et al. 2000, , 528, 96 —–. 2001, , 554, 981 —–. 2002, , 569, 742 Phillips, A. C. 1993, , 105, 486 Quilis, V., Bower, R. G., & Balogh, M. L. 2001, , 328, 1091 Radovich, M., Kahanpää, J., & Lemke, D. 2001, , 377, 73 Rupke, D. S, Veilleux, S., & Sanders, D. B. 2002, , 570, 588 Ruszkowski, M., & Begelman, M. C. 2003, , 586, 384 Schiano, A. V. R. 1985, , 299, 94 Shapley, A. E., et al. 2003, , 588, 65 Shopbell, P. L., & Bland-Hawthorn, J. 1998, , 493, 129 Silich, S. A., & Tenorio-Tagle, G. 2001, , 552, 91 Sofue, Y. 2000, , 540, 224 Steidel, C. C., Pettini, M., & Adelberger, K. L. 2001, ApJ, 546, 665 Stevens, I. R., Read, A. M., & Bravo-Guerrero, J. 2003, , preprint (astro-ph/0306334) Stickel, M., et al. 1998, , 329, 55 Strickland, D. K. 2002, in ASP Conf. Ser. 253, Chemical Enrichment of the Intracluster and Intergalactic Medium, eds. R. Fusco-Femiano and F. Matteucci (San Francisco: ASP), 387 Strickland, D. K., et al. 2000, , 120, 2965 —–. 2002, , 568, 689 —–. 2003, ApJS, preprint (astro-ph/0306592) Tenorio-Tagle, G., et al. 1999, , 309, 332 Trager, S. C., et al. 1998, , 116, 1 Veilleux, S. 2003, in The Neutral ISM in Starburst Galaxies, ASP Conf. Series, eds. S. Aalto, S. Hüttemeister, and A. Pedlar, in press. Veilleux, S., Kim, D.-C., & Sanders, D. B. 2002b, , 143, 315 Veilleux, S., & Rupke, D. S. 2002, , 565, L63 Veilleux, S., et al. 1994, , 433, 48 —–. 1995, , 98, 171 —–. 2002a, RMxAC, 13, 222 —–. 2003, , in press (astro-ph/0308330) Walter, F., Weiss, A., & Scoville, N. 2002, , 580, L21 Yamauchi, S., et al. 1990, , 365, 532 Yun, M. S., Ho, P. T. P., & Lo, K. Y. 1994, Nature, 372, 530 Zaritsky, D., Kennicutt, R. C. Jr., & Huchra, J. P. 1994, , 420, 87 Zensus, J. A. 1997, , 35, 607	{ "pile_set_name": "ArXiv" }
--- abstract: \| Simultaneous measurements of giant pulses from the Crab pulsar were taken at two widely spaced frequencies using the real-time detection of a giant pulse at 1.4GHz at the Very Large Array to trigger the observation of that same pulse at 0.6GHz at a 25-m telescope in Green Bank, WV. Interstellar dispersion of the signals provided the necessary time to communicate the trigger across the country [via]{} the Internet. About 70% of the pulses are seen at both 1.4GHz and 0.6GHz, implying an emission mechanism bandwidth of [at least]{} 0.8GHz at 1GHz for pulse structure on time scales of one to ten microseconds. The giant pulse spectral indices fall between $-2.2$ and $-4.9$, which may be compared to the average main pulse value for this pulsar of $-3.0$. The arrival times at both frequencies display a jitter of 100$\mu$s within the window defined by the average main pulse profile and are tightly correlated. This tight correlation places limits on both the emission mechanism and on frequency dependent propagation within the magnetosphere. At 1.4GHz the giant pulses are resolved into several, closely spaced components. Simultaneous observations at 1.4GHz and 4.9GHz show that the component splitting is frequency independent. We conclude that the multiplicity of components is intrinsic to the emission from the pulsar, and reject the hypothesis that this is the result of multiple imaging as the signal propagates through the perturbed thermal plasma in the surrounding nebula. At both 1.4GHz and 0.6GHz the pulses are characterized by a fast rise time and an exponential decay time which are correlated. At 0.6GHz the rise time is not resolved. The rise and fall times of the 1.4-GHz components vary from component to component and from pulse to pulse. The pulse broadening with its exponential decay form is most likely the result of multipath propagation in intervening ionized gas. These decay times, and that seen in contemporaneous 0.3-GHz average pulse data, are large compared to normal conditions for the Crab pulsar. The most likely location for the perturbed plasma is the interface region between the pulsar-driven synchrotron nebula and the overlying supernova ejecta. author: - 'Shauna Sallmen & D. C. Backer' - 'T. H. Hankins' - 'D. Moffett' - 'S. Lundgren' nocite: - '[@mh96]' - '[@lcu+95]' - '[@cc+69]' - '[@gm69]' - '[@ssp71]' - '[@hr71]' - '[@m97]' - '[@han71]' - '[@hr75]' - '[@eil96]' - '[@hm98]' - '[@ir77]' - '[@lt75]' - '[@chm98]' - '[@lcu+95]' - '[@hr71]' - '[@m97]' - '[@lun94]' - '[@lcu+95]' - '[@mel96]' - '[@a92]' - '[@a94]' - '[@lm95]' - '[@eil98]' title: Simultaneous Dual Frequency Observations of Giant Pulses from the Crab Pulsar --- Introduction {#secintro} ============ The Crab pulsar was discovered in 1968 by the detection of its extremely strong individual pulses ([@sr68]). Such pulses, which are 100’s of times stronger than the average, are not seen in most pulsars. The properties of these giant pulses have been explored for many years ([e.g.]{}, [@hcr70]; [@ss70]; [@fb90]; [@lcu+95]). Giant pulses in the Crab pulsar occur at all radio frequencies, but only [at the rotational phase of]{} the main pulse and interpulse components. These two components have counterpart nonthermal emission at high frequencies – from the infrared to gamma ray energies – and may be associated with the outer voltage gaps in the pulsar magnetosphere ([@ry95]). Giant pulses are not seen in the radio precursor nor at the phases of the high radio frequency components recently described by Moffett and Hankins (1996). The radio precursor is identified as being more typical of the emission from a conventional pulsar and is believed to originate at, and be aligned with, the magnetic pole. Lundgren [[et al. ]{}]{}(1995) found that two separate distributions were required to describe the fluctuations of single pulse energies[^1] for the Crab pulsar main pulse and interpulse components at 0.8GHz. About 2.5% of the pulses lie in the giant pulse distribution with a power law slope extending to high values and a low energy cutoff of 20 times the average of all pulse energies. The distinct distributions suggest different emission mechanisms for the giant and weak pulses and possibly different emission locations within the magnetosphere. However, the lack of an offset in the timing residuals between giant pulses and the average pulse profile ([@lun94]; for opposing evidence see [@fb90]) suggests that the emission region is the same. The frequency of occurrence of pulses ($f_\circ$) with energy greater than 20 times the average increases with frequency, from $10^{-4}$ at 0.146GHz ([@ag72]), to 0.025 at 0.8GHz ([@lcu+95]). The contribution of the giant pulses to the average energy of all pulses also increases with radio frequency, although not as quickly. The probability distribution of the giant pulse energies can be written as $$P(E_{\rm GP}>E_{\rm o})=f_{\rm o}\left ({E_{\rm o}\over E_{\rm min}}\right )^{-\alpha},$$ where $f_{\rm o}$ is the frequency of occurrence of the giant pulses, and $E_{\rm min}$ is the minimum energy. Correspondingly the probability density function is $$p(E_{\rm GP})={f_{\rm o}\alpha\over E_{\rm min}}\left ({E_{\rm GP}\over E_{\rm min}}\right )^{-\alpha-1},$$ and the corresponding mean giant pulse energy averaged over all pulses is ${f_{\rm o}\alpha\over (\alpha-1)}E_{\rm min}$. The probability distribution $P$ has a slope $\alpha = 2.3\pm 0.15$ at 0.8GHz ([@lcu+95]), and $\alpha=2.5$ with significant errors at 0.146GHz ([@ag72]). At 1.4GHz and 0.43GHz the overall slope is roughly consistent with these, but is not the same for all energies ([@fb90]; [@m97]). Using the scaling law $\alpha \approx 2.5$ at all radio frequencies below 0.8GHz, we find that the contribution of giant pulses with energy more than 20 times the average of all pulses, $E_{\rm GP}>20\,E_{\rm avg}$, is 89% of the average energy at 0.8GHz ([@lcu+95]), 9% at 0.43GHz ([@fb90]), and only about 1% at 0.146GHz ([@ag72]). At 1.4GHz, a similarly large fraction of the energy comes from the approximately 2% of pulses with greater than 20 times the average energy, although a single slope $\alpha$ does not accurately describe the distribution. There is no evidence of increased flux density in pulses near the giant pulses ([@ssp71]; [@lun94]), nor is there any correlation between giant pulses. [We note that many pulsars do show pulse to pulse correlation indicating a memory process with a duration of many rotational periods.]{} The timescale of giant pulses is, in contrast, less than a single period. In addition, the time separation distribution for giant pulses is consistent with Poisson process ([@lun94]). Despite all these studies, the emission bandwidth of the giant pulses has been poorly determined. Comella [[et al. ]{}]{}(1969) found that 50% of giant pulses were seen simultaneously at 0.074GHz and 0.111GHz. Goldstein & Meisel (1969) also found that some but not all pulses were correlated between 0.112GHz and 0.170GHz. Sutton [[et al. ]{}]{}(1971) noted that there was no evidence that the largest pulses at 0.16GHz and 0.43GHz were correlated. Heiles & Rankin (1971) observed giant pulses simultaneously at 0.318GHz and 0.111GHz, for a bandwidth spread of about 3:1. They found that pulses classified as giant at one frequency were stronger than the average at the other, but not usually classified as giant. Much more recently, Moffett (1997) reported that fully 90% of the giant pulses detected at 4.9GHz were also detected at 1.4GHz, implying an emission bandwidth of 3.5GHz at high radio frequencies. In this paper, we report on giant pulses observed simultaneously at 1.4GHz and 0.6GHz to explore the correlation in this intermediate range of frequencies. Section \[sec:gpobs\] describes the observations, while analysis of the simultaneous pulses lies in Section \[sec:gpanal\]. Observations {#sec:gpobs} ============ The data shown here were recorded on 1996 May 21 at UT $17^{\rm h}45^{\rm m}-19^{\rm h}15^{\rm m}$ (1.4/0.6 GHz) and 1996 Oct 12 at UT $11^{\rm h}30^{\rm m}-12^{\rm h}05^{\rm m}$ (4.9/1.4 GHz). We also refer to observations at 1.4-GHz earlier in 1996 and in 1997 November. The 1.4-GHz and 4.9-GHz data were taken at the NRAO Very Large Array (VLA).[^2] For the simultaneous 1.4/0.6-GHz observations all 27 VLA antennas were phased to create the equivalent [sensitivity]{} of a 130-m antenna, while the 0.6-GHz data were taken using a 25-m telescope at the NRAO Green Bank, WV, site. For the simultaneous 4.9/1.4-GHz observations the VLA was split into two equal subarrays. The peculiar phases of each antenna at the VLA were determined by observing a standard point-source calibrator. These phases were then applied to the antennas to synthesize a pencil beam pointed at the Crab pulsar, which essentially resolves out the bright Crab Nebula and vastly improves the signal to noise ratio compared to a single-dish antenna. The received voltages from each antenna are summed to form orthogonally circularly polarized 50-MHz bandwidth signals centered at precisely 4.8851GHz or 1.4351GHz. The signals are then detected and summed with a 100-$\mu$s time constant. The detector rms noise power was determined using an rms to DC converter. A detector threshold was set at either 5 or 6 times the running average of this rms noise level. Pulses that exceeded this threshold generated a trigger pulse that was sent to the data recorder, and were then saved to disk and archived to tape using the full 50-MHz bandwidth. In an off-line computer the data were coherently dedispersed using the method developed by Hankins (1971) and described by Hankins and Rickett (1975). Although the ultimate time resolution of the dedispersed data is 10ns, for the analyses described here the data were smoothed to 0.5–1.0$\mu$s after software detection. The two linearly polarized signals centered at exactly 0.610GHz were converted to 90MHz and 110MHz center frequencies, respectively. The two intermediate frequency signals were then summed and transmitted on a single fiber optic link from the 25-m telescope to the Green Bank–Berkeley Pulsar Processor (GBPP)[^3] which was located at the 140-ft telescope. The GBPP converted the signals to baseband, split these into 32 0.5-MHz channels, and dedispersed the pulsar signal in each channel by (de-)convolution in the time domain. The dispersion delay across the 16-MHz bandwidth of the GBPP at 0.6GHz is about 33ms, or one pulse period for the dispersion measure of the Crab pulsar [(${\rm DM}\approx 56.8$ pccm$^{-3}$)]{}. Full Stokes information for 982 samples across the pulsar period was recorded with an accurate UTC start time for each pulse. The 25-m telescope also monitors the Crab pulsar daily with 0.327-GHz observations which are valuable for their sensitivity to scatter broadening by intervening plasma. The 1.4/0.6-GHz part of this experiment utilized the difference in pulse arrival time between the two frequencies due to interstellar dispersion to provide the time interval needed to communicate the trigger information between the sites. At the VLA, the same trigger pulse that was sent to the data recorder was also sent, as an interrupt, to the SUN workstation used for experiment control and recording. The program that received the interrupt had a socket link open over the Internet to a slave program running on another SUN workstation in Green Bank, and communicated the 1.4356-GHz arrival time of the pulse to Green Bank. We arranged for the GBPP to begin taking data just before the giant pulse reached the upper edge of the 0.6-GHz band, in order to obtain data for the same main pulse across the entire band. The dispersion time delay between 1.4351GHz and 0.618GHz (the top of our 0.6-GHz band) allowed a half second (0.503s) to arrange this. The SUN workstations at the two sites were synchronized to the local versions of UTC which were derived from accurate atomic clocks using the [xntp]{} protocol. Although both remote sites had 56-kB links to the Internet, the typical Internet transfer time delay was 200ms during our observation. The program running in Green Bank received the trigger message with its VLA time stamp, calculated the transit time of the trigger, and compared that to the dispersion delay difference of 0.503s. In addition to this delay, the program included other factors such as the difference in pulse arrival due to the separation between observing sites on the Earth and the latency in the GBPP hardware, both of which were of order 1–3ms. [ If sufficient time remained, the program waited until the appropriate time and issued a trigger to the GBPP [via]{} the SUN parallel port to take data for the next pulse period.]{} Owing to the slow rate of data transfer from the GBPP, it could only accept such a command approximately every 12 s. Some [VLA-initiated]{} triggers were therefore missed by the GBPP. For the 4.9/1.4-GHz measurements the recording system was triggered by a 4.9-GHz pulse, and then automatically triggered again after the appropriate dispersion delay to record the 1.4-GHz signals. The Stokes parameters for the high-time resolution data from the VLA were formed from the dedispersed voltages. The necessary 90$^\circ$ phase shift was obtained using an finite impulse response approximation to the Hilbert Transform and was applied to the right circular polarization signal before forming the Stokes parameters. No instrumental polarization corrections were made other than bandpass leveling and gain matching. Concurrent calibration (Moffett 1997) has shown that the polarization cross-coupling is less than 10% for the VLA phase array. The polarization error then is comparable with the radiometer uncertainty imposed by the limited number of degrees of freedom in the data ($\sigma_I/I = (4\Delta \nu \,\Delta \tau)^{-1/2} = 0.1)$. The polarization profiles at 0.6GHz were calibrated using factors derived from pulsed noise observations. The receiver introduces a relative phase between the two linear polarizations which couples the Stokes parameters $U$ and $V$. This phase was determined and removed using nearby observations of the Vela pulsar, and comparison to a template polarization profile. No attempt has been made to remove coupling between the two polarizations. The error in polarization due to improper calibration is estimated at 10%. For each pulse, the relative dispersion between the 32 channels was removed, and the resultant data were summed over channels, after first removing the effects of the pulsar’s rotation measure (${\rm RM}=-42.3$ rad m$^{-2}$) across the band. The unknown Faraday rotation from the ionosphere causes negligible rotation of 1 to 10 degrees across the total band. Analysis {#sec:gpanal} ======== Wide Bandwidth Correlation -------------------------- Of the 85 trigger events initiated at the VLA a total of 77 events reached Green Bank within the required time, and were accepted by the GBPP. The [xntp]{} protocol requires 24 hours to stabilize to the accuracy required by our experiment. [The minimum time stabilization period]{} was not available for the SUN at the VLA. Consequently the VLA clock used to identify the time at which the trigger was sent drifted by a small amount. We are currently certain that the received trigger events allowed capture of the correct period in the GBPP for 42 pulses. The arrival times and pulse energy amplitudes were determined at both frequencies for each of these pulses. The 0.6-GHz arrival times were determined by cross-correlation with a model template, which consisted of a single-sided exponential with a decay time scale of 3 time bins, about 100$\mu$s. Owing to complex structure discussed below, the 1.4-GHz pulse arrival times were obtained by computing the location of the centroid of the pulses. Pulse amplitudes in units of Jy-s were determined by removing an “off-pulse” baseline, and then summing the flux in the time bins that comprise the “on-pulse” window. The measurement uncertainty of these amplitudes was determined from the “off-pulse” noise distribution. We definitely detected 29 of the 42 [correctly timed]{} pulses at both radio frequencies. These detections correspond to a 0.6-GHz pulse energy threshold of about 4.5 times the typical measurement uncertainty, or 0.075 Jy-s using 0.14 K Jy$^{-1}$ for the 25-m telescope. This gain factor was determined using on and off measurements of the Crab nebula (which is 1208 Jy at 0.6GHz), and has an estimated [uncertainty]{} of 50%. We conclude that about 70% of the pulses are detected at both frequencies. This statistic is used to discuss the spectral index distribution below. The data for a single giant pulse at 1.4GHz and 0.6GHz are displayed in Figures \[vlabest\] and \[gbbest\], respectively. This pulse is strongly polarized at both frequencies, although fully two thirds of the giant pulses at the lower frequency are consistent with zero polarization. The 0.6-GHz data has a low number of degrees of freedom, and so the [polarization estimation uncertainty]{} is about 5%. At 1.4GHz, the typical polarization is about 8% although at least one pulse is 50% polarized. The position angle of the linear polarization generally varies significantly across the pulse, as is seen in Figure \[vlabest\]. The arrival times for the 29 giant pulses detected at the two frequencies were separately compared to a single model for this pulsar using the TEMPO program developed for pulsar timing ([@tw89]). For each radio frequency, the arrival times are well-represented by the model, leaving timing residuals of order $\pm 100\,\mu$s. The residuals are comparable to the pulse width of the average profile during periods of low scattering, which is $275 \pm 50\,\mu$s (FWHM) at 0.6GHz, and $257 \pm 50\,\mu$s at 1.4GHz. These widths were estimated using GBPP data obtained with the 25-m and 140-ft telescopes, respectively. The timing residuals for 1.4GHz and 0.6GHz are plotted against one another in the top panel of Figure \[resids\], which shows that they are highly correlated. The solid line has a slope of one and goes through the origin. In order for the points to fall along this line, the 1.049-ms digital latency of the GBPP and the 235.42-$\mu$s latency of the VLA samplers and delay lines were removed, and a further fit for dispersion measure was done in TEMPO. The derived DM is 56.830 pc cm$^{-3}$ although systematic errors may remain in the arrival times from the two sites. The bottom panel displays the same 1.4-GHz residuals with the solid line removed. Eilek (1996) has shown that the dispersion law in the polar cap is proportional to $\nu^{-1}$, as opposed to $\nu^{-2}$ for the cold interstellar medium (ISM). No systematic trends remain in the data in the lower panel of Figure \[resids\], indicating that systematic variations with pulse phase are less than $\pm 15\,\mu$s between our two bands. This places limits on the differential effects of propagation through the magnetosphere. Geometrically, emission in the two bands must originate within 0.16$^\circ$ of rotational phase, or a range of 4.5km in altitude. It would have been possible to have correlated emission from subpulses at different pulse longitudes at each frequency. In this case, the radiation at the two frequencies need not have come from the same radiating unit of charges. The observed rms jitter in arrival time at either frequency is $\approx$100$\mu$s, so the fact that the difference between the residuals has such a small dispersion indicates that the emission must be from the same radiating unit at both frequencies. This means that at least 70% of the giant pulses must have a bandwidth of [at least]{} 0.8GHz at 1GHz. The emission is clearly broadband for these cases.[^4] Pulse Shape Model {#secshape} ----------------- The giant pulses at 0.6GHz all have profiles that display a fast rise followed by an exponential decay, similar to the profile shown in Figure \[gbbest\]. The exponential decay time scale is $\tau_{\rm ISS}(0.6\,{\rm GHz})= 95 \pm 5 \mu$s. Monitoring of the exponential broadening of the average pulse profile at 0.3GHz using the 25-m telescope ([@bw96]) provides a contemporaneous decay time scale of $\tau_{\rm ISS} (0.3\,{\rm GHz}) = 1.3 \pm 0.2$ms. The 0.3-GHz and 0.6-GHz pulse broadening time scales are consistent with the $\nu^{-4\,{\rm to}\,-4.4}$ dependence expected from scattering by an intervening, turbulent plasma screen. At the time of these observations the Crab was undergoing a period of unusually large scattering. The contemporaneous value of $\tau_{\rm ISS} (0.3\,{\rm GHz})$ may be compared to 0.28ms at an earlier epoch. Enhanced scattering of the Crab pulsar radiation also occurred in 1975 ([@ir77]; [@lt75]). A likely site of the perturbed plasma that causes the enhanced pulse broadening is in the interface between the synchrotron nebula and the supernova ejecta ([@hetal96]). This region is proposed because such enhanced scattering events are not seen for any other pulsar. Furthermore, this location is known to contain dense thermal plasma with structure on a variety of length scales. The transverse velocity of the line of sight with respect to the interface material is sufficient to bring occasional regions which produce strong perturbations into view. The fast rise of the 0.6-GHz giant pulses indicates that any intrinsic time scale of the pulse is unresolved, [.5ex]{}10$\mu$s. The giant pulses at 1.4GHz have a wide variety of shapes. Figures \[pulse69\] and \[pulse56\] display two further single pulse profiles. The first pulse is extremely narrow and is dominated by a single component, while the second has several components contributing to the emission. The darker solid line is a fit to the data where the flux model $S(t)$ consists of up to six components and is of the form $$S(t) = \sum_{i=1}^n a_{1_i} (t-a_{2_i}) e^{-(t-a_{2_i})/a_{3_i}},~~~n\leq 6. \label{eq:xfred_model}$$ These components rise to their peak in a time $a_3$, fall by $e^{-1}$ in a further $2.15~a_3$, and have a pulse energy amplitude of $A=a_1 a_3^2$. At 1.4GHz, the majority of giant pulse components are well represented by this model, with widths $a_3$ ranging from 1.2$\mu$s to 10$\mu$s. Components with widely varying decay time scales may be superposed within a single giant pulse, as shown in Figure \[pulse56\]. The narrow component of the giant pulse shown in Figure \[pulse69\] has a rise time of 1.2$\mu$s, and a decay time scale of 2.5$\mu$s. The weak and broad second component in Figure \[pulse69\] is clearly necessary to account for the emission on the trailing edge of the pulse, which does not follow an exponential tail. Pulse Shape Interpretation {#secinterp} -------------------------- The 1.4-GHz observations raise fundamental issues about the nature of the giant pulse emission. Is the multiplicity of the components and the characteristic shape of the components a result of propagation through a turbulent screen, or are these effects intrinsic to pulse formation and propagation in the pulsar magnetosphere, or is there a mixture of effects present? Two additional results from further observations at the VLA and the 0.3-GHz monitoring in Green Bank provide important constraints to aid in answering these questions. First, our broadening measurement of 1.3ms at 0.3GHz scales to $1.9-3.5~\mu$s at 1.4GHz using the $\nu^{-4{\rm~to~}-4.4}$ dependence expected for a spectrum of turbulence filling the intervening screen. This range is consistent with the typical [minimum]{} broadening time of approximately 2.5$\mu$s that we measured for the giant pulses at 1.4GHz. During a later epoch (1997 Nov 26) when the scattering broadening of the average pulse at 0.3GHz had increased dramatically by a factor of about 5, the minimum pulse widths of the 1.4-GHz giant pulses increased by a similar factor. At an earlier epoch in 1996 both the 0.3-GHz average pulse broadening and the minimum width of the 1.4-GHz giant pulses were reduced. The second observational constraint on our interpretation of the 1.4-GHz giant pulse shapes comes from consideration of the multiple component spacing. These do not scale in the same way as the component broadening. During an earlier epoch of low 0.3-GHz scattering, the typical 1.4-GHz component spacing is similar to that for 1996 May 21. In addition, simultaneous VLA measurements of the giant pulses at 1.4GHz and 4.9GHz (1996 Oct 12) show a frequency [independence]{} of the multiple component spacing (Hankins & Moffett 1998). At 4.9GHz the pulse components are intrinsically very short, typically 0.1 to 0.4$\mu$s wide, while the corresponding components are broadened at 1.4GHz. Simultaneous 4.9-GHz and 1.4-GHz measurements of a single giant pulse are shown in Figures \[5GHz\]a and \[5GHz\]b. The separation of the two main components is similar at both frequencies, in that the onsets of the pulse components (the fiducial points in the fits discussed in section \[secshape\]) are separated by the same amount. A similar conclusion of frequency independence of structure can be inferred from the 1.4GHz and 0.6GHz measurements of 1996 May 21; that is, 1.4-GHz giant pulses with multiple components have a spread of tens of microseconds, which is consistent with the limit on structure at 0.6GHz. The correlation of broadening time scales over the range from 0.3GHz to 1.4GHz provides support for diffractive scattering in the nebular material. Although the simple thin screen prediction for interstellar scattering is a single sided exponential, $\exp(-t)$, models including two widely separated screens or a single thick screen both predict pulse shapes similar to the observed $(t/a_3)\exp(-t/a_3)$ form of the components ([@wil74]). Isaacman & Rankin (1977) have derived parameters for a two screen model from studies of the average pulse profile of the Crab pulsar. A scatter plot of the 1.4-GHz component energies, $A=a_1 a_3^2$ from the fitted parameters in Equation \[eq:xfred\_model\], [vs.]{} component width, $a_3$, shows that the energies are independent of the pulse width, have an average of about $6.3 \times 10^{-3}$Jy-s, and are scattered over two orders of magnitude. This means that the peak flux is approximately inversely proportional to the width; stronger pulses are narrower. The data from other observing sessions show that although the broadening times change, the pulse energies remain within the same range; [i.e.]{}, when the scattering times are longer, the peak amplitudes are correspondingly smaller. This multi-epoch result supports the hypothesis that, in spite of questions about the pulse component shape and the frequency dependence of the exponential decay time scale, the shape of the components at 1.4GHz is most likely the result of scattering by the intervening medium. The wide variations in pulse broadening seen within a single pulse are, however, difficult to explain unless the scattering screen is illuminated differently by each component. While we expect variation in the pulse broadening from component to component and from pulse to pulse owing to their being instantaneous samplings of the time variable diffractive effects, why does the [minimum]{} exponential time scale of the components at 1.4GHz agree with that extrapolated from 0.3GHz? We have considered the possibility that the components of giant pulses are intrinsically single impulsive events that are multiply refracted and/or diffracted by discrete regions of enhanced plasma density and/or turbulence, respectively. Lyne & Thorne (1975) invoked similar intermittency in the wave front perturbing medium to explain the irregular and rapidly varying effects of the 1974 event, and Cordes, Hankins & Moffett (1998) have studied the effects of discrete, small scale “screenlets” on giant pulses. An intermittent scattering medium that multiply refracts and/or diffracts a single impulsive signal appears to solve the problem, but has serious difficulties explaining the frequency independence of the spread and multiplicity of pulse components. Consider a screen at a distance $xD$ from the pulsar and $(1-x)D$ from the observer. Diffraction leads to an expansion of the angular spectrum of the intensity by an angle $\theta_{\rm d}\equiv\lambda/l_\circ$ where $l_\circ$ is the coherence scale for one radian of phase difference across the wavefront. An impulse which passes through this screen is broadened by the multipath diffractive time scale, $\tau_{\rm d}$. If the scattering material covers a transverse scale of $l_{\rm d}=xD\theta_{\rm d}$ centered on the geometric line of sight, then $\tau_{\rm d}=x(1-x)D\theta_{\rm d}^2/2c$. A measure of the apparent angular size of the pulsar is $\theta_{\circ,d}=x\theta_{\rm d}$. If $\tau_{\rm d}$ is 4$\mu$s at 1GHz and the screen is located 2 pc from the pulsar, $x=10^{-3}$, $\theta_{\rm d}=60$mas, $\theta_{\circ,d}=60\,\mu$as, $l_\circ=10^7$cm, and $l_{\rm d}=1.8\times 10^{11}$cm. If the diffracting material covers only a fraction of $l_{\rm d}$, then the pulse broadening time will be [reduced]{} relative to $\tau_{\rm d}$ owing to the reduction in the multipath propagation ([@chm98]). The observed frequency scaling between 0.3-GHz and 1.4-GHz broadening times reported above is not consistent with this result. In addition, a diffracting region not located along the line of sight will result in components disappearing from view as the frequency increases, owing to the reduced viewing zone set by $l_{\rm d}$ centered on the line of sight. This is in conflict with the current observational results. Refraction in an intermittent medium has similar limitations to that of diffraction. Consider a uniform density, spherical plasma lens at a transverse distance $b$ from the line of sight with a transverse dimension significantly less than $b$. The excess propagation delay from the pulsar to the observer is dominated by the geometric delay $\tau_{\rm g}=b^2/(2cx(1-x)D,$ and is independent of frequency owing to the highly aberrant lens. While multiple lenses of this form appear to satisfy the frequency independence of giant pulse component spacings, they will [not]{} satisfy the frequency independence of the number of components and their overall distribution in longitude. More components over a wider range of longitude would be expected at lower frequencies as is the case for diffraction discussed above. We conclude that the multiplicity of components is intrinsic to the pulsar emission mechanism or to propagation effects within the pulsar magnetosphere. Multiple scatterings of a single emitted pulse component cannot easily explain the observed spread and multiplicity of pulse components. Despite this, we favor propagation effects in the Crab Nebula as an explanation for the shape of the components at 1.4GHz and below. Alternatively, the observed component shape may be intrinsic to the emission. Each component is emitted with its own time scale, but the characteristic $(t/a_3)\exp(-t/a_3)$ shape. Components with time scales shorter than the interstellar broadening time scale are broadened by scattering in the ISM. This model explains the observed correlation between the minimum time scale and the low-frequency scattering, but involves two separate explanations for the characteristic shape of the components. Determining whether the long-$\tau$ end of the distribution scales with the low-frequency scattering or is independent of it would distinguish between the two possibilities. Observations at multiple frequencies during times of minimal scattering at 0.3GHz are critical to further inquiry. Energies and Spectral Indices ----------------------------- The distribution of pulse energy amplitudes from the 1.4/0.6-GHz experiment are displayed along with the corresponding detection thresholds in Figure \[gbvla\]. The 1.4-GHz amplitudes of the 13 pulses that were not detected at 0.6GHz are also shown. Solid lines corresponding to spectral indices $-2.2$ and $-4.9$ are shown, where spectral index $q$ is defined by $A_{\rm GB}/A_{\rm VLA} = (0.6/1.4)^q$. The overall spectral index for the Crab pulsar is $-3.1$, while the spectral index for the average main pulse, which is shown as a dotted line in Figure \[gbvla\] ([@m97]), is $-3.0$. The pulse amplitude of the average main pulse is 5.4 $\times$ 10$^{-3}$ Jy-s at 0.6GHz. The largest 0.6-GHz giant pulse therefore has a pulse amplitude of about 150 times the amplitude of the average pulse. The giant pulses are narrower than the average pulse, and so are even stronger relative to the average pulse within this window. Lundgren [[et al. ]{}]{}(1995) concluded from their analysis of the relative contribution of giant pulses to the average as a function of frequency ([cf.]{} section \[secintro\]) that if the emission is narrow band, the rate of giant pulses must increase with frequency, while if the emission is broadband, then the giant pulses must have flatter spectra than the weak pulses. Heiles & Rankin (1971) found that their measured spectral indices at low radio frequencies ranged from nearly 0 to less than $-3.0$. Similarly, Moffett (1997) found that between 1.4 and 4.9GHz, giant pulse spectral indices ranged from 0 to $-4$, with an average of about $-2$. These two studies corroborate Lundgren’s analysis. We find that at least 70% of the pulses are broadband, and so we expect that their spectral indices are, on average, flatter than the average main pulse spectral index. The average spectral index of the giant pulses in Figure 7 which were detected at both frequencies is $-3.4$, which is comparable to that of the average main pulse. This estimate is biased by the fact that the pulses that were not detected at 0.6 GHz have flatter spectral indices. In addition there may be systematic errors in flux calibration that could change the average spectral index by up to $0.4$. The individual giant pulse spectral indices display a relatively large scatter. Individual pulse spectral indices are known for two other pulsars. The distributions of spectral indices for pulsars B0329$+$54 and B1133$+$16 have a standard deviation of 0.2, and range from -1.6 to -3.1 and -2 to 0, respectively ([@bs78]). One contribution to the spectral index variations is the stochastic uncertainty in the determination of the amplitudes that is introduced by the low number of degrees of freedom in narrow band observations of intrinsically short duration pulses. We estimate this [uncertainty]{} to be of order 10% (20%) at 1.4 (0.6) GHz for an intrinsic pulse width of 1$\mu$s. This is not large enough to explain the scatter in the spectral indices. The scatter could also be intrinsic to the radiation emission process. The signal could consist of a randomly occurring series of nanosecond impulses whose combined power spectrum is irregular. This would cause scatter in the observed spectral indices. But with the radiation extending over 1$\mu$s or more, there are many nanosecond pulses which would smooth out this distribution. Alternatively, one might expect the spectral index to vary due to properties of the emission beam: frequency dependence and orientation with respect to the observer. Future observations with higher time resolution and also with many samples across the spectrum will provide critical new insight into the giant pulse emission process. Models of the Emission Beam {#secwithwidth} --------------------------- We consider two general models for giant pulse emission. In the temporal model the giant pulse emission components are impulses in time ($<$1$\mu$s) with angular beam widths comparable to that of the average pulse, $\sim 3^\circ$. Following Lundgren (1994), we also consider a model in which the enhanced emission during a giant pulse results from a steady narrow beam ($<$1$^\prime$) whose position wanders on time scales 300$\mu$s $\ll t \ll P$. In this case intrinsic structure within a particular giant pulse is due to structure within this narrow beam. In both models, the average beam may be circular, as in polar cap models, or fan-shaped, as is likely if the emission comes from the outer gaps. In the angular model, the narrow beam of emission may wobble in either the $l$ direction (along the trajectory of the line of sight), or the $\phi$ direction (perpendicular to the trajectory of the line of sight). The width of the giant pulses corresponds to the size of the beam in this model, while the jitter in arrival times $\sigma_{\rm toa}$ corresponds to the wobbling of the beam along $l$. For giant pulses which occur a fraction $f$ of the time, the wobble in $\phi$ is then $w_{\rm gp}/(P f)$ where $P$ is the pulse period (following Lundgren [[et al. ]{}]{}1995). Lundgren was able to separate the giant pulse and normal pulse distributions at 0.8GHz, and found that one of 40 pulses is giant. Since the giant pulses form a separate distribution, then if they are all broadband, they will all also appear at higher radio frequencies. Then at 1.4GHz or 4.9GHz, one of 40 pulses should be giant. In fact, Moffett (1997) finds that one of 50 pulses at 1.4GHz has an energy greater than 20 times the average. At 4.9GHz the intrinsic width of the giant pulses is 0.1$\mu$s to 0.4$\mu$s. We find $\sigma_{\rm toa} \approx 100\,\mu$s, so the $0.001^\circ$ to $0.004^\circ$ beam must wobble $1.1^\circ$ in $l$ and $0.05^\circ$ to $0.2^\circ$ in $\phi$. This is not consistent with a narrow beam wobbling within a roughly circular average beam. The Emission Mechanism {#secemission} ---------------------- Radio emission from pulsars must come from a coherent emission process ([@cor81]). The exact process is very uncertain, as is the location(s) of the emission. It is not necessary for the giant pulse emission to originate at the same place or in the same way as the ordinary pulse emission. The broadband nature of the giant pulse emission provides the main constraint on its origin. According to Melrose (1996), broadband emission is traditionally associated with models in which the emission occurs at a pair production front in the polar cap, or by Schott radiation from a corotating charge and current distribution outside the light cylinder ([e.g.]{}, [@dk85]; Ardavan 1992, 1994). The emission process itself could rely on plasma instabilities ([@cr77]; [@a93]; [@mu79]; [@k+91]; [@wea96]). Alternatively, other maser processes such as linear acceleration emission ([@mel78]; [@row95]) or maser curvature emission ([@lm92], 1995) could produce the radiation. In any case, if the giant pulses are a temporal effect, this variability in radio emission could be the result of the statistics of a small number of coherently emitting regions which are incoherently summed, or an increase in the coherence within a single emission region. Although we believe that the $(t/a_3)\exp(-t/a_3)$ characteristic shape of the giant pulse components at 1.4GHz is most likely the effect of propagation through the Crab nebula interface, the remaining questions ([cf.]{} section \[secinterp\]) lead to consideration of effects intrinsic to the pulsar. An asymmetric shape is not expected for a simple narrow beam with an angular wobble. In either the temporal or angular beam model, this shape might be explained by effects that occur as the signals traverse the pulsar magnetosphere (Eilek 1998). The effects of aberration are too small to produce the broadening and asymmetry seen in these pulses, if one confines the range of emission altitudes to 4.5 km, the limit obtained from the timing residual differences at the two frequencies. In the temporal model, the asymmetric shape is consistent with any emission process which turns on with a rapid nearly linear rise, then saturates and decays. In this case one might expect the peak energy to be independent of width, whereas we have seen that it is the total pulse energy which is independent of width. If the model is truly temporal, then the angular size of the beam does not affect the observed pulse width. The radiation is beamed into a beam width $\theta \approx \gamma^{-1}$. The beam must be wider than any given 1.4-GHz pulse component, 50$\mu$s, and therefore $\gamma$[.5ex]{}100. We can use this value of $\gamma$ to estimate particle properties using a simplistic model of coherent curvature radiation. The power lost by the $N$ excess charged particles in the bunch will be $$P_{\rm curv} = N^2 \left(\frac{2 e^2 \gamma^4 c}{3 \rho_{\rm c}^2}\right),$$ where $e$ is the charge on an electron, $\gamma$ is the relativistic factor $(1-v^2/c^2)^{-1/2}$, and $\rho_{\rm c}$ is the radius of curvature of the magnetic field. We observe $6.3 \times 10^{-3}$ Jy-s in a 50-MHz band, so $P_{\rm curv}$ must equal the measured luminosity, which is therefore greater than $3.9 \times 10^{23}$ erg s$^{-1}$, assuming a distance of 2 kpc, and a circular beam 300$\mu$s = 3$^\circ$ wide. Then the number of particles in the bunch must be at least $$N=9.2 \times 10^{19} \left(\frac{\gamma}{100}\right)^{-2}\left( \frac{\rho_{\rm c}}{10^8 \rm ~cm}\right).$$ These particles must fit within a cube with volume $\le \lambda_{\rm obs}^3$ in order to maintain the observed coherence, so the number density of excess charges must be at least $\delta n_{\rm e} = N/\lambda_{\rm obs}^3 = 9.9 \times 10^{15} ({\gamma \over 100})^{-2}$cm$^{-3}$ for the parameters used above, and a wavelength of 21cm. For comparison, the Goldreich-Julian density in the observer’s frame is ${n_{\rm G-J}} ={\Omega B / e 2 \pi c } = 8.3 \times 10^{12} ({R \over R_{\rm NS}})^{-3}$cm$^{-3}$, using a surface magnetic field of $4 \times 10^{12}$G. The excess charge density can be further reduced if several bunches are radiating in a periodic structure. Summary ======= The giant pulse emission from the Crab pulsar is broadband since 70% of the pulses are observed in our 1.4/0.6-GHz experiment. The strong correlation in arrival times at the two frequencies implies that the same radiating unit is operating at both frequencies. The giant pulses display a scatter in spectral index that is consistent with or flatter than the spectral index of the average main pulse component. Pulsar emission models are restricted to those that can explain the broadband nature of the giant pulse radiation on intrinsic observed time scales of one to 10 microseconds. Above 1GHz a multiplicity of components is observed with frequency independent spacing and number. The 1.4/0.6-GHz data are also consistent with frequency independence. We conclude that the multiplicity is intrinsic to the pulsar emission process and not the result of multiple imaging in the intervening plasma. The exponential pulse broadening time scale of the average pulse at 0.3GHz and of the giant pulses at 0.6GHz are consistent with multipath propagation effects. The large values and rapid variations indicate a special scattering region. We identify this region with the interface between the Crab synchrotron nebula and the surrounding supernova ejecta. At 1.4GHz the observed giant pulse component shapes are characterized by fast rise and exponential decay time scales that are correlated. The minimum time scale is consistent with extrapolation of the pulse broadening at lower frequencies with a filled turbulent scattering screen. The shape and the distribution of broadening at 1.4-GHz are not fully understood. Multi-frequency simultaneous observations of giant pulses with higher time resolution at epochs of minimal scattering at low frequency will provide critical new insights into the emission processes and subsequent propagation effects. We thank John Ford for assistance with the trigger reception software. Thanks to the observatory staff at both sites for minimizing internet traffic during the two hours of the experiment. T.H. Hankins and D.A. Moffett thank the NSF for partial funding of this work under grant AST9315285. D.C. Backer and S. Sallmen thank the NSF for partial funding of this work under grant AST9307913. We are indebted to NASA for maintenance of the 25-m telescope at Green Bank. We thank J. Cordes, R. Jenet, M. Kramer, A. Melatos, and J. Weatherall for conversations and comments on the early versions of the paper, and the referee for attentive reading and useful criticism. [zzzz99]{} Ardavan, H. 1992, in Proc. IAU Colloq. 128, Magnetospheric Structure and Emission Mechanisms of Radio Pulsars, ed. T. H. Hankins, J. M. Rankin, & J. A. Gil (Poland: Pedagogical University Press), 316 Ardavan, H. 1994, , 268, 361 Argyle, E., & Gower, J. F. R. 1972, , 175, L89 Asséo, E. 1993, , 264, 940 Backer, D. C., Dexter, M., Zepka, A., Ng, D., Wertheimer, D. J., Ray, P. S., & Foster, R. S. 1997, , 109, 61 Backer, D. C., & Wong, T. 1996, in Proc. IAU Colloq. 160, Pulsars: Problems and Progress, ed. S. Johnston, M. A. Walker, & M. Bailes (San Francisco: ASP), 87 Bartel, N., & Sieber, W. 1978, , 70, 307 Cheng, A., & Ruderman, M. A. 1977, , 212, 800 Comella, J. M., Craft, H. D., Lovelace, R. V. E., Sutton, J. M., & Tyler, G. L. 1969, Nature, 221, 453 Cordes, J. M. 1981, in IAU Symp. 95, Pulsars: 13 Years of Research on Neutron Stars, ed. W. Seiber & R. Wielebinski (Dordrecht: D. Reidel), 115 Cordes, J. M., Hankins, T. H., & Moffett, D. 1998, in preparation da Costa, A. A., & Kahn, F. D. 1985, , 215,701 Eilek, J. A. 1996, in Proc. IAU Colloq. 160, Pulsars: Problems and Progress, ed. S. Johnston, M. A. Walker, & M. Bailes (San Francisco: ASP), 155 Eilek, J. A. 1998, preprint, , submitted Friedman, J., & Boriakoff, V. 1990, in Proc. IAU Colloq. 128, Magnetospheric Structure and Emission Mechanisms of Radio Pulsars, ed. T. H. Hankins, J. M. Rankin, & J. A. Gil (Poland: Pedagogical University Press), 347 Goldstein, S. J., & Meisel, D. D. 1969, Nature, 224 Hankins, T. H. 1971, , 169, 487 Hankins, T. H., & Moffett, D. A. 1998, , 30, 903 Hankins, T. H., & Rickett, B. J. 1975, Meth. in Comp. Phys., 14, 55 Heiles, C., Campbell, D. B., & Rankin, J. M. 1970, Nature, 226, 529. Heiles, C., & Rankin, J. M. 1971, Nature, Physical Science, 231, 97 Hester, J. J. [[et al. ]{}]{}1996, , 456, 225 Isaacman, R., & Rankin, J. M. 1977, , 214, 214 Kazbegi, A., Machabeli, G., Melikidze, G., & Shukre, C. 1991, , 309, 515 Lundgren, S. C. 1994, Ph.D. Thesis, Cornell University Lundgren, S., Cordes, J. C., Ulmer, M., Foster, R., & Hankins, T. H. 1995, , 453, 433 Luo, Q., & Melrose, D. B. 1995, , 276, 372 Luo, Q., & Melrose, D. B. 1992, , 258, 616 Lyne, A. G., & Thorne, D. J. 1975, , 172, 97L Machabeli, G. Z., & Usov, V. V. 1979, Sov.Astron.Lett., 5, 238 Melrose, D. B. 1996, in Proc. IAU Colloq. 160, Pulsars: Problems and Progress, ed. S. Johnston, M. A. Walker, & M. Bailes (San Francisco: ASP), 139 Melrose, D. B. 1978, , 225, 557 Moffett, D. A., & Hankins, T. H. 1996, , 468, 779 Moffett, D. A. 1997, Ph.D. Thesis, New Mexico Institute of Mining and Technology Romani, R., & Yadigaroglu, I.-A. 1995, , 438, 314 Rowe, E. T. 1995, , 296, 275 Staelin, D. H., & Reifenstein, E. C. 1968, Science, 162, 1481 Staelin, D. H., & Sutton, J. M. 1970, Nature, 226, 69 Sutton, J. M., Staelin, D. H., & Price, R. M. 1971, in IAU Symp. 46, The Crab Nebula, ed. R.D. Davies & F.G. Smith (Dordrecht: Reidel), 97 Taylor, J. H., & Weisberg, J. M. 1989, , 345, 434 Weatherall, J. C. 1996, in Proc. IAU Colloq. 160, Pulsars: Problems and Progress, ed. S. Johnston, M. A. Walker, & M. Bailes (San Francisco: ASP), 205 Williamson, I. P. 1974, , 166, 499 [^1]: Pulsar emission profiles are generally given in units of [flux]{} (Jy) even though in the context of rotating neutron stars one actually samples a one dimensional cut of the [specific intensity]{} pattern (Jy sr$^{-1}$). The integral of emission over a pulse component in the latter case would be its flux, while in the former and conventional case one quantifies the integrated component emission in units of energy (Jy-s). [^2]: The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc. [^3]: A partial technical description of the GBPP is given in ([@b97]). [^4]: Note that in this paper the term “broadband” connotes simultaneous emission over wide range of radio frequencies with a ratio of amplitudes comparable to that of the average pulse. However, the observed emission has a very steep spectrum, even if weighted by frequency to obtain total power, and is therefore narrow band in an absolute sense.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Quantum Weak Energy Inequalities (QWEIs) are results which limit the extent to which the smeared renormalised energy density of a quantum field can be negative. On globally hyperbolic spacetimes the massive quantum Dirac field is known to obey a QWEI in terms of a reference state chosen arbitrarily from the class of Hadamard states; however, there exist spacetimes of interest on which state-dependent bounds cannot be evaluated. In this paper we prove the first QWEI for the massive quantum Dirac field on four dimensional globally hyperbolic spacetime in which the bound depends only on the local geometry; such a QWEI is known as an absolute QWEI.' author: - \| Calvin J. Smith[^1]\ School of Mathematical Sciences, University College Dublin,\ Belfield, Dublin 4, Ireland date: 15 May 2007 title: 'An absolute quantum energy inequality for the Dirac field in curved spacetime' --- \[section\] \[theorem\][Lemma]{} \[theorem\][Corollary]{} \[theorem\][Claim]{} \[theorem\][Proposition]{} \[theorem\][Proposal]{} \[theorem\][Example]{} \[theorem\][Summary]{} Introduction ============ When formulating the classical theory of general relativity it is necessary to impose certain energy conditions on the source matter fields being considered. The most commonplace of these energy conditions is the weak energy condition, $T_{ab}k^ak^b \geq 0$ for every timelike vector field $k$, which entails that observers only encounter positive energy densities. However, it has been known since 1965 that, unlike most classical physics models, no (Wightman) quantum field theory can obey pointwise energy conditions [@epstein]. Moreover, it is possible to show that the negative energy density arising from a quantum field theoretic source is unbounded (from below) in magnitude [@fewster; @lisbon]. This startling feature of quantum field theory is often used, in the context of the semi-classical Einstein equation $G_{ab}=8\pi G\langle T^{\mathrm{ren}}_{ab}\rangle_\omega$, to support so-called ‘designer spacetimes’ like Alcubierre’s warp drive [@alcubierre] or traversable worm hole geometries. Following Ford’s [@ford] observation that it is possible to bound the magnitude and duration of the flux of negative energy of a quantum field source, work began in earnest to prove that the averaged (expectation value of the) energy density of a quantum field was bounded from below. A suitable definition, sufficient for our purposes[^2], is that a worldline quantum weak energy inequality (QWEI) is a result of the form $$\label{wqwei} \int_{\mathbb{R}}{\mathrm{d}}\tau \, \langle \rho^\mathrm{ren} \rangle_\omega (\gamma(\tau)) F(\tau) \geq - \mathcal{B} > -\infty$$ where $F$ is some appropriately chosen sampling function, $\gamma:{\mathbb{R}}\mapsto {\mathcal{M}}$ is a timelike worldline and $\langle \rho^\mathrm{ren}\rangle_\omega$ is the (expectation value of the) energy density of the quantum field in a state $\omega$. In this discussion we shall exclusively consider the massive quantum Dirac field in a smooth four-dimensional globally hyperbolic spacetime $({\mathcal{M}},g)$. Moreover, we shall only consider the Hadamard states of the Dirac field as this is a sufficient class of states to renormalise the stress energy density. Typically, the bound $\mathcal{B}$ featuring in (\[wqwei\]) is a function of another state of the theory usually called a reference state; these QWEIs are known as difference QWEIs. Due to the work of Fewster and his collaborators [@dawson; @fewster; @fewster; @verch], difference QWEIs are known in great generality for the Dirac field in curved spacetime. (For a brief review of QWEIs for other fields the reader is directed to section one of [@fewster; @smith] and the references therein). Difference QWEIs have been instrumental in constraining the likelihood of designer spacetime manifestation; however, there exist spacetimes on which one does not know how to write down the closed form expressions for states necessary for the evaluation of the difference QWEI bound. Indeed, the warp drive is an example of a spacetime on which it is not currently known how to obtain explicit expressions for Hadamard states. Therefore, it is desirable to have a lower bound $\mathcal{B}$ which is state independent and constructed only from the local geometry; such a bound is known as an absolute QWEI. Currently, for the Dirac field, absolute QWEIs are known only for the conformally invariant [@vollick] and massive field in two-dimensions [@dawson] and the massive field in four-dimensional flat spacetime [@fewster; @mistry]. In this discussion we state and prove the first absolute QWEI for the massive Dirac field in arbitrary four-dimensional globally hyperbolic spacetime. The argument is an adaptation of Fewster’s earlier work with Verch [@fewster; @verch] and Dawson [@dawson; @fewster] and is to be viewed as a companion to the analogous result for the Klein-Gordon field [@fewster; @smith]. Our result may be stated as follows: Let $({\mathcal{M}},g)$ be a classical curved four-dimensional spacetime. Here $\mathcal{M}$ is a four-dimensional smooth manifold (assumed Hausdorff, paracompact and without boundary) with a Lorentz metric $g_{ab}$ of signature ($+---$). Furthermore, we require $({\mathcal{M}},g)$ to be globally hyperbolic, that is ${\mathcal{M}}$ contains a Cauchy surface. In addition we assume that an orientation, time orientation and spin structure have been chosen. It may be shown that on such a background one may formulate the quantum Dirac field and a notion of Hadamard states. The essential feature of Hadamard states is that they all share a common singularity structure; in particular their two-point functions, and their Dirac adjoints, have a local and covariantly determined singular expansion. We denote the Hadamard series corresponding to the singularity structure of the Dirac two-point function ${\mathcal{W}}_\omega$ by ${\,}^\psi H_k^{(+)}$ and that corresponding to the singular structure of the adjoint Dirac two-point function ${\mathcal{W}}^\Gamma_\omega$ by ${\,}^\psi H^{(-)}_k$. The salient feature of such states is that one may define a finite stress energy density $\langle \rho^{\mathrm{fin}}\rangle_\omega(x)$ by using ${\mathcal{W}}_\omega-{\,}^\psi H_k^{(+)}$ and a point-splitting prescription. Indeed, define an operator $\rho^\mathrm{split}$ such that the finite contribution to the energy density is given by $\rho^\textrm{split}$ acting on the regularised two-point function of the Dirac field ${\mathcal{W}}_\omega - {\,}^\psi H_k^{(+)}$; i.e. $$\langle \rho^{\mathrm{fin}}\rangle_\omega (x) := \lim_{x'\rightarrow x} [\rho^\mathrm{split} ( {\mathcal{W}}_\omega -{\,}^\psi H_1^{(+)} ) ](x,x') \, .$$ The precise form of $\rho^\textrm{split}$ is given in §\[energy section\]. The quantity $\langle \rho^{\mathrm{fin}}\rangle_\omega$ is equal to the renormalised energy density modulo a local curvature term; we shall return to this issue later after proving our main result in theorem \[main result\]. Our result then reads for any real valued $f\in C^\infty_0({\mathbb{R}})$ and Hadamard state $\omega$ $$\int_{\mathbb{R}}{\mathrm{d}}\tau \, \langle \rho^\mathrm{ren} \rangle_\omega (\gamma(\tau)) f^2(\tau) \geq - \mathcal{B} \, ,$$ modulo local curvature terms, where $\mathcal{B}$ is of the form $$\begin{aligned} \mathcal{B} &=& \int_{{\mathbb{R}}^+}\frac{{\mathrm{d}}\xi}{2\pi}\xi \bigg[ f\otimes f \, \vartheta^{\,}^\psi{\mathrm{H}}^{(+)}_4 \bigg]^\wedge(-\xi,\xi) {\nonumber \\}&\,& \quad - \int_{{\mathbb{R}}^-}\frac{{\mathrm{d}}\xi}{2\pi} \xi \bigg[ f\otimes f \,\vartheta^\big(i\mathrm{S}_{\mathrm{sp}}-{\,}^\psi{\mathrm{H}}^{(+)}_4 \big) \bigg]^\wedge(-\xi,\xi) \, .\end{aligned}$$ Here $\vartheta=\gamma\otimes\gamma$, $S_{\mathrm{sp}}$ is the fundamental solution to the Dirac equation, ${\,}^\psi{\mathrm{H}}^{(\pm)}_k$ are scalar distributions created from ${\,}^\psi H^{(\pm)}_k$ and $\hat{\,}$ denotes the Fourier transform which in our conventions is given by $$\hat{f}(\xi) = \int {\mathrm{d}}x \, f(x) \, e^{i\xi\cdot x} \, .$$ The structure of this paper is as follows: In section \[dirac background\] we present a review of the formulation of the classical (§\[dirac background 1\]-\[dirac background 2\]) and quantum (§\[quantum dirac hadamard\]) Dirac fields and their Hadamard states. We then direct our attention to a microlocal description of the Hadamard series for the Dirac field in section \[microlocal review\]; in particular we review the Sobolev wave-front set and its properties (§\[microlocal review 1\]) before applying the theory to the matter in hand and obtaining estimates on the singularities of the Hadamard series (§\[microlocal results\]). Finally, a point-splitting lemma is presented in section \[splitting section\] before our main result is stated in section \[punch line\]. The Dirac field in curved spacetime {#dirac background} =================================== The reader who is familiar with the formalism necessary to describe the classical Dirac field on a curved background is encouraged to skip ahead to section §\[quantum dirac hadamard\]. Spin structures and spinors on curved spacetimes {#dirac background 1} ------------------------------------------------ We begin by reviewing the geometry necessary to discuss the Dirac field in a curved spacetime. We shall employ the algebraic framework for describing the Dirac quantum field in a classical curved four-dimensional spacetime $(\mathcal{M},g)$. Here $\mathcal{M}$ is a four-dimensional smooth manifold (assumed Hausdorff, paracompact and without boundary) with a Lorentz metric $g_{ab}$ of signature ($+---$). Furthermore, we require $({\mathcal{M}},g)$ to be globally hyperbolic, that is ${\mathcal{M}}$ contains a Cauchy surface. Where index notation is used, Latin indices will run over the range $0,1,2,3$ unless explicitly stated otherwise, while Greek characters will denote frame indices and also run over $0,1,2,3$ unless explicitly stated otherwise. We employ units in which $c=\hbar=1$. In Minkowski spacetime the spinors are nothing more than the spin-half representation of the Poincaré group, however, a general manifold does not exhibit this symmetry globally: Therefore, the usual (i.e. Minkowski spacetime) interpretation of a spinor as being a quadruple of complex numbers at each point in spacetime does not generalise under the replacement $({\mathbb{R}}^4,\eta)\mapsto ({\mathcal{M}},g)$. A rigorous formulation of spinors on a manifold is given in terms of fibre bundles where the spin group is the structure group. (For a review of the necessary concepts related to fibre bundles, and in particular spinor bundles, the reader is directed to [@nakahara]). We shall review basic facts about the Dirac matrices and the Lorentz and spin groups in Minkowski spacetime, and use a local frame to generalise the results to a curved spacetime. What follows is based on [@dawson; @fewster] and benefits from the elaborations in [@dimock; @dirac; @fewster; @verch]. We begin by summarising several groups which appear in our discussion. The Lorentz group $O(1,3) = \{ \Lambda \in \mathrm{GL}_4({\mathbb{R}}) \mid \eta_{\alpha\beta}\Lambda^\alpha_{\,\,\,\gamma}\Lambda^\beta_{\,\,\,\delta}=\eta_{\gamma\delta} \}$ has the subgroup $\mathfrak{L}^\uparrow_+$, $$\mathfrak{L}^\uparrow_+ = \{ \Lambda \in O(1,3) \mid \det\Lambda=1 \, \& \, \Lambda^0_{\,\,\,0}>0 \}$$ called the proper orthochronous Lorentz group. The Dirac gamma matrices $\gamma_\alpha$ satisfy the Clifford algebra relation $\{\gamma_\alpha , \gamma_\beta \} = 2\eta_{\alpha\beta}\mathbbm{1}$ and are said to belong to a standard representation if $\gamma^\dagger_0=\gamma_0$ and $\gamma^\dagger_{\alpha}=-\gamma_{\alpha}$ for $\alpha=1,2,3$. From here on we shall assume that our Dirac matrices belong to a standard representation. The spin group, ${\mathrm{Spin }}(1,3)$, is defined by $${\mathrm{Spin }}(1,3) = \{ S \in \mathrm{SL}_4({\mathbb{C}}) \mid S\gamma_\alpha S^{-1} = \gamma_\beta \Lambda^\beta_{\,\,\,\alpha} \textrm{ for some }\Lambda\in\mathcal{L} \} \ ,$$ and is known to be a two-to-one cover of $\mathfrak{L}^\uparrow_+$, i.e. the mapping $S\mapsto \Lambda(S)$ is a two-to-one covering homomorphism from the identity connected component ${\mathrm{Spin }}_0(1,3)$, of ${\mathrm{Spin }}(1,3)$, to $\mathfrak{L}^\uparrow_+$ with kernel $\{\mathbbm{1},-\mathbbm{1} \}$. We now direct our attention to a curved spacetime setting. The frame bundle ${F{\mathcal{M}}}$ is the bundle of oriented and time-oriented tetrads $\{ e_\alpha^a \}_{\alpha=0,1,2,3}$ over spacetime $({\mathcal{M}},g)$ with the convention that $e_0^a$ is a future pointing timelike vector; moreover, ${F{\mathcal{M}}}$ is a principal $\mathcal{L}^\uparrow_+$ bundle whose right action is given by $(R_\Lambda e)_\alpha = e_\beta \Lambda^\beta_{\,\,\,\alpha}$. A spin structure on $({\mathcal{M}},g)$ is a principal ${\mathrm{Spin }}_0(1,3)$ bundle, ${S{\mathcal{M}}}$, over $({\mathcal{M}},g)$ equipped with a fibre homomorphism $\varphi:{S{\mathcal{M}}}\mapsto {F{\mathcal{M}}}$ such that $\varphi\circ R_S=R_{\Lambda(S)}\circ\varphi$, i.e. $\varphi$ intertwines the right action of the structure group on these bundles. Spin structures are not unique, however two such structures, ${S{\mathcal{M}}}$ and $\widetilde{{S{\mathcal{M}}}}$ equipped with $\varphi$ and $\widetilde\varphi$ respectively can be said to be equivalent if there is an isomorphism $\iota : {S{\mathcal{M}}}\mapsto \widetilde{{S{\mathcal{M}}}}$ such that $\varphi=\widetilde{\varphi}\circ\iota$. It is worth pointing out that spin structures do not exist in general for an arbitrary manifold; their existence is determined by the second Stiefel-Whitney class. In essence, the requirement that the second Stiefel-Whitney class vanishes ensures consistency between the (transition functions of the) fibre group of the tangent bundle and the (lift to the transition functions of the) spin group. It is known that there exist spin structures over orientable manifolds if and only if the second Stiefel-Whitney class vanishes and that every four-dimensional globally hyperbolic manifold admits a spin structure. We now assume that an arbitrary spin structure has been chosen and is fixed for the remainder of this discussion. We may now define spinor fields on a curved manifold by saying they are sections of the associated ${\mathrm{Spin }}_0(1,3)$ bundle $${D{\mathcal{M}}}= {S{\mathcal{M}}}\ltimes_{{\mathrm{Spin }}_0(1,3)}{\mathbb{C}}^4 \, .$$ The fibre of ${D{\mathcal{M}}}$ at $x\in{\mathcal{M}}$ is the equivalence class $[T,z]_x$ where $T\in S_x{\mathcal{M}}$ and $z\in{\mathbb{C}}^4$ is a column vector and the equivalence relation is: $[\widetilde{T},\widetilde{z}]_x=[T,z]_x$ if and only if $\widetilde{T}=R^{-1}_ST$ and $\widetilde{z}=Sz$ for some $S\in {\mathrm{Spin }}_0(1,3)$. The bundle $D{\mathcal{M}}$ has fibre ${\mathbb{C}}^4$ at every point and left action given by $L_S [T,z]_x = [T,Sz]_x$. The dual bundle $$D^{\mathcal{M}}=S{\mathcal{M}}\ltimes_{{\mathrm{Spin }}_0({\mathcal{M}},g)}{{\mathbb{C}}}^4 \, ,$$ where ${{\mathbb{C}}}^4$ is the set of complex row* 4-vectors, is constructed similarly and its fibres are the equivalence classes $[T,z^{\mathrm{t}}]_x^$, $z^{\mathrm{t}}\in{{\mathbb{C}}}^4$ a row vector, such that $[\widetilde{T},\widetilde{z^{\mathrm{t}}}]_x^ = [T,z^{\mathrm{t}}]_x^$ if and only if $\widetilde{T}=R^{-1}_ST$ and $\widetilde{z^{\mathrm{t}}}=z^{\mathrm{t}}S^{-1}$ for some $S\in{\mathrm{Spin }}_0(1,3)$. Just as the sections of ${D{\mathcal{M}}}$ are called spinors, the sections of $D^{\mathcal{M}}$ are called cospinors. We shall refer to test spinors as being the smooth and compactly supported sections of $D{\mathcal{M}}$, the space of which we denote $C^\infty_0(D{\mathcal{M}})$; test cospinors are similarly defined and are elements of $C^\infty_0(D^{\mathcal{M}})$. As expected, there exists a natural pairing between spinors and cospinors: Set $v_x = [T,z_1^{\mathrm{t}}]_x^$ and $u_x\in [T,z_2]_x$, then $v_x(u_x) = z_1^{\mathrm{t}}z_2$ is a scalar. We are now in a position to define the Dirac adjoint operation ${\,}^+ : D{\mathcal{M}}\mapsto D^{\mathcal{M}}$ which is given by $$[T,z]^+_x = [T,z^\dagger \gamma_0]^_x \, .$$ Any local section $E:{\mathcal{M}}\mapsto S{\mathcal{M}}$ of $S{\mathcal{M}}$ determines a local frame $e^a_\alpha$ by $\varphi\circ E$ and local sections $E_A$ of $D{\mathcal{M}}$, such that $E_A(x)=[E_x,z_A]$ where $\{ z_A \}_{A=0,1,2,3}$ is the canonical basis of ${\mathbb{C}}^4$. The dual frames $e_a^\alpha$, $E^A$ are defined through $e^\alpha \cdot e_\beta=\delta^\alpha_\beta$ and $E^A ( E_B)=\delta^A_B$. One may define a mixed tensor-spinor object $\gamma \in C^\infty(T^{\mathcal{M}})\otimes C^\infty(D{\mathcal{M}})\otimes C^\infty(D^{\mathcal{M}})$ by setting its components $\gamma^{\,\,\,A}_{\alpha\,\,\,\,\,B}$ in the frame $e^\alpha_a\otimes E_A\otimes E^B$ equal to the matrix elements $(\gamma_\alpha)^{A}_{\,\,\,\,B}$. For example, it can be shown that $$\label{split gamma 0} \gamma_0=\delta^{AB}E_A\otimes E^+_B \, .$$ The Dirac equation {#dirac background 2} ------------------ The metric $g$ determines a connection $\Gamma$ in the usual way via the covariant derivative operator $\nabla: C^\infty(T{\mathcal{M}})\mapsto C^\infty(T^{\mathcal{M}}\otimes T{\mathcal{M}})$. One may equally define a connection $\sigma$ and covariant derivative on the spinor and cospinor bundle, which we also denote by $\nabla$, $$\nabla: \left\{ \begin{array}{c} C^\infty(D{\mathcal{M}}) \\ C^\infty(D^{\mathcal{M}}) \end{array}\right. \mapsto \left\{ \begin{array}{c} C^\infty(T^{\mathcal{M}}\otimes D{\mathcal{M}}) \\ C^\infty(T^{\mathcal{M}}\otimes D^{\mathcal{M}}) \end{array}\right. \, .$$ Given a local section $E$, $f\in C^\infty(D{\mathcal{M}})$ may be decomposed $f=f^AE_A$, then $\nabla_af$ has components $$\nabla_\alpha f^A = \partial_\alpha f^A + \sigma^A_{\alpha B}f^B$$ in the frame $e^\alpha_a\otimes E_A$ where the connection $\sigma$ has elements given by $\sigma^A_{\alpha B} = -\frac{1}{4}\Gamma^\beta_{\alpha\delta}\gamma_{\beta\,\,\,\,C}^{\,\,A}\gamma^{\delta C}_{\,\,\,\,\,\, B}$. We are now in a position to define the equation of motion the spinors will satisfy, i.e. the Dirac equation. The Dirac operator $${\nabla\hspace{-3mm}{/}}: \left\{ \begin{array}{l} C^\infty(D{\mathcal{M}}) \\ C^\infty(D^{\mathcal{M}}) \end{array} \right. \mapsto \left\{ \begin{array}{l} C^\infty (D{\mathcal{M}}) \\ C^\infty(D^{\mathcal{M}}) \end{array} \right.$$ maps (co)spinor fields into (co)spinor fields by $$\begin{aligned} {\nabla\hspace{-3mm}{/}}f &=& ({\nabla\hspace{-3mm}{/}}f)^A E_A = \eta^{\alpha\beta}\gamma^{\,\,\, A}_{\alpha\,\,\,\, B}(\nabla_\beta f^B) E_A \quad \forall f \in C^\infty(D{\mathcal{M}}) \\ {\nabla\hspace{-3mm}{/}}h &=& ({\nabla\hspace{-3mm}{/}}h)_B E^B = \eta^{\alpha\beta}(\nabla_\beta h_C)\gamma_{\alpha\,\,\,\, B}^{\,\,\, C} E^B \quad \forall h \in C^\infty(D^{\mathcal{M}}) \, .\end{aligned}$$ The spinor field $f\in C^\infty(D{\mathcal{M}})$ is said to satisfy the Dirac equation if $(-i{\nabla\hspace{-3mm}{/}}+\mu)f=0$ where the constant $\mu\geq 0$ is interpreted as the mass of the field. Similarly, the cospinor field $h \in C^\infty(D^{\mathcal{M}})$ is said to satisfy the Dirac equation if $(i{\nabla\hspace{-3mm}{/}}+\mu)h=0$. Even though the Dirac operator is not normally hyperbolic it is possible to find unique advanced and retarded fundamental solutions on arbitrary globally hyperbolic spacetimes. The key element in this analysis is the Lichnérowicz identity, $$\label{lich} P = (-i{\nabla\hspace{-3mm}{/}}+\mu)(i{\nabla\hspace{-3mm}{/}}+\mu)$$ where $P=\nabla^2+R/4+\mu^2$ is the so-called supersymmetrically coupled Klein-Gordon operator for spinors, which relates Dirac operators to normally hyperbolic ones. The $R$ featuring in (\[lich\]) is the Ricci scalar. It is known that there exist unique advanced $E^-_P$ and retarded $E^+_P$ fundamental solutions to any normally hyperbolic operator $P$ on globally hyperbolic spacetimes. Hence, for the spinor field, one has the following fundamental solutions: $S^\pm_\textrm{sp}=(i{\nabla\hspace{-3mm}{/}}+\mu)E^\pm_P$. To be explicit, $S^\pm_\textrm{sp}$ are continuous operators $S^\pm_\mathrm{sp}: C^\infty_0(D{\mathcal{M}})\mapsto C^\infty(D{\mathcal{M}})$ such that $$(-i{\nabla\hspace{-3mm}{/}}+\mu)S^\pm_\mathrm{sp} f = S^\pm_\mathrm{sp}(-i{\nabla\hspace{-3mm}{/}}+\mu)f =f$$ satisfying ${\textrm{supp }}(S^\pm_\mathrm{sp} f)\subset J^\pm ({\textrm{supp }}f)$. Clearly, there is a similar construction for the cospinor field resulting in $S^\pm_\textrm{cosp}$. The advanced-minus-retarded fundamental solution for spinors is $S_\mathrm{sp}=S^-_\mathrm{sp}-S^+_\mathrm{sp}$ and for cospinors $S_\mathrm{cosp}=S^-_\mathrm{cosp}-S^+_\mathrm{cosp}$. An additional antilinear map $\Gamma :D^{\mathcal{M}}\oplus D{\mathcal{M}}\mapsto D^{\mathcal{M}}\oplus D{\mathcal{M}}$ acts by $$\Gamma\left( \begin{array}{c} h \\ f \end{array} \right) = \left( \begin{array}{c} f^+ \\ h^+\end{array} \right) \, ,$$ where $f\in C^\infty_0(D{\mathcal{M}})$ and $h\in C^\infty_0(D^{\mathcal{M}})$. The map $\Gamma$ makes explicit the symmetry between spinors and cospinors in this dual setting. The quantum Dirac field and Hadamard states {#quantum dirac hadamard} ------------------------------------------- To define the field algebra we denote by $\mathcal{D}(D^{\mathcal{M}}\oplus D{\mathcal{M}}) = C^\infty_0(D^{\mathcal{M}})\oplus C^\infty_0(D{\mathcal{M}})$ the space of all test cospinors and test spinors on which the operators $$D := i\left( \begin{array}{cc} i{\nabla\hspace{-3mm}{/}}+\mu & 0 \\ 0 & -i{\nabla\hspace{-3mm}{/}}+ \mu \end{array}\right) \, \quad S := i\left( \begin{array}{cc} S_\mathrm{cosp} & 0 \\ 0 & S_\mathrm{sp} \end{array}\right)$$ act. The elements $\mathcal{F}\in\mathcal{D}(D^{\mathcal{M}}\oplus D{\mathcal{M}})$ may be used to label a set of abstract objects $\{ \Psi(\mathcal{F}) \mid \mathcal{F} \in \mathcal{D}(D^{\mathcal{M}}\oplus D{\mathcal{M}})\}$ which, when equipped with $\mathbbm{1}$, generates a unital \-algebra $\mathfrak{F}$. We define the algebra of smeared fields* $\mathfrak{F}({\mathcal{M}},g)$ to be $\mathfrak{F}$ quotiented by\ ------ -------------------------------------------------------------------------------------------------------------------------------- i) Adjoint, $\Psi (\mathcal{F})^* = \Psi(\Gamma\mathcal{F})$; ii) Linearity, $\Psi(\alpha_1 \mathcal{F}_1 + \alpha_2\mathcal{F}_2) = \alpha_1\Psi(\mathcal{F}_1)+\alpha_2\Psi(\mathcal{F}_2)$; iii) The field equation, $\Psi(D\mathcal{F})=0$; iv) Canonical anticommutation relation, $\{\Psi(\mathcal{F}_1),\Psi(\mathcal{F}_2) \}=iS(\mathcal{F}_1,\mathcal{F}_2)\mathbbm{1}$. ------ -------------------------------------------------------------------------------------------------------------------------------- \ Here $\mathcal{F}, \mathcal{F}_1,\mathcal{F}_2\in\mathcal{D}(D^{\mathcal{M}}\oplus D{\mathcal{M}})$ and $\alpha_1,\alpha_2\in{\mathbb{C}}$. It is relation (iv) that quantises the theory. The usual Dirac field $\psi$ and its adjoint $\psi^+$ are special cases of the above construction. For $h\in C^\infty_0(D^{\mathcal{M}})$, $f\in C^\infty_0(D{\mathcal{M}})$ we define $$\psi(h) := \Psi\left( \begin{array}{c} h \\ 0 \end{array} \right) \quad \textrm{and} \quad \psi^+(f) = \Psi \left( \begin{array}{c} 0 \\ f \end{array} \right)$$ which we interpret as smeared fields. A state $\omega : \mathfrak{F}({\mathcal{M}},g)\mapsto {\mathbb{C}}$ is a linear functional which is positive, i.e. $\omega(A^A)\geq 0$ $\forall A\in\mathfrak{F}({\mathcal{M}},g)$, and normalised such that $\omega(\mathbbm{1})=1$. We shall restrict our attention to states for which the two-point function, defined by $$\omega(\Psi(\mathcal{F}_1)\Psi(\mathcal{F}_2)) \quad \forall \mathcal{F}_1,\mathcal{F}_2 \in {\mathcal{D}}(D^{\mathcal{M}}\oplus D{\mathcal{M}}) \, ,$$ is a distribution on $\mathcal{D}(D^{\mathcal{M}}\oplus D{\mathcal{M}})\otimes \mathcal{D}(D^{\mathcal{M}}\oplus D{\mathcal{M}})$. Associated to each state $\omega$ we define what we shall call the Dirac and Dirac adjoint two-point functions ${\mathcal{W}}_\omega, {\mathcal{W}}^\Gamma_\omega \in {\mathcal{D}}'(D{\mathcal{M}}\times D^{\mathcal{M}})$ respectively by $${\mathcal{W}}_\omega(f,h)=\omega (\psi^+(f) \psi(h)) \quad \textrm{and} \quad {\mathcal{W}}^\Gamma_\omega(f,h)=\omega(\psi(h)\psi^+(f)) \, .$$ As a consquence of the positivity of states we immediately have that $${\mathcal{W}}(f,f^+) \geq 0 \quad \forall f \in C^\infty_0(D{\mathcal{M}}) \quad \textrm{and} \quad {\mathcal{W}}^\Gamma(h^+,h)\geq 0 \quad \forall h\in C^\infty(D^{\mathcal{M}}) \, .$$ The covariant anticommutation relation in terms of ${\mathcal{W}}_\omega$ and ${\mathcal{W}}^\Gamma_\omega$ is equally expressed $${\mathcal{W}}_\omega + {\mathcal{W}}^\Gamma_\omega = iS_{\mathrm{sp}}\, .$$ It is clear from our discussion concerning the advanced and retarded fundamental solutions of the Dirac field that one may use the Lichnérowicz identity to define a notion of Hadamard state for the Dirac field. In order to give the precise formulation of the Hadamard series construction we must first discuss some geometry and here we follow [@sahlmann2]. We denote by $\mathfrak{X}\subset {\mathcal{M}}\times{\mathcal{M}}$ the set $$\begin{aligned} \mathfrak{X} &=& \{ (x,x^\prime)\in{\mathcal{M}}\times{\mathcal{M}}\mid x,x^\prime \text{ are causally related and} {\nonumber \\}&\,& \qquad J^+(x)\cap J^-(x^\prime) \,\textrm{and}\, J^-(x)\cap J^+(x^\prime) \text{ are contained} {\nonumber \\}&\,& \qquad \text{within a convex normal neighbourhood} \} \, .\end{aligned}$$ Let $X\subset \mathfrak{X}$ be an open subset of $\mathfrak{X}$ such that between any pair $(x,x^\prime)\in X$ there exists a unique geodesic connecting them such that the (signed) geodesic separation of points defines a smooth function $\sigma$ on $X$. We make the additional requirement that the Hadamard construction (to be described shortly) can be carried out on $X$. Subject to all these requirements, we call $X$ \[notation reg dom\]a regular domain. We define two sequences of distributions $\{ H_k^{(\pm)}\}_{k=0,1,2,\dots} \in {\mathcal{D}}^\prime (X)$ by $$\begin{aligned} \label{parametrix} H_k^{(\pm)}(x,x^\prime) &=& \frac{1}{4\pi^2} \bigg\{ \frac{\Delta^\frac{1}{2}(x,x^\prime)}{\sigma_\pm (x,x^\prime)} + \sum^k_{j=0}v_j (x,x^\prime)\frac{\sigma^j(x,x^\prime)}{\ell ^{2(j+1)}}\ln\bigg(\frac{\sigma_\pm (x,x^\prime)}{\ell^2}\bigg) \nonumber \\ &\,& \qquad +\sum^k_{j=0}w_j(x,x^\prime)\frac{\sigma^j(x,x^\prime)}{\ell^{2(j+1)}}\bigg\} \, \end{aligned}$$ where we have introduced a length scale $\ell$ to make $\sigma/\ell^2$ dimensionless. By $F(\sigma_\pm)$, $F$ some function, we mean $$F(\sigma_\pm) = \lim_{\epsilon\rightarrow 0^+}F(\sigma_{\pm\epsilon})$$ in the sense of distributions, where $\sigma_{\pm\epsilon}(x,x^\prime) = \sigma(x,x^\prime)\pm 2i\epsilon(t(x)-t(x^\prime))+\epsilon^2$ and $t$ is a time function on $X$. The functions $\Delta$, known as the van Vleck-Morette determinant, $v_j$ and $w_j$ are found by fixing $x^\prime$ and applying $P\otimes\mathbbm{1}$ to $H_k^{(+)}$ and equating all the coefficients of $1/\sigma_+$, $1/\sigma^2_+$, $\ln \sigma_+$ etc to zero; moreover, they are all spinors (i.e., they carry internal indices). This determines a system of differential equations known as the Hadamard recursion relations[^3]. In $X$ the system of differential equations uniquely determines the $\{ v_j \}_{j=0,\dots,k}$ series. The $\{ w_j\}_{j=0,\dots,k}$ series is specified once the value of $w_0$ is fixed; we adopt Wald’s prescription that $w_0=0$ [@wald]. Let $u$ be of Hadamard form for the operator $P$, i.e. within a regular domain $X$ one has $u = H_k^{(+)}$ modulo $C^k(X)$ for each $k\in{\mathbb{N}}$; the distribution $u$ is sometimes referred to as the auxiliary two-point function. A state $\omega$ on the algebra of smeared (Dirac) fields is said to be Hadamard if its associated two-point function ${\mathcal{W}}_\omega$ is of the form ${\mathcal{W}}_\omega = (i{\nabla\hspace{-3mm}{/}}+\mu)\otimes\mathbbm{1}u$ where $u$ is an auxiliary two-point function. Consequently, we have, within a regular domain $X$, that $${\mathcal{W}}_\omega = {\,}^\psi H_k^{(+)} \quad \textrm{modulo} \quad C^k(X)$$ where ${\,}^\psi H^{(+)}_k = (i{\nabla\hspace{-3mm}{/}}+\mu)\otimes\mathbbm{1}H^{(+)}_{k+1}$. We also define ${\,}^\psi H^{(-)}_k = (i{\nabla\hspace{-3mm}{/}}+\mu) \otimes\mathbbm{1}H^{(-)}_{k+1}$. Hence, within a regular domain $X$, for any Hadamard state $\omega$ we have the following identities: $$\begin{aligned} \label{dirac hadamard relations} {\mathcal{W}}_\omega &=& {\,}^\psi H_k^{(+)} \quad \textrm{modulo } C^k(X) \, , \label{dirac hadamard relations 1} \\ {\mathcal{W}}^\Gamma_\omega &=& -{\,}^\psi H^{(-)}_k \quad \textrm{modulo } C^k(X) \label{dirac hadamard relations 2}\, , \label{sign note}\\ iS_{\mathrm{sp}}&=& {\,}^\psi H^{(+)}_k - {\,}^\psi H^{(-)}_k \quad \textrm{modulo } C^k(X) \, . \label{dirac hadamard relations 3}\end{aligned}$$ Two remarks are in order: First, note that we require the sign in (\[sign note\]) so as to ensure that the anticommutation relation holds. Second, observe that as $k$ increases ${\mathcal{W}}_\omega - {\,}^\psi H_k^{(+)}$ becomes more regular and that, for sufficiently high $k$, ${\mathcal{W}}_\omega - {\,}^\psi H_k^{(+)}$ has a well defined coincidence limit. The stress energy tensor {#energy section} ------------------------ We open this section with a few remarks about obtaining scalar distributions from spinorial ones as this will be the basis of our analysis of the energy density for the remainder of our discussion. Let $E_A$ be the spinor field derived from a local section $E$ of the spin bundle ${S{\mathcal{M}}}$. Then one can derive matrices ${\mathcal{W}}_{\omega AB}$ and ${\mathcal{W}}^\Gamma_{\omega AB}$ from ${\mathcal{W}}_\omega$ and ${\mathcal{W}}^\Gamma_\omega$ via $${\mathcal{W}}_{\omega AB}(f,f') = {\mathcal{W}}_\omega(fE_A,f'E_B^+) \quad {\mathcal{W}}_{\omega AB}^\Gamma (f,f') = {\mathcal{W}}^\Gamma_\omega (fE_A,f'E^+_B)$$ for all $f,f'\in C^\infty_0({\mathcal{M}})$. Scalar bi-distributions ${\mathrm{W}}_\omega$ and ${\mathrm{W}}^\Gamma_\omega$ (in ${\mathcal{D}}'({\mathcal{M}}\times{\mathcal{M}})$) may be constructed by taking the traces of the matrices, i.e. ${\mathrm{W}}_\omega=\delta^{AB}{\mathcal{W}}_{\omega AB}$ and ${\mathrm{W}}^\Gamma_\omega = \delta^{AB}{\mathcal{W}}^\Gamma_{\omega AB}$. Similarly, one may define a scalar version ${\,}^\psi{\mathrm{H}}_k^{(\pm)}$ of the Hadamard series ${\,}^\psi H_k^{(\pm)}$. The stress energy tensor of the classical spin-1/2 field $\psi$ is given by $$T_{ab} = \frac{i}{2}\bigg( \psi^+\gamma_{(a}\nabla_{b)}\psi - (\nabla_{(a}\psi^+)\gamma_{b)}\psi \bigg) \,$$ where the subscript parentheses denote symmetrisation, i.e $\tau_{(ab)}=(\tau_{ab}+\tau_{ba})/2$. As advertised, we shall concentrate exclusively on proving an absolute QWEI along a timelike worldline. Therefore, we pick a properly parametrised smooth timelike worldline $\gamma:{\mathbb{R}}\mapsto {\mathcal{M}}$ and consider a spacetime tube $\tau_\gamma\subset {\mathcal{M}}$ centred about it, a precise construction of this tube will be given shortly. In $\tau_\gamma$ we may construct a tetrad $\{e^a_\alpha \}_{\alpha=0,1,2,3}$ such that $e^a_0 = \dot{\gamma}^a$ and the remaining $\{ e^a_\alpha\}_{\alpha=1,2,3}$ are orthonormal to this vector. Then, in the (dual) frame $e^\alpha_a\otimes e^\beta_b$, the tensor $T_{ab}$ has components $T_{\alpha\beta}$ given by $$T_{\alpha\beta} = \frac{i}{2}\bigg( \psi^+\gamma_{(\alpha}\nabla_{\beta)}\psi - (\nabla_{(\alpha}\psi^+)\gamma_{\beta)}\psi \bigg) \,$$ from which follows the energy density $\rho$: $$\rho = \frac{i}{2}\bigg( \psi^+\gamma_0\nabla_{0}\psi - (\nabla_{0}\psi^+)\gamma_{0}\psi \bigg) \, .$$ Recall from (\[split gamma 0\]), that the spin frame $E_A$ and its Dirac conjugate satisfy $\delta^{AB}E_A\otimes E_B^+ =\gamma_0$. This enables us to write the classical point-split stress energy density $\rho^\mathrm{split}$ of the Dirac field as $$\begin{aligned} \rho^\mathrm{split}(x,x') &=& \frac{i}{2}\delta^{AB}\bigg( \big( \psi^+E_A \big)\otimes \big( E^+_B e_{0'}\cdot\nabla\psi\big) {\nonumber \\}&\,& \qquad - \big( [e_0\cdot\nabla\psi^+]E_A \big)\otimes \big( E^+_B\psi \big) \bigg)(x,x') \, .\end{aligned}$$ One may then use $\rho^\mathrm{split}$ to define the distributional point-split energy density, also denoted $\rho^\mathrm{split}$: Let $f,f' \in C^\infty_0({\mathcal{M}})$ then $$\begin{aligned} \rho^\mathrm{split}(f',f) &=& \frac{i}{2}\delta^{AB}\int_{{\mathcal{M}}\times{\mathcal{M}}}{\mathrm{dvol}}(x){\mathrm{dvol}}(x') {\nonumber \\}&\,& \quad (f'\otimes f)\bigg( \big( \psi^+E_A \big)\otimes \big( E^+_B e_{0'}\cdot\nabla\psi\big) {\nonumber \\}&\,& \qquad - \big( [e_0\cdot\nabla\psi^+]E_A \big)\otimes \big( E^+_B\psi \big) \bigg)(x,x') \\ &=& \frac{i}{2}\delta^{AB} \bigg( \psi^+(f'E_A)\psi(\nabla\cdot(e_{0'}fE^+_B)) {\nonumber \\}&\,& \qquad - \psi^+(\nabla\cdot(e_0f'E_A))\psi(fE^+_B) \bigg) \label{class rho split dirac}\end{aligned}$$ where we have expressed the right hand side in more traditional distributional language and $u\nabla\cdot v$ denotes the distributional dual of $(\nabla u)\cdot v$. It is now clear that the replacement of $\psi^+\otimes\psi$ in (\[class rho split dirac\]) by the two-point function of any Hadamard state $\omega$ defines the expectation point-split energy density $\langle \rho^\mathrm{split}\rangle_\omega$ of the Dirac field in that state. For notational convenience, we decompose $\langle \rho^\mathrm{split}\rangle_\omega$ into the operator $(T^\mathrm{split}_{00'})^{AB}$ acting on ${\mathcal{W}}_{\omega AB} - {\,}^\psi H_{1 AB}^{(+)}$ where $(T^\mathrm{split}_{00'})^{AB}$ is given by $$\begin{aligned} (T^\mathrm{split}_{00'})^{AB} = \frac{i}{2}\delta^{AB}\bigg(\mathbbm{1} \otimes e_0\cdot\nabla - e_0\cdot\nabla\otimes\mathbbm{1} \bigg)+ \Theta^{AB}\end{aligned}$$ and $\Theta^{AB}$ is a term which depends only on the spin-connection. The precise form of $\Theta^{AB}$ may be found in [@fewster; @verch] (eqn. (3.10) of that reference) but does not affect our discussion due to a useful result[^4] which shows it is identically zero on a timelike worldline under certain conditions which we shall now motivate: Let $\gamma:{\mathbb{R}}\mapsto {\mathcal{M}}$ be a timelike worldline with unit tangent vector $\dot\gamma$. Pick a point $x$ on $\gamma\subset{\mathcal{M}}$ and construct a local frame $\{ e^a_\alpha \}_{\alpha=0,1,2,3}$ subject to $e^a_0=\dot\gamma^a$ at $x$. As we shall only be concerned with averaging along a compact subset of $\gamma$ we shall fix a closed interval $I\subset\gamma({\mathbb{R}})$ such that $x\in I$. One may utilise Fermi-Walker transport to move $\{ e^a_\alpha \}_{\alpha=0,1,2,3}$ along $I$ keeping $e^a_0\|_\gamma=\dot\gamma^a$. The salient feature of Fermi-Walker transport is that it preserves angles, i.e. $\{ e^a_\alpha\}_{\alpha=0,1,2,3}$ remains an orthonormal family along $I$. Next, at each point $\gamma(s)$ along $I$ consider the convex normal neighbourhood ${\mathcal{U}}$ orthogonal to $\dot\gamma(s)$ and for each $y\in{\mathcal{U}}$ parallel transport $\{ e^a_\alpha \}_{\alpha=0,1,2,3}$ along the unique geodesic connecting $y$ to $\gamma(s)$. In this manner we ‘sweep out’ a tube $\tau_\gamma\subset {\mathcal{M}}$ in spacetime. Importantly, the local frame (throughout $\tau_\gamma)$ is a local section of $F{\mathcal{M}}$ which may be identified with a smooth section $E$ of $S{\mathcal{M}}$. The details of this identification may be found in §3 of [@fewster; @verch]. For our purposes it is sufficient to know that, as a consequence of this construction, the form of $(T^\mathrm{split}_{00'})^{AB}$ simplifies when restricted to the diagonal. The precise statement, quoted from [@dawson; @fewster] (lem.4) is: If $E$ is any local section of $S{\mathcal{M}}$ obtained in the above fashion from the curve $\gamma$, then $\Theta^{AB}\|_\gamma=0$. We shall assume that $E$ has been obtained in this way. Therefore, the finite contribution $\langle \rho^{\mathrm{fin}}\rangle_\omega$ to the energy density is given by: $$\langle \rho^{\mathrm{fin}}\rangle_\omega (\gamma(t)) = \frac{i}{2}\vartheta^\bigg(\big( \mathbbm{1}\otimes e_0\cdot\nabla -e_0\cdot\nabla \otimes \mathbbm{1} \big) \big({\mathrm{W}}_\omega - {\,}^\psi{\mathrm{H}}^{(+)}_1 \big)\bigg)(t,t)$$ where $\vartheta=\gamma\otimes\gamma$. Finally, it may be argued in analogy with [@fewster; @verch] that one may re-express $\langle \rho^{\mathrm{fin}}\rangle_\omega$ as $$\label{Had regularised energy} \langle \rho^{\mathrm{fin}}\rangle_\omega (\gamma(t)) = \frac{1}{2}\bigg( \mathbbm{1} \otimes D - D\otimes\mathbbm{1} \bigg)\vartheta^\big( {\mathrm{W}}_\omega-{\,}^\psi{\mathrm{H}}^{(+)}_1 \big)(t,t)$$ where $D$ is the distributional dual to $-i{\mathrm{d}}/{\mathrm{d}}t$. Microlocal analysis applied to quantum field theory {#microlocal review} =================================================== The Sobolev wave-front set {#microlocal review 1} -------------------------- Since the publication of Radzikowski’s equivalence theorem [@radzikowski] microlocal analysis, in particular Hörmander’s concept of wave-front set, has been successfully applied to quantum field theory[^5]. The proof of the most general QWEIs rely on microlocal analysis at various stages in their argument, e.g. [@dawson; @fewster; @fewster; @fewster; @pfenning]. A refined version of the usual (smooth) wave-front set, an exposition of which may be found in chapter VIII of [@hormander], has already been employed in the proof of an absolute quantum energy inequality for the Klein-Gordon field [@fewster; @smith]. We shall briefly review the necessary details of this refinement, known as the Sobolev wave-front set. For $s \in {\mathbb{R}}$, the Sobolev space $H^s({\mathbb{R}}^n)$ is the set of all tempered distributions $u$ such that $\widehat{u}$ is a measurable function and $${\parallel}u {\parallel}_{H^s({\mathbb{R}}^n)}^2=\int_{{\mathbb{R}}^n} {\mathrm{d}}^n\xi \, (1+\|\xi\|^2)^s \|\widehat{u}(\xi)\|^2 < \infty \, .$$ It is clear that ${\mathcal{S}}({\mathbb{R}}^n)\subset H^s({\mathbb{R}}^n)$ for each $s\in{\mathbb{R}}$. Moreover, one may show that ${\mathcal{S}}({\mathbb{R}}^n)$ is dense in $H^s({\mathbb{R}}^n)$, see chapter 1 §3 of [@taylor] for a brief argument. Sobolev space theory is usually introduced into the study of distributional solutions to partial differential equations by asking when such a solution is an honest function; this is the subject of the embedding theorems. We summarise the following useful properties of the $H^s$ spaces, the first of which is a relevant embedding theorem: \[Sobolev prop\] The Sobolev spaces $H^s({\mathbb{R}}^n)$ have the following properties: $$\begin{array}{cl} i) & \textrm{Let $k\in \{ 0\} \cup{\mathbb{N}}$ and $s\in{\mathbb{R}}$ satisfy $s>k+n/2$ then } H^s({\mathbb{R}}^n) \subset C^k({\mathbb{R}}^n) \\ & \text{ is a continuous embedding} \\ ii) & H^s({\mathbb{R}}^n) \subset H^{s^\prime}({\mathbb{R}}^n) \, \forall s\geq s^\prime; \\ iii) & \textrm{if } u \in H^s({\mathbb{R}}^n), f \in C^k({\mathbb{R}}^n) \textit{ and } D^\alpha f \in L^\infty({\mathbb{R}}^n) \, \forall \|\alpha\|\leq k, \textit{ where $D$ is a} \\ & \textit{partial derivative operator and $\alpha$ is a multi-index, then for all } \|s\|\leq k \\ & u \mapsto fu \textit{ is a bounded linear map of $H^s({\mathbb{R}}^n)$ into $H^s({\mathbb{R}}^n)$. In particular,} \\ & \textit{$H^s({\mathbb{R}}^n)$ is closed under multiplication by smooth functions.} \end{array}$$ The concept of the Sobolev wave-front set will give a concise way of saying what it means for a distribution to microlocally fail to be an element of a Sobolev space. For convenience, we adopt the notation that $\dot{T}^{\mathbb{R}}^n$ (similarly $\dot{T}{\mathbb{R}}^n$, $\dot{T}{\mathcal{M}}$ etc) is the bundle $T^{\mathbb{R}}^n$ ($T{\mathbb{R}}^n$, $T{\mathcal{M}}$, etc) with the zero section removed. A distribution $u \in {\mathcal{D}}^\prime ({\mathbb{R}}^n)$ is said to be microlocally $H^s$ at $(x,\xi)\in \dot{T}^{\mathbb{R}}^n$ if there exists an open cone $\Gamma \subset {\mathbb{R}}^n\setminus 0$ about $\xi$ and a smooth function $\varphi \in C^\infty_0 ({\mathbb{R}}^n)$, $\varphi(x)\not=0$, such that $$\int_\Gamma {\mathrm{d}}^n\zeta \, (1+\|\zeta\|^2)^s \|[\varphi u]^\wedge(\zeta)\|^2 < \infty \, .$$ The Sobolev wave-front set $WF^s(u)$ of a distribution $u \in {\mathcal{D}}^\prime ({\mathbb{R}}^n)$ is the complement, in $\dot{T}^{\mathbb{R}}^n$, of the set of all pairs $(x,\xi)$ at which $u$ is microlocally $H^s$. To define the Sobolev wave-front set of a distribution on a manifold one works locally. Let ${\mathcal{U}}$ be an open patch of a manifold ${\mathcal{M}}$ with associated coordinate map $\kappa:{\mathcal{U}}\mapsto {\mathbb{R}}^n$. If $(\kappa(x),\xi)\in WF^s(u\circ\kappa^{-1})$ then $(x,\kappa^{-1}_(\xi))\in WF^s(u)\subset \dot{T}^{\mathcal{M}}$. We shall occasionally use the notation $u\in H^s_{\mathrm{loc}}({\mathcal{M}})$ if $WF^s(u)=\emptyset$ for a distribution $u\in{\mathcal{D}}'({\mathcal{M}})$ and direct the reader to the remarks following definition 8.2.5 of [@hormander; @hyper] to justify this notation. The Sobolev wave-front set is a closed cone in $\dot{T}^{\mathcal{M}}$. Furthermore, we have the following properties[^6] of $WF^s$:\ ------ ------------------------------------------------------------------------------------------------- i) The smooth wave-front set is related to the Sobolev wave-front set via $WF(u)=\overline{\bigcup_{s\in{\mathbb{R}}} WF^s(u)}$. ii) If $\varphi \in C^\infty_0 ({\mathbb{R}}^n)$ does not vanish in a neighbourhood of $x$ then $(x,\xi)\in WF^s(u)$ if and only if $(x,\xi)\in WF^s(\varphi u)$. iii) $(x,\xi)\in WF^s(u)$ if and only if, for all $v\in H^s_{\mathrm{loc}}$, $(x,\xi)\in WF^s(u-v)$. iv) $WF^s(u+w)\subset WF^s(u)\cup WF^s(w)$. v) The nesting property: $WF^s(u)\subset WF^{s'}(u)$ $\forall s\leq s'$. ------ ------------------------------------------------------------------------------------------------- It is also possible to see explicitly in $WF^s$ what effect partial differential operators have on the singularities of distributions. For a general $m-$dimensional smooth manifold ${\mathcal{M}}$, let $P$ be a partial differential operator of order $r$, i.e. in local coordinates on ${\mathcal{M}}$ $$P = \sum_{\|\alpha\|\leq r} p_\alpha(x)(-i\partial_a)^{\alpha}$$ where $\alpha$ is a multi-index and $p_\alpha$ are smooth functions, then the principal symbol, $p_r(x,\xi)$, of $P$ is $$p_r(x,\xi) = \sum_{\|\alpha\|=r}p_\alpha(x)\xi^{\alpha}_a \, .$$ The characteristic set, ${\mathrm{Char \,}}P$, of a partial differential operator $P$ is the set of $(x,\xi)\in \dot{T}^{\mathcal{M}}$ such that the principal symbol vanishes. We may now quote corollaries 8.4.9-10 of [@hormander; @hyper] which show the effect differential operators have on the Sobolev wave-front set of a distribution: \[order\] Let ${\mathcal{M}}$ be a smooth manifold. For $u \in {\mathcal{D}}^\prime ({\mathcal{M}})$ and any linear partial differential operator $P$ of order $r$ with smooth coefficients then $WF^s(Pu)\subset WF^{s+r}(u)$ and $WF^{s+r}(u)\subset WF^{s}(Pu)\cup {\mathrm{Char \,}}P$. We close this section with the statement of Beal’s restriction theorem, which tells us under what circumstances a distribution may be restricted to a submanifold, and a result about the implications for the positivity of states under such a restriction. Such results are of interest to us as we need to understand how to restrict those distributions which make up the Hadamard series (and ones derived from it such as the point-split energy density) to timelike worldlines. Beal’s restriction theorem tells us that, for certain well behaved restrictions, the Sobolev grading on the wave-front set is reduced by an amount proportional to the codimension of the restriction. The result discusses the case of restricting a distribution on a $m$ dimensional manifold ${\mathcal{M}}$ to a smoothly embedded submanifold $\Sigma$ of dimension $n$, writing the embedding as $\iota : \Sigma \rightarrow {\mathcal{M}}$. The embedding function $\iota$ has associated conormal bundle $N^\Sigma$ given by $$N^ \Sigma = \{ (\iota(x),\xi)\in T^{\mathcal{M}}; \, x \in \Sigma, \, \iota^(\xi) = 0 \} \, .$$ We wish to formulate a statement of the restriction theorem for product manifolds[^7]. As usual we let $({\mathcal{M}},g)$ denote a smooth $m$ dimensional spacetime, $\Sigma\subset {\mathcal{M}}$ an $n\leq m$ dimensional submanifold embedded using $\iota : \Sigma \mapsto {\mathcal{M}}$. Then we define the map $\vartheta : \Sigma \times\Sigma \mapsto {\mathcal{M}}\times {\mathcal{M}}$ by $\vartheta = \iota\otimes\iota$, the pull back $\vartheta^$ may sometimes be referred to as a restriction map. \[beal\] Let $u \in {\mathcal{D}}^\prime ({\mathcal{M}}\times{\mathcal{M}})$ and $\vartheta$ be defined as above. If $\big( N^\Sigma\times N^\Sigma\big)\cap WF^s(u)=\emptyset$ for some $s>m-n$ then the restriction $\vartheta^u$ of $u$ to $\Sigma\times\Sigma$ is a well defined distribution in ${\mathcal{D}}^\prime(\Sigma\times\Sigma)$. Moreover, $$WF^{s-(m-n)}(\vartheta^* u)\subset \vartheta^WF^s(u)$$ where the set $\vartheta^WF^s(u)$ is defined to be $$\begin{aligned} \vartheta^WF^s (u) &=& \{ (t,\iota^(\xi);t^\prime,\iota^(\xi^\prime)) \in (T^\Sigma\times T^\Sigma) \mid \nonumber \\ &\,& \qquad (\iota(t),\xi;\iota(t^\prime),\xi^\prime) \in WF^s(u) \} \, .\end{aligned}$$ Finally, we state a result[^8] which asserts that the positivity of states is preserved under the restrictions carried out by Beal’s theorem. \[res pos\] Let ${\mathcal{M}}$ and $\Sigma$ be smooth manifolds each equipped with smooth positive densities, and suppose $\iota:\Sigma\mapsto{\mathcal{M}}$ is smooth. If $u\in{\mathcal{D}}'({\mathcal{M}}\times{\mathcal{M}})$ is positive in the sense of states and $WF(u)\cap (N^\Sigma\times N^\Sigma)=\emptyset$, then $\vartheta^u=u\circ(\iota\otimes\iota)\in{\mathcal{D}}'(\Sigma\times\Sigma)$ is also positive. A microlocal description of the Hadamard series {#microlocal results} ----------------------------------------------- Denote by $\mathcal{R} = \{ (x,\xi)\in \dot{T}^{\mathcal{M}}\, \mid \, g^{ab}(x)\xi_a\xi_b =0 \}$ the bundle of null covectors over ${\mathcal{M}}$. Since $({\mathcal{M}},g)$ is time orientable we may decompose $\mathcal{R}$ into two disjoint sets $\mathcal{R}^\pm$ defined by $\mathcal{R}^\pm = \{ (x,\xi) \in \mathcal{R} \, \mid \, \pm\xi \rhd 0 \}$ where by $\xi\rhd 0$ ($\xi\in T^_x{\mathcal{M}}$) we mean that $\xi_a$ is in the dual of the future light cone at $x$. We define the notation $(x,\xi)\sim (x',\xi')$ to mean that there exists a null curve $\gamma : [0,1]\mapsto {\mathcal{M}}$ such that $\gamma(0)=x$, $\gamma(1)=x^\prime$ and $\xi_a = \dot{\gamma}^b(0)g_{ab}(x)$, $\xi^\prime_a = \dot{\gamma}^b(1)g_{ab}(x^\prime)$. In the instance where $x=x'$, $(x,\xi)\sim (x,\xi')$ shall mean that $\xi=\xi'$ is null. Then, for convenience, define the set $$C = \{ (x,\xi;x',\xi') \in \mathcal{R} \times \mathcal{R} \mid (x,\xi) \sim (x',\xi') \} \, .$$ The set $C^{+-}$ is defined to be $$\label{C+-} C^{+-} = \{ (x,\xi;x^\prime,-\xi^\prime) \in C \mid \xi \rhd 0 \} \, .$$ An occasionally useful set will be $C^{-+}$ defined by $$C^{-+} = \{ (x,-\xi;x^\prime,\xi^\prime) \in C \mid \xi\rhd 0 \} \, .$$ Junker & Schrohe [@junker] have proven that the quantum Klein-Gordon field, whose two-point function we denote $\Lambda_\omega$, obeys the following condition for all Hadamard states $$WF^s(\Lambda_\omega) = \left\{ \begin{array}{cl} C^{+-} & s\geq -1/2 \\ \emptyset & s<-1/2 \end{array} \right. \, .$$ In [@fewster; @smith] an analysis of the Hadamard series $H^{(+)}_k$ was given and concluded that $$WF^{s+j+1}(\sigma^j\ln\sigma_+)\subset WF^s(1/\sigma_+) = \left\{ \begin{array}{cc} C^{+-} & s \geq -1/2 \\ \emptyset & s<-1/2 \end{array} \right. \,$$ for $j\in\{0\}\cup{\mathbb{N}}$ and where $1/\sigma_+$ and $\sigma^j\ln\sigma_+$ are the singular constituents of the Hadamard series $H^{(+)}_k$ (\[parametrix\]). Since it is known that the the coefficients $\Delta^\frac{1}{2}$ and $v_j$ appearing in the $H^{(+)}_k$ series are symmetric it follows from the simple symmetry argument $H^{(+)}_k(x,x')=H^{(-)}_k(x',x)$ modulo smooth functions that $$WF^{s+j+1}(\sigma^j\ln\sigma_\pm)\subset WF^s(1/\sigma_\pm) = \left\{ \begin{array}{cc} C^{\pm\mp} & s \geq -1/2 \\ \emptyset & s<-1/2 \end{array} \right. \, .$$ Therefore, lemma \[order\] implies that if $\omega$ is a Hadamard state for the Dirac field then within a regular domain $$WF^s({\mathcal{W}}_\omega) \subset \left\{ \begin{array}{cl} C^{+-} & s\geq 1/2 \\ \emptyset & s<1/2 \end{array}\right. \, .$$ A similar condition holds for ${\mathcal{W}}^\Gamma_\omega$ under the replacement ${\mathcal{W}}_\omega \mapsto {\mathcal{W}}^\Gamma_\omega$ and $C^{+-}\mapsto C^{-+}$. Moreover, one may use the relations (\[dirac hadamard relations 1\]-\[dirac hadamard relations 3\]) to conclude that within a regular domain $$\begin{aligned} WF^s({\mathcal{W}}_\omega - {\,}^\psi H^{(+)}_k) &\subset& \left\{ \begin{array}{cc} C^{+-} & s \geq k+3/2 \\ \emptyset & s < k+3/2 \end{array} \right. \, ,\\ WF^s({\mathcal{W}}_\omega^\Gamma - {\,}^\psi H^{(-)}_k) &\subset& \left\{ \begin{array}{cc} C^{-+} & s \geq k+3/2 \\ \emptyset & s < k+3/2 \end{array} \right. \, ,\end{aligned}$$ where $\omega$ is a Hadamard state. Consequently, if $\omega$ is a Hadamard state then (because ${\mathrm{W}}_\omega,{\mathrm{W}}^\Gamma_\omega$ are formed from linear combinations of ${\mathcal{W}}_\omega$ and ${\mathcal{W}}^\Gamma_\omega$) it follows that the Sobolev wave-front set conditions which apply to ${\mathcal{W}}_\omega$ and ${\mathcal{W}}^\Gamma_\omega$ also apply to ${\mathrm{W}}_\omega$ and ${\mathrm{W}}^\Gamma_\omega$ respectively. We encapsulate these findings in the following corollary: \[dirac grading\] Let $\omega$ be a Hadamard state for the Dirac field; then within a regular domain $$\begin{aligned} WF^s({\mathrm{W}}_\omega - {\,}^\psi {\mathrm{H}}^{(+)}_k) &\subset& \left\{ \begin{array}{cc} C^{+-} & s \geq k+3/2 \\ \emptyset & s < k+3/2 \end{array} \right. \, ,\\ WF^s({\mathrm{W}}_\omega^\Gamma - {\,}^\psi {\mathrm{H}}^{(-)}_k) &\subset& \left\{ \begin{array}{cc} C^{-+} & s \geq k+3/2 \\ \emptyset & s < k+3/2 \end{array} \right. \, .\end{aligned}$$ A Sobolev point-splitting result {#splitting section} ================================ We quote a result, taken from [@fewster; @verch], for smooth functions: \[smooth splitting\] If $f\in C^\infty_0({\mathbb{R}})$ and $u\in C^\infty_0({\mathbb{R}}\times{\mathbb{R}})$ then the following identity holds: $$\label{rhs} \int_{\mathbb{R}}{\mathrm{d}}t \, f^2(t)u(t,t) = \int_{{\mathbb{R}}\times{\mathbb{R}}} \frac{{\mathrm{d}}\xi{\mathrm{d}}\xi'}{(2\pi)^2}\widehat{f^2}(\xi-\xi')\hat{u}(-\xi,\xi') \, .$$ This result forms the basis of the analysis in [@fewster; @verch; @dawson; @fewster] where the authors relate $u$ to the energy density obtained from normal ordering, i.e. $$\label{normal ordered energy} \frac{1}{2}\bigg( \mathbbm{1} \otimes D - D\otimes\mathbbm{1} \bigg)\vartheta^\big( {\mathrm{W}}_\omega-{\mathrm{W}}_{\omega_0} \big)(t,t)$$ where $\omega_0$ is another Hadamard state of the Dirac field. Since the difference between any two two-point functions arising from Hadamard states is smooth the quantity (\[normal ordered energy\]) is readily identifiable with $u$ in the hypothesis of lemma \[smooth splitting\]. However, ${\mathrm{W}}_\omega - {\,}^\psi {\mathrm{H}}_1^{(+)}$ featuring in (\[Had regularised energy\]) is not smooth so we need to relax the hypothesis of lemma \[smooth splitting\] in order to proceed. In particular we wish to show that one has the conclusion of lemma \[smooth splitting\] under the weaker assumption that $u\in H^{s}({\mathbb{R}}\times{\mathbb{R}})\cap {\mathcal{E}}'({\mathbb{R}}\times{\mathbb{R}})$ for $s>1$. Let $u\in C^\infty_0({\mathbb{R}}\times{\mathbb{R}})$; then, applying the Hölder inequality we have $$\begin{aligned} \int_{\mathbb{R}}{\mathrm{d}}t \, \big\| f^2(t)u(t,t) \big\| &\leq& {\parallel}f^2 {\parallel}_{L^1({\mathbb{R}})} \, \sup_{t\in{\mathbb{R}}}\|u(t,t)\| \\ &\leq& {\parallel}f^2 {\parallel}_{L^1({\mathbb{R}})} \, {\parallel}u {\parallel}_{L^\infty({\mathbb{R}}\times{\mathbb{R}})} \end{aligned}$$ for all $u\in C^\infty_0({\mathbb{R}}\times{\mathbb{R}})$; we also remark that ${\parallel}\cdot{\parallel}_{L^\infty({\mathbb{R}}\times{\mathbb{R}})}$ is the natural norm on $C({\mathbb{R}}\times{\mathbb{R}})$. Since the embedding of $H^{s}({\mathbb{R}}\times{\mathbb{R}})$ into $C({\mathbb{R}}\times{\mathbb{R}})$ is continuous for $s>1$ there exists a constant $c>0$ such that ${\parallel}u {\parallel}_{L^\infty({\mathbb{R}}\times{\mathbb{R}})}\leq c {\parallel}u{\parallel}_{H^{s}({\mathbb{R}}\times{\mathbb{R}})}$ for all $u\in H^{s}({\mathbb{R}}\times{\mathbb{R}})$ and $s>1$. Hence, $$\label{BLT2} \int_{\mathbb{R}}{\mathrm{d}}t \, \|f^2(t)u(t,t)\| \leq c{\parallel}f^2 {\parallel}_{L^1({\mathbb{R}})} \, {\parallel}u {\parallel}_{H^{s}({\mathbb{R}}\times{\mathbb{R}})} \quad \forall u \in C^\infty_0({\mathbb{R}}\times{\mathbb{R}}) \quad s>1 \, .$$ Moreover, as $C^\infty_0({\mathbb{R}}\times{\mathbb{R}})$ is dense in $H^{s}({\mathbb{R}}\times{\mathbb{R}})$, the Bounded Linear Transform theorem implies that (\[BLT2\]) holds for all $u\in H^{s}({\mathbb{R}}\times{\mathbb{R}})$ for $s>1$. Equally, we may apply the Hölder inequality to the right-hand-side of (\[rhs\]) to obtain $$\bigg\| \int_{{\mathbb{R}}\times{\mathbb{R}}}\frac{{\mathrm{d}}\xi{\mathrm{d}}\xi'}{(2\pi)^2} \widehat{f^2}(\xi-\xi')\widehat{u}(-\xi,\xi') \bigg\| \leq {\parallel}F {\parallel}_{L^\infty({\mathbb{R}}\times{\mathbb{R}})} \, {\parallel}\widehat{u}{\parallel}_{L^1({\mathbb{R}}\times{\mathbb{R}})} \, ,$$ where $F(\xi,\xi')=\widehat{f^2}(\xi-\xi')$. A factor of $(1+\|\xi\|^2+\|\xi'\|^2)^{(1+\varepsilon)/2}(1+\|\xi\|^2+\|\xi'\|^2)^{-(1+\varepsilon)/2}$ is then introduced into the $L^1$ norm and the Cauchy-Schwarz inequality applied to obtain $$\begin{aligned} \label{BLT1} {\parallel}\widehat{u}{\parallel}_{L^1({\mathbb{R}}\times{\mathbb{R}})} &\leq& \sqrt{\frac{\pi}{\epsilon}} {\parallel}u {\parallel}_{H^{1+\varepsilon}({\mathbb{R}}\times{\mathbb{R}})} \, \end{aligned}$$ where we have written $s>1$ as $s=1+\epsilon$. Again the Bounded Linear Transform theorem implies that (\[BLT1\]) holds for all $u\in H^{1+\epsilon}({\mathbb{R}}\times{\mathbb{R}})$ where the $\,\widehat{\,}\,$ now refers to the continuous extension of the Fourier transform to $H^{1+\epsilon}({\mathbb{R}}\times{\mathbb{R}})$ (although this must agree with the usual Fourier transform on $L^2({\mathbb{R}}\times{\mathbb{R}})$ or ${\mathcal{S}}'({\mathbb{R}}\times{\mathbb{R}})$). Therefore, since (\[rhs\]) holds for a dense subset of $H^{1+\epsilon}({\mathbb{R}}\times{\mathbb{R}})$ and may be extended continuously onto the whole of the space, we have proven: \[dirac point split 1\] Let $f\in C^\infty_0({\mathbb{R}})$ and $u\in H^s({\mathbb{R}}\times{\mathbb{R}})\cap{\mathcal{E}}'({\mathbb{R}}\times{\mathbb{R}})$, $s >1$, then the following identity holds: $$\int_{\mathbb{R}}{\mathrm{d}}t \, f^2(t) u(t,t) = \int_{{\mathbb{R}}\times{\mathbb{R}}} \frac{{\mathrm{d}}\xi{\mathrm{d}}\xi'}{(2\pi)^2} \widehat{f^2}(\xi-\xi')\hat{u}(-\xi,\xi') \, .$$ It is clear under the identification $u(t,t')=\langle \rho^{\mathrm{fin}}\rangle_\omega(\gamma(t))$ (cf. (\[Had regularised energy\])) that $$\begin{aligned} &\,& \int_{\mathbb{R}}{\mathrm{d}}t \, f^2(t)\langle \rho^{\mathrm{fin}}\rangle_\omega (\gamma(t)) {\nonumber \\}&\,& \qquad = \frac{1}{2}\int_{{\mathbb{R}}\times{\mathbb{R}}}\frac{{\mathrm{d}}\xi {\mathrm{d}}\xi'}{(2\pi)^2} (\xi+\xi')\widehat{f^2}(\xi-\xi') \big[\vartheta^( {\mathrm{W}}_\omega - {\,}^\psi {\mathrm{H}}^{(+)}_1 )\big]^\wedge(-\xi,\xi') \label{dirac int} \, .\end{aligned}$$ We now prove a result (similar to lemma 5 of [@dawson; @fewster] in all but one detail of the proof[^9]) which will enable us to relate the right hand side of (\[dirac int\]) to an integral over the diagonal: \[dirac point split 2\] If $f\in C^\infty_0({\mathbb{R}})$ is real valued and $u\in H^s({\mathbb{R}}\times{\mathbb{R}})$, $s>2$, is compactly supported then $$\begin{aligned} \int_{{\mathbb{R}}\times{\mathbb{R}}}\frac{{\mathrm{d}}\xi\,{\mathrm{d}}\xi'}{(2\pi)^2} \, (\xi+\xi')\widehat{f^2}(\xi-\xi')\hat{u}(-\xi,\xi') = \frac{1}{\pi}\int_{\mathbb{R}}{\mathrm{d}}\xi \, \xi \, u(\overline{f^\xi},f^\xi)\end{aligned}$$ where $f^\xi(t)=e^{i\xi t}f(t)$. Lemma 6.1 of [@fewster; @verch] states that $$(\xi+\xi')\widehat{f^2}(\xi-\xi')=\frac{1}{\pi}\int_{\mathbb{R}}{\mathrm{d}}\zeta\, \zeta\hat{f}(\xi-\zeta)\overline{\hat{f}(\xi'-\zeta)}$$ and therefore $$\begin{aligned} &\,& \int_{{\mathbb{R}}\times{\mathbb{R}}}\frac{{\mathrm{d}}\xi\,{\mathrm{d}}\xi'}{(2\pi)^2}(\xi+\xi')\widehat{f^2}(\xi-\xi')\hat{u}(-\xi,\xi') {\nonumber \\}&\,& \quad = \frac{1}{\pi}\int_{{\mathbb{R}}\times{\mathbb{R}}}\frac{{\mathrm{d}}\xi\,{\mathrm{d}}\xi'}{(2\pi)^2}\int_{\mathbb{R}}{\mathrm{d}}\zeta \, \zeta \hat{f}(\xi-\zeta)\overline{\hat{f}(\xi'-\zeta)}\hat{u}(-\xi,\xi') \, . \label{thingy}\end{aligned}$$ We also note, by a simple application of the convolution theorem, that $$\begin{aligned} u(\overline{f^\xi},f^{\xi}) &=& [(f\otimes f)u](e^{-i\xi\cdot},e^{i\xi\cdot}) \\ &=& \int_{{\mathbb{R}}\times{\mathbb{R}}}\frac{{\mathrm{d}}\zeta\,{\mathrm{d}}\zeta'}{(2\pi)^2}\hat{f}(-\xi-\zeta)\hat{f}(\xi-\zeta')\hat{u}(\zeta,\zeta') \\ &=& \int_{{\mathbb{R}}\times{\mathbb{R}}}\frac{{\mathrm{d}}\zeta \, {\mathrm{d}}\zeta'}{(2\pi)^2}\hat{f}(-\xi+\zeta)\hat{f}(\xi-\zeta')\hat{u}(-\zeta,\zeta') \\ &=& \int_{{\mathbb{R}}\times{\mathbb{R}}}\frac{{\mathrm{d}}\zeta\,{\mathrm{d}}\zeta'}{(2\pi)^2}\,\hat{f}(\zeta-\xi)\overline{\hat{f}(\zeta'-\xi)} \, \hat{u}(-\zeta,\zeta')\end{aligned}$$ and, therefore, that the statement of the theorem will be established if the integrals in (\[thingy\]) can be reordered. If we estimate $\|\hat{f}(x)\|\leq c/(1+\|x\|^2)$ then, by the arithmetic-geometric mean inequality $$\begin{aligned} &\,& \int_{\mathbb{R}}{\mathrm{d}}\zeta \, \big\| \zeta \hat{f}(\xi-\zeta)\overline{\hat{f}(\xi'-\zeta)} \big\| {\nonumber \\}&\,& \qquad \leq \frac{c^2}{2}\int_{\mathbb{R}}{\mathrm{d}}\zeta \,\bigg( \frac{\|\zeta\|}{(1+\|\xi-\zeta\|^2)^2}+\frac{\|\zeta\|}{(1+\|\xi'-\zeta\|^2)^2} \bigg)\\ &\,& \qquad \leq \frac{c^2}{2}(2+\|\xi\arctan\xi\|+\|\xi'\arctan\xi'\|) \\ &\,& \qquad \leq \frac{c^2\pi}{4}(2+\|\xi\|+\|\xi'\|) \, .\end{aligned}$$ Then $$\begin{aligned} &\,& \frac{1}{\pi}\int_{{\mathbb{R}}\times{\mathbb{R}}\times{\mathbb{R}}} \frac{{\mathrm{d}}\xi\,{\mathrm{d}}\xi'}{(2\pi)^2} \, {\mathrm{d}}\zeta \, \bigg\| \zeta \, \hat{f}(\xi-\zeta)\overline{\hat{f}(\xi'-\zeta)}\hat{u}(-\xi,\xi') \bigg\| {\nonumber \\}&\,& \quad \leq \frac{c^2}{4}\int_{{\mathbb{R}}\times{\mathbb{R}}} \frac{{\mathrm{d}}\xi\,{\mathrm{d}}\xi'}{(2\pi)^2} (2+\|\xi\|+\|\xi'\|) \|\hat{u}(-\xi,\xi')\| \\ &\,& \quad \leq \frac{c^2}{16\pi^2}{\parallel}p_s {\parallel}_{L^2({\mathbb{R}}\times{\mathbb{R}})} \, {\parallel}u {\parallel}_{H^s({\mathbb{R}}\times{\mathbb{R}})} \label{distinction}\end{aligned}$$ where we have used the Cauchy-Schwarz inequality and written $$p_s(\xi,\xi')=\frac{2+\|\xi\|+\|\xi'\|}{(1+\|\xi\|^2+\|\xi'\|^2)^{s/2}} \, .$$ The $L^1$ norm of $\zeta\hat{f}(\xi-\zeta)\overline{\hat{f}(\xi'-\zeta)}\hat{u}(-\xi,\xi')$ will be finite if $u\in H^s({\mathbb{R}}\times{\mathbb{R}})$ and $s>2$. Under these conditions, Fubini’s theorem implies that the integrals can be reordered to obtain the desired result. A worldline absolute quantum weak energy inequality {#punch line} =================================================== We are now in a position to state our result concerning the Dirac field. \[main result\] Let $\omega$ be a Hadamard state for the Dirac field, $\gamma$ be a timelike worldline and $f\in C^\infty_0({\mathbb{R}})$ be real valued; then $$\int_{\mathbb{R}}\, {\mathrm{d}}t \, f^2(t) \langle \rho^{\mathrm{fin}}\rangle_\omega (\gamma(t)) \geq - B$$ where $$\begin{aligned} B &=& \int_{{\mathbb{R}}^+}\frac{{\mathrm{d}}\xi}{2\pi}\xi \bigg[ f\otimes f \, \vartheta^{\,}^\psi{\mathrm{H}}^{(+)}_4 \bigg]^\wedge(-\xi,\xi) {\nonumber \\}&\,& \quad - \int_{{\mathbb{R}}^-}\frac{{\mathrm{d}}\xi}{2\pi} \xi \bigg[ f\otimes f \,\vartheta^\big(i\mathrm{S}_{\mathrm{sp}}-{\,}^\psi{\mathrm{H}}^{(+)}_4 \big) \bigg]^\wedge(-\xi,\xi)\end{aligned}$$ and $\vartheta =\gamma\otimes\gamma$. Corollary \[dirac grading\] implies that ${\mathrm{W}}_\omega-{\,}^\psi {\mathrm{H}}_4^{(+)} \in H^{5+\varepsilon}_{\mathrm{loc}}(X)$ from which it follows that $$\label{dirac suff reg} f\otimes f\vartheta^\bigg(\big(\mathbbm{1}\otimes ie_0\cdot\nabla -ie_o\cdot\nabla \otimes \mathbbm{1} \big) \big( {\mathrm{W}}_\omega-{\,}^\psi {\mathrm{H}}^{(+)}_4 \big)\bigg) \in H^{1+\varepsilon}_{\mathrm{loc}}({\mathbb{R}}\times{\mathbb{R}}) \, ,$$ where we have lost one Sobolev order as a result of differentiation and a further three from the restriction to ${\mathbb{R}}\times{\mathbb{R}}$. Lemma \[dirac point split 1\], and the remarks following the proof, enable us to write $$\begin{aligned} &\,& \int_{\mathbb{R}}{\mathrm{d}}t \, f^2(t)\langle \rho^{\mathrm{fin}}\rangle_\omega (\gamma(t)) {\nonumber \\}&\,& \qquad = \frac{1}{2}\int_{{\mathbb{R}}\times{\mathbb{R}}}\frac{{\mathrm{d}}\xi{\mathrm{d}}\xi'}{(2\pi)^2}(\xi+\xi')\widehat{f^2}(\xi-\xi')\big[ \vartheta^({\mathrm{W}}_\omega - {\,}^\psi{\mathrm{H}}^{(+)}_4) \big]^\wedge (-\xi,\xi') \, .\end{aligned}$$ and as $\vartheta^({\mathrm{W}}_\omega-{\,}^\psi{\mathrm{H}}^{(+)}_4)\in H^{2+\varepsilon}_{\mathrm{loc}}({\mathbb{R}}\times{\mathbb{R}})$ we can employ lemma \[dirac point split 2\] to obtain: $$\int_{\mathbb{R}}{\mathrm{d}}t \, f^2(t)\langle \rho^{\mathrm{fin}}\rangle_\omega(t) = \int_{\mathbb{R}}\frac{{\mathrm{d}}\xi}{2\pi}\xi \vartheta^({\mathrm{W}}_\omega-{\,}^\psi{\mathrm{H}}^{(+)}_4)(\overline{f^\xi},f^\xi) \, .$$ We decompose the integral into its positive and negative frequency components and appeal to the anticommutation relation (in scalar form) ${\mathrm{W}}_\omega+{\mathrm{W}}_\omega^\Gamma = i\mathrm{S}_{\mathrm{sp}}$ to write $$\begin{aligned} \int_{\mathbb{R}}{\mathrm{d}}t \, f^2(t) \langle \rho^{\mathrm{fin}}\rangle_\omega(t) &=& \int_{{\mathbb{R}}^+}\frac{{\mathrm{d}}\xi}{2\pi}\xi\vartheta^\big({\mathrm{W}}_\omega-{\,}^\psi{\mathrm{H}}^{(+)}_4\big)(\overline{f^\xi},f^\xi) {\nonumber \\}&\,& \qquad + \int_{{\mathbb{R}}^-}\frac{{\mathrm{d}}\xi}{2\pi}\xi\vartheta^\big( i\mathrm{S}_{\mathrm{sp}}-{\mathrm{W}}_\omega^\Gamma - {\,}^\psi{\mathrm{H}}^{(+)}_4 \big)(\overline{f^\xi},f^\xi) \, .\end{aligned}$$ Recall that ${\mathrm{W}}_\omega$ and ${\mathrm{W}}^\Gamma_\omega$ are distributions of positive type, hence $$\begin{aligned} \int_{\mathbb{R}}{\mathrm{d}}t \, f^2(t) \langle \rho^{\mathrm{fin}}\rangle_\omega(t) &\geq& -\int_{{\mathbb{R}}^+}\frac{{\mathrm{d}}\xi}{2\pi}\xi\vartheta^{\,}^\psi{\mathrm{H}}^{(+)}_4(\overline{f^\xi},f^\xi) {\nonumber \\}&\,& \qquad + \int_{{\mathbb{R}}^-}\frac{{\mathrm{d}}\xi}{2\pi}\xi\vartheta^\big( i\mathrm{S}_{\mathrm{sp}}- {\,}^\psi{\mathrm{H}}^{(+)}_4 \big)(\overline{f^\xi},f^\xi) \, \end{aligned}$$ and it remains to show that this lower bound is finite. We make the replacements of ${\,}^\psi{\mathrm{H}}^{(+)}_4={\mathrm{W}}_{\omega_0}-F$ and $i\mathrm{S}_{\mathrm{sp}}-{\,}^\psi{\mathrm{H}}^{(-)}_4={\mathrm{W}}^\Gamma_{\omega_0}-G$ where $F,G\in C^4(X)$ and ${\mathrm{W}}_{\omega_0},{\mathrm{W}}^\Gamma_{\omega_0}$ arise from some arbitrary Hadamard state $\omega_0$. We remark that this replacement is a technical device only which we introduce to prove finiteness: The bound is still independent of any state. Hence, we have $$\begin{aligned} B &=& \int_{{\mathbb{R}}^+}\frac{{\mathrm{d}}\xi}{2\pi} \xi \bigg[ f\otimes f \, \vartheta^{\mathrm{W}}_{\omega_0} \bigg]^\wedge(-\xi,\xi) -\int_{{\mathbb{R}}^-}\frac{{\mathrm{d}}\xi}{2\pi}\xi \bigg[ f\otimes f \, \vartheta^{\mathrm{W}}^\Gamma_{\omega_0} \bigg]^\wedge(-\xi,\xi) {\nonumber \\}&\,& \quad -\int_{{\mathbb{R}}^+}\frac{{\mathrm{d}}\xi}{2\pi}\xi\bigg[ f\otimes f \vartheta^F \bigg]^\wedge(-\xi,\xi) + \int_{{\mathbb{R}}^-}\frac{{\mathrm{d}}\xi}{2\pi}\xi\bigg[ f\otimes f \vartheta^G \bigg]^\wedge(-\xi,\xi) \, .\end{aligned}$$ The finiteness of the ${\mathrm{W}}_{\omega_0}$ and ${\mathrm{W}}^\Gamma_{\omega_0}$ pieces is proven by the wave-front set conditions, $WF({\mathrm{W}}_{\omega_0})\subset C^{+-}$ and $WF({\mathrm{W}}^\Gamma_{\omega_0})\subset C^{-+}$, which imply[^10] that $\vartheta^{\mathrm{W}}_{\omega_0}$ and $\vartheta^{\mathrm{W}}_{\omega_0}^\Gamma$ are rapidly decaying in the directions they are being integrated in. Finally, one may use the following estimates $$\begin{aligned} [f\otimes f \, \vartheta^F ]^\wedge(\xi,\xi') &\leq& \frac{c}{(1+\|\xi\|^2+\|\xi'\|^2)^2} \, , \\ \left[f\otimes f \, \vartheta^G \right]^\wedge(\xi,\xi') &\leq& \frac{c'}{(1+\|\xi\|^2+\|\xi'\|^2)^2} \, .\end{aligned}$$ Hence, our bound is finite. Theorem \[main result\] enables us to finally formulate our absolute QWEI for the Dirac field. Wald’s uniqueness theorem implies that the regularised energy density $\langle \rho^{\mathrm{fin}}\rangle_\omega$ we have computed is equal to the renormalised energy density $\langle \rho^\textrm{ren}\rangle_\omega$ up to the addition of a local curvature term $C$. Hence, our result reads: $$\int_{\mathbb{R}}{\mathrm{d}}t \, f^2(t) \, \langle \rho^\textrm{ren}\rangle_\omega (t) \geq - \mathcal{B}$$ where $\mathcal{B}$ is given by $$\begin{aligned} \mathcal{B} &=& \int_{{\mathbb{R}}^+}\frac{{\mathrm{d}}\xi}{2\pi}\xi \bigg[ f\otimes f \, \vartheta^{\,}^\psi{\mathrm{H}}^{(+)}_4 \bigg]^\wedge(-\xi,\xi) {\nonumber \\}&\,& \quad - \int_{{\mathbb{R}}^-}\frac{{\mathrm{d}}\xi}{2\pi} \xi \bigg[ f\otimes f \,\vartheta^\big(i\mathrm{S}_{\mathrm{sp}}-{\,}^\psi{\mathrm{H}}^{(+)}_4 \big) \bigg]^\wedge(-\xi,\xi) {\nonumber \\}&\,& \qquad - \int_{\mathbb{R}}\, {\mathrm{d}}t \, f^2(t)C(t) \end{aligned}$$ As reported in [@fewster; @smith], where a more complete discussion of the renormalisation of the stress tensor of a quantum field may be found, the view may be held that the value of this curvature term (alongside the mass and curvature coupling) is an essential detail in the specification of the theory and that $C$ should be, at least in principle, measurable. Alternatively, one may hold the view that this unavoidable ambiguity is a manifestation of a breakdown of the semi-classical theory and that a more complete theory of quantum gravity is needed. Conclusion ========== We have succeeded in proving a new absolute QWEI for the Dirac field under general circumstances. By exploiting a Sobolev graded refinement of Hörmander’s wave-front set we have been able to modify the proof Fewster and Dawson [@dawson; @fewster] give for their difference QWEI and remove any reference to a state in the bound. Moreover, it is straightforward to use the techniques of [@fewster; @smith] to obtain additional $WF^s$ information of the constituents of the Dirac Hadamard series ${\,}^\psi H^{(\pm)}_k$. Acknowledgements The author would like to thank C.J. Fewster for his guidance over the course of this research. Additional thanks go to J.A. Sanders, L.W. Osterbrink, S.P. Dawson and P. Watts for their helpful comments on the manuscript. [99]{} Alcubierre M., 1994, The warp drive: hyper-fast travel within general relativity, Class. Quantum Grav. 11, L73-L77 Dawson S.P., 2006, A quantum weak energy inequality for the Dirac field in two-dimensional flat spacetime, Class. Quantum Grav. 23, 287-293 Dawson S.P. & Fewster C.J., 2006, An explicit quantum weak energy inequality for Dirac fields in curved spacetimes, Class. Quantum Grav. 23, 6659-6681 Dimock J., 1982, Dirac quantum fields on a manifold, Trans. Am. Math. Soc. Volume 232 No. 1 133-147 Epstein H., Jaffe A. & Glaser V., 1965, Nonpositivity in the energy density in quantised field theories, Nuovo Cimento 36, 1016-1022 Fewster C.J., 2000, A general worldline quantum inequality, Class. Quantum Grav. 17, 1897-1911 Fewster C.J. & Verch R., 2002, A quantum weak energy inequality for Dirac fields in curved spacetime, Commun. Math. Phys. 225, 331-359 Fewster C.J. & Mistry B., 2003, Quantum weak energy inequalities for the Dirac field in flat spacetime, Phys. Rev. D 68, 105010 Fewster C.J. & Pfenning M.J., 2003, A weak quantum energy inequality for spin-one fields in curved spacetime, J. Math. Phys. 44, 4480-4513 Fewster C.J., 2005, Energy inequalities in quantum field theory, in [XIVth International Congress on Mathematical Physics]{}, ed. J.C. Zambrini (World Scientific, Singapore, 2005). See [math-ph/0501073]{} for an expanded and updated version. Fewster C.J., 2005, Quantum energy inequalities and stability conditions in quantum field theory, in [Rigorous Quantum Field Theory: A Festschrift for Jacques Bros]{}, A. Boutet de Monvel, D. Buchholz, D. Iagolnitzer, U. Moschella (Eds.) Progress in Mathematics, Vol. 251. (Birkhäuser, Boston, 2006), [math-ph/0502002]{}. Fewster C.J. & Smith C.J., 2007, Absolute quantum energy inequalities in curved spacetime, [gr-qc/0702056]{} Ford L.H., 1978, Quantum coherence effects and the second law of thermodynamics, Proc. R. Soc. Lond. A. 364, 227-236 Hollands S., 2000, DPhil. thesis Aspects of quantum field theory in curved spacetime, Department of Mathematics, University of York, United Kingdom Hörmander L., 1989, The analysis of linear partial differential operators I second edition, Springer-Verlag, New York Hörmander L., 1996, Lectures on nonlinear hyperbolic differential equations, Springer, New York Junker W. & Schrohe E., 2002, Adiabatic vacuum states on general spacetime manifolds: Definition, construction, and physical properties, Annales Poincaré Phys. Theor. 3, 1113-1182 Kay B.S. & Wald R.M., 1991, Theorems on the uniqueness and thermal properties of stationary, nonsingular, quasifree states on spacetimes with a bifurcate Killing horizon, Phys. Rep. 207, 49-136 Köhler M., 1995, PhD. dissertation The stress energy tensor of a locally supersymmetric quantum field on a curved spacetime, [gr-qc/9505014]{} Kratzert K., 2000, Singularity structure of the two point function of the free Dirac field on globally hyperbolic spacetime, [math-ph/0003015]{} Nakahara M., 1989, Geometry, topology and physics, Institute of Physics publishing Radzikowski M., 1996, Micro-local approach to the Hadamard condition in quantum field theory on curved space-time, Commun. Math. Phys. 179, 529-553 Sahlmann H. & Verch R., 2001, Microlocal spectral condition and Hadamard form for vector-valued quantum fields in curved spacetime, Rev. Math. Phys. 13, 1203-1246 Synge J., 1960, Relativity: The general theory, North Holland, Amsterdam Taylor M.E., 1981, Pseudodifferential operators, Princeton University Press, New Jersey Vollick D.N., 2000, Quantum inequalities in curved two dimensional spacetimes, Phys. Rev. D 61, 084022 Wald R.M., 1978, Trace anomaly of a conformally invariant quantum field in curved spacetime, Phys. Rev. D 17, 1477-1484 Wald R.M., 2006, The history and present status of quantum field theory in curved spacetime, [gr-qc/0608018]{} [^1]: Electronic address: [[email protected]]{} [^2]: A more general, and rigorous, definition of a quantum energy inequality is given in [@fewster; @paris]. [^3]: The Hadamard recursions relations for the scalar field can be found in [@fewster; @smith]. [^4]: Lemma 4 of [@dawson; @fewster]. [^5]: For a readable account of the general significance of microlocal analysis in quantum field theory the reader is directed to [@wald; @review]. [^6]: Taken from the remarks following definition B.1 of [@junker]. [^7]: The following result is adapted from lemma 11.6.1 of [@hormander; @hyper] which is a refinement of the standard restriction theorem which may be found presented as theorem 8.2.4 of [@hormander]. We have used the notation of prop. B7 of [@junker]. [^8]: Theorem 2.2 of [@fewster]. [^9]: The distinct step is contained within line (\[distinction\]). [^10]: For the full details of this step the reader is directed to §3.2 of [@dawson; @fewster].	{ "pile_set_name": "ArXiv" }
--- abstract: 'We review applicability of QCD factorization theorem to multiple scattering in deeply inelastic lepton-nucleus scattering. We show why $A^{1/3}$-type nuclear enhancement can be calculated consistently in perturbative QCD. We derive the transverse momentum broadening of the leading pions in deeply inelastic lepton-nucleus collisions. We argue that the measurement of such transverse momentum broadening can provide direct information on the multiple-parton correlation functions inside a nucleus. We also estimate the numerical values of the broadening at different values of $x_B$ and $Q^2$.' address: \| $^1$Department of Physics and Astronomy, University of Kentucky,\ Lexington, Kentucky 40506, USA\ $^2$Department of Physics and Astronomy, Iowa State University\ Ames, Iowa 50011, USA\ $^3$Physics Department, Brookhaven National Laboratory\ Upton, New York 11973-5000, USA author: - 'Xiaofeng Guo$^1$ and Jianwei Qiu$^{2,3}$' date: 'November 2, 1999' title: 'Probing Quark-Gluon Correlation Functions' --- Introduction ============ Multiparton correlation functions inside a large nucleus are extremely important and useful for understanding nuclear dependence in relativistic heavy ion collisions. When a quark or a gluon passes through the nuclear matter, it loses energy via radiation, and it picks up extra transverse momentum because of multiple scattering [@Mueller; @GW; @AHMetal; @BBL; @LQS1]. The multiple scattering is directly associated with multiparton correlation functions [@LQS1]. Inside a large nucleus, multiple scattering can take place within one nucleon or between different nucleons. At high energy, a single hard scattering is always localized within one nucleon, and it cannot generate any large dependence on the nuclear size. Similarly, multiple scattering within one nucleon cannot provide much dependence on the nuclear size either. Therefore, the large dependence on the nuclear size is an unique signal of multiple scattering between nucleons. Measurements of such anomalous dependence on the nuclear size for any physical observable will provide direct information on multiparton correlation functions in a nucleus. However, there are many multiparton correlation functions in QCD [@LQS1]. It is therefore important to identify physical observables which depend only on a small number of multiparton correlation functions. Otherwise, data will not be able to separate contributions from different correlation functions. It was pointed out recently in Ref. [@Guo2] that the $A^{1/3}$-type enhancement of the transverse momentum broadening of Drell-Yan pairs and the jet broadening in lepton-nucleus deeply inelastic scattering (DIS) are good observables, because at the leading order of $\alpha_s$, these two observables depend only on one type of multiparton correlation functions. It represents the correlations between hard quarks and soft-gluons, and has the following operator definition [@LQS1; @Guo2], $$\begin{aligned} T_{qF}^{A}(x,\mu^2)&=& \int \frac{dy^-}{2\pi} e^{ix p^+ y^-} \frac{dy_1^- dy_2^-}{2\pi} \theta(y_1^--y^-)\theta(y_2^-) \nonumber \\ &\times & \langle P_A\| \bar{\psi}_q(0)\, \frac{\gamma^+}{2}\, F^+_{\ \alpha}(y_2^-) \, F^{+\alpha}(y_1^-)\, \psi_q(y^-)\, \|P_A\rangle \, , \label{T4qA}\end{aligned}$$ with $A$ the atomic weight, and $q$ is the quark flavor and $F$ is the gluon field strength. In Eq. (\[T4qA\]), $x$ is the effective momentum fraction and $\mu$ represents the scale-dependence of the correlation functions. According to QCD factorization theorem [@CSS_fac; @QS_fac], all information on the identity of the target is factorized into the matrix elements. The $A^{1/3}$-type enhancement must be a property of the relevant nuclear matrix elements, like the one in Eq. (\[T4qA\]). Experimentally, the $A^{1/3}$-type nuclear enhancement have been observed in the transverse momentum broadening of Drell-Yan pairs in hadron-nucleus collisions [@DY_PT]. Theoretically, Eq. (\[T4qA\]) shows how such enhancement can occur, through integrals over the relative positions of the fields, $y_i$, that appear in the expectation values. For example, enhancement can occur when the two quark fields are close together and the two gluon fields are close together, and they pair off into color singlets that are separated by no more than a nucleon diameter in the $y_i$ integrals; but the two pairs are separated by a distance that varies up to the nuclear size. In this manner, aside from an overall factor of $A$ which reflects growth with the nuclear volume, we anticipate an $A^{1/3}$ nuclear enhancement [@LQS2]. Direct measurement of the multiparton correlation functions is of great importance for testing the generalized QCD factorization theorem [@QS_fac], which allows us to apply QCD perturbation theory for studying the collisions involving nuclei. It was observed in Ref. [@Guo2] that measuring the jet transverse momentum broadening in DIS, $\Delta\langle p_T^2\rangle$, provides a [direct]{} measurement of the quark-gluon correlation functions, $T^A_{qF}(x_B,Q^2)$, inside a nucleus. At the lowest order, the jet transverse momentum broadening in DIS can be expressed as [@Guo2] $$\begin{aligned} \Delta\langle p_T^2\rangle &\equiv & \langle p_T^2\rangle^{eA} - \langle p_T^2\rangle^{eN} \nonumber \\ &\approx & \frac{4\pi^2\alpha_s(Q^2)}{3}\, \frac{\sum_q e_q^2\,T_{qF}^{A}(x_B,Q^2)} {\sum_q e_q^2\,q^{A}(x_B,Q^2)}\, , \label{jetdis}\end{aligned}$$ where $p_T$ is the transverse component of the jet momentum in the photon-nucleus center of mass frame in DIS. In Eq. (\[jetdis\]), $x_B$ is the Bjorken variable and $Q^2=-q^2$ with $q^2$ is the invariant mass of the virtual photon in DIS. However, other than a future HERA with a nuclear beam, existing fixed target facilities can not provide good measurements on jets in lepton-nucleus DIS. It is the purpose of this paper to show that by measuring the leading pions in DIS and their transverse momentum broadening, fixed target facilities, such as CEBAF [@Brodsky], can provide good information on the quark-gluon correlation functions defined in Eq. (\[T4qA\]). This paper is organized as follows. In the next section, we derive the transverse momentum broadening of the leading pions in lepton-nucleus DIS. We demonstrate analytically the direct correspondence between the transverse momentum broadening of the leading pions and multiparton correlation functions. In Sec. III, we estimate the numerical values of such broadening at different values of $x_B$ and $Q^2$, by using the model of the quark-gluon correlation functions introduced in Ref. [@LQS2]. We also derive explicit relations of the transverse momentum broadening for $\pi^{\pm}$ and $\pi^0$, and discuss the $x_B$ and $Q^2$ dependence of the correlation functions. Transverse momentum broadening of pions in DIS ============================================== Consider leading pion production in the deeply inelastic lepton-nucleus scattering, $$e(k_1) + A(p) \longrightarrow e(k_2) + \pi(\ell) +X \ , \label{process}$$ where $k_1$ and $k_2$ are the four momenta of the incoming and the outgoing leptons respectively, $p$ is the momentum per nucleon for the nucleus with the atomic number $A$, and $\ell$ is the observed pion momentum. In order to extract useful information on multiple scattering from the pion production, it is natural to think that the $A$-dependence of the differential cross section $d\sigma_{eA\rightarrow e\pi X} /dx_BdQ^2d\ell_T^2$ is the physical observable to study, because the transverse momentum spectrum is more sensitive to multiple scattering [@LQS1]. But, other than a future HERA with a nuclear beam, existing fixed target facilities can not produce good data on the production of leading pions at [large]{} transverse momenta in DIS. On the other hand, at the small transverse momentum $\ell_T^2$, the differential cross section $d\sigma/dx_BdQ^2d\ell_T^2$ is proportional to $1/\ell_T^2$, and is not perturbatively stable. A nontrivial all order resummation of the large logarithms $\log(Q^2/\ell_T^2)$ is necessary in order to get a reliable prediction of the transverse momentum spectrum at small $\ell_T^2$ [@Guo1]. However, the inclusive moments of the transverse momentum spectrum $\int d\ell_T^2 (\ell_T^2)^n\, d\sigma_{eA\rightarrow e\pi X}/dx_BdQ^2d\ell_T^2$ with $n\ge 1$ are perturbatively stable, and also enhance the information in the region of large $\ell_T^2$, where the multiple scattering is most relevant. Therefore, the $A$-dependence for the moments of the transverse momentum spectrum of the leading pions is a good observable for extracting information on multiple scattering. We define the first moment of the transverse momentum spectrum - the averaged pion transverse momentum squared, as $$\langle \ell_{T}^2\rangle^{eA} = \left. \int d\ell_{T}^2 \, \ell_{T}^2 \, \frac{d\sigma_{eA\rightarrow e\pi X}}{dx_B dQ^2 d\ell_{T}^2} \right/ \frac{d\sigma_{eA\rightarrow eX}}{dx_B dQ^2} \, ; \label{LTavg}$$ where $d\sigma_{eA\rightarrow eX}/dx_B dQ^2$ is the total inclusive DIS cross section normalized by the atomic weight $A$. In Eq. (\[LTavg\]), the Bjorken variable $x_B=Q^2/(2p\cdot q)$, and $q=k_1-k_2$ is the four-momentum of the virtual photon. The transverse momentum $\ell_{T}$ of the leading pion depends on our choice of the frame. In this paper, we choose the photon-nucleus frame, and choose the $z$-axis along the momentum line of the nucleus and the virtual photon. To separate the multiple scattering contribution from the single scattering contribution, we define the nuclear broadening of the transverse momentum square as $$\Delta\langle \ell_T^2\rangle \equiv \langle \ell_T^2\rangle^{eA} - \langle \ell_T^2\rangle^{eN} \, . \label{DLT2}$$ As we will demonstrate below, the nuclear broadening of the transverse momentum square defined in Eq. (\[DLT2\]) can be parameterized as [@E683; @E609] $$\Delta\langle \ell_T^2\rangle = a + b\, A^{1/3}\ , \label{BA13}$$ where $bA^{1/3}$ represents the contribution directly from the multiple scattering which is explicitly proportional to the nuclear size ($\propto A^{1/3}$) and term $a$ includes all contributions from the [localized]{} hard scattering as well as those suppressed by high power of $1/Q^2$. In principle, the parameter $a$ in Eq. (\[BA13\]) can also depend on the atomic weight $A$. But, as we will explain below, its dependence on $A$ should be very weak (e.g., $A^{\alpha}$ dependence with $\alpha \simeq \pm 0.02$). In this paper, we are interested in the second term in Eq. (\[BA13\]), and we show that experimental measurement of the constant $b$ in Eq. (\[BA13\]) will provide direct information on the quark-gluon correlation functions shown in Eq. (\[T4qA\]). To derive explicit expressions for the $a$ and $b$ in Eq. (\[BA13\]), we expand both numerator and the denominator in Eq. (\[LTavg\]) in terms of a power series of $1/Q^2$. In addition to the terms of leading power, we keep only the power suppressed terms that are explicitly enhanced by a factor $A^{1/3}$. Since the denominator, $d\sigma_{eA\rightarrow eX}/dx_B dQ^2$, is the total inclusive DIS cross section, the operator product expansion (OPE) allows us to expand it in terms of the power series of $1/Q^2$ [@CFP; @Ellis], $$\begin{aligned} \frac{d\sigma_{eA\rightarrow eX}}{dx_B dQ^2} &=& \sum_a \int dx\, \phi_{a/A}(x,\mu^2)\, \frac{d\hat{\sigma}^{(0)}_{ea\rightarrow eX}}{dx_B dQ^2} \left(x_B/x,\mu^2/Q^2,\alpha_s(\mu^2)\right) \left[\, 1 + O\left(\frac{1}{Q^2}\right)\right]\ \nonumber \\ &\equiv & D_A^{(0)}\left[\, 1 + O\left(\frac{1}{Q^2}\right)\right]\ , \label{D0}\end{aligned}$$ where $\mu$ is the factorization and/or renormalization scale, $\phi_{a/A}(x,\mu^2)$ is the leading twist parton distribution of flavor $a$ in a nucleus normalized by the atomic weight $A$, and $d\hat{\sigma}^{(0)}_{ea\rightarrow eX}/dx_B dQ^2$ is a perturbatively calculable short-distance hard part, which is independent of the nuclear medium. The [total]{} DIS cross section normalized by the atomic weight $A$ is an inclusive quantity and depends only on one hard scale $Q^2$. The QCD factorization theorem [@CFP; @Ellis] allows us to factorize each term in Eq. (\[D0\]) into a convolution of a short-distance coefficient function and a corresponding matrix element. The $1/Q^2$ term in Eq. (\[D0\]) can be expressed in terms of twist-4 matrix elements [@Ellis]. As demonstrated in Refs. [@QS_fac; @Qiu], integration of all position variables ($y_i$) of the field operators, which define these twist-4 matrix elements, are bounded by $1/x_Bp$ due to the oscillating exponential. Let $m$ and $r$ be nucleon mass and radius, respectively. If $1/x_Bp < 2r(m/p)$ (i.e. $x_B > 1/2mr \approx 0.1$), all the fields in these twist-4 matrix elements are bounded within the size of individual nucleon. Therefore, the power suppressed terms in Eq. (\[D0\]) is of the order of $O(1/Q^2)$ if $x_B > 0.1$, not $O(A^{1/3}/Q^2)$. On the other hand, the numerator in Eq. (\[LTavg\]) depends on two physically measured hadrons: the pion and the nucleus, and therefore, the OPE alone cannot be used to expand the numerator. However, the generalized factorization theorem introduced in Ref. [@QS_fac] can be used to factorize the numerator up to the $1/Q^2$ power corrections. Similarly to the Eq. (1) of Ref. [@LQS2], we expand the numerator in Eq. (\[LTavg\]) as $$\begin{aligned} \int d\ell_{T}^2 \, \ell_{T}^2 \, \frac{d\sigma_{eA\rightarrow e\pi X}}{dx_B dQ^2 d\ell_{T}^2} &=& \sum_{a,c} \phi_{a/A}(x) \otimes C^{(0)}_{ea\rightarrow ecX}\left(x_B/x,z=\ell/p_c,Q^2\right) \otimes D_{c\rightarrow\pi}(z) \nonumber \\ &+&\frac{1}{Q^2} \sum_{a,c} \left[ T_{a/A}(x) \otimes C^{(2)}_{ea\rightarrow ecX}\left(x_B/x,z=\ell/p_c,Q^2\right) \otimes D_{c\rightarrow\pi}(z) \right. \nonumber \\ && {\hskip 0.4in} \left. +\phi_{a/A}(x) \otimes \bar{C}^{(2)}_{ea\rightarrow ecX}\left(x_B/x,z=\ell/p_c,Q^2\right) \otimes d_{c\rightarrow\pi}(z) \right] \nonumber \\ &+& ... \nonumber \\ &\equiv & H^{(0)}_A + H^{(2)}_A + \bar{H}^{(2)}_A + ...\, , \label{Hexp}\end{aligned}$$ where $...$ represents terms further suppressed in $1/Q^2$, $\otimes $ represents the convolution over partons’ momentum fractions, and explicit dependence on the factorization and/or renormalization scale is suppressed for simplicity. In Eq. (\[Hexp\]), $C^{(0)}$, $C^{(2)}$, and $\bar{C}^{(2)}$ are perturbatively calculable short-distance hard parts [@QS_fac], and $D_{c\rightarrow \pi}$ and $d_{c\rightarrow \pi}$ are twist-2 and twist-4 parton-to-pion fragmentation functions, respectively. $\phi_{a/A}$ are nuclear parton distributions, and the $T_{a/A}$ in Eq. (\[Hexp\]) represents the twist-4 quark-gluon correlation functions (e.g., the one defined in Eq. (\[T4qA\])). Both $\phi_{a/A}$ and $T_{a/A}$ are normalized by the atomic weight $A$. Effective nuclear parton distributions $\phi_{a/A}(x)$ have been measured [@EMC; @SLAC; @NMC; @E665] and are known to have the nuclear shadowing for small $x$ ($x\simeq 0.1$), the EMC effect for intermediate $x$ values ($0.3\simeq x \simeq 0.7$), and Fermi motion effect for the large $x$ region. By fitting recent high precision DIS and Drell-Yan data on various nuclear targets, Eskola [et al.]{} [@Eskola] extracted a set of effective scale dependent nuclear parton distributions. In order to estimate the $A$-dependence of the nuclear parton distributions, we introduce a parameter $\alpha(A,x_B)$ as $$F_2^A(x_B,Q^2) \equiv A^{\alpha(A,x_B)}\, F_2^N(x_B,Q^2)\ , \quad\quad \mbox{or} \quad\quad \alpha(A,x_B) = \frac{\ln\left[F_2^A(x_B,Q^2)\left/F_2^N(x_B,Q^2)\right.\right]} {\ln\left[A\right]}\, , \label{F2}$$ where $F_2^A(x_B,Q^2)$ is a nuclear structure function normalized by $A$ and $F_2^N(x_B,Q^2)$ is an isoscalar nucleon structure function. Using the lowest order expression for the structure function: $F_2(x_B,Q^2)=\sum_q\, e_q^2\, x_B\, \phi_{q}(x_B,Q^2)$ with $e_q$ being a quark’s fractional charge and $\phi_{q}$ the effective nuclear parton distributions of Ref. [@Eskola], we plot the parameter $\alpha(A,x_B)$ as a function of $x_B$ for $A$=12, 64, and 207 at $Q^2=4$ and 9 GeV$^2$ in Fig. \[fig0\]a and \[fig0\]b, respectively. Fig. \[fig0\] shows that although the exact $x_B$-dependence and the $A$-dependence of the structure functions are non-trivial, the overall $A$-dependence of the structure functions is limited to $A^{\alpha}$ with a value of $\alpha\simeq \pm 0.02$ for $x_B$ values between two vertical lines in Fig. \[fig0\], which are relevant to pion production in this paper. Actually, as shown in Fig. \[fig0\], the absolute value of $\alpha$ is limited to 0.05 for any practical value of $x_B$ accessible at fixed target energies. Such $A$-dependence in the structure function as well as in the $\phi_{a/A}(x)$ is much weaker than the $A^{1/3}$-dependence caused by multiple scattering. Since the parton-to-pion fragmentation functions, $D_{c\rightarrow \pi}$ and $d_{c\rightarrow \pi}$, should not have explicit $A^{1/3}$-dependence, the term $\bar{H}^{(2)}_A$ is of the order of $O(1/Q^2)$, not $O(A^{1/3}/Q^2)$. On the other hand, the term $H^{(2)}_A$ depends on the multi-parton correlation function in a nucleus $T_{a/A}(x)$, and therefore, it will have an $A^{1/3}$-type enhancement [@LQS2]. Substituting Eqs. (\[D0\]) and (\[Hexp\]) into Eq. (\[LTavg\]), and keeping terms up to $O(A^{1/3}/Q^2)$, we obtain $$\langle \ell_{T}^2\rangle^{eA} \approx \frac{H^{(0)}_A+H^{(2)}_A}{D^{(0)}_A} + O\left(\frac{A^0}{Q^2}\right)\ . \label{expLTA}$$ Similarly, for a nucleon target, we have $$\langle \ell_{T}^2\rangle^{eN} \approx \frac{H^{(0)}_N}{D^{(0)}_N} + O\left(\frac{A^0}{Q^2}\right)\ . \label{expLTN}$$ Substituting above Eqs. (\[expLTA\]) and (\[expLTN\]) into our definition of the nuclear broadening of the transverse momentum square in Eq. (\[DLT2\]), we derive $$\Delta\langle \ell_T^2\rangle \approx \left[\frac{H^{(0)}_A}{D^{(0)}_A} - \frac{H^{(0)}_N}{D^{(0)}_N} \right] + \frac{H^{(2)}_A}{D^{(0)}_A} + O\left(\frac{A^0}{Q^2}\right)\ , \label{DLT2com}$$ where the power suppressed terms $O(A^0/Q^2)$ should be very small due to the cancellation between $\langle \ell_{T}^2\rangle^{eA}$ and $\langle \ell_{T}^2\rangle^{eN}$ and the fact that they are not enhanced by the $A^{1/3}$. The first term in Eq. (\[DLT2com\]) would be exactly equal to zero if the normalized effective nuclear parton distributions $\phi_{a/A}(x)=\phi_{a/N}(x)$. However, due to the well-known nuclear effects in the parton distributions, the first term in Eq. (\[DLT2com\]) does not have to vanish. But, because of very weak $A$-dependence of nuclear parton distributions shown in Fig. \[fig0\], we identify this term with the $a$ in Eq. (\[BA13\]), $$a \approx \frac{H^{(0)}_A}{D^{(0)}_A} - \frac{H^{(0)}_N}{D^{(0)}_N}\ . \label{DLTa}$$ It is the second term in Eq. (\[DLT2com\]) that is responsible for the $A^{1/3}$-type dependence of the nuclear broadening. We introduce $\Delta \langle \ell_T^2\rangle_{1/3}$ as $$\begin{aligned} \Delta \langle \ell_T^2\rangle_{1/3} &\equiv & b\, A^{1/3} \approx \frac{H^{(2)}_A}{D^{(0)}_A} \nonumber \\ &=& \left. \int d\ell_T^2 \, \ell_T^2 \, \frac{d\sigma^{(D)}_{eA\rightarrow e\pi X}}{dx_B dQ^2 d\ell_T^2} \right/ \frac{d\sigma^{(0)}_{eA\rightarrow eX}}{dx_B dQ^2} \, , \label{dlT20}\end{aligned}$$ where superscript “(D)” represents the double scattering contribution, and “(0)” stands for the leading power contribution. Our formalism in Eqs. (\[DLTa\]) and (\[dlT20\]) can be verified from the $A$-dependence of the measured transverse momentum broadening. From the factorized formula in Eq. (\[Hexp\]), the double scattering contribution $H_A^{(2)}$ depends only on the twist-2 parton-to-pion fragmentation function $D(z)$. Therefore, the differential cross section in Eq. (\[dlT20\]) for the pion production in DIS can be expressed as $$\frac{d\sigma^{(D)}_{eA\rightarrow e\pi X}}{dx_B dQ^2 d\ell_T^2} =\sum_{c}\, \int \frac{d{\sigma}^{(D)}_{eA\rightarrow ecX}}{dx_B dQ^2 dp_{c_T}^2} \cdot D_{c \rightarrow \pi} (z) \cdot \frac{dz}{z^2} \ , \label{sigma-pi}$$ where $\sum_c$ runs over all parton flavors $c$, $z=\ell/p_c$ is the momentum fraction carried by the produced pion, and $D_{c\rightarrow\pi}(z)$ is the fragmentation function for the parton of momentum $p_c$ to become a pion of momentum $\ell$. In Eq. (\[sigma-pi\]), $d{\sigma}^{(D)}_{eA\rightarrow ecX}/ dx_B dQ^2 dp_{c_T}^2$ is the double scattering contributions to the differential cross section to produce a parton of momentum $p_c$ in DIS, and it can be written as $$\left[ \frac{d{\sigma}^{(D)}_{eA\rightarrow ecX}}{dx_B dQ^2 dp_{c_T}^2} \right]\, dp_{c_T}^2 =\frac{1}{8\pi}\, \frac{e^2}{x_B^2s^2Q^2} \, L^{\mu\nu}(k_1,k_2)\, W^{(D)}_{\mu\nu}(x_B,Q^2,p_c)\ , \label{sigma-c}$$ where $s=(p+k_1)^2$ is the total invariant mass of the lepton-nucleon system. In Eq. (\[sigma-c\]), the leptonic tensor $L^{\mu\nu}$ is given by the diagram in Fig. \[fig1\]a, $$L^{\mu\nu}(k_1,k_2) =\frac{1}{2}\, {\rm Tr}(\gamma \cdot k_1 \gamma^{\mu} \gamma \cdot k_2 \gamma^{\nu}) \ , \label{L}$$ and $W^{(D)}_{\mu\nu}$ is the hadronic tensor due to double scattering, which is given by the diagram shown in Fig. \[fig1\]b. For comparison, the lowest order differential cross section for producing a parton of momentum $p_c$ due to single scattering is given by $$\left[ \frac{d{\sigma}^{(0)}_{eA\rightarrow ecX}}{dx_B dQ^2 dp_{c_T}^2} \right]\, dp_{c_T}^2 = \left[ \frac{d{\sigma}^{(0)}_{eA\rightarrow ecX}}{dx_B dQ^2}\, \delta(p_{c_T}^2) \right]\, dp_{c_T}^2 \label{single}$$ where $d{\sigma}^{(0)}/dx_B dQ^2$ is the leading order inclusive DIS cross section, which appears in the definition of $\Delta\langle\ell^2_T\rangle_{1/3}$ in Eq. (\[dlT20\]). The complete double scattering contributions to the production of a leading quark of momentum $p_c$ in DIS at the leading order of $\alpha_s$ come from the diagrams shown in Fig. \[fig2\], which represent the final-state interactions, plus the same order diagrams involving initial-state interactions shown in Fig. \[fig2p\]. In the following, we first discuss the role of initial-state interactions, and argue that although required by gauge invariance, they contribute to the $a$ term of Eq. (\[BA13\]) only. After the collinear expansion of the parton momenta, the gluon interactions on the initial quark lines, as shown in Fig. \[fig2p\], can be reduced into two categories: the long-distance and the short-distance contributions due to the fact that every “on-shell” propagator can have a pole contribution as well as a contact contribution [@Qiu]. For example, a quark propagator of momentum $k$ can always be written as $$\frac{i\gamma\cdot k}{k^2} = \frac{i\gamma\cdot \hat{k}}{k^2} + \frac{i\gamma\cdot n}{2k\cdot n}\ , \label{DF0}$$ where $\hat{k}^2 = 0$ and $n^\mu$ is any auxiliary vector with $k\cdot n \neq 0$. The first term in the right-hand-side of Eq. (\[DF0\]) corresponds the pole contribution, while the second term is the contact contribution [@Qiu]. Attaching one gluon to the initial quark line introduces a quark propagator, and this propagator will have both the pole and contact contributions. The pole contribution is long-distance in nature, representing the interactions between the quark and the gluon long before the hard collision between the quark and the virtual photon. The pole contribution is partially responsible for the $A$-dependence of the leading-twist parton distributions in a nucleus [@CQR], and therefore, is one of the sources for the weak $A$-dependence appeared in the $a$ term in Eq. (\[BA13\]). Because of the nature of the contact term, its contribution is localized in space, and therefore, it does not contribute to the $A^{1/3}$-type of nuclear enhancement [@QS_fac]. Since the short distance contributions of Feynman diagrams in Fig. \[fig2p\] have at least one propagator given by the contact term, these diagrams do not contribute to the $A^{1/3}$-type of nuclear enhancement. At the same time, the contact term of the initial-state interaction is important for the gauge invariance of the complete double scattering (twist-4) process [@Qiu], but only at order $A^0$. Although the Feynman diagrams with the final-state interactions in Fig. \[fig2\] are not gauge invariant in general, their contributions to the $A^{1/3}$-behavior of nuclear enhancement are observable and hence gauge invariant [@LQS2]. The three lowest order diagrams in Fig. \[fig2\] are all convoluted with the same two-quark-two-gluon matrix element through three independent parton momenta. The leading power contributions of these diagrams come from the region phase space when all partons are collinear to the direction of the target, as shown in Fig. \[fig2\]. In order to convert the gluon field $A^+$ to the corresponding field strength $F^{+\perp}$ in covariant gauge, we expand the gluon momenta in an extra transverse component $k_T$ in Fig. \[fig2\] [@LQS2]. All three Feynman diagrams in Fig. \[fig2\] have two potential poles due to the interactions between the final-state quark and gluons. These two poles (which are not pinched) can be used to carry out the momentum fraction $dx_1$ and $dx_2$ integrals [@LQS1; @LQS2]. After taking the poles, the hard-scattering factor (the interaction between the virtual photon and the initial quark) and the final-state interaction between the “on-shell” outgoing quark and the gluon are separately gauge invariant. To derive the leading contribution from the double scattering, we follow the derivation in Ref. [@Guo2]. By taking the poles to fix the integrations of momentum fractions $x_1$ and $x_2$, we derive contributions of all three diagrams in Fig. \[fig2\] to the hadronic tensor $W^{(D)}_{\mu\nu}$ in Eq. (\[sigma-c\]) as $$\begin{aligned} W^{(D)}_{\mu\nu}(x_B,Q^2,p_c) &=& \frac{1}{4\pi}\, \int \frac{dy^-}{4\pi}\, \frac{dy_1^-dy_2^-}{(2\pi)^2}\, \frac{d^2y_T}{(2\pi)^2} \, \int d^2 k_T\, e^{ik_T \cdot y_{1T}} \, e^{i \frac{k_T^2-2k_T \cdot q}{2p \cdot q} p^+ (y_1^--y_2^-)} \, \nonumber \\ &\ & \times\, \langle p_A \| \bar{\psi}_q(0)\,\gamma^+ \, A^+(y_{1}^{-},y_{1T})\, A^+(y_{2}^{-}) \psi_q(y^{-})\, \| p_{A}\rangle \nonumber \\ &\ & \times (2\pi) \theta (y_1^--y^-) \, (2\pi) \theta (y_2^-) \, e^{i x_B p^+ y^- } \nonumber \\ &\ & \times \frac{2\pi}{(2p\cdot q)^3} \cdot \hat{H}^{(D)}_{\mu\nu}(p_c) \cdot \left[ \delta (p_{c_T}^2-k_T^2) - \delta (p_{c_T}^2) \right] dp_{c_T}^2 \\ & \approx & \left[ \frac{\pi}{(2p\cdot q)^3}\, \hat{H}^{(D)}_{\mu\nu}(p_c) \, \left[-\delta ' (p_{c_T}^2) \right] T_{qF}^A(x_B,\mu^2) \right] \, dp_{c_T}^2 \ . \label{W-c}\end{aligned}$$ In Eq. (\[W-c\]), $T_{qF}^A(x_B,\mu^2)$ is defined by Eq. (\[T4qA\]) with the factorization scale $\mu^2$. $\hat{H}^{(D)}_{\mu\nu}(p_c)$ represents the spinor trace of the partonic part of the double scattering and is given by $$\hat{H}^{(D)}_{\mu\nu}(p_c) =\frac{4}{3} \pi^2 \alpha_s \alpha_{em} e_q^2 \, {\rm Tr}\left[\gamma \cdot p \gamma_{\mu} \gamma \cdot (xp+q) \gamma^{\sigma} \gamma \cdot p_c \gamma^{\rho} \gamma \cdot (xp+q) \gamma_{\nu}\right] \, p_{\rho}\,p_{\sigma}\ , \label{hatH}$$ where a color factor $1/2N_c$ with $N_c=3$ was included. Using the definition $z=\ell/p_c$, see Eq. (\[z\]) below, we can reexpress the momentum $p_c$ in $\delta'(p^2_{c_T})$ in terms of the momentum $\ell$ of the observed pion, $$-\delta '(p_{c_T}^2) =- \frac{d}{d(\frac{\ell_{T}^2}{z^2})} \left( \delta (\frac{\ell_{T}^2}{z^2}) \right) = - z^4 \delta '(\ell_{T}^2) \ . \label{dpi}$$ Combining Eqs. (\[sigma-pi\]), (\[sigma-c\]), (\[W-c\]), and (\[dpi\]), we obtain $$\frac{d\sigma^{(D)}_{eA\rightarrow e\pi X}}{dx_B dQ^2 d\ell_T^2} =\frac{e^2}{8\pi x_B^2 s^2 Q^2} \, \frac{\pi}{(2p\cdot q)^3} \sum_q\int dz\, z^2 \, D_{q\rightarrow \pi}(z)\, T_{qF}^A(x_B,Q^2)\, L^{\mu\nu} \hat{H}^{(D)}_{\mu\nu}(\ell/z) \, \left[-\delta ' (\ell_{T}^2) \right]\, , \label{sigma-pi2}$$ where the $\sum_q$ runs over all quark and antiquark flavors, and $z=\ell/p_q$ is the momentum fraction of the quark flavor $q$ carried by the observed pion. At the lowest order in $\alpha_s$, hard gluon initiated double scattering, which is proportional to the four-gluon correlation function $T_{gF}$, does not contribute [@LQS1; @LQS2]. In the photon-nucleus frame, we choose the target momentum $p$ along the $-\vec{z}$-axis, such that $p^{\mu}=(p^0,p^x,p^y,p^z)=(P,0,0,-P)$, and only $p^-=(p^0-p^z)/\sqrt{2}$ is nonvanishing. In this frame, we have the following expression for the momentum fraction $z=\ell/p_c$, $$\begin{aligned} z &\equiv& \frac{\ell}{p_{c}}=\frac{\ell^{+}}{p_{c}^{+}} =\frac{p\cdot \ell }{p \cdot p_{c}} \nonumber \\ &\approx & \frac{p\cdot \ell }{p\cdot q} \ . \label{z}\end{aligned}$$ In deriving the last equation, we used $p_c=x_Bp+q$ and $p^2\approx 0$. After working out the algebra for $L^{\mu\nu} \hat{H}^{(D)}_{\mu\nu}$, and substituting Eq. (\[sigma-pi2\]) into Eq. (\[dlT20\]), we obtain $$\Delta\langle \ell_{T}^2\rangle_{1/3} \left(\ell_{\rm min}\right) = \frac{4\pi^2\alpha_s(\mu^2)}{3}\, \frac{\sum_q e_q^2\,\int_{z_{\rm min}}^1 dz\, z^2 D_{q\rightarrow\pi}(z) T_{qF}^{A}(x_B,\mu^2)} {\sum_q e_q^2\,q^{A}(x_B,\mu^2)}\, , \label{pilt}$$ where $D_{q\rightarrow\pi}(z)$ are the quark-to-pion fragmentation functions, and $q^A(x_B,\mu^2)=\phi_{q/A}(x_B,\mu^2)$ are the effective quark distributions in the nucleus normalized by the atomic weight $A$. In Eq. (\[pilt\]), $z_{\rm min}=p\cdot \ell_{\rm min}/(p\cdot q)$ is defined in terms of $\ell_{\rm min}$, which is the momentum cut for the pions measured in the experiment. One can choose $\ell_{\rm min}$ to be large enough to ensure that the observed pions are from the fragmentation of energetic quarks. In Eq. (\[pilt\]), the $\mu^2$ in $\alpha_s(\mu^2)$ is the renormalization scale, and the $\mu^2$ in $T_{qF}^{A}(x_B,\mu^2)$ and $q^{A}(x_B,\mu^2)$ are the factorization scale. Since the short-distance part of the transverse momentum broadening defined in this paper is an inclusive and infrared safe calculable quantity, the scale $\mu^2$ should be chosen to be the order of the physically measured momentum scales. As shown in Eq. (\[pilt\]), the only physically measured momentum scales for the transverse momentum broadening are $Q$ and $\ell_{\rm min}$. Since $\|\vec{\ell}_{\rm min}\|$ are of the same order as $\sqrt{Q^2}$, we choose $\mu^2=Q^2$ for the numerical calculations below, and we believe that such a choice will not result into the large logarithmic high order corrections. Since the quark-to-pion fragmentation functions have been measured [@newDD], Eq. (\[pilt\]) shows that the transverse momentum broadening of pions provides direct information on the quark-gluon correlation functions inside a nucleus. The measured size of the transverse momentum broadening, $\Delta\langle \ell_{T}^2\rangle_{1/3}$, depends on the choice of the momentum cut $\ell_{\rm min}$ (or equivalently $z_{\rm min}$). Our predictions should be more reliable for the leading pions, or pions with relatively large momenta. By measuring the transverse momentum broadening for $\pi^{\pm}$ and $\pi^0$, and keeping a reasonable large value of $z_{\rm min}$, we can extract the quark flavor dependence of the correlation functions. Numerical Results and Discussions ================================= The numerical values of the pion transverse momentum broadening, $\Delta\langle \ell_{T}^2\rangle_{1/3} (\ell_{\rm min})$, depend on the explicit functional form of the quark-to-pion fragmentation functions and the quark-gluon correlation functions. In this section, without assuming any specific form for the quark-gluon correlation functions, we derive relations for the transverse momentum broadening between $\pi^{\pm}$ and $\pi^0$ based on isospin symmetry and the charge conjugation invariance. Furthermore, by using the simple model proposed in Ref. [@LQS1] for the quark-gluon correlation functions, we explore both the normalization and functional dependence of the momentum broadening. Although new parameterizations of the quark-to-pion fragmentation functions were obtained recently [@newDD], we will use the simple parameterizations of Ref.[@Dpi] for the following analytical derivations and discussions on the general features of the momentum broadening. Later, when we present our figures for the numerical values of the momentum broadening, we will use the parameterizations from both Refs. [@Dpi] and [@newD], and demonstrate the similarities and differences. Like the parton distributions, the quark-to-pion fragmentation functions have scaling violation (or $Q^2$-dependence). To simplify our discussion on the general features of transverse momentum broadening, we ignore the scaling violation of the fragmentation functions, and take for the fragmentation functions the input distributions of Ref. [@Dpi], which are given as: \[D\] $$\begin{aligned} D^+(z)=D_u^{\pi ^+}=D_d^{\pi ^-}=D_{\bar{u}}^{\pi ^-} =D_{\bar{d}}^{\pi ^+} &=&\frac{1-z^2}{4z} \ ; \label{D+} \\ D^-(z)=D_u^{\pi ^-}=D_d^{\pi ^+}=D_{\bar{u}}^{\pi ^+} =D_{\bar{d}}^{\pi ^-} &=& \frac{(1-z)^2}{4z} \ ; \label{D-} \\ D^0(z)=D_u^{\pi ^0}=D_d^{\pi ^0}=D_{\bar{u}}^{\pi ^0} =D_{\bar{d}}^{\pi ^0} &=& \frac{1-z}{4z} \ ; \label{D01} \end{aligned}$$ with $D^0(z)=[D^+(z)+D^-(z)]/2$. For the strange quark fragmentation functions, we use $D_s^{\pi ^+}=D_s^{\pi ^-}=D_s^{\pi ^0}=D^-$. Notice that the fragmentation functions given in Eq. (\[D\]), as well as those given in Ref. [@newDD; @newD], violate Gribov-Lipatov reciprocity, which requires that the power of $(1-z)$ must be even. However, since it is not our purpose to invent better fragmentation functions in this paper, we will first use the fragmentation functions in Eq. (\[D\]) in our numerical calculation to illustrate the size and the general features of the transverse momentum broadening. Our predictions can be easily adjusted for other sets of fragmentation functions. In order to evaluate the transverse momentum broadening, Eq. (\[pilt\]) requires the moments of the fragmentation functions. We therefore introduce $$D^{(i)}(n,z_{\rm min}) \equiv \int_{z_{\rm min}}^1 \, dz\, z^n\, D^{(i)}(z) \ , \label{Dzn}$$ where $i=+, -$ and $0$ for $\pi^+$, $\pi^-$ and $\pi^0$, respectively. Using the fragmentation functions defined in Eq. (\[D\]), we obtain the following identity, $$D^{+}(n,z_{\rm min}) - D^{0}(n,z_{\rm min}) = D^{0}(n,z_{\rm min}) - D^{-}(n,z_{\rm min}) \ , \label{Dp0n}$$ for all $n$. This identity is useful for the following discussions on the general features of the transverse momentum broadening. From Eq. (\[pilt\]), the transverse momentum broadening, $\Delta\langle \ell_{T}^2\rangle_{1/3} (\ell_{\rm min})$, depends on the second moments $D^{(i)}(2,z_{\rm min})$, which are plotted in Fig. \[fig3\]. Given the quark-to-pion fragmentation functions, we derive the following general relations between the transverse momentum broadening of $\pi^+$, $\pi^-$, and $\pi^0$ particles: $$\Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^+} - \Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^0} = \Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^0} - \Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^-} \ , \label{symmetry}$$ and $$\Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^+} > \Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^0} > \Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^-} \ , \label{p0m}$$ independent of the details of the correlation functions $T_{qF}^A$. Eq. (\[symmetry\]) is a result of the isospin symmetry of the fragmentation functions given in Eq. (\[D\]) and the identity given in Eq. (\[Dp0n\]). The inequality given in Eq. (\[p0m\]) results from the fact that the contribution of each quark flavor is weighted by the square of the quark’s fractional charge, $e_q^2$, and the relation $D^+(z)>D^0(z)>D^-(z)$. Since the correlation functions $T_{qF}^A$ are not known, in order to obtain numerical estimates of the transverse momentum broadening of pions, we adopt the following model for quark-gluon correlation functions [@LQS1; @LQS2]: $$T_{qF}^A(x,Q^2)=\lambda^2 A^{1/3} q^A(x,Q^2) \ , \label{TiM}$$ where $q^A(x)$ is the effective twist-2 quark distribution of a nucleus normalized by the atomic weight $A$, and $\lambda$ is a free parameter to be fixed by experimental data. This model was proposed after comparing the operator definitions of the four-parton correlation functions and the definitions of the normal twist-2 parton distributions. Substituting Eq. (\[TiM\]) into Eq. (\[pilt\]), we have $$\Delta\langle \ell_{T}^2\rangle_{1/3} = \frac{4\pi^2\alpha_s(Q^2)}{3}\, A^{1/3}\, \lambda^2\, \frac{\sum_q e_q^2\,q^{A}(x_B,Q^2)\, D_{q \rightarrow \pi}(2,z_{\rm min})} {\sum_q e_q^2\,q^{A}(x_B,Q^2)}\, . \label{pilt2}$$ From Eq. (\[pilt2\]), we can deduce a very simple formula for the transverse momentum broadening of $\pi^0$ by using the fragmentation functions given in Eq. (\[D01\]): $$\Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^0} \approx \frac{4\pi^2\alpha_s(Q^2)}{3}\, A^{1/3}\, \lambda^2\, D^0(2,z_{\rm min})\ . \label{ltpi0}$$ In deriving Eq. (\[ltpi0\]), we neglected the strange quark contribution which is much smaller. This simple expression shows that $\Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^0}$ has a scaling behavior as $x_B$ varies. The approximate $x_B$-scaling behavior of $\Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^0}$ is a direct consequence of the model for $T_{qF}^A(x)$, given in Eq. (\[TiM\]). If the future experimental data on the transverse momentum broadening of $\pi^0$ shows a strong violation of the $x_B$-scaling, the model for the quark-gluon correlation function $T_{qF}^A(x_B)$, given in Eq. (\[TiM\]), will have to be modified. In addition, Eq. (\[ltpi0\]) also shows that $Q^2$ dependence of $\Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^0}$ is mainly from the $Q^2$ dependence of $\alpha_s(Q^2)$. Since the $Q^2$ dependence of $\alpha_s(Q^2)$ is already known, we can learn the $Q^2$ dependence of the correlation function $T_{qF}^A(x_B,Q^2)$ by measuring the $Q^2$ dependence of $\Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^0}$. If the measured $Q^2$ dependence of $\Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^0}$ is consistent with the $Q^2$ dependence of $\alpha_s(Q^2)$, it means the correlation function $T_{qF}^A(x_B,Q^2)$ has a similar scale-dependence as a normal parton distribution. By examining the $x_B$-scaling property and $Q^2$ dependence of $\Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^0}$, we can test whether the model for $T_{qF}^A(x_B,Q^2)$, given in Eq. (\[TiM\]), is reasonable. It should be emphasized that our general conclusions given in last paragraph, as well as numerical estimates given below, are the consequences of our lowest order calculations. High order corrections in $\alpha_s$ and/or in inverse powers of $Q^2$ can certainly change the dependence on $x_B$ and $Q^2$, as well as the absolute values of our numerical estimates. But, because of the inclusivity and the infrared safety of the transverse momentum broadening, and the lack of the large logarithms of the ratios of the physically measured scales, we believe that the change should not be very dramatic, and that our predictions should not be off by orders of magnitudes. In the following, we obtain our numerical estimates of transverse momentum broadening for $\pi^+$, $\pi^-$, and $\pi^0$ particles by evaluating Eq. (\[pilt2\]). We normalize $\Delta\langle \ell_{T}^2\rangle_{1/3}$ by $A^{1/3}$, and plot our results in Fig. \[fig5\] to Fig. \[fig6\]. In view of the limited range of $Q^2$ available at CEBAF, we neglect the scaling violation of the fragmentation functions for our numerical estimates. We used the simple fragmentation functions given in Eq. (\[D\]) to produce Fig. \[fig5\] to Fig. \[fig6\]. More realistic parameterizations for quark-to-pion fragmentation functions were obtained in Refs. [@newDD] and [@newD]. In particular, those in Ref. [@newD] are extracted from DIS data. We present both results for $\Delta\langle \ell_{T}^2\rangle_{1/3}^{\pi^\pm} / A^{1/3}$ by using two different sets of the fragmentation functions at $Q^2=4$ GeV$^2$ in Fig. \[fig7\] to illustrate the uncertainties in our numerical estimates due to different choices of the fragmentation functions. Although numerical details of the $A^{1/3}$-type transverse momentum broadening depend on the fragmentation functions used, the overall shape and magnitude of the broadening are consistent. The most important feature is that the size of the $A^{1/3}$-type broadening is large enough to be measured experimentally [@E683; @E609]. In obtaining our numerical results, we used CTEQ3L parton distributions of Ref. [@CTEQ3] for the quark distributions in the nucleons. We took an average of the quark distributions of the proton and the neutron for the effective quark distribution in the nucleus, $q^A(x_B)$. If the $A$-dependence of the nuclear quark distributions is independent of the flavor, from Eq. (\[pilt2\]), such nuclear dependence can be factorized and canceled between the numerator and denominator. However, there is no obvious reason why the nuclear dependence of quark distributions is flavor independent. But, as we discussed above, in the range of $x$ probed in these experiments the effective nuclear parton distribution should have extremely weak nuclear dependence in comparison with the $A^{1/3}$-type enhancement that we discussed in this paper. In order to verify this feature, we used both $q^A(x_B)=q^N(x_B)$ and the realistic parameterizations for $q^A(x_B)$ given in Ref. [@Eskola] to calculate the transverse momentum broadening $\Delta\langle \ell_{T}^2\rangle_{1/3}^{\pi^\pm} / A^{1/3}$. We found no noticeable difference for the transverse momentum broadening presented in Figs. \[fig5\], \[fig6\], and \[fig7\], which were obtained with $q^A(x_B)=q^N(x_B)$. Figs. \[fig5\]a and \[fig5\]b shows $\Delta \langle \ell_{T}^2\rangle_{1/3} / A^{1/3}$ as functions of $z_{\rm min}$ at $Q^2=4$ GeV$^2$, and $x_B=0.2$ and $0.3$, respectively. Figs. \[fig6\]a and \[fig6\]b show $\Delta \langle \ell_{T}^2\rangle_{1/3} / A^{1/3}$ as functions of $z_{\rm min}$ at $Q^2=9$ GeV$^2$, and $x_B=0.2$ and $0.3$, respectively. In these figures, the relations $\Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^+} > \Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^0} > \Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^-}$ and $\Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^+} - \Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^0} = \Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^0} - \Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^-}$ are clearly demonstrated. Furthermore, we have $x_B$-scaling for $\Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^0}$, and $\Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^+} / \Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^-} \approx 2$ for small $z_{\rm min}$. The approximate ratio, $\Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^+} / \Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^-} \approx 2$, is a result of the isospin averaged targets, and can be easily verified by using the fragmentation functions in Eq. (\[D\]), keeping only the valence quark contributions and $z_{\rm min}^2 \ll 1$. Comparing Figs. \[fig5\] and \[fig6\], it is evident that the absolute sizes of the $\Delta \langle \ell_{T}^2\rangle_{1/3}$ decrease as $Q^2$ increases. This is caused by the $Q^2$ dependence of the $\alpha_s(Q^2)$ in the overall factor of the transverse momentum broadening. Due to the available energy at CEBAF, we plotted the broadening at $Q^2=4$ GeV$^2$ in Fig. \[fig5\]. Even though $Q^2=4$ GeV$^2$ may not be large enough to apply for the fragmentation analysis, our results can be tested at CEBAF with its future upgrade, and can be easily tested at RHIC with its future option of electron-ion collider. From Figs. \[fig5\] and \[fig6\], we noticed that the identity, $\Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^+} - \Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^0} = \Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^0} - \Delta \langle \ell_{T}^2\rangle_{1/3}^{\pi^-}$, has some weak violation on the value of $x_B$, which is due to the fact that the number of sea quark increases at smaller $x_B$ in comparison with the valence quarks. The value of $\lambda^2$ for the quark-gluon correlation function $T_{qF}^A$, given in Eq. (\[TiM\]), has not been well determined. It was estimated in Ref. [@LQS2] by using the nuclear enhancement of the momentum imbalance of two jets in photon-nucleus collisions [@E683; @E609], and was found to be the order of $\lambda^{2} \sim 0.05 - 0.1 \mbox{GeV}^{2}$. However, the nuclear enhancement of the Drell-Yan transverse momentum [@NA10; @E772] indicates a much smaller value of $\lambda^2$ [@Guo2]. Although the momentum imbalance of the di-jet data depends on the final-state multiple scattering, and the Drell-Yan data is an effect of initial-state multiple scattering, the leading order calculation indicates that $\lambda^2$ from the two data sets to be the same [@Guo2]. Therefore, this work on the transverse momentum broadening in pion productions should provide a new and independent measurement of the size of the quark-gluon correlation functions, and the value of the $\lambda^2$, which is extremely important for understanding the nuclear dependence and the event rates in the relativistic heavy ion collisions. Since both di-jet momentum imbalance and the transverse momentum broadening are caused by the final state multiple scattering, we used $\lambda^2=0.05$ GeV$^2$ for Figs. \[fig5\] to \[fig7\]. Any change in $\lambda^2$ will only change the overall normalization factor in $\Delta\langle \ell_{T}^2\rangle_{1/3} /A^{1/3}$, and hence does not affect the general features of $\Delta\langle \ell_{T}^2\rangle_{1/3} /A^{1/3}$. We conclude that the transverse momentum broadening of leading pions in deeply inelastic lepton-nucleus scattering is an excellent observable to probe the parton correlation functions in the nucleus. Measuring the $A^{1/3}$-type transverse momentum broadening $\Delta\langle \ell_{T}^2\rangle_{1/3}$ in the leading pion production in DIS provides an independent test for the existing model of the quark-gluon correlation functions [@LQS1]. More importantly, by measuring such broadening, we can directly measure the $x$ and $Q^2$ dependence of the correlation functions, which is very useful for predicting the event rates [@FMSS] and for understanding the nuclear dependence in relativistic heavy ion collisions. Acknowledgment {#acknowledgment .unnumbered} ============== We thank G. Sterman for very helpful discussions. This work was supported in part by the U.S. Department of Energy under Contract No. DE-AC02-98CH10886, and Grant Nos. DE-FG02-87ER40731 and DE-FG02-96ER40989. A.H. Mueller, in the proceeding of the [Workshop on CEBAF at Higher Energies]{}, ed. N. Isgur and P. Stoler (Newport News, USA, 1994). M. Gyulassy and X.-N. Wang, Nucl. Phys. [B420]{}, 583 (1994); X.-N. Wang, M. Gyulassy and M. Plümer, Phys. Rev. [D51]{}, 3436 (1995). R. Baier, Yu.L. Dokshitzer, A.H. Mueller, and D. Schiff, hep-ph/9907267, and references therein. G.T. Bodwin, S.J. Brodsky and G.P. Lepage, Phys. Rev. D[39]{}, 3287 (1989). M. Luo, J.-W. Qiu and G. Sterman, Phys. Lett. [B279]{}, 377 (1992); M. Luo, J.-W. Qiu and G. Sterman, Phys. Rev. D[50]{}, 1951 (1994). X.-F. Guo, Phys. Rev. D [58]{}, 114033 (1998). J.C. Collins, D.E. Soper, and G. Sterman, in [Perturbative Quantum Chromodynamics]{}, edited by A.H. Mueller, 1989 (World Scientific, Singapore); and references therein. J.-W. Qiu and G. Sterman, Nucl. Phys. [B353]{}, 105 (1991); [B353]{}, 137 (1991). P. Bordalo [et al.]{}, Phys. Lett. B [193]{}, 373 (1987); D.M. Alde [et al.]{}, Phys. Rev. Lett. [66]{}, 2285 (1991). M. Luo, J.-W. Qiu and G. Sterman, Phys. Rev. D[49]{}, 4493 (1994). S.J. Brodsky, in the proceeding of the [Workshop on CEBAF at Higher Energies]{}, ed. N. Isgur and P. Stoler (Newport News, USA, 1994). X.-F. Guo, Phys. Rev. D[58]{}, 036001 (1998). D. Naples, [et al.]{}, Phys. Rev. Lett. [72]{}, 2341 (1994). M. D. Corcoran [et al.]{}, Phys. Lett. [B259]{} (1991), 209; T. Fields, Nucl. Phys. [A544]{}, 565 (1992). G. Curci, W. Furmanski, and R. Petronzio, Nucl. Phys. [B175]{}, 27 (1980). R.K. Ellis, W. Furmanski, and R. Petronzio, Nucl. Phys. [B207]{}, 1 (1982); [B212]{}, 29 (1983). J.-W. Qiu, Phys. Rev. D[42]{}, 30 (1990); and X.-F. Guo and J.-W. Qiu, hep-ph/9810548. EMC Collaboration, J.J. Aubert, [et al.]{}, Phys. Lett. [B123]{}, 275 (1983). A. Bodek, [et al.]{}, Phys. Rev. Lett. [50]{}, 1431 (1983); [51]{}, 534 (1983). NMC Collaboration, P. Amaudruz, [et al.]{}, Nucl. Phys. [B441]{}, 3 (1995); NMC Collaboration, M. Arneodo, [et al.]{}, Nucl. Phys. [B441]{}, 12 (1995); Nucl. Phys. [B481]{}, 3 (1996); Nucl. Phys. [B481]{}, 23 (1996). E665 Collaboration, M.R. Adams, [et al.]{}, Phys. Rev. Lett. [68]{}, 3266 (1992); Z. Phys. C67, 403(1995). K.J. Eskola, V.J. Kolhinen, and C.A. Salgado Eur. Phys. J. [C9]{}, 61 (1999). K.J. Eskola, V.J. Kolhinen, and P.V. Ruuskanen, Nucl. Phys. [B535]{}, 351 (1998). For example, see F.E. Close, J.-W. Qiu and D.G. Roberts, Phys. Rev. D[40]{}, 2820 (1989). P. Chiappetta, M. Greco, J.P. Guillet, S. Rolli, and M. Werlen, Nucl. Phys. [B412]{}, 3 (1994); and references therein. R. Baier, J. Engles and B. Petersson, Z. Phys. C2, 265(1979). J.J. Aubert, [et al.]{}, Phys. Lett. [B160]{}, 417(1985); M. Arneodo, [et al.]{}. Nucl. Phys. [B321]{}, 541(1989). H. L. Lai, [et al.]{}, Phys. Rev. D[51]{}, 4763 (1995); and references therein. P. Bordalo, et al, Phys. Lett. [B193]{}, 373(1987). D. M. Alde, et al, Phys. Rev. Lett. [66]{}, 2285(1991) R.J. Fries, B. Müller, A. Schäfer, and E. Stein, hep-ph/9907567.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We discuss the interplay between transport and intrinsic dissipation in quantum Hall bilayers, within the framework of a simple thought experiment. We compute, for the first time, quantum corrections to the semiclassical dynamics of this system. This allows us to re-interpret tunnelling measurements on these systems. We find a strong peak in the zero-temperature tunnelling current that arises from the decay of Josephson-like oscillations into incoherent charge fluctuations. In the presence of an in-plane field, resonances in the tunnelling current develop an asymmetric lineshape.' author: - 'Robert L. Jack' - 'Derek K. K. Lee' - 'Nigel R. Cooper' title: Dissipation and Tunnelling in Quantum Hall Bilayers --- Transport in quantum Hall bilayers has been the subject of much recent interest[@qh_pers; @Moon1995; @Balents2001; @bl_theory; @Fertig2003; @Spielman2000on; @Kellogg2004]. The bilayers consist of two closely-spaced parallel two-dimensional electron layers in a double quantum well. If the Landau level fillings of the layers are $\nu_1=\nu_2=1/2$, then Coulomb interactions between the layers drive a transition to a ground state in which the bilayer as a whole exhibits the quantum Hall effect[@qh_pers]. This ground state is believed to have a broken U(1) symmetry[@bl_json]; there is a macroscopically coherent phase associated with the electrons’ layer degree of freedom. The ground state can be viewed as an easy-plane ferromagnet[@Moon1995] or, equivalently, as an excitonic condensate[@Fertig1989; @Balents2001]. There are also analogies with Josephson junctions[@bl_json]. A series of remarkable experiments[@Eisenstein92-94; @Spielman2000on; @Kellogg2004] have been used to probe the internal degrees of freedom of this strongly correlated system. They show evidence for interlayer coherence and the linearly dispersing Goldstone mode resulting from the broken U(1) symmetry[@Spielman2000on], and also, most recently, “excitonic” superfluidity[@Kellogg2004]. There still remain questions concerning the interlayer tunnelling spectrum of the bilayers. One prominent feature in the $IV$ characteristic is a sharp peak in the tunnelling current for small biases (between 10 and 100 $\mu$V)[@Spielman2000on]. At low temperatures, reducing the bias leads to a sharp rise in the tunnelling current. This increase is cut off below 10 $\mu$V so that the current falls to zero at zero bias. Existing theories for interlayer tunnelling are restricted to the classical limit of the underlying spin model, and do not produce this feature at the low temperatures relevant to experiment ($k_BT\sim 2\mu \mathrm{eV}$). In this paper we study interlayer tunnelling within the framework of a simple “thought experiment”. We follow the relaxation of an initial charge imbalance across the bilayer. Including quantum ($1/S$) corrections to the dynamics of the pseudospin model of the bilayer[@Moon1995], we see that electron tunnelling across the bilayer generates density waves. This quantum dissipative process leads to a zero-temperature tunnelling current of the form $I\propto 1/V$ for a bias voltage $V$ above a threshold $V_0$; this is consistent with experimental measurements. The long-wavelength density fluctuations in the bilayer have an energy gap ${\Delta_\mathrm{sw}}$; in the absence of disorder the threshold $V_0$ corresponds to an energy of this order. Introducing an in-plane magnetic field to our calculations causes the small bias feature to develop into an asymmetric resonant peak in the tunnelling current, similar to that reported in [@Spielman2000on]. Below the threshold $V_0$, we find that the intrinsic dissipation causes any macroscopically coherent charge oscillations to decay in time, consistent with the absence of Josephson-like oscillations in experiments. We work in the pseudospin picture of the bilayer[@Moon1995]. The charge imbalance on the bilayer is given by the $z$-component of the magnetisation of the pseudospins. The system is a ferromagnet due to Coulomb exchange. Since a $z$ component in the magnetisation incurs a capacitative energy cost, the ferromagnet has easy-plane anisotropy. We use a model of discrete spins on a lattice, beginning with a spin-1/2 system in which each spin represents a local single-particle state within the lowest Landau level. We work with a large-$S$ version of this system. This large-$S$ generalisation may be treated as a coarse-graining procedure. The limit of large $S$ is the classical limit for the spin system. We work with the Hamiltonian: $$\frac{H}{2S} = - \frac{\rho_E}{2} \sum_{\langle ij \rangle} \vec{m}_i \cdot \vec{m}_j + \frac{D}{4} \sum_i ( m_i^z )^2 - \frac{\Delta_\mathrm{SAS}}{2} \sum_i m^x_i \label{equ:H_spins}$$ where $\vec{S}_i=S\vec{m}_i=S(m_i^x,m_i^y,m_i^z)$ is the spin operator on site $i$ of a square lattice with spacing $c_0=\sqrt{2\pi} l_B$ where $l_B=(\hbar c/eB)^{1/2}$ is the magnetic length. The exchange $\rho_E$ and the strength of the on-site repulsion $D$ were derived from microscopic considerations by Moon et al.[@Moon1995]. The tunnelling between the layers enters the problem through ${\Delta_\mathrm{SAS}}$: the splitting of the “bonding” and “anti-bonding” single-particle states in the double well. In our thought experiment, a gate is used to control the charge imbalance on the bilayer: this adds a term $H_V = - SV \sum_i m^z_i$ to the Hamiltonian. Typical values for the model parameters in physical bilayer systems are $l_B\simeq20\mathrm{nm}$, ${\Delta_\mathrm{SAS}}\simeq90\mu\mathrm{K}$, $\rho_E\simeq0.5\mathrm{K}$, $D\simeq 30\mathrm{K}$. We consider the properties of the ferromagnetic phase of this model, which is believed to describe the experimental systems. In particular, the observation of a linearly dispersing peak in the tunnelling conductance[@Spielman2000on] indicates that the experimentally accessible state has ferromagnetic order — the peak arises from the Goldstone mode of the system. We therefore ignore the quantum disordered phase that exists for $D/\rho_E S^2 \gg 1$ [@Chubukov1994]. We will now outline the dynamics of the classical ferromagnetic system before we discuss the quantum effects which are responsible for the existence of a dissipative tunnelling current. In the classical system, a spatially uniform spin configuration will remain uniform forever. The total magnetisation precesses along a trajectory of constant energy. We sketch the trajectories in Fig. \[fig:q0\_spins\] for ${\Delta_\mathrm{SAS}}\ll D$ which is the physically relevant regime. Near the ($x$-polarised) ground state, the precession frequency is given by ${\Delta_\mathrm{sw}}/\hbar$ where ${\Delta_\mathrm{sw}}= [{\Delta_\mathrm{SAS}}( {\Delta_\mathrm{SAS}}+ D )]^{1/2}$ is the energy gap for spinwave excitations (density waves of charge imbalance across the bilayer). As the magnetisation precesses along these trajectories, the charge imbalance on the bilayer oscillates around zero. Far away from the ground state, the spin precesses around one of the two maximal energy states which lie close to the $S^z$ axis. This yields a Josephson-like alternating current $I\simeq e{\Delta_\mathrm{SAS}}\cos (eVt/\hbar)$ where $V$ is the voltage across the bilayer due to capacitative charging. We stress that this is valid only for large charge imbalance (large $V$). Trajectories through the saddle point at $\vec{m}=(-1,0,0)$ mark the boundary between oscillations around the ground state and oscillations around the maximal energy states. The saddle point trajectory crosses the $xz$-plane at three points, as shown in Fig. \[fig:q0\_spins\]. The angle $\theta_0$ satisfies $\cos\theta_0=1-2({\Delta_\mathrm{SAS}}/D)$. It corresponds to a voltage difference of $V_0 = 2{\Delta_\mathrm{sw}}/e$ across the layers. In our thought experiment, we imagine using a gate to induce a uniform charge imbalance on the bilayer. In the spin picture, the magnetisation is tilted out of the easy ($xy$) plane. The bias is then instantaneously removed and the bilayer finds itself in a highly excited state. The resulting behaviour depends crucially on whether the initial gate voltage $V$ is above or below the saddle point value $V_0$. We treat the two regimes in separate perturbation theories. In both calculations, we expand around the classical limit using spinwave theory in a $1/S$ expansion. The leading terms in this expansion result in the leading terms in the $I$-$V$ relation. We begin by considering the situation for initial energies above the saddle point ($V>V_0$). When the gate voltage is removed, the charge imbalance results in a potential difference across the bilayer equal to the original gate voltage $V$. As discussed above, this causes an alternating current with a frequency $(eV/\hbar)$. In this regime, we can treat the tunnelling ${\Delta_\mathrm{SAS}}$ perturbatively. The quantum spin system is distinct from the classical one in that spatially uniform oscillations do not persist — the tunnelling term breaks the global spin rotation symmetry so that long-wavelength modes are no longer protected from decay by Goldstone’s theorem. The uniform mode decays by transferring energy into spin waves with finite wavevectors. The magnetisation falls to a lower trajectory on the spin sphere, corresponding to a net transfer of charge across the bilayer. We can therefore obtain the d.c. tunnelling current by calculating the rate of this dissipative process. It is straightforward to derive a bosonic spinwave theory for our model. We use the Holstein-Primakov representation: $S^x_j+iS^y_j=(2S-a^\dag_j a_j)^{1/2} a_j$, $S^z_j = S-a^\dag_j a_j$. Expanding around $m^z=(eV/D)=\sin\theta$, the quadratic part of the Hamiltonian is then easily diagonalised in the Fourier basis to give $$H^{(0)}_{\theta>\theta_0} = (D\sin\theta) \delta n_{q=0} + \sum_{\bm{q}\neq0} \varepsilon_{\bm{q}} \alpha_{\bm{q}}^\dag \alpha_{\bm{q}}$$ where the $\alpha_{\bm{q}}$ are bosonic operators describing the spinwave modes. The spinwave dispersion is given by $\varepsilon_{\bm{q}} = [\rho_E \gamma(\bm{q}) ( D + \rho_E \gamma(\bm{q}))]^{1/2}$ where $\gamma(\bm{q})=4-2\cos(q_x c_0) - 2\cos(q_y c_0)$. The dispersion is linear and gapless at small $q$. (Tunnelling should produce a small energy gap but this does not affect the perturbative calculation we describe here.) The spinwave velocity is $v=l_B (2\pi D\rho_E)^{1/2}$. Observe that the $q=0$ mode has been singled out in the Hamiltonian, and its energy is not given by the long wavelength limit of $\varepsilon_{\bm{q}}$. The quanta of this mode carry $S^z=1$ while spin waves with finite wavevector have $S^z=0$. Starting from an initial state $\|i\rangle$, the energy dissipation rate is given by $$\Gamma = \partial_t \Big< i \Big\| e^{iHt} \sum_{\bm{q}\neq0} \varepsilon_{\bm{q}} \alpha^\dag_{\bm{q}} \alpha_{\bm{q}} e^{-iHt} \Big\| i \Big>$$ The dissipation arises from the destruction of one quantum in the $q=0$ mode, and the generation of multiple spinwaves during tunnelling across the bilayer. To leading order in $1/S$, a pair of spin waves with opposite momenta is excited (Fig. \[fig:diag\_asp\]a). The relevant vertex is: $$H^{(1)}_{\theta>\theta_0} = (-{\tilde\Delta_\mathrm{SAS}}/8) e^{-i\phi_{q=0}} \sum_{\bm{q}} \gamma_{2,\bm{q}} \alpha^\dag_{\bm{q}} \alpha^\dag_{-\bm{q}} + \hbox{h.c.}$$ in which ${\tilde\Delta_\mathrm{SAS}}= {\Delta_\mathrm{SAS}}\exp({-\sqrt{D/\rho_E S^2}})$ is a renormalised tunnelling amplitude, and $\gamma_{2,\bm{q}} = \cos\theta [ (u_{\bm{q}}+v_{\bm{q}})^2 + 2\sin\theta\sec^2\theta- (u_{\bm{q}}-v_{\bm{q}})^2\sec^4\theta ] $ where $u_{\bm{q}}$ and $v_{\bm{q}}$ are coherence factors: $(u_{\bm{q}}+v_{\bm{q}})^2=(D+\rho_E\gamma(\bm{q}))/\varepsilon_{\bm{q}}$ with $u_{\bm{q}}^2- v_{\bm{q}}^2 =1$. The $\theta$-dependence arises because the $x$-component of the pseudospin depends on its angle with the easy plane, $\theta$, as well as on the azimuthal angle $\phi$. This dependence is weak for $\theta\ll 1$ when the charge imbalance is small compared to the Landau level filling. We now calculate the power dissipation $\Gamma$ for a given initial voltage $V$; the steady-state tunnelling current density at a bias $V$ is $I=\Gamma/VL^2 S$. The result is: $$\frac{\Gamma_{\theta>\theta_0}}{L^2 S} = I_{\theta>\theta_0}V = \frac{D {\tilde\Delta_\mathrm{SAS}}^2}{32 \pi \rho_E l_B^2\hbar S} \left[ 1 + X(\theta) \right]^2 \label{equ:I_asp}$$ where $X(\theta)=(\sec\theta-1)(\sec^2\theta-2\sec\theta-1)$ is small for a small charge imbalance ($\theta\ll 1$). Thus, the power dissipation into spinwave pairs is independent of the voltage, and so the tunnelling current has a $1/V$ divergence. This should be cut off at low bias when the bias $V$ becomes comparable to $V_0$ and perturbation theory breaks down. This contribution to the tunnelling current is significant at low temperatures: it is our main result, consistent with the experimental observation[@Spielman2000on] of a peak at a bias close to the spinwave gap. Its contribution to the tunnelling current density is shown in Fig. \[fig:iv\]. We note that the region of the experimental sample in which the tunnelling current flows remains an open question[@Balents2001]. Estimating the power dissipated in the experiments[@Spielman2000on] to be $10^{-16}W$, and neglecting renormalisation of the tunnelling (so $\tilde{\Delta_\mathrm{SAS}}={\Delta_\mathrm{SAS}}$), we obtain an estimate of $50\mu\mathrm{m}^2$ for this area. This estimate is increased if we use a renormalised ${\Delta_\mathrm{SAS}}$. The result is consistent with tunnelling taking place near the contacts to the bilayer, rather than over its entire area. In the experimental data of [@Spielman2000on], the current decays more slowly than $1/V$ away from the resonance, so that the power increases with increasing applied voltage. We attribute this increase to other dissipative channels, due to disorder or finite temperature. Processes at higher order in $1/S$ also affect the current in this way (see (\[equ:I\_BR\]) below), but are not large enough to explain the differences between theory and experiment. It should be noted that the dissipation rate (\[equ:I\_asp\]) is averaged over the period of the Josephson oscillations. As with other theories in which the tunnelling is treated perturbatively[@Balents2001; @bl_theory], we also expect an oscillatory a.c. component with frequency $eV/\hbar$. At this point, we make contact with previous calculations of the tunnelling current in [@Balents2001]. If we ignore the weak $\theta$-dependence of the vertex factor, the dynamics depend only on the azimuthal angle $\phi$ of the pseudospin. This is the same theory as in [@Balents2001]. However, that work gives a vanishing tunnelling current at zero temperature. Our calculation is different from that one since we include quantum corrections involving multi-spinwave processes. More quantitatively, the current is calculated in [@Balents2001] from $I_\phi \sim \int d^2{\bm{r}} dt \exp(ieVt/\hbar -G_Q)$ where the quantum propagator $G_Q$ is, in our notation: $$G_Q(\bm{r},t) = -i (v\hbar/2\pi S \rho_E) \left[ (vt-ia)^2 - r^2 \right]^{-1/2}\,.$$ We have explicitly included a lattice cutoff, $a\sim l_B$, that reflects the finite bandwidth of the spin waves. We recall that our parameters satisfy $(\hbar v/\rho_E S l_B) < 1$ to avoid quantum disordering. We can therefore perform the integral order by order in $(\hbar v/\rho_E S a)$, finding: $$I_\phi = \frac{ \exp(-eVa/\hbar v)}{V} \frac{D{\tilde\Delta_\mathrm{SAS}}^2} {16\pi \hbar l_B^2\rho_E} \frac{\mathrm{d}^2}{\mathrm{d}\Omega^2} I_0(2\sqrt{\Omega}) \label{equ:I_BR}$$ where $\Omega=eV/4\pi S \rho_E$ and $I_0$ is a modified Bessel function. The leading terms in the voltage $V$ are also the leading ones in $1/S$, justifying the expansion about the classical limit. In an expansion in $eV/\rho_E S$, the leading term in eq. (\[equ:I\_BR\]) coincides with eq. (\[equ:I\_asp\]) if we ignore $X(\theta)$. In fact, the $n$th term in the expansion arises from the decay of a single quantum of the $q=0$ mode into $n+2$ finite-momentum spinwaves. Thus, the zero-temperature tunnelling current arises from multi-spinwave processes. These contributions have been ignored in [@Balents2001] although they are, in principle, contained in the framework of that paper. We now generalise to the case in which a magnetic field is applied in the plane of the bilayer. In that case, the tunnelling term in the Hamiltonian acquires a spatial dependence: ${\Delta_\mathrm{SAS}}m^x\to {\Delta_\mathrm{SAS}}[m^x \cos(Qx) + m^y\sin(Qx)]$, where $Q=(e B_y d/\hbar c)^{1/2}$ is the wavelength associated with the in-plane field $B_y$. The spacing, $d$, between the two electron layers is approximately $30\mathrm{nm}$. The tunnelling vertex now involves a change of momentum $Q$ in the $x$-direction. A new process appears: a single $q=0$ quantum may decay into a single spinwave (Fig. \[fig:diag\_asp\]b). The contribution from this process leads to a $\delta$ function peak in the dissipation. The two-spinwave process contributes to the dissipation rate only for $eV > vQ$. The leading contributions to the current density are $$\begin{aligned} I_{\theta>\theta_0,Q} &=& \frac{D {\tilde\Delta_\mathrm{SAS}}^2}{16\pi\hbar l_B^2 } \times \Big[ 2\pi V^{-1} \delta(eV-vQ) + \nonumber\\ & & \phantom{hi} \;\frac{1}{2 \rho_E S} \frac{e}{\sqrt{(eV)^2-(vQ)^2}} \Theta(eV-vQ) \Big]\end{aligned}$$ where $\Theta(x)$ is the step function. The peak at $eV=vQ$ is asymmetric. The lineshape is controlled by the multi-spinwave processes. For $eV<vQ$, there is no available spinwave channel for dissipation and so there is no tunnelling current: at larger voltages, the current decays as a power law. This is shown in Fig. \[fig:iv\], in which we have broadened the delta function by introducing low-momentum scattering with momentum spread $\sigma_Q$. This may arise from disorder or thermal effects. In the absence of these effects, the delta function will be broadened by the intrinsic lifetimes of the final states (for example, by further tunnelling or spinwave/spinwave scattering). The treatment thus far has been valid only for voltages greater than $V_0=2{\Delta_\mathrm{sw}}/e$. We now discuss charging voltages smaller than this value. In this regime, our thought experiment does not result in a steady tunnelling current. Instead, the current and charge imbalance oscillate coherently around zero with frequency ${\Delta_\mathrm{sw}}/\hbar$. We now evaluate the rate at which these oscillations decay. We treat the parameter $\theta$ perturbatively in this regime. In the Holstein-Primakov representation, the quadratic part of the Hamiltonian takes the form: $$H_{\theta<\theta_0}^{(0)} = \sum_{\bm{q}} \omega_{\bm{q}} \beta^\dag_{\bm{q}} \beta_{\bm{q}}$$ where $\omega_q = \left[ ( {\Delta_\mathrm{SAS}}+ \rho_E \gamma(\bm{q}) )( D + {\Delta_\mathrm{SAS}}+ \rho_E \gamma(\bm{q}) ) \right]^{1/2}$ and $\beta_{\bm{q}}$ annihilates a spinwave with wavevector $\bm{q}$. The initial state of our thought experiment has a charge imbalance. This corresponds to a condensate of bosons with $q$=0. These bosons have the small concentration $n_0/S=\theta^2 [1+(D/{\Delta_\mathrm{SAS}})]^{1/2}/4\pi l_B^2$. The simplest dissipative process involves four $q$=0 quanta annihilating to form a pair of spinwaves with finite momenta. We defer details to a later paper[@Jack]: the dissipation rate is $$\frac{\Gamma_{\theta<\theta_0}}{L^2} = \left(\frac{\theta}{\theta_0}\right)^8\!\! \frac{{\Delta_\mathrm{SAS}}^3 {\Delta_\mathrm{sw}}^2 }{ 512\pi \rho_E D^2 l_B^2\hbar} f\left(\frac{{\Delta_\mathrm{SAS}}}{D}\right)^2$$ where $f(x)=(1+x)^{-4}[3-4x-8x^2-x\sqrt{1+16x+16x^2}]$ is approximately equal to 3 since ${{\Delta_\mathrm{SAS}}}\ll D$. This dissipation corresponds a decay of the density of $q=0$ modes: $\dot n = -\Gamma(\theta)/4\hbar{\Delta_\mathrm{sw}}L^2$. Since the density $n$ is proportional to $\theta^2$, we see that $d\theta^2/dt \sim -\theta^8$. The amplitude $A$ of the oscillations in the tunnelling current is proportional to $\sin\theta$, and so it decays in time as a power law: $A(t) \sim 1/t^{1/6}$. The dissipation is very weak at small initial angles partly due to the kinematic constraint for a 4-spinwave collision. One might expect stronger dissipation in the presence of disorder or inelastic scattering, e.g., with phonons or charged quasiparticles. Nevertheless, the multi-spinwave processes provide an intrinsic damping mechanism that has not previously been identified. We note that this treatment does not lead to a state with a direct tunnelling current at $V<V_0$. Therefore we cannot address the lineshape of the zero-bias peak in $\mathrm{d}I/\mathrm{d}V$. It is worth noting that this feature has sharpened with improved experimental conditions. It may be controlled by low-energy states introduced by disorder or by effects arising from coupling to external leads. An approach similar to [@Fertig2003] may be needed. In conclusion, we find that the zero-temperature tunnelling current has a strong $1/V$ peak arising from the decoherence of Josephson-like oscillations by the generation of electron density fluctuations in the bilayer. We also predict a non-trivial lineshape for the dispersing feature that appears in the presence of an in-plane field. These effects resemble features reported in experiments[@Spielman2000on]. We acknowledge the financial support of EPSRC (in part through GR/S61263/01) and of the Royal Society. [10]{} S. M. Girvin, A. H. MacDonald and J. P. Eisenstein, in Perspectives in Quantum Hall Effects, ed. S. Das Sarma and A. Pinczuk (Wiley, New York, 1997). I. B. Spielman [et al.]{}, Phys. Rev. Lett. [84]{}, 5808 (2000); Phys. Rev. Lett. [87]{}, 036803 (2001) K. Moon [et al.]{}, Phys. Rev. B [51]{}, 5138 (1995). L. Balents and L. Radzihovsky, Phys. Rev. Lett. [86]{}, 1825 (2001). H. A. Fertig and J. P. Straley, Phys. Rev. Lett. [91]{}, 046806 (2003). A. Stern [et al.]{}, Phys. Rev. Lett. [86]{}, 1829 (2001); M. Fogler and F. Wilczek, Phys. Rev. Lett. [86]{}, 1833 (2001); Y. N. Joglekar and A. H. MacDonald, Phys. Rev. Lett. [87]{}, 196802 (2001); M. Kellogg [et al.]{} Phys. Rev. Lett. [93]{}, 036801 (2004). X.-G. Wen and A. Zee, Phys. Rev. B [47]{}, 2265 (1993); Z. F. Ezawa and A. Iwazaki, Phys. Rev. B [48]{}, 15189 (1993). H. A. Fertig, Phys. Rev. B [40]{}, 1087 (1989) J. P. Eisenstein [et al.]{}, Phys. Rev. Lett. [68]{}, 1383 (1992); S. Q. Murphy [et al.]{}, Phys. Rev. Lett. [72]{}, 728 (1994). A. V. Chubukov, S. Sachdev, and J. Ye, Phys. Rev. B [49]{}, 11919 (1994). R. L. Jack, D. K. K. Lee and N. R. Cooper, in preparation	{ "pile_set_name": "ArXiv" }
--- abstract: 'A perturbative quantization procedure for Lie bialgebras is introduced and used to classify all three dimensional complex quantum algebras compatible with a given coproduct. The role of elements of the quantum universal enveloping algebra that, analogously to generators in Lie algebras, have a distinguished type of coproduct is discussed, and the relevance of a symmetrical basis in the universal enveloping algebra stressed. New quantizations of three dimensional solvable algebras, relevant for possible physical applications for their simplicity, are obtained and all already known related results recovered. Our results give a quantization of all existing three dimensional Lie algebras and reproduce, in the classical limit, the most relevant sector of the complete classification for real three dimensional Lie bialgebra structures given in [@gomez].' --- A. Ballesteros$^1$, E. Celeghini$^2$ and M.A. del Olmo$^3$ $^1$[Departamento de Física, Universidad de Burgos,\ E-09006, Burgos, Spain.]{}\ $^2$[Departimento di Fisica, Universitá di Firenze and INFN–Sezione di Firenze\ I50019 Sesto Fiorentino, Firenze, Italy]{}\ $^3$[Departamento de Física Teórica, Universidad de Valladolid,\ E-47011, Valladolid, Spain.]{}\ [e-mail: [email protected], [email protected], [email protected]]{} MSC: 81R50, 81R40, 17B37 0.4cm Keywords: Lie bialgebras, quantization, quantum algebras Introduction\[introduccion\] ============================ Cartan classification of semisimple Lie algebras has facilitated their applications in Physics. Quantum algebras are not been classified in a similar way and their physical applications are far to be developed or understanding. For these reasons, this paper deals with some facets of the problem of the construction and classification of quantum universal enveloping algebras (hereafter, quantum algebras) [@Dri]-[@Tjin]. It is well-known that any quantum algebra $(U_z(\mathfrak a),\Delta)$ with deformation parameter $z$ defines a unique Lie bialgebra structure $(\mathfrak a,\eta)$, a pair determined by the Lie algebra $\mathfrak a$ and a skew-symmetric linear map (cocommutator) $\eta:\mathfrak a \to \mathfrak a\times \mathfrak a$ . Such cocommutator $\eta$ is defined as the first order skewsymmetric part of the coproduct $$\label{bialgebra} \eta (X)=\frac 1 2 (\Delta (X)-\sigma\circ \Delta (X))\quad \mbox{mod}\,z^2\qquad\forall X\in \mathfrak a$$ where $ \sigma$ is the flip operator $ \sigma (X\otimes Y)=Y\otimes X$. Therefore, quantum deformations of a given Lie algebra $\mathfrak a$ can be classified according to this “semiclassical" limit. Moreover, we recall that Lie bialgebras are in one to one correspondence with Poisson-Lie structures on the group $Lie(\mathfrak a)$ [@DriPL] that arise again as the first order in $z$ of the quantum groups dual to the quantum algebras $(U_z(\mathfrak a),\Delta)$. With this in mind, some classifications of Lie bialgebra structures for several physically relevant Lie algebras have been obtained. For the three dimensional (3D) case we will refer to the full classification of Lie bialgebras given in [@gomez] and for some higher dimensional Lie bialgebras we refer to [@galileo] and references therein. However, it is clear that the inverse problem has also to be faced: that is, to obtain general recipes for the “quantization" (i.e. the associated quantum algebra) of a given Lie bialgebra $(\mathfrak a,\eta)$. Although the existence of such an object is indeed guaranteed (see [@charipressley], Chapter 6), only for coboundary triangular bialgebras the Drinfel’d twist operator gives rise to the associated quantum algebra [@majid]. However, quasitriangular and even non-coboundary Lie bialgebras do exist and for them the twist operator approach to quantization is not available. We recall that some attempts have been performed in order to get structural properties of the quantization of arbitrary Lie bialgebras (see [@LM; @mudrov] for a prescription to get the quantum coproduct –but not the deformed commutation rules– for a wide class of examples). Moreover, to our knowledge, a general investigation concerning the uniqueness of this Lie bialgebra quantization process has not been given yet and only restrictive results for certain deformations of simple Lie algebras have been obtained (see [@schnider], Chapter 11). As a consequence of the abovementioned facts, a complete classification of quantum algebras in the spirit of the Cartan classification for Lie algebras is still far to be reached. In this paper we present a “direct approach" to the quantization problem, together with an operational procedure [à la Cartan]{} (in a sense that will be made more explicit in the sequel) for the classification of quantum algebras. The essential ingredients of the quantization method are described in Section 2. In particular, given a Lie bialgebra we shall firstly obtain a coassociative quantum coproduct with some outstanding symmetry properties and, afterwards, we shall solve order by order the compatibility equations for the deformed commutation rules. Lie-Poisson brackets are immediately recover using a completely symmetrized basis for $U_z(\mathfrak a)$ non previously considered in the literature. In Section 3 this approach will be explicitly developed for a relevant type of cocommutator $\eta$ that is compatible with all the non-isomophic 3D complex Lie algebras, as described -for instance- in Jacobson [@Jac]. This cocommutator gives rise to a coproduct very simple and deformed commutators that are not specially complex. In this way we obtain three different families of quantum algebras whose deformed commutation rules are very general and depend on several structure constants. Another point to stress is that our quantization method has been made using a symmetrized basis instead the usual Poincaré-Birkhoff-Witt (PBW) basis. This procedure has not been previously considered in the literature and can be relevant in quantum physics applications and in the semiclassical limit. In order to classify such quantizations, in Section 4 we consider the equivalence transformations on the basis of the quantum algebra that are defined through invertible maps in $U_z(\mathfrak a)$ that leave formally invariant the coproduct for the elements of the basis. This constraint regarding the formal invariance of the coproduct can be understood as a way to identify certain basis elements as some sort of “generators" of a quantum algebra, thus following for quantum algebras the “Cartan" approach to the classification of Lie algebras within $U(\mathfrak a)$, where only linear transformations in the space of generators leaving invariant their (primitive) coproduct are performed. After a such classification is systematically developed, new quantizations of 3D solvable algebras are obtained, and already known results are recovered. The above mentioned simplicity of the algebraic and coalgebraic structures of this new quantizations increases notably their possible interest in Physics . In order to make more precise the range of 3D deformations that have been covered, a detailed comparison with the complete classification of 3D real Lie bialgebra structures presented in [@gomez] is explicitly given within a final Section which also includes some further comments and conclusions. The quantization method ======================= \[generalquantization\] Let us consider the Lie algebra $\mathfrak a$ and its universal enveloping algebra $U(\mathfrak a)$ [@Jac], an associative algebra that is obtained as the quotient $T(\mathfrak a)/I$, where $T(\mathfrak a)$ is the tensor algebra of $\mathfrak a$ and $I$ is the ideal generated by the elements $XY-YX-[X,Y]$ ($X,Y\,\in \mathfrak a$). If we define the coproduct, counit and antipode ($\forall\, X \,\in \mathfrak a$) $$\co_0(X)=1\otimes X + X\otimes 1,\quad \co_0(1)=1\otimes 1,\quad \epsilon(X)=0,\quad \epsilon(1)=1,\quad \gamma(X)=-X,$$ and we extend by linearity all these (anti)automorphisms to the full $U(\mathfrak a)$, we shall endow the universal enveloping algebra with a Hopf algebra structure. In general, an element $Y$ of a Hopf algebra is called [primitive]{} if $$\co(Y)=1\otimes Y + Y\otimes 1.$$ Within $U(\mathfrak a)$, it can be shown that the only primitive generators under the coproduct $\co_0$ are the generators of $\mathfrak a$ (this result is known as the Fiedrichs theorem [@Post]). Note that an essential property of the Hopf algebra $U(\mathfrak a)$ is its cocommutativity, since $\co_0$ is invariant under the action of the flip operator $\sigma$. Let us now consider the Hopf algebra $U( \mathfrak a)$ and a deformation parameter $z$. A quantum algebra $(U_z(\mathfrak a),\co)$ is a Hopf algebra of formal power series in the deformation parameter $z$ with coefficients in $U( \mathfrak a)$ and such that [@charipressley] $$U_z(\mathfrak a)/zU_z(\mathfrak a)\simeq U(\mathfrak a).$$ Therefore, $U(\mathfrak a)$ is obtained (as Hopf algebra) in the limit $z\to 0$, and the first order in $z$ of the coproduct is directly related to the cocommutator of an underlying Lie bialgebra $(\mathfrak a,\eta)$ through (\[bialgebra\]). In this way, Lie bialgebras can be used to characterize quantum deformations. Amongst all the Hopf algebra axioms to be imposed, we recall that the quantum coproduct $\Delta$ has to be a coassociative map, namely: $$\label{coassociative} (\Delta\otimes id)\circ\Delta=(id\otimes \Delta)\circ\Delta.$$ From now on, we shall refer to the “quantization" of a given Lie bialgebra $(\mathfrak a,\eta)$ as the problem of finding a quantum algebra $(U_z(\mathfrak a),\co)$ such that (\[bialgebra\]) is fulfilled (we recall that the uniqueness of such construction cannot be take for granted). The general quantization procedure that we propose is based on three essential ingredients that can be considered whatever the dimension of the Lie bialgebra is. The first one is a generalized cocommutativity property for the coproduct, the second one is related to the choice of a basis in the universal enveloping algebra and the third one selects a given type of power series expansion for the deformed commutation rules. [1. Generalized cocommutativity]{}. In general, we impose the (noncocommutative) quantum coproduct $\Delta$ to be invariant under the composition $\tilde \sigma=\sigma\circ T$ of the flip operator $ \sigma $ and a change of sign of (all) the deformation parameter(s):$$\tilde \sigma \circ \Delta = \Delta \qquad \mbox{where}\quad \tilde \sigma=T\circ \sigma \quad \mbox{and}\quad T(z)=-z.$$ We point out that the definition (\[bialgebra\]) for the underlying cocommutator implies that the deformation parameter appear explicitly as multiplicative factors within the cocommutator. This symmetry property of the coproduct can be imposed in any dimension and makes much easier the procedures of symmetrization and ‘hermitation’ [@enrico03]. In particular, given a certain Lie bialgebra $(\mathfrak a,\eta)$, the above assumption implies that the first order deformation of the coproduct will be just given by the (skewsymmetric) Lie bialgebra cocommutator. $$\label{first} \Delta(X)=\Delta_0 (X) + \eta(X) + O[z^2], \qquad X\in\mathfrak{g}.$$ Moreover, the invariance of $\Delta$ under the transformation $\tilde \sigma$ together with the fact that the coproduct is an algebra homomorphism with respect to the deformed commutation rules $[\cdot,\cdot]_z$, i. e. $\Delta ([\cdot,\cdot]_z)=[\Delta (\cdot), \Delta (\cdot)]_z$ implies that $[\cdot,\cdot]_z$ has to be an even function in the deformation parameter $z$. [2. The choice of a basis in $U_z(\mathfrak a)$]{}. In contradistinction with previous works on this subject in which the PBW basis $X_1^\alpha\,X_2^\beta\,\dots\,X_l^\zeta$ is considered, we introduce using the operator $\rm Sym$ a basis in $U_z(\mathfrak a)$ given by the completely symmetrized monomials. We, thus, define the linear operator ${\rm Sym}$ by $${\rm Sym}\;\left\{\sum c_{\alpha\beta\dots\zeta}\; X_1^\alpha\,X_2^\beta\,\dots\,X_l^\zeta\right\} = \sum c_{\alpha\beta\dots\zeta}\, {\rm Sym}\,\{X_1^\alpha\,X_2^\beta\,\dots\,X_l^\zeta\}$$ $${\rm Sym}\;\{A_1\dots A_n\} := \frac 1{n!}\sum_{p\in {\rm S_n}} p (A_1\dots A_n),$$ with ${\rm S_n}$ the group of permutations of n elements . Note that Sym is the identity for commuting operators. This symmetrization Ansatz, although very convenient in quantum mechanical terms, has not been previously considered in the literature, and turns out to be very efficient in order to get the explicit form of the deformed commutation rules. We remark that one of the main advantages of the symmetrized basis in the quantum algebra is the fact that if we replace the deformed commutation rules by Poisson brackets, the correspondent Poisson-Hopf algebra is uniquely and immediately defined. [3. Deformed commutation rules]{}. We will assume $$\label{hipotesis3} [X,Y]_z = \frac{1}{z} {\rm Sym}\;(f( zX_1,zX_2,\dots ,zX_l)),$$ where $f$ is a meromorphic function at $z=0$ and also odd in $z$; We extend the discussion to the multiparametric case considering meromorphic deformations of Lie algebras with $z_i/z_j$ fixed. Under all these assumptions, given a (family of) Lie bialgebra $(\mathfrak a,\eta)$, the “direct" quantization procedure that we propose would be sketched as follows: - Assume that the first order coproduct is of the form (\[first\]). - Order by order in the deformation parameter(s), get the relations coming from the coassociativity constraint (\[coassociative\]) and solve them recursively by taking into account the invariance under $\tilde\sigma$ of the solution, thus obtaining the full quantum coproduct. - Obtain, again order by order, the deformed commutation rules by solving the compatibility equations coming from the fact that the coproduct has to be an algebra homomorphism. Three dimensional quantum algebras ================================== Let $\{ A,B,C\}$ be the generators of an arbitrary complex 3D Lie algebra ${\cal L}$ with commutators $$\label{conmutaciones} \begin{array}{lll} [A,B] &=& c_1 A + c_2 B + c_3 C \\[0.3cm] [A,C] &=& b_1 A + b_2 B + b_3 C \\[0.3cm] [B,C] &=& a_1 A + a_2 B + a_3 C , \end{array}$$ where the structure constants are complex numbers subjected to some (nonlinear) relations coming from Jacobi identity. We recall that the complete classification of the 3D complex Lie algebras is given in [@Jac] (see for instance [@pavel] for the real case). According to the dimension of the derived algebra ${\cal L}'=[{\cal L},{\cal L}]$, the nonisomorphic classes of 3D Lie algebras read: - Type I: ${\rm dim}\;{\cal L}'=0$. Then ${\cal L}$ is abelian. We shall not consider this case from the point of view of quantum deformations, since any coassociative coalgebra with primitive non-deformed limit is compatible with the abelian commutation rules. - Type II: ${\rm dim}\;{\cal L}'=1$. We have two algebras, the Heisenberg-Weyl algebra and a central extension ${\cal L}={\cal B}\oplus{\cal C}$ of the Borel algebra, where $ {\cal B}$ is a Borel algebra and ${\cal C}$ commutes with ${\cal B}$. - Type III: ${\rm dim}\;{\cal L}'=2$. We have the family on non-isomorphic Lie algebras labeled by the nonzero complex number $\alpha$ with commutators $$\label{jacobson1} [X_1,X_2]=0,\qquad [X_1,X_3]=X_1,\qquad [X_2,X_3]=\alpha X_2 ,$$ and the Lie algebra $$\label{32} [X_1,X_2]=0,\qquad [X_1,X_3]=X_1+X_2,\qquad [X_2,X_3]= X_2 .$$ - Type IV: ${\rm dim}\;{\cal L}'=3$. The only element in this class is the simple Lie algebra ${\cal A}_1$. Note that in the Type III the algebra of dilations and the 2D Euclidean algebra for $\alpha=1$ and $\alpha=-1$, respectively, are included. A complete classification of the Lie bialgebra structures of the [real]{} 3D Lie algebras has been given in [@gomez], and the Lie bialgebra classification for the complex case can be extracted from there. Obviously, once a set of values for the structure constants $\{a_i,b_i,c_i\}$ is given, a suitable linear transformation $X_i=X_i(A,B,C)$ with complex coefficients can be found in such a way that the Lie algebra (\[conmutaciones\]) is reduced to one of the Jacobson cases. However, while in (\[conmutaciones\]) all $X_i$ are primitive and, thus, equivalent, this is not longer true in presence of deformation. In particular, by looking at the complexified form of the classification of 3D Lie bialgebras given in [@gomez], one can realize that a cocommutator of the form $$\label{coconmutador} \eta(A)=0 \qquad \eta(B)= z\, A \wedge B , \qquad \eta(C)=\rho\,z\, A \wedge C , \qquad z,\rho\in \mathbb C$$ defines a Lie bialgebra structure for each of the types of Lie algebras given in Jacobson’s classification provided that different linear transformations $X_i=X_i(A,B,C)$ are defined and that $\rho$ takes certain appropriate values. Since we are interested in obtaining the most general types of deformed commutation rules arising in 3D quantizations, we shall apply the perturbative quantization procedure described in the previous Section in order to get a quantum coproduct coming from the cocommutator (\[coconmutador\]) together with a compatible deformation of the commutation rules (\[conmutaciones\]). We remark that the cocommutator (\[coconmutador\]) can be thought of as a two parameter structure in $z$ and $\chi=\rho\,z$. Moreover, complex values for all the parameters (including the structure constants $\{a_i,b_i,c_i\}$ and $\rho$) will be considered, and the results here presented will also contain the quantizations in which the deformation parameter is a root of unity. Let us follow step by step the procedure introduced in Section 2. Firstly, we assume that the quantum coproduct will be of the form (\[first\]), namely $$\label{assumption} \begin{array}{lll} \Delta(A) &=& 1\otimes A + A \otimes 1 , \\[0.3cm] \Delta(B) &=& 1\otimes B + B \otimes 1 + z\, A \wedge B + O[z^2], \\[0.3cm] \Delta(C) &=& 1\otimes C + C \otimes 1 + \rho\,z\, A \wedge C + O[z^2]. \end{array}$$ In this case, it is straightforward to prove that the following coassociative coproduct has a first order given by the cocommutators (\[coconmutador\]) and is invariant under the transformation $\tilde\sigma$: $$\label{coproductos} \begin{array}{lll} \Delta(A) &=& 1\otimes A + A \otimes 1 , \\[0.3cm] \Delta(B) &=& e^{z A}\otimes B + B \otimes e^{-z A} , \\[0.3cm] \Delta(C) &=& e^{ \rho\, z A}\otimes C + C \otimes e^{- \rho\, z A}, \end{array}$$ Now we have to obtain the deformed commutation rules by solving the compatibility equations coming from the fact that the coproduct has to be an algebra homomorphism. Since the $\tilde\sigma$ invariance of the coproduct implies that the first deformed term in the commutation rules has to be of the order $z^2$, the first order coproduct (\[assumption\]) has to be compatible with the [non deformed]{} commutation rules. This condition leads to the following relations between the structure constants and $\rho$: $$\label{constantesp} b_2 (1-\rho )=0,\quad c_3 (1-\rho)=0, \quad b_1=-\rho\,a_2,\quad a_3=\rho\,c_1,\quad b_3 = - \rho\, c_2.$$ Jacobi identity requires the relation $$\label{jacobi0} (1-\rho) [a_2 c_1(1+\rho) -a_1 c_2] =0$$ Now we can distinguish the following cases: 1. Case $\rho \neq \pm 1$: $$\begin{array}{c}\label{constantesp0} b_2 =0,\qquad c_3=0, \quad b_1=-\rho\,a_2,\quad a_3=\rho\,c_1,\quad b_3 = - \rho\, c_2 ; \\[0.3cm] a_2 c_1(1+\rho) -a_1 c_2=0. \end{array}$$ Note that the space of parameters is 3D 2. Case $\rho =+1$: $$\label{constantesp1} b_1=-\,a_2,\quad a_3=\,c_1,\quad b_3 = - \, c_2.$$ In this case the space of parameters is 6D. 3. Case $\rho =-1$: $$\begin{array}{c}\label{constantesm1} b_2=0,\qquad c_3=0 , \qquad b_1=\,a_2,\quad a_3=-\,c_1,\quad b_3 = \, c_2 ; \\[0.3cm] a_1 c_2=0 . \end{array}$$ Here, the space of parameters is 3D. In all the cases no more conditions are found between the structure constants to higher orders, in spite of the fact that we have assumed no dependence on the deformation parameter of the structure constants. Note that the order by order procedure has to be solved, in general, simultaneously for both the quantum coproduct and the deformed commutation rules. Now, we are in conditions to obtain the deformed commutators. So, the integration of the above equations to all orders gives the general $q$–algebra of three generators compatible with the deformed coproduct. We find: 1. Case $\rho\neq \pm1$: - $c_2\neq 0$: $$\label{zconmutaciones01} \begin{array}{lll} [A,B] &=& c_1 \,\sinh (zA)/z +c_2\, B ,\\[0.3cm] [A,C] &=& - a_2\,\sinh (z\,\rho\, A)/z - \rho\,c_2 \,C ,\\[0.3cm] [B,C] &=& \displaystyle \frac{a_2 \, c_1}{c_2} \frac{\sinh [z(1+\rho) A]}{z} \\[0.3cm] && \qquad+a_2\, {\rm Sym}\;[B\cosh (z\,\rho\, A)] + \rho\,c_1\, {\rm Sym}\;[C\cosh (z A)] . \end{array}$$ - $c_2=0,\; a_2=0$: $$\label{zconmutaciones02} \begin{array}{lll} [A,B] &=& c_1 \,\sinh (zA)/z ,\\[0.3cm] [A,C] &=& 0 ,\\[0.3cm] [B,C] &=&\displaystyle a_1\,\frac{\sinh [z(1+\rho)\,A]}{z(1+\rho)} + \rho\,c_1\, C \,\cosh (zA) . \end{array}$$ $(c_2=0,\; a_2=0)$. 2. Case $\rho =+1$: - $$\label{zconmutaciones1} \begin{array}{lll} [A,B] &=& c_1 \,\sinh (zA)/z + c_2\, B + c_3 \,C ,\\[0.3cm] [A,C] &=& -a_2 \,\sinh (zA)/z + b_2 \,B - c_2 \,C ,\\[0.3cm] [B,C] &=& a_1\, \sinh (2zA)/(2z) + {\rm Sym}\;\{(a_2\, B + c_1\, C) \,\cosh (zA)\} \end{array}$$ 3. Case $\rho =-1$: - $c_2\neq 0$: $$\label{zconmutaciones2} \begin{array}{lll} [A,B] &=& c_1\,\sinh (zA)/z + c_2\, B ,\\[0.3cm] [A,C] &=& b_1 \,\sinh (zA)/z + c_2 \,C ,\\[0.3cm] [B,C] &=& {\rm Sym}\;\{(b_1\, B - c_1\, C) \,\cosh (zA)\} . \end{array}$$ - $c_2=0$: $$\label{zconmutaciones3} \begin{array}{lll} [A,B] &=& c_1 \,\sinh (zA)/z ,\\[0.3cm] [A,C] &=& b_1 \,\sinh (zA)/z ,\\[0.3cm] [B,C] &=& a_1\, A +{\rm Sym}\;\{(b_1\, B - c_1\, C) \,\cosh (zA)\} . \end{array}$$ It is worthy to notice that in many of these case there is a form invariance of the commutators related with the interchange of the generators $B$ and $C$. Equivalence and classification ============================== \[equivalence\] In general, two quantum algebras are said to be equivalent (isomorphic) if there exists an invertible (nonlinear in many cases) map between their corresponding quantum universal enveloping algebras as Hopf algebras. In this way, equivalence classes of quantum deformations can be defined. But it is clear that, due to the infinite number of possibilities given by arbitrary nonlinear maps, such equivalence classes are huge, and presumably some general criteria for the choice of “canonical" representatives of each of them should be helpful for classification purposes. Moreover, the definition of a such canonical representative for the Hopf algebra structure would be given, indeed, in terms of what could be properly called as “generators" of the quantum algebra. In fact, the standard classification of (non deformed) Lie algebras can be understood as a particular application of the abovementioned procedure to their corresponding universal enveloping algebras as Hopf algebras, since Friedrichs theorem states that the only primitive elements in $U(\mathfrak a)$ under the coproduct $\co_0$ are just the generators of the Lie algebra $\mathfrak a$. Therefore, we can define the generators of a Lie algebra $\mathfrak a$ as those elements of $U(\mathfrak a)$ which have a primitive coproduct. In this way, the (Cartan) classification of Lie algebras is performed by obtaining appropriate generators (primitive elements) having the “simplest" commutation rules (minimum number of non-vanishing structure constants). We stress that, in order to find such “irreducible" commutation rules only equivalence transformations leaving the coproduct invariant are allowed (in the Lie case, these are just linear transformations). From this perspective, we propose a definition of the “canonical" representatives of quantum algebras by following a similar procedure. In this case is essential to realize that a quantum algebra $U_z(\mathfrak a)$ is endowed with a deformed coproduct $\Delta$ which is no longer cocommutative, but $\Delta$ is (through our quantization procedure) invariant under $\tilde{\sigma}$ (generalized cocommutativity). Thus, in order to find the “canonical" generators we shall move within the equivalence subclass defined through the restricted set of Hopf algebra isomorphisms of quantum universal enveloping algebras (always in a symmetrized basis) that leave the coproduct formally invariant. Through such restricted isomorphisms we shall look for representatives with “irreducible" deformed commutation rules having a minimal number of non-zero terms. By proceeding in this way we have succeeded in classifying all the non-isomorphic quantum algebras that are contained in the three multiparameter families given in the previous section. Let us explicitly obtain them by eliminating many irrelevant parameters through coproduct-preserving mappings. Case 1: $\rho \neq \pm 1$ ------------------------- Let us start with the case $\rho \neq \pm1$. Let us consider the following transformation [@dobrev]: $$\begin{array}{l}\label{cambio1} \displaystyle \A=\alpha A,\quad \B= \beta B + \delta \frac{\sinh (A)}{z} ,\quad \C= \nu C +\eta \frac{\sinh (z\,\rho\,A)}{z\,\rho} ,\\[0.3cm] \z= \alpha ^{-1} z ,\qquad \alpha ,\; \beta ,\; \delta ,\; \nu ,\; \eta \in \mathbb C. \end{array}$$ After this transformation the coproduct (\[coproductos\]) becomes $$\label{coproductos1} \begin{array}{lll} \Delta(\A) &=& 1\otimes \A + \A \otimes 1 , \\[0.3cm] \Delta(\B) &=& e^{\z \A}\otimes \B + \B \otimes e^{-\z \A} , \\[0.3cm] \Delta(\C) &=& e^{ \z\rho \A}\otimes \C + \C \otimes e^{- \z\rho \A} , \end{array}$$ i.e., it remains formally invariant. So, all the elements of the form (\[cambio1\]) belong of the same class. - $c_2 \neq 0$ In this case the above mentioned change of basis (\[cambio1\]) can be reduced to $$\label{cambio11} \begin{array}{l}\A= A / c_2,\qquad \B=c_2 B + c_1 \frac{\sinh (zA)}{z} ,\qquad \C= c_2 C + a_2 \frac{\sinh (z\rho A)}{z\rho} ,\\[0.3cm] \z= c_2 z. \end{array}$$ - Under this change of basis we obtain the new Lie commutators $$\label{algebra1.1.1} [\A,\B] = \B ,\qquad [\A,\C] = - \,\rho\, \C ,\qquad [\B,\C] = 0 ,$$ which correspond to a quantization of the Lie algebra (\[jacobson1\]). Its associated bialgebra is non-coboundary, i.e., there is no classical-$r$ matrix. - $c_2 = 0$ We have two quantum algebras. Following a procedure analogous to (\[cambio11\]) they can be written as - $a_2=0, \; c_1 \neq 0$ $$\label{algebra1.1.2} [\A,\B] = \sinh (\z\A)/\z ,\qquad [\A,\C] = 0 ,\qquad [\B,\C] = \rho\, \C \,\cosh (\z\A) .$$ The bialgebra is coboundary: the classical $r$-matrix is $r=\hat z\,\A\wedge \B$. It is non-standard, i.e. it verifies the classical Yang-Baxter equation. This is another quantization of the Lie algebra (\[jacobson1\]). - $a_2=0, \; a_1 \neq 0 ,\; c_1 = 0$ $$\label{algebra1.2.2} [\A,\B] = 0 ,\qquad [\A,\C] = 0 ,\qquad [\B,\C] =\displaystyle \frac{\sinh [\z (1+\rho) \A]}{ \z(1+\rho)} .$$ This is also a coboundary deformation with standard $r$-matrix, i.e. it verifies the modified classical Yang-Baxter equation, $r=\hat z \B\wedge\C$. Note that for $\rho=0$ we have obtained two deformations of the extended Borel algebra and one deformation of the Heisenberg-Weyl algebra, both of them of Type II in Jacobson. Case 2: $\rho =+1$ ------------------ The equivalence classes are defined by applying to (\[zconmutaciones1\]) the transformation: $$\begin{array}{l}\label{cambio2} \A=\alpha A,\qquad \B= \beta B + \gamma C +\delta \frac{\sinh (zA)}{z} ,\qquad \C= \mu B + \nu C +\eta \frac{\sinh (zA)}{z} ,\\[0.3cm] \z= \alpha ^{-1} z \qquad\qquad \alpha ,\; \beta ,\; \gamma,\; \delta ,\; \mu ,\; \nu ,\; \eta \in \mathbb C. \end{array}$$ This transformation allows us to distinguish the quantum algebras characterized by $b_2 c_3 + {c_2}^2 \neq 0$ and those in which $b_2 c_3 + {c_2}^2 = 0$. - $b_2 c_3 + {c_2}^2 \neq 0$ - $$\label{algebra2.1.1} [\A, \B] =\B ,\qquad [\A, \C] = -\C ,\qquad [\B, \C ]=\frac{\sinh (2 \z \A)}{2 \z} .$$ This quantum algebra is just ${\cal A}_1(q)$. The classical $r$-matrix is $r= z\,\B\wedge \C$, and is standard. - $$\label{algebra2.1.2} [\A, \B] =\B ,\qquad [\A, \C] = -\C ,\qquad [\B, \C ] = 0 .$$ This is the complexification of the first discovered contraction of $su_q(2)$ deformation of the Euclidean algebra ${\cal E}(2)$ [@rusos] in two dimensions. It is a non-coboundary one. - $b_2 c_3 + {c_2}^2 = 0$ - $a_2 c_2+b_2 c_1\neq 0$ $$\label{algebra2.2.1} [\A, \B] = - \frac{\sinh(\z\A)}{\z},\qquad [\A, \C] = \B,\qquad [\B, \C] =- {\rm Sym} \{\C \cosh(\z \A)\}$$ This is the symmetrized version of the well-known Jordanian deformation of ${\cal A}(1)$ [@ohn] with non-standard classical $r$-matrix $r= \z\, \A\wedge \B$. According to the commutation relations (\[algebra2.2.1\]) we obtain that $${\rm Sym}(\C\,\cosh(\z\A))=\frac12(\C\,\cosh(\z\A)+\cosh(\z\A)\,\C))+ \frac{1}{12}\,\z^2\,\sinh\frac{2\z\A}{2\z},$$ thus, the quantum algebra presented in [@ohn] is a case of (\[zconmutaciones1\]) with $z$-dependent parameters. - $a_2 c_2+b_2 c_1=0$ This condition implies that the commutators $[\A,\B]$ and $[\A,\C]$ are proportional. We obtain the following algebras: - $$\label{algebra2.2.2} [\A, \B] = \frac{\sinh(\z\A)}{\z},\qquad [\A, \C] = 0,\qquad [\B, \C] = \C \cosh(\z \A)$$ We recover a non-standard deformation of the Euclidean group in two dimensions $E(2)$ [@angel95]. The classical $r$-matrix is non-standard, $r= \z\A\wedge \B$. - $$\label{algebra2.2.3} [\A, \B] = 0,\qquad [\A, \C] = -\B,\qquad [\B, \C] = \frac{\sinh(2\z\A)}{2 \z}$$ We have the standard deformation of $E(2)$ with classical $r$-matrix $r= \z \,\B\wedge \C$. - $$\label{algebra2.2.4} [\A, \B] = 0 ,\qquad [\A, \C] = 0 ,\qquad [\B, \C] = \frac{\sinh(2\z\A)}{2 \z}$$ It corresponds to a deformation of the Heisenberg-Weyl algebra with classical $r$-matrix is standard, $r=\z \,\C\wedge \B$. - $$\label{algebra2.2.5} [\A,\B] = 0 ,\qquad [\A,\C] = \B ,\qquad [\B, \C] = 0$$ This is a non-coboundary deformation of the Heisenberg-Weyl algebra. Case 3: $\rho =-1$ ------------------ The classification of the quantum algebras corresponding to the case $\rho =-1$ can be made considering the transformation: $$\label{cambio3} \A=\alpha A,\qquad \B=\beta B + \delta \frac{\sinh (zA)}{z} ,\qquad \C= \nu C +\eta \frac{\sinh (zA)}{z} ,\qquad \z= \alpha ^{-1} z ,$$ where $\alpha ,\;\beta ,\; \delta ,\; \nu ,\,\eta$ are complex numbers. - $c_2\neq 0$ $$\label{algebra3.1.1} [\A ,\B]=\B, \qquad [\A, \C]=\C, \qquad [ \B,\C]=0 .$$ We have a deformation of the dilations algebra such that there is not $r$-matrix. - $c_2= 0$: - $$\label{algebra3.2.1} \begin{array}{lll} [\A,\B] &=& - \,\sinh (\z\A)/\z ,\\[0.3cm] [\A,\C] &=& \,\sinh (\z\A)/\z ,\\[0.3cm] [\B,\C] &=& \, \A + ( \B + \C) \,\cosh (\z\A) . \end{array}$$ Like in the previous case there is not $r$-matrix. Note that in the limit of $\z \to 0$ we recover $$\label{algebra3.2.1limite} [\A,\B] = - \A ,\qquad [\A,\C] = \A ,\qquad [\B,\C] = \, \A + ( \B + \C) .$$ that can be rewritten under a change ($ \B+\C \to \B $) like $$\label{algebra3.2.1limitec} [\A,\B] = 0 ,\qquad [\A,\C] = \A ,\qquad [\B,\C] = \, \A + \B .$$ Therefore we have obtained a quantum deformation of the Lie algebra (\[32\]). - Other deformation of (\[32\]), now non-standard, is $$\label{algebra3.2.2} [\A,\B] = 0 ,\qquad [\A,\C] = \sinh (\z\A)/\z ,\qquad [\B,\C] = \A + \B \,\cosh (\z\A) .$$ The $r$-matrix is $r= \z\,\A\wedge \C$. - $$\label{algebra3.2.3} \begin{array}{lll} [\A,\B] &=& - \sinh (\z\A)/\z ,\\[0.3cm] [\A,\C] &=& \sinh (\z\A)/\z ,\\[0.3cm] [\B,\C] &=& ( \B + \C) \,\cosh (\z\A). \end{array}$$ This is a deformation of the dilation algebra in two dimensions without $r$-matrix. - $$\label{algebra3.2.4} [\A,\B] = 0 ,\qquad [\A,\C] = 0 ,\qquad [\B,\C] = \A .$$ It corresponds to a coboundary deformation of the Heisenberg-Weyl algebra with standard $r$-matrix, $r= \z\,\B\wedge \C$. - $$\label{algebra3.2.5} [\A,\B] = -\sinh (\z\A)/\z ,\qquad [\A,\C] = 0 \, ,\qquad [\B,\C] = \C \,\cosh (\z\A) .$$ We have a deformation of the dilatation algebra in two dimensions. In this case the classical $r$-matrix is non-standard, $r= \z \A\,\wedge \B$. Conclusions and remarks ======================= \[conclusiones\] This quantization method can be simultaneously applied and successfully solved for a multiparameter family of Lie bialgebras that share some structural properties. Throughout the paper we have considered certain 3D complex Lie bialgebras. In particular, the parameter $\rho$ is complex but we do also have isolated solutions for $\rho=\pm 1$. On the other hand, the comparison with the complete classification of 3D real Lie bialgebras given in [@gomez] can be worked out by considering the isomorphisms among the complexified versions of 3D real Lie algebras. In this way it can be shown that all the quantum algebras that we have obtained in section 4 are quantizations of the complexifications of the dual version of the Lie bialgebras given in [@gomez]. In order to find out the correspondence explicitly, in Table III of [@gomez] we have to identify a given algebra $\mathfrak a$ with the complex version of the [dual]{} Lie algebra $\mathfrak g^\ast$ and, consequently, the dual of the cocommutator $\eta$ will have to be isomorphic to one of the algebras $\mathfrak g$ in the first row of such Table. By proceeding in this way we find the following correspondences (we write first the Lie bialgebras $(\mathfrak g^\ast,\mathfrak g\equiv \eta^\ast)$ as labeled in Table III of [@gomez] and afterwards the corresponding quantum algebra according to our classification): - 5 $\; \rightarrow \;$ 1.2.2) ; 6 $\; \rightarrow \;$ 1.2.1) ; 7 $\; \rightarrow \;$ 1.1.1) ; - \(1) $\; \rightarrow \;$ 2.1.1) ; (2), (4) $\; \rightarrow \;$ 2.1.1) ; (3) $\; \rightarrow \;$ 2.2.1) ; 9 $\; \rightarrow \;$ 2.1.2) ; 11, 11’ $\; \rightarrow \;$ 2.2.2.2) ; - 10 $\; \rightarrow \;$ 2.2.2.4) ; 5$_{\rho=1}$ $\; \rightarrow \;$ 2.2.2.3) ; 6$_{\rho=1}$ $\; \rightarrow \;$ 2.2.2.1)$ $ ; 7$_{\rho=1}$ $\; \rightarrow \;$ 2.1.2)$_{\rho=1}$ ; - 5’ $\; \rightarrow \;$ 3.2.4) ; 8 $\; \rightarrow \;$ 3.2.1) ; (14) $\; \rightarrow \;$ 3.2.2) ; (11) $\; \rightarrow \;$ 3.2.3) ; 6$_{\rho=-1}$ $\; \rightarrow \;$ 3.2.5)$_{\rho=-1}$ ; 7$_{\rho=-1}$ $\; \rightarrow \;$ 3.1)$_{\rho=-1}$ . In this way we can realize that our choice (\[coconmutador\]) for the cocommutator implies that we have just obtained the quantizations for the full set of dual Lie bialgebras $(\mathfrak g^\ast,\mathfrak g\equiv \eta)$ of [@gomez] such that $\eta^\ast\equiv{\mathfrak r}_3(\rho)$ for all values of $\rho$. In fact, the $\rho$ parameter in (\[coconmutador\]) is identified with the one appearing in Gomez’s classification. Acknowledgments {#acknowledgments .unnumbered} =============== This work has been partially supported by DGI of the Ministerio de Ciencia y Tecnología (Projects BMF2002-0200 and BFM2000-1055), the FEDER Programme and Junta de Castilla y León (Projects VA085/02 and BU04/03) (Spain). The visit of E.C. to Valladolid have been financed by Universidad de Valladolid, by CICYT-INFN and by Ministerio de Educación y Cultura (Spain). [99]{} X. Gomez, [J. Math. Phys]{}. [41]{}. (2000) 4939 V.G. Drinfel’d, [Quantum Groups]{} in “Proceedings of the International Congress of Mathematicians”, Berkeley, 1986, A.M. Gleason (ed.) (AMS, Providence, 1987) V. Chari and A. Pressley, A Guide to Quantum Groups (Cambridge University Press, Cambridge, 1995) S. Majid, Foundations of Quantum Group Theory (Cambridge University Press, Cambridge, 1995) T. Tjin, Int. [J. Mod. Phys. A: Math. Gen.]{} [7]{} (1992) 6175 V.G. Drinfel’d, [Soviet Math. Dokl]{}. [27]{} (1983) 68 A. Ballesteros, E. Celeghini and F.J. Herranz, [J. Phys. A: Math. Gen.]{} [33]{} (2000) 3431 V. Lyakhovsky and A.I. Mudrov, [J. Phys. A: Math. Gen.]{} [25]{} (1992) L1139 V. Lyakhovsky, [Group-like structures in quantum-Lie algebras and the procedure of quantization]{} in [“Quantum Groups, Formalism and Applications”]{} ed. J Lukierski [et al]{} (Warsaw: Polish Scientific Publishers, 1995) A.I. Mudrov, [J. Math. Phys.]{} [38]{} (1997) 476 S. Schnider and S. Sternberg, Quantum Groups from Coalgebras to Drinfeld algebras (Int. Press, Cambridge (Mass.), 1993) N. Jacobson, Lie algebras (Dover, New York, 1979) M. Postnikov, Lectures in Geometry: Lie Groups and Lie Algebras (Mir, Moscow, 1982) E. Celeghini and M.A. del Olmo, [Europhys. Lett]{}. [61]{} (2003) 438 J. Patera, R.T. Sharp, P. Winternitz and H. Zassenhaus, [J. Math. Phys]{}. [17]{} (1976) 986 B.L. Aneva, D. Arnaudon, A. Chakrabarti, V.K.Dobrev and S.G. Mihov, [J. Math. Phys]{}. [42]{} (2001) 1236 M. Jimbo, [Lett. Math. Phys]{}. [10]{} (1985) 63 L.L. Vaksman and L.I. Korogodski, [Sov. Math. Dokl]{}. [39]{} (1989) 173 Ch. Ohn, [Lett. Math. Phys]{}. [25]{} (1992) 85 A. Ballesteros, E. Celeghini, F.J. Herranz, M.A. del Olmo and M. Santander, [J. Phys. A: Math. Gen.]{} [28]{} (1995) 3129	{ "pile_set_name": "ArXiv" }
--- abstract: 'It has been suggested that the observed rotation periods of radio pulsars might be induced by a non-axisymmetric spiral-mode instability in the turbulent region behind the stalled supernova bounce shock, even if the progenitor core was not initially rotating. In this paper, using the three-dimensional AMR code CASTRO with a realistic progenitor and equation of state and a simple neutrino heating and cooling scheme, we present a numerical study of the evolution in 3D of the rotational profile of a supernova core from collapse, through bounce and shock stagnation, to delayed explosion. By the end of our simulation ($\sim$420 ms after core bounce), we do not witness significant spin up of the proto-neutron star core left behind. However, we do see the development before explosion of strong differential rotation in the turbulent gain region between the core and stalled shock. Shells in this region acquire high spin rates that reach $\sim$$150\,$ Hz, but this region contains too little mass and angular momentum to translate, even if left behind, into rapid rotation for the full neutron star. We find also that much of the induced angular momentum is likely to be ejected in the explosion, and moreover that even if the optimal amount of induced angular momentum is retained in the core, the resulting spin period is likely to be quite modest. Nevertheless, induced periods of seconds are possible.' author: - 'Emmanouela Rantsiou, Adam Burrows, Jason Nordhaus, Ann Almgren' title: 'Induced Rotation in 3D Simulations of Core Collapse Supernovae: Implications for Pulsar Spins' --- Introduction ============ One of the key questions associated with radio pulsars is the origin of their spin periods (Lorimer 2009,2010). In two recent papers (Blondin & Mezzacappa 2007 \[hereafter BM07\]; Blondin & Shaw 2007), it has been suggested that the growth in 3D of an $\ell = 1$ non-axisymmetric ($m = \{-1,1\}$) spiral mode (Fernàndez 2010) of the so-called standing accretion shock instability (“SASI") during the post-bounce delay phase seen in current supernova theory can spin up matter in the turbulent region behind the stalled shock. BM07 and Blondin & Shaw (2007) find that the accretion of this spinning matter (and its associated angular momentum) onto the inner core can spin up the nascent neutron star to rotation periods that compare favorably with the observed/inferred values for pulsars at birth. More precisely, they report final periods of $\sim$70 milliseconds (ms) for moderately rotating progenitors and $\sim$50 ms for non-rotating progenitors. The measured period range of birth periods is from $\sim$10 milliseconds to seconds, with an average “injection period" near $\sim$0.5 seconds (Chevalier & Emmering 1986; Narayan 1987)[^1]. If this mechanism were to obtain, it would obviate the need to invoke a spinning progenitor core. However, rotation is a natural feature of stars, and there is no obvious reason the progenitor “Chandrasekhar" mass in the center of a massive star at death wouldn’t be rotating with the spins and angular momenta needed to leave a compact object with the measured spin frequencies (Ott et al. 2006ab). This is a function of the angular momentum distribution in the cores of massive stars when they collapse (Heger, Woosley, & Spruit 2005; Maeder & Meynet 2000,2004; Hirschi, Meynet, & Maeder 2004). Hence, the possibility that the 3D dynamics of the core itself could generate countervailing flows with opposite signs of angular momentum, “one stream" of which could leave the residue spinning at interesting rates while conserving overall angular momentum, is intriguing. If this were true, to zeroth order all cores at collapse could be non-rotating, and the spins could be a function of collapse dynamics alone. This would eliminate one of the current ambiguities in massive star progenitor models, whose core spins at collapse are not observable. The progenitor cores could all be born with “zero" angular momentum, while giving birth to proto-neutron stars with respectable spins. However, BM07 and Blondin & Shaw (2007) employed various simplifications in their simulations. They did not follow collapse itself, but started in a steady-state post-bounce configuration. They used a simple “gamma-law" equation of state (EOS), and not a full nuclear EOS. They set the mass accretion rate through the shock to a constant value and did not perform their simulations in the context of a realistic massive-star progenitor. Importantly, they excised the inner core whose very spin-up was being studied. Rather, they inferred from ostensible accretion of matter and angular momentum at their inner boundary an irreversible accumulation of spin in the proto-neutron star. Finally, they did not consider the potential role of the supernova explosion itself on the magnitude of the residual angular momentum and final induced core period. They did, however, investigate the potential role of initial rotation in the final outcome. In this paper, we investigate the claims of BM07 and Blondin & Shaw (2007), improving upon their study in several ways. Our 3D simulation is described in Nordhaus et al. (2010a) and was carried out using CASTRO (Almgren et al. 2010), an Eulerian Adaptive Mesh Refinement (AMR) hydrodynamics code with hierarchically nested rectangular grids that are refined both in space and time. We employed a realistic EOS (Shen et al. 1998ab) and a simple neutrino transfer scheme (see Nordhaus et al. 2010a and Murphy & Burrows 2008). Both Nordhaus et al. (2010a) and Murphy & Burrows (2008) incorporated the Liebendörfer (2005) prescription for electron capture on infall and a simple neutrino heating and cooling algorithm after bounce. The driving electron and anti-electron neutrino luminosities were kept constant at $1.9\times 10^{52}$ ergs s$^{-1}$. This enabled us to heat the so-called “gain region" (Bethe & Wilson 1985) and investigate the pre-bounce infall, bounce, post-bounce delay, and explosion phases. In this simulation (model 3d:L\_1.9 of Nordhaus et al. 2010a), a neutrino-driven explosion occurred approximately $\sim$200$-$250 ms after bounce and we carried the simulation out to $\sim$420 ms after bounce. Our initial model was the 15-M$_\odot$ solar-metallicity, non-rotating red-supergiant model of [@woos95]. An important aspect of our simulation is that the proto-neutron star itself is at all times present on the grid (which is Cartesian) and is followed throughout the run in a hydrodynamically consistent fashion. The cubic grid in our simulations extended to $10^4\,$ kilometers (km) on each side, with $304^3$ cubic cells covering the grid’s volume at its coarsest level. Three levels of refinement of a factor of $4$ in each dimension were followed. As a result, the resolution in the central regions (the inner $200\,$km) was $\sim$0.5 km. The monopole approximation for self-gravity was adopted. By the end of our simulation, we find that there is almost no spin up of the proto-neutron star. During the post-bounce phase and before explosion, the region exterior to the core and interior to the shock ($\sim$60 to $\sim$250 km) does witness the growth of spiral modes, some spin up (to $\sim$150 Hz), and the emergence of countervailing rotational flows. However, the explosion ejects not only mass, but its associated angular momentum, and depending upon where the mass cut is finally imposed, the residual core can end up with some net rotation. In our most optimistic scenario the core achieves a period of only $1.2\,$seconds. This fastest possible spin does not alone seem adequate to explain the measured and/or inferred pulsar birth periods without significant initial rotation in the progenitor itself. In §\[results\], we present our quantitative results and we follow in §\[discussion\] with an extended discussion and our summary conclusions. Results ======= Profiles and Temporal Evolution of Angular Momentum {#profiles_L} --------------------------------------------------- We begin our analysis by calculating the amount of angular momentum present in our computational domain. We divide the innermost $2000\,$km of the grid into concentric spheres of increasing radius, and we monitor the magnitude of each of the three components ($x, y, {\rm and}\ z$) of the angular momentum ($L_i$) enclosed within each sphere for the duration of the calculation after core bounce. The panels in Fig. \[plot1\] depict three-dimensional surface representations of the temporal evolution of the three orthogonal components of the total angular momentum enclosed within a certain radius, as a function of that radius. Such panels summarize what we find in our simulation. As Fig. \[plot1\] indicates, we do indeed see an initial increase of angular momentum, but only between $\sim$60 km and $\sim$250 km, and quasi-oscillatory behavior with a period near $\sim$30 ms. The innermost zones that might be identified with the proto-neutron star are not spun up to periods faster than $\sim$5$-$10 seconds. At later times (the last $0.17\,$ s after bounce), we note that the angular momentum that originally resided in this inner region is ejected with the exploding mass and follows the outward progress of the reenergized shock. This ejected angular momentum “bump” can also be seen clearly in Fig. \[plot2\], where we show the spatial profile of all three components of angular momentum for four different times near the end of our simulation. We address the subsequent fate of this ejected angular momentum in §\[final\_state\] and §\[discussion\]. At the end of our simulation, very little angular momentum is left in the central region of our grid, indicating that the inner core is left rotating very slowly. At the final timestep, the spin periods in the inner ten kilometers are all greater than $\sim$10 seconds. Note that the flatness of the regions in Fig. \[plot1\] at large distances demonstrates that total angular momentum is very well conserved during the simulation. In CASTRO, angular momentum is not conserved by construction, so that its global conservation during this simulation is encouraging. Specifically, before and after bounce the total angular momentum of matter on the grid is conserved to better than $\sim$1%, staying very near “zero" for the duration of the calculation. Next, we divide our grid into thin spherical shells that cover the innermost $2000\,$km of the computational domain. For the inner $250\,$km the width of the shells was chosen to be $1\,$km and for radii between $250$ and $2000\,$km the width was increased to $10\,$km. For each of the shells we calculated the average specific angular momentum ($\ell_i$, not the total enclosed angular momentum), for all three components of the angular momentum. Figure \[plot3\] depicts their associated spatial and temporal profiles. As discussed above and can be seen in Figs. \[plot1\] and \[plot2\], regions interior to $\sim$200 km experience some growth of specific angular momentum during the earlier stages after core bounce, whereas the specific angular momentum at larger radii remains close to zero until after the onset of explosion. As Fig. \[plot3\] makes clear, the sign of the specific angular momentum for each component varies with time and radius, conserving total angular momentum after bounce to high precision. After the onset of explosion, angular momentum is seen to propagate outward with the ejected mass and the shock wave. An interesting feature is observed for radii between $60$ and $250\,$km. As time progresses in the simulation, the magnitudes of the specific angular momenta on shells in this region grow with time, indicating that in this region there is indeed some spin up by the spiral mode identified in BM07, Blondin & Shaw (2007), and Fernàndez (2010). We proceed in §\[rot\_freq\] with a discussion of this phenomenon and the associated spin periods. Rotational Period and Frequency Profiles {#rot_freq} ---------------------------------------- Having calculated the angular momentum contained in spherical shells at various radii, we proceed to extract the rotational profiles of those shells. Figure \[plot4\] shows the temporal evolution of the average spin frequency ($f_i$, in Hz) of each shell, for rotation around the x, y and z axes (the three panels in Fig. \[plot4\]). We see that shells with radii $R<60\,$km and $R>250\,$km do not develop any significant rotational motion during the entire simulation. However, at the very early times after core bounce and for radii between $\sim$60 km and $\sim$250 km in the shocked, turbulent regions, some rotational motion does develop, which with time is amplified. By the end of the simulation, those shells have acquired spin frequencies that reach $\sim$150 Hz at a radius of $\sim$100 km, corresponding to a spin period of $\sim$7 ms. Furthermore, we notice that although at early post-bounce times the sign of the spin vector of those shells kept flipping (as indicated by the positive and negative frequency values shown in Fig. \[plot4\]), at the end of the simulation neighboring shells tended to align their rotational motion. The left panel of Fig. \[plot5\] depicts the velocity field on a ball at radius 90 km near the end of the simulation. As this figure clearly shows, coherent fields of rotation have emerged in this region (see also the panels in Fig. \[plot4\]). Near the same final time, the shock has been launched into explosion and has the multi-turbule morphology portrayed in the right panel of Fig. \[plot5\] as an isoentropy surface. In Fig. \[plot6\] we show the temporal evolution of the magnitude of the specific angular momentum (top) for various shells at different radii, and the corresponding magnitude of the rotational frequency (bottom) of those shells. As in Fig. \[plot4\], the development of rotational motion from $\sim$60 to $\sim$250 km is obvious and the high spin rate of $\sim$150 Hz in this narrow range of radii is manifest. However, there is very little mass ($\sim$0.01$-$0.03 M$_{\odot}$) and, hence, total angular momentum in this fastest spinning region, and this has major consequences for the spin evolution of the entire residual neutron star. Rotational Periods for the Final State {#final_state} -------------------------------------- Our simulation shows not only that the inner core isn’t spun up significantly, but that the mass that does experience an interesting degree of induced rotation is likely to be ejected with the subsequent blast. An interesting question arises associated with this ejected angular momentum “bump," shown in Figs. \[plot1\] and \[plot2\]. If that angular momentum were to be accreted later on onto the central regions (in a “fallback" scenario for angular momentum), what would that imply for the terminal rotation rate of the nascent neutron star? We address this issue by applying various mass cuts to the final state of our simulation. We calculate the total angular momentum enclosed within a given spherical mass cut (given in column 2 of Table \[table1\]) and then divide it by a representative moment of inertia for the final-state neutron star (taken to be $2 \times 10^{45}\,$ g cm$^{2}$ and assuming solid-body rotation) to derive a final angular frequency. Our results are summarized in Fig. \[plot7\] and Table \[table1\]. For a range of mass cuts between $1.2$ and $1.6$ M$_\odot$ (baryonic) for our “$1.9\times 10^{52}$ ergs s$^{-1}$" explosion model the final inferred rotational periods vary between 1.2 and 47 seconds (as can be seen in Fig. \[plot7\]). The fastest inferred period of $1.2\,$s is seen for a spherical mass cut of 1.53M$_\odot$ baryonic. Note that many measured neutron star masses are near 1.35 M$_{\odot}$ gravitational, which translates into $\sim$1.5 M$_{\odot}$ baryonic. Apart for the magnitude of such final-state rotational periods, we would also like to know the direction of rotation and how that varies with the different spherical mass cuts. We define two angles to describe the direction of the total angular momentum enclosed within a certain mass cut. Angle $\theta$ ($0^{\circ}\leqslant\theta\leqslant180^{\circ}$) is the polar angle, measured with respect to the y-axis, and angle $\phi$ ($0^{\circ}\leqslant\phi\leqslant360^{\circ}$) is the azimuthal angle, defined with respect to the z-axis. The two last columns of Table \[table1\] show the values of $\theta$ and $\phi$ for the different mass cuts. The direction of the angular momentum varies little for spherical mass cuts from $\sim$1.2 to $\sim$1.4 M$_\odot$). However, as Table \[table1\] indicates, it does shows significant variation for mass cuts in the range $1.45$ to 1.6 M$_\odot$. Discussion and Conclusions {#discussion} ========================== In this paper, we have studied the temporal and spatial evolution of the rotational profile of a newly-formed neutron star. The goal was to test the possibility that the remnant neutron star created in the context of core-collapse supernova could be induced to rotate rapidly due to the growth of a spiral, non-axisymmetric mode of the so-called SASI, even if the progenitor was initially non-rotating. We employed the 3D hydrodynamics code CASTRO to carry out the simulation, starting from the infall phase of a 15-M$_{\odot}$ non-rotating progenitor, and followed the core bounce, convection, and explosion stages. The core, whose induced rotation was being studied, was included fully on the computational grid, a realistic EOS was used, and a simple neutrino heating and cooling algorithm enabled us to simulate in approximate fashion the turbulent post-bounce delay and subsequent explosive phases. Our main findings are: - We did not find that a significant spin rate was induced in the nascent neutron star. By the end of our simulation the inner core is seen to rotate very slowly, with spin periods no faster than $\sim$5$-$10 seconds. - However, in the unstable gain region between the shock and the inner core (60 km $< R <$ 250 km), counter-rotating shells can acquire enough specific angular momentum during the stall phase before explosion that they can achieve interestingly high spin rates, reaching frequencies around $\sim$150 Hz (Fig. \[plot4\]) and spin periods between $\sim$6 and $\sim$100 ms. However, the mass in this region, and the corresponding angular momentum are rather low. - At the later stages in the evolution of our simulation, the explosion and ejecta transport angular momentum outward behind the exploding shock wave. Even if some of this angular momentum were to fall back at later times, the minimum spin period that the residual neutron star could achieve is likely to be quite modest (Fig. \[plot7\] and Table \[table1\]). Blondin & Mezzacappa (2007) employed a gamma-law equation of state, did not follow the hydrodynamics of the inner core, did not include in their simulations the collapse or the explosion phases, and did not include any neutrino interactions. Yet, they concluded that even a non-rotating progenitor could in principle leave behind a rotating remnant. In this qualitative conclusion we agree with Blondin & Mezzacappa (2007). However, we used a general nuclear equation of state, included the entire core all the way to the center in the hydrodynamic simulation, included in a crude fashion the effects of neutrino heating, cooling, and electron capture, and followed both the collapse and the explosion phases. What we found was that the degree of induced central rotation of the residual proto-neutron star, after explosion and for any given degree of fallback, produces periods between $\sim$1 and $\sim$10 seconds. Nevertheless, those authors suggest that a longer delay could lead to even greater net spin rates and see in their simulation a significant ramp up from $\sim$0.4 to $\sim$0.9 seconds after bounce. However, they concede that the spiral mode spin-up phenomenon should not continue beyond the onset of explosion. Since our explosion occurs near $\sim$250 ms, one could argue that a longer delay could result in a larger induced spin. While this seems plausible, [@wongwa] have performed a new set of approximate 3D simulations of core-collapse supernovae with non-rotating progenitors, focusing on the remnant neutron star kick velocities (see also Nordhaus et al. 2010b). They also show that after as much as $\sim$1.4 seconds after bounce the induced spin rates are still fairly low, with final periods in the range $\sim$500 to $\sim$1000 ms. Such long periods are approximately in accord with those we derive here in our “optimum" case. In summary, the increase in the spin rate from the optimum derived from our simulation, which terminated at $\sim$0.422 seconds, to that seen in the simulations of [@wongwa], which terminated at $\sim$1.4 seconds, was a factor of $\sim$2. The calculations performed by Blondin & Mezzacappa (2007) implied that the enhancement in delaying from $\sim$0.422 seconds to $\sim$0.9 seconds might be a factor of $\sim$10 or more. We take this as a further indication that induced rotation by spiral modes may not be adequate to explain the observed (or inferred) rotation rates of most pulsars. Nevertheless, induced spin remains an intriguing, if sub-dominant, potential contribution to the overall angular momentum budget of nascent neutron stars. Where final neutron-star spin periods of seconds are observed, the spiral-mode spin-up mechanism may remain viable. One could argue that the delay to explosion we witness is already rather long. We suggest that in order for the neutrino mechanism to be robust and to generate explosion energies sufficient to overcome the binding energies of the progenitor mantle, while at the same time yielding explosion energies at infinity of $\sim$10$^{51}$ ergs, the onset of explosion should probably be rather early after bounce. However, this has yet to be demonstrated. We emphasize that our models start with zero angular momentum. A longer delay and total angular momentum conservation would naturally make it less and less likely that any significant [net]{} angular momentum would be left behind. This effect is implicit in Fig. \[plot7\], for which the residual net spin period jumps up on the right-hand side at greater mass cuts. Individual zones could initially be left with interesting spins, but others would have to be in the opposite directions required by net angular momentum conservation. At very late times, the residue would have to be in solid-body rotation at low net angular momentum. In this paper, we have not explored various progenitor masses, nor a range of initial spin structures. Our goal was to determine whether a sizable rotation rate could be induced in the residual proto-neutron star when the progenitor itself was non-rotating. The behavior of the modest spiral mode we do see, and the overall role of rotation in the supernova phenomenon itself, remain to be fully mapped out. As computational capabilities improve, and codes acquire more physics and sophistication, all these issues can and will be readdressed. The authors would like to acknowledge Rodrigo Fernàndez, Timothy Brandt, John Bell, and Brian Metzger for fruitful discussions and input. A.B. and J.N. are supported by the Scientific Discovery through Advanced Computing (SciDAC) program of the DOE, under grant number DE-FG02-08ER41544. E.R. is supported by the NSF under the subaward no. ND201387 to the Joint Institute for Nuclear Astrophysics (JINA, NSF PHY-0822648), and A.B. receives support from the NSF PetaApps program, under award OCI-0905046 via a subaward no. 44592 from Louisiana State University to Princeton University. Work at LBNL was supported in part by the SciDAC Program under contract DE-AC02-05CH11231. The authors would like to thank the members of the Center for Computational Sciences and Engineering (CCSE) at LBNL for their invaluable support for CASTRO. The authors employed computational resources provided by the TIGRESS high performance computer center at Princeton University, which is jointly supported by the Princeton Institute for Computational Science and Engineering (PICSciE) and the Princeton University Office of Information Technology; by the National Energy Research Scientific Computing Center (NERSC), which is supported by the Office of Science of the US Department of Energy under contract DE-AC03-76SF00098 and on the Kraken and Ranger supercomputers, hosted at NICS and TACC and provided by the National Science Foundation through the TeraGrid Advanced Support Program under grant number TGAST100001. Almgren, A.S., Beckner, V.E., Bell, J.B., Day, M.S., Howell, L.H., Joggerst, C.C., Lijewski, M.J., Nonaka, A., Singer, M., & Zingale, M. 2010, , 715, 1221 Atoyan, B. M. 1999, , 346, L49 Blondin, J. M., & Mezzacappa, A. 2007, Nature, 445, 58 Blondin, J. M., & Shaw, S. 2007, , 656, 366 Chevalier, R. & Emmering, R.T. 1986, , 304, 140 Fernàndez, R. 2010, arXiv:1003.1730 Heger, A., Woosley, S. E., & Spruit, H. C. 2005, , 626, 350 Hirschi, R., Meynet, G., & Maeder, A. 2004, , 425, 649 Liebendörfer, M. 2005, , 633, 1042 Lorimer, D. R. 2009, Astrophys. & Space Sci. Lib., 357, 1 Lorimer, D. R. 2010, to appear in the Proceedings of the ICREA Workshop on The High-Energy Emission from Pulsars and their Systems, held in Sant Cugat, Spain, 2010 April 12-16 (Springer), arXiv:1008.1928 Maeder, A. & Meynet, G. 2000, ARA&A, 38, 143 Maeder, A. & Meynet, G. 2004, , 422, 225 Murphy, J. W., & Burrows, A. 2008, ApJ, 688, 1159 Narayan, R. 1987, , 319, 162 Nordhaus, J., Burrows, A., Almgren, A., & Bell, J. 2010a, , 720, 694 Nordhaus, J., Brandt, T., Burrows, A., Livne, E. & Ott, C.D. 2010b, accepted to , arXiv:1010.0674 Ott, C. D., Burrows, A., Thompson, T. A., Livne, E., & Walder, R.2006a, , 164, 130 Ott, C. D., Burrows, A., Livne, E., & Walder, R. 2006b, , 600, 834 Shen, H., Toki, H., Oyamatsu, K., & Sumiyoshi, K. 1998a, Nucl. Phys. A, 637, 435 \|. 1998b, Prog. Th. Phys., 100, 1013 Wongwathanarat, A., Janka, H.-Th. & Müller, E. 2010, submitted to , arXiv:1010.0167 Woosley, S. E. & Weaver, T. A. 1995, , 101, 181 [cccccc]{} 1.20 & 0.501 & 2.50 &0.40 & 97.6 & 67.2\ 1.25& 0.531 & 2.37 &0.42 & 92.1 & 67.5\ 1.30 & 0.479 & 2.83 &0.35 & 89.7& 70.5\ 1.35 & 0.380 & 3.30 &0.30 & 96.7& 84.2\ 1.40 & 0.252 & 4.99 &0.20 &111.6 & 272.3\ 1.45 & 0.125 & 10.17 &0.1 & 152.5&0.25\ 1.50 & 0.619 & 2.02 &0.49 & 128.6& 51.6\ 1.55 & 0.790 & 1.59 &0.63 & 21.6 & 5.3\ 1.60 & 0.027 & 46.89 &0.02 &125.5 & 191.9 \[table1\] ![Three-dimensional surface representation of the temporal evolution (after bounce) of the components ($L_i$) of the total angular momentum (vertical axes) enclosed within a spherical radius R (in km). The three panels, from top to bottom, correspond to the x-, y-, and z-components of the enclosed angular momentum. The two-dimensional plot under each surface shows the projection of the upper surface onto the plane, with the color variation indicating the value of the projected point, going from black (negative) to white (positive) through a scale of red shades. We notice the initial increase of the angular momentum components in the very inner regions of our computational domain, and the ejection (with the ejected mass) of the bulk of angular momentum outwards during the later stages of the simulation. See text in §\[profiles\_L\] for further discussion.[]{data-label="plot1"}](f1a.pdf "fig:"){height="0.4\columnwidth" width="0.75\columnwidth"}\ ![Three-dimensional surface representation of the temporal evolution (after bounce) of the components ($L_i$) of the total angular momentum (vertical axes) enclosed within a spherical radius R (in km). The three panels, from top to bottom, correspond to the x-, y-, and z-components of the enclosed angular momentum. The two-dimensional plot under each surface shows the projection of the upper surface onto the plane, with the color variation indicating the value of the projected point, going from black (negative) to white (positive) through a scale of red shades. We notice the initial increase of the angular momentum components in the very inner regions of our computational domain, and the ejection (with the ejected mass) of the bulk of angular momentum outwards during the later stages of the simulation. See text in §\[profiles\_L\] for further discussion.[]{data-label="plot1"}](f1b.pdf "fig:"){height="0.4\columnwidth" width="0.75\columnwidth"}\ ![Total angular momentum enclosed within a spherical radius R and its evolution with time after bounce, for four different timesteps. The three panels correspond to the three components of the angular momentum vector (x, y and z components from top to bottom, respectively). Lines of different color indicate different post-bounce times during the simulation. We notice again (as in Fig. \[plot1\]) an initial increase of the angular momentum components between $\sim$60 and $\sim$200 km. With time, much of the angular momentum is ejected with the exploding mass. This can be seen here in the propagating angular momentum “bump." As discussed in the text and also in the captions to Table \[table1\] and Fig. \[plot7\], this ejected angular momentum, if actually all accreted onto the neutron star at later time, would induce a rotational period no faster than $1.2\,$s.[]{data-label="plot2"}](f2a.pdf "fig:"){height="0.4\columnwidth" width="0.75\columnwidth"}\ ![Total angular momentum enclosed within a spherical radius R and its evolution with time after bounce, for four different timesteps. The three panels correspond to the three components of the angular momentum vector (x, y and z components from top to bottom, respectively). Lines of different color indicate different post-bounce times during the simulation. We notice again (as in Fig. \[plot1\]) an initial increase of the angular momentum components between $\sim$60 and $\sim$200 km. With time, much of the angular momentum is ejected with the exploding mass. This can be seen here in the propagating angular momentum “bump." As discussed in the text and also in the captions to Table \[table1\] and Fig. \[plot7\], this ejected angular momentum, if actually all accreted onto the neutron star at later time, would induce a rotational period no faster than $1.2\,$s.[]{data-label="plot2"}](f2b.pdf "fig:"){height="0.4\columnwidth" width="0.75\columnwidth"}\ ![Temporal evolution (in seconds after bounce) of the average specific angular momentum ($\ell_i$, vertical axes) of spherical shells at various enclosed spherical radii (R, in centimeters). R is logaritmically distributed. From top to bottom, the three panels correspond to the x-, y-, and z-component of the specific angular momentum. To determine the averages, the computational grid was divided into thin spherical shells of width $1\,$km within the inner $250\,$km of the grid and of width $10\,$km for radii between $250$ and $2000\,$km. The two-dimensional surfaces under each upper surface are projections of the upper surfaces, with the color variation indicating the value of the projected point, going from black (negative) to white (positive) through a scale of red shades. The propagation outward at later times of the induced angular momentum is manifest in these panels as well. See the text in §\[profiles\_L\] for a discussion.[]{data-label="plot3"}](f3a.pdf "fig:"){height="0.4\columnwidth" width="0.75\columnwidth"}\ ![Temporal evolution (in seconds after bounce) of the average specific angular momentum ($\ell_i$, vertical axes) of spherical shells at various enclosed spherical radii (R, in centimeters). R is logaritmically distributed. From top to bottom, the three panels correspond to the x-, y-, and z-component of the specific angular momentum. To determine the averages, the computational grid was divided into thin spherical shells of width $1\,$km within the inner $250\,$km of the grid and of width $10\,$km for radii between $250$ and $2000\,$km. The two-dimensional surfaces under each upper surface are projections of the upper surfaces, with the color variation indicating the value of the projected point, going from black (negative) to white (positive) through a scale of red shades. The propagation outward at later times of the induced angular momentum is manifest in these panels as well. See the text in §\[profiles\_L\] for a discussion.[]{data-label="plot3"}](f3b.pdf "fig:"){height="0.4\columnwidth" width="0.75\columnwidth"}\ ![The temporal evolution of the spin frequency profiles of thin spherical shells at various spherical radii R (in centimeters, shown logarithmically). For each of the individual shells we have calculated the corresponding average rotational frequency, shown here on the vertical axis of each plot. Time is the post-bounce time in seconds. The three panels correspond to rotation around the x, y and z axes (from top to bottom). The positive and negative values of the frequency indicate opposite rotation directions. The computational grid was divided into thin spherical shells of width $1\,$km within the inner $250$ km of the grid and of width $10\,$km for radii between $250$ and $2000\,$km. We notice that for radii in the range $R<60\,$km and $R>250\,$km, the resulting frequencies remain near “zero" throughout the entire simulation. Interestingly, in the region $60<R<250\,$km, some rotational motion develops even during the early stages of the simulation and amplifies during the very late stages, producing shell rotational frequencies that reach $\sim$$150\,$Hz at a radius of $100\,$km. See text in §\[rot\_freq\] for a discussion.[]{data-label="plot4"}](f4a.pdf "fig:"){height="0.4\columnwidth" width="0.75\columnwidth"}\ ![The temporal evolution of the spin frequency profiles of thin spherical shells at various spherical radii R (in centimeters, shown logarithmically). For each of the individual shells we have calculated the corresponding average rotational frequency, shown here on the vertical axis of each plot. Time is the post-bounce time in seconds. The three panels correspond to rotation around the x, y and z axes (from top to bottom). The positive and negative values of the frequency indicate opposite rotation directions. The computational grid was divided into thin spherical shells of width $1\,$km within the inner $250$ km of the grid and of width $10\,$km for radii between $250$ and $2000\,$km. We notice that for radii in the range $R<60\,$km and $R>250\,$km, the resulting frequencies remain near “zero" throughout the entire simulation. Interestingly, in the region $60<R<250\,$km, some rotational motion develops even during the early stages of the simulation and amplifies during the very late stages, producing shell rotational frequencies that reach $\sim$$150\,$Hz at a radius of $100\,$km. See text in §\[rot\_freq\] for a discussion.[]{data-label="plot4"}](f4b.pdf "fig:"){height="0.4\columnwidth" width="0.75\columnwidth"}\ ![[Left:]{} Velocity vector field profile on a spherical shell of radius $90\,$ km towards the end of the simulation ($\sim$370 ms after core bounce). Both the size and color of the vectors represent the magnitude of the velocity field, with darker (lighter) shades corresponding to higher (lower) values. [Right:]{} 3D rendering of an isentropic surface with entropy equal to 10 $k_b$ baryon$^{-1}$. The colormap represents the magnitude of the entropy gradient on the surface, with values increasing going from orange to purple. The surface spans 1600 km and corresponds to a time 380 ms after core bounce.[]{data-label="plot5"}](f5a.pdf "fig:"){height="0.55\columnwidth" width="0.55\columnwidth"} ![[Left:]{} Velocity vector field profile on a spherical shell of radius $90\,$ km towards the end of the simulation ($\sim$370 ms after core bounce). Both the size and color of the vectors represent the magnitude of the velocity field, with darker (lighter) shades corresponding to higher (lower) values. [Right:]{} 3D rendering of an isentropic surface with entropy equal to 10 $k_b$ baryon$^{-1}$. The colormap represents the magnitude of the entropy gradient on the surface, with values increasing going from orange to purple. The surface spans 1600 km and corresponds to a time 380 ms after core bounce.[]{data-label="plot5"}](f5b.pdf "fig:"){height="0.55\columnwidth" width="0.55\columnwidth"} ![Temporal evolution (in seconds) of the magnitude of the specific angular momentum (top) and of the magnitude of the rotational frequency (bottom) of spherical shells at various spherical radii R (in centimeters, shown logarithmically). As discussed in the caption to Fig. 4 and in the text, individual shells at radii $60<R<250\,$km develop significant rotational frequencies over time. However, outside this range of radii the rotation induced by non-axisymmetric instabilities is quite modest at all times during the simulation.[]{data-label="plot6"}](f6a.pdf){height="0.6\columnwidth" width="0.75\columnwidth"} -4in ![Rotation periods of inferred final-state neutron stars as a function of various possible mass cuts (in M$_{\odot}$), as described in the caption to Table 1 and in the text. To derive these approximate final spin periods, the enclosed total angular momentum is divided by a final neutron star moment of inertia of $2\times 10^{45}$ g cm$^{2}$. Solid-body rotation for the final state is assumed. The fastest induced spin up is seen to be for a mass cut of $1.53$ M$_\odot$ (baryonic mass), resulting in a slow rotational period of $\sim$1.2 seconds.[]{data-label="plot7"}](f7.pdf){width="0.90\columnwidth"} [^1]: The inferred initial spin period of the Crab pulsar is 15$-$16 milliseconds (Atoyan 1999), but the average for radio pulsars is considerably longer.	{ "pile_set_name": "ArXiv" }
--- abstract: 'NGC 4636, an X-ray bright elliptical galaxy, was observed for 70 ks with Suzaku. The low background and good energy resolution of the XIS enable us to estimate the foreground Galactic emission accurately and hence measure, for the first time, the O, Mg, Si and Fe abundances out to a radius of $\sim$28 arcmin ($\simeq$ 140 kpc). These metal abundances are as high as $>$1 solar within the central 4’ and decrease by $\sim$50% towards the outer regions. Further, the O to Fe abundance ratio is about 0.60–1.0 solar in all regions analyzed, indicating that the products of both SNe II and SNe Ia have mixed and diffused to the outer regions of the galaxy. The O and Fe metal mass-to-light-ratios (MLR) of NGC 4636 are 2–3 times larger than those of NGC 1399 implying that metal distributions in NGC 4636 are less extended than those in NGC 1399, possibly due to environmental factors, such as frequency of galaxy interaction. We also found that the MLRs of NGC 4636 at 0.1 $r_{180}$ are $\sim$5 times smaller than those of clusters of galaxies, possibly consistent with the correlation between temperature and MLR of other spherically symmetric groups of galaxies. We also confirmed a resonant scattering signature in the Fe ${\rm_{XV\hspace{-.1em}I\hspace{-.1em}I}}$ line in the central region, as previously reported using the XMM-Newton RGS.' author: - \| Katsuhiro H$^1$, Yasushi F$^1$, Miyako T$^1$\ Sho N$^1$, Kyoko M$^2$, Yoh T$^3$,and Keith A. A$^{4,5}$\ \ \ \ \ \ ****** title: \| Suzaku Observation of the Metallicity Distribution\ in the Elliptical Galaxy NGC 4636 --- Introduction ============ The universe is filled with matter in various physical states. One of these is the hot interstellar medium (ISM) in elliptical galaxies. The ISM is gravitationally bound by the mass of stars, dark matter and itself, and its temperature is about 0.5–1 keV. Emission from the ISM is observed in the X-ray band and its luminosity reaches about 10$^{39}$–10$^{42}$ erg s$^{-1}$. By accurate analysis of X-ray spectra, we can estimate the temperature and metallicity distributions of the ISM and reveal the metal enrichment history of the universe. In the past, X-ray imaging detectors, such as Chandra ACIS or XMM-Newton EPIC, have observed the ISM with high angular resolution and high signal-to-noise ratio (S/N) or wide field of view, and revealed temperature structures, surface brightness and mass distributions (e.g., Buote 2002; Buote et al. 2003; Fukazawa et al. 2006; Humphrey et al. 2006). In addition, observations by XMM-Newton have suggested similar metal abundance patterns to those in the centers of clusters or groups of galaxies (e.g., Matsushita et al. 2003; Matsushita et al. 2007b). The excellent energy resolution of XMM-Newton RGS has been used to precisely measure metal abundances (e.g., Xu et al. 2002; Tamura et al. 2003) and show that Si and Fe abundances in the central region are similar, while the O abundance is in general about half that of Fe. To understand metal enrichment processes, measurements of O, Ne, or Mg abundances in the outer regions are needed. To unambiguously constrain the O abundance in elliptical galaxies it is also necessary to have an accurate estimate of the foreground Galactic thermal emission. The best detector to satisfy these conditions is the Suzaku XIS which has a better energy resolution and a lower background than any previous X-ray CCD camera. The launch of the Suzaku XIS (Koyama et al. 2007) with high energy resolution in the lower energy X-ray band allows us to obtain O and Mg abundance distributions through accurate measurements of their emission lines. Past studies (e.g., NGC1399 ; Matsushita et al. 2007a, NGC 5044 ; Komiyama et al. 2008) have shown that the O to Fe abundance ratio, which can be used to estimate the number ratio of Ia to II supernovae, is about 0.6–0.8 solar in all regions. However, observations in the outer regions, where metals ejected from galaxies have remained, are still poor. So far, Suzaku measurements of abundance distribution have been reported for three elliptical galaxies. NGC 5044 and NGC 507 exhibit a steeper radial O abundance profile than NGC 1399 while there is a large scatter among their O mass-to-light ratios. (Matsushita et al. 2007a; Komiyama et al. 2009; Sato et al. 2009a). Studies of these properties in additional galaxies will thus aid understanding of metal enrichment processes. In this paper, we report the results of spectral analysis of NGC 4636 observed by the XIS detector onboard Suzaku (Mitsuda et al. 2007). NGC 4636 (z = 0.00313; $\simeq$ 17 Mpc, Smith et al. 2000) is an X-ray bright elliptical galaxy belonging to the Virgo cluster. Metal-enriched hot gas is extended over the entire galaxy group (Matsushita et al. 1998). This galaxy is located at the south end of the cluster, where the X-ray emission from the Virgo center is negligible, so we can study the metal enrichment history of a giant elliptical galaxy (Matsushita et al. 1997). So far, observations by ASCA (Matsushita et al. 1997), Chandra (Ohto et al. 2003; Humphrey et al. 2006), and XMM-Newton (e.g. Xu et al. 2002) have been used to measure metal abundances. The temperature of hot gas in this galaxy is 0.6–0.7 keV, which is lower than the 1.0–1.5 keV of NGC 1399, NGC 507, and NGC 5044. Therefore, we can also discuss the temperature dependence of metal properties. In addition, NGC 4636 is quite isolated in comparison with the above 3 galaxies, and thus the comparison will aid investigations of environmental effects. We describe the observation and data reduction in section 2. Section 3 on the analysis and results describes the methods we used to estimate background and fit spectra as well as the resulting temperature and metallicity distributions. Additionally, we discuss the confirmation of resonance scattering in the Fe ${\rm_{XV\hspace{-.1em}I\hspace{-.1em}I}}$ line in the spectrum of the central region. In section 4, we discuss the O, Ne, Mg, Si abundance ratios to Fe, their distributions, the metal transport process, and resonance scattering of the Fe ${\rm_{XV\hspace{-.1em}I\hspace{-.1em}I}}$ line. A summary is given in section 5. Throughout this paper, we use a Hubble constant value of $H_{0}$ = 70 km s$^{-1}$Mpc$^{-1}$ or $h_{100}$ = 0.7 (1 arcmin $\simeq$ 5 kpc), and a virial radius of [r]{}$_{180}$ = 1.95$h^{-1}_{100}$ $\sqrt[]{\mathstrut k{\langle T\rangle}/10{\rm keV}}$ Mpc (Markevitch et al. 1998) where $kT$ is the gas temperature. The Galactic hydrogen column density is $N_{\rm H}$ = 1.8$\times$10$^{20}$ cm$^{-2}$ in the direction of NGC 4636 (Stark et al. 1992). We use the Anders and Grevesse (1989) definition of Solar abundance ratios for easier comparison with other Suzaku papers which use this definition. Other recent papers often use instead Grevesse and Sauval (1998). The former has 26% and 48% higher abundances for O and Fe, respectively. All errors are given as a 90% confidence range. Observation and Data Reduction ============================== A pointing observation of NGC 4636 by Suzaku was carried out on December 6–7, 2005 for about 70 ks. In order to measure metal abundances in the outer region of NGC 4636, a 10’ north offset observation was carried out on December 7–9, 2007. Another north-west offset observation at 3.2$^{\circ}$ away from the NGC 4636 center was also carried out on June 17–18, 2007 to estimate the foreground Galactic emission – this region contains no other bright objects. All the XIS data were taken with both 3$\times$3 and 5$\times$5 pixel mode. The observation log is summarized in Table 1. [cccc]{} Target & NGC 4636 & $^1$NGC 4636-NORTH & $^2$NGC 4636-GALACTIC\ z & 0.00313 & - & -\ position (Ra, Dec) in J2000 & 190.7250, 2.7520 & 190.6828, 3.0503 & 192.5138, 5.4608\ exposure time (ks) & 70.3 & 51.1 & 34.9\ date (y-m-d h:m) & 05-12-06 06:38 & 07-12-07 05:01 & 07-06-17 20:11\ \ \ We analyzed the public rev 1.2 data and screened them with cut-off rigidity (COR) $>$ 8 GV, target elevation angle above the earth rim (ELV) $>$ 5$^{\circ}$, and elevation angle above the day earth rim (DYE\_ELV) $>$ 20$^{\circ}$. For the 2005 observation, we extracted spectra of six annular regions centered on the X-ray peak (Figure 1(a)). For the north offset observation, we analyzed the spectrum of a 8’-radius aperture centered on a position at 20’ away from the galaxy center (Figure 1(b)). In extracting the above spectra, we excluded the region which is contaminated by the Fe calibration source. We used xisrmfgen version 2006-11-26 and xissimarfgen version 2006-10-26 to make response matrix files (RMF) and ancillary response files (ARF), respectively. The source mode of xissimarfgen was set to SKYFITS, which is appropriate for analyzing extended objects, and the XIS image file was input. Data from the hard X-ray detector, HXD, are not treated here because no signal was detected from NGC 4636; an upper limit of the hard X-ray emission is $9\times10^{-12}$ erg cm$^{-2}$ s$^{-1}$ (10–50 keV) for the 3% background uncertainty (Fukazawa et al. 2009). [cc]{} (a) \ (b) \ Spectral Analysis and Results ============================= Estimation of Background ------------------------ In order to study the emission from NGC 4636 correctly, we have to estimate the background accurately. This consists of the instrumental non X-ray background (NXB), the cosmic X-ray background (CXB) and the foreground Galactic X-ray emission (GXE). First, we estimated the NXB spectra in each analysis region using the tool “xisntebgdgen”, which adds COR-sorted ($<$ 4GV, 4–13 GV in 1 GV step, and $>$ 13 GV) night-Earth spectra weighted by the exposure time for each COR in the NGC 4636 observation. Second, using the sky-averaged CXB parameters from HEAO-1 (Boldt et al. 1987) summarized in table 2, we estimated the CXB contribution by using an uniform-sky ARF, which is generated with “xissimarfgen” and assumes that the emission is uniform within 20’. Finally, we estimated the GXE, using the 3.2$^{\circ}$ offset observation by extracting spectra within 8’ ($\simeq$ 40 kpc) of the center of the field of view. We used a uniform-sky ARF as for the CXB estimation. The GXE consists of a Local Hot Bubble (LHB) component and a Milky Way Halo (MWH) component. These have temperatures of approximately $\sim$0.08 keV and $\sim$0.2–0.3 keV, respectively (e.g., Lumb et al. 2002). We fitted the GXE spectrum with a two-temperature APEC model, where temperatures are set to be free and metal abundances are fixed to one solar. At these temperatures, the continuum emission is vanishingly small compared to the line emission. Therefore, even if this assumption for metal abundance is not completely correct, the estimation of the GXE will not change significantly. The results of the fit are shown in table 2, along with the fixed parameter values used for the CXB model. Figure 2 shows the best fit model and spectra. The temperatures obtained are consistent within errors with the results from the ROSAT-PSPC all-sky survey (Kuntz & Snowden 2000). Using the best-fit parameters, we modeled the Galactic emission in each spectral region of the NGC 4636 observations by assuming the same surface brightness and spectral parameters as those of the 3.2$^{\circ}$ offset region. [ccccccc]{}\ & & &\ & 1.29 & & 40.0\ \ & & &\ & $kT$ (keV) & normalization$^{\S}$ & $kT$ (keV) & normalization$^{\S}$ & reduced-$\chi^2$ (d.o.f.)\ & 0.326$\pm$0.013 & 2.31$\pm$0.37 & 0.094$\pm$0.004 & 2.02$\pm$0.84 & 1.48 (206)\ \ \ \ \ \ \ \ [c]{} Spectral Fitting ---------------- We fitted the spectra of each region of NGC 4636, based on the above estimation of the NXB, the CXB, and the GXE. The CXB and the GXE are included as models in the spectral fit with their spectral parameters fixed, and the NXB was subtracted. We applied a two-component model for the emission from NGC 4636 : a single-temperature vAPEC model for the ISM, and a BREMSS model with a temperature fixed to 7 keV (Matsushita et al. 1994) for the hard component due to low-mass X-ray binaries (LMXB). For the all regions, the fitting is performed in the energy range of 0.3–5 keV because the spectrum below 0.3 keV is difficult to analyze due to large systematic errors, and the NXB dominates above 5 keV. The energy range around the Si K-edge (1.82–1.841 keV) was ignored due to response problems. We divided the metals into several elemental groups (He = C = N, O, Ne, Mg = Al, Si, S = Ar = Ca, Fe = Ni), and fixed the abundance of the He=C=N group to solar. Abundances of other groups were treated as free parameters. [ccc]{} & & \ & & \ [cccccccc]{} region & model & $kT$ (keV) & O (solar) & Ne (solar) & Mg (solar) & Si (solar) & Fe (solar)\ 0–2’ & 1T$^a$ & 0.642$\pm$0.002 & 0.75$\pm$0.06 & 1.54$\pm$0.12 & 1.50$\pm$0.11 & 1.51$\pm$0.12 & 1.14$\pm$0.07\ 0–2’ & 2T$^b$ & 1.56$\pm$0.33, 0.641$\pm$0.002 & 0.79$\pm$0.07 & 1.47$\pm$0.15 & 1.52$\pm$0.11 & 1.54$\pm$0.12 & 1.20$\pm$0.07\ 2–4’ & 1T & 0.733$\pm$0.004 & 1.01$\pm$0.13 & 2.14$\pm$0.28 & 2.33$\pm$0.24 & 1.81$\pm$0.19 & 1.41$\pm$0.12\ 4–6’ & 1T & 0.792$\pm$0.005 & 0.41$\pm$0.06 & 0.62$\pm$0.11 & 0.86$\pm$0.08 & 0.57$\pm$0.06 & 0.49$\pm$0.03\ 6–8’ & 1T & 0.811$\pm$0.006 & 0.41$\pm$0.08 & 0.45$\pm$0.13 & 0.74$\pm$0.10 & 0.50$\pm$0.07 & 0.43$\pm$0.03\ 8–10’ & 1T & 0.796$\pm$0.008 & 0.34$\pm$0.10 & 0.60$\pm$0.17 & 0.86$\pm$0.14 & 0.53$\pm$0.11 & 0.43$\pm$0.05\ 10–12’ & 1T & 0.774$\pm$0.013 & 0.58$\pm$0.28 & 0.67$\pm$ 0.42 & 1.20$\pm$0.40 & 0.84$\pm$0.39 & 0.68$\pm$0.17\ 12–28’ & 1T & 0.746$\pm$0.016 & $<$1.20 & $<$1.60 & 1.13$\pm$0.41 & 0.87$\pm$0.31 & 0.63$\pm$0.17\ region & & &\ 0–2’ & & &\ 0–2’ & & &\ 2–4’ & & &\ 4–6’ & & &\ 6–8’ & & &\ 8–10’ & & &\ 10–12’ & & &\ 12–28’ & & &\ \ \ \ \ \ Table 3 shows the best-fit results for each region. The fit is acceptable for all the regions. The O abundance is measured up to 10–12’ and Mg, Si and Fe to 12–28’ for the first time. However, the Ne abundance is not reliable because Ne line measurements are influenced by nearby Fe-L lines, which cannot be resolved with the resolution of the XIS. It can be seen from Figure 3 that the Galactic emission is well reproduced and significantly contributes to the spectra in the outer region, and thus it must be accurately estimated to derive the O abundance of NGC 4636. As a result, even for O, the abundances in the outer regions are determined up to 10’ with $\sim$30% accuracy. Even if we allow the temperature of the BREMSS (7keV) component to vary, the metal abundances only change within errors so we believe that they are reliable. The luminosity of the hard X-ray component of the BREMSS model is 1.5$\times$10$^{40}$ erg s$^{-1}$ over 2–10 keV within 6’ of NGC 4636. This is consistent with the ASCA results (Matsushita et al. 1994), indicating that the hard X-ray component is due to LMXB. We also confirmed that the abundance results are almost the same when we fit the hard component with a power-law of index 1.5, which was used for the LMXB estimation in Loewenstein et al. (2001). In the central region, the reduced-$\chi^2$ is a little larger than that in the outer region. This might be due to temperature structure in the center region (Ohto et al. 2003) so we tried fitting with a two-temperature model only in the 0–2’ region. We assumed that the metal abundances are the same for both temperature components. Comparing the result with that of the single-temperature model, the metal abundances do not greatly change and the temperature of the hot component in the two-temperature model is almost the same as that in the single-temperature. The F-test probability is 0.097% and thus the improvement is marginal. Therefore, hereafter we focus on the results with the single-temperature model. A fit residual around 0.7 keV in the 0–2’ region is discussed and analyzed in §3.4. Temperature and Metal Abundances -------------------------------- Distributions of temperature and metal abundances are plotted against radius (kpc) in Figures 4 and 5, respectively. Systematic errors due to the GXE, CXB, and NXB are also shown. The systematic errors due to the GXE are estimated by varying the normalizations of the APEC model for the MWH and LHB within their error ranges, which are comparable to the X-ray count fluctuation (15%) of the ROSAT All-sky X-ray background survey (band 4)[^1] for the NGC 4636 region. The systematic errors due to the CXB and NXB are at most 5% (Tawara et al. 2008), respectively, and they are dominant in the hard X-ray band. Therefore, the systematic errors on the temperature and metal abundances are mostly caused by the GXE uncertainty. A 15% uncertainty in the estimate of the XIS optical blocking filter thickness introduces systematic errors comparable to those from the GXE. The temperature in the central region is 0.64 keV, rising to 0.7–0.8 keV. This profile is consistent with that from ASCA (Matsushita et al. 1997), XMM-Newton (Xu et al. 2002) and Chandra (Ohto et al. 2003) and the temperature distribution is barely affected by any systematic errors in the GXE. For the inner 50 kpc most metal abundances are tightly restricted thanks to the high energy resolution of the Suzaku XIS while in the outer regions the abundances are less well constrained due to the GXE systematic errors. However, the Ne abundance is not reliable because Ne lines are confused with Fe-L lines. Furthermore, only upper limits could be obtained on the O and Ne abundances in the 60–140 kpc region, since their uncertainties are very large due to contamination by the Galactic emission. It can been seen that all the metal abundances are 0.7–2.5 solar in the central region, and decrease down to 0.3–1.0 solar towards the outer region. In the central region, the O and Fe abundances are about 1.2 times larger than that obtained with the RGS (Xu et al. 2002). Figure 6 shows confidence contours of O and Mg abundances against the Fe abundance. The O and Mg to Fe abundance ratio are particularly important to constrain the contribution of different types of supernovae. Although the individual errors for each metal abundance are large, the abundance ratios are constrained well, especially for the Mg to Fe abundance ratio in the outermost region. [c]{} -- -- -- -- -- -- -- -- -- -- Resonance Scattering in the Central Region ------------------------------------------ In Figure 7, we show the spectra in the 0–2’and 2–4’ regions. In the spectrum of the 0–2’ region, the fit residual has a negative line feature around 0.85 keV. This feature disappears in the 2–4’ region and other outer regions (Figure 3). This is possibly due to resonance scattering of the Fe ${\rm_{XV\hspace{-.1em}I\hspace{-.1em}I}}$ line, which was reported with the RGS (Xu et al. 2002; Werner et al. 2009). When we introduce a negative gaussian line in the spectral model, the fit significantly improves with a F-test probability of $<0.01$%. In that case, the Fe abundance increases by 15%. We compared the XIS result in the central region with that using the RGS and looked into whether the features are consistent. We modeled the XIS spectrum with two BREMSS for the continuum thermal emission from the plasma and the LMXBs, and add each emission line with a gaussian. The temperatures of these BREMSS models are fixed to 0.642 keV and 7.0 keV for the thermal and LMXB emission, respectively. The former is the best-fit vAPEC temperature in table 3. The lines considered here are listed in Table 4. Since the XIS cannot resolve individual lines, we summed the flux of neighboring lines within 30 eV in the RGS paper (Xu et al. 2002), and assumed the strength-weighted line energy. Furthermore, we refer to the ATOMDB line table [^2] for the energy and intensity of the 0.961 and 0.963 keV lines, which were not analyzed in the RGS spectrum, but are significant in the XIS spectra. The Ne lines around this energy range are not shown because their fluxes are not reliable due to the influence of Fe-L lines. The line width of all lines is fixed to 0 keV. The line intensity ratio against the summed flux of the strongest line (0.737 and 0.728 keV ) of Fe ${\rm_{XV\hspace{-.1em}I\hspace{-.1em}I}}$ is summarized in Table 4 and Figure 8. In Figure 8, the predicted line intensity ratios are also shown. These were obtained by simulating spectra with the emission models and parameters from Table 3 and fitting them with the same model used for the data. The result indicates that the flux intensity of the Fe ${\rm_{XV\hspace{-.1em}I\hspace{-.1em}I}}$ line (0.825 keV), which was suggested to be affected by resonance scattering (Xu et al. 2002), is about 0.75 times lower than that predicted by the model. This might indicate that the resonance scattering occurs for the Fe ${\rm_{XV\hspace{-.1em}I\hspace{-.1em}I}}$ ion. The Suzaku intensity ratio is about 1.4 times higher than that from the RGS. This might be because the RGS photons are accumulated from a smaller central region than the XIS where the resonance scattering is more effective. Doron and Behar (2002) indicated the problem of modeling the Fe+16/Fe+17 ion lines, but their recalculation showed that the line intensity ratio of the 0.825 keV line to the sum of the 0.737 and 0.728 keV lines changes by 30–40%, mostly independent of the temperature in the range of 0.6–0.9 keV. Their paper does not refer to the APEC model, and thus this factor of change is not necessarily accepted for our cases. Our results show that the discrepancy of the line intensity ratio between the data and the model is seen only at the central region. Therefore, we conclude that the resonance scattering is robust. [cc]{} [cccccccc]{} Ion & &\ & $E$ (keV) & \flux & \sum of flux & intensity ratio & $E$ (keV) & \*flux & intensity ratio\ N ${\rm_{V\hspace{-.1em}I}}$ & 0.418-0.428 & 0.43 & 0.43 & 0.05 & 0.428 & 0 & 0\ N ${\rm_{V\hspace{-.1em}I\hspace{-.1em}I}}$ & 0.497 & 0.20 & 0.20 & 0.03 & 0.497 & 0 & 0\ O ${\rm_{V\hspace{-.1em}I\hspace{-.1em}I}}$ & 0.557 & 0.21 & & & & &\ O ${\rm_{V\hspace{-.1em}I\hspace{-.1em}I}}$ & 0.565 & 0.09 & & & & &\ O ${\rm_{V\hspace{-.1em}I\hspace{-.1em}I}}$ & 0.570 & 0.13 & 0.43 & 0.05 & 0.563 & 0 & 0\ O ${\rm_{V\hspace{-.1em}I\hspace{-.1em}I\hspace{-.1em}I}}$ & 0.651 & 1.96 & 1.96 & 0.245 & 0.651 & 1.03$\pm$0.18 & 0.21$\pm$0.04\ Fe ${\rm_{XV\hspace{-.1em}I\hspace{-.1em}I}}$ & 0.728 & 5.66 & & & & &\ Fe ${\rm_{XV\hspace{-.1em}I\hspace{-.1em}I}}$ & 0.737 & 2.33 & 7.99 & 1 & 0.731 & 4.85$\pm$0.26 & 1.00$\pm$0.08\ Fe ${\rm_{XV\hspace{-.1em}I\hspace{-.1em}I\hspace{-.1em}I}}$ Fe ${\rm_{XI\hspace{-.1em}X}}$ O ${\rm_{V\hspace{-.1em}I\hspace{-.1em}I\hspace{-.1em}I}}$ & 0.769 & 1.99 & 1.99 & 0.249 & 0.769 & 2.94$\pm$0.28 & 0.61$\pm$0.07\ Fe ${\rm_{XV\hspace{-.1em}I\hspace{-.1em}I}}$ & 0.809 & 1.66 & & & & &\ Fe ${\rm_{XV\hspace{-.1em}I\hspace{-.1em}I}}$ & 0.825 & 3.56 & 5.22 & 0.653 & 0.820 & 4.46$\pm$0.26 & 0.92$\pm$0.07\ Fe ${\rm_{XV\hspace{-.1em}I\hspace{-.1em}I\hspace{-.1em}I}}$ & 0.853 & 2.25 & & & & &\ Fe ${\rm_{XV\hspace{-.1em}I\hspace{-.1em}I\hspace{-.1em}I}}$ & 0.859 & 1.85 & & & & &\ Fe ${\rm_{XV\hspace{-.1em}I\hspace{-.1em}I\hspace{-.1em}I}}$ Fe ${\rm_{XX}}$ & 0.865 & 1.45 & & & & &\ Fe ${\rm_{XV\hspace{-.1em}I\hspace{-.1em}I\hspace{-.1em}I}}$ & 0.871 & 1.90 & 7.45 & 0.932 & 0.861 & 4.98$\pm$0.23 & 1.03$\pm$0.07\ Fe ${\rm_{XX}}$ & 0.961 & - & & & & &\ Fe ${\rm_{XX}}$ & 0.963 & - & - & - & 0.962 & 1.73$\pm$0.15 & 0.36$\pm$0.04\ Mg ${\rm_{XI}}$ & 1.35 & 0.59 & 0.59 & 0.070 & 1.35 & 0.61$\pm$0.04 & 0.13$\pm$0.01\ \ [c]{} Discussion ========== We measured the temperature and metal abundance distributions of NGC 4636 out to 28’ and confirmed the resonance scattering in the central region. In this section we compare the Suzaku results with those for other X-ray bright elliptical galaxies (NGC 1399, NGC 507, and NGC 5044), and discuss the metal enrichment process in elliptical galaxies. Metallicity Distribution ------------------------ In Figures 4 and 5, observational results by the Suzaku XIS for NGC 1399 (Matsushita et al. 2007a) are also shown. These are the results with a two-temperature model within 4’ and one-temperature model beyond 4’. The temperature of NGC 1399 is about twice as high as that of NGC 4636, indicating that the gravitational potential of NGC 1399 is deeper. In Figure 5, the Ne abundance of NGC 1399 was not measured because its line was polluted by Fe-L lines. The radial distributions of the O, Mg, Si and Fe abundances are similar between the two galaxies, but metal abundances are about twice as high as those of NGC 1399 in all radial regions. In Figure 9, radial profiles of the abundance ratios of O, Ne, Mg, and Si against Fe are shown, with error bars estimated from the allowed slope of the confidence contours in Figure 6. For comparison, the abundance ratios for a model of SN II (for all elements) and SN Ia (for Si) are also shown. The abundance ratios for SN II are calculated by integrating the metal mass ejected by SNe II (Iwamoto et al. 2006) with a progenitor mass from 10M$_{\odot}$ to 50M$_{\odot}$, weighting by the Salpeter initial mass function, and calculated, for example, as O: 1.80M$_{\odot}$ and Fe: 0.0907M$_{\odot}$. The W7 model (O: 0.143M$_{\odot}$ and Fe: 0.749M$_{\odot}$ ; Iwamoto et al. 1999) is used for SN Ia. The result indicates that all metal abundances can be explained by a mixture of SN II and SN Ia products. This indicates that, in NGC 4636 and NGC 1399, each metal diffused to the outside of the galaxy after mixing products from both SN II and SN Ia. For NGC 4636, all the abundance ratios are constant in the radial direction. The abundance ratios against Fe in the central region (0–10 kpc) are 0.66$\pm^{0.07}_{0.06}$ solar, 1.35$\pm^{0.12}_{0.10}$ solar, 1.32$\pm^{0.06}_{0.07}$ solar, 1.32$\pm^{0.11}_{0.10}$ solar for O, Ne, Mg, Si, respectively. The O/Fe ratio is similar to that of NGC 1404 (Matsushita et al. 2007a) and NGC 720 (Tawara et al. 2006) measured by the XIS, and also with the RGS results for NGC 4636 (Xu et al. 2002) and NGC 5044 (Tamura et al. 2003). This indicates that a common metal production process has taken place in elliptical galaxies. However, the Mg/Fe ratio of NGC 4636 is higher than that of NGC 1399 in the outer region. This difference might be caused by different populations of SN II progenitor stars between the two galaxies, but problems with the plasma emission model cannot be ruled out (Matsushita et al. 2003). It is noticeable that the radial profile of the O abundance of NGC 4636 exhibits a steeper gradient than other elliptical galaxies, NGC 1399 , NGC 5044, and NGC 507. This indicates that NGC 4636 keeps metals more tightly concentrated. NGC 4636 is rather isolated and has never been identified as a galaxy group; the other galaxies are classified as the center galaxy of their galaxy group, which contains many dwarf galaxies (Ferguson and Sandage 1990; Tifft et al. 1975). Therefore, the differences in the abundance gradient could be explained by NGC 4636 experiencing fewer galaxy interactions leading to reduced diffusion of metals to the outer region. We consider this further in the next section. -- -- -- -- Metal-mass-to-light-ratio ------------------------- The gas mass distributions of NGC 4636 and NGC 1399 are shown in Figure 10, which also shows the Chandra results within 0.022 $r_{180}$ ($\simeq$ 15.6 kpc) (Fukazawa et al. 2006) and the ASCA results beyond this radius (Matsushita et al. 1998). These results show that the gas mass of NGC 4636 is about half as much as that of NGC 1399. This is possibly because NGC 1399 sits at the center of the Fornax cluster, where the potential is deeper than that for NGC 4636 (Figure 4), so the gas is more tightly bound by NGC 1399. The NGC 4636 metal mass distributions and metal-mass-to-light-ratios (MLR) of O and Fe are shown in Figures 11(a) and (b), respectively. The metal mass is obtained by multiplying the metal abundance value in the outermost region in Figure 5 with the gas mass distribution (Figure 10). The stellar light distribution is quoted from Fukazawa et al. (2006). [c]{} [cc]{} Figure 11(a) shows that the O and Fe mass distributions are similar between the two galaxies and the metal mass of NGC 4636 is almost the same as that of NGC 1399 at all radii. Combining this with the difference of the gas mass, it can be understood that the abundance and the gas mass of NGC 4636 are about twice and half those of NGC 1399, respectively. As seen in Figure 11(b), MLR of NGC 4636 is also 2$\sim$3 times larger than that of NGC 1399 in all regions for both metals. This indicates that the metals in the ISM are distributed more compactly in NGC 4636; the metals in NGC 4636 do not diffuse so widely. This difference might be caused by environmental effects; galaxy interaction is frequent for NGC 1399 and thus metal-rich gas could be stirred by galaxy motions. Finally, we compared the MLR of NGC 4636 with Suzaku results of other groups and clusters. These objects are galaxy groups centered on NGC 1399, NGC 5044, and NGC 507. They are X-ray bright elliptical galaxies like NGC 4636, but the galaxy density is high around them and thus they are identified as centers of galaxy groups. We also compare with galaxy clusters such as Abell 262, Abell 1060, AWM 7, and Centaurus. Following Sato at al. (2009b), we plot the correlation between temperature and MLR at 0.1 $r_{180}$ in figure 12. This result shows that MLR of groups and galaxies with a temperature of $\sim$1 keV, including NGC 4636, is $\sim$5 times lower than that for clusters. Among galaxies and groups, NGC 4636, NGC 5044 and HCG 62 exhibit a spherically symmetric X-ray emission, and NGC 4636 has the lowest MLR among them. A possible correlation between temperature and MLR is suggested only for these spherically symmetric objects. Such a correlation could be explained by higher temperature systems possessing a deeper gravitational potential which prevents dispersal of the metals. Although galaxy velocity dispersion is larger for higher temperature clusters, there are no big galaxies around cD galaxies. Therefore, we think that the deep potential can confine the metals even if a small galaxy passes around the cD galaxy. On the other hand, the X-ray asymmetric groups NGC 1399 and NGC 507 show a relatively low MLR. These imply that NGC 1399 and NGC 507 experience galaxy interactions frequently and metal-rich gas could be stirred by galaxy motions. [c]{} Optical depth of Fe ${\rm_{XV\hspace{-.1em}I\hspace{-.1em}I}}$ in the central region ------------------------------------------------------------------------------------ We calculate the optical depth of resonance scattering for the Fe ${\rm_{XV\hspace{-.1em}I\hspace{-.1em}I}}$ line in the central region mentioned in $\S$3.4. We assume that the gas density distribution follows the $\beta$ model. The scattering cross section $\sigma_\nu$ is represented by the following formula (Shigeyama et al. 1998), $$\begin{aligned} %\sigma_R=\frac{\pi e^2}{m_i c} gf \frac{h}{\sqrt[]{\mathstrut {2\pi}}\Delta % E} \sigma_\nu=\frac{\pi e^2}{m_e \nu_0} gf \sqrt[]{\mathstrut \frac{m_i}{2\pi kT}}\end{aligned}$$ $m_i$ is the mass of the ion, $gf$ is oscillator strength, $\nu_0$ is the frequency of the emission and $T$ is the temperature of the system. We set $T$ to 0.642 keV which is the temperature in the 0–2’ region and the value of $gf$ to 2.5 (Mewe et al. 1985). As a result, $\sigma_\nu$ becomes 1.20$\times$10$^{-15}$ cm$^2$. For the density distribution $n(r)$, we use the $\beta$ model, $$\begin{aligned} n(r)=n_0[1+({\frac{r}{r_c})^2}]^{-{\frac{3}{2}}\beta}\end{aligned}$$ $n_0$ is the core density and $r_c$ the core radius. For NGC 4636, the core density, the core radius and $\beta$ are 0.2 cm$^{-3}$, 1.1 kpc and 0.47, respectively (Ohto et al. 2003). Then, the optical depth for the emission line from the circular region with projected radius $R$ is calculated by the following formula, assuming that the Fe abundance in the central region A$_{\rm Fe}$ is 1.14 solar from the value summarized in Table 3, the Fe abundance solar ratio [Fe/H]{} is 4.68$\times10^{-5}$ (Anders & Grevesse 1989), ionic fraction $f_i$ of Fe ${\rm_{XV\hspace{-.1em}I\hspace{-.1em}I}}$ is 0.23 (SPEX) and gas exists up to 300 kpc. $$\begin{aligned} \tau(r)=\int^{300}_R A_{\rm Fe} {\rm(Fe/H)} f_i \sigma_{\nu_0} n(r) \frac{r}{\sqrt[]{\mathstrut {r^2-R^2}}} dr\end{aligned}$$ As a result of this calculation, the value of $\tau$ in the central region within 2’ becomes 24.1 $<$ $\tau$ $<$ 56.3. Note that those values are not influenced by thermal Doppler broadening and extent of the diffuse source. Even if we consider them, $\tau$ $\gg$ 1 is secure. Our result is somewhat larger than that of Werner et al. (2009), and the difference is possibly explained by the assumption of the Fe abundance. Therefore, it is optical thick there and resonance scattering will easily occur. Turbulence could reduce the resonance scattering (Werner et al. 2009), but the scattering is significant for NGC 4636. Summary ======= In this paper, we performed spectral analysis out to 28’($\simeq$ 140 kpc) from the center of NGC 4636, using data from the XIS detectors onboard Suzaku. We obtained the following results. - [The temperature distribution shows a smooth gradient from the center (0.642 keV) to the outer regions. For the metal abundance, we found that the distributions of O, Mg, Si, Fe are all higher in the central 4’ region and decrease by 50% toward the outer region. It is noticeable that O and Mg abundances are determined accurately, even out to 28’ from the center.]{} - [The abundance ratio of O/Fe is about 0.6$\sim$1.0 solar independent of the radius. This means that O and Fe in the gas diffused to the outside of the galaxy after the products from both SNe II and SNe Ia were well mixed.]{} - [We found out that the O and Fe radial mass distributions are similar between NGC 4636 and NGC 1399. This indicates that both objects have similar metal diffusion processes. However the MLR of NGC 4636 is 2$\sim$3 times larger than that of NGC 1399, possibly due to NGC 4636 confining more metal-rich gas. The MLR of NGC 4636 is $\sim$5 times smaller than that of groups and clusters of galaxies and there may be a correlation between the temperature and MLR in other spherical groups.]{} - [We confirmed resonance scattering in the Fe ${\rm_{XV\hspace{-.1em}I\hspace{-.1em}I}}$ line in the central region of NGC 4636, first suggested by the XMM-Newton RGS results. The optical depth in the central region is, 24.1 $<\tau<$ 56.3, and thus it is extremely optically thick in the central region if there is no turbulence. ]{} We thank to the anonymous referee for careful reading and many helpful comments. We also thank to the Suzaku team for development of hardware/software and operation. [00]{} Anders, E., & Grevesse, N. 1989, Geochim. Cosmochim. Acta, 53, 197 Boldt, E., 1987, Phys. Rep., 146, 215 Buote, D. A. 2002, ApJ, 574, L135 Buote, D. A. 2003, Lewis, A. D., Brighenti, F., & Mathews, W. G. ApJ, 594, 741 Doron, R., & Behar, E. 2002, ApJ, 574, 518 Ferguson, H. C., & Sandage, A. 1990, AJ, 100, 1 Fukazawa, Y., Botoya-Nonesa, J. G., Pu, J., Ohto, A., & Kawano, N. 2006, ApJ, 636, 698 Fukazawa, Y. et al. 2009, PASJ, 61, S17 Grevesse, N., & Sauval, A. J. 1998, Space Science Reviews, 85, 161 Humphrey, P. J., & Buote, D. A. 2006, ApJ, 639, 136 Ikebe, Y. 1995, Ph.D. thesis, Univ. Tokyo Iwamoto, K., Brachwitz, F., Nomoto, K., Kishimoto, N,. Umeda, H., Hix, W. R., & Thielemann, F. -K. 1999, ApJ, 125, 439 Komiyama, M., Sato, K., Nagino, R., Ohashi, T., & Matsushita, K. 2009, PASJ, 61, S337 Koyama, K., et al. 2007, PASJ, 59, S23 Kuntz & Snowden 2000, ApJ, 543, 196 Loewenstein, M., Mushotzky, R. F., Angelini, Lorella., & Arnaud, K. A. 2001, ApJ, 555, L23 Lumb, D. H., Warwick, R. S., Page, M., & De Luca, A. 2002, A&A, 389, 102 Markevitch, M., Forman, W. R., Sarazin, C, L., & Vikhlinin, A. 1998, ApJ, 503, 91 Matsushita, K., Makishima, K., Awaki, H., Canizares, C. R., Fabian, A. C, Fukazawa, Y., Loewenstein, M., Matsumoto, H., Mihara, T., Mushotzky, R. F., Ohashi, T., Ricker, G. R., Serlemitsos, P. J., Tsuru, T., Tsusaka, Y., & Yamazaki, T. 1994, ApJ, 436, L44 Matsushita, K., Makaishima, K., Rokutandai, E., Yamasaki, Y. N., & Ohashi, T. 1997, ApJ, 488, L125 Matsushita, K., Makishima, K., Ikebe, Y., Rokutandai, E., Yamasaki, Y. N., & Ohashi, T. 1998, ApJ, 499, L13 Matsushita, K., Finoguenov, A., & B$\rm{\ddot{o}}$hringer, H. 2003, A&A, 401, 445 Matsushita, K., B$\rm{\ddot{o}}$hringer, H., Takahashi, I., Ikebe, Y. 2007b, A&A, 462, 954 Matsushita, K., Fukazawa, Y., Hughes, J. P., Kitaguchi, T., Makishima, K., Nakazawa, K., Ohashi, T., Ota, Naomi., Tamura, T., Tozuka, M., Tsuru, T, G., Urata, Y., Yamasaki, N, Y. 2007a, PASJ, 59, S327 Mewe, R., Gronenschild, E. H. B. M., van den Oord, G. H. J, 1985, A&A, 62, 213 Mitsuda, K., et al. 2007, PASJ, 59, 1 Nomoto, K., Tominaga, N., Umeda, H., Kobayashi, C., & Maeda, K. 2006, Nuclear Physics A, 777, 424 Fukazawa, Y., Botoya-Nonesa, J. G., Pu, J., Ohto, A., & Kawano, N. 2006, ApJ, 636, 698 Ohto, A., Kawano, N., Fukazawa, Y. 2003, PASJ, 55, 819 Shigeyama, T., et al. 1998, ASJ, 497, 587 Sato, K., Yamasaki, N. Y., Ishida, M., Ishisaki, Y., Ohashi, T., Kawahara, H., Kitaguchi, T., Kawaharada, M., Kokubun, M., Makishima, K., Ota, N., Nakazawa, K., Tamura, T., Matsushita, K., Kawano, N., Fukazawa, Y., Hughes, J. P. 2007, PASJ, 59, 305 Sato, K., Matsushita, K., Ishisaki, Y., Sasaki, S., Ohashi, T., Yamasaki, N. Y., & Ishida, M. 2008, PASJ, 60, 333 Sato, K., Matsushita, K., Ishisaki, Y., Yamasaki, N. Y., Ishida, M., & Ohashi, T. 2009a, PASJ, 61, S353 Sato, K., Matsushita, K., & Gastaldello, F. 2009b, PASJ, 61, S365 Tamura, T., Kaastra, J. S., Makishima, K., & Takahashi, I. 2003, A&A, 399, 497 Tawa, N. et al. 2008, PASJ, 60, S11 Terada, Y. 2001, Ph.D. thesis, Univ. Tokyo Tifft, W. G., Hilsman, K. A., & Corrado, L. C. 1975, ApJ, 199, 16 Tokoi, K, Sato, K., Ishisaki, Y., Ohashi, T., Yamasaki, N. Y., Nakazawa, K., Matsushita, K., Fukazawa, Y. et al. 2008b, PASJ, 60, S317 Werner, N. et al. 2009, astro-ph/0904.0254 Xu, H., Kahn, S. M., Peterson, J. R., Behar, E., Paerels, F. B. S., Mushotzky, R. F., Jernigan, J. G., Brinkman, A. C., & Makishima, K. 2002, ApJ, 579, 600 [^1]: SkyView: [http://skyview.gsfc.nasa.gov/]{} [^2]: http://cxc.harvard.edu/atomdb/WebGUIDE/index.html	{ "pile_set_name": "ArXiv" }
