Beruflich Dokumente
Kultur Dokumente
Mathias Raer
Universittsverlag Potsdam
Mathias Rafler
Universittsverlag Potsdam
Die Deutsche Nationalbibliothek verzeichnet diese Publikation in der Deutschen Nationalbibliografie; detaillierte bibliografische Daten sind im Internet ber http://dnb.d-nb.de/ abrufbar.
http://info.ub.uni-potsdam.de/verlag.htm Am Neuen Palais 10, 14469 Potsdam Tel.: +49 (0)331 977 4623 / Fax: 3474 E-Mail: verlag@uni-potsdam.de Dieses Werk ist unter einem Creative Commons Lizenzvertrag lizenziert: Namensnennung - Keine kommerzielle Nutzung - Weitergabe unter gleichen Bedingungen 3.0 Deutschland Um die Bedingungen der Lizenz einzusehen, folgen Sie bitte dem Hyperlink: http://creativecommons.org/licenses/by-nc-sa/3.0/de/ Zugl.: Potsdam, Univ., Diss., 2009 Online verffentlicht auf dem Publikationsserver der Universitt Potsdam: URL http://pub.ub.uni-potsdam.de/volltexte/2009/3870/ URN urn:nbn:de:kobv:517-opus-38706 http://nbn-resolving.org/urn:nbn:de:kobv:517-opus-38706 Zugleich gedruckt erschienen im Universittsverlag Potsdam: ISBN 978-3-86956-029-8
Zusammenfassung
Zuf allige Punktprozesse beschreiben eine (zuf allige) zeitliche Abfolge von Ereignissen oder eine (zuf allige) r aumliche Anordnung von Objekten. Deren wichtigster Vertreter ist der Poissonproze. Der Poissonproze zum Intensit atsma , das Lebesgue-Ma ordnet jedem Gebiet sein Volumen zu, erzeugt lokal, d.h in einem beschr ankten Gebiet B , gerade eine mit dem Volumen von B poissonverteilte Anzahl von Punkten, die identisch und unabh angig voneinander in B plaziert werden; im Mittel ist diese Anzahl pB q. Ersetzt man durch ein Vielfaches a, so wird diese Anzahl mit dem a-fachen Mittelwert erzeugt. Poissonprozesse, die im gesamten Raum unendlich viele Punkte realisieren, enthalten bereits in einer einzigen Stichprobe gen ugend Informationen, um Statistik betreiben zu k onnen: Bedingt man lokal bzgl. der Anzahl der Teilchen einer Stichprobe, so fragt man nach allen Punktprozessen, die eine solche Beobachtung h atten liefern k onnen. Diese sind Limespunktprozesse zu dieser Beobachtung. Kommt mehr als einer in Frage, spricht man von einem Phasen ubergang. Da die Menge dieser Limespunktprozesse konvex ist, fragt man nach deren Extremalpunkten, dem Rand. Im ersten Teil wird ein Poissonproze f ur ein physikalisches Teilchenmodell f ur Bosonen konstruiert. Dieses erzeugt sogenannte Loops, das sind geschlossene Polygonz uge, die dadurch charakterisiert sind, da man an einem Ort mit einem Punkt startet, den mit einem normalverteilten Schritt l auft und dabei nach einer gegebenen, aber zuf alligen Anzahl von Schritten zum Ausgangspunkt zur uckkehrt. F ur verschiedene Beobachtungen von Stichproben werden zugeh orige Limespunktprozesse diskutiert. Diese Beobachtungen umfassen etwa das Z ahlen der Loops gem a ihrer L ange, das Z ahlen der Loops insgesamt, oder das Z ahlen der von den Loops gemachten
Schritte. Jede Wahl zieht eine charakteristische Struktur der invarianten Punktprozesse nach sich. In allen hiesigen F allen wird ein charakteristischer Phasen ubergang gezeigt und Extremalpunkte werden als spezielle Poissonprozesse identiziert. Insbesondere wird gezeigt, wie die Wahl der Beobachtung die L ange der Loops beeinut. Geometrische Eigenschaften dieser Poissonprozesse sind der Gegenstand des zweiten Teils der Arbeit. Die Technik der Palmschen Verteilungen eines Punktprozesses erlaubt es, unter den unendlich vielen Loops einer Realisierung den typischen Loop herauszupicken, dessen Geometrie dann untersucht wird. Eigenschaften sind unter anderem die euklidische L ange eines Schrittes oder, nimmt man mehrere aufeinander folgende Schritte, das Volumen des von ihnen denierten Simplex. Weiterhin wird gezeigt, da der Schwerpunkt eines typischen Loops normalverteilt ist mit einer festen Varianz. Der dritte und letzte Teil befat sich mit der Konstruktion, den Eigenschaften und der Statistik eines neuartigen Punktprozesses, der P olyascher Summenproze genannt wird. Seine Konstruktion verallgemeinert das Prinzip der P olyaschen Urne: Im Gegensatz zum Poissonproze, der alle Punkte unabh angig und vor allem identisch verteilt, werden hier die Punkte nacheinander derart verteilt, da der Ort, an dem ein Punkt plaziert wird, eine Belohnung auf die Wahrscheinlichkeit bekommt, nach der nachfolgende Punkte verteilt werden. Auf diese Weise baut der P olyasche Summenproze T urmchen, indem sich verschiedene Punkte am selben Ort stapeln. Es wird gezeigt, da dennoch grundlegende Eigenschaften mit denjenigen des Poissonprozesses u bereinstimmen, dazu geh oren unendliche Teilbarkeit sowie Unabh angigkeit der Zuw achse. Zudem werden sein LaplaceFunktional sowie seine Palmsche Verteilung bestimmt. Letztere zeigt, da die H ohe der T urmchen gerade geometrisch verteilt ist. Abschlieend werden wiederum Statistiken, nun f ur den Summenproze, diskutiert. Je nach Art der Beobachtung von der Stichprobe, etwa Anzahl, Gesamth ohe der T urmchen oder beides, gibt es in jedem der drei F alle charakteristische Limespunktprozesse und es stellt sich heraus, da die zugeh origen Extremalverteilungen wiederum P olyasche Summenprozesse sind.