--- abstract: 'In this note we present some remarks on big Cohen-Macaulay algebras and almost Cohen-Macaulay algebras via closure operations on the ideals and the modules. Our methods for doing these are inspired by the notion of the tight closure and the dagger closure and by the ideas of Northcott on dropping of the Noetherian assumption of certain homological properties.' address: - 'School of Mathematics, Institute for Research in Fundamental Sciences (IPM), P.O. Box 19395-5746, Tehran, Iran.' - 'Dinabandhu Andrews College, Kolkata 700084, India' author: - Mohsen Asgharzadeh - 'Rajsekhar Bhattacharyya $^\ast$' title: 'Applications of closure operations on big Cohen-Macaulay Algebras' --- Introduction ============ Throughout this paper all rings are commutative, associative, with identity, and all modules are unital. We always denote a commutative Noetherian ring by $R$ and an $R$-algebra (not necessarily Noetherian) by $A$. The perfect closure of a reduced ring $A$ of prime characteristic $p$ is defined by adjoining all the higher $p$-power roots of all the elements of $A$, to $A$ itself. We denote it by $A^\infty$. Observe the definition of coherent ring and non-Noetherian regular ring in Section 2. In Section 2, we show that the Frobenius map is flat over a coherent regular ring which is of prime characteristic. In view of [@E page 36], one may ask for a generalization of Tight Closure via Colon-Capturing with respect to some non-Noetherian Cohen-Macaulay rings. Let $R$ be a Noetherian local F-coherent domain which is either excellent or homomorphic image of a Gorenstein local ring. If ${\operatorname{p.grade}}_A({\frak{a}},M)$ is the polynomial grade of an ideal ${\frak{a}}$ on an $A$-module $M$ (see Section 2 for the definition) then by applying a version of colon-capturing for $R^{\infty}$, in Theorem \[main\] of Section 3, we show that ${\operatorname{ht}}({\frak{a}})={\operatorname{p.grade}}_{R^{\infty}}({\frak{a}}, R^{\infty})$ for every ideal ${\frak{a}}$ of $R^{\infty}$. On regular rings, this result follows easily by using a result of Kunz (which is true only for regular rings), see e.g. [@AT Theorem 4.5]. Following [@Sh], a ring $R$ is called F-coherent if $R^{\infty}$ is a coherent ring. In particular, Theorem \[main\] provides an evidence for Shimomoto’s question: Does the perfect closure of an F-coherent domain coincide with the perfect closure of some regular Noetherian rings? The Dagger closure was introduced by Hochster and Huneke in [@HH]. It has an interesting connection with the theory of vector bundles, see e.g. [@BS] and the theory of tight closure, see e.g. [@HH] and [@HH2]. Let $I$ be an ideal of $R$ and let $R^{+}$ be its integral closure in an algebraic closure of its fraction field. We recall that an element $x\in R$ is in $I^{\dag}$, the dagger closure of $I$, if there are elements $\epsilon_n\in R^{+}$ of arbitrarily small order such that $\epsilon_nx\in IR^{+}$. By the main result of [@HH], the tight closure of an ideal coincides with the dagger closure, where $R$ is complete and of prime characteristic $p$. In Section 4, we extend that notion to the submodules of finitely generated modules over $R$ to prove that if a complete local domain is contained in an almost Cohen-Macaulay domain then there exists a balanced big Cohen-Macaulay module over it (see Corollary \[main2\]). In [@Ho], Hochster showed the existence of a big Cohen-Macaulay algebra from the existence of an almost Cohen-Macaulay algebra in dimension 3. Corollary \[main2\] extends this result, by proving the existence of big Cohen-Macaulay modules from the existence of almost Cohen-Macaulay algebras in any dimension $d$. In [@S1], certain weakly almost Cohen-Macaulay $R^+$-algebra is constructed and it has been asked whether it is almost Cohen-Macaulay or not. In Section 5, we obtain that such a weakly almost Cohen-Macaulay $R^+$-algebra will be almost Cohen-Macaulay if and only if it maps to some big Cohen-Macaulay $R^+$-algebra. For more details see Remark 5.4. Preliminary Notations ===================== We begin this section by exploring the following definitions. Let $A$ be an algebra equipped with a map $v:A\to{{\mathbb R}}\cup \{\infty\}$ satisfying $(1)$ $v(ab)=v(a)+v(b)$ for all $a,b \in A$; $(2)$ $v(a+b) \ge \min\{v(a),v(b)\}$ for all $a,b \in A$ and $(3)$ $v(a)=\infty$ if and only if $a=0$. The map $v$ is called the value map on $A$. Moreover, if $v(a) \ge 0$ for every $a \in T$ and $v(a) > 0$ for every non-unit $a \in A$, then we say that $v$ is normalized. If $A$ is equipped with a normalized value map, then by [@AS Proposition 3.1], one has $\bigcap_{n=1}^{\infty} {\frak{a}}^n=0$ where ${\frak{a}}$ is proper and finitely generated ideal of $A$. Let $A$ be of prime characteristic $p$ and $I$ be an ideal of $A$. Set $A^0$ to be the complement of the set of all minimal primes of $A$ and $I^{[q]}$ to be the ideal generated by $q=p^e$-th powers of all elements of $I$. Then the tight closure $I^\ast$ of $I$ is the set of $x\in A$ such that there exists $c \in A^0$ such that $cx^q\in I^{[q]}$ for $q\gg 0$. We also recall the following definitions: a ring is called coherent if each of its finitely generated ideal is finitely presented and a ring is called regular if each of its finitely generated ideal is of finite projective dimension. \[sh1\] Let $A$ be a coherent regular ring of prime characteristic $p$. Then the following statements hold. 1. The Frobenius map is flat. 2. If $A$ is equipped with a normalized value map, then all finitely generated ideals of $A$ are tightly closed. $\mathrm{(i)}$: This proof is inspired by a famous result of Kunz [@BH Corollary 8.2.8], which characterizes Noetherian regular rings of prime characteristic in the term of Frobenius map. First recall that the assignment $a\mapsto a^p$ defines the Frobenius map $F:A\to A$ which is a ring homomorphism. By $F(A)$, we mean $A$ as a group equipped with left and right scalar multiplication from $A$ given by $$a.r\star b = ab^pr, \ \ where \ \ a,b\in A \ \ and \ \ r\in F(A),$$ see also [@BH Section 8.2]. We show that for all $i > 0$, ${\operatorname{Tor}}^A_i(A/{\frak{a}}, F(A))=0$ for all finitely generated ideals ${\frak{a}}\subset A$. Note that $A/{\frak{a}}$ has a free resolution $(F_\bullet,d_\bullet)$ consisting of finitely generated modules, since $A$ is coherent. Such a resolution is bounded, because $A$ is regular. Then $(F_\bullet,d_\bullet)\otimes_A F(A)=(F_\bullet,d_\bullet^p)$. Consider the ideal $I_{t}((a_{ij}))$ generated by the $t\times t$ minors of a matrix $(a_{ij})$ and recall the definition of Koszul grade, which is denoted by ${\operatorname{K.grade}}$. One can show that it is unique up to the radical, see [@BH Proposition 2.2 (vi)]. Let $r_i$ be the expected rank of $d_\bullet$, see [@BH Section 9.1]. Clearly, $r_i$ is the expected rank of $d_\bullet^p$. Now, all these facts together imply, $${\operatorname{K.grade}}_A(I_{r_i}(d_i), A)={\operatorname{K.grade}}_A(I_{r_i}(d_i^p ),A).$$ In view of [@BH Theorem 9.1.6], which is a beautiful theorem of Buchsbaum-Eisenbud-Northcott, we find that $(F_\bullet,d_\bullet^p)$ is exact and so ${\operatorname{Tor}}^A_i(A/{\frak{a}}, F(A))=0$. $\mathrm{(ii)}$: This proof is inspired by [@Sh Lemma 4.1]. Let ${\frak{a}}$ be a proper and finitely generated ideal of $A$. Suppose that ${\frak{a}}^\ast\neq{\frak{a}}$. Take $x\in {\frak{a}}^\ast\setminus {\frak{a}}$. From the definition and from part $(i)$ of the proof, it follows that there exists an element $c \in A^0$ such that $c\in( {\frak{a}}^{[q]}:_Ax^q)=({\frak{a}}:_Ax)^{[q]}$ for $q\gg 0$. We recall that $\bigcap_{n=1}^{\infty} {\frak{b}}^n=0$ for any proper and finitely generated ideal of $A$, since $A$ is equipped with a normalized value map. Here $({\frak{a}}:_Ax)$ is proper, since $x\notin {\frak{a}}$ and it is also finitely generated, since $A$ is coherent. These facts together imply $c\in\bigcap ({\frak{a}}:_Ax)^{[q]}\subseteq\bigcap ({\frak{a}}:_Ax)^q=0$ and thus we arrive at a contradiction. It is to be noted that the finitely generated assumption of the previous result is really needed. \[exam\] Let $(R,{\frak{m}})$ be a Noetherian regular local ring of prime characteristic $p$ which is not a field. Let $R^\infty$ be its perfect closure. In view of [@A Theorem 1.2], $R^\infty$ is coherent and of finite global dimension. This is well-known that, $R^\infty$ is equipped with a normalized value map. To construct such a value map, consider a discreet valuation valuation ring $(V,{\frak{n}})$, which birationally dominates $R$, i.e., $R\subseteq V$, ${\frak{n}}\cap R={\frak{m}}$ and both $R$ and $V$ have the same field of fractions. We know that the above situation exists. Let $v$ be a value map of $V$ and take $y\in R$ such that $v(y)=\ell\in \mathbb{N}\setminus\{0\}$. Let $r\in R^\infty$. Then $r^{p^n}\in R$ for some $n$. The assignment $r\mapsto v(r^{p^n})/p^n$ defines a normalized value map on $R^\infty$. Consider the ideal ${\frak{a}}:=\{x\in R^\infty:v(x)>\ell/p\}$. Here, we show that ${\frak{a}}^\ast\neq {\frak{a}}$. To see this, let $x\in R^\infty$ be the p-th root of $y\in R$. Here $v(x)=\ell/p$. Take $c\in R^\infty$ with $v(c)>0$. Clearly, $v(c^{1/q}x)>\ell/p$ and so $c^{1/q}x\in {\frak{a}}.$ Thus $cx^q=\prod_q c^{1/q}x\in{\frak{a}}^{[q]}$ for $q\gg 0$. Therefore, $x\in {\frak{a}}^\ast\setminus {\frak{a}}$. By Lemma \[sh1\], ${\frak{a}}$ can not be finitely generated. Here, we show (directly) that ${\frak{a}}$ is not finitely generated. The proof of this fact goes along the same line of that of [@A Proposition 6.8], which gives a similar result over the absolute integral closure of $R$. We leave the details to the reader. Cohen-Macaulayness of Minimal Perfect Algebras ============================================== We begin this section with the following lemma. \[sh\] Let $(R,{\frak{m}})$ be a Noetherian local F-coherent domain which is either excellent or homomorphic image of a Gorenstein local ring. Then $R^{\infty}$ is balanced big Cohen-Macaulay. This is proved in [@Sh Theorem 3.10] when $R$ is a homomorphic image of a Gorenstein ring. The argument of [@Sh Theorem 3.10] based on Almost Ring Theory. Assume that $R$ is excellent. For each n, set $R_n:=\{x\in R^\infty\|x^{p^n}\in R\}$. The assignment $x\mapsto x^{p^n}$ shows that $R\simeq R_n$, which implies that $R_n$ is excellent. We recall that over excellent domains of prime characteristic one can use the colon capturing property of tight closure theory, see e.g. [@HH2 Theorem 1.7.4]. Let $\underline{x}:=x_1,\ldots,x_d$ be a system of parameters for $R$, where $d:=\dim R$ and let $r\in R^{\infty}$ be such that $rx_{i+1}=\sum_{j=1}^i r_jx_j$ for some $r_j\in R^{\infty}$. Then $r,r_j\in R_n$ for $n\gg 0$. So $r\in((x_1,\ldots, x_i)R_n :_{R_n} x_{i+1})$. Thus from Lemma \[sh1\], $$\begin{array}{ll} ((x_1,\ldots, x_i)R^{\infty} :_{R^{\infty}} x_{i+1})&=\bigcup_n((x_1,\ldots, x_i)R_n :_{R_n} x_{i+1})\\ &\subseteq \bigcup_n((x_1,\ldots, x_i)R_n)^\ast\\ &\subseteq ((x_1,\ldots, x_{i})R^\infty)^\ast\\ &=(x_1,\ldots, x_{i})R^\infty. \end{array}$$ Clearly, ${\frak{m}}R^{\infty}\neq R^{\infty}$. So, any system of parameters of $R$ is regular sequence on $R^{\infty}$, i.e., $R^{\infty}$ is balanced big Cohen-Macaulay. It is to be noted that in the case of non-Noetherian rings, tight closure may behave nicely, since from the above proof we find that it exhibits the colon capturing property. We recall the following definitions. Let ${\frak{a}}$ be an ideal of a ring $A$ and $M$ be an $A$-module. A finite sequence $\underline{x}:=x_{1},\ldots,x_{r}$ of elements of $A$ is called $M$-sequence if $x_i$ is a nonzero-divisor on $M/(x_1,\ldots, x_{i-1})M$ for $i=1,\ldots,r$ and $M/\underline{x}M\neq0$. The classical grade of ${\frak{a}}$ on $M$, denoted by ${\operatorname{c.grade}}_A({\frak{a}},M)$, is defined by the supremum length of maximal $M$-sequences in ${\frak{a}}$. The polynomial grade of ${\frak{a}}$ on $M$ is defined by $${\operatorname{p.grade}}_A({\frak{a}},M):=\underset{m\rightarrow\infty}{\lim}{\operatorname{c.grade}}_{A[t_1, \ldots,t_m]}({\frak{a}}A[t_1, \ldots,t_m],A[t_1,\ldots,t_m]\otimes_A M)$$ and we will use the following well-known properties of the polynomial grade. \[pro\] (see e.g. [@AT]) Let ${\frak{a}}$ be an ideal of a ring $A$ and $M$ an $A$-module. Then the following statements hold. 1. If ${\frak{a}}$ is finitely generated, then$${\operatorname{p.grade}}_A({\frak{a}},M)=\inf\{{\operatorname{p.grade}}_{A_{{\frak{p}}}}({\frak{p}}A_{{\frak{p}}},M_{{\frak{p}}})\|{\frak{p}}\in {\operatorname{V}}({\frak{a}})\cap{\operatorname{Supp}}_A M\}.$$ 2. Let $\Sigma$ be the family of all finitely generated ideals ${\frak{b}}\subseteq{\frak{a}}$. Then$${\operatorname{p.grade}}_A({\frak{a}},M)=\sup\{{\operatorname{p.grade}}_A({\frak{b}},M):{\frak{b}}\in\Sigma\}.$$ 3. ${\operatorname{p.grade}}_A({\frak{a}},M)\leq{\operatorname{ht}}_{M}({\frak{a}}).$ Now, we are ready to prove the following: \[main\] Let $R$ be a Noetherian local F-coherent domain which is either excellent or homomorphic image of a Gorenstein local ring. Then ${\operatorname{ht}}({\frak{a}})={\operatorname{p.grade}}_{R^{\infty}}({\frak{a}}, R^{\infty})$ for all ideal ${\frak{a}}$ of $R^{\infty}$. Let $\underline{x}:=x_1,\ldots,x_d$ be a system of parameters for $R$, where $d:=\dim R$. By Lemma \[sh\], $\underline{x}$ is regular sequence on $R^{\infty}$. So $$d\leq{\operatorname{p.grade}}(\underline{x}R^{\infty},R^{\infty})\leq{\operatorname{p.grade}}({\frak{m}}_{R^{\infty}},R^{\infty})\leq{\operatorname{ht}}({\frak{m}}_{R^{\infty}})=d.$$This shows that ${\operatorname{p.grade}}({\frak{m}}_{R^{\infty}},R^{\infty})={\operatorname{ht}}({\frak{m}}_{R^{\infty}})$ for the maximal ideal ${\frak{m}}_{R^{\infty}}$ of $R^{\infty}$. At first, we assume that ${\frak{a}}$ is finitely generated and let $P$ be a prime ideal of $R^{\infty}$ such that $P\supseteq {\frak{a}}$. Set ${\frak{p}}:=P\cap R$. Take $x\in(R_{{\frak{p}}})^{\infty}$. Then $x^{p^n}\in R_{{\frak{p}}}$ for some $n$, where $p$ is the characteristic of $R$. Thus $x^{p^n}=a/b$ for some $a\in R$ and $b \in R\setminus{\frak{p}}$. Consider $\frac{a^{1/p^n}}{b^{1/p^n}}$ as an element of $R^{\infty}_P$. Let $m\geq n$. Then $\frac{a^{1/p^n}}{b^{1/p^n}}=(\frac{a^{1/p^{m-n}}}{b^{1/p^{m-n}}})^{1/p^m}$. Keep it in the mind that $R$ is of prime characteristic. If we assume that $a/b=c/d$ then $\frac{a^{1/p^n}}{b^{1/p^n}}=\frac{c^{1/p^n}}{d^{1/p^n}}$. Putting these facts together, we see that the assignment $x\mapsto \frac{a^{1/p^n}}{b^{1/p^n}}$ is independent of the presentation of $x$ and of the choice of $n$. So, it defines a ring homomorphism $\varphi: (R_{{\frak{p}}})^{\infty} \to R^{\infty}_P$. Clearly, $\ker \varphi =0$. Let $x\in R^{\infty}_P$. Then $x=\alpha /\beta$, where $\alpha\in R^{\infty}$ and $\beta\in R^{\infty}\setminus P$. Take $n$ to be such that $a:=\alpha^{p^n}\in R$ and $b:=\beta^{p^n}\in R\setminus {\frak{p}}$. Set $c:=a/b\in R_{{\frak{p}}}$. Consider $\gamma:=c^{1/p^n}\in (R_{{\frak{p}}})^{\infty}$. Then $\varphi(\gamma)=x$, and thus $\varphi$ is an isomorphism. We recall that $R_{{\frak{p}}}$ is F-coherent. Clearly, $R_{{\frak{p}}}$ is either excellent or homomorphic image of a Gorenstein local ring. By the case of maximal ideals, $${\operatorname{p.grade}}(PR^{\infty}_P,R^{\infty}_P)={\operatorname{p.grade}}({\frak{m}}_{(R_{{\frak{p}}})^{\infty}},(R_{{\frak{p}}})^{\infty})={\operatorname{ht}}({\frak{p}})={\operatorname{ht}}(P).$$ This fact along with Lemma \[pro\] yields that $$\begin{array}{ll} {\operatorname{p.grade}}({\frak{a}},R^{\infty})&=\inf\{{\operatorname{p.grade}}(PR^{\infty}_P,R^{\infty}_P)\|P\in {\operatorname{V}}({\frak{a}})\}\\ &=\inf\{{\operatorname{ht}}(P)\|P\in {\operatorname{V}}({\frak{a}})\}\\&={\operatorname{ht}}({\frak{a}}), \end{array}$$ which shows that ${\operatorname{p.grade}}({\frak{a}},R^{\infty})={\operatorname{ht}}({\frak{a}})$ for every finitely generated ideal ${\frak{a}}$ of $R^{\infty}$. Finally we assume that ${\frak{a}}$ is an arbitrary ideal of $R^{\infty}$ and let $\Sigma$ be the family of all finitely generated ideals ${\frak{b}}\subseteq{\frak{a}}$. We bring the following claim: Claim: One has ${\operatorname{ht}}({\frak{a}})=\sup\{{\operatorname{ht}}({\frak{b}}):{\frak{b}}\in \Sigma\}$. To see this, let $P\in {\operatorname{Spec}}(R^{\infty})$ be such that ${\operatorname{ht}}(P)={\operatorname{ht}}({\frak{a}})$ and set ${\frak{p}}:=P\cap R$ and we already have that $R^{\infty}_P=(R_{{\frak{p}}})^{\infty}$. Set $n:={\operatorname{ht}}({\frak{p}})={\operatorname{ht}}(P)={\operatorname{ht}}({\frak{a}}\cap R)$. Due to [@BH Theorem A.2], there exists a sequence $\underline{x}:=x_1,\ldots,x_n$ of elements of ${\frak{a}}\cap R$ such that ${\operatorname{ht}}(x_1,\ldots,x_i)R=i$ for all $1\leq i \leq n$. Since $R$ is catenary, $\underline{x}$ is part of a system of parameters for $R$. By Lemma \[sh\], $\underline{x}$ is a regular sequence on $R^{\infty}$. So $${\operatorname{ht}}(P)\geq {\operatorname{ht}}(\underline{x}R^{\infty})\geq{\operatorname{p.grade}}(\underline{x}R^{\infty},R^{\infty})=n= {\operatorname{ht}}({\frak{p}})={\operatorname{ht}}(P),$$ which shows that ${\operatorname{ht}}(\underline{x}R^{\infty})={\operatorname{ht}}({\frak{a}})$. This completes the proof of the claim. From the results of finitely generated ideals, we have that ${\operatorname{p.grade}}({\frak{b}},R^{\infty})={\operatorname{ht}}({\frak{b}})$ for all ${\frak{b}}\in\Sigma$. Thus from Lemma \[pro\] and from the above claim we see that $$\begin{array}{ll} {\operatorname{p.grade}}({\frak{a}},R^{\infty})&=\sup\{{\operatorname{p.grade}}({\frak{b}},R^{\infty}):{\frak{b}}\in\Sigma\}\\ &=\sup\{{\operatorname{ht}}({\frak{b}}):{\frak{b}}\in \Sigma\}\\&={\operatorname{ht}}({\frak{a}}),\\ \end{array}$$ and this is precisely what we wish to prove. Let $M$ be an $A$-module. Recall that a prime ideal ${\frak{p}}$ is weakly associated to $M$ if ${\frak{p}}$ is minimal over $(0 :_{A} m)$ for some $m\in M $. Adopt the assumption of Theorem \[main\] and let ${\frak{a}}$ be an ideal of $R^{\infty}$ which is generated by ${\operatorname{ht}}({\frak{a}})$ elements. Then all the weakly associated prime ideals of $R^{\infty}/ {\frak{a}}$ have the same height. See [@AT Corollary 4.6] and its proof. \[rem\] $\mathrm{(i)}$: Concerning the proof of Theorem \[main\], we can ask the following question which has a positive answer for several classes of commutative rings such as valuation domains. Let $A$ be a commutative ring and ${\frak{a}}$ an ideal of $A$. Let $\Sigma$ be the family of all finitely generated ideals ${\frak{b}}\subseteq{\frak{a}}$. Is ${\operatorname{ht}}({\frak{a}})=\sup\{{\operatorname{ht}}({\frak{b}}):{\frak{b}}\in \Sigma\}$? $\mathrm{(ii)}$: Let A be a ring with the property that ${\operatorname{p.grade}}_A({\frak{m}};A) = {\operatorname{ht}}({\frak{m}})$ for all maximal ideals ${\frak{m}}$ of $A$. One may ask whether ${\operatorname{p.grade}}_A({\frak{a}};A) = {\operatorname{ht}}({\frak{a}})$ holds for all ideals ${\frak{a}}$ of $A$. In view of [@AT Example 3.11], this is not true in general. Dagger Closure and Big Cohen-Macaulay algebras ============================================== In this section we extend the notion of dagger closure to the submodules of a finitely generated module over a Noetherian local domain $(R,{\frak{m}})$ and we present Corollary \[main2\]. \[def:almost closure of an ideal\] Let $A$ be a local algebra with a normalized valuation $v:A\rightarrow {{\mathbb R}}\cup\{\infty\}$ and $M$ be an $A$-module. Consider a submodule $N\subset M$. We say $x\in N^{v}_M$ if for every $\epsilon> 0$, there exists $a\in A$ such that $v(a)< \epsilon$ and $ax\in N$. \[almcm\] Let $A$ be an algebra over a Noetherian local domain $(R,{\frak{m}})$ of $\dim d$. Assume that $A$ is equipped with a normalized value map $v : A\to {{\mathbb R}}\cup\{\infty\}$. From [@RSS], we recall that $A$ is called almost Cohen-Macaulay if for $i=1,\ldots, d$, each element of $((x_{1},\ldots, x_{i-1})A:_{A}x_{i})/(x_{1},\ldots,x_{i-1})A$ is annihilated by elements of sufficiently small order with respect to $v$ for all system of parameters $x_1,\ldots, x_d$ of $A$. \[def\] Let $(R,{\frak{m}})$ be a Noetherian local domain and let $A$ be a local domain containing $R$ with a normalized valuation $v:A\rightarrow {{\mathbb R}}\cup\{\infty\}$. For any finitely generated $R$-module $M$ and for its submodule $N$ we define submodule $N_{M}^{\bold{v}}$ such that $x\in N_{M}^{\bold{v}}$ if $x\otimes 1 \in {\operatorname{im}}(N\otimes A\to M\otimes A)_{M\otimes A}^{v}$. Let $A$ be a perfect domain and $\frak a$ be a nonzero radical ideal of $A$. Then $\frak a^v=A$. To see this, let $x$ be a nonzero element of $\frak a$. Since $v(x^{1/n})=v(x)/n$ and $x^{1/n}\in \frak a$, the conclusion follows. \[pr\] Let $(R,{\frak{m}})$ be a Noetherian local domain and let $A$ be a local domain containing $R$ with a normalized valuation $v:A\rightarrow {{\mathbb R}}\cup\{\infty\}$. Let $M$, $M'$ be finitely generated $R$-modules. Consider the submodules $N$, $W$ of $M$. Then the following statements are true: 1. $N^{\bold{v}}_M$ is a submodule of $M$ containing $N$. 2. $(N^{\bold{v}}_M)^{\bold{v}}_M= N^{\bold{v}}_M$. 3. If $N\subset W\subset M$, then $N^{\bold{v}}_M\subset W^{\bold{v}}_M$. 4. Let $f:M\to M'$ be a homomorphism. Then $f(N^{\bold{v}}_M)\subset f(N)^{\bold{v}}_{M'}$. 5. If $N^{\bold{v}}_M= N$, then $0^{\bold{v}}_{M/N}= 0$. 6. We have $0^\bold{v}_R=0$ and ${\frak{m}}^\bold{v}_R={\frak{m}}$. In addition to, if $A$ is almost Cohen-Macaulay, then the following is true: 1. Let $x_1, \ldots , x_{k+1}$ be a partial system of parameters for $R$, and let $J = (x_1, \ldots , x_k)R$. Suppose that there exists a surjective homomorphism $f:M\to R/J$ such that $f(u) = \bar{x}_{k+1}$, where $\bar{x}$ is the image of $x$ in $R/J$. Then $(Ru)^{\bold{v}}_M\cap \ker f \subset (Ju)^{\bold{v}}_M$. The proof is straightforward and we leave it to the reader. In [@Ho], Hochster showed the existence of a big Cohen-Macaulay algebra from the existence of an almost Cohen-Macaulay algebra in dimension 3. The following Corollary extends this result, by proving the existence of big Cohen-Macaulay modules from the existence of almost Cohen-Macaulay algebras in any dimension $d$. \[main2\] For a complete Noetherian local domain, if it is contained in an almost Cohen-Macaulay domain, then there exists a balanced big Cohen-Macaulay module over it. Due to [@D Theorem 3.16], we know that there exists a list of seven axioms for a closure operation defined for finitely generated modules over complete local domains. By the main result of [@D], one can see that a closure operation satisfying those axioms implies the existence of a balanced big Cohen-Macaulay module over the base ring. In view of Proposition \[pr\], we see that the closure operation defined in Definition \[def\] satisfies all the seven axioms of [@D Theorem 3.16]. This completes the proof. The above result can be used to prove [@RSS Proposition 1.3] by a completely different argument, since the existence of almost Cohen-Macaulay domain impiles the existence of big Cohen-Macaulay module. Concluding Remarks ================== For Noetherian local domain $R$ of $\dim d$, the above results can be extended to an almost Cohen-Macaulay $R^{+}$-algebra where $R^{+}$ be its integral closure in an algebraic closure of its fraction field. Fixed a local $R^{+}$-algebra $A$ and let $M$ be an $A$-module. Consider a submodule $N\subset M$. In view of Definition \[def:almost closure of an ideal\], define $ N^{v}_M$ via a normalized valuation $v:R^{+}\rightarrow {{\mathbb R}}\cup\{\infty\}$. For a finitely generated $R$-module $M$ and its submodule $N$ we define submodule $N_{M}^{\bold{v}}$ such that $x\in N_{M}^{\bold{v}}$ if $x\otimes 1 \in {\operatorname{im}}(N\otimes A\to M\otimes A)_{M\otimes A}^{v}$. From [@S1], we recall that $A$ is almost Cohen-Macaulay if for every system of parameters $x_1,\ldots, x_d$, $((x_1,\ldots, x_{i-1})A:_Ax_i)/(x_1,\ldots, x_{i-1})A$ is almost zero $R^{+}$-module for $i=1,\ldots,d$ and $A/{\frak{m}}A$ is not almost zero. Let $(R,{\frak{m}})$ be a Noetherian local domain and let $R^{+}$ be equipped with a normalized valuation $v:R^{+}\rightarrow {{\mathbb R}}\cup\{\infty\}$. Consider a local $R^{+}$-algebra $A$ and for every $I\subset R$, we define $I^{\bold{v}}$ as above. Under this situation, closure operation satisfies all the properties given in Proposition \[pr\] if and only if $A$ is an almost Cohen-Macaulay $R^{+}$-algebra where $R^{+}$ is contained in it as a subdomain. We show $0^{\bold{v}}= 0$ implies $R^{+}\to A$ is injective. To see this, take $0\neq a\in R^{+}$ such that its image in $A$ is zero. Since $a$ is integral over $R$, there exists a minimal monic expression $a^n+ r_1a^{n-1}+\cdots +r_n= 0$, where each $r_i\in R$ with $r_n\neq 0$. Take the image of the expression in $A$ which gives that the image of $r_n$ is zero in $A$. So $r_n\in 0^{\bold{v}}= 0$. This gives $r_n= 0$. So we arrive at a contradiction and $a$ is zero in $R^{+}$. By using straightforward modification of the proof of Corollary \[main2\], we get the rest of the claim. The proof of the converse is left it to the reader. The following result provides an example of an almost Cohen-Macaulay $R^{+}$-algebra where $R^{+}$ is contained in it as a subdomain. We recall that an $R$-algebra $T$ is called a seed, if it maps to a big Cohen-Macaulay $R$-algebra (see [@D1]). \[rem\] Let $(R,{\frak{m}})$ be a Noetherian complete local domain in mixed characteristic $p>0$, with a system of parameters $p, x_2,\ldots, x_d$. Let $B$ be as [@S1 Theorem 5.3]. Then the following are equivalent. 1. $B/{\frak{m}}B$ is not almost zero when viewed as $R^{+}$-module. 2. $p^{\epsilon}\notin (p, x_2,\ldots, x_d)B$ for some rational number $\epsilon>0$. $\mathrm{(i)}\Rightarrow\mathrm{(ii)}$: Clearly ${\frak{m}}B$ does not contain elements of $R^{+}$ of arbitrary small order. If $p^{\epsilon}\in (p, x_2,\ldots, x_d)B$ for every rational number $\epsilon$ then $p^{1/k}\in (p, x_2,\ldots, x_d)B$ for every positive integer $k$. Since $v(p)< \infty$, $p^{1/k}$ are the elements of $R^{+}$ of arbitrarily small order and ${\frak{m}}B$ contains all of them. This is a contradiction. $\mathrm{(ii)}\Rightarrow\mathrm{(i)}$: From [@S1 Corollary 6.1] it follows that $R^{+}$ is a seed and using similar argument of [@D1 Example 5.2] we find that $R^{+}$ is actually a minimal seed. Thus $R^{+}$ can be thought as a subdomain of $B$. Let $\{x_{\lambda}\}= B-R^{+}$. Take $C= R^{+}[\{X_{\lambda}\}]$ where $X_{\lambda}$’s are the indeterminates. One can extend the valuation $R\to {{\mathbb Z}}\cup \{\infty\}$ to $C\to {{\mathbb R}}\cup \{\infty\}$ such that $v$ is non negative on $C$ and positive on non-units of $C$. Also, we choose values of $X_{\lambda_i}$’s greater than some $N> 0$. For $B= C/JC$, we have $y\in JC$ implies $v(y)> N> 0$, and since ${\frak{m}}$ is also finitely generated same is true for ${\frak{m}}C$. Thus $C/(J+{\frak{m}})C$ is not almost zero viewing as $C$-module. Since $B/{\frak{m}}B= C/(J+{\frak{m}})C$, we find that $B/{\frak{m}}B$ is not almost zero viewing as $C$-module. Thus $B/{\frak{m}}B$ is not almost zero viewing as $R^{+}$-module. In [@S1], it has been asked that whether the $R^+$-algebra $B$ which satisfies conditions of Theorem 5.3 (there) is almost Cohen-Macaulay or not (i.e $B/mB$ is almost zero or not). From Corollary 5.2 and Proposition \[rem\], we get the following: Let $B$ be a weakly almost Cohen-Macaulay $R^+$-algebra satisfies Theorem 5.3 of [@S1]. Then $B$ is almost Cohen-Macaulay $R^+$-algebra (i.e. $B/mB$ is almost zero) if and only if it maps to big Cohen-Macaulay $R^+$-algebra. It is to be noted that here $R^+$ is contained in $B$ as a subdomain. We would like to thank K. Shimomoto for his comments on the earlier version of this paper. We also thank the referee for careful reading of the paper and for the valuable suggestions. [99]{} M. Asgharzadeh, Homological properties of the perfect and absolute integral closure of Noetherian domains, Math. Annalen [348]{} (2010), 237–-263. M. Asgharzadeh and M. Tousi, On the notion of Cohen-Macaulayness for non-Noetherian rings, J. Algebra [[322]{}]{} (2009), 2297–2320. M. Asgharzadeh and K. Shimomoto, Almost Cohen-Macaulay and almost regular algebras via almost flat extensions, to appear in J. Commutative Algebra. W. Bruns and J. Herzog, Cohen-Macaulay rings, Cambridge University Press [[39]{}]{}, Cambridge, (1998). H. Brenner and A. Stabler, Dagger closure and solid closure in graded dimension two, Preprint (2011), Arxiv math.AG/1104.3748. G. Dietz, A characterization of closure operations that induce big Cohen-Macaulay modules, Proc. AMS [[138]{}]{} (2010), 3849–3862. G. Dietz, Big Cohen-Macaulay algebras and seeds, Trans. Amer. Math. Soc. [[359]{}]{}, (2007), 5959–-5989. J. Elliott, Prequantales and applications to semistar operations and module systems, arXiv:1101.2462 \[math AC\]. M. Hochster, Big Cohen-Macaulay algebras in dimension three via Heitmann’s theorem, J. Algebra [[254]{}]{} (2002), 395–408. M. Hochster and C. Huneke, Tight closure and elements of small order in integral extensions, J. Pure and Applied Algebra [[71]{}]{} (1991), 233-247. M. Hochster and C. Huneke, Tight closure in equal characteristic zero, Preprint. P. Roberts, A. Singh and V. Srinivas, Annihilators of local cohomology in characteristic zero, Illinois J. Math. [[51]{}]{} (2007), 237-254. K. Shimomoto, F-coherent rings with applications to tight closure theory, Journal of Algebra [[338]{}]{}, (2011), 24–34. K. Shimomoto, Almost Cohen-Macaulay algebras in mixed characteristic via Fontaine rings, to appear in Illinois J. Math.	{ "pile_set_name": "ArXiv" }
--- author: - \| [^1] [^2], Chris Schroeder, Joseph Wasem\ Physical Sciences Directorate, Lawrence Livermore National Laboratory\ Livermore, California 94550, USA\ E-mail: title: 'Neutron-antineutron oscillations on the lattice' --- ł Ł ø Ø § \#1 \#1[[(\[\#1\])]{}]{} Introduction ============ One unanswered mystery of the universe is the process that led to the abundance of observed baryons as compared to their antibaryon counterparts. The source of this baryon number violation, which is expected to come from beyond the Standard Model (BSM) physics, can be realized in low-energy processes such as proton decay (if baryon number is violated by 1 unit) or transitions between neutrons and antineutrons (if baryon number is violated by 2 units). The latter case, often referred to as neutron-antineutron oscillation (akin to neutral meson mixing), proves to be an intriguing scenario when considering the usual sphaleron picture of baryogenisis (which violates baryon number, $B$, and lepton number, $L$, but conserves $B-L$) coupled with Majorana neutrinos [@Mohapatra:1980qe] (whose transition between neutrinos and antineutrinos leads to $\Delta L = 2$). Additionally, neutron-antineutron oscillations do not suffer from kinematic suppressions that can restrict proton decay if there is little overlap with the initial state proton and final state electron or muon [@Nussinov:2001rb]. To that end, neutron-antineutron oscillations have been explored experimentally with intriguing prospects for $\mathcal{O}(1000)$ improvements in upcoming experimental efforts [@Proj_X]. Any discussion of neutron-antineutron oscillations starts with assuming the existence of some BSM process that leads to a $\Delta B = 2$ operator in the low-energy effective field theory. This operator will lead to off-diagonal elements of the Hamiltonian of the neutron-antineutron system $$\begin{aligned} H = \begin{pmatrix} E_n & \delta m \\ \delta m & E_{\bar{n}} \end{pmatrix}=\begin{pmatrix} E+V & \delta m \\ \delta m & E-V \end{pmatrix},\end{aligned}$$ where $V$ is the potential difference between the neutron and the antineutron (a magnetic field can lead to a non-zero $V$ since the magnetic moments have opposite signs) and $V=0$ in a free system. Upon solving the Schrödinger equation for the system, one finds the transition probability between neutrons and antineutrons given by $$P_{n\rightarrow \bar{n}}(t) = \frac{\delta m^2}{\delta m^2 + V^2}\sin^2\Big[\sqrt{\delta m^2 + V^2} \ t \Big].$$ While this equation is true for a given $V$, it is standard to define the period of free neutron oscillations due to the BSM physics as $$\tau_{n\overline{n}} = \frac{1}{\delta m}.$$ The value for $\tau_{n\overline{n}}$ greatly depends on which BSM scenario is being explored. It has been estimated that a bound of $\tau_{n\overline{n}} \gtrsim 10^{10}-10^{11}\ \text{seconds}$ is sufficient to rule out many of the current models[@Mohapatra:2009wp]. For example, TeV-scale seesaw mechanisms for neutrino masses in $SU(2)_L \times SU(2)_R \times SU(4)_c$ are expected to be ruled out at $\tau_{n\overline{n}} \gtrsim 10^{10}-10^{11}\ \text{seconds}$ [@Babu:2008rq]; and $SO(10)$ seesaw mechanisms with adequate baryogenisis, at $\tau_{n\overline{n}} \gtrsim 10^{9}-10^{12}\ \text{seconds}$ [@Babu:2012pp]. Current experimental limits can even restrict extra-dimensional models with new particles with masses below a TeV [@Nussinov:2001rb; @Winslow:2010wf]. It should be emphasized that these are order of magnitude estimates with the QCD input coming from naïve dimensional analysis. Future estimates will require rigorous and precise lattice calculations to keep pace with experimental precision. The detection mechanism for these transitions is the cold annihilation of a newly formed antineutron with a nearby neutron (for more details, see W. M. Snow’s plenary at PXPS 2012 [@Proj_X]). The primary channel for this cold annihilation is $n\overline{n} \rightarrow 5\pi$, and this unique signature allows for experimental signals with little or no background. Generally, there are two sets of experimental searches. The first, which comes for free with large proton decay detectors such as Super-K, is based on neutron-antineutron annihilation within nuclei. Naïvely, one might expect this to occur quite frequently, as the number of nuclei far exceeds the expected bound $\tau_{n\overline{n}} \gtrsim 10^{11}$; however, the oscillation period within nuclei is highly suppressed, with a magnitude of roughly [@Friedman:2008es] $$\tau_{Nucl}=(3 \times 10^{22})\frac{\tau_{n\overline{n}}^2}{\text{sec}}.$$ As a result, the bound is suppressed compared to the free expectation, and one must rely on model estimations and extrapolations to extract it. To date, the most stringent bound from experiments of this kind, $\tau_{n \overline{n}} > 3.5 \times 10^8 \ \text{seconds}$, comes from Super-K (2011) [@Super-K]. The second type of experiment explores the annihilation of free, cold neutrons with a target after a significant time of flight. This type of experiment is free of the model-dependent estimations required for annihilations within nuclei and allows for greater control of systematics. To date, the most stringent bound comes from the ILL experiment (1993): $\tau_{n \overline{n}} > 0.86 \times 10^8 \ \text{seconds}$ [@BaldoCeolin:1994jz]. A factor of $\mathcal{O}(1000)$ increase is estimated for future experiments of this kind, but the bounds to rule out various BSM theories could be altered significantly depending on QCD enhancement or suppression of the neutron-antineutron matrix elements. Oscillations and matrix elements ================================ The observed value of the mixing arises from three inputs $$\frac{1}{\tau_{n\overline{n}}}=\delta m = c_{BSM}(\mu_{BSM},\mu_W)c_{QCD}(\mu_W,\Lambda_{QCD}) \langle \overline{n}\| \mathcal{O} \| n \rangle,$$ where $c_{BSM}$ is the running of the BSM theory to the weak interaction scale, $c_{QCD} $ is the QCD running from the weak to the nuclear scale, and $\langle \overline{n}\| \mathcal{O} \| n \rangle$ is the non-perturbative matrix element mixing the neutron and antineutron states. The one-loop perturbative QCD running, $c_{QCD}$, is known [@Winslow:2010wf; @Ozer:1982qh], and $c_{BSM}$ has been calculated for multiple theories [@Nussinov:2001rb; @Babu:2008rq; @Babu:2012pp; @Winslow:2010wf]. The operator $\mathcal{O}$ contains two up quarks and four down quarks and is composed of three pairs of quarks from the possible forms $$u^T C u \quad , \quad u^T C d \quad , \quad d^T C d \quad,$$ where $C$ is the charge conjugation matrix. Additionally, these terms always come in chiral pairs, $$u^T_L C d_L \quad , \quad u^T_R C d_R \quad,$$ since the mixed chirality terms are zero. Lastly, these operators are invariant under color symmetry, $SU(3)_c$, which leads to two color tensors $$\begin{aligned} \Gamma^{s}_{ijklmn}&=& \epsilon_{mik}\epsilon_{njl}+\epsilon_{nik}\epsilon_{mjl}+\epsilon_{mjk}\epsilon_{nil}+\epsilon_{njk}\epsilon_{mil} ,\quad\Gamma^{a}_{ijklmn}= \epsilon_{mij}\epsilon_{nkl}+\epsilon_{nij}\epsilon_{mkl},\end{aligned}$$ where $i,j,k,l,m,n$ are color indices. These three conditions lead to three types of operators [@Rao:1982gt]: $$\begin{aligned} \mathcal{O}^1_{\chi_1\chi_2\chi_3}&=&(u_{i\chi_1}^{T} C u_{j\chi_1})(d_{k\chi_2}^{T} C d_{l\chi_2})(d_{m\chi_3}^T C d_{n\chi_3})\Gamma^{s}_{ijklmn},\nonumber\\ \mathcal{O}^2_{\chi_1\chi_2\chi_3}&=&(u_{i\chi_1}^{T} C d_{j\chi_1})(u_{k\chi_2}^{T} C d_{l\chi_2})(d_{m\chi_3}^T C d_{n\chi_3})\Gamma^{s}_{ijklmn},\nonumber\\ \mathcal{O}^3_{\chi_1\chi_2\chi_3}&=&(u_{i\chi_1}^{T} C d_{j\chi_1})(u_{k\chi_2}^{T} C d_{l\chi_2})(d_{m\chi_3}^T C d_{n\chi_3})\Gamma^{a}_{ijklmn},\end{aligned}$$ where $\chi_i=L,R$. At first glance, there would appear to be 24 independent operators, but there are several additional symmetries. The first set of symmetries due to the flavor structure is $$\begin{aligned} \label{Sym_1} \mathcal{O}^1_{\chi_1 L R}&=&\mathcal{O}^1_{\chi_1 R L},\quad \mathcal{O}^{2,3}_{L R \chi_3}=\mathcal{O}^{2,3}_{R L \chi_3},\end{aligned}$$ which reduces the set to 18 independent operators. An additional symmetry that emerges from antisymmetrizing pairs of epsilon tensors over four indices leads to (with $\sigma=L,R$ and $\rho=L,R$)[@Caswell:1982qs] $$\label{Sym_2} \mathcal{O}^2_{\sigma \sigma \rho} - \mathcal{O}^1_{\sigma \sigma \rho} = 3\mathcal{O}^3_{\sigma \sigma \rho},$$ which reduces the set to 14 operators. In addition to enforcing $SU(3)_c$, it is also expected that the operators should be invariant under $SU(2)_L \otimes U(1)_Y$. This gauge symmetry and the symmetries in Eq. leave only six operators: $$\begin{aligned} \ \mathcal{P}_1&=&\mathcal{O}^1_{R R R},\quad\mathcal{P}_2=\mathcal{O}^2_{R R R},\quad\mathcal{P}_3=\mathcal{O}^3_{R R R},\nonumber\\ \quad\mathcal{P}_4&=&2\mathcal{O}^3_{L R R},\quad\mathcal{P}_5=4\mathcal{O}^3_{L L R},\quad\mathcal{P}_6=4(\mathcal{O}^1_{L L R}-\mathcal{O}^2_{L L R}).\end{aligned}$$ We will present results for these operators; however, including the symmetry in Eq. leads to the conditions, $$\begin{aligned} \label{check} \mathcal{P}_6&=&-3\mathcal{P}_5,\quad\quad\mathcal{P}_2-\mathcal{P}_1=3\mathcal{P}_3,\end{aligned}$$ which reduces the number of independent operators to four. We will use these last two equations for a consistency check of our calculation. Lattice formalism and contraction details ========================================= The mechanism to extract the neutron-antineutron matrix elements follows the common practice of taking ratios of three-point to two-point correlation functions. In particular, the three correlation functions of interest and their large Euclidean time behavior are given by $$\begin{aligned} C_{NN}(t) = \langle N(t) \overline{N}(0) \rangle &\rightarrow& \|\langle N\|n\rangle \|^2 e^{-m_n t}, \quad C_{\overline{N}\overline{N}}(t) = \langle \overline{N}(t) N(0) \rangle \rightarrow \|\langle \overline{N}\|\overline{n}\rangle \|^2 e^{-m_n t}, \nonumber\\ C_{\overline{N}\mathcal{O}N}(t_1,t_2) &=& \langle N(t_2)\mathcal{O}(0) \overline{N}(-t_1) \rangle \rightarrow \langle \overline{n}\|\overline{N}\rangle \langle N\|n\rangle e^{-m_n (t_1+t_2)} \langle n \|\mathcal{O}\|\overline{n}\rangle.\end{aligned}$$ The desired quantity of interest, $\langle n \|\mathcal{O}\|\overline{n}\rangle$, is the long time asymptote of a combination of these correlation functions, $$\label{ratio} \mathcal{R}=\frac{C_{\overline{N}\mathcal{O}N}(t_1,t_2)}{C_{\overline{NN}}(t_1+t_2)}\Bigg[\frac{C_{NN}(t_1)C_{\overline{NN}}(t_2)C_{\overline{NN}}(t_1+t_2)}{C_{\overline{NN}}(t_1)C_{NN}(t_2)C_{NN}(t_1+t_2)}\Bigg]^\frac{1}{2}\rightarrow \langle \overline{n}\| \mathcal{O} \| n \rangle.$$ ![Comparison of neutron-antineutron three-point contractions (left) to typical bilinear three-point contractions (right). One propagator is required for a measurements at all $(t_1,t_2)$ for the left diagram and two propagators are required for one measurement at a single $t_1$-value on the right diagram.[]{data-label="fig:contract"}](NNBar_Contract_Crop.pdf "fig:"){width="3in"} ![Comparison of neutron-antineutron three-point contractions (left) to typical bilinear three-point contractions (right). One propagator is required for a measurements at all $(t_1,t_2)$ for the left diagram and two propagators are required for one measurement at a single $t_1$-value on the right diagram.[]{data-label="fig:contract"}](Three_Point_Crop.pdf "fig:"){width="2.1in"} The six-quark neutron-antineutron three-point correlation function has several key advantages over typical bi-linear or four-quark nucleon operators. First, if the starting point for the propagator is at the operator insertion (as shown in Fig. \[fig:contract\]), only one propagator is needed per measurement, whereas the typical nucleon three-point function requires two propagators, one that starts from the source and one that starts from the operator. Second, because the propagator starts at the operator, one can acquire all the source-operator separations (given by $t_1$ in Fig. \[fig:contract\]) and operator-sink separations (given by $t_2$ in Fig. \[fig:contract\]), which allows for a two-dimensional analysis to quantify the excited state effect. Alternatively, typical three point functions require far more computational resources to quantify excited state effects. Lastly, the neutron-antineutron matrix element contains no disconnected or quark loop contractions, which removes the need for costly all-to-all propagators. One possible disadvantage is that multiplying six propagators together (as done for the neutron-antineutron correlator) could increase the signal-to-noise degradation as compared to bilinear or four-quark matrix elements; however, we find a reasonably good signal-to-noise ratio, as shown in Fig. \[fig:R\_Plots\]. Lattice Details =============== The lattice calculations were performed with Chroma [@Edwards:2004sx] using the $32^3 \times 256$ anisotropic clover-Wilson lattices defined in Ref. [@Lin:2008pr] with a pion mass of 390 MeV. The temporal and spatial lattice spacings are roughly 0.035 and 0.123 fm, respectively, and the total spatial extent is roughly 4 fm ($m_\pi L \sim 7.8$). For this preliminary calculation, we use a total of 159 configurations, each separated by 4 trajectories, to calculate 7268 propagators with Gaussian-smeared sources. Contractions of these propagators lead to the same number of measurements at all source-operator and operator-sink separations. Preliminary results =================== ![Plots of $\mathcal{R}$ vs. $t_2$ for six values of $t_1=5,10,15,20,25,30$. The large $t_1$ and $t_2$ behavior of $\mathcal{R}$ should approach the neutron-antineutron matrix elements of interest. For $t_1>30$ ($t_1\gtrsim1\ \text{fm}$), the plateau does not change appreciably, but signal-to-noise decreases.[]{data-label="fig:R_Plots"}](R_Plots_Crop.pdf){width="5in"} ![Two-dimensional plot of $\mathcal{R}$ (left) and $\|\mathcal{R}\|$ (right) as a function of $t_1$ and $t_2$ ranging from time slices 0 to 60. The lighter colors represent larger values of $\mathcal{R}$. The two-dimensional plateau is achieved for $10<t_2<25$ and $30<t_1<40$.[]{data-label="fig:2D_Plots"}](O1RRR_Both_Crop.pdf){width="5.1in"} The desired matrix elements, $\langle \overline{n} \| \mathcal{P}_i \| n \rangle$, can be extracted from the long Euclidean time behavior of Eq. . For each ratio $\mathcal{R}$, there are two time inputs, the source-operator separation ($t_1$) and the operator-sink separation ($t_2$). In Fig. \[fig:R\_Plots\], $\mathcal{R}$ for the $\mathcal{P}_1$ operator is plotted against $t_2$ for six different values of $t_1$. Two features stand out from these plots. First, there is a significant range of time slices where a signal can be extracted and the signal-to-noise degradation is not overly restrictive. Second, it is evident that there is significant excited state dependence as $t_1$ is varied (for example, the plateaux extracted for $t_1 = 10$ and $t_1=30$ are significantly different). For this reason, it is very important to use all information available to explore the full behavior of $\mathcal{R}$ as a function of both $t_1$ and $t_2$. In Fig. \[fig:2D\_Plots\], the 2D plot of $\mathcal{R}$ and $\|\mathcal{R}\|$ are plotted against $t_1$ and $t_2$. Again, it is clear that there is a significant amount of non-trivial behavior due to excited states. To that end, a 2D correlated fit has been performed over the time slices $10<t_2<25$ and $30<t_1<40$. For this preliminary calculation, systematic errors are estimated by adjusting 2D fit window $\pm 1$ on all sides. Operator Lattice Calculation ($10^{-5}\ \text{GeV}$) MIT Bag Model Calculation ($10^{-5}\ \text{GeV}$) ---------------------------------------------------- --------------------------------------------- --------------------------------------------------- $\langle \overline{n} \| \mathcal{P}_1 \| n \rangle$ $1.57\pm 0.85^{+0.25}_{-0.30}$ -6.56 $\langle \overline{n} \| \mathcal{P}_2 \| n \rangle$ $-0.20\pm 0.14^{+0.14}_{-0.12}$ 1.64 $\langle \overline{n} \| \mathcal{P}_3 \| n \rangle$ $-0.24\pm 0.26^{+0.10}_{-0.07}$ 2.73 $\langle \overline{n} \| \mathcal{P}_4 \| n \rangle$ $-0.02\pm 0.39^{+0.07}_{-0.18}$ -6.36 $\langle \overline{n} \| \mathcal{P}_5 \| n \rangle$ $0.34\pm 0.82^{+0.27}_{-0.57}$ 9.64 $\langle \overline{n} \| \mathcal{P}_6 \| n \rangle$ $-2.07\pm 1.10^{+1.28}_{-0.77}$ -28.92 : Table of results for neutron-antineutron matrix elements from the bare, non-renormalized lattice calculation and MIT bag model calculations [@Rao:1982gt].[]{data-label="tab:results"} The bare (unrenormalized) results for $\langle \overline{n} \| \mathcal{P}_i \| n \rangle$ are shown in Table \[tab:results\]. Eq. is satisfied exactly, configuration by configuration, but only stochastically in Table \[tab:results\] due to the bootstrapping in the analysis. The corresponding values calculated from the MIT bag model are also displayed for comparison. The magnitude of each operator as computed on the lattice is below that derived using the MIT bag model; however, it should be emphasized that the lattice results are very preliminary and still require renormalization factors. Systematic effects ================== The primary systematic uncertainty in comparing the results of Table \[tab:results\] to experiment is the unphysically large pion mass used. While it is not clear that an IR quantity such as the pion mass should dramatically effect the short distance six-quark vertex, it is a distinct possibility given that contractions of this system are similar to those for low-energy $NN$ scattering, where physical quark masses are expected to lead to a dramatic increase in the scattering length [@Beane:2006mx]. The second source of systematic uncertainty is the lattice cutoff (i.e., discretization) and matching the lattice regularization to the usual $\overline{MS}$ scheme used in the perturbative running [@Ozer:1982qh; @Winslow:2010wf]. Generically, operators of interest might mix with lower dimensional operators with the same symmetries, leading to diverging $1/a$ corrections (where $a$ is the lattice spacing). However, this is not an issue for these operators since the lowest dimension operator that can lead to a $\Delta B=2$ interaction requires six quarks. Regardless, there are expected to be $\mathcal{O}(a)$ corrections and renormalization coefficients that should be quantified. The third systematic which was clear from Fig. \[fig:R\_Plots\] and Fig. \[fig:2D\_Plots\] is excited state contamination. The calculation of the six-quark neutron-antineutron correlator gives us a unique view of these contaminations as a 2D function in $t_1$ and $t_2$, which is difficult to come by for any other nucleon three-point function. For this reason, we should be able to accurately quantify these contaminations. Finally, finite volume effects should be quantified as well, though their impact is expected to be insignificant given the $m_\pi L \sim 7.8$ lattice used. Future prospects ================ We are in the process of taking several steps to improve upon this very preliminary work: we are extending the calculation presented here as well as repeating it for a lighter, $240 \ \text{MeV}$ pion mass at the same volume and for the current pion mass with a smaller, $2.5 \ \text{fm}$ spatial extent with a larger ensemble. We are exploring perturbative and non-perturbative lattice renormalization to properly match onto the perturbative QCD running previously calculated. We are also refining our analysis procedures to better quantify excited state effects. Within the next year or two, we hope to carry out this calculation both with physical pion masses and with a chiral fermion discretization (domain-wall fermions). Both calculations are numerically expensive, but within reach of the LLNL 20 PetaFlops Sequoia BG/Q. [99]{} R. N. Mohapatra and R. E. Marshak, Phys. Rev. Lett. [44]{}, 1316 (1980). S. Nussinov and R. Shrock, Phys. Rev. Lett. [88]{}, 171601 (2002) \[hep-ph/0112337\]. 2012 Project X Physics Study, https://indico.fnal.gov/event/projectxps12 R. N. Mohapatra, J. Phys. G G [36]{}, 104006 (2009). K. S. Babu, P. S. Bhupal Dev and R. N. Mohapatra, Phys. Rev. D [79]{}, 015017 (2009). K. S. Babu and R. N. Mohapatra, arXiv:1206.5701 \[hep-ph\]. P. T. Winslow and J. N. Ng, Phys. Rev. D [81]{}, 106010 (2010) \[arXiv:1003.1424 \[hep-th\]\]. E. Friedman and A. Gal, Phys. Rev. D [78]{}, 016002 (2008) \[arXiv:0803.3696 \[hep-ph\]\]. K. Genezer, Proc. of Workshop on “B-L Violation", LBL (2007); http://inpa.lbl.gov/BLNV/blnv.htm M. Baldo-Ceolin [et al.]{}, Z. Phys. C [63]{}, 409 (1994). M. Ozer, Phys. Rev. D [26]{}, 3159 (1982). S. Rao and R. Shrock, Phys. Lett. B [116]{}, 238 (1982). W. E. Caswell, J. Milutinovic and G. Senjanovic, Phys. Lett. B [122]{}, 373 (1983). R. G. Edwards [et al.]{}, Nucl. Phys. Proc. Suppl. [140]{}, 832 (2005) \[hep-lat/0409003\]. H. -W. Lin [et al.]{} \[Hadron Spectrum Collaboration\], Phys. Rev. D [79]{}, 034502 (2009). S. R. Beane, P. F. Bedaque, K. Orginos and M. J. Savage, Phys. Rev. Lett. [97]{}, 012001 (2006). [^1]: This work performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344. This work was partially supported by LDRD 10-ERD-033 and computing provided by LLNL Institutional Computing program. [^2]: Preprint numbers: LLNL-PROC-563736	{ "pile_set_name": "ArXiv" }
--- abstract: 'We analyze the charge-noise induced coherence time $T_2$ of the fluxonium qubit as a function of the number of array junctions in the device, $N$. The pure dephasing rate decreases with $N$, but we find that the relaxation rate increases, so $T_2$ achieves an optimum as a function of $N$. This optimum can be much smaller than the number typically chosen in experiments, yielding a route to improved fluxonium coherence and simplified device fabrication at the same time.' author: - Ari Mizel and Yariv Yanay bibliography: - '../bibliography/physics.bib' title: 'Right-sizing fluxonium against charge noise' --- #### Introduction— One of the earliest superconducting qubits was the Cooper-pair box [@Schnirman97; @Bouchiat_1998; @NakamuraNat99]. This qubit design is intuitive, with the $\left\| 0 \right>$ and $\left\| 1 \right>$ states corresponding physically to an excess Cooper-pair residing on or off a small superconducting island. However, it did not take long before experiments demonstrated its acute vulnerability to ambient charge noise [@NakamuraPRL02]. One popular modification of the Cooper-pair box is the transmon qubit [@KochPRA2007], which adds a capacitive shunt to increase robustness against charge noise at the cost of reduced spectral anharmonicity. The superconducting flux qubit [@orlandoPRB99] offers an alternative that can exhibit large anharmonicity, and researchers continue to refine its design [@Yan2016]. One innovative reconsideration of the flux qubit, called “fluxonium,” was proposed [@manucharyanSCI09] to suppress charge-noise sensitivity in all of the eigenstates of the system. In this paper, we present a potent optimization of the fluxonium design that minimizes its charge-noise decoherence. Fluxonium exploits the fact that a single piece of metal naturally keeps its voltage uniform even in the presence of static external electric fields. So, instead of a standard flux qubit comprised of three or four superconducting islands, one imagines a qubit constructed from a single, annulus-shaped, island. The annulus is interrupted by a Josephson junction, and the body of the annulus shunts that junction with a large inductance, as in Fig. \[Fig:Cartoon\]. Since it is made of a single island of metal, such a qubit should remain indifferent to low-frequency charge noise. In practice, the inductance of such a loop is too small to permit a good qubit. To produce the required large inductance, fluxonium adds a long chain of islands to the loop [@manucharyanSCI09], strongly coupled via Josephson junctions, as in Fig. \[Fig:circuit\]. This design choice requires deliberation – our qubit was motivated by the robustness of a continuous piece of metal, so it seems counterproductive to incorporate a large number of islands. In the following, we confirm that, as long as the islands are coupled together sufficiently strongly, they can behave like a single piece of superconductor as far as low-frequency charge noise is concerned. However, using a standard model of charge noise [@Yan2016], we show that the qubit relaxation rate scales with the number of islands. This leads to our main result: for given fluxonium qubit parameters, there is an optimal number of islands that maximizes the qubit’s decoherence time $T_2$. We focus here on the original fluxonium proposal [@manucharyanSCI09] in which the inductor is formed by a chain of coupled superconducting islands, but our findings may be relevant for alternative realizations of the inductor [@Hazard2019; @Grunhaupt2019; @Niepce2019] provided they can be modeled [@matveevPRL02; @Maleeva2018] by such a chain. #### Hamiltonian — To calculate the charge-noise decoherence rate of fluxonium, consider the superconducting circuit depicted in Fig. \[Fig:circuit\]. The loop is pierced by time-independent flux $\Phi$, so that $\varphi \equiv 2e \Phi/\hbar$ is dimensionless. We have labelled the gauge-invariant phase drops as shown. The Lagrangian associated with this circuit is $ \mathcal{L} = \mathcal{T} + \mathcal{T}_d - \mathcal{U} $. Here, the Josephson energy is $$\mathcal{ U} = E_J^a \sum_{i=1}^N (1 - \cos \Theta_i)+ E_J^b (1-\cos(\sum_{i=1}^N \Theta_i - \varphi). \label{Eq:LU}$$ The capacitative energy is composed of two parts. The first describes the capacitors around the superconducting loop, $$\mathcal{T} = \frac{1}{2} C^a \sum_{i=1}^N\left(\frac{\hbar \dot{\Theta}_i}{2e}\right)^2 + \frac{1}{2} C^b \left(\sum_{i=1}^N \frac{\hbar \dot{\Theta}_i}{2e}\right)^2. \label{Eq:LT}$$ The second part, $\mathcal{T}_d$, describes capacitive coupling to dissipative elements. These dissipative elements, modeled as impedances [@Devoret1997; @Caldeira1983], are shown in red in Fig. \[Fig:circuit\]. Placed in series with small capacitances to ground $C_d^a$ and $C_d^b$, they produce voltage fluctuations $V_i$ that model background charge noise [@Yan2016]. The associated energy is $$\begin{split} & {\cal T}_d = \sum_{i=1}^{N-1} \frac{1}{2} C_d^a \Big(\frac{\hbar \dot{\tau}}{2e} + \sum_{j=1}^{i}\frac{\hbar \dot{\Theta}_j}{2e} - V_i\Big)^2 \\ & + \frac{1}{2} C_d^b \Big(\frac{\hbar \dot{\tau}}{2e}- V_0\Big)^2 + \frac{1}{2} C_d^b \Big(\frac{\hbar \dot{\tau}}{2e} + \sum_{j=1}^{N}\frac{\hbar \dot{\Theta}_j}{2e} - V_N\Big)^2. \label{Eq:LTd} \end{split}$$ We define (dimensionless) canonical momenta $\mathcal{N}_\tau = \partial \mathcal{L}/\partial \hbar \dot{\tau}$ and $\mathcal{N}_i = \partial \mathcal{L}/\partial \hbar \dot{\Theta}_i$. Physically, the $\mathcal{N}_i$ are integer-valued variables determined by the number of Cooper pairs residing on islands of the circuit. A standard Legendre transformation yields the Hamiltonian $$\mathcal{H} = \frac{1}{2} \left[\begin{array}{c} 2e \mathcal{N}_\tau + Q_\tau\\ 2e \mathcal{N}_1 + Q_1\\ \vdots \\ 2e \mathcal{N}_N + Q_N \end{array} \right]^T {\cal C}^{-1} \left[\begin{array}{c} 2e \mathcal{N}_\tau + Q_\tau\\ 2e \mathcal{N}_1 + Q_1\\ \vdots \\ 2e \mathcal{N}_N + Q_N \end{array} \right] + {\cal U} \label{Eq:H}$$ where $\mathcal{C}$ is the capacitance matrix obtained from \[Eq:LT,Eq:LTd\] and the offset charges have the form $$\label{Eq:Qi} Q_i = C_d^a \sum_{j=i}^{N-1} V_j + C_d^b V_N\,\,\,\,\,\, \text{ and }\,\,\,\,\,\,\, Q_\tau = C_d^b V_0 + Q_1.$$ The fact that (\[Eq:Qi\]) involves sums of voltages leads to larger offset charge fluctuations than one might naively assume. This plays a central role in making the fluxonium relaxation rate increase with $N$ as we show below. To analyze $\mathcal{H}$, first note [@FergusonPRX2013] that $\mathcal{N}_{\tau}$ is a conserved quantity since $\tau$ is absent from the Josephson energy $\mathcal{U}$. We can therefore set $\mathcal{N}_\tau$ to zero in $\mathcal{H}$, restricting our attention to eigenstates of $\mathcal{H}$ that are independent of $\tau$. Next, we specialize to the case of large $E_J^a$, which is suitable for fluxonium [@manucharyanSCI09]. The low-energy eigenstates then reside in the region $\Theta_{i}\ll 1$, and we approximate $\cos \Theta_{i} \approx 1 - \Theta_{i}^{2}/2$. This renders the Hamiltonian mostly harmonic. Following the usual procedure for harmonic Hamiltonians, we introduce a real unitary (orthogonal) transformation $U$ to define $N$ new variables $n_i = \sum_{j=1}^N U_{i j} \mathcal{N}_j$, $q_i = \sum_{j=1}^N U_{i j} Q_j$, and $\theta_i =\sum_{j=1}^N U_{i j} \Theta_j$. We set $U_{1 j} = 1/\sqrt{N}$, so that $\theta_1 = \sum_{j=1}^N \Theta_j/\sqrt{N}$ is an equal superposition mode. Then, the Hamiltonian decomposes to $\mathcal{H}_{\rm eff} = \sum_{i=1}^{N} H_{i}$, where $$\begin{split} H_1 &= \frac{\left(2e n_1 + q_1 - \sqrt{N} Q_\tau/2\right)^2 }{2 (C^a + N C^b)}\\ & + \frac{E_J^a}{2} \theta_1^2 + E_J^b (1 - \cos(\sqrt{N} \theta_1 - \varphi)), \end{split}$$ $$H_{i\ne 1} = 4E_{C}^{a} (n_i+\frac{q_i}{2e})^2 + \frac{1}{2}E_{J}^{a} \theta_i^2. \label{Eq:initialHi}$$ Here, we have neglected the effect of $C_d^a$ and $C_d^b$ on the capacitance denominators and have defined $E_{C}^{a} = e^2/2C^a$. The form of $H_{1}$ becomes more familiar if we set ${\theta = \sqrt{N} \theta_1}$, ${n = n_1/\sqrt{N}}$, ${q = q_1/\sqrt{N} - Q_\tau/2}$. Then, $$\label{Eq:h} H_{1} = 4 E_C \left(n +\frac{q}{2e}\right)^2 + E_J (1 - \cos(\theta - \varphi)) + \frac{E_L \theta^2}{2},$$ where $E_C = e^2 / 2(C^a/N + C^b)$, $E_J = E_J^b$, and ${E_L = E_J^a/N}$. These effective parameters determine the physics of the qubit. Keeping them fixed, the same Hamiltonian (\[Eq:h\]) can be realized for different $N$ provided the array junction parameters vary as $C^a = N (e^2/2E_C - C^b)$ and $E_J^a = N E_L $. Physically, $C^a$ and $E_J^a$ can be tuned this way by simply changing the area of the array junctions. The central problem we address in this paper is to optimize the charge-noise robustness of fluxonium as a function of $N$. To make the dependence on $N$ explicit, we rewrite \[Eq:initialHi\] as $$H_{i\ne 1} = \frac{4}{N} \mathcal{E}_C^a (n_i+\frac{q_i}{2e})^2 + \frac{ N E_L }{2} \theta_i^2 \label{Eq:Hi}$$ with $\mathcal{E}_C^a \equiv N E_{C}^{a} = 1/(1/E_C - 2C^b/e^2)$ independent of $N$. #### Approximate solution— The approximate Hamiltonian $\mathcal{H}_{\text{eff} }$ is conveniently separated in terms of the new variables $\theta, \theta_2,\dots,\theta_N$, so we can find the eigenstates of each term individually. We denote the Gaussian ground state of the Hamiltonian (\[Eq:Hi\]) by $e^{-i\theta_{i} q_{i}/2e} \phi_0(\theta_{})$, and the ground state and first excited state of \[Eq:h\] by $e^{-i\theta q/2e} \psi_0(\theta)$ and $e^{-i\theta q/2e}\psi_1(\theta)$, respectively. These states satisfy the usual boundary conditions, vanishing as $\theta \rightarrow \pm \infty$. At first, it appears that the exact ground state of $\mathcal{H}_{\text{eff} }$ is $$\begin{split}\label{Eq:chi0} \chi_0 & (\theta,\{\theta_i\}) = e^{-i(q \theta +\sum_{i=2}^{N} q_i \theta_i)/2e} \psi_0(\theta) \prod_{j=2}^{N} \phi_0(\theta_j ). \end{split}$$ The phase factor in front removes the background charges from $\mathcal{H}_{\text{eff}}$, so that its ground state energy is perfectly independent of low-frequency charge noise. In fact, this phase factor can be placed in front of every eigenstate of $\mathcal{H}_{\text{eff}}$, so the entire energy spectrum seems to be independent of low-frequency charge noise, the goal described in the second paragraph of this paper. Upon reflection, we realize that $\chi_0$ unfortunately does not satisfy the correct boundary conditions. Physically, each superconducting island of the circuit must house an integral number of Cooper pairs. It follows that $\Theta_i$, since it is conjugate to the discrete variable $\mathcal{N}_i$, must be a compact variable. In other words, changing the value of $\Theta_i$ by $2 \pi$ does not describe a different state of the system, so the quantum mechanical wavefunctions of the circuit must satisfy periodic boundary conditions in $\Theta_i$. To address this, we impose the correct boundary conditions using an (unnormalized) tight-binding ansatz [^1], $$\begin{split} \label{Eq:Psi0} & \Psi_0 (\theta, \{\theta_{i}\})= \\ & \sum_{\mathclap{k_1= -\infty}}^\infty\;\dotsi\; \sum_{\mathclap{k_{N}= -\infty}}^\infty \chi_0 ( \theta+2 \pi \sum_{j=1}^{N} k_j ,\{\theta_i+ 2 \pi \sum_{j=1}^{N} U_{i,j} k_j \}). \end{split}$$ The $k_1 = \dotsi = k_N=0$ term of $\Psi_0$ is our earlier ground state $\chi_0$. Since the remaining terms overlap relatively weakly with it (recall $\phi_0$ is strongly localized), $\Psi_0$ is approximately an eigenstate of $\mathcal{H}_{\text{eff}}$. We have argued that the sum of terms in $\Psi_0$ is essential in order to enforce the periodic boundary conditions, without which the spectrum would be independent of the charge-offsets $Q_i$ [@KochPRA2007]. An alternative perspective is that these terms describe coherent phase-slips in which $\theta$ jumps by a multiple of $2 \pi$. It is important to stress that these are two ways of looking at the same physical effect: fluxonium phase-slip physics [@ManucharyanPRB2012] is properly incorporated in our analysis. Note that by substituting $\psi_1$ for $\psi_0$ in \[Eq:chi0\] to define $\chi_1$, and then $\chi_{1}$ for $\chi_{0}$ in \[Eq:Psi0\], we can find the wavefunction of the first excited state, $\Psi_1$. This ansatz for $\Psi_1$ is not perfectly orthogonal with $\Psi_{0}$, but orthogonalizing it leads to negligible corrections. #### Pure dephasing— With the approximate wavefunction (\[Eq:Psi0\]), we can calculate the qubit decoherence time. We first quantify the pure dephasing of the qubit by low-frequency charge noise. Dephasing occurs when a shift in the charge parameters $Q_i$ alters the transition frequency of the qubit, ${\omega_{01} = (E_1-E_0)/\hbar}$. To find the dependence $\omega_{01}$ on the offset charges, we calculate the expectation of the original, periodic Hamiltonian $\mathcal{H}$ in state $\Psi_n$ and find that it varies with $Q_i$ as $$\begin{aligned} \nonumber & E_{n}(Q_\tau,\dots,Q_N) = \\ & E_{n}(0,\frac{e}{2},\dots,\frac{e}{2}) - \frac{\epsilon_n}{2 } \sum_{j=1}^N \cos \frac{2 \pi Q_j - \pi Q_\tau}{2e}, \label{Eq:E0}\end{aligned}$$ for $n=0,1$, where $$\begin{split} \epsilon_n= 4 N E_L e^{-\pi^2 \sqrt{E_L/8{\cal E}_{C}^{a}} (N-1)} & \int_{-\infty}^{\infty} d\theta \psi_n^{}(\theta + 2 \pi) \psi_n(\theta) \left[ \frac{\pi^2}{2} \left( 1 - \frac{1}{N}\right) + \left(\frac{\theta^2}{2N} - \int_{-\infty}^{\infty} d \bar{\theta} \left\| \psi_n(\bar{\theta}) \right\|^2 \frac{\bar{\theta}^2}{2N}\right) \right. \\ \label{Eq:epsilon0} & \left. \qquad + e^{-\sqrt{{\cal E}_{C}^{a}/2E_L} (N-1)/N^2} \left((N-2) \cos \frac{\theta + \pi}{N} - \int_{-\infty}^{\infty} d \bar{\theta} \left\| \psi_n(\bar{\theta}) \right\|^2 N \cos \frac{\bar{\theta}}{N}\right) \right] . \end{split}$$ This expression is derived within the tight-binding approximation, neglecting matrix elements between next nearest neighbor terms of \[Eq:Psi0\] and beyond. We can now find the qubit’s dephasing rate. For simplicity, suppose all of the voltages $V_i$ have the same noise power spectrum $S_V(\omega)$, and let the associated charge fluctuation be $$S_{\text{charge}}(\omega) \equiv (C_d^a)^2 S_{V}(\omega). \label{Eq:S}$$ Assume a $1/f$ form for the low-frequency power spectrum $S_{\text{charge}}(\omega) = 2 \pi A^2_{\text{charge}}/\| \omega\|$. Then, reasoning as in [@KochPRA2007], we find $$\frac{1}{T_\phi} \sim \sqrt{\frac{N}{2}\left(\frac{N-1}{2} + \left(\frac{C_d^b}{C_d^a}\right)^2 \right)} \frac{\left\| \epsilon_1 - \epsilon_0\right\|}{\hbar}\frac{A _{\text{charge}} \pi}{2 e}. \label{Eq:Tphi}$$ The pure dephasing time rapidly increases with $N$ because of the exponential factor in \[Eq:epsilon0\], as predicted in [@manucharyanSCI09]. This is because the ratio $E_J^a /E_{C}^{a} = N^{2} E_L/{\cal E}_C^a$ increases with $N$ for fixed $E_L$ and ${\cal E}_{C}^{a}$, carrying the superconducting islands further into the transmon regime [@KochPRA2007]. Alternatively, the increasing value of $E_J^a$ means stronger coupling between superconducting islands, which therefore better approximate the single piece of metal discussed in the second paragraph of this paper. #### Relaxation rate— The other source of decoherence is unwanted transitions between the two computational states. To compute the rate of this relaxation, consider the term $4 E_C n q /e$ obtained by expanding the square in \[Eq:h\]. The qubit lifetime is determined by the matrix element of this term between $\Psi_0$ and $\Psi_1$. The offset charge $q$ varies with the fluctuating voltages $V_i$ [^2], leading to $$\begin{split} \frac{1}{T_1} &= \frac{8E_C^2}{\hbar^2} \left\| \left<\psi_0\right\|n\left\|\psi_1\right>\right\|^2 \times\\ & \left( \frac{(N-2)(N-1)}{6N}+ \left(\frac{C_d^b}{C_d^a}\right)^2\right) \frac{S_{\text{charge}}(\omega_{01})}{e^2}. \label{Eq:T1} \end{split}$$ within the tight-binding approximation. In contrast to the dephasing time, we discover that $T_1$ decreases* with $N$. #### Net decoherence rate— We incorporate \[Eq:Tphi,Eq:T1\] into the net decoherence rate using the standard relation $1/T_2 = 1/T_\phi + 1/2T_1$. Since the relaxation rate increase with the number of voltages while the pure dephasing rate decreases, $T_2$ has a maximum with respect to $N$. In Figs. \[Fig:T2versusNoriginal\] and \[Fig:T2versusNdeviceC\], we show the dependence of $T_2$ upon $N$ using values of $E_C$, $E_J$, and $E_L$, and $C^b$ from fluxonium experiments [@manucharyanSCI09; @Nguyen2018]. We have chosen ${C^b= (e^2/2E_C)/(1 + E_L/E_J)}$, independent of $N$. This follows from ${C^b/C^a = E_J^b/E_J^a = E_J/N E_L}$, which is true since capacitance and Josephson energy both scale with junction area. The flux through the loop is set to $\varphi= \pi$. For the low-frequency noise spectrum in \[Eq:Tphi\], we set $A_{\rm charge} = 10^{-3} e$ [@Zorin1996; @KochPRA2007; @Krantz2019]. For the high-frequency power spectrum in \[Eq:T1\], we adopt the ohmic charge noise model [@Yan2016] $S_{\text{charge}}(\omega) = {\cal A}_{\text{charge}}^2 \omega/(2 \pi \times \text{1 GHz})$, with ${{\cal A}_{\text{charge}}^2 = (5.2 \times 10^{-9} e)^2/\text{Hz}}$. For simplicity, we set $C^b_d = C^a_d$. The rates (\[Eq:Tphi\]) and (\[Eq:T1\]) are evaluated using numerically computed fluxonium wavefunctions $\psi_i$. ![Plot of $T_2$ versus number of islands $N$. Fluxonium parameters are fixed at $E_C = 2.5$, $E_J = 9.0$, and $E_L = 0.52$ GHz, as in experiment [@manucharyanSCI09]. These imply $e^2/2C^b = 2.64$ GHz. Blue arrow indicates optimal choice $N=68$. The yellow curve shows a recalculation to check the effect of reduced wavefunction confinement, see \[Eq:broadphi\], with the arrow indicating optimal choice $N=90$.[]{data-label="Fig:T2versusNoriginal"}](T2versusNoriginal.pdf){width="2.75in"} Fig. \[Fig:T2versusNoriginal\] considers the early fluxonium experiment [@manucharyanSCI09]. The blue curve indicates that, for the experimentally chosen parameters of $E_C$, $E_J$, $E_L$, and $C^b$, the optimal value of $N$ is $68$. The original device, with $N=43$, had a ratio of only $E_J^a/E_{C}^{a}\approx 22$, leading to excessive low-frequency charge noise dephasing. At the optimal $N$, $E_J^a/E_{C}^{a} \approx 53$, so the array junctions are deeper in the transmon regime, leading to better suppression of charge-noise dephasing. ![Plot of $T_2$ versus number of islands $N$. Fluxonium parameters are fixed at $E_C = 0.55$, $E_J =2.2$, and $E_L =0.72$ GHz, as in device C of experiment [@Nguyen2018]. These imply $e^2/2C^b = 0.73$ GHz. Blue arrow indicates optimal choice $N=12$. The yellow curve shows a recalculation to check the effect of reduced wavefunction confinement, see \[Eq:broadphi\], with an arrow indicating optimal choice $N=18$. Similar results are obtained for the other devices in [@Nguyen2018].[]{data-label="Fig:T2versusNdeviceC"}](T2versusNdeviceC.pdf){width="2.75in"} Fig. \[Fig:T2versusNdeviceC\] provides an analogous plot for a recent experiment [@Nguyen2018]. The blue curve shows that the optimal choice is the relatively small value $N=12$, far less than the experimental value $N=102$. This striking reduction arises since the original device had array junctions with $E_J^a/E_{C}^{a}>3000$, far larger than needed to protect against low-frequency charge noise. The satisfactory value $E_J^a/E_{C}^{a}\approx 47$ is achieved at $N=12$; further increasing $N$ just brings about a faster relaxation rate $1/T_1$. This kind of argument gives a general rule-of-thumb for the optimal $N$. Because the pure dephasing rate in \[Eq:epsilon0,Eq:Tphi\] drops exponentially with $N$ while the relaxation rate in \[Eq:T1\] increases only polynomially, the optimal number of junctions is just large enough to suppress the former. As shown in \[Fig:optimumNversusratio\], this means that up to a logarithmic correction, $N_{\text{optimal}} \sim 5-10\times 1/\sqrt{E_{L}/{\cal E}_{C}^{a}}$. This value ensures that the array junctions are sufficiently “transmon-like” with $E_J^a/E_{C}^{a} \gtrsim 50$. The optimal fluxonium qubit incorporates the minimal number of junctions consistent with this constraint and the desired $E_C$, $E_J$, and $E_L$. ![The optimal number of junctions $N$ as a function of the ratio $E_L/{\cal E}_C^a$. The square yellow data point on the left is taken from [@manucharyanSCI09] while the remaining 8 data points are taken from [@Nguyen2018]. The region between the curves $5/\sqrt{E_L/{\cal E}_C^a}$ and $10/\sqrt{E_L/{\cal E}_C^a}$ is shaded, showing agreement with the rule-of-thumb for $N_{\text{optimal}}$.[]{data-label="Fig:optimumNversusratio"}](optimumNversusratio.pdf){width="2.75in"} Naturally, the millisecond-scale $T_2$ times in Figs. \[Fig:T2versusNoriginal\] and \[Fig:T2versusNdeviceC\] exceed the much shorter values measured experimentally. This is unsurprising, because the figures only consider charge noise, neglecting all other mechanisms of decoherence. One expects that such mechanisms, like flux noise or Purcell emission, are functions of $E_C$, $E_L$, and $E_J$ that probably do not depend sensitively on $N$. It is plausible that some mechanisms could favor smaller fluxonium designs, in which case the optimal value $N = 12$ found in Fig. \[Fig:T2versusNdeviceC\] could only decrease. #### Discussion — The optimizations here should not require experimentally unrealistic parameters. We have fixed $C^b$, $E_J^b$ near their experimental values [@manucharyanSCI09; @Nguyen2018] while the values required for $C^a$ and $E_J^a$ should be attainable by scaling the area of the array junctions. For example, to realize the optimal point $N=12$ in Fig. \[Fig:T2versusNdeviceC\], the array junctions should be modified to have an area about $12/102$ times what was chosen in the $N=102$ experiment. The resulting junctions would still be larger than the “black-sheep” junction (i.e., at the optimum $N$, $C^a > C^b$ and $E_J^a > E_J^b$). In any case, the superconducting qubit platform is characterized by remarkable experimental flexibility. There are many possibilities one could imagine to realize specified junction parameters, such as shunting each of the array junctions with its own transmon-style capacitor [@KochPRA2007]. One expects generally correct answers from the harmonic approximation (\[Eq:initialHi\]) and the tight-binding approximation (\[Eq:Psi0\]) that underlie our calculations. Symmetry considerations decidedly limit the effect of corrections to the harmonic approximation, as investigated thoroughly in [@FergusonPRX2013]. In addition, as discussed above, the optimal $N$ depends only logarithmically on most of the parameters in the equation. However, some caution is appropriate – the harmonic approximation does exaggerate the confining potential, since $1-\cos \Theta_i \le \Theta_i^2/2$. The Gaussian ground state wavefunctions $\phi_0$ are thus overly localized, suppressing the overlap between neighboring terms in \[Eq:Psi0\]. As a result, the matrix element (\[Eq:epsilon0\]) and dephasing rate (\[Eq:Tphi\]) are somewhat underestimated. To assess the amount of error that results, we consider the broadened Gaussian wavefunctions $$\phi_0(\theta_i)\to \sqrt{\lambda} \phi_0(\lambda\theta_i), \label{Eq:broadphi}$$ with $\lambda = 2/\pi$. These would be the eigenstates of \[Eq:initialHi\] if we changed $\Theta_{i}^{2}/2 \to 2\Theta_{i}^{2}/\pi^{2}\le 1- \cos\Theta_{i}$ to bound the Josephson potential from below. We use these revised $\phi_0$ to recalculate \[Eq:epsilon0\] [^3]. The result, shown as yellow curves in \[Fig:T2versusNoriginal,Fig:T2versusNdeviceC\], is an increase of the optimum $N$ by around ${\pi/2 - 1\approx 50\%}$ and a modest decrease in the associated $T_2$. Of course, the yellow curves significantly overestimate the dephasing rate, and we expect the real value to be closer to the blue curves [^4]. It is also instructive to note that the transmon case, solved exactly using a Mathieu function [@KochPRA2007] without a tight-binding approximation, exhibits $\epsilon_{0}\sim e^{-\sqrt{8 E_L /\mathcal{E}_C^a }}$, implying $\lambda \sim 8/\pi^2 \approx 0.8$. This would increase our optimal values of $N$ in Figs. \[Fig:T2versusNoriginal\] and \[Fig:T2versusNdeviceC\] by around $20\%$, intermediate between the blue and yellow curves. Taken together, these checks show that our approximations give the correct picture and only lead to modest quantitative errors. The findings presented here identify an important potential optimization of the fluxonium design. Our prediction follows from the familiar charge-noise model described in Fig. \[Fig:circuit\] and is absent from earlier studies of fluxonium decoherence [@ManucharyanPRB2012; @Viola2015] that considered different forms of environmental noise. For instance, [@ManucharyanPRB2012] instead assumed an admitttance in parallel with each of the Josephson junctions of the circuit, which naturally models dissipative current fluctuations (such as quasiparticle tunneling [@Catelani2011; @Yan2016]) across the junctions. Thus, an experimental test of our results could shed light on the charge-noise model of superconducting qubits. Most importantly, a significant gain in device performance could result from our proposed optimization. #### Acknowledgements— We are grateful to Ben Palmer for generously sharing his expertise. [^1]: To carefully verify that $\Psi_0$ satisfies the correct boundary conditions, regard the new variables $\theta, \theta_2, \dots, \theta_N$ as functions of the original variables $\Theta_i$. Add $2 \pi$ to any $\Theta_i$, and replace the index $k_i$ everywhere with $k^\prime_i -1= k_i$; $\Psi_0$ returns to itself. [^2]: Note $q = C_d^a \sum_{k=1}^{N-1} (k/N-1/2) V_k + C_d^b(V_N-V_0)/2$. This is unchanged, as to be expected, by a constant shift of all $V_i$. [^3]: In each of the exponentials of \[Eq:epsilon0\], but not the prefactor, $E_L$ gets multiplied by $\lambda^2$. The term $\frac{\pi^2}{2} \left( 1 - \frac{1}{N}\right)$ also gets multiplied by $\lambda^2$. [^4]: It is possible to treat $\lambda$ as a variational parameter in \[Eq:Psi0\] and minimize the energy of $\mathcal{H}$ to determine $\lambda$ at each $N$. The obtained $\lambda$ is generally quite close to 1, leading to curves that hug the blue curves in Figs. \[Fig:T2versusNoriginal\] and \[Fig:T2versusNdeviceC\]	{ "pile_set_name": "ArXiv" }
--- abstract: \| Let $V$ be a complex nonsingular projective 3-fold of general type with $\chi(\omega_V)\geq 0$ (resp. $>0$). We prove that the m-canonical map $\Phi_{\|mK_V\|}$ is birational onto its image for all $m\ge 14$ (resp. $\geq 8$). Known examples show that the lower bound $r_3=14$ (resp. $=8$) is optimal. address: - 'Institute of Mathematics, School of Mathematical Sciences, Fudan University, Shanghai, 200433, China' - 'Faculty of Mathematics & Computer Science, Johannes Gutenberg Universitaet Mainz, Staudingerweg 9, D-55128 Mainz, Germany' author: - Meng Chen and Kang Zuo title: 'Complex projective threefolds with non-negative canonical Euler-Poincare characteristic' --- Introduction ================ We work over the complex number field ${{\mathbb C}}$. In this paper we study multi-canonical systems $\|mK_V\|$ on complex projective threefolds $V$ of general type. On a smooth complex complete curve $C$ of genus $g(C)\ge 2$ it is well-known that the 3-canonical map $\varphi_3:=\Phi_{\|3K_C\|}$ is always an embedding. On a smooth minimal projective surface $S$ of general type Bombieri [@Bom] proved that the m-canonical map is a birational morphism for all $m\ge 5$. The work of Tsuji [@Tsuji], Hacon-M$^{\rm c}$Kernan [@H-M] and Takayama [@Tak] says that there exists a universal constant $r_3$ such that the $r_3$-canonical map $\Phi_{r_3}$ is birational for all smooth projective 3-folds of general type. We note that Tsuji [@Tsuji] announced a very large $r_3$. Recently J.A. Chen and the first author [@Jungkai-Meng] have given an explicit $r_3$ ($\leq 77$) . Under extra assumptions there have been already the following optimal results about $r_3$ for minimal projective 3-folds $X$ of general type with $\mathbb{Q}$-factorial terminal singularities: > $\bullet$ $\varphi_5:=\Phi_{\|5K_X\|}$ is birational if either $X$ is Gorenstein (by J.A. Chen, M.Chen and D-Q.Zhang [@Crelle]) or $p_g(X)\ge 4$ (by M. Chen [@IJM]) or $K_X^3\gg 0$ (by G. T. Todorov [@Tod]). $r_3=5$ is optimal. > > $\bullet$ $\varphi_8:=\Phi_{\|8K_X\|}$ is birational if $p_g(X)\ge 2$ (by M. Chen [@IJM]). $r_3=8$ is optimal. > > $\bullet$ $\varphi_7$ is birational if $q(X)>0$ (by J. A. Chen and C. D. Hacon [@C-H]). > > $\bullet$ $\varphi_5$ is birational if $\chi(\omega_X)\geq 0$ and $q(X)>0$ (by J. A. Chen and C. D. Hacon ([@C-H06]). $r_3=5$ is optimal. Going on the study of 3-folds with $\chi(\omega)\geq 0$, we prove the following: \[main\] Let $V$ be a nonsingular projective 3-fold of general type with $\chi(\omega_V):=\sum_{i=0}^3(-1)^ih^i(V,\omega_V)\ge 0$ where $\omega_V$ is the canonical line bundle of $V$. Then the m-canonical map $\varphi_m$ is birational onto its image for all $m\geq 14$. The following example of A. R. Iano-Fletcher shows that $r_3=14$ in Theorem \[main\] is optimal. \[examp1\] (See No.19 at page 151 in [@C-R]) The canonical hypersurface $X_{28}\subset \mathbb{P}(1,3,4,5,14)$ has 3 terminal quotient singularities, $p_g(X)=1$, $q(X)=h^2(\mathcal{O}_X)=0$ and $\chi(\omega_X)=-\chi(\mathcal {O}_X)=0$. A smooth model $V$ of $X_{28}$ has the invariant: $\chi(\omega_V)=0$ and $V$ is of general type. It is clear that $\varphi_m$ is birational for all $m\ge 14$ and that $\varphi_{13}$ is not birational. So the lower bound $r_3=14$ in Theorem \[main\] is sharp. Our method has a direct consequence: \[corollary\] Let $V$ be a smooth projective 3-fold of general type with $\chi(\omega_V)> 0$. Then the m-canonical map $\Phi_{\|mK_V\|}$ is birational onto its image for all $m\ge 8$. (See No.12 at page 151 in [@C-R]) Fletcher has a 3-fold $X_{16}\subset \mathbb{P}(1,1,2,3,8)$ on which the 8-canonical map is birational and the 7-canonical map is not birational. A smooth model $V$ of $X_{16}$ has the invariant $\chi(\omega)=1$ and $V$ is of general type. $V$ has $r_3(V)=8$. Thus the lower bound $r_3=8$ in Corollary \[corollary\] is again optimal. Note that 3-folds of general type with $\chi(\omega)\ge 0$ form an interesting class, around which there have been already some established works: > $\bullet$ Gorenstein minimal 3-folds of general type have $\chi(\omega)>0$; > > $\bullet$ Any 3-fold of general type admitting a generically finite cover over an Abelian variety has $\chi(\omega)>0$ (see Green-Lazarsfeld [@G-L]); > > $\bullet$ Any 3-fold of general type, admitting a representation $\rho: \pi_1(X)\rightarrow \text{GL}$ such that the Shafarevich map $\text{Sh}_\rho$ is generically finite, has $\chi(\omega)\ge 0$ (see Jost-Zuo [@J-Z]); > > $\bullet$ Any 3-fold of general type with generically large fundamental group is conjectured to have $\chi(\omega)>0$ (see Kollár [@Kollar], Conjecture 18.12.1). It is clear that the key point in studying pluricanonical maps is to compute $P_m$ for $m>1$. One may do this on a minimal model according to Reid [@YPG]. When $\chi(\omega)=-\chi({\mathcal O})\geq 0$, it is clear from Reid’s formula that $P_2>0$. This makes it possible for us to prove effective birationality of $\varphi_m$. But to prove the optimal bound $r_3=14$ need more evolved argument, which is exactly what we have done in this paper. The case $\chi(\omega)<0$ has been treated with quite a different approach very recently by J. A. Chen and the first author [@Jungkai-Meng]. Notations and set up ======================== Since both the birationality of pluricanonical maps and $\chi(\omega)$ are birationally invariant we may study a minimal model. By the 3-dimensional MMP (see [@K-M; @KMM] for instance) we only have to consider a minimal 3-fold $X$ of general type with $\mathbb{Q}$-factorial terminal singularities. Denote the Cartier index of $X$ by $r:=r(X)$ which is the minimal positive integer with $rK_X$ a Cartier divisor, where $K_X$ is a canonical divisor on $X$. The symbol $\equiv$ stands for the numerical equivalence of divisors, whereas $\sim$ denotes the linear equivalence and $=_{\mathbb{Q}}$ denotes the $\mathbb{Q}$-linear equivalence. \[setup\][Set up for $\|m_0K_X\|$.]{} Assume $P_{m_0}(X):=h^0(X, \mathcal {O}_X(m_0K_X))\ge 2$ for some positive integer $m_0>0.$ We study the $m_0$-canonical map $\varphi_{m_0}$ which is a rational map. First we fix an effective Weil divisor $K_{m_0}\sim m_0K_X$. Take successive blow-ups $\pi: X'\rightarrow X$ (along nonsingular centers), which exists by Hironaka’s big theorem, such that: \(i) $X'$ is smooth; \(ii) the movable part of $\|m_0K_{X'}\|$ is base point free; \(iii) the support of $\pi^(K_{m_0})$ is of simple normal crossings. Denote by $g_{m_0}$ the composition $\varphi_{m_0}\circ\pi$. So $g_{m_0}: X'\longrightarrow W'\subseteq{\mathbb P}^{P_{m_0}(X)-1}$ is a morphism. Let $X'\overset{f_{m_0}}\longrightarrow B\overset{s}\longrightarrow W'$ be the Stein factorization of $g_{m_0}$. We have the following commutative diagram: (50,80) (100,0)[$X$]{} (100,60)[$X'$]{} (170,0)[$W'$]{} (170,60)[$B$]{} (112,65)[(1,0)[53]{}]{} (106,55)[(0,-1)[41]{}]{} (175,55)[(0,-1)[43]{}]{} (114,58)[(1,-1)[49]{}]{} (112,2.6)(5,0)[11]{}[-]{} (162,5)[(1,0)[4]{}]{} (133,70)[$f_{m_0}$]{} (180,30)[$s$]{} (92,30)[$\pi$]{} (135,-8)[$\varphi_{m_0}$]{}(136,40)[$g_{m_0}$]{} We recall the definition of $\pi^(K_X)$. One has $r(X)K_{X'}=\pi^(r(X)K_X)+E_{\pi}$ where $E_{\pi}$ is a sum of effective exceptional divisors. One defines $\pi^(K_X):=K_{X'}-\frac{1}{r(X)}E_{\pi}$. So, whenever we take the round up of $m\pi^(K_X)$, we always have ${\ulcorner{m\pi^(K_X)}\urcorner}\leq mK_{X'}$ for any integer $m>0$. We may write $m_0\pi^(K_X)=_{\mathbb Q} M_{m_0}+E_{m_0}',$ where $E_{m_0}'$ is an effective ${{\mathbb Q}}$-divisor and $M_{m_0}$ is the movable part of $\|m_0K_{X'}\|$. On the other hand, one has $m_0K_{X'}=_{\mathbb Q}\pi^(m_0K_X)+E_{\pi, m_0}=M_{m_0}+Z_{m_0},$ where $Z_{m_0}$ is the fixed part and $E_{\pi, m_0}$ an effective ${\mathbb Q}$-divisor which is a ${\mathbb Q}$-sum of distinct exceptional divisors. Clearly $Z_{m_0}=E_{m_0}'+E_{\pi, m_0}$. If $\dim\varphi_{m_0}(X)=2$, a general fiber of $f_{m_0}$ is a smooth projective curve of genus $\ge 2$. We say that $X$ is [$m_0$-canonically fibred by curves]{}. If $\dim\varphi_{m_0}(X)=1$, a general fiber $S$ of $f_{m_0}$ is a smooth projective surface of general type. We say that $X$ is [$m_0$-canonically fibred by surfaces]{} with invariants $(c_1^2(S_0), p_g(S)),$ where $S_0$ is the minimal model of $S$. We may write $M_{m_0}\equiv a_{m_0}S$ where $a_{m_0}\ge P_{m_0}(X)-1$ by considering the degree of a curve in a projective space. [A generic irreducible element $S$ of]{} $\|M_{m_0}\|$ means either a general member of $\|M_{m_0}\|$ whenever $\dim\varphi_{m_0}(X)\ge 2$ or, otherwise, a general fiber of $f_{m_0}$. \[generic\] Assume that $\|M'\|$ is movable on $V$. By abuse of concepts, we also define [a generic irreducible element]{} $S'$ of an arbitrary linear system $\|M'\|$ on an arbitrary variety $V$ in a similar way. [A generic irreducible element]{} $S'$ of $\|M'\|$ is defined to be a generic irreducible component in a general member of $\|M'\|$. A technical theorem ======================= Believing that Theorem 2.2 in [@IJM] is quite effective in treating 3-folds $X$ with $p_g(X)\ge 2$, we extend the technique there to build a parallel theorem so as to study those $X$ with $p_g(X)\le 1$, but with $P_{m_0}(X)\ge 2$ for some integer $m_0>0$. \[a\][Assumptions]{}. We need to make the following assumptions to explain our key method. Keep the same notation as in $\ref{setup}$ above. Let $m>0$ be certain integer: - Either $m\geq m_0+2$ and $p_g(X)>0$ or $\|mK_{X'}\|$ separates different irreducible elements $S$ of $\|M_{m_0}\|$ (namely, $\Phi_{\|mK_{X'}\|}(S')\neq \Phi_{\|mK_{X'}\|}(S'')$ for two different irreducible elements $S'$, $S''$ of $\|M_{m_0}\|$) and $p_g(X)=0$. - Assume that, on the smooth surface $S$, there is a movable linear system $\|G\|$ and that $C$,as a generic irreducible element of $\|G\|$, is smooth. The linear system $\|mK_{X'}\|\|_S$ on $S$ (as a sub-linear system of $\|mK_{X'}\|_S\|$) separates different generic irreducible elements of $\|G\|$. Or sufficiently, the complete linear system $$\|K_{S} +{\ulcorner{(m-1)\pi^(K_X)-S-\frac{1}{p}E_{m_0}'}\urcorner}\|_{S}\|$$ separates different generic irreducible elements of $\|G\|$. - There is a rational number $\beta>0$ such that ${\pi}^(K_X)\|_{S}-\beta C$ is numerically equivalent to an effective ${\mathbb Q}$-divisor; [Set]{} $$\alpha:=(m-1-\frac{m_0}{p}-\frac{1}{\beta})\xi$$ and $\alpha_0:={\ulcorner{\alpha}\urcorner}$. - Either the inequality $\alpha > 1$ holds; or $C$ is non-hyperelliptic, $m-1-\frac{m_0}{p}-\frac{1}{\beta}>0$ and $C$ is an even divisor on $S$. - Either $\alpha>2$; or $\alpha_0\ge 2$ and $C$ is non-hyperelliptic; or $C$ is non-hyperelliptic, $m-1-\frac{m_0}{p}-\frac{1}{\beta}>0$ and $C$ is an even divisor on $S$. Set $\xi:=(\pi^(K_X)\cdot C)_{X'}$ which is a positive rational number and define $$p:=\begin{cases} 1 &\text{if}\ \dim (B)\ge 2\\ a_{m_0} (\text{see}\ \ \ \ref{setup}\ \text{for the definition}) &\text{otherwise.} \end{cases}$$ Let $f:=f_{m_0}: X'\longrightarrow B$ be an induced fibration by $\varphi_{m_0}$. \[Key\] Let $X$ be a minimal projective $3$-fold of general type with $P_{m_0}(X)\ge 2$ for some integer $m_0>0$. Keep the same notation as in $\ref{setup}$ above. Then the inequality $$m\xi\ge 2g(C)-2+\alpha_0$$ holds under Assumption \[a\] $(iii)$ and $(iv)$. Furthermore $\varphi_m$ of $X$ is birational onto its image under Assumption \[a\] $(i), (ii), (iii)$ and $(v)$. First we shall show that $\|mK_{X'}\|$ can separate different irreducible elements of $\|M_{m_0}\|$. When $p_g(X)=0$, $\|mK_{X'}\|$ can separate different irreducible elements of $\|M_{m_0}\|$ by assumption (i). When $p_g(X)>0$, we consider the sub-system $$\|K_{X'}+{\ulcorner{(m-1)\pi^(K_X)-\frac{1}{p}E_{m_0}'}\urcorner}\|\subset \|mK_{X'}\|.$$ Let $S'$ and $S''$ be two different generic irreducible elements of $\|M_{m_0}\|$. Clearly one has $$\begin{aligned} &&K_{X'}+{\ulcorner{(m-1)\pi^(K_X)-\frac{1}{p}E_{m_0}'}\urcorner}\\ &\ge& K_{X'}+{\ulcorner{(m-m_0-1)\pi^(K_X)}\urcorner}+S\ge S\end{aligned}$$ by assumption (i). So $\|K_{X'}+{\ulcorner{(m-1)\pi^(K_X)-\frac{1}{p}E_{m_0}'}\urcorner}\|$ can separate $S'$ and $S''$ if either $\dim (B)\ge 2$ (cf. Lemma 2 of [@T]) or $\dim (B)=1$ and $g(B)=0$ (cf. 2.1(P2) of [@MPCPS]). For the case $\dim (B)=1$ and $g(B)>0$, one has $a_{m_0}\ge P_{m_0}\ge 2$. Thus $p\ge 2$. Since $$m-1-\frac{2m_0}{p}\ge 1+(1-\frac{2}{p})m_0>0$$ and then $$(m-1)\pi^(K_X)-\frac{2}{p}E_{m_0}'-S'-S''\equiv (m-1-\frac{2m_0}{p})\pi^(K_X)$$ is nef and big, the Kawamata-Viehweg vanishing theorem ([@Kav; @V]) gives a surjective map: $$\begin{aligned} && H^0(X',K_{X'}+{\ulcorner{(m-1)\pi^(K_X)-\frac{2}{p}E_{m_0}'}\urcorner})\\ &\longrightarrow& H^0(S', K_{S'}+{\ulcorner{(m-1)\pi^(K_X)-\frac{2}{p}E_{m_0}'}\urcorner}\|_{S'})\oplus\\ &&H^0(S'', K_{S''}+{\ulcorner{(m-1)\pi^(K_X)-\frac{2}{p}E_{m_0}'}\urcorner}\|_{S''}).\end{aligned}$$ The last two groups are non-zero because $p_g(X)>0$ (so $p_g(S')$, $p_g(S'')>0$) and $m-1-m_0> 0$. Therefore $\|K_{X'}+{\ulcorner{(m-1)\pi^(K_X)-\frac{2}{p}E_{m_0}'}\urcorner}\|$ can separate $S'$ and $S''$ and so can $\|mK_{X'}\|$. By the birationality principle (P1) and (P2) of [@MPCPS], it suffices to prove that $\|mK_{X'}\|\|_S$ on $S$ gives a birational map onto its image. Practically we may study a smaller linear system than $\|mK_{X'}\|\|_S$ on $S$. Noting that $(m-1)\pi^(K_X)-\frac{1}{p}E_{m_0}'-S\equiv (m-1-\frac{m_0}{p})\pi^(K_X)$ is nef and big under the assumptions (iv) or (v), the vanishing theorem gives a surjective map $$H^0(X',K_{X'}+{\ulcorner{(m-1)\pi^(K_X)-\frac{1}{p}E_{m_0}'}\urcorner})$$ $$\longrightarrow H^0(S, K_{S}+{\ulcorner{(m-1)\pi^(K_X)-S-\frac{1}{p}E_{m_0}'}\urcorner}\|_{S}).\eqno (3.1)$$ Note that $\|K_{X'}+{\ulcorner{(m-1)\pi^(K_X)-\frac{1}{p}E_{m_0}'}\urcorner}\|\subset \|mK_{X'}\|$. It suffices to prove that $\|K_{S}+{\ulcorner{(m-1)\pi^(K_X)-S-\frac{1}{p}E_{m_0}'}\urcorner}\|_{S}\|$ gives a birational map. The birationality principle again allows us to study the restriction to curves by assumption (ii). Now consider a generic irreducible element $C\in \|G\|$. By assumption (iii), there is an effective ${\mathbb Q}$-divisor $H$ on $S$ such that $$\frac{1}{\beta}\pi^(K_X)\|_{S}\equiv C+H.$$ By the vanishing theorem, whenever $m-1-\frac{m_0}{p}-\frac{1}{\beta}>0$, we have the surjective map $$H^0(S, K_{S}+{\ulcorner{((m-1)\pi^(K_X)-S-\frac{1}{p}E_{m_0}')\|_{S}-H}\urcorner})$$ $$\longrightarrow H^0(C, K_C + D)\hskip6.8cm \eqno (3.2)$$ where $D:={\ulcorner{((m-1)\pi^(K_X)-S-\frac{1}{p}E_{m_0}')\|_{S}-C-H}\urcorner}\|_C$ is a divisor on $C$. Noting that $$((m-1)\pi^(K_X)-S-\frac{1}{p}E_{m_0}')\|_{S}-C-H\equiv (m-1- \frac{m_0}{p}-\frac{1}{\beta})\pi^(K_X)\|_{S}$$ and that $C$ is nef on $S$, we have $\deg(D)\geq\alpha$ and thus $\deg(D)\geq \alpha_0$. Whenever $C$ is non-hyperelliptic, $m-1-\frac{m_0}{p}-\frac{1}{\beta}>0$ and $C$ is an even divisor on $S$, $\deg(D)\geq 2$ automatically follows and thus $\|K_C+D\|$ gives a birational map. Whenever $\deg(D)\ge 3$, then $$\|K_{S}+{\ulcorner{((m-1)\pi^(K_X)-S-\frac{1}{p}E_{m_0}')\|_{S}-H}\urcorner}\|\|_C$$ gives a birational map. Since $$\begin{aligned} &&\|K_{S}+{\ulcorner{((m-1)\pi^(K_X)-S-\frac{1}{p}E_{m_0}')\|_{S}-H}\urcorner}\|\\ &\subset &\|K_{S}+{\ulcorner{(m-1)\pi^(K_X)-S-\frac{1}{p}E_{m_0}'}\urcorner}\|_{S}\|,\end{aligned}$$ the latter linear system gives a birational map. So $\varphi_m$ of $X$ is birational. Finally we show the inequality for $\xi$. Whenever we have $\deg(D)\ge 2$, $\|K_C+D\|$ is base point free by the curve theory. Denote by $\|M_m\|$ the movable part of $\|mK_{X'}\|$ and by $\|N_m\|$ the movable part of $\|K_{S}+{\ulcorner{((m-1)\pi^(K_X)-S-\frac{1}{p}E_{m_0}')\|_{S}-H}\urcorner}\|$. Applying Lemma 2.7 of [@MPCPS] to surjective maps (3.1) and (3.2), we have $$m\pi^(K_X)\|_{S}\ge N_m\ \ \text{and}\ \ (N_m\cdot C)_{S}\ge 2g(C)-2 +\deg (D).$$ Note that the above inequality holds without conditions (i) and (ii). We are done. \[remark\] If we replace $M_{m_0}$ in Theorem \[Key\] by any divisor $N_{m_0}\le M_{m_0}$ with $h^0(X', N_{m_0})\ge 2$, Theorem \[Key\] is still true accordingly. This is clear by the proof. The main point is that it suffices to prove that a sub-linear system of $\|mK_{X'}\|$ gives a birational map. To avoid frustrating setting up and more complicated notations, we omit the proof in details. The idea is, however, trivially similar. While applying Theorem \[Key\], one has to choose a suitable movable system $\|G\|$ on $S$. Then quite a technical problem is to find a suitable $\beta$ as in Theorem \[Key\](iii). The following lemma presents the way for the most difficult case - the rational pencil case. \[beta\] Keep the same notation as in \[setup\] and Theorem \[Key\]. Assume $B=\mathbb{P}^1$. Let $f:X'\longrightarrow \mathbb{P}^1$ be an induced fibration of $\varphi_{m_0}$. Denote by $F:=S$ a general fiber of $f$. Then one can find a sequence of rational numbers $\{\beta_n\}$ with $\lim_{n\mapsto +\infty} \beta_n = \frac{p}{m_0+p}$ such that $\pi^(K_X)\|_F-\beta_n\sigma^(K_{F_0})$ is ${{{\mathbb Q}}}$-linearly equivalent to an effective ${{{\mathbb Q}}}$-divisor $N_n$, where $\sigma:F\longrightarrow F_0$ is the blow down onto the smooth minimal model. One has $\mathcal {O}_{B}(p)\hookrightarrow {f}_\omega_{X'}^{m_0}$ and therefore ${f}_\omega_{X'/B}^{t_0p}\hookrightarrow {f}_\omega_{X'}^{t_0p+2t_0m_0}$ for any big integer $t_0$. For any positive integer $k$, denote by $M_k$ the movable part of $\|kK_{X'}\|$. Note that ${f}_\omega_{X'/B}^{t_0p}$ is generated by global sections since it is semi-positive according to E. Viehweg ([@VV]). So any local section can be extended to a global one. On the other hand, $\|t_0p\sigma^(K_{F_0})\|$ is base point free and is exactly the movable part of $\|t_0pK_F\|$ by Bombieri [@Bom] or Reider [@Reider]. Clearly one has the following relation: $$a_0\pi^(K_X)\|_F\ge M_{t_0p+2t_0m_0}\|_F\ge b_0\sigma^(K_{F_0})$$ where $a_0:=t_0p+2t_0m_0$ and $b_0:=t_0p$. This means that there is an effective $\mathbb{Q}$-divisor $E_0'$ on $F$ such that $$a_0\pi^(K_X)\|_F=_{{{\mathbb Q}}} b_0\sigma^(K_{F_0})+E_0'.$$ Thus $\pi^(K_X)\|_F =_{{{\mathbb Q}}} \frac{p}{p+2m_0}\sigma^(K_{F_0})+E_0$ with $E_0=\frac{1}{a_0}E_0'$. We consider the case $p\ge 2$. Assume that we have defined $a_n$ and $b_n$ such that the following is satisfied with $l = n:$ $$a_{l}\pi^(K_X)\|_F \ge b_{l}\sigma^(K_{F_0}).$$ We will define $a_{n+1}$ and $b_{n+1}$ inductively such that the above inequality is satisfied with $l = n+1$. One may assume from the beginning that $a_n\pi^(K_X)$ supports on a divisor with normal crossings. Then the Kawamata-Viehweg vanishing theorem implies the surjective map $$H^0(K_{X'}+{\ulcorner{a_n\pi^(K_X)}\urcorner}+F)\longrightarrow H^0(F, K_F+ {\ulcorner{a_n\pi^(K_X)}\urcorner}\|_F).$$ One has the relation $$\begin{aligned} \|K_{X'}+{\ulcorner{a_n\pi^(K_X)}\urcorner}+F\|\|_F&=&\|K_F+{\ulcorner{a_n\pi^(K_X)}\urcorner}\|_F\|\\ &\supset& \|K_F+b_n\sigma^(K_{F_0})\|\\ &\supset& \|(b_n+1)\sigma^(K_{F_0})\|.\end{aligned}$$ Denote by $M_{a_n+1}'$ the movable part of $\|(a_n+1)K_{X'}+F\|$. Applying Lemma 2.7 of [@MPCPS] again, one has $M_{a_n+1}'\|_F\ge (b_n+1)\sigma^(K_{F_0}).$ Re-modifying our original $\pi$ such that $\|M_{a_n+1}'\|$ is base point free. In particular, $M_{a_n+1}'$ is nef. Since $X$ is of general type $\|mK_X\|$ gives a birational map whenever $m$ is big enough. Thus we see that $M_{a_n+1}'$ is big if we fix a very big $t_0$ in advance. Now the Kawamata-Viehweg vanishing theorem again gives $$\begin{aligned} \|K_{X'}+M_{a_n+1}'+F\|\|_F&=&\|K_F+M_{a_n+1}'\|_F\|\\ &\supset& \|K_F+(b_n+1)\sigma^(K_{F_0})\|\\ &\supset& \|(b_n+2)\sigma^(K_{F_0})\|.\end{aligned}$$ We may repeat the above procedure inductively. Denote by $M_{a_n+t}'$ the movable part of $\|K_{X'}+M_{a_n+t-1}'+F\|$ for $t\ge 2$. For the same reason, we may assume $\|M_{a_n+t}'\|$ to be base point free. Inductively one has: $$M_{a_n+t}'\|_F\ge (b_n+t)\sigma^(K_{F_0}).$$ Applying the vanishing theorem once more, we have $$\begin{aligned} \|K_{X'}+M_{a_n+t}'+F\|\|_F&=&\|K_F+M_{a_n+t}'\|_F\|\\ &\supset& \|K_F+(b_n+t)\sigma^(K_{F_0})\|\\ &\supset& \|(b_n+t+1)\sigma^(K_{F_0})\|.\end{aligned}$$ Take $t=p-1$. Noting that $$\|K_{X'}+M_{a_n+p-1}'+F\|\subset \|(a_n+p+m_0)K_{X'}\|$$ and applying Lemma 2.7 of [@MPCPS] again, one has $$a_{n+1}\pi^(K_X)\|_F\ge M_{a_n+p+m_0}\|_F\ge M'_{a_n+p}\|_F\ge b_{n+1} \sigma^(K_{F_0}).$$ Here we set $a_{n+1}:=a_n+p+m_0$ and $b_{n+1}=b_n+p$. Set $\beta_n = \frac{b_{n}}{a_{n}}.$ Clearly $\lim_{n\mapsto +\infty} \beta_n = \frac{p}{m_0+p}$. The case $p=1$ can be proved similarly, but with a simpler induction. We omit the details. The following Lemma is needed in our proof. Though similar one has already been established in several papers of the first author, we include it here for the convenience to readers. \[b>0\] Keep the same notation as in \[setup\]. Let $f:X'\longrightarrow B$ be the induced fibration of $\varphi_{m_0}$. Denote by $F:=S$ a general fiber of $f$. If $\dim(B)=1$ and $g(B)>0$, then $\pi^(K_X)\|_F\sim\sigma^(K_{F_0})$ where $\sigma:F\longrightarrow F_0$ is the blow down onto the smooth minimal model. We shall use the idea of Lemma 14 in Kawamata’s paper [@KA]. By Shokurov’s theorem in [@Sho](see also [@HsM]), each fiber of $\pi:X'\longrightarrow X$ is rationally chain connected. Therefore, $f(\pi^{-1}(x))$ is a point for all $x\in X$. Considering the image $G\subset (X \times B)$ of $X'$ via the morphism $(\pi\times f)\circ \triangle_{X'}$ where $\triangle_{X'}$ is the diagonal map $X'\longrightarrow X'\times X'$, one knows that $G$ is a projective variety. Let $g_1:G\longrightarrow X$ and $g_2:G\longrightarrow B$ be two projections. Since $g_1$ is a projective morphism and even a bijective map, $g_1$ must be both a finite morphism of degree 1 and a birational morphism. Since $X$ is normal, $g_1$ must be an isomorphism. So $f$ factors as $f_1 \circ \pi$ where $f_1:=g_2\circ g_1^{-1} : X \rightarrow B$ is a well defined morphisms. In particular, a general fiber $F_0$ of $f_1$ must be smooth minimal. So it is clear that $\pi^(K_X)\|_F\sim\sigma^(K_{F_0})$ where $\sigma$ is nothing but $\pi\|_F$. Proof of the main theorem* ============================= We begin to prove Theorem \[main\]. Let $X$ be a complex minimal projective 3-fold of general type with $\mathbb{Q}$-factorial terminal singularities and $\chi(\omega_X)\ge 0$. \[reduction\][Reduction to the case $p_g=1$.]{} If $p_g(X)>1$, then $\varphi_8$ is birational by [@IJM]. If $q(X)>0$, then $\varphi_m$ is birational for all $m\ge 7$ by [@C-H]. Thus we assume, from now on, that $p_g(X)\le 1$ and $q(X)=0$. The assumption $$0\le \chi(\omega_X)=-\chi(\mathcal {O}_X)=-1+q(X)-h^2(\mathcal {O}_X)+p_g(X)$$ implies $p_g(X)\ge h^2(\mathcal {O}_X)+1\ge 1$. Then we clearly have $p_g(X)=1$, $h^2(\mathcal {O}_X)=0$ and $\chi(\omega_X)=0$. According to [@Crelle], we only have to study non-Gorenstein minimal 3-folds $X$. In practice we may assume the Cartier index $r(X)>1$. \[plurigenus\][Plurigenus.]{} let $X$ be a minimal 3-fold pf general type with terminal singularities. Recall Reid’s plurigenus formula (at page 413 of [@YPG]): $$P_m(X)=\frac{1}{12}m(m-1)(2m-1)K_X^3-(2m-1)\chi(\mathcal {O}_X)+\sum_{Q} R_m(Q) \eqno (4.1)$$ where $m>1$ is an integer, the correction term $$R_m(Q):=\frac{r^2-1}{12r}(m-\bar{m})+\sum_{j=0}^{\bar{m}-1} \frac{\overline{bj}(r-\overline{bj})}{2r}$$ and the sum $\sum_{Q}$ runs through all baskets of singularities $Q$ of type $\frac{1}{r}(a,-a,1)$ with the positive integer $a$ coprime to $r$, $0<a<r$, $0<b<r$, $ab\equiv 1$ ( $r$) and $\bar{m}$ the smallest residue of m $r$. [Reid’s result (Theorem 10.2 in [@YPG]) says that the above baskets $\{Q\}$ of singularities are in fact virtual (!) though $X$ may have non-quotient terminal singularities. Iano-Fletcher ([@Fletcher]) actually shows that the set of baskets $\{Q\}$ in Reid’s formula is uniquely determined by $X$]{}. \[p3\] For all basket $Q$, $R_5(Q)\ge R_4(Q)\ge R_3(Q)\ge R_2(Q)$. In particular, $$P_5(X)>P_4(X)>P_3(X)>P_2(X)\ge 1$$ for all 3-fold $X$ with $\chi(\mathcal {O}_X)=0$. Suppose that $Q$ is of type $\frac{1}{r}(a,-a,1)=\frac{1}{r}(1,-1, b)$ with $r>1$, $a$ coprime to $r$, $ab\equiv 1$ ( r) and $0<b<r$. When $r=2$, one has $$R_3(Q)=\frac{r^2-1}{6r}=\frac{1}{4}=R_2(Q).$$ When $r=3$, one has $$R_3(Q)=\frac{r^2-1}{4r}=\frac{2}{3}>\frac{1}{3}=\frac{b(r-b)}{2r}=R_2(Q).$$ When $r>3$, one has $m=\bar{m}$ for $m=2,3$ and $$R_3(Q)=\sum_{j=0}^2 \frac{\overline{bj}(r-\overline{bj})}{2r}\ge \sum_{j=0}^1 \frac{\overline{bj}(r-\overline{bj})}{2r}=R_2(Q).$$ If $\chi(\mathcal {O}_X)=0$, then Reid’s formula (4.1) gives $$P_3(X)=\frac{5}{2}K_X^3+\sum_Q R_3(Q)>\frac{1}{2}K_X^3+\sum_Q R_2(Q) =P_2(X)>0.$$ In particular, $P_3(X)\ge 2$. Similarly one can verify the inequality $P_5(X)>P_4(X)>P_3(X)$. So $P_5(X)\ge 4$. Now Theorem \[main\] essentially follows from the following theorem by letting $m_0=2, 3$. \[m\_0\] Let $X$ be a minimal projective 3-fold of general type with $\mathbb{Q}$-factorial terminal singularities, $p_g(X)=1$ and $P_{m_0}(X)\ge 2$. Then $\varphi_{m}$ is birational onto its image under one of the following conditions: \(1) $m\ge 4m_0+3$ and $2\le m_0\le 3$; \(2) $m\ge 4m_0+2$ and $m_0\ge 4$; \(3) $m\ge 14$, $\chi(\mathcal {O}_X)=0$ and $m_0=3$. Set $d_{m_0}:=\dim\varphi_{m_0}(X)$. We discuss according to the value of $d_{m_0}$. We shall mainly apply Theorem \[Key\]. Because $p_g(X)>0$, Theorem \[Key\](i) is always satisfied whenever $m\ge m_0+2$. [Case 1]{}. $d_{m_0}=3$. Recall that $S$ is a generic irreducible element of $\|M_{m_0}\|$. We have $p=1$. On $S$ we take $G:=S\|_S$. Clearly $\|G\|$ is base point free and is not composed of a pencil. So a generic irreducible element $C$ of $\|G\|$ is a smooth curve. Also under the assumption $m\ge m_0+2$ one has $$\begin{aligned} &&K_S+{\ulcorner{(m-1)\pi^(K_X)-S-E_{m_0}'}\urcorner}\|_S\\ &\ge &(K_{X'}\|_S+{\ulcorner{(m-1)\pi^(K_X)-S-E_{m_0}'}\urcorner}\|_S)+G\ge G.\end{aligned}$$ Thus Theorem \[Key\](ii) is also satisfied. Because $m_0\pi^(K_X)\|_S\ge C$, we may take $\beta=\frac{1}{m_0}$ and so Theorem \[Key\](iii) is satisfied. Note that $C^2\ge 2$ because $\|G\|$ is not composed of a pencil. So $m_0\pi^(K_X)\|_S\cdot C\ge C^2\ge 2$, which implies $\xi\ge \frac{2}{m_0}$. Now we take $m\ge 3m_0+2$ and run Theorem \[Key\]. One has $\alpha=(m-1-\frac{m_0}{p}-\frac{1}{\beta})\xi\ge 2+\frac{2}{m_0}>2$. This means that $\varphi_m$ is birational for all $m\ge 3m_0+2$. This is not the best. In fact, Theorem \[Key\] already gives $\xi\ge \frac{2g(C)+1}{3m_0+2}$. Note that $2g(C)-2=(K_S+C)\cdot C=(K_{X'}\|_S+2C)\cdot C> 4$. One has $\xi\ge \frac{9}{3m_0+2}$. Now take $m>\frac{8}{3}m_0+\frac{13}{9}$. One has $\alpha=(m-1-2m_0)\xi>2$. Theorem \[Key\] says that $\varphi_m$ is birational whenever $m>\frac{8}{3}m_0+\frac{13}{9}$. (One may go on optimizing the estimate. We stop here since we have already proved the theorem.) [Case 2]{}. $d_{m_0}=2$. We have $p=1$. On $S$ we take $G:=S\|_S$. Clearly $\|G\|$ is base point free and is composed of a pencil. So a generic irreducible element $C$ of $\|G\|$ is a smooth curve. One has $G\equiv tC$ for $t\ge P_{m_0}-2\ge 1$. Also under the assumption $m\ge m_0+2$ one has $$\begin{aligned} &&K_S+{\ulcorner{(m-1)\pi^(K_X)-S-E_{m_0}'}\urcorner}\|_S\\ &\ge &(K_{X'}\|_S+{\ulcorner{(m-1)\pi^(K_X)-S-E_{m_0}'}\urcorner}\|_S)+G\ge C.\end{aligned}$$ So $\|K_S+{\ulcorner{(m-1)\pi^(K_X)-S-E_{m_0}'}\urcorner}\|_S\|$ can separate different generic irreducible elements of $\|G\|$ provided that $\|G\|$ is composed of a rational pencil. When $\|G\|$ is composed of an irrational pencil, we need the assumption $m\ge 2m_0+2$. In fact, we have $S\|_S\equiv tC$ with $t\ge 2$ and $m_0\pi^(K_X)\|_S\equiv tC+E_{m_0}'\|_S$. Take two generic irreducible elements $C'$, $C''$ of $\|G\|$. Because $$(m-m_0-1)\pi^(K_X)\|_S-C'-C''-\frac{2}{t}E_{m_0}'\equiv (m-m_0-\frac{2}{t}m_0-1)\pi^(K_X)\|_S$$ is nef and big, the Kawamata-Viehweg vanishing theorem gives a surjective map $$\begin{aligned} & &H^0(S, K_S+{\ulcorner{(m-m_0-1)\pi^(K_X)\|_S-\frac{2}{t}E_{m_0}'\|_S}\urcorner})\\ &\longrightarrow & H^0(C', K_{C'}+D')\oplus H^0(C'', K_{C''}+D'')\end{aligned}$$ where $D'$, $D''$ are divisors of positive degree. Besides the last groups are nonzero. Noting that $$\begin{aligned} &&K_S+{\ulcorner{(m-1)\pi^(K_X)-S-E_{m_0}'}\urcorner}\|_S\\ &\ge & K_S+{\ulcorner{(m-m_0-1)\pi^(K_X)\|_S}\urcorner}\\ &\ge &K_S+{\ulcorner{(m-m_0-1)\pi^(K_X)\|_S-\frac{2}{t}E_{m_0}'\|_S}\urcorner},\end{aligned}$$ $\|K_S+{\ulcorner{(m-1)\pi^(K_X)-S-E_{m_0}'}\urcorner}\|_S\|$ separates $C'$ and $C''$. So we see that Theorem \[Key\](ii) is satisfied whenever $m\ge 2m_0+2$. Because $m_0\pi^(K_X)\|_S\ge C$, we may take $\beta=\frac{1}{m_0}$ and so Theorem \[Key\](iii) is satisfied. If we take a big $m$ such that $\alpha$ is big enough, then Theorem \[Key\] gives $$m\xi\ge 2g(C)-2+(m-1-\frac{m_0}{p}-\frac{1}{\beta}),$$ which says $\xi\ge \frac{2}{2m_0+1}$. This is only an initial estimate. Take $m> 4m_0+2$. Then $\alpha=(m-2m_0-1)\xi>2$. Theorem \[Key\] says that $\varphi_m$ is birational and that $\xi\ge \frac{5}{4m_0+3}$. Take $m=3m_0+2$. Then $\alpha=(m-2m_0-1)\xi>1$. Theorem \[Key\] gives $\xi\ge \frac{4}{3m_0+2}$. Finally take $m>\frac{7}{2}m_0+2$. Then $\alpha=(m-2m_0-1)\xi>2$. Theorem \[Key\] says that $\varphi_m$ is birational whenever $m>\frac{7}{2}m_0+2$. [Case 3]{}. $d_{m_0}=1$. We have an induced fibration $f:X'\longrightarrow B$. Denote by $F:=S$ a general fiber of $f$. Note that $F$ is a surface of general type. Denote by $\sigma:F\longrightarrow F_0$ the blowing down onto the minimal model. [Subcase 3.1]{}. $g(B)>0$. By Lemma \[b>0\], we have $\pi^(K_X)\|_F\sim \sigma^(K_{F_0})$. We need to study the condition (ii) in Theorem \[Key\]. For all $m\ge m_0+5$ one has $$\begin{aligned} && K_F+{\ulcorner{(m-1)\pi^(K_X)-F-\frac{1}{p}E_{m_0}'}\urcorner}\|_F\\ &\ge &K_F+{\ulcorner{(m-m_0-1)\pi^(K_X)\|_F}\urcorner}\\ &\ge & K_F+(m-m_0-1)\pi^(K_X)\|_F\\ &\ge &(m-m_0)\pi^(K_X)\|_F\ge 5\sigma^(K_{F_0}).\end{aligned}$$ So $\Phi_{\|K_F+{\ulcorner{(m-1)\pi^(K_X)-F-\frac{1}{p}E_{m_0}'}\urcorner}\|_F\|}$ already gives a birational map. The proof of Theorem \[Key\] tells that we may omit the conditions $(iii)\sim (v)$ in Theorem \[Key\]. Thus $\varphi_m$ is birational for all $m\ge m_0+5$. [Subcase 3.2]{}. $g(B)=0$. For a general $F$ the natural map $$H^0(X', K_{X'}-F)\longrightarrow H^0(X', K_{X'})$$ is a strict inclusion simply because $F$ is movable. This means that $p_g(F)>0$ since $p_g(X')=p_g(X)>0$. For our purpose we classify $F$ into the following three types: > \(I) $(K_{F_0}^2, p_g(F)=(1,2)$; > > \(II) $(K_{F_0}^2, p_g(F)=(2,3)$; > > \(III) $F$ does not belong to types (I) and (II). We first study the type (III) case. By the results of Bombieri [@Bom], Reider [@Reider], Catanese-Ciliberto [@C-C] and P. Francia [@Francia], $\|2\sigma^(K_{F_0})\|$ is always base point free whenever $p_g(F)>0$. We set $G:=2\sigma^(K_{F_0})$ to run Theorem \[Key\]. The inclusion $\mathcal {O}(1)\hookrightarrow f_\omega_{X'}^{m_0}$ implies the inclusion $$f_\omega_{X'/\mathbb{P}^1}^2\hookrightarrow f_\omega_{X'}^{4m_0+2}.$$ Viehweg ([@VV]) first showed that $f_\omega_{X'/\mathbb{P}^1}^2$ is semi-positive. So it is also generated by global sections. Thus it is clear that $$\|M_{4m_0+2}\|\|_F\supset \|2\sigma^(K_{F_0})\|$$ where $M_{4m_0+2}$ is the movable part of $\|(4m_0+2)K_{X'}\|$. > (\#) So Theorem \[Key\](ii) is satisfied for all $m\ge 4m_0+2$ and for all $F$ with $p_g(F)>0$, $G\le 2\sigma^(K_{F_0})$ and $\|G\|$ not an irrational pencil. In type (III) case, $G$ is an even divisor under our setting and is not composed of a pencil of curves. On the other hand the birationality of $\Phi_{\|3K_F\|}$ implies that a general $C\in \|G\|$ is non-hyperelliptic. By Lemma \[beta\] we can take a $\beta\mapsto \frac{1}{2m_0+2}$ such that $\pi^(K_X)\|_F-\beta C$ is numerically equivalent to an effective $\mathbb{Q}$-divisor. By our definition, one has $p\ge P_{m_0}-1\ge 1$. For all $m\ge 3m_0+4$, $\alpha=(m-1-\frac{m_0}{p}-\frac{1}{\beta})\xi>0$. In general, whenever $m\ge 4m_0+2$, Theorem \[Key\] asserts that $\varphi_m$ is birational. Next we study the type (II) case. We have $(K_{F_0}^2, p_g(F))=(2,3)$. By [@BPV] (page 227) we know that $\|\sigma^(K_{F_0})\|$ is base point free and $\Phi_{\|K_F\|}$ is finite of degree 2. We set $G:=\sigma^(K_{F_0})$. Then $C$, as a generic irreducible element of $G$, is smooth and of genus 3. We have already showed (statement (\#) in type (III) case) that Theorem \[Key\](ii) is satisfied for all $m\ge 4m_0+2$ because $G\le 2\sigma^(K_{F_0})$. By Lemma \[beta\] we can take a $\beta\mapsto \frac{1}{m_0+1}$ such that $\pi^(K_X)\|_F-\beta C$ is numerically equivalent to an effective $\mathbb{Q}$-divisor. So $\xi=\pi^(K_X)\|_F\cdot C\ge \beta C^2\mapsto \frac{2}{m_0+1}$. Taking the limit one has $\xi\ge \frac{2}{m_0+1}$. Similarly we have $p\ge 1$. Take $m=3m_0+2$. Then $\alpha=(m-1-\frac{m_0}{p}-\frac{1}{\beta})\xi\ge \frac{2m_0}{m_0+1}>1$. Theorem \[Key\] gives $\xi\ge \frac{6}{3m_0+2}.$ In order to get the birationality we need the assumption $m\ge 4m_0+2$, under which one has $$\alpha=(m-1-\frac{m_0}{p}-\frac{1}{\beta})\xi\ge \frac{12m_0}{3m_0+2}>2.$$ Thus $\varphi_m$ is birational whenever $m\ge 4m_0+2$. Finally we study the type (I) case. We have $(K_{F_0}^2, p_g(F))=(1,2)$. By [@BPV] we know that the movable part of $\|K_F\|$ has one simple base point. Take $\|G\|$ to be the movable part of $\|K_F\|$. Then a generic irreducible element $C$ of $\|G\|$ is a smooth curve of genus 2. Similarly we need the assumption $m\ge 4m_0+2$ to secure the condition Theorem \[Key\](ii) (by virtue of the statement (\#) in the type (III) case because $\|G\|$ is a rational pencil). By Lemma \[beta\] we can take a $\beta\mapsto \frac{1}{m_0+1}$ such that $\pi^(K_X)\|_F-\beta C$ is numerically equivalent to an effective $\mathbb{Q}$-divisor. Clearly we have $p\ge 1$. If we take a big $m$ such that $\alpha$ is big enough, Theorem \[Key\] gives $$m\xi\ge 2g(C)-2+(m-2m_0-2)\xi$$ which implies $\xi\ge \frac{1}{m_0+1}$. Take $m=4m_0+5$. Then $\alpha\ge \frac{2m_0+3}{m_0+1}>2$. Theorem \[Key\] gives $\xi\ge \frac{5}{4m_0+5}$ and so does the birationality of $\varphi_m$. Take $m=4m_0+4$. Then $\alpha\ge \frac{10m_0+10}{4m_0+5}>2$. One has the birationality of $\varphi_m$ and $\xi\ge \frac{5}{4m_0+4}$. Take $m=4m_0+3$. Then $\alpha\ge \frac{10m_0+5}{4m_0+4}>2$ whenever $m_0\ge 2$. So $\varphi_{4m_0+3}$ is birational and $\xi\ge \frac{5}{4m_0+3}$. We have already proved Theorem \[m\_0\](1). Assume $m_0\ge 4$ and take $m=4m_0+2$. Then $\alpha\ge \frac{10m_0}{4m_0+3}>2$. Theorem \[Key\] gives the birationality of $\varphi_{4m_0+2}.$ We have proved Theorem \[m\_0\](2). The only left case to verify is $m_0=3$ and $\chi(\mathcal {O}_X)=0$ for which we have already the birationality of $\varphi_{4m_0+3}.$ We have $$\xi\ge \frac{1}{3}\eqno (4.2)$$ as shown above. > (\\)Furthermore the assumption $m_0=3$, $\chi(\mathcal {O}_X)=0$ and $\xi>\frac{1}{3}$ gives $\alpha\ge 2m_0\xi>2$, which means that $\varphi_{4m_0+2}=\varphi_{14}$ is birational by Theorem \[Key\]. \[claim\] When $m_0=3$ and $\chi(\mathcal {O}_X)=0$, $\varphi_{4m_0+2}=\varphi_{14}$ is birational. To prove the claim, we need to study the 5-canonical system $\|5K_{X'}\|$. By Lemma \[p3\], one has $P_5(X)\ge 4$. Denote $d_5:=\dim\varphi_5(X)$. Keep the same notation as in \[setup\] and in Theorem \[Key\]. Recall that $M_5$ is the movable part of $\|5K_{X'}\|$. The induced fibration from $\varphi_5$ is denoted by $f_5:X'\longrightarrow W_5$. (In fact we can take further modifications over $X'$ (still denote by $X'$ the final modification) such that the final $X'$ dominates both $\varphi_3$ and $\varphi_5$.) If $d_5=3$, then $\dim \varphi_5(F)=2$ and $\dim\varphi_5(C)=1$ for a general curve $C\in \|G\|$ in a general fiber $F$. This means that the linear system $\|M_5\|\|_C\ (\subset \|M_5\|_C\|)$ gives a finite map from $C$ onto a curve and so does $\|M_5\|_C\|$. The Riemann-Roch and Clifford theorem on $C$ says that $5\pi^(K_X)\cdot C\ge M_5\cdot C\ge 2$, i.e. $\xi\ge \frac{2}{5}>\frac{1}{3}$. Statement (\\) tells that $\varphi_{14}$ is birational. If $d_5=1$, then $\|5K_{X'}\|$ induces the same fibration $f:X'\longrightarrow W_5=B$ simply because $5K_{X'}\ge 3K_{X'}$. The typical property here is that $5\pi^(K_X)\ge 3F$ for a general fiber $F$. By Lemma \[beta\] (just take $m_0'=5$ and $p'=3$) we can find a $\beta\mapsto \frac{3}{8}$ such that $\pi^(K_X)\|_F-\beta C$ is numerically equivalent to an effective $\mathbb{Q}$-divisor. Going on the argument just before the claim, we have $m_0=3$, $p=1$, $\beta=\frac{3}{8}$ and $\xi\ge \frac{1}{3}$ as in (4.2). Take $m=14$. Then $\alpha=(m-1-m_0-\frac{1}{\beta})\xi>2$. So Theorem \[Key\] gives the birationality of $\varphi_{14}$. Finally if $d_5=2$, we need to study the image surface $W_5'$ of $X'$ through the morphism $\Phi_{\|M_5\|}$. In fact we have the decomposition $$\Phi_{\|M_5\|}: X' \overset{f_5} \longrightarrow W_5 \overset{s_5}\longrightarrow W_5'\subset \mathbb{P}^{P_5(X)-1}.$$ There is a very ample divisor $H_5$ on $W_5'$ such that $M_5=\Phi_{\|M_5\|}^(H_5)$. Furthermore one has $M_5\|_{S_5}\equiv a_5\tilde{C}$ for a general member $S_5\in \|M_5\|$ and an integer $a_5\ge \deg{s_5}\deg(W_5')\ge \deg(W_5')\ge P_5(X)-2$, $\tilde{C}$ is a general fiber of $f_5$. If $a_5\ge 3$, we may utilize our argument in [Case 2]{} replacing $m_0$ by $m_0'=5$. Over there we have shown $\xi\ge \frac{4}{3m_0'+2}=\frac{4}{17}$. But we can take a better $\beta$, namely $\beta=\frac{3}{5}$. We have $p=1$. Take $m=12$. Then $\alpha=(12-1-5-\frac{5}{3})\xi>1$. Theorem \[Key\] gives $\xi\ge \frac{1}{3}$. Take $m=14$. Then $\alpha=(14-1-5-\frac{5}{3})\xi\ge \frac{19}{9}>2$. Theorem \[Key\] says that $\varphi_{14}$ is birational. So we are left to study what happens if $a_5=2$. This means that $\deg(W_5')=2$, $P_5(X)=4$ and $\deg(s_5)=1$. Recall that a degree 2 irreducible surface in $\mathbb{P}^3$ must be one of the following type of surfaces (see for instance Reid’s lecture notes [@Park], Ex. 19 at page 30): [(a)]{} $W_5'$ is the cone $\overline{\mathbb{F}}_2$ by blowing down the unique section with the self-intersection $-2$ on the Hirzebruch surface $\mathbb{F}_2$ (a ruled surface); [(b)]{} $W_5'=\mathbb{P}^1\times \mathbb{P}^1$. We study these two cases separately in the following propositions. \[a\] For case ${\bf (a)}$, $\varphi_{14}$ is birational. We know that $M_5=g_5^(H_5')$ for a very ample divisor $H_5'$ on $W_5'$ with ${H_5'}^2=2$ and $g_5:X'\longrightarrow W_5'$ is the birational morphism. Because $W_5'$ is already normal (which is a cone), by taking further modification to $X'$, we can assume that $g_5$ factors through the minimal resolution $\mathbb{F}_2$ of $W_5'$. So we have the map $g_5:X'\overset{h_5}\longrightarrow \mathbb{F}_2\overset{\nu}\longrightarrow W_5'$ where $h_5$ is a fibration and $\nu$ is the minimal resolution. Set $\bar{H_5}=\nu^(H_5')$. Then $\bar{H}^2= (H_5')^2 = 2$. Noting that $\bar{H_5}$ is nef and big on $\mathbb{F}_2$, we can write $$\bar{H}\sim \mu G_0+nT$$ where $G_0$ is the unique section with $G_0^2=-2$, $\mu$ and $n$ are integers and $T$ is the general fiber of the ruling on $\mathbb{F}_e$. The property of $\bar{H_5}$ being nef and big implies $\mu>0$ and $n\ge 2\mu\ge 2$. Now let $\theta_2:\mathbb{F}_2\longrightarrow \mathbb{P}^1$ be the ruling, whose fibers are all smooth rational curves. Set $f_0:=\theta_2\circ h_5: X'\longrightarrow \mathbb{P}^1$, which is a fibration with connected fibers. Denote by $\tilde{F}$ a general fiber of $f_0$ and by $\tilde{\sigma}:\tilde{F}\longrightarrow \tilde{F_0}$ the contraction onto the minimal model. Clearly $p_g(\tilde{F})>0$. We have $$M_5\sim g_5^(H_5')=h_5^(\bar{H})\ge 2\tilde{F}.$$ Set $N_5=2\tilde{F}$. Replace $\|M_5\|$ by the sub-linear system $\|N_5\|$ in Theorem \[Key\]. We can also study the birationality of $\varphi_m$ as remarked in \[remark\]. Note that the fibration $f_0:X'\longrightarrow \mathbb{P}^1$ is the induced fibration by $\Phi_{\|2\tilde{F}\|}$. We can repeat a similar argument to that in [Subcase 3.2]{} above. We can verify those conditions in Theorem \[Key\]. Because $14K_{X'}\ge 2\tilde{F}$, Theorem \[Key\](i) is satisfied. On the other hand, we have $\mathcal {O}(2)\hookrightarrow {f_0}_\omega_{X'}^5$ which implies the inclusion $${f_0}_\omega_{X'/\mathbb{P}^1}^2\hookrightarrow {f_0}_\omega_{X'}^{12}.$$ Similarly Viehweg’s semi-positivity and the base point freeness of $\|2\sigma^(K_{\tilde{F_0}})\|$ say that $\|12K_{X'}\|\|_{\tilde{F}}$ can separate different generic irreducible elements of $\|2\sigma^(K_{\tilde{F_0}})\|$. So can $\|14K_{X'}\|\|_{\tilde{F}}$. So Theorem \[Key\](ii) is satisfied. Now we take $m_0'=5$ to run Theorem \[Key\] with $\|M_5\|$ replaced by $\|N_5\|$. One has $p=2$. If $\tilde{F}$ is of type (III), we take $\tilde{G}:=2\tilde{\sigma}^(K_{\tilde{F_0}})$. Take $m=14$. The proof of Lemma \[beta\] says that one can take a $\beta\mapsto \frac{1}{7}$ such that $\pi^(K_X)\|_{\tilde{F}}-2\beta \tilde{\sigma}^(K_{\tilde{F_0}})$ is numerically equivalent to an effective $\mathbb{Q}$-divisor. So $\alpha=(m-1-\frac{m_0'}{p}-\frac{1}{\beta})\tilde{\xi}>0$. Noting that a generic irreducible element of $\|\tilde{G}\|$ is non-hyperelliptic and even, $\varphi_{14}$ is birational by Theorem \[Key\]. If $\tilde{F}$ is of type (II), we set $\tilde{G}:=\tilde{\sigma}^(K_{\tilde{F_0}})$. Then $\tilde{C}$, as a generic irreducible element of $\|\tilde{G}\|$, is smooth and of genus 3. We have already showed above that Theorem \[Key\](ii) is satisfied for $m=14$. By Lemma \[beta\] we can take a $\beta\mapsto \frac{2}{7}$ such that $\pi^(K_X)\|_{\tilde{F}}-\beta \tilde{C}$ is numerically equivalent to an effective $\mathbb{Q}$-divisor. So $\tilde{\xi}=\pi^(K_X)\|_{\tilde{F}}\cdot \tilde{C}\ge \beta \tilde{C}^2\mapsto \frac{4}{7}$. Taking the limit one has $\tilde{\xi}\ge \frac{4}{7}$. We have $p=2$. Take $m=14$. Then $\alpha=(m-1-\frac{m_0'}{p}-\frac{1}{\beta})\tilde{\xi}\ge 4$. Theorem \[Key\] says that $\varphi_{14}$ is birational. If $\tilde{F}$ is of type (I), we take $\tilde{G}$ to be the movable part of $\|\tilde{\sigma}^(K_{\tilde{F_0}})\|$. Then $\tilde{C}$, as a generic irreducible element of $\|\tilde{G}\|$, is smooth and of genus 2. We have already showed above that Theorem \[Key\](ii) is satisfied for $m=14$. By Lemma \[beta\] we can take $\beta=\frac{2}{7}$ such that $\pi^(K_X)\|_{\tilde{F}}-\beta \tilde{\sigma}^(K_{\tilde{F_0}})$ is numerically equivalent to an effective $\mathbb{Q}$-divisor. So $\tilde{\xi}=\pi^(K_X)\|_{\tilde{F}}\cdot C\ge \beta= \frac{2}{7}$. Take $m=12$. Then $\alpha=(m-1-\frac{m_0'}{p}-\frac{1}{\beta})\tilde{\xi}>1$. Theorem \[Key\] gives $\tilde{\xi}\ge \frac{1}{3}$. Take $m=14$. Then $\alpha=(m-1-\frac{m_0'}{p}-\frac{1}{\beta})\tilde{\xi}\ge \frac{7}{3}>2$. Theorem \[Key\] says that $\varphi_{14}$ is birational. \[b\] For case ${\bf (b)}$, $\varphi_{14}$ is birational. Recall that we have two fibrations $f:X'\longrightarrow B$ and $f_5:X'\longrightarrow W_5'=\mathbb{P}^1\times \mathbb{P}^1$. For a general canonical curve $C$ in the a general fiber $F$ of $f$, we can study $\dim\varphi_5(C)$. If $\dim\varphi_5(C)=1$, then $\Phi_{\|M_5\|}$ maps $C$ onto a curve. Clearly $5\pi^(K_X)\cdot C\ge M_5\cdot C\ge 2$. So $\xi\ge \frac{2}{5}>\frac{1}{3}$ and $\varphi_{14}$ is birational according to the statement (\\). From now on we assume that $\dim\varphi_5(C)=0$ for a general $C$. We may take further blowing ups to $X'$ so that a general $C$ is simply a generic fiber of $f_5$. Because the only very ample divisor $H_5'$ on $W_5'$ with ${H_5'}^2=2$ is $L_1+L_2=q_1^(\text{point})+q_2^(\text{point})$ where $q_1, q_2$ are projection maps from $\mathbb{P}^1\times \mathbb{P}^1$ to $\mathbb{P}^1$. Set $\tilde{f_i}:=q_i\circ f_5: X'\longrightarrow \mathbb{P}^1$, $i=1,2$. Then $\tilde{f_1}$ and $\tilde{f_2}$ are two fibrations onto $\mathbb{P}^1$. Let $F_1$ and $F_2$ are respectively general fibers of $\tilde{f_1}$ and $\tilde{f_2}$. Then $F_1\cap F_2$ is simply a general fiber $C$ of $f_5$. We will prove alternatively that $\varphi_{14}$ is birational. Consider the sub-liner system $\|K_{X'}+{\ulcorner{8\pi^(K_X)}\urcorner}+F_1+F_2\|\subset \|14K_{X'}\|$ which clearly separates different fibers of $\tilde{f_1}$. Take a general $F_1$ as a fiber of $\tilde{f_1}$. Because $8\pi^(K_X)+F_2$ is nef and big, the Kawamata-Viehweg vanishing theorem gives a surjective map $$H^0(K_{X'}+{\ulcorner{8\pi^(K_X)}\urcorner}+F_2+F_1)\longrightarrow H^0(F_1, K_{F_1}+{\ulcorner{8\pi^(K_X)}\urcorner}\|_{F_1}+C).$$ We hope to prove that $\|K_{F_1}+{\ulcorner{8\pi^(K_X)}\urcorner}\|_{F_1}+C\|$ gives a birational map for a general $F_1$. Note that $\|C\|$ is a rational pencil on the surface $\|F_1\|$. Clearly $\|K_{F_1}+{\ulcorner{8\pi^(K_X)}\urcorner}\|_{F_1}+C\|$ separate different $C$ simply because $p_g(F_1)>0$ and ${\ulcorner{8\pi^(K_X)}\urcorner}\|_{F_1}\ge 0$. Take a general curve $C$ in the family $\|F_2\|_{F_1}\|$. The vanishing theorem again gives the surjective map $$H^0(F_1, K_{F_1}+{\ulcorner{8\pi^(K_X)\|_{F_1}}\urcorner}+C)\longrightarrow H^0(C, K_C+D)$$ where $D:=({\ulcorner{8\pi^(K_X)\|_{F_1}}\urcorner})\|_C$ is a divisor of $$\deg(D)\ge (8\pi^(K_X)\|_{F_1})\cdot C=8\pi^(K_X)\cdot C=8\xi\ge \frac{8}{3}>2$$ recalling the inequality (4.2) and that $C$ is also a canonical curve in the surface $F$. So $\Phi_{\|K_C+D\|}$ is an embedding and the birationality principle says that $\varphi_{14}$ is birational. We have completely showed Theorem \[m\_0\]. \[examp2\] Example \[examp1\] shows that Theorem \[m\_0\] is optimal for $m_0=3$. It is also optimal for $m_0=2$. In fact, Fletcher has an example (page 151, No. 18 in [@C-R]): the canonical hyper-surface $X_{22}\subset \mathbb{P}(1,2,3,4,11).$ It is clear $m_0=2$ and that $\varphi_{11}$ is birational, but $\varphi_{10}$ is not birational. Recently we were informed of a new construction of a 3-fold $Y$ by E. Stagnaro ([@S]). $Y$ also has $p_g=1$, $P_2=2$ and $\varphi_{10}$ is not birational. By a similar method we are able to prove Corollary \[corollary\]. To avoid unnecessary redundancy we omit some details. \[sketch\][A sketch proof of Corollary \[corollary\]]{} First one has $P_2(X)\ge 4$ by a similar application of Reid’s formula (see the proof of Lemma \[p3\]). So one may take $m_0=2$. Set $d_2:=\dim \varphi_2(X)$. If $d_2\ge 2$, it is clear that Theorem \[Key\](i), (ii) are satisfied for $m\ge 8$. Now the situation $d_2=3$ follows directly from the argument in [Case 1]{} of Theorem \[m\_0\]. In fact, one get the birationality of $\varphi_{7}$. In the situation $d_2=2$, one may also follow the argument in [Case 2]{} of Theorem \[m\_0\]. The only difference is that we can take $p=2$ here. So it can be verified that $\varphi_8$ is birational. If $d_2=1$, one has an induced fibration $f:X'\longrightarrow B$. When $g(B)>0$, the argument in [Subcase 3.1]{} of Theorem \[m\_0\] shows that $\varphi_{m}$ is birational for all $m\ge 7$. We assume $g(B)=0$ from now on. Let $F$ be a general fiber of $f$. If $p_g(F)=0$, then $q(X)=q(F)=0$ and the assumption $\chi(\omega_X)>0$ implies $p_g(X)\ge 2$. So $\varphi_{m}$ is birational for all $m\ge 8$ by the main theorem in [@IJM]. We are reduced to study the situation $p_g(F)>0$. Now we can follow the argument of [Subcase 3.2]{} in Theorem \[m\_0\]. The typical property here is the inclusion $$\mathcal {O}(3)\hookrightarrow f_\omega_{X'}^2.$$ We can still choose $G\le 2\sigma^(K_{F_0})$ on $F$. We can verify that Theorem \[Key\] (i), (ii) are satisfied for all $m\ge 6$. Also we can choose much better $p$ ($p\ge 2$) and $\beta$, respectively, than those in [Subcase 3.2]{} in Theorem \[m\_0\]. We can verify with less difficulties, case by case (for type (I), (II), (III)) by applying Theorem \[Key\], that $\varphi_{m}$ is birational for all $m\ge 8$. In a word, Corollary \[corollary\] is true. [Final remarks.]{} Clearly, for the case $p_g=0$, Theorem \[Key\] will generate some new results which improves a corollary of J. Kollár ([@Kol]) and the first author on the birationality of $\varphi_m$, where $m$ is certain function in terms of $m_0$. We do not include our result here because we are not sure whether that would be optimal. [99]{} W. Barth, C. Peter, A. Van de Ven, [Compact complex surface]{}, Springer-Verlag, 1984. E. Bombieri, [Canonical models of surfaces of general type]{}. Inst. Hautes Études Sci. Publ. Math. [42]{} (1973), 171–219. F. Catanese, C. Ciliberto, [Surfaces with $p\sb g=q=1$]{}. Problems in the theory of surfaces and their classification (Cortona, 1988), 49-79, Sympos. Math., XXXII, Academic Press, London, 1991. J. A. Chen, M. Chen, D.-Q. Zhang, [The 5-canonical system on threefolds of general type]{}, J. Reine. Angew Math. 603(2007), 165-181. J. A. Chen, M. Chen, [Explicit birational geometry of threefolds of general type]{}, preprint. arXiv:0706.2987 J. A. Chen and C. Hacon, [Linear series of irregular varieties]{}, Algebraic Geometry in East Asia, Japan, 2002, World Scientific Press. J. A. Chen and C. Hacon, [Pluricanonical systems on irregular 3-folds of general type]{}, Math. Z. [255]{}(2007), no. 2, 343-355.. M. Chen, [Canonical stability of 3-folds of general type with $p_g\ge 3$]{}, Internat. J. Math. [14]{}(2003), 515-528. M. Chen, [Canonical stability in terms of singularity index for algebraic threefolds]{}, Math. Proc. Camb. Phil. Soc. [131]{}(2001), 241-264. A. Corti, M. Reid, [Explicit birational geometry of 3-folds]{}. London Mathematical Society, Lecture Note Series, 281. Cambridge University Press, Cambridge, 2000. P. Francia, [On the base points of the bicanonical system]{}. Problems in the theory of surfaces and their classification (Cortona, 1988), 141-150, Sympos. Math., XXXII, Academic Press, London, 1991. M. Green, R. Lazarsfeld, [Deformation theory, generic vanishing theorems and some conjectures of Enriques, Catanese and Beauville]{}, Invent. Math. [90]{}(1987), 389-407. C. D. Hacon and J. M$^{\text{\rm c}}$Kernan, [Boundedness of pluricanonical maps of varieties of general type]{}, Invent. Math. [166]{}(2006), no. 1, 1-25 C. D. Hacon, J. M$^{\rm c}$Kernan, [Shokurov’s rational connectedness conjecture]{}, Duke Math. J. [138]{}(2007), no. 1, 119-136 A. R. Iano-Fletcher, [Inverting Reid’s exact plurigenera formula]{}. Math. Ann. 284 (1989), no. 4, 617-629. J. Jost, K. Zuo, [Vanishing theorems for $L\sp 2$-cohomology on infinite coverings of compact Kahler manifolds and applications in algebraic geometry]{}. Comm. Anal. Geom. [8]{}(2000), no. 1, 1-30. Y. Kawamata, [A generalization of Kodaira-Ramanujam’s vanishing theorem]{}, Math. Ann. [261]{}(1982), 43-46. Y. Kawamata, [Characterization of abelian varieties.]{} Compositio Math. [43]{}(1981), no. 2, 253-276 Y. Kawamata, K. Matsuda, K. Matsuki, [Introduction to the minimal model problem]{}, Adv. Stud. Pure Math. [10]{}(1987), 283-360. J. Kollár, [Shafarevich maps and automorphic forms]{}, Princeton University Press, Princeton, New Jersey, 1995. J. Kollár, [Higher direct images of dualizing sheaves I]{}, Ann. Math. [123]{}(1986), 11-42; II, ibid. [124]{}(1986), 171-202. J. Kollár, S. Mori, Birational geometry of algebraic varieties, 1998, Cambridge Univ. Press. M. Reid, [Young person’s guide to canonical singularities]{}, Proc. Symposia in pure Math. [46]{}(1987), 345-414. M. Reid, [Chapters on algebraic surfaces.]{} Complex algebraic geometry (Park City, UT, 1993), 3–159, IAS/Park City Math. Ser., 3, Amer. Math. Soc., Providence, RI, 1997. I. Reider, [Vector bundles of rank 2 and linear systems on algebraic surfaces]{}, Ann. Math. [127]{}(1988), 309-316. V. V. Shokurov, [On rational connectedness]{}, Math. Notes [68]{}(2000), 652-660. E. Stagnaro, [Threefolds with Kodaira dimension 0 or 3]{}, preprint, 2006. S. Takayama, [Pluricanonical systems on algebraic varieties of general type]{}, Invent. Math. [165]{}(2006), no. 3, 551-587. S. G. Tankeev, [On n-dimensional canonically polarized varieties and varieties of fundamental type]{}, Izv. A. N. SSSR, Ser. Math. [35]{}(1971), 31-44. G. T. Todorov, [Pluricanonical maps of threefolds of general type]{}, preprint. arXiv: math.AG/0512346 H. Tsuji, [Pluricanonical systems of projective 3-folds of general type]{}, preprint. arXiv: math.AG/0204096 E. Viehweg, [Vanishing theorems]{}, J. reine angew. Math. [335]{}(1982), 1-8. E. Viehweg, [Weak positivity and the additivity of the Kodaira dimension for certain fibre spaces]{}. Proc. Algebraic Varieties and Analytic Varieties, Tokyo 1981. Adv. Studies in Math. [1]{}, Kinokunya-North-Holland Publ. 1983, 329-353	{ "pile_set_name": "ArXiv" }
--- abstract: 'This paper introduces the notion of [[cache-tapping]{}]{} into the information theoretic models of coded caching. In particular, the wiretap II channel model in the presence of multiple receivers equipped with fixed-size cache memories, and an adversary who is able to choose symbols to tap into from cache placement, in addition to or in lieu of, delivery transmission, is introduced. The model is hence termed the caching broadcast channel with a [[wire and cache tapping adversary]{}]{} of type II. The legitimate parties know neither whether cache placement, delivery, or both phases are tapped, nor the positions in which they are tapped. Only the size of the overall tapped set is known. For the instance of two receivers and two library files, the strong secrecy capacity of the model, i.e., the maximum achievable file rate while keeping the overall library strongly secure, is identified. Lower and upper bounds on the achievable strong secrecy file rate are derived when the library has more than two files. Achievability schemes in this paper rely on a code design which combines wiretap coding, security embedding codes, one-time pad keys, and coded caching. A genie-aided upper bound, in which a genie provides the transmitter with user demands before cache placement, establishes the converse for the two files instance. For the library of more than two files, the upper bound is constructed by three successive channel transformations. Our results establish that strong information theoretic security is possible against a powerful adversary who optimizes its attack over both phases of communication in a cache-aided system.' bibliography: - 'MyLib.bib' title: The Caching Broadcast Channel with a Wire and Cache Tapping Adversary of Type II --- Introduction {#Int} ============ Caching is proposed to efficiently reduce network traffic congestion by storing partial contents at the cache memories of end users earlier during off-peak times, providing local caching gain [@dowdy1982comparative; @almeroth1996use; @borst2010distributed]. More recently, reference [@maddah2014fundamental] has shown that a careful design of cache contents at the end users in a multi-receiver setting allows the transmitter to send delivery transmissions that are simultaneously useful for many users, providing a [[global caching gain]{}]{}. This gain depends on the aggregate cache memory of the network and demonstrates the ability of coding over delivery transmission and/or cache contents to offset lack of cooperation between the receivers. In numerous works to date, coded caching has been studied under various modeling assumptions and for various network configurations, including online, decentralized, hierarchical caching [@pedarsani2016online; @maddah2015decentralized; @amiri2017decentralized; @karamchandani2016hierarchical], non-uniform demands [@niesen2017coded], more users than files [@wan2016caching; @sahraei2016k], heterogeneous cache sizes [@ibrahim2017centralized; @bidokhti2017benefits], improved bounds [@lim2017information; @amiri2017fundamental; @yu2018exact], interference networks [@maddah2015cache; @hachem2016degrees; @naderializadeh2017fundamental], combination networks [@ji2015fundamental; @zewail2017combination], device-to-device communication [@ji2016fundamental; @ibrahim2018device], and broadcast channels [@bidokhti2016noisy; @ghorbel2016content; @amiri2017cache; @zhang2017fundamental]. Coded caching with confidentiality requirements has recently been studied in references [@sengupta2015fundamental; @awan2015fundamental; @zewail2016fundamental; @ravindrakumar2017private; @zewail2017combination; @kamel2018secrecy; @zewail2018wiretap]. These references assume secure cache placement, i.e., the adversary cannot tap into the cache contents or into the communication which performs the cache placement. At the other extreme, if cache placement were to be public, i.e., the adversary were to have perfect access to the cache contents, it follows in a straightforward fashion from [@shannon1949communication; @ahlswede1993common] that the presence of cache memories cannot increase the secrecy capacity. Given the results of these two extreme settings, this paper considers an intermediate scenario in which the adversary may have [[partial access]{}]{} to cache placement. The wiretap channel II, introduced in [@WTCII_Wyner], provides a model for an adversary who has partial access to the legitimate communication, in the form of a threshold on the time fraction during which the adversary is capable of tapping into the communication. In particular, the model considers a noiseless legitimate channel and an adversary who [[chooses]{}]{} a [[fixed-size]{}]{} subset of the transmitted symbols to noiselessly observe. Reference [@WTCII_Wyner] showed that, despite this ability to choose the wiretapped symbols, with proper coding, this adversary can be made no more powerful than nature, i.e., the secrecy capacity of the wiretap II model is identical to that of a binary erasure wiretapper channel with the same fraction of erasures. Reference [@nafea2015wiretap] has generalized the wiretap II model to one with a discrete memoryless, i.e., noisy, legitimate channel, and derived inner and outer bounds for its capacity-equivocation region. The secrecy capacity for this model was identified in [@goldfeld2015semantic]. Reference [@nafea2018new] has introduced a generalized wiretap model which subsumes both the classical wiretap [@WTCWyner] and wiretap II [@nafea2015wiretap] models as special cases. This generalized model was extended to multi-transmitter and multi-receiver networks in [@nafea2018generalizing; @nafea2017new1; @nafea2017new2]. In all these models, the common theme is the robustness of stochastic wiretap encoding [@WTCWyner] against a type II adversary who is allowed to choose where to tap. In this paper, we consider an adversary model of type II as in [@WTCII_Wyner; @nafea2015wiretap; @goldfeld2015semantic; @nafea2018new; @nafea2018generalizing; @nafea2017new1; @nafea2017new2], but in a cache-aided communication setting. In particular, the adversary noiselessly observes a partial subset of its choosing of the transmitted symbols over the cache placement and/or delivery phases. Thus we term this model the caching broadcast channel with a [[wire and cache tapping]{}]{} adversary of type II (CBC-WCT II). The legitimate parties do not know whether the cache placement, delivery, or both transmissions are tapped, the relative fractions of tapped symbols in each, or their positions. Only the knowledge of the overall size of the tapped set by the adversary is available to the legitimate terminals. The challenge in caching stems from the fact that the transmitter, who has access to a library of files, has no knowledge about the future demands of end users when designing their cache contents. This remains to be the case when security against an external adversary is concerned. Additionally, for the adversary model in consideration, the adversary might tap into cache placement, delivery, or both, and where the tapping occurs is [[unknown]{}]{} to the legitimate parties. We show that even under these unfavorable conditions, strong secrecy guarantees can be provided that are invariant to the positions of the tapped symbols varying between cache placement, delivery, or both. In coded caching literature up to date, the physical communication which populates the cache memories at end users does not need to be considered in the problem formulation, due to the assumption of secure cache placement. By contrast, in order to model cache placement that is tapped by an adversary, we consider a length-$n$ communication block over a two-user broadcast wiretap II channel [@nafea2017new1]. The sizes of cache memories at the receivers are fixed in this model. We note that introducing variable memory sizes for which a rate-memory tradeoff can be characterized, as in the usual setup for caching, requires considering additional communication blocks for cache placement. Being of future interest, we comment on this extension to multiple communication blocks for cache placement in the Discussion section, Section \[Discussion\]. We as well provide reasoning for our choice of the broadcast setting for cache placement. In summary, the contributions of this work are summarized as follows: 1. We introduce the notion of [[cache-tapping]{}]{} into the information theoretic models of coded caching, in which an adversary of type II is able to tap into a fixed-size subset of symbols of its choosing either from cache placement, delivery, or both transmissions. 2. We characterize the strong secrecy capacity of the model, i.e., the maximum achievable file rate which keeps the overall library strongly secure, for the instance of a transmitter’s library with two files: - We devise an achievability scheme which integrates wiretap coding [@WTCII_Wyner], security embedding codes [@ly2012security; @liang2014broadcast], one-time pad keys [@shannon1949communication], [[coded]{}]{} cache placement and [[uncoded]{}]{} delivery [@maddah2014fundamental]. - We utilize a genie-aided upper bound, in which a genie provides the transmitter with user demands before cache placement, rendering the model to a two-user broadcast wiretap II channel [@nafea2017new1], in order to establish the converse for this case. 3. We derive lower and upper bounds on the strong secrecy file rate when the transmitter’s library has more than two files: - We utilize a coding scheme that is similar to the scheme we used for two files. However, the cache placement and delivery schemes we employ to achieve the rates differ from those utilized for two files. In particular, we utilize here [[uncoded]{}]{} cache placement and a [[partially coded]{}]{} delivery. - We derive the upper bound in three steps. First, we consider a transformed channel with an adversary who can tap an equal fraction of symbols to our model, but is only allowed to tap into the delivery phase. Since this adversary has a more restricted strategy space than the original one, the corresponding secrecy capacity is at least as large as our original model. Next, we utilize Sanov’s theorem in method of types [@cover2006elements Theorem 11.4.1] to further upper bound the secrecy capacity for the restricted adversary model by the secrecy capacity when the adversary encounters a discrete memoryless binary erasure channel. Finally, the secrecy capacity of the discrete memoryless model is upper bounded by the secrecy capacity of a single receiver setting in which the receiver requests two files from the library. The remainder of the paper is organized as follows. Section \[ChannelModel\] describes the communication system proposed in this paper. Section \[MainResult\] presents the main results. The proofs for these results are provided in Sections \[Proof\_Thm1\], \[Proof\_Thm2\], and \[Proof\_Thm3\]. Section \[Discussion\] provides a discussion about the communication model in question and the presented results. Section \[Con\] concludes the paper. ![The caching broadcast channel with a wire and cache tapping adversary of type II (CBC-WCT II). The adversary chooses tapping sets $S_1$ and $S_2$ in placement and delivery.[]{data-label="fig:sysmodel_1"}](Fig1.eps) System Model ============ [\[ChannelModel\]]{} We first remark the notation we utilize throughout the paper. $\mathbb{N}$, $\mathbb{Z}$, and $\mathbb{R}$ denote the sets of natural, integer, and real numbers, respectively. For $a,b\in\mathbb{R}$, $[a:b]$ denotes the set of integers $\{i\in\mathbb{N}: a\leq i\leq b\}$. We use $A_{[1:n]}$ to denote the sequence of variables $\{A_1,A_2,\cdots,A_n\}$. For two sets ${\mathcal{A}}_1$ and ${\mathcal{A}}_2$, ${\mathcal{A}}_1\times {\mathcal{A}}_2$ denotes their Cartesian product. ${\mathcal{A}}^T$ denotes the $T$-fold Cartesian product of the set ${\mathcal{A}}$. For $W_1,W_2\in[1:M]$, $W_1\oplus W_2$ denotes the bit-wise XOR on the binary bit strings that correspond to $W_1$, $W_2$. We use $\mathbbm{1}_{{\mathcal{A}}}$ to denote the indicator function for the event ${\mathcal{A}}$. $\mathbb{D}(p_x\|\|q_x)$ denotes the Kullback-Leibler divergence between the probability distributions $p_x$, $q_x$, defined on the same probability space. $\{\epsilon_n\}_{n\geq 1}$ denotes a sequence of positive real numbers such that $\epsilon_n\rightarrow 0$ as $n\rightarrow\infty$. Consider the communication system depicted in Fig. \[fig:sysmodel\_1\], in which the adversary has the ability to tap into both the cache placement and delivery transmissions. The transmitter observes $D\geq 2$ independent messages (files), $W_1,W_2,\cdots,W_D$, each of which is uniformly distributed over $[1:2^{nR_s}]$. Each receiver has a cache memory of size $\frac{n}{2}$ bits. The communication occurs over two phases: placement and delivery. The broadcast channel is noiseless during both phases. The communication model is described as follows: [[Cache placement phase:]{}]{} During this phase, the transmitter broadcasts a length-$n$ binary signal, ${\bold{X}}_c^n\in\{0,1\}^n$, to both receivers. The codeword ${\bold{X}}_c^n$ is a function of the library files, i.e., ${\bold{X}}_c^n\triangleq f_c\left(W_{[1:D]}\right)$. The transmitter does not know the receiver demands during cache placement [@maddah2014fundamental]. Each receiver has a cache memory of size $\frac{n}{2}$ bits in which they store a function of ${\bold{X}}_c^n$, $M_{c,j}\triangleq f_{c,j}({\bold{X}}_c^n)$, where $f_{c,j}: \{0,1\}^n\mapsto [1:2^{\frac{n}{2}}]$ and $j=1,2$. [[Delivery phase:]{}]{} At the beginning of the delivery phase, the two receivers announce their demands ${\bold{d}}\triangleq (d_1,d_2)\in [1:D]^2$ to the transmitter. The transmitter, in order to satisfy the receiver demands, encodes $W_{[1:D]}$ and ${\bold{d}}$ into the binary codeword ${\bold{X}}_{{\bold{d}}}^n\in\{0,1\}^n$. In particular, for each ${\bold{d}}$, the transmitter uses the encoder $f_{{\bold{d}}}: [1:2^{nR_s}]^D \mapsto\{0,1\}^n$ and sends ${\bold{X}}_{{\bold{d}}}^n\triangleq f_{{\bold{d}}}\left(W_{[1:D]}\right)$. [[Decoding:]{}]{} Receiver $j$ utilizes the decoder $g_{{\bold{d}},j}:[1:2^{\frac{n}{2}}]\times \{0,1\}^n \mapsto [1:2^{nR_s}]$, in order to output the estimate $\hat{W}_{d_j}\triangleq g_{{\bold{d}},j}\left(f_{c,j}({\bold{X}}_c^n),{\bold{X}}_{{\bold{d}}}^n\right)$ of its desired message $W_{d_j}$, where $j=1,2$. [[Adversary model:]{}]{} The adversary chooses two subsets $S_1,S_2\subseteq [1:n]$. The size of the sum of cardinalities of $S_1$ and $S_2$ is fixed. That is, for $\|S_1\|=\mu_1$, $\|S_2\|=\mu_2$, $\mu_1,\mu_2\leq n$, we have $\mu_1+\mu_2= \mu$. The subsets $S_1$ and $S_2$ indicate the positions tapped by the adversary during the cache placement and delivery phases, respectively. Over the two phases, the adversary observes the length-$2n$ sequence ${\bold{Z}}_{S}^{2n}= [{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n]\in{\mathcal{Z}}^{2n}$; ${\bold{Z}}_{S_j}^n\triangleq [Z_{S_j,1},Z_{S_j,2},\cdots,Z_{S_j,n}]\in{\mathcal{Z}}^{n}$, $j=1,2$, $$\begin{aligned} \label{eq:Z_S_1_2} &Z_{S_1,i}=\begin{cases} X_{c,i},\; i\in S_1\\ ?,\; i\notin S_1 \end{cases},\text{ and } \qquad Z_{S_2,i}=\begin{cases} X_{{\bold{d}},i},\; i\in S_2\\ ?,\; i\notin S_2. \end{cases}\end{aligned}$$ The alphabet is ${\mathcal{Z}}=\{0,1,?\}$, where $``?"$ denotes an erasure. The legitimate terminals know neither the realizations of $S_1$ and $S_2$, nor the values of $\mu_1$ and $\mu_2$. Only $\mu$ is known. Let us define $\alpha_1=\frac{\mu_1}{n}$ and $\alpha_2=\frac{\mu_2}{n}$ as the fractions of the tapped symbols in the cache placement and delivery phases, and let $\alpha=\alpha_1+\alpha_2$ be the overall tapped ratio. Notice that $\alpha_1,\alpha_2\in [0,1]$ and $\alpha\in (0,2]$. We consider that $\alpha$ is strictly greater than zero, i.e., the adversary is present. For $\alpha=0$, i.e., no adversary, the problem considered in this paper has been extensively studied in the literature, see for example [@maddah2014fundamental; @tian2016symmetry; @ibrahim2017optimization; @cao2018coded]. A channel code ${\mathcal{C}}_{2n}$ for this model consists of - $D$ message sets; ${\mathcal{W}}_{l}\triangleq[1:2^{nR_s}]$, $l=1,2,\cdots,D$, - Cache encoder; $f_c:[1:2^{nR_s}]^D\mapsto \{0,1\}^n$, - Cache decoders; $f_{c,j}:\{0,1\}^n\mapsto [1:2^{\frac{n}{2}}]$, $j=1,2$, - Delivery encoders; $\left\{f_{{\bold{d}}}:{\bold{d}}\in [1:D]^2\right\}$, $f_{{\bold{d}}}:[1:2^{nR_s}]^D\mapsto \{0,1\}^n$, - Decoders; $\left\{g_{{\bold{d}},j}: j=1,2,\; {\bold{d}}\in [1:D]^2\right\}$, $g_{{\bold{d}},j}:[1:2^{\frac{n}{2}}]\times\{0,1\}^n\mapsto [1:2^{nR_s}]$. The file rate $R_s$ is said to be achievable with strong secrecy if there exists a sequence of channel codes $\{{\mathcal{C}}_{2n}\}_{n\geq 1}$ satisfying $$\begin{aligned} \label{eq:reiaility} &{{\underset{n\rightarrow\infty}\lim}}\max_{{\bold{d}}\in [1:D]^2}\mathbb{P}\left(\bigcup_{j=1,2}(\hat{W}_{d_j}\neq W_{d_j})\right)=0,\\ \label{eq:secrecy} \text{and }\quad&{{\underset{n\rightarrow\infty}\lim}}\;\underset{\begin{subarray}{c} S_1,S_2\subseteq [1:n]:\\\|S_1\|+\|S_2\|\leq \mu\end{subarray}}\max\; I\left(W_{[1:D]};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)=0.\end{aligned}$$ That is, $R_s$ is the symmetric secure file rate, under any demand vector and adversarial strategy. The strong secrecy capacity $C_s$ is the the supremum of all achievable $R_s$. While we consider the file rate $R_s$ which guarantees reliability for the worst-case demand vector, the average rate for which there exists a prior distribution on the demands has been studied in coded caching literature as well; see for example [@niesen2017coded; @lim2017information; @ji2017order]. The condition in (\[eq:secrecy\]) guarantees strong secrecy against all possible strategies for the adversary, i.e., choices of the subsets $S_1$ and $S_2$ which satisfy the condition $\|S_1\|+\|S_2\|\leq \mu$. Main Results {#MainResult} ============ For clarity of exposition, we first study the model described in Section \[ChannelModel\] when the transmitter has two files in its library, i.e., $D=2$. We then extend the ideas and the analysis to the case of a library with more than two files, i.e., $D>2$. For $D>2$, we utilize a channel coding scheme that is similar to the scheme we construct for $D=2$, but the cache placement and delivery schemes that achieve the best rates are different from those used for $D=2$. The following theorem presents the strong secrecy capacity for $D=2$. \[thm:Thm1\] For $0<\alpha\leq 2$ and $D=2$, the strong secrecy capacity for the caching broadcast channel with a wire and cache tapping adversary of type II (CBC-WCT II), described in Section \[ChannelModel\], is given by $$\begin{aligned} \label{eq:sum_secrecy_capacity_D=2} C_s(\alpha)=1-\frac{\alpha}{2}.\end{aligned}$$ The proof is provided in Section \[Proof\_Thm1\]. Theorem \[thm:Thm2\] below presents an achievable strong secrecy file rate for $D>2$. \[thm:Thm2\] For $0<\alpha\leq 2$ and $D>2$, the achievable strong secrecy file rate for the CBC-WCT II is $$\begin{aligned} \label{eq:sum_secrecy_rate_D>2} R_s(\alpha)\geq \begin{cases} \frac{1}{2}+\frac{3(1-\alpha)}{4D},\qquad 0<\alpha<1 \\ 1-\frac{\alpha}{2},\qquad 1\leq \alpha\leq 2.\end{cases}\end{aligned}$$ The proof is provided in Section \[Proof\_Thm2\]. The following theorem upper bounds the secure file rate when $D>2$. \[thm:Thm3\] For $0<\alpha\leq 2$ and $D>2$, the achievable strong secrecy file rate for the CBC-WCT II is upper bounded as $$\begin{aligned} \label{eq:sum_secrecy_bound_D>2} R_s(\alpha)\leq \begin{cases} \frac{1}{2}+\frac{2D-1}{2D(D-1)}(1-\alpha),\qquad 0< \alpha<1\\ 1-\frac{\alpha}{2},\qquad 1\leq \alpha\leq 2.\end{cases}\end{aligned}$$ The proof is provided in Section \[Proof\_Thm3\]. ![Bounds on the achievable strong secrecy file rate $R_s$, when $\alpha=0.4$ and $D\geq 3$.[]{data-label="fig:lower_upper_bounds"}](Fig2.eps) The following corollary is immediate from Theorems \[thm:Thm1\], \[thm:Thm2\], and \[thm:Thm3\]. \[Cor:Cor1\] For $1\leq\alpha\leq 2$, that is when the adversary can tap longer than one phase of communication, the strong secrecy capacity for the CBC-WCT II is $$\begin{aligned} \label{eq:sum_secrecy_capacity_alpha>1} C_s(\alpha)=1-\frac{\alpha}{2}.\end{aligned}$$ When $\alpha\in [1,2]$, i.e., $n\leq \mu\leq 2n$, two possible strategies for the adversary are $\{S_1=[1:n], S_2\subset [1:n]\}$ or $\{S_1\subset [1:n], S_2=[1:n]\}$. That is, the adversary can tap into all symbols in one phase and a subset of symbols in the other phase. Interestingly, a positive strong secrecy capacity is achievable against such an adversary. We elaborate more on the intuition behind this result in the Discussion section. Unlike for $1\leq \alpha\leq 2$, for $0<\alpha<1$, the lower and upper bounds in (\[eq:sum\_secrecy\_rate\_D>2\]) and (\[eq:sum\_secrecy\_bound\_D>2\]) have a gap. For illustration purposes, these bounds are plotted for $\alpha=0.4$ in Fig. \[fig:lower\_upper\_bounds\]. When $\alpha=0$, i.e., no adversary, our achievability scheme for $D>2$ described in Section \[Proof\_Thm2\] reduces to the achievability scheme in [@maddah2014fundamental], which is shown to achieve the optimal rate-memory tradeoff for the case of two users and a library size of three or larger [@tian2016symmetry; @cao2018coded]. The upper bound for $D>2$ derived in this work is to address the intricacies of the adversarial model and is useful only when the adversary is present ($\alpha>0$), i.e., (\[eq:sum\_secrecy\_bound\_D>2\]) is loose when $\alpha=0$. Proof of Theorem \[thm:Thm1\] {#Proof_Thm1} ============================= In this section, we prove Theorem \[thm:Thm1\], which identifies the strong secrecy capacity for the model in Section \[ChannelModel\] when $D=2$. Recall that the demand vector is denoted by ${\bold{d}}=(d_1,d_2)$, where $d_1,d_2\in\{1,2\}$. Converse {#Converse} -------- For the model in Theorem \[thm:Thm1\], when the demand vector ${\bold{d}}$ is known to the transmitter during cache placement, the model reduces to a broadcast wiretap channel II, over a length-$2n$ communication block. The strong sum secrecy rate for that model, $2R_s$, is upper bounded by $$\begin{aligned} \label{eq:known_demand_tapped_caching} 2R_s\leq 2-\alpha,\end{aligned}$$ which follows from our recent work [@nafea2017new1 Theorem 1]. Notice that (\[eq:known\_demand\_tapped\_caching\]) holds for any ${\bold{d}}=(d_1,d_2)$ such that $d_1\neq d_2$, which represents the worst-case demands. Since the demand vector is unknown for the model in consideration, $1-\frac{\alpha}{2}$ is an upper bound for its strong secrecy capacity. Restricted Adversary Models as Building Blocks {#Restricted_Models} ---------------------------------------------- Before proceeding with the achievability proof, it is relevant to take a step back and investigate the secrecy capacity when a known fraction of cache placement, a known fraction of delivery, or both, is tapped. In particular, we consider that the adversary taps into (i) cache placement only, (ii) delivery only, or (iii) both and the relative fractions of tapped symbols in each are known. For these three models, we show that the strong secrecy file rate in (\[eq:sum\_secrecy\_capacity\_D=2\]), i.e., $1-\frac{\alpha}{2}$, is achievable, and hence determines their strong secrecy capacities. We then use these models as building blocks for when the relative fractions are [[unknown]{}]{}, and provide the achievability proof in Sections \[Achievability\_1\] and \[Achievability\_2\]. ### $\\\;$ Let $\alpha_1=\alpha$ ($\alpha_2=0$), and $\|S_1\|=\mu$ ($S_2=\varnothing$). That is, the adversary taps into cache placement only. Consider that the transmitter and the receivers know that $\alpha_1=\alpha$. We show that $1-\frac{\alpha}{2}$ is an achievable strong secrecy file rate for this setting. The transmitter divides the message $W_l,\;l=1,2,$ into three independent messages, $W_l^{(1)}$, $W_l^{(2)}$, and $W_{l,s}$, where $W_l^{(1)}$, $W_l^{(2)}$, are uniform over $\left[1:2^{n\frac{1-\alpha-\epsilon_n}{2}}\right]$, and $W_{l,s}$ is uniform over $\left[1:2^{n\frac{\alpha+\epsilon_n}{2}}\right]$. Define $$\begin{aligned} \label{eq:example1_1} &M_c=\left\{M_{c,1},M_{c,2}\right\};\quad M_{c,1}=W_1^{(1)}\oplus W_2^{(1)},\quad M_{c,2}=W_1^{(2)}\oplus W_2^{(2)},\\ \label{eq:example1_2} &M_{{\bold{d}}}=\left\{W_{d_1}^{(2)},W_{d_2}^{(1)},W_{d_1,s},W_{d_2,s}\right\}.\end{aligned}$$ During cache placement, the transmitter maps $M_c$ into ${\bold{X}}_c^n$ using stochastic encoding, i.e., wiretap coding [@WTCWyner]. Since the rate of $M_c$ is less than $1-\alpha$, $M_c$ is strongly secure from the adversary who observes $n\alpha$ symbols of ${\bold{X}}_c^n$ [@nafea2018new; @goldfeld2015semantic]. During the delivery phase, the transmitter sends ${\bold{X}}_{{\bold{d}}}^{n}$ as the binary representation of $M_{{\bold{d}}}$ which is of length $n$ bits, since the delivery phase is noiseless and secure. Using ${\bold{X}}_c^n$, noiselessly received during cache placement, receiver $j$, $j=1,2,$ recovers $M_{c,j}$ and stores it in its cache memory. Notice that the size of $M_{c,j}$, for $j=1,2,$ is smaller than $\frac{n}{2}$ bits, i.e., the cache size at each receiver. Using ${\bold{X}}_{{\bold{d}}}^n$, received noiselessly during delivery, both receivers perfectly recover $M_{{\bold{d}}}$. Using $M_{{\bold{d}}}$ along with its cache contents, i.e., $M_{c,j}$, and for $n$ sufficiently large[^1], receiver $j$ correctly recovers its desired message $W_{d_j}$, $j=1,2$. For secrecy, we show in Appendix \[AppendixA\] that (\[eq:secrecy\]), which reduces to $$\begin{aligned} \label{eq:example1_3} {{\underset{n\rightarrow\infty}\lim}}\;\max_{S_1\subseteq [1:n]:\;\|S_1\|=\mu}\; I(W_1,W_2;{\bold{Z}}_{S_1}^n)=0,\end{aligned}$$ is satisfied. Since $\epsilon_n\rightarrow 0$ as $n\rightarrow\infty$, the achievable strong secrecy file rate is given by $$\begin{aligned} \label{eq:example1_4} R_s(\alpha)=2\times \frac{1-\alpha}{2}+\frac{\alpha}{2}=1-\frac{\alpha}{2}.\end{aligned}$$ ### $\\\;$ This setting corresponds to $\alpha_1=0$ and $\alpha_2=\alpha$, and the transmitter and receivers possess this knowledge. Once again, we show that $1-\frac{\alpha}{2}$ is an achievable strong secrecy file rate. The transmitter performs the same division of $W_l,\;l=1,2,$ as in Setting $1$. In addition, the transmitter generates the keys $K_1$, $K_2$, each is uniform over $\left[1:2^{n\frac{\alpha+\epsilon_n}{2}}\right]$, independent from one another and from $W_1$, $W_2$. In this case, we define $M_c$, $M_{{\bold{d}}}$, and $\tilde{M}_{{\bold{d}}}$, as follows. $$\begin{aligned} \label{eq:example2_2} &M_c=\left\{M_{c,1},M_{c,2}\right\};\;\;M_{c,1}=\left\{W_1^{(1)}\oplus W_2^{(1)},K_1\right\},\;\; M_{c,2}=\left\{W_1^{(2)}\oplus W_2^{(2)},K_2\right\},\\ \label{eq:example2_3} &M_{{\bold{d}}}=\left\{W_{d_1}^{(2)},W_{d_2}^{(1)}\right\},\\ \label{eq:example2_4} &\tilde{M}_{{\bold{d}}}=\left\{W_{d_1,s}\oplus K_1,W_{d_2,s}\oplus K_2\right\}.\end{aligned}$$ During cache placement, the transmitter sends ${\bold{X}}_c^n$ as the binary representation of $M_c$, and receiver $j$, $j=1,2,$ stores $M_{c,j}$ in its cache memory. During delivery, the transmitter encodes $M_{{\bold{d}}}$ into ${\bold{X}}_{{\bold{d}}}^n$ using wiretap coding, while using $\tilde{M}_{{\bold{d}}}$ as the randomization message. Receiver $j$ recovers $M_{{\bold{d}}}$ and $\tilde{M}_{{\bold{d}}}$, using which, along with $M_{c,j}$, it correctly decodes $W_{d_j}$, for sufficiently large $n$. By contrast, the adversary can only retrieve $\tilde{M}_{{\bold{d}}}$ using which it can gain no information about $W_1$ and $W_2$. We show in Appendix \[AppendixB\] that (\[eq:secrecy\]), i.e., $$\begin{aligned} \label{eq:example2_5} {{\underset{n\rightarrow\infty}\lim}}\;\max_{S_2\subseteq [1:n]:\;\|S_2\|=\mu}\; I(W_1,W_2;{\bold{Z}}_{S_2}^n)=0,\end{aligned}$$ is satisfied. The achievable strong secrecy file rate is again $1-\frac{\alpha}{2}$. ### $\\\;$ For this setting, neither $\alpha_1=0$ nor $\alpha_2=0$. However, the transmitter and receivers know the values of $\alpha_1$ and $\alpha_2$. Under these assumptions, the scheme which achieves the strong secrecy rate of $1-\frac{\alpha}{2}$ depends on whether $\alpha_1\geq \alpha_2$. For $\alpha_1\geq \alpha_2$, we utilize an achievability scheme similar to Setting $1$; for $\alpha_1<\alpha_2$, we utilize an achievability scheme similar to Setting $2$. [[Case $1$: $\alpha_1\geq \alpha_2$:]{}]{} The transmitter divides $W_l,\;l=1,2,$ into the independent messages $\left\{W_l^{(1)},W_l^{(2)},W_{l,s}\right\}$; $W_l^{(1)},W_l^{(2)}$ are uniform over $\left[1:2^{n\frac{1-\alpha_1-\epsilon_n}{2}}\right]$ and $W_{l,s}$ is uniform over $\left[1:2^{n\frac{{\alpha_1-\alpha_2}}{2}}\right]$. The transmitter forms $M_c$ and $M_{{\bold{d}}}$ as in (\[eq:example1\_1\]) and (\[eq:example1\_2\]), and uses wiretap coding to map them into ${\bold{X}}_c^n$ and ${\bold{X}}_{{\bold{d}}}^n$, respectively. As in setting $1$, receiver $j$ correctly decodes $W_{d_j}$. For the secrecy constraint, notice that $M_c$ and $M_{{\bold{d}}}$ are independent, and their rates are $1-\alpha_1-\epsilon_n$ and $1-\alpha_2-\epsilon_n$, respectively. Thus, wiretap coding strongly secures both $M_c$ and $M_{{\bold{d}}}$ against the adversary. We show in Appendix \[AppendixC\] that (\[eq:secrecy\]) is satisfied. The achievable strong secrecy file rate is $$\begin{aligned} \label{eq:example3_2} R_s(\alpha)=2\times \frac{1-\alpha_1}{2}+\frac{\alpha_1-\alpha_2}{2}=\frac{2-\alpha_1-\alpha_2}{2}=1-\frac{\alpha}{2}.\end{aligned}$$ [[Case $2$: $\alpha_1<\alpha_2$:]{}]{} The transmitter (i) divides $W_l$, $l=1,2$, into $\left\{W_l^{(1)},W_l^{(2)},W_{l,s}\right\}$; $W_l^{(1)},W_l^{(2)}$ are uniform over $\left[1:2^{n\frac{1-\alpha_2-\epsilon_n}{2}}\right]$ and $W_{l,s}$ is uniform over $\left[1:2^{n\frac{{\alpha_2-\alpha_1}}{2}}\right]$, (ii) generates the keys $K_1,K_2$, uniform over $\left[1:2^{n\frac{\alpha_2-\alpha_1}{2}}\right]$, independently from $W_1,W_2$, (iii) forms $M_c$ as in (\[eq:example2\_2\]) and encodes it into ${\bold{X}}_c^n$ using wiretap coding, (iv) forms $M_{{\bold{d}}}$ as in (\[eq:example2\_3\]) and forms $\tilde{M}_{{\bold{d}}}$ as $$\begin{aligned} &\tilde{M}_{{\bold{d}}}=\left\{W_{d_1,s}\oplus K_1,W_{d_2,s}\oplus K_2,\tilde{W}\right\},\end{aligned}$$ where $\tilde{W}$ is independent from all other variables and uniform over $\left[1:2^{n(\alpha_1+\epsilon_n)}\right]$, and (v) encodes $M_{{\bold{d}}}$ into ${\bold{X}}_{{\bold{d}}}^n$ using wiretap coding, while using $\tilde{M}_{{\bold{d}}}$ as the randomization message. As in Setting $2$, for $n$ sufficiently large, receiver $j$, $j=1,2,$ correctly decodes $W_{d_j}$, and the adversary can only retrieve $\tilde{M}_{{\bold{d}}}$ using which it can gain no information about $W_1$, $W_2$. In Appendix \[AppendixD\], we show that (\[eq:secrecy\]) is satisfied. The achievable secrecy rate is $$\begin{aligned} R_s(\alpha)=2\times \frac{1-\alpha_2}{2}+\frac{\alpha_2-\alpha_1}{2}=1-\frac{\alpha}{2}.\end{aligned}$$ With the aforementioned settings, we showed that the same secrecy rate, i.e., $1-\frac{\alpha}{2}$, is achievable irrespective of where the adversary taps as long as $\alpha_1$ and $\alpha_2$ are known. The question then arises whether the lack of knowledge about relative fractions of tapped symbols would decrease the secrecy capacity. The following setting we propose provides a hint on the answer. ### $\\\;$ The adversary taps into either cache placement or delivery, but not both. The legitimate terminals [[do not]{}]{} know which phase is tapped. We show that the strong secrecy rate $1-\frac{\alpha}{2}$ is again achievable. The transmitter performs the same division of $W_1$, $W_2$ as in Settings $1$, $2$, and generates independent keys $K_1$, $K_2$ as in Setting $2$. Let us define $$\begin{aligned} \label{eq:example4_1} &M_c=\left\{M_{c,1},M_{c,2}\right\};\quad M_{c,1}=W_1^{(1)}\oplus W_2^{(1)},\quad M_{c,2}=W_1^{(2)}\oplus W_2^{(2)},\\ \label{eq:example4_2} &\tilde{M}_c=\left\{K_1,K_2\right\},\\ \label{eq:example4_3} &M_{{\bold{d}}}=\left\{W_{d_1}^{(2)},W_{d_2}^{(1)}\right\},\\ \label{eq:example4_4} &\tilde{M}_{{\bold{d}}}=\left\{W_{d_1,s}\oplus K_1,W_{d_2,s}\oplus K_2\right\}.\end{aligned}$$ During cache placement, the transmitter encodes $M_c$ into ${\bold{X}}_c^n$ using wiretap coding, while using $\tilde{M}_c$ as the randomization message. Receiver $j$, $j=1,2,$ stores $M_{c,j}$, $\tilde{M}_{c,j}$, in its cache memory. During delivery, the transmitter uses wiretap coding to encode $M_{{\bold{d}}}$ into ${\bold{X}}_{{\bold{d}}}^n$, while using $\tilde{M}_{{\bold{d}}}$ as the randomization message. Using its cache contents, along with $M_{{\bold{d}}}$ and $\tilde{M}_{{\bold{d}}}$, receiver $j$ correctly decodes $W_{d_j}$. By contrast, the adversary can only retrieve either $\left\{K_1,K_2\right\}$ or $\left\{W_{d_1,s}\oplus K_1,W_{d_2,s}\oplus K_2\right\}$, but not both, using which it can obtain no information about $W_1$ and $W_2$. We show in Appendix \[AppendixE\] that (\[eq:secrecy\]) is satisfied. The achievable strong secrecy rate is $1-\frac{\alpha}{2}$. The lack of knowledge about which phase is tapped is countered by encrypting pieces of information, $\left\{W_{d_1,s},W_{d_2,s}\right\}$, with one-time pad keys $K_1$ and $K_2$, while ensuring that the adversary only retrieves either the keys or the encrypted bits but not both; using which it can gain no information about the messages $W_1$ and $W_2$. In the following subsection, we generalize this idea to tackle the case when the adversary gets to tap into both phases, with no knowledge about the relative fractions of tapped symbols in each, i.e., the model in Fig. \[fig:sysmodel\_1\]. In particular, similar to [@liang2014broadcast], in which the uncertainty about the wiretapper’s channel is treated by using a security embedding code [@ly2012security], here, in each phase, we construct an embedding code in which $n\alpha$ single-bit layers are embedded into one another. Doing so, we ensure that, no matter what the values for $\alpha_1$ and $\alpha_2$ are, the adversary can retrieve no more than $n\alpha_1$ bits from cache placement, and $n\alpha_2$ bits from delivery. By designing what the adversary retrieves to be either a set of key bits and/or information bits encrypted with a distinct set of key bits, we guarantee no information on the messages is asymptotically leaked to the adversary. We thus prove that the lack of knowledge about relative fractions of tapped symbols [[does not decrease]{}]{} the secrecy capacity. Achievability for $\alpha\in(0,1)$: {#Achievability_1} ----------------------------------- We are now ready to present the achievability for the general model considered in this paper. Consider first $\alpha\in(0,1)$. For simplicity, assume that $n\frac{\alpha_1}{2}=\frac{\mu_1}{2}$ and $n\frac{\alpha_2}{2}=\frac{\mu_2}{2}$ are integers. A minor modification to the analysis can be adopted otherwise. The transmitter divides $W_l$, $l=1,2,$ into the independent messages $W_l^{(1)}$, $W_l^{(2)}$, $W_{l,s}$; $W_l^{(1)}$, $W_l^{(2)}$ are uniform over $\left[1:2^{n\frac{1-\alpha}{2}}\right]$, and $W_{l,s}$ is uniform over $\left[1:2^{n\frac{\alpha}{2}}\right]$. The transmitter generates the independent keys $K_1,K_2$, uniform over $\left[1:2^{n\frac{\alpha}{2}}\right]$, and independent from $W_1,W_2$. For simplicity of exposition, we have ignored the small rate reduction $\epsilon_n$ at this stage, as we will introduce this later into the security analysis. The main ideas of the achievability proof are: 1. The transmitter uses wiretap coding with a randomization message of size $n(\alpha_1+\alpha_2)=n\alpha$ bits in [[both the cache placement and delivery]{}]{} phases. As the adversary does not tap into more than $n\alpha$ bits in each phase, a secure transmission rate of $1-\alpha$ is achievable in each phase, as long as the randomization messages in the two phases are independent. Using coded placement for $W_1^{(1)}$, $W_1^{(2)}$, $W_2^{(1)}$, $W_2^{(2)}$, a secure file rate of $1-\alpha$ can be achieved. 2. The randomization messages over the two phases can deliver additional secure information, of rate $\frac{\alpha}{2}$ per file, via encryption. The overall achievable file rate is thus $R_s=1-\frac{\alpha}{2}$. In particular, we utilize the keys $K_1$, $K_2$, as the randomization message for cache placement. Along with wiretap coding, we employ a security embedding code [@ly2012security], by using bits of $K_1$, $K_2$, in a manner that allows the adversary to be able to retrieve only the last $n\frac{\alpha_1}{2}$ bits from each. In the delivery phase, we encrypt additional pieces of information, $W_{d_1,s}$ and $W_{d_2,s}$, with the keys $K_1$ and $K_2$, and utilize this encrypted information as the randomization message. We employ again a security embedding code, in the [[reverse order]{}]{}, such that the adversary can only retrieve the first $n\frac{\alpha_2}{2}$ bits from each of $W_{d_1,s}\oplus K_1$ and $W_{d_2,s}\oplus K_2$. 3. With the aforementioned construction, the adversary, for any values of $\alpha_1$ and $\alpha_2$ it chooses, can only retrieve a set of key bits and/or a set of information bits encrypted with other key bits. In particular, due to the [[reversed embedding order]{}]{}, the adversary does not obtain, in the delivery phase, any message bits encrypted with key bits it has seen during cache placement. In addition, since $\{K_1,K_2\}$ is independent from $\left\{W_{d_1,s}\oplus K_1,W_{d_2,s}\oplus K_2\right\}$, and is an independent sequence, the adversary can not use the revealed key bits in the cache placement to obtain any information about the bits of $W_{d_1,s}\oplus K_1$ and $W_{d_2,s}\oplus K_2$ that need to be securely transmitted in the delivery phase. ![Codebook construction for the cache placement phase, ${\mathcal{C}}_{c,n}$.[]{data-label="fig:code_construction_1"}](Fig3.eps) We now explain the achievability scheme in more detail. Let us define $M_c$ and $\tilde{M}_c$ as in (\[eq:example4\_1\]) and (\[eq:example4\_2\]). In particular, let $$\begin{aligned} \label{eq:M_c} &M_c=\left\{M_{c,1},M_{c,2}\right\};\quad M_{c,1}=W_1^{(1)}\oplus W_2^{(1)},\quad M_{c,2}=W_1^{(2)}\oplus W_2^{(2)},\\ \label{eq:M_c_tilde} &\tilde{M}_c=\left\{\tilde{M}_{c,1},\tilde{M}_{c,2}\right\};\qquad \tilde{M}_{c,1}=K_1,\qquad \tilde{M}_{c,2}=K_2.\end{aligned}$$ $M_c$ in (\[eq:M\_c\]) represents the message to be securely transmitted during cache placement, regardless of the adversary’s choice of $\alpha_1$. $\tilde{M}_c$ in (\[eq:M\_c\_tilde\]) represents the randomization message utilized for wiretap coding in the cache placement. The transmitter further divides $\tilde{M}_{c,1}$ and $\tilde{M}_{c,2}$ into sequences of independent binary bits, $\left\{\tilde{M}_{c,1}^{(1)},\tilde{M}_{c,1}^{(2)},\cdots,\tilde{M}_{c,1}^{\left(n\frac{\alpha}{2}\right)}\right\}$ and $\left\{\tilde{M}_{c,2}^{(1)},\tilde{M}_{c,2}^{(2)},\cdots,\tilde{M}_{c,2}^{\left(n\frac{\alpha}{2}\right)}\right\}$, and generates ${\bold{X}}_c^n$ as follows: [[Cache Placement Codebook Generation:]{}]{} Let $m_c$, $\tilde{m}_{c,1}=\left\{\tilde{m}_{c,1}^{(1)},\tilde{m}_{c,1}^{(2)},\cdots,\tilde{m}_{c,1}^{\left(n\frac{\alpha}{2}\right)}\right\}$, and $\tilde{m}_{c,2}=\left\{\tilde{m}_{c,2}^{(1)},\tilde{m}_{c,2}^{(2)},\cdots,\tilde{m}_{c,2}^{\left(n\frac{\alpha}{2}\right)}\right\}$ be the realizations of $M_c$, $\tilde{M}_{c,1}$, and $\tilde{M}_{c,2}$ in (\[eq:M\_c\]) and (\[eq:M\_c\_tilde\]). We construct the cache placement codebook ${\mathcal{C}}_{c,n}$, from which ${\bold{X}}_c^n$ is drawn, as follows. We randomly and independently distribute all the possible $2^{n}$ length-$n$ binary sequences into $2^{n(1-\alpha)}$ bins, indexed by $m_c \in \left[1:2^{n\frac{1-\alpha}{2}}\right]^2$. Each bin $m_c$ contains $2^{n\alpha}$ binary sequences (codewords). Further, we randomly and independently divide each bin $m_c$ into two sub-bins, indexed by $\tilde{m}_{c,1}^{(1)}$, and each contains $2^{n\alpha-1}$ codewords. The two sub-bins $\tilde{m}_{c,1}^{(1)}$ are further divided into smaller bins, indexed by $\tilde{m}_{c,2}^{(1)}$, and each contains $2^{n\alpha-2}$ codewords. The process continues, going over $\tilde{m}_{c,1}^{(2)}$, $\tilde{m}_{c,2}^{(2)}$, $\cdots$, $\tilde{m}_{c,1}^{\left(n\frac{\alpha}{2}-1\right)}$, $\tilde{m}_{c,2}^{\left(n\frac{\alpha}{2}-1\right)}$, $\tilde{m}_{c,1}^{\left(n\frac{\alpha}{2}\right)}$, until the remaining two codewords, after each sequence of divisions, are indexed by $\tilde{m}_{c,2}^{\left(n\frac{\alpha}{2}\right)}$. The codebook ${\mathcal{C}}_{c,n}$ is described in Fig. \[fig:code\_construction\_1\]. An alternative representation of the binning procedure described above is that, each of the $2^{n\alpha}$ binary codewords in the bin $m_c$, where $m_c \in \left[1:2^{n\frac{1-\alpha}{2}}\right]^2$, is randomly assigned to an index $\left\{\tilde{m}_{c,1}^{(1)},\tilde{m}_{c,2}^{(1)},\tilde{m}_{c,1}^{(2)},\tilde{m}_{c,2}^{(2)},\cdots,\tilde{m}_{c,1}^{\left(n\frac{\alpha}{2}\right)},\tilde{m}_{c,2}^{\left(n\frac{\alpha}{2}\right)}\right\}$. We chose however to present the former description in order to provide a more detailed explanation of the embedding structure; in particular, the order of embedding, which is a critical component in the achievability scheme. [[Cache Encoder:]{}]{} Given the messages $w_1$, $w_2$, i.e., $\left\{w_1^{(1)},w_1^{(2)},w_{1,s}\right\}$, $\left\{w_2^{(1)},w_2^{(2)},w_{2,s}\right\}$, the transmitter generates $m_c$, $\tilde{m}_c=\left\{\tilde{m}_{c,1},\tilde{m}_{c,2}\right\}$ as in (\[eq:M\_c\]), (\[eq:M\_c\_tilde\]). Using the codebook ${\mathcal{C}}_{c,n}$, the transmitter sends ${\bold{x}}_c^n$ which corresponds to $m_c$, $\tilde{m}_{c,1}$, $\tilde{m}_{c,2}$, i.e., ${\bold{x}}_c^n\Big(m_c,\tilde{m}_{c,1}^{(1)},\tilde{m}_{c,2}^{(1)},\cdots,\tilde{m}_{c,1}^{\left(n\frac{\alpha}{2}\right)},\tilde{m}_{c,2}^{\left(n\frac{\alpha}{2}\right)}\Big)$. For the delivery phase, as in (\[eq:example4\_3\]) and (\[eq:example4\_4\]), define $$\begin{aligned} \label{eq:M_d} &M_{{\bold{d}}}=\left\{W_{d_1}^{(2)},W_{d_2}^{(1)}\right\},\\ \label{eq:M_d_tilde} &\tilde{M}_{{\bold{d}}}=\left\{\tilde{M}_{{\bold{d}},1},\tilde{M}_{{\bold{d}},2}\right\};\qquad \tilde{M}_{{\bold{d}},1}=W_{d_1,s}\oplus K_1,\qquad \tilde{M}_{{\bold{d}},2}=W_{d_2,s}\oplus K_2.\end{aligned}$$ $M_{{\bold{d}}}$ in (\[eq:M\_d\]) represents the message to be securely transmitted during the delivery phase no matter what the adversary’s choice of $\alpha_2$ is. $\tilde{M}_{{\bold{d}}}$ in (\[eq:M\_d\_tilde\]) represents the randomization message utilized for the wiretap coding in the delivery phase. Similar to cache placement, the transmitter further divides $\tilde{M}_{{\bold{d}},1}$, $\tilde{M}_{{\bold{d}},2}$ into sequences of independent binary bits, $\big\{\tilde{M}_{{\bold{d}},1}^{(1)},\cdots,\tilde{M}_{{\bold{d}},1}^{\left(n\frac{\alpha}{2}\right)}\big\}$, $\big\{\tilde{M}_{{\bold{d}},2}^{(1)},\cdots,\tilde{M}_{{\bold{d}},2}^{\left(n\frac{\alpha}{2}\right)}\big\}$, and generates ${\bold{X}}_{{\bold{d}}}^n$ as follows. [[Delivery Codebook Generation:]{}]{} Let $m_{{\bold{d}}}$, $\tilde{m}_{{\bold{d}},1}=\big\{\tilde{m}_{{\bold{d}},1}^{(1)},\cdots,\tilde{m}_{{\bold{d}},1}^{\left(n\frac{\alpha}{2}\right)}\big\}$, $\tilde{m}_{{\bold{d}},2}=\big\{\tilde{m}_{{\bold{d}},2}^{(1)},\cdots,\tilde{m}_{{\bold{d}},2}^{\left(n\frac{\alpha}{2}\right)}\big\}$ be the realizations of $M_{{\bold{d}}}$, $\tilde{M}_{{\bold{d}},1}$, $\tilde{M}_{{\bold{d}},2}$ in (\[eq:M\_d\]), (\[eq:M\_d\_tilde\]). We construct the delivery codebook ${\mathcal{C}}_{{\bold{d}},n}$, from which ${\bold{X}}_{{\bold{d}}}^n$ is drawn, in a similar fashion as the codebook ${\mathcal{C}}_{c,n}$, but with a reversed indexing of the sub-bins. In particular, we randomly and independently divide all the $2^{n}$ binary sequences into $2^{n(1-\alpha)}$ bins, indexed by $m_{{\bold{d}}} \in \left[1:2^{n\frac{1-\alpha}{2}}\right]^2$, and each contains $2^{n\alpha}$ codewords. We further randomly and independently divide each bin $m_{{\bold{d}}}$ into two sub-bins, indexed by $\tilde{m}_{{\bold{d}},1}^{\left(n\frac{\alpha}{2}\right)}$, and each contains $2^{n\alpha-1}$ codewords. The process continues, going in reverse order over $\tilde{m}_{{\bold{d}},2}^{\left(n\frac{\alpha}{2}\right)},\tilde{m}_{{\bold{d}},1}^{\left(n\frac{\alpha}{2}-1\right)}$, $\tilde{m}_{{\bold{d}},2}^{\left(n\frac{\alpha}{2}-1\right)}$, $\cdots$, $\tilde{m}_{{\bold{d}},1}^{(1)}$, until the remaining two codewords, after each sequence of divisions, are indexed by $\tilde{m}_{{\bold{d}},2}^{(1)}$. The codebook ${\mathcal{C}}_{{\bold{d}},n}$ is described in Fig. \[fig:code\_construction\_2\]. ![Codebook construction for the delivery phase, ${\mathcal{C}}_{{\bold{d}},n}$.[]{data-label="fig:code_construction_2"}](Fig4.eps) [[Delivery Encoder:]{}]{} Given $w_1$, $w_2$, i.e., $\big\{w_1^{(1)},w_1^{(2)},w_{1,s}\big\}$, $\big\{w_2^{(1)},w_2^{(2)},w_{2,s}\big\}$, and ${\bold{d}}=(d_1,d_2)$, the transmitter generates $m_{{\bold{d}}}$, $\tilde{m}_{{\bold{d}}}=\{\tilde{m}_{{\bold{d}},1},\tilde{m}_{{\bold{d}},2}\}$ as in (\[eq:M\_d\]), (\[eq:M\_d\_tilde\]). The transmitter sends ${\bold{x}}_{{\bold{d}}}^n$, from ${\mathcal{C}}_{{\bold{d}},n}$, which corresponds to $m_{{\bold{d}}}$, $\tilde{m}_{{\bold{d}},1}$, and $\tilde{m}_{{\bold{d}},2}$, i.e., ${\bold{x}}_{{\bold{d}}}^n\Big(m_{{\bold{d}}},\tilde{m}_{{\bold{d}},1}^{\left(n\frac{\alpha}{2}\right)},\tilde{m}_{{\bold{d}},2}^{\left(n\frac{\alpha}{2}\right)},\cdots,\tilde{m}_{{\bold{d}},1}^{(1)},\tilde{m}_{{\bold{d}},2}^{(1)}\Big)$. [[Decoding:]{}]{} Using ${\bold{X}}_c^n$, receiver $j$, $j=1,2,$ recovers $M_{c,j}$, $\tilde{M}_{c,j}$, and stores them in its cache memory. For $j=1,2,$ the combined size of $M_{c,j}$ and $\tilde{M}_{c,j}$ does not exceed $\frac{n}{2}$ bits. Using ${\bold{X}}_{{\bold{d}}}^n$, both receivers recover $M_{{\bold{d}}}$, $\tilde{M}_{{\bold{d}}}$. Using $M_{{\bold{d}}}$, $\tilde{M}_{{\bold{d}}}$, $M_{c,j}$, $\tilde{M}_{c,j}$, and for $n$ sufficiently large, receiver $j$ correctly decodes $W_{d_j}$. [[Security Analysis:]{}]{} Let us first slightly modify the construction above as follows. Recall that $\{\epsilon_n\}_{n\geq 1}$ is a sequence of positive real numbers such that $\epsilon_n\rightarrow 0$ as $n\rightarrow\infty$. Define $$\begin{aligned} \label{eq:alpha_epsilon} \alpha_{\epsilon}=\alpha+2\epsilon_n,\qquad \alpha_{1,\epsilon}=\alpha_1+\epsilon_n,\qquad \alpha_{2,\epsilon}=\alpha_2+\epsilon_n.\end{aligned}$$ That is, $\alpha_{1,\epsilon}+\alpha_{2,\epsilon}=\alpha_{\epsilon}$. We increase the sizes of $K_1$ and $K_2$ into $\frac{n\alpha_{\epsilon}}{2}$ bits, from $n\frac{\alpha}{2}$, and zero-pad the bit strings of $W_{d_1,s}$ and $W_{d_2,s}$ accordingly. Additionally, we decrease the sizes of $W_l^{(1)}$, $W_l^{(2)}$, $l=1,2,$ to $n\frac{1-\alpha_{\epsilon}}{2}$ bits, instead of $n\frac{1-\alpha}{2}$. Once again, we assume that $\frac{n\alpha_{\epsilon}}{2}$ and $\frac{n\alpha_{1,\epsilon}}{2}$ are integers; as minor modifications can be adopted otherwise. Let us fix the subsets $S_1,S_2\subseteq [1:n]$. For the corresponding (fixed) values of $\alpha_1$ and $\alpha_2$, the cache placement codebook ${\mathcal{C}}_{c,n}$ can be viewed as a wiretap code with $2^{n(1-\alpha_{1,\epsilon})}$ bins. Each bin is indexed by the message $$\begin{aligned} \label{eq:w_c} w_c=\left(m_c,\tilde{m}_{c,1}^{(1)},\tilde{m}_{c,2}^{(1)},\tilde{m}_{c,1}^{(2)},\tilde{m}_{c,2}^{(2)},\cdots,\tilde{m}_{c,1}^{\left(n\frac{\alpha_{2,\epsilon}}{2}\right)},\tilde{m}_{c,2}^{\left(n\frac{\alpha_{2,\epsilon}}{2}\right)}\right).\end{aligned}$$ Each bin $w_c$ contains $2^{n\alpha_{1,\epsilon}}$ binary codewords which are indexed by the randomization message $$\begin{aligned} \label{eq:w_c_tilde} \tilde{w}_c=\left(\tilde{m}_{c,1}^{\left(n\frac{\alpha_{2,\epsilon}}{2}+1\right)},\tilde{m}_{c,2}^{\left(n\frac{\alpha_{2,\epsilon}}{2}+1\right)},\tilde{m}_{c,1}^{\left(n\frac{\alpha_{2,\epsilon}}{2}+2\right)},\tilde{m}_{c,2}^{\left(n\frac{\alpha_{2,\epsilon}}{2}+2\right)},\cdots,\tilde{m}_{c,1}^{\left(n\frac{\alpha_{\epsilon}}{2}\right)},\tilde{m}_{c,2}^{\left(n\frac{\alpha_{\epsilon}}{2}\right)}\right).\end{aligned}$$ Similarly, the delivery codebook ${\mathcal{C}}_{{\bold{d}},n}$ can be seen as a wiretap code with $2^{n(1-\alpha_{2,\epsilon})}$ bins, each of which is indexed by the message $$\begin{aligned} \label{eq:w_d} w_{{\bold{d}}}=\left(m_{{\bold{d}}},\tilde{m}_{{\bold{d}},1}^{\left(n\frac{\alpha_{\epsilon}}{2}\right)},\tilde{m}_{{\bold{d}},2}^{\left(n\frac{\alpha_{\epsilon}}{2}\right)},\tilde{m}_{{\bold{d}},1}^{\left(n\frac{\alpha_{\epsilon}}{2}-1\right)},\tilde{m}_{{\bold{d}},2}^{\left(n\frac{\alpha_{\epsilon}}{2}-1\right)},\cdots,\tilde{m}_{{\bold{d}},1}^{\left(n\frac{\alpha_{2,\epsilon}}{2}+1\right)},\tilde{m}_{{\bold{d}},2}^{\left(n\frac{\alpha_{2,\epsilon}}{2}+1\right)}\right).\end{aligned}$$ Each bin $w_{{\bold{d}}}$ contains $2^{n\alpha_{2,\epsilon}}$ codewords, indexed by the randomization message $$\begin{aligned} \label{eq:w_d_tilde} \tilde{w}_{{\bold{d}}}=\left(\tilde{m}_{{\bold{d}},1}^{\left(n\frac{\alpha_{2,\epsilon}}{2}\right)},\tilde{m}_{{\bold{d}},2}^{\left(n\frac{\alpha_{2,\epsilon}}{2}\right)},\tilde{m}_{{\bold{d}},1}^{\left(n\frac{\alpha_{2,\epsilon}}{2}-1\right)},\tilde{m}_{{\bold{d}},2}^{\left(n\frac{\alpha_{2,\epsilon}}{2}-1\right)},\cdots,\tilde{m}_{{\bold{d}},1}^{(1)},\tilde{m}_{{\bold{d}},2}^{(1)}\right). \end{aligned}$$ Let $\left\{{\mathcal{B}}_{w_c}:\;w_c=1,2,\cdots,2^{n(1-\alpha_{1,\epsilon})}\right\}$ and $\big\{{\mathcal{B}}_{w_{{\bold{d}}}}:\;w_{{\bold{d}}}=1,2,\cdots, 2^{n(1-\alpha_{2,\epsilon})}\big\}$ denote the partition, i.e., bins, of the codebooks ${\mathcal{C}}_{c,n}$ and ${\mathcal{C}}_{{\bold{d}},n}$, which correspond to the messages $w_c$ and $w_{{\bold{d}}}$ in (\[eq:w\_c\]) and (\[eq:w\_d\]), respectively. Let ${\bold{x}}^{2n}\triangleq({\bold{x}}_{c}^n,{\bold{x}}_{{\bold{d}}}^n)$ denote the concatenation of the two length-$n$ binary codewords ${\bold{x}}_c^n$, ${\bold{x}}_{{\bold{d}}}^n$. Define the Cartesian product of the bins ${\mathcal{B}}_{w_c}$ and ${\mathcal{B}}_{w_{{\bold{d}}}}$, as $$\begin{aligned} \label{eq:partition_B_w} &{\mathcal{B}}_{w_c,w_{{\bold{d}}}}\triangleq \left\{{\bold{x}}^{2n}=({\bold{x}}_{c}^n,{\bold{x}}_{{\bold{d}}}^n):\;{\bold{x}}_c^n\in{\mathcal{B}}_{w_c},\;{\bold{x}}_{{\bold{d}}}^n\in {\mathcal{B}}_{w_{{\bold{d}}}}\right\}.\end{aligned}$$ Since the partitioning of the codebooks ${\mathcal{C}}_{c,n}$ and ${\mathcal{C}}_{{\bold{d}},n}$ is random, for every $w_c$ and $w_{{\bold{d}}}$, ${\mathcal{B}}_{w_c,w_{{\bold{d}}}}$ is a random codebook which results from the Cartesian product of the random bins ${\mathcal{B}}_{w_c}$, ${\mathcal{B}}_{w_{{\bold{d}}}}$. Recall that ${\mathcal{B}}_{w_c}$ contains $2^{n\alpha_{1,\epsilon}}$ and ${\mathcal{B}}_{w_{{\bold{d}}}}$ contains $2^{n\alpha_{2,\epsilon}}$ length-$n$ binary codewords. Thus, the product ${\mathcal{B}}_{w_c,w_{{\bold{d}}}}$ contains $2^{n\alpha_{\epsilon}}$ length-$2n$ binary codewords. Let $\Big\{W_{d_l,s}^{(1)},W_{d_l,s}^{(2)},\cdots,W_{d_l,s}^{\left(n\frac{\alpha_{\epsilon}}{2}\right)}\Big\}$ and $\Big\{K_l^{(1)},K_l^{(2)},\cdots,K_l^{\left(n\frac{\alpha_{\epsilon}}{2}\right)}\Big\}$ denote the binary bit strings of $W_{d_l,s}$ and $K_l$, $l=1,2$. In addition, for notational simplicity, define $$\begin{aligned} \label{eq:Ws-1} {\bold{W}}_{s}^{(1)}&=\left\{W_{d_1,s}^{(1)},W_{d_2,s}^{(1)},\cdots,W_{d_1,s}^{\left(n\frac{\alpha_{2,\epsilon}}{2}\right)},W_{d_2,s}^{\left(n\frac{\alpha_{2,\epsilon}}{2}\right)}\right\},\\ \label{eq:Ws-2} {\bold{W}}_{s}^{(2)}&=\left\{W_{d_1,s}^{\left(n\frac{\alpha_{2,\epsilon}}{2}+1\right)},W_{d_2,s}^{\left(n\frac{\alpha_{2,\epsilon}}{2}+1\right)},\cdots,W_{d_1,s}^{\left(n\frac{\alpha_{\epsilon}}{2}\right)},W_{d_2,s}^{\left(n\frac{\alpha_{\epsilon}}{2}\right)}\right\},\\ \label{eq:K-1} {\bold{K}}^{(1)}&=\left\{K_1^{(1)},K_2^{(1)},\cdots,K_1^{\left(n\frac{\alpha_{2,\epsilon}}{2}\right)},K_2^{\left(n\frac{\alpha_{2,\epsilon}}{2}\right)}\right\},\\ \label{eq:K-2} {\bold{K}}^{(2)}&=\left\{K_1^{\left(n\frac{\alpha_{2,\epsilon}}{2}+1\right)},K_2^{\left(n\frac{\alpha_{2,\epsilon}}{2}+1\right)},\cdots,K_1^{\left(n\frac{\alpha_{\epsilon}}{2}\right)},K_2^{\left(n\frac{\alpha_{\epsilon}}{2}\right)}\right\},\\ \label{eq:K-Ws-1} {\bold{W}}_{\oplus{\bold{K}}}^{(1)}&=\left\{W_{d_1,s}^{(i)}\oplus K_1^{(i)},W_{d_2,s}^{(i)}\oplus K_2^{(i)}:\; \; i=1,2,\cdots,n\frac{\alpha_{2,\epsilon}}{2}\right\},\\ \label{eq:K-Ws-2} {\bold{W}}_{\oplus{\bold{K}}}^{(2)}&=\left\{W_{d_1,s}^{(i)}\oplus K_1^{(i)},W_{d_2,s}^{(i)}\oplus K_2^{(i)}:\; \; i=n\frac{\alpha_{2,\epsilon}}{2}+1,\;n\frac{\alpha_{2,\epsilon}}{2}+2,\cdots,n\frac{\alpha_{\epsilon}}{2}\right\}.\end{aligned}$$ Let $W_c$, $\tilde{W}_c$, $W_{{\bold{d}}}$, and $\tilde{W}_{{\bold{d}}}$ denote the random variables that correspond to the realizations defined in (\[eq:w\_c\])–(\[eq:w\_d\_tilde\]). Using (\[eq:M\_c\])–(\[eq:M\_d\_tilde\]), (\[eq:w\_c\])–(\[eq:w\_d\_tilde\]), and (\[eq:K-1\])–(\[eq:K-Ws-2\]), we have $$\begin{aligned} \label{eq:Wc-Wc_tilde} &W_c=\left\{M_c,{\bold{K}}^{(1)}\right\}=\left\{W_1^{(1)}\oplus W_2^{(1)},W_1^{(2)}\oplus W_2^{(2)},{\bold{K}}^{(1)}\right\},\qquad \tilde{W}_c={\bold{K}}^{(2)}\\ \label{eq:Wd-Wd_tilde} &W_{{\bold{d}}}=\left\{M_{{\bold{d}}},{\bold{W}}_{\oplus{\bold{K}}}^{(2)}\right\}=\left\{W_{d_1}^{(2)},W_{d_2}^{(1)},{\bold{W}}_{\oplus{\bold{K}}}^{(2)}\right\},\qquad \tilde{W}_{{\bold{d}}}={\bold{W}}_{\oplus{\bold{K}}}^{(1)}.\end{aligned}$$ Notice that $\tilde{W}_c$ and $\tilde{W}_{{\bold{d}}}$ are independent, and each is uniformly distributed. $\{\tilde{W}_c,\tilde{W}_{{\bold{d}}}\}$ is thus jointly uniform. In addition, $\{\tilde{W}_c,\tilde{W}_{{\bold{d}}}\}$ is independent from $\{W_c,W_{{\bold{d}}}\}$. Thus, we can apply the analysis in [@goldfeld2015semantic ($94$)-($103$)] to show that, for every $S_1$, $S_2$, $w_c$, and $w_{{\bold{d}}}$, and every $\epsilon>0$, there exists $\gamma(\epsilon)>0$ such that $$\begin{aligned} \label{eq:Proof_1_1} \mathbb{P}_{{\mathcal{B}}_{w_c,w_{{\bold{d}}}}}\left(\mathbb{D}\left(P_{{\bold{Z}}_{S_1}^n{\bold{Z}}_{S_2}^n\big\|W_c=w_c,W_{{\bold{d}}}=w_{{\bold{d}}}}\|\|P_{{\bold{Z}}_{S_1}^n{\bold{Z}}_{S_2}^n}\right)>\epsilon\right)\leq \exp\left(-e^{n\gamma(\epsilon)}\right).\end{aligned}$$ $P_{{\bold{Z}}_{S_1}^n{\bold{Z}}_{S_2}^n\big\|W_c=w_c,W_{{\bold{d}}}=w_{{\bold{d}}}}$ is the induced distribution at the adversary when ${\bold{x}}_c^n(w_c,\tilde{w}_c)$ and ${\bold{x}}_{{\bold{d}}}^n(w_{{\bold{d}}},\tilde{w}_{{\bold{d}}})$ are the transmitted codewords over cache placement and delivery phases. $P_{{\bold{Z}}_{S_1}^n{\bold{Z}}_{S_2}^n}$ is the output distribution at the adversary. The number of the messages $\{w_c,w_{{\bold{d}}}\}$ is $2^{n(2-\alpha_{\epsilon})}$. Additionally, the number of possible choices for the subsets $S_1$ and $S_2$ is $\binom{2n}{\alpha n}<2^{2n}$. Thus, the combined number of the messages and the subsets is at most exponential in $n$. Using (\[eq:Proof\_1\_1\]) and the union bound, as in [@goldfeld2015semantic; @nafea2018new], we have $$\begin{aligned} \label{eq:Proof_1_2} {{\underset{n\rightarrow\infty}\lim}}\max_{S_1,S_2}I\left(W_c,W_{{\bold{d}}};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)=0.\end{aligned}$$ For the sake of completeness, we provide the full proofs for (\[eq:Proof\_1\_1\]) and (\[eq:Proof\_1\_2\]) in Appendix \[AppendixF\]. We also have, for any ${\bold{d}}=(d_1,d_2)$, $d_1,d_2\in\{1,2\}$, $$\begin{aligned} \label{eq:Proof_1_3_0} \nonumber &I\left(W_1,W_2;{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\\ &=I\left(W_1^{(1)},W_1^{(2)},W_2^{(1)},W_2^{(2)},W_{1,s},W_{2,s};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\\ \label{eq:Proof_1_3} &= I\left(W_1^{(1)},W_1^{(2)},W_2^{(1)},W_2^{(2)},{\bold{W}}_{s}^{(1)},{\bold{W}}_{s}^{(2)};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\\ \label{eq:Proof_1_4} &=I\left(W_1^{(1)}\oplus W_2^{(1)},W_1^{(2)}\oplus W_2^{(2)},W_{d_1}^{(2)},W_{d_2}^{(1)},{\bold{W}}_{s}^{(1)},{\bold{W}}_{s}^{(2)};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\\ \label{eq:Proof_1_5} &=I\left(M_c,M_{{\bold{d}}},{\bold{W}}_{s}^{(1)},{\bold{W}}_{s}^{(2)};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\\ \label{eq:Proof_1_6} &\leq I\left(M_c,M_{{\bold{d}}},{\bold{W}}_{s}^{(1)},{\bold{W}}_{\oplus{\bold{K}}}^{(2)};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\\ \label{eq:Proof_1_7} &=I\left(M_c,{\bold{W}}_s^{(1)},W_{{\bold{d}}};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\\ \label{eq:Proof_1_8} &= H\left({\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)-H\left({\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\Big\|M_c,{\bold{W}}_s^{(1)},W_{{\bold{d}}}\right).\end{aligned}$$ Equation (\[eq:Proof\_1\_3\]) follows since, for ${\bold{d}}=(d_1,d_2)$, ${\bold{Z}}_{S_1}^n$ and ${\bold{Z}}_{S_2}^n$ depend only on $W_1^{(1)}$, $W_1^{(2)}$, $W_2^{(1)}$, $W_2^{(2)}$, $W_{d_1,s}$, and $W_{d_2,s}$, and by using (\[eq:Ws-1\]) and (\[eq:Ws-2\]). Equation (\[eq:Proof\_1\_4\]) follows because there exists a bijection between $\{W_1^{(1)},W_1^{(2)},W_2^{(1)},W_2^{(2)}\}$ and $\left\{W_1^{(1)}\oplus W_2^{(1)},W_1^{(2)}\oplus W_2^{(2)},W_{d_1}^{(2)},W_{d_2}^{(1)}\right\}$. Equation (\[eq:Proof\_1\_5\]) follows from (\[eq:M\_c\]) and (\[eq:M\_d\]). The inequality in (\[eq:Proof\_1\_6\]) follows due to the Markov chain ${\bold{W}}_s^{(2)}-\left\{M_c,M_{{\bold{d}}},{\bold{W}}_s^{(1)},{\bold{W}}_{\oplus{\bold{K}}}^{(2)}\right\}-\left\{{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right\}$, and the data processing inequality. This Markov chain holds because $\left\{M_c,M_{{\bold{d}}},{\bold{W}}_s^{(1)}\right\}$ are independent from $\left\{{\bold{W}}_s^{(2)},{\bold{K}}^{(2)}\right\}$, and only the encrypted information ${\bold{W}}_{\oplus{\bold{K}}}^{(2)}$ is transmitted. Equation (\[eq:Proof\_1\_7\]) follows from (\[eq:Wd-Wd\_tilde\]). The second term on the right hand side of (\[eq:Proof\_1\_8\]) can be lower bounded as $$\begin{aligned} \label{eq:Proof_1_9} H&\left({\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\Big\|M_c,{\bold{W}}_s^{(1)},W_{{\bold{d}}}\right)=H\left({\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n,{\bold{W}}_s^{(1)}\Big\|M_c,W_{{\bold{d}}}\right)-H\left({\bold{W}}_s^{(1)}\big\|M_c,W_{{\bold{d}}}\right)\\ \label{eq:Proof_1_10} \nonumber&=H\left({\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n,{\bold{W}}_s^{(1)},{\bold{W}}_{\oplus K}^{(1)}\Big\|M_c,W_{{\bold{d}}}\right)\\ &\qquad\qquad\qquad\qquad -H\left({\bold{W}}_{\oplus K}^{(1)}\Big\|M_c,W_{{\bold{d}}},{\bold{W}}_s^{(1)},{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)-H\left({\bold{W}}_s^{(1)}\right)\\ \label{eq:Proof_1_11} \nonumber&=H\left({\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n,{\bold{K}}^{(1)},{\bold{W}}_{\oplus K}^{(1)}\Big\|M_c,W_{{\bold{d}}}\right)\\ &\qquad\qquad\qquad\qquad -H\left({\bold{K}}^{(1)}\Big\|M_c,W_{{\bold{d}}},{\bold{W}}_s^{(1)},{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)-H\left({\bold{W}}_s^{(1)}\right)\\ \label{eq:Proof_1_12} &\geq H\left({\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n,{\bold{K}}^{(1)},{\bold{W}}_{\oplus K}^{(1)}\Big\|M_c,W_{{\bold{d}}}\right)-H\left({\bold{W}}_s^{(1)}\right)-\epsilon'_n\\ \label{eq:Proof_1_13} &\geq H\left({\bold{K}}^{(1)}\Big\|M_c,W_{{\bold{d}}}\right)+H\left({\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\Big\|M_c,{\bold{K}}^{(1)},W_{{\bold{d}}}\right)-H\left({\bold{W}}_s^{(1)}\right)-\epsilon'_n\\ \label{eq:Proof_1_14} &=H\left({\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\Big\|W_c,W_{{\bold{d}}}\right)+ H\left({\bold{K}}^{(1)}\right)-H\left({\bold{W}}_s^{(1)}\right)-\epsilon'_n\\ \label{eq:Proof_1_15} &= H\left({\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\Big\|W_c,W_{{\bold{d}}}\right)-\epsilon'_n,\end{aligned}$$ where $\epsilon'_n\rightarrow 0$ as $n\rightarrow \infty$. Equation (\[eq:Proof\_1\_10\]) follows since ${\bold{W}}_s^{(1)}$ is independent from $\{M_c,W_{{\bold{d}}}\}$. Equation (\[eq:Proof\_1\_11\]) follows because there exists a bijection between $\left\{{\bold{W}}_s^{(1)},{\bold{W}}_{\oplus K}^{(1)}\right\}$ and $\left\{{\bold{K}}^{(1)},{\bold{W}}_{\oplus K}^{(1)}\right\}$. Equation (\[eq:Proof\_1\_14\]) follows from (\[eq:Wc-Wc\_tilde\]), and since ${\bold{K}}^{(1)}$ is independent from $\{M_c,W_{{\bold{d}}}\}$. Equation (\[eq:Proof\_1\_15\]) follows since ${\bold{K}}^{(1)}$ and ${\bold{W}}_s^{(1)}$ are independent and identically distributed. The inequality in (\[eq:Proof\_1\_12\]) follows because, given $\left\{M_c,{\bold{W}}_s^{(1)},W_{{\bold{d}}}\right\}$, and for $n$ sufficiently large, the adversary can decode ${\bold{K}}^{(1)}$ using its observations ${\bold{Z}}_{S_1}^n$ and ${\bold{Z}}_{S_2}^n$. In particular, $\left\{M_c,{\bold{W}}_s^{(1)},W_{{\bold{d}}}\right\}$ determine a partition of the codebook into bins, each of which contains $2^{n\alpha_{\epsilon}}$ binary codewords. For $n$ sufficiently large, and given the values of $M_c$, ${\bold{W}}_s^{(1)}$, $W_{{\bold{d}}}$, i.e., the bin index, the adversary can determine the codeword index inside the bin, and hence decode ${\bold{K}}^{(1)}$. We conclude that, $H\left({\bold{K}}^{(1)}\Big\|M_c,W_{{\bold{d}}},{\bold{W}}_s^{(1)},{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\leq \epsilon'_n$, where $\epsilon'_n\rightarrow 0$ as $n\rightarrow\infty$. Substituting (\[eq:Proof\_1\_15\]) in (\[eq:Proof\_1\_8\]) gives $$\begin{aligned} \label{eq:Proof_1_16} I\left(W_1,W_2;{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\leq I\left(W_c,W_{{\bold{d}}};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)+\epsilon'_n.\end{aligned}$$ Using (\[eq:Proof\_1\_2\]) and (\[eq:Proof\_1\_16\]), the secrecy constraint in (\[eq:secrecy\]) is satisfied. Since $\epsilon_n\rightarrow 0$ as $n\rightarrow\infty$, we conclude that, the achievable strong secrecy file rate is $$\begin{aligned} \label{eq:Proof_1_18} R_s(\alpha)=2\times \frac{1-\alpha}{2}+ \frac{\alpha}{2}=1-\frac{\alpha}{2}.\end{aligned}$$ Although the cache placement and delivery codebooks, ${\mathcal{C}}_{c,n}$ and ${\mathcal{C}}_{{\bold{d}},n}$, are designed and generated disjointly, in the security analysis, we have considered the Cartesian products of the individual bins of the two codebooks. We were able to do so since the input distributions for generating the two codebooks are identical, i.e., uniform binary. Achievability for $\alpha\in[1,2]$: {#Achievability_2} ----------------------------------- For $\alpha\in [1,2]$, we adapt the achievability scheme described in Section \[Achievability\_1\] as follows. The messages $W_1$ and $W_2$ are uniform over $\left[1:2^{n\frac{2-\alpha_{\epsilon}}{2}}\right]$; $\alpha_{\epsilon}$ is defined in (\[eq:alpha\_epsilon\]). The transmitter generates the independent keys $K_1$, $K_2$, uniform over $\left[1:2^{n\frac{2-\alpha_{\epsilon}}{2}}\right]$, and independent from $W_1$, $W_2$. In addition, the transmitter, independently from $W_1$, $W_2$, $K_1$, $K_2$, generates the independent randomization messages $\tilde{W}$ and $\tilde{W}_K$, uniformly over $\left[1:2^{n(\alpha_{\epsilon}-1)}\right]$. The messages for cache placement at receivers $1$ and $2$ are $$\begin{aligned} \label{eq:Mc1-Mc2} M_{c,1}=K_1,\qquad M_{c,2}=K_2.\end{aligned}$$ That is, receiver $j$, $j=1,2,$ stores the key $K_j$ in its cache memory. The message to be securely transmitted during delivery is $$\begin{aligned} \label{eq:Md1-Md2} M_{{\bold{d}}}=\left\{M_{{\bold{d}},1},M_{{\bold{d}},2}\right\};\;\;M_{{\bold{d}},1}=W_{d_1}\oplus K_1,\;\; M_{{\bold{d}},2}=W_{d_2}\oplus K_2.\end{aligned}$$ Let $\Big\{W_{d_l}^{(1)},\cdots,W_{d_l}^{\left(n\frac{2-\alpha_{\epsilon}}{2}\right)}\Big\}$, $\Big\{K_l^{(1)},\cdots,K_l^{(n\frac{2-\alpha_{\epsilon}}{2})}\Big\}$, and $\Big\{M_{{\bold{d}},l}^{(1)},\cdots,M_{{\bold{d}},l}^{(n\frac{2-\alpha_{\epsilon}}{2})}\Big\}$ denote the bit strings of $W_{d_l}$, $K_l$, and $M_{{\bold{d}},l}$, $l=1,2$. Notice that, for $\alpha\in[1,2]$, the adversary can see all symbols in at least one of the phases. Hence, unlike Section \[Achievability\_1\], we cannot utilize randomization messages, $\tilde{W}$ and $\tilde{W}_K$, to carry any information; only keys are stored in the cache memories of the receivers. Additionally, the cache placement and delivery codebooks for this case have a different embedding structure than for $\alpha\in(0,1)$ in Section \[Achievability\_1\]. In particular, the embedding here is performed on the bits of the messages $M_c$ and $M_{{\bold{d}}}$, while the embedding in Section \[Achievability\_1\] is performed on the bits of the randomization messages $\tilde{M}_c$ and $\tilde{M}_{{\bold{d}}}$. [[Cache Placement Codebook Generation:]{}]{} During cache placement, the transmitter generates ${\mathcal{C}}_{c,n}$ as follows. The transmitter randomly and independently divides all the $2^n$ length-$n$ binary sequences into $2$ bins, indexed by $K_1^{(1)}$, and each contains $2^{n-1}$ codewords. These two bins are further randomly and independently divided into two sub-bins, indexed by $K_2^{(1)}$, and each contains $2^{n-2}$ codewords. The process continues, going over $K_1^{(2)}$, $K_2^{(2)}$, $\cdots$, $K_1^{\left(n\frac{2-\alpha_{\epsilon}}{2}\right)}$, $K_2^{\left(n\frac{2-\alpha_{\epsilon}}{2}\right)}$, until the remaining $2^{n(\alpha_{\epsilon}-1)}$ codewords, after each sequence of divisions, are indexed by $\tilde{W}_{K}$. [[Cache Encoder:]{}]{} The transmitter sends $\bold{X}_c^n$ which corresponds to the keys $K_1$, $K_2$, and the randomization message $\tilde{W}_K$, i.e., $\bold{X}_c^n\left(K_1^{(1)},K_2^{(1)},\cdots,K_1^{\left(n\frac{2-\alpha_{\epsilon}}{2}\right)}, K_2^{\left(n\frac{2-\alpha_{\epsilon}}{2}\right)},\tilde{W}_K\right)$. [[Delivery Codebook Generation:]{}]{} In the delivery phase, the transmitter generates ${\mathcal{C}}_{{\bold{d}},n}$ as follows. The transmitter randomly and independently divides all the $2^n$ length-$n$ binary sequences into two bins, indexed by $M_{{\bold{d}},1}^{\left(n\frac{2-\alpha_{\epsilon}}{2}\right)}$, and each contains $2^{n-1}$ codewords. These two bins are further randomly and independently divided into two sub-bins, indexed by $M_{{\bold{d}},2}^{\left(n\frac{2-\alpha_{\epsilon}}{2}\right)}$, and each contains $2^{n-2}$ codewords. The process continues, going in reverse order over $M_{{\bold{d}},1}^{\left(n\frac{2-\alpha_{\epsilon}}{2}-1\right)}$, $M_{{\bold{d}},2}^{\left(n\frac{2-\alpha_{\epsilon}}{2}-1\right)}$, $\cdots$, $M_{{\bold{d}},1}^{(1)}$, $M_{{\bold{d}},2}^{(1)}$, until the remaining $2^{n(\alpha_{\epsilon}-1)}$ codewords, after each sequence of divisions, are indexed by the randomization message $\tilde{W}$. [[Delivery Encoder:]{}]{} Given $W_1$, $W_2$, $K_1$, $K_2$, $\tilde{W}$, and ${\bold{d}}=(d_1,d_2)$, the transmitter forms $M_{{\bold{d}},1}$ and $M_{{\bold{d}},2}$ as in (\[eq:Md1-Md2\]) and sends $\bold{X}_{{\bold{d}}}^n$ which corresponds to $M_{{\bold{d}},1}$, $M_{{\bold{d}},2}$, and $\tilde{W}$, i.e.,\ $\bold{X}_{{\bold{d}}}^n\Big(M_{{\bold{d}},1}^{\left(n\frac{2-\alpha_{\epsilon}}{2}\right)},M_{{\bold{d}},2}^{\left(n\frac{2-\alpha_{\epsilon}}{2}\right)},\cdots,M_{{\bold{d}},1}^{(1)}, M_{{\bold{d}},2}^{(1)},\tilde{W}\Big)$. [[Decoding:]{}]{} Using ${\bold{X}}_c^n$, receiver $j$, $j=1,2,$ recovers $M_{c,j}=K_j$ and stores it in its cache memory. Using ${\bold{X}}_{{\bold{d}}}^n$, both receivers recover $M_{{\bold{d}}}=\{M_{{\bold{d}},1},M_{{\bold{d}},2}\}$. Using $M_{{\bold{d}},j}$, $K_j$, and for $n$ sufficiently large, receiver $j$ correctly decodes $W_{d_j}$. [[Security Analysis:]{}]{} Fix the subsets $S_1$, $S_2$. Recall that $\alpha_1,\alpha_2\leq 1$. Since $\alpha\geq 1$, $\alpha_1,\alpha_2\geq \alpha -1$. If $\alpha_1=1$, then $\alpha_2=\alpha-1$, and vice versa. In addition, notice that $1-\alpha_1,1-\alpha_2\leq 2-\alpha$. As in Section \[Achievability\_1\], for a fixed value of $\alpha_1$, the codebook ${\mathcal{C}}_{c,n}$ is a wiretap code with $2^{n(1-\alpha_{1,\epsilon})}$ bins, indexed by $$\begin{aligned} \label{eq:Wc} W_c=\left(K_1^{(1)},K_2^{(1)},\cdots,K_1^{\left(n\frac{1-\alpha_{1,\epsilon}}{2}\right)},K_2^{\left(n\frac{1-\alpha_{1,\epsilon}}{2}\right)}\right).\end{aligned}$$ Each bin $W_c$ contains $2^{n\alpha_{1,\epsilon}}$ binary codewords, indexed by $$\begin{aligned} \label{eq:Wctilde} \tilde{W}_c=\left(K_1^{\left(n\frac{1-\alpha_{1,\epsilon}}{2}+1\right)},K_2^{\left(n\frac{1-\alpha_{1,\epsilon}}{2}+1\right)},\cdots,K_1^{\left(n\frac{2-\alpha_{\epsilon}}{2}\right)},K_2^{\left(n\frac{2-\alpha_{\epsilon}}{2}\right)},\tilde{W}_K\right).\end{aligned}$$ Similarly, for a fixed value of $\alpha_2$, the codebook ${\mathcal{C}}_{{\bold{d}},n}$ is a wiretap code with $2^{n(1-\alpha_{2,\epsilon})}$ bins, each is indexed by $$\begin{aligned} \label{eq:Wd} W_{{\bold{d}}}=\left(\tilde{M}_{{\bold{d}},1}^{\left(n\frac{2-\alpha_{\epsilon}}{2}\right)},\tilde{M}_{{\bold{d}},2}^{\left(n\frac{2-\alpha_{\epsilon}}{2}\right)},\cdots,\tilde{M}_{{\bold{d}},1}^{\left(n\frac{1-\alpha_{1,\epsilon}}{2}+1\right)},\tilde{M}_{{\bold{d}},2}^{\left(n\frac{1-\alpha_{1,\epsilon}}{2}+1\right)}\right).\end{aligned}$$ Each bin $W_{{\bold{d}}}$ contains $2^{n\alpha_{2,\epsilon}}$ codewords, indexed by $$\begin{aligned} \label{eq:Wdtilde} \tilde{W}_{{\bold{d}}}=\left(\tilde{M}_{{\bold{d}},1}^{\left(n\frac{1-\alpha_{1,\epsilon}}{2}\right)},\tilde{M}_{{\bold{d}},2}^{\left(n\frac{1-\alpha_{1,\epsilon}}{2}\right)},\cdots,\tilde{M}_{{\bold{d}},1}^{(1)},\tilde{M}_{{\bold{d}},2}^{(1)},\tilde{W}\right). \end{aligned}$$ Let us re-define $$\begin{aligned} \label{eq:K1} {\bold{K}}^{(1)}&=\left\{K_1^{(i)},K_2^{(i)}:\;\; i=1,\cdots,n\frac{1-\alpha_{1,\epsilon}}{2}\right\},\\ \label{eq:K2} {\bold{K}}^{(2)}&=\left\{K_1^{(i)},K_2^{(i)}:\;\; i=n\frac{1-\alpha_{1,\epsilon}}{2}+1,\cdots,n\frac{2-\alpha_{\epsilon}}{2}\right\},\\ \label{eq:K-W-1} {\bold{W}}_{\oplus{\bold{K}}}^{(1)}&=\left\{W_{d_1}^{(i)}\oplus K_1^{(i)},W_{d_2}^{(i)}\oplus K_2^{(i)}:\; \; i=1,\cdots,n\frac{1-\alpha_{1,\epsilon}}{2}\right\},\\ \label{eq:K-W-2} {\bold{W}}_{\oplus{\bold{K}}}^{(2)}&=\left\{W_{d_1}^{(i)}\oplus K_1^{(i)},W_{d_2}^{(i)}\oplus K_2^{(i)}:\; \; i=n\frac{1-\alpha_{1,\epsilon}}{2}+1,\cdots,n\frac{2-\alpha_{\epsilon}}{2}\right\},\end{aligned}$$ and define $$\begin{aligned} \label{eq:W-1} {\bold{W}}^{(1)}&=\left\{W_{d_1}^{(i)},W_{d_2}^{(i)}:\; \; i=1,\cdots,n\frac{1-\alpha_{1,\epsilon}}{2}\right\},\\ \label{eq:W-2} {\bold{W}}^{(2)}&=\left\{W_{d_1}^{(i)},W_{d_2}^{(i)}:\; \; i=n\frac{1-\alpha_{1,\epsilon}}{2}+1,\cdots,n\frac{2-\alpha_{\epsilon}}{2}\right\},\end{aligned}$$ From (\[eq:Wc\])-(\[eq:K-W-2\]), we have $$\begin{aligned} \label{eq:Wc-Wct-Wd-Wdt} &W_c={\bold{K}}^{(1)},\quad \tilde{W}_c=\left\{{\bold{K}}^{(2)},\tilde{W}_K\right\}, \quad W_{{\bold{d}}}={\bold{W}}_{\oplus{\bold{K}}}^{(2)},\quad \tilde{W}_{{\bold{d}}}=\left\{{\bold{W}}_{\oplus{\bold{K}}}^{(1)},\tilde{W}\right\}.\end{aligned}$$ Similar to Section \[Achievability\_1\], $\tilde{W}_c$, $\tilde{W}_{{\bold{d}}}$, are independent and uniformly distributed, and hence $\{\tilde{W}_c,\tilde{W}_{{\bold{d}}}\}$ is jointly uniform. Additionally, $\{\tilde{W}_c,\tilde{W}_{{\bold{d}}}\}$ is independent from $\{W_c,W_{{\bold{d}}}\}$. Thus, (\[eq:Proof\_1\_2\]) is satisfied. We also have, for any ${\bold{d}}=(d_1,d_2)$, $$\begin{aligned} \label{eq:Proof_2_2} I&\left(W_1,W_2;{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)=I\left({\bold{W}}^{(1)},{\bold{W}}^{(2)};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\\ \label{eq:Proof_2_3} &\leq I\left({\bold{W}}^{(1)},{\bold{W}}_{\oplus{\bold{K}}}^{(2)};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\\ \label{eq:Proof_2_4} &=I\left({\bold{W}}^{(1)},W_{{\bold{d}}};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\\ \label{eq:Proof_2_5} &= H\left({\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)-H\left({\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\Big\|{\bold{W}}^{(1)},W_{{\bold{d}}}\right)\\ \label{eq:Proof_2_6} &\leq H\left({\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)-H\left({\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\Big\|{\bold{K}}^{(1)},W_{{\bold{d}}}\right)+\epsilon'_n\\ \label{eq:Proof_2_7} &= I(W_c,W_{{\bold{d}}};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n)+\epsilon'_n.\end{aligned}$$ The inequality in (\[eq:Proof\_2\_3\]) follows due to the Markov chain ${\bold{W}}^{(2)}-\left\{{\bold{W}}^{(1)},{\bold{W}}_{\oplus{\bold{K}}}^{(2)}\right\}-\left\{{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right\}$. Equations (\[eq:Proof\_2\_4\]) and (\[eq:Proof\_2\_7\]) follow from (\[eq:Wc-Wct-Wd-Wdt\]). The inequality in (\[eq:Proof\_2\_6\]) follows by using similar steps as in (\[eq:Proof\_1\_9\])-(\[eq:Proof\_1\_15\]). Using (\[eq:Proof\_1\_2\]) and (\[eq:Proof\_2\_7\]), the secrecy constraint in (\[eq:secrecy\]) is satisfied. Since $\epsilon_n\rightarrow 0$ as $n\rightarrow\infty$, the achievable strong secrecy file rate is $$\begin{aligned} \label{eq:Proof_2_9} R_s(\alpha)= \frac{2-\alpha}{2}=1-\frac{\alpha}{2}.\end{aligned}$$ This completes the proof for Theorem \[thm:Thm1\]. Proof of Theorem \[thm:Thm2\] {#Proof_Thm2} ============================= In this section, we extend the achievability scheme presented in Section \[Proof\_Thm1\] and provide a lower bound on the achievable strong secrecy file rate when $D>2$. The demand vector is again denoted by ${\bold{d}}=(d_1,d_2)$, where $d_1,d_2\in[1:D]$. As in Section \[Proof\_Thm1\], we divide the proof into two cases for the ranges $\alpha\in(0,1)$ and $\alpha\in[1,2]$. $\alpha\in[1,2]$ {#Proof_Thm2_1} ---------------- For $\alpha\in[1,2]$, we utilize the same achievability scheme in Section \[Achievability\_2\]. The reason behind this is, for this range of $\alpha$, only the keys $K_1$, $K_2$, are transmitted in the cache placement, and stored in receivers $1$ and $2$ cache memories, respectively. That is, no information messages are stored in the caches, and the user demands are known during the delivery phase. The achievable strong secrecy file rate is $1-\frac{\alpha}{2}$. $\alpha\in(0,1)$ {#Proof_Thm2_2} ---------------- The achievability scheme for this case has the same channel coding structure as in the scheme described in Section \[Achievability\_1\]. The difference however lies in generating the messages to be securely communicated over cache placement and delivery phases, i.e., $M_c$ and $M_{{\bold{d}}}$. In particular, we utilize here uncoded placement for designing the cache contents, and a partially coded delivery transmission that is simultaneously useful for both receivers. The transmitter divides $W_l$, $l\in[1:D]$, into the independent messages $\left\{W_l^{(1)},W_l^{(2)},W_{l,t},W_{l,s}\right\}$. $W_l^{(1)}$, $W_l^{(2)}$, are uniform over $\left[1:2^{n\frac{1-\alpha_{\epsilon}}{2D}}\right]$; $\alpha_{\epsilon}$ is defined in (\[eq:alpha\_epsilon\]). $W_{l,t}$ is uniform over $\left[1:2^{n\frac{(2D-1)(1-\alpha_{\epsilon})}{4D}}\right]$, and $W_{l,s}$ is uniform over $\left[1:2^{n\frac{\alpha_{\epsilon}}{2}}\right]$. The transmitter, independently from $W_{[1:D]}$, generates the independent keys $K_1$, $K_2$, uniformly distributed over $\left[1:2^{n\frac{\alpha_{\epsilon}}{2}}\right]$. Let $M_c= \{M_{c,1},M_{c,2}\}$. Unlike (\[eq:M\_c\]), we utilize here [[uncoded placement]{}]{} for designing $M_{c,1}$ and $M_{c,2}$. We have, $$\begin{aligned} \label{eq:Mc1_LB_1} &M_{c,1}=\left\{W_1^{(1)},W_2^{(1)},\cdots,W_D^{(1)}\right\},\\ \label{eq:Mc2_LB_1} &M_{c,2}=\left\{W_1^{(2)},W_2^{(2)},\cdots,W_D^{(2)}\right\}.\end{aligned}$$ The randomization message for cache placement, $\tilde{M}_c=\{\tilde{M}_{c,1},\tilde{M}_{c,2}\}$, is identical to (\[eq:M\_c\_tilde\]). That is, $\tilde{M}_{c,1}=K_1$ and $\tilde{M}_{c,2}=K_2$. Receiver $j$ stores $M_{c,j}$ and $\tilde{M}_{c,j}$ in its cache memory. Unlike (\[eq:M\_d\]), we utilize here [[partially coded]{}]{} delivery. The message to be securely transmitted during the delivery phase is $$\begin{aligned} \label{eq:Md_LB_1} M_{{\bold{d}}}=\left\{W_{d_2}^{(1)}\oplus W_{d_1}^{(2)}, W_{d_1,t}, W_{d_2,t}\right\}.\end{aligned}$$ The randomization message for delivery, $\tilde{M}_{{\bold{d}}}$, is identical to (\[eq:M\_d\_tilde\]). Notice that the sizes of $M_c$, $M_{{\bold{d}}}$, $\tilde{M}_c$, and $\tilde{M}_{{\bold{d}}}$, are the same as in Section \[Achievability\_1\]. In particular, the sizes of $\tilde{M}_c$ and $\tilde{M}_{{\bold{d}}}$ are both $n\alpha_{\epsilon}$ bits. The size of $M_c$ is given by $$\begin{aligned} \label{eq:Proof_LB_1} 2\times D\times n\frac{1-\alpha_{\epsilon}}{2D}=n(1-\alpha_{\epsilon})\quad\text{bits},\end{aligned}$$ and the size of $M_{{\bold{d}}}$ is given by $$\begin{aligned} \label{eq:Proof_LB_2} n\frac{1-\alpha_{\epsilon}}{2D}+2\times n\frac{(2D-1)(1-\alpha_{\epsilon})}{4D}=n(1-\alpha_{\epsilon})\quad\text{bits.}\end{aligned}$$ [[Codebooks Generation and Encoders:]{}]{} For the messages $M_c$, $M_{{\bold{d}}}$, $\tilde{M}_c$, and $\tilde{M}_{{\bold{d}}}$ defined above, the cache placement and delivery codebooks, ${\mathcal{C}}_{c,n}$ and ${\mathcal{C}}_{{\bold{d}},n}$, and the cache and delivery encoders, are designed in the same exact manner as in Section \[Achievability\_1\], see Figures \[fig:code\_construction\_1\] and \[fig:code\_construction\_2\]. [[Decoding:]{}]{} As in Section \[Achievability\_1\], using $M_{{\bold{d}}}$, $\tilde{M}_{{\bold{d}}}$, $M_{c,j}$, $\tilde{M}_{c,j}$, and for $n$ sufficiently large, receiver $j$ correctly decodes $W_{d_{j}}$, $j=1,2$. [[Security analysis:]{}]{} Let $W_c$, $\tilde{W}_c$, $W_{{\bold{d}}}$, and $\tilde{W}_{{\bold{d}}}$, be defined as in (\[eq:w\_c\])-(\[eq:w\_d\_tilde\]), (\[eq:Wc-Wc\_tilde\]), and (\[eq:Wd-Wd\_tilde\]). Once again, $\tilde{W}_c$ and $\tilde{W}_{{\bold{d}}}$ are independent and uniform, and hence $\{\tilde{W}_c,\tilde{W}_{{\bold{d}}}\}$ is jointly uniform. In addition, $\{W_c,W_{{\bold{d}}}\}$ are independent from $\{\tilde{W}_c,\tilde{W}_{{\bold{d}}}\}$. Thus, (\[eq:Proof\_1\_2\]) holds for this case. For any ${\bold{d}}=(d_1,d_2)$, we have $$\begin{aligned} \label{eq:Proof_LB_4} &I\left(W_{[1:D]};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)=I\left(\left\{W_l^{(1)},W_l^{(2)},W_{l,t},W_{l,s}\right\}_{l=1}^D;{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\\ \label{eq:Proof_LB_5} &\leq I\left(M_c,\left\{W_l^{(1)},W_l^{(2)},W_{l,t},W_{l,s}\right\}_{l=1}^D;{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\\ \label{eq:Proof_LB_6} &\leq I\left(M_c,W_{d_2}^{(1)}\oplus W_{d_1}^{(2)},W_{d_1,t},W_{d_2,t},W_{d_1,s},W_{d_2,s};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\\ \label{eq:Proof_LB_7} &=I\left(M_c,M_{{\bold{d}}},W_{d_1,s},W_{d_2,s};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\\ \label{eq:Proof_LB_8} &\leq I\left(W_c,W_{{\bold{d}}};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)+\epsilon'_n,\end{aligned}$$ where (\[eq:Proof\_LB\_6\]) follows form the Markov chain $W_{[1:D]}-\left\{M_c,W_{d_2}^{(1)}\oplus W_{d_1}^{(2)},W_{d_1,t},W_{d_2,t},W_{d_1,s},W_{d_2,s}\right\}-\left\{{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right\}$; (\[eq:Proof\_LB\_7\]) follows from (\[eq:Md\_LB\_1\]), and (\[eq:Proof\_LB\_8\]) follows using similar steps as in (\[eq:Proof\_1\_4\])-(\[eq:Proof\_1\_15\]). Using (\[eq:Proof\_1\_2\]) and (\[eq:Proof\_LB\_8\]), the secrecy constraint in (\[eq:secrecy\]) is satisfied. With $\epsilon_n\rightarrow 0$ as $n\rightarrow\infty$, the achievable strong secrecy file rate is $$\begin{aligned} \label{eq:Proof_LB_9} R_s(\alpha)&=2\times \frac{1-\alpha}{2D}+\frac{(2D-1)(1-\alpha)}{4D}+\frac{\alpha}{2}\\ \label{eq:Proof_LB_10} &=\frac{1}{2}+\frac{3(1-\alpha)}{4D}.\end{aligned}$$ This completes the proof for Theorem \[thm:Thm2\]. For $D=2$, the achievable secrecy rate in (\[eq:Proof\_LB\_10\]) is strictly smaller than the secrecy rate obtained by [[coded]{}]{} placement and [[uncoded]{}]{} delivery in Section \[Achievability\_1\], i.e., $1-\frac{\alpha}{2}$. An achievability scheme which utilizes coded placement and uncoded delivery, as in Section \[Achievability\_1\], achieves the same secure file rate as (\[eq:Proof\_LB\_10\]) for $D=3$. However, this scheme achieves a strictly smaller secure file rate for $D\geq 4$. In this scheme, $W_l^{(1)}$ and $W_l^{(2)}$ are uniform over $\left[1:2^{n\frac{1-\alpha_{\epsilon}}{2(D-1)}}\right]$. $W_{l,t}$ is uniform over $\left[1:2^{n\frac{(D-2)(1-\alpha_{\epsilon})}{2(D-1)}}\right]$. $W_{l,s}$, $K_1$ and $K_2$, are uniform over $\left[1:2^{n\frac{\alpha_{\epsilon}}{2}}\right]$. $M_c=\{M_{c,1},M_{c,2}\}$ and $M_{{\bold{d}}}$ are given by $$\begin{aligned} \label{eq:Mc1_LB_2} M_{c,1}&=\left\{W_1^{(1)}\oplus W_2^{(1)}, W_2^{(1)}\oplus W_3^{(1)},\cdots,W_{D-1}^{(1)}\oplus W_D^{(1)}\right\},\\ \label{eq:Mc2_LB_2} M_{c,2}&=\left\{W_1^{(2)}\oplus W_2^{(2)}, W_2^{(2)}\oplus W_3^{(2)},\cdots,W_{D-1}^{(2)}\oplus W_D^{(2)}\right\},\\ \label{eq:Md_LB} M_{{\bold{d}}}&=\left\{W_{d_2}^{(1)}, W_{d_1}^{(2)}, W_{d_1,t}, W_{d_2,t}\right\}.\end{aligned}$$ Without loss of generality, let $d_1<d_2$. For any ${\bold{d}}=(d_1,d_2)$, using $M_{c,j}$ in (\[eq:Mc1\_LB\_2\]) and (\[eq:Mc2\_LB\_2\]), receiver $j$, can restore $W_{d_1}^{(j)}\oplus W_{d_2}^{(j)}$ by xor-ing $\big\{W_{d_1}^{(j)}\oplus W_{d_1+1}^{(j)}\big\}$, $\big\{W_{d_1+1}^{(j)}\oplus W_{d_1+2}^{(j)}\big\}$,$\cdots$, $\big\{W_{d_2-1}^{(j)}\oplus W_{d_2}^{(j)}\big\}$. The achievable strong secrecy file rate using this scheme is $R_s(\alpha)=\frac{1}{2}+\frac{1-\alpha}{2(D-1)}$. Proof of Theorem \[thm:Thm3\] {#Proof_Thm3} ============================= When $\alpha\in[1,2]$, the upper bound on $R_s$, stated in Theorem \[thm:Thm3\] for $D>2$, follows as in Section \[Converse\]. Thus, it remains to prove the upper bound for $\alpha\in(0,1)$. The proof is divided into the three following steps. [[Step $1$:]{}]{} We first upper bound $R_s$ by the secrecy capacity when the adversary is restricted to tap into the delivery transmission only, denoted as $C_s^{\rm{Res}}$. That is, $C_s^{\rm{Res}}$ is the maximum achievable file rate when $\alpha_1=0$ and $\alpha_2=\alpha$. Restricting the adversary to only tap into the delivery phase cannot decrease the secrecy capacity, i.e., $R_s\leq C_s^{\rm{Res}}$, since this setting is included in the feasible strategy space for the adversary. The cache placement transmission is secure, and each receiver has a secure cache memory of size $\frac{n}{2}$ bits. [[Step $2$:]{}]{} We upper bound $C_s^{\rm{Res}}$ by the secrecy capacity, i.e., the maximum achievable file rate, when the delivery channel to the adversary is replaced by a discrete memoryless binary erasure channel, with erasure probability $1-\alpha$, denoted as $C_s^{\rm{DM}}$. The proof for this step follows the same lines as in [@nafea2018new Section V]. The idea is when the binary erasure channel produces a number of erasures greater than or equal to $(1-\alpha)n$, the adversary’s channel in this discrete memoryless setup is worse than its channel in the former model, i.e., when it encounters exactly $(1-\alpha)n$ erasures and is able to select their positions. Hence, $C_s^{\rm{Res}}\leq C_s^{\rm{DM}}$ for this case. The result follows by utilizing Sanov’s theorem in method of types [@cover2006elements Theorem 11.4.1] to show that the probability of the binary erasure channel causing a number of erasures less than $(1-\alpha)n$ goes to zero as $n\rightarrow\infty$. [[Step $3$:]{}]{} From Step $1$, each receiver has a secure cache of size $\frac{n}{2}$ bits. Since increasing the cache sizes cannot decrease the achievable file rate, we further upper bound $C_s^{\rm{DM}}$ with the maximum achievable file rate when each receiver has a cache memory of size $n$ bits, in which it stores ${\bold{X}}_c^n$. Receiver $j$, $j=1,2,$ utilizes both ${\bold{X}}_c^n$ and ${\bold{X}}_{{\bold{d}}}^n$ in order to decode its desired message $W_{d_j}$, i.e., $\hat{W}_{d_j}=g_{{\bold{d}},j}\left({\bold{X}}_c^n,{\bold{X}}_{{\bold{d}}}^n\right)$, ${\bold{d}}=(d_1,d_2)$. This setup is thus equivalent to a single receiver, with a cache of size $n$ bits, who demands two files $W_{d_1}$, $W_{d_2}$, and utilizes the decoder $g_{{\bold{d}}}\triangleq \{g_{{\bold{d}},1},g_{{\bold{d}},2}\}$. Let us denote the maximum achievable file rate for this single receiver model as $C_s^{\rm{SR}}$. We have $C_s^{\rm{DM}}\leq C_s^{\rm{SR}}$. In the following, we upper bound $C_s^{\rm{SR}}$. Let $M_{\rm{D}}$ denote the fraction of the size-$n$ bits cache memory dedicated to store (coded or uncoded) information bits, and let $M_{\rm{K}}$ denote the fraction dedicated to store key bits. That is, $M_{\rm{D}}+M_{\rm{K}}=1$. Let $S_{\rm{D}}$ denote the information bits stored in this memory, i.e., $S_{\rm{D}}=f_{\rm{D}}(W_{[1:D]})$ and $H(S_{\rm{D}})=n M_{\rm{D}}$. We utilize the following lemma in order to upper bound $C_s^{\rm{SR}}$. [@zewail2018wiretap Lemma 1] For a fixed allocation of $M_{\rm{D}}$ and $M_{\rm{K}}$, and a receiver who demands the files $W_{d_1}$ and $W_{d_2}$, the secrecy rate for the single receiver model is upper bounded as $$\begin{aligned} \label{eq:UB_1} 2R_s^{\rm{SR}}\leq \min\left\{1,1-\alpha+M_{\rm{K}}\right\}+\frac{1}{n} I\left(W_{d_1},W_{d_2};S_{\rm{D}}\right).\end{aligned}$$ Notice that (\[eq:UB\_1\]) holds for any demand pair ${\bold{d}}=(d_1,d_2)$ such that $d_1\neq d_2$, i.e., the worst-case demands. Summing over all such demands, we have $$\begin{aligned} \label{eq:UB_2} 2R_s^{\rm{SR}}\leq \min\left\{1,1-\alpha+M_{\rm{K}}\right\}+\frac{1}{nD(D-1)} \sum_{d_1,d_2\in[1:D],\;d_1\neq d_2} I\left(W_{d_1},W_{d_2};S_{\rm{D}}\right).\end{aligned}$$ The second term on the right hand side of (\[eq:UB\_2\]) can be written as $$\begin{aligned} \label{eq:UB_3_1} \nonumber &\frac{1}{nD(D-1)} \sum_{d_1,d_2\in[1:D],\;d_1\neq d_2} I\left(W_{d_1},W_{d_2};S_{\rm{D}}\right)\\ &\qquad=\frac{1}{nD} \sum_{d_1\in[1:D]}\;I\left(W_{d_1};S_{\rm{D}}\right)+\frac{1}{nD(D-1)}\sum_{d_1,d_2\in[1:D],\; d_1\neq d_2}\;I\left(W_{d_2};S_{\rm{D}}\|W_{d_1}\right)\\ \label{eq:UB_3} &\qquad\leq \frac{1}{nD} \sum_{d_1\in[1:D]}\;I\left(W_{d_1};S_{\rm{D}}\right)+\frac{1}{nD(D-1)}\sum_{d_1\in[1:D]}\left(\sum_{d_2\in[1:D]}\;I\left(W_{d_2};S_{\rm{D}}\|W_{d_1}\right)\right).\end{aligned}$$ For any $d_1\in[1:D]$, we have $$\begin{aligned} \label{eq:UB_4} &\sum_{d_2\in[1:D]}\;I\left(W_{d_2};S_{\rm{D}}\big\|W_{d_1}\right)=\sum_{d_2=1}^D\;\left[H\left(W_{d_2}\big\|W_{d_1}\right)-H\left(W_{d_2}\big\|W_{d_1},S_{\rm{D}}\right)\right]\\ \label{eq:UB_5} &\leq \sum_{d_2=1}^D\;\left[H\left(W_{d_2}\big\|W_1,W_2,\cdots,W_{d_2-1},W_{d_1}\right)-H\left(W_{d_2}\big\|W_1,W_2,\cdots,W_{d_2-1},W_{d_1},S_{\rm{D}}\right)\right]\\ \label{eq:UB_6} &=I\left(W_1,W_2,\cdots,W_{\rm{D}};S_{\rm{D}}\big\|W_{d_1}\right)\\ \label{eq:UB_7} &\leq H(S_{\rm{D}})=nM_{\rm{D}},\end{aligned}$$ where (\[eq:UB\_5\]) follows because when $d_2=d_1$, $H\left(W_{d_2}\|W_{d_1}\right)=H\left(W_{d_2}\|W_1,W_2,\cdots,W_{d_2-1},W_{d_1}\right)=0$, and when $d_2\neq d_1$, $H\left(W_{d_2}\|W_{d_1}\right)=H\left(W_{d_2}\|W_1,W_2,\cdots,W_{d_2-1},W_{d_1}\right)=H(W_{d_2})$. Similarly, we have $$\begin{aligned} \label{eq:UB_8} &\sum_{d_1\in[1:D]}\;I\left(W_{d_1};S_{\rm{D}}\right)\leq H(S_{\rm{D}})=nM_{\rm{D}}.\end{aligned}$$ Substituting (\[eq:UB\_7\]) and (\[eq:UB\_8\]) in (\[eq:UB\_3\]) gives $$\begin{aligned} \label{eq:UB_9} &\frac{1}{nD(D-1)} \sum_{d_1,d_2\in[1:D],\;d_1\neq d_2} I\left(W_{d_1},W_{d_2};S_{\rm{D}}\right)\leq \frac{2D-1}{D(D-1)}M_{\rm{D}}.\end{aligned}$$ Thus, using (\[eq:UB\_2\]) and (\[eq:UB\_9\]), $R_s^{\rm{SR}}$ is further upper bounded as $$\begin{aligned} \label{eq:UB_10} R_s^{\rm{SR}}\leq \frac{1}{2}\left[\min\left\{1,1-\alpha+M_{\rm{K}}\right\}+\frac{2D-1}{D(D-1)}M_{\rm{D}}\right].\end{aligned}$$ Finally, by maximizing over all possible allocations for $M_{\rm{D}}$ and $M_{\rm{K}}$ such that $M_{\rm{D}}+M_{\rm{K}}=1$, we obtain $$\begin{aligned} \label{eq:UB_11} C_s^{\rm{SR}}&\leq \frac{1}{2}\underset{\begin{subarray}{c}M_{\rm{D}},M_{\rm{K}}:\\ M_{\rm{D}}+M_{\rm{K}}=1\end{subarray}}\max\left\{\min\left\{1,1-\alpha+M_{\rm{K}}\right\}+\frac{2D-1}{D(D-1)}M_{\rm{D}}\right\}\\ \label{eq:UB_12} &=\frac{1}{2}\left[1+\frac{2D-1}{D(D-1)}(1-\alpha)\right].\end{aligned}$$ Equation (\[eq:UB\_12\]) follows because, for $D\geq 3$, the maximum occurs at $M_{\rm{K}}=\alpha$ and $M_{\rm{D}}=1-\alpha$. This completes the proof for Theorem \[thm:Thm3\]. An upper bound considering uncoded placement only can be derived as follows. The same analysis as in (\[eq:UB\_1\])-(\[eq:UB\_12\]) carries through with $I(W_{d_2};S_{\rm{D}}\|W_{d_1})$ in (\[eq:UB\_3\_1\]) is equal to $I(W_{d_2};S_{\rm{D}})$. Hence the right hand side of (\[eq:UB\_9\]) is replaced by $\frac{2M_{\rm{D}}}{D}$. The resulting bound $R_s\leq \frac{1}{2}+\frac{(1-\alpha)}{D}$ is tighter than (\[eq:UB\_12\]). Discussion {#Discussion} ========== While the fixed-size cache memory setup considered in this paper can be seen as a clean basic model for the intricate problem in consideration, it also allows us to obtain results and insights that are generalizable to more involved cache memory models. In particular, the extension to variable memory sizes can be done by considering multiple communication blocks for cache placement. Our results and coding scheme readily apply to an adversary model whose tapping capability during the delivery is normalized with respect to tapping during cache placement, i.e., $\mu_1+B\mu_2\leq \mu$; $B$ is the number of communication blocks for cache placement. This is a reasonable assumption given that cache placement generally takes place in a longer period than delivery. The problem turns to be more challenging when the adversary optimizes its tapping uniformly over the multiple blocks for cache placement as well as the delivery phase. This is left for future investigation. It is typical to model the cache placement as a noiseless channel since placement is assumed to occur when networks are not congested and their rates are assumed to be large enough. Here however we model the cache placement as a broadcast channel communication. The broadcast model avails a clean and tractable solution without compromising its generalizability. A time division multiple access (TDMA) model for cache placement is a special case by imposing an additional constraint in which each receiver has to decode its desired file using only one half of the transmitted codeword. Additionally, the broadcast model is in line with the network information theory literature and it does not limit the cache placement to occur over low rate traffic. With the ever-growing user demands, placement and delivery occurring in less asymmetric network loads is likely to be expected in the near future. Corollary \[Cor:Cor1\] demonstrates that, for the model considered in this paper, when $\alpha\in [1,2]$, the strong secrecy capacity is equal to $1-\frac{\alpha}{2}$ for any library size. For $\alpha\in [1,2]$, $\{S_1=[1:n], S_2\subset [1:n]\}$ and $\{S_1\subset [1:n], S_2=[1:n]\}$ are two possible strategies for the adversary. In other words, the adversary can tap into either all transmitted symbols in cache placement and a subset of symbols in the delivery, or all transmitted symbols in the delivery and a subset of symbols in cache placement. Such an adversary limits the communication for cache placement, i.e., the use of cache memories, to exchanging additional randomness (key bits) that allows for communicating a positive secure rate over the two phases. In other words, the cache memories are not utilized to store any data bits, and hence the lack of knowledge of user demands during cache placement is immaterial. For a library with two files, if the receivers were to have cache memories of size $n$ bits in which they store the transmitted signal during cache placement, the strong secrecy file rate in Theorem \[thm:Thm1\] is achievable using a simple wiretap code. In particular, the transmitter encodes $W=(W_1,W_2)\in[1:2^{n2R}]$ into a length-$2n$ binary codeword using a wiretap code, and sends the first $n$ bits of this codeword during cache placement and the last $n$ bits during delivery. Each receiver can thus decode both files, and the secrecy of $W_1$ and $W_2$ against the adversary follows by the results in [@goldfeld2015semantic; @nafea2018new]. In caching problems, the relevant setup however is when the receivers have cache memories of limited size with respect to the overall transmission during cache placement. This calls for the limited size cache memory model considered in this paper, which in turn necessitates the use of the more elaborate coding scheme in Section \[Proof\_Thm1\]. Conclusion {#Con} ========== We have introduced the caching broadcast channel with a [[a wire and cache]{}]{} tapping adversary of type II. In this broadcast model, each receiver is equipped with a fixed-size cache memory, and the adversary is able to tap into a subset of its choosing of the transmitted symbols during cache placement, or delivery, or both. The legitimate terminals have no knowledge about the fractions of the tapped symbols in each phase, or their positions. Only the size of the overall tapped set is known. We have identified the strong secrecy capacity of this model, i.e., the maximum achievable file rate while keeping the overall library secure, when the transmitter’s library has two files. We have derived lower and upper bounds for the strong secrecy file rate when the transmitter has more than two files in its library. We have devised an achievability scheme which combines wiretap coding, security embedding codes, one-time pad keys, and coded caching techniques. The results presented in this paper highlight the robustness of (stochastic) coding against a smart adversary who performs a chosen attack, jointly optimized over both cache placement and delivery phases. Future directions include investigating a tighter upper bound for a library with more than two files, and exploring the extensions of this work to variable cache memory sizes, more than two users, and a noisy legitimate channel. Acknowledgment {#acknowledgment .unnumbered} ============== The authors would like to thank Prof. Abbas El Gamal for useful discussions about the modeling details considered in this work, and for the suggestion to develop the results building on the simpler cases. Secrecy Constraint for Setting $1$: Proof of (\[eq:example1\_3\]) {#AppendixA} ================================================================= For every $S_1\subseteq [1:n]$ satisfying $\|S_1\|=\mu$, we have $$\begin{aligned} \label{eq:AppendixA_1} {{\underset{n\rightarrow\infty}\lim}}I\left(W_1,W_2;{\bold{Z}}_{S_1}^n\right)&={{\underset{n\rightarrow\infty}\lim}}I\left(W_1^{(1)}, W_1^{(2)},W_2^{(1)},W_2^{(2)},W_{1,s},W_{2,s};{\bold{Z}}_{S_1}^n\right)\\ \label{eq:AppendixA_2} &={{\underset{n\rightarrow\infty}\lim}}I\left(W_1^{(1)}, W_1^{(2)},W_2^{(1)},W_2^{(2)};{\bold{Z}}_{S_1}^n\right)\\ \label{eq:AppendixA_3} &\leq {{\underset{n\rightarrow\infty}\lim}}I\left(W_1^{(1)}\oplus W_2^{(1)},W_1^{(2)}\oplus W_2^{(2)};{\bold{Z}}_{S_1}^n\right)\\ \label{eq:AppendixA_4} &={{\underset{n\rightarrow\infty}\lim}}I\left(M_c;{\bold{Z}}_{S_1}^n\right)=0.\end{aligned}$$ Recall that the adversary’s observation over cache placement, ${\bold{Z}}_{S_1}^n$, results from sending $M_c=\{M_{c,1},M_{c,2}\}$, where $M_{c,1}=W_1^{(1)}\oplus W_2^{(1)}$ and $M_{c,2}=W_1^{(2)}\oplus W_2^{(2)}$. Thus, (\[eq:AppendixA\_2\]) follows because ${\bold{Z}}_{S_1}^n$ does not depend on $\{W_{1,s},W_{2,s}\}$ and (\[eq:AppendixA\_3\]) follows due to the Markov chain $\left\{W_1^{(1)}, W_1^{(2)},W_2^{(1)},W_2^{(2)}\right\}-\left\{W_1^{(1)}\oplus W_2^{(1)},W_1^{(2)}\oplus W_2^{(2)}\right\}-{\bold{Z}}_{S_1}^n$. The second equality in (\[eq:AppendixA\_4\]) follows from [@goldfeld2015semantic Theorem 2], and the fact that the rate of $M_c$ is less than $1-\alpha$. Secrecy Constraint for Setting $2$: Proof of (\[eq:example2\_5\]) {#AppendixB} ================================================================= For every $S_2\subseteq [1:n]$ satisfying $\|S_2\|=\mu$, and any ${\bold{d}}=(d_1,d_2)$, $d_1,d_2\in\{1,2\}$, we have $$\begin{aligned} \label{AppendixB_1} I\big(W_1,&W_2;{\bold{Z}}_{S_2}^n\big)=I\left(W_{d_1}^{(2)},W_{d_2}^{(1)},W_{d_1,s}, W_{d_2,s};{\bold{Z}}_{S_2}^n\right)\\ \label{AppendixB_2} &=I\left(W_{d_1}^{(2)},W_{d_2}^{(1)};{\bold{Z}}_{S_2}^n\right)+I\left(W_{d_1,s}, W_{d_2,s};{\bold{Z}}_{S_2}^n\Big\|W_{d_1}^{(2)},W_{d_2}^{(1)}\right)\\ \label{AppendixB_3} &\leq I\left(W_{d_1}^{(2)},W_{d_2}^{(1)};{\bold{Z}}_{S_2}^n\right)+I\left(W_{d_1,s}, W_{d_2,s};W_{d_1,s}\oplus K_1, W_{d_2,s}\oplus K_2\Big\|W_{d_1}^{(2)},W_{d_2}^{(1)}\right)\\ \label{AppendixB_4} &=I\left(W_{d_1}^{(2)},W_{d_2}^{(1)};{\bold{Z}}_{S_2}^n)+I(W_{d_1,s},W_{d_2,s};W_{d_1,s}\oplus K_1,W_{d_2,s}\oplus K_2\right)\\ \label{AppendixB_5} &=I\left(W_{d_1}^{(2)},W_{d_2}^{(1)};{\bold{Z}}_{S_2}^n\right)\\ \label{AppendixB_6} &=I\left(M_{{\bold{d}}};{\bold{Z}}_{S_2}^n\right).\end{aligned}$$ The adversary’s observation over the delivery phase, ${\bold{Z}}_{S_2}^n$, results from sending $M_{{\bold{d}}}=\{W_{d_1}^{(2)},W_{d_2}^{(1)}\}$ and the randomization message $\tilde{M}_{{\bold{d}}}=\left\{W_{d_1,s}\oplus K_1,W_{d_2,s}\oplus K_2\right\}$. Equation (\[AppendixB\_1\]) thus follows because ${\bold{Z}}_{S_2}^n$ depends only on $W_{d_1}^{(2)}$, $W_{d_2}^{(1)}$, $W_{d_1,s}$, and $W_{d_2,s}$. The inequality in (\[AppendixB\_3\]) follows from the Markov chain $\left\{W_{d_1,s},W_{d_2,s}\right\}-\left\{W_{d_1}^{(2)},W_{d_2}^{(1)},W_{d_1,s}\oplus K_1,W_{d_2,s}\oplus K_2\right\}-{\bold{Z}}_{S_2}^n$. Equation (\[AppendixB\_4\]) follows because $\left\{W_{d_1}^{(2)},W_{d_2}^{(1)}\right\}$ are independent from $\left\{W_{d_1,s},W_{d_2,s},K_1,K_2\right\}$. The randomization message of the wiretap code in the delivery phase, $\tilde{M}_{{\bold{d}}}$, is independent from the message $M_{{\bold{d}}}$. Thus, using (\[AppendixB\_6\]) and [@goldfeld2015semantic Theorem 2], we have $$\begin{aligned} \label{AppendixB_7} \nonumber {{\underset{n\rightarrow\infty}\lim}}\; &\max_{S_2\subseteq [1:n]: \|S_2\|=\mu}\;I(W_1,W_2;{\bold{Z}}_{S_2}^n)\\ &\leq {{\underset{n\rightarrow\infty}\lim}}\max_{S_2\subseteq [1:n]:\;\|S_2\|=\mu}\;I(M_{{\bold{d}}};{\bold{Z}}_{S_2}^n)=0.\end{aligned}$$ Secrecy Constraint for Setting $3$ when $\alpha_1\geq \alpha_2$ {#AppendixC} =============================================================== Recall that $M_c$ and $M_{{\bold{d}}}$ are defined as in (\[eq:example1\_1\]) and (\[eq:example1\_2\]), respectively. For a fixed choice of the subsets $S_1,S_2\subseteq [1:n]$ such that $\|S_1\|+\|S_2\|=\mu$, and any ${\bold{d}}=(d_1,d_2)$, $d_1,d_2\in\{1,2\}$, we have $$\begin{aligned} \label{AppendixC_1} I(W_1,&W_2;{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n)=I\left(W_1^{(1)},W_1^{(2)},W_2^{(1)},W_2^{(2)},W_{1,s},W_{2,s};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\\ \label{AppendixC_2} &=I\left(W_1^{(1)}\oplus W_2^{(1)},W_1^{(2)}\oplus W_2^{(2)},W_{d_1}^{(2)},W_{d_2}^{(1)},W_{d_1,s},W_{d_2,s};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\\ \label{AppendixC_3} &=I\left(M_c,M_{{\bold{d}}};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\\ \label{AppendixC_4} &=I\left(M_c;{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)+I\left(M_{{\bold{d}}};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\big\|M_c\right)\\ \label{AppendixC_5} &=I\left(M_c;{\bold{Z}}_{S_1}^n\right)+I\left(M_c;{\bold{Z}}_{S_2}^n\big\|{\bold{Z}}_{S_1}^n\right)+I\left(M_{{\bold{d}}};{\bold{Z}}_{S_2}^n\big\|M_c\right)+I\left(M_{{\bold{d}}};{\bold{Z}}_{S_1}^n\big\|M_c,{\bold{Z}}_{S_2}^n\right),\end{aligned}$$ where (\[AppendixC\_2\]) follows because, for any $d_1,d_2\in\{1,2\}$, there exists a bijective map between $\left\{W_1^{(1)},W_1^{(2)},W_2^{(1)},W_2^{(2)}\right\}$ and $\left\{W_1^{(1)}\oplus W_2^{(1)},W_1^{(2)}\oplus W_2^{(2)},W_{d_1}^{(2)},W_{d_2}^{(1)}\right\}$. From (\[eq:example1\_1\]) and (\[eq:example1\_2\]), $M_c$ and $M_{{\bold{d}}}$ are independent. The adversary’s observation in cache placement, ${\bold{Z}}_{S_1}^n$, results from sending $M_c$, while its observation in the delivery phase, ${\bold{Z}}_{S_2}^n$, results from sending $M_{{\bold{d}}}$. Thus, for a fixed choice of the subsets $S_1$ and $S_2$, $\left\{M_c,{\bold{Z}}_{S_1}^n\right\}$ are independent from ${\bold{Z}}_{S_2}^n$. We thus have $$\begin{aligned} \label{AppendixC_6} I\left(M_c;{\bold{Z}}_{S_2}^n\|{\bold{Z}}_{S_1}^n\right)=0.\end{aligned}$$ In addition, $\left\{M_{{\bold{d}}},{\bold{Z}}_{S_2}^n\right\}$ are independent from $M_c$. Thus, $$\begin{aligned} \label{AppendixC_7} I\left(M_{{\bold{d}}};{\bold{Z}}_{S_2}^n\|M_c\right)&=H\left({\bold{Z}}_{S_2}^n\|M_c\right)-H\left({\bold{Z}}_{S_2}^n\|M_c,M_{{\bold{d}}}\right)\\ \label{AppendixC_8} &=H\left({\bold{Z}}_{S_2}^n\|M_c\right)-H\left({\bold{Z}}_{S_2}^n\|M_{{\bold{d}}}\right)\\ \label{AppendixC_9} &\leq I\left(M_{{\bold{d}}};{\bold{Z}}_{S_2}^n\right).\end{aligned}$$ Finally, using the Markov chain $\left\{M_{{\bold{d}}},{\bold{Z}}_{S_2}^n\right\}-M_c-{\bold{Z}}_{S_1}^n$, we have $$\begin{aligned} \label{AppendixC_10} I\left(M_{{\bold{d}}};{\bold{Z}}_{S_1}^n\|M_c,{\bold{Z}}_{S_2}^n\right)&=H\left({\bold{Z}}_{S_1}^n\|M_c,{\bold{Z}}_{S_2}^n\right)-H\left({\bold{Z}}_{S_1}^n\|M_c,{\bold{Z}}_{S_2}^n,M_{{\bold{d}}}\right)\\ \label{AppendixC_11} &\leq H\left({\bold{Z}}_{S_1}^n\right)-H\left({\bold{Z}}_{S_1}^n\|M_c\right)=I\left(M_c;{\bold{Z}}_{S_1}^n\right).\end{aligned}$$ Substituting (\[AppendixC\_6\]), (\[AppendixC\_9\]), and (\[AppendixC\_11\]) in (\[AppendixC\_5\]), $$\begin{aligned} \label{AppendixC_12} I\left(W_1,W_2;{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\leq 2I\left(M_c;{\bold{Z}}_{S_1}^n\right)+I\left(M_{{\bold{d}}};{\bold{Z}}_{S_2}^n\right).\end{aligned}$$ The rates of $M_c$ and $M_{{\bold{d}}}$ are $1-\alpha_1-\epsilon_n$ and $1-\alpha_2-\epsilon_n$, respectively. By applying [@goldfeld2015semantic Theorem 2] to (\[AppendixC\_12\]), we have $$\begin{aligned} \nonumber {{\underset{n\rightarrow\infty}\lim}}&\;\underset{\begin{subarray}{c} S_1,S_2\subseteq [1:n]:\\\|S_1\|+\|S_2\|=\mu\end{subarray}}\max I\left(W_1,W_2;{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\\ \label{AppendixC_13} &\leq 2 {{\underset{n\rightarrow\infty}\lim}}\;\underset{S_1\subseteq[1:n]:\;\|S_1\|=\mu_1}\max\;I(M_c;{\bold{Z}}_{S_1}^n)+{{\underset{n\rightarrow\infty}\lim}}\;\underset{S_2\subseteq[1:n]:\;\|S_2\|=\mu_2}\max\;I\left(M_{{\bold{d}}};{\bold{Z}}_{S_2}^n\right)\\ \label{AppendixC_14} &=0.\end{aligned}$$ Secrecy Constraint for Setting $3$ when $\alpha_1< \alpha_2$ {#AppendixD} ============================================================ For this case, $M_c$ and $M_{{\bold{d}}}$ are defined in (\[eq:example2\_2\]) and (\[eq:example2\_3\]) and the randomization message $\tilde{M}_{{\bold{d}}}$ is defined in (\[eq:example2\_4\]). For notational simplicity, let us define $$\begin{aligned} \label{AppendixD_0_1} &M_{c,1\setminus K_1}=W_1^{(1)}\oplus W_2^{(1)},\qquad M_{c,2 \setminus K_2}=W_1^{(2)}\oplus W_2^{(2)}\\ \label{AppendixD_0_2} &M_{c\setminus K}=\left\{M_{c,1\setminus K_1},M_{c,2\setminus K_2}\right\}.\end{aligned}$$ For a fixed choice of $S_1,S_2\subseteq [1:n]$ such that $\|S_1\|+\|S_2\|=\mu$, and any ${\bold{d}}=(d_1,d_2)$, $d_1,d_2\in\{1,2\}$, we have $$\begin{aligned} \label{AppendixD_1} &I\left(W_1,W_2;{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right) =I\left(W_1^{(1)}\oplus W_2^{(1)},W_1^{(2)}\oplus W_2^{(2)},W_{d_1}^{(2)},W_{d_2}^{(1)},W_{d_1,s},W_{d_2,s};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\\ \label{AppendixD_2} &=I\left(M_{c\setminus K},M_{{\bold{d}}},W_{d_1,s},W_{d_2,s};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\\ \label{AppendixD_3} &=I\left(M_{c\setminus K};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)+I\left(M_{{\bold{d}}};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\|M_{c\setminus K}\right)+I\left(W_{d_1,s},W_{d_2,s};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\|M_{{\bold{d}}},M_{c\setminus K}\right).\end{aligned}$$ From (\[eq:example2\_2\]), (\[eq:example2\_3\]), and (\[eq:example2\_4\]), $M_c$ is independent from $\{M_{{\bold{d}}},\tilde{M}_{{\bold{d}}}\}$. The adversary’s observation in cache placement, ${\bold{Z}}_{S_1}^n$, results from sending $M_{c}=\left\{M_{c\setminus K}, K_1,K_2\right\}$, and its observation in the delivery results from sending $M_{{\bold{d}}}=\left\{W_{d_1}^{(2)},W_{d_2}^{(1)}\right\}$ and the randomization message $\tilde{M}_{{\bold{d}}}=\left\{W_{d_1,s}\oplus K_1, W_{d_2,s}\oplus W_2\right\}$. We now upper bound each term on the right hand side of (\[AppendixD\_3\]). For the third term, we have $$\begin{aligned} \label{AppendixD_4} I\Big(W_{d_1,s},W_{d_2,s};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\big\|&M_{{\bold{d}}},M_{c\setminus K}\Big)\leq I\left(W_{d_1,s},W_{d_2,s};\tilde{M}_{{\bold{d}}}\big\|M_{{\bold{d}}},M_{c\setminus K}\right)\\ \label{AppendixD_5} &=I\left(W_{d_1,s},W_{d_2,s};W_{d_1,s}\oplus K_1,W_{d_2,s}\oplus K_2\right)=0,\end{aligned}$$ where (\[AppendixD\_4\]) follows due to the Markov chain $\{W_{d_1,s},W_{d_2,s}\}-\{M_{c\setminus K},M_{{\bold{d}}},\tilde{M}_{{\bold{d}}}\}-\{{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\}$, and (\[AppendixD\_5\]) follows because $\tilde{M}_{{\bold{d}}}$ is independent from $\{W_{d_1,s},W_{d_2,s},M_{{\bold{d}}},M_{c\setminus K}\}$. For a fixed choice of the subsets $S_1$ and $S_2$, ${\bold{Z}}_{S_2}^n$ is independent from $\left\{M_c, {\bold{Z}}_{S_1}^n\right\}$. Thus, the first term on the right hand side of (\[AppendixD\_3\]) is bounded as $$\begin{aligned} \label{AppendixD_6} I&\left(M_{c\setminus K};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\leq I\left(M_{c};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\\ \label{AppendixD_7} &=I\left(M_{c};{\bold{Z}}_{S_1}^n\right)+I\left(M_{c};{\bold{Z}}_{S_2}^n\big\|{\bold{Z}}_{S_1}^n\right)= I\left(M_{c};{\bold{Z}}_{S_1}^n\right).\end{aligned}$$ For the second term on the right hand side of (\[AppendixD\_3\]), we have $$\begin{aligned} \label{AppendixD_9} I\left(M_{{\bold{d}}};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\big\|M_{c\setminus K}\right)&=I\left(M_{{\bold{d}}};{\bold{Z}}_{S_2}^n\big\|M_{c\setminus K}\right)+I\left(M_{{\bold{d}}};{\bold{Z}}_{S_1}^n\big\|M_{c\setminus K},{\bold{Z}}_{S_2}^n\right).\end{aligned}$$ Notice that $M_{c\setminus K}$ and ${\bold{Z}}_{S_2}^n$ are conditionally independent given $M_{{\bold{d}}}$. Thus, $$\begin{aligned} \label{AppendixD_10} I\left(M_{{\bold{d}}};{\bold{Z}}_{S_2}^n\big\|M_{c\setminus K}\right)=H\left({\bold{Z}}_{S_2}^n\big\|M_{c \setminus K}\right)-H\left({\bold{Z}}_{S_2}^n\big\|M_{{\bold{d}}}\right)\leq I\left(M_{{\bold{d}}};{\bold{Z}}_{S_2}^n\right).\end{aligned}$$ In addition, using the independence between $\left\{M_{{\bold{d}}},{\bold{Z}}_{S_2}^n\right\}$ and $\left\{M_c,{\bold{Z}}_{S_1}^n\right\}$, we have $$\begin{aligned} \label{AppendixD_11} I\left(M_{{\bold{d}}};{\bold{Z}}_{S_1}^n\big\|M_{c\setminus K},{\bold{Z}}_{S_2}^n\right)&=H\left({\bold{Z}}_{S_1}^n\big\|M_{c\setminus K},{\bold{Z}}_{S_2}^n\right)-H\left({\bold{Z}}_{S_1}^n\big\|M_{c\setminus K},M_{{\bold{d}}},{\bold{Z}}_{S_2}^n\right)\\ \label{AppendixD_12} &\leq H\left({\bold{Z}}_{S_1}^n\right)-H\left({\bold{Z}}_{S_1}^n\big\|M_{c\setminus K},K_1,K_2,M_{{\bold{d}}},{\bold{Z}}_{S_2}^n\right)\\ \label{AppendixD_13} &= H\left({\bold{Z}}_{S_1}^n\right)-H\left({\bold{Z}}_{S_1}^n\big\|M_{c}\right)=I\left(M_c;{\bold{Z}}_{S_1}^n\right).\end{aligned}$$ Substituting (\[AppendixD\_10\]) and (\[AppendixD\_13\]) in (\[AppendixD\_9\]) gives $$\begin{aligned} \label{AppendixD_14} I\left(M_{{\bold{d}}};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\big\|M_{c\setminus K}\right)\leq I\left(M_{{\bold{d}}};{\bold{Z}}_{S_2}^n\right)+I\left(M_c;{\bold{Z}}_{S_1}^n\right).\end{aligned}$$ Finally, substituting (\[AppendixD\_5\]), (\[AppendixD\_7\]), (\[AppendixD\_14\]) in (\[AppendixD\_3\]), and applying [@goldfeld2015semantic Theorem 2], we have $$\begin{aligned} \label{AppendixD_15} {{\underset{n\rightarrow\infty}\lim}}&\;\underset{\begin{subarray}{c} S_1,S_2\subseteq [1:n]:\\\|S_1\|+\|S_2\|=\mu\end{subarray}}\max I\left(W_1,W_2;{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)=0,\end{aligned}$$ since the rates of $M_c$ and $M_{{\bold{d}}}$ are $1-\alpha_1-\epsilon_n$ and $1-\alpha_2-\epsilon_n$, respectively. Secrecy Constraint for Setting $4$ {#AppendixE} ================================== For this setting, $M_c$ and $M_{{\bold{d}}}$ are defined in (\[eq:example4\_1\]) and (\[eq:example4\_3\]), and the randomization messages $\tilde{M}_c$ and $\tilde{M}_{{\bold{d}}}$ are defined in (\[eq:example4\_2\]) and (\[eq:example4\_4\]). Notice that, $M_c$ is independent from $\tilde{M}_c$; $M_{{\bold{d}}}$ is independent from $\tilde{M}_{{\bold{d}}}$, and $\{M_c,\tilde{M}_c\}$ are independent from $\{M_{{\bold{d}}},\tilde{M}_{{\bold{d}}}\}$. Conditioned on a fixed choice of the subsets $S_1$ and $S_2$, which satisfies the conditions for this setting, i.e., either $\left\{\|S_1\|=\mu,\; \|S_2\|=0\right\}$ or $\left\{\|S_1\|=0,\;\|S_2\|=\mu\right\}$, define the random variable $$\begin{aligned} \label{AppendixE_1} {\bold{Z}}_S^n\triangleq {\bold{Z}}_{S_1}^n\;\mathbbm{1}_{\left\{\|S_2\|=0\right\}}+{\bold{Z}}_{S_2}^n \;\mathbbm{1}_{\left\{\|S_1\|=0\right\}}.\end{aligned}$$ Notice that the random variable ${\bold{Z}}_S^n$ only have a well-defined probability distribution when conditioned on the event $\{S_1,S_2\}$, since a prior distribution on these subsets is not defined. For this fixed choice of the subsets, and any ${\bold{d}}=(d_1,d_2)$, $d_1,d_2\in\{1,2\}$, we have $$\begin{aligned} \label{AppendixE_3} I\big(W_1,&W_2;{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\big) =I\left(W_1^{(1)}\oplus W_2^{(1)},W_1^{(2)}\oplus W_2^{(2)},W_{d_1}^{(2)},W_{d_2}^{(1)},W_{d_1,s},W_{d_2,s};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\\ \label{AppendixE_4} &=I\left(M_c,M_{{\bold{d}}},W_{d_1,s},W_{d_2,s};{\bold{Z}}_S^n\right)\\ \label{AppendixE_5} \nonumber &=\mathbbm{1}_{\left\{\|S_2\|=0\right\}}\;I\left(M_c,M_{{\bold{d}}},W_{d_1,s},W_{d_2,s};{\bold{Z}}_S^n\Big\|\left\{\|S_2\|=0\right\}\right)\\ &\qquad\qquad\qquad +\mathbbm{1}_{\left\{\|S_1\|=0\right\}}\;I\left(M_c,M_{{\bold{d}}},W_{d_1,s},W_{d_2,s};{\bold{Z}}_S^n\Big\|\left\{\|S_1\|=0\right\}\right)\\ \label{AppendixE_6} &=\mathbbm{1}_{\left\{\|S_2\|=0\right\}}\;I\left(M_c,M_{{\bold{d}}},W_{d_1,s},W_{d_2,s};{\bold{Z}}_{S_1}^n\right)+\mathbbm{1}_{\left\{\|S_1\|=0\right\}}\;I\left(M_c,M_{{\bold{d}}},W_{d_1,s},W_{d_2,s};{\bold{Z}}_{S_2}^n\right)\\ \label{AppendixE_7} &=\mathbbm{1}_{\left\{\|S_2\|=0\right\}}\;I\left(M_c;{\bold{Z}}_{S_1}^n\right)+\mathbbm{1}_{\left\{\|S_1\|=0\right\}}\;I\left(M_{{\bold{d}}},W_{d_1,s},W_{d_2,s};{\bold{Z}}_{S_2}^n\right)\\ \label{AppendixE_8} &\leq \mathbbm{1}_{\left\{\|S_2\|=0\right\}}\;I\left(M_c;{\bold{Z}}_{S_1}^n\right)+\mathbbm{1}_{\left\{\|S_1\|=0\right\}}\;I\left(M_{{\bold{d}}};{\bold{Z}}_{S_2}^n\right).\end{aligned}$$ Equation (\[AppendixE\_7\]) follows because (i) ${\bold{Z}}_{S_1}^n$ results from $\{M_c,\tilde{M}_c\}$ which are independent from $\{M_{{\bold{d}}},W_{d_1,s},W_{d_2,s}\}$, and (ii) ${\bold{Z}}_{S_2}^n$ is conditionally independent from $M_c$ given $\{M_{{\bold{d}}},W_{d_1,s},W_{d_2,s}\}$, due to the Markov chain $M_c-\{M_{{\bold{d}}},W_{d_1,s},W_{d_2,s}\}-\{M_{{\bold{d}}},\tilde{M}_{{\bold{d}}}\}-{\bold{Z}}_{S_2}^n$. The inequality in (\[AppendixE\_8\]) follows using the same steps in (\[AppendixB\_1\])–(\[AppendixB\_6\]), in Appendix \[AppendixB\]. Finally, since $\tilde{M}_c$ is independent from $M_c$; $\tilde{M}_{{\bold{d}}}$ is independent from $M_{{\bold{d}}}$, and the rates of $M_c$ and $\tilde{M}_{{\bold{d}}}$ are both equal to $1-\alpha-\epsilon_n$, we have $$\begin{aligned} \label{AppendixE_9} &{{\underset{n\rightarrow\infty}\lim}}\;\underset{\begin{subarray}{c} S_1,S_2\subseteq [1:n]:\\\|S_1\|+\|S_2\|=\mu\end{subarray}}\max I\left(W_1,W_2;{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)= {{\underset{n\rightarrow\infty}\lim}}\;\underset{\begin{subarray}{c} S_1,S_2\subseteq [1:n]:\\\|S_i\|=0,\;\|S_j\|=\mu\\i,j\in\{1,2\},\;i\neq j\end{subarray}}\max I\left(W_1,W_2;{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\\ \label{AppendixE_10} &\leq {{\underset{n\rightarrow\infty}\lim}}\;\max\left\{\underset{S_1\subseteq[1:n]:\;\|S_1\|=\mu}\max\;I(M_c;{\bold{Z}}_{S_1}^n),\;\underset{S_2\subseteq[1:n]:\;\|S_2\|=\mu}\max\;I\left(M_{{\bold{d}}};{\bold{Z}}_{S_2}^n\right)\right\}\\ \label{AppendixE_11} &=\max\left\{{{\underset{n\rightarrow\infty}\lim}}\;\;\underset{S_1\subseteq[1:n]:\;\|S_1\|=\mu}\max\;I(M_c;{\bold{Z}}_{S_1}^n), {{\underset{n\rightarrow\infty}\lim}}\;\;\underset{S_2\subseteq[1:n]:\;\|S_2\|=\mu}\max\;I\left(M_{{\bold{d}}};{\bold{Z}}_{S_2}^n\right)\right\}\\ \label{AppendixE_12} &=0,\end{aligned}$$ where (\[AppendixE\_10\]) follows from (\[AppendixE\_8\]), and (\[AppendixE\_11\]) follows because both limits exist and equal to zero; by using [@goldfeld2015semantic Theorem 2]. Proofs for (\[eq:Proof\_1\_1\]) and (\[eq:Proof\_1\_2\]) {#AppendixF} ======================================================== Let us fix the subsets $S_1$ and $S_2$, and the messages $w_c$ and $w_{{\bold{d}}}$. Consider the Cartesian product of the random bins ${\mathcal{B}}_{w_c}$ and ${\mathcal{B}}_{w_{{\bold{d}}}}$, i.e., ${\mathcal{B}}_{w_c,w_{{\bold{d}}}}$, defined in (\[eq:partition\_B\_w\]). Recall that $P_{{\bold{Z}}_{S_1}^n{\bold{Z}}_{S_2}^n\big\|W_c=w_c,W_{{\bold{d}}}=w_{{\bold{d}}}}$ denotes the induced distribution at the adversary’s output when the transmitted codewords over cache placement and delivery phases are ${\bold{x}}_c^n\left(w_c,\tilde{w_c}\right)$ and ${\bold{x}}_{{\bold{d}}}^n\left(w_{{\bold{d}}},\tilde{w}_{{\bold{d}}}\right)$, i.e., when $\left({\bold{x}}_c^n,{\bold{x}}_{{\bold{d}}}^n\right)$ belongs to ${\mathcal{B}}_{w_c,w_{{\bold{d}}}}$. In addition, $P_{{\bold{Z}}_{S_1}^n{\bold{Z}}_{S_2}^n}$ denotes the output distribution at the adversary, induced by the cache placement and delivery codebooks, ${\mathcal{C}}_{c,n}$ and ${\mathcal{C}}_{{\bold{d}},n}$, defined in Figures \[fig:code\_construction\_1\] and \[fig:code\_construction\_2\]. Let ${\bold{z}}_1^n,{\bold{z}}_2^n\in{\mathcal{Z}}^n$, where ${\mathcal{Z}}\triangleq \{0,1\}\cup\{?\}$. Define the distribution $Q_{{\bold{Z}}_{S_1}^n{\bold{Z}}_{S_2}^n}$ as follows: $$\begin{aligned} \label{eq:AppendixF_1} Q_{{\bold{Z}}_{S_1}^n {\bold{Z}}_{S_2}^n}({\bold{z}}_1^n,{\bold{z}}_2^n)=\prod_{i\notin S_1,j\notin S_2}\mathbbm{1}\{z_{1,i}=?,z_{2,j}=?\}\prod_{i\in S_1,j\in S_2}U_X(z_{1,i})\;U_{X}(z_{2,i}),\end{aligned}$$ where $U_{X}(z)$ is a uniform binary distribution when $z=0,1,$ and $U_X(z)=0$ when $z=?$. We thus have $$\begin{aligned} \label{eq:AppendixF_2} &I\left(W_c,W_{{\bold{d}}};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)=\mathbb{D}\left(P_{W_c W_{{\bold{d}}}{\bold{Z}}_{S_1}^n{\bold{Z}}_{S_2^n}}\big\|\big\|P_{W_c W_{{\bold{d}}}}P_{{\bold{Z}}_{S_1}^n{\bold{Z}}_{S_2^n}}\right)\\ \label{eq:AppendixF_3} &=\sum_{w_c,w_{{\bold{d}}}}P_{W_c W_{{\bold{d}}}}(w_c,w_{{\bold{d}}})\sum_{{\bold{z}}_1^n,{\bold{z}}_2^n} P_{{\bold{Z}}_{S_1}^n {\bold{Z}}_{S_2}^n\big\|W_c W_{{\bold{d}}}} ({\bold{z}}_1^n,{\bold{z}}_2^n\|w_c,w_{{\bold{d}}})\log \left(\frac{P_{W_c W_{{\bold{d}}}{\bold{Z}}_{S_1}^n {\bold{Z}}_{S_2}^n} (w_c,w_{{\bold{d}}},{\bold{z}}_1^n,{\bold{z}}_2^n)}{P_{{\bold{Z}}_{S_1}^n {\bold{Z}}_{S_2}^n} ({\bold{z}}_1^n,{\bold{z}}_2^n)P_{W_c W_{{\bold{d}}}}(w_c,w_{{\bold{d}}})}\right)\\ \label{eq:AppendixF_4} \nonumber &=\sum_{w_c,w_{{\bold{d}}}}P_{W_c W_{{\bold{d}}}}(w_c,w_{{\bold{d}}})\sum_{{\bold{z}}_1^n,{\bold{z}}_2^n} P_{{\bold{Z}}_{S_1}^n {\bold{Z}}_{S_2}^n\big\|W_c W_{{\bold{d}}}} ({\bold{z}}_1^n,{\bold{z}}_2^n\|w_c,w_{{\bold{d}}})\\ &\qquad\qquad\qquad \qquad\qquad \times \log\left( \frac{P_{{\bold{Z}}_{S_1}^n {\bold{Z}}_{S_2}^n\big\|W_c W_{{\bold{d}}}} ({\bold{z}}_1^n,{\bold{z}}_2^n\|w_c,w_{{\bold{d}}})}{Q_{{\bold{Z}}_{S_1}^n{\bold{Z}}_{S_2}^n}({\bold{z}}_1^n,{\bold{z}}_2^n)}\times \frac{Q_{{\bold{Z}}_{S_1}^n{\bold{Z}}_{S_2}^n}({\bold{z}}_1^n,{\bold{z}}_2^n)}{P_{{\bold{Z}}_{S_1}^n {\bold{Z}}_{S_2}^n}({\bold{z}}_1^n,{\bold{z}}_2^n)}\right)\\ \label{eq:AppendixF_5} &=\sum_{w_c,w_{{\bold{d}}}}P_{W_c W_{{\bold{d}}}}(w_c,w_{{\bold{d}}})\left[\mathbb{D}\left(P_{{\bold{Z}}_{S_1}^n {\bold{Z}}_{S_2}^n \big\|W_c=w_c,W_{{\bold{d}}}=w_{{\bold{d}}}}\big\|\big\|Q_{{\bold{Z}}_{S_1}^n{\bold{Z}}_{S_2}^n}\right)-\mathbb{D}\left(P_{{\bold{Z}}_{S_1}^n {\bold{Z}}_{S_2}^n }\big\|\big\|Q_{{\bold{Z}}_{S_1}^n{\bold{Z}}_{S_2}^n}\right)\right]\\ \label{eq:AppendixF_6} &\leq \sum_{w_c,w_{{\bold{d}}}}P_{W_c W_{{\bold{d}}}}(w_c,w_{{\bold{d}}})\;\mathbb{D}\left(P_{{\bold{Z}}_{S_1}^n {\bold{Z}}_{S_2}^n\big\|W_c=w_c,W_{{\bold{d}}}=w_{{\bold{d}}}}\big\|\big\|Q_{{\bold{Z}}_{S_1}^n{\bold{Z}}_{S_2}^n}\right).\end{aligned}$$ Define ${\bold{Z}}^{S_1}\triangleq\{Z_{S_1,i}: i\in S_1\}$, ${\bold{Z}}^{S_2}\triangleq\{Z_{S_2,i}: i\in S_2\}$, ${\bold{Z}}^{S_1^c}\triangleq\{Z_{S_1,i}: i\notin S_1\}$, ${\bold{Z}}^{S_2^c}\triangleq\{Z_{S_2,i}: i\notin S_2\}$, and let ${\bold{z}}^{S_1}$, ${\bold{z}}^{S_2}$, ${\bold{z}}^{S_1^c}$, ${\bold{z}}^{S_2^c}$ be the corresponding realizations. Note that ${\bold{Z}}_{S_1}^n=\{{\bold{Z}}^{S_1},{\bold{Z}}^{S_1^c}\}$ and ${\bold{Z}}_{S_2}^n=\{{\bold{Z}}^{S_2},{\bold{Z}}^{S_2^c}\}$. For each $S_1$, $S_2$, $w_c$, and $w_{{\bold{d}}}$, we have $$\begin{aligned} \label{eq:AppendixF_7} &\mathbb{D}\left(P_{{\bold{Z}}_{S_1}^n {\bold{Z}}_{S_2}^n \big\|W_c=w_c,W_{{\bold{d}}}=w_{{\bold{d}}}}\big\|\big\|Q_{{\bold{Z}}_{S_1}^n{\bold{Z}}_{S_2}^n}\right)=\mathbb{D}\left(P_{{\bold{Z}}^{S_1}{\bold{Z}}^{S_1^c}{\bold{Z}}^{S_2}{\bold{Z}}^{S_2^c}\big\|W_c=w_c,W_{{\bold{d}}}=w_{{\bold{d}}}}\big\|\big\|Q_{{\bold{Z}}^{S_1}{\bold{Z}}^{S_1^c}{\bold{Z}}^{S_2}{\bold{Z}}^{S_2^c}}\right)\\ \label{eq:AppendixF_8} \nonumber &=\sum_{{\bold{z}}^{S_1},{\bold{z}}^{S_1^c},{\bold{z}}^{S_2},{\bold{z}}^{S_2^c}}P_{{\bold{Z}}^{S_1}{\bold{Z}}^{S_1^c}{\bold{Z}}^{S_2}{\bold{Z}}^{S_2^c}\big\|W_c W_{{\bold{d}}}}\left({\bold{z}}^{S_1},{\bold{z}}^{S_1^c},{\bold{z}}^{S_2},{\bold{z}}^{S_2^c}\big\|w_c,w_{{\bold{d}}}\right)\\ &\qquad\qquad\qquad\qquad \times \log\left(\frac{P_{{\bold{Z}}^{S_1}{\bold{Z}}^{S_1^c}{\bold{Z}}^{S_2}{\bold{Z}}^{S_2^c}\big\|W_c W_{{\bold{d}}}}\left({\bold{z}}^{S_1},{\bold{z}}^{S_1^c},{\bold{z}}^{S_2},{\bold{z}}^{S_2^c}\|w_c,w_{{\bold{d}}}\right)}{Q_{{\bold{Z}}^{S_1}{\bold{Z}}^{S_1^c}{\bold{Z}}^{S_2}{\bold{Z}}^{S_2^c}}({\bold{z}}^{S_1},{\bold{z}}^{S_1^c},{\bold{z}}^{S_2},{\bold{z}}^{S_2^c})}\right)\\ \label{eq:AppendixF_9} \nonumber &=\sum_{{\bold{z}}^{S_1},{\bold{z}}^{S_1^c},{\bold{z}}^{S_2},{\bold{z}}^{S_2^c}} P_{{\bold{Z}}^{S_1}{\bold{Z}}^{S_2}\big\|W_c W_{{\bold{d}}}}\left({\bold{z}}^{S_1},{\bold{z}}^{S_2}\big\|w_c,w_{{\bold{d}}}\right)P_{{\bold{Z}}^{S_1^c}{\bold{Z}}^{S_2^c}\big\|W_c W_{{\bold{d}}}{\bold{Z}}^{S_1}{\bold{Z}}^{S_2}}\left({\bold{z}}^{S_1^c},{\bold{z}}^{S_2^c}\big\|w_c,w_{{\bold{d}}},{\bold{z}}^{S_1},{\bold{z}}^{S_2}\right)\\ &\qquad \times \log\left(\frac{P_{{\bold{Z}}^{S_1}{\bold{Z}}^{S_2}\big\|W_c W_{{\bold{d}}}}\left({\bold{z}}^{S_1},{\bold{z}}^{S_2}\big\|w_c,w_{{\bold{d}}}\right)P_{{\bold{Z}}^{S_1^c}{\bold{Z}}^{S_2^c}\big\|W_c W_{{\bold{d}}}{\bold{Z}}^{S_1}{\bold{Z}}^{S_2}}\left({\bold{z}}^{S_1^c},{\bold{z}}^{S_2^c}\big\|w_c,w_{{\bold{d}}},{\bold{z}}^{S_1},{\bold{z}}^{S_2}\right)}{Q_{{\bold{Z}}^{S_1}{\bold{Z}}^{S_2}}\left({\bold{z}}^{S_1},{\bold{z}}^{S_2}\right)Q_{{\bold{Z}}^{S_1^c}{\bold{Z}}^{S_2^c}\big\|{\bold{Z}}^{S_1}{\bold{Z}}^{S_2}}\left({\bold{z}}^{S_1^c},{\bold{z}}^{S_2^c}\big\|{\bold{z}}^{S_1},{\bold{z}}^{S_2}\right)}\right)\\ \label{eq:AppendixF_10} \nonumber &= \mathbb{D}\left(P_{{\bold{Z}}^{S_1}{\bold{Z}}^{S_2}\big\|W_c=w_c,W_{{\bold{d}}}=w_{{\bold{d}}}}\big\|\big\|Q_{{\bold{Z}}^{S_1}{\bold{Z}}^{S_2}}\right)+\sum_{{\bold{z}}^{S_1},{\bold{z}}^{S_2}} P_{{\bold{Z}}^{S_1}{\bold{Z}}^{S_2}\big\|W_c=w_c, W_{{\bold{d}}}=w_{{\bold{d}}}}\left({\bold{z}}^{S_1},{\bold{z}}^{S_2}\right)\\ &\qquad\qquad\qquad\times \mathbb{D}\left(P_{{\bold{Z}}^{S_1^c}{\bold{Z}}^{S_2^c}\big\|W_c=w_c, W_{{\bold{d}}}=w_{{\bold{d}}},{\bold{Z}}^{S_1}={\bold{z}}^{S_1},{\bold{Z}}^{S_2}={\bold{z}}^{S_2}}\big\|\big\|Q_{{\bold{Z}}^{S_1^c}{\bold{Z}}^{S_2^c}\|{\bold{Z}}^{S_1}={\bold{z}}^{S_1},{\bold{Z}}^{S_2}={\bold{z}}^{S_2}}\right)\\ \label{eq:AppendixF_11} &=\mathbb{D}\left(P_{{\bold{Z}}^{S_1}{\bold{Z}}^{S_2}\big\|W_c=w_c,W_{{\bold{d}}}=w_{{\bold{d}}}}\big\|\big\|Q_{{\bold{Z}}^{S_1}{\bold{Z}}^{S_2}}\right),\end{aligned}$$ where (\[eq:AppendixF\_11\]) follows because $$\begin{aligned} \label{eq:AppendixF_12} \nonumber P&_{{\bold{Z}}^{S_1^c}{\bold{Z}}^{S_2^c}\big\|W_c=w_c, W_{{\bold{d}}}=w_{{\bold{d}}},{\bold{Z}}^{S_1}={\bold{z}}^{S_1},{\bold{Z}}^{S_2}={\bold{z}}^{S_2}}=Q_{{\bold{Z}}^{S_1^c}{\bold{Z}}^{S_2^c}\big\|{\bold{Z}}^{S_1}={\bold{z}}^{S_1},{\bold{Z}}^{S_2}={\bold{z}}^{S_2}}\\ &=\prod_{i\notin S_1,j\notin S_2} \mathbbm{1}\{z_{1,i}=?, z_{2,j}=?\}.\end{aligned}$$ By applying the stronger version of Wyner’s soft covering lemma in [@goldfeld2015semantic Lemma 1] to (\[eq:AppendixF\_11\]), for every $\epsilon>0$, there exists a $\gamma(\epsilon)>0$ such that $$\begin{aligned} \label{eq:AppendixF_13} \nonumber \mathbb{P}_{{\mathcal{B}}_{w_c,w_{{\bold{d}}}}}&\left(\mathbb{D}\left(P_{{\bold{Z}}_{S_1}^n {\bold{Z}}_{S_2}^n \big\|W_c=w_c,W_{{\bold{d}}}=w_{{\bold{d}}}}\big\|\big\|Q_{{\bold{Z}}_{S_1}^n{\bold{Z}}_{S_2}^n}\right)>\epsilon\right)\\ &=\mathbb{P}_{{\mathcal{B}}_{w_c,w_{{\bold{d}}}}}\left(\mathbb{D}\left(P_{{\bold{Z}}^{S_1}{\bold{Z}}^{S_2}\big\|W_c=w_c,W_{{\bold{d}}}=w_{{\bold{d}}}}\big\|\big\|Q_{{\bold{Z}}^{S_1}{\bold{Z}}^{S_2}}\right)>\epsilon\right)\leq \exp\left(-e^{n\gamma(\epsilon)}\right),\end{aligned}$$ since the rate of ${\mathcal{B}}_{w_c,w_{{\bold{d}}}}$ is slightly greater than $\alpha$, i.e., ${\mathcal{B}}_{w_c,w_{{\bold{d}}}}$ contains $2^{n\alpha_{\epsilon}}$ codewords. Using (\[eq:AppendixF\_6\]) and (\[eq:AppendixF\_11\]), we have $$\begin{aligned} \label{eq:AppendixF_14} I\left(W_c,W_{{\bold{d}}};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\leq \sum_{w_c,w_{{\bold{d}}}}P_{W_c W_{{\bold{d}}}}(w_c,w_{{\bold{d}}})\;\mathbb{D}\left(P_{{\bold{Z}}^{S_1}{\bold{Z}}^{S_2}\big\|W_c=w_c,W_{{\bold{d}}}=w_{{\bold{d}}}}\big\|\big\|Q_{{\bold{Z}}^{S_1}{\bold{Z}}^{S_2}}\right).\end{aligned}$$ Thus, $$\begin{aligned} \label{eq:AppendixF_15} \nonumber &\mathbb{P}_{{\mathcal{B}}_{w_c,w_{{\bold{d}}}}}\left(\underset{\begin{subarray}{c} S_1,S_2\subseteq [1:n]:\\\|S_1\|+\|S_2\|=\mu\end{subarray}}\max\;I\left(W_c,W_{{\bold{d}}};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)\geq \epsilon\right)\\ &\leq \mathbb{P}_{{\mathcal{B}}_{w_c,w_{{\bold{d}}}}}\left(\max_{S_1,S_2}\;\sum_{w_c,w_{{\bold{d}}}}P_{W_c W_{{\bold{d}}}}(w_c,w_{{\bold{d}}})\;\mathbb{D}\left(P_{{\bold{Z}}^{S_1}{\bold{Z}}^{S_2}\big\|W_c=w_c,W_{{\bold{d}}}=w_{{\bold{d}}}}\big\|\big\|Q_{{\bold{Z}}^{S_1}{\bold{Z}}^{S_2}}\right)>\epsilon\right)\\ \label{eq:AppendixF_16} &\leq \mathbb{P}_{{\mathcal{B}}_{w_c,w_{{\bold{d}}}}}\left(\underset{\begin{subarray}{c} w_c,w_{{\bold{d}}}\\S_1,S_2\end{subarray}}\max\;\mathbb{D}\left(P_{{\bold{Z}}^{S_1}{\bold{Z}}^{S_2}\big\|W_c=w_c,W_{{\bold{d}}}=w_{{\bold{d}}}}\big\|\big\|Q_{{\bold{Z}}^{S_1}{\bold{Z}}^{S_2}}\right)>\epsilon\right)\\ \label{eq:AppendixF_17} &= \mathbb{P}_{{\mathcal{B}}_{w_c,w_{{\bold{d}}}}}\left(\underset{\begin{subarray}{c} w_c,w_{{\bold{d}}}\\S_1,S_2\end{subarray}}\bigcup\;\mathbb{D}\left(P_{{\bold{Z}}^{S_1}{\bold{Z}}^{S_2}\big\|W_c=w_c,W_{{\bold{d}}}=w_{{\bold{d}}}}\big\|\big\|Q_{{\bold{Z}}^{S_1}{\bold{Z}}^{S_2}}\right)>\epsilon\right)\\ \label{eq:AppendixF_18} &\leq \underset{\begin{subarray}{c} w_c,w_{{\bold{d}}}\\S_1,S_2\end{subarray}}\sum\;\mathbb{P}_{{\mathcal{B}}_{w_c,w_{{\bold{d}}}}}\left(\mathbb{D}\left(P_{{\bold{Z}}^{S_1}{\bold{Z}}^{S_2}\big\|W_c=w_c,W_{{\bold{d}}}=w_{{\bold{d}}}}\big\|\big\|Q_{{\bold{Z}}^{S_1}{\bold{Z}}^{S_2}}\right)>\epsilon\right)\end{aligned}$$ where (\[eq:AppendixF\_15\]) follows from (\[eq:AppendixF\_14\]), and (\[eq:AppendixF\_18\]) follows from the union bound. Since the combined number of the messages $w_c,w_{{\bold{d}}},$ and the subsets $S_1,S_2$ is at most exponential in $n$, using the super-exponential decay rate in (\[eq:AppendixF\_13\]), the probability term on the right hand side of (\[eq:AppendixF\_15\]) goes to zero as $n$ goes to infinity. Thus, $\max_{S_1,S_2}I\left(W_c,W_{{\bold{d}}};{\bold{Z}}_{S_1}^n,{\bold{Z}}_{S_2}^n\right)$ converges to zero almost surely. This completes the proof of (\[eq:Proof\_1\_2\]). [^1]: Large block-length $n$ is needed in order to ensure a valid subpacketization of the file $W_l$ into the sub-files $\{W_l^{(1)},W_l^{(2)},W_{l,s}\}$, for $l=1,2$. That is, a bijective map between the file and its sub-files is preserved.	{ "pile_set_name": "ArXiv" }
--- abstract: \| A semi-simple tensor extension of the Poincaré algebra is proposed for the arbitrary dimensions $D$. A supersymmetric also semi-simple generalization of this extension is constructed in the $D=4$ dimensions. This paper is dedicated to the memory of Anna Yakovlevna Gelyukh. [PACS:]{} 02.20.Sv; 11.30.Cp; 11.30.Pb [Keywords:]{} Poincaré algebra, Tensor, Extension, Casimir operators, Supersymmetry --- \ [Kharkov Institute of Physics and Technology,\ 1, Akademicheskaya St., 61108 Kharkov, Ukraine]{}\ 0 Introduction ============ In the papers [@gios1; @gios2; @cj; @ss1; @dss0; @dss] the Poincaré algebra for the generators of the rotations $M_{ab}$ and translations $P_a$ in $D$ dimensions $$\begin{aligned} [M_{ab},M_{cd}]=(g_{ad}M_{bc}+g_{bc}M_{ad})-(c\leftrightarrow d),\nonumber\end{aligned}$$ $$\begin{aligned} [M_{ab},P_c]=g_{bc}P_a-g_{ac}P_b,\nonumber\end{aligned}$$ $$\begin{aligned} \label{1.1} [P_a,P_b]=0\end{aligned}$$ has been extended by means of the second rank tensor generator $Z_{ab}$ in the following way: $$\begin{aligned} [M_{ab},M_{cd}]=(g_{ad}M_{bc}+g_{bc}M_{ad})-(c\leftrightarrow d),\nonumber\end{aligned}$$ $$\begin{aligned} [M_{ab},P_c]=g_{bc}P_a-g_{ac}P_b,\nonumber\end{aligned}$$ $$\begin{aligned} [P_a,P_b]=cZ_{ab},\nonumber\end{aligned}$$ $$\begin{aligned} [M_{ab},Z_{cd}]=(g_{ad}Z_{bc}+g_{bc}Z_{ad})-(c\leftrightarrow d),\nonumber\end{aligned}$$ $$\begin{aligned} [P_a,Z_{bc}]=0,\nonumber\end{aligned}$$ $$\begin{aligned} \label{1.2} [Z_{ab},Z_{cd}]=0,\end{aligned}$$ where $c$ is some constant[^1]. Such an extension makes common sense, since it is homomorphic to the usual Poincaré algebra (\[1.1\]). Moreover, in the limit ${c\to 0}$ the algebra (\[1.2\]) goes to the semi-direct sum of the commutative ideal $Z_{ab}$ and Poincaré algebra (\[1.1\]). It is remarkable enough that the momentum square Casimir operator of the Poincaré algebra under this extension ceases to be the Casimir operator and it is generalized by adding the term linearly dependent on the angular momentum $$\begin{aligned} \label{1.3} P^aP_a+cZ^{ab}M_{ba}\mathrel{\mathop=^{\rm def}}X_kh^{kl}X_l,\end{aligned}$$ where $X_k=\{P_a, Z_{ab}, M_{ab}\}$. Due to this fact, an irreducible representation of the extended algebra (\[1.2\]) has to contain the fields of the different masses [@ss1; @ss2]. This extension with non-commuting momenta has also something in common with the ideas of the papers [@sn; @ya; @hl] and with the non-commutative geometry idea [@c]. It is interesting to note that in spite of the fact that the algebra (\[1.2\]) is not semi-simple and therefore has a degenerate Cartan-Killing metric tensor nevertheless there exists another non-degenerate invariant tensor $h_{kl}$ in adjoint representation which corresponds to the quadratic Casimir operator (\[1.3\]), where the matrix $h^{kl}$ is inverse to the matrix $h_{kl}$: $h^{kl}h_{lm}=\d_m^k$. There are other quadratic Casimir operators \[1.4\] c\^2Z\^[ab]{}Z\_[ab]{}, \[1.5\] c\^2\^[abcd]{}Z\_[ab]{}Z\_[cd]{}. Note that the Casimir operator (\[1.5\]), dependent on the Levi-Civita tensor $\e^{abcd}$, is suitable only for the $D=4$ dimensions. It has also been shown that for the dimensions $D=2,3,4$ the extended Poincaré algebra (\[1.2\]) allows the following supersymmetric generalization: $$\begin{aligned} \{Q_\k,Q_\l\}=-d(\sigma^{ab}C)_{\k\l}Z_{ab},\nonumber\end{aligned}$$ $$\begin{aligned} [M_{ab},Q_\k]=-(\sigma_{ab}Q)_\k,\nonumber\end{aligned}$$ $$\begin{aligned} [P_a,Q_\k]=0,\nonumber\end{aligned}$$ $$\begin{aligned} \label{1.6} [Z_{ab},Q_\k]=0\end{aligned}$$ with the help of the super-translation generators $Q_\k$. In (\[1.6\]) $C$ is a charge conjugation matrix, $d$ is some constant and $\sigma_{ab}={1\over4}[\gamma_a,\gamma_b]$, where $\g_a$ is the Dirac matrix. Under this supersymmetric generalization the quadratic Casimir operator (\[1.3\]) is modified into the following form: $$\begin{aligned} \label{1.7} P^aP_a+cZ^{ab}M_{ba}-{c\over2d}Q_\k(C^{-1})^{\k\l}Q_\l,\end{aligned}$$ while the form of the rest quadratic Casimir operators (\[1.4\]), (\[1.5\]) remains unchanged. In the present paper we propose another possible semi-simple tensor extension of the Poincaré algebra (\[1.1\]) and for the case $D=4$ dimensions we give a supersymmetric generalization of this extension. In the limit this supersymmetrically generalized extension go to the Lie superalgebra (\[1.2\]), (\[1.6\]). Semi-simple tensor extension ============================ Let us extend the Poincaré algebra (\[1.1\]) in the $D$ dimensions by means of the tensor generator $Z_{ab}$ in the following way: $$\begin{aligned} [M_{ab},M_{cd}]=(g_{ad}M_{bc}+g_{bc}M_{ad})-(c\leftrightarrow d),\nonumber\end{aligned}$$ $$\begin{aligned} [M_{ab},P_c]=g_{bc}P_a-g_{ac}P_b,\nonumber\end{aligned}$$ $$\begin{aligned} [P_a,P_b]=cZ_{ab},\nonumber\end{aligned}$$ $$\begin{aligned} [M_{ab},Z_{cd}]=(g_{ad}Z_{bc}+g_{bc}Z_{ad})-(c\leftrightarrow d),\nonumber\end{aligned}$$ $$\begin{aligned} [Z_{ab},P_c]={4a^2\over c}(g_{bc}P_a-g_{ac}P_b),\nonumber\end{aligned}$$ $$\begin{aligned} \label{2.1} [Z_{ab},Z_{cd}]={4a^2\over c}[(g_{ad}Z_{bc}+g_{bc}Z_{ad}) -(c\leftrightarrow d)],\end{aligned}$$ where $a$ and $c$ are some constants. This Lie algebra, when the quantities $P_a$ and $Z_{ab}$ are taken as the generators of a homomorphism kernel, is homomorphic to the usual Lorentz algebra. It is remarkable that the Lie algebra (\[2.1\]) is [semi-simple]{} in contrast to the Poincaré algebra (\[1.1\]) and extended Poincaré algebra (\[1.2\]). The extended Lie algebra (\[2.1\]) has the following quadratic Casimir operators: $$\begin{aligned} \label{2.2} C_1=P^aP_a+cZ^{ab}M_{ba}+2a^2M^{ab}M_{ab}\mathrel{\mathop=^{\rm def}} X_kH_1^{kl}X_l,\end{aligned}$$ \[2.3\] C\_2=c\^2Z\^[ab]{}Z\_[ab]{}+8a\^2(cZ\^[ab]{}M\_[ba]{}+2a\^2M\^[ab]{}M\_[ab]{}) X\_kH\_2\^[kl]{}X\_l, \[2.4\] C\_3=\^[abcd]{}\[c\^2Z\_[ab]{}Z\_[cd]{}+8a\^2(cZ\_[ba]{}M\_[cd]{}+2a\^2M\_[ab]{}M\_[cd]{})\]. Note that in the limit $a\to0$ the algebra (\[2.1\]) tend to the algebra (\[1.2\]) and the quadratic Casimir operators (\[2.2\]), (\[2.3\]) and (\[2.4\]) are turned into (\[1.3\]), (\[1.4\]) and (\[1.5\]), respectively. The symmetric tensor \[2.5\] H\^[kl]{}=sH\_1\^[kl]{}+tH\_2\^[kl]{}=H\^[lk]{} with arbitrary constants $s$ and $t$ is invariant with respect to the adjoint representation H\^[kl]{}=H\^[mn]{}[U\_m]{}\^k[U\_n]{}\^l.Conversely, if we demand the invariance with respect to the adjoint representation of the second rank contravariant symmetric tensor, then we come to the structure (\[2.5\]) (see also the relation (32) in [@dss]). The semi-simple algebra (\[2.1\]) =[f\_[kl]{}]{}\^mX\_mhas the non-degenerate Cartan-Killing metric tensor g\_[kl]{}=[f\_[km]{}]{}\^n[f\_[ln]{}]{}\^m,which is invariant with respect to the co-adjoint representation g\_[kl]{}=[U\_k]{}\^m[U\_l]{}\^ng\_[mn]{}.With the help of the inverse metric tensor $g^{kl}$: $g^{kl}g_{lm}=\d_m^k$ we can construct the quadratic Casimir operator which, as it turned out, has the following expression in terms of the quadratic Casimir operators (\[2.2\]) and (\[2.3\]): \[2.6\] X\_kg\^[kl]{}X\_l=[18a\^2(D-1)]{}, that corresponds to the particular choice of the constants $s$ and $t$ in (\[2.5\]). The extended Poincaré algebra (\[2.1\]) can be rewritten in the form: \[2.7\] \[N\_[ab]{},N\_[cd]{}\]=(g\_[ad]{}N\_[bc]{}+g\_[bc]{}N\_[ad]{})-(cd), \[2.8\] \[L\_[AB]{},L\_[CD]{}\]=(g\_[AD]{}L\_[BC]{}+g\_[BC]{}L\_[AD]{})-(CD), \[2.9\] \[N\_[ab]{},L\_[CD]{}\]=0, where the metric tensor $g_{AB}$ has the following nonzero components: g\_[AB]{}={g\_[ab]{}, g\_[D+1D+1]{}=-1}. The generators \[2.10\] N\_[ab]{}=M\_[ab]{}-[c4a\^2]{}Z\_[ab]{} form the Lorentz algebra $so(D-1, 1)$ and the generators \[2.11\] L\_[AB]{}={L\_[ab]{}=[c4a\^2]{}Z\_[ab]{}, L\_[aD+1]{}=-L\_[D+1a]{}=[12a]{}P\_a, L\_[D+1D+1]{}=0} form the algebra $so(D-1,2)$[^2]. The algebra (\[2.7\])-(\[2.9\]) is a direct sum $so(D-1,1)\oplus so(D-1,2)$. The quadratic Casimir operators $N_{ab}N^{ab}$, $L_{AB}L^{AB}$ and $\e^{abcd}N_{ab}N_{cd}$ of the algebra (\[2.7\])-(\[2.9\]) are expressed in terms of the operators $C_1$ (\[2.2\]), $C_2$ (\[2.3\]) and $C_3$ (\[2.4\]) in the following way: \[2.12\] N\_[ab]{}N\^[ab]{}-L\_[AB]{}L\^[AB]{}=[12a\^2]{}C\_1, \[2.13\] N\_[ab]{}N\^[ab]{}=[116a\^4]{}C\_2, \[2.14\] \^[abcd]{}N\_[ab]{}N\_[cd]{}=[116a\^4]{}C\_3. Supersymmetric generalization ============================= In the case $D=4$ dimensions the extended Poincaré algebra (\[2.1\]) admits the following supersymmetric generalization: $$\begin{aligned} \{Q_\k,Q_\l\}=-d\left[{2a\over c}(\g^aC)_{\k\l}P_a +(\sigma^{ab}C)_{\k\l}Z_{ab}\right],\nonumber\end{aligned}$$ $$\begin{aligned} [M_{ab},Q_\k]=-(\sigma_{ab}Q)_\k,\nonumber\end{aligned}$$ $$\begin{aligned} [P_a,Q_\k]=a(\g_aQ)_\k,\nonumber\end{aligned}$$ $$\begin{aligned} \label{3.1} [Z_{ab},Q_\k]=-{4a^2\over c}(\s_{ab}Q)_\k,\end{aligned}$$ where $Q_\k$ are the super-translation generators. Under such a generalization the Casimir operator (\[2.2\]) is modified by adding a term quadratic in the super-translation generators $$\begin{aligned} \label{3.2} {\tilde C}_1=P^aP_a+cZ^{ab}M_{ba}+2a^2M^{ab}M_{ab} &-&{c\over2d}Q_\k(C^{-1})^{\k\l}Q_\l\cr{\nonumber\\}&\stackrel{\rm def}{=}& X_KH_1^{KL}X_L,\end{aligned}$$ whereas the form of the rest quadratic Casimir operators (\[2.3\]) and (\[2.4\]) is not changed. In (\[3.2\]) $X_K=\{P_a, Z_{ab}, M_{ab}, Q_\k\}$ is a set of the generators for the also semi-simple extended superalgebra (\[2.1\]), (\[3.1\]). The tensor \[3.3\] H\^[KL]{}=vH\_1\^[KL]{}+wH\_2\^[KL]{}=(-1)\^[p\_Kp\_L+p\_K+p\_L]{}H\^[LK]{} is invariant with respect to the adjoint representation H\^[KL]{}=(-1)\^[(p\_K+p\_M)(p\_L+1)]{}H\^[MN]{}[U\_N]{}\^L[U\_M]{}\^K,where $p_K=p(K)$ is a Grassmann parity of the quantity $K$. In (\[3.3\]) $v$ and $w$ are arbitrary constants and nonzero elements of the matrix $H_2^{KL}$ equal to the elements of the matrix $H_2^{kl}$ followed from (\[2.3\]). Again, by demanding the invariance with respect to the adjoint representation of the second rank contravariant tensor $H^{KL}=(-1)^{p_Kp_L+p_K+p_L}H^{LK}$, we come to the structure (\[3.3\]) (see also the relation (32) in [@dss]). The semi-simple Lie superalgebra (\[2.1\]) (\[3.1\]) has the non-degenerate Cartan-Killing metric tensor $G_{KL}$ (see the relation (A.5) in the Appendix A) which is invariant with respect to the co-adjoint representation G\_[KL]{}=(-1)\^[p\_K(p\_L+p\_N)]{}[U\_L]{}\^N[U\_K]{}\^MG\_[MN]{}.With the use of the inverse metric tensor $G^{KL}$ G\^[KL]{}G\_[LM]{}=\_M\^Kwe can construct the quadratic Casimir operator (see the relation (A.8) in the Appendix A) which takes the following expression in terms of the Casimir operators (\[2.3\]) and (\[3.2\]): \[3.4\] X\_KG\^[KL]{}X\_L=[120a\^2]{}([C]{}\_1-[932a\^2]{}C\_2), that meets the particular choice of the constants $v$ and $w$ in (\[3.3\]). In the $D=4$ case the extended superalgebra (\[2.1\]), (\[3.1\]) can be rewritten in the form of the relations (\[2.7\])-(\[2.9\]) and the following ones: \[3.5\] {Q\_,Q\_ł}=-[4a\^2dc]{}(§\^[AB]{}C)\_[ł]{}L\_[AB]{}, \[3.6\] \[L\_[AB]{},Q\_\]=-(§\_[AB]{}Q)\_, \[3.7\] \[N\_[ab]{},Q\_\]=0, where §\_[AB]{}=[14]{}\[\_A,\_B\],\_A={\_a\_5,\_5},{\_a,\_b}=2g\_[ab]{},g\_[ab]{}=diag(-1,1,1,1),\_5=\_0\_1\_2\_3.The generators $N_{ab}$ (\[2.10\]) form the Lorentz algebra $so(3,1)$ and the generators $L_{AB}$ (\[2.11\]), $Q_\k$ form the orthosymplectic algebra $osp(1,4)$. We see that superalgebra (\[2.7\])-(\[2.9\]), (\[3.5\])- (\[3.7\]) is a direct sum $so(3,1)\oplus osp(1,4)$. In this case the Casimir operator (\[2.12\]) is modified by adding a term quadratic in the super-translation generators N\_[ab]{}N\^[ab]{}-L\_[AB]{}L\^[AB]{}-[c4a\^2d]{}Q\_(C\^[-1]{})\^[ł]{}Q\_ł= [12a\^2]{}[C]{}\_1,while the form of the quadratic Casimir operators (\[2.13\]) and (\[2.14\]) is not changed. Conclusion ========== Thus, we proposed the semi-simple second rank tensor extension of the Poincaré algebra in the arbitrary dimensions $D$ and super Poincaré algebra in the $D=4$ dimensions. It is very important, since under construction of the models it is more convenient to deal with the non-degenerate space-time symmetry. We also constructed the quadratic Casimir operators for the semi-simple extended Poincaré and super Poincaré algebras. It is interesting to develop the models based on these extended algebras. The work in this direction is in progress. Acknowledgments {#acknowledgments .unnumbered} =============== We are grateful to J.A. de Azcarraga for the valuable remark. One of the authors (V.A.S.) thanks the administration of the Office of Associate and Federation Schemes of the Abdus Salam ICTP for the kind hospitality at Trieste where this work has been completed. The research of V.A.S. was partially supported by the Ukrainian National Academy of Science and Russian Fund of Fundamental Research, Grant No 38/50-2008. Appendix: Properties of Lie superalgebra ======================================== Permutation relations for the generators $X_K$ of Lie superalgebra are \[A.1\] \[X\_K,X\_L} X\_KX\_L-(-1)\^[p\_Kp\_L]{}X\_LX\_K=[f\_[KL]{}]{}\^MX\_M. Structure constants ${f_{KL}}^M$ have the Grassmann parity \[A.2\] p([f\_[KL]{}]{}\^M)=p\_K+p\_L+p\_M=02, following symmetry property: \[A.3\] [f\_[KL]{}]{}\^M=-(-1)\^[p\_Kp\_L]{}[f\_[LK]{}]{}\^M and obey the Jacobi identities \[A.4\] \_[(KLM)]{}(-1)\^[p\_Kp\_M]{}[f\_[KN]{}]{}\^P[f\_[LM]{}]{}\^N=0, where the symbol $(KLM)$ means a cyclic permutation of the quantities $K$, $L$ and $M$. In the relations (\[A.1\])-(\[A.4\]) an every index $K$ takes either a Grassmann-even value $k$ $(p_k=0)$ or a Grassmann-odd one $\k$ $(p_\k=1)$. The relations (\[A.1\]) have the following components: =[f\_[kl]{}]{}\^mX\_m,{X\_,X\_ł}=[f\_[ł]{}]{}\^mX\_m,=[f\_[kł]{}]{}\^X\_. The Lie superalgebra possesses the Cartan-Killing metric tensor \[A.5\] G\_[KL]{}=(-1)\^[p\_N]{}[f\_[KM]{}]{}\^N[f\_[LN]{}]{}\^M=(-1)\^[p\_Kp\_L]{}G\_[LK]{} =(-1)\^[p\_K]{}G\_[LK]{}=(-1)\^[p\_L]{}G\_[LK]{}, which components are G\_[kl]{}=[f\_[km]{}]{}\^n[f\_[ln]{}]{}\^m-[f\_[k]{}]{}\^\^,G\_[ł]{}=[f\_]{}\^m[f\_[łm]{}]{}\^-[f\_[m]{}]{}\^\^m,G\_[kł]{}=0. As a consequence of the relations (\[A.3\]) and (\[A.4\]) the tensor with low indices \[A.6\] f\_[KLM]{}=[f\_[KL]{}]{}\^NG\_[NM]{} has the following symmetry properties: \[A.7\] f\_[KLM]{}=-(-1)\^[p\_Kp\_L]{}f\_[LKM]{}=-(-1)\^[p\_Lp\_M]{}f\_[KML]{}. For a semi-simple Lie superalgebra the Cartan-Killing metric tensor is non-degenerate and therefore there exists an inverse tensor $G^{KL}$ G\_[KL]{}G\^[LM]{}=\_K\^M.In this case, as a result of the symmetry properties (\[A.7\]), the quantity \[A.8\] X\_KG\^[KL]{}X\_L is a Casimir operator =0. [999]{} A. Galperin, E. Ivanov, V. Ogievetsky and E. Sokatchev, Ann. Phys. 185 (1988) 1. A. Galperin, E. Ivanov, V. Ogievetsky and E. Sokatchev, Ann. Phys. 185 (1988) 22. D. Cangemi and R. Jackiw, Phys. Rev. Lett. 69 (1992) 233. D.V. Soroka and V.A. Soroka, Phys. Lett. B607 (2005) 302; hep-th/0410012. S.A. Duplij, D.V. Soroka and V.A. Soroka, J. Kharkov National Univ. No. 664 (2005), Physical series “Nuclei, Particles, Fields”, Issue 2/27/, p. 12. S.A. Duplij, D.V. Soroka and V.A. Soroka, J. Zhejiang Univ. SCIENCE A 7 (2006) 629. D.V. Soroka and V.A. Soroka, Problems of Atomic Science and Technology (Ukraine) 3(1) (2007) 76; hep-th/0508141. H.S. Snyder, Phys. Rev. 71 (1947) 38. C.N. Yang, Phys. Rev. 72 (1947) 874. V.V. Khruschev and A.N. Leznov, Grav. Cosmol. 9 (2003) 159; hep-th/0207082. A. Connes, Publ. math. Inst. hautes étud. sci. 62 (1985) 41. [^1]: Note that, to avoid the double count under summation over the pair antisymmetric indices, we adopt the rules which illustrated by the following example: =cZ\_[ab]{}=[c2]{}(\_a\^c\_b\^d-\_a\^d\_b\^c)Z\_[cd]{} =\_[c<d]{}[f\_[ab]{}]{}\^[cd]{}Z\_[cd]{}=[12]{}[f\_[ab]{}]{}\^[cd]{}Z\_[cd]{},where ${f_{ab}}^{cd}$ are structure constants, and so on. [^2]: Note that in the case $D=4$ we obtain the anti-de Sitter algebra $so(3,2)$.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The Bañados-Teitelboim-Zanelli (BTZ) black hole metric solves the three-dimensional Einstein’s theory with a negative cosmological constant as well as all the generic higher derivative gravity theories based on the metric; as such it is a universal solution. Here, we find, in all generic higher derivative gravity theories, new universal non-Einsteinian solutions obtained as Kerr-Schild type deformations of the BTZ black hole. Among these, the deformed non-extremal BTZ black hole loses its event horizon while the deformed extremal one remains intact as a black hole in any generic gravity theory.' author: - Metin Gürses - Tahsin Çağri Şişman - Bayram Tekin title: 'Non-Einsteinian Black Holes in Generic 3D Gravity Theories' --- Introduction ============ The black hole in 2+1 dimensions, the BTZ metric [@btz; @BHTZ], as a solution to vacuum Einstein’s gravity with a negative cosmological constant, shares many of the features of the $\left(3+1\right)$-dimensional realistic Kerr black hole. Due to the local triviality of Einstein’s gravity in 2+1 dimensions, the BTZ solution has been a remarkable tool in exploring the quantum nature of the black hole geometry such as microscopic description of black hole entropy (see the review [@Carlip] and the references therein). Three important features of the BTZ geometry should be stressed. First, being a locally Einstein metric, it solves all the metric based higher curvature gravity equations derived from the most general action $$I=\int d^{3}x\sqrt{-g}\,{\cal L}\left(\text{Riem},\nabla\text{Riem},\cdots\right).\label{eq:Action}$$ Such metrics are called universal which are unaffected by the quantum effects [@GG1; @GG2]. Generically, for dimensions greater than three, Einstein metrics fail to solve higher derivative theories but in three dimensions since the Riemann tensor can be written in terms of the Einstein tensor $G_{\mu\nu}$ as $R_{\mu\alpha\nu\beta}=\epsilon_{\mu\alpha\sigma}\epsilon_{\nu\beta\sigma}G^{\sigma\rho}$, any Einsteinian solution also solves the higher derivative theory as long as the cosmological constant is tuned accordingly. This fact is quite important and paves way to study the Einstein metrics such as the BTZ black hole as solutions to the low energy quantum theory of gravity at any scale defined by the action (\[eq:Action\]) where the nonmetric fields are set to zero or constant values. Secondly, the BTZ geometry can be dressed with two arbitrary functions to represent all the locally Einsteinian metrics yielding the Bañados geometry as [@Banados] $$\begin{aligned} ds^{2}=\ell^{2}\Biggl[\frac{{\rm d}r^{2}}{r^{2}}+ & \left(r{\rm d}u+\frac{1}{r}f\left(v\right){\rm d}v\right)\nonumber \\ & \times\left(r{\rm d}v+\frac{1}{r}g\left(u\right){\rm d}u\right)\Biggr],\label{eq:Banados_geom}\end{aligned}$$ where $u$ and $v$ are null coordinates. The geometry corresponds to the non-extremal rotating BTZ black hole for constant nonvanishing values of $f$ and $g$; and to the extremal rotating BTZ black hole when one of these constants becomes zero. Thirdly, within the cosmological Einstein’s theory, the BTZ black hole has the uniqueness property under the conditions described in [@Rooman; @li]. Due to the importance of the BTZ black hole, one would like to know its uniqueness and also whether it is preserved as a black hole under the deformations described as $g_{\mu\nu}=\bar{g}_{\mu\nu}+h_{\mu\nu}$ in the generic higher derivative theory (\[eq:Action\]). Here, $h_{\mu\nu}$ is not a small perturbation, hence just like the BTZ black hole $\bar{g}_{\mu\nu}$, the deformed metric $g_{\mu\nu}$ is expected to solve the full field equations with the condition that the black hole property is kept intact. Without a further specification of the field equations of the theory, one cannot proceed further with this most general deformation in a theory independent way. Therefore, to keep the universal nature of the BTZ black hole under this deformation in the setting of the most general higher derivative theory, we shall consider a specific deformation which is called the Kerr-Schild–Kundt (KSK) type whose universality; i.e. it solves the generic gravity theory once a linear scalar partial differential equation is solved, has been shown in [@Gurses-PRL; @AdS-plane_pp-wave; @KSK_universal]. The KSK metric is in the form $$g_{\mu\nu}=\bar{g}_{\mu\nu}+2V\lambda_{\mu}\lambda_{\nu},\label{eq:AdS-waveKS}$$ where $V$ is a scalar field and $\lambda$ is a null vector field which satisfy the properties $$\begin{aligned} \lambda^{\mu}\lambda_{\mu} & =0,\quad\nabla_{\mu}\lambda_{\nu}\equiv\xi_{(\mu}\lambda_{\nu)},\nonumber \\ \xi_{\mu}\lambda^{\mu} & =0,\quad\lambda^{\mu}\partial_{\mu}V=0,\label{eq:AdS-wave_prop}\end{aligned}$$ for both the background and the full metric. The $\xi$ vector is defined via the second equation in (\[eq:AdS-wave\_prop\]) once the $\lambda$ null vector is chosen (a way to generate viable $\lambda$ vectors from smooth curves was given in [@smooth]). For the KSK metrics, the Ricci tensor becomes $$R_{\mu\nu}=\left(\mathcal{Q}V\right)\lambda_{\mu}\lambda_{\nu}-\frac{2}{\ell^{2}}g_{\mu\nu},$$ where $\ell$ is the AdS length and the operator $\mathcal{Q}$ is defined as $$\mathcal{Q}V\equiv\left(\bar{g}^{\mu\nu}\bar{\nabla}_{\mu}\bar{\nabla}_{\nu}+2\xi^{\mu}\partial_{\mu}+\frac{1}{2}\xi^{\mu}\xi_{\mu}-\frac{2}{\ell^{2}}\right)V.$$ Then, for the pure cosmological Einstein theory, the nonlinear field equations $R_{\mu\nu}=2\Lambda g_{\mu\nu}$ become linear in $V$ and boil down to [@gurses1] $$\mathcal{Q}V=0,\label{eq:Einstein_operator}$$ once the trace of the field equationsis is solved as $\Lambda=-1/\ell^{2}$. Given the background metric in some local coordinates, one can find the local solution. For a general gravity theory with the highest derivative order of $\left(2N+2\right)$ in the field equations with $N\ge0$, the field equations reduce to [@Gurses-PRL; @AdS-plane_pp-wave; @GravWaves3D] $$\prod_{n=1}^{N}\,\big({\cal Q}-m_{n}^{2}\big)\,{\cal Q}\,V=0,\label{denk2}$$ whose generic solution is $V=V_{E}+\sum_{n=1}^{N}\,V_{n}$ where the Einsteinian part ($V_{E}$) and the other (massive) parts, assuming nondegeneracy, satisfy the following equations $$\begin{aligned} {\cal Q}V_{E}=0,\qquad\big({\cal Q}-m_{n}^{2}\big)\,V_{n}=0.\label{denk3}\end{aligned}$$ One can also interpret these equations as transverse-traceless perturbations of the background space, therefore they correspond to massless and massive gravitons. In three dimensional Einstein’s theory, since there are no gravitons, $V_{E}$ corresponds to pure gauge transformations when the deformation $h_{\mu\nu}$ is assumed to be a perturbation about the exact background. On the other hand, the $V_{n}$ solutions are the non-Einsteinian solutions with the Ricci tensor $R_{\mu\nu}=\left(\sum_{n=1}^{N}m_{n}^{2}V_{n}\right)\lambda_{\mu}\lambda_{\nu}-2/\ell^{2}g_{\mu\nu}$. Deformations of BTZ =================== Along the lines described above, let us consider the deformations of the BTZ black hole $$d\bar{s}^{2}=-h{\rm d}t^{2}+\frac{{\rm d}r^{2}}{h}+r^{2}\left({\rm d}\phi-\frac{j}{2r^{2}}{\rm d}t\right)^{2},\label{btz1}$$ with $h\left(r\right)=-m+\frac{r^{2}}{\ell^{2}}+\frac{j^{2}}{4r^{2}}$. We shall call the generic deformation as BTZ-waves since the general solution will be of the wave form depending on the null coordinates. As we shall show below, among these only a subclass will remain a black hole. In (\[btz1\]), $m$ and $j$ are constants representing the mass and angular momentum, respectively. The outer and inner horizons of the black hole are located at $$r_{\pm}^{2}=\frac{m\ell^{2}}{2}\left(1\pm\sqrt{1-\frac{j^{2}}{m^{2}\ell^{2}}}\right).$$ which coalesce for the extremal case $j=\pm m\ell$ at $r_{0}^{2}=m\ell^{2}/2$. To understand if and how the black hole nature of the BTZ metric is changed by the KSK deformation, let us study the event horizon. In the generic case, the symmetries of the BTZ geometry are no longer symmetries of the KSK geometry. Hence, the detection of the event horizon cannot be done with the Killing vectors; instead, since the horizons will be null hypersurfaces defined as level sets of $r$, let us consider where the surface normal $\partial_{\mu}r$ becomes a null vector in the BTZ-wave geometry as $$\Omega\equiv g^{\mu\nu}\partial_{\mu}r\partial_{\nu}r=0.$$ Using (\[eq:AdS-waveKS\]) and (\[btz1\]), $\Omega$ becomes $$\begin{aligned} \Omega & =h\left(r\right)-2V\left(\lambda^{\mu}\partial_{\mu}r\right)^{2}\nonumber \\ & =2V\left(t,r_{\pm},\phi\right)\left(\left.\lambda^{r}\right\|_{r=r_{\pm}}\right)^{2}.\end{aligned}$$ Here, to have $\Omega=0$, $V\left(t,r_{\pm},\phi\right)=0$ is a possibility but recall that the metric function $V$ must satisfy a theory dependent differential equation. Then, to keep the BTZ black hole intact in a theory independent way, $$\left.\lambda^{r}\right\|_{r=r_{\pm}}=0,\label{eq:BH_cond}$$ must be satisfied. In this way, one has the deformed black hole solutions for all generic gravity theories. There can be other black hole solutions where the profile function $V$ satisfies the condition $V\left(t,r_{\pm},\phi\right)=0$ or $h\left(r\right)-2V\left(\lambda^{\mu}\partial_{\mu}r\right)^{2}=0$ for different $r$ values. In these cases, since $V$ takes different functional forms in different gravity theories then such black holes solutions will be theory dependent; and given the theory, one can construct these. Since we are interested in the KSK-type deformations of the BTZ black hole which keep the event horizon intact, we considered a null hypersurface of constant $r$ to locate the event horizon. However, for the KSK metric (\[eq:AdS-waveKS\]), in general, a null hypersurface of the form $F\left(t,r,\theta\right)={\rm constant}$ should be considered to locate a horizon as $$\bar{g}^{\mu\nu}\partial_{\mu}F\partial_{\nu}F-2V\left(t,r,\theta\right)\left(\lambda^{\mu}\partial_{\mu}F\right)^{2}=0.\label{eq:General_horizon}$$ In addition to the undeformed event horizon given with (\[eq:BH\_cond\]), to have a deformed horizon for the KSK geometry, this equation should be considered which will be studied elsewhere [@BTZwaves]. The analysis of (\[eq:General\_horizon\]) in its full generality is a tedious task; however, to get some understanding, for the $\lambda_{\mu}=\partial_{\mu}u$ case,[^1] let us consider the $r=f\left(u\right)$ hypersurface which becomes null if $$0=h\left(r\right)-2\lambda^{r}\left(\frac{df}{du}+V\left(t,r,\theta\right)\lambda^{r}\right).\label{eq:r=00003Df(u)_hypersurf}$$ To have an equation in $r$ and $u$ with a solution $r=f\left(u\right)$, one must have $V=V\left(u,r\right)$ with a $\lambda^{r}$ depending only $r$. Then, the KSK property $\lambda^{\mu}\partial_{\mu}V=0$ reduces to $$\lambda^{r}\frac{\partial V}{\partial r}=0,$$ which requires either $\lambda^{r}=0$ or $V=V\left(u\right)$. For $\lambda^{r}=0$, (\[eq:r=00003Df(u)\_hypersurf\]) becomes $h\left(r\right)=0$ so it does not provide a generalization as $r=f\left(u\right)$. Thus, one needs to have $V=V\left(u\right)$ in general. For this case, $\partial_{\mu}V=V_{u}\lambda_{\mu}$ and $\bar{g}^{\mu\nu}\bar{\nabla}_{\mu}\bar{\nabla}_{\nu}V=0$, so $\mathcal{Q}V$ reduces $$\mathcal{Q}V=\left(\frac{1}{2}\xi^{\mu}\xi_{\mu}-\frac{2}{\ell^{2}}\right)V.$$ To obtain an Einsteinian solution, $\mathcal{Q}V=0$ must be satisfied which is the case for any $V=V\left(u\right)$ if $\xi^{2}=4/\ell^{2}$. The condition $\xi^{2}=4/\ell^{2}$ is satisfied for the BTZ-waves constructed in the next section. Thus, one may find a solution for (\[eq:r=00003Df(u)\_hypersurf\]) indicating a null hypersurface of the form $r=f\left(u\right)$ exists if $V=V\left(u\right)$. However, the $V=V\left(u\right)$ solution is an Einsteinian metric which is already represented in the Bañados geometry. Note that this case also covers the shifted horizons, that is $r={\rm constant}$ but $r\ne r_{\pm}$, by having $f={\rm constant}$ and $V={\rm constant}$. On the other hand, to have a non-Einsteinian KSK geometry for an horizon of the form $r=f\left(u\right)$, the metric function $V=V\left(u\right)$ must satisfy $$\big({\cal Q}-m_{n}^{2}\big)\,V_{n}=\left(\frac{1}{2}\xi^{\mu}\xi_{\mu}-\frac{2}{\ell^{2}}-m_{n}^{2}\right)V=0,\label{eq:Non-Einsteinian_eqn_for_V(u)}$$ where $m_{n}$ depends on the parameters of the higher derivative theory. Since $\xi^{\mu}$ is theory independent, (\[eq:Non-Einsteinian\_eqn\_for\_V(u)\]) cannot be satisfied in general. Therefore, it is not possible to obtain a non-Einsteinian KSK geometry that has a horizon of the form $r=f\left(u\right)$. In [@BTZwaves], we will study more general horizon forms such as $r=f\left(u,\psi\right)$ with $\xi_{\mu}=\partial_{\mu}\psi$ which require more general $V$ beyond $V=V\left(u\right)$. In the discussion below, we will show that the condition (\[eq:BH\_cond\]), which keeps the BTZ event horizon intact, can be satisfied if and only if the BTZ seed is extremal so that a subclass of the BTZ-waves will be a deformed version of the extremal BTZ black hole. BTZ-wave construction ===================== Now, let us obtain the BTZ-wave metrics by a direct construction. As a consequence of the second property in (\[eq:AdS-wave\_prop\]), let us choose the null one-form field $\lambda_{\mu}$ to be exact, $\lambda_{\mu}=\partial_{\mu}u\left(t,r,\phi\right)$. Then, the condition that $\lambda_{\mu}$ be null yields $$\begin{aligned} -\left(\frac{\partial u}{\partial t}\right)^{2}-\frac{j}{r^{2}}\frac{\partial u}{\partial t}\frac{\partial u}{\partial\phi}\label{eq:Nullity}\\ +\left(\frac{h}{r^{2}}-\frac{j^{2}}{4r^{4}}\right)\left(\frac{\partial u}{\partial\phi}\right)^{2}+h^{2}\left(\frac{\partial u}{\partial r}\right)^{2} & =0.\nonumber \end{aligned}$$ Notice that all coefficients are a function of $r$, so the easiest way to satisfy the nullity condition is to consider a $u$ whose derivatives are either a function of $r$ or a constant as [^2] $$u\left(t,r,\phi\right)=c_{1}t+c_{2}\phi+w\left(r\right).\label{eq:ansatz}$$ This ansatz provides a solvable set of differential equations for the KSK metric properties. The solution can be put in a simpler form if the BTZ metric is written in terms of $r_{\pm}$ with $h\left(r\right)=\frac{\left(r^{2}-r_{+}^{2}\right)\left(r^{2}-r_{-}^{2}\right)}{r^{2}\ell^{2}}$ and $j=\frac{2\sigma r_{+}r_{-}}{\ell}$ where $\sigma$ represents the direction of rotation which we choose to be $\sigma=+1$. For this non-extremal BTZ seed, the $\lambda_{\mu}$ and $\xi_{\mu}$ one-forms are found to be $$\lambda_{\mu}=\left(1,\frac{\ell^{2}r\left(r_{+}+\epsilon r_{-}\right)}{\left(r^{2}-r_{+}^{2}\right)\left(r^{2}-r_{-}^{2}\right)},\epsilon\ell\right),\label{eq:lambda_nonextremal}$$ and $$\xi_{\mu}=\left(-\frac{r_{+}+\epsilon r_{-}}{\ell^{2}},-\frac{r\left(\alpha+\beta\right)}{\ell^{2}\alpha\beta},\frac{\epsilon r_{+}+r_{-}}{\ell}\right),$$ where $\epsilon$ is equal to $\pm1$, $\alpha$ and $\beta$ are defined as $\alpha\left(r\right)=\left(r^{2}-r_{+}^{2}\right)/\ell^{2}$ and $\beta\left(r\right)=\left(r^{2}-r_{-}^{2}\right)/\ell^{2}$. From (\[eq:lambda\_nonextremal\]), $\lambda^{r}$ can be calculated to be $$\lambda^{r}=h\left(r\right)\lambda_{r}=\frac{r_{+}+\epsilon r_{-}}{r}.$$ The black hole event horizon condition (\[eq:BH\_cond\]) is not satisfied, so the BTZ deformation for the nonextremal case is not a black hole in the generic theory. Yet, the resulting metric is a solution to the generic theory if $V$ satisfies the constraint $\lambda^{\mu}\,\partial_{\mu}V=0$ and (\[denk2\]) for the specific theory. The constraint can be solved in a theory independent way and the solution is $$\begin{aligned} V\left(t,r,\phi\right)=\mathcal{F}\Biggl( & t+\frac{r_{+}\ln\alpha-\epsilon r_{-}\ln\beta}{2\left(\beta-\alpha\right)},\nonumber \\ & \phi+\frac{r_{-}\ln\alpha-\epsilon r_{+}\ln\beta}{2\left(\beta-\alpha\right)}\Biggr),\label{eq:Mthd_char_soln}\end{aligned}$$ where $\mathcal{F}$ is a smooth function. Above, we discussed the nonextremal case, now let us focus to the extremal case $j=m\ell$ with $h\left(r\right)=\frac{\left(r^{2}-r_{0}^{2}\right)^{2}}{\ell^{2}r^{2}}$ and $j=\frac{2r_{0}^{2}}{\ell}$. For this case, the sign choice $\epsilon$ becomes important as one arrives at two different metrics. For $\epsilon=+1$, with a similar construction as in the nonextremal case, the $\lambda_{\mu}$ and $\xi_{\mu}$ one-forms become $$\lambda_{\mu}=\left(1,\frac{2rr_{0}\ell^{2}}{\left(r^{2}-r_{0}^{2}\right)^{2}},\ell\right),\label{eq:Extremal_lambda_nonzero_r_comp}$$ and $$\xi_{\mu}=\left(-\frac{2r_{0}}{\ell^{2}},-\frac{2r}{r^{2}-r_{0}^{2}},\frac{2r_{0}}{\ell}\right).$$ From (\[eq:Extremal\_lambda\_nonzero\_r\_comp\]), $\lambda^{r}$ can be calculated to be $$\lambda^{r}=\frac{2r_{0}}{r}.$$ Again, the black hole event horizon condition (\[eq:BH\_cond\]) is not satisfied, so the BTZ deformation for the extremal case with $\epsilon=+1$ is not a black hole in the generic theory. For $\epsilon=-1$, the KSK metric construction for the extremal case differs in a subtle way from the nonextremal construction such that (\[eq:Nullity\]) requires $w\left(r\right)$ in (\[eq:ansatz\]) to be constant. As a result, the $\lambda_{\mu}$ and $\xi_{\mu}$ one-forms become $$\lambda_{\mu}=\left(1,0,-\ell\right),\label{eq:Extremal_lambda_zero_r_comp}$$ and $$\xi_{\mu}=\left(0,\frac{2r}{r_{0}^{2}-r^{2}},0\right).\label{eq:ksi}$$ From (\[eq:Extremal\_lambda\_zero\_r\_comp\]), $\lambda^{r}$ can simply be found to be $$\lambda^{r}=0.$$ This time, the black hole event horizon condition (\[eq:BH\_cond\]) is satisfied, so the BTZ deformation for the extremal case with $\epsilon=-1$ is a black hole in the generic theory. Here, the metric function $V$ must satisfy $\lambda^{\mu}\partial_{\mu}V=0$ yielding $$\frac{\ell^{2}}{r^{2}-r_{0}^{2}}\left(\ell\frac{\partial V}{\partial t}+\frac{\partial V}{\partial\phi}\right)=0.$$ with the solution $$V=V\left(t-\ell\phi,r\right).$$ The explicit form of $V$ will be given below for Einstein’s theory and the new massive gravity (NMG) [@NMG]. Extremal-BTZ Wave Solution of Einstein’s Gravity ------------------------------------------------ We showed that the only possible KSK deformation of BTZ black hole which keeps the black hole nature intact is the extremal BTZ black hole deformed with the constant null vector field of $\lambda_{\mu}=\left(1,0,-\ell\right)$. Now, let us find the metric function $V$ for the cosmological Einstein’s gravity by solving (\[eq:Einstein\_operator\]). With (\[eq:ksi\]), the field equation for $V$ becomes $$r\frac{\partial^{2}}{\partial r^{2}}V_{E}\left(u,r\right)-\frac{\partial}{\partial r}V_{E}\left(u,r\right)=0,$$ where we defined $u=t-\ell\phi$ which is in fact the generating function for $\lambda_{\mu}$ as $\lambda_{\mu}=\partial_{\mu}u$. If $r\ne r_{0}$, the Einsteinian solution becomes $$V_{E}\left(u,r\right)=c_{1}\left(u\right)r^{2}+c_{2}\left(u\right),\label{eq:V_E}$$ yielding the metric $$\begin{aligned} ds^{2}= & d\bar{s}^{2}+2\left(c_{1}\left(u\right)r^{2}+c_{2}\left(u\right)\right)\left({\rm d}t-\ell\,{\rm d}\phi\right)^{2},\end{aligned}$$ where $d\bar{s}^{2}$ is the extremal BTZ seed. This result is consistent with the Bañados geometry (\[eq:Banados\_geom\]) and the analysis of [@li]. As in the case of the Bañados geometry which dresses the BTZ black hole with two arbitrary functions, our generic solution with arbitrary $c_{1}\left(u\right)$ and $c_{2}\left(u\right)$ are of the nonlinear wave type which we called the BTZ wave. To understand this solution better, we can compute its mass and angular momentum using the Abbott-Deser approach [@Abbott]. Assuming $c_{1}\left(t-\ell\phi\right)=c_{2}\left(t-\ell\phi\right)=0$ and $r_{0}=0$ to be the background, the mass corresponding to the background time-like Killing vector $\zeta^{\mu}=\left(-1,0,0\right)$ is $M=m+\frac{2}{\pi}\int_{0}^{2\pi}d\phi\,c_{2}\left(t-\ell\phi\right)$; and the angular momentum corresponding to the background Killing vector $\zeta^{\mu}=\left(0,0,1\right)$ is $J=m\ell+\frac{2\ell}{\pi}\int_{0}^{2\pi}d\phi\,c_{2}\left(t-\ell\phi\right)$. We have kept mass and angular momentum computation with generic $c_{1}\left(u\right)$ and $c_{2}\left(u\right)$. Since this solution is no longer stationary, its mass angular momentum are time dependent via these functions. Note that the extremality condition is intact as $J=M\ell$. The function $c_{1}\left(t-\ell\phi\right)$ corresponds to a pure gauge and does not appear in the mass and angular momentum expressions. Of course, for a stationary black hole solution, the arbitrary $u$ dependent functions should be taken as constants as we mentioned for (\[eq:Banados\_geom\]). Then, one obtains time-independent mass and angular momentum. The discussion is exactly like the case of Bañados metric [@Banados; @li]. Extremal-BTZ Wave Solution of ----------------------------- NMG Now, we study the solution of cosmological new massive gravity (NMG) given with the action $$I=-\frac{1}{\kappa^{2}}\int d^{3}x\,\sqrt{-g}\left(R-2\Lambda_{0}+L^{2}K\right),$$ whose field equations are $$G_{\mu\nu}+\Lambda_{0}g_{\mu\nu}-\frac{L^{2}}{2}K_{\mu\nu}=0,\label{eq:NMG_eom}$$ where $K_{\mu\nu}=2\square R_{\mu\nu}-\frac{1}{2}\left(\nabla_{\mu}\nabla_{\mu}+g_{\mu\nu}\square\right)R+4R_{\mu\alpha\nu\beta}R^{\alpha\beta}-\frac{3}{2}RR_{\mu\nu}-g_{\mu\nu}K$ and the trace $K=g^{\mu\nu}K_{\mu\nu}=R_{\mu\nu}R^{\mu\nu}-\frac{3}{8}R^{2}$. Putting the metric of the extremal BTZ wave defined by $\lambda_{\mu}$ given in (\[eq:Extremal\_lambda\_zero\_r\_comp\]) yields the field equations $$\begin{aligned} \frac{1}{\ell^{2}}+\Lambda_{0}+\frac{L^{2}}{4\ell^{4}} & =0,\label{eq:Tr_eqn}\\ \big({\cal Q}-m_{g}^{2}\big){\cal Q}V & =0,\label{eq:Trless_eqn}\end{aligned}$$ where $m_{g}^{2}$ is the mass of the spin-2 graviton of the NMG theory given as $$m_{g}^{2}=\frac{1}{L^{2}}-\frac{1}{2\ell^{2}}.$$ The first equation determines the effective cosmological parameter $\ell$. The second equation (\[eq:Trless\_eqn\]) determines the metric function $V$ and has the general solution $$V\left(u,r\right)=V_{E}\left(u,r\right)+V_{p}\left(u,r\right),\label{eq:General_soln}$$ where $u=t-\ell\phi$ and $V_{E}$ is the Einsteinian solution (\[eq:V\_E\]) while $V_{p}$ is the solution of the massive operator $\big({\cal Q}-m_{g}^{2}\big)$ which can be found as $$\begin{aligned} V_{p}\left(u,r\right)= & c_{3}\left(u\right)\left(r^{2}-r_{0}^{2}\right)^{\left(1+p\right)/2}\nonumber \\ & +c_{4}\left(u\right)\left(r^{2}-r_{0}^{2}\right)^{\left(1-p\right)/2},\end{aligned}$$ with $p\equiv\sqrt{m_{g}^{2}\ell^{2}+1}$. The reality of $p$ is equivalent to the Breitenlohner-Freedman (BF) bound [@BF]. It is important to note that the solution (\[eq:General\_soln\]) to this quadratic theory solves all higher curvature theories as long as the corresponding effective cosmological constant equation is satisfied. Using the construction of [@Deser_Tekin-PRL], one can show that the finiteness of mass and angular momentum requires $c_{3}\left(u\right)=c_{4}\left(u\right)=0$ for $0<p<1$, $c_{3}\left(u\right)=0$ for $1<p$, or $c_{4}\left(u\right)=0$ for $p<-1$ yielding the mass $M=m\left(1+\frac{2}{2p^{2}-1}\right)$ and the angular momentum $J=M\ell$ such that extremality is kept intact. Conclusions =========== We have studied the exact deformation of the BTZ black hole in the context of generic gravity; and showed that the non-extremal black hole loses its exact horizon and the resulting deformed metric is of wave type, which we called the BTZ wave. Surprisingly, the deformed extremal black hole remains a black hole. There are several ways to read this result: First, the non-extremal BTZ is unique in generic gravity while the extremal one is not as in the case of Einstein’s theory; secondly, considering the deformations as generic quantum or classical corrections, the non-extremal BTZ is not preserved as a black hole solution to the generic gravity while the extremal one remains a black hole in any generic gravity theory. Lastly, regarding the $r=0$ singularity after the KSK deformation, note that all the curvature invariants of the KSK metrics are constant; therefore, there is no curvature singularity. [10]{} M. Banados, C. Teitelboim, and J. Zanelli, “The Black hole in three-dimensional space-time,” Phys. Rev. Lett. 69, 1849 (1992). M. Banados, M. Henneaux, C. Teitelboim and J. Zanelli, “Geometry of the (2+1) black hole,” Phys. Rev. D 48, 1506 (1993) Erratum: [\[]{}Phys. Rev. D 88, 069902 (2013)[\]]{}. S. Carlip, “Black Hole Thermodynamics,” Int. J. Mod. Phys. D 23, 1430023 (2014). G. W. Gibbons, “Quantized Fields Propogating in Plane Waves Spacetimes”, Commun. Math. Phys. 45, 191-202 (1975). A. A. Coley, G. W. Gibbons, S. Hervik and C. N. Pope, “Metrics with vanishing Quantum Corrections”, Class. Quantum Grav. 25, 145017 (2008). M. Banados, “Three-dimensional quantum geometry and black holes,” AIP Conf. Proc. 484, no. 1, 147 (1999). M. Rooman and P. Spindel, “Uniqueness of the asymptotic AdS(3) geometry,” Class. Quant. Grav. 18, 2117 (2001). C. Li and J. Lucietti, “Three-dimensional black holes and descendants,” Phys. Lett. B 738, 48 (2014). M. Gurses, S. Hervik, T. C. Sisman, and B. Tekin, “Anti-de Sitter-Wave Solutions of Higher Derivative Theories”, Phys. Rev. Lett. 111, 101101 (2013). M. Gurses, T. C. Sisman and B. Tekin, “AdS-plane wave and pp-wave solutions of generic gravity theories,” Phys. Rev. D 90, 124005 (2014). M. Gurses, T. C. Sisman and B. Tekin, “Kerr-Schild--Kundt Metrics are Universal,” Class. Quant. Grav. 34, 075003 (2017). M. Gurses, T. C. Sisman and B. Tekin, “From Smooth Curves to Universal Metrics,” Phys. Rev. D 94, 044042 (2016). M. Gurses, T. C. Sisman and B. Tekin, “New Exact Solutions of Quadratic Curvature Gravity”, Phys. Rev. D 86, 024009 (2012). M. Gurses, T. C. Sisman and B. Tekin, “Gravity Waves in Three Dimensions,” Phys. Rev. D 92, 084016 (2015). M. Gurses, T. C. Sisman and B. Tekin, in progress. E. A. Bergshoeff, O. Hohm and P. K. Townsend, “Massive Gravity in Three Dimensions,” Phys. Rev. Lett. 102, 201301 (2009). L. F. Abbott and S. Deser, “Stability Of Gravity With A Cosmological Constant,” Nucl. Phys. B 195, 76 (1982). P. Breitenlohner and D. Z. Freedman, “Positive Energy in anti-De Sitter Backgrounds and Gauged Extended Supergravity,” Phys. Lett. 115B, 197 (1982). S. Deser and B. Tekin, “Gravitational Energy in Quadratic Curvature Gravities,” Phys. Rev. Lett. 89, 101101 (2002). [^1]: This choice is motivated at the beginning of the next section. [^2]: There can be other choices for the function $u$ providing different solutions which will be discussed elsewhere [@BTZwaves].	{ "pile_set_name": "ArXiv" }