ii
Vorwort
Diese Arbeit fat die Ergebnisse meines Studiums der Punktprozesse von April 2006 bis Juni 2009 an der Universi at Potsdam unter der Betreuung von Prof. Dr. S. Rlly und Prof. Dr. H. Zessin (Universit at Bielefeld) zusammen. Dabei war ich seit Oktober 2006 Stipendiat der International Research Training Group (IRTG) Stochactic Models of Complex Processes Berlin-Z urich (SMCP). Mein Dank gilt vor allem Prof. Dr. S. Rlly f ur die Betreuung, ihr umfangreiches Engagement, ihre wissenschaftliche Unterst utzung sowie eine fortw ahrende Bereitschaft zur Diskussion u ber den aktuellen Stand der Arbeit. Des Weiteren m ochte ich bei Prof. Dr. H. Zessin bedanken, dessen Impulse und anregende Fragestellungen zu den beiden Themenbereichen f uhrten. Ausgesprochen wertvoll waren ihre zahlreichen, kritischen Anmerkungen in der letzten Phase der Arbeit. F ur die nanzielle Unterst utzung sowie die F orderung bin ich dem IRTG sehr dankbar. Viele Veranstaltungen trugen wesentlich zu einer Erweiterung meines Horizonts sowie Einblicken in weitere Gebiete der Wahrscheinlichkeitstheorie bei, dazu m ochte ich neben den regelm aigen Seminaren und Minikursen die Fr uhlingsschule in Potsdam sowie die Sommerschule in Disentis hervorheben. Das IRTG erm oglichte mir einen erfahrungsreichen Aufenthalt an der Universit at Z urich bei Prof. Dr. Bolthausen. Die in diesem Rahmen entstandenen Kontakte haben mich sowohl fachlich als auch pers onlich sehr bereichert. Dar uber hinaus sch atze ich die stets sehr angenehme und oene Atmosph are, die sowohl im IRTG, als auch in der Arbeitsgruppe in Potsdam herrschte. Insbesondere gilt mein Dank meiner Familie, meinen Eltern, die mich
iii
w ahrend des gesamten Studiums sowie der Promotion in jeglicher Hinsicht unterst utzten.
Mathias Raer
iv
Je planm aiger der Mensch vorgeht, um so wirkungsvoller trit ihn der Zufall. Friedrich D urrenmatt 21 Punkte zu den Physikern, Punkt 8
Contents
0. Introduction 1
I.
1. Point Processes 11 1.1. Point Processes . . . . . . . . . . . . . . . . . . . . . . . . 12 1.1.1. Basic Notions . . . . . . . . . . . . . . . . . . . . . 12 1.1.2. Moment Measures and Functionals of Point Processes 16 1.1.3. The Poisson Process . . . . . . . . . . . . . . . . . . 18 1.2. The Campbell Measure of a Point Process . . . . . . . . . 23 1.2.1. Disintegration with respect to the Intensity Measure: Palm Distributions . . . . . . . . . . . . . . . . . . . 24 1.2.2. Disintegration with respect to the Point Process: Papangelou Kernels . . . . . . . . . . . . . . . . . . . . 26 1.2.3. Construction of Point Processes from Papangelou Kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 2. Deviation Principles 2.1. General Large Deviation Principles . . . . . . . . . . . . . . 2.2. Large Deviations for Poisson Processes at increasing Intensity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3. Deviations for Brownian Motions and Brownian Bridges . . 39 40 42 45
vii
Contents
49
51 52 63 66 66 67 69 75 78 79 91 94 98 98 102 108 108 110 112 115 116 122
III. A Generalisation of the P olya Urn Schemes: the P olya Sum Process 127
6. The P olya Sum Processes 129 6.1. The Denition of the P olya Sum Process . . . . . . . . . . . 130
viii
Contents 6.2. Laplace Functionals . . . . . . . . . . . . . . . . . . . . . . 132 6.3. Disintegration and Partial Integration . . . . . . . . . . . . 135 7. Limit Theorems for Conditioned 7.1. The Turret Ensemble . . . 7.2. The Brick Ensemble . . . . 7.3. The General Ensemble . . 7.4. Large Deviations . . . . . . P olya Sum Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 140 142 145 147
8. Concluding Remarks
155
ix
0. Introduction
The accidental occurrence of certain events in time, like incoming phone calls in a call centre, the growth of a queue, impulses of nerve bres or detection of ionising radiation by a Geiger-M uller counter, is a very fundamental problem in probability. At discrete, but random times specied events occur, which leads to the rst idea of counting processes; processes which count the number of certain events in some time interval. Such processes may be easily described by the (random) waiting time between two events. An important role play exponential waiting times, since this distribution is known to be memoryless. If these waiting times are in addition assumed to be independently and identically distributed, the number of events in some given time interval has a Poisson distribution proportional to the length of the interval. Moreover, the quantities of dierent periods of time are independent. These are the characterising properties of the Poisson process. In focusing the events as points in time, temporal counting processes are extended to spatial counting processes. Events are now points in space, which may be counted in any bounded domain. A primer example is given by the positions of physical particles, or even individuals, animals, plants, stars. In any case a realisation of a point process is a snapshot of some situation. Moreover, points may be replaced by geometric objects, such as spheres representing holes in some porous medium or hard-core particles as well as line segments representing fractures of the earths surface. Natural questions refer for instance to the size of clusters built through overlapping objects. Such percolation problems were addressed by e.g. Hall [Hal85] and will occur in the second part of this work. In general point processes can be dened on polish spaces, compact presentations are Kerstan, Matthes and
0. Introduction Mecke [KMM74] and Kallenberg [Kal83], Daley and Vere-Jones [DVJ08a, DVJ08b]. Very often earthquakes cause further earthquakes nearby their epicentre, osprings of a tree grow not too far away from their parental tree or settlements are very unlikely to be isolated. The rst of the examples may be considered in time as well, an earthquake causes further earth tremors. These examples show that naturally dependencies between points arise: points aliate to clusters and dene some kind of families of a population. Objects are divided into classes of related objects: the set of points is partitioned. Considerations about the sizes of families of a population, though without any spatial or temporal component, can be found in the works of Ewens and Kingman [Kin78a, Kin78b]. Starting from a sample of a population of size N , the dierent species dene a (random) integer partition of N , and by the demand for a consistent sampling procedure, the latter author gave a characterisation of the limiting proportions of these species whenever the proportions are in descending order. Two interesting aspects of Kingmans results should be pointed out, and this work shows comparable results: Firstly, the set of limits is convex and the limits themselves have a representation as mixtures of extremal points of the whole set. Therefore the analysis may be restricted to the set of extremal points. Secondly, the limiting proportions are not necessarily proportions in the sense that they sum up to 1. In fact there is the possibility that really many small families get lost in the limiting procedure. If the amount of such families is suciently large, then the small families all together may contribute to the whole population. Chapter 4, in particular section 4.5 touches related questions. Permutations on a nite set of N elements dene integer partitions by determining the cycle sizes of the cycle decomposition of a permutation. In this manner random permutations lead to random integer partitions. It is a highly non-trivial task to nd out about limiting objects and sizes of cycles when expanding the set of permuted elements; such questions were addressed by Tsilevich [Tsi97], Vershik and Schmidt [VS77, VS78]. They studied the asymptotic behaviour of functionals on symmetric groups which
only depends of the length on the cycles. An important application relating random integer partitions and their extremal limits on the one hand and point processes on the other hand are quantum particle systems. Consider a nite system in equilibrium described by a Hamiltonian H . A pair p, q satisfying H , where is a real number and a square integrable, normalised function with square integrable second derivative, represents the system at energy . Such solutions satisfy the so-called Boltzmann statistics. Combinatorial diculties enter as soon as one demands additional symmetry properties, which are invariance of under any permutation of the particles, , for Bose statistics, and invariance under any permutation with an added minus for odd permutations, sgnp q , for Fermi-Dirac statistics. Feynman [Fey48, Fey90] introduced functional integration, which was treated rigorously by Kac. His method is applied to the object of interest, the statistical operator exppH q, where 0 is the inverse temperature. Ginibre [Gin71] carried out this analysis and obtained an integration on closed trajectories, i.e. Brownian bridges, also named loops. For Boltzmann statistics these loops are exactly of length . The introduction of the invariance under permutations for the other two statistics is a sophisticated part, but has an interesting eect; its treatment was Ginibres important step. While for the Boltzmann statistics the end point of each loop is equal to its starting point, in Bose or Fermi-Dirac statistics starting point and end point are not obliged to be identical. Indeed, the symmetrisation of N elementary trajectories means to obtain the end points of these trajectories from a permutation of their starting points. Since every permutation decomposes into cycles, the set of elementary trajectories decomposes into classes of connected trajectories, where two trajectories and I are connected if and only if there exists a sequence of successive trajectories with the rst being and the nal one being I . These classes are called composite loops. Therefore in a natural way the equivalence relation on the set of elementary trajectories denes an integer partition of N . One starting point of this work will be the interpretation of Ginibres Feynman-Kac representation of exppH q as a Poisson process P z on the space of composite
0. Introduction loops. A basic question originates in the pioneering work of Bose and Einstein in the 1920s about Bosons. They proposed a curious phase transition, nowadays named Bose-Einstein condensation. They showed that if the particle density exceeds some critical value, a positive fraction of the whole amount of particles conglomerates or condenses in the lowest eigenstate. In 1938 London proposed that a phase transition between He I and He II is related to the Bose-Einstein condensation. But not until 1995 BoseEinstein condensation was observed experimentally in a gas of Rubidium and Natrium. The physicists Cornell, Ketterle and Wieman received the Nobel price for that experiment in 2001. Feynman [Fey53b, Fey53a] again proposed that Bose-Einstein condensation occurs if and only if innitely long loops occur with positive probability. The connection between Bose-Einstein condensation and cycle percolation has been established in two articles by S ut o [S ut93, S ut02] and Benfatto et al [BCMP05] in the mean eld. S ut o considers a model on random integer partitions and Benfatto et al a mean eld model, both did not take spatial relations into account. Fichtner pointed out the connection between random permutations of countable subsets of Rd and its decomposition into nite clusters in [Fic91b] and moreover gave a characterisation of the position distribution of the Bose gas in terms of its moment measures of a point process on Rd in [Fic91a]. Later Ueltschi [Uel06a, Uel06b] examined lattice models on the basis of S ut os work and thereby introduced so-called spatial permutations. Very recently, Ueltschi and Betz [BU09, Uel08] generalised the lattice model to models of random point congurations in a continuous space. By symmetrising initial and terminal conditions of Brownian bridges of a given length , Adams and K onig [AK07] construct for each Brownian bridge a successor starting at the terminal point, and a predator ending at the starting point. In that way connected bridges dene loops (as classes of connected Brownian bridges). In chapters 3 5 a related model is considered. Initial point is the already mentioned Feynman-Kac representation of the Bose gas obtained by Ginibre. The specic model is constructed in chapter 3, which contains