--- abstract: 'We present a formulation of gravity in terms of a theory based on complex $SU(2)$ gauge fields with a general coordinate invariant action functional quadratic in the field strength. Self-duality or anti-self-duality of the field strength emerges as a constraint from the equations of motion of this theory. This in turn leads to Einstein gravity equations for a dilaton and an axion conformally coupled to gravity for the self-dual constraint. The analysis has also been extended to $N=1$ and $~ 2$ super Yang-Mills theory of complex $SU(2)$ gauge fields. This leads, besides other equations of motion, to self-duality/anti-self-duality of generalized supercovariant field-strengths. The self-dual case is then shown to yield as its solutions $N~=~1,~2$ supergravity equations respectively.' author: - \| Romesh K. Kaul\ The Institute of Mathematical Sciences,\ Chennai 600 113, India.\ [email protected] title: GAUGE THEORY OF GRAVITY AND SUPERGRAVITY --- section1 section2 section3 section4 section5 1.5cm acknowledgements references	{ "pile_set_name": "ArXiv" }
--- bibliography: - 'bib.bib' title: 'A review of the Dividend Discount Model: from deterministic to stochastic models' --- \[chap-struct\] Introduction {#Sec:Intro} ============ This chapter presents a review of the dividend discount models starting from the basic models [@williams1938theory; @gordon1956capital] to more recent and complex models [@ghezzi2003stock; @barbu2017novel; @dmultivariate] with a focus on the modelling of the dividend process rather than the discounting factor, that is assumed constant in most of the models. The Chapter starts with an introduction of the basic valuation model with some general aspects to consider when performing the computation. Then, Section \[Sec:Gordon\] presents the Gordon growth model [@gordon1962investment] with some of its extensions [@malkiel1963equity; @fuller1984simplified; @molodovsky1965common; @brooks1990n; @barsky1993does], and reports some empirical evidences. Extended reviews of the Gordon stock valuation model and its extensions can be found in [@kamstra2003pricing] and [@damodaran2012investment]. In Section \[Sec:Markov\], the focus is directed to more recent advancements which make us of the Markov chain to model the dividend process [@hurley1994realistic; @yao1997trinomial; @hurley1998generalized; @ghezzi2003stock; @barbu2017novel; @dmultivariate]. The advantage of these models is the possibility to obtain a different valuation that depends on the state of the dividend series, allowing the model to be closer to reality. In addition, these models permit to obtain a measure of the risk of the single stock or a portfolio of stocks. General Model {#Sec:General} ============= Stock valuation is one of the basic aspects of financial markets. Discussions about the fair price of a stock, or its overpricing and underpricing, have always been of paramount importance to investors. [@williams1938theory] was the first to recognise that market prices and fundamental values are . In his work, he states that an asset’s intrinsic long-term value is the present value of all future cash flows, i.e., dividends and future selling price. Let $P(t)$ be the random variable giving the fundamental value of a stock at time $t\in \mathbb{N}$. Let $D(t)$ be the dividend at time $t\in \mathbb{N}$, also assumed to be a random variable, and denote by $k_e(t)$ the required rate of return on the stock at time $t$. If we buy a stock at time $t$ and plan to sell it at time $t+1$, the price $p(t):=\mathbb{E}_{(t)}[P(t)]$ that we pay is the expected value of the stock price at time $t+1$ plus the cash flows distributed by the company, all discounted at an appropriate measure of risk $k_e(t)$, $$\label{Eq:DDMonePeriod} p(t)=\mathbb{E}_{(t)}\Bigg[\frac{P(t+1)+D(t+1)}{1+k_e(t)}\Bigg],$$ If we buy and hold the stock indefinitely, and assuming [see, e.g., @samuelson1973proof] $$\label{Eq:BubbleCondition} \lim_{i\rightarrow +\infty}\mathbb{E}_{(t)}\Bigg[\frac{P(t+i)}{\prod_{j=0}^i\big[1+k_e(t+j)\big]}\Bigg]=0,$$ then the price we pay is the expected value of all future cash flows in the form of dividends, $$\label{Eq:DDM} p(t)=\sum_{i=0}^{+\infty}\mathbb{E}_{(t)}\Bigg[\frac{D(t+i+1)}{\prod_{j=0}^i\big[1+k_e(t+j)\big]}\Bigg].$$ If condition (\[Eq:BubbleCondition\]) is not assumed, then [@blanchard1982bubbles] proved that there could exist different solutions of the fundamental equation, i.e., there is the presence of bubbles in the stock market. To solve equation (\[Eq:DDM\]), we have to identify two inputs, namely future dividends and the required measure of risk. When estimating future dividends, because of the impossibility of making predictions through to infinity, many models make assumptions about the dividend growth. The basic Gordon model [@gordon1962investment] is based on a constant dividend growth rate, while multistage models are advanced by [@brooks1990n] and [@barsky1993does] to better describe the dividend growth series. [@donaldson1996new] generalise the Gordon growth model to allow for arbitrary dividend growth and discount rates using a Monte Carlo simulation. On the contrary, others models apply specific stochastic processes to forecast dividends. [@gutierrez2004switching] propose a model which allows a regime switching in the dividend process and [@korn2005stocks] model dividends as a deterministic transformation of a Levy process. [@hurley2013calculating] introduces dividends modelled as a Bernoulli process with a continuous set of values, while [@eisdorfer2014pricing] model the time series behaviour of dividend growth rates with a first-order autoregressive process. In this Chapter we focus on how the various models make assumptions about the dividend process, with a particular attention to the Markov chain based models. The second input of the equation is the discount factor $k_e(t)$, or cost of equity, that represents a measure of the asset’s riskiness. In most of the dividend discount models, it is assumed to be constant, $k_e$. Traditionally, the estimation of $k_e$ has been performed using the Capital Asset Pricing Model (CAPM). This model originates from the idea of mean-variance efficient portfolio of [@markowitz1952portfolio], and it is formalised by [@sharpe1964capital] and [@lintner1965security] and extended by [@black1972capital]. The rationale of the model is that risky investments $R_{i}$, e.g., stocks in financial markets, are expected to be more remunerating than the risk-free assets $$\begin{aligned} & \mathbb{E}[R_i]=R_f+\beta_{im}(\mathbb{E}[R_m]-R_f),\\ & \beta_{im}=\frac{Cov[R_i,R_m]}{Var[R_m]}, \end{aligned}$$ where $R_m$ is the return on the market portfolio, and $R_f$ is the return on the risk-free asset. The [@black1972capital] version substitutes the risk-free rate with a zero-beta portfolio uncorrelated with the market. The coefficient $\beta_{im}$ represents the correlation of the stock with the market, and can be estimated as the slope coefficient of the OLS regression $$Z_{it} = \alpha_{im}+\beta_{im}Z_{mt}+\epsilon_{it},$$ where $Z_{it}$ is the excess return of the stock on the risk-free asset, or equity premium, and $Z_{mt}$ is the market risk premium, $\mathbb{E}[R_i]-R_f$. In practical applications, the market return and the risk-free rate are proxied by market indices, e.g., S&P 500 Index, and government treasury bonds, respectively. The estimation is based on a period of time that generally extends to about five years of historical data [@campbell1997econometrics]. Many authors provides empirical evidence on the CAPM application [see, e.g., @jensen1972capital; @fama1973risk; @blume1973new; @basu1977investment; @fama1992cross; @fama1993common], while [@roll1977critique] criticise it because the market portfolio is not observable and therefore the model is not testable. For a comprehensive description of the CAPM models and its variations with econometrics analysis see, e.g., [@campbell1997econometrics; @cochrane2009asset]. In general, the dividend discount model is a very attractive model because it is intuitive and easy to implement. Nevertheless, it encounters much criticism because of the limits it poses. The main argument is the applicability of the model only to certain firms with stable, high-paying dividend policy. Moreover, the firms’ recent practice of performing share buybacks instead of paying dividends, for obvious tax reasons, reduces the dividend cash flow and the application of the dividend discount model results in an underestimation of the value of the firm. The same principle applies to other assets that are ignored in the model, e.g., the value of brand names. However, share buybacks and values of other assets can be included in the dividends flow and treated as such with adequate adjustments [see, e.g., @damodaran2012investment]. Gordon Growth Model and Extensions {#Sec:Gordon} ================================== Equation (\[Eq:DDM\]) can be rewritten in terms of dividend growth, defining $$\label{Eq:DivGrowth} g(t)=\frac{D(t+1)-D(t)}{D(t)},$$ as the growth rate of dividends from time $t$ to time $t+1$, so that $D(t+1)=D(t)(1+g(t))$ and $D(t+2)=D(t)(1+g(t))(1+g(t+1))$. Then, the price becomes, $$\label{Eq:GeneralDivDiscount} p(t)=D(t)\sum_{i=0}^{+\infty}\mathbb{E}_{(t)}\Bigg[\prod_{j=0}^{i}\frac{1+g(t+j)}{1+k_e(t+j)}\Bigg].$$ Assuming a constant dividend growth rate $g(t+j)=g$ and a constant discounting factor $k_e(t+j)=k_e$, equation (\[Eq:GeneralDivDiscount\]) becomes $$\label{Eq:GeneralDivDiscountConstant} p(t)=D(t)\sum_{i=0}^{+\infty}\frac{(1+g)^i}{(1+k_e)^i},$$ and summing the geometric progression, we obtain the Gordon fundamental price estimate [@gordon1962investment] $$\label{Eq:GordonGrowthModel} p^G(t)=D(t)\frac{1+g}{k_e-g}, \quad\text{or}\quad p^G(t)=\frac{D(t+1)}{k_e-g},$$ with the constraint $g<k_e$ to obtain a finite price. The Gordon model is straightforward because it requires only estimates of the dividend growth rate and discount rate, that are both easily obtained from a company’s historical data. Nevertheless, it has some limitations. The model can result in incorrect estimations of the price when the growth rate approaches the discount rate, as the price tends to grow up to infinity. Therefore, this model is more suitable for companies with a stable dividend policy with a growth that is less than the growth of the economy. Moreover, empirical applications of the Gordon model show that dividends tend to grow exponentially, meaning that a linear growth model is not suitable for the stock valuation [see, e.g., @campbell1987cointegration; @west1988dividend]. The assumption of constant growth of the dividends forever is not realistic. To relax this assumption, [@malkiel1963equity] introduces a 2-stage model, with the first period of $n$ years of extraordinary growth followed by a stable growth forever. The value of a stock can be obtained as the sum of first years values, calculated from the general model plus a discounted value of the Gordon growth model at year $n$: $$\label{Eq:2stageDDM} p^{2st}(t)=\mathbb{E}_{(t)}\Bigg[\sum_{i=0}^n\frac{D(t+i+1)}{\prod_{j=0}^i\big[1+k_e(t+j)\big]}+\frac{P^G(n)}{\prod_{j=0}^n\big[1+k_e(t+j)\big]}\Bigg],$$ where $P^G(n)$ is the Gordon growth fundamental price estimate (\[Eq:GordonGrowthModel\]) at year $n$. A further assumption of constant growth in the first phase, $g_h$, and constant discount rate $k_{e,h}$, simplifies equation (\[Eq:2stageDDM\]) to $$\label{Eq:2stageDDMconstant} p^{2st}(t)=\frac{D(t)(1+g_h)\bigg[1-\frac{(1+g_h)^n}{(1+k_{e,h})^n}\bigg]}{k_{e,h}-g}+\frac{P^G(n)}{(1+k_{e,h})^n},$$ This model is suitable for valuing companies that expect to have an initial growth period higher than normal, because of a specific investment or a patent right, that will result in higher profits. At the same time, it presents some limits. First, the growth rate is expected to drop drastically from high to normal level, and second, it is hard to define the length of the high growth period in practical terms. To avoid the sharp drop from high to stable growth rate, [@fuller1984simplified] propose a linear decline of the growth in their model. The high growth phase with decline is assumed to last $2H$ periods up to the stable growth phase $g_n$, with an initial growth rate $g_a$. The model assumes that the discount rate $k_e$ is constant over time, as well as the dividend payout ratio. $$\label{Eq:H-DDM} p^{H}(t)=\frac{D(t)(1+g_a)}{k_e-g_n}+\frac{D(t)H(g_a-g_n)}{k_e-g_n},$$ A constant payout ratio assumption poses some limits to this model. Generally, a company is expected to have lower payout ratios in high growth phases and higher payout ratios in the stable growth phase, as shown in Figure \[fig:3-stage\]. ![Expected growth in a 3-stage dividend discount model [Figure from @damodaran2012investment p. 341][]{data-label="fig:3-stage"}](3-stage.png){width="0.9\columnwidth"} A 3-stage model, initially formulated by [@molodovsky1965common] and derived from a combination of the H model and the 2-stage model, with the inclusion of a variable payout policy and different discount factors for the various phases, overcomes the limits of previous models, but it requires a larger number of inputs. Let $k_{e,h}$, $k_{e,d}$, and $k_{e,st}$ be the discount factors for high, declining, and stable phases, respectively. Let $g_a$ and $g_n$ be the growth rate at the beginning and the end of the period. Let $EPS$ be the earnings per share, and $\Pi_a$ and $\Pi_n$ the payout ratios at the beginning and end of the period, respectively. The stock valuation for the 3-stage model is $$\label{Eq:3stageDDM} \begin{aligned} p^{3st}(t)=\sum_{i=0}^{n1}\frac{EPS(t)\Pi_a(1+g_a)^i}{(1+k_{e,h})^i}+\sum_{i=n1+1}^{n2}\frac{D(t+i)}{(1+k_{e,d})^i}\\ +\frac{EPS(t+n2)\Pi_n(1+g_n)}{(k_{e,st}-g_n)(1+k_{e,h})^{n1}(1+k_{e,d})^{n2-n1}}, \end{aligned}$$ An empirical comparison of the Gordon model and its variations is in [@sorensen1985some]. The authors analyse the intrinsic value of a random sample of 150 firms from the S&P 400 using data available in 1981, from four different valuation models, price/earning model, constant growth model, two-period, and three-period model. They base the analysis on normalised earnings and a dividend payout ratio of approximately 45 per cent. The discount factor is calculated using the CAPM model for the growth period, according to the beta of the stock and the high growth period is assumed to last five years for all the stock. Then, based on the assumption that all mature firms look alike, an equal risk measure of 8% among all the stocks is adopted for the stable phase. For every model, the authors generate five portfolios of 30 stocks each, ordered from undervalued to overvalued securities, estimating returns for two years. Results show that the increased complexity of the model improves the annualised returns. As well as looking at the risk characteristics of the portfolios, the 3-stage model outperforms the other model. [@brooks1990n] generalise the 2-stage model from [@malkiel1963equity]. They propose an N-stage model, with quarterly dividends and fractional periods. Within each stage, dividends growth is assumed constant, and the discount rate is based on quarterly compounding $r_e = (1-k_e)^{\frac{1}{4}}-1$. They test the model on the case of Commonwealth Edison Company (CWE), an electricity supplier, estimating the required rate of return for three cases: (a) annual dividends, no fractional period; (b) quarterly dividends, no fractional periods; and (c) quarterly dividends, fractional periods. They show that ignoring quarterly compounding and fractional periods the results present a downward bias. Another extension of the Gordon growth models is given in [@barsky1993does]. The authors propose to model the permanent dividend growth as a geometric average of past dividend changes: $$\label{Eq:IMA} g(t)=(1-\theta)\sum_{i=0}^t\theta^i \Delta D(t-i) + \theta^tg(0)$$ with $g(t)$ following a random walk process and, thus, change in dividends following an IMA(1,1). [@donaldson1996new] generalise the Gordon growth model allowing for arbitrary dividend growth and discount rates. Their methodology involves a Monte Carlo simulation and numerical integration of the random joint process of dividend growth and discount rates $$\label{Eq:DKprocess} y(t+1)=\frac{1+g(t+j)}{1+k_e(t+j)}.$$ They forecast a range of possible evolution of the process $y(t+1)$ up to a certain point in the future, $t+I$, and calculate the average of several estimations of the present stock value $$\label{Eq:DKprice} p(t) = D(t)\sum_{i=0}^I\prod_{j=0}^iy(t+1).$$ Markov chain stock models {#Sec:Markov} ========================= According to equation (\[Eq:GeneralDivDiscount\]), the stock valuation is obtained through two inputs, namely the dividend growth and the discount factor. The idea of the Markov chain stock models is to describe the dividend growth rate as a sequence of independent, identically distributed, discrete random variables, and model it as a Markov process. In all these models, the discount factor $k_e$ is kept constant. [@hurley1994realistic] model the dividend growth as a Markov dividend stream. They assume that in each period the dividend can increase with probability $q$, be the same with probability $1-q$, to resemble a step pattern in the long term. Moreover, they include the possibility for the firm to go bankrupt, with probability $q_B$. They propose two variations of the model, an additive model and a geometric model, both giving an estimation of the value, along with a lower bound estimation for each of these values. In the additive model, the dividend at time $t+1$ increase by the amount $\Delta$ with probability $q$, and assuming a constant discount rate $k_e$, the value of the firm is $$\label{Eq:HurleyAdditiveValue} p(t)=\begin{cases} D(t) + \Delta + p(t+1)\frac{D(t)+\Delta}{1+k_e} &\mbox{with prob } q\\ D(t) + p(t+1)\frac{D(t)}{1+k_e} &\mbox{with prob } 1-q-q_B\\ 0 &\mbox{with prob } q_B\\ \end{cases},$$ and the closed form solutions for the value and the lower bound are $$\label{Eq:HurleyAddClosed} p^A(t)=\frac{D(t)}{k_e}+\bigg[\frac{1}{k_e}+\frac{1}{k_e^2}\bigg] q\Delta,$$ and $$\label{Eq:HurleyAddLower} p^{A}_{low}(t)=\frac{D(t)(1-q_B)}{k+q_B}+\bigg[\frac{1}{k+q_B}+\frac{1}{(k+q_B)^2}\bigg] q\Delta.$$ Note that, when $q_b=0$, $p^{A}_{low}=p^A$. The geometric model assumes that the dividend increases with a growth rate $g$ and with a probability $q$ $$\label{Eq:HurleyGeometricDividend} D(t+1)=\begin{cases} D(t)(1+g) &\mbox{with prob } q\\ D(t) &\mbox{with prob } 1-q-q_B \end{cases}.$$ The closed form solutions for the value and the lower bound become $$\label{Eq:HurleyGeoClosed} p^G(t)=\frac{D(t)(1+qg)}{k_e-qg},$$ and $$\label{Eq:HurleyGeoLower} p^{G}_{low}(t)=D(t)\bigg[\frac{1+qg-q_B}{k_e-(qg-q_B)}\bigg].$$ It is worth noting that the geometric model reduces to the Gordon model, setting the expected growth rate to $qg-q_B$, or, if we exclude the possibility of bankruptcy, setting the expected growth rate to $qg$. An empirical application to three stocks, provided in [@hurley1994realistic], shows that the geometric method performs well when the dividend series is erratic and does not always show increases. The model gives an estimation that is very close to the actual stock prices. [@hurley1998generalized] formulate a generalised version of their model to include the possibility of a decrease in the dividends, so the dividend at time $t$ is $D(t)=D(t-1)+\Delta_i$ for the additive model, and $D(t)=D(t-1)(1+g_i)$ with probability $q_i$ for the geometric model. Both $\Delta_i$ and $g_i$ include the possibility of dividends reduction, or suspensions. Under the condition $q_0+\sum_{i=1}^nq_i=1$, the closed form solution for both models are $$\label{Eq:HurleyAddGeneral} p^A(t)=\frac{D(t)}{k_e}+\bigg[\frac{1}{k_e}+\frac{1}{k_e^2}\bigg] \sum_{i=1}^nq_i\Delta_i,$$ and $$\label{Eq:HurleyGeoGeneral} p^G(t)=D(t)\frac{1+\sum_{i=1}^nq_ig_i}{k_e-\sum_{i=1}^nq_ig_i}.$$ When $n=1$, the models reduces to [@hurley1994realistic] models. [@yao1997trinomial] advances the same proposal of a dividend reduction extending [@hurley1994realistic] models. The author introduces a trinomial dividend valuation model and extends the additive model, where the dividend at time $t+1$ is $$\label{Eq:YaoAdditiveDividend} D(t+1)=\begin{cases} D(t) + \Delta &\mbox{with prob } q^u\\ D(t) - \Delta &\mbox{with prob } q^d\\ D(t) &\mbox{with prob } q^c=1-q^u-q^d \end{cases},$$ with closed solution for the stock value $$\label{Eq:YaoAddValue} p^A(t)=\frac{D(t)}{k_e}+\bigg[\frac{1}{k_e}+\frac{1}{k_e^2}\bigg] (q^u-q^d)\Delta.$$ Then, the geometric model, with $$\label{Eq:YaoGeometricDividend} D(t+1)=\begin{cases} D(t)(1+g) &\mbox{with prob } q^u\\ D(t)(1-g) &\mbox{with prob } q^d\\ D(t) &\mbox{with prob } q^c=1-q^u-q^d \end{cases},$$ and closed solution $$\label{Eq:YaoGeoValue} p^G(t)=D(t)\frac{1+(q^u-q^d)g}{k_e-(q^u-q^d)g}.$$ Lower bounds for both models are also given by the author. Moreover, a practical application on five firms, provided in [@yao1997trinomial], shows that the model produces better estimates than [@hurley1994realistic]. [@ghezzi2003stock] start from the previous Markov models to formulate a more general Markov chain stock model. The authors begin with a description of the simple model for the dividend growth rate using a 2-state discrete Markov chain, and a constant discount rate $r=1+k_e$. Finally, they extend the model to an n-state Markov chain and define a vector of price-dividend ratios as the solution of a system of linear equations. In previous models, [@hurley1994realistic; @hurley1998generalized] and [@yao1997trinomial] assume that the dividend growth rates are independent, identically distributed, discrete random variables, thus obtaining one closed form solution irrespective of the state of the dividend. On the contrary, [@ghezzi2003stock] relax the i.i.d. assumption and obtain a different price-dividend solution for each state of the dividends. This variety allows the Markov chain stock model to be closer to reality. The dividend series obeys the difference equation $$\label{Eq:GhezziDividend} D(k+1)=G(k+1)D(k), \quad k=t,t+1,\hdots,$$ where $G(k+1)$ is the dividend growth factor described by a Markov chain. The dividend series relation (\[Eq:GhezziDividend\]) states that given the initial dividend value $D(0)=d\in \mathbb{R}$, we can compute the next random dividend $D(1)=G(1)D(0)=G(1)d$, and the next $D(2)=G(2)D(1)=G(2)G(1)d$, and so on. Generally $D(n)=\prod_{i=1}^{n}G(i)d$. The combination of the dividend discount model equation (\[Eq:GeneralDivDiscount\]) and (\[Eq:GhezziDividend\]), with a constant discount factor $r$, i.e., one plus the required rate of return, yields $$\label{Eq:GhezziValue} p(k)=d(k)\sum_{i=1}^{+\infty}\frac{\mathbb E_{(k)}[\prod_{j=1}^{i}G(k+j)]}{r^{i}} =: d(k)\psi_1(g(k)),$$ where $d(k)$ and $g(k)$ are the values at time $k$ of the dividend process and of the growth dividend process, respectively, and $\psi_1(g(k))$ is the price-dividend ratio. The simple case is modelled with a 2-state Markov chain taking values in the state space $E=\{g_1,g_2\}$. Let $\mathbf{P}=(p_{ij})_{i,j\in E}$ be the one-step transition probability matrix of this Markov chain, and let $$\label{Eq:GhezziCondition} {\bf A1}: \overline{g}: = \max(p_{11}g_{1}+p_{12}g_{2},p_{21}g_{1}+p_{22}g_{2})<r,$$ be the largest one step conditional expectation on the dividend growth rate. If A1 holds true, then the series $p(k)=\sum_{i=1}^{+\infty}\frac{\mathbb E_{(k)}[D(k+i)]}{r^{i}}$ converges and satisfies the asymptotic condition in (\[Eq:BubbleCondition\]), and the pair $(\psi_1(g_{1}),\psi_1(g_{2}))$ is the unique and non-negative solution of the linear system $$\label{Eq:Ghezzi2stateSystem} \begin{aligned} & \psi_1(g_{1})=p_{11}\frac{\psi_1(g_{1})g_{1}+g_{1}}{r}+p_{12}\frac{\psi_1(g_{2})g_{2}+g_{2}}{r}\\ & \psi_1(g_{2})=p_{21}\frac{\psi_1(g_{1})g_{1}+g_{1}}{r}+p_{22}\frac{\psi_1(g_{2})g_{2}+g_{2}}{r}. \end{aligned}$$ Assuming that for any given $D(k)$ we obtain the same $\mathbb{E}[D(t+1)]$, irrespective of the initial states $g_1,g_2$, then $p_{11}=q$ and $p_{22}=1-q$, therefore the solution to (\[Eq:Ghezzi2stateSystem\]) becomes $$\label{Eq:GhezziReduction} \psi_1(g_{1})=\psi_1(g_{2})=\frac{qg_1+(1-q)g_2}{r-1g_1-(1-q)g_2},$$ thus implying that the same price-dividend ratio is attached to each state, sharing the same results as [@hurley1994realistic; @hurley1998generalized] and [@yao1997trinomial]. Results can be easily extended to the case of an s-state Markov chain with state space $E=\{g_{1},g_{2}, \ldots, g_{s}\}$, where assumption A1 becomes, $$\label{Eq:GhezziConditionNstate} \overline{g}:=\max_{i\in E}\bigg(\sum_{j=1}^{s}p_{ij}g_{j}\bigg)<r.$$ If $\overline{g}<r$ the series (\[Eq:GhezziValue\]) converges and the unique and non-negative solution to the linear system is $$\label{Eq:GhezziNstateSystem} \psi(g_{i})=\sum_{j=1}^s p_{ij}\frac{\psi(g_{j})g_{j}+g_{j}}{r}, \quad i=1,2,\hdots,s.$$ This model has the advantage of assigning a different price-dividend ratio to each value of the states, that does not depend on the time. Forecasts on the dividend growth rate are updated based on the previous value of the state, according to the Markov property, thus the price of the stock is updated according to the state of the dividend process. On the contrary, all previous models make fixed assumptions on forecasts and obtain a unique valuation. [@agosto2015variance] complement the model calculating a closed-form expression for the variance of random stock prices in a multinomial setting. The authors argue that for proper investment decisions a measure of risk should be taken into consideration. Thus applying the standard mean-variance analysis, an investor can deal with financial decisions under uncertainty. In their model, they relate the variance of stock prices with the variance of the dividend rate of growth, obtaining a measure of the stock riskiness. An extension of the Markov stock model is given in [@barbu2017novel]. The authors provide a formula for the computation of the second order moment of the fundamental price process in the case of a 2-state Markov chain, $$\label{Eq:BarbuSeries} \begin{aligned} & p^{(2)}(k)=\sum_{i=1}^{+\infty}\frac{\mathbb{E}_{k}[D^{2}(k+i)]}{r^{2i}}+ \\ & \qquad 2\sum_{i=1}^{+\infty}\sum_{j>i}\frac{\mathbb{E}_{k}[D(k+i)D(k+j)]}{r^{i+j}}=\psi_{2}(g(k)) \: d^{2}(k). \end{aligned}$$ To obtain the convergence of the series (\[Eq:BarbuSeries\]) and to satisfy the asymptotic condition in (\[Eq:BubbleCondition\]) and $$lim_{N\rightarrow +\infty} \sum_{i=1}^{N}\frac{\mathbb{E}_{k}[D(k+i)P(k+N)]}{r^{i+N}}=0,$$ the authors introduce a further assumption that avoids the presence of speculative bubbles, $$\label{Eq:BarbuCondition} {\bf A2}: \overline{g}^{(2)}:=max(p_{11}g_{1}^{2}+p_{12}g_{2}^{2},p_{21}g_{1}^{2}+p_{22}g_{2}^{2})<r^{2},$$ where $\overline{g}^{(2)}$ is the largest one step second order moment of the dividend growth rate. If assumptions A1 and A2 hold true, the pair $(\psi_{2}(g_{1}),\psi_{2}(g_{2}))$ is the unique and nonnegative solution of the linear system $$\label{Eq:Barbu2stateSystem} \begin{aligned} & \psi_{2}(g_{1})\big(r^{2}-p_{11}g_{1}^{2}\big)-\psi_{2}(g_{2})p_{12}g_{2}^{2}=p_{11}g_{1}^{2}\big(1+2\psi_{1}(g_{1})\big)+p_{12}g_{2}^{2}\big(1+2\psi_{1}(g_{2})\big)\\ & \psi_{2}(g_{2})\big(r^{2}-p_{22}g_{2}^{2}\big)-\psi_{2}(g_{1})p_{21}g_{1}^{2}=p_{21}g_{1}^{2}\big(1+2\psi_{1}(g_{1})\big)+p_{22}g_{2}^{2}\big(1+2\psi_{1}(g_{2})\big). \end{aligned}$$ To extend the results to an s-state Markov chain with state space $E=\{g_{1},g_{2}, \ldots, g_{s}\}$, assumptions A1 should be formulated as (\[Eq:GhezziConditionNstate\]) and A2 as follows: $$\label{Eq:BarbuConditionNstate} \overline{g}^{(2)}:=\max_{i\in E}\bigg(\sum_{j=1}^{s}p_{ij}g_{j}^{2}\bigg)<r^{2}.$$ In this general case, the systems (\[Eq:Ghezzi2stateSystem\]) and (\[Eq:Barbu2stateSystem\]) can be conveniently represented in matrix form, $$\label{matrixsystem_Psi} \left(\mathbf{I}_{r}-\mathbf{P} \cdot \mathbf{I}_{\mathbf{g}}\right)\cdot \mathbf{\Psi}_{1}=\mathbf{P}\cdot \mathbf{g},$$ $$\label{matrixsystem} \bigg(\mathbf{I}_{r}^{2}-\mathbf{P} \cdot \mathbf{I}_{\mathbf{g}}^{2}\bigg)\cdot \mathbf{\Psi}_{2}=\mathbf{P}\cdot \left(\big(\mathbf{g} \diamond \mathbf{g}\big)+2\mathbf{\Psi}_{1}\diamond \big(\mathbf{g} \diamond \mathbf{g}\big)\right),$$ where: - $\mathbf{\Psi}_{1}=(\psi_{1}(g_{1}), \ldots,\psi_{1}(g_{n}))^{\top}$ and $\mathbf{\Psi}_{2}=(\psi_{2}(g_{1}), \ldots, \psi_{2}(g_{n}))^{\top}$, - $\mathbf{I}_{r}:=r \mathbf{I}$, for any $ r \in \mathbb R^{*}:=\mathbb{R}- \{0\}$, and, more generally, $\mathbf{I}_{r}^n = \mathbf{I}_{r^n}$, - $ \mathbf{I}_{\mathbf{g}}=\label{eq_Ig} (I_{g}(i,j))_{i,j\in E},\,\, I_{\mathbf{g}}(i,j)= \left\{ {\begin{array}{ll} g_{i},\,\, \text{if}\,\,i=j \\ 0,\,\, \text{if}\,\,\,\,i\neq j \end{array} } \right. $, - $\mathbf{I}$ is the identity matrix of dimension $s \times s$, - $\cdot$ denotes the row by column matrix product and $\diamond $ denotes the Hadamard element by element product. The matrix $\left(\mathbf{I}_{r}-\mathbf{P} \cdot \mathbf{I}_{g}\right)$ is invertible, therefore the system (\[matrixsystem\_Psi\]) has a unique solution, $$\label{matrixsystem_Psi_sol} \mathbf{\Psi}_{1}=\left(\mathbf{I}_{r}-\mathbf{P} \cdot \mathbf{I}_{\mathbf{g}}\right)^{-1} \cdot \mathbf{P} \cdot \mathbf{g}.$$ Similarly, the matrix $\left(\mathbf{I}_{r}^{2}-\mathbf{P} \cdot \mathbf{I}_{g}^{2}\right)$ is invertible and the solution to the system (\[matrixsystem\]) is $$\begin{aligned} \mathbf{\Psi}_{2}&=& \nonumber \left(\mathbf{I}_{r}^{2}-\mathbf{P} \cdot \mathbf{I}_{\mathbf{g}}^{2}\right)^{-1} \cdot \mathbf{P}\cdot \left(\big(\mathbf{g} \diamond \mathbf{g}\big)+2\mathbf{\Psi}_{1}\diamond (\mathbf{g} \diamond \mathbf{g})\right)\\ &=& \label{matrixsystem_Psi2_sol}\left(\mathbf{I}_{r^{2}}-\mathbf{P} \cdot \mathbf{I}_{\mathbf{g}^{2}}\right)^{-1} \cdot \mathbf{P}\cdot \left(\big(\mathbf{g} \diamond \mathbf{g}\big)+2\mathbf{\Psi}_{1}\diamond \big(\mathbf{g} \diamond \mathbf{g}\big)\right), \end{aligned}$$ Relation (\[matrixsystem\_Psi2\_sol\]) represents an explicit formula for the second-order price-dividend ratio, that multiplied by $d^{2}(t)$ results in the second moment of the price process that is expressed in function of the model parameters $\mathbf{P}$ and $\mathbf{g}$. [@barbu2017novel] completed the Markov stock model framework developing non parametric statistical techniques for the inferential analysis of the model where they propose estimators of price, risk and forecasted prices. For each estimator they demonstrate that they are strongly consistent and that, after proper centralisation and normalisation, they converge in distribution to normal random variables. Finally, they give the interval estimators. A further generalisation of [@ghezzi2003stock] is available in [@d2013semi]. The author models the dividend growth rate as a semi-Markov chain. In this setting, prices become duration dependent. Therefore, they are influenced by the current state of the dividend growth process and by the elapsed time in the state. The same author proposes another extension of the model describing the dividend growth series via a continuous state space semi-Markov model [@d2017stochastic]. Multivariate Markov chain stock model {#Sec:Multivariate} ------------------------------------- The previous analysis of the dividend discount model focused on the valuation of a single firm based on its dividend process. In this section, we analyse the problem of valuating multiple stocks when they constitute a financial portfolio. When dealing with more than one price series, it is important to consider the possible dependencies that characterise the pool of stocks. In a recent paper, [@agosto2018stochastic] compute the covariance between two stocks that may be held in a portfolio. They consider a Markov chain with state space equal to the set of possible couples of the growth-dividend values for both stocks. However, this strategy cannot be easily implemented in real applications, especially when we introduce dependencies between more than two stocks as the number of parameters to estimate increase drastically. [@dmultivariate] propose and extension of the Markov stock model to a multivariate setting, computing the first and the second order price-dividend ratios. Moreover, the authors provide a formula for the computation of the variances and covariances between stocks in a portfolio. The model belongs to the class of mixture transition distribution models originated by [@raftery1985model] in a high order Markov chain setting and further extended in [@ching2006markov] to a multivariate Markov chain setting. This approach permits to overcome the limitations of [@agosto2018stochastic] because it reduces the number of parameters to estimate. With a portfolio of multiple stocks, $\alpha=1,2,\hdots,\gamma$, the dividend series expressed in (\[Eq:GhezziDividend\]) becomes $$\label{Eq:MultivariateDividendSeries} D^{(\alpha)}(k+1)=G^{(\alpha)}(k+1)\cdot D^{(\alpha)}(k),$$ where $\{G^{(\alpha)}\}_{k\in \mathbb{N}}$ is the growth-dividend random process for stock $\alpha$, and the multivariate Markov chain model follows the relationship, $$\label{Eq:MTDmodel} \mathbf{A}^{(\alpha)}(k+1)=\sum_{\beta=1}^{\gamma}\mathbf{A}^{(\beta)}(k)\cdot \lambda_{\beta,\alpha}\cdot \mathbf{P}^{(\beta,\alpha)},$$ where: - $\mathbf{A}^{\alpha}(k):=[A_{1}^{(\alpha)},\hdots,A_{m}^{(\alpha)}]$ is a probability distribution vector with $A_{i}^{(\alpha)}(k):=\mathbb{P}[G^{(\alpha)}(k)=i]$ being the probability of growth-dividend of stock $\alpha$ to be at time $k$ in state $i$, - $\lambda_{\beta,\alpha}\in [0,1]$, $\sum_{\beta=1}^{\gamma}=1$, - $\mathbf{P}^{(\beta,\alpha)}$ is the transition probability matrix of stock $\alpha$ given the state occupied one time step before by stock $\beta$, i.e. $$\label{Eq:MultiTransitionProb} \mathbf{P}^{(\beta,\alpha)}_{i,j}=\mathbb{P}[G^{(\alpha)}(k+1)=j \mid G^{(\beta)}(k)=i].$$ According to equation (\[Eq:MTDmodel\]) the probability distribution function of the growth-dividend process at time $k+1$ for the stock $\alpha$ depends on the state of the growth-dividend process of the same stock at time $k$, and, at the same time, on the set of states visited by each stock in the portfolio at time $k$. To extend the model to a multivariate setting, the price process series (\[Eq:GhezziValue\]) and (\[Eq:BarbuSeries\]) can be rewritten as, $$\label{Eq:MultivariateSeries1} \begin{aligned} p^{(\alpha)}(\bm{g}(k))&=\sum_{i=1}^{+\infty}\frac{\mathbb{E}_{(k)}[D^{(\alpha)}(k+i)]}{r_{\alpha}^{i}}\\ &=\sum_{i=1}^{+\infty}\Big(\frac{\mathbb{E}_{(k)}[\prod_{j=1}^{i}G^{(\alpha)}(k+j)]}{r_{\alpha}^{i}}\Big)d^{(\alpha)}(k), \end{aligned}$$ $$\label{Eq:MultivariateSeries2} \begin{aligned} p_{2}^{(\alpha,\beta)}(k)&:=\sum_{i=1}^{+\infty}\mathbb{E}_{(k)}\Big[\frac{D^{(\alpha)}(k+i)D^{(\beta)}(k+i)}{r_{\alpha}^{i}\cdot r_{\beta}^{i}}\Big]\\ & +\sum_{i=1}^{+\infty}\sum_{j>i}\mathbb{E}_{(k)}\Big[\frac{D^{(\alpha)}(k+i)D^{(\beta)}(k+j)}{r_{\alpha}^{i}\cdot r_{\beta}^{j}}\Big]\\ & + \sum_{i=1}^{+\infty}\sum_{j>i}\mathbb{E}_{(k)}\Big[\frac{D^{(\alpha)}(k+j)D^{(\beta)}(k+i)}{r_{\alpha}^{j}\cdot r_{\beta}^{j}}\Big], \end{aligned}$$ respectively. Equation (\[Eq:MultivariateSeries2\]) represents the fundamental formula of the price-product and reduces to the second order moment of the price process when considering the same price series, $\alpha=\beta$. To guarantee the convergence of the series (\[Eq:MultivariateSeries1\]) and (\[Eq:MultivariateSeries2\]) in the multivariate setting, [@dmultivariate] extend the transversality conditions in (\[Eq:GhezziConditionNstate\]) and (\[Eq:BarbuConditionNstate\]), $$\begin{aligned} & \overline{g}^{(\alpha;1)}:=\max_{\bm{e}^{(1)},\hdots,\bm{e}^{(\gamma)}}\Bigg(\sum_{j=1}^{m}\sum_{\beta=1}^{\gamma}\sum_{h=1}^{m}e_{h}^{(\beta)}\lambda_{\beta,\alpha}\mathbf{P}_{h,j}^{(\beta,\alpha)}g_{j}\Bigg)<r_{\alpha}, \label{Eq:MultivariateConditions1}\\ & \overline{g}^{(\alpha;2)}:=\max_{\bm{e}^{(1)},\hdots,\bm{e}^{(\gamma)}}\Bigg(\sum_{j=1}^{m}\sum_{\beta=1}^{\gamma}\sum_{h=1}^{m}e_{h}^{(\beta)}\lambda_{\beta,\alpha}\mathbf{P}_{h,j}^{(\beta,\alpha)}(g_{j})^{2}\Bigg)<r_{\alpha}^{2}. \label{Eq:MultivariateConditions2} \end{aligned}$$ If assumptions in (\[Eq:MultivariateConditions1\]) holds true, then the first order price-dividend ratio, $\psi_{1}^{(\alpha)}(\bm{g}(k))$, can be computed as a linear system of $m^{\gamma}$ equations in $m^{\gamma}$ unknown that admits a unique solution, $$\label{Eq:MultivariateFirstOrder} \begin{aligned} \psi_{1}^{(\alpha)}(g_{a_{1}}^{(1)},\ldots,g_{a_{\gamma}}^{(\gamma)})=\frac{1}{r_{\alpha}}\Big\{\sum_{j_{\alpha}=1}^{m}\sum_{\beta=1}^{\gamma}\sum_{h=1}^{m}e_{h}^{(\beta)}(k)\lambda_{\beta,\alpha}\mathbf{P}_{h,j_{\alpha}}^{(\beta,\alpha)}g_{j_{\alpha}}^{(\alpha)}+\\ \sum_{j_{1},\hdots,j_{\gamma}=1}^{m}\psi_{1}^{(\alpha)}(g_{j_{1}}^{(1)},\hdots,g_{j_{\gamma}}^{(\gamma)})\cdot g_{j_{\alpha}}^{(\alpha)}\cdot \prod_{f=1}^{\gamma}\sum_{w=1}^{\gamma}\sum_{c=1}^{m}e_{c}^{(w)}(k)\lambda_{w,f}\mathbf{P}_{c,j_{f}}^{(w,f)}\Big\}. \end{aligned}$$ Correspondingly, if assumptions in (\[Eq:MultivariateConditions1\]) and (\[Eq:MultivariateConditions2\]) hold true, then the second order price-dividend ratio, $\psi_{2}^{(\alpha)}(\bm{g}(k))$, can be computed as a linear system of $m^{\gamma}$ equations in $m^{\gamma}$ unknown that admits a unique solution, $$\label{Eq:MultivariateSecondOrder} \begin{aligned} & r_{\alpha}^{2}\psi_{2}^{(\alpha)}(g_{a_{1}}^{(1)},\ldots,g_{a_{\gamma}}^{(\gamma)})-\\ & \sum_{j_{1},\hdots,j_{\gamma}=1}^{m}\psi_{2}^{(\alpha)}(g_{j_{1}}^{(1)},\hdots,g_{j_{\gamma}}^{(\gamma)})(g_{j_{\alpha}}^{(\alpha)})^{2}\big(\prod_{f=1}^{\gamma}\sum_{w=1}^{\gamma}\sum_{c=1}^{m}e_{c}^{(w)}(k)\lambda_{w,f}\mathbf{P}_{c,j_{f}}^{(w,f)}\big)\\ & =2\sum_{j_{1},\hdots,j_{\gamma}=1}^{m}\psi_{1}^{(\alpha)}(g_{j_{1}}^{(1)},\hdots,g_{j_{\gamma}}^{(\gamma)})\cdot (g_{j_{\alpha}}^{(\alpha)})^{2}\cdot \big(\prod_{f=1}^{\gamma}\sum_{w=1}^{\gamma}\sum_{c=1}^{m}e_{c}^{(w)}(k)\lambda_{w,f}\mathbf{P}_{c,j_{f}}^{(w,f)}\big)\\ & + \sum_{j=1}^{m}\sum_{\beta=1}^{\gamma}\sum_{h=1}^{m}e_{h}^{(\beta)}(k)\lambda_{\beta,\alpha}\mathbf{P}_{h,j}^{(\beta,\alpha)}(g_{j}^{(\alpha)})^{2}. \end{aligned}$$ The solutions of the first and second order price-dividend ratio in (\[Eq:MultivariateFirstOrder\]) and (\[Eq:MultivariateSecondOrder\]) present a different price-dividend ratio attached to each combination of states of the growth-process of each stock. Finally, considering the possible correlation between the stocks and holding assumptions in (\[Eq:MultivariateConditions1\]) and (\[Eq:MultivariateConditions2\]), it is possible to compute the product price-dividend ratio, $\psi_{2}^{(\alpha;\beta)}(\bm{g}(k))$, $$\label{Eq:MultivariatePriceProduct} \begin{aligned} & r_{\alpha}r_{\beta}\psi_{2}^{(\alpha,\beta)}(g_{a_{1}}^{(1)},\ldots,g_{a_{\gamma}}^{(\gamma)})=\sum_{j_{\alpha},j_{\beta}=1}^{m}g_{j_{\alpha}}^{(\alpha)}g_{j_{\beta}}^{(\beta)}\big(\prod_{f\in \{\alpha,\beta\}}\sum_{w=1}^{\gamma}\sum_{c=1}^{m}e_{c}^{(w)}(k)\lambda_{w,f}\mathbf{P}_{c,j_{f}}^{(w,f)}\big)\\ & +\!\!\!\!\!\sum_{j_{1},\hdots,j_{\gamma}=1}^{m}\!\!\!\psi_{1}^{(\beta)}(g_{j_{1}}^{(1)},\hdots,g_{j_{\gamma}}^{(\gamma)})(g_{j_{\alpha}}^{(\alpha)})(g_{j_{\beta}}^{(\beta)})\big(\prod_{f=1}^{\gamma}\sum_{w=1}^{\gamma}\sum_{c=1}^{m}e_{c}^{(w)}(k)\lambda_{w,f}\mathbf{P}_{c,j_{f}}^{(w,f)}\big)\\ & +\!\!\!\!\!\sum_{j_{1},\hdots,j_{\gamma}=1}^{m}\!\!\!\psi_{1}^{(\alpha)}(g_{j_{1}}^{(1)},\hdots,g_{j_{\gamma}}^{(\gamma)})(g_{j_{\alpha}}^{(\alpha)})(g_{j_{\beta}}^{(\beta)})\big(\prod_{f=1}^{\gamma}\sum_{w=1}^{\gamma}\sum_{c=1}^{m}e_{c}^{(w)}(k)\lambda_{w,f}\mathbf{P}_{c,j_{f}}^{(w,f)}\big)\\ & +\!\!\!\!\!\sum_{j_{1},\hdots,j_{\gamma}=1}^{m}\!\!\!(g_{j}^{(\alpha)})(g_{j}^{(\beta)})\psi_{1}^{(\alpha)}(g_{j_{1}}^{(1)},\hdots,g_{j_{\gamma}}^{(\gamma)})\psi_{1}^{(\beta)}(g_{j_{1}},\hdots,g_{j_{\gamma}})\big(\prod_{f=1}^{\gamma}\sum_{w=1}^{\gamma}\sum_{c=1}^{m}e_{c}^{(w)}(k)\lambda_{w,f}\mathbf{P}_{c,j_{f}}^{(w,f)}\big). \end{aligned}$$ Knowing the product price-dividend ratio for any couple $(\alpha, \beta)$ of stocks, it is simple to compute the covariance function between the prices of two stocks: $$\label{Eq:Covariance} \begin{aligned} &Cov(\mathcal{P}^{(\alpha)}(\bm{g}(k)),\mathcal{P}^{(\beta)}(\bm{g}(k)))\\ &=\mathbb{E}_{(k)}[\mathcal{P}^{(\alpha)}(\bm{g}(k))\cdot \mathcal{P}^{(\beta)}(\bm{g}(k))]-\mathbb{E}_{(k)}[\mathcal{P}^{(\alpha)}(\bm{g}(k))]\cdot \mathbb{E}_{(k)}[\mathcal{P}^{(\beta)}(\bm{g}(k))]\\ & =d^{(\alpha)}(k)d^{(\beta)}(k)\bigg(\psi_{2}^{(\alpha,\beta)}(\bm{g}(k))-\psi_{1}^{(\alpha)}(\bm{g}(k))\psi_{1}^{(\beta)}(\bm{g}(k))\bigg). \end{aligned}$$ The authors apply the model to a portfolio of three US stock with a stable dividend policy with a long history and compare results with other valuation models. Finally, they show how to obtain the risk of the portfolio for different combinations of the stocks. Conclusion {#Sec:Conclusion} ========== This Chapter presented a review of the dividend discount model from its basic formulation to more recent and advanced stochastic models based on the Markov chain modelling of the dividend process. As the fundamental valuation of the firms represents an important function in the financial markets, especially for long-term investments, the Markov stock model clearly show some advantages over the Gordon model and its extensions. In particular, the Markov stock model permits to obtain a different valuation depending on the state of the growth-dividend process, or on a combination of the states of the various series in the multivariate case. However, the Markov stock model presents some limitations, that are shared with the other cited models. First, the valuation is based on the dividend process, therefore it is only applicable to companies that pay dividend and with a long history of payments. Second, the discounting factor, $k_e$, is considered constant, thus it is not a realistic assumptions when considering the very long timeframe. Future extensions of the Markov stock model could consider the inclusion of some restrictions on the estimation of the transition probability matrix to reduce the number of parameters to estimate and permit the use of shorter dividend series. Moreover, the cost of equity could be modelled as a stochastic process interdependent with the dividend process. Finally, the model could be extended to companies without a dividend policy, perhaps using the earnings or similar cash flows.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We explicitly produce symplectic genus-$3$ Lefschetz pencils (with base points), whose total spaces are homeomorphic but not diffeomorphic to rational surfaces ${{\mathbb{CP}}{}^{2}}\# p \, {\overline{\mathbb{CP}}{}^{2}}$ for $p= 7, 8, 9$. We then give a new construction of an infinite family of symplectic Calabi-Yau surfaces with first Betti number $b_1=2,3$, along with a surface with $b_1=4$ homeomorphic to the $4$-torus. These are presented as the total spaces of symplectic genus-$3$ Lefschetz pencils we construct via new positive factorizations in the mapping class group of a genus-$3$ surface. Our techniques in addition allow us to answer in the negative a question of Korkmaz regarding the upper bound on $b_1$ of a genus-$g$ fibration.' address: 'Department of Mathematics and Statistics, University of Massachusetts, Amherst, MA 01003-9305, USA' author: - 'R. İnanç Baykur' title: \| Small symplectic Calabi-Yau surfaces\ and exotic $4$-manifolds via genus-$3$ pencils --- Introduction ============ Since the advent of Gauge theory, many new construction techniques, such as rational blowdowns, generalized fiber sum, knot surgery and Luttinger surgery, have been introduced and successfully employed to produce symplectic $4$-manifolds homeomorphic but not diffeomorphic to rational surfaces ${{\mathbb{CP}}{}^{2}}\# p \, {\overline{\mathbb{CP}}{}^{2}}$, for $p \geq 2$. (e.g. [@ABP; @AP; @BK; @Donaldson87; @FSrationalblowdown; @FSKnotsurgery; @FSPinwheels; @FSReverseEngineering; @FriedmanMorgan; @Gompf; @Kotschick; @ParkD; @ParkJ; @StipsiczSzabo].) All these $4$-manifolds should admit symplectic Lefschetz pencils (which, by definition, always have base points) by the ground-breaking work of Donaldson [@Donaldson]. Despite this a priori existence result, no explicit Lefschetz pencils on these small (as in small second homology) exotic $4$-manifolds were known up to date. The main goal of this article is to introduce novel ways of constructing small positive factorizations (as in small number of positive Dehn twists) of the boundary multi-twist in the mapping class group $\Gamma_g^m$ of a compact genus-$g$ surface with $m>0$ boundary components, which correspond to monodromy factorizations of genus-$g$ Lefschetz pencils on small symplectic $4$-manifolds. A first application yields the first explicit examples of pencils on exotic rational surfaces: \[main1\] There are symplectic genus-$3$ Lefschetz pencils $(X_i, f_i$) whose total spaces are homeomorphic but not diffeomorphic to ${{\mathbb{CP}}{}^{2}}\# (6+i) {\overline{\mathbb{CP}}{}^{2}}$ for $i=1, 2, 3$. They realize the smallest possible genera pencils in these homeomorphism classes. A key result leading to Theorem \[main1\] is that ${{\mathbb{CP}}{}^{2}}\# \, p {\overline{\mathbb{CP}}{}^{2}}$, for $p \leq 9$, does not admit a genus $g \geq 2$ Lefschetz fibration or a genus-$g$ Lefschetz pencil with $m < 2g-2$ base points (Lemma \[notrational\]), which provides a new criterion to argue that the total spaces of such pencils/fibrations like the ones in Theorem \[main1\] are exotic. We moreover observe that $g=3$ is the smallest possible genus for a Lefschetz pencil on these exotic $4$-manifolds (Remark \[smallest\]). We hope that the additional information on these exotic symplectic $4$-manifolds we construct, namely the existence of genus-$3$ pencils on them, will help with understanding whether there are distinct minimal symplectic $4$-manifolds in the homeomorphism classes ${{\mathbb{CP}}{}^{2}}\# \, p {\overline{\mathbb{CP}}{}^{2}}$, for $p < 9$ (a question which is still open as of today; see Remark \[distinctsymp\]). To produce the positive factorizations for the pencils in Theorem \[main1\], we employ a new way of deriving a positive factorization ${\widetilde}{W}$ in $\Gamma_g^m$, $m \geq 0$ from a collection of positive factorizations $W_1, \ldots, W_k$ of boundary twists in $\Gamma_h^n$, with $h< g$ and $n>0$. For $(X,f)$ the Lefschetz pencil/fibration corresponding to ${\widetilde}{W}$ and $(X_i, f_i)$ to $W_i$, we say $(X,f)$ is obtained by breeding $(X_1,f_1), \ldots, (X_k, f_k)$. This involves embedding the latter into $\Gamma_g^m$ as factorizations for achiral Lefschetz pencils/fibrations, and then canceling matching pairs of positive and negative Dehn twists. The idea for this construction scheme comes from the smallest hyperelliptic genus-$3$ Lefschetz fibration Mustafa Korkmaz and the author produced in [@BaykurKorkmazGenus3]. A careful variation of our construction of $(X_i, f_i)$ yield to our next theorem: \[main2\] For any pair of non-negative integers $M=(m_1, m_2)$, there exists a symplectic genus-$3$ Lefschetz pencil $(X_{i,M}, f_{i,M})$ with $c_1^2(X_{i,M})=3-i$, , and $\pi_1(X_{i,M})= ({\mathbb{Z}}\, / {m_1 \, {\mathbb{Z}}}) \oplus ({\mathbb{Z}}\, / {m_2 \, {\mathbb{Z}}})$, for each $i=1,2,3$. Infinitely many of these pencils have total spaces homotopy inequivalent to a complex surface. In contrast, I know of only one explicit example of a non-holomorphic genus-$3$ Lefschetz pencil in the literature, due to Ivan Smith. ([@SmithGenus3]\[Theorem 1.3\]; see also Remark \[nonholompencil\].) Our construction technique yielding the pencils in Theorem \[main2\] can be easily generalized to obtain many other non-holomorphic pencils. The second part of our article will deal with new constructions of symplectic Calabi-Yau surfaces. Recall that a symplectic $4$-manifold $(X, \omega)$ is called a symplectic Calabi-Yau surface (SCY) if it has trivial canonical class $K_{X} \in H^2(X ; {\mathbb{Z}})$, in obvious analogy with complex Calabi-Yau surfaces. As shown by Tian-Jun Li, any minimal symplectic $4$-manifold with Kodaira dimension zero is a symplectic Calabi-Yau surface or has trivial canonical class [@LiSCY]. Up to date, the only known examples of symplectic Calabi-Yau surfaces that are not diffeomorphic to complex Calabi-Yau surfaces are torus bundles over tori (at least virtually)[^1] , leading to the question [@LiSCY; @Donaldson_SCY]: \[SCYclassification\] Is every symplectic Calabi-Yau surface with $b_1> 0$ diffeomorphic to an oriented torus bundle over a torus? Works of Tian-Jun Li and Stefan Bauer independently established that any symplectic Calabi-Yau surface with $b_1 >0$ has the same rational homology as a torus bundle (and that of ${{\rm K3}}$ and the Enriques surface when $b_1=0$) [@LiSCY; @LiQuaternionic; @Bauer]. All torus bundles over tori have solvmanifold fundamental groups [@Hillman], and notably, it was shown by Stefan Friedl and Stefano Vidussi that the homeomorphism type of a symplectic Calabi-Yau surface with a solvmanifold fundamental group is uniquely determined [@FriedlVidussi]. As stated by Tian-Jun Li [@LiSCYSurvey], what seems to commonly provide a posteriori reasoning for a positive answer to the Question \[SCYclassification\] is the lack of any new constructions of symplectic $4$-manifolds of Kodaira dimension zero. Surgical operations widely used to construct most interesting symplectic $4$-manifolds in the past, such as Luttinger surgery, generalized fiber sums, knot surgery, or simplest rational blow-down operations, do not produce any new SCYs [@HoLi; @LiSCYSurvey; @UsherKodaira; @Dorfmeister]. We will present a new construction of symplectic Calabi-Yau surfaces with via Lefschetz pencils, where the genus of the pencil $(X,f)$ versus $c_1^2(X)$, building on the work of Cliff Taubes, will determine the Kodaira dimension of $X$ (Lemma \[SCYpencil\]; also [@SatoKodaira]). Once again, we will focus on generating the smallest genus examples: genus-$3$ pencils. To produce our examples, we will use a version of the aforementioned breeding technique for producing new positive factorizations in $\Gamma_3^m$, which we will call inbreeding, where this time copies of the same positive factorization in $\Gamma_2^n$ will be embedded into $\Gamma_3^m$ by identifying some of the boundary components of the fiber $\Sigma_2^n$ and introducing extra $\pi_1$-generators this way. The result is a family of positive factorizations $$W_{\phi}= t_{\phi(B_{0})} t_{\phi(B_{1})} t_{\phi(B_{2})} \, t_{\phi(A_{0})} t_{\phi(A_{1})} t_{\phi(A_{2})} \, t_{B'_{0}} t_{B'_{1}} t_{B'_{2}} \, t_{A'_{0}} t_{A'_{1}} t_{A'_{2}} = 1$$ in $\Gamma_3$, where $\phi \in \Gamma_3$ is parametrized by pairs of mapping classes $\phi_1, \phi_2 \in \Gamma_1^2$, and the curves $B_{j}, A_{j}$ are as shown in Figure \[SCYcurves\]. Here we inbreed the monodromy factorizations of Matsumoto’s well-known genus-$2$ Lefschetz fibration [@Matsumoto], and using further lifts of it obtained by Noriyuki Hamada [@Hamada], we can even derive an explicit lift of $W_{\phi}$, which is a positive factorization of the boundary multi-twist $\Delta = t_{\partial_1} t_{\partial_2} t_{\partial_3} t_{\partial4}$ in $\Gamma_3^4$. Letting $(X_{\phi}, f_{\phi})$ denote the corresponding symplectic genus-$3$ Lefschetz pencil, we obtain (Theorem \[SCYfamily\]): \[main3\] Positive factorizations $W_{\phi}$ prescribe symplectic genus-$3$ Lefschetz pencils on a family of symplectic Calabi-Yau surfaces $X_{\phi}$ in all rational homology classes of torus bundles over tori. This provides a first step towards analyzing the monodromies of pencils on a rich family of SCYs, as proposed by Simon Donaldson in [@Donaldson_MCG]\[Problem 5\]. The first question that arises here is whether or not all $X_{\phi}$ are diffeomorphic to torus bundles over tori, in reference to Question \[SCYclassification\]. Torus bundles over tori with sections do admit symplectic genus-$3$ pencils [@SmithTorus]; however the uncanny freedom we have in the choice of $\phi$ appears to exceed that of a torus bundle with a section (see the discussion in Remark \[CompareSmith\]). The explicit nature of our construction allows us to derive explicit presentations for $\pi_1(X_{\phi})$, which for instance yields an $X$ (when $\phi$ is trivial) that is –at least– homeomorphic to the $4$-torus. We will discuss this example in complete detail in Section \[Sec:4torus\]. It is not clear to us at this point whether or not all $\pi_1(X_{\phi})$ are $4$-dimensional solvmanifold groups (some of them certainly are), which we hope to understand in future work. On the other hand, we observe that all $X_{\phi}$ are obtained from $X$, our SCY homeomorphic to the $4$-torus, via fibered Luttinger surgeries along Lagrangian tori or Klein bottles (Remark \[CompareHoLi\]). So if our family of $X_{\phi}$ is equal to the family of torus bundles over tori, it immediately confirms an improved version[^2] of an interesting conjecture by Tian-Jun Li and Chung-I Ho [@HoLi]\[Conjecture 4.9\]: Any smooth oriented torus bundle $X$ possesses a symplectic structure $\omega$ such that $(X, \omega)$ can be obtained by applying Luttinger surgeries to $(T^4, \omega_{\rm{std}})$ (Remark \[CompareHoLi\]). Finally, Theorem \[main3\] implies that we have symplectic genus-$3$ pencils $(X_\phi, f_{\phi})$ with $H_1(X_{\phi}) \cong {\mathbb{Z}}^2 \oplus ({\mathbb{Z}}\, / {m_1 \, {\mathbb{Z}}}) \oplus ({\mathbb{Z}}\, / {m_2 \, {\mathbb{Z}}})$ for any $m_1, m_2 \in {\mathbb{N}}$. This yields further (infinitely many) examples of non-holomorphic genus-$3$ pencils, in the same fashion as in our earlier examples in Theorem \[main2\]. The techniques we employed to produce the genus-$3$ pencils in Theorems \[main1\], and \[main3\] can be easily adapted to produce higher genera pencils on similar $4$-manifolds; we can in fact derive many of these by repeatedly breeding/inbreeding the same collection of positive factorizations used in this paper. This direction will be explored elsewhere. We will however discuss one immediate generalization of our construction of a genus-$3$ pencil to higher genera. It will allow us to address a problem raised by Mustafa Korkmaz [@KorkmazProblems]\[Problem 2.7\], which we re-express here for pencils (which are even more constrained than fibrations): Is $b_1(X) \leq g$ for any non-trivial genus-$g$ Lefschetz pencil $(X,f)$? Is there an upper bound on $b_1(X)$ linear in $g$, and sharper than $2g-1$? In Section \[Sec:bound\], generalizing our construction of a genus-$3$ pencil on $X$ homeomorphic to the $4$-torus, we will produce genus-$g$ Lefschetz pencils violating the above bound for every odd $g> 1$. Though not much; we will indeed revamp Korkmaz’s question by raising the upper bound to $g+1$ (Question \[b1RefinedQuestion\]). Acknowledgements. We would like to thank Weimin Chen, Bob Gompf, Jonathan Hillman, Tian-Jun Li, and Stefano Vidussi for helpful conversations related to our constructions in this paper. Thanks to Noriyuki Hamada for telling us about his preprint on the sections of the Matsumoto fibration. Our results on small symplectic Calabi-Yau surfaces via Lefschetz fibrations with exceptional sections were first presented at the Great Lakes Geometry Conference in Ann Arbor in March 2015; we would like to thank the organizers for motivating discussions. Last but not least, we are grateful to Mustafa Korkmaz for his interest in our work and numerous stimulating discussions. The author was partially supported by the NSF Grant DMS-$1510395$. Preliminaries ============= Here we quickly review the definitions and basic properties of Lefschetz pencils and fibrations, Dehn twist factorizations in mapping class groups of surfaces, and symplectic Calabi-Yau surfaces. The reader can turn to [@GompfStipsicz; @LiKodairaSurvey; @BaykurKorkmaz] for more details. Lefschetz pencils and fibrations -------------------------------- A Lefschetz pencil on a closed, smooth, oriented $4$-manifold $X$ is a smooth surjective map $f: X \setminus \{ b_j \} \to S^2$, defined on the complement of a non-empty finite collection of points $\{b_j \}$, such that around every base point $b_j$ and critical point $p_i$ there are local complex coordinates (compatible with the orientations on $X$ and $S^2$) with respect to which the map $f$ takes the form $(z_1,z_2) \mapsto z_1/z_2$ and $(z_1, z_2) \mapsto z_1 z_2$, respectively. A Lefschetz fibration is defined similarly for $\{ b_j \} = \emptyset$. Blowing-up all the base points $b_j$ in a pencil $(X,f)$, one obtains a Lefschetz fibration $({\widetilde}{X},{\widetilde}{f})$ with disjoint $(-1)$-sphere sections $S_j$ corresponding to $b_j$, and vice versa. We say $(X,f)$ is a genus $g$ Lefschetz pencil or fibration for $g$ the genus of a regular fiber $F$ of $f$. The fiber containing the critical point $p_i$ has a nodal singularity at $p_i$, which locally arises from shrinking a simple loop $c_i$ on $F$, called the vanishing cycle. A singular fiber of $(X,f)$ is called reducible if $c_i$ is separating. When $c_i$ is null-homotopic on $F$, one of the fiber components becomes an exceptional sphere, an embedded $2$-sphere of self-intersection $-1$, which we can blow down without changing the rest of the fibration. In this paper we use the term Lefschetz fibration only when the set of critical points $\{p_i\}$ is non-empty, i.e. when the Lefschetz fibration is non-trivial. We moreover assume that the fibration is relatively minimal, i.e. it there are no exceptional spheres contained in the fibers, and also that the points $p_i$ lie in distinct singular fibers, which can be always achieved after a small perturbation. A widely used way of constructing Lefschetz fibrations is the fiber sum operation: if $(X_i, f_i)$ are genus-$g$ Lefschetz fibrations with regular fiber $F_i$ for $i=1,2$, then their fiber sum is a genus-$g$ Lefschetz fibration $(X,f)$, where $X$ is obtained by removing a fibered tubular neighborhood of each $F_i$ from $X_i$ and then identifying the resulting boundaries via a fiber-preserving, orientation-reversing self-diffeomorphism, and where $f$ restricts to $f_i$ on each $X_i \setminus \nu F_i$. A Lefschetz fibration $(X,f)$ that can be expressed in this way is called decomposable, and each $(X_i, f_i)$ are called its summands. Allowing local models $(z_1, z_2) \mapsto z_1 \bar{z}_2$ around the critical points $p_i$, called negative critical points, we can extend all of the above to achiral Lefschetz pencils and fibrations. Positive factorizations ----------------------- Let $\Sigma_g^m$ denote a compact, connected, oriented surface genus $g$ with $m$ boundary components. We denote by $\Gamma_g^m$ its mapping class group; the group composed of orientation-preserving self homeomorphisms of $\Sigma_g^m$ which restrict to the identity along $\partial \Sigma_g^m$, modulo isotopies that also restrict to the identity along $\partial \Sigma_g^m$. We write $\Sigma_g = \Sigma_g^0$, and $\Gamma_g=\Gamma_g^0$ for simplicity. Denote by $t_c \in \Sigma_g^m$ the positive (right-handed) Dehn twist along the simple closed curve $c \subset \Sigma_g^m$. The inverse $t^{-1}_c$ denotes the negative (left-handed) Dehn twist along $c$. Let $\{c_i\}$ be a non-empty collection of simple closed curves on $\Sigma_g^m$, which do not become null-homotopic when $\partial \Sigma_g^m$ is capped off by disks, and let $\{\delta_j\}$ be a collection of curves parallel to distinct boundary components of $\Sigma_g^m$. If the relation $$\label{factorization} t_{c_l} \cdots t_{c_2} t_{c_1} = t_{\delta_1} \cdots t_{\delta_m} \, \,$$ holds in $\Gamma_g^m$, we call the word $W$ on the left-hand side a positive factorization of the boundary multi-twist $\Delta = t_{\delta_1} \cdots t_{\delta_m}$ in $\Gamma_g^n$. (We will also use $\partial_i$ instead of $\delta_i$ to denote boundary components at times when there are several surfaces with boundaries involved in our discussion.) Capping off $\partial \Sigma_g^m$ induces a homomorphism $\Gamma_g^m \to \Gamma_g$, under which $W$ maps to a similar positive factorization of the identity element $1 \in \Gamma_g$. The positive factorization in (\[factorization\]) gives rise to a genus-$g$ Lefschetz fibration $({\widetilde}{X},{\widetilde}{f})$ with $m$ disjoint $(-1)$-sections $S_j$, and equivalently a genus-$g$ Lefschetz pencil $(X,f)$ with $m$ base points. Identifying the regular fiber $F$ with $\Sigma_g$, we can view the vanishing cycles as $c_i$. In fact, every Lefschetz fibration or pencil can be described by a positive factorization $W$ as in (\[factorization\]) [@GompfStipsicz; @Kas; @Matsumoto], which is called the monodromy factorization of ${\widetilde}{f}$ or $f$. Let $W$ be a positive factorization of the form $W= P P'$ in $\Gamma_g^m$, where $P, P'$ are products of positive Dehn twists along curves which do not become null-homotopic when $\partial \Sigma_g^m$ is capped off by disks. If $P= \prod t_{c_i}$, as a mapping class, commutes with some $\phi \in \Gamma_g^m$, we can produce a new positive factorization $W_{\phi}= P^{\phi} P'$, where $P^{\phi}$ denotes the conjugate factorization $\phi P \phi^{-1} = \prod (\phi \, t_{c_i} \, \phi^{-1}) = \prod t_{\phi(c_i)} \, . $ In this case, we say $W_{\phi}$ is obtained from $W$ by a partial conjugation $\phi$ along the factor $P$. When $P, P \in \Gamma_g$ are positive factorizations, $W=P P'$ is another positive factorization, Lefschetz fibration corresponding to which is a fiber sum of the fibrations corresponding to $W'$ and $W''$. Allowing negative Dehn twists, which correspond to negative critical points, we can more generally work with factorizations for achiral Lefschetz fibrations and pencils. All of the above extend to this general case. Symplectic $4$-manifolds and the Kodaira dimension -------------------------------------------------- It was shown by Donaldson that every symplectic $4$-manifold $(X, \omega)$ admits a symplectic Lefschetz pencil whose fibers are symplectic with respect to $\omega$ [@Donaldson]. Conversely, generalizing a construction of Thurston, Gompf showed that the total space of a (nontrivial) Lefschetz fibration, and in particular blow-up of any pencil, always admits a symplectic form $\omega$ with respect to which all regular fibers and any preselected collection of disjoint sections are symplectic [@GompfStipsicz]. Whenever we take a symplectic form $\omega$ on a Lefschetz pencil or fibration $(X,f)$ we will assume that it is of Thurston-Gompf type, with respect to which any explicitly discussed sections of $f$ will always be assumed to be symplectic as well. The Kodaira dimension for projective surfaces can be extended to symplectic $4$-manifolds as follows: Let $K_{X_{\text{min}}}$ be the canonical class of a minimal model $(X_{\text{min}}, \omega_{\text{min}})$ of $(X, \omega)$. The symplectic Kodaira dimension of $(X, \omega)$, denoted by $\kappa=\kappa(X,\omega)$ is defined as $$\kappa(X,\omega)=\left\{\begin{array}{rl}-\infty& \mbox{\ \ if }K_{X_{\text{min}}}\cdot[\omega_{\text{min}}]<0 \mbox{ or } K_{X_{\text{min}}}^{2}<0 \\ 0 & \mbox{\ \ if } K_{X_{\text{min}}}\cdot[\omega_{\text{min}}]= K_{X_{\text{min}}}^{2}=0\\ 1 & \mbox{\ \ if } K_{X_{\text{min}}}\cdot[\omega_{\text{min}}]>0\mbox{ and } K_{X_{\text{min}}}^{2}=0\\2& \mbox{\ \ if } K_{X_{\text{min}}}\cdot[\omega_{\text{min}}]>0\mbox{ and } K_{X_{\text{min}}}^{2}>0\end{array}\right .$$ Remarkably, $\kappa$ is not only independent of the minimal model $(X_{\text{min}}, \omega_{\text{min}})$ but also of the chosen symplectic form $\omega$ on $X$: it is a smooth invariant of the $4$-manifold $X$ [@LiSCY]. Symplectic $4$-manifolds with $\kappa = -\infty$ are classified up to symplectomorphisms, which are precisely the rational and ruled surfaces [@LiKodairaSurvey]. Symplectic $4$-manifolds with $\kappa=0$ are the analogues of the Calabi-Yau surfaces that have torsion canonical class [@LiSCY]. It has been shown by Tian-Jun Li, and independently by Stefan Bauer [@LiSCY; @Bauer], that the rational homology type of any minimal symplectic Calabi-Yau surface is that of either a torus bundle over a torus, the ${{\rm K3}}$ surface, or the Enriques surface. Exotic symplectic rational surfaces via genus-$3$ pencils ========================================================= Here we construct explicit positive factorizations for symplectic genus-$3$ Lefschetz pencils, whose total spaces are homeomorphic but not diffeomorphic to the complex rational surfaces. Breeding genus-$3$ Lefschetz pencils from genus-$2$ pencils ----------------------------------------------------------- In [@BaykurKorkmaz], we have obtained the following relation in $\Gamma_2^1$ $$\begin{aligned} t_ e t_{x_1} t_{x_2} t_{x_3} t_d t_C t_{ x_4 } &=& t_\delta \, ,\end{aligned}$$ which can be rewritten as $$\begin{aligned} t_ e t_{x_1} t_{x_2} t_{x_3} t_d t_{B_2} t_C t_\delta^{-1} &=& 1 \, ,\end{aligned}$$ where $B_2= t_c(x_4)$. Embed this relation into $\Gamma_3^1$ so that $\delta=\partial \Gamma_2^1$ is mapped to $c'$ and the remaining curves are as in Figure \[Exoticcurves\], denoted by the same letters. (So, along with the embedding of $x_4$, the original curves on $\Gamma_2^1$ can be all seen in the obvious subsurface of $\Gamma_3^1$ in Figure \[Exoticcurves\].) We thus get the following relation in $\Gamma_3^1$: $$\begin{aligned} \label{embed1} t_ e t_{x_1} t_{x_2} t_{x_3} t_d t_{B_2} t_C t_{C'}^{-1} &=& 1 \, ,\end{aligned}$$ or $P_1 \, t_C t_{C'}^{-1} = 1$, where we set $\underline{P_1= t_ e t_{x_1} t_{x_2} t_{x_3} t_d t_{B_2}}$. Note that $P_1, t_C$ and $t_{C'}$ all commute with each other. ![Curves $C, x_1, x_2, x_3, x_4, d, e$ of the first embedding are given on the left. Curves $B_0, B_1, B_2, A_0, A_1, A_2$ of the second embedding (along with the same $C$) and $C', B'_0, B'_1, B'_2, A'_0, A'_1, A'_2$ of the third embedding are on the right. []{data-label="Exoticcurves"}](ExoticGenus3curves.eps){width="12cm"} Our second building block will be a positive factorization of the boundary multi-twist in $\Gamma_2^2$, which is a lift of the monodromy factorization of Matsumoto’s well-known genus-$2$ Lefschetz fibration [@Matsumoto]. Such a lift to $\Gamma_2^1$ was first found by Ozbagci and Stipsicz in [@OzbagciStipsicz2] and a further lift to $\Gamma_2^2$ by Korkmaz in [@Korkmaz2]. We will however work with another lift discovered by Noriyuki Hamada, which yields the following relation in $\Gamma_2^2$ [@Hamada]: $$\begin{aligned} \label{firstlift} (t_{B_0} t_{B_1} t_{B_2} t_{C} )^2 &=& t_{\delta_1} t_{\delta_2} \, ,\end{aligned}$$ where $\delta_i$ are the boundary parallel curves, and the curves $B_i$ and $C$ are as shown on the left-hand side of Figure \[Matsumoto\]. We can rewrite this relation in $\Gamma_2^2$ as $$\begin{aligned} t_{B_0} t_{B_1} t_{B_2} t_{A_0} t_{A_1} t_{A_2} t_{C}^2 t_{\delta_1}^{-1} t_{\delta_2}^{-1} &=& 1 \, ,\end{aligned}$$ where each $A_j= t_C(B_j)$ for $j=0,1,2$ (Figure \[Matsumoto\]). Let us first view this relation in $\Gamma_2^1$ by capping off one of the boundary components, say $\delta_1$. Embed the resulting relation into $\Gamma_3^1$, as induced by the same embedding of $\Gamma_2^1$ into $\Gamma_3^1$ we used above so that the boundary is mapped to $c'$, and all the other curves are as shown in Figure \[Exoticcurves\], once again denoted by the same letters for simplicity. So the following holds in $\Gamma_3^1$: $$\begin{aligned} \label{embed2} t_{B_0} t_{B_1} t_{B_2} t_{A_0} t_{A_1} t_{A_2} t_{C}^2 t_{C'}^{-1} &=& 1 \, ,\end{aligned}$$ or $P_2 \, t_{C}^2 t_{C'}^{-1}= 1$, where $\underline{P_2= t_{B_0} t_{B_1} t_{B_2} t_{A_0} t_{A_1} t_{A_2}} $. Here $P_2, t_C$ and $t_{C'}$ all commute with each other. Finally, we will embed the lat relation we had above in $\Gamma_2^2$ into $\Gamma_3^1$ using an embedding of $\Sigma_2^2$ into $\Sigma_3^1$ such that $\delta_1$ is mapped to $c$, $\delta_2$ is mapped to $\delta=\partial \Sigma_3^1$, and the remaining curves are as shown in Figure \[Exoticcurves\], this time their labels are decorated by prime. This gives a third relation in $\Gamma_3^1$: $$\begin{aligned} \label{embed3} t_{B'_0} t_{B'_1} t_{B'_2} t_{A'_0} t_{A'_1} t_{A'_2} t_{C'}^2 t_{C}^{-1} &=& t_{\delta} \, ,\end{aligned}$$ or $P'_2 \, t_{C'}^2 t_{C}^{-1}= t_{\delta} $, where $\underline{P'_2=t_{B'_0} t_{B'_1} t_{B'_2} t_{A'_0} t_{A'_1} t_{A'_2}} $. Similar to the above, $P'_2, t_C$ and $t_{C'}$ all commute with each other. Now for $\phi= t_{b_1}^{-1} t_{a_2}$ (where $b_1, a_2$ are as in Figure \[pi1generators\]), and in fact for any mapping class in $\Gamma_3^1$ that keeps $C$ and $C'$ fixed, we have $$(P_1)^{\phi} P_1 P'_2 \, t_C = (P_1)^{\phi} t_C t_{C'}^{-1} \, P_1 t_C t_{C'}^{-1} \, P'_2 \, t_{C'}^2 t_{C}^{-1} = (P_1 t_C t_{C'}^{-1})^{\phi} \, t_{\delta} =t_\delta \, ,$$ where the first equality follows from the commutativity relations noted above, and the second equality follows from the relations $(\ref{embed1})$–$(\ref{embed3})$. So $\underline{W_1=(P_1)^{\phi} P_1 P'_2 t_C}$ is a positive factorization of $t_{\delta}$ in $\Gamma_3^1$. Similarly, we get two more positive factorizations $\underline{W_2=(P_1)^{\phi} P_2 P'_2 \, t^2_C}$ and $\underline{W_3=(P_2)^{\phi} P_2 P'_2 \, t^3_C}$ of $t_{\delta}$ in $\Gamma_3^1$. Each $W_i$ prescribes a symplectic genus-$3$ Lefschetz fibration $(\widetilde{X_i}, \widetilde{f_i})$ with a $(-1)$-section, which can be blown-down to arrive at a symplectic genus-$3$ Lefschetz pencil $(X_i, f_i)$ with one base point. They are all obtained by breeding the genus-$2$ Lefschetz pencils whose monodromies we have lifted to $\Gamma_3^1$. Homeomorphism type of the pencils $(X_i, f_i)$ ---------------------------------------------- \[homeo\] We will prove the following: \[lemhomeo\] $X_i$ is homeomorphic to ${{\mathbb{CP}}{}^{2}}\# \, (6+i) \, {\overline{\mathbb{CP}}{}^{2}}$, for $i=1,2,3$. The proof will follow from Freedman’s celebrated work [@Freedman], for which we will determine the Euler characteristic, the signature, and the spin type of $X_i$, while showing that $X_i$ is simply-connected for all $i=1,2,3$. The Euler characteristic of $X_i$ is given by $$\eu(X_i)= \eu(\widetilde{X_i})-1= (4-4\cdot 3+\ell(W_i))-1=\ell(W_i)-9 \, ,$$ where $\ell(W_i)$ is the number of Dehn twists in the positive factorization $W_i$. Since $\ell(P_1)=\ell(P_2)=\ell(P'_1)=6$, we have $\ell(W_i)=18+i$, and in turn, $\eu(X_i)= 9+i$. The signature of $X_i$ can be calculated as $\sigma(X_i)=-5-i$ using Ozbagci’s algorithm [@Ozbagci]. Here is a quick argument for this calculation: we have obtained the monodromy factorization of $(\widetilde{X_1}, \widetilde{f_1})$ after canceling a pair of $t_C$ and ${t_C}^{-1}$, and two pairs of $t_{C'}$ and ${t^{-1}_{C'}}$ in the factorization $$(P_1 t_C t_{C'}^{-1})^{\phi} \, P_1 t_C t_{C'}^{-1} \, P'_2 \, t_{C'}^2 t_{C}^{-1} = 1 \,$$ in $\Gamma_3$ (where we capped off the boundary). The latter is a factorization of a genus-$3$ achiral Lefschetz fibration, which is a fiber sum of three achiral Lefschetz fibrations prescribed by $P_1 t_C t_{C'}^{-1}$. $P_1 t_C t_{C'}^{-1}$, and $P'_2 \, t_{C'}^2 t_{C}^{-1}$. Each one of these three fibrations is obtained by a section sum of an achiral genus-$2$ Lefschetz fibration and a trivial torus fibration along a section of self-intersection zero, so by Novikov additivity, its signature is equal to that of the achiral genus-$2$ fibration. The hyperelliptic signature formula for genus-$2$ fibrations [@Matsumoto; @Endo] allows us to easily calculate these as $-2$, $-2$, and $-3$, which add up to give signature $-7$ for the original genus-$3$ achiral Lefschetz fibration. The signature of $(\widetilde{X_1}, \widetilde{f_1})$ is the same as the signature of this achiral fibration, since canceling a pair of separating Dehn twists with opposite signs amounts to taking out a pair of $\pm 1$ contribution to the signature. Hence, we get $\sigma(X_1) = \widetilde{X_1}+1=-6$. By identical arguments, we get $\sigma(X_2)=-7$ and $\sigma(X_3)=-8$. Clearly each $X_i$ has an odd intersection form: this can be easily seen from the presence of the reducible fiber component split off by the separating vanishing cycle $C$ and not hit by the $(-1)$-sphere. (Alternatively, we can use Rokhlin’s theorem after showing that $\pi_1(X_i)=1$.) Lastly, we will show that $\pi_1(X_i)=\pi_1(\widetilde{X_i})=1$, by calculating the latter as the quotient of $\pi_1(\Sigma_3)$ by the subgroup normally generated by the vanishing cycles of $(\widetilde{X_i}, \widetilde{f_i})$, for each $i=1,2,3$. It suffices to find enough relations that kill the generators $a_1, b_1, a_2, b_2, a_3, b_3$ of $\pi_1(\Sigma_3)$ (which are as shown in Figure \[pi1generators\]), so we do not need to work with the full presentation. Each positive factorization $W_i$ contains the factor $P'_2 \, t_{C}$. So the following relations hold for the finite presentations we have for each $\pi_1(\widetilde{X_i})$: $$\begin{aligned} & \phantom{o} & [a_1,b_1][a_2,b_2][a_3,b_3]=1,\label{eqn:0} \\ & \phantom{o} & [a_1,b_1]=1, \label{eqn:C} \\ & \phantom{o} & a_2a_3=1, \label{eqn:B'0} \\ & \phantom{o} & a_2 \bar{b}_2 a_3 \bar{b}_3 =1, \label{eqn:B'1} \\ & \phantom{o} & b_3 b_2 =1, \label{eqn:B'2} \end{aligned}$$ where the relators $(\ref{eqn:C})$–$(\ref{eqn:B'2})$ come from the vanishing cycles $C, B'_0, B'_1, B'_2$, respectively. We have $a_2= \bar{a}_3$ from $(\ref{eqn:B'0})$ and $b_2= \bar{b}_3$ from $(\ref{eqn:B'2})$. Together with $(\ref{eqn:B'1})$, these imply $[a_2, b_2]=[a_3,b_3]=1$. Together with $(\ref{eqn:C})$, we conclude $[a_j, b_j]=1$ for every $j=1,2,3$ (and $(\ref{eqn:0})$ becomes a trivial relation). From the factor $P_1$, we get the following relations (among many others) $$\begin{aligned} & \phantom{o} & a_1 (\bar{b}_1a_2b_2)^2=1, \label{eqn:x1} \\ & \phantom{o} & a_1 \bar{b}^3_1 a_2 b_2 a_2=1, \label{eqn:x2} \\ & \phantom{o} & a_1 \bar{b}^5_1 a_2 [b_2, a_2] b_1 a_2 =1, \label{eqn:x3} \\ & \phantom{o} & b_2 b_1 [b_3, a_3] =1, \label{eqn:B2}\end{aligned}$$ induced by the vanishing cycles $x_1, x_2, x_3$ and $B_2$, respectively. Adding these to the previous relators from the factor $P'_2$, we immediately see that $[a_3,b_3]=1$ and $(\ref{eqn:B2})$ imply $b_1=\bar{b}_2$. So $(\ref{eqn:x1})$ implies that $a_1$ can be generated by $a_2$ and $b_2=\bar{b}_1$. From the factor $P_2$ we get the following relations (again, among many others) $$\begin{aligned} & \phantom{o} & a_1 a_2=1, \label{eqn:B0} \\ & \phantom{o} & b_2 \bar{a}_2 b_1 \bar{a}_1 [b_3, a_3] =1, \label{eqn:B1} \\ & \phantom{o} & b_2 b_1 [b_3, a_3] =1, \nonumber\end{aligned}$$ induced by the vanishing cycles $B_0, B_1$ and $B_2$, respectively. We get $a_1=\bar{a}_2$, and together with the relators from $P'_2$ we once again get $b_1=\bar{b}_2$, since $[a_3, b_3]=1$. Since the positive factorization $W_1$ contains the factor $P_1 P'_2 \, t_{C}$ and $W_2, W_3$ both contain the factor $P_2 P'_2 \, t_{C}$, the above discussion shows that every $\pi_1(\widetilde{X_i})$ is a factor of the abelian group ${\mathbb{Z}}^2$ generated by $a_2$ and $b_2$. (In fact no additional relators come from the remaining twists in $P_1, P_2$ or $P'_2$.) For the remaining relators coming from the conjugated factors $P_1^{\phi}$ or $P_2^{\phi}$, it therefore suffices to look at the relators they induce in $H_1(\Sigma_3)$. Moreover, this allows us to simply look at the homology classes of the vanishing cycles in these conjugated factors. (We haven’t bothered to give explicitly for this very reason.) Without the conjugated factor, in each case we have the abelianized relations $$\label{ab1} a_3=-a_2 \, \text{ and } \, b_3=-b_2=b_1,$$ and depending on whether $W_i$ contains the factor $P_1$ or $P_2$, either $$\begin{aligned} & \phantom{o} & a_1+ 2a_2 +4b_2 &= 0 \ \ \ \text{for $W_1$, or} \label{ab2-1} \\ & \phantom{o} & a_1+a_2 &= 0 \ \ \ \text{for $W_2$ and $W_3$}, \label{ab2-2}\end{aligned}$$ where we used $(\ref{ab1})$ to simplify. For our conjugation $\phi= t_{b_1}^{-1} t_{a_2}$, we easily check using the Picard-Lefschetz formula that we get the additional relator $$\label{ab3} b_1 + a_2 +b_2 =0,$$ from both $P_1^{\phi}$ and $P_2^{\phi}$ whereas, depending on whether it is $P_1^{\phi}$or $P_2^{\phi}$ we get either $$\begin{aligned} & \phantom{o} & a_1 + b_1 + 6 a_2 + 4 b_2 &=0 \ \ \ \text{for $W_1$ and $W_2$, or} \label{ab4-1} \\ & \phantom{o} & a_1+ b_1+a_2 &= 0 \ \ \ \text{for $W_3$}. \label{ab4-2}\end{aligned}$$ Therefore, $(\ref{ab1})$ and $(\ref{ab3})$, which hold for all $\pi_1(\widetilde{X_i})$, imply $a_2=0$. The relators in $(\ref{ab2-1})$, $(\ref{ab2-2})$ and $(\ref{ab4-1})$, $(\ref{ab4-2})$ involved in a given $W_i$ then easily give $b_2=-b_1=0$, for each $i=1,2,3$. Hence in all cases we have got $\pi_1(X_i)=\pi_1(\widetilde{X_i})=1$. By Freedman, we have $X_i$ homeomorphic to ${{\mathbb{CP}}{}^{2}}\# \, (6+i) \, {\overline{\mathbb{CP}}{}^{2}}$, for $i=1,2,3$. Diffeomorphism type of the pencils $(X_i, f_i)$ ----------------------------------------------- A quick amendment to the title: we will determine the “non-diffeomorphism type” of $(X_i, f_i)$, namely, we will prove that $X_i$ are not diffeomorphic to rational surfaces. This will follow from the next lemma. \[notrational\] The rational surface ${{\mathbb{CP}}{}^{2}}\# p {\overline{\mathbb{CP}}{}^{2}}$, for $p \leq 9$, does not admit a genus $g \geq 2$ Lefschetz fibration or a genus-$g$ Lefschetz pencil with $m < 2g-2$ base points. It suffices to show this for $X= {{\mathbb{CP}}{}^{2}}\# 9 {\overline{\mathbb{CP}}{}^{2}}$, since we can always blow-up on the fibers to produce a genus-$g$ Lefschetz fibration (resp. pencil) on $X$ from a given genus-$g$ Lefschetz fibration (resp. pencil) on ${{\mathbb{CP}}{}^{2}}\# p {\overline{\mathbb{CP}}{}^{2}}$, for any $p < 9$. Assume that $X$ is equipped with a genus $g \geq 2$ Lefschetz fibration or a pencil with $m < 2g-2$ base points. For our arguments to follow, it will be convenient to allow $m$ to be non-negative so that $m=0$ marks the fibration case. Let $F= a H - \sum_{i=1}^9 c_i E_i$ be the fiber class, where $H_2(X)$ is generated by $H$ and $E_1, \ldots, E_9$, with $H^2=1$, $E_i \cdot E_j=- \delta_{ij}$, and $H \cdot E_i=0$. Since $F^2=m$, we have $$a^2= m+ \sum_{i=1}^9 c_i^2 \, .$$ We can equip $X$ with a Thurston-Gompf symplectic form $\omega$ which makes the fibers symplectic. Moreover, we can choose an $\omega$-compatible almost complex structure $J$, even a generic one in the sense of Taubes (see e.g [@UsherDS]), with respect to which $f$ is $J$-holomorphic (for a suitable choice of almost complex structure on the base $2$-sphere). It was shown by Li and Liu [@LiLiu] that for a generic $\omega$-compatible $J$, the class $H$ in the rational surface $X$ has a $J$-holomorphic representative . Hence, $F$ and $H$ both have $J$-holomorphic representatives, which implies that $F \cdot H = a \geq 0$. There is a unique symplectic structure on $X$ up to deformation and symplectomorphisms [@LiLiu], so we can apply the adjunction formula to get $$2g-2= F^2 + K \cdot F = m +(-3H + \sum_{i=1}^9 E_i) \cdot (aH - \sum_{i=1}^9 c_i E_i) = m -3a + \sum_{i=1}^9 c_i \, .$$ Since $a, m \geq 0$, and $g \geq 2$, from the above equalities we have $$3 a = \sqrt{9 a^2}=\sqrt{9 (m+\sum_{i=1}^9 c_i^2) } \geq \sqrt{9 (\sum_{i=1}^9 c_i^2) } = \sqrt{(\sum_{i=1}^9 1) (\sum_{i=1}^9 c_i^2)} \geq \sqrt{\|\sum_{i=1}^9 c_i\|^2} \, ,$$ where the last inequality is by Cauchy-Schwartz, and in turn $$3a \geq \sqrt{\|\sum_{i=1}^9 c_i\|^2} = \|\sum_{i=1}^9 c_i\| = \| 2g-2-m +3a \| = 2g-2 -m +3a\, ,$$ implying $m \geq 2g-2$. The contradiction shows that there exists no such symplectic surface $F$. In turn, there is no such fibration or pencil. In fact, if $X$ is a symplectic $4$-manifold in the homeomorphism class of a rational surface ${{\mathbb{CP}}{}^{2}}\# p {\overline{\mathbb{CP}}{}^{2}}$ with $p \leq 9$, $X$ is exotic if and only if it admits a genus $g \geq 2$ pencil with $m < 2g-2$ base points. This follows from Lemma \[notrational\] and Donaldson’s result on the existence of Lefschetz pencils on symplectic $4$-manifolds (together with the fact there is a unique genus-$1$ pencil, which is on $CP$). Combining the lemmas \[lemhomeo\] and \[notrational\] we conclude: \[exoticrational\] $(X_i,f_i)$ are symplectic genus-$3$ Lefschetz pencils whose total spaces are homeomorphic but not diffeomorphic to ${{\mathbb{CP}}{}^{2}}\# (6+i) {\overline{\mathbb{CP}}{}^{2}}$ for $i=1,2, 3$. \[smallest\] These are the smallest genera examples one can find. Any Lefschetz pencil of genus $0$ or $1$ has a total space diffeomorphic to (a blow-up of) ${{\mathbb{CP}}{}^{1}}\x {{\mathbb{CP}}{}^{1}}$, ${{\mathbb{CP}}{}^{2}}\# {\overline{\mathbb{CP}}{}^{2}}$ or $E(1)={{\mathbb{CP}}{}^{2}}\# 9 {\overline{\mathbb{CP}}{}^{2}}$ [@Kas; @Moishezon]. After blowing-up all the base points of a genus-$2$ Lefschetz pencil we arrive at a genus-$2$ Lefschetz fibration on a non-minimal symplectic $({\widetilde}{X},{\widetilde}{f})$. Now if ${\widetilde}{X}$ is an exotic rational surface, it cannot have the rational homology type of a (blow-up of a) symplectic Calabi-Yau surface either. Thus there could be at most one base point, and as observed by Sato, $({\widetilde}{X},{\widetilde}{f})$ has at most one reducible fiber; see Theorem 5-5(iii) and Theorem 5-12(iii) in [@SatoKodaira]. By Lemmas 4 and 5 in [@BaykurKorkmaz], there is no genus-$2$ Lefschetz fibration $({\widetilde}{X},{\widetilde}{f})$ with at most one reducible fiber and say $\eu({\widetilde}{X}) \leq 13$. In fact, in [@BaykurKorkmaz], we have constructed minimal genus-$2$ Lefschetz fibrations whose total spaces are homeomorphic to rational surfaces ${{\mathbb{CP}}{}^{2}}\# \, p {\overline{\mathbb{CP}}{}^{2}}$, for $p=7,8,9$ [@BaykurKorkmaz]. These fibrations decompose as fiber sums, which implies their minimality [@Usher; @BaykurPAMS], and in turn why they are not diffeomorphic to blown-up rational surfaces. The above observation generalizes to any genus $g \geq 3$ Lefschetz pencil $(X,f)$ whose total space is an exotic ${{\mathbb{CP}}{}^{2}}\# p {\overline{\mathbb{CP}}{}^{2}}$, for $p < 9$, allowing us to conclude that $f$ has at most $2g-4$ base points. It follows from the following more general observation: A collection of exceptional classes in the corresponding Lefschetz fibration $({\widetilde}{X}, {\widetilde}{f})$ can be represented by disjoint multisections $S_j$, so each one intersects the regular fiber $F$ positively at least once. However, $(\sum S_j) \cdot F \geq 2g-3$ implies that $\kappa(X)= \kappa({\widetilde}{X} \leq 1$ by the work of Sato [@SatoKodaira]. Since $X$ is not rational (nor can possibly be ruled), we already know that $\kappa(X) \neq -\infty$. Further, $\kappa(X)=0$ or $1$ implies that the minimal model of $X$ has $c_1^2=0$, which is impossible for any $X$, as $c_1^2(X)=c_1^2({{\mathbb{CP}}{}^{2}}\# p\, {\overline{\mathbb{CP}}{}^{2}})= 9-p >0$ and blow-down increases $c_1^2$. Hence $\kappa(X) = 2$ and has at most $2g-4$ disjoint exceptional spheres. In particular we see that each $X_i$ in Theorem \[exoticrational\] is either minimal or it is only once blown-up of a minimal symplectic $4$-manifold. Using a variation of our construction with $3$ different embeddings of $P_1$ into $\Gamma_3$, we can get a genus-$3$ Lefschetz fibration $({\widetilde}{X}_0, {\widetilde}{f}_0)$ where ${\widetilde}{X}_0$ is an exotic symplectic ${{\mathbb{CP}}{}^{2}}\# 7 {\overline{\mathbb{CP}}{}^{2}}$. I expect this monodromy lifts to $\Gamma_3^1$, and gives another pencil $(X_0, f_0)$ whose total space is an exotic ${{\mathbb{CP}}{}^{2}}\# 6 {\overline{\mathbb{CP}}{}^{2}}$. \[distinctsymp\] The existence of exotic symplectic $4$-manifolds in the homeomorphism classes in Theorem \[exoticrational\] was already established. The first example of an exotic symplectic ${{\mathbb{CP}}{}^{2}}\# \, p \, {\overline{\mathbb{CP}}{}^{2}}$, for $p=9$, was the complex Dolgachev surface $E(1)_{2,3}$, as shown by Donaldson [@Donaldson87]. For $p=8$, the first example was Barlow’s complex surface, as proved by Kotschick [@Kotschick]. For $p=7$, the first examples were constructed via generalized rational blowdowns by Jongil Park [@ParkJ], which are symplectic by the work of Symington [@Symington]. Infinitely many distinct smooth structures in these homeomorphism classes were constructed using logarithmic transforms, knot surgeries and Luttinger surgeries [@FSKnotsurgery; @Friedman; @Szabo; @FSDoublenode; @ABP] (all of which are indeed instances of surgeries along tori [@BaykurSunukjian].) Notably, only for $p=9$ it is known that there is an infinite family of pairwise non-diffeomorphic symplectic $4$-manifolds in the homeomorphism class of ${{\mathbb{CP}}{}^{2}}\# \, p \, {\overline{\mathbb{CP}}{}^{2}}$ (e.g. [@FSKnotsurgery]). Moreover, it was observed by Stipsicz and Szabo that Seiberg-Witten invariants cannot distinguish any two minimal symplectic $4$-manifolds homeomorphic to ${{\mathbb{CP}}{}^{2}}\# \, p {\overline{\mathbb{CP}}{}^{2}}$ for $p <9$ [@StipsiczSzabo]\[Corollary 4.4\]. It remains an open question whether there are two distinct symplectic $4$-manifolds in these homeomorphism classes [@Stern]\[Problem 11\]. It is thus desirable to have examples with more structure in order to address this intriguing question. For instance, one can use other $\phi$ to obtain simply-connected symplectic genus-$3$ pencils in the same homeomorphism classes, possibly not isomorphic to the ones in Theorem \[exoticrational\]. An infinite family of genus-$3$ pencils with $c_1^2=0,1,2$ and $\chi_h=1$ ------------------------------------------------------------------------- By varying our construction of the positive factorizations $W_i$, we can also produce symplectic genus-$3$ Lefschetz pencils with any prescribed abelian group of rank at most $2$. Let $M=(m_1,m_2)$, where $m_1$ and $m_2$ are nonnegative integers. For $i=1,2,3,$ let $(\widetilde{X}_{i,M},\widetilde{f}_{i,M})$ be the Lefschetz fibration prescribed by the positive factorization $W_i$ with $\phi_i= t^{-m_1}_{b_1} t_{a_2}^{m_2} $, and $({X}_{i,M},{f}_{i,M})$ be the corresponding pencil. Note that we still get a positive factorization $W_i$ since $\phi$ keeps $C$ and $C'$ fixed. We claim that $\pi_1(\widetilde{X}_{i,M}) \cong ({\mathbb{Z}}\, / {m_1 \, {\mathbb{Z}}}) \oplus ({\mathbb{Z}}\, / {m_2 \, {\mathbb{Z}}})$. To verify it, let’s go back to our fundamental group calculation via the finite presentation which has the generators $a_j, b_j$ of $\pi_1(\Sigma_3)$, and relators induced by the Dehn twist curves in $W_i$. The arguments we gave in Section \[homeo\] for the relators coming from the factors $P_1, P_2$ and $P'_2$ appearing in $W_i$ apply here as well. In particular, $\pi_1(\widetilde{X}_{i,M}))$ is also a quotient of the abelian group ${\mathbb{Z}}^2$ generated by $a_2$ and $b_2$, and the following abelianized relations hold: $$\begin{aligned} & \phantom{o} & a_3=-a_2 \, \text{ and } \, b_3 =-b_2 &=b_1 \ &\text{for all $W_i$}, \\ & \phantom{o} & a_1+ 2a_2 +4 b_2 &= 0 \ \ \ &\text{for $W_1$,} \\ & \phantom{o} & a_1+a_2 &= 0 \ \ \ &\text{for $W_2$ and $W_3$}. \end{aligned}$$ Importantly, there are no other relations coming from the factors $P_1, P_2$ and $P'_2$: This is easy to see by abelianizing the relators $(~\ref{eqn:0})$–$(~\ref{eqn:B1})$, which include all the relators induced by the curves $x_1, x_2, x_3, B_0, B_1, B_2, B'_0, B'_1, B'_2$. Missing are the relators induced by the separating curves $d, e$ from $P_1$, the curves $A_0, A_1, A_2$ from $P_1$, and the curves $A'_0, A'_1, A'_2$ from $P'_2$. The first two are trivial in homology, so they have no contribution to the list of relators we already have. On the other hand, for each $j=0,1,2$, $A_j$ is homologous to $B_j$, because $[A_j]= [t_C(B_j)]=[B_j] + (C \cdot B_j) [C]$ by the Picard-Lefschetz formula, where $C$ is a separating cycle. Similarly each $A'_j$ is homologous to $B'_j$. Therefore the abelianized relations they induce are identical to those we already had from $B_j, B'_j$. For our conjugation $\phi= t_{b_1}^{-m_1} t_{a_2}^{m_2}$, we get the additional relators $$\begin{aligned} & \phantom{o} & b_1 + m_2 a_2 +b_2 =0 &=0 \ \ \ \text{for all $W_i$,} \label{abconj1} \\ & \phantom{o} & a_1 + m_1 b_1 + (2+4m_2) a_2 + 4 b_2 &=0 \ \ \ \text{for $W_1$ and $W_2$,} \label{abconj2} \\ & \phantom{o} & a_1+ m_1 b_1+a_2 &= 0 \ \ \ \text{for $W_3$}. \label{abconj3}\end{aligned}$$ Similar to before, $(\ref{abconj1})$ and $(\ref{abconj3})$, which hold for all $\pi_1(\widetilde{X}_{i,M})$, imply $m_2 a_2=0$. The remaining relators involved in a given $W_i$ then easily give $m_1 b_2=-m_1 b_1=0$, for each $i=1,2,3$. Hence, $\pi_1(X_{i,M})= \pi_1(\widetilde{X}_{i,M})= ({\mathbb{Z}}\, / {m_1 \, {\mathbb{Z}}}) \oplus ({\mathbb{Z}}\, / {m_2 \, {\mathbb{Z}}})$, as claimed. Since the monodromy of each $(\widetilde{X}_{i,M}, \widetilde{f_{i,M}})$ is equivalent to that of $(\widetilde{X_i}, \widetilde{f_i})$ by a partial conjugation with $t_{b_1}^{-m_1+1} t_{a_2}^{m_2-1}$, they have the same Euler characteristic and signature. So $c_1^2(X_{i,M})=3-i$ and $\chi(X_{i,M})=1$, for $i=1,2,3$. For each $i$, we clearly have an infinite family of symplectic genus-$3$ pencils among $$\{ ({X}_{i,M},{f}_{i,M}) \ \| \ M = (m_1, m_2) \in {\mathbb{N}}\times {\mathbb{N}}\ \} ,$$ with pairwise homotopy inequivalent total spaces easily distinguished by $\pi_1(X_{i,M})$. By Parshin and Arakelov’s proofs of the Geometric Shafarevich Conjecture, there are finitely many holomorphic fibrations with fixed fiber genus $g\geq 2$ and degeneracy [@Parshin; @Arakelov]. This implies that all but finitely many of these symplectic pencils are non-holomorphic. Furthermore, for $M=(m_1, 0)$ with $m_1 \geq 2$, we get an infinite subfamily of pencils, whose total spaces cannot be complex surfaces, since there is no complex surface $Z$ with $b_1(Z)=1$ and $b^+(Z)>0$ (e.g. cf. [@BaykurHolomorphic]\[Lemma 2\]). (Note that the same argument applies to the infinite family of symplectic Calabi-Yau surfaces we will construct in the next section.) We have proved: \[nonholom\] There exists a symplectic genus-$3$ Lefschetz pencil $(X_{i,M}, f_{i,M})$ with $c_1^2(X_{i,M})=3-i$, $\chi(X_{i,M})=1$, and $\pi_1(X_{i,M})= ({\mathbb{Z}}\, / {m_1 \, {\mathbb{Z}}}) \oplus ({\mathbb{Z}}\, / {m_2 \, {\mathbb{Z}}})$, for any pair of non-negative integers $M=(m_1, m_2)$, and for each $i=1,2,3$. Infinitely many of these pencils have total spaces homotopy inequivalent to a complex surface. \[nonholompencil\] There is only one example of a non-holomorphic genus-$3$ Lefschetz pencil in the literature, which was given by Smith in [@SmithGenus3]\[Theorem 1.3\] without appealing to Donaldson’s theorem.[^3] (We learn from Naoyuki Monden that in a joint work with Hamada and Kobayashi they produce two more examples.) Theorem \[nonholom\] improves the situation, at least quantitatively. \[pi1GenusInvariant\] Following the works of Donaldson [@Donaldson] and Gompf [@Gompf], every finitely presented group is the fundamental group of a symplectic Lefschetz pencil. (Also see [@ABKP] and [@Korkmaz2].) An invariant of finitely presented group $G$ is the $m_g(G)$, the minimal genus of a genus-$g$ pencil $(X,f)$ with $\pi_1(X)=G$, which is not as easy to calculate in general. However, Theorem \[nonholom\] shows that for $G \cong ({\mathbb{Z}}\, / {m_1 \, {\mathbb{Z}}}) \oplus ({\mathbb{Z}}\, / {m_2 \, {\mathbb{Z}}})$, $m(G) \leq 3$. Well-known examples of genus-$1$ and genus-$2$ pencils show that $m_g(G)=1$ for $G=1$ (the pencil on ${{\mathbb{CP}}{}^{2}}$), and $m_g(G)=2$ for $G= {\mathbb{Z}}_2$ (the pencil on the Enriques surface) or ${\mathbb{Z}}\oplus {\mathbb{Z}}$ (e.g. Matsumoto’s pencil on $S^2 \x T^2$). It seems likely that for all the other $G$ as above, $m_g(G)=3$. A new construction of symplectic Calabi-Yau surfaces with $b_1>0$ ================================================================= Here we will give a new construction of an infinite family of symplectic Calabi-Yau surfaces with $b_1=2$ and $3$, along with a symplectic Calabi-Yau surface with $b_1=4$, homeomorphic to the $4$-torus. This will follow from our construction of new positive factorization of boundary multi-twists in $\Gamma_3^4$ corresponding to symplectic genus-$3$ Lefschetz pencils. We find it more convenient for presentation purposes to discuss a special example first, and discuss the most general construction afterward. These will correspond to our examples with $b_1=4$, and then the others with $b_1=2$ and $3$. A symplectic Calabi-Yau homeomorphic to the $4$-torus {#Sec:4torus} ----------------------------------------------------- Our main building block will be the following relation in $\Gamma_2^2$ [@Hamada]: $$\begin{aligned} \label{firstlift} (t_{B_0} t_{B_1} t_{B_2} t_{C} )^2 &=& t_{\delta_1} t_{\delta_2} \, ,\end{aligned}$$ where $\delta_i$ are the boundary parallel curves, and the curves $B_i$ and $C$ are as shown on the left-hand side of Figure \[Matsumoto\]. ![Vanishing cycles $B_j, C, B_{j,i}$ and $C_i$ in Hamada’s lift of Matsumoto’s fibration. On the left are the curves of the positive factorization in $\Gamma_2^2$, along with the curves $A_j$ we get after Hurwitz moves. On the right are the curves of the further lift in $\Gamma_2^4$. []{data-label="Matsumoto"}](MatsumotoLifts.eps){width="10cm"} The relation $(\ref{firstlift})$ can be rewritten as $$\begin{aligned} t_{B_0} t_{B_1} t_{B_2} t_{A_0} t_{A_1} t_{A_2} t_{C}^2 t_{\delta_1}^{-1} t_{\delta_2}^{-1} &=& 1 \, ,\end{aligned}$$ where each $A_j= t_C(B_j)$ for $j=0,1,2$ are given on the left-hand side Figure \[Matsumoto\]). We can embed this relation into $\Gamma_3$ by gluing along the two boundary components $\delta_1$ and $\delta_2$ so that $\delta_1=\delta_2$ is mapped to $C'$ and the curves $C, B_j, A_j$ are as shown in Figure \[SCYcurves\], denoted by the same letters for simplicity. Thus the following holds in $\Gamma_3$: $$\begin{aligned} \label{embedSCY1} t_{B_0} t_{B_1} t_{B_2} t_{A_0} t_{A_1} t_{A_2} t_{C}^2 t_{C'}^{-2} &=& 1 \, ,\end{aligned}$$ or $P \, t_{C}^2 t_{C'}^{-2}= 1$, where $\underline{P= t_{B_0} t_{B_1} t_{B_2} t_{A_0} t_{A_1} t_{A_2}} $. Clearly $P, t_C$ and $t_{C'}$ all commute with each other. A similar embedding into $\Gamma_3$ can be given by mapping $\delta_1=\delta_2$ to $C$ instead, where the curves $C, B_j, A_j$ are now as shown in Figure \[SCYcurves\], all decorated by a prime notation this time (so $C$ indeed maps to $C'$ above). We thus get another relation in $\Gamma_3$ of the form: $$\begin{aligned} \label{embedSCY2} t_{B'_0} t_{B'_1} t_{B'_2} t_{A'_0} t_{A'_1} t_{A'_2} t_{C'}^2 t_{C}^{-2} &=& 1 \, ,\end{aligned}$$ or $P' \, t_{C'}^2 t_{C}^{-2}= 1$, where $\underline{P'= t_{B'_0} t_{B'_1} t_{B'_2} t_{A'_0} t_{A'_1} t_{A'_2}} $. Similarly, $P', t_C$ and $t_{C'}$ all commute with each other. Combining the above, we get a positive factorization $\underline{W=P P'}$ in $\Gamma_3$, since $$P P' = P \, t_{C}^2 t_{C'}^{-2} \, P' \, t_{C'}^2 t_{C}^{-2} = 1 \, .$$ Let $(\widetilde{X}, \widetilde{f})$ be the genus-$3$ Lefschetz fibration prescribed by $W$. We can equip $({\widetilde}{X}, {\widetilde}{f})$ with a Gompf-Thurston symplectic form ${\widetilde}{\omega}$, in particular ${\widetilde}{X}$ is a symplectic $4$-manifold. We will show that $\widetilde{X}$ is the $4$ times blow-up of a symplectic Calabi-Yau surface $X$ homeomorphic to the $4$-torus. We will indeed show that it is a blow-up of a symplectic genus-$3$ pencil $(X,f)$, which we will first argue implicitly below, and then explicitly in the next section. \[SCYpencil\] Let $({\widetilde}{X},{\widetilde}{f})$ be a genus $g \geq 2$ Lefschetz fibration and $c_1^2({\widetilde}{X})=2-2g$. If ${\widetilde}{X}$ is not a rational or ruled surface, then $({\widetilde}{X},{\widetilde}{f})$ is a blow-up of a symplectic Lefschetz pencil $(X,f)$, where $X$ is a (minimal) symplectic Calabi-Yau surface. We can equip $({\widetilde}{X},{\widetilde}{f})$ with a Gompf-Thurston symplectic form ${\widetilde}{\omega}$. By the work of Taubes [@Taubes; @Taubes2], any minimal symplectic $4$-manifold has non-negative $c_1^2$. So ${\widetilde}{X} \cong X_0 \# m {\overline{\mathbb{CP}}{}^{2}}$, where $X_0$ is a minimal model for ${\widetilde}{X}$ and $m \geq 2g-2$, which means that ${\widetilde}{X}$ contains at least $2g-2$ disjoint exceptional spheres $S_j$, $j=1, \ldots, 2g-2$. There is an ${\widetilde}{\omega}$-compatible almost complex structure $J$ with respect to which both a regular fiber $F$ of ${\widetilde}{f}$ and all $S_j$ are $J$-holomorphic [@SatoKodaira; @BaykurPAMS], which implies that $(\sum_{j=1}^{2g-2} S_j) \cdot F \geq 2g-2$. By the work of Sato [@SatoKodaira] (also see [@BaykurHayano]), this is only possible if the symplectic Kodaira dimension $\kappa({\widetilde}{X}) \leq 0$. As we assumed ${\widetilde}{X}$ is not a rational or a ruled surface, we have $\kappa({\widetilde}{X}) \neq - \infty$ [@LiSCY], leaving the only possibility as $\kappa({\widetilde}{X})=0$ and also that there are no other disjoint exceptional spheres than $S_j$ in ${\widetilde}{X}$. Thus $(\sum_{j=1}^{2g-2} S_j) \cdot F = 2g-2$, so one can blow-down these exceptional spheres (which can be assumed to be symplectic after deforming ${\widetilde}{\omega}$ if needed [@GompfStipsicz]) to arrive at a symplectic genus-$3$ Lefschetz pencil $(X,f)$, where $X= X_0$ is a minimal symplectic Calabi-Yau surface. We show that ${\widetilde}{X}$ satisfies the hypotheses of this lemma. The Euler characteristic of ${\widetilde}{X}$ is given by $\eu({\widetilde}{X})= 4-4 \cdot 3 + \ell(W)= 4$, where $\ell(W)=12$ is the number of Dehn twists in the factorization $W$. The signature of ${\widetilde}{X}$ can be calculated as $\sigma({\widetilde}{X})=-4$ using the local signatures and Ozbagci’s algorithm [@Ozbagci]. Since $({\widetilde}{X}, {\widetilde}{f})$ is not hyperelliptic, we do not have the luxury of employing Endo’s closed formula for a direct calculation, however the following observation allows us to deduce the signature from a calculation in $\Gamma_2$, where any factorization is hyperelliptic: We have obtained $W$ from the factorization which had two pairs of $t_C$ and $t_{C'}$ with opposite signs. This is an achiral genus-$3$ Lefschetz fibration, which is a fiber sum of two achiral fibrations prescribed by the factorizations $P \, t_{C}^2 t_{C'}^{-2}$ and $P' \, t_{C'}^2 t_{C}^{-2}$. Each one is a section sum of an achiral genus-$2$ Lefschetz fibration and a trivial torus fibration along a section of self-intersection zero, which, by Novikov additivity, has signature equals to that of the achiral genus-$2$ Lefschetz fibration. By the hyperelliptic signature formula for genus-$2$ fibrations [@Matsumoto; @Endo], the signature of the latter is $-2$. Invoking Novikov additivity again, we conclude that the decomposable achiral genus-$3$ fibration has signature $-4$, the sum of the signature of its summands. Lastly, we claim that canceling the pairs $t^{\pm 2}_C$ and $t^{\pm 2}_{C'}$ does not change the signature. This can be easily seen by splitting the fibration as $t^2_C t^{-2}_C$ (or $t^2_{C'} t^{-2}_{C'}$) and the rest. The piece with monodromy factorization $t^2_C t^{-2}_C$ has zero signature, which is easily calculated from the induced handle decomposition, or by observing that it can be further split into two identical fibrations whose total spaces have opposite orientations (and thus opposite signatures). By Novikov additivity once again, the rest of the fibration has the same signature as the total signature of the achiral fibration. So $\sigma({\widetilde}{X})=-4$. It follows that $c_1^2({\widetilde}{X})= 2 \eu({\widetilde}{X}) + 3\sigma({\widetilde}{X})=-4$. Assuming $\pi_1({\widetilde}{X})={\mathbb{Z}}^4$ for the moment, let us complete our arguments regarding the homeomorphism type of the minimal model $X$ of ${\widetilde}{X}$. By Lemma \[SCYpencil\] $({\widetilde}{X}, {\widetilde}{f})$ is a blow-up of a symplectic genus-$3$ Lefschetz pencil $(X,f)$, where $X$ is a symplectic Calabi-Yau surface. Since $X$ is a symplectic Calabi-Yau surface with $b_1(X)>0$, it has the rational homology type of a torus bundle over a torus [@Bauer; @LiSCY]. Moreover, since $\pi_1(X)={\mathbb{Z}}^4$ is a virtually poly-${\mathbb{Z}}$ group, the Borel conjecture holds in this case by the work of Farrell and Jones [@FJ90], and as observed by Friedl and Vidussi, this implies that a symplectic Calabi Yau surface with $\pi_1({\mathbb{Z}}^4)$ is unique up to homeomorphism [@FriedlVidussi]. Hence $X$ is homeomorphic to the well-known symplectic Calabi-Yau surface, the $4$-torus. We now calculate $\pi_1(X)=\pi_1({\widetilde}{X})$, which will constitute the rest of this section. Using the standard handlebody decomposition for $({\widetilde}{X}, {\widetilde}{f})$, we can obtain a finite presentation for $\pi_1({\widetilde}{X}) \cong \pi_1(\Sigma_3) \, / \, N$, where $N$ denotes the subgroup of $\pi_1(\Sigma_3)$ generated normally by the vanishing cycles of ${\widetilde}{f}$. For $\{a_j, b_j\}$ standard generators of $\pi_1(\Sigma_g)$ as given in Figure \[pi1generators\], we get $$\pi_1(X) \cong \, \langle \, a_1, b_1, a_2, b_2, a_3, b_3 \, \| \, [a_1, b_1] [a_2, b_2] [a_3, b_3] , R_1, \ldots, R_{12} \, \rangle \, ,$$ where each $R_i$ is a relation obtained by expressing a corresponding vanishing cycle (oriented arbitrarily) in $\{a_j, b_j\}$. Recall that $({\widetilde}{X}, {\widetilde}{f})$ has a section (in fact $4$ $(-1)$-sections), so the above presentation is complete, there is no other relation. ![Generators $a_j, b_j$ of $\pi_1(\Sigma_3)$[]{data-label="pi1generators"}](pi1generators.eps){width="5cm"} We have the following relations, the ones after the first one coming from $\Sigma_3$ all induced by the vanishing cycles $B_0$, $B_1$, $B_2$, $A_0$, $A_1$, $A_2$, $B'_0$, $B'_1$, $B'_2$, $A'_0$, $A'_1$, $A'_2$ (which are all given in Figure \[SCYcurves\], in the given order: $$\begin{aligned} & \phantom{o} & [a_1, b_1][a_2, b_2][a_3,b_3]=1 , \label{0} \\ & \phantom{o} & a_1 a_3=1, \label{B0} \\ & \phantom{o} & a_1 \bar{b}_1 a_2 b_2 \bar{a}_2 a_3 \bar{b}_3=1, \label{B1} \\ & \phantom{o} & \bar{b}_1 a_2 b_2 \bar{a}_2 \bar{b}_3 =1, \label{B2} \\ & \phantom{o} & a_1 [b_3, a_3] \, b_2 a_3 \bar{b}_2 \, [a_3, b_3]=1, \label{A0} \\ & \phantom{o} & a_3 \bar{b}_3 \bar{b}_2 [a_3, b_3] \, a^2_1 \bar{b}_1 \bar{a}_1 [b_3, a_3] \, b_2 [b_3, a_3] \, b_2=1, \label{A1} \\ & \phantom{o} & a_1 \bar{b}_1 \bar{a}_1 [b_3, a_3] \, b_2 [b_3, a_3] \, b_2 \bar{b}_3 \bar{b}_2 [a_3, b_3] =1, \label{A2} \\ & \phantom{o} & \bar{a}_2 a_1 a_2 a_3=1, \label{B'0} \\ & \phantom{o} & a_1 \bar{b}_1 a_2 a^2_3 \, \bar{b}_3 \bar{a}_3 b_2 \bar{a}_2 =1, \label{B'1} \\ & \phantom{o} & b_1 a_2 \bar{b}_2 a_3 b_3 \bar{a}_3 \bar{a}_2 =1, \label{B'2} \\ & \phantom{o} & a_1 a_2 \bar{b}_2 a_3 b_2 \bar{a}_2=1, \label{A'0} \\ & \phantom{o} & a_1 \bar{b}_1 a_2 \bar{b}_2 a^2_3 \, \bar{b}_3 \bar{a}_3 b^2_2 \, \bar{a}_2=1, \label{A'1} \\ & \phantom{o} & \bar{b}_1 a_2 \bar{b}_2 a_3 \bar{b}_3 \bar{a}_3 b^2_2 \, \bar{a}_2 =1, \label{A'2} \end{aligned}$$ Here $\bar{x}$ denotes the inverse of the element $x$ in the group. ![Vanishing cycles $B_j, A_j, B'_j, A'_j$ of the genus-$3$ fibration $(X,f)$. On the left are the curves coming from the factorization $P$ and on the right are those coming from the factorization $P'$, which correspond to the two different embeddings of the factorization in $\Gamma_2^2$ into $\Gamma_3$. (Dotted lines drawn on the surfaces on the top are the images of $\delta_1=\delta_2$ under these embeddings.)[]{data-label="SCYcurves"}](SCYGenus3curves.eps){height="18.5cm"} When abelianized, the relations coming from each one of the triples $\{B_0, B_1, B_2\}$, $\{A_0, A_1, A_2\}$, $\{B'_0, B'_1, B'_2\}$, $\{A'_0, A'_1, A'_2\}$ give the same three relations $$\begin{aligned} & \phantom{o} & a_1 + a_3= 0, \\ & \phantom{o} & a_1 -b_1 + b_2 + a_3 - b_3= 0, \\ & \phantom{o} & -b_1 +b_2 -b_3 =0, \end{aligned}$$ where we identified the images of the generators with the same letters. Any $2$ of these relations imply the other. So we can eliminate $a_1=-a_3$ and $b_1=b_2-b_3$, and get a free abelian group of rank $4$ generated by $a_2, b_2, a_3$ and $b_3$. Now, going back to the presentation we have above for $\pi_1({\widetilde}{X})$, we easily see that it is also generated by $a_2, b_2, a_3, b_3$, since $a_1= \bar{a}_3$ by $(\ref{B0})$ and $b_1 = a_2 b_2 \bar{a}_2 \bar{b}_3$ by $(\ref{B2})$. To prove $\pi_1({\widetilde}{X})={\mathbb{Z}}^4$, it therefore suffices to show that $a_2, b_2, a_3, b_3$ all commute with each other. Replacing $a_1$ with $\bar{a}_3$ in $(\ref{B'0})$ gives $[a_3, a_2]=1$. From $(\ref{B2})$ we have $\bar{b}_1 a_2 b_2 \bar{a}_2 = b_3$. Substituting this in $(\ref{B1})$, and replacing $a_1$ with $\bar{a}_3$, we get $[a_3, b_3]=1$. With $a_1=\bar{a}_3$ and $[a_3, b_3]=1$, the relation $(\ref{A0})$ simplifies to $[a_3, b_2]=1$. So $a_3$ commute with each one of $a_2, b_2, b_3$. We can rewrite $(\ref{A'0})$ as $a_1 a_2 \bar{b}_2 a_3 b_2 \bar{a}_2= a_2 \bar{b}_2 a_3 b_2 \bar{a}_2 =1$, since $a_1=\bar{a}_3$ and $a_3$ commutes with $a_2$ and $\bar{b}_2$. So $[a_2, b_2]=1$. It implies that $b_1 = a_2 b_2 \bar{a}_2 \bar{b}_3 = b_1 \bar{b}_3$. Note that $[a_3, b_3]=1$ means $a_3 b_3 \bar{a}_3 = b_3$. Using these last two identities and $(\ref{B'2})$ we derive $b_2 \bar{b}_3 a_2 \bar{b}_2 b_3 \bar{a}_2 =1$. So $[b_2, \bar{b}_3 a_2] =[a_2, \bar{b}_2 b_3]=1$. These commuting relations, together with $[a_2, b_2]=1$, give us the last two commuting relations we need: $[b_2, b_3]=[a_2, b_3]=1$. Hence $\pi_1({\widetilde}{X}) \cong {\mathbb{Z}}^4$, generated by $a_2, b_2, a_3, b_3$. This completes our claim that $(X,f)$ is a symplectic genus-$3$ Lefschetz pencil, where $X$ is a symplectic Calabi-Yau surface homeomorphic to the $4$-torus. Symplectic Calabi-Yau surfaces with $b_1=2$ and $3$ via pencils --------------------------------------------------------------- We will now generalize the construction in the previous section, and construct an infinite family of examples of genus-$3$ Lefschetz pencils on symplectic Calabi-Yau surfaces with $b_1=2$ and $3$. We will produce these pencils explicitly via positive factorizations of the boundary multi-twist $\Delta= t_{\partial_1} t_{\partial_2} t_{\partial_3} t_{\partial_4}$ in $\Gamma_2^4$, where $\partial_i$ are boundary parallel curves for distinct boundary components of $\Sigma_3^4$. For this, we will use a further lift of the monodromy factorization of Matsumoto’s genus-$2$ fibration to $\Gamma_2^4$ given by Hamada, which will allow us to obtain a nice and symmetric presentation at the end. The factorization $(\ref{firstlift})$ has a further lift to $\Gamma_2^4$ [@Hamada] of the form: $$\begin{aligned} \label{secondlift} t_{B_{0,1}} t_{B_{1,1}} t_{B_{2,1}} t_{C_1} \, t_{B_{0,2}} t_{B_{1,2}} t_{B_{2,2}} t_{C_2} &=& t_{\delta_1} t_{\delta_2} t_{\delta_3} t_{\delta_4} \, ,\end{aligned}$$ where again $\delta_i$ are the boundary parallel curves, and the curves $B_{j,i}, C_i$ are given on the right-hand side of Figure \[Matsumoto\]. When the two boundary components $\delta_3, \delta_4$, drawn in the middle of $\Sigma_2^4$ in Figure \[Matsumoto\], are capped off, the curves $B_{j,i}$ descend to $B_j$ and $C_i$ to $C$ for each $j=0,1,2$ and $i=1,2$. ![Curves involved in our embeddings of $\partial \Sigma_4^2$ into $\partial \Sigma_3^4$.[]{data-label="SCYgluing"}](SCYgluing.eps){width="5cm"} The relation $(\ref{secondlift})$ can be rewritten as the following relation in $\Gamma_2^4$: $$\begin{aligned} t_{B_{0,1}} t_{B_{1,1}} t_{B_{2,1}} \, t_{A_{0,2}} t_{A_{1,2}} t_{A_{2,2}} t_{C_1} t_{C_2} t^{-1}_{\delta_1} t^{-1}_{\delta_2} &=& t_{\delta_3} t_{\delta_4} \, ,\end{aligned}$$ where each $A_{j,2}= t_{C_1}(B_{j,2})$ for $j=0,1,2$. Embed this configuration of curves on $\Sigma_2^4$ into $\Sigma_3^4$ by gluing a cylinder with $2$ holes along two of the boundary components of $\Sigma_3^4$. Choose the embedding so that $\delta_1$ maps to $C'_2$, $\delta_2$ to $C'_1$, $\delta_3$ to $\partial_2$ and $\delta_4$ to $\partial_1$. (See Figure \[SCYgluing\].) We then map the interior of $\Sigma_2^4$ so that the curves $B_{j,1}$ and $A_{j,2}$ all map to the curves $B_j$ and $A_j$ in Figure \[SCYcurves\] when the boundary components $\partial_3$ and $\partial_4$ are capped off. (That is, we map the interior so that when $\partial \Sigma_3^4$ is capped off, we get the first embedding we had in the previous section.) Thus the following holds in $\Gamma_3^4$: $$\begin{aligned} \label{embedSCY3} t_{B_{0,1}} t_{B_{1,1}} t_{B_{2,1}} \, t_{A_{0,2}} t_{A_{1,2}} t_{A_{2,2}} t_{C_1} t_{C_2} t^{-1}_{C'_2} t^{-1}_{C'_1} &=& t_{\partial_1} t_{\partial_2} \, ,\end{aligned}$$ or ${\widetilde}{P} \, t_{C_1} t_{C_2} t^{-1}_{C'_2} t^{-1}_{C'_1} = 1$, where $\underline{{\widetilde}{P}=t_{B_{0,1}} t_{B_{1,1}} t_{B_{2,1}} \, t_{A_{0,2}} t_{A_{1,2}} t_{A_{2,2}}}$. A similar embedding into $\Gamma_3^4$ can be given by mapping $\delta_1$ to $C_2$, $\delta_2 \to C_1$, $\delta_3 \to \partial_3$ and $\delta_4 \to \partial_4$ (see Figure \[SCYgluing\], where the interior is mapped in a similar fashion as above so as to get the curves $B'_j$ and $A'_j$ of Figure \[SCYcurves\]. We get another relation in $\Gamma_3^4$ of the form: $$\begin{aligned} \label{embedSCY2} t_{B'_{0,1}} t_{B'_{1,1}} t_{B'_{2,1}} \, t_{A'_{0,2}} t_{A'_{1,2}} t_{A'_{2,2}} t_{C'_1} t_{C'_2} t^{-1}_{C_2} t^{-1}_{C_1} &=& t_{\partial_3} t_{\partial_4} \, ,\end{aligned}$$ or ${\widetilde}{P}' \, t_{C'_1} t_{C'_2} t^{-1}_{C_2} t^{-1}_{C_1} = 1$, where $\underline{{\widetilde}{P}'=t_{B'_{0,1}} t_{B'_{1,1}} t_{B'_{2,1}} \, t_{A'_{0,2}} t_{A'_{1,2}} t_{A'_{2,2}}}$. Now let $\phi$ be any mapping class in $\Gamma_3^4$ which fix the curves $C_1, C_2, C'_1, C'_2$. Then the product of ${\widetilde}{P}^{\phi}$ and ${\widetilde}{P}'$ then give: $$\begin{aligned} & \phantom{o} & {\widetilde}{P}^{\phi} {\widetilde}{P}' &= {\widetilde}{P}^{\phi} \, t_{C_1} t_{C_2} t^{-1}_{C'_2} t^{-1}_{C'_1} \ {\widetilde}{P}' t_{C'_1} t_{C'_2} t^{-1}_{C_2} t^{-1}_{C_1} \\ & \phantom{o} & &= ({\widetilde}{P} \, t_{C_1} t_{C_2} t^{-1}_{C'_2} t^{-1}_{C'_1})^{\phi} \ {\widetilde}{P}' t_{C'_1} t_{C'_2} t^{-1}_{C_2} t^{-1}_{C_1} \\ & \phantom{o} & &= t_{\partial_1} t_{\partial_2} t_{\partial_3} t_{\partial_4} \, .\end{aligned}$$ In the first equality we used the commutativity of disjoint Dehn twists $t_{C_1}$, $t_{C_2}$, $t_{C'_1}$, $t_{C'_2}$ and that they all commute with ${\widetilde}{P}$ (and ${\widetilde}{P}'$). The second equality comes from our choice of $\phi$ so that $\phi$ commutes with Dehn twists along $C_1$, $C_2$, $C'_1$, $C'_2$. Therefore $\underline{W_{\phi}= {\widetilde}{P}^\phi {\widetilde}{P}'}$ is a positive factorization of the boundary multi-twist $\Delta=t_{\partial_1} t_{\partial_2} t_{\partial_3} t_{\partial_4}$ in $\Gamma_3^4$ for any $\phi$ in $\Gamma_3^4$ fixing the curves $C_1, C_2, C'_1, C'_2$ on $\Sigma_3^4$. Letting $(X_{\phi}, f_{\phi})$ denote the symplectic genus-$3$ Lefschetz pencil corresponding to the positive factorization $W_{\phi}$, we can now state our general result as: \[SCYfamily\] Each $(X_{\phi}, f_{\phi})$ is a symplectic genus-$3$ Lefschetz pencil on a minimal symplectic Calabi-Yau surface, where $X$ is homeomorphic to the $4$-torus for $\phi=1$. The family of genus-$3$ Lefschetz pencils $$\{ (X_{\phi}, f_{\phi}) \ \| \ \phi \in \Gamma_3^4 \text{ and keeps } C_1, C_2, C'_1, C'_2 \text{ fixed} \, \}$$ contains a subfamily $$\{ (X_M, f_M) \ \| \ H_1(X_M) \cong {\mathbb{Z}}^2 \oplus ({\mathbb{Z}}\, / {m_1 \, {\mathbb{Z}}}) \oplus ({\mathbb{Z}}\, / {m_2 \, {\mathbb{Z}}}) \, , \text{ for } \, M=(m_1, m_2) \in {\mathbb{N}}\x {\mathbb{N}}, \}.$$ We will appeal to the following in the proof of our theorem: \[ruledno\] The manifolds $S^2 \x T^2$ and $S^2 {\widetilde}{\x} T^2$ do not admit any genus-$3$ Lefschetz pencils with $4$ base points. We will show that neither one of these ruled surfaces contains an embedded symplectic surface $F$ of genus $3$ and self-intersection $4$, whereas the fiber of a pencil as in the statement of lemma would give such an $F$. For $X= S^2 \x T^2$, $H_2(X) \cong {\mathbb{Z}}^2$ is generated by $S= S^2 \x \{ pt\}$ and $T= \{ pt\} \x T^2$, where $S \cdot S=0$, $T \cdot T=0$, and $S \cdot T=1$. Now let $F= a S + bT$ be a symplectic curve of genus $3$ with $F^2=4$. So $4= F^2= 2ab$, which implies $ab=2$. Since there is a unique symplectic structure on a minimal ruled surface up to deformations and symplectomorphisms [@LiLiu], we can apply the adjunction formula and derive $$4= \eu(F)= F^2 + K_{X} \cdot F = 4 + (-2T) \cdot (aS+bT) = 4- 2a \, ,$$ and get $a=0$. It however contradicts with $ab=2$ above, so there is no such $F$. For $X=S^2 {\widetilde}{\x} T^2$, $H_2(X) \cong {\mathbb{Z}}^2$ is generated by the fiber $S$ and section $T$ of the degree-$1$ ruling on $X$, where $S \cdot S=0$, $T \cdot T=1$, and $S \cdot T=1$. Let $F= a S + bT$ be a symplectic curve of genus $3$ with $F^2=4$. So $4= F^2= 2ab +b^2$, which implies $b(2a+b)=4$. By the same argument as above, we can apply the adjunction to derive $$4= \eu(F)= F^2 + K_{X} \cdot F = 4 + (S-2T) \cdot (aS+bT) = 4 -2a-b \, ,$$ and get $2a+b=0$, which contradicts with $b(2a+b)=4$. So there is no such $F$ in this case either. We can now prove our theorem: For $\phi=1$ we get the pencil $(X,f)$ constructed in the previous section, now with an explicit lift of its monodromy to $\Gamma_3^4$ as a positive factorization ${\widetilde}{P} {\widetilde}{P}'= \Delta$. Since any $(X_{\phi}, f_{\phi})$ can be derived from $(X,f)$ by a partial conjugation along the first factor ${\widetilde}{P}$ in this factorization, $\eu(X_{\phi})= \eu(X)=4$, $\sigma(X_{\phi})=\sigma(X)=-4$, and $c_1^2(X_{\phi})= c_1^2(X)=-4$. By Lemma \[SCYpencil\], we conclude that $X_{\phi}$ is a minimal symplectic Calabi-Yau surface, provided it is not rational or ruled. From $2 - 2 b_1(X_{\phi})+ b^+(X_{\phi}) + b^-(X_{\phi}) = \eu(X_{\phi}) = 4$ and $b^+(X_{\phi})-b^-(X_{\phi}) = \sigma(X_{\phi}) = -4$ we see that $b^+(X_{\phi})=1$ only if $b_1(X_{\phi})=2$. So the only rational or ruled surface with these characteristic numbers can be $S^2 \x T^2$ or $S^2 {\widetilde}{\x} T^2$. However, by the previous lemma, neither one of these surfaces admit a genus-$3$ pencil, so any $X_{\phi}$ is a symplectic Calabi-Yau surface. To prove the second part of the theorem, consider genus-$3$ pencils $(X_{\phi}, f_{\phi})$ with $\phi= t_{b_1}^{-m_1} t_{a_3}^{m_2}$ (where $b_1$ and $a_3$ are as in Figure \[pi1generators\]). Note that $b_1$ and $a_3$ are disjoint from $C_1, C_2, C'_1, C'_2$, so $\phi$ fixes all. For $m= (m_1, m_2) $ any pair of non-negative integers, we denote this pencil by $(X_m, f_m)$, and its monodromy factorization by $W_m= {\widetilde}{P}^\phi {\widetilde}{P}'$, $\phi= t_{b_1}^{-m_1} t_{a_3}^{m_2}$. As we observed in the previous section, every triple of vanishing cycles $\{B_0, B_1, B_2\}$, $\{A_0, A_1, A_2\}$, $\{B'_0, B'_1, B'_2\}$, $\{A'_0, A'_1, A'_2\}$ give the same three relations in $H_1$: $$\begin{aligned} & \phantom{o} & a_1 + a_3= 0, \label{varyingab1} \\ & \phantom{o} & a_1 -b_1 + b_2 + a_3 - b_3= 0, \label{varyingab2} \\ & \phantom{o} & b_1 -b_2 +b_3 =0, \label{varyingab3}\end{aligned}$$ where the first and the third relations imply the second. We can draw two conclusions: First, the vanishing cycles coming from the non-conjugated factor ${\widetilde}{P}'$ induce exactly these relations in $H_1(X_m)$. Second, the vanishing cycles coming from the conjugated factor ${\widetilde}{P}^{\phi}$ induce the following relations we can easily calculate using the Picard-Lefschetz formula: $$\begin{aligned} & \phantom{o} & a_1 + m_1 b_1 + a_3= 0, \label{varyingab4} \\ & \phantom{o} & a_1 + m_1 b_1 -b_1 + b_2 + a_3 - m_2 a_3 - b_3= 0, \label{varyingab5} \\ & \phantom{o} & b_1 -b_2 + m_2 a_3+ b_3 =0 \, . \label{varyingab6}\end{aligned}$$ We easily see that $(\ref{varyingab1})$ and $(\ref{varyingab4})$ imply $m_1 b_1=0$, and $(\ref{varyingab3})$ and $(\ref{varyingab6})$ imply $m_2 a_3=0$, whereas the others are already implied by these relations. From $a_3= -a_1$ and $b_3= b_2-b_1$, we conclude that $H_1(X_m)$ is generated by $a_1, b_1, a_2, b_2$ with $m_1 b_1=0$ and $m_2 a_1 = 0$. Hence $H_1(X_m) \cong {\mathbb{Z}}^2 \oplus ({\mathbb{Z}}\, / {m_1 \, {\mathbb{Z}}}) \oplus ({\mathbb{Z}}\, / {m_2 \, {\mathbb{Z}}})$, as promised. \[CompareSmith\] The very first question here is whether or not every $X_{\phi}$ is a torus bundle over a torus, as they are commonly conjectured to exhaust the list of all symplectic Calabi-Yau surfaces with $b_1 >0$. If for any $\phi$, $\pi_1(X_{\phi})$ is not a $4$-dimensional solvmanifold group [@Hillman; @FriedlVidussi], this would imply that $X_{\phi}$ is not a torus bundle over a torus, and it is a new symplectic Calabi-Yau surface. As our arguments in the proof of Theorem \[SCYfamily\] show, if any partial conjugation along any Hurwitz equivalent factorization to $W_{\phi}$ results in a pencil with a fundamental group which is not a solvmanifold group, we can arrive at a similar conclusion as well. I do not know for the moment if either one of these can be realized or ruled out for all possible $\phi$. It is worth comparing our pencils with those on torus bundles over tori. In [@SmithTorus] Ivan Smith constructs genus-$3$ pencils on torus bundles admitting sections (not all do), by generalizing the algebraic geometric construction of holomorphic genus-$3$ pencils on abelian surfaces. A natural question is whether or not our examples overlap with Smith’s. (Well, if they do, the explicit factorizations $W_{\phi}$ can be seen to prescribe very nice handle decompositions for such torus bundles.) Note that that any torus bundle over a torus with a section $S$ would admit a second disjoint section as well; for any section of a surface bundle over a torus has self-intersection zero [@BaykurKorkmazMonden] and can be pushed-off of itself. So any genus-$3$ Lefschetz fibration of Smith would be determined by a pair $\phi_1, \phi_2 \in \Gamma_1^2$ subject to $[\phi_1, \phi_2]=1$. Now for a comparison, let us note that the stabilizer subgroup of $\Gamma_3^4$ which fix $C_1$, $C_2$, $C'_1$, $C'_2$, is generated by Dehn twists along curves which do not intersect the stabilized collection, and the obvious involution swapping the genus one surfaces bounded by $C_1, C'_2$ and by $C'_1, C_2$. (This involution maps each $A_j, B_j, A'_j, B'_j$ to itself!) So any $\phi$ in Theorem \[SCYfamily\] can be expressed as $\phi= t_{z_k} \cdots t_{z_1}$, where each $z_i$, for $i=1, \ldots, k$, is either contained in the subsurface $F_1 \cong \Gamma_1^2$ in $\Gamma_3^4$ bounded by $C_1$ and $C'_2$, or $F_2 \cong \Gamma_1^2$ bounded by $C_2$ and $C'_1$. Clearly these have disjoint supports, so we can rearrange the indices so that $\phi = \phi_1 \phi_2$, where each $\phi_i$ is compactly supported on $F_i$. Hence we see that $(X_{\phi}, f_{\phi})$ is determined by a pair of mapping classes $\phi_1, \phi_2 \in \Gamma_1^2$, but with no relation to each other whatsoever!. \[CompareHoLi\] As observed in the previous remark, for each $\phi$ in Theorem \[SCYfamily\], we have factorization $\phi= t_{z_k} \cdots t_{z_1}$, where $z_i$ are disjoint from $C_1, C_2, C'_1, C'_2$. So $$W_{\phi}= {\widetilde}{P}^{\phi} {\widetilde}{P}' = ( \phi {\widetilde}{P} \phi^{-1}) \, {\widetilde}{P}' = (t_{z_k} ( \cdots (t_{z_1} {\widetilde}{P} t^{-1}_{z_1}) \cdots) t^{-1}_{z_k}) \, {\widetilde}{P}'$$ is obtained by a sequence of partial conjugations by $t_{z_1}, \ldots, t_{z_k}$, which amounts to performing fibered Luttinger surgeries along Lagrangian tori or Klein bottles that are swept off by $z_i$ on the fibers over a loop on the base [@Auroux; @BaykurLuttingerLF]. The reason we can break down the conjugation of ${\widetilde}{P}$ by $\phi$ to a sequence of partial conjugations by Dehn twists is that at each step the conjugated mapping class $t_{z_{i}}\cdots t_{z_1} {\widetilde}{P}\,t^{-1}_{z_1} \cdots t^{-1}_{z_{i}}$ indeed commutes with $t_{z_{i+1}}$. This is easy to see once we recall that as a mapping class ${\widetilde}{P}= t_{C}^{-2} t_{C'}^2$ and that each $z_i$ is disjoint from $C$ and $C'$. The mapping class $t_{z_{i}}\cdots t_{z_1} {\widetilde}{P}\,t^{-1}_{z_1} \cdots t^{-1}_{z_{i}} = t_{C}^{-2} t_{C'}^2$ certainly commutes with $t_{z_{i+1}}$. Moreover, each $z_i$ is supported away from $\partial_1, \ldots, \partial_4$, so these surgeries can be viewed in the minimal models $X_{\phi}$. Since Luttinger surgery do not change $\omega$ away from these Lagrangians and since the canonical class $K$ can be supported away from them [@ADK], $K^2$ or $K \cdot [\omega]$ do not change under these surgeries. Thus the symplectic Kodaira dimension does not change under these surgeries. This gives an alternative proof (without using Lemma \[ruledno\]) of all $(X_{\phi}, f_{\phi})$ being symplectic Calabi-Yau surfaces, since all are obtained from the symplectic Calabi-Yau $(X,f)$ we constructed in the previous section by partial conjugations. With this said, we observe that if our family $\{ X_{\phi} \}$ coincides with torus bundles over tori, we immediately get a proof of an improved version of a conjecture by Chung-I Ho and Tian-Jun Li: that every symplectic torus bundle over a torus can be obtained from the $4$-torus via Luttinger surgeries along tori [@HoLi], or Klein bottles, as we add. (This conjecture is easier to verify for Lagrangian torus bundles over tori, which however only constitute a subfamily of torus bundles with $b_1 \geq 3$.) All our examples are clearly obtained from $X$, our symplectic Calabi-Yau surface homeomorphic to the $4$-torus, via Luttinger surgeries. The upper bound on the first Betti number of a genus-$g$ fibration ------------------------------------------------------------------ \[Sec:bound\] In [@KorkmazProblems], Korkmaz asked if it is true that $b_1(X) \leq g$ for any non-trivial genus-$g$ Lefschetz fibration $(X,f)$, noting that this is the case for all the examples he knows of —with a single counter-example provided by Smith’s genus-$3$ fibration on $T^4 \# 4 {\overline{\mathbb{CP}}{}^{2}}$. Here we will sketch a generalizing of our construction of a genus-$3$ Lefschetz pencil $(X,f)$ with $b_1=4$ in Section \[Sec:4torus\] to higher genera, which provides a negative answer to Korkmaz’s question, for all odd $g> 1$. ![Vanishing cycles $B_j, C$ in Hamada’s lift of the generalized Matsumoto fibration of genus $2h$.](GeneralizedMatsumotoLifts.eps){width="8cm"} \[GeneralizedMatsumoto\] Our input will be the generalization of Matsumoto’s genus-$2$ fibration constructed by Korkmaz himself [@Korkmaz] (and also by Cadavid [@Cadavid]): for every $h>0$, there is a genus-$2h$ Lefschetz fibration on the ruled surface $S^2 \x \Sigma_h \# 4 {\overline{\mathbb{CP}}{}^{2}}$. As shown by Hamada, the monodromy factorization of these can be lifted to $\Gamma_{2h}^2$ to obtain: $$(t_{B_0} t_{B_1} t_{B_2} \cdots t_{B_{2h-1}} t_{B_{2h}} \, t_C )^2 t^{-1}_{\delta_1} t^{-1}_{\delta_2} = 1 \, ,$$ where the curves $B_0, \ldots, B_{2h}, C$ and the boundary parallel curves $\delta_1, \delta_2$ are as shown in Figure \[GeneralizedMatsumoto\]. Embed this relation into $\Gamma_g$, for $g=2h+1$, by gluing along the two boundary components $\delta_1$ and $\delta_2$ of $\Sigma_{2h}^2$. We can also embed it by rotating the resulting surface by $180$ degrees as before. The product of the two factorizations we get in $\Gamma_{g}$ through these two embeddings prescribes a fiber sum decomposable achiral genus-$g$ Lefschetz fibration. Removing pairs of matching Dehn twists with opposite signs (coming from the two embeddings of $t^2_C$ and $t_{\delta_1}= t_{\delta_2}$) gives us a genus-$g$ Lefschetz fibration $({\widetilde}{X_h},{\widetilde}{f_h})$ with $H_1({\widetilde}{X_h}) \cong {\mathbb{Z}}^{2h+2}$, which can be seen from a straightforward calculation as before. Moreover, similar to our refined construction in Theorem \[SCYfamily\], we could the further lifts of the Cadavid-Korkmaz fibration in $\Gamma_{2h}^4$ to obtain a pencil $(X_h, f_h)$ with $4$ base points, blowing-up which yields $({\widetilde}{X_h},{\widetilde}{f_h})$. We have $b_1(X_h)=b_1({\widetilde}{X_h})= 2h+2=g+1 \nleq g$. With our marginal examples in mind, it seems reasonable to revamp Korkmaz’s estimate and ask if every non-trivial Lefschetz pencil $(X,f)$ has $b_1(X) \leq g+1$. We can thus ask: \[b1RefinedQuestion\] Is the first Betti number $b_1(X)$ of any non-trivial genus-$g$ Lefschetz pencil $(X,f)$ bounded above by $g+1$? What is the supremum for $b_1(X)$ of all genus-$g$ Lefschetz pencils $(X,f)$ for a given $g$? It has been known for a while by now that $b_1(X) \leq 2g-1$ for any non-trivial genus-$g$ Lefschetz pencil/fibration $(X,f)$ [@ABKP]. The first part of Question \[b1RefinedQuestion\] asks if $g+1$ can be a sharper bound, still linear in $g$. The second part of Question \[b1RefinedQuestion\], which is formulated in the most general way, can already be answered for low genera pencils: for $g=1$ this supremum is clearly $0$, and it is $2$ for $g=2$, which can be easily shown using the Kodaira dimension (invoking [@SatoKodaira], as in Remark \[smallest\]). For $g=3$, we conjecture that the supremum is indeed $4=g+1$. Note that, the analogous question for the supremum for the second Betti number $b_2(X)$ of relatively minimal genus-$g$ pencils $(X,f)$ is now fully answered. It is $11$ for $g=1$, realized by the pencil on ${{\mathbb{CP}}{}^{2}}\# \, 8 {\overline{\mathbb{CP}}{}^{2}}$; $37$ for $g=2$, following from Smith’s analysis of genus-$2$ pencils in [@SmithGenus3]; whereas the supremum does not exist when $g \geq 3$, which is implied by the arbitrarily long positive factorizations of the boundary multi-twist produced in [@BaykurVHM; @DalyanKorkmazPamuk; @BaykurMondenVHM]. However, constructions of fixed genus pencils with large $b_2$ is rather antagonistic to that of fixed genus pencils with large $b_1$: larger $b_2$ means more vanishing cycles (typically) killing $b_1$. [99999]{} A. Akhmedov, R. I. Baykur and D. Park, Constructing infinitely many smooth structures on small $4$-manifolds, J. Topol. 1 (2008), no. 2, 409–-428. A. Akhmedov and B. D. Park, Exotic smooth structures on small $4$-manifolds with odd signatures, Invent. Math. 181 (2010), no. 3, 577–-603. J. Amoros, F. Bogomolov, L. Katzarkov, and T. Pantev. Symplectic Lefschetz fibrations with arbitrary fundamental groups, with an appendix by I. Smith, J. Differential Geom. vol 54 (2000): 489–-545. S.Ju. Arakelov, Families of algebraic curves with fixed degeneracies, Izv. Akad. Nauk. SSSR Ser. Mat. 35 (1971) 1269–1293. D. Auroux, Mapping class group factorizations and symplectic $4$-manifolds: some open problems, Problems on mapping class groups and related topics, 123–132, Proc. Sympos. Pure Math. 74, Amer. Math. Soc., Providence, RI, 2006. D. Auroux, S. K. Donaldson and L. Katzarkov, Luttinger surgery along Lagrangian tori and non-isotopy for singular symplectic plane curves, Math. Ann. 326 (2003), no. 1, 185–-203. S. Baldridge and P. Kirk, A symplectic manifold homeomorphic but not diffeomorphic to ${{\mathbb{CP}}{}^{2}}\# 3 {\overline{\mathbb{CP}}{}^{2}}$, Geom. Topol. 12, 919–-940. S. Bauer, Almost complex $4$-manifolds with vanishing first Chern class, J. Differential Geom. 79 (2008), no. 1, 25–32. R. I. Baykur, Minimality and fiber sum decompositions of Lefschetz fibrations, Proc. Amer. Math. Soc.; proc/12835 (2015). R. I. Baykur, Non-holomorphic surface bundles and Lefschetz fibrations, Math. Res. Lett. 19 (2012), no. 3, 567–-574. R. I. Baykur, Inequivalent Lefschetz fibrations and surgery equivalence of symplectic $4$-manifolds, to appear in J. Symplectic Geom. ; http://arxiv.org/abs/1408.4869. R. I. Baykur and K. Hayano, Multisections of Lefschetz fibrations and topology of symplectic $4$-manifolds, to appear in Geom. Topol.; http://arxiv.org/abs/1309.2667. R. Inanc Baykur and M. Korkmaz, Small Lefschetz fibrations and exotic $4$-manifolds, preprint; http://arxiv.org/abs/1510.00089. R. Inanc Baykur and M. Korkmaz, . R. I. Baykur, M. Korkmaz and N. Monden, Sections of surface bundles and Lefschetz fibrations, Trans. Amer. Math. Soc. 365 (2013), no. 11, 5999–6016. R. I. Baykur, N. Monden and M. Korkmaz, Positive factorizations of mapping classes, preprint; http://arxiv.org/abs/1412.0352. R. I. Baykur and N. Sunukjian, Round handles, logarithmic transforms, and smooth $4$-manifolds, J. Topol., Vol. 6 No 1 (2013), 49–63. R. I. Baykur and J. Van Horn-Morris, Topological complexity of symplectic $4$-manifolds and Stein fillings, to appear in J. Symplectic Geom. C. Cadavid, A remarkable set of words in the mapping class group, Dissertation, Univ. of Texas, Austin, 1998. E. Dalyan, M. Pamuk and M. Korkmaz, Arbitrarily long factorizations in mapping class groups, to appear in Int. Math. Res. Notices. S. K. Donaldson, Irrationality and the $h$-cobordism conjecture, J. Differential Geom. 26 (1987), 141–-168. S. K. Donaldson, Lefschetz pencils on symplectic manifolds, J. Differential Geom. vol. 53 (1999), no.2, 205–236. S. K. Donaldson, Lefschetz pencils and mapping class groups, Problems on mapping class groups and related topics, 151–-163, Proc. Sympos. Pure Math., 74, Amer. Math. Soc., Providence, RI, 2006. S. K. Donaldson, Some problems in differential geometry and topology, Nonlinearity 21 (2008), no. 9, T157–T164. J. Dorfmeister, Kodaira dimension of fiber sums along spheres, Geom. Dedicata, March (2014), DOI 10.1007/s10711-014-9974-2. H. Endo, Meyer’s signature cocycle and hyperelliptic fibrations, Math. Ann., 316 (2000), 237–257. F. T. Farrell and L. E. Jones, Classical aspherical manifolds, CBMS Regional Conference Series in Mathematics, 75. Published for the Conference Board of the Mathematical Sciences, Washington, DC; by the American Mathematical Society, Providence, RI (1990). R. Fintushel and R. Stern, Rational blowdowns of smooth $4$-manifolds, J. Differential Geom. 46 (1997), no. 2, 181–235. R. Fintushel and R. Stern, Knots, links, and $4$-manifolds, Invent. Math. vol. 134 (1998), no. 2, 363–400. R. Fintushel and R. J. Stern, Double node neighborhoods and families of simply connected $4$-manifolds with $ b^+=1$, [J. Amer. Math. Soc.]{} [19]{} (2006), 171–180. R. Fintushel, B. D. Park and R. Stern, Reverse engineering small $4$-manifolds, Algebr. Geom. Topol. 7 (2007) 2103-–2116. R. Fintushel and R. Stern, Pinwheels and nullhomologous surgery on $4$-manifolds with $b^+=1$, Algebr. Geom. Topol. 11 (2011), no. 3, 1649–-1699. M. H. Freedman, The topology of four-dimensional manifolds, J. of Differential Geometry. 17 (3): 357–-453. S. Friedl and S. Vidussi, On the topology of symplectic Calabi-Yau $4$-manifolds, J. Topol. 6 (2013), no. 4, 945–-954. R. Friedman, Vector bundles and ${\rm SO}(3)$-invariants for elliptic surfaces, J. Amer. Math. Soc. [8]{} (1995), 29–139. R. Friedman and J. W. Morgan, Smooth four-manifolds and complex surfaces, Ergeb. Math. Grenzgeb. (3), 27, Springer-Verlag, Berlin, 1994. R. E. Gompf, A new construction of symplectic manifolds, Ann. of Math. (2) 142 (1995), no. 3, 527–595. R. Gompf and A. Stipsicz, $4$-Manifolds and Kirby Calculus, Graduate Studies in Mathematics, vol. 20, American Mathematical Society, Rhode Island, 1999. N. Hamada, Sections of the Matsumoto-Cadavid-Korkmaz Lefschetz fibration, preprint. C.-I. Ho and T.-J. Li, Luttinger surgery and Kodaira dimension, Asian J. Math. 16 (2012), no. 2, 299–-318. J. A. Hillman, Four-manifolds, geometries and knots. Geom. & Topol. Monographs, 5. Geom. & Topol. Publications, Coventry, 2002. xiv$+379$ pp. A. Kas, On the handlebody decomposition associated to a Lefschetz fibration, Pacific J. Math. 89 (1980), 89–104. M. Korkmaz, Noncomplex smooth $4$-manifolds with Lefschetz fibrations, Internat. Math. Res. Notices 2001, no. 3, 115–128. M. Korkmaz, Lefschetz fibrations and an invariant of finitely presented groups, Internat. Math. Res. Notices 2009, no. 9, 1547–1572. M. Korkmaz, Problems on homomorphisms of mapping class groups, Problems on mapping class groups and related topics, 81–-89, Proc. Sympos. Pure Math., 74, Amer. Math. Soc., Providence, RI, 2006. D. Kotschick, On manifolds homeomorphic to ${{\mathbb{CP}}{}^{2}}\# 8 {\overline{\mathbb{CP}}{}^{2}}$, Invent. Math. 95 (1989), 591–-600. T.-J. Li, Symplectic $4$-manifolds with Kodaira dimension zero, J. Differential Geom. 74 (2006), no. 2, 321–352. T.-J. Li, The Kodaira dimension of symplectic $4$-manifolds, Floer homology, gauge theory, and low-dimensional topology, 249–261, Clay Math. Proc., 5, Amer. Math. Soc., Providence, RI, 2006. T.-J. Li, Quaternionic vector bundles and Betti numbers of symplectic $4$-manifolds with Kodaira dimension zero, Int. Math. Res. Not. 2006, 1–28. T.-J. Li, Symplectic Calabi-Yau surfaces, Handbook of geometric analysis, No. 3, 231–356, Adv. Lect. Math. (ALM), 14, Int. Press, Somerville, MA, 2010. T.-,J. Li and A. Liu, Symplectic structures on ruled surfaces and a generalized adjunction formula, Math. Res. Lett. 2 (1995), 453–471. Y. Matsumoto, Lefschetz fibrations of genus two - a topological approach -, Proceedings of the 37th Taniguchi Symposium on Topology and Teichmüller Spaces, (S. Kojima, et. al., eds.), World Scientific, 1996, 123–148. B. Moishezon, Complex surfaces and connected sums of complex projective planes, Lecture Notes in Math. 603, Springer-Verlag, 1977. B. Ozbagci, Signatures of Lefschetz fibrations, Pacific J. Math. 202 (2002), no.1, 99–118. B. Ozbagci and A. Stipsicz, Noncomplex smooth $4$-manifolds with genus-$2$ Lefschetz fibrations, Proc. Amer. Math. Soc. 128 (2000), no. 10, 3125–-3128. B. Ozbagci and A. Stipsicz, Contact $3$-manifolds with infinitely many Stein fillings, Proc. Amer. Math. Soc. 132 (2004), no. 5, 1549–-1558. J. Park, Simply connected symplectic $4$-manifolds with $b^+_2 = 1$ and $c_1^2 = 2$, Invent. Math. 159 (2005), 657-–667. B. D. Park, Exotic smooth structures on $3 {{\mathbb{CP}}{}^{2}}\# n {\overline{\mathbb{CP}}{}^{2}}$, Proc. Amer. Math. Soc. 128 (2000), 3057–-3065. A. N. Parshin, Algebraic curves over function fields, I, Izv. Akad. Nauk. SSSR Ser. Mat. 32 (1968) 1191–1219. Y. Sato, Canonical classes and the geography of nonminimal Lefschetz fibrations over $S^2$, Pacific J. Math., 262 (2013), no. 1, 191–226. I. Smith, Lefschetz pencils and divisors in moduli space, Geom. Topol. 5 (2001), 579-–608. I. Smith, Torus fibrations on symplectic four-manifolds, Turk. J. Math. 25 (2001) , 69–95. R. Stern, Will we ever classify simply-connected smooth 4-manifolds?, Floer homology, gauge theory, and low-dimensional topology, 225–239, Clay Math. Proc., 5, Amer. Math. Soc., Providence, RI, 2006. A. Stipsicz, Indecomposability of certain Lefschetz fibrations, Proceedings of the American Mathematical Society 129 (2001), no. 5, 1499–1502. A. I. Stipsicz and Z. Szabó, An exotic smooth structure on ${{\mathbb{CP}}{}^{2}}\#6{\overline{\mathbb{CP}}{}^{2}}$, Geom. Topol. [9]{} (2005), 813–832. M. Symington, Symplectic rational blowdowns, J. Differential Geom. 50 (1998), no. 3, 505-–518. Z. Szabó, Exotic $4$-manifolds with $b_2 ^+ =1$, Math. Res. Lett. [3]{} (1996), 731–741. C. H. Taubes, The Seiberg–Witten and Gromov invariants, Math. Res. Lett. 2:2 (1995), 221–238. C. H. Taubes, SW $\Rightarrow$ Gr: From the Seiberg-Witten equations to pseudo-holomorphic curves, Journal of the AMS, 9 (1996), 845–918. M. Usher, The Gromov invariant and the Donaldson-Smith standard surface count, Geom. Topol. 8 (2004), 565–-610. M. Usher, Minimality and symplectic sums, Int. Math. Res. Not. 2006, Art. ID 49857, 17 pp. M. Usher, Kodaira dimension and symplectic sums, Comment. Math. Helv. 84 (2009), no. 1, 57–85. [^1]: A couple other examples of SCYs, such as symplectic $S^1$-bundles over $T^2$-bundles over $S^1$, although not known to be torus bundles themselves, are finitely covered by them [@FriedlVidussi]. [^2]: Only Luttinger surgeries along tori are considered in [@HoLi], but it is natural to include surgeries along Klein bottles, too. [^3]: In contrast, there are many non-holomorphic fibrations obtained by twisted fiber sums; e.g. [@OzbagciStipsicz; @Korkmaz; @BaykurHolomorphic]. However decomposable fibrations do not admit $(-1)$-sections —which could be blown-down to produce pencils [@StipsiczFiberSum].	{ "pile_set_name": "ArXiv" }
--- abstract: 'With a simple phenomenological metapopulation model, which characterizes the invasion process of an influenza pandemic from a source to a subpopulation at risk, we compare the efficiency of inter- and intra-population interventions in delaying the arrival of an influenza pandemic. We take travel restriction and patient isolation as examples, since in reality they are typical control measures implemented at the inter- and intra-population levels, respectively. We find that the intra-population interventions, e.g., patient isolation, perform better than the inter-population strategies such as travel restriction if the response time is small. However, intra-population strategies are sensitive to the increase of the response time, which might be inevitable due to socioeconomic reasons in practice and will largely discount the efficiency.' author: - 'Lin Wang, Yan Zhang, Tianyi Huang, and Xiang Li[^1]' title: 'Estimating the value of containment strategies in delaying the arrival time of an influenza pandemic: A case study of travel restriction and patient isolation' --- Introduction ============ During the past decades, extensive efforts have been made to investigate the spread of epidemics. Besides various epidemiological models having been proposed to explore virus transmission in a closed population[@ANDERSONMAY], the study of network spreading uses structured populations to understand the evolution of epidemics in more realistic social settings[@PR424175; @BBV08; @RMP801275; @NP832]. These studies have contributed a great deal of insightful findings, such as the absence of epidemic threshold in scale-free networks[@PRE63066117], the reaction-diffusion process, and metapopulation[@MATHBIO753; @PRE84041936], to name a few. These significant advances have raised new issues on how to limit or control the spread of infectious diseases in human society. To curb the spatial spread of diseases from city to city, a variety of strategies are recommended according to World Health Organization(WHO) or United States(US) response plans[@WHO]: (i) Vaccination of prior groups or dynamic mass vaccination; (ii) antiviral drugs for prophylaxis and treatment; (iii) community-based prevention and control; and (iv) travel-related containment measures. Except for the fact that travel-related measures are implemented at the inter-city level, other strategies are mainly performed at the intra-city level. The first two pharmaceutical interventions cut down the number of potential susceptibles or allay the virus transmission rate, respectively. Community-based strategies might affect individuals (e.g., patient isolation, self-isolation, quarantine), groups, or entire communities (e.g., cancellation of public gatherings, school closures) in a city. Travel-related measures mainly result in the restriction or cancellation of nonessential trips. By supposing that the outbreak of a pandemic is underway, many works have studied the efficiency of strategies by using the metapopulation model, which harnesses the reaction-diffusion framework to sketch human daily contacts and mobility. The epidemic reaction takes place inside each subpopulation due to personal contacts, and the infectious disease cascades subpopulation by subpopulation via the travel of individuals (here each city is represented by a subpopulation). The importance of various strategies in decreasing the attack rate or prevalence has been extensively studied in Refs. [@SCI326729; @SCI3091083] mainly by computational simulations. Particularly, by analyzing the delay of arrival time of the disease [@PLOSMED3e212; @NatMed12497; @PLoSONE2e401; @PLoSONE6e16591; @MB21470], it has been shown that the efficiency of travel restriction in slowing down the international spread of pandemic influenza is limited. In these seminal works, the intra- and inter-population interventions are seldom compared with each other to provide a holistic picture about their value in delaying disease invasion. This should give us pause for thought. Whether it is reasonable to discard the tactic of travel restriction might also depend on how good the intra-population strategies perform. In an attempt to study this issue, we theoretically analyze the efficiency of two kinds of typical containment strategies, namely, travel restriction and patient isolation, which are implemented at the inter- and intra-population levels, respectively. We mainly use a simple phenomenological model following Refs. [@MB21470; @PLoSONE2e143], which considers the importation of an infectious disease from a source to a region at risk during the early stage of a pandemic outbreak. Since the spreading process cascades subpopulation by subpopulation, this two-subpopulation version[@PLoSONE6e16591] is a simple model but rational approximation of the initial stage of the pandemic. We mainly focus on the impact of strategies to delay the arrival time of disease in the subpopulation at risk, because no outbreak will occur in an unaffected region before the introduction of infectious seeds. After the disease lands in the subpopulation, the ongoing endogenous transmission will become the mainstream of infections[@SCI3091083; @PLoSONE6e16591]. Thus the first arrival time of infectious travelers is an important quantity characterizing the timing of the disease outbreak[@MB21470; @JSM09001; @JTB251509]. Model description ================= To build the model, we first specify the mechanism of individual mobility between subpopulations $x,y$. Following Refs. [@MB21470; @PLoSONE2e143; @JSM09001; @JTB251509], at every time step, each individual may travel from his current location $x$($y$) to a neighboring subpopulation $y$($x$) with a per capita diffusion rate $\omega_{xy}$($\omega_{yx}$). We define the unit time as 1 day. The model proceeds with discrete time steps. In reality, the amount of transportation flows, e.g., air traffic, between cities is often symmetric[@JSM09001; @JTB251509; @PNAS1068847], which indicates a detailed balance for the traffic flows. For simplicity, we assume that the subpopulations $x,y$ have the same population size $N_{x}=N_{y}=N$ and diffusion rate $\omega_{xy}=\omega_{yx}=\omega$. Thus there are on average $\omega N$ individuals departing from each subpopulation per day. Note that relaxing these two restrictions does not change the main results of this Brief Report as long as we maintain a detailed balance condition. While mobility couples different locations, the epidemic reaction process occurs in each subpopulation, where the population is mixing homogeneously. We consider a standard susceptible-infective-removed (SIR) compartment model to represent the influenza-like illness[@MATHBIO753; @PRE84041936; @SCI326729]. At a given time $t$, the number of susceptible, infectious, and recovered individuals in $x(y)$ are defined as $S_x(t),I_x(t),R_x(t)(S_y(t),I_y(t),R_y(t))$, respectively. The SIR reaction is governed by the transition rates $\mu$ and $\beta$[@ANDERSONMAY]. In a unit time, an infectious one recovers and becomes immune at the rate $\mu$. The parameter $\beta$ characterizes disease transmissibility, which reflects the combined factors of the virus transmission rate and individual contact rate per unit time[@PRE84041936]. A susceptible individual might acquire infection by contact with infectious ones staying in the same subpopulation. With the mean-field approximation, at time $t$, the probability for a susceptible one in subpopulation $x(y)$ to acquire infection is found by multiplying the density of infectious $I_{x}(t)/N(I_y(t)/N)$ by $\beta$[@ANDERSONMAY]. In this baseline case, the transfer of susceptible and infectious individuals is ruled by the diffusion rate $\omega$. The epidemic threshold is determined by the basic reproductive number $R_0=\beta/\mu$, which identifies the expected number of secondary infections produced by an infected individual during his infectious period in an entire susceptible population[@ANDERSONMAY]. We next specify the dynamics under interventions. Since many socioeconomic factors might defer the implementation of strategies, we define a response time $t_0$ representing the time interval between the actual inception of an outbreak and the time when the strategies become available. Travel restriction (TR) mainly affects individual mobility between two subpopulations. We define the parameter $\alpha$ as the intensity of TR, which means that a reduction of fraction $\alpha$ in travel begins at time $t_0$, i.e., in the model, we decrease the diffusion rate from $\omega$ to $(1-\alpha)\omega$ after time $t_0$. Patient isolation (PI) mainly impacts individual compartment transitions. The effect of PI may relate to enforcement by local authorities, or is attributed to the self-isolation of infected individuals. For simplicity, we do not distinguish between these two aspects. The parameter $\eta$ is defined to reflect the intensity of PI. It means that on average a fraction $\eta$ of infectious persons will be isolated per unit time after $t_0$. We introduce the PI by adding an isolation process that each infectious one has a likelihood to be isolated with rate $\eta$ per unit time. Since these isolating individuals have little chance to cause infection, we remove them as long as they are isolated. ANALYTICAL AND SIMULATION RESULTS ================================= Initially, an infectious individual is introduced into subpopulation $x$. Thus the initial condition is $I_x(0)$=1,$I_y(0)$=0. We first analyze the efficiency of TR in slowing down disease invasion to subpopulation $y$. The key issue is to evaluate its impact on delaying the first arrival time(FAT) of infectious travelers from $x$. With the Poisson process assumption that the diffusion of any individual is independent from that of others, the probability that the first infectious individual arrives in subpopulation $y$ at time $t^{y}=t$ is $$P(t^{y}=t)=[1-(1-\omega)^{I_{x}(t)}]\prod^{t-1}_{t_{i}=1}(1-\omega)^{I_{x}(t_{i})},\label{eq.1}$$ which describes that at least one successful transfer of infectious individuals from subpopulation $x$ to $y$ occurs at time $t$, and none at previous time steps[@JSM09001; @JTB251509]. In reality, it is general that the number of travelers per day is several orders of magnitude smaller than the total population of a city, where only small amounts of people leave to travel per day. Empirical evidence of worldwide or US domestic air transportation[@MATHBIO753] suggests that the daily diffusion rate of individuals on each flight route is of the order $10^{-4}$ or less. We here assume $\omega=10^{-4}$,$N=10^6$. Using the Taylor expansion, Eq. (\[eq.1\]) becomes $P(t^{y}=t)=\omega I_{x}(t)\exp[-\omega\sum_{0<t_{i}<t}I_{x}(t_{i})]$. Based on many seminal works[@SCI326729; @SCI3091083; @PLOSMED3e212; @NatMed12497; @PLoSONE2e401; @PLoSONE6e16591; @MB21470], we assume a pandemic influenza with $R_0=1.75$ and the infectious period $\mu^{-1}=3$ days. In this case, the Malthusian parameter $\lambda$, the real-time exponential growth rate at the early stage of an outbreak[@PRL103038702; @Interface7873], is $\beta-\mu=0.25$. Since $\omega\!\ll\!\lambda$, the SIR reaction happens at a time scale much faster than the diffusion process, thus the number of infectious individuals in subpopulation $x$ grows sufficiently before subpopulation $y$ is invaded. Meanwhile, at this early stage, the infectious ones only make up a small fraction of the total population in $x, I_x(t)\!\ll\!N$. With a mean-field approximation for the evolution of infectious individuals, we have[@BBV08; @JSM09001; @JTB251509] $I_{x}(t_{i})\!\simeq\!I_{x}(0)\exp(\lambda t_{i})$, $t_{i}\!\le\!t^{y}$. Using the continuum approximation $\sum_{0<t_{i}<t}I_{x}(t_{i})\!=\!\int^{t}_{0}d\tau I_{x}(\tau)$, we obtain the probability density of FAT, $P(t)\!=\!\omega\exp[\lambda t-(\omega/\lambda)\exp(\lambda t)]$, with the mean value $<\!t^{F}\!>=(1/\lambda)(\ln(\lambda/\omega)-\gamma)$[@JSM09001; @JTB251509], where $\gamma$ is the Euler constant. With the above given parameters, this characteristic time scale of FAT is $<\!t^F\!>\simeq29$ days. ![image](Graph1.eps){width="5in"} In the TR scenario, when the FAT is smaller than the response time $t_0$, the probability density of FAT is still $P(t)$; however, when the FAT is larger than $t_0$, this probability density becomes $P_{\alpha}(t)=\big{(}1-(1-(1-\alpha)\omega)^{I_{x}(t)}\big{)}\prod_{0<t_{i}<t_{0}}(1-\omega)^{I_{x}(t_{i})}\prod_{t_{0}\le t_{j}<t}\big{(}1-(1-\alpha)\omega)^{I_{x}(t_{j})}\simeq(1-\alpha)\omega\exp[\lambda t-(1-\alpha)\omega\exp(\lambda t)/\lambda -\alpha\omega\exp(\lambda t_{0})/\lambda]$. We numerically calculate the average FAT through $<\!t_{\alpha}^{F}\!>=\!\int_0^{t_0}\tau P(\tau)d\tau+\int_{t_{0}}^{\infty}\tau P_\alpha(\tau)d\tau$, and get the delay of FAT, $\Delta t(\alpha)$, by solving $$\Delta t(\alpha)=<\!t^{F}_{\alpha}\!>-<\!t^{F}\!>.\label{eq.2}$$ If the response time $t_0$ is negligible ($t_0\!=\!0$), Eq.(\[eq.2\]) is simplified as $$\Delta t(\alpha)\|_{t_{0}=0}\!=\!-\ln(1-\alpha)/\lambda,\label{eq.3}$$ which recovers the results obtained by the cumulative probability $P(t^{y}\leq t)$ in Refs. [@PLoSONE6e16591; @MB21470]. Note that Eq (\[eq.3\]) is independent from the values of $\omega,N$. With $\lambda$=0.25, unless the intensity $\alpha$ is increased to an unpractically high level ($\alpha>0.97$), $\Delta t(\alpha)$ cannot be longer than 2 weeks. To study the PI scenario, we first consider the case where the FAT is larger than $t_0$. At this early stage, we still have the approximation $I_x(t_{i})\!\simeq\!\exp(\lambda t_{i})$ when time $t_i\!\le\!t_0$; after $t_0$, the Malthusian parameter becomes $\lambda_{\eta}\!=\!\lambda-\eta$, and thus we have $I'_x(t_j)\!\simeq\!\exp(\eta t_0)\exp(\lambda_{\eta}t_{j})$ when $t_{0}\!<\!t_{j}\!\le\!t^y$. The probability density in this case is $P_{\eta}(t)=(1-(1-\omega)^{I'_{x}(t)})\prod_{0<t_{i}\le t_{0}}(1-\omega)^{I_{x}(t_{i})}\prod_{t_{0}<t_{j}<t}(1-\omega)^{I'_{x}(t_{j})}\simeq\omega I'_{x}(t)\exp(-\omega\int^{t_{0}}_{0}I_{x}(\tau)d\tau)\exp(-\omega\int_{t_{0}}^{t}I'_{x}(\tau)d\tau) =\omega\exp[\Theta(t_{0})]\exp[\lambda_{\eta} t\!-\!\omega\exp(\eta t_{0}\!+\!\lambda_{\eta}t)/\lambda_{\eta}]$, where $\Theta(t_{0})\! =\!\eta t_{0}\!-\!\omega\exp(\lambda t_{0})/\lambda\!+\!\omega\exp(\lambda t_{0})/\lambda_{\eta}$. If the response time is negligible ($t_{0}$=0), we simplify the former expression as $P_{\eta}(t)\|_{t0=0}\!\simeq\!\omega\exp[\lambda_{\eta}t\!-\!\omega \exp(\lambda_{\eta}t)/\lambda_{\eta}]$, which leads to the average FAT, $<t_{\eta}^{F}>\|_{t_{0}=0}\!=\!\int_{0}^{\infty}\tau P_{\eta}(\tau)\|_{t0=0}d\tau\!\simeq\!(\ln(\lambda_{\eta}/\omega)-\gamma)/\lambda_{\eta}$. In this case, we get the relation between $\Delta t$ and $\eta$ by solving the equation $$\Delta t(\eta)\|_{t_{0}=0}=<\!t^{F}_{\eta}\!>\|_{t_{0}=0}\ -<\!t^{F}\!>\label{eq.4}$$ If $t_0>0$, the average FAT is numerically integrated via the equation $<\!t_{\eta}^{F}\!>=\int_0^{t_0}\tau P(\tau) d\tau+\int_{t_{0}}^{\infty}\tau P_{\eta}(\tau)d\tau$. We therefore have the relation between $\Delta t$ and $\eta$ as $$\Delta t(\eta)=<\!t_{\eta}^{F}\!>-<\!t^F\!>.\label{eq.5}$$ With Eq. (\[eq.4\]) and $\lambda=0.25$, we find that an intermediate level of the strategy intensity $\eta=0.12$ can adequately suspend the arrival of disease to subpopulation $y$ for more than 3 weeks. When the response time $t_{0}=0$, we conclude that the strategy of PI performs better than the TR. This is mainly because the TR alone can not mitigate the initial exponential growth of infectious ones in the source. However, the strategy of PI is highly sensitive to the increase of the response time $t_{0}$. As shown in Fig. \[fig.1\], when $t_0$ increases from 0 to 20 days, there is an evident decline for the delay $\Delta t(\eta)$ in the PI scenario, while the delay $\Delta t(\alpha)$ actualized by implementing the TR is robust to the increase of $t_{0}$. ![(Color online) The relation between the delay of FAT, $\Delta t(\alpha)$, and the intensity of travel restriction. The gray stars are the analytical results with $t_0=0$. The other colored symbols are the simulation results with various response times $t_0$=0,10,15,20, and 25 days.[]{data-label="fig.2"}](Graph2.eps){width="2.5in"} We further use the dynamic Monte Carlo method to simulate the epidemic evolution under different interventions. The simulations are performed with discrete time steps, and we update each individual’s behavior in parallel per unit time. The parameters are $N=10^6$, $\omega=10^{-4}$, $R_0=1.75$, and $\mu^{-1}=3$ days. Initially, an infectious individual is introduced into subpopulation $x$, and thus the initial condition is $I_{x}(0)$=1,$I_{y}(0)$=0. When the containment strategies are excluded, the epidemic reaction and diffusion at each unit time proceed as follows. (i) Reaction: Inside each subpopulation, individuals are mixing homogeneously. At time $t$, the probability for any susceptible in subpopulation $x(y)$ to acquire infection is $\beta I_{x}(t)/N(\beta I_{y}(t)/N)$. The number of new infections in $x(y)$ at time $t$ is extracted from a binomial distribution with probability $\beta I_{x}(t)/N(\beta I_{y}(t)/N)$ and the number of trials $S_{x}(t)(S_y(t))$. The number of recovered individuals in $x(y)$ is also extracted from a binomial distribution with probability $\mu$ and the number of trials $I_x(t)(I_y(t))$. (ii) Diffusion: After all individuals have been updated for the reaction, we simulate their diffusion. The number of susceptible travelers departing from each subpopulation per unit time is also extracted from a binomial distribution with probability $\omega$ and the number of trials $S_{x}(t)(S_{y}(t))$. The number of infectious and recovered travelers is obtained in the same way. We first study the effects of TR in delaying the arrival of disease to subpopulation $y$. To assemble this factor into the simulation, we rescale the per capita diffusion rate $\omega$ by a multiplier $1-\alpha$, where the parameter $\alpha$ reflects the intensity of TR. The strategy is activated after a given response time $t_{0}$. Figure \[fig.2\] provides a holistic view about the relation between the delay of FAT, $\Delta t(\alpha)$, and the restriction intensity $\alpha$. Since the disease might die out due to randomness, every data point is obtained by averaging the simulations with the successful transfer of infectious ones among $10^4$ times of Monte Carlo random experiments, each of which is simulated with 500 time steps. The gray stars are the analytical results obtained by Eq. (\[eq.3\]), which agree well with the simulations. When $\alpha=0.3,0.6$, and 0.9 and $t_0=0$, the simulations show that $\Delta t(\alpha)\simeq 2,4$, and 11 days, respectively. Even if the restriction intensity is elevated to an unpractically high level, e.g., $\alpha=0.97$, $\Delta t(\alpha)$ is still less than 3 weeks. It is clear that $\Delta t$ is small if the time scale of the initial exponential growth $1/\lambda$ is small \[see Eq. \[eq.3\]\]. We further study the impact of the response time on the efficiency of TR. In Fig. \[fig.2\], unless $t_{0}$ approaches $<\!t^{F}\!>$, which is the average FAT without TR, and $\alpha$ is large, there is no evident decline for the simulation results of $\Delta t(\alpha)$. We next study the effects of PI in delaying disease invasion. To introduce this factor in the model, we add an isolation process before the reaction process at each time step after $t_0$. The parameter $\eta$ reflects the intensity of PI. Per unit time, the number of newly isolated individuals in subpopulation $x(y)$ is extracted from a binomial distribution with probability $\eta$ and the number of trials $I_x(t)(I_y(t))$. Figure \[fig.3\](a) presents the relation between the delay of FAT, $\Delta t$, and the isolation intensity $\eta$ with $t_0=0$. For each $\eta$, we perform $10^4$ times of Monte Carlo random experiments, each of which is simulated with 500 time steps. Due to the randomness embedded in the dynamical process, the infectious individuals in source $x$ might be totally eradicated before traveling to subpopulation $y$. With a given $\eta$, we measure $\Delta t(\eta)$ by averaging the simulations that the infectious ones from source $x$ successfully jump to subpopulation $y$. The results are highlighted by the red squares in Fig. \[fig.3\](a). The gray stars are the analytical results obtained by Eq.(\[eq.4\]). If the isolation intensity $\eta$ is at a small or intermediate level($\eta\le0.18$), the simulation results agree well with the theoretical predications. However, if the intensity $\eta$ is extremely large, the simulations obviously deviate from the analytical results. In this latter case, since the Malthusian parameter $\lambda_{\eta}$ is quite small, there is a huge likelihood of totally eradicating the infectious individuals at the early stage of an outbreak due to randomness. For instance, when $\eta=0.2,0.22$, the fraction of eradication in all independent modeling realizations reaches 97.7$\%$ and 99.2$\%$, respectively, while for $\eta=0.12$, the fraction of eradication is only 75.6$\%$ \[see the dark cyan diamonds in Fig.\[fig.3\](a)\]. If $\eta\ge0.25$, the Malthusian parameter $\lambda\leq 0$, the disease hardly persists in the population. With the same condition that $t_0=0$, the strategy of PI is more efficient than TR: An intermediate level of isolation intensity $\eta$ can adequately delay the arrival of disease for about 1 month. ![image](Graph3.eps){width="5.5in"} Figure \[fig.3\](b) shows the impact of the response time $t_0$ on the delaying effects of PI. For a small $t_0$, e.g., $t_0=10$, which is much smaller than $<\!t^F\!>$, an intermediate level of PI(e.g., $\eta=0.14$) still suspends the arrival of disease for about 3 weeks. This achievement exceeds the performance of TR even with an extremely high restriction intensity. The simulations also illuminate that the PI is sensitive to the increase of $t_0$. There is a remarkable decline in the simulation results of $\Delta t(\eta)$ when $t_0$ approaches $<\!t^F\!>$. For instance, 25 days of waiting to implement the strategy ($t_{0}=25$) will only postpone the arrival of disease in subpopulation $y$ for about 2 weeks at most. Actually, other intra-population interventions can also be analyzed under this framework. For instance, social distancing limits public activities to reduce personal contacts, which can be reflected by rescaling the disease transmission rate $\beta$ with a multiplier $1-\varphi$ when time $t\ge t_0$. At the initial stage of an outbreak, the Malthusian parameter becomes $\lambda_{\varphi}\!=\!(1-\varphi)\beta-\mu$. From a mathematical point of view, we can adjust the parameters $\varphi,\eta$ to allow $\lambda_\varphi\!=\!\lambda_{\eta}$. Therefore, the above analysis can cover this scenario. SUMMARY ======= In sum, the intra-population interventions, e.g., patient isolation, perform better than the inter-population strategies such as travel restriction if the response time is small. Therefore, the intra-population strategies are more beneficial in delaying the spatial spread of pandemic influenza if they are implemented very promptly. However, the intra-population measures are sensitive to the increase of response time, which might be inevitable due to miscellaneous socioeconomic reasons in reality and largely discounts the efficiency. ACKNOWLEDGMENTS =============== This study is supported by National Key Basic Research and Development Program (No.2010CB731403), the NCET program (No.NCET-09-0317), and the National Natural Science Foundation(No. 61273223) of China. [57]{} R.M. Anderson, R.M. May, Infectious Diseases of Humans: Dynamics and Control (Oxford University Press, Oxford, U.K., 1991). S. Boccaletti, et al., Phys. Rep. [424]{}, 175(2006). A. Barrat, et al. Dynamical Processes on Complex Networks(Cambridge University Press, Cambridge, U.K., 2008). S.N. Dorogovtsev, et al., Rev. Mod. Phys. [80]{}, 1275(2008). A. Vespignani, Nat. Phys. [8]{}, 32(2012). R. Pastor-Satorras and A. Vespignani, Phys. Rev. Lett. [86]{}, 3200(2001); Y. Moreno, et al., Eur. Phys. J B [26]{}, 521(2002); Boguñá M., et al., Phys. Rev. Lett. [90]{}, 028701(2003); Xia C.Y., et al., Int. J Mod. Phys. B [23]{}, 2303(2009); B. Guerra, J. Gómez-Gardeñes, Phys. Rev. E [82]{}, 035101(2010); C. Castellano and R. Pastor-Satorras. et al., Phys. Rev. Lett. [105]{}, 218701(2010). L.A. Rvachev, and I.M. Longini, Jr., Math. Biosci. [75]{}, 3(1985); C. Viboud, et al., Science [312]{}, 447(2006); V. Colizza, et al., Nat. Phys. [3]{}, 276(2007); D. Balcan, et al., Proc. Natl. Acad. Sci. U.S.A. [106]{}, 21484(2009); Wang L., et al., PLoS ONE [6]{}, e21197(2011). H.H.K Lentz, et al., Phys. Rev. E [85]{}, 066111(2012). C. Poletto, et al., Sci. Rep. [2]{}, 476(2012). L. Cao, et al., Phys. Rev. E [84]{}, 041936(2011) World Health Organization, Pandemic Influenza Preparedness and Response (WHO, Geneva, 2009); United States Department of Health and Human Services, HHS Pandemic Influenza Plan (HSS, Washington, D.C., 2005). Y. Yang, et al., Science [326]{}, 729(2009); J. Gómez-Gardeñes, et al., Proc. Natl. Acad. Sci. U.S.A. [105]{}, 1399(2008); P. Holme, Europhys. Lett. [68]{}, 908(2004); M.E. Halloran, et al., Proc. Natl. Acad. Sci. U.S.A.S [105]{}, 4639(2008);Y. Chen, et al., Phys. Rev. Lett. [101]{}, 058701(2008); L. Hufnagel, et al., Proc. Natl. Acad. Sci. U.S.A. [101]{}, 15124(2004); V. Colizza, et al., PLoS Med. [4]{}, e13(2007); I.M. Longini Jr., et al., Science [309]{}, 1083(2005). B.S. Cooper, et al., PLoS Med. [3]{}, e212(2006). T.D. Hollingsworth, et al., Nat. Med. [12]{}, 497(2006). J.M. Epstein, et al., PLoS ONE [2]{}, e401(2007). P. Bajardi, et al., PLoS ONE [6]{}, e16591(2011). G.S. Tomba, J. Wallinga, Math. Biosci. [214]{}, 70(2008). P. Caley, N.G. Becker, D.J. Philp, PLoS ONE [2]{}: e143(2007). A. Gautreau, et al., J. Stat. Mech. L09001(2007). A. Gautreau, et al., J. Theor. Biol. [251]{}, 509(2008). A. Gautreau, et al., Natl. Acad. Sci. U.S.A. [106]{}, 8847(2009). J.L. Iribarren, E. Moro, Phys. Rev. Lett. [103]{}, 038702(2009). O. Diekmann, et al., J. R. Soc. Interface [7]{}, 873(2010). [^1]: [email protected]	{ "pile_set_name": "ArXiv" }
--- abstract: \| The specific frequencies of globular cluster systems, $S_N\propto {\cal N}_ {\rm tot}/L_{V,{\rm gal}}\propto M_{\rm gcs}/M_{\rm stars}$, are discussed in terms of their connection to the efficiency of globular cluster formation in galaxy halos, which is claimed to reflect a generic aspect of the star formation process as it operates even at the current epoch. It is demonstrated that the total [masses]{} of GCSs are little affected by the dynamical destruction of low-mass clusters at small galactocentric radii. This permits direct, empirical estimates of the cluster formation efficiency by mass, $\epsilon_{\rm cl}\equiv M_{\rm gcs}^{\rm init}/M_{\rm gas}^{\rm init}$, even after $10^{10}$ years of GCS evolution. However, the standard practice of using only the stellar luminosities of galaxies as indicators of their initial total gas masses (and thus relating $S_N$ to $\epsilon_{\rm cl}$ in one step) leads to serious conceptual problems, that are reviewed here. The [first specific frequency problem]{}, which is the well known tendency for many brightest cluster galaxies to have higher than average $S_N$, is a global one; the [second specific frequency problem]{} is a local one, in which the more extended spatial distribution of GCSs relative to halo stars in some (not all) bright ellipticals leads to $S_N$ values that increase with radius inside the galaxies. Extending similar suggestions in the recent literature, it is argued that these trends in $S_N$ do not reflect any such behavior in the underlying $\epsilon_{\rm cl}$; rather, [both]{} of these problems stem from neglecting the hot, X-ray emitting gas in and around many large ellipticals, and both may be alleviated by including this component in estimates of $M_{\rm gas}^{\rm init}$. This claim is checked and confirmed in each of M87, M49, and NGC 1399, all of which have been thought to suffer from one or the other of these $S_N$ problems. Existing data are combined to construct GCS surface density profiles that extend over nearly the whole extents of these three galaxies, and a non-parametric, geometrical deprojection algorithm is developed to afford a direct comparison between the [volume]{} density profiles of their GCSs, stars, and gas. It is found, in each case, that $\rho_{\rm cl}\propto(\rho_ {\rm gas}+\rho_{\rm stars})$ at radii beyond roughly a stellar effective radius, inside of which dynamical evolution may have depeleted the initial GCSs. The constant of proportionality is the same in all three galaxies: $\epsilon_{\rm cl}=0.0026\pm0.0005$. Taken together, these results suggest that GCSs generally should be more spatially extended than stellar halos only in gas-rich galaxies that also have a high global specific frequency. The implication that $\epsilon_{\rm cl}$ might have had a universal value is supported by global GCS data for a sample of 97 giant ellipticals, brightest cluster galaxies, and faint dwarfs. The total globular cluster populations in all of these early-type systems are in excellent agreement with the predictions of a constant $\epsilon_{\rm cl}$ at the level observed directly in M87, M49, and NGC 1399; all [systematic]{} variations in GCS specific frequency between galaxies are shown to result entirely from different relations, in different magnitude ranges, between $M_{\rm gas}^{\rm init}$ and the present-day $L_{V,{\rm gal}}$. An identical $\epsilon_{\rm cl}$ is also calculated for the Pop. II spheroid of the Milky Way, and is indicated (although less conclusively) for the ongoing formation of open clusters. The inferred universal cluster formation efficiency, of $\simeq$0.25% by mass, should serve as a strong constraint on general theories of star and cluster formation. The associated inference of a non-universal formation efficiency for unclustered stars is considered, particularly in terms of the suggestion that this might result, both in dwarf galaxies and at large galactocentric radii in the brightest ellipticals, from feedback and galactic winds. Implications for a merger-formation model of early-type GCSs, and for the proposed existence of intergalactic globulars in clusters of galaxies, are briefly discussed. author: - 'Dean E. McLaughlin' title: THE EFFICIENCY OF GLOBULAR CLUSTER FORMATION --- Introduction ============ As the study of galaxy formation comes more and more to require a reliable description of star-gas interactions and feedback on relatively small spatial scales, it is forced towards questions that overlap with fundamental issues in star formation theory. Conversely, as evidence accumulates that most stars today—in environments ranging from the disk of the Milky Way to circumnuclear starbursts—are born not in isolation but in groups, the study of star formation in general seems increasingly to demand a focus on larger scales than might previously have been considered. Although these two disciplines have not yet converged—the smallest scales currently of importance for galaxy formation can still be one or two orders of magnitude greater than the “large” scales in star formation—they will necessarily continue to approach one another as further progress is made. It is therefore important to identify and understand those areas of truly common interest, on the largest stellar scales and the smallest galactic ones. In many ways, the systems of old globular clusters in galaxy halos are positioned just at this interface; they can potentially be exploited to constrain aspects of both local and global star-formation processes, in both protogalactic and present-day settings. It has, indeed, long been recognized that the global properties of globular cluster systems (GCSs) might be used to good effect as probes of galaxy formation and evolution. (This is especially true of bright early-type galaxies, whose GCSs are more populous and better studied than those in typical spirals like the Milky Way.) For example, following the suggestion that large fractions of the GCSs in ellipticals may have formed in major merger events (e.g., Schweizer 1987; Ashman & Zepf 1992; Zepf & Ashman 1993), and the related discovery of young super star clusters in systems like the Antennae galaxies (Whitmore & Schweizer 1995), much recent discussion has centered on the interpretation of the distribution of metallicities (or broad-band colors) among E-galaxy globulars as clues to their hosts’ dynamical histories (see Geisler, Lee, & Kim 1996; Forbes, Brodie & Grillmair 1997; Côté, Marzke, & West 1998; Kissler-Patig, Forbes, & Minniti 1998). A somewhat older approach has used just the total population of GCSs, as a function of parent galaxy luminosity, Hubble type, and local environment, to make arguments relating to issues such as biasing in a cold dark matter cosmology (West 1993); the merger history of ellipticals (Harris 1981; van den Bergh 1984; Ashman & Zepf 1992; Forbes et al. 1997); the formation of cD galaxies and their envelopes (McLaughlin, Harris, & Hanes 1993, 1994); star formation in galaxies at the centers of cluster cooling flows (Harris, Pritchet, & McClure 1995); the influence of early galactic winds on the evolution of giant ellipticals (Harris, Harris, & McLaughlin 1998); and even larger-scale aspects of galaxy clusters such as populations of intergalactic stars (e.g., West et al. 1995; Harris et al. 1998). Yet a third characteristic of GCSs—their radial distributions (the projected number density of globulars around a galaxy as a function of galactocentric radius)—has also served as the basis for discussions of elliptical galaxy formation (e.g., Harris 1986; Ashman & Zepf 1992; Kissler-Patig 1997; van den Bergh 1998). Rather less developed, however, is a clear sense of the relevance of GCSs to [local]{} star formation on smaller scales: What relation, if any, do the fundamental properties of globular cluster systems bear to the way stars form now, in the dense clumps within Galactic giant molecular clouds? Can the study of one of these phenomena possibly shed light on the other? Larson (1988, 1993, 1996), has emphasized the possible parallels between the basic pattern of local star formation and the birth of globular clusters within large protogalactic clouds in a scenario for the collapse of galaxies from clumpy initial conditions (as in, e.g., the classic picture of Searle & Zinn 1978, and current models of hierarchical galaxy formation). A more quantitative discussion along these lines, including an exposition of links to several of the gross characteristics of GCSs, is given by Harris & Pudritz (1994). Most recently, McLaughlin & Pudritz (1996) and Elmegreen & Efremov (1997) have focussed in detail on the globular cluster luminosity function (the number of clusters per unit magnitude, from which may be derived the mass function $d{\cal N}/dm$ for the entire system of globulars in a given galaxy) in attempts to model this property of GCSs using our current understanding of present-day star formation in molecular clouds and their clumps. The work presented in this paper further explores the connections between the formation of stars, globular clusters, and galaxies, by concentrating on two of the primary attributes of GCSs: their specific frequencies (or total populations) and radial distributions. The next Section discusses the significance of these GCS properties as they relate to general questions on the formation of stellar clusters. In §3 below, data are compiled from the literature, and a geometrical deprojection algorithm applied, to construct [volume]{} density profiles for the stars, the X-ray gas, and the globular clusters in three galaxies with well studied GCSs: M87 (the cD at the center of the Virgo Cluster), M49 (giant elliptical in Virgo), and NGC 1399 (cD galaxy in the Fornax Cluster). An intercomparison of the density profiles of the three halo components yields a point-by-point measure of the GCS contribution to the total luminous mass in each system, as a function of [three-dimensional]{} galactocentric radius, $r_{\rm gc}$. These GCS mass ratios, which are constant throughout each galaxy (at large $r_{\rm gc}$) and the same in all three, are interpreted as an indication of the formation efficiency of globular clusters. In §4, these results are supplemented with broader considerations of the GCSs in other early-type galaxies and in the Milky Way halo, and with a direct estimate for the relative formation rate of open clusters in the Galactic disk, to argue for the existence of a [universal efficiency, or probability, for the formation of a bound stellar cluster from a dense cloud of gas]{}. Section 5 discusses the implications of this for a few specific issues in galaxy and GCS formation and evolution, and §6 concludes with a summary. The Efficiency of Cluster Formation from Globular Cluster Systems ================================================================= Cluster Formation Efficiencies ------------------------------ The possibility that globular cluster systems can be connected not only to galaxy formation, but to ongoing star formation as well, is suggested by the fact that this latter process operates largely in a [clustered mode]{}: by mass, most new stars in the Milky Way appear in groups within the largest clumps in molecular clouds, and only rarely in true isolation. This is, by now, an established empirical fact (e.g., Lada et al. 1991; Zinnecker et al. 1993), and it is easily seen (Patel & Pudritz 1994) to be a direct consequence of the different power-law slopes in the mass function of molecular clumps ($d{\cal N}/dm\propto m^{-1.6}$, so that the largest clumps, which weigh in at $10^2$–$10^3\,M_\odot$, contain most of the star-forming gas mass in any molecular cloud) and the stellar initial mass function ($d{\cal N}/dm\propto m^{-2.35}$, putting most of the mass in young stars into $\la1\,M_\odot$ objects). Similarly, young super star clusters, which tend to have sizes and masses that are generally comparable to the old globulars in the Milky Way, can account for as much as $\sim$20% of the UV light in starburst galaxies (Meurer et al. 1995). This is, however, [not]{} to say that all, or even most, stars are born into bona fide clusters that exist as coherent dynamical units for any length of time; on the contrary, it would appear that very few multiple-star systems actually form as gravitationally bound clusters. (A quantification of “very few” is one of the goals of this paper.) At some point during the collapse and fragmentation of a cluster-sized cloud of gas, the massive stars which it has formed will expel any remaining gas by the combined action of their stellar winds, photoionization, and supernova explosions. If the [cumulative]{} star formation efficiency of the cloud, ${\rm SFE}=M_{\rm stars}/ (M_{\rm stars}+M_{\rm gas})$, is below a critical threshold when the gas is lost, then the blow-out carries away sufficient energy that the stellar group remaining is unbound, and disperses into the field. The precise value of this threshold SFE depends on details of the dynamics and magnetic field in the gas cloud before its self-destruction, and on the timescale over which the massive stars dispel the gas; but various estimates place it in the range ${\rm SFE}_{\rm crit}\sim 0.2$–0.5 (e.g., Hills 1980; Mathieu 1983; Elmegreen & Clemens 1985; Verschueren 1990; Lada, Margulis, & Dearborn 1984). A complete theory of star formation must therefore be able to anticipate the final cumulative SFE in any single piece of gas with (say) a given mass and density, and thereby predict whether or not it will form a bound cluster. No such theory yet exists. More within reach are considerations of a statistical nature, in which a [distribution]{} of cumulative SFEs (and, thus, the fraction of clouds that achieve ${\rm SFE}\ge {\rm SFE}_{\rm crit}$ before self-destruction) is calculated for an ensemble of star-forming clouds, in either a protogalactic or a present-day context (see, e.g., Elmegreen 1983; Elmegreen & Clemens 1985; Elmegreen & Efremov 1997). But even here, the systematic effects of a range in the initial masses of the gaseous clouds, or of changes in the assumed star formation law (which is probably the most outstanding “free” parameter here) have not been explored fully. It is therefore not clear that the dominant factor governing the formation of bound stellar clusters vs. unbound associations has been identified. It is nonetheless possible to constrain any forthcoming theories of star and cluster formation with empirical evaluations of the probability that a cluster-sized cloud of gas is able to achieve a cumulative SFE of at least the critical 20%–50%. This probability—which can equally well be interpreted as that fraction of an ensemble of massive star-forming clouds which manages to produce bound stellar systems—is referred to here as the efficiency of cluster formation. As was indicated above, an estimate of this will be taken first from a detailed investigation of the old globular cluster populations in three giant elliptical galaxies; general arguments will then be used to show that the result also applies to globular cluster formation in other galaxies, and to the formation of disk clusters today. A first (and very rough) assessment of the formation efficiency of globular clusters appeals to the most basic and most easily measured observable of a GCS: namely, its total population, ${\cal N}_{\rm tot}$, summed over all cluster luminosities and galactocentric positions. This is usually compared to the integrated $V$-band luminosity of the parent galaxy’s halo (as a measure of its unclustered field star population) through the so-called specific frequency (Harris & van den Bergh 1981): $$S_N\equiv {\cal N}_{\rm tot}\times10^{0.4(M_V^T+15)} \propto {\cal N}_{\rm tot}/L_{V,{\rm gal}}\ . \label{eq:21}$$ Even though any observational derivation of $S_N$ necessarily involves (possibly large) extrapolations of data that directly count only the brightest clusters in some spatially limited subset of a GCS, these estimates tend usually to be surprisingly robust. To date, specific frequencies have been obtained for of order 100 galaxies, some 80% of which are early-type systems (see the compilations of Harris 1991; Durrell et al. 1996; Kissler-Patig 1997; Blakeslee 1997; Blakeslee, Tonry, & Metzger 1997; Harris et al. 1998). Given this bias in the observational database, some of what follows will make use of results that are really specific to elliptical galaxies; spirals are brought more explicitly into the discussion in §4 below. Specific frequency is easily related to the total masses of a galaxy’s GCS and of its unclustered halo stars, if both the mean mass of the individual globulars in the system and the mass-to-light ratio of the overall [stellar]{} component are known. As is discussed further in §3.1 below, the basic distribution of globular cluster masses is quite similar from galaxy to galaxy, so that the mean cluster mass in the Milky Way GCS—$\langle m\rangle_ {\rm cl}=2.4\times10^5\,M_\odot$—serves as a reliable estimate for other systems as well. Meanwhile, the stellar mass-to-light ratios of large galaxies are essentially equal to the dynamical $M/L$ in their [cores]{}, which are baryon-dominated. As is also discussed below (§4.1), observations show that the core $M/L$ of hot galaxies increases with luminosity; but, even though this turns out to be rather an important consideration in detail, it will suffice for now to use a single value of $\Upsilon_V\equiv M/L_V=7\,M_\odot\, L_\odot^{-1}$ as a rough characterization of large ellipticals (van der Marel 1991; cf. Faber & Gallagher 1979; Binney & Tremaine 1987). The definition of $S_N$ in equation (\[eq:21\]) may then be rewritten as (cf. Zepf & Ashman 1993; Harris & Pudritz 1994; Harris et al. 1998) $$S_N\simeq2500\,\left({{\langle m\rangle_{\rm cl}}\over{2.4\times10^5\,M_\odot}} \right)^{-1}\,\left({{\Upsilon_V}\over{7\,M_\odot\,L_\odot^{-1}}}\right)\, {{M_{\rm gcs}}\over{M_{\rm stars}}}\ . \label{eq:22}$$ To be clear, it is stressed again that the mass-to-light ratio applied here refers to measurements in the cores of ellipticals, and is not meant to include any dark matter other than stellar remnants. In any event, an inversion of this last expression yields $${{M_{\rm gcs}}\over{M_{\rm stars}}} = 4.0\times10^{-4}\,S_N\, \left({{\langle m\rangle_{\rm cl}}\over{2.4\times10^5\,M_\odot}}\right)\, \left({{\Upsilon_V}\over{7\,M_\odot\,L_\odot^{-1}}}\right)^{-1}\, \sim\,2\times10^{-3}\,\left({{S_N}\over{5}}\right)\ , \label{eq:23}$$ where $S_N\approx5$ is suggested by the studies cited above as a representative value for early-type galaxies ranging in luminosity from $M_V^T\simeq-15$ to $M_V^T\simeq-22$. Of course, there is some uncertainty in this “mean” $S_N$, and some intrinsic scatter about it. Moreover, the data suggest that $S_N$ varies systematically with $L_{V,{\rm gal}}$—the simple scaling ${\cal N}_{\rm tot}\propto L_{V,{\rm gal}}$ that corresponds to a constant specific frequency is not strictly obeyed—particularly among dwarf ellipticals and between brightest cluster galaxies. These issues have been discussed in detail in the literature, and they are revisited here, in §2.2. With this caveat in mind, however, $M_{\rm gcs}/M_{\rm stars}\sim 2\times 10^{-3}$ is useful as a crude, order-of-magnitude guide to a “typical” global GCS mass ratio in galaxy halos. The advantage of working in terms of this [mass]{} ratio is that, whatever its precise value in any one system at the current epoch, it is expected to be a fairly well conserved quantity that has not changed drastically over the course of GCS and galaxy evolution in a Hubble time. [This is true even though the total population ${\cal N}_ {\rm tot}$ of any GCS is bound to decrease over time]{}, as individual globulars are disrupted by the cumulative effects of external gravitational shocks and internal two-body relaxation (see, e.g., Aguilar, Hut, & Ostriker 1988; Ostriker, Binney, & Saha 1989; Capuzzo-Dolcetta 1993; Murali & Weinberg 1997a, 1997b, 1997c; Gnedin & Ostriker 1997; Vesperini & Heggie 1997; but cf. Portegies Zwart et al. 1998). These dynamical processes discriminate preferentially against low-mass and low-density globular clusters, and it turns out that even if large [numbers]{} of such objects are lost, no great change is effected in the total [mass]{} of the remaining GCS. This conclusion rests on the fact that the mass functions of GCSs—particularly those in early-type galaxies—are shallow enough that most of the total $M_{\rm gcs}$ is contained in the most massive members of the system; but these are the least susceptible to dynamical destruction. The implication is that the total mass of a GCS may not change much over its lifetime; thus, [if]{} the current stellar mass of a galaxy were a reliable indication of the total amount of gas that was initially available to form stars, then the presently observed ratio $M_{\rm gcs}/M_{\rm stars}\sim2\times10^{-3}$ would serve as a rough, [global]{} estimate of the efficiency of globular cluster formation. Note that this reasoning also shows that galactic halos cannot be built up from disrupted globulars; see, e.g., Ashman & Zepf (1992) and Harris & Pudritz (1994). (One of the main conclusions of this paper will be that, in fact, $M_{\rm stars}$ is [not]{} always an adequate measure of an initial gas supply, and that $M_{\rm gcs}/M_{\rm stars}$ therefore cannot always be equated directly to a cluster formation efficiency. However, since most previous discussions of GCS specific frequencies do draw on this simple assumption at some level, it is of interest to develop some of its implications.) A simple numerical example should emphasize the robustness of $M_{\rm gcs}$. Thus, consider a newly formed system of globulars whose masses are distributed according to a power-law mass function between some lower and upper limits: $$\left({{d{\cal N}}\over{dm}}\right)_{\rm init}\,=\,K\,m^{-\gamma_2}\,, \ \ \ \ \ \ \ \ m_\ell \leq m\leq m_u\ , \label{eq:24}$$ where $\gamma_2>0$ and reasonable choices for $m_\ell$ and $m_u$ may be $500\,M_\odot$ and $5\times10^6\,M_\odot$. (The precise values are not critical here.) Now let this GCS be eroded over $10^{10}$ yr, keeping in mind that each cluster in the system has a definite, mass-dependent lifetime against the dynamical mechanisms mentioned above. In a rough approximation to a combination of complicated processes, there is then a mass scale (say $m_$) above which few or no globulars will have disappeared by the current epoch. Thus, the GCS mass function at $m>m_$ is effectively unchanged, even after a Hubble time, from the initial distribution. However, clusters with initial masses $m<m_$ [can]{} have been removed from the system, and the GCS mass function below $m_$ may differ significantly from its initial form. If this low-mass end of the evolved $d{\cal N}/dm$ can also be described by a power-law, then the full, present-day GCS mass function is just $$\left({{d{\cal N}}\over{dm}}\right)_{\rm obs}\,=\,\left\{ \begin{array}{cl} K\,m_^{\gamma_1-\gamma_2}\,m^{-\gamma_1}\,, & \ \ \ \ m_\ell\leq m\leq m_ \\ K\,m^{-\gamma_2}\,, & \ \ \ \ m_\leq m\leq m_u\ , \end{array} \right. \label{eq:25}$$ where $\gamma_1<\gamma_2$ for consistency. Observed GCSs do, in fact, have mass functions that conform to roughly this sort of double power law (Harris & Pudritz 1994; McLaughlin 1994; McLaughlin & Pudritz 1996)—although there is somewhat more structure than this at the highest cluster masses in $d{\cal N}/dm$, it can be safely ignored for the purposes of the present argument. Empirically, the mass scale $m_$ is [always]{} about $1.6\times10^5\,M_\odot$; at this mass, the logarithmic slope of $d{\cal N}/ dm$ changes rather abruptly from $\gamma_2>1$ to $\gamma_1<1$, with precise values that can vary slightly from galaxy to galaxy. There are strong arguments to support the notion that observed GCS mass functions at $m>m_$ are indeed accurate reflections of the formation distributions (namely, the high-mass sides of $d {\cal N}/dm$ in observed GCSs show no significant variation with galactocentric radius inside a single system, and are quite similar—though not identical—from galaxy to galaxy; see Harris & Pudritz 1994; McLaughlin & Pudritz 1996; Elmegreen & Efremov 1997). In addition, some theoretical studies (e.g., Surdin 1979; Okazaki & Tosa 1995; Vesperini 1997; Elmegreen & Efremov 1997) have suggested that the shallower slopes at lower masses might be caused entirely by dynamical evolution from an initial $d{\cal N}/dm\propto m^{-\gamma_2}$, roughly as described above. In reality, this view could very well be too extreme: empirically, the system of young super star clusters in the Antennae, which is not likely to have been severely affected by dynamical destruction, may have a mass function of the form (\[eq:25\]) rather than (\[eq:24\]) (Fritze-von Alvensleben 1999); and theoretically, such an initial shape to $d {\cal N}/dm$ may in fact be [preserved]{} during a Hubble time of dynamical evolution (Vesperini 1997). Thus, the intent here is not necessarily to argue for an explanation of present-day GCS mass spectra in purely evolutionary terms, but to show that even if such is adopted as an extreme possibility, the implied time dependence of total GCS masses is quite weak. Given equations (\[eq:24\]) and (\[eq:25\]), then, straightforward integrations allow for comparisons of the present GCS population to the initial one: $${{{\cal N}_{\rm tot}^{\rm obs}}\over{{\cal N}_{\rm tot}^{\rm init}}} = \left({{m_u}\over{m_}}\right)^{\gamma_2-1}\,\left\{ {{1-(m_/m_u)^{\gamma_2-1}+[(\gamma_2-1)/(1-\gamma_1)][1-(m_\ell/m_)^ {1-\gamma_1}]}\over{(m_u/m_\ell)^{\gamma_2-1}-1}}\right\}\ , \label{eq:26}$$ and of the current total mass to its initial value: $${{M_{\rm gcs}^{\rm obs}}\over{M_{\rm gcs}^{\rm init}}} = \left({{m_}\over{m_u}}\right)^{2-\gamma_2}\,\left\{ {{(m_u/m_)^{2-\gamma_2}-1+[(2-\gamma_2)/(2-\gamma_1)][1-(m_\ell/m_)^ {2-\gamma_1}]}\over{1-(m_\ell/m_u)^{2-\gamma_2}}}\right\}\ . \label{eq:27}$$ Values of $\gamma_1=0.4$ and $\gamma_2=1.7$ may be taken as representative of the typical situation in elliptical galaxies (Harris & Pudritz 1994; McLaughlin 1994, 1995; McLaughlin & Pudritz 1996). With $m_\ell=500\,M_\odot$, $m_=1.6\times10^5\,M_\odot$, and $m_u=5\times10^6\,M_\odot$, it is then easily seen that $${{{\cal N}_{\rm tot}^{\rm obs}}\over{{\cal N}_{\rm tot}^{\rm init}}}=0.036 \ \ \ \ \ \ \ \ {\rm and}\ \ \ \ \ \ \ \ \ {{M_{\rm gcs}^{\rm obs}}\over{M_{\rm gcs}^{\rm init}}}=0.76 \label{eq:28}$$ in this simplistic scenario. Thus, [even though dynamical evolution reduces the total population of this hypothetical GCS by more than an order of magnitude, it diminishes the total mass by only]{} $\sim$25%. (The same line of argument also shows that the often appreciable uncertainties in the exact form of $d{\cal N}/dm$ at cluster masses $m<m_$—or even blind extrapolations from observations of only brighter clusters—do not seriously bias observational estimates of $M_{\rm gcs}$ in any galaxy.) This result is best interpreted as an estimate for the possible depletion of a GCS [globally]{} (i.e., spatially averaged over the whole of a parent galaxy), because the values of $\gamma_1$, $\gamma_2$, and $m_$ used are taken from observations of $d{\cal N}/dm$ in projection, often over fairly large areas on the sky. Thus, the normalization $K$ in equations (\[eq:24\]) and (\[eq:25\]) can be taken to include an implicit integration over galactocentric radius, and the exponents $\gamma$ and mass scale $m_$ to be appropriate averages. The importance of these considerations is not actually clear, since radial dependences are not expected theoretically in current models for the initial GCS mass function (McLaughlin & Pudritz 1996; Elmegreen & Efremov 1997) and observational evidence for such trends at the current epoch is slight to non-existent (e.g., McLaughlin et al. 1994 and Harris et al. 1998; though see also Kavelaars & Hanes 1997 and Murali & Weinberg 1997b). Nevertheless, it should be noted that dynamical friction, gravitational shocks, and even evaporation couple to galactic tidal fields in such a way that if a GCS’s present-day mass function [were]{} sculpted to any great extent by these processes, then the [local]{} ratios ${\cal N}_{\rm tot}^{\rm obs}/{\cal N}_{\rm tot}^{\rm init}$ and $M_{\rm gcs}^{\rm obs}/M_{\rm gcs}^{\rm init}$ would be expected to increase outwards from the center of the system, from values less than equation (\[eq:28\]) at small galactocentric radii to near unity at large distances (e.g., beyond a stellar effective radius; see especially Murali & Weinberg 1997b). Returning now to the main issue, the global efficiency of globular cluster formation in a galaxy may be defined as $$\epsilon_{\rm cl}\,\equiv\, M_{\rm gcs}^{\rm init}/M_{\rm gas}^{\rm init}\ , \label{eq:29}$$ where $M_{\rm gas}^{\rm init}$ refers to the total gas supply that was available to the protogalaxy, whether in a rapid monolithic collapse or, more likely, in a slower assembly (at high redshift) of many distinct, subgalactic clouds. From the preceding discussion, it is clear that if a newly formed GCS were able to evolve “passively”—that is, if its mass were potentially vulnerable only to reductions caused by forces internal to an isolated parent galaxy—then a present-day observation of $M_{\rm gcs}$ would suffice, to at least a $\sim$25% level of accuracy globally, as a measure of $M_{\rm gcs}^ {\rm init}$. Passive evolution also implies mass conservation for the galaxy as a whole, so that a reliable value for $M_{\rm gas}^{\rm init}$ could be obtained, at any epoch, from an inventory of the total mass in gas plus unclustered stars and stellar remnants. Thus, a good [observational estimate]{} of the global cluster formation efficiency in this simple situation is provided by $${\hbox{$\widehat\epsilon$}}_{\rm cl}\,\equiv\,{{M_{\rm gcs}}\over{M_{\rm gas}+M_{\rm stars}}}\ , \label{eq:210}$$ which is expected to be nearly a time-independent quantity. Realistically, of course, galaxies are not such pristine entities; interactions between systems may effect significant changes in both $M_{\rm gcs}$ and $(M_{\rm gas}+M_{\rm stars})$ over time, making it difficult to evaluate (or even clearly define) initial GCS and gas masses. For example, the formation history of an elliptical may include major mergers with gas-rich spiral systems, and it has been proposed that these might involve the formation of new globular clusters and substantial increases in $M_{\rm gcs}$ (e.g., Ashman & Zepf 1992; Kumai, Basu, & Fujimoto 1993a). Alternatively, additional globulars might be captured by a large galaxy that accretes small, gas-poor systems (e.g., Côté et al. 1998) or accumulates tidal debris from other members in a cluster of galaxies (see Muzzio 1987). However, processes such as these also increase the total masses of gas and/or stars in the final system. In [large]{} galaxies (i.e., in those that do not suffer significant gas loss from strong galactic winds; cf. §4.2), any changes in the ratio $M_{\rm gcs}/(M_{\rm gas}+M_{\rm stars})$ are bound to be smaller than changes in either of these quantities individually. The presently observed ${\hbox{$\widehat\epsilon$}}_ {\rm cl}$ in any large galaxy with a complex dynamical history is then a mass-weighted average of the true $\epsilon_{\rm cl}$ in some number of discrete star-formation events and individual accreted systems; and in the limit that the underlying efficiency of cluster formation is independent of environment and epoch (as the results of §§3 and 4 will suggest is nearly correct), it is still true that ${\hbox{$\widehat\epsilon$}}_{\rm cl}\simeq\epsilon_{\rm cl}$—just as in the case of passive evolution, and subject only to the small error associated there with the effects of GCS erosion by cluster evaporation and tidal disruption. In gas-poor elliptical galaxies and the halos of spirals, equation (\[eq:210\]) reduces to ${\hbox{$\widehat\epsilon$}}_{\rm cl}\simeq M_{\rm gcs}/M_{\rm stars}$, so that—as was mentioned above—the cluster formation efficiency can be obtained directly from obervations of the global GCS specific frequency (cf. eq. \[\[eq:23\]\]). More generally, however, ellipticals can contain large amounts of hot, X-ray emitting gas (e.g., Forman, Jones, & Tucker 1985), and when it happens that $M_{\rm gas}$ is non-negligible relative to $M_{\rm stars}$, measurements of $S_N$ are only low-order approximations to ${\hbox{$\widehat\epsilon$}}_{\rm cl}$ (which is itself only an estimate of the real efficiency $\epsilon_{\rm cl}$). Keeping this in mind, the observed systematics of both global [and local]{} GCS specific frequencies provide a useful starting point for an examination of the extent to which cluster formation efficiencies might vary from galaxy to galaxy, or from place to place within a single system. The Specific Frequency “Problems” --------------------------------- As was intimated in the discussion around equation (\[eq:23\]), not all galaxies have the same globular cluster specific frequency. There are several aspects to this: (1) Spiral galaxies tend to have lower $S_N$—at least when their GCS populations are normalized to their [total]{} (bulge plus disk) light—than do average ellipticals (Harris 1991). (2) Although a constant $S_N\approx5$ is a fair [approximation]{} for many normal giant ellipticals, fits to the data suggest that, in fact, $S_N\propto L_{V,{\rm gal}}^{0.2-0.3}$ (Kissler-Patig 1997; cf. Santiago & Djorgovski 1993). At the extremes of the early-type galaxy sequence, moreover, (3) the $S_N$ values for centrally dominant or first-ranked members of galaxy clusters show a much steeper dependence on galaxy luminosity—$S_N\propto L_{V,{\rm gal}}^{0.8}$ or so is implied by the correlations of Blakeslee (1997) and Harris et al. (1998)—and (4) specific frequencies actually appear to increase towards [lower]{} luminosities in dwarf ellipticals: $S_N\propto L_{V,{\rm gal}}^{-0.4}$ (Miller et al. 1998; see also Durrell et al. 1996). Thus, some brightest cluster galaxies, and some of the faintest dwarfs, have global specific frequencies that exceed the “average” by factors of 3 or 4. All in all, observed $S_N$ values span a range from $\sim$1 to $\sim$20. As will be shown in §4 below, the first two of these points can be understood fairly easily. Here, it suffices to note that (1) when only the [halo]{} stars and globular clusters in the Milky Way are considered, and allowances made for the lower Pop. II mass-to-light ratio in our Galaxy vs. large ellipticals, ${\hbox{$\widehat\epsilon$}}_{\rm cl}$ here is no different from that in “normal” early-type systems. It is reasonable to expect that this result—which is contrary to the claim made by Zepf & Ashman (1993)—should hold in other spirals as well. Also, (2) the observed dependence of $S_N$ on $L_{V,{\rm gal}}$ among non-brightest cluster ellipticals is fully consistent with that expected from equation (\[eq:22\]) for gas-poor systems with a uniform efficiency of cluster formation (${\hbox{$\widehat\epsilon$}}_{\rm cl}\simeq M_{\rm gcs}/ M_{\rm stars}$) and a mass-to-light ratio ($\Upsilon_V$) that scales with luminosity in the way observed for galaxies in the fundamental plane. However, if taken at face value, items (3) and (4) just above appear to imply that GCS formation efficiencies can vary by as much as an order of magnitude from galaxy to galaxy, in a non-monotonic fashion and for reasons that have never been understood. These issues are not so readily dismissed. The high $S_N$ “phenomenon” in many first-ranked cluster ellipticals, especially, has been discussed at length in the literature (e.g., McLaughlin et al. 1994; Harris et al. 1995; West et al. 1995; Forbes et al. 1997; Blakeslee 1997; Blakeslee et al. 1997; Harris et al. 1998), and is referred to here as the [first specific frequency problem]{}. Equally non-trivial, though only more recently recognized, is the [inverse]{} variation of $S_N$ with $L_{V,{\rm gal}}$ among dwarf ellipticals. Local specific frequencies, which are obtained for a single GCS by comparing its projected radial distribution to its galaxy’s surface brightness profile (normalized as in eq. \[\[eq:21\]\]), are also of interest here; they are related in the obvious way to the efficiency of cluster formation as a function of (projected) galactocentric radius, $R_{\rm gc}$. Of particular importance is the fact that $S_N$ is seen to be an [increasing]{} function of $R_{\rm gc}$ in many galaxies. This is equivalent to the well known fact that GCSs are often less centrally concentrated than the stellar halos of their parent galaxies. This statement has two contexts, which must be separated. First, it seems that the GCS of most any galaxy—spiral or elliptical, small or large—has an apparent core radius that is significantly larger than that of the unclustered field stars (e.g., Djorgovski & Meylan 1994; Lauer & Kormendy 1986; McLaughlin 1995; Forbes et al. 1996; Minniti, Meylan, & Kissler-Patig 1996; Côté et al. 1998); that is, the [central regions]{} of GCSs show projected radial distributions with slopes that are significantly shallower than those of the light profiles on kpc scales, so that the density ratio of the two halo components is an increasing function of radius. Second, the GCSs of some large galaxies are spatially more extended than the stellar halos on [larger scales]{}, $R_{\rm gc}\sim10$–100 kpc (Harris & Racine 1979; Harris 1986). However, this is a less general result than that relating to the large core radii of globular cluster systems. It is true of many dominant cluster galaxies like M87 in Virgo (see the discussion in McLaughlin et al. 1995), and also of the M31 halo (Pritchet & van den Bergh 1994); but many elliptical systems (e.g., Durrell et al. 1996; Kissler-Patig et al. 1997), and the Milky Way itself (Harris 1976; Zinn 1985), show GCS and halo-star distributions that are in quite close agreement (and, thus, a constant local $S_N$) outside of any core regions. The physical factors that govern whether the GCS density falls off at the same rate as or more slowly than the unclustered halo density in a given system have not yet been identified. (No case has ever been found in which a GCS radial distribution is [steeper]{} than a galaxy’s light profile.) The first of these points is not easily related to the radial dependence of cluster formation efficiencies because, as was suggested in §2.1, the dynamical mechanisms that act to destroy globulars work most efficiently at small $R_{\rm gc}$, where the stellar densities are highest. Although the cores of GCSs could quite plausibly have always been somewhat larger than those of halos or bulges generally (e.g., Harris & Pudritz 1994; Côté et al. 1998), it seems inevitable that any initial discrepancies must have been augmented by an inside-out erosion of the cluster system from tidal shocking, dynamical friction, and evaporation (Ostriker et al. 1989; Capuzzo-Dolcetta 1993; Murali & Weinberg 1997b). Detailed modelling is therefore required to disentangle the effects of formation and evolutionary processes in this instance. At galactocentric radii greater than a few kpc, however, the timescales on which the evolutionary mechanisms operate can exceed a Hubble time, and the studies cited here show that dynamics will not significantly affect GCS radial distributions beyond about one effective radius in the stellar halo (see also Aguilar et al. 1988). Thus, to the extent that the assumption ${\hbox{$\widehat\epsilon$}}_{\rm cl}\simeq M_{\rm gcs}/M_{\rm stars}$ is valid, the distension of GCSs relative to stellar halos and the concomitant systematic increase in local $S_N$ in the outer parts of [some]{} galaxies would seem to say that bound clusters were sometimes (but not always, and for unknown reasons) [more]{} likely to form at larger $R_{\rm gc}$, in gas that was presumably at lower ambient densities and pressures. This situation will be referred to as the [second specific frequency problem]{}. It is clearly of interest, from the points of view of both galaxy and star formation, to verify whether these global and local $S_N$ variations really do reflect differences in $\epsilon_{\rm cl}$—and, if so, to understand their origins. Accordingly, much attention has been paid to the first $S_N$ problem in particular, usually in efforts to empirically correlate the $S_N$ measurements for brightest cluster galaxies (BCGs) with global properties of the galaxies themselves or of the clusters they inhabit. One such correlation does appear to be of fundamental importance, as will be discussed momentarily, but it suggests that $S_N$ varies in the face of [constant]{} $\epsilon_{\rm cl}$. Indeed, no proposed solution of the first specific frequency problem has identified a physical mechanism that could possibly lead to the superefficient production of bound star clusters in some galaxies (see the references given above). In fact, one suggestion—that high global specific frequencies (meaning $S_N\ga5$) in centrally located BCGs result from the acquisition of globulars and stars which are tidally stripped from the outer, high-$S_N$ regions of other galaxies in a cluster—actually appears to rely on the existence of the [second]{} specific frequency problem (see Forbes et al. 1997; Côté et al. 1998); but this is itself essentially unexplained. Nor, to date, has any explanation (or prediction) been given for the high $S_N$ that have now been observed in many faint dwarf galaxies. The view is taken here that [the second specific frequency problem is inseparable from the first]{}, in the specific sense that both relate to the possible influence of environment on the probability of bound cluster formation. Any real insight into either one of these issues must therefore include a better understanding of the other. This view is in keeping with recent evidence (Forbes et al. 1997; Kissler-Patig 1997; van den Bergh 1998) for at least a rough correlation between global $S_N$ values and GCS density-profile slopes in elliptical galaxies: those GCSs with the highest specific frequencies tend also to have the shallowest radial distributions. Such a correspondence hints strongly at (although it does not directly show) an intimate connection between an increasing local $S_N$ and a high global $S_N$. A significant development in this field has come with the recent recognition of a correlation between the global specific frequencies of BCGs and the X-ray luminosity of the hot gas in their parent clusters (West et al. 1995; Blakeslee 1997; Blakeslee et al. 1997). A relatively large and homogeneous sample of BCGs shows $S_N$ values that increase from about 4, in clusters with X-ray luminosities (on 500-kpc scales) $L_X\sim10^{42}$ erg s$^{-1}$ at 0.5–4.5 keV, to $S_N\simeq12$ for $L_X\sim10^{45}$ erg s$^{-1}$. Three different interpretations have already been attached to this result: West et al. (1995) propose that it points to the existence of intergalactic globulars which collect naturally at the centers of galaxy clusters and are superimposed in projection on any galaxy there. This would certainly boost the apparent total population ${\cal N}_{\rm tot}$ of a central BCG, so that globulars need not have formed in situ with abnormally high specific frequency; but they would still have had to do so elsewhere in the cluster (see also Côté et al. 1998; Harris et al. 1998). In a different approach, Blakeslee (1997) and Blakeslee et al. (1997) consider $L_X$ to be mainly a reflection of the temperature of intracluster gas, and thus a measure of a cluster’s mass and the depth of its potential. They then claim that the observed $S_N$–$L_X$ correlation is consistent with the globulars in BCGs having formed in numbers directly proportional to the [total]{} mass, including dark matter, of their [entire]{} galaxy clusters (see Kavelaars 1998 for a similar suggestion on the scale of individual galaxies). In this scenario, cluster formation would obey a universal efficiency of sorts—although one which is quite different from that discussed here—and the first $S_N$ problem would result from some process that caused the BCGs in more massive clusters to be systematically underluminous (i.e., underefficient in the production of unclustered field stars). Here, however, the role that the dark matter content of a galaxy cluster might play in the formation of globular clusters is not clear. Moreover, any rule such as $M_{\rm gcs}^{\rm init}/(M_{\rm gas}^{\rm init}+ M_{\rm dark})=constant$ clearly can apply over only a limited range of conditions. Thus, Harris et al. (1998) instead pursue the hypothesis that GCSs formed with a constant $\epsilon_{\rm cl}$ defined essentially as in equation (\[eq:29\]) above. This is adopted as a working [assumption]{} by Harris et al., who construct plausibility arguments for a universal value of $\epsilon_{\rm cl}\sim10^{-3}$ (as could be expected) and suggest that the higher global $S_N$ in more X-ray luminous BCGs might have resulted from stronger (partial) galactic winds which heated larger fractions of their initial gas supplies to keV temperatures after an early bout of star and globular cluster formation. Such hot gas would be largely removed from further star formation—but still confined to the vicinity of these central cluster galaxies—and the final BCGs would be optically underluminous as a result. It is this basic idea that is closest to the discussion in this paper (although many aspects of the quantitative analysis here differ significantly from those in Harris et al. 1998). That is, the fundamental point of the global $S_N$–$L_X$ correlation for BCGs is the associated implication that the first specific frequency problem may stem, not from intrinsic variations in the global $\epsilon_{\rm cl}$ of these systems, but from the fact that they can contain large amounts of gas. High gas mass ratios invalidate the assumption that ${\hbox{$\widehat\epsilon$}}_{\rm cl}\simeq M_{\rm gcs}/ M_{\rm stars}$, and therefore cause $S_N$ to overestimate the true globular cluster formation efficiency. If it is correct to infer from this that the “excess” globulars in systems with high global $S_N$ are associated with hot gas, then the fact that the distribution of gas is generally more extended than that of the stars in ellipticals (e.g., Trinchieri, Fabbiano, & Canizares 1986; Fabbiano 1989) suggests a possible explanation for the second specific frequency problem as well as the first. These two issues would then be closely related—possibly even just different aspects of a single phenomenon—as was suggested above; and both might be settled simply by measuring the full estimator ${\hbox{$\widehat\epsilon$}}_{\rm cl}=M_{\rm gcs} /(M_{\rm gas}+M_{\rm stars})$ both locally and globally in individual systems. In the next Section, this universal-$\epsilon_{\rm cl}$ hypothesis is [tested]{}, for the first time, by evaluating the observable ${\hbox{$\widehat\epsilon$}}_{\rm cl}$ as a function of galactocentric radius in each of the galaxies M87, M49, and NGC 1399. It turns out that the inclusion of the X-ray gas in M87 is indeed key to the alleviation of both the first and the second $S_N$ problems there: once this is done, ${\hbox{$\widehat\epsilon$}}_{\rm cl}$ is constant locally, and essentially the same globally, in all three of these systems. The discussion is expanded to include other giant ellipticals and BCGs, as well as dwarf ellipticals, in §4. It is shown directly that all available GCS data are consistent with a single efficiency of cluster formation. Notably, the increase of $S_N$ towards low $L_{V,{\rm gal}}$ in the dwarfs (Miller et al. 1998; Durrell et al. 1996) can be explained in terms of this universal $\epsilon_{\rm cl}$ once allowances are made for the effects of the violent gas outflows (a more extreme version of the feedback envisioned by Harris et al. 1998 for BCGs) that were driven by early bursts of star formation in these low-mass systems. Case Studies of Three Galaxies ============================== M87, M49, and NGC 1399 are ideally suited to a closer examination of these issues, for three reasons: (1) They are relatively nearby, making it possible to observe internal trends within the systems in the first place. M87 and M49 both are assumed here to lie at a distance of $D=15$ Mpc, appropriate for the core of the Virgo Cluster (Pierce et al. 1994; Freedman et al. 1994). A relative distance modulus of $\Delta (m-M)[{\rm Fornax}-{\rm Virgo}]=0.2$ (e.g., McMillan, Ciardullo, & Jacoby 1993; Tonry et al. 1997) then puts NGC 1399 at $D\simeq16.5$ Mpc. (2) Each is a giant elliptical (M87 and NGC 1399 are actually type cD), and is therefore accompanied by a large GCS: ${\cal N}_{\rm tot}\simeq13\,500$ for M87; $6\,900$ for M49; and $4\,700$ in NGC 1399 (e.g., Harris 1991; see also §5 below). And, most importantly, (3) even this small sample of three galaxies suffices to bring out all aspects of the two specific frequency problems summarized in §2.2: M87, as is well known, suffers from both; M49 does not show the first, but has long been thought to exhibit the second; and NGC 1399 has been claimed as an example of the first problem, but not the second. Because of their proximity, these systems have been well studied in all respects of importance here. $B$-band surface photometry extends in M87 to a projected galactocentric radius of about $22\arcmin\simeq95$ kpc (de Vaucouleurs & Nieto 1978, 1979); in M49, to a similar distance (King 1978; Caon, Capaccioli, & D’Onofrio 1994); and in NGC 1399, to $R_{\rm gc}\simeq14\farcm5=70$ kpc (Killeen & Bicknell 1988; Caon et al. 1994). These measurements are easily converted to projected stellar mass density profiles $\Sigma_{\rm stars}(R_{\rm gc})$, as mass-to-light ratios have been obtained from a spectroscopic analysis applied uniformly to all three systems (van der Marel 1991). X-ray observations, which have been made with several different satellites, yield gas densities over still larger areas around each galaxy. Here, use will be made of the [ROSAT]{} data for M87 (taken from Nulsen & Böhringer 1995); of [ROSAT]{} data for M49 (Irwin & Sarazin 1996); and of [Einstein]{} observations in NGC 1399 (Killeen & Bicknell 1988). These studies all give the [volume]{} electron density $n_e(r_{\rm gc})$, as a function of [three-dimensional]{} galactocentric radius, derived either from fits of a parametric function to the projected X-ray brightness profile or from a direct deprojection of the same. Finally, integrated optical magnitudes have been measured for globular clusters in these galaxies at radii $R_{\rm gc}\sim1-100$ kpc—generally, over most of the areas covered by the surface photometry of unresolved halo stars. These are used to construct projected GCS number densities, $N_{\rm cl}(R_{\rm gc})$, which are trivially converted to mass profiles, $\Sigma_{\rm cl}$, once a mean globular cluster mass is specified. For M87, relevant investigations are those of Harris & Smith (1976), Harris (1986), Grillmair, Pritchet, & van den Bergh (1986), Lauer & Kormendy (1986), Cohen (1988), McLaughlin et al. (1993, 1994), McLaughlin (1995), Whitmore et al. (1995), and Harris et al. (1998). Data on M49 can be found in Harris & Petrie (1978), Harris & van den Bergh (1981), Harris (1986), Harris et al. (1991), and Lee, Kim, & Geisler (1998). The GCS of NGC 1399 has been observed by Hanes & Harris (1986), Bridges, Hanes, & Harris (1991), Wagner et al. (1991), and Kissler-Patig et al. (1997). While surface brightness profiles and X-ray gas densities often can be obtained for most of an entire galaxy with a single set of observations—as is the case in the studies cited above—this is generally [not]{} so for GCS radial distributions. Older, photographic surveys, which did have large ($R\sim100$ kpc) fields of view, were hampered by rather bright limiting magnitudes that became systematically more so towards smaller galactocentric radii. Newer, CCD photometry achieves deeper limiting magnitudes, and digital data reduction techniques can quantify and correct for any spatial variations; but until very recently, CCD fields of view were quite small, and best suited to analyses of the central regions ($R_{\rm gc}\la10$ kpc) of GCSs. Thus, the first step in comparing the stellar, gaseous, and globular cluster contents of any large galaxy is to combine existing photographic and CCD GCS data into a single, composite radial distribution that gives the projected density $N_{\rm cl}(R_{\rm gc})$, to a uniform limiting magnitude, over as large a range of galactocentric radius as possible. This is done for each of M87, M49, and NGC 1399 in §3.1. But then, the surface densities of stars and GCSs are related to their volume densities by an integral of the form $\Sigma\propto \int\rho\,dz$, while the observed X-ray brightness along some line of sight is proportional to $\int\rho_{\rm gas}^2\,dz$. A direct intercomparison of the [raw]{} GCS, optical, and X-ray imaging data in a galaxy therefore makes little sense. Thus, to discuss these three components simultaneously in M87, M49, and NGC 1399, either (1) their GCS and stellar surface densities can be deprojected and compared to their derived gas volume densities, or (2) the gas surface densities, $\Sigma_{\rm gas}(R_{\rm gc})\propto\int\rho_{\rm gas} (r_{\rm gc})\,dz$, can be obtained from the published volume density profiles and compared to the directly observed $\Sigma_{\rm stars}(R_{\rm gc})$ and $\Sigma_{\rm cl}(R_{\rm gc})$. The results of these two separate exercises are presented in §§3.2 and 3.3. GCS Surface Densities --------------------- When joining GCS radial distributions from two or more separate studies of the same galaxy, two important factors must be taken into account. First, the GCS of a distant galaxy (essentially, any with $D\ga5$ Mpc) is usually identified as a centrally concentrated excess of point-like sources, above a more uniform distribution of unresolved background galaxies and foreground stars. The surface density of this “background,” $N_b$, is subtracted from the total density of point sources at any $R_{\rm gc}$ to form the surface density profile $N_{\rm cl}$ of the GCS alone; but $N_b$ is a function of seeing conditions, detector characteristics, and limiting magnitude. Background corrections must therefore be applied separately to each of the datasets to be combined. Second, differences in limiting magnitude (call this $V_{\rm lim}$) are of further importance because a deep survey will clearly see more globulars, at a given $R_{\rm gc}$, than one with a bright limit. Before they can be simply matched onto one another, then, distinct background-corrected profiles $N_{\rm cl}(R_{\rm gc})$ must be normalized to the same effective $V_{\rm lim}$. The published GCS observations that will be used here for M87, M49, and NGC 1399 are summarized in Table \[tab1\]. These papers were selected from the more numerous references cited above, simply because they tabulate all of the relevant data in full. The third column of Table \[tab1\] gives the minimum and maximum $R_{\rm gc}$ contained in each study’s field of view (note that $1\arcmin=4.363$ kpc for a distance of $D=15$ Mpc to M87 and M49, while $D=16.5$ Mpc for NGC 1399 implies $1\arcmin=4.800$ kpc); Column 4 gives the background surface density to the limiting magnitude in Column 5; and Column 6 lists the scalings applied to each dataset to bring them all to a common $V_{\rm lim}$. [llcccl]{} M87 & McLaughlin (1995) & $0.153-1.965$ & $3.5\pm0.8$ & 23.9 & 1.000 $V^0=23.70\ \ \sigma_V=1.39$ & McLaughlin et al. (1993) & $1.210-9.090$ & $6.3\pm0.4$ & 24.0 & 0.954 & Harris (1986) & $1.000-22.52$ & $5.8\pm0.3$ & 23.4 & 1.35 & & & & & & & & & & M49 & Harris & Petrie (1978) & $2.000-19.69$ & $4.3\pm0.3$ & 23.3 & 1.000 $V^0=23.75\ \ \sigma_V=1.30$ & Harris & van den Bergh (1981) & $0.732-12.15$ & $3.3\pm0.5$ & $\ldots$ & 0.86 & Harris et al. (1991) & $0.600-2.304$ & $2.88\pm0.45$ & 23.8 & 0.707 & Lee et al. (1998) & $0.900-7.167$ & $2.95\pm0.30$ & 23.45 & 0.892 & & & & & & & & & & NGC 1399 & Kissler-Patig et al. (1997) & $0.761-9.463$ & $6.1\pm0.3$ & 24.0 & 1.000 $V^0=23.90\ \ \sigma_V=1.20$ & Hanes & Harris (1986) & $0.734-13.97$ & $5.40\pm0.45$ & 23.4 & 1.575 Generally, the background levels $N_b$ have been determined and applied by the original authors, either from observations of a blank field near the observed GCS, or from “on-frame” extrapolations of counts at the edges of the GCS field itself (see Harris 1986); these values are reproduced in Table \[tab1\] only for completeness. The single exception to this is the M49 study of Lee et al. (1998), for which $N_b$ has been derived here, using the method of Harris (1986). (It is possible, particularly with on-frame estimates, to overestimate $N_b$ and produce spurious curvature at large radii in a subtracted $N_{\rm cl}$ profile. As is discussed below, however, this is not a major concern in the composite GCS radial distributions obtained here.) The limiting magnitudes $V_{\rm lim}$ in Table \[tab1\] are also those quoted in the original papers. These have been used to renormalize the individual $N_{\rm cl}$ profiles, by adopting—for computational convenience—a standard Gaussian approximation to the globular cluster luminosity function (GCLF, $d{\cal N}/ d\,\ln\,m$; cf. §2.1). That is, with the number of globulars at a given apparent magnitude given roughly by $d{\cal N}/dV\propto\exp [-(V-V^0)^2/2 \sigma_V^2]$ (see Harris 1991), then the density of clusters to some limiting magnitude, say $N_{\rm cl}(V\le V_{{\rm lim},2})$, is referred to the density up to some other limit, i.e., $N_{\rm cl}(V\le V_{{\rm lim},1})$, through $${{N_{\rm cl}(V\le V_{{\rm lim},1})}\over{N_{\rm cl}(V\le V_{{\rm lim},2})}} \,=\, {{1+{\rm erf}\left[(V_{{\rm lim},1}-V^0)/\sqrt{2}\,\sigma_V\right]}\over {1+{\rm erf}\left[(V_{{\rm lim},2}-V^0)/\sqrt{2}\,\sigma_V\right]}}\ . \label{eq:31}$$ The right-hand side of this expression is just the scaling factor given in Column 6 of Table \[tab1\], where $V_{{\rm lim},1}$ for each GCS has been taken as the observed limiting magnitude in the first of each set of references. Thus, the published radial distributions for M87 have all been normalized to an effective $V_{\rm lim}=23.9$; for M49, to $V_{\rm lim}=23.3$; and to $V_{\rm lim}=24.0$ in NGC 1399. The Gaussian peaks $V^0$ and dispersions $\sigma_V$ used in equation (\[eq:31\]) are given in the first column of Table \[tab1\]. These parameters are taken from the fits of Whitmore et al. (1995) and Harris et al. (1998) to the M87 GCLF, from Lee et al. (1998) for M49, and from Kohle et al. (1996) for NGC 1399. Note that all of the scale factors derived here are rather modest, and thus do not introduce significant uncertainty into the final, combined $N_{\rm cl}$ profiles. =4.0truein A few practical notes should be made regarding this scaling procedure. First, it is applied to the GCS densities at every observed galactocentric radius, and it therefore assumes that there is no intrinsic radial variation in the GCLF or the underlying GCS mass function. As was also mentioned in §2.1 above, this is a reasonable assumption that is directly supported by the observations themselves. Second, the limiting magnitudes for the photographic studies of Harris & Petrie (1978) and Harris (1986) are rather poorly determined; Harris quotes only approximate ranges for $V_{\rm lim}$. The values adopted here, which are fully consistent with these ranges, have been chosen to give scaled $N_{\rm cl}(R_{\rm gc})$ that agree well with more recent CCD studies with accurately known $V_{\rm lim}$. And third, the scaling applied to the M49 GCS counts of Harris & van den Bergh (1981) has not been calculated in this way at all; rather, the factor of 0.86 is derived, by Harris (1986), by requiring a good overall agreement with the $N_{\rm cl}$ profile of Harris & Petrie (1978). The results of this exercise are illustrated in Fig. \[fig1\]. Clearly, the different projected density profiles in each GCS corroborate each other. It remains only to extract a single $N_{\rm cl}$ at each $R_{\rm gc}$, in each system, from the overlapping data that have been collected. It is not generally possible to simply average different densities at a given $R_{\rm gc}$, since profiles such as these are constructed by counting clusters in a series of wide, circular annuli that need not coincide from study to study. Instead, Fig. \[fig1\] shows that a perfectly acceptable approach is to simply adopt one of the (scaled) profiles over some range in $R_{\rm gc}$, and then switch to any other that might extend beyond that range. At this point, it must be kept in mind that ultimately the data will be deprojected to yield volume GCS densities, $n_{\rm cl}(r_{\rm gc})$. The method developed in the Appendix to do this requires that the inner and outer radii be known for the annuli used to define the composite $N_{\rm cl}$. This further restricts the useful published datasets to those which contain this information on the radial bins as well (that is, [in addition]{} to just a density and a mean radius for each annulus). In M87, then, a composite GCS profile is derived by using the data of McLaughlin (1995) from $R_{\rm gc}=0\farcm153$ to $R_{\rm gc}= 1\farcm870$; those of McLaughlin et al. (1993) from $1\farcm870$ to $7\farcm870$; and those of Harris (1986) from $7\farcm870$ to $22\farcm52$. (In order to make annuli that do not overlap in the final $N_{\rm cl}$, some interpolation is applied to the datasets around the matching points at $1\farcm870$ and $7\farcm870$.) For M49, the scaled profile of Harris & van den Bergh (1981) is used over $0\farcm732\le R_{\rm gc}\le 10\farcm96$, and that of Harris & Petrie (1978) for $10\farcm96\le R_{\rm gc}\le 19\farcm69$. And in NGC 1399, the densities of Kissler-Patig et al. (1997) are adopted for $0\farcm761\le R_{\rm gc}\le 7\farcm571$, with the data of Hanes and Harris (1986) continuing on out to $R_{\rm gc}=13 \farcm97$. In all cases, CCD data have usually been chosen over photographic data in any regions of overlap. However, superseding this preference is a concern over possible errors in the background estimation $N_b$; where this seems too high—for example, when the background-subtracted $N_{\rm cl}$ at the largest radii in a smaller-scale CCD study drop significantly below the densities indicated by a more extensive photographic survey (as in the outermost data points of Kissler-Patig et al. 1997 for NGC 1399)—the photographic profiles are given precedence. Enforcing consistency in this way between independent GCS datasets ensures that the question of potentially overestimated $N_b$ values remains an issue only at the largest radii (beyond the regime of any overlap) in the widest-field observations of each galaxy. The composite radial distributions thus obtained are listed in the third columns of Tables \[tab2\] through \[tab4\], as average surface densities in series of concentric circular annuli (with inner, outer, and average projected radii given in the first and second columns of the Tables). They are also drawn as the bold lines in the three panels of Fig. \[fig1\]. There are two details of particular note. First, the densities for the innermost two annuli in M49 have been scaled up from the (already renormalized) GCS profile of Harris & van den Bergh (1981), by a further factor of 1.1, to put the composite $N_{\rm cl}$ at small radii into better agreement with the CCD data of Harris et al. (1991) and Lee et al. (1998). This ad hoc correction most likely just reflects some residual incompleteness in the photographic GCS counts at small $R_{\rm gc}$, and in any event does not significantly affect the results which follow. Second, all of the $N_{\rm cl}$ profiles have been smoothed somewhat by rebinning, in order to allow for a more stable deprojection in §3.2 below: the outermost six annuli of Harris (1986) are combined into one for the M87 profile; the outermost 11 rings of Harris & Petrie (1978) are blended into 4 for M49; and the last 8 radial bins of Hanes & Harris (1986) become 4 in the NGC 1399 GCS. (As an added benefit of this smoothing, any possible errors in the subtracted $N_b$ are expected to affect only the outermost [one or two points]{} in the final $N_{\rm cl}$ profiles.) Similarly, counts in the “Northeast” and “Northwest” sectors of NGC 1399 in the analysis of Kissler-Patig et al. (1997; see their Table 8) have been added together in every annulus they define; and these annuli have been combined in pairs to give the coarser profile in Table \[tab4\] here. [ccccc]{} 0.153 – 0.328 & 0.244 & 282.6$\pm$55.8 & 0.245 & 243.7$\pm$155.6 0.328 – 0.655 & 0.494 & 197.2$\pm$19.9 & 0.505 & 89.34$\pm$34.01 0.655 – 0.983 & 0.808 & 145.3$\pm$16.5 & 0.827 & 68.58$\pm$20.55 0.983 – 1.310 & 1.157 & 89.10$\pm$12.40& 1.152 & 18.16$\pm$12.94 1.310 – 1.870 & 1.500 & 74.46$\pm$8.78 & 1.602 & 19.33$\pm$5.23 1.870 – 2.160 & 2.010 & 48.92$\pm$4.25 & 2.018 & 11.36$\pm$3.56 2.160 – 2.490 & 2.320 & 37.81$\pm$3.26 & 2.328 & 4.015$\pm$2.391 2.490 – 2.880 & 2.680 & 35.75$\pm$2.80 & 2.689 & 6.404$\pm$1.722 2.880 – 3.320 & 3.090 & 26.93$\pm$2.05 & 3.104 & 3.423$\pm$1.183 3.320 – 3.830 & 3.570 & 21.74$\pm$1.65 & 3.580 & 2.875$\pm$0.818 3.830 – 4.430 & 4.120 & 16.78$\pm$1.34 & 4.135 & 0.926$\pm$0.587 4.430 – 5.110 & 4.760 & 16.38$\pm$1.23 & 4.776 & 1.447$\pm$0.469 5.110 – 5.900 & 5.490 & 13.26$\pm$1.13 & 5.512 & 1.202$\pm$0.360 5.900 – 6.820 & 6.350 & 9.79$\pm$0.93 & 6.368 & 0.768$\pm$0.271 6.820 – 7.870 & 7.330 & 7.16$\pm$0.96 & 7.354 & 0.442$\pm$0.219 7.870 – 8.973 & 8.403 & 5.82$\pm$0.73 & 8.431 & 0.184$\pm$0.166 8.973 – 9.962 & 9.454 & 5.68$\pm$0.69 & 9.474 & 0.346$\pm$0.154 9.962 – 10.95 & 10.44 & 4.20$\pm$0.65 & 10.46 & 0.222$\pm$0.135 10.95 – 11.93 & 11.43 & 3.32$\pm$0.61 & 11.45 & 0.079$\pm$0.124 11.93 – 12.92 & 12.42 & 3.23$\pm$0.61 & 12.43 & 0.156$\pm$0.120 12.92 – 13.90 & 13.40 & 2.55$\pm$0.58 & 13.41 & 0.008$\pm$0.113 13.90 – 14.87 & 14.38 & 3.05$\pm$0.57 & 14.39 & 0.067$\pm$0.104 14.87 – 15.84 & 15.35 & 3.27$\pm$0.57 & 15.36 & 0.143$\pm$0.103 15.84 – 16.81 & 16.32 & 2.59$\pm$0.55 & 16.33 & 0.212$\pm$0.093 16.81 – 22.52 & 19.46 & 1.01$\pm$0.43 & 19.77 & 0.024$\pm$0.010 [ccccc]{} 0.732 – 1.464 & 1.035 & 29.83$\pm$2.53 & 1.129 & 8.041$\pm$1.684 1.464 – 2.196 & 1.793 & 18.00$\pm$1.61 & 1.848 & 3.818$\pm$0.874 2.196 – 2.928 & 2.536 & 11.40$\pm$1.11 & 2.575 & 1.163$\pm$0.543 2.928 – 3.660 & 3.274 & 9.82$\pm$1.06 & 3.304 & 1.307$\pm$0.426 3.660 – 4.392 & 4.009 & 7.02$\pm$0.80 & 4.034 & 0.759$\pm$0.292 4.392 – 4.996 & 4.684 & 5.48$\pm$0.63 & 4.699 & 0.464$\pm$0.279 4.996 – 5.992 & 5.471 & 4.51$\pm$0.59 & 5.505 & 0.384$\pm$0.169 5.992 – 6.988 & 6.471 & 3.29$\pm$0.52 & 6.500 & 0.276$\pm$0.138 6.988 – 7.981 & 7.467 & 2.30$\pm$0.48 & 7.493 & 0.120$\pm$0.120 7.981 – 8.973 & 8.462 & 1.97$\pm$0.45 & 8.484 & 0.092$\pm$0.097 8.973 – 10.95 & 9.912 & 1.62$\pm$0.31 & 9.985 & 0.070$\pm$0.044 10.95 – 13.90 & 12.33 & 1.18$\pm$0.23 & 12.47 & 0.054$\pm$0.026 13.90 – 16.81 & 15.28 & 0.66$\pm$0.22 & 15.39 & 0.028$\pm$0.023 16.81 – 19.69 & 18.19 & 0.37$\pm$0.21 & 18.28 & 0.010$\pm$0.005 [ccccc]{} 0.761 – 1.518 & 1.075 & 54.88$\pm$4.60 & 1.171 & 13.74$\pm$0.31 1.518 – 2.274 & 1.858 & 35.30$\pm$3.10 & 1.915 & 7.719$\pm$0.163 2.274 – 3.031 & 2.625 & 20.64$\pm$2.11 & 2.666 & 4.288$\pm$0.949 3.031 – 3.788 & 3.388 & 10.81$\pm$1.48 & 3.420 & 1.141$\pm$0.607 3.788 – 4.544 & 4.149 & 8.77$\pm$1.27 & 4.175 & 0.957$\pm$0.472 4.544 – 5.301 & 4.908 & 6.72$\pm$1.10 & 4.930 & 0.618$\pm$0.393 5.301 – 6.058 & 5.667 & 5.31$\pm$0.97 & 5.686 & 0.496$\pm$0.314 6.058 – 6.814 & 6.425 & 3.93$\pm$0.86 & 6.442 & 0.402$\pm$0.267 6.814 – 7.571 & 7.183 & 2.61$\pm$0.79 & 7.197 & 0.108$\pm$0.261 7.571 – 8.104 & 7.833 & 2.60$\pm$0.99 & 7.840 & 0.093$\pm$0.332 8.104 – 9.573 & 8.808 & 2.32$\pm$0.71 & 8.854 & 0.175$\pm$0.134 9.573 – 11.04 & 10.21 & 1.39$\pm$0.68 & 10.32 & 0.045$\pm$0.111 11.04 – 12.50 & 11.75 & 1.32$\pm$0.55 & 11.78 & 0.095$\pm$0.097 12.50 – 13.97 & 13.21 & 0.65$\pm$0.62 & 13.24 & 0.024$\pm$0.023 Given these radial distributions, it is a simple matter to derive the total [mass]{} surface density profiles of the GCSs. This requires, first, another multiplicative scaling of $N_{\rm cl}(R_{\rm gc})$ in each case, so that the number density of [all]{} globulars is properly reflected. (Note that the observations used here reach limiting magnitudes which are just at, or brighter than, the peak of the symmetric Gaussian that best fits the GCLF. This implies that most of the faint globulars are lost in the noise in these studies, and that half or fewer have been directly observed at any $R_{\rm gc}$.) These final correction factors are obtained from equation (\[eq:31\]) above, by setting $V_{{\rm lim},1}=\infty$ and putting $V_{{\rm lim},2}$ at the effective limit for each composite $N_{\rm cl}$. In this way, it is found that the M87 distribution in Table \[tab2\] should be multiplied by 1.795; the M49 densities in Table \[tab3\], by 2.743; and the NGC 1399 numbers, by 1.875. Once this is done, the total number densities are multiplied by a mean globular cluster mass to obtain the mass densities $\Sigma_{\rm cl}(R_{\rm gc})$. The mean cluster mass adopted here is $\langle m\rangle_{\rm cl}=2.4\times 10^5\,M_\odot$, which (as was mentioned in §2.1 above) is the value for the Milky Way GCS. In both its shape [and]{} its mass scale, the Galactic GCLF, or mass function, bears enough similarity to those of the cluster systems being discussed here (and, indeed, to most others: Harris 1991) that the same $\langle m\rangle_{\rm cl}$ may safely be applied to them as well. The only issue is whether or not there is any systematic radial trend in mean globular cluster masses. It was also pointed out in §2.1 that there are no strong radial dependences observed in the GCLF of M87 (McLaughlin et al. 1994; Harris et al. 1998), suggesting that $\langle m\rangle_{\rm cl}$ is constant with $r_{\rm gc}$ in this galaxy, and thus likely also in similar systems like M49 and NGC 1399. The question can be addressed in more detail only for the Milky Way, which—although clearly an imperfect substitute for these giant ellipticals—is the only galaxy where accurate luminosities and three-dimensional galactocentric distances both can be assigned to globulars on an individual basis. Figure \[fig2\] shows that there is no dependence on radius in the observed $\langle m\rangle_{\rm cl}$ of the Milky Way. The top panel plots the volume number density of globulars in the Galactic halo, derived from the database of Harris (1996); drawn as the solid line is a least-squares fit of a parametric density distribution in the family discussed by Dehnen (1993) and Tremaine et al. (1994). (Only metal-poor clusters have been counted, as these are the ones that display true halo kinematics: Zinn 1985; Minniti 1995.) In the bottom panel, the total $V$-band luminosities of the clusters (taken again from Harris 1996) have been summed separately in each radial bin, and converted to total cluster masses by applying a mass-to-light ratio of $\Upsilon_V=2$ (Mandushev, Spassova, & Staneva 1991; Pryor & Meylan 1993). The mass densities $\rho_{\rm cl}$ that result are shown as the circular points. The solid line now corresponds to the same $n_{\rm cl}$ fit from the top panel, but multiplied by the average mass calculated for all globulars with $2\,{\rm kpc}\le r_{\rm gc}\le40\,{\rm kpc}$. This shows that $\rho_{\rm cl}=\langle m\rangle_{\rm cl}\,n_{\rm cl}$ to a good approximation at every $r_{\rm gc}$, and therefore that the mean globular cluster mass is not a strong function of galactocentric radius. The other curves in the bottom panel of this Figure are discussed in §4.3. For now, the point is simply to lend some circumstantial support to the claim that $\langle m \rangle_{\rm cl} \simeq2.4\times10^5\,M_\odot$ should not vary significantly with galactocentric radius in M87, M49, and NGC 1399. =4.0truein The total projected mass densities for these three GCSs are shown as the filled circles in Fig. \[fig3\], where angular distances and areas have been converted to linear units using $D({\rm Virgo})=15$ Mpc and $D({\rm Fornax})= 16.5$ Mpc. The final $\Sigma_{\rm cl}$ have also been scaled upwards by additional factors of $300-400$ for this Figure, to facilitate direct comparisons with the mass surface densities of the galaxies’ stars (drawn as lines). These latter distributions are derived from published $B$-band surface brightness profiles (M87: de Vaucouleurs & Nieto 1978, 1979; M49: King 1978, Caon et al. 1994; NGC 1399: Caon et al. 1994), according to the standard relation $$\Sigma_{\rm stars}/(M_\odot\,{\rm pc}^{-2})= \Upsilon_B\times10^{0.4(27.05+A_B-\mu_B)} \label{eq:32}$$ for $\mu_B$ in units of mag arcsec$^{-2}$. (Extinctions of $A_B=0.09$ mag, 0, and 0 are adopted for M87, M49, and NGC 1399; see Burstein & Heiles 1984.) The stellar mass-to-light ratios $\Upsilon_B$ for all three galaxies come from van der Marel’s (1991) kinematic analysis of their cores; his values have been corrected for the distances $D$ assumed here, and are taken to be independent of $R_{\rm gc}$. =4.0truein The comparisons of $\Sigma_{\rm cl}$ and $\Sigma_{\rm stars}$ in Fig. \[fig3\] illustrate the $S_N$ “problems” in M87 and show that, contrary to previous claims, [they do not exist]{} in M49 and NGC 1399. As was indicated above, the global specific frequency of M49 is quite normal; the most recent estimate places it at $S_N=4.7\pm0.6$ (Lee et al. 1998). Given a stellar mass-to-light ratio of $\Upsilon_V=7.5$ \[which follows from $\Upsilon_B=10.4$ and $(B-V)=1.0$\], equation (\[eq:23\]) then implies that $M_{\rm gcs}/M_{\rm stars}=(1.7\pm0.2)\times10^{-3}$ globally in this galaxy. In keeping with this, the middle panel of Fig. \[fig3\] shows directly that $\Sigma_{\rm cl}/\Sigma_{\rm stars}\simeq2.3\times10^{-3}$ at essentially any $R_{\rm gc}\ga 15$ kpc. This is slightly higher than the global GCS mass fraction because the globular cluster surface density profile is shallower than the stellar distribution at [small]{} galactocentric radii. As was discussed in §2.2, such a feature might result, at least in part, from the dynamical destruction of globulars in the densest regions of the galaxy. \[Again, the theoretical work of Murali & Weinberg 1997b, or of Aguilar et al. 1988, suggests that the important spatial scale is generally of order a stellar effective radius, $R_{\rm eff}$. These are marked by vertical lines for the galaxies in Fig. \[fig3\]; note that $R_{\rm eff}({\rm M49})\simeq16$ kpc.