the construction of the space of composite loops and the Poisson process P z thereon. Furthermore properties of the intensity measure z are shown, such as a factorisation and asymptotics. Chapter 4 is devoted to limit theorems of local specications derived from P z conditioned on dierent z to obtain representation theorems for P invariant elds. The rst section of this chapter introduces the notion of local specications and the Martin-Dynkin boundary technique. Dierent ways of counting lead to dierent invariant elds and they are introduced and studied in the following sections. Firstly by counting loops for the microcanonical, canonical and grand canonical ensemble, loop representations of their corresponding Martin-Dynkin boundary is obtained in terms of extremal points by direct computations. The most delicate part consists of the determination of the Martin-Dynkin boundary in the canonical ensemble of elementary components in section 4.5. The large deviation principle from section 2.2 allows the representation of limits of random integer partitions in terms of a variational problem with constraints, which is solved afterwards. This procedures allows the determination of the essential part of the canonical Martin-Dynkin boundary. The complex of the limits of integer partitions gives insight into a global property with no focus on spatial properties of the loops. Chapter 5 faces geometric properties of congurations weighted by P z . The main object is the typical loop under P z . Since P z realises congurations of a countably innite number of loops, and due to the lack of an uniform distribution on countably innite sets, there is no natural denition of a typical loop. This implies a change of the point of view on the point process: from the number of points in some region to the single point. The modern denition of the typical point has its origin in the work of Kummer and Matthes [KM70] with the introduction of the Campbell measure, which is also developed in the monograph of Kerstan, Matthes and Mecke [KMM74]. In using this concept of the typical loop, properties such as its barycentre, its euclidean length and the number of its extremal points are considered. Furthermore a percolation problem of the congurations is treated. A fundamental characteristic of the Poisson process is that points are
0. Introduction placed independently and foremost each one with the same intensity. Papangelou processes, apart from the Poisson process, contrast this construction. In [MWM79] Papangelou processes were characterised by a partial integration formula. Recently, Zessin [Zes09] gave a direct construction of these larger class of processes. Particularly the points are placed according to conditional intensities. Zessins construction is reproduced in subsection 1.2.3 and simplied under an additional assumption. These preparations unfold their relevance in chapter 6, where the so-called P olya sum process, which rstly occurred in [Zes09], is constructed. Instead of placing the points independently and, most notably, identically, the mechanism of placing the points makes use of P olyas urn dynamics: points are placed successively and each location, at which a point is placed, gets a reward on the probability to get another point. That way turrets are built from bricks. Apart from this building brick construction, the P olya sum process is shown to share many properties with the Poisson process, particularly innite divisibility and complete randomness. Moreover, its Palm kernels are characterised. In chapter 7 again limit theorems for local specications are shown, this time for the P olya sum process and dierent limiting stochastic elds obtained from dierent observations: rstly by counting turrets, then by counting bricks and nally by counting turrets and bricks. Particularly the methods used to obtain the last two ensembles are related to the methods used in section 4.5. The fundament for these discussions is laid in chapters 1 and 2. Basic tools are introduced in required generality and discussed. Section 1.1 deals with the denition of point processes and Poisson processes on complete, separable metric spaces. Their basic construction via Laplace functionals is recalled as well as the moment measures dened. Important properties of the Poisson process, such as complete randomness and innite divisibility, and their classication into larger classes of point processes follow. Further properties which are not shared by all Poisson processes, like orderliness and stationarity, are reviewed. Particularly stationarity, the invariance under translations, leads to helpful factorisations. In section 1.2 the concept of
the Campbell measure is introduced, and its disintegrations, which lead to Palm and Papangelou kernels, are recalled. Chapter 2 recalls deviation principles, which are used to obtain properties of the Poisson process P z constructed in chapter 3 and the limits of converging sequences of measures as solution of certain minimisation problems. An important role for the former application play large and small deviations of Brownian bridges. They drive the asymptotic behaviour of the intensity measure z . The latter application is prepared in section 2.2 and applied in chapters 4 and 7.
Part I.
1. Point Processes
In this very rst part the theoretical background is developed, the framework of point processes on complete, separable metric spaces (c.s.m.s.). This covers the basic objects in subsection 1.1.1, particularly random measures and random point measures following the books of Daley and VereJones [DVJ08a, DVJ08b] as well as Kerstan et al. [KMM74] and Kallenberg [Kal83]. Thereafter in subsection 1.1.2 point processes are focused. Besides the introduction of the intensity measure and higher moment measures, the main question is to characterise point processes as in von Waldenfels [vW68]. Since random measures are non-negative, the Laplace functional turns out to be sucient. Subsection 1.1.3 starts with the denition of the Poisson process on a c.s.m.s. X in terms of its Laplace functional and classies its most important basic properties, which will be needed in this work and underline the nature of the Poisson process. These are complete randomness given by Kingman [Kin67] and innite divisibility. Further properties like stationarity and orderliness do not follow directly from the general denition of the Poisson process, but need additional assumptions. If X is an Abelian group with corresponding Haar measure (in fact an arbitrary, but xed Haar measure), then a Poisson process is stationary if and only if its intensity measure is a multiple of . Even more generally, if T is an Abelian group acting measurably on X , then the factorisation theorem yields that a T -invariant Poisson processs intensity measure disintegrates into a multiple m of and a probability measure s on a spce of marks located at s, pdxq ms pdxq pdsq. Moreover the simplicity of a point process is related to properties of the
11
1. Point Processes second order moment measure. The orderliness of a Poisson process also addresses the multiplicity of points. For the Poisson process this shows that simplicity is equivalent to the absence of atoms of its intensity measure. A fundamental tool in point process theory are Campbell and reduced Campbell measure recalled in section 1.2, which allows to change the point of view by disintegration. While the Laplace functional describes the nitedimensional distributions, the disintegration of the Campbell measure with respect to the intensity measure of the point process yields the Palm kernel, which is the point process conditioned on the event that at a certain site a point is present. The result is the point process from the point of view of a single point which is almost surely present in the realisations. Palm distributions are the important tool in chapter 5 and occur a second time in chapter 6. Under additional assumptions, the reduced Campbell measure is absolutely continuous with respect to the product of the intensity measure of and the point process itsself. The disintegration, which yields the Papangelou kernel, allows a further change of the point of view: the Papangelou kernel is the intensity of a point conditioned on the presence of a certain conguration. These relations were established in [MWM79]. Very recently Zessin [Zes09] started with some kernel and constructed the point process for which the kernel is a Papangelou kernel. In section 1.2.3 we reproduce his proof, and give in theorem 1.31 a simpler construction under an additional measurability assumption. Moreover we generalise Zessins proof and correct an inaccuracy. A primer example, the P olya sum process dened in [Zes09], is presented and studied in chapters 6 and 7.
12
1.1. Point Processes great importance is the ring B0 pX q of bounded Borel sets allowing us to dene locally nite measures on the measurable space X, B pX q . Denition 1.1 (Locally nite measures). A Borel measure on the c.s.m.s. X is locally nite if pB q V for every B B0 pX q. These measures may contain an innite mass, but locally only a nite amount is allowed. Point congurations, i.e. generalised subsets of X , which are locally nite, play a central role. They are expressed as measures which only take non-negative integer values on bounded sets. Denition 1.2 (Measure spaces). Dene the following spaces of measures: i) MpX q is the space of locally nite Borel measures on B pX q, ii) M pX q MpX q : pB q N dB iii) M pX q M pX q : {x}
B0pX q
,
1 dx X
iv) Mf pX q MpX q : pX q V , analogously M f pX q. Hence M pX q is the set of all locally nite point measures on X , M pX q the set of all locally nite, simple point measures and likewise Mf pX q and M f pX q the corresponding sets of measures of nite total mass. Any locally nite subset of X can be represented as an element of M pX q. With M pX q we thus allow multiple points. For measurable functions f : X R write pf q : f d.
We say that a sequence of nite measures pn qn Mf pX q converges weakly if n pf q pf q for any bounded, continuous f : X R. This concept carries over to locally nite measures with the additional demand that f has bounded support.
13
1. Point Processes Denition 1.3 (Vague convergence). A sequence pn qn of locally nite measure converges vaguely to if for any continuous f with bounded support n pf q pf q
BpX q dene the evaluation mapping B as B : MpX q R {V}, B : pB q. (1.1) Later B is only considered as a mapping on M pX q, and therefore takes values in N {V}. The next proposition shows their fundamental role,
For B see [DVJ08b, prop. 9.1.IV]. Proposition 1.4. Let X be a c.s.m.s. ii) The Borel -algebra B MpX q is the smallest -algebra on MpX q generated by the mappings {B }B B0 pxq . i) MpX q is a c.s.m.s. when endowed with the vague topology.
iii) M pX q is a c.s.m.s. under the vague topology and its Borel sets agree with the ones inherited from MpX q. In particular the last statement follows directly from the following lemma, see [DVJ08b, lemma 9.1.V]. Lemma 1.5. M pX q is a closed subset of MpX q. x pAq : 1 0 if x A . if x A Let x be the Dirac measure, that is for A B pX q
The next proposition [DVJ08b, prop. 9.1.III] shows that measures MpX q decompose into an atomic and a diuse part, i.e. the former having non-negative masses on singletons and the latter not. Point measures are particular examples of purely atomic measures and permit a representation as a sum of Dirac measures. Simple point measures relate locally nite sets of points of X and point measures of X ; in fact this correspondence is one-to-one. General point measures M pX q allow multiple points.
14
i) obeys a unique decomposition into a d with a purely atomic part a , which permits a representation a
i
ki xi ,
where the ki s are positive real numbers, pxi qi countable set and d is a diuse measure.