\] This aside, M49 has been put forward as an example of the second specific frequency problem (beginning with Harris 1986, and continuing on to Lee et al. 1998) on the basis of comparisons at [large]{} $R_{\rm gc}\ga R_{\rm eff}$ between $\Sigma_{\rm cl}$ and the $\Sigma_{\rm stars}$ [derived from the surface photometry of King (1978)]{}. King’s data are shown as the dashed line in the middle of Fig. \[fig3\], and they clearly do suggest that the projected stellar density falls off more steeply with radius than the GCS density in M49. However, the solid line in this panel represents the more recent $\mu_B$ data of Caon et al. (1994), which include many more measurements at $R_{\rm gc}\ga30$ kpc and do [not]{} show any GCS-halo discrepancy. The local specific frequency, which is proportional to $\Sigma_{\rm cl}/\Sigma_{\rm stars}$, then takes on a roughly constant value throughout the outer halo of M49, and the second $S_N$ problem does not appear there. The differences between the light profiles of King (1978) and Caon et al. (1994) illustrate the difficulty of accurate sky subtraction at faint surface brightness levels, and make the point that background corrections are an important consideration in any discussion of galaxy (or GCS) structure on large spatial scales. The surface photometry of Caon et al. (1994) is adopted here for M49, and thus $\Sigma_{\rm cl}\propto\Sigma_{\rm stars}$ at large $R_{\rm gc}$; but this important caveat must be kept in mind. In §3.2, it is shown that the gas in M49 adds up to a small fraction of its stellar mass, in which case the local cluster formation efficiency can be estimated by ${\hbox{$\widehat\epsilon$}}_{\rm cl}=\rho_{\rm cl}/(\rho_{\rm gas}+\rho_{\rm stars})\simeq\rho_{\rm cl}/\rho_{\rm stars}$. The fact that $\Sigma_{\rm cl} \propto\Sigma_{\rm stars}$ therefore implies that $\epsilon_{\rm cl}$ was essentially independent of galactocentric radius beyond about $R_{\rm eff} \simeq16$ kpc. An extrapolation of this result to smaller distances requires that the systematic decrease in $\Sigma_{\rm cl}/\Sigma_{\rm stars}$ there has resulted entirely from the gradual destruction of (initially bound) globulars over a Hubble time. As [possible]{} anecdotal support for this interpretation, recall that globally $M_{\rm gcs}/M_{\rm stars}\simeq0.0017$ for M49, while locally $\Sigma_{\rm cl}/\Sigma_{\rm stars}\simeq0.0023$ at large radii. If the latter number is assumed to have held all the way to the center of the galaxy initially, then a global $M_{\rm gcs}^{\rm init}/M_{\rm stars}^{\rm init}\simeq0.0023$ is implied. In such a scenario, if $M_{\rm stars}$ were roughly conserved over a Hubble time while globulars were dynamically disrupted (their remains adding nothing significant to the total stellar mass), the global ratio of current to initial GCS masses would be $M_{\rm gcs}/M_{\rm gcs}^{\rm init}\simeq0.0017/0.0023\sim0.75$. This is surprisingly close to the result of the general GCS mass-function arguments in §2.1 (see eq. \[\[eq:28\]\]). However, this could be just a coincidence; it remains possible that $\epsilon_{\rm cl}$ really was lower at small galactocentric radii, i.e., that the GCS [formed]{} with a “core” which was somewhat larger than that of the unclustered stellar distribution, and which was only enhanced over time. The current data cannot settle the question either way. The bottom panel of Fig. \[fig3\] shows that the second $S_N$ problem does not exist beyond a stellar $R_{\rm eff}$ in NGC 1399, either. Every study of this GCS has appreciated this point, and it is not a new result (see the references cited above). (Note also that the surface brightness profile used here for NGC 1399, from Caon et al. 1994, is consistent with that of Killeen & Bicknell 1988 at large $R_{\rm gc}$.) On the other hand, the global $S_N$ in this galaxy—and therefore the local $\Sigma_{\rm cl}/\Sigma_{\rm stars}$ at any single $R_{\rm gc}$—has always been thought to be significantly higher than the normal, M49-like value of $\sim$5: the latest calculation claims $S_N=12\pm3$ (Kissler-Patig et al. 1997). However, it is immediately apparent from Fig. \[fig3\] that this is [not]{} the case: the bulk scaling of $1/0.0029$ required to match the GCS radial distribution to the stellar mass profile in NGC 1399 is identical, within the observational uncertainties, to the $1/0.0023$ needed in M49. Since the gas mass is negligible compared to the stellar mass on $<70$-kpc scales in NGC 1399, as it is in M49, ${\hbox{$\widehat\epsilon$}}_{\rm cl}\simeq\Sigma_{\rm cl}/\Sigma_{\rm stars}$ holds for both systems, and their cluster formation efficiencies are inferred to have been essentially identical. The spuriously high global $S_N$ values typically quoted for NGC 1399 can be traced back to the practice of assigning an absolute magnitude of $M_V\simeq-21.6$ to the galaxy (from the Third Reference Catalogue: de Vaucouleurs et al. 1991); but, in fact, $M_V\simeq-22.1$ is indicated by direct integration (to $R_{\rm gc}= 70$ kpc, assuming $D=16.5$ Mpc) of the surface brightness profiles of either Caon et al. (1994) or Killeen & Bicknell (1988). Given this brighter magnitude, and an estimated GCS population of ${\cal N}_{\rm tot}=4\,700\pm400$ projected to within 70 kpc of NGC 1399 (see §5), equation (\[eq:21\]) yields $S_N({\rm NGC\ 1399})=7.0\pm0.6$, much closer to the value for M49 (see also Ostrov et al. 1998). This is consistent with the alternate representation of the data in Fig. \[fig3\], and confirms that the [first]{} specific frequency problem is not an issue in NGC 1399. It appears, therefore, that there are no $S_N$ problems in either M49 or NGC 1399. There is no evidence for any fundamental variation in $\epsilon_{\rm cl}$ from the total populations of these two GCSs, and the local ${\hbox{$\widehat\epsilon$}}_{\rm cl}$ in each is independent of galactocentric radius beyond $\simeq1\,R_{\rm eff}$. These results lend some weight to the claim (§2.2) that the first and second specific frequency problems seem likely to appear, and should be solved, only together. Both of these problems are genuine in M87, as the top panel of Fig. \[fig3\] attests. (The surface photometry of de Vaucouleurs & Nieto 1978, which is the standard, is well sampled and is supported by the independent studies of Oemler 1976 and King 1978; and the total $M_V$ used to calculate the galaxy’s $S_N$ is taken from integration of its $\mu_B$ profile.) The second specific frequency problem is evidenced by the shallower decline of $\Sigma_{\rm cl}$ compared to $\Sigma_{\rm stars}$ at $R_{\rm gc}\ga R_{\rm eff}\simeq7$ kpc. This also causes the first problem, since scaling the total GCS density upwards by the factors appropriate for M49 or NGC 1399 obviously puts $\Sigma_{\rm cl}$ well in excess of $\Sigma_{\rm stars}$ throughout most of the galaxy. Indeed, the most recent estimate for the global $S_N$ of M87 is $14.1\pm1.6$ (Harris et al. 1998), which is fully a factor of three higher than that of M49. Here it is necessary to consider the distribution of hot gas—which is known to exist in great quantities around M87—in evaluating ${\hbox{$\widehat\epsilon$}}_{\rm cl}$ from equation (\[eq:210\]). Volume Densities ---------------- The volume density profiles of circumgalactic and intracluster gas are routinely derived in the course of X-ray studies of elliptical galaxies. To include such a halo component in this discussion, it is first necessary either to deproject observed GCS and stellar surface densities, or to compute the gas surface density via a standard projection integral. In general, the first of these options potentially has the most insight to offer, since projected quantities always mix information from a wide range of physical radii. An Appendix therefore develops a nonparametric deprojection algorithm which is explicitly designed to handle GCS data that have been binned in projected radius, $R_{\rm gc}$ (i.e., counted in circular annuli). This procedure returns a coarse volume density profile—$n_{\rm cl}$ vs. three-dimensional $r_{\rm gc}$—expressed as a series of average densities in a number of concentric spherical shells. The algorithm is preferably applied to surface density profiles that extend to fairly large radii (so that corrections for cluster populations beyond the observed field of view can be estimated with confidence, and are as small as possible), and it makes most sense for $N_{\rm cl}$ distributions that are monotonically decreasing functions of $R_{\rm gc}$ (otherwise, the derived $n_{\rm cl}$ can be zero or negative in some shells). After re-binning to smooth some of the original data, as described above, the composite GCS radial distributions in Tables \[tab2\] through \[tab4\] satisfy each of these criteria. Equation (\[eq:37\]) of the Appendix has therefore been applied to each of these $N_{\rm cl}$ profiles, with $f(R_{i-1},R_i)$ (a sort of “residual background” correction) calculated by assuming that $n_{\rm cl}\propto r_{\rm gc}^{-3}$ beyond the last observed radius of each GCS. (This reflects the fact that the projected $N_{\rm cl}$ shows roughly an $R_{\rm gc}^{-2}$ behavior at large $R_{\rm gc}$ in each case. In any event, only the outermost one or two points in each deprojected profile are affected—and then not greatly—by this choice.) The resulting $n_{\rm cl}$ are listed in the fifth columns of Tables \[tab2\], \[tab3\], and \[tab4\]. The radii in the first columns, which delimit projected annuli in reference to the surface density data, are identified with the inner and outer three-dimensional radii of spherical shells in reference to the volume densities. =4.0truein The deprojection algorithm is an iterative one: the volume density in a spherical shell with some mean radius is computed by first subtracting, from the projected density at that radius, the contribution of any globulars that actually lie at larger physical $r_{\rm gc}$. Thus, the uncertainties of the derived $n_{\rm cl}$ at any two radii in one system are not independent. The (1-$\sigma$) errorbars quoted in Tables \[tab2\], \[tab3\], and \[tab4\] are the standard deviations of results from 1000 different deprojections in each case. Each trial deprojection in each GCS begins with a different $N_{\rm cl}$ profile, in which the surface density for any given annulus is drawn at random from a Gaussian distribution that is centered on the value given in Column 3 of the appropriate Table, and that has a dispersion equal to the observational uncertainty there. =4.0truein The tabulated number densities $n_{\rm cl}$ have the same limiting magnitudes as the corresponding surface densities, and are therefore multiplied by the same two factors (so effectively scaling to $V_{\rm lim}=\infty$, and applying the mean globular cluster mass $\langle m\rangle_{\rm cl}=2.4\times10^5\, M_\odot$) to find the total GCS mass densities $\rho_{\rm cl}(r_{\rm gc})$. These distributions are plotted as the filled circles in Figs. \[fig4\], \[fig5\], and \[fig6\], where the density in each spherical shell has been put at a radius $\overline{r}_{\rm gc}$, given in Columns 4 of Tables \[tab2\]–\[tab4\], calculated according to equation (\[eq:39\]). As in Fig. \[fig3\], the $\rho_{\rm cl}$ profiles have been scaled up by factors of a few hundred in order to compare them directly with the stellar mass densities $\rho_{\rm stars}(r_{\rm gc})$ (thin solid lines in Figs. \[fig4\] to \[fig6\]). These have been obtained by applying the algorithm of the Appendix to the projected profiles $\Sigma_{\rm stars}(R_{\rm gc})$ derived above from $B$-band surface photometry (see eq. \[\[eq:32\]\]). It is immediately apparent that the good agreement seen in Fig. \[fig3\] between the GCS and stellar densities in NGC 1399 and M49 (for the data of Caon et al. 1994) persists after deprojection, as should be the case. More noteworthy is the fact that the discrepancy between $\rho_{\rm cl}$ and $\rho_{\rm stars}$ in M87, while certainly still present, is somewhat less pronounced than that between $\Sigma_{\rm cl}$ and $\Sigma_ {\rm stars}$. For example, the truly local picture given by Fig. \[fig4\] shows that the second $S_N$ problem here is confined to galactocentric radii $r_{\rm gc}\ga30$–40 kpc, and does not extend inward to the $\sim15$ kpc scales suggested by Fig. \[fig3\]. This just reflects the fact that the projection of any two volume densities inevitably amplifies any small differences that may actually exist between them, and artificially associates disparities at large $r_{\rm gc}$ with smaller projected radii. A similar comment applies to the comparison of $\rho_{\rm cl}$ with $\rho_{\rm stars}$ in M49: it is now clear, from Fig. \[fig5\], that any argument for the second $S_N$ problem in this galaxy could only be made for three-dimensional radii greater than about 50 kpc, and then (as in §3.1) only if King’s (1978) $\mu_B$ profile—which includes just [a single data point]{} at such large $r_{\rm gc}$—were preferred over that of Caon et al. (1994). =4.0truein The gas densities in these systems have been obtained from published [electron]{} number densities by way of the relation $${{\rho_{\rm gas}}\over{M_\odot\,{\rm pc}^{-3}}}= 2.865\times10^{-5}\,\left({\mu\over{0.60}}\right)\left({{n_e}\over{10^{-3}\, {\rm cm}^{-3}}}\right)\ , \label{eq:310}$$ where a mean particle mass $\mu\simeq0.6$ (in units of $m_H$) is equally appropriate for hot ($T\sim10^7$–$10^8$ K) and highly ionized plasmas of any composition from primordial to solar. The open squares for the M87 gas in Fig. \[fig4\] come from scaling the electron densities of Nulsen & Böhringer (1995) for the different Virgo distance assumed here (15 Mpc vs. their 20), according to $n_e\propto D^{-1/2}$. (These data were published as volume densities after Nulsen & Böhringer applied a deprojection algorithm, similar to that derived in the Appendix, to their [ROSAT]{} spectra.) The dotted line in Fig. \[fig4\] is a fit taken from McLaughlin (1999): $$\rho_{\rm gas}({\rm M87})=5.60\times10^{-6}\,M_\odot\,{\rm pc}^{-3}\, \left({r_{\rm gc}\over{1070\,{\rm kpc}}}\right)^{-1} \left(1+{r_{\rm gc}\over{1070\,{\rm kpc}}}\right)^{-3}\ , \label{eq:311}$$ so that, essentially, $\rho_{\rm gas}\propto r_{\rm gc}^{-1}$ over the area of interest here. This fit has been used to compute the total $\rho_{\rm gas}+ \rho_{\rm stars}$ at every radius with a measured stellar density; that sum is drawn as the bold solid line in the Figure. Similarly, the open squares in Fig. \[fig5\] here use the $n_e$ read from Irwin & Sarazin’s (1996) Figure 6, after scaling to $D=15$ Mpc from their assumed M49 distance of 25.8 Mpc. The dotted line in this Figure is a scaled fit (from Brighenti & Mathews 1998) to these data plus those of Trinchieri et al. (1986): $$\rho_{\rm gas}({\rm M49})=\left[ {{3.57\times10^{-3}}\over{{1+(r_{\rm gc}/0.807\,{\rm kpc})^2}}}\,+\, {{2.24\times10^{-4}}\over{{1+(r_{\rm gc}/7.18\,{\rm kpc})^{1.14}}}}\,-\, {{1.50\times10^{-5}}\over{{1+(r_{\rm gc}/75.6\,{\rm kpc})^{1.19}}}}\right] \ \ M_\odot\,{\rm pc}^{-3}\ . \label{eq:312}$$ The heavy black line again represents the sum of $\rho_{\rm gas}+\rho_{\rm stars}$. The broken line for $\rho_{\rm gas}$ in NGC 1399 (Fig. \[fig6\]) comes from Tsai’s (1993) fit to the X-ray observations of Killeen & Bicknell (1988), corrected for $D({\rm Fornax})= 16.5$ Mpc: $$\rho_{\rm gas}({\rm NGC\ 1399})=1.03\times10^{-2}\,M_\odot\,{\rm pc}^{-3}\, \left[1+\left({r_{\rm gc}\over{0.304\,{\rm kpc}}}\right)^2\right]^{-0.615}\ . \label{eq:313}$$ It is clear that stars outweigh the gas everywhere in M49 and NGC 1399: locally, $\rho_{\rm gas}/\rho_{\rm stars}\la0.3$ even at the largest radii in Figs. \[fig5\] and \[fig6\]; globally, $M_{\rm gas}/M_{\rm stars}\sim0.05$ within $r_{\rm gc}=100$ kpc and 70 kpc respectively. In both systems, then, $M_{\rm gcs}/(M_{\rm gas}+M_{\rm stars})\simeq M_{\rm gcs}/ M_{\rm stars}$ and (albeit to slightly lower accuracy at very large $r_{\rm gc}$) $\rho_{\rm cl}/(\rho_{\rm gas}+\rho_{\rm stars})\simeq\rho_{\rm cl}/ \rho_{\rm stars}$. It is indeed valid to use standard specific frequencies as direct measures of globular cluster formation efficiencies in these two cases, as was done at the end of §3.1 above. To be complete, however, all of the stars, gas, and globular clusters can now be used to estimate $\epsilon_{\rm cl}$ more precisely, according to the basic equation (\[eq:210\]). By summing the data at large $r_{\rm gc}$ in each of M49 and NGC 1399, then, average ${\hbox{$\widehat\epsilon$}}_{\rm cl}$ values are obtained: $${{M_{\rm gcs}(r_{\rm gc}>12.5\,{\rm kpc})}\over{M_{\rm gas}(r_{\rm gc}>12.5\, {\rm kpc})+M_{\rm stars}(r_{\rm gc}>12.5\,{\rm kpc})}}= 0.0023\pm0.0005\ \ \ \ \ \ {\rm (M49)} \label{eq:314}$$ and $${{M_{\rm gcs}(r_{\rm gc}>11\,{\rm kpc})}\over{M_{\rm gas}(r_{\rm gc}>11\,{\rm kpc})+M_{\rm stars}(r_{\rm gc}>11\,{\rm kpc})}}= 0.0029\pm0.0008\ \ \ \ \ \ {\rm (NGC\,1399)}\ , \label{eq:315}$$ which account for the scalings applied to $\Sigma_{\rm cl}$ and $\rho_{\rm cl}$ in Figs. \[fig3\], \[fig5\], and \[fig6\]. As has already been discussed, these are equally well interpreted as either local or global formation efficiencies, since the ratio $\rho_{\rm cl}/(\rho_{\rm gas}+\rho_{\rm stars})$ shows no variations with $r_{\rm gc}$ beyond a stellar effective radius (or even inside this, in NGC 1399), i.e., outside those regions of the GCSs which could be significantly dynamically evolved. (As in Fig. \[fig3\], the stellar $R_{\rm eff}$ are marked with vertical broken lines in Figs. \[fig4\], \[fig5\], and \[fig6\].) The situation is somewhat different for M87: referring to Fig. \[fig4\], $\rho_{\rm gas}/\rho_{\rm stars}>1$ at all radii $r_{\rm gc}\ga40$ kpc (which is where the second $S_N$ problem begins in the deprojected data), and $M_{\rm gas}/M_{\rm stars}\simeq0.35$–$0.40$ globally (i.e., integrated over scales $r_{\rm gc}\leq100$ kpc). The hot gas around M87 is therefore an important factor in calculating its cluster formation efficiency. As before, outside of the central regions of the galaxy it is found that $${{M_{\rm gcs}(r_{\rm gc}>11\,{\rm kpc})}\over{M_{\rm gas}(r_{\rm gc}>11\, {\rm kpc})+M_{\rm stars}(r_{\rm gc}>11\,{\rm kpc})}}= 0.0032\pm0.0005\ \ \ \ \ \ {\rm (M87)}\ , \label{eq:316}$$ explaining the scaling factor applied to the M87 GCS densities in Figs. \[fig3\] and \[fig4\]. Figure \[fig4\] shows that this average value is valid also as the local ${\hbox{$\widehat\epsilon$}}_{\rm cl}=\rho_{\rm cl}/(\rho_{\rm gas}+\rho_{\rm stars})$ at any galactocentric position beyond about $R_{\rm eff}=7$ kpc; as in M49 and NGC 1399, the inferred globular cluster formation efficiency in M87 was independent of $r_{\rm gc}$ at large radii. In this case, however, if the gas were not included in the calculation, much larger—and radially varying—formation efficiencies would have been indicated: for example, $M_{\rm gcs}(r_{\rm gc}>11\,{\rm kpc})/M_{\rm stars} (r_{\rm gc}>11\,{\rm kpc})=0.0056\pm0.0008$, and $M_{\rm gcs}(r_{\rm gc}>40\, {\rm kpc})/M_{\rm stars}(r_{\rm gc}>40\,{\rm kpc})=0.0095\pm0.0023$. Note that the first of these ratios, $M_{\rm gcs}/M_{\rm stars}=0.0056$ for $r_{\rm gc}> 11$ kpc, is a factor of 2.4 higher than the corresponding quantity in M49. In combination with the different stellar mass-to-light ratios in these two galaxies ($\Upsilon_B=14.6$ vs. 10.4; see van der Marel 1991, or Fig. \[fig3\] above), this just corresponds to the factor of 3 difference observed in their global $S_N$ values. The high local and global specific frequencies in M87 are thus seen to result from the presence of a population of globular clusters that are intimately associated with the hot gas there. Gas Surface Densities --------------------- As was mentioned earlier, an alternate comparison of a galaxy’s gas distribution with its stars and globular clusters could make use of only surface densities—$\Sigma_{\rm stars}$, $\Sigma_{\rm cl}$, and a derived $\Sigma_{\rm gas}$—as functions of projected $R_{\rm gc}$. The advantage of such an approach is that $\Sigma_{\rm gas}$ is on an equal physical footing with $\Sigma_{\rm stars}$ and $\Sigma_{\rm cl}$, and the latter are basically unmanipulated data with smaller relative uncertainties than the derived $\rho_{\rm stars}$ and $\rho_{\rm cl}$ profiles. The cost, aside from possible uncertainties in any extrapolations required to obtain $\Sigma_{\rm gas}$ from $\rho_{\rm gas}$, is that differences between projected distributions are not easily related to local effects at some true, three-dimensional radius (e.g., recall the comparison between projected $\Sigma_{\rm cl}$ vs. $\Sigma_{\rm stars}$ and deprojected $\rho_{\rm cl}$ vs. $\rho_{\rm stars}$ in M87). More to the point, if the ratio $\Sigma_{\rm cl}/(\Sigma_{\rm gas}+\Sigma_{\rm stars})$ varies as a function of $R_{\rm gc}$ in a galaxy, then variations in the physical quantity of real interest, the local ${\hbox{$\widehat\epsilon$}}_{\rm cl}=\rho_{\rm cl}/(\rho_{\rm gas}+\rho_{\rm stars})$, are clearly indicated, but [cannot be properly quantified]{} without resorting to a deprojection of some sort. For this reason, the procedure of §3.2 must be favored as the best way, generally speaking, to measure local ${\hbox{$\widehat\epsilon$}}_{\rm cl}$. That said, however, in the galaxies studied here, ${\hbox{$\widehat\epsilon$}}_{\rm cl}$ is found to be independent of radius at large $r_{\rm gc}$, i.e., $\rho_{\rm cl}\propto (\rho_{\rm gas}+\rho_{\rm stars})$. In this specific instance, it should then also happen that $\Sigma_{\rm cl}\propto(\Sigma_{\rm gas}+\Sigma_{\rm stars})$ at large $R_{\rm gc}$, with a constant of proportionality that is exactly ${\hbox{$\widehat\epsilon$}}_{\rm cl}$. It is therefore useful to compute $\Sigma_{\rm gas}$ and the projected analogue of ${\hbox{$\widehat\epsilon$}}_{\rm cl}$ in M87, M49, and NGC 1399, in order both to verify the results of §3.2 (which depend in good part on the behavior of $\rho_{\rm cl}$ at large $r_{\rm gc}$, where the deprojected data are noisiest) and to obtain potentially higher-precision estimates of the cluster formation efficiencies in these systems. First, then, the projected mass densities of gas are obtained from the volume density profiles in equations (\[eq:311\]), (\[eq:312\]), and (\[eq:313\]), by way of the standard integral $$\Sigma_{\rm gas}(R_{\rm gc})=2\int_0^{\sqrt{r_{\rm max}^2-R_{\rm gc}^2}} \rho_{\rm gas}(r_{\rm gc})\,dz\ , \label{eq:317}$$ where the coordinate $z=(r_{\rm gc}^2-R_{\rm gc}^2)^{1/2}$ measures distance along the line of sight, and $r_{\rm max}$ represents the total physical extent of the X-ray corona. For M87, this is taken to be the virial radius of the Virgo Cluster: $r_{\rm max}\simeq1.5$ Mpc (see McLaughlin 1999). For M49, which is not at the center of Virgo, $r_{\rm max}=500$ kpc is adopted. NGC 1399 lies at the center of the Fornax Cluster, which is somewhat less massive than Virgo (e.g., Jones et al. 1997), and $r_{\rm max}$ is set at 1 Mpc there. There is a certain arbitrariness to each of these choices for $r_{\rm max}$, but any of them can be changed by a factor of a few without affecting the conclusions which follow. =4.0truein The resulting $\Sigma_{\rm gas}$ profiles are shown in Fig. \[fig7\], along with the same $\Sigma_{\rm stars}$ and $\Sigma_{\rm cl}$ profiles plotted in Fig. \[fig3\] above (although the $\Sigma_{\rm cl}$ have been scaled up by slightly different factors in the current Figure). It is easily seen that $\Sigma_{\rm cl}\propto(\Sigma_{\rm gas}+\Sigma_{\rm stars})$ at large $R_{\rm gc}$ in each galaxy, just as expected following the analysis in §3.2. Moreover, the constants of proportionality are in reasonable agreement with the ${\hbox{$\widehat\epsilon$}}_{\rm cl}$ derived from the deprojected data: by taking the direct mean of the surface-density ratios beyond the galaxies’ effective radii (so as once again to avoid a zone of potential GCS erosion), it is found that $$\left\langle{{\Sigma_{\rm cl}}\over{\Sigma_{\rm gas}+\Sigma_{\rm stars}}} \right\rangle_{R_{\rm gc}\ga R_{\rm eff}}\,=\,\left\{ \begin{array}{ll} 0.0026\pm0.0003\ , & \ \ {\rm M87} \\ 0.0022\pm0.0004\ , & \ \ {\rm M49} \\ 0.0026\pm0.0006\ , & \ \ {\rm NGC\,1399}\ . \end{array} \right. \label{eq:318}$$ The result for M49 is indistinguishable from that obtained in either §3.1 or §3.2, owing to the truly low gas densities in this galaxy. Similarly, the projected ${\hbox{$\widehat\epsilon$}}_{\rm cl}$ in NGC 1399 is identical, within the uncertainties, to the deprojected quantity $\rho_{\rm cl}/(\rho_{\rm gas}+ \rho_{\rm stars})=0.0029\pm0.0008$. The number in equation (\[eq:318\]) is formally 10% smaller because of the apparently large projected gas densities at radii $R_{\rm gc}\ga40$–50 kpc in NGC 1399. However, these values are almost certainly overestimates: in deriving $\Sigma_{\rm gas}$, the shallow volume density profile of equation (\[eq:312\]), which roughly has $\rho_{\rm gas}\propto r_{\rm gc}^{-1.2}$, has been assumed to hold even on Mpc spatial scales in Fornax—well beyond the range of actual X-ray observations—and this is likely unrealistic. On the other hand, the projected ${\hbox{$\widehat\epsilon$}}_{\rm cl}=0.0026\pm0.0003$ indicated for M87 in Fig. \[fig7\] might genuinely be slightly smaller than the $0.0032\pm0.0005$ inferred from deprojected quantities. The 20% difference between these two numbers is significant at just the 1-$\sigma$ level, however, and—although it is interesting that the former is actually closer to the values for M49 and NGC 1399—it is not obvious which is more reliable. The X-ray gas density around M87 (eq. \[\[eq:311\]\]) has again been extrapolated to unobserved radii to compute $\Sigma_{\rm gas}$, but in this case the basic profile actually follows the dark matter distribution in Virgo (McLaughlin 1999), and exhibits a realistic, steepening slope at large $r_{\rm gc}$. Thus, although it is possible that $\Sigma_{\rm gas}$ is overestimated at large $R_{\rm gc}$ in M87, any argument for this is not as compelling as in NGC 1399. The best measure of ${\hbox{$\widehat\epsilon$}}_{\rm cl}$ in M87 will therefore be taken as the error-weighted mean of the values (eqs. \[\[eq:316\]\] and \[\[eq:318\]\]) obtained by analyzing the deprojected and projected data there; in M49, the choice is of no consequence, but for definiteness the deprojected ${\hbox{$\widehat\epsilon$}}_{\rm cl}$ will be adopted; and the deprojected estimate will also be adopted for NGC 1399, to avoid any concerns that its gas surface density might have been overestimated at large $R_{\rm gc}$. Thus, $${\hbox{$\widehat\epsilon$}}_{\rm cl}=\left\{ \begin{array}{ll} 0.0028\pm0.0004\ , & \ \ {\rm M87} \\ 0.0023\pm0.0005\ , & \ \ {\rm M49} \\ 0.0029\pm0.0008\ , & \ \ {\rm NGC\ 1399}\ . \end{array} \right. \label{eq:319}$$ Once again, these can be seen as estimates of either local or global cluster formation efficiencies if the inner several kpc of M87 and M49 are excluded from the discussion. A Universal Efficiency for Cluster Formation ============================================ The results of the previous Section clearly show that [working in terms of stellar masses rather than luminosities and including the gas in M87 removes both the first and second specific frequency problems there]{}: in all three of M87, M49, and NGC 1399, it is found that $\rho_{\rm cl}\propto (\rho_{\rm gas}+\rho_{\rm stars})$, in identical proportions within the uncertainties, outside of the central, possibly dynamically depleted regions of the GCSs. (At the most, differences between the constants of proportionality in M87 and M49 might remain at the $\sim$20% level; but again, this is easily taken up in the observational uncertainties, and in any case is a significant improvement over the 300% difference in their specific frequencies.) The interpretation put to this is that globular clusters formed with an essentially constant efficiency throughout all of these galaxies, at the level of $$\epsilon_{\rm cl}({\rm globular})\equiv {{M_{\rm gcs}^{\rm init}}\over{M_{\rm gas}^{\rm init}}}\simeq {{M_{\rm gcs}}\over{M_{\rm gas}+M_{\rm stars}}}= 0.0026\pm0.0005\ , \label{eq:41}$$ which is the error-weighted mean of equation (\[eq:319\]). It should be kept in mind that these are all large galaxies with complex evolutionary histories. Thus, as was discussed in §2.1, what has been measured here is necessarily a [mass-weighted average]{} of $\epsilon_{\rm cl}$, possibly over multiple distinct episodes of in situ cluster formation and/or of the accretion of globulars formed elsewhere. Nevertheless, the uniformity of the observational estimates ${\hbox{$\widehat\epsilon$}}_{\rm cl}$ in these systems—both as a function of $r_{\rm gc}$ within each and from one entire system to another—suggests that these concerns may not be too severe, and that equation (\[eq:41\]) may still serve as a reliable guide to a single, possibly universal efficiency for cluster formation. This impression can now be checked quantitatively, albeit to rather lower precision than has proved possible with the excellent datasets for M87, M49, and NGC 1399, by appealing to observations of many other GCSs. All indications are that $\epsilon_{\rm cl}$ is indeed similar, at least to first order, from galaxy to galaxy and from the protogalactic epoch to the present. Other Ellipticals and BCGs -------------------------- Given the definition for ${\hbox{$\widehat\epsilon$}}_{\rm cl}$ (eq. \[\[eq:210\]\]), and denoting by $G$ the global ratio $M_{\rm gas}/M_{\rm stars}$ in a galaxy, the total population of any GCS may be written as $${\cal N}_{\rm tot}=2.92\times10^6\ {\hbox{$\widehat\epsilon$}}_{\rm cl}\,(1+G)\, \left({{L_{V,{\rm gal}}}\over{10^{11}\,L_\odot}}\right)\, \left({{\Upsilon_V}\over{7\,M_\odot\,L_\odot^{-1}}}\right)\, \left({{{\langle m\rangle}_{\rm cl}}\over{2.4\times10^5\,M_\odot}}\right)^{-1} \ . \label{eq:42}$$ This can, of course, also be derived from equation (\[eq:22\]) above, which explicitly relates $S_N$ to ${\hbox{$\widehat\epsilon$}}_{\rm cl}$ when $G=0$. Now, in the general discussion of §2, the stellar mass-to-light ratio was taken to be roughly similar from galaxy to galaxy. This is not exactly true, however: as was seen in §3, for example, the recovery of similar ${\hbox{$\widehat\epsilon$}}_{\rm cl}$ in M87 and M49 from their significantly different global and local $S_N$ depends on the $\simeq$40% larger $\Upsilon_B$ of M87 as well as on its greater gas mass. Thus, if other ellipticals are to be brought properly into this discussion, it is necessary first to account for the fact that $\Upsilon$ depends systematically on galaxy luminosity. This is a well known result of many fundamental-plane analyses of bright ellipticals, and it can be incorporated quantitatively in this analysis by referring to the study of van der Marel (1991; and references therein). Specifically, for a sample of 37 early-type systems, van der Marel obtains a mean $R$-band mass-to-light ratio of $\Upsilon_R=3.32h_{50}\,M_\odot\,L_ \odot^{-1}$, corresponding to $\Upsilon_B=5.93h_{50}$, at a mean magnitude of $M_B\simeq-22.22+5\,\log\,h_{50}$. For a galaxy color of $(B-V)=1.0$ (typical of large ellipticals) and $H_0=70$ km s$^{-1}$ Mpc$^{-1}$, this result corresponds to $\Upsilon_V=6.0$ at an absolute magnitude of $M_V\simeq-22.5$, or $L_{V,{\rm gal}}=8.5\times10^{10}\,L_\odot$. In addition to this, van der Marel (1991) finds that $\Upsilon$ increases with $L_{\rm gal}$ as a power law with exponent $0.35\pm0.05$. Since other studies of the fundamental plane have claimed a slightly shallower exponent of 0.25–0.30 (e.g., Faber et al. 1987; Pahre, Djorgovski, & de Carvalho 1995, and references therein), $\Upsilon_V\propto L_{V,{\rm gal}}^{0.3}$ is adopted here. Thus, $${{\Upsilon_V}\over{M_\odot\,L_\odot^{-1}}}=6.3\left({{L_{V,{\rm gal}}} \over{10^{11}\,L_\odot}}\right)^{0.3}\ . \label{eq:43}$$ It should be noted again that these mass-to-light ratios refer to the stellar populations in the cores of the galaxies, and are not unduly influenced by any dark matter on larger scales. If equation (\[eq:43\]) is used in (\[eq:42\]), the dependence ${\cal N}_{\rm tot}\propto L_{V,{\rm gal}}^ {1.3}$, or $S_N\propto L_{V,{\rm gal}}^{0.3}$, emerges naturally for constant ${\hbox{$\widehat\epsilon$}}_{\rm cl}$ and $G$. This is essentially the scaling found empirically by Kissler-Patig (1997). The other point of concern is, of course, the hot gas content of ellipticals in general. To get a rough handle on this, notice that the (soft) X-ray luminosities of early-type systems scale with their blue luminosities as $L_X\propto L_B^x$, with $x\simeq2$–3 (and with significant scatter about the mean trend: e.g., Forman et al. 1985; Brown & Bregman 1998). This scaling can be related to the ratio $G=M_{\rm gas}/M_{\rm stars}$ by noting that $L_X\propto \int \rho_{\rm gas}^2T_{\rm gas}^{1/2}\,dV \propto M_{\rm gas}^2T_{\rm gas}^{1/2}R^{-3}$ for the bremsstrahlung emission that dominates $L_X$ in this context. Also, the slope $x$ of the $L_X$–$L_B$ relation is the same if $L_V$ is used instead, so the implication is that $G^2 \propto R^3\, T_{\rm gas}^{-1/2} \Upsilon_V^{-2} L_V^{x-2}$. To go further, $T_{\rm gas}$ bears some relation to the stellar velocity dispersion: $T_{\rm gas}\propto \sigma^{2y}$, with $y\simeq0.7$–1.5 (Davis & White 1996; Brown & Bregman 1998); the $V$-band fundamental plane gives for the stellar quantities, $\sigma\propto L_V^{0.64}R^{-0.48}$ and $R\propto L_V^{1.1}$ (e.g., de Carvalho & Djorgovski 1989, 1992; Pahre et al. 1995); and, from equation (\[eq:43\]) just above, $\Upsilon_V\propto L_V^{0.3}$. Putting all of this together, $G^2\propto L_V^{x+0.7-0.11y}$. Evidently, the uncertainties in $x$ and $y$ are not crippling to this analysis, as $G\propto L_V^{1.5\pm0.25}$ is fairly representative of the range of possibilities. The scaling can be normalized to apply specifically to brightest cluster galaxies (BCGs) by using the observations of M87, which lies at the spatial and dynamical center of the Virgo Cluster. As was mentioned in §3.2, the global ($r_{\rm gc}\leq100$ kpc) ratio of gas to stars in M87 is $\simeq0.4$ by mass. Since the $V$-band luminosity of the galaxy is about $8\times10^{10}L_\odot$ (from de Vaucouleurs & Nieto 1978, for $D=15$ Mpc), it is then reasonable to set $$G\approx0.55\left({{L_{V,{\rm gal}}}\over{10^{11}\,L_\odot}}\right)^{1.5}\ . \label{eq:44}$$ Putting this and equation (\[eq:43\]) both into equation (\[eq:42\]), and holding ${\hbox{$\widehat\epsilon$}}_{\rm cl}$ fixed, it can be seen that the mean trend expected for ${\cal N}_{\rm tot}$ between $L_{V,{\rm gal}}=3\times10^{10} L_\odot$ and $2\times10^{11}L_\odot$ (so between an implied $G$ of $\simeq$0.1 and 1.6) is roughly ${\cal N}_{\rm tot}\propto L_{V,{\rm gal}}^{1.75}$. Despite the crudeness of the estimate of $G$ used here, the power 1.75 is just that inferred from observational scalings presented by Harris et al. (1998) for a large sample of BCGs. The full relation between ${\cal N}_{\rm tot}$ and $L_{V,{\rm gal}}$ is drawn as the heavy black line in Fig. \[fig8\], where the total GCS populations of 65 early-type galaxies are plotted against their $V$ luminosities. These data are the same as those plotted in Figure 11 of Harris et al. (1998), and the original references are given in that paper. The open circles in Fig. \[fig8\] here denote BCGs in a large number of rich clusters and a few poor groups; the filled symbols refer to other ellipticals, some in clusters and some more isolated. The bold curve has ${\hbox{$\widehat\epsilon$}}_{\rm cl}=M_{\rm gcs}/(M_{\rm gas}+M_{\rm stars})=0.0026$, constant in all galaxies. Also shown, as the dashed line, is the result of keeping this uniform ${\hbox{$\widehat\epsilon$}}_{\rm cl}$, but setting $G\equiv0$ in equation (\[eq:42\]) while still having $\Upsilon_V$ increase with luminosity as in equation (\[eq:43\]). The three dotted straight lines are loci of constant specific frequency, ${\cal N}_{\rm tot}\propto L_{V,{\rm gal}}$, as given by equation (\[eq:21\]) for (from top to bottom in the Figure) $S_N=15$, 5, and 1.5. =4.0truein It is immediately obvious why a roughly constant $S_N\simeq5$ is often assumed for large ellipticals such as these. It is equally clear, however, that such a simple characterization tends to overpredict ${\cal N}_{\rm tot}$ at low $L_{V,{\rm gal}}$, and systematically underestimates it at high luminosities. Accounting for $\Upsilon_V\propto L_{V,{\rm gal}}^{0.3}$ goes some of the way towards explaining this while allowing for a universal formation efficiency $\epsilon_{\rm cl}$; and referring the GCS populations in BCGs to the galaxies’ total stellar [plus gas]{} masses seems indeed to remove the rest of the first specific frequency problem in general. It is encouraging that these rather broad considerations lead to a significantly improved description of the observed ${\cal N}_{\rm tot}$–$L_{V, {\rm gal}}$ dependence, in which there are no obvious [systematic]{} deviations from a uniform ${\hbox{$\widehat\epsilon$}}_{\rm cl}\simeq0.0026$ (see also §4.4). There remains a good deal of scatter about the mean line in Fig. \[fig8\], but this could well reflect departures of individual galaxies from the fundamental plane, the mean $L_X$–$L_B$ (or $G$–$L_{V,{\rm gal}}$) relation, or the assumed $(B-V)=1.0$, rather than any large variations in the fundamental efficiency of cluster formation. Note, for example, that the observed mass-to-light ratio of M87 is $\Upsilon_V\simeq10.5$, some 80% higher than that expected from equation (\[eq:43\]). Correcting for this would bring the ${\hbox{$\widehat\epsilon$}}_{\rm cl}=0.0026$ curve in Fig. \[fig8\] just up to the M87 data point there. To rather a lesser degree, some of the scatter could also arise from GCS-to-GCS differences in the relative magnitude of any dynamical reduction of $M_{\rm gcs}^{\rm init}$ over a Hubble time. It should be noted that this explanation of the first $S_N$ problem differs somewhat from those previously proposed (see §2.2). In particular, Blakeslee (1997) and Blakeslee et al. (1997) showed a correlation between the specific frequencies on 40-kpc scales in BCGs, and the X-ray luminosities over much larger volumes ($R_{\rm gc}\le500$ kpc) in their host clusters. These authors, along with West et al. (1995) and Harris et al. (1998), therefore interpret the phenomenon of high specific frequency as one associated in some way with entire clusters of galaxies. The analysis presented here, however, has attempted to address the problem more locally, by referring to the gas and stellar masses on similar spatial scales (typically $r_{\rm gc}\la100$ kpc) to estimate $G$ (eq. \[\[eq:44\]\]) in individual systems; nowhere has any appeal been made to more global properties of galaxy clusters. This general approach is motivated, of course, by the observed situation in M87 specifically (§3). The correlation that does exist between BCG $S_N$ and cluster-wide $L_X$ (and other such relations; see the references cited) may simply stem from the fact that the gas around BCGs in rich clusters blends smoothly into the intracluster medium as a whole. Note that a relation of some kind between “local” $L_X$ and ${\cal N}_{\rm tot}$ is also implied by Santiago & Djorgovski (1993), although they attach no specific interpretation to it. Interestingly, the bold line in Fig. \[fig8\] departs significantly from linearity (i.e., from a simple power-law scaling between ${\cal N}_{\rm tot}$ and $L_{V,{\rm gal}}$) for $M_{\rm gas}/M_{\rm stars}\ga 0.1$–0.2, corresponding to $L_V\ga3$–$5\times10^{10}L_\odot$ or $M_V\la-21.5$. This magnitude is similar to that which appears to divide ellipticals into two broad classes in terms of their kinematics, isophote shapes, and light distributions. (Lower-luminosity ellipticals generally show rapid rotation, flattened and disky isophotes, and steep cusps in their central densities, while brighter systems are slowly rotating, with round and boxy isophotes and shallower central density profiles; see, e.g., Davies et al. 1983; Bender et al. 1989; Nieto, Bender, & Surma 1991; Tremblay & Merritt 1996; Gebhardt et al. 1996; Faber et al. 1997.) This bears directly on the recent discussion of Kissler-Patig (1997; see also van den Bergh 1998), who used observed trends in GCS specific frequencies and radial distributions as further evidence for the empirical “dichotomy” of ellipticals. In particular, Kissler-Patig noted that the cluster systems in early-type galaxies with $M_V\ga-21.5$ tend to have lower specific frequencies and steeper $N_{\rm cl}$ profiles than those in brighter galaxies. Although he went on to argue for a discontinuous change in the cluster formation efficiency for galaxies with $L_V$ below and above $\approx5\times10^{10} L_\odot$, Fig. \[fig8\] now shows that all of the data are well represented by a [continuously]{} increasing $S_N$ that derives from a [constant]{} $\epsilon_{\rm cl}$ and a [smoothly increasing]{} gas mass fraction, $G$; the nonlinearity intrinsic to the resulting ${\cal N}_{\rm tot}$–$L_{V,{\rm gal}}$ relation just happens to roughly mimic a break of sorts around $M_V\simeq-21.5$. This interpretation can also account for another correlation suggested by Kissler-Patig (1997), namely, that $S_N$ appears to be lower in (faint) ellipticals whose isophotes show disk-like deviations from pure ellipses, and higher in those (brighter) galaxies with more box-like shapes. These isophotal perturbations are themselves correlated with the X-ray luminosities of ellipticals (Bender et al. 1989), in the sense required qualitatively to explain this observation of GCSs again in terms of a constant $\epsilon_{\rm cl}$ and a systematically increasing gas-to-stellar mass ratio. This raises the interesting question of which, of gas content or isophote shape, is the more fundamental property of elliptical galaxies; the present discussion, at least, would appear to favor the first alternative. Finally, although the slopes of GCS radial distributions cannot be addressed directly by Fig. \[fig8\], it seems natural that, given a fixed $\epsilon_{\rm cl}$, ellipticals with large gas contents (and thus high global $S_N$) should generally have shallower GCS density distributions, relative to their stellar profiles, at large galactocentric radii (simply because $\rho_{\rm gas}/\rho_{\rm stars}$ usually increases with $r_{\rm gc}$ in early-type galaxies). As was suggested in §2.2, then, the first and second $S_N$ problems should be closely related, with $G=M_{\rm gas}/M_{\rm stars}$ the controlling factor in both. Again, this is consistent with the observed trend (noted by Forbes et al. 1997; Kissler-Patig 1997; and van den Bergh 1998) towards weaker dependences of $N_{\rm cl}$ on $R_{\rm gc}$ in brighter and higher-$S_N$ galaxies. This proposal should, of course, be checked on a case-by-case basis, as it was for M87 in §3.2 above, in as many systems as possible. Dwarf Ellipticals ----------------- If the bold line in Fig. \[fig8\] is extrapolated into the regime of dwarf ellipticals, $L_{V,{\rm gal}}\la2\times10^9L_\odot$, it clearly implies that global $S_N<2$ should be observed in these faint galaxies. This is completely at odds with the data of Durrell et al. (1996) and Miller et al. (1998), who find an average $S_N\approx5$ for a large sample of dE’s in the Local Group, in other groups nearby, and in the Virgo and Fornax Clusters. In addition, Miller et al. claim evidence for an [increasing $S_N$ towards fainter $L_{V,{\rm gal}}$]{}. This can be explained by first recalling that dwarf ellipticals do not fall on the fundamental plane of bright galaxies; thus, the scalings (\[eq:43\]) and (\[eq:44\]) do not apply to them. Also, these galaxies are bluer than their giant counterparts, so that the conversion to $V$-band mass-to-light ratios from measured $\Upsilon_R$, as described above, is again inapplicable; and in any case, even the cores of dwarfs appear to be dominated by dark matter, so that their dynamical $\Upsilon_V$ are not necessarily representative of the stellar values. Most importantly—and likely at the heart of these other points—dE’s present sufficiently shallow potential wells that they should have suffered significant amounts of mass loss during supernova-driven winds in early bouts of star formation. Thus, even though these galaxies are gas-poor at the present, the appropriate observable ${\hbox{$\widehat\epsilon$}}_{\rm cl}=M_{\rm gcs}/ M_{\rm stars}$ [is not an accurate estimator of the true cluster formation efficiency]{} $\epsilon_{\rm cl}=M_{\rm gcs}^{\rm init}/M_{\rm gas}^{\rm init}$ (see Durrell et al. 1996 for the first suggestion that this is an important consideration in understanding dwarf GCSs). Even still assuming that the GCSs now have $M_{\rm gcs}\simeq M_{\rm gcs}^{\rm init}$, allowance has to be made for an additional factor $M_{\rm stars}/M_{\rm gas}^ {\rm init}<1$. The relation analogous to equation (\[eq:42\]) in this case is therefore $${\cal N}_{\rm tot}\simeq1.7\times10^4\, {{M_{\rm gcs}}\over{M_{\rm gas}^{\rm init}}}\, \left({{M_{\rm gas}^{\rm init}}\over{M_{\rm stars}}}\right)\, \left({{L_{V,{\rm gal}}}\over{2\times10^9\,L_\odot}}\right)\, \left({{\Upsilon_V}\over{2\,M_\odot\,L_\odot^{-1}}}\right)\, \left({{\langle m\rangle_{\rm cl}}\over{2.4\times10^5\,M_\odot}}\right)^{-1}\ , \label{eq:45}$$ where $\Upsilon_V\simeq2$ should be a good representation of the stellar populations in systems fainter than about $2\times10^9L_\odot$. =4.0truein The specific model of Dekel & Silk (1986), which considers supernova-driven winds in dwarf galaxies that are dynamically dominated by dark matter halos, can be used to evaluate the ratio of initial gas to present stellar mass. That study finds that the winds from systems with velocity dispersions less than about 100 km s$^{-1}$, corresponding to $M_V\ga-18.5$ at the current epoch, should have expelled whatever gas remained after a single burst of star formation. Moreover, as one moves towards shallower potential wells and fainter $L_{V,{\rm gal}}$, the fractional mass lost, relative to the initial total, is naturally expected to increase. The predicted scaling, in the limit of very large mass loss and for a power spectrum similar to that in a standard cold dark matter cosmology, is roughly $M_{\rm gas}^{\rm init}/ M_{\rm stars}\propto L_{\rm gal}^{-0.4}$. Taking the crude step of extrapolating this all the way to the critical velocity dispersion (i.e., to $L_{V,{\rm gal}}\simeq2\times10^9L_\odot$), and assuming that any gas driven from brighter and more massive galaxies was a negligible fraction of the total, the implication is that $${{M_{\rm gas}^{\rm init}}\over{M_{\rm stars}}}\simeq \left({{L_{V,{\rm gal}} \over{2\times10^{9}\,L_\odot}}}\right)^{-0.4}\ ,\ \ \ \ \ L_{V,{\rm gal}}\le2\times10^9\,L_\odot\ . \label{eq:46}$$ Substituting this in equation (\[eq:45\]), and assuming a constant stellar $\Upsilon_V=2$, yields ${\cal N}_{\rm tot}\propto L_{V,{\rm gal}}^{0.6}$ and $S_N\propto L_{V,{\rm gal}}^{-0.4}$ for dwarf ellipticals. Although the picture of dE’s as single-burst populations is now known to be incomplete (see, e.g., Mateo 1998), this result is nevertheless consistent with the observations of Miller et al. (1998). Figure \[fig9\] uses the data from Durrell et al. (1996) and Miller et al. (1998) to place 32 nucleated and non-nucleated dwarf ellipticals into the $L_{V,{\rm gal}}-{\cal N}_{\rm tot}$ plane with the brighter galaxies from Fig. \[fig8\]. (Of these, 16 are taken from Durrell et al., and 16 from Miller et al. Only GCSs detected at the $\ge1$-$\sigma$ level are used.) The bold dashed line in the lower left of this plot comes from equations (\[eq:45\]) and (\[eq:46\]), assuming a constant $\epsilon_{\rm cl}=M_{\rm gcs}/M_{\rm gas}^{\rm init}=0.0026$. The heavy solid line is the same as that in Fig. \[fig8\]—i.e., it also has $\epsilon_{\rm cl}=0.0026$, but applies to fundamental-plane ellipticals—and the light, dotted lines correspond, as before, to constant $S_N=15$, 5, and 1.5. Once again, all the data appear consistent with a universal globular cluster formation efficiency. It will be important to properly confirm this with detailed models for the dE’s in particular, and to flesh it out with observations of systems at $L_{V,{\rm gal}}\sim(1-5)\times10^9L_\odot$ in general. Globular Clusters in the Milky Way ---------------------------------- One object that falls just inside this observational “gap” in luminosity is the stellar [spheroid]{}—disk excluded—of the Milky Way. The total $V$-band luminosity here has been estimated by de Vaucouleurs & Pence (1978) to be $L_V\simeq 5.1\times10^9 L_\odot$ \[after applying a color of $(B-V)=0.65$ to their $M_B=-18.8$\]. At the same time, the total number of globular clusters that are genuine halo objects, rather than part of the (thick) disk, is ${\cal N}_{\rm tot}\simeq110$–120. (Halo globulars are identified on the basis of their metallicity, ${\rm [Fe/H]}\leq-0.80$; refer back to Fig. \[fig2\] above.) This is the point plotted as the large open square in Fig. \[fig9\]. In fact, de Vaucouleurs & Pence (1978) model the Galactic spheroid with an $R^{1/4}$ density law with effective radius $R_{\rm eff}= 2.67$ kpc. Thus, in addition to predicting the absolute magnitude that would be seen by an external observer, it is also possible to parametrize the density profile of the entire stellar halo. To do this in three dimensions, Young’s (1976) deprojection of the $R^{1/4}$ law may be used: $${{\rho_{\rm stars}(r_{\rm gc})}\over{M_\odot\,{\rm pc}^{-3}}}=52.195\, \left({{\Upsilon_B}\over{M_\odot\,L_\odot^{-1}}}\right)\, \left({{L_B^{\rm tot}}\over{L_\odot}}\right)\, \left({{R_{\rm eff}}\over{{\rm pc}}}\right)^{-3}\, \exp\left[-7.669\left({{r_{\rm gc}}\over{R_{\rm eff}}}\right)^{1/4}\right]\, \left({{r_{\rm gc}}\over{R_{\rm eff}}}\right)^{-7/8}\ , \label{eq:47}$$ which is valid for $r_{\rm gc}\ga0.2 R_{\rm eff}$. Substitutions can be made for $L_B^{\rm tot}$ and $R_{\rm eff}$ directly from the work of de Vaucouleurs & Pence. One way to estimate the $B$-band mass-to-light ratio of the Galactic Population II is to adopt the same $\Upsilon$–$L_B$ scaling used in §4.1, but to normalize the relation by adopting solar colors rather than the redder ones that apply in ellipticals. van der Marel’s (1991) mean $\Upsilon_R=3.32h_{50}$ then corresponds to $\Upsilon_B=4.65 M_\odot L_\odot^{-1}$ at $M_B=-21.49$ (for $H_0=70$ km s$^{-1}$ Mpc$^{-1}$ assumed), and $\Upsilon_B\propto L_B^{0.3}$. The implied $\Upsilon_B({\rm halo})\simeq2.2 M_\odot L_\odot^{-1}$ for $M_B=-18.8$ is consistent with the mean $\Upsilon_V=2$ for globular clusters, and with the $\Upsilon_V=1.75$ adopted for the spheroid by Bahcall & Soneira (1980) in their model of the Galaxy. Equation (\[eq:47\]) is therefore evaluated here with $\Upsilon_B=2 M_\odot L_\odot^{-1}$. The dashed line in the bottom panel of Fig. \[fig2\] in §3.1 above shows this density profile, scaled down by a factor of $$\langle {\hbox{$\widehat\epsilon$}}_{\rm cl}\rangle_{\rm MW}=0.0027\pm0.0004\ . \label{eq:48}$$ This value is obtained as the mean of the ratio $\rho_{\rm cl}/\rho_{\rm stars}$ over all galactocentric radii $2\le r_{\rm gc}\le 40$ kpc, and it accounts for the position of the Milky Way halo in Fig. \[fig9\]. Such a basic correspondence between our Galaxy and early Hubble types was previously suggested, although developed from a somewhat different argument, by de Vaucouleurs (1993). It is also evident from Fig. \[fig2\] that there is no second $S_N$ problem in the Milky Way: $\rho_{\rm cl}\propto \rho_{\rm stars}$ obtains throughout the entire halo at $r_{\rm gc}\ga3$ kpc—so again, beyond roughly an effective radius, within which the original GCS might have suffered significant dynamical depletion. This point has long been appreciated, although it is usually discussed in the context of the fact that the same rough rule $\rho \propto r_{\rm gc}^{-3.5}$ is inferred both for the spheroid, from counts of RR Lyrae and horizontal-branch stars (e.g., Preston, Schectman, & Beers 1991; Kinman, Suntzeff, & Kraft 1994), and for the GCS, from simple power-law fits to its density profile (Harris 1976; Zinn 1985; Djorgovski & Meylan 1994). This power law is shown as the dotted line in the bottom panel of Fig. \[fig2\]. Scatter in the $L_V$–${\cal N}_{\rm tot}$ Plane ----------------------------------------------- =4.0truein As a final test of the general viability of a single globular cluster formation efficiency, Fig. \[fig10\] plots the ratio of observed to predicted GCS ${\cal N}_{\rm tot}$ as a function of host galaxy luminosity for all of the 98 systems discussed thus far. The predicted ${\cal N}_{\rm tot}$ for fundamental-plane ellipticals and the Milky Way spheroid (filled circles and open square in the Figure) are given by equation (\[eq:42\]), with the gas mass fraction $G$ set identically to zero; for BCGs (open circles), by the same relation but with $G$ as in equation (\[eq:44\]); and for dE’s (filled squares), by equation (\[eq:45\]). In all cases, it has been assumed that $\epsilon_{\rm cl}=0.0026$. The median of the resulting ratios ${\cal R}\equiv {\cal N}_{\rm tot}^{\rm obs}/{\cal N}_{\rm tot}^{\rm pred}$ is 1.04, and the geometric mean is 0.98. Close to 80% (77/98) of these galaxies have $\frac{1}{2}\le{\cal R}\le2$. As was also mentioned in §4.1, but can perhaps be seen more clearly here than in Fig. \[fig8\] or Fig. \[fig9\], the deviations of observed ${\cal N}_ {\rm tot}$ from the uniform-$\epsilon_{\rm cl}$ prediction show no significant dependence on either galaxy luminosity or Hubble type. There are possibly two exceptions to this claim: (1) the GCS populations in the brightest BCGs, $L_{V,{\rm gal}}\ga 10^{11}\,L_\odot$, appear to fall below the values predicted with $\epsilon_{\rm cl}=0.0026$, by perhaps $\sim50\%$ on average; and (2) the two faintest dwarfs, which happen to be the Local Group spheroidals Fornax and Sagittarius, both have ${\cal N}_{\rm tot}$ greater than predicted. These rough impressions may well not remain, in the mean, after future additions of GCS data from other very high- and low-luminosity galaxies; but even if they do, they need not imply the existence of fundamental variations in $\epsilon_{\rm cl}$. Rather, there is likely enough uncertainty in the adopted relation between gas mass fraction and $L_{V,{\rm gal}}$ (eq. \[\[eq:44\]\]) to account for the first item; and the second could reflect either a deficiency in the very simple correction (§4.2) for gas blow-out from very faint dwarfs, or the neglect of possibly extensive post-formation dynamical evolution in such low-mass galaxies (such as tidal stripping of their field stars; recall, for example, that Sagittarius is in the process of being accreted by the Milky Way). Even if these issues are ignored completely, a least-squares fit to all of the data shown in Fig. \[fig10\] reveals an exceedingly weak dependence of ${\cal N}_{\rm tot}^{\rm obs}/{\cal N}_{\rm tot}^{\rm pred}$ on galaxy luminosity: ${\cal R}\propto L_{V,{\rm gal}}^{-0.04\pm0.03}$. Fitting to only the fundamental-plane ellipticals and BCGs yields a steeper ${\cal R}\propto L_{V,{\rm gal}}^{-0.2}$ or so, but this result appears to be driven by just three gE’s with anomalously low or high GCS populations (cf. Fig. \[fig8\]): NGC 3557, with $L_V=9.4\times10^{10}\,L_\odot$ and $S_N=0.4\pm0.3$; NGC 5018, which has $L_V=7.8\times10^{10}\,L_\odot$ and $S_N=0.9\pm0.3$; and NGC 4278, with $L_V=7.1\times10^9\,L_\odot$ but $S_N= 12.3\pm1.4$. If these systems are excluded, then ${\cal R}({\rm gE + BCG}) \propto L_{V,{\rm gal}}^{0.02\pm0.04}$ is indicated instead. Similarly, the slopes of ${\cal R}$ vs. $L_{V,{\rm gal}}$ do not differ significantly from 0 for any of the individual Hubble-type samples represented in Fig. \[fig10\]. Thus, current observations of total GCS populations do show some scatter, and it will be important to understand—likely through detailed analyses, along the lines of §3, of more systems [individually]{}—whether or not this reflects an intrinsic scatter in the underlying globular cluster formation efficiency (see also §4.1). However, there is no clear evidence at present for any systematic departures from a mean global efficiency of $\epsilon_{\rm cl}\approx0.0026$. Open Clusters ------------- As was discussed in §2.1, the formation of stellar clusters continues to be an important element of star formation in general, and the existence of an apparently universal formation efficiency for globular clusters must reflect a generic piece of this process. However, the picture may still be incomplete: although the discussion to this point has included clusters that formed in a wide variety of environments—at galactocentric distances $r_{\rm gc}\sim 10$–100 kpc, in galaxies spanning four orders of magnitude in total luminosity and inhabiting both poor groups and rich clusters—all of these formed at essentially a single epoch, and under physical conditions that were possibly much more extreme than those in, say, the disk of the Milky Way today. It is important, therefore, to compare the protogalactic $\epsilon_{\rm cl} \simeq0.0026\pm0.0005$ to the frequency with which bound star clusters form in a typical quiescent galaxy at the current epoch. Disk clusters in the Milky Way present one opportunity to do this. The rate of open cluster formation in the solar neighborhood has been estimated by Elmegreen & Clemens (1985) as ${\cal S}_{\rm cl} \simeq(2.5\pm1)\times10^{-7}$ clusters kpc$^{-2}$ yr$^{-1}$. If the masses of the clusters range between $100\,M_\odot$ and $5\,000\,M_\odot$ according to $d{\cal N}/dm\propto m^{-1.5}$ (from the luminosity spectrum $d{\cal N}/dL$ of van den Bergh & Lafontaine 1984), then a mean mass of $\simeq700\,M_\odot$ is indicated, and the formation rate becomes ${\cal S}_ {\rm cl}\simeq(1.8\pm0.7)\times10^{-4}M_\odot$ kpc$^{-2}$ yr$^{-1}$. This is to be compared with the [total]{} star formation rate in the solar neighborhood, which includes those stars appearing in loose groups and unbound associations (or even in relative isolation) as well as in bound clusters. This rate is ${\cal S}_{\rm tot}\simeq4.5\times10^{-3}M_\odot$ kpc$^{-2}$ yr$^{-1}$ (e.g., McKee 1989). Since the mean stellar mass is no different in clusters than in the field, the ratio of these two rates can be taken directly, to assess the local [number]{} fraction of stars born in clusters: ${\cal S}_{\rm cl}/{\cal S}_{\rm tot}\sim0.04$. This is consistent with the classic estimate (e.g., Roberts 1957) of about 10%. However, the globular cluster ${\hbox{$\widehat\epsilon$}}_{\rm cl}$ has been obtained here as a [mass]{} ratio, since this is the observable quantity that can be related directly to its formation value (see eq. \[\[eq:28\]\]). To make a proper comparison with open cluster formation, it is therefore necessary to refer the mass going into newborn clusters (per unit area, per unit time) to the total mass of gas available for any kind of star formation. This latter figure can be derived from ${\cal S}_ {\rm tot}$ using the mean star formation efficiency, $\langle {\rm SFE}\rangle= \langle M_{\rm stars}/M_{\rm gas}\rangle$, averaged over many of the dense, star-forming (not necessarily cluster-forming) cores in giant molecular clouds. This number is only poorly known, but statistical arguments (e.g., Elmegreen 1983) suggest that it lies at the level of $\langle {\rm SFE}\rangle\sim5$%–10%. Thus, $$\epsilon_{\rm cl}({\rm open})\simeq {{{\cal S}_{\rm cl}}\over{{\cal S}_{\rm tot}/\langle {\rm SFE}\rangle}} \sim0.002-0.004\ , \label{eq:49}$$ which is in reasonably good agreement with the result for globular cluster systems. The numerical value in equation (\[eq:49\]) clearly should not be taken as definitive, given the very rough, order-of-magnitude nature of the argument that has produced it. That said, however, it is just the order-of-magnitude agreement between equations (\[eq:49\]) and (\[eq:41\]) which suggests that the basic mechanism of cluster formation may not have changed appreciably from protogalactic times to the present. Another obvious place to check the current level of $\epsilon_{\rm cl}$ is in starbursts and merging galaxies, which are forming massive young clusters in substantial numbers (e.g., Whitmore & Schweizer 1995; Meurer et al. 1995). Complete inventories of the star-forming gas mass in individual systems, and of the total mass in those of their young clusters that can be shown to be gravitationally bound (which to date has been done rigorously for only two objects: Ho & Filippenko 1996a, 1996b), would be of considerable interest. The results should be of particular relevance to the oft-made claim, which is now at the center of some debate (Sternberg 1998; Brodie et al. 1998), that these super star clusters are the modern-day equivalents of young globulars. If it turns out, for example, that strongly bound clusters in starbursts are forming in numbers very much in excess of those expected on the basis of $\epsilon_{\rm cl}\simeq0.25\%$ (as, e.g., the UV observations of Meurer et al. 1995 appear to suggest), then these spectacular events might [not]{} be representative of a ubiquitous phase in the formation of average elliptical galaxies. Discussion ========== The fact that the $\epsilon_{\rm cl}$ values inferred here are everywhere so similar—over large ranges in radius inside M87, M49, and NGC 1399; from dwarf ellipticals to brightest cluster galaxies; and from massive, old globular clusters to smaller, much younger open clusters—should serve as a strong constraint on theories of star formation. The simplest interpretation is that the probability of attaining a cumulative star formation efficiency of ${\rm SFE}\ga20\%$–50% (see §2.1) in any dense clump of gas more massive than $\sim10^2$–$10^3\,M_\odot$ (so from open clusters on up) can depend only weakly, on average, on local environment (i.e., the density and pressure of ambient gas) or on factors such as global background density (meaning, e.g., in the protogalactic context, a 1-$\sigma$ vs. a 3-$\sigma$ density fluctuation; or a poor group of galaxies vs. a rich cluster). Moreover, implicit in plots like Figs. \[fig8\] and \[fig9\] is the fact that cluster formation efficiencies are also rather insensitive to metallicity: the mean abundances of the globulars in the galaxies represented there range over two orders of magnitude, from $[{\rm Fe/H}]\simeq-2$ to $[{\rm Fe/H}]\simeq0$ (see Durrell et al. 1996; Forbes et al. 1997). Over a remarkably broad spectrum of physical conditions, therefore, it appears that similar fractions of star-forming gas—always about 0.25% by mass—are able to produce bound stellar clusters. The implicitly mass-averaged nature of the observational estimates of $\epsilon_{\rm cl}$ has to be kept in mind, but would seem to be of little concern in most cases. It is therefore reasonable to place the efficiency of cluster formation alongside the globular cluster mass function $d{\cal N}/dm$ (McLaughlin & Pudritz 1996) and the mean globular cluster mass $\langle m\rangle_{\rm cl}$ (e.g., Fig. \[fig2\] above) as robust and nearly universal physical quantities that must be derived from the most general aspects of the star formation process. As a corollary to this, one implication of Figs. \[fig8\] and \[fig9\] above is that the efficiency of [unclustered]{} star formation could [not]{} have been universal. That is, in both dwarf elliptical and brightest cluster galaxies, where the data are consistent with $\epsilon_{\rm cl}=M_{\rm gcs}^{\rm init}/M_{\rm gas}^{\rm init}$ having been constant in the mean, globular clusters apparently did form in precisely the numbers expected for the total amount of gas initially available to these systems. But then, the higher than average specific frequencies there are inferred to reflect larger ratios of $M_{\rm gas}^{\rm init}/M_{\rm stars}$, so that lower than average amounts of this gas were converted into field stars. The data for dwarfs agree with a simple scenario in which most of the unused gas supply is driven entirely out of the galaxies by a strong wind following one major burst of star and cluster formation. If this is basically correct, then the globulars in dE’s quite possibly were able to form completely before the onset of such a wind, and they could conceivably have been instrumental in driving it; at the very least, if the protoclusters were still largely gaseous, they must have been already sufficiently dense and well defined that they could survive such a catastrophic event essentially intact. At the same time, field star formation must have been truncated by the wind, before reaching what would have been a “normal” efficiency of $M_{\rm stars}/M_{\rm gas}^{\rm init}=1-\epsilon_{\rm cl}$. Thus, [the massive and dense clumps of gas which ultimately formed bound star clusters had to have collapsed more rapidly than those which produced unbound groups and associations]{}. This conclusion may hold quite generally—one possible interpretation being that only those gas clouds in the highest-density tail of some distribution, in any setting, are able to achieve a high cumulative SFE—and it may be related to the fact that the globular clusters in dwarfs are generally somehat bluer, and presumably more metal-poor, than the bulk of the field stars there (Durrell et al. 1996; Miller et al. 1998) A similar feedback argument could also apply to BCGs, where it may be that especially strong early bursts of star and cluster formation led to the premature virialization of large amounts of protogalactic gas, and perhaps drove slow, [partial]{} galactic winds, most effectively in the low-density environs at large galactocentric radii (cf. Harris et al. 1998). The unused gas would have to remain hot to the present day, and more or less in the vicinity of the parent galaxy, in order to enter the observed $L_X$–$L_B$ relation as inferred in §4.1. This requirement is consistent with the fact that BCGs are situated at or near the centers of their galaxy clusters, and thus at the bottoms of very deep potential wells. In this scenario, again, globulars would have had to form, in just the numbers expected of them, somewhat before field stars were able to do the same; and this truncation of unclustered star formation would have had to be more severe at large galactocentric radii (where the feedback is strongest). This could then account explicitly for the connection between the first and second specific frequency problems, i.e., between high global $S_N$ and local $\Sigma_{\rm cl}/\Sigma_{\rm stars}$ ratios that increase with $R_{\rm gc}$. It also gives a more specific context to the general claim (Harris 1986) that the shallowness of GCS radial distributions relative to the stellar light profiles in some ellipticals resulted from the globulars having formed slightly in advance of the stars there. The main demand on this picture for BCG formation is that it must be able to explain why $M_{\rm gas}/M_{\rm stars}$ increases with galaxy luminosity, i.e., why more massive BCGs apparently had larger fractions of their initial gas mass virialized early on. As Harris et al. (1998) discuss in some detail, this issue is likely related to the fact that brighter BCGs are generally found in more massive galaxy clusters, which may have presented higher-density and more turbulent environments that led to systematically higher star formation rates, and more violent feedback. Although speculative, these ideas are suggested directly by the GCS data, and are consistent with all those available. Since they are also obviously related to questions on the star-formation histories of dE’s, on the origin of the $L_X$–$L_B$ correlation in the brightest ellipticals, and on the nature of the intracluster medium, they should be addressed with considerably more rigor than has been applied here. It is particularly striking that the cluster formation efficiency derived in §4.3 for the Milky Way halo is essentially indistinguishable from that which applied in M87, M49, and NGC 1399, and which appears to have held quite generally in ellipticals of most any description. This calls into serious question one of the primary motivations for the model of Ashman & Zepf (1992) and Zepf & Ashman (1993), who posit that ellipticals are formed primarily by the mergers of gas-rich spirals, and invoke preferential cluster formation during these events to account for what they claim is a typical factor of two difference in the observed ratios $M_{\rm gcs}/M_{\rm stars}$ of late- and early-type systems. Although it is true that spirals have specific frequencies that are lower than those in bright E galaxies, by factors of perhaps 2–3 on average (as is exemplified by the Milky Way in Fig. \[fig9\] above; see also Harris 1991), it is now apparent that this is not necessarily a reflection of similar discrepancies in their more fundamental ${\hbox{$\widehat\epsilon$}}_{\rm cl}$. If the Milky Way is typical, it suggests that—even though stellar clusters certainly can form in merging galaxies—[mergers are not required to explain the relative GCS populations of spirals and ellipticals in general]{}. Conversely, whenever an elliptical does form by the major merger of two spirals, this is allowed to include cluster formation only at the standard efficiency of $\epsilon_{\rm cl}\simeq0.0025$. Ashman & Zepf (1992) and Zepf & Ashman (1993) also attribute the second specific frequency problem to the formation of ellipticals by mergers of spirals with pre-existing GCSs. However, it is not clear that this is necessary either, if the extended spatial distribution of globulars in the brightest ellipticals is generally related to their association with hot gas at rather large galactocentric radii. In fact, a basic expectation of the original Ashman-Zepf scenario is that the second $S_N$ problem should be [less]{} pronounced in galaxies with higher global $S_N$ (which they hypothesize to have suffered greater numbers of major mergers). The sense of this predicted trend is [opposite]{} to the one implied here—with empirical support directly from the evidence in M87, and indirectly from the correlations of Forbes et al. (1997) and Kissler-Patig (1997)—in which a constant $\epsilon_{\rm cl}$ combines with the larger $M_{\rm gas}/M_{\rm stars}$ in higher-$S_N$ ellipticals, to produce greater contrasts in their projected $\Sigma_{\rm cl}$ and $\Sigma_{\rm stars}$ profiles. Although radial-profile and $S_N$ data by themselves may not rule out the merger model for GCSs altogether, a revision to take account of the hot gas in large ellipticals is essential if it is to be at all viable. (See also Forbes et al. 1997 and Kissler-Patig et al. 1998 for critical discussions of this and other aspects of the basic idea.) [lccclcclcc]{} M87 & 2.53$\pm$0.17 & 0.73$\pm$0.27 & 2.177$^{+0.248}_{-0.193}$ & & 2.10$\pm$0.40 & 1.522$^{+ 0.220}_{- 0.180}$ & & $\ldots$ & $\ldots$ & & & & & & & & & M49 & 2.73$\pm$0.27 & 1.41$\pm$0.56 & 1.170$^{+0.356}_{-0.187}$ & & 2.25$\pm$0.45 & 0.985$^{+0.285}_{-0.219}$ & & 6.18$\pm$1.32 & 3260$\pm$475 & & & & & & & & & N1399 & 3.15$\pm$0.35 & 1.54$\pm$0.46 & 1.466$^{+0.233}_{-0.124}$ & & 1.15$\pm$0.45 & 1.859$^{+0.543}_{-0.351}$ & & 3.78$\pm$0.92 & 3630$\pm$460 Finally, the GCS density profiles derived specifically for M87, M49, and NGC 1399 in §3 also are relevant to issues of the generic spatial structure of GCSs, and to the specific matter of the possible existence of intergalactic globulars in clusters of galaxies. To facilitate a discussion of these points, and for reference, Table \[tab5\] presents the results of least-squares fits of a few different parametric functions to the spatial distributions of the M87, M49, and NGC 1399 GCSs, as they are given in Tables \[tab2\], \[tab3\], and \[tab4\] above. The functions include (from left to right in Table \[tab5\]) a power law with a constant-density core region; the cusped density profile suggested by Navarro, Frenk, & White (1996, 1997) as appropriate for cold dark matter halos; and the law popularized by Hernquist (1990) and applied widely to the light profiles of elliptical galaxies. In all cases, fits were performed by minimizing the residuals of the functions’ [projected]{} density profiles relative to the $N_{\rm cl}$ data, which have smaller observational uncertainties than the volume densities do. The best-fit parameters thus obtained were then used to overlay the original, three-dimensional functions on the deprojected $n_{\rm cl}$ and verify that they provided good descriptions of those as well. One consequence of this procedure is that no fit is given for the Hernquist (1990) profile, $n_{\rm cl}\propto (r/a)^{-1}(1+r/a)^{-3}$, to the M87 GCS. Although it is possible to fit this function directly to the $n_{\rm cl}$ data, its projection affords a poor description of the full $N_{\rm cl}$ distribution. This almost certainly reflects the fact that the GCS profile in M87 appears about to become significantly shallower, at least temporarily, beyond the last radius ($r_{\rm gc}\simeq100$ kpc) for which cluster counts exist (because Fig. \[fig4\] suggests essentially that $\rho_{\rm cl}\propto\rho_{\rm gas}$ there). Such a change in slope is obviously not allowed by Hernquist’s density profile. Thus, the failure of this model in this case makes the point that [any]{} of the density fits in Table \[tab5\] should be taken at face value only over the ranges of $r_{\rm gc}$ that have been directly observed; extrapolations to much larger radii could potentially introduce appreciable systematic errors. Table \[tab5\] has three interesting implications: (1) None of the available data [require]{} the existence of truly constant-density “cores” in GCSs. This was also mentioned by McLaughlin (1995) and it is, of course, plainly seen in Figs. \[fig4\] through \[fig6\]. Evidence is emerging for a similar situation in elliptical galaxies generally (Crane et al. 1993; Gebhardt et al. 1996), and apparently also in galaxy clusters (Merritt & Tremblay 1994; Carlberg et al. 1997; McLaughlin 1999). (2) It has been suggested, particularly in the contexts of some dynamical analyses (e.g., Huchra & Brodie 1987; Weil, Bland-Hawthorn, & Malin 1997) that the M87 GCS directly reflects the distribution of dark matter in and around that galaxy. However, this is not true in any simple sense; neither of the $n_{\rm cl}$ fits given in Table \[tab5\] provides an adequate description of $\rho_{\rm dark}(r_{\rm gc})$ in M87. Instead, it is found there that $\rho_{\rm dark}\propto\rho_{\rm gas}$ (see McLaughlin 1999), in which case Fig. \[fig4\] shows that the GCS becomes a reliable tracer of the dark matter halo [alone]{} only on fairly large spatial scales (where the number densities become quite low). (3) The analytical fits to $N_{\rm cl}$ and $n_{\rm cl}$ offer a convenient way to extrapolate the observed densities to (slightly) smaller and larger radii than those observed, so as to compute the total GCS populations of these three galaxies. Thus, in M87, it is found that ${\cal N}_{\rm tot}= 13\,600\pm500$ clusters (over all magnitudes) are [projected]{} to distances $R_{\rm gc}\leq25\arcmin\simeq110$ kpc from the galaxy’s center; in M49, there are ${\cal N}_{\rm tot}=6\,850\pm550$ globulars within the same projected $R_{\rm gc}$; and in NGC 1399, ${\cal N}_{\rm tot}=4\,700\pm400$ clusters have $R_{\rm gc}\leq14\farcm5\simeq70$ kpc. These total “projected populations” are in good agreement with those estimated in many other studies (see the references cited in §3). However, it is now possible to determine the three-dimensional radius of the sphere which actually contains all of these clusters in each system. From either of the M87 fits in Table \[tab5\], it is found that 13600 clusters are contained within $r_{\rm gc}\simeq150$–160 kpc; in M49, 6850 globulars come from the sphere $r_{\rm gc}\leq140$–150 kpc; and the 4700 globulars in NGC 1399 all have true galactocentric positions $r_{\rm gc}\leq90$–95 kpc. To put this another way, all three GCSs are sufficiently centrally concentrated that, in each case, some 85%–90% of the globular clusters projected onto the galaxy—so again, to within $R_{\rm gc}\leq110$ kpc for M87 and M49, and $R_{\rm gc}\leq70$ kpc for NGC 1399—really are located at those galactocentric distances in three dimensions. This last point specifically can clarify and constrain the suggestion (White 1987; West et al. 1995; Côté et al. 1998) that the high specific frequencies in BCGs like M87 are due to the “contamination” of their GCSs, in projection, by a significant number of globular clusters that are associated with a galaxy cluster as a whole, rather than bound to the central galaxy itself. Harris et al. (1998) tried to test this idea in some detail for M87 in particular, working from the assumption (following White 1987) that any population of intergalactic globular clusters (IGCs) should generally be associated with the diffuse stellar light of the cD envelopes that surround many BCGs (i.e., these structures were considered to comprise primarily intracluster stars). This hypothesis leads to inconsistencies that prompted Harris et al. to reject IGCs as the main cause of the first specific frequency problem. While their arguments still stand within the framework that they constructed, it is now seen that the high $S_N$ of M87 certainly, and likely those of other BCGs as well, can be traced to subsets of their GCSs that are associated with hot gas, and not necessarily with stars at all. This does appear to favor the basic notion of IGCs, since the gas around M87 extends to very large spatial scales, and traces out the potential well of the Virgo Cluster as a whole (McLaughlin 1999). If a constant ${\hbox{$\widehat\epsilon$}}_{\rm cl}$ holds all the way out, then, a cluster-wide population of globulars could be expected. However, since [all]{} of the 13600 globular clusters which give M87 its high specific frequency in projection are physically associated with a volume of radius $\sim$150 kpc around the galaxy, it is not obvious that any of them are truly “intergalactic” objects at the current epoch. Projection effects are [not]{} solely responsible for the high $S_N$ of M87, and although intergalactic globular clusters may well exist, it is not clear that any have yet been observed. Of course, since M87 is at the dynamical center of Virgo, it can be difficult to distinguish meaningfully between the galaxy and the cluster on 100-kpc scales. In the context of IGCs, the question evidently comes down to one of origin: did the globulars now seen right around M87 actually form there, or were many of them somehow brought in from much further than 150 kpc away? Any self-consistent answer to this must be able to account for the observed fact that $\rho_{\rm cl}=0.003\,(\rho_{\rm gas}+\rho_{\rm stars})$ at galactocentric radii $r_{\rm gc}\sim10$–100 kpc in M87; that is, both the very existence and the normalization of this proportionality have to be explained. Although essentially any scenario which is able to do this would be allowed by the current data, none could be [preferred]{} over what remains the simplest option: that the bulk of the M87 GCS formed in situ, with a thoroughly standard and spatially constant $\epsilon_{\rm cl}$. Ultimately, only detailed evolutionary models, tailored specifically to the properties of M87 and Virgo and considering both stars [and gas]{}, can really decide what fraction of the galaxy’s GCS might have come originally from the cluster at large. Summary ======= Stars form mainly in groups, some of which emerge from their natal clouds of gas as gravitationally bound clusters. An understanding of cluster formation will therefore be integral to any full theory of star formation, and an important element of this is simply the frequency with which it occurs. That is, what is the likelihood that a star-forming cloud of gas—whether this be a dense clump in a Galactic giant molecular cloud, or a larger one in a protogalactic fragment—will achieve the high cumulative star formation efficiency (viz. ${\rm SFE}\ga20\%$–50%) that allows its stars to remain bound as a cluster after they have cleared away the gas? This paper has used observations of the globular cluster systems in galaxy halos to address this question, and to empirically evaluate the efficiency of cluster formation: by mass, $\epsilon_{\rm cl}\simeq0.25$%. This is, by all appearances, nearly a universal number, and it should therefore serve as a strong constraint on theories of star and cluster formation in any context. To arrive at this result, it was first shown (§2.1) that the total (global) [masses]{} of GCSs are quite robust, even over a Hubble time, against the dynamical processes that work to destroy low-mass globular clusters in the central regions of galaxies. With an observed $M_{\rm gcs}$ thus useful as a measure of the initial quantity, the specific frequencies of GCSs, $S_N \propto {\cal N}_{\rm tot}/L_{V,{\rm gal}}\propto M_{\rm gcs}/M_{\rm stars}$, serve as [crude]{} first estimates of the basic formation efficiency $\epsilon_{\rm cl}\equiv M_{\rm gcs}^{\rm init}/M_{\rm gas}^{\rm init}$. To the [limited]{} extent that $S_N\approx5$ is an adequate description of early-type galaxies, the implication is that $\epsilon_{\rm cl}\sim2\times 10^{-3}$. The limiting factor in this simplest treatment of GCS data is the assumption that a galaxy’s current stellar mass always suffices as an estimate of its total initial gas mass; if taken too far, this quickly leads to some perplexing conclusions. As was reviewed in §2.2, systematic variations in $S_N$ exist along the entire sequence of early-type galaxies, from dwarf ellipticals (where ${\cal N}_{\rm tot}\propto L_{V,{\rm gal}}^{0.6}$ or so) through normal giants (${\cal N}_{\rm tot}\propto L_{V,{\rm gal}}^{1.3}$) and on to the brightest ellipticals, including many of the central galaxies in groups and rich clusters (${\cal N}_{\rm tot}\propto L_{V,{\rm gal}}^{1.8}$). If specific frequency were indeed a good estimator of $\epsilon_{\rm cl}$ in general, the implication would seem to be that this also varied, for unknown reasons and in a non-monotonic fashion, as a function of galaxy luminosity. Closely related to this is the fact that the GCSs of some, but not all, large ellipticals are spatially more extended than their stellar halos. That is, local $S_N$ values increase with projected $R_{\rm gc}$ in some systems, which naively would suggest that—again, for unknown reasons, and counter to any intuition—the efficiency of cluster formation [sometimes]{} increased with galactocentric radius. An important clue to the true nature of both these “specific frequency problems”—which are most likely two aspects of a single phenomenon—lies in the observation that the global $S_N$ in central cluster galaxies increases with the X-ray luminosity of their parent clusters. Some of the globulars in these systems ought then to be associated with the hot gas there; and $M_{\rm gas}^{\rm init}$ is approximated better by the present-day $(M_{\rm gas}+M_{\rm stars})$ than by $M_{\rm stars}$ alone. $S_N$ then systematically overestimates the true cluster formation efficiency $\epsilon_{\rm cl}$, which instead is given empirically by ${\hbox{$\widehat\epsilon$}}_{\rm cl}=M_{\rm gcs}/(M_{\rm gas}+M_{\rm stars})$. In addition, since a galaxy’s hot gas generally follows a shallower density profile than its stellar halo does, the “extra” clusters in very gas-rich (and high-$S_N$) ellipticals might well be expected to also show a radial distribution, $\rho_{\rm cl}$ or $\Sigma_{\rm cl}$, that is distended relative to $\rho_{\rm stars}$ or $\Sigma_{\rm stars}$. These ideas, which extend similar suggestions in the recent literature, have been tested in detail for each of M87, M49, and NGC 1399. Data were collected from the literature on the GCSs, surface brightness profiles, and X-ray gas densities of these galaxies, covering spatial scales $r_{\rm gc}\sim1$–100 kpc. In §3.1, new comparisons of the projected densities $\Sigma_{\rm cl}$ and $\Sigma_{\rm stars}$ in M49 and NGC 1399 (which are gas-poor) showed that, contrary to previous claims, neither galaxy suffers from either of the two specific frequency problems: The global $S_N$ of M49 has a “standard” value of $\simeq$5, and the local ratio $\Sigma_{\rm cl}/\Sigma_{\rm stars}$ is constant with radius (outside of a “core” region, $R_{\rm gc}\la R_{\rm eff}$, that possibly has been dynamically depleted) if the recent surface-brightness profile of Caon et al. (1994) is used in the comparison rather than the older $\mu_B$ measurements of King (1978). The global specific frequency of NGC 1399 is similarly “average”, $S_N=7.0\pm0.6$, if its GCS population is normalized to an absolute magnitude for the galaxy ($M_V\simeq-22.1$) that is significantly brighter than the value more commonly adopted. The ratio $\Sigma_{\rm cl}/ \Sigma_{\rm stars}$ is also constant with radius in this galaxy, and has a value equal to that found in M49. M87, at the center of the Virgo Cluster, is much richer in gas than either of M49 or NGC 1399, and both $S_N$ problems are in evidence there. A discrete, geometrical deprojection algorithm was developed to enable a direct comparison between the volume densities $\rho_{\rm cl}$, $\rho_{\rm stars}$, and $\rho_{\rm gas}$ at all three-dimensional galactocentric radii in M87. It was found in §3.2 that $\rho_{\rm cl}\propto(\rho_{\rm gas}+\rho_{\rm stars})$, beyond $r_{\rm gc}\sim8$ kpc, and that the constant of proportionality is the same as that in M49 and NGC 1399. [The cluster formation efficiency was constant with galactocentric radius inside each of these systems, and the same from one galaxy to another]{}. Specifically, $$\epsilon_{\rm cl}\simeq {{\rho_{\rm cl}}\over{\rho_{\rm gas}+\rho_{\rm stars}}}= {{M_{\rm gcs}}\over{M_{\rm gas}+M_{\rm stars}}}= 0.0026\pm0.0005\ .$$ (Observed departures from this ratio in the central regions $r_{\rm gc}\la R_{\rm eff}$ of M87 and M49 may have resulted, at least in part, from the dynamical destruction of globular clusters there.) Many globulars in M87 are indeed associated with its X-ray gas, and this is in fact responsible for the appearance of both specific frequency problems. This single case leads to the basic expectation—which should be tested in as many individual systems as possible—that real discrepancies between $\Sigma_{\rm cl}$ and $\Sigma_{\rm stars}$ should appear only in gas-rich ellipticals that also have high global specific fequencies. The obvious generalization of these results is that most galaxies might have been subject to a single, common $\epsilon_{\rm cl}$. This possibility was explored, and borne out, in §4. Observations of global $S_N$ (i.e., total GCS populations and galaxy luminosities) in 97 giant ellipticals, brightest cluster galaxies, and dwarf ellipticals [all]{} were shown to be consistent with the predictions of a universal $\epsilon_{\rm cl}$ at the same level as that observed directly in M87, M49, and NGC 1399. The different scalings of $S_N$ with $L_{V,{\rm gal}}$ among normal giant ellipticals, central cluster galaxies, and dwarfs stem entirely from fundamentally different relations between $L_{V,{\rm gal}}$ and $M_{\rm gas}^{\rm init}$ in these different physical regimes. These relations have been obtained from scaling arguments that employ the optical fundamental plane of bright ellipticals; the observed correlation between X-ray and optical luminosities in early-type systems; and the fractional gas loss expected in a specific model (Dekel & Silk 1986) for winds from a single burst of star formation in dark-matter dominated dwarf ellipticals. Observations of ${\cal N}_{\rm tot}$ vs. $L_{V, {\rm gal}}$ still show some [scatter]{} about the mean prediction of constant $\epsilon_{\rm cl}$, and it will be important to clarify the significance and the origins of this. However, there are no [systematic]{} deviations from a universal cluster formation efficiency, in galaxies that span four orders of magnitude in total luminosity and GCS population. Exactly the same $\epsilon_{\rm cl}$ was shown, in §4, to have obtained for globular clusters in the spheroid, or stellar halo, of the Milky Way; and it was argued that it appears also to apply now, to the formation of [open]{} clusters locally in the Galactic disk. This remarkable robustness, in which a single cluster formation efficiency was and is realized in a great variety of local and global environments, demands a very general theoretical explanation. From a more empirical standpoint, it bears directly on current questions involving GCS and galaxy formation and evolution on large scales. Some of these were discussed in §5. Of considerable interest is the fact that, with $\epsilon_{\rm cl}$ constant but $M_{\rm gcs}/M_{\rm stars}$ not, the efficiency of [unclustered]{} star formation could [not*]{} have been universal; field star formation must have been suppressed, in a sense, in dwarf ellipticals and at large galactocentric radii in the brightest, most gas-rich ellipticals. The possibility that these points could be understood in terms of feedback processes seems promising (§5), and should be explored in quantitative models. I am grateful to Bill Harris for providing the data plotted in Figures \[fig8\] and \[fig9\]; to Ivan King and Chris McKee for helpful comments on an earlier draft of the paper; and to an anonymous referee for several valuable suggestions, including the addition of Figures \[fig7\] and \[fig10\]. This work was supported by NASA through grant number HF-1097.01-97A awarded by the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., for NASA under contract NAS5-26555. Deprojection Algorithm ====================== The surface density data in Tables \[tab2\], \[tab3\], and \[tab4\] lend themselves well to a geometrical deprojection of the sort that is often applied to the X-ray emission from ellipticals and galaxy clusters (e.g., Fabian et al. 1981; Kriss, Cioffi, & Canizares 1983; Nulsen & Böhringer 1995). This technique makes only one assumption—that of circular symmetry—and is fully nonparametric. It also avoids the explicit differentiation of $N_{\rm cl}$ that is required in the standard Abel-transform approach to deprojection. Such a differentiation can be performed if an ad hoc parametrization is adopted for the global form of the density profile from the outset, but it is clearly desireable to minimize any such a priori restrictions. Although sophisticated, nonparametric algorithms that resort finally to an Abel integral have also been developed (e.g., Merritt & Tremblay 1994), they present a greater degree of complexity than is necessary here; and, in any case, they seem poorly suited to density profiles that are constructed from multiple observations with different fields of view and statistical corrections for different levels of background contamination and photometric incompleteness. The assumption of spherical symmetry made here is not perfect (the isopleths of the M87 GCS, for one, are clearly non-circular in projection; see McLaughlin et al. 1994), but neither is it nonsensical: it takes azimuthally averaged surface densities and returns volume densities that are properly interpreted as angular averages at various galactocentric radii. A simple example of the basic method, which should serve to define its geometrical set-up and mathematical notation, is illustrated in Fig. \[fig11\]. The plane of the sky there is perpendicular to the plane of the page, and the two intersect in the line $z=0$. Thus, the lines on the right, labelled by projected radius $R_0$, $R_1$, and $R_2$, are lines of sight for an observer situated off the bottom of the page. The bands $R_0\le R\le R_1$ and $R_1\le R\le R_2$, along with their mirror images on the left side of the Figure, are axial cross sections of cylinders that extend to $z=\pm\infty$; in the plane of the sky, of course, the corresponding cross sections are circular annuli. The circular annuli shown here are not those on the plane of the sky, however; rather, they are equatorial cross sections of spherical shells with three-dimensional radii $r_0\le r\le r_1$ and $r_1\le r\le r_2$. Evidently, these shells have been chosen such that $r_0=R_0$, and so on. The origin $r=0$ (through which passes the axis $R=0$) represents the center of a spherical distribution of globular clusters, which are supposed to have been counted up in discrete cylinders $R_0\le R\le R_1$ and $R_1\le R\le R_2$ on the sky, to form the surface densities $N_{\rm cl}(R_0,R_1)$ and $N_{\rm cl}(R_1,R_2)$. From such data, it is possible to compute the fraction of clusters in each cylinder that actually reside in the spherical shells $r_0\le r\le r_1$ and $r_1\le r\le r_2$, and thus to form the average volume densities $n_{\rm cl}(r_0,r_1)$ and $n_{\rm cl}(r_1,r_2)$. =4.0truein Suppose, for the moment (this will be relaxed just below), that $r_2=R_2$ marks the outer edge of the GCS in question, i.e., $n_{\rm cl}(r>r_2)\equiv0$ and $N_{\rm cl}(R>R_2)\equiv0$. (It is assumed that the density $N_b$ of any uniform distribution of background objects has already been subtracted to form $N_{\rm cl}$ from the raw data.) In this case, all of the objects seen in the outermost cylinder actually come from the outer spherical shell, i.e., from the hatched regions between $R_1$ and $R_2$ in Fig. \[fig11\]. Since the number of globulars in this cylinder is just $${\cal N}_{\rm cl}(R_1,R_2)=N_{\rm cl}(R_1,R_2)\times\pi(R_2^2-R_1^2)\ ,$$ all that is required is a calculation of the volume of the shell $r_1\le r\le r_2$ which is intersected by the cylinder $R_1\le R\le R_2$. Denoting this by $V_{\rm int}(r_1,r_2;R_1,R_2)$, the volume density is obviously $$n_{\rm cl}(r_1,r_2)={{{\cal N}_{\rm cl}(R_1,R_2)}\over{V_{\rm int}(r_1,r_2; R_1,R_2)}}\ .$$ Moving inwards to the cylinder $R_0\le R\le R_1$, the total number of clusters observed there includes a contribution from the spherical shell $r_1 \le r\le r_2$ (hatched regions between $R_0$ and $R_1$), as well as one from the shell of interest, $r_0\le r\le r_1$ (solid region). The volume density in the outer shell has already been determined, and the volume of intersection between that shell and the inner cylinder can be calculated \[call this $V_{\rm int} (r_1,r_2;R_0,R_1)$\], so the volume density in the inner shell is given by $$n_{\rm cl}(r_0,r_1)={{{\cal N}_{\rm cl}(R_0,R_1)\ -\ n_{\rm cl}(r_1,r_2)\,V_{\rm int}(r_1,r_2;R_0,R_1)}\over {V_{\rm int}(r_0,r_1;R_0,R_1)}}\ ,$$ where ${\cal N}_{\rm cl}(R_0,R_1)$ and $V_{\rm int}(r_0,r_1;R_0,R_1)$ have the same meanings as the corresponding quantities in the first cylinder, $R_1\le R\le R_2$. Clearly, this procedure can be extended and applied iteratively, from the outside in, to the projected number counts in any series of concentric, non-overlapping, circular annuli; it is only required that there be no gaps between any two consecutive annuli. Thus, let the annuli be defined on the plane of the sky by projected radii $R_0<R_1<R_2<\ldots<R_m$, i.e., set up $m$ rings with $R_0\le R\le R_1$; $R_1\le R\le R_2$; and so on. Given average surface densities $N_{\rm cl}$ in each of these rings, the volume densities $n_{\rm cl}$ in the corresponding spherical shells $r_0<r_1<r_2<\ldots<r_m$ can be obtained once the volume of intersection between some generic shell $r_{j-1}\le r\le r_j$ and a cylinder $R_{i-1}\le R\le R_i$ is known. A sketch similar to Fig. \[fig11\] readily shows that this volume is, quite generally, $$V_{\rm int}(r_{j-1},r_j;R_{i-1},R_i)={{4\pi}\over{3}}\, \left[(r_j^2-R_{i-1}^2)^{3/2}-(r_j^2-R_i^2)^{3/2} +(r_{j-1}^2-R_i^2)^{3/2}-(r_{j-1}^2-R_{i-1}^2)^{3/2}\right], \label{eq:33}$$ for $i,j=1,\ldots,m$. This holds for any set of radii $\{r_{j-1},r_j,R_{i-1}, R_i\}$, with the understanding that any term in parentheses which evaluates as negative is to be set to 0. An important application is to the case shown in Fig. \[fig11\], where, e.g., $r_1=R_1$ and $r_2=R_2$ and the volume common to the shell $R_1\le r\le R_2$ and the cylinder $R_1\le R\le R_2$ is $V_{\rm int} (R_1,R_2;R_1,R_2)=(4\pi/3)(R_2^2-R_1^2)^{3/2}$. It is most appropriate to define shells such that $r_i=R_i$ for all $i$, in which case the results given just above for $m=2$ generalize to $$\begin{aligned} n_{\rm cl}^{\prime}(R_{i-1},R_i) & = & {{{\cal N}_{\rm cl}(R_{i-1},R_i)-\sum_{j=i+1}^{m} \left[n_{\rm cl}(R_{j-1},R_j) \,V_{\rm int}(R_{j-1},R_j;R_{i-1},R_i)\right]} \over{V_{\rm int}(R_{i-1},R_i;R_{i-1},R_i)}}\nonumber \\ & = & {3\over{4(R_i^2-R_{i-1}^2)^{1/2}}}\left[ N_{\rm cl}(R_{i-1},R_i)-\sum_{j=i+1}^{m} {{n_{\rm cl}(R_{j-1},R_j)\,V_{\rm int} (R_{j-1},R_j;R_{i-1},R_i)}\over{\pi(R_i^2-R_{i-1}^2)}}\right]. \nonumber \\ & & \label{eq:34}\end{aligned}$$ Again, this assumes that any uniform background has already been subtracted to obtain the surface densities $N_{\rm cl}$ in every annulus $i=1,\ldots,m$. This equation is not quite final, however, since its derivation has been facilitated by treating the outermost annulus seen in projection, $R_{m-1}\le R\le R_m$ (or the outermost spherical shell, $R_{m-1}\le r\le R_m$), as if it ended just at the “true edge” of some GCS. Of course, this will never actually be the case—if not because it makes little sense to associate sharp boundaries with such distributions in the first place, then because real observations and number counts can cover small fields of view within GCSs that are known to extend considerably further. This means that there is a second type of “background” contamination to be dealt with in any dataset, quite apart from that (already accounted for) due to a uniform distribution of unassociated galaxies and stars; corrections have to be made, in every annulus or shell, for the possible presence of bona fide globular clusters at radii $r>R_m$, beyond the last annulus for which a surface density can be directly measured. If the density distribution of such objects is denoted by $n_{\rm cl}(r)$, their contribution to the projected number density in a cylinder $R_{i-1}\le R\le R_i$ is given by $$N_{\rm cl}^{\rm back}(R_{\rm i-1},R_i) = {4\over{R_i^2-R_{i-1}^2}}\, \int_{R_{i-1}}^{R_i} R\,dR\,\int_{z_0(R)}^{\infty} n_{cl}(r)\,dz \ ;\ \ \ \ \ \ z_0(R)\equiv(R_m^2-R^2)^{1/2}\ , \label{eq:35}$$ which is just an average of the projected densities along every line of sight contained in the cylinder. \[As usual, the coordinate $z=(r^2-R^2)^ {1/2}$ here measures distance along these lines of sight; $z=0$ defines the plane of the sky, as in Fig. \[fig11\].\] It is most convenient to express this “background” surface density as a fraction of the observed $N_{\rm cl}$ in the outermost annulus $R_{m-1}\le R\le R_m$: $$f(R_{i-1},R_i)\equiv{{N_{\rm cl}^{\rm back}(R_{i-1},R_i)}\over{N_{\rm cl} (R_{m-1},R_m)}}={{R_m^2-R_{m-1}^2}\over{R_i^2-R_{i-1}^2}}\,\left[ {{\int_{R_{i-1}}^{R_i} R\,dR\,\int_{z_0(R)}^{\infty} n_{\rm cl}(r)\, dz}\over{\int_{R_{m-1}}^{R_m} R\,dR\,\int_{0}^{\infty} n_{\rm cl}(r)\,dz}} \right], \label{eq:36}$$ where $z_0(R)$ is defined as in equation (\[eq:35\]), and the second equality relies on the basic definition of average surface density in a circular annulus of finite width. The advantage of this formulation is that the form of $n_{\rm cl}(r)$ cannot be directly measured for $r>R_m$, so the calculation of $N_{\rm cl}^{\rm back}$ necessarily depends on an extrapolation of some sort. Without any constraints on this extrapolation, it is conceivable that an artificial and unphysical situation such as $N_{\rm cl}^{\rm back}(R_{m-1},R_m) >N_{\rm cl}(R_{m-1},R_m)$ could arise. This and similar difficulties are avoided, however, by only ever evaluating $N_{\rm cl}^{\rm back}$ in terms of $f$ (which is guaranteed always to be $<$1 when $i=m$) and the known $N_{\rm cl}$ at the largest projected radii observed. Equation (\[eq:36\]) should therefore be used to subtract $N_{\rm cl}^{\rm back}$ from the total $N_{\rm cl}$ in every annulus before deprojecting the distribution according to equation (\[eq:34\]). The corrected volume densities in the spherical shells $i=1,\ldots,m$ may then be written as $$n_{\rm cl}(R_{i-1},R_i)=n_{\rm cl}^{\prime}(R_{i-1},R_i)\ -\ {3\over{4(R_i^2-R_{i-1}^2)^{1/2}}}\,f(R_{i-1},R_i)N_{\rm cl}(R_{m-1},R_m)\ . \label{eq:37}$$ Typically, when measured GCS surface densities extend to sufficiently large radius, an adequate estimate of $f$ can be made by assuming a power-law distribution $n_{\rm cl}\propto r^{-\alpha}$ in the outer reaches of the system, $r>R_m$; and in the case that $\alpha$ is an integer, equation (\[eq:36\]) can be evaluated analytically. With $\alpha=2$, for example, $$f(R_{i-1},R_i)={{R_{m-1}+R_m}\over{R_{i-1}+R_i}} \left[1-{2\over{\pi}}\, {{\left(\cos^{-1}{{R_i}\over{R_m}}-{{R_{i-1}}\over{R_i}}\cos^{-1}{{R_{i-1}} \over{R_m}}\right)-\left(\sqrt{{{R_m^2}\over{R_i^2}}-1}-\sqrt{{{R_m^2}\over {R_i^2}}-{{R_{i-1}^2}\over{R_i^2}}}\right)}\over{1-R_{i-1}/R_i}} \right] . \label{eq:a1}$$ For $\alpha=3$, $$f(R_{i-1},R_i)={{R_m^2-R_{m-1}^2}\over{R_i^2-R_{i-1}^2}}\, \left[{{\sqrt{1-{{R_{i-1}^2}\over{R_m^2}}}-\sqrt{1-{{R_i^2}\over{R_m^2}}} -\ln\left({{1+\sqrt{1-R_{i-1}^2/R_m^2}}\over{1+\sqrt{1-R_i^2/R_m^2}}}\right)} \over{\ln(R_m/R_{m-1})}} \right] . \label{eq:a2}$$ And for $\alpha=4$, $$f(R_{i-1},R_i)= \left[{{(R_{m-1}+R_m)R_mR_{m-1}}\over{(R_{i-1}+R_i)R_iR_{i-1}}}\right]\ \ \ \ \times \qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad$$ $$\qquad\qquad\qquad\qquad \left[1-{2\over{\pi}}\, {{\left({{R_i}\over{R_{i-1}}}\cos^{-1}{{R_{i-1}}\over{R_m}}-\cos^{-1}{{R_i} \over{R_m}}\right)-{{R_i}\over{R_m}}\left(\sqrt{1-{{R_{i-1}^2}\over{R_m^2}}}- \sqrt{1-{{R_i^2}\over{R_m^2}}}\right)}\over{R_i/R_{i-1}-1}} \right] . \label{eq:a3}$$ These three power-law exponents are expected to bracket the conditions that hold in real GCSs. Finally, the volume densities derived in this way are a series of averages of a continuous $n_{\rm cl}(r)$ in spherical shells of finite thickness. The question therefore arises as to how a single radius $\overline{r}_i$ should be assigned to each shell so that the discrete distribution, $n_{\rm cl}(R_{i-1}, R_i)$ vs. $\overline{r}_i$, represents the continuous one as faithfully as possible. (See Harris 1986 or King 1988 for a discussion of this issue in the two-dimensional context.) Clearly, the answer is to choose each $\overline{r}_i\in [R_{i-1},R_i]$ such that $$n_{\rm cl}(R_{i-1},R_i)={3\over{R_i^3-R_{i-1}^3}} \int_{R_{i-1}}^{R_i} n_{\rm cl}(r)\,r^2\,dr\,=\, n_{\rm cl}(\overline{r}_i)\ . \label{eq:38}$$ If it is assumed that $n_{\rm cl}\propto r^{-\alpha}$, the second equality is easily solved for $\overline{r}_i$ in every shell. However, $\alpha$ will generally not be a constant from shell to shell; that is, a real GCS need not obey a single power-law density distribution over its entire extent (as is evident, for example, in Fig. \[fig3\] above). Moreover, $\alpha$ cannot be known a priori, and any attempts to derive it from the data must rely on fits to the average deprojected densities—fits which themselves depend on the values of $\overline{r}_i$. Fortunately, a complicated iterative procedure can be avoided by noting that observed radial distributions suggest that $0\la \alpha\la3$ anywhere in a typical GCS. This in turn suggests, and numerical experiments confirm, that simply using $\alpha=3/2$ everywhere should lead to $\overline{r}_i$ estimates that are acceptable substitutes for the exact solutions of equation (\[eq:38\]) in any realistic situation. Thus, the formula adopted here is $$\overline{r}_i=\left[(R_{i-1}^{3/2}+R_i^{3/2})/2\right]^{2/3}\ . \label{eq:39}$$ Aguilar, L., Hut, P., & Ostriker, J. P. 1988, ApJ, 335, 720 Ashman, K. M., & Zepf, S. E. 1992, ApJ, 384, 50 Bahcall, J. N., & Soneira, R. M. 1980, ApJS, 44, 73 Bender, R., Surma, P., Döbereiner, S., Möllenhoff, C., & Madejsky, R. 1989, A&A, 217, 35 Binney, J., & Tremaine, S. 1987, Galactic Dynamics (Princeton: Princeton Univ. Press) Blakeslee, J. P. 1997, ApJ, 481, L59 Blakeslee, J. P., Tonry, J. L., & Metzger, M. R. 1997, AJ, 114, 482 Bridges, T. J., Hanes, D. A., & Harris, W. E. 1991, AJ, 101, 469 Brighenti, F., & Mathews, W. G. 1998, ApJ, 495, 239 Brodie, J. P., Schroder, L. L., Huchra, J. P., Phillips, A. C., Kissler-Patig, M., & Forbes, D. A. 1998, AJ, 116, 691 Brown, B. A., & Bregman, J. N. 1998, ApJ, 495, L75 Burstein, D., & Heiles, C. 1984, ApJS, 54, 33 Caon, N., Capaccioli, M., & D’Onofrio, M. 1994, A&AS, 106, 199 Capuzzo-Dolcetta, R. 1993, ApJ, 415, 616 Carlberg, R. G., et al. 1997, ApJ, 485, L13 Cohen, J. G. 1988, AJ, 95, 682 Côté, P., Marzke, R. O., & West, M. J. 1998, ApJ, 501, 554 Crane, P., et al. 1993, AJ, 106, 1371 Davies, R. L., Efstathiou, G., Fall, S. M., Illingworth, G., & Schecter, P. L. 1983, ApJ, 266, 41 Davis, D. S., & White, R. E. 1996, ApJ, 470, L35 de Carvalho, R. R., & Djorgovski, S. 1989, ApJ, 341, L37 de Carvalho, R. R., & Djorgovski, S. 1992, ApJ, 389, L49 de Vaucouleurs, G. 1993, ApJ, 415, 40 de Vaucouleurs, G., & Nieto, J.-L. 1978, ApJ, 220, 449 de Vaucouleurs, G., & Nieto, J.-L. 1979, ApJ, 230, 697 de Vaucouleurs, G., & Pence, W. D. 1978, AJ, 83, 1163 de Vaucouleurs, G., de Vaucouleurs, A., Corwin, Jr., H. G., Buta, R. J., Paturel, G., & Fouqué, P. 1991, Third Reference Catalogue of Bright Galaxies (New York: Springer) Dehnen, W. 1993, MNRAS, 265, 250 Dekel, A.., & Silk, J. 1986, ApJ, 303, 39 Djorgovski, S., & Meylan, G. 1994, AJ, 108, 1292 Durrell, P. R., Harris, W. E., Geisler, D., & Pudritz, R. E. 1996, AJ, 112, 972 Elmegreen, B. G. 1983, MNRAS, 203, 1011 Elmegreen, B. G., & Clemens, C. 1985, ApJ, 294, 523 Elmegreen, B. G., & Efremov, Y. N. 1997, ApJ, 480, 235 Fabbiano, G. 1989, ARA&A, 27, 87 Faber, S. M., & Gallagher, J. 1979, ARA&A, 17, 135 Faber, S. M., Dressler, A., Davies, R. L., Burstein, D., Lynden-Bell, D., Terlevich, R., & Wegner, G. 1987, in Nearly Normal Galaxies, ed. S. M. Faber (New York: Springer), 175 Faber, S. M., et al. 1997, AJ, 114, 1771 Fabian, A. C., Hu, E. M., Cowie, L. L., & Grindlay, J. 1981, ApJ, 248, 47 Forbes, D. A., Brodie, J. P., & Grillmair, C. J. 1997, AJ, 113, 1652 Forbes, D. A., Franx, M., Illingworth, G. D., & Carollo, C. M. 1996, ApJ, 467, 126 Forman, W., Jones, C., & Tucker, W. 1985, ApJ, 293, 102 Freedman, W., et al. 1994, Nature, 371, 757 Fritze-von Alvensleben, U. 1999, A&A, in press (astro-ph/9812285) Gebhardt, K., et al. 1996, AJ, 112, 105 Geisler, D., Lee, M. G., & Kim, E. 1996, AJ, 111, 1529 Gnedin, O. Y., & Ostriker, J. P. 1997, ApJ, 474, 223 Grillmair, C., Pritchet, C., & van den Bergh, S. 1986, AJ, 91, 1328 Hanes, D. A., & Harris, W. E. 1986, ApJ, 309, 564 Harris, W. E. 1976, AJ, 81, 1095 Harris, W. E. 1981, ApJ, 251, 497 Harris, W. E. 1986, AJ, 91, 822 Harris, W. E. 1991, ARA&A, 29, 543 Harris, W. E. 1996, AJ, 112, 1487 Harris, W. E., & Petrie, P. L. 1978, ApJ, 223, 88 Harris, W. E., & Pudritz, R. E. 1994, ApJ, 429, 177 Harris, W. E., & Racine, R. 1979, ARA&A, 17, 241 Harris, W. E., & Smith, M. G. 1976, ApJ, 207, 1036 Harris, W. E., & van den Bergh, S. 1981, AJ, 86, 1627 Harris, W. E., Harris, G. L. H., & McLaughlin, D. E. 1998, AJ, 115, 1801 Harris, W. E., Pritchet, C. J., & McClure, R. D. 1995, ApJ, 441, 120 Harris, W. E., Allwright, J. W. B., Pritchet, C. J., & van den Bergh, S. 1991, ApJS, 76, 115 Hernquist, L. 1990, ApJ, 356, 359 Hills, J. G. 1980, ApJ, 225, 986 Ho, L. C., & Filippenko, A. V. 1996a, ApJ, 466, L83 Ho, L. C., & Filippenko, A. V. 1996b, ApJ, 472, 600 Huchra, J. P., & Brodie, J. P. 1987, AJ, 93, 779 Irwin, J. A., & Sarazin, C. 1996, ApJ, 471, 683 Jacoby, G. H., Branch, D., Ciardullo, R., Davies, R. L., Harris, W. E., Pierce, M. J., Pritchet, C. J., Tonry, J. L., & Welch, D. L. 1992, PASP, 104, 599 Jones, C., Stern, C., Forman, W., Breen, J., David, L., Tucker, W., & Franx, M. 1997, ApJ, 482, 143 Kavelaars, J. J., & Hanes, D. A. 1997, MNRAS, 285, L31 Kavelaars, J. J. 1998, preprint (astro-ph/9806094) Killeen, N. E. B., & Bicknell, G. V. 1988, ApJ, 325, 165 King, I. R. 1978, ApJ, 222, 1 King, I. R. 1988, PASP, 100, 999 Kinman, T. D., Suntzeff, N. B., & Kraft, R. P. 1994, AJ, 108, 1722 Kissler-Patig, M. 1997, A&A, 319, 83 Kissler-Patig, M., Forbes, D. A., & Minniti, D. 1998, MNRAS, 298, 1123 Kissler-Patig, M., Kohle, S., Hilker, M., Richtler, T., Infante, L., & Quintana, H. 1997, A&A, 319, 470 Kohle, S., Kissler-Patig, M., Hilker, M., Richtler, T., Infante, L., & Quintana, H. 1996, A&A, 309, L39 Kriss, G. A., Cioffi, D. F., & Canizares, C. R. 1983, ApJ, 272, 439 Kumai, Y., Basu, B., & Fujimoto, M. 1993a, ApJ, 404, 144 Kumai, Y., Hashi, Y., & Fujimoto, M. 1993b, ApJ, 416, 576 Lada, C. J., Margulis, M., & Dearborn, D. 1984, ApJ, 285, 141 Lada, E. A., DePoy, D. L., Evans, N. J., & Gatley, I. 1991, ApJ, 371, 171 Larson, R. B. 1988, in IAU Symposium No. 126, Globular Cluster Systems in Galaxies, ed. J. E. Grindlay and A. G. D. Philip (Dordrecht: Kluwer), 311 Larson, R. B. 1993, in ASP Conf. Ser. 48, The Globular Cluster-Galaxy Connection, ed. G. H. Smith and J. P. Brodie (San Francisco: ASP), 675 Larson, R. B. 1996, in ASP Conf. Ser. 92, Formation of the Galactic Halo...Inside and Out, ed. H. Morrison and A. Sarajedini (San Francisco: ASP), 241 Lauer, T. R., & Kormendy, J. 1986, ApJ, 303, L1 Lee, M. G., Kim, E., & Geisler, D. 1998, AJ, 115, 947 Mandushev, G., Spassova, N., & Staneva, A. 1991, A&A, 252, 94 Mathieu, R. D. 1983, ApJ, 267, L97 Mateo, M. L. 1998, ARA&A, 36, 435 McKee, C. F. 1989, ApJ, 345, 782 McLaughlin, D. E. 1994, PASP, 106, 47 McLaughlin, D. E. 1995, AJ, 109, 2034 McLaughlin, D. E. 1999, ApJ Letters, 512, in press (astro-ph/9812242) McLaughlin, D. E., & Pudritz, R. E. 1996, ApJ, 457, 578 McLaughlin, D. E., Harris, W. E., & Hanes, D. A. 1993, ApJ, 409, L45 McLaughlin, D. E., Harris, W. E., & Hanes, D. A. 1994, ApJ, 422, 486 McLaughlin, D. E., Secker, J., Harris, W. E., & Geisler, D. 1995, AJ, 109, 1033 McMillan, R., Ciardullo, R., & Jacoby, G. 1993, ApJ, 416, 62 Merritt, D., & Tremblay, B. 1994, AJ, 108, 514 Meurer, G. R., Heckman, T. M., Leitherer, C., Kinney, A., Robert, C., & Garnett, D. R. 1995, AJ, 110, 2665 Miller, B. W., Lotz, J. M., Ferguson, H. C., Stiavelli, M., & Whitmore, B. C. 1998, ApJ, 508, L133 Minniti, D. 1995, AJ, 109, 1663 Minniti, D., Meylan, G., & Kissler-Patig, M. 1996, A&A, 312, 49 Murali, C., & Weinberg, M. D. 1997a, MNRAS, 288, 749 Murali, C., & Weinberg, M. D. 1997b, MNRAS, 288, 767 Murali, C., & Weinberg, M. D. 1997c, MNRAS, 291, 717 Muzzio, J. C. 1987, PASP, 99, 245 Navarro, J. F., Frenk, C. S., & White, S. D. M. 1996, ApJ, 462, 563 Navarro, J. F., Frenk, C. S., & White, S. D. M. 1997, ApJ, 490, 493 Nieto, J.-L., Bender, R., & Surma, P. 1991, A&A, 244, L37 Nulsen, P. E. J., & Böhringer, H. 1995, MNRAS, 274, 1093 Oemler, A. 1976, ApJ, 209, 693 Okazaki, T., & Tosa, M. 1995, MNRAS, 274, 48 Ostriker, J. P., Binney, J., & Saha, P. 1989, MNRAS, 241, 849 Ostrov, P. G., Forte, J. C., & Geisler, D. 1998, AJ, 116, 2854 Pahre, M. A., Djorgovski, S. G., & de Carvalho, R. R. 1995, ApJ, 453, L17 Patel, K., & Pudritz, R. E. 1994, ApJ, 424, 688 Pierce, M. J., Welch, D. L., McClure, R. D., van den Bergh, S., Racine, R., & Stetson, P. B. 1994, Nature, 371, 385 Portegies Zwart, S. F., Hut, P., Makino, J., & McMillan, S. L. W. 1998, A&A, 337, 363 Preston, G. W., Schectman, S. A., & Beers, T. C. 1991, ApJ, 375, 121 Pritchet, C. J., & van den Bergh, S. 1994, AJ, 107, 1730 Pryor, C., & Meylan, G. 1993, in ASP Conf. Ser. 50, Structure and Dynamics of Globular Clusters, ed. S. G. Djorgovski and G. Meylan (San Francisco: ASP), 357 Roberts, M. S. 1957, PASP, 69, 59 Santiago, B. X., & Djorgovski, S. 1993, MNRAS, 261, 753 Schweizer, F. 1987, in Nearly Normal Galaxies, ed. S. M. Faber (New York: Springer), 18 Searle, L., & Zinn, R. 1978, ApJ, 225, 357 Sternberg, A. 1998, ApJ, 506, 721 Surdin, V. G. 1979, Sov. Astron., 23, 648 Tonry, J. L., Blakeslee, J. P., Ajhar, E. A., & Dressler, A. 1997, ApJ, 475, 399 Tremaine, S., Richstone, D. O., Byun, Y.-I., Dressler, A., Faber, S. M., Grillmair, C., Kormendy, J., & Lauer, T. R. 1994, AJ, 107, 634 Tremblay, B., & Merritt, D. 1996, AJ, 111, 2243 Trinchieri, G., Fabbiano, G., & Canizares, C. R. 1986, ApJ. 310, 637 Tsai, J. C. 1993, ApJ, 413, L59 van den Bergh, S. 1984, PASP, 96, 329 van den Bergh, S. 1998, ApJ, 492, 41 van den Bergh, S., & Lafontaine, A. 1984, AJ, 89, 1822 van der Marel, R. P. 1991, MNRAS, 253, 710 Verschueren, W. 1990, A&A, 234, 156 Vesperini, E. 1997, MNRAS, 287, 915 Vesperini, E., & Heggie, D. C. 1997, MNRAS, 289, 898 Wagner, S., Richtler, T., & Hopp, U. 1991, A&A, 241, 399 Weil, M. L., Bland-Hawthorn, J., & Malin, D. F. 1997, ApJ, 490, 664 West, M. J. 1993, MNRAS, 265, 755 West, M. J., Côté, P., Jones, C., Forman, W., & Marzke, R. O. 1995, ApJ, 453, L77 White, R. E. 1987, MNRAS, 227, 185 Whitmore, B. C. 1997, in The Extragalactic Distance Scale, ed. M Livio, M. Donahue, and N. Panagia (Baltimore: STScI), 254 Whitmore, B. C., & Schweizer, F. 1995, AJ, 109, 960 Whitmore, B. C., Sparks, W. B., Lucas, R. A., Macchetto, F. D., & Biretta, J. A. 1995, ApJ, 454, L73 Young, P. J. 1976, AJ, 81, 807 Zepf, S. E., & Ashman, K. M. 1993, MNRAS, 264, 611 Zinn, R. 1985, ApJ, 293, 424 Zinnecker, H., McCaughrean, M. J., & Wilking, B. A. 1993, in Protostars and Planets III, ed. E. H. Levy and J. I. Lunine (Tucson: Univ. of Arizona Press), 429	{ "pile_set_name": "ArXiv" }
--- abstract: 'Both the weighted and unweighted Unifrac distances have been very successfully employed to assess if two communities differ, but do not give any information about how two communities differ. We take advantage of recent observations that the Unifrac metric is equivalent to the so-called earth mover’s distance (also known as the Kantorovich-Rubinstein metric) to develop an algorithm that not only computes the Unifrac distance in linear time and space, but also simultaneously finds which operational taxonomic units are responsible for the observed differences between samples. This allows the algorithm, called EMDUnifrac, to determine why given samples are different, not just if they are different, and with no added computational burden. EMDUnifrac can be utilized on any distribution on a tree, and so is particularly suitable to analyzing both operational taxonomic units derived from amplicon sequencing, as well as community profiles resulting from classifying whole genome shotgun metagenomes. The EMDUnifrac source code (written in python) is freely available at: <https://github.com/dkoslicki/EMDUnifrac>.' address: '$^{1}$Mathematics Department, Oregon State University, Corvallis OR' author: - 'Jason McClelland$^{1}{}^$, David Koslicki$^{1}$' bibliography: - 'bibliography.bib' title: 'EMDUnifrac: Exact linear time computation of the Unifrac metric and identification of differentially abundant organisms' --- [^1] Introduction ============ An important first step in comparative microbial ecology studies is the assessment of if and how two communities of microorganisms differ. Unifrac [@lozupone2007quantitative; @lozupone2005unifrac; @hamady2010fast], in its various implementations, is a commonly utilized distance metric that quantifies if two communities do indeed differ. In the field of metagenomics, this phylogentic-aware distance has been used to effectively cluster many 16S rRNA samples and distinguish between them based on a given environmental factor [@ley2006ecological; @frank2007molecular; @rawls2006reciprocal]. However, a recognized disadvantage to the Unifrac distance is that it only quantifies if* two communities differ and gives no indication of how they differ [@white2009statistical]. Typically, to answer the question of how two communities differ, further statistical or computational methods are employed [@white2009statistical; @schloss2006introducing; @wooley2010primer; @parks2010identifying]. In this article, we demonstrate that in viewing the Unifrac distance as the so-called Kantorovich-Rubinstein metric (also known as the earth mover’s distance [@rubner2000earth]), one can obtain exactly how two communities differ and which operational taxonomic units (OTUs) or taxa are responsible for the observed Unifrac distance. This equivalence between the Unifrac distance and the earth mover’s distance was demonstrated recently [@evans2012phylogenetic] and while this equivalence greatly improved the understanding of the Unifrac distance, the authors of [@evans2012phylogenetic] were primarily concerned with assessing statistical significance of Unifrac distances and not with detailing how this view can be used to returning differentially abundant OTUs. We begin first by detailing how using the earth mover’s distance to compute the Unifrac distance can identify differentially abundant OTUs. We then introduce a linear time algorithm, called EMDUnifrac, that computes the Unifrac distance and also returns the differentially abundant OTUs that contributed to this distance. Finally, after demonstrating its usefulness on previously published biological data, we prove the correctness of EMDUnifrac and calculate its time and space complexity. Identifying differentially abundant OTUs ======================================== To demonstrate how viewing the Unifrac distance as the earth mover’s distance (EMD) identifies differentially abundant OTUs, we first need to define the EMD. We focus here on the weighted (normalized) Unifrac distance, as the unweighted Unifrac can be obtained by appropriately modifying the underlying distributions utilized. Given two sample communities and the associated abundances of microorganisms therein, we can associate to these a phylogenetic tree $T$ and two probability distributions $P$ and $Q$ that represent the fraction of a given sample that appears at each node of the phylogenetic tree (not necessarily restricted to the leaves). As the phylogenetic tree $T$ has associated branch lengths, we can find the minimal distance between any two nodes of the tree. Let $D$ be the matrix of all pairwise distances between nodes in $T$. We use the notation $\Gamma(P,Q)$ to describe the space of all ways in which one community can be transformed into the other. The elements $M\in \Gamma(P,Q)$ are referred to as flows and are matrices indexed by the nodes in the tree $T$ with the stipulation that the row sums of $M$ are equal to $P$ and the column sums of $M$ are equal to $Q$. The $(i,j)^{\rm th}$ entry of such an $M$ indicates that a total abundance of $M_{i,j}$ has been moved from node $i$ in the sample $P$ to node $j$ in sample $Q$. With these conventions, we can define the earth mover’s distance on this tree, which we refer to herein as EMDUnifrac: $$\label{eq:EMD} {\rm EMDUnifrac}(P,Q) = \underset{M\in \Gamma(P,Q)}{\rm minimize} \sum_{i,j\in T} D_{i,j} M_{i,j}.$$ Informally, the quantity ${\rm EMDUnifrac}(P,Q)$ represents the minimum amount of “work” required to transform the distribution of one sample $P$ into the distribution of the other sample $Q$ along the phylogenetic tree. It has been previously shown, using different notation, that ${\rm EMDUnifrac}(P,Q)$ is equivalent to the weighted (normalized) Unifrac distance [@evans2012phylogenetic]. Equivalence can be shows for the unweighted Unifrac distance by modifying the distributions $P$ and $Q$ to be binary vectors on the same original support and redefining the space of all flows $\Gamma(P,Q)$. A toy example is given in Figure \[fig:Example\] that details the previously defined quantities. We concentrate on the flow $M^$ that minimizes the right hand side of the expression in and call this the minimizing flow. This matrix represents where the abundance of one sample was moved when it was being transformed into the other sample, and this quantity precisely describes how* the two samples differ and which OTUs contributed to the computed Unifrac value. For example, in Figure \[fig:Example\], the entry $M^_{2,1}=\frac{1}{3}$ indicates that $1/3^{\rm rd}$ of the abundance of the first sample was moved from node 2 to node 1. A little care must be taken, though, as it is not guaranteed that there is one unique* minimizing flow. In all cases, the elements on the diagonal of any minimizing flow can be ignored (as this only indicates the abundances that were the same between the two samples). However, we can define a vector indexed by the edges of our phylogenetic tree called the differential abundance vector, which is the same no matter which minimizing flow is chosen. Letting $E$ denote the edges of our phylogenetic tree, $T_e$ the nodes of the subtree below an edge $e \in E$ and $T_{e'}$ the remaining nodes of $T$, so that $T = T_e \cup T_{e'}$, we have that ${\rm DiffAbund}(e) = l(e)\sum_{i \in T_e}\sum_{j \in T_{e'}} M_{i,j}-M_{j,i}$. Normalizing this vector so its sum is 1 leads to the following biological interpretation: > The normalized differential abundance vectors indicate which taxa contributed to the Unifrac distance and by what percentage. For typical metagenomics and metatranscriptomic studies, the distributions $P$ and $Q$ are supported on the leaves of the tree $T$. In this case, minimizing flows and differential abundance vectors can be defined for all nodes, as well as at any fixed taxonomic rank simply by summing over the lower taxa. Figure \[fig:RealFlow\] gives such an example at the phylum level. Application to real data ======================== To demonstrate the utility of EMDUnifrac on real data, we evaluate it on the 16S rRNA data from a previous study[@willing2010pyrosequencing]. This data consists of 454 pyrosequenced fecal samples from a cohort of 40 twin pairs. The RDPII [@maidak2001rdp] and BLAST [@altschul1990basic] classifications were accessed via QIIME/QIITA [@caporaso2010qiime]. For simplicity, we focus here on the phylum level, and so summed these classifications to this level. We selected a subset of the data consisting of 49 healthy samples and 16 ulcerative colitis samples and used the silva taxonomic tree[@yilmaz2013silva] for the EMDUnifrac computation. We evaluated the EMDUnifrac algorithm on all 2,080 pairs of samples and performed a principle coordinate analysis (PCoA) on the resulting distance matrix (disregarding the flows for each pair). The result of this is contained in part (A) of Figure \[fig:RealFlow\]. Next, we combined all the healthy samples and combined all the ulcerative colitis samples and evaluated EMDUnifrac on these two combined samples. The returned minimizing flow is depicted in part (B) of Figure \[fig:RealFlow\]. The corresponding differential abundance vector is shown in part (C). Even though upon visual inspection, the PCoA plot in part (A) does not show much distinction between healthy and ulcerative colitis samples (compare to the similar plot contained in Figure 2 of [@willing2010pyrosequencing]), the differential abundance vector immediately leads to the conclusion that the ulcerative colitis samples are primarily enriched for Actinobacteria and Proteobacteria, while being deficient in Bacteroidetes. This is consistent with other studies where the same trend was observed in irritable bowel disease subjects, but using more intricate analysis techniques [@frank2007molecular; @spor2011unravelling; @manichanh2012gut], and demonstrates how utilizing the minimizing flow results in more information than simply using a dimension reduction technique (here PCoA) on the pairwise Unifrac distances. \#1[m[\#1]{}]{} [@cXX@]{} [cc]{} &\ Speed comparison to Unifrac --------------------------- As modern comparative metagenomics studies often perform all pair-wise Unifrac distance computations for datasets consisting of tens to thousands of samples, it is important to compute such distances in an efficient manner. We show in Theorem \[thm:speed\] below that our Algorithm \[alg:EMDUnifrac\] to compute EMDUnifrac runs in space and time complexity linear in the total support of the input vectors (so less than or equal to the number of nodes in the tree). To assess practical performance of Algorithm \[alg:EMDUnifrac\], we compared it to the fastest previous implementation of Unifrac, called FastUnifrac [@hamady2010fast]. We randomly generated trees (using the ete2 toolkit [@ete3]) with the number of leaf nodes ranging from 10 to 90,000. We then randomly produced pairs of distributions on the leaves using an exponential distribution with scale parameter 1. Importantly, EMDUnifrac can handle distributions with weights on leaf nodes as well as internal nodes while FastUnifrac only allows distributions with weights on the leaf nodes. We performed 10 replicates for each number of tree leaves and 10 replicates for each tree topology. Using the same fixed computational resources, we then ran FastUnifrac, EMDUnifrac in a mode that computes and returns the computed flow, and EMDUnifrac in a mode that just calculates the distance (and does not return an optimal flow, returning identical output to FastUnifrac). The average timings (over each number of tree leaves) are depicted in Figure \[fig:timing\]. These results indicate that in either mode, EMDUnifrac is more computationally efficient than FastUnifrac, and when just the resulting distance is desired, takes less than half a second to run, even on trees with 90,000 leaves (noting that our implementation is a non-optimized, Python implementation). Proof of correctness ==================== In this section, we detail our algorithm to compute EMDUnifrac, prove its correctness, and assess its computational complexity. Definitions and algorithm ------------------------- We begin with some definitions. Let $P$ and $Q$ be probability distributions on a tree $T$ with distance matrix $D_{i,j}$ and edge set $E$. Recall that $\Gamma(P,Q)$ is the set of all flows from $P$ to $Q$ in $T$. By an abuse of notation, we write $i \in T$ to denote a vertex of our tree. For such a vertex $i \in T$ we will say $i$ is a source if $P_i \geq Q_i$ and say $i$ is a sink otherwise. Let $T_{source}$ and $T_{sink}$ denote the sets of sources and sinks, respectively. Next, we select an arbitrary vertex of $T$ and distinguish it as the root $\rho$ of $T$. The choice is a convenience of notation. For each $i \in T$ let $a(i)$ be the unique neighbor of $i$ in $T$ which lies on the path from $i$ to $\rho$ in $T$. Thus the edges of $T$ are determined by the set of ordered pairs $(i,a(i)))$ for $i \in T$. Let $e_i$ denote the edge $(i,a(i))$. As $T$ is a tree, each edge $e \in E$ is a bridge. Thus its removal partitions the vertices into two disjoint subsets. We denote the subset containing $\rho$ by $T_e$ and the other by $T'_e$. Let $l: E \rightarrow \mathbb{R}_{\geq0}$ define a set of edge weights or lengths on $E$. For $i,j \in T$, define $\pi(i,j)$ to be the set of edges comprising the unique minimal path from $i$ to $j$ in $T$ and let $D_{i,j}= \sum_{e \in \pi(i,j)}{l(e)}$ be the distance from $i$ to $j$ in $T$. The pseudocode for EMDUnifrac is contained in Algorithm \[alg:EMDUnifrac\]. Intuitively, the algorithm begins at the leaves of the tree and “pushes" mass toward the root; satisfying the sources and sinks for each subtree encountered in the progression. The matrix $G$ tracks the mass still needed to be moved to or from each vertex by the algorithm, while the vector $w$ tracks the length of paths traversed by mass at each step. To implement EMDUnifrac, we first choose an ordering on the set of vertices of $T$ such that for $i,j \in T$, $i$ is an element of the path from $j$ to $\rho$ only if $i \geq j$. A natural such ordering is defined by partitioning the vertices of $T$ by the number of edges in the path to $\rho$, and then ordering vertices such that increasing indices correspond to decreasing path lengths to $\rho$. We then let $G$ and $M$ be a pair of matrices whose rows and columns are indexed by the vertices of $T$ with respect to an ordering as above. Let $G_{i,\cdot}$ denote the $i$-th row of the matrix $G$. Initialize both $G$ and $M$ to be the zero matrix. Let $w$ be a vector indexed by the vertices of $T$, initialized to be the zero vector. For any vector $u$, define ${\rm skel}(u)$ to be the binary vector of the same dimension as $u$ such for all $i$, ${\rm skel}(u(i)) = 1$ if $u(i) \neq 0$ and ${\rm skel}(u(i)) = 0$ otherwise. \[section:Algorithm\] $P,Q,\rho, T, E = \{i,a(i)\} \mbox{ for } i\in T, l$ $M,G = \mathbf{0}$ ${\rm EMDUnifrac}(P,Q) = 0$ ${\rm DiffAbund} = \vec{0}$ $M_{i,i} = \min(P_i,Q_i)$ $G_{i,i} = P_i - Q_i$ $M_{j,k} = \min(G_{i,j},-G_{i,k})$ $G_{i,j} = G_{i,j} - M_{j,k}$ $G_{i,k} = G_{i,k} + M_{j,k}$ ${\rm EMDUnifrac}(P,Q) = {\rm EMDUnifrac}(P,Q) + (w_j+w_k) M_{j,k}$ $G_{a(i),\cdot} = G_{a(i),\cdot} + G_{i,\cdot}$ ${\rm DiffAbund}({(i,a(i))}) = l(i,a(i))\sum_{t\in T}{G_{i,t}}$ $G_{i,\cdot} = \vec{0}$ $w = w + l(i,a(i)){\rm skel}(G_{i,\cdot})$ $M$, ${\rm EMDUnifrac}(P,Q)$ and ${\rm DiffAbund}$ Proof of correctness -------------------- We first prove an alternate characterization of the earth movers distance for probability distributions on a tree $T$. We have that $${\rm EMDUnifrac}(P,Q) = \min_{M \in \Gamma(P,Q)}{\sum_{e \in E}\sum_{i\in{T_e}}\sum_{j\in {T'_e}}{l(e)\left(M_{i,j}+M_{j,i}\right)}}.$$ Let $1_{\pi(i,j)}(e):E \rightarrow \{0,1\}$ be the indicator function of the path from $i$ to $j$ in $T$. That is, $1_{\pi(i,j)}(e) = 1$ if $e$ is an edge in the path from $i$ to $j$ and is $0$ otherwise. We then have that for any flow $M \in \Gamma(P,Q)$ $$\begin{aligned} \sum_{i,j \in T}{D_{i,j}M_{i,j}} &= \sum_{i\in T}\sum_{j\in T}\left(\sum_{e\in E}{l(e)1_{\pi(i,j)}(e)}\right)M_{i,j} \label{lem1_1}\\ &= \sum_{e\in E}\sum_{i\in T}\sum_{j \in T}{l(e)1_{\pi(i,j)}(e)}M_{i,j}\label{lem1_2}\\ &= \sum_{e\in E}\:\sum_{\substack{i\in \\ T_e \cup T'_e}}\:\sum_{\substack{i\in \\ T_e \cup T'_e}}{l(e)1_{\pi(i,j)}(e)}M_{i,j}\label{lem1_3}\\ &= \sum_{e\in E}\left(\sum_{i\in T_e}\sum_{j \in T'_e}l(e)M_{i,j}+\sum_{i\in T'_e}\sum_{j \in T_e}l(e)M_{i,j}\right)\label{lem1_4}\\ &= \sum_{e\in E}\sum_{i\in T_e}\sum_{j \in T'_e}l(e)\left(M_{i,j}+M_{j,i}\right).\label{lem1_5}\end{aligned}$$ The above equalities are justified as follows. To begin, (\[lem1\_1\]) follows from the definition of the distance function and the use of the characteristic function of the path between vertices to expand the summation over all edges of the graph. Next, (\[lem1\_2\]) and (\[lem1\_3\]) reorder the summation and express the vertex set in terms of the partitions defined above by edge deletion. We have that $1_{\pi(i,j)}(e)=1$ if and only if the vertices $i$ and $j$ belong to distinct partitions $T_e$ and $T'_e$, from which (\[lem1\_4\]) follows. Finally, in (\[lem1\_5\]) we condense the summation notation by reordering the last sum and grouping terms. Taking the minimum over all $M \in \Gamma(P,Q)$ yields the earth mover’s distance on the left hand side, and thus the desired result is obtained. Next, we prove a lower bound on the summands involved in the above definition of the earth mover’s distance. For any flow $M \in \Gamma(P.Q)$ and any $e \in E$ we have that $$\sum_{i\in T_e}\sum_{j \in T'_e}l(e)(M_{i,j}+M_{j,i}) \geq l(e)\left\|\sum_{i \in T_e}P(i)-Q(i)\right\|.$$ Further, the differential abundance vector, indexed by the edges of $T$ and having entries ${\rm DiffAbund}_e = l(e)\sum_{i \in T_e}\sum_{j \in T_{e'}} M_{i,j}-M_{j,i}$ is unique, regardless of the minimizing flow $M$. We have that $$\begin{aligned} l(e)\left\|\sum_{i \in T_e} P_i-Q_i\right\| &= \left\|l(e)\sum_{i \in T_e}\left(\sum_{j \in T}M_{i,j}-\sum_{j \in T}M_{j,i}\right)\right\|\label{lem2_1}\\ &= \left\|\sum_{i \in T_e}l(e)\sum_{j \in T}M_{i,j}-M_{j,i}\right\|\label{lem2_2}\\ &= \left\|\sum_{i \in T_e}\left(\sum_{j \in T_e}l(e)(M_{i,j}-M_{j,i})+\sum_{j \in T'_e}l(e)(M_{i,j}-M_{j,i})\right)\right\|\label{lem2_3}\\ &= \left\|\sum_{i \in T_e}\sum_{j \in T_e}l(e)(M_{i,j}-M_{j,i})+\sum_{i \in T_e}\sum_{j \in T'_e}l(e)(M_{i,j}-M_{j,i})\right\|\label{lem2_4}\\ &= \left\|\sum_{i \in T_e}\sum_{j \in T'_e}l(e)(M_{i,j}-M_{j,i})\right\|\label{lem2_5}\\ &\leq \sum_{i \in T_e}\sum_{j \in T'_e}l(e)(M_{i,j}+M_{j,i})\label{lem2_6}.\end{aligned}$$ Equations (\[lem2\_1\]) and (\[lem2\_2\]) above follow from expanding $P_i$ and $Q_i$ in terms of the row and column sums of $M$. Equations (\[lem2\_3\]) and (\[lem2\_4\]) reorganize the inner sums by way of the partitions $T_e$ and $T'_e$ and then group terms. Next we note that $\sum_{i \in T_e}\sum_{j \in T_e}l(e)(M_{i,j}-M_{j,i}) = 0$, as each term $M_{i,j}$ occurs precisely twice, once with each sign, which is reflected in (\[lem2\_5\]) above. This line also demonstrates the uniqueness of ${\rm DiffAbund}_e$, as the quantity is here shown to be equal to $\sum_{i \in T_e} P_i-Q_i$, which depends on the distributions $P$ and $Q$. Finally, we apply the triangle inequality to yield our result. What follows is a brief technical lemma used to prove that the matrix $M$ produced by\ EMDUnifrac is indeed a flow. Let $m \in T$ be arbitrary. Then for all $n \in T$ such that $n$ is a vertex along the path from $m$ to $\rho$, when $i = n$ in the loop beginning at line 1 of Algorithm (\[alg:EMDUnifrac\]) we have that one of the following hold: If $m$ is a source, then at the beginning of line 4 of algorithm \[alg:EMDUnifrac\] we have that $$\begin{aligned} P_m &= G_{n,m} + \sum_{k \in T} M_{m,k}\\ Q_m &= \sum_{k_ \in T} M_{k,m}.\end{aligned}$$ Alternately, if $m$ is a sink, then at the beginning of line 4 of Algorithm (\[alg:EMDUnifrac\]) we have that $$\begin{aligned} P_m &= \sum_{k \in T} M_{m,k} \\ Q_m &= -G_{n,m} + \sum_{k \in T} M_{k,m}.\end{aligned}$$ This follows by induction. Suppose $m$ is a source and let $i=m$ in the loop at line 1 of Algorithm \[alg:EMDUnifrac\]. Then $\min(P_m,Q_m) = Q_m$ and hence, by construction, $M_{m,m} = Q_m, G_{m,m} = P_m - Q_m$. Further, before beginning the loop at line 4 of Algorithm \[alg:EMDUnifrac\], every other entry of the $m$-th row of $M$ and $G$ are zero. This is because the elements of these rows are first potentially assigned non-zero values for $i=m$ in the midst of lines 6, 7 or 8. Thus at the beginning of line 4 of Algorithm \[alg:EMDUnifrac\], we have $$\begin{aligned} P_m &= G_{m,m} + \sum_{k \in T} M_{m,k}, \\ Q_m &= \sum_{k \in T} M_{k,m}.\end{aligned}$$ Thus the claim holds for $i=m$. Now suppose inductively that the above equalities holds when $i=j$ for some vertex $j \geq m$ on the path from $m$ to $\rho$ in $T$. We shall show the equalities holds for $i=a(j)$. As Algorithm \[alg:EMDUnifrac\] proceeds in the loop at line 1 to the vertex for $i = a(j)$, we have that $G_{a(j),m} \geq 0$ and thus by line 5 of Algorithm \[alg:EMDUnifrac\], the $m$-th column of $M$ is left unchanged. Hence the sum $\sum_{k \in T} M_{k,m}$ remains unchanged. Additionally, any change to $G_{a(j),m}$ during the loop at line 5 is compensated by a change to $\sum_{k \in T} M_{m,k}$, thus $$G_{a(j),m} + \sum_{k \in T} M_{m,k} = G_{j,m} + \sum_{k \in T} M_{m,k} = P_m.$$ Thus, inductively, the claims holds for all vertices along the path from $m$ to $\rho$ in $T$ and $m$ a source. Symmetric reasoning holds for the case of $m$ a sink. We now prove our main result. The EMDUnifrac algorithm in Algorithm \[alg:EMDUnifrac\] produces the earth mover’s distance $EMDUnifrac(P,Q)$ and a corresponding minimizing flow $M$. We first show that $M$ is indeed a flow. Upon the algorithm reaching the root $\rho$, that is when $i=\|T\|$ in line 4 of Algorithm \[alg:EMDUnifrac\], we have traversed every vertex of $T$, so that $$\begin{aligned} 0 &= 1-1\label{prf1_1}\\ &= \sum_{k \in T} P_k - Q_k \label{prf1_2}\\ &= \sum_{k \in T_{source}}P_k - Q_k + \sum_{k \in T_{sink}}P_k - Q_k\label{prf1_=3}\\ &= \sum_{k \in T_{source}}\left(G_{\rho,k} + \sum_{l \in T} M_{k,l} - \sum_{l \in T}M_{l,k}\right) + \sum_{k \in T_{sink}}\left(\sum_{l \in T} M_{k,l} - \left(-G_{\rho,k} + \sum_{l \in T}M_{l,k}\right)\right)\label{prf1_4}\\ &= \sum_{k \in T }\sum_{l \in T}M_{l,k} - \sum_{k \in T }\sum_{l \in T}M_{k,l} + \sum_{k \in T_{source}}G_{\rho,k} + \sum_{k \in T_{source}}G_{\rho,k}\label{prf1_5}\\ &= \sum_{k \in T} G_{\rho,k}. \label{prf1_6}\end{aligned}$$ The above equalities are justified as follows. In (\[prf1\_4\]) we expand the terms $P_k$ and $Q_k$ in terms of the matrices $G$ and $M$, as shown in Lemma $3$, since $\rho$ is an element of the path from any vertex to $\rho$. We then group terms in (\[prf1\_5\]) and (\[prf1\_6\]) by repeatedly using that $T_{source} \cup T_{sink} = T$, before canceling the symmetric summations of the elements of $M$. It then follows that the sum of the positive elements of $G_{\rho,\cdot}$ is equal to the sum of the negative elements of $G_{\rho,\cdot}$, and thus, by construction of the loops at lines 4 and 5 of Algorithm \[alg:EMDUnifrac\], the algorithm must terminate with $G_{\rho,\cdot}$ identically zero. As we still have that for each $i \in T$, $P_i = \sum_{k \in T} M_{j,k}, Q_i = \sum_{k \in T} M_{k,j}$, up to the addition or subtraction of $G_{\rho,i}=0$, $M$ must be a flow. Now we show that $M$ minimizes the sum defining the earth mover’s distance. By Lemmas $1$ and $2$, it suffices to show that $\sum_{i \in T_e}\sum_{j \in T'_e}l(e)(M_{i,j}+M_{j,i}) = \|\sum_{i \in T_e}P_i-Q_i\|$ for all $e \in E$. Given the ordering of the vertices chosen for the algorithm above, let $n \in T-\{\rho\}$ be arbitrary. To begin, we make some observations regarding the structure of the matrix $G$ and its relationship to $M$ in the algorithm. Note, that by construction, at the termination of the loop at line 4 of Algorithm \[alg:EMDUnifrac\] for $i=n$, the entries of $G_{n,\cdot}$ all have the same sign, as the the loops at lines 4 and 5 have the effect of pairwise choosing elements of opposite signs and using one to eliminate the other. This process terminates when elements of one or the other sign are exhausted. Second, note that for $k \in T'_{e_n}$ and $m > n$, either $G_{m,k} = 0$ or has the same sign as $G_{n,k}$, as any change to the entries of $G_{\cdot, k}$ is made to move the value toward zero by a quantity bounded by the magnitude of the entry. This again follows from examination of the inner most loop of the algorithm, as well as the evolution of rows of $G$. Finally, note that across all $i \in T'_{e_n}, j \in T_{e_n}$ either $M_{j,i} = 0$ or $M_{j,i} = 0$. This follows since $M_{i,j}$, respectively $M_{j,i}$, is only assigned a non-zero value in the case of $G_{m,i} > 0$, respectively $G_{m,i} < 0$. By the above observation regarding the signs of the elements of $G_{n,\cdot}$, only one of these conditions holds across $i,j$. Now, without loss of generality, assume $$\left\|\sum_{i \in T_{e_n}} P_i - Q_i\right\| = \sum_{i \in T_{e_n}} P_i - Q_i$$ as the argument for the alternate case is analogous. We then have that $$\begin{aligned} \left\|\sum_{i \in T_{e_n}} P_i - Q_i\right\| &= \sum_{i \in T_{e_n}} P_i - Q_i\\ &= \sum_{i \in T_{e_n}} \sum_{j \in T} M_{i,j} - M_{j,i}\\ &= \sum_{i \in T_{e_n}} \sum_{j \in T'_{e_n}} M_{i,j} - M_{j,i}\label{prf2_3}\\ &= \sum_{i \in T_{e_n}} \sum_{j \in T'_{e_n}} M_{i,j} + M_{j,i}.\label{prf2_4}\end{aligned}$$ The change of sign in moving from (\[prf2\_3\]) to (\[prf2\_4\]) follows from the above observation that at least one of $M_{i,j}$ or $M_{j,i}$ must be identically zero, and that the sum must be non-negative. Hence $-M_{j,i} = 0 = M_{j,i}$. Scaling the above equality by $l({e_n})$ yields $$\left\|\sum_{i \in T_{e_n}} P_i - Q_i\right\| = \sum_{i \in T_{e_n}} \sum_{j \in T'_{e_n}} M_{i,j} + M_{j,i}.$$ Having achieved the lower bound established in Lemma 2, we must have that the flow $M$ is a minimizer for the sum defining ${\rm EMDUnifrac}(P,Q)$. \[thm:speed\] Let $\|{\rm supp }\ P\|,\|{\rm supp }\ Q\|$ denote the number of elements in the support of the probability distributions $P$ and $Q$, respectively. Let $s = \|{\rm supp }\ P\|+\|{\rm supp }\ Q\|$. Then the EMDUnifrac algorithm has time and space complexity $O(s)$. We first consider the time complexity of EMDUnifrac. Note that each iteration of the loop at line 5 of Algorithm \[alg:EMDUnifrac\] has the effect of satisfying a source $i$ or sink $j$, that is, establishing the appropriate row sum $i$ or column sum $j$ of the matrix $M$. Further, the loop at line 5 only visits a pair of vertices $(i,j)$ in the case that both source $i$ and sink $j$ have not been satisfied, that is, that both $P(i) \neq \sum_{k\in T} M_{i,k}$ and $Q(i) \neq \sum_{k\in T} M_{k,i}$. As there are $s$ such row or column sums to satisfy, the loop at line 5 is evaluated at most $s$ times. Hence the time complexity of the algorithm is, in total, linear in $s$. Now we examine the space requirements of EMDUnifrac. By the above, the matrix $M$ is sparse. That is, there are most $s$ evaluations of the loop at line 5 of Algorithm \[alg:EMDUnifrac\] and thus, including the assignment of values to $M$ at line 2 of the algorithm, at most $2s$ non-zero entries in $M$. Additionally, line 3 of the algorithm assigns a non-zero entry to $G$ at most $n$ times, while line 12 has the effect of passing non-zero entries of $G$ from one row to another prior to being removed in line 13. Thus the number of non-zero entries of $G$ is bounded by $s$. Finally, the vector $w$ in Algorithm \[alg:EMDUnifrac\] is one dimensional, having at most $s$ non-zero entries. Hence the total space requirements of the algorithm are also linear in $s$. Conclusion ========== This paper implements the ideas of [@evans2012phylogenetic] to capitalize on the characterization of the Unifrac distance as the earth mover’s distance on weighted phylogenetic trees. The EMDUnifrac algorithm developed, and proved correct, allows for extremely rapid computation of weighted and unweighted Unifrac distances between biological communities. In particular, computations times are much faster than FastUnifrac when producing identical outputs, as seen in Figure \[fig:timing\]. These very rapid computation times and the minimal storage requirements, both linear in the number of taxa present, allow for all pairwise comparisons in large-scale studies. An example of this sort of implementation is seen in Figure \[fig:RealFlow\]. In addition to the Unifrac distance, EMDUnifrac is capable of producing both a minimizing flow and a differential abundance vector. The minimizing flow and differential abundance vector can be viewed as partitions of the numeric Unifrac distance, partitions which describe how operational taxonomic units present in biological communities contribute to their measured dissimilarity. The results shown in Figure \[fig:RealFlow\] demonstrate an application in which the raw Unifrac value has less apparent discerning power than achieved by an analysis of the differential abundance vector. Finally, EMDUnifrac algorithm is capable of computing the Unifrac distance for any weighted tree, not merely those trees weighted at their leaves. This allows for the comparison of whole genome shotgun metagenomes, an application in which weights are assigned at various levels of phylogenetic specificity. This is a capability apparently lacking in FastUnifrac, which combined with the ability to produce differential abundance vectors gives EMDUnifrac broader utility than current computational tools for measuring Unifrac distances. The EMDUnifrac algorithm itself is an extension of the ideas presented in the [@mangul2015reference] which considered De Bruijn graphs. Both leverage the earth mover’s distance to compute biologically relevant metrics on graphs. In EMDUnifrac, the topological benefits of a tree are exploited to speed computation in ways which are not possible under the more complicated topology of a De Bruijn graph. Acknowledgement {#acknowledgement .unnumbered} =============== #### Funding: None. [^1]: ${}^*$[[email protected]]([email protected])	{ "pile_set_name": "ArXiv" }
--- abstract: 'We suggest that nonequilibrium Bose-Einstein condensates may occur in time-periodically driven interacting Bose gases. Employing the model of a periodically forced bosonic Josephson junction, we demonstrate that resonance-induced ground state-like many-particle Floquet states possess an almost perfect degree of coherence, as corresponding to a mesoscopically occupied, explicitly time-dependent single-particle orbital. In marked contrast to the customary time-independent Bose-Einstein condensates, the order parameter of such systems is destroyed by violent fluctuations when the particle number becomes too large, signaling the non-existence of a proper mean field limit.' author: - Bettina Gertjerenken - Martin Holthaus date: 'October 23, 2014' title: Fluctuations of the order parameter of a mesoscopic Floquet condensate --- Nonequilibrium condensates ========================== In the wake of traditional textbook teaching, Bose-Einstein condensation usually is associated with thermal equilibrium: At sufficiently low temperatures a Bose gas “condenses” into the lowest single-particle state [@LandauLifshitz75; @Huang87; @PathriaBeale11]. In the present paper we take a theoretical step towards the exploration of nonequilibrium condensates [@VorbergEtAl13]. The possible existence of such nonequilibrium condensates is reflected in the fundamental Penrose-Onsager criterion [@PenroseOnsager56] for Bose-Einstein condensation in a system of $N$ repulsively interacting Bose particles, where $N$ is large: This criterion does neither require thermal equilibrium nor even steady states [@Leggett01]. Instead, it takes recourse to the one-particle reduced density matrix $$\varrho(\bm r, \bm r';t) = \langle \Psi_N(t) \| \widehat{\psi}^\dagger(\bm r) \widehat{\psi}(\bm r') \| \Psi_N(t) \rangle \; , \label{eq:OPR}$$ where $\| \Psi_N(t) \rangle$ denotes the state of the $N$-Boson system at time $t$, and $\widehat{\psi}^\dagger(\bm r)$ and $\widehat{\psi}(\bm r)$ are the usual creation and annihilation operators, obeing the Bose commutation relation $\left[ \widehat{\psi}(\bm r), \widehat{\psi}^\dagger(\bm r') \right] = \delta(\bm r - \bm r')$. Considered as a matrix with indices $\bm r$ and $\bm r'$, its diagonal elements $\varrho(\bm r, \bm r;t)$ provide the particle density of the system at the position $\bm r$. Because at each moment this matrix is Hermitian, it can be decomposed in terms of a complete set of orthonormal single-particle functions $\chi_j^{\phantom }(\bm r;t)$ with eigenvalues $n_j(t)$, such that $$\varrho(\bm r, \bm r';t) = \sum_j n_j(t) \chi_j^{\phantom }(\bm r,t) \chi_j^(\bm r', t) \; . \label{eq:SDD}$$ According to Penrose and Onsager one has a simple Bose-Einstein condensate when the largest eigenvalue $n_{\max}(t)$ is on the order of $N$, all others being of order $1$; the corresponding eigenfunction $\chi_{\max}(\bm r;t)$ then is the condensate wave function [@PenroseOnsager56]. In the most favorable case where $n_{\max}(t) = N$, the density matrix (\[eq:SDD\]) reduces to a projector, times $N$, onto the $N$-fold occupied single-particle orbital $\chi_{\max}(\bm r;t)$. As a matter of principle, this orbital can have an arbitrarily strong time-dependence. Here we suggest that a particular type of nonequilibrium condensate may become experimentally accessible when an interacting Bose gas is subjected to a resonant time-periodic force. In general, when a quantum system evolves according to a Hamiltonian $H(t) = H(t+T)$ which depends periodically on time with period $T$, and remains bounded, the Floquet theorem asserts that there exists a complete set of solutions to the time-dependent Schrödinger equation which possess the particular form $\|\psi_j(t)\rangle = \| u_j(t) \rangle \exp(-{{\mathrm i}}\varepsilon_j t/\hbar)$, where the Floquet functions $\| u_j(t) \rangle = \| u_j(t+T) \rangle$ inherit the imposed periodicity in time, and the quantities $\varepsilon_j$ which determine the growth rates of the accompanying phases are known as quasienergies [@Shirley65; @Zeldovich67; @Sambe73; @FainshteinEtAl78]. Each solution to the time-dependent Schrödinger equation can be expanded in this Floquet-state basis with constant coefficients, implying that one can describe, e.g., a time-periodically driven ideal Bose gas by means of single-particle Floquet orbitals which carry constant occupation numbers [@VorbergEtAl13]. In particular, it makes sense to introduce the notion of a macroscopically occupied Floquet state. Recent experiments with Bose-Einstein condensates in optical lattices subjected to strong time-periodic forcing already have demonstrated dynamic localization [@LignierEtAl07; @EckardtEtAl09; @ArimondoEtAl12], coherent control of the superfluid-to-Mott insulator transition [@ZenesiniEtAl09], giant Bloch oscillations [@AlbertiEtAl09; @HallerEtAl10], frustrated classical magnetism [@StruckEtAl11], controlled correlated tunneling [@ChenEtAl11], artificial tunable gauge fields [@StruckEtAl12; @StruckEtAl13], and effective ferromagnetic domains [@ParkerEtAl13]. Without claiming completeness of this list, these experiments testify that a macroscopic matter wave persists in the presence of strong time-periodic forcing. Appearance of new ground state ============================== For our theoretical considerations we employ the model of a periodically driven bosonic Josephson junction, which can be realized, for instance, with Bose-Einstein condensates in optical double-well potentials [@GatiMKO07]. The junction itself is described by the Lipkin-Meshkov-Glick Hamiltonian [@LipkinEtAl65] $$H_0 = -\frac{\hbar\Omega}{2} \left(a^{\phantom{\dagger}}_1 a_2^{\dagger} + a_1^{\dagger}a^{\phantom{\dagger}}_2\right) + \hbar \kappa \left( a_1^{\dagger} a_1^{\dagger} a^{\phantom{\dagger}}_1 a^{\phantom{\dagger}}_1 + a_2^{\dagger} a_2^{\dagger} a^{\phantom{\dagger}}_2 a^{\phantom{\dagger}}_2 \right) \; , \label{eq:UDJ}$$ where the operators $a_j^{\dagger}$ and $a^{\phantom{\dagger}}_j$ create and annihilate, respectively, a Bose particle in the $j$th well ($j=1,2)$, obeying the commutation relation $\left[ a^{\phantom{\dagger}}_j, a_k^{\dagger} \right] = \delta_{jk}$. Moreover, $\hbar\Omega$ is the single-particle tunneling splitting, and $2\hbar\kappa$ quantifies the repulsion energy of each pair of bosons occupying the same well. This Hamiltonian (\[eq:UDJ\]) had originally been devised for testing many-body approximation schemes [@LipkinEtAl65]; its paradigmatic importance as a nontrivial, but well tractable model for interacting Bose gases has been realized shortly after experiments with ultracold atomic vapors became standard practice [@MilburnEtAl97; @ParkinsWalls98]. We extend this model by assuming that the two wells are time-periodically shifted with frequency $\omega$ in phase opposition to each other, giving rise to the total Hamiltonian [@HolthausStenholm01; @JinasunderaEtAl06] $$H(t) = H_0 + \hbar\mu_1 \cos(\omega t) \left( a_1^{\dagger} a^{\phantom{\dagger}}_1 - a_2^{\dagger} a^{\phantom{\dagger}}_2 \right) \; . \label{eq:DJJ}$$ Here the driving amplitude $\hbar\mu_1$ denotes the maximum shift in energy; bosonic Josephson junctions with different driving schemes have also been considered in the literature [@MahmudEtAl05; @BoukobzaEtAl10]. With the spatial degree of freedom being restricted to two discrete sites, the one-particle reduced density matrix (\[eq:OPR\]) becomes the $2\times 2$ matrix $$\varrho = \left( \begin{array}{cc} \langle a_1^{\dagger} a_1^{\phantom{\dagger}} \rangle & \langle a_1^{\dagger} a_2^{\phantom{\dagger}} \rangle \\ \langle a_2^{\dagger} a_1^{\phantom{\dagger}} \rangle & \langle a_2^{\dagger} a_2^{\phantom{\dagger}} \rangle \end{array} \right) \; , \label{eq:TBT}$$ where the expectation values are taken with respect to the state under consideration. If the junction is filled with $N$ particles, the Penrose-Onsager criterion now always confirms the existence of a condensate, but the question is whether this condensate is simple or fragmented: In the former case the larger eigenvalue of the matrix (\[eq:TBT\]) is close to $N$, while the smaller is close to zero, thus indicating that there exists one single-particle state which is almost $N$-fold occupied. In contrast, the condensate is fragmented when both eigenvalues are close to $N/2$. Therefore, Leggett has introduced the quantity [@Leggett01] $$\eta =2 N^{-2} \, {\rm tr} \, \varrho^2 - 1 \, , \label{eq:ETA}$$ computed from the trace of the squared density matrix, as an invariant measure of the degree of the system’s coherence: One has $\eta = 1$ for a pure simple condensate, whereas $\eta = 0$ in the case of maximum fragmentation. In Fig. \[F\_1\] we plot $\eta$ for the lowest five energy eigenstates of the undriven junction (\[eq:UDJ\]). Here the scaled interaction strength $N\kappa/\Omega = 0.95$ is kept fixed as the particle number $N$ is varied, as is required for approaching the mean field limit: In a rigorous mathematical setting, that limit, which is described by the Gross-Pitaevskii theory, requires $N \to \infty$ such that the product of particle number and interaction strength remains constant [@LiebEtAl00]. Evidently, the ground state $\| 0 \rangle$ is almost fully coherent when $N$ becomes sufficiently large, thereby indicating the existence of a [bona fide]{} order parameter, namely, of a single-particle orbital which is occupied by almost all of the $N$ particles when the system (\[eq:UDJ\]) is in its ground state, and which thus constitutes the macroscopic wave function. It is well known that the exact ground state of the Hamiltonian (\[eq:UDJ\]) coincides with an exact coherent state only when $\hbar\kappa = 0$. However, the difference between the exact ground state $\| 0 \rangle$ and an exactly coherent state here becomes insignificant when approaching the mean field limit, when $\hbar\kappa$ vanishes proportionally to $1/N$. We now extend this analysis to the driven junction (\[eq:DJJ\]). Here we focus on [resonant]{} driving, [i.e.]{}, we choose the frequency $\omega$ such that $\hbar\omega$ equals the spacing $E_{r+1} - E_r$ of the unperturbed energy eigenvalues $E_j$ of the junction (\[eq:UDJ\]) at a particular state label $j = r$. Figure \[F\_2\] (a) shows the exact quasienergies of the system for $N = 100$ particles, scaled interaction strength $N\kappa/\Omega = 0.95$, and scaled driving frequency $\omega/\Omega = 1.62$. This implies $r = 8$, so that the unperturbed $N$-particle energy eigenstates $\| 8 \rangle$ and $\| 9 \rangle$ are almost exactly on resonance. Note that a Floquet state can be factorized according to $$\begin{aligned} & & \| u_j(t) \rangle\exp(-{{\mathrm i}}\varepsilon_j t/\hbar) \nonumber \\ & = & \| u_j(t) {{\mathrm e}}^{i m \omega t}\rangle \exp(-{{\mathrm i}}[\varepsilon_j + m\hbar\omega] t/\hbar) \label{eq:BZS}\end{aligned}$$ with an arbitrary positive or negative integer $m$, so that the Floquet function $\|u_j(t) {{\mathrm e}}^{i m \omega t}\rangle$ remains $T$-periodic, with $T = 2\pi/\omega$. This means, loosely speaking, that “the quasienergies are defined only up to an integer multiple of $\hbar\omega$.” More precisely, the quasienergy of a Floquet state labeled by $j$ has to be regarded as an infinite set of representatives $\varepsilon_j + m\hbar\omega$ spaced by $\hbar\omega$, implying that each Brillouin zone of the quasienergy spectrum of width $\hbar\omega$ contains precisely one representative of each state. The Brillouin zone of quasienergies displayed in Fig. \[F\_2\] (a) features a regular fan of almost equidistant lines, which can be explained analytically by means of a standard resonance approximation [@BermanZaslavsky77; @Holthaus95; @GertjerenkenHolthaus14; @GertjerenkenHolthaus14b]. In the vicinity of the state $\| r \rangle$ singled out by the condition $\hbar\omega = E_{r+1} - E_r$, the dynamics of the driven $N$-particle system can be mapped to that of an effective quasiparticle, named “floton”, which moves in a cosine potential well without external driving, such that the energies of this quasiparticle yield the quasienergies of the near-resonant Floquet states [@GertjerenkenHolthaus14; @GertjerenkenHolthaus14b]: $$\varepsilon_{k} = E_r + \frac{1}{8}E''_r\alpha_k(q) \quad \mathrm{mod}\; \hbar\omega \; , \label{eq:MAP}$$ where $E''_r$ denotes the formal (discrete) second derivative of the unperturbed eigenvalues $E_j$ with respect to the state label $j$, evaluated at the resonant state $j = r$, and $\alpha_k(q)$ is a characteristic value of the Mathieu equation. Using the notation of Ref. [@AbramowitzStegun72], one has $\alpha_k(q) = a_k(q)$ for quantum numbers $k = 0,2,4\ldots$ labeling the even eigenstates of the floton quasiparticle, while $\alpha_k(q) = b_k(q)$ for $k = 1,3,5,\ldots\;$. The Mathieu parameter $q$ is proportional to the driving amplitude, $$q = \frac{2}{E''_r/(\hbar\omega)} \frac{2\mu_1}{\omega} \langle r \| a_1^{\dagger} a^{\phantom{\dagger}}_1 - a_2^{\dagger} a^{\phantom{\dagger}}_2 \| r - 1 \rangle \; .$$ The important feature here is the appearance of a new quantum number $k$: The resonant state $\| r \rangle$ turns into the floton ground state $k = 0$; the neighboring states of the unperturbed junction (\[eq:UDJ\]) are transformed into its excitations $k > 0$. In Fig. \[F\_3\] we depict the degree of coherence (\[eq:ETA\]) for the exact near-resonant Floquet states, computed numerically, with floton quantum numbers $k = 0, \ldots, 4$. The similarity to the previous Fig. \[F\_1\] is striking: Indeed the “resonant ground state” $k = 0$ is an almost coherent state, in the sense that it corresponds to an $N$-fold occupied, periodically time-dependent single-particle orbital. Thus, here we encounter an example of Floquet engineering: The driving is not employed primarily to excite the system, but rather to create a new effective Hamiltonian [@GoldmanDalibard14], describing the floton quasiparticle, and providing a new ground state into which the actual particles can condense. This Floquet condensate constitutes a collective mode of response to the drive which remains perfectly coherent in the course of time. Order parameter fluctuations ============================ However, there is a fundamental difference between such Floquet condensates and the customary, time-independent Bose-Einstein condensates which shows up if one tries to recover the mean field regime: In Fig. \[F\_4\] we show the maximum degree of coherence $\eta_{\rm max}$, taken over all Floquet states of the driven Josephson junction (\[eq:DJJ\]) with $\omega/\Omega = 1.62$, vs. the scaled driving strength; again the interaction strength is adjusted such that $N\kappa/\Omega = 0.95$. In panel (a) we take $N = 100$: Here we observe extended intervals where $\eta_{\rm max} = 1$ with high accuracy, caused by the floton state $k = 0$, and large fluctuations occurring when $2\mu_1/\omega \approx 0.9$. The interval magnified in the inset is scanned again in panel (b), but now with $N = 500$; here additional small fluctuations appear. Iterating this procedure, the interval framed in the inset of panel (b) is evaluated in panel (c) with $N = 1000$; here the fluctuations become more violent. In panel (d), where $N = 2000$, even the baseline of the fluctuations is shifted downward. These results indicate that the size of a resonant Floquet condensate remains restricted to mesoscopically large particle numbers, while its order parameter would be destroyed for high $N$ by large fluctuations. The origin of these fluctuations is closely related to Eq. (\[eq:BZS\]), that is, to the Brillouin-zone structure of the quasienergy spectrum: Each zone contains $N+1$ quasienergy eigenvalues, as corresponding to the dimension of the junction’s Hilbert space when there are $N$ Bose particles, so that the eigenvalue density is proportional to $N$. On the other hand, eigenvalues falling into the same symmetry class are not allowed to cross. The quasienergy operator of the driven junction (\[eq:DJJ\]) remains invariant when the site labels are exchanged and simultaneously time is shifted by half a period; the Floquet functions therefore are even or odd under this generalized parity. Hence, neither “odd” nor “even” quasienergies may cross each other, which necessarily leads to a vast multitude of anticrossings when $N$ becomes large, each one indicating hybridization of the participating Floquet states. This mechanism effectuates a degradation of the order parameter; each dip seen in panel (b) can be traced to an isolated avoided quasienergy crossing. The Mathieu approximation (\[eq:MAP\]) locally reduces the driven $N$-particle system to an almost equivalent, integrable single-particle one, neglecting, in the sense of the rotating-wave approximation, fast-oscillating coupling terms [@GertjerenkenHolthaus14; @GertjerenkenHolthaus14b]. While for low driving amplitudes these couplings only produce anticrossings which are too small to detect on the scale of Fig. \[F\_2\] (a), their effect becomes stronger when $2\mu_1/\omega$ is increased. This eventually leads to a chaotic spectrum, as exemplified in Fig. \[F\_2\] (b). In the sequence shown in Fig. \[F\_4\], there are two opposing tendencies: On the one hand, the eigenvalue density increases by a factor of $20$ when enhancing $N$ from $100$ to $2000$; on the other, the interaction strength $\hbar\kappa$ is reduced by $1/20$. But evidently, this reduction is over-compensated by the growth of the particle number. While the individual anticrossings tend to become smaller upon reducing $\hbar\kappa$, they proliferate and overlap upon increasing $N$ to such an extent that the resulting multiple hybridizations forbid the formation of an order parameter: When the system becomes too complex, it does not possess a simple mean field description. This absence of a proper mean field limit is closely related to the absence of an adiabatic limit in periodically driven quantum systems [@HoneEtAl97]. Conclusions =========== Since the appearance of resonances is a generic feature of driven nonlinear quantum systems, we anticipate that the findings reported in this work are not restricted to our particular model (\[eq:DJJ\]). Thus, we may summarize our main results as follows: [(i)]{} Resonantly driven Bose gases allow the formation of nonequilibrium Bose-Einstein condensates, with the resonance-induced effective ground state corresponding to a mesoscopically occupied, periodically time-dependent single-particle orbital; [(ii)]{} the coherence of such condensates is destroyed when the particle number becomes large, a mean field limit cannot be reached. This non-existence of a mean field limit should be detectable through large fluctuations of the system’s coherence in a series of measurements in which the particle number varies slightly from shot to shot. We acknowledge support from the Deutsche Forschungsgemeinschaft (DFG) through grant No. HO 1771/6-2. The computations were performed on the HPC cluster HERO, located at the University of Oldenburg and funded by the DFG through its Major Research Instrumentation Programme (INST 184/108-1 FUGG), and by the Ministry of Science and Culture (MWK) of the Lower Saxony State. [99]{} L. D. Landau and E. M. Lifshitz, [Statistical Physics: Part 1]{} (Third edition, Butterworth-Heinemann, Oxford, 1975). K. Huang, [Statistical Mechanics]{} (Second edition, Wiley, New York, 1987). R. K. Pathria and P. D. Beale, [Statistical Mechanics]{} (Third edition, Butterworth-Heinemann, Oxford, 2011). D. Vorberg, W. Wustmann, R. Ketzmerick, and A. Eckardt, Phys. Rev. Lett. [111]{}, 240405 (2013). O. Penrose and L. Onsager, Phys. Rev. [104]{}, 576 (1956). A. J. Leggett, Rev. Mod. Phys. [73]{}, 307 (2001). J. H. Shirley Phys. Rev. [138]{}, B979 (1965). Ya. B. Zel’dovich, Sov. Phys. JETP [24]{}, 1006 (1967) \[Zh. Eksp. Teor. Fiz. [51]{}, 1492 (1966)\]. H. Sambe, Phys. Rev. A [7]{}, 2203 (1973). A. G. Fainshtein, N. L. Manakov, and L. P. Rapoport, J. Phys. B [11]{}, 2561 (1978). H. Lignier, C. Sias, D. Ciampini, Y. Singh, A. Zenesini, O. Morsch, and E. Arimondo, Phys. Rev. Lett. [99]{}, 220403 (2007). A. Eckardt, M. Holthaus, H. Lignier, A. Zenesini, D. Ciampini, O. Morsch, and E. Arimondo, Phys. Rev. A [79]{}, 013611 (2009). E. Arimondo, D. Ciampini, A. Eckardt, M. Holthaus, and O. Morsch, Adv. At. Mol. Opt. Phys. [61]{}, 515 (2012). A. Zenesini, H. Lignier, D. Ciampini, O. Morsch, and E. Arimondo, Phys. Rev. Lett. [102]{}, 100403 (2009). A. Alberti, V. V. Ivanov, G. M. Tino, and G. Ferrari, Nature Physics [5]{}, 547 (2009). E. Haller, R. Hart, M. J. Mark, J. G. Danzl, L. Reichsöllner, and H.-C. Nägerl, Phys. Rev. Lett. [104]{}, 200403 (2010). J. Struck, C. Ölschläger, R. Le Targat, P. Soltan-Panahi, A. Eckardt, M. Lewenstein, P. Windpassinger, and K. Sengstock, Science [333]{}, 996 (2011). Y.-A. Chen, S. Nascimbène, M. Aidelsburger, M. Atala, S. Trotzky, and I. Bloch, Phys. Rev. Lett. [107]{}, 210405 (2011). J. Struck, C. Ölschläger, M. Weinberg, P. Hauke, J. Simonet, A. Eckardt, M. Lewenstein, K. Sengstock, and P. Windpassinger, Phys. Rev. Lett. [108]{}, 225304 (2012). J. Struck, M. Weinberg, C. Ölschläger, P. Windpassinger, J. Simonet, K. Sengstock, R. Höppner, P. Hauke, A. Eckardt, M. Lewenstein, and L. Mathey, Nature Physics [9]{}, 738 (2013). C. V. Parker, L.-C. Ha, and C. Chin, Nature Physics [9]{}, 769 (2013). R. Gati and M. K. Oberthaler, J. Phys. B [40]{}, R61 (2007). H. J. Lipkin, N. Meshkov, and A. J. Glick, Nuc. Phys. [62]{}, 188 (1965). G. J. Milburn, J. Corney, E. M. Wright, and D. F. Walls, Phys. Rev. A [55]{}, 4318 (1997). A. S. Parkins and D. F. Walls, Phys. Rep. [303]{}, 1 (1998). M. Holthaus and S. Stenholm, Eur. Phys. J. B [20]{}, 451 (2001). T. Jinasundera, C. Weiss, and M. Holthaus, Chem. Phys. [322]{}, 118 (2006). K. W. Mahmud, H. Perry, and W. P. Reinhardt, Phys. Rev. A [71]{}, 023615 (2005). E. Boukobza, M. G. Moore, D. Cohen, and A. Vardi, Phys. Rev. Lett. [104]{}, 240402 (2010). E. H. Lieb, R. Seiringer, and J. Yngvason, Phys. Rev. A [61]{}, 043602 (2000). G. P. Berman and G. M. Zaslavsky, Phys. Lett. A [61]{}, 295 (1977). M. Holthaus, Chaos, Solitons & Fractals [5]{}, 1143 (1995). B. Gertjerenken and M. Holthaus, New J. Phys. [16]{}, 093009 (2014). B. Gertjerenken and M. Holthaus, preprint (submitted to Phys. Rev. A). M. Abramowitz and I. A. Stegun (Eds.), [Handbook of Mathematical Functions]{} (Dover, New York, 1972), ch. 20. N. Goldman and J. Dalibard, Phys. Rev. X [4]{}, 031027 (2014). D. W. Hone, R. Ketzmerick, and W. Kohn, Phys. Rev. A [56*]{}, 4045 (1997).	{ "pile_set_name": "ArXiv" }
CERN-PH-TH/2015-068 [Non-linear curvature inhomogeneities]{}\ \ Massimo Giovannini [^1]\ \ 0.3cm Abstract 0.1cm The non-perturbative curvature inhomogeneities induced by relativistic viscous fluids are not conserved in the large-scale limit. However when the bulk viscosity is a function of the total energy density of the plasma (or of the trace of the extrinsic curvature) the relevant evolution equations develop a further symmetry preventing the non-linear growth of curvature perturbations. In this situation the fully inhomogeneous evolution can be solved to leading order in the gradient expansion. Over large-scales both the acceleration and the curvature inhomogeneities are determined by the bulk viscosity coefficients. Conversely the shear viscosity does not affect the evolution of the curvature and does not produce any acceleration. The curvature modes analyzed here do not depend on the choice of time hypersurfaces and are invariant for infinitesimal coordinate transformations in the perturbative regime. Introduction {#sec1} ============ The temperature and the polarization anisotropies of the cosmic microwave background are adequately described in the framework of the linear theory (see, e.g. the WMAP or WMAP9 data essentially consistent with the Planck explorer data [@data1; @data2; @data3]). However throughout the whole history of the plasma a non-linear growth of the curvature inhomogeneities cannot be excluded. The initial conditions of the temperature and polarization anisotropies of the cosmic microwave background are customarily set after neutrino decoupling [@data1; @data2; @peebles] but the evolution of curvature perturbations starts at a much earlier epoch namely during inflation and possibly even before. It is then particularly interesting, from the theoretical viewpoint, to define and discuss plausible non-perturbative generalisations of the curvature inhomogeneities in different physical situations going beyond the conventional cases of a single scalar field and of a perfect barotropic fluid. The theme of this paper concerns the non-perturbative generalization of the curvature inhomogeneities, their evolution and their physical relevance when the energy-momentum tensor is dominated by relativistic viscous effects at large-scales. Prior to the formulation of the inflationary paradigm the non-linear evolution of curvature perturbations has been analyzed in various frameworks. The approach followed here can be traced back to the expansion in spatial gradients of the geometry [@bel1; @bel2] (see also [@LL]). The gradient expansion has been used either in the proximity of the big-bang singularity or away from it. Close to the singularity the geometry may be highly anisotropic but it turns out to be rather homogeneous [@bel1; @bel2]. As soon as an inflationary event horizon is formed [@star; @salope; @tomita] any finite portion of the event horizon gradually loses the memory of an initially imposed anisotropy or inhomogeneity so that the metric attains the observed regularity regardless of the initial boundary conditions as hypothesized in the past [@hoyle; @zel1; @misner]. One of the central themes of the inflationary paradigm [@inf1; @inf2] is to wash out primeval anisotropies in the expansion right after the formation of the inflationary event horizon (see, however, Ref. [@barrow]). Probably the first non-linear generalization of inflationary curvature perturbations has been proposed in [@salope] after the pioneering analyses on the gauge-invariant treatment of linearised cosmological perturbations [@pio]. These have been subsequently scrutinized and rediscovered by different authors [@shell]. The non-perturbative generalizations discussed here may also have some impact on the neighbouring problems such as the higher-order approaches to cosmological perturbations (see [@NH] and references therein). It is appropriate to gauge the effects of the viscosity coefficients that can play a relevant role both in the early and in the late Universe. Since this investigation addresses the evolution of the non-linear curvature perturbation in the relativistic theory of viscous fluids (see, for instance, [@visc1] and references therein) the assumption of the he strict reversibility of the system will be dropped. The shear viscosity suppresses exponentially the traceless part of the extrinsic curvature. The bulk viscosity enters the definition of the curvature perturbations and may cause their non-conservations for typical length-scales larger than the Hubble radius. While the bulk viscosity does affect directly the non-linear deceleration parameter (possibly causing accelerated expansion) the opposite is true for the shear viscosity. From the technical viewpoint the variables introduced in the present analysis do not depend on the choice of time hypersurfaces and they are exactly invariant for infinitesimal coordinate transformations in the perturbative regime. It will be argued that the gravitating viscous fluids lead, under certain conditions, to a further symmetry preventing the non-linear growth of curvature inhomogeneities. The same set of conditions will also be shown to be compatible with a quasi-de Sitter stage of expansion. In the reversible limit our variables coincide with the ones conventionally defined in the case of a single scalar field or for a perfect barotropic fluid. The results pursued here can also be used to deduce the corrections to the linear theory without going through second or even higher-order calculations [@NH]. Finally, in the single inflaton case the large-scale cosmological perturbations are sometimes treated within the so-called $\delta {\mathcal N}$ formalism [@salope; @shell] where ${\mathcal N}$ denotes the inhomogeneous expansion rate integrated in time generalizing to the non-perturbative level the total number of inflationary efolds. The presence of non-adiabatic fluctuations of the pressure make the formalism less appealing but we will show that, in some specific cases, the $\delta {\mathcal N}$ formalism can also be applied in the case of relativistic viscous fluid. The layout of the investigation is the following. In section \[sec2\] we shall discuss the basic aspects of the geometry and of its coordinate transformations. The interplay between relativistic viscous fluids and general relativistic gradient expansion is addressed in section \[sec3\]. Section \[sec4\] is devoted to the non-perturbative evolution of large-scale curvature inhomogeneities in the viscous case. The perturbative limit of our considerations is discussed in section \[sec5\]. Section \[sec6\] contains the concluding remarks. Nonlinear gauge transformations {#sec2} =============================== In the Arnowitt-Deser-Misner formalism [@ADM1] (ADM in what follows) the line element is expressed in terms of the conventional $(3+1)$-dimensional decomposition: $$ds^2 = g_{\mu\nu}(\tau,\vec{x})\, dx^{\mu} \, dx^{\nu} =N^2 d\tau^2 - \gamma_{ij} (d x^{i} + N^{i} d\tau) ( d x^{j} + N^{j} d\tau), \label{adm1}$$ where $N=N(\tau,\vec{x})$ denotes the lapse function, $N^{i}=N^{i}(\tau,\vec{x})$ is the shift vector and $\gamma_{ij}=\gamma_{ij}(\tau, \vec{x})$ is the spatial three metric[^2]. In the ADM variables of Eq. (\[adm1\]) the extrinsic curvature of the spatial slices (i.e. $K_{ij}(\tau,\vec{x})$) and the components of the Ricci tensor of the spatial slices (i.e. $r_{ij}(\tau,\vec{x})$) become: $$\begin{aligned} K_{ij}(\tau,\vec{x}) &=& \frac{1}{2 N} \biggl[- \partial_{\tau}\gamma_{ij} + \nabla_{i}N_{j} + \nabla_{j} N_{i} \biggr], \label{ADM1a}\\ r_{ij}(\tau,\vec{x}) &=& \partial_{m} \, ^{(3)}\Gamma^{m}_{ij} -\partial_{j} ^{(3)}\Gamma_{i m}^{m} + ^{(3)}\Gamma_{i j}^{m} \,^{(3)}\Gamma_{m n}^{n} - ^{(3)}\Gamma_{j n}^{m} \,^{(3)}\Gamma_{i m}^{n}, \label{ADM1b}\end{aligned}$$ where, for short, $^{(3)}\nabla_{i}= \nabla_{i}$ is the covariant derivative defined[^3] with respect to the metric $\gamma_{ij}$; $\partial_{\tau}$ denotes a derivation with respect to the time coordinate $\tau$ and $^{(3)}\Gamma_{i j}^{m}$ are the Christoffel symbols computed from $\gamma_{ij}$. Note that $\Gamma_{ij}^{m} = ^{(3)}\Gamma_{i j}^{m}$ but only in the case $N_{i}=0$. From Eq. (\[adm1\]) the unit time-like vector normal to the $x^{0} = \tau = {\mathrm constant}$ hypersurface is $n_{\mu} = (N, 0)$ and $n^{\mu} = (1/N, \, - N^{i}/N)$. The choices $N=1$ and $N_{i}=0$ corresponds to the geodesic slicing leading to the Gaussian normal coordinates, a popular gauge in numerical relativity [@shap]. The condition $N_{i}=0$ implies that coordinate observers coincide with normal observers: the normal vector $n^{\mu}$ has vanishing spatial component. Without positing a specific gauge choice, Eqs. (\[ADM1a\]) and (\[ADM1b\]) lead to the following expressions for the components of the Ricci tensor: $$\begin{aligned} R_{0}^{0} &=& \frac{\partial_{\tau} K}{N}- \mathrm{Tr}K^2 + \frac{\nabla^2 N}{N} - \frac{N^{m}}{N} \nabla_{m} K + \frac{N^{q}}{N} \biggl(\nabla_{q} K - \nabla_{k} K^{k}_{q} \biggr), \label{ADM7}\\ R_{i}^{0} &=& \frac{1}{N} \biggl(\nabla_{i} K - \nabla_{k} K^{k}_{i} \biggr), \label{ADM8}\\ R_{i}^{j} &=& \frac{1}{N} \partial_{\tau} K_{i}^{j} - K K_{i}^{j} - r_{i}^{j} + \frac{1}{N} \nabla_{i} \nabla^{j} N - \frac {N^{m}}{N} \nabla_{m} K_{i}^{j} \nonumber\\ &+& \frac{1}{N} \nabla_{m} N^{j} K^{m}_{i}- \frac{1}{N} \nabla_{i} N^{m} K_{m}^{j} - \frac{N^{j}}{N} \biggl(\nabla_{i} K - \nabla_{k} K^{k}_{i} \biggr), \label{ADM9}\end{aligned}$$ where the notations $K = \gamma^{ij} K_{ij}$ and $\mathrm{Tr}K^2 = K_{i}^{j} K_{j}^{i}$ have been adopted. Notice that in the linear theory, for infinitesimal coordinate transformations of the type: $$x^{\mu} \to \tilde{x}^{\mu} = x^{\mu} + \epsilon^{\mu}, \qquad \epsilon^{\mu} = (\epsilon^{0} , \, \epsilon^{i}), \label{INF}$$ the metric fluctuations change as the Lie derivative in the direction $\epsilon_{\mu}$ [@pio] (see also section \[sec5\]). The perturbative expansion underlying Eq. (\[INF\]) assumes the separation of the geometry into a background value supplemented by a perturbation. The choice of the temporal gauge defines the spatial hypersurface of fixed coordinate time while the choice of the spatial gauge determines the worldlines of fixed spatial coordinates. The coordinate system is completely specified when both $\epsilon^{0}$ and $\epsilon^{i}$ are assigned. It is however possible to define various sets of gauge-invariant variables that do not change under Eq. (\[INF\]). In the linearized treatment, the gauge parameters are of the same order of the metric perturbations: as soon as, in some gauge, the perturbation variables grow non-linear and affect the background geometry the linearised approximation is no longer tenable. The approach pursued here does not assume the validity of the perturbative expansion insofar as the geometry is not split into a background value supplemented by a perturbation with small amplitude: this treatment is arguably the most suitable for the unambiguous analysis of some backreaction problems. When the metric is not linearized around a specific background, the coordinate transformations will not necessarily be infinitesimal and shall be parametrized as $x^{\mu} \to x^{\mu} = Y^{\mu}(x)$ or, in more explicit terms, as: $$\tau \to T= T(\tau, \vec{x}), \qquad x^{i} \to Y^{i} = Y^{i}(\tau, \vec{x}). \label{adm2}$$ Under the transformation of Eq. (\[adm2\]) the metric components of Eq. (\[adm1\]) will change as $$g_{\alpha\beta}(\tau, \vec{x}) = G_{\mu\nu}(T,\vec{Y}) \, \biggl(\frac{\partial X^{\mu}}{\partial x^{\alpha}}\biggr) \, \biggl(\frac{\partial X^{\nu}}{\partial x^{\beta}}\biggr). \label{admin3}$$ The explicit form of Eq. (\[admin3\]) can be written, schematically, as[^4]: $$\begin{aligned} (N^2 - N_{k} N^{k}) &=& (\alpha^2 - \beta_{k} \beta^{k}) (\partial_{\tau} T)^2 - 2 \beta_{i} \partial_{\tau} T \partial_{\tau} Y^{i} - \overline{\gamma}_{ij} \partial_{\tau} Y^{i} \partial_{\tau} Y^{j}, \label{admin4a}\\ N_{i} &=& - (\alpha^2 - \beta_{k} \beta^{k}) \partial_{\tau} T \partial_{i} T + 2 \beta_{k} \partial_{\tau} T \, \partial_{i} Y^{k} + \overline{\gamma}_{\ell k} \partial_{\tau} Y^{\ell} \partial_{i} Y^{k}, \label{admin4b}\\ \gamma_{ij} &=& - (\alpha^2 - \beta_{k} \beta^{k}) \partial_{j} T \partial_{i} T + 2 \beta_{k} \partial_{i} T \, \partial_{j}Y^{k} + \overline{\gamma}_{\ell k} \partial_{i} Y^{k} \partial_{j} Y^{\ell}, \label{admin4c} \end{aligned}$$ where the lapse function, the shift vectors and the spatial three metric in the transformed system have been denoted, respectively, by $\alpha= \alpha(T,\vec{Y})$, $\beta_{i}= \beta_{i}(T,\vec{Y})$ and $\overline{\gamma}_{\ell k} = \overline{\gamma}_{\ell k}(T,\vec{Y})$; Eq. (\[adm1\]) becomes, in the transformed frame, $$ds^2 = G_{\mu\nu}(T,\vec{Y})\,dY^{\mu} \,dY^{\nu} =\alpha^2 d T^2 - \overline{\gamma}_{ij} ( d Y^{i} + \beta^{i} d T) ( d Y^{j} + \beta^{j} d T). \label{adm11}$$ We shall often refer to the concept of non-linear gauge-invariant variables. This terminology refers to the possibility of finding specific quantities that do not depend on the choice of time hypersurfaces and that are exactly invariant for infinitesimal coordinate transformations in the perturbative regime. Gravitating viscous fluids and gradient expansion {#sec3} ================================================= Whenever dissipative effects are included both in the energy-momentum tensor and in the particle current the physical meaning of he four-velocity $u^{\mu}$ must be specified. In the Eckart approach $u^{\mu}$ coincides with the velocity of particle transport. Conversely, in the Landau approach the velocity $u^{\mu}$ coincides with the velocity of the energy transport defined by the $(0 i)$ component of the energy-momentum tensor giving the energy flux. The Landau approach shall be privileged mainly for practical reasons. First and second viscosity in the Landau frame ---------------------------------------------- The total energy-momentum tensor of the problem is given as the perfect field contribution (characterized by a total pressure $p_{t}$ and a total energy density $\rho_{t}$) supplemented by the irreversible contribution: $$T_{\mu}^{\nu} = (p_{t} + \rho_{t}) u_{\mu} u^{\nu} - p_{t} \delta_{\mu}^{\nu} + {\mathcal T}_{\mu}^{\nu}(\eta,\xi), \label{T1}$$ where ${\mathcal T}_{\mu}^{\nu}(\eta,\xi)$ denotes the viscous energy momentum tensor depending on the first and second viscosities (i.e. $\eta$ and $\xi$): $${\mathcal T}_{\mu}^{\nu}(\eta,\xi) = 2 \eta \sigma_{\mu}^{\nu} + \xi {\mathcal P}_{\mu}^{\nu} \nabla_{\alpha} u^{\alpha},\qquad \sigma_{\mu\nu} = \frac{1}{2} {\mathcal P}_{\mu}^{\gamma} \,{\mathcal P}_{\nu}^{\delta} \,{\mathcal W}_{\gamma\delta}. \label{T2}$$ As usual, ${\mathcal P}_{\beta}^{\nu} = (\delta_{\beta}^{\nu} - u_{\beta} u^{\nu})$ and the tensor ${\mathcal W}_{\gamma\delta}$ appearing in Eq. (\[T2\]) is defined as: $${\mathcal W}_{\gamma\delta} = \nabla_{\gamma} u_{\delta} + \nabla_{\delta} u_{\gamma} - \frac{2}{3} \, g_{\gamma\delta} \, \nabla_{\lambda} u^{\lambda}. \label{T2a}$$ The total particle current will be denoted by $j^{\mu} = n_{t} u^{\mu} + \nu^{\mu}$ (where $n_{t}$ is the total concentration of the fluid while $\nu^{\mu}$ denotes the diffusion current). Note that $\nu^{\mu}$ will denote hereunder the relativistic thermal conduction four-vector. As we shall see explicitly $\nu^{\mu} u_{\mu} =0$ since $\nu_{\mu}$ can be written as $\nu_{\mu} = f(T, \rho_{t},p_{t} ) {\mathcal P}_{\mu}^{\alpha} \partial_{\alpha} \overline{\mu}$ where $\overline{\mu} = \mu/T$ and $\mu$ is the chemical potential. The viscous energy-momentum tensor at large-scales can be evaluated in the Landau-Lifshitz or in the Eckart frames. In the Eckart case the four-velocity $u_{\mu}$ appearing in Eq. (\[T1\]) denotes the velocity of the particle transport. The total[^5] particle current $j^{\mu}$ vanishes in the comoving frame. Consequently the Eckart frame is fixed by requiring that $j^{\alpha} \, u_{\alpha} =0$ while ${\mathcal T}^{\mu\nu} u_{\nu} \neq 0$. The Eckart approach [@visc3] (see also second paper of Ref. [@visc1]), seems preferable when the concentration of radiation quanta exceeds the concentration of the other species. Conversely, in the Landau-Lifshitz approach [@visc2] pure thermal conduction corresponds to an energy flux without particles: the four-velocity $u_{\mu}$ coincides with the velocity of the energy transport implying ${\mathcal T}^{\mu\nu} u_{\nu} =0$ . The two approaches are largely equivalent but the Landau-Lifshitz approach seems slightly more convenient, in the present context. Both the Eckart and the Landau approaches are suitable for the present class of problems where the typical scales are much larger than the mean free path. The second-order dissipative effects [@isr] become particularly relevant in the collisions of heavy ions [@RHIC] where, however, not all the numerous second-order terms have been so far included in the available analytical and numerical discussions. The covariant conservation of the total energy-momentum tensor (i.e. $\nabla_{\mu} T^{\mu\nu}=0$) can be projected along $u_{\nu}$ and along ${\mathcal P}_{\nu}^{\alpha}$; the two obtained equations together with the covariant conservation of the particle current are given hereunder: $$\begin{aligned} && \nabla_{\mu}[ (p_{t} + \rho_{t}) u^{\mu}] - u_{\alpha} \partial^{\alpha} p_{t} + u_{\beta} \nabla_{\alpha} {\mathcal T}^{\alpha\beta} =0, \label{c1}\\ && (p_{t} + \rho_{t}) u^{\beta} \nabla_{\beta} u^{\alpha} - \partial^{\alpha}p_{t} + u^{\alpha} u_{\beta} \partial^{\beta} p_{t} + {\mathcal P}^{\alpha}_{\nu} \nabla_{\mu} {\mathcal T}^{\mu\nu} =0, \label{c2}\\ && \nabla_{\alpha}( n_{t} u^{\alpha} + \nu^{\alpha}) =0. \label{c3}\end{aligned}$$ Using then Eqs. (\[c1\]) and (\[c3\]) together with the first principle of thermodynamics, the evolution of the entropy[^6] can be easily derived: $$\nabla_{\alpha} [ s u^{\alpha} - \overline{\mu} \nu^{\alpha} ] + \nu^{\alpha} \partial_{\alpha} \overline{\mu} = \nabla_{\alpha} u_{\beta} \, {\mathcal T}^{\alpha\beta}/T. \label{entro}$$ In Eq. (\[entro\]), as already mentioned after Eq. (\[T2a\]), $\overline{\mu} = \mu/T$ is the chemical potential rescaled through the temperature, $s$ is the entropy density and $\nu_{\alpha}$ is given by: $$\nu_{\alpha} = \chi \biggl(\frac{n_{t} T}{\rho_{t} + p_{t}}\biggr)^2 \biggl[ \partial_{\alpha} \overline{\mu} - u_{\alpha} u^{\beta}\partial_{\beta} \overline{\mu} \biggr], \label{nunu}$$ where $\chi$ denotes the heat transfer coefficient. From the definition of the viscous energy-momentum tensor of Eqs. (\[T2\]) and (\[T2a\]) we can also explicitly write the term at the right hand side of Eq. (\[entro\]) $$(\nabla_{\alpha} u_{\beta})\,{\mathcal T}^{\alpha\beta}/T = (\xi/T) (\nabla_{\alpha} u^{\alpha})^2 + 2 (\eta/T)\, \sigma_{\mu\nu} \,\sigma^{\mu\nu}. \label{entro2}$$ The adiabatic limit is recovered when the viscous contributions are neglected and the total entropy four-vector is conserved. The right hand side of Eq. (\[entro2\]) is positive semi-definite provided $\xi$ and $\eta$ are both positive semi-definite. It is appropriate to remark, at this point, that the perfect fluid contribution is characterized by the barotropic index $w= p_{t}/\rho_{t}$ and by the related sound speed $c_{st}^2 = p_{t}^{\prime}/\rho^{\prime}$. In linear theory the fluctuations of the total pressure of the fluid are customarily decomposed into an adiabatic component supplemented by the entropic (or simply non-adiabatic) contributions $\delta p_{t} = c_{st}^2 \delta \rho_{t} + \delta p_{\mathrm{nad}}$ (see, for instance, [@nonad] and the Eqs. (\[RPR\])–(\[zetapr\]) in section \[sec5\]). This occurrence would correspond, at the level of the non-linear discussion, to the case $w \neq c_{st}^2$ where $w$ may be a space-time dependent function (see section \[sec4\]). Even if the conventional terminology might suggest otherwise, the non-adiabatic modes have nothing to do with the global viscosity of the system and may even arise in a globally inviscid fluid. This potential confusion of the standard terminology should be borne in mind to avoid unwanted misunderstandings. Gradient expansion of the Einstein equations -------------------------------------------- The Ricci tensor reported in Eqs. (\[ADM7\]), (\[ADM8\]) and (\[ADM9\]) have been already written in a form where the terms containing spatial gradients are distinguished from all the other. The same criteria must be adopted when expressing the explicit components of the total energy-momentum tensor so that, at the very end, we shall be able to write down the Einstein equations in their contracted form: $$R_{\mu}^{\nu} = \ell_{P}^2 \biggl[ T_{\mu}^{\nu} - \frac{T}{2} \delta_{\mu}^{\nu} \biggr],\qquad \ell_{P} = \sqrt{8 \pi G}, \label{T6}$$ where $T = T_{\alpha}^{\alpha}$ is the trace of the total energy-momentum tensor and must not be confused with the effective temperature of the fluid appearing in the previous subsection. The parameter counting the gradients can be formally indicated as the gradient itself in units of the trace of the extrinsic curvature, i.e. as $\lambda= \nabla/K(\tau,\vec{x})$. From Eq. (\[T1\]) and bearing in mind the right hand side of Eq. (\[T6\]) we have: $$\begin{aligned} T_{\mu\nu} - \frac{T}{2} g_{\mu\nu} &=& (\rho_{t} + P_{eff}) u_{\mu} u_{\nu} - \frac{\rho_{t} - P_{eff}}{2} g_{\mu\nu} \nonumber\\ &+& \eta\biggl\{ \nabla_{\mu} u_{\nu} + \nabla_{\nu} u_{\mu} - u^{\alpha} \biggl[ u_{\mu} \nabla_{\alpha} u_{\nu} + u_{\nu} \nabla_{\alpha} u_{\mu} \biggr] - \frac{2}{3} {\mathcal P}_{\mu\nu} \nabla_{\lambda} u^{\lambda} \biggr\}. \label{EM1}\end{aligned}$$ The trace $T$ has been expressed in terms of the effective pressure $P_{eff}$ defined, in our case, as: $$T = T_{\alpha}^{\alpha} = \rho_{t} - 3 P_{eff}, \qquad P_{eff} = p_{t} - \xi \nabla_{\alpha} u^{\alpha}.$$ The term $\nabla_{\alpha} u^{\alpha}$ can be easily expanded in spatial gradients and the result is: $$\nabla_{\alpha} u^{\alpha} = - K - \frac{1}{N \sqrt{\gamma}} \partial_{k} [ N \sqrt{\gamma} u^{k}] + \frac{\partial_{\tau} u^2}{2 N} + {\mathcal O}(\lambda^3).$$ From Eqs. (\[ADM7\]), (\[ADM8\]) and (\[ADM9\]), the terms containing the shift vectors turn out to be ${\mathcal O}(\lambda)$ (see also [@salope; @shell]). It seems therefore appropriate to select the gauge $N_{i} =0$ where the coordinate observers coincide with normal observers. Equivalent choices, for the present purposes, include the coordinate system where the expansion is uniform (i.e. $K = K(\tau)$, in our notations) or the gauge where the energy density is uniform. The full expression of the various components of Eq. (\[EM1\]) are necessarily lengthy so we shall just exemplify the $(00)$ component and then mention the leading order results for the other components. The $(00)$ component of Eq. (\[EM1\]) is given by: $$T_{00} - \frac{T}{2} g_{00} = \frac{N^2}{2} ( \rho_{t} + 3 P_{eff}) + N^2 u^2 ( \rho_{t} + P_{eff}) - 2 \eta N u^2 {\mathcal F}(N,\gamma_{ij}, u_{k}), \label{inter1}$$ where ${\mathcal F}(N,\gamma_{ij}, u_{k})$ is defined as: $$\begin{aligned} {\mathcal F}(N,\gamma_{ij}, u_{k}) &=& \biggl\{ \partial_{\tau}[\sqrt{1 + u^2}] - u_{i} \partial^{i} N - N \sqrt{ 1 + u^2} \biggl[ u^{k} \partial_{k} \sqrt{1 + u^2} + u^{k} u^{j} K_{k j} \biggr] \nonumber\\ &-& \frac{1}{3 \sqrt{\gamma}} [ \partial_{\tau} ( \sqrt{\gamma} \sqrt{1 + u^2}) - \partial_{k}(N \sqrt{\gamma} u^{k})]\biggr\}. \label{inter1a}\end{aligned}$$ Except for the first term at the right hand side, all the remaining contributions are ${\mathcal O}(\lambda^2)$ since they contain, at least, two spatial gradients. From the lowest-order form of the momentum constraint (see below Eqs. (\[inter2\]) and (\[T8\])) the leading contribution of the spatial part of the velocity is clearly ${\mathcal O}(\lambda)$ so that $u^2 = \gamma_{ij} u^{i} u^{j} = {\mathcal O}(\lambda^2)$. The same procedure outlined in the case of the $(00)$ component of the energy-momentum tensor can be discussed for the remaining components and the leading order results are: $$\begin{aligned} T_{i}^{0} &=& (\rho_{t} + P_{eff}) u_{i} \, u^{0} + 2 \eta u_{k} u^{0} \overline{K}_{i}^{k} + {\mathcal O}(\lambda^3), \label{inter2}\\ T_{i}^{j} - \frac{T}{2} \delta_{i}^{j} &=& -\frac{1}{2} ( \rho_{t} - P_{eff}) \delta_{i}^{j} - 2 \eta \overline{K}_{i}^{j} + {\mathcal O}(\lambda^2), \label{inter3}\end{aligned}$$ where the traceless part of the extrinsic curvature $\overline{K}_{i}^{j} = K_{i}^{j} - \delta_{i}^{j} K/3$ has been explicitly introduced. Using the results of Eqs. (\[inter1\]), (\[inter2\]) and (\[inter3\]) together with Eqs. (\[ADM7\]), (\[ADM8\]) and (\[ADM9\]), Eq. (\[T6\]) will become: $$\begin{aligned} && \frac{1}{N}\partial_{\tau} K - \mathrm{Tr} K^2 = \frac{\ell_{P}^2}{2} ( \rho_{t} + 3 P_{eff}), \label{T7}\\ && \nabla_{i} K - \nabla_{k} K^{k}_{i} = N \ell_{P}^2 \biggl[ ( \rho_{t} + P_{eff}) u_{i} u^{0} + 2\eta \overline{K}_{i}^{j} u_{j} u^{0} \biggr], \label{T8}\\ && \frac{1}{N}\partial_{\tau} K_{i}^{j} - K K_{i}^{j} - r_{i}^{j} = \ell_{P}^2 \biggl[ \frac{(P_{eff} - \rho_{t})}{2} \delta_{i}^{j} - 2 \eta \overline{K}_{i}^{j} + \Pi_{i}^{j} \biggr], \label{T9}\end{aligned}$$ where $\Pi_{i}^{j}$ denotes the anisotropic stress which is by definition a traceless rank-two tensor in three dimensions. We have chosen to keep generic the form of $\Pi_{i}^{j}$ since it may contain all the potential sources of anisotropic stress not necessarily connected to the the fluid sector such as scalar fields or even gauge fields. In all cases $\Pi_{i}^{j}$ contains at least two spatial gradients. The anisotropic stress and the curvature $r_{i}^{j}$ are of higher order in the gradients but have been kept for the benefit of the forthcoming discussion aimed at showing the the traceless part of the extrinsic curvature is of higher order in the gradients and it the only component affected by the presence of shear viscosity. Finally, the evolution equations stemming from the covariant conservation of the energy-momentum tensor and of the particle current of Eqs. (\[c1\]), (\[c2\]) and (\[c3\]) are given by $$\begin{aligned} && \frac{1}{N} \partial_{\tau} \rho_{t} - K ( \rho_{t} + P_{eff}) =0 \label{EN}\\ && \frac{1}{N}\partial_{\tau} u^{i} + u^{k} \biggl[ \frac{\partial_{\tau} P_{eff}}{N(\rho_{t} + P_{eff})} \delta_{k}^{i} - 2 K_{k}^{i} \biggr] = \frac{\partial^{i} N}{N} - \frac{\partial^{i} P_{eff}}{\rho + P_{eff}}, \label{VEL}\\ && \frac{1}{N} \partial_{\tau} n_{t} - K n_{t} =0. \label{dens}\end{aligned}$$ The absence of the dissipative coefficients arising in the diffusion current is justified since these terms are of higher order in the spatial gradients, as it can be easily appreciated from Eq. (\[nunu\]). Decoupling of the shear contribution ------------------------------------ From Eq. (\[T9\]) the shear contribution only affects the evolution of the traceless part of the extrinsic curvature $\overline{K}_{i}^{j}$ and does not enter the deceleration parameter whose sign is solely determined by the bulk viscosity coefficient. Indeed, after taking the the traceless part of Eq. (\[T9\]) the following equation is obtained: $$\partial_{\tau} \overline{K}_{i}^{j} - N K \overline{K}_{i}^{j} = - 2 \eta N \ell_{P}^2 \overline{K}_{i}^{j} + N \ell_{P}^2 \Pi_{i}^{j} + N \overline{r}_{i}^{j} \label{S1}$$ where $\overline{r}_{i}^{j} = r_{i}^{j} - \delta_{i}^{j} \, r/3$ is the traceless part of the spatial curvature. In the general situation where $\eta(\tau, \vec{x})$ Eq. (\[S1\]) implies $$\begin{aligned} \overline{K}_{i}^{j}(\tau,\vec{x}) &=& \frac{\sqrt{\gamma(\tau_{}, \vec{x})}}{\sqrt{\gamma(\tau,\vec{x})}} \, \overline{K}_{i}^{j}(\tau_{},\vec{x}) e^{- 2 {\mathcal A}(\tau_{}, \tau, \vec{x})} \nonumber\\ &+& \frac{\ell_{P}^2}{\sqrt{\gamma(\tau,\vec{x})}} \int_{\tau_{}}^{\tau} d\tau^{\prime\prime} \sqrt{\gamma(\tau^{\prime\prime},\vec{x})} N(\tau^{\prime\prime},\vec{x})\, e^{ - 2 {\mathcal A}(\tau^{\prime\prime},\tau,\vec{x})}\, \Pi_{i}^{j}(\tau^{\prime\prime}, \vec{x}) d\tau^{\prime\prime} \nonumber\\ &+& \frac{1}{\sqrt{\gamma(\tau,\vec{x})}} \int_{\tau_{}}^{\tau} d\tau^{\prime\prime} \sqrt{\gamma(\tau^{\prime\prime},\vec{x})} N(\tau^{\prime\prime},\vec{x})\, e^{ - 2 {\mathcal A}(\tau^{\prime\prime},\tau,\vec{x})}\, \overline{r}_{i}^{j}(\tau^{\prime\prime}, \vec{x}) d\tau^{\prime\prime} \label{S2}\end{aligned}$$ where $\gamma = \mathrm{det}(\gamma_{ij})$. In Eq. (\[S2\]) $\tau_{}= \tau_{}(\vec{x})$ denotes some arbitrary integration time while for two generic times $\tau_{1}$ and $\tau_{2}$, ${\mathcal A}(\tau_{1},\tau_{2},\vec{x})$ is defined as: $${\mathcal A}(\tau_{1},\tau_{2},\vec{x}) = \ell_{P}^2 \int_{\tau_1}^{\tau_{2}} \eta(\tau^{\prime},\vec{x}) \, N(\tau^{\prime},\vec{x}) \, d\tau^{\prime}. \label{S3}$$ Equations (\[S2\]) and (\[S3\]) show that the traceless part of the extrinsic curvature is determined by the anisotropic stress and by the traceless part of the intrinsic curvature. Both quantities are of higher order in the gradient expansion[^7]. Equation (\[S3\]) shows that the shear viscosity suppresses the traceless part of the extrinsic curvature even further in comparison with the case $\eta \to 0$. The features of the damping are determined by the explicit expression of $\eta$. For a system dominated by radiation $\eta \sim \ell_{mfp} \, \rho_{t}$ where $\ell_{mfp}$ denotes the mean free path (for instance the Thomson mean free path prior to photon decoupling). In this case ${\mathcal A} \simeq K \ell_{mfp} \ll 1$. In more general terms, however, $\eta$ can depend on $\rho_{t}$ on the trace of the extrinsic curvature, on the total particle concentration and in all these cases $\overline{K}_{ij}$ may even be much smaller than the anisotropic stress. Since $\eta$ decouples from the trace of the extrinsic curvature, it does not contribute to the inhomogeneous generalization of the deceleration parameter. For the sake of comparison with the fully homogenous case we choose Gaussian normal coordinates and set $N=1$; in this situation Eq. (\[T7\]) can be written as: $$q(t,\vec{x}) \mathrm{Tr} K^2 = \ell_{P}^2 \biggl[ (\rho_{t} + P_{eff}) u_{0} u^{0} + \frac{P_{eff} - \rho_{t}}{2}\biggr] - 2 \eta u^2 {\mathcal F}(1,\gamma_{ij},u_{k}), \label{00cont}$$ where $q(\vec{x},t) = -1 + \dot{K}/{\rm Tr} K^2$ is the inhomogeneous generalisation of the deceleration parameter[^8] and the overdot denotes the derivative with respect to the cosmic time coordinate $t$ which coincides with $\tau$ in the case $N=1$. The function ${\mathcal F}(1,\gamma_{ij},u_{k})$ (defined in Eq. (\[inter1a\])) accounts for the higher-order corrections. In general terms ${\rm Tr} K^2 \geq K^2/3 \geq 0$, where the sign of equality (in the first relation) is reached in the isotropic limit. Since $\gamma^{ij}$ is always positive semi-definite, it is also clear that $u_{0}\,u^{0} = 1 + \gamma^{ij} u_{i} u_{j} \geq 1$. From Eq. (\[00cont\]) it also follows that $q(t,\vec{x})$ is always positive semi-definite as long as $(\rho + 3 P_{eff}) \geq 0$. This means that the sign of the generalized deceleration parameter only depends on $P_{eff}$ (and hence on the bulk viscosity) while the shear viscosity does not play any role. According to Eq. (\[00cont\]) the correction of the bulk viscosity only arises to second order in the gradient expansion where, however, the bulk viscosity also contributes through the term $(\rho + P_{eff}) u^2$ implicitly contained in $ (\rho_{t} + P_{eff}) u_{0} u^{0}$. Gauge invariant variables and their evolution {#sec4} ============================================= Generalities ------------ Under Eq. (\[adm2\]) the energy-momentum tensor transforms as $$T_{\mu\nu}(\tau,\vec{x}) \to \overline{T}_{\mu\nu}(T, \vec{Y}) = \biggl(\frac{\partial x^{\alpha}}{\partial Y^{\mu}} \biggr) \biggl(\frac{\partial x^{\beta}}{\partial Y^{\nu}} \biggr) T_{\alpha\beta}(\tau,\vec{x}), \label{TTtr}$$ where $T$ is the transformed time coordinate and will not be confused with the trace of the energy-momentum tensor. Equations (\[admin4a\])–(\[admin4b\]) and (\[admin4c\]) hold for the metric; similar expressions hold for the transformed components of the energy-momentum tensor. The analog of Eq. (\[admin4a\]) will be reported to fix the notations $$T_{00}(\tau,\vec{x}) = (\partial_{\tau} T)^2 \overline{T}_{00}(T,\vec{Y}) + 2 (\partial_{\tau} T) (\partial_{\tau} Y^{i}) \overline{T}_{0i}(T,\vec{Y}) +(\partial_{\tau} Y^{i}) (\partial_{\tau} Y^{j}) \overline{T}_{ij}(T,\vec{Y}). \label{TTtr2}$$ The explicit expressions of the $(0i)$ and $(ij)$ components can be easily written in terms of the notations of Eq. (\[TTtr2\]) and will be employed below. The coordinate transformation must preserve the order of the gradient expansion. This implies, from Eq. (\[admin4b\]), that $\beta_{i} = N_{i} =0$ and the coordinate transformation demands: $$\alpha^2 \partial_{\tau} T \partial_{i} T = \overline{\gamma}_{\ell k} \partial_{\tau} Y^{\ell} \partial_{i} Y^{k}. \label{first}$$ The transformations preserving the order of the gradient expansion [@salope] can be written as follows: $$\tau \to T= T(\tau, \vec{x}), \qquad x^{k} \to Y^{k}(\tau,\vec{x}) = f^{k}(\tau,\vec{x}) + F^{k}(\tau,\vec{x}). \label{trans}$$ By construction the function $f^{i}(\vec{x},\tau)$ does not contain any gradient while $F^{i}(\tau,\vec{x})$ contains at least one spatial gradient; $f^{k}(\tau, \vec{x})$ can then be parametrized as $f^{k}(\tau,\vec{x}) = x^{k} g(\tau,r)$ where $r = \sqrt{x_{i} x^{i}}$. Since in the transformation all the spatial gradients of $g(\tau,r)$ will automatically contribute to $F^{k}(\tau, \vec{x})$, the effect of $g(\tau,r)$ boils down to a redefinition of $\alpha$ in the transformed frame. For this reason and for the sake of simplicity we shall set $g(r,\tau)=1$. Equations (\[admin4a\]) and (\[admin4c\]), thanks to Eq. (\[first\]), will then lead, respectively, to the following pair of conditions: $$\begin{aligned} && N^2 = \alpha^2 (\partial_{\tau} T)^2 - \overline{\gamma}_{ij} \partial_{\tau} Y^{i} \partial_{\tau} Y^{j}, \label{second}\\ && \gamma_{ij} = - \alpha^2 \partial_{i}T \partial_{j} T + \overline{\gamma}_{k\ell} \partial_{i} Y^{k} \, \partial_{j} Y^{\ell}. \label{third}\end{aligned}$$ Recalling the explicit form of Eq. (\[trans\]), to lowest order in the spatial gradients, Eq. (\[third\]) implies that $\gamma_{ij}(\tau, \vec{x})= \gamma_{ij}(T, \vec{Y})$ while Eqs. (\[first\]) and (\[second\]) determine the explicit form of $F^{k}(T,\vec{Y})$; the explicit results are: $$F^{k}(T,\vec{Y}) = \int d T \, N^2(\tau) \frac{\partial^{k} T}{(\partial_{\tau} T)^2}, \qquad \gamma_{ij}(\tau, \vec{x})= \gamma_{ij}(T, \vec{Y}). \label{fourth}$$ Equations (\[trans\]) and (\[fourth\]) can be inserted into the various components of Eq. (\[TTtr\]) to obtain the transformation properties of the pressure, of the energy density and of the velocity: $$\begin{aligned} && \rho(\tau,\vec{x})=\overline{\rho}(T,\vec{Y}) , \qquad P_{eff}(\tau,\vec{x})= \overline{P}_{eff}(T,\vec{Y}), \nonumber\\ && \qquad u_{i} = \overline{u}_{i} + \alpha \partial_{i}T, \qquad N = \alpha \,\partial_{\tau} T. \label{fifth}\end{aligned}$$ Non-linear curvature inhomogeneities ------------------------------------ In the viscous case the non-linear generalization of the curvature perturbations on comoving orthogonal hypersurfaces is: $${\mathcal R}_{i}(\tau,\vec{x}) = \frac{1}{3} \nabla_{i} [\ln{(\sqrt{\gamma})}] - \frac{1}{3N} \partial_{\tau}[ \ln{(\sqrt{\gamma})}] \,\, u_{i}. \label{defR}$$ Using the transformation properties defined by Eqs. (\[first\]), (\[second\]), (\[third\]) and (\[fifth\]) we have that ${\mathcal R}_{i}$ transforms as: $$\begin{aligned} {\mathcal R}_{i}(\tau, \vec{x}) &\to& \overline{{\mathcal R}}_{i}(T, \vec{Y}) = \frac{1}{3} \frac{\partial [\ln{(\sqrt{\overline{\gamma}})}]}{\partial Y^{j}} \frac{\partial Y^{j}}{\partial x^{i}} + \frac{1}{3} \frac{\partial[ \ln{(\sqrt{\overline{\gamma}})}]}{\partial T} \frac{\partial T}{\partial x^{i}} \nonumber\\ &-& \frac{1}{3 \alpha} \biggl(\overline{u}_{i} + \alpha \frac{\partial T}{\partial x^{i}} \biggr) \frac{\partial [\ln{(\sqrt{\overline{\gamma}})}]}{\partial T}; \label{transR1}\end{aligned}$$ since the two intermediate terms simplify in Eq. (\[transR1\]), we have that the curvature inhomogeneities are invariant i.e. ${\mathcal R}_{i}(\tau,\vec{x}) = \overline{{\mathcal R}}_{i}(T,\vec{Y})$. Using the same strategy applied in the case of Eq. (\[defR\]), the non-linear generalization of the density contrast on uniform curvature hypersurfaces becomes[^9]: $$\zeta_{i}(\tau,\vec{x}) = \frac{1}{3} \nabla_{i}[ \ln{(\sqrt{\gamma})} ]+ \frac{\nabla_{i} \rho}{ 3(\rho + P_{eff})}. \label{defZ}$$ The same analysis leading to Eq. (\[transR1\]) can be performed in the case of the $\zeta_{i}(\tau,\vec{x})$: $$\begin{aligned} \zeta_{i}(\tau, \vec{x}) &\to& \overline{\zeta}_{i}(T, \vec{Y}) =\frac{1}{3} \frac{\partial [\ln{(\sqrt{\overline{\gamma}})}]}{\partial Y^{j}} \frac{\partial Y^{j}}{\partial x^{i}} + \frac{1}{3} \frac{\partial[ \ln{(\sqrt{\overline{\gamma}})}]}{\partial T} \frac{\partial T}{\partial x^{i}} \nonumber\\ &+& \frac{1}{3 (\overline{\rho}_{t} + \overline{P}_{eff})} \,\frac{\partial \overline{\rho}}{\partial Y^{j}} \frac{\partial Y^{j}}{\partial x^{i}} + \frac{1}{3 (\overline{\rho}_{t}+ \overline{P}_{eff})} \frac{\partial \overline{\rho}}{\partial T} \frac{\partial T}{\partial x^{i}}. \label{transzeta1}\end{aligned}$$ Equation (\[EN\]) (stemming from the covariant conservation of the total energy-momentum tensor in the transformed frame) implies that the derivative of $\overline{\rho}_{t}$ with respect to $T$ equals $\alpha \overline{K} ( \overline{\rho}_{t} + \overline{P}_{eff})$. Thus the second and fourth terms at the right hand side of Eq. (\[transzeta1\]) cancel since $\alpha \overline{K}= - \partial \ln{\sqrt{\overline{\gamma}}}/\partial T$. As in the case of Eq. (\[transR1\]) the invariance of $\zeta_{i}$ is manifest since $\zeta_{i}(\tau,\vec{x}) = \overline{\zeta}_{i}(T,\vec{Y})$. Since, in the general situation, the bulk viscosity coefficient is a space-time scalar function its derivative transforms non-trivially under Eq. (\[trans\]): $$\frac{\partial \xi }{\partial x^{i}}= \frac{\partial \overline{\xi}}{\partial Y^{i}} + \partial_{T} \overline{\xi} \partial_{i} T, \qquad \partial_{\tau} \xi = \frac{\partial \overline{\xi}}{\partial T} \partial_{\tau} T. \label{Z2}$$ Thanks to Eq. (\[Z2\]) we can obtain a further non-linear variable invariant under Eqs. (\[first\])–(\[second\]) and (\[third\])–(\[fifth\]) that has no analogue in the inviscid case: $${\mathcal Z}_{i}(\tau, \vec{x}) = \frac{1}{3} \nabla_{i} [\ln{(\sqrt{\gamma})}] + \frac{K N}{3} \frac{\partial_{i} \xi}{\partial_{\tau}\xi}. \label{Z1}$$ From the gauge-transformed expression of Eq. (\[Z1\]), using Eqs. (\[second\]) and (\[third\]) the gauge-invariance of ${\mathcal Z}_{i}(\tau, \vec{x})$ is easily demonstrated. Let us finally mention, for the sake of comparison, that the general form of Eq. (\[transR1\]) can be used to recover the well known results obtainable in the case of the single scalar field [@salope]. The $(0i)$ component of the energy-momentum tensor of a minimally coupled scalar field $T_{i}^{0}(\varphi)=\partial_{i} \varphi \partial_{\tau}\varphi/N^2$ implies that $u_{i} = N \nabla_{i}\varphi /(\partial_{\tau}\varphi)$. Therefore Eq. (\[defR\]) implies $${\mathcal R}_{i}(\tau,\vec{x}) = \frac{1}{3} \nabla_{i}[\ln {(\sqrt{\gamma})}] + \frac{K N}{3} \frac{\nabla_{i} \varphi}{\partial_{\tau}\varphi}, \label{conv1}$$ which is also invariant since, from Eq. (\[fifth\]), the transformation of $\varphi$ will be given by: $$\partial_{\tau} \varphi= \partial_{T} \overline{\varphi} (\partial_{\tau} T),\qquad \frac{\partial \varphi}{\partial x^{i}} = \frac{\partial \overline{\varphi}}{\partial Y^{i}} + \partial_{T} \overline{\varphi} \partial_{i} T. \label{conv2}$$ Equations (\[conv1\]) and (\[conv2\]) reproduce the standard results of Refs. [@salope; @shell]. Evolution of the gauge-invariant variables ------------------------------------------ According to the momentum constraint of Eq. (\[T8\]) the combination $K u_{i}$ appearing in Eq. (\[defR\]) is expressible in terms of the gradients of the extrinsic curvature as: $$K u_{i} = \frac{1}{3 \ell_{P}^2 ( \rho_{t}+ P_{eff})} \biggl[ \partial_{i} K^2 - 3K \nabla_{k} \overline{K}_{i}^{k}\biggr]. \label{A1}$$ In Eq. (\[A1\]) and in the forthcoming discussion we shall keep the dependence on the traceless part of the extrinsic curvature just to keep track of the difference between ${\mathcal R}_{i}$ and $\zeta_{i}$. From Eq. (\[T7\]) and from the trace of Eq. (\[T9\]) we obtain the following pair of equations: $$2 \ell_{P}^2 \rho_{t} = K^2 - \mathrm{Tr}K^2,\qquad 3 N \ell_{P}^2 (\rho_{t} + P_{eff}) = 2 \partial_{\tau}K - 3 N \mathrm{Tr}K^2 + N K^2, \label{A2}$$ implying that Eq. (\[A1\]) can be finally expressed as: $$K u_{i} = \frac{\partial_{i} \rho_{t}}{ \rho_{t} + P_{eff}} + {\mathcal G}_{i},\qquad {\mathcal G}_{i}= \frac{1}{6 \ell_{P}^2( \rho_{t} + P_{eff})} \biggl\{ 3\partial_{i}[ \mathrm{Tr} \overline{K}^2] - 6 K\, \nabla_{k} \overline{K}^{k}_{i} \biggr\}. \label{A3}$$ After inserting Eq. (\[A3\]) into Eq. (\[defR\]), the partial time derivative of ${\mathcal R}_{i}$ becomes: $$\partial_{\tau} {\mathcal R}_{i} = - \frac{1}{3} \partial_{i}( N K) + \frac{1}{3} \partial_{\tau} \biggl(\frac{\partial_{i} \rho_{t} }{\rho_{t} + P_{eff}}\biggr) + \frac{\partial_{\tau} {\mathcal G}_{i}}{3}. \label{A3a}$$ To leading order in the spatial gradients, the same kind of evolution equation reported in Eq. (\[A3a\]) is derivable for $\zeta_{i}$ starting directly from the definition Eq. (\[defZ\]) and using Eq. (\[A2\]). The leading terms of the evolution equation will be the same and the rationale for this occurrence is that $\zeta_{i}$ and ${\mathcal R}_{i}$ differ by terms that are of higher order in the gradient expansion, i.e. $6 \ell_{P}^2 \,(\zeta_{i} - {\mathcal R}_{i}) = - [ 3 \partial_{i}( \mathrm{Tr}\overline{K}^2) - 6 K \nabla_{k} \overline{K}^{k}_{i}]$. Inserting now Eq. (\[EN\]) into the first term at the right hand side of Eq. (\[A3a\]) we arrive at the following result: $$\partial_{\tau} {\mathcal R}_{i} = \frac{1}{3} \partial_{\tau} \biggl(\frac{\partial_{i}\rho_{t}}{\rho_{t} + P_{eff}}\biggr) - \frac{1}{3} \partial_{i} \biggl(\frac{\partial_{\tau}\rho_{t}}{\rho_{t} + P_{eff}}\biggr) + \frac{\partial_{\tau} {\mathcal G}_{i}}{3}. \label{A4}$$ The third term at the right hand side of Eq. (\[A4\]) will now be dropped since it is of higher order in the gradients. Equation (\[A4\]) can be expressed in a physically more significant form by separating the viscous contributions from the conventional non-adiabatic terms that normally appear even in the absence of irreversible contributions: $$\partial_{\tau} {\mathcal R}_{i} = {\mathcal S}_{nad}(\tau,\vec{x}) + {\mathcal S}_{viscous}(\tau,\vec{x}), \label{A4a}$$ where the two source terms are given, respectively, by $$\begin{aligned} {\mathcal S}_{nad}(\tau, \vec{x}) &=& \frac{ K N}{ 3 (\rho_{t} + P_{eff})} \bigl( \partial_{i} p_{t} - c_{st}^2 \partial_{i} \rho_{t}\bigr), \label{A4b}\\ {\mathcal S}_{viscous}(\tau, \vec{x}) &=& \frac{K}{3 ( \rho_{t} + P_{eff})^2} \bigl[ (\partial_{\tau} \rho_{t}) \partial_{i} \xi - \partial_{i} \rho_{t} (\partial_{\tau} \xi)\bigr] \nonumber\\ &+& \frac{\xi}{3 (\rho_{t} + P_{eff})^2} \bigl[(\partial_{\tau} \rho_{t})\partial_{i} K - (\partial_{i} \rho_{t})\partial_{\tau} K \bigr]. \label{A4c}\end{aligned}$$ As already mentioned, the total sound speed is $c_{st}^2 = \partial_{\tau} p_{t}/\partial_{\tau}\rho_{t}$. Let us consider first ${\mathcal S}_{nad}(\tau,\vec{x})$ and show that it is nothing but the standard adiabatic contribution. Broadly speaking the barotropic index is a space-time function, i.e. $w(\tau,\vec{x}) = p_{t}/\rho_{t}$ and Eq. (\[A4b\]) implies: $${\mathcal S}_{nad}(\tau, \vec{x}) = \frac{K N \rho_{t}}{3 ( \rho_{t} + P_{eff})} \partial_{i} w - \frac{\rho_{t} \partial_{i} \rho_{t}}{3 ( \rho_{t} + P_{eff})^2} \partial_{\tau} w. \label{A4d}$$ In linear theory when $w$ is a space-time constant the sound speed equals $\sqrt{w}$ and the non-adiabatic contribution is absent. In this situation ${\mathcal S}_{nad}(\tau, \vec{x}) \to 0$, as expected. According to Eqs. (\[A4a\]), (\[A4b\]) and (\[A4c\]) the relativistic viscous fluids lead to a source term implying that ${\mathcal R}_{i}$ is not constant in general terms. This may happen in various situations where, for instance, the viscosity coefficients depend in time and in space. For instance, across the matter-radiation transition the shear viscosity coefficient $\eta$ determines the optical depth, the Silk damping scale and, ultimately, the shape of the visibility function [@peebles; @visibility]. The scaling properties of $\eta$ and $\xi$ can be expressed $\xi/\eta \simeq \biggl( \frac{1}{3}- c_{\mathrm{st}}^2 \biggr)^{2 q}$ for $q\geq 1$. Across the matter-radiation transition the sound speed of the plasma interpolates between $1/\sqrt{3}$ and $0$. Adopting the viewpoint of linear theory (see section \[sec5\]) and separating the background from the fluctuations the sound speed can be computed as $c_{st}^2 = 4/[3 ( 4 + 3 \alpha)]$ where $\alpha= a/a_{eq}$ denotes the scale factor normalized at equality. This dependence in $\xi$ implies the generation of non-adiabatic modes (see, in particular, the last paper of [@nonad]). Conservation of curvature perturbations --------------------------------------- If $c_{st}^2 \neq w$ (or if $w$ is a space-time function) the curvature perturbations are non conserved even in the limit $\xi\to 0$, and this is nothing but the standard situation of the conventional non-adiabatic modes. To exclude all the potential sources that could make ${\mathcal R}_{i}$ time dependent besides the ones we ought to investigate specifically, namely the relativistic viscous contributions we shall posit that $w$ is constant and that $ c_{st} = \sqrt{w}$. This choice implies, according to Eq. (\[A4b\]), that ${\mathcal S}_{nad}=0$. The only contribution remaining at the right hand side of Eq. (\[A4a\]) is the one coming from ${\mathcal S}_{viscous}$. Furthermore, by focussing on Eq. (\[A4c\]) we see that the second term (proportional to $\xi$) can be rewritten as $$\frac{\xi}{3 (\rho_{t} + P_{eff})^2} \bigl[(\partial_{\tau} \rho_{t})\partial_{i} K - (\partial_{i} \rho_{t})\partial_{\tau} K \bigr] = \frac{\xi N}{6 (\rho_{t} + P_{eff})} \partial_{i}\bigl( K^2 - 3 \ell_{P}^2 \rho_{t}), \label{van1}$$ but the term at the right hand side vanishes because of the first of Eq. (\[A2\]); in fact, $\mathrm{Tr}K^2 = K^2/9 + {\mathrm Tr} \overline{K}^2$ and ${\mathrm Tr}\overline{K}^2$ is of higher order being proportional to the square of the total anisotropic stress. We stress that the result of Eq. (\[van1\]) holds non-perturbatively; it does not assume a separation between the background space-time and its perturbative fluctuations. Equation (\[A4a\]) becomes then: $$\partial_{\tau} {\mathcal R}_{i} = \frac{K}{3 ( \rho_{t} + P_{eff})^2} \bigl[ (\partial_{\tau} \rho_{t}) \partial_{i} \xi - \partial_{i} \rho_{t} (\partial_{\tau} \xi)\bigr]. \label{A5}$$ If the source term in Eq. (\[A5\]) vanishes the curvature inhomogeneities will be conserved and the equations of motion will enjoy a further symmetry since ${\mathcal R}_{i}(\tau,\vec{x})$ can be shifted by a a term constant in time (but not in space). The bulk viscosity $\xi$ can depend, in principle, on five quantities, namely $K$, $\mathrm{Tr}K^2$, $\rho_{t}$, $p_{t}$ and $n_{t}$. The dependence on $p_{t}$ can be traded for $\rho_{t}$ since $p_{t} = w \rho_{t}$ with constant $w$. Since the traceless part of the extrinsic curvature is of higher order we can also drop the dependence on $\mathrm{Tr}K^2$ that coincides, to leading order, with $K^2/9$. Equation (\[A2\]) can be finally used to relate $K^2$ and $\rho_{t}$. To lowest order in the gradient expansion we have therefore only two qualitatively different cases: $\xi = \xi(\rho_{t})$ and $\xi=\xi(\rho_{t},n_{t})$; the case $\xi=\xi(n_{t})$ is indeed the same as the one where $\xi(\rho_{t},n_{t})$. When $\xi = \xi(\rho_{t})$ Eq. (\[A5\]) implies that $\partial_{\tau} {\mathcal R}_{i} =0$: in this case the two terms at the right hand side simplify because $\partial_{i} \xi = (\partial \xi/\partial \rho_{t})\partial_{i} \rho_{t}$ and $\partial_{\tau}\xi = (\partial \xi/\partial \rho_{t})\partial_{\tau} \rho_{t}$. Putting together the results obtained so far, we can therefore say that Eq. (\[A4\]) is invariant for ${\mathcal R}_{i}(\tau, \vec{x}) \to {\mathcal R}_{i}(\tau, \vec{x}) + {\mathcal Q}(\vec{x})$ provided $\xi$ is either a space-time constant or a function of the total energy density. The requirements of the previous paragraph correspond to the situation where Eq. (\[A2\]) admit a fully inhomogeneous solution whose homogeneous limit is of quasi-de Sitter type. Let us therefore show explicitly that this is indeed the case in the simplest situation where $\xi$ is a space-time constant. Equations (\[A2\]) implies the following decoupled equation[^10] for $K$: $$\dot{K} - \frac{w+1}{2} K^2 = \frac{3}{2} \ell_{P}^2 \xi K, \label{pos5}$$ The general solution of Eq. (\[pos5\]) is: $$K(t,\vec{x}) = \frac{K_{0}(\vec{x}) \, e^{K_{\xi} [t - t_{0}(\vec{x})]/2}}{ (w+1) K_{0}(\vec{x}) \biggl[1 - e^{K_{\xi}[t - t_{0}(\vec{x})]/2}\biggr] + K_{\xi} }, \qquad K_{\xi} = 3 \ell_{P}^2 \xi = \frac{1}{t_{\xi}}, \label{pos6}$$ where $K_{0}(\vec{x})$ constant in time but not in space and $K_{\xi}$ is just a parameter of the solution. For $K_{\xi} [t - t_{0}(\vec{x})] \ll 1$ we have that $K(t,\vec{x})$ is singular while in the opposite limit it goes to a negative constant. In the homogenous limit $t_{0}$ is constant also in space. Equations (\[pos5\]) and (\[pos6\]) show that when $\xi$ is a space-time constant we can derive an inhomogeneous solution whose homogeneous limit interpolates between a perfect fluid solutions and a quasi-de Sitter solution. These solutions are the inhomogeneous counterpart of various quasi-de Sitter solutions derivable in the fully homogenous limit [@BV]. The generalization of Eq. (\[pos5\]) to the case when $\xi= \xi(\rho_{t})$ is straightforward since the dependence on $\rho_{t}$ can be eliminated through Eq. (\[A2\]). Some cases where $\xi$ has a power-law dependence on $\rho_{t}$ are even analytically solvable. We can therefore conclude that the large-scale curvature inhomogeneities are non-perturbatively conserved in numerous cases where the non-perturbative solution, in its fully homogeneous limit, admits a set of quasi-de Sitter backgrounds. Non-conservation of curvature perturbations ------------------------------------------- Let us finally consider the case $\xi= \xi(\rho_{t},n_{t})$ where Eq. (\[A5\]) becomes: $$\partial_{\tau} {\mathcal R}_{i} = \frac{K}{3 (\rho_{t} + P_{eff})^2} \biggl(\frac{\partial \xi}{\partial n_{t}}\biggr)\biggl[ \partial_{\tau} \rho_{t} \partial_{i} n_{t} - \partial_{i} \rho_{t} \partial_{\tau} n_{t}\biggr]. \label{A6}$$ The term at the right hand side does not vanish, in general. To lowest order in the gradient expansion the diffusion current $\nu^{\alpha}$ does not contribute to the evolution of $n_{t}$. Equation (\[A6\]) can then be rewritten by using Eqs. (\[EN\]) and (\[dens\]). The result is: $$\partial_{\tau} {\mathcal R}_{i} = \frac{K^2 N}{3 (\rho_{t} + P_{eff})} n_{t} \biggl(\frac{\partial \xi}{\partial n_{t}}\biggr)\biggl[ \frac{\partial_{i} n_{t}}{n_{t}} - \frac{\partial_{i} \rho_{t}}{\rho_{t} + P_{eff}}\biggr] \label{A6a}$$ Similar effects are expected when electromagnetic fields or scalar fields are present together with a dissipative fluid (see, for instance, the first two papers of Ref. [@sc1] for the case of scalar fields and the third paper of Ref. [@sc1] for the case of electromagnetic fields). The extension of the present considerations to a multicomponent viscous system is beyond the scope of this paper but it is conceptually feasible. In summary, the viscous fluids do not necessarily jeopardize the large-scale conservation of the curvature inhomogeneities at least as long as $\xi$ is a function of the total energy density or of the trace of the extrinsic curvature. Conversely the large-scale conservation of the non-linear curvature perturbations is invalidated whenever $\xi= \xi(\rho_{t}, n)$ \[or when $\xi= \xi(K, n)$\]. These observations have a counterpart in linear theory which will be discussed in the following section. Back to linear theory {#sec5} ===================== Scalar modes in linear theory ----------------------------- In linear theory variables ${\mathcal R}_{i}$ and $\zeta_{i}$ defined in Eqs. (\[defR\]) and (\[defZ\]) have a well defined limit. Let us choose conformally Newtonian frame where the gauge freedom is removed and the coordinate system completely fixed $$N^2(\tau,\vec{x}) = a^2(\tau) [ 1 + 2 \phi(\tau,\vec{x})],\qquad \gamma_{i j}(\tau,\vec{x}) = a^2(\tau)[ 1 - 2 \psi(\tau,\vec{x})]\delta_{ij}, \label{lim0}$$ and expand Eqs. (\[defR\]) and (\[defZ\]) by assuming that $\phi$ and $\psi$ are both smaller than one. The result of this limit is given by: $${\mathcal R}_{i} \to \partial_{i} {\mathcal R},\qquad \qquad \zeta_{i} \to \partial_{i} \zeta, \label{lim1}$$ implying that ${\mathcal R}_{i}$ and $\zeta_{i}$ are, respectively, the spatial gradients of the curvature perturbation on comoving orthogonal hypersurfaces and of the density contrast on uniform curvature hypersurfaces, i.e.[^11] $${\mathcal R}= - \psi - \frac{{\mathcal H} ( \psi^{\prime} + {\mathcal H} \phi)}{{\mathcal H}^2 - {\mathcal H}^{\prime}}, \qquad \zeta= - \psi + \frac{\delta\rho_{t}}{3 (\rho_{t} + P_{eff})}. \label{lim1a}$$ In this section $\rho_{t}$ and $P_{eff}$ denote the background values of the corresponding quantity while $\delta\rho$ is the first-order fluctuation of the energy density and so on and so forth. In other words the conventions will be such that $$\rho(\tau, \vec{x}) = \rho_{t}(\tau) + \delta \rho_{t}(\tau,\vec{x}), \qquad \xi(\tau,\vec{x}) = \xi(\tau) +\delta\xi(\tau,\vec{x}). \label{lim2}$$ where $\delta$ will denote the first-order fluctuation of the corresponding quantity. The same conventions will be employed for all the other variables involved in the discussion. The linear order form of the evolution equation for ${\mathcal R}$ and $\zeta$ can be derived in perturbation theory and then compared with limit of Eq. (\[A4\]). The result is is given by the following pair of equations where we have included, for the sake of comparison, also the terms that are of higher order in the gradient expansion but are consistent with the linearized approximation: $$\begin{aligned} {\mathcal R}^{\prime} &=& \frac{3 {\mathcal H} }{a (\rho_{t} + P_{eff})} \xi^{\prime} ({\mathcal R} + \psi)- \frac{{\mathcal H}}{\rho_{t} + P_{eff}} \delta p_{nad} \nonumber\\ &+& \frac{3 {\mathcal H}^2}{a( \rho_{t} + P_{eff})} \delta \xi + \frac{\xi {\mathcal H} }{a(\rho_{t} + P_{eff})}\theta_{t} - \frac{3{\mathcal H} c_{st}^2 }{2\ell_{P}^2 (\rho + P_{eff})} \nabla^2 \psi, \label{RPR}\\ \zeta^{\prime} &=& \frac{3 {\mathcal H} }{a(\rho_{t} + P_{eff})} \xi^{\prime} (\zeta + \psi)- \frac{{\mathcal H}}{\rho_{t}+ P_{eff}} \delta p_{nad} \nonumber\\ &+& \frac{3 {\mathcal H}^2}{a(\rho_{t} + P_{eff})} \delta \xi + \theta_{t} \biggl[\frac{{\mathcal H} \xi}{a(\rho_{t} + P_{eff})} - \frac{1}{3}\biggr] \nonumber\\ &-& \frac{{\mathcal H}}{ 2 \ell_{P}^2 ( \rho_{t} + P_{eff})} \nabla^2 (\phi - \psi) - \frac{\xi}{a(\rho_{t} + P_{eff})}\nabla^2 \psi. \label{zetapr}\end{aligned}$$ In Eqs. (\[RPR\]) and (\[zetapr\]) the term $\theta_{t}$ denotes the three-divergence of the total velocity field. Equations (\[RPR\]) and (\[zetapr\]) have been derived directly in the linear theory and they reproduce the results obtainable by linearizing . Consider then, for the sake of comparison, Eqs. (\[RPR\]) and (\[A5\]) in the limit $\delta p_{nad} \to 0$. From Eq. (\[RPR\]) we will have $${\mathcal R}^{\prime} = \frac{3 {\mathcal H} }{a (\rho_{t} + P_{eff})} \xi^{\prime} ({\mathcal R} + \psi) +\frac{3 {\mathcal H}^2}{a(\rho_{t} + P_{eff})}\, \delta \xi, \label{lim6}$$ where $\theta_{t}$ has been neglected since it is of higher order in the gradients thanks to the momentum constraint (implying $\theta_{t} = \nabla^2({\mathcal R} + \psi)/{\mathcal H}$). To see how things work in linear theory we can verify explicitly that the right hand side of Eq. (\[lim6\]) vanishes when $\xi= \xi(\rho_{t})$. From the definition of $\zeta$ in linear theory we have that the momentum constraint can be expressed as: $$\zeta = {\mathcal R} + \frac{\nabla^2\psi}{2 \ell_{P}^2 (\rho_{t} + P_{eff})}. \label{lim7}$$ Neglecting the gradients, Eq. (\[lim6\]) is expressible, in the case $\xi= \xi(\rho_{t})$, as: $${\mathcal R}^{\prime} = \frac{3 {\mathcal H}}{a(\rho_{t} + P_{eff})}\biggl(\frac{\partial \xi}{\partial\rho_{t}} \biggr) \biggl[ \rho_{t}^{\prime} ( \zeta + \psi) + {\mathcal H} \delta \rho_{t} \biggr] \label{lim8}$$ where we used that $\delta \xi = (\partial \xi/\partial \rho_{t}) \delta \rho_{t}$. But now thanks to the definition of $\zeta$ we have that $(\zeta+ \psi) = 3 (\rho_{t} + P_{eff}) \delta \rho_{t}$. Thus Eq. (\[lim8\]) implies ${\mathcal R}^{\prime} =0$ since, by covariant conservation of the background energy-momentum tensor, $\rho_{t}^{\prime} = - 3 {\mathcal H} ( \rho_{t}+ P_{eff})$. This result does not hold if $\xi = \xi(\rho_{t},n_{t})$ so that, in general, the curvature perturbations induced by relativistic viscous fluids are not conserved. This derivation is the linear order counterpart of the discussion presented after Eq. (\[A5\]). Standard gauge-invariant variables ---------------------------------- For infinitesimal coordinate transformations we have that $\phi\to \widetilde{\phi} = \phi - {\mathcal H} \epsilon_{0} - \epsilon_{0}'$ and that $\psi \to \widetilde{\psi} = \psi + {\mathcal H} \epsilon_{0}$. The bulk viscosity coefficient transforms instead as: $$\delta\xi \to \widetilde{\delta\xi} = \delta \xi - \xi^{\prime} \epsilon_{0}.$$ Including the gradients in the appropriate entries of the perturbed metric (i.e. $\delta g_{ij} = 2 a^2( \psi \delta_{ij} - \partial_{i} \partial_{j} E)$ and $\delta g_{i0} = - a^2 \partial_{i} B$) and recalling that they transform as $B \to \widetilde{B} = B + \epsilon_{0} - \epsilon'$ and as $E \to \widetilde{E} = E - \epsilon$ the gauge-invariant fluctuations of the bulk viscosity fluctuations are: $$\Xi = \delta \xi +\xi^{\prime} ( B - E^{\prime}),$$ while the Bardeen potentials are, as usual, $\Phi = \phi + ( B - E')' + {\cal H} ( B - E')$ and $\Psi= \psi - {\cal H} ( B - E')$. In terms of the explicitly gauge-invariant fluctuations Eq. (\[RPR\]) becomes $$\begin{aligned} {\mathcal R}^{\prime} &=& \frac{3 {\mathcal H} }{a (\rho_{t} + P_{eff})} \xi^{\prime} ({\mathcal R} + \Psi)- \frac{{\mathcal H}}{\rho_{t} + P_{eff}} \delta p_{nad} \nonumber\\ &+& \frac{3 {\mathcal H}^2}{a( \rho + P_{eff})} \Xi + \frac{\xi {\mathcal H} }{a(\rho_{t} + P_{eff})}\Theta_{t} - \frac{3{\mathcal H} c_{st}^2 }{2\ell_{P}^2 (\rho + P_{eff})} \nabla^2 \Psi, \label{RPR2}\end{aligned}$$ where $\Theta_{t} = \theta_{t} + \nabla^2 E^{\prime}$ is the gauge-invariant expression of the total velocity field. Fluctuations of the expansion and $\delta {\mathcal N}$ formalism ----------------------------------------------------------------- We can also compute the total expansion rate with the aim of showing that the presence of the bulk viscosity affects the basis of the so-called $\delta {\mathcal N}$ formalism stipulating that $\zeta$ can be related to the scalar-field perturbations at the initial time computed in the uniform curvature gauge once we know the derivatives of the number of efolds with respect to the initial values of the unperturbed scalar fields and their derivatives. Let us therefore introduce the inhomogeneous generalization of the total number of efolds $${\mathcal N}(\vec{x},\tau_{}, \tau_{f}) = \frac{1}{3} \int_{\tau_{}}^{\tau_{f}} \nabla_{\alpha} u^{\alpha} \, N \, d\tau, \label{dN1}$$ where $\tau_{}$ denotes the initial time;all the quantities of the integrand are space-time dependent. We recall that in linear theory we can define $$\nabla_{\mu} u^{\mu} = (\nabla_{\mu} u^{\mu})^{(0)} + \delta^{(1)}(\nabla_{\mu} u^{\mu}) + \delta^{(2)}(\nabla_{\mu} u^{\mu}) +\,... \label{dN1a}$$ Without committing ourselves to a specific gauge choice, from Eq. (\[dN1a\]) and from the fluctuations of the lapse function, Eq. (\[dN1\]) becomes: $$\begin{aligned} {\mathcal N}(\vec{x},\tau_{}, \tau_{f}) &=& \overline{{\mathcal N}}(\tau_{},\tau_{f}) + \frac{1}{3} \int_{\tau_{}}^{\tau_{f}} (\theta_{t} + \nabla^2 E') \, d\tau - \int_{\tau_{}}^{\tau_{f}} \psi^{\prime} \, d\tau, \nonumber\\ \overline{{\mathcal N}}(\tau_{},\tau_{f}) &=& \int_{\tau_{}}^{\tau_{f}} \,{\mathcal H} d \tau = \int_{t_{}}^{t_{f}} H d t, \label{dN2}\end{aligned}$$ where ${\mathcal H} = H a$ and $dt = a d\tau$. From the first-order fluctuation of the covariant conservation of the total energy-momentum tensor we can express $\psi^{\prime}$ as $$\begin{aligned} \psi^{\prime} &=& \frac{1}{3} [ \theta_{t} + \nabla^2 E^{\prime}] + \frac{\delta \rho_{t}^{\prime} + 3 {\mathcal H} ( \delta \rho_{t} + \delta p_{t})}{3 ( \rho_{t} + P_{eff})} - \frac{3 {\mathcal H}^2 }{a ( \rho_{t} + P_{eff})} \delta\xi \nonumber\\ &-& \frac{{\mathcal H} \xi}{a(\rho + P_{eff})}[ \theta_{t} + \nabla^2 E^{\prime} - 3 ( \psi^{\prime} + {\mathcal H} \phi)]. \label{dN3}\end{aligned}$$ Inserting Eq. (\[dN3\]) into Eq. (\[dN2\]) we can easily obtain $$\begin{aligned} {\mathcal N}(\vec{x},\tau_{}, \tau_{f}) &=& \overline{{\mathcal N}}(\tau_{},\tau_{f}) + \int_{\tau_{}}^{\tau_{f}} \biggl\{ \frac{3 {\mathcal H}}{a (\rho_{t} + P_{eff})} [ {\mathcal H} \delta \xi + \xi^{\prime} ( \psi + \zeta)] - (\zeta^{\prime} + \psi^{\prime}) \nonumber\\ &+& \frac{\xi}{a (\rho_{t} + P_{eff})} \nabla^2 {\mathcal R} - \frac{{\mathcal H}}{(\rho_{t} + P_{eff})} \delta p_{nad}\biggr\}, \label{dN4}\end{aligned}$$ where, according to Eq. (\[lim1a\]), we used that $\delta \rho = 3 (\zeta + \psi) (\rho + P_{eff})$. When $\delta p_{nad}\to 0$ and $\xi= \xi(\rho_{t})$ we have: $${\mathcal N}(\vec{x},\tau_{}, \tau_{f}) = \overline{{\mathcal N}}(\tau_{},\tau_{f}) - \int_{\tau_{}}^{\tau_{f}} \biggl( \zeta^{\prime} + \psi^{\prime} \biggr) d\tau, \label{dN5}$$ implying $${\mathcal N}(\vec{x},\tau_{}, \tau_{f}) = \overline{{\mathcal N}}(\tau_{},\tau_{f})+ [\zeta(\vec{x},\tau_{})+ \psi(\vec{x},\tau_{})] - [\zeta(\vec{x},\tau_{f})+ \psi(\vec{x},\tau_{f})] . \label{dN6}$$ Let us then evaluate $(\zeta + \psi)_{\tau_{f}}$ in the uniform density gauge (i.e. $\delta \rho_{t} =0$) while $(\zeta + \psi)_{\tau_{}}$ is evaluated in the uniform curvature gauge (i.e. $\psi =0$). This choice implies that: $$[\zeta(\vec{x},\tau_{f})+ \psi(\vec{x},\tau_{f}) ]\to 0, \qquad [\zeta(\vec{x},\tau_{})+ \psi(\vec{x},\tau_{})] = \frac{\delta\rho_{t}}{3 ( \rho_{t} + P_{eff})}. \label{dN7}$$ Thus, from Eqs. (\[dN6\]) and (\[dN7\]) we have that $$\delta {\mathcal N} = {\mathcal N}(\vec{x},\tau_{}, \tau_{f}) - \overline{{\mathcal N}}(\tau_{},\tau_{f})= \frac{\delta\rho_{t}}{3 ( \rho_{t} + P_{eff})}. \label{dN8}$$ The same formulas derived above hold in the single scalar field case (at least up to some point) by setting $\xi =0$ and by recalling that, in the single scalar field case, ${\mathcal R} = - \psi - ({\mathcal H}/\varphi^{\prime}) \delta \varphi$ where $\delta\varphi$ denotes the scalar field fluctuation. Thus, in the single scalar field case, we will have $$[\zeta(\vec{x},\tau_{})+ \psi(\vec{x},\tau_{})] =\biggl[ {\mathcal R}_{} + \psi_{} + \frac{\nabla^2 \psi_{}}{ 2 \ell_{P}^2 (\rho_{\varphi} + p_{\varphi}) }\biggr] \to - \frac{{\mathcal H}}{\varphi^{\prime}} \delta \varphi. \label{dN9}$$ In analogy with Eq. (\[dN8\]), Eq. (\[dN9\]) can be expressed as $\delta {\mathcal N} \simeq - (\partial {\mathcal N}/\partial \varphi)_{} \delta \varphi(\vec{x},\tau_{})$, which is the standard result of the single scalar field case. We can therefore conclude that the presence of bulk viscous stresses affects the $\delta {\mathcal N}$ only if $\xi = \xi(\rho_{t},n_{t})$. Conversely, if $\xi = \xi(\rho_{t})$ both ${\mathcal R}$ and $\zeta$ are constant and $\delta {\mathcal N}$ only depends on the values of $\zeta + \psi$ between the initial and final times characterizing the integrated (total) expansion. Concluding remarks {#sec6} ================== The hypothesis of reversibility of the plasma has been dropped by allowing for a relativistic dissipative fluid as the dominant source of curvature inhomogeneity in the non-perturbative regime. In the general situation the non-linear evolution does not preserve the curvature inhomogeneities. However if the bulk viscosity is either a space-time constant or if it depends solely on the total energy density of the plasma, then a further symmetry prevents the growth of curvature inhomogeneities. To lowest order in the gradient expansion the dependence on the energy density can be traded for a dependence on the trace of the extrinsic curvature. In all these cases the curvature inhomogeneities are non-perturbatively conserved. The validity of the linear theory has not been posited as a necessary requirement of the derivation. Nonetheless the non-perturbative results have a perturbative counterpart describable in terms of the conventional gauge-invariant variables namely the curvature perturbations on comoving orthogonal hypersurfaces and the density contrast on uniform curvature hypersurfaces. The only contribution to the non-perturbative evolution of curvature inhomogeneities comes from the bulk viscous stresses while the shear viscosity affects the evolution of the traceless part of the extrinsic curvature and it appears to higher order in the gradient expansion. The only contribution to the inhomogeneous deceleration parameter comes from the bulk viscosity. Beyond linear theory the large-scale acceleration can only be driven by the bulk pressure and no acceleration can take place thanks to the shear viscosity. The results of this investigation confirm that the non-linear conservation of curvature perturbations is non a generic phenomenon. However when the curvature inhomogeneities are conserved the evolution equations of the extrinsic curvature can be solved in fully inhomogeneous terms. These solutions correspond, in the homogeneous and isotropic limit, to a quasi-de Sitter stage of expansion. One could therefore speculate that the non-perturbative constancy of the curvature inhomogeneities pins down a class of viscous coefficients and, ultimately, a specific set of physical properties of the geometry. It is amusing that this logic is opposite to the one commonly pursued in linear theory where a quasi-de Sitter stage of expansion is postulated to insure, somehow, the perturbative conservation of the linearized fluctuations. [99]{} D. N. Spergel [et al.]{}, Astrophys. J. Suppl. [148]{}, 175 (2003); D. N. Spergel [et al.]{}, [ibid.]{} [170]{}, 377 (2007); L. Page [et al.]{}, [ibid.]{} [170]{}, 335 (2007). B. Gold [et al.]{}, Astrophys. J. Suppl. [192]{}, 15 (2011); D. Larson, [et al.]{}, [ibid.]{} [192]{}, 16 (2011); C. L. Bennett [et al.]{}, [ibid.]{} [192]{}, 17 (2011); G. Hinshaw [et al.]{}, [ibid.]{} [208]{} 19 (2013); C. L. Bennett [et al.]{}, [ibid.]{} [208]{} 20 (2013). P. A. R. Ade [et al.]{} \[Planck Collaboration\], Astron. Astrophys. [571]{}, A22 (2014); [ibid.]{} [571]{}, A16 (2014); P. A. R. Ade [et al.]{} \[Planck Collaboration\], arXiv:1502.02114 \[astro-ph.CO\]. P. J. E. Peebles and J. T. Yu, Astrophys. J. [162]{}, 815 (1970); G. Efstathiou and J. R. Bond, Mon. Not. Roy. Astron. Soc. [218]{}, no. 1, 103 (1986); R. y. Cen, J. P. Ostriker and P. J. E. Peebles, Astrophys. J. [415]{}, 423 (1993). E. M. Lifshitz and I. M. Khalatnikov, Sov. Phys. JETP [12]{}, 108 (1960); Sov. Phys. JETP [12]{}, 558 (1961); Phys. Rev. Lett. [24]{}, 76 (1970). V. A. Belinskii and I. M. Khalatnikov, Sov. Phys. JETP [30]{}, 1174 (1970); Sov. Phys. JETP [36]{}, 591 (1973). L. D. Landau and E. M. Lifshitz, [The Classical Theory of Fields]{}, (Pergamon Press, New York, 1971). A. A. Starobinsky, JETP Lett. [37]{}, 66 (1983); R. M. Wald, Phys. Rev. D [28]{}, 2118 (1983). D. S. Salopek and J. R. Bond, Phys. Rev. D [42]{}, 3936 (1990); D. S. Salopek and J. M. Stewart, Class. Quant. Grav. [9]{}, 1943 (1992); J. Parry, D. S. Salopek and J. M. Stewart, Phys. Rev. D [49]{}, 2872 (1994). J. Bardeen, Phys. Rev. D [22]{}, 1882 (1980); J. Bardeen, P. Steinhardt, and M. Turner, Phys. Rev. D [28]{}, 679 (1983); J. A. Frieman and M. S. Turner, Phys. Rev. D [30]{}, 265 (1984); D. S. Salopek, J. R. Bond and J. M. Bardeen, Phys. Rev. D [40]{}, 1753 (1989). K. Tomita, Prog. Theor. Phys. [67]{}, 1076 (1982); Phys. Rev. D [48]{}, 5634 (1993); N. Deruelle and K. Tomita, Phys. Rev. D [50]{}, 7216 (1994); N. Deruelle and D. Goldwirth, Phys. Rev. D [51]{}, 1563 (1995); M. Giovannini, JCAP [0509]{}, 009 (2005); Phys. Lett. B [746]{}, 159 (2015); arXiv:1504.07624 \[gr-qc\]. F. Hoyle and J. V. Narlikar, Proc. R. Soc. A, [273]{}, 1 (1963); F. Hoyle, G.Burbidge, and J. V. Narlikar, [A different approach to cosmology]{}, (Cambridge University Press, Cambridge 2000). Ya. Zeldovich, Sov. Phys. JETP [21]{}, 656 (1965); Sov. Astron. [13]{}, 608 (1970); Ya. Zeldovich and I. Novikov, [The Structure and Evolution of the Universe]{}, (Chicago University Press, Chicaggo, 1971), Vol.2. C. W. Misner, Astrophys. J. [151]{}, 431 (1968); Phys. Rev. Lett. [22]{}, 1071 (1969); M. J. Rees, Phys. Rev. Lett. [28]{}, 1669 (1972). A.H. Guth, Phys. Rev. D [23]{} 347 (1981); A.D. Linde, Phys. Lett. B [108]{}, 389 (1982); A. Albrecht, P.J. Steinhardt, Phys. Rev. Lett. [48]{}, 1220 (1982). S. Weinberg, [Cosmology]{} (Oxford University Press, Oxford 2008). J. D. Barrow and O. Gron, Phys. Lett. B [182]{}, 25 (1986); J. D. Barrow, Phys. Lett. B [187]{}, 12 (1987); Phys. Rev. D [55]{}, 7451 (1997); M. Giovannini, Phys. Rev. D [89]{}, 063512 (2014). N. Afshordi and R. H. Brandenberger, Phys. Rev. D [63]{}, 123505 (2001); G. I. Rigopoulos and E. P. S. Shellard, Phys. Rev. D [68]{}, 123518 (2003); JCAP [0510]{}, 006 (2005); M. Shibata and H. Asada, Prog. Theor. Phys. [94]{}, 11 (1995). H. Noh and J. c. Hwang, Phys. Rev. D [69]{}, 104011 (2004); J. c. Hwang and H. Noh, Phys. Rev. D [72]{}, 044012 (2005); J. C. Hwang, H. Noh and D. Puetzfeld, JCAP [0803]{}, 010 (2008). L. Tisza, Phys. Rev. [61]{}, 531 (1942); S. Weinberg, Astrophys. J. [168]{}, 175 (1971); M. Giovannini,Phys. Lett. B [622]{}, 349 (2005); Class. Quant. Grav. [22]{}, 5243 (2005). R. Arnowitt and S. Deser Phys. Rev. [113]{}, 745 (1959); R. Arnowitt, S. Deser, and C. W. Misner Phys. Rev. [117]{}, 1595 (1960). T. Baumgarte and S. L. Shapiro, [Numerical Relativity]{}, (Cambridge University Press, Cambridge 2010). S. Weinberg, [Gravitation and cosmology]{} (Wiley, New York, 1972). S. R. de Groot, V. A. van Leeuwen, and Ch. G. van Weert, [Relativistic Kinetic Theory]{} (North Holland, Amsterdam, 1980). W. Israel Ann. Phys. [100]{}, 310 (1976); W. Israel and J. M. Stewart Ann. Phys. [118]{}, 341 (1979). A. Muronga, Phys. Rev. C [69]{}, 034903 (2004); U. Heinz, H. Song, and A. K. Chaudhuri Phys. Rev. C [73]{} 034904 (2006); H. Song and U. Heinz Phys. Rev. C [78]{}, 024902 (2008); G. S. Denicol, T. Koide and D. H. Rischke Phys. Rev. Lett. [105]{}, 162501 (2010). K. Enqvist, H. Kurki-Suonio and J. Valiviita, Phys. Rev. D [62]{}, 103003 (2000); J. Valiviita and V. Muhonen, Phys. Rev. Lett. [91]{}, 131302 (2003); H. Kurki-Suonio, V. Muhonen and J. Valiviita, Phys. Rev. D [71]{}, 063005 (2005); M. Giovannini, Class. Quant. Grav. [23]{}, 4991 (2006); R. Keskitalo, H. Kurki-Suonio, V. Muhonen and J. Valiviita, JCAP [0709]{}, 008 (2007); J. Valiviita, M. Savelainen, M. Talvitie, H. Kurki-Suonio and S. Rusak, Astrophys. J. [753]{}, 151 (2012); M. Giovannini, Phys. Rev. D [91]{}, 061304 (2015). P. Naselsky and I. Novikov, Astrophys. J. [413]{}, 14 (1993); H. Jorgensen, E. Kotok, P. Naselsky, and I. Novikov, Astron. Astrophys. [294]{}, 639 (1995); U. Seljak, Astrophys. J. [435]{}, L87 (1994); C. P. Ma and E. Bertschinger, Astrophys. J. [455]{}, 7 (1995). R. Treclokas and G. Ellis, Commun. Math. Phys. [23]{}, 1 (1971); G. L. Murphy, Phys. Rev. D [8]{}, 4231 (1973); G. L. Murphy, Phys. Lett. A [62]{}, 75 (1977); V. A. Belinskii and I. M. Khalatnikov, JETP Lett. [21]{}, 99 (1975) \[Zh. Eksp. Teor. Fiz. Pis. Red. [21]{}, 223 (1975)\]; N. G. Turok, Phys. Rev. Lett. [60]{}, 549 (1988); J. Barrow, Nucl. Phys. B [310]{}, 743 (1988); B. Li and J. D. Barrow, Phys. Rev. D [79]{}, 103521 (2009). M. Bastero-Gil, A. Berera and R. O. Ramos, JCAP [1107]{}, 030 (2011); M. Bastero-Gil, A. Berera, R. Cerezo, R. O. Ramos and G. S. Vicente, JCAP [1211]{}, 042 (2012); M.Giovannini, Phys. Rev. D [88*]{}, 063536 (2013). [^1]: Electronic address: [email protected] [^2]: Note that the spatial indices are lowered and raised using $\gamma_{ij}(\tau,\vec{x})$. The Greek indices will take the values $(\mu,\, \nu )= 0,\, 1\, 2\, 3$. [^3]: We warn the reader that this identification will be followed throughout the paper. According to some, this notation may lead to potential ambiguities but we hope that, with this note, confusions will be avoided. In this paper $\nabla_{i}$ will denote the covariant derivative on the spatial slices and not the spatial component of a (four-dimensional) covariant derivative. [^4]: The partial derivatives with respect to $\tau$ and $x^{i}$ shall be denoted, respectively, by $\partial_{\tau}$ and by $\partial_{i}$. [^5]: In this paper we discuss the global evolution of large-scale curvature perturbations. We shall therefore deal with global quantities such as the total pressure, the total energy density of the system, the total viscosity of the fluid and so on. Of course the same analysis can be extended to the case where the fluid is composed by a number of fluids interacting among themselves as it happens, for instance, prior to photon decoupling. [^6]: The explicit form of Eq. (\[entro\]) has been obtained by trading the term $u_{\nu} \nabla_{\mu} {\mathcal T}^{\mu\nu}$ for $(\nabla_{\mu} u_{\nu}) {\mathcal T}^{\mu\nu}$ since, in the Landau frame, $\nabla_{\mu} ( u_{\nu} {\mathcal T}^{\mu\nu} ) = 0$. [^7]: The higher order terms in the gradient expansion can be computed by following iterative methods where the spatial geometry is reconstructed, order by order, starting from a seed metric that do not contain any spatial gradient [@salope; @tomita] but this is not our primary goal in this investigation. [^8]: In the homogeneous and isotropic limit, $ \gamma_{ij} = a^2(t) \delta_{ij}$, $K_{i}^{j} = - H \delta_{i}^{j}$ and, as expected, $q(t) \to - \ddot{a} a/ \dot{a}^2$. [^9]: In linear theory the density contrast on uniform curvature hypersurfaces is invariant under infinitesimal coordinate transformations. Since, by definition, it has the same value in different gauges it can be also interpreted as the curvature perturbation on the hypersurfaces where the energy density is unperturbed. These two physical interpretations are relevant when discussing the so-called $\delta {\mathcal N}$ formalism (see section \[sec5\]). [^10]: In Eq. (\[pos5\]) we have chosen the geodesic slicing with $N=1$; in this case $\tau=t$ where $t$ denotes the cosmic time coordinate and the overdot denotes a derivation with respect to $t$. [^11]: The prime will denote a derivation with respect to $\tau$ and ${\mathcal H} =a^{\prime}/a$.	{ "pile_set_name": "ArXiv" }
--- author: - 'Francesco Solera, Simone Calderara, and Rita Cucchiara, [^1]' bibliography: - 'IEEEabrv.bib' - 'main.bib' title: Socially Constrained Structural Learning for Groups Detection in Crowd --- [Solera : Socially Constrained Structural Learning for Groups Detection in Crowd]{} Conclusion ========== In this work, we pointed out the need to approach the task of detecting social groups in crowds from a learning perspective. Many existing methods rely on specifically tuned parameters that limit their applicability in real world scenarios. Our intuition is that there are crowds that preserve the same concept of social group, but in many cases this concept cannot be distilled from spatial consideration only. We thus defined a set of social-inspired and strongly motivated features able to capture and characterize different groups peculiarities. To learn a socially meaningful clustering rule to group pedestrians, we relied on the Structural SVM framework and designed a peculiar loss function able to account for singletons as well as for group errors. Even though the algorithm was originally designed to work with exact trajectories, we replicated the experiments on noisy tracklets extracted by a detector/tracker obtaining state-of-the-art results. Moreover, we proposed an online training version of the method, able to achieve superior generalization performances on crowds with variable density. We did note, however, that as we consider wider portions of the scene, the chance that many different densities groups coexist in different locations increases, leading to the necessity to learn more than one clustering rule per scene. To resolve this problem we plan, as future work, to learn a set of different distance measures and use latent variables to choose the most appropriate given a particular zone. Code and datasets are made publicly available[^2] in order to reproduce this paper results and allow the community to improve the proposed method. [Francesco Solera]{} obtained a master degree in computer engineering from the University of Modena and Reggio Emilia in 2013. He is now a PhD candidate within the ImageLab group in Modena, researching on applied machine learning and social computer vision. [Simone Calderara]{} received a computer engineering master degree in 2004 and a PhD degree in 2009 from the University of Modena and Reggio Emilia, where he is now an assistant professor within the Imagelab group. His current research interests include computer vision and machine learning applied to human behavior analysis, visual tracking in crowded scenarios and time series analysis for forensic applications. [Rita Cucchiara]{} received her master degree in electronic engineering and the PhD degree in computer engineering from the University of Bologna, Italy, in 1989 and 1992 respectively. Since 2005, she is a full professor at University of Modena and Reggio Emilia, Italy, where she heads the ImageLab group and the SOFTECH-ICT research center. Her research focuses on pattern recognition, computer vision and multimedia. [^1]: Authors are with the Department of Engineering Enzo Ferrari, University of Modena and Reggio Emilia, Italy e-mail: [email protected] [^2]: `http://imagelab.ing.unimore.it/group-detection`	{ "pile_set_name": "ArXiv" }
--- abstract: 'To explain Gamma Ray Bursts, Ruffini argued that the event horizon of a charged black hole is surrounded by a special region called, the Dyadosphere where electric field exceeds the critical value for $e^+$ $e^-$ pair production. In the present work, we construct a thin shell wormhole by performing a thought surgery between two dadospheres. Several physical properties of this thin shell wormhole have been analyzed.' author: - 'F.Rahaman$^$ , M.Kalam$^{\ddag}$ and K A Rahman$^$' title: - Thin shell wormhole due to dyadosphere of a charged black hole - 'Introduction: ' - '2. The Dyadosphere - a prelude:' - '3. Thought surgery and thin shell wormhole construction: ' - '4. Stability Analysis: ' - '5. Discussions: ' --- $ $ To theoretical support of recent experimental evidence of gamma ray bursts is an intriguing research area in modern astrophysics. Some peoples believe that collapses of massive stars could be responsible for these bursts. Recently, Ruffini and Collaborators \[1,2\] have proposed an alternative explanation of gamma ray bursts by introducing a new concept of dyadosphere of an electromagnetic black hole. They have claimed that the event horizon of charged black hole is encircled by a special region called dyadosphere where the electromagnetic field strength exceeds the well known Heisenberg - Euler critical value for the electron-positron pair creation $\epsilon_{crit} = \frac{m_e^2c^3}{\hbar e } $ ( $m_e$ and e are mass and charge of an electron respectively ). By considering the dyadosphere corresponding to Reissner-Nordström spacetime, Ruffini \[3\] and Preparata et al \[4\] have shown that the electron positron pair creation process occurs over the entire dyadosphere outside the Reissner-Nordström horizon. They have also given a measure of total energy of electron-positron pairs created within the dyadosphere. It is proved that in presence of strong electromagnetic field, the velocity of light propagation is affected by vacuum polarization states which lead to super luminal photon propagation \[ 5 \]. During the investigation of photon propagation around Reissner-Nordström black hole, Daniels and Shore \[6\] have shown that the super luminal photon propagation is possible due the effect of one loop vacuum polarization on photon propagation. So it is crucial to find the region where electric field exceeds its classical limit and vacuum fluctuations take place. Recently, Delorenci et al \[7\] have computed the correction for the Reissner-Nordström metric from the first contribution of the Euler-Heisenberg Lagrangian. In 1989, Visser \[8\] had performed a thought surgery between two Schwarzschild black holes and glued together in such a way that no event horizon is permitted to form. The resultant structure leads to a specific geometrical structure known as thin shell wormhole. Recently, several authors have constructed thin shell wormholes by surgically grafting of different black holes following Visser’s approach \[9 - 18\]. This approach is important as because the exotic matter required for the creation of wormhole structure is confined within the shell. Also, this novel approach gives a way of minimizing the usage of exotic matter to construct a wormhole. Our purpose with this paper is to present a new thin shell wormhole whose required exotic matter could be collected from cosmic mine i.e. from dyadosphere. According to Ruffini, the sources of gamma ray bursts are dyadospheres, so one can imagine, these sources could be used by an advanced civilization to construct and sustain a wormhole. We will discuss different characteristics of this thin shell wormhole namely, time evolution of the throat, stability, total amount of exotic matter. The paper is organized as follows : In section 2, the reader is reminded about dyadosphere proposed by Ruffini. In section 3, thin shell wormhole has been constructed following Visser’s techniques. The linearized stability analysis is studied in section 4. Section 5 is devoted to a brief discussion. Ruffini and collaborators have proposed that there exists a region outside the event horizon of a charged black hole, called dyadosphere. In this region, the electromagnetic field is greater than the Euler-Heisenberg critical value of electro-positron pair production. According to them, this newly designed region is responsible for the gamma ray bursts. The simplest charged balck hole is the Reissner-Nordström black hole which is described by the line element $$ds^2= f(r) dt^2 - \frac{dr^2}{ f(r)} - r^2 (d\theta^2 + \sin^2 \theta d\phi^2)$$ with $$f(r)= 1 - \frac{2GM}{c^2 r} + \frac{Q^2G}{c^4 r^2}$$ where, M and Q are mass and charge parameters. It is known that the electric field in the Reissner-Nordström geometry is given by $ \epsilon = \frac{Q}{r^2}$ and this is larger than $\epsilon_{crit}$ in the dyadosphere region. For the Reissner-Nordström black hole, the dyadosphere is defined by the radial interval $r_+ \leq r \leq r_{ds}$ where, $$r_+ = \frac{GM}{c^2 }\left( 1 + \sqrt{ 1 - \frac{Q^2}{G M^2}} \right )$$ is the inner radius of the dyadosphere and $r_{ds}$ is the outer radius defined by $$r_{ds} = \sqrt{ \left(\frac{\hbar}{c m_e }\right) \left(\frac{GM}{c^2 } \right) \left(\frac{m_p}{ m_e }\right) \left(\frac{e}{q_p }\right) \left(\frac{Q}{\sqrt{G}M}\right)}$$ where $m_p = \sqrt{ \left(\frac{\hbar c}{G }\right)} $ and $q_p = \sqrt{ \hbar c} $ are Planck mass and Planck charge. It is shown that the dyadospheres exist for the charged black holes whose masses lie within the range, $ 3.2 M_{\bigodot} < M_{dyado} < 6\times 10^5 M_{\bigodot} $. Ruffini et al \[2\] hava found that total number of electron-positron pairs in the dyadosphere region is ( in the limit, $ r_{rs} > > \frac{GM}{c^2}$ ) $$N_{e^+ e^-} = \frac{Q-Q_{critical}}{e }\left[ 1 + \frac{(r_{ds} - r_+)}{\frac{\hbar}{m_e c }}\right]$$ During vacuum polarization process, the total energy of electron-positron pairs from the static electric energy and deposited within the dyadosphere is $$E_{dyado} = \frac {Q^2}{2 r_+ } \left( 1 - \frac {r_+}{r_{ds}} \right)\left[ 1 - \left(\frac {r_+}{r_{ds}}\right )^4 \right]$$ Delorenci et al \[7\] have computed the correction for the Reissner-Nordström metric from the first contribution of the Euler-Heisenberg Lagrangian and obtained the following metric as $$ds^2= \left[1 - \frac{2M}{ r} + \frac{Q^2}{ r^2} - \frac{\sigma Q^4}{5r^6}\right]dt^2 - \frac{dr^2}{ \left[1 - \frac{2M}{ r} + \frac{Q^2}{ r^2} - \frac{\sigma Q^4}{5r^6}\right]} - r^2 (d\theta^2 + \sin^2 \theta d\phi^2)$$ Here we use $ G = c = 1 $ and $\sigma$ is a parameter occurring due to vacuum fluctuation effects ( i.e. the last term coming from the one loop QED in the first order of the approximation ). One can note that when $ \sigma \rightarrow 0$, Reissner-Nordström solution is recovered. DeLorenci \[7\] have also shown that the correction term $\frac{\sigma Q^4}{5r^6}$ is of the same order of magnitude as the classical Reissner-Nordström charge term $\frac{Q^2}{ r^2}$. Let us cut out two slices of region from the dyadosphere geometry ( 4 - spaces ) described by $ \Omega^\pm = ( x \mid r \leq a ) $, where $ a\geq r_h$ ( position of event horizon of Reissner-Nordström black hole ). Now taking two copies of the remaining regions, $ M^\pm = ( x \mid r \geq a ) $, we paste the two pieces together at the hypersurface $ \Sigma = \Sigma^\pm = ( x \mid r = a ) $. Thus 3 - spaces $ \Sigma $ divides thew spacetime into two distinct four dimensional Manifold $M^+$ ( inner spacetime ) and $M^-$ ( exterior spacetime ). Thus one gets a geodesically complete manifold $ M = M^+ \bigcup M^- $ with a matter shell at the surface $ r = a $ , where the throat of the wormhole is located. This new construction implies that M is a manifold with two asymptotically flat regions connected by the throat. Since the boundary surface $ \Sigma $, is a 3 - spaces, we take the intrinsic coordinates in $\Sigma$ as $ \xi^i = ( \tau, {\theta}, \phi)$ with $\tau$ is the proper time on the junction shell. To understand the dynamics of the wormhole, we assume the radius of the throat be a function of the proper time $ a = a(\tau)$. The parametric equation for $\Sigma$ is defined by $$\Sigma : F(r,\tau ) = r - a(\tau)$$ The extrinsic curvature associated with the two sides of the shell are $$K_{ij}^\pm = - n_\nu^\pm\ [ \frac{\partial^2X_\nu} {\partial \xi^i\partial \xi^j } + \Gamma_{\alpha\beta}^\nu \frac{\partial X^\alpha}{\partial \xi^i} \frac{\partial X^\beta}{\partial \xi^j }] \|_\Sigma$$ where $ n_\nu^\pm\ $ are the unit normals to $\Sigma$, $$n_\nu^\pm = \pm \| g^{\alpha\beta}\frac{\partial F}{\partial X^\alpha} \frac{\partial F}{\partial X^\beta} \|^{-\frac{1}{2}} \frac{\partial F}{\partial X^\nu}$$ with $ n^\mu n_\mu = 1 $. The intrinsic metric on $\Sigma$ is given by $$ds^2 = - d\tau^2 + a(\tau)^2 d\Omega_2^2$$ From Lanczos equation, one can obtain the surface stress energy tensor $ S_j^i = diag ( - \sigma_s , -v_{\theta}, -v_{\phi}) $ (where $ \sigma$ is the surface energy density and $ v_{\theta,\phi} $ , the surface tensions) as $$\sigma_s = - \frac{1}{2 \pi a} \sqrt{1-\frac{2M}{a} + \frac{Q^2}{a^2}- \frac { \sigma Q^4}{5 a^6} + \dot{a}^2}$$ $$- v_{\theta} = - v_{\phi} = - v = \frac{1}{4\pi a} \frac{1 - \frac{M}{a} + \frac{2 \sigma Q^4}{5 a^6} + \dot{a}^2 + a \ddot{a} }{\sqrt{1-\frac{2M}{a} + \frac{Q^2}{a^2}- \frac{ \sigma Q^4}{5 a^6} + \dot{a}^2}}$$ where over dot means the derivative with respect to $\tau$. Negative surface energy density in (12) implies the existence of exotic matter at the shell. The negative signs of the tensions mean that they are indeed pressures ($ - v_{\theta} = - v_{\phi} = - v = p $). Here the radius of the shell is given by $a(\tau)$. For the static solution of the shell, we assume $\dot{a} = \ddot{a} = 0 $. The surface mass of this thin shell can be defined as $ M_{shell} = 4 \pi a^2 \sigma_s$ or $$M_{shell} = - 2 \pi a \sqrt{1-\frac{2M}{a} + \frac{Q^2}{a^2}- \frac { \sigma Q^4}{5 a^6}}$$ Here the term M could be interpreted as the total mass of the system i.e. total mass of the wormhole with two asymptotic regions connected by the throat at thin shell boundary surface $ \Sigma $. The above equation implies, $$M = \frac{a}{2} + \frac{Q^2}{2a} - \frac { \sigma Q^4}{10 a^5} - \frac { M_{shell}^2}{8 a}$$ It is interesting to note that $ M_{shell}$ is decreasing with increases of M and this indicates that one could reduce the exotic mass confined within the thin shell by increasing the mass of the black hole. So the minimizing of usage of exotic matter lies on the fact that how large Reissner-Nordström black hole we have considered. One can also find where the pressureless dust shell will occur. From equation (13) , $p = 0$ implies $$h(a) \equiv 1 - \frac{M}{a} + \frac { 2 \sigma Q^4}{5 a^6} = 0$$ For the suitable choices of parameters, the graph of the function $h(a)$ indicates the point $a_d$ where $h(a)$ cuts $'a'$ axis (see fig - 1 ). ![ Dust shell occurs at $a_d$ where $h(a)$ cuts $'a'$ axis ( choosing suitably the parameters as $ M = 10$ and $ \frac { 2 \sigma Q^4}{5} =.9 $ ).](dyado1.eps) We note that the matter on the junction surface shows peculiar behavior. This matter violates null energy and weak energy conditions but obeys strong energy condition. Here, $$\sigma_s < 0$$ $$\sigma_s + p = - \frac{\left( 1-\frac{3M}{a} + \frac{2Q^2}{a^2}- \frac { 3\sigma Q^4}{5 a^6}\right)}{4 \pi a \sqrt{1-\frac{2M}{a} + \frac{Q^2}{a^2}- \frac { \sigma Q^4}{5 a^6} }} < 0$$ $$\sigma_s + 3 p = \frac{\left( 1+\frac{M}{a} - \frac{2Q^2}{a^2}+ \frac { 7\sigma Q^4}{5 a^6}\right)}{4 \pi a \sqrt{1-\frac{2M}{a} + \frac{Q^2}{a^2}- \frac { \sigma Q^4}{5 a^6} }} > 0$$ Now we measure the Average Null Energy Condition (ANEC) violating matter present in the shell. This can be quantified by the following integrals\[19-20\]: $$\Omega_1 = \int \rho \sqrt{-g}d^3x , \Omega_j = \int [\rho + p_j] \sqrt{-g}d^3x$$ where, $\rho = \sigma_s$, the energy condition given in (12) and $p_j$, the principal pressures (here, radial pressure $p_r$ is zero and transverse pressures $ p_t = p_\theta = p_\phi = - v = p = - v_\theta = - v_\phi$ given in (13)). Following Eiroa and Simone \[11\] , we introduce a new radial coordinate $ R = \pm ( r -a ) $ in M ( $\pm $ for $M^{\pm}$ respectively ) so that $$\Omega_1 = \int_0^{2\pi} \int_0^{\pi}\int_{-\infty}^\infty \rho \sqrt{-g}dRd{\theta}d{\phi}$$ $$\Omega_j = \int_0^{2\pi} \int_0^{\pi}\int_{-\infty}^\infty [\rho + p_j] \sqrt{-g}dRd{\theta}d{\phi}$$ Since the shell does not exert radial pressure and the energy density is located on a thin shell surface, so that $ \rho = \rho + p_r = \delta(R)\sigma_s$, $ \rho + p_t = \delta(R)(\sigma_s + p_t)$, Hence, one gets, $$\Omega_r = \Omega_1 = \int_0^{2\pi} \int_0^\pi [\sigma_s \sqrt{-g} ]\|_{r=a} d{\theta}d{\phi} = 4\pi a^2\sigma(a) = -2a \sqrt{1-\frac{2M}{a} + \frac{Q^2}{a^2}- \frac { \sigma Q^4}{5 a^6} }$$ $$\Omega_t = \int_0^{2\pi} \int_0^\pi [(\sigma_s + p_t )\sqrt{-g} ]\|_{r=a} d{\theta}d{\phi} = - a\frac{\left( 1-\frac{3M}{a} + \frac{2Q^2}{a^2}- \frac { 3\sigma Q^4}{5 a^6}\right)}{ \sqrt{1-\frac{2M}{a} + \frac{Q^2}{a^2}- \frac { \sigma Q^4}{5 a^6} }}$$ one can see that if the charge of the Reissner-Nordström black hole is kept fixed, then total amount of ANEC violating matter present in the shell is reduced by increasing the mass of the black hole. Obviously this supports our previous analysis ( see eq.(15) ). Rearranging equation (12), we obtain the thin shell’s equation of motion $$\dot{a}^2 + V(a)= 0$$ Here the potential is defined as $$V(a) = 1-\frac{2M}{a} + \frac{Q^2}{a^2} - \frac{ \sigma Q^4}{5 a^6}- 4\pi^2 a^2\sigma_s^2(a)$$ Linearizing around a static solution situated at $a_0$, one can expand V(a) around $a_0$ to yield $$V = V(a_0) + V^\prime(a_0) ( a - a_0) + \frac{1}{2} V^{\prime\prime}(a_0) ( a - a_0)^2 + 0[( a - a_0)^3]$$ where prime denotes derivative with respect to $a$. Since we are linearizing around a static solution at $ a = a_0 $, we have $ V(a_0) = 0 $ and $ V^\prime(a_0)= 0 $. The stable equilibrium configurations correspond to the condition $ V^{\prime\prime}(a_0)> 0 $. Now we define a parameter $\beta$, which is interpreted as the speed of sound, by the relation \[7\] $$\beta^2(\sigma_s) = \frac{ \partial p}{\partial \sigma_s}\|_{\sigma_s}$$ Here, $$V^{\prime\prime}(a) = -\frac{4M}{a^3} + \frac{6Q^2}{a^4} - \frac{42 \sigma Q^4}{5a^8} - 8\pi^2 \sigma_s^2 - 32\pi^2 a \sigma_s \sigma_s^\prime - 8\pi^2 a^2 (\sigma_s^{\prime})^2 - 8\pi^2 a^2\sigma_s \sigma_s^{\prime\prime}$$ From equations (12) and (13), one can write energy conservation equation as $$\dot{\sigma_s} + 2\frac{\dot{a}}{a}( p + \sigma_s ) = 0$$ or $$\frac {d}{d \tau} (4 \pi \sigma_s a^2) + p \frac{d}{d \tau}(4\pi a^2)= 0$$ From equation (30) ( by using (28) ), we obtain, $$\sigma_s^{\prime\prime} + \frac{2}{a} \sigma_s ^{\prime} ( 1 + \beta^2) - \frac{2}{a^2} ( p + \sigma_s) = 0$$ The wormhole solution is stable if $ V^{\prime\prime}(a_0)> 0 $ i.e. if $$\beta_0^2 < \frac{\left( \frac{3}{2}-\frac{6M}{a_0} + \frac{5Q^2}{a_0^2} - \frac { 18\sigma Q^4}{5 a_0^6}\right)}{1-\frac{2M}{a} + \frac{Q^2}{a^2}- \frac { \sigma Q^4}{5 a^6} } - \frac{\left( 1-\frac{3M}{a_0} + \frac{2Q^2}{a_0^2} - \frac { 4\sigma Q^4}{5 a_0^6}\right)}{2 \left( 1-\frac{2M}{a} + \frac{Q^2}{a^2}- \frac { \sigma Q^4}{5 a^6}\right) }-\frac{3}{2}$$ If one treats $a_0$, M , Q and $\sigma $ are specified quantities, then the stability of the configuration requires the above restriction on the parameter $\beta_0$. This means there exists some part of the parameter space where the throat location is stable. For a lot of useful information, we show the stability regions graphically ( see figure 2). ![ Here we plot $ z = \beta^2_{\|{(a=a_0)}} $ $ Vs.$ $ x= \frac{a_0}{M}$ ( choosing suitably the parameters as $ (\frac{Q}{a_0})^2 = .12 $ and $\frac { \sigma Q^4}{5 a_0^6} = .1$ ). The stability region is situated below the curve.[]{data-label="fig:stability"}](dyado2.eps) The minimizing of usage exotic matter needed to construct a wormhole remains an encouraging research area to the scientists working in wormhole physics. Several models and ideas are proposed time to time. Recent observations of gamma ray bursts confirmed that there should exist some sources that produce these bursts. Ruffini and his collaborators argued that these sources are nothing but the dyadospheres. In this work, we have considered a dyadosphere of Reissner-Nordström black hole to develop thin shell wormhole. We have constructed thin shell wormhole by surgically grafting two dyadosphere spacetimes. This study is interesting and more physical since we have used the spacetimes of the astrophysical sources of cosmic gamma ray bursts. We have analyzed the dynamical stability of thin shell wormhole considering linearized radial perturbation around the stable solution. We have shown that there exist some part of the parametric space where the thin shell wormhole is stable. We have also discussed the stability graphically. The most important part of this study is the minimizing of usage of exotic matter confined within the shell. We have seen that Reissner-Nordström black hole mass plays a crucial role to minimize the usage of exotic matter. The Reissner-Nordström black hole with larger mass reduces the amount of exotic matter needed for their construction. Finally, we note that the parameter $\sigma$ coming from the one loop QED in the first order of the approximation is also responsible for reducing the exotic matter. [ Acknowledgments ]{} F.R. is thankful to DST , Government of India for providing financial support.\ [99]{} R Ruffini, in XLLX Yamada Conference on Black holes and High Energy Astrophysics, edited by H Salto ( Univ.Acad.Press., Tokyo , 1998 ) R Ruffini et al, Int.J.Mod.Phys. D 12, 173 (2003) R Ruffini, ArXiv: astr-ph/9905072 G Preparata, R Ruffini and S Xue, Astronmy and Astrophysics L 87, 338 (1998) L T Drummond et al, Phys.Rev.D 22, 343 (1980) R D Daniels et al, Nucl.Phys.B 425, 635 ( 1994) V A De Lorenci et al, Phys.Lett.B 482, 134 (2000) M Visser, Nucl.Phys.B 328, 203 ( 1989) E Poisson and M Visser, Phys.Rev.D 52, 7318 (1995) \[arXiv: gr-qc / 9506083\] E Eiroa and G Romero, Gen.Rel.Grav. 36, 651 (2004)\[arXiv: gr-qc / 0303093\] E Eiroa and C Simeone, Phys.Rev.D 71, 127501 (2005) \[arXiv: gr-qc / 0502073\] M Thibeault , C Simeone and E Eiroa, arXiv: gr-qc / 0512029 F.Rahaman , M.Kalam and S.Chakraborty, Gen.Rel.Grav.38:1687-1695,2006. e-Print: gr-qc/0607061 F.Rahaman , M.Kalam and S.Chakraborti, Int.J.Mod.Phys.D16:1669-1681,2007. e-Print: gr-qc/0611134 F.Rahaman et al, Gen.Rel.Grav.39:945-956,2007. e-Print: gr-qc/0703143 F.Rahaman et al, arxiv:0804.3852\[gr-qc\] E Eiroa et al, gr-qc/0404050 F Lobo, gr-qc/0311002 M Visser, S Kar and N Dadhich, Phys. Rev.Lett. 90, 201102(2003)\[arXiv: gr-qc / 0301003\] K K Nandi et al , Phys.Rev.D70, 127503(2004)	{ "pile_set_name": "ArXiv" }
--- author: - Tuhin Sahai title: 'Dynamical Systems Theory and Algorithms for NP-hard Problems' --- Computational Complexity, Dynamical Systems Theory, NP-hardness, Heuristic Algorithms, Combinatorial Optimization. Introduction {#sec:intro} ============ Dynamical systems theory and computational complexity have, predominantly, been developed as independent areas of research over the last century with little interaction and mutual influence. Dynamical systems theory has its origins in the seminal work of Henri Poincaré [@Cit:Poincare] on celestial mechanics. Computational complexity theory, on the other hand, originated in the works of Alan Turing [@Cit:Turing] and Alonzo Church [@Cit:Church] in the 1930s and has played an intimate role in the computing revolution of the twentieth century. Eventually, dynamical systems theory (or nonlinear dynamics) found broad application beyond celestial mechanics. In particular, it has been used extensively to model and analyze engineering systems [@Cit:Stro], physics of natural phenomena, biological [@Cit:Bio] and chemical processes [@Cit:chem], fluid dynamics [@Cit:turb], and epidemiology [@Cit:epi] to name a few. Moreover, the analysis of dynamical systems is typically intimately tied to numerical methods [@Cit:set_oriented; @Cit:Igor] and scientific computation [@Cit:scientific_comp]. Links to the applications (outlined in the previous paragraph) have played a critical role in the theoretical development of the field. For example, they have influenced the development of various sub-areas within nonlinear dynamics such as ergodicity [@Cit:ergodicity], chaos theory [@Cit:Lorenz], and symbolic dynamics [@Cit:symb] to name a few. For a broad overview of the theoretical approaches to dynamical systems, we refer the reader to [@Cit:Gucken]. Although, dynamical systems theory has found wide application in engineering and the sciences, it has received scant attention from the computer science community. Local continuous optimization techniques such as Nesterov’s method [@Cit:Nesterov] have recently been analyzed from a dynamical systems perspective [@Cit:boyd_candes]. Nesterov’s method is an optimal gradient descent algorithm in terms of convergence rate. In [@Cit:boyd_candes], the authors derive a dynamical system by invoking a continuous time limit of the optimization step size. They then analyze the resulting ordinary differential equations (ODEs) to provide valuable insight into the algorithm and its associated optimality. Additionally, in [@Cit:wibisono] the authors use calculus of variations to gain additional insight into the convergence rates of accelerated gradient descent schemes. Although, this body of work does fall under the category of novel application of dynamical systems theory to optimization methods, we will not discuss it at length in this paper for two reasons a) this work has sparked extensive follow-on work and consequently, various summary articles and presentations are already available, and b) they appear to be restricted to accelerated gradient methods with no clear extension to the broader theory of computational complexity. Non-deterministic polynomial-time (NP)-hard and -complete complexity classes can be traced to seminal work by Cook in 1971 [@Cit:cook]. The broad applicability of this work was outlined in a highly influential publication by Karp [@Cit:karp]. NP-hard problems such as the traveling salesman problem (TSP) [@Cit:TSP] and lattice-based vector problems [@Cit:lattice] arise in a wide variety of applications ranging from DNA sequencing and astronomy [@Cit:TSP_book] to encryption [@Cit:lattice]. In essence, the computation of optimal solutions for these problems quickly becomes intractable with the size of the instance (unlike problems that lie in the $P$ complexity class). Note that some problems such as graph isomorphism [@Cit:GI_klus] lie in the NP complexity class but are not expected to be NP-complete or NP-hard. Over the last few years, several efficient heuristic algorithms for approximating the solutions of NP-hard problems have been developed. For example, careful implementations of the Lin-Kernighan [@Cit:LKH] and branch-and-bound [@Cit:Concorde] heuristics have been successful in computing optimal solutions of several large instances of the TSP. However, most NP-hard problems suffer from a lack of scalable approaches. Moreover, as long as $P\neq NP$ (where $P$ is the complexity class of problems that can be solved in poylnomial time on a deterministic Turing machine), even efficient heuristics for some of these problems will remain elusive. See Fig. \[fig:complexity\_map\] for the hypothesized relationship between the most popular classes. ![The computational complexity map for the most popular complexity classes.[]{data-label="fig:complexity_map"}](ComplexityMap2){width="65.00000%"} In this work, we start by surveying the use of dynamical systems in the context of constructing state-of-the-art algorithms for NP-hard problems. In particular, we will cover the use of dynamical systems theory for constructing decentralized graph clustering algorithms [@Cit:cluster1; @Cit:cluster2], solutions for the TSP [@Cit:TSP_sahai], and quantum-inspired networks of Duffing oscillators for solving the MAX-CUT problem [@Cit:quantum_net]. We then switch to the use of dynamical systems theory for analysis of algorithms [@Cit:koopman] and the underlying problems [@Cit:zoltan; @Cit:zoltan2]. The goal of this survey paper is to highlight the potential application of dynamical systems theory for optimization of complex functions and analysis of computational complexity theory. This is a nascent field which presents the possibility of tremendous impact. Additionally, we expect this area to lead to new theoretical developments in nonlinear dynamics theory and novel algorithms for computationally intractable problems. Ziessler, Surana, Speranzon, Klus, Dellnitz, and Banaszuk have all served as co-authors in my efforts in this area. However, my extensive discussions with Prof. Michael Dellnitz inspired me to delve deeper into the area of dynamical systems and the analysis of algorithms! Novel algorithm construction: decentralized graph clustering {#sec:graph} ============================================================ [Overview]{} Algorithms for graph analysis have a wide variety of applications such as routing, pattern recognition, database searches, network layout, and Internet PageRank to name a few [@Cit:graph_algos]. Although some of these problems can solved efficiently on present day computing devices, several graph analysis problems are computationally intractable [@Cit:cormen]. For example, the problem of partitioning graphs into equal size sets while minimizing the weights of cut edges arises in a range of settings such as social anthropology, gene networks, protein sequences, sensor networks, computer graphics, and Internet routing algorithms [@Cit:cluster2]. To avoid unbalanced cuts, size restrictions are typically placed on the clusters; instead of minimizing inter-connection strength, if one minimizes the ratio of the inter-connection strength to the size of individual clusters, the problem becomes NP-complete [@Cit:tutorial; @Cit:npcomp]. In [@Cit:cluster1; @Cit:cluster2], a novel decentralized algorithm for clustering/partitioning graphs that exploits fundamental properties of a dynamically evolving networked system was constructed. In particular, by propagating waves in a graph, one can compute partitions or clusters in a completely decentralized setting. The method is orders of magnitude faster than existing approaches [@Cit:kempe]. This is our first example of a dynamical systems theory based algorithm for a combinatorial optimization problem. We now discuss the details of the approach. Let $\mathcal{G}=(V,E)$ be a graph with vertex set $V = \{1,\dots,N\}$ and edge set $E\subseteq V\times V$, where a weight $\W_{ij} \geq 0$ is associated with each edge $(i,j)\in E$, and $\W$ is the $N\times N$ weighted adjacency matrix of $\mathcal{G}$. We assume that $\W_{ij}=0$ if and only if $(i,j) \notin E$. The (normalized) graph Laplacian is defined as, $$\begin{aligned} \L_{ij} = \begin{cases} 1 & \mbox{if}\: i = j\\ -\W_{ij}/\sum_{\ell=1}^N \W_{i\ell} & \mbox{if}\: (i,j) \in E\\ 0 & \mbox{otherwise}\,, \end{cases} \label{eq:ldef}\end{aligned}$$ or equivalently, $\L = \I-\D^{-1}\W$ where $\D$ is the diagonal matrix with the row sums of $\W$. Note that in [@Cit:cluster2], only undirected graphs were considered. The smallest eigenvalue of the Laplacian matrix is $\lambda_1 = 0$, with an associated eigenvector $\v^{(1)}=\1=\left[1,1,\dots,1\right]^T$. Eigenvalues of $\L$ can be ordered as, $ 0 = \lambda_1 \leq \lambda_2 \leq \lambda_3 \leq \cdots \leq \lambda_N$ with associated eigenvectors $\1, \v^{(2)}, \v^{(3)}\cdots \v^{(N)}$ [@Cit:tutorial]. It is well known that the multiplicity of $\lambda_1$ is the number of connected components in the graph [@Cit:tutorial]. Given the Laplacian matrix $\L$, associated with a graph $\mathcal{G}= (V,E)$, spectral clustering divides $\mathcal{G}$ into two clusters by computing the signs of the $N$ elements of the second eigenvector $\v^{(2)}$, or Fiedler vector. For further details about the computation of two or more clusters see [@Cit:tutorial]. There are many algorithms to compute eigenvectors, such as the Lanczos method or orthogonal iteration [@GolubVanLoan96]. Although some of these methods are distributable, convergence is slow [@GolubVanLoan96] and these algorithms do not consider/take advantage of the fact that the matrix for which the eigenvalues and eigenvectors need to be computed is the adjacency matrix of the underlying graph. In [@Cit:kempe], the authors propose an algorithm to compute the first $k$ largest eigenvectors (associated with the first $k$ eigenvalues with greatest absolute value)[^1] of a symmetric matrix. The algorithm in [@Cit:kempe] emulates the behavior of orthogonal iteration using a decentralized process based on gossip algorithms or deterministic random walks on graphs. This approach can be slow as it converges after $O(\tau\log^2 N)$ iterations [@Cit:kempe] where where $\tau$ is the mixing time for the random walk on the graph and $N$ is the number of nodes. This procedure is equivalent to evolving the discretized heat equation on the graph and can be demonstrated as follows. The heat equation is given by, $$\begin{aligned} \frac{\partial u}{\partial t} = \Delta u\,,\end{aligned}$$ where $u$ is a function of time and space, $\partial u/\partial t$ is the partial derivative of $u$ with respect to time, and $\Delta$ is the Laplace operator [@Cit:cluster2]. When the above equation is discretized on a graph $\mathcal{G}=(V,E)$ one gets the following equation: $$\begin{aligned} \u_{i}(t+1) = \u_{i}(t) - \displaystyle\sum_{j\in\mathcal{N}(i)}\L_{ij}\u_{j}(t)\,,\end{aligned}$$ for $i,j \in V$. Here $\u_{i}(t)$ is the scalar value of $u$ on node $i$ at time $t$ and $\mathcal{N}(i)$ are the neighbors of node $i$ in the graph. The graph Laplacian $\L=[\L_{ij}]$ is the discrete counterpart of the $\Delta$ operator. The above iteration can be re-written, in matrix form, $\u(t+1) = (\I-\L)\,\u(t)$ where $\u(t) = (\u_{1}(t),\dots,\u_{N}(t))^{T}$. The solution of this iteration is, $$\begin{aligned} \u(t) = C_0\1 + C_1 (1-\lambda_2)^t \v^{(2)} + \dots +C_N (1-\lambda_N)^t \v^{(N)}\,, \label{eq:HeatSol}\end{aligned}$$ where constants $C_{j}$ depend on the initial condition $\u(0)$. It is interesting to note that in Eqn. \[eq:HeatSol\], the dependence of the solution on higher eigenvectors and eigenvalues of the Laplacian decays with increasing iteration count. Thus, it is difficult to devise a fast and distributed method for clustering graphs based on the heat equation. In [@Cit:cluster1; @Cit:cluster2], a novel algorithm based on the idea of permanent excitation of the eigenvectors of $\I-\L$ using dynamical systems theory is constructed. In a theme similar to Mark Kac’s question “Can one hear the shape of a drum?” [@DrumShape], it was demonstrated that by evolving the wave equation in the graph, nodes can “hear” the eigenvectors of the graph Laplacian using only local information. Moreover, it was shown, both theoretically and on examples, that the wave equation based algorithm is orders of magnitude faster than random walk based approaches for graphs with large mixing times. The overall idea of the wave equation based approach is to simulate, in a distributed fashion, the propagation of a wave through the graph and capture the frequencies at which the graph “resonates”. In other words, it was shown that by using these frequencies one can compute the eigenvectors of the Laplacian, thus clustering the graph. The wave equation based clustering approach can be described as follows. Analogous to the heat equation case (Eq. \[eq:HeatSol\]), the solution of the wave equation can be expanded in terms of the eigenvectors of the Laplacian. However, unlike the heat equation where the solution eventually converges to the first eigenvector of the Laplacian, in the wave equation all the eigenvectors remain eternally excited (a consequence of the second derivative of $u$ with respect to time). This observation is used to develop a simple, yet powerful, distributed eigenvector computation algorithm. The algorithm involves evolving the wave equation on the graph and then computing the eigenvectors using local FFTs. The graph decomposition/partitioning algorithm based on the discretized wave equation on the graph, given by $$\begin{aligned} \u_{i}(t) = 2\u_{i}(t-1) - \u_{i}(t-2) - c^2\displaystyle\sum_{j\in\mathcal{N}(i)}\L_{ij} \u_{j}(t-1)\,, \label{onenodewave}\end{aligned}$$ where $\sum_{j\in\mathcal{N}(i)}\L_{ij}\u_{j}(t-1)$ originates from the discretization of the spatial derivatives in the wave equation. The rest of the terms originate from discretization of the $\partial^{2} u/\partial t^{2}$ term in the wave equation. To update $\u_{i}$ using Eq. \[onenodewave\], one needs only the value of $\u_{j}$ at neighboring nodes and the connecting edge weights (along with previous values of $\u_{i}$). The main steps of the algorithm are shown as Algorithm \[alg:WaveAlg\]. Note that at each node (node $i$ in the algorithm) one only needs nearest neighbor weights $\L_{ij}$ and the scalar quantities $\u_{j}(t-1)$ also at nearest neighbors. We emphasize, again, that $\u_{i}(t)$ is a scalar quantity and `Random`($[0,1]$) is a random initial condition on the interval $[0,1]$. The vector $\v^{(j)}_{i}$ is the $i$-th component of the $j$-th eigenvector, $T_{max}$ is a positive integer derived in [@Cit:cluster1; @Cit:cluster2], $\texttt{FrequencyPeak(Y,j)}$ returns the frequency at which the $j$-th peak occurs and $\texttt{Coefficient}(\omega_{j})$ return the corresponding Fourier coefficient. $\u_{i}(0) \leftarrow \texttt{Random}\:([0,1])$ $\u_{i}(-1) \leftarrow \u_{i}(0)$ $t\leftarrow 1$ ----------------------------------------------------------------- $\u_{i}(t) \leftarrow 2\u_{i}(t-1)-\u_{i}(t-2) -$ $\qquad \qquad c^2 \sum_{j\in\mathcal{N}(i)}\L_{ij}\u_{j}(t-1)$ ----------------------------------------------------------------- $t\leftarrow t+1$ $Y\leftarrow \texttt{FFT}\:(\left[\u_{i}(1),\dots\dots,\u_{i}(T_{max})\right])$ $\omega_{j} \leftarrow \texttt{FrequencyPeak}\:(Y,j)$ $\v^{(j)}_{i} \leftarrow \texttt{Coefficient}(\omega_{j})$ $A_j \leftarrow 1$ $A_j \leftarrow 0$ ClusterNumber $\leftarrow \sum_{j=1}^k A_j 2^{j-1}$ The clusters of graph $\mathcal{G}$, determined by the signs of the elements of the eigenvectors of $\L$, can be computed using the frequencies and coefficients obtained from the Fast Fourier Transform of $(\u_{i}(1),\dots,\u_{i}(T_{max}))$, for all $i$ and some $T_{max}>0$. Here $\u_i$ is governed by the wave equation on the graph (shown in Eqn. \[onenodewave\]) with the initial condition $\u(-1)=\u(0)$ and $0<c < \sqrt{2}$. For the proofs see [@Cit:cluster1; @Cit:cluster2]. The above proof demonstrates that the approach is fundamentally decentralized. Moreover, it is shown in [@Cit:cluster1; @Cit:cluster2] that the convergence of the wave equation based eigenvector computation depends on the mixing time of the underlying Markov chain on the graph, and is given by, $$T_{max} = O\left(\arccos\left(\cfrac{2+c^2 (e^{-1/\tau}-1)}{2}\right)^{-1} \right) + O(N)\,, \label{eq:Waveconvfinal}$$ where $\tau$ is the mixing time of the Markov chain. Thus, the wave equation based algorithm has better scaling with $\tau$ for graphs of any size (given by $N$, see Fig. \[ConvComp\]). ![Comparison of convergence rates between the distributed algorithm in [@Cit:kempe] and our proposed wave equation algorithm for $c^{2}=1.99$. The wave equation based algorithm has better scaling with $\tau$ for graphs of any size (given by $N$). The plots are upper bounds on the convergence speed. For more details see [@Cit:cluster2].\[ConvComp\]](TauComp1.eps "fig:"){width="0.85\hsize"}\ The above work is an example of the construction of a state-of-the-art algorithm using dynamical systems theory. This work has also found application in distributed numerical computations [@Cit:Num_klus] and uncertainty quantification [@Cit:UQ_surana]. We now present another example of constructing novel algorithms for NP-hard problems using the theory of nonlinear dynamics and invariant manifold computations. Novel algorithm construction: invariant manifolds and the traveling salesman problem {#sec:tsp} ==================================================================================== [Overview]{} Recently, dynamical systems theory was used to construct novel algorithms for another iconic NP-hard problem [@Cit:TSP_sahai]. The traveling salesman problem (TSP) has a long and rich history in the areas of computer science, optimization theory, and computational complexity, and has received decades of interest [@Cit:cook]. This combinatorial optimization problem arises in a wide variety of applications related to genome map construction, telescope management, and drilling circuit boards. The TSP also naturally occurs in applications related to target tracking [@Cit:target_tracking], vehicle routing, and communication networks to name a few. We refer the reader to [@Cit:cook; @Cit:TSP_sahai] for further details. In its basic form, the statement of the TSP is exceedingly simple. The task is to find the shortest Hamiltonian circuit through a list of cities, given their pairwise distances. Despite its simplistic appearance, the underlying problem is NP-hard [@Cit:karp]. Several heuristics have been developed over the years to solve the problem [@Cit:cook] including ant colony optimization, cutting plane methods, Christofides heuristic algorithm, and the Lin–Kernighan heuristic. In [@Cit:TSP_sahai], inspired by dynamical systems theory, the authors construct novel orthogonal relaxation based approximations to the TSP. In particular, the constructed dynamical system captures the flow on the manifold of orthogonal matrices and ideally converges to a permutation matrix that minimizes the tour length. However, in general, the flow typically converges to local minima that are not competitive when compared to state-of-the-art heuristics. Inspired by this continuous relaxation, the authors compute the solution to a two-sided orthogonal Procrustes problem [@Cit:Procrustes-book] that relaxes the TSP to the manifold of orthogonal matrices. They then combine the Procrustes approach with the Lin–Kernighan heuristic [@Cit:LKH] for computing solutions of the TSP. Additionally, the authors use set-oriented methods to study the stability of optimal solutions and their stable manifolds, thereby providing insight into the associated basins of attraction and the resulting computational complexity of the problem. Given a list of $ n $ cities $ \{ C_{1}, C_{2}, \dots, C_{n} \} $ and the associated distances between cities $ C_i $ and $ C_j $, denoted by $ d_{ij} $, the TSP aims to find an ordering $ \sigma $ of $ \{ 1, 2, \dots, n \} $ such that the tour cost, given by $$\label{eq:basic_cost} c = \sum_{i=1}^{n-1} d_{\sigma(i), \sigma(i+1)} + d_{\sigma(n), \sigma(1)},$$ is minimized. For the Euclidean TSP, for instance, $ d_{ij} = \|\|x_i - x_j\|\|_2 $, where $ x_i \in \R^d $ is the position of $ C_i $. In general, however, the distance matrix $ D = (d_{ij}) $ does not have to be symmetric (for example see [@Cit:tsp_asymm]). The ordering $ \sigma $ can be represented as a unique permutation matrix $ P $. Note, however, that due to the underlying cyclic symmetry, multiple orderings – corresponding to different permutation matrices – have the same cost. There are several equivalent ways to define the cost function of the TSP. The authors restrict themselves to the trace[^2] formulation. Let $ \mathcal{P}_n $ denote the set of all $ n \times n $ permutation matrices, then the TSP can be written as a combinatorial optimization problem of the form $$\label{eq:TSPCost} \min_{P \in \mathcal{P}_n} \mbox{tr} \left( A^T P^T B P \right),$$ where $ A = D $ and $ B = T $. Here, $ T $ is defined to be the adjacency matrix of the cycle graph of length $ n $. One uses the undirected cycle graph adjacency matrix for symmetric TSPs and the one corresponding to the directed cycle graphs for asymmetric TSPs. The matrices are defined as, $$\label{eq:tmatrix} T_\text{dir} = \begin{pmatrix} 0 & 1 & & & \\ & 0 & 1 & & \\ & & \ddots & \ddots & \\ & & & 0 & 1 \\ 1 & & & & 0 \end{pmatrix} \; \text{ or } \; T_\text{undir} = \begin{pmatrix} 0 & 1 & & & 1 \\ 1 & 0 & 1 & & \\ & \ddots & \ddots & \ddots & \\ & & 1 & 0 & 1 \\ 1 & & & 1 & 0 \end{pmatrix}.$$ The work in [@Cit:TSP_sahai] focuses on the undirected version of the TSP. By relaxing the TSP problem to the manifold of orthogonal matrices (since permutation matrices are orthogonal matrices restricted to $0$ or $1$ entries), one can use the two sided Procrustes problem to solve the problem exactly, as outlined in the theorem below. Given two symmetric matrices $ A $ and $ B $, whose eigenvalues are distinct, let $ {A = V_A \Lambda_A V_A^T} $ and $ B = V_B \Lambda_B V_B^T $ be eigendecompositions, with $ \Lambda_A = \diag\left(\lambda_{A}^{(1)}, \dots, \lambda_{A}^{(n)}\right) $, $ \Lambda_B = \diag\left(\lambda_{B}^{(1)}, \dots, \lambda_{B}^{(n)}\right) $, and $ \lambda_{A}^{(1)} \geq \dots \geq \lambda_{A}^{(n)} $ as well as $ \lambda_{B}^{(1)} \geq \dots \geq \lambda_{B}^{(n)} $. Then every orthogonal matrix $ P^* $ which minimizes $$\min_{P \in \mathcal{O}_n} \|\| A - P^T B P \|\|_F \label{eq:RelTSPCost}$$ has the form $$P^* = V_B S V_A^T,$$ where $ S = \diag(\pm 1, \dots, \pm 1) $. A proof of this theorem can be found in [@Cit:Sch68]. If the eigenvalues of $ A $ and $ B $ are distinct, then there exist $ 2^n $ different solutions with the same cost. If one or both of the matrices possess repeated eigenvalues, then the eigenvectors in the matrices $V_A$ and $V_B$ are determined only up to basis rotations, which further increases the size of the solution space. The Procrustes problem is related to a dynamical system formulation of the TSP as outlined below. The orthogonal relaxation of the combinatorial optimization problem , given by , can be solved using a steepest descent method on the manifold of orthogonal matrices. For more details about this formulation see [@Cit:TSP_sahai]. One can pose the TSP as a constrained optimization problem of the form, $$\begin{aligned} \label{eq:minPwithEqualityConstraints} \min_{P \in \mathcal{O}_n} & \; \mbox{tr} \left( A^T P^T B P \right), \\ s.t. \; & \; G(P) = 0.\end{aligned}$$ This formulation gives rise to the following set of equations, $$\label{eq:TSP flow} \begin{aligned} \dot{P} &= -P \left( \left\{ P^T B P, A \right\} + \left \{P^T B^T P, A^T \right\} \right) - \lambda P \left( (P \circ P)^T P - P^T(P \circ P) \right), \\ \dot{\lambda} &= \frac{1}{3} \mbox{tr} \left( P^T \left( P - (P \circ P) \right) \right). \end{aligned}$$ The above set of equations are obtained by using gradient descent on the Lagrangian cost function. \[ex:P flow\] In order to illustrate the gradient flow approach, let us consider a simple TSP with 10 cities. Using , we obtain the results shown in Figure \[fig:P flow\]. In this example, the dynamical system converges to the optimal tour. ![Traveling salesman problem with 10 cities solved using the gradient flow . The original positions of the cities are shown in black, the positions transformed by the orthogonal matrix $ P $ in red. a) Initial trivial tour given by $ \sigma = (1, \dots, 10) $. b–d) Intermediate solutions. e) Convergence to an orthogonal matrix which is “close” to a permutation matrix with respect to any matrix norm. f) Extraction of the corresponding permutation matrix. The initial tour was transformed into the optimal tour by the gradient flow.[]{data-label="fig:P flow"}](P1){width="80.00000%"} ![Traveling salesman problem with 10 cities solved using the gradient flow . The original positions of the cities are shown in black, the positions transformed by the orthogonal matrix $ P $ in red. a) Initial trivial tour given by $ \sigma = (1, \dots, 10) $. b–d) Intermediate solutions. e) Convergence to an orthogonal matrix which is “close” to a permutation matrix with respect to any matrix norm. f) Extraction of the corresponding permutation matrix. The initial tour was transformed into the optimal tour by the gradient flow.[]{data-label="fig:P flow"}](P2){width="80.00000%"} ![Traveling salesman problem with 10 cities solved using the gradient flow . The original positions of the cities are shown in black, the positions transformed by the orthogonal matrix $ P $ in red. a) Initial trivial tour given by $ \sigma = (1, \dots, 10) $. b–d) Intermediate solutions. e) Convergence to an orthogonal matrix which is “close” to a permutation matrix with respect to any matrix norm. f) Extraction of the corresponding permutation matrix. The initial tour was transformed into the optimal tour by the gradient flow.[]{data-label="fig:P flow"}](P3){width="80.00000%"} \ ![Traveling salesman problem with 10 cities solved using the gradient flow . The original positions of the cities are shown in black, the positions transformed by the orthogonal matrix $ P $ in red. a) Initial trivial tour given by $ \sigma = (1, \dots, 10) $. b–d) Intermediate solutions. e) Convergence to an orthogonal matrix which is “close” to a permutation matrix with respect to any matrix norm. f) Extraction of the corresponding permutation matrix. The initial tour was transformed into the optimal tour by the gradient flow.[]{data-label="fig:P flow"}](P4){width="80.00000%"} ![Traveling salesman problem with 10 cities solved using the gradient flow . The original positions of the cities are shown in black, the positions transformed by the orthogonal matrix $ P $ in red. a) Initial trivial tour given by $ \sigma = (1, \dots, 10) $. b–d) Intermediate solutions. e) Convergence to an orthogonal matrix which is “close” to a permutation matrix with respect to any matrix norm. f) Extraction of the corresponding permutation matrix. The initial tour was transformed into the optimal tour by the gradient flow.[]{data-label="fig:P flow"}](P5){width="80.00000%"} ![Traveling salesman problem with 10 cities solved using the gradient flow . The original positions of the cities are shown in black, the positions transformed by the orthogonal matrix $ P $ in red. a) Initial trivial tour given by $ \sigma = (1, \dots, 10) $. b–d) Intermediate solutions. e) Convergence to an orthogonal matrix which is “close” to a permutation matrix with respect to any matrix norm. f) Extraction of the corresponding permutation matrix. The initial tour was transformed into the optimal tour by the gradient flow.[]{data-label="fig:P flow"}](P6){width="80.00000%"} The dynamical system without constraints converges to equilibria that are given by the Procrustes solutions. To shed light on the stability and local dynamics around the optimal TSP solutions one can approximate subsets of the stable manifold of the Procrustes solutions such that two permutation matrices are inside these sets. This numerical study enables the analysis of the robustness of Procrustes solutions under small perturbations of the initial permutation matrix and the assessment of the ‘closeness’ the Procrustes solution is to the optimal permutation matrix. In order to compute the sets of interest, set-oriented continuation techniques developed in [@DH96] are used in [@Cit:TSP_sahai]. An example computation is depicted in Fig. \[fig:mostAttractingSets\]. ![Three-dimensional projection of two subsets of the stable manifold. The omega-limit sets of a small neighborhood of the permutation matrices $P_1$ and $P_2$ form a half circle on their corresponding Procrustes set.[]{data-label="fig:mostAttractingSets"}](AttractingSet2PermutationMatrices){width=".8\textwidth"} Moreover, one can also use set oriented methods to compute basins of attraction of optimal permutation matrices for small instances of the TSP. These basins (subsets of the stable manifold) are computed by perturbing the optimal solutions and integrating the flow backward in time [@Cit:TSP_sahai]. The solutions are shown in Fig. \[fig:polar\]. These computations are interesting and capture the “hardness” of the problem. In particular, one can see that the solutions of relaxed versions of the problem (such as the relaxations to the manifolds of orthogonal matrices) do not, in general, lie in the basin of attraction of the optimal solutions of the original problem. Other such instances of analysis of relaxed solutions of the TSP using dynamical systems theory are outlined in [@Cit:TSP_sahai]. Although, dynamical systems theory demonstrates that the Procrustes solutions do not typically lie in the basin of attraction of the optimal solutions of the TSP, a new biasing scheme for the Lin–Kernighan heuristic is constructed using the aforementioned relaxation [@Cit:TSP_sahai]. The Lin–Kernighan heuristic is a popular heuristic for the TSP [@Cit:LKH]. Starting from an initial tour, the approach progresses by extracting edges from the tour and replacing them with new edges, while maintaining the Hamiltonian cycle constraint. If $k$ edges in the tour are simultaneously replaced, this is known as the $k$-opt move [@Cit:LKH]. To prune the search space, the algorithm relies on minimum spanning trees to identify edges that are more likely to be in the tour. This “importance” metric for edges is called $\alpha$-nearness and described in [@Cit:LKH; @Cit:TSP_sahai]. The algorithm has found great success on large instances of the TSP, see [@Cit:cook] for more details. In [@Cit:TSP_sahai], the $\alpha$-nearness metric is replaced with a new Procrustes solution–based metric that prunes/identifies important potential edges to include in the “candidate set list”. This list is then used to generate the $k$-opt moves. The metric is captured in Fig. \[fig:distance\_vs\_procrustes\]. The Procrustes solution tends to capture the longer edges that are important. To increase the inclusion of the short edges, the approach in [@Cit:TSP_sahai] constructs a homotopy between the Procrustes ($P$-nearness) solution and the distance matrix. Using a graph Laplacian approach, the mixture of the two matrices is compared to the $\alpha$-nearness approach on 22 well-known instances of the TSP. $ P $-nearness based LKH converges to lower cost values in $18$ of the instances when compared to $\alpha$-nearness based LKH. Moreover, for $50$ random TSP instances of size $1000$ (cities) it is found that $P$-nearness has lower tour costs after a fixed number of $k$-opt moves in $31$ of the instances, translating into an improvement for $62\%$ of the instances. ![Illustration of $ P $-nearness for random TSP instances of size $50$ and $100$. The left column contains the edges with shortest distance, the center column has the optimal tour for the instances, and the right column contains the edges with the highest $ P $-nearness values for each city. For each city, we plotted the three edges with the highest nearness values.[]{data-label="fig:distance_vs_procrustes"}](Distance_vs_Procrustes){width="95.00000%"} Thus, this is another example that demonstrates that dynamical systems theory can be used to analyze NP-hard problems and construct improved heuristics. We now show how networks of Duffing oscillators can be used to construct a new algorithm for the iconic MAX-CUT problem. Novel algorithm construction: network of Duffing oscillators for the MAX-CUT problem {#sec:maxsat_quantum} ==================================================================================== [Overview]{} MAX-CUT [@Cit:karp] is a well-known NP-hard problem that arises in graph theory. Simply stated, the goal is to compute a subset $S$ of the vertex set in a graph $\mathcal{G}$, such that the number of edges between $S$ and the rest of the graph are maximized. The best known approximation ratio of $0.878$ can be achieved in polynomial time using semi-definite programming [@Cit:Goemans]. The problem naturally arises in VLSI design and statistical physics, and has been extensively studied. In [@Cit:quantum_net], the authors construct an optimization algorithm by simulating the adiabatic evolution of Hamiltonian systems which can be used to approximate the MAX-CUT solution of an all-to-all connected graph. The approach is inspired by quantum adiabatic optimization for Ising systems [@Cit:Ising] with the following energy, $$\begin{aligned} E_{Ising}(s) = -\frac{1}{2}\displaystyle\sum_{i=1}^{N}\sum_{j=1}^{N}J_{ij}s_{i}s_{j}, \label{eq:ising}\end{aligned}$$ where $s_i$ are the spins which can take values {-1,1} and $J_{ij}$ is the coupling coefficient. Finding the lowest energy state of the Ising system is computationally challenging (for a system with N spins, the potential number of states is $2^N$). Note that one can map the Ising problem to the MAX-CUT problem by setting $J_{ij}=-w_{ij}$, where $w_{ij}$ is the weight of the edge that connects nodes $i$ and $j$. It is easy to show that minimizing the energy in Eqn. \[eq:ising\] corresponds computing the solution of the MAX-CUT problem. The approach outlined in [@Cit:quantum_net] relies on the adiabatic evolution of a network of nonlinear oscillators. This system exhibits a bifurcation (called “simulated bifurcation”) for each nonlinear oscillator. The two branches correspond to the $-1$ and $+1$ values for each spin. The authors exploit Graphical Processing Units (GPU) and Field Programmable Gate Array (FPGA) platforms to compute the solution of these Hamiltonian systems. This method is compared to existing methods and displays orders-of-magnitude improvement. The approach is demonstrated on an Ising system with 100,000 spins. Consider the Hamiltonian that arises in Kerr-nonlinear parametric oscillators, $$\begin{aligned} H(x,y,t) &= \displaystyle\sum_{i=1}^{N}\left[\frac{K}{4}(x_{i}^2 + y_{i}^2)^2 - \frac{p(t)}{2}(x_{i}^2 - y_{i}^2) + \frac{\Delta_{i}}{2}(x_{i}^2 + y_{i}^2)\right] \nonumber\\ &-\frac{\xi_{0}}{2}\displaystyle\sum_{i=1}^N\sum_{j=1}^N J_{ij}(x_i x_j + y_i y_j). \label{eq:ham}\end{aligned}$$ Here $x_i$ and $y_i$ are position and momentum of the $i$-th oscillator respectively, $K$ is the Kerr coefficient, $p(t)$ is the parametric pumping amplitude, and $\Delta_i$ is the detuning frequency between the natural frequency of the $i$-th oscillator and half the pumping frequency. Using the standard Hamiltonian formulation, one can derive equations of motion for each oscillator. Evolving these system of equations for $x_i$ and $y_i$ converges to low energy solutions of an Ising system (Eqn. \[eq:ising\]) with high probability. Thus, the sign of $x_i$ at the end of the simulation determines the $i$-th spin. However, the above equations are computationally challenging to simulate from a numerical perspective. Instead of using the equations that arise from the “full” Hamiltonian in Eqn. \[eq:ham\], the authors (in [@Cit:quantum_net]) construct a simplified Hamiltonian of the form, $$\begin{aligned} H(x,y,t) &= \displaystyle\sum_{i=1}^{N} \frac{\Delta}{2}y_i^2 + V(x,t) \nonumber\\ &= \displaystyle\sum_{i=1}^{N} \frac{\Delta}{2}y_i^2 + \left[\frac{K}{4}x_{i}^4 + \frac{\Delta - p(t)}{2}x_{i}^2\right] - \frac{\xi_{0}}{2}\displaystyle\sum_{i=1}^N\sum_{j=1}^N J_{ij}x_i x_j. \label{eq:ham2}\end{aligned}$$ The above Hamiltonian corresponds to the following system of equations, $$\begin{aligned} \dot x_{i} &= \Delta y_i \nonumber\\ \dot y_{i} &= -\left[Kx_{i}^2 -p(t) + \Delta\right]x_{i} + \xi_{0} \displaystyle\sum_{i=1}^N\sum_{j=1}^N J_{ij}x_j. \label{eq:duffing_sys}\end{aligned}$$ It is easy to see that the above system of equations are a network of Duffing oscillators [@Cit:Gucken]. The separability of the Hamiltonian (Eqn. \[eq:ham2\]) makes the numerical integration of the system of equations easier. In particular, the authors use an explicit symplectic Euler scheme which makes it amenable for one to hard wire the resulting computational circuits on an FPGA platform. The computation proceeds as follows: all $x$ and $y$ variables are initially set to zero, $p(t)$ is then increased from $0$ and the system in Eqn. \[eq:duffing\_sys\] is evolved. The sign of the final value of $x_i$ serves as an approximation of the $i$-th spin of the associated Ising system. The system in Eqn. \[eq:duffing\_sys\] has two branches of solutions as $p(t)$ is increased from zero. It is easy to see that these branches correspond to $\pm\sqrt{p-\Delta/K}$ for each oscillator and, consequently, leads to a $2^N$ solution space. If one varies $p(t)$ slowly, the adiabatic theorem ensures that if one converges to a low energy solution for $p(t)$ close to $0$, the final solutions (for large $p(t)$) will also correspond to low energy. This method is compared to state-of-the-art approaches for two instances of the MAX-CUT problem. In the first instance, an all-to-all 2000 node MAX-CUT problem is solved on an FPGA using the above approach. The authors demonstrate that the above framework successfully converges to the best known solutions [@Cit:Goemans] very quickly. Moreover, they test the approach on a 100,000 size problem (with $5\times10^9$ edges) and find that their approach converges to the answer $100-1000$ times faster than existing software on GPU hardware. For more details of the work and associated results we refer the reader to [@Cit:quantum_net]. Thus far, we have summarized three examples in which dynamical systems theory was used to construct novel algorithms for NP-hard problems. We now discuss approaches that exploit nonlinear dynamics theory to analyze optimization algorithms for NP-hard problems. Analysis of algorithms: Koopman operators based analysis of algorithms {#sec:koopman} ====================================================================== [Overview]{} Koopman operator theory is one of the most active and exciting sub-areas within dynamical systems theory [@Cit:koop_mez; @Cit:koop_klus; @Cit:koop_kev; @Cit:koop_kutz]. The approach is based on the construction of an infinite dimensional linear operator that captures the evolution of the observables of the underlying nonlinear system. Consequently, the spectra and eigenfunctions of the operator, capture system dynamics. This methodology has been used used in a wide variety of settings, including system control and identification. An advantage of Koopman operator based methodologies is that the computations are typically based on time trace data of system evolution [@Cit:DMD]. In recent work [@Cit:koopman], Koopman operator theory was used to analyze algorithms. In particular, the authors consider optimization algorithms that evolve their state in the form of iterations. An assumption is made that the algorithm state spaces $X\subseteq \mathbb{R}^d$ are smooth $k$-dimensional Riemannian submanifolds in $d$-dimensional Euclidean spaces. A single iteration of the state $x_n$ is represented as, $$x_{n+1} = a(x_n), n\in\mathbb{N},$$ where $n$ is the iteration count. These iterative algorithms can sometimes be represented in continuous form (akin to the process used in [@Cit:Nesterov]). In other words, one can (in the limit) represent the algorithm as a continuous vector field $v:X\rightarrow\mathbb{R}^k$. If one starts at an initial condition $x_0$, the continuous time representation is of the form [@Cit:koopman], $$\frac{dc(s)}{ds} = v(c(s)),\quad c(0) = x_{0}.$$ As shown in [@Cit:koopman], this continuous form can be approximated using a Koopman operator framework. The authors then use this approach to study gradient descent and Newton-Raphson from a global dynamics perspective. Although the examples fall under the category of continuous optimization, the approach can certainly be used to study combinatorial optimization algorithms in the future. Using the same nomenclature as in [@Cit:koopman], consider a dynamical system of the form, $$\frac{dS_{t}(x_0)}{dt} = v(S_{t}(x_0)), \label{eq:dyn_algo}$$ then the family of Koopman operators $\mathscr{K}^t$ acts on the function space of observables $g:X\rightarrow\mathbb{C}$ as follows, $$[\mathscr{K}^t g](x) = (g\circ S_{t})(x). \label{eq:koopman}$$ An $L^2$ function space with an inner product is typically chosen for the space of observables. For more details, on the approach and choice of function space see [@Cit:koopman]. Note that the Koopman operator is the adjoint of the Perron-Frobenius operator [@Cit:Perron_Frob]. Letting $t=1$, without loss of generality, The Koopman operator can be expanded in terms of its spectrum, $$\mathscr{K} = \displaystyle\sum_k\lambda_k P_{\lambda_k} + \int_{\sigma_{ac}}\lambda dE(\lambda) \label{eq:spectrum}$$ where $\lambda_k$ lie in the discrete part and $\sigma_{ac}$ is the continuous spectrum of the operator. $P_{\lambda}$ and $dE(\lambda)$ are projection operators for their corresponding eigenspaces. The eigenfunctions of the Koopman operator are, $$[\mathscr{K}^t \phi_{\lambda}](x) = (\phi_{\lambda}\circ S_{t})(x) = \lambda^t\phi_{\lambda}(x).$$ Thus, one can predict the evolution of observables, $$\mathscr{K}^t g = \displaystyle\sum_{k}c_{k}\lambda^t\phi_{\lambda,k}. \label{eq:observerable}$$ The operator can be numerically approximated using a data driven approach as outlined in [@Cit:koop_kev; @Cit:koop_kutz]. The popular extended dynamic mode decomposition (EDMD) methodology introduced in [@Cit:koop_kev] is used to analyze algorithms using the Koopman operator lens [@Cit:koopman]. The EDMD approach approximates the action of the infinite dimensional operator using a finite set of real-valued functions (also called “dictionary”). In particular, given a smooth manifold $M \subset \mathbb{R}^d$ sampled by a finite set of points $X=\{x_{i}\in M\}$, the EDMD approach computes the action of the Koopman operator on the dictionary of points in $X$. The operator itself is approximated using a least squares approach [@Cit:koop_kev] as outlined below. Given a dictionary of $N_D$ observables $D=\{d_i:M\rightarrow\mathbb{R}\|\,\,i=1,\hdots,N_{D}\}$ one can define a matrix of the form $G=[d_{1}(X),d_{2}(X),\hdots,d_{N_D}(X)]$. Then the Koopman operator $\mathscr{K}^t$ can be approximated as, $$K = \frac{1}{N_X^2}(G^{T}G)^{\dagger} (A^{T}A), \label{eq:koop_num}$$ where $N_X$ is the size of the dataset and $A=\left[\mathscr{K}^t d_{1}(X), \mathscr{K}^t d_{2}(X),\hdots,\mathscr{K}^t d_{N_D}(X)\right]$. For more details see [@Cit:koop_kev]. The operator gives a local approximation of the underlying algorithm applied to a specific instance of a problem. In particular, one can use the data of a short burst of computation to compute a local approximation of the dynamics of the algorithm to accelerate its convergence. The eigenvalues, vectors, and modes are computed using EDMD. This approximation is used as a data-driven surrogate for the system to accelerate optimization. In [@Cit:koopman], the authors use the following cost function, $$f(x_{1},x_{2}) = (x_{1}^2 + x_{2} - 11)^2 + (x_{1} + x_{2}^2 -7)^2,$$ to demonstrate utility of the Koopman approach. The function has one local maximum and four local minima [@Cit:koopman]. The authors study the the popular gradient descent algorithm using the Koopman operator framework. In particular, they use radial basis functions to form a dictionary and compute $503$ eigenvalues and eigenfunctions. They show that one can construct an ergodic decomposition of the state space, thereby separating the different basins of attraction [@Cit:koopman]. The approach is able to capture the global dynamics of the algorithm in this setting, providing valuable insight regarding the performance and limitations of the algorithm. Additionally, the authors demonstrate the use of Koopman operators for studying global dynamics of algorithms in high-dimensional spaces. They take the example of a $100$-dimensional problem and show that the dynamics quickly contracts to a low dimensional subset. They demonstrate that one can accelerate the prediction of trajectories of gradient descent with high accuracy. The work concludes with the illustration of the utility of Koopman operators for analyzing the iconic Newton-Raphson method for root finding [@Cit:koopman]. For a complex polynomial of degree two, they show that the eigenfunction diverges at the roots. They also show that in cases of chaotic behavior of Newton-Raphson, one can use approximations [@Cit:Mezic_cont] of the continuous spectrum of the Koopman operator to study statistical properties of the emergent chaos. Although the above Koopman methodology was demonstrated on problems of continuous optimization, it provides a new technique with which one can study combinatorial optimization problems. We anticipate that the Koopman operator approach will be a new tool with which to study algorithms for NP-hard problems and improve their performance. Analysis of algorithms: Chaos and dynamical systems for analyzing the satisfiability (SAT) problem {#sec:dyn_sat} ================================================================================================== [Overview]{} The satisfiability problem is another iconic problem that frequently arises in the study of computational complexity theory. The challenge here is to find a satisfying assignment for a logical formula. In particular, a $k$-SAT Boolean formula $\phi(x)$ of $N$ Boolean variables and $m$ clauses, $\phi:\{0,1\}^{N}\rightarrow\{0,1\}$, is written in the conjunctive normal form (CNF) [@biere2009handbook] as follows, $$\label{eq:sat} \phi(x) = \bigwedge_{i=1}^m C_i = \bigwedge_{i=1}^m (x_{i_1} \lor x_{i_2} \lor\hdots\lor x_{i_{k}}),$$ where $x_{i_l}$ is the $l^{\rm th}$ literal in clause $C_{i}$. A SAT formula is said to be satisfiable if there exists an assignment for the binary variables $\x$ such that $\phi(\x)=1\,\, (\text{true})$. It is well known that the satisfiability problem is NP-complete [@Cit:cook]. A critical parameter associated with the satisfiability problem is the clause density $\alpha = m/N$ [@biere2009handbook]. In particular, the probability that a random $k$-SAT instance is satisfiable undergoes a phase transition as a function of $\alpha$ ($N\rightarrow\infty$) [@biere2009handbook]. The MAX-SAT problem (and the corresponding weighted version) [@krentel1988complexity] requires one to find that assignment (or assignments) that maximize the number (or the cumulative weights) of satisfied clauses. Consider a SAT formula $\phi$, then every assignment $x$ can be mapped to an “energy” $\Phi(x)$ such that, $$\label{eq:1} \Phi(\x) = \sum_{i=1}^{m} C_{i},$$ where $C_i=1$, if the $i$-th clause evaluates to $\text{true}$. In other words, the goal under the MAX-SAT problem is to find the assignment for $\x$ such that the number of satisfied clauses (or energy) is maximized. The MAX-SAT problem is harder (from a computational standpoint) than the SAT problem. In particular, it is known to be strongly NP-hard (there are no polynomial time approximation schemes). The problem of computing density of states (DOS) encompasses the the SAT and MAX-SAT problems. Classical and quantum algorithms for estimating the DOS of logical formulae were constructed in [@Cit:SAT_sahai]. In a series of seminal papers [@Cit:zoltan; @Cit:zoltan2; @Cit:zoltan3], the authors construct a dynamical systems approach to study satisfiability problems. They construct a dynamical system that computes the solutions of SAT instances. Here, the equilibria of the dynamical system correspond to literal values for which the SAT formula in Eqn. \[eq:sat\] evaluates to $\texttt{true}$. They prove that the dynamical system admits no false equilibria or limit cycles. Additionally, they relate the emergence of transient chaos and fractal boundaries with optimization hardness of the problem instance [@Cit:zoltan], pointing to a deep connection between dynamical systems theory and computational complexity. As mentioned in the overview, in [@Cit:zoltan], the authors embed SAT equations into a system of ordinary differential equations using the the following mapping: let $s_{x_{i}} = \left[-1,1\right]$, i.e. $s_{x_{i}}$ can take values between $-1$ and $1$ such that, $$\begin{aligned} s_{x_i} = \begin{cases} -1, & \text{if}\ x_i= \mbox{FALSE} \\ 1, & \text{if}\ x_i= \mbox{TRUE}. \end{cases}\end{aligned}$$ Generalizing the dynamical system for satisfiability problems constructed in [@Cit:zoltan], one can define $c_{mi}$ and ${K_m}$ as follows, $$\begin{aligned} c_{mi} = \begin{cases} -1, & \text{if}\ s_{x_i}\,\, \text{appears in negated form in m-th clause} \\ 1, & \text{if}\ s_{x_i}\,\, \text{appears in direct form in m-th clause} \\ 0, & \text{otherwise}, \end{cases}\end{aligned}$$ $$\begin{aligned} K_m(s) = 2^{-k} \prod_{j=0}^{k-1}\prod_{i=1}^{N} (1-c_{mi}s_{x_i}) \quad \forall m=1,2,\hdots,M.\end{aligned}$$ Note that $K_m(s)=0$, if and only if $m$-th clause is satisfied i.e. $c_{mi}s_{x_i}=1$ for at least one variable $x_i$ that appears in clause $m$. In [@Cit:zoltan], the authors define an energy function of the form $V(s)=\sum_{m=1}^{M}a_{m}K_{m}(s)^2$ such that $V(s^{})=0$ only at a solution $s^{}$ of the satisfiability problem. The auxiliary variables $a_{m}\in (0,\infty)$ prevent the non-solution attractors from trapping the search dynamics (for more information see [@Cit:zoltan]). In [@Cit:zoltan], the authors find that as the constraint density of the k-SAT problem increases, the trajectories of the dynamical system display intermittent chaos with fractal basin boundaries [@Cit:zoltan]. Note that, in this work, the existence of chaos is associated with the emergence of positive finite size Lyapunov exponents (FSLE) [@Cit:gucken] and the emergence of chaos corresponds to the well known phase transitions in the $k$-SAT problem [@Cit:zoltan]. In [@Cit:zoltan2], the authors further exploit the above system of equations to study the $k$-SAT problem with increasing constraint density. They find that hardness appears as a second order phase transition and discover that the resulting transient chaos displays a novel exponential-algebraic scaling. In [@Cit:zoltan3], the authors exploit the above framework to construct novel solvers for the MAX-SAT problem. This body of work demonstrates that dynamical systems theory can, in fact, be used to simultaneously study computational complexity theory and construct novel algorithms for NP-hard problems. Conclusion ========== Combinatorial optimization is a wide and important area of research with numerous applications. For decades, computer scientists have developed novel algorithms for addressing these problems. Some problems are amenable to algorithms and theory developed thus far (examples include graph routing and sorting), while others, in general, remain intractable from a computational standpoint (such as the traveling salesman problem and MAX-SAT) despite significant efforts. The classification of problems into different classes (such as NP, NP-hard, and PSPACE) and associated analyses has given rise to the field of computational complexity theory. Nonlinear dynamics, on the other hand, arises in a multitude of engineering and scientific settings. The theory has been used to explain system behavior in a diverse set of fields such as fluidics, structural mechanics, population dynamics, epidemiology, optics, and aerospace propulsion. However, the application of the theory of dynamical systems to combinatorial optimization and computational complexity remains limited. In this survey article, we summarize five recent examples of using dynamical systems theory for constructing and analyzing combinatorial optimization problems. In particular, we cover a) a novel approach for clustering graphs using the wave equation partial differential equation (PDE), b) invariant manifold computations for the traveling salesman problem, c) novel approaches for building quantum networks of Duffing oscillators to solve the MAX-CUT problem, d) applications of the Koopman operator for analyzing optimization algorithms, and e) the use of dynamical systems theory to analyze computational complexity of the SAT problem. We note that the above set of examples are not comprehensive and there are several examples that have been omitted in this survey. However, the goal of this article is not to provide a complete list of all such examples but to demonstrate that dynamical systems theory can be exploited to construct algorithms and approaches for optimization problems. Even more importantly, we hope to inspire others to extend existing dynamical systems approaches to construct the next-generation of techniques and insights for combinatorial optimization. This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA) and Space and Naval Warfare Systems Center, Pacific (SSC Pacific) under Contract No. N6600118C4031. [99.]{} Poincar[é]{}, H.: Sur le probl[è]{}me des trois corps et les [é]{}quations de la dynamique. Acta Mathematica 13, A3–A270 (1890) Turing, A.: On computable problems with an application to the entscheidungsproblem. Proc. London Math. Society 2, 42 (1936) Church, A.: A set of postulates for the foundation of logic. Annals of mathematics, 346–366 (1932) Strogatz, S. H.: Nonlinear dynamics and chaos: with applications to physics, biology, chemistry, and engineering. CRC press (2018). Walleczek, J. (Ed.): Self-organized biological dynamics and nonlinear control: toward understanding complexity, chaos and emergent function in living systems. Cambridge University Press (2006). Kevrekidis, I. G., Schmidt, L. D., and Aris, R.: On the dynamics of periodically forced chemical reactors. Chemical engineering communications, 30(6), 323-330 (1984). Holmes, P., Lumley, J. L., Berkooz, G., and Rowley, C. W. (2012). Turbulence, coherent structures, dynamical systems and symmetry. Cambridge University Press. Liu, W. M., Hethcote, H. W., and Levin, S. A.: Dynamical behavior of epidemiological models with nonlinear incidence rates. Journal of mathematical biology, 25(4), 359-380 (1987). Dellnitz, M. and Junge, O. Set oriented numerical methods for dynamical systems. Handbook of dynamical systems, 2, 221-264 (2002). Mezić, I. Analysis of fluid flows via spectral properties of the Koopman operator. Annual Review of Fluid Mechanics, 45, 357-378 (2013). Hummer, G. and Kevrekidis, I. G. Coarse molecular dynamics of a peptide fragment: Free energy, kinetics, and long-time dynamics computations. The Journal of chemical physics, 118(23), 10762-10773 (2003). Hasselblatt, B. and Katok, A. (Eds.). Handbook of dynamical systems. Elsevier (2002). Lorenz, E. N. Deterministic nonperiodic flow. Journal of the atmospheric sciences, 20(2), 130-141 (1963). Morse, M. and Hedlund, G. A. Symbolic dynamics. American Journal of Mathematics, 60(4), 815-866 (1938). Guckenheimer, J., and Holmes, P. (2013). Nonlinear oscillations, dynamical systems, and bifurcations of vector fields (Vol. 42). Springer Science & Business Media. Nesterov, Y. A method for unconstrained convex minimization problem with the rate of convergence $O (1/k^ 2)$. In Doklady an ussr (Vol. 269, pp. 543-547) (1983). Su, W., Boyd, S., and Candes, E. A differential equation for modeling Nesterov’s accelerated gradient method: Theory and insights. In Advances in Neural Information Processing Systems (pp. 2510-2518) (2014). Wibisono, A., Wilson, A. C., and Jordan, M. I. A variational perspective on accelerated methods in optimization. Proceedings of the National Academy of Sciences, 113(47), E7351-E7358 (2016). Cook, S. A. The complexity of theorem-proving procedures. In Proceedings of the third annual ACM Symposium on Theory of Computing, 151-158 (1971). Karp, R. M. Reducibility among combinatorial problems. In Complexity of computer computations, 85-103. Springer, Boston, MA (1972). Lin, S. Computer solutions of the traveling salesman problem. Bell System Technical Journal, 44(10), 2245-2269 (1965). Cook, W. J. In pursuit of the traveling salesman: Mathematics at the limits of computation. Princeton University Press (2011). Peikert, C. A decade of lattice cryptography. Foundations and Trends® in Theoretical Computer Science, 10(4), 283-424 (2016). Klus, S., and Sahai, T. A spectral assignment approach for the graph isomorphism problem. Information and Inference: A Journal of the IMA, 7(4), 689-706 (2018). Helsgaun, K. An effective implementation of the Lin–Kernighan traveling salesman heuristic. European Journal of Operational Research, 126(1), 106-130 (2000). Applegate, D., Bixby, R., Chvatal, V., and Cook, W. Concorde TSP solver (2006). Sahai, T., Speranzon, A., and Banaszuk, A. Wave equation based algorithm for distributed eigenvector computation. In 49th IEEE Conference on Decision and Control (CDC) (pp. 7308-7315) (2010, December). Sahai, T., Speranzon, A., and Banaszuk, A. Hearing the clusters of a graph: A distributed algorithm. Automatica, 48(1), 15-24 (2012). Sahai, T., Ziessler, A., Klus, S., and Dellnitz, M. Continuous relaxations for the traveling salesman problem. Nonlinear Dynamics, 97(4), 2003-2022 (2019). Goto, H., Tatsumura, K., and Dixon, A. R. Combinatorial optimization by simulating adiabatic bifurcations in nonlinear Hamiltonian systems. Science Advances, 5(4), eaav2372 (2019). Dietrich, F., Thiem, T. N., and Kevrekidis, I. G. On the Koopman operator of algorithms. arXiv preprint arXiv:1907.10807 (2019). Ercsey-Ravasz, M. and Toroczkai, Z. Optimization hardness as transient chaos in an analog approach to constraint satisfaction. Nature Physics, 7(12), 966-970 (2011). Varga, M., Sumi, R., Toroczkai, Z., and Ercsey-Ravasz, M. Order-to-chaos transition in the hardness of random Boolean satisfiability problems. Physical Review E, 93(5), 052211 (2016). Even, S. Graph algorithms. Cambridge University Press (2011). Cormen, T. H., Leiserson, C. E., Rivest, R. L., and Stein, C. Introduction to algorithms. MIT press (2009). Von Luxburg, U. A tutorial on spectral clustering. Statistics and computing, 17(4), 395-416 (2007). Wagner, D. and Wagner, F. Between min cut and graph bisection. In International Symposium on Mathematical Foundations of Computer Science (pp. 744-750). Springer, Berlin, Heidelberg (1993, August). Golub, G. H. and Van Loan, C. F. Matrix Computations Johns Hopkins University Press. Baltimore and London (1996). Kempe, D. and McSherry, F. A decentralized algorithm for spectral analysis. Journal of Computer and System Sciences, 74(1), 70-83 (2008). Kac, M. Can one hear the shape of a drum?. The American Mathematical Monthly, 73(4P2), 1-23 (1966). Klus, S., Sahai, T., Liu, C., and Dellnitz, M. An efficient algorithm for the parallel solution of high-dimensional differential equations. Journal of computational and applied mathematics, 235(9), 3053-3062 (2011). Surana, A., Sahai, T., and Banaszuk, A. Iterative methods for scalable uncertainty quantification in complex networks. International Journal for Uncertainty Quantification, 2(4) (2012). Englot, B., Sahai, T., and Cohen, I. Efficient tracking and pursuit of moving targets by heuristic solution of the traveling salesman problem. In 52nd IEEE Conference on Decision and Control (pp. 3433-3438). IEEE (2013, December). J.C. Gower and G.B. Dijksterhuis. Procrustes problems, volume 3. Oxford University Press New York (2004). Sahai, T., Klus, S., and Dellnitz, M. A Traveling Salesman Learns Bayesian Networks. arXiv preprint arXiv:1211.4888 (2012). Schönemann, P. H. On two-sided orthogonal Procrustes problems. Psychometrika, 33(1), 19-33 (1968). Dellnitz, M. and Hohmann, A. The computation of unstable manifolds using subdivision and continuation. In Nonlinear dynamical systems and chaos (pp. 449-459). Birkhäuser, Basel (1996). Goemans, M. X., and Williamson, D. P. Improved approximation algorithms for maximum cut and satisfiability problems using semidefinite programming. Journal of the ACM (JACM), 42(6), 1115-1145 (1995). Glauber, R. J. Time‐dependent statistics of the Ising model. Journal of mathematical physics, 4(2), 294-307 (1963). Guckenheimer, J. and Holmes, P. Nonlinear oscillations, dynamical systems, and bifurcations of vector fields (Vol. 42). Springer Science & Business Media (2013). Budisić, M., Mohr, R., and Mezić, I. Applied Koopmanism. Chaos: An Interdisciplinary Journal of Nonlinear Science, 22(4), 047510 (2012). Klus, S., Koltai, P., and Schütte, C. On the numerical approximation of the Perron-Frobenius and Koopman operator. arXiv preprint arXiv:1512.05997 (2015). Williams, M. O., Kevrekidis, I. G., and Rowley, C. W. A data–driven approximation of the Koopman operator: Extending dynamic mode decomposition. Journal of Nonlinear Science, 25(6), 1307-1346 (2015). Lusch, B., Kutz, J. N., and Brunton, S. L. Deep learning for universal linear embeddings of nonlinear dynamics. Nature communications, 9(1), 1-10 (2018). Kutz, J. N., Brunton, S. L., Brunton, B. W., and Proctor, J. L. Dynamic mode decomposition: data-driven modeling of complex systems. Society for Industrial and Applied Mathematics (2016). Dellnitz, M., Hohmann, A., Junge, O., and Rumpf, M. Exploring invariant sets and invariant measures. Chaos: An Interdisciplinary Journal of Nonlinear Science, 7(2), 221-228 (1997). Korda, M., Putinar, M., and Mezić, I. Data-driven spectral analysis of the Koopman operator. Applied and Computational Harmonic Analysis (2018). Molnár, B., Molnár, F., Varga, M., Toroczkai, Z., and Ercsey-Ravasz, M. A continuous-time MaxSAT solver with high analog performance. Nature communications, 9(1), 1-12 (2018). Biere, A., Heule, M., and van Maaren, H. (Eds.). Handbook of satisfiability (Vol. 185). IOS press (2009). Krentel, M. W. The complexity of optimization functions. Journal of Computer and System Sciences, 36(490-509), 10-1016 (1988). Sahai, T., Mishra, A., Pasini, J. M., and Jha., S. Estimating the Density of States of Boolean Satisfiability Problems on Classical and Quantum Computing Platforms. 34th AAAI Conference on Artificial Intelligence (AAAI) 2020. [^1]: Note that in the case of spectral clustering we desire to compute the smallest $k$ eigenvectors of $\L$. The algorithm is still applicable if we consider the matrix $\I-\L$. [^2]: The trace of a matrix $ A \in \R^{n \times n} $ is defined to be the sum of all diagonal entries, i.e., $ \mbox{tr}(A) = \sum_{i=1}^n a_{ii} $.	{ "pile_set_name": "ArXiv" }