X is an at most
1 for any i. Write x for some M pX q and x X if p{x}q 0, therefore p{x}q 1, and say that x is contained in . Thereby any locally nite
point measure can be represented as
x
ii) If M pX q, then coincides with its atomic part a with additionally the ki s being non-negative integers and pxi qi X has the property that pxi qi B is a nite set for any bounded B B0 pX q.
with the factors p{x}q 1 if and only if is a simple point measure. is called a conguration of elements of X . The basic terms are dened and point processes may now be dened. Denition 1.7 (Random measure, Point process). i) A probability measure on MpX q, B pX q is called a random measure on MpX q, B pX q .
p{x}qx
ii) A probability measure P on M pX q, B pX q is called a point process. iii) A simple point process P is a point process which is concentrated on M pX q, P M pX q 1.
15
1. Point Processes
P pB q :
B dP.
(1.2)
inherits the nite additivity property and monotone convergence property for increasing sequences Bn B from P and therefore is a measure. Denition 1.8 (Intensity measure). Let P be a point process. If MpX q, then P is of rst order and is called the intensity measure of P . Also is named rst moment measure. Since pB q is the expectation of B , pB q is the expected number of points of P inside B whether nite or not. Suppose that the intensity measure exists as a locally nite measure. Let f : X R be a positive and measurable function, then the random integral f : f pxqpdxq
can be constructed in the usual way as limit of simple functions, i.e. linear combinations of B s. Their expectation with respect to a point process P then is P pf q f pxqpdxqP pdq f pxqpdxq (1.3) Related integrals will appear in section 1.2. The postponed basic discussion of the so-called Campbell measure will in fact allow the function f to depend on the conguration . Instead of integrating B , products of the form B1 . . . Bn for not necessarily disjoint, measurable B1 , . . . , Bn may be integrated, pnq : B pX qn
R ,
: B1 Bn
B1 Bn dP.
16
1.1. Point Processes pnq can be extended in the usual way from rectangles to general sets in B pX n q, which yields Denition 1.9 (Higher order moment measures). Let P be a point process and n N. If pnq MpX n q, then P is of n-th order and pnq is called the n-th order moment measure of P . Like moments of random variables extend to moment measures, the characterisation of non-negative random variables by its Laplace transform carries over to the Laplace functional of a point process. Let f : X R be non-negative, measurable and bounded with bounded support. Then the Laplace functional LP of P is LP pf q : P ef
exp
pf q P pdq.
The importance of the Laplace functional is due to the one-to-one correspondence between random measures and functionals which occur as Laplace functionals [DVJ08b, prop. 9.4.II]. Theorem 1.10. Let the functional L be dened for all non-negative, measurable and bounded functions f : X R with bounded support. Then L is the Laplace functional of a random measure P on X if and only if i) for f1 , . . . , fk non-negative, measurable and bounded with bounded support the functional Lk pf1 , . . . , fk ; s1 , . . . , sk q L
k m 1
is the multivariate Laplace transform of random vector pY1 , . . . , Yk q. ii) for every sequence fn monotonously converging to f pointwise Lpfn q Lpf q iii) Lp0q 1 Moreover, if these conditions are satised, the functional L uniquely determines P .
sm fm
17
1. Point Processes
exp 1 ef
for any continuous, non-negative f with bounded support. Write P for this process. Putting f B with B B1 . . . Bk for pairwise disjoint and bounded B1 , . . . , Bk , we get that the family B1 , . . . , Bk is mutually independent with each Bm having Poisson distribution with intensity pBm q. The independence property of P is known as complete randomness, in particular Poisson processes are prime examples for completely random measures. An important representation theorem was given by Kingman [Kin67], also in [DVJ08b, thm. 10.1.III]. Theorem 1.12 (Kingman). The log-Laplace functional of a completely random measure is of the form
log LP pf q pf q
V
k 1
where MpX q is a xed, non-atomic measure, pxk qk enumerates an at most countable, locally nite set of atoms of P , pk qk is a family of log-Laplace transforms of positive random variables and is the intensity measure of a Poisson process on X R which satises for every 0 the integrability conditions for each B B0 pX q, pB, p, Vqq V
0
k f pxk q
(1.4)
u pB, duq V.
Conversely, each log-Laplace functional satisfying equation (1.4) denes a completely random measure.
18
1.1. Point Processes Thus completely random measures may consist of a xed, non-atomic part , an atomic part at the sites pxk qk with random weights and a compound Poisson part independent of the atomic part. For a completely random measure to be a point process, needs to vanish, each k needs to be the log-Laplace transform of an integer-valued random variable and pB q needs to be a measure on the positive integers. Additionally, to be Poisson, k psq uk p1 es q and d 1 . The intensity measure of a Poisson process P then is d k uk xk , which is exactly the decomposition in theorem 1.6 of into its diuse and atomic part, respectively. Sums of independently Poisson distributed random variables again have a Poisson distribution, and the latter intensity is exactly the sum of the former ones. Vice versa, each Poisson random variable can be represented as a sum of any given number of independently, identically Poisson distributed random variables. This property is known as innite divisibility. Consequently, a point process, or more general a random measure, is innitely divisible, if it can be represented as a superposition of k independent, identically distributed point processes or random measures for any k . For a Poisson process with intensity choose k for a given non-negative integer k to obtain the representation of the innitely divisible Poisson process, see e.g. [DVJ08b, thm. 10.2.IX]. Theorem 1.13 (L evy-Khinchin representation). A random measure P is innitely divisible if and only if its log-Laplace functional permits a representation
log LP pf q pf q
1 exp
pf q
pd q
(1.5)
where MpX q and is a -nite measure on MpX qz{r} such that for every B B0 pX q the integrability condition
R
1 pduq V 1 eu B
is satised. Conversely, such measures and dene via equation (1.5) an innitely divisible random measure.