--- abstract: \| In this note we classify all triples $(a,b,i)$ such that there is a convex lattice polygon $P$ with area $a$ which has $b$ and $i$ lattice points on the boundary and in the interior, respectively. The crucial lemma for the classification is the necessity of $b \le 2 \ i + 7$. We sketch three proofs of this fact: the original one by Scott [@Scott], an elementary one, and one using algebraic geometry. As a refinement, we introduce an onion skin parameter $\ell$: how many nested polygons does $P$ contain? and give sharper bounds. address: - \| Fachbereich Mathematik & Informatik\ Freie Universität Berlin\ 14195 Berlin\ Germany - \| RICAM\ Austrian Academy of Sciences\ Altenberger Straße 69\ 4040 Linz\ Austria author: - Christian Haase - Josef Schicho bibliography: - 'alles.bib' - 'josef.bib' title: 'Lattice polygons and the number $\mathbf{2i+7}$' --- [^1] Introduction ============ How it all began ---------------- When the second author translated a result on algebraic surfaces into the language of lattice polygons using toric geometry, he obtained an inequality for lattice polygons. This inequality had originally been discovered by Scott [@Scott]. The first author then found a third proof. Subsequently, both authors went through a phase of polygon addiction. Once you get started to draw lattice polygons on graph paper and to discover relations between their numerical invariants, it is not so easy to stop! (The gentle reader has been warned.) Thus, it was just unavoidable that the authors came up with new inequalities: Scott’s inequality can be sharpened if one takes another invariant into account, which is defined by peeling off the skins of the polygons like an onion (see Section \[sec:onions\]). Lattice polygons ---------------- We want to study convex lattice polygons: convex polygons all whose vertices have integral coordinates. As it turns out, we need to consider nonconvex polygons as well. Even nonsimple polygons – polygons with self intersection – will prove useful later on. In what follows, we will abbreviate “polygon” := “convex lattice polygon”, and we will emphasize when we allow nonintegral or nonsimple situations. \#1\#2\#3\#4\#5[ @font ]{} (7366,2581)(0,-10) (7283,1283) (7283,1283) (7283,1883) (7283,1883) (7283,683) (7283,683) (6083,1283) (6083,1283) (6083,1883) (6083,1883) (6083,683) (6083,683) (6683,1283) (6683,1283) (6683,1883) (6683,1883) (6683,683) (6683,683) (7283,1283)(6083,1883) (6324.495,1829.334)(6083.000,1883.000)(6270.830,1722.003) (6083,1883)(6683,1883) (6443.000,1823.000)(6683.000,1883.000)(6443.000,1943.000) (6683,1883)(6083,683) (6136.666,924.495)(6083.000,683.000)(6243.997,870.830) (6083,683)(7283,1283) (7095.170,1122.003)(7283.000,1283.000)(7041.505,1229.334) (3083,1283) (3083,1283) (3083,1883) (3083,1883) (3083,683) (3083,683) (3683,1283) (3683,1283) (3683,1883) (3683,1883) (3683,683) (3683,683) (3683,2483) (3683,2483) (3083,2483) (3083,2483) (3683,83) (3683,83) (3083,83) (3083,83) (83,1283) (83,1283) (83,1883) (83,1883) (83,683) (83,683) (683,1283) (683,1283) (683,1883) (683,1883) (683,683) (683,683) (683,2483) (683,2483) (83,2483) (83,2483) (683,83) (683,83) (83,83) (83,83) (1283,1283) (1283,1283) (1283,1883) (1283,1883) (1283,683) (1283,683) (1883,1283) (1883,1283) (1883,1883) (1883,1883) (1883,683) (1883,683) (1883,2483) (1883,2483) (1283,2483) (1283,2483) (1883,83) (1883,83) (1283,83) (1283,83) (4283,1283) (4283,1283) (4283,1883) (4283,1883) (4283,683) (4283,683) (4283,2483) (4283,2483) (4883,2483) (4883,2483) (4883,1883) (4883,1883) (4883,1283) (4883,1283) (4883,683) (4883,683) (4883,83) (4883,83) (4283,83) (4283,83) (4883,2483)(4283,1883) (4410.279,2095.132)(4283.000,1883.000)(4495.132,2010.279) (4283,1883)(3083,1883) (3323.000,1943.000)(3083.000,1883.000)(3323.000,1823.000) (3083,1883)(4883,83) (4670.868,210.279)(4883.000,83.000)(4755.721,295.132) (4883,83)(4883,2483) (4943.000,2243.000)(4883.000,2483.000)(4823.000,2243.000) (1583,83)(83,83) (323.000,143.000)(83.000,83.000)(323.000,23.000) (383,2258)(1883,1283) (1649.074,1363.491)(1883.000,1283.000)(1714.473,1464.104) (83,83)(373,2187) (399.668,1941.055)(373.000,2187.000)(280.792,1957.440) (1883,1283)(1592,124) (1592.251,371.386)(1592.000,124.000)(1708.639,342.164) Denote the area enclosed by a polygon $P$ by $a=a(P)$, the number of lattice points on the boundary by $b=b(P)$, and the number of lattice points strictly inside of $P$ by $i=i(P)$. A classic result relates these data. $$\label{eq:pick} a = i + \frac{b}{2} - 1$$ A thorough discussion of this theorem – including an application in forest industry! – can be found in [@Gruenbaum_Shepard:93]. Pick’s theorem is not the only relation between the three parameters $a$, $b$, $i$ of polygons. There is the rather obvious constraint $b\ge 3$. From Pick’s formula we obtain immediately $a \ge i + \frac{1}{2}$ and $a \ge \frac{b}{2} - 1$. Are there other constraints? For the sake of suspense, we do not want to reveal the final inequalities just yet. We refer the impatient reader to the conclusion in Section \[sec:conclusion\] which contains a summary of the main results. Lattice equivalence {#sec:equivalence} ------------------- Clearly, the area $a(P)$ is invariant under rigid motions of the plane. On the other hand, the numbers $i(P)$, $b(P)$ are not concepts of Euclidean geometry, because they are not preserved by rigid motions. But they [are]{} preserved under [ lattice equivalences]{}: affine maps $\Phi \colon {{\mathbb R}}^2 \rightarrow {{\mathbb R}}^2$ of the plane that restrict to isomorphisms of the lattice ${{\mathbb Z}}^2$. Orientation preserving lattice equivalences form a group, the semi direct product ${\operatorname{SL_2 {{\mathbb Z}}}}\ltimes {{\mathbb Z}}^2$. \#1\#2\#3\#4\#5[ @font ]{} (2566,1981)(0,-10) (683,83)(83,83)(83,683) (683,683)(683,83) (683,83)(83,83)(83,683) (683,683)(683,83) (1883,1883) (1883,1883) (1283,1883) (1283,1883) (1883,683) (1883,683) (683,683) (683,683) (1883,1283) (1883,1283) (2483,1283) (2483,1283) (1283,1283) (1283,1283) (83,683) (83,683) (683,83) (683,83) (1283,83) (1283,83) (683,1283) (683,1283) (83,83) (83,83) (2483,1883) (2483,1883) (1883,83) (1883,83) (2483,83) (2483,83) (2483,683) (2483,683) (83,1283) (83,1283) (83,1883) (83,1883) (683,1883) (683,1883) (1283,683) (1283,683) \#1\#2\#3\#4\#5[ @font ]{} (2566,1981)(0,-10) (1883,1283)(83,83)(683,683) (2483,1883)(1883,1283) (1883,1283)(83,83)(683,683) (2483,1883)(1883,1283) (1883,1883) (1883,1883) (1283,1883) (1283,1883) (1883,683) (1883,683) (683,683) (683,683) (1283,683) (1283,683) (1883,1283) (1883,1283) (2483,1283) (2483,1283) (1283,1283) (1283,1283) (83,683) (83,683) (683,83) (683,83) (1283,83) (1283,83) (683,1283) (683,1283) (83,83) (83,83) (2483,1883) (2483,1883) (1883,83) (1883,83) (2483,83) (2483,83) (2483,683) (2483,683) (83,1283) (83,1283) (83,1883) (83,1883) (683,1883) (683,1883) So $\Phi$ has the form $\Phi({\mathbf{x}}) = A {\mathbf{x}} + {\mathbf{y}}$ for a matrix $A$, and a vector ${\mathbf{y}}$. The lattice preservation property $\Phi({{\mathbb Z}}^2)={{\mathbb Z}}^2$ implies that both $A$ and ${\mathbf{y}}$ have integral entries, and the same is true for the inverse transformation $\Phi^{-1}({\mathbf{x}}) = A^{-1} {\mathbf{x}} - A^{-1} {\mathbf{y}}$. Hence $\det A = \pm 1$, and $a(P)$ is preserved under $\Phi$ as well. In all our arguments, we will treat lattice equivalent polygons as indistinguishable. For example, the quadrangle in Figure \[fig:square\] on the right looks to us like a perfect square. We see that angles and Euclidian lengths are not preserved. A lattice geometric substitute for the length of a lattice line segment is the number of lattice points it contains minus one. In this sense, $b$ is the perimeter of $P$. Here is an exercise that helps to get a feeling for what lattice equivalences can do and cannot do.[^2] \[ex:angle\] Given a vertex ${\mathbf{x}}$ of a polygon $P$, show that there is a unique orientation preserving lattice equivalence $\Phi$ so that - $\Phi({\mathbf{x}})=(0,0){^t}$, and - there are (necessarily unique) coprime $0 < p \le q$ so that the segments $[(1,0){^t},(0,0){^t}]$ and $[(0,0){^t}, (-p,q){^t}]$ are contained in edges of $\Phi(P)$. Why algebraic geometry? ----------------------- Toric geometry is a powerful link connecting discrete and algebraic geometry (see e.g. [@Sturmfels:99]). At the heart of this link is the simple correspondence $$\begin{aligned} \text{ lattice point \qquad } & & \text{ \qquad Laurent monomial } \\ {\mathbf{p}}=(p_1,\ldots,p_m) \in {{\mathbb Z}}^m & \longleftrightarrow & {\mathbf{x}}^{{\mathbf{p}}} = x_1^{p_1} \cdot\ldots\cdot x_m^{p_m} \in \mathbb{C}[x_1^{\pm1},\ldots,x_m^{\pm1}]\end{aligned}$$ It was invented by M. Demazure [@Demazure:70] for a totally different purpose (to study algebraic subgroups of the Cremona group) in algebraic geometry. R. Stanley used it in combinatorics to classify the possible face numbers of simplicial convex polytopes [@gThmS]. R. Krasauskas [@Krasauskas:02] used it in geometric modeling to construct surfaces with new control structure (see Figure \[fig:rimas\]). ![Toric surface with hexagonal control structure.[]{data-label="fig:rimas"}](rimas.ps){height="6cm" width="8cm"} For any polygon $P$, the Laurent monomials corresponding to its lattice points define a [toric surface]{} $X_P$ in a projective space of dimension $b+i-1$ as follows. Number the lattice points $P \cap {{\mathbb Z}}^2=\{{\mathbf{p}}_0,\dots,{\mathbf{p}}_n\}$ (where $n=b+i-1$). Then $X_P$ is the closure of the image of the map $({\mathbb C}^)^2\to{\mathbb P}^n$ defined by ${\mathbf{x}} \mapsto ({\mathbf{x}}^{{\mathbf{p}}_0} : \dots : {\mathbf{x}}^{{\mathbf{p}}_n})$. Lattice equivalent polygons define the same toric surface. As to be expected, there is a dictionary translating toric geometry to lattice geometry: the degree of the toric surface is equal to twice the area, and the number of interior points is equal to the [sectional genus]{} of the surface. For instance, let $\Gamma$ be the triangle with corners $(0,0)^t,(1,2)^t,(2,1)^t$ (it has one interior point $(1,1)^t$). Then the toric surface is given by $(1 : x_1x_2^2 : x_1^2x_2 : x_1x_2) \in {{\mathbb P}}^3$, its degree is 3 – this is also reflected by its implicit equation $y_1y_2y_3-y_4^3=0$, which has also degree $3$ – and its sectional genus is $1$, i.e., if we intersect with a generic hyperplane in ${{\mathbb P}}^3$, we obtain a genus one Riemann surface. In the context of toric geometry, Pick’s formula appears as a consequence of the Riemann-Roch Theorem. Examples ======== Let us approach the question which parameters are possible for polygons by looking at some examples. Can we bound $i$ or $a$ in terms of $b$? Figure \[fig:bequal3\] shows examples with $b=3$ and arbitrarily high $a$ and $i$. So there is no lattice geometric analogue of the isoperimetric inequality. \#1\#2\#3\#4\#5[ @font ]{} (9766,3181)(0,-10) (1883,983) (1883,983) (2483,983) (2483,983) (3083,983) (3083,983) (683,983) (683,983) (1283,983) (1283,983) (83,2183) (83,2183) (3683,2183) (3683,2183) (3683,1583) (3683,1583) (83,983) (83,983) (83,1583) (83,1583) (1883,1583) (1883,1583) (2483,1583) (2483,1583) (3083,1583) (3083,1583) (683,1583) (683,1583) (1283,1583) (1283,1583) (1883,2183) (1883,2183) (2483,2183) (2483,2183) (3083,2183) (3083,2183) (683,2183) (683,2183) (1283,2183) (1283,2183) (3683,983) (3683,983) (8483,1883) (8483,1883) (9083,1883) (9083,1883) (9683,1883) (9683,1883) (7283,1883) (7283,1883) (7883,1883) (7883,1883) (6683,1883) (6683,1883) (8483,2483) (8483,2483) (9083,2483) (9083,2483) (9683,2483) (9683,2483) (7283,2483) (7283,2483) (7883,2483) (7883,2483) (6683,2483) (6683,2483) (8483,3083) (8483,3083) (9083,3083) (9083,3083) (9683,3083) (9683,3083) (7283,3083) (7283,3083) (7883,3083) (7883,3083) (6683,3083) (6683,3083) (8483,1283) (8483,1283) (9083,1283) (9083,1283) (9683,1283) (9683,1283) (8483,83) (8483,83) (9083,83) (9083,83) (9683,83) (9683,83) (8483,683) (8483,683) (9083,683) (9083,683) (9683,683) (9683,683) (7283,1283) (7283,1283) (7283,83) (7283,83) (7283,683) (7283,683) (7883,1283) (7883,1283) (7883,683) (7883,683) (6683,1283) (6683,1283) (6683,83) (6683,83) (6683,683) (6683,683) (7883,83) (7883,83) (83,983)(683,2183)(3683,1583)(83,983) (7283,3083)(9683,83)(6683,683)(7283,3083) What about bounds in the opposite direction? Can we bound $b$ in terms of $i$? Well, there is the family of Figure \[fig:iequal0\] with $i=0$ and arbitrary $b$. \#1\#2\#3\#4\#5[ @font ]{} (3766,781)(0,-10) (83,683) (83,683) (3683,683) (3683,683) (3683,83) (3683,83) (83,83) (83,83) (1883,83) (1883,83) (2483,83) (2483,83) (3083,83) (3083,83) (683,83) (683,83) (1283,83) (1283,83) (1883,683) (1883,683) (2483,683) (2483,683) (3083,683) (3083,683) (683,683) (683,683) (1283,683) (1283,683) (83,683)(83,83)(3683,83)(83,683) Perhaps surprisingly, for $i>0$ no such families exist. For $i=1$, there are precisely the $16$ lattice equivalence classes depicted in Figure \[fig:reflexive\]. We see that all values $3 \le b \le 9 = 2i+7$ occur. The polygon labeled $3\Delta$ is the $3$-fold dilation of the standard triangle $\Delta$ which is the convex hull ${\mbox{\rm conv}}[(0,0){^t}, (1,0){^t}, (0,1){^t}]$ of the origin together with the standard unit vectors. It will play an important rôle later on. \#1\#2\#3\#4\#5[ @font ]{} (10366,4981)(0,-10) (83,1883)(83,83)(1883,83)(83,1883) (83,1883)(83,83)(1883,83)(83,1883) (83,4883) (83,4883) (1283,4883) (1283,4883) (2483,4883) (2483,4883) (3683,4883) (3683,4883) (4883,4883) (4883,4883) (6083,4883) (6083,4883) (7283,4883) (7283,4883) (8483,4883) (8483,4883) (9683,4883) (9683,4883) (683,4883) (683,4883) (1883,4883) (1883,4883) (3083,4883) (3083,4883) (4283,4883) (4283,4883) (5483,4883) (5483,4883) (6683,4883) (6683,4883) (7883,4883) (7883,4883) (9083,4883) (9083,4883) (10283,4883) (10283,4883) (83,4283) (83,4283) (1283,4283) (1283,4283) (2483,4283) (2483,4283) (3683,4283) (3683,4283) (4883,4283) (4883,4283) (6083,4283) (6083,4283) (7283,4283) (7283,4283) (8483,4283) (8483,4283) (9683,4283) (9683,4283) (683,4283) (683,4283) (1883,4283) (1883,4283) (3083,4283) (3083,4283) (4283,4283) (4283,4283) (5483,4283) (5483,4283) (6683,4283) (6683,4283) (7883,4283) (7883,4283) (9083,4283) (9083,4283) (10283,4283) (10283,4283) (83,3683) (83,3683) (1283,3683) (1283,3683) (2483,3683) (2483,3683) (3683,3683) (3683,3683) (4883,3683) (4883,3683) (6083,3683) (6083,3683) (7283,3683) (7283,3683) (8483,3683) (8483,3683) (9683,3683) (9683,3683) (683,3683) (683,3683) (1883,3683) (1883,3683) (3083,3683) (3083,3683) (4283,3683) (4283,3683) (5483,3683) (5483,3683) (6683,3683) (6683,3683) (7883,3683) (7883,3683) (9083,3683) (9083,3683) (10283,3683) (10283,3683) (83,3083) (83,3083) (1283,3083) (1283,3083) (2483,3083) (2483,3083) (3683,3083) (3683,3083) (4883,3083) (4883,3083) (6083,3083) (6083,3083) (7283,3083) (7283,3083) (8483,3083) (8483,3083) (9683,3083) (9683,3083) (683,3083) (683,3083) (1883,3083) (1883,3083) (3083,3083) (3083,3083) (4283,3083) (4283,3083) (5483,3083) (5483,3083) (6683,3083) (6683,3083) (7883,3083) (7883,3083) (9083,3083) (9083,3083) (10283,3083) (10283,3083) (83,2483) (83,2483) (1283,2483) (1283,2483) (2483,2483) (2483,2483) (3683,2483) (3683,2483) (4883,2483) (4883,2483) (6083,2483) (6083,2483) (7283,2483) (7283,2483) (8483,2483) (8483,2483) (9683,2483) (9683,2483) (683,2483) (683,2483) (1883,2483) (1883,2483) (3083,2483) (3083,2483) (4283,2483) (4283,2483) (5483,2483) (5483,2483) (6683,2483) (6683,2483) (7883,2483) (7883,2483) (9083,2483) (9083,2483) (10283,2483) (10283,2483) (83,1883) (83,1883) (1283,1883) (1283,1883) (2483,1883) (2483,1883) (3683,1883) (3683,1883) (4883,1883) (4883,1883) (6083,1883) (6083,1883) (7283,1883) (7283,1883) (8483,1883) (8483,1883) (9683,1883) (9683,1883) (683,1883) (683,1883) (1883,1883) (1883,1883) (3083,1883) (3083,1883) (4283,1883) (4283,1883) (5483,1883) (5483,1883) (6683,1883) (6683,1883) (7883,1883) (7883,1883) (9083,1883) (9083,1883) (10283,1883) (10283,1883) (83,1283) (83,1283) (1283,1283) (1283,1283) (2483,1283) (2483,1283) (3683,1283) (3683,1283) (4883,1283) (4883,1283) (6083,1283) (6083,1283) (7283,1283) (7283,1283) (8483,1283) (8483,1283) (9683,1283) (9683,1283) (683,1283) (683,1283) (1883,1283) (1883,1283) (3083,1283) (3083,1283) (4283,1283) (4283,1283) (5483,1283) (5483,1283) (6683,1283) (6683,1283) (7883,1283) (7883,1283) (9083,1283) (9083,1283) (10283,1283) (10283,1283) (83,683) (83,683) (1283,683) (1283,683) (2483,683) (2483,683) (3683,683) (3683,683) (4883,683) (4883,683) (6083,683) (6083,683) (7283,683) (7283,683) (8483,683) (8483,683) (9683,683) (9683,683) (683,683) (683,683) (1883,683) (1883,683) (3083,683) (3083,683) (4283,683) (4283,683) (5483,683) (5483,683) (6683,683) (6683,683) (7883,683) (7883,683) (9083,683) (9083,683) (10283,683) (10283,683) (83,83) (83,83) (1283,83) (1283,83) (2483,83) (2483,83) (3683,83) (3683,83) (4883,83) (4883,83) (6083,83) (6083,83) (7283,83) (7283,83) (8483,83) (8483,83) (9683,83) (9683,83) (683,83) (683,83) (1883,83) (1883,83) (3083,83) (3083,83) (4283,83) (4283,83) (5483,83) (5483,83) (6683,83) (6683,83) (7883,83) (7883,83) (9083,83) (9083,83) (10283,83) (10283,83) (683,4883)(1283,4283)(83,3683)(683,4883) (1883,3683)(2483,4883)(3083,3683)(1883,3683) (3683,4283)(4283,3683)(4883,3683) (4283,4883)(3683,4283) (5483,4883)(5483,3683)(7283,3683) (6083,4883)(5483,4883) (7883,3683)(9083,4883)(10283,3683)(7883,3683) (83,3083)(83,2483)(683,1883) (1283,1883)(1283,2483)(683,3083)(83,3083) (1883,2483)(2483,3083)(3083,2483) (3083,1883)(2483,1883)(1883,2483) (3683,1883)(3683,3083)(4883,3083) (4883,2483)(4283,1883)(3683,1883) (5483,1883)(6083,3083)(6683,3083) (6683,1883)(5483,1883) (7283,1883)(7283,2483)(7883,3083) (8483,2483)(8483,1883)(7283,1883) (9083,1883)(9083,3083)(10283,3083) (10283,1883)(9083,1883) (2483,83)(2483,683)(3083,1283) (4283,83)(2483,83) (4883,83)(5483,1283)(6683,83)(4883,83) (7283,83)(7883,1283)(8483,683) (8483,83)(7283,83) (9083,683)(9683,1283)(10283,683) (9683,83)(9083,683) (233,308)[(0,0)\[lb\]]{} What can we do for $i \ge 2$? The family shown in Figure \[fig:shoe\] yields all $4\le b\le 2i+6$. In fact, Scott [@Scott] showed that $2i+6$ is how far we can get. \#1\#2\#3\#4\#5[ @font ]{} (6766,2211)(0,-10) (1883,1083) (1883,1083) (2483,1083) (2483,1083) (3083,1083) (3083,1083) (683,1083) (683,1083) (1283,1083) (1283,1083) (3683,1083) (3683,1083) (83,1083) (83,1083) (1883,1683) (1883,1683) (2483,1683) (2483,1683) (3083,1683) (3083,1683) (683,1683) (683,1683) (1283,1683) (1283,1683) (3683,1683) (3683,1683) (83,1683) (83,1683) (1883,483) (1883,483) (2483,483) (2483,483) (3083,483) (3083,483) (683,483) (683,483) (1283,483) (1283,483) (3683,483) (3683,483) (83,483) (83,483) (6083,1083) (6083,1083) (6683,1083) (6683,1083) (4883,1083) (4883,1083) (5483,1083) (5483,1083) (4283,1083) (4283,1083) (6083,1683) (6083,1683) (6683,1683) (6683,1683) (4883,1683) (4883,1683) (5483,1683) (5483,1683) (4283,1683) (4283,1683) (6083,483) (6083,483) (6683,483) (6683,483) (4883,483) (4883,483) (5483,483) (5483,483) (4283,483) (4283,483) (83,483)(4883,483)(6683,1083) (83,1683)(83,483) (83,108)[(0,0)\[b\]]{} (6683,1308)[(0,0)\[b\]]{} (83,1908)[(0,0)\[b\]]{} (4883,108)[(0,0)\[b\]]{} Scott’s proof is elementary and short enough to be included in this paper. We give two other proofs for the same result. One of them uses toric geometry; it is merely the observation that a well-known inequality [@Schicho:98d Theorem 6] in algebraic geometry translates to Scott’s inequality when applied to toric surfaces. The third proof is again elementary, and it was the search for this proof which sparked polygon addiction in the first author. For the inequality $b\le 2\ i +6$, we have arbitrary large examples where equality holds (see Figure \[fig:shoe\]); but for all these examples, all interior points are collinear. Under the additional assumption that the interior points are not collinear, the inequality can be strengthened to $b \le i + 9$ (see the remark after Lemma \[lemma:step\]). The coefficient in front of the $i$ can be improved further by introducing the [level]{} of a polygon: roughly speaking, this is the number of times one can pass to the convex hull of the interior lattice points. Before we really get going, here is a little caveat. Most of our considerations break down in dimension $3$. Pick’s formula has no analogue. Already tetrahedra with no boundary or interior lattice points except the vertices can have arbitrary volume. This was first pointed out by J. Reeve [@Reeve] (see Figure \[fig:reeve\]). Nevertheless, the phenomenon that for given $i>0$, the volume is bounded occurs in arbitrary dimension [@lagariasZiegler]. ![Reeve’s simplices[]{data-label="fig:reeve"}](reeve3.eps) Three proofs of $\mathbf{b \le 2i+7}$ ===================================== Let $P$ be a polygon with interior lattice points. Denote $a$ its surface area, $i$ the number of interior lattice points, and $b$ the number of lattice points on $P$’s boundary. In view of Pick’s Theorem , the following three inequalities are equivalent. \[prop:main\] If $i>0$, then $$\begin{aligned} b \ &\le \ 2 \ i \ + \ 7 \label{eq:main:bi} \\ a \ &\le \ 2 \ i \ + \ 5/2 \label{eq:main:ai} \\ b \ &\le \ a \ + \ 9/2 \label{eq:main:ab} \end{aligned}$$ with equality only for the triangle $3\Delta$ in Figure \[fig:reflexive\]. Scott’s proof {#ssec:scott} ------------- Apply lattice equivalences to $P$ so that $P$ fits tightly into a box $[0,p'] \times [0,p]$ with $p$ as small as possible. Then $2 \le p \le p'$ (remember, $i>0$). If $P$ intersects the top and the bottom edge of the box in segments of length $q \ge 0$ and $q' \ge 0$ respectively, then (See Figure \[fig:box\].) $$\begin{aligned} b &\le q+q'+2p \ , \ \text{ and } \label{eq:sc1} \\ a &\ge p\,(q+q')/2 \ . \label{eq:sc2}\end{aligned}$$ (0,0)![$P$ in a box.[]{data-label="fig:box"}](box.pstex "fig:") \#1\#2\#3\#4\#5[ @font ]{} (3383,2489)(518,-6967) (3901,-5761)[(0,0)\[lb\]]{} (1876,-6211)[(0,0)\[lb\]]{} (1726,-6886)[(0,0)\[lb\]]{} (1876,-5011)[(0,0)\[lb\]]{} We distinguish three cases - $p=2$, or $q+q' \ge 4$, or $p=q+q'=3$ - $p=3$, and $q+q' \le 2$ - $p \ge 4$, and $q+q' \le 3$. The above inequalities , and are already sufficient to deal with the first two cases.\ ($i$) We have $$\begin{aligned} 2b-2a \ &\le \ 2(q+q'+2p) - p\,(q+q') \\ &= \ (q+q'-4)(2-p) + 8 \quad \le \ 9 ,\end{aligned}$$ which shows in Proposition \[prop:main\]. (With equality if and only if $p=q+q'=3$, $a=9/2$, $b=9$.)[^3]\ ($ii$) The estimate $b \le q+q'+2p \le 8$ together with $i \ge 1$ show that inequality in Proposition \[prop:main\] is strictly satisfied.\ ($iii$) The only case where we have to work a little is case three. Choose points ${\mathbf{x}}=(x_1,p)^t$, ${\mathbf{x}}'=(x'_1,0)^t$, ${\mathbf{y}}=(0,y_2)^t$, and ${\mathbf{y}}'=(p',y'_2)^t$ in $P$ so that $\delta = \|x_1-x'_1\|$ is as small as possible. Then $a \ge p (p'- \delta)/2$ (see Figure \[fig:case3\]). (0,0)![[]{data-label="fig:case3"}](case3a.pstex "fig:") \#1\#2\#3\#4\#5[ @font ]{} (7080,5691)(436,-5329) (7501,-2911)[(0,0)\[lb\]]{} (451,-1111)[(0,0)\[lb\]]{} (2776,-61)[(0,0)\[lb\]]{} (4051,-5161)[(0,0)\[lb\]]{} (3601,-2536)[(0,0)\[b\]]{} (0,0)![[]{data-label="fig:case3"}](case3b.pstex "fig:") \#1\#2\#3\#4\#5[ @font ]{} (6024,4911)(1189,-5260) (4051,-5161)[(0,0)\[lb\]]{} Now the task is to apply lattice equivalences so that $\delta$ becomes small. After applying a lattice equivalence of the form $\left[ \begin{smallmatrix} 1&k\\0&1 \end{smallmatrix} \right]$ it is possible to choose $\delta \le (p-q-q')/2$. This lattice equivalence will leave $q$, $q'$, $p$ unchanged, because it fixes the $x_1$-axis. We still have $p \le p'$ because $p$ was supposed to be minimal. Thus, we obtain $$\label{eq:sc3} a \ge p(p+q+q')/4 ,$$ and $$\begin{gathered} 4(b-a) \le 8p+4q+4q' \ - \ p(p+q+q') \\ = p(8-p) - (p-4)(q+q') \le p(8-p) \le 16\end{gathered}$$ because $p\ge4$ in case three. This proves that inequality in Proposition \[prop:main\] is strictly satisfied. Clipping off vertices --------------------- This proof proceeds by induction on $i$. If $i=1$, we can check the inequalities on all $16$ lattice equivalence classes of such $P$. (See Figure \[fig:reflexive\].) For the induction step, we want to “chop off a vertex”. If $i \ge 2$, and $b \le 10$, nothing is to show. So assume $b \ge 11$. By applying a lattice equivalence, we may assume without loss of generality that ${{\mathbf{0}}}$ and $(1,0){^t}$ lie in the interior of $P$. Reflect in the $x_1$-axis if necessary in order to assure that there are $\ge 5$ boundary lattice points with positive second coordinate. First, suppose there is a vertex $v$ with positive second coordinate which is not unimodular. That is, the triangle formed by $v$ together with its two neighboring lattice points $v'$ and $v''$ on the boundary has area $> 1/2$. Denote $P'$ the convex hull ${\mbox{\rm conv}}(P \cap {{\mathbb Z}}^2 \setminus \{v\})$. This omission affects our parameters as follows: $b' = b+k-2$, $i' = i - k + 1$, and, by Pick’s formula, $a' = a-k/2$. Here $k$ is the lattice length of the boundary of $P'$ that is visible from $v$. Because $v$ was not unimodular, there is an additional lattice point in the triangle $vv'v''$. Thus, we have $k \ge 2$. Because there are other lattice points with positive second coordinate, at least one of ${{\mathbf{0}}}$ or $(1,0){^t}$ remains in the interior of $P'$, and we can use induction. ![Clipping off a nonunimodular vertex.[]{data-label="fig:omit1"}](omit1.eps) Now, if all vertices with positive second coordinate are unimodular, similiarly omit one vertex $v$ together with its two boundary neighbors $v'$ and $v''$: $P'={\mbox{\rm conv}}(P \cap {{\mathbb Z}}^2 \setminus \{v,v',v''\})$. The parameters change as follows: $b' = b+k-4$, $i' = i - k + 1$, and $a' = a-k/2-1$, where $k$ is the lattice length of the boundary of $P'$ that is visible from the removed points. In order to see that $k \ge 2$, observe that the point $v'''=v'+v''-v$ belongs to the interior of $P$, and the two adjacent segments of $P'$ are both visible from the removed points. As observed above, there remain lattice points with positive second coordinate in $P'$ so that at least one of ${{\mathbf{0}}}$ or $(1,0){^t}$ stays in the interior of $P'$. $\Box$ (0,0)![Clipping off a unimodular vertex (and its neighbors).[]{data-label="fig:omit2"}](omit2.pstex "fig:") \#1\#2\#3\#4\#5[ @font ]{} (3174,3099)(814,-4348) (2401,-2011)[(0,0)\[b\]]{} (1201,-2686)[(0,0)\[rb\]]{} (1726,-2086)[(0,0)\[rb\]]{} (1876,-3136)[(0,0)\[lb\]]{} Algebraic geometry ------------------ We use the letters $d$ and $p$ to denote the degree and the sectional genus of an algebraic surface. The inequality $p\le (d-1)(d-2)/2$ holds for arbitrary algebraic surfaces. If the surface is rational, i.e. if it has a parametrization by rational functions, then there are more inequalities. \[thm:ag\] - If $p=1$, then $d\le 9$. - If $p\ge 2$, then $d\le 4p+4$. Rational surfaces with $p=1$ are called Del Pezzo surfaces. The degree bound $9$ is due to del Pezzo [@Pezzo:87]. The bound $d\le 4p+4$ was shown by Jung [@Jung:90], hence this proof is actually the oldest one! A modern proof can be found in Schicho [@Schicho:98d]. Toric surfaces are rational, and Scott’s inequality is equivalent to Theorem \[thm:ag\] for toric surfaces. $\Box$ Onion skins {#sec:onions} =========== ![image](zwiebel3.eps){height="43mm"} In flatland, take a solid polygon $P$ into your hand and peel off the shell. You get another convex polygon $P^{(1)}$, the convex hull of its interior lattice points. Except, of course, if $i=0$ then $P$ was an empty nut, and if all interior lattice points are collinear then $P^{(1)}$ is a “degenerate polygon”, namely a line segment or a single point. Repeat this process as long as it is possible, peeling off the skins of the polygon one after the other: $P^{(k+1)} := (P^{(k)})^{(1)}$. After $n$ steps you arrive at the nucleus $P^{(n)}$, which is either a degenerate polygon or an empty nut. We define the [level]{} $\ell=\ell(P)$ in the following way: - $\ell(P)=n$ if the nucleus is a degenerate polygon, - $\ell(P)=n+1/3$ if the nucleus is $\Delta$, - $\ell(P)=n+2/3$ if the nucleus is $2\Delta$, and - $\ell(P)=n+1/2$ if the nucleus is any other empty nut. Here $\Delta$ stands for (a polygon lattice equivalent to) the standard triangle ${\mbox{\rm conv}}[(0,0){^t}, (1,0){^t}, (0,1){^t}]$. The purpose of this weird definition is to ensure the second statement in the exercise below. Show that $\ell$ is uniquely defined by - $\ell(P) = \ell(P^{(1)})+1$ if $P^{(1)}$ is $2$-dimensional, and - $\ell(kP)=k\ell(P)$ for positive integers $k$. ![Polygons of levels []{data-label="fig:P(ell)"}](cases.eps) The level of a polygon is an analogue of the radius of the in-circle in Euclidean geometry. There we have the equation $2a=\ell b$. In lattice geometry, we have an inequality. \[thm:onion\] Let $P$ be a convex lattice polygon of area $a$ and level $\ell \ge 1$ with $b$ and $i$ lattice points on the boundary and in the interior, respectively. Then $(2\ell-1) b \le 2i + 9\ell^2 - 2$, or equivalently $2\ell b\le 2a+9\ell^2$, or equivalently $(4l-2)a\le 9\ell^2+4l(i-1)$, with equality if and only if $P$ is a multiple of $\Delta$. For $\ell>1$, these inequalities really strengthen the old $b \le 2i+7$. We give two elementary proofs. One is similar to Scott’s proof. The other is a bit longer, but it gives more insight into the process of peeling onion skins. For instance, it reveals that the set of all polygons $P$ such that $P^{(1)}=Q$ for some fixed $Q$ is either empty or has a largest element. Moving out edges ---------------- Using this technique, it is actually possible to sharpen the bound in various (sub)cases. E.g., - if $P^{(\ell)}$ is a point, but $P^{(\ell-1)} \neq 3 \Delta$, then $(2\ell-1) b \le 2i + 8 \ell^2 - 2$; - if $P^{(\ell)}$ is a segment, then $(2\ell-1) b \le 2i + 8\ell^2 - 2$ with equality if and only if $P$ is lattice equivalent to a polygon with vertices ${{\mathbf{0}}}, (r,0){^t}, (2pq+r,2p){^t}, (0,2p){^t}$ for integers $p \ge 1$, $q,r \ge 0$ such that $pq+r \ge 3$; - if $P^{(\ell)}$ has no interior lattice points but is not a multiple of $\Delta$, then $(2\ell-1) b \le 2i + 8 \ell^2 - 2$. We reduce the proof to the case that $P$ is obtained from $P^{(1)}$ by “moving out the edges by one”. This is done in the following three lemmas. Finally, Lemma \[lemma:step\] yields the induction step in the proof of the Onion–Skin Theorem. We say that an inequality $\langle {\mathbf{a}}, {\mathbf{x}} \rangle = a_1 x_1 + a_2 x_2 \le b$ with coprime $(a_1,a_2)$ defines an edge of a polygon $Q$ if it is satisfied by all points ${\mathbf{x}} \in Q$, and there are two distinct points in $Q$ satisfying equality. Then, moving out this edge by one means to relax the inequality to $\langle {\mathbf{a}}, {\mathbf{x}} \rangle \le b + 1$. \[lemma:shift\] Suppose that the inequality $\langle {\mathbf{a}}, {\mathbf{x}} \rangle \le b$ defines an edge of $P^{(1)}$. Then $\langle {\mathbf{a}}, {\mathbf{x}} \rangle \le b + 1$ is valid for $P$. That means, if we move all the edges of $Q=P^{(1)}$ out by one, we obtain a superset $Q^{(-1)}$ of $P$. ![If $Q=P^{(1)}$ then $P \subseteq Q^{(-1)}$.[]{data-label="fig:shift"}](shift.eps) We may apply a lattice equivalence to reduce to the case where the edge is defined by $x_2 \le 0$, and that $(0,0){^t}$ and $(1,0){^t}$ are two lattice points of $P^{(1)}$ lying on this edge. Suppose indirectly that $P$ has a vertex ${\mathbf{v}}$ with $v_2 > 1$. Then the triangle formed by the three points has area $v_2/2 \ge 1$. It must therefore contain another lattice point which lies in the interior of $P$, and has positive second coordinate. For arbitrary $Q$, this $Q^{(-1)}$ does not necessarily have integral vertices. But then, not every polygon arises as $P^{(1)}$ for some $P$. A necessary condition is that the polygon has good angles. ![$Q^{(-1)}$ may be nonintegral.[]{data-label="fig:badshift"}](Qminus1.alt.eps) \[lemma:angles\] If $P^{(1)}$ is $2$–dimensional, then for all vertices ${\mathbf{v}}$ of $P^{(1)}$, the cones generated by $P^{(1)} - {\mathbf{v}}$ are lattice equivalent to a cone generated by $(1,0){^t}$ and $(-1,k){^t}$, for some integer $k \ge 1$. ![Good angles and a bad angle.[]{data-label="fig:goodAngles"}](goodAngles.eps) Assume, after a lattice equivalence, that ${\mathbf{v}} = {{\mathbf{0}}}$, and the rays of the cone in question are generated by $(1,0){^t}$ and $(-p,q){^t}$, with coprime $0 < p \le q$ (see Exercise \[ex:angle\]). By Lemma \[lemma:shift\], all points of $P$ satisfy $x_2 \ge -1$ and $q x_1 + p x_2 \ge -1$. But this implies $x_1 + x_2 \ge -1 + \frac{p-1}{q}$. So, if $p > 1$, because $P$ has integral vertices, we have $x_1 + x_2 \ge 0$ for all points of $P$. This contradicts the fact that ${{\mathbf{0}}}\in P^{(1)}$. For a vertex ${\mathbf{v}}$ of a polygon define the shifted vertex ${\mathbf{v}}^{(-1)}$ as follows. Let $\langle {\mathbf{a}}, {\mathbf{x}} \rangle \le b$ and be the two edges that intersect in ${\mathbf{v}}$. The unique solution to $\langle {\mathbf{a}}, {\mathbf{x}} \rangle = b + 1$ and $\langle {\mathbf{a}}', {\mathbf{x}} \rangle = b' + 1$ is denoted ${\mathbf{v}}^{(-1)}$. According to Lemma \[lemma:angles\], when we deal with $P^{(1)}$ then ${\mathbf{v}}^{(-1)}$ is a lattice point. (In the situation of the lemma, it is $(0,-1){^t}$.) We obtain a characterization of when $Q = P^{(1)}$ for some $P$. \[lemma:Qminus1integral\] For a polygon $Q$, the following are equivalent: - $Q = P^{(1)}$ for some polygon $P$. - $Q^{(-1)}$ has integral vertices. Thus, given $Q$, [the]{} maximal polygon $P$ with $P^{(1)} = Q$ is $P = Q^{(-1)}$. We will (and can) restrict to this situation when we prove the induction step $\ell \leadsto \ell + 1$ for the Onion–Skin Theorem. If $Q^{(-1)}$ has integral vertices, then its interior lattice points span $Q$. For the converse direction, if $Q = P^{(1)}$ then we claim that $$\label{eq:Qminus1} Q^{(-1)} = {\mbox{\rm conv}}\{ {\mathbf{v}}^{(-1)} {\ : \ }{\mathbf{v}} \text{ vertex of } Q \} \ .$$ To this end, denote ${\mathbf{a_1}}, \ldots, {\mathbf{a_n}}$ and $b_1, \ldots, b_n$ the normal vectors respectively right hand sides of the edges of $Q$ in cyclic order. Also, denote ${\mathbf{v}}_1, \ldots, {\mathbf{v}}_n$ the vertices of $Q$ so that edge number $k$ is the segment ${[ {{\mathbf{v_\text{$k$}}},{\mathbf{v_\text{$k\negthinspace+\negthinspace1$}}}}]}$ ($k \mod n$). “$\subseteq$”: this inclusion holds for arbitrary $Q$. For a point ${\mathbf{y}} \in Q^{(-1)}$, let $\langle {\mathbf{a}}_k, {\mathbf{x}} \rangle \le b_k$ be an edge of $Q$ that maximizes $\langle {\mathbf{a}}, {\mathbf{y}} \rangle - b$ over all edges. So if $\langle {\mathbf{a}}_k, {\mathbf{y}} \rangle - b_k \le 0$ then ${\mathbf{y}} \in Q$. Otherwise we have - $b_k \le \langle {\mathbf{a}}_k, {\mathbf{y}} \rangle \le b_k + 1$, - $\langle {\mathbf{a}}_k, {\mathbf{y}} \rangle - b_k \ge \langle {\mathbf{a}}_{k-1}, {\mathbf{y}} \rangle - b_{k-1}$, and - $\langle {\mathbf{a}}_k, {\mathbf{y}} \rangle - b_k \ge \langle {\mathbf{a}}_{k+1}, {\mathbf{y}} \rangle - b_{k+1}$, which describes (a subset of) the convex hull of ${\mathbf{v}}_k, {\mathbf{v}}_k^{(-1)}, {\mathbf{v}}_{k+1}, {\mathbf{v}}_{k+1}^{(-1)}$. “$\supseteq$”: For this inclusion we actually use $Q = P^{(1)}$. Figure \[fig:Qminus1\] shows how Equation can fail in general. (0,0)![Equation can fail in general.[]{data-label="fig:Qminus1"}](Qminus1.bad.pstex "fig:") \#1\#2\#3\#4\#5[ @font ]{} (6766,2435)(2918,-7044) (6601,-4936)[(0,0)\[lb\]]{} (6001,-6136)[(0,0)\[lb\]]{} In our situation, $Q = P^{(1)}$, and we need to show that $v_k^{(-1)}$ satisfies all inequalities $\langle {\mathbf{a}}_j, \cdot \rangle \le b_j + 1$ for $Q^{(-1)}$. Our assumption implies that $P$ (and therefore $Q^{(-1)}$ by Lemma \[lemma:shift\]) contains points ${\mathbf{w}}_j$ with $\langle {\mathbf{a}}_j, {\mathbf{w}}_j \rangle = b_j + 1$. None of the other edge normals belongs to the cone generated by ${\mathbf{a}}_{k-1}$ and ${\mathbf{a}}_k$. So for $j \neq k,k-1$, $$\begin{array}{lcccccc} \text{either } & \langle {\mathbf{a}}_j, {\mathbf{v}}_k^{(-1)} \rangle &\le& \langle {\mathbf{a}}_j, {\mathbf{w}}_{k-1} \rangle &\le& b_m + 1 , \\ \text{or } & \langle {\mathbf{a}}_j, {\mathbf{v}}_k^{(-1)} \rangle &\le& \langle {\mathbf{a}}_j, {\mathbf{w}}_k \rangle &\le& b_m + 1 & \text{ (or both).} \end{array}$$ Finally, we can prove the key lemma for our induction step. \[lemma:step\] Let $b^{(1)}$ denote the number of lattice points on the boundary of $P^{(1)}$. Then $b \le b^{(1)} + 9$, with equality if and only if $P$ is a multiple of $\Delta$. This immediately shows that $b \le b^{(1)} + 9 \le i + 9$ if $P^{(1)}$ is $2$–dimensional. For the proof, we need a result of B. Poonen and F. Rodriguez-Villegas [@Poonen12]. Consider a primitive oriented segment $s={[ {{\mathbf{x}},{\mathbf{y}}}]}$, i.e., ${\mathbf{x}}$ and ${\mathbf{y}}$ are the only lattice points $s$ contains. Call $s$ admissible if the triangle ${\mbox{\rm conv}}({{\mathbf{0}}}, {\mathbf{x}}, {\mathbf{y}})$ contains no other lattice points. Equivalently, $s$ is admissible if the determinant ${\operatorname{sign}}(s) = \left\| \begin{smallmatrix} x_1&y_1 \\ x_2&y_2 \end{smallmatrix} \right\|$ is equal to $\pm 1$. The length of a sequence $(s^{(1)}, \ldots, s^{(n)})$ of admissible segments is $\sum {\operatorname{sign}}(s^{(k)})$. The dual of an admissible segment is the unique integral normal vector ${\mathbf{a}}={\mathbf{a}}(s)$ such that $\langle {\mathbf{a}}, {\mathbf{x}} \rangle = \langle {\mathbf{a}}, {\mathbf{y}} \rangle = 1$. For a closed polygon with segments $(s^{(1)}, \ldots, s^{(n)})$, the dual polygon walks through the normal vectors ${\mathbf{a}}(s^{(k)})$. \#1\#2\#3\#4\#5[ @font ]{} (5570,2581)(0,-10) (4515,1883) (4515,1883) (4515,683) (4515,683) (3315,1283) (3315,1283) (3315,1883) (3315,1883) (3315,683) (3315,683) (3915,1283) (3915,1283) (3915,1883) (3915,1883) (3915,683) (3915,683) (3915,2483) (3915,2483) (4515,2483) (4515,2483) (3315,2483) (3315,2483) (5115,2483) (5115,2483) (5115,1883) (5115,1883) (5115,1283) (5115,1283) (5115,683) (5115,683) (5115,83) (5115,83) (4515,83) (4515,83) (3915,83) (3915,83) (3315,83) (3315,83) (1515,1283) (1515,1283) (1515,1883) (1515,1883) (1515,683) (1515,683) (315,1283) (315,1283) (315,1883) (315,1883) (315,683) (315,683) (915,1883) (915,1883) (915,683) (915,683) (4515,1283) (4515,1283) (915,1283) (915,1283) (1515,1283)(315,1883) (556.495,1829.334)(315.000,1883.000)(502.830,1722.003) (315,1883)(915,1883) (675.000,1823.000)(915.000,1883.000)(675.000,1943.000) (915,1883)(315,683) (368.666,924.495)(315.000,683.000)(475.997,870.830) (315,683)(1515,1283) (1327.170,1122.003)(1515.000,1283.000)(1273.505,1229.334) (5115,2483)(4515,1883) (4642.279,2095.132)(4515.000,1883.000)(4727.132,2010.279) (4515,1883)(3315,1883) (3555.000,1943.000)(3315.000,1883.000)(3555.000,1823.000) (3315,1883)(5115,83) (4902.868,210.279)(5115.000,83.000)(4987.721,295.132) (5115,83)(5115,2483) (5175.000,2243.000)(5115.000,2483.000)(5055.000,2243.000) (1140,1583)[(0,0)\[lb\]]{} (540,2108)[(0,0)\[lb\]]{} (15,833)[(0,0)\[lb\]]{} (540,308)[(0,0)\[lb\]]{} (2865,1958)[(0,0)\[lb\]]{} (5340,158)[(0,0)\[lb\]]{} (4215,1958)[(0,0)\[lb\]]{} (5340,2258)[(0,0)\[lb\]]{} The sum of the lengths of an admissible polygon and its dual is $12$ times the winding number of the polygon. Heuristically, the winding number counts how many times a polygon winds around the origin. Dual polygons will have equal winding number. In this article, we will only be concerned with polygons of winding number one. \#1\#2\#3\#4\#5[ @font ]{} (1366,1381)(0,-10) (1283,683) (1283,683) (1283,1283) (1283,1283) (1283,83) (1283,83) (83,683) (83,683) (83,1283) (83,1283) (83,83) (83,83) (683,683) (683,683) (683,1283) (683,1283) (683,83) (683,83) (83,683)(1283,1283) (1095.170,1122.003)(1283.000,1283.000)(1041.505,1229.334) (1283,1283)(683,83) (736.666,324.495)(683.000,83.000)(843.997,270.830) (683,83)(83,1283) (243.997,1095.170)(83.000,1283.000)(136.666,1041.505) (83,1283)(1283,683) (1041.505,736.666)(1283.000,683.000)(1095.170,843.997) (1283,683)(83,83) (270.830,243.997)(83.000,83.000)(324.495,136.666) (83,83)(83,683) (143.000,443.000)(83.000,683.000)(23.000,443.000) Let $Q = P^{(1)}$, and note that by Lemma \[lemma:shift\], $P \subseteq Q^{(-1)}$, and by Lemma \[lemma:Qminus1integral\], $Q^{(-1)}$ has integral vertices. Notice that the number of boundary lattice points of $Q$ is $b^{(1)}$, and let $b'$ be the number of boundary lattice points of $Q^{(-1)}$. Since $P$ and $Q^{(-1)}$ have the same interior lattice points and $P \subseteq Q^{(-1)}$, by Pick’s Theorem $Q^{(-1)}$ has at least as many boundary lattice points as $P$; in other words, $b' \ge b$. For each of the vertices $v^{(1)}_1, \ldots, v^{(1)}_n$ of $Q$ there is a corresponding vertex $v_1, \ldots, v_n$ of $Q^{(-1)}$. Consider the (possibly nonconvex, nonsimple) admissible polygon with vertices $v_1-v^{(1)}_1, \ldots, v_n-v^{(1)}_n$. It is admissible because there are no lattice points between $Q$ and $Q^{(-1)}$. One can think of it as what remains of $Q^{(-1)}$ when $Q$ shrinks to a point. Each segment measures the difference (with the correct sign) between the corresponding edges of $Q^{(-1)}$ and $Q$. I.e., the length of that polygon is precisely $b'-b^{(1)}$. ![An admissible polygon from $(P,P^{(1)})$, and its dual.[]{data-label="fig:shrink"}](shrink.eps) Now the dual polygon will walk through the normal vectors of $Q$. Therefore all segments will count with positive length, and there cannot be less than $3$. Also, there is a unique one with $3$ segments, which is the dual to $3 \Delta$. Thus $b-b^{(1)} \le b'-b^{(1)} \le 12-3$ with equality only for multiples of $\Delta$. Induction on $\ell$. - For $\ell=1$, the inequality $b \le 2i + 7$ was proved earlier. - For $\ell=4/3$, we have $i=3$, and $P \subseteq 4 \Delta$. So $b \le 12$. - For $\ell=5/3$, we have $i=6$, and $P \subseteq 5 \Delta$. So $b \le 15$. - For $\ell=3/2$, Lemma \[lemma:step\] reads $b \le i + 8$ which is stronger than what we need. If $\ell \ge 2$, we have $$\begin{aligned} (2\ell-1) b &\le (2\ell-1) b^{(1)} + 9(2\ell-1) \\ &= 2 b^{(1)} + (2(\ell-1)-1) b^{(1)} + 9(2\ell-1) \\ &\le 2 b^{(1)} + 2i^{(1)} + 9(\ell-1)^2 - 2 + 9(2\ell-1) \\ &= 2i + 9\ell^2 - 2 \end{aligned}$$ Generalizing Scott’s proof -------------------------- As in Subsection \[ssec:scott\], we tightly fit $P$ into a box $[0,p'] \times [0,p]$, with $p\le p'$. Let $q$ and $q'$ be the length of the top and bottom edge (see Figure \[fig:box\]). We again apply lattice equivalence transformations such that $p$ is as small as possible, and that $P$ has points on the top and bottom edges with horizontal distance smaller than or equal to $(p-q-q')/2$. Again, we obtain the following inequalities: $$\begin{aligned} b \le q+q'+2p \tag{\ref{eq:sc1}} \\ a \ge p(q+q')/2 \tag{\ref{eq:sc2}} \\ a \ge p(p+q+q')/4 \tag{\ref{eq:sc3}}\end{aligned}$$ Set $x:=p/\ell$ and $y:=(q+q')/\ell$. Then $x\ge 2$, because passing to $P^{(1)}$ reduces the height at least by 2.[^4] From and , we get $$\begin{aligned} \frac{2\ell b-2a-9\ell^2}{\ell^2} \le 2(q+q'+2p)/\ell - p(q+q')/\ell^2-9 \\ = -xy+4x+2y-9 ,\end{aligned}$$ and from and , we get $$\begin{aligned} \frac{4\ell b-4a-18\ell^2}{\ell^2} \le 4(q+q'+2p)/\ell - p(p+q+q')/\ell^2-18 \\ = -x^2-xy+8x+4y-18.\end{aligned}$$ ![At least one of the polynomials is $\le 0$.[]{data-label="fig:plot"}](plot.ps) For $x\ge 2$ and $y\ge 0$, at least one of the two polynomials $p_1(x,y) = -xy+4x+2y-9$ and $p_2(x,y) = -x^2-xy+8x+4y-18$ is zero or negative, as it can be seen in Figure \[fig:plot\]. (The two shaded regions are where $p_1$ and $p_2$, respectively, take non-negative values.) There is only one point where both upper bounds reach zero, namely $(x,y)=(3,3)$, and this is the only case where equality can hold in the Onion–Skin Theorem. It is an exercise to show that equality actually holds only for multiples of $\Delta$. Conclusion {#sec:conclusion} ========== Summary of results. ------------------- For a triple $(a,b,i)$ of numbers the following are equivalent. - There is a convex lattice polygon $P$ with $(a,b,i)=(a(P),b(P),i(P))$. - $b \in {{\mathbb Z}}_{\ge 3}$, $i \in {{\mathbb Z}}_{\ge 0}, a = i + b/2 -1$, and - $i=0$ or - $i=1$ and $b \le 9$ or - $i \ge 2$ and $b \le 2i+6$. Furthermore, if $\ell = \ell(P)$, then $(2\ell-1) b \le 2i + 9\ell^2 - 2$. Outlook. -------- Is there a proof of the Onion–Skin Theorem using algebraic geometry? Currently not (yet). The toric dictionary between polygons and algebraic varieties also does not (yet) have an algebraic geometry term for the level of a polygon. A first step in this direction is the use of the [process]{} of peeling off onion skins – or rather its algebraic geometry analogue – for the simplification of the rational parametrization of an algebraic surface [@Schicho:03a]. In any case, the Onion–Skin Theorem gives rise to a [conjecture]{} in algebraic geometry, namely the inequality $(2\ell-1)d\le 9\ell^2+4\ell(p-1)$ for any rational surface of degree $d$, sectional genus $p$, and level $\ell$. Here the level of an algebraic surface should be defined via the process of peeling mentioned above. For toric surfaces, the inequality holds by the Onion–Skin Theorem, but for nontoric rational surfaces we do not have a proof (nor a counterexample). We thank Ricky Pollack for his encouragement to write up this story. We also thank Daniel J. Velleman and two anonymous referees who helped us do so in a comprehensible way. [^1]: Work by the first author was supported by NSF grant DMS–0200740 and by Emmy Noether fellowship HA 4383/1 of the German Research Foundation (DFG). Work by the second author was supported by the FWF in the frame of the SFB F013 and the project P15551. [^2]: Hint: if you know [Chinese]{}, not much [remains*]{} to be done. [^3]: It is an exercise to show that the only $P$ with these parameters is the triangle $3\Delta$ in Figure \[fig:reflexive\]. [^4]: We also have $x \le 3$ with equality only for multiples of $\Delta$.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Using the Hartree-Fock plus random-phase-approximation (HF+RPA), we study the impurity effect of $\Lambda$ hyperon on the collective vibrational excitations of double-$\Lambda$ hypernuclei. To this end, we employ a Skyrme-type $\Lambda N$ and $\Lambda\Lambda$ interactions for the HF calculations, and the residual interactions for RPA derived with the same interactions. We find that inclusion of two $\Lambda$ hyperons in $^{16}$O shifts the energy of the collective states towards higher energies. In particular, the energy of the giant monopole resonance of $^{\,\,18}_{\Lambda\Lambda}$O, as well as that of $^{210}_{\Lambda\Lambda}$Pb, becomes larger. This implies that the effective incompressibility modulus increases due to the impurity effect of $\Lambda$ particle, if the $\beta$-stability condition is not imposed.' author: - 'F. Minato' - 'K. Hagino' title: ' Application of random phase approximation to vibrational excitations of double-$\Lambda$ hypernuclei' --- Introduction ============ Information on the interaction between a $\Lambda$ hyperon and a nucleon deepens our understanding of baryon-baryon forces and the equation of state (EOS) of nuclear matter. In principle, the interaction between two particles can be investigated with a measurement of their scattering. However, due to the short life-time of $\Lambda$ hyperon, it has yet been difficult to perform a direct scattering experiment of nucleon and $\Lambda$ hyperon. Therefore, the $\Lambda N$ interaction has been mainly investigated by $\gamma$ spectroscopy of single-$\Lambda$ hypernuclei [@Tanida2001; @Ukai2006; @Tamura2005; @Kohri2002; @Ukai2004]. Such measurements have revealed the $\Lambda$-impurity effect, that is, the change of several properties of atomic nuclei, such as excitation energies and transition probabilities of $\gamma$-ray, due to an addition of $\Lambda$ particle. Apparently, high-resolution $\gamma$-ray measurements are vital in investigating $\Lambda$ hypernuclei. In addition to the existing experimental data, research projects currently planned at the J-PARC facility using new Ge detector arrays (Hyperball-J)[@Tamura] aim at obtaining new data on the low-lying energy level scheme of $\Lambda$ hypernuclei in the $sd$ shell region, that will lead to further understanding of the $\Lambda N$ and $\Lambda\Lambda$ interactions. Several theoretical calculations have been carried out to analyze the relation between low-lying energy levels of single-$\Lambda$ hypernuclei and the $\Lambda N$ interaction [@Nemura2002; @Millener2008; @Umeya2009; @Motoba1983; @Hiyama1997; @Hiyama2010; @Isaka2011; @Isaka2011b]. These calculations have not only contributed to identification of energy level schemes of single-$\Lambda$ hypernuclei, but have also predicted the $\Lambda$-impurity effect on the structure of single-$\Lambda$ hypernuclei, [e.g.]{}, shrinkage of the radius of $^{7}_{\Lambda}$Li from $^6$Li[@Motoba1983], which was subsequently observed experimentally [@Tanida2001]. Besides single-$\Lambda$ hypernuclei, double-$\Lambda$ hypernuclei have also been studied both experimentally and theoretically. Similar to the $\Lambda N$ interaction, information on the $\Lambda\Lambda$ interaction can be deduced from observation of $\gamma$-rays emitted from double-$\Lambda$ hypernuclei. However, until now double-$\Lambda$ hypernuclei have been produced only in an emulsion, and at present emitted $\gamma$-rays are difficult to detect experimentally with high precision. In addition, so far observed double-$\Lambda$ hypernuclei in the emulsion have been limited to five cases($^{\,\,\,\,\,6}_{\Lambda\Lambda}$He and $^{10-13}_{\quad\,\Lambda\Lambda}$B [@Danysz1963; @Nakazawa2010]), and the experimental data have been scarce. Therefore the theoretical approaches make an important tool to assess the $\Lambda$ impurity effect on the structure of double-$\Lambda$ hypernuclei as well as appropriate selection of a target nucleus for future experiments. Theoretically, the double-$\Lambda$ hypernuclei have been investigated within the frameworks of ab-initio few body model[@Nemura2005], shell model[@Gal2011] and cluster model[@Hiyama2002]. However, these theoretical approaches demand a huge computational power, and they may be difficult to apply to heavy hypernuclei. In order to study systematically the $\Lambda$-impurity effect, from light to heavy nuclei, a Hartree-Fock (HF) plus random-phase-approximation (RPA) method provides one of the most suitable tools. This approach has been applied to study vibrational excitations of normal nuclei (without hyperons) throughout the nuclear chart, starting with a single energy functional applicable in the whole range of the nuclear chart. In particular, the RPA has been successfully applied to descriptions of giant resonances of atomic nuclei. See Refs.[@AGL77; @AG80; @LA95; @MDMC08] for earlier applications of Tamm-Dancoff approximation to (K$^-$,$\pi^-$) and ($\pi^+$,K$^+$) reactions, and of an RPA-like model to single-particle spectra of single-$\Lambda$ hypernuclei. So far, the mean field approach has been extended to $\Lambda$ hypernuclei in order to study the ground state properties [@Rayet; @Yamamoto1988; @Lanskoy1998; @Tretyakova1999; @Schulze1995; @Vretenar1998], the potential energy surface in the deformation plane [@Myaing2008; @Myaing2011; @LZZ11] and fission barrier heights [@Minato]. Concerning excited states, the low-lying excited states of $^{25}_{\Lambda}$Mg have recently been calculated with a 5-dimensional (5D) collective Bohr Hamiltonian on the basis of the Skyrme-Hartree-Fock method [@Yao2011] (see also Ref.[@Isaka2011] for a recent application of anti-symmetrized molecular dynamics to the $^{25}_{\Lambda}$Mg hypernucleus). Although the Bohr Hamiltonian approach can handle a large amplitude collective motion, it is much easier to employ the RPA approach to describe collective vibrations with several multipolarities, including giant resonances as well. In this paper, we extend the Skyrme-HF plus RPA (SHF+RPA) scheme to hypernuclei. Skyme-type $\Lambda$N and $\Lambda\Lambda$ interactions, similarly to the Skyrme nucleon-nucleon interaction, are used in this work. The residual interactions for RPA are derived self-consistently from the second derivative of the energy functional with respect to densities. In this study, we shall focus on the double-$\Lambda$ hypernuclei rather than single-$\Lambda$ hypernuclei, partly because the description is much simpler due to the time-reversal symmetry. The $\Lambda$-impurity effect is expected to be stronger in double-$\Lambda$ hypernuclei, and such calculations will provide the upper limit of the impurity effect for single-$\Lambda$ hypernuclei. The paper is organized as follows. In Sec. II, we describe the formalism of the SHF+RPA for hypernuclei. In Sec. III, we apply the SHF+RPA method to $^{\,\,18}_{\Lambda\Lambda}$O hypernuclei and present the results for the strength distributions and the transition densities for the isoscalar giant monopole resonance (IS $0^+$), the electric dipole (E1), quadrupole (E2) and octupole (E3) transitions. We also calculate the isoscalar giant monopole resonance of $^{210}_{\Lambda\Lambda}$Pb hypernucleus and discuss the nuclear incompressibility in the presence of $\Lambda$ hyperon. We give a summary of the paper in Sec. IV. RPA FOR HYPERNUCLEI =================== In order to describe the ground state and excited states of double-$\Lambda$ hypernuclei, we adopt the Skyrme-type zero-range force for the $\Lambda N$ and $\Lambda\Lambda$ interactions. The $\Lambda N$ and 3-body $\Lambda NN$ interactions of this type were first introduced by Rayet as[@Rayet], $$\begin{split} &v_{\Lambda N}({\mbox{\boldmath $r$}}_\Lambda-{\mbox{\boldmath $r$}}_N) = t_0^\Lambda(1+x_0^\Lambda P_\sigma)\delta({\mbox{\boldmath $r$}}_{\Lambda}-{\mbox{\boldmath $r$}}_N)\\ &+\frac{1}{2}t_1^\Lambda\left({\mbox{\boldmath $k'$}}^2\delta({\mbox{\boldmath $r$}}_{\Lambda}-{\mbox{\boldmath $r$}}_N) +\delta({\mbox{\boldmath $r$}}_{\Lambda}-{\mbox{\boldmath $r$}}_N){\mbox{\boldmath $k$}}^2\right)\\ &+t_2^\Lambda{\mbox{\boldmath $k'$}}\delta({\mbox{\boldmath $r$}}_\Lambda-{\mbox{\boldmath $r$}}_N)\cdot{\mbox{\boldmath $k$}} +iW_0^\Lambda{\mbox{\boldmath $k'$}}\delta({\mbox{\boldmath $r$}}_\Lambda-{\mbox{\boldmath $r$}}_N)\cdot(\sigma\times{\mbox{\boldmath $k$}}), \end{split} \label{LN}$$ and $$v_{\Lambda NN}({\mbox{\boldmath $r$}}_\Lambda,{\mbox{\boldmath $r$}}_{N_1},{\mbox{\boldmath $r$}}_{N_2})= t_3^\Lambda\delta({\mbox{\boldmath $r$}}_\Lambda-{\mbox{\boldmath $r$}}_{N_1}) \delta({\mbox{\boldmath $r$}}_{\Lambda}-{\mbox{\boldmath $r$}}_{N_2}), \label{LNN}$$ respectively. In a similar way, Lanskoy introduced the $\Lambda\Lambda$ interaction as [@Lanskoy1998], $$\begin{split} v_{\Lambda\Lambda}&({\mbox{\boldmath $r$}}_{\Lambda_1}-{\mbox{\boldmath $r$}}_{\Lambda_2}) = \lambda_0\delta({\mbox{\boldmath $r$}}_{\Lambda_1}-{\mbox{\boldmath $r$}}_{\Lambda_2})\\ &+\frac{1}{2} \lambda_1\left({\mbox{\boldmath $k'$}}^2\delta({\mbox{\boldmath $r$}}_{\Lambda_1}-{\mbox{\boldmath $r$}}_{\Lambda_2}) +\delta({\mbox{\boldmath $r$}}_{\Lambda_1}-{\mbox{\boldmath $r$}}_{\Lambda_2}){\mbox{\boldmath $k$}}^2\right)\\ &+\lambda_2{\mbox{\boldmath $k'$}}\delta({\mbox{\boldmath $r$}}_{\Lambda_1}-{\mbox{\boldmath $r$}}_{\Lambda_2})\cdot{\mbox{\boldmath $k$}}\\ &+\lambda_3\, \left[\rho_N\left(\frac{{\mbox{\boldmath $r$}}_{\Lambda_1}+{\mbox{\boldmath $r$}}_{\Lambda_2}}{2}\right) \right]^{\alpha_\Lambda}\,\delta({\mbox{\boldmath $r$}}_{\Lambda_1}-{\mbox{\boldmath $r$}}_{\Lambda_2}). \end{split} \label{LL}$$ The operator ${\mbox{\boldmath $k'$}}=-({\mbox{\boldmath $\nabla$}}_1-{\mbox{\boldmath $\nabla$}}_2)/2i$ acts on the left hand side while ${\mbox{\boldmath $k$}}=({\mbox{\boldmath $\nabla$}}_1-{\mbox{\boldmath $\nabla$}}_2)/2i$ acts on the right hand side. $\rho_N({\mbox{\boldmath $r$}})$ is the density distribution for the nucleons. The last term in Eq. corresponds to the three-body $\Lambda\Lambda N$ interaction, which originates mainly from the $\Lambda\Lambda-\Xi N$ coupling [@Hiyama2004]. Together with the Skyrme $NN$ interaction[@Vautherin72], the total energy $E_{\rm tot}$ in the Hartree-Fock approximation is given by $$E_{\rm tot}=E_N+E_\Lambda, \label{energy}$$ where $$E_N=\int H_N({\mbox{\boldmath $r$}})\,d{\mbox{\boldmath $r$}}, \label{energy2}$$ is the energy for the core nucleus without $\Lambda$ hyperons while $$E_\Lambda=\int [H_{N\Lambda}({\mbox{\boldmath $r$}})+H_{\Lambda\Lambda}({\mbox{\boldmath $r$}})]\,d{\mbox{\boldmath $r$}}, \label{energy3}$$ is due to the $\Lambda N$ and $\Lambda\Lambda$ interactions (see Appendix A). The kinetic energy density for $\Lambda$ particles is included in the energy density $H_{N\Lambda}({\mbox{\boldmath $r$}})$. The SHF equations are obtained by taking variation of the total energy $E_{\rm tot}$ with respect to the densities for neutrons, protons and $\Lambda$ hyperons. These are given as $$\left(-{\mbox{\boldmath $\nabla$}}\cdot \frac{\hbar^2}{2m_b^({\mbox{\boldmath $r$}})}{\mbox{\boldmath $\nabla$}}+U_{bN}({\mbox{\boldmath $r$}}) +U_{b\Lambda}({\mbox{\boldmath $r$}})\right)\phi_b({\mbox{\boldmath $r$}})=\epsilon_b\phi_b({\mbox{\boldmath $r$}}), \label{shfeq1}$$ where the index $b$ refers to proton, neutron or $\Lambda$, and $\epsilon_b$ is the single-particle energy. The explicit forms for the mean-field potentials $U_{bN}({\mbox{\boldmath $r$}})$ and $U_{b\Lambda}({\mbox{\boldmath $r$}})$, and the effective mass $m_b^({\mbox{\boldmath $r$}})$ are given in Appendix A. After we construct the ground state in the Hartree-Fock approximation, we describe excited states with RPA as a linear superposition of 1 particle-1 hole (1p1h) configurations. That is, the excitation operator $Q_k^\dagger$ for the $k$-th RPA phonon is assumed to be, $$Q_k^\dagger = \sum_{p,h\in n,p,\Lambda}\left( X_{ph}^{(k)}a_p^\dagger a_h - Y_{ph}^{(k)}a_h^\dagger a_p\right),$$ where $X_{ph}^{(k)}$ and $Y_{ph}^{(k)}$ are the forward and backward amplitudes, respectively. $a_p^\dagger$ and $a_h^\dagger$ are the creation operators for a particle state $p$ and for a hole state $h$, respectively. The excitation energy $E_k$ is obtained by diagonalizing the $2\nu$-dimensional RPA equation, $$\left(\begin{tabular}{cc} $A$ & $B$\\ $-B^$ & $-A^$ \end{tabular}\right) \left(\begin{tabular}{c} $X^{(k)}$\\$Y^{(k)}$ \end{tabular}\right) =E_k \left(\begin{tabular}{c} $X^{(k)}$\\$Y^{(k)}$ \end{tabular}\right), \label{rpa}$$ where $\nu$ is the number of 1p1h configurations. Here, $A$ and $B$ are RPA matrices given by, $$\begin{split} A_{ph,p'h'}&=(\epsilon_p-\epsilon_h)\delta_{pp'}\delta_{hh'}+v_{ph'hp'}\\ B_{ph,p'h'}&=v_{pp'hh'}, \end{split} \label{abmatrix}$$ where $v$ is the residual interaction derived from the energy functional, $E_{\rm tot}$. The formalism is almost the same as the standard RPA found [e.g.,]{} in Refs.[@RingSchuck; @Rowe], but the particle-hole configurations run over not only protons and neutron but also $\Lambda$ hyperons. The interaction matrix elements $v_{ph'hp'}$ and $v_{pp'hh'}$ include the $\Lambda N$ and $\Lambda\Lambda$ interactions as well as the $NN$ interaction (see Appendix B). The external fields for electric multipole excitations with multipolarities $L\neq$ 0 and 1 are defined as $$\hat{F}_{EL}=e\sum_{i\in p} r_i^LY_{LM}(\hat{r_i}),$$ while that for the “isoscalar” monopole transition is $$\hat{F}_{0^+}=\sum_{i\in p,n,\Lambda} r_i^2.$$ For the electric dipole response, we take into account the center of mass motion of the whole hypernucleus and use the operator $$\begin{split} \hat{F}_{E1} &=e\sum_{i\in p}(r_iY_{1M}(\hat{r}_i)-RY_{1M}({\mbox{\boldmath $R$}})), \\ &=e\frac{Nm_N+N_\Lambda m_\Lambda}{M}\sum_{i\in p}r_iY_{1M}(\hat{r}_i)\\ &-e\frac{Z}{M}\left(m_N\sum_{i\in n}r_iY_{10}(\hat{r}_i) +m_\Lambda\sum_{i\in \Lambda} r_i Y_{10}(\hat{r}_i)\right), \end{split}$$ where $${\mbox{\boldmath $R$}}=\frac{1}{M}\,\left(m_N\sum_{i\in n,p} {\mbox{\boldmath $r$}}_i + m_\Lambda\sum_{i\in \Lambda} {\mbox{\boldmath $r$}}_i\right),$$ is the center of mass of the hypernucleus, and $M \equiv m_N(Z+N)+m_\Lambda N_\Lambda$ is the total mass, $m_N=(m_p+m_n)/2=938.92$ MeV/$c^2$ and $m_\Lambda=1115.68$ MeV/$c^2$ being the mass of nucleon and $\Lambda$ hyperon, respectively. $N$, $Z$ and $N_\Lambda$ are the number of neutron, proton and $\Lambda$ hyperon, respectively. RESULTS ======= single-particle level of $^{\,\,18}_{\Lambda\Lambda}$O ------------------------------------------------------ ![(Color online) Neutron and proton single-particle levels of $^{16}$O (the solid lines) and $^{\,\,18}_{\Lambda\Lambda}$O (the dashed lines) obtained with the Skyrme-Hartree-Fock method.[]{data-label="single-particle"}](splevel.eps){width="0.7\linewidth"} We now numerically solve the RPA equation and discuss the collective excitations of double-$\Lambda$ hypernuclei. Before we show the results for multipole vibrations, we first discuss the single particle levels of double-$\Lambda$ hypernucleus $^{\,\,18}_{\Lambda\Lambda}$O and $^{16}$O, which will help to understand the $\Lambda$-impurity effect on the giant resonances. To this end, we assume spherical symmetry, and solve the SHF equation in the coordinate space with a grid size of $dr$=0.1 fm. We use the SkM$^$ parameter set for $NN$ interaction [@SkM], while the No.5 parameter set in Ref.[@Yamamoto1988] for the $\Lambda N$ interaction, whose parameters were determined by fitting the Hartree-Fock calculations to the experimental binding energies of single-$\Lambda$ hypernuclei[@Yamamoto1988]. For the $\Lambda\Lambda$ interaction, we use the S$\Lambda\Lambda1$ parameter set evaluated by Lanskoy[@Lanskoy1998]. This parameter set was obtained by fitting to the $\Lambda\Lambda$ bond energy [@Lanskoy1998], $\Delta B_{\Lambda\Lambda}=B_{\Lambda\Lambda}-2B_{\Lambda}$, where $B_{\Lambda}$ is the one-$\Lambda$ hyperon separation energy from a $^{\rm{A}+1}_{\quad\Lambda}$Z hypernucleus and $B_{\Lambda\Lambda}$ is the two-$\Lambda$ hyperon separation energy of $^{\rm{A}+2}_{\,\,\,\Lambda\Lambda}$Z. As we will show in the next subsection, the dependence of giant resonances on a choice of parameter sets for the $\Lambda N$ and the $\Lambda\Lambda$ interactions is weak, and any significant change in the results is not obtained even if we use different parameter sets for the interactions. Figure \[single-particle\] shows the neutron and proton single-particle levels of the $^{16}$O (the solid lines) and the $^{\,\,18}_{\Lambda\Lambda}$O (the dashed lines). The single-particle energies of the $^{\,\,18}_{\Lambda\Lambda}$O hypernucleus are smaller than those of $^{16}$O, since the depths of the central part of the mean-field potentials are deepened both for protons and neutrons due to the attractive $\Lambda N$ interaction, as shown in Fig.2 of Ref.[@Tretyakova1999]. The lowest level ($1s_{1/2}$) is the most sensitive to the addition of $\Lambda$ particles, for which the single particle levels are shifted by $\Delta\epsilon_n=-1.4$ MeV for neutron and $\Delta\epsilon_p=-1.3$ MeV for proton. This $\Lambda$-impurity effect becomes weaker as the energy of a single-particle state increases. If the continuum spectra are discretized within a large box, the difference of single-particle energies between the $^{16}$O and the $^{\,\,18}_{\Lambda\Lambda}$O nuclei is much smaller as compared to the bound levels. Consequently, in the independent-particle approximation, the excitation energies increase relatively in the double-$\Lambda$ hypernucleus as compared to those of the normal nucleus. In the next subsections, we will see that this is the case even in the presence of the residual particle-hole interaction. Low-lying excitations --------------------- We next solve the RPA equation in order to discuss collective excitations of the $^{\,\,18}_{\Lambda\Lambda}$O hypernucleus. To this end, we discretize the single-particle continuum states with the box boundary condition with the box size of $16.0$ fm. We take into account the continuum states up to $\epsilon$=30 MeV, and consider the 1p1h configurations whose unperturbed energy, $\epsilon_p-\epsilon_h$, is smaller than 60 MeV. For the residual interactions, we neglect the Coulomb and the spin-orbit terms for simplicity, although we include all the other terms self-consistently. Therefore, our RPA calculations are not fully self-consistent, and the spurious translational motion appears at a finite excitation energy. In order to recover effectively the self-consistency, we introduce a scaling factor $f$ to the residual interaction $v_{res}$ so as to produce the spurious translational mode at zero energy. Table \[1stE2E3\] shows the results of such RPA calculations for the lowest quadrupole and the octupole states of $^{16}$O and $^{\,\,18}_{\Lambda\Lambda}$O. For both the 2$_1^+$ and 3$_1^-$ states, the impurity effects of $\Lambda$ particles slightly reduces the collectivity, that is, the excitation energies are increased while the electromagnetic transition probabilities are decreased by 26-28 %. The increase of the excitation energies is consistent with the increase of unperturbed particle-hole energies discussed in the previous subsection. ------------------------------- ----------- ----------------------- ----------- ----------------------- nucleus $E$ (MeV) $B(E2)$ (e$^2$fm$^4$) $E$ (MeV) $B(E3)$ (e$^2$fm$^6$) $^{16}$O $13.1$ $0.726$ $6.06$ $91.1$ $^{\,\,18}_{\Lambda\Lambda}$O $13.8$ $0.529$ $6.32$ $67.7$ ------------------------------- ----------- ----------------------- ----------- ----------------------- : The excitation energies and the electromagnetic transition probabilities, $B(E2)$ and $B(E3)$, for the first 2$^+$ and 3$^-$ states of $^{16}$O and $^{\,\,18}_{\Lambda\Lambda}$O nuclei obtained with the Skyrme-HF+RPA method. \[1stE2E3\] Giant resonances ---------------- ![(Color online) The strength distributions for the electric dipole (E1, the top panel), the electric quadrupole (E2, the middle panel) and the electric octupole (E3, the bottom panel) excitations of the $^{16}$O nucleus (the dashed lines) and of the double-$\Lambda$ hypernucleus $^{\,\,18}_{\Lambda\Lambda}$O (the solid and the dotted lines). The solid lines are obtained by including the residual $NN$, $\Lambda N$ and $\Lambda\Lambda$ interactions, while the dotted lines are obtained by including only the residual $NN$ interactions. The strength distributions are weighted by the Lorentzian function with a width of $1.0$ MeV. For the peaks indicated by the arrows, the transition densities are shown in Fig.\[drho\_O16\].[]{data-label="multi"}](O16dipo2.eps "fig:"){width="0.70\linewidth"} ![(Color online) The strength distributions for the electric dipole (E1, the top panel), the electric quadrupole (E2, the middle panel) and the electric octupole (E3, the bottom panel) excitations of the $^{16}$O nucleus (the dashed lines) and of the double-$\Lambda$ hypernucleus $^{\,\,18}_{\Lambda\Lambda}$O (the solid and the dotted lines). The solid lines are obtained by including the residual $NN$, $\Lambda N$ and $\Lambda\Lambda$ interactions, while the dotted lines are obtained by including only the residual $NN$ interactions. The strength distributions are weighted by the Lorentzian function with a width of $1.0$ MeV. For the peaks indicated by the arrows, the transition densities are shown in Fig.\[drho\_O16\].[]{data-label="multi"}](O16quad2.eps "fig:"){width="0.70\linewidth"} ![(Color online) The strength distributions for the electric dipole (E1, the top panel), the electric quadrupole (E2, the middle panel) and the electric octupole (E3, the bottom panel) excitations of the $^{16}$O nucleus (the dashed lines) and of the double-$\Lambda$ hypernucleus $^{\,\,18}_{\Lambda\Lambda}$O (the solid and the dotted lines). The solid lines are obtained by including the residual $NN$, $\Lambda N$ and $\Lambda\Lambda$ interactions, while the dotted lines are obtained by including only the residual $NN$ interactions. The strength distributions are weighted by the Lorentzian function with a width of $1.0$ MeV. For the peaks indicated by the arrows, the transition densities are shown in Fig.\[drho\_O16\].[]{data-label="multi"}](O16octu2.eps "fig:"){width="0.70\linewidth"} The RPA method has been successfully applied to giant resonances of the normal nuclei. We therefore apply it in this subsection to the giant resonances of the double-$\Lambda$ hypernucleus $^{\,\,18}_{\Lambda\Lambda}$O, although they may not be easy to access experimentally at present. The top, the middle, and the bottom panels of Fig. \[multi\] show the strength distributions for the electric dipole (E1), quadrupole (E2) and octupole (E3) excitations, respectively, weighted by the Lorentzian function with a width of $1.0$ MeV. The solid and the dashed lines denote the results for the $^{\,\,18}_{\Lambda\Lambda}$O and $^{16}$O nuclei, respectively. In order to assess the role of $\Lambda$ hyperon, we also show the results for the $^{\,\,18}_{\Lambda\Lambda}$O hypernuclei in which the $\Lambda N$ and $\Lambda\Lambda$ interactions are taken into account only in the ground state, that is, the results obtained by switching off the residual $\Lambda N$ and $\Lambda\Lambda$ interactions (the dotted lines). The figure indicates that the addition of the $\Lambda$ hyperons shifts the peaks of the strength functions towards high energy for all the modes of excitations. This is similar to the results for the low-lying modes of excitations discussed in the previous subsection, and is again caused mainly by the change of the single-particle energies. On the other hand, the difference between the solid and the dotted lines is relatively small, except for the low-lying dipole state at $E=12.8$ MeV. We have confirmed that the strength functions remain almost the same, including the low-lying dipole state, even if other parameter sets for the $\Lambda N$ and $\Lambda\Lambda$ interactions are employed. This suggests that the main effect of $\Lambda$ particles on the collective vibrational excitations is indeed attributed to the change in the single-particle energies, rather than the residual $\Lambda N$ and $\Lambda\Lambda$ interactions, although the low-lying dipole state may require a separate analysis (see Fig. 4 below). In order to see the $\Lambda$-impurity effect quantitatively, we show in Table \[centroid\] the centroid energy defined as $\overline{E}=m_1/m_0$, where $m_k$ is $k$-th energy-weighted sum-rule, $$m_k=\sum_{\nu}(E_i)^k \Big\|\langle i\|F\|0\rangle\Big\|^2, \label{ewsr}$$ for the unperturbed (HF) and the perturbed (RPA) strength functions. We also list the difference of the centroid energy, $\delta\overline{E}$, between the $^{\,\,18}_{\Lambda\Lambda}$O and the $^{16}$O nuclei. The values in the parentheses for the RPA E1 response are the results obtained by excluding the contribution of the low-lying dipole peak at $E=12.8$ MeV. As is expected, the centroid energies $\overline{E}$ for the HF calculations shift to higher energies when the $\Lambda$ hyperons are added to $^{16}$O. For the E2 and E3 responses, this remain the same even if the residual interactions are taken into account in RPA. For the E1 response, the energy shift is negative, but it turs to positive if the low-lying peak is excluded. As we will show below, this low-lying peak corresponds to the dipole motion of a $\Lambda$ hyperon against the core nucleus. We thus conclude that the $\Lambda$ hyperons generally increases the centroid energy for collective motions of the core nucleus, not only for the quadrupole and the octupole states but also for the dipole states. E1 E2 E3 ------- ------------------------------- ------------------- --------- --------- (HF) $^{16}$O $13.76$ $25.57$ $26.53$ $^{\,\,18}_{\Lambda\Lambda}$O $14.34$ $26.63$ $27.74$ $\delta \overline{E}$ $+0.58$ $+1.06$ $+1.21$ (RPA) $^{16}$O $19.92$ $19.55$ $22.32$ $^{\,\,18}_{\Lambda\Lambda}$O $19.68$ ($20.95$) $20.09$ $24.05$ $\delta \overline{E}$ $-0.24$ ($+1.03$) $+0.54$ $+1.73$ : The centroid energy $\overline{E}$ for the E1, E2 and E3 modes of excitations for $^{16}$O and $^{\,\,18}_{\Lambda\Lambda}$O nuclei. Those are given in units of MeV, and the results of both the unperturbed (HF) and the perturbed (RPA) calculations are shown. $\delta \overline{E}$ denotes the difference of the centroid energies between $^{\,\,18}_{\Lambda\Lambda}$O and $^{16}$O. The values in the parentheses for the E1 mode are the results obtained by excluding the contribution of the low-lying peak at $E=12.8$ MeV. \[centroid\] Figure \[drho\_O16\] shows the transition densities for the giant dipole and quadrupole resonances and for the high-lying octupole state, which are indicated by the arrows in Fig.\[multi\]. The top, the middle and the bottom panels denote the transition densities for the neutrons, the protons and the $\Lambda$ hyperons, respectively. Those transition densities are computed as $$\begin{split} \delta\rho_i(r) =\sum_{ph}&(X^{(i)}_{ph}+(-1)^LY^{(i)}_{ph})\\ &\times\varphi_p(r)\varphi_h(r)\langle j_pl_p\|\|Y_L\|\|j_hl_h\rangle, \end{split}$$ where $j_i$, $l_i$ are the total and the orbital angular momenta for a single-particle state $i$, respectively, while $\varphi_i(r)$ is the radial part of the wave function, normalized as $\int_0^\infty \varphi_i^(r)\varphi_i(r)\,r^2dr=1$. The peaks of the transition densities for the neutrons and protons slightly move to a small distance and are shifted towards inside for all the multipolarities when two $\Lambda$ hyperons are added. The amplitude of the transition density for the $\Lambda$ hyperons is about $1/10-1/100$ smaller than that for the protons and neutrons, so that the $\Lambda$ hyperons do not contribute much to these giant resonances. For the E2 and E3 states, the neutrons and the protons oscillate in phase as is expected for isoscalar motions, while they oscillate out of phase for the E1 state ([i.e.,]{} the isovector motion). In addition, the $\Lambda$ hyperons oscillate in phase with the protons and the neutrons for the E2 and the E3 modes, while they oscillate in phase with the protons for the E1 mode. When the Coulomb force is turned off completely, that is, when the single-particle levels for the protons are identical to those for the neutrons, the amplitude of the E1 transition density for the $\Lambda$ hyperons vanish. ![image](drho_O16GDR.eps){width="0.300\linewidth"} ![image](drho_O16GQR.eps){width="0.286\linewidth"} ![image](drho_O16GOR2.eps){width="0.292\linewidth"} ![(Color online) The transition density for the low-lying E1 state at $E=12.8$ MeV in the $^{\,\,18}_{\Lambda\Lambda}$O hypernucleus. The dashed, the dotted, and the solid lines show the contributions of the neutrons, the protons, and the $\Lambda$ hyperons, respectively.[]{data-label="drho_low"}](drho_O16GDR_low.eps){width="0.80\linewidth"} The low-lying dipole state at $E=12.8$ MeV deserves a special attention. This peak appears only when the $\Lambda$ hyperons are added to the $^{16}$O nucleus, and a similar peak does not seen in other modes of excitations. Figure \[drho\_low\] shows the transition density for this state. In contrast to the giant resonance shown in Fig.\[drho\_O16\], the amplitude of the transition density for the $\Lambda$ hyperons is higher than that for the protons and the neutrons. The strongest RPA amplitude, $\xi\equiv X^2-Y^2$, contributing to this peak is the $[1p\,(1s)^{-1}]$ configuration of the $\Lambda$ particles ($\xi=0.873$). The total RPA amplitudes for the neutrons and the protons are small ($\xi=0.050$ for the neutrons and $\xi=0.071$ for the protons), and these values become entirely zero when the $\Lambda N$ interaction is switched off. The neutrons and the protons oscillate in phase, and the $\Lambda$ particles oscillate out of phase with the nucleons. We can thus interpret this mode as a dipole oscillation of the $\Lambda$ particles against the core nucleus $^{16}$O, similar to the soft dipole motion of a valence neutron in halo nuclei [@Ikeda10]. Giant monopole resonance and incompressibility ---------------------------------------------- Giant monopole resonances, the so called “breathing mode”, are intimately related to the incompressibility of nuclear matter [@Blaizot1995; @Li2010; @K09; @Colo04; @Khan2010; @Piekarewicz2009; @Blaizot1980], which plays an important role in neutron stars. It has been shown that the EOS of infinite nuclear matter is softened when hyperons($\Lambda, \Xi, \Sigma$) emerge at high densities, and as a consequence the maximum mass of neutron stars becomes smaller [@Dexheimer2008; @Glendenning1991; @ST94]. It is thus of interest to investigate the $\Lambda$-impurity effect of giant monopole resonances in finite nuclei. Figure \[monopole\] shows the strength function for the isoscalar monopole responses of $^{16}$O and $^{\,\,18}_{\Lambda\Lambda}$O nuclei. The meaning of each line is the same as in Fig. \[multi\]. For a comparison, the figure also shows the strength function for $^{208}$Pb and $^{208}_{~\Lambda\Lambda}$Pb. As in the other multipolarities discussed in the previous subsection, the strength distributions are shifted towards high energies when $\Lambda$ hyperons are added, and also the difference between the solid and the dotted lines is small, indicating that the residual $\Lambda N$ and $\Lambda\Lambda$ interactions play a minor role. ![(Color online) The strength distributions for the isoscalar monopole mode for the $^{16}$O and $^{\,\,18}_{\Lambda\Lambda}$O nuclei (the top panel) and for the $^{208}$Pb and $^{210}_{\Lambda\Lambda}$Pb nuclei (the bottom panel). The meaning of each line is the same as in Fig. \[multi\]. []{data-label="monopole"}](O16mono2.eps "fig:"){width="0.70\linewidth"} ![(Color online) The strength distributions for the isoscalar monopole mode for the $^{16}$O and $^{\,\,18}_{\Lambda\Lambda}$O nuclei (the top panel) and for the $^{208}$Pb and $^{210}_{\Lambda\Lambda}$Pb nuclei (the bottom panel). The meaning of each line is the same as in Fig. \[multi\]. []{data-label="monopole"}](Pb208mono2.eps "fig:"){width="0.70\linewidth"} Figure \[drho\_O16GMR\] shows the transition densities for the giant monopole resonances corresponding to the states indicated by the arrows in the Fig.\[monopole\] (that is, those states at $E=22.2$ MeV and 21.1 MeV for $^{\,\,18}_{\Lambda\Lambda}$O and $^{16}$O, respectively, and at 14.6 MeV and 14.2 MeV for $^{210}_{\Lambda\Lambda}$Pb and $^{208}$Pb, respectively). The meaning of each line is the same as in Fig. \[drho\_O16\]. For the oxygen nuclei, when $\Lambda$ hyperons are added, the amplitude of the transition density for the neutrons decreases by about $20\%$ while that for the protons remains almost the same. The amplitude of the $\Lambda$ transition density is about 10 times smaller than that of the nucleons. It is interesting to notice that the $\Lambda$ hyperons oscillate out of phase with the nucleons. These features are qualitatively the same for the lead nuclei as well, although the changes in the transition densities are much smaller compared to the oxygen isotopes. ![(Color online) The transition densities for the isoscalar giant monopole resonances of $^{\,\,18}_{\Lambda\Lambda}$O (the left panels) and $^{210}_{\Lambda\Lambda}$Pb (the right panels), with comparisons to the transition densities for $^{16}$O and $^{208}$Pb. The meaning of each line is the same as in Fig. \[drho\_O16\].[]{data-label="drho_O16GMR"}](drho_O16GMR.eps "fig:"){width="0.48\linewidth"} ![(Color online) The transition densities for the isoscalar giant monopole resonances of $^{\,\,18}_{\Lambda\Lambda}$O (the left panels) and $^{210}_{\Lambda\Lambda}$Pb (the right panels), with comparisons to the transition densities for $^{16}$O and $^{208}$Pb. The meaning of each line is the same as in Fig. \[drho\_O16\].[]{data-label="drho_O16GMR"}](drho_Pb208GMR.eps "fig:"){width="0.48\linewidth"} According to Blaizot, the effective incompressibility modulus $K_A$ for finite nuclei without $\Lambda$ hyperons is defined as [@Blaizot1980], $$K_A=\frac{m_N}{\hbar^2}\mathcal{E}^2\langle r^2 \rangle, \label{blaizot1}$$ where $\sqrt{\langle r^2 \rangle}$ is the root mean square radius, and $\mathcal{E}^2=m_{1}/m_{-1}$ (see Eq. (\[ewsr\]) for the definition of the $k$-th energy-weighted sum-rule, $m_k$). When the $\Lambda$ hyperons are present, this formula is modified as, $$K_A=\frac{m_N}{\hbar^2}\mathcal{E}^2\langle r^2 \rangle \left( \frac{N+Z}{A}\frac{\langle r^2 \rangle_{\rm n+p}} {\langle r^2 \rangle} + \frac{m_NN_\Lambda}{m_\Lambda A}\frac{\langle r^2 \rangle_\Lambda} {\langle r^2 \rangle} \right)^{-1}, \label{blaizot2}$$ where we have used Eqs. (3.45) and (3.47) in Ref.[@Blaizot1980] and the energy-weighted sum-rule for the isoscalar monopole transition, $$m_1(L=0)= \frac{2\hbar^2}{m_N}(N+Z)\langle r^2 \rangle_{\rm{n+p}} + \frac{2\hbar^2}{m_\Lambda}N_\Lambda\langle r^2 \rangle_{\Lambda}. \label{ewsr1}$$ In Eq. (\[blaizot2\]), $\sqrt{\langle r^2 \rangle}_{\rm{n+p}}$ is the root mean square radius of the core nucleus. Notice that Eq. is reduced to Eq. when $N_\Lambda=0$. In Table \[centroid0\], we list the centroid energy $\overline{E}_{0^+}$, $\mathcal{E}$ for the isoscalar monopole modes, the root-mean-square radii, $\sqrt{\langle r^2 \rangle}$ and $\sqrt{\langle r^2 \rangle_{\rm n+p}}$, and the effective incompressibility, $K_A$, calculated according to Eqs. (\[blaizot1\]) and (\[blaizot2\]). When $\Lambda$ hyperons are added, the centroid energies increase by $1.9$ MeV for $^{16}$O and $0.4$ MeV for $^{208}$Pb, and the rms radii for the core nucleus, $\sqrt{\langle r^2\rangle_{\rm{n+p}}}$, decrease by $0.04$ fm for $^{16}$O and $0.01$ fm for $^{208}$Pb. As we have shown, the increase of the centroid energies is mainly due to the change of single-particle levels, and the residual $\Lambda N$ and $\Lambda\Lambda$ interactions give only a minor effect. The decrease of the rms radii is attributed to the attractive $\Lambda N$ interaction, that is, the shrinkage effect of $\Lambda$ hyperons. The effective incompressibility, $K_A$, increases for both the nuclei studied when $\Lambda$ hyperons are added. $\overline{E}_{0^+}$(MeV) $\mathcal{E}$(MeV) $\sqrt{\langle r^2 \rangle_{\rm{n+p}}}$ $\sqrt{\langle r^2 \rangle}$(fm) $K_A$(MeV) ------------------------------- --------------------------- -------------------- ----------------------------------------- ---------------------------------- ------------ $^{16}$O $22.4$ $21.7$ $2.68$ $2.68$ $81.6$ $^{\,\,18}_{\Lambda\Lambda}$O $24.3$ $23.5$ $2.64$ $2.58$ $90.0$ $^{208}$Pb $14.1$ $14.0$ $5.56$ $5.56$ $146$ $^{210}_{\Lambda\Lambda}$Pb $14.5$ $14.4$ $5.55$ $5.53$ $153$ : Properties of the isoscalar monopole responses obtained with the Skyrme HF+RPA method. $\overline{E}_{0^+}$ is the centroid energy, and $\mathcal{E}$ is defined as $\mathcal{E}=\sqrt{m_1/m_{-1}}$. $\sqrt{\langle r^2 \rangle_{\rm{n+p}}}$ and $\sqrt{\langle r^2 \rangle}$ are the root mean square radii for the core nuclei and that for the total densities, respectively. $K_A$ is the effective nuclear incompressibility defined by Eqs. (\[blaizot1\]) and (\[blaizot2\]). \[centroid0\] ![(Color online) The binding energy per particle in infinite nuclear matter for the fraction of $\Lambda$ particle of $x_\Lambda$=0 (the dashed line) and $x_\Lambda$=0.11 (the solid line). The neutron and the proton densities are set to be equal, that is, $\rho_n=\rho_p=(1-x_\Lambda)\rho/2$. ](fig7_2.eps) The increase of the effective incompressibility should reflect the properties of infinite nuclear matter. In order to assess this, Fig. 7 shows the binding energy per particle in infinite nuclear matter, $E/A$, given as $$\begin{aligned} \frac{E}{A}&=&\frac{H}{\rho} =\frac{H_N}{\rho}+\frac{1}{\rho}\left[\frac{\hbar^2}{2m_\Lambda}\tau_\Lambda +t_0^\Lambda\left(1+\frac{x_0^\Lambda}{2}\right)\rho_N\rho_\Lambda \right.\nonumber \\ &&+\frac{3}{8}t_3^\Lambda\rho_\Lambda\rho_N^2 +\frac{1}{4}(t_1^\Lambda+t_2^\Lambda)(\tau_\Lambda\rho_N+\tau_N\rho_\Lambda) \nonumber \\ &&+\frac{\lambda_0}{4}\rho_\Lambda^2+\frac{1}{8}(\lambda_1+3\lambda_2)\rho_\Lambda\tau_\Lambda \nonumber \\ &&\left.+\frac{\lambda_3}{4}\rho_\Lambda^2\rho_N^\alpha\right],\end{aligned}$$ where $H_N$ is the nucleon part of the energy density evaluated in infinite matter. $\rho$ is the total density, while $\rho_\Lambda\equiv x_\Lambda \rho$ and $\rho_N\equiv (1-x_\Lambda) \rho$ are the densities of $\Lambda$ particles and nucleons, respectively. The kinetic energy densities $\tau_N$ and $\tau_\Lambda$ are evaluated as $\tau_N=3\rho_N k_{FN}^2/5$ and $\tau_\Lambda=3\rho_\Lambda k_{F\Lambda}^2/5$, respectively, where the Fermi momenta are given by $k_{FN}=(3\pi^2\rho_N/2)^{1/3}$ and $k_{F\Lambda}=(3\pi^2\rho_\Lambda)^{1/3}$. The dashed and the solid lines in Fig. 7 show the results with the $\Lambda$ fraction of $x_\Lambda$=0 and $x_\Lambda$=2/18, respectively. Here, we have assumed that the neutron and the proton densities are the same, $\rho_n=\rho_p=\rho_N/2$. One can see that the addition of $\Lambda$ particles shifts the equilibrium density $\rho_0$ towards a high density, that is, $\rho_0=0.161$ fm$^{-3}$ for $x_\Lambda$=0 and $\rho_0=0.185$ fm$^{-3}$ for $x_\Lambda$=2/18. The incompressibility for infinite nuclear matter, $$K_\infty=9\rho_0^2\,\left(\frac{d^2(E/A)}{d\rho^2}\right)_{\rho=\rho_0},$$ is $K_\infty=217$ MeV for $x_\Lambda$=0 and $K_\infty=239$ MeV for $x_\Lambda=2/18$, agreeing with the increase of effective incompressibility of the finite hypernuclei. Notice that this observation does not contradict to the fact that hyperons soften the equation of state for infinite nuclear matter relevant to neutron stars. That is, in neutron star calculations, the emergence of hyperons takes place at high densities and nucleons are the only constituents at the normal density, when the beta stability condition is imposed. In contrast, Fig. 7 shows the effect of hyperons on the incompressibility defined at the equilibrium density. Even though the beta stability condition does not hold there if the $\Lambda$ fraction is finite, this EOS is more relevant to giant monopole resonances of finite hypernuclei. SUMMARY ======= We have extended the Skyrme-HF plus RPA schemes to calculations for vibrational excitations of double-$\Lambda$ hypernuclei. We have applied it to the electric dipole (E1), the quadrupole (E2), and the octupole (E3) modes of excitations in the $^{\,\,18}_{\Lambda\Lambda}$O hypernucleus. We have shown that the strength distributions shift towards high energies for all the modes when the $\Lambda$ hyperons are added to $^{16}$O. This is the case both for the low-lying quadrupole and octupole states and for giant resonances. At the same time, the electromagnetic transition probabilities also decrease. We have argued that these features are mainly caused by the change in the single-particle energies, whereas the residual $\Lambda N$ and $\Lambda\Lambda$ interactions play a minor role. The calculated transition densities show that the peak of the transition densities are shifted towards inside and the height of the peaks slightly changes due to the impurity effect of $\Lambda$ hyperons. For the E1 strength, we have found a new peak at low-lying energy, that is absent in the E1 response of the core nucleus. From the analysis of the transition density, we have shown that this state corresponds to an oscillation of the $\Lambda$ particles against the core nucleus. We have also discussed the $\Lambda$-impurity effect on the isoscalar monopole vibration of $^{\,\,18}_{\Lambda\Lambda}$O and $^{210}_{\Lambda\Lambda}$Pb, and the incompressibility of infinite nuclear matter in the presence of $\Lambda$ hyperons. When the $\Lambda$ hyperons are added, the strength distributions are shifted to higher energies and thus the centroid energies increase, similarly to the other multipole transitions. We have shown that the transition density for the $\Lambda$ particles behave rather differently from the transition densities for the neutrons and the protons. The increase of the centroid energy for the giant monopole resonance implies that $\Lambda$ particles increase the nuclear incompressibility, when hyperons were emerged at the equilibrium density. In this paper, we have studied several collective vibrational motions, taking the $^{\,\,18}_{\Lambda\Lambda}$O hypernucleus as examples. It would be an interesting future work to study systematically the $\Lambda$-impurity effect on the collective excitations of other double-$\Lambda$ hypernuclei. In particular, the low-lying dipole mode, originated from a dipole oscillation of the $\Lambda$ particles against the core nucleus, would be interesting to study. For this purpose, we would have to extend our formalism by including the pairing correlations with the quasi-particle RPA (QRPA). Another interesting extension is to study the collective excitations of single-$\Lambda$ hypernuclei, although the broken time-reversal symmetry will have to be taken into account correctly there. We thank G. Colo and H. Sagawa for discussions. This work was supported by the Japanese Ministry of Education, Culture, Sports, Science and Technology by Grant-in-Aid for Scientific Research under the program number (C) 22540262. Energy densities and mean-field potentials for hypernuclei ========================================================== In this Appendix A, we summarize the explicit formulae for the energy densities $H_{N\Lambda}$ and $H_{\Lambda\Lambda}$ in Eq. and the mean-field potentials in Eq.. The energy density $H_{N\Lambda}$, due to the Skyrme-type $\Lambda N$ interaction given by Eq., reads [@Rayet], $$\begin{split} &H_{N\Lambda}({\mbox{\boldmath $r$}})= \frac{\hbar^2}{2m_\Lambda}\tau_\Lambda+t_0^\Lambda \left(1+\frac{1}{2}x_0^\Lambda\right) \rho_N\rho_\Lambda\\ &+\frac{1}{4}(t_1^\Lambda+t_2^\Lambda)(\tau_\Lambda\rho_N+\tau_N\rho_\Lambda) +\frac{1}{8}(3t_1^\Lambda-t_2^\Lambda) ({\mbox{\boldmath $\nabla$}}\rho_N\cdot{\mbox{\boldmath $\nabla$}}\rho_\Lambda)\\ &+\frac{1}{2}W_0^\Lambda({\mbox{\boldmath $\nabla$}}\rho_N\cdot{\mbox{\boldmath $J$}}_\Lambda +{\mbox{\boldmath $\nabla$}}\rho_\Lambda\cdot{\mbox{\boldmath $J$}}_N) +\frac{1}{4}t_3^\Lambda\rho_\Lambda(\rho_N^2+2\rho_n\rho_p), \end{split} \label{eln}$$ while the $\Lambda\Lambda$ part, $H_{\Lambda\Lambda}$, originated from the $\Lambda\Lambda$ interaction given by Eq. reads[@Lanskoy1998], $$\begin{split} &H_{\Lambda\Lambda}({\mbox{\boldmath $r$}})= \frac{1}{4}\lambda_0\rho_\Lambda^2 +\frac{1}{8}(\lambda_1+3\lambda_2)\rho_\Lambda\tau_\Lambda\\ &+\frac{3}{32}(\lambda_2-\lambda_1)\rho_\Lambda{\mbox{\boldmath $\nabla$}}^2\rho_\Lambda +\frac{1}{4}\lambda_3\rho_\Lambda^2\rho_N^\alpha. \end{split} \label{ell}$$ Here, $\rho_b=\rho_b({\mbox{\boldmath $r$}})$, $\tau_b=\tau_b({\mbox{\boldmath $r$}})$, ${\mbox{\boldmath $J$}}_b={\mbox{\boldmath $J$}}_b(r)$ are the number, the kinetic energy and the spin densities, respectively ($b=p,n$, or $\Lambda$). The indices $N$, $p$, $n$ and $\Lambda$ are the nucleon, the proton, the neutron and the $\Lambda$ hyperon, respectively. After taking variation of the energy in Eq. with respect to the densities, we obtain the Skyrme-Hartree-Fock equation given by Eq.. The mean-field potentials in Eq. are given by $$\begin{split} &U_{q\Lambda}({\mbox{\boldmath $r$}}) =t_0^\Lambda\left(1+\frac{1}{2}x_0^\Lambda\right)\rho_\Lambda +\frac{1}{4}\left(t_1^\Lambda+t_2^\Lambda\right)\tau_\Lambda\\ &-\frac{1}{8}\left(3t_1^\Lambda-t_2^\Lambda\right){\mbox{\boldmath $\nabla$}}^2\rho_\Lambda -\frac{1}{2}W_0^\Lambda{\mbox{\boldmath $\nabla$}}\cdot{\mbox{\boldmath $J$}}_\Lambda\\ &+\frac{1}{2}W_0^\Lambda{\mbox{\boldmath $\nabla$}}\rho_\Lambda\cdot (-i{\mbox{\boldmath $\nabla$}}\times{\mbox{\boldmath $\sigma$}}) +\frac{1}{2}t_3^\Lambda\rho_\Lambda(2\rho_N-\rho_q), \end{split}$$ $$\begin{split} &U_{\Lambda N}({\mbox{\boldmath $r$}}) =t_0^\Lambda\left(1+\frac{1}{2}x_0^\Lambda\right)\rho_N +\frac{1}{4}\left(t_1^\Lambda+t_2^\Lambda\right)\tau_N\\ &-\frac{1}{8}\left(3t_1^\Lambda-t_2^\Lambda\right){\mbox{\boldmath $\nabla$}}^2\rho_N -\frac{1}{2}W_0^\Lambda{\mbox{\boldmath $\nabla$}}\cdot{\mbox{\boldmath $J$}}_N\\ &+\frac{1}{2}W_0^\Lambda{\mbox{\boldmath $\nabla$}}\rho_N\cdot (-i{\mbox{\boldmath $\nabla$}}\times{\mbox{\boldmath $\sigma$}}) +\frac{1}{4}t_3^\Lambda(\rho_N^2+2\rho_n\rho_p), \end{split}$$ and $$\begin{split} U_{\Lambda\Lambda}({\mbox{\boldmath $r$}}) &=\frac{1}{2}\lambda_0\rho_\Lambda +\frac{1}{8}(\lambda_1+3\lambda_2)\tau_\Lambda\\ &+\frac{3}{16}(\lambda_2-\lambda_1){\mbox{\boldmath $\nabla$}}^2\rho_\Lambda +\frac{1}{2}\lambda_3\rho_\Lambda\rho_N^\alpha. \end{split}$$ Note that the index $q$ refers only to the proton and the neutron. The effective mass for the nucleons and the $\Lambda$ hyperons in Eq. are given by $$\frac{\hbar^2}{2m_q^}=\frac{\hbar^2}{2m_N} +\frac{1}{4}(t_1^\Lambda+t_2^\Lambda)\rho_\Lambda({\mbox{\boldmath $r$}}), \label{effect_m1}$$ and $$\frac{\hbar^2}{2m_{\Lambda}^}=\frac{\hbar^2}{2m_\Lambda} +\frac{1}{4}(t_1^\Lambda+t_2^\Lambda)\rho_N({\mbox{\boldmath $r$}}) +\frac{1}{8}(\lambda_1+3\lambda_3)\rho_\Lambda({\mbox{\boldmath $r$}}), \label{effect_m2}$$ respectively. $\Lambda N$ and $\Lambda\Lambda$ residual interactions ====================================================== \[appb\] The matrix elements for a particle-hole residual interaction $v_{\rm res}$ are given as[@RingSchuck; @Rowe] $$\begin{split} v_{ph'hp'}&=\langle p(h)^{-1}LK\|v_{\rm{res}}\|p'(h')^{-1}LK\rangle,\\ v_{pp'hh'}&=\langle p(h)^{-1}LK,p'(h')^{-1}L\bar{K}\|v_{\rm{res}}\|\rangle, \end{split}$$ where $L$ is the multipolarity for the particle-hole excitations and $K$ is its $z$-component. For hypernuclei, the residual interaction can be separated into two parts, $v_{\rm{res}}=v_{\rm{res}}^{b_1b_2}(N)+v_{\rm{res}}^{b_1b_2}(\Lambda)$, where the indices $b_1$ and $b_2$ denote $p$, $n$ or $\Lambda$. The interaction $v_{\rm{res}}^{b_1b_2}(N)$ is due to the $NN$ residual interaction, whose explicit form can be found in [e.g.*]{} Refs.[@Liu1991; @Hamamoto1999; @Terasaki2005]. $v_{\rm{res}}^{b_1b_2}(\Lambda)$ is the additional term due to the $\Lambda N$ and the $\Lambda\Lambda$ residual interactions. These are given in the form of, $$\begin{split} &v_{\rm{res}}^{b_1b_2}(\Lambda)=\delta({\mbox{\boldmath $r$}}_{b1}-{\mbox{\boldmath $r$}}_{b2})\Big(a_{b_1b_2}\\ &+b_{b_1b_2}\Big[{\mbox{\boldmath $\nabla$}}_1^2 +{\mbox{\boldmath $\nabla$}}_2^2 +{\mbox{\boldmath $\nabla$}}_1'^2+{\mbox{\boldmath $\nabla$}}_2'^2 -({\mbox{\boldmath $\nabla$}}_1-{\mbox{\boldmath $\nabla$}}_1')({\mbox{\boldmath $\nabla$}}_2-{\mbox{\boldmath $\nabla$}}_2')\Big]\\ &+c_{b_1b_2}({\mbox{\boldmath $\nabla$}}_1+{\mbox{\boldmath $\nabla$}}_1')({\mbox{\boldmath $\nabla$}}_2+{\mbox{\boldmath $\nabla$}}_2')\Big), \end{split} \label{ph-int}$$ where $a_{b_1b_2}$, $b_{b_1b_2}$ and $c_{b_1b_2}$ are given by $$\begin{split} a_{qq'}=&\frac{t_3^\Lambda}{4}\rho_{\Lambda} \Big(3-{\mbox{\boldmath $\sigma$}}\cdot{\mbox{\boldmath $\sigma$}}'-{\mbox{\boldmath $\tau$}}\cdot{\mbox{\boldmath $\tau$}}' -{\mbox{\boldmath $\sigma$}}\cdot{\mbox{\boldmath $\sigma$}}'{\mbox{\boldmath $\tau$}}\cdot{\mbox{\boldmath $\tau$}}'\Big)\\ &+\frac{\lambda_3}{4}\alpha(\alpha-1)\rho_N^{\alpha-2}\rho_\Lambda^2\\ b_{qq'}&=c_{qq'}=0, \end{split}$$ $$\begin{split} a_{\Lambda q}&=t_0^\Lambda\left(1+\frac{x_0^\Lambda}{2}\right) +\frac{t_0^\Lambda x_0^\Lambda}{2}{\mbox{\boldmath $\sigma$}}_1\cdot{{\mbox{\boldmath $\sigma$}}_2}\\ &+t_3^\Lambda\left(\rho_N-\frac{\rho_q}{2}\right) +\frac{\lambda_3}{2}\alpha\rho_N^{\alpha-1}\rho_\Lambda\\ b_{\Lambda q}&=-\frac{1}{8}(t_1^\Lambda+t_2^\Lambda),\quad c_{\Lambda q}=\frac{1}{8}(t_1^\Lambda-3t_2^\Lambda).\\ \end{split}$$ for ($q,q'=p$ or $n$), and $$\begin{split} a_{\Lambda\Lambda}& =\frac{1}{2}\lambda_0(1-{\mbox{\boldmath $\sigma$}}_1\cdot{\mbox{\boldmath $\sigma$}}_2) +\frac{1}{2}\lambda_3\rho_N^\alpha(1-{\mbox{\boldmath $\sigma$}}_1\cdot{\mbox{\boldmath $\sigma$}}_2)\\ b_{\Lambda\Lambda}&=-\frac{1}{16}\Big( \lambda_1(1-{\mbox{\boldmath $\sigma$}}_1\cdot{\mbox{\boldmath $\sigma$}}_2) +\lambda_2(3+{\mbox{\boldmath $\sigma$}}_1\cdot{\mbox{\boldmath $\sigma$}}_2)\Big)\\ c_{\Lambda\Lambda}&=\frac{1}{16}\Big( \lambda_1(1-{\mbox{\boldmath $\sigma$}}_1\cdot{\mbox{\boldmath $\sigma$}}_2) -3\lambda_2(3+{\mbox{\boldmath $\sigma$}}_1\cdot{\mbox{\boldmath $\sigma$}}_2)\Big), \end{split}$$ for the $\Lambda\Lambda$ terms.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Understanding the particle-scale transition from elastic deformation to plastic flow is central to making predictions about the bulk material properties and response of disordered materials. To address this issue, we perform experiments on flow-stabilized solids composed of micron-scale spheres within a microfluidic channel, in a regime where particle inertia is negligible. Each solid heap exists within a stress gradient imposed by the flow, and we track the positions of particles in response to single impulses of fluid-driven compression or decompression. We find that the resulting deformation field is well-decomposed into an affine field, with a constant strain profile throughout the solid, and a non-affine field. The magnitude of this non-affine response decays with the distance from the free surface in the long-time limit, suggesting that the distance from jamming plays a significant role in controlling the length scale of plastic flow. Finally, we observe that compressive pulses create more rearrangements than decompressive pulses, an effect that we quantify using the $D^2_\mathrm{min}$ statistic for non-affine motion. Unexpectedly, the time scale for the compression response is shorter than for decompression at the same strain (but unequal pressure), providing insight into the coupling between deformation and cage-breaking.' address: 'Department of Physics, North Carolina State University, Raleigh, NC, 27695, USA' author: - 'Carlos P. Ortiz, Robert Riehn and Karen E. Daniels' title: 'Nonaffine deformation under compression and decompression of a flow-stabilized solid' --- Introduction ============ Understanding how structural rearrangements in disordered solids differ from crystalline solids is central [@Schall2007; @Falk1998; @Falk2011] to achieving control of material properties such as resistance to flow [@Brady1993], sound propagation [@Kriegs2004], heat capacity [@Lubchenko2007] , and dielectric constants [@Bradshaw-Hajek2009]. For large deformations, the microscopic response differs non-perturbatively from the predictions of linear elasticity [@Bocquet2009]. Instead of linear deformations, phenomena such as shear banding [@Hays2000], yielding and plastic rearrangements [@Hebraud1997], and non-local effects [@Lu2000] are present. Recent experiments have explored non-affine deformations in 3D sheared colloidal glasses [@Chikkadi2012], 3D emulsions [@Knowlton2014], and 2D foams [@Twardos2005]. For sufficiently slow deformations, it is an open question whether the flow behavior [@Lerner2012] is controlled by the jamming transition, where moduli vanish as the packing approaches a critical packing fraction [@Liu2010]. In this paper, we present experiments quantifying the particle-scale deformation of flow-stabilized solids: particle heaps formed under controlled hydrodynamic stress (see Fig. \[fig:setup\]). These quasi-2D heaps are assembled via the slow accumulation of micron-scale particles against a barrier within a microchannel, and are found to be stable above Péclet number $\sim 1$ [@Ortiz2013]. For lower Péclet numbers (slower flows) the particles reversibly evaporate away from the solid. We have previously observed that the elastic modulus of the solid is proportional to the confining stress provided by the fluid flow [@Ortiz2014]. However, the amount of deformation of the solid in response to a flow perturbation is dependent on the sign of the perturbation: for piles prepared under identical conditions, compressions result in smaller strains than decompressions [@Ortiz2014]. At the bulk scale, this effect can be understood by considering an excluded volume equation of state, as in thermal systems, under the assumption of locally affine deformations. In this paper, we investigate how the particle-scale dynamics lead to deviations from the excluded volume argument at high deformations. ![(a) Schematic of channel geometry, not to scale, illustrating barrier with overflow. Streamlines are scaled versions of a calculation solving Stokes’ equations on the true device dimensions. (b) Image of central region of a heap of bidisperse particles. Dark regions are dimly fluorescent 600 nm particles and bright regions are brightly fluorescent 710 nm particles. White box indicates the location of the region in which deformations are analyzed. \[fig:setup\]](Fig1-setup.pdf){width="80.00000%"} Our experiments use sterically- and electrostatically-stabilized Brownian microspheres with a short Debye length (3 nm), so that the net interparticle interaction is well-approximated by a hard-sphere potential except near contact. A bidisperse mixture of particle sizes (5:4 diameter ratio) suppresses the nucleation of crystal domains. Using fluorescence microscopy, we measure the particle-scale deformation fields and characterize the response of the heap under compressive and decompressive loads created by changing the hydrodynamic stress. We characterize the influence of cooperative motions by measuring the degree to which the the deformation field locally deviates from global affine deformations. The affine (or homogeneous) component of the deformation field is the part that can be described by an affine transformation: rotation, shear, extension, or compression [@Wu2005]. After identifying the affine component of the deformation, the residual (or inhomogeneous) component is the non-affine deformation. For linearly-elastic materials, only affine deformations are present, but non-affine deformations can arise due to either thermally-driven cage-breaking events [@Weeks2002] or local rearrangements [@Falk1998]. Here, we quantify two effects: the total non-affine rearrangements, and the spatiotemporal dynamics of the response. We observe, as expected [@Leonforte2006; @Ellenbroek2009], that non-affine deformation fields typically exhibit mesoscale correlations. Furthermore, the non-affine deformations are about twice as large for compressive deformations as compared with decompressive deformations of similar size, and happen over a shorter time scale. For both compressive and decompressive deformations, non-affine deformations continue after affine deformations have completed. Experimental Setup ================== Our experiments begin by assembling a microsphere heap by flowing a dilute suspension against a barrier (see Fig. \[fig:setup\]a). The microchannel is fabricated to have a height $H = 897$ nm, higher than the height of a barrier ($h_b=694$ nm), so that the fluid overflow accumulates particles against the barrier of width $W=512~{\ensuremath{{\mu}\text{m}}}$. The heights are chosen to create a quasi-2D heap, shallow enough to suppress both stacked and non-stacked bilayer phases [@VanWinkle1986]. The suspension is pumped into the channel by compressing a reservoir at the inlet using a low pressure, piezoelectrically actuated, digital regulator (AirCom PRE1-UA1), at $P_0 = 10$ kPa above atmospheric pressure. After two hours of equilibration, the heap is $154~{\ensuremath{{\mu}\text{m}}}$ deep ($30^\circ$ angle of repose) and contains approximately 40,000 particles, as shown in Fig. \[fig:setup\]b. The coordinate system takes $\hat x$ parallel to the barrier and $\hat y$ perpendicular to the barrier, with the origin at center of the barrier; the fluid flow is in the $-\hat y$ direction. The dilute, aqueous suspension is prepared at a concentration of $\rho=180/(100~{\mu}\text{m})^2$ fluorescent microspheres. The particles are a bidisperse mixture of equal concentrations of 600 nm and 710 nm polystyrene microspheres ($\approx$6% polydispersity, elastic modulus $4$ GPa from Bangs Laboratories). We use steric and electrostatic stabilization (sulfate functionalized surface with $\zeta$-potential $=-60$ mV and coated with Triton X-100) to provide reversible inter-particle and channel-particle interactions. The suspending fluid is a density-matched aqueous solvent at pH 5.4, buffered by citric acid to prevent crystallization and with $17\%(w/v)$ sucrose to provide density-matching. The later suppresses segregation and sedimentation effects, important both at the barrier and at the inlet reservoir. Because the total particle brightness scales approximately with the particle volume, and we are working near the diffraction limit, the 600 nm particles appear dimmer than the 710 nm particles. This effect aids in tracking the motion of the particles. We quantify the affine and non-affine deformation due to a pulse of either compression ($\Delta P > 0$) or decompression ($\Delta P < 0$). The decompression pulse results from a change from 10 kPa to 0.5 kPa over a time scale of 10 ms, and the compression pulse returns the pressure to 10 kPa. This corresponds to $\Delta P/P_0 = 19$ for the compression pulse and $\Delta P/P_0 = 0.95$ for the decompression pulse. Imaging occurs in two phases. A full view of the initial heap is accessible with a $10\times$ objective, while experiments quantifying the particle motions require visualizing a zoomed-in region using a $60\times$ objective with a $4\times$ beam expander. The image was recorded by a CCD camera with $10 \times 10$ $\mu$m$^2$ pixels and an exposure time 36$\mu s$. As shown in Fig. \[fig:setup\]b, the zoomed measurement region is of size $40\,d \times 30\,d$ and is located adjacent to the barrier. For each pulse, we first acquire an image of a region of interest at the center of the heap prior to the pressure change, allowing us to extract the initial configuration of particle positions. Additional images, taken at 27 Hz, characterize the particle-scale response of the heap to the change in pressure. After a wait of 100 seconds, long enough for particles to settle onto a new, equilibrated configuration, we repeat this process for the compression pulse. Image Processing \[sec:imageproc\] ================================== ![Illustration of image analysis process. (a) Sample micrograph, with grayscale indicating raw data. (b) Edge detection between bright and dark regions.(c) Wiener Deconvolved Image. []{data-label="fig:imanalysis"}](Fig2-featdetect.pdf){width="\textwidth"} For either compressive or decompressive pulses, we first compare the initial and final configurations (separated by 100 sec), and quantify both the total deformation and the non-affine deformation. Second, using the series of frames immediately following the pulse, we track individual particles to identify non-affine effects on the local scale. Below, these are referred to as “long-time” and “short-time” dynamics, and require slightly different image-processing. For the long-time dynamics, the total distance traveled is on the order of a few particle diameters. Therefore, we first subtract the total affine deformation before performing particle tracking using the Blair-Dufresne implementation [@MatlabTracking] of the Grier-Crocker particle tracking algorithm [@Crocker1996]. #### Particle identification: Fig. \[fig:imanalysis\] summarizes how we obtain particle positions beginning from a raw image. We identify the location of each particle by performing a Wiener deconvolution on the raw image, using a Gaussian approximation to the point-spread function with full-width at half maximum of 540 nm. This value is found to maximize the contrast in the output image, as measured from the ratio of the standard deviation to the mean intensity, but is more effective at locating the large (bright) particles than the small (dim) particles. The resulting deconvolved image allows us to detect the centroid of each particle using Matlab’s local extended maxima method. #### Total deformation: We estimate the total affine deformation $\Delta y$ due to a single pulse by making a coarse-grained measurement of the particle displacements between an initial image and a final image. These two images are created by averaging 10 initial images $I_i(x,y)=\langle I_i(x,y,t)\rangle_t$ and 10 final images $I_f(x,y)=\langle I_f(x,y,t)\rangle_t$. We divide $I_f(x,y)$ into horizontal strips of width $2d$ and compute a cross-correlation with $I_i(x,y)$ to determine its displacement. We find that the cross-correlation is sharply-peaked function for strips of at least this width. Due to the large total strains, we perform particle pair matching between initial and final configurations based on particle positions from which the total affine deformation $\Delta y$ has already been subtracted. After this adjustment, pair identification proceeds as in the one-step particle tracking [@MatlabTracking], with the size of the search region selected to correspond to the estimate of the maximum non-affine displacement amplitude, plus an estimate of the error in the affine strain. #### Short-time particle tracking: In order to obtain particle trajectories during the full duration of the dynamics, we make several assumptions about the nature of valid trajectories. We limit the displacement per frame to $0.5$ [${\mu}\text{m}$]{}; this value is consistent with the total affine deformation rate determined above. In addition, we consider a particle’s identified size (brightness) in order to either split incorrect trajectories or or reconnect broken trajectories. Results ======= In previous experiments [@Ortiz2014], we observed that flow-stabilized solids exhibit a nonlinear stress-strain relationship in which the magnitude of the deformation of the [surface]{} of the flow-stabilized solid is well-described by $$\frac{\Delta y_\text{surface}}{y_\text{surface}} \propto \left(\frac{1}{1+\frac{\Delta P}{P_0}}-1\right). \label{eqn:gamma}$$ The success of this description is somewhat surprising, as Eqn. \[eqn:gamma\] does not contain any information about the distribution of stresses or strains [throughout]{} the flow-stabilized solid. The stress field within the solid is anticipated to be similar to that in a sedimentation experiment where particles “on top” of the sedimented material apply some stress on lower layers (in the limit of shallow sediments without side walls). We believe that the success in describing our experiments is due to the universality of the van-der-Waals thermal argument. However, that argument breaks down if non-affine motions occur, and we anticipate that the anticipated lower stress at the upstream (“top”) surface of the flow-stabilized solid is not fully characterized by the van-der-Waals argument. In the following, we first identify the distributions of particle displacements in the asymptotic long-term limit, before following individual trajectories through compression and decompression. Our particular interest is in the associated particle-scale non-affine motions and their dependence on the sign of $\Delta P$. Total deformation \[sec:longTdyn\] ---------------------------------- ![Determination of the total (long-time) affine field from particle tracks for $\Delta P=+9.5$ kPa. (a) Image difference of blobs with size corresponding to $1\sigma$ of the position uncertainty in initial (blue) and final (yellow) configuration. (b) The associated affine deformation field. The dashed line is a linear regression with $\gamma = 3.4\pm 0.2$%. []{data-label="fig:longTaffine"}](Fig3-compress.pdf){width="5in"} We find displacements of individual particles in the heap immediately before and 100 s after a compressive/decompressive pulse through a two-step analysis. Following the homogeneous strain field assumption from our prior work [@Ortiz2014], we first use the image cross-correlation analysis of images before and after deformation to obtain a global estimate of the affine strain field. We then use the affine transformation identified by the cross-correlation analysis as a scaffold for the matching of particles in the images before and after deformation. In Fig. \[fig:longTaffine\]a, we show an example of the particle locations after Wiener deconvolution and prior to finding the centroids, for both $I_i$ (red, before compression) and $I_f$ (white, after compression). By tracking each centroid, we can plot the local displacement $\Delta y$ as a function of $y$-position within the heap. As shown in In Fig. \[fig:longTaffine\]b, the mean behavior is linear, confirming that the overall assumption of an affine deformation was sufficiently accurate. The best fit line to these points provides a measure of the strain: $\Delta y = \gamma_\infty y$ with $\gamma_\infty=-3.8\%$. The precision of these measurements is insufficient to estimate the expected higher-order (quadratic) term, although we expect one to be present due to a depth-dependent stress field. The observed linear behavior, combined with Eqn. \[eqn:gamma\], suggests that the packing fraction is close to invariant along the $y$-direction. ![Deformation field for a compression given by a pressure change at the inlet of $\Delta P=+9.5$ kPa. The axes in both figures are the $x$-position and the $y$-position in units of particle diameters. (a) Full deformation field. (b) Affine deformation field. (b) Non-affine deformation field, magnified by a factor of two.[]{data-label="fig:longTcompress"}](Fig4-compression.pdf){width="6in"} ![Deformation field for a decompression given by a pressure change at the inlet of $\Delta P=-9.5$ kPa. The axes in both figures are the $x$-position and the $y$-position in units of particle diameters. (a) Full deformation field. (b) Affine deformation field. (b) Non-affine deformation field, magnified by a factor of two.[]{data-label="fig:longTdecompress"}](Fig5-decompression.pdf){width="6in"} To obtain the total non-affine deformation field, we subtract the local affine motion from each displacement vector, as in [@Falk1998]. Figs. \[fig:longTcompress\] and \[fig:longTdecompress\] (both compression and decompression) show the total, affine, and non-affine displacement fields, for comparison. Importantly, we observe bands of correlated motions, as expected from [@Leonforte2006; @Ellenbroek2009]. Because the total deformation field is not robust in tracking individual bead pairs over long times, we next examine the short-time dynamics. Dynamics of individual particle tracks \[sec:shortTdyn\] --------------------------------------------------------- ![Particle Tracks and Packing Fraction. (a) Particle tracks during compression. Tracks are colored to introduce contrast. (b) Particle tracks during decompression. []{data-label="fig:ptracks"}](Fig6-tracks.pdf){width="5in"} Using the estimate of strain provided by Fig. \[fig:longTaffine\]b, we track the fast dynamics arising from a compressive/decompressive pulse. For each frame, we first subtract the estimated affine deformation, based on the fraction of total strain which should have accumulated so far (see §\[sec:imageproc\]). This temporary adjustment allows for the correct association of particle centroids from frame to frame. Once the particle trajectories have been assembled based on these adjusted trajectories, we return to using the original positions detected for each particle centroid. The analyses that follow are based on the non-adjusted deformations that result from that tracking. Sample trajectories are shown in Fig. \[fig:ptracks\]. While the total deformation field is of the similar magnitude under compression and decompression, we find a more pronounced scrambling of the particle trajectories under compression, as compared to decompression. Below, we quantify both the affine and non-affine contributions to these trajectories. #### Affine deformations: Data was binned within strips along the $x$-direction, providing ensembles of particles-dynamics sampled as a function of depth. Figs. \[fig:shortT-decompress\]a and \[fig:shortT-compress\]a show the mean deformation field $\Delta y$ as a function of $y$-position for compression and decompression, respectively. For both deformation directions, we find an exponential-like asymptotic approach to the final displacement magnitude. The depth-dependence of the asymptotic value of $\Delta y$ (Fig. \[fig:shortT-decompress\]b and \[fig:shortT-compress\]b) demonstrates the same linear relationship originally shown in Fig. \[fig:longTaffine\]b. The resulting slope ($\gamma_\infty = \Delta y/y$) quantifies the dynamics of affine reorganization. We find a marked difference between decompression (Fig. \[fig:shortT-decompress\]c) and compression (Fig. \[fig:shortT-compress\]c) in that decompressions are far slower than compressions, and that the strain curves for decompression collapse better onto a single dynamic curve. To quantify the difference, we make the Ansatz of a single-exponential approach to the asymptotic deformation $$\label{eqn:affinevt} \Delta y = \gamma_\infty \, y \left( 1-e ^{- \frac{\Delta t}{\tau}} \right)$$ where $\Delta y$ is the particle displacement after a time interval $\Delta t$, $\tau$ is a characteristic time scale of particle rearrangements, and $\gamma_\infty$ is the asymptotic strain. Note that the value $\gamma_\infty$ here is a fitting parameter; we find its value to be consistent with the estimate from the long-time dynamics. As shown in both panels (d), this exponential form is a good fit for the decompression pulses with $\tau_\text{affine,decompression}=0.28\pm 0.05$ s. For compressive deformations, a single-exponential form is less consistent with the observed dynamics. Instead, there appears to be a two-step process of compression in which the viscous stress increase acts nearly instantaneously throughout the solid, while stresses due to particle-particle contacts propagate at a distinct speed of sound from the immobile barrier on which the solid is formed. Given the two-step nature of the process under compression, we establish an upper bound on the relaxation time scale of $\tau_\text{affine,compression}=0.11\pm 0.05$ s. ![Affine Deformation Time-Series Analysis During Decompression. (a) Magnitude of correlated displacement field as a function of time at varying distances upstream of the barrier. (b) Long-time displacement amplitude $\Delta y$, as a function of distance upstream of the barrier, both in units of particle diameters. (c) Correlated displacement amplitude normalized by the long-time displacement amplitude. (d) Log-linear plot of growth curves in (c). []{data-label="fig:shortT-decompress"}](Fig7-affine_decompress.pdf){width="4in"} ![Affine Deformation Time-Series Analysis During Compression. (a) Magnitude of correlated displacement field as a function of time at varying $y$-positions. (b) Long-time displacement amplitude $\Delta y$, as a function of distance upstream of the barrier, both in units of particle diameters. (c) Correlated displacement amplitude normalized by the long-time displacement amplitude. (d) Log-linear plot of growth curves in (c). []{data-label="fig:shortT-compress"}](Fig8-affine_compress.pdf){width="4in"} ![Magnitude of non-affine deformation field $D^2_\text{min}$ as a function of time and depth for (a) compression and (b) decompression.[]{data-label="fig:D2minim"}](Fig9-D2minim.pdf){width="5in"} #### Non-affine deformations: We identify the non-affine contribution to the deformation field by subtracting the affine portion associated with the best-fit instantaneous value of the strain, which we designate $$\gamma(t)=\gamma_\infty\left( 1-e ^{- \frac{\Delta t}{\tau}} \right).$$ To quantify the resulting non-affinity, we use the $D_\text{min}^2$ measure [@Falk1998] defined by $$D_\text{min}^2(t)\equiv\sum_\text{neighbors}\left(\Delta \vec r(t) - \gamma(t)\Delta \vec r(t_0)\right)^2 .$$ Here, $\vec r(t)$ is the set of local displacement vectors connection nearest neighbors, and $t_0$ is the time immediately before the pressure step was applied. Fig. \[fig:D2minim\] shows the time-evolution of the non-affine displacement as a function of $y$-position during compression and decompression, respectively. In both graphs, $D^2_\mathrm{min}$ grows and ultimately saturates. Interestingly, the magnitude of the non-affine field scales linearly with depth as demonstrated by the collapse of $D^2_\text{min}/y^2$ data series shown in Fig. \[fig:D2minvt\]. This a surprising finding in light of the assumed constant strain throughout the flow-stabilized solid. Furthermore, the magnitude of non-affine deformations is approximately twice as large under compression than under decompression at near identical asymptotic strain $\gamma$. ![Plots of $D^2_\text{min}$ as a function of time, with the $y$-axis scaled either by the size of the large particles or the $y$-position at which the non-affine deformation is being probed. The legend in panel b applies to all panels. (a) Compression. $D^2_\text{min}$ scaled by $d^2$. (b) Compression. $D^2_\text{min}$ scaled by $y^2$. (c) Decompression. $D^2_\text{min}$ scaled by $d^2$. (d) Decompression. $D^2_\text{min}$ scaled by $y^2$. \[fig:D2minvt\]](Fig10-D2min.pdf){width="5in"} We observe that the growth of $D^2_\text{min}$ with time is smooth. We are able to determine a characteristic time for the approach to the asymptotic value of $D^2_\text{min}$ by fitting a single exponential approach, as we did for the affine deformation field. In doing so, we neglect the low background value of $D^2_\text{min}$ in steady-state flow-stabilized solids arising from Brownian motion. We find $\tau_\text{non-affine,compression}=0.36\pm0.08$ s, and $\tau_\text{non-affine,decompression}=0.44\pm 0.1$ s. Therefore, the non-affine field significantly lags the affine field for decompression ($\tau_\text{affine,decompression}=0.28\pm0.05$ s). For compression, where a single time scale is less well defined, and an upper bound on the affine time scale is $\tau_\text{affine,compression}=0.11\pm 0.05$ s, the non-affine field also lags the affine deformation. Discussion ========== We have observed particle-scale non-affine motions within flow-stabilized solids, and examined how their spatiotemporal dynamics depend on whether the deformation is compressive or decompressive. We observed the typical swirling regions often associated with non-affine deformations, arising through cooperatively rearranging regions. The magnitude of these effects is nearly twice as large under compression than under decompression, in spite of very similar total strains. We observe that compressive pulses (large $\Delta P/P_0$) generate more non-affine deformation, which is able to dissipate the effect of the pulse more quickly. Because the non-affine fields for both compressive/decompressive deformations occur after similar delays with respect to the affine deformations, is suggests that they are triggered by the affine deformations. In the context of caging behavior, this suggests that the affine deformation distorts the cages provided by the neighboring particles and thereby makes Brownian cage-breaking (non-affine deformation) more likely. Remarkably, this is the case even though the strain is approximately the same for decompression and compression. In probing the spatial dependence to the magnitude of the non-affine deformations (Fig. \[fig:D2minim\]), we observe that the degree of non-affinity increases with distance from the barrier. This effect can be rescaled by the position to indicate a universal behavior. The form of this dependence suggests $D^2_\text{min}\propto \frac{1}{p^2} \propto \frac{1}{K}$, for pressure $p$ and modulus $K$ [@Ortiz2014]. One interpretation is that the surface of the heap is less rigid (smaller $K$), and therefore more prone to undergoing non-affine deformations (higher $D^2_\text{min}$). Similar effects have been observed in numerical simulations [@Ellenbroek2009], where increasingly non-affine displacements are present in proximity to unjamming. The significance of the above conclusions to soft-matter particle assemblies is to reinforce the centrality of understanding non-affine rearrangements to link bulk properties of the material, such as its modulus and global stability, to local properties about the typical particle geometry and rearrangement timescales. Based on these results, this experimental setup opens the possibility to explore this connection, by studying multiple orders of magnitude of heap sizes, under dynamically tunable interaction potentials and heap geometry, maintaining the ability relate particle-scale rearrangement dynamics to bulk properties. By doing so, it should be possible to determine length and time scales at which localized and collective rearrangements have the greatest impact on bulk properties, and shed light on the general mechanisms by which it is feasible to control the bulk properties of soft matter systems. Acknowledgments {#acknowledgments .unnumbered} =============== We are grateful for support from the National Science Foundation through an NSF Graduate Fellowship, grants DMR-0644743, DMS-0968258, DMR-1121107, MRSEC/DMR-112090, and INSPIRE/EAR-1344280. Research was also supported by US Army Research Office–Division of Earth Materials and Processes grant 64455EV. This work was performed in part at the Cornell NanoScale Facility, a member of the National Nanotechnology Infrastructure Network, which is supported by the National Science Foundation (Grant ECCS-0335765). This work was also performed in part at North Carolina State University facilities: Nanofabrication Facility, Advanced Instrumentation Facility, and Education and Research Laboratory. References {#references .unnumbered} ========== [10]{} url \#1[[\#1]{}]{}urlprefix\[2\]\[\][[\#2](#2)]{} Schall P, Weitz D A and Spaepen F 2007 [Science]{} [318]{} 1895–9 Falk M L and Langer J S 1998 [Phys. Rev. E]{} [57]{} 7192–7205 Falk M L and Langer J 2011 [Annu. Rev. Condens. Matter Phys.]{} [2]{} 353–373 Brady J F 1993 [J. Chem. Phys.]{} [99]{} 567 Kriegs H, Petekidis G, Fytas G, Penciu R S, Economou E N and Schofield A B 2004 [J. Chem. Phys.]{} [121]{} 7849–54 Lubchenko V and Wolynes P G 2007 [Annu. Rev. Phys. Chem.]{} [58]{} 235–266 Bradshaw-Hajek B H, Miklavcic S J and White L R 2009 [Langmuir]{} [25]{} 1961–9 Bocquet L, Colin A and Ajdari A 2009 [Phys. Rev. Lett.]{} [103]{} 036001 Hays C C, Kim C P and Johnson W L 2000 [Phys. Rev. Lett.]{} [84]{} 2901–2904 Hébraud P, Lequeux F, Munch J and Pine D 1997 [Phys. Rev. Lett.]{} [ 78]{} 4657–4660 Lu C Y D, Olmsted P D and Ball R C 2000 [Phys. Rev. Lett.]{} [84]{} 642–645 Chikkadi V and Schall P 2012 [Phys. Rev. E]{} [85]{} 031402 Knowlton E D, Pine D J and Cipelletti L 2014 [Soft Matter]{} [10]{} 6931–6940 Twardos M and Dennin M 2005 [Phys. Rev. E]{} [71]{} 061401 Lerner E, D[ü]{}ring G and Wyart M 2012 [Proceedings of the National Academy of Sciences of the United States of America]{} [109]{} 4798–803 Liu A J and Nagel S R 2010 [Annual Review of Condensed Matter Physics]{} [1]{} 347–369 Ortiz C P, Riehn R and Daniels K E 2013 [Soft Matter]{} [9]{} 543 Ortiz C P, Daniels K E and Riehn R 2014 [Physical Review E]{} [90]{} 022304 Wu H C 2005 [Continuum Mechanics and Plasticity]{} (Chapman and Hall/CRC Press) Weeks E and Weitz D 2002 [Phys. Rev. Lett.]{} [89]{} 095704 Léonforte F, Tanguy A, Wittmer J and Barrat J L 2006 [Phys. Rev. Lett.]{} [97]{} 055501 Ellenbroek W G, Hecke M V and Saarloos W V 2009 [Physical Review E]{} [ 80]{} 61307 D and Murray C 1986 [Phys. Rev. A]{} [34]{} 562–573 Blair D and Dufresne E The matlab particle tracking code repository <http://physics.georgetown.edu/matlab/> Crocker J C and Grier D G 1996 [J. Colloid Interface Sci.]{} [179]{} 298–310	{ "pile_set_name": "ArXiv" }
--- abstract: 'Horizontal Federated learning (FL) handles multi-client data that share the same set of features, and vertical FL trains a better predictor that combine all the features from different clients. This paper targets solving vertical FL in an asynchronous fashion, and develops a simple FL method. The new method allows each client to run stochastic gradient algorithms without coordination with other clients, so it is suitable for intermittent connectivity of clients. This method further uses a new technique of perturbed local embedding to ensure data privacy and improve communication efficiency. Theoretically, we present the convergence rate and privacy level of our method for strongly convex, nonconvex and even nonsmooth objectives separately. Empirically, we apply our method to FL on various image and healthcare datasets. The results compare favorably to centralized and synchronous FL methods.' --- Introduction ============ Federated learning (FL) is an emerging machine learning framework where a central server and multiple clients (e.g., financial organizations) collaboratively train a machine learning model [@Konecn2016Fed; @mcmahan2017; @bonawitz2017]. Compared with existing distributed learning paradigms, FL raises new challenges including the difficulty of synchronizing clients, the heterogeneity of data, and the privacy of both data and models. Most of existing FL methods consider the scenario where each client has data of a different set of subjects but their data share many common features. Therefore, they can collaboratively learn a joint mapping from the feature space to the label space. This setting is also referred to data-partitioned or horizontal FL [@konevcny2016federated; @mcmahan2017]. Unlike the data-partitioned setting, in many learning scenarios, multiple clients handle data about the same set of subjects, but each client has a unique set of features. This case arises in e-commerce, financial, and healthcare applications [@hardy2017private]. For example, an e-commerce company may want to predict a customer’s credit using her/his historical transactions from multiple financial institutions; and, a healthcare company wants to evaluate the health condition of a particular patient using his/her clinical data from various hospitals [@sun2019privacy]. In these examples, data owners (e.g., financial institutions and hospitals) have different records of those users in their joint user base, so by combining their features, they can establish a more accurate model. We refer to this setting as feature-partitioned or vertical FL [@yang2019fed]. Compared to the relatively well-studied horizontal FL setting [@mcmahan2017communication], the vertical FL setting has its unique features and challenges [@hu2019fdml; @kairouz2019advances]. In horizontal FL, the global model update at a server is an additive aggregation of the local models, which are updated by each client using its own data. In contrast, the global model in vertical FL is the concatenation of local models, which are coupled by the loss function, so updating a client’s local model requires the information of the other clients’ models. Stronger model dependence in the vertical setting leads to challenges on privacy protection and communication efficiency. Prior art --------- We review prior work from the following three categories. [Federated learning.]{} Since the seminal work [@Konecn2016Fed; @mcmahan2017], there has been a large body of studies on FL in diverse settings. The most common FL setting is the horizontal setting, where a large set of data are partitioned among clients that share the same feature space [@konevcny2016federated]. To account for the personalization, multi-task FL has been studied in [@smith2017] that preserves the specialty of each client while also leveraging the similarity among clients, and horizontal FL with local representation learning has been empirically studied in [@liang2020think]. Agnostic FL has also been proposed in [@mohri2019icml], where the federated model is optimized for any target distribution formed by a mixture of the client distributions. Communication efficiency has been an important issue in FL. Popular methods generally aim to: c1) reduce the number of bits per communication round, including [@seide20141; @alistarh2017qsgd; @strom2015scalable; @aji2017sparse], to list a few; and, c2) save the number of communication rounds [@chen2018lag; @wang2018coop; @liu2019communication; @chen2020lasg]. [Privacy-preserving learning.]{} More recently, feature-partitioned vertical FL has gained popularity in the financial and healthcare applications [@hardy2017private; @yang2019fed; @niu2019; @kairouz2019]. Different from the aggregated gradients in the horizontal case, the local gradients in the vertical FL may involve raw data of those features owned by other clients, which raises additional concerns on privacy. Data privacy has been an important topic since decades ago [@yao1982]. But early approaches typically require expensive communication and signaling overhead when they are applied to the FL settings. Recently, the notion of differential privacy becomes popular because i) it is a quantifiable measure of privacy [@dwork2014algorithmic; @abadi2016; @dong2019gaussian]; and, ii) many existing learning algorithms can achieve differential privacy via simple modifications. In the context of learning from multiple clients, it has been studied in [@bonawitz2017; @hamm2016]. But all these approaches are not designed for the vertical FL models and the flexible client update protocols. [Asynchronous and parallel optimization.]{} Regarding methodology, asynchronous and parallel optimization methods are often used to solve problems with asynchrony and delays, e.g., [@recht2011]. For the feature-partitioned vertical FL setting in this paper, it is particularly related to the Block Coordinate Descent (BCD) method [@xu2013; @razaviyayn2013opt]. The asynchronous BCD and its stochastic variant have been developed under the condition of bounded delay in [@peng2016; @lian2017nips; @cannelli2016]. The Recent advances in this direction established convergence under unbounded delay with blockwise or stochastic update [@sun2017nips; @dutta2018slow]. However, all the aforementioned works consider the shared memory structure so that each computing node can access the entire dataset, which significantly alleviates the negative effect of asynchrony and delays. Moreover, the state-of-the-art asynchronous methods cannot guarantee i) the convergence when the loss is nonsmooth, and, ii) the privacy of the local update which is at the epicenter of FL. This work --------- The present paper puts forth an optimization method for vertical FL, which is featured by three main components. 1\. A general optimization formulation for vertical FL that consists of a global model and one local embedding model for each client. The local embedding model can be linear or nonlinear, or even nonsmooth. It maps raw data to compact features and, thus, reduces the number of parameters that need to be communicated to and from the global model. 2\. Flexible federated learning algorithms that allow intermittent or even strategic client participation, uncoordinated training data selections, and data protection by differential-privacy based methods (for specific loss functions, one can instead apply multiple-party secure computing protocols). 3\. Rigorous convergence analysis that establishes the performance lower bound and the privacy level. We have also numerically validated our vertical FL algorithms and their analyses on federated logistic regression and deep learning. Tests on image and medical datasets demonstrate the competitive performance of our algorithms relative to centralized and synchronous FL algorithms. Vertical federated learning =========================== This section introduces the formulation of vertical FL. Problem statement ----------------- Consider a set of $M$ clients: ${\cal M}:=\{1,\ldots,M\}$. A dataset of $N$ samples, $\{\mathbf{x}_n, y_n\}_{n=1}^N$, are maintained by $M$ local clients. Each client $m$ is also associated with a unique set of features. For example, client $m$ maintains feature $x_{n,m}\in\mathbb{R}^{p_m}$ for $n=1,\ldots,N$, where $x_{n,m}$ is the $m$-th block of $n$-th sample vector $\mathbf{x}_n:=[x_{n,1}^{\top},\cdots,x_{n,M}^{\top}]^{\top}$ at client $m$. Suppose the $n$-th label $y_n$ is stored at the server. To preserve the privacy of data, the client data $x_{n,m}\in\mathbb{R}^{p_m}$ are not shared with other clients as well as the server. Instead, each client $m$ learns a local (linear or nonlinear) embedding $h_m$ parameterized by $\theta_m$ that maps the high-dimensional vector $x_{n,m}\in\mathbb{R}^{p_m}$ to a low-dimensional one $h_{n,m}:=h_m(\theta_m;x_{n,m})\in\mathbb{R}^{\underline{p_m}}$ with $\underline{p_m}\ll p_m$. Ideally, the clients and the server want to solve $$\begin{aligned} \label{eq.prob} &F(\theta_0,{\boldsymbol{\theta}}):=\frac{1}{N}\sum\limits_{n=1}^N\ell\left(\theta_0, h_{n,1},\ldots, h_{n,M}; y_n\right)+\sum_{m=1}^M r(\theta_m)\nonumber\\ & \text{with}~~~ h_{n, m}:=h_m(\theta_m;x_{n,m}),~~~m=1,\cdots,M$$ where $\theta_0$ is the global model parameter kept at and learned by the server, and ${\boldsymbol{\theta}}:=[\theta_1^{\top},\cdots,\theta_M^{\top}]^{\top}$ concatenates the local models kept at and learned by local clients, $\ell$ is the loss capturing the accuracy of the global model parameters $\theta_0, \theta_1,\ldots, \theta_M$, and $r$ is the per-client regularizer that confines the complexity of or encodes the prior knowledge about the local model parameters. For problem , the local information of client $m$ is fully captured in the embedding vector $h_{n,m},~\forall n=1,\cdots, N$. Hence, the quantities that will be exchanged between server and clients are $\{h_{n,m}\}$ and the gradients of $\ell(\theta_0, h_{n,1},\ldots, h_{n,M}; y_n)$ with respect to (w.r.t.) $\{h_{n,m}\}$. See a diagram for VAFL implementation in Figure \[fig:a-diag\]. $^\dag$We can let Step 5 also send $h_{n,m}$ for those $n$ not selected in Step 4. We can re-order Steps 4–7 as 6, 7, then 4, and 5. They reduce information delay, yet analysis is unchanged. ![A diagram for VAFL. The local model at client $m$ is denoted as $\theta_m$ which generates the local embedding $h_{n,m}$.[]{data-label="fig:a-diag"}](global.pdf){width=".45\textwidth"} Asynchronous client updates {#subsec.avfl} --------------------------- For FL, we consider solving without coordination among clients. Asynchronous optimization methods have been used to solve such problems. However, state-of-the-art methods cannot guarantee i) the convergence when the mapping $h_{n,m}$ is nonlinear (thus the loss is nonsmooth), and, ii) the privacy of the update which is at the epicenter of the FL paradigm. We first describe our vertical asynchronous federated learning (VAFL) algorithm in a high level as follows. During the learning process, from the server side, it waits until receiving a message from an active client $m$, which is either\ i) a query of the loss function’s gradient w.r.t. to the embedding vector $h_{n,m}$; or,\ ii) a new embedding vector $h_{n,m}$ calculated using the updated local model parameter $\theta_m$. To response to the query i), the server calculates the gradient for client $m$ using its current $\{h_{n,m}\}$, and sends it to the client; and, upon receiving ii), the server computes the new gradient w.r.t. $\theta_0$ using the embedding vectors it currently has from other clients and updates its model $\theta_0$. For each interaction with server, each active client $m$ randomly selects a datum $x_{n,m}$, queries the corresponding gradient w.r.t. $h_{n,m}$ from server, and then it securely uploads the updated embedding vector $h_{n,m}$, and then updates the local model $\theta_m$. The mechanism that ensures secure uploading will be described in Section 4. Without introducing cumbersome iteration, client, and sample indexes, we summarize the asynchronous client updates in Algorithm \[alg:Async-SGD-FL\]. Specifically, since clients update the model without external coordination, we thereafter use $k$ to denote the global counter (or iteration), which increases by one whenever i) the server receives the new embedding vector $h_{n,m}$ from a client, calculates the gradient, and updates the server model $\theta_0$; and, ii) the corresponding client $m$ obtains the gradient w.r.t. $h_{n,m}$, and updates the local model $\theta_m$. Accordingly, we let $m_k$ denote the client index that uploads at iteration $k$, and $n_k$ denote the sample index used at iteration $k$. For notation brevity, we use a single datum $n_k$ for each uncoordinated update in the subsequent algorithms, but the algorithm and its analysis can be easily generalized to a minibatch of data ${\cal N}_k$. Let $\hat{g}_0^k$ denote the stochastic gradients of the loss at $n_k$-th sample w.r.t. server model $\theta_0$ as \[eqn:gradient\] $$\begin{aligned} \hat{g}_0^k:=\nabla_{\theta_0}\ell\big(\theta_0^k,h_{n_k,1}^{k-\tau_{n_k,1}^k},\ldots, h_{n_k,M}^{k-\tau_{n_k,M}^k}; y_{n_k}\big) \end{aligned}$$ and the gradients w.r.t. the local model $\theta_m$ as $$\begin{aligned} \hat{g}_m^k\!:=&\nabla_{\theta_m}\ell\big(\theta_0^k,h_{n_k,1}^{k-\tau_{n_k,1}^k},\ldots, h_{n_k,M}^{k-\tau_{n_k,M}^k}; y_{n_k}\big) \\ =&\nabla_{\theta_m} h_{n_k,m}^k\! \nabla_{h_{n_k,m}}\ell \big(\theta_0^k,h_{n_k,1}^{k-\tau_{n_k,1}^k}\!\!\!,\ldots, h_{n_k,M}^{k-\tau_{n_k,M}^k}\!; y_{n_k} \big) . \nonumber\end{aligned}$$ The delay for client $m$ and sample $n$ will increase via $$\begin{aligned} \label{eq.tau} \tau_{n,m}^{k+1} = \begin{cases} ~~~ 1, \quad\quad &m= m^k,\, n=n^k, \\ ~~~\tau_{n,m}^k+1, \quad\quad &{\rm otherwise}. \end{cases}\end{aligned}$$ With the above short-hand notation, at iteration $k$, the update at the [server side]{} is $\theta_0^{k+1}=\theta_0^k-\eta_0^k \hat{g}_0^k$. For the [active local client $m_k$]{} at iteration $k$, its update is $$\theta_{m_k}^{k+1}= \theta_{m_k}^k-\eta_{m_k}^k \hat{g}_{m_k}^k-\eta_{m_k}^k \nabla r(\theta_{m_k}^k),$$ and for the [other clients $m\neq m_k$]{}, the update is $$\theta_m^{k+1}=\theta_m^k,$$ where $\eta_m^k$ is the stepsize and $m_k$ is the index of the client responsible for the $k$th update. Types of flexible update rules ------------------------------ As shown in , the stochastic gradients are evaluated using delayed local embedding information [$h_m^{k-\tau_{n_k,m}^k}$]{} from each client $m$, where $\tau_{n_k,m}^k$ is caused by both asynchronous communication and stochastic sampling. To ensure convergence, we consider two reasonable settings on the flexible update protocols:\ 1. Uniformly bounded delay $D$. We can realize this by modifying the server behavior. During the training process, whenever the delay of $\tau_{n_k,m}^k$ exceeds $D(>0)$, the server immediately queries fresh $h_{n,m}$ from client $m$ before continuing the server update process.\ 2. Stochastic unbounded delay. In this case, the activation of each client is a stochastic process. The delays is determined by the hitting times of the stochastic processes. For example, if the activation of all the clients follows independent Poisson processes, the delays will be geometrically distributed.\ 3. $t$-synchronous update, $t>0$. While fully asynchronous update is most flexible, $t$-synchronous update is also commonly adopted. In this case, the server computes the gradient w.r.t. $\theta_0$ until receiving $\{h_{n,m}\}$ from $t$ different clients, and then updates the server’s model using the newly computed gradient. The $t$-synchronous updates have more stable performance empirically, which is listed in Algorithm \[alg:ksync-SGD-FL\]. Convergence analysis {#sec:convergence} ==================== We present the convergence results of our VAFL method for the nonconvex and strongly convex cases and under different update rules. Due to space limitation, this section mainly presents the convergence rates for fully asynchronous version of VAFL (Algorithm \[alg:Async-SGD-FL\]), and the convergence results for $t$-synchronous one (Algorithm \[alg:ksync-SGD-FL\]) are similar, and thus will be given in the supplementary materials. To analyze the performance of Algorithm \[alg:Async-SGD-FL\], we first make the following assumptions on sampling and smoothness. \[assump:sample\] Sample indexes $\{n_k\}$ are i.i.d. And the variance of gradient follows $ {\mathbb{E}}\left[\\|g_m^k-\nabla_{\theta_m}F(\theta_0^k,{\boldsymbol{\theta}}^k]\\|^2\right]\leq\sigma_m^2, ~ \forall m$, where $g_m^k$ is the stochastic gradient $\hat{g}_m^k$ without delay, e.g., $g_m^k:=\nabla_{\theta_m}\ell\left(\theta_0^k,h_{n_k,1}^k,\ldots, h_{n_k,M}^k; y_{n_k}\right)$. \[assump:lip\] The optimal loss is lower bounded $F^>-\infty$. The gradient $\nabla F$ is $L$-Lipschitz continuous, and $\nabla_{\theta_m}F$ is $L_m$-Lipschitz continuous. Generally, assumption \[assump:lip\] cannot be satisfied under our general vertical FL formulation with nonsmooth* local embedding functions such as neural networks. However, techniques that we call perturbed local embedding will be introduced to enforce smoothness in Section \[sec:randomneuron\]. To handle asynchrony, we need the following assumption, which is often seen in the analysis of asynchronous BCD. \[assump:client\] The uploading client $m_k$ is independent of $m_0,\ldots, m_{k-1}$ and satisfies ${\mathbb{P}}(m_k=m):= q_m$. A simple scenario satisfying this assumptions is that the activation of all clients follows independent Poisson processes. That is, if the time difference between two consecutive activations of client $m$ follows exponential distribution with parameter $\lambda_m$, then the activation of client $m$ is a Possion process with $q_m=\lambda_m^{-1}/{\textstyle\sum_{j=1}^M}\lambda_j^{-1}$. We first present the convergence results for bounded $\tau_{n_k,m}^k$. Convergence under bounded delay {#subsec.bounded} ------------------------------- We make the following assumption only for this subsection. \[assump:boundeddelay\] For each client $m$ and each sample $n$, the delay $\tau_{n,m}^k$ at iteration $k$ is bounded by a constant $D$, i.e., $\tau_{n,m}^k\leq D$. We first present the convergence for the nonconvex case. \[thm:nc1\] Under Assumptions \[assump:sample\] – \[assump:boundeddelay\], if $\eta_0^k=\eta_m^k=\min\{\frac{1}{4(1+D)L}, \frac{c_{\eta}}{\sqrt{K}}\}$ with $c_{\eta}\!>\!0$, then we have $$\small \frac{1}{K}\sum\limits_{k=0}^{K-1}{\mathbb{E}}[\\|\nabla F(\theta_0^k,{\boldsymbol{\theta}}^k)\\|^2]={\cal O}\left({1}/{\sqrt{K}}\right).$$ Under the additional assumption of strong convexity, the convergence rate is improved. \[thm:sc1\] Under Assumptions \[assump:sample\] – \[assump:boundeddelay\], and the additional assumption that $F$ is $\mu$-strongly convex in $(\theta_0,{\boldsymbol{\theta}})$, if $\eta^k=\frac{4}{\mu\min\limits_m\sqrt{q_m}(k+K_0)}$ with the constant $K_0>0$, then $$\small {\mathbb{E}}F\left(\theta_0^K,{\boldsymbol{\theta}}^K\right)-F^={\cal O}\left({1}/{K}\right).$$ Convergence under stochastic unbounded delay -------------------------------------------- We make the following assumption only* for this subsection. \[assump:unboundeddelay\] For each client $m$, the delay $\tau_{n_k,m}^k$ is an random variable with unbounded support. And there exists $\bar p_m, \rho>0$ such that ${\mathbb{P}}\left(\tau_{n_k,m}^k=d\right)\leq \bar p_m\rho^d:=p_{m,d}$. Under Assumption \[assump:unboundeddelay\], we obtain the convergence rates of the same order as those the under bounded delay assumption. \[thm:nc2\] Under Assumptions \[assump:sample\]-\[assump:client\] and \[assump:unboundeddelay\], if $\eta_0^k=\eta_m^k=\min\big\{\frac{1}{4(1+\min_m\sqrt{c_m})L}, \frac{c_{\eta}}{\sqrt{K}}\big\}$, then we have $$\small \frac{1}{K}\sum\limits_{k=0}^{K-1}{\mathbb{E}}[\\|\nabla F(\theta_0^k,{\boldsymbol{\theta}}^k)\\|^2]={\cal O}\left({1}/{\sqrt{K}}\right).$$ Under the additional assumption of strong convexity, the convergence rate is improved. \[thm:sc2\] Assume that $F$ is $\mu$-strongly convex in $(\theta_0,{\boldsymbol{\theta}})$. Then under Assumptions \[assump:sample\]-\[assump:client\] and \[assump:unboundeddelay\], if $\eta_0^k=\eta_m^k=\frac{2}{\nu(k+K_0)}$ where $K_0$ is a positive constant, then it follows that $$\small {\mathbb{E}}F(\theta_0^K,{\boldsymbol{\theta}}^K)-F^={\cal O}\left({1}/{K}\right).$$ It worth mentioning that under the assumption of bounded delay and unbounded but stochastic delay, without even coordinating clients’ gradient samples and local model updates, our algorithm achieves the same order of convergence as that of block-wise SGD in both cases [@XuYin2015_block]. Perturbed local embedding: Enforcing differential privacy and smoothness {#sec:randomneuron} ======================================================================== In this section, we introduce a local perturbation technique that is applied by each client to enforce the differential privacy of local information, which also smoothes the otherwise nonsmooth mapping of local embedding. Local perturbation ------------------ Recall that $h_m$ denotes a local embedding function of client $m$ with the parameter $\theta_m$ which embeds the information of local data $x_{n,m}$ into its outputs $h_{n, m}:=h_m(\theta_m;x_{n,m})$. When $h_m$ is linear embedding, it is as simple as $h_m(\theta_m;x_{n,m})=x_{n,m}^{\top}\theta_m$. To further account for nonlinear embedding such as neural networks, we represent $h_{n,m}$ in the following composite form \[eqn:localembed\] $$\begin{aligned} &u_0=x_{n,m}\\ &u_l=\sigma_l(w_lu_{l-1}+b_l),~~~l=1,\ldots, L\label{eqn:localem2}\\ &h_{n,m}=u_L \end{aligned}$$ where $\sigma_l$ is a linear or nonlinear function, and $w_l,b_l$ corresponds to the parameter $\theta_m$ of $h_m$, e.g., $\theta_m:=[w_1,\cdots,w_L, b_1,\cdots,b_L]^{\top}$. Here we assume that $\sigma_l$ is $L_{\sigma_l}^0$-Lipschitz continuous. Specially, when $h_m$ is linear, the composite embedding corresponds to $L=1, \sigma_1(z)=z$. We perturb the local embedding function $h_m$ by adding a random neuron with output $Z_l$ at each layer $l$ (cf. ) $$\begin{aligned} \label{eqn:pertlocalembed} &u_l=\sigma_l(w_lu_{l-1}+b_l+Z_l),~~~l=1,\ldots, L\end{aligned}$$ where $Z_1,\ldots, Z_L$ are independent random variables. With properly chosen distributions of $Z_l, l=1,\ldots, L$, we show below $h_m$ is smooth and enables differential privacy. While it does not exclude other options, our choice of the perturbation distributions is \[eq.pertb\] $$\begin{aligned} &Z_L\sim\mathcal N\left(0,c^2\right)\\ &Z_l\sim\mathcal U[-\sqrt{3}c_l,\sqrt{3}c_l],~~~l=1,\cdots, L-1\end{aligned}$$ where $\mathcal N\left(0,c^2\right)$ denotes the Gaussian distribution with zero mean and variance $c^2$, and $\mathcal U[-\sqrt{3}c_l,\sqrt{3}c_l]$ denotes the uniform distribution over $[-\sqrt{3}c_l,\sqrt{3}c_l]$. ------------------------------------------------------------------------------------------------------------------------------------------------------ -- -- -- ![image](mnist_test){width="4.8cm"} ![image](fashion_mnist_test){width="4.8cm"} ![image](cifar_test){width="4.8cm"} ![image](PD_test){width="4.8cm"} ------------------------------------------------------------------------------------------------------------------------------------------------------ -- -- -- Enforcing smoothness -------------------- The convergence results in Section \[sec:convergence\] hold under Assumption \[assump:lip\] which requires the smoothness of the overall loss function. Inspired by the randomized smoothing technique [@duchi2012randomized; @nesterov2017random], we are able to smooth the objective function by taking expectation with respect to the random neurons. Intuitively this follows the fact that the smoothness of a function can be increased by convolving with proper distributions. By adding random neuron $Z_l$, the landscape of $\sigma_l$ will be smoothed in expectation with respect to $Z_l$. And by induction, we can show the smoothness of local embedding vector $h_m$. Then so long as the loss function $\ell$ is smooth w.r.t. the local embedding vector $h_m$, the global objective $F$ is smooth by taking expectation with respect to all the random neurons. We formally establish this result in the following theorem. \[thm:smoothness\] For each embedding function $h_m$, if the activation functions follow $\sigma_l=\sigma, \,\forall l$, and the variances of the random neurons follow , and assume $\\|w_l\\|$ is bounded, then with $\mathbf{Z}:=[Z_1^{\top},\cdots, Z_L^{\top}]^{\top}$, the perturbed loss satisfies Assumption \[assump:lip\], which is given by $$\label{eq.perloss} \small F_c(\theta_0,{\boldsymbol{\theta}}):=\mathbb{E}_{\mathbf{Z}}[F(\theta_0,{\boldsymbol{\theta}};\mathbf{Z})].$$ Starting from $L_{b_L}^{h}={L_{\sigma}^0d}/{c}$, the smoothness constants of the local model $\theta_m$ denoted as $L^{F_c}_{\theta_m}$ satisfy the following recursion ($l=1,\cdots,L-1$) [$$\begin{aligned} \label{eq.smoothness} &\!\!L_{b_l}^{h}=L_{b_{l+1}}^{h}\\|w_{l+1}\\|(L_{\sigma}^0)^2+L_{\sigma}^0\\|w_L\\|\cdots L_{\sigma}^0\\|w_{l+1}\\|L_{\bar\sigma}(c_l)\nonumber\\ &\!\!L_{w_l}^{h}={\mathbb{E}}[\\|u_{l-1}\\|]L_{b_l}^{h}\nonumber\\ &\!\!L^{F_c}_{\theta_m}=L^{\ell}_{h_m}(L_{h_m}^0)^2+L_{\ell}^0\sum_{l=1}^L(L_{w_l}^{h}+L_{b_l}^{h})+L^{r}_{\theta_m}\end{aligned}$$]{} where $L^{r}_{\theta_m}$ is the smoothness constant of the regularizer w.r.t. $\theta_m$; $L_{b_l}^{h}$ and $L_{w_l}^{h}$ are the smoothness constants of the perturbed local embedding $h$ w.r.t. the bias $b_l$ and weight $w_l$; and $L_{\bar{\sigma}}(c_l):={2\sqrt{d}L_{\sigma}^0}/{c_l}$ is the smoothness constant of the neuron at $l$th layer under the uniform perturbation. Theorem \[thm:smoothness\] implies that the perturbed loss is smooth w.r.t. the local model $\theta_m$, and a large perturbation (large $c_l$ or $c$) will lead to a smaller smoothness constant. Enforcing differential privacy ------------------------------ We now connect the perturbed local embedding technique with the private information exchange in Algorithms \[alg:Async-SGD-FL\]-\[alg:ksync-SGD-FL\]. As local clients keep sending out embedded information, it is essential to prevent any attacker to trace back to a specific individual via this observation. Targeting a better trade-off between the privacy and the accuracy, we leverage the Gaussian differential privacy (GDP) developed in [@dong2019gaussian]. A mechanism $\mathcal M$ is said to satisfy $\mu$-GDP if for all neighboring datasets $S$ and $S'$, we have $$\label{def.dp} T(\mathcal M(S), \mathcal M(S'))\geq T(\mathcal N(0,1),\mathcal N(\mu, 1))$$ where the trade-off function $ T(P,Q)(\alpha)=\inf\{\beta_{\phi}:\alpha_{\phi}\leq\alpha\}$, and $\alpha_{\phi}, \beta_{\phi}$ are type I and II errors given a threshold $\phi$. Intuitively, $\mu$-GDP guarantees that distinguishing two adjacent datasets via information revealed by $\mathcal M$ is at least as difficult as distinguishing the two distributions $\mathcal N(0,1)$ and $\mathcal N(\mu, 1)$. Smaller $\mu$ means less privacy loss. To characterize the level of privacy of our local embedding approaches, we build on the moments accountant technique originally developed in [@abadi2016] to establish that adding random neurons endows Algorithm \[alg:Async-SGD-FL\] with GDP. \[thm:privacy\] Under the same set of assumptions as those in Theorem \[thm:smoothness\], for client $m$, if we set the variance of the Gaussian random neuron at the $L$-th layer as $$\label{eqn:privacy} \small c={\cal O}\left({N_m\sqrt{K}}/(\mu N)\right)$$ where $N_m$ is the size of minibatch used at client $m$, $N$ is the size of the whole batch, $K$ is the number of queries (i.e. the number of data samples processed by $h_m$ at client $m$), then VAFL satisfies ${\mu}$-GDP for the dataset of client $m$. Theorem \[thm:privacy\] demonstrates the trade-off between accuracy and privacy. To increase privacy, i.e., decrease $\mu$ in , the variance of random neurons needs to be increased (cf. ). However, as the variance of random neurons increases, the variance of the stochastic gradient also increases, which will in turn lead to slower convergence. ![Testing accuracy of VAFL with nonlinear local embedding on ModelNet40* dataset.[]{data-label="fig:Nonlinear_acc"}](model40_test){width=".38\textwidth"} ![AUC curve of VAFL with local LSTM embedding on MIMIC-III clinical care dataset.[]{data-label="fig:mortality_auc"}](mimic_test){width=".38\textwidth"} Numerical tests and remarks =========================== We benchmark the fully asynchronous version of VAFL ([async]{}) in Algorithm \[alg:Async-SGD-FL\], and $t$-synchronous version of VAFL ([t-sync]{}) in Algorithm \[alg:ksync-SGD-FL\] with the synchronous block-wise SGD ([sync]{}), which requires synchronization and sample index coordination among clients in each iteration. We also include private versions of these algorithms via perturbed local embedding technique in Section \[sec:randomneuron\]. VAFL for federated logistic regression. We first conduct logistic regression on MNIST, Fashion-MNIST, CIFAR10 and Parkinson disease [@sakar2019comparative] datasets. The $l_2$-regularizer coefficient is set as $0.001$. We select $M=7$ for MNIST and MNIST, $M=8$ for CIFAR10 and $M=3$ for PD dataset. The testing accuracy versus wall-clock time is reported in Figures \[fig:test1\]. The dashed horizontal lines represent the results trained on the centralized (non-federated) model, and the dashed curves represent private variants of considered algorithms with variance $c=0.1$. In all cases, VAFL learns a federated model with accuracies comparable to that of the centralized model that requires collecting raw data. VAFL for federated deep learning. We first train a neural network modified from MVCNN with 12-view data [@wu20153d]. We use $M=4$ clients, and each client has 3 views of each object and use a 7-layer CNN as local embedding functions, and server uses a fully connected network to aggregate the local embedding vectors. Results are plotted in Figure \[fig:Nonlinear\_acc\]. We also test our VAFL algorithm in MIMIC-III — an open dataset comprising deidentified health data [@johnson2016mimic]. We perform the in-hospital mortality prediction as in [@Harutyunyan2019] among $M=4$ clients. Each client uses LSTM as the embedding function. In Figure \[fig:mortality\_auc\], we can still observe that async and $t$-sync VAFL learn a federated model with accuracies comparable to that of the centralized model, and requires less time relative to the synchronous FL algorithm. [43]{} \[1\][\#1]{} \[1\][`#1`]{} urlstyle \[1\][doi: \#1]{} Abadi, M., Chu, A., Goodfellow, I., McMahan, H. B., Mironov, I., Talwar, K., and Zhang, L. Deep learning with differential privacy. In Proc. ACM SIGSAC Conference on Computer and Communications Security, pp. 308–318, Vienna, Austria, October 2016. Aji, A. F. and Heafield, K. Sparse communication for distributed gradient descent. In Proc. Conf. Empirical Methods Natural Language Process., pp. 440–445, Copenhagen, Denmark, Sep 2017. Alistarh, D., Grubic, D., Li, J., Tomioka, R., and Vojnovic, M. via gradient quantization and encoding. In Proc. Advances in Neural Info. Process. Syst., pp. 1709–1720, Long Beach, CA, Dec 2017. Bonawitz, K., Ivanov, V., Kreuter, B., Marcedone, A., McMahan, H. B., Patel, S., Ramage, D., Segal, A., and Seth, K. Practical secure aggregation for privacy-preserving machine learning. In Proc. ACM Conf. on Comp. and Comm. Security, pp. 1175–1191, Dallas, TX, October 2017. Cannelli, L., Facchinei, F., Kungurtsev, V., and Scutari, G. Asynchronous parallel algorithms for nonconvex big-data optimization: Model and convergence. arXiv preprint:1607.04818, July 2016. Chen, T., Giannakis, G., Sun, T., and Yin, W. azily aggregated gradient for communication-efficient distributed learning. In Proc. Advances in Neural Info. Process. Syst., pp. 5050–5060, Montreal, Canada, Dec 2018. Chen, T., Sun, Y., and Yin, W. azily aggregated stochastic gradients for communication-efficient distributed learning. arXiv preprint:2002.11360, February 2020. Dong, J., Roth, A., and Su, W. J. Gaussian differential privacy. arXiv preprint:1905.02383, May 2019. Duchi, J. C., Bartlett, P. L., and Wainwright, M. J. Randomized smoothing for stochastic optimization. SIAM Journal on Optimization, 220 (2):0 674–701, 2012. Dutta, S., Joshi, G., Ghosh, S., Dube, P., and Nagpurkar, P. Slow and stale gradients can win the race: [Error-runtime trade-offs in distributed SGD]{}. In Proc. Intl. Conf. on Artif. Intell. and Stat., Lanzarote, Spain, 2018. Dwork, C., Roth, A., et al. The algorithmic foundations of differential privacy. Foundations and Trends in Theoretical Computer Science, 90 (3–4):0 211–407, 2014. Hamm, J., Cao, Y., and Belkin, M. Learning privately from multiparty data. In Proc. Intl. Conf. Machine Learn., pp. 555–563, New York, NY, June 2016. Hardy, S., Henecka, W., Ivey-Law, H., Nock, R., Patrini, G., Smith, G., and Thorne, B. Private federated learning on vertically partitioned data via entity resolution and additively homomorphic encryption. arXiv preprint:1711.10677, November 2017. Harutyunyan, H., Khachatrian, H., Kale, D. C., Ver Steeg, G., and Galstyan, A. Multitask learning and benchmarking with clinical time series data. Scientific Data, 60 (1):0 96, 2019. ISSN 2052-4463. [doi: ]{}[10.1038/s41597-019-0103-9]{}. URL <https://doi.org/10.1038/s41597-019-0103-9>. Hu, Y., Niu, D., Yang, J., and Zhou, S. Fdml: A collaborative machine learning framework for distributed features. In Proc. of ACM SIGKDD Intl. Conf. Knowledge Discovery & Data Mining, pp. 2232–2240, 2019. Johnson, A. E., Pollard, T. J., Shen, L., Li-wei, H. L., Feng, M., Ghassemi, M., Moody, B., Szolovits, P., Celi, L. A., and Mark, R. G. . Scientific data, 30 (160035), 2016. Kairouz, P., McMahan, H. B., Avent, B., Bellet, A., Bennis, M., Bhagoji, A. N., Bonawitz, K., Charles, Z., Cormode, G., Cummings, R., et al. Advances and open problems in federated learning. arXiv preprint:1912.04977, December 2019. Kairouz, P., McMahan, H. B., Avent, B., Bellet, A., Bennis, M., Bhagoji, A. N., Bonawitz, K., Charles, Z., Cormode, G., Cummings, R., et al. Advances and open problems in federated learning. arXiv preprint:1912.04977, December 2019. Kone[č]{}n[y]{}, J., McMahan, H. B., Ramage, D., and Richt[á]{}rik, P. Federated optimization: Distributed machine learning for on-device intelligence. arXiv preprint:1610.02527, October 2016. Kone[č]{}n[y]{}, J., McMahan, H. B., Yu, F. X., Richt[á]{}rik, P., Suresh, A. T., and Bacon, D. Federated learning: Strategies for improving communication efficiency. arXiv preprint:1610.05492, Oct 2016. Lian, X., Zhang, C., Zhang, H., Hsieh, C.-J., Zhang, W., and Liu, J. case study for decentralized parallel stochastic gradient descent. In Proc. Advances in Neural Info. Process. Syst., pp. 5330–5340, Long Beach, CA, Dec 2017. Liang, P. P., Liu, T., Ziyin, L., Salakhutdinov, R., and Morency, L.-P. Think locally, act globally: Federated learning with local and global representations. arXiv preprint:2001.01523, January 2020. Liu, Y., Kang, Y., Zhang, X., Li, L., Cheng, Y., Chen, T., Hong, M., and Yang, Q. A communication efficient vertical federated learning framework. arXiv preprint arXiv:1912.11187, 2019. McMahan, B., Moore, E., Ramage, D., Hampson, S., and y Arcas, B. A. Communication-efficient learning of deep networks from decentralized data. In Proc. Intl. Conf. Artificial Intell. and Stat., pp. 1273–1282, Fort Lauderdale, FL, April 2017. McMahan, B., Moore, E., Ramage, D., Hampson, S., and y Arcas, B. A. Communication-efficient learning of deep networks from decentralized data. In Proc. Intl. Conf. on Artif. Intell. and Stat., pp. 1273–1282, Fort Lauderdale, Florida, Apr 2017. Mohri, M., Sivek, G., and Suresh, A. T. Agnostic federated learning. In Proc. Intl. Conf. Machine Learn., pp. 4615–4625, Long Beach, CA, June 2019. Nesterov, Y. and Spokoiny, V. Random gradient-free minimization of convex functions. Foundations of Computational Mathematics, 170 (2):0 527–566, 2017. Niu, C., Wu, F., Tang, S., Hua, L., Jia, R., Lv, C., Wu, Z., and Chen, G. Secure federated submodel learning. arXiv preprint:1911.02254, November 2019. Peng, Z., Xu, Y., Yan, M., and Yin, W. Arock: an algorithmic framework for asynchronous parallel coordinate updates. SIAM J. Sci. Comp., 380 (5):0 2851–2879, September 2016. Razaviyayn, M., Hong, M., and Luo, Z.-Q. A unified convergence analysis of block successive minimization methods for nonsmooth optimization. SIAM Journal on Optimization, 230 (2):0 1126–1153, June 2013. Recht, B., Re, C., Wright, S., and Niu, F. Hogwild: A lock-free approach to parallelizing stochastic gradient descent. In Proc. Advances in Neural Info. Process. Syst., pp. 693–701, Granada, Spain, December 2011. Sakar, C. O., Serbes, G., Gunduz, A., Tunc, H. C., Nizam, H., Sakar, B. E., Tutuncu, M., Aydin, T., Isenkul, M. E., and Apaydin, H. A comparative analysis of speech signal processing algorithms for parkinson’s disease classification and the use of the tunable q-factor wavelet transform. Applied Soft Computing, 74:0 255–263, 2019. Seide, F., Fu, H., Droppo, J., Li, G., and Yu, D. 1-bit stochastic gradient descent and its application to data-parallel distributed training of speech dnns. In Proc. Conf. Intl. Speech Comm. Assoc., Singapore, Sept 2014. Smith, V., Chiang, C.-K., Sanjabi, M., and Talwalkar, A. S. Federated multi-task learning. In Proc. Advances in Neural Info. Process. Syst., pp. 4427–4437, Long Beach, CA, December 2017. Strom, N. Scalable distributed [DNN]{} training using commodity gpu cloud computing. In Proc. Conf. Intl. Speech Comm. Assoc., Dresden, Germany, Sept 2015. Sun, C., Ippel, L., van Soest, J., Wouters, B., Malic, A., Adekunle, O., van den Berg, B., Mussmann, O., Koster, A., van der Kallen, C., et al. A privacy-preserving infrastructure for analyzing personal health data in a vertically partitioned scenario. Studies in health technology and informatics, 264:0 373–377, 2019. Sun, T., Hannah, R., and Yin, W. Asynchronous coordinate descent under more realistic assumptions. In Proc. Advances in Neural Info. Process. Syst., pp. 6183–6191, Long Beach, CA, December 2017. Wang, J. and Joshi, G. Cooperative [SGD: A]{} unified framework for the design and analysis of communication-efficient [SGD]{} algorithms. arXiv preprint:1808.07576, August 2018. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X., and Xiao, J. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 1912–1920, 2015. Xu, Y. and Yin, W. A block coordinate descent method for regularized multiconvex optimization with applications to nonnegative tensor factorization and completion. SIAM Journal on Imaging Sciences, 60 (3):0 1758–1789, 2013. Xu, Y. and Yin, W. Block stochastic gradient iteration for convex and nonconvex optimization. SIAM Journal on Optimization, 250 (3):0 1686–1716, 2015. Yang, Q., Liu, Y., Chen, T., and Tong, Y. Federated machine learning: Concept and applications. ACM Trans. Intelligent Systems and Technology, 100 (2), January 2019. Yao, A. C. Protocols for secure computations. In Annual Symposium on Foundations of Computer Science, pp. 160–164, Chicago, Illinois, 1982. [Supplementary materials for\ “VAFL: a Method of Vertical Asynchronous Federated Learning"]{} In this supplementary document, we first present some supporting lemmas that will be used frequently in this document, and then present the proofs of all the lemmas and theorems in the paper, which is followed by details on our experiments. The content of this supplementary document is summarized as follows. Supporting Lemmas {#sec.sup_lem} ================= For notational brevity, we define $$\begin{aligned} &G_{0}^k:=\nabla_{\theta_0}F(\theta_0^k, {\boldsymbol{\theta}}^k)\\ & G_m^k:=\nabla_{\theta_m}F(\theta_0^k, {\boldsymbol{\theta}}^k)\\ &g_0^k:=\nabla_{\theta_0}\ell\left(\theta_0^k,h_{n_k,1}^k,\ldots, h_{n_k,M}^k; y_{n_k}\right)\\ & g_m^k:=\nabla_{\theta_m}\ell\left(\theta_0^k,h_{n_k,1}^k,\ldots, h_{n_k,M}^k; y_{n_k}\right)\\ &\hat g_0^k:=\nabla_{\theta_0}\ell(\theta_0,h_{n_k,1}^{k-\tau_{n_k,1}^k},\ldots, h_{n_k,M}^{k-\tau_{n_k,M}^k};y_{n_k})\\ &\hat g_m^k:=\left\{\begin{array}{ll} \nabla h_m(\theta_m^k;x_{n_k,m})\nabla_{h_m}\ell(\theta_0^k,h_{n_k,1}^{k-\tau_{n_k,1}^k}\ldots, h_{n_k,M}^{k-\tau_{n_k,M}^k};y_{n_k}) & \text{if }m=m_k\\ 0 & \text{else}. \end{array}\right.\\ &\hat{\boldsymbol{\theta}}:=[\theta_1^{k-\tau_{n_k,1}^k};\ldots;\theta_M^{k-\tau_{n_k,M}^k}]\label{eqn:defg}.\end{aligned}$$ To handle the delayed information, we leverage the following Lyapunov function for analyzing VAFL $$\label{eqn:lyapunov1} V^k:=F(\theta_0^k, {\boldsymbol{\theta}}^k)+\sum\limits_{d=1}^D\gamma_d\\|{\boldsymbol{\theta}}^{k-d+1}-{\boldsymbol{\theta}}^{k-d}\\|^2$$ where $\{\gamma_d\}$ are a set of constants to be determined later. \[lemma:l1descent\] Under Assumptions \[assump:sample\] – \[assump:boundeddelay\], for $\eta_0^k\leq\frac{1}{4L},\eta_m^k\leq\frac{1}{4(L+2\gamma_1)}$, it follows that (with $\gamma_{D+1}=0$) $$\begin{aligned} \label{eqn:l1descent1} \small &{\mathbb{E}}V^{k+1}-{\mathbb{E}}V^k\leq -\left(\frac{\eta_0^k}{2}-L(\eta_0^k)^2\right){\mathbb{E}}[\\|\nabla_{\theta_0}F(\theta_0^k,{\boldsymbol{\theta}}^k)\\|^2]\nonumber\\ &\!-\!\!\sum\limits_{m=1}^M\!q_m\!\left(\!\frac{\eta_{m}^k}{2}-L(\eta_m^k)^2-2\gamma_1(\eta_m^k)^2\!\right)\!{\mathbb{E}}[\\|\nabla_{\theta_m}F(\theta_0^k,{\boldsymbol{\theta}}^k)\\|^2]\nonumber\\ &\!+\sum\limits_{d=1}^D\left(\!Dc^k\!+\!D\gamma_1\max\limits_m2(\eta_m^k)^2L_m^2\!+\!\gamma_{d+1}\!-\!\gamma_d\right)\nonumber\\ &\!\times{\mathbb{E}}[\\|{\boldsymbol{\theta}}^{k+1-d}\!-\!{\boldsymbol{\theta}}^{k-d}\\|^2]\nonumber\\ &\!+\!L(\eta_0^k)^2\sigma_0^2+\sum\limits_{m=1}^Mq_m(L+2\gamma_1)(\eta_m^k)^2\sigma_m^2.\end{aligned}$$ If $\{\gamma_d\}$ are chosen properly as specified in the supplementary materials, the first three terms in the right hand side of is negative. By carefully choosing $\{\eta_0^k,\eta_m^k\}$, we can ensure the convergence of Algorithm \[alg:Async-SGD-FL\]. We first quantify the descent amount in the objective value. \[lemma:Fdescent\] Under Assumptions \[assump:sample\]-\[assump:client\], the iterates $\{\theta_0^k, {\boldsymbol{\theta}}^k\}$ generated by Algorithm \[alg:Async-SGD-FL\] satisfy $$\begin{aligned} \label{eqn:Fdescent} {\mathbb{E}}[F(\theta_0^{k+1}, {\boldsymbol{\theta}}^{k+1})\|\Theta^k] & \leq F(\theta_0^k, {\boldsymbol{\theta}}^k)+c_k{\mathbb{E}}[\\|\hat {\boldsymbol{\theta}}^k-{\boldsymbol{\theta}}^k\\|^2\|\Theta^k]+L(\eta_0^k)^2\sigma_0^2\nonumber\\ &+\sum\limits_{m=1}^Mq_mL(\eta_m^k)^2\sigma_m^2 -(\frac{\eta_0^k}{2}-L(\eta_0^k)^2)\\|G_0^k\\|^2-\sum\limits_{m=1}^Mq_m(\frac{\eta_{m}^k}{2}-L(\eta_m^k)^2)\\|G_m^k\\|^2\end{aligned}$$ where $\Theta^k$ is the $\sigma$-algebra generated by $\{\theta_0^0,{\boldsymbol{\theta}}^0, \ldots, \theta_0^k,{\boldsymbol{\theta}}^k\}$, and $c_k$ is defined as $$\label{eqn:const-c} c_k:=\left(\frac{\eta_0^k}{2}+L(\eta_0^k)^2\right)L_0^2+\max\limits_m\left(\frac{\eta_{m}^k}{2}+L(\eta_m^k)^2\right)L_m^2.$$ By Assumption \[assump:lip\], we have $$\begin{aligned} \label{sec3.eq1} &F(\theta_0^{k+1}, {\boldsymbol{\theta}}^{k+1})\nonumber\\ =&F(\theta_0^k-\eta_0^k\hat g_0^k, \ldots, \theta_{m_k}^k-\hat g_{m_k}^k, \ldots)\nonumber\\ \leq &F(\theta_0^k, {\boldsymbol{\theta}}^k)-\eta_0^k\dotp{G_0^k, \hat g_0^k)}-\eta_{m_k}^k\dotp{G_{m_k}^k, \hat g_{m_k}^k}+\frac{L(\eta_0^k)^2}{2}\\|\hat g_{0}^k\\|^2+\frac{L(\eta_{m_k}^k)^2}{2}\\|\hat g_{m_k}^k\\|^2\nonumber\\ \leq & \! F(\theta_0^k, {\boldsymbol{\theta}}^k)\!-\eta_0^k\dotp{G_0^k, g_0^k}\!-\eta_0^k\dotp{G_0^k, \hat g_{0}^k-g_{0}^k}\!-\eta_{m_k}^k\dotp{G_{m_k}^k, g_{m_k}^k}\!-\eta_{m_k}^k\dotp{G_{m_k}^k,\hat g_{m_k}^k-g_{m_k}^k}\!+\frac{L(\eta_0^k)^2}{2}\\|\hat g_{0}^k\\|^2\!+\!\frac{L(\eta_{m_k}^k)^2}{2}\\|\hat g_{m_k}^k\\|^2\nonumber\\ \leq& F(\theta_0^k, {\boldsymbol{\theta}}^k)-\eta_0^k\dotp{G_0^k,g_0^k}-\eta_{m_k}^k\dotp{G_{m_k}^k,g_{m_k}^k}+\frac{\eta_0^k}{2}\\|G_0^k\\|^2+\frac{\eta_{m_k}^k}{2}\\|G_{m_k}^k\\|^2+L(\eta_0^k)^2\\|g_0^k\\|^2 + L(\eta_{m_k}^k)^2\\|g_{m_k}^k\\|^2\nonumber\\ & +(\frac{\eta_0^k}{2}+L(\eta_0^k)^2)\\|\hat g_{0}^k-g_{0}^k\\|^2+(\frac{\eta_{m_k}^k}{2}+L(\eta_{m_k}^k)^2)\\|\hat g_{m_k}^k-g_{m_k}^k\\|^2.\end{aligned}$$ Note that we have $$\begin{aligned} \label{sec3.eq2} {\mathbb{E}}\left[\\|g_{m_k}^k\\|^2\|\Theta^k\right]&={\mathbb{E}}\left[\\|g_{m_k}^k-G_{m_k}^k+G_{m_k}^k\\|^2\|\Theta^k\right]\nonumber\\ &={\mathbb{E}}\left[\\|g_{m_k}^k-G_{m_k}^k\\|^2\|\Theta^k\right]+2{\mathbb{E}}\left[\dotp{g_{m_k}^k-G_{m_k}^k, G_{m_k}^k}\|\Theta^k\right]+\\|G_{m_k}^k\\|^2\nonumber\\ &=\sigma_{m_k}^2+\\|G_{m_k}^k\\|^2\end{aligned}$$ where the last equality follows from Assumption \[assump:sample\]. First we take expectation on with respect to $n_k$, conditioned on $\Theta^k$ and $m_k=m$, we have $$\begin{aligned} \label{sec3.eq3} {\mathbb{E}}[F(\theta_0^{k+1}, {\boldsymbol{\theta}}^{k+1})\|m_k=m, \Theta^k] \leq& F(\theta_0^k, {\boldsymbol{\theta}}^k)-(\frac{\eta_0^k}{2}-L(\eta_0^k)^2)\\|G_0^k\\|^2-(\frac{\eta_{m}^k}{2}-L(\eta_m^k)^2)\\|G_m^k\\|^2+L(\eta_0^k)^2\sigma_0^2+L(\eta_m^k)^2\sigma_m^2\nonumber\\ & +(\frac{\eta_0^k}{2}+L(\eta_0^k)^2){\mathbb{E}}[\\|\hat g_{0}^k-g_{0}^k\\|^2\|m_k=m]+(\frac{\eta_{m}^k}{2}+L(\eta_m^k)^2){\mathbb{E}}[\\|\hat g_{m}^k-g_{m}^k\\|^2\|m_k=m].\end{aligned}$$ Then taking expectation with respect to $m_k$, it follows that $$\begin{aligned} {\mathbb{E}}[F(\theta_0^{k+1}, {\boldsymbol{\theta}}^{k+1})\|\Theta^k] \leq &F(\theta_0^k, {\boldsymbol{\theta}}^k)-(\frac{\eta_0^k}{2}-L(\eta_0^k)^2)\\|G_0^k\\|^2-\sum\limits_{m=1}^Mq_m(\frac{\eta_{m}^k}{2}-L(\eta_m^k)^2)\\|G_m^k\\|^2+L(\eta_0^k)^2\sigma_0^2+\sum\limits_{m=1}^Mq_mL(\eta_m^k)^2\sigma_m^2\\ &+(\frac{\eta_0^k}{2}+L(\eta_0^k)^2)L_0^2{\mathbb{E}}[\\|\hat {\boldsymbol{\theta}}^k-{\boldsymbol{\theta}}^k\\|^2\|\Theta^k]+\sum\limits_{m=1}^Mq_m(\frac{\eta_{m}^k}{2}+L(\eta_m^k)^2)L_m^2{\mathbb{E}}[\\|\hat {\boldsymbol{\theta}}^k-{\boldsymbol{\theta}}^k\\|^2\|m_k=m, \Theta^k]\\ \leq &F(\theta_0^k, {\boldsymbol{\theta}}^k)-(\frac{\eta_0^k}{2}-L(\eta_0^k)^2)\\|G_0^k\\|^2-\sum\limits_{m=1}^Mq_m(\frac{\eta_{m}^k}{2}-L(\eta_m^k)^2)\\|G_m^k\\|^2+L(\eta_0^k)^2\sigma_0^2+\sum\limits_{m=1}^Mq_mL(\eta_m^k)^2\sigma_m^2\\ &+\underbrace{\left(\left(\frac{\eta_0^k}{2}+L(\eta_0^k)^2\right)L_0^2+\max\limits_m\left(\frac{\eta_{m}^k}{2}+L(\eta_m^k)^2\right)L_m^2\right)}_{:=c^k}{\mathbb{E}}[\\|\hat {\boldsymbol{\theta}}^k-{\boldsymbol{\theta}}^k\\|^2\|\Theta^k]\end{aligned}$$ which completes the proof. Convergence under bounded delay {#convergence-under-bounded-delay} =============================== Recalling the definition of $\hat{\boldsymbol{\theta}}^k$ in , if $\tau_{n_k,m}^k\leq D$, then it can be derived that $$\\|\hat{\boldsymbol{\theta}}^k-{\boldsymbol{\theta}}^k\\|^2\leq\sum\limits_{d=1}^DD\left\\|{\boldsymbol{\theta}}^{k+1-d}-{\boldsymbol{\theta}}^{k-d}\right\\|^2.$$ Proof of Lemma \[lemma:l1descent\] ---------------------------------- Recall the definition of $V^k$, that is $$\begin{aligned} V^k=F(\theta_0^k, {\boldsymbol{\theta}}^k)+\sum\limits_{d=1}^D\gamma_d\\|{\boldsymbol{\theta}}^{k+1-d}-{\boldsymbol{\theta}}^{k-d}\\|^2\end{aligned}$$ where we initialize the algorithm with ${\boldsymbol{\theta}}^{-D+1}=\cdots={\boldsymbol{\theta}}^{-1}={\boldsymbol{\theta}}^0$. We first decompose $\\|{\boldsymbol{\theta}}^{k+1}-{\boldsymbol{\theta}}^k\\|^2$ as $$\begin{aligned} \label{eqn:decompose1} \\|{\boldsymbol{\theta}}^{k+1}-{\boldsymbol{\theta}}^k\\|^2=(\eta_{m_k}^k)^2\\|\hat g_{m_k}^k\\|^2\leq2(\eta_{m_k}^k)^2\\|g_{m_k}^k\\|^2+2(\eta_{m_k}^k)^2\\|\hat g_{m_k}^k-g_{m_k}^k\\|^2.\end{aligned}$$ Taking expectation on both sides of , and applying leads to $$\begin{aligned} \label{eqn:decompose2} {\mathbb{E}}\left[\\|{\boldsymbol{\theta}}^{k+1}-{\boldsymbol{\theta}}^k\\|^2\|\Theta^k\right]&=\sum\limits_{m=1}^M{\mathbb{E}}\left[\\|{\boldsymbol{\theta}}^{k+1}-{\boldsymbol{\theta}}^k\\|^2\|m_k=m,\Theta^k\right]{\mathbb{P}}(m_k=m)\nonumber\\ &\leq \sum\limits_{m=1}^M2q_m(\eta_m^k)^2\\|G_m^k\\|^2+\sum\limits_{m=1}^M2q_m(\eta_m^k)^2\sigma_m^2+2\max\limits_m(\eta_m^k)^2L_m^2{\mathbb{E}}[\\|\hat{\boldsymbol{\theta}}^k-{\boldsymbol{\theta}}^k\\|^2\|\Theta^k].\end{aligned}$$ Following Lemma \[lemma:Fdescent\] in Appendix \[sec.sup\_lem\] and , the Lyapunov function $V^k$ satisfies $$\begin{aligned} \label{eqn:l1descent2} {\mathbb{E}}[V^{k+1}\|\Theta^k]-V^k=&{\mathbb{E}}[F(\theta_0^{k+1}, {\boldsymbol{\theta}}^{k+1})\|\Theta^k]-F(\theta_0^k, {\boldsymbol{\theta}}^k)+\gamma_1{\mathbb{E}}[\\|{\boldsymbol{\theta}}^{k+1}-{\boldsymbol{\theta}}^k\\|^2\|\Theta^k]\nonumber\\ &+\sum\limits_{d=1}^{D-1}(\gamma_{d+1}-\gamma_d)\\|{\boldsymbol{\theta}}^{k+1-d}-{\boldsymbol{\theta}}^{k-d}\\|^2-\gamma_D\\|{\boldsymbol{\theta}}^{k+1-D}-{\boldsymbol{\theta}}^{k-D}\\|^2\nonumber\\ \leq &-\left(\frac{\eta_0^k}{2}-L(\eta_0^k)^2\right)\\|G_0^k\\|^2-\sum\limits_{m=1}^Mq_m\left(\frac{\eta_{m}^k}{2}-L(\eta_m^k)^2-2\gamma_1(\eta_m^k)^2\right)\\|G_m^k\\|^2\nonumber\\ & +L(\eta_0^k)^2\sigma_0^2+\sum\limits_{m=1}^Mq_m(L+2\gamma_1)(\eta_m^k)^2\sigma_m^2\nonumber\\ & +\sum\limits_{d=1}^{D-1}\left(Dc^k+D\gamma_1\max\limits_m2(\eta_m^k)^2L_m^2+\gamma_{d+1}-\gamma_d\right)\\|{\boldsymbol{\theta}}^{k+1-d}-{\boldsymbol{\theta}}^{k-d}\\|^2\nonumber\\ &+\left(Dc^k+D\gamma_1\max\limits_m2(\eta_m^k)^2L_m^2-\gamma_D\right)\\|{\boldsymbol{\theta}}^{k+1-D}-{\boldsymbol{\theta}}^{k-D}\\|^2. $$ Since we choose $\eta_0^k,\eta_{m}^k\leq\bar\eta\leq\frac{1}{4(L+2\gamma_1)}$, it follows that $c^k\leq\frac{3}{2}\bar\eta L^2$. By taking expectation on both sides of , we have $$\begin{aligned} \label{eqn:l1descent} {\mathbb{E}}V^{k+1}-{\mathbb{E}}V^k\leq& -\frac{1}{4}\min\{\eta_0^k,q_m\eta_m^k\}{\mathbb{E}}[\\|\nabla F(\theta_0^k,{\boldsymbol{\theta}}^k)\\|^2]+L(\eta_0^k)^2\sigma_0^2+(2\gamma_1+L)\sum\limits_{m=1}^Mq_m(\eta_m^k)^2\sigma_m^2\nonumber\\ &-\sum\limits_{d=1}^{D-1}\left(\gamma_d-\gamma_{d+1}-\frac{3}{2}D\bar\eta L^2-2D\gamma_1\bar\eta^2L^2\right){\mathbb{E}}\\|{\boldsymbol{\theta}}^{k+1-d}-{\boldsymbol{\theta}}^{k-d}\\|^2\nonumber\\ &-\left(\gamma_D-\frac{3}{2}D\bar\eta L^2-2D\gamma_1\bar\eta^2L^2\right){\mathbb{E}}\\|{\boldsymbol{\theta}}^{k+1-D}-{\boldsymbol{\theta}}^{k-D}\\|^2.\end{aligned}$$ Proof of Theorem \[thm:nc1\] ---------------------------- Define $\gamma_1=\frac{\frac{3}{2}\bar\eta D^2L^2}{1-2D^2\bar\eta^2L^2}\leq\frac{1}{2}DL$, and $\eta_0^k=\eta_m^k=\eta=\min\{\frac{1}{4(D+1)L}, \frac{c_{\eta}}{\sqrt{K}}\}$. Select $\gamma_2,\ldots,\gamma_D$ as follows $$\begin{aligned} \gamma_{d+1}=\gamma_d-\frac{3}{2}D\eta L^2-2D\gamma_1\eta^2L^2,~~~ d=1,\ldots, D-1.\end{aligned}$$ It can be verified that $\gamma_D-\frac{3}{2}D\eta L^2-2D\gamma_1\eta^2L^2\geq 0$. Then $\eqref{eqn:l1descent}$ reduces to $$\begin{aligned} \label{eqn:l1descent3} {\mathbb{E}}V^{k+1}-{\mathbb{E}}V^k\leq& -\frac{1}{4}\min_m q_m\eta{\mathbb{E}}[\\|\nabla F(\theta_0^k,{\boldsymbol{\theta}}^k)\\|^2]+\eta^2L\sigma_0^2+\eta^2(2\gamma_1+L)\sum\limits_{m=1}^Mq_m\sigma_m^2.\end{aligned}$$ By summing over $k=0,\cdots,K-1$ and using $\eta\leq \frac{c_{\eta}}{\sqrt{K}}$, it follows that $$\begin{aligned} \frac{1}{K}\sum\limits_{k=0}^{K-1}{\mathbb{E}}[\\|\nabla F(\theta_0^k,{\boldsymbol{\theta}}^k)\\|^2]&\leq \frac{F^0-F^+K\eta^2L\sigma_0^2+K\eta^2(2\gamma_1+L)\sum\limits_{m=1}^Mq_m\sigma_m^2}{\frac{1}{4}\min\limits_mq_m\eta K}\\ &\leq\frac{16DL(F^0-F^)}{\min\limits_mq_mK}+\frac{4c_{\eta}(F^0-F^)}{\min\limits_mq_m\sqrt{K}}+\frac{4c_{\eta}L\sigma_0^2+c_{\eta}(8\gamma_1+4L)\sum\limits_{m=1}^Mq_m\sigma_m^2}{\min\limits_mq_m\sqrt{K}}.\end{aligned}$$ Proof of Theorem \[thm:sc1\] ---------------------------- By the $\mu$-strong convexity of $F(\theta_0,{\boldsymbol{\theta}})$, we have $$\begin{aligned} \label{eqn:strongcvx} 2\mu(F(\theta_0,{\boldsymbol{\theta}})-F^)\leq\\|\nabla F(\theta_0,{\boldsymbol{\theta}})\\|^2.\end{aligned}$$ Choose $\gamma_d$ such that $$\begin{aligned} &\gamma_d-\gamma_{d+1}-\frac{3}{2}D\bar\eta L^2-2D\gamma_1\bar\eta^2L^2=\frac{\mu}{2}\min\limits_mq_m\bar\eta\gamma_1, ~~~d=1,\ldots, D-1\\ &\gamma_D-\frac{3}{2}D\bar\eta L^2-2D\gamma_1\bar\eta L^2=\frac{\mu}{2}\min\limits_m q_m\bar\eta\gamma_1.\end{aligned}$$ Solve the above linear equations above and get $$\begin{aligned} \gamma_1=\frac{\frac{3}{2}\bar\eta D^2L^2}{1-2D^2\bar\eta^2L^2-\frac{\mu}{2}\min\limits_mq_mD\bar\eta},~~~ \gamma_d=(D+1-d)(\frac{3}{2}D\bar\eta L^2+2D\gamma_1\bar\eta L^2+\frac{\mu}{2}\min\limits_mq_m\bar\eta\gamma_1), ~~~d=1,\ldots, D-1.\end{aligned}$$ If we choose $\eta_0^k=\eta_m^k=\eta^k\leq\bar\eta\leq\frac{1}{4(D+1)L+2\mu\min\limits_mq_mD}$, $\gamma_1\leq 2\bar\eta D^2L^2$, then $\eqref{eqn:l1descent}$ reduces to $$\begin{aligned} {\mathbb{E}}V^{k+1}\leq(1-\frac{\mu}{2}\eta^k\min\limits_mq_m){\mathbb{E}}V^k+(\eta^k)^2\left(L\sigma_0^2+(2\gamma_1+L)\sum\limits_{m=1}^Mq_m\sigma_m^2\right).\end{aligned}$$ Defining $R:=\big(L\sigma_0^2+(2\gamma_1+L)\sum\limits_{m=1}^Mq_m\sigma_m^2\big)$ and $\eta^k=\frac{4}{\mu\min\limits_mq_m(k+K_0)}$, where $K_0=\frac{4(4(D+1)L+2\mu\min\limits_mq_mD)}{\mu\min\limits_mq_m}$, we have $$\begin{aligned} {\mathbb{E}}V^k&\leq V^0\prod\limits_{k=0}^{K-1}(1-\frac{\mu}{2}\min\limits_mq_m\eta^k)+R\sum\limits_{k=0}^{K-1}(\eta^k)^2\prod\limits_{j=k+1}^{K-1}(1-\frac{\mu}{2}\min\limits_mq_m\eta^j)\\ &=V^0\prod\limits_{k=0}^{K-1}\frac{k+K_0-2}{k+K_0}+\frac{16R}{\mu^2\min\limits_mq_m}\sum\limits_{k=0}^{K-1}\frac{1}{(k+K_0)^2}\prod\limits_{j=k+1}^{K-1}\frac{j+K_0-2}{j+K_0}\\ &\leq \frac{(K_0-2)(K_0-1)}{(K+K_0-2)(K+K_0-1)}V^0+\frac{16R}{\mu^2\min\limits_mq_m}\sum\limits_{k=0}^{K-1}\frac{1}{(k+K_0)^2}\frac{(k+K_0-1)(k+K_0)}{(K+K_0-2)(K+K_0-1)}\\ &\leq\frac{(K_0-1)^2}{(K+K_0-1)^2}(F(\theta_0^0,{\boldsymbol{\theta}}^0)-F^)+\frac{16RK}{\mu^2\min\limits_mq_m(K+K_0-1)^2}.\end{aligned}$$ Convergence under stochastic unbounded delay ============================================ We first present a useful fact. Given the definition of $\bar p_m, p_{m,d}$ in Assumption \[assump:unboundeddelay\], it can be shown that $$\begin{aligned} &\sum\limits_{s=d}^{\infty}sp_{m,s}=\bar p_m\left(\frac{d\rho^d}{1-\rho}+\frac{\rho^{d+1}}{(1-\rho)^2}\right):=c_{m,d}\\ &\sum\limits_{s=d}^{\infty}c_{m,s}=\bar p_m\left(\frac{d\rho^d}{(1-\rho)^2}+\frac{2\rho^{d+1}}{(1-\rho)^3}\right)\\ &\sum\limits_{d=1}^{\infty}c_{m,d}=\bar p_m\left(\frac{\rho}{(1-\rho)^2}+\frac{2\rho^2}{(1-\rho)^3}\right):=c_m.\end{aligned}$$ For unbounded delay, we have the following relation $$\begin{aligned} {\mathbb{E}}[\\|\hat{\boldsymbol{\theta}}^k-{\boldsymbol{\theta}}^k\\|^2\|\Theta^k] &=\sum\limits_{m=1}^M{\mathbb{E}}[\\|\theta_m^{k-\tau_{n_k,m}^k}-\theta_m^k\\|^2\|\Theta^k]\\ &=\sum\limits_{m=1}^M\sum\limits_{s=1}^{\infty}{\mathbb{E}}[\\|\theta_m^{k-s}-\theta_m^k\\|\|\Theta^k]{\mathbb{P}}(\tau_{n_k,m}^k=s)\\ &\leq\sum\limits_{m=1}^M\sum\limits_{s=1}^{\infty}\sum\limits_{d=1}^ssp_{m,s}\\|\theta_m^{k+1-d}-\theta_m^{k-d}\\|^2\\ &=\sum\limits_{m=1}^M\sum\limits_{d=1}^{\infty}c_{m,d}\\|\theta_m^{k+1-d}-\theta_m^{k-d}\\|^2.\end{aligned}$$ Similar to , we can decompose the difference term as $$\begin{aligned} \label{eqn:decompose3} {\mathbb{E}}\left[\\|\theta_m^{k+1}-\theta_m^k\\|^2\|\Theta^k\right]\leq 2q_m(\eta_m^k)^2\\|G_m^k\\|^2+2q_m(\eta_m^k)^2\sigma_m^2+2q_m(\eta_m^k)^2L^2{\mathbb{E}}[\\|\hat\theta_m^k-\theta_m^k\\|^2\|\Theta^k].\end{aligned}$$ Following Lemma \[lemma:Fdescent\] and , we have $$\begin{aligned} &{\mathbb{E}}[V^{k+1}\|\Theta^k]-V^k\\ =&{\mathbb{E}}[F(\theta_0^{k+1}, {\boldsymbol{\theta}}^{k+1})\|\Theta^k]-F(\theta_0^k, {\boldsymbol{\theta}}^k)+\sum\limits_{m=1}^M\gamma_{m,1}{\mathbb{E}}[\\|\theta_m^{k+1}-\theta_m^k\\|^2\|\Theta^k]+\sum\limits_{m=1}^M\sum\limits_{d=1}^{\infty}(\gamma_{d+1}-\gamma_d)\\|\theta_m^{k+1-d}-\theta_m^{k-d}\\|^2\\ \leq &-\left(\frac{\eta_0^k}{2}-L(\eta_0^k)^2\right)\\|G_0^k\\|^2-\sum\limits_{m=1}^Mq_m\left(\frac{\eta_m^k}{2}-(2\gamma_{m,1}+L)(\eta_m^k)^2\right)\\|G_m^k\\|^2+L(\eta_0^k)^2\sigma_0^2\\ &+\sum\limits_{m=1}^Mq_m(\eta_m^k)^2(2\gamma_{m,1}+L)\sigma_m^2 +\sum\limits_{m=1}^M\sum\limits_{d=1}^{k}\left(\left(c^k+2q_m\gamma_{m,1}(\eta_m^k)^2L^2\right)c_{m,d}+\gamma_{m,d+1}-\gamma_{m,d}\right)\\|\theta_m^{k+1-d}-\theta_m^{k-d}\\|^2.\end{aligned}$$ If we choose $\eta_0^k,\eta_m^k\leq\bar\eta\leq\frac{1}{4(L+2\max_m\gamma_{m,1})}$, then $c^k\leq \frac{3}{2}\bar\eta L^2$. By direct calculation, we have $$\begin{aligned} \label{eqn-pf:lemma3} {\mathbb{E}}V^{k+1}-{\mathbb{E}}V^k\leq&-\frac{1}{4}\min\{\eta_0^k,q_m\eta_m^k\}{\mathbb{E}}[\\|\nabla F(\theta_0^k,{\boldsymbol{\theta}}^k)\\|^2]+L(\eta_0^k)^2\sigma_0^2+\sum\limits_{m=1}^Mq_m(\eta_m^k)^2(2\gamma_{m,1}+L)\sigma_m^2\nonumber\\ &-\sum\limits_{m=1}^M\sum\limits_{d=1}^{\infty}\left(\gamma_{m,d}-\gamma_{m,d+1}-c_{m,d}\left(\frac{3}{2}\bar\eta L^2+2q_m\gamma_{m,1}\bar\eta^2L^2\right)\right){\mathbb{E}}[\\|\theta_m^{k+1-d}-\theta_m^{k-d}\\|^2].\end{aligned}$$ If we select $\gamma_{m,d}$ such that $$\begin{aligned} \left(\frac{3}{2}\bar\eta L^2+2q_m\gamma_{m,1}\bar\eta^2L^2\right)c_{m,d}+\gamma_{m,d+1}-\gamma_{m,d}=-\xi_m c_{m,d}, ~~~m=1,\ldots, M, ~d=1,\ldots, \infty\end{aligned}$$ then it remains that $$\begin{aligned} \gamma_{m,d}=\sum\limits_{s=d}^{\infty}c_{m,s}\left(\frac{3}{2}\bar\eta L^2+2q_m\gamma_{m,1}\bar\eta^2L^2+\xi_m\right).$$ Proof of Theorem \[thm:nc2\] ---------------------------- We set the parameters as $$\xi_m=0,\qquad \qquad c_m=\sum\limits_{d=1}^{\infty}c_{m,d},\qquad\qquad \gamma_{m,1}=\frac{\frac{3}{2}c_m\bar\eta L^2}{1-2c_mq_m\bar\eta^2L^2}\leq2\bar\eta c_mL^2\leq\frac{1}{2}\sqrt{c_m}L$$ and $$\eta_0^k=\eta_m^k=\eta=\min\left\{\frac{1}{4(1+\max_m\sqrt{c_m})L}, \frac{c_{\eta}}{\sqrt{K}}\right\}.$$ Plugging these constants into , we have $$\begin{aligned} \label{eqn:pf-thm3} {\mathbb{E}}V^{k+1}-{\mathbb{E}}V^k\leq\frac{1}{4}\min\limits_mq_m\eta{\mathbb{E}}[\\|\nabla F(\theta_0^k,{\boldsymbol{\theta}}^k)\\|^2]+\eta^2\sum\limits_{m=1}^Mq_m(2\gamma_{m,1}+L)\sigma_m^2.\end{aligned}$$ By summing over $k=0,\cdots,K-1$, it follows that $$\begin{aligned} \frac{1}{K}\sum\limits_{k=0}^{K-1}{\mathbb{E}}[\\|\nabla F(\theta_0^k,{\boldsymbol{\theta}}^k)\\|^2]&\leq\frac{F^0-F^+K\eta^2\sum\limits_{m=1}^M(2\gamma_{m,1}+L)q_m\sigma_m^2}{\frac{1}{4}\min\limits_mq_m\eta K}\\ &\leq\frac{16(1+\max_m\sqrt{c_m})L(F^0-F^)}{\min\limits_m q_mK}+\frac{4c_{\eta}(F^0-F^)}{\min\limits_mq_m\sqrt{K}}+\frac{4c_{\eta}\sum\limits_{m=1}^M(2\gamma_{m,1}+L)q_m\sigma_m^2}{\min\limits_mq_m\sqrt{K}}.\end{aligned}$$ Proof of Theorem \[thm:sc2\] ---------------------------- If we set $\xi_m=\frac{1}{4}c_m\bar\eta L^2$ and $\eta^k\leq\bar\eta=\frac{1}{4(1+\max_m\sqrt{c_m})L}$, then $$\gamma_{m,1}=\frac{\frac{3}{2}c_m\bar\eta L^2+\xi_m}{1-2c_m\bar\eta^2L^2}= 2c_m\bar\eta L^2\leq\frac{1}{2}\sqrt{c_m}L.$$ Plugging the parameters in and using the strong convexity in , we have $$\begin{aligned} {\mathbb{E}}V^{k+1}\leq(1-\nu\eta^k){\mathbb{E}}V^k+(\eta^k)^2R\end{aligned}$$ where $\nu=\inf\limits_{m,d}\{\frac{\xi_m c_{m,d}}{\bar\eta\gamma_{m,d}},\frac{\mu q_m}{2}\}$ and $R:=\big(L\sigma_0^2+(2\gamma_{m,1}+L)\sum\limits_{m=1}^Mq_m\sigma_m^2\big)$. Choosing $\eta^k=\frac{2}{\nu(k+K_0)}$ with $K_0=\frac{4(1+\max_m\sqrt{c_m})L}{\nu}$, it follows that $$\begin{aligned} {\mathbb{E}}V^k&\leq\prod\limits_{k=0}^{K-1}(1-\nu\eta^k)V^0+R\sum\limits_{k=0}^{K-1}(\eta^k)^2\prod\limits_{j=k+1}^{K-1}(1-\nu\eta^j)\\ &=V^0\prod\limits_{k=0}^{K-1}\frac{k+K_0-2}{k+K_0}+\frac{16R}{\mu^2\min\limits_mq_m}\sum\limits_{k=0}^{K-1}\frac{1}{(k+K_0)^2}\prod\limits_{j=k+1}^{K-1}\frac{j+K_0-2}{j+K_0}\\ &\leq \frac{(K_0-2)(K_0-1)}{(K+K_0-2)(K+K_0-1)}V^0+\frac{16R}{\mu^2\min\limits_mq_m}\sum\limits_{k=0}^{K-1}\frac{1}{(k+K_0)^2}\frac{(k+K_0-1)(k+K_0)}{(K+K_0-2)(K+K_0-1)}\\ &\leq\frac{(K_0-1)^2}{(K+K_0-1)^2}(F^0-F^)+\frac{16RK}{\mu^2\min\limits_mq_m(K+K_0-1)^2}.\end{aligned}$$ To verify the existence of $\nu>0$, we have $$\begin{aligned} \frac{\xi_m c_{m,d}}{\bar\eta\gamma_{m,d}}&=\frac{c_{m,d}}{\sum\limits_{s=d}^{\infty}c_{m,s}}\frac{1}{\bar\eta}\frac{\xi_m}{\frac{3}{2}\bar\eta L^2+2q_m\gamma_{m,1}\bar\eta^2L^2+\xi_m} \geq\frac{1}{\bar\eta}\frac{(1-\rho)\xi_m}{3\bar\eta L^2+4q_m\gamma_{m,1}\bar\eta^2L^2+2\xi_m}\\ &\geq\frac{1}{\bar\eta}\frac{(1-\rho)\xi_m}{5\bar\eta L^2+\xi_m}=\frac{1}{\bar\eta}\frac{(1-\rho)c_m}{20+c_m}\end{aligned}$$ where we use the fact that $\frac{c_{m,d}}{\sum\limits_{s=d}^{\infty}c_{m,s}}\geq\frac{1-\rho}{2}$. Then $\nu=\min\{\frac{1}{\bar\eta}\frac{(1-\rho)c_m}{20+c_m},\frac{\mu q_m}{2}\}$. Convergence results of vertical $t$-synchronous federated learning ================================================================== In the $t$-synchronous, we use ${\cal M}^k$ to denote the set of clients that upload at iteration $k$. For notational brevity, we define $$\begin{aligned} & \hat g_0^k=\frac{1}{t}\sum\limits_{m\in\mathcal M^k}\nabla_{\theta_0}\ell(\theta_0,h_{n_k,1}^{k-\tau_{n_k(m),1}^k},\ldots, h_{n_k(m),M}^{k-\tau_{n_k(m),M}^k};y_{n_k})\\ & \hat g_m^k=\left\{\begin{array}{ll} \nabla h_m(\theta_m^k;x_{n_k(m),m})\nabla_{h_m}\ell(\theta_0^k,h_{n_k(m),1}^{k-\tau_{n_k(m),1}^k},\ldots, h_{n_k(m),M}^{k-\tau_{n_k(m),M}^k};y_{n_k(m)}), & \text{if } m\in\mathcal M^k;\\ 0, & \text{else.} \end{array}\right.\end{aligned}$$ Similar to Assumption \[assump:client\], we assume that \[assump:tclient\] The probability of client $m$ in the set of uploading clients ${\cal M}_k$ at iteration $k$ is independent of ${\cal M}^{k-1}, \cdots, {\cal M}^1$, and it satisfies $$\begin{aligned} {\mathbb{P}}(m\in\mathcal M^k):=q_m.\end{aligned}$$ Connecting with asynchronous case --------------------------------- Similar to the previous analysis, the objective value satisfies the following inequality $$\begin{aligned} F(\theta_0^{k+1},{\boldsymbol{\theta}}^{k+1}) &\leq F(\theta_0^k,{\boldsymbol{\theta}}^k)+\dotp{G_0^k,\theta_m^{k+1}-\theta_0^k}+\dotp{G_m^k, \theta_m^{k+1}-\theta_m^k}+\frac{L}{2}\\|\theta_0^k-{\boldsymbol{\theta}}^k\\|^2+\sum\limits_{m=1}^M\frac{L}{2}\\|\theta_m^{k+1}-\theta_m^k\\|^2\\ &=F(\theta_0^k,{\boldsymbol{\theta}}^k)+\dotp{G_0^k, \hat g_0^k}+\sum\limits_{m\in\mathcal M^k}\dotp{G_m^k, \hat g_m^k}+\frac{L(\eta_0^k)^2}{2}\\|\hat g_0^k\\|^2+\frac{L}{2}\sum\limits_{m\in\mathcal M^k}(\eta_m^k)^2\\|\hat g_m^k\\|^2\\ &\leq F(\theta_0^k,{\boldsymbol{\theta}}^k)-\eta_0^k\dotp{G_0^k,g_0^k}-\sum\limits_{m\in\mathcal M^k}\eta_m^k\dotp{G_m^k,g_m^k}+\frac{\eta_0^k}{2}\\|G_0^k\\|^2+\sum\limits_{m\in\mathcal M}\frac{\eta_m^k}{2}\\|G_m^k\\|^2+L(\eta_0^k)^2\\|G_0^k\\|^2\\ &~~~~~+\sum\limits_{m\in\mathcal M^k}L(\eta_m^k)^2\\|G_m^k\\|^2+(\frac{\eta_0^k}{2}+L(\eta_0^k)^2)\\|\hat g_0^k-g_0^k\\|^2+\sum\limits_{m\in\mathcal M^k}(\frac{\eta_m^k}{2}+L(\eta_m^k)^2)\\|\hat g_m^k-g_m^k\\|^2.\end{aligned}$$ And by taking expectation with respect to $\mathcal M^k, n_k(m)$, it follows that (with $t=\sum\limits_{m=1}^Mq_m$) $$\begin{aligned} &{\mathbb{E}}[F(\theta_0^{k+1}, {\boldsymbol{\theta}}^{k+1})\|\Theta^k]\\ \leq &F(\theta_0^k, {\boldsymbol{\theta}}^k)-(\frac{\eta_0^k}{2}-L(\eta_0^k)^2)\\|G_0^k\\|^2-\sum\limits_{m=1}^Mq_m(\frac{\eta_{m}^k}{2}-L(\eta_m^k)^2)\\|G_m^k\\|^2+\frac{1}{t^2}L(\eta_0^k)^2\sigma_0^2+\sum\limits_{m=1}^Mq_mL(\eta_m^k)^2\sigma_m^2\\ &+\left(\left(\frac{\eta_0^k}{2}+L(\eta_0^k)^2\right)L_0^2+t\max\limits_m\left(\frac{\eta_{m}^k}{2}+L(\eta_m^k)^2\right)L_m^2\right){\mathbb{E}}[\\|\hat {\boldsymbol{\theta}}^k-{\boldsymbol{\theta}}^k\\|^2\|\Theta^k].\end{aligned}$$ Following the Lyapunov analysis of the asynchronous case, it can be shown that the vertical $t$-synchronous federated learning achieves the same order of convergence rate as in Theorems \[thm:nc1\]-\[thm:sc2\]. Convergence results ------------------- For completeness, we state the convergence results for the vertical $t$-synchronous federated learning as follows. \[thm:nc1t\] Under Assumptions \[assump:sample\],\[assump:lip\],\[assump:boundeddelay\] and \[assump:tclient\], if $\eta_0^k=t\eta_m^k=\min\{\frac{1}{4(1+D)L}, \frac{c_{\eta}}{\sqrt{K}}\}$ with $c_{\eta}\!>\!0$, then we have $$\frac{1}{K}\sum\limits_{k=0}^{K-1}{\mathbb{E}}[\\|\nabla F(\theta_0^k,{\boldsymbol{\theta}}^k)\\|^2]={\cal O}\left({1}/{\sqrt{K}}\right).$$ \[thm:sc1t\] Assume that $F$ is $\mu$-strongly convex in $(\theta_0,{\boldsymbol{\theta}})$. Then under the same assumptions of Theorem \[thm:nc1t\], if $\eta^k=\frac{4}{\mu\min\limits_m\sqrt{q_m}(k+K_0)}$ with $K_0=\frac{4(4(D+1)L+\mu\min\limits_m\sqrt{q_m}D)}{\mu t\min\limits_m\sqrt{q_m}}$, then $${\mathbb{E}}F\left(\theta_0^K,{\boldsymbol{\theta}}^K\right)-F^={\cal O}\left({1}/{K}\right).$$ \[thm:nc2t\] Under Assumptions \[assump:sample\],\[assump:lip\],\[assump:unboundeddelay\] and \[assump:tclient\], if we choose $\eta_0^k=t\eta_m^k=\min\big\{\frac{1}{4(1+\min_m\sqrt{c_m})L}, \frac{c_{\eta}}{\sqrt{K}}\big\}$, then we have $$\frac{1}{K}\sum\limits_{k=0}^{K-1}{\mathbb{E}}[\\|\nabla F(\theta_0^k,{\boldsymbol{\theta}}^k)\\|^2]={\cal O}\left({1}/{\sqrt{K}}\right).$$ \[thm:sc2t\] Assume that $F$ is $\mu$-strongly convex in $(\theta_0,{\boldsymbol{\theta}})$. Then under the same assumptions of Theorem \[thm:nc2t\], if $\eta_0^k=\eta_m^k=\frac{2}{\nu(k+K_0)}$ where $K_0=\frac{4(1+\max_m\sqrt{c_m})L}{t\nu}$ and $\nu$ is a positive constant depending on $\mu, L, \bar p_m, \rho$, then it follows that $${\mathbb{E}}F\left(\theta_0^K,{\boldsymbol{\theta}}^K\right)-F^={\cal O}\left({1}/{K}\right).$$ Proof of Theorem \[thm:smoothness\] =================================== Before proceeding to the proof of Theorem \[thm:smoothness\], we first present the smoothness of a single neuron in the following lemma. \[lemma:singlesmoothness\] If $\sigma(x)$ is $L_{\sigma}^0$-Lipschitz continuous and differentiable almost everywhere, $Z$ is a continuous random variable with pdf $\mu(Z)$, then $\bar\sigma(x):={\mathbb{E}}\sigma(x+Z)$ is differentiable with Lipschitz continuous gradient $\nabla\bar\sigma(x)={\mathbb{E}}\nabla\sigma(x+Z)$. We first prove that $\bar\sigma(x)$ is smooth and ${\mathbb{E}}\nabla\sigma(x+Z)=\nabla\bar\sigma(x)$. $$\begin{aligned} \frac{{\mathbb{E}}_Z\sigma(x+\delta v+Z)-{\mathbb{E}}_Z\sigma(x+Z)}{\delta}=\int_{{\mathbb{R}}^d}\frac{\sigma(x+\delta v+Z)-\sigma(x+Z)}{\delta}\mu(Z)dZ\end{aligned}$$ Since $\sigma$ is differentiable almost everywhere, for any fixed $x\in{\mathbb{R}}^d$ and directional vector $v\in{\mathbb{R}}^d$, we have $$\begin{aligned} \lim_{\delta\rightarrow 0}\frac{\sigma(x+\delta v+Z)-\sigma(x+Z)}{\delta}=v^{\top}\nabla\sigma(x+Z)=\sum\limits_{i=1}^n\frac{\partial \sigma}{\partial x_i}(x+Z)v_i\quad\quad a.e.\end{aligned}$$ and $$\begin{aligned} \int_{{\mathbb{R}}^d}\left\|\frac{\sigma(x+\delta v+Z)-\sigma(x+Z)}{\delta}\right\|\mu(Z)dZ\leq\int_{{\mathbb{R}}^d}L_{\sigma}^0\mu(Z)dZ=L_{\sigma}^0.\end{aligned}$$ Then by dominated convergence theorem, when taking $\delta\rightarrow 0$, it follows that $$\begin{aligned} &\frac{\partial\bar\sigma}{\partial x_i}(x)=\int_{{\mathbb{R}}^d}\frac{\partial\sigma}{\partial x_i}(x+Z)\mu(Z)dZ,\\ &\frac{\partial\bar\sigma}{\partial v}(x)=\int_{{\mathbb{R}}^d}\sum\limits_{i=1}^d\frac{\partial \sigma}{\partial x_i}(x+Z)v_i\mu(Z)dZ=\sum\limits_{i=1}^n\frac{\partial\bar\sigma}{\partial x_i}(x)v_i.\end{aligned}$$ Therefore, $\bar\sigma(x)$ is differentiable, that is $$\begin{aligned} \nabla\bar\sigma(x)=\int_{{\mathbb{R}}^d}\nabla\sigma(x+Z)\mu(Z)dZ={\mathbb{E}}_Z\nabla\sigma(x+Z).\end{aligned}$$ Next we derive the smoothness constant of $\bar\sigma(x)$. We focus on the uniform distribution and the Gaussian distribution. Case I.* Assume that the uniform distribution $Z\sim\mathcal U[-\frac{c}{2},\frac{c}{2}]^d$, i.e., $\mu(Z)=\frac{1}{c^d}\mathbbm{1}_{\{-\frac{c}{2}\leq Z_i\leq \frac{c}{2},1\leq i\leq d\}}(Z)$. $$\begin{aligned} \\|\nabla\bar\sigma(x)-\nabla\bar\sigma(x')\\|&=\left\\|\int_{{\mathbb{R}}^d}\nabla\sigma(x+y)\mu(Z)dZ-\int_{{\mathbb{R}}^d}\nabla\sigma(x'+Z)\mu(Z)dZ\right\\|\\ &=\left\\|\int_{{\mathbb{R}}^d}\nabla\sigma(y)(\mu(y-x)-\mu(y-x'))dy\right\\|\leq L_{\sigma}^0\int_{{\mathbb{R}}^d}\left\|\mu(y-x)-\mu(y-x')\right\|dy\\ &\leq \frac{2\sqrt{d}L_{\sigma}^0}{c}\\|x-x'\\|:=L_{\bar{\sigma}}(c)\\|x-x'\\|\end{aligned}$$ where the smoothness constant is defined as $L_{\bar{\sigma}}(c):=\frac{2\sqrt{d}L_{\sigma}^0}{c}$. Case II. Assume that the Gaussian distribution $Z\sim\mathcal N(0, c^2I_d)$, i.e., $\mu(Z)=\frac{1}{\sigma\sqrt{2\pi}}e^{-\frac{\\|Z\\|^2}{2\sigma^2}}$. $$\begin{aligned} \bar\sigma(x)=\int_{{\mathbb{R}}^d}\sigma(x+Z)\mu(Z)dZ=\int_{Z\in{\mathbb{R}}^d}\sigma(y)\mu(y-x)dy.\end{aligned}$$ By the Leibniz rule, we have $$\begin{aligned} \nabla\bar{\sigma}(x)=-\int_{{\mathbb{R}}^d}\sigma(y)\nabla\mu(y-x)dy=-\int_{{\mathbb{R}}^d}\sigma(y+x)\nabla\mu(y)dy.\end{aligned}$$ Then it follows $$\begin{aligned} \\|\nabla\bar\sigma(x)-\nabla\bar\sigma(x')\\|&=\left\\|\int_{{\mathbb{R}}^d}(\sigma(y+x)-\sigma(y+x'))\nabla\mu(y)dy\right\\|\\ &\leq L_{\sigma}^0\left(\int_{{\mathbb{R}}^d}\\|\nabla\mu(y)\\|dy\right)\\|x-x'\\|\\ &=\frac{L_{\sigma}^0d}{c}\\|x-x'\\|:= L_{\bar{\sigma}}(c)\\|x-x'\\|\end{aligned}$$ where the smoothness constant is $L_{\bar{\sigma}}(c):=\frac{L_{\sigma}^0d}{c}$. Proof of Theorem \[thm:smoothness\] ----------------------------------- Building upon Lemma \[lemma:singlesmoothness\], we next prove Theorem \[thm:smoothness\]. For simplicity, we assume that all the activation functions are same, e.g., $\sigma_l=\sigma,\,\forall l=1,\cdots,L$. We use $L_{f}$ to denote the lipschitz constant of a function $f$. In the following proof, we change the order of differentiation and integration (expectation) as it is supported by Leibniz integral rule. We also let $\bar f={\mathbb{E}}f$. Since $\nabla_{b_L}\bar h={\mathbb{E}}[\nabla\bar\sigma], \nabla_{w_L}\bar h={\mathbb{E}}[\nabla\bar\sigma]u_{L-1}^{\top}, \nabla_{u_{L-1}}{\mathbb{E}}_{Z_L}h=w_L^{\top}{\mathbb{E}}[\nabla\bar\sigma]$. The smoothness of $\bar\sigma$ implies that $\nabla_{b_L}\bar h, \nabla_{w_L}\bar h, \nabla_{u_{L-1}}\bar h$ are $L_{b_L}^{\bar h}, L_{w_L}^{\bar h}, L_{u_{L-1}}^{\bar h}$-Lipschitz continuous respectively, with $$\begin{aligned} &L_{b_L}^{\bar h}:=L_{\bar{\sigma}}(c),\\ &L_{w_L}^{\bar h}:=L_{b_L}^{\bar h}{\mathbb{E}}[\\|u_{L-1}\\|],\\ &L_{u_{L-1}}^{\bar h}:=L_{b_L}^{\bar h}\\|w_L\\|,\\ &\\|\nabla_{u_{L-1}}\bar h\\|\leq L_{\sigma}^0\\|w_L\\|.\end{aligned}$$ Since $\sigma$ is differentiable almost everywhere, $\bar\sigma(w_L\sigma(\cdot + Z_{L-1})+b_L)$ is differentiable almost everywhere and thus is smooth in expectation of $Z_{L-1}$. By some calculation, we can show that $$\begin{aligned} &L_{b_{L-1}}^{\bar h}=L_{u_{L-1}}^{\bar h}(L_{\sigma}^0)^2+L_{\sigma}^0\\|w_L\\|L_{\bar{\sigma}}(c_l)\\ &L_{w_{L-1}}^{\bar h}=L_{b_{L-1}}^{\bar h}{\mathbb{E}}[\\|u_{L-1}\\|]\\ &L_{u_{L-2}}^{\bar h}=L_{b_{L-1}}^{\bar h}\\|w_{L-1}\\|.\end{aligned}$$ Following the similar steps, we can obtain that $$\begin{aligned} &L_{b_l}^{\bar h}=L_{u_l}^{\bar h}(L_{\sigma}^0)^2+L_{\sigma}^0\\|w_L\\|\cdots L_{\sigma}^0\\|w_{l+1}\\|L_{\bar\sigma}(c_l),\\ &L_{w_l}^{\bar h}=L_{b_l}^{\bar h}{\mathbb{E}}[\\|u_{l-1}\\|],\\ &L_{u_{l-1}}^{\bar h}=L_{b_l}^{\bar h}\\|w_l\\|. $$ As long as the overall loss $\ell(\theta_0, h_1,\ldots, h_M;y)$ is smooth w.r.t. $\theta_0, h_1,\ldots, h_M$, we can extend our results to show that it is smooth in the local parameters $\theta_1,\ldots,\theta_M$. Taking $u_l$ from $h_m$ as example, that is $$\begin{aligned} L^{\bar\ell}_{\theta_m}=L^{\bar\ell}_{h_m}(L_{h_m}^0)^2+L_{\ell}^0L^{\bar h_m}_{\theta_m}\end{aligned}$$ we can extend our results to show that $F_c(\theta_0,{\boldsymbol{\theta}})=\frac{1}{N}\sum\limits_{n=1}^N{\mathbb{E}}\ell(\theta_0,h_{n,1},\ldots,h_{n,M};y_n)+\sum\limits_{m=1}^Mr(\theta_m)$ is smooth, where the expectation is taken with respect to all the random neurons in local embedding vectors $h_1,\ldots, h_M$. Specifically, the smoothness of $F_c$ is given by $$L^{F_c}_{\theta_m}=L^{\bar\ell}_{h_m}(L_{\bar h_m}^0)^2+L_{\ell}^0\sum_{l=1}^L(L_{w_l}^{\bar h}+L_{b_l}^{\bar h})+L^{r}_{\theta_m}$$ where $L^{r}_{\theta_m}$ is the smoothness constant of the regularizer w.r.t. $\theta_m$; $L_{b_l}^{\bar h}$ and $L_{w_l}^{\bar h}$ are the smoothness constants of the perturbed local embedding $h$ w.r.t. the bias $b_l$ and weight $w_l$. The objective difference after local perturbation ------------------------------------------------- Now we evaluate the difference between $F_{c}(\theta_0,{\boldsymbol{\theta}})$ and $F(\theta_0,{\boldsymbol{\theta}})$. Note that $$\begin{aligned} \label{eqn:smoothdiff} \|F_{c}(\theta_0,{\boldsymbol{\theta}})-F(\theta_0,{\boldsymbol{\theta}})\|^2 &=\left\|\frac{1}{N}\sum_{n=1}^N({\mathbb{E}}_{\mathbf{Z}}\ell(\theta_0,h_{n,1}',\ldots,h_{n,M}’;\mathbf {Z})-\ell(\theta_0,h_{n,1},\ldots,h_{n,M}))\right\|^2\nonumber\\ &\leq\frac{1}{N}\sum_{n=1}^N{\mathbb{E}}_{\mathbf{Z}}[\|\ell(\theta_0,h_{n,1}',\ldots,h_{n,M}';\mathbf{Z})-\ell(\theta_0,h_{n,1},\ldots,h_{n,M})\|^2]\nonumber\\ &\leq\frac{M}{N}\sum_{n=1}^NL_{\ell}^2{\mathbb{E}}_{\mathbf{Z}}[\\|h_{n,m}'-h_{n,m}\\|^2]\end{aligned}$$ where $h_{n,m}$ and $h_{n,m}'$ correspond to the outputs of and , respectively. Since we have that $$\begin{aligned} &\\|h_{n,m}'-h_{n,m}\\|=\\|u_L'-u_L\\|\leq L_{\sigma_L}(\\|w_L\\|\\|u_{L-1}'-u_{L-1}\\|+\\|Z_L\\|)\leq\cdots\leq\sum\limits_{j=1}^L\left(\prod\limits_{l=j}^{L} L_{\sigma_l}\\|w_l\\|\right)\\|Z_l\\|\end{aligned}$$ and thus it follows that $$\begin{aligned} &\\|h_{n,m}'-h_{n,m}\\|^2\leq\left(\sum\limits_{j=1}^L\prod\limits_{l=j}^LL_{\sigma_l}^2\\|w_l\\|^2\right)\left(\sum\limits_{j=1}^L\\|Z_j\\|^2\right).\end{aligned}$$ Taking expectation on both side, we have $$\begin{aligned} &{\mathbb{E}}[\\|h_{n,m}'-h_{n,m}\\|^2]\leq\left(\sum\limits_{j=1}^L\prod\limits_{l=j}^LL_{\sigma_l}^2\\|w_l\\|^2\right)\left(\sum\limits_{j=1}^{L-1}{c_l^2}+c^2\right).\end{aligned}$$ Plugging into , we arrive at $$\begin{aligned} \|F_{c}(\theta_0,{\boldsymbol{\theta}})-F(\theta_0,{\boldsymbol{\theta}})\|\leq M\left(\sum\limits_{j=1}^L\prod\limits_{l=j}^LL_{\sigma_l}^2\\|w_l\\|^2\right)^{\frac{1}{2}}\left(\sum\limits_{j=1}^{L-1}{c_l^2}+c^2\right)^{\frac{1}{2}}.\end{aligned}$$ Proof of Theorem \[thm:privacy\] ================================ Let $u_l,u_l'$ denote the the outputs of $l$-th layer with inputs $u_0=x,x'$. Under the assumptions that $Z_l\sim\mathcal U[-c_l/2,c_l/2]$,$\sigma_l$ is $L_{\sigma_l}$-Lipschitz continuous for $l=1,\ldots, L-1$, we can derive that $$\begin{aligned} \\|w_Lu_{L-1}-w_Lu_{L-1}'\\| &\leq \\|w_L\\|L_{\sigma_{L-1}}(\\|w_{L-1}\\|\\|u_{L-2}-u_{L-2}'\\|+\sqrt{d_{L-1}}c_{L-1})\\ &\leq \\|w_L\\|\prod\limits_{l=1}^{L-1}L_{\sigma_l}\\|w_l\\|\\|x-x'\\|+\\|w_L\\|\sum\limits_{l=1}^{L-1}\left(\prod\limits_{j=1}^{l}L_{\sigma_j}\sqrt{d_j}\right)c_{l}\\ &:=\bar B.\end{aligned}$$ Consider the linear operation of $L$-th layer $\mathcal M(u_{L-1})=w_Lu_{L-1}+b_L+Z_L$ which is a random mechanism defined by $Z_L\sim\mathcal N(0,\nu^2)$. Since differential privacy is immune to post-processing [@dwork2014algorithmic], $\sigma_L\circ\mathcal{M}$ does not increase the privacy loss compared with $\mathcal M$. According to Theorem 1 in [@abadi2016], Algorithm \[alg:Async-SGD-FL\] is $(\varepsilon,\delta)$-differentially private if $\nu=c\frac{q\sqrt{T\log(1/\delta)}}{\varepsilon}$. Simulation details ================== In this section, we present the details of our simulations, and provide the additional test results. Simulation environment ---------------------- We conducted our simulations on a deep learning workstation with 2 and 2 GPUs. Codes are written using and . VAFL for federated logistic regression -------------------------------------- Data allocation. The datasets we choose are CIFAR-10, Parkinson Disease, MNIST and Fashion MNIST. The batch size is selected to be approximate 0.01 fraction of the entire training dataset. The data are uniformly distributed among $M=8$ clients for CIFAR-10, $M=3$ for Parkinson Disease, and $M=7$ for both MNIST and Fashion MNIST. Stepsize. The stepsize is $\eta=1\times10^{-2}$ for Parkinson Disease, $\eta=2\times10^{-4}$ for CIFAR-10, and $\eta=1\times10^{-4}$ for both MNIST and Fashion MNIST. Random delay. The random delay follows a Poisson distribution with client-specific parameters to reflect heterogeneity. The delay on each worker $m$ follows the Poisson distribution with parameter $2m$ and scaled by $1/2M$, where $M$ is the number of workers and $m$ is the worker index. The expectation of maximum worker delay is one second. Perturbation. The noise added to the output of each local client follows the Gaussian distribution of each task is $\mathcal N(0,0.01)$ for CIFAR-10, MNIST and Fashion MNIST and $\mathcal N(0,1)$ for Parkinson Disease. For each task, we run the algorithms sufficiently many epochs and record the training loss. Testing accuracy and wall clock time are recorded at the end of each epoch. ![Training loss of VAFL with nonlinear local embedding on ModelNet40 dataset.[]{data-label="fig:Nonlinear_loss2"}](model40_train_1){width=".45\textwidth"} ![Testing accuracy of VAFL with nonlinear local embedding on ModelNet40 dataset.[]{data-label="fig:Nonlinear_acc2"}](model40_test_1){width=".45\textwidth"} ![Training loss of VAFL with local LSTM embedding on MIMIC-III critical care dataset.[]{data-label="fig:mortality_loss2"}](mimic_train_1){width=".45\textwidth"} ![AUC curve of VAFL with local LSTM embedding on MIMIC-III clinical care dataset.[]{data-label="fig:mortality_auc2"}](mimic_test_1){width=".45\textwidth"} VAFL for federated deep learning -------------------------------- ### Training on ModelNet40 dataset Local embedding structure. We train a convolutional neural network-based model consisting of two parts: the local embedding models and the server model. Each local model is a 7-layer convolutional neural network. The server part is a centralized 3-layer fully connect neural network. Vertical data allocation. The data we choose is ModelNet40 and we vertically distributed images of the objects in the dataset from 12 angles and assign to each local client. Each local client deals with the data assigned by their local convolutional network and generate a vector whose dimension is $512$ as the local output. Random delay. The random delay follows exponential distribution with client-specific parameters to reflect heterogeneity. For each worker $m$, the delay follows the exponential distribution with parameter $m$. Random perturbation. We use ReLU as the local embedding activation function. We add a random noise on the output of each local embedding convolutional layer. The noises follow the following distributions: $\mathcal U(-0.1,0.1)$ (the first two layers), $\mathcal U(-0.01,0.01)$ (the other convolutional layers except the last layer) and $\mathcal N(0,1)$ (the last convolutional layers). Server structure. The server then combines the $12$ vectors linearly and pass them into the three-layer fully connected neural network and classify into 40 classes. The number of nodes of each layer is $256$, $100$ and $40$. Learning rate. The stepsize of the local embedding update $\eta_m$ is $10^{-3}$ and the server stepsize $\eta_0 = \frac{\eta_m}{M}$ where $M$ is the number of workers. ### Training on MIMIC-III dataset MIMIC is an open dataset comprising deidentified health data associated with 60,000 intensive care unit admissions [@johnson2016mimic]. The data are allocated into 4 workers having different feature dimensions. Local embedding structure. The local embedding part is a two layer LSTM models and the server part is a fully connected layer. The first layer is a bidirectional LSTM and the number of units is 16. The second layer is a normal LSTM layer and the number of units is also 16. Random delay. The random delay follows an exponential distribution with client-specific parameters to reflect heterogeneity. The delay on each worker $m$ follows an exponential distribution with parameter $m$. Random perturbation. A random noise following Gaussian distribution $\mathcal N(0,10^{-4})$ is also added on the output of each local embedding layer. We have also added simulation results that compare all the algorithms with their private counterparts on both ModelNet40 and MIMIC-III datasets.	{ "pile_set_name": "ArXiv" }
--- abstract: \| While direct social ties have been intensely studied in the context of computer-mediated social networks, indirect ties (e.g., friends of friends) have seen little attention. Yet in real life, we often rely on friends of our friends for recommendations (of good doctors, good schools, or good babysitters), for introduction to a new job opportunity, and for many other occasional needs. In this work we attempt to 1) quantify the strength of indirect social ties, 2) validate it, and 3) empirically demonstrate its usefulness for distributed applications on two examples. We quantify social strength of indirect ties using a(ny) measure of the strength of the direct ties that connect two people and the intuition provided by the sociology literature. We validate the proposed metric experimentally by comparing correlations with other direct social tie evaluators. We show via data-driven experiments that the proposed metric for social strength can be used successfully for social applications. Specifically, we show that it alleviates known problems in friend-to-friend storage systems by addressing two previously documented shortcomings: reduced set of storage candidates and data availability correlations. We also show that it can be used for predicting the effects of a social diffusion with an accuracy of up to 93.5%. author: - - - - bibliography: - 'sigproc.bib' title: The Power of Indirect Social Ties ---	{ "pile_set_name": "ArXiv" }
--- abstract: 'We obtain Taylor approximations for functionals $V\mapsto{{\mathrm{Tr}}}\big(f(H_0+V)\big)$ defined on the bounded self-adjoint operators, where $H_0$ is a self-adjoint operator with compact resolvent and $f$ is a sufficiently nice scalar function, relaxing assumptions on the operators made in [@vanS], and derive estimates and representations for the remainders of these approximations.' address: 'A.S., Department of Mathematics and Statistics, University of New Mexico, 400 Yale Blvd NE, MSC01 1115, Albuquerque, NM 87131, USA' author: - 'Anna Skripka$^{}$' title: Asymptotic expansions for trace functionals --- [^1] Introduction ============ Let $H_0$ be an unbounded self-adjoint operator, $V$ a bounded self-adjoint operator on a separable Hilbert space $\mathcal{H}$, $f$ a sufficiently nice scalar function, and let $f(H_0+V)$ be defined by the standard functional calculus. The functionals $f\mapsto {{\mathrm{Tr}}}\big(f(H_0+V)\big)$ and $V\mapsto {{\mathrm{Tr}}}\big(f(H_0+V)\big)$ or their modifications have been involved in problems of perturbation theory (of, for instance, differential operators) and noncommutative geometry since as early as 1950’s (see, e.g., [@Azamov0; @ACS; @B; @CC; @DK; @Hansen; @L; @vanS]). The latter functional in the context of noncommutative geometry is called the spectral (action) functional [@CC]. Assume that the resolvent of $H_0$ belongs to some Schatten ideal (or, more generally, ${{\mathrm{Tr}}}\big(e^{-t H_0^2}\big)<\infty$, for any $t>0$), $\\|\delta(V)\\|<\infty$, $\\|\delta^2(V)\\|<\infty$, where $\delta(\cdot)=[\|H_0\|,\cdot]$, and $f$ is a sufficiently nice even function. Let $\{\mu_k\}_{k=1}^\infty$ be a sequence of eigenvalues of $H_0$ counting multiplicity and let $\{\psi_k\}_{k=1}^\infty$ be an orthonormal basis of the respective eigenvectors. The asymptotic expansion of the spectral action functional $$\begin{aligned} \label{2} &{{\mathrm{Tr}}}\big(f(H_0+V)\big)={{\mathrm{Tr}}}\big(f(H_0)\big)\\ \nonumber &\quad+\sum_{p=1}^\infty \frac1p\sum_{i_1,\ldots,i_p}(f')^{[p-1]}(\mu_{i_1},\ldots,\mu_{i_p})\, \left<V\psi_{i_1},\psi_{i_2}\right>\cdots\left<V\psi_{i_{p-1}},\psi_{i_p}\right>\left<V\psi_{i_p},\psi_{i_1}\right>,\end{aligned}$$ where $(f')^{[p-1]}$ is the divided difference of order $p-1$ of the function $f'$, was derived in [@vanS], extending the results of [@Hansen] for finite-dimensional operators. (The precise assumptions on $H_0$, $V$, and $f$ can be found in [@vanS Theorem 18].) In this paper, we obtain the asymptotic expansion under relaxed assumptions on $H_0$ and $V$ and find bounds for the remainders of the respective approximations by taking a different approach to the problem. Specifically, we assume that $H_0=H_0^$ has compact resolvent, $V=V^\in{\mathcal{B}(\mathcal{H})}$ (where ${\mathcal{B}(\mathcal{H})}$ is the algebra of bounded linear operators on $\mathcal{H}$), and $f$ is a sufficiently nice compactly supported function (but no summability restriction on $H_0$ is made, $H_0$ is not assumed to be positive, and $f$ is not assumed to be even). Let $$\begin{aligned} \label{eq:sum} &R_{H_0,f,n}(V):={{\mathrm{Tr}}}\big(f(H_0+V)\big)-{{\mathrm{Tr}}}\big(f(H_0)\big)\\ \nonumber &\quad-\sum_{p=1}^{n-1} \frac1p\sum_{i_1,\ldots,i_p}(f')^{[p-1]}(\mu_{i_1},\ldots,\mu_{i_p})\, \left<V\psi_{i_1},\psi_{i_2}\right>\cdots\left<V\psi_{i_{p-1}},\psi_{i_p}\right>\left<V\psi_{i_p},\psi_{i_1}\right>.\end{aligned}$$ In Theorem \[asexp\] and Corollary \[asexpev\], we establish the bound $$\begin{aligned} \label{r1} \|R_{H_0,f,n}(V)\|=\mathcal{O}\big(\\|V\\|^n\big)\end{aligned}$$ and find an explicit estimate for $\mathcal{O}\big(\\|V\\|^n\big)$ in Theorem \[derbound\] and Remark \[remarks\]. (The case $n=1$ also follows from [@ACS].) If, in addition, $H_0$ has Hilbert-Schmidt resolvent, we refine the bound of Theorem \[asexp\] in Theorem \[asexpL2\]. In Theorem \[ac\], we show that the functional $C_c^{3}({{\mathbb R}})\ni f''\mapsto R_{H_0,f,2}(V)$ is given by a locally finite absolutely continuous measure. (An analogous result for the functional $f'\mapsto R_{H_0,f,1}(V)$) was obtained in [@ACS].) Preliminaries ============= The asymptotic expansion can be rewritten as $$\begin{aligned} \label{3} \nonumber &{{\mathrm{Tr}}}\big(f(H_0+V)\big)={{\mathrm{Tr}}}\big(f(H_0)\big)\\ &\quad+\sum_{p=1}^\infty \frac1p\sum_{{\lambda}_1,\ldots,{\lambda}_p\in \text{\rm spec}(H) }(f')^{[p-1]}({\lambda}_1,\ldots,{\lambda}_p)\, {{\mathrm{Tr}}}\big(E_{H_0}({\lambda}_1)V\ldots VE_{H_0}({\lambda}_p)V\big),\end{aligned}$$ where $E_{H_0}$ is the spectral measure of $H_0=H_0^$. In this section, we justify that the traces in are well defined and prepare a technical base for the derivation of . #### Functional calculus We start with recalling some useful features of functional calculus for self-adjoint operators with compact resolvents. Note that if the resolvent of an operator is compact at one point, then it is compact at all points of its domain. Note also that $({\text{\rm i}}+H_0)^{-1}$ is compact if and only if $\|({\text{\rm i}}+H_0)^{-1}\|=(1+H_0^2)^{-1/2}$ is compact. By standard properties of the resolvent, we have $($[@ACS Lemma 1.3]$)$ \[compres\] If $H_0=H_0^$ is defined in $\mathcal{H}$ and has compact resolvent and if $W=W^\in{\mathcal{B}(\mathcal{H})}$, then $H_0+W$ also has compact resolvent. $($[@CPh Appendix B, Lemma 6]$)$ \[resest\] If $H_0=H_0^$ is defined in $\mathcal{H}$ and if $W=W^\in{\mathcal{B}(\mathcal{H})}$, then $$\begin{aligned} \big(1+(H_0+W)^2\big)^{-1}\leq\left(1+\\|W\\|+\\|W\\|^2\right)(1+H_0^2)^{-1}.\end{aligned}$$ The following consequence was essentially established in [@ACS Lemma 1.4]. \[lemma:Efinite\] Let $H_0=H_0^$ have compact resolvent and let $W=W^$ be bounded. Then, for any compact subset $\delta$ of ${{\mathbb R}}$, the spectral projection $E_{H_0+W}(\delta)$ has finite rank and $$\begin{aligned} \label{eq:Efinite} E_{H_0+W}(\delta)\leq\left(1+\max_{s\in\delta}\|s\|^2\right)\cdot \left(1+\\|W\\|+\\|W\\|^2\right)(1+H_0^2)^{-1}.\end{aligned}$$ From the spectral theorem we have $$E_{H_0+W}(\delta)\leq\left(1+\max_{s\in\delta}\|s\|^2\right)\cdot\big(1+(H_0+W)^2\big)^{-1}.$$ Application of Lemma \[resest\] gives , which, in particular, implies that $E_{H_0+W}$ has finite rank. Note that for a compact subset $\delta$ of ${{\mathbb R}}$, ${{\mathrm{Tr}}}\big(E_H(\delta)\big)$ equals the number of eigenvalues of $H$, counting multiplicities, in the set $\delta$. If $H_0=H_0^$ satisfies $\big(1+H_0^2\big)^{-1/2}\in {\mathcal{S}}^p$, with $p\geq 1$, and $W=W^$ is bounded, then $\big(1+(H_0+W)^2\big)^{-1/2}\in {\mathcal{S}}^p$. The result follows from Lemma \[resest\] and operator monotonicity of the function $t\mapsto t^{p/2}$. Let ${\mathcal{S}}^\alpha$ denote the Schatten ideal of order $\alpha$, that is, $${\mathcal{S}}^\alpha=\left\{A\in{\mathcal{B}(\mathcal{H})}:\; \\|A\\|_\alpha:=\big({{\mathrm{Tr}}}(\|A\|^\alpha)\big)^\frac{1}{\alpha}<\infty\right\}.$$ By standard properties of the Schatten norms, Corollary \[lemma:Efinite\], Theorem \[HSest\], and by the spectral theorem, we have the following lemma. \[fH\] Let $H=H^$ and let $f$ be a continuous compactly supported function on ${{\mathbb R}}$. (i) If $H$ has compact resolvent, then $f(H)\in {\mathcal{S}}^1$ and $$\begin{aligned} \label{straight} \\|f(H)\\|_1\leq \\|f\\|_\infty\cdot{{\mathrm{Tr}}}\big(E_H({\text{\rm supp}}f)\big).\end{aligned}$$ (ii) Let $u(t)=(1+t^2)^{1/2}$. If $(1+H^2)^{-1/2}\in {\mathcal{S}}^2$, then $$\begin{aligned} \label{straight2} \\|(fu)(H)\\|_2\leq \\|fu^2\\|_\infty\cdot\big\\|(1+H^2)^{-1/2}\big\\|_2.\end{aligned}$$ #### Operator derivatives* Let $H_0=H_0^$ be defined in $\mathcal{H}$ and let $V=V^\in{\mathcal{B}(\mathcal{H})}$. Denote $$\begin{aligned} \label{eq:sumOp} \mathcal{R}_{H_0,f,p}(V):=f(H_0+V)-\sum_{k=0}^{p-1}\frac{1}{k!}\cdot\frac{d^k}{ds^k}\bigg\|_{s=0} f(H_0+sV),\end{aligned}$$ provided the Gâteaux derivatives exist in the operator norm. We will see in the proof of Theorem \[asexp\] that $R_{H_0,f,p}(V)$ from equals ${{\mathrm{Tr}}}\big(\mathcal{R}_{H_0,f,p}(V)\big)$. Now we list results that guarantee the estimate $$\begin{aligned} \\|\mathcal{R}_{H_0,f,p}(V)\\|=\mathcal{O}\big(\\|V\\|^p\big)\end{aligned}$$ for the operator norm of the remainder and help to establish the estimate for the trace of the remainder. Recall that the divided difference of order $p$ is an operation on functions $f$ of one (real) variable, which we will usually call ${\lambda}$, defined recursively as follows: $$\begin{aligned} &f^{[0]}({\lambda}_0):=f({\lambda}_0),\\ &f^{[p]}({\lambda}_0,\ldots,{\lambda}_p):= \begin{cases} \frac{f^{[p-1]}({\lambda}_0,\ldots,{\lambda}_{p-2},{\lambda}_{p-1})- f^{[p-1]}({\lambda}_0,\ldots,{\lambda}_{p-2},{\lambda}_p)}{{\lambda}_{p-1}-{\lambda}_p}& \text{ if } {\lambda}_{p-1}\neq{\lambda}_p\\[2ex] \frac{\partial}{\partial t}\big\|_{t={\lambda}_{p-1}}f^{[p-1]}({\lambda}_0,\ldots,{\lambda}_{p-2},t)& \text{ if } {\lambda}_{p-1}={\lambda}_p. \end{cases}\end{aligned}$$ Denote $$\mathcal{W}_p=\big\{f:\, f^{(j)},\widehat{f^{(j)}}\in L^1({{\mathbb R}}), j=0,\ldots,p\big\}.$$ It is known (see, e.g., [@Azamov0 Lemma 2.3]) that for $f\in\mathcal{W}_p$, $$\begin{aligned} f^{[p]}({\lambda}_0,\ldots,{\lambda}_p)=\int_{\Pi^{(p)}}e^{{\text{\rm i}}(s_0-s_1){\lambda}_0}e^{{\text{\rm i}}(s_1-s_2){\lambda}_1}\dots e^{{\text{\rm i}}s_p {\lambda}_p}\,d\sigma_f^{(p)}(s_0,\dots,s_p),\end{aligned}$$ where $$\begin{aligned} &{\Pi^{(p)}}=\{(s_0,s_1,\dots,s_p)\in{{\mathbb R}}^{p+1}{\,:\,}\|s_p\|\leq\dots\leq\|s_1\|\leq\|s_0\|, \text{ \rm sign}(s_0)=\dots=\text{\rm sign}(s_p)\},\\ &d\sigma_f^{(p)}(s_0,s_1,\dots,s_p)={\text{\rm i}}^p\hat f(s_0)\,ds_0\,ds_1\dots ds_p.\end{aligned}$$ Let $H_0=H_0^,\ldots,H_p=H_p^$ be defined in $\mathcal{H}$ and let $V_1,\ldots,V_p\in{\mathcal{B}(\mathcal{H})}$. If $f\in \mathcal{W}_p$, then the Bochner integral $$\begin{aligned} \label{trans} T_{f^{[p]}}^{H_0,\ldots,H_p}(V_1,\ldots,V_p)y:=\int_{\Pi^{(p)}} e^{{\text{\rm i}}(s_0-s_1)H_0}V_1 e^{{\text{\rm i}}(s_1-s_2)H_1}V_2\ldots V_p\, e^{{\text{\rm i}}s_p H_p}y\,d\sigma_f^{(p)}(s_0,\dots,s_p)\end{aligned}$$ exists for every $y\in\mathcal{H}$ and thus defined operator has the norm bound $$\begin{aligned} \label{nest} \big\\|T_{f^{[p]}}^{H_0,\ldots,H_p}(V_1,\ldots,V_p)\big\\|\leq \frac{1}{p!}\cdot\big\\|\widehat{f^{(p)}}\big\\|_1\cdot\\|V_1\\|\cdot\ldots\cdot\\|V_p\\|\end{aligned}$$ (see [@Azamov0 Lemma 4.5]), which follows from the bound for the total variation of the measure $\big\\|\sigma_f^{(p)}\big\\|\leq \frac{1}{p!}\big\\|\widehat{f^{(p)}}\big\\|_1$. Similarly to [@Azamov0 Theorem 5.7], we have the following differentiation formula for an operator function. \[dermoi\] Let $H$ be a self-adjoint (unbounded) operator in $\mathcal{H}$, $V=V^\in{\mathcal{B}(\mathcal{H})}$, $p\in{{\mathbb N}}$, and $f\in \mathcal{W}_p$. Then, $$\begin{aligned} \frac{1}{p!}\cdot\frac{d^p}{dt^p}\bigg\|_{t=0} f(H+tV)=T_{f^{[p]}}^{H,\ldots,H}\underbrace{(V,\ldots,V)}_{p \text{ \rm times}}.\end{aligned}$$ We will work with the subspace $G_p$ of $C^{p+1}({{\mathbb R}})$, $p\in{{\mathbb N}}$, consisting of functions $f$ such that $f^{(p)}, f^{(p+1)}\in L^2({{\mathbb R}})$. Let $\\|\cdot\\|_{G_p}$ denote the semi-norm $$\\|f\\|_{G_p}=\frac{\sqrt{2}}{p!}\big(\big\\|f^{(p)}\big\\|_2+\big\\|f^{(p+1)}\big\\|_2\big).$$ It is known that $\frac{1}{p!}\big\\|\widehat{f^{(p)}}\big\\|_1\leq\\|f\\|_{G_p}$ (see, e.g., [@PS-Crelle Lemma 7]). In particular, we have $C_c^{p+1}({{\mathbb R}})\subset \cap_{j=0}^p G_j\subset \mathcal{W}_p$. Since all the scalar functions we consider are defined on ${{\mathbb R}}$, we will use the shortcut $C_c^{p+1}:=C_c^{p+1}({{\mathbb R}})$. We will need the following version of the well known integral representation for the remainder of the Taylor approximation. \[integral\] If $f\in \cap_{j=0}^p G_j$, $H_0=H_0^$ is defined in $\mathcal{H}$, and $V=V^\in{\mathcal{B}(\mathcal{H})}$, then $$\begin{aligned} \label{eq:integral} \mathcal{R}_{H_0,f,p}(V)=\frac{1}{(p-1)!}\int_0^1 (1-t)^{p-1}\frac{d^p}{ds^p}\bigg\|_{s=t}f(H_0+sV)\,dt,\end{aligned}$$ where the integral is defined for every $y\in{\mathcal{B}(\mathcal{H})}$ by $$\begin{aligned} \left(\int_0^1 (1-t)^{p-1}\frac{d^p}{ds^p}\bigg\|_{s=t}f(H_0+sV)\,dt\right) y =\int_0^1 (1-t)^{p-1}\frac{d^p}{ds^p}\bigg\|_{s=t}f(H_0+sV)\,y\,dt.\end{aligned}$$ By Theorem \[dermoi\], the definition , the continuity of the function $t\mapsto e^{{\text{\rm i}}s(H+tV)}$, $s\in{{\mathbb R}}$, in the strong operator topology, Theorem \[moib\], and the Lebesgue dominated convergence for Bochner integrals, the function $t\mapsto \frac{d^n}{ds^n}\big\|_{s=t}f(H_0+sV)$ is continuous in the strong operator topology. (More details in a slightly modified setting can be found in the proof of Theorem \[asexp\] below.) For any $y,g\in\mathcal{H}$, define $\phi_{y,g}\in({\mathcal{B}(\mathcal{H})})^$ by $\phi_{y,g}(A)=\left<Ay,g\right>$, for all $A\in{\mathcal{B}(\mathcal{H})}$. The function $$t\mapsto \phi_{y,g}\left(\frac{d^n}{ds^n}\bigg\|_{s=t}f(H_0+sV)\right)=\frac{d^n}{ds^n}\bigg\|_{s=t}\phi_{y,g}\big(f(H_0+sV)\big)$$ is continuous. By the fundamental theorem of calculus and integration by parts, $$\begin{aligned} \left<\mathcal{R}_{H_0,f,n}y,g\right>&=\frac{1}{(p-1)!}\int_0^1 (1-t)^{p-1}\frac{d^p}{ds^p}\bigg\|_{s=t}\phi_{y,g}\big(f(H_0+sV)\big)\,dt\\ &=\left<\frac{1}{(p-1)!}\int_0^1 (1-t)^{p-1}\frac{d^p}{ds^p}\bigg\|_{s=t}f(H_0+sV)y\,dt,g\right>,\quad \text{for all } y,g\in\mathcal{H},\end{aligned}$$ completing the proof. The inequality has analogs for Schatten norms, as it is stated in of the theorem below. \[moib\] Let $H=H^$ be defined in $\mathcal{H}$, $p\in{{\mathbb N}}$, and $f\in \cap_{k=1}^p G_k$. Let $\alpha,\alpha_1,\ldots,\alpha_p\in [1,\infty]$ be such that $\frac{1}{\alpha}=\frac{1}{\alpha_1}+\ldots +\frac{1}{\alpha_p}$ and let $V_j\in {\mathcal{S}}^{\alpha_j}$, $j=1,\ldots,p$. Then,[^2] $$\begin{aligned} \label{Spest} \big\\|T_{f^{[p]}}^{H,\ldots,H}(V_1,\ldots,V_p)\big\\|_\alpha\leq \\|f\\|_{G_p}\cdot\\|V_1\\|_{\alpha_1}\cdot\ldots\cdot\\|V_p\\|_{\alpha_p}.\end{aligned}$$ In the particular case of $p=1$ and $V$ a Hilbert-Schmidt perturbation, we have a stronger estimate, which holds for a more general $T_{f^{[1]}}^{H,H}(V)$ than the one defined above. Let $\phi$ be a bounded continuous function on ${{\mathbb R}}^2$ and let $$\begin{aligned} \label{TR} &\hat T_\phi^{H,H}(V)\\ \nonumber &:=\\|\cdot\\|_2\,\text{-}\!\lim_{m\rightarrow\infty}\;\lim_{N\rightarrow\infty}\sum_{\|l_0\|,\|l_1\|\leq N} \phi\left(\frac{l_0}{m},\frac{l_1}{m}\right)E_H\left(\left[\frac{l_0}{m},\frac{l_0+1}{m}\right)\right) VE_H\left(\left[\frac{l_1}{m},\frac{l_1+1}{m}\right)\right).\end{aligned}$$ The iterated limit above exists and defines a bounded operator on ${\mathcal{S}}^2$ (with the bound as in the theorem below). The proof can be found on pp. 5–6 of [@HS-compatible] or, for a slightly different construction and more general $\phi$, in [@BS]. \[HSest\] Let $H$ be a self-adjoint (unbounded) operator in $\mathcal{H}$ and let $V\in {\mathcal{S}}^2$. Then, for $\phi\in C_b({{\mathbb R}}^2)$, $$\big\\|\hat T_{\phi}^{H,H}(V)\big\\|_2\leq \\|\phi\\|_\infty\cdot\\|V\\|_2.$$ In particular, for $f\in C^1({{\mathbb R}})$, with $f'\in L^\infty({{\mathbb R}})$, $$\big\\|\hat T_{f^{[1]}}^{H,H}(V)\big\\|_2\leq \\|f'\\|_\infty\cdot\\|V\\|_2.$$ If $f\in\mathcal{W}_1$, then $\hat T_{f^{[1]}}^{H,H}(V)=T_{f^{[1]}}^{H,H}(V)$ by [@PSS Lemma 3.5]. It is easy to see that if $V$ is a trace-class operator on ${\mathcal{H}}$ and $E$ is a spectral measure (of a self-adjoint operator) acting on ${\mathcal{H}}$, then the measure ${{\mathrm{Tr}}}\big(E(\cdot)V\big)$ has finite total variation. It is also known (see, e.g., [@ds Section 4] for references and details) that for $V_1,\ldots,V_p \in {\mathcal{S}}^2$ and $E_1,\ldots,E_p$ spectral measures, with $p\geq 2$, the set function $$A_1\times\cdots\times A_p\mapsto {{\mathrm{Tr}}}\big(E_1(A_1)V_1\ldots E_p(A_p)V_p\big),$$ where $A_1,\ldots,A_p$ are Borel subsets of ${{\mathbb R}}$, uniquely extends to a measure on ${{\mathbb R}}^p$ of finite total variation. These observations are core for the following useful representations for operator derivatives. \[moimeasure\] Let $H$ be a self-adjoint operator, $p\in{{\mathbb N}}$, and let $f\in \cap_{k=1}^p G_k$. (i) If $p=1$ and $V\in {\mathcal{S}}^1$, then $$\begin{aligned} {{\mathrm{Tr}}}\left(\frac{d}{dt}\bigg\|_{t=0}f(H+tV)\right)=\int_{{\mathbb R}}f'({\lambda})\,{{\mathrm{Tr}}}\big(E_H(d{\lambda})V\big).\end{aligned}$$ (ii) $($[@moissf Theorem 3.12]$)$ If $p\geq 2$ and $V\in {\mathcal{S}}^2$, then $$\begin{aligned} &{{\mathrm{Tr}}}\left(\frac{d^p}{dt^p}\bigg\|_{t=0}f(H+tV)\right)\\ &\quad=(p-1)!\int_{{{\mathbb R}}^p}(f')^{[p-1]}({\lambda}_1,\ldots,{\lambda}_p)\,{{\mathrm{Tr}}}\big(E_H(d{\lambda}_1)V\ldots E_H(d{\lambda}_p)V \big). $$ It was proved in [@PSS Lemma 3.5] that for $f\in\mathcal{W}_p$ and $V_j\in {\mathcal{S}}^{\alpha_j}$, $j=1,\ldots,p$, with $\alpha_1,\ldots,\alpha_p\in [1,\infty]$, the operator $T_{f^{[n]}}^{H,\ldots,H}(V_1,\ldots,V_p)$ given by coincides with the operator $$\begin{aligned} \label{TRmoi} &\hat T_\phi^{H,\ldots,H}(V_1,\ldots,V_p)\\ \nonumber &:=s\,\text{-}\!\lim_{m\rightarrow\infty}\;\\|\cdot\\|_\alpha\,\text{-}\!\lim_{N\rightarrow\infty}\sum_{\|l_0\|,\ldots,\|l_p\|\leq N} \phi\left(\frac{l_0}{m},\frac{l_1}{m},\ldots,\frac{l_p}{m}\right)E_{H,l_0,m}V_1E_{H,l_1,m}V_2\ldots V_p E_{H,l_p,m},\end{aligned}$$ where $\phi=f^{[p]}$, $E_{H,l_k,m}=E_H\left(\left[\frac{l_k}{m},\frac{l_k+1}{m}\right)\right)$, for $k=0,\ldots,p$, $\frac{1}{\alpha}=\frac{1}{\alpha_1}+\ldots+\frac{1}{\alpha_p}$, and $s$-$\lim$ denotes a limit in the strong operator topology on the tuples $(V_1,\ldots,V_p)\in {\mathcal{S}}^{\alpha_1} \times \ldots \times {\mathcal{S}}^{\alpha_p}$. When $p=0$, we will use the symbol $\hat T_{f^{[0]}}$ (or $T_{f^{[0]}}$) to refer to the operator $\phi(H)$. We need the following algebraic properties of $\hat T_\phi$, which can be derived straightforwardly from the definition . \[properties\] Let $H=H^$ be defined in $\mathcal{H}$. Let $\alpha,\alpha_1,\ldots,\alpha_p\in [1,\infty]$ be such that $\frac{1}{\alpha}=\frac{1}{\alpha_1}+\ldots +\frac{1}{\alpha_p}$ and let $V_j\in {\mathcal{S}}^{\alpha_j}$, $j=1,\ldots,p$. (i) \[property1\] Let $\phi_1$ and $\phi_2$ be bounded Borel functions on ${{\mathbb R}}^p$. If the polylinear operators $\hat T_{\phi_1}^{H,\ldots,H},\hat T_{\phi_2}^{H,\ldots,H}:{\mathcal{S}}^{\alpha_1} \times \ldots \times {\mathcal{S}}^{\alpha_p}\mapsto{\mathcal{S}}^\alpha$ exist and are bounded, then $\hat T_{\phi_1+\phi_2}^{H,\ldots,H}:{\mathcal{S}}^{\alpha_1} \times \ldots \times {\mathcal{S}}^{\alpha_p}\mapsto {\mathcal{S}}^\alpha$ is also bounded and $$\hat T_{\phi_1+\phi_2}^{H,\ldots,H}=\hat T_{\phi_1}^{H,\ldots,H}+\hat T_{\phi_2}^{H,\ldots,H}.$$ (ii) $($[@PSS Lemma 3.2(iii)]$)$ \[property2\] Let $\phi_1: {{\mathbb R}}^{k + 1} \mapsto {{\mathbb C}}$ and $\phi_2 : {{\mathbb R}}^{p - k + 1} \mapsto {{\mathbb C}}$ be bounded Borel functions such that the operators $\hat T_{\phi_1}^{H,\ldots,H}$ and $\hat T_{\phi_2}^{H,\ldots,H}$ exist and are bounded on ${\mathcal{S}}^{\alpha_1} \times \ldots \times {\mathcal{S}}^{\alpha_k}$ and ${\mathcal{S}}^{\alpha_{k + 1}} \times \ldots \times {\mathcal{S}}^{\alpha_p}$, respectively. If $$\phi(\lambda_0, \ldots, \lambda_p) := \phi_1 \left( \lambda_0, \ldots, \lambda_k \right) \cdot \phi_2 \left( \lambda_k, \ldots, \lambda_p \right),$$ then the operator $\hat T_{\phi}^{H,\ldots,H}:{\mathcal{S}}^{\alpha_1} \times \ldots \times {\mathcal{S}}^{\alpha_p}\mapsto {\mathcal{S}}^\alpha$ is bounded and $$\hat T_{\phi}^{H,\ldots,H} \left( V_1, \ldots, V_p \right) = \hat T_{\phi_1}^{H,\ldots,H} \left( V_1, \ldots, V_k \right) \cdot \hat T_{\phi_2}^{H,\ldots,H} \left( V_{k + 1}, \ldots, V_p \right).$$ (iii) $($[@HS-compatible Lemma 2.9]$)$ \[property3\] Let $\phi:{{\mathbb R}}^{p}\mapsto{{\mathbb C}}$ and $\psi_1,\psi_2:{{\mathbb R}}\mapsto{{\mathbb C}}$ be bounded Borel functions. Denote $$(\psi_1\phi\psi_2)({\lambda}_0,\ldots,{\lambda}_p):=\psi_1({\lambda}_0)\phi({\lambda}_0,\ldots,{\lambda}_p)\psi_2({\lambda}_p).$$ If $\hat T_\phi^{H,\ldots,H}:{\mathcal{S}}^{\alpha_1} \times \ldots \times {\mathcal{S}}^{\alpha_p}\mapsto{\mathcal{S}}^\alpha$ exists and is bounded, then the operator $\hat T_{\psi_1\phi\psi_2}^{H,\ldots,H}:{\mathcal{S}}^{\alpha_1} \times \ldots \times {\mathcal{S}}^{\alpha_p}\mapsto{\mathcal{S}}^\alpha$ is also bounded and $$\hat T_{\psi_1\phi\psi_2}^{H,\ldots,H}(V_1,\ldots,V_p)=\hat T_\phi^{H,\ldots,H}(\psi_1(H)V_1,\ldots,V_p\psi_2(H)).$$ Asymptotic expansions ===================== In this section we prove the Taylor asymptotic expansion for $H_0$ having compact resolvent and find bounds for the remainder $R_{H_0,f,n}$.\ #### Compact resolvent We start with deriving estimates for the transformations , which will imply estimates for directional operator derivatives and the remainders $\mathcal{R}_{H_0,f,n}(V)$ defined in . \[constants\] Let $H=H^$ be defined in $\mathcal{H}$ and have compact resolvent and let $V=V^\in{\mathcal{B}(\mathcal{H})}$. Denote[^3] $j_n=1+\lfloor\log_2(n)\rfloor$. Then, for each function $0\leq f \in C_c^{n+1}$ with $f^{2^{-j_n}}\in C_c^{n+1}$, the transformation $T_{f^{[n]}}^{H,\ldots,H}$ is a bounded polylinear operator from ${\mathcal{B}(\mathcal{H})}\times\cdots\times {\mathcal{B}(\mathcal{H})}$ to ${\mathcal{S}}^1$ and $$\begin{aligned} \label{compactest} \left\\|T_{f^{[n]}}^{H,\ldots,H}(V,\ldots,V)\right\\|_1&\leq a_n\cdot \\|V\\|^n\cdot {{\mathrm{Tr}}}\big(E_H({\text{\rm supp}}f)\big)\\ \nonumber &\quad\times\max_{1\leq k\leq j_n}\big\\|f^{2^{-k}}\big\\|_\infty\cdot\left(\max_{\substack{1\leq k\leq j_n\\1\leq d\leq n}}\left\{1,\;\big\\|f^{2^{-k}}\big\\|_{G_d}\right\}\right)^n,\end{aligned}$$ where[^4] $$\begin{aligned} \label{an} a_1=2,\quad a_k=\begin{cases}a_{k-1}+a_{\frac{k}{2}} & \text{if } k \text{ is even}\\ a_{k-1}+a_{\frac{k-1}{2}} & \text{if } k\geq 3 \text{ is odd}.\end{cases}\end{aligned}$$ Note that by the Leibnitz formula for the divided difference, $$\begin{aligned} \label{obs0} f^{[n]}=\big(\sqrt{f}\cdot\sqrt{f}\big)^{[n]} &=\sum_{k=0}^{\lfloor\frac{n-1}{2}\rfloor} \sqrt{f}^{[k]}({\lambda}_0,\ldots,{\lambda}_k)\sqrt{f}^{[n-k]}({\lambda}_k,\ldots,{\lambda}_n)\\ \nonumber &\quad+\sum_{k=0}^{\lfloor\frac{n-1}{2}\rfloor}\sqrt{f}^{[n-k]}({\lambda}_0,\ldots,{\lambda}_{n-k})\sqrt{f}^{[k]}({\lambda}_{n-k},\ldots,{\lambda}_n)\\ \nonumber &\quad+\begin{cases} \sqrt{f}^{[\frac{n}{2}]}({\lambda}_0,\ldots,{\lambda}_{\frac{n}{2}})\sqrt{f}^{[\frac{n}{2}]}({\lambda}_{\frac{n}{2}},\ldots,{\lambda}_n) & \text{if } n \text{ is even}\\ 0 & \text{if } n \text{ is odd}. \end{cases}\end{aligned}$$ Hence, by Theorem \[properties\] (and the equality $\hat T_{f^{[n]}}=T_{f^{[n]}}$), we have $$\begin{aligned} \label{obsT} & T_{f^{[n]}}^{H,\dots,H}(V,\ldots,V)=\sum_{k=0}^{\lfloor\frac{n-1}{2}\rfloor} T_{\sqrt{f}^{[k]}}^{H,\ldots,H}(V,\ldots,V)\cdot T_{\sqrt{f}^{[n-k]}}^{H,\ldots,H}(V,\ldots,V)\\ \nonumber &\quad+\sum_{k=0}^{\lfloor\frac{n-1}{2}\rfloor}T_{\sqrt{f}^{[n-k]}}^{H,\ldots,H}(V,\ldots,V)\cdot T_{\sqrt{f}^{[k]}}^{H,\ldots,H}(V,\ldots,V)+\begin{cases} \left(T_{\sqrt{f}^{[\frac{n}{2}]}}^{H,\ldots,H}(V,\ldots,V)\right)^2 & \text{if } n \text{ is even}\\ 0 & \text{if } n \text{ is odd}. \end{cases}\end{aligned}$$ Recall that when $k=0$, the operator $T_{\sqrt{f}^{[k]}}^{H,\ldots,H}(V,\ldots,V)$ degenerates to the operator $\sqrt{f}(H)$. If $n=1$, then reduces to $$\begin{aligned} \label{T1} T_{f^{[1]}}^{H,H}(V)=\sqrt{f}(H)\cdot T_{\sqrt{f}^{[1]}}^{H,H}\big(V\big)+T_{\sqrt{f}^{[1]}}^{H,H}\big(V\big)\cdot\sqrt{f}(H).\end{aligned}$$ From Theorem \[moib\] and the straightforward inequality applied to $\sqrt{f}(H)$, we derive $$\begin{aligned} \label{T1n} \left\\|T_{f^{[1]}}^{H,H}(V)\right\\|_1\leq 2\cdot\big\\|\sqrt{f}\big\\|_\infty\cdot\big\\|\sqrt{f}\big\\|_{G_1} \cdot\\|V\\|\cdot{{\mathrm{Tr}}}\big(E_H({\text{\rm supp}}f)\big).\end{aligned}$$ If $n=2$, then reduces to $$\begin{aligned} \label{T2} T_{f^{[2]}}^{H,H,H}(V,V)&=\sqrt{f}(H)\cdot T_{\sqrt{f}^{[2]}}^{H,H,H}(V,V)+T_{\sqrt{f}^{[2]}}^{H,H,H}(V,V)\cdot \sqrt{f}(H)\\ \nonumber &\quad+T_{\sqrt{f}^{[1]}}^{H,H}(V)\cdot T_{\sqrt{f}^{[1]}}^{H,H}(V).\end{aligned}$$ Hence, $$\begin{aligned} \left\\|T_{f^{[2]}}^{H,H,H}(V,V)\right\\|_1\leq 2\,\big\\|\sqrt{f}(H)\big\\|_1\cdot \left\\|T_{\sqrt{f}^{[2]}}^{H,H,H}(V,V)\right\\|+\left\\|T_{\sqrt{f}^{[1]}}^{H,H}(V)\right\\|_1\cdot \left\\|T_{\sqrt{f}^{[1]}}^{H,H}(V)\right\\|.\end{aligned}$$ Applying, in addition, Theorem \[moib\] and the estimates and , we obtain $$\begin{aligned} \label{T2n} \left\\|T_{f^{[2]}}^{H,H,H}(V,V)\right\\|_1 &\leq 4\,\\|V\\|^2\cdot{{\mathrm{Tr}}}\big(E_H({\text{\rm supp}}f)\big)\cdot\max\left\{\big\\|\sqrt{f}\big\\|_\infty, \big\\|\sqrt[4]{f}\big\\|_\infty\right\}\\ \nonumber &\quad\times\left(\max\left\{1,\;\big\\|\sqrt{f}\big\\|_{G_1}, \big\\|\sqrt[4]{f}\big\\|_{G_1}, \big\\|\sqrt{f}\big\\|_{G_2}\right\}\right)^2.\end{aligned}$$ Application of Theorem \[properties\] and the decomposition gives $$\begin{aligned} \label{rec} \left\\|T_{f^{[n]}}^{H,\ldots,H}(V,\ldots,V)\right\\|_1 &\leq 2\sum_{k=0}^{\lfloor\frac{n-1}{2}\rfloor}\left\\|T_{\sqrt{f}^{[k]}}^{H,\ldots,H}(V,\ldots,V)\right\\|_1 \cdot\left\\|T_{\sqrt{f}^{[n-k]}}^{H,\ldots,H}(V,\ldots,V)\right\\|\\ \nonumber &\quad+\begin{cases} \left\\|T_{\sqrt{f}^{[\frac{n}{2}]}}^{H,\ldots,H}(V,\ldots,V)\right\\|_1 \cdot \left\\|T_{\sqrt{f}^{[\frac{n}{2}]}}^{H,\ldots,H}(V,\ldots,V)\right\\| & \text{if } n \text{ is even}\\ 0 & \text{if } n \text{ is odd}. \end{cases}\end{aligned}$$ where the involved transformations are bounded by Theorem \[moib\]. We will prove by induction on $n$ that the right hand side RHS of satisfies $$\begin{aligned} \label{indineq} \text{RHS} \leq a_n\cdot \\|V\\|^n\cdot {{\mathrm{Tr}}}\big(E_H({\text{\rm supp}}f)\big)\cdot \max_{1\leq k\leq j_n}\big\\|f^{2^{-k}}\big\\|_\infty\cdot\left(\max_{\substack{1\leq k\leq j_n\\1\leq d\leq n}}\left\{1,\;\big\\|f^{2^{-k}}\big\\|_{G_d}\right\}\right)^n.\end{aligned}$$ Suppose that the estimate is proved for $n-1$ (and for all $1\leq m\leq n-1$). Then we have $$\begin{aligned} \label{obs1} &\left\\|T_{\sqrt{f}^{[p]}}^{H,\ldots,H}(V,\ldots,V)\right\\|_1\cdot \left\\|T_{\sqrt{f}^{[q]}}^{H,\ldots,H}(V,\ldots,V)\right\\|\\ \nonumber &\quad\leq a_p\cdot{{\mathrm{Tr}}}\big(E_H({\text{\rm supp}}f)\big)\cdot\\|V\\|^n\cdot\max_{1\leq k\leq j_n}\big\\|f^{2^{-k}}\big\\|_\infty\cdot\left(\max_{\substack{1\leq k\leq j_n\\1\leq d\leq n}}\left\{1,\;\big\\|f^{2^{-k}}\big\\|_{G_d}\right\}\right)^n,\end{aligned}$$ where $p=\frac{n-1}{2}$, $q=\frac{n+1}{2}$ if $n$ is odd and $p=q=\frac{n}{2}$ if $n$ is even. Similarly, we have the bound $$\begin{aligned} \label{obs2} &\left\\|T_{f^{[n]}}^{H,\ldots,H}(V,\ldots,V)-T_{\sqrt{f}^{[p]}}^{H,\ldots,H}(V,\ldots,V)\cdot T_{\sqrt{f}^{[q]}}^{H,\ldots,H}(V,\ldots,V)\right\\|_1\\ \nonumber &\quad \leq 2\sum_{k=0}^{\lfloor\frac{n-2}{2}\rfloor}\left\\|T_{\sqrt{f}^{[k]}}^{H,\ldots,H}(V,\ldots,V)\right\\|_1 \cdot\left\\|T_{\sqrt{f}^{[n-k]}}^{H,\ldots,H}(V,\ldots,V)\right\\|\\ \nonumber &\quad\leq a_{n-1}\cdot{{\mathrm{Tr}}}\big(E_H({\text{\rm supp}}f)\big)\cdot\\|V\\|^n\cdot \max_{1\leq k\leq j_n}\big\\|f^{2^{-k}}\big\\|_\infty\cdot\left(\max_{\substack{1\leq k\leq j_n\\1\leq d\leq n}}\left\{1,\;\big\\|f^{2^{-k}}\big\\|_{G_d}\right\}\right)^n.\end{aligned}$$ Combining and completes the proof of the estimate. The value of $j_n$ is defined as follows. We repeat recursively the decomposition until each summand in the sum representing $T_{f^{[n]}}^{H,\ldots,H}(V,\ldots,V)$ decomposes into a product of $f^{2^{-i}}(H)\in {\mathcal{S}}^1$ (see Lemma \[fH\]) and operators in the form $T_{(f^{2^{-l}})^{[m]}}^{H,\ldots,H}(V,\ldots,V)$, for some $1\leq i, l\leq j_n$ and $1\leq m\leq n$. We have derived in and that $j_1=1$ and $j_2=1+j_1$. By the analogous reasoning, $j_n=1+j_{\lfloor\frac{n}{2}\rfloor}=\ldots=r+j_{\lfloor\frac{n}{2^r}\rfloor}$. The recursive procedure stops when $\lfloor\frac{n}{2^r}\rfloor=1$. Hence, $j_n=1+\lfloor\log_2(n)\rfloor$. \[derbound\] Let $H_0=H_0^$ be defined in $\mathcal{H}$ and have compact resolvent and let $V=V^\in{\mathcal{B}(\mathcal{H})}$. Then, for each function $f \in C_c^{n+1}$, $$\begin{aligned} \label{O} \nonumber &\left\\|\frac{1}{n!}\cdot\frac{d^n}{ds^n}\bigg\|_{s=t}f(H_0+sV)\right\\|_1\leq C_{H_0,f,n}\cdot{{\mathrm{Tr}}}\big(E_{H_0+tV}({\text{\rm supp}}f)\big)\cdot\\|V\\|^n,\quad t\in [0,1],\\ &\big\|{{\mathrm{Tr}}}\left(\mathcal{R}_{H_0,f,n}(V)\right)\big\|\leq C_{H_0,f,n}\cdot\sup_{t\in [0,1]}{{\mathrm{Tr}}}\big(E_{H_0+tV}({\text{\rm supp}}f)\big)\cdot\\|V\\|^n.\end{aligned}$$ Decompose the function $f$ into $f=f_1-f_2$, where $0\leq f_i,f_i^{2^{-j_n}}\in C_c^{n+1}$ for $i=1,2$. From Theorem \[dermoi\] we have $$\begin{aligned} \frac{1}{n!}\cdot\frac{d^n}{ds^n}\bigg\|_{s=t}f(H_0+sV)= T_{f_1^{[n]}}^{H_0+tV,\ldots,H_0+tV}(V,\ldots,V)-T_{f_2^{[n]}}^{H_0+tV,\ldots,H_0+tV}(V,\ldots,V).\end{aligned}$$ By Lemma \[compres\], $H_0+tV$ has compact resolvent. Hence, for each $T_{f_i^{[n]}}$, $i=1,2$, we have the bound as in of Lemma \[constants\]. Note that the bound for $\mathcal{R}_{H_0,f,n}(V)$ would follow from the integral representation for the remainder $$\begin{aligned} \label{eq:integralTr} {{\mathrm{Tr}}}\big(\mathcal{R}_{H_0,f,n}(V)\big)=\frac{1}{(n-1)!}\int_0^1 (1-t)^{n-1}\,{{\mathrm{Tr}}}\left(\frac{d^n}{ds^n}\bigg\|_{s=t}f(H_0+sV)\right)dt\end{aligned}$$ and the estimate for the derivatives established above. By the argument given in the proof of Theorem \[integral\], the functions $$t\mapsto \frac{d^n}{ds^n}\bigg\|_{s=t}f(H_0+sV)\quad\text{and}\quad t\mapsto \left(\frac{d^n}{ds^n}\bigg\|_{s=t}f(H_0+sV)\right)^$$ are continuous in the strong operator topology. These functions are also uniformly ${\mathcal{S}}^1$-bounded; therefore, implies on the strength of [@Azamov0 Lemma 3.10]. \[remarks\] (i) \[remarksi\] If $f\geq 0$ and $f^{2^{-j_n}}\in C_c^{n+1}$, then $$\begin{aligned} C_{H_0,f,n}&\leq a_n\cdot\max_{1\leq k\leq j_n}\big\\|f^{2^{-k}}\big\\|_\infty\cdot\left(\max_{\substack{1\leq k\leq j_n\\1\leq d\leq n}}\left\{1,\;\big\\|f^{2^{-k}}\big\\|_{G_d}\right\}\right)^n,\end{aligned}$$ where $a_n$ is given by . (ii) The case $n=1$ was handled in [@ACS] and it inspired decomposition of $f$ into positive and negative parts and use of dyadic roots of $f$ in the proof of Lemma \[constants\]. It was established in [@ACS Theorem 1.23] that the function $f: H\in H_0+({\mathcal{B}(\mathcal{H})})_{sa}\mapsto f(H)$, with $f\in C_c^3$, is Fréchet differentiable and the derivative is continuous in the trace norm. The $n$th order Fréchet differentiability in the trace norm can also be established, provided we take $f\in C_c^{n+2}$. \[asexp\] Let $H_0=H_0^$ be defined in $\mathcal{H}$ and have compact resolvent and $V=V^\in{\mathcal{B}(\mathcal{H})}$. Then, for each function $f \in C_c^{n+1}$, $$\begin{aligned} \label{eq:asexp} &{{\mathrm{Tr}}}(f(H_0+V))\\ \nonumber &\quad={{\mathrm{Tr}}}(f(H_0))+\sum_{k=1}^{n-1}\;\frac1k\sum_{{\lambda}_1,\ldots,{\lambda}_k\in\text{\rm spec}(H_0)} (f')^{[k-1]}({\lambda}_1,\ldots,{\lambda}_k)\, {{\mathrm{Tr}}}\big(E_{H_0}({\lambda}_1)V\ldots E_{H_0}({\lambda}_k)V\big)\\ \nonumber &\quad\quad+{{\mathrm{Tr}}}\big(\mathcal{R}_{H_0,f,n}(V)\big),\end{aligned}$$ with $${{\mathrm{Tr}}}\big(\mathcal{R}_{H_0,f,n}(V)\big)=\mathcal{O}\big(\\|V\\|^n\big)$$ satisfying . By Theorem \[dermoi\], $$f(H_0+V)=f(H_0)+\sum_{k=1}^{n-1}T_{f^{[k]}}^{H_0,\ldots,H_0}(V,\ldots,V) +\mathcal{R}_{H_0,f,n}(V),$$ where each summand is in ${\mathcal{S}}^1$ by Theorem \[derbound\]. Hence, $${{\mathrm{Tr}}}(f(H_0+V))={{\mathrm{Tr}}}(f(H_0))+\sum_{k=1}^{n-1}{{\mathrm{Tr}}}\big(T_{f^{[k]}}^{H_0,\ldots,H_0}(V,\ldots,V)\big) +{{\mathrm{Tr}}}\big(\mathcal{R}_{H_0,f,n}(V)\big).$$ The bound for the remainder is provided by Theorem \[derbound\], so we are left to prove the representation $$\begin{aligned} \label{obs3} \nonumber &{{\mathrm{Tr}}}\big(T_{f^{[k]}}^{H_0,\ldots,H_0}(V,\ldots,V)\big)\\ &\quad=\frac1k\sum_{{\lambda}_1,\ldots,{\lambda}_k\in\text{\rm spec}(H_0)} (f')^{[k-1]}({\lambda}_1,\ldots,{\lambda}_k)\, {{\mathrm{Tr}}}\big(E_{H_0}({\lambda}_1)V\ldots E_{H_0}({\lambda}_k)V\big),\end{aligned}$$ for any $k=1,\ldots,n-1$. Let $E_m:=E_{H_0}([-m,m])$. Clearly, $E_m$ converges to the identity in the strong operator topology and, by Corollary \[lemma:Efinite\], $V_m:=E_m V E_m\in {\mathcal{S}}^1$. Theorem \[moimeasure\] implies $$\begin{aligned} \label{obs4} \nonumber &{{\mathrm{Tr}}}\big(T_{f^{[k]}}^{H_0,\ldots,H_0}(V_m,\ldots,V_m)\big)\\ &\quad=\frac1k\sum_{\substack{{\lambda}_1,\ldots,{\lambda}_k\in\text{\rm spec}(H_0)\\ \|{\lambda}_1\|,\ldots,\|{\lambda}_k\| \leq m}} (f')^{[k-1]}({\lambda}_1,\ldots,{\lambda}_k)\, {{\mathrm{Tr}}}\big(E_{H_0}({\lambda}_1)V\ldots E_{H_0}({\lambda}_k)V\big).\end{aligned}$$ As it was noted in the proof of Lemma \[constants\], $T_{f^{[k]}}^{H_0,\ldots,H_0}(V,\ldots,V)$ is decomposable into a sum where each summand is a product of $f^{2^{-i}}(H_0)\in {\mathcal{S}}^1$ and operators in the form $T_{(f^{2^{-l}})^{[p]}}^{H_0,\ldots,H_0}(V,\ldots,V)$, for some $1\leq i, l\leq j_n$ and $1\leq p\leq k$. We also have the completely analogous decomposition for $T_{f^{[k]}}^{H_0,\ldots,H_0}(V_m,\ldots,V_m)$. Firstly, we verify that $T_{(f^{2^{-l}})^{[p]}}^{H_0,\ldots,H_0}(V_m,\ldots,V_m)$ converges to $T_{(f^{2^{-l}})^{[p]}}^{H_0,\ldots,H_0}(V,\ldots,V)$ in the strong operator topology as $m\rightarrow\infty$ by the Lebesgue dominated convergence theorem for Bochner integrals. Indeed, define $$\begin{aligned} &h(\omega)=e^{{\text{\rm i}}(s_0-s_1)H}V e^{{\text{\rm i}}(s_1-s_2)H}V\ldots V e^{{\text{\rm i}}s_p H},\\ & h_m(\omega)=e^{{\text{\rm i}}(s_0-s_1)H}V_m e^{{\text{\rm i}}(s_1-s_2)H}V_m\ldots V_m e^{{\text{\rm i}}s_p H}.\end{aligned}$$ Then we have convergence of the integrands in $$\lim_{m\rightarrow\infty}h_m(\omega) y=h(\omega)y,\quad \text{for every }\omega=(s_0,\dots,s_p)\in\Omega,$$ and we also have $$\sup_m\\|h_m(\cdot)\\|\in L^1\big(\Omega,\sigma_f^{(p)}\big),$$ which implies $$\begin{aligned} \label{lprop} \lim_{m\rightarrow\infty}T_{(f^{2^{-l}})^{[p]}}^{H_0,\ldots,H_0}(V_m,\ldots,V_m)y =T_{(f^{2^{-l}})^{[p]}}^{H_0,\ldots,H_0}(V,\ldots,V)y,\quad y\in\mathcal{H}.\end{aligned}$$ Since we have uniform boundedness $$\sup_m\big\\|T_{(f^{2^{-l}})^{[p]}}^{H_0,\ldots,H_0}(V_m,\ldots,V_m)\big\\|\leq \big\\|f^{2^{-l}}\big\\|_{G^p}\\|V\\|^p$$ (see Theorem \[moib\]), the convergence in along with $f^{2^{-i}}(H_0)\in {\mathcal{S}}^1$ implies that $$\lim_{m\rightarrow\infty}{{\mathrm{Tr}}}\big(T_{f^{[k]}}^{H_0,\ldots,H_0}(V_m,\ldots,V_m)\big)= {{\mathrm{Tr}}}\big(T_{f^{[k]}}^{H_0,\ldots,H_0}(V,\ldots,V)\big)$$ (see, e.g., [@Azamov0 Lemma 2.5]), which also implies convergence of the sequence on the right hand side of to the expression on the right hand side of . Thus, is proved. Since can be written as , we have the following consequence of Theorem \[asexp\]. \[asexpev\] Let $H_0=H_0^$ be defined in $\mathcal{H}$ and have compact resolvent and let $V=V^\in{\mathcal{B}(\mathcal{H})}$. Let $\{\mu_k\}_{k=1}^\infty$ be a sequence of eigenvalues of $H_0$ counting multiplicity and let $\{\psi_k\}_{k=1}^\infty$ be an orthonormal basis of the respective eigenvectors. Then, for each function $f \in C_c^{n+1}$, $$\begin{aligned} &{{\mathrm{Tr}}}\big(f(H_0+V)\big)-{{\mathrm{Tr}}}\big(f(H_0)\big)\\ \nonumber &\quad=\sum_{p=1}^{n-1} \frac1p\sum_{i_1,\ldots,i_p}(f')^{[p-1]}(\mu_{i_1},\ldots,\mu_{i_p})\, \left<V\psi_{i_1},\psi_{i_2}\right>\cdots\left<V\psi_{i_p},\psi_{i_1}\right>+{{\mathrm{Tr}}}\big(\mathcal{R}_{H_0,f,n}(V)\big),\end{aligned}$$ with $${{\mathrm{Tr}}}\big(\mathcal{R}_{H_0,f,n}(V)\big)=\mathcal{O}\big(\\|V\\|^n\big)$$ satisfying . #### Hilbert-Schmidt resolvent* Under the assumption $(1+H_0^2)^{-1/2}\in {\mathcal{S}}^2$, in Theorem \[asexpL2\], we improve the bound for the remainder obtained in Corollary \[asexpev\] by eliminating $\sup\limits_{t\in [0,1]}{{\mathrm{Tr}}}\big(E_{H_0+tV}({\text{\rm supp}}f)\big)$ and, consequently, eliminating $\sup\limits_{t\in [0,1]}\max\limits_{s\in{\text{\rm supp}}f}(1+\|s\|^2)$ (see connection between these expressions in ). \[CL2\] Let $H=H^$ satisfy $(1+H^2)^{-1/2}\in {\mathcal{S}}^2$ and let $V=V^$ be bounded. Denote $u(t)=(1+t^2)^{1/2}$. Then, for every $n\in{{\mathbb N}}$ and $f\in C_c^{n+1}$, the transformation $T_{f^{[n]}}^{H,\ldots,H}$ is a bounded polylinear mapping from ${\mathcal{B}(\mathcal{H})}\times\cdots\times {\mathcal{B}(\mathcal{H})}$ to ${\mathcal{S}}^1$ and $$\left\\|T_{f^{[n]}}^{H,\ldots,H}(V,\ldots,V)\right\\|_1\leq c_{f,n}\cdot\big\\|(1+H^2)^{-\frac12}\big\\|_2^2\cdot\\|V\\|^n,$$ where $$\begin{aligned} \label{eq:CL2} c_{f,n}\leq \begin{cases} \\|fu^2\\|_{G_1}+2\,\\|fu^2\\|_\infty\quad\text{\rm if } n=1\\ \\|fu^2\\|_{G_n}+\frac{n(n+3)}{2}\max\limits_{1\leq k\leq n}\big\{\\|f\\|_\infty,\\|fu\\|_\infty, \\|f\\|_{G_k},\\|fu\\|_{G_k}\big\}\cdot\max\limits_{2\leq l\leq n}\\|u\\|_{G_l}^2\quad\text{\rm if } n\geq 2. \end{cases}\end{aligned}$$ We need the following routine lemma. \[longL\] Let $f,u\in C^n$. Then, $$\begin{aligned} &u({\lambda}_0)\,f^{[n]}({\lambda}_0,\ldots,{\lambda}_n)\,u({\lambda}_n)\\ &\quad=(fu^2)^{[n]}({\lambda}_0,\ldots,{\lambda}_n)-\psi_1({\lambda}_0,\ldots,{\lambda}_n) -\psi_2({\lambda}_0,\ldots,{\lambda}_n)-\psi_3({\lambda}_0,\ldots,{\lambda}_n),\end{aligned}$$ where $$\begin{aligned} \psi_1({\lambda}_0,\ldots,{\lambda}_n)&=\sum_{k=1}^n (fu)^{[n-k]}({\lambda}_0,\ldots,{\lambda}_{n-k})\,u^{[k]}({\lambda}_{n-k},\ldots,{\lambda}_n),\end{aligned}$$ $$\begin{aligned} \psi_2({\lambda}_0,\ldots,{\lambda}_n)&=\sum_{k=1}^n u^{[k]}({\lambda}_0,\ldots,{\lambda}_k)\,(fu)^{[n-k]}({\lambda}_k,\ldots,{\lambda}_n),\end{aligned}$$ $$\begin{aligned} \psi_3({\lambda}_0,\ldots,{\lambda}_n)&=\sum_{k=1}^{n-1} u^{[k]}({\lambda}_0,\ldots,{\lambda}_k)\sum_{j=1}^{n-k}\,f^{[n-k-j]}({\lambda}_k,\ldots,{\lambda}_{n-j})\,u^{[j]}({\lambda}_{n-j},\ldots,{\lambda}_n).\end{aligned}$$ By the Leibnitz formula for the divided difference, $$\begin{aligned} &u({\lambda}_0)\,f^{[n]}({\lambda}_0,\ldots,{\lambda}_n)\,u({\lambda}_n)\\ &=(uf)^{[n]}({\lambda}_0,\ldots,{\lambda}_n)\,u({\lambda}_n)-\sum_{k=1}^n u^{[k]}({\lambda}_0,\ldots,{\lambda}_k)\,f^{[n-k]}({\lambda}_k,\ldots,{\lambda}_n)\,u({\lambda}_n),\end{aligned}$$ and applying the Leibnitz formula one more time completes the proof. It is easy to see that $fu^2,fu,f\in G_n$, for any natural $n$, and $u\in G_k$, for any $k\geq 2$. Note also that $\\|u'\\|_\infty\leq 1$. Denote $\widetilde V:=(1+H^2)^{-1/2}V(1+H^2)^{-1/2}\in {\mathcal{S}}^1$. For brevity, we denote the function $({\lambda}_0,\ldots,{\lambda}_n)\mapsto u({\lambda}_0)\,f^{[n]}({\lambda}_0,\ldots,{\lambda}_n)\,u({\lambda}_n)$ by $uf^{[n]}u$. In case $n=1$, Lemma \[longL\] and Theorem \[properties\], along with the equality $\hat T_{f^{[1]}}=T_{f^{[1]}}$, ensure the decomposition $$\begin{aligned} T_{f^{[1]}}^{H,H}(V)&=\hat T_{u f^{[1]}u}^{H,H}(\widetilde V)\\ &=T_{(fu^2)^{[1]}}^{H,H}(\widetilde V)-\big((fu)(H)\big)\cdot \hat T_{u^{[1]}}^{H,H}(\widetilde V) -\hat T_{u^{[1]}}^{H,H}(\widetilde V)\cdot\big((fu)(H)\big).\end{aligned}$$ Theorem \[HSest\] implies $$\left\\|\hat T_{u^{[1]}}(\widetilde V)\right\\|_2\leq \\|u'\\|_\infty\\|\widetilde V\\|_2\leq \\|\widetilde V\\|_2.$$ Applying also Theorem \[moib\] and Lemma \[fH\] gives $$\begin{aligned} \left\\|T_{f^{[1]}}^{H,H}(V)\right\\|_1 &\leq \\|fu^2\\|_{G_1}\cdot\big\\|(1+H^2)^{-1/2}V(1+H^2)^{-1/2}\big\\|_1\\ &\quad+2\\|fu^2\\|_\infty\cdot \big\\|(1+H^2)^{-1/2}\big\\|_2\cdot\big\\|(1+H^2)^{-1/2}V(1+H^2)^{-1/2}\big\\|_2\end{aligned}$$ Let now $n\geq 2$ and denote $W=(1+H^2)^{-1/2}V$. Since the operator $H$ is fixed, to lighten the notation, we omit the superscript when refer to the transformation $T_{f^{[n]}}(V,\ldots,V)$ and similar ones. Applying Lemma \[longL\] and Theorem \[properties\] leads to the decomposition $$\begin{gathered} \label{Tfn} T_{f^{[n]}}(V,\ldots,V)=T_{u f^{[n]}u}(W,V,\ldots,V,W^) =T_{(fu^2)^{[n]}}(W,V,\ldots,V,W^)\\-T_{\psi_1}(W,V,\ldots,V,W^) -T_{\psi_2}(W,V,\ldots,V,W^)-T_{\psi_3}(W,V,\ldots,V,W^),\end{gathered}$$ where $$\begin{aligned} &T_{\psi_1}(W,V,\ldots,V,W^)=-T_{(fu)^{[n-1]}}(W,V,\ldots,V)\cdot \hat T_{u^{[1]}}(W^)\\ &\quad-\sum_{k=2}^{n-1} T_{(fu)^{[n-k]}}(W,V,\ldots,V)\, T_{u^{[k]}}(V,\ldots,V,W^) -(fu)(H)\cdot T_{u^{[n]}}(W,\ldots,W^),\end{aligned}$$ $$\begin{aligned} &T_{\psi_2}(W,V,\ldots,V,W^)=-\hat T_{u^{[1]}}(W)\cdot T_{(fu)^{[n-1]}}(V,\ldots,V,W^)\\ &\quad\quad-\sum_{k=2}^{n-1} T_{u^{[k]}}(W,V,\ldots,V)\cdot T_{(fu)^{[n-k]}}(V,\ldots,V,W^) -T_{u^{[n]}}(W,\ldots,W^)\cdot(fu)(H),\end{aligned}$$ $$\begin{aligned} &T_{\psi_3}(W,V,\ldots,V,W^)=-\hat T_{u^{[1]}}(W)\cdot T_{f^{[n-2]}}(V,\ldots,V)\cdot \hat T_{u^{[1]}}(W^)\\ &\quad-\hat T_{u^{[1]}}(W)\cdot\sum_{j=2}^{n-1}T_{f^{[n-1-j]}}(V,\ldots,V)\cdot T_{u^{[j]}}(V,\ldots,V,W^)\\ &\quad-\sum_{k=2}^{n-1} T_{u^{[k]}}(W,V,\ldots,V)\cdot T_{f^{[n-1-k]}}(V,\ldots,V)\cdot \hat T_{u^{[1]}}(W^)\\ &\quad-\sum_{k=2}^{n-2} T_{u^{[k]}}(W,V,\ldots,V)\cdot\sum_{j=2}^{n-k}T_{f^{[n-k-j]}}(V,\ldots,V)\cdot T_{u^{[j]}}(V,\ldots,V,W^).\end{aligned}$$ Application of Theorems \[moib\] and \[HSest\] implies the bounds $$\begin{aligned} \label{e1} \left\\|T_{(fu^2)^{[n]}}(W,V,\ldots,V,W^)\right\\|_1\leq\\|fu^2\\|_{G_n}\cdot\\|W\\|_2^2 \cdot\\|V\\|^{n-2},\end{aligned}$$ $$\begin{aligned} \label{e2} \nonumber &\left\\|T_{\psi_i}(W,V,\ldots,V,W^)\right\\|_1\\ \nonumber &\quad\leq \bigg(\\|fu\\|_{G_{n-1}}+\sum_{k=2}^{n-1}\\|fu\\|_{G_{n-k}}\\|u\\|_{G_k}+\\|fu\\|_\infty\\|u\\|_{G_n}\bigg)\cdot\\|W\\|_2^2 \cdot\\|V\\|^{n-2}\\ &\quad\leq n\cdot\max_{1\leq k\leq n-1}\big\{\\|fu\\|_\infty,\\|fu\\|_{G_k}\big\}\cdot\max_{2\leq l\leq n}\\|u\\|_{G_l}\cdot \\|W\\|_2^2\cdot\\|V\\|^{n-2},\end{aligned}$$ for $i=1,2$, and $$\begin{gathered} \label{e3} \left\\|T_{\psi_3}(W,V,\ldots,V,W^)\right\\|_1 \leq \bigg(\\|f\\|_{G_{n-2}}+\sum_{j=2}^{n-1}\\|f\\|_{G_{n-1-j}}\\|u\\|_{G_j} +\sum_{k=2}^{n-1}\\|u\\|_{G_k}\\|f\\|_{G_{n-1-k}}\\ +\sum_{k=2}^{n-2}\\|u\\|_{G_k}\sum_{j=2}^{n-k}\\|f\\|_{G_{n-k-j}}\\|u\\|_{G_j}\bigg)\cdot\\|W\\|_2^2\cdot\\|V\\|^{n-2}\\ \leq \bigg(n-1+n-2+\sum_{k=2}^{n-2}(n-k-1)\bigg)\cdot\max_{1\leq k\leq n-2}\big\{\\|f\\|_\infty,\\|f\\|_{G_k}\big\}\cdot\max_{2\leq l\leq n-1}\\|u\\|_{G_l}^2\cdot \\|W\\|_2^2\cdot\\|V\\|^{n-2},\end{gathered}$$ where $\\|f\\|_{G_0}$ stands for $\\|f\\|_\infty$. Combining - gives . \[asexpL2\] Let $H_0=H_0^$ be defined in $\mathcal{H}$, let $V=V^\in{\mathcal{B}(\mathcal{H})}$, and suppose that $(1+H_0^2)^{-1/2}\in {\mathcal{S}}^2$. Let $\{\mu_k\}_{k=1}^\infty$ be a sequence of eigenvalues of $H_0$ counting multiplicity and let $\{\psi_k\}_{k=1}^\infty$ be an orthonormal basis of the respective eigenvectors. Then, for $n\in{{\mathbb N}}$ and $f\in C_c^{n+1}$, $$\begin{aligned} &{{\mathrm{Tr}}}\big(f(H_0+V)\big)-{{\mathrm{Tr}}}\big(f(H_0)\big)\\ \nonumber &\quad=\sum_{p=1}^{n-1} \frac1p\sum_{i_1,\ldots,i_p}(f')^{[p-1]}(\mu_{i_1},\ldots,\mu_{i_p})\, \left<V\psi_{i_1},\psi_{i_2}\right>\cdots\left<V\psi_{i_p},\psi_{i_1}\right>+{{\mathrm{Tr}}}\big(\mathcal{R}_{H_0,f,n}(V)\big)\end{aligned}$$ and $$\begin{aligned} \left\|{{\mathrm{Tr}}}\big(\mathcal{R}_{H_0,f,n}(V)\big)\right\|\leq c_{f,n}\cdot\big\\|(1+H_0^2)^{-1}\big\\|_1\cdot \left(1+\\|V\\|+\\|V\\|^2\right)\cdot\\|V\\|^n,\end{aligned}$$ where $c_{f,n}$ is as in . The result follows upon applying Lemma \[resest\] to $W=tV$, Lemma \[CL2\] to $H=H_0+tV$, $t\in[0,1]$, repeating the approximation argument in the proof of Theorem \[asexp\], and using the integral representation for the remainder as in the proof of Theorem \[derbound\]. We conclude with the discussion of the integral representations for $R_{H_0,V,1}(f)$ and $R_{H_0,V,2}(f)$. Let $C_c^3((a,b))$ denote the set of $C^3$-functions whose closed supports are compact subsets of $(a,b)$. $($[@ACS Theorem 2.5]$)$\[R1\] Let $H_0=H_0^$ have a compact resolvent and let $V=V^\in{\mathcal{B}(\mathcal{H})}$. Then, for $f\in C_c^3((a,b))$, $$\begin{aligned} {{\mathrm{Tr}}}\big(f(H_0+V)\big)={{\mathrm{Tr}}}\big(f(H_0)\big)+\int_{{\mathbb R}}f'({\lambda}){{\mathrm{Tr}}}\big(E_{H_0}((a,{\lambda}])-E_{H_0+V}((a,{\lambda}])\big)\,d{\lambda}.\end{aligned}$$ Applying the spectral theorem, Corollary \[lemma:Efinite\], and performing integration by parts gives $$\begin{aligned} &{{\mathrm{Tr}}}\big(f(H_0+V)\big)-{{\mathrm{Tr}}}\big(f(H_0)\big)\\ &\quad=\int_{{\mathbb R}}f({\lambda})\,d\,{{\mathrm{Tr}}}\big(E_{H_0+V}((a,{\lambda}])\big) -\int_{{\mathbb R}}f({\lambda})\,d\,{{\mathrm{Tr}}}\big(E_{H_0}((a,{\lambda}])\big)\\ &\quad=\int_{{\mathbb R}}f'({\lambda}){{\mathrm{Tr}}}\big(E_{H_0}((a,{\lambda}])-E_{H_0+V}((a,{\lambda}])\big)\,d{\lambda}.\end{aligned}$$ \[ac\] Let $H_0=H_0^$ satisfy $(1+H_0^2)^{-1/2}\in {\mathcal{S}}^2$ and let $V=V^\in{\mathcal{B}(\mathcal{H})}$. Denote $u(t)=(1+t^2)^{1/2}$. Then, there is a locally integrable function $\eta=\eta_{H_0,V}$ such that $$\begin{aligned} \label{trf} R_{H_0,f,2}(V)=\int_{{\mathbb R}}f''(t)\eta(t)\,dt,\quad\text{for }f\in C_c^3,\end{aligned}$$ and $$\int_{[a,b]}\|\eta(t)\|\,dt\leq C_{a,b}\cdot\\|(1+H_0^2)^{-1}\big\\|_1\cdot \left(1+\\|V\\|+\\|V\\|^2\right)\cdot\\|V\\|^2,$$ where $$\begin{aligned} C_{a,b}\leq 9\cdot\max\big\{1,(b-a)^2\big\}\cdot\max\big\{2,\\|u\\|_{L^\infty([a,b])},\\|u^2\\|_{L^\infty([a,b])}, \\|(u^2)'\\|_{L^\infty([a,b])}\big\}\end{aligned}$$ and $R_{H_0,f,2}(V)$ is given by . Let $H_t=H_0+tV$ and $W_t=(1+H_t^2)^{-1/2}V$, for $t\in [0,1]$. As a particular case of , we have $$\begin{aligned} T_{f^{[2]}}^{H_t,H_t,H_t}(V,V)&=T_{(fu^2)^{[2]}}^{H_t,H_t,H_t}(W_t,W_t^)\\ &\quad-\hat T_{(fu)^{[1]}}^{H_t,H_t}(W_t)\cdot \hat T_{u^{[1]}}^{H_t,H_t}(W_t^)-(fu)(H)\cdot T_{u^{[2]}}^{H_t,H_t,H_t}(W_t,W_t^)\\ &\quad-\hat T_{u^{[1]}}^{H_t,H_t}(W_t)\cdot \hat T_{(fu)^{[1]}}^{H_t,H_t}(W_t^)- T_{u^{[2]}}^{H_t,H_t,H_t}(W_t,W_t^)\cdot (fu)(H)\\ &\quad-\hat T_{u^{[1]}}^{H_t,H_t}(W_t)\cdot f(H)\cdot \hat T_{u^{[1]}}^{H_t,H_t}(W_t^).\end{aligned}$$ Therefore, with application of Theorem \[HSest\], Hölder’s inequality, and $\\|u'\\|_\infty\leq 1$ we have $$\begin{aligned} \label{Tf2} \nonumber \left\|{{\mathrm{Tr}}}\left(T_{f^{[2]}}^{H_t,H_t,H_t}(V,V)\right)\right\|&\leq \left\|{{\mathrm{Tr}}}\left(T_{(fu^2)^{[2]}}^{H_t,H_t,H_t}(W_t,W_t^)\right)\right\|\\ \nonumber &+2\cdot\\|(fu)'\\|_\infty\cdot\\|W_t\\|_2^2+2\cdot\\|fu\\|_\infty\cdot \left\|{{\mathrm{Tr}}}\left(T_{u^{[2]}}^{H_t,H_t,H_t}(W_t,W_t^)\right)\right\|\\ &+\\|f\\|_\infty\cdot\\|W_t\\|_2^2.\end{aligned}$$ One can derive from that $${{\mathrm{Tr}}}\left(T_{u^{[2]}}^{H_t,H_t,H_t}(W_t,W_t^)\right)={{\mathrm{Tr}}}\left(T_\phi^{H_t,H_t}(W_t)W_t^\right), \quad\text{with}\quad\phi({\lambda}_0,{\lambda}_1)=u^{[2]}({\lambda}_0,{\lambda}_0,{\lambda}_1)$$ (for more details, see, e.g., [@moissf Lemma 3.8]). Hence, by Hölder’s inequality, the equality $T_\phi=\hat T_\phi$ (see [@PSS Lemma 3.5]), and Theorem \[HSest\], $$\begin{aligned} \label{u2} \nonumber \left\|{{\mathrm{Tr}}}\left(T_{u^{[2]}}^{H_t,H_t,H_t}(W_t,W_t^)\right)\right\|&\leq \big\\|\hat T_\phi^{H_t,H_t}(W_t)\big\\|_2\cdot\\|W_t\\|_2 \leq \big\\|u^{[2]}\big\\|_\infty\cdot\\|W_t\\|_2^2\\&\leq\frac12\cdot\\|u''\\|_\infty\cdot\\|W_t\\|_2^2\leq\\|W_t\\|_2^2.\end{aligned}$$ Similarly, $$\begin{aligned} \label{fu2} \left\|{{\mathrm{Tr}}}\left(T_{(fu^2)^{[2]}}^{H_t,H_t,H_t}(W_t,W_t^)\right)\right\|\leq \frac12\cdot\\|(fu^2)''\\|_\infty\cdot\\|W_t\\|_2^2.\end{aligned}$$ Since $$\begin{aligned} \big\\|(fg)^{(k)}\big\\|_\infty=\left\\|\sum_{j=0}^k\begin{pmatrix}k\\j\end{pmatrix}f^{(j)}g^{(k-j)}\right\\|_\infty &\leq \sum_{j=0}^k \begin{pmatrix}k\\j\end{pmatrix}\big\\|f^{(j)}\big\\|_\infty\big\\|g^{(k-j)}\big\\|_\infty\\ &\leq 2^k\cdot\max_{0\leq j\leq k}\big\\|f^{(j)}\big\\|_\infty\cdot\max_{0\leq l\leq k}\big\\|g^{(l)}\big\|\big\\|_\infty,\end{aligned}$$ $$\big\\|f^{(j)}\big\\|_\infty\leq \big\\|f^{(n)}\big\\|_\infty\cdot (b-a)^{n-j},\quad 0\leq j\leq n,$$ and $$(u^2)''\equiv 2,$$ we have that for $f\in C_c^3((a,b))$, $$\begin{aligned} \label{fi} &\\|(fu)'\\|_\infty\leq 2\cdot\\|f''\\|_\infty\cdot\max\big\{1,(b-a)^2\big\}\cdot\\|u\\|_{L^\infty([a,b])},\\ \nonumber &\\|(fu^2)''\\|_\infty\leq 4\cdot\\|f''\\|_\infty\cdot\max\big\{1,(b-a)^2\big\} \cdot\max\big\{2,\\|u^2\\|_{L^\infty([a,b])},\\|(u^2)'\\|_{L^\infty([a,b])}\big\}.\end{aligned}$$ Combination of the inequalities - ensures the bound $$\begin{gathered} \label{of} \left\|{{\mathrm{Tr}}}\left(T_{f^{[2]}}^{H_t,H_t,H_t}(V,V)\right)\right\|\leq \\|f''\\|_\infty\cdot\\|W_t\\|_2^2\\ \times 9\cdot\max\big\{1,(b-a)^2\big\}\cdot\max\big\{2,\\|u\\|_{L^\infty([a,b])},\\|u^2\\|_{L^\infty([a,b])}, \\|(u^2)'\\|_{L^\infty([a,b])}\big\}.\end{gathered}$$ Applying , Lemma \[resest\], and Theorems \[integral\] and \[dermoi\] gives $$\begin{aligned} \label{RC} \big\|R_{H_0,f,2}(V)\big\|\leq C_{a,b}\cdot\big\\|f''\big\\|_\infty\cdot\big\\|(1+H_0^2)^{-1}\big\\|_1\cdot \left(1+\\|V\\|+\\|V\\|^2\right)\cdot\\|V\\|^2.\end{aligned}$$ Hence, by the Riesz representation theorem for a functional in $\big(C_c({{\mathbb R}})\big)^$, there is a locally finite measure $\nu=\nu_{H_0,V}$, with $$\int_{[a,b]}d\|\nu\|\leq C_{a,b}\cdot\\|(1+H_0^2)^{-1}\big\\|_1\cdot \left(1+\\|V\\|+\\|V\\|^2\right)\cdot\\|V\\|^2,$$ such that $$\begin{aligned} \label{Rtr} R_{H_0,f,2}(V)=\int_{{\mathbb R}}f''(t)\,d\nu(t),\quad\text{for }f\in C_c^3.\end{aligned}$$ We are left to prove absolute continuity of $\nu$. By adjusting the proof of [@HS-compatible Theorem 2.25], we derive the representation $$\begin{aligned} T_{f^{[1]}}^{H_0,H_0}(V)=\hat T_F^{H_0,H_0}\big((1+H_0^2)^{-1/2}V\big)\big((1+H_0^2)^{-1/2},\quad\text{for } f\in C_c^3((a,b)),\end{aligned}$$ where $$F({\lambda}_0,{\lambda}_1)=u({\lambda}_0)f^{[1]}({\lambda}_0,{\lambda}_1)u({\lambda}_1),\quad \\|F\\|_\infty\leq \widetilde C_{a,b}\cdot\\|f'\\|_\infty.$$ Hence, by Theorem \[dermoi\], Hölder’s inequality, and Theorem \[HSest\], $$\begin{aligned} \left\|{{\mathrm{Tr}}}\left(\frac{d}{dt}\bigg\|_{t=0}f(H_0+tV)\right)\right\|\leq \widetilde C_{a,b}\cdot\big\\|f'\big\\|_\infty\cdot\big\\|(1+H_0^2)^{-1}\big\\|_1\cdot\\|V\\|.\end{aligned}$$ Therefore, there exists a locally finite measure $\mu=\mu_{H_0,V}$ such that $$\begin{aligned} \label{Dtr} {{\mathrm{Tr}}}\left(\frac{d}{dt}\bigg\|_{t=0}f(H_0+tV)\right)=\int_{{\mathbb R}}f'(t)\,d\mu(t),\quad\text{for }f\in C_c^3.\end{aligned}$$ Let $$\begin{aligned} \label{xic} \xi({\lambda}):={{\mathrm{Tr}}}\big(E_{H_0}((a,{\lambda}])-E_{H_0+V}((a,{\lambda}])\big).\end{aligned}$$ By Theorem \[R1\] and by , we have $$\begin{aligned} R_{H_0,f,2}(V)&={{\mathrm{Tr}}}\big(f(H_0+V)\big)-{{\mathrm{Tr}}}\big(f(H_0)\big)-{{\mathrm{Tr}}}\left(\frac{d}{dt}\bigg\|_{t=0}f(H_0+tV)\right)\\ &=\int_{{\mathbb R}}f'({\lambda})\xi({\lambda})\,d{\lambda}-\int_{{\mathbb R}}f'({\lambda})\,d\mu({\lambda}).\end{aligned}$$ Integrating by parts yields $$\begin{aligned} R_{H_0,f,2}(V)=\int_{{\mathbb R}}f''({\lambda})\left(\mu((a,{\lambda}))-\int_a^{\lambda}\xi(t)\,dt\right)d{\lambda},\quad\text{for }f\in C_c^3((a,b)),\end{aligned}$$ completing the proof of the absolute continuity of $\nu$. Analogs of the function $\xi$ given be and the function $\eta$ given by have long history in perturbation theory. It was established in [@Krein], [@Kop], and [@PSS] for $n=1$, $n=2$, and $n\geq 3$, respectively, that there exists an integrable function $\eta_n=\eta_{n,H_0,V}$ such that $$\begin{aligned} {{\mathrm{Tr}}}\big(\mathcal{R}_{H_0,f,n}(V)\big)=\int_{{\mathbb R}}f^{(n)}(t)\eta_n(t)\,dt\end{aligned}$$ for sufficiently nice functions $f$ (including $f\in C_c^{n+1}$), provided $V=V^\in {\mathcal{S}}^n$ and $H_0=H_0^$ (without restrictions on the resolvent of $H_0$). If the resolvent of $H_0$ is compact and $V\in{\mathcal{B}(\mathcal{H})}$, then $\xi$ essentially coincides with the spectral flow (see [@ACS]). An analog of (with substantially modified left hand side) was obtained for $H_0=H_0^$ and $V=V^$ satisfying $(1+H_0^2)^{-1/2}V\in{\mathcal{S}}^2$ in [@HS-compatible Theorem 4.9]. [99]{} N. A. Azamov, A. L. Carey, P. G. Dodds, F. A. Sukochev, [Operator integrals, spectral shift, and spectral flow,]{} Canad. J. Math. [61]{} (2009), no. 2, 241–263. N. A. Azamov, A. L. Carey, F. A. Sukochev, [The spectral shift function and spectral flow,]{} Comm. Math. Phys. [276]{} (2007), no. 1, 51–91. M. Sh. Birman, M. Solomyak, [Double operator integrals in a Hilbert space,]{} Integral Equations Operator Theory [47]{} (2003), no. 2, 131-–168. J.-M. Bouclet, [Spectral distributions for long range perturbations,]{} J. Funct. Anal. [212]{} (2004), 431–471. A. L. Carey, J. Phillips, [Unbounded Fredholm modules and spectral flow,]{} Canad. J. Math. [50]{} (1998), 673–-718. A. H. Chamseddine, A. Connes, [The spectral action principle,]{} Comm. Math. Phys. [186]{} (1997), 731–750. Yu. L. Daleckii, S. G. Krein, [Integration and differentiation of functions of Hermitian operators and applications to the theory of perturbations,]{} (Russian) Voronež. Gos. Univ. Trudy Sem. Funkcional. Anal. [1956]{} (1956), no. 1, 81-–105. K. Dykema, A. Skripka, [Higher order spectral shift,]{} J. Funct. Anal. [257]{} (2009), 1092–1132. F. Hansen, [Trace functions as Laplace transforms,]{} J. Math. Phys. [47]{} (2006) 043504, 11 pp. M. G. Krein, On a trace formula in perturbation theory, Matem. Sbornik [33]{} (1953), 597–626 (Russian). L. S. Koplienko, Trace formula for perturbations of nonnuclear type, Sibirsk. Mat. Zh. [25]{} (1984), 62–71 (Russian). English transl. in Siberian Math. J. [25]{} (1984), 735–743. I. M. Lifshits, [On a problem of the theory of perturbations connected with quantum statistics,]{} Uspehi Matem. Nauk (N.S.), [7]{} (1952), no. 1 (47), 171–180. (Russian) D. Potapov, A. Skripka, F. Sukochev, [Spectral shift function of higher order,]{} Invent. Math., [193]{} (2013), no. 3, 501–538. D. Potapov, A. Skripka, F. Sukochev, [On Hilbert-Schmidt compatibility,]{} Oper. Matrices, [7]{} (2013), no. 1, 1–34. D. Potapov and F. Sukochev, Unbounded Fredholm modules and double operator integrals, J. reine. angew. Math. [626]{} (2009), 159–185. A. Skripka, [Multiple operator integrals and spectral shift,]{} Illinois J. Math., 55 (2011), no. 1, 305–324. W. D. van Suijlekom, [Perturbations and operator trace functions,]{} J. Funct. Anal. [260]{} (2011), no. 8, 2483–-2496. [^1]: Research supported in part by NSF grant DMS-1249186. [^2]: $\\|\cdot\\|_\infty$ denotes the operator norm. [^3]: As usually, $x\mapsto \lfloor x\rfloor$ denotes the floor function. [^4]: $\{a_n\}_{n=1}^\infty=\{2,4,6,10,14,20,26,36,46,60,74,94,114,140,\ldots\}$	{ "pile_set_name": "ArXiv" }
--- abstract: 'We present the results of an ongoing weak lensing survey conducted with the Subaru telescope whose initial goal is to locate and study the distribution of shear-selected structures or [halos]{}. Using a Suprime-cam imaging survey spanning 21.82 deg$^2$, we present a catalog of 100 candidate halos located from lensing convergence maps. Our sample is reliably drawn from that subset of our survey area, (totaling 16.72 deg$^2$) uncontaminated by bright stars and edge effects and limited at a convergence signal to noise ratio of 3.69. To validate the sample detailed spectroscopic measures have been made for 26 candidates using the Subaru multi-object spectrograph, FOCAS. All are confirmed as clusters of galaxies but two arise as the superposition of multiple clusters viewed along the line of sight. Including data available in the literature and an ongoing Keck spectroscopic campaign, a total of 41 halos now have reliable redshifts. For one of our survey fields, the XMM LSS [@pierre04] field, we compare our lensing-selected halo catalog with its X-ray equivalent. Of 15 halos detected in the XMM-LSS field, 10 match with published X-ray selected clusters and a further 2 are newly-detected and spectroscopically confirmed in this work. Although three halos have not yet been confirmed, the high success rate within the XMM-LSS field (12/15) confirms that weak lensing provides a reliable method for constructing cluster catalogs, irrespective of the nature of the constituent galaxies or the intracluster medium.' author: - \| Satoshi Miyazaki, Takashi Hamana, Richard S. Ellis,\ Nobunari Kashikawa, Richard J. Massey, James Taylor, Alexandre Refregier title: 'A Subaru Weak Lensing Survey I: Cluster Candidates and Spectroscopic Verification ' --- Introduction ============ Clusters of galaxies represent the most massive bound systems in the cosmos. Although they result from non-linear structure evolution, the departure from linear growth is modest compared to that for less massive objects. As a result, simple analytic models can provide an accurate indication of their expected number density at various redshifts. This is the primary reason why cluster of galaxies are considered to be valuable cosmological probes. Cosmological attention has focused on the redshift-dependent number of clusters, $N(z)$, whose mass exceeds a certain threshold. This is one of the most straightforward observables, and is a function of the cluster mass function, $d^2n/dMdz$, and the evolution of comoving volume $dV/d\Omega(z)$. The mass function is obtained from the growth rate of density fluctuations, $\delta(z)$, numerically [@jenkinsetal01] under the assumption of a particular theory of structure formation, e.g. the currently-popular cold dark matter (CDM) model. Since both $\delta(z)$ and $dV/d\Omega(z)$ are dependent on the cosmological model, useful constraints could be estimated by comparing $N(z)$ with various model predictions. To make progress, e.g. on the dark energy equation of state parameter $w$, data on several thousand clusters to $z\simeq$1 is thought to be required [@levineetal02; @wangetal04], and maintaining an accurate and uniform mass threshold is critical. Most early work focused on selecting clusters optically, with detection techniques that have improved over the decades: e.g. matched-filter [@postmanetal96], red-sequence [@gladdersetal00], cut-and-enhance [@gotoetal02]. Optically selected samples have traditionally suffered from uncertainties in the optical richness - mass relation, although there has been recent progress in calibrating the closely related richness-velocity dispersion relation using the large maxBCG sample of clusters identified in the Sloan Digital Sky Survey [@beckeretal07]. So far, X-ray samples have been the most popular cosmological probes e.g. [@bohringeretal01], [@ikebeetal02], [@reiprichetal02]. Luminosity ($L_X$) or temperature ($T_X$)-limited samples offer simpler selection functions because the observables, $L_X$, $T_X$, are a fair estimate of the cluster mass, calibrated through empirical scaling relations. The derived mass does depend, however, on the assumed dynamical state of the system. Unrelaxed clusters, arising for example from recent mergers, will introduce scatter in the scaling relation. A recent study by [@smithetal05] points out that at least half of 10 z $\sim$ 0.2 cluster cores show unrelaxed features and a scatter of $\sigma \sim 0.4$ around the mean scaling relation. Weak gravitational lensing, which analyses the coherent shear pattern of background galaxies, can potentially provide estimates of the cluster mass [regardless of its dynamical state or the properties of the constituent galaxies]{}. For some years, the method has been used to calibrate mass obtained from X-ray data [@smail97]. [@allenetal03] concluded that X-ray mass measurements are consistent with those from weak lensing, particularly for relaxed systems which thus offer a useful cosmological probe[@allenetal04]. A natural extension of this progress is thus to consider selecting clusters directly from weak lensing signals. The development of panoramic imaging surveys has now made this a practical proposition. [@wittmanetal01] reported the first discovery of a cluster located from a weak lensing convergence map, during the course of conducting the Deep Lens Survey [@wittmanetal02]. [@miyazakietal02a] later undertook a systematic search of mass concentrations on a 2 deg$^2$ field using the Suprime-Cam imager on Subaru. They detected several significant (S/N$>4$) candidates, one of which was later spectroscopically identified as a cluster at $z$ =0.41. [@hetterscheidtetal05] investigated 50 randomly-selected VLT FORS1 fields, spanning 0.64 deg$^2$ in total, and found 5 shear-selected candidates, each associated with an overdensity in luminosity. The first results from the Deep Lens Survey, based on an area of 8.6 deg$^2$ have also recently emerged [@wittmanetal06]. The above pioneering studies have demonstrated that clusters can be located directly via weak lensing. However, key issues, including the optimum selection threshold, the rate of spurious detection and the degree of mass completeness at a given redshift, crucial for any eventual cosmological application, remain unresolved. At the present time, theoretical studies offer the only insight into these issues. Projection is one of the most troublesome aspects of a weak lensing survey, because of the relatively broad window function. Unrelated structures contributing to the signal would lead to an overestimate of the cluster mass. Moreover, as the noise in the convergence map arises largely from shot noise in the ellipticity distribution of background galaxies, some fraction of genuine clusters might be missed in a shear selected catalog. N-body and ray-tracing simulations [@whiteetal02; @hamanaetal04; @hennawiandspergel05] have concluded that, for systems whose convergence signal lie above a 4 standard deviation ($\sigma$) threshold, 60-75% of clusters can reliably recovered (completeness). Likewise, for peaks detected in the simulated data using typically-used algorithms, 60-75 % represent genuine clusters (efficiency). The difference in these figures between the various studies is largely due to differences in the lower mass limit adopted in the studies. This series of papers is motivated by the need to address these key issues observationally. The survey we describe is a natural and ongoing extension of the 2 deg$^2$ survey of [@miyazakietal02a]; to date a total field of 21.82 deg$^2$ has been imaged. This first paper describes the imaging survey and discusses the validation of the candidates found, both via spectroscopic verification and comparison with clusters located via X-ray techniques. Later papers in the series will extend the spectroscopic follow-up to the full sample and will consider the feasibility of deriving cosmological constraints from both this survey and future enhanced versions. We note that a similary motivated program has been initiated by [@maturietal06] , [@schirmeretal06] and [@dietrichetal07], all of which made used of the imaging data taken by 2.5 m VLT survey telescope. We compare the their conclusions with our own in this paper. We note that a similarly motivated program has been initiated by [@maturietal06] and @schirmeretal06 whose conclusions we compare with our own in this paper. A plan of the paper follows. We discuss the imaging strategy and data analysis in §2, including construction of the cluster catalog and its reliability. In §3, we describe our initial spectroscopic follow-up with Subaru and address the completeness by comparing X-ray selected clusters on one of our survey area where relevant X-ray data is available. We summarize our conclusions in §4. Imaging Observations & Data Analysis ==================================== Survey Fields ------------- -------------- ------- ----- --------- ------------- ---------------- ----------------- --------- --------- --------- Field RA DEC Area Secure Area Seeing $\rho_{gal}$ $T_{R}$ $T_{C}$ $T_{N}$ deg$^2$ deg$^2$ arcsec arcmin$^2$ ksec ksec ksec DEEP02 02:30 00 1.39 0.73 $0.70\pm 0.06$ $33.5\pm 6.1 $ SXDS 02:18 –05 1.12 0.83 $0.68\pm 0.06$ $47.7 \pm 5.7 $ 100 XMM-LSS 02:26 –04 2.80 2.24 $0.55\pm 0.07$ $46.0 \pm 6.7 $ 10 Lynx 08:49 +45 1.76 1.28 $0.80\pm 0.08$ $30.7\pm 7.3 $ 64 300 140 COSMOS 10:02 +01 1.92 1.41 $0.54\pm 0.03$ $37.1\pm 2.1$ 30 Lockman Hole 10:52 +57 1.85 1.57 $0.60\pm 0.14$ $39.3\pm 7.8 $ 200 300 100 GD140 11:36 +30 1.83 1.50 $0.71\pm 0.17$ $29.3\pm 12.9 $ 33 PG1159-035 12:04 –04 1.43 1.19 $0.75\pm 0.05$ $23.4\pm 3.6 $ 51 13 hr Field 13:34 +38 2.06 1.72 $0.74\pm 0.17$ $29.6\pm 9.6 $ 110 120 130 GTO2deg$^2$ 16:04 +43 2.01 1.53 $0.67\pm 0.04$ $38.0\pm 3.6 $ 26 CM DRA 16:34 +57 1.38 0.99 $0.72\pm 0.12$ $28.4\pm 8.4 $ 47 DEEP16 16:52 +36 1.20 0.93 $0.76\pm 0.08$ $26.4\pm 4.0 $ DEEP23 23:30 00 1.07 0.80 $0.58\pm 0.01$ $36.3\pm 1.3 $ Total 21.82 16.72 -------------- ------- ----- --------- ------------- ---------------- ----------------- --------- --------- --------- In order to evaluate the efficiency of our weak lensing survey for locating cluster halos, we considered that a comparison with a sample of X-ray selected clusters would be highly advantageous [@henry00]. Therefore, our survey fields were primarily selected to contain X-ray data as shown in Table \[tab:survey\_field\]. We set a minimum ROSAT exposure time, $T_R$, of 25 ksec ensuring a detection limit of $L_x(0.5-2.0 keV) \sim 2 \times 10^{43}$erg/s at $z\simeq$0.5 (for $H_0=75$,$\Omega_M=0.3$, $\Omega_{\Lambda}=0.7$). This corresponds to $M \sim 10^{14}M_{\odot}$, which is well matched to the likely mass detection limit of our weak lensing survey [@miyazakietal02a]. More sensitive X-ray missions, XMM-Newton and Chandra, have been surveying ROSAT fields, in part, to deeper limits. These include the “Lynx”, “Lockman Hole” and “UK 13 hr deep field” in our target list. XMM is also actively involved in international campaigns of multi-wavelength wide field ($> 1 deg^2$) observations such as the “COSMOS” , “XMM-LSS” and “SXDS” fields which we also included. Among these, the XMM-LSS field [@pierre04] is of particular interest given its panoramic area and published cluster catalogs [@valtchanovetal04; @willisetal05; @pierreetal06]. Finally, we included the DEIMOS DEEP2 survey fields [@davisetal03] where spectra of $\sim$ 50,000 faint galaxies will eventually become available. This will enable close correlations of lensing mass and various measures of the luminosity density as was recently pioneered for the COSMOS field [@massey07]. Although X-ray data is not available for these DEEP fields at the current time, such observations are planned and will likely become available soon. Observations ------------ The imaging observations were largely carried out on May 1-4 and September 27, 2003 as part of an “Intensive Program” of Subaru Telescope. The Lynx field was observed on January 29-30, 2003, SXDS data was obtained from the public archive and the COSMOS field observed on February 18 and 21, 2004. The Suprime-Cam field size is 0.25 deg$^2$ and all observations were conducted with the $R_c$ filter (except the COSMOS field which was observed in $i'$-band to enable direct comparison with ACS$/$HST’s F814W images). The total exposure time was 30 minutes for each pointing, taken via four 7.5 minute exposures in a dither pattern of spacing $\sim$ 1 arcmin. Fig. \[fig:fwhmee\] (top) shows the seeing (FWHM) of each 7.5 minute exposure over a typical five night observing sequence. The bottom panel shows ellipticities of moderately bright unsaturated stars which provide a measure of the PSF anisotropy; the raw ellipticity is mostly 2 - 4 % (and occasionally $>$ 5 %). We discuss the question of the PSF anisotropy in Appendix \[sec:imagequality\]. Thus far, we have surveyed over 21.82 deg$^2$ as summarized in Table \[tab:survey\_field\]. The useful field area excludes the surroundings of bright stars and galaxies and field boundaries. Individual pointings whose seeing was worse than 0.9 arcsec were also excluded; this occurred for only 5 % of the clear time (see section \[sec:wlana\]). Data Reduction -------------- The data reduction procedures for the present survey closely followed those described in Section 3.1 of [@miyazakietal02b], enhanced as discussed below. Normally, with Suprime-Cam images, each CCD exposure is “mosaic-stacked” to yield a single image of a particular pointing. A simple geometrical model is used for the focal plane astrometry whose parameters include the effects of camera distortion, the displacement and rotation of each detector from a defined fiducial location and the offset and the rotation of the dithered exposures. The best fit parameters are obtained by minimizing the positional difference of control stars (70$\sim$100 stars per CCD) held common for each exposure. The residual alignment error in this procedure is $\sim$ 0.5 pixel rms (0.1 arcsec rms). Such a residual is sufficiently small for most imaging applications. However, in seeing better than 0.6 arcsec (FWHM), a misalignment of 0.1 arcsec between two images introduces a $\simeq$2 % ellipticity on the stacked image which is a serious issue for weak lensing studies. A further improvement is thus necessary. The residual $(\Delta x, \Delta y)$ from a reference frame is parameterized as a polynomial function of field position $\vec{x} =(x, y)$ as: $$\Delta x = \sum^{6}_{l=0}\sum^{l}_{m=0} a_{lm,e} x^{l-m}y^m \;\;\;\; \Delta y = \sum^{6}_{l=0}\sum^{l}_{m=0} b_{lm,e} x^{l-m}y^m$$ Pixel values are estimated via linear interpolation of neighboring four pixels. The coefficients $a_{lm,e}$ and $b_{lm,e}$ are then obtained by minimizing the variance of the residuals. A sixth order polynomial is usually sufficient for this purpose. Each individual CCD image is ‘warped’ using this polynomial correction prior to stacking. This process reduces the alignment error to $\sim$ 0.07 pixel (0.014 arcsec) and is similar to the “Jelly CCD” model described in [@kaiseretal99]. Although an external stellar catalog would ideally be used for accurate astrometry, no such data is available at the relevant faint limits (R $>$ 22). We therefore employ the first exposure, corrected by the simple geometrical model discussed above, as the basic reference frame. We noticed that this refined procedure still introduces some artificial deformation. Fig. \[fig:warp\](a) represents the raw image of one CCD whereas Fig. \[fig:warp\](b) is that slightly rotated by $7\times 10^{-4}$ rad, a typical value, using the mapping described above. Clearly some degree of artificial deformation is introduced. After some experimentation, we found this deformation arises from the undersampled nature of the Suprime-Cam images. By adopting 2$\times$2 oversampling prior to rotation, the resulting ellipticity field shows no sign of image deformation. However, this is a computationally a very time-consuming solution. Accordingly, in our final analysis, instead of using oversampled images, we modified the mapping procedure itself. We estimate the pixel values of the target images from 3rd order bi-linear polynomial interpolation of 4$\times$4 source pixels [^1] rather than the linear interpolation four neighboring pixels in the previous procedure. This mapping process avoids introducing image deformation and is significantly quicker computationally. Galaxy Catalogs {#sec:galaxycatalogs} --------------- Object finding and shape measurement was executed on the mosaic-stacked images using the [imcat]{} software suite developed by Nick Kaiser. A threshold [nu]{}=10 was adopted. Photometric calibration used Landolt standard stars [@landolt92] and the faint standards of [@majewskietal94]. We adopt the Vega magnitude system in the following. Galaxies are distinguished from stars via their half light radius, $r_h$, viz: $$r_h > r_{h}^{} + \sigma_{r_{h}^{}}$$ where $r_{h}^{}$ and $\sigma_{r_{h}^{}}$ are the half light radius of a stellar image and its rms respectively. The galaxy size distribution is shown in Fig. \[fig:size\_gd140\]. Fig. \[fig:ncnt\_gd140\] shows the cumulative number density of galaxies as a function of $R_c$-band magnitude. The surface density exceeds 50 arcmin$^{-2}$ when the seeing is superb (0.47 arcsec) and is $\sim15$ arcmin$^{-2}$ in those poor seeing images ($>$0.9 arcsec) discarded from our analysis (Fig. \[fig:seeing\_gdens\_pgamma\](a)). Table \[tab:survey\_field\] lists the seeing and the galaxy density for each field. Finally, we masked all objects close to bright stars (within 18 arcsec for $b_{USNO-A} < 15$, 90 arcsec for $b_{USNO-A} < 11.7$). Light halos around bright stars can introduce spurious galaxies. Weak Lensing Analysis {#sec:wlana} --------------------- ### Shape Measurements Object shapes are represented by the ellipticities, $\vec{e} = (e_1, e_2) = \{I_{11} - I_{22}, 2I_{12}\} / (I_{11}+I_{22})$ where $I_{ij}$ are Gaussian-weighted quadrupole moments of the surface brightness distribution. The point spread function (PSF) of the images is usually smeared by various instrumental effects such as optical aberrations and the tracking error of the telescope. The PSF anisotropy is estimated based on images of stars, and the galaxy images are corrected so that images of neighboring stars are re-circularized. Galaxy ellipticities are then corrected as: $$\label{Psmcorrection} \vec{e}' = \vec{e} - \frac{P_{sm}}{P_{sm}^{}}\vec{e}^{},$$ where the asterisk designates a stellar value, $P_{sm}$ is the smear polarisability tensor and is mostly diagonal [@ksb95]. $(P_{sm}^{-1}\vec{e})^{}$ is evaluated using stars in the field of view and modeled as 5th order bi-polynomial function of position. Eqn.\[Psmcorrection\] then applies this for the galaxy images. This correction is carried out independently on each pointing. We further justify the correction procedure in Appendix \[sec:imagequality\]. ### Shear Estimate The shear induced by gravitational lensing, $\vec{\gamma}$, is diluted by atmospheric seeing. [@lk97] developed a prescription to convert the observed ellipticities to a ’pre-seeing shear’ as $$\label{pgammacorrection} \vec{\gamma} = (P_{\gamma})^{-1}\vec{e}'$$ where $P_{\gamma}$ is the pre-seeing shear polarisability tensor defined as $$\label{pgamma} P_{\gamma} = P_{sh} - P_{sm} (P_{sm}^{})^{-1}P_{sh}^{} .$$ $P_{sh}$ is the shear polarisability tensor defined in [@ksb95], and $P_{sh}^{}$ is the stellar shear polarisability tensor. Note that (the inverse of) $P_{\gamma}$ represents the degree of dilution. Since the $P_{sh}$ and $P_{sm}$ are mostly diagonal, we replace the tensors in Eqn. \[pgamma\] with their trace and evaluate $P_{\gamma}$ as a scalar. The average value, $<P_{\gamma}>$, over all galaxies ($23 < R_c < 26 $) of each pointing is shown in Fig \[fig:seeing\_gdens\_pgamma\](b). $<P_{\gamma}>$ decreases slightly as the seeing worsens but the change is not very large (0.4 to 0.3). This is because we only select larger galaxies compared with the seeing size. Thus, the dilution factor is 30$\sim$40 % regardless of the seeing. In the mean time, we compare the first component of galaxy ellipticities, $e_1$, of Suprime-Cam and ACS/HST images taken in the COSMOS field, and the result is shown in Fig \[fig:suprime\_acs\]. The ellipticities observed by Suprime-Cam are in fact diluted by 36 % compared with those of ACS, which is consistent with Fig \[fig:seeing\_gdens\_pgamma\](b). To calculate $P_{\gamma}$ as a function of position we employed a “smoothing” scheme [@waerbeke00; @erbenetal01; @hamanaetal03]. We took the median $P_{\gamma}$ for 20 neighboring galaxies on the $r_g$-magnitude plane (where $r_g$ is a measure of object size adopted in the [imcat]{} suite). In deriving the mean, the weight $w$ on an individual measure is taken to be: $$\label{weight} w = \frac{1}{\sigma_{\gamma}^2 + \alpha^2},$$ where $\sigma_{\gamma}$ is the variance of the raw $\gamma$ of those 20 neighbors, obtained using the raw $P_{\gamma}$. $\alpha$ is the variance of all of the galaxies in the catalog ($\sim$ 0.4). In general, the weighted value of a quantity $<A>$ is calculated as $<A> = \Sigma_{i=1}^N w_i A_i/\Sigma_{i=1}^Nw_i$ . The method we adopt is based on that adopted by @ksb95. More sophisticated methods have since been developed and the variants are summarized by @heymansetal06. In their notation, our method is very similar to the procedure termed “LV”. Based on the results of the STEP simulation study, [@heymansetal06] concluded that both the “KSB+” method, modified by [@hoekstraetal98] and implementations of “BJ02” method [@bernsteinandjarvis02] are able to reconstruct input shears to a few percent level. To calibrate our method, we analyzed the simulated data provided by [@heymansetal06]. For the model designated “PSF3”, we underestimate the input shear by 5% for $\gamma_{true}$ = 0.1 and 0.05, whereas for $\gamma_{true} \le 0.01$, the difference, $\gamma_{obs} - \gamma_{true}$ is insignificant. An error of 5 % in the recovered shear is competitive with most of the methods discussed by [@heymansetal06] ($\sim$ 7 % is a typical error). A 5% shear error would induce a similar uncertainty in the mass estimate of a typical halo. Such an error is considered adequate for the applications envisaged. ### Kappa Map The dimensionless surface mass density, $\kappa(\vec{\theta})$, is estimated from the shear field $\gamma(\vec{\theta})$ by the [@kaiserandsquires93] inversion algorithm as: $$\label{eqn:ks93} \kappa(\vec{\theta}) = \frac{1}{\pi} \int d^{2}\theta' Re[D^{}(\vec{\theta} - \vec{\theta'})\gamma(\vec{\theta '})],$$ where $D(\vec{\theta}) $ is defined as $$\label{eqn:complexkernel} D(\vec{\theta}) = \frac{-1}{(\theta_1 - i \theta_2)^2},$$ which is a complex convolution kernel for $\kappa$ to obtain the shear $\gamma$. In the actual implementation, we smoothed to avoid the effect of noise. A smoothing scale of $\theta_G = 1$ arcmin was chosen following the discussion given by @hamanaetal03. We adopt 15$\times$15 arcsec$^2$ square grids and then calculate $\kappa(\vec{\theta})$ on each grid using Eqn.\[eqn:ks93\] to obtain the $\kappa$ map. In order to estimate the noise of the $\kappa$ field, we randomized the orientations of the galaxies in the catalog and created a $\kappa_{\rm noise}$ map. We repeated this randomization 100 times and computed the rms value at each grid point where $\kappa$ is computed. Assuming the $\kappa$ error distribution is Gaussian, this rms represents the 1-$\sigma$ noise level, and thus the measured signal divided by the rms gives the signal/noise ratio, $\nu(\theta)$, of the $\kappa$ map at that point. The Gaussian-smoothed signal/noise map is then searched for mass concentrations. [@hennawiandspergel05] concluded that a ’truncated’ NFW filter applied to the aperture mass map, $M_{ap}$, is the most efficient detection technique. Such optimization may be necessary to improve the efficiency of future very wide field surveys where thousands of clusters are sought. Here we adopt a simpler technique in order to evaluate its effectiveness in a direct comparison of lensing and X-ray techniques. Halo Catalog ------------ Figure \[fig:saclay\_halomap\] shows the results of a halo search in one of our fields: the XMM-Wide field. The red contour shows the $\kappa$ S/N map where the threshold and increment are set at 2 and 0.5, respectively. The blue contour shows the surface number density of moderately bright galaxies ($21 < R_c < 23$). Local peaks are searched on the $\kappa$ S/N map and their positions are marked as open and filled circles. Figure \[fig:deep02\_halomap\] to \[fig:deep23\_halomap\] show the $\kappa$ S/N maps for the remainder of our survey fields. While visually inspecting galaxy concentrations around the detected halos, we noticed that less concentrated halos tend to occur preferentially near bright stars and field boundaries. Since regions near bright stars are masked (section \[sec:galaxycatalogs\]) conceivably the discontinuity in the faint background galaxy distribution could cause spurious peaks. To avoid this, we reject halos occurring within a 4 arcmin radius of bright stars ($b_{USNO-A}$ $<$ 11) and within 2.3 arcmin of the field boundary. These restrictions reduce the survey area by 23 % to what we will refer to as the [secure survey area]{} (16.72 deg$^2$). Halos found within the secure area are termed the [secure sample]{}. It is certainly possible that a significant fraction of halos lying in the non-secure area are genuine clusters. We will discuss this further in a later paper concerned with their spectroscopic follow-up (Green et al, in preparation). Detailed inspection of the halo candidates and the spectroscopic follow-up discussed below revealed our completeness is high to a limiting signal to noise in the convergence map of 3.69. Table \[tab:halocat\], \[tab:halocat2\] lists the secure sample with S/N $>$ 3.69. In this table, N$_g$ represents the number of moderately bright ($R_c < 22$) galaxies within 2 arcmin, indicative of the galaxy concentration. Concerning the optimum threshold for the significance, decreasing it will increase the halo sample but likely introduce more false detections. The optimum value should be set based on the spectroscopically-observed true/false rate. Investigating the rate is a major goal of our study. In this work, the least significant spectroscopically-identified halos have S/N = 3.69 (XMM-Wide n=23). We adopted the threshold of 3.69 for this work so that all the spectroscopically followed-up samples are included in the table. We adopted significance maps for the selection of candidates rather than kappa maps. This is because we would like to minimize contamination by the false peaks. However, the effective kappa threshold varies over the field, so we may encounter a “completeness” problem; i.e. halos that have high kappa value are lost from the list. We investigated such omissions in the XMM-Wide field (Figure \[fig:saclay\_halomap\]), and only one halo is found in this category, with $S/N < 3.69$ & $\kappa > \kappa_{thres}$ where the $\kappa_{thres}$ is calculated as $\kappa_{thres} = 3.69 \times Noise_{global}$. The “$Noise_{global}$” is estimated globally over the entire XMM-Wide field kappa map, and is 0.018 here. This halo is lost because the local noise is as high as 0.022. In practice, it will be very hard to generate completely uniform data sets over the entire field of a survey because weather and seeing conditions will vary. We will have to optimize the $\kappa_{thres}$ on a field by field basis based on the data quality of each field. Therefore, our strategy is the following: at first, we collect reliable cluster samples based on the significance, and then users of the catalog can set their own kappa threshold or mass threshold to carry out their studies. This work represents the results of the first step above. We list the kappa values in Table \[tab:halocat\], \[tab:halocat2\] for reference. ![image](f7.eps){height="17.078cm"} \[tab:halocat\] [llllllllllll]{} Field & n & ID & RA & DEC & $\kappa$S/N & $\kappa$ & N$_g$ & FOCAS & Known & NEDG & Note\ DEEP02 & 00 & - & 37.32 & 0.63 & 4.44 & 0.099 & 16 & - & - & 1.35 &\ & 01 & - & 37.87 & 0.51 & 4.39 & 0.083 & 24 & - & - & - &\ & 02 & - & 37.38 & 0.41 & 4.25 & 0.120 & 19 & - & - & 0.73 &\ & 04 & - & 37.15 & 0.73 & 4.11 & 0.086 & 25 & - & - & 0.10 &\ & 05 & - & 37.31 & 0.44 & 4.07 & 0.131 & 20 & - & - & 1.03 &\ & 06 & - & 37.72 & 0.69 & 3.97 & 0.081 & 18 & - & - & 0.86 &\ & 07 & - & 37.86 & 0.57 & 3.95 & 0.074 & 18 & - & - & 0.92 &\ & 08 & SL J0228.4+0030 & 37.12 & 0.51 & 3.93 & 0.085 & 49 & - & 0.46(P) & - & VGCF 46\ & 09 & SL J0228.2+0033 & 37.07 & 0.55 & 3.84 & 0.102 & 33 & - & 0.50(P) & - & [SDSS CE J037.099808+00.540769 ]{}\ SXDS & 00 & - & 34.29 & -5.59 & 5.33 & 0.057 & 26 & - & - & - &\ & 01 & - & 34.38 & -4.86 & 4.14 & 0.084 & 24 & - & - & - &\ & 02 & - & 34.61 & -4.41 & 3.96 & 0.059 & 16 & - & - & - &\ & 03 & - & 34.74 & -4.70 & 3.90 & 0.044 & 19 & - & - & - &\ & 04 & - & 34.96 & -5.12 & 3.86 & 0.069 & 21 & - & - & - &\ & 05 & - & 34.41 & -4.50 & 3.77 & 0.056 & 26 & - & - & - &\ XMM-Wide & 00 & SL J0221.7-0345 & 35.44 & -3.77 & 8.15 & 0.156 & 72 & - & 0.43 & - & XLSSC 006\ & 01 & SL J0225.7-0312 & 36.43 & -3.21 & 5.72 & 0.108 & 41 & 0.14 & - & - & LRIS z = 0.14\ & 02 & SL J0224.4-0449 & 36.10 & -4.82 & 5.06 & 0.074 & 40 & 0.49 & - & - &\ & 04 & - & 35.34 & -3.50 & 4.91 & 0.082 & 21 & - & - & - &\ & 08 & SL J0222.3-0446 & 35.48 & -3.80 & 4.33 & 0.081 & 29 & - & - & - & LRIS z = 0.41\ & 10 & - & 36.25 & -4.25 & 4.20 & 0.062 & 23 & - & - & - &\ & 12 & SL J0224.5-0414 & 36.13 & -4.24 & 4.06 & 0.057 & 70 & 0.26 & - & - & LRIS z = 0.26\ & 15 & SL J0225.3-0441 & 36.34 & -4.70 & 3.94 & 0.091 & 34 & 0.26 & - & - &\ & 16 & SL J0228.1-0450 & 37.03 & -4.84 & 3.94 & 0.072 & 31 & 0.29 & - & - &\ & 17 & SL J0226.5-0401 & 36.63 & -4.02 & 3.90 & 0.079 & 37 & - & 0.34 & - & XLSSC 014\ & 19 & SL J0227.7-0450 & 36.94 & -4.85 & 3.81 & 0.064 & 43 & - & 0.29 & - & Pierre et al. (2006)\ & 20 & - & 35.98 & -3.77 & 3.81 & 0.048 & 20 & - & - & - &\ & 21 & SL J0228.4-0425 & 37.12 & -4.43 & 3.80 & 0.055 & 49 & - & 0.43 & - & XLSSC 012\ & 22 & SL J0225.4-0414 & 36.36 & -4.25 & 3.72 & 0.073 & 43 & 0.14 & - & - &\ & 23 & SL J0222.8-0416 & 35.71 & -4.27 & 3.69 & 0.049 & 52 & 0.43,0.19,0.23 & - & - &\ Lynx & 00 & - & 131.91 & 44.80 & 5.84 & 0.121 & 20 & - & - & - &\ & 01 & - & 132.59 & 44.07 & 5.01 & 0.083 & 43 & - & - & - &\ & 03 & - & 131.83 & 44.86 & 4.57 & 0.139 & 23 & - & - & - &\ & 05 & - & 131.77 & 44.85 & 4.37 & 0.110 & 13 & - & - & - &\ & 07 & - & 132.69 & 44.95 & 4.15 & 0.105 & 31 & - & - & - &\ & 08 & SL J0850.5+4512 & 132.64 & 45.20 & 4.02 & 0.085 & 53 & 0.19 & 0.24(P) & - & [NSC J085029+451141,LRIS$z=0.19$]{}\ & 09 & - & 131.47 & 44.96 & 4.02 & 0.076 & 26 & - & - & - &\ & 10 & - & 133.02 & 44.14 & 4.00 & 0.108 & 23 & - & - & - &\ & 12 & - & 132.37 & 44.38 & 3.90 & 0.081 & 36 & - & - & - &\ & 13 & - & 132.41 & 44.37 & 3.90 & 0.072 & 31 & - & - & - & Part of n=12\ & 14 & - & 132.54 & 44.07 & 3.86 & 0.066 & 48 & - & - & - &\ & 15 & - & 132.81 & 44.35 & 3.77 & 0.077 & 39 & - & - & - &\ & 16 & - & 132.31 & 44.30 & 3.75 & 0.072 & 44 & - & - & - &\ & 17 & - & 131.40 & 44.94 & 3.74 & 0.084 & 25 & - & - & 0.15 &\ COSMOS & 00 & SL J1000.7+0137 & 150.19 & 1.63 & 6.11 & 0.113 & 64 & 0.22 & 0.20(P) & - & NSC J100047+013912\ & 01 & SL J1001.4+0159 & 150.35 & 1.99 & 5.64 & 0.098 & 32 & - & 0.85(P)& - & [@finoguenovetal06]\ & 02 & SJ J0959.6+0231 & 149.92 & 2.52 & 4.74 & 0.067 & 83 & - & 0.73(P) & - & [@finoguenovetal06]\ & 05 & - & 149.65 & 1.55 & 3.92 & 0.078 & 47 & - & - & - &\ & 07 & - & 150.19 & 2.01 & 3.88 & 0.070 & 36 & - & - & - &\ \[tab:halocat2\] [llllllllllll]{} Field & n & ID & RA & DEC & $\kappa$S/N & $\kappa$ & N$_g$ & FOCAS & Known & NEDG & Note\ Lockman & 00 & SL J1057.5+5759 & 164.39 & 58.00 & 6.28 & 0.109 & 68 & 0.60 & - & - &\ & 03 & SL J1051.5+5646 & 162.88 & 56.77 & 4.97 & 0.082 & 31 & 0.33, 0.35 & - & - &\ & 05 & SL J1047.3+5700 & 161.84 & 57.01 & 4.56 & 0.103 & 56 & 0.30, 0.24 & - & - &\ & 06 & SL J1049.4+5655 & 162.35 & 56.93 & 4.51 & 0.095 & 47 & 0.42 & - & - & LRIS z = 0.31\ & 09 & SL J1055.4+5723 & 163.86 & 57.38 & 4.22 & 0.086 & 20 & - & - & - & LRIS z = 0.38\ & 10 & SL J1051.6+5647 & 162.92 & 56.78 & 4.20 & 0.068 & 49 & 0.33, 0.25 & - & 0.05 & part of SL J1051.5+5646\ & 11 & SL J1053.4+5720 & 163.35 & 57.34 & 4.08 & 0.064 & 50 & - & 0.34 & - & RX J1053.3+5719\ & 12 & - & 163.69 & 57.55 & 4.07 & 0.068 & 26 & - & - & - &\ & 13 & - & 162.91 & 58.02 & 4.04 & 0.089 & 21 & - & - & 0.08 &\ & 14 & - & 162.54 & 57.28 & 3.93 & 0.070 & 38 & - & - & - &\ & 15 & SL J1048.1+5730 & 162.04 & 57.51 & 3.89 & 0.071 & 35 & 0.32 & - & - &\ & 16 & - & 163.85 & 57.95 & 3.83 & 0.072 & 24 & - & - & 0.02 &\ & 18 & - & 163.16 & 57.88 & 3.77 & 0.069 & 26 & - & - & - &\ & 19 & - & 164.21 & 57.70 & 3.77 & 0.070 & 22 & - & - & - &\ & 20 & - & 163.23 & 57.84 & 3.73 & 0.102 & 29 & - & - & - &\ & 21 & - & 163.14 & 57.82 & 3.72 & 0.111 & 19 & - & - & - &\ GD140 & 00 & SL J1135.6+3009 & 173.91 & 30.16 & 4.98 & 0.126 & 35 & 0.21 & - & - &\ & 01 & - & 173.89 & 30.21 & 4.19 & 0.086 & 23 & - & - & - &\ & 02 & - & 173.96 & 29.81 & 4.09 & 0.069 & 17 & - & - & - &\ & 03 & SL J1136.3+2915 & 174.09 & 29.26 & 4.03 & 0.100 & 24 & - & - & - & LRIS z = 0.20\ & 05 & - & 174.77 & 29.89 & 3.86 & 0.111 & 26 & - & - & - &\ & 06 & - & 174.86 & 30.33 & 3.83 & 0.080 & 21 & - & - & - &\ PG1159-035 & 05 & SL J1201.7-0331 & 180.44 & -3.53 & 4.42 & 0.077 & 49 & 0.52 & - & - &\ & 06 & - & 180.99 & -3.09 & 3.90 & 0.119 & 21 & - & - & 0.09 &\ & 08 & - & 181.76 & -3.27 & 3.71 & 0.102 & 37 & - & - & - &\ 13 hr Field & 00 & SL J1334.3+3728 & 203.60 & 37.47 & 4.33 & 0.128 & 74 & 0.30 & 0.48(P) & - & NSCS J133424+372822\ & 01 & SL J1335.7+3731 & 203.94 & 37.53 & 4.10 & 0.091 & 65 & 0.41 & - & - &\ & 04 & SL J1337.7+3800 & 204.43 & 38.01 & 3.85 & 0.080 & 34 & 0.18 & - & - &\ & 06 & - & 203.85 & 37.90 & 3.78 & 0.068 & 27 & - & - & - &\ & 07 & - & 204.22 & 37.54 & 3.77 & 0.078 & 30 & - & - & - &\ GTO 2deg$^2$ & 00 & SL J1602.8+4335 & 240.72 & 43.59 & 6.65 & 0.110 & 56 & 0.42 & - & - &\ & 01 & SL J1603.1+4245 & 240.78 & 42.76 & 5.47 & 0.106 & 57 & - & - & - & LRIS z = 0.18\ & 02 & - & 241.82 & 43.19 & 5.17 & 0.133 & 21 & - & - & - &\ & 04 & - & 241.95 & 43.60 & 4.49 & 0.093 & 17 & - & - & - &\ & 07 & SL J1604.1+4239 & 241.04 & 42.65 & 4.19 & 0.083 & 39 & - & - & - & LRIS z = 0.30\ & 08 & - & 241.63 & 43.61 & 4.16 & 0.093 & 25 & - & - & - &\ & 09 & SL J1605.4+4244 & 241.36 & 42.74 & 4.09 & 0.064 & 36 & 0.22 & - & - &\ & 10 & - & 241.18 & 43.46 & 3.85 & 0.092 & 29 & - & - & - &\ & 11 & - & 241.70 & 43.64 & 3.84 & 0.077 & 33 & - & - & - &\ & 12 & SL J1603.1+4243 & 240.78 & 42.72 & 3.82 & 0.091 & 37 & - & - & - & Part of SL J1603.1+4245\ CM DRA & 04 & SL J1639.9+5708 & 249.98 & 57.15 & 4.15 & 0.073 & 32 & 0.20 & - & - &\ & 06 & SL J1634.1+5639 & 248.55 & 56.66 & 3.97 & 0.104 & 31 & 0.24 & - & - &\ DEEP16 & 00 & SL J1647.7+3455 & 251.94 & 34.93 & 4.30 & 0.084 & 42 & 0.26 & - & - &\ & 01 & - & 253.54 & 34.98 & 4.24 & 0.078 & 24 & - & - & - &\ & 02 & - & 252.36 & 35.02 & 3.75 & 0.123 & 25 & - & - & - &\ & 03 & - & 251.79 & 35.04 & 3.72 & 0.087 & 21 & - & - & - &\ DEEP23 & 00 & SL J2326.4+0012 & 351.61 & 0.20 & 4.41 & 0.076 & 19 & 0.28 & - & - &\ & 01 & - & 351.75 & 0.00 & 4.37 & 0.095 & 20 & - & - & - &\ & 02 & - & 352.33 & 0.15 & 4.27 & 0.107 & 33 & - & - & 1.38 &\ & 03 & - & 352.47 & -0.06 & 4.11 & 0.079 & 34 & - & - & 0.07 &\ & 04 & - & 352.22 & 0.09 & 3.87 & 0.084 & 21 & - & - & 1.37 &\ & 05 & - & 353.09 & 0.04 & 3.70 & 0.066 & 22 & - & - & - &\ Here we compare the detected halo number density \[deg$^{-2}$\] with the prediction of numerical simulations done by [@hamanaetal04], who have attempted to reproduce a survey such as ours. Figure \[fig:seeingdens\] shows the halo density of our thirteen survey fields. The horizontal axis shows the density of faint galaxies ($\rho_{gal}$) used for weak lensing analysis. The error bars are based only on $\sqrt{N}$ estimates and do not include the effects of cosmic variance. Solid lines in Figure \[fig:seeingdens\] show the prediction of the simulation where three different mean redshifts for the background galaxies are assumed (from bottom to top: 0.8, 1.0 and 1.2). Although the scatter is large, there is reasonable agreement between the observations and the prediction. We also see a gradual increase of the observed number density over the $\rho_{gal}$ range sampled as expected. These comparisons validate our observational procedures. We have, however, two outliers in Figure \[fig:seeingdens\]. These could be caused by cosmic variance or may arise from some other reasons. Further follow-up studies will be important to clarify the issue. Cluster Identification and Verification ======================================= Armed with the shear-selected halo catalog, we now discuss the tests we have made to verify its reliability, using both spectroscopic observations and comparisons with X-ray data in the fields where the overlap of targets can be studied. Spectroscopic Follow-up ------------------------ Because the lensing kernel (or window function) has a relatively broad redshift range, the superposition of multiple low mass halos with different redshifts could yield highly significant weak lensing signals. Such ‘superposition halos’ are unwelcome in a mass-selected cluster catalog. In order to identify such superposition halos we undertook a more comprehensive spectroscopic survey using a multi-object spectrograph for selected candidates. This also gives us the velocity dispersion of member galaxies, which is an estimate of dynamical mass of clusters. By comparing dynamical mass with weak lensing mass we will be able to discuss the dynamical state of the clusters (Hamana et al. in preparation). We used the FOCAS spectrograph on Subaru whose multi-object mode permits the simultaneous observation of 25 - 30 galaxies in the field of a particular halo. We used the 150/mm grating and a SY47 order sorting filter. This configuration spans the wavelength range $4700-9400$Å [@kashikawaetal02]. Target selection was based primarily on apparent magnitude and the color information was used when available. The exposure time was 45 $\sim$ 70 minutes depending on the magnitude of the selected galaxies and the observing conditions. The spectroscopic data was reduced using standard IRAF procedures (Hamana et al. in preparation). Figure \[fig:focasobs\] shows a typical FOCAS observation where the target is identified as a $z$ = 0.6 cluster. ![image](f9.eps){height="8.5cm"} Since May 2004 we have observed 26 halos from the secure sample with FOCAS. Higher priority was given to more significant halos, except in the early stages of program (for example in the Lynx and PG1159-035 fields) when the follow-up strategy was still being evaluated. Each of the 26 halos has been reliably identified with a cluster of galaxies. The redshift so determined is shown in the column labeled “FOCAS” of Table \[tab:halocat\],\[tab:halocat2\]. In parallel with the FOCAS follow-up, long slit observations with Keck/LRIS have been carried out to enlarge the sample of redshifts as quickly as possible. We have to keep in mind that this method cannot discriminate the projected samples, and some statistical consideration is necessary in dealing with the data for further studies. A complete discussion of that aspect of our program is discussed in a separate paper (Green et al, in prep.). In the “Note” column of Table \[tab:halocat\],\[tab:halocat2\], we list the preliminary Keck/LRIS results as “LRIS z = zvalue” whose identification is already robust (e.g. at least two galaxy redshifts agree) including one of the halos in the XMM-Wide field (SL J0222.3-0446, $z$=0.41) where we discuss the reliability of our catalog (see section \[sec:reliability\]). Correlation with Published Data ------------------------------- In addition to undertaking our own spectroscopic observations, we also searched the NASA Extragalactic Database (NED) for clusters and groups with published redshifts close to the position of our detected halos. We included recently-published cluster catalogs in the XMM-wide survey by [@valtchanovetal04], [@willisetal05] and [@pierreetal06] ([VWP]{} data hereafter). In making assignments, we considered a cluster to be associated with a halo if the angular separation was less than 2 arcmin and the 3 dimensional distance less than 1 $h^{-1}$ Mpc. 13 out of 100 halos can be identified in this way with rich clusters in the literature and the published redshift is shown in the column labeled “Known” of Table \[tab:halocat\],\[tab:halocat2\]. Three of these thirteen clusters were also observed by FOCAS in multi-object mode and we note that the mean FOCAS redshift is different from the published value for these systems ($\delta z = 0.02 \sim 0.18$). These discrepancies can be explained by the fact that NED redshift of these clusters are all photometric, which have larger uncertainties. Therefore, we adopted the FOCAS redshifts for these three cases. We searched the NED galaxy group catalog in the same manner, but no group matched any of our halos. Next, we searched for individual galaxies with published redshifts within 3 arcmin of our halos. If suitable data was found, we grouped them together using a “friend of friend” algorithm with a linking length of 1 $h^{-1}$ Mpc. If we found a clustering of more than two members, we calculated the average redshift and the mean astrometric position. Using the same proximity criterion above, we assigned a redshift to 14 further halos. These are shown in the column labeled “NEDG”. In order to estimate the probability of a chance coincidence of the matching procedures, we first randomized the position of the detected halos in the field, and applied the same matching procedure described above to the randomized halo catalog. This process suggests that the probability of a chance coincidence is roughly 10 % and 25 % for the cluster and galaxy clustering searches, respectively. Clearly care should be taken in analyzing such data. The total number of reliably identified clusters are 41 (out of 100 in Table \[tab:halocat\],\[tab:halocat2\]) where we do not include those identified by “NEDG” galaxies because the chance coincidence is not negligible. We assign ID labels only for these reliable halos in the tables. Reliability of the Shear-Selected Halo Catalog {#sec:reliability} ---------------------------------------------- We now turn to the important question of the evaluating the reliability of our shear-selected catalog. We will do this by examining both the success rate of our identifications and by comparing with X-ray samples obtained via the XMM-Wide field survey. VWP have confirmed spectroscopic redshifts for 28 X-ray selected clusters in the 3.5 deg$^2$ survey field. Their locations are plotted as squares on Figure \[fig:saclay\_halomap\]. Our secure sample of 15 shear-selected halos with S/N $>$ 3.69 is plotted as solid circles. Those we have spectroscopically confirmed using Subaru’s FOCAS and/or Keck LRIS are marked by large open circles. We find that three halos (1, 16 and 23) are confirmed as clusters by FOCAS and are not reported in VWP. Among our 15 shear-selected halo samples, only three (4, 10 and 20) have yet to be identified. This identification success rate (80 %) can be regarded as a lower limit given future observations may yet locate an associated cluster. Although affected by small number statistics, this minimum efficiency is already higher than expected by simulation studies[@whiteetal02; @hamanaetal04; @hennawiandspergel05]. It is interesting to compare our identification success rate with those found in the [GaBoDS]{} survey. [@maturietal06] found 14 significant halos in their 18 deg$^2$ area. Among them, 5 halos turn out to be [known]{} clusters of galaxies, 2 seem to have associated [light]{} concentrations but no spectroscopic confirmation, and the remaining [uncertain]{} 7 (50 %) had no apparent counterpart in either optical nor X-ray data. [@schirmeretal06] undertook a search on the same data adopting a different peak selection algorithm and a lower detection threshold. They found 158 “possible mass concentration” on the 18 deg$^2$ field. If those halos are divided into the same classes as above ([known, light, uncertain]{}), the ratio is almost the same as [@maturietal06]. Regardless, almost half of the [GaBoDS]{} halos could be considered uncertain at this point. Meanwhile, we find 15 halos in a 2.24 deg$^2$ field (XMM-wide) and show that 80 % of have been already identified as clusters. It is too early to make any definite conclusion because their spectroscopic is underway. However, our (tentatively) higher success rate could be explained at least in part by the larger number density of faint galaxies, $\rho_{gal}$, usable for the weak lensing analysis owing to larger aperture and better average seeing. In the case of [GaBoDS]{}, $\rho_{gal}$ spans from 6 to 28 arcmin$^{-2}$ depending on the field; the average value is 11 arcmin$^{-2}$. This is generally smaller than our $\rho_{gal}$ shown in Table \[tab:survey\_field\]. As a result one can expect a lower angular resolution of the $\kappa$ map, reducing the S/N ratio for a fixed smoothing scale and possibly an increased degree of contamination in the resulting halo catalog. Another possibility to explain the increased contamination is that their sample consists of a combination of different sets of catalogs, each of which is selected by different methods. This could decrease the significance threshold effectively, and could introduce more false peaks. Superposition of Multiple Clusters ---------------------------------- [@hamanaetal04] estimated, based on their simulation, that the halo superposition rate in survey such as ours should be roughly 3 % - a small but not negligible effect. A longslit spectroscopic survey, such as that undertaken with LRIS (Green et al, in prep) might be poorly-suited for locating such cases. However, the FOCAS multi-object survey reported here should reliably find them. In fact, we have found three apparent superposition halos (SL J0222.8-0416, SL J1051.5+5456 and SL J1047.3+5700) out of 26 examined. The overlap rate is broadly consistent with expectation considering the small number so far sampled. Conclusions and Future Prospects ================================ We have introduced a new Subaru imaging survey and described techniques for locating and verifying shear-selected halos. Across a search area of 16.72 deg$^2$ we have found 100 halos whose $\kappa$ S/N exceeds 3.69. We have described the first phase of a detailed follow-up campaign based on multi-object spectroscopy of 26 halos using FOCAS on the Subaru telescope. A later paper in this series (Green et al, in prep) will extend the spectroscopic survey to the full sample using a longslit approach. Detailed studies on one of our fields, the XMM-wide field, show that 80% of the shear selected 15 halos in the 2.2 deg$^2$ area can be confirmed as genuine clusters of galaxies. 10 overlap with X-ray detections and two are new systems confirmed spectroscopically. The overall success rate and reliability of our sample provides convincing proof that, with care, a weak lensing survey can provide a large sample of mass-selected halos. We compare our success rate and the reliability of our catalog with that of the [GaBoDS]{} survey and conclude a major advantage of our approach is the superior imaging depth which leads to a high surface density of usable galaxies. This suggests future, more ambitious, surveys for shear-selected halos will be more effective if undertaken with large aperture telescopes. It is interesting to use our results to estimate the requirements for a future survey motivated by the need to constrain dark energy. [@kolbetal06] discuss hypothetical missions which would aim to analyze the redshift distribution $N(z)$ of 10,000 clusters. Figure \[fig:seeingdens\] shows that our survey technique typically finds 5 halos deg$^{-2}$. Because the field size of Suprime-Cam is $\sim$ 0.25 deg$^2$, it takes two hours to cover 1 deg$^2$ assuming the exposure time of 30 minutes of each field as adopted in this study. A survey of 10,000 clusters would be prohibitive even in terms of imaging alone ($\sim$ 500 clear nights) even before contemplating the follow-up spectroscopy. The Hyper Suprime-Cam (HSC) project [@hsc06] has been proposed to remedy this shortcoming. HSC is expected to have a field of view ten times larger than Suprime-Cam while maintaining the same image quality. This new facility will makes the 2000 deg$^2$ scale imaging survey within a reasonable number of clear nights. Spectroscopic follow-up might be enabled by the proposed prime focus multi object optical spectrograph WFMOS (Wide Field Fiber Multi Object Spectrograph) whose field is 1.5 deg$^2$ in diameter. Typically we can expect $\simeq$10 shear-selected clusters in each spectroscopic field. With only 20 targets per halo, only a small fraction of the several thousand fibers envisaged for WFMOS need be allocated to the halo verification program. We are very grateful to Subaru astronomers: Y. Oyama, K. Aoki and T. Hattori for their dedicated supports of the FOCAS observing. Numerical computations presented in this paper were carried out at the Astronomical Data Center (ADC) and at the Center for Computational Astrophysics (CfCA) of the National Astronomical Observatory of Japan. This work is supported in part by Grant-in-Aid for Scientific Research (Kaken-hi) of Japan Society for the Promotion of Science (JSPS): Project number 15340065 (SM&TH) and 17740116 (TH) . [999]{} Allen, S.W., Schmidt, R.W., Fabian, A.C., Ebeling, H. 2003 342, 287 Allen, S.W., Schmidt, R.W., Ebeling, H., Fabian, A.C., van Speybroeck, L. 2004 , 353, 457 Becker, M.R. et al. 2007 submitted to ApJ (astro-ph/0704.3614) Bernstein, G. & Jarvis, M. 2002, , 123, 583 Böhringer, H. et al. 2001 A&A, 369, 826 Davis, M., Faber, S.M., Newman, J. et al. 2003, SPIE, 4834, 161 Dietrich, J.P., Erben, T., Lamer, G., Schneider, P., Schwope, A., Hartlap, J., Maturi, M (2007) A&A in press (astro-ph/0705.3455) Erben, T., van Waerbeke, L., Bertin, E., Mellier, & Y. & Schneider, P. 2001, A&A, 366, 717 Finoguenov, A., Guzzo, L., Hasinger, G., Scoville, N.Z. et al. (2006) ApJS in press (astro-ph/0612360) Gladders, M.D. & Yee, H.K.C. 2000, , 120, 2148 Goto, T. et al. 2002, , 123, 1807 Hamana, T., Miyazaki, S. et al. 2003, , 597, 98 Hamana, T., Takada, M., Yoshida, N. 2004, , 350, 893 Hennawi, J.F. & Spergel, D.N. 2005 , 624, 59 Henry, J.P. 2000, , 534, 565 Hetterscheidt, M., Erben, T., Schneider, P., Maoli, R., van Waerbeke, L., Mellier, Y. 2005, A&A, 442, 43. Heymans, C. et al. 2006, , 368, 1323 Hokekstra, H., Franx, M., Kuijken, K., Squires, G. 1998 , 504. 636 Hoekstra, H. 2004, , 347, 1337 Ikebe, Y., Reiprich, T.H., Böhringer, H., Tanaka, Y., Kitayama, T. 2002 A&A, 383, 773 Jenkins, A., Frenk, C. S., White, S. D. M., Colberg, J. M., Cole, S., Evrard, A. E., Couchman, H. M. P. & Yoshida, N. 2001, , 324, 450 Kaiser, N. & Squires, G. 1993, , 404, 441 Kaiser, N., Squires, G. & Broadhurst, T. 1995, , 449, 460 Kaiser, N., Wilson, G., Luppino, G., Dahle, H. 1999 submitted to PASP (astro-ph/9907229) Kashikawa, N. et al. 2002, , 54, 819 Kolb et al. 2006, US Dark Energy Task Force Report Landolt, A.U. 1992, , 104, 340 Levine, E.S., Schulz, A.E., White, M. 2002 , 577, 569 Luppino, G.A. & Kaiser, N. 1997, , 475, 20L Majewski, S.R., Kron, R.G., Koo, D.C., Bershady, M.A. 1994, , 106, 1258 Massey, R.J, Rhodes, J., Ellis, R.S. et al 2007, Nature, 445, 286 Maturi, M., Schirmer, M., Meneghetti, M., Bartelmann, M., Moscardini, L. 2006, A&A in press (astro-ph/0607254) Miyazaki, S., Hamana, T., Shimasaku, Furusawa, H., Doi, M., Hamabe, M., Imi, K., Kimura, M., Komiyama, Y., Nakata, F., Okada, N., Okamura, S., Ouchi, M., Sekiguchi, M., Yagi, M., Yasuda, N. 2002a , 580, L97 Miyazaki, S., Komiyama, Y., Okada, N., Imi, K., Yagi, M., Yasuda, N., Sekiguchi, M., Kimura, M, Doi, M., Hamabe, M., Nakata, F., Shimasaku, K., Furusawa, H., Ouchi, M. & Okamura, S. 2002b, , 54, 833 Miyazaki, S., Komiyama, Y., Nakaya, H., Doi, Y., Furusawa, H., Gillingham, P., Kamata Y., Takeshi, K., Nariai, K. 2006, SPIE, 6269, 9 Pierre, M. et al. 2004 J. Cosmol. Astropart. Phys, 09, 011 Pierre, M., Pacaud, F. et al. 2006, , 372, 591 Postman, M., Lubin, L.M., Gunn, J.E., Oke, J.B., Hoessel, J.G., Schneider, D.P., Christensen, J.A. 1996, , 111, 615 Press, W.H., Flannery, B.P., Teukolsky, S.A., Vetterling, W.T. 1993, “Numerical Recipes in C”, ISBN-13: 9780521431088 Reprich, T.H. & Böhringer, H. 2002 , 567, 716 Schirmer, M., Erben, T., Hetterscheidt, M., Schneider, P. 2006, Submitted to A&A (astro-ph/0607022) Smail, I., Ellis, R.S., Dressler, A. et al. 1997 , 479, 70 Smith, G.P., Kneib, J., Smail, I., Mazzotta, P., Ebeling, H., Czoske, O. 2005, , 359, 417 Valtchanov, I. et al. 2004, A&A, 423, 75 Wang, S., Khoury, J., Haiman, Z. May, M. 2004 PhRvD, 70, 123008 Van Waerbeke, L., Mellier, Y., Erben, T. et al. 2000, A&A, 358, 30 White, M., van Waerbeke, L, Mackey, J. 2002, , 575, 640 Willis, J.P., Pacaud, F., Valtchanov, I. et al 2005, , 363, 675 Wittman, D., Tyson, J.A., Margoniner, V.E., Cohen, J.G., Dell’Antonio, I.P. 2001 , 557, L89. Wittman, D, Tyson, J.A., Dell’Antonio, I.P. et al. 2002, Proc, SPIE, 4836, 73 Wittman, D, Dell’Antonio, I.P, Hughes, J.P. et al. 2006, , 643, 128 Probability to obtain spurious peaks due to insufficient correction of the PSF anisotropy {#sec:imagequality} ========================================================================================= The shear induced by massive cluster of galaxies is expected to be 7$\sim$ 10 %. This is actually larger than raw ellipticities due to PSF anisotropy (2 $\sim$ 4 % refer to Fig. \[fig:fwhmee\]), but the difference is not so significant. Therefore, the correction of the anisotropy is very important to make a precise kappa map. Here, we examine the PSF anisotropy Suprime-Cam, and evaluate the effect of imperfect correction in this work. In order to characterize the anisotropy of the Suprime-Cam images, we obtained sequences of i’-band short exposures (1 min) of dense stellar fields with various telescope pointings over one night long. The seeing was mostly $\sim$ 0.65 arcsec (FWHM). The PSF anisotropy is approximated as an elliptical, and the field position dependence of the ellpticities are investigated. Two examples of such ellipticity fields are shown in Fig. \[fig:starfield\_e\_examples\]. General tendency that we note is that the fields can be represented as a super-position of (1) the radial field at four corners (almost invariant) and (2) almost unidirectional field (variable). The variable components is most likely due to the telescope shaking whereas the invariant component can be explained by the optical aberration of the corrector. Slight asymmetry of the radial component is visible (i.e. ellipticities near the lower left corner is larger than other corners), and this would be a sign of imperfect optical alignment. We also note that the discontinuity of ellipticities is not visible across the boundary of the CCDs, and the ellipticities can be represented as a single continuous function of field position. We adopt 5-th order polynomial function here. In order to simulate the actual science field analysis, we randomly select controls stars and “galaxy role” stars from the star catalogs with an appropriate density; 2 arcmin$^{-1}$ and 40 arcmin$^{-1}$, respectively. Using the control stars, we obtain the best fit coefficients of the polynomial, and the anisotropy of “galaxy role” stars is corrected using Eqn. \[Psmcorrection\]. Because the intrinsic ellipticities of “galaxy role” stars is all zero, the residual ellipticities after the correction is a estimate of incompleteness of the correction. Fig \[fig:radreal\](a) shows the azimuthally averaged ellipticities of stars; before the correction, the typical 3 % raw ellipticity is seen. It might be interesting to note that the ellipticities of 3 %. can be induced by rms pointing error of merely 0.1 arcsec under the seeing of 0.65 arcsec. The ellipticities is reduced down to about 0.75 % after the correction. Beyond the field angle of 18 arcmin, the correction does not work fine and we decided to eliminate the field r $>$ 18 from lensing analysis; which, however, results in only a few percent loss of FOV. We estimated a PSF anisotropy caused by the optical aberration using a ray-tracing code, [zemax]{}. The calculated PSF is convolved with 0.65 arcsec FWHM gaussian, and the shape is evaluated by elliptical. The result is shown in filled square in Fig \[fig:radreal\](a). Compared with this ideal case, the observed ellipticities is large even after the correction. This shows a limit of the adoption of single polynomial function as a representatives of the PSF anisotropy, where the residual is still locally correlated correlated and not completely random. We now want to evaluate the impact of the incompleteness onto the $\kappa$ S/N map. The map is created based by the residual ellipticities after the correction. We simulate the galaxy ellipticities using the following “conversion” formula [@hoekstra04], $$\label{conversion} e^{gal} = \frac{P_{sm}^{gal}}{P_{sm}^{}}e^{},$$ which is essentially a sensitivity correction against the PSF anisotropy. This is necessary because galaxies are larger and their shape is more insensitive to the deformation compared with stars. We evaluate $\frac{P_{sm}^{gal}}{P_{sm}^{}}$ with $\frac{\langle Tr(P_{sm}^{gal})\rangle}{\langle Tr(P_{sm}^{})\rangle}$ where we adopt $\langle Tr(P_{sm}^{gal}) \rangle = 0.1 $ which is a typical value under the typical 0.7 arcsec seeing. Since the $\langle Tr(P_{sm}^{})\rangle \sim 0.2 $ here, the conversion factor is roughly $\frac{1}{2}$. We omit $P_{\gamma}$ correction because it cancels out in this case when we calculate the S/N. We created twelve such maps from independent exposures, and co-added the peak distribution functions to obtain Fig \[fig:radreal\](b). It is obvious that the incompleteness of the anisotropy correction is quite unlikely to create any significant (say S/N $>$ 3) fake peaks. [^1]: In the actual implementation, we employ the Numerical Recipes code [polin2]{} [@pressetal93].	{ "pile_set_name": "ArXiv" }