19
1. Point Processes In the point process case, necessarily 0 and is concentrated on M pX q. Such innitely divisible random measures play an important role in limit theorems. Now assume that in addition X is an Abelian group with the commutative group operation . There exists a measure MpX q which is invariant under the group action, i.e. pA xq pAq for every A B pX q, x X and A x : {a x : a A}. is called Haar measure and is uniquely determined up to a positive constant, see e.g. Stroppel [Str06, thm. 12.23]. Every x X induces an automorphism Tx on M pX q by Tx pB q : pB xq. Because of Txy Tx Ty , the set of automorphisms T {Tx }xX is an Abelian group itself and in a natural way homeomorphic to X . A point process P is T -invariant, if the action of T conserves the distribution, P pTx Aq P pAq for every x X . In case of T being a translation group, P is also called stationary. Since a point process P is determined by its Laplace functional LP , LP must be stationary itself. Particularly the log-Laplace functional of a Poisson process satises
log LP
ef p xq
1 ef p xq 1 ef
and since the measure is uniquely determined by the set of continuous functions and is stationary, can only be a multiple of the Haar measure on X, B pX q . Therefore a Poisson process is stationary if and only if its intensity measure is a multiple of the Haar measure on X . More generally, T {Ts }sG can be an Abelian group dening transformations on X , where G is a complete, separable metric group, which is locally compact. At least partial results carry over to this more general case and a factorisation theorem 1.14 below states that a T -invariant measure
20
1.1. Point Processes decomposes into a measure which is a multiple of a Haar measure on G and a second measure on some other space, see e.g. [DVJ08a, prop. A2.7.III]. Proposition 1.14 (Factorisation). Let X be a c.s.m.s, T {Ts }sG a complete, separable, locally compact metric group of transformations acting measurably on X . Furthermore suppose that there exists a one-to-one, both ways measurable and bounded sets conserving mapping : G Y X with some c.s.m.s. Y , which preserves the shifts Tg in the sense that Tg ph, y q pg h, y q. Then any T -invariant measure MpX q can be represented as pf q
Y G
f pg, y q
pdgqpdyq,
where is the Haar measure and is up to a constant a unique measure on Y for measurable, non-negative functions f . Such a situation will occur in section 3.1, where a group of translations acts on a space of functions and a translation invariant measure is decomposed into such two parts. A further application will be in section 5.1 the disintegration of the so-called Campbell measure of a stationary point process with respect to its intensity measure. An important question are criteria for the simplicity of point processes, particularly of Poisson processes. A rst characterisation involves the second order moment measure p2q of a point process P , see e.g. [DVJ08b, prop. 9.5.II]. Let the diagonal of a set A B pX q be diag Ak : {px1 , . . . , xk q X k : x1
. . . xk A}.
Proposition 1.15. A point process P of second order satises p2q pdiag B 2 q pB q for all B B0pX q with equality if and only if P is simple.
21
1. Point Processes Since the Poisson process P is completely random, p2q pB1 B2 q
pB zB B B qpB zB B B qdP pB1zB2qpB2zB1q pB1 B2q pB1zB2q pB2zB1q 2 B B pB1 B2q 2dP pB1 B2 q
B1 B2 dP
1 2 1 2 2 2 1 2 1 2
Hence the second order moment measure of P for a product is the product of the intensities plus the variance on the common part of these two sets. Particularly on products of disjoint sets the last term vanishes. The following concept of orderliness also addresses the multiplicity of points of a point process and particularly for Poisson processes. A point process P is orderly, if the probability of nding many points in a sequence of shrinking spheres vanishes suciently fast compared to the probability of nding some point, P S pxq
1 o
P S pxq
as 0 for every x X ,
where S pxq is the sphere with centre x and radius . For a Poisson process P
1 1 e S pxq Spxq e S pxq P S pxq 0 1 e S pxq vanishes as 0 if and only if x is not an atom of , that is p{x}q 0.
P S pxq
Theorem 1.16. A Poisson process P is orderly if and only if P is a simple point process. Even more, orderliness and complete randomness together characterise Poisson processes, see [DVJ08b, thm. 2.4.V]:
22
1.2. The Campbell Measure of a Point Process Theorem 1.17. A point process P without xed atoms is a Poisson process if and only if P is orderly and completely random. Directly from the denition of the Poisson process and the subsequent discussion we get a local representation of a Poisson process P , i.e. if is EB -measurable for some B B0 pX q, a Poisson process P can be written as
pBq P ,B pq e
V 1
n 0
n!
Bn
px1
. . . x qpdx1q . . . pdxnq,
n
hence can be interpreted as rst choosing an integer n according to a Poisson distribution with intensity and then placing n points independently according to in .
px, q X MpX q : p{x}q 0 , then according to [KMM74, prop. 2.5.1], C is B pX q B MpX q -measurC able, and [KMM74, thm. 2.5.2] states Theorem and Denition 1.18 (Campbell measure). For any integrable or non-negative function h : X MpX q R the mapping hpx, qpdxq
is measurable and the Campbell measure CP of a point process P on X is given by CP phq hpx, qpdxqP pdq.