--- abstract: \| Motivated by the closest major merger, the Antennae Galaxies (NGC4038/4039), we want to improve our genetic algorithm based modeling code [Minga]{} (Theis [@theis99]). The aim is to reveal the major interaction and galaxy parameters, e.g.orbital information and halo properties of such an equal mass merger system. Together with the sophisticated search strategy of [Minga]{}, one needs fast and reliable models in order to investigate the high dimensional parameter space of this problem. Therefore we use a restricted [N]{}-body code which is based on the approach by Toomre & Toomre ([@toomre72]), however with some refinements like consistent orbits of extended dark matter halos. Recently also dynamical friction was included to this code (Petsch [@petsch07]). While a good description for dynamical friction was found for mass ratios up to $q=1/3$ (Petsch & Theis [@petsch08]), major merger systems were only imperfectly remodeled. Here we show recent improvements for a major merger system by including mass-loss and using NFW halos.\ address: 'Institut für Astronomie, Universität Wien, Türkenschanzstraße 17, 1180 Wien, Austria' author: - 'Hanns P. Petsch$^1$' - Christian Theis title: 'Determining properties of the Antennae system - Merging ability for restricted N-body' --- Introduction ============ The Antennae System (NGC4038/4039) is the closest observable major merger. It serves as a prototype for interacting galaxies and tidally induced activity. Current high precision observations over a large range of wavelengths give insight to the dynamical and morphological properties. If one wants to perform state-of-the-art high resolution self-consistent models, a reliable set of orbital and structural parameters is required. Due to the high dimensional parameter space of this problem, one needs fast and reliable models as well as a sophisticated strategy to pick out the best one. Therefore we use the program [Minga]{} (Theis [@theis99]), where an improved restricted [N]{}-body code is coupled to a genetic algorithm. It was shown in Theis & Kohle ([@theis01]), that it is possible to reproduce the large scale features of an observed system with this method. A similar approach was successfully used by Wahde & Donner ([@wahde01]) and Ruzicka ([@ruzicka07]).\ Method ====== The introduction of dynamical friction in our restricted [N]{}-body code allowed us to model close interactions and late stages of merging. Up to a mass ratio of $q=1/3$, the orbital decay and the basic tidal features of a satellite merging with an isothermal halo can be reliably reproduced (Petsch & Theis [@petsch08]). While a mass and distance dependent description for the Coulomb logarithm together with an orientation correction of the frictional force was applicable for these minor mergers, it did not perfectly work for a major merging system. In this previous work we have focused on the determination and parameterisation of the Coulomb logarithm $\ln \Lambda$. We found the best description of the dynamical friction using a modified Chandrasekhar approach (Eq.\[petsch\_eq\_4\]). $$\frac{d \vec{v}_M}{dt} = - F(v_M,\sigma) C_f \frac{\rho M}{v_M^2} \left(\hat{v}_M \cos{(\beta)} + \hat{e}_\bot \sin{(\beta)} \right) \ln \Lambda \label{petsch_eq_4}$$ $$\ln \Lambda=\ln \left[1 + \frac{M_{\mathrm{halo}}(r_M)}{M} \right] \label{petsch_eq_6}$$ The acceleration $d\vec{v}_M / dt$ of a massive particle $M$ (“satellite”) depends on the background mass-density $\rho$, the mass of the perturber $M$ and its velocity $v_M$. For more details on the function $F(v_M,\sigma)$, refer to Binney & Tremaine ([@binney87]) Eq.(7-18). $C_f$ is a scaling factor of order unity. As galaxies have density gradients, the force might point not exactly opposite to the velocity. Therefore we introduced an orthogonal component which is adjustable via an angle $\beta$. The Coulomb logarithm $\ln \Lambda$ is the relation between the maximum impact parameter $b_{\mathrm{max}}$ and the impact parameter $b_0$ that leads to a $90^\circ$ degree deflection. We found its best parameterisation to be mass and distance dependent – cf.Eq.(\[petsch\_eq\_6\]), where $M_\mathrm{halo}(r_M)$ is the mass of the host halo enclosed within the satellite’s distance $r_M$.\ As our simulation uses static dark matter (DM) halos, the structural change due to the interaction (e.g. mass-loss and memory effect in the perturbed region) is not included. As a first test we decided to directly implement mass-loss to our restricted [N-]{}body code. At the time of the pericenter we are instantaneously reducing the mass of both interacting galaxies by the same factor. Certainly this is a very rude and heuristic approach, but it will already give a hint, if the shortcomings of the dynamical friction equation for equal mass merger can be reduced by including mass-loss.\ Another important fact is, that the internal structure of the involved systems has a large impact on the merging process. E.g. the merging time is strongly correlated to the velocity dispersion. Therefore we exchanged our static isothermal halo by a static NFW (Navarro [@navarro97]) halo. Results ======= It is shown in Fig.\[petsch\_fig\_4a\], that the internal structure of a halo has a strong impact on the merging time. If the velocity dispersion within an isothermal halo is reduced by a factor of $1.22$, the merging time can drop by a factor of up to $5$. If disk-like galaxies are used instead of pure isothermal halos, the merging time gets even smaller. In order to achieve reasonalbe merging times, we have decided to use NFW halos for future restricted N-body models, in addition they are able to describe the halo properties of a disk galaxy better than isothermal spheres. In order to quantify both processes from a stellar-dynamical point of view, we performed simulations with NFW profiles and the direct ad-hoc implementation of mass-loss described above. The results are shown in Fig.\[petsch\_fig\_4b\], for an appropriate strength of the frictional force, the radial decay can be modeled convincingly, though the implemented mass-loss model is very simple. At the moment we are improving the approach by measuring the mass-loss in self-consistent simulations and gauging the simplified approach. Conclusions and outlook ======================= In order to model observed major merger systems, we have implemented dynamical friction and a heuristic approach for the mass-loss. We have found, that NFW halos are more suitable to serve as host halos for our restricted [N-]{}body models than isothermal halos (merging time, description of disk galaxy halo). Our results show, that implementation of mass-loss is necessary when modeling $1:1$ mergers in restricted [N-]{}body models. At the moment we improve our implementation of mass-loss, which will be calibrated by measuring the mass-loss of self-consistent simulations. This work was supported by the German Science Foundation (DFG) under the grant TH 511/9-1, which is part of the DFG priority program 1177. [99]{} Binney, J., Tremaine, S. 1987, Galactic Dynamics (Princeton Univ. Press) Chandrasekhar, S. 1942, Principles of stellar dynamics (Dover: New York, 1960) Dehnen, W. 2000, ApJ 536, L39 Kuijken, K., Dubinski, J. 1995, MNRAS 277, 1341 Navarro, J.F., Frenk, C.S., & White, S.D.M. 1997, ApJ, 490, 493 Petsch, H.P. 2007, Thesis, Univ.of Vienna, Austria Petsch, H.P. & Theis, C. 2008, AN 329, 1046, ArXiv e-prints:0810.0625 Ruzicka, A., Palous, J., Theis, C. 2007, A&A 461, 155-169. Theis, C. 1999, Reviews in Modern Astronomy, Vol. 12, 309 Theis, C., & Kohle, S. 2001, A&A, 370, 365-383 Toomre, A., Toomre, J. 1972, ApJ 178, 623 Wahde, M. & Donner, K.J. 2001, A&A, 379, 115	{ "pile_set_name": "ArXiv" }
--- abstract: 'Bayesian networks are probabilistic graphical models widely employed to understand dependencies in high dimensional data, and even to facilitate causal discovery. Learning the underlying network structure, which is encoded as a directed acyclic graph (DAG) is highly challenging mainly due to the vast number of possible networks. Efforts have focussed on two fronts: constraint-based methods that perform conditional independence tests to exclude edges and score and search approaches which explore the DAG space with greedy or MCMC schemes. Here we synthesise these two fields in a novel hybrid method which reduces the complexity of MCMC approaches to that of a constraint-based method. Individual steps in the MCMC scheme only require simple table lookups so that very long chains can be efficiently obtained. Furthermore, the scheme includes an iterative procedure to correct for errors from the conditional independence tests. The algorithm offers markedly superior performance to alternatives, particularly because DAGs can also be sampled from the posterior distribution, enabling full Bayesian model averaging for much larger Bayesian networks.' author: - \| Jack Kuipers [email protected]\ Polina Suter\ D-BSSE, ETH Zurich, Mattenstrasse 26, 4058 Basel, Switzerland Giusi Moffa\ Division of Psychiatry, University College London, London, UK\ Department of Mathematics and Computer Science, University of Basel, Basel, Switzerland bibliography: - 'mcmcdags.bib' title: \| Efficient Sampling and Structure Learning\ of Bayesian Networks[^1] --- Bayesian Networks, Structure Learning, MCMC. Introduction {#intro} ============ Bayesian networks are statistical models to describe and visualise in a compact graphical form the probabilistic relationships between variables of interest. The nodes of a graphical structure correspond to the variables, while directed edges between the nodes encode conditional independence relationships between them. The most important property of the digraphs underlying a Bayesian network is that they are acyclic, i.e. there are no directed paths which revisit any node. Such graphical objects are commonly denoted DAGs (directed acyclic graphs). Alongside their more canonical use for representing multivariate probability distributions, DAGs also play a prominent role in describing causal models [@greenland99causal; @pearl00; @hr06instruments; @wr07four] and facilitating causal discovery from observational data [@mkb09; @moffaetal17], though caution is warranted in their use and interpretation [@dawid10]. However, the potential for causal discovery and uncovering the mechanisms underlying scientific phenomena in disciplines ranging from the social sciences [@elwert13graphical] to biology [@friedmanetal00; @friedman04; @kuipersetal18] has driven interest in DAG inference. To fully characterise a Bayesian network we need the DAG structure and the parameters which define an associated statistical model to explicitly describe the probabilistic relationships between the variables. To make any inference about the variables in the network, we need both components, the structure and the parameters, which we may need to estimate from data. Given sample data from the joint probability distribution on the node variables, learning the graphical structure is in general the more challenging task. The difficulty mostly rests with the mere size of the search space, which grows super-exponentially with the number of nodes $n$, since the logarithm grows quadratically [@robinson70; @robinson73]. A curious illustration of this growth is that the number of DAGs with 21 nodes is approximately the estimated number of atoms in the observable universe ($\approx10^{80}$). Bayesian network notation ------------------------- Bayesian networks represent a factorisation of multivariate probability distributions of $n$ random variables ${\boldsymbol{X}}= \{ X_1, \ldots, X_n\}$ by encoding conditional dependencies in a graphical structure. A Bayesian network ${\mathcal{B}}= ({\mathcal{G}}, {\theta})$ consists of a DAG ${\mathcal{G}}$ and an associated set of parameters ${\theta}$ which define the conditional distribution $P(X_i \mid {\mbox{{\bf Pa}}}_i)$ of each variable $X_i$ on its parents ${\mbox{{\bf Pa}}}_i$. The distribution represented by the Bayesian networks is then assumed to satisfy the Markov property [described for example in @bk:kf09] that each variable $X_i$ is independent of its non-descendants given its parents ${\mbox{{\bf Pa}}}_i$, allowing the joint probability distribution to factorise as $$\label{distfactorisation} P( X_1, \ldots, X_n ) = \prod_i^n P(X_i \mid {\mbox{{\bf Pa}}}_i) \,$$ Learning the parameters ${\theta}$ which best describe a set of data ${D}$ for a given graph ${\mathcal{G}}$ is generally straightforward for complete data, with the main difficulty in learning the structural dependence in ${\boldsymbol{X}}$ and the DAG ${\mathcal{G}}$ itself. Due to the symmetry of conditional independence relationships, the same distribution might factorise according to different DAGs. DAGs encoding the same probability distribution constitute an equivalence class: they share the v-structures [two unconnected parents of any node, @vp90] and the skeleton (the edges if directions were removed). The equivalence class of DAGs can be represented as an essential graph [@amp97] or a completed partially directed acyclic graph (CPDAG) [@chickering02CPDAG]. Based purely on probabilistic properties, Bayesian networks can therefore only be learned from data up to an equivalence class. DAG posteriors {#sec:DAGposterior} -------------- In inferring Bayesian networks, the dual challenge consists of learning the graph structure (or its equivalence class) which best fits and explains the data ${D}$, and accounting for the uncertainty in the structure and parameters given the data. A natural strategy in Bayesian inference consists of sampling and averaging over a set of similarly well fitting networks. Each DAG ${\mathcal{G}}$ is assigned a score equal to its posterior probability given the data ${D}$ $$\nonumber P( {\mathcal{G}}\mid {D}) \propto P({D}\mid {\mathcal{G}}) P({\mathcal{G}}) \,$$ where the likelihood $P({D}\mid {\mathcal{G}})$ has been marginalised over the parameter space. When the graph and parameter priors satisfy certain conditions of structure modularity, parameter independence and parameter modularity [@hg95; @friedmanetal00; @fk03] then the score decomposes as $$\label{scoredecomp} P( {\mathcal{G}}\mid {D}) \propto P(D \mid {\mathcal{G}}) P({\mathcal{G}}) = \prod_{i=1}^{n} {S}(X_i, {\mbox{{\bf Pa}}}_i \mid {D}) \, ,$$ involving a function ${S}$ which depends only on a node and its parents. For discrete categorical data, a Dirichlet prior is the only choice satisfying the required conditions for decomposition, leading to the BDe score of [@hg95]. For continuous multivariate Gaussian data, the inverse Wishart prior leads to the BGe score ([@gh02]; corrected in [@cr12; @kmh14]). In this manuscript we focus on the continuous case with the BGe score when evaluating the complexity of our approach, and discuss the discrete categorical case in [Appendix \[catdata\]]{}. State of the art structure learning ----------------------------------- Traditional approaches to structure learning fall into two categories (and their combination): - constraint-based methods, relying on conditional independence tests - score and search algorithms, relying on a scoring function and a search procedure Below we provide a brief review of these concepts and algorithms of each class pertinent to this work. ### Constraint-based methods Constraint-based methods exploit the property of Bayesian networks that edges encode conditional dependencies. If a pair of variables can be shown to be independent of each other when conditioning on at least one set (including the empty set) of the remaining variables, then a direct edge between the corresponding nodes in the graph can be excluded. The most prominent example of constraint-based methods is the well known PC algorithm of [@bk:sgs00], more recently popularised by [@art:KalischB2007], who provided consistency results for the case of sparse DAGs and an [[R]{}]{} [@R17] implementation within the [[pcalg]{}]{} package [@art:KalischMCMB2012]. Rather than exhaustively search the $2^{(n-2)}$ possible conditioning sets for each pair of nodes, the crucial insight of the PC algorithm is to perform the tests in order of increasing complexity. Namely, starting from a fully connected (undirected) graph, the procedure tests marginal independence (conditioning on the empty set) for all pairs of nodes. Then it performs pairwise conditional independence tests between pairs of node which are still directly connected, conditioning on each adjacent node of either node in the pair, and so on conditioning on larger sets. Edges are always deleted when a conditional independence test cannot be rejected. This strategy differs from the typical use of hypothesis testing since edges are assumed to be present by default, but the null hypothesis is taken as conditional independence. Edges which are never deleted through the process form the final skeleton of the PC algorithm. The conditional independencies which are not rejected identify all v-structures of the graph, fixing the direction of the corresponding edges. At this point it may still be possible to orient some edges, to ensure that no cycles are introduced and no additional v-structures are created [@meek95]. The algorithm finally returns the CPDAG of the Markov equivalence class consistent with the conditional dependencies compatible with the data. In the final skeleton, for each node the remainder of its adjacent neighbourhood will have been conditioned upon. For the node with largest degree $K$, at least $K2^{K-1}$ tests will have been performed and the algorithm is of exponential complexity in the largest degree. In the best case the algorithm may still run with $K$ around 25–30, but in the worst case the base can increase giving a complexity bound of $O(n^K)$ making the algorithm infeasible even for low $K$ for large DAGs [@art:KalischB2007]. For sparse large graphs, the PC algorithm can be very efficient. Despite the efficiency, since edges can only be deleted and many (correlated) independence tests are performed, the PC algorithm tends to have a high rate of false negatives and hence lose edges, so that it finds only a fraction of those in the true network [@uhleretal13]. Increasing the threshold for the conditional independence tests does little to alleviate the problem of false negatives while increasing runtime substantially. Another aspect of the problem of repeated tests is that the output of the PC algorithm can depend on the order of the tests (or the ordering of the input data), leading to unstable estimates in high dimensions, although modifications have been proposed to mollify this effect [@cm14]. ### Score and search methods On the other side of the coin are score and search methods and MCMC samplers. Each DAG gets a score, typically a penalised likelihood or a posterior probability ([Section \[sec:DAGposterior\]]{}). An algorithm then searches through the DAG space for the structure (or ensemble thereof) which optimises the score, or to return a sample proportional to the score. The most basic sampler is structure MCMC where each step involves adding, deleting or reversing an edge [@my95; @gc03] and accepting the move according to a Metropolis-Hastings probability. The scheme is highly flexible – amplifying the score leads to simulated annealing while modulating the amplification through the acceptance rate leads to adaptive tempering, to speed up the traversal and exploration of the DAG space. Sampling from the neighbourhood of DAGs with one edge added, removed or reversed, proportional to their scores results in faster MCMC convergence [as for example in @jc18]. Greedy hill climbing instead proceeds by choosing the highest scoring DAG in the neighbourhood as the new starting point. Greedy equivalence search (GES) [@chickering02] is a popular algorithm for a greedy search approach on the space of Markov equivalent DAGs. When the score is decomposable (as in [Eq. (\[scoredecomp\])]{}), only nodes whose parents change need to be rescored providing an $O(n)$ speedup for structure based methods. Since the decomposability in [Eq. (\[scoredecomp\])]{} mimics the factorisation in [Eq. (\[distfactorisation\])]{} it is a property possessed by commonly used scores like the BIC penalised likelihood or the BDe or BGe, and an essential property for more advanced algorithms. A direction spearheaded by order MCMC [@fk03] reduces the search space by combining large collections of DAGs together, namely all DAGs sharing the same topological ordering of the nodes. The score of the order consists of the sum of the scores of the DAGs consistent with it, and an MCMC scheme runs on the space of orders which are simply permutations of the $n$ nodes. Although the number of DAGs compatible with each order also grows super-exponentially, the sum of all their scores involves $\approx 2^{n}$ different contributions and can be evaluated in exponential time [@buntine91]. For larger $n$ computations become quickly prohibitive and the complexity is artificially reduced to polynomial $O(n^{K+1})$ by imposing a hard limit $K$ on the number of parents allowed for each node. Nevertheless the strategy of combining large sets of DAGs and working on the much smaller (though still factorial) space of orders, enormously improves convergence with respect to structure MCMC, allowing the search and sampling of much larger graphs (for moderate or low $K$). Score decomposability is also necessary for order-based greedy search [@tk05] as well as for dynamic or integer linear programming methods [@ks04; @em07; @htw16; @cussens11; @cussensetal17] for structure learning. For Bayesian model averaging, one limitation with order MCMC derives from the fact that DAGs may belong to multiple orders, introducing a bias in the sampling. The bias can be avoided by working on the space of ordered partitions [@km17] which provide a unique representation of each DAG. Other MCMC alternatives include large scale edge reversal [@gh08] and Gibbs sampling [@gm16] moves. Unbiased MCMC schemes, such as these, are currently the only viable approaches to sampling and accounting for structure uncertainty, though still limited to smaller or relatively sparse graphs. As the size and connectivity of the target DAGs increase, the wide spectrum of constraint-based and score and search algorithms, cannot but fail to converge or discover optimally scoring graphs. To limit the extent to which the search space grows with the number of nodes, [@fnp99] pruned it by only allowing edges from selected candidate parents and performing a greedy search in the restricted search space. They also iteratively updated the set of candidate parents based on the current best DAG discovered. With the same aim of limiting the search space, [@tba06] brought together the ease of conditional independence testing and the performance of DAG searching, to benefit from their individual advantages. First a constraint-based method, akin to the PC algorithm, identifies a (liberal) undirected skeleton. A greedy search then acts on the restricted search space defined by excluding edges which are not included in the reference skeleton. Since score and search, when feasible, tends to perform better [@heckerman2006bayesian] than constraint-based methods, the hybrid approach of [@tba06] outperformed previous methods. [@nhm18] recently investigated the consistency properties of hybrid approaches using GES in high dimensional settings. Original contribution --------------------- In this work, we bring the power and sophistication of order and partition MCMC to the hybrid framework for structure learning. The result is a highly efficient algorithm for search and sampling with marked improvements on current state of the art methods. The key is to observe that the exponential complexity in $K$ of $n^{K}$ for order or partition MCMC [@fk03; @km17] derives from allowing among the potentially $K$ parents of each node any of the other $(n-1)$. If the set of $K$ potential parents is pre-selected, for example through a constraint-based method relying on conditional independence tests, the complexity reduces to $2^K$ for searching of an optimal structure, and $3^K$ for unbiased structure sampling. The complexity of the search then matches that of the testing component of the PC algorithm. Along with the standard pre-computation of parents set scores, which are exponentially costly, we introduce a method to also precompute tables of partial sums of parent set score with no complexity overhead. In particular we tabulate every score quantity needed for the MCMC scheme. During each MCMC step we then simply need to look up the relevant scores providing a very efficient sampler. A distinctive feature of our method is not to fully trust the search space provided by the initial constraint-based method. Each node is entitled to have as parent any of the permissible nodes in the search space, and an additional arbitrary one from outside that set. Accounting for the expansion to the potential parent set, each MCMC step takes an expected time of order $O(K)$, despite scoring vast sets of DAGs at a time, and therefore comparable or even lower complexity than structure MCMC moves [between $O(n)$ and $O(n^2)$, though some rejected moves can be $O(1)$; @gc03]. The expansion beyond the skeleton provides a mechanism to iteratively improve the search space until it includes the maximally scoring DAG encountered, or the bulk of the posterior weight. Based on our simulations the results strongly outperform currently available alternatives, enabling efficient inference and sampling of much larger DAGs. In [Section \[sec:order\]]{} we develop efficient algorithms for order-based sampling for a known search space and examine their convergence. Then in [Section \[sec:spacemax\]]{} we demonstrate how to improve the search space iteratively, and show how this offers notably improved performance in finding the best DAG. Finally, we extend the scheme to the space of partitions to provide an unbiased sampler in [Section \[sec:partition\]]{}, with discussions in [Section \[discussion\]]{}. The algorithms introduced here are all implemented in the [[R]{}]{} [@R17] package [[BiDAG]{}]{}. Order-based DAG sampling on a given search space {#sec:order} ================================================ In the order MCMC algorithm of [@fk03], the $n$ nodes of a DAG are arranged in topological order ${\prec}$. We associate a permutation $\pi_{{\prec}}$ with each order. For a DAG to be compatible with an order, the parents of each node must have a higher index in the permutation $${\mathcal{G}}\in {\prec}\overset{\mathrm{def}}{\iff} \forall i \, , \; \forall \left\{j : X_j \in {\mbox{{\bf Pa}}}_i \right\} , \; \pi_{{\prec}}[i] < \pi_{{\prec}}[j]$$ Visually, when we place the nodes in a linear chain from left to right according to $\pi_{{\prec}}$, edges may only come from nodes further to the right ([Figure \[fig:dagsorderpartition\]]{}b). With the rows and columns labelled following $\pi_{{\prec}}$, the adjacency matrix of a compatible DAG is lower triangular so that a total of $2^{\binom{n}{2}}$ DAGs are compatible with each order. ![The DAG in (a) is compatible with the order depicted in (b), as edges only originate from nodes further right in the chain, along with 8 other orders. The DAG however can be uniquely assigned to a labelled partition by collecting outpoints into partition elements to arrive at the representation in (c).[]{data-label="fig:dagsorderpartition"}](./dagsorderpartition){width="\textwidth"} Order MCMC ---------- The idea of order MCMC is to combine all DAGs consistent with an order to reduce the problem to the much smaller space of permutations instead of working directly on the space of DAGs. Each order ${\prec}$ receives a score ${R}({\prec}\mid {D})$ equal to the sum of the scores of all DAGs in the order $${R}({\prec}\mid {D}) = \sum_{{\mathcal{G}}\in {\prec}} P( {\mathcal{G}}\mid {D}) \,$$ Instead of naively scoring all DAGs in an order, [@fk03] used the factorisation in [Eq. (\[scoredecomp\])]{} $$\label{orderScore} {R}({\prec}\mid {D}) \propto \sum_{{\mathcal{G}}\in {\prec}} \prod_{i=1}^{n} {S}(X_i, {\mbox{{\bf Pa}}}_i \mid {D}) = \prod_{i=1}^{n} \sum_{{\mbox{{\bf Pa}}}_i \in {\prec}} {S}(X_i, {\mbox{{\bf Pa}}}_i \mid {D}) \,$$ to exchange the sum and product [following @buntine91]. The sum is restricted to parent subsets compatible with the node ordering $${\mbox{{\bf Pa}}}_i \in {\prec}\overset{\mathrm{def}}{\iff} \forall \left\{j : X_j \in {\mbox{{\bf Pa}}}_i \right\} , \; \pi_{{\prec}}[i] < \pi_{{\prec}}[j]$$ The score of the order therefore reduces to sums over all compatible parent subsets, eliminating the need of summing over DAGs. For a node with $k$ possible parents further along the order, there are $2^k$ parents subsets. Evaluating the score of the order therefore requires $\sum_{k=0}^{n-1}2^{k}=(2^{n}-1)$ evaluations of the score function ${S}$. This provides a massive reduction in complexity compared to scoring all $2^{\binom{n}{2}}$ DAGs in the order individually. The exponential complexity in $n$ is still too high for larger DAGs so a hard limit $K$ on the size of the parent sets is typically introduced to obtain polynomial complexity of $O(n^{K+1})$ evaluations of ${S}$. For larger DAGs however, $K$ must be rather small in practice, so that the truncation runs the risk of artificially excluding highly scoring DAGs. As a remedy, we start by defining order MCMC on a given search space, which may be selected for example based on prior subject knowledge or from a skeleton derived through constraint-based methods. Restricting the search space ---------------------------- The search space can be defined by a directed graph ${\mathcal{H}}$, which is not necessarily acyclic, or its adjacency matrix, ${H}$: $${H}_{ji} = 1 \mbox{ if } \{j,i\} \in {\mathcal{H}}\,$$ One advantage with respect to simply using an undirected skeleton, which corresponds to a symmetric matrix, is that the directed graph naturally allows for prior information about edge directionality to be included. Prior beliefs about undirected edges can also be included by treating both directions equally. In the search space, each node has the following set of permissible parents $${\boldsymbol{h}}^i = \{X_j : H_{ji} =1 \} \,$$ For a set of size $K$ we evaluate the score of each possible combination and store the scores in a table (as in the example on the left of [Table \[scoretableexample\]]{}) as is standard practice. Since there are $2^{K}$ possible combinations, and for the BGe score each involves taking the determinant of a matrix, the complexity of building this table is $O(K^{3}2^{K})$. For convenience later, we label the different scores with a binary mapping from a parent subset ${\boldsymbol{Z}}$ to the integers: $$f({\boldsymbol{Z}})=\sum_{j=1}^{K}I({h}^{i}_j \in {\boldsymbol{Z}})2^{j-1}$$ using the indicator function $I$. Efficient order scoring ----------------------- To score all the DAGs compatible with a particular order we still need to select and sum the rows in the score tables where the parent subset respects the order $$\label{orderScoresdg} {R}_{{\mathcal{H}}}({\prec}\mid {D}) \propto \prod_{i=1}^{n} \sum_{\substack{{\mbox{{\bf Pa}}}_i \subseteq {\boldsymbol{h}}^i \cr {\mbox{{\bf Pa}}}_i \in {\prec}} } {S}(X_i, {\mbox{{\bf Pa}}}_i \mid {D}) \,$$ with the additional constraint that all elements in the parent sets considered must belong to the search space defined by ${\mathcal{H}}$. From the precomputed score table of all permissible parent subsets in the search space, we select those compatible with the order constraint. Simply running through the $2^K$ rows takes exponential time (in $K$) for each node. Unlike other order-based schemes, we can avoid this by building a second table: the summed score table (see the example on the right of [Table \[scoretableexample\]]{}). In [Appendix \[sumtableapp\]]{} we detail the algorithmic steps which allow us to build the summed score table for each variable with a complexity of $O(K^{2}2^{K})$. Compared to the complexity of building the original score table of $O(K^{3}2^{K})$, this step adds no complexity overhead to the method. The summed score table however provides the means to efficiently score each order. For each node we look up the relevant row in the summed score table and by moving linearly along the order we can compute [Eq. (\[orderScoresdg\])]{} in $O(Kn)$ as $$\label{orderScoresdgtable} {R}_{{\mathcal{H}}}({\prec}\mid {D}) \propto \prod_{i=1}^{n} {\Sigma}^{i}_{f(\substack{{\boldsymbol{h}}^i \in {\prec}})} \,$$ where $${\boldsymbol{h}}^i \; \in {\prec}\; \overset{\mathrm{def}}{=} \; \left\{X_j \in {\boldsymbol{h}}^i : \pi_{{\prec}}[i] < \pi_{{\prec}}[j] \right\}$$ are the elements of ${\boldsymbol{h}}^i$ compatible with the order constraints and $${\Sigma}^{i}_{f(\substack{{\boldsymbol{h}}^i \in {\prec}})} = \sum_{\substack{{\mbox{{\bf Pa}}}_i \subseteq {\boldsymbol{h}}^i \cr {\mbox{{\bf Pa}}}_i \in {\prec}} } {S}(X_i, {\mbox{{\bf Pa}}}_i \mid {D})$$ is the sum of scores of all the parent sets in the search space respecting the order, which is precomputed using [Algorithm \[parentsumalg\]]{}. Order MCMC moves {#sec:orderMCMCmoves} ---------------- The strategy of order MCMC is to build a Markov chain on the space of orders. From the order ${\prec}_{t}$ at the current iteration ${t}$, a new order ${\prec}'$ is proposed and accepted with probability $$\label{orderMHaccprob} \rho = \min \left \{ 1 , \frac{{R}_{{\mathcal{H}}}({\prec}' \mid {D}) }{{R}_{{\mathcal{H}}}({\prec}_{t}\mid {D})} \right \} \,$$ to provide a chain with a stationary distribution proportional to the score ${R}_{{\mathcal{H}}}({\prec}\mid {D})$ for symmetric proposals. The standard proposal [@fk03] is to swap two nodes in the order while leaving the others fixed. We will denote this move as a global swap. The set of permissible parents of all nodes between the two swapped ones may change, requiring them to be rescored. Computing the score of the proposed order then has a complexity $O(n)$. It is possible to move from one order to any other in $(n-1)$ steps (assuming non-zero order scores), making the chain irreducible. A more local move of only transposing two adjacent nodes allows the proposed order to be scored in $O(1)$. We will denote this move as a local transposition which takes $O(n^2)$ steps to have access to any order. ### A Gibbs move in order space ![For any randomly chosen node, here 4, we score the entire neighbourhood of new positions in the order by performing a sequence of local transpositions and include the known score of the current position. The new placement is sampled proportionally to the scores of the different orders in the neighbourhood.[]{data-label="ordermovepic"}](./ordermove){width="50.00000%"} On the space of orders we define a new move with the same complexity of the standard global swap move and denote it as a node relocation. From the current state in the chain, ${\prec}_{t}$, first sample a node uniformly from the $n$ available, say node $i$. For the move we sample the position of node $i$ conditional on keeping the relative ordering of the remaining nodes unchanged. Define the neighbourhood of the order under this move, ${\mathrm{nbd}^{i}({\prec}_{t})}$, to be all orders with node $i$ placed elsewhere in the order or at its current position, as in the example in [Figure \[ordermovepic\]]{} with node 4 chosen. To move through the full neighbourhood of size $n$, we can sequentially transpose node $i$ with adjacent nodes. Since each local transposition takes a time $O(1)$ to compute the score of the next order, scoring the whole neighbourhood takes $O(n)$. Finally sample a proposed order ${\prec}'$ proportionally to the scores of all the orders in the neighbourhood. As a consequence the move is always accepted and the next step of the chain set to the proposed order ${\prec}_{{t}+1}={\prec}'$. We summarise the move in [Algorithm \[ordermovealg\]]{}. input The order ${\prec}_{t}$ at step ${t}$ of the chain Sample node $i$ uniformly from the $n$ Build and score all orders in the neighbourhood ${\mathrm{nbd}^{i}({\prec}_{t})}$,\ $\quad$ through consecutive local transpositions of node $i$ Sample proposed order ${\prec}'$ from ${\mathrm{nbd}^{i}({\prec}_{t})}$ proportionally to ${R}_{{\mathcal{H}}}({\prec}' \mid {D})$ Set ${\prec}_{{t}+1}={\prec}'$ return ${\prec}_{{t}+1}$ The newly defined node relocation move satisfies detailed balance $$\label{detailedbalance} P({\prec}' \mid {\prec}) {R}_{{\mathcal{H}}}({\prec}\mid {D}) = P({\prec}\mid {\prec}') {R}_{{\mathcal{H}}}({\prec}' \mid {D}) \,$$ where $P({\prec}' \mid {\prec})$ is the transition probability from ${\prec}$ to ${\prec}'$. The transition involves first sampling a node $i$ and then the order proportionally to its score so that (for orders not connected by a local transposition) $$\label{transprobeqn} P({\prec}' \mid {\prec}) = \frac{1}{n}\frac{{R}_{{\mathcal{H}}}({\prec}' \mid {D})}{\sum_{{\prec}'' \in {\mathrm{nbd}^{i}({\prec})}}{R}_{{\mathcal{H}}}({\prec}'' \mid {D})} \,$$ The reverse move needs the same node $i$ to be selected, and as ${\mathrm{nbd}^{i}({\prec})}={\mathrm{nbd}^{i}({\prec}')}$ the denominators cancel when substituting into [Eq. (\[detailedbalance\])]{}. For orders connected by a local transposition, say node $i$ swapped with the adjacent node $j$, there are two possible paths connecting the orders and a transition probability of $$\label{transprobadjeqn} P({\prec}' \mid {\prec}) = \frac{1}{n}\frac{{R}_{{\mathcal{H}}}({\prec}' \mid {D})}{\sum_{{\prec}'' \in {\mathrm{nbd}^{i}({\prec})}}{R}_{{\mathcal{H}}}({\prec}'' \mid {D})} + \frac{1}{n}\frac{{R}_{{\mathcal{H}}}({\prec}' \mid {D})}{\sum_{{\prec}'' \in {\mathrm{nbd}^{j}({\prec})}}{R}_{{\mathcal{H}}}({\prec}'' \mid {D})}\,$$ Since the reverse move involves the same pair of nodes, we can again directly verify detailed balance by substituting into [Eq. (\[detailedbalance\])]{}. The move is aperiodic since the original order is included in the neighbourhood. It is possible to reach any order from any other by performing $(n-1)$ steps making the chain also irreducible. Therefore the newly defined move satisfies the requirements for the chain to converge and provide order samples from a probability distribution proportional to the score ${R}_{{\mathcal{H}}}({\prec}\mid {D})$. The node relocation move naturally provides a fixed scan Gibbs sampler by cycling through the nodes sequentially, rather than sampling at each step. ### Chain complexity We mix the three moves into a single scheme. Since the global swap involves rescoring $\approx\frac{n}{3}$ nodes on average at each step, while the local transposition involves $2$ and the node relocation $2n$ we can keep the average complexity at $O(1)$ if we sample the more expensive moves with a probability $\propto \frac{1}{n}$. In this way, we can also balance their computational costs. For simplicity we assign each move equal average computational time by selecting them with a probability of $(\frac{6}{n},\frac{n-7}{n}, \frac{1}{n})$ respectively. With the mixture of moves, the number of steps to reach any order is $O(n)$ so following the heuristic reasoning of [@km17] we would expect convergence of the chain to take $O(n^2\log(n))$ steps. This complexity is consistent with our simulation results ([Figure \[fig:R2convergence\]]{}). Once the score tables have been computed, the complexity of running the whole chain is also $O(n^2\log(n))$. Utilising the lookup table of summed scores therefore reduces the complexity substantially by a factor of $O(n^K)$ compared to standard order MCMC where one simply restricts the size of parent sets to $K$ and only utilises the basic score tables. ### DAG sampling To draw a DAG consistent with a given sampled order, we can sample the parents of each node proportionally to the entries respecting the order in the score table. Complexity remains exponential, of $O(2^{K})$, so the scheme should be thinned such that DAG sampling only happens periodically in the order chain. The frequency should be set so that sampling takes computational time at most comparable to running the order chain inbetween. Extending the search space {#sec:expand} -------------------------- The restricted search space, derived for example through constraint-based methods, may exclude relevant edges. To address this problem, we extend our approach by softening the restrictions. In particular we allow each node to have one additional parent from among the nodes outside its permissible parent set. The score of each order becomes $$\label{orderScoresdgx} {R}_{{\mathcal{H}}}^{+}({\prec}\mid {D}) \propto \prod_{i=1}^{n} \sum_{\substack{{\mbox{{\bf Pa}}}_i \subseteq {\boldsymbol{h}}^i \cr {\mbox{{\bf Pa}}}_i \in {\prec}} } \Bigg[{S}(X_i, {\mbox{{\bf Pa}}}_i \mid {D}) + \sum_{\substack{X_j \notin {\boldsymbol{h}}^i \cr \pi_{{\prec}}[i] < \pi_{{\prec}}[j]} }{S}(X_i, \{{\mbox{{\bf Pa}}}_i, X_j\} \mid {D}) \Bigg] \,$$ For the efficient computation of the sum, we build a score table for each node and each additional parent. For a node with $K$ parents this leads to $(n-K+1)$ tables in total and we perform [Algorithm \[parentsumalg\]]{} of [Appendix \[sumtableapp\]]{} on each of them. The time and space complexity of building these tables is therefore an order $n$ more expensive than using the restricted search space. Given the tables, however, an order can again be scored with a simple lookup $$\label{orderScoresdgxtable} {R}_{{\mathcal{H}}}^{+}({\prec}\mid {D}) \propto \prod_{i=1}^{n} \Bigg[ {\Sigma}^{i}_{f(\substack{{\boldsymbol{h}}^i \in {\prec}})} + \sum_{\substack{X_j \notin {\boldsymbol{h}}^i \cr \pi_{{\prec}}[i] < \pi_{{\prec}}[j]} }{\Sigma}^{ij}_{f(\substack{{\boldsymbol{h}}^i \in {\prec}})} \Bigg] \,$$ where we also index the summed scores with the additional parent $${\Sigma}^{ij}_{f(\substack{{\boldsymbol{h}}^i \in {\prec}})} = \sum_{\substack{{\mbox{{\bf Pa}}}_i \subseteq {\boldsymbol{h}}^i \cr {\mbox{{\bf Pa}}}_i \in {\prec}} } {S}(X_i, \{{\mbox{{\bf Pa}}}_i, X_j\} \mid {D}) \, , \qquad X_j \notin {\boldsymbol{h}}^i$$ The complexity of scoring the order is $O(n^2)$. ### Move complexity For the local transposition where we swap two adjacent nodes in the order, if neither is in the permissible parent set of the other we simply update one element of the sum in [Eq. (\[orderScoresdgxtable\])]{} in $O(1)$. If either is in the permissible parent set, all terms need to be replaced in $O(n)$. However since the nodes have up to $K$ parents, the probability of a permissible parent being affected is $\propto \frac{K}{n}$ giving an average complexity of $O(K)$. For the global swap the maximum complexity is $O(n^2)$ when many of the intermediate nodes have among their permissible parents either of the swapped nodes, but on average the complexity is $O(Kn)$ following the same argument as above. The node relocation move has the same complexity, so that the weighted mixture of moves typically takes just $O(K)$. Convergence {#sec:convergence} ----------- To examine the convergence of our MCMC scheme, we ran simulations as described in [Appendix \[app:simulations\]]{}, and for each simulation repetition we compared two independent runs with different random initial points in the final search space for each dataset. We computed the squared correlation $\rho^2$ of the two runs between the posterior probabilities of each edge, after excluding a burn-in period of 20%. Since most edges are absent, only edges with a posterior probability greater than 5% in at least one run were included. The median $\rho^2$ along with the first and third quartiles are displayed in [Figure \[fig:R2convergence\]]{}. By scaling the number of MCMC steps with $n^2\log(n)$ we observe the correlation approaching 1. To examine the behaviour in more detail, we can consider $(1-\rho^2)^{-1}$ which increases roughly linearly with the scaled number of steps. We see, especially, that there is little dependence on the size of the network, apart from a slower convergence for the smallest network size. With reasonable consistency, and importantly no obvious decrease in scaled performance as the number of variables $n$ increases, the simulation results of [Figure \[fig:R2convergence\]]{} are in line with the estimate of requiring $O(n^2\log(n))$ steps for the MCMC convergence. Similar insensitivity to the network size is also visible for the root mean square error between runs ([Figure \[fig:RMSEconvergence\]]{}). ![The correlation between edge probabilities from different runs as the size of the network increases. The simulation setting is described in [Appendix \[app:simulations\]]{}. The number of steps in the chain is scaled by $n^2\log(n)$. The transformation on the right highlights the roughly linear improvement in the convergence measure with the number of steps and that there is little dependence on the network size.[]{data-label="fig:R2convergence"}](./R2convergence){width="85.00000%"} Search space and maximal DAG discovery {#sec:spacemax} ====================================== In order to sample DAGs effectively, our search space needs to cover the bulk of the posterior weight. We describe here an iterative scheme to search for the highest scoring DAG in the current extended search space, and use it to update and improve the search space itself. Maximal DAG discovery --------------------- In addition to sampling DAGs, we can also employ the MCMC scheme to search for the maximally scoring or maximum a posteriori (MAP) DAG. To this end [and analogously to @tk05] we replace the score of each order by the score of the highest scoring DAG in that order $$\label{orderScoresdgmax} {Q}_{{\mathcal{H}}}({\prec}\mid {D}) = \max_{\substack{{\mathcal{G}}\subseteq {\mathcal{H}}\cr {\mathcal{G}}\in {\prec}}} P( {\mathcal{G}}\mid {D}) \propto \prod_{i=1}^{n} \max_{\substack{{\mbox{{\bf Pa}}}_i \subseteq {\boldsymbol{h}}^i \cr {\mbox{{\bf Pa}}}_i \in {\prec}} } {S}(X_i, {\mbox{{\bf Pa}}}_i \mid {D}) \,$$ To compute the terms on the right we again follow the steps detailed in [Appendix \[sumtableapp\]]{} using the Hasse power set representation of the permissible parent set of each node and propagating the maximum of scores down the power set following [Algorithm \[parentmaxalg\]]{} of [Appendix \[sumtableapp\]]{}: $$\label{orderScoresdgmaxtable} {Q}_{{\mathcal{H}}}({\prec}\mid {D}) \propto \prod_{i=1}^{n} {M}^{i}_{f(\substack{{\boldsymbol{h}}^i \in {\prec}})} \, , \qquad {M}^{i}_{f(\substack{{\boldsymbol{h}}^i \in {\prec}})} = \max_{\substack{{\mbox{{\bf Pa}}}_i \subseteq {\boldsymbol{h}}^i \cr {\mbox{{\bf Pa}}}_i \in {\prec}} } {S}(X_i, {\mbox{{\bf Pa}}}_i \mid {D})$$ ### Stochastic search Finding the order with the highest ${Q}$ directly provides a MAP DAG. A stochastic search based on the order MCMC scheme with score ${Q}_{{\mathcal{H}}}({\prec}\mid {D})$ can tackle the problem. Running the scheme, we keep track of the highest scoring order, and hence the highest scoring DAG, encountered. The convergence time to sample orders from a distribution proportional to ${Q}_{{\mathcal{H}}}({\prec}\mid {D})$ is again expected to be $O(n^2\log(n))$. To perform adaptive tempering and speed up discovery of a MAP DAG, we can transform the score by raising it to the power of $\gamma$ and employ ${Q}_{{\mathcal{H}}}({\prec}\mid {D})^{\gamma}$. This transformation smooths ($\gamma<1$) or amplifies ($\gamma>1$) the score landscape and the value of $\gamma$ can be tuned adaptively depending on the acceptance probability of the MCMC moves while running the algorithm. To effectively explore local neighbourhoods, the target acceptance of swaps may scale $\propto \frac{1}{n}$. Alternatively, simulated annealing can be performed by sending $\gamma\to\infty$. ### Greedy search The order-based scheme can also be adapted to perform a greedy search [@tk05]. For example we score all possible $(n-1)$ local transpositions of adjacent nodes in $O(n)$ and select the highest scoring at each step. Since it takes $O(n^2)$ steps to be able to reach any order with this move, we would expect $O(n^3)$ complexity to find each local maximum. For the global swap of two random nodes, scoring the neighbourhood itself is $O(n^3)$ so that the $O(n)$ to traverse the space makes this move more expensive than local transpositions. Local transpositions would therefore be generally preferable for greedy search, although global swaps may be useful to escape from local maxima. The new node relocation move of moving a single node at a time ([Figure \[ordermovepic\]]{}) requires only $O(n^2)$ to score all the possible new placements of all nodes. With $O(n)$ steps to move between any pair of DAGs, we are again looking at $O(n^3)$ complexity for each search. The new move also contains all local transpositions in its neighbourhood and so provides a complementary alternative to a greedy search scheme purely based on local transpositions. Iteratively improving the search space -------------------------------------- Extending the search space, so that each node may have an additional parent outside, allows us to discover edges which improve the score of DAGs, or edges with a high posterior weight (those which occur in a large fraction of sampled DAGs), which were previously excluded from the core search space defined by ${\mathcal{H}}$. Incorporating these edges into the core search space, we can iteratively improve the search space. This is analogous to the iterative updating of [@fnp99], but adapted to the order-based setting. Starting with the original core search space ${\mathcal{H}}_0 = {\mathcal{H}}$ we expand to allow one additional parent and search for the maximally scoring DAG in that space and convert it to the CPDAG ${\mathcal{G}}^{}_0$. For the next core search space ${\mathcal{H}}_1$ we take the union of the edges in ${\mathcal{H}}$ and ${\mathcal{G}}^{\star}_0$, expand and search for the highest scoring CPDAG ${\mathcal{G}}^{}_1$ in the new space. We iteratively repeat this procedure $${\mathcal{H}}_{i+1} = {\mathcal{H}}\cup {\mathcal{G}}^{}_i$$ until no higher scoring CPDAG is uncovered and the last ${\mathcal{G}}^{}_i$ is entirely included in the core search space (${\mathcal{H}}_i = {\mathcal{H}}\cup {\mathcal{G}}^{}_i$). Performance {#sec:performance} ----------- To illustrate the performance of our structure learning and sampling, we performed detailed simulations in [Appendix \[app:simulations\]]{}, and highlight here in [Figure \[fig:ROC10n\]]{} the improvement that iteratively updating the search space and then sampling from it has over alternatives which apply to the size and density of networks we consider. ![The performance in recovering the underlying DAG skeleton of our MCMC scheme (purple squares for different posterior thresholds) after converging to a core search space which contains the maximally scoring DAG encountered, compared to the PC algorithm (blue triangles for different significance levels) and GES (green stars for different likelihood penalisations). For completeness we include the results of our MCMC scheme when forced to use the expanded initial search skeleton from the PC algorithm (pink circles).[]{data-label="fig:ROC10n"}](./ROC10n){width="85.00000%"} Unbiased DAG sampling on a fixed search space {#sec:partition} ============================================= The order-based sampling scheme with a reduced search space ([Section \[sec:order\]]{}) can be modified to work with partition MCMC [@km17] instead of order MCMC to obtain an unbiased sample of DAGs. For partition MCMC, the nodes of a DAG are assigned to a labelled ordered partition ${\Lambda}=({\lambda},{\pi}_{\lambda})$ consisting of a partition ${\lambda}$ of the $n$ nodes and a permutation ${\pi}_{\lambda}$ where the nodes within each partition element take ascending order. This provides a unique representation of each DAG unlike the simpler order representation which has a bias towards DAGs which belong to multiple orders. The assignment can be performed by recursively tracking nodes with no incoming edges, called outpoints [@robinson70; @robinson73]. In [Figure \[fig:dagsorderpartition\]]{} the outpoints are nodes 1 and 5 which are placed in the rightmost partition element. With these nodes and their outgoing edges removed, nodes 3 and 4 become outpoints placed into the second partition element and finally node 2 fills the remaining partition element. The partition is $\lambda=[1,2,2]$ with permutation ${\pi}_{\lambda}=2,3,4,1,5$. When reversing the process and building a DAG recursively, the outpoints at each stage must be connected to from outpoints at the next stage. Each node in any partition element must have at least one incoming edge from nodes in the adjacent partition element to the right, if there is one. For example, node 2 in any DAG compatible with the partition in [Figure \[fig:dagsorderpartition\]]{}(c) must have an edge from either node 3 or node 4 (or both). Additional edges may only come from nodes further right. There are then 12 possible incoming edge combinations for node 2, three for each of node 3 and 4, giving a total of 108 DAGs compatible with the labelled ordered partition of the DAG in [Figure \[fig:dagsorderpartition\]]{}(c). In partition MCMC the sum of the scores of all these DAGs is assigned to the partition. We now describe an efficient implementation when the search space is restricted. Scoring partitions on a restricted search space ----------------------------------------------- The posterior probability of a labelled partition is the sum of posterior probabilities of DAGs within the search space compatible with the partition $$\label{partitionscoreeqdg} P_{{\mathcal{H}}}({\Lambda}\mid {D}) = \sum_{\substack{{\mathcal{G}}\subseteq {\mathcal{H}}\cr {\mathcal{G}}\in {\Lambda}}} P({\mathcal{G}}\mid {D}) \propto \prod_{i=1}^{n} \sum_{\substack{{\mbox{{\bf Pa}}}_i \subseteq {\boldsymbol{h}}^i \cr {\mbox{{\bf Pa}}}_i \in {\Lambda}}} {S}(X_i, {\mbox{{\bf Pa}}}_i \mid {D})$$ where the restriction on parents sets induced by the partition is that they must contain at least one node from the adjacent partition element to the right. To evaluate the sums in [Eq. (\[partitionscoreeqdg\])]{} for each subset of banned nodes (belonging to the same partition element or further left) we need to keep track of the subset of needed* nodes belonging to the partition element immediately to the right to ensure at least one is in the parent set. With $K$ permissible parents for node $i$ we have $3^{K}$ possible subset pairs for which we use the ternary mapping: $$\label{ternarymappingeq} g({\boldsymbol{Z}},{\boldsymbol{W}})=\sum_{j=1}^{K}I({h}^{i}_j \in {\boldsymbol{Z}})3^{j-1} + 2\sum_{j=1}^{K}I({h}^{i}_j \in {\boldsymbol{W}})3^{j-1}$$ with ${\boldsymbol{Z}}$ representing the permissible parents, and ${\boldsymbol{W}}$ those of which at least one must be present. We detail the algorithmic steps to compute these sums efficiently in [Appendix \[sumtableapp\]]{} and [Algorithm \[partparentsumalg\]]{} with a complexity of $O(K^23^K)$. The restriction encoded by partitions to remove the bias of order MCMC therefore increases the computational cost of building the lookup tables for the partition MCMC sampler. However, once the score table is built, computing the score of any partition from [Eq. (\[partitionscoreeqdg\])]{} reduces to $$\label{partitionscoreeqdgtable} P_{{\mathcal{H}}}({\Lambda}\mid {D}) \propto \prod_{i=1}^{n} {\tilde{\Sigma}}^{i}_{g(\substack{{\boldsymbol{h}}^i \in {\Lambda}, \, {\boldsymbol{h}}^i \in {\lambda}_i})} \, , \qquad {\tilde{\Sigma}}^{i}_{g(\substack{{\boldsymbol{h}}^i \in {\Lambda}, \, {\boldsymbol{h}}^i \in {\lambda}_i})} = \sum_{\substack{{\mbox{{\bf Pa}}}_i \subseteq {\boldsymbol{h}}^i \cr {\mbox{{\bf Pa}}}_i \in {\Lambda}}} {S}(X_i, {\mbox{{\bf Pa}}}_i \mid {D})$$ where ${\lambda}_i$ represents the partition element containing node $i$. The score of the partition can be evaluated in $O(Kn)$ from the lookup tables. Partition MCMC moves -------------------- The simplest move in the partition space consists of splitting partition elements, or joining adjacent ones. Proposing such a move from ${\Lambda}$ to ${\Lambda}'$ and accepting with probability $$\label{partitionacceptratio} \rho = \min \left \{ 1 , \frac{{\#\mathrm{nbd}({\Lambda})} P_{{\mathcal{H}}}( {\Lambda}' \vert {D}) }{{\#\mathrm{nbd}({\Lambda}')} P_{{\mathcal{H}}}( {\Lambda}\vert {D})} \right \} \, ,$$ while accounting for the neighbourhood sizes [following @km17] is sufficient to sample partitions proportionally to their posterior probability in the search space. Nodes in two partition elements need to be rescored by looking up new values in their restricted sum tables. Although partition elements can get to a size $O(n)$, on average they contain around $1.5$ nodes [@km15] so we would expect $O(1)$ for this move. Once a partition has been sampled, a compatible DAG can be sampled conditionally. To speed up convergence, additional permutation moves were included in [@km17], either between two nodes in adjacent partition elements, requiring again the rescoring of nodes in two partition elements, or between any two nodes in different partition elements requiring the rescoring of all nodes inbetween as well. We would typically expect $O(1)$ for the local swapping of nodes and $O(n)$ for the global swapping. The global swap is picked with a probability $\propto \frac{1}{n}$ to contain the average complexity. ![For any randomly chosen node, here 4, we score the entire neighbourhood of new positions by sequentially moving the node into different partition elements or the gaps inbetween. The new placement is sampled proportionally to the scores of the different labelled ordered partitions in the neighbourhood.[]{data-label="partitionmovepic"}](./partitionmove){width="30.00000%"} A final move is to select a single node and to move it elsewhere in the partition, or as a new partition element. Analogously to the single node move introduced in [Section \[sec:orderMCMCmoves\]]{} we can score the entire neighbourhood for any node selected at random by sequentially moving it through the partition, as in [Figure \[partitionmovepic\]]{}. Since each other node has its needed or banned parent sets essentially affected twice, scoring the neighbourhood takes $O(n)$. We always accept this move as it is sampled directly from the neighbourhood (which includes the starting partition), further aiding convergence. This move is also selected with a probability $\propto \frac{1}{n}$. Expanding the search space -------------------------- When expanding the search space, for each node we simply create further summed score tables where each time we include one other node from outside the search space as an additional parent. The space and time complexity increase by a factor of $n$ when building these tables. Since only one element is required from the needed node subsets defined by the adjacent partition element to the right, we sum over the relevant entries for all nodes outside the search space but in the adjacent partition element. For the MCMC moves, the time complexity can increase by a factor $O(n)$ but the typical increase is again $O(K)$ on average. Discussion ========== In this work we presented a novel and computationally efficient approach for learning and sampling the DAGs underlying high dimensional Bayesian network models. Two main original features are worth highlighting: First, the computational efficiency we gain by observing that every score quantity needed for the MCMC scheme can be effectively precomputed and stored in lookup tables. This goes beyond the common strategy in DAG inference of simply storing the scores of individual parent sets by additionally storing all sums of parent set scores, with no complexity overhead. This allows us to sink the complexity class of methods where we reduce the search space by grouping DAGs to score collectively, as in order-based approaches. Specifically, we reduce the main complexity bottleneck of $n^K$ to just $2^K$ or $3^K$ which provides a massive computational advantage for larger and denser networks. Order and partition MCMC constitute the building blocks of our procedure with the further advantage that each step in the chain now takes minimal computational time. Second, the improved accuracy in the inference of the network structure, achieved by means of an iterative expansion of the search space beyond the preliminary skeleton obtained through constraint-based methods. In fact the pre-defined search space may not include DAGs corresponding to the mode of the posterior distribution, so that hybrid methods can heavily benefit from the additional flexibility. The simulation studies ([Appendix \[app:simulations\]]{}) extensively demonstrate the improved performance we can achieve with respect to current mainstream approaches. When iteratively updating the search space, we include edges ensuring that the highest scoring DAG found at each stage belongs to the core search space for the next iteration. Alternatively we could sample from the posterior distribution and update the search space by adding edges with a certain posterior probability. The order-based sample is of course biased, but the additional restriction in the combinatorics of partition MCMC, which ensures a unique representation of each DAG, increases the complexity of building the necessary lookup tables. For denser networks it may be preferable to pursue bias removal only at later iterations, once the search space has already converged under order. Finding the highest scoring DAG or sampling with order MCMC share the same complexity. We chose to update the search space based on the highest scoring DAG since order MCMC may find a maximal score faster than sampling, thanks to the possibility of tempering. The freedom to add edges beyond a pre-defined skeleton, allows for the correction of errors where edges may be missed. The iterative approach is, aside from stochastic fluctuations in the search or sampling, mainly deterministic. However, since we only consider the addition of a single parent at a time for each node, the algorithm may not pick up missing correlated edges, which would only improve the score if added at the same time. Allowing for the concomitant addition of edge pairs increases the overall space complexity by a factor $n$ which can be computationally prohibitive. On the other hand we could view the search space itself, or the lists of permitted parents, as a random variable and implement a stochastic updating scheme. Especially for sparser graphs, such a scheme may be effective at extending the posterior sample outside of a fixed search space. As the initial core search space we adopted the undirected skeleton obtained from the PC algorithm, without accounting for any orientations. The iterative steps of building the score tables have exponential complexity in the number of parents. In the case of nodes with many children, which will be included as potential parents, ignoring the direction will lead to increased costs in building the lookup tables. In certain cases it may therefore be convenient to limit permissible parent sets of particular nodes to those compatible with directed or undirected edges in the CPDAG learned through the PC algorithm. Despite our focus on taking the skeleton given by the PC algorithm as the initial core search space, our approach is agnostic to the method used to define the starting point, although obviously performance will improve the closer the initial search space is to the target space containing the bulk of the posterior distribution. If relevant edges are missing in the initial search space, our algorithm can add them though it may take a few iterations to do so. False positive edges in the search space do not affect the MCMC search, but do increase the time and space needed for computing the lookup tables. In our simulations, the PC algorithm was quite conservative, even when relaxing the significance threshold, missing many edges but introducing few false positives. Due to the large number of missing edges, improving the search space tended to require quite a few iterations, which were however reasonably fast. Defining the initial core search space by GES would include more of the important edges to start with, but also many false positives. As a consequence the algorithm would potentially require fewer steps to improve the search space, at the expense of higher computational cost of each step. In the context of GES, the number of false positives is sensitive to the penalisation parameter in the score, so ideally we should optimally tune it if using GES to define the initial search space. Order-based conditional independence tests also offer another option [@ru18]. For Gaussian models, the Markov random field or conditional independence graph defined by the precision matrix [as used for example in @nhm18] is also a possibility. Theoretically the conditional independence graph should contain all edges present in the PC algorithm skeleton, potentially including more true positive edges, while most likely also introducing additional false positives. In principle one may even combine search spaces from different approaches. Interesting direction may also come from the ILP method of [@cussens11] and [@cussensetal17], if the solver manages to complete and the number of parents in the maximally scoring DAG is less than the low limit needed for their input score tables. By expanding such a DAG appropriately, we may obtain a good starting point for the full sampling. Conversely, the final search space obtained by our search could be an interesting input for the ILP, or may be determined by combining elements of both approaches. Similarly one may investigate whether one can modify dynamic programming approaches for exhaustively searching orders [@ks04; @sm06; @em07; @htw16] to work on restricted search spaces and be efficient enough to replace the MCMC search. Regardless of how we define the initial search space, or how we discover the maximal DAG, our hybrid scheme is the only one capable of efficiently sampling larger and denser graphs. Sampling from the posterior distribution not only improves structure learning, but is vital for understanding the uncertainty in the graph structure itself. To achieve robust inference we need to account for the structure uncertainty in analyses further downstream [@kuipersetal18], for example for causal interpretations and in the estimation of intervention effects [@moffaetal17]. Acknowledgements {#acknowledgements .unnumbered} ================ The authors would like to thank Markus Kalisch and Niko Beerenwinkel for useful comments and discussions. Supplementary Material {#supplementary-material .unnumbered} ====================== Algorithmic details for computing summed score tables {#sumtableapp} ===================================================== ----------------------------- ------------------------------------------------------------------ -- ----------------------------- ----------------------------------------------------------------------- Parent nodes Node score Banned parents Summed node score $\emptyset$ ${S}^{i}_{0} = {S}(X_i, \{\emptyset\} \mid {D})$ $\emptyset$ ${\Sigma}^{i}_{7} = \sum_{j=0}^{7}{S}^{i}_{j}$ ${h}^i_1$ ${S}^{i}_{1} = {S}(X_i, \{{h}^i_1\}\mid {D})$ ${h}^i_1$ ${\Sigma}^{i}_{6} = {S}^{i}_{0}+{S}^{i}_{2}+{S}^{i}_{4}+{S}^{i}_{6}$ ${h}^i_2$ ${S}^{i}_{2} = {S}(X_i, \{{h}^i_2\} \mid {D})$ ${h}^i_2$ ${\Sigma}^{i}_{5} = {S}^{i}_{0}+{S}^{i}_{1}+{S}^{i}_{4}+{S}^{i}_{5}$ ${h}^i_3$ ${S}^{i}_{4} = {S}(X_i, \{{h}^i_3\} \mid {D})$ ${h}^i_3$ ${\Sigma}^{i}_{3} = {S}^{i}_{0}+{S}^{i}_{1}+{S}^{i}_{2}+{S}^{i}_{3}$ ${h}^i_1, {h}^i_2$ ${S}^{i}_{3} = {S}(X_i, \{{h}^i_1,{h}^i_2\} \mid {D})$ ${h}^i_1, {h}^i_2$ ${\Sigma}^{i}_{4} = {S}^{i}_{0}+{S}^{i}_{4}$ ${h}^i_1, {h}^i_3$ ${S}^{i}_{5} = {S}(X_i, \{{h}^i_1,{h}^i_3\} \mid {D})$ ${h}^i_1, {h}^i_3$ ${\Sigma}^{i}_{2} = {S}^{i}_{0}+{S}^{i}_{2}$ ${h}^i_2, {h}^i_3$ ${S}^{i}_{6} = {S}(X_i, \{{h}^i_2,{h}^i_3\} \mid {D})$ ${h}^i_2, {h}^i_3$ ${\Sigma}^{i}_{1} = {S}^{i}_{0}+{S}^{i}_{1}$ ${h}^i_1, {h}^i_2, {h}^i_3$ ${S}^{i}_{7} = {S}(X_i, \{{h}^i_1,{h}^i_2, {h}^i_3\} \mid {D})$ ${h}^i_1, {h}^i_2, {h}^i_3$ ${\Sigma}^{i}_{0} = {S}^{i}_{0}$ ----------------------------- ------------------------------------------------------------------ -- ----------------------------- ----------------------------------------------------------------------- : An example score table of a node with 3 permissible parents in the search space (left). For each possible list of excluded parents, we also create a second table (right) containing the sum of scores of all subsets of remaining parents.[]{data-label="scoretableexample"} Given the scores of all permissible parent sets of a node ([Table \[scoretableexample\]]{}, left), we detail how to compute the summed score ([Table \[scoretableexample\]]{}, right). The first column indicates which nodes are banned as parents and the second column reports the sum of scores over all parent subsets excluding those nodes. For the indexing of the sums we negate the indicator function: $${\bar{f}}({\boldsymbol{Z}})=\sum_{j=1}^{K}I({h}^{i}_j \notin {\boldsymbol{Z}})2^{j-1} = 2^{K}-f({\boldsymbol{Z}})-1$$ Power set representation ------------------------ A Hasse diagram ([Figure \[possepic\]]{}) visualises the power set of the permissible parents with layers ranked by the size of the parent subsets, and helps develop a strategy to efficiently evaluate the partial sums over parent subsets. Directed edges indicate the addition of another parent to each subset, while the corresponding scores of each parent subset are attached to the nodes in [Figure \[possepic\]]{}. The advantage of the Hasse representation is that each element in the summed score table (right of [Table \[scoretableexample\]]{}) is the sum of the scores of a node and all its ancestors in the network. Power set representations have also been previously used for Bayesian network inference, for example by [@ks04] to sum over orders. ![In the Hasse diagram the permissible parent subsets can be arranged by their size and connected as a network where parents are added along each directed edge. The subsets are indicated on the right side of each node. Inside the nodes of the network are the scores of that particular parent subset. To the left of each node we place the sum of scores of that node and all its ancestors in the network. This sum encompass all possible subsets which exclude any members of the complement.[]{data-label="possepic"}](./possepic){width="65.00000%"} Score propagation ----------------- To actually perform the sums we utilise the separation of the power set into $(K+1)$ layers of differently sized parent subsets and implement [Algorithm \[parentsumalg\]]{}. The partial sums at each layer are propagated to their children in the network. To avoid overcounting contributions from ancestors which are propagated along all $d!$ paths connecting nodes $d$ layers apart, we divide by the corresponding factorials to obtain the required sums. This division is separated over the layers by dividing by one of the factorial terms each time. For each end layer there are a different number of ancestral paths to the nodes in previous layers leading to different correction factors, so we need to repeat the propagation $K$ times. Building the summed score table for each variable has a complexity of $O(K^{2}2^{K})$: during each propagation, the value at each of the $2^{K}$ elements of the power set is created by adding its starting value with the values of its parents of which there can be at most $K$ giving a complexity of $O(K2^{K})$, while the propagation is repeated $K$ times. Input The power set network of the $K$ permissible parents of variable $i$ Input The table of scores of each parent subset ${S}^{i}_j$, $j=0,\ldots,(2^{K}-1)$ Label the network nodes $Y_{f({\boldsymbol{Z}})}$ for each ${\boldsymbol{Z}}$ in the power set Initialise the node at layer 0: ${\Sigma}^{i}_{0} = Y_{0} = {S}^{i}_{0}$ Initialise the value of all nodes in layer $(m+1)$: $Y_{j} = {S}^{i}_{j}$ Add the value of nodes $Y_j$ at layer $m$, divided by $(l-m)$, to all children in the power set network at layer $(m+1)$: $Y_{j'} = Y_{j'}+\frac{Y_j}{(l-m)}$ Read off sum scores at layer $l$: ${\Sigma}^{i}_{j} = Y_{j}$ return Table of summed scores: ${\Sigma}^{i}_j$, $j=0,\ldots,(2^{K}-1)$ MAP DAG targetting ------------------ When assigning the score of each order to be the maximum score of DAGs in the order, we do not need to worry about the overcounting and can propagate only once in $O(K2^{K})$, see [Algorithm \[parentmaxalg\]]{}. Input The power set network of the $K$ permissible parents of variable $i$ Input The table of scores of each parent subset ${S}^{i}_j$, $j=0,\ldots,(2^{K}-1)$ Label the network nodes $Y_{f({\boldsymbol{Z}})}$ for each ${\boldsymbol{Z}}$ in the power set Initialise all nodes: $Y_{j} = {S}^{i}_{j}$, $j=0,\ldots,(2^{K}-1)$ Replace the value of nodes $Y_j$ at layer $l$ by the maximum of itself and all its parents in the power set network at layer $(l-1)$: $Y_{j} = \max(Y_{j},Y_{j'})$ return Table of maximal scores: ${M}^{i}_{j}=Y_j$, $j=0,\ldots,(2^{K}-1)$ Restricted sums --------------- For the partition based sampling, we need to ensure that nodes receive at least one edge from the adjacent partition element. For the example with 3 permissible parents, there are the 8 values calculated in [Table \[scoretableexample\]]{} where there was no restriction on enforcing the presence of a member of the needed parent subset (which we regard as the empty set). Additionally there are the 19 combinations in [Table \[partitionscoretableexample\]]{} where we index the sums with the ternary mapping of [Eq. (\[ternarymappingeq\])]{}. We also define the mapping back to the banned parent set $$\label{ternarymappingbackeq} {\tilde{g}}(j)= {\boldsymbol{Z}}\mid g({\boldsymbol{Z}},{\boldsymbol{W}}) = j$$ -------------------- --------------------------- -------------------------------------------------------------------------------------------------------------------- Banned parents Needed parents Summed node score $\emptyset$ ${h}^i_1$ ${\tilde{\Sigma}}^{i}_{2} = {S}^{i}_{1}+{S}^{i}_{3}+{S}^{i}_{5}+{S}^{i}_{7}$ $\emptyset$ ${h}^i_2$ ${\tilde{\Sigma}}^{i}_{6} = {S}^{i}_{2}+{S}^{i}_{3}+{S}^{i}_{6}+{S}^{i}_{7}$ $\emptyset$ ${h}^i_3$ ${\tilde{\Sigma}}^{i}_{18} = {S}^{i}_{4}+{S}^{i}_{5}+{S}^{i}_{6}+{S}^{i}_{7}$ $\emptyset$ ${h}^i_1,{h}^i_2$ ${\tilde{\Sigma}}^{i}_{8} = {S}^{i}_{1}+{S}^{i}_{2}+{S}^{i}_{3}+{S}^{i}_{5}+{S}^{i}_{6}+{S}^{i}_{7}$ $\emptyset$ ${h}^i_1,{h}^i_3$ ${\tilde{\Sigma}}^{i}_{20} = {S}^{i}_{1}+{S}^{i}_{3}+{S}^{i}_{4}+{S}^{i}_{5}+{S}^{i}_{6}+{S}^{i}_{7}$ $\emptyset$ ${h}^i_2,{h}^i_3$ ${\tilde{\Sigma}}^{i}_{24} = {S}^{i}_{2}+{S}^{i}_{3}+{S}^{i}_{4}+{S}^{i}_{5}+{S}^{i}_{6}+{S}^{i}_{7}$ $\emptyset$ ${h}^i_1,{h}^i_2,{h}^i_3$ ${\tilde{\Sigma}}^{i}_{26} = {S}^{i}_{1}+{S}^{i}_{2}+{S}^{i}_{3}+{S}^{i}_{4}+{S}^{i}_{5}+{S}^{i}_{6}+{S}^{i}_{7}$ ${h}^i_1$ ${h}^i_2$ ${\tilde{\Sigma}}^{i}_{7} = {S}^{i}_{2}+{S}^{i}_{6}$ ${h}^i_1$ ${h}^i_3$ ${\tilde{\Sigma}}^{i}_{19} = {S}^{i}_{4}+{S}^{i}_{6}$ ${h}^i_1$ ${h}^i_2, {h}^i_3$ ${\tilde{\Sigma}}^{i}_{25} = {S}^{i}_{2}+{S}^{i}_{4}+{S}^{i}_{6}$ ${h}^i_2$ ${h}^i_1$ ${\tilde{\Sigma}}^{i}_{5} = {S}^{i}_{1}+{S}^{i}_{5}$ ${h}^i_2$ ${h}^i_3$ ${\tilde{\Sigma}}^{i}_{21} = {S}^{i}_{4}+{S}^{i}_{5}$ ${h}^i_2$ ${h}^i_1, {h}^i_3$ ${\tilde{\Sigma}}^{i}_{23} = {S}^{i}_{1}+{S}^{i}_{4}+{S}^{i}_{5}$ ${h}^i_3$ ${h}^i_1$ ${\tilde{\Sigma}}^{i}_{11} = {S}^{i}_{1}+{S}^{i}_{3}$ ${h}^i_3$ ${h}^i_2$ ${\tilde{\Sigma}}^{i}_{15} = {S}^{i}_{2}+{S}^{i}_{3}$ ${h}^i_3$ ${h}^i_1, {h}^i_2$ ${\tilde{\Sigma}}^{i}_{17} = {S}^{i}_{1}+{S}^{i}_{2}+{S}^{i}_{3}$ ${h}^i_1, {h}^i_2$ ${h}^i_3$ ${\tilde{\Sigma}}^{i}_{22} = {S}^{i}_{4}$ ${h}^i_1, {h}^i_3$ ${h}^i_2$ ${\tilde{\Sigma}}^{i}_{16} = {S}^{i}_{2}$ ${h}^i_2, {h}^i_3$ ${h}^i_1$ ${\tilde{\Sigma}}^{i}_{14} = {S}^{i}_{1}$ -------------------- --------------------------- -------------------------------------------------------------------------------------------------------------------- : An example sum score table for each possible list of excluded parents, where at least one member of the needed parents subset must be included.[]{data-label="partitionscoretableexample"} Again we build a network representation of the possibilities by replicating each node in the power set representation according to the number of choices of possible needed parent subsets in the complement of the banned parent subsets. We rank the nodes by the size of the banned node subsets as in [Figure \[partpossepic\]]{}, and assign to the nodes the score corresponding to the complement of the banned node subset. The connections in the network represent either removing an element from the banned parent subset, or moving it to the needed parent subset. For any $j$ such that ${\boldsymbol{h}}^i_j$ is in the banned parent subset there is then an edge to the node indexed by $3^{j-1}$ more and the node indexed by $3^{j-1}$ less using the ternary mapping of [Eq. (\[ternarymappingeq\])]{}. ![The banned parent subsets can be arranged by their size and expanded to include all needed parent subsets in the complement. Inside the nodes of the network are the scores of the complement of the banned parent subset, and both the banned and needed parents subsets are indicated on the side of each node The nodes are connected as a network where parents are deleted from the banned parent subsets or moved into the needed parent subsets. The sum of all scores which do not involve certain banned parents but do include at least one member of the needed parent subset is simply the sum of scores associated with a node and all its ancestors in the network.[]{data-label="partpossepic"}](./partpossepic){width="\textwidth"} The sums in [Table \[partitionscoretableexample\]]{} are the sums of the scores of each node in the network in [Figure \[partpossepic\]]{} and its ancestors. To compute these sums efficiently we again propagate through the network using [Algorithm \[partparentsumalg\]]{} whose complexity is $O(K^23^K)$. Input The network of the banned and needed parent subsets of variable $i$ from the $K$ permissible parents Input The table of scores of each parent set ${S}^{i}_j$, $j=0,\ldots,(2^{K}-1)$ Input The table of summed scores for each banned parent subset ${\Sigma}^{i}_j$, $j=0,\ldots,(2^{K}-1)$ Label the network nodes $Y_{g({\boldsymbol{Z}},{\boldsymbol{W}})}$ Initialise the restricted summed scores for empty needed nodes: ${\tilde{\Sigma}}^{i}_{g(f^{-1}(j),\emptyset)} = {\Sigma}^{i}_{j}$ Initialise the nodes at layer 1: ${\tilde{\Sigma}}^{i}_{j} = Y_{j} = {S}^{i}_{{\bar{f}}({\tilde{g}}(j))}$ Initialise the value of all nodes in layer $(m+1)$: $Y_{j} = {S}^{i}_{{\bar{f}}({\tilde{g}}(j))}$ Add the value of nodes $Y_j$ at layer $m$, divided by $(l-m)$, to all children in the network at layer $(m+1)$: $Y_{j'} = Y_{j'}+\frac{Y_j}{(l-m)}$ Read off restricted sum scores at layer $l$: ${\tilde{\Sigma}}^{i}_{j} = Y_{j}$ return Table of restricted summed scores: ${\tilde{\Sigma}}^{i}_j$, $j=0,\ldots,(3^{K}-1)$ Simulation studies {#app:simulations} ================== To examine the performance and convergence of our method, we performed a simulation study for 4 network sizes $n\in\{20, 80, 140, 200\}$ and 2 sample sizes $N\in\{2n, 10n\}$. For each combination of network and sample sizes, 100 random graphs and corresponding data were generated using the functions `randomDAG` and `rmvDAG` from the [[pcalg]{}]{} package. The strengths of the edges were in the range $[0.4, 2]$. The edge probability parameter was set to $\frac{4}{n}$, so that the average number of edges in the DAG equals $2(n-1)$ and the average parent set size per node is just under $2$. Although the average number of parents is around 2, the determining factor for the run time is the maximal number of parents. With 20 nodes the maximal number of parents is 6 on average, rising to 8 on average with 80 nodes (with more than 5% of cases having 10 parents of more) and further increases for larger networks. ![The root mean square error (RMSE) between edge probabilities from pairs of different runs for each simulation as the size of the network increases. The runs are those from [Figure \[fig:RMSEconvergence\]]{} where the correlation $\rho^2$ was displayed.[]{data-label="fig:RMSEconvergence"}](./RMSEconvergence){width="75.00000%"} Methods which impose a hard limit on the number of parents, as is the case with leading score-based schemes [@fk03; @tk05; @gh08; @km17] including ILP solvers [@cussens11; @cussensetal17], scale in complexity as $O(n^{K+1})$ and simply do not scale to our simulation setting. Order-based schemes for MAP DAG discovery, can prune the sets of permissible parents to reduce this computational burden [@fk03; @tk05]. A recent implementation [@scanagatta2015learning] can only handle categorical variables, so we compare to that approach in [Section \[catsim\]]{}. For continuous data we therefore compare only to GES [@chickering02], a greedy structure-based search, and the PC algorithm [@bk:sgs00], a constraint-based method. The hybrid approach of [@tba06] of a greedy structure search in a constraint-based skeleton performs very similarly to the PC algorithm and is not included in the comparisons. Skeleton inference ------------------ To assess the performance we first considered the number of true positives (TP) and false positives (FP) in the undirected skeletons of the networks inferred and scaled them by the number of edges (P) in the true DAG $$\mathrm{TPR} = \frac{\mathrm{TP}}{\mathrm{P}} \, \qquad \mathrm{FPRn} = \frac{\mathrm{FP}}{\mathrm{P}}$$ We computed the median TPR along with the first and third quartiles and plotted ([Figures \[fig:ROC10n\]]{} and \[fig:ROC2n\]) against the median FPRn over the 100 realisations for our MCMC scheme for two search spaces: the initial search space defined by the PC algorithm skeleton, and expanded to include an additional parent; and the final search space which is improved iteratively until it contains the MAP DAG discovered. Also plotted are the results from GES [@chickering02] and the PC algorithm using Fisher’s $z$ test for conditional independence [@bk:sgs00; @art:KalischMCMB2012]. The range of discrimination thresholds for plotting points in the ROC curves were: - penalisation parameter $\frac{\lambda}{\log (N)} \in \{1, 2, 5, 7, 9, 11, 15, 25\}$ for GES - significance level $\alpha \in \{0.01, 0.05, 0.1, 0.2, 0.35, 0.45\}$ for the PC algorithm (the highest threshold may result in too many false positives to display in the plots) - posterior probability $\rho \in \{0.2, 0.3, 0.4, 0.5, 0.6, 0. 7, 0.8, 0.9, 0.95, 0.99\}$ for MCMC With the initial search space, we see a distinct improvement with our MCMC scheme (pink circles in [Figures \[fig:ROC10n\]]{} and \[fig:ROC2n\]), while when we improve the search space iteratively we observe a strong advantage over the alternative methods (purple squares in [Figures \[fig:ROC10n\]]{} and \[fig:ROC2n\]) and approach perfect recovery of the skeleton for the larger sample size ([Figure \[fig:ROC10n\]]{}). In the simulations, increasing the significance level of the conditional independence tests of the PC algorithm does not really improve the recovery of true edges, while the additional false positives start to dramatically increase the algorithm’s runtime. ![The performance in recovering the underlying DAG skeleton. The graph is as [Figure \[fig:ROC10n\]]{} but with a smaller sample size of $N=2n$.[]{data-label="fig:ROC2n"}](./ROC2n){width="75.00000%"} ![How the iterative search improves performance, using the setting of [Figure \[fig:ROC10n\]]{} as an example. Starting from the search skeleton defined by the PC algorithm, expanded to include an possible additional parent (pink circles) we plot the highest scoring DAG discovered in that search space and during subsequent iterations (red plusses). When no better DAG is discovered, sampling from the final search space provides the purple squares.[]{data-label="fig:iterations10n"}](./iterations10n){width="75.00000%"} Iterative steps {#moresimsapp} --------------- To explore how the iterative search leads to an improvement in performance we keep track of the highest scoring DAG uncovered at each iteration, and used to update the core search space for the next iteration. In [Figure \[fig:iterations10n\]]{}, we overlay these intermediate results on the MCMC lines of [Figure \[fig:ROC10n\]]{}. Each iteration, and especially the earlier ones leads to an improvement in the search space allowing the MCMC search to find better DAGs which were previously excluded. Finally, utilising the posterior probability of edges in the sample from the final search space, we can remove some of the false positive edges in the point estimate of the highest scoring DAG uncovered. Direction inference ------------------- Along with inferring the undirected skeleton, we also assess the performance in recovering the correct directions and compute the structural Hamming distance (SHD) between the true generating DAG and those inferred by the different methods. In all cases we convert to CPDAGs before computing the distances. For GES we used the penalisation $\lambda = 2\log(N)$ while for the PC algorithm we used a significance level of $\alpha=0.05$. To condense the sample of DAGs from our MCMC schemes to a single graph, we converted the sample to CPDAGs and retained edges occurring with a posterior probability greater than $0.5$. The result for targeting a MAP DAG correspond to the highest scoring DAG discovered in the final search space, again transformed into a CPDAG. The results ([Figures \[fig:SHD10n\]]{} and \[fig:SHD2n\]) again show a strong improvement of our MCMC approach over the alternative algorithms. Sampling and performing Bayesian model averaging also offers a consistent advantage over taking a MAP point estimate. ![The performance in recovering the CPDAG measured by the structural Hamming distance (SHD). We compare the performance of our MCMC sampler (purple) and the highest scoring (MAP) DAG (red) from the final search space to the PC algorithm (blue) and GES (green), along with the MCMC sampler from a search space of the the PC algorithm skeleton expanded to include one additional parent (pink).[]{data-label="fig:SHD10n"}](./SHD10n){width="75.00000%"} ![The performance in recovering the CPDAG. The graph is as [Figure \[fig:SHD10n\]]{} but with a smaller sample size of $N=2n$.[]{data-label="fig:SHD2n"}](./SHD2n){width="75.00000%"} Other DAG types --------------- For the same settings as in [Figure \[fig:ROC10n\]]{}, of graphs of size $n\in\{20, 80, 140, 200\}$ with a sample size of $N=10n$ and an average parent set size of 2, which we denote by $\nu=2$, we sample Erdős-Rényi DAGs instead using that option in the `randDAG` function of the [[pcalg]{}]{} package. We observe ([Figure \[fig:ROCer2\]]{}) a slight worsening in the GES results and a slight improvement in the PC algorithm results (compared to the previous simulation, [Figure \[fig:ROC10n\]]{}), and again a strong advantage for our MCMC approach. ![The performance in recovering the underlying DAG skeleton for Erdős-Rényi DAGs with 2 parents on average. We compare the performance of our MCMC scheme (purple squares) on its final search space, to the PC algorithm (blue triangles) and GES (green stars).[]{data-label="fig:ROCer2"}](./ROCer2){width="75.00000%"} For Barabási-Albert scale-free or power-law DAGs, again with an average of 2 parents ([Figure \[fig:ROCbarab2\]]{}), there is a further decrease in the performance of GES and a slight increase in false positives of our MCMC scheme. Nonetheless, it still recovers almost all true positives, unlike the alternatives. By sparsifying the graphs and setting the average number of parents to 1 instead, our MCMC scheme performs very well, the PC algorithm captures most true edges, but GES obtains essentially perfect performance ([Figure \[fig:ROCbarab1\]]{}). ![The performance in recovering the underlying DAG skeleton for Barabási-Albert DAGs with 2 parents on average.[]{data-label="fig:ROCbarab2"}](./ROCbarab2){width="75.00000%"} ![The performance in recovering the underlying DAG skeleton for Barabási-Albert DAGs with 1 parent on average.[]{data-label="fig:ROCbarab1"}](./ROCbarab1){width="75.00000%"} Increasing the density instead to have an average of 3 parents per node, and having a network made up of two Erdős-Rényi islands with an interconnectivity parameter of 0.1 ([Figure \[fig:ROCinterer3\]]{}), we observe that our MCMC scheme has a marked increase in false positive edges (up to around 40% of the number of true positives) and that we also miss out on some true edges. We can track this loss of performance down to the fact that the final search space does not include the true DAG, for when we artificially add it to the search space (grey diamonds, [Figure \[fig:ROCinterer3\]]{}) we again have near perfect performance for larger networks. Modifications to improve finding the best search space for our MCMC sampler could therefore significantly improve its performance for denser networks. Although our MCMC scheme does miss out on some true edges, the PC algorithm struggle to find even half of them while GES finds some more, but at the cost of a large number of false positives. Both alternatives perform substantially worse than MCMC. ![The performance in recovering the underlying DAG skeleton for a pair of Erdős-Rényi islands with an interconnectivity coefficient of 0.1 and 3 parents on average. The MCMC performance on a search space artificially including the true DAG is illustrated with grey diamonds.[]{data-label="fig:ROCinterer3"}](./ROCinterer3){width="75.00000%"} In terms of run times, the final MCMC sampling is roughly an order of magnitude slower than finding the initial search space with the PC algorithm, but the bulk of the time is in the iterative steps to find the final search space, which can be roughly an order of magnitude slower again ([Figure \[fig:runtime\]]{}). As the final search space can have a strong effect on the performance for denser networks ([Figure \[fig:ROCinterer3\]]{}), improvements in the speed of finding it could also be beneficial. The actual run times of the full algorithm averages around 10 seconds for 20 nodes, around 10 minutes for 80 nodes, and up to a few hours for 200 nodes. ![The average runtimes of finding the initial search space with the PC algorithm (blue), iteratively improving the search space (red) and the final MCMC sampling (purple) for the four different network settings.[]{data-label="fig:runtime"}](./runtime){width="75.00000%"} Categorical simulations {#catsim} ----------------------- To compare to the order-based search implemented in [[r.blip]{}]{} [@scanagatta2015learning], we had to simulate categorical data. First we considered the ANDES network with $n=223$ nodes for which we generated random samples [with the [[bnlearn]{}]{} package; @bnlearn] for two sample sizes $N=2n$ and $N=10n$. Fixing the score parameters to those in [[r.blip]{}]{} and using the same runtime (1500 seconds) we find similar high scores (slightly higher for [[BiDAG]{}]{} for $N=2n$: -42844.62 against -42856.39; and slightly lower for $N=10n$: -209833.8 against -209811.9). Irrespective of the score, we find a much better structure with [[BiDAG]{}]{} with notably fewer false positive edges ([Figure \[fig:andes\]]{}). Moreover, by accounting for the uncertainty in the edges and considering a posterior sample as opposed to a point estimate, we can drastically reduce the number of false positive edges while retaining true edges. ![The accuracy in learning the ANDES network. We compare the performance of [[r.blip]{}]{} to the MAP DAG search of [[BiDAG]{}]{} and its posterior sampling.[]{data-label="fig:andes"}](./ROCbidagvsblipandes){width="75.00000%"} A similar pattern is repeated in a larger scale simulation of 100 binary power-law networks (with 1 parent on average and a maximum of 5). In terms of score, we routinely find higher scoring graphs at the larger sample size, but mostly lower scores at the lower sample size ([Figure \[fig:binsimscore\]]{}). The higher score at the lower sample size is however driven by false positives and does not lead to a more accurate graph structure ([Figure \[fig:binsim\]]{}). Indeed running [[r.blip]{}]{} for longer times, while increasing the score on average, solely increases the false positive rate with no improvement in finding true edges. Although the highest scoring DAGs returned by [[BiDAG]{}]{} is more accurate than those of [[r.blip]{}]{}, the real advantage comes again by sampling and using a posterior threshold to remove false positive edges, which heavily improves the accuracy in network inference. ![The relative log scores of the highest scoring DAG returned by [[r.blip]{}]{} and by [[BiDAG]{}]{} for random networks of $n=100$ nodes. The runtime for [[r.blip]{}]{} was fixed to 240 seconds, while [[BiDAG]{}]{} had an average time of 150 seconds.[]{data-label="fig:binsimscore"}](./scoreshiftblip){width="50.00000%"} ![The accuracy in learning random networks with $n=100$ nodes. We compare the performance of [[r.blip]{}]{} to the MAP DAG search of [[BiDAG]{}]{} and its posterior sampling. The 3 points (left to right) for [[r.blip]{}]{} corresponds to runtimes of 60, 120 and 240 seconds, while [[BiDAG]{}]{} had an average time of 150 seconds for MAP discovery, and 30 seconds for sampling.[]{data-label="fig:binsim"}](./ROCbidagvsblip){width="75.00000%"} Categorical data {#catdata} ================ For categorical data, we employ the BDe score [@hg95]. For ease of presentation, we details the binary case here. For any node $X_i$, its contribution to the score involves computing the number of times $X_i$ takes the value 1, or the value 0, for each of the $2^K$ possible configurations of its $K$ parents ${\mbox{{\bf Pa}}}_i$. All the parents for each of the ${N}$ observations must be run through in a complexity of $O(K{N})$. As there is a parameter associated with each of the $2^K$ parent configurations, we assume ${N}\gg 2^K$. Building the score table of node $X_i$ by naively running through all parent configurations would take $O(K2^K{N}) \gg O(K4^K)$. However the parent configurations are connected via the poset structure of [Figure \[possepic\]]{}, so we can build the score table more efficiently by only looking at the raw data once. For the BDe score, for the case when all $K$ parents are present we build the vectors ${\boldsymbol{N}}_{1}^{i}({\boldsymbol{h}}^{i})$ and ${\boldsymbol{N}}_{0}^{i}({\boldsymbol{h}}^{i})$ whose $2^K$ elements count the number of times $X_i$ takes the value 1 and 0 for each parent state, in time $O(K{N})$. We employ a binary mapping of the parent states to elements of the vectors using $$\sum_{j=1}^{\vert {\boldsymbol{Z}}\vert }I(Z_j =1)2^{j-1}$$ where for the full set of parents, ${\boldsymbol{Z}}={\boldsymbol{h}}^{i}$. When we remove one of the parents to compute the score table entry at layer $(K-1)$ in the poset of [Figure \[possepic\]]{} we simply combine elements of the vector ${\boldsymbol{N}}_{\{1,0\}}^{i}$ where the removed parent takes the value 0 with the corresponding elements where it takes the value 1. In general we can create the vectors at each level from any connected at a higher level with $$\label{eqcombinebinaryvectors} {\boldsymbol{N}}_{\{1,0\}}^{i}({\boldsymbol{Z}}\setminus {\boldsymbol{Z}}_j)[t] = {\boldsymbol{N}}_{\{1,0\}}^{i}({\boldsymbol{Z}})[v(t,j)] + {\boldsymbol{N}}_{\{1,0\}}^{i}({\boldsymbol{Z}})[v(t,j)+2^{j-1}]$$ where the square brackets indicate the elements of the vectors and we employ the mapping $$v(t,j) = t + \left(\left\lceil \frac{t}{2^{j-1}} \right\rceil -1 \right)2^{j-1}$$ From the pair of vectors for any set of permissible parent nodes ${\boldsymbol{N}}_{\{1,0\}}^{i}({\boldsymbol{Z}})$ we can compute the entry in the score table according to the BDe score [@hg95] $$\label{BDescoreeqn} {S}^{i}_{f({\boldsymbol{Z}})} = \sum_{t=1}^{2^{m}} \frac{\Gamma\left(\frac{\chi}{2^{m}}\right)}{\Gamma\left(\frac{\chi}{2^{m+1}}\right)\Gamma\left(\frac{\chi}{2^{m+1}}\right)} \frac{\Gamma\left({\boldsymbol{N}}_{1}^{i}({\boldsymbol{Z}})[t]+\frac{\chi}{2^{m+1}}\right)\Gamma\left({\boldsymbol{N}}_{0}^{i}({\boldsymbol{Z}})[t]+\frac{\chi}{2^{m+1}}\right)}{\Gamma\left({\boldsymbol{N}}_{1}^{i}({\boldsymbol{Z}})[t]+{\boldsymbol{N}}_{0}^{i}({\boldsymbol{Z}})[t]+\frac{\chi}{2^{m}}\right)}$$ with $m=\vert {\boldsymbol{Z}}\vert$ and $\chi$ the hyperparameters of the beta distributions which correspond to pseudocounts in the score. Input The power set network of the $K$ permissible parents of variable $i$ Input The count vectors of the full parent set ${\boldsymbol{N}}_{\{1,0\}}^{i}({\boldsymbol{h}}^{i})$ Label the network nodes $Y_{f({\boldsymbol{Z}})}$ for each ${\boldsymbol{Z}}$ in the power set Compute the score ${S}^{i}_{2^{K}-1}$ at layer $K$ from the ${\boldsymbol{N}}_{\{1,0\}}^{i}({\boldsymbol{h}}^{i})$ Choose any child in the network Compute ${\boldsymbol{N}}_{\{1,0\}}^{i}(f^{-1}(j))$ from the child Compute ${S}^{i}_{j}$ from the ${\boldsymbol{N}}_{\{1,0\}}^{i}(f^{-1}(j))$ return Table of scores: ${S}^{i}_j$, $j=0,\ldots,(2^{K}-1)$ Repeating the creation of the count vectors and computation of the score by moving up the layers in the poset of [Figure \[possepic\]]{}, as summarised in [Algorithm \[binarydataalg\]]{}, we efficiently build the score table for each node in the data. For each term at layer $l$ we look at vectors from the layer above of size $2^{l+1}$ so that filling out the score tables takes $O(3^K)$. Combining with the initial step leads to an overall complexity of $O(\max\{K{N},3^{K}\})$ which is a significant improvement on the naive implementation of $O(\max\{K2^{K}{N},K4^{K}\})$. For categorical data, the same approach is followed, although with mixed radix indexing for different sized categories rather than the simple binary mapping discussed above. With more possible states, the complexity also increases. For example if all categories have $C$ levels, the complexity is $O((C+1)^K)$. 0.2in [^1]: [[R]{}]{} package [[BiDAG]{}]{} is available at <https://CRAN.R-project.org/package=BiDAG>	{ "pile_set_name": "ArXiv" }
--- bibliography: - 'RP.bib' --- [Zeros of a random analytic function approach perfect spacing under repeated differentiation]{}\ Robin Pemantle[^1]$^,$ [^2], Sneha Subramanian[^3] [Abstract:]{} We consider an analytic function $f$ whose zero set forms a unit intensity Poisson process on the real line. We show that repeated differentiation causes the zero set to converge in distribution to a random translate of the integers. [Key words and phrases:]{} Poisson, coefficient, saddle point, lattice, Cauchy integral, random series, translation-invariant.\ [Subject classification]{} 30B20, 60G55; Secondary: 30C15.\ Introduction {#sec:intro} ============ Study of the relation of the zero set of a function $f$ to the zero set of its derivative has a rich history. The Gauss-Lucas theorem (see, e.g., [@marden1949 Theorem 6.1]) says that if $f$ is a polynomial then the zero set of $f'$ lies in the convex hull of the zero set of $f$. Another property of the differentiation operator is that it is [complex zero decreasing]{}: the number of non-real zeros of $f'$ is at most the number of non-real zeros of $f$. This property is studied by [@craven-csordas] in the more general context of [P[ó]{}lya-Schur]{} operators, which multiply the coefficients of a power series by a predetermined sequence. Much of the recent interest in such properties of the derivative and other operators stem from proposed attacks on the Riemann Hypothesis involving behavior of zeros under these operators [@LeMo1974; @Conr1983]. See also [@pemantle-hyperbolicity Section 4] for a survey of combinatorial reasons to study locations of zeros such as log-concavity of coefficients [@brenti-LC] and negative dependence properties [@borcea-branden-liggett]. The vague statement that differentiation should even out the spacings of zeros is generally believed, and a number of proven results bear this out. For example, a theorem attributed to Riesz (later rediscovered by others) states that the minimum distance between zeros of certain entire functions with only real zeros is increased by differentiation; see [@farmer-rhoades Section 2] for a history of this result and its appearance in [@Stoy1926] and subsequent works of J. v. Sz.–Nagy and of P. Walker. The logical extreme is that repeated differentiation should lead to zeros that are as evenly spaced as possible. If the original function $f$ has real zeros, then all derivatives of $f$ also have all real zeros. If the zeros of $f$ have some long-run density on the real line, then one might expect the zero set under repeated differentiation to approach a lattice with this density. A sequence of results leading up to this was proved in [@farmer-rhoades]. They show that the gaps between zeros of $f' + af$ are bounded between the infimum and supremum of gaps between consecutive zeros of $f$ and generalize this to a local density result that is applicable to the Riemann zeta function. They claim a result [@farmer-rhoades Theorem 2.4.1] that implies the convergence of spacings of zeros to a constant (their Theorem 2.4.2) but a key piece of their proof, Proposition 5.2.1, has a hole that seemingly cannot be fixed (D. Farmer, personal communication). The central object of this paper is a random analytic function $f$ whose zeros form a unit intensity point process. We construct such a function and prove translation invariance in Proposition \[th:f and TI\]. Our main result is that as $k \to \infty$, the zero set of the $k^{th}$ derivative of $f$ approaches a random translate of the integers. Thus we provide, for the first time, a proof of the lattice convergence result in the case of a random zero set. The remainder of the paper is organized as follows. In the next section we give formal constructions and statements of the main results. We also prove preliminary results concerning the contruction, interpretation and properties of the random function $f$. At the end of the section we state an estimate on the Taylor coefficients of $f$, Theorem \[th:a\] below, and show that Theorem \[th:main\] follows from Theorem \[th:a\] without too much trouble. In Section \[sec:sigma\] we begin proving Theorem \[th:a\], that is, estimating the coefficients of $f$. It is suggested in [@farmer-rhoades] that the Taylor series for $f$ might prove interesting, and indeed our approach is based on determination of these coefficients. We evaluate these via Cauchy’s integral formula. In particular, in Theorem \[th:sigma\], we locate a saddle point $\sigma_k$ of $z^{-k} f$. In Section \[sec:f\] we prove some estimates on $f$, allowing us to localize the Cauchy integral to the saddle point and complete the proof of Theorem \[th:a\]. We conclude with a brief discussion. Statements and preliminary results {#sec:statements} ================================== We assume there may be readers interested in analytic function theory but with no background in probability. We therefore include a couple of paragraphs of formalism regarding random functions and Poisson processes, with apologies to those readers for whom it is redundant. Formalities {#ss:formal} ----------- A random object $X$ taking values in a $S$ endowed with a $\sigma$-field $\S$ is a map $X : (\Omega , \F) \to (S , \S)$ where $(\Omega , \F , \P)$ is a probability space. We will never need explicitly to name the $\sigma$-field $\S$ on $\S$, nor will we continue to say that maps must be measurable, though all maps are assumed to be. If $S$ is the space of analytic functions, the map $X$ may be thought of as a map $f : \Omega \times \C \to \C$. The statement “$f$ is a random analytic function” means that for any fixed $\omega \in \Omega$, the function $z \mapsto f(\omega , z)$ is an analytic function. The argument $\omega$ is always dropped from the notation, thus, e.g., one may refer to $f'(z)$ or $f(\lambda z)$, and so forth, which are also random analytic functions. A unit intensity Poisson process on the real lines is a random counting measure $N$ on the measurable subsets of $\R$ such that for any disjoint collection of sets $\{ A_1 , \ldots , A_n \}$, of finite measure, the random variables $\{ N(A_1) , \ldots , N(A_n) \}$ are a collection of independent Poisson random variables with respective means $\|A_1\| , \ldots , \|A_n\|$ (here $\|B\|$ denotes the measure of $B$). The term “counting measure” refers to a measure taking values in the nonnegative integers; there is a random countable set $E$ such that the measure of any set $A$ is the cardinality of $A \cap E$. We informally refer to the set $E := \{ x \in \R : N( \{ x \}) = 1 \}$ as the “points of the Poisson process.” Let $\Omega$ henceforth denote the space of counting measures on $\R$, equipped with its usual $\sigma$-field $\F$, and let $\P$ denote the law of a unit intensity Poisson process. This simplifies our notation by allowing us to construct a random analytic function $f : \Omega \times \C \to \C$ by a formula for the value of $f(N , z)$, guaranteeing that the random function $f$ is determined by the locations of the points of the Poisson process $N$. For $N \in \Omega$ and $\lambda \in \R$, let $\tau_\lambda N$ denote the shift of the measure $N$ that moves points to the right by $\lambda$; in other words, $\tau_\lambda N (A) := N (A - \lambda)$ where $A - \lambda$ denotes the leftward shift $\{ x - \lambda : x \in A \}$. A unit intensity point process is translation invariant. This means formally that $\P \circ \tau_\lambda = \P$ for any $\lambda$. If $X$ is a random object in a space $S$ admitting an action of the group $(\R , +)$, we say that $X$ is constructed in a translation invariant manner from $N$ if $X(\tau_\lambda N) = \lambda X(N)$. This implies that the law of $X$ is invariant under the $(\R , +)$-action but not conversely. In what follows we will construct a random analytic function $f$ which is translation invariant up to constant multiple. Formally, for any function $g$ let $[g]$ denote the set of functions $\{ \lambda g : g \in \R \}$. Let $(\R , +)$ act on the set of analytic functions by translation in the domain: $\lambda * g(z) := g(z - \lambda)$. This commutes with the projection $g \mapsto [g]$. Our random analytic function $f$ will have the property that $[f]$ is constructed in a translation invariant manner from $N$. Construction of $f$ {#ss:f} ------------------- Various quantities of interest will be defined as sums and products over the set of points of the Poisson process $N$. The sum of $g$ evaluated at the points of the counting measure $N$ is more compactly denoted $\int g \, dN$. If $\int \|g\| \, dN < \infty$ then this is an absolutely convergent sum and its meaning is clear. Because many of these infinite sums and products are not absolutely convergent, we introduce notation for some symmetric sums that are conditionally convergent. Let $g : \R \to \C$ be any function. Let $N_M$ denote the restriction of $N$ to the interval $[-M,M]$. Thus, $\int g \, dN_M$ denotes the sum of $g(x)$ over those points of the process $N$ lying in $[-M,M]$. Define the symmetric integral $\int_* g \, dN$ to be equal to $\lim_{M \to \infty} \int g \, dN_M$ when the lmit exists. It is sometimes more intuitive to write such an integral as a sum over the points, $x$, of $N$. Thus we denote $$\sum_* g(x) := \int_* g(x) \, dN (x) = \lim_{M \to \infty} \int g(x) \, dN_M (x)$$ when this limit exists. Similarly for products, we define the symmetric limit by $$\prod_* g(s) := \lim_{M \to \infty} \exp \left ( \int \log g \, dN_M \right ) \, .$$ Note that although the logarithm is multi-valued, its integral agains a counting measure is well definedup to multiples of $2 \pi i$, whence such an integral has a well defined exponential. \[th:f and TI\] Except for a set of values of $N$ of measure zero, the symmetric product $$\label{eq:f} f(z) := \prod_* \left ( 1 - \frac{z}{x} \right )$$ exists. The random function $f$ defined by this product is analytic and translation invariant. In particular, $$\label{eq:const} f(\tau_\lambda N , z) = \frac{f(N , z - \lambda)}{f(N , -\lambda)}$$ which implies $[f(\tau_\lambda N , \cdot)] = [f(N , \cdot - \lambda)]$. We denote the $k^{th}$ derivative of $f$ by $f^{(k)}$. The following is an immediate consequence of Theorem \[th:f and TI\]. \[cor:TI\] For each $k$, the law of the zero set of $f^{(k)} (z)$ is translation invariant. $\Cox$ Translation invariance of $f$ is a little awkward because it holds only up to a constant multiple. It is more natural to work with the logarithmic derivative $$h(z) := \sum_* \frac{1}{z-x} \, .$$ \[lem:h TI\] The random function $h$ is meromorphic and its poles are precisely the points of the process $N$, each being a simple pole. Also $h$ is translation invariant and is the uniform limit on compact sets of the functions $$h_M (z) := \int \frac{1}{z-x} \, dN_M (x) \, .$$ [Proof:]{} Let $\Delta_M := h_{M+1} (0) - h_M (0)$. It is easily checked that (i) $\P (\Delta_M > \ee)$ is summable in $M$; (ii) $\E \Delta_M = 0$; (iii) $\E \Delta_M^2$ is summable. By Kolmogorov’s three series theorem, it follows that $\lim_{M \to \infty} h_M (0)$ exists almost surely. To improve this to almost sure uniform convergence on compact sets, define the $M^{th}$ tail remainder by $T_M (z) := h(z) - h_M (z)$ if the symmetric integral $h$ exists. Equivalently, $$T_M (z) := \lim_{R \to \infty} \int \frac{1}{z-x} d(N_R - N_M) (x)$$ if such a limit exists. Let $K$ be any compact set of complex numbers. We claim that the limit exists and that $$\label{eq:claim} G(M) := \sup_{z \in K} \left \| T_M(z) - T_M (0) \right \| \to 0 \mbox{ almost surely as } M \to \infty .$$ To see this, assume without loss of generality that $\disp M \geq 2 \sup \{ \|\Re \{ z \}\| : z \in K \}$. Then $$\label{eq:T} T_M (z) - T_M (0) = \lim_{R \to \infty} \int \left ( \frac{1}{z-x} - \frac{1}{-x} \right ) \, d (N_R - N_M) (x) \, .$$ Denote $C_K := \sup_{z \in K} \|z\|$. As long as $z \in K$ and $\|x\| \geq M$, the assumption on $M$ gives $$\label{eq:C_K} \left \| \frac{1}{z-x} - \frac{1}{-x} \right \| = \left \| \frac{z}{x(z-x)} \right \| \leq \frac{2C_K}{x^2} \, .$$ This implies that the integral in is absolutely integrable with probability 1. Thus, almost surely, $T_M (z) - T_M (0)$ is defined by the convergent integral $$T_M (z) - T_M (0) = \int \left ( \frac{1}{z-x} - \frac{1}{-x} \right ) \, d (N - N_M) (x) \, .$$ Plugging in , we see that $\disp G(M) \leq 2 C_K \int x^{-2} \, d(N - N_M) (x) \, ,$ which goes to zero (by Lebesgue dominated convergence) except on the measure zero event that $\int \|x\|^{-2} \, dN(x) = \infty$. This proves . The triangle inequality then yields $\sup_{z \in K} \|T_M (z)\| \leq G(M) + \|T_M (0)\|$, both summands going to zero almost surely. By definition of $T_M$, this means $h_M \to h$ uniformly on $K$. The rest is easy. For fixed $K$ and $M$, $h = h_M + \lim_{R \to \infty} (h_R - h_M)$. When $M$ is sufficiently large and $R > M$, the functions $h_R - h_M$ are analytic on $K$. Thus $h$ is the sum of a meromorphic function with simple poles at the points of $N$ in $K$ and a uniform limit of analytic functions. Such a limit is analytic. Because $K$ was arbitrary, $h$ is meromorphic with simple poles exactly at the points of $N$. The final conclusion to check is that $h$ is translation invariant. Unraveling the definitions gives $$h (\tau_\lambda N , z) = \int_{* \lambda} \frac{1}{(z - \lambda) - x} \, d N (x)$$ where $\int_{* \lambda}$ is the limit as $M \to \infty$ of the integral over $[-\lambda - M , - \lambda + M]$. Translation invariance then follows from checking that $\int_{M-\lambda}^M \frac{1}{z-x} \, dN(x)$ and $\int_{-M-\lambda}^{-M} \frac{1}{z-x} \, dN(x)$ both converge almost surely to zero. This follows from the large deviation bound $$\P \left ( \left \| \int_{M-\lambda}^M \frac{1}{z-x} \, dN(x) \right \| \geq \ee \right ) = O \left ( e^{-c M} \right )$$ and Borel-Cantelli. $\Cox$ [Proof of Theorem]{} \[th:f and TI\]: The antiderivative of the meromorphic function $h$ is an equivalence class (under addition of constants) of functions taking values in $\C \mod (2 \pi i)$. Choosing the antiderivative of $h_m$ to vanish at the origin and exponentiating gives the functions $f_M$, whose limit as $M \to \infty$ is the symmetric product, $f$. Analyticity follows because $f$ is the uniform limit of analytic functions. Translation invariance up to constant multiple follows from translation invariance of $h$. The choice of constant follows from the definition, which forces $f(0) = 1$. $\Cox$ Before stating our main results, we intoduce a few properties of the random analytic function $f$. \[pr:increasing\] $f(\overline{z}) = \overline{f(z)}$ and $\|f (a + bi)\|$ is increasing in $\|b\|$. [Proof:]{} Invariance under conjugation is evident from the construction of $f$. For $a,b \in \R$, $$\begin{aligned} \log \|f (a + bi)\| & = & \sum_* \log \left \| 1 + \frac{a+bi}{x} \right \| \\[2ex] & = & \frac{1}{2} \sum_* \log \left [ \left ( 1 + \frac{a}{x} \right )^2 + \left ( \frac{\|b\|}{x} \right )^2 \right ] \, .\end{aligned}$$ Each term of the sum is increasing in $\|b\|$. $\Cox$ The random function $f$, being almost surely an entire analytic function, almost surely possesses an everywhere convergent power series $$f(z) = \sum_{n=0}^\infty e_n z^n \, .$$ By construction $f(0) = 1$, hence $e_0 = 1$. The function $f$ is the uniform limit on compact sets of $\disp f_M := \exp \left ( \int \log (1 - z/x) \, dN_M (x) \right )$. The Taylor coefficients $e_{M,n}$ of $f_M$ are the elementary symmetric functions of the negative reciprocals of the points of $N_M$: $$e_{M,k} = e_k \left ( \{ -1/x : N_M (x) = 1 \} \right ) \, .$$ It follows that $e_{M,k} \to e_k$ as $M \to \infty$ for each fixed $k$. Thus we may conceive of $e_k$ as the $k^{th}$ elementary symmetric function of an infinite collection of values, namely the negative reciprocals of the points of the Poisson process. The infinite sum defining this symmetric function is not absolutely convergent but converges conditionally in the manner described above. We do not know a simple form for the marginal distribution of $e_k$ except in the case $k=1$. To see the distribution of $e_1$, observe that the negative reciprocals of the points of a unit intensity Poisson process are a point process with intensity $dx/x^2$. Summing symmetrically in the original points is the same as summing the negative reciprocals, excluding those in $[-\ee , \ee]$, and letting $\ee \to 0$. By a well known construction of the stable laws (see, e.g. [@durrett4 Section 3.7]), this immediately implies: \[pr:cauchy\] The law of $e_1$ is a symmetric Cauchy distribution. $\Cox$ While we have not before seen a systematic study of symmetric functions of points of an infinite Poisson process, symmetric functions of IID collections of variables have been studied before. These were first well understood in Rademacher variables (plus or minus one with probability $1/2$ each). It was shown in [@mori-szekely Theorem 1] that the marginal of $e_k$, suitably normalized, is the value of the $k^{th}$ Hermite polynomial on a standard normal random input. This was extended to other distributions, the most general result we know of being the one in [@major1999]. Main result and reduction to coefficient analysis {#ss:reduction} ------------------------------------------------- The random analytic function $f$ is the object of study for the remainder of the paper. Our main result is as follows, the proof of which occupies most of the remainder of the paper. \[th:main\] As $k \to \infty$, the zero set of $f^{(k)}$ converges in distribution to a uniform random translate of the integers. We prove the main result via an analysis of the Taylor coefficients of $f$, reducing Theorem \[th:main\] to the following result. \[th:a\] Let $a_{k,r} := [z^r] f^{(k)} (z)$. There are random quantities $\{ A_k \}_{k \geq 1}$ and $\{ \theta_k \}_{k \geq 1}$ such that $$\begin{aligned} a_{k,r} & = & A_k \left [ \cos \left ( \theta_k - \frac{r \pi}{2} \right ) + o_k(1) \right ] \, \cdot \frac{\pi^r}{r!} \mbox{ in probability } \label{eq:a1} \\[1ex] \sum_{r = 1}^\infty M^r \frac{ \|a_{k,r}\| }{A_k} & < & \infty \mbox{ with probability } \, 1 - o(1), \label{eq:a2}\end{aligned}$$ for any $M > 0$. The use of the term “in probability” in the first statement means that for every $\ee > 0$ the quantity $$\P \left ( \left \| \frac{r!}{\pi^r A_k} a_{k,r} - \cos \left( \theta_k - \frac{r \pi}{2} \right) \right \| > \ee \right )$$ goes to zero for fixed $r$ as $k \to \infty$. A surprising consequence of this result is that the signs of the coefficients $\{ e_k \}$ are periodic with period 4. In particular, $e_k$ and $e_{k+2}$ have opposite signs with probability approaching 1 as $k \to \infty$. It is interesting to compare this with simpler models, such as the Rademacher model in [@mori-szekely] in which a polynomial $g$ has $n$ zeros, each of them at $\pm 1$, with signs chosen by independent fair coin flips. The number of positive roots will be some number $b = n/2 + O(\sqrt{n})$. Once $n$ and $b$ are determined, the polynomial $g$ is equal to $(z-1)^b (z+1)^{n-b}$. The coefficients of $g$ are the elementary symmetric functions of $b$ ones and $n-b$ negative ones. The signs of these coefficients have 4-periodicity as well ([@mori-szekely Remark 4]). An analogue of Theorem \[th:a\] in the case of IID variables with a reasonably general common distribution appears in [@major1999] (see also [@sneha-thesis] for extensions). The proofs, in that case as well as in the present paper, are via analytic combinatorics. We know of no elementary argument for the sign reversal between $e_k$ and $e_{k+2}$. [Proof of Theorem \[th:main\] from Theorem]{} \[th:a\]: We assume the conclusion of Theorem \[th:a\] holds and establish Theorem \[th:main\] in the following steps. Let $\theta_k$ and $A_k$ be as in the conclusion of Theorem \[th:a\]. [^4] Let $\psi_k (x) := \cos (\pi x - \theta_k)$. Fix any $M > 0$. Then $$\label{eq:a3} \sup_{x \in [-M,M]} \left \| \frac{f^{(k)} (x)}{A_k} - \psi_k (x) \right \| \to 0 \mbox{ in probability as } k \to \infty \, .$$ To prove this, use the identity $\cos (\theta_k - r \pi / 2) = (-1)^j \cos (\theta_k)$ when $r = 2j$ and $(-1)^j \sin (\theta_k)$ when $r = 2j+1$ to write $$\begin{aligned} \psi_k (x) & = & \cos(\theta_k) \cos (\pi x) + \sin (\theta_k) \sin (\pi x) \\[2ex] & = & \cos (\theta_k) \left [ 1 - \frac{\pi^2 x^2}{2!} + \cdots \right ] + \sin (\theta_k) \left [ \pi x - \frac{\pi^3 x^3}{3!} + \cdots \right ] \\[2ex] & = & \sum_{r=0}^\infty \cos \left ( \theta_k - \frac{r\pi}{2} \right ) \frac{\pi^r}{r!} x^r \, .\end{aligned}$$ This last series is uniformly convergent on $[-M,M]$. Therefore, given $\ee > 0$ we may choose $L$ large enough so that $$\label{eq:a1proof} \sup_{x \in [-M,M]} \left \| \psi_k (x) - \sum_{r=0}^L \cos \left ( \theta_k - \frac{r\pi}{2} \right ) \frac{\pi^r}{r!} x^r \right \| < \frac{\ee}{3} \, .$$ By , we may choose $L$ larger if necessary, in order to ensure that $$\label{eq:a2proof} \left \| \sum_{r=L+1}^\infty \frac{ a_{k,r} }{A_k} x^r \right \| < \frac{\ee}{3}$$ for all $x \in [-M,M]$. Fix such an $L$ and use the power series for $f^{(k)}$ to write $$\label{eq:parts} \frac{f^{(k)} (x)}{A_r} - \psi_k (x) = \left ( \sum_{r=0}^L \frac{a_{k,r}}{A_k} x^r - \psi_k (x) \right ) + \sum_{r=L+1}^\infty \frac{a_{k,r}}{A_k} x^r \, .$$ Putting together with shows that the first term on the right-hand side of is at most $\ee / 3 + \sum_{r=0}^L \xi_r$ where $\xi_r$ is the term of that is $o_k (1)$ in probability. By we may choose $k$ large enough so that $\ee / 3 + \sum_{r=0}^L \xi_r < 2 \ee / 3$ with probability at least $1 - \ee / 2$. Thus, we obtain $$\sup_{x \in [-M,M]} \left \| \frac{f^{(k)} (x)}{A_k} - \psi_k (x) \right \| \leq \ee$$ with probability at least $1 - \ee$, etablishing . Let $\eta_k (x) := - \pi \sin (\pi x - \theta_k)$. Fix any $M > 0$. Then $$\label{eq:a4} \sup_{x \in [-M,M]} \left \| \frac{f^{(k+1)} (x)}{A_k} - \eta_k (x) \right \| \to 0 \mbox{ in probability as } k \to \infty \, .$$ The argument is the same as in Step 1, except that we use the power series $f^{(k+1)} (x) =\sum_{r=1}^\infty r a_{k,r} x^{r-1}$ in place of $f^{(k)} (x) = \sum_{r=0}^\infty a_{k,r} x^r$ and $\disp \eta_k (x) = \sum_{r=1}^\infty \cos (\theta_k - r\pi/2) \frac{\pi^r}{(r-1)!} x^{r-1}$. On any interval $[-M,M]$, the zero set of $f^{(k)}$ converges to the zero set of $\psi_k$ in probability. More precisely, for each $\ee > 0$, if $k$ is large enough, then except on a set of probability at most $\ee$, for each zero of $f^{(k)}$ in $[-M+2\ee , M-2\ee]$ there is a unique zero of $\psi_k$ within distance $2 \ee$ and for each zero of $\psi_k$ in $[-M+2\ee , M-2\ee]$ there is a unique zero of $f^{(k)}$ within distance $2 \ee$. This follows from Steps 1 and 2 along with the following fact applied to $\psi = \psi_k$, $\psit = f^{(k)}$, $I = [-M,M]$ and $c = 1/2$. \[lem:unif zeros\] Let $\psi$ be any function of class $C^1$ on an interval $I := [a,b]$. Suppose that $\min \{ \\|psi\| , \|\psi'\| \} \geq c$ on $I$. For any $\ee > 0$, let $I^\ee$ denote $[a+\ee , b - \ee]$. Let $\ee < c^2$ be positive, and suppose that a $C^1$ function $\psit$ satisfies $\|\psit - \psi\| \leq \ee$ and $\|\psit' - \psi'\| \leq c/2$ on $I$. Then the zeros of $\psi$ and $\psit$on $I$ are in correspondence as follows. (i) For every $x \in I^{\ee/c}$ with $\psi (x) = 0$ there is an $\xt \in I$ such that $\psit (\xt) = 0$ and $\|\xt - x\| \leq \ee / c$. This $\xt$ is the unique zero of $\psit$ in the connected component of $\{ \|\psi\| < c \}$ containing $x$. (ii) For every $\xt \in I^{\ee/c}$ with $\psit (\xt) = 0$ there is a $x \in I$ with $\psi (x) = 0$. This $x$ is the unique zero of $\psi$ in the connected component of $\{ \|\psi\| < c \}$ containing $x$. [Proof:]{} For $(i)$, pick any $x \in I^{\ee / c}$ with $\psi (x) = 0$. Assume without loss of generality that $\psi' (x) > 0$ (the argument when $\psi' (x) < 0$ is completely analogous). On the connected component of $\|\psi\| \leq c$ one has $\psi' > c$. Consequently, moving to the right from $x$ by at most $\ee / c$ finds a value $x_2$ such that $\psi (x_2) \geq \ee$, moving to the left from $x$ by at most $\ee / c$ finds a value $x_2$ such that $\psi (x_2) \leq -\ee$, and $\psi'$ will be at least $c$ on $[x_1 , x_2]$. We have $\|\psit - \psi\| \leq \ee$, whence $\psit (x_1) \leq 0 \leq \psit (x_2)$, and by the Intermediate Value Theorem $\psit$ has a zero $\xt$ on $[x_1,x_2]$. To see uniqueness, note that if there were two such zeros, then there would be a zero of $\psit'$, contradicting $\|\psit' - \psit\| < c/2$ and $\|\psi'\| > c$. To prove $(ii)$. pick $\xt \in I^{\ee / c}$ with $\psit (\xt) = 0$. Then $\|\psi (\xt)\| \leq \ee \leq c$ whence $\|\psi' (\xt)\| > c$. Moving in the direction of decrease of $\|\psi (\xt)\|$, $\|\psi'\|$ remains at least $c$, so we must hit zero within a distance of $\ee / c$. Uniqueness follows again because another such zero would imply a critical point of $\psi$ in a region where $\|\psi\| < c$. $\Cox$ Because convergence in distribution is a weak convergence notion, it is equivalent to convergence on every $[-M,M]$. We have therefore proved that the zero set of $f^{(k)}$ converges in distribution to a random translate of the integers. On the other hand, Corollary \[cor:TI\] showed that the zero set of $f^{(k)}$ is translation invariant for all $k$. This implies convergence of the zero set of $f^{(k)}$ to a uniform random translation of $\Z$, and completes the proof of Theorem \[th:main\] from Theorem \[th:a\]. $\Cox$ Estimating coefficients {#sec:sigma} ======================= Overview -------- The coefficients $e_k := [z^k] f(z)$ will be estimated via the Cauchy integral formula $$\label{eq:cauchy} e_k = \frac{1}{2 \pi i} \int z^{-k} f(z) \frac{dz}{z} \, .$$ Denote the logarithm of the integrand by $\phi_k (z) := \log f (z) - k \log z$. Saddle point integration theory requires the identification of a saddle point $\sigma_k$ and a contour of integration $\Gamma$, in this case the circle through $\sigma_k$ centered at the origin, with the following properties. 1. $\sigma_k$ is a critical point of $\phi$, that is, $\phi' (\sigma_k) = 0$. 2. The contribution to the integral from a arc of $\Gamma$ of length of order $\phi'' (\sigma_k)^{-1/2}$ centered at $\sigma_k$ is asymptotically equal to $e^{\phi (\sigma_k)} \sqrt{2 \pi / \phi''(\sigma_k)}$. 3. The contribution to the integral from the complement of this arc is negligible. In this case we have a real function $f$ with two complex conjugate saddle points $\sigma_k$ and $\overline{\sigma_k}$. Accordingly, there will be two conjugate arcs contributing two conjugate values to the integral while the complement of these two arcs contributes negligibly. One therefore modifies $(i)$–$(iii)$ to: 1. $\sigma_k$ and $\overline{\sigma_k}$ are critical points of $\phi$, and there are no others on the circle $\Gamma$, centered at the origin, of radius $\|\sigma_k\|$. 2. The contribution to the integral from a arc of $\Gamma$ of length of order $\phi'' (\sigma_k)^{-1/2}$ centered at $\sigma_k$ is asymptotically equal to $e^{\phi (\sigma_k)} \sqrt{2 \pi / \phi''(\sigma_k)}$. 3. The contribution to the integral from the complement of the two conjugagte arcs is negligible compared to the contribution from either arc. Note that $(iii')$ leaves open the possibility that the two contributions approximately cancel, leaving the supposedly negligible term dominant. Locating the dominant saddle point {#ss:sigma} ---------------------------------- The logarithm of the integrand in , also known as the phase function, is well defined up to multiples of $2 \pi i$. We denote it by $$\phi_k (z) := - k \log z + \sum_* \log \left ( 1 - \frac{z}{x} \right ) \, .$$ When $k=0$ we denote $\sum_* \log (1 - z/x)$ simply by $\phi (z)$. \[pr:derivs\] For each $k,r$, the $r^{th}$ derivative $\phi_k^{(r)}$ of the phase function $\phi_k$ is the meromorphic function defined by the almost surely convergent sum $$\label{eq:phi r} \phi_k^{(r)} (z) = (-1)^{r-1} (r-1)! \left [ - \frac{k}{z^r} + \sum_* \frac{1}{(z-x)^r} \right ] \, .$$ Thus in particular, $$\phi_k' (z) = - \frac{k}{z} + \sum_* \frac{1}{z-x} \, .$$ [Proof:]{} When $r=1$, convergence of and the fact that this is the derivative of $\phi$ is just Lemma \[lem:h TI\] and the subsequent proof of Theorem \[th:f and TI\] in which $f$ is constructed from $h$. For $r \geq 2$, with probability 1 the sum is absolutely convergent. $\Cox$ The main work of this subsection is to prove the following result, locating the dominant saddle point for the Cauchy integral. \[th:sigma\] Let $E_{M,k}$ be the event that $\phi_k$ has a unique zero, call it $\sigma_k$, in the ball of radius $M k^{1/2}$ about $i k / \pi$. Then $\P (E_{M,k}) \to 1$ as $M,k \to \infty$ with $k \geq 4\pi^{2}M^{2}$. This is proved in several steps. We first show that $\phi_k' (i k / \pi)$ is roughly zero, then use estimates on the derivatives of $\phi$ and Rouch[é]{}’s Theorem to bound how far the zero of $\phi_k'$ can be from $i k / \pi$. The function $\phi_k'$ may be better understood if one applies the natural scale change $z = ky$. Under this change of variables, $$\phi_k' (z) = - \frac{1}{y} + \sum_* \frac{1/k}{y - x/k} \, .$$ Denote the second of the two terms by $$h_k (y) := \sum_* \frac{1/k}{y - x/k} \, .$$ We may rewrite this as $\disp h_k (Y) = \int \frac{1}{y - x} dN^{(k)} (x)$ when $N^{(k)}$ denotes the rescaled measure defined by $N^{(k)} (A) = k^{-1} N(kA)$. The points of the process $N^{(k)}$ are $k$ times as dense and $1/k$ times the mass of the points of $N$. Almost surely as $k \to \infty$ the measure $N^{(k)}$ converges to Lebesgue measure. In light of this it is not surprising that $h_k (y)$ is found near $\disp \int \frac{1}{z-y} \, dy$. We begin by rigorously confirming this, the integral being equal to $- \pi \, {\rm sgn} \, \Im \{ z \}$ away from the real axis. \[lem:interchange\] If $z$ is not real then $$\E \int_{} \frac{1}{\|z - x\|^{m}} dN(x) = \lim_{M\to\infty} \E \int \frac{1}{\|z - x\|^{m}} dN_{M}(x),$$ for $m \geq 2$, and $$\E \int_{} \frac{1}{(z - x)^{m}} dN(x) = \lim_{M\to\infty} \E \int \frac{1}{(z - x)^{m}} dN_{M}(x),$$ for $m \geq 1$. [Proof:]{} The first equality holds trivially by Monotone Convergence Theorem. Next, write $\RM$ as the number of points of the process $N$ withing $[-M,M]$, and $L = 2\Im(z)$. Then, for $m = 2$, $$\begin{aligned} \E \int \frac{1}{\|z - x\|^{2}} dN_{M}(x) & = & \E \sum_{j: \|X_{j}\| \leq M} \frac{1}{\Im(z)^{2} + (\Re(z) - X_{j})^{2}} \\[1ex] & \leq & \E \left( \frac{\mathcal{R}_{L}}{\Im(z)^{2}} \right) + \, \E \sum_{j: L \leq \|X_{j}\| \leq M} \frac{4}{\|X_{j}\|^{2}} \\[1ex] & \leq & \frac{2L}{\Im(z)^{2}} + \frac{4}{L}.\end{aligned}$$ Therefore, as $\Im(z) \neq 0$, $\E \int_{} \frac{1}{\|z - x\|^{2}} dN(x) < \infty$, and moreover, $\E \int_{} \frac{1}{\|z - x\|^{m}} dN(x) < \infty, \forall m \geq 2$. Thus, by Dominated Convergence Theorem, $$\E \int_{} \frac{1}{(z - x)^{m}} dN(x) = \lim_{M\to\infty} \E \int \frac{1}{(z - x)^{m}} dN_{M}(x)$$ holds for $m \geq 2$. We shall now show the above to hold true for $m = 1$. Note that $$\E \left[ \left\| \int \frac{1}{z - x} dN_{M}(x) \right\|^{2} \right] = \E \int \frac{1}{\|z - x\|^{2}} dN_{M}(x) \, + \, \E \sum_{j \neq k: \|X_{j}\|, \|X_{k}\| \leq M} \frac{1}{(z - X_{j})(\overline{z} - X_{k})}.$$ The first term in the above equation converges to $\E \int_{} \frac{1}{\|z - x\|^{2}} dN(x)$ as $M \to \infty$. As for the second part, $$\E \sum_{j \neq k: \|X_{j}\|, \|X_{k}\| \leq M} \frac{1}{(z - X_{j})(\overline{z} - X_{k})} = \E \left[ \RM (\RM - 1) \cdot \E \left( \frac{1}{(z - \Uni_{2})(\overline{z} - \Uni_{2})} \right) \right],$$ where $\Uni_{1}$ and $\Uni_{2}$ are i.i.d. Uniform$(-M,M)$ random variables. So, $$\begin{aligned} \E \sum_{j \neq k: \|X_{j}\|, \|X_{k}\| \leq M} \frac{1}{(z - X_{j})(\overline{z} - X_{k})} & = & \left( \int_{-M}^{M} \frac{1}{z - u} du \right)^{2} \\[1ex] & = & \left[ - \log \left\| \frac{M - z}{M + z} \right\| - i \arctan \left ( \frac{M - \Re(z)}{\Im(z)} \right) \right. \\ & & \left. + \, i \arctan \left( \frac{- M - \Re(z)}{\Im(z)} \right) \right]^{2} \\[1ex] & \longrightarrow & - \pi^{2}, \text{ as } M \to \infty.\end{aligned}$$ Thus the quantities $\left \{ \E \left[ \left\| \int \frac{1}{z - x} dN_{M}(x) \right\|^{2} \right], M > 0 \right \}$ have a uniform upper bound - let us call it $B(z)$. Then, given $\ee > 0$, $$\begin{aligned} \E \left[ \left\| \int \frac{1}{z - x} dN_{M}(x) \right\| \cdot \one_{\left\| \int \frac{1}{z - x} dN_{M}(x) \right\| \geq K}\right] & \leq & \frac{1}{K} \cdot \E \E \left[ \left\| \int \frac{1}{z - x} dN_{M}(x) \right\|^{2} \right] \\[1ex] & \leq & \frac{B(z)}{K} < \ee,\end{aligned}$$ for $K > \frac{B(z)}{\ee}$. Thus, if $z$ is not real, $\left \{ \E \left[ \int \frac{1}{z - x} dN_{M}(x) \right], M > 0 \right \}$ is a uniformly integrable collection, and hence, converges in $L_{1}$. $\Cox$ \[pr:1\] If $z$ is not real then $$\label{eq:1} \E \left [ \int_* \frac{1}{z - x} \, dN(x) \right ] = \mp i \pi$$ with the negative sign if $z$ is in the upper half plane and the positive sign if $z$ is in the lower half plane. If $z$ is not real and $m \geq 2$ then $$\label{eq:2} \E \left [ \int_* \frac{1}{(z - x)^m} \, dN(x) \right ] = 0 \, .$$ [Proof:]{} If $\RM$ denotes the number of Poisson points in \[-M,M\], then conditioning on $\RM$, the poisson points $X_j$ that are contained in $[-M,M]$ are identically and independently distributed as Uniform$[-M,M]$. Then, $$\E \left[ \left. \int \frac{1}{z - x} \, dN_M(x) \right\| \RM \right] = \RM \cdot \E \left ( \frac{1}{z-\Uni} \right ),$$ where $\Uni \sim$ Uniform$[-M,M]$. Writing $z = r e^{i\theta}$, we get, $$\begin{aligned} \E \left[ \left. \int \frac{1}{z - x} \, dN_M(x) \right\| \RM \right] & = & \frac{\RM}{2M} \int_{x \in [-M,M]} \frac{1}{r\cos\theta + i r\sin\theta - x}dx\\[1ex] & = & \RM\left[ \frac{-1}{2M} \log \left\| \frac{M - z}{M + z} \right\| - \frac{i}{2M} \arctan \left ( \frac{M - r\cos\theta}{r\sin\theta} \right ) \right.\\ & & + \, \left. \frac{i}{2M} \arctan \left( \frac{- M - r\cos\theta}{r\sin\theta} \right) \right]\\[1ex] \implies \E \left[ \int \frac{1}{z - x} \, dN_M(x) \right] & = & - \log \left\| \frac{M - z}{M + z} \right\| - i \arctan \left ( \frac{M - r\cos\theta}{r\sin\theta} \right )\\ & & + \, i \arctan \left ( \frac{- M - r\cos\theta}{r\sin\theta} \right )\end{aligned}$$ since, $\RM \sim $ Poisson$(2M)$. Taking $M \to \infty$, by Lemma \[lem:interchange\] we get, $$\E \left [ \int_* \frac{1}{z - x} \, dN(x) \right ] = - \pi i,$$ for $z$ in the upper half plane, and, $$\E \left [ \int_* \frac{1}{z - x} \, dN(x) \right ] = \pi i,$$ for $z$ in the lower half plane, where the interchange of limits and expectation is by Lemma \[lem:interchange\]. Now fix $m \geq 2$ and $z \notin \R$. $$\begin{aligned} \E \left[ \left. \int \frac{1}{(z - x)^m} \, dN_M(x) \right\| \RM \right] & = & \RM \cdot \E \left[ \frac{1}{(z - \Uni)^m} \right]\\[1ex] & = & \frac{\RM}{2M} \cdot \frac{1}{m - 1} \left\{ \frac{1}{(z-M)^{m-1}} - \frac{1}{(M+z)^{m-1}} \right\}.\\[1ex] \implies \E \left[ \int \frac{1}{(z - x)^m} \, dN_M(x) \right] & = & \frac{1}{m-1}\left\{ \frac{1}{(z-M)^{m-1}} - \frac{1}{(M+z)^{m-1}} \right\}.\end{aligned}$$ Thus, using Lemma \[lem:interchange\], $$\E \left[ \int_* \frac{1}{(z - x)^m} \, dN(x) \right] = \lim_{M \to \infty} \frac{1}{m-1}\left\{ \frac{1}{(z-M)^{m-1}} - \frac{1}{(M+z)^{m-1}} \right\} = 0.$$ $\Cox$ The next proposition and its corollaries help us to control how much the functions $\phi_k$ and $h_k$ can vary. These will be used first in Lemma \[lem:4.3.6\], bounding $h_k$ over a ball, then in Section \[ss:saddle est\] to estimate Taylor series involving $\Phi_k$. We begin with a general result on the variance of a Poisson integral. \[pr:3.5a\] Let $\psi : \R \to \C$ be any bounded function with $\int \|\psi (x)\|^2 \, dx < \infty$. Let $Z$ denote the compensated Poisson integral of $\psi$, namely $$Z := \lim_{M \to \infty} \left [ \int \psi (x) \, dN_M (x) - \int_{-M}^M \psi (x) \, dx \right ] \, .$$ Then $Z$ is well defined and has finite variance given by $$\E \|Z\|^2 = \int \|\psi (x)\|^2 \, dx \, .$$ [Proof:]{} This is a standeard result but the proof is short so we supply it. Let $$Z_M := \int \psi (x) \, dN_M (x) - \int_{-M}^M \psi (x) \, dx$$ and let $\Delta_M := Z_M - Z_{M-1}$ denote the increments. We apply Kolmogorov’s Three Series Theorem to the independent sum $\sum_{M=1}^\infty \Delta_M$, just as in the proof of Lemma \[lem:h TI\]. Hypothesis (i) is satisfied because $\int_M^{M+1} \|\psi\|$ goes to zero. Hypothesis (ii) is satisfied because $\E \Delta_M = 0$ for all $M$. To see that hypothesis (iii) is satisfied, observe that $\E \|\Delta_M\|^2 = \int \|\psi (x)\|^2 \one_{M-1 \leq \|x\| \leq M} \, dx$, the summability of which is equivalent to our assumption that $\psi \in L^2$. We conclude that the limit exists almost surely. By monotone convergence as $M \to \infty$, $\Var (Z) = \int \|\psi\|^2$. $\Cox$ Define $$W_r (z) := \int_* (z-x)^{-r} \, dN(x) \, .$$ If $\alpha > 1$ and $\lambda$ is real, the intergal $\int \|z-x\|^{-\alpha} \, dx$ is invariant under $z \mapsto z + \lambda$ and scales by $\lambda^{1-\alpha}$ under $z \mapsto \lambda z$. Plugging in $\psi (x) = (z - x)^{-r}$ therefore yields the following immediate corollary. \[cor:3.5b\] Let $z$ have nonzero imaginary part and let $r \geq 2$ be an integer. Then $W_r (z)$ is well defined and there is a positive constant $\gamma_r$ such that $$\E \|W_r (z)\|^2 = \frac{\gamma_r}{\|\Im \{ z \}\|^{2r-1}} .$$ $\Cox$ In the case of $r=1$ we obtain the explicit constant $\gamma_1 = 1$: $$\E \|W_1(z) \mp \pi i \|^2 = \frac{\pi}{\|\Im(z)\|}.$$ To see this, compute $$\begin{aligned} \E \left[ \left. \int \frac{1}{\| z - x \|^2} \, dN_M(x) \right\| \RN \right] & = & \RN \cdot \frac{1}{2N} \int_{x \in [- N, N]} \frac{1}{(z - x) \cdot (\bar{z} - x)} dx\\[1ex] & = & \frac{\RN}{2N (\bar{z} - z)} \left [ \int_{x \in [- N, N]} \frac{1}{z - x}dx - \int_{x \in [- N, N]} \frac{1}{\bar{z} - x}dx \right ]\\[1ex] & = & \frac{1}{\bar{z} - z} \left \{ \E \left [ \left. \int \frac{1}{z - x} \, dN_M(x) \right\| \RN \right ] - \E \left [ \left. \int \frac{1}{\bar{z} - x} \, dN_M(x) \right\| \RN \right ] \right \}.\end{aligned}$$ Thus, taking expectations and by Lemma \[lem:interchange\] $$\E \left [ \int_* \frac{1}{\|z - x\|^2} \, dN(x) \right ] = \frac{1}{\bar{z} - z} \left \{ \E \left [ \int_* \frac{1}{z - x} \, dN(x) \right ] - \E \left [ \int_* \frac{1}{\bar{z} - x} dN(x) \right ] \right \}.$$ Proposition \[pr:1\] shows the difference of expectations on the right-hand side to be $- 2 i \pi$, yielding $\gamma_1 = \pi$. \[cor:3.5c\] For $y$ with nonzero imaginary part and $r \geq 1$, $W_r (ky)$ has variance $\E [ \Re \{ W - \overline{W} \}^2 + \Im \{ W - \overline{W} \}^2] = k^{-1/2} \gamma_r (y)$. It follows (with $\delta_{1,r}$ denoting the Kronecker delta), that $$\label{eq:W_r} \phi_k^{(r)} (ky) = - i \pi \delta_{1,r} + (r-1)! k^{1-r} \left ( \frac{-1}{y} \right )^r + O \left ( k^{1/2 - r} \right )$$ in probability as $k \to \infty$. [Proof:]{} Let $N^{(k)}$ denote a Poisson law of intensity $k$, rescaled by $k^{-1}$. In other words, $N^{(k)}$ is the average of $k$ independent Poisson laws of unit intensity. Under the change of variables $u = x/k$, the Poisson law $dN (x)$ becomes $k dN^{(k)} (u)$. Therefore, $$\begin{aligned} W_r (ky) & = & \int_* \frac{1}{(ky-x)^r} \, dN(x) \\[2ex] & = & k^{1-r} \int_* \frac{1}{(y-u)^r} \, dN^{(k)} (u) \\[2ex] & = & k^{1-r} \left ( \frac{1}{k} \sum_{j=1}^k W_r^{[j]} \right )\end{aligned}$$ where $\{ W_r^[1] , \ldots , W_r^{[k]} \}$ are $k$ independent copies of $W_r (y)$. Because $W_r (y)$ has mean $- i \pi \delta_{1,r}$ and variance $\gamma_r (y)$, the variance of the average is $k^{-1/2} \gamma_r (y)$. The remaining conclusion follows from the expression for $\phi_k^{(r)}$ and the fact that a random variable with mean zero and variance $V$ is $O(V^{1/2})$ in probability. $\Cox$ Uniformizing the estimates -------------------------- At some juncture, our pointwise estimates need to be strengthened to uniform estimates. The following result is a foundation for this part of the program. \[lem:lipschitz\] Fix a compact set $K$ in the upper half plane and an integer $r \geq 1$. There is a constant $C$ such that for all integers $k \geq 1$, $$\E \sup_{z \in K} \left \| h_k^{(r)} (z) \right \| \leq C k^{-1/2} \, .$$ [Proof:]{} Let $F^{(k)}$ denote the CDF for the random compensated measure $N^{(k)} - dx$ on $\R^+$, thus $F(x) = N^{(k)}[0,x] - x$ when $x > 0$ and $F(x) = x - N^{(k)}[x,0]$ when $x < 0$. We have $$h_k^{(r)} (z) = \int_* C (z-x)^{-r-1} \, dN(x) = \int_* C (z-x)^{-r-1} \, dF^{(k)} (x)$$ because $\int_* (z-x)^{-r-1} \, dx = 0$. This leads to $$\begin{aligned} && \E \sup_{z \in K} \left \| h_k^{(r)} (z) \right \| \\[1ex] & \leq & \lim_{M \to \infty} \E \sup_{z \in K} \left \| \int_0^M \frac{1}{(z-x)^r} \, dF^{(k)}(x) \right \| + \E \sup_{z \in K} \left \| \int_{-M}^0 \frac{1}{(z-x)^r} \, dF^{(k)}(x) \right \| \, .\end{aligned}$$ The two terms are handled the same way. Integrating by parts, $$\int_0^M (z-x)^{-r} \, dF^{(k)}(x) = (z-x)^{-r} N[0,M] - \int - r (z-x)^{-r-1} \, F^{(k)} (x) \, dx .$$ This implies that $$\begin{aligned} && \E \sup_{z \in K} \left \| h_k^{(r)} (z) \right \| \\[2ex] & \leq & \lim_{M \to \infty} \left [ \E \|F^{(k)}(M)\| \sup_{z \in K} \|z-x\|^{-r} + \int_0^M \sup_{z \in K} r \|z-x\|^{-r-1} \|F^{(k)} (x)\| \right ] \, dx \\[2ex] & \leq & C_K \lim_{M \to \infty} \left ( M^{-r + 1/2} + k^{-1/2} \right ) \, .\end{aligned}$$ Sending $M$ to infinity gives the conclusion of the lemma. $\Cox$ \[cor:lipschitz\] \ (i) $\disp \sup_{z \in K}\|h_k^{(r)} (z)\| = O(k^{-1/2})$ in probability. (ii) $h_k$ and its derivatives are Lipschitz on $K$ with Lipschitz constant $O(k^{-1/2})$ in probability. (iii) For $r \geq 2$, the $O(k^{-1/2})$ term in the expression for $\phi_k^{(r)} (ky)$ is uniform as $y$ varies over compact sets of the upper half plane. [Proof:]{} Conclusion $(i)$ is Markov’s inequality. Conclusion $(ii)$ follows because any upper bound on a function $\|g'\|$ is is a Lipschitz constant for $g$. Conclusion $(iii)$ follows from the relation between $h_k$ and $\phi_k$. $\Cox$ \[lem:4.3.6\] For any $c > 0$, $$\P \left [ \sup \left \{ \left \|h_k (y) + i \pi \right\| \, : \, \|y - \frac{i}{\pi}\| \leq M k^{-1/2} \right \} \geq c M k^{-1/2} \right ] \to 0$$ as $M \to \infty$ with uniformly in $k \geq 4 \pi^2 M^2$. [Proof:]{} Fix $c , \ee > 0$. Choose $L > 0$ such that the probability of the event $G$ is at most $\ee/2$, where $G$ is the event that the Lipschitz constant for some $h_k$ on the ball $B(i\pi , 1/(2\pi))$ is greater than $L$. Let $B$ be the ball of radius $M k^{-1/2}$ about $i/\pi$. The assumption $k \geq 4 \pi^2 M^2$ guarantees that $B$ is a subset of the ball $B(i\pi , 1 / (2\pi))$ over which the Lipschitz constant was computed. Let $y$ be any point in $B$. The ball of radius $\rho := c M k^{-1/2} \ee / (2L)$ about $y$ intersects $B$ in a set whose area is at least $\rho^2 \sqrt{3}/2$, the latter being the area of two equilateral triangles of side $\rho$. If $\|h_k (y) + i / \pi\| \geq c M k^{-1/2}$ and $G$ goes not occur, then $\|h_k (u) + i / \pi\| \geq (1/2) c M k^{-1/2}$ on the ball of radius $\rho$ centered at $y$. Now we compute in two ways the expected measure $\E \|S\|$ of the set $S$ of points $u \in B$ such that $\|h_k (u) + i \pi\| \geq \rho$. First, by what we have just argued, $$\label{eq:Q1} \E \|S\| \geq \frac{\sqrt{3}}{2} \rho^2 \left ( Q - \frac{\ee}{2} \right ) = \left ( Q - \frac{\ee}{2} \right ) \sqrt{\frac{3 c^2 \ee^2}{16 L^2}} \frac{M^2}{k}$$ where $Q$ is the probability that there exists a $y \in B$ such that $\|h_k (y) + i / \pi\| \geq c M k^{-1/2}$. Secondly, by Proposition \[pr:1\] and the computation of $\gamma_1$, for each $u \in B$, $\E h_k (u) + i/\pi = 0$ and $\E \|h_k (u)\|^2 = \pi / k$, leading to $\E \|h_k (u) + i/\pi\| \leq \sqrt{2 \pi / k}$ and hence $$\begin{aligned} \P \left ( \left \| h_k (u) + \frac{i}{\pi} \right \| \geq \rho \right ) & \leq & \frac{\sqrt{2 \pi/k}}{\rho} \\ & = & \frac{\sqrt{2 \pi / k}}{c M k^{-1/2} \ee / (2L)} \\[1ex] & = & \sqrt{\frac{32 \pi L^2}{c^2}} M^{-1/2} \, .\end{aligned}$$ By Fubini’s theorem, $$\label{eq:Q2} \E \|S\| \leq \|B\| \sup_{u \in B} \P \left ( \left \| h_k (u) + \frac{i}{\pi} \right \| \geq c M k^{-1/2} \right ) \leq \pi \frac{M^2}{k} \sqrt{\frac{32 \pi L^2}{c^2}} M^{-1/2} \, .$$ Putting together the inequalities and gives $$Q - \frac{\ee}{2} \leq \sqrt{\frac{512 \pi^3 L^4}{3 c^4 \ee^2}} M^{-1/2} \, .$$ Once $M$ is sufficiently larger that the radical is at most $\ee / 2$, this implies that $Q \leq \ee$. Because $\ee > 0$ was arbitrary, we have shown that $Q \to 0$ as $M \to \infty$ uniformly in $k$, as desired. $\Cox$ [Proof of Theorem]{} \[th:sigma\]: Using Lemma \[lem:4.3.6\] for $c < 1$, we know that $$\P \left [ \sup \left \{ \|h_k (y) + i \pi\| \, : \, \|y - \frac{i}{\pi}\| \leq M k^{-1/2} \right \} \leq c M k^{- 1/2} \right ] \longrightarrow 1, \text{ as, } k\to\infty.$$ Writing $$A_{M,k} = \left \{ \omega : \sup \left \{ \|h_k (y) + i \pi\| \, : \, \|y - \frac{i}{\pi}\| \leq M k^{-1/2} \right \} \leq c M k^{- 1/2} \right \},$$ $\forall \omega \in A_{M,k}$, and all $y$ such that $\|y - \frac{i}{\pi}\| = Mk^{-1/2}$, $$\begin{aligned} \left\| \phi_k(y) (\omega) - \left( - i\pi - \frac{1}{y} \right) \right\| & = & \|h_k(y)(\omega) + i\pi\| \\[1ex] & \leq & cMk^{-1/2} \\[1ex] & = & c \left \| y - \frac{i}{\pi} \right \| \\[1ex] & < & \left \| y - \frac{i}{\pi} \right \| \end{aligned}$$ for $k$ sufficiently large. Thus, by Rouche’s theorem, $ \phi_k(y) (\omega) $ and $y - \frac{i}{\pi}$ have the same number of zeros inside the disc centered at $i/\pi$ of radius $Mk^{- 1/2}$, i.e. exactly one. This implies that, $\P (E_{M,k}) \to 1$ as $M,k \to \infty$ with $k \geq 4\pi^{2}M^{2}$. $\Cox$ The Cauchy integral =================== Dominant arc: saddle point estimate {#ss:saddle est} ----------------------------------- We sum up those facts from the foregoing subsection that we will use to estimate the Cauchy integral in the dominant arc near $\sigma_k$. \[lem:summarizing phi\] \ (i) $\phi' (\sigma_k) = 0$. (ii) $\sigma_k^2 \phi'' (\sigma_k) = k + O(k^{1/2})$ in probability as $k \to \infty$. (iii) If $K$ is the set $\{ z : \|z - \sigma_k\| \leq k/2$ then $\disp \sup_{z \in K} k^3 \phi^{(3)} (z) = O(k)$ in probability. [Proof:]{} The first is just the definition of $\sigma_k$. For the second, using Corollary \[cor:lipschitz\] for $r = 2$ and $y = \frac{i}{\pi}$, the estimate is uniform, hence $$\phi'' (\sigma_k) = \phi'' \left ( \frac{ik}{\pi} \right ) + O \left ( k^{-3/2} \right ) = \frac{-\pi^2}{k} + O \left ( k^{-3/2} \right )$$ in probability. Multiplying by $\sigma_k^2 \sim -k^2 / \pi^2$ gives $(ii)$. The argument for part $(iii)$ is analogous to the argument for part $(ii)$. $\Cox$ For the remainder of the paper, fix a number $\delta \in (1/3 , 1/2)$. Parametrize the circle $\Gamma$ through $\sigma_k$ in several pieces, all oriented counterclockwise, as follows (see Figure \[fig:Gamma\]). Define $\Gamma_1$ to be the arc $\{z: z = \sigma_{k}e^{it}, - k^{-\delta} \leq t \leq k^{-\delta} \}$. Define $\Gamma_1'$ to be the arc $\{z: z = \overline{\sigma}_{k}e^{it}, - k^{-\delta} \leq t \leq k^{-\delta} \}$, so that the arc is conjugate to $\Gamma_1$ but the orientation remains counterclockwise. Define $\Gamma_2$ to be the part of $\Gamma$ in the second quadrant that is not part of $\Gamma_1$, define $\Gamma_3$ to be the part of $\Gamma$ in the first quadrant not in $\Gamma_1$, and define $\Gamma_2'$ and $\Gamma_3'$ to be the respective conjugates. Define the phase function along $\Gamma$ by $$g_k (t) := \phi_{k} (\sigma_k e^{it}) \, .$$ \[fig:Gamma\] ![Parametrization of the circular contour $\Gamma$](Gamma.eps "fig:"){width="2.8in"} \[th:good arcs\] For any integer $r \geq 0$, $$\frac{\disp{\int_{\Gamma_1} \frac{f(z)}{z^{k + r + 1}} dz}} {\disp{k^{-1/2} f(\sigma_k) \sigma_k^{-k-r}}} \longrightarrow i \sqrt{2 \pi}$$ in probability as $k \to \infty$. [Proof:]{} For fixed $k$, the Taylor’s expansion of $g_{k}(t)$ gives us, $$g_k(t) = g_{k}(0) + t g_k'(0) + \frac{t^2}{2} g_k^{(2)}(0) + \frac{t^3}{6} \left( \Re g_{k}^{(3)}(t_1) + i \Im g_{k}^{(3)}(t_2)\right),$$ where $t_1$ and $t_2$ are points that lie between $0$ and $t$.\ By Lemma \[lem:summarizing phi\], $g_{k}'(0) = 0$ and $$g_k^{2}(0) = k + O\left ( k^{1/2} \right )$$ in probability. Thus, $$\sup_{\|t\| \leq k^{- \delta}} \sqrt{k} \left[ \exp \left( \frac{t^2}{2} g_k^{(2)}(0) \right) - \exp \left( - \frac{kt^2}{2} \right) \right] \longrightarrow 0.$$ In addition, $$\sup_{z \in \Gamma_1} \left\| \frac{\sigma_k^r}{z^r} - 1 \right\| \longrightarrow 0,$$ while, Lemma \[lem:summarizing phi\] also gives us $$\sup_{\|t\| \leq k^{- \delta}} \frac{t^3}{6} g_k^{(3)}(t) \longrightarrow 0.$$ Thus, $$\int_{\Gamma_1} \frac{f(z)}{z^{k + r + 1}} dz \; = \; i \int_{- k^{- \delta}}^{k^{- \delta}} \sigma_k^{- r} \exp \left[ g_{k}(0) + \frac{t^2}{2} g_k^{(2)}(0) + \frac{t^3}{6} \left( \Re g_{k}^{(3)}(t_1) + i \Im g_{k,N}^{(3)}(t_2)\right) - i \, r t \right] dt,$$ whence, as $k \to \infty$, $$\sqrt{k} \frac{\int_{\Gamma_1} \frac{f(z)}{z^{k + r + 1}} dz}{\sigma_k^{- r} \exp(g_{k}(0))} - i \sqrt{k} \int_{- k^{- \delta}}^{k^{- \delta}} \exp \left(- \frac{kt^2}{2} \right) dt \longrightarrow 0.$$ Changing variables to $t = u/\sqrt{k}$ shows that when $\delta < 1/2$, the integral is asymptotic to $\sqrt{2 \pi/k}$. Plugging in $g_k (0) = f(\sigma_k) \sigma_k^{- k}$ completes the proof. $\Cox$ Negligble arcs and remainder of proof of {#sec:f} ----------------------------------------- We now show that the Cauchy integral receives negligible contributions from $\Gamma_2, \Gamma_2', \Gamma_3$ and $\Gamma_3'$. By conjugate symmetry we need only check $\Gamma_2$ and $\Gamma_3$; the arguments are identical so we present only the one for $\Gamma_2$. Let $R := \|\sigma_k\|$ and let $\beta$ denote the polar argument of $\sigma_k$, that is, $\beta := \arg (\sigma_k) - \pi / 2$, so that $\sigma_k = i R e^{i \beta}$. By Theorem \[th:sigma\], $\beta = O(k^{-1/2})$ in probability. We define an exceptional event $G_k$ of probability going to zero as follows. > Let $G_k$ be the event that either $R \notin [k/(2\pi) , 2 k/\pi]$ or $\beta > k^{-\delta} / 2$. If $z = i R e^{i \theta}$ is a point of $\Gamma_2$ with polar argument $\theta$, then $\theta$ is at least $k^{-\delta} - \|\beta\|$, hence is at least $(1/2) k^{-\delta}$ on $G_k^c$. Note that the notation suppresses the dependence of $R$ and $\beta$ on $k$, which does not affect the proof of the in-probability result in Lemma \[lem:bad arcs\]. \[lem:bad arcs\] $$\label{eq:big} \frac{\disp{\int_{\Gamma_2} \frac{f(z)}{z^{k + r + 1}} dz}} {\disp{k^{-1/2} f(\sigma_k) \sigma_k^{-k-r}}} \longrightarrow 0$$ in probability as $k \to \infty$. [Proof:]{} Let $z = i R e^{i \theta} \in \Gamma_2$. Our purpose is to show that $\|f(z) z^{-k}\|$ is much smaller that $\|f (\sigma_k) \sigma_k^{-k}\|$. On $\Gamma_2$ we are worried only about the magnitude, not the argument, so we may ignore the $z^{-k}$ and $\sigma^{-k}$ terms, working with $\phi$ rather than with $\phi_k$. This simplifies to $$\label{eq:Re} \phi' (z) = - i \pi + O \left ( k^{-1/2} \right ) \,$$ the estimate being uniform on the part of $\Gamma_2$ with polar argument less than $\pi/2 - \ee$ by part $(iii)$ of Corollary \[cor:lipschitz\]. Let $H_k$ be the exceptional event where the constant in the uniform $O(k^{-1/2})$ term is greater than $k^{1/2-\delta} / 100$, the probability of $H_k$ going to zero according to the corollary. Integrating the derivative of $\Re \{ \phi (z) \}$ along $\Gamma$ then gives $$\label{eq:logs} \log \left \| \frac{f(z)}{f(\sigma_k)} \right \| = \pi \left ( \Im \{ z \} - \Im \{ \sigma_k \} \right ) + O \left ( k^{-1/2} \|z - \sigma_k\| \right ) \, .$$ The first of the two terms is $\pi R (\cos (\theta) - \cos (\beta))$ which is bounded from above by $- (R/2) (\theta^2 - \beta^2)$ which is at most $-(R/4) \theta^2$ on $G_k^c$. The second term is at most $$\frac{k^{1/2 - \delta}}{100} k^{-1/2} (2 R \theta)$$ on $G_k^c \cap H_k^c$, provided that $\theta \leq \pi/2 - \ee$. Combining yields $$\log \left \| \frac{f(z)}{f(\sigma_k)} \right \| \leq - \frac{R}{4} \theta^2 + \frac{k^{-\delta}}{100} (2 R \theta) \, \leq \, - R \theta (\left ( \frac{\theta}{4} - \frac{k^{-\delta}}{50} \right ) \, \leq \, - \frac{R \theta^2}{8}$$ on $\Gamma_2$ as long as the polar argument of $z$ is at most $\pi/2 - \ee$. Decompose $\Gamma_2 = \Gamma_{2,1} + \Gamma_{2,2}$ according to whether $\theta$ is less than or greater than $\pi/2 - \ee$. On $G_k^c$ we know that $\theta \geq (1/2) k^{-\delta}$ and $R \geq k/(2\pi)$, hence on $\Gamma_{2,1}$, $$\log \left \| \frac{f(z)}{f(\sigma_k)} \right \| \leq - \frac{k^{1 - 2\delta}}{64 \pi} \, .$$ Using $d\theta = dz / z$ we bound the desired integral from above by $$\left \| \frac{\disp{\int_{\Gamma_2} \frac{f(z)}{z^{k + r + 1}} dz}} {\disp{k^{-1/2} f(\sigma_k) \sigma_k^{-k-r}}} \right \| \leq \sqrt{k} \int_{\Gamma_2} \left \| \frac{f(z)}{f(\sigma_k)} \right \| \, d\theta \, .$$ On $G_k^c \cap H_k^c$, the contribution from $\Gamma_{2,1}$ is at most $$\label{eq:Gamma_1} \sqrt{k} \, \|\Gamma_2\| \, \exp \left [ - \frac{k^{1 - 2\delta}}{64 \pi} \right ] \, .$$ Finally, to bound the contribution from $\Gamma_{2,2}$, use Proposition \[pr:increasing\] to deduce $\|f(z)\| \leq \|f(z')\|$ where $\Re \{ z' \} = \Re \{ z \}$ and $\Im \{ z' \} = k/(4 \pi)$. Integrating on the line segment between $\sigma_k$ and $z'$ now gives again, and on $G_k^c \cap H_k^c$ the right-hand side is at most $- (k/4) + k^{-\delta} k < -k/8$ once $k \geq 8$. This shows the contribution from $\Gamma_{2,2}$ to be at most $\ee R e^{-k/8}$. Adding this to and noting that $\P (G_c \cup H_k) \to 0$ proves the lemma. $\Cox$ \[th:e\_k\] For fixed $r$ as $k \to \infty$, $$e_{k+r} = 2 (- 1)^{k + r} \, \Re \left \{ (1 + o(1)) \sigma_k^{-k-r} \, f(\sigma_k) \, \sqrt{\frac{1}{2 \pi k}} \right \}$$ in probability as $k \to \infty$. [Proof:]{} By Cauchy’s integral theorem, $$e_{k+r} = \frac{(- 1)^{k + r}}{2 \pi i} \int_{\Gamma} f(z) z^{-k-r-1} \, dz \, .$$ By Theorem \[th:good arcs\] and the fact that the contributions from $\Gamma_1$ and $\Gamma_1'$ are conjugate, their sum is twice the real part of a quantity asymptotic to $$\label{eq:leading} \frac{1}{\sqrt{2 \pi k}} f(\sigma_k) \sigma_k^{-k-r} \, .$$ By Lemma \[lem:bad arcs\], the contributions from the remaining four arcs are negligible compared to . The theorem follows. $\Cox$ [Proof of Theorem]{} \[th:a\]: By the definition of $a_{k,r}$, using Theorem \[th:e\_k\] to evaluate $e_k$, $$\begin{aligned} a_{k, r} & = & (- 1)^{k + r} \, e_{k + r} \, \frac{(k+r)!}{r!} \\[1ex] & = & 2 \, k! \, \frac{(k+r)!}{k!} \, \frac{1}{r!} \, \Re \left \{ (1 + o(1)) \sigma_k^{-k-r} \, f(\sigma_k) \, \sqrt{\frac{1}{2 \pi k}} \right \} \, .\end{aligned}$$ For fixed $r$ as $k \to \infty$ asymptotically $(k+r)! / k! \sim k^r$. Setting $\disp A_k = k! \sqrt{\frac{2}{\pi k}} \left\| \sigma_k^{-k} f(\sigma_k) \right\|$ and $\disp \theta_k = \arg \{ \sigma_k^{- k} f(\sigma_k) \}$ simplifies this to $$A_{k} \frac{k^r}{\|\sigma_k\|^r} \left [ \cos \left ( \theta_k - r \arg (\sigma_k) \right ) \right ] \, .$$ Because in probabiltiy $\arg (\sigma_k) = \pi / 2 + o(1)$ while $\|\sigma_k\| \sim k / \pi$, this simplifies finally to $$a_{k,r} = A_k \left [ \cos \left ( \theta_k - \frac{r \pi}{2} \right ) + o(1) \right ] \cdot \frac{\pi^r}{r!} \mbox{ in probability},$$ proving the first part of the theorem. Next, from the proof of Theorem \[th:good arcs\] it is clear that $$\left\| \frac{ \int_{\Gamma_1} \frac{f(z)}{z^{k + r + 1}} dz} {\frac{f(\sigma_k)}{\sigma_k^{k + r}}} \right\| \leq \int_{\Gamma_1} \| \exp(g_k(t) - g_k(0)) \| dt$$ is bounded above in probability, and this bound is independent of $r$. Also the convergence in the proof of Lemma \[lem:bad arcs\] is independent of $r$. Therefore, $$\left\| \frac{a_{k,r}}{A_k} \right\| = O \left( \frac{(k+r)!}{k!} \frac{1}{r! \|\sigma_k\|^r} \right).$$ Since, for any $M > 0$, $$\sum_{r=1}^\infty \frac{(k+r)!}{k!} \frac{\pi^r}{r!} \frac{M^r}{k^r} < \infty, \, \forall \, k > M\pi,$$ with the convergence being uniform over $k \in [T, \infty)$, with $T > M\pi$, we have our result. $\Cox$ [^1]: Department of Mathematics, University of Pennsylvania, 209 South 33rd Street, Philadelphia, PA 19104, USA, [[email protected]]{} [^2]: Research supported in part by NSF grant \# DMS-1209117 [^3]: Department of Mathematics, University of California at Irvine, 340 Rowland Hall (Bldg. \# 400), University of California, Irvine, CA 92697-3875, [[email protected]]{} [^4]: This step is analogous to [@farmer-rhoades Theorem 2.4.1], the correctness of which is unknown to us at this time.	{ "pile_set_name": "ArXiv" }
--- abstract: 'In this paper, we present a continuum model for the dynamics of low angle grain boundaries in two dimensions based on the motion of constituent dislocations of the grain boundaries. The continuum model consists of an equation for the motion of grain boundaries (i.e., motion of the constituent dislocations in the grain boundary normal direction) and equations for the dislocation structure evolution on the grain boundaries. This model is derived from the discrete dislocation dynamics model. The long-range elastic interaction between dislocations is included in the continuum model, which ensures that the dislocation structure on a grain boundary is consistent with the Frank’s formula. These evolutions of the grain boundary and its dislocation structure are able to describe both normal motion and tangential translation of the grain boundary and grain rotation due to both coupling and sliding. Since the continuum model is based upon dislocation structure, it naturally accounts for the grain boundary shape change during the motion and rotation of the grain boundary by motion and reaction of the constituent dislocations. Using the derived continuum grain boundary dynamics model, simulations are performed for the dynamics of circular and non-circular two dimensional grain boundaries, and the results are validated by discrete dislocation dynamics simulations.' address: 'Department of Mathematics, The Hong Kong University of Science and Technology, Clear Water Bay, Kowloon, Hong Kong' author: - Luchan Zhang - Yang Xiang bibliography: - 'science\_v4\_rev.bib' title: Motion of grain boundaries incorporating dislocation structure --- Grain boundary dynamics; dislocation dynamics; long-range elastic interaction; grain rotation; coupling and sliding Introduction ============ Grain boundaries are the interfaces of grains with different orientations and they play essential roles in the polycrystalline materials [@Sutton1995]. Grain boundaries migrate under various driving forces such as the capillarity force, the bulk energy difference, the concentration gradients across the boundary, and the applied stress field. The motion of grain boundaries crucially determines the mechanical and plastic behaviors of the materials. The classical grain boundary dynamics models are based upon the motion by mean curvature to reduce the total interfacial energy [@Herring1951; @Mullins1956; @Sutton1995] using the misorientation-dependent grain boundary energy [@ReadShockley1950]. There are extensive studies in the literature on such motion of grain boundaries by using molecular dynamics or continuum simulations, e.g. [@Chenlq1994; @Upmanyu1998; @Kobayashi2000; @Kazaryan2000; @Upmanyu2002; @Zhang2005; @Upmanyu2006; @Kirch2006; @Selim2009; @Srolovitz2010; @Selim2016]. It has been shown that the grain boundary normal motion can induce a coupled tangential motion which is proportional to the normal motion, as a result of the geometric constraint that the lattice planes must be continuous across the grain boundary [@Li1953223; @Cahn20021; @Cahn20044887]. Besides the tangential motion coupled with normal motion, there is another type of tangential motion that is the relative rigid-body translation of the grains along the boundary by sliding to reduce the grain boundary energy [@Li1962; @Shewmon1966; @Harris19982623; @Kobayashi2000; @Upmanyu2006; @Selim2016]. This sliding motion is independent of all other grain boundary motions. When a grain is embedded in another one, the tangential motions along a grain boundary give rise to a relative rotation between the two grains, leading to change of the misorientation of the grain boundary. In the grain rotation due to sliding, the misorientation angle goes to the nearby local energy minimum state (decreases for a low angle grain boundary), whereas in the grain rotation due to coupling, the misorientation angle increases. The coupling and sliding motions depend on the grain boundary structure and mechanisms of the dynamics. @Rath2007 showed by a simple dislocation model and experimental observations that grain boundary motion does not have to couple with tangential motion. In fact, the coupling and sliding effects cancel out in their case. Cahn and Taylor [@Cahn20044887; @Taylor2007493] proposed a unified approach to the phenomena of the coupling and sliding associated with the grain boundary motion. They formulated the total tangential velocity $v_{\parallel}$ as the superposition of the coupling and sliding effects: $v_{\parallel}=\beta v_{\perp}+v_s$, where the tangential velocity induced by coupling effect is proportional the normal velocity $v_{\perp}$ with the coupling parameter $\beta$, and $v_s$ is the tangential velocity produced by sliding effect. They discussed different cases for the rotation of a circular cylindrical grain embedded in another one [@Cahn20044887]. When the grain does not have such symmetry, they proposed a generalized theory based on mass transfer by diffusion confined on the grain boundary [@Taylor2007493]. Molecular dynamics simulations have been performed to validate the theory of Cahn and Taylor on the coupling grain boundary motion to shear deformation for planar grain boundary [@Cahn20064953; @Cahn20063965], and grain boundary migration and grain rotation for closed circle cylindrical grain boundaries [@Cahn20021; @Trautt20122407]. Experimental observations have also been reported on the migration of low angle planar tilt boundaries coupled to shear deformation in Al bicrystal with stress [@Molodov20071843; @Molodov20095396]. The ratios of the normal to the lateral motion that they measured are complied with the coupling factors in the theory and atomistic simulations by Cahn et al [@Cahn20044887; @Cahn20064953; @Cahn20063965]. Phase field crystal model (an atomistic-level model) was employed to simulate the dynamics of a two-dimensional circular grain, and grain rotation and translation by motion and reaction of the constituent dislocations were observed [@Wu2012407]. Phase field crystal simulations also showed that the coupling of grain boundary motion in polycrystalline systems can give rise to a rigid body translation of the lattice as a grain shrinks and that this process is mediated by dislocation climb and dislocation reactions [@Voorhees2016264]. Three-dimensional phase field crystal simulations were further performed to investigate the motion, rotation and dislocation reactions on a spherical grain in a BCC bicrystal [@Voorhees2017]. Numerical simulations based upon the generalization of the Cahn-Taylor theory to noncircular grains [@Taylor2007493] were performed using the level set method [@Gupta2014]. In this paper, we present a continuum model for the dynamics of low angle grain boundaries in two dimensions based on the motion of constituent dislocations of the grain boundaries. The continuum model consists of an equation for the motion of grain boundaries (i.e., motion of the constituent dislocations in the grain boundary normal direction) and equations for the dislocation structure evolution on the grain boundaries (Eqs. and , or Eqs. and in Sec. \[sec:model\]). The long-range elastic interaction between dislocations is included in the continuum model, which ensures that the dislocation structures on the grain boundaries are consistent with the Frank’s formula for grain boundaries (the condition of cancellation of the far-field elastic fields). These evolutions of the grain boundary and its dislocation structure are able to describe both normal motion and tangential translation of grain boundaries and grain rotation due to both coupling and sliding effects. Since the continuum model is based upon dislocation structure, it naturally accounts for the grain boundary shape change during the motion and rotation of the grain boundary by motion and reaction of the constituent dislocations without explicit mass transfer. Unlike the Cahn-Taylor theory [@Cahn20044887] in which the coupling effect is an assumption, our model is based on the motion of the grain boundary dislocations and the coupling effect is a result. Our model also generalizes the Cahn-Taylor theory by incorporating detailed formulas of the driving forces for the normal and tangential grain boundary velocities that depend on the constituent dislocations, their Burgers vectors, and the grain boundary shape, as well as the shape change of the grain boundaries. Our model is different from their earlier generalization based on the assumption of the coupling effect and mass transfer via surface diffusion [@Taylor2007493]. Note that in some existing continuum models for the motion of grain boundaries and grain rotation [@Li1962; @Shewmon1966; @Kobayashi2000; @Upmanyu2006; @Selim2016], evolution of misorientation angle was included to reduce the grain boundary energy density. These models able to capture the grain boundary sliding but not coupling. In our continuum model, we use dislocation densities on the grain boundary as variables instead of the misorientation angle, which enables the incorporation both the grain boundary coupling and sliding motions. Using the derived continuum grain boundary dynamics model, simulations are performed for the dynamics of circular and non-circular two dimensional grain boundaries. We also perform discrete dislocation dynamics simulations for the dynamics of these grain boundaries and the simulation results using the two models agree excellently with each other. In particular, both our continuum and discrete dislocation dynamics simulations show that without dislocation reaction, a non-circular grain boundary shrinks in a shape-preserving way due to the coupling effect, which is consistent with the prediction of the continuum model in @Taylor2007493 based on the assumption of the coupling effect and mass transfer via surface diffusion. Our continuum grain boundary dynamics model is based upon the continuum framework for grain boundaries in @Zhu2014175 derived rigorously from the discrete dislocation dynamics model. Previously, a continuum model for the energy and dislocation structures on static grain boundaries has been developed [@Zhang2017] using this framework. In fact, the energetic and dynamic properties of grain boundaries were understood based on the underlying dislocation mechanisms in many of the available theories, simulations and experiments [@ReadShockley1950; @Li1953223; @Cahn20021; @Cahn20044887; @Cahn20064953; @Cahn20063965; @Molodov20071843; @Molodov20095396; @Trautt20122407; @Wu2012407; @Voorhees2016264; @Voorhees2017]. Although some direct discrete dislocation dynamics simulations are able to provide detailed information of grain boundary or interface structures and dynamics [@Lim2009; @Quek2011; @Lim2012], continuum models of the dynamics of grain boundaries incorporating their dislocation structures are still desired for larger scale simulations without tracking individual dislocations. The rest of the paper is organized as follows. In Sec. \[sec:settings\], we describe the two dimensional settings of the grain boundaries with their dislocation structures, and review the continuum framework in @Zhu2014175 based on which our continuum model will be developed. In Sec. \[sec:model\], we present our continuum model for the motion of grain boundaries. The variational derivation method of the continuum model is presented in \[sec:variation\]. In Sec. \[sec:Frank\], we discuss the calculation of the misorientation angle based on the Frank’s formula, which is maintained by the long-range dislocation interaction of the grain boundary during its motion. In Sec. \[sec:compare\], we present the formula for the change of misorientation angle (grain rotation) derived from our continuum dynamics model. We also derive the formulas for the tangential motions of grain boundaries due to the coupling and sliding effects based on our continuum model, and make comparisons with the formulas in the Cahn-Taylor theory [@Cahn20044887; @Trautt20122407]. Simulation results for the dynamics of circular and non-circular two dimensional grain boundaries using our continuum model and comparisons with discrete dislocation dynamics simulation results are presented in Sec. \[sec:simulation\]. Conclusions and discussion are made in Sec. \[sec:conclusion\]. Grain boundaries in two dimensions and review of the continuum framework in @Zhu2014175 {#sec:settings} ======================================================================================= We consider the two dimensional problem that one cylindrical grain is embedded in another grain with arbitrary cross-section shape. The inner grain has a misorientation angle $\theta$ relative to the outer grain, and the rotation axis is parallel to the cylindrical axis. The grain boundary is then a pure tilt boundary. ![The cross-section of a cylindrical grain boundary with geometric center $O$. A point $P$ on the grain boundary can be written in polar coordinates as $(R(\alpha),\alpha)$, where $R(\alpha)$ is the radius and $\alpha$ is the polar angle, and its Cartesian coordinate $\mathbf R(\alpha)= (x(\alpha),y(\alpha))=(R(\alpha)\cos\alpha,R(\alpha)\sin\alpha)$. []{data-label="fig:geometry"}](demo_GB_v1.eps){width=".6\linewidth"} We assume that the cross-section curve of the grain boundary $\Gamma$ is in the $xy$ plane and the rotation axis is in $z$ direction, and the geometric center (mass center) of the cross-section of the inner grain is the origin $O$ of the $xy$ plane. We parameterize the two dimensional grain boundary $\Gamma$ by the polar angle $\alpha$, see Fig. \[fig:geometry\]. All the functions defined on $\Gamma$, such as the radius $R$ and the arclength parameter $s$ of the grain boundary, are functions of the parameter $\alpha$. The derivative of a function $g$ defined on the grain boundary with respect to $\alpha$ is denoted by $g'$, i.e., $g'=dg/d\alpha$. See \[sec:geometryformulas\] for the formulas of the grain boundary tangent direction $\mathbf T$, the grain boundary normal direction $\mathbf n$, the curvature of the grain boundary $\kappa$, and $ds/d\alpha$. Assume that on the grain boundary, there are $J$ dislocation arrays corresponding to $J$ different Burgers vectors $\mathbf{b}^{(j)}=(b_1^{(j)},b_2^{(j)},b_3^{(j)})$ with length $b^{(j)}=\\|\mathbf b^{(j)}\\|$ , $j=1,2,\cdots,J$, respectively. All the dislocations are parallel to the $z$ axis, i.e., they are points in the $xy$ plane. Our continuum model is based on the continuum framework proposed in @Zhu2014175. The dislocation densities on the grain boundaries are described by the dislocation density potential functions. A dislocation density potential function $\eta$ is a scalar function defined on the grain boundary such that the constituent dislocations of the same Burgers vector are given by the integer-valued contour lines of $\eta$: [$\eta=i$, where $i$ is an integer]{}. The dislocation structure can be described in terms of the surface gradient of $\eta$ on the grain boundary, which is $ \nabla_s \eta=\frac{d\eta}{ds}\mathbf T =\frac{{\eta}'(\alpha)}{\sqrt{{{R(\alpha)}^2+R'(\alpha)}^2}}\mathbf T$. In particular, the local dislocation line direction is $\mathbf t=\frac{\nabla_s \eta}{\\|\nabla_s \eta\\|} \times \mathbf n$, and the inter-dislocation distance along the grain boundary is $D=1/ \\|\nabla_s \eta\\|$. Accordingly, the dislocation density per unit length on the grain boundary is $$\label{eqn:densityperlength} \rho(\alpha)=\nabla_s \eta \cdot\mathbf T=\frac{d\eta}{ds} =\frac{{\eta}'(\alpha)}{\sqrt{{{R(\alpha)}^2+R'(\alpha)}^2}}.$$ Here we allow the dislocation density $\rho$ to be negative to also include dislocations with opposite line directions. In our continuum model, it is more convenient to use the dislocation density per unit polar angle, which is $$\label{eqn:density_angle} \varrho(\alpha)=\rho(\alpha)\frac{ds}{d\alpha}=\rho(\alpha)\sqrt{R(\alpha)^2+R'(\alpha)^2}=\eta'(\alpha).$$ Here we have used Eqs. and . The surface gradient $\nabla_s \eta$ can be expressed in terms of $\varrho$ as $$\label{eqn:eta_by_rho} \nabla_s\eta=\rho(\alpha) \mathbf T=\frac{\varrho(\alpha)}{\sqrt{R(\alpha)^2+R'(\alpha)^2}}\mathbf T.$$ The dislocation line direction defined by $\mathbf t=\frac{\nabla_s \eta}{\\|\nabla_s \eta\\|} \times \mathbf n$ becomes $$\label{eqn:ddirection_rho} \mathbf t=\frac{\varrho}{\|\varrho\|} \mathbf T\times \mathbf n.$$ When there are $J$ dislocation arrays on the grain boundary, these constituent dislocations are represented by $\eta^{(j)},j=1,2,\cdots,J$, corresponding to $J$ different Burgers vectors $\mathbf{b}^{(j)}$, $j=1,2,\cdots,J$, respectively. In the continuum framework presented in @Zhu2014175, the total energy associated with grain boundaries can be written as $$\label{eqn:totalenergy} E=E_{\rm long}+E_{\rm local}+E_{\rm other},$$ where $E_{\rm long}$ is the energy due to the long-range elastic interaction between the constituent dislocations of the grain boundaries, $E_{\rm local}$ is the line energy of the constituent dislocations corresponding the commonly used grain boundary energy in the literature [@Sutton1995; @ReadShockley1950] and is a generalization of the classical Read-Shockley energy formula [@ReadShockley1950], and $E_{\rm other}$ includes the energy due to the interactions between the dislocation arrays and other stress fields such as the applied stress. The variations of this total energy with respect to the change of the grain boundary and the change of dislocation structure on the grain boundary are $$\begin{aligned} \frac{\delta E}{\delta r}&=&- \sum_{j=1}^J\\|\nabla_s\eta^{(j)}\\| \ \mathbf f_{\rm total}^{(j)}\cdot \mathbf n \label{eqn:var_n}, \ \ {\rm when} \ \delta\mathbf r = \mathbf n \delta r, \vspace{1ex} \\ \frac{\delta E}{\delta \eta^{(j)}}&=&\mathbf f_{\rm total}^{(j)}\cdot\frac{\nabla_s\eta^{(j)}}{\\|\nabla_s\eta^{(j)}\\|}, \ \ j=1,2,\cdots,J,\label{eqn:var_eta}\end{aligned}$$ respectively, where $$\mathbf f_{\rm total}^{(j)}={\mathbf f}_{\rm long}^{(j)}+{\mathbf f}_{\rm local}^{(j)}+{\mathbf f}_{\rm other}^{(j)}.\label{eqn:fs}$$ Here $\mathbf f_{\rm total}^{(j)}$ is the total force on the $j$-th dislocation array, ${\mathbf f}_{\rm long}^{(j)}$ is the force that comes from the long-range interaction energy $E_{\rm long}$, ${\mathbf f}_{\rm local}^{(j)}$ is the force that comes from the local grain boundary energy $E_{\rm local}$, and ${\mathbf f}_{\rm other}^{(j)}$ is the force due to other stress fields (such as the applied stress field), for $j=1,2,\cdots,J$. See Sec. \[sec:model\] for more discussion on these continuum forces. Following the discrete dislocation dynamics model [@Xiang2003; @Xiang2010; @Zhu2014175], the continuum dynamics of these dislocation arrays along the grain boundary is given by $$\begin{aligned} \frac{\partial\eta^{(j)}}{\partial t}+\mathbf v^{(j)}\cdot \nabla_s\eta^{(j)}=0,\ \ \ \ j=1,2,\cdots,J,\label{eqn:evolution_eta}\end{aligned}$$ where the dislocation velocity $\mathbf v^{(j)}=\mathbf M_{\rm PK}^{(j)} \cdot \mathbf f^{(j)}_{\rm total}$, and $\mathbf M_{\rm PK}^{(j)}$ is the mobility associated with the total Peach-Koehler force $\mathbf f^{(j)}_{\rm total}$ of the $j$-th dislocation array. If in the dynamics process, generation and removal of dislocations are critical, e.g. [@Li1953223; @Shewmon1966; @Cahn20021; @Upmanyu2006; @Wu2012407; @Trautt20122407], using Eq. , the dislocation evolution equations are $$\begin{aligned} \frac{\partial \eta^{(j)}}{\partial t}=-m_j\frac{\delta E}{\delta \eta^{(j)}} =-m_j\left(\mathbf f^{(j)}\cdot\frac{\nabla_s\eta^{(j)}}{\\|\nabla_s\eta^{(j)}\\|}\right), \ \ \ \ j=1,2,\cdots,J, \label{eqn:evolution3}\end{aligned}$$ where $\mathbf f^{(j)}$ is the total force on the $j$-th dislocation array in Eq. or some of its contributions on the right-hand side, $m_j$ is some positive constant. The choice of Eqs. and/or depends on the physics of the dynamics process. The motion of the grain boundary in general is, following Eq. , $$\begin{aligned} v_n=-m_n\frac{\delta E}{\delta r} =m_n \sum_{j=1}^J\\|\nabla_s\eta^{(j)}\\| \ \mathbf f_{\rm total}^{(j)}\cdot \mathbf n, \label{eqn:evolution_S}\end{aligned}$$ where $m_n$ is the mobility. This continuum framework is general and applies to any dislocation arrays. Based on this framework, we will develop a two dimensional continuum model for the dynamics of grain boundaries including the motion and tangential translation of the grain boundaries and grain rotation that is consistent with the discrete dislocation dynamics model and atomistic simulations, see the next section. In the two dimensional problems, this new dislocation representation method based on dislocation density potential functions $\eta$’s and the classical method using the scalar dislocation densities are equivalent, see Eqs. and . Continuum dynamics model for grain boundaries in three dimensions is being developed and will be presented elsewhere, in which it is more convenient to use dislocation density potential functions for the structure of the constituent dislocations. The continuum model for grain boundary motion {#sec:model} ============================================= In this section, we present our continuum model for grain boundary motion based on densities of the constituent dislocations. The grain boundary motion is coupled with the evolution of the constituent dislocations on the grain boundary, under both the long-range and local interactions of dislocations. The misorientation angle changes due to the coupling and sliding motions and the shape change of the grain boundary are naturally accommodated by the motion and reaction of the constituent dislocations. A critical idea to incorporate the coupling motion in the continuum model is to include the driving forces due to energy variations under conservation of the constituent dislocations [@Zhu2014175], instead of the fixed grain boundary energy density as did in the literature. An example of such variational derivation is presented in \[sec:variation\] for the two dimensional problem in terms of the contribution of the (local) grain boundary energy. The continuum model also incorporates the driving forces associated with energy variations with respect to the changes of dislocation densities due to dislocation reactions or the applied stress on fixed grain boundary, leading to sliding motion of the grain boundary. We consider the dynamics of a closed curved grain boundary $\Gamma$ with radius function $R(\alpha)$ and $J$ dislocation arrays on $\Gamma$ with Burgers vector $\mathbf{b}^{(j)}$, $j=1,2,\cdots,J$, as described in the previous section. The motion of the grain boundary and the evolution of the dislocation structure on it are described by v\_n=&M\_[d]{}\_[j=1]{}\^J([f]{}\_[long]{}\^[(j)]{}+[f]{}\_[local]{}\^[(j)]{}+[f]{}\_[app]{}\^[(j)]{})n +M\_[b]{}p,\[eqn:vn\]\ =&-( M\_[d]{} [f]{}\_[long]{}\^[(j)]{}T)’ -M\_[t]{} -M\_[a]{}( [f]{}\_[app]{}\^[(j)]{}T)’, \[eqn:vp\]\ &j=1,2,,J. Note that with the grain boundary normal velocity in Eq. , Eq. holds when every point on the grain boundary moves in the normal direction of the grain boundary. In the simulations of a finite grain embedded in another one, instead of using the grain boundary normal velocity $v_n$ in Eq. directly, it is more convenient to move each point on the grain boundary in the radial direction: =&-v\_n/,\[eqn:vr\]\ =&-’-M\_[t]{} -M\_[a]{}( [f]{}\_[app]{}\^[(j)]{}T)’,\[eqn:vpr\]\ & j=1,2,,J, where $\lambda$ is the angle between the normal direction of the grain boundary and its inward radial direction (see Fig. \[fig:geometry\]), and $\cos\lambda=R/\sqrt{R^2+R'^2}$, $\sin\lambda=R'/\sqrt{R^2+R'^2}$. With Eq. , Eq. holds when every point on the grain boundary moves in the radial direction of the grain boundary, i.e., the value of the parameter $\alpha$ for each material point on the grain boundary does not change. Eq. gives the normal velocity of the grain boundary $v_n$, which represents the motion of the constituent dislocations in the normal direction of the grain boundary. In Eq. , ${\mathbf f}_{\rm long}^{(j)}$, ${\mathbf f}_{\rm local}^{(j)}$ and ${\mathbf f}_{\rm app}^{(j)}$ are the continuum long-range dislocation interaction force, the continuum local force, and the force due to applied stress field, respectively, on a dislocation with Burgers vector $\mathbf{b}^{(j)}$, and $M_{\rm d}$ is the mobility of the dislocations. The velocity of a point on the grain boundary is the weighted average of the velocities of dislocations on the grain boundary with different Burgers vectors. This modification has been made from the variational normal velocity formula in Eq. so that the velocity of the grain boundary is consistent with the dynamics of its constituent dislocations [@Cahn20044887]. Recall that the available atomistic simulations [@Cahn20021] and phase field crystal simulations [@Wu2012407; @Voorhees2016264; @Voorhees2017] were performed at high temperatures to examine the coupling motion of curved grain boundaries, which is purely geometric [@Cahn20044887]. At these temperatures, the dislocation climb mobility is comparable with that of dislocation glide. Moreover, the atomistic simulations of @Trautt20122407 showed that the grain boundary dislocations do not necessarily move by dislocation climb, which is in general required (simply based on the dislocation motion direction) for the motion of curved low angle grain boundaries and is very small at a not very high temperature; instead, the dislocations move by a chain of dissociation/association processes. Such dislocation reaction mechanism leads to an effectively large mobility of the grain boundary dislocations in their climb direction. Following these simulation results, we assume that the dislocation mobility $M_{\rm d}$ is isotropic in our continuum model to fully examine the purely geometric coupling motion of low angle grain boundaries. Influences of the dislocation glide, climb and cross-slip mobilities and dislocation dissociation/association reactions on the grain boundary motion can be further incorporated in the continuum model by using different mobilities for these dislocation motions. These will be explored in the future work. Another driving force for the grain boundary motion is the difference between the bulk energy densities of the two grains denoted by $p$ [@Cahn20044887; @Trautt20122407], and $M_{\rm b}$ is the grain boundary mobility associated with this driving force. Eq. describes the evolution of the dislocation structure on the grain boundary. (Recall that $g'=dg/d\alpha$ for a function $g(\alpha)$.) These equations are based on the dislocation motion along the grain boundary, driven by the continuum long-range interaction force, dislocation reaction, and the applied stress field. The first term in Eq. is the motion of dislocations along the grain boundary following conservation law in Eq. driven by the continuum long-range elastic force. The importance of this term is to maintain the dislocation structure and the condition of cancellation of the far-field stress fields for the grain boundary (or the Frank’s formula [@Frank; @Bilby]) in a way that is consistent with the discrete dislocation dynamics. The second term in Eq. is the driving force due to variation of the grain boundary energy density $\gamma$ (when the long-range elastic interaction vanishes) with respect to the change of dislocation density $\varrho$ on the fixed grain boundary. The third term in Eq. is due to the effect of the applied stress. Since the equilibrium dislocation structure on a grain boundary is stable except for the change of the misorientation angle (i.e. sliding along a fixed grain boundary) [@Xiang-Yan2017], the major influences of the local energy and the applied stress on a fixed grain boundary are to change the dislocation structure by dislocation reactions, leading to grain boundary sliding [@Li1953223; @Shewmon1966; @Cahn20021; @Upmanyu2006; @Wu2012407; @Trautt20122407], see Sec. \[sec:rotation\] for more discussion. In these processes, the dislocations on the grain boundary are not conserved and the last two terms in Eq. account for such changes of dislocation structure due to these two driving forces (see also Eq. ), in which $M_{\rm t}$ and $M_{\rm a}$ are the mobilities associated with the driving forces of the local energy and the applied stress, respectively. Note that these grain boundary dislocation reactions are not energetically favorable and energy barriers have to be overcome during the reaction processes [@Cahn20021; @Trautt20122407]. The rates of these dislocation reactions are reflected in the mobilities $M_{\rm t}$ and $M_{\rm a}$. In these evolution equations, ${\mathbf f}_{\rm long}^{(j)}$ is the force that comes from the long-range interaction energy $E_{\rm long}$ in Eq. . This continuum long-range interaction force on a dislocation of the $j$-th dislocation arrays located at the point $(x,y)$ is $$\begin{aligned} \label{eqn:flong0} {\mathbf f}_{\rm long}^{(j)}(x,y)=\sum_{k=1}^J \int_\Gamma\mathbf{\tilde{f}}^{(j,k)}(x,y;x_1,y_1)ds,\end{aligned}$$ where $ds$ is the line element of the integral along the grain boundary $\Gamma$, the point $(x_1,y_1)$ varies along $\Gamma$ in the integral, and $\int_\Gamma\mathbf{\tilde{f}}^{(j,k)}(x,y;x_1,y_1)ds$ is the force acting on a dislocation of the $j$-th dislocation array at the point $(x,y)$ generated by the $k$-th dislocation array, with \^[(j,k)]{}&=(\_1\^[(j,k)]{}, \_2\^[(j,k)]{}), \_1\^[(j,k)]{}(x,y;x\_1,y\_1) &=[ ]{}\ & { \[(x-x\_1)\^3-(x-x\_1)(y-y\_1)\^2\]b\_1\^[(k)]{}b\_1\^[(j)]{}.\ &+\[(x-x\_1)\^2(y-y\_1) -(y-y\_1)\^3\]b\_2\^[(k)]{}b\_1\^[(j)]{}\ &+ \[(x-x\_1)\^2(y-y\_1)- (y-y\_1)\^3\]b\_1\^[(k)]{}b\_2\^[(j)]{}\ &.+ \[(x-x\_1)\^3+3(x-x\_1)(y-y\_1)\^2\]b\_2\^[(k)]{} b\_2\^[(j)]{} }, \_2\^[(j,k)]{}(x,y;x\_1,y\_1) &=[ ]{}\ &{ \[3(x-x\_1)\^2(y-y\_1)+(y-y\_1)\^3\]b\_1\^[(k)]{} b\_1\^[(j)]{}.\ &+ \[-(x-x\_1)\^3+ (x-x\_1)(y-y\_1)\^2\]b\_2\^[(k)]{}b\_1\^[(j)]{}\ & + \[-(x-x\_1)\^3 +(x-x\_1)(y-y\_1)\^2\]b\_1\^[(k)]{}b\_2\^[(j)]{}\ &.+ \[-(x-x\_1)\^2(y-y\_1) + (y-y\_1)\^3\]b\_2\^[(k)]{}b\_2\^[(j)]{}}.\[eqn:flong1\] Here $\mu$ is the shear modulus and $\nu$ is the Poisson ratio. The integral in Eq. is over all the grain boundaries if there are multiple grain boundaries in the system. The long-range interaction force ${\mathbf f}_{\rm long}$ and the long-range interaction energy $E_{\rm long}$ vanish for an equilibrium grain boundary [@Sutton1995; @HL; @Zhu2014175]. The continuum local force on a dislocation of the $j$-th dislocation arrays in these evolution equations is $$\begin{aligned} \label{eqn:local_force} {\mathbf f}_{\text{local}}^{(j)}\cdot\mathbf{n}=\frac{\mu (b^{(j)})^2}{4\pi(1-\nu)}\kappa,\end{aligned}$$ where $\kappa$ is the curvature of the grain boundary. This local force comes from the grain boundary $E_{\rm local}$ (the line energy of the constituent dislocations) in Eq. . The energy density of $E_{\rm local}$ is [@Zhu2014175; @Zhang2017] =[\_[j=1]{}\^J ]{},\[eqn:gb\_density\] where $r_g$ is a dislocation core parameter and $\epsilon$ is a numerical cutoff parameter. This is a generalization of the classical Read-Shockley energy formula [@ReadShockley1950]. These force formulas are the two-dimensional forms of the continuum forces on low angle grain boundaries derived by @Zhu2014175 based on the discrete dislocation model [@HL]. It has been shown that the summation of the continuum long-range and local forces provides a good approximation to the Peach-Koehler force on a dislocation in the discrete dislocation dynamics model. The Peach-Koehler force due to the applied stress $\pmb\sigma_{\rm app}$ is ${\mathbf f}_{\rm app}^{(j)}=(\pmb\sigma_{\rm app}\cdot \mathbf{b}^{(j)})\times{\mathbf t}^{(j)}$, where ${\mathbf t}^{(j)}$ is the line direction of a dislocation in the $j$-th dislocation array on the grain boundary. Using Eq. , the contribution of the applied stress the grain boundary velocity given in Eq. can be written as $$\begin{aligned} \label{eqn:fapp_dot_n} {\mathbf f}_{\rm app}^{(j)}\cdot \mathbf n = -\frac{\varrho^{(j)}}{\|\varrho^{(j)}\|} (\mathbf{b}^{(j)}\cdot \pmb\sigma_{\rm app}\cdot \mathbf T),\end{aligned}$$ and the contribution of the applied stress to the dislocation structure evolution equation can be written as \[eqn:contributionofapplied\] ( [f]{}\_[app]{}\^[(j)]{}T)’ =(\^[(j)]{}\_[app]{}n)’=- (\^[(j)]{}\_[app]{}T), where the last expression holds for a constant applied stress $\pmb\sigma_{\rm app}$, otherwise the expression in the middle should be used. [Remark:]{} Essentially, it is equivalent to use the dislocation density per unit length $\rho^{(j)}$, $j=1,2,\cdots,J$, as the variables in the continuum grain boundary dynamics model in Eqs. and . However, such evolution equations of $\rho^{(j)}$’s will no longer be as simple as Eq. because the arclength of the grain boundary also evolves as the grain boundary migrates. Misorientation angle and effect of long-range dislocation interaction {#sec:Frank} ===================================================================== In our continuum model, the misorientation angle $\theta$ between the two grains can be calculated based on the Frank’s formula, which is maintained during the motion of the grain boundary by the long-range elastic interaction between the constituent dislocations of the grain boundary. Frank’s formula and misorientation angle ---------------------------------------- With the continuum model in Eqs. and (or Eqs. and ) for the motion of the grain boundary and evolution of the dislocation structure on it, the misorientation angle $\theta$ between the two grains at any point on the grain boundary can be calculated based on the Frank’s formula [@Frank; @Bilby], which is a condition for an equilibrium grain boundary dislocation structure and is equivalent to the cancellation of the long-range elastic fields [@Frank; @Bilby; @Zhu2014175]. Using the representation of dislocation density potential functions, the Frank’s formula is [@Zhu2014175] (va)- \_[j=1]{}\^J b\^[(j)]{}(\_s\^[(j)]{}v)=0, where $\mathbf a$ is the rotation axis and $\mathbf v$ is any vector tangent to the grain boundary. In the case of two dimensional tilt boundaries being considered in this paper, the rotation axis $\mathbf a$ is in the $+z$ direction, and the Frank’s formula can be written as \[eqn:frank1\] n+\_[j=1]{}\^J \^[(j)]{}b\^[(j)]{}=0, or \[eqn:frank2\] n+\_[j=1]{}\^J =0. An equivalent form is \[eqn:frank3\] ()n+\_[j=1]{}\^J \^[(j)]{}b\^[(j)]{}=0. Using the Frank’s formula in Eq. or , the misorientation angle $\theta$ at each point on the grain boundary can be calculated by \[theta\_calculation1\] =-\_[j=1]{}\^J \^[(j)]{}(b\^[(j)]{}n) =-\_[j=1]{}\^J . Note that an alternative formula to calculate the misorientation angle is $\theta= \\|\sum_{j=1}^J \rho^{(j)}\mathbf b^{(j)}\\|=\\|\sum_{j=1}^J \varrho^{(j)}\mathbf b^{(j)}\\|/\sqrt{R^2+R'^2}$. Integrating Eq. over the grain boundary, we have \[eqn:theta\_noncircular\] =-\_0\^[2]{}\_[j=1]{}\^J \^[(j)]{}(b\^[(j)]{})d, where $L$ is the circumference of the grain boundary. It is more convenient to use this formula to evaluate $\theta$ during the evolution of the grain boundary in which the shape of the grain boundary and the dislocation densities on it change. Frank’s formula maintained by the long-range dislocation interaction {#sec:Frank-longrange} -------------------------------------------------------------------- It has been shown that the Frank’s formula is equivalent to the cancellation of the long-range elastic fields generated by the constituent dislocations of the grain boundary [@Frank; @Bilby; @Zhu2014175]. In the continuum model in Eqs. and (or Eqs. and ), the Frank’s formula is maintained by the continuum long-range dislocation interaction force. This is because the continuum long-range dislocation force is much stronger than the local force except for very small size grain (comparable with the inter-dislocation distance on the grain boundary) [@Zhu2014175]. Starting from an equilibrium grain boundary with vanishing continuum long-range interaction, approximately during the motion of the grain boundary, the continuum long-range interaction will remain vanishing and the Frank’s formula will always hold (except when the grain size is very small), otherwise the long-range interaction would lead to an increase in the total energy. A proof of the equivalence of the Frank’s formula and cancellation of the long-range elastic fields generated by the constituent dislocations for a curved grain boundary can be found in @Zhu2014175. Here we give an alternative calculation in two dimensions to show their equivalence, see \[sec:equivalenceproof\]. Grain rotation and coupling and sliding motions {#sec:compare} =============================================== In this section, we present formulas for the rate of change of the misorientation angle $d\theta/dt$ (grain rotation) and the tangential velocities of the coupling and sliding motions of the grain boundary. Derivation of these formulas are based on the Frank’s formula, which is maintained by the long-range dislocation interaction in our continuum model in Eqs. and (or Eqs. and ). Grain rotation and coupling and sliding velocities {#sec:rotation} -------------------------------------------------- Using the Frank’s formula in Eq. , it can be calculated that the rate of change of the misorientation angle $\theta$ is \[eqn:dtheta\_dt\] =- +\_[j=1]{}\^J (b\^[(j)]{}) , where $\hat{\mathbf R}=\mathbf R/R$ is the outward radial direction of the grain boundary. The time derivatives $\partial R/\partial t$ and $\partial \varrho^{(j)}/\partial t$ in this grain rotation formula are given by our continuum model in Eqs. and . Derivation of this grain rotation formula from Frank’s formula is given in \[eqn:grainrotationderivation\]. It is interesting to see that the grain rotation formula in Eq. depends on the net Burgers vector in the radial direction, whereas the formulation of the misorientation angle in Eq. depends on the net Burgers vector in the normal direction of the grain boundary. The tangential velocity of the grain boundary is calculated based on the change of misorientation angle and the fixed center, which is $v_{\\|}=R\cos\lambda\frac{d\theta}{dt}=\frac{R^2}{\sqrt{R^2+R'^2}}\frac{d\theta}{dt}$, where $\lambda$ is recalled to be the angle between the normal direction of the grain boundary and its inward radial direction, and $\cos\lambda=R/\sqrt{R^2+R'^2}$, see Fig. \[fig:geometry\]. Using the grain rotation formula in Eq. , the tangential velocity can be written as \[eqn:coupling0\] v\_=v\_n+\_[j=1]{}\^J (b\^[(j)]{}) . Here we have used $v_n=-(\partial R/\partial t)\cos\lambda$. The grain rotation formula in Eq. and the tangential velocity formula in Eq. show that the grain rotation (or the tangential velocity) can be decomposed into the rotation (or the tangential velocity) due to the normal motion of the grain boundary with dislocation conservation (the first term in Eq. or ) and that due to the evolution of the constituent dislocations on the grain boundary (the second term in Eq. or ). The former is recalled to be the coupling motion of the grain boundary and is purely geometric [@Cahn20021; @Cahn20044887]. The latter includes the sliding motion of the grain boundary which is recalled to be the rigid-body translation independent of the normal motion of the grain boundary [@Shewmon1966]. The grain boundary dislocations are not conserved during the sliding motion [@Cahn20044887; @Trautt20122407]. This can be understood as follows. When the misorientation angle $\theta$ changes, say decreases, and the grain boundary profile remains the same, the inter-dislocation distance increases, which means the total number of the grain boundary dislocations decreases. This can also be seen from our formulation that the change of dislocation densities $\partial \varrho^{(j)}/\partial t$ in the second term in the grain rotation formula in Eq. or the tangential velocity formula in Eq. includes the motion of existing dislocations on the grain boundary (the first term on the right-hand side of Eq. ) and the changes of dislocation densities due to dislocation reactions (the second and third terms on the right-hand side of Eq. ). The effect of the motion of existing dislocations on the grain boundary is to maintain the Frank’s formula along with the motion of the grain boundary and reactions of dislocations. The integral form the the grain rotation formulation is: \[eqn:theta\_t\_noncircular\] =--\_0\^[2]{}\_[j=1]{}\^J (b\^[(j)]{})d. Using Eq. , it can be written as \[eqn:theta\_t\_noncircular1\] =&- +\_0\^[2]{}\_[j=1]{}\^J \^[(j)]{} ([f]{}\_[long]{}\^[(j)]{}T)(b\^[(j)]{})d\ &+\_0\^[2]{}\_[j=1]{}\^J (b\^[(j)]{})d+\_0\^[2]{}\_[j=1]{}\^J ( [f]{}\_[app]{}\^[(j)]{}T)’(b\^[(j)]{})d. Here we have used the equation $\sum_{j=1}^J \varrho^{(j)} (\mathbf b^{(j)}\cdot{\mathbf T})=0$ by using Eq. . Comparisons with the models of Cahn-Taylor and Shewmon ------------------------------------------------------ Our grain rotation and tangential velocity formulas in Eqs. and are derived based on the motion and reaction of the grain boundary dislocations and Frank’s formula. They can be considered as justifications in the case of low angle grain boundaries for the unified approach to the grain boundary coupling and sliding motions of @Cahn20044887, in which existence of the coupling effect and superposition of the coupling and sliding motions are assumptions. Our continuum model also generalizes the Cahn-Taylor model by incorporating detailed formulas of the driving forces for the grain boundary coupling and sliding tangential motions that depend on the constituent dislocations, their Burgers vectors, and the grain boundary shape, as well as the shape change of the grain boundaries. For the special case of a circular grain boundary with a single Burgers vector, our tangential velocity and grain rotation formulas reduce to the equations in the Cahn-Taylor model as illustrated below. When the grain boundary is circular, $R(\alpha)$ is constant, $R'(\alpha)=0$, $\kappa=1/R$, and $\mathbf n=-\hat{\mathbf R}$. At a point on the circular grain boundary where all the dislocations have the same Burgers vector $\mathbf b$ that is in the direction $\hat{\mathbf R}$, i.e., $b=\mathbf b\cdot \hat{\mathbf R}$, our grain rotation and tangential velocity formulas in Eqs. and become &=- - \[eqn:circularsinglerot\] +\_[rt]{},\ v\_&=v\_n-M\_[t]{} b+M\_[a]{}b\^2 \_[rt]{},\[eqn:circularsingle\] where $\sigma_{\rm rt}=\hat{\mathbf R}\cdot \pmb\sigma_{\rm app}\cdot \mathbf T$ is the shear component of the applied stress along the grain boundary. For this special case of circular grain boundary with single Burgers vector, the tangential velocity of the grain boundary given previously by the Cahn-Taylor model [@Cahn20044887; @Trautt20122407] is &=- + (-),\[mishin\_eqn13rot\]\ v\_&=v\_n + S(-),\[mishin\_eqn13\] where $\beta=\theta$ is the coupling factor, $\gamma'(\theta)=d\gamma/d\theta$, $S$ is the sliding coefficient, and $\sigma$ is the applied shear stress. Comparing Eqs. and (same for the comparison between Eqs. and ), we can see that our formula gives the same coupling term (the first term in Eq. ) as that in the Cahn-Taylor model. The last two terms in our formula in Eq. that accounts for the grain boundary sliding are also the same as the corresponding terms in the Cahn-Taylor model in Eq. based on the relations $\theta=b/D=\varrho b/R$ and $\partial\gamma/\partial \varrho=\gamma'(\theta)b/R$ if we further assume $\varrho>0$ and $S=M_{\rm t}b^2=M_{\rm a}b^2$. Note that at an arbitrary point on the circular low angle grain boundary, dislocation densities of multiple Burgers vectors are in general nonzero, and our full formulas in Eqs. and should be used instead of the above simple formulas. Moreover, the circular grain boundary may change its shape during its evolution with dislocation reactions, see the simulations in Sec. \[sec:simulation\]. We also remark that in the generalized Cahn-Taylor theory based on the assumptions of coupling effect and mass transfer via surface diffusion along the grain boundary [@Taylor2007493], alternative formulations for grain rotation and grain boundary tangential velocities due to the coupling and sliding as well as the shape-preserving motion for a non-circular grain boundary have been derived. In the model of @Shewmon1966 for grain rotation by sliding, the rate of grain rotation is proportional to $\gamma'(\theta)$. This is included in the unified model by @Cahn20044887 as shown in Eq. . In our grain rotation formula in Eq. , the rate of rotation due to sliding is proportional to the change rates of the grain boundary dislocation densities, which together with the Frank’s formula (maintained by the long-range elastic interaction) lead to the change of misorientation angle. Using densities of the constituent dislocations instead of the misorientation angle enables us to incorporate both the coupling and sliding motions of the grain boundary in the continuum model as discussed previously. In the simple case of a point on a circular grain boundary where all the dislocations have the same Burgers vector, our grain rotation formula includes the sliding-included rotation proportional to $\partial\gamma/\partial \varrho=\gamma'(\theta)b/R$ as shown in Eq. and discussed above. Shape-preserving grain boundary motion under pure coupling {#sec:preserving} ---------------------------------------------------------- Assume that the Frank’s formula holds during the evolution of the grain boundary. Consider the case when the tangential motion of the grain boundary is pure coupling, i.e. its constituent dislocations are conserved. From the simulations of discrete dislocation dynamics and our continuum model shown in Sec. \[sec:simulation\], the constituent dislocations are moving in the inward radial direction, i.e. $\partial \varrho^{(j)}/\partial t=0$, $j=1,2,\cdots,J$. Under these conditions, the grain rotation formula in Eq. gives =-. Integration of this equation gives \[eqn:Rfixshape\] R(,t)=R(,0). That is, the shape of the grain boundary does not change during the evolution. This was solved previously by @Taylor2007493 based on the assumption of the coupling effect and mass transfer via surface diffusion along the grain boundary. In addition to this shape-preserving property, we also have the following identities during the grain boundary motion under conservation of dislocations by using Eqs. and : (t)R(,t)&=(0)R(,0),\[eqn:Rtheta\]\ (t)L(t)&=(0)L(0).\[eqn:Ltheta\] Recall that $L=\int_0^{2\pi}\sqrt{R^2+R'^2}d\alpha$ is the circumference of the grain boundary. Finally, we remark that by Eq. , \[eqn:theta\_noncircular1\] L=-\_0\^[2]{}\_[j=1]{}\^J \^[(j)]{}(b\^[(j)]{})d, which holds generally no matter whether the dislocation densities $\varrho^{(j)}$, $j=1,2,\cdots,J$ change or not. Numerical Simulations {#sec:simulation} ===================== In this section, we perform numerical simulations using our continuum model in Eqs. and to study the motion of grain boundaries and evolution of their dislocation structure in two dimensions. We focus on the grain boundaries in fcc crystals, in which there are $J=6$ Burgers vectors of the $<110>$ type with the same magnitude $b$. The cross-section plane of the grain boundaries is the $(111)$ plane, and the directions $[\bar{1}10], [\bar{1}\bar{1}2]$ and $[111]$ are chosen to be the $x, y, z$ directions, respectively. In this coordinate system, the six Burgers vectors are $\mathbf{b}^{(1)} = \left(1,0,0\right)b$, $\mathbf{b}^{(2)} = \left(\frac{1}{2},\frac{\sqrt{3}}{2},0\right)b$, $\mathbf{b}^{(3)} = \left(\frac{1}{2},-\frac{\sqrt{3}}{2},0\right)b$, $\mathbf{b}^{(4)} = \left(0,\frac{\sqrt{3}}{3},\frac{\sqrt{6}}{3}\right)b$, $\mathbf{b}^{(5)} = \left(\frac{1}{2},\frac{\sqrt{3}}{6},-\frac{\sqrt{6}}{3}\right)b$, and $\mathbf{b}^{(6)} = \left(-\frac{1}{2},\frac{\sqrt{3}}{6},-\frac{\sqrt{6}}{3}\right)b$. The rotation axis is fixed as the $[111]$ direction. The misorientation angle of the initial grain boundary in the simulations is $5^\circ$. We choose the Poisson’s ratio $\nu=0.347$, which is value of aluminum. In our simulations, we neglect the applied stress and the bulk energy difference, that is, $\pmb \sigma_{\rm app}=\mathbf 0$ (accordingly, $\mathbf f^{(j)}_{\rm app}=0$, $j=1,2,\cdots,J$) and $p=0$ in Eqs. and . The grain boundary motion by Eq. is solved using standard forward Euler scheme. The singular integrals in the long-range force $\mathbf f_{\rm long}^{(j)}$ in Eqs. – are calculated using the trapezoidal rule with local approximation at the singular point [@SI1988]. The evolution of the dislocation structure in Eq. is solved by the following splitting scheme in time: &=-M\_[d]{} \_[t=t\_n]{}, \[eqn:vp11\]\ &=.-M\_[t]{}\|\_[t=t\_n]{}, where $\Delta t$ is the time step. Here recall that $\varrho(\alpha)=\eta'(\alpha)$ or $\eta(\alpha)=\int^\alpha_0 \varrho(\alpha_1)d\alpha_1$. The spatial derivative ${\eta^{(j)}}'$ in Eq. is calculated using the upwind scheme, and a cutoff parameter is used to set ${\eta^{(j)}}'=0$ when the calculated value is very small. We start from a grain boundary with the equilibrium dislocation structure that satisfies the Frank’s formula and has the lowest energy, which can be calculated pointwisely using the method in @Zhang2017. We also perform discrete dislocation dynamics simulations to validate our continuum model. The formulation of the interaction between dislocations can be found, e.g., in @HL. In fact, the Peach-Koehler force on a dislocation with Burgers vector $\mathbf b^{(i)}$ and line direction $\mathbf t^{(i)}$ is $(\pmb\sigma\cdot {\mathbf b^{(i)}}) \times \mathbf t^{(i)}$, where $\pmb \sigma$ is the stress. The stress at a point $(x,y)$ generated by a dislocation located at point $(x_1,y_1)$ with Burgers vector $\mathbf b^{(j)}=(b^{(j)}_1,b^{(j)}_2)$ and $+z$ line direction is $\pmb\sigma^{(j)}(x,y)=\mathbf G_1(x,y;x_1,y_1)b^{(j)}_1+\mathbf G_2(x,y;x_1,y_1)b^{(j)}_2$, where $\mathbf G_1$ and $\mathbf G_2$ are given in Eqs. and . Circular grain boundary {#sec:circle} ----------------------- We firstly consider the motion of an initially circular grain boundary with radius $R_0=140b$. The calculated equilibrium dislocation structure [@Zhang2017] on this initial grain boundary is shown in Fig \[circle\_motion\]a, which contains three arrays of dislocations with Burgers vectors $\mathbf{b}^{(1)}$, $\mathbf{b}^{(2)}$, and $\mathbf{b}^{(3)}$, respectively. The distribution of these dislocations are represented by the dislocation density potential functions $\eta^{(1)}$, $\eta^{(2)}$, and ${\eta^{(3)}}$, respectively, see Fig \[circle\_motion\]b. Recall that the dislocation density potential functions are defined on the grain boundary such that the constituent dislocations are located at their integer-valued contour lines. Also recall that the dislocation density per unit polar angle $\varrho^{(j)}$ and the dislocation density per unit length $\rho^{(j)}$ can be calculated from $\eta^{(j)}$ using Eqs. and . It can be seen that in this equilibrium dislocation structure, each array of dislocations (with the same Burgers vector) concentrate in two portions of the grain boundary, and on each point on the grain boundary, there are nonzero dislocation densities of two different Burgers vectors (except for some extreme points). In Fig \[circle\_motion\]b, the locations of dislocations using discrete dislocation model are also shown using the dislocation density potential functions. ### Motion under dislocation conservation {#sec:circle_conservation} ![Grain boundary motion under dislocation conservation. The initial grain boundary is circular. (a) Equilibrium dislocation structure on the initial circular grain boundary. There are three arrays of dislocations with Burgers vectors $\mathbf{b}^{(1)}$ (red), $\mathbf{b}^{(2)}$ (black), and $\mathbf{b}^{(3)}$ (blue), respectively. (b) Evolution of dislocation density potential functions of these three arrays of dislocations on the grain boundary (same colors as in (a)). The curves show the results of the continuum model and the dots show the results using the discrete dislocation model. (c) Motion of the grain boundary (shrinkage) by using our continuum model. (d) Motion of the grain boundary by using the discrete dislocation model. In (c) and (d), the grain boundary is plotted at uniform time intervals starting with the outer most one. []{data-label="circle_motion"}](fig2_circle_c.eps){width="0.65\linewidth"} In the continuum model in Eqs. and (or Eqs. and ), recall that the mobility $M_{\rm d}$ is associated with the conservative law motion of dislocations and the mobility $M_{\rm t}$ is associated with dislocation reactions. Recall that the dislocation structure on the grain boundary is stable and energy barriers have to be overcome for dislocations to react. In the simulations in this subsection, we set $M_{\rm t}=0$, i.e., the dislocations move entirely by the conservation law without dislocation reaction. This can happen to low angle grain boundaries whose sizes are not very small, as observed in atomistic [@Cahn20021] and phase-field crystal [@Wu2012407] simulations. Figure \[circle\_motion\]c and \[circle\_motion\]d show the evolution of the grain boundary obtained by using our continuum model and the discrete discrete model, respectively, under the same initial configuration and dislocation mobility. In the simulation results of the continuum model shown in Fig \[circle\_motion\]c, the circular grain boundary shrinks with increasing rate and keeps the circular shape during the shrinkage. This evolution is in excellent agreement with the simulation results using the discrete dislocation model shown in Fig \[circle\_motion\]d. The evolutions of the dislocation structure during the shrinking of this grain boundary by using the continuum model and the discrete dislocation model are shown in Fig \[circle\_motion\]b, based on the dislocation density potential functions $\eta^{(1)}$, $\eta^{(2)}$, and ${\eta^{(3)}}$ for dislocations with Burgers vectors $\mathbf{b}^{(1)}$, $\mathbf{b}^{(2)}$, and $\mathbf{b}^{(3)}$, respectively, as illustrated at the beginning of this section. We can see from this figure that the dislocation density potential functions using both the continuum model and the discrete model maintain their initial profiles during the evolution, meaning that the dislocation densities per unit polar angle do not change. This shows that all the dislocations move in the inward radial direction of the grain boundary using both models, as can also be seen from the results of the discrete model shown in Fig \[circle\_motion\]d. The Frank’s formula always holds and the long-range dislocation interaction vanishes during the motion of this circular grain boundary. ![Grain boundary motion under dislocation conservation. The initial grain boundary is circular. (a) The misorientation angle $\theta$ as a function of evolution time. (b) The radius $R$ as a function of evolution time. (c) $R\theta$ keeps constant during the evolution. The curves show the results of the continuum model and the dots show the results using the discrete dislocation model.[]{data-label="circle_R_rotation"}](R_theta_comparison_circle1_c.eps){width="0.9\linewidth"} In this simulation, the grain rotates by pure coupling effect as it shrinks. We calculate the misorientation angle $\theta$ during the shrinkage of the circular grain boundary by Eq. (where $R'\equiv0$). Fig. \[circle\_R\_rotation\]a shows the misorientation angle $\theta$ and $1/{\theta}^2$ as functions of the evolution time $t$. In Fig. \[circle\_R\_rotation\]b, the radius $R$ of the evolving circular grain boundary, as well as the quantity $R^2$ are shown in terms of $t$. It can be seen that the misorientation angle $\theta$ increases as the grain boundary shrinks, and the changing rates of both $\theta$ and $R$ are increasing. More precisely, the quantities $1/{\theta}^2$ and $R^2$ decrease linearly with $t$. We also examine the relation between the radius $R$ and misorientation $\theta$, and it can be seen that their multiplication $R\theta$ is constant during the evolution as shown in Fig. \[circle\_R\_rotation\]c. This relation agrees with the theoretical prediction by Eq. . Moreover, by the grain boundary evolution equation in Eq. , for this circular grain boundary, the long-range force ${\mathbf f}_{\rm long}$ vanishes as discussed above, and the evolution equation is reduced to $v_n=\frac{M_{\rm d}\mu b^2}{4\pi(1-\nu)R}$ and $\frac{dR}{dt}=-v_n$. This implies that $R(t)^2$ is a linear function of $t$: $R(t)^2=R(0)^2-\frac{M_{\rm d}\mu b^2}{2\pi(1-\nu)}t$. Accordingly, $1/\theta(t)^2$ is also a linear function of $t$ due to the fact that $R(t)\theta(t)$ remains constant. The simulation results discussed above agree with these theoretical predictions. Especially, both the slope of the numerical $R^2(t)$ curve in Fig. \[circle\_R\_rotation\]b and the theoretical formula of the slope $\frac{d R^2}{dt}=-\frac{M_{\rm d}\mu b^2}{2\pi(1-\nu)}$ give the same value $-0.2437M_{\rm d}\mu b^2$. ![Grain boundary motion under dislocation conservation. The initial grain boundary is circular. (a) The tangential velocity $v_{\\|}$ and normal velocity $v_{n}$ of the grain boundary. (b) The ratio $v_{\\|}/v_{n}$. (c) Verification of the coupling relation $v_{\\|}=\theta v_n$.[]{data-label="circle_v"}](v_comparison_circle2_c.eps){width="0.9\linewidth"} During the shrinkage of this circular grain boundary, the tangential velocity $v_{\\|}$ and the normal velocity $v_n$ of the grain boundary should satisfy the coupling relation $v_{\\|}=\theta v_n$ by our Eq. (with $\partial \varrho^{(j)}/\partial t=0$) or the theory of @Cahn20044887. We examine this relation by our continuum simulation as shown in Fig. \[circle\_v\]. The tangential velocity $v_{\\|}$ and the normal velocity $v_n$ during the evolution of the circular grain boundary are plotted in Fig. \[circle\_v\]a, where $v_{\\|}$ is calculated by $v_{\\|}=R d\theta/dt$. The relation of the two velocities obtained from our simulation results are shown in Fig. \[circle\_v\]b and Fig. \[circle\_v\]c. Excellent agreement can be seen between the numerical and theoretical values for the coupling relation $v_{\\|}=\theta v_n$. These simulation results of the continuum model and the discrete dislocation model agree excellently with the available atomistic [@Cahn20021; @Trautt20122407] and phase-field crystal [@Wu2012407] simulations, and the Cahn-Taylor theory [@Cahn20044887] which is a mechanism without details of the dislocation structure on the grain boundary. ### With dislocation reaction {#sec:circle_reaction} ![Grain boundary motion with dislocation reaction: $M_{\rm t}b/ M_{\rm d}=0.286\times 10^{-4}$. The initial grain boundary is circular. (a) Grain boundary motion using our continuum model. The grain boundary is plotted at uniform time intervals starting with the outer most one. (b) Grain boundary motion by using discrete dislocation dynamics model, where two pairs of dislocations in each of the three arrays of dislocations with different Burgers vectors are removed in the initial dislocation structure of the circular grain boundary. (c) Evolution of the dislocation density potential functions $\eta^{(1)}$, $\eta^{(2)}$, and ${\eta^{(3)}}$ of the three arrays of dislocations with Burgers vectors $\mathbf{b}^{(1)}$ (red), $\mathbf{b}^{(2)}$ (black), and $\mathbf{b}^{(3)}$ (blue), respectively. []{data-label="circle_motion1"}](fig5_circle_flocal_c.eps){width="0.6\linewidth"} In this subsection, we perform simulations using our continuum model for the case in which dislocation reaction is not negligible during the motion of the grain boundary, i.e., $M_{\rm t}\neq 0$. Recall again that the dislocation structure on the grain boundary is stable and energy barriers have to be overcome for dislocations to react. The simulation results from the initial, circular grain boundary with $M_{\rm t}b/ M_{\rm d}=0.286\times 10^{-4}$ are shown in Fig. \[circle\_motion1\]a. The initial circular grain boundary gradually changes to hexagonal shape as it shrinks. Evolution of the dislocation density potential functions $\eta^{(j)}$, $j=1,2,3$, is shown in Fig. \[circle\_motion1\]c. The amplitude of each $\eta^{(j)}$ is decreasing, meaning that the dislocations react and the number of dislocations of each Burgers vector is reduced. For the purpose of comparison, in the discrete dislocation dynamics simulation, we manually remove two pairs of dislocations in each of the three arrays of dislocations with different Burgers vectors in the initial dislocation structure of the circular grain boundary, and the simulation results are shown in Fig. \[circle\_motion1\]b. The grain boundary profile obtained in this discrete dislocation dynamics simulation also shows that the grain boundary evolves towards a hexagonal shape. Excellent qualitative agreement can be seen from the simulation results using these two models. Note that in the simulations using our continuum model, dislocations are removed continuously during the motion of the grain boundary. The reason to manually remove dislocations in the discrete dislocation dynamics simulations is that in order for the grain boundary dislocations to react, some energy barriers have to be overcome and these events are not included in the standard rules of discrete dislocation dynamics. In fact, the mechanisms of grain boundary dislocation reactions are non-trivial, e.g. by annihilation of dislocation pairs with opposite Burgers vectors located on opposite ends of the grain [@Cahn20021], or by a chain of dissociation and association of the GB dislocations [@Trautt20122407]. The evolution of this initially circular grain boundary into a hexagonal shape with dislocation reactions can be understood as follows. For a dislocation density $\varrho>0$ (same when $\varrho<0$), Eq. gives $\frac{\partial \gamma}{\partial \varrho}=\frac{\mu b^2}{4\pi(1-\nu)\sqrt{R^2+R'^2}}\log\frac{\sqrt{R^2+R'^2}}{r_g\varrho}$ (neglecting the numerical cutoff parameter $\epsilon$). Thus by Eq. with constant $M_{\rm t}\neq 0$, the rate of decrease of $\varrho$ due to $\frac{\partial \gamma}{\partial \varrho}$ is the largest when the density $\varrho$ is approaching $0$. It can be seen from Fig. \[circle\_motion1\]c (or Fig. \[circle\_motion\]a) that on the initial, circular grain boundary, dislocations with each Burgers vector occupy two arcs of the grain boundary, and these arcs of dislocations with different Burgers vectors overlap. During the evolution with $\frac{\partial \gamma}{\partial \varrho}$, all the arcs of dislocations shorten as discussed above and soon become disjoint. At this stage, all the dislocations on each arc have the same Burgers vector and the arc becomes planar (symmetric tilt) under the long-range elastic interaction, leading to a hexagonal shape of the grain boundary. ![Grain boundary motion with dislocation reaction for different reaction mobility $M_{\rm t}$: $M_{\rm t}b/ M_{\rm d}=0, 2.86\times 10^{-5}, 5.762\times 10^{-5}$, and $2.86\times 10^{-4}$ (from the top to the bottom). The initial grain boundary is circular. (a) Evolution of the misorientation angle $\theta$. (b) Evolution of the circumference of the grain boundary $L$. (c) Evolution of $L\theta$. []{data-label="circle_theta_R_flocal"}](R_theta_comparison_circle_eta_flocal1_c.eps){width="0.9\linewidth"} Evolution of the misorietation angle $\theta$ of this initial circular grain boundary with different values of reaction mobility $M_{\rm t}$ is shown in Fig. \[circle\_theta\_R\_flocal\]a, using the formula in Eq. . In this case, the evolution of the misorientation angle $\theta$ is controlled by both the coupling effect that depends on the dislocation mobility $M_{\rm d}$ and the sliding effect that depends mainly on the mobility $M_{\rm t}$ due to dislocation reaction. The coupling effect increases the misorientation angle $\theta$ while the sliding effect decreases $\theta$. This can be understood as follows. At a point on the initial circular grain boundary where all the dislocations have the same Burgers vector $\mathbf b$ that is in the outward radial direction $\hat{\mathbf R}$, by Eqs. , , and together with the cancellation of the long-range force in the initial dislocation structure, we have $$\frac{d\theta}{dt}=\frac{M_{\rm d}\mu b^2\theta}{4\pi(1-\nu)R^2} -\frac{M_{\rm t}\mu b^3}{4\pi(1-\nu)R^2}\log \left(\frac{1}{e r_g\sqrt{\varrho^2/R^2 +\epsilon}}\right).$$ The first term in this equation is due to the coupling effect and increases the misorientation angle $\theta$, whereas the second term is due to the sliding effect and decreases $\theta$. Note that the above equation is just for qualitative understanding because here we have dislocations with three different Burgers vectors and the grain boundary shape also changes during the evolution. As can be seen from the simulation results in Fig. \[circle\_theta\_R\_flocal\]a, when the dislocation reaction mobility $M_{\rm t}$ increases, meaning the sliding effect due to dislocation reaction is becoming stronger, the increase rate of $\theta$ is decreasing during the motion of the grain boundary, and when the sliding effect is strong enough, the misorientation angle $\theta$ is decreasing. Evolutions of the circumference of the grain boundary $L$ and the product $L\theta$ are shown in Fig. \[circle\_theta\_R\_flocal\]b and Fig. \[circle\_theta\_R\_flocal\]c. Note that here we consider the circumference for a non-circular grain boundary instead of the radius $R$ for a circular one. Recall that in the case without dislocation reaction discussed previously, $L\theta=2\pi R\theta$ keeps constant during the evolution of the grain boundary. When sliding effect is not negligible, $L\theta$ is decreasing during the evolution of the grain boundary due to the fact that the sliding effect decreases the misorientation angle, as shown in Fig. \[circle\_theta\_R\_flocal\]c, which is also in agreement with Eq. . However, the evolution of the circumference of the grain boundary $L$ does not change much for these values of $M_{\rm t}b/M_{\rm d}$ compared with the case of $M_{\rm t}=0$, see Fig. \[circle\_theta\_R\_flocal\]b. Elliptic grain boundary {#sec:ellipse} ----------------------- We then consider the motion of a grain boundary with initial shape of ellipse. The long axis of the ellipse is $140b$ and the ratio of the long axis to the short axis is $1.5$, see the outer most curve in Fig \[ellipse\_motion\]a. As in the circular grain boundary case, the equilibrium dislocation configuration on this elliptic grain boundary also consists of dislocations with Burgers vectors $\mathbf{b}^{(1)}$, $\mathbf{b}^{(2)}$, and $\mathbf{b}^{(3)}$, represented by dislocation density potential functions $\eta^{(1)}$, $\eta^{(2)}$, and ${\eta^{(3)}}$, respectively, as shown in Fig \[ellipse\_motion\]c. The locations of dislocations on this grain boundary using discrete dislocation model are also shown by the outer most curve in Fig \[ellipse\_motion\]b and in Fig \[ellipse\_motion\]c based on the dislocation density potential functions. ### Motion under dislocation conservation {#sec:ellipse_conservation} ![Grain boundary motion under dislocation conservation. The initial grain boundary is an ellipse. (a) Motion of the grain boundary (shrinkage) by using our continuum model. (b) Motion of the grain boundary by using the discrete dislocation model. In (a) and (b), the grain boundary is plotted at uniform time intervals starting with the outer most one. (c) Evolution of dislocation density potential functions $\eta^{(1)}$, $\eta^{(2)}$, and ${\eta^{(3)}}$, which represent dislocations on the grain boundary with Burgers vectors $\mathbf{b}^{(1)}$ (red), $\mathbf{b}^{(2)}$ (black), and $\mathbf{b}^{(3)}$ (blue), respectively. The curves show the results of the continuum model and the dots show the results using the discrete dislocation model.[]{data-label="ellipse_motion"}](fig7_ellip1_c.eps){width="0.6\linewidth"} We first consider the motion of this initially elliptical grain boundary under dislocation conservation without dislocation reaction, i.e., $M_{\rm t}=0$. In this case, the grain boundary motion is pure coupling and there is no sliding effect. Simulation result using our continuum model is shown in Fig \[ellipse\_motion\]a. We can see that the elliptic grain boundary shrinks with increasing rate and its shape does not change during the evolution. We also perform simulation using the discrete dislocation dynamics model to validate our continuum model, and the result is plotted in Fig \[ellipse\_motion\]b, which shows that all the dislocations move in the inward radial direction and the shape of the grain boundary is also preserved. This is in agreement with the theoretical prediction in Sec. \[sec:preserving\] and the prediction of the generalization of the Cahn-Taylor theory to noncircular grains based on assumptions of the coupling effect and mass transfer by surface diffusion [@Taylor2007493]. Evolutions of the dislocation structure on this grain boundary represented by dislocation density potential functions $\eta^{(1)}$, $\eta^{(2)}$, and ${\eta^{(3)}}$ are shown in Fig \[ellipse\_motion\]c. It can be seen that these functions do not change as the grain boundary shrinks, by using both the discrete dislocation model and the continuum model up to numerical errors. These simulation results indicate that the dislocation densities per unit polar angle, $\varrho^{(j)}={\eta^{(j)}}'$, $j=1,2,3$, do not change during the evolution of the grain boundary. Since such dislocation structure on the evolving grain boundary always satisfies the Frank’s formula, the simulation results using both the discrete and continuum models justify the assumption that the Frank’s formula holds during the evolution of the grain boundary, based on which the formulas and properties of the misorientation angle, grain rotation and coupling and sliding motions are derived in Secs. \[sec:Frank\] and \[sec:compare\]. We believe the small errors in the numerical results of the continuum model compared with the discrete model seen in Fig \[ellipse\_motion\]c are mainly due to the forward Euler type numerical schemes used in our simulations, see the beginning of this section. Note that this paper focus on derivation of the continuum model. Efficient and accurate numerical schemes for this continuum model will be explored in the future work. ![Grain boundary motion under dislocation conservation. The initial grain boundary is an ellipse. (a) The misorientation angle $\theta$ as a function of evolution time. (b) The circumference of the grain boundary $L$ as a function of evolution time. (c) $L\theta$ keeps constant during the evolution. The curves show the results of the continuum model and the dots show the results using the discrete dislocation model.[]{data-label="ellipse_R_rotation"}](R_theta_comparison_ellipse1_c.eps){width="0.9\linewidth"} Next, we examine the evolution of the misorientation angle $\theta$, which is calculated by using Eq. during the evolution of the grain boundary. Fig. \[ellipse\_R\_rotation\]a shows the misorientation angle $\theta$ as a function of the evolution time $t$. It can be seen that as in the simulation of a circular grain boundary, $\theta$ increases during the evolution and $1/\theta^2$ is a linear function of $t$. Evolution of the circumference of the elliptic grain boundary, $L$, is shown in Fig. \[ellipse\_R\_rotation\]b. The simulation results show that $L$ is decreasing and $L^2$ is a linear function of $t$, as the behavior of the radius of the circular grain boundary $R$ in the simulations in Sec. \[sec:circle\_conservation\]. That is L(t)\^2=L(0)\^2-Bt, where $B$ is some constant. These results are in excellent agreement with the discrete dislocation dynamics simulations. In this case, the quantity $L\theta$ does not change during the evolution, as the quantity $R\theta$ in the simulations of a circular grain in Sec. \[sec:circle\_conservation\]. This also agrees excellently with the discrete dislocation dynamics simulations as shown in Fig. \[ellipse\_R\_rotation\]c and the theoretic prediction in Eq. . ### With dislocation reaction {#sec:ellipse_reaction} In this subsection, we perform simulations from the initially elliptical grain boundary for the case in which dislocation reaction is not negligible during the motion of the grain boundary, i.e., $M_{\rm t}\neq 0$. The simulation results of the circular grain boundary with $M_{\rm t}b/ M_{\rm d}=0.286\times 10^{-4}$ are shown in Fig. \[ellipse\_motion1\]a. The initial circular grain boundary gradually changes to hexagonal shape as it shrinks, as the simulations from a circular grain boundary with dislocation reaction presented in Sec. \[sec:circle\_reaction\]. Evolution of the dislocation density potential functions $\eta^{(j)}$, $j=1,2,3$, is shown in Fig. \[ellipse\_motion1\]b. Again as the simulations from a circular grain boundary with dislocation reaction presented in Sec. \[sec:circle\_reaction\], the amplitude of each $\eta^{(j)}$ is decreasing, meaning that the dislocations react and the number of dislocations of each Burgers vector is reduced. ![Grain boundary motion with dislocation reaction: $M_{\rm t}b/ M_{\rm d}=0.286\times 10^{-4}$. The initial grain boundary is an ellipse. (a) Grain boundary motion using our continuum model. The grain boundary is plotted at uniform time intervals starting with the outer most one. (b) Evolution of the dislocation density potential functions $\eta^{(1)}$, $\eta^{(2)}$, and ${\eta^{(3)}}$ of the three arrays of dislocations with Burgers vectors $\mathbf{b}^{(1)}$ (red), $\mathbf{b}^{(2)}$ (black), and $\mathbf{b}^{(3)}$ (blue), respectively. []{data-label="ellipse_motion1"}](fig9_flocal_ellip_c.eps){width="0.6\linewidth"} ![Grain boundary motion with dislocation reaction for different reaction mobility $M_{\rm t}$: $M_{\rm t}b/ M_{\rm d}=0, 2.86\times 10^{-5}, 5.762\times 10^{-5}$, and $2.86\times 10^{-4}$ (from the top to the bottom). The shape of the initial grain boundary is an ellipse. (a) Evolution of the misorientation angle $\theta$. (b) Evolution of the circumference of the grain boundary $L$. (c) Evolution of $L\theta$. []{data-label="ellipse_angle"}](fig10_flocal_ellip_c.eps){width="0.9\linewidth"} Evolution of the misorietation angle $\theta$ from this initial elliptic grain boundary with different values of the reaction mobility $M_{\rm t}$ is shown in Fig. \[ellipse\_angle\]a, compared with the evolution of $\theta$ with $M_{\rm t}=0$ discussed previously. As discussed in the simulations of circular grain boundaries, the coupling effect that depends on the dislocation mobility $M_{\rm d}$ increases the misorientation angle $\theta$, while the sliding effect that depends mainly on the mobility $M_{\rm t}$ due to dislocation reaction decreases $\theta$. When the dislocation reaction mobility $M_{\rm t}$ increases, meaning the sliding effect due to dislocation reaction is becoming stronger, the increase rate of $\theta$ is decreasing during the motion of the grain boundary, and when the sliding effect is strong enough, the misorientation angle $\theta$ is decreasing. Evolutions of the circumference of the grain boundary $L$ and the product $L\theta$ are shown in Fig. \[ellipse\_angle\]b and Fig. \[ellipse\_angle\]c. Again as in the simulations of a circular grain boundary, the product $L\theta$ keeps constant in the case of $M_{\rm t}=0$ due to the pure coupling effect; whereas with dislocation reaction, the sliding effect is no longer negligible, and $L\theta$ is decreasing during the evolution of the grain boundary due to the fact that the sliding effect decreases the misorientation angle, as shown in Fig. \[ellipse\_angle\]c. The evolution of the circumference of the grain boundary $L$ does not change much for these values of $M_{\rm t}$ compared with the case of $M_{\rm t}=0$, see Fig. \[ellipse\_angle\]b. Pointwise misorientation angle $\theta$ --------------------------------------- In the simulations presented above, we calculate the misorientation angle $\theta$ of the grain boundary based on the formula in Eq. , which is an average of the pointwise misorientation angle along the grain boundary given by Eq. based on the Frank’s formula. In this subsection, we examine the pointwise misorientation angle formula in Eq. using the simulation results of our continuum model. ![Pointwise misorientation angle $\theta$ during the motion of the grain boundary, plotted at uniform time interval. The red line shows the values of $\theta$ on the initial grain boundary. (a) The shape of the initial grain boundary is a circle, under dislocation conservation. (b) The shape of the initial grain boundary is a circle, with dislocation reaction. (c) The shape of the initial grain boundary is an ellipse, under dislocation conservation. (d) The shape of the initial grain boundary is an ellipse, with dislocation reaction. In (b) and (d), $M_{\rm t}b/M_{\rm d}=0.286\times10^{-4}$. []{data-label="misorientation_angle"}](fig11_theta_evolution_c.eps){width="0.6\linewidth"} Figs. \[misorientation\_angle\]a and \[misorientation\_angle\]c show the pointwise misorientation angle $\theta$ for the simulation of the motion of the grain boundary under dislocation conservation for the circular and elliptic grain boundaries presented in Secs. \[sec:circle\_conservation\] and \[sec:ellipse\_conservation\], respectively. It can be seen that $\theta$ increases as the grain boundary evolves due to the pure coupling motion. In this case, $\theta$ is constant throughout the grain boundary during its motion, perfectly for the circular grain boundary and approximately for the elliptic one. Figs. \[misorientation\_angle\]b and \[misorientation\_angle\]d show the calculated pointwise value of $\theta$ for the simulation of the motion of the grain boundary with dislocation reaction for the intitially circular and elliptic grain boundaries presented in Secs. \[sec:circle\_reaction\] and \[sec:ellipse\_reaction\], respectively, where the mobility $M_{\rm t}$ associated with dislocation reaction is relatively large and $\theta$ decreases as the grain boundary evolves due to the strong sliding effect. In this case, $\theta$ is also a constant approximately throughout the grain boundary during its motion. Note that the grain boundary shape is changed once it evolves. The small deviations from a perfect constant of the pointwise misorientation angle $\theta$ on the grain boundary in Figs. \[misorientation\_angle\]b and d are due to the fact that the long-range elastic interaction between the constituent dislocations of the grain boundary that maintains the Frank’s formula is finite, thus is not able to make the Frank’s formula hold instantly after the dislocation reactions during the evolution. These numerical results show that the pointwise misorientation angle formula in Eq. indeed holds throughout the grain boundary during its evolution. Since Eq. is derived from the Frank’s formula, these simulation results also provide a justification that the Frank’s formula holds pointwise during the motion of the grain boundary, which is the basis for the formulas in Eqs. and for the misorientation angle (pointwise and averaged, respectively), Eq. for the grain rotation, and Eq. for the tangential velocity of the grain boundary. Conclusions and discussion {#sec:conclusion} ========================== In this paper, we present a continuum model for the dynamics of low angle grain boundaries in two dimensions based on the motion of constituent dislocations of the grain boundaries. The continuum model consists of an equation for the motion of the grain boundary that describes the motion of the constituent dislocations in the grain boundary normal direction, and equations for the dislocation structure evolution on the grain boundary (Eqs. and , or Eqs. and in Sec. \[sec:model\]). The long-range elastic interaction between dislocations is included in the continuum model, which maintains the dislocation structures on the grain boundaries and ensures that they are consistent with the condition of Frank’s formula for grain boundaries, i.e., the condition of cancellation of the far-field elastic fields. These evolutions of the grain boundary and its dislocation structure are able to describe both motion and tangential translation of grain boundaries and grain rotation due to both coupling and sliding. Since the continuum model is based upon dislocation structure, it naturally accounts for the grain boundary shape change during the motion and rotation of the grain boundary by motion and reaction of the constituent dislocations without explicit mass transfer. See Eqs. and for the pointwise and averaged misorientation angle formulas, Eq. for the grain rotation formula, and Eq. for the tangential velocity formula, derived using this continuum grain boundary dynamics model. Our continuum grain boundary dynamics model is based upon the continuum framework for grain boundaries in @Zhu2014175 derived from the discrete dislocation dynamics model, which is the basis for the development of grain boundary models and understanding of simulation and experimental results for grain boundaries [@ReadShockley1950; @Li1953223; @Cahn20021; @Cahn20044887; @Cahn20064953; @Cahn20063965; @Molodov20071843; @Molodov20095396; @Trautt20122407; @Wu2012407; @Voorhees2016264; @Voorhees2017; @Lim2009; @Quek2011; @Lim2012]. Compared with the theory of Cahn-Taylor [@Cahn20044887] in which the coupling effect is an assumption, our model is based on the motion of the grain boundary dislocations and the coupling effect is a result. Our continuum model generalizes the theory of Cahn-Taylor by (1) incorporating detailed formulas of the driving forces for the normal and tangential grain boundary velocities that depend on the constituent dislocations, their Burgers vectors, and the grain boundary shape, and (2) incorporating the shape change of the grain boundaries. Our model is different from their earlier generalized model based on mass transfer via surface diffusion [@Taylor2007493]. Compared with the existing continuum models for the motion of grain boundaries and grain rotation based on evolution of the misorientation angle [@Li1962; @Shewmon1966; @Kobayashi2000; @Upmanyu2006; @Selim2016], our dislocation-structure based continuum model is able to further describe the grain rotation due to the coupling of normal and tangential motions of the grain boundaries, which is missing in these existing models. Using the derived continuum grain boundary dynamic model, simulations are performed for the dynamics of circular and non-circular two dimensional grain boundaries, and the results are validated by discrete dislocation dynamics simulations. Simulations of both continuum and discrete models show that Frank’s formula holds approximately during the evolution of the grain boundary, and is maintained by the long-range elastic interaction of the constituent dislocations. For a finite grain embedded in another one with a low misorientation angle $\theta$, the coupling motion increases $\theta$ while the sliding motion reduces $\theta$. Simulations also show that when the grain boundary motion is pure coupling, its shape is preserved (which agrees with prediction of the model of [@Taylor2007493] based on mass transfer via surface diffusion), and the product $L\theta$ keeps constant, where $L$ is the circumference of the grain boundary. The product $L\theta$ is decreasing when the sliding effect is not negligible during the motion of the grain boundary. In general, a grain boundary is not able to maintain a circular shape during its evolution except for some special cases. The continuum model for the dynamics of low angle grain boundary presented in this paper can be generalized to multiple low angle grain boundaries with junctions. Generalizations can also be made to grain boundaries in three dimensions. These will be explored in the future work. Finally, we remark that continuum model for the dynamics of high angle grain boundaries has also been developed [@ZhangLuchan2] based on a disconnection model and atomistic simulations [@Thomas2017]. Parametrization of the grain boundary using polar angle $\alpha$ {#sec:geometryformulas} ================================================================ We parametrize the grain boundary by the polar angle $\alpha$ as shown in Fig. \[fig:geometry\]. A point on the grain boundary $\Gamma$ has the coordinate $\mathbf R(\alpha)= (x(\alpha),y(\alpha))=(R(\alpha)\cos\alpha,R(\alpha)\sin\alpha)$. The grain boundary tangent direction $\mathbf T$, the grain boundary normal direction $\mathbf n$, and the curvature of the grain boundary $\kappa$ at a point $(x(\alpha),y(\alpha))$ on the grain boundary can be calculated as T&=[(x’(),y’())]{}, \[eqn:tangent\] n&=[(-y’(),x’())]{}, \[eqn:normal\] &= . \[eqn:curvature\] The normal direction of the grain boundary $\mathbf n$ in Eq. is defined such that $\mathbf T\times \mathbf n=\mathbf k$, where $\mathbf k$ is the unit vector in the $+z$ direction. We have the relation =, =-T, where $s$ is the arclength parameter of the grain boundary $\Gamma$. We also have the relation =, and ==. \[eqn:arclength\] When the grain boundary is evolved with a velocity in the radial direction $\partial R/\partial t$, the evolution of the normal direction of the grain boundary is \[eqn:dndt\] =\^[-1]{}()T =T. Variational derivation of the continuum model for grain boundary motion {#sec:variation} ======================================================================= In the classical theories for the motion of grain boundaries, the driving force is obtained by variation of the grain boundary energy whose density does not evolve as the grain boundary migrates [@Herring1951; @Mullins1956; @Sutton1995]. Although in some models, further variations have been taken to reduce the grain boundary energy density (grain boundary sliding) [@Li1962; @Kobayashi2000; @Upmanyu2006; @Selim2016], these models are not able to include the grain boundary coupling motion in which the grain boundary energy density may increase [@Cahn20021; @Cahn20044887]. In the continuum framework for grain boundaries and dislocation arrays proposed in @Zhu2014175, the driving force for grain boundary or dislocation array motion was obtained by taking variation of the total energy with respect to the change of the grain boundary profile under the conservation of the constituent dislocations. This enables the incorporation of both the coupling and sliding motions of the grain boundary when it evolves. This continuum framework is general: it applies to dislocation structures on grain boundaries in three dimensions and includes both the long-range and short-range interactions of the constituent dislocations. In this section of Appendix, we present an example of such variation method for the derivation of the continuum grain boundary motion model in two dimensions, in terms of the contribution of the grain boundary energy due to the local interaction of dislocations which is the commonly used grain boundary energy in the literature. This approach is a generalization of that by @Cahn20021 and @Cahn20044887 for two dimensional circular grain boundaries, and is different from the model by Talor-Cahn based on explicit mass transfer along the grain boundaries [@Taylor2007493]. Consider a grain boundary $\Gamma$ as shown in Fig. \[fig:geometry\]. For the purpose of demonstration of the method, we consider the simple case where all the dislocations have the same Burgers vector $\mathbf b$ which is also in the $xy$ plane. The local grain boundary energy in Eq. in this case is, E\_[local]{}&=\_ds,\[eqn:gb\_energyA\]\ &= \|\| , \[eqn:gb\_densityA\] using the dislocation density per unit length $\rho$. The variation with respect to the change of $\Gamma$ is taken under the conservation of the constituent dislocations. Assuming that there is a small change in the location of the grain boundary $\Gamma$ in its normal direction $\mathbf n$, denoted by $\delta r \mathbf n$, the change of energy is \[eqn:delta\_E\] E\_[local]{}=(\_ds) =\_() ds+\_ (ds). In the classical theory of motion by curvature [@Herring1951; @Mullins1956; @Sutton1995], it has been shown that the change of a small arclength $ds$ is \[eqn:delta\_ds\] (ds)=-r ds, where $\kappa$ is the curvature of the grain boundary. For the small arclength $ds$, the total number of dislocations along $ds$ is $\|\rho\| ds$. (Note: Using the dislocation density potential function $\eta$, it is $d\eta=\|\rho\| ds$, which is very convenient in the derivation in three dimensions [@Zhu2014175].) Conservation of the constituent dislocations gives (\|\| ds)=\|\| ds+\|\| (ds)=0. Further using Eq. , we have $\delta\|\rho\|\! ds+\|\rho\| (-\kappa \delta r ds)=0$, or \|\|=\|\|r. Using the energy density formula in Eq. , it can be calculated that \[eqn:delta\_gamma\] ()=\|\| =\|\|r. Substituting Eqs. and into Eq. , we have E\_[local]{} =-\_(-\|\| )r ds. Therefore, we have the variation of the energy with respect to the change of the grain boundary profile: \[eqn:variation\_delta\_r\] =- (-\|\| )=- \|\|. This is a special case of the variation obtained in @Zhu2014175, see Eq. with in Sec. \[sec:settings\]. Therefore, the normal velocity of the grain boundary due to the driving force of the variation of the local grain boundary energy is v\_n=- = M\_[d]{}, where $M_{\rm d}$ is the mobility of the constituent dislocations of the grain boundary. Here a factor $1/\|\rho\|$ has been included with the dislocation mobility $M_{\rm d}$ so that the velocity of the grain boundary is consistent with the dynamics of its constituent dislocations [@Cahn20044887]. This can be seen more explicitly from the full variation formula in Eqs. and with all types of driving forces. Next we consider the variation of the local grain boundary energy in Eqs. and with respect to the change of the dislocation density per unit polar angle $\varrho$ on the fixed grain boundary. It is easy to calculate that =. Therefore, the evolution of the dislocation density $\varrho$ on the grain boundary due to this variation is \_t=-M\_[t]{} =-M\_[t]{}, where $M_{\rm t}$ is the mobility. Proof of equivalence of Frank’s formula and cancellation of the long-range elastic fields {#sec:equivalenceproof} ========================================================================================= A proof of the equivalence of the Frank’s formula and cancellation of the long-range elastic fields generated by the constituent dislocations for a curved grain boundary can be found in @Zhu2014175. Here we present an alternative calculation in two dimensions to show their equivalence. In fact, the long-range stress field generated by the constituent dislocations of the grain boundary $\Gamma$, in the infinite two dimensional space, can be written as \[generalstress\] \_[i]{}(x,y)=&\_[j=1]{}\^J\_(G\_1(x,y;x\_1,y\_1)b\^[(j)]{}\_1\^[(j)]{}(x\_1,y\_1)+G\_2(x,y;x\_1,y\_1)b\^[(j)]{}\_2\^[(j)]{}(x\_1,y\_1))ds, where the point $(x_1,y_1)$ varies along the grain boundary in the integrals, $ds$ is the line element of the integrals, and G\_1(x,y;x\_1,y\_1)&= ( [cc]{}\[G1\] -&\ & ),\ G\_2(x,y;x\_1,y\_1)&= ( [cc]{} &\ & ).\[G2\] Using divergence theorem and the fact that $\frac{\partial}{\partial x_1}\mathbf G_1+ \frac{\partial}{\partial y_1}\mathbf G_2=\mathbf 0$, we have \[Gdivergence\] \_(G\_1n\_1+G\_2n\_2)ds =\_(+ )dx\_1dy\_1=0, where $\mathbf n=(n_1,n_2)$ is recalled to be the unit normal vector of the grain boundary and $\Omega$ is the inner grain. Using Eqs. and , the stress field due to the long-range dislocation interaction can be written as \[generalstress2\] \_[i]{}(x,y)=&\_[j=1]{}\^J\_ds. Therefore, the long-range stress field $\pmb\sigma_{\rm i}$, and accordingly the long-range elastic interaction energy of the constituent dislocations, vanish when the Frank’s formula in Eq. holds. Derivation of the grain rotation formula in Eq. {#eqn:grainrotationderivation} ================================================ The grain rotation formula in Eq. is derived as follows. Taking time derivative in Eq. , we have \[eqn:dFrankdt\] n+ +\_[j=1]{}\^J \^[(j)]{}b\^[(j)]{}() +\_[j=1]{}\^J =0. Taking dot product with $\mathbf n$ and $\mathbf T$, respectively, in Eq. and using Eq. and the Frank’s formula in Eq. , we have =- -\_[j=1]{}\^J\ = -\_[j=1]{}\^J . The second equation gives = -\_[j=1]{}\^J (b\^[(j)]{}T), and inserting it into the first equation, we have =- +\_[j=1]{}\^J (b\^[(j)]{}T-b\^[(j)]{}n). This gives Eq. by using the equation that $\frac{R'}{\sqrt{R^2+R'^2}}\mathbf T-\frac{R}{\sqrt{R^2+R'^2}}\mathbf n=\hat{\mathbf R}$, which can be obtained using the formulas of $\mathbf T$ and $\mathbf n$ in Eqs. and . Acknowledgments {#acknowledgments .unnumbered} =============== This work was partially supported by the Hong Kong Research Grants Council General Research Fund 606313.	{ "pile_set_name": "ArXiv" }
--- author: - 'L. F. F. Freitas,' - 'G. Alencar' - 'and R. R. Landim' title: Consistency Conditions for Fields Localization on Braneworlds --- Introduction {#ASec-1} ============ Braneworld scenarios gained prominence after the emergence of the $5-$dimensional warped models introduced by Randall-Sundrum (RS) [@RS1; @RS2]. Since then, many other models with localized gravity have been proposed. Some of them also in $5$D, such as: thick brane generated by scalar fields with different potential functions [@Gremm01; @Bazeia04]; deformed brane models with internal structure [@Bazeia05; @Bazeia02]; thick brane with purely geometric features [@Cendejas]; beyond proposals in cosmological contexts [@Singh]; or $f(R)$ theories [@Hoff], among others. A comprehensive and more detailed review of thick braneworld can be found in Ref. [@Dzhunushaliev]. Besides these $5$D scenarios, other proposals in higher-dimensional configuration were presented. For example, in $6-$dimension, braneworld models generated by string-like topological defect with a scalar field [@Gregory; @Gherghetta]; vortex defect in the context of abelian Higgs model [@Giovannini01]; or cigar-like thick braneworld [@Carlos; @Silva]. And also, versions of these models in cosmological context [@Cline; @Kofinas], among others [@Kim; @Coley; @Choudhury01]. There are also proposals in higher-dimensional scenarios [@Arkani02; @Oda01; @Bazeia01; @Navarro01]. Amid this large variety of works, we can find some with the focus on the general features of braneworlds, as the Refs. [@Csaba; @Gibbons]. In these papers, it is studied the consistency of the gravitational field, Newton’s law on the braneworld, search for resonant gravitational modes, and other related issues. In this braneworld context, beyond the gravitational field, the issue of the Standard Model (SM) fields localization should also be verified. Some general studies have been performed in the literature. The Ref. [@Barut] presents a detailed study of the free spin $0,\ \frac{1}{2}$ and $1$ fields localization in RS-II delta-like brane model. Among these fields only the free scalar field and the left-handed spinor can be confined on that model. In Refs. [@Oda02; @Liu01; @Liu03; @Liu04] was analyzed the localization of the above fields for thick branes embedded in $AdS_{5}$ space. For these cases, the scalar and the left-handed spinor can be confined, and again the $U(1)$ gauge field is not localized. Another study, performed for thick branes embedded in $dS_{5}$ space [@Liu02; @Guo], showed the same results for the scalar and the spinor fields, however, unlike the early models, the free vector field can be confined in such models. This same analysis was also performed for other dimensional configurations. In $6$D string-like models, for example, the Refs. [@Oda03; @Oda06] show that the free spin $0$ and $1$ fields can be localized. In Ref. [@Midodashvili], the author shows that, beyond the scalar and the $U(1)$ vector fields, the free spin $\frac{1}{2}$ field can also be confined on the brane. In arbitrary higher-dimensions, the Refs. [@Oda04; @Oda05] obtained the same results above for the Standard Model fields. With this, we already have an indication that the confinement of fields, mainly the spinor and the gauge vector fields, is closely related to the geometric features of space. In another direction, aiming to obtain the localization of fields, some mechanisms were proposed. For example, in Refs. [@Koley01; @Ringeval; @Melfo; @Castro; @Mendes] the localization of spin $\frac{1}{2}$ field is obtained for various braneworld scenarios by proposing a Yukawa-like coupling. The Ref. [@Batell] provides the confinement of a non-abelian Yang-Mills field by introducing non-minimal couplings with gravity. Among others, for these and other fields, for those and various other braneworld models [@Casadio; @Alencar01; @Germani; @GLandim01; @GLandim02]. In most $5-$dimensional braneworlds, the free $U(1)$ gauge field cannot be confined, as mentioned above. However, by adding some suitable interaction terms, that field can be localized. For example, in RS-II like models, the localization can be obtained by adding an interaction term between a $3$-form field (topological term) and the vector field [@Oda07]; or by proposing a non-covariant mechanism with two mass terms, one in the bulk and another just in brane [@Ghoroku], the confinement is obtained by fine-tuning these mass parameters; or still, by proposing non-minimal interactions between the gravity and the vector field through the Ricci scalar [@Alencar02; @Zhao]. For $5$D thick brane models, the Refs. [@Kehagias; @Fu01; @Chumbes; @Alencar03; @Cruz01; @Cruz02] proposed a modified kinetic term of the vector gauge field by adding couplings with a scalar field function. In $6$D models, the Ref. [@Flachi] proposed a delta-like brane generated by brane intersection and, to confine the vector field, the author proposes interaction terms between this and the scalar Ricci, and/or the Ricci tensor. In higher-dimensional models, the Ref. [@Freitas] carried out a general study of the abelian vector field localization through the couplings with the scalar and the Ricci tensor. Despite these various results indicate that the Standard Model fields can be well-defined on the braneworld scenarios, there is not yet a study on the consistency of the localization procedure. Generally, when we talk about fields localization on braneworlds, it means that we wish to factor out an action $S^{(D)}$ defined on the bulk into a sector containing an effective action $S^{(4)}_{eff}$ on the $3$-brane and an integral $K$ in the coordinates of the extra dimensions. Thus, we say that the theory is well-defined, [i.e.]{}, the field is localized on the brane, when the integral $K$ is finite. In this manuscript, we aim to analyze the consistency of this localization procedure with the Einstein’s equations for the SM fields. Special attention is paid to the $U(1)$ gauge field, both the free field and the cases where some localization mechanism is used. This work is organized as follows: In section (\[ASec-2\]), we obtain two general consistency conditions that must be satisfied by any field in order that the localization procedure to be consistent. We apply these conditions for spin $0$ and $\frac{1}{2}$ fields in sections (\[ASec-3\]) and (\[ASec-4\]), respectively. In section (\[ASec-5\]), we carry out the consistency analysis for the vector field, free and with some localization mechanism. Conclusions are left for section (\[ASec-6\]). Einstein Equations - Consistency Conditions {#ASec-2} =========================================== To perform a more general and comprehensive discussion, let us consider a braneworld in $(D=d+n)-$dimensions with metric given by the generic ansatz $$\label{2-01} ds^{2}=g_{MN}dx^{N}dx^{M}=e^{2\sigma(y)}\hat{g}_{\mu\nu}(x)dx^{\mu}dx^{\nu}+\Bar{g}_{jk}(y)dy^{j}dy^{k}.$$ Where $\underline{d}$ is the brane dimension, indexed by $(\mu,\nu,..)$, and $\underline{n}$ is the number of extra dimensions, labeled by $(j,k,..)$. Beyond this, the metric (\[2-01\]) will be considered diagonal with signature $(-,+,+,..)$. As mentioned before, in the study of fields localization on braneworld scenarios like (\[2-01\]), we wish to factor out a matter Lagrangian like $$\begin{aligned} \label{2-02} S=\int d^{d}xd^{n}y\sqrt{-g^{(D)}}\mathcal{L}^{(D)}(x,y),\end{aligned}$$ into an effective action on the brane and an integral in the extra dimension, i.e., $$\begin{aligned} \label{2-03} S=\int d^{n}yf(y)\int d^{d}x\sqrt{-\hat{g}^{d}(x)}\mathcal{L}_{ef}^{(d)}(x)=K\int d^{d}x\sqrt{-\hat{g}^{d}(x)}\mathcal{L}_{ef}^{(d)}(x).\end{aligned}$$ From this, we say that the theory is well-defined, or the field is localized, if $K$ is finite. In doing this, the metric (\[2-01\]) is considered only as a background previously determined by some process. And also, it is assumed that the matter Lagrangian (\[2-02\]) will not change the background geometry. Here, we will discuss exactly the consistency of this last assumption by studying the Einstein’s equations. The Einstein-Hilbert action for an arbitrary braneworld model, with a matter Lagrangian like (\[2-02\]), can be written as $$\label{2-04} S=S_{brane}+S_{matter}=\int d^{d}xd^{n}y\sqrt{-g}\left[\frac{1}{2\kappa^{2}}\left(R-2\Lambda\right)+\mathcal{L}^{b}(y)+\mathcal{L}^{m}(x,y)\right].$$ Here, $\mathcal{L}^{b}$ is related to the brane generation mechanism, and it is a function only of extra dimensions. The term $\mathcal{L}^{m}$ is the matter Lagrangian related to the Standard Model fields, which the confinement should be studied. By performing the variation of action (\[2-04\]) with respect to the metric, we get the bellow Einstein’s equations $$\begin{aligned} \label{2-05} R_{MN}-\frac{1}{2}g_{MN}R+g_{MN}\Lambda=\kappa^{2}\left(T^{b}_{MN}+T^{m}_{MN}\right).\end{aligned}$$ We should add to Eq. (\[2-05\]) the equations of motion (EOM) related to the fields in the Lagrangians $\mathcal{L}^{b}$, and $\mathcal{L}^{m}$. Fortunately, these EOM’s are not important to our discussion, and they will not be written here. From metric (\[2-01\]), we can get the Ricci tensor components, $$\begin{aligned} R_{\mu\nu}(x,y)&=&\hat{R}_{\mu\nu}(x)-\frac{1}{d}\hat{g}_{\mu\nu}(x)e^{-(d-2)\sigma(y)}\nabla_{k}\nabla^{k}e^{d\sigma(y)},\label{2-06}\\ R_{jk}(x,y)&=&\Bar{R}_{jk}(y)-de^{-\sigma(y)}\nabla_{j}\nabla_{k}e^{\sigma(y)},\label{2-07}\end{aligned}$$ and the components $R_{j\mu}=0$. From these, the Ricci scalar can be written as $$\begin{aligned} R(x,y)=\hat{R}(x)e^{-2\sigma}+\Bar{R}(y)-de^{-\sigma(y)}\nabla_{k}\nabla^{k}e^{\sigma(y)}-e^{-d\sigma(y)}\nabla_{k}\nabla^{k}e^{d\sigma(y)},\label{2-08}\end{aligned}$$ where $\Bar{R}=\Bar{g}^{jk}\Bar{R}_{jk}$ and $\hat{R}=\hat{g}^{\mu\nu}\hat{R}_{\mu\nu}$. That way, the Eq. (\[2-05\]) can be separated in the following two equations $$\begin{aligned} R_{\mu\nu}(x,y)-\frac{1}{2}g_{\mu\nu}(x,y)R(x,y)+g_{\mu\nu}(x,y)\Lambda=\kappa^{2}\left[T^{b}_{\mu\nu}(x,y)+T^{m}_{\mu\nu}(x,y)\right],\label{2-09}\\ R_{jk}(x,y)-\frac{1}{2}g_{jk}(x,y)R(x,y)+g_{jk}(x,y)\Lambda=\kappa^{2}\left[T^{b}_{jk}(x,y)+T^{m}_{jk}(x,y)\right].\label{2-10}\end{aligned}$$ By using Eqs. (\[2-06\]), (\[2-07\]) and (\[2-08\]), the equation (\[2-09\]) can be written as $$\begin{aligned} \label{2-11} \hat{G}_{\mu\nu}(x)+\hat{g}_{\mu\nu}(x)\!\left[\!\frac{1}{2}S(y)-\frac{1}{2}\Bar{R}(y)+\Lambda\!\right]\!e^{2\sigma(y)}=\kappa^{2}\!\left[T^{b}_{\mu\nu}(x,y)+T^{m}_{\mu\nu}(x,y)\right],\end{aligned}$$ where $S(y)=(d-1)\left[2\nabla_{k}\nabla^{k}\sigma+d\nabla^{k}\sigma\nabla_{k}\sigma\right]$. When the equation (\[2-11\]) is solved to obtain the vacuum braneworld metric, the matter Lagrangian $\mathcal{L}^{m}$ is considered equal to zero, and the Lagrangian $\mathcal{L}^{b}$ is a function only of the extra dimensions. In this setup, $T^{b}_{\mu\nu}(x,y)=\hat{g}_{\mu\nu}(x)e^{2\sigma(y)}\mathcal{L}^{b}(y)$, and, with this, we can perform the separation of variables in Eq. (\[2-11\]) as $$\begin{aligned} \label{2-12} \left[\frac{1}{2}S(y)-\frac{1}{2}\Bar{R}(y)+\Lambda-\kappa^{2}\mathcal{L}^{b}(y)\right]\!e^{2\sigma(y)}=\alpha,\end{aligned}$$ with $\alpha$ a constant that will be interpreted as the cosmological constant on the brane. Thus, braneworld metric in the vacuum $(\mathcal{L}^{m}=0)$ should satisfy the equation (\[2-12\]). In the study of fields localization, where the starting point is the matter Lagrangian $\mathcal{L}^{m}$, the metric is exactly that obtained in the vacuum, and therefore it satisfies (\[2-12\]). Now, let us assume that Eq. (\[2-12\]) is still valid even after adding Lagrangian $\mathcal{L}^{m}$. With this, Eq. (\[2-11\]) can be written as, $$\begin{aligned} \hat{G}_{\mu\nu}(x)+\hat{g}_{\mu\nu}(x)\alpha=\kappa^{2}T^{m}_{\mu\nu}(x,y). \label{2-13}\end{aligned}$$ Thus, we observe that the left-hand side of (\[2-13\]) does not depend on the extra dimensions $y^{j}$, therefore, for consistency reasons, the energy-momentum tensor of matter field should satisfy the following condition $$\begin{aligned} \label{2-14} T^{m}_{\mu\nu}(x,y)=\hat{T}^{m}_{\mu\nu}(x).\end{aligned}$$ About equation (\[2-10\]), it can also be simplified by using Eqs. (\[2-06\]), (\[2-07\]) and (\[2-08\]). By doing this, we get a condition on the components $T^{m}_{jk}$ given by $$\begin{aligned} \label{2-15} T^{m}_{jk}(x,y)=\left\lbrace\begin{array}{c} 0, \ \mbox{or} \\ \Bar{g}_{jk}(y)e^{-2\sigma(y)}\Bar{T}^{m}(x). \end{array} \right.\end{aligned}$$ These consistency conditions, (\[2-14\]) and (\[2-15\]), are completely independent of the brane model, the number of extra dimensions, and also of the matter field considered in $\mathcal{L}^{m}$. That way, such conditions have a general valid, and should be satisfied for any model with the features considered above. Note that these conclusions are closely related to the possibility of the metric not changing by the presence of matter fields. In other words, these consistency conditions mean that back-reaction effects from the matter fields on the bulk geometry can be eliminated. Let us apply these results for some known cases. Applications - Scalar field {#ASec-3} =========================== Let us start by discussing the scalar field localization. The Lagrangian for a massless scalar field in the braneworld model (\[2-01\]) is given by $$\begin{aligned} \label{3-01} \mathcal{L}^{m}(x,y)=-\frac{1}{2}\partial_{M}\Phi(x,y)\partial^{M}\Phi(x,y).\end{aligned}$$ In studying the localization of this field, we can obtain the equation of motion $$\begin{aligned} \label{3-02} \partial_{M}\left[\sqrt{-g}\partial^{M}\Phi\right]=e^{(d-2)\sigma}\sqrt{\Bar{g}}\partial_{\mu}\left[\sqrt{-\hat{g}}\hat{g}^{\mu\nu}\partial_{\nu}\Phi\right]+\sqrt{-\hat{g}}\partial_{j}\left[\sqrt{\Bar{g}}e^{d\sigma}\Bar{g}^{jk}\partial_{k}\Phi\right]=0,\end{aligned}$$ where $\Bar{g}$ is the determinant of $\Bar{g}_{jk}(y)$, and $\hat{g}$ is the determinant of $\hat{g}_{\mu\nu}(x)$. With this, by proposing $\Phi(x,y)=\phi(x)\xi(y)$, it is possible to separate the variables for the zero-mode as $$\begin{aligned} \frac{1}{\sqrt{-\hat{g}}}\partial_{\mu}\left[\sqrt{-\hat{g}}\hat{g}^{\mu\nu}\partial_{\nu}\phi_{0}(x)\right]=0,\label{3-03}\\ -\frac{e^{-(d-2)\sigma}}{\sqrt{\Bar{g}}}\partial_{j}\left[\sqrt{\Bar{g}}e^{d\sigma}\Bar{g}^{jk}\partial_{k}\xi_{0}(y)\right]=0.\label{3-04}\end{aligned}$$ From this, a solution for (\[3-04\]) can be obtained, and the localization can be analyzed. As discussed in many Refs. [@Barut; @Kehagias; @Oda02; @Oda03; @Liu01; @Liu02; @Fu01; @Midodashvili], there is a constant solution for (\[3-04\]) that can be confined for a wide variety of models. In order to test the consistency conditions (\[2-14\]) and (\[2-15\]) for zero-mode, let us calculate the energy-momentum tensor from (\[3-01\]). In doing this, we get $$\begin{aligned} \label{3-05} T^{m}_{MN}(x,y)=-\frac{2}{\sqrt{-g}}\frac{\delta\left(\sqrt{-g}\mathcal{L}^{m}\right)}{\delta g^{MN}}=\partial_{M}\Phi(x,y)\partial_{N}\Phi(x,y)+g_{MN}\mathcal{L}^{m}(x,y).\end{aligned}$$ By using the constant solution $\xi_{0}(y)=c_{0}$ for (\[3-04\]), the components $T^{m}_{\mu\nu}$, obtained from equation (\[3-05\]), can be written as $$\begin{aligned} \label{3-06} T^{m}_{\mu\nu}(x,y)=c^{2}_{0}\left[\partial_{\mu}\phi(x)\partial_{\nu}\phi(x)+\hat{g}_{\mu\nu}(x)\hat{L}_{0}^{m}(x)\right].\end{aligned}$$ Therefore, the consistency condition (\[2-14\]) is immediately satisfied. About condition (\[2-15\]), we can get, from (\[3-05\]), for the zero-mode, that $$\begin{aligned} \label{3-07} T^{m}_{jk}(x,y)=\Bar{g}_{jk}(y)e^{-2\sigma(y)}c^{2}_{0}\hat{L}_{0}^{m}(x).\end{aligned}$$ Thus, by comparing this with (\[2-15\]), we conclude that both consistency conditions are satisfied. Therefore, the free scalar field (zero-mode) localization is consistent with Einstein’s equations, and any possible back-reaction effect from the scalar field on the background metric must be caused by the massive modes. Note that nowhere was it necessary to define the braneworld model, or the number of extra dimensions, for the consistency conditions to be met. In this way, these results for the zero-mode of scalar field are valid for a wide variety of models, whether for those with thin or thick brane, and for arbitrary co-dimension. Applications - Spinor field {#ASec-4} =========================== Now let us see briefly the spin $\frac{1}{2}$ field localization for an arbitrary co-dimension $1$ model. In this particular configuration, the metric (\[2-01\]) will be written as $$\label{3-08} ds^{2}=e^{2\sigma_{1}(y)}\hat{g}_{\mu\nu}(x)dx^{\mu}dx^{\nu}+e^{2\sigma_{2}(y)}dy^{2},$$ with $\mu,\nu=1,2,...,d$ (even). Beyond this, the consistency conditions are given by $$\begin{aligned} \label{3-09} T^{m}_{\mu\nu}(x,y)=\hat{T}^{m}_{\mu\nu}(x), \ \ \ \mbox{and}\ \ \ T^{m}_{jk}(x,y)=\left\lbrace\begin{array}{c} 0, \ \mbox{or} \\ e^{2(\sigma_{2}-\sigma_{1})}\Bar{T}^{m}(x). \end{array} \right.\end{aligned}$$ We will consider the spinor field coupled to an arbitrary scalar function $f(y)$ through a Yukawa-like interaction term. In order that the Lagrangian for this case will be written as $$\begin{aligned} \label{3-10} \mathcal{L}^{m}(x,y)=-i\Bar{\Psi}\Gamma^{M}D_{M}\Psi+\lambda f(y)\Bar{\Psi}\Psi,\end{aligned}$$ where $D_{M}=\partial_{M}+\omega_{M}$, and $\omega_{M}=\frac{1}{4}\omega^{ab}_{M}\Gamma_{a}\Gamma_{b}$ are the spin connections. The Gamma matrix in curved space $\Gamma^{M}$ are related to those in a local flat frame by[^1] $\Gamma^{M}(x,y)=E^{M}_{a}(x,y)\Gamma^{a}$. From (\[3-10\]), the following equation of motion can be obtained, $$\begin{aligned} \label{3-11} \left[i\Gamma^{M}D_{M}-\lambda f(y)\right]\Psi=\left[i\Gamma^{\mu}D_{\mu}+i\Gamma^{y}D_{y}-\lambda f(y)\right]\Psi=0.\end{aligned}$$ By defining[^2] $E^{\mu}_{a}(x,y)=e^{-\sigma_{1}}\hat{e}^{\mu}_{a}(x)$, $E^{y}_{a}=0$ ($a=1,2,..,d$), $E^{\mu}_{d+1}=0$ and $E^{y}_{d+1}=e^{-\sigma_{2}}\delta^{y}_{d+1}$, the spin connections can be calculated $$\begin{aligned} \label{3-12} \omega_{\mu}(x,y)=\hat{\omega}_{\mu}(x)+\frac{1}{2}\partial_{y}\sigma_{1}\Gamma_{\mu}\Gamma^{y},\ \ \ \omega_{y}(x,y)=0.\end{aligned}$$ And with this, the equation (\[3-11\]) gives us $$\begin{aligned} \label{3-13} i\hat{\Gamma}^{\mu}(x)\hat{D}_{\mu}\Psi-i\frac{d}{2}e^{\sigma_{1}}\partial_{y}\sigma_{1}\Gamma^{y}\Psi+ie^{\sigma_{1}}\Gamma^{y}\partial_{y}\Psi-\lambda e^{\sigma_{1}}f(y)\Psi=0,\end{aligned}$$ where $\hat{D}_{\mu}=\partial_{\mu}+\hat{\omega}_{\mu}(x)$. Here, to solve the above equation, let us consider that the zero-mode solution satisfies $-i\hat{\Gamma}^{\mu}\hat{D}_{\mu}\Psi_{0}=0$. That way, we get for the massless mode $$\begin{aligned} \label{3-14} i\frac{d}{2}e^{\sigma_{1}-\sigma_{2}}\partial_{y}\sigma_{1}\Gamma^{d+1}\Psi_{0}-ie^{\sigma_{1}-\sigma_{2}}\Gamma^{d+1}\partial_{y}\Psi_{0}+\lambda e^{\sigma_{1}}f(y)\Psi_{0}=0.\end{aligned}$$ At this point, we will use a Gamma matrix representation such that, $\Gamma^{d+1}$ is a $d\times d$ diagonal matrix (remember that $d$ is even) in the following shape $$\begin{aligned} \label{3-15} \Gamma^{d+1}=i\left[\begin{array}{cc} \mathbf{I}_{\frac{d}{2}} & \ \mathbf{0}_{\frac{d}{2}} \\ \mathbf{0}_{\frac{d}{2}} & -\mathbf{I}_{\frac{d}{2}} \end{array}\right],\end{aligned}$$ where $\mathbf{I}_{\frac{d}{2}}$ is a $\frac{d}{2}\times\frac{d}{2}$ identity matrix. Thus, we can define the $\frac{d}{2}$-dimensional spinors $\Psi_{0}^{\pm}(x,y)$, such that, $$\begin{aligned} \label{3-16} \Psi_{0}(x,y)= \left[\begin{array}{c} \Psi_{0}^{+}(x,y) \\ \Psi_{0}^{-}(x,y) \end{array}\right]= \left[\begin{array}{c} \psi_{0}^{+}(x)\xi^{+}(y) \\ \psi_{0}^{-}(x)\xi^{-}(y) \end{array}\right].\end{aligned}$$ And, Eq. (\[3-14\]) can be split as $$\begin{aligned} \label{3-17} \pm \frac{d}{2}\partial_{y}\sigma_{1}\xi^{\pm}(y)\mp \partial_{y}\xi^{\pm}(y)-\lambda e^{\sigma_{2}}f(y)\xi^{\pm}(y)=0.\end{aligned}$$ The zero-mode solutions for (\[3-17\]) are given by $$\begin{aligned} \xi^{+}(y)&=&c_{1}e^{-\lambda \int_{y} dy'e^{\sigma_{2}(y')}f(y')+\frac{d}{2}\sigma_{1}},\label{3-18}\\ \xi^{-}(y)&=&c_{2}e^{\lambda \int_{y} dy'e^{\sigma_{2}(y')}f(y')+\frac{d}{2}\sigma_{1}}.\label{3-19}\end{aligned}$$ Therefore, by specifying the function $f(y)$, and the warp factors $\sigma_{1}$ and $\sigma_{2}$, the localization discussion can be performed. As we are interested in verify the consistency conditions (\[3-09\]), let us calculate the energy-momentum tensor for these zero-mode solutions. From Lagrangian (\[3-10\]), we get the components $$\begin{aligned} \label{3-20} \mathcal{T}^{m}_{\mu\nu}(x,y)=e^{\sigma_{1}(y)}\xi^{2}_{0}\left[\Bar{\psi}\hat{\Gamma}_{(\mu}\hat{D}_{\nu)}\psi+\hat{g}_{\mu\nu}(x)\hat{L}_{0}^{m}(x)\right].\end{aligned}$$ Then, the first consistency condition in (\[3-09\]) is satisfied if $e^{\sigma_{1}(y)}\xi^{2}_{0}(y)$ is a constant quantity. By using the solutions (\[3-18\]) and (\[3-19\]), we get $$\begin{aligned} \label{3-21} \pm 2\lambda \int_{y} dy'e^{\sigma_{2}(y')}f(y')+(d+1)\sigma_{1}(y)=\kappa_{\pm},\ \mbox{or}\ \pm 2\lambda e^{\sigma_{2}(y)}f(y)+(d+1)\sigma_{1}'(y)=0.\end{aligned}$$ Here, $\kappa_{\pm}$ are constants, and prime is the derivative with respect to the extra dimension. These relations in (\[3-21\]) should be valid for any value of $y$. Note that, already for the free case ($\lambda=0$), consistency cannot be obtained. In fact, we should have $\sigma'(y)=0$, and this is not satisfied for any non-factorizable braneworld model. For models with $\lambda\neq 0$, we conclude that the spinor field localization can be made consistent with Einstein’s equations only if $f(y)\propto e^{-\sigma_{2}(y)}\sigma_{1}'(y)$. When we analyze some models in the literature, the mechanism presented in Refs. [@Barut; @Oda04], for RS-II model [@RS2], can be made consistent. However, for thick brane models, the localization mechanisms presented in Refs. [@Koley01; @Kehagias; @Melfo; @Castro; @Mendes; @Ringeval] are not consistent, in such way those mechanisms should be reviewed. Otherwise, the braneworld metric should be modified to take into account the presence of the spinor field. Applications - Vector field {#ASec-5} =========================== Now, we will discuss the consistency of the vector field localization. This subject was already treated early in the literature for co-dimension 1 delta-like models [@Duff]. In such reference, the authors show that the free vector field localization (zero-mode) is not consistent with Einstein’s equations. Here, let us discuss this issue for the free field in an arbitrary braneworld and also for some localization mechanisms commonly used in the literature. Free vector field localization {#ASubsec-5-1} ------------------------------ To start the discussion, we will consider the free field in a brane model with metric given by (\[2-01\]). The Lagrangian for this case can be written as $$\begin{aligned} \label{3-22} \mathcal{L}^{m}(x,y)=-\frac{1}{4}\mathcal{F}_{MN}\mathcal{F}^{MN},\end{aligned}$$ with $\mathcal{F}_{MN}=\partial_{M}\mathcal{A}_{N}-\partial_{N}\mathcal{A}_{M}$. By calculating the equations of motion, we get $$\begin{aligned} \label{3-23} \partial_{M}\left[\sqrt{-g}\mathcal{F}^{MN}\right]=\partial_{\mu}\left[\sqrt{-g}\mathcal{F}^{\mu N}\right]+\partial_{k}\left[\sqrt{-g}\mathcal{F}^{kN}\right]=0.\end{aligned}$$ Here, we can propose $\mathcal{A}_{N}(x,y)=\left(\mathcal{A}^{T}_{\mu}+\partial_{\mu}\theta,\mathcal{B}_{k}\right)$ with $\partial^{\mu}\mathcal{A}^{T}_{\mu}=0$, and thus the equation (\[3-23\]) can be split, for the components $\mathcal{A}^{T}_{\mu}$, as $$\begin{aligned} \label{3-24} \partial_{\mu}\left[\sqrt{-g}g^{\mu\nu}g^{\lambda\rho}\mathcal{F}^{T}_{\nu\lambda}\right]+\partial_{k}\left[\sqrt{-g}g^{kj}g^{\nu\rho}\partial_{j}\mathcal{A}^{T}_{\nu}\right]=0.\end{aligned}$$ The equations of motion for the fields $\theta$ and $\mathcal{B}_{k}$ will not be important to our discussion, thus they will not be written here. From Eq. (\[3-24\]) and by proposing $\mathcal{A}^{T}_{\mu}(x,y)=\hat{A}^{T}_{\mu}(x)\xi(y)$, we can get $$\begin{aligned} \frac{1}{\sqrt{-\hat{g}}}\partial_{\mu}\left[\sqrt{-\hat{g}}\hat{F}_{T}^{\mu\rho}(x)\right]=m^{2}\hat{A}_{T}^{\rho}(x),\label{3-25}\\ -\frac{e^{-(d-4)\sigma(y)}}{\sqrt{\Bar{g}}}\partial_{k}\left[\sqrt{\Bar{g}}\Bar{g}^{kj}e^{(d-2)\sigma(y)}\partial_{j}\xi(y)\right]=m^{2}\xi(y).\label{3-26}\end{aligned}$$ Now, as well as we did for the scalar and spinor fields, the equation (\[3-26\]) can be solved for $m^{2}=0$ and, with these zero-mode solutions, the vector field localization can be studied. Equation (\[3-26\]) has a constant, $\xi_{0}(y)=c_{0}$, and also a non-constant solution for the zero-mode. This last one is closely related to the braneworld and its specific form is model dependent. Fortunately, in most cases, the constant zero-mode solution is the only one that can be confined [@Oda03; @Oda04; @Midodashvili; @Freitas]. We are interested in studying the consistency of the confinement, therefore let us obtain the energy-momentum tensor for the Lagrangian (\[3-22\]). By doing this, we get $$\begin{aligned} \label{3-27} \mathcal{T}^{m}_{MN}(x,y)=\mathcal{F}_{MP}(x,y)\mathcal{F}_{N}^{\ \ P}(x,y)+g_{MN}\mathcal{L}^{m}(x,y).\end{aligned}$$ And, for the zero-mode, the components $\mathcal{T}^{m}_{\mu\nu}$ can be written as $$\begin{aligned} \label{3-28} \mathcal{T}^{m}_{\mu\nu}(x,y)=e^{-2\sigma(y)}\xi^{2}_{0}(y)\left[\hat{F}^{T}_{\mu\rho}(x)\hat{F}_{\nu}^{T\ \rho}(x)+\hat{g}_{\mu\nu}(x)\hat{L}_{0}^{m}(x)\right].\end{aligned}$$ From this, the consistency condition (\[2-14\]) is satisfied only if $e^{-2\sigma(y)}\xi^{2}_{0}(y)=const$. As a first result, we get that the constant zero-mode solution cannot satisfy this requirement, and this is co-dimension independent. Therefore, for all those models where the confinement is performed with such solution, the localization is not consistent. For $5$D models, the free gauge field cannot be confined because the $K$ integral in (\[2-03\]) is not finite [@Barut; @Kehagias]. In this way, the result obtained from (\[3-28\]) just confirms the non-localization of this field. However, most interesting results are obtained for co-dimension 2 and higher-dimensional models. In the literature, there are a large variety of models in these dimensional configurations where the metric (\[2-01\]) gets the particular shape [@Gregory; @Gherghetta; @Carlos; @Silva; @Oda04] $$\begin{aligned} \label{3-29} ds^{2}=e^{2\sigma(r)}\hat{g}_{\mu\nu}(x)dx^{\mu}dx^{\nu}+dr^{2}+e^{2\sigma_{2}(r)}d\Omega^{2}_{n-1}.\end{aligned}$$ Localization for free vector field in such scenarios was already studied [@Oda03; @Oda05; @Oda06; @Midodashvili; @Costa01; @Costa02]. Generally, for all these references, the constant zero-mode solution is confined given that the $K$ integral, in (\[2-03\]), is finite. However, by using our consistency conditions, we get that the localization of the free gauge field in such scenarios is not consistent. Therefore, even with a localized zero-mode, it is not possible to ignore back-reaction effects from the $U(1)$ gauge field on the backgorund geometry. Of course, there are still other braneworld models where the metric is not like that in (\[3-29\]), as those in Refs. [@Choudhury01; @Arkani02; @Flachi]. However, the conclusion for these cases is the same: the constant zero-mode solution is not consistent with Einstein’s equations. That way, the vector field localization really seems to need a mechanism to be consistent, and in the next section we will discuss some of them. Vector field localization through mechanisms {#ASubsec-5-2} -------------------------------------------- Now, we will review some mechanisms used to confine the abelian vector field. Again, the focus is to verify the consistency of localization procedure with Einstein’s equations. Let us first consider the co-dimension 1 braneworld models. In this dimensional configuration, the metric (\[2-01\]) can be written as $$\begin{aligned} \label{3-30} ds^{2}=e^{2\sigma(y)}\hat{g}_{\mu\nu}dx^{\mu}dx^{\nu}+e^{2\sigma_{2}(y)}dy^{2}.\end{aligned}$$ With this, we can discuss the consistency for a wide variety of models, whether with thin or thick brane, with or without internal structure.\ [i) Scalar field coupling]{}\ Let us start by considering a localization mechanism commonly used to confine the spin $1$ field on thick branes [@Kehagias; @Fu01; @Chumbes; @Cruz01]. In these models, the gauge field is coupled to some scalar function $G\left(y\right)$ through an action like the one given below $$\label{3-31} S=-\frac{1}{4}\int d^{d}xdy\sqrt{-g}G\left(y\right)\mathcal{F}_{MN}\mathcal{F}^{MN}.$$ The function $G\left(y\right)$ will be defined later for some known cases. By calculating the equation of motion and assuming the gauge $\mathcal{A}_{M}=(\mathcal{A}_{\mu},\mathcal{A}_{d+1}=0)$, we get the separated equations $$\begin{aligned} \frac{1}{\sqrt{-\hat{g}}}\partial_{\mu}\left[\sqrt{-\hat{g}}\hat{g}^{\mu\rho}\hat{g}^{\nu\lambda}\mathcal{F}_{\rho\lambda}\right]=m^{2}\hat{g}^{\nu\lambda}\hat{A}_{\lambda}(x),\label{3-32}\\ -\frac{e^{-(d-4)\sigma(y)-\sigma_{2}(y)}}{G\left(y\right)}\partial_{y}\left[e^{(d-2)\sigma(y)-\sigma_{2}(y)}G\left(y\right)\partial_{y}\xi(y)\right]=m^{2}\xi(y),\label{3-33}\end{aligned}$$ where, we already used the metric (\[3-30\]) and also $\mathcal{A}_{\mu}(x,y)=\hat{A}_{\mu}(x)\xi(y)$. Beyond this, the effective action on the brane can be written as $$\begin{aligned} \label{3-34} S=-\frac{1}{4}\int d^{d}x\sqrt{-\hat{g}(x)}\left[\hat{F}_{\mu\nu}(x)\hat{F}^{\mu\nu}(x)+2m^{2}\hat{A}_{\mu}(x)\hat{A}^{\mu}(x)\right]K.\end{aligned}$$ where, Neumann boundary conditions was used for $\xi(y)$, and the quantity $K$ is given by $$\begin{aligned} \label{3-35} K=\int dye^{(d-4)\sigma(y)+\sigma_{2}(y)}G\left(y\right)\xi^{2}(y).\end{aligned}$$ With this, by observing (\[3-34\]), a gauge field (massless mode) on the brane can be obtained by doing $m^{2}=0$. Moreover, by properly choosing the function $G(y)$, this massless mode can be confined. In Refs. [@Kehagias; @Fu01] this function is chosen in the form $G(y)=e^{\lambda\pi(y)}$, where $\pi(y)$ is a scalar field, namely the dilaton. From this, the zero-mode localization is obtained for some values $\lambda$. In Ref. [@Chumbes] the authors get a general form for $G(y)$, that should be valid for an arbitrary thick brane with $\sigma_{2}(y)=0$. In another direction, Ref. [@Zhao02] proposes $G(y)$ as a function of the Ricci scalar, and also for this case the localization can be obtained. Let us turn back to the consistency conditions (\[2-14\]) and (\[2-15\]). From action (\[3-35\]), the energy-momentum tensor can be written as $$\begin{aligned} \label{3-36} T_{MN}^{(\mbox{mat})}=G(y)\left[\mathcal{F}_{MP}\mathcal{F}_{N}^{\ \ P}-\frac{1}{4}g_{MN}\mathcal{F}_{PQ}\mathcal{F}^{PQ}\right].\end{aligned}$$ Thus, by using the above configuration and considering only the zero-mode solution, we get $$\begin{aligned} \label{3-37} T_{\mu\nu}^{(\mbox{mat})}=G(y)e^{-2\sigma(y)}\xi_{0}^{2}(y)\left[\hat{F}_{\mu \rho}(x)\hat{F}_{\nu}^{\ \ \rho}(x)-\frac{1}{4}\hat{g}_{\mu\nu}(x)\hat{F}_{\mu\nu}(x)\hat{F}^{\mu\nu}(x)\right].\end{aligned}$$ where $\xi_{0}^{2}(y)$ is the zero-mode ($m^{2}=0$) solution of (\[3-33\]). From expression (\[3-37\]), we get that the consistency condition (\[2-14\]) is satisfied only if $G(y)e^{-2\sigma(y)}\xi_{0}^{2}(y)=const.$. That way, by using the constant solution for the massless mode $\xi_{0}(y)$, we conclude that $G(y)$ only can assume the specific form $G(y)=e^{2\sigma(y)}$. The same conclusion is obtained from the condition (\[2-15\]). That condition on the function $G(y)$ considerably restricts the allowed models for this type of coupling. For example, in Ref. [@Kehagias], the authors define $G(y)=e^{\frac{\lambda}{2}\sigma(y)}$. In this case, we get that the coupling constant $\lambda$ must be defined as $\lambda=4$, for the localization to be well-defined. In Ref. [@Fu01] is used $G(y)=e^{\tau \sqrt{3r}\sigma(y)}$, and the localization is obtained for $\tau\geq -\sqrt{\frac{r}{3}}$, with $0<r<1$; or $\tau>-\sqrt{\frac{1}{3r}}$, with $r>1$. By using our consistence condition, we get that $\tau=\frac{2}{\sqrt{3r}}$. Therefore, for both references localization can be done consistently. There is a very interesting reference, namely [@Chumbes], where function $G(y)$ is defined as $G(y)=G(\phi)\propto \frac{\partial W^{2s}\left(\phi\right)}{\partial \phi}$, with $\underline{s}$ a constant, and $W(\phi)$ the superpotential related to the scalar field $\phi(y)$ which generate the braneworld. The authors show that for a brane model generated by scalar field with Sine-Gordon potential [@Gremm01], i.e., $$\begin{aligned} \label{3-38} W(\phi)=3bc\sin\left(\sqrt{\frac{2}{3b}}\phi\right),\ \ \mbox{with}\ \ \phi(y)=\sqrt{6b}\arctan\left(cy\right),\end{aligned}$$ function $G(\phi)$ is given by $G(y)=\mbox{sech}^{2s}\left(2cy\right)$ and localization can be obtained. When we compare this with our result, which for the specific model [@Gremm01] stays $G(y)=e^{2\sigma(y)}=\mbox{sech}^{2b}\left(2cy\right)$, the constant $s$ should be $s=b$. On the other hand, for the brane model [@Kehagias], where $$\begin{aligned} \label{3-39} W(\phi)=av\left(\phi-\frac{\phi^{3}}{3v^{2}}\right),\ \ \mbox{with}\ \ \phi(y)=v\arctan\left(ay\right),\end{aligned}$$ the Ref. [@Chumbes] obtain $G(y)=\mbox{sech}^{4s}\left(ay\right)$. Now, by using our result $G(y)=e^{2\sigma(y)}=\mbox{sech}^{4b}\left(ay\right)e^{-b\tanh^{2}(ay)}$, we see that functions $G(y)$ does not match, and the superpotential (\[3-39\]) does not allow a consistent localization. Similar result can be obtained for deformed thick brane models [@Bazeia05]. This conclusion indicates that $G(y)=G(\phi)\propto \frac{\partial W^{2s}\left(\phi\right)}{\partial \phi}$ have not a general validity as localization mechanism, i.e., it does not work for any braneworld model. Therefore, except for (\[3-38\]), the function $G(y)=G(\phi)\propto \frac{\partial W^{2s}\left(\phi\right)}{\partial \phi}$ does not provide a consistent localization for the gauge field (zero-mode). Another interesting model is presented in Ref. [@Zhao02], where $G(y)$ is function of the Ricci scalar, namely, $$\label{3-31a} S=-\frac{1}{4}\int d^{4}xdy\sqrt{-g}G\left(R\right)\mathcal{F}_{MN}\mathcal{F}^{MN}.$$ The authors argue that, if $G(R)$ is a continuous function, the zero-mode localization of the vector field is determined by the behavior of $G(R)$ when $y\to\infty$. They get that this function must be, asymptotically, something like $G(R_{\infty})\propto \|y\|^{-p}$, with $p$ a positive value. Considering the AdS feature of the space, the authors show that the warp factor must be, asymptotically, in the shape $$\begin{aligned} e^{2\sigma(y\to\pm\infty)}\to \|y\|^{-2}.\end{aligned}$$ Therefore, the consistency condition obtained by us for models like (\[3-31\]), i.e., $G(y)=e^{2\sigma(y)}$, can be satisfied for $G(R_{\infty})\propto \|y\|^{-p}$, if $p=2$. However, we cannot say that this is valid for another range of the variable $y$. Moreover, there seems to be a contradiction in the arguments used by the authors themselves. They propose a localization mechanism in an asymptotically AdS space-time, thus, the Ricci scalar is, in that limit, $R(\|y\|\to\infty)\propto -C_{R}$ (constant). Therefore, $G(R)$ should go to a constant value at that limit, and the localization cannot be reached. Anyway, the requirement of $G(R)=e^{2\sigma(y)}$ does not seem so easy to meet for an arbitrary model. Other interesting points can also be discussed. For example, for models like [@Kehagias; @Fu01], we obtain that the coupling parameter ($\lambda$ or $\tau$) are not free, they must be fixed by consistency reasons. In this way, the analysis performed in Refs. [@Landim01; @Landim02; @Landim03] by research resonances of the gauge field with action like (\[3-35\]) should be reevaluated. Since, there is no freedom in choosing the parameters $\lambda$ or $\tau$, used to plot the graphics in those references.\ [ii) G-N localization mechanism]{}\ Now, let us verify the non-covariant mechanism proposed by K. Ghoroku and A. Nakamura (G-N) in Ref. [@Ghoroku]. In this paper is used a metric like (\[3-30\]) with the warp factors given by $\sigma_{2}(y)=\sigma(y)=-\ln\left(1+k\|y\|\right)$. The Lagrangian for the vector field with G-N mechanism is written as $$\label{3-40} \mathcal{L}^{m}=-\frac{1}{4}\mathcal{F}_{MN}\mathcal{F}^{MN}-\frac{1}{2}\left[M^{2}+c\delta(y)\right]\mathcal{A}_{M}\mathcal{A}^{M}.$$ This model, although not being gauge invariant or even covariant, the effective theory on the brane has the desired features: a massless vector field with gauge symmetry. After some steps like those performed in the previous case, we can get an EOM for the effective vector field $\mathcal{A}^{T}_{\mu}$, and by proposing the separation of variables $\mathcal{A}^{T}_{\mu}=\hat{A}^{T}_{\mu}(x)\xi(y)$, the localization discussion can be carried out. By doing this, the zero-mode solution can be obtained with the ansatz $$\begin{aligned} \label{3-41} \xi_{0}(y)=c_{0}e^{a\sigma(y)},\ \mbox{where}\ a=-\frac{c}{2k}=\frac{1}{k}\left(\sqrt{k^{2}+M^{2}}-k\right),\end{aligned}$$ and it will be localized if $a>0$. With this, we can analyze the consistency conditions (\[2-14\]) and (\[2-15\]) for the energy-momentum tensor. From the Lagrangian (\[3-40\]), we get for the zero-mode $$\begin{aligned} \label{3-42} T_{\mu\nu}^{(\mbox{mat})}(x,y)=e^{-2\sigma}\xi^{2}_{0}\left[\hat{F}^{T}_{\mu\rho}(x)\hat{F}_{\nu}^{T\ \rho}(x)-\frac{1}{4}\hat{g}_{\mu\nu}(x)\hat{F}^{T}_{\rho\alpha}(x)\hat{F}_{T}^{\rho\alpha}(x)\right].\end{aligned}$$ Therefore, by using the zero-mode solution, we get that $e^{-2\sigma}\xi^{2}_{0}=e^{2(a-1)\sigma}=const.$ and the consistency condition (\[2-14\]) will be satisfied when $a=1$. This value of $a$ fix all parameters in the Lagrangian (\[3-40\]) in the following shape, $c=-2k$ and $\|M\|=\sqrt{3}\|k\|$. By a similar analysis, we show that consistency condition (\[2-15\]) gives us the same result and this localization mechanism can be performed consistently.\ [iii) Non-minimal coupling with gravity]{}\ Finally, let us discuss the localization mechanism proposed in Refs. [@Flachi; @Zhao; @Alencar02]. In these works, the G-N mechanism is used as a motivation to propose the vector field coupling with gravity through the scalar and the Ricci tensor. The action for this case is given by $$\label{3-43} S=\frac{1}{2}\int d^{d}xdy\sqrt{-g}\left[-\frac{1}{2}\mathcal{F}_{MN}\mathcal{F}^{MN}+\lambda_{1}R\mathcal{A}_{M}\mathcal{A}^{M}+\lambda_{2}R_{MN}\mathcal{A}^{M}\mathcal{A}^{N}\right].$$ Where $\underline{d}$ is the brane dimension, $R$ and $R_{MN}$ are the scalar and the Ricci tensor, respectively. By proposing $\mathcal{A}_{M}=\left(\mathcal{A}^{T}_{\mu}+\partial_{\mu}\theta,\mathcal{B}\right)$, and after some steps, it is possible to obtain an EOM for the transverse field $\mathcal{A}^{T}_{\mu}$. In order that, by using the separation of variables $\mathcal{A}^{T}_{\mu}(x,y)=\hat{A}^{T}_{\mu}(x)\xi(y)$, we obtain a zero-mode solution for $\xi(y)$ given by $$\begin{aligned} \label{3-44} \xi_{0}(y)=e^{a\sigma(y)},\ \mbox{with}\ a=-2\lambda_{1}\left(D-1\right)-\lambda_{2}.\end{aligned}$$ Beyond the additional conditions $$\begin{aligned} \left(\frac{D-4}{2}+a\right)^{2}&=&\frac{\left(D-4\right)^{2}}{4}-\left(\lambda_{1} \left(D-1\right)+\lambda_{2}\right)(D-2),\label{3-45}\\ \frac{D-4}{2}+a &>&\frac{1}{2},\label{3-46}\end{aligned}$$ where $D=d+1$. The conditions, (\[3-44\]) and (\[3-45\]), are required for a zero-mode solution to exist, and the condition (\[3-46\]) should be satisfied for that the solution $\xi_{0}$ be confined on the brane. By doing a similar analysis like that in (\[3-42\]), we conclude that the consistency conditions (\[2-14\]) and (\[2-15\]) can be satisfied for $a=1$. Therefore, this localization mechanism can provide a confinement of the vector field consistent. By eliminating higher-order terms, the brane components of energy-momentum tensor for the action (\[3-43\]) can be written, for the zero-mode, as $$\begin{aligned} \label{3-47} T_{\mu\nu}^{(\mbox{mat})}(x,y)&=&e^{-2\sigma}\xi^{2}_{0}\left[\hat{F}^{T}_{\mu\rho}(x)\hat{F}_{\nu}^{T\ \rho}(x)-\frac{1}{4}\hat{g}_{\mu\nu}(x)\hat{F}^{T}_{\rho\alpha}(x)\hat{F}_{T}^{\rho\alpha}(x)\right]\nonumber\\ &=&e^{2(a-1)\sigma}\left[\hat{F}^{T}_{\mu\rho}(x)\hat{F}_{\nu}^{T\ \rho}(x)-\frac{1}{4}\hat{g}_{\mu\nu}(x)\hat{F}^{T}_{\rho\alpha}(x)\hat{F}_{T}^{\rho\alpha}(x)\right].\end{aligned}$$ Thus, we see that the condition (\[2-14\]) is satisfied when $a=1$. When we use $a=1$, the localization condition (\[3-46\]) can be written as $d-2>0$, and it is always satisfied for models with $d\geq 4$. By doing $a=1$ in Eqs. (\[3-44\]) and (\[3-45\]), we get $$\begin{aligned} \lambda_{1}=-\frac{1}{(D-2)(D-1)}.\label{3-48}\end{aligned}$$ This result shows that vector field localization, by using the mechanism (\[3-43\]), is consistent with Einstein equation only when both interaction terms are switch on simultaneously. Beyond this, the two parameters $\lambda_{1}$ and $\lambda_{2}$ are completely fixed by consistency reasons. This result allows us to comment briefly that one presented in Ref. [@Sui], where the authors plot some graphics of the potential and the relative probability for various values of $a$. As we obtained from the consistency conditions, the parameters are fixed and such freedom for the parameter $a$ does not exist. In fact, the authors argue that massive resonant modes can exist if $a>3$, which, when compared with our results $a=1$, shows that resonant modes cannot exist.\ [iv) Localization in co-dimension $2$ or higher models]{}\ Generally, for most of the models in co-dimendion $2$ or higher, the free $U(1)$ gauge field is already naturally confined just by minimum couplings with gravity [@Oda03; @Oda04; @Oda06; @Midodashvili; @Costa02; @Torrealba; @Giovannini; @Fu02]. Thus, there are not many localization mechanisms for this field in those dimensional configurations. However, as we saw in subsection (\[ASubsec-5-1\]), the free field case is already not consistent with Einstein’s equations, so some localization mechanism really seems to be necessary. In Ref. [@Flachi], the vector field is confined in co-dimension 2 intersecting delta-like branes by proposing a mechanism like that in equation (\[3-43\]). In Ref. [@Freitas], this study is performed for a generic model with arbitrary co-dimension embedded in asymptotically AdS space. For both cases, the results are similar to these obtained early in item (iii). In other words, the consistency with Einstein’s equations is obtained just when both interaction terms are switch on simultaneously. In fact, this conclusion is co-dimension independent for this localization mechanism. Let the action for the vector field in a generic model with the localization mechanism like (\[3-43\]) be given by $$\label{3-51} S=\frac{1}{2}\int d^{d}xdy^{n}\sqrt{-g}\left[-\frac{1}{2}\mathcal{F}_{MN}\mathcal{F}^{MN}+\lambda_{1}R\mathcal{A}_{M}\mathcal{A}^{M}+\lambda_{2}R_{MN}\mathcal{A}^{M}\mathcal{A}^{N}\right],$$ where, $\underline{d}$ is the brane dimension and $\underline{n}$ is the co-dimension. By performing some steps like that in the item (iii), we can get a zero-mode solution given by $$\begin{aligned} \label{3-52} \xi_{0}(y)=e^{a\sigma(y)},\ \mbox{with}\ a=-2\lambda_{1}\left(D-1\right)-\lambda_{2}.\end{aligned}$$ With the additional conditions $$\begin{aligned} \left(\frac{D-4}{2}+a\right)^{2}&=&\frac{\left(D-4\right)^{2}}{4}-\left(\lambda_{1} \left(D-1\right)+\lambda_{2}\right)(D-2),\label{3-53}\\ \frac{D-4}{2}+a &>&\frac{n}{2},\label{3-54}\end{aligned}$$ where $D=d+n$. Again, to verify the consistency with Einstein’s equations, we need the energy-momentum tensor. By eliminating higher order terms, the brane components of energy-momentum tensor, for the action (\[3-51\]), can be written for the zero-mode as, $$\begin{aligned} \label{3-55} T_{\mu\nu}^{(\mbox{mat})}(x,y)&=&e^{-2\sigma}\xi^{2}_{0}\left[\hat{F}^{T}_{\mu\rho}(x)\hat{F}_{\nu}^{T\ \rho}(x)-\frac{1}{4}\hat{g}_{\mu\nu}(x)\hat{F}^{T}_{\rho\alpha}(x)\hat{F}_{T}^{\rho\alpha}(x)\right]\nonumber\\ &=&e^{2(a-1)\sigma}\left[\hat{F}^{T}_{\mu\rho}(x)\hat{F}_{\nu}^{T\ \rho}(x)-\frac{1}{4}\hat{g}_{\mu\nu}(x)\hat{F}^{T}_{\rho\alpha}(x)\hat{F}_{T}^{\rho\alpha}(x)\right].\end{aligned}$$ Thus, we see that the condition (\[2-14\]) is satisfied when $a=1$. If we use $a=1$, the localization condition (\[3-54\]) can be written as $d-2>0$, which is always satisfied for models with $d\geq 4$. By doing $a=1$ in Eqs. (\[3-52\]) and (\[3-53\]), we get $$\begin{aligned} \lambda_{1}=-\frac{1}{(D-2)(D-1)}.\label{3-56}\end{aligned}$$ Therefore, again, we get the same conclusion that presented in the item (iii): the localization mechanism (\[3-51\]) is consistent with Einstein’s equations only when both interaction terms are switch on simultaneously. In this way, any back-reaction effect from the vector field on the background metric can be eliminated, at least for the zero-mode. Final Remarks {#ASec-6} ============= In this work, we discussed the consistency of fields localization in braneworld models. By studying Einstein’s equations in the presence of matter fields, we obtained the constraints (\[2-14\]) and (\[2-15\]) for the energy-momentum tensor that should be valid for any brane model. Such constraints are a consequence of the assumption used in fields localization, where a confined matter field does not modify the bulk metric. In this way, the localization procedure used in the literature will be consistent only if such conditions are satisfied. We applied these consistency conditions for some cases, namely, the spin $0$, $\frac{1}{2}$ and $1$ Standard Model fields, with and without localization mechanisms. For the scalar field, as discussed in many references [@Barut; @Kehagias; @Oda02; @Oda03; @Liu01; @Liu02; @Fu01; @Midodashvili], there is a constant zero-mode solution that can be localized. By using this confined constant solution, we showed in Eqs. (\[3-06\]) and (\[3-07\]) that the energy-momentum tensor for the scalar field satisfies the consistency conditions (\[2-14\]) and (\[2-15\]). Therefore, the scalar field localization (zero-mode) is consistent with Einstein’s equations. After, we apply those conditions for the spin $\frac{1}{2}$ field in co-dimension 1 models, with a Yukawa-like interaction given by $\mathcal{L}^{m}_{in}=\lambda f(y)\Bar{\Psi}\Psi$. Also, for this field, there is a variety of models with this kind of coupling [@Koley01; @Kehagias; @Melfo; @Castro; @Mendes; @Ringeval]. For each of these cases, a different Yukawa interaction is proposed and the spinor zero-mode localization, actually one of the chiralities, is obtained for some properly condition. About consistency conditions (\[2-14\]) and (\[2-15\]), we obtained the energy-momentum tensor (\[3-20\]). And from this, the spinor field localization is consistent with Einstein’s equations only if $f(y)\propto e^{-\sigma_{2}(y)}\sigma_{1}'(y)$, with $\sigma_{1}$ and $\sigma_{2}$ the warp factors in (\[3-08\]). As discussed in section (\[ASec-4\]), this relation eliminates the freedom to choose the function $f(y)$. With this, the Yukawa interaction used in Refs. [@Barut; @Oda04], for RS-II type braneworlds, is consistent with the Einstein’s equations by properly choose the interacting parameter $\lambda$. However, for those functions $f(y)$ used in thick brane models like [@Koley01; @Kehagias; @Melfo; @Castro; @Mendes; @Ringeval], the localization is not consistent and it should be reviewed. We must to stress that the analysis performed for the spinor field considered only Yukawa-like interactions, and also only in co-dimension 1 models. There are still other localization mechanisms and other dimensional configurations where this analysis could be carried out as for exemple [@GAlencar]. Finally, we verified the consistency condition for the vector field. As discussed widely in the literature [@Barut; @Oda03; @Oda04; @Midodashvili; @Freitas], the free gauge field (zero-mode) cannot be confined in $5$D, however, for some higher co-dimension models it can be localized. In section (\[ASubsec-5-1\]), we obtained the energy-momentum tensor (\[3-27\]) for the gauge field and, from this, the consistency conditions were analyzed. As a general result, the conditions (\[2-14\]) and (\[2-15\]) are consistent for the zero-mode vector field only if $e^{-2\sigma(y)}\xi^{2}_{0}(y)$ is a constant. However, as discussed in that section, there is not zero-mode solution $\xi_{0}$ confined that satisfies the above condition. Such result is independent of the braneworld model or the number of extra dimension, thus the localization of this field is not consistent with Einstein’s equations and a mechanism to confine it really seems necessary. In this direction, we analyzed the consistency of some localization mechanisms in section (\[ASubsec-5-2\]). For example, that mechanism proposed in Refs. [@Kehagias; @Fu01; @Chumbes; @Zhao02], where the action for the gauge field is given by something like (\[3-31\]). For these kinds of coupling, there is a zero-mode constant solution that can always be localized when $G(y)$ is like Gaussian. By using the consistency conditions, the localization of $\xi_{0}=c_{0}$ is consistent with Einstein’s equations just if $G(y)=e^{2\sigma(y)}$ in Eq. (\[3-37\]). In this way, the Gaussian feature is confirmed, however, such expression does not present any free coupling parameter. Beyond this, that function $G(y)$ eliminates some mechanism proposed in the literature, like that in Ref. [@Zhao02], where $G(y)=G(R)$ is a function of Ricci scalar. Other interesting points can still be discussed here. As there is not a free parameter in $G(y)=e^{2\sigma(y)}$, the analysis performed in Refs. [@Landim01; @Landim02; @Landim03] about resonant modes of the gauge field with action given by (\[3-31\]), for models like [@Kehagias; @Fu01], should be reevaluated. Since some results are obtained by using a free coupling parameter, which, by our analysis, does not exist. We also discussed the non-covariant mechanism proposed in Ref. [@Ghoroku]. For this case, the zero-mode solution for the gauge field sector is given by $\xi_{0}(y)=e^{a\sigma(y)}$, and the consistency conditions (\[2-14\]) and (\[2-15\]) were satisfied when $a=1$. With this value of $a$, all parameters in (\[3-40\]) were fixed, namely, $c=-2k$ and $\|M\|=\sqrt{3}\|k\|$. In Ref. [@Ghoroku] is also discussed the localization of the scalar component, and the authors conclude that both sectors cannot be confined simultaneously. Moreover, the theory does not indicate what these sectors should be confined. Maybe, the consistency condition could be used to solve this, but we did not perform such study here. Inspired by this mechanism, we analyzed the localization mechanism proposed in Refs. [@Zhao; @Alencar02]. And by starting from (\[3-43\]), the consistency with Einstein’s equations was obtained only if both interaction terms are present. And just like in the previous cases, the parameters $\lambda_{1}$ and $\lambda_{2}$ in (\[3-43\]) were completely fixed by consistency reasons. Beyond all these co-dimension 1, higher co-dimension models were investigated. And also for this dimensional configuration, by using the localization mechanism (\[3-51\]), the consistency with Einstein’s equations is possible only if both interaction terms are switch on simultaneously. In this way, we believe that a comprehensive analysis was performed about the consistency of fields localization, and such study can be used as a guide for building new confining mechanisms. Acknowledgment {#acknowledgment .unnumbered} ============== The authors would like to thank Alexandra Elbakyan and sci-hub, for removing all barriers in the way of science. We acknowledge the financial support provided by the Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq) and Fundação Cearense de Apoio ao Desenvolvimento Científico e Tecnológico (FUNCAP) through PRONEM PNE-0112-00085.01.00/16. [00]{} L. Randall and R. Sundrum, [A Large mass hierarchy from a small extra dimension]{}, [Phys. Rev. Lett. 83 (1999) 3370.](https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.83.3370) [\[hep-ph/9905221\]](https://arxiv.org/abs/hep-ph/9905221) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/499284) L. Randall and R. Sundrum, [An Alternative to compactification]{}, [Phys. Rev. Lett. 83 (1999) 4690.](https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.83.4690) [\[hep-th/9906064\]](https://arxiv.org/abs/hep-th/9906064) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/501443) M. Gremm, [Four-dimensional gravity on a thick domain wall]{}, [Phys. Lett. B. 478 (2000) 434.](https://www.sciencedirect.com/science/article/abs/pii/S0370269300003038?via%3Dihub) [\[hep-th/9912060\]](https://arxiv.org/abs/hep-th/9912060) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/511276) D. Bazeia, A. R. Gomes, L. Losano and R. Menezes, [Braneworld Models of Scalar Fields with Generalized Dynamics]{}, [Phys. Lett. B 671, (2009) 402-410.](https://doi.org/10.1016/j.physletb.2008.12.039) [\[hep-th/0808.1815\]](http://arxiv.org/abs/arXiv:0808.1815) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/793014) D. Bazeia and A. R. Gomes, [Bloch brane]{}, [JHEP. 0405 (2004) 012.](https://iopscience.iop.org/article/10.1088/1126-6708/2004/05/012/meta) [\[hep-th/0403141\]](https://arxiv.org/abs/hep-th/0403141) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/646356) D. Bazeia, C. Furtado and A. R. Gomes, [Brane structure from scalar field in warped space-time]{}, [JCAP. 0402 (2004) 002.](https://doi.org/10.1088/1475-7516/2004/02/002) [\[hep-th/0308034\]](http://arxiv.org/abs/hep-th/0308034) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/624977) N. Barbosa-Cendejas and A. Herrera-Aguilar, [Localization of $4$D gravity on pure geometrical thick branes]{}, [Phys. Rev. D 73, (2006) 084022.](https://doi.org/10.1103/PhysRevD.73.084022) [Erratum: Phys.Rev. D77 (2008) 049901.](https://doi.org/10.1103/PhysRevD.77.049901) [\[hep-th/0603184\]](http://arxiv.org/abs/hep-th/0603184) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/713093?ln=pt) P. Singh and N. Dadhich, [Localization of gravity in brane world cosmologies]{}, [Mod. Phys. Lett. A 18, (2003) 983-992.](https://doi.org/10.1142/S0217732303009757) [\[hep-th/0204190\]](https://arxiv.org/abs/hep-th/0204190) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/585769) J. M. Hoff da Silva and M. Dias, [Five Dimensional $f(R)$ Branewointeractionrld Models]{}, [Phys. Rev. D 84, (2011) 066011.](https://doi.org/10.1103/PhysRevD.84.066011) [\[hep-th/1107.2017\]](http://arxiv.org/abs/arXiv:1107.2017) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/917641) V. Dzhunushaliev, V. Folomeev and M. Minamitsuji, [Thick brane solutions]{}, [Rept. Prog. Phys. 73, (2010).](https://iopscience.iop.org/article/10.1088/0034-4885/73/6/066901) [\[gr-qc/0904.1775\]](https://arxiv.org/abs/0904.1775) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/535852) R. Gregory, [Nonsingular Global String Compactifications]{}, [Phys. Rev. Lett. 84 (2000) p. 2564-7.](https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.84.2564) [\[hep-th/9911015\]](https://arxiv.org/abs/hep-th/9911015) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/509535) T. Gherghetta and M. Shaposhnikov, [Localizing Gravity on a Stringlike Defect in Six Dimensions]{}, [Phys. Rev. Lett., v. 85 (2000).](https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.85.240) [\[hep-th/000401\]](https://arxiv.org/abs/hep-th/0004014) M. Giovannini, H. Meyer and M. E. Shaposhnikov, [Warped compactification on Abelian vortex in six-dimensions]{}, [Nucl. Phys. B 619 (2001) 615-645.](https://doi.org/10.1016/S0550-3213(01)00520-X) [\[hep-th/0104118\]](http://arxiv.org/abs/hep-th/0104118) [\[[IN]{}SPIRE\]](https://inspirehep.net/record/555412) B. Carlos and J. M. Moreno, [A cigar-like Universe]{}, [JHEP, v. 11 (2003).](https://iopscience.iop.org/article/10.1088/1126-6708/2003/11/040/meta) [\[hep-th/030925\]](https://arxiv.org/abs/hep-th/0309259) J. E. G. Silva, V. Santos and C. A. S. Almeida, [Gravity localization in a string-cigar braneworld]{}, [Class. Quant. Grav. 30 (2013) 025005.](https://iopscience.iop.org/article/10.1088/0264-9381/30/2/025005/meta) [\[hep-th/1208.2364\]](https://arxiv.org/abs/1208.2364) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/1127317) G. Kofinas, [On braneworld cosmologies from six dimensions, and absence thereof]{}, [Phys. Lett. B 633, (2006) 141-148.](https://doi.org/10.1016/j.physletb.2005.11.064) [\[hep-th/0506035\]](http://arxiv.org/abs/hep-th/0506035) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/684131) J. M. Cline, J. Descheneau, M. Giovannini1 and J. Vinet, [Cosmology of co-dimension-two braneworlds]{}, [JHEP. 0306 (2003) 048.](https://iopscience.iop.org/article/10.1088/1126-6708/2003/06/048/meta) [\[hep-th/0304147\]](https://arxiv.org/abs/hep-th/0304147) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/617181) J. E. Kim, B. Kyae and H. M. Lee, [Localized gravity and mass hierarchy in D = 6 with a Gauss-Bonnet term]{}, [Phys. Rev. D. 64 (2001).](https://journals.aps.org/prd/abstract/10.1103/PhysRevD.64.065011) [\[hep-th/0104150\]](https://arxiv.org/abs/hep-th/0104150) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/555592) R. Koley and S. Kar, [Braneworlds in six dimensions: New models with bulk scalars]{}, [Class. Quant. Grav. 24, (2007) 79-94.](https://iopscience.iop.org/article/10.1088/0264-9381/24/1/004) [\[hep-th/0611074\]](http://arxiv.org/abs/hep-th/0611074) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/731121) D. Choudhury and S. Sengupta, [Generalization of the Randall-Sundrum warped braneworld model to higher dimensions]{}, [Phys. Rev. D, v. 76, (2007) 064030.](https://journals.aps.org/prd/abstract/10.1103/PhysRevD.76.064030) N. Arkani-Hamed, S. Dimopoulos, G. R. Dvali and N. Kaloper, [Infinitely large new dimensions]{}, [Phys. Rev. Lett. 84 (2000) 586.](https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.84.586) [\[hep-th/9907209\]](https://arxiv.org/abs/hep-th/9907209) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/504757) I. Oda, [Locally Localized Gravity Models in Higher Dimensions]{}, [Phys. Rev. D 64 (2001).](http://dx.doi.org/10.1103/PhysRevD.64.026002) [\[hep-th/0102147\]](https://arxiv.org/abs/hep-th/0102147) [\[[IN]{}SPIRE\]](https://inspirehep.net/record/553376) D. Bazeia, J. Menezes and R. Menezes, [New global defect structures]{}, [Phys. Rev. Lett. 91 (2003) 241601.](https://journals.aps.org/prl/abstract/10.1103/PhysRevLett.91.241601) [\[hep-th/0305234\]](https://arxiv.org/abs/hep-th/0305234) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/619637) I. Navarro and J. Santiago, [Higher co-dimension brane worlds from intersecting branes]{}, [JHEP 0404 (2004) 062.](https://doi.org/10.1088/1126-6708/2004/04/062) [\[hep-th/0402204\]](http://arxiv.org/abs/hep-th/0402204) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/645146) C. Csaki, J. Erlich, T. J. Hollowood and Y. Shirman, [Universal aspects of gravity localized on thick branes]{}, [Nucl. Phys. B 581 (2000) 309.](https://doi.org/10.1016/S0550-3213(00)00271-6) [\[hep-th/0001033\]](https://arxiv.org/abs/hep-th/0001033) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/522837) G. W. Gibbons, R. Kallosh and A. D. Linde, [Brane world sum rules]{}, [JHEP, 0101 (2001) 022.](https://doi.org/10.1088/1126-6708/2001/01/022) [\[hep-th/0011225\]](http://arxiv.org/abs/hep-th/0011225) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/537397/?ln=pt) B. Bajc and G. Gabadadze, [Localization of matter and cosmological constant on a brane in anti-de Sitter space]{}, [Phys. Lett. B. 474 (2000) 282.](https://www.sciencedirect.com/science/article/abs/pii/S0370269300000551?via%3Dihub) [\[hep-th/9912232\]](https://arxiv.org/abs/hep-th/9912232) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/512450) I. Oda, [Localization of bulk fields on AdS$_{4}$ brane in AdS$_{5}$]{}, [Phys. Lett. B 508, (2001) 96-102.](https://doi.org/10.1016/S0370-2693(01)00376-8) [\[hep-th/0012013\]](http://arxiv.org/abs/hep-th/0012013) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/538001/?ln=pt) Y-X. Liu, X-H. Zhang, L-D. Zhang and Y-S. Duan, [Localization of Matters on Pure Geometrical Thick Branes]{}, [JHEP, 0802, (2008) 067.](https://doi.org/10.1088/1126-6708/2008/02/067) [\[hep-th/0708.0065\]](http://arxiv.org/abs/arXiv:0708.0065) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/757213) Y-X. Liu, C-E Fu, H. Guo, S-W. Wei and Z-H. Zhao, [Bulk Matters on a GRS-Inspired Braneworld]{}, [JCAP, 1012 (2010) 031.](https://doi.org/10.1088/1475-7516/2010/12/031) [\[hep-th/1002.2130\]](http://arxiv.org/abs/arXiv:1002.2130) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/845514) Y-X. Liu, H. Guo, C-E Fu and H-T. Li, [Localization of Bulk Matters on a Thick Anti-de Sitter Brane]{}, [Phys. Rev. D 84, (2011) 044033.](https://doi.org/10.1103/PhysRevD.84.044033) [\[hep-th/1101.4145\]](http://arxiv.org/abs/arXiv:1101.4145) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/885091) Y. X. Liu, Z. H. Zhao, S. W. Wei and Y. S. Duan, [Bulk Matters on Symmetric and Asymmetric de Sitter Thick Branes]{}, [JCAP, 0902 (2009) 003.](https://iopscience.iop.org/article/10.1088/1475-7516/2009/02/003/meta) [\[hep-th/0901.0782\]](https://arxiv.org/abs/0901.0782) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/810441) H. Guo, A. Herrera-Aguilar, Y. X. Liu, D. Malagon-Morejon, and R. R. Mora-Luna, [Localization of bulk matter fields, the hierarchy problem and corrections to Coulomb’s law on a pure de Sitter thick braneworld]{}, [Phys. Rev. D. 87 (2013) 095011.](https://journals.aps.org/prd/abstract/10.1103/PhysRevD.87.095011) [\[hep-th/1103.2430\]](https://arxiv.org/abs/1103.2430) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/892511) I. Oda, [Localization of Matters on a String-like Defect]{}, [Phys. Lett. B, v. 496, (2000), p. 113-121.](https://www.sciencedirect.com/science/article/abs/pii/S0370269300012843?via%3Dihub) [\[hep-th/0006203\]](https://arxiv.org/abs/hep-th/0006203) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/529325) I. Oda, [Bosonic fields in the stringlike defect model]{}, [Phys. Rev. D, v. 62 (2000).](https://journals.aps.org/prd/abstract/10.1103/PhysRevD.62.126009) [\[hep-th/0008012\]](https://arxiv.org/abs/hep-th/0008012) P. Midodashvili, [Brane in 6D and localization of matter fields]{}, (2003). [\[hep-th/0308003\]](https://arxiv.org/abs/hep-th/0308003) [\[[IN]{}SPIRE\]](https://inspirehep.net/record/624796) I. Oda, [Localization of Various Bulk Fields on a Brane]{}, (2000). [\[hep-th/0009074\]](https://arxiv.org/abs/hep-th/0009074) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/533253) I. Oda, [Gravitational localization of all local fields on the brane]{}, [Phys. Lett. B 571 (2003) 235-244.](https://doi.org/10.1016/j.physletb.2003.07.076) [\[hep-th/0307119\]](http://arxiv.org/abs/hep-th/0307119) [\[[IN]{}SPIRE\]](https://inspirehep.net/record/623330) R. Koley and S. Kar, [Scalar kinks and fermion localization in warped spacetimes]{}, [Class. Quant. Grav. 22 (2005) 753-768.](https://iopscience.iop.org/article/10.1088/0264-9381/22/4/008) [\[hep-th/0407158\]](https://arxiv.org/abs/hep-th/0407158) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/654792) C. Ringeval, P. Peter and J. P. Uzan, [Localization of massive fermions on the brane]{}, [Phys. Rev. D. 65 (2002) 044016.](https://journals.aps.org/prd/abstract/10.1103/PhysRevD.65.044016) [\[hep-th/0109194\]](https://arxiv.org/abs/hep-th/0109194) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/563412/) A. Melfo, N. Pantoja and J. D. Tempo, [Fermion localization on thick branes]{}, [Phys. Rev. D. 73 (2006) 044033.](https://journals.aps.org/prd/abstract/10.1103/PhysRevD.73.044033) [\[hep-th/0601161\]](https://arxiv.org/abs/hep-th/0601161) [\[[IN]{}SPIRE\]](https://inspirehep.net/record/709070) L. B. Castro and L. A. Meza, [Fermion localization on branes with generalized dynamics]{}, [EPL. 102 (2013).](https://doi.org/10.1209/0295-5075/102/21001) [\[hep-th/1011.5872\]](http://arxiv.org/abs/arXiv:1011.5872) [\[[IN]{}SPIRE\]](https://inspirehep.net/record/879077) W. M. Mendes, G. Alencar and R. R. Landim, [Spinors Fields in Co-dimension One Braneworlds]{}, [JHEP. 1802 (2018) 018.](https://doi.org/10.1007/JHEP02(2018)018) [\[hep-th/1712.02590\]](http://arxiv.org/abs/arXiv:1712.02590) [\[[IN]{}SPIRE\]](https://inspirehep.net/record/1641797) B. Batell and T. Gherghetta, [Yang-Mills Localization in Warped Space]{}, [Phys. Rev. D 75 (2007) 025022.](https://doi.org/10.1103/PhysRevD.75.025022) [\[hep-th/0611305\]](http://arxiv.org/abs/hep-th/0611305) [\[[IN]{}SPIRE\]](https://inspirehep.net/record/732870) R. Casadioa and A. Gruppuso, [Discrete symmetries and localization in a brane world]{}, [Phys. Rev. D 64 (2001) 025020.](https://doi.org/10.1103/PhysRevD.64.025020) [\[hep-th/0103200\]](http://arxiv.org/abs/hep-th/0103200) [\[[IN]{}SPIRE\]](https://inspirehep.net/record/554478) G. Alencar, R. R. Landim, M. O. Tahim, C. R. Muniz and R. N. Costa Filho, [Antisymmetric Tensor Fields in Randall Sundrum Thick Branes]{}, [Phys. Lett. B 693 (2010) 503-508.](https://doi.org/10.1016/j.physletb.2010.09.005) [\[hep-th/1008.0678\]](http://arxiv.org/abs/arXiv:1008.0678) [\[[IN]{}SPIRE\]](https://inspirehep.net/record/864444) C. Germani, [Spontaneous localization on a brane via a gravitational mechanism]{}, [Phys. Rev. D 85 (2012) 055025.](https://doi.org/10.1103/PhysRevD.85.055025) [\[hep-th/1109.3718\]](http://arxiv.org/abs/arXiv:1109.3718) [\[[IN]{}SPIRE\]](https://inspirehep.net/record/927734) R. G. Landim, [Gauge field and brane-localized kinetic terms on the chiral square]{}, [Eur. Phys. J. C 79 (2019) n. 10, 862.](https://doi.org/10.1140/epjc/s10052-019-7376-1) [\[hep-th/1907.10460\]](http://arxiv.org/abs/arXiv:1907.10460) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/1746091) R. G. Landim and T. G. Rizzo, [Thick Branes in Extra Dimensions and Suppressed Dark Couplings]{}, [JHEP 1906 (2019) 112.](https://doi.org/10.1007/JHEP06(2019)112) [\[hep-th/1902.08339\]](http://arxiv.org/abs/arXiv:1902.08339) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/1721405) I. Oda, [A new mechanism for trapping of photon]{}, (2001). [\[hep-ph/0103052\]](http://arxiv.org/abs/hep-th/0103052) [\[[IN]{}SPIRE\]](https://inspirehep.net/record/553860) K. Ghoroku and A. Nakamura, [Massive vector trapping as a gauge boson on a brane]{}, [Phys. Rev. D. 65 (2002) 084017.](https://doi.org/10.1103/PhysRevD.65.084017) [\[hep-th/0106145\]](https://arxiv.org/abs/hep-th/0106145) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/558602) G. Alencar, R. R. Landim, M. O. Tahim and R. N. Costa Filho, [Gauge Field Localization on the Brane Through Geometrical Coupling]{}, [Phys. Lett. B. 739 (2014) 125.](https://doi.org/10.1016/j.physletb.2014.10.040) [\[hep-th/1409.4396\]](https://arxiv.org/abs/1409.4396) [\[[IN]{}SPIRE\]](https://inspirehep.net/record/1316548) Z. H. Zhao, Q. Y. Xie and Y. Zhong, [New localization method of $U(1)$ gauge vector field on flat branes in (asymptotic) $AdS_{5}$ spacetime]{}, [Class. Quant. Grav. 32 (2015) 035020.](https://doi.org/10.1088/0264-9381/32/3/035020) [\[hep-th/1406.3098\]](https://arxiv.org/abs/1406.3098) [\[[IN]{}SPIRE\]](https://inspirehep.net/record/1300372?ln=pt) A. Kehagias and K. Tamvakis, [Localized gravitons, gauge bosons and chiral fermions in smooth spaces generated by a bounce]{}, [Phys. Lett. B. 504 (2001) 38.](https://www.sciencedirect.com/science/article/abs/pii/S037026930100274X?via%3Dihub) [\[hep-th/0010112\]](https://arxiv.org/abs/hep-th/0010112) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/535084) C-E. Fu, Y-X. Liu and H. Guo, [Bulk matter fields on two-field thick branes]{}, [Phys. Rev. D 84 (2011) 044036.](https://doi.org/10.1103/PhysRevD.84.044036) [\[hep-ph/1101.0336\]](http://arxiv.org/abs/arXiv:1101.0336) [\[[IN]{}SPIRE\]](https://inspirehep.net/record/883262) A. E. R. Chumbes, J. M. Hoff da Silva and M. B. Hott, [A model to localize gauge and tensor fields on thick branes]{}, [Phys. Rev. D 85 (2012) 085003.](https://doi.org/10.1103/PhysRevD.85.085003) [\[hep-ph/1108.3821\]](http://arxiv.org/abs/arXiv:1108.3821) [\[[IN]{}SPIRE\]](https://inspirehep.net/record/924358) G. Alencar, I. C. Jardim, R. R. Landim, C. R. Muniz and R. N. Costa Filho, [Generalized nonminimal couplings in Randall-Sundrum scenarios]{}, [Phys. Rev. D 93 (2016) 124064.](https://doi.org/10.1103/PhysRevD.93.124064) [\[hep-th/1506.00622\]](https://arxiv.org/abs/1506.00622) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/1373906/?ln=pt) W. T. Cruz, M. O. Tahim and C. A. S. Almeida, [Gauge field localization on a dilatonic deformed brane]{}, [Phys. Lett. B. 686 (2010) 259-263.](https://doi.org/10.1016/j.physletb.2010.02.064) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/852111) W. T. Cruz, A. R. P. Lima and C. A. S. Almeida, [Gauge field localization on the Bloch Brane]{}, [Phys. Rev. D. 87 (2012) 045018.](https://doi.org/10.1103/PhysRevD.87.045018) [\[hep-th/1211.7355\]](https://arxiv.org/abs/1211.7355) [\[[IN]{}SPIRE\]](https://inspirehep.net/record/1204986) A. Flachi and M. Minamitsuji, [Field localization on a brane intersection in anti-de Sitter spacetime]{}, [Phys. Rev. D. 79 (2009) 104021.](https://journals.aps.org/prd/abstract/10.1103/PhysRevD.79.104021) [\[hep-th/0903.0133\]](https://arxiv.org/abs/0903.0133) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/814467/?ln=pt) L. F. Freitas, G. Alencar and R. R. Landim, [Universal Aspects of $U(1)$ Gauge Field Localization on Branes in $D$-dimensions]{}, [JHEP. 1902 (2019) 035.](https://doi.org/10.1007/JHEP02(2019)035) [\[hep-ph/1809.07197\]](http://arxiv.org/abs/arXiv:1809.07197) [\[[IN]{}SPIRE\]](https://inspirehep.net/record/1694700) M. J. Duff and J. T. Liu, [Hodge duality on the brane]{}, [Phys. Lett. B 508 (2001) 381.](https://www.sciencedirect.com/science/article/pii/S0370269301005202?via%3Dihub) [\[hep-th/0010171\]](https://arxiv.org/abs/hep-th/0010171) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/535448) F. W. V. Costa, J. E. G. Silva and C. A. S. Almeida, [Gauge vector field localization on a $3$-brane placed in a warped transverse resolved conifold]{}, [Phys. Rev. D 87 (2013) n. 12, 125010.](https://doi.org/10.1103/PhysRevD.87.125010) [\[hep-ph/1304.7825\]](http://arxiv.org/abs/arXiv:1304.7825) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/1231520/?ln=pt) F. W. V. Costa, J. E. G. Silva, D. F. S. Veras and C. A. S. Almeida, [Gauge fields in a string-cigar braneworld]{}, [Phys. Lett. B 747 (2015) 517-522.](https://doi.org/10.1016/j.physletb.2015.06.042) [\[hep-ph/1501.00632\]](http://arxiv.org/abs/arXiv:1501.00632) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/1336325?ln=pt) Z-H. Zhao and Q-Y. Xie, [Localization of $U(1)$ gauge vector field on flat branes with five-dimension (asymptotic) AdS$_5$ spacetime]{}, [JHEP, 1805 (2018) 072.](https://doi.org/10.1007/JHEP05(2018)072) [\[hep-ph/1712.09843\]](http://arxiv.org/abs/arXiv:1712.09843) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/1645225/?ln=pt) R. R. Landim, G. Alencar, M. O. Tahim and R. N. Costa Filho, [A Transfer Matrix Method for Resonances in Randall-Sundrum Models]{}, [JHEP, 1108 (2011) 071.](https://doi.org/10.1007/JHEP08(2011)071) [\[hep-ph/1105.5573\]](http://arxiv.org/abs/arXiv:1105.5573) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/901766?ln=pt) R. R. Landim, G. Alencar, M. O. Tahim and R. N. Costa Filho, [A Transfer Matrix Method for Resonances in Randall-Sundrum Models II: The Deformed Case]{}, [JHEP, 1202 (2012) 073.](https://doi.org/10.1007/JHEP02(2012)073) [\[hep-ph/1110.5855\]](http://arxiv.org/abs/arXiv:1110.5855) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/943169) G. Alencar, R. R. Landim, M. O. Tahim and R. N. Costa Filho, [A Transfer Matrix Method for Resonances in Randall-Sundrum Models III: An analytical comparison]{}, [JHEP, 1301 (2013) 050.](https://doi.org/10.1007/JHEP01(2013)050) [\[hep-ph/1207.3054\]](http://arxiv.org/abs/arXiv:1207.3054) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/1122173) T-T. Sui, W-D. Guo, Q-Y. Xie and Y-X Liu, [Generalized geometrical coupling for vector field localization on thick brane in asymptotic Anti-de Sitter spacetime]{}, [\[hep-th/2001.02154\]](http://arxiv.org/abs/arXiv:2001.02154) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/1774329) R. S. Torrealba, [Localizing Gauge Fields on a Topological Abelian String and the Coulomb Law]{}, [Phys. Rev. D 82 (2010) 024034.](https://doi.org/10.1103/PhysRevD.82.024034) [\[hep-ph/1003.4199\]](http://arxiv.org/abs/arXiv:1003.4199) [\[[IN]{}SPIRE\]](https://inspirehep.net/record/849592) M. Giovannini, [Gauge field localization on Abelian vortices in six dimensions]{}, [Phys. Rev. D. 66 (2002).](https://journals.aps.org/prd/abstract/10.1103/PhysRevD.66.044016) [\[hep-th/0205139\]](https://arxiv.org/abs/hep-th/0205139) [\[[IN]{}SPIRE\]](http://inspirehep.net/record/586803) C.-E. Fu, Y. Zhong and Y.-X. Liu, [$U(1)$ gauge vector field on a co-dimension-2 brane]{}, [JHEP 1901 (2019) 021.](https://doi.org/10.1007/JHEP01(2019)021) [\[hep-th/1810.02081\]](http://arxiv.org/abs/arXiv:1810.02081) [\[[IN]{}SPIRE\]](https://inspirehep.net/record/1696988) G. Alencar, [Hidden conformal symmetry in Randall–Sundrum $2$ model: Universal fermion localization by torsion]{}, [Phys. Lett. B 773 (2017) 601-603.](https://doi.org/10.1016/j.physletb.2017.09.014) [\[hep-th/1705.09331\]](http://arxiv.org/abs/arXiv:1705.09331) [\[[IN]{}SPIRE\]](https://inspirehep.net/record/1601374) [^1]: Here, the index $M,N,..$ are related to the curved space, and $a,b,..$ are related to the flat space. Beyond this, the vierbein satisfies $E^{M}_{a}E^{Na}=g^{MN}$, and the Gamma matrix satisfy $\left\lbrace\Gamma^{M},\Gamma^{N}\right\rbrace=-2g^{MN}$. [^2]: The vierbein $\hat{e}^{\mu}_{a}(x)$ should satisfy $\hat{e}^{\mu}_{a}(x)\hat{e}^{\nu a}(x)=\hat{g}^{\mu\nu}(x)$.	{ "pile_set_name": "ArXiv" }
--- abstract: \| We consider the semilinear elliptic problem $$\label{problemAbstract} \left\{\begin{array}{lr}-\Delta u= \|u\|^{p-1}u\qquad \mbox{ in }B\\ u=0\qquad\qquad\qquad\mbox{ on }\partial B \end{array}\right.\tag{$\mathcal E_p$}$$ where $B$ is the unit ball of ${\mathbb{R}}^2$ centered at the origin and $p\in (1,+\infty)$. We prove the existence of non-radial sign-changing solutions to which are quasi-radial, namely solutions whose nodal line is the union of a finite number of disjoint simple closed curves, which are the boundary of nested domains contained in $B$. In particular the nodal line of these solutions doesn’t touch $\partial B$.\ The result is obtained with two different approaches: via nonradial bifurcation from the least energy sign-changing radial solution $u_p$ of at certain values of $p$ and by investigating the qualitative properties, for $p$ large, of the least energy nodal solutions in spaces of functions invariant by the action of the dihedral group generated by the reflection with respect to the $x$-axis and the rotation about the origin of angle $\frac{2\pi}{k}$ for suitable integers $k$.\ We also prove that for certain integers $k$ the least energy nodal solutions in these spaces of symmetric functions are instead radial, showing in particular a breaking of symmetry phenomenon in dependence on the exponent $p$. address: - 'Francesca Gladiali, Polcoming, Università di Sassari - Via Piandanna 4, 07100 Sassari, Italy' - 'Isabella Ianni, Dipartimento di Matematica e Fisica, Università degli Studi della Campania Luigi Vanvitelli, V.le Lincoln 5, 81100 Caserta, Italy' author: - 'F. Gladiali, I. Ianni' title: \| Quasi-radial nodal solutions\ for the Lane-Emden problem in the ball --- [^1] [^2] [^3] Introduction {#Introduction} ============ We consider the semilinear Lane-Emden problem $$\label{problem} \left\{\begin{array}{lr}-\Delta u= \|u\|^{p-1}u\qquad \mbox{ in }B\\ u=0\qquad\qquad\qquad\mbox{ on }\partial B\\ \end{array}\right.$$ where $B\subset{\mathbb{R}}^2 $ is the unit ball centered at the origin and $p>1$.\ It is well known that admits a unique positive ground state solution which is radially symmetric. Observe that the oddness of the nonlinearity implies that $u$ is a solution of if and only if $-u$ is a solution, so there is also a unique negative solution to .\ Moreover, due to the oddness of the nonlinear term, standard variational methods give the existence of infinitely many sign-changing solutions. While the ground state solution of has been widely investigated, not much is known for nodal ones. Among these one can select the least energy nodal solutions, which can be obtained by minimizing the associated energy functional $$\label{functional}E_p(u)\ : \ =\ \frac 12 \int_{B}\|\nabla u \|^2 -\frac 1{p+1}\int_B \|u\|^{p+1}$$ on the nodal Nehari set in the Sobolev space $H^1_0(B)$ (see [@CastroCossioNeuberger] for details). We denote a least energy sign-changing solution by $\widetilde u_p$. In [@BartschWeth] it has been shown that $$\label{leastRegioEMorse} \sharp(\widetilde u_p)=2\quad \mbox{ and }\quad m(\widetilde u_p)=2,$$ where $\sharp(u)$ is the number of nodal regions of $u$ and $m(u)$ is the Morse index of the solution $u$ (see Section \[section:linearizedOperator\] for the definition). Moreover in [@BartschWethWillem] it has been proved that $\widetilde u_p$ partially inherits the symmetries of the domain, being foliated Schwarz symmetric, namely axially symmetric with respect to an axis passing through the origin and nonincreasing in the polar angle from this axis (see also [@PacellaWeth]). Since the domain $B$ is radially symmetric one can restrict to the Sobolev space of radial functions $H^1_{0, {{\text{\upshape rad}}}}(B)$ and prove the existence of infinitely many sign-changing radial solutions for . More precisely it can be proved that for every $m\in\mathbb N_0:=\mathbb N\setminus \{0\}$ there exists a unique radial solution to that satisfies $$\label{segno-in-zero} u(0)>0$$ and such that $\sharp(u)=m$ (see [@NN], [@KAJIKIYA]). We denote by $u_p$ the unique radial least energy sign changing solution to which satisfies , clearly $$\label{radduenod} \sharp(u_p)=2.$$ Morover it has been proved in [@AftalionPacella] that $$\label{StimaAPMorseR} m(u_p)\geq4$$ (see also [@DeMarchisIanniPacellaMathAnn] where $m(u_p)$ has been explicitly computed for $p$ large and also [@DeMarchisIanniPacellaAdv] where the previous estimate on the Morse index has been generalized to any radial solutions with $m$ nodal regions, with bound given by the number $3m-2$). Comparing the information on the Morse index in with the one in one gets that the radial solution $u_p$ is not the least energy sign-changing solution in the whole space $H^1_0(B)$, namely that $u_p\neq \widetilde u_p$. As a consequence the monotonicity of $\widetilde u_p$ with respect to the polar angle (as recalled above $\widetilde u_p$ is foliated Schwarz symmetric) must be strict at some region, and in [@PacellaWeth] it is actually proved that, for $p>2$, the monotonicity is always strict. Moreover in [@AftalionPacella] it has been also proved that the nodal set of $\widetilde u_p$ touches the boundary of $B$. One can also restrict to the Sobolev space $H^1_{0,k}(B)$ of the functions in $H^1_0(B)$ which are even and $\frac{2\pi}{k}$-periodic in the angular variable, for $k\in\mathbb N_0$ and similarly show the existence of infinitely many sign-changing symmetric solutions in $H^1_{0,k}(B)$, among which we denote by $u_p^k$ the least energy ones. Anyway a priori it is not clear whether this procedure produces new solutions or not. Indeed, clearly $u_p^1=\widetilde u_p$ (since $\widetilde u_p $ is axially symmetric) and, even though $u_p^k\neq \widetilde u_p$ for $k\geq 2$ (since if they coincide then $\widetilde u_p$ would be $\frac{2\pi}{k}$-periodic in the angular variable and so necessarily radial by the Schwarz symmetry, getting a contradiction), $u_p^k$ could be radial. In particular it would be interesting to show the existence of sign-changing solutions to which belong to $H^1_{0,k}(B)$ but are not radially symmetric, having nevertheless a quasi-radial shape, in the sense of the following definition: \[def:quasiradial\] We say that a solution of is quasi-radial if its nodal set is the union of a finite number of disjoint simple closed curves which are the boundary of nested domains contained in $B$. Observe that the nodal line of a quasi-radial solution doesn’t touch the boundary of the ball $B$. Clearly any radial solution is quasi-radial. By the asymptotic estimates for the energy of the solutions of in [@RenWei], the obvious inequality $E_p(u_p^k)\leq E_p(u_p)$ and the upper bound $$pE_p(u_p)\leq \alpha\cdot 4\pi e, \quad\mbox{ for }p \mbox{ large},$$ proved in [@GGP2] for a certain value $\alpha\in (4.5,5)$, one derives the following upper bound on the number of nodal regions of $u_p^k$: $$\sharp(u_p^k)\leq 4 \quad \forall k\in\mathbb N_0, \mbox{ for }p \mbox{ large}.$$ Combining this bound with the results in [@DeMarchisIanniPacellaJDE] (which hold in symmetric and simply connected domains, more general than the ball) it then follows that the least energy symmetric solution $$\label{primaquadratino} u_p^k \mbox{ is \emph{quasi-radial} when }k\geq 4\mbox{ and }p\mbox{ is large,}$$ from which in particular one also derives $$\label{quadratino} \sharp(u_p^k)=2 \quad \mbox{ and }\quad m(u_p^k)\geq4, \quad \mbox{ for }k\geq 4, \mbox{ for }p \mbox{ large.}$$ Observe that the properties in , are satisfied also by $u_p$ (see , ), hence the question of the existence of symmetric but non-radial solutions of which are quasi-radial is still open.\ Moreover, as $p\in (1,+\infty)$ and $k\in\mathbb N_0$ vary, one would like to investigate whether $u_p^k$ coincides with the radial least energy nodal solution $u_p$ or not. We start by giving a positive answer to the first question, showing the existence of three distinct solutions to which belong respectively to $H^1_{0,k}(B)\setminus H^1_{0,{{\text{\upshape rad}}}}(B)$, for $k=3,4,5$. Each solution bifurcates from the least energy radial nodal solution $u_p$ at certain values of $p$ and close to the bifurcation point it is quasi-radial. The result is the following, where $\mathcal X_k:=H^1_{0,k}(B)\cap C^{1,\alpha}(\bar B)$ (and $C^{1,\alpha}(\bar B)$ denotes the space of $C^1(\bar B)$ functions with Hölder derivatives): \[teo1\] For any $k=3,4,5$ there exists at least one exponent $p^k\in(1,+\infty)$ such that $(p^k, u_{p^k})$ is a nonradial bifurcation point for problem . The bifurcating solutions are sign-changing, belong to $ \mathcal X_k$ and close to the bifurcation point they have two nodal domains and are quasi-radial. Moreover the bifurcation is global and, letting $\mathcal{C}_k$ be the continuum that branches out of $( p^k, u_{p^k})$, then either $\mathcal{C}_k$ is unbounded in $(1,+\infty)\times\mathcal X_k$ or it intersects $\{1\}\times\mathcal X_k$. Finally at any point along each branch $\mathcal C_k$ either the solution belongs to $\mathcal X_k\setminus\mathcal X_j$, $\forall j>k$ or it is radial, in particular the continua bifurcating from different values of $k$ can intersect only at radial solutions. Our second goal is to understand whether the least energy symmetric solution $u_p^k$, $k\in\mathbb N_0$, $p\in (1,+\infty)$, coincides with the radial least energy nodal solution $u_p$ or not, and this is analyzed in next result: \[prop1.4\] Let $u_p^k$ be the least energy sign-changing solution of in the space $H^1_{0,k}(B)$, $k\in\mathbb N_0$, then there exist $\delta >0$ and $p^{\star} >1$ such that: - for $k=2$: $\quad u_p^k$ is non-radial both for $p\in (1,1+\delta)$ and $p\geq p^{\star}$; - for $k=3,4,5$: $\quad u_p^k$ is radial for $p\in (1,1+\delta)$ and non-radial when $p\geq p^{\star}$; - for $k\geq 6$: $\quad u_p^k$ is radial for $p\in (1,1+\delta)$. Clearly when $u_p^k$ is radial then it coincides with $u_p$ (up to the sign). The fact that the symmetry of the domain is not totally caught by these least energy solutions is reasonable, since we are dealing with sign-changing solutions, anyway the symmetry breaking phenomenon when $k=3,4,5$ (case $ii)$) and its dependence on the value of the exponent $p$ were totally unexpected. It is also interesting that we can identify the symmetries of the solutions at which this symmetry breaking phenomenon occurs. Theorem \[prop1.4\]-$ii)$ combined with and provides another example for non-radial symmetric sign-changing solution of which are quasi-radial in the sense of Definition \[def:quasiradial\]. Differently with respect to Theorem \[teo1\], this result is now for any $p$ large enough: \[cor:teoleast\] Let $k=4,5$, then $u_p^k$ is not radial but it is quasi-radial for $p$ large enough. In particular $u_p^k\neq u_p$. Moreover $u_p^k\neq\widetilde u_p$ and holds. We conjecture that the bifurcating solution in $\mathcal X_k$ found in Theorem \[teo1\] not only exists for any $p\geq p^k$ but also coincides with $u_p^k$, when $k=4,5$ and even $3$. Differently from the higher symmetry cases considered in Corollary \[cor:teoleast\], when $k=3$ we do not expect $u_p^k$ to keep the quasi-radial shape for large $p$. For $k=2$ we believe that $u_p^k$ is not radial for all $p$ and also not quasi-radial (when $p$ is close to $1$ it could be proved rigorously, see Remark \[RemarkNoQuaiRadk=2\]), for $k=1$ we recall that $u_p^k=\widetilde u_p$ for any $p\in (1,+\infty)$. The case $k\geq 6$ and $p$ large is not covered by the previous result, we conjecture that $u_p^k$ is radial, observe that this is not in contrast with . The asymptotic behavior, as $p\rightarrow +\infty$, of the least energy sign-changing solution $u_p^k$ of in the spaces $H^{1}_{0,k}(B)$ will be object of a subsequent paper [@GIP]. Next we briefly explain the main ideas to get Theorem \[teo1\] and Theorem \[prop1.4\]. The bifurcation in Theorem \[teo1\] is with respect to the exponent $p$ of the nonlinearity, previous results in this direction can be found for instance in [@GladialiGrossiPacellaSrikanth] and [@G10]. Observe that the bifurcation can occur only at values $p$ at which the least energy nodal radial solution $u_p$ is degenerate and that a sufficient condition to identify degeneracy points is to have a change in the Morse index of $u_p$.\ This paper starts then from the recent results in [@DeMarchisIanniPacellaMathAnn] where the Morse index of the radial least energy sign-changing solution $u_p$ is computed for large values of $p$, proving the existence of an exponent $p^{\star}>1$ such that: $$\label{morseplargeintro} m(u_p)=12 \qquad \forall\ p\geq p^{\star}.$$ This result is only for large $p$ and it strongly relies on the asymptotic behavior of $u_p$ as $p\rightarrow +\infty$, which has been described in [@GGP2]. Indeed, an asymptotic analysis of the behavior of the solution $u_p$ as $p\rightarrow 1$ shows that a suitable re-normalization of $u_p$ converges to the second radial eigenfunction of the Laplace operator with Dirichlet boundary conditions (see Lemma \[lemma-pvicino1\]) and this allows to compute the Morse index of $u_p$ for $p$ close to $1$, showing that it has a different value in this range. More precisely in Proposition \[risultatoMorse\_pvicino1\] we get the existence of $\delta>0$ such that $$\label{morse-index-6} m(u_p)=6 \qquad \forall\ p\in (1,1+\delta).$$ Hence and prove that along the branch of radial solutions $(p,u_p)$ of there should be points at which the Morse index increases and this change of the Morse index of $u_p$ in the interval $(1,+\infty)$ suggests bifurcation from $u_p$. We underline that in the convex domain $B$ this phenomenon is specific of sign-changing solutions, since the positive solution in $B$ is unique and non-degenerate (for the uniqueness in more general convex domains see [@DeMarchisGrossiIanniPacellaprogress]).\ Anyway this is the first time that a non-radial bifurcation result from sign-changing solutions in convex domains is observed and there was no chance to get it before the study of the Morse index in [@DeMarchisIanniPacellaMathAnn]. To prove the result in Theorem \[teo1\] we need first to analyze the degeneracies of the solution $u_p$. This is the goal of Sections \[section:generale\], \[section p large\] and \[sse:pvicino1\]. We first consider in Section \[se:auxiliary\] an auxiliary singular weighted eigenvalue problem $$\label{problemaPesatoPallaINTRO} \begin{cases} \begin{array}{ll} -\Delta \psi - p\|u_p(x)\|^{p-1} \psi =\frac{\beta}{\|x\|^2} \psi\qquad & \text{ in } B\setminus\{0\}, \\ \psi= 0 & \text{ on } \partial B\\ \int_B\|\nabla \psi\|^2+\frac{\psi^2}{\|x\|^2}<+\infty, \end{array} \end{cases}$$ which has the same kernel and the same number of negative eigenvalues of the linearized operator at $u_p$ (see Lemma \[lemma:Morse=numeroAutovaloriesatoPalla\]) and whose main advantage relies on the fact that, in addition, a classical spectral decomposition into radial and angular part may be applied to it (Lemma \[lemma:decomposizionePalla\]). The weighted eigenvalue problem belongs to the class of eigenvalue problems which has been studied in [@GGN2], where the eigenvalues for have been variationally characterized in the case when they are negative.\ Since $u_p$ is the radial least energy nodal solution, then in the space of radial functions its Morse index is $2$, in Section \[subsec:mo\], in view of the spectral decomposition, we estimate the two negative radial eigenvalues of problem from above and from below by certain consecutive eigenvalues of $-\Delta_{S^1}$. The proof is based on the approximation of the negative eigenvalues of problem by the negative eigenvalues of a family of weighted eigenvalue problems in annuli already studied in [@DeMarchisIanniPacellaMathAnn], in particular we can extend some previous estimates in [@DeMarchisIanniPacellaMathAnn] related to the negative radial eigenvalues in annuli to the negative eigenvalues for the singular problem . As a consequence of our estimates we get information about the Morse index of the solution $u_p$ (Lemma \[proposition:autovaloriRadiaiGenerale\]) and a general characterization of its degeneracy (Proposition \[p4.7\]), for any $p>1$. Finally, thanks to and , we get more specific results both in the case $p$ large and $p$ close to $1$ (see Sections \[section p large\] and \[sse:pvicino1\]).\ Observe that, due now to the spectral decomposition, we can decompose any solution of the linearized equation at $u_p$ (and more in general each solution of the eigenvalue problem ) along spherical harmonics, which in ${\mathbb{R}}^2$ are the functions $\cos ( j\theta), \ \sin (j \theta)$ with $j\in {\mathbb{N}}$, getting in particular an explicit representation of the solutions of the linearized equation when they are nontrivial (and more in general of the eigenfunctions of associated with negative eigenvalues). As a consequence we can then identify the symmetries of those functions which are responsible of the degeneracy of $u_p$ (or which give rise to negative eigenvalues for the linearized operator at $u_p$). This aspect has been investigated in Section \[se:symmspaces\], where the symmetric spaces ($H^1_{0,k}(B)$ and) $\mathcal X_k$ have been introduced and the degeneracy and Morse index of $u_p$ in these spaces studied (see Proposition \[Morse-simmetrico1\], \[Morse-simmetrico2\] and \[lemma:degenerazioneSimmetria\]).\ The reason for restricting to the spaces $\mathcal X_k$ is to isolate a unique function in the kernel of the linearized operator; more precisely, on one side it allows to select one suitable spherical harmonic (between $\sin$ and $\cos$) that produces degeneracy and, on the other side it avoids a possible double degeneracy due to the contemporary vanishing of two eigenvalues, possibility that cannot be ruled out and it is specific of sign-changing solutions. Since we do not know explicitly the solution $u_p$, it is not clear whether the transversality condition of the well-known Crandall-Rabinowitz Theorem (for one dimensional kernel) is satisfied or not. Anyway the bifurcation result may be obtained here using a degree argument. The separation of the branches is obtained defining suitable cones $\mathcal K_k\subset\mathcal X_k$ of monotone functions introduced by Dancer in [@D] and using the degree in cones, see [@Amman] (see Section \[se:bifurcation\] for the definitions of the cones). The quasi-radiality is inherited from the radial least energy solution $u_p$, since near the bifurcation point the bifurcating solution is a small perturbation of it (see Remark \[remark:shape\]).\ Along the branch instead the number of nodal regions and the shape of the solutions may change, anyway the characterization of the behavior for branches of non-radial solutions may be a very difficult task to investigate, we also conjecture that the branches exist for every $p\geq p^k$. In this paper we have focused on the radial least energy sign-changing solution $u_p$ of . A bifurcation result similar to Theorem \[teo1\] could be obtained from any nodal radial solution $u_p^m$ of with $m> 2$ nodal regions, provided information about its Morse index when $p$ is large is available. In this case we expect that the symmetries which cause the degeneracy and hence produce branches of bifurcating solutions, should be of the same type of the one for functions in $\mathcal X_k$ (which derive by the symmetry groups of spherical harmonics), but with different values of $k$, probably $k\geq 6$. Moreover one could think to extend the bifurcation result in Theorem \[teo1\] also to higher dimension $N\geq 3$, when $p\in (1,\frac{N+2}{N-2})$. Indeed the behavior of all the radial sign-changing solutions of has been studied in [@DeMarchisIanniPacellaAdv] and in particular their Morse index has been explicitly computed when $p$ is sufficiently close to $\frac{N+2}{N-2}$, giving for instance, for the radial least energy sign-changing solution $u_p$: $$m(u_p)=2+N, \quad \mbox{ for $p$ close to $\frac{N+2}{N-2}$. }$$ Similarly as in the $2$-dimensional case, we expect a change in the Morse index of $u_p$ as $p$ varies from $1$ to $\frac{N+2}{N-2}$. Indeed $u_p$ should converge as $p\rightarrow 1$ to the radial Dirichlet eigenfunction with $2$ nodal regions of the Laplace operator in $B$ and this would imply $$m(u_p)= 2+N+\frac{(N+2)(N-1)}2, \quad\mbox{ for $p$ close to $1$.}$$ Again a change in the Morse index should give a nonradial bifurcation result. An extra difficulty in dimension $N\geq 3$ would be to identify the symmetry groups of the spherical harmonics, which are much more involved than those of the $2$-dimensional spherical harmonics, see for instance [@AG]. Next we discuss the main ideas behind the proof of Theorem \[prop1.4\], which is contained in Section \[section:proofTeoLeastEnergy\]. The non-radial part is a byproduct of the study of the symmetry groups that cause the degeneracy and the bifurcation from $u_p$. Indeed in order to prove that $u_p$ and $u_p^k$ do not coincide one would like to compare their Morse indexes and show that they are different. However the computation of $m(u_p^k)$ may be very difficult, but if we restrict to the symmetric spaces $H^1_{0,k}(B)$ then the Morse index of $u_p^k$ is always $2$ (see Lemma \[lem-10-1\]). On the other side we are able to compute the symmetric Morse index also for the radial solution $u_p$ (Proposition \[Morse-simmetrico1\] and \[Morse-simmetrico2\]). Observe that it is computed only for $p$ close to $1$ and $p$ large since it is deduced, among other things, from the asymptotic analysis of $u_p$ as $p\rightarrow 1$ and as $p\rightarrow +\infty$ respectively. The proof of the [radial part]{} of Theorem \[prop1.4\] is more involved. It relies on a careful blow-up procedure in the spirit of [@GS] for showing $L^{\infty}$ bounds for the solutions $u_p^k$ (see Proposition \[prop7.4\]). Once an $L^{\infty}$ bound is available one can deduce the result by studying the asymptotic behavior of the solutions $u_p^k$ as $p\rightarrow 1$ (see the proof of Proposition \[leastRadiale\]). In particular a delicate expansion of ${\lVertu_p^k\rVert}_{\infty}$ at $p=1$ up to the second order is needed.\ Getting a uniform $L^{\infty}$ bound is somehow standard for solutions with uniformly bounded Morse index, since one shows that the bound on the Morse index is preserved as $p\rightarrow 1$, while the blow-up analysis of unbounded solutions in $L^{\infty}$-norm leads to solutions to limit problems in unbounded domains, whose Morse index is not finite, thus reaching a contradiction.\ The main problem here is that for the least energy symmetric solutions $u_p^k$ we do not have a bound for the full Morse index, but [only for the $k$-Morse index]{} (see Lemma \[lem-10-1\]), while in the rescaling procedure the symmetries are not preserved.\ To overcome this technical difficulty we exploit the symmetry of $u_p^k$ and reduce problem to the circular sector $S_k$ of the ball of amplitude $\frac{\pi}k$, for $k\in {\mathbb{N}}_0$. In particular we are able to convert the bound on the $k$-Morse index to a bound on the full Morse index of $u_p^k$ in the sector $S_k$ (Morse index for a mixed Dirichlet-Neumann problem, see Lemma \[lem-10-nuovo\]) and finally we perform the blow-up argument in $S_k$.\ Also the blow-up procedure in $S_k$ requires special care, since we have to deal with mixed boundary conditions and, above all, with the angular points of $S_k$. For these reasons the analysis of the rescaled solutions includes several different cases, depending upon the location of the maximum points in the sector. Anyway in all the cases we end-up with solutions to a limit linear problem in unbounded domains with either Dirichlet or Neumann or mixed boundary conditions, whose Morse index is finite. Finally studying the Morse index of solutions for these limit problems (Proposition \[prop7.3\]) we get a contradiction. Preliminary results {#section:preliminaries} =================== \[PropositionUnicoMaxeMin\] admits a unique radial solution $(u_p)$ having $2$ nodal regions and satisfying . Moreover: - $u_p(0)=\\|u\\|_{\infty}$ - in each nodal region there is exactly one critical point (namely the maximum and the minimum points) In [@HRS Lemma 5.2] the authors proved the following estimate that can be useful in the sequel: For any $p_\in (1,+\infty)$ there exist constants $m,M$ such that, for any $p\in(1,p_]$ $$\label{}m\leq \left({\lVert u_p\rVert}_{\infty} \right)^{p-1}\leq M.$$ Finally we state a Proposition which provides the behavior, at the singularity, of solutions to a singular ordinary differential equation. This result is partially contained in [@GGN2 Lemma 2.4], although one implication is new and proved here. \[p2.2\] Let $\psi$ be a solution to $$\label{f1a} \begin{cases} -\psi'' - \frac{1}{r}\psi' +\beta^2\frac{\psi}{r^2}=h\psi, \quad \text{in} \ \ (0,1)\\ \psi(1)=0, \ \int_0^1 r (\psi')^2dr< \infty \end{cases}$$ with $h\in L^\infty(0,1)$ and $\beta>0$. Assume that $\psi$ satisfies one of the following conditions: $$\begin{aligned} a) & \psi(0)=0\\ b) & \int_0^1\frac{\psi^2}{r} dr< \infty.\end{aligned}$$ Then $\psi\in L^\infty(0,1)$ and $$\label{psi-in-zero} \psi(r)=O(r^{\beta}) \ \ \text{ as } \ r{\rightarrow}0.$$ When $\psi$ satisfies condition $b)$ then the thesis follows from Lemma 2.4 in [@GGN2] (see estimate (2.28)). When $\psi$ satisfies condition $a)$ we observe that since $\psi$ solves then $\psi\in L^{\infty}(0,1)$. We can proceed as in the proof of Lemma 2.4 in [@GGN2]. Then, multiplying by $r_n$ and integrating in $(r_n,1)$ we get $$r_n^{\beta+1}\psi'(r_n)-r_n^\beta \psi'(1)+\beta^2r_n^{\beta}\int_{r_n}^1\frac{\psi}r \ dr=r_n^{\beta}\int_{r_n}^1r h(r)\psi(r)\ dr.$$ Using the fact that along a sequence $r_n{\rightarrow}0$ it holds $$\big\| r_n^{\beta}\int_{r_n}^1 \frac {\beta^2}s\psi(s) \ ds \Big\| \leq C r_n^{\beta}\|\log{r_n}\|=o(1)$$ we get as $n{\rightarrow}\infty$ $$r_n^{\beta+1}\psi'(r_n)=o(1).$$ Observe now that the function $v(r)=r^{\beta}$ satisfies $$\label{inter} -v''-\frac 1r v'+\frac{\beta^2}{r^2}v=0 \ \text{ in }(0,1) \ , \ v(0)=0$$ We multiply by $v$, we multiply by $\psi$, we integrate on $(r_n,R)$, with $R\in (0,1)$, we subtract the two equations and we get $$\int_{r_n}^R r^{\beta+1}h(r)\psi(r) \ dr=r_n^{\beta+1}\psi'(r_n)-\beta r_n^{\beta}\psi(r_n)-R^{\beta+1}\psi'(R)+\beta R^{\beta}\psi(R)$$ and, passing to the limit as $n{\rightarrow}\infty$ $$\int_0^R r^{\beta +1}h(r)\psi(r) \ dr=-R^{\beta+1}\psi'(R)+\beta R^{\beta}\psi(R)$$ which implies for any $t\in(0,1)$ $$\label{fra1} \frac{\psi(t)}{t^{\beta}}=\int_t^1\frac1{R^{2\beta+1}} \left(\int_0^Rs^{\beta+1}h(s)\psi(s)ds\right)dR.$$ The boundedness of $h(s)$ and $\psi(s)$ then gives $$\label{fra2} \Big\|\int_0^R s^{\beta +1}h(s)\psi(s)ds\Big\|\leq CR^{\beta+2}$$ which, together with gives $$\frac{\|\psi(t)\|}{t^{\beta}}\leq \begin{cases} C \|1-t^{2-\beta}\| & \text{ if }\beta\neq 2\\ C(1-\log t) & \text{ if }\beta= 2 \end{cases}$$ and this implies the thesis in case $\beta<2$. When $\beta\geq 2$ instead we have $ \|\psi(t)\|\leq Ct^2$ for $\beta>2$ and $ \|\psi(t)\|\leq Ct^{\beta-\varepsilon}$ for $\beta=2$ where $0<\varepsilon<<1$. Inserting these estimates into then we have $$\Big\|\int_0^R s^{\beta +1}h(s)\psi(s)ds\Big\|\leq \begin{cases} CR^{\beta+4} & \text{ if } \beta>2\\ CR^{2\beta+1-\varepsilon}& \text{ if } \beta=2 \end{cases}$$ which, together with gives $$\frac{\|\psi(t)\|}{t^{\beta}}\leq \begin{cases} C \|1-t^{4-\beta}\| & \text{ if }\beta\neq 4\\ C(1-\log t) & \text{ if }\beta= 4\\ C(1-t^{1-\varepsilon})& \text{ if }\beta= 2 \end{cases}$$ which implies the thesis when $\beta<4$. We can repeat the procedure. At each step the set of values of $\beta$ at which is satisfied increases by $2$. Then for every value of $\beta$ the thesis follows after a finite number of steps. Linearized operator {#section:linearizedOperator} =================== Let $L_{p}: H^2(B)\cap H^1_0(B)\rightarrow L^2(B)$ be the linearized operator at $u_p$, namely $$\label{linearizedOperator} L_{p} v: = -\Delta v-p\|u_p(x)\|^{p-1}v.$$ It is well known that $L_p$ admits a sequence of eigenvalues which, counting them according to their multiplicity, we denote by $$\mu_1(p)< \mu_2(p)\leq\ldots\leq\mu_i(p)\leq\ldots,\quad \mu_i(p)\rightarrow +\infty \mbox{ as }i\rightarrow +\infty,$$ where the first inequality is strict because it is known that $\mu_1(p)$ is simple. We also recall their min-max characterization $$\begin{aligned} \label{CourantCharEigenv} \mu_i(p) &=& \min_{\substack{ W\subset H^1_{0}(B)\\ dim W=i}} \max_{\substack{v\in W\\v\neq 0}}\ \ \ R_p[v],\qquad i\in{\mathbb{N}}_0\end{aligned}$$ where $R_p[v]$ is the Rayleigh quotient $$\label{Rayleigh} R_p[v]:=\frac{Q_p(v)}{\int_B v(x)^2 dx}$$ and $Q_p: H^1_0(B)\rightarrow \mathbb R$ denotes the quadratic form associated to $L_p$, namely $$\label{formaQuadratica} Q_p (v):=\int_B \left[\|\nabla v(x)\|^2 -p\|u_p(x)\|^{p-1}v(x)^2 \right]dx.$$ Since $u_p$ is a radial solution to we can also consider the subsequence of $(\mu_i(p))_{i\in\mathbb N_0}$ of the radial eigenvalues of $L_p$ (i.e. eigenvalues which are associated to a radial eigenfunction) that we denote by $$\mu_{i,{{\text{\upshape rad}}}}(p), \quad i\in{\mathbb{N}}_0$$ and which are all simple in the space of radial functions. For the eigenvalues $\mu_{i,{{\text{\upshape rad}}}}(p)$ an analogous characterization holds: $$\begin{aligned} \label{CourantCharEigenvRad} \mu_{i,{{\text{\upshape rad}}}}(p) &=& \min_{\substack{ W\subset H^1_{0,rad}(B)\\ dim W=i}} \max_{\substack{v\in W\\v\neq 0}}\ \ \ R_p[v]\end{aligned}$$ where $R_p$ is as in and $H^1_{0,{{\text{\upshape rad}}}}(B)$ is the subspace of the radial functions of $H^1_0(B)$. Moreover it is known that $\mu_{1,{{\text{\upshape rad}}}}(p)=\mu_1(p).$ The [Morse index of $u_p$]{}, denoted by $m(u_p)$, is the maximal dimension of a subspace $X\subseteq H^1_0(B)$ such that $Q_p(v)<0, \ \forall v\in X\setminus\{0\}$. Since $B$ is a bounded domain this is equivalent to say that $m(u_p)$ is the number of the negative eigenvalues of $L_p$ counted with their multiplicity.\ The [radial Morse index of $u_p$]{}, denoted by $m_{{{\text{\upshape rad}}}}(u_p)$, is instead the number of the negative radial eigenvalues $\mu_{i,{{\text{\upshape rad}}}}(p)$ of $L_{p}$. By the results in [@AftalionPacella] we have \[LemaAftalionPacella\] For any $p>1$ $$(+\infty >)\ m(u_p)\geq 4.$$ Moreover it is well known (see for instance [@BartschWeth], see also [@HRS]) the following \[LemmaMorseIndexRadiale\] For any $p>1$ $$m_{{{\text{\upshape rad}}}}(u_p)=2.$$ The previous lemma means that for any $p>1$ $$\mu_{1,{{\text{\upshape rad}}}}(p)<\mu_{2,{{\text{\upshape rad}}}}(p)<0\leq \mu_{3,{{\text{\upshape rad}}}}(p)<\ldots,$$ next we show that $$\mu_{3,{{\text{\upshape rad}}}}(p)>0,$$ namely that the problem $$\label{linearizedProblem} \left\{\begin{array}{lr} L_pv=0 & \mbox{ in }B\\ v=0 & \mbox{ on }\partial B \end{array} \right.$$ doesn’t admit nontrivial radial solutions, indeed the following result holds: \[lemma:radiallyNonDeg\] For any $p>1$ $u_p$ is radially non-degenerate. Given a solution $w_{\alpha}$ for the problem $$\label{problvaliniz} \left\{ \begin{array}{lr} w_{\alpha}''+\frac{1}{r}w_{\alpha}'+\|w_{\alpha}\|^{p-1}w_{\alpha}=0 &\quad \mbox{ in } (0,T)\\ w_{\alpha}(0)=\alpha>0 \\ w_{\alpha}'(0)=0\\ w_{\alpha} \quad \mbox{ has exactly $1$ zero in $(0,T)$}\\ w_{\alpha}(T)=0 \end{array} \right.$$ where $T>0$, it is not difficult to see (see [@SmollerWassarman]) that $w_{\alpha}$ is differentiable with respect to $\alpha$ and that it is radially non-degenerate in $(0, T)$ if and only if $\frac{\partial w_{\alpha}}{\partial\alpha}\|_{r=T}\neq 0$.\ Observe that $u_p$ solves with $\alpha=u_p(0)>0$ and $T=1$.\ Moreover for any $\alpha>0$ has a unique solution $w_{\alpha}$ which is obtained by scaling $u_p$ as $$w_{\alpha}(r):=T(\alpha)^{-\frac{2}{p-1}}u_p(\frac{r}{T(\alpha)}),$$ where $T=T(\alpha):=\left(\frac{u_p(0)}{\alpha}\right)^{\frac{p-1}{2}}$.\ Hence it is immediate to check that $\frac{\partial w_{\alpha}}{\partial\alpha}\|_{r=T(\alpha)}\neq 0$, from which it then follows that $u_p$ is radially non-degenerate. Morse index and degeneracy of $u_p$ {#section:generale} =================================== The section is organized as follows: we first consider an auxiliary weighted eigenvalue problem (problem below), whose main advantage, as we will see, relies on the fact that it shares with $L_p$ the same spectral properties (see Lemma \[lemma:Morse=numeroAutovaloriesatoPalla\]) and, in addition, a classical spectral decomposition into radial and angular part may be applied to it (Lemma \[lemma:decomposizionePalla\] in the section). The study of the auxiliary problem is carried out for any $p>1$, getting information about the Morse index of the solution $u_p$ (Lemma \[proposition:autovaloriRadiaiGenerale\]) and a general characterization of its degeneracy (Proposition \[p4.7\]). An auxiliary weighted eigenvalue problem {#se:auxiliary} ---------------------------------------- We consider the auxiliary eigenvalue problem $$\label{problemaPesatoPalla} \begin{cases} \begin{array}{ll} -\Delta \psi - p\|u_p(x)\|^{p-1} \psi =\frac{\beta}{\|x\|^2} \psi\qquad & \text{ in } B\setminus\{0\}, \\ \psi= 0 & \text{ on } \partial B\\ \int_B\|\nabla \psi\|^2+\frac{\psi^2}{\|x\|^2}<+\infty, \end{array} \end{cases}$$ where $\beta\in\mathbb R$ and $p>1$.\ Observe that, since $p\|u_p\|^{p-1}\in L^{\infty}(B)$, belongs to the class of eigenvalue problems which has been studied in [@GGN2], where the eigenvalues for have been variationally characterized in the case when they are negative. In the following we recall the variational characterization obtained in [@GGN2]. In particular they have observed that when the associated Rayleigh quotient is greater or equal than zero there is a compactness problem, but as far as the quotient is strictly negative, the eigenvalues and eigenfunctions maintain the usual properties of the classical ones.\ Let us denote by $\mathcal H$ the closure of $C^{\infty}_0(B)$ with respect to the norm $\\|v\\|^2_{\mathcal H}=\int_{B}\left( \|\nabla v\|^2+\frac{v^2}{\|x\|^2}\right) dx$. Notice that $\mathcal H \subset H^1_0(\Omega)$ and the inclusion is strict (consider for instance the function $w(x)=1-\|x\|^2$).\ For $\eta, \xi\in\mathcal H$ we write $$\label{perpH}\eta\perp_{\mathcal H}\xi \quad\Leftrightarrow \quad \int_{B}\frac{\eta\xi}{\|x\|^2}dx=0.$$ Observe that if $\psi,\widetilde{\psi}\in\mathcal H$ are weak solutions to related respectively to the eigenvalues $\beta$ and $\widetilde{\beta}$, $\beta\neq\widetilde{\beta}$ then $$\label{ortogonalita_auto_deboli} \psi\perp_{\mathcal H}\widetilde{\psi}$$ (just multiply by $\widetilde{\psi}$, the equation for the eigenvalue $\widetilde{\beta}$ by $\psi$, integrate and subtract). We define $$\begin{aligned} \label{CourantCharEigenvpesatipalla1} \beta_1(p) &:=& \inf_{\substack{v\in \mathcal H,\,\,v\neq 0}}\ \ \ \widetilde{R_p}[v]\end{aligned}$$ where $\widetilde{R_p}[v]$ is the Rayleigh quotient $$\label{Rayleigh2} \widetilde{R_p}[v]:=\frac{Q_p(v)}{\int_B \frac{v(x)^2}{\|x\|^2} dx}$$ and $Q_p$ is as in . From [@GGN2 Proposition 2.1] we know that when $\beta_1(p)<0$ then this infimum is achieved at a radial function $\psi_1 \in \mathcal H$, $\psi_1>0$ in $B\setminus\{0\}$, which solves $$\label{problemaPesatoPallaDebole1} \int_B\nabla\psi_1\nabla v-p\|u_p\|^{p-1}\psi_1v\, dx= \beta_1(p)\int_B \frac{\psi_1 v}{\|x\|^2}\, dx, \qquad \forall v\in \mathcal H.$$ Moreover $\beta_1(p)$ is simple (in $\mathcal H$). In this case we can then define $$\begin{aligned} \label{CourantCharEigenvpesatipalla2} \beta_2(p) &:=& \inf_{\substack{v\in \mathcal H,\,\,v\neq 0\\ v\perp_{\mathcal H}\psi_1}}\ \ \ \widetilde{R_p}[v]\end{aligned}$$ which again is achieved when it is negative (see [@GGN2 Proposition 2.3]) and any function $\psi_2\in\mathcal H$ at which $\beta_2(p)$ is achieved solves $$\label{problemaPesatoPallaDebole2} \int_B\nabla\psi_2\nabla v-p\|u_p\|^{p-1}\psi_2v\, dx= \beta_2(p)\int_B \frac{\psi_2 v}{\|x\|^2}\, dx, \qquad \forall v\in \mathcal H,$$ and by definition $\psi_1\perp_{\mathcal H}\psi_2$, then $\psi_2$ must change sign.\ \ More in general, by iterating, if $\beta_j(p)<0$ and $\psi_j\in\mathcal H$ is a function where it is achieved, for $j=1,\ldots, i-1$, we can define $$\begin{aligned} \label{CourantCharEigenvpesatipalla3} \beta_i(p) &:=& \inf_{\substack{v\in \mathcal H,\,\,v\neq 0\\ v\perp_{\mathcal H}span\{\psi_1,\ldots,\psi_{i-1}\}}}\ \ \ \widetilde{R_p}[v],\qquad i\in{\mathbb{N}}, \ i\geq 2\end{aligned}$$ which (again [@GGN2 Proposition 2.3]) is achieved if it is negative and, in such a case, any function $\psi_i\in\mathcal H$ at which $\beta_i(p)$ is achieved solves $$\label{problemaPesatoPallaDebolei} \int_B\nabla\psi_i\nabla v-p\|u_p\|^{p-1}\psi_iv\, dx= \beta_i(p)\int_B \frac{\psi_i v}{\|x\|^2}\, dx, \qquad \forall v\in \mathcal H,$$ and changes sign. Similarly, restricting to the subspace $\mathcal H_{{{\text{\upshape rad}}}}$ of the radial functions of $\mathcal H$, we can also define: $$\begin{aligned} \label{CourantCharEigenvpesatipallaRADIALE1} \beta_{1,{{\text{\upshape rad}}}}(p) &:=& \inf_{\substack{v\in \mathcal H_{{{\text{\upshape rad}}}},\,\,v\neq 0}}\ \ \ \widetilde{R_p}[v] \ (= \beta_1(p))\end{aligned}$$ and, if $\beta_{j,{{\text{\upshape rad}}}}(p)<0$ for $j=1,\ldots, i-1$ $$\begin{aligned} \label{CourantCharEigenvpesatipallaRADIALE} \beta_{i,{{\text{\upshape rad}}}}(p) &:=& \inf_{\substack{v\in \mathcal H_{{{\text{\upshape rad}}}},\,\,v\neq 0\\ v\perp_{\mathcal H}span\{\phi_{1},\ldots,\phi_{i-1}\}}}\ \ \ \widetilde{R_p}[v],\qquad i\in{\mathbb{N}}, \ i\geq 2\end{aligned}$$ where $\phi_j\in\mathcal H_{{{\text{\upshape rad}}}}$ is the function where $\beta_{j,{{\text{\upshape rad}}}}(p)$ is achieved for $j=1,\ldots, i-1$ (observe that $\phi_1=\psi_1$) and solve $$\label{problemaPesatoPallaDeboleiRadiale} \int_B\nabla\phi_j\nabla v-p\|u_p\|^{p-1}\phi_jv\, dx= \beta_{j,{{\text{\upshape rad}}}}(p)\int_B \frac{\phi_j v}{\|x\|^2}\, dx, \qquad \forall v\in \mathcal H_{{{\text{\upshape rad}}}}.$$ \[valoriVariaz=AutovaloriesatoPalla\] The negative eigenvalues (resp. negative radial eigenvalues) of problem coincide with the negative numbers $\beta_i(p)$’s defined in - (resp. with the numbers $\beta_{i,{{\text{\upshape rad}}}}(p)$’s defined in -). Moreover, by , the corresponding eigenfunctions, which solve , are in $\mathcal H$ and can be chosen to be orthogonal in the sense of . The following relation holds between the Morse index of $u_p$ and the number of negative eigenvalues of the weighted problem : \[lemma:Morse=numeroAutovaloriesatoPalla\] The Morse index (resp. radial Morse index) of $u_p$ coincides with the number of negative eigenvalues (resp. negative radial eigenvalues) of problem counted according to their multiplicity. As a consequence we have: \[2AutovRadNegativiPesatoPalla\] For any $p>1$ $$\beta_{1,{{\text{\upshape rad}}}}(p)<\beta_{2,{{\text{\upshape rad}}}}(p)<0.$$ Moreover $\beta_{3,{{\text{\upshape rad}}}}(p)=0$ and it is not an eigenvalue for . The first statement is a consequence of Lemma \[LemmaMorseIndexRadiale\] and Lemma \[lemma:Morse=numeroAutovaloriesatoPalla\].\ Observe that the value $\beta_{3,{{\text{\upshape rad}}}}(p)$ is well defined by , being both $\beta_{1,{{\text{\upshape rad}}}}(p)$ and $\beta_{2,{{\text{\upshape rad}}}}(p)$ negative, moreover $\beta_{3,{{\text{\upshape rad}}}}(p)\geq 0$ from Lemma \[valoriVariaz=AutovaloriesatoPalla\] and Lemma \[lemma:Morse=numeroAutovaloriesatoPalla\], since $m_{{{\text{\upshape rad}}}}(u_p)=2$ by Lemma \[LemmaMorseIndexRadiale\]. In particular even if $\beta_{3,{{\text{\upshape rad}}}}(p)=0$ it cannot be an eigenvalue for because $\mathcal H\subset H^1_0(B)$ and $u_p$ is radially nondegenerate by Lemma \[lemma:radiallyNonDeg\].\ To show that $\beta_{3,{{\text{\upshape rad}}}}(p)=0$ we let $\phi_j\in \mathcal H_{{{\text{\upshape rad}}}}$ be the function where $\beta_{j,{{\text{\upshape rad}}}}(p)$ is achieved for $j=1,2$, we choose the test functions $$\eta_{{\varepsilon}}(x):= \begin{cases} 1-\|x\|&\hbox{if }{\varepsilon}\le\|x\|\le1\\ \frac{2(1-{\varepsilon})}{\varepsilon}\|x\|+{\varepsilon}-1&\hbox{if }\frac{\varepsilon}2\le\|x\|\le{\varepsilon}\\ 0&\hbox{if }\|x\|\le\frac{\varepsilon}2 \end{cases}$$ defined for $0<{\varepsilon}<1$ and we let $$\widetilde\eta_{\varepsilon}(x):=\eta_{\varepsilon}(x)-a_{\varepsilon}\phi_1-b_{\varepsilon}\phi_2$$ where $a_{\varepsilon},b_{\varepsilon}\in {\mathbb{R}}$ are given by $$a_{\varepsilon}:=\frac{\int_B \frac{\eta_{\varepsilon}\phi_1} {\|x\|^{2}}}{\int_B \frac{\phi_1^2}{\|x\|^{2}} }\quad , \quad b_{\varepsilon}:=\frac{\int_B \frac{\eta_{\varepsilon}\phi_2}{\|x\|^{2}}}{\int_B \frac{\phi_2^2}{\|x\|^{2}}}$$ so that $\widetilde\eta_{\varepsilon}$ is orthogonal in the sense of to $\phi_j$, $j=1,2$ for any ${\varepsilon}\in (0,1)$.\ Moreover observe that by our choice of the test functions $\eta_{\varepsilon}$ there exists $C=C_p>0$ such that $$\label{etaepUnifBounded} \int_B \left(\|\nabla \eta_{\varepsilon}\|^2-p\|u_p\|^{p-1}\eta_{\varepsilon}^2\right)\leq C,$$ for any ${\varepsilon}\in (0,1)$.\ Since $\beta_{j,{{\text{\upshape rad}}}}(p)<0$ for $j=1,2$, by Proposition \[p2.2\] we have that $$\label{numero} \phi_j(r)=O\left( r^{\sqrt {-\beta_{j,{{\text{\upshape rad}}}}(p)}}\right) \quad \text{ as }r\rightarrow 0.$$ This last estimate together with the definition of $\eta_{\varepsilon}$ then implies that $$\begin{split} \int_0^1 \frac { \eta_{\varepsilon}\phi_j}r\, dr=&\frac {2(1-{\varepsilon})} {{\varepsilon}}\int_{\frac {{\varepsilon}}2}^{{\varepsilon}} \phi_j(r)\, dr+({\varepsilon}-1)\int_{\frac {{\varepsilon}}2}^{{\varepsilon}}\frac { \phi_j(r)}r\, dr+\int_{{\varepsilon}}^1\frac{(1-r)\phi_j(r)}r\, dr\\ &\hspace{-4em}\overset{\eqref{numero}} {\leq} C+O\left({\varepsilon}^{\sqrt {-\beta_{j,{{\text{\upshape rad}}}}(p)}}\right)\leq C \end{split}$$ so that $a_{\varepsilon}$ and $b_{\varepsilon}$ are uniformly bounded. From and the orthogonality between $\widetilde\eta_{\varepsilon}$ and $\phi_j$, $j=1,2$ then $\beta_{3,{{\text{\upshape rad}}}}(p)\leq \widetilde{R_p}[\widetilde\eta_{\varepsilon}]$ where $$\label{R_peta-tilde} \widetilde{R_p}[\widetilde\eta_{\varepsilon}]=\frac{Q_p(\widetilde\eta_{\varepsilon})}{\int_B \frac{\widetilde\eta_{\varepsilon}^2}{\|x\|^2} dx}.$$ An easy computation shows that $$\begin{aligned} &Q_p(\widetilde\eta_{\varepsilon})=&\int_B \left(\|\nabla \eta_{\varepsilon}\|^2-p\|u_p\|^{p-1}\eta_{\varepsilon}^2\right)+a_{\varepsilon}^2\int_B\left(\|\nabla \phi_1\|^2 -p\|u_p\|^{p-1}\phi_1^2\right)\\ &&+b_{\varepsilon}^2\int_B\left(\|\nabla \phi_2\|^2 -p\|u_p\|^{p-1}\phi_2^2\right)-2a_{\varepsilon}\int_B\left(\nabla \eta_{\varepsilon}\cdot \nabla \phi_1-p\|u_p\|^{p-1}\eta_{\varepsilon}\phi_1\right)\\ &&-2b_{\varepsilon}\int_B\left(\nabla \eta_{\varepsilon}\cdot \nabla \phi_2-p\|u_p\|^{p-1}\eta_{\varepsilon}\phi_2\right) -2a_{{\varepsilon}}b_{\varepsilon}\int_B\left(\nabla \phi_1\cdot \nabla \phi_2-p\|u_p\|^{p-1}\phi_1\phi_2\right)\end{aligned}$$ and, using that $\phi_j$, $j=1,2$ solves , that $\phi_1\perp_{\mathcal H}\phi_2$ and recalling the definition of $a_{\varepsilon},b_{\varepsilon}$, we then get $$Q_p(\widetilde\eta_{\varepsilon})= \int_B \left(\|\nabla \eta_{\varepsilon}\|^2-p\|u_p\|^{p-1}\eta_{\varepsilon}^2\right)-a_{\varepsilon}^2\beta_{1,{{\text{\upshape rad}}}}(p)\int_B\frac{\phi_1^2}{\|x\|^2}-b_{\varepsilon}^2\beta_{2,{{\text{\upshape rad}}}}(p)\int_B\frac{\phi_2^2}{\|x\|^2}.$$ The last equality, together with and the boundedness of $a_{\varepsilon},b_{\varepsilon}$ implies that $$Q_p(\widetilde\eta_{\varepsilon})\leq C$$ for any ${\varepsilon}\in (0,1)$. Finally, using again the definition of $a_{\varepsilon},b_{\varepsilon}$ we have $$\begin{aligned} \int_B \frac{\widetilde\eta_{\varepsilon}^2}{\|x\|^2} dx &= &\int_B \frac{\eta_{\varepsilon}^2}{\|x\|^2} -a_{\varepsilon}^2 \int_B\frac{\phi_1^2}{\|x\|^2}-b_{\varepsilon}^2\int_B\frac{\phi_2^2}{\|x\|^2}\\ &\overset{\mbox{$a_{\varepsilon}, b_{\varepsilon}$ bounded}}{\geq} & \int_B \frac{\eta_{\varepsilon}^2}{\|x\|^2}-C\\ &= & 2\pi \left(\frac{(1-{\varepsilon})^2}{{\varepsilon}^2} \int_{\frac {{\varepsilon}}2}^{{\varepsilon}}\frac {(2r-{\varepsilon})^2}r \, dr+\int_{{\varepsilon}}^1\frac {(1-r)^2}r\, dr\right)-C\\ &= & 2\pi\left(-\log {\varepsilon}+{\varepsilon}\log 2+(1-{\varepsilon})({\varepsilon}-2)\right)-C\\ &=& -2\pi \log {\varepsilon}\left(1+o(1)\right) \quad \text{ as }{\varepsilon}\rightarrow 0.\end{aligned}$$ The conclusion then follows using and $0\leq \beta_{3,{{\text{\upshape rad}}}}(p)\leq \widetilde{R_p}[\widetilde\eta_{\varepsilon}]$. Here and in the following we denote by $\alpha_k$, $k\in\mathbb N$ the [spherical harmonics]{} in dimension $2$, namely the homogeneous harmonic polynomials of degree $k$ considered on the unit sphere $S^1\subset\mathbb R^2$. They can be written explicitly, using the polar coordinates $x=(r\cos\theta, r\sin\theta)$ $$\label{espressioneArmonicheSferiche} \alpha_k(\theta)=\left\{\begin{array}{ll}c & k=0 \\c_1\cos (k\theta)+c_2\sin (k\theta) & k= 1,2,3,\ldots\end{array} \right.$$ for $c,c_1,c_2\in\mathbb R$. Recall that the set $(\alpha_k)_{k\in\mathbb N}$ is a complete orthogonal system for $L^2(S^1)$, hence any function $v\in L^2(B)$ can be written as $$\label{spectralDec1} v(r,\theta)=\sum_{k=0}^{+\infty}h_k(r)\alpha_k(\theta)$$ where $$\label{spectralDec2} h_k(r):=\int_0^{2\pi}\alpha_k(\theta)v(r,\theta)d\theta, \qquad r\in (0,1).$$ Moreover if $v(r,\theta)$ is continuous in the origin, then $2 \pi c v(0)=h_0(0)$ (where $c$ is the constant in ) and $$\label{h_kInZero} h_k(0)=0, \quad \forall k\geq1.$$ Recall also that the eigenvalues of the Laplace-Beltrami operator $-\Delta_{S^1}$ on the unit sphere $S^1$ are the numbers $k^2$, $k\in\mathbb N$, that they have multiplicity $1$ if $k=0$ and multiplicity $2$ if $k\geq 1$, and that the spherical harmonics $\alpha_k$ are the eigenfunctions associated to the eigenvalue $k^2$. For the negative eigenvalues of we then have the following spectral decomposition into radial and angular part, where the angular part is given by the eigenvalues of $-\Delta_{S^1}$: \[lemma:decomposizionePalla\] Let $p>1$. For any $i=1,\ldots, m(u_p)$ there exists $(j,k)\in \{1,2\}\times\mathbb N$ ($(j,k)$ depending also on $p$) such that $$\label{spectralDecomposition} \beta_i(p)= \beta_{j,{{\text{\upshape rad}}}}(p)+k^2.$$ Conversely for every $(j,k)\in \{1,2\}\times\mathbb N$ such that $\beta_{j,{{\text{\upshape rad}}}}(p)+k^2<0$ there exists $i\in\{1,\ldots, m(u_p)\}$ ($i$ depending also on $p$) for which holds.\ Moreover the eigenspace associated to each negative eigenvalue $\beta(p)$ of is spanned by the functions $$\label{eigenspaceSpectralDec} \phi_j(r)\cos (k\theta)\ \mbox{ and }\ \phi_j(r)\sin (k\theta),\quad \forall\, (j,k) \mbox{ such that }\ \beta_{j,{{\text{\upshape rad}}}}(p)+k^2=\beta(p),$$ where $\phi_j$ is the radial eigenfunction to associated to the radial eigenvalue $\beta_{j,{{\text{\upshape rad}}}}(p)$ (which is simple in the space of radial functions). [Step 1.]{} [We show the first statement.]{}\ By Lemma \[valoriVariaz=AutovaloriesatoPalla\] and Lemma \[lemma:Morse=numeroAutovaloriesatoPalla\] the value $\beta_i(p)$, for any $i=1,\ldots, m(u_p)$, is a (negative) eigenvalue for problem and so there exists a function $\psi\neq 0$ which satisfies with $\beta=\beta_i(p)$. Decomposing $\psi$ along spherical harmonics (see , ), we write $$\psi(r,\theta)=\sum_{k=0}^{+\infty} h_k(r)\alpha_k(\theta)$$ where $$\label{decompos} h_k(r):=\int_0^{2\pi}\alpha_k(\theta)\psi(r,\theta)d\theta, \qquad r\in (0,1).$$ Then, since $\psi\neq 0$ and $(\alpha_k)_k$ is a complete orthogonal system for $L^2(S^1)$, it follows that $h_k\neq 0$ for some $k\in\mathbb N$, moreover it satisfies $$\begin{aligned} -h_k''-\frac{1}{r}h_k'&=& \int_0^{2\pi}\left(-\psi_{rr}-\frac{1}{r}\psi_r\right)\alpha_k \,d\theta \\ &=&\int_{0}^{2\pi} \left(-\Delta\psi+\frac{1}{r^2}\Delta_{S^1}\psi\right)\alpha_k\,d\theta \\ &=& p \|u_p\|^{p-1}h_k +\frac{\beta_i(p)}{r^2} h_k+ \frac{1}{r^2}\int_0^{2\pi} \left(\Delta_{S^1}\psi\right) \alpha_k\,d\theta.\end{aligned}$$ Integrating the last term by parts we get $$\label{A} \begin{cases} -h_k''-\frac 1r h_k'-p\|u_p\|^{p-1}h_k =\frac{\beta_i(p)-k^2}{r^2} h_k & \text{ in }(0,1)\\ h_k(1)=0, \end{cases}$$ where $\beta_i(p)-k^2\leq \beta_i(p)<0$. Next we show that it satisfies also the condition $$\label{radsphint} \int_0^1 r(h_k')^2+\frac{h^2_k}{r}<+\infty.$$ Indeed using we get $$\begin{aligned} \label{primop} \int_0^1 \frac{h_k(r)^2}{r}\; dr &=&\int_0^1 \frac 1 r \left(\int_{0}^{2\pi}\alpha_k(\theta)\psi(r,\theta) \;d \theta\right)^2 dr\\\nonumber &\overset{\text{Jensen ineq.}}{\leq}& \int_0^1\frac 1 r\int_{0}^{2\pi}\alpha^2_k(\theta)\psi^2(r,\theta) \; d \theta\, dr\\ \nonumber &\overset{\text{ $\alpha_k$ are bounded}}{\leq}& C\int_0^1\int_0^{2\pi} \frac{\psi^2(r,\theta)}{r^2} r\; dr\, d\theta = C\int_B \frac{\psi^2(x)}{\|x\|^2}<\infty,\end{aligned}$$ where last estimate follows from . In the same way we obtain $$\begin{aligned} \label{secondop} \int_0^1r\left(h'_k(r)\right)^2dr &=&\int_0^1 r \left(\int_{0}^{2\pi}\alpha_k(\theta)\frac {\partial \psi(r,\theta)}{\partial r}d \theta\right)^2dr\\\nonumber &\leq & C\int_0^1\int_0^{2\pi}r \left\|\frac {\partial \psi(r,\theta)}{\partial r}\right\|^2 dr d\theta\leq C\int_B\|\nabla \psi(x)\|^2 dx<\infty,\end{aligned}$$ showing .\ By Lemma \[valoriVariaz=AutovaloriesatoPalla\], Lemma \[lemma:Morse=numeroAutovaloriesatoPalla\] and Lemma \[2AutovRadNegativiPesatoPalla\] problem - admits only two negative eigenvalues which coincide with $\beta_{1,{{\text{\upshape rad}}}}(p)$ and $\beta_{2,{{\text{\upshape rad}}}}(p)$. Then - has a nontrivial solution $h_k$ (related to a negative eigenvalue) if and only if $\beta_{j,{{\text{\upshape rad}}}}(p)=\beta_i(p)-k^2$ for some $j=1,2$. This ends the proof of the existence of $(j,k)\in \{1,2\}\times \mathbb N$ which satisfies .\ Let $(j,k)\in \{1,2\}\times\mathbb N$ be such that $\beta_{j,{{\text{\upshape rad}}}}(p)+k^2<0$, let $\phi_j$ be an eigenfunction associated to the radial eigenvalue $\beta_{j,{{\text{\upshape rad}}}}(p)$ (which is simple in the space of the radial functions) and $\alpha_k$ be an eigefunction of $-\Delta_{S^1}$ associated to the eigenvalue $k^2$ (see ). Then easy computation shows that the number $\beta_{j,{{\text{\upshape rad}}}}(p)+k^2$ is a negative eigenvalue for the weighted problem with eigenfunction given by $$\label{moltiplicazione} \psi_{j,k}(x):=\phi_j(\|x\|)\alpha_k(\frac{x}{\|x\|}).$$ As a consequence, by Lemma \[valoriVariaz=AutovaloriesatoPalla\] and Lemma \[lemma:Morse=numeroAutovaloriesatoPalla\], there exists $i\in\{1,\ldots, m(u_p)\}$ for which holds. \ Let $m\in{\mathbb{N}}_0$ be the multiplicity of $\beta(p)$, so there exists an index $\ell\in{\mathbb{N}}$, $\ell\geq 1$ such that $$\beta(p)=\ \beta_{\ell}(p)=\beta_{\ell+1}(p)=\cdots\beta_{\ell+m-1}(p)\ < \beta_{\ell+m}(p)$$ and if $\ell \geq 2$ also $$\beta_{\ell-1}(p)<\ \beta(p)$$ ($m$ is the number of subsequent indexes $i$ in our notation).\ By [Step 1.]{} for every $i=\ell,\ldots, \ell+m-1$ there exists a couple $(j,k)\in \{1,2\}\times{\mathbb{N}}$ for which holds (some of the couples may coincide).\ Then considering the set $$\mathcal I:=\{(j,k)\in\{1,2\}\times{\mathbb{N}}\ :\ \beta_i(p)=\beta(p)=\beta_{j,{{\text{\upshape rad}}}}(p)+k^2,\quad i=\ell,\ldots \ell +m\},$$ as seen in [Step 2.]{} all the functions in with $(j,k)\in\mathcal I$ are eigenfunctions for . Observe that since $\beta_{j,{{\text{\upshape rad}}}}(p)$ is simple in the space of radial functions and $\alpha_k$ are the functions in one obtains all the functions in , which are linearly independent.\ Last we prove by contradiction that the eigenspace of $\beta(p)$ consists only of the functions in . So let us assume the existence of another eigenfunction $\psi\neq 0$, $$\label{orto} \psi\perp_{\mathcal H} \mbox{ span}\left\{\phi_j(r)\cos (k\theta),\ \phi_j(r)\sin (k\theta)\ :\ (j,k) \in\mathcal I\right\},$$ then similarly as in [Step 1.]{} we can write $$\label{expa} \psi(r,\theta)=\sum_{s=0}^{+\infty} h_s(r)\alpha_s(\theta) $$ where $$h_s(r):=\int_0^{2\pi}\alpha_s(\theta)\psi(r,\theta)d\theta, \qquad r\in (0,1).$$ Since $\psi\neq 0$ then there exists $s\in{\mathbb{N}}$ such that $h_s\neq 0$. Then, as in [Step 1.]{} we can prove that for any $s$ such that $h_s\neq 0$ there exists $t_s\in \{1,2\}$ such that $$\label{PerAssurdo} \beta(p)=\beta_{t_s,{{\text{\upshape rad}}}} +s^2\quad \mbox{ and }\quad h_s=\phi_{t_s}.$$ As a consequence becomes $$\psi(r,\theta)=\sum_{s=0, \ h_s\neq 0}^{+\infty} \phi_{t_s}(r)\alpha_s(\theta)$$ and so the orthogonality condition gives $$0=\sum_{s=0}^{\infty} \int_0^{1}\frac{\phi_{t_s}\phi_j}{r} dr\ \int_{0}^{2\pi}\alpha_s\alpha_k d\theta=\sum_{s=0, \ h_s\neq 0}^{+\infty} \delta_{t_s,j}\delta_{s,k}, \qquad \forall (j,k)\in\mathcal I.$$ As a consequence, for any $(j,k)\in\mathcal I$ either $s\neq k$ or if $s=k$ then necessarily $t_s\neq j$, namely the couple $(t_s,s)\not\in\mathcal I$. Since holds this contradicts the definition of the set $\mathcal I$. Morse index and characterization of the degeneracy of $u_p$ {#subsec:mo} ----------------------------------------------------------- In the next result we estimate the two negative radial eigenvalues of the auxiliary weighted eigenvalue problem from above and from below by consecutive eigenvalues of $-\Delta_{S^1}$. The proof is based on the approximation of problem by a family of weighted eigenvalue problems in annuli already studied in [@DeMarchisIanniPacellaMathAnn], in particular we exploit some previous estimates in [@DeMarchisIanniPacellaMathAnn] related to the negative radial eigenvalues to this family of approximating auxiliary problems. As a consequence of our estimates we also get that the Morse index of $u_p$ is even for any $p>1$ and uniformly bounded in $p$. Moreover the estimate of the two negative radial eigenvalues of is the starting point to characterize the degeneracy of $u_p$, this last result is contained in Proposition \[p4.7\] at the end of the section. \[proposition:autovaloriRadiaiGenerale\] $$\label{nu2controllato} -1\leq \beta_{2,{{\text{\upshape rad}}}}(p)<0 \quad \forall p>1.$$ For any $p>1$ there exists a unique $j=j(p)\in\mathbb N$, $j\geq 2$ such that $$\label{nu1controllato} -j^2\leq \beta_{1,{{\text{\upshape rad}}}}(p)<-(j-1)^2$$ and $$\label{MorsePari} m(u_p)=2j$$ Moreover $j(p)\leq C$ for any $p>1$, where the constant $C>0$ does not depend on $p$. Let us consider the set $$\label{def:anello} A_n:=\{x\in \mathbb R^2\ :\ \frac{1}{n}<\|x\|<1 \}, \quad n\in \mathbb N_0$$ and the weighted eigenvalue problem $$\label{problemaPesatoAnello} \begin{cases} \begin{array}{ll} -\Delta \psi - p\|u_p(x)\|^{p-1} \psi =\frac{\beta}{\|x\|^2} \psi\qquad & \text{ in } A_n, \\ \psi= 0 & \text{ on } \partial A_n \end{array} \end{cases}$$ and let us denote by $$\beta_i^n(p), \quad \beta_{i,{{\text{\upshape rad}}}}^n(p), \quad i\in{\mathbb{N}}_0$$ its eigenvalues, counted with their multiplicity, and the radial eigenvalues, which are simple in the space of radial functions, respectively. Clearly the following characterizations hold $$\begin{aligned} \label{defmutilde1n} \beta_i^n(p)&=& \min_{\substack{ V\subset H^1_{0}(A_n)\\ dim V=i}} \max_{\substack{v\in V\\v\neq 0}} \widetilde{R^n_p}[v]\end{aligned}$$ $$\begin{aligned} \label{defbetatilde1n} \beta_{i,{{\text{\upshape rad}}}}^n(p)&=& \min_{\substack{ V\subset H^1_{0,rad}(A_n)\\ dim V=i}} \max_{\substack{v\in V\\v\neq 0}} \widetilde{R^n_p}[v]\end{aligned}$$ where $\widetilde{R^n_p}$ is the Rayleigh quotient $$\label{Rayleigh3} \widetilde{R^n_p}[v]:=\frac{Q^n_p(v)}{\int_{A_n} \frac{v(x)^2}{\|x\|^2} dx}$$ and $Q^n_p(v):=\int_{A_n}\left(\|\nabla v(x)\|^2-p\|u_p(x)\|^{p-1}v(x)^2\right) dx$. [Step 1.]{} [We show that for any $p>1$ there exists a unique $j=j(p)\in\mathbb N$, $j\geq 2$ and $n_p\in\mathbb N_0$ such that $$\label{morsePari}m(u_p)=2j,$$ and for $n\geq n_p$ $$\label{posizionebetatilden1} -j^2\leq\beta_{1,{{\text{\upshape rad}}}}^n(p)<-(j-1)^2$$ $$\label{posizionebetatilden2} -1< \beta_{2,{{\text{\upshape rad}}}}^n(p)<0.$$* ]{} As already proved in [@DeMarchisIanniPacellaMathAnn Proposition 4.3] there exists $n_p'\in\mathbb N_0$ such that $$\label{morseradCoincide} 2\overset{Lemma \ \ref{LemmaMorseIndexRadiale}}{=} m_{{{\text{\upshape rad}}}}(u_p)=\#\{\mbox{negative eigenvalues $\beta_{i,{{\text{\upshape rad}}}}^n(p)$}\}, \quad\mbox{ for }n\geq n_p',$$ namely $$\label{ahah}\beta_{1,{{\text{\upshape rad}}}}^n(p)< \beta_{2,{{\text{\upshape rad}}}}^n(p)<0\leq \beta_{3,{{\text{\upshape rad}}}}^n(p)<\ldots, \quad\mbox{ for }n\geq n_p',$$ where the strict inequalities are due to the fact that the radial eigenvalues are simple in the space of the radial functions.\ From [@DeMarchisIanniPacellaMathAnn Proposition 4.5] we also know that for any $p>1$ there exists $n_p''\in\mathbb N_0$ such that $$\label{dip} \beta_{2,{{\text{\upshape rad}}}}^n(p)> -1,\ \ \mbox{ for any }n\geq n_p''.$$ Hence follows immediately from and . In order to conclude the proof observe that from [@DeMarchisIanniPacellaMathAnn Proposition 4.3] we also know that $$\label{morseCoincide} 4\overset{ Lemma \ \ref{LemaAftalionPacella}}{\leq } m(u_p)=\#\{\mbox{negative eigenvalues $\beta_i^n(p)$}\}, \quad\mbox{ for }n\geq n_p'.$$ We show that, as a consequence of , using - and a spectral decomposition for the eigenvalues $\beta_i^n(p)$ in $A_n$, necessarily also and hold. Indeed recall that for the eigenvalues in $A_n$ the spectral decomposition holds: $$\label{spectralDecompositionAnello} \beta_i^n(p)= \beta_{j,{{\text{\upshape rad}}}}^n(p)+k^2\quad i,j \in\mathbb N_0,\ k\in\mathbb N,$$ where as before $k^2$ are the eigenvalues of the Laplace-Beltrami operator on the unit sphere $S^1$, and the eigenfunctions $\psi$ associated to the eigenvalue $\beta_i^n(p)$ may be obtained by multiplying the spherical harmonics $\alpha_k$ associated to $k^2$ (given in ) with the radial $j$-th eigenfunction $\phi_j^n$ for problem , similarly as we already did in : $$\label{moltiplicazione2} \psi(x)=\phi_j^n(\|x\|)\alpha_k(\frac{x}{\|x\|}).$$ By and it follows that the modes $k$ that contributes to the Morse index of $u_p$ are those such that $$\label{spectralDecompositionAnelloModiCheContano} \beta_i^n(p)= \beta_{j,{{\text{\upshape rad}}}}^n(p)+k^2 <0,\qquad j=1,2.$$ The case $j=2$ in is possible only when $k=0$ by . Hence by and recalling that there is only $1$ spherical harmonic for $k=0$ (see ) we get only $1$ contribution to the Morse index in this case.\ The case $j=1$ gives instead $1$ contribution (for $k=0$) and moreover, by , it must also give other contribution ($k\geq 1$). As a consequence holds. Hence by and recalling that there are two spherical harmonics for $k\geq 1$, and only $1$ for $k=0$ (see ) we get in this case that the total contribution of $\beta_{1,{{\text{\upshape rad}}}}^n(p)$ to the Morse index is then $2(j-1)+1$.\ Summing up all the contributions from both $j=1$ and $j=2$ we get . [Step 2.]{} [We show that for any $p>1$ the sequence $(\beta_{i,{{\text{\upshape rad}}}}^n(p))_n$ is monotone non-increasing and $$\beta_{i,{{\text{\upshape rad}}}}(p)=\lim_{n\rightarrow +\infty}\beta_{i,{{\text{\upshape rad}}}}^n(p)=\inf_n \beta_{i,{{\text{\upshape rad}}}}^n(p), \quad i=1,2.$$ ]{} The monotonicity of $(\beta_{i,{{\text{\upshape rad}}}}^n(p))_n$ follows by the variational characterization of these eigenvalues and by the canonical embeddings $H^1_0(A_{n})\subset H^1_0(A_{n+1})$, $\forall n\in{\mathbb{N}}_0$.\ By the monotonicity of $(\beta_{i,{{\text{\upshape rad}}}}^n(p))_n$ we can define the values $$\label{deflim}\delta_i(p):=\lim_{n\rightarrow +\infty} \beta_{i,{{\text{\upshape rad}}}}^n(p)=\inf_n \beta_{i,{{\text{\upshape rad}}}}^n(p)\overset{\mbox{{\scriptsize{{\bf Step 1}}}}}{<}0, \quad i=1,2.$$ Then the proof of [Step 2.]{} consists in proving that $$\label{tesiSTEP2} \delta_i(p)=\beta_{i,{{\text{\upshape rad}}}}(p), \quad i=1,2.$$ Let $\widetilde{\phi _i^n}$ be the radial eigenfunction of problem corresponding to the radial eigenvalue $\beta_{i,{{\text{\upshape rad}}}}^n(p)$ and normalized so that $$\label{normalizzazione} \int_{A_n} \left(\widetilde{\phi _i^n}\right)^2 =1.$$ Then, extending $ \widetilde{\phi _i^n}$ to zero in $B\setminus A_n$, from we have $$\label{deltaip}\int _{B} \frac{\left(\widetilde{\phi _i^n}\right)^2}{\|x\|^2}>\int_{B} \left(\widetilde{\phi _i^n}\right)^2 =1.$$ By [Step 1.]{} we know in particular that $\beta_{i,{{\text{\upshape rad}}}}^n(p)<0$, so from we get $$\label{betaiequ} \int_{B} \| \nabla \widetilde{\phi _i^n}\|^2-p \int_{B}\|u_p\|^{p-1} \left(\widetilde{\phi _i^n}\right)^2=\beta_{i,{{\text{\upshape rad}}}}^n(p)\int _{B} \frac{\left(\widetilde{\phi _i^n}\right)^2}{\|x\|^2}< 0,$$ from which it follows that the sequence $(\widetilde{\phi _i^n})_n$ is bounded in $H^1_0(B)$, indeed $$\label{c} \int_{B} \| \nabla \widetilde{\phi _i^n}\|^2<p \int_{B}\|u_p\|^{p-1} \left(\widetilde{\phi _i^n}\right)^2\leq p\\|u_p\\|_{\infty}^{p-1}\int_{B}\left(\widetilde{\phi _i^n}\right)^2\overset{\eqref{normalizzazione}}{\leq} C_p.$$ Hence, up to a subsequence, $\widetilde{\phi _i^n}\rightarrow \widetilde{\phi_i}$ as $n\rightarrow +\infty$ weakly in $H^1_0(B)$, strongly in $L^2(B)$ and pointwise a.e. in $B$ and then as a consequence $$\label{stainH10} \int_{B} \| \nabla {\widetilde{\phi_i}}\|^2\leq \liminf_{n\rightarrow +\infty}\int_{B} \| \nabla \widetilde{\phi _i^n}\|^2\overset{\eqref{c}}{\leq} C_p$$ and $$\label{convergenzapartel2} p\int_{B}\|u_p\|^{p-1} \left({\widetilde{\phi_i}}\right)^2=\lim_{n\rightarrow +\infty} p \int_{B}\|u_p\|^{p-1} \left(\widetilde{\phi _i^n}\right)^2.$$ $\widetilde{\phi_i}$ is radial and moreover, since $\int_{B} \left({\widetilde{\phi_i}}\right)^2 =1$, we have $$\label{nonZ}\widetilde{\phi_i}\neq 0.$$ Furthermore, since $\widetilde{\phi _1^n}\geq 0$, we also have $$\label{nonegp}\widetilde{\phi_1}(x)\geq 0 \mbox{ in }B.$$ Moreover there exists $C>0$, independent on $n\in\mathbb N$, such that $$\label{normpesfin} \int_B \frac{\left({\widetilde{\phi_i}}\right)^2}{\|x\|^2}\overset{\mbox{\footnotesize{Fatou}}}{\leq} \liminf_{n\rightarrow +\infty} \int _{B} \frac{\left(\widetilde{\phi _i^n}\right)^2}{\|x\|^2}\leq C,$$ indeed, if by contradiction we have that $\int _{B} \frac{\left(\widetilde{\phi _i^n}\right)^2}{\|x\|^2}\rightarrow +\infty$, then by and we derive $$\delta_i(p)=\lim_{n\rightarrow +\infty} \beta_{i,{{\text{\upshape rad}}}}^n(p)\overset{\eqref{betaiequ}}{=}\lim_{n\rightarrow +\infty}\frac{\int_{B} \| \nabla \widetilde{\phi _i^n}\|^2-p \int_{B}\|u_p\|^{p-1} \left(\widetilde{\phi _i^n}\right)^2}{\int _{B} \frac{\left(\widetilde{\phi _i^n}\right)^2}{\|x\|^2}}=0,$$ which is in contradiction with . Observe that by the bounds in and and the estimate in we also get that $\delta_i(p)>-\infty$. By and it follows that $$\label{limInH} \widetilde{\phi_i}\in\mathcal H_{{{\text{\upshape rad}}}}.$$ Moreover implies that, up to a subsequence, $\widetilde{\phi _i^n}$ converges to $\widetilde{\phi_i}$ also weakly in $L^2_{\frac 1{\|x\|^2}}(B)$. Multiplying by $\varphi\in \mathcal{H}$ and integrating on $A_n$, we have $$\label{d} \int_{A_n}\nabla \widetilde{\phi _i^n}\cdot \nabla \varphi -\int_{\|x\|=\frac{1}{n}}\left(\nabla \widetilde{\phi _i^n}\cdot \nu \right) \varphi-p \int_{A_n} \|u_p\|^{p-1}\widetilde{\phi _i^n}\varphi=\beta_{i,{{\text{\upshape rad}}}}^n(p)\int_{A_n} \frac{\widetilde{\phi _i^n}\varphi}{\|x\|^2}$$ (where the only boundary term is the one on the interior part of $\partial A_n$ since $\varphi(x)=0$ when $\|x\|=1$).\ Now by the weak convergence of $\widetilde{\phi _i^n}\rightarrow \widetilde{\phi_i}$ in $H^1_0(B)$ and in $L^2_{\frac 1{\|x\|^2}}(B)$, we have $$\label{term1} \int_{A_n}\nabla \widetilde{\phi _i^n}\cdot \nabla \varphi=\int_{B}\nabla \widetilde{\phi _i^n}\cdot \nabla \varphi\underset{n\rightarrow +\infty}{\longrightarrow} \int_{B}\nabla \widetilde{\phi_i}\cdot \nabla \varphi,$$ $$\label{term2} \int_{A_n} \|u_p\|^{p-1}\widetilde{\phi _i^n}\varphi=\int_{B} \|u_p\|^{p-1}\widetilde{\phi _i^n}\varphi\underset{n\rightarrow +\infty}{\longrightarrow} \int_{B} \|u_p\|^{p-1}\widetilde{\phi_i}\varphi$$ and $$\label{term3} \int_{A_n} \frac{\widetilde{\phi _i^n}\varphi}{\|x\|^2}=\int_{B} \frac{\widetilde{\phi _i^n}\varphi}{\|x\|^2}\underset{n\rightarrow +\infty}{\longrightarrow}\int_{B} \frac{\widetilde{\phi_i}\varphi}{\|x\|^2}.$$ Furthermore $$\label{convTermBordo} \int_{\|x\|=\frac{1}{n}}\left(\nabla \widetilde{\phi _i^n}\cdot \nu \right) \varphi \underset{n\rightarrow +\infty}{\longrightarrow} 0,$$ indeed $$\begin{aligned} \left\|\int_{\|x\|=\frac{1}{n}}\left(\nabla \widetilde{\phi _i^n}\cdot \nu \right) \varphi \right\| &\leq & \frac{1}{n}\int_{\|x\|=\frac{1}{n}} \|\nabla \widetilde{\phi _i^n}\|\frac{\|\varphi\|}{\|x\|} \leq \frac{1}{n}\int_{ A_n}\|\nabla \widetilde{\phi _i^n}\|\frac{\|\varphi\|}{\|x\|}\nonumber\\ &\overset{\mbox{{\footnotesize H\"older}}}{\leq} & \frac{1}{n}\left(\int_{ B}\|\nabla \widetilde{\phi _i^n}\|^2\right)^{\frac 12}\left(\int_B \frac{\|\varphi\|^2}{\|x\|^2}\right)^{\frac{1}{2}} \nonumber\\ &\overset{\eqref{c}}{\leq} & \frac{1}{n}\sqrt{C_p}\\|\varphi\\|_{\mathcal H}\underset{n\rightarrow +\infty}{\longrightarrow} 0.\end{aligned}$$ Passing to the limit into and using , , and , we get that $\widetilde{\phi_i}$ satisfies $$\label{equadelta} \int_{B}\nabla {\widetilde{\phi_i}}\cdot \nabla \varphi-p\int_{B} \|u_p\|^{p-1}{\widetilde{\phi_i}}\varphi=\delta_i(p)\int_{B} \frac{{\widetilde{\phi_i}}\varphi}{\|x\|^2}, \quad\mbox{ for any }\varphi \in \mathcal{H},$$ in particular by we can choose $\varphi=\widetilde{\phi_i}\neq 0$ by , so since $\delta_i(p)<0$ we have from that the quadratic form evaluated at $\widetilde{\phi_i}$ is negative $$Q_p(\widetilde{\phi_i})<0.$$ As a consequence, since $m_{{{\text{\upshape rad}}}}(u_p)=2$, it must be $$\widetilde{\phi_i}\in span\{\phi_1,\phi_2\}$$ where $0\neq \phi_j\in\mathcal H_{{{\text{\upshape rad}}}}\subset H^1_0(B)$ is the function where the negative weighted radial eigenvalue $\beta_{j,{{\text{\upshape rad}}}}(p)$ is achieved for $j=1,2$, which satisfies (and so $Q_p(\phi_j)<0$) and such that $\phi_1\geq 0$ and $\phi_1\perp_{\mathcal H}\phi_2$ (so $\phi_2$ changes sign).\ Choosing the test function $\varphi=\phi_j$, $j=1,2$ into and using the equation for $\phi_j$ we also get $$[\delta_i(p)-\beta_{j,{{\text{\upshape rad}}}}(p)]\int_B\frac{\widetilde{\phi_i}\phi_j}{\|x\|^2}=0\quad\mbox{ for }j=1,2,$$ hence the only possibility is that there exists $\alpha\in\mathbb R$ such that $$\label{alternativa}\mbox{ either }\ \left\{\begin{array}{lr}\widetilde{\phi_i}=\alpha\phi_1\\\delta_i(p)=\beta_{1,{{\text{\upshape rad}}}}(p)\end{array}\right. \ \mbox{ or }\ \left\{\begin{array}{lr}\widetilde{\phi_i}=\alpha\phi_2 \\\delta_i(p)=\beta_{2,{{\text{\upshape rad}}}}(p)\end{array}\right.$$ for $i=1,2$. By it follows that necessarily there exists $\alpha\in\mathbb R$ such that $$\left\{\begin{array}{lr}\widetilde\phi_1=\alpha \phi_1\\\delta_1(p)=\beta_{1,{{\text{\upshape rad}}}}(p).\end{array}\right.$$ Moreover and proved in [Step 1.]{} and the definition of $\delta_i(p)$ in imply that $$\delta_1(p)< -1 \leq \delta_2(p),$$ hence $\delta_1(p)\neq\delta_2(p)$ and so by necessarily there exists $\beta\in\mathbb R$ such that $$\left\{\begin{array}{lr}\widetilde\phi_2=\beta\phi_2 \\\delta_2(p)=\beta_{2,{{\text{\upshape rad}}}}(p)\end{array}\right.$$ which concludes the proof of .\ [Step 3.]{} [Conclusion.]{}\ is the same as in [Step 1]{}. Moreover passing to the limit in and in proved in [Step 1]{} and using the results of [Step 2]{} we obtain and respectively, where the strict inequalities are due to the monotonicity of the sequences $(\beta_{1,{{\text{\upshape rad}}}}^n(p))_n$ and $(\beta_{2,{{\text{\upshape rad}}}}^n(p))_n$.\ Last we show that there exists $C>0$ independent of $p$ such that $$\label{nuUnoLimitato} -C\leq \beta_{1,{{\text{\upshape rad}}}}(p)\ (< 0)\quad \mbox{ for any $p>1$}$$ from which the uniform bound on $j(p)$ then follows and this concludes the proof. Let $\phi_p\in\mathcal H$ be a function where $\beta_{1,{{\text{\upshape rad}}}}(p)$ is achieved, then by , choosing $v=\phi_p$, we have: $$\begin{aligned} 0\leq \int_{B}\|\nabla\phi_p(y)\|^2dy&=&\int_{B}p\|{u_p}(y)\|^{p-1}\phi_p(y)^2 dy+\beta_{1,{{\text{\upshape rad}}}}(p)\int_{B}\frac{\phi_p(y)^2}{\|y\|^2}dy\\ &=&\int_{B}\left(p\|{u_p}(y)\|^{p-1}\|y\|^2+\beta_{1,{{\text{\upshape rad}}}}(p)\right)\frac{\phi_p(y)^2}{\|y\|^2} dy\\ &\leq & \left[\max_{y\in B} \left(p\|{u_p}(y)\|^{p-1}\|y\|^2\right)+ \beta_{1,{{\text{\upshape rad}}}}(p)\right] \int_{B}\frac{\phi_p(y)^2}{\|y\|^2} dy,\end{aligned}$$ As a consequence $$\label{semifin} \beta_{1,{{\text{\upshape rad}}}}(p)\geq -\max_{y\in B} \left( p\|{u_p}(y)\|^{p-1}\|y\|^2\right).$$ We recall the following pointwise estimate for $u_p$ which has been proved in [@DeMarchisIanniPacellaJEMS]: $$\label{Q3} p\|u_p(x)\|^{p-1}\|x\|^2\leq C, \quad \forall p>1, \ \forall x\in B,$$ for a certain $C>0$ (see property $(P_3^k)$ in [@DeMarchisIanniPacellaJEMS Proposition 2.2], observing that in the radial case the origin is the only absolute maximum point of $\|u_p\|$ and that $k=1$ by [@DeMarchisIanniPacellaJEMS Proposition 3.6]). The conclusion follows combining with . \[Remark:convergenzaAutovaloriDAnelloAPalla\] In the proof of Lemma \[proposition:autovaloriRadiaiGenerale\] we have introduced the auxiliary weighted eigenvalue problems in the annuli $A_n$, $n\in\mathbb N_0$ and, as an intermediate step, we have shown that for any fixed $p>1$ the sequence $(\beta_{i,{{\text{\upshape rad}}}}^n(p))_n$ of the $i$-th radial eigenvalues for these problems is monotone non-increasing and $$\label{convergenzaAutovRadiali} \beta_{i,{{\text{\upshape rad}}}}(p)=\lim_{n\rightarrow +\infty} \beta_{i,{{\text{\upshape rad}}}}^n(p)=\inf_n \beta_{i,{{\text{\upshape rad}}}}^n(p), \quad i=1,2.$$ We stress that by the spectral decomposition in we also get the non-increasing monotonicity of the sequence of the $j$-th eigenvalues $(\beta_j^n(p))_n$ of the problems . Moreover combining , the spectral decomposition in and Lemma \[lemma:decomposizionePalla\] we also know the limit for the negative ones: $$\beta_j(p)=\lim_{n\rightarrow +\infty} \beta_j^n(p)= \inf_n \beta_j^n(p), \quad j=1,\ldots, m(u_p).$$ The auxiliary weighted eigenvalue problems in the annuli $A_n$, $n\in\mathbb N_0$ is the same already introduced and studied in [@DeMarchisIanniPacellaMathAnn] when computing the Morse index of $u_p$ for large $p$. Next we investigate the degeneracy of the solution $u_p$, for any $p>1$. This result will be useful to characterize the degeneracy of $u_p$ in the case of large $p$. Moreover we will need it to identify the possible bifurcation points and select the eigenfunctions related to them. \[p4.7\] For any $p\in (1, +\infty)$ let $j=j(p)\in\mathbb N$, $j\geq 2$ be as in Lemma \[proposition:autovaloriRadiaiGenerale\]. The solution $u_p$ is degenerate if and only if $$\label{1deg}\beta_{1,{{\text{\upshape rad}}}}(p)=-j^2$$ or $$\label{2deg}\beta_{2,{{\text{\upshape rad}}}}(p)=-1.$$ Moreover the space of the solutions to the linearized problem at a value $p$ that satisfies and/or is spanned by $$\label{autof1} v_{j}(r,\theta)=\phi_1(r) \left(A \sin(j\theta)+B\cos(j\theta)\right) \qquad A,B\in {\mathbb{R}}$$ and/or $$\label{autof2} v(r,\theta)=\phi_2(r) \left(A \sin(\theta)+B\cos(\theta)\right) \qquad A,B\in {\mathbb{R}},$$ where $\phi_1, \phi_2$ are the eigenfunctions associated to the first and second radial eigenvalue $\beta_{1,{{\text{\upshape rad}}}}(p)$, $\beta_{2,{{\text{\upshape rad}}}}(p)$ respectively.\ Hence $Ker (L_p)$ has dimension $0$ when neither nor hold, dimension $2$ when either or holds, and dimension $4$ when and hold. $u_p$ is degenerate if and only if there exists $v\in H^1_0(B)$, $v\neq 0$ such that $$\label{nucleoNonVuoto}\left\{ \begin{array}{ll} -\Delta v - p\|u_p\|^{p-1} v =0 \qquad & \text{ in } B, \\ v= 0 & \text{ on } \partial B. \end{array} \right.$$ [Step 1.]{} [We show that if $u_p$ is degenerate then or hold.]{}\ If $u_p$ is degenerate, problem admits a solution $v$ which is continuous in $B$ by elliptic regularity. Then we can decompose $v$ along spherical harmonics, namely for $k\in\mathbb N$ we consider the radial function $$\label{defhk} h_k(r):=\int_0^{2\pi}\alpha_k(\theta)v(r,\theta)\, d\theta,\qquad r\in [0,1)$$ where $\alpha_k$ is an eigefunction of $-\Delta_{S^1}$ associated to the eigenvalue $k^2$ (see —). Since $(\alpha_k)_k$ is a complete orthogonal system for $L^2(S^1)$ and $v\neq 0$, then necessarily $h_k\neq 0$ for some $k\in\mathbb N$. Moreover, similarly as in [Step 1]{} in the proof of Lemma \[lemma:decomposizionePalla\], it is easy to show that $h_k$, for these values of $k$, is a nontrivial solution to the problem $$\label{radsph} \left\{ \begin{array}{lr} -h_k''-\frac 1r h_k'-p\|u_p\|^{p-1}h_k =\frac{-k^2}{r^2} h_k & \text{ in }(0,1)\\ h_k(1)=0 \end{array}\right.$$ Observe that $k\geq 1$, since $u_p$ is radially nondegenerate by Lemma \[lemma:radiallyNonDeg\], so (see ), one has also $$\label{zeroh}h_k(0)=0.$$ Next we show that $h_k$ satisfies also the condition $$\int_0^1 r(h_k')^2+\frac{h^2_k}{r}<+\infty.$$ Indeed, since $v\in H^1_0(B)$, we can argue as in the proof of to get $$\label{radsphint1} \int_0^1 r(h_k')^2<+\infty$$ and moreover, using Proposition \[p2.2\], we also have that $h_k(r)=O(r^k)$, as $r {\rightarrow}0$, which implies $$\label{sommability} \int_0^1\frac{h^2_k}{r}<+\infty.$$ By Lemma \[valoriVariaz=AutovaloriesatoPalla\], Lemma \[lemma:Morse=numeroAutovaloriesatoPalla\] and Lemma \[2AutovRadNegativiPesatoPalla\] problem -- admits only two negative eigenvalues which coincide with $\beta_{1,{{\text{\upshape rad}}}}(p)$ and $\beta_{2,{{\text{\upshape rad}}}}(p)$. Hence we conclude that $h_k$ is nontrivial if and only if $\beta_{i,{{\text{\upshape rad}}}}(p)=-k^2$ for some $i=1,2$ and $k\geq 1$. The equalities and then follow remembering that, by Lemma \[proposition:autovaloriRadiaiGenerale\], $-1\leq\beta_{2,{{\text{\upshape rad}}}}(p)<0$ and $-j^2\leq \beta_{1,{{\text{\upshape rad}}}}(p)<-(j-1)^2$ for some $j=j(p)\in{\mathbb{N}}$, $j\geq 2$.\ [Step 2.]{} [We show that if or hold then $u_p$ is degenerate.]{}\ Let $$\label{nuclei} v_{i,k}(x):=\phi_i(\|x\|)\alpha_k(\frac{x}{\|x\|}),$$ where $\phi_i$ is an eigenfunction associated to the radial eigenvalue $\beta_{i,{{\text{\upshape rad}}}}(p)$ and $\alpha_k$ is an eigefunction of $-\Delta_{S^1}$ associated to the eigenvalue $k^2$ (see ). Then easy computation shows that if (resp. ) holds then $v_{i,k}$ with $i=1$ and $k=j$ (resp. $i=2$ and $k=1$) solves .\ [Step 3.]{} [We show that the space of solutions of at a value $p$ that satisfies (resp. ) is given by (resp. ).]{}\ The functions in (resp. ) clearly solve . This follows from [Step 2]{}, recalling the explicit expression of $\alpha_k$ (see ).\ To prove that the space of solutions to is spanned by the functions in and/or , recall that $\alpha_k$ is an orthogonal basis for $L^2(S^1)$, hence any nontrivial solution $v$ to may be written in $L^2(B)$ as $$\label{serie}v(r,\theta)=\sum_{k=0}^{+\infty}h_k(r)\alpha_k(\theta)$$ with $h_k$ defined as in . Then the same arguments used in [Step 1]{} imply that when only holds then $h_k=0$ for any $k\neq j$ and so reduces to $$v(r,\theta)=h_j(r)\alpha_j(\theta)$$ with $h_j$ eigenfunction associated to the radial eigenvalue $\beta_{1,{{\text{\upshape rad}}}}(p)$, namely $h_j=\phi_1$; similarly when only holds then $h_k=0$ for any $k\neq 1$ and so reduces to $$v(r,\theta)=h_1(r)\alpha_1(\theta)$$ where $h_1$ is now the eigenfunction associated to the radial eigenvalue $\beta_{2,{{\text{\upshape rad}}}}(p)$, namely $h_1=\phi_2$. The case $p$ large {#section p large} ================== In [@DeMarchisIanniPacellaMathAnn], exploiting the asymptotic analysis of $u_p$ for $p\rightarrow +\infty$, it has been already proved that \[risultatoMorse\_pGrande\] There exists $\hat{p}>1$ such that $$m(u_p)=12 \qquad \forall\ p\geq \hat{p}.$$ Moreover one can also improve some partial result in [@DeMarchisIanniPacellaMathAnn] about the asymptotic behavior of the first eigenvalue $(\beta^{n}_{1,{{\text{\upshape rad}}}}(p))_n$ of the auxiliary weighted problem (cfr. [@DeMarchisIanniPacellaMathAnn Theorem 6.1]) and deduce the following asymptotic result for the first eigenvalue $\beta_{1,{{\text{\upshape rad}}}}(p)$ in the ball (whose proof is sketched at the end of the section, see also [@AG2] for a detailed proof.) \[lemma:limiteBeta1\] $$\lim_{p\rightarrow +\infty}\beta_{1}(p)=-\frac{\ell^2+2}{2}\sim -26,9$$ where $\ell=\lim_{p{\rightarrow}\infty}\frac{s_p}{\varepsilon _p^-}\simeq 7.1979$ and $s_p\in (0,1)$ is the minimum point of $u_p$, (i.e. the point such that ${\lVertu_p^-\rVert}_{\infty}=u_p^-(s_p)=-u_p(s_p)$) and $\varepsilon _p^-$ is such that $\left(\varepsilon_p^-\right)^{-2}=p\|u_p(s_p)\|^{p-1}$. Using the general analysis previously done in Section \[section:generale\] (Lemma \[proposition:autovaloriRadiaiGenerale\] and Proposition \[p4.7\]), combining it with Proposition \[risultatoMorse\_pGrande\] above and with the asymptotic result in Lemma \[lemma:limiteBeta1\], we completely characterize the degeneracy of the solution $u_p$ when $p$ is large. Our result reads as follows: \[lemma:autovaloriRadialipgrande\] There exists $p^{\star}>1$ such that for any $p\geq p^{\star}$ $$\label{nu1controllatopgrande} -36 \ <\ \beta_{1,{{\text{\upshape rad}}}}(p)<-25$$ $$\label{nu2controllatopgrande} -1\leq \beta_{2,{{\text{\upshape rad}}}}(p)<0$$ Moreover for $p\ge p^{\star}$ the solution $u_p$ is degenerate if and only if $$\label{2degpgrande}\beta_{2,{{\text{\upshape rad}}}}(p)=-1.$$ The space of solutions to the linearized problem at a value $p\geq p^{\star}$ that satisfies is spanned by $$v(r,\theta)=\phi_2(r) \left(A \sin(\theta)+B\cos(\theta)\right) \qquad A,B\in {\mathbb{R}},$$ where $\phi_2$ is the eigenfunction associated to the second radial eigenvalue $\beta_{2,{{\text{\upshape rad}}}}(p)$ .\ Hence $Ker (L_p)$ for $p\geq p^{\star}$ has dimension $0$ when does not hold and dimension $2$ when it holds. The proof follows from Lemma \[proposition:autovaloriRadiaiGenerale\], Proposition \[p4.7\] and observing that by Proposition \[risultatoMorse\_pGrande\] $j(p)\equiv 6$ for $p\geq \hat{p}$ and that moreover by Lemma \[lemma:limiteBeta1\] there exists $p^{\star}(\geq\hat{p})$ such that the equality $$\beta_{1,{{\text{\upshape rad}}}}(p)=36$$ is never attained when $p\geq p^{\star}$. We conclude the section with: $\;$\ In [@DeMarchisIanniPacellaMathAnn Theorem 6.1] it has been proved that, if $\beta^n_1(p)$ is the first eigenvalue of the auxiliary weighted problem in the annulus $A_n$, then there exists a sequence $n_p$ such that $n_p{\rightarrow}\infty$ as $p{\rightarrow}\infty$ and $$\label{provvisoria} \tag{A.1} \lim_{p{\rightarrow}\infty}\beta_1^{n_p}(p)=-\frac{\ell^2+2}2.$$ Here we want to show that the proof for the annulus $A_{n_p}$ in [@DeMarchisIanniPacellaMathAnn] can be adapted to the ball $B$, so that one gets the same asymptotic result for the first eigenvalue $\beta_{1}(p)$ in the ball. The proof of the convergence in [@DeMarchisIanniPacellaMathAnn] deeply relies on the study of the behavior of the radial solution $u_p$ as $p{\rightarrow}\infty$, it is quite long and involved and requires several steps, which we now try to retrace.\ Let us first recall that $u_p$ admits two limit problems, one obtained when rescaling $u_p$ with respect to its maximum point, which is $0$ (the scaling parameter in this case is $\varepsilon^+_p$ defined by $\left(\varepsilon_p^+\right)^{-2}=pu_p(0)^{p-1}$) and the second one obtained rescaling $u_p$ with respect to its minimum point $s_p$ (with scaling parameter given by $\varepsilon_p^-$) (see [@GGP2 Theorem 1] for the rigorous statement of the result).\ As in [@DeMarchisIanniPacellaMathAnn] the idea to prove the result is now to consider the eigenvalue problem associated to $\beta_1(p)$, rescale properly the first eigenfunction $\psi_{1,p}$ using the scaling parameters $\varepsilon_p^\pm$ and pass to the limit into the rescaled equations. More precisely, similarly as in [@DeMarchisIanniPacellaMathAnn], we will obtain again that the rescaled eigenfunction $\widetilde \psi_{1,p}^{+}(x):=\psi_{1,p}\left(\varepsilon_p^{+} x\right)$ vanishes, while the other rescaled eigenfunction $\widetilde \psi_{1,p}^{-}(x):=\psi_{1,p}\left(\varepsilon_p^{-} x\right)$ converges (in some sense) to the first eigenfunction of some eigenvalue limit problem, whose first eigenvalue is exactly the value $ -\frac{\ell^2+2}2$ in .\ One of the main point, crucial to pass to the limit in the rescaled equation and get the limit eigenvalue problem, is to prove the analogous of Lemma 6.2, 6.3 and 6.4 in [@DeMarchisIanniPacellaMathAnn], where some estimates on the first eigenvalue $\beta_1^{n_p}(p)$ and on the associated rescaled eigenfunction are obtained. Similar estimates can be easily obtained exactly in the same way as in [@DeMarchisIanniPacellaMathAnn] directly for the first rescaled eigenfunctions $\widetilde \psi_{1,p}^{\pm}$ in the ball (without restricting to $A_{n_p}$) and imply in turn the convergence $$\label{provvisoria2}\tag{A.2} \widetilde \psi_{1,p}^{\pm}{\rightarrow}C^{\pm}\widetilde \psi_{1}^{\pm}\ \text{ as }p{\rightarrow}\infty$$ in some sense (in particular uniformly on compact sets of ${\mathbb{R}}^2$), where $\widetilde\psi_{1}^{\pm}$ are the first eigenfunctions of certain limit eigenvalue problems, associated respectively to eigenvalues $\widetilde \beta_1^{\pm}$. In particular, since we can prove that $\widetilde \beta_1^+=-1$ and we already know that $$\beta_1(p)\leq \beta_1^{n_p}(p)<-25 \ \text{ as $p$ large},$$ then necessarily $$\widetilde \psi_{1,p}^{+}{\rightarrow}0 \ \text{ as }p{\rightarrow}\infty.$$ The other main point, following the proof of , is to show that $\widetilde \psi_{1,p}^{-}$ does not vanishes. This step requires a deep analysis on the behavior of the function $$[0,1]\ni r\mapsto p\|u_p(r)\|r^2 \ \ \text{ as }p\rightarrow \infty$$ and luckily this behavior does not depend on the annulus $A_{n_p}$ and this produces estimates in all of the ball $B$. As a consequence, we can follow step by step the proof of Proposition 6.6 in [@DeMarchisIanniPacellaMathAnn], getting analogously that $$\liminf_{p{\rightarrow}\infty}\int_{\{\frac 1K<\|x\|<K\}}\frac{\left(\widetilde \psi_{1,p}^-\right)^2}{\|x\|^2}\ dx>0$$ for some $K>0$. The rest of the proof then follows similarly as in [@DeMarchisIanniPacellaMathAnn]. One can find in [@AG2] all the details. The case $p$ close to $1$ {#sse:pvicino1} ========================= Let us fix some notation. We denote by $(\lambda_i)_i$ the sequence of the Dirichlet eigenvalues of $-\Delta$ in $B$, counted with their multiplicity. Moreover let $(\varphi_i)_i$ be a basis of eigenfunctions in $L^2(B)$ associated to $\lambda_i$.\ We also denote by $(\lambda_{i,{{\text{\upshape rad}}}})_i$ and $(\varphi_{i, {{\text{\upshape rad}}}})_i$ the subsequences of the radial eigenvalues and eigenfunctions respectively (it is well known that $\lambda_{i,{{\text{\upshape rad}}}}$ are simple in the space of radial functions and that $\varphi_{i,{{\text{\upshape rad}}}}$ has $i-1$ zeros).\ The main result of this section is the following: \[risultatoMorse\_pvicino1\] There exists $\delta>0$ such that $$\label{morseindex-pvicino1} m(u_p)=6 \qquad \forall\ p\in (1,1+\delta)$$ and $u_p$ is nondegenerate for $p\in (1,1+\delta)$ (namely $\mu_7(p)>0$).\ Moreover $$\begin{aligned} \label{iRadiali} && \mu_{i}(p) \underset{p\rightarrow 1}{\longrightarrow} \lambda_i- \lambda_{2,{{\text{\upshape rad}}}} <0,\qquad i=1,\ldots, 5 \\ \nonumber &&\mu_{6}(p)=\mu_{2,{{\text{\upshape rad}}}}(p)\underset{p\rightarrow 1}{\longrightarrow} \lambda_6- \lambda_{2,{{\text{\upshape rad}}}}= 0^-\end{aligned}$$ and, up to a subsequence $$\label{convAutofDefinitivaTeo} v_{i,p}\underset{p\rightarrow 1}{\longrightarrow} C\frac{\varphi_i}{\\|\varphi_i\\|_\infty }\ \mbox{ in } C(\bar B), \qquad i=1,\ldots, 6$$ where $C=\pm1$ and $\mu_i(p)$, $\mu_{i,{{\text{\upshape rad}}}}(p)$ are the Dirichlet eigenvalues and radial eigenvalues respectively of the linearized operator $L_p$ at $u_p$ (see , and ) and $v_{i,p}$ are the eigenfunctions of $L_p$ associated to the eigenvalues $\mu_{i,p}$ and normalized in $L^{\infty}(B)$ ($\\|v_{i,p}\\|_\infty =1$). We observe that, combining with the general results about the Morse index of $u_p$ and the characterization of its degeneracy given in Section \[section p large\] for any $p>1$ (Proposition \[p4.7\] and Lemma \[proposition:autovaloriRadiaiGenerale\] respectively), we also have the following estimates for the $2$ negative radial eigenvalues of the auxiliary problem , when $p$ is close to $1$: \[lemma:autovaloriRadialipvicino1\] Let $\delta>0$ be as in Proposition \[risultatoMorse\_pvicino1\]. Then for any $p\in (1,1+\delta)$ $$\label{nu1controllatopvicino1} -9 < \beta_{1,{{\text{\upshape rad}}}}(p)<-4$$ $$\label{nu2controllatopvicino1} -1 < \beta_{2,{{\text{\upshape rad}}}}(p)<0.$$ From Lemma \[proposition:autovaloriRadiaiGenerale\], observing that implies $j(p)\equiv 3$ for $p\in (1,1+\delta)$, we have that $$\begin{aligned} && -9 \leq \beta_{1,{{\text{\upshape rad}}}}(p)<-4 \\ && -1 \leq \beta_{2,{{\text{\upshape rad}}}}(p)<0,\end{aligned}$$ for $p\in (1,1+\delta)$. The strict inequalities in the left hand sides follow from the nondegeneracy of $u_p$ in $(1,1+\delta)$ (see Proposition \[risultatoMorse\_pvicino1\]) and from the characterization of the degeneracy in Proposition \[p4.7\]. In order to obtain the previous result we need to analyze the behavior of the solution $u_p$, as $p$ is close to $1$. We will show that $u_p$ converges, as $p\rightarrow 1$, to the second radial Dirichlet eigenfunction of $-\Delta$ in the ball $B$ (Lemma \[lemma-pvicino1\] below).\ Hence let us recall some useful result for the Dirichlet eigenvalues and for the second radial eigenfunction of $-\Delta$ in $B$. \[lemma:morseSecAutofRadialeLap\] One has $$m\left(\varphi_{2,{{\text{\upshape rad}}}}\right)=5$$ and in particular $$\label{ordinamentol}\lambda_1=\lambda_{1,{{\text{\upshape rad}}}}<\lambda_2=\lambda_3<\lambda_4=\lambda_5<\lambda_6=\lambda_{2,{{\text{\upshape rad}}}}< \lambda_7\leq \ldots.$$ This proof is classical, we write it for completeness. The eigenfunctions of the Laplace operator $-\Delta$ with Dirichlet boundary conditions in $B$ are given, in radial coordinates, by $$\label{autof-laplace} \widetilde\varphi_{n,k}(r,\theta)=J_n(\nu_{nk}r)\times \left\{\begin{array}{ll} \cos (n\theta) &\\ \sin(n\theta) &\hbox{ for } n\neq 0 \end{array}\right.$$ for $n\in{\mathbb{N}},$ $k\in{\mathbb{N}}_0$, where $J_n$ are the Bessel functions of the first kind (see for instance [@Watson]) and $\nu_{nk}$ is the $k$-th positive root of $J_n$ (for any fixed $n$ there are infinitely many roots). The corresponding eigenvalues are given by $$\label{eigenLap} \widetilde\lambda_{nk}= \nu_{nk}^2,$$ hence they are simple for $n=0$ and have multiplicity $2$ when $n\geq 1$.\ From it follows that the second radial eigenfunction is $$\varphi_{2,{{\text{\upshape rad}}}}(r)=J_0(\nu_{02}r)$$ and so by the second radial eigenvalue is $$\label{eigenLap2}\lambda_{2,{{\text{\upshape rad}}}} = \nu_{02}^2 .$$ The Morse index of $\varphi_{2,{{\text{\upshape rad}}}}$ is the number of eigenvalues (counted with multiplicity) of the Laplace operator $-\Delta$ with Dirichlet boundary conditions in $B$ which are strictly less than $\lambda_{2,{{\text{\upshape rad}}}}$. By and this is equivalent to compute the number of the zeros $\nu_{nk}$ which are strictly less than $\nu_{02}$, recalling that when $n\geq 1$ each eigenvalue has multiplicity $2$.\ It is known (see [@Watson]) that $$\label{ordine} \nu_{01}<\nu_{11}<\nu_{21}<\nu_{02},$$ while $$\label{ordineancora} \nu_{12},\nu_{22}, \nu_{h1}>\nu_{02},\qquad \forall h\geq 3$$ hence the Morse index of $\varphi_{2,{{\text{\upshape rad}}}}$ is $5$.\ By , and (recalling the multiplicities) it follows that $$\begin{aligned} \lambda_1 &=& \widetilde\lambda_{01},\\ \lambda_2=\lambda_3 &=& \widetilde\lambda_{11}, \\ \lambda_4=\lambda_5 &=&\widetilde\lambda_{21},\\ \lambda_6 &=& \widetilde\lambda_{02} \ <\ \lambda_7,\end{aligned}$$ and that holds. Asymptotic behavior of $u_p$ as $p\rightarrow 1$ ------------------------------------------------ We now analyze the asymptotic behavior of $u_p$, as $p\rightarrow 1$. In particular we obtain an expansion of its $L^{\infty}$-norm up to the second order which will be useful for the proof of Theorem \[prop1.4\] (see Proposition \[leastRadiale\]). \[lemma-pvicino1\] Let $p_n$ be any sequence converging to $1$. Then $$\label{u-n} \bar{u}_n:=\frac {u_{p_n}}{{\lVertu_{p_n}\rVert}_{\infty}}\rightarrow \varphi_{2,{{\text{\upshape rad}}}} =J_0(\nu_{02}\|x\|)\quad \text{ in }C( \bar B)$$ (recall that, by the definition of $J_0$, we have that ${\lVert\varphi_{2,{{\text{\upshape rad}}}}\rVert}_{\infty}=\varphi_{2,{{\text{\upshape rad}}}}(0)=J_0(0)=1$) and $$\label{gammanp-1} {\lVertu_{p_n}\rVert}_{\infty}^{p_n-1}= \lambda_{2,{{\text{\upshape rad}}}}\left(1-\widetilde c (p_n-1)\right)+o(p_n-1) \text{ as }n{\rightarrow}\infty$$ where $$\label{c-tilde} \widetilde c:=\frac{\int_B\ \varphi_{2,{{\text{\upshape rad}}}}^2\log\|\varphi_{2,{{\text{\upshape rad}}}}\|dx}{\int_B\ \varphi_{2,{{\text{\upshape rad}}}}^2dx}$$ The function $\bar{u}_n$ defined in satisfies $$\label{eq:u-n} \left\{\begin{array}{lr} -\Delta \bar{u}_n= \gamma_n^{p_n-1}\|\bar{u}_n \|^{p_n-1}\bar{u}_n\qquad \mbox{ in }B\\ \bar{u}_n =0\qquad\qquad\qquad\qquad\qquad\mbox{ on }\partial B\\ \bar{u}_n(0)=1 \end{array}\right.$$ where $\gamma_n:={\lVertu_{p_n}\rVert}_{\infty}$. From it easily follows $$\Arrowvert \gamma_n^{p_n-1}\|\bar{u}_n \|^{p_n-1}\bar{u}_n\Arrowvert_{\infty}\leq M,$$ from which $$\label{bounGradL2} \\|\nabla\bar u_n\\|_{L^2(B)}\leq M.$$ Moreover we have the following estimate $$\label{stima-agg} \|\left( \|\bar{u}_n \|^{p_n-1}-1\right)\bar{u}_n\|\leq c(p_n-1)$$ in $\bar B$, with $c$ independent on $n$. Estimate obviously holds, for any fixed $n$, at the points at which $\bar{u}_n=0$. When $\bar{u}_n\neq 0$ instead it comes as in [@AGG (3.10)] from the identity $e^x-1=x\int_0^1e^{tx}\ dt$, from which $$\label{6.12-bis} \|\bar{u}_n \|^{p_n-1}-1=(p_n-1)\log \|\bar{u}_n \|\int_0^1 \left(\|\bar{u}_n \|^{p_n-1}\right)^t\ dt,$$ so that $$\big\|\|\bar{u}_n \|^{p_n-1}-1\big\|\leq (p_n-1)\big\|\log \|\bar{u}_n \|\big\|,$$ which implies by the boundedness of the function $x\mapsto x\log x$ in $(0,1)$. From we get $$\label{6.17} \left(\|\bar{u}_n \|^{p_n-1}-1\right)\bar{u}_n\rightarrow 0\quad \text{ as }n{\rightarrow}+\infty$$ uniformly in $\bar B$. Then, by and , $\bar{u}_n$ converges, up to a subsequence, in $C(\bar B)$ to a solution to $$\left\{\begin{array}{lr} -\Delta \bar{u}= \gamma\bar{u}\qquad\qquad \mbox{ in }B\\ \bar{u} =0\qquad\qquad\qquad\mbox{ on }\partial B\\ \bar{u}(0)=1 \end{array}\right.$$ where $\gamma:=\lim_{n\rightarrow +\infty} \gamma_n^{p_n-1}> 0$ by . Moreover $\bar{u}$ is radial and we will prove that it has two nodal regions. This implies that $\bar{u}=\varphi_{2,{{\text{\upshape rad}}}}$ showing and consequently $\gamma=\lambda_{2,{{\text{\upshape rad}}}}$. Since the convergence in holds for every subsequence, then it holds directly for the sequence $\bar{u}_n$.\ Next we show that $\bar u$ has $2$ nodal regions. Observe that the number of nodal regions of $\bar u$ cannot be grater then $2$ since $\bar u_n$ has $2$ nodal regions and it converges uniformly to $\bar u$. Let $r_n$ be the unique zero of $\bar{u}_n$ in $(0,1)$, up to a subsequence $r_n\rightarrow r_0$, then $\bar{u}$ has $2$ nodal regions if we show that $r_0\in (0,1)$. The $C^0$ convergence of $\bar{u}_n$ to $\bar{u}$ easily implies that $r_0>0$ since $\bar{u}(0)=1$. So by contradiction let us assume $r_n\rightarrow 1$ as $n\rightarrow +\infty$. By Rolle Theorem there exists $\xi_n\in (r_n,1)$ such that $\bar{u}_n'(\xi_n)=0$ for any $n$. By assumption $\xi_n\rightarrow 1$ as $n\rightarrow +\infty$. Moreover observe that the convergence in holds also in $C^1(B)$, by standard regularity theory, so it follows that $\bar{u}'(\xi_n)\rightarrow 0$ and this is not possible since the Hopf boundary Lemma implies $\bar{u}'(r)\neq 0$ in a neighborhood of $r=1$. We have shown so far that $ \gamma_n^{p_n-1}{\rightarrow}\lambda_{2,{{\text{\upshape rad}}}}$ as $n{\rightarrow}\infty$. To conclude we have to prove the expansion in . Let us multiply by $\varphi_{2,{{\text{\upshape rad}}}}$ and integrate over $B$. We get $$\gamma_n^{p_n-1}\int_B \|\bar{u}_n \|^{p_n-1}\bar{u}_n \varphi_{2,{{\text{\upshape rad}}}}=\int_B \nabla \bar{u}_n\nabla \varphi_{2,{{\text{\upshape rad}}}}=\lambda_{2,{{\text{\upshape rad}}}}\int_B \bar{u}_n \varphi_{2,{{\text{\upshape rad}}}}$$ where last equality follows by the definition of $ \varphi_{2,{{\text{\upshape rad}}}}$. This implies that $$\label{8.19} \lambda_{2,{{\text{\upshape rad}}}} \int_B\left(\|\bar{u}_n \|^{p_n-1}-1\right)\bar{u}_n \varphi_{2,{{\text{\upshape rad}}}}=\left(\lambda_{2,{{\text{\upshape rad}}}}-\gamma_n^{p_n-1}\right) \int_B \|\bar{u}_n \|^{p_n-1}\bar{u}_n \varphi_{2,{{\text{\upshape rad}}}}.$$ By using the identity , which holds a.e. in $B$, we also have $$\int_B\left(\|\bar{u}_n \|^{p_n-1}-1\right)\bar{u}_n \varphi_{2,{{\text{\upshape rad}}}}= (p_n-1)\int_B\bar{u}_n \varphi_{2,{{\text{\upshape rad}}}} \log\| \bar{u}_n\|\int_0^1 \|\bar{u}_n\|^{t(p_n-1)}dt \ dx$$ and so from we get $$\label{eqRapporto} \frac{\lambda_{2,{{\text{\upshape rad}}}}-\gamma_n^{p_n-1}}{\lambda_{2,{{\text{\upshape rad}}}}(p_n-1)}=\frac{\int_B \bar{u}_n \varphi_{2,{{\text{\upshape rad}}}} \log\| \bar{u}_n\|\int_0^1 \|\bar{u}_n\|^{t(p_n-1)}dt \ dx} {\int_B \|\bar{u}_n \|^{p_n-1}\bar{u}_n \varphi_{2,{{\text{\upshape rad}}}}\ dx}.$$ To conclude the proof we show that the right hand side of converges to the constant $\widetilde c$ in . First we observe that the uniform convergence of $\bar u_n$ to $\varphi_{2,{{\text{\upshape rad}}}}$ in $B$ implies $$\label{laPrima} \int_B \|\bar{u}_n \|^{p_n-1}\bar{u}_n \varphi_{2,{{\text{\upshape rad}}}}{\rightarrow}\int_B \varphi_{2,{{\text{\upshape rad}}}}^2\neq 0\ \text{ as }n{\rightarrow}\infty$$ (recall that $\varphi_{2,{{\text{\upshape rad}}}}(x)=J_0(\nu_{02}\|x\|)$). Moreover, since $\\|\bar u_n\\|_{\infty}\leq 1$, ($\bar u_n\neq 0$ q.o.) and the function $x\mapsto x\log x $ is bounded in $(0,1)$, then the term $\bar{u}_n \varphi_{2,{{\text{\upshape rad}}}} \log\| \bar{u}_n\| \int_0^1 \|\bar{u}_n\|^{t(p_n-1)}dt \in L^{\infty}(B)$ and $${\lVert\bar{u}_n \varphi_{2,{{\text{\upshape rad}}}} \log\| \bar{u}_n\|\int_0^1 \|\bar{u}_n\|^{t(p_n-1)}dt \rVert}_{L^{\infty}(B)}\leq C,$$ so by the convergence of $\bar u_n$ to $\varphi_{2,{{\text{\upshape rad}}}}$ and the dominated convergence theorem we also get $$\label{laSeconda}\int_B\bar{u}_n \varphi_{2,{{\text{\upshape rad}}}} \log\| \bar{u}_n\|\int_0^1 \|\bar{u}_n\|^{t(p_n-1)}dt \ dx{\rightarrow}\int_B \varphi_{2,{{\text{\upshape rad}}}}^2\log\|\varphi_{2,{{\text{\upshape rad}}}}\|dx \ \text{ as }n{\rightarrow}\infty.$$ Then, from , by and , it follows that $\frac{\lambda_{2,{{\text{\upshape rad}}}}-\gamma_n^{p_n-1}}{\lambda_{2,{{\text{\upshape rad}}}}(p_n-1)}$ is bounded and, up to a subsequence, $$\frac{\lambda_{2,{{\text{\upshape rad}}}}-\gamma_n^{p_n-1}}{\lambda_{2,{{\text{\upshape rad}}}} (p_n-1)}{\rightarrow}\widetilde c \ \text{ as }n{\rightarrow}\infty.$$ Since this convergence holds for every subsequence, then it holds for the sequence concluding the proof. Proof of Proposition \[risultatoMorse\_pvicino1\] ------------------------------------------------- Using Lemma \[lemma-pvicino1\] and Lemma \[lemma:morseSecAutofRadialeLap\] we can finally prove Proposition \[risultatoMorse\_pvicino1\]. The proof of consists in showing that for $p$ sufficiently close to $1$ $$\label{tesiMorse} m(u_p)=m\left(\varphi_{2,{{\text{\upshape rad}}}}\right)+1,$$ where $m\left(\varphi_{2,{{\text{\upshape rad}}}}\right)=5$ by Lemma \[lemma:morseSecAutofRadialeLap\]. We divide it into three steps. First observe that for $\bar u_p$ defined from $u_p$ as in $$\label{stessoLinearizzato}\|u_p\|^{p-1}={\lVert u_p\rVert}_{\infty}^{p-1}\|\bar {u}_p\|^{p-1}.$$ [Step 1.]{} [We show that $m(u_p)\geq m\left(\varphi_{2,{{\text{\upshape rad}}}}\right)+1$, for $p$ sufficiently close to $1$.]{}\ Let $Q_p\ : H^1_0(B)\rightarrow {\mathbb{R}}$ be the quadratic form in and let us consider the first $5$ Dirichlet eigenfunctions $\varphi_1,\dots,\varphi_5$ of $-\Delta$ in $B$ and the corresponding eigenvalues $\lambda_1,\dots,\lambda_5$. Then by we have that $$\begin{aligned} Q_p(\varphi_i) & = \int_B \left[\|\nabla \varphi_i\|^2 -p\|u_p\|^{p-1}\varphi_i^2 \right]dx\\ & \overset{\eqref{stessoLinearizzato}}{=}\int_{B}\left[\|\nabla \varphi_i\|^2-p{\lVert u_p\rVert}_{\infty}^{p-1}\|\bar{u}_p\|^{p-1}\varphi_i^2\right]\, dx\\ &=\lambda_i \int_B\varphi_i^2\, dx - p{\lVert u_p\rVert}_{\infty}^{p-1} \int_B \|\bar u_p\|^{p-1}\varphi_i^2\, dx \\ & \overset{(\star)}{=}\left(\lambda_i-\lambda_{2,{{\text{\upshape rad}}}}\right)\int_B\varphi_i^2\, dx +o_p(1)<0\end{aligned}$$ for $i=1,\dots,5$ and $p$ sufficiently close to $1$, since $\lambda_i<\lambda_{2,{{\text{\upshape rad}}}}$ by Lemma \[lemma:morseSecAutofRadialeLap\], where for the equality in $(\star)$ we have used and the Lebesgue dominated convergence theorem thanks to . Recalling that the eigenfunctions $\varphi_i$ are orthogonal in $L^2(B)$ and hence in $H^1_0(B)$ this means that the Morse index of $u_p$ is at least $5$ for $p$ sufficiently close to $1$. But from in Lemma \[proposition:autovaloriRadiaiGenerale\] we already know that $m(u_p)$ must be always even, then the Morse index of $u_p$ is at least $6$ for $p$ sufficiently close to $1$. [Step 2.]{} [Let $\mu_i(p)\leq 0$ be a non-positive Dirichlet eigenvalue of the operator $L_p$ for $p\in (1,1+\delta)$ and let $v_{i,p}$ be an associated eigenfunction with $\\|v_{i,p}\\|_{\infty}=1$. We prove that as $p\rightarrow 1$ $$\begin{aligned} && \mu_{i}(p) \ra \lambda_j- \lambda_{2,{{\text{\upshape rad}}}} \label{convAuto} \\ && v_{i,p}\rightarrow {C_j}\varphi_j \ \mbox{ in } C(\bar B)\ \mbox{ up to a subsequence,}\label{convAutof}\end{aligned}$$ for a certain $j=j(i)\in\{1,2,3,4,5,6\}$, where $C_j:=\pm {\lVert\varphi_j\rVert}^{-1}_\infty$. Moreover we also show that if $l\in\mathbb N$, $l\neq i$ and $\mu_l(p)\leq 0$ for $p\in (1,1+\delta)$, then $$\label{distinti}j(l)\neq j(i)$$ (we stress that under condition it is nevertheless possible to have $\lambda_{j(l)}=\lambda_{j(i)}$). ]{}\ Observe that the non-positive eigenvalue $\mu_{i}(p)$ is bounded for $p$ close to $1$, indeed by the standard variational characterization of $\mu_1(p)$ $$\begin{aligned} \mu_{i}(p)&>\mu_{1}(p)=\mu_{1,{{\text{\upshape rad}}}}(p) \overset{\eqref{stessoLinearizzato}}{=}\inf_{\substack{v\in H^1_{0,{{\text{\upshape rad}}}}(B)\\v\neq 0}} \frac{\int_0^1\left(r\left(v'\right)^2-p{\lVert u_p\rVert}_{\infty}^{p-1}\|\bar{u}_p\|^{p-1}r v^2\right) dr}{\int_0^1 rv^2}\\ &\geq -p{\lVert u_p\rVert}_{\infty}^{p-1}\overset{\eqref{gammanp-1}}{\geq}-(\lambda_{2,{{\text{\upshape rad}}}}+{\varepsilon}) \end{aligned}$$ for $p$ close to $1$. Let $p_n$ be a sequence converging to $1$, then the eigenfunction $v_{i,n}:=v_{i,p_n}$ satisfies $$\label{eigen} \left\{ \begin{array}{ll} L_pv_{i,n}\overset{\eqref{stessoLinearizzato}}{=}-\Delta v_{i,n} - p_n {\lVert u_{p_n}\rVert}_{\infty}^{p_n-1}\|\bar u_{p_n}\|^{p_n-1} v_{i,n} =\mu_i(p_n) v_{i,n} & \text{ in } B \\ {\lVert v_{i,n}\rVert}_{\infty}=1\\ v_{i,n}= 0 & \text{ on } \partial B. \end{array} \right.$$ Moreover $$\big\| p_n {\lVert u_{p_n}\rVert}_{\infty}^{p_n-1}\|\bar u_{p_n}\|^{p_n-1} v_{i,n} +\mu_i(p_n) v_{i,n}\big\|\leq C$$ and then, up to a subsequence, $v_{i,n}{\rightarrow}\tilde{\varphi_i}$ in $C(\bar B)$ where ${\lVert\tilde \varphi_i\rVert}_{\infty}=1$ by the uniform convergence and, using and , it follows that $\tilde \varphi_i$ solves $$\label{new} \begin{cases} \begin{array}{ll} -\Delta \tilde\varphi_{i} =\left(\lambda_{2,{{\text{\upshape rad}}}} +\tilde \mu_{i}\right) \tilde \varphi_{i}\qquad & \text{ in } B \\ {\lVert \tilde \varphi_{i}\rVert}_{2}=1\\ \tilde \varphi_{i}= 0 & \text{ on } \partial B, \end{array} \end{cases}$$ where $\tilde \mu_i=\lim_{n\rightarrow +\infty} \mu_{i}(p_n)\leq 0$. This means that $\tilde \varphi_i$ is an eigenfunction of the Laplace operator associated to the eigenvalue $ \lambda_{2,{{\text{\upshape rad}}}} +\tilde \mu_{i}$, namely there exists $j=1,2,\ldots$ such that $$\tilde \mu_{i}=\lambda_j-\lambda_{2,{{\text{\upshape rad}}}}$$ and $$\tilde \varphi_i={C_j}\varphi_j$$ where $C_j=\pm{\lVert\varphi_j\rVert}_{\infty}^{-1}$. Since $ \tilde \mu_{i}\leq 0$, by Lemma \[lemma:morseSecAutofRadialeLap\] we have necessarily that $j\in\{1,2,3,4,5,6\}$. Moreover, since the convergence in holds for any subsequence, then it also holds for the sequence. Last we prove . Let $l\neq i$ be such that $\mu_l(p)\leq 0$. We can take $v_{l,p}$ orthogonal in $L^2(B)$ to $v_{i,p}$. The uniform convergence in $\bar B$ implies then that $$0=\int_B v_{i,p}v_{l,p}={C_{j(i)}C_{j(l)}}\int _B \varphi_{j(i)} \varphi_{j(l)},$$ hence $j(i)\neq j(l)$. [Step 3.]{} [Conclusion]{}\ From [Step 2]{} we deduce that the operator $L_p$, for $p$ close to $1$, may have at most $6$ non-positive eigenvalues $\mu_i(p)\leq 0$, namely that $\mu_7(p)>0$.\ Indeed if we assume by contradiction that $\mu_7(p)\leq 0$ for $p$ close to $1$, then holds for all $i=1,2,\ldots, 7$ and so necessarily $j(7)=j(\hat i)$ for some $\hat i\in\{1,\ldots 6\}$, a contradiction with .\ From [Step 1]{}, we also know that the operator $L_p$ for $p$ close to $1$ has at least $6$ negative eigenvalues $\mu_i(p)< 0$.\ Combining both the information we get: $$\label{dimostratoPrimaParte} \mu_1(p)<\mu_2(p)\leq \mu_3(p)<\mu_4(p)\leq\mu_5(p)< \mu_6(p)<0<\mu_7(p)\leq\ldots$$ (the strict inequalities are a consequence of and of the convergence in ), which proves both and the nondegeneracy of $u_p$ for $p$ close to $1$.\ It remains to prove . It is well known that $\mu_1(p)=\mu_{1,{{\text{\upshape rad}}}}(p)$. Moreover $m_{{{\text{\upshape rad}}}}(u_p)=2$ by Lemma \[LemmaMorseIndexRadiale\], hence there exists a unique $l\in\{2,3,4,5,6\}$ such that $\mu_l(p)=\mu_{2,rad}(p)$. We denote by $v_{l,p}$ a radial eigenfunction associated to $\mu_l(p)$. Next we show that $l=6$.\ Observe that as a consequence of and of the monotonicity property of the limit, we can take $j=i$ in the convergences already proved in [Step 2]{}, namely and become respectively: $$\begin{aligned} && \mu_{i}(p) \ra \lambda_i- \lambda_{2,{{\text{\upshape rad}}}} \label{convAutoDefinitiva} \\ && v_{i,p}\rightarrow {C_i}\varphi_i \label{convAutofDefinitiva}\end{aligned}$$ as $p\rightarrow 1$, for any $i=1,\ldots, 6$.\ Obviously $\varphi_1=\varphi_{1,{{\text{\upshape rad}}}}$ and moreover, since $\lambda_6=\lambda_{2,{{\text{\upshape rad}}}}$ by Lemma \[lemma:morseSecAutofRadialeLap\], we can take $\varphi_6=\varphi_{2,{{\text{\upshape rad}}}}$, while $\varphi_i$ is surely not radial for $i=2,3,4,5$. Observe now that $\varphi_l$ is radial, being obtained in the limit of the radial eigenfunction $v_{l,p}$ in , this proves that $l=6$. Last in the case $i=6$ also gives the limit $ \mu_{6}(p)=\mu_{2,{{\text{\upshape rad}}}}(p) \ra 0^-$ as $p\rightarrow 1$. Regularity of the eigenvalues and the set $\mathcal P^j$ ======================================================== By Proposition \[p4.7\] the sets of the exponents $p$ at which $u_p$ is degenerate are $$\begin{aligned} \label{setdegeneratep} &&\{p\in(1,+\infty) : \beta_{1,{{\text{\upshape rad}}}}(p)=-j^2\},\ \mbox{ for }j\in{\mathbb{N}}_0\nonumber \\ &&\{p\in(1,+\infty) : \beta_{2,{{\text{\upshape rad}}}}(p)=-1\}.\end{aligned}$$ In particular we will be interested in the subset $$\label{def:S_j} \mathcal P^j:=\left\{p\in(1,+\infty) :\ p\mapsto \beta_{1,{{\text{\upshape rad}}}}(p)+j^2\ \mbox{ changes sign}\right\}, \quad j\in {\mathbb{N}}_0.$$ Clearly $\mathcal P^j\subseteq \left\{p\in(1,+\infty) :\beta_{1,{{\text{\upshape rad}}}}(p)=-j^2\right\}.$ \[l5.1\] The maps $p\mapsto\beta_{i,{{\text{\upshape rad}}}}(p)$ are analytic in $p$ and the sets of degenerate points in , when not empty, consist of only isolated points.\ Moreover $\mathcal P^j\not=\emptyset$, for $j=3,4,5$ and there exists an odd number $s_j(\geq 1)$ of isolated values $p_1^j,\ldots, p^j_{s_j}\in (1+\delta,p^{\star})$ (where $\delta$ and $p^{\star}$ are as in Proposition \[risultatoMorse\_pvicino1\] and Proposition \[lemma:autovaloriRadialipgrande\] respectively) such that $$\mathcal P^j=\{p_1^j,\ldots, p^j_{s_j}\} \qquad j=3,4,5.$$ In [@D] it is proved that for any smooth bounded domain $\Omega\subset {\mathbb{R}}^2$ for any $p>1$ except possibly for isolated $p$ the equation $-\Delta u=u^p$ in $\Omega$, $u=0$ on $\partial \Omega$ has a non-degenerate positive solution. The proof relies on the fact that the map $(u,p)\longrightarrow \left(-\Delta\right)^{-1}(u^p)$ is real analytic when considered in a suitable cone of positive weighted functions.\ This proof cannot be directly applied for sign-changing solutions, and so we need to adapt the proof of the analyticity for sign-changing radial fast decay solutions in the exterior of the ball used in [@DW], which holds in ${\mathbb{R}}^N$, with $N\geq 3$.\ Following [@DW] we let $\tilde w_p(s)=r^{\frac 2{p-1}}u_p(r)$, for $r=e^s$. This function satisfies $$\tilde{w}_p''-\frac 4{p-1}\tilde{w}_p'+\left(\frac 2{p-1}\right)^2\tilde w_p+\|\tilde w_p\|^{p-1}\tilde w_p=0$$ for $s\in(-\infty,0)$ with the conditions $$\label{**} \tilde w_p(0)=0\quad ,\quad \lim_{s{\rightarrow}-\infty}\tilde w_p(s)=0.$$ We consider, for $z>0$, the rescaled function $w(t)=\tilde w_p(z^{-1}t)$ that satisfies $$\label{riscalato} {w}''-\frac 4{p-1}z{w}'+\left(\frac 2{p-1}\right)^2z^2 w+z^2\| w\|^{p-1} w=0$$ in $(-\infty,0)$ with the boundary conditions in . We let $s_1$ be the unique zero of $w(t)$ in $(-\infty, 0)$ and we consider problem in one of the intervals $(-\infty,s_1)$ or $(s_1,0)$ with Dirichlet boundary conditions (also at infinity). Of course we have that $r_1=e^{z^{-1}s_1}$ is the unique zero of $u_p$. Problem is equivalent to solve $$\begin{cases} -\Delta u=u^p & \text{ in }\Omega_i\\ u>0 & \text{ in }\Omega_i\\ u=0 & \text{ on }\partial\Omega_i \end{cases}$$ where $\Omega_1=B(0,e^{z^{-1}s_1})$ or $\Omega_2=B\setminus B(0,e^{z^{-1}s_1})$ and $u$ is radial. The Dancer result for positive solutions in [@D] implies then that the positive solutions $w^1_{z,p}$ and $w^2_{z,p}$ to , in $(-\infty,s_1)$ and $(s_1,0)$ respectively, depend analytically on $p$ and $z$.\ Lastly, following the proof of Lemma 3.2 part c) in [@DW], one can show the existence of $z_p$ close to $1$ and analytic in $p$ such that the function $$\tilde{w}_p(s) =\left\{ \begin{split} w^1_{z_p,p}(z_ps) \quad \text{ for }s\in (-\infty,z_p^{-1}s_1]\\ -w^2_{z_p,p}(z_ps) \quad \text{ for }s\in (z_p^{-1}s_1,0) \end{split} \right.$$ is $C^1$ in $s=z_p^{-1}s_1$. This proves that $p\mapsto u_p$ is analytic.\ The fact that $u_p$ is analytic with respect to $p$ implies that the eigenvalues $\beta_{1,{{\text{\upshape rad}}}}(p)$, $\beta_{2,{{\text{\upshape rad}}}}(p)$ are analytic [@Kato]. Moreover by and it follows that $p\mapsto\beta_{1,{{\text{\upshape rad}}}}(p)$ is not constant in $(1,+\infty)$ and so the solutions $p\in (1,+\infty)$ to $\beta_{1,{{\text{\upshape rad}}}}(p)=-j^2$ are isolated and can accumulate only at $+\infty$. By and instead, either $p\mapsto\beta_{2,{{\text{\upshape rad}}}}(p)$ is constant and there are no solutions $p\in (1,+\infty)$ to $\beta_{2,{{\text{\upshape rad}}}}(p)=-1$ or it is not constant in $(1,+\infty)$ and in this case the solutions $p\in (1,+\infty)$ to $\beta_{2,{{\text{\upshape rad}}}}(p)=-1$ are isolated and can accumulate only at $+\infty$. Finally and imply also that $\beta_{1,{{\text{\upshape rad}}}}(p)+j^2$ changes sign for some $p\in(1+\delta,p^{\star})$ (precisely at an odd number of values of $p$), when $j=3,4,5$. Morse index and degeneracy in symmetric functions spaces {#se:symmspaces} ======================================================== To prove the bifurcation result in Theorem \[teo1\] and also to prove Theorem \[prop1.4\] we need to introduce some spaces of symmetric functions. To this end we let $O(2)$ be the orthogonal group in ${\mathbb{R}}^2$, $O_k\subset O(2)$, for $k\in\mathbb N_0$, be the subgroup of rotations of angle $\frac {2\pi}k$ and $\tau\in O(2)$ be the reflection with respect to the $x$-axis, i.e. $\tau(x,y)=(x,-y)$ for any $(x,y)\in {\mathbb{R}}^2$. For any $k\in {\mathbb{N}}_0$, we denote by $$\label{g-k} \mathcal{G}_k\subset O(2) \ \text{ the subgroup generated by the elements of $O_k$ and by $\tau$ }$$ and by $$\label{Hk} H^1_{0,k}(B)\ : =\{v\in H^1_0(B)\ \text{ such that } v(g(x))=v(x), \ \ \forall g\in \mathcal{G}_k, \ \forall x\in B \}.$$ The functions in the spaces $H^1_{0,k}(B)$ clearly possess the following invariances (in polar coordinates $(x,y)=(r\cos \theta,r\sin \theta)$): $$\begin{aligned} \label{fi00} &v(r,\theta)=v(r,2\pi-\theta)\\\label{fi0} &v(r,\theta)=v(r, \theta+\frac{2\pi}k) \end{aligned}$$ and so also $$\label{fi1} v(r,\frac \pi k+\theta)=v(r,\frac \pi k-\theta) $$ for every $r\in (0,1]$ and for every $\theta\in [0,2\pi]$. Note that in general $\theta+\frac{2\pi}k\notin[0,2\pi]$, if this occurs we mean that $v(r,\theta)=v(r, \theta+\frac{2\pi}k-2\pi)$ and similarly we do when $\frac{\pi}k\pm\theta\notin[0,2\pi]$.\ Observe that when $k=1$ then $O_1$ is the trivial subgroup of $O(2)$ given by the identity map and the functions in $H^1_{0,1}(B)$ are only invariant by the reflection $\tau$. Clearly the radial solution $u_p\in H^1_{0,k}(B)$, for every $k\in{\mathbb{N}}_0$.\ As a consequence, letting as before $\left(\mu_i(p)\right)_{i\in{\mathbb{N}}_0}$ be the sequence of the eigenvalues of the linearized operator $L_p$ at $u_p$ (see Section \[section:linearizedOperator\]), we can consider its subsequence $\left( \mu_{i,k}(p)\right)_{i\in {\mathbb{N}}_0}$ of the $\mathcal G_k$-symmetric eigenvalues (i.e. eigenvalues associated to an eigenfunction that belongs to $H^1_{0,k}(B)$) for any $k\in{\mathbb{N}}_0$, which can be characterized as $$\begin{aligned} \nonumber \mu_{i,k}(p) &=& \min_{\substack{ W\subset H^1_{0,k}(B)\\ dim W=i}} \max_{\substack{v\in W\\v\neq 0}}\ \ \ R_p[v],\end{aligned}$$ where $R_p$ is the usual Rayleigh quotient as in . By the principle of symmetric criticality the functions $v_i$ that attains $\mu_{i,k}(p)$ are indeed solutions to the eigenvalue problem associated to the linearized operator, i.e. they satisfy $$\begin{cases} \begin{array}{ll} -\Delta v_i - p\|u_p(x)\|^{p-1} v_i =\mu_{i,k}(p) v_i\qquad & \text{ in } B, \\ v_i= 0 & \text{ on } \partial B \end{array} \end{cases}$$ and are invariant by the action of $\mathcal{G}_k$. It is known that $\mu_{1,k}(p)=\mu_{1,{{\text{\upshape rad}}}}(p)=\mu_1(p)$, for any $k\in{\mathbb{N}}_0$, since $v_1$ is a radial function. We then define the [$k$-Morse index of $u_p$]{}, that we denote by $m_k(u_p)$, as the number of the negative $\mathcal G_k$-symmetric eigenvalues $\mu_{i,k}(p)$ of $L_p$ counted with multiplicity. To compute the $k$-Morse index of $u_p$ it is useful the following result, analogous to the one in Lemma \[lemma:Morse=numeroAutovaloriesatoPalla\]: \[lemma:Morse=numeroAutovaloriesatoPallaSIMMETRIA\] The $k$-Morse index of $u_p$ coincides with the number of the negative $\mathcal G_k$-symmetric eigenvalues of the weighted problem counted according to their multiplicity. The proof of the previous result is an easy adaptation of the arguments in [@GGN2 Lemma 2.6] and relies on the variational characterization of the negative $\mathcal G_k$-symmetric eigenvalues of the weighted problem (i.e. the eigenvalues whose eigenfuntions belong to $H^1_{0,k}(B)$). Indeed observe that they are a subsequence of the eigenvalues of the weighted problem and that, as we have already seen in Section \[se:auxiliary\], they can be variationally characterized exactly when they are [negative]{}. More precisely, by the principle of symmetric criticality, we can now restrict to the subspace $\mathcal H_{k}$ of the $\mathcal G_k$-symmetric functions of $\mathcal H$ ($\mathcal H_{k}\subset H^1_{0,k}(B)$) and define $$\begin{aligned} \label{CourantCharEigenvpesatipallaSIMMETRICO1} \beta_{1,k}(p) &:=& \inf_{\substack{v\in \mathcal H_{k},\,\,v\neq 0}}\ \ \ \widetilde{R_p}[v] \ (= \beta_1(p)=\beta_{1,{{\text{\upshape rad}}}}(p))\end{aligned}$$ and, if $\beta_{j,k}(p)<0$ for $j=1,\ldots, i-1$ $$\begin{aligned} \label{CourantCharEigenvpesatipallaSIMMETRICO} \beta_{i,k}(p) &:=& \inf_{\substack{v\in \mathcal H_{k},\,\,v\neq 0\\ v\perp_{\mathcal H}span\{\phi_{1},\ldots,\phi_{i-1}\}}}\ \ \ \widetilde{R_p}[v],\qquad i\in{\mathbb{N}}, \ i\geq 2,\end{aligned}$$ where $\phi_j\in\mathcal H_{k}$ is the function where $\beta_{j,k}(p)$ is achieved for $j=1,\ldots, i-1$ and solve $$\label{problemaPesatoPallaDeboleiSIMMETRICO} \int_B\nabla\phi_j\nabla v-p\|u_p\|^{p-1}\phi_jv\, dx= \beta_{j,k}(p)\int_B \frac{\phi_j v}{\|x\|^2}\, dx, \qquad \forall v\in \mathcal H.$$ So similarly as in Lemma \[valoriVariaz=AutovaloriesatoPalla\] one can prove the following variational characterization, which then gives the characterization of the $k$-Morse index in Lemma \[lemma:Morse=numeroAutovaloriesatoPallaSIMMETRIA\] above: \[valoriVariaz=AutovaloriesatoPallaSIMMETRICO\] The negative $\mathcal G_k$-symmetric eigenvalues of problem coincide with the negative numbers $\beta_{i,k}(p)$’s in -. Moreover the corresponding eigenfunctions, which solve , are in $\mathcal H_k$ and can be chosen to be orthogonal in the sense of . \[remark:simmetriaAutofunzioni\] Recall that, according to the spectral decomposition result in Lemma \[lemma:decomposizionePalla\] and using Lemma \[2AutovRadNegativiPesatoPalla\], we can decompose the negative eigenvalues of the weighted problem as $$\label{autov-minori-zero} \beta_{n,{{\text{\upshape rad}}}}(p)+j^2<0$$ for some $n=1,2$ and some $j\in {\mathbb{N}}$, where $\beta_{n,{{\text{\upshape rad}}}}(p)$ are the negative radial weighted eigenvalues as defined in Section \[se:auxiliary\].\ Moreover the eigenfunctions associated to each $(n,j)\in\{1,2\}\times{\mathbb{N}}$ in the decomposition are explicitly known by Lemma \[lemma:decomposizionePalla\], indeed they are: $$\phi_n(r)\cos (j\theta)\ \mbox{ and }\ \phi_n(r)\sin (j\theta) $$ where $\phi_n(r)$ is a radial eigenfunction associated to the simple radial eigenvalue $\beta_{n,{{\text{\upshape rad}}}}(p)$.\ Recall also that, by , the eigenspace related to each negative eigenvalue of problem is generated by these eigenfunctions, with $(n,j)$ varying among all the possible associated decompositions.\ Hence the $\mathcal G_k$-invariance of the eigenfunctions is known, precisely one has that: - for $j=0$, the eigenvalues $\beta_{1,{{\text{\upshape rad}}}}(p)<\beta_{2,{{\text{\upshape rad}}}}(p)<0$ are simple in the space of the radial functions and each one produces $1$ radial eigenfunction $\phi_{n}$ ($n=1,2$ respectively) of problem , which belongs to $H^1_{0,k}(B)$ for every $k\geq1$; - for every $j\geq 1$, the eigenfunction $\phi_n(r)\sin (j \theta)$ doesn’t belong to any space $H^1_{0,k}(B)$, $k\geq 1$ (since the reflection $\tau\in\mathcal G_k$); - for every $j\geq 1$, the eigenfunction $\phi_n(r)\cos (j \theta)$ is in $H^1_{0,j}(B)$; - for every $j\geq 2$, the eigenfunction $\phi_n(r)\cos (j \theta)$ belongs also to the spaces $H^1_{0,k}(B)$ such that $k\in{\mathbb{N}}_0$ is a factor of $j$ (we write $k\mid j$) (in particular it always belongs to $H^1_{0,1}(B)$), while it doesn’t belong to the spaces $H^1_{0,k}(B)$ when $k\in{\mathbb{N}}_0$ is not a factor of $j$. In the next section we will use the following result: \[lemma:legameAutovSimmNegativoLinEPesato\] Let $p\in (1,+\infty)$. The linearized operator $L_p$ has a negative eigenvalue with eigenfunction in $H^1_{0,k}(B)\setminus H^1_{0,{{\text{\upshape rad}}}}(B)$ if and only if $$\label{negativo-k} \beta_{1,{{\text{\upshape rad}}}}(p)+k^2<0 \quad$$ Lemma \[lemma:Morse=numeroAutovaloriesatoPallaSIMMETRIA\] implies that $L_p$ has a negative eigenvalue in $H^1_{0,k}(B)\setminus H^1_{0,{{\text{\upshape rad}}}}(B)$ if and only if the weighted problem has a negative eigenvalue in the space $\mathcal H_k\setminus \mathcal H_{{{\text{\upshape rad}}}}$. By the spectral decomposition given in Lemma \[lemma:decomposizionePalla\] then, when holds problem has the negative eigenvalue $\beta(p)=\beta_{1,{{\text{\upshape rad}}}}(p)+k^2$ with corresponding eigenfunctions $\phi_1(r)\sin (k\theta)$ and $\phi_1(r)\cos (k\theta)$, the second of which belonging to $\mathcal H_{k}\setminus \mathcal H_{{{\text{\upshape rad}}}}$. When, instead $\beta_{1,{{\text{\upshape rad}}}}(p)+k^2\geq 0$ the negative eigenvalues of problem are: $\beta_{i,{{\text{\upshape rad}}}}(p)$, for $i=1,2$ with corresponding eigenfunctions $\phi_i(r)\in \mathcal H_{{{\text{\upshape rad}}}}$ so that they do not belong to $\mathcal H_k\setminus\mathcal H_{{{\text{\upshape rad}}}}$ and $\beta_{1,{{\text{\upshape rad}}}}(p)+j^2$ for some $j\in \{1,\dots,k-1\}$ with corresponding eigenfunctions $\phi_1(r)\sin (j\theta)$ and $\phi_1(r)\cos (j\theta)$ neither of which belong to $\mathcal H_k$ since $j<k$, by Remark \[remark:simmetriaAutofunzioni\]. This means that when is not satisfied then the linearized operator does not admit any negative eigenvalue in $H^1_{0,k}(B)\setminus H^1_{0,{{\text{\upshape rad}}}}(B)$ concluding the proof. By exploiting the information about the location of the weighted radial eigenvalues $\beta_{n,{{\text{\upshape rad}}}}(p)$, $n=1,2$ obtained in the previous sections we can also derive information about the $k$-Morse index of the radial solution $u_p$ which will be useful to prove the non-radial part in Theorem \[prop1.4\] (see Section \[section:proofTeoLeastEnergy\]). Indeed using the results in Section \[section p large\] and Section \[sse:pvicino1\], we can explicitly compute the $k$-Morse index of $u_p$, for $p$ large enough and for $p$ close to $1$ respectively: \[Morse-simmetrico1\] Let $p^{\star}>1$ be as in Proposition \[lemma:autovaloriRadialipgrande\]. Then for any $p\geq p^{\star}$ $$\label{Morse-symm} m_k(u_p)=\begin{cases} 7 & \text{ for }k=1\\ 4 & \text{ for }k=2\\ 3& \text{ for }k=3,4,5\\ 2& \text{ for }k\geq 6 \end{cases}$$ By Lemma \[lemma:Morse=numeroAutovaloriesatoPallaSIMMETRIA\] in order to compute $m_k(u_p)$ we have to count the linearly independent eigenfunctions to the weighted problem which are associated to a negative eigenvalue and belong to the symmetric space $H^1_{0,k}(B)$.\ From Proposition \[lemma:autovaloriRadialipgrande\] we know that for $p\geq p^{\star}$ it holds $$-36 < \beta_{1,{{\text{\upshape rad}}}}(p)<-25 ,\ \ \ -1\leq \beta_{2,{{\text{\upshape rad}}}}(p)<0.$$ Then all the negative eigenvalues are given by with $$j=\left\{ \begin{array}{lr} 0 & \mbox{ for }n=2\\ 0,1,2,3,4,5 & \text{ for }n=1 \end{array} \right.$$ The conclusion follows by $a)$, $b)$, $c)$ and $d)$ in Remark \[remark:simmetriaAutofunzioni\]. Analogously for $p$ close to $1$ one has: \[Morse-simmetrico2\] Let $\delta>0$ be as in Proposition \[risultatoMorse\_pvicino1\]. Then for any $p\in (1,1+\delta)$ $$\label{Morse-symm2} m_k(u_p)=\begin{cases} 4& \text{ for }k=1\\ 3 & \text{ for }k=2\\ 2 & \text{ for }k\geq 3 \end{cases}$$ We reason as in the proof of the previous lemma. From Corollary \[lemma:autovaloriRadialipvicino1\] we know that for $p\in (1,1+\delta) $ it holds $$-9< \beta_{1,{{\text{\upshape rad}}}}(p)<-4 , \ \ \ -1< \beta_{2,{{\text{\upshape rad}}}}(p)<0.$$ Then all the negative eigenvalues are given by with $$j=\left\{ \begin{array}{lr} 0 & \mbox{ for }n=2\\ 0,1,2 & \text{ for }n=1 \end{array} \right.$$ The conclusion follows again by Remark \[remark:simmetriaAutofunzioni\]. Finally we can characterize the degeneracy of $u_p$ in the symmetric spaces. We know from Proposition \[p4.7\] that $u_p$ is degenerate if and only if $$\begin{aligned} &\beta_{1,{{\text{\upshape rad}}}}(p)+j^2=0 \quad \text{ for some }j=j(p)>1 \ \text{ or }\\ &\beta_{2,{{\text{\upshape rad}}}}(p)+1=0\end{aligned}$$ and these equalities can hold at the same time. As we will see in next result, the restriction to the symmetric spaces on one side rules out the degeneracy due to the second case, on the other side reduces the kernel of $L_p$ to be $1$-dimensional. \[lemma:degenerazioneSimmetria\] Let $\delta>0$ and $p^{\star}>1$ be as in Proposition \[risultatoMorse\_pvicino1\] and Proposition \[lemma:autovaloriRadialipgrande\] respectively. Let $k\in{\mathbb{N}}_0$. - if $p\in (1,1+\delta)$ then $u_p$ is non-degenerate in $H^1_{0,k}(B)$ for any $k\geq 1$; - if $p\geq p^{\star}$ then $u_p$ is non-degenerate in $H^1_{0,k}(B)$ for any $k\geq 2$; - if $p\in (1+\delta,p^{\star})$ then $u_p$ is degenerate in $H^1_{0,k}(B)$ for $k\geq 2$ if and only if there exists $j\geq 2$ such that $$\beta_{1,{{\text{\upshape rad}}}}(p)=-j^2\quad \mbox{ and }\quad k\mid j.$$ In this case the kernel of $L_p$ in $H^1_{0,k}(B)$ is one dimensional and it is spanned by the function $\phi_1(r)\cos(j\theta)$. $i)$ is obvious, since $u_p$ is non-degenerate in $H^1_0(B)$ when $p\in (1,1+\delta)$ (Proposition \[risultatoMorse\_pvicino1\]).\ $ii)$ follows from the characterization of the degeneracy of $u_p$ in $H^1_0(B)$ for $p$ large. Indeed when $u_p$ is degenerate in $H^1_0(B)$ by Proposition \[lemma:autovaloriRadialipgrande\] the kernel of $L_p$ is spanned by the two functions $\phi_2(r)\sin (\theta)$ and $\phi_2(r)\cos(\theta)$ and neither of the two belong to $H^1_{0,k}(B)$, when $k\geq 2$.\ $iii)$ follows from the characterization of the degeneracy of $u_p$ in $H^1_0(B)$ given in Proposition \[p4.7\]. Indeed, observe that the solution to the linearized equation $v$ in do not belong to $H^1_{0,k}(B)$ when $k\geq 2$, hence $Ker(L_p)\not = \{0\}$ in $H^1_{0,k}(B)$ if and only if $p$ satisfies the equation . To conclude let us recall that in this case $Ker( L_p)$ is spanned by the functions $\phi_1(r) \sin(j\theta)$ and $\phi_1(r)\cos(j\theta)$ (see ) and that $\phi_1(r) \sin(j\theta)\not\in H^1_{0,k}(B)$ for $k\geq 2$, while $\phi_1(r)\cos(j\theta)\in H^1_{0,k}(B)$ for any $k\mid j$. \[rmk:OddChange\] From Proposition \[lemma:degenerazioneSimmetria\] - $iii)$, Lemma \[lemma:Morse=numeroAutovaloriesatoPallaSIMMETRIA\] and the usual spectral decomposition of the negative eigenvalues of the weighted problem it follows that $p\in (1, +\infty)$ is a value at which the $k$-Morse index $m_k(u_p)$, $k\geq 2$ changes if and only if there exists $j\geq 2$ such that $k\mid j$ and $p\in \mathcal P^j$, where $\mathcal P^j$ is defined in .\ Moreover the change in the $k$-Morse index is always odd (precisely $\pm 1$).\ In particular a sufficient condition for $p$ to be a $k$-Morse index odd changing point is that $p\in \mathcal P^k$. The bifurcation result {#se:bifurcation} ====================== In this section we will find nonradial solutions to bifurcating from the curve of radial solutions $(p,u_p)$, looking for fixed points of the operator $T\ :\ (1,+\infty)\times C^{1,\alpha}_0(\bar B)\longrightarrow C^{1,\alpha}_0(\bar B)$ defined by $$\label{T} T(p,u)\ :=\left(-\Delta \right)^{-1}\left(\|u\|^{p-1}u\right).$$ We will restrict to the $\mathcal G_k$-invariant functions introduced in Section \[se:symmspaces\], in particular let us define the spaces $$\label{X} \mathcal X_k\ := C^{1,\alpha}(\bar B)\cap H^1_{0,k}(B), $$ where $H^1_{0,k}(B)$ is the symmetric space in . We also set $$\label{Xrad}\mathcal X_{{{\text{\upshape rad}}}}:=C^{1,\alpha}(\bar B)\cap H^1_{0,{{\text{\upshape rad}}}}(B).$$ Obviously $u_p\in\mathcal X_{{{\text{\upshape rad}}}}\subset \mathcal X_k$, for every $p\in (1,\infty)$ and for every $k\geq 1$.\ We will look for solutions in $\mathcal X_k$ which bifurcate at some point $(p^k, u_{p^k})$. Proposition \[lemma:degenerazioneSimmetria\]-$iii)$ characterizes the values of $p$ at which $u_p$ is degenerate in $\mathcal X_k$, we will show bifurcation for any $p$ in the subset $\mathcal P_k$ (see ) of degenerate values, for $k=3,4,5$. Observe that for any fixed $p$ the operator $T(p,\cdot)$ is compact and continuous in $p$ and that also its restriction to the subspaces $\mathcal X_k$, $k\geq 2$ is still compact (and continuous in $p$). In particular we will prove that the continuum of bifurcating solutions belongs to $\mathcal X_k\setminus\mathcal X_j$, $\forall j> k$ until they are non-radial, thus separating the branches related to different values of $k$. In order to get this property we restrict the operator $T$ to suitable cones $\mathcal K_k$ in $\mathcal X_k$, defined, similarly as in [@D2], by imposing some angular monotonicity to the $\mathcal G_k$-symmetric functions. Hence for $k\in\mathbb N_0$ let us define the cone: $$\label{K-k} \mathcal K_k\ : \ =\{v\in \mathcal X_k \ \text{ s.t. $v_{\theta}(r,\theta)\leq 0$ for $0\leq \theta\leq\frac{\pi}k$ , $0<r<1$} \},$$ where $ v_{\theta}$ denotes the derivative with respect to the angle $\theta$ of the polar coordinates. By definition $\mathcal X_{{{\text{\upshape rad}}}}\subset\mathcal K_k\subset \mathcal X_k$ for any $k\geq 1$ and the monotonicity in the definition implies the following separation property: $$\label{quelloCheGuadagnoConConi}\mathcal K_k\cap \mathcal K_h=\mathcal X_{{{\text{\upshape rad}}}},\quad \forall h\neq k,$$ which will be crucial in order to separate the branches. The complete statement of our bifurcation result is contained in Theorem \[teo-bif\] below, which is the main result of the section, Theorem \[teo1\] in the introduction follows from it. Let $\mathcal P^k$, $k\in\mathbb N_0$ be the subset of degenerate exponents defined in . By Lemma \[l5.1\] we know that $$\emptyset\neq \mathcal P^k=\{p_1^k,\ldots, p_{s_k}^k\}, \ \mbox{ when $\ k=3,4,5$}$$ (where $s_k\geq 1$ is an odd integer). We then have: \[teo-bif\] The points $(p^k_h, u_{p^k_h})$, $h\in \{1,\dots,s_k\}$ for $k=3,4,5$ are nonradial bifurcation points from the curve of radial solutions $(p,u_p)$ and the bifurcating solutions belong to the cone $\mathcal K_k$. The bifurcation is global and the Rabinowitz alternative holds. Moreover, for every $k=3,4,5$ there exists at least one exponent $p^k\in\{p^k_1,\dots,p^k_{s_k}\}$ such that, letting $\mathcal{C}_k$ be the continuum that branches out of $(p^k, u_{p^k})$ then either it is unbounded in $(1,+\infty)\times \mathcal K_k$ or it intersects $\{1\}\times \mathcal K_k$. Finally $\mathcal C_k\cap\mathcal C_j\subset \mathcal X_{{{\text{\upshape rad}}}}$ for any $j=3,4,5$, $j\neq k$. The proof of Theorem \[teo-bif\] can be found at the end of the section. The core of the proof consists in getting bifurcation at the degenerate points at which there is a change in the fixed point index of $T(p,\cdot)$ at $u_p$ relative to the cone $\mathcal K_k$ (index introduced in [@D83]). These degenerate points $(p,u_p)$ are given by any $p\in\mathcal P^k$ (see Proposition \[prop:ChangeIndexCone\]). Observe that at $p\in\mathcal P^k$ also the $k$-Morse index of $u_p$ has a (odd) change (see Remark \[rmk:OddChange\]). First we show that: \[lemma:TMandaConoInCono\] The operator $T(p,\cdot)$ maps $\mathcal X_k$ into $\mathcal X_k$ and in particular $\mathcal K_k$ into $\mathcal K_k$. Let $w\in \mathcal X_k$ and let $z=T(p,w)$. Since $w\in C^{1,\alpha}(B)$ then $z\in C^{3,\alpha}(B)$ and by definition of $T$, it is a classical solution to $$\label{z} \begin{cases} -\Delta z= \|w\|^{p-1}w & \text{ in } B, \\ z= 0 & \text{ on } \partial B. \end{cases}$$ Let $\tilde z(x)=z(g(x))$, for $g\in\mathcal G_k$. Then $\tilde z$ is a solution to , because $w\in \mathcal X_k$ and $-\Delta$ is invariant by the action of $\mathcal G_k$. This implies $\tilde z=z$ getting that $z\in \mathcal X_k$.\ It remains to show that when $w\in\mathcal K_k$ also the monotonicity assumption on $w$ is preserved by $T$. Since $z\in C^{3,\alpha}(B)$ we can compute $z_\theta=\frac{\partial z}{\partial \theta}$ and letting $w_\theta=\frac{\partial w}{\partial \theta}$, we have that $z_\theta$ is a classical solution to $$\begin{cases} -\Delta z_\theta = p\|w\|^{p-1}w_\theta & \text{ in } (0,1)\times (0,\frac \pi k), \\ z_\theta(1,\theta)= 0 & \text{ on } \partial B. \end{cases}$$ By assumption $w\in \mathcal K_k$ so that $w_\theta\leq 0$ in $(0,1)\times (0,\frac \pi k)$. Moreover $z_\theta(r,0)=0$ since $z$ is even in $\theta$ (see ) and moreover $z_\theta(r,\frac \pi k)=0$ by . The maximum principle then yields $z_{\theta}\leq 0$ for $0\leq \theta\leq\frac{\pi}k$ , $0<r<1$, concluding the proof. When $u_p$ is an isolated fixed point for $T(p,\cdot)$ we can consider its index relative to the cone $\mathcal K_k$ (see [@D83]), which we denote by $\mathit{ind}_{\mathcal K_k}\left(T(p,\cdot),u_p\right)$.\ We can compute $\mathit{ind}_{\mathcal K_k}\left(T(p,\cdot),u_p\right)$ when $u_p$ is non-degenerate in $\mathcal X_k$. In this case the characterization in Proposition \[lemma:degenerazioneSimmetria\]-$iii)$ implies in particular that $\beta_{1,{{\text{\upshape rad}}}}(p)+k^2\neq 0$, we then have: \[lemma:calcoloIndexCono\] Let $k\geq 2$ and $p$ be such that $u_p$ is non-degenerate in $\mathcal X_k$ then $$\mathit{ind}_{\mathcal K_k}\left(T(p,\cdot),u_p\right)=\left\{ \begin{array}{ll} 0 & \text{ if }\beta_{1,{{\text{\upshape rad}}}}(p)+k^2<0\\ {1}& \text{ if }\beta_{1,{{\text{\upshape rad}}}}(p)+k^2>0 \end{array}\right.$$ By Lemma \[lemma:TMandaConoInCono\] we can consider the operator $T$ restricted to the space $\mathcal X_k$, namely $T : (1,+\infty)\times \mathcal X_k\longrightarrow \mathcal X_k$ for some $k\geq 2$. Let us denote by $T'_u$ the Frechét derivative of $T$ with respect to $u$. Since $u_p$ is non-degenerate in $\mathcal X_k$, then $I-T'_u(p,u_p) : \mathcal X_k\longrightarrow \mathcal X_k$ is invertible. We can then apply Theorem 1 in [@D83] getting that $$\label{o2} \!\!\!\!\!\!\!\!\mathit{ind}_{\mathcal K_k}\left(T(p,\cdot),u_p\right)=\left\{ \begin{array}{ll} 0 & \text{ if } T'_u \text{ has the property } \alpha\\ \mathit{ind}_{\mathcal X_k}\left(T'_u(p,u_p),0\right) & \text{ if } T'_u \text{ does not have the property } \alpha \end{array}\right.$$ where we refer to [@D83] for the definition of the property $\alpha$. Moreover, since $u_p$ is isolated in $\mathcal X_k$ (again by its nondegeneracy) and since $I-T'_u(p,u_p)$ is invertible we have $$\label{o1} \mathit{ind}_{\mathcal X_k}\left(T'_u(p,u_p),0\right)=\lim_{r{\rightarrow}0}\mathit{deg}_{\mathcal X_k} \left(I-T'_u(p,\cdot), U_r(u_p),0\right)=(-1)^{m_k(u_p)}$$ where $\mathit{deg}$ is the usual Leray-Schauder degree in the Banach space $\mathcal X_k$, $U_r(u_p):=\{w\in \mathcal X_k\ : \ \\|u_p-w\\|<r\}$ and the last equality follows by standard results for the Leray Schauder degree of linear, compact, invertible maps (see for instance [@AM]). The characterization of the degeneracy in $\mathcal X_k$ (see Proposition \[lemma:degenerazioneSimmetria\]-$iii)$) implies in particular that $\beta_{1,{{\text{\upshape rad}}}}(p)+k^2\neq 0$ at the non-degenerate point $p$, the rest of the proof is devoted to show that $$\label{propertyalpha} \mbox{$T'_u$ has the property $\alpha$ if and only if $\beta_{1,{{\text{\upshape rad}}}}(p)+k^2<0$.}$$ In this case indeed and implies the result since by Lemma \[lemma:legameAutovSimmNegativoLinEPesato\] and Lemma \[LemmaMorseIndexRadiale\] one has $$m_k(u_p)=2,\mbox{ when }\beta_{1,{{\text{\upshape rad}}}}(p)+k^2>0.$$ The property $\alpha$ in is stated in [@D83 Lemma 2]. Following the same notations we have that the linear map $T'_u(p,u_p)$ has the property $\alpha$ if and only if (Lemma 2-(a) of [@D83]) the spectral radius of $T'_u(p,u_p)$ is greater than 1 when restricted to the orthogonal complement to $\mathcal X_{{{\text{\upshape rad}}}}$ in $\mathcal X_k$, which we denote by $\mathcal X_{{{\text{\upshape rad}}}}^\perp$ (observe that in our case the subspace $S_{u_p}$ in [@D83] is $\mathcal X_{{{\text{\upshape rad}}}}$). Equivalently, as observed also in [@D2 proof of Theorem 1], $T'_u(p,u_p)$ has the property $\alpha$ if and only if there exist $t\in(0,1)$ and $h\in \mathcal X_{{{\text{\upshape rad}}}}^\perp$ such that $h=tT'_u(p,u_p)h$, namely, recalling the definition of $T$, such that the linear equation $$\label{o3} \begin{cases} -\Delta h-tp\|u_p\|^{p-1}h=0 & \text{ in }B\\ h=0 &\text{ on }\partial B \end{cases}$$ admits a nontrivial solution $h\in \mathcal X_{{{\text{\upshape rad}}}}^{\perp}$ for some $t\in(0,1)$. This is equivalent to say that zero is an eigenvalue of the problem $$\begin{cases} -\Delta h-tp\|u_p\|^{p-1}h=\mu h & \text{ in }B\\ h=0 &\text{ on }\partial B \end{cases}$$ with eigenfunction in $\mathcal X_{{{\text{\upshape rad}}}}^\perp$ for some $t\in(0,1)$. We denote by $\mu_t$ the smallest eigenvalue of this problem in $\mathcal X_{{{\text{\upshape rad}}}}^\perp$, which depends on $t$. By the variational characterization of the eigenvalues $\mu_t$ is decreasing in $t$. Moreover $\mu_0>0$, since when $t=0$ $\mu_0$ is the first Dirichlet eigenvalue in $\mathcal X_{{{\text{\upshape rad}}}}^\perp$ of the Laplace operator in $B$ which is strictly positive. When $t=1$ instead $\mu_1$ is the smallest eigenvalue in $\mathcal X_{{{\text{\upshape rad}}}}^\perp$ of the linearized operator $L_p$. When $\mu_1$ is negative then there exists a $t\in(0,1)$ such that has a solution in $\mathcal X_k\setminus\mathcal X _{{{\text{\upshape rad}}}}$. When $\mu_1$ is positive instead then $\mu_t>\mu_1>0$ for any $t\in (0,1)$ and equation does not have a solution in $\mathcal X_k\setminus\mathcal X _{{{\text{\upshape rad}}}}$. Finally from Lemma \[lemma:legameAutovSimmNegativoLinEPesato\] we have that $\mu_1<0$ if and only if $\beta_{1,{{\text{\upshape rad}}}}(p)+k^2<0$ and this concludes the proof of . As a consequence one can characterize the set of the points $p$ at which the index $\mathit{ind}_{\mathcal K_k}\left(T(p,\cdot),u_p\right)$ changes: \[prop:ChangeIndexCone\] $p\in (1,+\infty)$ is a value at which $\mathit{ind}_{\mathcal K_k}\left(T(p,\cdot),u_p\right)$ changes, for $k\geq 2$ if and only if $p\in\mathcal P^k$, where the set $\mathcal P^k$ is the one defined in . If $p\in \mathcal P^k$ then $(p,u_p)$ is an isolated degenerate point (Lemma \[l5.1\]), as a consequence the values $p=p_h^k\pm\delta$ are non-degenerate for any $\delta>0$ small and by definition of $\mathcal P_k$ we also have $[\beta_{1,{{\text{\upshape rad}}}}(p+\delta)+k^2][\beta_{1,{{\text{\upshape rad}}}}(p-\delta)+k^2]<0$. The conclusion then follows by Lemma \[lemma:calcoloIndexCono\] applied at the points $p=p_h^k\pm\delta$.\ Viceversa if $\mathit{ind}_{\mathcal K_k}\left(T(p,\cdot),u_p\right)$ changes at $p$ then by Lemma \[lemma:calcoloIndexCono\] $p$ satisfies $\beta_{1,{{\text{\upshape rad}}}}(p)=-k^2$ and $\beta_{1,{{\text{\upshape rad}}}}(p)+k^2$ changes sign at $p$. This implies that necessarily $p\in\mathcal P^k$. Proof of Theorem \[teo-bif\] ---------------------------- [Step 1.]{} [Non-radial local bifurcation in $\mathcal K_k$]{}\ Let us consider $p_h^k$ for a certain $h\in\{1,\ldots, s_k\}$. By Proposition \[prop:ChangeIndexCone\] we know that $ \mathit{ind}_{\mathcal K_k}\left(T(p,\cdot),u_p\right)$ changes as $p$ crosses $p_h^k$, namely that for any $\delta>0$ small $$\label{indiceDiv} \mathit{ind}_{\mathcal K_k} \left(T( p^k_h-\delta ,\cdot),u_{p^k_h-\delta}\right) \neq \mathit{ind}_{\mathcal K_k} \left(T( p^k_h+\delta ,\cdot),u_{p^k_h+\delta}\right),$$ we now show that $(p^k_h, u_{ p^k_h})$ is a bifurcation point in $(1,+\infty)\times \mathcal K_k$.\ Hence let us assume by contradiction that $(p^k_h, u_{ p^k_h})$ is not a bifurcation point in $(1,+\infty)\times \mathcal K_k$, then we can find $\delta>0$ and a neighborhood $\mathcal O$ of $\{(p,u_p): p\in ( p^k_h-\delta, p^k_h+\delta)\} $ in $( p^k_h-\delta, p^k_h+\delta)\times \mathcal K_k$ such that $u-T(p,u)\neq 0$ for every $(p,u)$ in $\mathcal O$ different from $(p,u_p)$. We can choose $\delta>0$ such that holds. Letting $\mathcal O_p:=\{v\in \mathcal K_k\ : \ (p,v)\in \mathcal O\}$, it then follows that there are no solutions to $u-T(p,u)= 0$ on $\cup_{p\in (p^k_h-\delta, p^k_h+\delta) }\{p\} \times \partial \mathcal{O}_p$ and there is only the radial solution $(p,u_p)$ in $\left(\{ p^k_h-\delta\}\times \mathcal{O}_{ p^k_h-\delta }\right)\cup \left(\{ p^k_h+\delta \}\times \mathcal{O}_{ p^k_h+\delta }\right)$. By the homotopy invariance of the fixed point index in the cone, see [@D83], then we have that $$\mathit{ind}_{\mathcal K_k}\left(T(p,\cdot),u_p\right) \ \ \hbox{ is constant for } \ p\in( p^k_h-\delta, p^k_h+\delta),$$ which is in contradiction with . This proves the local bifurcation. The bifurcating solutions belong to $\mathcal K_k$ since $T$ maps the cone in itself (Lemma \[lemma:TMandaConoInCono\]) and are non-radial for $p$ close to $p_h^k$ since $u_p$ is radially non-degenerate by Lemma \[lemma:radiallyNonDeg\]. \ We can adapt the proof of Theorem 3.3 in [@G10]. One of the main differences is that now, since the cone $\mathcal K_k$ is not a Banach space, we substitute the Leray-Schauder degree used in [@G10] with the degree in the convex cone $\mathcal K_k$, which we denote by $\mathit{deg}_{\mathcal K_k}(I-T(p,\cdot), \mathcal O,0)$, for any open (with the induced topology) set $\mathcal O$ in $\mathcal K_k$. The degree in the convex cone has been introduced in [@Amman] (where it is called [index]{}), its definition arises directly from the Leray-Schauder degree (to which it coincides when the cone is a Banach space) and in particular it admits the same properties of the Leray-Schauder degree (normalization, additivity, homotopy invariance, permanence, excision, solution property, etc, see [@Amman Theorem 11.1 and 11.2]).\ Following [@G10], let $\mathcal S\ : =\left\{(p,u_p): p\in (1,+\infty)\right\} \subseteq (1,+\infty)\times \mathcal K_k$ be the curve of radial least-energy solutions, let $\Sigma_k$ be the closure of the set $ \{(p,v)\in \left((1,+\infty)\times \mathcal K_k\right)\setminus \mathcal S : v \text{ solves }\eqref{problem}\}$ and let $\mathcal C_k$ be the closed connected component of $\Sigma_k$ bifurcating from $( p^k_h , u_{ p^k_h})$. Assume by contradiction that the Rabinowitz alternative, namely one of the following, does not occur: - $\mathcal C_k$ is unbounded in $(1,+\infty)\times \mathcal K_k$; - $\mathcal C_k$ intersects $\{1\}\times \mathcal K_k$; - there exists $ p_l^k$ with $l\neq h$ such that $( p^k_l, u_{ p^k_l})\in\mathcal C_k\cap \mathcal S$. Then as in Step 2 in the proof of [@G10 Theorem 3.3] we can then construct a suitable neighborhood $\mathcal O$ of $\mathcal C_k$ in $\mathcal K_k$ such that $\partial \mathcal O\cap \Sigma_k=\emptyset$, $\mathcal O\cap \mathcal S\subset (p^k_h-\delta, p^k_h+\delta)\times \mathcal K_k$ for $\delta$ such that $ u_{p_h^k\pm \delta}$ is nondegenerate and moreover there exists $c_0>0$ such that $\\| v-u_p\\|_{\mathcal X_k}\geq c_0$ for $(p,v)\in \mathcal O$ such that $\|p-p^k_h\|\geq \delta$. Then we can follow the proof of Step 3 and Step 4 in [@G10 Theorem 3.3], recalling now that, for $\Lambda_c:=\{(p,v)\in (1,+\infty)\times \mathcal X_k : \\| v-u_p\\|_{\mathcal X_k}< c\} $ one has $$\mathit{deg}_{\mathcal K_k}(I-T(p_h^k\pm \delta,\cdot), \left(\mathcal O\cap \Lambda_c\right)_{p_h^k\pm \delta},0)=\mathit{ind}_{\mathcal K_k}(T(p_h^k\pm \delta,\cdot),u_{p_h^k\pm \delta})$$ for any $c<c_0$. The fixed point index relative to the cone $\mathcal K_k$ can be then computed in $p_h^k\pm \delta$ and it assumes either the value $0$ or $1$ (Lemma \[lemma:calcoloIndexCono\]). The proof of Step 3 and 4 of [@G10 Theorem 3.3] can be repeated and so we get a contradiction.\ We can now adapt the proof of [@G16 Proposition 2.3], again using the degree in the convex cone $\mathcal K_k$ which is, as already observed, either $0$ or $1$ in a neighborhood of the isolated (in $\mathcal X_k$) solution $u_p$. The main difference is that, in the final part of the proof of [@G16 Proposition 2.3] we now obtain, following the notations of [@G16], that $$\mathit{deg}_{\mathcal K_k}(S_r(p,v), \mathcal O\cap B_r(p_l^k,u_{ p_l^k}),0)=\pm 1$$ for every $p_l^k\in \mathcal P^k$. This implies again that the number of points $p_l^k\in \mathcal P^k$ which belong to $\mathcal C_k$, including $( p^k_h, u_{ p^k_h})$, has to be even if $\mathcal C_k$ is bounded. Since the total number $s_k$ of points in $\mathcal P^k$ is odd (see Lemma \[l5.1\]), then there exist at least one value $p^k\in\{p^k_h\}_{h=1,\ldots, s_k}$ at which either $i)$ or $ii)$ holds. \ Since the bifurcating solutions are not radial for $p$ close to $p_h^k$, the separation property implies that near the bifurcation points $\mathcal{C}_k\neq \mathcal{C}_i$ if $k\neq i$. Moreover $(\mathcal{C}_k\cap \mathcal{C}_i)\subset(\mathcal K_k\cap\mathcal K_j)$ hence it contains only radial solutions. \[remark:shape\] Observe that from the definition of the space $\mathcal X_h$ and from the separation property of $\mathcal K_k$ it follows that $$\label{quelloCheGuadagnoConConiSeparazione}\mathcal K_k\cap \mathcal X_h=\mathcal X_{{{\text{\upshape rad}}}},\quad \forall h> k$$ and so, as stated in Theorem \[teo1\] in the introduction, either the bifurcating solution belongs to $\mathcal X_k\setminus\mathcal X_j$, $\forall j>k$ or it is radial.\ Moreover, since the kernel of the linearized operator is one dimensional when restricted to the spaces $\mathcal X_k$ (Proposition \[lemma:degenerazioneSimmetria\]-iii)), we can get an expansion of the bifurcating solution found in Theorem \[teo-bif\] near the bifurcation point $(p^k, u_{p^k})$, even if we cannot apply the Crandall-Rabinowitz result to obtain some regularity on the solutions set. Indeed, applying Proposition 2.4 in [@G16] we know that there exists $\varepsilon_0>0$ such that for any $0<\varepsilon<\varepsilon_0$ if $(p,v) \in \mathcal{C}_k\cap \left(B_{\varepsilon}( p^k, u_{p^k})\setminus\{( p^k, u_{p^k})\}\right)$, then $$v(r,\theta)= u_p(r)+\alpha_{\varepsilon} \phi_1(r) \cos \left(k \theta\right)+ \psi_{\varepsilon}(r,\theta)$$ where $\alpha_{\varepsilon}{\rightarrow}0$ as $\varepsilon{\rightarrow}0$, $\phi_1(r)>0$ is a first eigenfunction of the weighted eigenvalue problem as defined in Proposition \[p4.7\] and $\psi_{\varepsilon}(r,\theta)\in \mathcal X_k$ is such that $\\|\psi_{\varepsilon}\\|_{\infty}=o(\alpha_{\varepsilon})$ as $\varepsilon{\rightarrow}0$. As a consequence, near the bifurcation point, the solutions we found not only are in $\mathcal X_k\setminus \mathcal X_{{{\text{\upshape rad}}}}$ but, being small perturbation of the radial least energy solution $u_p$, they also inherit from $u_p$ the property of having two nodal domains and of being quasi-radial in the sense of Definition \[def:quasiradial\].\ We remark that along the branch the number of nodal regions of the solutions may change and that moreover far from the bifurcation point they may also loose the quasi-radial shape* and their nodal line could touch the boundary. Observe that we can obtain a solution to by rotating the solution $v$ in Theorem \[teo-bif\] of an angle $\alpha$. This solution coincides with the one bifurcating from $u_p$ in the direction $$w(r,\theta)=\phi_1(r)\left(a\sin (k\theta)-b\cos (k\theta)\right) \in Ker(L_p)$$ with $\alpha=\arctan (-a/b)$, letting $\hat \tau$ be the reflection with respect to the hyperplane $ax+by=0$ and restrincting to the spaces $$\widehat{\mathcal X}_k \ := C^{1,\alpha}_0( B)\cap \widehat{H}^1_{0,k}(B),$$ where $\widehat{H}^1_{0,k}(B)\ : =\{v\in H^1_0(B)\ \text{ such that } v(g(x))=v(x), \ \ \forall g\in \widehat{\mathcal{G}}_k , \ \forall x\in B \}$ and $\widehat{\mathcal{G}}_k\subset O(2)$ is the group generated by $O_k$ and by the reflection $\hat \tau$. We stress that in order to get the bifurcation result one could work directly in the space $\mathcal X_k$, $k=3,4,5$ without restricting to the cones $\mathcal K_k\subset\mathcal X_k$ substituting the degree in the cone $\mathcal K_k$ with the usual Leray-Schauder degree in $\mathcal X_k$.\ Anyway the bifurcation result obtained in this way is only partial, since a priori different branches of solutions could coincide.\ The advantage of restricting to the cones $\mathcal K_k$ in the proof of Theorem \[teo-bif\] is that set $\mathcal K_k\cap\mathcal K_j$ contains only radial functions when $k\neq j$, and this allow to separate the branches. The proof of Theorem \[prop1.4\] {#section:proofTeoLeastEnergy} ================================ Let us consider the functional $E_p$ in restricted to the space $H^1_{0,k}(B)$, $k\in\mathbb N_0$ defined in . By the principle of symmetric criticality critical points of $E_p$ in $H^1_{0,k}(B)$ are solutions to which are invariant by the action of $\mathcal{G}_k$. In particular (see [@CastroCossioNeuberger]) to produce nodal solutions to which are invariant by the action of $\mathcal{G}_k$ one can minimize the functional $E_p$ on the nodal symmetric Nehari set $$M_k\ :=\{v\in H^1_{0,k}(B)\ : \ v^+\neq 0,\ v^-\neq 0,\ E_p'(u)u^+=E_p'(u)u^-=0\}$$ where $E_p'$ is the Fréchet derivative of $E_p$. Then a function $\bar u$ such that $$E_p(\bar u)=\inf_{u\in M_k}E_p(u)$$ is a least energy sign-changing $\mathcal{G}_k$-invariant solution to and we denote it by $u_p^k$, for $k=1,2,\dots$. The proof of the [non-radial part]{} in Theorem \[prop1.4\] follows directly by comparing the $k$-Morse index of the radial solution $u_p$ (computed in Section \[se:symmspaces\]) with the $k$-Morse index of the least energy symmetric solution $u_p^k$. Indeed, following the same arguments in [@BartschWeth Theorem 1.3] and working in the space of symmetric functions $H^1_{0,k}(B)$, one can prove the following result: \[lem-10-1\] Let $u_p^k$ be a least energy sign-changing solution to in the space $H^1_{0,k}(B)$. Then $$\label{k-morse-upk} m_k\left( u_p^k\right)=2, \quad\forall p\in (1,+\infty),$$ where $m_k$ denotes the $k$-Morse index of $u_p^k$. By comparing with the information in Propositions \[Morse-simmetrico1\] and \[Morse-simmetrico2\] we get then the proof of the non-radial part of Theorem \[prop1.4\] since necessarily $u_p^k$ is not radial for any $p$ and $k$ such that $m_k(u_p)>2$. The proof of the [radial part]{} of Theorem \[prop1.4\] is more involved and is the goal of the rest of this section where first we show an $L^{\infty}$ bound for the solution $u_p^k$ for $p$ close to $1$ (Proposition \[prop7.4\]) and then, using this bound, we deduce the result by studying the asymptotic behavior of the solutions $u_p^k$ as $p\rightarrow 1$ (this is done in the proof of Proposition \[leastRadiale\]). As already discussed in the introduction we do not have a bound for the full Morse index of $u_p^k$, but [only for the $k$-Morse index]{} (Lemma \[lem-10-1\] above), for this reason, exploiting the symmetry of $u_p^k$, we reduce problem from the ball $B$ to the circular sector $S_k$ of the ball defined in polar coordinates as $$S_k:=\{(r,\theta)\ : \ 0<r<1 \ \ ,\ \ 0<\theta<\frac{\pi}k\}.$$ Indeed setting $\Gamma_1:=\{(r,\theta): \ r=1, \ \theta\in(0,\frac \pi k)\}$, $\Gamma_2 :=\{(r,\theta): \ \theta=0, \ r\in(0,1)\}$, $\Gamma_3:= \{(r,\theta): \ \theta=\frac \pi k,\ r\in(0,1)\}$, $A=(\cos \frac{\pi}k, \sin \frac{\pi}k)$ and $B=(1,0)$, one has $\partial S_k=\Gamma_1\cup \Gamma_2\cup \Gamma_3\cup \{O,A,B\}$ and any regular function $v$ to which is invariant by the action of the group $\mathcal{G}_k$, satisfies $$v\in C^1(S_k\cup\Gamma_2\cup \Gamma_3\cup O)\ \ , \ \ \frac{\partial v }{\partial \nu}=0 \ \ \ \text{ on } \Gamma_2\cup \Gamma_3$$ where $\nu$ denotes the outer normal vector to the boundary of $S_k$. Hence $u_p^k$ is a classical solution to $$\label{sector} \begin{cases} -\Delta u_p^k=\|u_p^k\|^{p-1}u_p^k & \text{ in }S_k\\ u_p^k=0 & \text{ on } \Gamma_1\\ \frac{\partial u_p^k }{\partial \nu}=0 & \text{ on } \Gamma_2\cup \Gamma_3. \end{cases}$$ In next result we convert the bound on the $k$-Morse index in into a bound on the full [[mixed-Morse index]{}]{} of $u_p^k$ in the sector $S_k$. \[lem-10-nuovo\] Let $u_p^k$ be the least energy sign-changing solution to in the space $H^1_{0,k}(B)$. Then for any $p\in (1,+\infty)$ the mixed eigenvalue problem $$\label{sector-lin} \begin{cases} -\Delta v=p\|u_p^k\|^{p-1}v +\mu v& \text{ in }S_k\\ v=0 & \text{ on } \Gamma_1\\ \frac{\partial v}{\partial \nu}=0 & \text{ on } \Gamma_2\cup \Gamma_3 \end{cases}$$ admits only $2$ negative eigenvalues $\mu$. Because of Lemma \[lem-10-1\] the Dirichlet eigenvalue problem $$\label{nuovo-lin} \begin{cases} -\Delta v=p\|u_p^k\|^{p-1}v +\mu v& \text{ in }B\\ v=0 & \text{ on } \partial B \end{cases}$$ admits only two linearly independent eigenfunctions $\tilde \psi_1$ and $\tilde\psi_2$ which are invariant by the action of $\mathcal{G}_k$, are regular, by elliptic regularity theory, and which correspond to a negative eigenvalue, say $\mu_1^k$ and $\mu_2^k$. By the symmetry properties of $\tilde \psi_i$ it is straightforward to see, that, the restriction of $\tilde \psi_i$ to the sector $S_k$ satisfies corresponding to the same eigenvalue $\mu_i^k<0$ for $i=1,2$. This shows that the number of negative eigenvalues of is at least two. Viceversa, if problem possess $m>2$ negative eigenvalues $\mu_i$ corresponding to the eigenfunctions $\psi_1,\dots,\psi_m$ (that we take orthogonal in $L^2(S_k)$), then, denoting by $\tilde \psi_1,\dots,\tilde \psi_m$ the extension of $\psi_1,\dots,\psi_m$ to $B$ under the action of $\mathcal{G}_k$, it is easy to see that $\tilde \psi_1,\dots,\tilde \psi_m\in H^1_{0,k}(B)$ solve corresponding to the eigenvalues $\mu_1<\dots\leq \mu_m<0$ and are orthogonal in $L^2(B)$ contradicting Lemma \[lem-10-1\]. This shows that the number of negative eigenvalues for problem is at most two concluding the proof. In order to get an uniform $L^{\infty}$ bound for the solution $u_p^k$ we want to perform a blow-up argument in the sector $S_k$ exploiting the uniform bound of the mixed Morse index in Lemma \[lem-10-nuovo\].\ This blow-up procedure in $S_k$ requires special care, since we have to deal with mixed boundary conditions and above all with the angular points of $S_k$. For these reasons the analysis of the rescaled solutions includes several different cases, depending on the location of the maximum points in the sector which gives different shapes of the limiting domain. Anyway in all the cases we end-up with solutions to a limit linear problem in unbounded domains with either Dirichlet or Neumann or mixed boundary conditions, whose Morse index (or symmetric Morse index) is finite. In order to rule-out this possibility we will need the following symmetric version of a well known non-existence result: \[prop7.3\] Let $\Sigma$ be either ${\mathbb{R}}^2$ or ${\mathbb{R}}^2_{+}:=\{(x,y)\in {\mathbb{R}}^2\ : y> 0\}$ and let $\mathcal{G}$ be any subgroup of $O(2)$ which preserves $\Sigma$. Let $u$ be any nontrivial solution to the problem $$\label{prob-limite} -\Delta u-u=0 \ \ \text{ in }\Sigma$$ and when $ \Sigma={\mathbb{R}}^2_{+}$ assume also that $$\label{boundary-cond} u=0 \ \ \text{ on }\partial \Sigma.$$ Then, the $\mathcal{G}$-Morse index of $u$ is not finite.\ Here the $\mathcal{G}$-Morse index of a solution $u$ to is the maximal dimension of a subspace $X\subseteq C^{\infty}_{0,\mathcal{G}}(\Sigma)$ such that $$\label{quad} Q(v):=\int_{\Sigma} \left[\|\nabla v\|^2-\|v\|^2\right]dx <0, \ \forall v\in X\setminus\{0\},$$ where $C^{\infty}_{0,\mathcal{G}}(\Sigma)$ denotes the subspace of $C^{\infty}_0(\Sigma)$ of the functions invariant with respect to the action of $\mathcal{G}$. Let us consider first the case of $\Sigma={\mathbb{R}}^2$. Let us denote, as usual, by $\lambda_j$, $j\in\mathbb N$, the Dirichlet eigenvalues of $-\Delta$ in $B$, since $\mathcal{G} $ preserves $B$, we can consider among them the subsequence $\lambda_j^{\mathcal{G}}$ of the eigenvalues corresponding to $\mathcal G$-invariant eigenfunctions.\ Let $\psi_j^\mathcal{G}$ be the $\mathcal G$-invariant eigenfunction associated to $\lambda_j^{\mathcal{G}}$, then it is easy to see that the function $\widehat \psi_j^\mathcal{G}(x):=\psi_j^\mathcal{G}\left(\frac xR\right)$, where $R>0$, solves $$\label{autof-riscalate} \begin{cases} -\Delta\widehat \psi_j^{\mathcal{G}} =\frac{\lambda_j^{\mathcal{G}}}{R^2} \widehat \psi_j^\mathcal{G} & \text{ in }B_R\\ \widehat \psi_j^{\mathcal{G}}=0 & \text{ on }\partial B_R, \end{cases}$$ where $B_R$ is the ball centered at the origin with radius $R$.\ Observe that for any integer $m>0$ and for any subgroup $\mathcal{G}$ of $ O(2)$ there exists $R>0$ such that $\frac{\lambda_1^{\mathcal{G}}}{R^2}< \dots \leq \frac{\lambda_m^{\mathcal{G}}}{R^2}<1$, so that by we get $$Q\left(\widehat\psi_j^{\mathcal{G}}\right)=\int_{\Sigma} \left[\|\nabla \widehat\psi_j^{\mathcal{G}} \|^2-\|\widehat \psi_j^{\mathcal{G}} \|^2\right]dx=\left(\frac{\lambda_j^{\mathcal{G}}}{R^2}-1\right) \int_{\Sigma}\| \widehat\psi_j^{\mathcal{G}} \|^2dx<0, \mbox{ for }j=1,\dots,m$$ Since the functions $\widehat \psi_1^{\mathcal{G}}, \dots, \widehat \psi_m^{\mathcal{G}}\in C^{\infty}_{0,\mathcal G}(\Sigma)$ and are linearly independent (and orthogonal in $L^2(B_R)$), this means that the $\mathcal{G}$-Morse index of any nontrivial solution $u$ to is greater or equal than $m$, for any $m\in\mathbb N$ showing the result in case of $\Sigma={\mathbb{R}}^2$. When $\Sigma={\mathbb{R}}^2_{+}$ we let $\lambda_j^+$ be the sequence of Dirichlet eigenvalues of $-\Delta$ in $B\cap {\mathbb{R}}^2_+$ and $(\lambda_j^+)^{\mathcal G}$ the subsequence of the eigenvalues invariant with respect to the action of $\mathcal{G}$ with associated $\mathcal G$-invariant eigenfunctions $\psi_j^{\mathcal{G}}$. Then defining as before the rescaled function $\widehat \psi_j^\mathcal{G}$, it solves $$\begin{cases} -\Delta\widehat \psi_j^{\mathcal{G}} =\frac{(\lambda_j^+)^{\mathcal{G}}}{R^2} \widehat \psi_j^\mathcal{G} & \text{ in }B_R\cap {\mathbb{R}}^2_+ \\ \widehat \psi_j^{\mathcal{G}}=0 & \text{ on }\partial \left(B_R\cap {\mathbb{R}}^2_+\right) \end{cases}$$ and the thesis follows similarly as in the previous case. We are now ready to perform the blow-up analysis in $S_k$ to get a uniform $L^{\infty}$ bound for the solutions $u_p^k$. \[prop7.4\] Let $u_p^k$ be a least energy sign-changing solution to in the space $H^1_{0,k}(B)$ and let $\delta>0$. Then there exists $C>0$ such that $${\lVertu_p^k\rVert}_{\infty}^{p-1}\leq C,\quad \mbox{ for any }p\in (1,1+\delta).$$ Assume by contradiction that there exists a sequence $p_n{\rightarrow}1$ such that, letting $M_n:={\lVertu_n\rVert}_{\infty}$ with $u_n:=u_{p_n}^k$, $M_n^{p_n-1}{\rightarrow}\infty$ as $n{\rightarrow}\infty$. Let $P_n=(x_n,y_n)$ be the points at which $\|u_n(P_n)\|=M_n$. W.l.o.g. we can assume $u_n(P_n)=M_n$ and, by the symmetry properties of $u_n$, also that $P_n\in S_k\cup \Gamma_2\cup \Gamma_3\cup\{O\}$. We may also assume that $$P_n{\rightarrow}P_0:=(x_0,y_0)\in \bar S_k.$$ We restrict the functions $u_n$ to the sector $S_k$ and define the functions $$\widetilde{u}_n(x,y):=\frac 1{M_n}{u_n(M_n^{\frac {1-p_n}2}(x,y)+P_n)},$$ that satisfy $$-\Delta \widetilde{u}_n=\|\widetilde{u}_n\|^{p_n-1}\widetilde{u}_n$$ in $\Omega_n:=M_n^{\frac {p_n-1}2}\left( S_k-P_n\right)$. In the sequel we analyze the asymptotic behavior of the rescaled functions $\widetilde u_n$ and get a contradiction by mean of Proposition \[prop7.3\]. We need to consider several cases depending upon the localization of the limit point $P_0$ in $\bar S_k$. The underlying idea of each case is that the sequence of solutions $\widetilde{u}_n$ converges to a non-trivial solution $\widetilde u$ to either in ${\mathbb{R}}^2$ or in a halfplane with Dirichlet boundary conditions. Moreover the bound on the Morse index of $\widetilde{u}_n$ obtained in Lemma \[lem-10-nuovo\] is preserved when passing to the limit problem. This last property, together with Proposition \[prop7.3\], implies $\widetilde u=0$ giving always a contradiction. Thus $M_n^{p_n-1}$ is bounded and this ends the proof. Observe that by definition $(\widetilde x, \widetilde y)\in \Omega_n$ if and only if $$\widetilde x= M_n^{\frac {p_n-1}2}(x-x_n) \ \ \ \text{ and }\ \ \ \widetilde y= M_n^{\frac {p_n-1}2}(y-y_n)$$ for some $(x,y)\in S_k$, moreover a point $(x,y)$ belongs to $ S_k$ if and only if $$\label{AppS_k}x>0 \, , \ \ \ y> 0\, ,\ \ \ \frac yx<\tan \frac \pi k \ \ \ \text{ and }\ \ \ 0<x^2+y^2<1.$$ As a consequence we deduce that $(\widetilde x, \widetilde y)\in \Omega_n$ if and only if the following inequalities are all satisfied: $$\begin{aligned} &&M_n^{\frac {1-p_n}2}\widetilde x+x_n> 0 \, , \label{10-cc}\\ &&M_n^{\frac {1-p_n}2}\widetilde y+y_n> 0\, , \label{10-c}\\ &&\frac{M_n^{\frac {1-p_n}2}\widetilde y+y_n }{M_n^{\frac {1-p_n}2}\widetilde x+x_n }<\tan \frac \pi k \label{10-a}\\ &&0<x_n^2+y_n^2+M_n^{{1-p_n}}\left(\widetilde x^2+ \widetilde y^2\right)+2M_n^{\frac {1-p_n}2}\left(\widetilde xx_n+\widetilde yy_n\right)<1\label{10-b}\end{aligned}$$ From now on we denote by $d_n$ the distance between $P_n$ and $\partial S_k$, namely $$\label{dn} d_n:=\min_{P\in\partial S_k}\|P_n-P\|.$$ [Step 1.]{} [$P_0\in S_k$]{}\ Observe that in this case $d_nM_n^{\frac {p_n-1}2}\rightarrow +\infty$ as $n\rightarrow +\infty$. Indeed, since $P_0\in S_k$, by $x_0>0$, $y_0>0$, $x_0^2+y_0^2<1$ and $\frac{y_0}{x_0}<\tan \frac \pi k$, so that, since $M_n^{p_n-1}{\rightarrow}\infty$ as $n\rightarrow +\infty$, any point $(\widetilde x,\widetilde y)\in B_R$ satisfies , , and , for $n$ large enough, namely for any $R>0$ $B_R\subseteq \Omega_n$ for $n$ large enough.\ Elliptic estimates imply that, up to a subsequence $ \widetilde{u}_n{\rightarrow}\widetilde u$ uniformly on compact sets of ${\mathbb{R}}^2$. By the argument in [@GS] $\widetilde u$ is defined in all of ${\mathbb{R}}^2$, it is a nontrivial weak solution to in $\Sigma={\mathbb{R}}^2$ and satisfies $\widetilde u(0)=1$.\ Finally we show that the Morse index of the limit function $\widetilde u$ is less or equal than $2$, this contradicts Proposition \[prop7.3\] and proves the thesis in the case $P_0\in S_k$.\ Assume, by contradiction, that the Morse index of $\widetilde u$ as a solution to is greater than $2$. Then there exist at least $3$ functions $\widetilde\psi_1,\widetilde\psi_2,\widetilde\psi_3\in C^{\infty}_0({\mathbb{R}}^2)$ such that $\widetilde\psi_i$ are linearly independent (orthogonal in $L^2({\mathbb{R}}^2)$) and $$Q(\widetilde\psi_i)<0$$ where $Q$ is the quadratic form as defined in . Since $ \widetilde\psi_i$ are supported in a ball $B_R$ then, the uniform convergence of $\widetilde{u}_n{\rightarrow}\widetilde u$ on compact sets of ${\mathbb{R}}^2$ implies that $$\int_{{\mathbb{R}}^2}\|\nabla \widetilde\psi_i\|^2 - p_n\|\widetilde{u}_n\|^{p_n-1} \widetilde\psi_i^2 <0$$ for $n$ large enough. Then the functions $\widehat\psi_i(x,y):=\widetilde \psi_i\left(\frac{(x,y)-P_n}{M_n^{\frac{p_n-1}2}}\right)$ belong to $C^{\infty}_0(S_k)$ for $n$ large enough, are orthogonal in $L^2(S_k)$ and satisfy $$\int_{S_k}\|\nabla\widehat\psi_i\|^2 - p_n\|{u}_n\|^{p_n-1}\widehat\psi_i^2 <0$$ for $i=1,2,3$. Then, letting $\psi_i\in C^{\infty}_0(B) $ be the $\mathcal{G}_k$-invariant extension of $\widehat\psi_i$ to the ball $B$, it holds $$\int_{B}\|\nabla\psi_i\|^2 - p_n\|{u}_n\|^{p_n-1}\psi_i^2 <0$$ for $i=1,2,3$ contradicting the fact that the $k$-Morse index of ${u}_n$ is two (Lemma \[lem-10-1\]). \ In this case we have to consider the two possibilities either $d_nM_n^{\frac{p_n-1}2}{\rightarrow}\infty$ or $d_nM_n^{\frac{p_n-1}2}{\rightarrow}s>0$, for $d_n$ as in (the fact that $s>0$ is a consequence of the Dirichlet boundary conditions on $\Gamma_1$ and can be deduced exactly as in the paper [@GS]). Then, as in the proof in [@GS] the rescaled functions $ \widetilde{u}_n^k{\rightarrow}\widetilde u$ as $n{\rightarrow}\infty$ uniformly on compact sets of $\Sigma$, where $\widetilde u$ is a nontrivial solution (recall that $\widetilde{u}(0)=1$) either to in $\Sigma={\mathbb{R}}^2$ in the first case or in $\Sigma={\mathbb{R}}^2_{+}$ in the second case (up to a rotation and a translation) satisfying . Moreover one can prove similarly as in [Step 1]{} that $\widetilde u$ has finite Morse index, contradicting again Proposition \[prop7.3\]. \ We give the details of the proof only in the case $P_0\in \Gamma_2$ since the case $P_0\in \Gamma_3$ can be handled in a similar way. In this case $d_n=y_n{\rightarrow}0$ ($d_n$ as in ) and $x_n{\rightarrow}x_0$ as $n{\rightarrow}\infty$ with $0<x_0<1$, hence a point $(\widetilde x,\widetilde y)\in B_R$ satisfies , and for $n$ large enough, and so it belongs to $\Omega_n$ if and only if holds, namely when $$\widetilde y>-y_nM_n^{\frac{p_n-1}2}.$$ Two possibilities may hold: either $y_nM_n^{\frac{p_n-1}2}{\rightarrow}\infty$ or $y_nM_n^{\frac {p_n-1}2}{\rightarrow}s\geq 0$. \ In the first case it follows that any ball $B_R\subset \Omega_n$ for $n$ large enough, namely $\Omega_n{\rightarrow}\Sigma={\mathbb{R}}^2$ and so, as in [Step 1]{}, $\widetilde{u}_n{\rightarrow}\widetilde u$ uniformly on compact sets of $\Sigma$, where $\widetilde u$ is a nontrivial solution to in ${\mathbb{R}}^2$ that satisfies $\widetilde u(0)=1$ and that has finite Morse index, getting a contradiction. \ In this case instead $\Omega_n{\rightarrow}\Sigma:=\{(x,y)\in {\mathbb{R}}^2 : \ y> -s\} $ for some $s\geq 0$ and $ \widetilde{u}_n{\rightarrow}\widetilde u$ on compact sets of $\Sigma$ where $\widetilde u$ is a solution to in $\Sigma:=\{(x,y)\in {\mathbb{R}}^2 : \ y> -s\}$ that satisfies a Neumann boundary condition on $\partial \Sigma$.\ When $s>0$, $0\in \Omega_n$ for $n$ large enough, hence $\widetilde{u}$ is nontrivial since $\widetilde{u}(0)=1$ by the uniform convergence on compact sets. Finally by translating this limit nontrivial solution in the $y$-direction we then end-up, when $s>0$, with a nontrivial solution $\widetilde u$ to in $\Sigma=\mathbb R^2_+$ with Neumann boundary conditions on $\partial \Sigma$.\ Next we treat the case $s=0$ and show that again the limit solution $\widetilde u$ is non-trivial. Observe that $\widetilde y=-M_n^{\frac{p_n-1}2}y_n\in\partial\Omega_n$ and that in the case $s=0$ it belongs to a neighborhood of $0$ for $n$ large. By the elliptic regularity up to the boundary (see Lemma 6.18 in [@GT]) for the equation $-\Delta \widetilde u_n=f_n$ with $f_n=\|\widetilde u_n\|^{p_n-1}\widetilde u_n$, we obtain a uniform bound on the gradient of $\widetilde u_n$ in $\overline \Omega_n\cap B_{\rho}$, for $\rho$ sufficiently small (indeed by definition $\|\widetilde u_n\|\leq 1$ on $\partial \Omega_n$, hence $\|f_n(x)\|\leq 1$ and we use the fact that $u_n\in C^{2,\gamma}(\Gamma_2)$). This implies that $$\widetilde u_n(F)\geq \widetilde u_n(0)-C\|F-0\|=1-C\|F\|, \quad \forall F\in \Omega_n\cap B_{\rho}$$ where $C$ is the uniform bound on the gradient. Choosing $F$ in the set $\Sigma= \{(x,y)\in{\mathbb{R}}^2 : \ y> 0\}$ and sufficiently close to $0$ and passing to the limit in the previous inequality one then has $\widetilde u(F)>0$, namely $\widetilde u$ is non-trivial.\ Summarizing, for any $s\geq 0$, we have obtained a non-trivial solution $\widetilde u$ to in $\Sigma:=\mathbb R^2_+$ that satisfies a Neumann boundary condition on $\partial \Sigma$. Moreover, as a consequence of Lemma \[lem-10-nuovo\], similarly as in [Step 1]{}, one can easily prove that the maximal number of linearly independent functions $\widetilde\psi_i$ in the space $C^{\infty}_0(\overline{{\mathbb{R}}^2_+})\cap \{\frac{\partial \widetilde\psi_i}{\partial y}{\big\|_{y=0}}=0\}$ that make negative the quadratic form $Q$ is at most $2$. As a consequence, the even extension of $\widetilde u$ to the whole ${\mathbb{R}}^2$ is a nontrivial solution to in $\Sigma={\mathbb{R}}^2$ which has finite $\mathcal G$-Morse index, where $\mathcal{G}$ here is the group generated by the reflection with respect to the $x$-axis. Again this is not possible by Proposition \[prop7.3\]. .\ Since we are assuming that $M_n^{p_n-1}{\rightarrow}\infty$ and $(x_n,y_n){\rightarrow}(1,0)$ it is straightforward to see that a point $(\widetilde x, \widetilde y)\in B_R$ satisfies , and the first inequality in for large values of $n$ and so it belongs to $\Omega_n$ for large $n$ if and only if and the second inequality in are satisfied, namely: $$\label{10-5} \widetilde y>- y_nM_n^{\frac {p_n-1}2}$$ $$\label{10-6} M_n^{\frac {1-p_n}2}\left(\widetilde x^2+ \widetilde y^2\right)+2\left(\widetilde xx_n+\widetilde yy_n\right)<\left(1-x_n^2-y_n^2\right)M_n^{\frac {p_n-1}2}$$ Hence we have to to distinguish several possibilities: $$\begin{aligned} & \mbox{ either } & y_nM_n^{\frac {p_n-1}2}{\rightarrow}\infty\label{10-1}\\ & \mbox{ or } & y_nM_n^{\frac {p_n-1}2}{\rightarrow}\alpha\geq 0\label{10-2}\end{aligned}$$ as $n{\rightarrow}\infty$ and also $$\begin{aligned} & \mbox{ either } & \left(1-x_n^2-y_n^2\right)M_n^{\frac {p_n-1}2}{\rightarrow}\infty\label{10-3}\\ & \mbox{ or } & \left(1-x_n^2-y_n^2\right)M_n^{\frac {p_n-1}2}{\rightarrow}\beta> 0\label{10-4}\end{aligned}$$ as $n{\rightarrow}\infty$, where the case $\beta=0$ is ruled-out by the Dirichlet boundary conditions on $\Gamma_1$ (as in [Step 2]{}).\ Observe that implies for large $n$, while when holds then is satisfied for $n$ large if and only if $\widetilde y>-\alpha$. Similarly if holds then is satisfied when $n$ is large, while if holds then is satisfied for $n$ large if and only if $\widetilde x<\frac {\beta}2 $.\ Summarizing we have that $ \widetilde{u}_n{\rightarrow}\widetilde u$ uniformly on compact sets of $\Sigma$, where $\widetilde u$ is a solution to in $\Sigma$, more precisely: \ In this case $\Sigma ={\mathbb{R}}^2$, $\widetilde u$ is nontrivial (since $\widetilde u(0)=1$) and moreover, as in [Step 1]{} one can prove that $\widetilde u$ has finite Morse index contradicting Proposition \[prop7.3\]. \ In this case $\Sigma=\{(x,y)\in {\mathbb{R}}^2: \ x<\frac {\beta}2\}$, $\widetilde u$ is nontrivial (again $0\in \Omega_n$ when $n$ is large enough and then $\widetilde u(0)=1$), it satisfies Dirichlet boundary conditions on the hyperplane $x=\frac {\beta}2$ and has finite Morse index. This (up to a translation) contradicts again Proposition \[prop7.3\]. \ Now $\Sigma=\{(x,y)\in {\mathbb{R}}^2: \ y>-\alpha\}$, $\widetilde u$ satisfies Neumann boundary conditions on the hyperplane $y=-\alpha$. If $\alpha>0$ then, as before, $\widetilde u(0)=1$ and so it is nontrivial. In this case we translate this solution in the $y$-direction getting a solution to in ${\mathbb{R}}^2_+$ that satisfies Neumann boundary conditions and we obtain a contradiction as in [Step 3]{}-[Case 2]{}. In the case $\alpha=0$ we observe that $d_n=y_n$ (where $d_n$ as usual is the distance in ). Indeed $P_0=B$ implies that $d_n= \min\{dist (P_n,\Gamma_2), dist (P_n,\Gamma_1)\}$, where $dist (P_n,\Gamma_2)=y_n$ and $dist (P_n,\Gamma_1)= 1-\sqrt{x_n^2+y_n^2}$, moreover $1-\sqrt{x_n^2+y_n^2}\geq y_n$ if and only if $$\label{intermedia} y_n(2-y_n)\leq 1-x_n^2-y_n^2,$$ and holds for $n$ large, under the assumptions with $\alpha=0$ and . Since $d_n=y_n$, then $\widetilde y=-M_n^{\frac{p_n-1}2}y_n\in\partial\Omega_n$ and moreover it belongs to a neighborhood of $0$ for $n$ large, hence we can reason as in [Step 3]{}-[Case 2]{} and use the elliptic regularity up to the boundary to obtain a uniform estimate on the gradient of $\widetilde u_n$ in a neighborhood of $0$, showing that $\widetilde u$ is nontrivial. Again we obtain a contradiction as at the end of [Step 3]{}-[Case 2]{}. \ Now $\Sigma=\{(x,y)\in {\mathbb{R}}^2: \ y>-\alpha, \ x<\frac {\beta}2\}$, $\widetilde u$ satisfies Dirichlet boundary conditions on the hyperplane $x=\frac {\beta}2$ and Neumann boundary conditions on the hyperplane $y=-\alpha$. As before when $\alpha>0$ we have that $0\in \Omega_n$ when $n$ is large enough and then $\widetilde u(0)=1$, namely $\widetilde u$ is nontrivial and so we translate it ending with a nontrivial solution $\bar u$ to in $\bar{\Sigma}=\{(x,y)\in {\mathbb{R}}^2: \ y>0, \ x<0\}$, with Dirichlet boundary conditions on $x=0$ and Neumann boundary conditions on $y=0$. When $\alpha=0$ one proves as in the previous case, so again $d_n=y_n$ for large $n$. Then $\widetilde y=- M_n^{\frac {p_n-1}2}y_n\in\partial\Omega_n$ and it belongs to a neighborhood of $0$ for large $n$, so we can prove that $\widetilde u$ is nontrivial using again the elliptic regularity up to the boundary as in the previous situation. Also in this case we translate $\widetilde u$ ending with a nontrivial solution $\bar u$ to in $\bar{\Sigma}=\{(x,y)\in {\mathbb{R}}^2: \ y>0, \ x<0\}$, with Dirichlet boundary conditions on $x=0$ and Neumann boundary conditions on $y=0$.\ Finally observe that as a consequence of Lemma \[lem-10-nuovo\], using arguments similar to the ones in [[Step 1]{}]{}, one can prove that the maximal number of linearly independent functions $\widetilde\psi_i\in C^{\infty}_0(\{(x,y)\in {\mathbb{R}}^2: \ y\geq 0, \ x<0\})\cap \{\frac{\partial \widetilde\psi_i}{\partial y}{\big\|_{y=0}}=0\}$ that make negative the quadratic form $Q$ is at most $2$. Thus, by extending $\bar u$ to $\widetilde \Sigma :=\{(x,y)\in {\mathbb{R}}^2: \ x<0\}$ in an even way, we obtain a solution to in $\widetilde \Sigma$ which has finite $\mathcal{G}$-Morse index, where $\mathcal{G}$ here is the group generated by the reflection with respect to the $x$-axis. This is again in contradiction with Proposition \[prop7.3\]. [Step 5.]{} [$P_0=O$]{}\ In this case we can assume w.l.o.g. that $d_n=y_n$, since $P_0=O$ implies that $d_n=\min \{dist(P_n,\Gamma_2),dist(P_n,\Gamma_3)\}$, $dist(P_n,\Gamma_2)=y_n$ and w.l.o.g (up to rotation) we may consider only the case $dist(P_n,\Gamma_2)\leq dist(P_n,\Gamma_3)$. We may also assume that $y_n\leq x_n$ and $\frac{y_n}{x_n}\leq \tan \frac{\pi}{2k}$ (if $x_n\neq 0$). Then a point $(\widetilde x, \widetilde y)\in B_R(0)$ for some $R>0$ belongs to $\Omega_n$ if and only if conditions and are satisfied. Indeed is easily verified. We have to distinguish different cases, since $$\begin{aligned} &\mbox{either } & y_nM_n^{\frac {p_n-1}2}{\rightarrow}\infty\label{10-7}\\ &\mbox{or } &y_nM_n^{\frac {p_n-1}2}{\rightarrow}\alpha\geq 0\label{10-8}\end{aligned}$$ and $$\begin{aligned} & \mbox{either } & x_nM_n^{\frac {p_n-1}2}{\rightarrow}\infty\label{10-9}\\ &\mbox{or } &x_nM_n^{\frac {p_n-1}2}{\rightarrow}\beta\geq 0,\label{10-10}\end{aligned}$$ where it is obvious that implies and that implies with $\alpha\leq\beta$ (since $y_n\leq x_n$). \ In this case also holds and $d_nM_n^{\frac {p_n-1}2}{\rightarrow}\infty$, hence and are satisfied for large $n$ and so $\Omega_n{\rightarrow}{\mathbb{R}}^2$. Then $ \widetilde{u}_n{\rightarrow}\widetilde u$ uniformly on compact sets of ${\mathbb{R}}^2$ where $\widetilde u$ is a nontrivial (since $\widetilde u(0)=1$) solution to in ${\mathbb{R}}^2$ of finite Morse index, giving a contradiction to the results of Proposition \[prop7.3\]. \ is satisfied for large $n$ while is satisfied for large $n$ if and only if $\widetilde y>-\alpha$. Hence the limit domain is $\Sigma=\{(x,y)\in {\mathbb{R}}^2: \ y>-\alpha\}$ and $\widetilde{u}_n{\rightarrow}\widetilde u$ uniformly on compact sets of $\Sigma$ where $\widetilde u$ is a solution to in $\Sigma$ that satisfies a Neumann boundary condition on $y=-\alpha$ of finite Morse index, in the sense of [Step 3]{}. Moreover when $\alpha>0$ then $0\in \Omega_n$ and this implies that $\widetilde u$ is nontrivial getting a contradiction. When $\alpha=0$ we observe that $\widetilde y=- M_n^{\frac {p_n-1}2}y_n\in\partial\Omega_n$ and it belongs to a neighborhood of $0$. We can therefore apply the elliptic regularity up to the boundary as in [Step 3]{} getting that $\widetilde u$ is nontrivial. Thus a contradiction arises as in the previous case. \ In this case also condition holds with $0\leq\alpha\leq\beta$, which implies that is satisfied for large $n$ if and only if $\widetilde y>-\alpha$. Moreover by and it follows that $\widetilde x>- \beta$. Condition is satisfied for large $n$, instead, if and only if $$\frac{\widetilde y+ \alpha}{\widetilde x+ \beta}< \tan \frac{\pi}k.$$ Then the limiting domain $\Sigma$ is a positive cone in ${\mathbb{R}}^2$ with vertex in $(-\beta, -\alpha)$ and with amplitude $\frac{\pi}k$ (the same of $S_k$) $$\Sigma=\left\{ (r\cos\theta -\beta,r\sin\theta-\alpha)\ : \ r\in (0,+\infty), \ \theta \in [0,\frac{\pi}{k}] \right\}$$ Then $\widetilde{u}_n{\rightarrow}\widetilde u$ uniformly on compact sets of $\Sigma$ where $\widetilde u$ is a solution to in $\Sigma$ that satisfies a Neumann boundary condition on $\partial \Sigma$. When $\alpha,\beta\neq 0$ then $0\in \Sigma$ and we can infer that $\widetilde u$ is nontrivial. The same is true when $\alpha=0$, since $\beta>0$ and in this case we have that $\widetilde y=- M_n^{\frac {p_n-1}2}y_n\in\partial\Omega_n$ and belongs to a neighborhood of $0$, so we can reason as in [Step 3]{} the and show that $\widetilde u$ is nontrivial. Moreover in both the cases $\widetilde u$ has finite Morse index, since the maximal number of linearly independent functions $\widetilde \psi_i$ in $C^{\infty}_0(\overline \Sigma)\cap\{\frac{\partial\widetilde \psi_i}{\partial \nu}\|_{\partial \Sigma}=0\}$ ($\nu$ denotes the outer normal to $\partial \Sigma$) that make negative the quadratic form $Q$ is at most two due to Lemma \[lem-10-nuovo\]. Translating $\widetilde u$ with respect to one or both the axes we end-up with a function $\bar u$ that satisfies in $\{(x,y)\in {\mathbb{R}}^2: x>0, y>0, \frac yx<\tan \frac \pi k\}$ and Neumann boundary conditions. Finally the $\mathcal{G}_k$ extension of $ \bar u$ to the whole ${\mathbb{R}}^2$ (which is well defined due to the Neumann boundary conditions) is a non trivial $k$-symmetric solution to in $\mathbb R^2$ which has $k$-Morse index at most $2$. This contradicts the result in Proposition \[prop7.3\]. In this case also condition holds with $\alpha=0$. We consider the solution $u_n$ in the whole ball $B$ (without restricting it to the sector $S_k$) and we define $$\widetilde{v}_n(x,y):=\frac 1{M_n}{u_n(M_n^{\frac {1-p_n}2}(x,y))}$$ that satisfies $$-\Delta \widetilde{v}_n=\|\widetilde{v}_n\|^{p_n-1}\widetilde{v}_n$$ in $\widetilde B_n:=M_n^{\frac {p_n-1}2} B$ and also $\|\widetilde{v}_n\|\leq 1$. The rescaled domain $\widetilde B_n{\rightarrow}{\mathbb{R}}^2$ and $\widetilde{v}_n{\rightarrow}\widetilde v$ uniformly on compact sets of ${\mathbb{R}}^2$ where $\widetilde v$ is a solution to which has $k$-Morse index at most $2$ (observe that since we are rescaling with respect to the origin the symmetries are preserved). To obtain a contradiction via Proposition \[prop7.3\] we need to show that $\widetilde v$ is nontrivial. This easily follows since $\widetilde{v}_n(\widetilde P_n)=1$, where $\widetilde P_n=(M_n^{\frac {p_n-1}2}x_n, M_n^{\frac {p_n-1}2}y_n)$ and by assumption $\widetilde P_n{\rightarrow}0$, so that $\widetilde v(0)=1$. This end the proof. Now we are in the position to consider the asymptotic behavior of the nodal least energy solutions $u_p^k$ as $p{\rightarrow}1$ and to conclude the proof of the radial part of Theorem \[prop1.4\]. \[leastRadiale\] The least energy nodal solutions $u_p^k$ are radial for any $k\geq 3$ when $p$ is close to $1$. [Step 1.]{} [We show that for any sequence $p_n>1$ converging to $1$ $$\label{u-nkappaconv} \bar{u}_n^k:=\frac {u_{p_n}^k}{{\lVertu_{p_n}^k\rVert}_{\infty}}\rightarrow C\varphi_{2,{{\text{\upshape rad}}}} =J_0(\nu_{02}\|x\|)\quad \text{ in }C(\bar B)$$ up to a subsequence, where $C=\pm1$ and $$\label{8.14} {\lVertu_{p_n}^k\rVert}_{\infty}^{p_n-1}= \lambda_{2,{{\text{\upshape rad}}}}\left(1-\widetilde c (p_n-1)\right)+o(p_n-1) \text{ as }n{\rightarrow}\infty$$ where $\widetilde c$ is as in . ]{} Let $M_n:={\lVertu_{p_n}^k\rVert}_\infty$, we have shown in Proposition \[prop7.4\] that $M_n^{p_n-1}$ is bounded, we can then repeat the proof of Lemma \[lemma-pvicino1\] proving that $$M_n^{p_n-1}{\rightarrow}\lambda \ \ \text{and} \ \ \bar u_n^k{\rightarrow}C\varphi \text{ in } C(\bar B) \, \text{ up to a subsequence, with }C=\pm 1$$ where $\lambda$ is an eigenvalue of $-\Delta$ in $B$ with Dirichlet boundary conditions, $\varphi$ is a corresponding eigenfunction with $\\|\varphi\\|_{\infty}=1$. Moreover $\varphi$ is invariant by the action of $\mathcal{G}_k$ (since $\bar u_{n}^k$ are for every $n$) and, following the ideas in Step 1 in the proof of Proposition \[risultatoMorse\_pvicino1\] we can show that $m_k(\varphi)\leq m_k(u_{p_n}^k)$, hence $m_k(\varphi)\leq 2$ by Lemma \[lem-10-1\]. Since the $k$-symmetric eigenvalues of $-\Delta$ are known and since we are assuming $k\geq 3$, this means that necessarily either $\lambda=\lambda_{1,{{\text{\upshape rad}}}}$ or $\lambda=\lambda_{2,{{\text{\upshape rad}}}}$. We show that the case $\lambda=\lambda_{1,{{\text{\upshape rad}}}}$ cannot hold. Indeed, following similar ideas as in Step 2 of the proof of Proposition \[risultatoMorse\_pvicino1\], since $\varphi_{1,{{\text{\upshape rad}}}}$ has Morse index $0$, one gets that the $2$ negative $k$-symmetric eigenvalues of the linearized operator at $u_{p_n}^k$ (recall $m_k(u_{p_n}^k)=2$ by Lemma \[k-morse-upk\]) converge both to $0$ and that the corresponding eigenfunctions (that we can take to be orthogonal in $L^2(B)$) converge to two orthogonal solutions of $$\left\{ \begin{array}{lr} -\Delta v=\lambda_1 v & \text{ in }B\\ v=0 & \mbox{ on }\partial B. \end{array} \right.$$ This is not possible, since $\lambda_1$ is simple, so $\lambda=\lambda_{2,{{\text{\upshape rad}}}}$. Reasoning exactly as in the proof of Lemma \[lemma-pvicino1\], we can then prove . Assuming w.l.o.g. that $\bar u_n^k(0)>0$ for $n$ large, we also have $$\bar u_n^k{\rightarrow}\varphi_{2,{{\text{\upshape rad}}}}=J_0(\nu_{02}\|x\|) \text{ as }n{\rightarrow}\infty \ \text{ in } C(\bar B),$$ getting . Assume by contradiction that there exists a sequence $p_n>1$, $p_n\rightarrow 1$ as $n\rightarrow +\infty$ such that $u_n^k\neq u_n$, where $u_n^k:=u_{p_n}^k$ and $u_n:=u_{p_n}$, and define $w_n:=\frac{u_n^k- u_n}{{\lVertu_{n}^k- u_n\rVert}_{\infty}}$. $w_n$ satisfies $$\label{8.12} \left\{\begin{array}{lr} -\Delta w_n= p_nc_n(x)w_n \qquad\ \mbox{ in }B\\ w_n =0\qquad\qquad\qquad\qquad\mbox{ on }\partial B\\ {\lVertw_n\rVert}_{\infty}=1 \end{array}\right.$$ where, by the Mean value Theorem, $$\label{cn} c_n(x)=\int_0^1 \|tu_n^k+(1-t) u_n\|^{p_n-1}\, dt\leq {\lVertu_n^k\rVert}_{\infty}^{p_n-1}+{\lVertu_n\rVert}_{\infty}^{p_n-1}\leq \overset{{\eqref{8.14}-\eqref{gammanp-1}}}{C}\lambda_{2,{{\text{\upshape rad}}}}.$$ We show that $$\label{infinitob} c_n(x){\rightarrow}\lambda_{2,{{\text{\upshape rad}}}}\ \text{ almost everywhere in } B\ \mbox{ as } \ n\rightarrow \infty.$$ Indeed from and we have that $$\begin{aligned} \frac{u_n}{\lambda_{2,{{\text{\upshape rad}}}}^{\frac 1{p_n-1}}}&=&\frac{u_n}{{\lVertu_n\rVert}_{\infty}}\left(\frac{{\lVertu_n\rVert}_{\infty}^{p_n-1}}{\lambda_{2,{{\text{\upshape rad}}}}}\right)^{\frac 1{p_n-1}}=\bar{u}_n\left(1-\widetilde c(p_n-1)+o(p_n-1)\right)^{\frac 1{p_n-1}}\\ &=&\varphi_{2,{{\text{\upshape rad}}}}e^{-\widetilde c}(1+o(1))\end{aligned}$$ as $n{\rightarrow}\infty$, where $\widetilde c$ is as in , and the same holds for $u_n^k$ using and . Namely $$\frac{u_n}{e^{-\widetilde c}\lambda_{2,{{\text{\upshape rad}}}}^{\frac 1{p_n-1}}}{\rightarrow}\varphi_{2,{{\text{\upshape rad}}}}\ \text{ and }\ \frac{u_n^k}{e^{-\widetilde c}\lambda_{2,{{\text{\upshape rad}}}}^{\frac 1{p_n-1}}}{\rightarrow}\varphi_{2,{{\text{\upshape rad}}}}\quad \text{ in $C(\bar B)$ as }n{\rightarrow}\infty.$$ As a consequence, for any $x\in B$ we have $$\label{cuoreb} t \frac{u_n^k}{e^{-\widetilde c}\lambda_{2,{{\text{\upshape rad}}}}^{\frac 1{p_n-1}}} +(1-t) \frac{u_n}{e^{-\widetilde c}\lambda_{2,{{\text{\upshape rad}}}}^{\frac 1{p_n-1}}} {\rightarrow}\varphi_{2,{{\text{\upshape rad}}}}$$ and follows then from observing that $$\label{8.20}\begin{split} \frac {c_n(x)}{\lambda_{2,{{\text{\upshape rad}}}}}& =\int_0^1 \Big\|t \frac{u_n^k}{\lambda_{2,{{\text{\upshape rad}}}}^{\frac 1{p_n-1}}} +(1-t) \frac{u_n}{\lambda_{2,{{\text{\upshape rad}}}}^{\frac 1{p_n-1}}} \Big\|^{p_n-1}\, dt=\\ &=e^{-\widetilde c(p_n-1)}\int_0^1 \Big\|t \frac{u_n^k}{e^{-\widetilde c}\lambda_{2,{{\text{\upshape rad}}}}^{\frac 1{p_n-1}}} +(1-t) \frac{u_n}{e^{-\widetilde c}\lambda_{2,{{\text{\upshape rad}}}}^{\frac 1{p_n-1}}} \Big\|^{p_n-1}\, dt. \end{split}.$$ Passing to the limit in and using get that $w_n$ converges, up to a subsequence, in $C(\bar B)$ to a function $w$ which solves $$\label{8.13} \left\{\begin{array}{lr} -\Delta w= \lambda_{2,{{\text{\upshape rad}}}}w \qquad \mbox{ in }B\\ w =0\qquad\qquad\qquad\mbox{ on }\partial B\\ {\lVertw\rVert}_{\infty}=1 \end{array}\right.$$ so that $$\label{defC} w=C\varphi_{2,{{\text{\upshape rad}}}}, \quad\mbox{ with }\ C=\pm1\ \mbox{ depending on the sign of }w(0).$$ On the other side, multiplying by $\varphi_{2,{{\text{\upshape rad}}}}$ and integrating over $B$ we find $$\label{8.18} \begin{split} &\lambda_{2,{{\text{\upshape rad}}}} \int_B w_n \varphi_{2,{{\text{\upshape rad}}}}=\int_B \nabla w_n\nabla \varphi_{2,{{\text{\upshape rad}}}}=\lambda_{2,{{\text{\upshape rad}}}} p_n\int_B \frac{c_n(x)}{\lambda_{2,{{\text{\upshape rad}}}}} w_n\varphi_{2,{{\text{\upshape rad}}}}\\ &=\lambda_{2,{{\text{\upshape rad}}}}\int_B \frac{c_n(x)}{\lambda_{2,{{\text{\upshape rad}}}}} w_n\varphi_{2,{{\text{\upshape rad}}}}+\lambda_{2,{{\text{\upshape rad}}}}(p_n-1)\int_B \frac{c_n(x)}{\lambda_{2,{{\text{\upshape rad}}}}} w_n\varphi_{2,{{\text{\upshape rad}}}}. \end{split}$$ Using the trivial equality $e^x-1=x\int_0^1e^{sx}ds $ and , we write $$\begin{aligned} \frac{c_n(x)}{\lambda_{2,{{\text{\upshape rad}}}}}&=&\int_0^1 1+(p_n-1)\log \Big\| t\frac{u_n^k}{\lambda_{2,{{\text{\upshape rad}}}}^{\frac 1{p_n-1}}} +(1-t) \frac{u_n} {\lambda_{2,{{\text{\upshape rad}}}}^{\frac 1{p_n-1}}}\Big\| \int_0^1 \Big\| t\frac{u_n^k}{\lambda_{2,{{\text{\upshape rad}}}}^{\frac 1{p_n-1}}} +(1-t) \frac{u_n} {\lambda_{2,{{\text{\upshape rad}}}}^{\frac 1{p_n-1}}}\Big\|^{s(p_n-1)}\ ds\, dt\\ &=&1+(p_n-1)g_n(x),\end{aligned}$$ where $$g_n(x):=\int_0^1\log \Big\| t\frac{u_n^k}{\lambda_{2,{{\text{\upshape rad}}}}^{\frac 1{p_n-1}}} +(1-t) \frac{u_n} {\lambda_{2,{{\text{\upshape rad}}}}^{\frac 1{p_n-1}}}\Big\|\int_0^1 \Big\| t\frac{u_n^k}{\lambda_{2,{{\text{\upshape rad}}}}^{\frac 1{p_n-1}}} +(1-t) \frac{u_n} {\lambda_{2,{{\text{\upshape rad}}}}^{\frac 1{p_n-1}}}\Big\|^{s(p_n-1)}\ ds\, dt.$$ Equation then becomes $$\lambda_{2,{{\text{\upshape rad}}}} \int_B w_n \varphi_{2,{{\text{\upshape rad}}}}= \lambda_{2,{{\text{\upshape rad}}}}\int_B \left( 1+(p_n-1)g_n(x) \right) w_n\varphi_{2,{{\text{\upshape rad}}}}+\lambda_{2,{{\text{\upshape rad}}}}(p_n-1)\int_B \frac{c_n(x)}{\lambda_{2,{{\text{\upshape rad}}}}} w_n\varphi_{2,{{\text{\upshape rad}}}}.$$ so that, dividing by $\lambda_{2,{{\text{\upshape rad}}}}(p_n-1)$ we obtain $$\label{8.18bis} 0=\int_B g_n(x)w_n\varphi_{2,{{\text{\upshape rad}}}}+\int_B \frac{c_n(x)}{\lambda_{2,{{\text{\upshape rad}}}}} w_n\varphi_{2,{{\text{\upshape rad}}}}.$$ Observe now that, by , for any $x\in B$ such that $\varphi_{2,{{\text{\upshape rad}}}}\neq 0$ we have that $$\label{quadratinob} g_n(x){\rightarrow}\log \big\|\varphi_{2,{{\text{\upshape rad}}}}e^{-\widetilde c}\big\|=\log\big\|\varphi_{2,{{\text{\upshape rad}}}}\big\| -\widetilde c\ \text{ as }n{\rightarrow}\infty.$$ This implies that $g_n(x)\varphi_{2,{{\text{\upshape rad}}}}\in L^{\infty}(B)$ and $${\lVertg_n(x)\varphi_{2,{{\text{\upshape rad}}}}\rVert}_{\infty}\leq C.$$ We can then pass to the limit as $n{\rightarrow}\infty$ into and using and we get $$0=C\int_B\left( \log\big\|\varphi_{2,{{\text{\upshape rad}}}}\big\| -\widetilde c\right) \varphi_{2,{{\text{\upshape rad}}}}^2+C\int_B \varphi_{2,{{\text{\upshape rad}}}}^2$$ which implies, using the definition of $\widetilde{c}$ in , that $$0=C\int_B \varphi_{2,{{\text{\upshape rad}}}}^2$$ namely that $C=0$, contradicting the definition of $C$ in and ending the proof. \[RemarkNoQuaiRadk=2\] One could prove, reasoning as in the proof of Proposition \[leastRadiale\], that $\bar u_p^2\rightarrow \varphi$ in $C^1(\bar B)$ as $p\rightarrow 1$, where $\varphi$ is an eigenfunction of $-\Delta$ corresponding to the eigenvalue $\lambda_4=\lambda_5$, which is not quasi-radial. The convergence in $C^1(\bar B)$, by the Hopf lemma then implies that $u_p^2$ is not quasi-radial for $p$ close to $1$. [99]{} H. Amann, Fixed point equations and nonlinear eigenvalue problems in ordered Banach spaces. SIAM Rev. 18(4) (1976), 620-709. A. Aftalion, F. Pacella, Qualitative properties of nodal solutions of semilinear elliptic equations in radially symmetric domains, C. R. Acad. Sci. 339 (2004), 339-344. A. Amadori, F. Gladiali, Nonradial sign changing solutions to Lane-Emden problem in an annulus, Nonlinear Analysis, Theory, Methods and Applications 155(1) (2017), 294-305. A. Amadori, F. Gladiali, in progress. A. Amadori, F. Gladiali, M. Grossi, Nodal solutions for Lane-Emden problems in almost-annular domains, to appear in Differential and Integral Equations. A. Ambrosetti, A. Malchiodi, Nonlinear analysis and semilinear elliptic problems, Cambridge Studies in Advanced Mathematics, 104, Cambridge University Press, Cambridge, 2007. T. Bartsch, T. Weth, A note on additional properties of sign changing solutions to superlinear elliptic equations, Topological Methods in Nonlinear Analysis 22 (2003), 1-14. Bartsch, T., Weth, T. and Willem M., Partial symmetry of least energy nodal solutions to some variational problems, J. Anal. Math. 96 (2005), 1-18. A. Castro, J. Cossio, J.M. Neuberger, A sign-changing solution for a superlinear Dirichlet problem, Rocky Mountain J.Math. 27 (4) (1997), 1041-1053. E.N. Dancer, On the indices of fixed points of mappings in cones and applications, Journal of Math. Anal. and Appl. 91 (1983), 131-151. E.N. Dancer, Global breaking of symmetry of positive solutions on two-dimensional annuli, Differential Integral Equations 5 (1992), 903-913. E.N. Dancer, Real analyticity and non-degeneracy, Math. Ann. 325 (2003), 369-392. E.N. Dancer, J. Wei, Sign-changing solutions for supercritical elliptic problems in domains with small holes, Manuscripta Math. 123 (2007), 493-511. F. De Marchis, M. Grossi, I. Ianni and F. Pacella, Uniqueness for positive solutions of Lane-Emden problem in convex domain, in progress. F. De Marchis, I. Ianni and F. Pacella, Sign changing solutions of Lane Emden problems with interior nodal line and semilinear heat equations, Journal of Differential Equations 254 (2013), 3596-3614. F. De Marchis, I. Ianni and F. Pacella, Asymptotic analysis and sign changing bubble towers for Lane-Emden problems, Journal of the European Mathematical Society 17(8) (2015), 2037-2068. F. De Marchis, I. Ianni, F. Pacella, Morse index computation for nodal radial solutions of Lane-Emden problems, Mathematische Annalen, 367(1) (2017), 185-227. F. De Marchis, I. Ianni, F. Pacella, A Morse index formula for radial solutions of Lane-Emden problems, preprint. B. Gidas, J. Spruck, A priori bounds for positive solutions of nonlinear elliptic equations, Comm. in Part. Diff. Eq. 6 (1981), 883-901. D. Gilbarg, N.S. Trudinger, Elliptic partial differential equations of second order. Reprint of the 1998 edition. Classics in Mathematics. Springer-Verlag, Berlin, 2001. F. Gladiali, A global bifurcation result for a semilinear elliptic equation, Journal of Math. Anal. and Appl. 369(1) (2010), 306-311. F. Gladiali, Separation of branches of $O(N-1)$-invariant solutions for a semilinear elliptic equation, Journal of Math. Anal. and Appl. 453 (2017), 159-173. F. Gladiali, M. Grossi, S.L.N. Neves, Symmetry breaking and Morse index of solutions of nonlinear elliptic problems in the plane, Commun. Contemp. Math. 18 (2016). F. Gladiali, M. Grossi, F. Pacella, P.N. Srikanth, Bifurcation and symmetry breaking for a class of semilinear elliptic equations in an annulus, Calc. Var. 40 (2011), 295-317. F. Gladiali, I. Ianni, F. Pacella, in progress. M. Grossi, C. Grumiau, F. Pacella, Lane Emden problems with large exponents and singular Liouville equations, J. Math. Pures Appl. 101(9) (2014), 735-754. A. Harrabi, S. Rebhi, A. Selmi, Existence of radial solutions with prescribed number of zeros for elliptic equations and their Morse index, J. Differential Equations 251 (2011), 2409-2430. R. Kajikiya, Sobolev norms of radially symmetric oscillatory solutions for superlinear elliptic equations, Hiroshima Math. J. 20 (1990), 259-276. T. Kato, Perturbation Theory for Linear Operators, Springer-Verlag, Berlin, 1976. W.M. Ni, R. D. Nussbaum, Uniqueness and nonuniqueness for positive radial solutions of $\Delta u+f(u,r)=0$, Comm. Pure Appl. Math. 38 (1985), 67-108. F. Pacella, T. Weth, Symmetry of solutions to semilinear elliptic equations via Morse index, Proc. American Math. Soc. 135(6) (2007), 1753-1762. X. Ren, J. Wei, On a two dimensional elliptic problem with large exponent in nonlinearity, Trans. Amer. Math. Soc. 343 (1994), 749-763. G. N. Watson, A Treatise on the Theory of Bessel Functions, 2nd ed., Cambridge Math. Lib., Cambridge University Press, Cambridge, UK, 1995. J. A. Smoller, A. G. Wasserman, Existence, uniqueness and non-degeneracy of positive solutions of semi-linear elliptic equations, Comm. Math. Phys. 95 (1984), 129-159. [^1]: 2010 Mathematics Subject classification: [^2]: Keywords: nodal solutions, non-radial solutions, bifurcation, Morse index, least energy, symmetry, blow-up [^3]: Research partially supported by: PRIN $201274$FYK7$\_005$ grant and INDAM - GNAMPA	{ "pile_set_name": "ArXiv" }
=1200 =cmcsc10 at 12pt ‘=11 =msam10 =msam7 =msam5 =msbm10 =msbm7 =msbm5 == === =@\#1[\#1 0123456789 ABCDEF]{} =eufm10 =eufm7 =eufm5 === @\#1[[@@[\#1]{}]{}]{} @@\#1[\#1]{} @[@]{} @[@]{} =“0@03 =”3@28 =“2@6E =”2@6F @\#1[[@@[\#1]{}]{}]{} @@\#1[\#1]{} ‘=12 \#1 \#1 \#1[ ]{} \#1[ ]{} 6.5pc =cmbx10 scaled2 2.0pc An Asymptotically Good Tower of Curves Over the Field with Eight Elements =cmcsc10 at 11pt Gerard van der Geer and Marcel van der Vlugt 2.0pc Introduction In this note we construct an explicit asymptotically good tower of curves over the field $\F_8$. For a curve $C$ defined over a finite field $\F_q$ of cardinality $q$ we denote by $\#C(\F_q)$ the number of $\F_q$-rational points on $C$. If furthermore we denote, as usual, by $N_q(g)$ the function $$N_q(g)= \max \{ \# C(\F_q): C/F_q, \, g(C)=g \},$$ where $C$ runs through the set of smooth absolutely irreducible projective curves of genus $g$ defined over $\F_q$ then Drinfeld and Vladuts showed in \[D-V\] the inequality $$\lim\sup_{g \to \infty} {N_q(g) \over g} \leq \sqrt{q}-1, \eqno(1)$$ and Ihara constructed in \[I\] for $q$ a square a sequence of modular curves which attains the upper bound in (1). It then came as a surprise when in 1995 Garcia and Stichtenoth constructed in \[G-S1\] for $q$ a square a tower of Artin-Schreier covers $$\ldots \tto C_i \tto C_{i-1}\tto\ldots \tto C_1 \tto C_0$$ which is defined over $\F_q$ and given by a simple recursive equation such that $$\lim_{i\to \infty} g(C_i)=\infty \qquad {\rm and} \quad \lim_{i\to \infty} {\# C_i(\F_q) \over g(C_i)}= \sqrt{q}-1.$$ An infinite tower $C_{\bullet}$ of covers of curves over $\F_q$ $$\ldots \tto C_i \tto C_{i-1}\tto\ldots \tto C_1 \tto C_0$$ with $g(C_i)>1$ for some $i\geq 0$ is called an [ asymptotically good tower]{} if $$\ell(C_{\bullet})= \lim_{i \to \infty} { \# C_i(\F_q) \over g(C_i)} > 0.$$ Note that in \[G-S2\] it is shown that this limit exists for towers having at least one index $i$ with $g(C_i)>1$. Apart from having an evident charm of their own, asymptotically good towers are important for coding theory, since such towers enable the construction of long error correcting codes over $\F_q$ which can correct a fixed percentage of errors per codeword and have a positive transmission rate. However, for this application it is essential that the curves are in explicit form. For $q$ is not a square the results are much less complete. It is not known how good the Drinfeld-Vladuts upper bound (1) is for that case. For $q$ not a square asymptotically good towers of curves are mainly obtained by class field theory, see for example \[N-X\]. These constructions are not explicit. In 1985 Zink using certain Shimura surfaces constructed in \[Z\] a tower $C_{\bullet}$ of curves defined over $\F_{p^3}$ for $p$ a prime with limit $$\ell(C_{\bullet})\geq { 2(p^2-1)\over p+2}, \eqno(2)$$ but that construction is far from explicit. In \[G-S-T\] there is an explicit asymptotically good tower of Kummer covers over $\F_{q=p^m}$ for $m\geq 2$ with limit $$\ell(C_{\bullet}) \geq { 2 \over q-2}.$$ Here we present an explicit tower $C_{\bullet}$ of Artin-Schreier curves defined over $\F_8$ given by a simple recursive equation with limit $$\ell(C_{\bullet})= 3/2.$$ One should compare this with (2). It remains an interesting problem to see whether our explicit tower is related to that of Zink, cf. the remarks made by Elkies at the end of \[E\]. Another interesting problem is to extend our construction to other fields of odd degree over the prime field. We give explicit formulas for the genus and number of rational points for the curves $C_i$ in our tower. The ramification behavior turns out to be rather subtle with alternating ramification and non-ramification. §1 The basic equation In our search for curves over finite fields with many points we came across a curve defined over $\F_8$ with a remarkable property. The curve of genus $1$ given by the affine equation $$x_1^2+x_1=x_0+1+1/x_0$$ has $14$ $\F_8$-rational points and attains the Hasse-Weil bound for $\F_8$. To each $x_0\in \F_8-\F_2$ there correspond $2$ solutions $x_1 \in \F_8-\F_2$ and if $x_0$ runs through $\F_8-\F_2$ then so does $x_1$. This implies that the system of equations $$x_{i+1}^2+x_{i+1}= x_i + 1 +1/x_i \qquad i=0,1,2,\ldots$$ has sequences of solutions $(x_0,x_1,x_2,\ldots)$ for every $x_0 \in \F_8-\F_2$. Consider in $\PP^1\times \PP^1$ over the field $\F_2$ the closure of the affine curve given by the equation $$x_1^2+x_1= x_0+1+1/x_0.$$ This defines a smooth projective curve $C$ of genus $1$ together with two morphisms $ b_1: C \to \PP^1$, $(x_0,x_1) \mapsto x_0$, and $e_1: C \to \PP^1$, $(x_0,x_1) \mapsto x_1$ of degree $2$. The curve $C$ possesses $2$ points rational over $\F_2$, $8$ points rational over $\F_4$ and $14$ points rational over $\F_8$. The correspondence $C$ in $\PP^1 \times \PP^1$ preserves the points of $\PP^1(\F_8)-\PP^1(\F_2)$ and surprisingly also those of $\PP^1(\F_4)$. We consider the following infinite tower $C_{\bullet}$ of smooth projective curves defined over $\F_2$ $$\tto C_i {\buildrel \pi_i \over \tto} C_{i-1} {\buildrel \pi_{i-1} \over \tto} \ldots {\buildrel\pi_2 \over \tto} C_1 {\buildrel\pi_1 \over \tto} C_0= \PP^1,$$ where we take an affine coordinate $x_0$ on $C_0$ and where the cover $C_i \to C_{i-1}$ is given by the affine equation $$x_i^2+x_i= x_{i-1}+1+ {1 \over x_{i-1}}\qquad \hbox {\rm for $i\geq 1$}. \eqno(3)$$ Equivalently, we can describe the curve $C_i$ as the normalization of the curve $D_i$ defined by $$D_i= \{ (p_0,p_1,\ldots,p_{i})\in \PP^1 \times \ldots \times \PP^1 : (p_j,p_{j+1}) \in C \, \hbox {\rm for $j=0,\ldots,i-1$)} \}.$$ This shows that $C_i$ for $i\geq 1$ admits two maps $b_i: C_i \tto C_{i-1} $ (resp. $e_i: C_i \tto C_{i-1}$) given (on the model $D_i$) by $(p_0,\ldots , p_i) \mapsto (p_0,\ldots,p_{i-1})$ (resp. $\mapsto (p_1,\ldots,p_i)$). The curve $C_i$ is then the normalization of the fibre product $$C_{i-1} \times_{C_{i-2}} C_{i-1}$$ via the maps $b_{i-1}$ and $e_{i-1}$. We now work over the algebraic closure $\F$ of $\F_2$ and consider geometric points of $C_i\otimes \F$. It will turn out that ramification in $C_i/C_{i-1}$ can occur only in points $P$ that map to a point of $D_i$ with coordinates in $\PP^1(\F_4)$. We therefore introduce the following notation for such points. By $P=P(a_0,a_1,\ldots,a_i)$ with $a_j \in \PP^1(\F_4)$ we denote a point on $C_i$ such that $x_j(P)=a_j$ for $0 \leq j \leq i$. That is, the point $(a_0,\ldots, a_i)$ is the image point in $D_i$ and will be called the index sequence of the point. Note that because of the normalization, a point $P(a_0,\ldots,a_i)$ of $C_i$ is not necessarily uniquely determined by its index sequence $(a_0,\ldots,a_i)$ on $D_i$. We shall write $\rho$ for a primitive element of $\F_4$. Note that in an index sequence $(a_0,a_1,\ldots,a_i)$ of a point $P$ we have $$\matrix{\infty\cr 0 \cr 1 \cr \rho \cr \rho^2\cr}\qquad \hbox{\rm is followed by} \qquad\matrix{\infty \cr \infty \cr \hbox{\rm $\rho$ or $\rho^2$} \cr \hbox{\rm $0$ or $1$}\cr \hbox{\rm $0$ or $1$} \cr}$$ Sometimes we shall write $(a_0,\ldots,a_i,\infty^j)$ for a point $(a_0,\ldots,a_i,\underbrace{\infty, \ldots, \infty}_{j \times})$. §2 The principal part of the $x_i$ The main problem to find the limit $\ell(C_{\bullet})$ of our tower lies in the determination of the genus $g(C_{i})$. In order to compute it we have to find the ramification divisor of $C_{i+1}$ over $C_i$. Since we are dealing with Artin-Schreier equations we can restrict ourselves to the points which are poles of the function $f_i= x_i+1+1/x_i$. The contribution to the ramification is determined by the orders ${\rm ord}_P(f_{i}^)$ for the poles $P$ on $C_{i}$ of the Artin-Schreier reduction $f_i^$ of the function $f_{i}$. (2.1) Lemma. The zeros of $x_i$ on $C_i$ are of the form $P(a_0,a_1,\ldots,a_i)$ with $a_i=0$, $a_{i-j}\in {\F}_4-{\F}_2$ for $j\geq 1$ odd and $a_{i-j}=1$ for $j\geq 2$ even. The poles of $x_i$ are of the form $P(b_0,b_1,\ldots,b_j,\infty^{i-j})$ with $0\leq j\leq i-1$ and $(b_0,\ldots,b_j)$ an index sequence of a zero of $x_j$ or of the form $P(\infty^{i+1})$. [Proof.]{} By induction on $i$. The lemma is true for $x_0$. From the equation $$x_i^2+x_i= {x_{i-1}^2 +x_{i-1}+1\over x_{i-1}} =f_{i-1}$$ it follows that we have the equality of divisors on $C_i$ $$(f_{i-1})= (x_i)+(x_i+1)=(x_i)_0 + (x_i)_1 -2(x_i)_{\infty}.$$ So poles of $x_i$ lie above poles of $f_{i-1}$ and the points $P$ on $C_i$ with $x_i(P) \in \F_2$ lie above the zeros of $f_{i-1}$. Moreover, we have $$\eqalign{ (f_{i-1})&= (x_{i-1}+\rho)+(x_{i-1}+\rho^2)-(x_{i-1})\cr &= (x_{i-1})_{\rho} + (x_{i-1})_{\rho^2} -(x_{i-1})_0-(x_{i-1})_{\infty}.\cr }$$ which implies that the poles of $f_{i-1}$ are the zeros and poles of $x_{i-1}$, while the zeros of $f_{i-1}$ are the points $P$ on $C_{i-1}$ with $x_{i-1}(P)\in {\F}_4 - {\F}_2$. Hence the poles of $x_i$ are the points on $C_i $ above the zeros and the poles of $x_{i-1}$ on $C_{i-1}$, whereas the zeros of $x_i$ lie above points $P$ on $C_{i-1}$ with $x_{i-1}(P) \in {\F}_4 - {\F}_2$. So we obtain the index sequence of a pole of $x_i$ by adding $\infty$ to a zero or pole of $x_{i-1}$ and we obtain the index sequence of a zero of $x_i$ by adding a zero to an index sequence which ends with an element of ${\F}_4 - {\F}_2$ and in which $1$ and elements of $\{ \rho, \rho^2\}$ alternate. $\square$ In the following we shall develop rational functions on $C_i$ as a power series in a local parameter at a given point $P$, that is, we consider the function as an element of the quotient field of the completion of the local ring of $P$. Often we are only interested in the principal part and neglect elements that are regular, i.e. elements of (the completion of) the local ring. By the notation $$f=g+O(P)$$ we mean that $f-g$ is regular in $P$, that is, is an element of $O_P$ or of $\hat{O}_P$. Consider now a sequence of points $P_0 \in C_0, P_1 \in C_1, \ldots, P_i \in C_i$ with $\pi_{\ell}(P_{\ell})=P_{\ell-1}$ for $\ell=1,\ldots,i$ and with the property that $1$ and $\rho$ or $\rho^2$ alternate in the index sequence $(a_0,\ldots,a_i)$ of $P_i$. We shall first assume that $a_0=1$. Then the function $t=x_0+1$ provides a local parameter at $P_0$ on $C_0$ and the pull back (under the maps $\pi_{\ell}$) of this function (again denoted by $t$) is still a local parameter at the points $P_j$ on $C_j$ for $j\leq i$. In the completion of the local ring $\hat{O}_{P_j}\cong \F[[t]]$ the function $x_j$ can be written as a power series in $t$ $$x_j= a_j +m_j(t),$$ where $m_j(t) \in \F[[t]]$ has ${\rm ord}_t(m_j)\geq 1$. (2.2) Lemma. In the quotient field $\F((t))$ of the formal power series ring $\hat{O}_{P_j}\cong \F[[t]]$ the function $m_j(t)$ satisfies for $0\leq j \leq i$ the relations $$\eqalign{ {1 \over m_j}& = {a_{j-1} \over m_{j-1}}+O(P_j) \qquad \hbox{\rm for $j \geq 2$ even},\cr {1 \over m_j}& = {1 \over m_{j-1}^2}+ { 1 \over m_{j-1}} +O(P_j) \qquad \hbox{\rm for $j $ odd.}\cr }$$ [Proof.]{} We start with $m_0(t)=t$. For even $j\geq 2$ we have $a_{j-1} \in \{ \rho, \rho^2\}$ and $a_j=1$ since we assumed that $a_0=1$. From the relation $$x_j^2+x_j = x_{j-1}+ 1 + 1/x_{j-1}$$ we obtain $$\eqalign{ m_j^2+m_j &= a_{j-1}+m_{j-1} +1 + 1/(a_{j-1}+m_{j-1})\cr &= a_{j-1} + m_{j-1} + 1 + (1 /a_{j-1}) \sum_{n=0}^\infty (m_{j-1} / a_{j-1})^n\cr &= a_{j-1}^2m_{j-1} + m_{j-1}^2 + \hbox{\rm higher powers of $m_{j-1}$}.\cr }$$ This implies that $m_j$ is the product of $a_{j-1}^2m_{j-1}$ with a $1$-unit $u$ in $m_{j-1}$, i.e. a unit of the form $u=1+r$ with $r\in (m_{j-1})$. So we get $${1 \over m_{j}}= {a_{j-1} \over m_{j-1}} \cdot u = {a_{j-1} \over m_{j-1}} + O(P_{j}).$$ For $j$ odd we have $a_j \in \{ \rho, \rho^2\}$ and $a_{j-1}=1$. In the same way as for $j$ even we obtain $$a_j^2+m_j^2 +a_j+m_j= 1 + m_{j-1} +1 + {1 \over 1 + m_{j-1}}= 1 + \sum_{n=2}^{\infty} m_{j-1}^n,$$ that is, $ m_j^2+m_j= \sum_{n=2}^{\infty} m_{j-1}^n $, so that we have $$m_j= m_{j-1}^2+m_{j-1}^3 +\hbox{\rm higher powers of $m_{j-1}$}.$$ This means that $$\eqalign{ {1 \over m_{j}} &= {1 \over m_{j-1}^2} + { 1 \over m_{j-1}} + \hbox{\rm higher powers of $m_{j-1}$}\cr &= {1 \over m_{j-1}^2} + {1 \over m_{j-1}} + O(P_j). }$$ This completes the proof of the lemma. $\square$ We denote the principal part of $1/m_j$ by $F_j$. We now can deduce the following corollary. (2.3) Corollary. The principal part $F_j$ of $1/m_j$ satisfies: $$F_j=\cases{ F_{j-1}^2+F_{j-1} & for $j$ odd\cr a_{j-1}\cdot F_{j-1} & for $j\geq 2$ even.\cr}$$ Furthermore, $F_j$ is a $2$-linearized polynomial in $1/t$ of the form $$F_j= {b_{k} \over t^{2^k}} + {b_{k-1} \over t^{2^{k-1}}} + \ldots + {b_0 \over t},$$ where $k=[(j+1)/2]$, the coefficients $b_{\ell}$ are in $\F_4$ and $b_{k}\neq 0$. [Proof.]{} The relations for $F_j$ follow at once from Lemma.(2.2). We have $F_0=1/t$ and $F_1= 1/t^2+1/t$ from which the formula for $F_j$ follows by induction. $\square$ For an index sequence $(a_0,a_1,\ldots, a_i)$ where $a_0 \in \{ \rho, \rho^2\}$ we have a similar result. §3 The ramification behavior Now we study the ramification behavior in a point $P_i= P(a_0,\ldots,a_{i-1},a_i=0)$ which is a zero of $x_i$ on $C_i$ for $i\geq 2$. We assume $a_0=1$. Then $a_{\rm odd}\in \{ \rho, \rho^2\}$ and $i$ is even. In a point $P_i$ where $a_0 \in \{ \rho, \rho^2\} $ the ramification behavior is similar. Since we are working with Artin-Schreier covers here we introduce the standard notation $\wp(f)= f^2+f$ for an element $f$ in one of our function fields. (3.1) Lemma. A linear combination $\sum_{j=2, \, \rm even}^i B_{j,i} F_j$ with coefficients $B_{j,i}\in \F_4$ can be written as $$\sum_{j=2, \, \rm even}^i B_{j,i} F_j= \wp(\sum_{j=0, \, \rm even}^{i-2} B_{j,i-2}F_j) + B_i^* F_0 \eqno(4)$$ with $$B_i^= \wp(\sum_{j=2,\, \rm even}^i B_{j,i}a_{j-1} ) \eqno(5)$$ and $$B_{j,i-2}=\big( B_i^+\wp( \sum_{k=2,\, \rm even}^j B_{k,i}a_{k-1}) \big)a_{j+1}^2 + \wp(B_{j+2,i}). \eqno(6)$$ [Proof.]{} Using Corollary (2.3) we find for even $j \geq 2$ $$B_{j,i}F_j = B_{j,i}a_{j-1}F_{j-1} = B_{j,i}a_{j-1}\wp( F_{j-2})= \wp(B_{j,i}^2a_{j-1}^2 F_{j-2})+\wp(B_{j,i}a_{j-1})F_{j-2},$$ that is $$B_{j,i}F_j= \wp(B_{j,i}^2a_{j-1}^2 F_{j-2})+\wp(B_{j,i}a_{j-1})F_{j-2}. \eqno(7)$$ Applying (7) to the second term in the RHS of (7) we obtain $$\wp(B_{j,i}a_{j-1})F_{j-2} = \wp(( \wp(B_{j,i}a_{j-1})a_{j-3}^2F_{j-4})+ \wp(B_{j,i} a_{j-1})F_{j-4},$$ where we use $a_{\rm odd}^2+a_{\rm odd}=1$. Continuing this way we find an expression for $B_{j,i}F_j$ as $$\wp( \hbox{\rm linear combination of $F_{j-2},\ldots, F_0$ }) +\wp(B_{j,i}a_{j-1})F_0.$$ Adding these relations for all terms in $\sum_{j=2, \, \rm even}^i B_{j,i}F_j$ we find (4) with coefficients satisfying the equations (5) and (6). $\square$ Note that all coefficients are in $\F_4$ and that $B_i^$ is in $\F_2$. The cover $C_{i+1}/C_{i}$ is given by the equation $$x_{i+1}^2+x_{i+1}= x_i + 1 +1/x_i.$$ In a point $P_i $ with index sequence $(1,a_1,\ldots,a_{i-1},0)$ we have the relation $$x_{i+1}^2+x_{i+1} = {1 \over m_i} + O(P_i) = F_i+O(P_i). \eqno(8)$$ Therefore the principal part of $x_{i+1}^2+x_{i+1}$ in $P_i$ is $F_i$. According to Lemma (3.1) we have $$F_i= \wp(\sum_{j=0, \, \rm even}^{i-2} B_{j,i-2} F_j) + B_i^F_0$$ with $B_i^= a_{i-1}^2+a_{i-1}=1$ and $B_{j,i-2}=a_{j+1}^2$ for even $j=0, 2, \ldots, i-2$. By the substitution $$X_{i+1}= \sum_{j=0}^{i-2} B_{j,i-2} F_j + x_{i+1}$$ we can reduce the equation (8) to $$X_{i+1}^2+X_{i+1}= B_i^F_0 + O(P_i) = F_0 + O(P_i) \eqno(9)$$ with $F_0=1/t$. (3.2) Corollary. The point $P_i=P(a_0=1,a_1,\ldots,a_i=0)$ is totally ramified in $C_{i+1}/C_i$ and the contribution of $P_i$ to the ramification divisor of $C_{i+1}/C_i$ is $2$. In the pole $P_{i+1} =P_i(\infty)$ of $x_{i+1}$ we have $${\rm ord}_{P_{i+1}}(x_{i+1})= -2^{[(i+1)/2]}.$$ As the next step we consider the behavior of the pole $P_{i+1}=P_i(\infty)$ of the function $f_{i+1}=x_{i+1} + 1 + (1/x_{i+1})$ in the cover $C_{i+2}/C_{i+1}$. In $P_{i+1}$ the equation of $C_{i+2}/C_{i+1}$ is $$\eqalign{ x_{i+2}^2+x_{i+2}&= x_{i+1} + O(P_{i+1}) \cr &=\sum_{j=0,\, \rm even}^{i-2} B_{j,i-2}F_j + X_{i+1} + O(P_{i+1}).\cr } \eqno(10)$$ If we apply Lemma (3.1) to the linear combination $\sum_{j=2,\, \rm even}^{i-2} B_{j,i-2} F_j$ the RHS of (10) becomes $$\wp(\sum_{j=0, \, \rm even}^{i-4} B_{j,i-4}F_j) + B_{i-2}^* F_0 + B_{0,i-2}F_0 + X_{i+1} + O(P_{i+1}). \eqno(11)$$ Then by using (9), i.e., by substituting $F_0= X_{i+1}^2+X_{i+1} + O(P_{i+1})$, the expression (11) is converted to $$\wp(\sum_{j=0, \, \rm even}^{i-4} B_{j,i-4} F_j)+ \wp(B_{i-2}^* X_{i+1}) + \wp (B_{0,i-2}^2X_{i+1})+ (B_{0,i-2}^2+B_{0,i-2}+1)X_{i+1} + O(P_{i+1})$$ with $B_{0,i-2}^2+B_{0,i-2}+1= a_1^2+a_1+1=0$. Hence the equation of $C_{i+2}/C_{i+1}$ in $P_{i+1}$ is of the form $$x_{i+2}^2+x_{i+2} = \wp( \sum_{j=0}^{i-4} B_{j,i-4}F_j) + \wp((B_{i-2}^* + B_{0,i-2}^2)X_{i+1}) + O(P_{i+1}).\eqno(12)$$ (3.3) Corollary. The pole $P_{i+1}$ of $f_{i+1}$ is unramified in the cover $C_{i+2}/C_{i+1}$ and in a point $P_{i+2}=P_{i+1}(\infty)$ above $P_{i+1}$ we have $${\rm ord}_{P_{i+2}}(x_{i+2})= -2^{[(i+1)/2]-1}.$$ Note that (12) implies that $$x_{i+2}+(\sum_{j=0, \, \rm even}^{i-4} B_{j,i-4}F_j) + (B_{i-2}^+B_{0,i-2}^2)X_{i+1}$$ is integral in the point $P_{i+2}$. To analyze the situation in $P_{i+2}$ we start with the equation of $C_{i+3}/ C_{i+2}$ in this point: $$\eqalign{ x_{i+3}^2+x_{i+3}&= x_{i+2} + O(P_{i+2})\cr &=\big(\sum_{j=0,\, \rm even}^{i-4} B_{j,i-4}F_j\big)+ (B_{i-2}^ +B_{0,i-2}^2)X_{i+1} + O(P_{i+2}). \cr}\eqno(13)$$ Using Lemma (3.1) and (9) the RHS of (13) is of the form $$\eqalign{ \wp(\sum_{j=0, \, \rm even}^{i-6} B_{j,i-6}F_j) + & \wp((B_{i-4}^+B_{0,i-4}^2) X_{i+1})+ \cr + (\wp(B_{0,i-4})+ & B_{i-2}^+B_{0,i-2}^2) X_{i+1} + O(P_{i+2}). \cr} \eqno(14)$$ Since (6) implies $\wp(B_{0,i-4})=B_{i-2}^$ the coefficient of $X_{i+1}$ in (14) is $B_{0,i-2}^2=a_1^2$. So the right hand side of (13) has the form $$\wp(\gamma) + B_{0,i-2}^2X_{i+1} + O(P_{i+2})$$ for some $\gamma \in \F(C_{i+2})$. If we set $$X_{i+3}= \sum_{j=0}^{i-6} B_{j,i-6}F_j + (B_{i-4}^+B_{0,i-4}^2)X_{i+1}$$ the equation of $C_{i+3}/C_{i+2}$ becomes $$X_{i+3}^2+X_{i+3}= B_{0,i-2}^2 X_{i+1} + O(P_{i+2}). \eqno(15)$$ (3.4) Corollary. The point $P_{i+2}$ is totally ramified in the cover $C_{i+3}/C_{i+2}$ and the contribution to the ramification divisor is $2$. In $P_{i+3}=P_{i+2}(\infty)$ above $P_{i+2}$ we have $${\rm ord}_{P_{i+3}}(x_{i+3})= -2^{[(i+1)/2]-1}.$$ If we continue along these lines we obtain the following formulas: (3.5) Formula. For $t$ even and $2\leq t \leq i$ the equation of $C_{i+t}/C_{i+t-1}$ in a point $P_{i+t-1}$ is $$\eqalign{ x_{i+t}^2+x_{i+t}=\wp(\sum_{j=0,\, \rm even}^{i-2t} B_{j,i-2t}F_j) + \wp((B_{i-2t+2}^+B_{0,i-2t+2}^2)X_{i+1})+ &\cr \quad +\sum_{k=1}^{t/2-1} \wp((B_{2k-2,i-2t+4k}B_{2k-2,i-2}^2)X_{i+2k+1}) +& O(P_{i+t-1}). \cr }$$ (3.6) Formula. For $t$ odd and $3\leq t \leq i-1$ the equation of $C_{i+t}/C_{i+t-1}$ in a point $P_{i+t-1}$ is $$\eqalign{ x_{i+t}^2+x_{i+t}=\wp(\sum_{j=0, \, \rm even}^{i-2t} B_{j,i-2t}F_j) + \wp((B_{i-2t+2}^+B_{0,i-2t+2}^2)X_{i+1})+ &\cr \quad+\sum_{k=1}^{(t-3)/2} \wp((B_{2k-2,i-2t+4k}B_{2k-2,i-2}^2)X_{i+2k+1}) +B_{t-3,i-2}^2X_{i+t-2} +&O(P_{i+t-1}). \cr }$$ [(3.7) Remark.]{} The function $X_{i+2k+1}$ with $k\geq 1$ satisfies an equation of the form $$X_{i+2k+1}^2+X_{i+2k+1}=B_{2k-2,i-2}^2 X_{i+2k-1} + O(P_{i+2k}).$$ We also find $${\rm ord}_{P_{i+t}}(x_{i+t})=-2^{[(i+2)/2]-[t/2]} \qquad {\rm for} \quad 2 \leq t \leq i.$$ Hence ${\rm ord}_{P_{2i}}(x_{2i})=-1$ and this implies that from $P_{2i}$ on the extensions in the tower are totally ramified above $P_{2i}$ and their contribution to the ramification divisor is $2$. We summarize the preceding results in the following theorem. (3.8) Theorem. A pole $P_{i+j}=P(a_0,\ldots,a_{i-1},0,\infty^j)$ of $x_{i+j}$ on $C_{i+j}$ for $j\geq 1$ (resp. of $1/x_i$ for $j=0$) with $a_0=1$ is [1)]{} totally ramified in $C_{i+j+1}/C_{i+j}$ for $j$ with $j=0,2,4,\ldots,i-2$ or with $j\geq i$ and each of these contributes $2$ to the ramification divisor, [2)]{} unramified in $C_{i+j+1}/C_{i+j}$ for $j=1,3,5,\ldots,i-1$. For a pole whose index sequence starts with $a_0 \in \{ \rho, \rho^2\}$ there is the following similar result. (3.9) Theorem. A pole $P_{i+j}=P(a_0,\ldots,a_{i-1},0,\infty^j)$ of the function $x_{i+j}$ on $C_{i+j}$ with $a_0\in \{\rho, \rho^2\}$ is [1)]{} totally ramified in $C_{i+j+1}/C_{i+j}$ for all $j$ with $j=0,2,4,\ldots,i-3$ and for all $j$ with $j\geq i-1$ with contribution $2$ to the ramification divisor, [2)]{} unramified in $C_{i+j+1}/C_{i+j}$ for $j=1,3,5,\ldots,i-2$. [(3.10) Remark.]{} We always have the totally ramified points $P(\infty,\infty, \ldots,\infty)$ and also $P(0,\infty,\infty,\ldots,\infty)$ which contribute $2$ to the ramification divisor. §4 The genus and the number of points in the tower In order to compute the genus of our curve $C_i$ we have to count the number of points on the curve $C_i$ which contribute to the ramification divisor. We find using (3.8)–(3.10) (4.1) Theorem. Let $n_i$ be the number of points on $C_i$ which are totally ramified in $C_{i+1}/C_i$. Then $$n_i=\cases{ ([(i+2)/ 4]+2) 2^{i/2} & for $i$ even,\cr ([i/ 4]+2) 2^{(i+1)/2} & for $i$ odd.\cr }$$ Now it is not difficult to determine the genus $g(C_i)$ of $C_i$. From the Hurwitz formula it follows that $$g(C_i)=1+ \sum_{j=1}^{i-1} 2^{i-j-1} n_j. \eqno(16)$$ If we combine (16) with theorem (4.1) we get: (4.2) Theorem. The genus $g(C_i)$ of $C_i$ satisfies $$g(C_i)= 2^{i+2}+1-\cases{ (i+10)2^{(i/2)-1} & for $i$ even, \cr (i+2[i / 4]+15) 2^{(i-3)/2} & for $i$ odd.\cr }$$ Now we count the number $\#C_i(\F_8)$ of $\F_8$-rational points on $C_i$. (4.3) Theorem. We have $\# C_i(\F_8)= 6\cdot 2^i +2$. [Proof.]{} Let $\alpha$ be a primitive element of $\F_8$ which satisfies $\alpha^3+\alpha +1=0$. For $x \in \F_8-\F_2$ we find $$\{ x+{1 \over x} +1 : x \in \F_8-\F_2\} = \{ \alpha, \alpha^2,\alpha^4\},$$ but also $$\{ y^2+y: y \in \F_8-\F_2\} =\{ \alpha, \alpha^2,\alpha^4\}.$$ This means that a point $x \in \PP^1(\F_8)$ with $x \not\in \PP^1(\F_2)$ splits completely in the tower. This yields $6 \cdot 2^i$ rational points over $\F_8$ on $C_i$. Besides these we have two totally ramified points $P(0,\infty,\ldots,\infty)$ and $P(\infty, \ldots, \infty)$ defined over $\F_2$. $\square$ Combining this with the formula for the genus we obtain the following Theorem. (4.4) Theorem. The tower of curves $C_{\bullet}$ over $\F_8$ is asymptotically good with limit $$\lim_{i\to \infty} { \# C_i(\F_8) \over g(C_i)} = {3 \over 2}.$$ References \[D-V\] V.G. Drinfeld, S.G. Vladuts: Number of points of an algebraic curve. [Funct. Anal. 17*]{} (1983), p. 68–69. \[E\] N. Elkies: Explicit modular towers. Preprint Harvard University, 2000. \[G-S1\] A. Garcia, H. Stichtenoth: A tower of Artin-Schreier extensions of function fields attaining the Drinfeld-Vladut bound. [Inv. Math. 121]{} (1995), p. 211–222. \[G-S2\] A. Garcia, H. Stichtenoth: On the asymptotic behavior of some towers of function fields over finite fields. [J. of Number Th. 61]{} (1996), p. 248–273. \[G-S-T\] A. Garcia, H. Stichtenoth, M. Thomas: On towers and composita of towers of function fields over finite fields. [Finite Fields and their Appl. 3]{} (1997), p. 257–274. \[I\] Y. Ihara: Some remarks on the number of points of algebraic curves over finite fields. [J. Fac. Sci. Tokyo, Ser. Ia, 28]{} (1982), p. 721–724. \[N-X\] H. Niederreiter, C. Xing: Global function fields with many rational places and their applications. [Contemp. Math. 225]{} (1999), p. 87–111. \[Z\] Th. Zink: Degeneration of Shimura curves and a problem in coding theory. In [Fundamentals of Computation Theory*]{}, Springer LNCS 199, p. 503–511. Springer, Berlin 1985. 3 G. van der Geer &&M. van der VlugtFaculteit Wiskunde en Informatica &&Mathematisch InstituutUniversiteit van Amsterdam &&Rijksuniversiteit te Leiden Plantage Muidergracht 24&&Niels Bohrweg 1 1018 TV Amsterdam &&2333 CA Leiden The Netherlands &&The Netherlands &&[[email protected]]{}	{ "pile_set_name": "ArXiv" }
--- author: - 'K. Chilikin' - 'I. Adachi' - 'H. Aihara' - 'S. Al Said' - 'D. M. Asner' - 'V. Aulchenko' - 'R. Ayad' - 'V. Babu' - 'I. Badhrees' - 'A. M. Bakich' - 'V. Bansal' - 'E. Barberio' - 'D. Besson' - 'V. Bhardwaj' - 'B. Bhuyan' - 'J. Biswal' - 'A. Bobrov' - 'A. Bondar' - 'A. Bozek' - 'M. Bračko' - 'T. E. Browder' - 'D. Červenkov' - 'V. Chekelian' - 'A. Chen' - 'B. G. Cheon' - 'K. Cho' - 'Y. Choi' - 'D. Cinabro' - 'N. Dash' - 'S. Di Carlo' - 'Z. Doležal' - 'Z. Drásal' - 'D. Dutta' - 'S. Eidelman' - 'H. Farhat' - 'J. E. Fast' - 'T. Ferber' - 'B. G. Fulsom' - 'V. Gaur' - 'N. Gabyshev' - 'A. Garmash' - 'R. Gillard' - 'P. Goldenzweig' - 'J. Haba' - 'T. Hara' - 'K. Hayasaka' - 'W.-S. Hou' - 'K. Inami' - 'A. Ishikawa' - 'R. Itoh' - 'Y. Iwasaki' - 'W. W. Jacobs' - 'I. Jaegle' - 'H. B. Jeon' - 'Y. Jin' - 'D. Joffe' - 'K. K. Joo' - 'T. Julius' - 'K. H. Kang' - 'G. Karyan' - 'P. Katrenko' - 'D. Y. Kim' - 'H. J. Kim' - 'J. B. Kim' - 'K. T. Kim' - 'M. J. Kim' - 'S. H. Kim' - 'Y. J. Kim' - 'K. Kinoshita' - 'P. Kodyš' - 'S. Korpar' - 'D. Kotchetkov' - 'P. Križan' - 'P. Krokovny' - 'T. Kuhr' - 'R. Kulasiri' - 'A. Kuzmin' - 'Y.-J. Kwon' - 'J. S. Lange' - 'L. Li' - 'L. Li Gioi' - 'J. Libby' - 'D. Liventsev' - 'M. Lubej' - 'T. Luo' - 'M. Masuda' - 'T. Matsuda' - 'D. Matvienko' - 'K. Miyabayashi' - 'H. Miyata' - 'R. Mizuk' - 'G. B. Mohanty' - 'H. K. Moon' - 'T. Mori' - 'R. Mussa' - 'E. Nakano' - 'M. Nakao' - 'T. Nanut' - 'K. J. Nath' - 'Z. Natkaniec' - 'M. Nayak' - 'M. Niiyama' - 'N. K. Nisar' - 'S. Nishida' - 'S. Ogawa' - 'S. Okuno' - 'S. L. Olsen' - 'H. Ono' - 'P. Pakhlov' - 'G. Pakhlova' - 'B. Pal' - 'S. Pardi' - 'H. Park' - 'S. Paul' - 'R. Pestotnik' - 'L. E. Piilonen' - 'C. Pulvermacher' - 'M. Ritter' - 'H. Sahoo' - 'Y. Sakai' - 'M. Salehi' - 'S. Sandilya' - 'L. Santelj' - 'T. Sanuki' - 'O. Schneider' - 'G. Schnell' - 'C. Schwanda' - 'Y. Seino' - 'K. Senyo' - 'O. Seon' - 'M. E. Sevior' - 'V. Shebalin' - 'C. P. Shen' - 'T.-A. Shibata' - 'J.-G. Shiu' - 'A. Sokolov' - 'E. Solovieva' - 'M. Starič' - 'T. Sumiyoshi' - 'M. Takizawa' - 'U. Tamponi' - 'K. Tanida' - 'F. Tenchini' - 'K. Trabelsi' - 'M. Uchida' - 'S. Uehara' - 'T. Uglov' - 'S. Uno' - 'Y. Usov' - 'C. Van Hulse' - 'G. Varner' - 'A. Vinokurova' - 'A. Vossen' - 'C. H. Wang' - 'M.-Z. Wang' - 'P. Wang' - 'M. Watanabe' - 'Y. Watanabe' - 'E. Widmann' - 'E. Won' - 'H. Yamamoto' - 'Y. Yamashita' - 'H. Ye' - 'C. Z. Yuan' - 'Y. Yusa' - 'Z. P. Zhang' - 'V. Zhilich' - 'V. Zhulanov' - 'A. Zupanc' ---	{ "pile_set_name": "ArXiv" }
--- abstract: 'Classically, there exists a determinant map from the moduli space of semi-stable sheaves on a smooth, projective variety to the Picard scheme. Unfortunately, if the underlying variety is singular, then such a map does not exist. In the case the underlying variety is a nodal curve, a similar map was produced by Bhosle on a stratification of the moduli space of semi-stable sheaves. In this note, we generalize this result to the higher dimension case.' address: - 'BCAM - Basque Centre for Applied Mathematics, Alameda de Mazarredo 14, 48009 Bilbao, Spain' - 'Freie Universitaet Berlin, FB Mathematik und Informatik, Arnimallee 3, 14195 Berlin, Germany' author: - Ananyo Dan - Inder Kaur bibliography: - 'researchbib.bib' title: A note on the determinant map --- Introduction ============ Recall, the classical notion of determinant of a coherent sheaf. Given a projective scheme $X$ and a coherent sheaf ${\mathcal{F}}$ with a finite locally free resolution, $$0 \to {\mathcal{L}}_r \to {\mathcal{L}}_{r-1} \to ... \to {\mathcal{L}}_0 \to {\mathcal{F}} \to 0,$$ the determinant of ${\mathcal{F}}$, $\det({\mathcal{F}})$ is defined to be $\otimes \det({\mathcal{L}}_i)^{(-1)^i}$. If $X$ is not regular, one cannot guarantee the existence of such a finite locally free resolution. So, the classical definition of determinant cannot be extended to the general case. One of the first results in this direction was due to Bhosle (see [@bhos2 Proposition $4.7$]), where she considers moduli of semi-stable sheaves on nodal curves. She introduces the theory of parabolic bundles and their moduli spaces. Using this, she defines a determinant map from a stratification of a given moduli space of semi-stable sheaves on the nodal curve to certain moduli spaces of parabolic line bundles. The stratification of the moduli space arises from an explicit description of the stalk of a torsion-free sheaf at a node on a curve. Unfortunately, such a description does not exist for higher dimensional projective varieties. Hence, her techniques cannot be generalized to higher dimension. In this article, we use completely different techniques to obtain a similar result without any restriction on the dimension of the underlying scheme. In this article we introduce the notion of the alternating determinant of a rank $n$ coherent sheaf ${\mathcal{F}}$, denoted $\operatorname{\mathcal{A}lt}^n({\mathcal{F}})$. This sheaf is a semi-stable rank one sheaf (see Proposition \[comp9\]). We prove that, \[cn12\] There exists a stratification $$M_X(P)=\coprod_{i=1}^\infty V_{R_i},$$ by locally closed subscheme $V_{R_i}$ satisfying: $\dim V_{R_i}>V_{R_{i+1}}$ unless $V_{R_i}=\emptyset$ and for each $i$, there exists a Hilbert polynomial $L_i$ such that there exists an alternating determinant map $${\mathrm{Adet}}:\coprod V_{R_i} \to \coprod M_X(L_i)$$ with ${\mathrm{Adet}}\|_{V_{R_i}}$ taking values in $M_X(L_i)$. In the case, the underlying variety is a nodal curve, our stratification agrees with the one given in [@bhos2]. Moreover, there is a natural map from $\operatorname{\mathcal{A}lt}^n({\mathcal{F}})$ to the determinant of ${\mathcal{F}}$, defined by Bhosle in terms of parabolic line bundles. The two sheaves differ only at the nodes by certain mixed terms described in [@bhos2 §$4.6$]. Furthermore, if the underlying variety $X$ is smooth, we have the following: \[cn11\] Let $X$ be a smooth, projective variety and $P$ the Hilbert polynomial of a torsion-free semi-stable rank $n$ sheaf on $X$ with degree coprime to $n$. Denote by $M_X(P)$ the moduli space of torsion-free semi-stable sheaves on $X$ of rank $n$. Then, 1. there exists a Hilbert polynomial $L$ of a rank $1$ torsion-free sheaf on $X$ such that for any closed point $s \in M_X(P)$, the corresponding coherent sheaf ${\mathcal{F}}_s$ satisfies the condition: $\operatorname{\mathcal{A}lt}^n({\mathcal{F}}_s)$ has Hilbert polynomial $L$. Furthermore, $\operatorname{\mathcal{A}lt}^n({\mathcal{F}}_s)$ is an invertible sheaf. 2. there exists a natural map ${\mathrm{Adet}}:M_X(P) \to M_X(L)$ which maps a closed point $s \in M_X(P)$ to the point in $M_X(L)$ corresponding to $\operatorname{\mathcal{A}lt}^n({\mathcal{F}}_s)$. See Proposition \[comp12\] and Corollary \[comp19\] for a proof of the statement. We now discuss our strategy. We define the alternating determinant $\operatorname{\mathcal{A}lt}^n({\mathcal{F}})$ as the sheaf of alternating multilinear $n$-forms on ${\mathcal{F}}$ (see Definiton \[com05\]). Ofcourse, if a sheaf is locally free then its alternating determinant is the same as the dual of its determinant. One can check that in the case $X$ is non-singular and ${\mathcal{F}}$ is torsion-free, the alternating determinant of ${\mathcal{F}}$ is also isomorphic to the dual of the determinant i.e., $\operatorname{\mathcal{A}lt}^n({\mathcal{F}})=\det({\mathcal{F}})^\vee$ (see Proposition \[comp12\]). In general (when $X$ is just a projective variety), $\operatorname{\mathcal{A}lt}^n({\mathcal{F}})$ is a rank one semi-stable sheaf. Therefore, a map on a moduli space of semi-stable sheaves, induced by taking alternating determinant, must have image in a moduli space of rank one semi-stable sheaves. Using Yoneda embedding one observes that such a map must be induced by a natural transformation between the corresponding moduli functors. This is where the problem lies. The obstruction to defining such a natural tranformation is the fact that alternating determinant of a sheaf need not commute with pullback. However, we can stratify the moduli space such that there exists a well-defined alternating determinant map on each strata. We first prove that given any locally closed subscheme of the Quot-scheme parametrizing semi-stable quotient sheaves, the locus of points where the $\operatorname{\mathcal{H}om}$-functor commute with pull-back, is open (see Theorem \[comp6\] and Proposition \[cn4\]). This gives rise to the required stratification on the semi-stable locus of the Quot-scheme (see Notation \[cn13\]). One can then observe that this stratification induces a similar stratification on the moduli space (see Corollary \[comp18\] and Theorem \[cn8\]). The remaining statements of Theorem \[cn12\] is not hard. The proof of Theorem \[cn11\] is a direct application of reflexive sheaves and basic properties of locally free sheaves. We fix some notations that will be used throughout this article. Denote by $k$, an algebraically closed field of any characteristic. Given a projective $k$-variety $X$ and a $k$-algebra $A$, denote by $X \times A$ the scheme $X \times \operatorname{\mathrm{Spec}}(A)$. Given a sheaf ${\mathcal{F}}$ on $X \times A$ and an $A$-module $M$, denote by ${\mathcal{F}} \otimes_A M$, the sheaf associated to the presheaf which to an open set $U \subset X \times A$ associates the $\operatorname{\mathcal{O}}_{X \times A}(U)$-module ${\mathcal{F}}(U) \otimes_A M$. Acknowledgements The authors thank Carlos Simpson for a helpful conversation. This work was done when the first author was a post-doctoral fellow and the second author was a visiting researcher at the Tata Institute of Fundamental Research, Mumbai. The second author thanks TIFR for their hospitality. The first author is currently supported by ERCEA Consolidator Grant $615655$-NMST and also by the Basque Government through the BERC $2014-2017$ program and by Spanish Ministry of Economy and Competitiveness MINECO: BCAM Severo Ochoa excellence accreditation SEV-$2013-0323$ Hom functor on families of schemes ================================== The aim of this section is to study variation of dual of a coherent sheaf in flat families of noetherian schemes. We prove that given a flat family $f:X \to Y$ and a coherent sheaf ${\mathcal{F}}$ on $X$, the locus of points $u \in Y$ for which the natural map $({\mathcal{F}}^\vee)\|_{X_u} \to ({\mathcal{F}}\|_{X_u})^\vee$ is an isomorphism, is open in $Y$ (see Theorem \[comp6\]). This is a generalization of [@R1 Theorem III.$12.11$] with two major differences: the sheaf ${\mathcal{F}}$ need not be flat, contrary to the assumption in the reference. Furthermore, in the case $Y=\operatorname{\mathrm{Spec}}(A)$ for a noetherian ring $A$, the functor $T^i$ from the category of $A$-modules to itself (see [@R1 III. §$12$]) does not map a finitely generated $A$-module $M$ to another finitely generated $A$-module, hence differs from the setup in the reference. Due to these two properties, several conclusions in the reference fail. We circumvent these problems to prove Theorem \[comp6\]. The theorem plays an important role in the remaining part of the article. Let $f:X \to Y$ be a proper, flat surjective morphism between noetherian schemes. Assume $Y={\mathrm{Spec}}(A)$ for some noetherian ring $A$. Let ${\mathcal{F}}$ be a coherent sheaf on $X$ (not necessarily flat over $Y$). The main result of this section is the following: \[comp6\] Suppose there exists $t \in Y$ such that the natural morphism $$\operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_X}({\mathcal{F}},\operatorname{\mathcal{O}}_X)\|_{X_t} \to \operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_X}({\mathcal{F}},\operatorname{\mathcal{O}}_{X_t})$$ is isomorphic. Then, there exists an open neighbourhood $U$ in $Y$ containing $t$ such that for all $u \in U$, the morphism $$\operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_X}({\mathcal{F}},\operatorname{\mathcal{O}}_X)\|_{X_u} \to \operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_X}({\mathcal{F}},\operatorname{\mathcal{O}}_{X_u})$$ is isomorphic. \[com03\] Denote by $\mathfrak{M}_A$ the category of $A$-modules and ${\mathrm{Coh}}_X$ the category of coherent sheaves on $X$. Define the functor $T^i:\mathfrak{M}_A \to {\mathrm{Coh}}_X$ which associates to an $A$-module $M$, the sheaf ${\mathcal{E}}xt^i_{\operatorname{\mathcal{O}}_X}({\mathcal{F}},\operatorname{\mathcal{O}}_X \otimes_A M)$, where $\operatorname{\mathcal{O}}_X \otimes_A M$ is the sheaf defined by $\operatorname{\mathcal{O}}_X \otimes_A M(U)=\operatorname{\mathcal{O}}_X(U) \otimes_A M$ for any open set $U \subset X$. \[com04\] For a given $x \in X$, denote by $T^i_x$ the functor from the category of $\operatorname{\mathcal{O}}_{Y,f(x)}$-modules to finitely generated $\operatorname{\mathcal{O}}_{X,x}$-modules, which takes an $\operatorname{\mathcal{O}}_{Y,f(x)}$-module $M$ to ${\mathrm{Ext}}^i_{\operatorname{\mathcal{O}}_{X,x}}({\mathcal{F}}_x,\operatorname{\mathcal{O}}_{X,x} \otimes_{\operatorname{\mathcal{O}}_{Y,f(x)}} M)$. We say that $T^i$ is left exact (resp. right exact, exact) at some point $y_0 \in Y$ if for all points $x \in f^{-1}(y_0)$, $T^i_{x}$ is left exact (resp. right exact, exact) on the category of $\operatorname{\mathcal{O}}_{Y,y_0}$-modules. Given a complex $N^\bullet$, $$0 \to N^0 \xrightarrow{d_0} N^1 \xrightarrow{d_1} N^2 \xrightarrow{d_3} ...$$ denote by $W^i(N^\bullet):={\mathrm{coker}}(d_{i-1}:N^{i-1} \to N^{i})$. As $d_i^2=0$, we have a natural morphism $W^i(N^\bullet) \to N^{i+1}$. The kernel of this morphism is $H^i(N^\bullet)$. Fix a locally free resolution of ${\mathcal{F}}$, $$... \to {\mathcal{L}}_2 \to {\mathcal{L}}_1 \to {\mathcal{L}}_0 \to {\mathcal{F}} \to 0$$ Denote by $\operatorname{\mathcal{H}om}^\bullet$ the complex $$0 \to \operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_X}({\mathcal{L}}_0,\operatorname{\mathcal{O}}_X) \xrightarrow{d_0} \operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_X}({\mathcal{L}}_1,\operatorname{\mathcal{O}}_X) \xrightarrow{d_1} ...$$ Although the following Proposition is similar to [@R1 Proposition III.$12.7$], the proof in the reference does not hold in our setup as $W^i(\operatorname{\mathcal{H}om}^\bullet)_x$ is not a finitely generated $A$-module for any $x \in X$, which is used in an important step in the reference. \[comp4\] If $T^i$ is left exact (resp. right exact) at some point $y_0 \in Y$, then the same is true for all points $y$ in a suitable open neighbourhood $U$ of $y_0$. By Proposition \[comp7\], $T^i$ is left exact at $y_0$ if and only if $W^i(\operatorname{\mathcal{H}om}^\bullet)_x$ is $\operatorname{\mathcal{O}}_{Y,y_o}$-flat for all $x \in f^{-1}(y_0)$. By the open nature of flatness, there exists an open neighbourhood of $U$ containing $f^{-1}(y_0)$ such that $W^i(\operatorname{\mathcal{H}om}^\bullet)_u$ is $\operatorname{\mathcal{O}}_{Y,f(u)}$-flat for all $u \in U$. As $f$ is proper and $X \backslash U$ is closed, so is $f(X \backslash U)$. Denote by $V:=Y\backslash f(X \backslash U)$. Then, for all $v \in V$ and $v_x \in f^{-1}(v)$, $W^i(\operatorname{\mathcal{H}om}^\bullet)_{v_x}$ is $\operatorname{\mathcal{O}}_{Y,v}$-flat. Applying Proposition \[comp7\] once again, we conclude that $T^i$ is left-exact at $v$ for all $v \in V$. Given a short exact sequence $$0 \to M' \to M \to M'' \to 0$$ there exists an exact sequence by Lemma \[comp1\], $$T^i(M') \to T^i(M) \to T^i(M'') \to T^{i+1}(M') \to T^{+1}(M) \to T^{i+1}(M'').$$ Then, $T^i$ is right exact at a point $y$ if and only if $T^{i+1}$ is left exact at $y$. So, the second statement follows from the first applied to $T^{i+1}$. This proves the proposition. The following statement is similar to [@R1 Proposition III.$12.10$], but several steps in the proof given in the reference fails in our setup. As above this is because $T^0(M)_x$ is not a finitely generated $A$-module in our case (a fact used extensively in the reference). \[comp3\] Assume that for some $y \in Y$, the map $$\phi:T^0(A) \otimes k(y) \to T^0(k(y))$$ is surjective. Then, $T^0$ is right exact at $y$. By making a flat base extension, $\operatorname{\mathrm{Spec}}\operatorname{\mathcal{O}}_{Y,y} \to Y$ if necessary, we may assume that $y$ is a closed point of $Y$, $A$ is a local ring with maximal ideal $m$ and $k(y)=A/m$. By Proposition \[comp2\], it is sufficient to show that $\phi(M):T^0(A) \otimes M \to T^0(M)$ is surjective for all $A$-modules $M$. Since $T^0$ and tensor product commute with direct limits, it is sufficient to consider finitely generated $M$. First, we consider $A$-modules $M$ of finite length and we show that $\phi(M)$ is surjective, by induction on the length of $M$. If the length is $1$ then $M=k$ and $\phi(k)$ is surjective by hypothesis. Given a short exact sequence, $$0 \to M' \to M \to M'' \to 0,$$ of $A$-modules of finite length, we have $M'$ and $M''$ have length less than the length of $M$. As $T^0$ is exact in the middle (Lemma \[comp1\]), we have a commutative diagram with exact rows: $$\begin{diagram} T^0(A) \otimes M' &\rTo&T^0(A) \otimes M &\rTo&T^0(A) \otimes M''&\rTo&0\\ \dTo& &\dTo& &\dTo\\ T^0(M')&\rTo&T^0(M)&\rTo&T^0(M'') \end{diagram}$$ The two outside vertical arrows are surjective by the induction hypothesis, so the middle one is surjective also. Now let $M$ be any finitely generated $A$-module. For each $n$, $M/m^nM$ is a module of finte length, so by the previous case, $\phi_n:T^0(A) \otimes M/m^nM \to T^0(M/m^nM)$ and $T^0(A) \otimes (m^n/m^{n+1})M \to T^0((m^n/m^{n+1})M)$ are surjective. Denote by ${\mathcal{K}}_n$ the kernel of the induced morphism $\psi_n:T^0((m^n/m^{n+1})M) \to T^0(M/m^{n+1}M)$. Then, the morphism $\psi_n$ factors through $T^0((m^n/m^{n+1})M)/{\mathcal{K}}_n$. As $T^0$ is exact in the middle by Lemma \[comp1\], we have the following commutative diagram of exact sequences: $$\begin{diagram} & &T^0(A) \otimes m^n/m^{n+1}M &\rTo&T^0(A) \otimes M/m^{n+1}M &\rTo&T^0(A) \otimes M/m^nM&\rTo&0\\ & &\dOnto& &\dOnto^{\phi_{n+1}}& &\dOnto^{\phi_n}\\ 0&\rTo&\frac{T^0(m^n/m^{n+1}M)}{{\mathcal{K}}_n}&\rTo^{\psi_n}&T^0(M/m^{n+1}M)&\rTo&T^0(M/m^nM) \end{diagram}$$ Finally, using Snake lemma, this implies the natural morphism from $\ker(\phi_{n+1})$ to $\ker(\phi_n)$ is surjective. By [@R1 Example II.$9.1.1$] this implies $\ker(\phi_n)$ satisfies the Mittag-Leffler condition. Then, by [@R1 Proposition II.$9.1$], the map $$\varprojlim \phi_n: T^0(A) \otimes M^{\wedge} \to \varprojlim T^0(M/m^n M)$$ is surjective. Denote by $i:\hat{X} \to X$ the completion of $X$ with respect to $m$ (see [@R1 §II.$9$]). Now, $\operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_X}({\mathcal{F}},-)$ commutes with inverse limit i.e., $$\varprojlim T^0(M/m^nM)=\varprojlim \operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_X}({\mathcal{F}},\operatorname{\mathcal{O}}_X \otimes_A M/m^nM) \cong$$$$\cong \operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_X}({\mathcal{F}},\varprojlim \operatorname{\mathcal{O}}_X \otimes_A M/m^nM) \cong$$$$\cong \operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_X}({\mathcal{F}},i_(i^(\operatorname{\mathcal{O}}_X \otimes_A M))).$$By the adjunction property, $$\operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_X}({\mathcal{F}},i_(i^(\operatorname{\mathcal{O}}_X \otimes_A M))) \cong \operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_{\hat{X}}}(i^{\mathcal{F}},i^(\operatorname{\mathcal{O}}_X \otimes_A M))=\operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_{\hat{X}}}(\hat{{\mathcal{F}}},(\operatorname{\mathcal{O}}_X \otimes_A M)^\wedge)$$ which by EGA-III$_{\mathrm{I}}$ Proposition $12.3.5$, is isomorphic to $\operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_X}({\mathcal{F}},\operatorname{\mathcal{O}}_X \otimes_A M)^\wedge$ i.e., $$\varprojlim T^0(M/m^nM) \cong T^0(M)^\wedge.$$ In particular, the map $$\varprojlim \phi_n:\left(T^0(A) \otimes M\right)^\wedge \to \left(T^0(M)\right)^\wedge \mbox{ is surjective. }$$ Since completion is an exact, faithful functor for finitely generated $\operatorname{\mathcal{O}}_X$-modules (see [@R1 Theorem II.$9,3$A]), it follows that $$\phi(M):T^i(A) \otimes M \to T^i(M)$$ is surjective and we are done. Finally, using Propositions \[comp4\] and \[comp3\], we can prove Theorem \[comp6\]. As there exists $t \in Y$ such that the natural morphism $$\operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_X}({\mathcal{F}},\operatorname{\mathcal{O}}_X) \twoheadrightarrow \operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_X}({\mathcal{F}},\operatorname{\mathcal{O}}_X)\|_{X_t} \xrightarrow{\sim} \operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_X}({\mathcal{F}},\operatorname{\mathcal{O}}_{X_t})$$ is surjective, Proposition \[comp3\] implies $T^0$ is right exact at $t$. Then, Proposition \[comp4\] implies there exists an open neighbourhood $U$ containing $t$ such that for all $u \in U$, $T^0$ is right exact at $u$. By Proposition \[comp2\], this is equivalent to the morphism $$\operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_X}({\mathcal{F}},\operatorname{\mathcal{O}}_X)\|_{X_u} \to \operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_X}({\mathcal{F}},\operatorname{\mathcal{O}}_{X_u})$$ being an isomorphism. This proves the theorem. Alternating determinant of coherent sheaves =========================================== In this section, we give the definition of alternating determinant of coherent sheaves. This is the dual of the top wedge product of the coherent sheaf. We observe that the alternating determinant of a torsion-free sheaf on an integral scheme is a semi-stable rank one torsion-free sheaf (Proposition \[comp9\]). Moreover, if the underlying scheme is non-singular then the alternating determinant of a torsion-free sheaf is isomorphic to the dual of the determinant (Proposition \[comp12\]). Recall, the following definition: Let $E$ be a coherent sheaf of dimension $d$ on a projective scheme $X$. Write the Hilbert polynomial $P(E,m)$ of $E$ as follows: $$P(E,m)=\sum\limits_{i=0}^{d} \alpha_i(E) \frac{m^i}{i!}.$$ If $\dim X=d$ then, the rank of $E$ is defined as $${\mathrm{rk}}(E):=\frac{\alpha_d(E)}{\alpha_d(\operatorname{\mathcal{O}}_X)}$$ and the degree of $E$ is defined as $$\deg(E):=\alpha_{d-1}(E)-{\mathrm{rk}}(E).\alpha_{d-1}(\operatorname{\mathcal{O}}_X).$$ \[comp10\] Let $k$ be an algebraically closed field (of any characteristic), $X$ an integral $k$-scheme. Let $P$ be the Hilbert polynomial of a coherent sheaf on $X$ of rank $n$ with degree coprime to $n$. \[com05\] Let $R$ be a ring and $M$ a finitely generated $R$-module. For any integer $n>0$, denote by ${\mathrm{Alt}}^n_R(M,R)$, the $R$-submodule of ${\mathrm{Hom}}_R(M \otimes_R M \otimes_R ... \otimes_R M,R)$ consisting of alternating $R$-multilinear maps from $n$-copies of $M$ to $R$. Let $X$ be a projective scheme and ${\mathcal{F}}$ a pure coherent sheaf on $X$ of rank $n$. Let $\{U_i\}$ be an affine open covering of $X$. Define $\operatorname{\mathcal{A}lt}^n_{U_i}({\mathcal{F}})$, the coherent sheaf associated to the finitely generated $\operatorname{\mathcal{O}}_{U_i}(U_i)$-module ${\mathrm{Alt}}^n_{\operatorname{\mathcal{O}}_{U_i}(U_i)}({\mathcal{F}}(U_i))$. We call the the alternating determinant of ${\mathcal{F}}$, denoted $\operatorname{\mathcal{A}lt}^n({\mathcal{F}})$, the sheaf obtained via glueing $\operatorname{\mathcal{A}lt}_{U_i}^n({\mathcal{F}})$ \[comp11\] Let $X$ be an projective scheme, ${\mathcal{F}}$ a coherent pure sheaf on $X$. Then, for any positive integer $n$, the sheaf ${\mathcal{A}}lt^n({\mathcal{F}})$ is isomorphic (as $\operatorname{\mathcal{O}}_{X}$-modules) to $\operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_{X}}(\bigwedge^n {\mathcal{F}},\operatorname{\mathcal{O}}_{X})$. There is a natural surjective morphism $$i:{\mathcal{F}} \otimes_{\operatorname{\mathcal{O}}_X} {\mathcal{F}} \otimes_{\operatorname{\mathcal{O}}_X} ... \otimes_{\operatorname{\mathcal{O}}_X} {\mathcal{F}} (n-{\mathrm{copies}}) \to \bigwedge^n {\mathcal{F}}$$ defined on small enough open sets by $m_1 \otimes m_2 \otimes ... \otimes m_n \mapsto m_1 \wedge m_2 \wedge ... \wedge m_n$. This induces an injective morphism $$i^:\operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_X} (\bigwedge^n {\mathcal{F}},\operatorname{\mathcal{O}}_X) \to \operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_X}({\mathcal{F}} \otimes ... \otimes {\mathcal{F}},\operatorname{\mathcal{O}}_X).$$ Clearly, the image of $i^$ is contained in ${\mathcal{A}}lt^n({\mathcal{F}})$. It remain to prove that the image of $i^$ coincides with ${\mathcal{A}}lt^n({\mathcal{F}})$. It suffices to prove this on the level of stalks, which follows from the universal property of exterior powers: Indeed, for any $x \in X$ and an alternating $\operatorname{\mathcal{O}}_{X,x}$-multilinear $n$-form $\phi:{\mathcal{F}}_x \otimes {\mathcal{F}}_x \otimes ... \otimes {\mathcal{F}}_x \to \operatorname{\mathcal{O}}_{X,x}$, there exists an unique morphism $\psi:\bigwedge^n {\mathcal{F}}_x \to \operatorname{\mathcal{O}}_{X,x}$ such that $\psi \circ i_x=\phi$, where $i_x$ is the localization of the map $i$ at the point $x$. This proves the lemma. We see now that the alternating determinant of a torsion-free sheaf is semi-stable. \[comp9\] Let $X$ be an integral projective scheme and ${\mathcal{F}}$ a torsion-free sheaf on $X$ of rank $n$. Then, the alternating determinant $\operatorname{\mathcal{A}lt}^n({\mathcal{F}})$ is a rank $1$ semi-stable sheaf. As ${\mathcal{F}}$ is torsion-free, there exists a non-empty open, dense subset $U \subset X$ such that ${\mathcal{F}}\|_U$ is locally-free. Let ${\mathcal{G}}$ be a subsheaf of $\operatorname{\mathcal{A}lt}^n({\mathcal{F}})$. By Lemma \[comp11\], $\operatorname{\mathcal{A}lt}^n({\mathcal{F}})$ is torsion-free (dual of a coherent sheaf is torsion-free), hence so is ${\mathcal{G}}$. Denote by ${\mathcal{G}}'$ the cokernel $\operatorname{\mathcal{A}lt}^n({\mathcal{F}})/{\mathcal{G}}$. As ${\mathcal{F}}\|_U$ is locally free, $\operatorname{\mathcal{A}lt}^n({\mathcal{F}})\|_U$ is invertible, which means ${\mathcal{G}}'\|_U$ is supported in a codimension $1$ subscheme in $U$ (quotient of two rank $1$ torsion-free sheaves is torsion). Therefore, ${\mathcal{G}}'$ is supported in a codimension $1$ subscheme in $X$. This means, the Hilbert polynomial corresponding to ${\mathcal{G}}'$ is of degree at most dimension $\dim X-1$. Obviously, the leading coefficient of the Hilbert polynomial corresponding to ${\mathcal{G}}'$ is positive. It then follows from the definition of degree mentioned above that $\deg({\mathcal{G}}) \le \deg(\operatorname{\mathcal{A}lt}^n({\mathcal{F}}))$. As rank of $\operatorname{\mathcal{A}lt}^n({\mathcal{F}})$ and ${\mathcal{G}}$ are of rank $1$, $\operatorname{\mathcal{A}lt}^n({\mathcal{F}})$ is therefore semi-stable. This proves the proposition. Let $f:Y \to S$ be a flat morphism of schemes and ${\mathcal{E}}$ a coherent sheaf on $Y$ flat over $S$. We say that ${\mathcal{E}}$ is $S_r$ relative to* $f$ if the following holds: for each $x \in Y$, $s=f(x)$, we have $${\mathrm{depth}}_{\operatorname{\mathcal{O}}_{Y_s,x}}({\mathcal{E}}\|_{Y_s}) \ge \min(r, \dim \operatorname{\mathcal{O}}_{Y,x}-\dim \operatorname{\mathcal{O}}_{S,s}).$$ In other words, the restriction of ${\mathcal{E}}$ to each fiber is $S_r$. The following proposition tells us that the alternating determinant of a torsion-free sheaf on a smooth scheme coincides with the dual of its determinant. \[comp12\] Suppose that $X$ is a non-singular variety and $S$ is a $k$-scheme. For any coherent, torsion-free sheaf ${\mathcal{F}}_S$ of rank $n$ on $X_S$, flat over $S$, its alternating determinant $\operatorname{\mathcal{A}lt}^n({\mathcal{F}}_S)$ is isomorphic to $\det({\mathcal{F}}_S)^\vee$ (see [@huy §$1.1.17$] for the definition of determinant). Denote by $\pi:X_S \to S$ the natural projection map. Since $X$ is non-singular, $\pi$ is smooth. Hence, there exists a finite, locally free resolution of ${\mathcal{F}}_S$. Choose one such resolution, $$0 \to {\mathcal{L}}_m \xrightarrow{\phi_m} {\mathcal{L}}_{m-1} \xrightarrow{\phi_{m-1}} ... \xrightarrow{\phi_1} {\mathcal{L}}_0 \xrightarrow{\phi_0} {\mathcal{F}}_S \to 0.$$ Denote by $n_i$ the rank of ${\mathcal{L}}_i$ for $i=1,...,m$. By definition, $\det({\mathcal{F}})=\otimes_i \left(\bigwedge^{n_i} {\mathcal{L}}_i\right)^{(-1)^i}$. Consider now the short exact sequences: $$0 \to \ker \phi_i \to {\mathcal{L}}_i \xrightarrow{\phi_i} \ker \phi_{i-1} \to 0 \, \, \mbox{ for all } i \ge 1, \, \, \mbox{ and } 0 \to \ker \phi_0 \to {\mathcal{L}}_0 \to {\mathcal{F}}_S \to 0.$$ As ${\mathcal{L}}_i$ and ${\mathcal{F}}_S$ are flat over $S$, one can prove recursively that $\ker \phi_i$ is also flat over $S$ for all $i$. Since ${\mathcal{L}}_i$ are locally-free and ${\mathcal{F}}_S$ is torsion-free, [@stabhar Proposition $1.1$] implies $\ker \phi_0$ is $S_2$ relative to $\pi$. Hence, by recursion $\ker \phi_i$ is also $S_2$ relative to $\pi$ for all $i \ge 0$. By [@huy Lemma $2.1.8$], one can show that there exists an open set $U \subset X_S$ with $X_s\backslash U_s$ of codimension at least $2$ for all $s \in S$ and $\ker \phi_i\|_U$ and ${\mathcal{F}}_S\|_U$ are locally free of rank, say $n_i'$ and $n$, respectively. Then, restricting the above set of short exact sequences to $U$ we get, $$\left(\bigwedge^{n_i} {\mathcal{L}}_i\|_U\right) \cong \left(\bigwedge^{n_i'} \ker \phi_i\|_U\right) \otimes \left(\bigwedge^{n'_{i-1}} \ker \phi_{i-1}\|_U\right) \, \, \mbox{ for all } i>0 \mbox{ and }$$$$\left(\bigwedge^{n_0} {\mathcal{L}}_0\|_U\right) \cong \left(\bigwedge^{n_i'} \ker \phi_0\|_U\right) \otimes \left(\bigwedge^n {\mathcal{F}}_S\|_U\right).$$ As taking dual commutes with tensor product of locally free sheaves, we have $$\left(\bigwedge^{n_i} {\mathcal{L}}_i\|_U\right)^\vee \cong \left(\bigwedge^{n_i'} \ker \phi_i\|_U\right)^\vee \otimes \left(\bigwedge^{n'_{i-1}} \ker \phi_{i-1}\|_U\right)^\vee \, \, \mbox{ for all } i>0 \mbox{ and }$$$$\left(\bigwedge^{n_0} {\mathcal{L}}_0\|_U\right)^\vee \cong \left(\bigwedge^{n_i'} \ker \phi_0\|_U\right)^\vee \otimes \left(\bigwedge^n {\mathcal{F}}_S\|_U\right)^\vee.$$ This implies $$\left(\bigwedge^n {\mathcal{F}}_S\|_U\right)^\vee \cong \bigotimes_{i=0}^m \left(\bigwedge^{n_i} {\mathcal{L}}_i\|_U\right)^{(-1)^{i+1}}.$$ As $\operatorname{\mathcal{O}}_X$ is $S_2$ relative to $\pi$, we have $j_\operatorname{\mathcal{O}}_U \cong \operatorname{\mathcal{O}}_{X_S}$ (see [@hass Proposition $3.5$]) where $j:U \to X_S$ is the open immersion. Hence, by adjoint property of $\operatorname{\mathcal{H}om}$ and local freeness of ${\mathcal{F}}\|_U$, we have $$\operatorname{\mathcal{H}om}_{X_S}\left(\bigwedge^n {\mathcal{F}}_S,\operatorname{\mathcal{O}}_{X_S}\right) \cong j_\operatorname{\mathcal{H}om}_U\left(\bigwedge^n {\mathcal{F}}_S\|_U,\operatorname{\mathcal{O}}_U\right) \cong j_\left(\bigotimes_{i=0}^m \left(\bigwedge^{n_i} {\mathcal{L}}_i\|_U\right)^{(-1)^{i+1}}\right).$$ The projection formula implies that, $$j_\left(\bigotimes_{i=0}^m \left(\bigwedge^{n_i} {\mathcal{L}}_i\|_U\right)^{(-1)^{i+1}}\right) \cong \bigotimes_{i=0}^m \left(\bigwedge^{n_i} {\mathcal{L}}_i\right)^{(-1)^{i+1}}.$$ By Lemma \[comp11\], we finally conclude that ${\mathcal{A}}lt^n({\mathcal{F}}_S) \cong \det({\mathcal{F}}_S)^\vee$. This completes the proof of the proposition. Alternating determinant stratification on the moduli space ========================================================== In the introduction, we discussed the obstacle to defining an alternating determinant map on the entire moduli space of semi-stable sheaves. The aim of this section is to stratify the moduli space such that the alternating determinant map can be defined over each strata. This is done in the main theorem of this section (Theorem \[cn8\]). We use this in the next section to define the expected alternating determinant map. The key step is Proposition \[cn4\], which tells us that given any locally closed subscheme of the Quot-scheme, parametrizing semi-stable sheaves, we can find an open subscheme for which the alternating determinant functor commutes with pullback. We observe in Corollary \[comp18\] that under certain conditions, there exists universal geometric quotient of such an open subscheme. We combine this in Theorem \[cn8\] to give the required stratification and observe that on each strata the Hilbert polynomial of the alternating determinant, remains unchanged. We first recall certain standard results on moduli spaces of semi-stable sheaves. \[comp03\] Let $P$ be the Hilbert polynomial of a pure coherent sheaf on $X$ whose rank is coprime to its degree. We define a functor ${\mathcal{M}}_X(P)$ as follows: $${\mathcal{M}}_{X}(P) : {\mathrm{Sch}}/k \rightarrow {\mathrm{Sets}}$$ such that for a $k$-scheme $T$, $${\mathcal{M}}_{X}(P)(T):= \left\{ \begin{array}{l} \mbox{ isomorphism classes of } \mbox{ pure, coherent sheaves } {\mathcal{F}} \mbox{ on } X \times_k T\\ \mbox{ flat over } T \mbox{ and for every geometric point } t \in T, \, {\mathcal{F}}\|_{X_t} \mbox{ is a } \\ \mbox{ semi-stable sheaf with Hilbert Polynomial } P \mbox{ on } X_{t}. \end{array} \right\}$$ \[comp04\] Since $X$ is integral, [@huy Lemma $1.2.13$ and $1.2.14$] implies that a semi-stable sheaf on $X$ is also stable. Hence, the moduli functor ${\mathcal{M}}_X(P)$ coincides with the moduli functor defined by replacing in Defintion \[comp03\] the semi-stable condition by stable. Then, by [@langa Theorem $0.2$] the functor ${\mathcal{M}}_{X}(P)$ is universally corepresentable by a projective $k$-scheme $M_X(P)$. \[comp22\] Let ${\mathcal{F}}$ be a Giesekar semi-stable sheaf on $X$ with Hilbert polynomial $P$ and rank $n$. By [@huy Corollary $1.7.7$] there exists an integer $N_0$ depending only on $P$ such that ${\mathcal{F}}$ is $e$-regular for all $e \ge N_0$. Fix such an integer $e$. Denote by $V:=H^0({\mathcal{F}}(e))$ and ${\mathcal{H}}:=\operatorname{\mathcal{O}}_{X}(-e) \otimes_{k} V$. Denote by ${\mathrm{Quot}}_{{\mathcal{H}}/X/P}$ the scheme parametrizing all quotients of the form ${\mathcal{H}} \twoheadrightarrow {\mathcal{Q}}_0$, where ${\mathcal{Q}}_0$ has Hilbert polynomial $P$ with $H^0({\mathcal{Q}}_0(e))$ (non-canonically) isomorphic to $V$. Denote by ${\mathcal{Q}}$ the universal quotient on $X \times {\mathrm{Quot}}_{{\mathcal{H}}/X/P}$ associated to ${\mathrm{Quot}}_{{\mathcal{H}}/X/P}$. Denote by $R$ the subset of ${\mathrm{Quot}}_{{\mathcal{H}}/X/P}$ parametrizing coherent quotients of ${\mathcal{H}}$ which are semi-stable sheaves on $X$ with Hilbert polynomial $P$. By [@huy Proposition $2.3.1$] one notices that $R$ is an open subscheme in ${\mathrm{Quot}}_{{\mathcal{H}}/X/P}$. The group ${\mathrm{GL}}(V)={\mathrm{Aut}}({\mathcal{H}})$ acts on ${\mathrm{Quot}}_{{\mathcal{H}}/X/P}$ from the right by the composition $[\rho] \circ g=[\rho \circ g]$ for some $[\rho:{\mathcal{H}} \to {\mathcal{F}}] \mbox{ and } g \in {\mathrm{GL}}(V)$. By [@huy Theorem $4.3.3$] $R$ is the set of semi-stable points of ${\mathrm{Quot}}_{{\mathcal{H}}/X/P}$ under this group action. This induces a group action of ${\mathrm{GL}}(V)$ on $R$. By [@huy Lemma $4.3.1$], $M_X(P)$ is the geometric quotient of $R$ by this action. Denote by $$\pi:R \to M_X(P)$$ the corresponding geometric quotients. By [@huy Corollary $4.3.5$], the quotient $\pi$ is a ${\mathrm{PGL}}(V)$-bundle. \[comp14\] Given a morphism of finite type $f:Y \to Z$ between noetherian schemes and a coherent sheaf ${\mathcal{F}}$ on $Z$, there is a natural morphism from $f^\operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_Z}({\mathcal{F}},\operatorname{\mathcal{O}}_Z)$ to $\operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_Y}(f^{\mathcal{F}},\operatorname{\mathcal{O}}_Y)$. In particular, using Lemma \[comp11\], this induces a natural morphism from $f^\operatorname{\mathcal{A}lt}^n({\mathcal{F}})$ to $\operatorname{\mathcal{A}lt}^n(f^{\mathcal{F}})$, where $n={\mathrm{rk}}({\mathcal{F}})$. Denote by ${\mathcal{H}}_{R}$ the pullback of the sheaf ${\mathcal{H}}$ under the natural projection map $\operatorname{\mathrm{pr}}_1:X \times_k R \to X$. Recall, we have the universal quotient:$${\mathcal{H}}_{R} \to {\mathcal{Q}}\|_{X \times R} \to 0.$$ Taking the wedge powers, we get the surjective morphism $$\phi_n:\bigwedge^n {\mathcal{H}}_{R} \to \bigwedge^n {\mathcal{Q}}\|_{X \times R}.$$ Denote by ${\mathcal{G}}_n$ the kernel of the morphism $\phi_n$, by $Q_n:=\bigwedge^n {\mathcal{Q}}\|_{X \times R}$ and ${\mathcal{H}}_n:=\bigwedge^n {\mathcal{H}}_{R}$. The following proposition can be formulated more generally in terms of families of coherent sheaves parametrized by a noetherian scheme, however for simplicity, we restrict to the setup relevant for this article. \[cn4\] Let $B$ be a locally closed subscheme of $R$. Then, there exists an open dense subscheme $U_B$ of $B$ such that for all $u \in U_B$, the induced morphism $$\label{cn3} \operatorname{\mathcal{A}lt}^n({\mathcal{Q}}\|_{X_B}) \otimes k(u) \to \operatorname{\mathcal{A}lt}^n({\mathcal{Q}}\|_{X_u}) \, \mbox{ is an isomorphism.}$$ For simplicity, we abuse the notations to denote by ${\mathcal{Q}}_n, {\mathcal{H}}_n$ and ${\mathcal{G}}_n$ the restrictions ${\mathcal{Q}}_n\|_{X_B}, {\mathcal{H}}_n\|_{X_B}$ and ${\mathcal{G}}_n\|_{X_B}$, respectively. Since ${\mathcal{H}}_n$ is a $\operatorname{\mathcal{O}}_{X_B}$-free sheaf, we have the following $\operatorname{\mathcal{H}om}$-exact sequence: $$\label{cn2} 0 \to \left({\mathcal{Q}}_n\right)^\vee \to {\mathcal{H}}_n^\vee \to ({\mathcal{G}}_n)^\vee \xrightarrow{\phi_B} {\mathcal{E}}xt^1_{X_B}\left({\mathcal{Q}}_n,\operatorname{\mathcal{O}}_{X_B}\right) \to 0.$$ Denote by $U$ the open dense subscheme such that $(\ker \phi_B)\|_{X_U}$ and ${\mathcal{E}}xt^1_{X_B}\left({\mathcal{Q}}_n,\operatorname{\mathcal{O}}_{X_B}\right)\|_{X_U}$ are flat over $U$. The existence of such an open set follows from [@gortz Theorem $10.84$, pp. $271$]. Since ${\mathcal{H}}_n\|_{X_U}$ is already flat over $U$, by [@R1 Proposition III.$9.1$A], we have ${\mathcal{G}}_n^\vee\|_{X_U}$ and ${\mathcal{Q}}_n^\vee\|_{X_U}$ are also flat over $U$. Therefore, for each $u \in U$, we have a long exact sequence: $$0 \to \left({\mathcal{Q}}_n\right)^\vee\|_{X_u} \to {\mathcal{H}}_n^\vee\|_{X_u} \to ({\mathcal{G}}_n)^\vee\|_{X_u} \xrightarrow{\phi_B} {\mathcal{E}}xt^1_{X_B}\left({\mathcal{Q}}_n,\operatorname{\mathcal{O}}_{X_B}\right)\|_{X_u} \to 0.$$ Since ${\mathcal{G}}_n$ is a coherent sheaf, it can be written as a quotient of a locally free sheaf, say ${\mathcal{L}}_0$ on $X_B$ (use [@R1 Corollary II.$5.18$]). Denote by $j:{\mathcal{L}}_0 \twoheadrightarrow {\mathcal{G}}_n$ the corresponding quotient map. Arguing as before, there exists an open dense subset $V$ of $B$ such that for each $v \in V$, we have the exact sequence: $$0 \to ({\mathcal{G}}_n)^\vee\|_{X_v} \to ({\mathcal{L}}_0^\vee)\|_{X_v} \to (\ker j)^\vee\|_{X_u} \to {\mathcal{E}}xt^1_{X_B}\left({\mathcal{G}}_n,\operatorname{\mathcal{O}}_{X_B}\right)\|_{X_v} \to 0.$$ Since ${\mathcal{L}}_0$ is locally-free, $({\mathcal{L}}_0^\vee)\|_{X_v} \cong ({\mathcal{L}}_0\|_{X_v})^\vee$, which gives us the following commutative diagram of exact sequences: $$\begin{diagram} 0 &\rTo &({\mathcal{G}}_n)^\vee\|_{X_v} &\rTo &({\mathcal{L}}_0^\vee)\|_{X_v} \\ & &\dTo^{\phi_3}& \circlearrowright & \dTo^{\cong}\\ 0 &\rTo &\left({\mathcal{G}}_n\|_{X_v}\right)^\vee &\rTo &({\mathcal{L}}_0\|_{X_v})^\vee \end{diagram}$$ This implies that the morphism $\phi_3$ is injective. Therefore, for any point $u \in U \cap V$, we have the following diagram of exact sequences: $$\begin{diagram} 0 &\rTo &\left({\mathcal{Q}}_n\right)^\vee\|_{X_u} &\rTo &({\mathcal{H}}_n^\vee)\|_{X_u} &\rTo & ({\mathcal{G}}_n)^\vee\|_{X_u} & \rTo^{\phi_B} & {\mathcal{E}}xt^1_{X_B}\left({\mathcal{Q}}_n,\operatorname{\mathcal{O}}_{X_B}\right)\|_{X_u}& \rTo &0\\ & &\dTo^{\phi_1}& \circlearrowright & \dTo^{\phi_2}&\circlearrowright&\dTo^{\phi_3}&\circlearrowright&\dTo\\ 0 &\rTo &\left({\mathcal{Q}}_n\|_{X_u}\right)^\vee &\rTo &({\mathcal{H}}_n\|_{X_u})^\vee &\rTo & ({\mathcal{G}}_n\|_{X_u})^\vee & \rTo^{\phi_u} & {\mathcal{E}}xt^1_{X_u}\left({\mathcal{Q}}_n\|_{X_u},\operatorname{\mathcal{O}}_{X_u}\right)& \rTo &0\\ \end{diagram}$$ Since ${\mathcal{H}}_n$ is $\operatorname{\mathcal{O}}_{X_B}$-free, $\phi_2$ is an isomorphism. Since $\phi_3$ is injective, one uses the short five lemma or diagram chase to conclude that $\phi_1$ is an isomorphism. Since $U \cap V$ is dense in $B$, we have the proposition. \[cn5\] Let $B$ be as in Proposition \[cn4\]. Using Theorem \[comp6\] and Proposition \[cn4\], there exists an open dense subscheme $U_B$ of $B$ such that $u \in U_B$ if and only if $u \in B$ satisfies the property (\[cn3\]) stated in Proposition \[cn4\]. \[com22\] Suppose the ${\mathrm{PGL}}(V)$-bundle $\pi:R \to M_X(P)$ descends to $B$ in the sense that there exists a locally closed subscheme $B' \subset M_X(P)$ such that $\pi\|_B:B \to B'$ is a ${\mathrm{PGL}}(V)$-bundle. Then, for any geometric point $x_0 \in B', \pi^{-1}(x_0) \cap U_B \not= \emptyset$ if and only if $\pi^{-1}(x_0) \subset U_B$. By the definiton of geometric good quotient, $x_0$ corresponds to a semi-stable coherent sheaf ${\mathcal{F}}_0$ on $X$ with Hilbert polynomial $P$. Since $\pi\|_B$ is a ${\mathrm{PGL}}(V)$-bundle, $\pi\|_B^{-1}(x_0)=\pi^{-1}(x_0)$ which equals to quotients of the form $[{\mathcal{H}} \stackrel{\phi}{\cong} {\mathcal{H}} \twoheadrightarrow {\mathcal{F}}_0]$ as $\phi$ runs through all the automorphisms of ${\mathcal{H}}$. A quotient $[{\mathcal{H}} \twoheadrightarrow {\mathcal{F}}_0]$ is in $U_B$ if and only if it satisfies the relation (\[cn3\]). It then follows directly that $\pi^{-1}(x_0) \cap U_B \not= \emptyset$ if and only if $\pi^{-1}(x_0) \subset U_B$. This proves the lemma. \[comp05\] Hypothesis as in Lemma \[com22\]. For simplicity, we denote by $\pi$ the restriction $\pi\|_B:B \to B'$. The scheme $\pi^{-1}(\pi(B\backslash U_B))$ does not intersect $U_B$, where $\pi(B\backslash U_B)$ denotes the scheme-theoretic image of the closed subscheme $B\backslash U_B$. Denote by $V:=\pi^{-1}(\pi(B\backslash U_B))$. Suppose that $V \cap U_B \not= \emptyset$. Since $V \cap U_B$ is a non-empty quasi-compact scheme, it contains a closed point, say $x_0$. As the geometric quotient morphism $\pi$ is of finite type, $\pi(x_0)$ is closed. This means in particular that there exists a closed point $\pi(x_0) \in B'$ such that $\pi^{-1}(\pi(x_0))$ intersects $U_B$ but is not completely contained in $U_B$. But this contradicts Lemma \[com22\]. This proves the corollary. \[cn7\] Hypothesis as in Lemma \[com22\]. Denote by $V_B:=B' \backslash (\pi\|_B(B \backslash U_B))$. Observe that $V_B$ is an open subscheme in $B'$. This directly implies: \[comp25\] The scheme $U_B$ is isomorphic (as schemes) to $V_B \times_{B'} B \cong \pi^{-1}(V_B)$. By Corollary \[comp05\], $U_B$ is contained in $\pi^{-1}(V_B)$. We now prove that $U_B$ coincides with $\pi^{-1}(V_B)$. Suppose this is not the case. This means that there exists a closed point $x_0 \in \pi^{-1}(V_B)$ not in $U_B$. Since the morphism $\pi$ is of finite-type, $\pi(x_0)$ is closed. In other words, there exists a closed point $y_0 \in V_B$ such that the fiber $\pi^{-1}(y_0)$ is not entirely contained in $U_B$. But this contradicts Lemma \[com22\], hence proves the corollary. \[comp18\] The natural morphism $\pi\|_{U_B}$ factors through $V_B$. Furthermore, the morphism $\pi:U_B \to V_B$, makes $U_B$ into a ${\mathrm{PGL}}(V)$-bundle over $V_B$. Since pullback of a ${\mathrm{PGL}}(V)$-bundle is again a ${\mathrm{PGL}}(V)$-bundle, Corollary \[comp25\] implies this corollary. We can now define a stratification on $R$ induced by the alternating determinant: \[cn13\] We follow notations as in Notations \[cn5\]. Denote by $R_1:=U_R$, $R_1':=R\backslash R_1, R_2:=U_{R_1'}, R_2':=R_1'\backslash R_2$ and inductively denote by $R_i:=U_{R'_{i-1}}$ and $R'_i:=R'_{i-1}\backslash R_i$. Set $R'_0:=R$. \[cn8\] The following holds true: 1. there exists an integer $i_0$ such that for all $i \ge i_0$, $R_i= \emptyset$, 2. the set of locally closed subschemes, $\{R_i\}$ defines a stratification of $R$ i.e., $R=\bigcup\limits_{i=1}^{i_0} R_i$ and $R_i \cap R_j = \emptyset$ for $i \not= j$, 3. the stratification $\{R_i\}$ induces a stratification $\{V_{R_i}\}$ on $M_X(P)$ (notations as in Notation \[cn7\]) and for each $i$, $\pi_i:R_i \to V_{R_i}$ is a ${\mathrm{PGL}}(V)$-bundle morphism. 4. for a fixed $i$, there exists a Hilbert polynomial $L_i$ (depending only on $i$) such that for every $u \in R_i$ the corresponding quotient $[{\mathcal{H}}_u \twoheadrightarrow {\mathcal{Q}}_u]$ satisfies the property: $\operatorname{\mathcal{A}lt}^n({\mathcal{Q}}_u)$ has Hilbert polynomial $L_i$, Using Proposition \[cn4\], observe that $$\dim R=\dim R_1>\dim R_1'=\dim R_2, \, \mbox{ and } \, \dim R'_{i-1}=\dim R_i>\dim R'_i=\dim R_{i+1}.$$ Since $R$ is finite dimensional there exists an integer $i_0$ such that for all $i \ge i_0$, $R_i= \emptyset$, which proves $(1)$. Then, by construction, $(2)$ follows. To prove $(3)$, we need to prove that for each $i$, $\pi$ descends to $R'_i$ in the sense of Lemma \[com22\]. Then, by Corollary \[comp18\], $\pi$ induces a ${\mathrm{PGL}}(V)$-bundle morphism $\pi_{i+1}:R_{i+1} \to V_{R'_i}$. Denoting $V_{R'_i}$ by $V_{R_{i+1}}$, we will have $(3)$. We prove that $R'_i$ satisfies the hypothesis of Lemma \[com22\], inductively. Trivially, $R'_0=R$ satisfies the hypothesis. Suppose there exists $m$ such that for all $i<m$, $R'_i$ satisfies the hypothesis. This means in particular, by Corollaries \[comp25\] and \[comp18\] that for all $i \le m$, $\pi$ induces, by restriction, a ${\mathrm{PGL}}(V)$-bundles $\pi_i:R_{i} \to V_{R_{i}}$ and $R_{i}=\pi^{-1}(V_{R_{i}})$. Denote by $V'_{R_m}:=M_{X}(P)\backslash (V_{R_1} \coprod V_{R_2} \coprod ... \coprod V_{R_{m}})$. One can think of $V'_{R_m}$ as the locally closed subscheme obtained by recursiverly removing from $M_X(P)\backslash \left(V_{R_1} \coprod ... \coprod V_{R_{i-1}}\right)$ the open subscheme $V_{R_i}$, for $i \le m$. We then have $$\pi^{-1}(V'_{R_m})=R\backslash\left(\pi^{-1}(V_{R_1}) \coprod ... \coprod \pi^{-1}(V_{R_{m}})\right)=R\backslash\left(R_1 \coprod ... \coprod R_{m}\right)=R'_{m}.$$ Since base change of ${\mathrm{PGL}}(V)$-bundle is again a ${\mathrm{PGL}}(V)$-bundle, we have $\pi:R'_{m} \to V'_{R_m}$ is a ${\mathrm{PGL}}(V)$-bundle. This proves the induction step hence $(3)$. We now prove $(4)$. Since Hilbert polynomial remains unchanged in flat families, we are going to prove that the sheaf $\operatorname{\mathcal{A}lt}^n({\mathcal{Q}}\|_{X_{R'_{i-1}}})\|_{X_{R_{i}}}$ is flat over $R_{i}$, which will directly imply part $(4)$ of the theorem. As flatness is a local property, it suffices to prove the statement for any affine open subscheme $\operatorname{\mathrm{Spec}}A$ of $R_i$. It then suffices to prove for any short exact sequence of $A$-modules, $$\label{comp15} 0 \to M' \to M \to M'' \to 0$$ the functor $- \otimes_A (\operatorname{\mathcal{A}lt}^n({\mathcal{Q}}\|_{X_{R'_{i-1}}})\|_{X \times A})$ is exact. Consider the functor $$T:\mathfrak{M}_A \to {\mathrm{Coh}}_{X \times_k A} \, \mbox{ defined by } \, M \mapsto \operatorname{\mathcal{H}om}_{X \times A}\left(\bigwedge^n {\mathcal{Q}}\|_{X \times A}, \operatorname{\mathcal{O}}_{X \times A} \otimes_A M\right)$$ and the natural map $\phi_M: T(A) \otimes_A M \to T(M)$. As $\operatorname{\mathrm{Spec}}A \subset R'_{i-1}$ is open, $$T(A)=\operatorname{\mathcal{A}lt}^n({\mathcal{Q}}\|_{X \times A}) \cong \operatorname{\mathcal{A}lt}^n({\mathcal{Q}}\|_{X \times R_{i-1}'})\|_{X \times A}.$$ Hence, we need to prove that $T(A)$ is $A$-flat. Since $\operatorname{\mathcal{H}om}_{X \times A}(\bigwedge^n {\mathcal{Q}}\|_{X \times A}, \operatorname{\mathcal{O}}_{X \times A} \otimes -)$ is left-exact, we have the following diagram: $$\begin{diagram} & &T(A) \otimes_A M' &\rTo^{\rho} &T(A) \otimes_A M&\rTo&T(A) \otimes M'' &\rTo& 0\\ & &\dTo_{\phi_{M'}}&\circlearrowleft&\dTo_{\phi_M}&\circlearrowleft&\dTo_{\phi_{M''}}\\ 0&\rTo&T(M')&\rTo&T(M)&\rTo&T(M'')& \end{diagram}$$ By assumption, for all $u \in \operatorname{\mathrm{Spec}}(A)$, $\phi_{k(u)}$ is surjective, where $k(u)$ is the residue field corresponding to $u$. Then, by Proposition \[comp3\], $\phi_L$ is surjective for any $A$-module $L$. Proposition \[comp2\] then implies $\phi_L$ is an isomorphism for all $A$-modules $L$. The commutative diagram then implies that $\rho$ is injective. Hence, $T(A)$ is $A$-flat. This completes the proof of $(4)$. Alternating determinant map on the stratification ================================================= In this section, we finally give the alternating determinant map on the stratification of the moduli space (Theorem \[comp16\]). We observe that if the underlying scheme is smooth, then there exists no non-trivial strata and the alternating determinant map is simply the dual of the determinant map (Corollary \[comp19\]). \[comp16\] For the stratification $\{V_{R_i}\}$ of $M_X(P)$ as in Theorem \[cn8\], there exists a natural alternating determinant map: $${\mathrm{Adet}}:\coprod_i V_{R_i} \to \coprod_i M_X(L_i)$$ which to a geometric point $s \in M_X(P)$ associates its alternating determinant, where $L_i$ is as in Theorem \[cn8\]$(4)$. We use the notations as in [@huy §$4.2$]. By Theorem \[cn8\]$(3)$, $R_i$ is a ${\mathrm{PGL}}(V)$-bundle over $V_{R_i}$. Since ${\mathrm{PGL}}(V)$ bundles are universal categorical quotients, the functor $\underline{R_i}/\underline{{\mathrm{GL}}(V)}$ is corepresented by $\underline{V_{R_i}}$. By Theorem \[cn8\], there exists a natural tranformation from $$\underline{R_i}/\underline{{\mathrm{GL}}(V)} \to {\mathcal{M}}_X(L_i) \mbox{ defined by } [{\mathcal{H}} \to {\mathcal{E}}] \mapsto \operatorname{\mathcal{A}lt}^n({\mathcal{E}}).$$ By the universal property of corepresentation, there exists a natural transformation $F_i$ from $\underline{V_{R_i}}$ to $\underline{M_X(L_i)}$ such that the following diagram commute, $$\begin{diagram} \underline{R_i}/\underline{{\mathrm{GL}}(V)}&\rTo&\underline{V_{R_i}}\\ \dTo&\circlearrowleft&\dTo^{F_i}\\ {\mathcal{M}}_X(L_i)&\rTo&\underline{M_X(L_i)} \end{diagram}$$ By Yoneda lemma, there exists an unique morphism ${\mathrm{Adet}}\|_{V_{R_i}}:V_{R_i} \to M_X(L_i)$ which gives rise to the natural transformation $F_i$. This proves the theorem. We finally observe that in the classical case, when the underlying scheme is smooth, we get an alternating determinant map on the entire moduli space and not just on its stratification: \[comp19\] If $X$ is a smooth, projective variety then there exists no non-trivial stratification of $M_X(P)$. In particular, there exists an alternating determinant map ${\mathrm{Adet}}:M_X(P) \to M_X(L)$, which to a geometric point $s \in M_X(P)$ associates its alternating determinant. Since dual of locally free sheaves commutes with pull-back, Proposition \[comp12\] implies that for all $s \in R$, $$\operatorname{\mathcal{A}lt}^n({\mathcal{Q}})\|_{X_s} \cong \det({\mathcal{Q}})^\vee\|_{X_s} \cong \det({\mathcal{Q}}\|_{X_s})^\vee \cong \operatorname{\mathcal{A}lt}^n({\mathcal{Q}}\|_{X_s}).$$ This proves that $R_1=R$, which is the first part of the corollary. The second part follows identically as in the proof of Theorem \[comp16\] above. Generalities ============ Let $f:X \to Y$ be a proper, flat surjective morphism between noetherian schemes. Assume $Y={\mathrm{Spec}}(A)$ for some noetherian ring $A$. Let ${\mathcal{F}}$ be a coherent sheaf on $X$ (not necessarily flat over $Y$). We use the same notations as in Definitions \[com03\] and \[com04\]. \[comp1\] Each $T^i$ is an additive covariant functor from $A$-modules to itself which is exact in the middle. The collection $(T^i)_{i \ge 0}$ forms a $\delta$-functor. Furthermore, $T^i$ commutes with direct limit. Clearly, $T^i$ is an additive covariant $\delta$-functor. Using [@R1 III. Proposition $6.4$], we can conclude ${\mathcal{E}}xt^i_{\operatorname{\mathcal{O}}_X}({\mathcal{F}},- \otimes_A M)$ is exact in the middle. As ${\mathcal{F}}$ is a coherent sheaf, by [@R1 Corollary II.$5.18$], there exists a resolution of ${\mathcal{F}}$ by finite locally free sheaves ${\mathcal{L}}_i$: $$... \to {\mathcal{L}}_2 \to {\mathcal{L}}_1 \to {\mathcal{L}}_0 \to {\mathcal{F}} \to 0.$$ Hence, [@breaz Theorem $1.1$] implies ${\mathcal{E}}xt^i_{\operatorname{\mathcal{O}}_X}({\mathcal{F}},-)$ commutes with direct limit. This completes the proof of the lemma. The proof of Propositions \[comp7\] and and \[comp2\] below are very similar to the proofs [@R1 Proposition III.$12.4, 12.5$] but we reproduce them as our initial hypothesis is different from the ones in the reference. They will be used in the proof of Theorem \[comp6\]. \[comp7\] Let $x \in X$. The following are equivalent: 1. $T^i_x$ is left exact 2. $W^i(\operatorname{\mathcal{H}om}^\bullet)_x$ is $\operatorname{\mathcal{O}}_{Y,f(x)}$-flat, where $W^i(\operatorname{\mathcal{H}om}^\bullet)_x$ is the stalk of the sheaf at $x$. Given any injective morphism $0 \to M \to M'$ of $\operatorname{\mathcal{O}}_{Y,f(x)}$-modules, it suffices to prove that the induced morphism $T^i_x(M) \to T^i_x(M')$ is injective if and only if so is $W^i(\operatorname{\mathcal{H}om}^\bullet)_x \otimes M \to W^i(\operatorname{\mathcal{H}om}^\bullet)_x \otimes M'$. Indeed, this is due to the property of flatness that $W^i(\operatorname{\mathcal{H}om}^\bullet)_x$ is $\operatorname{\mathcal{O}}_{Y,f(x)}$-flat if and only if for all injective morphism $0 \to M \to M'$ of $\operatorname{\mathcal{O}}_{Y,f(x)}$-modules, the induced morphism $W^i(\operatorname{\mathcal{H}om}^\bullet)_x \otimes M \to W^i(\operatorname{\mathcal{H}om}^\bullet)_x \otimes M'$ is injective. Denote by $\operatorname{\mathcal{H}om}_M^\bullet$ (resp. $\operatorname{\mathcal{H}om}_{M'}^\bullet$) the complexes $$0 \to \operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_X}({\mathcal{L}}_0,\operatorname{\mathcal{O}}_X \otimes_A M) \to \operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_X}({\mathcal{L}}_1,\operatorname{\mathcal{O}}_X \otimes_A M) \to ....$$ $$(\mbox{resp}. \, \, \, 0 \to \operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_X}({\mathcal{L}}_0,\operatorname{\mathcal{O}}_X \otimes_A M') \to \operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_X}({\mathcal{L}}_1,\operatorname{\mathcal{O}}_X \otimes_A M') \to .... )$$ Using Definition \[com04\], we have an exact sequence: $$0 \to H^i(\operatorname{\mathcal{H}om}_M^\bullet) \to W^i(\operatorname{\mathcal{H}om}_M^\bullet)_x \xrightarrow{d_i} \operatorname{\mathcal{H}om}({\mathcal{L}}_{i+1},\operatorname{\mathcal{O}}_X \otimes_A M)_x$$ and $H^i(\operatorname{\mathcal{H}om}^\bullet_M) \cong mr{Ext}^i_{\operatorname{\mathcal{O}}_{X,x}}({\mathcal{F}}_x,M \otimes_{\operatorname{\mathcal{O}}_{Y,f(x)}} \operatorname{\mathcal{O}}_{X,x})$. Since $- \otimes_{\operatorname{\mathcal{O}}_{Y,f(x)}} M$ is right exact and $\operatorname{\mathcal{H}om}({\mathcal{L}}_j,M \otimes_A \operatorname{\mathcal{O}}_X)_x \cong \operatorname{\mathcal{H}om}({\mathcal{L}}_j,\operatorname{\mathcal{O}}_X)_x \otimes_{\operatorname{\mathcal{O}}_{Y,f(x)}} M$ for all $j$, one can easily check that $$W^i(\operatorname{\mathcal{H}om}^\bullet_M)_x \cong W^i(\operatorname{\mathcal{H}om}^{\bullet})_x \otimes_A M.$$ Similarly for the $\operatorname{\mathcal{O}}_{Y,f(x)}$-module $M'$. Therefore, we have the following diagram: $$\begin{diagram} & & & & & &0\\ & & & & & &\dTo\\ 0&\rTo&T^i_x(M')&\rTo&W^i(\operatorname{\mathcal{H}om}^\bullet)_x \otimes M'&\rTo&\operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_X}({\mathcal{L}}_{i+1},\operatorname{\mathcal{O}}_X)_x \otimes M'\\ & &\dTo^{\alpha}& &\dTo^{\beta}& &\dTo\\ 0&\rTo&T^i_x(M)&\rTo&W^i(\operatorname{\mathcal{H}om}^\bullet)_x \otimes M&\rTo&\operatorname{\mathcal{H}om}_{\operatorname{\mathcal{O}}_X}({\mathcal{L}}_{i+1},\operatorname{\mathcal{O}}_X)_x \otimes M\\ \end{diagram}$$ The third vertical arrow is injective as ${\mathcal{L}}_{i+1}$ is a locally-free sheaf, which implies ${\mathcal{L}}_{i+1}^\vee$ is $\operatorname{\mathcal{O}}_X$-flat and hence $A$-flat. A simple diagram chase shows us that $\alpha$ is injective if and only if so is $\beta$. This completes the proof of the proposition. \[comp2\] For any $A$-module $M$, there is a natural map $$\phi:T^i(A) \otimes M \to T^i(M).$$Furthermore, the following conditions are equivalent: 1. $T^i$ is right exact 2. $\phi$ is an isomorphism for all $M$ 3. $\phi$ is surjective for all $M$ Since $T^i$ is a functor, we have a natural map for any $M$:$$M={\mathrm{Hom}}(A,M) \xrightarrow{\psi} \operatorname{\mathrm{Hom}}(T^i(A),T^i(M)).$$ This gives $\phi$, by setting $\phi(\sum a_j \otimes m_j)=\sum \psi(m_j)a_j$ where $m_j \in M$ and $a_j \in T^i(A)$. Since $T^i$ and $\otimes$ commute with direct limits (see Lemma \[comp1\]), it will be sufficient to consider only finitely generated $A$-modules (every module can be written as a direct limit of finitely generated modules). Let $A^r \to A^s \to M \to 0$ be an exact sequence. Then we have a diagram: $$\begin{diagram} T^i(A) \otimes A^r &\rTo&T^i(A) \otimes A^s&\rTo&T^i(A) \otimes M&\rTo&0\\ \dTo& &\dTo& &\dTo^{\phi}\\ T^i(A^r) &\rTo&T^i(A^s)&\rTo&T^i(M) \end{diagram}$$ where the bottom row is not necessarily exact. The first two vertical arrows are isomorphisms. Thus, if $T^i$ is right exact, then $\phi$ is an isomorphism. This proves $(1) \Rightarrow (2)$. The implication $(2) \Rightarrow (3)$ is obvious, so we have only to prove $(3) \Rightarrow (1)$. We must show if $$0 \to M' \to M \to M'' \to 0$$ is an exact sequence of $A$-modules then $$T^i(M') \to T^i(M) \to T^i(M'') \to 0$$is exact. By Lemma \[comp1\], it is exact in the middle, so we have only to show that $T^i(M) \to T^i(M'')$ is surjective. This follows from the diagram $$\begin{diagram} T^i(A) \otimes M &\rTo&T^i(A) \otimes M''&\rTo&0\\ \dTo^{\phi(M)}& &\dTo^{\phi(M'')}\\ T^i(M)&\rTo&T^i(M'') \end{diagram}$$ and the fact that $\phi(M'')$ is surjective.	{ "pile_set_name": "ArXiv" }
--- author: - 'T. <span style="font-variant:small-caps;">Matsui</span>' title: \| Deciphering the Measured Ratios of Iodine-131 to\ Cesium-137 at the Fukushima Reactors --- 1\. [Introduction]{} The current state of the troubled Fukushima Dai-ichi nuclear power plant is of considerable concern. The plant, on the east shore of Fukushima prefecture was first hit, at 14:46 pm local time on March 11, 2011, by the magnitude 9.0 earthquake on the Richter scale about 150 km east of Honshu, the main island of Japan, and then hit by a giant tsunami, exceeding 14 m high, about one hour after the earthquake. Although three reactors (Units 1, 2, 3), which had been on operation at the plant, were automatically shut down immediately after the earthquake, the emergent electric power generators, working after the loss of external power supply, for cooling the reactor system were knocked out by the tsunami. Uncontrolled decay heat from the fuel rods has caused serious trouble in containing the radioactive materials which had been accumulated in the reactors and in the spent-fuel rods in the cooling pools. Explosions of hydrogen gas, created by the reaction of the heated Zirconium fuel cladding with water vapor, took place at the buildings of the Units 1, 3, and 4 reactors, severely damaging those buildings. At the moment, only very limited information is available about the interior conditions of the reactor complex owing to the highly radioactive environment, with intense gamma rays from radioactive materials that apparently escaped from the reactors preventing direct human access for inspections. However, certain information has been released by the Tokyo Electric Power Company (TEPCO) [@TEPCO] on measurements of the radioactivity at several monitoring posts near the troubled reactors and from a water sample taken from the spent-fuel cooling pool of the Unit-4 reactor building.[@site4] We call attention here to the significance of the relative strength of the reported radioactivities of different radioactive isotopes. In particular, we focus on the two radioactive isotopes abundantly produced by fission, the Iodine isotope I-131 and the Cesium isotope Cs-137; the former has the half-life of 8 days while for the latter it is 30 years. The ratio should decrease on the time scale of days after the termination of nuclear fission processes and hence may be used for measuring the age of the fission products, similar to the carbon dating method using the ratio of C-14 to C-12 in the remains of ancient life[@carbon]. One crucial difference between the proposed iodine-cesium dating and carbon dating is that while the initial condition of the latter is essentially fixed by the equilibrium C-14/C-12 ratio in atmosphere with respect to cosmic ray interactions, the initial condition of the former depends on the prehistory of controlled nuclear reactions in the reactor. Another potential problem, which does not exist for carbon dating, is the effect of the different chemical properties of iodine and cesium[^1] possibly reflected in the change of their relative abundance depending on the environment in which they are found, for example in the water (fresh or sea), in the air, or in aerosols[@chemistry]. With these caveats in mind, we make a simple estimate of time-dependent numbers of I-131 and Cs-137 nuclei produced from a uranium burning nuclear reactor and then examine the data of the water samples released by TEPCO taken at several monitoring posts near the troubled Fukushima nuclear reactor plant, including sea water at the Southern Discharge Channel, Unit-4 cooling pool, Sub-drain (under ground) water, and sea water taken at the Intake Canal. Some of these data show anomalously high iodine-cesium ratios compared to the values expected for the nuclear fuels in the reactors if the nuclear fission has indeed terminated before the outbreak of the accident. We draw tentative conclusions from these analyses of the data at the end.\ 2. [Iodine-cesium ratios in nuclear reactors]{} We assume that in an operating nuclear reactor, where criticality is sustained for chain nuclear fission reactions, these isotopes are produced with a constant production rate, while the produced isotopes decay exponentially with a constant decay rate. The number of I-131 isotopes created by a nuclear reactor which had been in operation from time $t_i$ to $t_f$ is then given at later time $t$ by[^2] $$N_{\rm I} (t) = \int_{t_i}^{t_f} d t' f_{\rm I} {\cal N}_0 e^{-\lambda_{\rm I} (t - t') },$$ where ${\cal N}_0$ is the number of fissions per unit time, $f_{\rm I}$ is the fraction of the I-131 produced per fission, and the decay rate $\lambda_{\rm I}$ is given in terms of the half-life by $\tau_{\rm I} = \ln 2/ \lambda_{\rm I} =$ 8 days. The integration gives $$\begin{aligned} N_{\rm I} (t) & = & f_{\rm I} {\cal N}_0 e^{-\lambda_{\rm I} (t - t_f)} \frac{1 - e^{-\lambda_I (t_f - t_i)} }{\lambda_{\rm I}} \nonumber \\ & \simeq & \frac{f_{\rm I} {\cal N}_0}{\lambda_{\rm I}} e^{-\lambda_{\rm I} (t - t_f) },\end{aligned}$$ where we have assumed that the working time of the reactor is much longer than the half-life of I-131, $\lambda_{\rm I} (t_f - t_i) \gg 1$. The abundance of I-131 saturates on a time scale $\tau_I$ and remains constant during the operation of the reactor. A similar estimate can be made for Cs-137. In this case, however, the decay of Cs-137 is very slow, with a long half-life $\tau_{\rm Cs} = \ln 2/ \lambda_{\rm Cs} =$ 30 years; hence $\lambda_{\rm Cs} (t_f - t_i) \ll 1$ and we obtain $$\begin{aligned} N_{\rm Cs} (t) & = & f_{\rm Cs} {\cal N}_0 e^{-\lambda_{\rm Cs} (t - t_f) } \frac{1 - e^{-\lambda_{\rm Cs} (t_f - t_i)} }{\lambda_{\rm Cs}} \nonumber \\ & \simeq & f_{\rm Cs} {\cal N}_0 \Delta t e^{-\lambda_{\rm Cs} (t - t_f) }, \end{aligned}$$ where $\Delta t = t_f - t_i$ is the time interval that the reactor has been in operation. The Cs-137 accumulates during the operation of the reactor due to its long lifetime, and hence becomes proportional to $\Delta t$. The radioactivity of each of these elements measured in becquerel (Bq) is given by the decay rates $$\Gamma_{\rm I} (t ) = - \frac{d N_{\rm I}}{dt } = f_{\rm I} {\cal N}_0 e^{-\lambda_{\rm I} (t - t_f) }$$ and $$\Gamma_{\rm Cs} (t) = - \frac{d N_{\rm Cs}}{dt } = f_{\rm Cs} {\cal N}_0 \lambda_{\rm Cs} \Delta t e^{-\lambda_{\rm Cs} (t - t_f) },$$ so that the ratio becomes $$R_{\rm I/Cs} (t) = \frac{f_{\rm I}}{f_{\rm Cs}} \frac{\tau_{\rm Cs} }{ \Delta t \ln 2} \left( \frac{1}{2} \right)^{(t - t_f)/\tau_{\rm I}}, \label{R}$$ where we have again used the condition $\lambda_{\rm Cs} (t - t_f) \ll 1$. This formula gives the average ratio of the radioactivities of I-131 and Cs-137 contained in the fuel rods after the termination of the nuclear reaction at $t_f$. The fractions of each fission product are $f_{\rm I} = 2.88 \times 10^{-2}$, $f_{\rm Cs} = 6.22 \times 10^{-2}$ for the cumulative thermal fission yields of U-235, taken from the IAEA data base[@IAEA]. For the three Fukushima Dai-ichi nuclear reactors in operation at the time $t_X$ of the earthquake, the ratio at $t = t_X$ was $$R_{\rm I/Cs} (t_X) = \frac{f_{\rm I}}{f_{\rm Cs}} \frac{\tau_{\rm Cs} }{ \Delta t \ln 2} = \frac{240 }{\Delta t},$$ where $\Delta t$ is measured in months. If we choose $\Delta t = 7 (12)$ months then $ R_{\rm I/Cs} (t_X) = 34 (20)$. On April 12, this ratio would be reduced by factor $2^{32/8} = 16$ to 2.1 (1.3).\ 3. [Deciphering the Fukushima nuclear accident data]{} We now examine the data released by TEPCO for the Fukushima nuclear accident in the light of our theoretical estimates of the ratios of the Iodine-131 and Cesium-137 accumulated in the nuclear reactors and attempt to read their implications. \(i) [Sea water at the Southern Discharge Channel:]{} In Fig. \[seawater\], we show a semi-log plot of the ratio (\[R\]) against the number of days past X-day (March 11, 2011) with two different values for $\Delta t$; the upper line corresponds to $\Delta t = 7$ months and the lower line to $\Delta t =1$ year. Note that the slopes of these lines are fixed by the half-life of I-131; only the normalization changes with a change of $\Delta t$. The larger the value of $\Delta t$, the lower the position of the line. The blue dots are the data of the sea water samples taken at the South Discharge Channel, a monitoring post located 330 m south of the Discharge Channel of the Unit-1 – 4 reactor sites.[@seawaterdata] The data are better fit by the lower line with $\Delta t = 1$ year, but the slope is a little steeper for the data taken for the first 50 days. As mentioned, the slope of the decay line is essentially fixed by the half-life of I-131, so this may at first look puzzling. The change of the slope may, however, well have been caused by the mixture of 9000 tons of old low level contaminated water which had been discharged from the Central Radioactive Waste Disposal Facility from April 4 to 10[@discharge]. In these wastes I-131 had been completely depleted and only long half-life isotopes like Cs-137 were present. The slope tends to decrease slightly later. This change may be caused by the leakage of contaminated water with large iodine-cesium ratio from Unit-2 reactor site, as we shall see later. Although we have ignored the effects caused by the chemical or physical differences of iodine and cesium in compound formation, ionization, diffusion, etc., the overall fit of the sea water data with the present theoretical calculation is very encouraging. \(ii) [Unit-4 cooling pool: ]{} We now turn our attention to the spent-fuel rods in the Unit-4 cooling pool. Note that the Unit-4 reactor was not in operation at the time of the earthquake. The same ratio for the fission products contained in the spent-fuel rods in the Unit-4 cooling pool should be smaller by an extra attenuation factor $$\left( \frac{1}{2} \right)^{(t_X - t_f)/\tau_{\rm I}} = 1/2^{90/8} \simeq 1/2400,$$ if all these fuel rods were intact since they had been removed from the Unit-4 reactor three months earlier (December 2010). Since some of the fuel could be older, this ratio may be even smaller[@skimmer]. The calculated ratio is not, however, consistent with what has been reported. According to the TEPCO press release on April 14[@site4], the radioactivities of 220 Bq/cc of I-131 and 93 Bq/cc of Cs-137 have been detected from the water sample taken on April 12 from the Unit-4 cooling pool, where more than 1500 spent-fuel rods were stored. The data were analyzed on April 13. The ratio 2.4 is closer to the values of the ratio of the radioactivities from the other reactors, as shown by the red dot in Fig. \[seawater\]. [@newdata] If the data are correct, they would imply that at least some of the radioactive fission products found in the Unit-4 cooling pool have been produced by nuclear reactions that took place at around time $t_X$ or later.[^3] One possible explanation could be contamination by fallout of the fission products created in the neighboring reactors.[^4] The iodine-cesium ratio of the water sample taken from the Unit-3 cooling pool and measured on May 8, however, has turned out to be 0.079[@pool3]; this value is about factor 3 smaller than the same ratio 0.24 for the water of Unit-4 cooling pool measured on the same day[@pool40], but instead fits to the decay line of the sea water data, suggesting that the radioactivity of the sea water has been caused by the fall-out of the radioactive material released by the hydrogen explosions. Although the difference of the ratios in Unit-3 and Unit-4 cooling pools may be caused by fluctuations of the distribution of the fallout, we may take this difference more seriously and seek other explanation for the origin of high iodine-cesium ratio found in the Unit-4 cooling pool. Another possible explanation could be that a nuclear chain reaction was reignited in the melted used fuel in the Unit-4 cooling pool for a certain period. We now explore whether criticality had been re-established only briefly, in a period $\delta t$ from time $t'_i$ to $t'_f = t'_i + \delta t$. For simplicity in the following analysis, we assume $\delta t \ll \tau_{\rm I} = 8$ days. This would be reasonable since the heat produced by a nuclear reaction in a chunk of the melted fuel would cause rapid disintegration of the chunk, terminating criticality. In this case, the number of newly produced I-131 nuclei increases in proportion to $\delta t$: $$\begin{aligned} \delta N_{\rm I} (t) & \simeq & f_{\rm I} {\cal N}_1 \delta t e^{-\lambda_{\rm I} (t - t'_f) }, \end{aligned}$$ where ${\cal N}_1 $ is the average number of nuclear fissions per unit time taking place during this brief time period. Similarly the number of newly produced Cs-137 nuclei is $$\begin{aligned} \delta N_{\rm Cs} (t) & \simeq & f_{\rm Cs} {\cal N}_1 \delta t e^{-\lambda_{\rm Cs} (t - t'_f) },\end{aligned}$$ so that the ratio of the radioactivities of the newly produced I-131 and Cs-137 would be $$R^{\rm new}_{\rm I/Cs} (t) = \frac{f_{\rm I}}{f_{\rm Cs}} \frac{\tau_{\rm Cs}}{\tau_{\rm I} } e^{-\lambda_{\rm I} (t - t'_f) } = 630 \times \left( \frac{1}{2} \right)^{(t - t'_f)/\tau_{\rm I}} . \label{Rmax}$$ This result with the replacement $t'_f \to t_X$ also gives an upper bound on the radioactivity ratio for the fission products created in the reactor under normal conditions before X-day for short $\Delta t$. The radioactivity measured in the sample water taken from the Unit-4 cooling pool could perhaps be from a mixture of the newly created radioactive fission products and the remnant of old fission products which is mostly composed of Cs-137, since I-131 has been depleted considerably, leading to the ratio 2.4 accidentally coincident with the ratio from the other reactors. In any event, the amount of new fission products may be small since the absolute value of the measured radioactivity in the sample is not significant.[^5] \(iii) [Sub-drain water:]{} The data from water samples taken from four sub-drains near the reactor buildings show even more puzzling features, as shown in Fig. \[drain\].[@sub-drain] In particular, the water samples from the sub-drain near the Unit-2 reactor building show an anomalously high radioactivity ratio[^6] even greater than the upper bound given by (\[Rmax\]) if the nuclear fission ends on X-day as indicated by the red solid line in the figure. If there is no strong chemical filtering effect in draining contaminated water from the reactor buildings, it would be difficult to understand the observed anomaly near the Unit-2 reactor without assuming that a significant amount of fission products were produced at least 10 – 15 days after X-day. We note that several step-wise sequential characteristic changes have been observed in the Unit-2 sub-drain data with largest change on May 11 (60 days from X-day). The causes of these changes are still unknown to the author, but it is tempting to speculate that they were caused by the mixing of the new fission products with the old ones released three times later from the remaining fuel rods in the Unit-2 reactor or from the water from other reactor site, as we shall see later. The data from the Unit-3 sub-drain before April 23 sit close to the decay line which fits the sea water data, hence they may be understood as due to radiation from fission products produced before the X-day. However, the data of the Unit-1 sub-drain and Unit-4 sub-drain give a high radioactivity ratio, even larger than that of the samples from the Unit-4 cooling pool and from the Unit-3 sub-drain. The data therefore cannot be explained by the contamination of the old fission products which had existed in the spent-fuel rods in the Unit-4 cooling pool. The data occasionally show very different characteristics which are difficult to be understood unless there was considerable mixing of waters belonging to different sub-drains due to in-flow and out-flow of water through underground water channels. We note that the anomalously large iodine-cesium ratios were observed also in the sub-drain water in Unit-3 and Unit-4 reactor sites in June, and I-131 has been detected in the sub-drain water at Unit-1 and Unit-3 reactor sites briefly even as late as August 31, July 28, respectively. It is difficult, however, to judge the significance of these data since there were considerably large fluctuations. \(iv) [Intake Canal water:]{} We show a few more plots of the data available for sea water samples taken near the Intake Canal of the Unit-1 – 4 reactor sites that may fill the gap between the previous sea water data (Fig. 1) and the sub-drain water data (Fig. 2).[^7] We note that a leakage of contaminated water from the water-intake of Unit-2 reactor had been reported on April 2nd, and this prompted to monitor the radioactive isotopes contained in the contaminated water near the intake of Unit-2 reactor.[@u2leakage] The left panel of Fig. 3 corresponds to the sea water at the Shallow Draft Quay, located just outside of the canal, while the right panel is for the samples taken at the north (green) and south (pink) sides of the Intake Canal: the north monitoring post is located near the Unit-1 reactor intake, while the south post is near the Unit-4 reactor intake. Although the available data is limited, all these ratios seem to have stayed very close to the values of the sea water data taken at the southern monitoring post 330 m away from the canal until April 10 (30 days from X-day) and then started to deviate; the deviation becomes bigger at the north monitoring post than at the south, indicating that it is caused by the change of character of the sea water due to mixing with new contaminated water leaking out from the Unit-2 reactor. A sudden drop of the ratio has been observed on May 10 (60 days after X day) in these plots. This change may be caused by the leakage of contaminated water found in the pit of Unit-3 reactor sites; the iodine-cesium ratio may have been small in the water leaking out from the Unit-3 reactor and hence considerably diluted the ratio of pre-existing water in the canal with large iodine-cesium ratio.[@u3leakage] After May 11 when the leakage from the intake of Unit-3 reactor was stopped, the ratios started to increase again and the Intake Canal data exceed the upper bound (red line) allowable, had the nuclear reaction terminated on the X day. We would also like to note that the iodine-cesium ratio in the contaminated water leaked from the Unit-3 reactor site is very close to the ratio found in the southern monitoring post 330 m south of these troubled reactors. This may not be just a coincidence. To locate the origin of the increase of the ratios, we plot in Fig. 4 the data taken at the bar screens of the individual water intakes for Unit-1 – 4 reactors: the left (right) panel is a plot of the data taken outside (inside) of the silt fences which have been installed at the bar screen of the intake of the Unit-2 reactor site since April 18, and at the bar screens of the other three reactor sites since May 12. The radioactivity ratios of iodine-131 and Cs-137 for the water outside of the fences show very similar behavior, with striking similarity to the north Intake Canal data (the green dots in the right panel of Fig. 3). On the other hand, the water sample taken inside the silt fences shows clear difference of characteristics from the water coming from individual reactor sites: the water coming from the pit of the Unit-2 reactor site has in particular an anomalously large iodine-cesium ratio. This is qualitatively similar to what we have observed for the sub-drain water, while the ratio fluctuates more rapidly in the water leaking out of the pit of Unit-2 reactor site. It is apparent that the contaminated water leaking out of the Unit-2 reactor site is responsible for the over-all changes of the iodine-cesium ratios of the other parts of the water in the Intake Canal. It is interesting to note that the air sample taken from a very humid (99.9%) air inside the Unit-2 reactor building has a very high iodine-cesium ratio of 0.51 even 99 days after X day, close to the value measured for the water coming out of the Unit-2 pit. [@air]\ 4. [Conclusions: ]{} In these notes we have estimated the radioactivity ratios of iodine isotope I-131 and cesium isotope Cs-137 accumulated in the nuclear reactor and compared the results with the data of the Fukushima nuclear reactor accidents released by TEPCO for the first 100 days after the outbreak of the accident. Although there are still many uncertainties in connecting our estimates of the ratio in the nuclear reactor to the ratios measured in the water sampled outside of of the reactors, as mentioned in the very beginning of these notes, the results are rather provocative. The ratio of the measured radioactivity of I-131 and Cs-137 may be used to extract useful information about when these fission products were produced in the nuclear reactor complex of the Fukushima Dai-ichi plant. The sea water data taken at southern monitoring post are consistent with the expected radioactivity by the fission products which had been produced before the earthquake and perhaps released by the hydrogen explosions of Unit-1 and -3 reactor buildings. The data of the water samples from the Unit-4 cooling pool and from areas near the Unit-2 reactor, however, show an anomaly which may indicate, if the data are correct, that some of these fission products were produced by chain nuclear reactions reignited after the earthquake. Some of these results may well be affected by chemical, thermodynamic and/or transport effects which are not taken into account in the present analysis. Estimates of these effects are rather difficult problems which would require detailed knowledge of the course of the accident in addition to various engineering problems related to the reactor design. Hopefully, more direct information concerning the interior conditions of the troubled reactors and the used fuel rods in the cooling pool will eventually become available and will clarify what actually had happened at the Fukushima Dai-Ichi reactor complex. The author thanks Professors Gordon Baym, Munetake Ichimura and Koichi Yazaki for their interest in this work and for helpful communications. In particular, he is grateful to Professor Ichimura for patiently elucidating some of the data released by TEPCO and to Professor Baym for carefully reading these notes. He is grateful to Professor Ryogo Hayano for sharing his insights on many related aspects of the data for the Fukushima nuclear accident and for calling the author’s attention to the intake canal data, and to Professor Mamoru Shimoi for information regarding the water-solubility of iodine and cesium elements. [99]{} `http://www.tepco.co.jp/en/index-e.html` `http://www.tepco.co.jp/en/press/corp-com/release/11041412-e.html` W. F. Libby, E. C. Anderson and J. R. Arnold, Science [109]{} (1949), 227.\ See also E. Fermi, [Nuclear Physics]{} (Univ. Chicago Press, 1949), p. 220. It is known that most of Iodine isotopes are released into the coolant water in the form of CsI salt which then dissolve into the water by forming Cs$^+$ and I$^-$ ions. However the amount of cesium isotopes release into the water is an order of magnitude greater than that of the iodine isotopes. See E. C. Beahm, C. F. Weber, T. S. Kress, [Iodine Forms in LWR Severe Accidents]{}, NUREG/TM-11861, Oak Ridge National Laboratory (1991). We note that most of cesium isotopes would react with water strongly by the chemical reaction: 2 Cs + 2 H$_2$O $\to$ 2 CsOH + H$_2$ where the produced cesium hydroxide (CsOH) would melt into the water forming Cs$^+$ and OH$^-$ions. INDC International Nuclear Data Committee, [Handbook of nuclear data for safeguards: Database extensions, August 2008]{} , INDC(NDS)-534. The sea water data taken at the Southern Monitoring Post is available since March 21 (`http://www.tepco.co.jp/en/press/corp-com/release/11032208-e.html`) until May 21 (`http://www.tepco.co.jp/en/press/corp-com/release/11052207-e.html`). We plot the ratios of the I-131 and Cs-137 estimated from the raw data which do not contain information on possible errors in the detections. The data is not available after May 22, due to decrease of the I-131 density in the sea water below detectable level. `http://www.tepco.co.jp/en/press/corp-com/release/11041507-e.html` The data taken from the Unit-2 skimmer surge tank near the cooling pool shows a very small I/Cs ratio of 0.0256 which may reflect mostly the value in the used fuel rods with perhaps with contamination by fresher fission products from the Unit-2 reactor.\ `http://www.tepco.co.jp/en/press/corp-com/release/betu11_e/images/110418e4.pdf` In Fig. 1, new data of the unit4 cooling pool, released on April 29\ (`http://www.tepco.co.jp/en/press/corp-com/release/11042909-e.html`) and May 9 (`http://www.tepco.co.jp/en/press/corp-com/release/11050905-e.html`) are also shown. They sit close to the same decay line confirming the previous data. This also implies that there has been no nuclear reaction ignited at least since April 14 in the Unit-4 cooling pool. `http://www.tepco.co.jp/en/press/corp-com/release/11051009-e.html` `http://www.tepco.co.jp/en/press/corp-com/release/11050905-e.html` The first sub-drain data was uploaded at\ `http://www.tepco.co.jp/en/press/corp-com/release/betu11_e/images/110331e18.pdf`,\ which has been updated since then. Leakage of the contaminated water from Unit-2 intake screen has been reported in\ `http://www.tepco.co.jp/en/press/corp-com/release/11040203-e.html`,\ `http://www.tepco.co.jp/en/press/corp-com/release/11040506-e.html`,\ `http://www.tepco.co.jp/en/press/corp-com/release/11040601-e.html`.\ See also the Japanese Government Report to IAEA released on June 7:\ `http://www.kantei.go.jp/foreign/kan/topics/201106/iaea_houkokusho_e.html` The leakage of contaminated water in the pit of Unit-3 reactor site has been reported on May 11 in `http://www.tepco.co.jp/en/press/corp-com/release/11051112-e.html` These data are contained in the attachment released by TEPCO on June 19 on the environment inside the Unit-2 reactor building:\ `http://www.tepco.co.jp/en/press/corp-com/release/11061903-e.html` [^1]: This problem can be avoided by using Cs-134 and Cs-137. However, Cs-134 has a half-life of 2 years which is too long to see observable effects in the currently available data. [^2]: This result may be obtained by integration of the equation, $ d N_{\rm I}/dt = f_{\rm I} {\cal N}_0 \theta (t; t_1, t_f ) - \lambda_{\rm I} N_{\rm I}$, where $\theta (t; t_i, t_f ) = 1$ for $t_i < t < t_f$ and $\theta (t; t_i, t_f ) = 0$ otherwise, with the initial condition $N_{\rm I} (t_i) = 0$. [^3]: We have tacitly assumed that the solubility of iodine or cesium in water is not changed considerably by boric acid, which has been added to the water to suppress chain nuclear reactions. [^4]: Actually, the hydrogen explosions of the Unit-4 reactor building took place on March 15 after the explosion of the Unit-3 reactor building on March 14. However, the explosion on March 15 may have sent fallout from the hydrogen explosion of the Unit-3 building on the roof of the Unit-4 reactor building into the pool. The author thanks Professor Ichimura for calling his attention to these subtle points. [^5]: The total amount of the radioactivity by I-131 contained in the water of Unit-4 cooling pool, which normally contains about 1500 tons of water, may be estimated as $220 \times 1.5 \times 10^9 = 3.3 \times 10^{11}$ Bq (on April 13) which is small compared to the radioactivity of fresh fission products contained in the nuclear reactor, which is the order of $10^{18}$ Bq (on X-day). If we assume that these I-131 isotopes detected in the Unit-4 cooling pool were produced by nuclear fissions reignited inside the pool after X-day, the upper bound of the total number of fissions occurred may be estimated as $N_f = 3.3 \times 10^{11} \times 2^{32/8} \times \tau_I / f_I = 1.3 \times 10^{20}$, which would have produced an extra heat $\Delta U = N_f \times 200~{\rm MeV} = 4.1 \times 10^9 $J. This extra heat produced is equivalent to a latent heat absorbed by about 2 tons of boiling water, which is not significant if one compares it to the decay heat produced daily during the period by the used fuel rods in the cooling pool and therefore may have been easily unnoticed. This, however, may not exclude a possibility that nuclear reaction had reignited on a larger scale and some extra fission products had escaped from the pool together with violently boiling water by sudden release of large amount of heat. [^6]: The author thanks Gordon Baym for calling his attention to this point. [^7]: The author thanks Ryogo Hayano for calling his attention to the Intake Canal data.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We introduce the class of rational plane curves parameterizable by conics as an extension of the family of curves parameterizable by lines (also known as monoid curves). We show that they are the image of monoid curves via suitable quadratic transformations in projective plane. We also describe all the possible proper parameterizations of them, and a set of minimal generators of the Rees Algebra associated to these parameterizations, extending well-known results for curves parameterizable by lines.' address: - 'Universitat de Barcelona, Departament d’[À]{}lgebra i Geometria. Gran Via 585, E-08007 Barcelona, Spain' - 'Universitat de Barcelona, Departament d’[À]{}lgebra i Geometria. Gran Via 585, E-08007 Barcelona, Spain' author: - Teresa Cortadellas Benítez - 'Carlos D’Andrea' title: rational plane curves parameterizable by conics --- [^1] Curves parameterizable by forms of low degree {#uno} ============================================= This article deals with algebraic and geometric features of a special familiy of rational plane curves. Let $\K$ be an algebraically closed field. For a positive integer $k,$ we will denote with $\P^k$ the $k$-dimensional projective space over $\K$. Let $\C\subset \P^2$ be an algebraic plane curve of degree $d$, that is the the zero locus of an irreducible homogeneous polynomial $E(X_1,X_2,X_3)\in\K[X_1,X_2,X_3]$ of degree $d$. A curve is rational if it is birationally equivalent to $\P^1$, i.e.there exist dominant rational maps $\phi:\P^1 \to \C$ and $\psi: \C\dasharrow \P^1$ such that $\psi \circ \phi =id_{\P^1}$ and $\phi \circ \psi =id_{\C};$ equivalently there is an open subset of $\C$ isomorphic to an open subset of $\P^1$ (or $\A^1$). If this is the case, the cardinality of the general fiber of $\phi $ and $\psi $ is equal to one. So, $\phi$ actually defines a proper (i.e. generically injective) parameterization of $\C$. Note that $\psi$’s domain (the largest set where the map is defined) coincides with the set of nonsingular points of $\C$. A pair $(F_1(X_1,X_2,X_3),\,F_2(X_1,X_2,X_3))$ of polynomials in $\K[X_1,X_2,X_3],$ homogeneous of the same degree $d'$ without common factors defines a birational map $$\label{inverse} \begin{array}{cccc} \psi:&\C&\dashrightarrow &\P^1\\ &(x_{1}:x_{2}:x_{3})&\mapsto&(F_1(x_{1},x_{2},x_{3}):F_2(x_{1},x_{2},x_{3})) \end{array}$$ if there exist a triple $(u_1(T_1,T_2),\,u_2(T_1,T_2),\,u_3(T_1,T_2))$ of homogeneous polynomials without common factors in $\K[T_1,T_2]$ defining a parameterization of $C$ of the form $$\label{param} \begin{array}{cccc} \phi:&\P^1&\to&\C\\ &(t_{1}:t_{2})&\mapsto&(u_1(t_{1},t_{2}):u_2(t_{1},t_{2}):u_3(t_{1},t_{2})), \end{array}$$ with $\phi =\psi^{-1}$ as rational maps. Note that $\phi$ is globally defined but $\psi$ not necessarily. In fact, it is well-known (see Lemma \[silverman\]) that the set of singular points of $\C$ is contained in the algebraic variety defined by $F_1(\X)$ and $F_2(\X)$ in $\P^2$. Note that the inclusion may be strict, see for instance Example \[ejj\]. Set $\t:=(t_1,t_2)$ and $\x:=(x_1,x_2,x_3)$. The birationality of $\psi$ is equivalent to the following two claims: $$\label{u2} (u_1(F_1(\x),F_2(\x)):u_2(F_1(\x),F_2(\x)):u_3(F_1(\x),F_2(\x)) )=(x_1:x_2:x_3)$$ for almost all $(x_1:x_2:x_3) \in \C $, and $$\label{u1} \big(F_1(u_1(\t),u_2(\t),u_3(\t)):F_2(u_1(\t),u_2(\t),u_3(\t))\big)=(t_1:t_2)\\$$ for almost all $(t_1:t_2) \in \P^1.$ Note that the expressions in the left hand side of and are well-defined as the families of polynomials are homogeneous. Set $\T:=(T_1,T_2),\,\X:=(X_1,X_2,X_3),\,\u(\T)=(u_1(\T),\,u_2(\T),\,u_3(\T))$ and $\F(\X)=(F_1(\X),F_2(\X)).$ If holds, we say that $\C$ is parameterizable by $\F(\X)$ (or by $\psi$) and that $\u(\T)$ (or $\phi$) is the proper parameterization induced by $\F(\X)$. Note that is equivalent to $$\label{u} T_1F_2(\u(\T))-T_2F_1(\u(\T))=0,$$ and it turns out that implies . This is clear if the characteristic of $\K$ is zero, and reasoning as in Proposition $2.1$ [@CD10] for $\psi$ birational, one gets the general case. In order to find $\u(\T)$ starting from the data $\psi$ given in , some geometry is needed. For a set of homogeneous elements $S\subset\K[\X]$, we denote with $\V(S)\subset\P^2$ the variety defined by it. The fact that $\C=\V(E(\X))$ is parameterizable by $\F(\X)$ means that the system $$\label{dprima} \left\{\begin{array}{lcc} E(\X)&=&0\\ T_1F_2(\X)-T_2F_1(\X)&=&0 \end{array} \right.$$ has only one solution in $\P^2_{\overline{\K(\T)}}\setminus \P^2$ counted with multiplicities, or equivalently has $dd'-1$ zeroes in $\P^2$ counted with multiplicities. Here, $\P^2_{\overline{\K(\T)}}$ is the projective plane over $\overline{\K(\T)}$, the algebraic closure of $\K(\T)$. Note that our definition is not the same as the one given in [@SWP08 Definition 4.51] but a more restrictive one as shown in [@SWP08 Theorem 4.54]. From a computational point of view, a curve in the plane is typically given by either its implicit equation $E(\X)$ or —if it is rational— a parameterization like . Whether there exists a proper parameterization of $\C$ and, if this is the case, the computation of $\psi$ having as input $\phi$ or viceversa, are typical problems of Computational Algebraic Geometry, see [@SWP08] and the references therein for more on the subject. Let us consider the situation from a more algebraic perspective. Set $R:=\K[\T],$ and let $I$ be the ideal $\langle u_1(\T),u_2(\T),u_3(\T) \rangle \subset R.$ The Rees Algebra associated to $I$ is defined as $\mbox{Rees}(I):=\K[\T][I\,Z],$ where $Z$ is a new variable. There is a graded epimorphism of $K[\T]$-algebras defined by $$\label{ris} \begin{array}{cccc} {\mathfrak h}:&\K[\T][\X]&\to&\mbox{Rees}({\mathcal I})\\ &X_i&\mapsto&u_i(\T)\,Z. \end{array}$$ Set $\kK:=\ker ({\mathfrak h})$. Note that a description of $\kK$ allows also a full characterization of $\mbox{Rees}({\mathcal I})$ via . This is why we call it the defining ideal of the Rees Algebra associated to $\u(\T)$. Condition is equivalent to the fact that $T_1F_2(\X)-T_2F_1(\X)\in\kK.$ Observe that $\kK$ is a bihomogeneous ideal, and that one has an identification of $\kK_{\ast,1}$ with $\mbox{Syz}(I),$ the first module of syzygies of $I$. It turns out that $\mbox{Syz}(I)$ is a free $R$-module of rank $2$ generated by two elements, one of $\T$-degree $\mu$ for an integer $\mu$ such that $0\leq\mu\leq\frac{d}{2}$, and the other of $\T$-degree $d-\mu$. In the Computer Aided Geometric Design community, such a basis is called a $\mu$-basis of $I$ (see for instance [@CSC98; @CGZ00; @CCL05]). Indeed, by the Hilbert-Burch Theorem, $I$ is generated by the maximal order minors of a $3\times 2$ matrix $\varphi$ and the homogeneous resolution of $I$ is $$\label{sec} 0\longrightarrow R(-d-\mu)\oplus R(-d-(d-\mu)) \stackrel{\varphi}{\longrightarrow} R(-d)^3 \stackrel{(u_1,u_2,u_3)}{\longrightarrow} I\longrightarrow 0.$$ This matrix is called the Hilbert-Burch matrix of $I$ and its columns describe the $\mu$-basis. Computationally, a $\mu$-basis provides simple (i.e. in both $(\T,\X)$-degrees) elements to describe the parameterization of $\C$ given in than the data $\u(\T)$. The search for more simple elements to describe $\C$ leads to the study of the minimal generators of $\kK$. Indeed, the so-called method of implicitization by using moving curves of low degrees (described in [@SC95; @SGD97; @ZCG99]) is just a first step into a more complex picture which was described by Cox in [@cox08], and subsequently worked out in [@CHW08; @HSV08; @KPU09; @HSV09; @bus09; @HW10; @CD10] among others. However, we are still far from being able to describe minimal generators of $\kK$ for a general ideal of a parametric plane curve $I$ as above. This paper is a contribution in that direction. We will make a detailed study of rational curves parameterizable by forms of degree $2$, i.e. the situation $\deg(\F(\X))=2$ in . The case of curves parameterizable by forms of degree $1$ has been completely described in [@cox08; @bus09]. Before starting, we present some results concerning existence and uniqueness of the polynomials $F_1(\X),\,F_2(\X)$ defining for a fixed $\C$. Any rational plane curve $\C$ is parameterizable by forms of degree $d'$ for some $d'$. As a matter of fact, the method of adjoint curves proposed in [@walker] to parameterize any rational curve produces a map $\psi$ as in , with $\F(\X)$ of degree less than or equal to $\deg(\C)-2$. The following result shows that if $d'<\frac{\deg(\C)}{2}$, then not only $d'$ is unique but also the ideal $\langle F_1(\X),\,F_2(\X)\rangle$. \[d’\] Let $\C$ be a curve of degree $d$ parameterizable by $(F_1(\X),\,F_2(\X))$, with $\deg(F_i(\X))=d'$. Suppose that $\C$ is also parameterizable by $(F^0_1(\X),\,F^0_2(\X))$, the latter being forms of degree $d_0'$ with both $d',\,d_0'<d$. Then, either $d'+d_0'\geq d$ or $d'=d_0'$ and $\langle F_1(\X),\,F_2(\X)\rangle=\langle F^0_1(\X),\,F^0_2(\X)\rangle$. Let $\phi(\t):=\u(\t)$ and $\phi^ 0(\t):=\u^0(\t),$ be the proper parameterizations of $\C$ induced respectively by $(F_1(\X),\,F_2(\X))$ and $(F^0_1(\X),F^0_2(\X))$. Denote with $\psi$ the inverse of $\phi$. Then, $\psi \circ \phi^0$ is an automorphism of $\P^1$, and hence there exists a pair $(\rho_1(\T),\rho_2(\T))=:\rho(\T)$ of $\K$-linearly independent linear forms such that $\u^0(\T)=\u(\rho(\T))$. Let $\u^ 0(\T):=(u^0_1(\T), u^0_2(\T), u_3^0(\T))$ be the parameterization induced by $(F^0_1,F^0_2)$. From we have $T_1F^0_2(\u(\rho(\T)))-T_2F^0_1(\u(\rho(\T)))=0.$ And by writing $T_1$ and $T_2$ as linear combinations of $\rho_1(\T)$ and $\rho_2(\T)$, we get $$\rho_1(\T)F'_2(\u(\rho(\T)))-\rho_2(\T)F'_1(\u(\rho(\T)))=0$$ with $\langle F^0_1(\X),\,F^0_2(\X)\rangle=\langle F'_1(\X),\,F'_2(\X)\rangle$. As $\rho$ is an automorphism, we deduce $$T_1F'_2(\u(\T))-T_2F'_1(\u(\T))=0.$$ This equality, combined with $T_1F_2(\u(\T))-T_2F_1(\u(\T))=0$ implies that the polynomial $F'_1(\X)F_2(\X)-F'_2(\X)F_1(\X)$ vanishes on $\C$. As this is an element of degree $d'+d_0'$ and $\C$ has degree $d$, if $d'+d_0'<d$, then we have that $$F'_1(\X)F_2(\X)-F'_2(\X)F_1(\X)=0.$$ Now, using the fact that $F_1(\X)$ and $F_2(\X)$ do not share any common factor, we deduce that $F_i(\X)$ divides $F'_i(\X)$ for $i=1,2$, so $d'\leq d_0'$ and $$\langle F_1^0(\X), F_2^0(\X) \rangle=\langle F_1'(\X), F_2'(\X) \rangle\subset\langle F_1(\X),\,F_2(\X) \rangle .$$ Applying the same argument symmetrically, we conclude that $d_0'\leq d'$, and hence $$\langle F_1(\X),F_2(\X)\rangle\subset\langle F^0_1(\X),\,F^0_2(\X)\rangle.$$ If we restrict our attention to the set of curves parameterizable by forms of degree $d'$ for a fixed value of $d'$, the following natural questions arise: - Can we describe geometrically all of them? - What does a proper parameterization of a curve in this family look like? - Given $u_1(\T),u_2(\T),u_3(\T)\in K[\T]$ parametrizing a plane curve parameterizable by forms of degree $d'$, can we describe the minimal homogeneous free resolution of $\langle u_1(\T),\,u_2(\T),u_3(\T) \rangle $? - Given $u_1(\T),u_2(\T),u_3(\T)\in K[\T]$ as above, can we describe a minimal set of generators of $\kK$ ? An already interesting case is when $d'=1$. Such curves are called in [@SWP08] parameterizable by lines. Other authors call them monoid curves ([@JLP08]). The answer to all these questions are well-known for them. We will review them along the text in order to compare them with the main focus of this paper, which is $d'=2$. We will refer to them as curves parameterizable by conics. In Section \[sconics\] we will describe all possible proper parameterizations of them, and also compute a non-trivial multiple of its implicit equation. Most of the time, this polynomial will actually be the one defining its implicit equation and, when it is not the case, the implicit equation will be given by its irreducible factor of largest degree (see Theorem \[complete\]). In Section \[geometry\], we describe geometrically the space of all curves parameterizable by conics. In Theorem \[mtmt\] we show that they are the image of curves parameterizable by lines via a quadratic birational transformation of $\P^2$. Not surprisingly, the type of quadratic transformation depends on the geometry of the variety defined by $F_1(\X),\,F_2(\X)$ in $\P^2$. Then we turn to study the last of the questions above. In Section \[s1\] we present an extension of some of the tools used in [@CD10] for curves parameterizable by lines, to a more general context. These extended tools will be used in Section \[Rees\] to exhibit a complete set of generators of $\kK$ for proper parameterizations of curves parameterizable by conics. Curiously, the description of the generators depends on whether the degree of $\C$ is even or odd. In the first case, a “moving conic” arising from the classical method of implicitization with the aid of moving curves comes into play (see Proposition \[mconic\]). It is worth mentioning here that the results in Sections \[s1\] and \[Rees\] are independent of the previous sections, so the reader interested in the questions related to the Rees Algebra can skip the first pages without harm. Of course it would be very interesting to get a further understanding of the situation for $d'\geq3$, but our techniques only allow us to deal with curves parameterizable by conics. In Section \[conclu\], we conclude with open questions and problems. We are very grateful to E. Casas-Alvero for a careful reading of a preliminary draft of this paper, and also for patiently explaining us several features of the geometry of plane curves, in particular for helping us work out the canonical forms of Lemma \[reduccion\], and the quadratic transformations appearing in Section \[geometry\]. We are also grateful to J. C. Naranjo, J. I. Burgos and the anonymous referee for helpful comments and suggestions, and to the anonymous referee for several suggestions and corrections in the final version of this text. Our computations have been done with the aid of the softwares [Macaulay 2]{}, [Maple]{}, and [Mathematica.]{} Parameterizations and implicit equations of curves parameterizable by lines and conics {#sconics} ====================================================================================== In this section we will explore algebraic aspects of curves parameterizable by forms of degrees $1$ and $2$. They will be useful when studying geometric properties of the singularities of these curves. The case of curves parameterizable by lines is well-known in the literature. We review it here in order to compare it with curves parameterizable by conics. Curves of degree $1$ (lines in $\P^2$) are easily to describe so we will assume from now on that $d\geq2$. Curves parameterizable by lines {#monoid} ------------------------------- We start with the following result which characterizes curves parameterizable by lines having $(0:0:1)$ as a point of maximal multiplicity. Without loss of generality, we can assume that the inverse $\psi$ defined in is given by $F_1(\X)=X_1,\,F_2(\X)=X_2.$ Let $a(\T),\,b(\T)\in\K[\T]$ be homogeneous polynomials without common factors, of degrees $d-1$ and $d>1$ respectively. Set $$\label{monn} \left\{\begin{array}{ccl} u_1(\T)&:=&T_1\,a(\T),\\ u_2(\T)&:=&T_2\,a(\T),\\ u_3(\T)&:=&b(\T). \end{array} \right.$$ Then, $\u(\T):=(u_1(\T),\,u_2(\T),\,u_3(\T))$ defines a proper parameterization of curve $\C$ of degree $d$ parameterizable by lines having $(0:0:1)\in\C$ of multiplicity $d-1.$ Moreover, $b(X_1,X_2)-a(X_1,X_2)X_3$ is an irreducible polynomial defining $\C.$ This curve is parameterizable by $(X_1,X_2)$. Reciprocally, any curve defined implicitly as $b(X_1,X_2)-a(X_1,X_2)X_3=0$ in $\P^2$ with $a(\T),\,b(\T)$ as above, is a curve parameterizable by lines with $(0:0:1)\in\C$ having multiplicity $d-1$. Write $b(\T)=b_1(\T)T_1+b_2(\T)T_2.$ It is then easy to see that the matrix $$\varphi:=\left(\begin{array}{cc} T_2& b_1(\T)\\ -T_1& b_2(\T)\\ 0& -a(\T) \end{array} \right)$$ is the Hilbert-Burch matrix of the ideal $\langle u_1(\T),\,u_2(\T),\,u_3(\T)\rangle\subset\K[\T]$, as in . By looking at the $\T$-degree of the first column, we get that $\mu=1$, i.e. there is a generator of the $\mbox{Syz}(I)$ of $\T$-degree one. Proposition $2.1$ in [@CD10] tell us then that $\u(\T)$ defines a birational map $\phi:\P^1\longrightarrow \C:=\phi (\P^1)$ whose inverse is given by $(X_1,X_2)$. In particular, $\phi$ is a proper parameterization of a curve of degree $d$ having with $(0:0:1)\in\C$ having multiplicity $d-1.$ The fact that the implicit equation is given by $b(X_1,X_2)-a(X_1,X_2)X_3$ was shown in [@CD10 Lemma $2.5$]. The rest of the proof follows straightforwardly: given $a(\T),\,b(\T)\in\K[\T]$ homogeneous without common factors and with respective degrees $d-1,\,d$. With this data we define the parameterization and then we will find that the implicit equation of $\C$ is given by the irreducible polynomial $b(X_1,X_2)-a(X_1,X_2)X_3$. Curves parameterizable by conics {#deg2} -------------------------------- In order to mimic the results obtained above, by making a linear change of coordinates in $\P^2$ we start by assuming that $(0:0:1)\in\V(\F(\X))$. Set ${\mathcal F}(\T,\X):=T_1F_2(\X)-T_2F_1(\X)$, and write $$\label{conicg} {\mathcal F}(\T,\X)=l_1(\T)X_1X_2+l_2(\T)X_1X_3+l_3(\T)X_2X_3+l_4(\T)X_1^2+l_5(\T)X_2^2,$$ with $l_i(\T)$ a homogeneous linear form in $\K[\T],\ i=1, 2, 3, 4, 5$. \[prop22\] The conic defined by ${\mathcal F}(\T,\X)$ in $\P^2_{\overline{\K(\T)}}$ is degenerate if and only if each $F_i(\X)$ is the product of two linear forms in $\K[X_1,X_2]$. If this is the case, there is a curve $\C$ parameterizable by $\F(\X)$ if and only if $\C$ is either a line or parameterizable by lines. If ${\mathcal F}(\T,\X)$ defines a degenerate conic then there exist ${\mathcal A}(\T,\X),\,{\mathcal B}(\T,\X)\in\overline{\K(\T)}[\X]$ homogeneous of $\X$-degree one such that $$\label{cua} {\mathcal F}(\T,\X)={\mathcal A}(\T,\X)\,{\mathcal B}(\T,\X).$$ As the left hand side has degree at most one in $X_3$, one of the factors on the right hand side do not depend on $X_3$. Suppose w.l.o.g. that $\deg_{X_3}({\mathcal A}(\T,\X))=0$, and write $$T_1F_2(\X)-T_2F_1(\X)={\mathcal F}(\T,\X)=Q(\T,X_1,X_2)+X_3\,L(\T,X_1,X_2),$$ with $Q(\T,X_1,X_2),\,L(\T,X_1,X_2)\in\K[\T,\X]$. If $L(\T,X_1,X_2)\neq0,$ then —due to — both polynomials $Q(\T,X_1,X_2)$ and $L(\T,X_1,X_2)$ will have a non trivial common factor in $\overline{\K(\T)}[X_1, X_2]$. But this implies that they also share a common factor in $\K[\T,X_1, X_2]$, so a factorization as in holds, with ${\mathcal A}(\T,\X)\,{\mathcal B}(\T,\X)\in \K[\T,X_1, X_2].$ Looking now at the degree in $\T$ in , we have that one of the two factors in the right hand side does not depend on $\T,$ which implies that $F_1(\X)$ and $F_2(\X)$ have a common factor of positive degree, a contradiction. Hence, $L(\T,X_1,X_2)=0$, which implies that $F_1(\X)$ and $F_2(\X)$ only depend on $X_1,\,X_2$, and they factorize as a product of linear forms, as $\K$ is algebraically closed. The converse follows straightforwardly as $T_1F_2(X_1,X_2)-T_2F_1(X_1,X_2)$ factorizes as a product of two linear forms with coefficients in $\overline{\K(\T)},$ and hence they define a product of lines in $\P^2_{\overline{\K(\T)}}.$ Now, suppose that $F_1(\X),\,F_2(\X)\in\K[X_1,X_2]$. It is easy to see that here is a curve parameterizable by these conics if and only if there is a solution in $\P^1_{\K(\T)}$ of the equation $T_1F_2(X_1,X_2)-T_2F_1(X_1,X_2)=0$. By dividing this equality by $X_2^2$, we get a quadratic equation in $\frac{X_1}{X_2}$ whose coefficients are linear forms in $\T$. By Gauss Lemma, any rational solution should have both numerator and denominator being of $\T$-degree at most one. By looking at the shape of the first two coordinates of , we conclude that $\C$ is either a line or parameterizable by lines. \[elunico\] If $T_1F_2(X_1,X_2)-T_2F_1(X_1,X_2)=0$ has no rational solutions in $\P^2_{\K(\T)},$ then there are no rational curves parameterizable by $\F(\X)$. We will see below that this is actually the only possible choice of a complete intersection of conics in $\P^2$ which does not parameterize a curve $\C$. Now we deal with nonsingular pencils of conics. We will describe all the rational plane curves they produce by means of the usual argument of cutting out the pencil with a moving line passing through $(0:0:1)$. \[paramparam\] Let $F_1(\X),\,F_2(\X)\in\K[\X]$ be homogeneous of degree $2$ without common factors such that $(0:0:1)\in\V(\F(\X))$. If the conic defined by ${\mathcal F}(\T,\X)$ in $\P^2_{\overline{\K(\T)}}$ is nondegenerate, then for any pair $a(\T),\,b(\T)\in\K[\T]$ of homogeneous elements of the same degree $d_0>1$ without common factors, the polynomials $$\label{paramm} \left\{ \begin{array}{ccl} u_1(\T)&=&-a(\T)\big(a(\T)l_2(\T)+b(\T)l_3(\T)\big) \\ u_2(\T)&=&-b(\T)\big(a(\T)l_2(\T)+b(\T)l_3(\T)\big)\\ u_3(\T)&=& a(\T)b(\T)l_1(\T)+a(\T)^2l_4(\T)+b(\T)^2l_5(\T). \end{array} \right.$$ define a proper parameterization of a curve $\C$ parameterizable by $\F(\X)$. Moreover, if $\gcd(X_1l_2(\F(\X))+X_2l_3(F(\X)), a(\F(\X))X_2-b(\F(\X))X_1)=1,$ then $\gcd(\u(\T))=1$, and $\deg(\C)=2d_0+1$. Moreover, $a(\F(\X))X_2-b(\F(\X))X_1$ is an irreducible polynomial defining the curve. As $(0:0:1)$ is a rational point of the nondegenerate conic in $\P^2_{\overline{\K(\T)}},$ we can describe all the other rational solutions by using a pencil of lines passing through this point. In order to do that, given $a(\T),\,b(\T)\in\K[\T]$ homogeneous elements of degree $d_0>1$ without common factors, consider the system $$\left\{\begin{array}{lcc} {\mathcal F}(\T,\X)&=&0,\\ b(\T)X_1-a(\T)X_2&=&0. \end{array} \right.$$ It has two solutions in $\P^2_{\K(\T)}$, one of them being $(0:0:1)$, so the other is also rational and by computing it explicitly we get that it is proportional to $\u(\T)$ in . As $\gcd(a(\T),\,b(\T))=1$ and due to the fact that at least one between $l_2(\T)$ and $l_3(\T)$ is not identically zero (this is because the conic defined by ${\mathcal F}(\T,\X)$ in $\P^2_{\overline{\K(\T)}}$ is nondegenerate), we then have that defines the parameterization of a rational plane curve $\C$, which turns out to be parameterizable by $\F(\X)$. Hence, the parameterization is proper. Let $E(\X)\in\K[\X]$ be an irreducible polynomial defining $\C$. For $(x_1:x_2:x_3)\in \C$ we have $b(\F(\x))x_1-a(\F(\x))x_2=0$, which implies that $b(\F(\X))X_1-a(\F(\X))X_2$ is a multiple of $E(\X).$ In order to show that they are equal, first we will prove that the latter is not identically zero. Indeed, if this were the case, then there would exist $C(\X)\in\K[\X],$ homogeneous of degree $2d_0-1>0$ such that $$\begin{array}{ccl} a(\F(\X))&=&C(\X)X_1,\\ b(\F(\X))&=&C(\X)X_2. \end{array}$$ As $C(\X)$ has positive degree, there are infinite points $(x_1:x_2:x_3)\in\P^2$ such that $C(\x)=0$. For those points we will have $a(\F(\x))=b(\F(\x))=0,$ but as $a(\T)$ and $b(\T)$ do not have common zeroes in $\P^1$, this then implies that the point $(x_1:x_2:x_3)\in\V(\F(\X))$, which contradicts the fact that $\V(\F(\X))$ is a complete intersection (hence finite). This shows that $b(\F(\X))X_1-a(\F(\X))X_2\neq0$. Suppose that $X_1l_2(\F(\X))+X_2l_3(F(\X)$ and $a(\F(\X))X_1-b(\F(\X))X_2$ have no common factors. Choose $(x_1:x_2:x_3)\in \P^2$ such that $b(\F(\x))x_1-a(\F(\x))x_2=0,$ with $(x_1:x_2:x_3)$ neither in $\V(\F(\X))$ nor in $\V(X_1l_2(\F(\X))+X_2l_3(F(\X)))$. By hypothesis, we still have an open set in $\V(a(\F(\X))X_2-b(\F(\X))X_1)$ to make such choices. From the first condition, we get $(x_1:x_2)=(a(\F(\x)):b(\F(\x)))$. From the second constraint we deduce that $a(\F(\x))l_2(\F(\x))+b(\F(\x))l_3(\x)\neq0.$ So, by using , we have that $$(x_1:x_2:x_3)=(u_1(\F(\x)):u_2(\F(\x)):u_3(\F(\x)))$$ and hence the point lies in the image of the parameterization. This can be done in an open set of this curve, and so it implies that $b(\F(\X))X_1-a(\F(\X))X_2$ defines $\C=V(E(\X))$. Algebraically we have that —up to a nonzero constant in $\K$— there exists $\nu\in\Z_{>0}$ such that $$\label{dee} b(\F(\X))X_1-a(\F(\X))X_2=E(\X)^\nu.$$ The polynomial on the left hand side has degree $2d_0+1$. By inspecting , and using the fact that $\gcd(a(\T),b(\T))=1$, we conclude that the degree of $\C$ (which is the degree of any proper parameterization of it) is equal to $$2d_0+1-\deg(\gcd(\u(\T)))=d_0+i,$$ with $0\leq i\leq d+1$. Computing degrees in we get $$2d_0+1=\nu(d_0+i).$$ This diophantine equation in $(\nu,i)$ has only two solutions: $\nu=1$ and $i=d_0+1$, i.e. there are no common factors, or $\nu=3,\,i=0$, which can only be possible if $d_0=1$. A quick glance at (\[paramm\]) may let the reader think that all curves parameterizable by conics have odd degree, but this is not always the case as $\deg(\gcd(\u(\T)))$ may be strictly positive. Also it is not true that all the curves parameterized by (\[paramm\]) pass through the point $(0:0:1)$ as the following cautionary example shows. \[ejj\] Set $F_1(\X):=X_1X_2-X_1X_3,\,F_2(\X):=X_1X_2-X_2X_3.$ We then have $l_1(\T)=T_1-T_2,\,l_2(\T)=T_2,\,l_3(\T)=-T_1,\,l_4(\T)=l_5(\T)=0.$ Set also $a(\T):=T_1^2,\,b(\T):=T_2^2$. We get $$\begin{array}{ccl} X_1l_2(\F(\X))+X_2l_3(\F(\X))&=&X_1X_2(X_1-X_2),\\ b(\F(\X))X_1-a(\F(\X))X_2&=&X_1X_2(X_1-X_2)(X_3^2-X_1X_2), \end{array}$$ and it is easy to see that the implicit equation of the curve defined by this data is given by $X_3^2-X_1X_2,$ which is a smooth conic. Note that $(0:0:1)$ is not a point of the curve. Next we will show that the case presented in Example \[ejj\] is somehow unusual in the sense that if $d_0>2$, then any curve being parameterized by actually passes through the point $(0:0:1)$ and moreover, if there is a common factor among the three polynomials defining the parameterization, then it has degree at most $2$. In order to show that, we present first a “canonical” form of the sequence $\{F_(\X),\,F_2(\X)\}$ which will depend on the geometry of $\V(\F(\X))$. \[reduccion\] Let $F_1(\X),\,F_2(\X)$ be a sequence of homogeneous forms of degree $2$ in $\K[\X]$ without common factors and such that the conic defined by ${\mathcal F}(\T,\X)$ is nondegenerate in $\P^2_{\overline{\K(\T)}}$. Assume also that $(0:0:1)\in\V(\F(\X))$. Then, after a linear change of coordinates in $\P^2,$ we can assume: $$\label{four} \F(\X)=(X_1X_2-X_2X_3,X_1X_3-X_2X_3) \quad \mbox{if}\ \|\V(\F(\X))\|=4,$$ $$\label{three} \F(\X)=(X_1X_2,X_1X_3-X_2X_3) \quad \mbox{if}\ \|\V(\F(\X))\|=3,$$ $$\label{twotwo} \F(\X)=(X_1^2,X_2X_3) \quad \mbox{if}\ \|\V(\F(\X))\|=2$$ and each of the points in $\V(\F(\X))$ has multiplicity two, $$\label{twothree} \F(\X)=(X_1^2-X_2X_3,X_1X_2) \quad \mbox{if}\ \|\V(\F(\X))\|=2$$ and one of the points in $\V(\F(\X))$ has multiplicity three, $$\label{one} \F(\X)=(X_1^2,\,X_2^2-X_1X_3) \quad \mbox{if}\ \|\V(\F(\X))\|=1.$$ This classification is classic and well-known in Projective Geometry, see for instance [@SK52 Chapter VII].[^2] \[boundegree\] Assuming the same hypothesis and notations of Proposition \[paramparam\], $\deg\big(\gcd(\u(\T))\big)\leq 3.$ Note that linear changes of coordinates in $\P^2$ amount to linear combinations of the $u_i(\T)$’s with coefficients in $\K$ which are invertible, i.e. one can use the canonical forms of the polynomials $F_1(\X),\,F_2(\X)$ given by Lemma \[reduccion\] without changing $\gcd(\u(\T))$. Note also that, as $a(\T),b(\T)$ have no common factors, then $$\gcd(\u(\T))=\gcd(l_2(\T)a(\T)+l_3(\T)b(\T),a(\T)b(\T)l_1(\T)+a(\T)^2l_4(\T)+b(\T)^2l_5(\T)).$$ In each of the cases described in Lemma \[reduccion\] we explicit the values of $l_i$ for $i=1,\dots ,5$ and bound the degree of the gcd. - In we have $$l_4(\T)=l_5(\T)=0,\, l_1(\T)=-T_2,\, l_2(\T)=T_1, \,l_3(\T)=T_2-T_1.$$ Hence, $\gcd(\u(\T))=\gcd(a(\T)T_1+b(\T)(T_2-T_1),a(\T)b(\T)T_2)$, and from here we can conclude that $\gcd(\u(\T))$ divides $T_1T_2(T_1-T_2)$. - In we have $$l_4(\T)=l_5(\T)=0,\, l_1(\T)=-T_2,\,l_2(\T)=T_1,\,l_3(\T)=-T_1.$$ In this case, $\gcd(\u(\T))=\gcd(a(\T)T_1-b(\T)T_1,a(\T)b(\T)T_2)$ divides $T_1T_2$. - In we have $$l_1(\T)=l_2(\T)=l_5(\T)=0, \,l_3(\T)=T_1,\,l_4(\T)=-T_2.$$ We get that $\gcd(\u(\T))=\gcd(b(\T)T_1,a(\T)^2T_2)$ divides $T_1T_2$. - In we have $$l_2(\T)=l_5(\T)=0,\, l_1(\T)=T_1,\, l_3(\T)=T_2,\, l_4(\T)=-T_2.$$ So, we deduce that $\gcd(\u(\T))=\gcd(b(\T)T_2,a(\T)b(\T)T_1-a(\T)^2T_2)$ divides $T_2$. - In we have $$l_1(\T)=l_3(\T)=0,\, l_2(\T)=-T_1,\,l_4(\T)=-T_2,\,l_5(\T)=T_1,$$ and we get that $\gcd(\u(\T))=\gcd(a(\T)T_1,b(\T)^2T_1-a(\T)^2T_2)$ divides $T_1$. In all of the cases, we get $\deg(\gcd(\u(\T)))\leq 3,$ which proves the claim. Now we can prove a complete version of Proposition \[paramparam\]. \[complete\] Let $F_1(\X),\,F_2(\X)$ be a sequence of quadratic forms in $\K[\X]$ without common factors such that $(0:0:1)\in\V(\F(\X))$ and ${\mathcal F}(\T,\X)$ defines a nondegenerate conic in $\P^2_{\overline{\K(\T)}}.$ For any $a(\T),\,b(\T)\in\K[\T]$ homogeneous of degree $d_0>2$ without common factors, either $a(\F(\X))X_2-b(\F(\X))X_1$ is an irreducible polynomial or it has a unique irreducible factor of degree larger than $1$. In both cases, this irreducible factor defines a rational curve $\C\subset\P^2$ parameterizable by $\F(\X)$ and passing through $(0:0:1)$. All the linear extraneous factors define equations of lines passing through the points of $\V(\F)$, and the degree of this factor is less than or equal to three. As shown in Proposition \[paramparam\], the pair $(a(\T),b(\T))$ defines the parameterization of a curve $\C$ parameterizable by $\F(\X)$. As $d_0>2,$ we then have $d_0+1>3$ and on the other hand if there is a nontrivial $\gcd(\u(\T))$ in , its degree -thanks to Proposition \[boundegree\]- cannot be larger than three. This shows that the factor $a(\T)l_2(\T)+b(\T)l_3(\T)$ cannot be completely cancelled when removing the $\gcd$ in , and hence $(0:0:1)$ is in the image of the parameterization. So, $\C$ passes through this point. If $\gcd(\u(\T))=1$, as the parameterization is proper, $a(\F(\X))X_2-b(\F(\X))X_1$ has the same degree as the curve $\C$. Hence, it is the irreducible polynomial defining it. Suppose then that this is not the case. Then there exist $H(\X)\in\K[X]$ homogeneous and coprime with $E(\X)$ such that $$\label{deee} a(\F(\X))X_2-b(\F(\X))X_1=E(\X)^\mu\,H(\X),$$ with $\mu\in\N,\,E(\X)$ being the irreducible polynomial defining $\C$. Let us say that $\deg(E(\X))=\varepsilon,\,\deg(H(\X))=\rho>0$. By computing degrees in , we get $$2d_0+1=\mu\,\varepsilon+\rho$$ Thanks to Proposition \[boundegree\], we know that $2d_0-2\leq \varepsilon\leq 2 d_0+1,$ so we have $\mu(2d_0-2)+\rho\leq 2d_0+1$. As $d_0>2,$ we can conclude from here that $\mu=1$. Moreover, we get that $\rho\leq 3,$ i.e. the degree of the extraneous factor $H(\X)$ is bounded. It remains to show that $H(\X)$ decomposes as a product of linear factors. The proof of Proposition \[paramparam\] actually shows that $$\V\big(a(\F(\X))X_2-b(\F(\X))X_1\big)\subset\V(E(\X))\cup\V(X_1l_2(\F(\X))+X_2l_3(\F(\X))),$$ and hence the factors of $H(\X)$ must be among the factors of $X_1l_2(\F(\X))+X_2l_3(\F(\X))$. One can show that in all the possible cases listed in Lemma \[reduccion\], the polynomial $X_1l_2(\F(\X))+X_2l_3(\F(\X))$ factorizes as a product of linear forms. Moreover, these linear forms can always be chosen in the set $\{X_1,\,X_2,\,X_1-X_2\},$ which are always lines passing through the points of $\V(\F).$ ### Examples {#example} Let $d_0\in\N$ and set $a(\T)=T_1^{d_0},\,b(\T):=T_2^{d_0}$. We will consider all the possible scenarios given by Lemma \[reduccion\]. - For $\F(\X)=(X_1X_2-X_2X_3,X_1X_3-X_2X_3)$, becomes $$\left\{\begin{array}{ccl} u_1(\T)&=&-T_1^{d_0} (T_1^{1 + d_0} - T_1 T_2^{d_0} + T_2^{1 + d_0})\\ u_2(\T)&=&-T_2^{d_0} (T_1^{1 + d_0} - T_1 T_2^{d_0} + T_2^{1 + d_0})\\ u_3(\T)&=&-T_1^{d_0} T_2^{1 + d_0}. \end{array}\right.$$ Note that $\gcd(\u(\T))=1$, hence $\C$ has degree $2d_0+1$. Computing explicitly the implicit equation we get $$E(\X)=X_2^{d_0+1} (X_1 - X_3)^{d_0} - X_1X_3^{d_0} (X_1 - X_2)^{d_0}.$$ - Set now $\F(\X)=(X_1X_2,X_1X_3-X_2X_3)$. The family $\u(\T)$ of is now $$\left\{\begin{array}{ccl} u_1(\T)&=&-T_1^{1 + d_0} (T_1^{d_0} - T_2^{d_0})\\ u_2(\T)&=&-T_1 T_2^{d_0} (T_1^{d_0} - T_2^{d_0})\\ u_3(\T)&=&-T_1^{d_0} T_2^{1 + d_0}. \end{array}\right.$$ Note that $\gcd(\u(\T))=T_1$ in this case, and hence $\deg(\C)=2d_0.$ Indeed, an explicit computation shows that $$a(\F(\X))X_2-b(\F(\X))X_1=X_1\big(X_1^{d_0-1}X_2^{d_0+1}-X_3^{d_0}(X_1-X_2)^{d_0}\big),$$ hence the implicit equation is defined by $X_1^{d_0-1}X_2^{d_0+1}-X_3^{d_0}(X_1-X_2)^{d_0}.$ Note that in this case $$\gcd(X_1l_2(\F(\X))+X_2l_3(F(\X)), a(\F(\X))X_2-b(\F(\X))X_1)=X_1,$$ (cf. Proposition \[paramparam\]). - Set now $\F(\X)=(X_1^2,X_2X_3)$.Then, $$\left\{\begin{array}{ccl} u_1(\T)&=&-T_1^{1 + d_0} T_2^{d_0}\\ u_2(\T)&=&-T_1 T_2^{2 d_0}\\ u_3(\T)&=&-T_1^{2 d_0} T_2, \end{array}\right.$$ with $\gcd(\u(\T))=T_1T_2$. Hence, $\deg(\C)=2d_0-1$ and computing explicitly $a(\F(\X))X_2-b(\F(\X))X_1$ we get that it is equal to $X_1X_2\,E(\X)$, with $$E(\X)=X_1^{2d_0-1} - X_2^{d_0-1} X_3^{d_0}.$$ - For $\F(\X)=(X_1^2-X_2X_3,X_1X_2)$, we have $$\left\{\begin{array}{ccl} u_1(\T)&=&-T_1^{d_0} T_2^{1+d_0}\\ u_2(\T)&=&-T_2^{1+2 d_0}\\ u_3(\T)&=&-T_1^{d_0+1} T_2(T_1^{d_0-1} - T_2^{d_0-1}), \end{array}\right.$$ with $\gcd(\u(\T))=T_2$. So, $\deg(\C)=2d_0$ and $a(\F(\X))X_2-b(\F(\X))X_1$ is equal to $X_2\,E(\X)$ with $$E(\X)= (X_1^2 - X_2 X_3)^{d_0} -X_1^{1+d_0} X_2^{d_0-1}.$$ - Finally, consider $\F(\X)=(X_1^2,\,X_2^2-X_1X_3).$ By computing explicitly, we get $$\left\{\begin{array}{ccl} u_1(\T)&=&T_1^{1+2d_0} \\ u_2(\T)&=&T_1^{1+ d_0}T_2^{d_0}\\ u_3(\T)&=&T_1T_2(T_2^{2d_0-1}-T_1^{2d_0-1}). \end{array}\right.$$ Here, we have $\gcd(\u(\T))=T_1$. Again we get $\deg(\C)=2d_0$ and $$a(\F(\X))X_2-b(\F(\X))X_1=X_1\,E(\X)$$ with $E(\X)= X_1^{2d_0-1}X_2 - (X_2^2 - X_1 X_3)^{d_0}.$ The geometry of curves parameterizable by conics {#geometry} ================================================ In this section, we will study geometric properties of plane curves parameterizable by conics. We will show that essentially they are the image of a curve parameterizable by lines via a quadratic transformation of the plane. Quadratic transformations in the plane -------------------------------------- \[qt\] A rational map $\Lambda:\P^2\dasharrow\P^2$ is called a quadratic transformation if $\Lambda$ is birational and there exist $Q_1(\X),\,Q_2(\X),\,Q_3(\X)\in\K[\X]$ homogeneous of degree $2$ without common factors such that $$\label{Lambda}\Lambda(x_1:x_2:x_3)=\big(Q_1(\x):Q_2(\x):Q_3(\x)\big).$$ One of the most well-known of these quadratic transformations is the following $$\label{cremona} \Lambda_{\bf0}(x_1:x_2:x_3)=(x_2x_3:x_1x_3:x_1x_2),$$ which is used for desingularization of curves, see for instance [@walker]. Even though there are birational automorphisms of $\P^2$ defined by homogeneous forms of arbitrary degree, we will focus here in those of degree $2,$ as they will be crucial when studying curves parameterizable by conics. \[quadratic\] Let $F_1(\X),\,F_2(\X)\in\K[\X]$ be a sequence of homogeneous forms of degree $2$ without common factors. If the conic defined by ${\mathcal F}(\T,\X)$ in $\P^2_{\overline{\K(\T)}}$ is nondegenerate, then there exists $F_3(\X)\in\K[\x]$ homogeneous of degree $2$ such that $$\label{Lambdaf} \begin{array}{cccc} \Lambda_F:&\P^2&\dasharrow&\P^2\\ &(x_1:x_2:x_3)&\mapsto&\big(F_1(\x):F_2(\x):F_3(\x)\big) \end{array}$$ is a quadratic transformation. Moreover, $\Lambda_F^{-1}$ is also a quadratic transformation. In characteristic zero, it is well-known that a birational transformation given by polynomials of degree $n$ has an inverse also given by forms of the same degree, see for instance [@alberich]. We will use the canonical forms given in Lemma \[reduccion\] in order to make explicit the polynomial $F_3(\X)$ in each of the possible cases. 1. If $\|\V(\F(\X)\|\geq3,$ we can suppose w.l.o.g. that $$\{(1:0:0),\,(0:1:0),\,(0:0:1)\}\subset\V(\F(\X)),$$ and hence by using or , it is easy to see that if we set $F_3(\X):=X_2X_3$, $\Lambda_F$ is actually the classical transformation $\Lambda_{\bf0}$ composed with an automorphism of $\P^2$. As $\Lambda_{\bf0}^{-1}=\Lambda_{\bf0},$ it is easy to see that $\Lambda_F^{-1}$ can be defined with linear combinations of $X_1X_2,\,X_1X_3,\,X_2X_3$, hence it is a quadratic transformation. 2. If $\|\V(\F(\X))\|=2$, each point with multiplicity two, then by using we can assume w.l.o.g. that $$\F(\X)= (X_1^2,\,X_2X_3).$$ We set $F_3(\X):=X_1X_2$ and get $$\Lambda_F^{-1}(y_1:y_2:y_3)=\big(y_1y_3: y_3^2: y_1y_2\big),$$ hence $\Lambda_F$ is birational with quadratic inverse. 3. If $\|\V(\F(\X))\|=2$ and one of the points in this set has multiplicity three, then by we can assume after a linear change of coordinates that $$\F(\X)= ( X_1^2-\,X_2X_3, X_1X_2).$$ Setting $F_3(\X):=X_2 X_3$ we get that $$\Lambda_F^{-1}(y_1:y_2:y_3)=\big(y_2(y_1+ y_3):y_2^2:y_3(y_1+ y_3)\big).$$ Hence, $\Lambda_F$ is birational and the inverse is quadratic, as claimed. 4. If $\{(0:0:1)\}=\V(\F(\X))$. We then use and suppose w.l.o.g. that $$\F(\X)=( X_1^2,\,X_2^2-X_1X_3).$$ Once more, by setting $F_3(\X):=X_1X_2$, we get $$\Lambda_F^{-1}(y_1:y_2:y_3)=\big(y_1^2: y_1y_3:y_3^2-y_1y_2\big),$$ so we conclude that $\Lambda_F$ is birational with quadratic inverse. This completes the proof. \[zulema\] For any curve $\C_0$ of degree $d^0>1$ parameterizable by lines, having $(0:0:1)\in\C_0$ with multiplicity $d^0-1$, and any quadratic transformation $\Lambda:\P^2\dasharrow\P^2$ whose inverse is defined by a sequence of quadratic forms $(F_1(\X),\,F_2(\X),\,F_3(\X)),\ \overline{\Lambda(\C_0)}$ is a curve parameterizable by $(F_1(\X),F_2(\X))$. Set $\C=\overline{\Lambda(\C_0)}.$ The fact that $\C_0$ is not a line implies that $\dim(\C)=1$. As $\C_0$ is parameterizable by $(X_1,X_2)$, then it is easy to verify then that $\C$ is parameterizable by $(F_1,F_2)$. We are not claiming in Lemma \[zulema\] that the first two coordinates of a quadratic transformation have a non trivial common factor. Also, it is not necessarily true that the image of a curve parameterizable by lines via a quadratic transformation cannot be a parameterizable by lines anymore. For instance, $\Lambda_{\bf0}$ has $\F(\X)=(X_2X_3,\,X_1X_3)$ which has $X_3$ as a common factor. Also, if $\C_0$ is any curve parameterizable by lines having its singularity at $(0:0:1)$, then it is easy to check that $\overline{\Lambda_{\bf0}(\C_0)}$ is again a curve parameterizable by lines having its singularity at the same point. Moreover, not necesarily the first two coordinates of a quadratic transformation define a polynomial ${\mathcal F}(\T,\X)$ whose set of zeroes in $\P^2_{\overline{\K(\T)}}$ is a nondegenerate conic, for instance $\Lambda(x_1:x_2:x_3):=(x_1^2:x_1x_2:(x_1+x_2)x_3)$ is a quadratic transformation with inverse $\Lambda^{-1}(x_1:x_2:x_3)=(x_1(x_1+x_2):x_2(x_1+x_2):x_1x_3)),$ but the conic defined by $T_2X_1^2-T_1X_1X_2$ is degenerate according to Proposition \[prop22\]. Now we proceed to compare the degrees of $\C_0$ and its transform $C=\overline{\Lambda(C_0)}$. We start with the following result, which will be of use in the sequel. \[auxxi\] Let $Q_1(\X),\,Q_2(\X),\,Q_3(\X)\in\K[\X]$ be a sequence of homogeneous quadratic forms such that $\Lambda$ defined in is a quadratic transformation, and $\C\subset\P^2$ any curve of degree $d$. Let $\C_Q$ be a generic conic in the linear system defined by $Q_1(\X),\,Q_2(\X),\,Q_3(\X).$ Then, for any point $p\in\C\cap\C_Q,$ we have $$m_p(\C\cap\C_Q)\leq d.$$ Moreover, the inequality is strict if $\C_Q$ and $\C$ do not have a common tangent at $p$. Suppose w.l.o.g. that $p=(0:0:1).$ A generic linear combination of the $Q_i(\X)$’s must have a non-zero linear term with respect to $X_3$ otherwise those three polynomials would depend only on $X_1$ and $X_2$ contradicting the fact that $\Lambda$ is a birational. This implies that $m_p(\C_Q)=1$. On the other hand, we always have $m_p(\C)<d$. If $\C$ and $\C_Q$ intersect transversally at $p,$ then we have (cf. [@HKT08 Proposition $3.6$]) $$m_p(\C\cap\C_Q)=m_p(\C)< d.$$ In case they do not intersect transversally, as $\C_Q$ has a tangent line $L_Q$ having multiplicity one at $p$, then we have $$m_p(\C\cap\C_Q)=m_p(\C\cap L_Q)\leq d,$$ the last inequality is due to Bézout’s Theorem applied to $\C$ and $L_Q$. \[num\] With notations and assumptions as in Lemma \[zulema\], denoting with $D^0$ the degree of $\overline{\Lambda(\C_0)}$, then we have $d^0-1\leq D^0\leq 2d^0, $ and the inequalities are sharp. As before, set $\C=\overline{\Lambda(\C_0)}.$ Its degree can be computed as the cardinality of $\C\cap L$, with $L$ a generic line in $\P^2,$ which we will choose as intersecting $\C$ in the (dense) open set of $\P^2$ where $\Lambda$ is bijective. As $\Lambda$ is birational, then we can compute this intersection number via $\Lambda^{-1}$. Then, $\C$ gets converted into $\C_0$ and $L$ in a generic linear combination of the quadratic polynomials $F_1(\X),\,F_2(\X),\,F_3(\X)$. We use then Bézout’s Theorem in order to count the number of intersections between $\C_0$ and the conic $\Lambda^{-1}(L)$ to get $$\label{numeromania} 2d^0=D^0+\sum_{p\in\V(F_1,F_2,F_3)}m_p(\C_0\cap\Lambda^{-1}(L)).$$ As the data $(F_1(\X),\,F_2(\X),\,F_3(\X))$ defines a birational transformation, it is easy to see that $\|\V(F_1,F_2,F_3)\|\leq 3.$ Moreover, the scheme of points defined by $\F(\X)$ in $\P^2$ must have degree less than or equal to three, otherwise one of these polynomials would be a linear combination of the others contradicting the fact that $\Lambda$ is a quadratic transformation. We have in addition that $\C_0$ is parameterizable by lines. This implies that there is one point of multiplicity $d^0-1$ and the remaining have multiplicity one. Hence, thanks to Lemma \[auxxi\], we have $$0\leq \sum_{p\in\V(F_1,F_2,F_3)}m_p(\C_0\cap\Lambda^{-1}(L))\leq \left\{\begin{array}{l} 1+1+(d^0-1)\\ 1+d^0,\end{array} \right.=d^0+1.$$ The first case is when $\V(\F(\X))$ has three points, hence there cannot be fixed tangential conditions in the linear system and this implies that we can choose the generic line in such a way that $\Lambda^{-1}(L)$ cuts transversally $C_0$; the second case is when the linear system defined by $\F(\X)$ has a fixed tangential condition. But then, we have that $\V(\F(\X))$ cannot have more than two points, and by using Lemma \[auxxi\] we are done. From here plus , we get the bounds of the claim. Now we will show that the bounds are sharp. For a generic quadratic transformation $\Lambda,$ we will have $\sum_{p\in\V(F_1,F_2,F_3)}m_p(\C_0\cap\Lambda^{-1}(L))=0$. Indeed, one only has to pick $(F_1,F_2,F_3)$ in such a way that $\V(\F(\X))\cap\C_0=\emptyset.$ So, the inequality at the left is generically an equality. In order to show that the other inequality can also become an equality, let $d_0>1$ and consider the following parameterization $$\left\{\begin{array}{ccl} u_1(\t)&=&t_1\,\alpha(\t),\,\\ u_2(\t)&=&t_2\,\alpha(\t),\,\\ u_3(\t)&=&t_1t_2\beta(\t) \end{array}\right.$$ with $\alpha(\T),\,\beta(\T)$ homogeneous of degrees $d_0-1$ and $d_0-2$ without common factors and also without common factors with neither $T_1$ nor $T_2$. Then, the curve $\C_0$ defined as the image of this parameterization is parameterizable by lines of degree $d_0$ with $p=(0:0:1)$ having multiplicity $d_0-1.$ Consider $\Lambda_{\bf0}$ defined in . Then, an straightforward computation shows that a proper parameterization of $\Lambda_0(\C_0)$ is given by $$\left\{\begin{array}{ccl} v_1(\t)&=&t_2\,\beta(\t),\,\\ v_2(\t)&=&t_1\,\beta(\t),\,\\ v_3(\t)&=&\alpha(\t); \end{array}\right.$$ i.e. $\Lambda_0(\C_0)$ is a curve of degree $d_0-1$. Note that this curve is either a line or again parameterizable by lines. We can now describe geometrically the plane curves parameterizable by conics via quadratic transformations of curves parameterizable by lines. Recall that thanks to Proposition \[prop22\], if $T_1F_2(\X)-T_2F_1(\X)$ defines a degenerate conic in $\P^2_{\overline{\K(\T)}},$ then any curve parameterizable by $\F(\X)$ is either a line or parameterizable by lines. Also, curves of degree $2$ are parameterizable by lines. \[mtmt\] Let $F_1(\X),\,F_2(\X)$ be sequence of homogeneous forms of degree $2$ without common factors such that $(0:0:1)\in\V(\F(\X))$ and $T_1F_2(\X)-T_2F_1(\X)$ does not define a degenerate conic in $\P^2_{\overline{\K(\T)}}$. Consider any quadratic transformation of the form $\Lambda_F$ defined in . A curve $\C$ such that $\deg(\C)\geq3$ is parameterizable by $\F(\X)$ if and only if there exist $\C_0$ parameterizable by lines having $(0:0:1)$ as its only singular point and $\overline{\Lambda_F(\C)}=\C_0$. As $T_1F_2(\X)-T_2F_1(\X)$ defines a nondegenerate conic in $\P^2_{\overline{\K(\T)}}$, we can find a quadratic transformation $\Lambda_F$ as in Proposition \[quadratic\]. Set $\C_0$ to be the Zariski closure of $\Lambda_F(\C)$ in $\P^2$. By Proposition \[num\], we have that $\deg(\C_0)\geq \frac{\deg(\C)}{2}>1$, hence $\C_0$ is not a line. We can also verify easily that $\C_0$ is parameterizable by $(X_1,X_2)$, hence it is parameterizable by lines and having $(0:0:1)\in\C_0$ with maximal multiplicity. In order to prove the converse, if we start with $\C_0$ as in the hypothesis and define $\C$ to be the Zariski closure of $\Lambda_F(\C_0),$ we can easily verify that $\C$ is parameterizable by $\F(\X)$. On the singularities of curves parameterizable by conics -------------------------------------------------------- There is an increasing interest in the analysis of singularities of rational curves by means of elements of small degree in the Rees Algebra of the parameterization, see for instance [@CKPU11]. Theorem \[mtmt\] above shows that curves parameterizable by conics are only “one quadratic transformation away” from curves parameterizable by lines, and in principle it may seem that the study of their singularities can be done straightforwardly, as for instance the transformation $\Lambda_{\bf0}$ defined in is the one used in the process of desingularization of curves. The main drawback here is that —as Theorem \[mtmt\] claims— a curve parameterizable by conics is the image of a curve parameterizable by lines with singular point in $(0:0:1)$ via any quadratic transformation, and $\Lambda_{\bf0}$ is known to “behave properly” if the curve is in a general position with respect to the coordinate axes (cf. the notions of “good” and “excellent” positions in [@ful69]). So, even if we use $\Lambda_{\bf0}$ to transform $\C$ into a curve parameterizable by lines, we cannot expect to get a straightforward dictionary between the only singularity of the curve parameterizable by lines and those of $\C$. The analysis of the singularities of these curves require a further study of properties of general quadratic transformations, which goes beyond the scope of this article. One case which is easy to tackle is when $\|\V(\F(\X))\|=4$, We will show that in this situation, all of the four points are multiple points of $\C$ and moreover, there are no infinitely near multiple points. We start by analyzing the only singularity of a curve parameterizable by lines. \[geomonoid\] Let $\C$ be a curve parameterizable by lines having $(0:0:1)\in\C$ with multiplicity $\deg(C)-1$, and implicit equation given by the polynomial $b(X_1,X_2)+X_3\,a(X_1,X_2)\in\K[\X]$, with $a(\T),\,b(\T)$ homogeneous elements of degrees $d-1$ and $d$ respectively. Write $$a(\T)=\cc_0\prod_{j=1}^{\tau}(\dd_jT_2-\ee_jT_1)^{\nu_j},$$ with $\cc_0\in\K\setminus\{0\},\,(\dd_j:\ee_j)\neq(\dd_k:\ee_k)$ if $j\neq k$, and $\nu_j\in\N$ for $j=1,\ldots, s$. Then, 1. there are $\tau$ different branches of $\C$ passing through $(0:0:1)$; 2. denote with $\gamma_j$ the branch of $\C$ at $\phi((\dd_j:\ee_j))$, here $\phi(t_1:t_2)$ is the parameterization of $\C$ given by . The tangent to $\gamma_j$ at $(t_1:t_2)=(\dd_j:\ee_j)$ is the line $\dd_jX_2-\ee_jX_1=0$. In particular, different branches have different tangents (i.e. there are no tacnodes); 3. the order of contact of $\C$ with the tangent line $\dd_jX_2-\ee_jX_1=0$ at $(0:0:1)$ is equal to $\nu_j+1$; 4. the multiplicity of $(0:0:1)$ in $\C$ is $d-1$, and there are no infinitely near multiple points of $\C$. The first three items follow straightforwardly from working out the parameterization in a neighborhood of the zeroes of $a(\T),$ plus the fact that for this proper parameterization we have $T_2u_1(\T)-T_1u_2(\T)=0$. In order to conclude, recall that $(0:0:1)$ is a point of multiplicity $d-1$. The genus formula shows that there cannot be no more singular points in $\C$. The following result about curves and rational maps is well-known. We record it here for the convenience of the reader. Denote with $\mbox{Sing}(C)$ the set of singular points of $\C$ in $\P^2$. \[silverman\] If $\C$ is parameterizable by $(F_1(\X),F_2(\X)),$ then $\mbox{Sing}(\C)\subset \V(\F(\X)).$ Let $\phi$ be as in a proper parameterization of $\C$ having as its inverse $\psi=(F_1:F_2)$ whenever it is defined, as in . As $\phi$ is defined on the whole $\P^1$, from $\psi\circ\phi=id_{\P^1},$ we have $$\label{identidad} \big(F_1(\underline{\phi}(t_1:t_2)):F_2(\underline{\phi}(t_1:t_2))\big)=(t_1:t_2)\ \ \mbox{for } \underline{\phi}(t_1:t_2)\notin\V(\F).$$ If $p=\phi(t_{01}:t_{02})\in\C$ is a singular point, and suppose that $(F_1(p):F_2(p))=(t_{01}:t_{02}),$ We then have two possible scenarios: - If the gradient of $\phi$ at $(t_{01}:t_{02})$ is equal to zero, by differentiating both sides of and specializing $(t_1:t_2)\mapsto(t_{01}:t_{02})$ we would get a contradiction. - If the gradient of $\phi$ is not zero at $(t_{01}:t_{02})$, then there must be another point $(t_{11}:t_{12})\in\P^1$ such that $\phi(t_{11}:t_{12})=p$ (i.e. the curve “passes” at least twice over $p$). But then, we will have $$\big(F_1(\underline{\phi}(t_{11}:t_{12})):F_2(\underline{\phi}(t_{11}:t_{12}))\big)=(t_{01}:t_{02})\neq(t_{11}:t_{12}),$$ a contradiction with . This shows that for such a singular point $p\in\C, \,\F(p)=(0,0)$ which proves the claim. \[4points\] Let $F_1(\X),\,F_2(\X)$ be a sequence of forms of degree $2$ in $\K[\X]$ without common factors, such that the conic defined by $T_1F_2(\X)-T_2F_1(\X)$ is nondegenerate in $\P^2_{\overline{\K(\T)}}$. If $\C$ is parameterizable by $\F(\X),$ not parameterizable by lines, and $\V(\F(\X))$ has four points, then $\V(\F(\X))=\mbox{Sing}(\C)$ and at each $p\in \V(\F(\X))$, $p\in C$ is locally isomorphic to the singular point of a curve parameterizable by lines. Thus $p$ is not a tacnode and has no infinitely near singular points. Hence $$\label{aguap} \sum_{p\in \V(\F(\X))}m_p(\C)(m_p(\C)-1)=(d-1)(d-2).$$ After a linear change of coordinates, we may assume that we are in the conditions of and hence $$\{(1:0:0),\,(0:1:0),\,(0:0:1),\,(1:1:1)\}=\V(\F(\X)).$$ As in the proof of Proposition \[quadratic\], by setting $F_3(\X):=X_2X_3,$ we get a quadratic transformation $\Lambda_F$ which is actually the composition of $\Lambda_{\bf0}$ with an automorphim of $\P^2$. It is easy to check that $p_0=(1:1:1)$ is not in the union of lines where $\Lambda_F$ is not invertible. Hence, in neighborhood of this point, $\Lambda_F$ is actually an algebraic isomorphism. As $\overline{\Lambda_F(\C)}$ is not a line (due to the fact that $\deg(\C)>2,$ otherwise it would be parameterizable by lines), and is parameterizable by $(X_1,X_2)$ with only singularity in $(0:0:1)=\Lambda_F(p_0)$, then properties (1) to (3) in Proposition \[geomonoid\] apply to $p_0$ with respect to $\C,$ due to the fact that $\Lambda_F$ is a local isomorphism around $p_0$ and its image. For the same reason, the fact that there are no infinitely near multiple points of $\overline{\Lambda_F(\C))}$ above $\Lambda_F(p_0)$ (this is property (4) in Proposition \[geomonoid\]) implies that there cannot be infinitely near multiple points of $\C$ above $p_0$. Making a linear change of coordinates, the role of $(1:1:1)$ can be played by the other three points of $\V(\F(\X))$, and this implies the claim for the other three points, i.e. we have shown $\V(\F(\X))\subset\mbox{Sing}(\C)$). The other inclusion follows by Lemma \[silverman\], hence we have the equality. As there cannot be more singular points, and none of the elements in $\V(\F(\X))$ has infinitely near multiple points of $\C$ above it, follows due to the genus formula. Note that the Theorem does not claim that the four singular points have the same multiplicity and character. Just that they “look like” (locally) like a multiple point in a curve parameterizable by lines. This curve is not necessarily the same for all the points, as the following example shows. Let $\C$ be the rational curve of degree $5$ defined by the polynomial $E(\X)=X_2^3(X_1-X_3)^2-X_1X_3^2(X_1-X_2)^2$ (this is the first bullet of Example \[example\] with $d_0=2$). Its four singular points are $(1:0:0),\,(0:1:0),\,(0:0:1)$ and $(1:1:1)$. By analyzing them explicitly, we get that - $(0:0:1)$ is an ordinary triple point; - $(0:1:0)$ and $(1:0:0)$ are cusps; - $(1:1:1)$ is an ordinary triple point. We can straightforwardly verify equality in this case: $$12=(5-1)(5-2)=3\times2+2\times1+2\times1+2\times1.$$ We have thus completed our study of the singularities of $\C$ in the case $\|\V(\F(\X))\|=4,$ which is somehow the generic case among curves parameterizable by conics. Now we turn into the question of how the singularities look like in the remaining cases. We will see in Section \[s1\] (Corollary \[maxmu\]) that the value of $\mu$ in is always equal to $\lfloor \frac{d}2\rfloor$. This information is enough to show that if $\|\V(\F(\X))\|\leq 3,$ then $\C$ has always infinitely near singular points if $\deg(\C)>6$. If $\C$ is parameterizable by a sequence of forms $(F_1(\X),\,F_2(\X))$ of degree $2$ without common factors, with $d=\deg(\C)>6$, and $\|\V(\F(\X))\|\leq3,$ then $\C$ has infinitely near singular points. If there are no infinitely near multiple points, due to the genus formula, we will have $$(d-1)(d-2)=\sum_{p\in\C}m_p(\C)\big(m_p(\C)-1\big).$$ In [@CWL08 Theorem 1], it is shown that there can only be one multiple point of multiplicity larger than $\mu$. Moreover, if this is the case, then the multiplicity of this point is actually $d-\mu$. Suppose then that $\|\V(\F(\X))\|\leq3.$ As $\mbox{Sing}(\C)\subset\V(\F(\X))$ (cf. Lemma \[silverman\]), we then conclude that there are at most $3$ singular points. One of them has its multiplicity bounded by $d-\mu\leq \frac{d+1}{2}$ and the other two have both multiplicities bounded by $\frac{d}{2}$. Hence, we get $$\begin{array}{ccl} 0&=&\sum_{p\in\C}m_p(\C)\big(m_p(\C)-1\big)-(d-1)(d-2)\\ \\ &\leq &(\frac{d+1}{2}\frac{d-1}{2}+2\ \frac{d}{2}\frac{d-2}{2})-(d-1)(d-2)\\ \\ &=&-\frac{d^2-8d+9}{4}. \end{array}$$ For $d\geq7$, the last expression is negative. This concludes the proof. The Rees Algebra of a rational parameterization {#s1} =============================================== Now we turn to the problem of computing a set of minimal generators for the presentation of the Rees Algebra associated to the ideal of a rational parameterization of a curve parameterizable by conics. This section may be considered an extension of the results given in [@CD10] (see also [@bus09; @CHW08]) for curves parameterizable by lines. Let $I$ be the ideal of $\K[T_1,T_2]$ generated by three homogeneous polynomials $u_1(T_1,T_2),\,u_2(T_1,T_2),\,u_3(T_1,T_2)$ of degree $d$ without common factors. Recall that $\mbox{Rees}(I)=\K[\T][I\,Z]$ is the Rees Algebra associated to $I.$ Let $\kK\subset R[\X]$ be the kernel of the graded morphism of $\K[\T]$-algebras ${\mathfrak h}$ defined in . It is a bigraded ideal (with grading given by total degrees in $\T$ and $\X$) characterized by $$P(\T,\X)\in \kK_{i,j}\iff \bdeg(P)=(i,j)\ \mbox{and} \ P(\T,\u(\T))=0.$$ Let $\phi:\P^1\to\P^2$ be the map given by (\[param\]), and set as before $\C:=\phi(\P^1)$. As we observed in Section \[uno\], $\phi$ admits a rational inverse $\psi:\C\dasharrow\P^1$ if and only if there exists an irreducible nonzero element in $\kK_{1,}:=\oplus_{j=0}^\infty \kK_{1,j}.$ Moreover, if $F_1(\X),\,F_2(\X)$ are coprime elements in $\K[\X]$ and $T_1F_2(\X)-T_2F_1(\X)\in\kK_{1,\nu},$ then $\F(\X)$ defines the inverse of $\phi$. In the terminology of [@cox08; @bus09], $\kK$ is the moving curve ideal* of the parameterization $\phi$. An element in $\kK_{\ast,j}$ is called a moving curve of degree $j$ that follows the parameterization. In this sense, moving lines that follow the parameterization are the elements of $\kK_{\ast,1}$ and there is an obvious isomorphism of $\K[\T]$-modules $$\label{corresp} \begin{array}{ccc} \kK_{,1}&\to&\mbox{Syz}(\I)\\ a(\T)X_1+b(\T)X_2+c(\T)X_3&\mapsto&\big(a(\T),b(\T),c(\T)\big). \end{array}$$ Recall from Section \[uno\] that the first module of syzygies of $I$ is a free $\K[\T]$-module generated by two elements, one in degree $\mu$ for a positive integer $\mu$ such that $0<\mu\leq\frac{d}{2}$, and the other of degree $d-\mu$. Such a basis is called a $\mu$-basis. In the sequel, we will denote with $p_{\mu ,1}(\T,\X),\,q_{d-\mu ,1}(\T,\X)\in \kK_{,1}$ a (chosen) set of two elements in $\mbox{Syz}(\I)$ which are a basis of this module. Note that with this language, we can say that there exists an irreducible element in $\kK_{1,1}$ if and only if $\C$ is parameterizable by lines, and this is equivalent also to $\mu =1$. We will see (for $d>3$) that if there exists an irreducible element in $\kK_{1,2}$ (that is $\C$ is parameterizable by conics and not by lines, cf. Proposition \[d’\]) then $\mu =\lfloor\frac{d}{2}\rfloor$. Before that, we present two results that will be useful in the sequel. The first of them is the analogue of Proposition $2.6$ in [@CD10]. \[multiple\] Suppose $T_1F_2(\X)-T_2F_1(\X)\in\kK$ is an irreducible polynomial. Then, $P_{i,j}(\T,\X)\in\kK_{i,j}$ if and only if $P_{i,j}(F_1(\X),F_2(\X),\X)$ is a multiple of $E(\X)$. We only have to show that $P_{i,j}(F_1(\X),F_2(\X),\X)$ vanishes on $\C$ if and only if $P_{i,j}(\T,\X)\in\kK_{i,j}$. Taking into account that $\C=\{\u(\t) \mid (t_1:t_2)\in \P^1\},$ and that $(t_1:t_2)=(F_1(\u(\t)):F_2(\u(\t))$ for almost all $(t_1:t_2)\in \P^1$, then $$\begin{array}{l} \\ P_{i,j}(F_1(\x),F_2(\x),\x)=0 \textrm{ for all } (x_1:x_2:x_3)\in \C \Leftrightarrow \\ \\ \Leftrightarrow P_{i,j}(F_1(\u(\t)),F_2(\u(\t)),\u(\t))=0 \textrm{ for all } (t_1:t_2)\in \P^1 \Leftrightarrow \\ \\ \Leftrightarrow P_{i,j}(F_1(\u(\t)),F_2(\u(\t)),\u(\t))=0 \textrm{ for almost all } (t_1:t_2)\in \P^1 \Leftrightarrow \\ \\ \Leftrightarrow P_{i,j}(t_1,t_2,\u(\t))=0 \textrm{ for almost all } (t_1:t_2)\in \P^1\Leftrightarrow\\ \\ \Leftrightarrow P_{i,j}(t_1,t_2,\u(\t))=0 \textrm{ for all } (t_1:t_2)\in \P^1\Leftrightarrow P_{i,j}(\T,\X)\in\kK_{i,j}. \end{array}$$ The following proposition is the analogue of Lemma $2.7$ in [@CD10]. \[lem1\] Suppose $F_1(\X),\,F_2(\X)$ are homogeneous of degree $j_0$. Let $P(\T,\X)$ be a bihomogeneous polynomial of bidegree $(i,j)\in\N^2$, with $i>0,\,j\geq j_0$. Then there exists $Q(\T,\X)$ bihomogeneous of bidegree $(i-1,(i-1)j_0+j)$ such that $$\label{connect} F_2(\X)^iP(\T,\X)-T_2^iP(F_1(\X),F_2(\X),\X)=(T_1F_2(\X)-T_2F_1(\X))Q(\T,\X).$$ $$\begin{array}{l} F_2(\X)^iP(\T,\X)-T_2^iP(F_1(\X),F_2(\X),\X)=\\ =P(T_1F_2(\X),T_2F_2(\X),\X)-P(T_2F_1(\X),T_2F_2(\X),\X). \end{array}$$ By applying on the polynomial $p(\theta):=P(\theta,T_2F_2(\X),\X)$ the first order Taylor formula, the claim follows straightforwardly. Assume that $\phi$ defined as in , is a proper parameterization of a curve of degree $d$. Let $\nu$ be the degree of a homogenous pair of polynomials in $\K[\X]$ defining the inverse of $\phi$. If $\mu>1$ then $$\mu \, \nu +1\geq d.$$ Let $p_{\mu,1}(\T,\X)$ be a nonzero element in $\kK_{\mu,1}$. Due to Proposition \[multiple\], we have that $p_{\mu,1}(F_1(\X),F_2(\X),\X)$ is a multiple of $E(\X)$, which has degree $d$. As $\deg(p_{\mu,1}(F_1(\X),F_2(\X),\X))=\mu \, \nu +1$, it turns out that if $\mu \, \nu +1<d,$ then $$p_{\mu,1}(F_1(\X),F_2(\X),\X)=0.$$ By (\[connect\]), we then have $F_2(\X)p_{\mu,1}(\T,\X)\in\langle T_1F_2(\X)-T_2F_1(\X)\rangle,$ which is a prime ideal and clearly $F_2(\X)$ does not belong to it. So we conclude that $p_{\mu,1}(\T,\X)$ is a multiple of $T_1F_2(\X)-T_2F_1(\X)$ which is impossible unless $\deg(F_1(\X))=\deg(F_2(\X))=1$ which is equivalent to $\mu=1$. \[maxmu\] If $\nu=2$ (i.e. $\phi$ parameterizable by conics) and there are no linear syzygies, then $\mu=\lfloor\frac{d}{2}\rfloor$, the maximum possible value. It was shown already in [@bus09] that for $\mu\geq2$ the description of generators of $\kK$ is much more complicated than in the case of curves parameterizable by lines, so there is little hope that the elementary methods applied in [@CD10] can be used in these cases. Next we will show that instead of looking at low degrees for the syzygies of $\phi$, if we try low degrees for the inverse of $\phi$, that the approach of [@CD10] can be adapted, and indeed produces a minimal set of generators of rational plane curves parameterizable by conics (i.e. , the degree of the inverse is equal to $2$). We start by recalling the following: \[1.1\] The sequence $p_{\mu,1}(\T,\X),\,q_{d-\mu, 1}(\T,\X)$ is regular in $\K[\T,\X]$ and $$\kK= \bigcup_{n\geq 0} \langle p_{\mu,1}(\T,\X),\,q_{d-\mu, 1}(\T,\X)\rangle:\langle T_1,T_2 \rangle^n.$$ As explained in [@bus09 Section $2$], in order to search for a set of generators of $\kK$, it is enough to consider forms of $\T$-degree lower than $d$. Our next result is a refinement of this bound, which essentially states that we can replace $d-1$ by $d-\mu$. \[mu\] Let $u_1(\T),\,u_2(\T),\,u_3(\T)\in\K[\T]$ be homogeneous polynomials of degree $d$ having no common factors. A minimal set of generators of $\kK$ can be found with all its elements having $\T$-degree strictly less than $d-\mu$ except for the generators of $\kK_{\ast,1}$ with $\T$-degree $d-\mu$. Let $P(\T,\X)\in\kK_{i,j}$ with $i\geq d-\mu$, and $\{p_{\mu,1}(\T,\X),\,q_{d-\mu,1}(\T,\X)\}$ as above, a $\K[\T]$-basis of $\kK_{,1}$. Let $L_\mu(\X)$ (resp. $M_{d-\mu}(\X)$) be the coefficient of $T_2^\mu$ (resp. $T_2^{d-\mu}$) in $p_{\mu,1}(\T,\X)$ (resp. $q_{d-\mu,1}(\T,\X)$). Also, let $W(\X)$ be the coefficient of $T_2^i$ in $P(\T,\X)$. As $P(\T,\X)\in\kK_{i,j},$ due to Proposition \[1.1\] we have that there exists $a\in\N,\, \alpha(\T,\X),\,\beta(\T,\X)\in\K[\T,\X]$ such that $$\label{t2} T_2^aP(\T,\X)=\alpha(\T,\X)p_{\mu,1}(\T,\X)+\beta(\T,\X)q_{d-\mu,1}(\T,\X).$$ We set $T_1=0$ in , and get an expression of the form $$W(\X)=A(\X)L_{\mu}(\X)+B(\X)M_{d-\mu}(\X),$$ with $A(\X),\,B(\X)\in\K[\X]$. Set then $$\label{recur} Q(\T,\X):=P(\T,\X)-T_2^{i-\mu}A(\X)p_{\mu,1}(\T,\X)-T_2^{i-d+\mu}B(\X)q_{d-\mu,1}(\T,\X)$$ By setting $T_1=0$ in , it is easy to see that $Q(\T,\X)$ vanishes, so we have that $$Q(\T,\X)=T_1\,\tilde{Q}(\T,\X)$$ with $\tilde{Q}(\T,\X)\in\kK_{i-1,j}$. If $i-1\geq d-\mu$, we have then that $$P(\T,\X)\in\langle \tilde{Q}(\T,\X),\,p_{\mu,1}(\T,\X),\,q_{d-\mu,1}(\T,\X)\rangle\subset\langle \cup_{\ell\leq i-1}\kK_{\ell,j}\rangle;$$ and by iterating this argument with $\tilde{Q}(\T,\X)$ instead of $P(\T,\X)$, we conclude that $P(\T,\X)\in\langle\cup_{\ell\leq d-\mu}\kK_{\ell,j}\rangle$. If $i=d-\mu$, reasoning as above we arrive to $$P(\T,\X)\in\langle\cup_{\ell\leq d-\mu-1}\kK_{\ell,j}\rangle+\langle p_{\mu,1}(\T,\X),q_{d-\mu,1}(\T,\X)\rangle,$$ and hence the claim follows. The Rees Algebra of curves parameterizable by conics {#Rees} ==================================================== All along this section we will assume that $\phi$ is parameterizable by conics and not by lines, (i.e. $d>3$, see Proposition \[d’\] ). Let $(F_1(\X),F_2(\X))$ be the pair of forms of degree $2$ without common factors defining the inverse of $\phi$. Then, due to Corollary \[maxmu\] we know that $\mu=\lfloor\frac{d}{2}\rfloor$. We will describe a set of minimal generators of $\kK$ by computing successive Morley forms —as in [@CD10]— between two generators of the $\mu$-basis and $T_1F_2(\X)-T_2F_1(\X)$. There will also be a moving conic that will come into play if $d$ is even. We start with the following proposition, which will be useful in the sequel. \[prop2\] If $2i+j<d$, then every nonzero element of $\kK_{i,j}$ is a polynomial multiple of $T_1F_2(\X)-T_2F_1(\X)$. Let $P(\T,\X)\in \kK_{i,j}$. Due to (\[connect\]) we have $$F_2(\X)^iP(\T,\X)-T_2^iP(F_1(\X),F_2(\X),\X)=(T_1F_2(\X)-T_2F_1(\X))Q(\T,\X)$$ with —thanks to Proposition \[multiple\]— $P(F_1(\X),F_2(\X),\X)$ a homogeneous polynomial multiple of $E(\X)$ of degree $2i+j<d=\deg(E(\X))$. As $E(\X)$ is irreducible, we have then $P(F_1(\X),F_2(\X),\X)=0$ and so $$F_2(\X)^iP(\T,\X)=(T_1F_2(\X)-T_2F_1(\X))Q(\T,\X),$$ which implies that there exists $Q_0(\T,\X)$ such that $$P(\T,\X)=(T_1F_2(\T,\X)-T_2F_1(\T,\X))Q_0(\T,\X).$$ $d$ odd {#odd} ------- In this section we will assume $d=2k+1$. By Corollary \[maxmu\], we then have $\mu=k$. Let $\{p_{k,1}(\T,\X),\,q_{k+1,1}(\T,\X)\}$ be a basis of $\mbox{Syz}(I)$. \[polar\] Up to a nonzero constant in $\K$, we have that $$\label{jj} p_{k,1}(F_1(\X),F_2(\X),\X)=E(\X).$$ The polynomial $p_{k,1}(F_1(\X),F_2(\X),\X)$ is either identically zero or has degree $d=\deg(E(\X))$ and, due to Proposition \[multiple\], we know that it is a multiple of $E(\X)$. If we show that it is not identically zero, then we are done. But if this were not the case, due to (\[connect\]) we would have to conclude that $p_{k,1}(\T,\X)$ is a multiple of $T_1F_2(\X)-T_2F_1(\X)$, which is impossible as the latter has degree $2$ in the variables $\X$’s. We will define one nonzero element in $P_j(\T,\X)\in\kK_{j,d-2j}$ for $j=0,1,\ldots, k-1$. We will do this recursively starting from $\kK_{k-1,2}$ and increasing the $\X$-degree at the cost of decreasing the $\T$-degree. This is the analogue of “computing Sylvester forms” in [@cox08; @bus09], and we will perform essentially the same operations we have done in [@CD10] in order to get a minimal set of generators of $\kK$ for curves parameterizable by lines. Set then $P_k(\T,\X):=p_{k,1}(\T,\X);$ and for $j$ from $0$ to $k-1$ do: - write $P_{k-j}(\T,\X)$ as $A_{k-j}(\T,\X)T_1+B_{k-j}(\T,\X)T_2$ (clearly there is more than one way of doing this, just choose one), - Set $P_{k-j-1}(\T,\X):=A_{k-j}(\T,\X)F_1(\X)+B_{k-j}(\T,\X)F_2(\X)$. We easily check that $P_j(\T,\X)\in \kK_{j,d-2j}$ for $j=0,\ldots, k-1$, and also that (up to a nonzero constant in $\K$), $$\label{remm} P_j(F_1 (\X),F_2(\X),\X)=E(\X).$$ In addition, it is easy to check that $P_0(\T,\X)=E(\X)$. \[mtodd\] A minimal set of generators of $\kK$ is $$J:=\{T_1F_2(\X)-T_2F_1(\X),\,q_{k+1,1}(\T,\X),\,P_0(\T,\X),\ldots,\,P_k(\T,\X)\}.$$ Let us first check that $J$ is a minimal set of generators of the ideal generated by its elements. The forms $T_1F_2(\X)-T_2F_1(\X)$, $q_{k+1,1}(\T,\X)$,$P_k(\T,\X),\ldots,\,P_0(\T,\X)$ have bidegrees $(1,2)$, $(k+1,1)$, $(k,1)$, $(k-1,3)$ $\dots$, $(1,2k-1)$, $(0,2k+1)$ respectively. Taking into account these bidegrees we observe that, since $k\geq 2$, it is clear that $T_1F_2(\X)-T_2F_1(\X)$ cannot be a polynomial combination of the others. Also, $q_{k+1,1}(\T,\X)$ can only be a multiple of $P_k(\T,\X),$ which is impossible since they are a basis of $\mbox{Syz}(I)$. Suppose now that $P_j(\T,\X)$ for some $j=0,\dots,k$ is a polynomial combination of the others; then $$P_j(\T,\X)=H_0(\T,\X)(T_1F_2(\X)-T_2F_1(\X))+H_1(\T,\X)q_{k+1,1}(\T,\X)+\sum_{i\neq j}G_i(\T,\X)P_i(\T,\X)$$ All the elements $\{P_j(\T,\X)$, $j=0,\dots ,k\}$, are nonzero and have different bidegrees $(j,d-2j).$ In addition, $\deg_{\T}(q_{k+1,1}(\T,\X))=k+1>j.$ Thus, $$H_1(\T,\X)= G_i(\T,\X)=0, \ i\neq j.$$ It remains to show that $P_j(\T,\X)$ is not a multiple of $T_1F_2(\X)-T_2F_1(\X)$. But if this were the case, then we would have that $P_j(F_1(\X),F_2(\X),\X)=0$, in contradiction with . We conclude then that $J$ is a set of minimal generators of $\langle J\rangle.$ Now we have to show that $\langle J\rangle =\kK$, one of the inclusions being obvious. Let $P(\T,\X)$ be a nonzero element in $\kK_{i,j}$. If $2i+j<d$ then due to Proposition \[prop2\], $P(\T,\X)$ is a multiple of $T_1F_2(\X)-T_2F_1(\X)$ and the claim follows straightforwardly. Suppose then $2i+j\geq d$. Thanks to Theorem \[mu\] we only have to look at $0\leq i\leq k$. As $P(F_1(\X),F_2(\X),\X)=E(\X)h(\X)$ (due to Proposition \[multiple\]), then by applying to both $P(\T,\X)$ and $P_{i}(\T,\X)h(\X)$ we will get $$F_2(\X)^i\left(P(\T,\X)-P_i(\T,\X)h(\X) \right)\in\langle T_1F_2(\X)-T_2F_1(\X)\rangle,$$ and from here we deduce $$P(\T,\X)\in\langle P_i(\T,\X),\,T_1F_2(\X)-T_2F_1(\X)\rangle\subset\langle J\rangle.$$ $d$ even {#even} -------- Suppose now that $d=2k$. In this case we have again $\mu=k$, but also $d-\mu=k$ and hence there are two generators of $\kK_{,1}$ with $\T$-degree $k$. As usual, denote with $\{p_k(\T,\X),\,q_k(\T,\X)\}$ a basis of $\mbox{Syz}(I)$. One can show easily now that there exist nonzero linear forms $L_F(\X),\,L_G(\X)\in\K[\X]$ such that $$p_k(F_1(\X),F_2(\X),\X)=E(\X)L_F(\X),\ \mbox{and} \ q_k(F_1(\X),F_2(\X),\X)=E(\X)L_G(\X),$$ so we cannot use neither of these elements to get something like . However, by applying some known results derived from the method of moving conics explored in [@SGD97; @ZCG99], it turns out that we can find a polynomial in $\kK_{k-1,2}$ which will play the role of $p_{k,1}(\T,\X)$ in Proposition \[polar\] for this case. \[mconic\] There exists a nonzero element $Q(\T,\X)\in \kK_{k-1,2}$ such that $$Q(F_1(\X),F_2(\X),\X)=E(\X).$$ Moreover, as $\K$-vector spaces we have $$\label{sumdir} \kK_{k-1,2}=Q(\T,\X)\cdot\K \oplus\langle T_1F_2(\X)-T_2F_1(\X)\rangle_{k-1,2}.$$ In the language of moving curves, the fact that $d$ is even and $\mu=k$ means that there are no moving lines of degree $k-1$ which follow the curve; that is, $\kK_{k-1,1}=0$. This condition implies (see for instance Theorem $5.4$ in [@SGD97]) that there exist $k$ linearly independent elements in $\kK_{k-1,2}$. One can easily check that if we multiply $T_1F_2(\X)-T_2F_1(\X)$ by a polynomial $r(\T)$ of degree $k-2$, we then get an element of $\kK_{k-1,2}$. The dimension of the $\K$-vector space generated by all these polynomials is then $k-1$. Hence, there is one form $Q(\T,\X)\in\kK_{k-1,2}$ which does not belong to this subspace, and holds. For this $Q(\T,\X)$ we easily get that $Q(F_1(\X),F_2(\X),\X)$ has to be a scalar multiple of $E(\X)$. If it were zero, then by using the same arguments as before we would have to conclude that $Q(\T,\X)\in\langle T_1F_2(\X)-T_2F_1(\X)\rangle$, which contradicts . Now we will define nonzero elements in $\kK_{j,d-2j}$ for $j=0,1,\ldots, k-1$. As before, we will do this recursively starting from $\kK_{k-1,2}$ and increasing the $\X$-degree by decreasing the $\T$-degree. Set $P_{k-1}(\T,\X):=Q(\T,\X)$ and, for $j$ from $0$ to $k-2$ do: - write $P_{k-1-j}(\T,\X)$ as $A_{k-j}(\T,\X)T_1+B_{k-j}(\T,\X)T_2$ (there is more than one way of doing this, just choose one), - Set $P_{k-j-2}(\T,\X):=A_{k-j}(\T,\X)F_1(\X)+B_{k-j}(\T,\X)F_2(\X)$. We easily check that $P_j(\T,\X)\in \kK_{j,d-2j}$ for $j=0,\ldots, d-1$, and also that (up to a nonzero constant in $\K$), $$\label{remm2} P_j(F_1(\X),F_2(\X),\X) = Q(F_1(\X),F_2(\X),\X) =E(\X)\ \forall j=0,\ldots, k-1.$$ Also, by construction we have that $P_0(\T,\X)=E(\X)$. \[mteven\] A minimal set of generators of $\kK$ is $$J:=\{T_1F_2(\X)-T_2F_1(\X),\,P_0(\T,\X),\ldots,\,P_{k-1}(\T,\X),\,p_{k,1}(\T,\X),\,q_{k,1}(\T,\X)\}.$$ As before, we first check that $J$ is a minimal set of generators of the ideal $\langle J\rangle$. We start again by verifying that $p_k(\T,\X)$ and $q_k(\T,\X)$ cannot be combination of other elements in the family due to the fact that they have minimal $\X$-degree and $\K[\T]$-linearly independent. All the other elements $P_j(\T,\X),\,j=0,\ldots, k-1$ are in different pieces of bidegrees $(j,d-2j)$ so neither of them can be a polynomial combination of the others. In addition, the form $T_2F_1(\X)-T_1F_2(\X)$ is minimal with respect to the $\T$-degree, so it is independent. It remains then show that $P_j(\T,\X)$ is not a multiple of $T_1F_2(\X)-T_2F_1(\X)$. But if this were the case, then we would have that $P_j(F_1(\X),F_2(\X),\X)=0$, which contradicts . In order to complete the proof, we must show that $\langle J\rangle =\kK$. As before, one of the inclusions is trivial. Let then $P(\T,\X)$ be a nonzero element in $\kK_{i,j}$. If $2i+j<d$ then due to Proposition \[prop2\], $P(\T,\X)$ is a multiple of $T_1F_2(\X)-T_2F_1(\X)$ and the claim follows. Suppose now $2i+j\geq d$. Thanks to Theorem \[mu\] we only have to look at $0\leq i\leq k-1$. As $P(F_1(\X),F_2(\X),\X)=E(\X)h(\X)$ (due to Proposition \[multiple\]), then by applying to both $P(\T,\X)$ and $P_{i}(\T,\X)h(\X)$ we will get $$F_2(\X)^i\left(P(\T,\X)-P_i(\T,\X)h(\X) \right)\in\langle T_1F_2(\X)-T_2F_1(\X)\rangle.$$ From here we deduce $P(\T,\X)\in\langle P_i(\T,\X),\,T_1F_2(\X)-T_2F_1(\X)\rangle\subset\langle J\rangle,$ and the claim follows. Note that the number of minimal generators in both cases $d=2k+1$ or $d=2k$ is always $k+3$, and also that a system of generators of $\kK$ includes a $\K[\T]$-basis of $\mbox{Syz}(\I)$ and the implicit equation as expected. (120,65) (0,0)[(1,0)[38]{}]{} (39,0)[$i$]{} (0,0)[(0,1)[60]{}]{} (0,61)[$j$]{} (3,6)[ ]{} (3,7) (27,3)[ ]{} (27,4) (24,9)[ ]{} (24,10) (21,15)[ ]{} (21,16) (30,3)[ ]{} (33,4) (0,57) (1,57) (24,3)(1,0)[2]{}[(1,0)[0.5]{}]{} (24,3)(0,2)[3]{}[(0,1)[0.5]{}]{} (21,9)(1,0)[2]{}[(1,0)[0.5]{}]{} (21,9)(0,2)[3]{}[(0,1)[0.5]{}]{} (0,51)(1,0)[2]{}[(1,0)[0.5]{}]{} (3,51)[ ]{} (3,52) (3,45)(1,0)[2]{}[(1,0)[0.5]{}]{} (3,45)(0,2)[3]{}[(0,1)[0.5]{}]{} (6,45)[ ]{} (12,33)[ ]{} (15,27)[ ]{} (18,21)[ ]{} (70,0)[(1,0)[38]{}]{} (109,0)[$i$]{} (70,0)[(0,1)[60]{}]{} (70,61)[$j$]{} (73,6)[ ]{} (73,7) (97,3)[ ]{} (97,4) (98,3) (94,6)[ ]{} (94,7) (91,12)[ ]{} (91,13) (88,18)[ ]{} (89,18) (70,54) (71,54) (91,6)(1,0)[2]{}[(1,0)[0.5]{}]{} (91,6)(0,2)[4]{}[(0,1)[0.5]{}]{} (88,12)(1,0)[2]{}[(1,0)[0.5]{}]{} (88,12)(0,2)[4]{}[(0,1)[0.5]{}]{} (73,48) (74,48) (76,42) (70,48)(1,0)[2]{}[(1,0)[0.5]{}]{} (73,42)(1,0)[2]{}[(1,0)[0.5]{}]{} (73,42)(0,2)[4]{}[(0,1)[0.5]{}]{} (85,24)[ ]{} (82,30)[ ]{} Consider the following parameterization $$\left\{\begin{array}{ccl} u_1(T_1,T_2)&=&T_1^5+T_2^5+T_1^4T_2\\ u_2(T_1,T_2)&=&T_1^3T_2^2\\ u_3(T_1,T_2)&=&T_1^5-T_2^5, \end{array}\right.$$ whose inverse can easily be found as $$\F(\X)=(4 X_1^2 +X_2 X_1 +4 X_1 X_3 +16 X_2^2 +X_2 X_3,\,4 X_1^2 + 6 X_1 X_2 + X_2^2 + 2 X_2 X_3 - 4 X_3^2).$$ We have here $d=5,\, \mu=2$ and with the aid of a computer software find the following $\mu$-basis: $$\begin{array}{ccl} p_{2,1}(\T,\X)&=& 2 T_1^2 X_2 + T_2 T_1 X_2 - T_2^2 X_1 - T_2^2 X_3,\\ q_{3,1}(\T,\X)&=& 8 T_1^3 X_1 - 8 T_1^3 X_3 - 4 T_1^2 T_2 X_1 - 4 T_1^2 T_2 X_3 + 2 T_1 T_2^2 X_1 + T_2 ^2 T_1 X_2 \\ & &+ 2 T_1 T_2^2 X_3 - T_2^3 X_1 - 16 T_2^3 X_2 - T_2^3 X_3. \end{array}$$ Now we can perform the algorithm given in Section \[odd\], write $$P_2(\T,\X):=p_{2,1}(\T,\X)=( 2 T_1 X_2 + T_2 X_2)T_1 +(- T_2 X_1 - T_2 X_3)T_2,$$ and set $$\begin{array}{ccl} P_1(\T,\X)&=&( 2 T_1 X_2 + T_2 X_2)F_1(\X) +(- T_2 X_1 - T_2 X_3)F_2(\X)\\ &=&32 T_1X_2^3 + 8 T_1 X_1^2 X_2 + 2 T_1X_1 X_2^2 + 8 T_1X_1 X_2 X_3 + 2 T_1X_2^2 X_3 \\ && + 16 T_2X_2^3 - 2 T_2X_1^2 X_2 - 4 T_2X_1 X_2 X_3 - 4 T_2X_1^3 + 4 T_2 X_1 X_3^2 \\ & &- 4 T_2 X_1^2 X_3 - 2 T_2X_2 X_3^2 + 4 T_2 X_3^3. \end{array}$$ We perform the same operations on $P_1(\T,\X)$ to get the implicit equation: $$\begin{array}{ccl} P_0(\T,\X)&=&16\big( -X_1^5 + 33 X_2^5 - X_1^4 X_3 + 3 X_1^2 X_2 X_3^2 + 16 X_1 X_2^3 X_3 \\ &&+ X_1^3 X_2 X_3 - X_3^5 - 4 X_2^3 X_3^2 - X_1 X_3^4 + 2 X_1^3 X_3^2 + 2 X_1^2 X_3^3 \\ &&+ 20 X_1^2 X_2^3 + 6 X_2^4 X_3 + 10 X_1 X_2^4 + X_2 X_3^4 + 3 X_1 X_2 X_3^3\big). \end{array}$$ By Theorem \[mtodd\], a minimal set of generators of $\kK$ is given by the five polynomials $p_{2,1}(\T,\X),\,q_{3,1}(\T,\X),\,P_1(\T,\X),\,P_0(\T,\X)$ and $T_1F_2(\X)-T_2F_1(\X).$ \[d6\] Set $d=6$ and consider $$\left\{\begin{array}{ccl} u_1(T_1,T_2)&=&T_1^6+T_1^5T_2\\ u_2(T_1,T_2)&=&T_1^3T_2^3\\ u_3(T_1,T_2)&=&T_2^6. \end{array}\right.$$ By computing explicitly a Gröbner basis of $\ker(\mathfrak{h})$ we get $\mu=3$ and the following $\mu$-basis: $$\begin{array}{ccl} p_{3,1}(\T,\X)&=& T_1^3X_3-T_2^3X_2,\\ q_{3,1}(\T,\X)&=& T_2^3X_1-T_1^3X_2-T_1^2T_2X_2. \end{array}$$ A quadratic inverse can be found also as part of the Gröbner basis of $\kK$: $$T_1F_2(\X)-T_2F_1(\X)= T_1(X_1X_3-X_2^2)-T_2X_2^2,$$ and we can also detect a moving conic of degree $2$ in $\T$ which is not a multiple of the latter: $$Q(\T,\X)= T_1^2 X_2 X_3 - T_2^2 X_1 X_3 + T_2^2 X_2^2.$$ Now we have all the ingredients to start with the algorithm presented in Section \[even\]: set $P_2(\T,X):=Q(\T,\X)$, and write $$P_2(\T,\X)=(T_1 X_2 X_3)T_1 +(- T_2 X_1 X_3 + T_2 X_2^2)T_2.$$ Then, we have $$\begin{array}{ccl} P_1(\T,\X)&:=&(T_1 X_2 X_3)F_1(\X) +(- T_2 X_1 X_3 + T_2 X_2^2)F_2(\X)\\ &=&(T_1 X_2 X_3)X_2^2 +(- T_2 X_1 X_3 + T_2 X_2^2)(X_1X_3-X_2^2)\\ &=&X_2^3X_3T_1-(X_1X_3-X_2^2)^2T_2; \end{array}$$ and finally we get $$\begin{array}{ccccl} E(\X)&=&P_0(\T,\X)&:=&X_2^3X_3F_1(\X)-(X_1X_3-X_2^2)^2F_2(\X)\\ &&&=&X_2^5X_3-(X_1X_3-X_2^2)^3 \end{array}$$ which is the implicit equation of the curve. Theorem \[mteven\] tells us now that a minimal set of generators of $\kK$ is given by $p_{3,1}(\T,\X),\,q_{3,1}(\T,\X),\,P_2(\T,\X),\,P_1(\T,\X),\,E(\X). $ Conclusions and Open Problems {#conclu} ============================= We have described in detail the geometric features of rational curves parameterizable by conics and the algebraic aspects of their parameterizations. It would be interesting to get a similar description of families of curves parameterizable by forms of low degree. For simplicity, we will set our open questions and remarks for curves parameterizable by cubics (i.e. $\deg(\F(\X))=3$ in ), but of course the interest is to get a description for general curves of degree $d$ parameterizable by forms of degree $d',\,d'\ll d$. - In Proposition \[deg2\] and Remark \[elunico\] we have shown that the only pairs of quadratic forms $(F_1(X_1,X_2),\,F_2(X_1,X_2))$ without common factors not inducing a birational application $\C\dasharrow\P^1$ for any plane curve $\C$ are those such that $T_1F_2(X_1,X_2)-T_2F_1(X_1,X_2)$ defines a degenerate conic in $\P^1_{\K(\T)}$. Is there a geometric or algebraic condition analogue to this for pairs of cubics in $\K(\X)$? - The description of the family of all rational parameterizations induced by a given inverse map $\psi$ in Proposition \[paramparam\] is based on the fact that the every nondegenerate conic in any projective plane over an algebraically closed field is parameterizable. Which is the analogue of this fact for cubics? What is the equivalent of “nondegenerate conic” in the case of cubics? - One can prove a more general statement in one of the directions of Theorem \[mtmt\]: if $\Lambda:\P^2\dasharrow\P^2$ is a birational transformation whose inverse is given by three cubics $F_1(\X),\,F_2(\X),\,F_3(\X)$, and $\C_0$ is a curve parameterizable by lines with singularity at $(0:0:1)$, then $\overline{\Lambda(\C_0)}$ is a curve parameterizable by cubics. Are these all of them? Note that for a regular sequence of homogeneous forms $F_1(\X),\,F_2(\X)$ of degree $3,$ the variety $\V(\F(\X))$ has cardinality $9$ counted with multiplicities. It turns out that $F_1(\X),\,F_2(\X),\,F_3(\X)$ defines a birational map if and only if $\V(F_1(\X),\,F_2(\X),\,F_3(\X))$ has cardinality $8$ (counted with multiplicities). But Cayley-Bacharach Theorem ([@EGH96]) implies that any form $F_3(\X)$ of degree $3$ vanishing in all but one point of $\V(\F(\X))$ must vanish in all of them. So, in principle “extending” a general regular sequence of two cubics to an automorphism of $\P^2$ given by cubics as it was done in Proposition \[quadratic\], cannot be done straightforwardly. - The computation of minimal generators of ${\mathfrak h}$ in Section \[Rees\] involved one moving conic that follows the parameterization whose knowledge comes from the method of moving conics. There is no known systematic method for moving cubics so far. How can we detect forms of lower degree in $\X$ in order to produce elements like the $Q(\T,\X)$ described in Proposition \[mconic\]? We hope that we shall be able to answer these questions in future papers. [XXXXXX]{} Alberich-Carramiñana, Maria. Lecture Notes in Mathematics, 1769. Springer-Verlag, Berlin, 2002. Busé, Laurent. J. Algebra 321 (2009), no. 8, 2317–2344. Busé, Laurent; Jouanolou, Jean-Pierre. J. Algebra 265 (2003), no. 1, 312–357. Chen, Falai; Cox, David; Liu, Yang. J. Symbolic Comput. 39 (2005), no. 6, 689–706. Chen, Falai; Wang, Wenping; Liu, Yang. J. Symbolic Comput. 43 (2008), no. 2, 92–117. Cortadellas Benítez, Teresa; D’Andrea, Carlos. Computer Aided Geometric Design, Volume 27, Issue 6, August 2010, Pages 461-473. Cox, David A. Theoret. Comput. Sci. 392 (2008), no. 1-3, 23–36. Cox, David; Goldman, Ronald; Zhang, Ming. J. Symbolic Comput. 29 (2000), no. 3, 419–440. Cox, David; Hoffman, J. William; Wang, Haohao. J. Pure Appl. Algebra 212 (2008), no. 7, 1787–1796. Cox,David; Kustin, Andrew; Polini, Claudia; Ulrich, Bernd. Cox, David A.; Sederberg, Thomas W.; Chen, Falai. Comput. Aided Geom. Design 15 (1998), no. 8, 803–827. Eisenbud, David; Green, Mark; Harris, Joe. Bull. Amer. Math. Soc. (N.S.) 33 (1996), no. 3, 295–324. Fulton, William. Notes written with the collaboration of Richard Weiss. Reprint of 1969 original. Advanced Book Classics. Addison-Wesley Publishing Company, Advanced Book Program, Redwood City, CA, 1989. Hirschfeld, J. W. P.; Korchmáros, G.; Torres, F. Princeton Series in Applied Mathematics. Princeton University Press, Princeton, NJ, 2008. Hoffman, J. William; Wang, Haohao. Journal of Algebra and its Applications, Volume: 9, Issue: 6(2010), 1033–1049 Hong, Jooyoun; Simis, Aron; Vasconcelos, Wolmer V. J. Symbolic Comput. 43 (2008), no. 4, 275–292. Hong, Jooyoun; Simis, Aron; Vasconcelos, Wolmer V. Johansen, Pål Hermunn; Løberg, Magnus; Piene, Ragni. Geometric modeling and algebraic geometry, 55–77, Springer, Berlin, 2008. Kustin, Andrew R.; Polini, Claudia; Ulrich, Bernd. To appear in J. Reine Angew. Math. Sederberg, Thomas; Chen, Falai. Proceedings of SIGGRAPH, 1995, 301–308. Sederberg, Tom; Goldman, Ron; Du, Hang. Parametric algebraic curves and applications (Albuquerque, NM, 1995). J. Symbolic Comput. 23 (1997), no. 2-3, 153–175. Semple, J. G.; Kneebone, G. T. Oxford, at the Clarendon Press, 1952. Sendra, J. Rafael; Winkler, Franz; Pérez-Díaz, Sonia. Algorithms and Computation in Mathematics, 22. Springer, Berlin, 2008. Walker, R.J. Princeton Univ. Press, 1950. Zhang, Ming; Chionh, Eng-Wee; Goldman, Ronald N. Comput. Aided Geom. Design 16 (1999), no. 6, 517–527. [^1]: Both authors are supported by the Research Project MTM2010–20279 from the Ministerio de Ciencia e Innovación, Spain [^2]: Even though most of the books in classic Projective Geometry deal with fields of characteristic zero, it is easy to see that the arguments leading to this classification are characteristic-free.	{ "pile_set_name": "ArXiv" }
--- abstract: 'The relativistic kinetic equations for the two domains separated by the hypersurface with both space- and time-like parts are derived. The particle exchange between the domains separated by the time-like boundaries generates source terms and modifies the collision term of the kinetic equation. The correct hydrodynamic equations for the “hydro+cascade” models are obtained and their differences from existing freeze-out models of the hadronic matter are discussed.' address: \| Bogolyubov Institute for Theoretical Physics, Kiev, Ukraine\ Gesellschaft für Schwerionenforschung (GSI), Darmstadt, Germany author: - 'K.A. Bugaev' title: ' Relativistic Kinetic Equations for Finite Domains and Freeze-out Problem ' --- \#1[\_[ ]{} [ ]{}\_[\#1]{}]{} [Key words:]{} freeze-out, kinetic equations with source terms,“hydro+cascade” equations [2]{} Introduction ============ In recent years an essential progress has been achieved in our understanding of the freeze-out problem in relativistic hydrodynamics, i.e. how to convert hydrodynamic solution into free streaming particles. Thus, in works [@BUG:96] the correct generalization of the famous Cooper-Frey formula [@COOP:74] for the time-like hypersurfaces, the formula, was derived for the first time. This was a necessary, but yet not sufficient step to formulate the relativistic hydrodynamics without causal paradoxes. The main problem was to formulate the energy-momentum and particle conservation not only for the expanding fluid alone, but to extend it onto the system consisting of the fluid and the gas of free streaming particles which are emitted from the freeze-out hypersurface. The principal solution of this task was given in [@BUG:96] and its further analysis is presented in [@BUG:99b]. The advantage of this approach is the absence of the logical and causal paradoxes usually arising, if emission of particles happens from the time-like hypersurfaces [@BUG:96; @BUG:99b]. The disadvantage of this approach lies in its tremendous numerical complexity. Numerous attempts to improve and to develop this approach further using primitive kinetic models [@MAGAS:99; @MAGAS:99b] were not very successful so far. Very recently a more fundamental treatment [@SIN:02] based on the analysis of the Boltzmann equation was suggested. This approach, however, did not overcome the usual difficulty of transport equations in describing the phase transition phenomenon. This difficulty has been overcome naturally within the “hydro + cascade” models suggested in Ref. [@BD:00] (BD) and further developed in [@SHUR:01] (TLS). The latter models assume that the nucleus-nucleus collisions proceed in three stages: hydrodynamic expansion (“hydro”) of quark gluon plasma (QGP), [phase]{} transition from QGP to hadron gas (HG) and the stage of hadronic rescattering and resonance decays (“cascade”). The switch from hydro to cascade modeling takes place at the boundary between the mixed and hadronic phases. The spectrum of hadrons leaving this hypersurface of the QGP–HG transition is taken as input for the cascade. Evidently, such an approach incorporates the most attractive features of both hydrodynamics, which describes the QGP–HG phase transition very well, and cascade, which works better during hadronic rescattering. However, both the BD and TLS models face some principal difficulties which cannot be ignored. Thus, within the BD approach the initial distribution for cascade is found by the formula [@COOP:74], which takes into account particles with all possible velocities, whereas in the TLS model the initial cascade distribution is given by the formula [@BUG:96; @BUG:99b], which accounts for only those particles that can leave the phase boundary. As shown below the formula will lead to causal and mathematical problems in the present version of BD model because the QGP–HG phase boundary inevitably has time-like parts. On the other hand the TLS model from the beginning does not conserve energy, momentum and number of charges and this, as demonstrated later, is due to the fact that the equations of motion used in [@SHUR:01] are not complete and, hence, should be modified. The main difficulty of the “hydro + cascade” approach looks very similar to the freeze-out problem in relativistic hydrodynamics. In both cases the finite domains (subsystems) have time-like boundaries through which the exchange of particles is occurring and this should be taken into account. In relativistic hydrodynamics the problem was solved by the constraints which appeared on the freeze-out hypersurface and provided the global energy-momentum and charge conservation [@BUG:96; @BUG:99b]. Similarly, in kinetic theory one has to modify the transport equations by the source terms which describe the exchange of the particles on the time-like parts of the boundary between domains. Therefore, we shall consider the two semi-infinite domains which are separated by the hypersurface $\Sigma^$ of general type and rederive the kinetic equations for this case in Sect. II. In Sect. III the modification of the collision terms is found and the relation between the system obtained and the Boltzmann equation is discussed. The correct equations of motion for the “hydro + cascade” approach are analyzed in Sect. IV. Drift Term for Semi-Finite Domain ================================= Let us consider the two semi-finite domains, “in” and “out”, separated by the hypersurface $\Sigma^$ which in (3+1) dimensions will be parameterized as $t = t^(\bar{x}) = x_0^(\bar{x})$. The distribution function $\phi_{in}(x,p)$ for $t \le t^(\bar{x})$ belongs to the “in” domain, whereas $\phi_{out}(x,p)$ denotes the distribution function of the “out” domain for $t \ge t^(\bar{x})$. Throughout this work it is assumed that the initial conditions for $\phi_{in}(x,p)$ are given, whereas the initial conditions for $\phi_{out}(x,p)$ are not specified yet and will be the theme of the subsequent discussion. For simplicity we consider a classical gas of point-like Boltzmann particles. Similarly to Ref. [@GROOT] we derive the kinetic equations for $\phi_{in}(x,p)$ and $\phi_{out}(x,p)$ from the requirement of particle number conservation. Therefore, the particles leaving one domain (and crossing hypersurface $\Sigma^$) should be subtracted from the corresponding distribution function and added to the other one. Now we consider the closed hypersurface of the “in” domain, $\Delta x^3$, which consists of two semi-planes $\sigma_{t1}$ and $\sigma_{t2}$ of constant time $t1$ and $t2$, respectively, that are connected from $t1$ to $t2 > t1$ by the part of the boundary $\Sigma^(t1,t2)$. The original number of particles on the hypersurface $\sigma_{t1}$ is given by the standard expression [@GROOT] $$\label{one} N_1 = - \int\limits_{\sigma_{t1}} d \Sigma_\mu \frac{d^3 p}{p^0}~ p^\mu~ \phi_{in}(x,p)\,,$$ where $d \Sigma_\mu$ is the external normal vector to $\sigma_{t1}$ and, hence, the product $p^\mu d \Sigma_\mu \le 0$ is non-positive. It is clear that without collisions these particles can cross either hypersurface $\sigma_{t2}$ or $\Sigma^(t1,t2)$. The corresponding numbers of particles are as follows $$\begin{aligned} \label{two} N_2 \hspace{0.0cm}= \hspace{-0.1cm}\int\limits_{\sigma_{t2}} \hspace{-0.1cm}&& d \Sigma_\mu \frac{d^3 p}{p^0}~ p^\mu~ \phi_{in}(x,p)\,,\\ % \label{three} % N_{loss}^* = \hspace{-0.2cm}\int\limits_{\Sigma^(t1,t2)}\hspace{-0.4cm}&& d\Sigma_\mu % \frac{d^3 p}{p^0}~ p^\mu~ \phi_{in}(x,p) ~\Theta(p^\nu d\Sigma_\nu) \,. %\end{aligned}$$ The $\Theta$-function in the [loss]{} term (\[three\]) is very important because it accounts for the particles leaving the “in” domain (see also discussion in [@BUG:96; @BUG:99b]). For the space-like parts of the hypersurface $\Sigma^(t1,t2)$ which are defined by negative sign $ds^2 < 0$ of the element square, $ds^2 = dt^(\bar{x})^2 - d\bar{x}^2 $, the product $p^\nu d\Sigma_\nu > 0$ is always positive and, therefore, particles with all possible momenta can leave the “in” domain through the $\Sigma^(t1,t2)$. For the time-like parts of $\Sigma^(t1,t2)$ (with sign $ds^2 > 0$) the product $p^\nu d\Sigma_\nu $ can have either sign, and the $\Theta$-function [cuts off]{} those particles which return to the “in” domain. Similarly one has to consider the particles coming to the “in” domain from outside. This is possible through the time-like parts of the hypersurface $\Sigma^(t1,t2)$, if particle momentum satisfies the inequality $ - p^\nu d\Sigma_\nu > 0$. In terms of the external normal $d \Sigma_\mu$ with respect to the “in” domain (the same as in (\[three\])) the number of gained particles $$\label{four} N_{gain}^* = - \int\limits_{\Sigma^(t1,t2)}\hspace{-0.4cm} d\Sigma_\mu % \frac{d^3 p}{p^0}~ p^\mu~ \phi_{out}(x,p) ~\Theta(-p^\nu d\Sigma_\nu) \,$$ is, evidently, non-negative (compare it with contribution of feed-back particles in [@BUG:96]). Since the total number of particles is conserved, i.e. $N_2 = N_1 - N_{loss}^* + N_{gain}^$, one can use the Gauss theorem to rewrite the obtained integral over the closed hypersurface $\Delta x^3$ as the integral over $4$-volume $\Delta x^4 $ surrounded by $\Delta x^3$ $$\begin{aligned} % &&\int\limits_{\Delta x^4} \hspace{-0.1cm} d^4 x % \frac{d^3 p}{p^0}~ p^\mu ~{\partial}_\mu ~ \phi_{in}(x,p) = \nonumber \\ % \label{five} % \vm &&\int\limits_{\Sigma^(t1,t2)}\hspace{-0.4cm} d\Sigma_\mu % \frac{d^3 p}{p^0}~ p^\mu \biggl[\phi_{in}(x,p) - \phi_{out}(x,p) \biggr] \Theta(-p^\nu d\Sigma_\nu) \,. %\end{aligned}$$ Note that in contrast to the usual case [@GROOT], i.e. in the absence of boundary $\Sigma^$, the r.h.s of Eq. (\[five\]) does not vanish identically. The r.h.s of Eq. (\[five\]) can be transformed further to the $4$-volume integral in the following sequence of steps. First we express the integration element $d\Sigma_\mu$ via the normal vector $n^_\mu$ as follows $(dx^j > 0,$ for $ j =1,2,3)$ $$\label{six} % d\Sigma_\mu = n^_\mu dx^1 dx^2 dx^3; \quad n^_\mu \equiv \delta_{\mu 0} - \frac{ \partial t^(\bar{x}) }{\partial x^\mu} (1 - \delta_{\mu 0} )\,, %$$ where $\delta_{\mu \nu}$ denotes the Kronecker $\delta$-function. Then, using identity $\int\limits_{t1}^{t2} dt\, \delta (t - t3) = 1$ for the Dirac $\delta$-function with $t1 \le t3 \le t2$, we rewrite the r.h.s. integral in (\[five\]) as $$\label{seven} % \int\limits_{\Sigma^(t1,t2)}\hspace{-0.4cm} d\Sigma_\mu \cdots \equiv % \int\limits_{V^4_\Sigma} d^4 x~\delta (t - t^(\bar{x}) )~ n^_\mu \cdots\,, %$$ where the $4$-dimensional volume $V^4_\Sigma$ is a direct product of the $3$- and $1$-dimensional volumes $\Sigma^(t1,t2)$ and $(t2-t1)$, respectively. Evidently, the Dirac $\delta$-function allows us to extend integration in (\[seven\]) to the unified $4$-volume $V^4_U = \Delta x^4 \cup V^4_\Sigma$ of $\Delta x^4$ and $V^4_\Sigma$. Finally, with the help of notations $$\label{eight} % \Theta_> \equiv \Theta (t - t^(\bar{x}) ); \quad \Theta_< \equiv 1- \Theta_> %$$ it is possible to extend the l.h.s. integral in Eq. (\[five\]) from $\Delta x^4$ to $ V^4_U$. Collecting all the above results, from Eq. (\[five\]) one obtains $$\begin{aligned} % &&\int\limits_{ V^4_U} \hspace{-0.1cm} d^4 x % \frac{d^3 p}{p^0}~ \Theta_<~ p^\mu ~{\partial}_\mu ~ \phi_{in} = \nonumber \\ % \label{nine} \vm % &&\int\limits_{V^4_U}\hspace{-0.1cm} d^4 x % \frac{d^3 p}{p^0}~ p^\mu n^_\mu \biggl[\phi_{in} - \phi_{out} \biggr] \Theta(-p^\nu n^_\nu) % ~\delta (t - t^(\bar{x}) ) \,. %\end{aligned}$$ Since volumes $\Delta x^4$ and $V^4_U$ are arbitrary, one obtains the collisionless kinetic equation for the distribution function of the “in” domain $$\begin{aligned} % && \Theta_<~ p^\mu ~{\partial}_\mu ~ \phi_{in} (x,p) = \nonumber \\ % \label{ten} % % && p^\mu n^_\mu \biggl[\phi_{in}(x,p) - \phi_{out}(x,p) \biggr] \Theta(-p^\nu n^_\nu) % ~\delta (t - t^(\bar{x}) ) \,. %\end{aligned}$$ Similarly one can obtain the equation for the distribution function of the “out” domain $$\begin{aligned} % && \Theta_>~ p^\mu ~{\partial}_\mu ~ \phi_{out} (x,p) = \nonumber \\ % \label{eleven} % % && p^\mu n^_\mu \biggl[\phi_{in}(x,p) - \phi_{out}(x,p) \biggr] \Theta(p^\nu n^_\nu) % ~\delta (t - t^(\bar{x}) ) \, %\end{aligned}$$ with the same normal vector $n^_\nu$ as in Eq. (\[ten\]). Note the asymmetry between the r.h.s. of Eqs. (\[ten\]) and (\[eleven\]): for the space-like parts of hypersurface $\Sigma^$ the source term with $\Theta(-p^\nu n^_\nu) $ vanishes identically because $p^\nu n^_\nu > 0$. This reflects the causal properties of the equations above: propagation of particles faster than light is forbidden, and hence no particle can (re)enter the “in” domain. Collision Term for Semi-Finite Domain ====================================== Since in the general case $\phi_{in}(x,p) \neq \phi_{out}(x,p)$ on $\Sigma^$, the r.h.s. of Eqs. (\[ten\]) and (\[eleven\]) cannot vanish simultaneously on this hypersurface. Therefore, functions $\Theta_<^ \equiv \Theta_<\|_{\Sigma^} \neq 0$ and $ \Theta_>^ \equiv \Theta_>\|_{\Sigma^} \neq 0$ do not vanish simultaneously on $\Sigma^$ as well. For definiteness it is assumed that $$\label{twelve} % \Theta_<^* = \Theta_>^* = \Theta (0) = \frac{1}{2}\,, %$$ but the final results are independent of this choice. Now the collision terms for Eqs. (\[ten\]) and (\[eleven\]) can be readily obtained. Adopting the usual assumptions on the distribution functions [@GROOT], one can repeat the standard derivation of the collision terms [@GROOT] and get the desired expressions. We shall not recapitulate this standard part, but only discuss how to modify the derivation for our purpose. First of all, one has to start the derivation in the $\Delta x^4$ volume of the “in” domain and then extend it onto the unified $4$-volume $V^4_U = \Delta x^4 \cup V^4_\Sigma$ similarly to the preceding section. Then the first part of the collision term for Eq. (\[ten\]) reads as $$\begin{aligned} \label{thirteen} % C_<^{I} (x,p) & = & \Theta_<^2 \left( I_G [\phi_{in}, \phi_{in}] - I_L [\phi_{in}, \phi_{in}] \right) \,, \\ % \label{fourteen} % % I_G [\phi_{A}, \phi_{B}] & \equiv & \frac{1}{2} \int D^9 P~ % % \phi_{A}(p^{\prime}_1 )~ \phi_{B}(p_1^{\prime})~ W_{p,p_1^{} \| p^{\prime}p_1^{\prime}} % \,, \\ % \label{fifteen} % I_L [\phi_{A}, \phi_{B}] & \equiv & \frac{1}{2} \int D^9 P~ % \phi_{A}(p)~ \phi_{B}(p_1)~ W_{p,p_1^{} \| p^{\prime}p^{\prime}_1}\,, %\end{aligned}$$ where the invariant measure of integration is denoted as $ D^9 P \equiv \frac{d^3 p_1}{p^0_1} \frac{d^3 p^{\prime} }{p^{\prime 0}} \frac{d^3 p^{\prime}_1 }{p^{\prime 0}_1} $ and $W_{p,p_1^{} \| p^{\prime}p^{\prime}_1}$ is the transition rate in the elementary reaction with energy-momentum conservation given in the form $p^\mu + p_1^\mu = p^{\prime \mu} + p^{\prime \mu}_1$. The r.h.s. of (\[thirteen\]) contains the square of $\Theta_<$-function because the additional $\Theta_<$ accounts for the fact that on the boundary hypersurface $\Sigma^$ one has to take only one half of the traditional collision term (due to Eq. (\[twelve\]) only one half of $\Sigma^$ belongs to the “in” domain). It is easy to understand that on $\Sigma^$ the second part of the collision term (according to Eq. (\[twelve\])) is defined by the collisions between particles of “in” and “out” domains $$\label{sixteen} % C_<^{II} (x,p) = \Theta_< \Theta_> \left( I_G [\phi_{in}, \phi_{out}] - I_L [\phi_{in}, \phi_{out}] \right) \,. %$$ Combining results (\[ten\]), (\[thirteen\]) and (\[sixteen\]), we obtain the kinetic equation for the semi-finite “in” domain $$\begin{aligned} % && \Theta_<~ p^\mu ~{\partial}_\mu ~ \phi_{in} (x,p) = C_<^{I} (x,p) + C_<^{II} (x,p) + \nonumber \\ % \label{seventeen} % % && p^\mu n^_\mu \biggl[\phi_{in}(x,p) - \phi_{out}(x,p) \biggr] \Theta(-p^\nu n^_\nu) % ~\delta (t - t^(\bar{x}) ) \,. %\end{aligned}$$ The corresponding equation for the “out” domain $$\begin{aligned} % && \Theta_>~ p^\mu ~{\partial}_\mu ~ \phi_{out} (x,p) = C_>^{I} (x,p) + C_>^{II} (x,p) + \nonumber \\ % \label{eighteen} % % && p^\mu n^_\mu \biggl[\phi_{in}(x,p) - \phi_{out}(x,p) \biggr] \Theta(p^\nu n^_\nu) % ~\delta (t - t^(\bar{x}) ) \, %\end{aligned}$$ can be derived similarly. In (\[eighteen\]) we used the evident notations $C_>^{I} \equiv \Theta_>^2 \left( I_G [\phi_{out}, \phi_{out}] - I_L [\phi_{out}, \phi_{out}] \right) $ and $C_>^{II} \equiv \Theta_< \Theta_> \left( I_G [\phi_{out}, \phi_{in}] - I_L [\phi_{out}, \phi_{in}] \right) $. For the continuous distribution functions on $\Sigma^$, i.e. $\phi_{out}\|_{\Sigma^} = \phi_{in}\|_{\Sigma^}$, the source terms in r.h.s. of Eqs. (\[seventeen\]) and (\[eighteen\]) vanish and one recovers the Boltzmann equations. With the help of the evident relations $$\begin{aligned} \label{nineteen} % &&- {\partial}_\mu ~ \Theta_< = {\partial}_\mu ~ \Theta_> = n_\mu^~ \delta (t - t^(\bar{x}) )\,, \\ % \label{tw} % &&C_<^{I} + C_<^{II} + C_>^{I} + C_>^{II} = % I_G [\Phi, \Phi] - I_L [\Phi, \Phi]\,, %\end{aligned}$$ where the notation $\Phi(x,p) \equiv \Theta_<~\phi_{in}(x,p) + \Theta_>~\phi_{out}(x,p) $ is used, one can get the following result for the sum of Eqs. (\[seventeen\]) and (\[eighteen\]) $$\label{twone} % p^\mu ~{\partial}_\mu ~ \Phi (x,p) = I_G [\Phi, \Phi] - I_L [\Phi, \Phi]\,. %$$ In other words, the usual Boltzmann equation follows from the system of Eqs. (\[seventeen\]) and (\[eighteen\]) automatically without [any assumption]{} about the behavior of $\phi_{in}$ and $\phi_{out}$ on the boundary hypersurface $\Sigma^$. In fact the system (\[seventeen\], \[eighteen\]) generalizes the relativistic kinetic equation to the case of the strong temporal and spatial inhomogeneity, i.e., if $\phi_{in}(x,p) \neq \phi_{out}(x,p)$ on $\Sigma^$. Of course, one has to be extremely careful while discussing the strong temporal inhomogeneity (or discontinuity on the space-like parts of $\Sigma^$) such as the so called [time-like shocks]{} [@TIMESHOCK] because their existence may contradict to the usual assumptions adopted for distribution function. Therefore, in what follows we shall discuss exclusively the spatial inhomogeneities or discontinuities on the time-like parts of $\Sigma^$ which are less restrictive because in some sense the equations above are delocalized in space. From the system (\[seventeen\]), (\[eighteen\]) it is possible to derive the macroscopic equations of motion by multiplying the corresponding equation with $p^\nu$ and integrating it over the invariant measure. Thus Eq. (\[seventeen\]) generates the following expression ($T^{\mu \nu}_{A} \equiv \int \frac{d^3 p }{p^ 0}~ p^\mu p^\nu \phi_{A}(x,p) $) $$\begin{aligned} % && \Theta_<~ {\partial}_\mu ~ T^{\mu \nu}_{in} = \int \frac{d^3 p }{p^ 0}~ p^\nu C_<^{II} (x,p) + \nonumber \\ % \label{twtwo} % % && \int \frac{d^3 p }{p^ 0}~ p^\nu p^\mu n^_\mu \biggl[\phi_{in} - \phi_{out} \biggr] \Theta(-p^\rho n^_\rho) % ~\delta (t - t^(\bar{x}) ) \,. %\end{aligned}$$ Similarly to the usual Boltzmann equation the momentum integral of the collision term $C_<^{I}$ vanishes due to its symmetries, but it can be shown that the integral of the second collision term $C_<^{II}$ does not vanish because it involves two different (and not identical) distribution functions. The corresponding equation for the “out” domain follows similarly $$\begin{aligned} % && \Theta_>~ {\partial}_\mu ~ T^{\mu \nu}_{out} = \int \frac{d^3 p }{p^ 0}~ p^\nu C_>^{II} (x,p) + \nonumber \\ % \label{twthree} % % && \int \frac{d^3 p }{p^ 0}~ p^\nu p^\mu n^_\mu \biggl[\phi_{in} - \phi_{out} \biggr] \Theta(p^\rho n^_\rho) % ~\delta (t - t^(\bar{x}) ) \,. %\end{aligned}$$ Note that similar equations (with $\delta$-like term) first were obtained within the relativistic hydrodynamics in [@BUG:96]. Discussion ========== It is clear that Eqs. (\[seventeen\]), (\[eighteen\]), (\[twtwo\]) and (\[twthree\]) remain valid for the finite domains as well. With their help we are ready to analyze the “hydro+cascade” models. In the TLS model the formula relates $\phi_{in}$ ($\equiv$ hydro, Eq. (\[twtwo\])) and $\phi_{out}$ ($\equiv$ cascade, Eq. (\[eighteen\])) on $\Sigma^$ as follows $$\label{twfoure} % {\rm TLS:} \quad \phi_{out}\biggl\|_{\Sigma^} \hspace{-0.2cm} = \Theta(p^\rho n^_\rho)~ \phi_{out}\biggl\|_{\Sigma^} \hspace{-0.2cm} = % \Theta(p^\rho n^_\rho)~ \phi_{in}\biggl\|_{\Sigma^} \hspace{-0.1cm}\,, %$$ i.e., for the space-like parts of hypersurface $\Sigma^$ these functions are identical, whereas for the time-like parts of $\Sigma^$ there are no returning particles to the “in” domain. In this case the source term in cascade Eq. (\[eighteen\]) is zero, while the source term in hydro Eq. (\[twtwo\]) does not vanish on the time-like parts of the boundary $\Sigma^$. Therefore, the main defect of the TLS model is not even the energy-momentum non-conservation, but the incorrect hydrodynamic equations. The absence of the $\delta$-like source term in [@SHUR:01] breaks the conservation laws (evidently, the system (\[twtwo\],\[twthree\]) obeys the conservation laws), but its inclusion into consideration will inevitably change the hydrodynamic solution of Ref. [@SHUR:01]. The full analysis of the possible solutions of the systems (\[seventeen\],\[eighteen\]) and (\[twtwo\],\[eighteen\]) requires a special consideration. We only mention that from the negative sign of the TLS source term in the r.h.s. of (\[twtwo\]) for equal indices $\nu = \mu$ one immediately can deduce that such a correction to the hydro equations should increase the degree of the fluid rarefaction in comparison with the standard hydrodynamic expansion. It is, therefore, quite possible that such a source term will generate a discontinuity between “in” and “out” domains. In the thermodynamically normal media [@BUG:88] the rarefaction shocks are mechanically unstable. However, it is well know that on the phase transition boundary between QGP and HG the properties of the mixed phase are thermodynamically anomalous [@BUG:88] and the usual rarefaction shocks are possible. Another possibility is the occurrence of the new type of the discontinuity, the [freeze-out shock]{} suggested in Refs. [@BUG:96; @BUG:99b], where the post freeze-out state is described by the distribution and, hence, is very similar to the TLS ansatz. It is clear that in both cases the additional rarefaction will reduce the mean transverse size and the life-time of the hadronizing QGP. Let us consider briefly the BD approach. Since in the BD model the hydro and cascade distributions on $\Sigma^$ are equal $\phi_{out}\|_{\Sigma^} = \phi_{in}\|_{\Sigma^}$, the corresponding source terms vanish in all equations. Therefore, at first glance the BD approach correctly conjugates the hydro and cascade solutions on the arbitrary hypersurface. For the oversimplified kinetics considered above it is so. However, the real situation differs essentially from our consideration. Thus, the hydro part in both the BD and TLS models is assumed to be in the local thermodynamic equilibrium, whereas the matter in the cascade domain can be far from equilibrium [@BD:00; @SHUR:01] (this was, actually, the main reason why both groups decided to use the cascade). Consequently, the BD transport equations for all hadrons are homogeneous in the hydro domain, whereas for the cascade domain they are inhomogeneous. Since the initial BD cascade distribution on the time-like parts of $\Sigma^$ contains the particles returning to the fluid domain ($ p^\nu n^_\nu < 0$), then these particles will move towards the space-like parts of the hypersurface $\Sigma^$ which are located inside of the light cone originated at each point of the time-like part of $\Sigma^$. Then for each hadron the inhomogeneous BD cascade equation will generate a different distribution function than the one already obtained from the hydro equations (or homogeneous transport ones) on these space-like parts of $\Sigma^$. Thus, one arrives at a causal paradox, the [recoil problem]{} [@BUG:96], and at a mathematical inconsistency. Evidently, the inclusion of the viscosity into the hydro equations (apart from its tremendous complexity for the mixture of about hundred hadrons) will not solve the problem because for the small deviations from equilibrium (and, hence, a small viscosity effect) the influence of the returning particles may remain essential. If, on the contrary, the matter in the hydro domain is far from equilibrium, then the usage of the hydro equations becomes problematic. Therefore, a more realistic way is to find the boundary conditions for $\phi_{in}$ and $\phi_{out}$ on the separating hypersurface $\Sigma^$ form the system of kinetic equations (\[seventeen\],\[eighteen\]), and then to apply these boundary conditions to the system (\[seventeen\],\[twtwo\]) which ensures the correct treatment of the relativistic nuclear collisions within the frame of the “hydro+cascade” model. [Acknowledgments.]{} The author thanks A. L. Blokhin, P. Braun-Munzinger, A. Dumitru, P. T. Reuter and D. H. Rischke for very useful discussions and valuable comments. \#1[[Nucl. Phys.]{} [\#1]{}]{} \#1[[Phys. Rev. Lett.]{} [\#1]{}]{} \#1[[J. of Phys.]{} [\#1]{}]{} \#1[[Z. Phys.]{} [\#1]{}]{} \#1[[Phys. Lett.]{} [\#1]{}]{} \#1[[Phys. Rev.]{} [\#1]{}]{} \#1[[Heavy Ion Physics]{} [\#1]{}]{} \#1[[Phys. Rep.]{} [\#1]{}]{} \#1[[Preprint]{} [\#1]{}]{} [99]{} K. A. Bugaev, (1996) 559; and\ K. A. Bugaev and M. I. Gorenstein, [nucl-th/9903072]{} (1999) 70 p. F. Cooper and G. Frye, (1974) 186. K. A. Bugaev, M. I. Gorenstein and W. Greiner, (1999) 2147; and (1999) 333. V. K. Magas [et al.]{}, (1999) 193. V. K. Magas [et al.]{}, (1999) 596. Yu. M. Sinyukov, S. V. Akkelin and Y. Hama, (2002) 052301. S. Bass and A. Dumitru, (2000) 064909; D. Teaney, J. Lauret and E. V. Shuryak, (2001) 4783; and [nucl–th/0110037]{} (2001) 28 p. S. R. de Groot, W. A. van Leeuwen and Ch. G. and Weert, “Relativistic Kinetic Theory”, North-Holland Publishing, Amsterdam, (1980). L. P. Csernai, Zh. Eksp. [Teor. Fiz. (Russ.)]{} [92]{} (1987) 379; and [Sov. Phys. JETP]{} [65]{} (1987) 216. K. A. Bugaev, M. I. Gorenstein and V. I. Zhdanov, Z. Phys. [C39]{} (1988) 365; K. A. Bugaev and M. I. Gorenstein, Z. Phys. [C43*]{} (1989) 261.	{ "pile_set_name": "ArXiv" }
--- abstract: 'Let $A$ be an artin algebra. We show that the bounded homotopy category of finitely generated right $A$-modules has Auslander-Reiten triangles. Two applications are given: (1) we provide an alternative proof of a theorem of Happel in \[H2\]; (2) we prove that over a Gorenstein algebra, the bounded homotopy category of finitely generated Gorenstein projective (resp. injective) modules admits Auslander-Reiten triangles, which improves a main result in \[G\].' author: - \| Yuefei Zheng[^1] and Zhaoyong Huang[^2]\ [Department of Mathematics, Nanjing University, Nanjing 210093, Jiangsu Province, China]{} title: 'Auslander-Reiten Triangles in Homotopy Categories [^3] [^4]' --- Introduction ============= Throughout this paper, $A$ is an artin algebra over a fixed commutative artin ring $R$ and $D:=\Hom_{R}(-,E(R/J))$ is the usual duality, where $J$ is the Jacobson radical of $R$ and $E(R/J)$ is the injective envelope of $R/J$. We denote by $\mod A$ the category of finitely generated right $A$-modules. As usual, we write $\proj A$ (resp. $\inj A$) the category of finitely generated projective (resp. injective) right $A$-modules. Auslander-Reiten sequences, also known as almost split sequences, are one of the central tools in the representation theory of artin algebras. Auslander-Reiten triangles can also be defined by almost split morphisms in $\Hom$-finite Krull-Schmidt triangulated $R$-categories. Happel proved in \[H3\] that the bounded homotopy category $K^b(\proj A)$ of finitely generated projective right $A$-modules has right Auslander-Reiten triangles if and only if the left self-injective dimension of $A$ is finite; and dually, the bounded homotopy category $K^b(\inj A)$ of finitely generated injective right $A$-modules has left Auslander-Reiten triangles if and only if the right self-injective dimension of $A$ is finite. There is a close relation between Auslander-Reiten triangles and Serre functors (\[RV\]): a $\Hom$-finite Krull-Schmidt triangulated $R$-category has right (resp. left) Auslander-Reiten triangles if and only if it has a right (resp. left) Serre functor. A Serre functor by definition is a right Serre functor which is an equivalence. In \[BJ\], Backelin and Jaramillo proved that the bounded homotopy category $K^b(\mod A)$ of $\mod A$ has a right Serre duality. Their method is based on the construction of a $t$-structure in $K^b(\mod A)$, and their proof is somewhat complicated although they obtained some more results. We use the terminology of Auslander-Reiten triangles to prove that the right Serre functor in $K^b(\mod A)$ is always an equivalence (Theorem 3.4). Our result is based on the fact that right (left) minimal almost split morphisms are stable under quotients. It seems more elementary. In particular, we determine Auslander-Reiten triangles admitting special ending (resp. starting) terms (Proposition 4.2). As in abelian categories, the existence of Auslander-Reiten triangles in subcategories were investigated by Jørgensen in \[J\]. Note that $K^{b}(\proj A)$ and $K^{b}(\inj A)$ can be embedded in $K^{b}(\mod A)$ naturally. By using the obtained result about the existence of Auslander-Reiten triangles in $K^b(\mod A)$, we reprove the Happel’s theorem mentioned above (Theorem 4.4). The steps we take seem more “categorization" and can be easily treated dually. The advantage here is that the Auslander-Reiten triangles we treat always lie in $K^{b}(\mod A)$. Similarly, we prove that over a Gorenstein algebra, the bounded homotopy category of finitely generated Gorenstein projective (resp. injective) modules admits Auslander-Reiten triangles, which improves a main result in \[G\]. Preliminaries ============= Recall that a right $A$-module $M$ is called [Gorenstein projective]{} if there exists an exact sequence $T^{\bullet}$ of projective modules which remains exact when applying the functor $\Hom_A(-,P)$ for any $P\in \proj A$ such that $M$ is isomorphic to some kenrel of $T^{\bullet}$. Dually, The notion of [Gorenstein injective modules]{} is defined. We denote the category of all finitely generated Gorenstein projective (resp. injective) modules by $\Gproj A$ (resp. $\Ginj A$). Note that we have $\proj A \subset \Gproj A$ and $\inj A \subset \Ginj A$. Let $f:M\to N$ be a morphism in $\mod A$. According to \[AR1\], $f$ is called [right almost split]{} if it is not a retraction, and any morphism $g:L\to N$ which is not a retraction factors through $f$; it is called [right minimal]{} if any morphism $h$ satisfying $f=f\cdot h$ is an automorphism of $M$; and it is called [right minimal almost split]{} if it is both right almost split and right minimal. The left versions are defined dually. An exact sequence $$0\to M \overset{\alpha} \to N \overset{\beta} \to L\to 0$$ in $\mod A$ is called an [almost split sequence]{} if either $\alpha$ is left minimal almost split or $\beta$ is right minimal almost split. This also means that $M$ and $L$ must be indecomposable. We say $\mod A$ has [almost split sequences]{} if for any indecomposable non-injective module $M$ in $\mod A$ there is an almost split sequence starting at $M$, and for any indecomposable non-projective module $N \in \mod A$ there is an almost split sequence ending at $N$. By \[AR1\], $\mod A$ always has almost split sequences. Let $\mathcal{T}$ be a $\Hom$-finite Krull-Schmidt triangulated $R$-category. The notion of almost split triangles in $\mathcal{T}$ was introduced by Happel in \[H1\]. A triangle $$X \overset{\alpha}\to Y \overset{\beta}\to Z \to X[1]$$ in $\mathcal{T}$ is called an [Auslander-Reiten triangle]{} (or [almost split triangle]{}) if either $\alpha$ is left minimal almost split or $\beta$ is right minimal almost split (see \[H2\], where the notions source morphisms and sink morphisms were used). A triangulated category is said to [have right $($resp. left$)$ Auslander-Reiten triangles]{} if for any indecomposable object $M$ there is an Auslander-Reiten triangle ending (resp. starting) at $M$. Let $F:\mathcal{T} \to \mathcal{T}$ be a triangulated functor. According to \[RV\], $F$ is called a [right Serre functor]{} if for any $X$, $Y$ in $\mathcal{T}$, there is an isomorphism $D\Hom_{\mathcal{T}}(X,Y)\cong \Hom_{\mathcal{T}}(Y,FX)$ which is natural in $X$ and $Y$. This $F$ is unique up to natural isomorphism. A left Serre functor is defined dually. $\mathrm{([BJ, Corollary~2.5 ~and~ Proposition~ 4.6])}$ Let $A$ be an Artin algebra. Then there is a right serre functor $S: K^{b}(\mod A)\to K^{b}(\mod A)$, equivalently, $K^{b}(\mod A)$ has right Auslander-Reiten triangles. AR-triangles in homotopy categories ==================================== Let $\mathcal{A}$ be an additive $R$-category. By an $\mathcal{A}$-module, we mean a contravariant $R$-linear functor from $\mathcal{A}$ to the category of $R$-modules. We denote by $\mod \mathcal{A}$ the category of finitely presented $\mathcal{A}$-modules. Note that, in general, $\mod \mathcal{A}$ is not an abelian category. It is an abelian category if and only if $\mathcal{A}$ has pseudo-kernels (\[A\]). We call $\mathcal{A}$ a [dualizing $R$-category]{} if $D$ gives a duality between $\mod \mathcal{A}$ and $\mod \mathcal{A}^{op}$. Note that, in this case, $\mod \mathcal{A}$ is always an abelian category, and hence the bounded complex category $C^{b}(\mod \mathcal{A})$ of $\mod \mathcal{A}$ is also an abelian category. We begin with a main theorem in \[BJR\]. $\mathrm{([BJR, Theorem~4.3])}$ Let $\mathcal{A}$ be a dualizing $R$-category. Then $C^{b}(\mod \mathcal{A})$ has almost split sequences. Note that $\mod A$ is equivalent to $\mod (\proj A)$ as additive $R$-categories (\[A\]). This means that $C^{b}(\mod A)$ always has almost split sequences. As a consequence, for any indecomposable non-projective (resp. non-injective) object $X$ in $C^{b}(\mod A)$, there is always an almost split sequence ending (resp. starting) at $X$. Let $\mathcal{B}$ be an additive category and $\mathcal{C}$ an additive full subcategory of $\mathcal{B}$ closed under summands. Then we can form the factor category $\mathcal{B}/\mathcal{C}$. The objects in $\mathcal{B}/\mathcal{C}$ are the same as in $\mathcal{B}$, and the morphisms are the morphisms in $\mathcal{B}$ modulo morphisms factor through an object of $\mathcal{C}$. There is a natural factor functor $\pi: \mathcal{B}\to \mathcal{B}/\mathcal{C}$. It is an additive functor. For both objects and morphisms, we denote their images under $\pi$ by adding $\ \widetilde{ }\ $ above. The following lemma is a well known fact. For the reader’s convenience we give a quick proof here. Let $\mathcal{B}$ be a Hom-finite Krull-Schmidt $R$-category and $\mathcal{C}$ an additive full subcategory of $\mathcal{B}$ closed under summands. If $f: M\to N$ is a right (resp. left) minimal almost split morphism in $\mathcal{B}$, then $\widetilde{f}: \widetilde{M}\to \widetilde{N}$ is also a right (resp. left) minimal almost split morphism in $\mathcal{B}/\mathcal{C}$. [Proof.]{} Obviously, $N$ is indecomposable; in particular, it has no nonzero summands in $\mathcal{C}$. By \[AR2, Lemma 1.1(c)\], $\widetilde{f}: \widetilde{M}\to \widetilde{N}$ is not a retraction. Let $\widetilde{g}: \widetilde{L}\to \widetilde{N}$ be not a retraction. Then it is induced by a morphism $g: L\to N$ which is also not a retraction. So $g$ factors through $f$ since $f$ is right almost split, and hence $\widetilde{g}$ factors through $\widetilde{f}$. If $\widetilde{f}$ is not minimal, then $\widetilde{f}$ has a direct summand of the form $\widetilde{W} \to 0$, and so $f$ has a direct summand of the form $W \to 0$ or $W \to C$ with $0 \neq C\in \mathcal{C}$. Note that the former one gives a contradiction to the minimality of $f$, and the latter one gives a contradiction to the indecomposableness of $N$. $\square$ A complex $X$ is called [contractible]{} if it is isomorphic to zero in $K^{b}(\mod A)$, that is, it is splitting exact. Note that a chain map of complexes is homotopic to zero if and only if it factors through some contractible complex. So $K^{b}(\mod A)$ is exactly the factor category of $C^{b}(\mod A)$ by modulo contractible complexes (\[H2, p.28\]). We also need the following A complex $X$ is a projective (resp. injective) object in $C^{b}(\mod A)$ if and only if it is a contractible complex consisting of projective (resp. injective) modules in $\mod A$. [Proof.]{} See for example \[EJ2, Theorem 1.4.7\]. $\square$ Now we can prove the following result, which improves Theorem 2.1. $K^{b}(\mod A)$ has Auslander-Reiten triangles. [Proof.]{} Let $0\neq \widetilde{X} \in K^{b}(\mod A)$ be indecomposable. Then it is induced by an indecomposable object $X$ in $C^{b}(\mod A)$. Note that $X$ is neither projective nor injective by Lemma 3.3. It follows from Lemma 3.1 that there is a minimal right almost split morphism $f:Y\to X$ in $C^{b}(\mod A)$. Then by Lemma 3.2, its image $\widetilde{f}: \widetilde{Y}\to \widetilde{X}$ is also right minimal almost split. Complete it to a triangle $$L \to\widetilde{Y}\to \widetilde{X} \to L[1]$$ in $K^{b}(\mod A)$. Then by definition, it is an Auslander-Reiten triangle in $K^{b}(\mod A)$. Dually, there is an Auslander-Reiten triangle starting at $\widetilde{X}$. $\square$ Applications ============ In this section, we will reprove the Happel’s theorem by using the main result in Section 3. Our proof is based on the restriction of Auslander-Reiten triangles in subcategories. Also one can see that using this technique, over a Gorenstein algebra the existence of Auslander-Reiten triangles in the bounded homotopy category of Gorenstein projective modules is valid. This improves a result by Gao in \[G\] where only the existence of right Auslander-Reiten triangles is proved and the condition CM-finiteness is necessary. Although $K^{b}(\mod A)$ has Auslander-Reiten triangles, it is difficult to compute the Auslander-Reiten translation. By \[RV, Theorem I.2.4\], $K^{b}(\mod A)$ has Auslander-Reiten triangles if and only if it has a Serre functor. Denote the Serre functor of $K^{b}(\mod A)$ by $S$. Then by \[RV, Proposition I.2.3\], we have that for any indecomposable object $X$, the left end term of an Auslander-Reiten triangle ending at $X$ is $S\cdot [-1]$, the other end term is its quasi-inverse $S^{-1}\cdot [1]$. Thanks to this result, we only need to compute the Serre dual object for an indecomposable object $X$. Let $X$ and $Y$ be in $C^{b}(\mod A)$. 1. If $X$ is degreewise projective, then we have a natural isomorphism $$D\Hom_{A}(X,Y)\cong \Hom_{A}(Y,X\otimes_{A}DA).$$ 2. If $Y$ is degreewise injective, then we have a natural isomorphism $$D\Hom_{A}(X,Y)\cong \Hom_{A}(\Hom_{A}(DA,Y),X).$$ [Proof.]{} (1) Note that for any $X,Y$ in $\mod A$ , we have a natural morphism $$\delta_{Y,X}:Y\otimes_A\Hom_A(X,A)\to \Hom_A(X,Y),~y\otimes f\longmapsto (x\longmapsto yf(x)).$$ So we have a natural morphism $$\eta_{X,Y}:D\Hom_A(X,Y)\overset{D\delta_{Y,X}}\longrightarrow D(Y\otimes_A\Hom_A(X,A))$$ $$\begin{aligned} &~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\cong\Hom_R(Y\otimes_A\Hom_A(X,A),E(R/J))\\ &~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\cong \Hom_A(Y,\Hom_R(\Hom_A(X,A),E(R/J))\\ &~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\cong \Hom_A(Y,X\otimes_ADA).\end{aligned}$$ It is known that if $X$ is projective, then $\delta_{Y,X}$ is an isomorphism. Thus $D\delta_{Y,X}$ is also an isomorphism. Therefore we have a natural isomorphism $$\eta_{X,Y}:D\Hom_{A}(X,Y)\cong \Hom_{A}(Y,X\otimes_{A}DA).$$Now the isomorphism can be extended to the desired situation. \(2) Let $Y$ be injective. Then $Y\cong DA\otimes_A P$ for some $P\in \proj A$. Then we have isomorphisms $$\begin{aligned} &\ \ \ \ \Hom_{A}(\Hom_{A}(DA,Y),X)\\ & \cong \Hom_{A}(\Hom_{A}(DA,DA\otimes_A P),X)\\ & \cong \Hom_A(P,X)\\ & \cong D\Hom_{A}(X,P\otimes_{A}DA)\ {\rm (by\ (1))}\\ & \cong D\Hom_{A}(X,Y).\end{aligned}$$ Similarly, the isomorphism can be extended to the desired situation. $\square$ Let $X$ and $Y$ be in $C^{b}(\mod A)$. 1. If $X$ is degreewise projective, then there is an Auslander-Reiten triangle in $K^{b}(\mod A)$ $$X[-1]\otimes _A DA\to M\to X\to X\otimes _A DA.$$ 2. If $Y$ is degreewise injective, then there is an Auslander-Reiten triangle in $K^{b}(\mod A)$ $$Y\to N \to \Hom_A(DA,Y[1])\to Y[1].$$ [Proof.]{} We only prove (1), and the proof of (2) is similar. Let $X \in K^{b}(\proj A)$. By Lemma 4.1, we have $D\Hom_{A}(X,Y)\cong \Hom_{A}(Y,X\otimes_{A}DA)$. Thus we have isomorphisms $$D\Hom_{K^{b}(\mod A)}(X,Y)\cong DH^{0}\Hom_{A}(X,Y) \cong H^{0}D\Hom_{A}(X,Y)\cong$$ $$H^{0}\Hom_{A}(Y,X\otimes_{A}DA) \cong \Hom _{K^{b}(\mod A)}(Y,X\otimes_{A}DA).$$ This holds for any $Y\in K^{b}(\mod A)$. Then by the Yoneda’s lemma, the Serre dual object for $X$ is $X\otimes_{A}DA$. Now by \[RV, Proposition I.2.3\], we have the desired triangle. $\square$ Let $\mathcal{B}$ be an additive category and $\mathcal{C}$ a full subcategory of $\mathcal{B}$. Recall that a morphism $f:B\to C$ with $B\in \mathcal{B}$ and $C\in \mathcal{C}$ is called a [$\mathcal{C}$-preenvelope]{} if the natural map $\Hom_{\mathcal{B}}(C,C')\to \Hom_{\mathcal{B}}(B,C')\to 0$ is exact for any $C'\in \mathcal{C}$. A $\mathcal{C}$-preenvelope $C$ is called a [$\mathcal{C}$-envelope]{} if it is left minimal. Dually, the notion of (pre)covers is defined (\[AR3, E\]). The following proposition involves Auslander-Reiten triangles in subcategories. $\mathrm{([J, Theorem~3.1~\mathrm{and}~Theorem~3.2])}$ Let $\mathcal{T}$ be a triangulated category and $\mathcal{C}$ a full subcategory of $\mathcal{T}$ closed under extensions. 1. Let $X \to Y \to Z \to X[1]$ be an Auslander-Reiten triangle in $\mathcal{T}$ with $Z \in \mathcal{C}$. If there is an object $A' \in \mathcal{C} $ with a nonzero morphism $Z\to A'[1]$, then the following are equivalent. $\bullet$ $X$ has a $\mathcal{C}$-cover of the form $A\to X$. $\bullet$ There is an Auslander-Reiten triangle $A \to B \to Z \to A[1]$ in $\mathcal{C}$. 2. Let $X \to Y \to Z \to X[1]$ be an Auslander-Reiten triangle in $\mathcal{T}$ with $X \in \mathcal{C}$. If there is an object $Z' \in \mathcal{C} $ with a nonzero morphism $Z'\to X[1]$, then the following are equivalent. $\bullet$ $Z$ has a $\mathcal{C}$-envelope of the form $Z\to N$. $\bullet$ There is an Auslander-Reiten triangle $X \to M \to N \to X[1]$ in $\mathcal{C}$. As an application of Theorem 3.4, we now are in a position to reprove the following Happel’s theorem. Our argument is very different from the original one. $\mathrm{([H2, Section~3.4])}$ 1. $K^{b}(\proj A)$ has right Auslander-Reiten triangles if and only if $\id A_{A^{op}}<\infty$. 2. $K^{b}(\inj A)$ has left Auslander-Reiten triangles if and only if $\id A_{A}<\infty$. [Proof.]{} We only prove (2), and (1) is its dual. Let $0\neq Y\in K^{b}(\inj A)$ be indecomposable. Then by Proposition 4.2, there is an Auslander-Reiten triangle $$Y\to L \to \Hom_{A}(DA,Y)[1] \to Y[1]$$ in $K^{b}(\mod A)$. Since $(Y[1])[-1]\overset{\Id_{Y}}\to Y$ is not homotopic to zero, by Lemma 4.3 we have that $K^{b}(\inj A)$ has left Auslander-Reiten triangles if and only if $\Hom_{A}(DA,Y)$ has a $K^{b}(\inj A)$-envelope for any $Y \in K^{b}(\inj A)$. Now it suffices to prove that $\id A_{A}<\infty$ if and only if $\Hom_{A}(DA,Y)$ has a $K^{b}(\inj A)$-envelope for any $Y \in K^{b}(\inj A)$. If $\id A_{A}<\infty$, then the injective dimension of any module in $\proj A$ is finite. Since $\Hom_{A}(DA,Y)$ consists of modules in $\proj A$, then by using induction on the width of $\Hom_{A}(DA,Y)$, one can get an injective resolution $f:\Hom_{A}(DA,Y)\to L$, where $f$ is a quasi-isomorphism and $L\in K^{b}(\inj A)$. If there is a chain map $\alpha :\Hom_{A}(DA,Y)\to I$ with $I\in K^{b}(\inj A)$, then $ f^{\ast}:\Hom_{A}(L,I) \to \Hom_{A}(\Hom_{A}(DA,Y),I)$ is a quasi-isomorphism and $H^{0}(f^{\ast})$ is an isomorphism. Hence there is some $\beta:L\to I$ such that $\alpha$ homotopic to $\beta \cdot f$. If there is some $g$ satisfying $g\cdot f$ homotopic to $f$, then it is a quasi-isomorphism, and hence a homotopic equivalence. It follows that $L$ is a $ K^{b}(\inj A)$-envelope of $\Hom_{A}(DA,Y)$. Now suppose that $\Hom_{A}(DA,Y)$ has a $K^{b}(\inj A)$-envelope for any $Y \in K^{b}(\inj A)$. Let $Y=DA$. Then $\Hom_{A}(DA,Y)\cong A$. Let $A\to I$ be the $K^{b}(\inj A)$-envelope of $A$. Complete it to a triangle $$A\overset{\alpha}\to I \to L \to A[1]$$ in $K^{b}(\mod A)$. Then we have that $\Hom_{K^{b}(\mod A)}(L,Z)\cong 0$ for any $Z\in K^{b}(\inj A)$ by the Wakamatsu lemma (see for example \[J, Lemma 2.1\]); in particular, $\Hom_{K^{b}(\mod A)}(L,(DA)[i])=0$ for all $i$, which implies that $L$ is exact. As a consequence, $\alpha$ is a quasi-isomorphism. Since any injective resolution of $A$ is homotopically equivalent to $I$. It follows that $\id A_{A}<\infty$. $\square$ Note that the functor $-\otimes_{A}DA: K^{b}(\proj A)\to K^{b}(\inj A)$ is an equivalence. Hence the Gorenstein symmetry conjecture, which states that $\id A _{A}=\id A_{A^{op}}$ for any artin algebra $A$, can be reformulated as follows. $\bullet K^{b}(\proj A)$ has right Auslander-Reiten triangles if and only if it has left Auslander-Reiten triangles. Dually, $\bullet K^{b}(\inj A)$ has right Auslander-Reiten triangles if and only if it has left Auslander-Reiten triangles. The following are equivalent. 1. $A$ is a Gorenstein algebra. 2. $K^{b}(\proj A)$ has Auslander-Reiten triangles. 3. $K^{b}(\inj A)$ has Auslander-Reiten triangles. Let $A$ and $B$ be artin algebras. According to \[R\], $A$ and $B$ are derived equivalent if and only if $K^{b}(\proj A)$ and $K^{b}(\proj B)$ are equivalent as triangulated categories. Let $A$ and $B$ be artin algebras. If $A$ and $B$ are derived equivalent, then $A$ is Gorenstein if and only if $B$ is Gorenstein. In general, Theorem $3.4$ only tells us the validity of Auslander-Reiten triangles in $K^{b}(\mod A)$. When consider some subcategory $\mathcal{C}$ of $K^{b}(\mod A)$, one usually relies on the restriction as in Lemma $4.3$. However, it is often difficult to compute the serre dual objects for $\mathcal{A}$. For example, the isomorphism in Lemma 4.2 for projective modules can not be extended to Gorenstein projective version unless $A$ is self-injective, see \[ASS, Lemma 2.12\]. In the following, we only consider the subcategory $K^{b}(\Gproj A$) (resp. $K^{b}(\Ginj A$)). It was proved in \[EJ1\] that over a Gorenstein algebra any finitely generated module admits a finitely generated Gorenstein projective precover. That is, for any $M\in \mod$ A, there is a complex $G_M$ consisting of modules in $\Gproj A$ and a chain map $G_M\to M$ which is a quasi-isomorphism after applying the functor $\Hom_A(G',-)$ for any $G'\in \Gproj A$. Since $M$ has finite Gorenstein projective dimension, $G_M$ can be selected to be in $K^{b}(\Gproj A)$ by \[Ho, Proposition 2.18\]. The dual version for finitely generated Gorenstein injective modules is also valid. Let $A$ be a Gorenstein algebra. Then $K^{b}(\Gproj A)$ has Auslander-Reiten triangles. [Proof.]{} The proof is similar to that of Theorem 4.4. First, we prove that $K^{b}(\Gproj A)$ has right Auslander-Reiten triangles. Let $0\neq X\in K^{b}(\Gproj A)$ be indecomposable. Then by Theorem 3.4, we have an Auslander-Reiten triangle $$Y\to L\to X\to Y[1]$$ in $K^{b}(\mod A)$. We only need to prove that $Y$ has a $K^{b}(\Gproj A)$-cover. In fact, we will prove that any $Y\in K^b(\mod A)$ has a $K^{b}(\Gproj A)$-cover. Note that for any $M\in\mod A$, there is a chain map $G_M\to M$ with $G_M\in K^{b}(\Gproj A)$, which is a quasi-isomorphism after applying the functor $\Hom_A(G',-)$ for any $G'\in \Gproj A$ as above. By using induction on the width of $Y$, we have that there is a chain map $f_Y:G_Y\to Y$ with $G_Y\in K^b(\Gproj A)$, which is also a quasi-isomorphism after applying the functor $\Hom_A(G',-)$ for any $G'\in \Gproj A$. Hence $\Hom_A(G',f_Y)$ is also a quasi-isomorphism for any $G'\in K^{b}(\Gproj A)$. It is easy to see that $G_Y$ is a $K^{b}(\Gproj A)$-cover of $Y$. If we consider the category $K^{b}(\Ginj A)$, we then obtain that $K^{b}(\Ginj A)$ admits left Auslander-Reiten triangles. Note that $-\otimes_{A}DA:K^b(\Gproj A)\to K^b(\Ginj A)$ is an equivalence by \[B\]. This completes the proof.$\square$ $\mathbf{Acknowledgements}$ This research was partially supported by NSFC (Grant Nos. 11171142 and 11571164) and an Innovation Programme of Jiangsu province (Grant No. 0203001505). The authors would like to express their sincere thanks to Professor Silvana Bazzoni and the referees for many considerable suggestions, which have greatly improved this paper. [101]{} I. Assem, D. Simson and A. Skowro$\acute{n}$ski, Elements of the representation theory of associative algebras. Vol. 1. London Mathematical Society Student Texts [65]{}, Cambridge University Press, Cambridge, 2006. M. Auslander, [Representation Dimension of Artin Algebras]{}, Queen Mary College Mathematics Notes, 1971. Republished in: Selected Works of Maurice Auslander, Amer. Math. Soc., Providence, 1999. M. Auslander and I. Reiten, [Representation theory of Artin Algebras III: Almost split sequences]{}, Comm. Algebra [3]{} (1975), 239–294. M. Auslander and I. Reiten, [Representation theory of Artin algebras V. Methods for computing almost split sequences and irreducible morphisms]{}, Comm. Algebra [5]{} (1977), 519–554. M. Auslander and I. Reiten, [Applications of contravariantly finite subcategories]{}. Adv. Math. [86]{} (1991), 111–152. E. Backelin and O. Jaramillo, [Auslander-Reiten sequences and t-structures on the homotopy category of an abelian category]{}, J. Algebra [339]{} (2011), 80–96. R. Bautista, M. José Souto Salorio and R. Zuazua, [Almost split sequences for complexes of fixed size]{}, J. Algebra [287]{} (2005), 140–168. A. Beligiannis, [Cohen-Macaulay modules, (co)torsion pairs and virtually Gorenstein algebras]{}, J. Algebra [288]{} (2005), 137–211. E.E. Enochs, [Injective and flat covers, envelopes and resolvents]{}, Israel J. Math. [39]{} (1981), 189–209. E.E. Enochs and O.M. Jenda, Relative Homological Algebra, Vol. 1. Second revised and extended edition. de Gruyter Expositions in Mathematics [30]{}, Walter de Gruyter GmbH $\&$ Co. KG, Berlin, 2011. E.E. Enochs and O.M. Jenda, Relative Homological Algebra, Vol. 2. de Gruyter Expositions in Mathematics [54]{}, Walter de Gruyter GmbH $\&$ Co. KG, Berlin, 2011. G. Nan, [Auslander-Reiten triangles on Gorenstein derived categories]{}, Comm. Algebra [40]{} (2012), 3912–3919. D. Happel, [On the derived category of a finite-dimensional algebra]{}, Comment. Math. Helvetici [62]{} (1987), 339–389. D. Happel, Triangulated categories in the representation theory of finite-dimensional algebras, London Mathematical Society Lecture Note Series, [119]{}, Cambridge University Press, Cambridge, 1988. D. Happel, [On Gorenstein algebras]{}, Representation theory of finite groups and finite-dimensional algebras (Bielefeld, 1991), 389–404, Progr. Math., [95]{}, Birkh$\mathrm{\ddot{a}}$user, Basel, 1991. H. Holm, [Gorenstein homological dimensions]{}, J. Pure Appl. Algebra [189]{} (2004), 167–193. P. Jørgensen, [Auslander-Reiten triangles in subcategories]{}, J. K-Theory [3]{} (2009), 583–601. J. Rickard, [Morita theory for derived categories]{}, J. London Math. Soc. [39]{} (1989), 436–456. I. Reiten and M. van den Bergh, [Noetherian hereditary abelian categories satisfying Serre duality]{}, J. Amer. Math. Soc. [15]{} (2002), 295–366. [^1]: [E-mail address]{}: [email protected] [^2]: [E-mail address]{}: [email protected] [^3]: [2010 Mathematics Subject Classification]{}: 16G10, 16G70, 18E30. [^4]: [[Keywords]{}: Bounded Homotopy Categories, Auslander-Reiten triangles, Serre duality, Gorenstein algebras.]{}	{ "pile_set_name": "ArXiv" }
--- address: - 'University College London, Department of Mathematics, Gower Street, London WC1E 6BT, UK' - 'TU Wien, Institute for Analysis and Scientific Computing, Wiedner Hauptstr. 8-10 / E101 / 4, 1040 Wien, Austria' author: - Timo Betcke - Alexander Haberl - Dirk Praetorius bibliography: - 'literature.bib' - 'literature\_dpr.bib' title: Adaptive boundary element methods for the computation of the electrostatic capacity on complex polyhedra ---	{ "pile_set_name": "ArXiv" }
--- abstract: 'A calculation of the spin-wave polarization operator is very important for the analysis of the magnetic structure of high temperature superconductors. We analyze the significance of the incoherent part of the spin-wave polarization operator within the framework of the $t$-$J$ model. This part is calculated analytically for small doping with logarithmic accuracy. We conclude that the incoherent part of the spin-wave polarization operator is negligible in comparison with the coherent part.' address: \| School of Physics, The University of New South Wales,\ Sydney, 2052, Australia author: - 'D. W. Murray and O. P. Sushkov$^{a}$' title: 'The incoherent part of the spin-wave polarization operator in the model' --- -1.5cm .5cm 25.cm 16.5cm keywords: antiferromagnetic order, spin fluctuations. Introduction ============ It is widely believed that the spin structure of the ground state of high-temperature copper oxide superconductors is of the type of the Anderson spin-liquid state[@And]. On the other hand it is well known that at zero doping the materials under consideration are insulators with long-range antiferromagnetic (AF) order. The instability of long-range AF order under doping in the $t$-$J$ model was probably first realized by Shraiman and Siggia[@Shr9]. It was later demonstrated in numerous works (see e.g. refs. [@Dom0; @Sin0; @Ede1; @Iga2; @Sus3]). The instability is due to the strong interaction of spin-waves with mobile holes. The spin-wave polarization operator is needed for the analysis of this instability and especially for the derivation of the spin-liquid ground state. The calculation of the so called coherent part of the polarization operator is straightforward. However there is also the incoherent part. The purpose of the present work is to calculate this part and to provide a rigorous proof of its smallness in comparison with the coherent part. It means that in the analysis of the magnetic properties one can neglect the incoherent contribution to the spin-wave polarization operator. This conclusion is in agreement with that of ref.[@Beck93] based on the numerical computation. Hamiltonian =========== The $t$-$J$ model is defined by the following Hamiltonian: $$\label{H} H = H_t + H_J = -t \sum_{<nm>\sigma} ( d_{n\sigma}^{\dag} d_{m\sigma} + \mbox{H.c.} ) + J \sum_{<nm>} {\bf S}_n \cdot {\bf S}_m,$$ where $d_{n\sigma}^{\dag}$ is the creation operator of a hole with spin $\sigma$ ($\sigma= \uparrow, \downarrow$) at site $n$ on a two-dimensional square lattice. The $d_{n\sigma}^{\dag}$ operators act in the Hilbert space with no double electron occupancy. The spin operator is ${\bf S}_n = {1 \over 2} d_{n \alpha}^{\dag} $[$\sigma$]{}$_{\alpha \beta} d_{n \beta}$. $<nm>$ are pairs of nearest-neighbor sites on the lattice. Below we set $J$ as well as the lattice spacing equal to unity. At half-filling (one hole per site) the $t$-$J$ model is equivalent to the Heisenberg AF model[@Hir5; @Gro7] which has long-range AF order in the ground state[@Oit1; @Hus8]. Let us denote the wave function of this ground state by $\|0\rangle$. This is the undoped system. We consider the doped system based on the ground state of the undoped system. In spite of the destruction of the long-range AF order it is convenient to use $\|0\rangle$ and corresponding quasiparticle excitations as a basis set for the doped system. The effective Hamiltonian for the $t$-$J$ model was derived in terms of these quasiparticles in the papers[@Suhf; @Cher3; @Kuch3]: $$\label{Heff} H_{eff}=\sum_{{\bf k}\sigma}\epsilon_{\bf k}h_{{\bf k}\sigma}^{\dag} h_{{\bf k}\sigma}+ \sum_{\bf q}\omega_{\bf q}(\alpha_{\bf q}^{\dag}\alpha_{\bf q} +\beta_{\bf q}^{\dag}\beta_{\bf q}) +H_{h,sw} + H_{hh}.$$ It is expressed in terms of normal spin-waves on an AF background $\alpha_{{\bf q}}, \beta_{{\bf q}}$ (see e.g. review[@Manousakis]) and composite hole operators $h_{{\bf k}\sigma}$ ($\sigma = \pm 1/2$). The summations over ${\bf k}$ and ${\bf q}$ are restricted inside the Brillouin zone of one sublattice where $\gamma_{\bf q}= {1\over 2} (\cos q_x+\cos q_y)\ge 0$. The spin-wave dispersion relation is $$\label{swdisp} \omega_{\bf q}=2\sqrt{1-\gamma_{\bf q}^2}\ , \ \ \ \ \ \omega_{\bf q} \approx \sqrt{2}\|{\bf q}\| \ \mbox{for} \ q \ll 1.$$ The properties of single holes are well established (for a review see ref.[@Dag4]). The wave function of a single hole can be represented as $\psi_{{\bf k}\sigma} = h_{{\bf k}\sigma}^{\dag} \|0\rangle$. At large $t$ the composite hole operator $h_{{\bf k}\sigma}^{\dag}$ has a complex structure. For example at $t/J=3$ the weight of the bare hole in $\psi_{{\bf k} \sigma}$ is about 25%, the weight of configurations of the type “bare hole + 1 magnon” is $\sim$50% and for configurations of the type “bare hole + 2 or more magnons” it is $\sim$25%. The dressed hole is a normal fermion. The hole energy $\epsilon_{\bf k}$ has minima at ${\bf k}={\bf k}_0$, where ${\bf k}_0=(\pm \pi/2, \pm\pi/2)$. For $t \le 5$ the dispersion can be well approximated by the expression[@Sus2] $$\label{hdisp} \epsilon_{\bf k} \approx E_0+2-\sqrt{0.66^2+4.56t^2-2.8t^2\gamma_{\bf k}^2} + {1\over4} \beta_2 (\cos k_x -\cos k_y)^2.$$ The numerical values in this formula are some combinations of the Heisenberg model spin correlators. The constant $E_0$ defines a reference level for the energy. To find $\epsilon_{\bf k}$ with respect to the undoped system one has to set $E_0=0$. However for the present work it is convenient to set $\epsilon_{\bf k_0}=0$ and therefore $E_0=\sqrt{0.66^2+4.56t^2}-2$. The coefficient $\beta_2$ is small and therefore the dispersion is almost degenerate along the face of the magnetic Brillouin zone $\gamma_{\bf k}=0$. According to refs. [@Mart1; @Giam3] $\beta_2 \approx 0.1\,t$ for $t \ge 0.33$. To avoid misunderstanding we note that formula (\[hdisp\]) is not valid for very large $t$ where the hole band width is saturated at the value of the order of unity and does not increase with $t$. However, physically we are interested in $t \approx 3$ (see e.g. refs.[@Esk0; @Fla1; @BCh3]) where (\[hdisp\]) works well. Near the band minima, ${\bf k}_0$ the dispersion (\[hdisp\]) can be presented in the usual quadratic form: $$\label{hdisp1} \epsilon_{\bf p} \approx {1\over2} \beta_1 p_1^2 + {1\over2} \beta_2 p_2^2, \hspace{1.0cm} \beta_2 \ll \beta_1,$$ where $p_1$ ($p_2$) is the projection of ${\bf k}-{\bf k}_0$ on the direction orthogonal (parallel) to the face of the magnetic Brillouin zone (fig. 1). From eq. (\[hdisp\]) for $t \gg 0.33$ we find $\beta_1 \approx 0.65 t$, hence the mass anisotropy is $\beta_1/\beta_2 \approx 7$. For small hole concentrations, i.e. $\delta \ll 1$, the holes are localized in momentum space in the vicinity of the minima of the band ${\bf k}_0 = (\pm \pi/2, \pm \pi/2)$ and the Fermi surface consists of ellipses. (We recall that $\delta=0$ corresponds to an insulator with long-range AF order.) The Fermi energy and Fermi momentum of non-interacting holes are $$\label{ep} \epsilon_F \approx \frac{1}{2} \pi (\beta_1 \beta_2)^{1/2} \delta \qquad \mbox{and} \qquad p_F \approx \sqrt{p_{1F}p_{2F}} \approx (\pi \delta)^{1/2}.$$ The Fermi momentum $p_F$ is measured from the center of the corresponding ellipse. The interaction of a composite hole with spin-waves is of the form (see e.g. refs.[@Mart1; @Liu2; @Suhf]) $$\begin{aligned} \label{hsw} && H_{h,sw} = \sum_{\bf k,q} g_{\bf k,q} \biggl( h_{{\bf k}+{\bf q}\downarrow}^{\dag} h_{{\bf k}\uparrow} \alpha_{\bf q} + h_{{\bf k}+{\bf q}\uparrow}^{\dag} h_{{\bf k}\downarrow} \beta_{\bf q} + \mbox{H.c.} \biggr),\\ && g_{\bf k,q} = 2\sqrt{2}f (\gamma_{\bf k} U_{\bf q} + \gamma_{{\bf k}+{\bf q}} V_{\bf q}) \to 2^{1/4}f\sqrt{{1\over{q}}} (q_x \sin k_x + q_y \sin k_y) \ \mbox{for} \ q \ll 1. \nonumber\end{aligned}$$ Here $U_{\bf q}=\sqrt{{1\over{\omega_{\bf q}}}+{1\over 2}}$ and $V_{\bf q}=-\mbox{sign}(\gamma_{\bf q})\sqrt{{1\over{\omega_{\bf q}}}-{1\over 2}}$ are the parameters of the Bogoliubov transformation diagonalizing the spin-wave Hamiltonian. The hole spin-wave coupling constant $f$ is a function of $t$ evaluated in the work[@Suhf]. For large $t$ the coupling constant is $t$-independent, $f \approx 2$. Let us stress that even for $t > J$ the quasihole-spin-wave interaction (\[hsw\]) has the same form as for $t \ll J$ (i.e. as for bare hole operators) with an added renormalization factor (of the order $J/t$ for $t \gg J$). This remarkable property of the $t$-$J$ model is due to the absence of a single loop correction to the hole-spin-wave vertex. It was first stated implicitly by Kane, Lee and Read[@Kan9]. In refs.[@Mart1; @Liu2; @Suhf] it was explicitly demonstrated that the vertex corrections with different kinematic structure are of the order of a few percent at $t/J \approx 3$. There is also some $q$-dependence of the coupling constant $f$. For example $f(q=\pi)\approx 1.15f(q=0)$ at $t/J = 3$ (see refs. [@Kuch3; @Bel95]). However this dependence is weak and is beyond the accuracy of the calculation of the renormalized value of $f$. Therefore we neglect this dependence. Finally there is the contact hole-hole interaction $H_{hh}$ in the effective Hamiltonian (\[Heff\]). $H_{hh}$ is discussed in detail in refs.[@Cher3; @Kuch3; @Bel95]. It is proportional to some function $A(t)$. For small $t$ this function approaches the value $-0.25$, which gives the well known hole-hole attraction induced by the reduction of the number of missing antiferromagnetic links. However for realistic superconductors $t\approx 3$ (see e.g. refs.[@Esk0; @Fla1; @BCh3]). Surprisingly the function $A(t)$ vanishes exactly at $t\approx 3$ and this means that the mechanism of contact hole-hole attraction is switched off. In contrast the spin-wave exchange mechanism $H_{h,sw}$ is negligible for small $t$ for which $f \sim t$ and it is most important for large $t$ for which $f$ approaches 2. We are interested in “physical” values of $t$: $t \approx 3$. Therefore in the present work we neglect the contact interaction ($H_{hh}=0$) and consider only the hole-spin-wave interaction at $t=3$. The corresponding value of $f$ according to[@Suhf] is $f=1.8$. Thus all numerical values in the present work are calculated at $t=3$, $\beta_1/\beta_2=7$ and $f=1.8$. Polarization Operator ===================== There are two types of spin-waves in the $t$-$J$ model, $\alpha$ and $\beta$. Therefore, one generally has to introduce a set of spin-wave Green’s functions[@Iga22; @Khal3]: $D_{\alpha \alpha}$, $D_{\alpha \beta}$, $D_{\beta \alpha}$, and $D_{\beta \beta}$. However if we restrict ourselves to the long wavelength limit, the consideration can be simplified. It is well known that in the long wavelength limit the Heisenberg model is equivalent to the nonlinear $\sigma$-model (see e.g. review paper[@Manousakis]). Therefore the usual field theory crossing symmetry is valid. The same is valid for the $t$-$J$ model. This is evident from eq. (\[hsw\]): at $q \ll 1$ the vertex $g_{\bf k,q} \approx g_{\bf k-q,q}$. This means that instead of considering a set of Green’s functions one can introduce a single combined Green’s function of vector excitation: $$\label{GF} D(\omega,{\bf q})={{2\omega_{\bf q}}\over{\omega^2-\omega_{\bf q}^2- 2\omega_{\bf q}\Pi(\omega,{\bf q})}},$$ where $\Pi(\omega,{\bf q})$ (or $P(\omega,{\bf q})$) is the mobile holes polarization operator. For the system to be stable the condition $$\label{stab} \omega_{\bf q} + 2 \Pi(0,{\bf q}) > 0$$ should be fulfilled. Otherwise the Green’s function (\[GF\]) would possess poles with imaginary $\omega$. In the single loop approximation for dressed quasiholes the polarization operator is given by the diagram in fig. 2. The quasihole Green’s function in this diagram is determined by the effective Hamiltonian (\[Heff\]). Let us call this contribution to the polarization operator the “coherent part”. The meaning of this term is elucidated below. In this paper we neglect the interaction between quasiholes and consider them as a normal Fermi-liquid. In this approximation the calculation of a single loop polarization operator is very simple. In the static long wavelength limit the result is[@Sus3] $$\label{P0q} \Pi_{coh}(0,{\bf q})\approx -{{\sqrt{2}f^2}\over{\pi\sqrt{\beta_1 \beta_2}}}q \ \ \ \mbox{for} \ \ \ q \ll p_F.$$ This is independent of the hole concentration! Using this result one can easily prove that the stability condition (\[stab\]) is violated because $-2\Pi(0,{\bf q})/\omega_{\bf q}= 2f^2/(\pi \sqrt{\beta_1 \beta_2}) \approx 2.8$. This is a well known fact concerning the instability of long-range AF order under doping (see e.g. refs.[@Shr9; @Dom0; @Sin0; @Ede1; @Iga2; @Sus3]). Let us now explain the obtained result in terms of bare $d_{n\sigma}$ which enter into the bare Hamiltonian (\[H\]). The Green’s function of a bare hole is of the form $$\label{hGF} G(\epsilon,{\bf k})={{Z}\over{\epsilon - \epsilon_{\bf k}+i0}} + G_{incoh},$$ where $Z$ is the quasiparticle residue. If we use the bare hole Green’s function we also have to use the bare hole-spin-wave coupling constant $f_{bare}=2t$ in the vertex $g_{\bf k,q}$ (\[hsw\]). Substitution of the pole (coherent) part of (\[hGF\]) into the single loop polarization operator gives the combination $f_{bare}Z$, but this is exactly the effective coupling constant $f$[@Suhf]. Thus in the single loop approximation the effective theory with Hamiltonian (\[Heff\]) only takes into account the coherent (pole) part of the bare hole Green’s function (\[hGF\]). Because of this the contribution (\[P0q\]) is called the coherent part. The incoherent part of the polarization operator which comes from $G_{incoh}$ in (\[hGF\]) is proportional to the hole concentration $\delta$. Therefore it is negligible for $\delta \ll 1$. This conclusion agrees with that of Becker and Muschelknautz[@Beck93]. Now we will calculate $P_{incoh}$ explicitly. We give two different calculations of $P_{incoh}$. The first calculation is based on the self-energy approach. It is well known that the self-energy as well as the polarization operator is in essence an energy shift of the system. Let us consider the diagram in fig. 3a which represents the energy shift of a hole (solid line) in an external spin-wave (wavy line). If we join up the wavy lines and integrate over the momenta of the spin-wave we get the usual hole self energy (fig. 3b) which is already taken into account in the correct (renormalized) hole dispersion (\[hdisp\]) and the renormalized value of the spin-wave coupling constant $f$. If we join up the solid lines and integrate over the momenta of the hole we get the coherent polarization operator (fig. 3c) which has also already been taken into account. Let us consider now the first correction to the diagram fig. 3a. It is given by the diagrams in fig. 4. (We remind the reader that there is no one loop correction to the spin-wave vertex.) This is the correction to the hole energy in an external spin-wave or the correction to the energy of the spin-wave in the presence of a single hole. The calculation of this correction is very simple if one uses Schrödinger perturbation theory. Consider for example the case when the external spin-wave is of the $\beta$-type, i.e. for this wave the z-projection of spin is $S_z=+1$. The arrows in fig. 4 indicate the projections of hole spin (note that there is a spin flip at each vertex). It is evident that for spin up diagram 4b vanishes and similarly for spin down diagram 4a vanishes. We will see that the integral over $Q$ is logarithmically divergent. Therefore we set $q \ll Q \ll 1$. Near the band minimum the hole energy is quadratic in momenta (\[hdisp1\]). Therefore it can be neglected in comparison with the energy of the intermediate spin-wave $\omega_{\bf Q}$ which is linear in momentum. This means that the energies of all three intermediate states in fig. 4a as well as in fig. 4b are approximately equal to $\omega_{\bf Q}$. Therefore the energy correction corresponding to fig. 4 is independent of the hole spin projection and equals $$\label{deps} \delta \epsilon = -g^2_{{\bf k},{\bf q}}\sum_{\bf Q} {{g^2_{{\bf k},{\bf Q}}}\over{\omega_{\bf Q}^3}} =-{{f^4}\over{\sqrt{2}q}}(q_x \sin k_x +q_y \sin k_y)^2 \int {{(Q_x \sin k_x +Q_y \sin k_y)^2}\over{Q^4}} {{d^2 Q}\over{(2\pi)^2}}.$$ The integration over $Q$ is restricted by the limits $Q_{min} \sim p_F =\sqrt{\pi \delta}$ and $Q_{max} \sim \pi$. The integral over $Q$ is logarithmically divergent and therefore it gives a big logarithm: $\ln (Q_{max}/Q_{min}) \approx 0.5\ln (1/\delta)$. When the momentum of the hole is close to the band minimum, ${\bf k} \approx {\bf k}_0=(\pi/2, \pm \pi/2)$, one gets $$\label{deps1} \delta \epsilon = {{f^4}\over{4\sqrt{2}\pi}} {{(q_x \pm q_y)^2}\over{q}} \ln{\delta}.$$ To find the incoherent spin-wave polarization operator we now need to sum (\[deps1\]) over all occupied hole states (i.e. sum over ${\bf k}$). Thus for $\omega \ll \epsilon_F$ and $q \ll p_F$ the result is $$\label{Pincoh} P_{incoh}(\omega,{\bf q}) \approx -{{f^4}\over{4\sqrt{2}\pi}} \cdot \delta \cdot \ln{{1}\over{\delta}} \cdot q.$$ There are also higher order diagrams which contribute to $P_{incoh}$. An example is presented in fig. 5. However all of these diagrams do not contain the big logarithm ($\ln(1/\delta)$) and therefore can be neglected according to the accepted accuracy. To elucidate the meaning of $P_{incoh}$ we will now give a different derivation of (\[Pincoh\]). The first correction to the single loop polarization operator arises from the diagram presented in fig. 6. After integration over the energies in the closed loops one gets the following expression for the contribution of this diagram: $$\begin{aligned} \label{P6} &&P_6(\omega,{\bf q})=\int {{d^2 k}\over{(2\pi)^2}}{{d^2 Q}\over{(2\pi)^2}} g^2_{{\bf k},{\bf Q}}g^2_{{\bf k},{\bf q}}\times\\ &&\biggl({{n_{\bf k-q}(1-n_{\bf k-Q})-n_{\bf k}(1-n_{\bf k-Q})}\over {(\epsilon_{\bf k-q}-\epsilon_{\bf k}+\omega)^2 (\epsilon_{\bf k-q}-\epsilon_{\bf k-Q}-\omega_{\bf Q}+\omega)}}+ {{n_{\bf k-q}n_{\bf k-Q}-n_{\bf k}n_{\bf k-Q}}\over {(\epsilon_{\bf k-q}-\epsilon_{\bf k}+\omega)^2 (\epsilon_{\bf k-q}-\epsilon_{\bf k-Q}+\omega_{\bf Q}+\omega)}}+\nonumber \\ &&{{-n_{\bf k}(1-n_{\bf k-Q})}\over {(\epsilon_{\bf k-Q}-\epsilon_{\bf k}+\omega_{\bf Q})^2 (\epsilon_{\bf k-Q}-\epsilon_{\bf k-q}+\omega_{\bf Q}-\omega)}}+ {{(1-n_{\bf k})n_{\bf k-Q}}\over {(\epsilon_{\bf k-Q}-\epsilon_{\bf k}-\omega_{\bf Q})^2 (\epsilon_{\bf k-Q}-\epsilon_{\bf k-q}-\omega_{\bf Q}-\omega)}} \biggr),\nonumber\end{aligned}$$ where $n_{\bf k}$ is the ground state occupation number. The first two terms in the expression (\[P6\]) have second powers of $(\epsilon_{\bf k-q}-\epsilon_{\bf k}+\omega)$ in the denominator. This factor is exactly the denominator of the one-loop polarization operator fig. 2. It means that the first two terms correspond to the renormalization of the pole (coherent) part of the hole Green’s function (\[hGF\]). They are actually already taken into account in $P_{coh}$ because for the calculation of $P_{coh}$ we have used the dressed hole Green’s function which absorbs all of the corrections to the pole part. Thus the new contribution, corresponding to $P_{incoh}$, arises from the last two terms in (\[P6\]). After carrying out some algebraic manipulation and using the fact that $q$ is small (and so, for example $\epsilon_{\bf k-q} \approx \epsilon_{\bf k}$) $P_{incoh}$ can be written as $$\label{Pincoh1} P_{incoh}(0,{\bf q})=-2\int {{d^2 k}\over{(2\pi)^2}}{{d^2 Q}\over{(2\pi)^2}} g^2_{{\bf k},{\bf Q}}g^2_{{\bf k},{\bf q}} {{n_{\bf k}(1-n_{\bf k-Q})}\over {(\omega_{\bf Q}+\epsilon_{\bf k-Q}-\epsilon_{\bf k})^3}}.$$ It is evident that this expression is equivalent to (\[deps\]). Neglecting $\epsilon_{\bf k-Q}-\epsilon_{\bf k}$ in comparison with $\omega_{\bf Q}$ and carrying out the integration over ${\bf Q}$ with logarithmic accuracy we once more come to the result (\[Pincoh\]) for $P_{incoh}$. Conclusion ========== In the present work we have analytically calculated the incoherent contribution to the spin-wave polarization operator. This analytical calculation is possible because of the logarithmic enhancement of the corresponding diagrams. The ratio of the incoherent contribution (\[Pincoh\]) to the coherent one (\[P0q\]) is $$\label{ratio} {{P_{incoh}(0,{\bf q})}\over{P_{coh}(0,{\bf q})}}= {1\over{8}}f^2 \sqrt{\beta_1 \beta_2} \cdot \delta \cdot \ln{{1}\over{\delta}} \ .$$ For the parameters of the $t$-$J$ model corresponding to realistic high-T$_c$ superconductors and for hole concentration $\delta \le 0.2$ this ratio does not exceed 10%. It means that in the analysis of the magnetic properties one can neglect the incoherent contribution to the spin-wave polarization operator. Also at the Budker Institute of Nuclear Physics, 630090 Novosibirsk, Russia. P. W. Anderson, Science 235 (1987) 1196. B. Shraiman and E. Siggia, Phys. Rev. Lett. 62 (1989) 1564. T. Dombre, J. Physique 51 (1990) 847. A. Singh and Z. Tesanovic, Phys. Rev. B 41 (1990) 614. R. Eder, Phys. Rev. B 43 (1991) 10706. J. Igarashi and P. Fulde, Phys. Rev. B 45 (1992) 10419. O. P. Sushkov and V. V. Flambaum, Physica C 206 (1993) 269. K. W. Becker and U. Muschelknautz, Phys. Rev. B 48 (1993) 13826. J. E. Hirch, Phys. Rev. Lett. 54 (1985) 1317. C. Gross, R. Joynt and T. M. Rice, Phys. Rev. B 36 (1987) 381. J. Oitmaa and D. D. Betts, Can. J. Phys. 56 (1971) 897. D. A. Huse, Phys. Rev. B 37 (1988) 2380. O. P. Sushkov, Phys. Rev. B 49 (1994) 1250. A. L. Chernyshev, A. V. Dotsenko and O. P. Sushkov, Phys. Rev. B 49 (1994) 6197. M. Yu. Kuchiev and O. P. Sushkov, Physica C 218 (1993) 197. E. Manousakis, Rev. Mod. Phys. 63 (1991) 1. E. Dagotto, Rev. Mod. Phys. 66 (1994) 763 and references therein. O. P. Sushkov, Solid State Communications 83 (1992) 303. G. Martínez and P. Horsch, Phys. Rev. B 44 (1991) 317. T. Giamarchi and C. Lhuillier, Phys. Rev. B 47 (1993) 2775. H. Eskes, G. A. Sawatzky and L. F. Feiner, Physica C 160 (1989) 424. V. V. Flambaum and O. P. Sushkov, Physica C 175 (1991) 347. V. I. Belinicher and A. L. Chernyshev, Phys. Rev. B 47 (1993) 390. Z. Liu and E. Manousakis, Phys. Rev. B 45 (1992) 2425. C. L. Kane, P. A. Lee and N. Read, Phys. Rev. B 39 (1989) 6880. V. I. Belinicher, A. L. Chernyshev, A. V. Dotsenko and O. P. Sushkov, Phys. Rev. B 51 (1995) 6076. J. Igarashi and P. Fulde, Phys. Rev. B 45 (1992) 12357. G. Khaliullin and P. Horsch, Phys. Rev. B 47 (1993) 463. [FIGURE CAPTIONS]{}\ FIG. 1. The magnetic Brillouin zone.\ FIG. 2. The coherent spin-wave polarization operator.\ Fig. 3. a) The leading order contribution to the energy shift of a hole (solid line) in an external spin-wave (dashed line). b) The hole self-energy. c)The coherent spin-wave polarization operator.\ FIG. 4. The fourth order correction to the energy of a hole in an external spin-wave.\ FIG. 5. A typical fifth order diagram for the energy of a hole in an external spin-wave.\ FIG. 6. The two loop correction to the spin-wave polarization operator.	{ "pile_set_name": "ArXiv" }
--- abstract: 'We give an explicit geometric description of the $\times2,\times3$ system, and use this to study a uniform family of Markov partitions related to those of Wilson and Abramov. The behaviour of these partitions is stable across expansive cones and transitions in this behaviour detects the non-expansive lines.' address: - 'School of Mathematics, University of East Anglia, Norwich, CV4 7AL, UK' - 'Centro de Modelamiento Matem[á]{}tico, Av. Blanco Encalada 2120 Piso 7, Santiago de Chile' author: - Thomas Ward - Yuki Yayama title: 'Markov partitions reflecting the geometry of $\times2,\times3$' --- [^1] Introduction {#sec:introduction} ============ Markov partitions are a powerful tool in the study of hyperbolic diffeomorphisms of manifolds. Explicit constructions of Markov partitions for hyperbolic toral automorphisms of the $2$-torus $\mathbb T^2$ in the work of Adler and Weiss [@MR0257315] are an important paradigmatic example, and in special situations the tight connection between the geometry of the map and the partition found in [@MR0257315] is extended to automorphisms of $\mathbb T^d$ with $d>2$ by Manning [@MR1895206]. On the other hand, maps of objects that are not quite manifolds arise naturally in dynamics, notably as attractors of smooth maps in work of Bowen [@MR0482842] and Williams [@MR0348794]. Thus a natural extension of the classical theory of smooth maps of compact manifolds concerns maps of solenoids; a useful overview and the history may be found in a paper of Takens [@MR2164385]. The simplest solenoids are algebraic: compact groups that are locally isometric to products of local fields. The structure of a tangent space comprising a product of local fields including non-Archimedean ones may be used to study various dynamical properties of automorphisms of solenoids: exotic orbit-growth properties by Chothi, Everest, Miles, Stevens and the first author [@MR1461206], [@MR2339472]; entropy and structure of $\mathbb Z^d$-actions of entropy rank one by Einsiedler and Lind [@MR2031042]; topological entropy by Lind and the first author [@MR961739], [@MR1882488]. Our purpose here is to study geometrically natural Markov partitions like those used by Abramov [@MR0123346] and Wilson [@MR0390181] for one of the simplest examples in which non-Archimedean directions arise in the tangent space, and to study how the structure of those partitions changes in expansive cones. This gives a simple geometrical instance of the ‘subdynamics philosophy’ of Boyle and Lind [@MR1355295]. A combinatorial instance of the same kind of structure appears in work of Miles and the first author [@MR2279271], where it is shown that directional zeta functions detect the non-expansive set for systems of entropy rank one. In order to do this, we describe the structure of the space obtained by taking the invertible extension of the $\mathbb N^2$-action generated by $x\mapsto2x\pmod{1}$ and $x\mapsto3x\pmod{1}$ on the additive circle in a geometric way. To simplify matters we concentrate on this one example: the same kind of construction works in those systems of entropy rank one with an adelic covering space, but is significantly more involved. In principle the Markov and generating properties of the partitions can be shown from our geometric description, but for brevity we deduce some of these properties from Wilson’s results. The geometry of $\times2,\times3$ ================================= We make use of a simple version of the adelic machinery; an elegant account may be found in Weil [@MR0234930 Ch. 4]. We wish to describe the group $X=\widehat{\mathbb Z[\frac{1}{6}]}$ of characters on $\mathbb Z[\frac{1}{6}]$ and its metric structure: this group carries the automorphisms $\alpha^{(1,0)}$ and $\alpha^{(0,1)}$ dual to the automorphisms $x\mapsto 2x$ and $x\mapsto3x$ on $\mathbb Z[\frac{1}{6}]$, and is a presentation of the invertible extension of the $\mathbb N^2$ action generated by $x\mapsto2x\pmod{1}$ and $x\mapsto3x\pmod{1}$ on $\mathbb T$. For a prime $p$, define the local field $\mathbb Q_p$ to be the set of formal power series $\sum_{n{\geqslant}k}a_np^n$ with digits $a_n\in\{0,1,\dots,p-1\}$ and $k\in\mathbb Z$, and with the non-Archimedean metric $\vert\cdot\vert_p$ induced by the $p$-adic norm $\vert\sum_{n{\geqslant}k}a_np^n\vert_p=p^{-k}$ if $a_k\neq0$. Notice that $\mathbb Q$ is a subfield of each $\mathbb Q_p$ and each $\mathbb Q_p$ has a maximal compact subring $\mathbb Z_p=\{x\in\mathbb Q_p\mid\vert x\vert_P{\leqslant}1\}$. The homomorphism $$\begin{aligned} \Delta:\mathbb Z[\textstyle\frac{1}{6}]&\longrightarrow&\mathbb R\times\mathbb Q_2\times\mathbb Q_3\\ r&\longmapsto&(r,r,r)\end{aligned}$$ embeds $\mathbb Z[\frac{1}{6}]$ as a discrete (and hence closed) subgroup of $\mathbb R\times\mathbb Q_2\times\mathbb Q_3$ with respect to the metric $\metric(x,y)=\max\left\{\vert x_{\infty}-y_{\infty}\vert,\vert x_2-y_2\vert_2,\vert x_3-y_3\vert_3\right\}$, where we write $x=(x_{\infty},x_2,x_3)\in\mathbb R\times\mathbb Q_2\times\mathbb Q_3$. Write $G=\mathbb R\times\mathbb Q_2\times\mathbb Q_3$ and $\Gamma=\Delta(\mathbb Z[\frac{1}{6}])$. The group $X$ is the quotient $G/\Gamma$ (this may be seen from Weil [@MR0234930 Ch. 4]), and we wish to describe this quotient space in a concrete way. In order to motivate this, notice that a toral automorphism may be constructed as follows. The identity map embeds $\mathbb Z^d$ as a discrete subgroup of $\mathbb R^d$, and a choice of coset representatives for $\mathbb R^d/\mathbb Z^d$ gives an explicit geometric description of the map induced on the torus by any automorphism of $\mathbb R^d$ preserving $\mathbb Z^d$. In order to make this note self-contained and to rehearse the kind of calculation needed later, we include the proof of the following two lemmas, which are simple instance of a well-known principle (see Weil [@MR0234930 Ch. 4] or Hewitt and Ross [@MR551496 § II.10, Th. 10.15]). The set $F=[0,1)\times\mathbb Z_2\times\mathbb Z_3$ is a fundamental domain for $\Gamma$ in $G$. The first step is to check that $F$ is big enough: given $x\in G$, can we find $\gamma=(r,r,r)\in\Gamma$ with $x-\gamma\in F$? To do this, write $\{\sum_{n{\geqslant}k}a_np^n\}=\sum_{n=k}^{-1}a_np^n$ for the fractional part of $x\in\mathbb Q_p$; $\{t\}$ for the fractional part and $\lfloor t\rfloor$ for the integer part of $t\in\mathbb R$. A calculation shows that if $$r=\{x_2\}+\{x_3\}+\lfloor(x_{\infty}-\{x_2\}-\{x_3\}\rfloor$$ then $r\in\mathbb Z[\frac{1}{6}]$ and $(x_{\infty}-r,x_2-r,x_3-r)\in F$ as required. The second step is to check that $F$ is small enough: if $x,y\in F$ define the same coset of $\Gamma$ then they are equal. Assume therefore that $x,y\in F$ and $x-y=(r,r,r)\in\Gamma$. Then $x_2-y_2\in\mathbb Z_2\cap\mathbb Z[\frac{1}{6}]=\mathbb Z[\frac{1}{3}]$ and $x_3-y_3\in\mathbb Z_3\cap\mathbb Z[\frac{1}{6}]=\mathbb Z[\frac{1}{2}]$, so $r\in\mathbb Z[\frac{1}{3}]\cap\mathbb Z[\frac{1}{2}]=\mathbb Z$, and therefore $\{x_{\infty}\}=\{y_{\infty}\}$, so $x_{\infty}=y_{\infty}$ and $r=0$ as required. This means that there is a bijection $G/\Gamma\longleftrightarrow F$; to go further we need to describe the image of the group operation on $G/\Gamma$ under this bijection. \[lemma:additionrule\] For $s,t\in G$, $$(t+\Gamma)+(s+\Gamma)=$\{t_{\infty}+s_{\infty}\},t_2+s_2-\lfloor t_{\infty}+s_{\infty}\rfloor,t_3+s_3-\lfloor t_{\infty}+s_{\infty}\rfloor$+\Gamma$$ is the unique coset representative for $t+s$ in $F$. We wish to find the unique $u\in F$ with the property that there is some $(r,r,r)$ in $\Gamma$ with $u=t+s-r$. We must have $u_{\infty}=\{t_{\infty}+s_{\infty}\}$, which forces $r$ to be $\lfloor t_{\infty}+s_{\infty}\rfloor$; notice that we also then have $$u_2=t_2+s_2-\lfloor t_{\infty}+s_{\infty}\rfloor\in\mathbb Z_2$$ and $$u_3=t_3+s_3-\lfloor t_{\infty}+s_{\infty}\rfloor\in\mathbb Z_3$$ since $\mathbb Z_2,\mathbb Z_3$ are rings. Lemma \[lemma:additionrule\] may be written as follows: the operation $$\label{eqn:additionformula} t\plus s=$\{t_{\infty}+s_{\infty}\},t_2+s_2-\lfloor t_{\infty}+s_{\infty}\rfloor,t_3+s_3-\lfloor t_{\infty}+s_{\infty}\rfloor$$$ makes $F$ into a group $X=(F,\plus)$ isomorphic to $G/\Gamma$. An explicit metric on $X$ is given by $$\overline{\metric}(x+\Gamma,y+\Gamma)=\min_{r\in\mathbb Z[\frac{1}{6}]}\max\{\vert x_{\infty}-y_{\infty}+r\vert_{\infty}, \vert x_{2}-y_{2}+r\vert_{2},\vert x_{3}-y_{3}+r\vert_{3}\}.$$ Wilson [@MR0390181] describes the same solenoid in a different way, as a projective limit of circles $$\label{equation:wilsonform} X\cong\{z\in\mathbb T^{\mathbb N_0}\mid 6z_{k+1}=z_k\pmod{1}\mbox{ for all }k{\geqslant}1\};$$ points $z,z'$ in this description are close if their coordinates $z_k,z'_k$ are close in $\mathbb T$ for $1{\leqslant}k{\leqslant}K$ for large $K$. The isomorphism in may be thought of as follows. A given point $z=(z_k)_{k{\geqslant}0}$ in the right-hand side of defines an element $z_0\in\mathbb T$; each choice of $z_{k+1}$ given $z_k$ defines a unique pair $x_2^{(k)}\in\{0,1\}$ and $x_3^{(k)}\in\{0,1,2\}$ satisfying $z_{k+1}=\frac16z_{k}+\frac{x_2^{(k)}}{2}+\frac{x_3^{(k)}}{6}$ (thinking of $z_{k+1}$ as a real number in $[0,1)$). The isomorphism is then defined by sending $z$ to the point $$z_0,\sum_{k{\geqslant}0}x_2^{(k)}2^k,\sum_{k{\geqslant}0}x_3^{(k)}3^k$\in X$. This isomorphism respects the metric structures (nearby points in $X$ correspond to nearby points in the projective limit) and is equivariant with respect to the automorphisms we study. The automorphisms $\alpha^{(1,0)}:x\mapsto 2x$ and $\alpha^{(0,1)}:x\mapsto3x$ on $G$ preserve $\Gamma$ and therefore define automorphisms of $X=(F,\plus)$. To see how the group $X$ works, we compute the automorphisms $\alpha^{(0,1)}$ (multiplication by $3$), $\alpha^{(-1,0)}$ (multiplication by $\frac{1}{2}$), and $\alpha^{(-1,1)}$ (multiplication by $\frac{3}{2}$) explicitly. By , $$\alpha^{(0,1)}(x)=x\plus x\plus x=$\{3x_{\infty}\},3x_2-\lfloor 3x_{\infty}\rfloor,3x_3-\lfloor 3x_{\infty}\rfloor$.$$ Notice that the map $\alpha^{(0,1)}$ locally expands the real component by a factor of $3$, locally contracts the $3$-adic component by a factor of $3$, and is a local isometry on the $2$-adic component. Write $x_p=\sum_{n{\geqslant}k}x_p^{(n)}p^n$ with digits $x_p^{(n)}\in\{0,1,\dots,p-1\}$ for $n{\geqslant}k$. Then $$\alpha^{(-1,0)}(x)=\left(\textstyle\frac{1}{2}+\frac{1}{2}x_2^{(0)},\frac{1}{2}x_2+\frac{1}{2}x_2^{(0)}, \frac{1}{2}x_3+\frac{1}{2}x_2^{(0)}\right)$$ (this is most easily verified by doubling the right-hand side). Finally, by combining the two calculations we see that $\alpha^{(-1,1)}(x)$ is $$$\left\{\textstyle\frac{3}{2}x_{\infty}+\frac{3}{2}x_2^{(0)}\right\}, \textstyle\frac{3}{2}x_2+\frac{3}{2}x_2^{(0)}-\left\lfloor\textstyle\frac{3}{2}x_{\infty}+\frac{3}{2}x_2^{(0)} \right\rfloor, \frac{3}{2}x_3+\frac{3}{2}x_2^{(0)}-\left\lfloor\textstyle\frac{3}{2}x_{\infty}+\frac{3}{2}x_2^{(0)} \right\rfloor$.$$ Locally the action of $\alpha^{(a,b)}$ multiplies by $2^a3^b$, and therefore acts on each of the three directions in the tangent space as shown in Table \[table\] ($u$, $s$ denote unstable and stable directions). [region]{} [$\mathbb R$]{} $\mathbb Q_2$ $\mathbb Q_3$ -------------------- ----------------- --------------- --------------- $a>0,b>0$ $u$ $s$ $s$ $a<0,b>0,2^a3^b>1$ $u$ $u$ $s$ $a>0,b<0,2^a3^b>1$ $u$ $s$ $u$ $a<0,b<0$ $s$ $u$ $u$ $a>0,b<0,2^a3^b<1$ $s$ $s$ $u$ $a<0,b>0,2^a3^b<1$ $s$ $u$ $s$ : \[table\]Stable and unstable directions. The first three regions shown in Table \[table\] are the expansive regions in the sense of [@MR1355295] and [@MR1869066] (expansive regions are defined in the Grassmannian space of lines in $\mathbb R^2$, of which the circle is a two-fold cover; the table shows the six regions in the cover). There are three non-expansive lines $a=0$ (containing maps like $\alpha^{(0,1)}$, which behaves like an isometry on the $2$-adic direction), $b=0$ (containing maps like $\alpha^{(1,0)}$, which behaves like an isometry on the $3$-adic direction) and $2^a3^b=1$ (which does not contain any lattice points, but has a sequence of lattice points $(a_k,b_k)$ converging to it with the property that the real Lyapunov exponent $\log\vert 2^{a_k}3^{b_k}\vert$ of the map $\alpha^{(a_k,b_k)}$ converges to zero as $k\to\infty$). Stable Markov partitions ======================== It is clear that there cannot be a single finite partition that is generating for all the maps $\alpha^{(a,b)}$ as $(a,b)$ varies inside an expansive cone because the set of topological entropies of the maps in a cone is unbounded. Thus, what we mean by “stable” is that the Markov partition for $\alpha^{(a,b)}$ is constructed in a uniform manner across all $(a,b)\in\mathbb Z^2$. We will see later that the geometry of how the map acts on an atom of the partition is uniform across each expansive cone but changes at each non-expansive direction. Recall that the naïve height (in the sense of Diophantine geometry) of a non-zero rational $r/s$ is defined to be $H(r/s)=\max\{\vert r\vert,\vert s\vert\}$. Thus Abramov’s formula [@MR0123346] for the entropy of an automorphism of a one-dimensional solenoid may be written $h(T)=\log H(r/s)$ if $T$ is the map dual to multiplication by $r/s$. \[theorem:stablemarkovpartitions\] For each $(a,b)\in\mathbb Z^2\setminus\{(0,0)\}$ let $\xi^{(a,b)}$ denote the partition $$\left\{ A_j=\left[\textstyle\frac{j}{H(2^a3^b)},\frac{j+1}{H(2^a3^b)}\right)\times\mathbb Z_2\times\mathbb Z_3\mid 0{\leqslant}j<H(2^a3^b) \right\}.$$ Then $\{\xi^{(a,b)}\}$ is a stable family of Markov partitions whose geometry detects the non-expansive directions of $\alpha$. The partition $\xi^{(a,b)}$ is generating for $\alpha^{(a,b)}$ if and only if $\alpha^{(a,b)}$ is expansive. The theory of Markov partitions in the topological (rather than smooth) setting is developed by Adler [@MR1477538]; by ‘Markov’ we mean that the partition obtained from $\xi^{(a,b)}$ by using open intervals in the real coordinate instead of half-open intervals satisfies [@MR1477538 Def. 6.1]. Much of the proof of Theorem \[theorem:stablemarkovpartitions\] will use results from Wilson [@MR0390181] that conceal the geometry of the actions. In order to see how the maps act geometrically, we illustrate the result by describing the partition and the action of the map on the partition in some representative directions. In each figure the image of the atom $A_0$ of the partition is shaded. \[example:one\]Consider the direction $(1,0)$, with corresponding map $$\alpha^{(1,0)}(x)=$\{2x_{\infty}\},2x_2-\lfloor 2x_{\infty}\rfloor,2x_2-\lfloor 2x_{\infty}\rfloor$.$$ The partition $\xi^{(1,0)}$ simply divides the real coordinate into $[0,\frac12)$ and $[\frac12,1)$. We compute $$\alpha^{(1,0)}(\xi^{(1,0)})= \{ [0,1)\times2\mathbb Z_2\times\mathbb Z_3, [0,1)\times(1+2\mathbb Z_2)\times\mathbb Z_3 \}$$ and $$\alpha^{(2,0)}(\xi^{(1,0)})= \{ [0,1)\times(4\mathbb Z_2\cup1+4\mathbb Z_2)\times\mathbb Z_3, [0,1)\times(2+4\mathbb Z_2\cup3+4\mathbb Z_2)\times\mathbb Z_3 \}.$$ Similarly, $$\alpha^{(-1,0)}(\xi^{(1,0)})= \{ ([0,\textstyle\frac14)\cup[\frac12,\frac34))\times\mathbb Z_2\times\mathbb Z_3, ([\frac14,\frac12)\cup[\frac34,1))\times\mathbb Z_2\times\mathbb Z_3 \}$$ and $\alpha^{(-2,0)}(\xi^{(1,0)})$ is the partition into the sets $$$ [0,\textstyle\frac18)\cup[\frac14,\frac38)\cup[\frac12,\frac58)\cup[\frac34,\frac78) $\times\mathbb Z_2\times\mathbb Z_3$$ and $$$[\textstyle\frac18,\frac14)\cup[\frac38,\frac12)\cup[\frac58,\frac34)\cup[\frac78,1) $\times\mathbb Z_2\times\mathbb Z_3.$$ These partitions are illustrated in Figure \[fig:multby2a\] for the forward direction and Figure \[fig:multby2b\] for the reverse direction. Notice that $\bigvee_{n=-\infty}^{\infty}\alpha^{(n,0)}(\xi^{(1,0)})$ does not separate the $\mathbb Z_3$ coordinate, so the partition is not generating for $\alpha^{(1,0)}$. However, this does show that the system $(X,\alpha^{(1,0)})$ may be realized as an isometric extension of a base system (which is an almost $1$-$1$ image of a full Bernoulli $2$-shift) by $\mathbb Z_3$. The expansive region $ab>0$ is particularly simple because the system $(X,\alpha^{(a,b)})$ is (at each point with $a>0,b>0$) simply the invertible extension of the map $x\mapsto2^a3^bx\pmod{1}$ on the circle, and $\xi^{(a,b)}$ is the usual partition into intervals of width $\frac{1}{2^a3^b}$ on $[0,1)$ lifted to $X$. The action of $\alpha^{(1,1)}$ (multiplication by $6$) is illustrated in Figure \[fig:multby6a\] for the forward direction and Figure \[fig:multby6b\] for the reverse direction. Now consider the map $\alpha^{(-1,1)}$ (multiplication by $\frac32$). For this map the real and the $2$-adic directions are unstable and the $3$-adic direction is stable. The partition $\xi^{(-1,1)}$ divides the real coordinate into three pieces. A calculation shows that $\alpha^{(-1,1)}(\xi^{(-1,1)})$ consists of the sets $$[0,\textstyle\frac12)\times\mathbb Z_2\times3\mathbb Z_3\cup [\frac12,1)\times\mathbb Z_2\times(3\mathbb Z_3+2),$$ $$[\textstyle\frac12,1)\times\mathbb Z_2\times3\mathbb Z_3\cup [0,\frac12)\times\mathbb Z_2\times(3\mathbb Z_3+1),$$ and $$[0,\textstyle\frac12)\times\mathbb Z_2\times(3\mathbb Z_3+2)\cup [\frac12,1)\times\mathbb Z_2\times(3\mathbb Z_3+1).$$ The image of $A_0$ under the maps $\alpha^{(-1,1)}$ and $\alpha^{(1,-1)}$ are shown in Figure \[fig:multby3over2a\]. A similar calculation shows that $\alpha^{(1,-1)}(\xi^{(-1,1)})$ consists of the sets $$[0,\textstyle\frac29)\times2\mathbb Z_2\times\mathbb Z_3\cup [\frac13,\textstyle\frac59)\times(1+2\mathbb Z_2)\times\mathbb Z_3\cup [\frac23,\textstyle\frac89)\times2\mathbb Z_2\times\mathbb Z_3,$$ $$[\textstyle\frac29,\textstyle\frac49)\times2\mathbb Z_2\times\mathbb Z_3\cup [\frac59,\textstyle\frac79)\times(1+2\mathbb Z_2)\times\mathbb Z_3\cup $ [\frac89,\textstyle1)\times2\mathbb Z_2\times\mathbb Z_3\cup [0,\textstyle\frac19)\times(1+2\mathbb Z_2)\times\mathbb Z_3 $,$$ and the complement of their union. Notice that (for example) $\alpha^{-1}A_0\cap A_1\cap\alpha A_0$ does not consist of a single rectangle. Assume first that $a>0$, $b>0$, so that $2^a3^b\in\mathbb N$, and write $\alpha=\alpha^{(a,b)}$, $\xi=\xi^{(a,b)}$ throughout; the partitions $\alpha^{-1}(\xi),\xi, \alpha(\xi)$ are illustrated in Figures \[fig:multby6a\] and \[fig:multby6b\] with the image and pre-image of $A_0$ shaded for the case $(a,b)=(1,1)$. We claim that the combinatorics of a full shift on $6$ symbols suggested by Figures \[fig:multby6a\] and \[fig:multby6b\] is indeed the case. This (and other steps flagged below) may in principle be extracted from Wilson’s paper [@MR0390181] but we prove it here to show how the map works. We first need to check that an atom of the form $$\alpha(A_{i_1})\cap\cdots\cap\alpha^{n}(A_{i_{n}}),$$ for any choice of $i_1,\dots,i_{n}\in\{0,1,\dots,2^a3^b-1\}$, is a rectangle of the shape $$[0,1)\times (t_n+2^{an}\mathbb Z_2)\times(s_n+3^{bn}\mathbb Z_3)$$ with an explicit description of $t_n\in\{0,1,\dots 2^{an}-1\}$ and $s_n\in\{0,1,\dots,3^{bn}-1\}$. In order to do this, we need some notation for the sets arising as the map is iterated. The first iteration is straightforward, and we can write $$\alpha(A_k)=[0,1)\times(2^a\mathbb Z_2-k)\times(3^b\mathbb Z_3-k)$$ for $0{\leqslant}k{\leqslant}2^a3^b-1.$ The next iteration is more complicated, because the image involves reduction modulo $\Gamma$. We compute $$\label{equation:suggestionE1fromyuki} \alpha^2(A_k)=\bigsqcup_{\ell_1=0}^{2^a3^b-1} A_{k,\ell_1},$$ where $$A_{k,\ell_1}= [0,1)\times$2^a3^b(2^a\mathbb Z_2-k)-\ell_1$\times $2^a3^b(3^b\mathbb Z_3-k)-\ell_1$$$ ($\sqcup$ denoting a disjoint union). Continue, arriving at the notation $$\label{equation:suggestionE2fromyuki} \alpha^n(A_k)=\bigsqcup_{\ell_1=0}^{2^a3^b-1} \cdots \bigsqcup_{\ell_{n-1}=0}^{2^a3^b-1}A_{k,\ell_1,\dots,\ell_{n-1}}$$ for $n{\geqslant}2$, in which each $A_{k,\ell_1,\dots,\ell_{n-1}}$ is a set of the form $$[0,1)\times $ 2^{an}\mathbb Z_2-C(k,\ell_1,\dots,\ell_{n-2})-\ell_{n-1} $ \times $ 3^{bn}\mathbb Z_3-C(k,\ell_1,\dots,\ell_{n-2})-\ell_{n-1} $$$ where $$C(k,\ell_1,\dots,\ell_{n-2})= k(2^{a}3^{b})^{n-1}+\ell_1 (2^{a}3^{b})^{n-2}+\ell_2(2^{a}3^{b})^{n-3}+\cdots+\ell_{n-2}2^{a}3^{b}.$$ Using this description, we claim that an atom in $\bigvee_{j=1}^{n}\alpha^j(\xi)$ can be written in the form $$\label{equation:nameofatomforwardpositivequadrant} \alpha(A_{i_1})\cap\alpha^2(A_{i_2})\cap\cdots\cap\alpha^n(A_{i_n})= A_{i_n,i_{n-1},\dots,i_1}$$ for $n{\geqslant}2$ and some $0{\leqslant}i_j<2^a3^b$, $1{\leqslant}j{\leqslant}n$ where the right-hand side is defined as above. We prove the claim in by induction on the length $n$ starting with $n=2$. Clearly $\alpha^2(A_{i_2})\supseteq A_{i_2,i_1}$ by definition. Now $$\begin{aligned} A_{i_2,i_1}&=&[0,1)\times$2^{2a}\mathbb Z_2-i_22^a3^b-i_1$\times$3^{2b}\mathbb Z_3-i_22^a3^b-i_1$\\ &\subseteq&[0,1)\times$2^a\mathbb Z_2-i_1$\times$3^b\mathbb Z_3-i_1$=\alpha(A_{i_1})\end{aligned}$$ since $i_22^a3^b\mathbb Z_2\subseteq 2^a\mathbb Z_2$, and similarly for the other terms, so $\alpha(A_{i_1})\supseteq A_{i_2,i_1}$. Thus $$\alpha(A_{i_1})\cap\alpha^2(A_{i_2})\supseteq A_{i_2,i_1}.$$ We now claim that $\alpha(A_{i_1})\cap\alpha^2(A_{i_2})= A_{i_2,i_1}$ by using and showing that $$A_{i_2,\ell}\cap \alpha(A_{i_1})\neq\emptyset$$ for $0{\leqslant}\ell<2^a3^b$ implies that $\ell=i_1$. To see this, note first that if $A_{i_2,\ell}\cap \alpha(A_{i_1})\neq\emptyset$ then $A_{i_2,\ell}\subset\alpha(A_{i_1})$. Suppose that there is some $i_1'\neq i_1$, both in $\{0,\dots,2^a3^b-1\}$, with $A_{i_2,i_1'}\cap\alpha(A_{i_1})\neq\emptyset$. Then $i_1-i_1'=2^ak_1$ and $i_1-i_1'=3^bk_2$ for some $k_1,k_2\in\mathbb Z$, so (since $2$ and $3$ are coprime), $i_1\equiv i_1'\pmod{2^a3^b}$ and therefore $i_1=i_1'$. Now assume that holds for $n{\leqslant}k$. First notice that $$\alpha^{k+1}(A_{i_{k+1}})\supset A_{i_{k+1},i_k,\dots,i_1},$$ and we claim that $$\label{equation:inductivesteppositivequadrantforward} A_{i_k,\dots,i_1}\supset A_{i_{k+1},i_k,\dots,i_1}.$$ Since $$\begin{aligned} A_{i_k,\dots,i_1}&=&[0,1)\times$ 2^{ak}\mathbb Z_2-i_k(2^a3^b)^{k-1}-i_{k-1}(2^a3^b)^{k-2}-\cdots- i_1 $\\ &&\qquad\qquad\qquad\times $ 3^{bk}\mathbb Z_3-i_k(2^a3^b)^{k-1}-i_{k-1}(2^a3^b)^{k-2}-\cdots-i_1 $,\end{aligned}$$ $$\begin{aligned} A_{i_{k+1},i_k,\dots,i_1}&=&[0,1)\times$2^{a(k+1)}\mathbb Z_2- i_{k+1}(2^a3^b)^{k}-i_k(2^a3^b)^{k-1}-\cdots-i_1$\\ &&\qquad\qquad\times $3^{b(k+1)}\mathbb Z_3- i_{k+1}(2^a3^b)^{k}-i_k(2^a3^b)^{k-1}-\cdots-i_1$,\end{aligned}$$ and $$\begin{aligned} 2^{a(k+1)}\mathbb Z_{2}&\subseteq& 2^{ak}\mathbb Z_{2}\\ 3^{b(k+1)}\mathbb Z_{3}&\subseteq& 3^{bk}\mathbb Z_{3}\\ i_{k+1}(2^a3^b)^{k}\mathbb Z_2&\subseteq&2^{ak}\mathbb Z_2,\\ i_{k+1}(2^a3^b)^{k}\mathbb Z_3&\subseteq&3^{bk}\mathbb Z_3,\end{aligned}$$ we have , and therefore $$\label{equation:onedirectionpositivequadrant} A_{i_k,\dots,i_1}\cap\alpha^{k+1}(A_{i_{k+1}})\supseteq A_{i_{k+1},\dots,i_1}.$$ We now claim that there is equality in . To see this, assume that there is a choice of $\ell_1,\dots,\ell_k\in\{0,\dots,2^a3^b-1\}$ with $A_{i_{k+1},\ell_k,\dots,\ell_1}\cap A_{i_{k},i_{k-1},\dots,i_1}\neq\emptyset.$ By , and noting that $A_{i_{k+1},\ell_{k},\dots,\ell_{1}}\cap A_{i_{k}, i_{k-1},\dots,i_{1}}\neq \emptyset$ implies that $A_{i_{k+1}, \ell_{k},\dots,\ell_{1}}$ is a subset of $A_{i_k, i_{k-1}, \dots,i_1}$, it follows that $$2^{a(k+1)}\mathbb Z_2-i_{k+1}(2^a3^b)^{k}-i_k(2^a3^b)^{k-1}-\cdots-i_1$$ and $$2^{a(k+1)}\mathbb Z_2-i_{k+1}(2^a3^b)^{k}-\ell_k(2^a3^b)^{k-1}-\cdots-\ell_1$$ are both subsets of $$2^{ak}\mathbb Z_2-i_k(2^a3^b)^{k-1}-\cdots-i_1,$$ and similarly for the $\mathbb Z_3$ component. Thus $$(i_k-\ell_k)(2^a3^b)^{k-1}+(i_{k-1}-\ell_{k-1})(2^a3^b)^{k-2}+\cdots+(i_1-\ell_1)\equiv0\pmod{(2^a3^b)^k}.$$ Reducing this identity modulo $2^a3^b$ shows that $i_1=\ell_1$, reducing modulo $(2^a3^b)^2$ shows that $i_2=\ell_2$, and so on. Using , it follows that there is equality in as required, proving . Now we consider an atom of the form $$A_{i_0}\cap\alpha^{-1}(A_{i_1})\cap\cdots\cap\alpha^{-n}(A_{i_n});$$ we wish to prove a statement like for these atoms, by showing that each such atom is a rectangle of the form $J\times\mathbb Z_2\times\mathbb Z_3$ for an explicitly described interval $J$ of width $\frac{1}{(2^a3^b)^{n+1}}$. A calculation shows that $$\alpha^{-1}(A_k)=\bigsqcup_{\ell=0}^{2^a3^b-1} \left[\textstyle\frac{k}{(2^a3^b)^2}+\textstyle\frac{\ell}{2^a3^b}, \frac{k+1}{(2^a3^b)^2}+\frac{\ell}{2^a3^b}\right) \times\mathbb Z_2\times\mathbb Z_3 =\bigsqcup_{\ell=0}^{2^a3^b-1}A^{k,\ell},$$ and in general we have $$\label{equation:suggestionfromyukiE3} \alpha^{-n}(A_k)=\bigsqcup_{\ell_1=0}^{2^a3^b-1}\cdots\bigsqcup_{\ell_n=0}^{2^a3^b-1} A^{k,\ell_1,\dots,\ell_n}$$ for $n{\geqslant}1$, with $$A^{k,\ell_1,\dots,\ell_n}= \left[ \textstyle\frac{k}{(2^a3^b)^{n+1}}+D(\ell_1,\dots,\ell_n), \textstyle\frac{k+1}{(2^a3^b)^{n+1}}+D(\ell_1,\dots,\ell_n)\right)\times\mathbb Z_2\times\mathbb Z_3$$ where $$D(\ell_1,\dots,\ell_n)=\frac{\ell_1}{(2^a3^b)^n}+\frac{\ell_2}{(2^a3^b)^{n-1}}+ \cdots+\frac{\ell_n}{2^a3^b}.$$ We claim that $$\label{equation:reversepositivequadrantclaim} A_{i_0}\cap\alpha^{-1}(A_{i_1})\cap\cdots\cap\alpha^{-n}(A_{i_n})= A^{i_n,i_{n-1},\dots,i_0}.$$ for $n{\geqslant}1$. For $n=1$, $$A_{i_0}\cap\alpha^{-1}(A_{i_1})\supseteq\left[\textstyle\frac{i_0}{2^a3^b},\frac{i_0+1}{2^a3^b}\right) \times\mathbb Z_2\times\mathbb Z_3\cap A^{i_1,i_0}=A^{i_1,i_0}$$ since $[\frac{i_0}{2^a3^b},\frac{i_0+1}{2^a3^b})\supseteq [\frac{i_1}{(2^a3^b)^2}+\frac{i_0}{2^a3^b},\frac{i_1+1}{(2^a3^b)^2}+\frac{i_0}{2^a3^b})$. Thus $A_{i_0}\cap\alpha^{-1}(A_{i_1})=A^{i_1,i_0}$ since the width of the real interval defining $A_{i_0}$ is $\frac{1}{2^a3^b}$ and by the real coordinates of the sets in $\alpha^{-1}(A_{i_1})$ are intervals, each of width $\frac{1}{(2^a3^b)^2}$ and with the property that the left end-points of distinct intervals are at least $\frac{1}{2^a3^b}$ apart. Now assume that holds for $n{\leqslant}k$, so that $\bigcap_{j=0}^{k+1}\alpha^{-j}(A_{i_j})$ can be written as the intersection of $$\left[ D(i_k,\dots,i_0),\textstyle\frac{1}{(2^a3^b)^{k+1}}+D(i_k,\dots,i_0) \right)\times\mathbb Z_2\times\mathbb Z_3=A^{i_{k},\dots,i_{0}}$$ with $$\bigsqcup_{0{\leqslant}j_1,\dots,j_{k+1}<2^a3^b}A^{i_{k+1},j_1,\dots,j_{k+1}}.$$ It follows that $$\bigcap_{j=0}^{k+1}\alpha^{-j}(A_{i_j})\negthinspace\supseteq\negthinspace\negthinspace \left[D(i_{k+1},\dots,i_0),\textstyle\frac{1}{(2^a3^b)^{k+2}}+ D(i_{k+1},\dots,i_0)\right)\negthinspace\times\mathbb Z_2\times\mathbb Z_3\negthinspace\negthinspace=\negthinspace A^{i_{k+1},\dots,i_{0}}.$$ Notice that the width of the real interval defining the set $A^{i_{k},\dots,i_{0}}$ is $\frac{1}{(2^a3^b)^{k+1}}$. Now by each member of the real projection of $\alpha^{-(k+1)}(A_{i_{k+1}})$ has length $\frac{1}{(2^a3^b)^{k+2}}$ and each of these intervals has the property that the left end-points of distinct intervals are at least distance $\frac{1}{(2^a3^b)^{k+1}}$ apart, showing for $n=k+1$ and hence for all $n$ by induction. By and , the atom $$\bigcap_{j=-n}^{n}\alpha^{j}(A_{i_j})= A^{i_{n},\dots,i_0}\cap A_{i_n,\dots,i_1}$$ is a single rectangle with real width $\frac{1}{(2^a3^b)^{n+1}}$, $2$-adic width $\frac{1}{(2^a)^n}$ and $3$-adic width $\frac{1}{(3^b)^n}$. It follows that $\xi$ satisfies a strong form of the condition [@MR1477538 Exercise 6.1]. Moreover, $$\operatorname{diam}$\vphantom{\sum}\right.\negmedspace\bigvee_{j=-n}^{n}\alpha^{j}(\xi)\negmedspace \left.\vphantom{\sum}$\rightarrow0$$ as $n\to\infty$, so $\xi$ is a generating Markov partition in the sense of [@MR1477538]. Away from the positive and negative quadrants $ab>0$ the behaviour of $\xi=\xi^{(a,b)}$ under the map $\alpha=\alpha^{(a,b)}$ is more complicated. In particular, as seen in Figure \[fig:multby3over2a\], an atom in $\xi\vee\alpha\xi$ need not be a rectangle even in expansive directions. However, in an expansive direction the partition $\xi$ corresponds under the map described after to the partition $\pi_0^{-1}S(H(2^a3^b))$ used by Wilson [@MR0390181 Th. 2.4]. Notice that for any $(a,b)$ in an expansive region, the group $\Sigma_{mn}$ in the notation of [@MR0390181], where $\frac{m}{n}=2^a3^b$, is $X$. Wilson shows that this partition is a Bernoulli generator, so $$\bigcap_{j=0}^{n}\alpha^{j}(A_{i_j})\neq\emptyset, \bigcap_{j=-n}^{0}\alpha^{j}(A_{i_j})\neq\emptyset\implies \bigcap_{j=-n}^{n}\alpha^{j}(A_{i_j})\neq\emptyset$$ (as in [@MR1477538 Exercise 6.1]); he also shows that an atom in $\bigvee_{j=-n}^{n}\alpha^{j}(\xi)$ lies inside a cylinder defined by small intervals in many coordinates in the description , so $$\operatorname{diam}$ \bigvee_{j=-n}^{n}\alpha^{j}(\xi) $\rightarrow0.$$ It follows that $\xi$ is a generating Markov partition for $\alpha^{(a,b)}$. There are three non-expansive directions, but only two of them contain non-trivial lattice points: Example \[example:one\] shows that $\xi^{(1,0)}$ is not generating under $\alpha^{(1,0)}$; the other direction $(0,1)$ is similar. An impression of the complexity of a generating Markov partition may be gained by comparing the dynamical zeta function of the resulting symbolic cover shift map $\sigma^{(a,b)}$ to the zeta function of the original map $\alpha^{(a,b)}$. In the positive quadrant $a>0,b>0$, where we have seen that the partition $\xi$ behaves very simply, we have $\zeta_{\sigma^{(a,b)}}(z)=\frac{1}{1-H(a,b)z}$ while $\zeta_{\alpha^{(a,b)}}(z)=\frac{1-z}{1-H(a,b)z}$ since only one pair of points of each period are identified by the factor map defined by the partition. In contrast, in the region $a<0,b>0,2^a3^b>1$ (for example) we have $\zeta_{\sigma^{(a,b)}}(z)=\frac{1}{1-3^bz}$ while $\zeta_{\alpha^{(a,b)}}(z)=\frac{1-2^az}{1-3^bz}$, reflecting the fact that more periodic points in the full $3^b$-shift are identified under the factor map. Finally, in a non-expansive direction (like $a=1,b=0$) the zeta function of $\alpha^{(a,b)}$ is not even a rational function (it is shown in [@MR2180241] that the zeta function has a natural boundary on the circle $\vert z \vert=\frac{1}{2}$ in this case; the influence on the zeta function of further directions in which an automorphism of a solenoid acts like an isometry is studied by Miles [@MR2308145] and the first author [@MR1458718]). The variation in geometrical properties of the partition $\xi^{(a,b)}$ across each expansive cone is illustrated in Figure \[fig:seven\]: a representative shape of $\alpha^{(a,b)}(A_0)$ is shown shaded in each expansive cone. The transitions across the axes are clear; at the line $2^x3^y=1$ all that changes is the sign of the real Lyapunov exponent. [10]{} L. M. Abramov, The entropy of an automorphism of a solenoidal group, Theor. Probability Appl. 4 (1959), 231–236. [MR ]{}[MR0123346 (23 \#A674)]{} R. L. Adler, Symbolic dynamics and [M]{}arkov partitions, Bull. Amer. Math. Soc. (N.S.) 35 (1998), no. 1, 1–56. [MR ]{}[MR1477538 (98j:58081)]{} R. L. Adler and B. Weiss, Similarity of automorphisms of the torus, Memoirs of the American Mathematical Society, No. 98, American Mathematical Society, Providence, R.I., 1970. [MR ]{}[MR0257315 (41 \#1966)]{} R. Bowen, On [A]{}xiom [A]{} diffeomorphisms, American Mathematical Society, Providence, R.I., 1978, Regional Conference Series in Mathematics, No. 35. [MR ]{}[MR0482842 (58 \#2888)]{} M. Boyle and D. Lind, Expansive subdynamics, Trans. Amer. Math. Soc. 349 (1997), no. 1, 55–102. [MR ]{}[MR1355295 (97d:58115)]{} V. Chothi, G. Everest, and T. Ward, [$S$]{}-integer dynamical systems: periodic points, J. Reine Angew. Math. 489 (1997), 99–132. [MR ]{}[MR1461206 (99b:11089)]{} M. Einsiedler and D. Lind, Algebraic [$\mathbb Z\sp d$]{}-actions of entropy rank one, Trans. Amer. Math. Soc. 356 (2004), no. 5, 1799–1831 (electronic). [MR ]{}[MR2031042 (2005a:37009)]{} M. Einsiedler, D. Lind, R. Miles, and T. Ward, Expansive subdynamics for algebraic [${\mathbb Z}\sp d$]{}-actions, Ergodic Theory Dynam. Systems 21 (2001), no. 6, 1695–1729. [MR ]{}[MR1869066 (2004c:37033)]{} G. Everest, R. Miles, S. Stevens, and T. Ward, Orbit-counting in non-hyperbolic dynamical systems, J. Reine Angew. Math. 608 (2007), 155–182. [MR ]{}[MR2339472]{} G. Everest, V. Stangoe, and T. Ward, Orbit counting with an isometric direction, Algebraic and topological dynamics, Contemp. Math., vol. 385, Amer. Math. Soc., Providence, RI, 2005, pp. 293–302. [MR ]{}[MR2180241 (2006k:37046)]{} Edwin Hewitt and Kenneth A. Ross, Abstract harmonic analysis. [V]{}ol. [I]{}, second ed., Grundlehren der Mathematischen Wissenschaften \[Fundamental Principles of Mathematical Sciences\], vol. 115, Springer-Verlag, Berlin, 1979, Structure of topological groups, integration theory, group representations. [MR ]{}[MR551496 (81k:43001)]{} D. A. Lind and T. Ward, Automorphisms of solenoids and [$p$]{}-adic entropy, Ergodic Theory Dynam. Systems 8 (1988), no. 3, 411–419. [MR ]{}[MR961739 (90a:28031)]{} A. Manning, A [M]{}arkov partition that reflects the geometry of a hyperbolic toral automorphism, Trans. Amer. Math. Soc. 354 (2002), no. 7, 2849–2863 (electronic). [MR ]{}[MR1895206 (2003g:37048)]{} R. Miles, Zeta functions for elements of entropy rank-one actions, Ergodic Theory Dynam. Systems 27 (2007), no. 2, 567–582. [MR ]{}[MR2308145]{} R. Miles and T. Ward, Periodic point data detects subdynamics in entropy rank one, Ergodic Theory Dynam. Systems 26 (2006), no. 6, 1913–1930. [MR ]{}[MR2279271]{} F. Takens, Multiplications in solenoids as hyperbolic attractors, Topology Appl. 152 (2005), no. 3, 219–225. [MR ]{}[MR2164385 (2006e:37044)]{} T. Ward, An uncountable family of group automorphisms, and a typical member, Bull. London Math. Soc. 29 (1997), no. 5, 577–584. [MR ]{}[MR1458718 (98k:22028)]{} [to3em]{}, Additive cellular automata and volume growth, Entropy 2 (2000), no. 3, 142–167 (electronic). [MR ]{}[MR1882488 (2004h:37014)]{} A. Weil, Basic number theory, Die Grundlehren der mathematischen Wissenschaften, Band 144, Springer-Verlag New York, Inc., New York, 1967. [MR ]{}[MR0234930 (38 \#3244)]{} R. F. Williams, Expanding attractors, Inst. Hautes Études Sci. Publ. Math. (1974), no. 43, 169–203. [MR ]{}[MR0348794 (50 \#1289)]{} A. M. Wilson, On endomorphisms of a solenoid, Proc. Amer. Math. Soc. 55 (1976), no. 1, 69–74. [MR ]{}[MR0390181 (52 \#11007)]{} [^1]: This research was supported by EPSRC grant EP/C015754/1. The authors thank Richard Miles for useful comments.	{ "pile_set_name": "ArXiv" }
--- abstract: \| We have started high precision photometric monitoring observations at the AIU Jena observatory in Gro[ß]{}schwabhausen near Jena in fall 2006. We used a 25cm Cassegrain telescope equipped with a CCD-camera mounted picky-pack on a 90cm telescope. To test the obtainable photometric precision, we observed stars with known transiting planets. We could recover all planetary transits observed by us.\ We observed the parent star of the transiting planet TrES-2 over a longer period in Gro[ß]{}schwabhausen. Between March and November 2007 seven different transits and almost a complete orbital period were analyzed. Overall, in 31 nights of observation 3423 exposures (in total 57.05h of observation) of the TrES-2 parent star were taken. Here, we present our methods and the resulting light curves. Using our observations we could improve the orbital parameters of the system. --- Introduction ============ The search for extrasolar planets is one of the most important research field in Astronomy today. Up to now the most successful method to detect exoplanet candidates is the radial velocity technique. But in the last ten years another indirect detection method has established itself as a highly successful technique in finding planets and confirming candidates - the transit method.\ The transit event is a strictly periodic phenomenon. In a systen where a known planet transits its host star, a second planet in that system will cause the time between transits to vary. It is becoming increasingly popular because even small ground-based observatories have already obtained the photometric precision necessary to detect sub-Earth mass planets by the transit timing variation method [@Steffen_etal07 (Steffen et al. 2007)].\ We have started high precision photometric monitoring observations at the AIU Jena observatory in Gro[ß]{}schwabhausen near Jena in fall 2006. In this work we use the transit method to observe the known transiting planet TrES-2. The aim is to test procedures with our telescope and determine the obtainable photometric precision with the currently existing camera. We paid special attention to the accurate determination of transit times in order to identify precise transit timing variations that would be indicative of perturbations from additional bodies and to refine the orbital parameters of the systems. Instruments and Observations ============================ We have three telescopes available in our observatory in Gro[ß]{}schwabhausen, a 90cm reflector, a 20cm refractor with a focal length of 3m and a 25cm Cassegrain telescope with a focal ratio f/D=9. The 90cm reflector telescope made by Zeiss Jena can be used in two modes -– either as Schmidt camera (diameter of the correction plate D=60cm, f/D=3) or as Nasmyth telescope with D=90cm of free opening and f/D=15.\ Because new motors for the movement of the telescope were installed, we currently test procedures with the 25cm Cassegrain telescope with the CCD-camera CTK (assegrain-elescop CCD-amera)([@Mugrauer_etal09 Mugrauer et al. (2009)], in preperation). In the course of the year 2006 we started our continuous observations.\ For our TrES-2 observations, started in March 2007, we used 34 nights from March to November 2007. Due to the weather conditions 3 nights were not analyzed. All TrES-2 observations were taken in I-band with 60 s exposure time. The mean photometric accuracy of the V=11.4mag bright TrES-2 host star is 0.007 mag. Data Reduction and analysis =========================== We calibrate the images of our target fields using the standard IRAF[^1] procedures darkcombine, flatcombine and ccdproc.\ After calibrating all images, we perform aperture photometry. Therefore we use the IRAF task chphot which is written by Christopher Broeg and based on the standard IRAF routine phot. With chphot it is possible to do the photometry on every star in the field at the same time. We found 1294 stars in our 37.7’$\times$ 37.7’field of view around TrES-2. We used an aperture of radius 5 pixels (11.03”) and an annulus for sky subtraction ranging in radius from 15 to 20 pixels, centered on each star.\ A problem of the differential photometry is the search for a good comparison star. [@Broeg_etal05 Broeg et al. (2005)] developed an algorithm which uses as many stars as possible (all available field stars) and calculate an artificial comparison star (cs). The algorithm decides which stars are the best and takes the weighted average of them. Then it computes the artificial cs with the best possible signal-to-noise ratio (S/N) by automatically weighting down the stars according to their variability.\ To get the best possible result for the transit light curve, we try to use only the best cs in the field. Therefore we reject all stars that could not be measured on every image, faint stars with low S/N and variable stars which could introduce disturbing signals to the data. With the remaining objects we calculated the artificial cs. Finally this artificial cs is compared to TrES-2 to get the differential magnitudes for every image.\ Then we correct for systematic effects by using the Sys-Rem detrending algorithm which is proposed by [@Tamuz_etal05 Tamuz et al. (2005)] and implemented by Johannes Koppenhoefer. Figure \[TrES\_Sysrem\] shows the same transit event of TrES-2 before and after using Sys-Rem. ![The same transit of TrES-2 observed on March 13 2007 before and after using Sys-Rem. The time series includes 148 I-band 60s exposures between 0.17 AM and 3.56 AM (UT)[]{data-label="TrES_Sysrem"}](TrES2_ohne_mit_sysrem.eps){width="100.00000%"} \ To determine the time of the centre of the transit of TrES-2 we approximate a theoretical light curve on the observed light curve (see the right panel in Figure \[TrES\_Sysrem\]). To get the best fit we compare the theoretical light curve with the observed light curve until the $\chi^{2}$ is minimal. Transit timing residuals ======================== In addition to the observation of TrES-2 from Gro[ß]{}schwabhausen we found four transit times in the literature. These altogether 11 transits are summarized in the Table \[transit\_times\]. Observer Date Time of Midtransit \[HJD\] error \[d\] --------------------- ------------ ---------------------------- ------------- TrES-Network$^{1}$ 10.08.2006 2453957.63580 0.00100 TLC-Project$^{2}$ 11.09.2006 2453989.75286 0.00029 16.09.2006 2453994.69393 0.00031 02.11.2006 2454041.63579 0.00030 Gsh 13.03.2007 2454172.57793 0.00170 03.05.2007 2454224.46077 0.00290 17.07.2007 2454298.57589 0.00240 26.07.2007 2454308.46621 0.00210 16.09.2007 2454360.34750 0.00106 21.09.2007 2454365.27832 0.00180 14.10.2007 2454387.52294 0.00190 : Summary of all known transit times of TrES-2.[]{data-label="transit_times"} \ $^{1}$from [@ODonovan_etal06 O’Donovan et al. (2006)]\ $^{2}$from [@Holman_etal07 Holman et al. (2005)] \ We used the ephemeris of [@Holman_etal07 Holman et al. (2005)] $T_{\mathrm{c}}(E)=(2453957.63479\,+\,E\cdot 2.470621)\,\mathrm{d}$ to compute ”observed minus calculated” (O-C) residuals for all 11 transit times. Figure \[O\_C\_TrES2\] shows the differences between the observed and predicted times of midtransit, as a function of epoch. The dashed line represents the ephemeris given by [@Holman_etal07 Holman et al. (2005)]. ![a.) Transit timing residuals for TrES-2. The dashed line shows the ephemeris given by [@Holman_etal07 Holman et al. (2005)]. The best-fitting line (solid line) is plotted, representing the updated ephemeris given in equation \[Elemente:TrES2\]. — b.) More than 3000 individual observations of TrES-2 from March to October 2007 calculated in one phase according to the updated ephemeris.[]{data-label="O_C_TrES2"}](TrES2_B_R_und_Phase.eps){width="100.00000%"} We found a negative trend in this (O-C)-diagram. Thus, we refine the ephemeris. The resulting ephemeris which represent our measurements best is $$\label{Elemente:TrES2} T_{\mathrm{c}} = T_{0}+P\cdot E$$ with: $$\begin{aligned} T_{0} &=& (2453989.75286 \pm 0.00029)\,\mathrm{d} \nonumber\\ P &=& (2.470615 \pm 0.00002)\,\mathrm{d} \nonumber\end{aligned}$$ Discussion and Outlook ====================== During the observations of the transiting extrasolar planet TrES-2 at our university observatory in Gro[ß]{}schwabhausen with the 25cm Cassegrain telescope equipped with the optical CCD camera CTK we obtained a timing accuracy of $\sim$2min. The timing residuals are not consistent with zero within the measurement errors. The second last data point is 3$\sigma$ deviant. The deviations to both sides of zero could be a first indication of timing anomalies caused by additional planets or moons. We will continue observing TrES-2 to confirm these transit time variations. Therefore, we work on methods to improve the accuracy of our transit times.\ This year we get a new CCD camera for the Schmidt focus of the 90cm reflector. This camera will have a smaller pixel scale and a higher sensitivity. Part of the preparation for the new camera is the improvement of the software for relative photometry. Our transit observations will benefit strongly from the new camera.\ The transit observations with the AIU Jena telescope in Gro[ß]{}schwabhausen provide anchors for future searches for transit time variations. 2005, AN, 326, 134 , 2007 ApJ, 664, 1185 2009, AN, in preperation 2006, ApJ (Letters),651, L61 2007, astro-ph, 0704.0632v1 2005, MNRAS, 356, 1466 [^1]: IRAF is distributed by the National Optical Astronomy Observatories, which are operated by the Association of Universities for Research in Astronomy, Inc., under cooperative agreement with the National Science Foundation.	{ "pile_set_name": "ArXiv" }
--- author: - 'E. de la Fuente, M. A. Trinidad, A. Porras, C. Rodríguez–Rico, E. D. Araya, S. Kurtz, P. Hofner, and A. Nigoche–Netro' title: 'Unveiling the Hot Molecular Core in the Ultracompact Region with Extended Emission G12.21–0.10' --- Introduction {#intro} ============ Hot molecular cores (HMCs) are considered tobe the birthplace of high–mass stars [e.g., @K00; @GL99]. Thus, to study their physical properties, dynamics, and evolution is crucial for advancing our understanding of high–mass star formation. HMCs are detected mainly by emission of high–excitation molecular lines (e.g., CH$_3$CN and NH$_{\rm 3}$) and by millimeter continuum emission from warm dust. A detailed overview of star formation is presented by @SP05, while a review of high–mass star formation was published by @ZY07. Water maser emission is a common phenomenon in the neighborhood of newly-formed high–mass stars [e.g., @C04; @T03; @HC96]. However, the relationship of the H$_2$O masers to the star formation environment is still not fully understood. It is accepted that their presence implies the existence of at least one embedded YSO [e.g., @GL99]. For ultra compact regions () with cometary morphologies, @HC96 showed that the water masers almost always located in clusters, near, yet offset, from the cometary arc; and for other morphologies the masers are often projected onto the ionized gas. @HC96 also found that water masers are related to hot NH$_{\rm 3}$ clumps (and YSOs) rather than to the ionized gas of the regions. The latter was also observed by @C94 in a sample of 4 molecular cores. G12.21–0.10 (G12 hereafter), located at a distance of 13.5 kpc [@C90], has been classified as a cometary region with extended emission [@F09a; @F07; @KK01]. C$^{17}$O observations finding a molecular clump coincident with the UC emission were reported by @Ho00. This region also presents emission of hot and dense chemically enriched gas tracers (see Table \[previous\]), and has been also observed in infrared (IR) surveys [@F09b; @F07; @D03]. Hence, it was deemed a prime candidate for a HMC by @D03. In G12, H$_2$O and CH$_{\rm 3}$OH maser emission is observed toward the molecular clump, no towards the ionized gas [@K04]. In particular, the water masers are offset by at $\sim$4${\ifmmode{^{\prime \prime}}\else{$^{\prime \prime}$}\fi}$ (or 0.26 pc at 13.5 kpc) from the region, suggesting the location of a deeply embedded, young high–mass star [@HC96]. In addition, @K04 found that a strong water maser feature in this cluster coincides spatially with both class I (44 GHz) and class II (6.67 GHz) CH$_{\rm 3}$OH masers. Thus, G12 is an interesting candidate for studying the nature and formation of water and methanol masers and for understanding their relation to molecular gas and high–mass star formation. In § \[obs\] we describe the radio continuum and spectral line observations and data reduction. In § \[results\] we present the observational results and analysis as follows: 1.- we show that the $^{13}$CS (1$-$0 and 2$-$1) transitions trace the hot and optically thin dust clump in G12; 2.- we calculate physical characteristics of this clump using NH$_{\rm 3}$ data confirm that it is a hot molecular core. In § \[disc\] we discuss the $^{13}$CS and NH$_{\rm 3}$ data in the context of the HMC interpretation. A summary is presented in § \[summary\]. Observations and Data Reduction {#obs} =============================== Owen’s Valley Radio Observatory (OVRO) Observations {#ovro} --------------------------------------------------- The $^{13}$CS(2$-$1) and H41$\alpha$ line observations were made with OVRO during March and April of 1996. The equatorial and high resolution configurations were used, providing baselines from 30 to 240 m. Cryogenically cooled SIS receivers provided typical SSB system temperatures of 220–350 K. The digital correlator was split into two bands to observe the $^{13}$CS(2$-$1) ($\nu_0 = 92.49430$ GHz) and H41$\alpha$ ($\nu_0 = 92.03445$ GHz) lines with 30 MHz ($\sim$100[ km s$^{-1}$]{}) and 60 MHz ($\sim$200[ km s$^{-1}$]{}) bandwidths, and 62 and 60 Hanning–smoothed channels, respectively. This configuration gives a spectral resolution of 0.5 MHz (1.6[ km s$^{-1}$]{}) for the $^{13}$CS(2$-$1) line and 1 MHz (3.3[ km s$^{-1}$]{}) for the H41$\alpha$ line. In addition, simultaneous continuum observations at 3 mm were made using a 1 GHz bandwidth analog correlator. The quasars NRAO 530, 3C 454.3, B1908$-$202, and B1749+096 were observed to track amplitude and phase variations. The bandpass calibration and flux scale were established by observing 3C 273. The absolute flux uncertainty is estimated to be $\sim$20 %, and the initial calibration was carried out using the Caltech MMA data reduction package [@S93]. Further data reduction and mapping were done with the AIPS software package. The continuum obtained from the line–free channels was subtracted from the $^{13}$CS and H41$\alpha$ emission lines in the $(u,v)$ plane using the program UVLIN in AIPS. Finally, Gaussian fits to the observed spectra were made with the CLASS software package. Karl G. Jansky Very Large Array (VLA) Observations {#vla} -------------------------------------------------- ### Radio–Continuum Emission {#RCE} High angular resolution continuum observations at 0.7, 2, and 3.6 cm toward G12 were made with the VLA[^1] in its CnB configuration on 1996 January 29. Two 50 MHz bands were observed, each one including both right and left circular polarizations. We used a subarray of 12 antennas for 7 mm observations and the remaining 15 antennas were employed for 2 and 3.6 cm observations. The flux and phase calibrators were 3C286 (1.45, 3.40 and 5.23 Jy at 0.7, 2, and 3.6 cm) and B1923+210 (1.0, 1.8 and 1.0 Jy; flux densities meassured at 0.7, 2, and 3.6 cm respectively). In addition, low angular resolution VLA observations at 3.6 cm in the D configuration toward G12 were made on 2005 April 03. As before two 50 MHz bands were observed, each one including both right and left circular polarizations. The flux and phase calibrators were 3C 286 (5.23 Jy) and J1832$-$105 (1.28 Jy), respectively. The calibration and imaging for both high and low angular resolution observations were made following standard procedures in AIPS. Self–calibration in phase only, and Multi–Resolution Cleaning were performed for the low angular resolution data. ### NH$_{\rm 3}$ and $^{13}$CS(1$-$0) Emission {#nh3cs10} Observations of NH$_{\rm 3}$ were made with the VLA in its DnC configuration on 2004 June 06. Both transitions (2,2) and (4,4), with rest frequencies of 23722.631 and 24139.417 MHz, respectively, were observed. In order to include the main line and one pair of hyperfine satellites, the spectral line correlator was split into two overlapping bands of 64 channels each, centered at 2.30 and 0.76 MHz higher and lower than the expected line frequency for (2,2) and 1.06 and 1.24 MHz for (4,4). The Doppler tracking center velocity was set to 24[ km s$^{-1}$]{} (V$_{\rm LSR}$) and the total velocity coverage was 94[ km s$^{-1}$]{}, obtaining a frequency resolution of 97.656 kHz (1.25[ km s$^{-1}$]{}). The flux, phase and bandpass calibrators were 3C 286 (2.6 Jy), J1820$-$254 (0.53 Jy) and J1924$-$292 (9.1 Jy), respectively. Calibration and data reduction were done following standard line procedures of the AIPS software package. All images were obtained using the task IMAGR with ROBUST = 0, while the spectra were obtained with task ISPEC after combining the two overlapping bands into a single 71 channel cube and the cube moment maps were made using the task XMOM. Gaussian fits to the observed spectra were made using the CLASS software package. During the same run, $^{\rm 13}$CS(1$-$0) observations at the rest frequency of 46.24757 GHz, were also carried out. A central bandpass velocity of 24[ km s$^{-1}$]{} was also used. Flux, phase and bandpass calibrators were the same as for NH$_{\rm 3}$, as well as the calibration and data reduction. A natural weighting was used to generate the images. Observational Results and Analysis {#results} ================================== Radio Continuum Emission {#resultsRCE} ------------------------ The 21 cm NVSS contour map [@Co98] superposed on the 3.6 cm low angular resolution VLA–D map in gray–scale toward G12 is shown in Figure \[fig1\]a [Left]{}. The morphology of both emissions match at different scales in this region, and are in agreement with 21 cm observations from @KK01 [their Figure 1g]. Six continuum peaks are detected at 3.6 cm, one of which corresponds to the UC source. The 3 mm continuum emission toward the UC source is split into two peaks (see Figure \[fig1\]a [Right]{}). They are separated $\sim$4$''$ (0.26 pc at a distance of 13.5 kpc), one of them corresponds to the region (labeled as 1), while the other (labeled as 2) coincides with a water maser group reported by @HC96. Due to the possible relation between these masers and the molecular clump reported by @Ha00, we will refer to peak–2 as the molecular clump. The peak positions of the 3 mm continuum emission from the region and the molecular clump are reported in Table \[pos\]. High angular resolution continuum observations at 0.7, 2, and 3.6 cm are shown in Figure \[fig1\]b. They show that the radio continuum emission is only associated with the region. Then, the observed emission at 3 mm and the lack of emission at 0.7, 2 and 3.6 cm toward the 3 mm peak–2, suggests that the 3 mm continuum emission toward the molecular clump could arise from dust rather than ionized gas. Integrated continuum flux densities of the ionized gas, rms noise, beam, and deconvolved sizes for these wavelengths are given in Table \[vla\_cont\]. Line Emission {#le} ------------- ### H41$\alpha$ {#h41a} The H41$\alpha$ ($\sim$3.3 mm) emission is only detected toward the region and no emission is detected toward the molecular clump. The H41$\alpha$ emission has a similar distribution as the radio continuum observed at 7 mm shown in Figure \[fig1\]b. The H41$\alpha$ spectrum toward G12 is shown in Figure \[fig2\] (upper panel). Line emission was detected in the velocity range of $\sim$10 to 45[ km s$^{-1}$]{}, with a peak at about 30[ km s$^{-1}$]{}. The results of the Gaussian fit of this recombination line are presented in Table \[h41cspar\]. The H41$\alpha$ velocity is similar to that of the molecular emission from $^{13}$CO(1$-$0), C$^{34}$S(2$-$1), and C$^{34}$S(3$-$2) of 20$-$25[ km s$^{-1}$]{}, both from @KK03 and this work (see below). In addition, the H41$\alpha$ LSR central velocity is in agreement with that reported by @KK01 using H76$\alpha$ for three positions in the extended emission ($\sim$27[ km s$^{-1}$]{} on average, see their Figure 1g). The ionized gas in G12, present at different scales (UC and extended emission; see Figure \[fig1\]), coexists with the molecular gas and originates in the same star forming region. ### $^{\rm 13}$CS(2$-$1) {#13cs21} Figure \[fig3\] shows the channel maps of the $^{13}$CS(2$-$1) line emission superposed on the 3.6 cm continuum emission. The deconvolved size of this molecular region is $\sim 4{\ifmmode{{\rlap.}^{\prime \prime}}\else{${\rlap.}^{\prime \prime}$}\fi}5 \times 2{\ifmmode{{\rlap.}^{\prime \prime}}\else{${\rlap.}^{\prime \prime}$}\fi}0$ (0.36 $\times$ 0.16 pc) with the major axis oriented approximately in the E$-$W direction, while the $^{13}$CS(2$-$1) peak emission observed at V$_{\rm LSR}$ = 24[ km s$^{-1}$]{} is blushifted ($\sim$5[ km s$^{-1}$]{}) with respect to the H41$\alpha$. The $^{13}$CS(2$-$1) line spectrum is shown in Figure \[fig2\] (lower panel). Observed parameters (flux densities, rms noise, beam, and deconvolved sizes) for $^{\rm 13}$CS(2$-$1) toward G12 are presented in Table \[h41cspar\]. Comparing the $^{13}$CS(2$-$1) line and the 3.6 cm (VLA–CnB) continuum emission, we note that the $^{13}$CS(2$-$1) emission is not spatially associated with the UC region in G12, but rather is associated is with the molecular clump and the group of water masers. This result suggests the presence of a molecular clump spatially associated with the 3 mm continuum emission (peak–2) not far from the region (see Table \[pos\]). Figure \[fig2\] (lower panel), show that in addition to the $^{13}$CS(2$-$1) line at V$_{\rm LSR}$ = 24[ km s$^{-1}$]{}, a weaker component at 92.49284 GHz is marginally detected at V$_{\rm LSR}$ = 4.5$\pm$0.5[ km s$^{-1}$]{}assuming the $^{13}$CS(2$-$1) rest frequency; and ${\rm \Delta}$V = 3.6$\pm$1.1[ km s$^{-1}$]{}. A possible candidate to explain this weak emission is the molecule C$_{\rm 3}$S [$\nu_0$ = 92.48849 GHz; @Mo03] at V$_{\rm LSR}$ = 14.1$\pm$0.5[ km s$^{-1}$]{}. Although this molecular line seems to be real, given the low signal–to–noise ($\sim$2.5$\sigma$ at the peak), we cannot determine the physical parameters of the molecular gas from which this emission arises. ### $^{\rm 13}$CS(1$-$0) {#13cs10} Figure \[fig4\] shows the channel maps of the $^{13}$CS(1$-$0) emission towards G12 superposed to the $^{13}$CS(2$-$1) integrated map as comparison. Although $^{13}$CS(1$-$0) and $^{13}$CS(2$-$1) present different morphologies, both are spatially associated with the molecular clump (3 mm peak–2). This morphological difference could in part be due to the different spatial resolution achieved for each molecular line. However, the $^{13}$CS(1$-$0) emission presents a complex morphology as revealed by each channel map (see also the integrated emission shown in the inset of the 27.2[ km s$^{-1}$]{} panel). This morphology shows a central double–component roughly in the East–West direction (labeled as E and W on the inset) with peak line emission observed at V$_{\rm LSR}$ = 20[ km s$^{-1}$]{} and 23[ km s$^{-1}$]{}, respectively. Another weak emission peak to the North (labeled as N) is also observed at V$_{\rm LSR}$ = 26[ km s$^{-1}$]{}. ### NH$_{\rm 3}$ {#nh3} The NH$_{\rm 3}$ emission is only detected toward the molecular clump (3 mm peak–2). The integrated NH$_{\rm 3}$ spectra of the (2,2) and (4,4) transitions are shown in Figure \[fig5\]. In the (2,2) transition the satellite lines are marginally resolved, and in (4,4) the satellite lines are blended into a single component. The two electric hyperfine components were fit with two Gaussians and the results of the fits are given in Table \[nh3obs\]. The (4,4) main peak line flux density is weaker than that measured in the (2,2) line. The observed widths of 8–9[ km s$^{-1}$]{} are evidence of highly supersonic turbulent motions in the molecular gas [@GL99]. The observed line parameters, flux densities, rms noise, beam, and deconvolved sizes for both transitions (2,2) and (4,4) are presented in Table \[nh3obs\]. Within the uncertainties, the (2,2) and (4,4) V$_{\rm LSR}$ and $\Delta {\rm V_{obs}}$ are in mutual agreement. The measured LSR velocities of both transitions are slightly smaller ($\sim$1[ km s$^{-1}$]{}) than those reported by @C90 and @C92 for (2,2) and (4,4) using the Effelsberg 100 m telescope, while the $\Delta {\rm V_{obs}}$ for (2,2) and (4,4) are $\sim$1[ km s$^{-1}$]{} and $\sim$4[ km s$^{-1}$]{} greater, respectively. The line width difference indicates motion of the hotter compact (4,4) gas relative to the cooler extended (2,2) gas as discussed in @C92. Neither the (2,2) nor the (4,4) spectra show detectable anomalies in the satellite hyperfine components like those observed toward S106 in the NH$_{\rm 3}$ (1,1) transition [@S82]. Hence we can assume that the populations in the metastable levels are in LTE. To calculate physical parameters of the region based on the NH$_{\rm 3}$ spectra, we also assume that true intrinsic widths, $\Delta {\rm V_{t}}$, are equal to the observed values, $\Delta {\rm V_{obs}}$. Optical depths for the main lines, ${\rm \tau_{main}}$, of 3.5$\pm$0.5 and 8.0$\pm$0.5 for (2,2) and (4,4) transitions, respectively, were calculated numerically [@HT83 equation 1]. The total optical depth can be calculated as ${\rm \tau_{tot}}$ = (1/b)${\rm \tau_{main}}$ where b represents the normalized (to total intensity) relative intensities of the main line. Parameter b values of 0.8 and 0.93 for (2,2) and (4,4) respectively are assumed [@H77], obtaining $\tau_{\rm tot}^{\rm 2,2}$ = 4.4$\pm$0.5 and $\tau_{\rm tot}^{\rm 4,4}$ = 8.6$\pm$0.5. In addition, excitation temperatures (T$_{\rm exc}$) of 78$\pm$5 K and 35$\pm$5 K for (2,2) and (4,4) were calculated in a standard way by solving the transfer equation [@H77; @HT83]. The column densities for (2,2) and (4,4) transitions in terms of ${\rm \tau_{\rm main}}$, T$_{\rm exc}$ and $\Delta {\rm V_{t}}$ were computed through the ${\rm \tau_{tot}}$ definition given by @U86 using the respective Einstein coefficient for spontaneous de–excitation: A$_{\pm}$(2,2) = 2.23$\times$10$^{-7}$ s$^{-1}$ and A$_{\pm}$(4,4) = 2.82$\times$10$^{-7}$ s$^{-1}$. Hence N(2,2) = 2.6$\pm$1.0$\times$10$^{16}$ cm$^{-2}$ and N(4,4) = 2.0$\pm$1.0$\times$10$^{16}$ cm$^{-2}$. The rotational temperature, T$_{\rm{42}}$ = 77$\pm$10 K, was computed in terms of the (2,2) and (4,4) peak flux densities and velocity widths (see Table \[nh3obs\]) considering the same solid angle for both emissions. Using this rotational temperature, a rotation level diagram for para–NH$_{\rm 3}$ [@WU83; @PK75] considering collisional transitions between K–ladders and transitions which follow the parity transition rule, and doing the calculations in statistical equilibrium derived by @WU83 with the numerical computations of @D88 in detailed balance, we obtain a value for the kinetic temperature of ${\rm {T_k}}$ = 86$\pm$12 K. Discussion {#disc} ========== The Hot Molecular Core {#g12hmc} ---------------------- The kinetic temperature (86 K), volumetric density (1.5$\times$10$^6$ cm$^{-3}$) and linear size ($\sim$0.22 pc, assuming a distance of 13.5 kpc) generally coincide with the operational definition of HMC; hot ($\gtrsim$ 100 K), dense ($\sim$10$^5$ to 10$^8$ cm$^{\rm -3}$), and small (size $\lesssim$ 0.1 pc) molecular gas structures [@C92; @GL99]. So, based on these characteristics, we suggest that the molecular gas clump detected in G12 is a Hot Molecular Core (G12–HMC hereafter). Several maps of the molecular line emission from the G12–HMC region are presented in Figure \[fig6\]. No radio continuum source was detected toward the G12–HMC position (Figure \[fig6\]a) where the group of H$_2$O masers is located. This could be explained following the suggestion of @W03 that 6.67 GHz CH$_{\rm 3}$OH masers appear before an region is created. However, higher sensitivity observations are needed, since @R16 have a 100% detection rate in the radio continuum toward 25 HMCs, when observed with much higher sensitivity of $\sim$3–10 $\mu$Jy beam$^{-1}$ rms. The NH$_{\rm 3}$(2,2) and (4,4) emission clearly coincides with the molecular clump observed at 3 mm (see Figure \[fig1\]a), as well as the group of masers. The distribution of molecular gas from both transitions, NH$_{\rm 3}$(2,2) and (4,4), is similar, although the NH$_{\rm 3}$(4,4) emission is slightly more compact than the NH$_{\rm 3}$(2,2) emission, suggesting that the hotter and denser gas is confined in a smaller–scale structure. Note that Figures \[fig6\]c and \[fig6\]d show the more extended and more compact emission, respectively. The first moment map of the (2,2) transition shows a velocity gradient from E to N (following the nomenclature of the inset in Figure \[fig4\]) from 21 to 28[ km s$^{-1}$]{} over $\sim$4${\ifmmode{^{\prime \prime}}\else{$^{\prime \prime}$}\fi}$ ($\sim$30 km s$^{-1}$ pc$^{-1}$). A similar behavior is observed for the (4,4) moment–1 map [@Fu10], but from 21 to 27[ km s$^{-1}$]{} over $\sim$3${\ifmmode{^{\prime \prime}}\else{$^{\prime \prime}$}\fi}$ corresponding to 25 km s$^{-1}$ pc$^{-1}$. For $^{13}$CS(1$-$0), from $\sim$20 to $\sim$26[ km s$^{-1}$]{}, we obtain a velocity gradient of $\sim$3[ km s$^{-1}$]{}arcsec$^{-1}$ (or 45[ km s$^{-1}$]{}pc$^{-1}$), and for the $^{13}$CS(2$-$1) emission $\sim$28[ km s$^{-1}$]{}pc$^{-1}$, which is similar to that of the NH$_{\rm 3}$(2,2) map (compare the grey–scale bars in Figures \[fig6\]a and \[fig6\]c). The $^{\rm 13}$CS and NH$_{\rm 3}$ emission coincides in position and kinematics, which suggests no chemical differentiation in the HMC. In addition to the velocity gradient from E to N for $^{13}$CS(1$-$0), a velocity gradient from W to N of $\sim$1.5[ km s$^{-1}$]{} arcsec$^{-1}$ (or 23[ km s$^{-1}$]{}pc$^{-1}$) is also estimated. A comparison of these results with those reported in Orion–IRc2 HMCs [@CW97] suggests that $^{13}$CS(1$-$0) is a good candidate line to trace the internal structure of HMCs as observed in Figure \[fig6\]d. Assuming that the 3 mm flux density is due to warm dust, we derive a clump mass M$_{clump}$ = $g S_{\nu} d^2$ / $\kappa_{\nu} B_{\nu}(T_d)$ [e.g., @B06] for the G12–HMC of $\sim$ 2.0$\times$10$^{3}$ M$_{\sun}$, where a gas–to–dust ratio of g $=$ 100 and a dust mass opacity coefficient of $\kappa_{112}~=~0.2~cm^2~g^{-1}$ [@OH94] were adopted. A flux density of 0.15 Jy (Table \[vla\_cont\]), a distance of 13.5 kpc, and a temperature T$_d$ = 86 K (Table \[hotcore\]). In addition, considering a 3 mm deconvolved angular size of 1.7$\pm$0.5$''$, and assuming that all hydrogen is in molecular form and the clump is spherical, we determine a n(H$_{\rm 2}$)$_{\rm 3mm}$ $\sim$ 6.3$\times$10$^6$ cm$^{-3}$, and thus a central hydrogen column density of N(H$_{\rm 2}$) $\sim$ 4.2$\times$10$^{24}$ cm$^{-2}$ [following e.g., @S-M17]. Both values of n(H$_{\rm 2}$) and N(H$_{\rm 2}$) are in agreement with those reported in Table \[previous\]. The gas mass can be estimated from the $^{13}$CS(2$-$1) integrated line flux density assuming LTE, optically thin emission, and a fractional abundance for the $^{13}$CS molecule. Then, assuming the $^{12}$CS abundance is CS/H$_2 =10^{-8}$, as estimated for the Orion Hot Core [@VB98], $^{12}$CS/$^{13}$CS ratio of 40 [@T73], and excitation temperature $\simeq$ 86 K from NH$_{\rm 3}$ data, we obtain a gas mass of 6.3$\times$10$^3$ M$_{\sun}$. Then, the mass ratio for G12–HMC is M($^{\rm 13}$CS)/M(H$_{\rm 2}$) $\sim$ 3. This discrepancy in the estimated mass can be explained in terms of the uncertainties in the adopted abundances, on the hot dust emission law and the gas to dust ratio. Under the consideration of uniform density, a $^{\rm 13}$CS(2$-$1) virial mass of $\sim$2200 M$_{\sun}$ is derived. This virial mass is in a better agreement with the NH$_{\rm 3}$ virial mass ($\sim$1500 M$_{\sun}$ and $\sim$1800 M$_{\sun}$) in comparison with that determined by @H98 via JCMT observations of CH$_{\rm 3}$OH (1300 M$_{\sun}$). The volumetric density of H$_{\rm 2}$ was also determined using equation 2 of @HT83. Considering the Einstein coefficient A(s$^{-1}$) for the NH$_{\rm 3}$(2,2) transition, the collision rate C(s$^{-1}$) from @D88, we obtain n(H$_{\rm 2}$) $\sim$ 1.5$\times$10$^6$ cm$^{-3}$. The total NH$_{\rm 3}$ column density was determined by summation of N(J,K) over all metastable rotational levels up to level J = 5, using Boltzmann’s Equation and assuming LTE. This density is somewhat less than that calculated using the 3 mm emission ($\sim$6.3$\times$10$^6$cm$^{-3}$). In addition, the H$_{\rm 2}$ column density was calculated as N(H$_{\rm 2}$) = ${\rm \Theta_S}$ n(H$_{\rm 2}$) where ${\rm \Theta_S}$ is the deconvolved linear size in cm ($3\rlap{.}{''}4 \pm$0.1 $\sim$ 0.22$\pm$0.01 pc at 13.5 kpc; using ${\rm \Theta_S}$ = $\sqrt{\theta_x \theta_y}$ from Table \[nh3obs\]). The relative abundance (X$_{\rm NH_3}$) between NH$_{\rm 3}$ and H$_{\rm 2}$ was calculated as N(NH$_{\rm 3}$)/N(H$_{\rm 2}$). A molecular gas mass of the clump using ammonia, M(NH$_{{\rm 3}}$), was determined using equation 12 of @GL99 with a radius R = 0.11$\pm$0.01 pc, giving a value of M(NH$_{\rm 3}$) = 830$\pm$40 M$_\odot$. Since the virial mass estimations (2200 M$_{\sun}$ from $^{13}$CS(2$-$1), and $\sim$1500 M$_{\sun}$ to $\sim$1800 M$_{\sun}$ from NH$_{\rm 3}$) are about twice the ammonia gas mass ($\sim$ 830 $\pm$40 M$_{\sun}$) we conclude that the G12–HMC may not be gravitationally bound. The physical parameters of this HMC are summarized in Table \[hotcore\]. HMC Internal Heating Source {#disnh3} --------------------------- From the flux density at 3 mm and the upper limit at 7 mm from the molecular clump emission (3$\sigma$ = 1.8 mJy), a spectral index of $\alpha{\ifmmode{_>\atop^{\sim}}\else{${_>\atop^{\sim}}$}\fi}$4.5 is obtained. This value suggests that the 3 mm emission arises from optically thin dust. If we consider that the dust in the molecular clump is only heated by external radiation, a lower limit to the dust temperature can be obtained using the formulation of @SK76 [equation 9], under the assumption that no absorption exists between the ionizing star of the region and the clump. Using the 3 mm integrated flux density (see Table \[vla\_cont\]) and assuming an external O6.5 ZAMS ionizing star [@KK01 L$_{\rm {star}}$ $\sim$ 1.5$\times$10$^5$L$_{\rm {\sun}}$], a dust absorptivity $\kappa_\nu \propto \nu^{+2}$, and a projected separation between the clump and the ionizing star (which is presumed to be at the focus of the cometary region) of 0.26 pc (4${\ifmmode{^{\prime \prime}}\else{$^{\prime \prime}$}\fi}$ at 13.5 kpc), we obtain T$_{\rm d}$ = 130 K. This temperature is slightly greater than the values obtained with other molecular tracers (see Table \[previous\], § \[nh3\] and § \[disnh3\]). However, not all luminosity of the O6.5 ZAMS ionizing star reaches the HMC. Considering a dilution factor of $\Omega$/4$\pi$ $\sim$0.06 (where $\Omega$ is the solid angle subtended by the ammonia core as seen from the ionizing star of the ),the fraction of incident luminosity from the on the HMC would be L$_{\rm inc}$ $\sim$ 9000 L$_{\rm {\sun}}$. Using this quantity as the energy budget of the HMC, a dust temperature of about 70 K is estimated, which is similar to the kinetic temperature obtained in section § \[nh3\] (86$\pm$12 K). This implies that an external heating source is not needed to explain the temperature of the core. However, the detection of water and methanol masers suggests the presence of a YSO in the HMC. HMC Kinematics {#disKin} -------------- There are different types of gas motion that can ocurr in HMCs, i.e., infall, rotation, and expansion/outflows. Outflows and rotation have been invoked to explain velocity gradients toward several HMCs [@C98]. In the case of rotation, flattened compact ‘disk–like’ structures with diameters greater than 0.1 pc [@GL99 and references therein] are observed, and the velocity gradient may indicate the kind of rotation: differential or rigid body. The NH$_{\rm 3}$(2,2) position–velocity (PV) diagrams of G12–HMC are presented in Figure \[fig7\]. We find no evidence of infall (e.g., redshifted absorption features). The central parts of the G12–HMC present larger velocity widths compared to the edges ($\sim$6 versus 2[ km s$^{-1}$]{}), and the overall structure of the HMC is elongated. The rotation scenario is possible. The velocity widths across the HMC in the PV diagrams are also consistent with expansion motions. Two components on each side of the HMC central position(systemic velocity of $\sim$23[ km s$^{-1}$]{}), have velocities of $\sim$26[ km s$^{-1}$]{} and $\sim$20[ km s$^{-1}$]{}, suggesting that G12–HMC could be a structure with expansion motions at velocities of a few km s$^{-1}$ in the SE–NW direction. We have obtained images of integrated velocities from $\sim$30 to 23[ km s$^{-1}$]{} and from 23 to 18[ km s$^{-1}$]{} using only the main line, and we do not see a clear spatial separation between this red–shifted and blue–shifted components. The opposite was found in the high–mass star forming regions G25.65+1.05 and G240.31+0.07, where high velocity molecular gas is associated with bipolar outflows [@SC96]. Higher angular resolution observations are needed to confirm the rotation hypothesis and whether an outflow is present in the core. Summary and Conclusions {#summary} ======================= We present observations of radio continuum (0.3, 0.7, 2 and 3.6 cm) and spectral lines (H41$\alpha$, $^{13}$CS(2$-$1) and (1$-$0), NH$_{\rm 3}$(2,2) and (4,4)) toward the cometary region G12.21–0.10. Continuum emission is detected at all wavelengths toward the UC H II region. We also find 3 mm continuum emission toward a molecular clump located about 4$''$ from the UC H II region. $^{\rm 13}$CS and NH$_{\rm 3}$ emission is only detected toward the molecular clump and we do not find any chemical differentiation between these species. We find evidence that the molecular clump is consistent with a hot (86 K), dense (1.5$\times$10$^6$ cm$^{-3}$), big (0.22 pc), and turbulent ($\rm \Delta V_{\rm obs} \sim$ 8 [ km s$^{-1}$]{}) molecular core. We also find that this hot molecular core shows a marginal velocity gradient in the SE–NW direction. Given that the HMC coincides with a water and methanol maser group, we speculate that a YSO, not detected in the centimeter radio continuum, could be forming inside the HMC. Finally, we find that $^{\rm 13}$CS and NH$_{\rm 3}$ are good candidates to trace the internal structure of HMCs, however, high angular resolution and higher sensitive observations are needed to fully characterize this HMC. The authors thank the anonymous referee for useful comments which improved this work. E. F. very gratefully acknowledges financial support from CONACyT (grant 124449, SNI III 1326 México), PROMEP$/$103.5$/$08$/$4722, and facilities from NRAO, NMT, IRyA–UNAM, INAOE. He also acknowledges Benemérita Universidad Autónoma de Puebla (BUAP); Arturo Fernández and Humberto Salazar, for several supports during a research stay on December, 2017. We would like to thank Laurant Loinard, Mayra Osorio, Guillem Anglada, Jana Benda, Simon Kemp, Susana Lizano, and Luis Felipe Rodríguez for useful discussions. We also thank Gilberto Zavala, Miguel Espejel, Alfonso Ginori and Andrés Rodríguez (IAM–UdeG) for computer support. P. H. acknowledges partial support from NSF grant AST-0908901 for this project. Beltr[á]{}n, M. T., Girart, J. M., & Estalella, R. 2006, , 457, 865 Cesaroni, R., Churchwell, E., Hofner, P., Walmsley, C. M., & Kurtz, S. 1994, , 288, 903 Cesaroni, R., Hofner, P., Walmsley, C. M., & Churchwell, E. 1998, , 331, 709 Cesaroni, R., Walmsley, C. M., & Churchwell, E. 1992, , 256, 618 Chandler, C. J., & Wood, D. O. S. 1997, , 287, 445 Churchwell, E., Walmsley, C. M., & Cesaroni, R. 1990, , 83, 119 Codella, C., Lorenzani, A., Gallego, A. T., Cesaroni, R., & Moscadelli, L. 2004, , 417, 615 Condon, J. J., Cotton, W. D., Greisen, E. W., et al. 1998, , 115, 1693 Danby, G., Flower, D. R., Valiron, P., Schilke, P., & Walmsley, C. M. 1988, , 235, 229 de la Fuente, E. 2007, Ph. D. Thesis, Departamento de Física, CUCEI, Universidad de Guadalajara, México. de La Fuente, E., Kurtz, S. E., Kumar, M. S. N., et al. 2009a, RMxAC, 37, 13 de La Fuente, E. et al. 2009b, Astrophysics and Space Science Proceedings, 7, 167 de La Fuente, E. et al. 2010, Proceedings of the International Astronomical Union, 262 319 De Buizer, J. M., Radomski, J. T., Telesco, C. M., & Pi[ñ]{}a, R. K. 2003, , 598, 1127 Garay, G., & Lizano, S. 1999, , 111, 1049 Hatchell, J., Thompson, M. A., Millar, T. J., & MacDonald, G. H. 1998, , 133, 29 Hatchell, J., Fuller, G. A., Millar, T. J., Thompson, M. A., & Macdonald, G. H. 2000, , 357, 637 Ho, P. T. P. 1977, Ph.D. Thesis, Massachusetts Institute of Technology, USA Ho, P. T. P., & Townes, C. H. 1983, , 21, 239 Hofner, P., & Churchwell, E. 1996, , 120, 283 Hofner, P., Wyrowski, F., Walmsley, C. M., & Churchwell, E. 2000, , 536, 393 Kim, K.-T., & Koo, B.-C. 2001, , 549, 979 Kim, K.-T., & Koo, B.-C. 2003, , 596, 362 Kurtz, S., Cesaroni, R., Churchwell, E., Hofner, P., & Walmsley, C. M. 2000, Protostars and Planets IV, 299 Kurtz, S., Hofner, P., & [Á]{}lvarez, C. V. 2004, , 155, 149 Morata, O., Girart, J. M., & Estalella, R. 2003, , 397, 181 Olmi, L., et al., 1993, , 276, 489 Ossenkopf, V., & Henning, T. 1994, , 291, 943 Poynter, R. L., & Kakar, R. K. 1975, , 29, 87 Plume, R., et al., 1997, , 476, 730 Rosero, V., Hofner, P., Claussen, M., et al. 2016, , 227, 25 S[á]{}nchez-Monge, [Á]{}., Schilke, P., Schmiedeke, A., et al. 2017, , 604, A6 Scoville, N. Z., & Kwan, J. 1976, , 206, 718 Scoville, N. Z., Carlstrom, J. E., Chandler, C. J., et al. 1993, , 105, 1482 Shepherd, D. S., & Churchwell, E. 1996, , 457, 267 Shirley, Y. L., Evans, N. J., II, Young, K. E., Knez, C., & Jaffe, D. T. 2003, , 149, 375 Stahler, S. W., & Palla, F. 2005, The Formation of Stars, by Steven W. Stahler, Francesco Palla, pp. 865. ISBN 3-527-40559-3. Wiley-VCH , January 2005., 865 Stutzki, J., Ungerechts, H., & Winnewisser, G. 1982, , 111, 201 Trinidad, M. A., Rojas, V., Plascencia, J. C., et al. 2003, RMxAA, 39, 311 Turner, B. E., Zuckerman, B., Palmer, P., & Morris, M. 1973, , 186, 123 Ungerechts, H., Winnewisser, G., & Walmsley, C. M. 1986, , 157, 207 van Dishoeck, E. F., & Blake, G. A. 1998, , 36, 317 Walmsley, C. M., & Ungerechts, H. 1983, , 122, 164 Walsh, A. J., Macdonald, G. H., Alvey, N. D. S., Burton, M. G., & Lee, J.-K. 2003, , 410, 597 Zinnecker, H., & Yorke, H. W. 2007, , 45, 481 [cccccc]{}Molecule & T$_{\rm k}$ & n(H$_{\rm 2}$) & Size & N(H$_{\rm 2}$) & Reference\ & (K) & (cm$^{-3}$) & (arcsec) & (cm$^{-2}$) &\ NH$_{\rm 3}$ & & 4.8$\times$10$^5$ & 4.0 & 5.2$\times$10$^{23}$ & (1)\ CS & & 8.1$\times$10$^5$ & & 2.0$\times$10$^{23}$ & (2),(5)\ CH$_{\rm 3}$CN & 80–90 & 10$^6$–10$^9$ & & 2.0$\times$10$^{25}$ & (3),(1)\ CH$_{\rm 3}$OH & 56${{+800}\atop{-24}}$ & 10$^7$ & 1.3 & 1.2 $\times$ 10$^{25}$ & (4)\ [ccccc]{}Observation & Wavelength & Right Ascencion & Declination & Source\ & [(cm)]{} & (J2000) & (J2000) &\ VLA continuum & 0.7, 2, & 3.6 & 18$^{h}$ 12$^{m}$ 39.8$^{s}$ & –18$^{\circ}$ 24$'$ 21$''$ & region\ OVRO H41$\alpha$ & 0.33 & 18$^{h}$ 12$^{m}$ 39.6$^{s}$ & –18$^{\circ}$ 24$'$ 21$''$ & region\ OVRO continuum & 0.30 & 18$^{h}$ 12$^{m}$ 39.8$^{s}$ & –18$^{\circ}$ 24$'$ 21$''$ & region (peak–1)\ OVRO continuum & 0.30 & 18$^{h}$ 12$^{m}$ 39.8$^{s}$ & –18$^{\circ}$ 24$'$ 17$''$ & molecular clump (peak–2)\ OVRO $^{13}$CS(2$-$1) & 0.32 & 18$^{h}$ 12$^{m}$ 39.7$^{s}$ & –18$^{\circ}$ 24$'$ 18$''$ & molecular clump\ VLA NH$_3$(2–2) & 1.26 & 18$^{h}$ 12$^{m}$ 39.8$^{s}$ & –18$^{\circ}$ 24$'$ 18$''$ & molecular clump\ VLA NH$_3$(4–4) & 1.24 & 18$^{h}$ 12$^{m}$ 39.8$^{s}$ & –18$^{\circ}$ 24$'$ 18$''$ & molecular clump\ VLA $^{13}$CS(1$-$0) & 0.65 & 18$^{h}$ 12$^{m}$ 39.8$^{s}$ & –18$^{\circ}$ 24$'$ 18$''$ & molecular clump\ VLA continuum & 1.3 & 18$^{h}$ 12$^{m}$ 39.7$^{s}$ & –18$^{\circ}$ 24$'$ 21$''$ & region\ [ccccccc]{} Region & Telescope & $\lambda$ & $S_\nu$ & Beam Size; P.A & RMS Noise & Deconvolved Size\ & & (cm) & (mJy) & (arcsec; deg) & (mJy beam$^{-1}$) & (arcsec)\ & VLA CnB & 3.6 & 220 & 2.31 $\times$ 0.91; $-70$ & 0.31 & 5 $\times$ 4\ & VLA CnB & 2.0 & 200 & 1.04 $\times$ 0.55; $+80$ & 0.21 & 8 $\times$ 7\ & VLA CnB & 0.7 & 160 & 1.65 $\times$ 0.83; $+77$ & 0.58 & 4 $\times$ 5\ Molecular & &&&&&\ clump & OVRO & 0.3 & 150 &2.82 $\times$ 2.32; $-43$ & 1.70 & 3 $\times$ 1\ \ Extended & &&&&&\ emission & VLA D & 3.6 & 1320 &12.50 $\times$ 7.30; $-13$ & 0.15 & 210 $\times$ 156\ [ccccccc]{} Line & Beam Size; P.A. & RMS noise & S$_\nu$ & V$_{\rm LSR}$ & $\Delta$V$_{\rm obs}$ & Deconvolved Size\ & (arcsec; deg) & (mJy beam$^{-1}$) & (mJy) & km s$^{-1}$ & km s$^{-1}$ & (arcsec)\ $^{13}$CS(2-1) & 3.18 $\times$ 2.51; $-43$ & 35 & 300 & 23.7$\pm$0.1 & 9.4$\pm$0.5 & 4.5 $\times$ 2.0\ H41$\alpha$ & 3.18 $\times$ 2.51; $-43$ & 35 & 335 & 29.4$\pm$1.2 & 23.9$\pm$3.1 &\ [ccccccc]{} Transition & Beam Size; P.A. & RMS Noise & Peak S$_\nu$ & V$_ {\rm {LSR}}$ & $\Delta {\rm V_{obs}}$ & Deconvolved Size\ (J,K; line) & (arcsec; deg) & (mJy beam$^{-1}$) & (mJy) & (km s$^{-1}$) & (km s$^{-1}$) & (arcsec)\ (2,2; main) & 3.54$\times$1.36; $+60$ & 4.4 & 260$\pm$5 & 23.1$\pm$0.3 & 8.1$\pm$0.7 & 3.8 $\times$ 3.0\ (2,2; inner satellites) & 3.54$\times$1.36; $+60$ & 5.7 & 59$\pm$2 & 6.5, 39.7 & 9.5 &\ (2,2; outer satellites) & 3.54$\times$1.36; $+60$ & 6.1 & 59$\pm$2 & –2.7, 48.9 & 8.5 &\ (4,4; main) & 4.29$\times$1.18; $+56$ & 3.2 & 106$\pm$5 & 23.3$\pm$0.3 & 8.9$\pm$0.7 & 3.0 $\times$ 1.0\ (4,4; inner satellites) & 4.29$\times$1.18; $+56$ & 3.3 & 16.0$\pm$0.8 & –0.7, 27.3 & 13.2 &\ (4,4; outer satellites) & 4.29$\times$1.18; $+56$ & 3.3 & 16.0$\pm$0.8 & –6.7, 53.3 & &\ [cccccc]{} T$_{42}$ & T$_{\rm k}$ & N(H$_{\rm 2}$) & M(NH$_{\rm 3}$) & N(NH$_{\rm 3}$) & X(NH$_{\rm 3}$)\ (K) & (K) & (10$^{\rm 24}$ cm$^{\rm -2}$) & (M$_\odot$) & (10$^{\rm 17}$ cm$^{\rm -2}$) & (10$^{\rm -7}$)\ $77\pm10$ & $86\pm12$ & 1.0$\pm$0.4 & 830$\pm$40 & $1.1\pm0.4$ & 1.1$\pm$0.7\ ![G12.21–0.10 (G12). [a)]{} [Left]{}: 21 cm NVSS map (contours) superposed with 3.6 cm VLA D-configuration image in gray-scale [@F07]. The NVSS beam is 45$''$, while the 3.6 cm beam, shown in the lower-left corner, is $12\rlap{.}{''}5$ $\times$ $7\rlap{.}{''}3$ at P.A. = $-$13$^{\circ}$. Contours are –3, 3, 6, 12, 24, 48, 68, 96 $\times$ 3 mJy beam$^{\rm -1}$. The gray-scale is 0.16 to 8 mJy beam$^{\rm -1}$. These maps represent the extended emission in G12. [Right]{}: A close up view of the UC component traced by radio continuum emission at 3 mm (see § 2). Contours are –3, 3, 9, 18, 27, 48 times RMS noise of 1.7 mJy beam$^{\rm -1}$. The beam size is shown in the bottom left. The positions of the the region and the molecular clump are labeled as 1 and 2 respectively. In all panels, the crosses mark the position of the water masers from @HC96. [b)]{} Radio continuum emission at 0.7, 2, and 3.6 cm (VLA-CnB) toward G12. The contour levels correspond to –1, 1, 2, 4, 8, 16, 32, 64 $\times$ RMS noise presented in Table \[vla\_cont\] for each wavelength. The synthesized beam for each map is shown in the lower left corner (see Table \[vla\_cont\]). []{data-label="fig1"}](Figure1.eps){width="\columnwidth"} ![H41$\alpha$ and $^{\rm 13}$CS(2$-$1) spectral lines. The solid line represents the best adjusted Gaussian in each spectrum. The CS spectrum also shows an unexpected weaker line at $\sim$4.5[ km s$^{-1}$]{}. A possible candidate for this line is C$_{\rm 3}$S.[]{data-label="fig2"}](Figure2.eps){width="\columnwidth"} ![$^{\rm 13}$CS(2$-$1) channel maps (contours) with channel width of 1.62[ km s$^{-1}$]{} are shown between 16 and 29.0[ km s$^{-1}$]{} superposed on the radio continuum emission at 3.6 cm (VLA-CnB, gray-scale) of G12. The gray-scale goes from $-$1.24 to 79.83 mJy beam$^{-1}$. Contours are –3, 3, 4, 5, 6, 7, 8, 9 $\times$ 35 mJy beam$^{-1}$ (RMS noise). The synthesized beam of the $^{\rm 13}$CS(2$-$1) observations ($3\rlap{.}{''}18$ $\times$ $2\rlap{.}{''}51$ with a P.A. of $-$43.3$^\circ$) is shown in top left panel. The crosses mark the position of water masers from @HC96. []{data-label="fig3"}](Figure3.eps){width="\columnwidth"} ![ $^{\rm 13}$CS(1$-$0) maps of individual velocity channels from the main line spectra of G12 (contours), superposed with the $^{\rm 13}$CS(2$-$1) integrated map (in gray scale). Velocities from 18.9 to 28.4 [ km s$^{-1}$]{} are shown in each map. The $^{\rm 13}$CS(1$-$0) integrated emission map (inset in the 27.2 [ km s$^{-1}$]{}panel) shows a complex edge-like structure morphology with three components labeled E (East), W (West), and N (North). The maximum of emission for each of these components corresponds to $\sim$20, 23 and 26 [ km s$^{-1}$]{} respectively. The contours are –3, 3, 4, 5 $\times$ 3.6 mJy (RMS noise). The beam size is $1\rlap{.}{''}44$ $\times$ $1\rlap{.}{''}26$ with a P.A. of –79$^\circ$, shown in the bottom left of the first map. The crosses mark the position of water masers from @HC96.[]{data-label="fig4"}](Figure4.eps){width="\columnwidth"} ![NH$_{\rm 3}$ integrated spectra of (2,2) and (4,4) transitions in the G12–HMC . The solid lines show Gaussian fits.[]{data-label="fig5"}](Figure5.eps){width="\columnwidth"} ![G12–HMC. [a)]{} NH$_{\rm 3}$(2,2) moment–1 map (gray-scale), superposed to the radio continuum emission at 1.3 cm (contours; region) obtained from line–free channels of the NH$_{\rm 3}$(2,2) observations. Gray–scale is in the interval $\sim$ 19–27[ km s$^{-1}$]{}. Contours are –4, 4, 6, 10, 15, 20, 25, 30, 35, 40 $\times$ 2 mJy beam$^{-1}$ (RMS noise). [b)]{} NH$_{\rm 3}$(2,2) moment–1 map (gray-scale) superposed to the NH$_{\rm 3}$(4,4) integrated emission (contours). Contours are 20, 50, 100, 140, 180, 220, 260 mJy beam$^{-1}$km s$^{-1}$. Axes labeled as X and Y at the lower right corner show the directions used to generate the PV diagrams presented in Figure \[fig7\]. [c)]{} $^{\rm 13}$CS (2$-$1) moment–1 map (gray-scale) of the emission presented in Figure \[fig4\], superposed to the NH$_{\rm 3}$(2,2) integrated emission with countours in 50, 150, 250, 350, 450, 550 mJy beam$^{-1}$km s$^{-1}$. [d)]{} $^{\rm 13}$CS (1$-$0) moment–1 map (gray-scale), superposed is the NH$_{\rm 3}$(4,4) integrated emission in contours as in b). Synthesized beams are shown at the bottom left of each plot and corresponds to NH$_{\rm 3}$ in a) and b), $^{\rm 13}$CS(2$-$1) in c), and $^{\rm 13}$CS (1$-$0) in d). In all figures, crosses mark the position of water masers [@HC96].[]{data-label="fig6"}](Figure6.eps){width="\columnwidth"} ![NH$_{\rm 3}$(2,2) position–velocity (PV) diagram of the HMC presented as grey-scale in Figure \[fig6\]a and \[fig6\]b. The plots are slices in the directions X and Y shown in the lower-right corner in Figure \[fig6\]b. The origin of the position axes is located at $\alpha_{2000}$ = 18$^{\rm h}$ 12$^{\rm m}$ 40$^{\rm s}$, $\delta_{2000}$ = –18$^{\circ}$ 24$'$ 15$''$. Contours are –4, 4, 5, 6, 7, 8, 9, 10, 11 $\times$ 1.6 mJy beam$^{-1}$. Dashed lines mark the systemic velocity of the core at $\sim$23[ km s$^{-1}$]{}.[]{data-label="fig7"}](Figure7.eps){width="\columnwidth"} [^1]: The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc.	{ "pile_set_name": "ArXiv" }

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.