23
1. Point Processes A characterisation of measures which may occur as a Campbell measure of a random measure is given by Wegmann [Weg77]. However, such a characterisation is not needed in this work. A special case of such integrals occurred in section 1.1.2 as the expectation of random integrals h with h only depending on x. Here even the dependence on the whole conguration is allowed. Recall also that in choosing hA px, q 1A pxq, CP php q q reduces to the intensity measure of the point process P . Hence the Campbell measure is an extension of the intensity measure. ! of a point process Closely related is the reduced Campbell measure CP P . Instead of attaching the whole conguration to x , is reduced beforehand by x , i.e. x is attached to x. Denition 1.19 (Reduced Campbell measure). The reduced Campbell mea! of a point process P on X is the measure on X M pX q given sure CP by
! CP phq
h 0 measurable.
24
CP phq
hpx, qP x pdqpdxq
for non-negative or CP -integrable h. The disintegration result of the Campbell measure with respect to the intensity measure leads to the interpretation that P x is the original process P conditioned on the event that there is at least a point at x, i.e. conditioned on the event {x} 0 . In the special case of the Poisson process these Palm kernels take a simple form and moreover characterise the Poisson process uniquely. Similar characterisations can be shown for a larger class of processes. Theorem 1.21 (Meckes characterisation of the Poisson process). There is exactly one point process P satisfying for any measurable, non-negative h CP phq P hpx, x qP pdqpdxq. (1.6)
(1.7)
i.e. the Poisson process P is the unique solution of the functional equation
! CP
P. P
x
(1.8)
This theorem was rstly given in a general form by Mecke [Mec67], see also [DVJ08b, prop. 13.1.VII], and leads to the interpretation that the local Palm distribution P x of the Poisson process P is the Poisson process with an additional point at x. Later this characterisation was generalised e.g. by Nguyen, Zessin [NZ79] to Gibbs processes.
25
1. Point Processes
pIq
holds.
! CP pB
q 3 P
dB B0
! pB q 3 P for each Condition pI q ensures the absolute continuity CP B B0 pX q and therefore the Radon-Nikod ym derivative can be computed [Kal78].
Theorem and Denition 1.23 (Papangelou kernel). Let the point process P satisfy pI q, then there exists a measurable mapping : M pX q MpX q, such that p,
! pB q dCP p , B q dP for B B0 pX q. Since the paper [MWM79] has been called Papangelou kernel for the point process P .
By the Radon-Nikod ym theorem the Papangelou kernel for P is P -a.s. unique. As in the previous subsection the Palm kernel is interpreted as the point process conditioned on the occurrence of a point at some site, the
26
1.2. The Campbell Measure of a Point Process Papangelou kernel is interpreted as the conditional intensity measure (on X ) conditioned on a given conguration. The following theorem from [MWM79] relates the Papangelou kernel with a partial integration formula for CP . Theorem 1.24 (Partial Integration). Let P be a point process and be a measurable mapping M pX q MpX q. Then the following statements are equivalent: i) is a Papangelou kernel for P ii) P satises the partial integration formula for non-negative, measurable h CP phq hpx, qpdxqP pdq hpx, x q p, dxqP pdq
By Matthes et al. [MWM79] the last equivalence leads to a nice characterisation of simple point processes involving the Papangelou kernel. Corollary 1.25 (Simplicity). Let be a Papangelou kernel for the point process P , then the mapping p, supp q is measurable and the simplicity of P is equivalent to P { p, supp q 0} 0. This can be seen by setting h the indicator on pairs px, q for which pxq 1. For Poisson processes P this corollary implies the known fact that P is a simple point process if and only if the intensity measure , which is a Papangelou kernel for P , is a diuse measure.
27
1. Point Processes , construct a point process P , such that is a Papangelou kernel for P . Throughout this subsection let : M pX q MpX q be a measurable mapping. Let us rstly derive some properties of the following iterated kernels. Denition 1.26 (Iterated kernel). For a Papangelou kernel : M pX q MpX q and m N let the iterated kernel pmq , m 1, of be pmq p, q : Set p0q : 1. px1
(1.10)
The mapping pmq is measurable, and moreover is nite on rectangles of bounded sets and for m 2 even a symmetric measure [MWM79]. Theorem 1.27. Let be a Papangelou kernel for a point process P . Then for every m 1 P pmq p , B1 . . . Bm q V
dB1, . . . , Bm B0pX q 1 1
and for m 2
for every permutation on {1, . . . , m} and bounded, measurable sets B1 , . . . Bm . In order to construct a point process P for which the measurable mapping : M pX q MpX q is a Papangelou kernel, the kernels pmq given by equation (1.10) at least need to satisfy the properties of the previous theorem P -a.s. Particularly the symmetry of is assumed for all , which is equivalent to the cocyle condition,
28
1.2. The Campbell Measure of a Point Process Lemma 1.28 (Cocycle condition). Let pmq be given by equation (1.10) for some measurable mapping : M pX q MpX q. Then pmq p, q is a symmetric measure for each m N, i.e. pmq , B1 . . . Bm
pmq
, Bp1q . . . Bpmq
for every permutation on {1, . . . , m} if and only if the cocycle condition holds p x , dy q p, dxq p y , dxq p, dy q. Proof. Assume rstly that pmq p, Then by choosing m 2,
p2qp, B2 B1q
for all bounded, measurable B1 , B2 . Hence the cocycle condition holds. Secondly assume that the cocyle condition holds for , then pmq p, B1 Bm q
. . . x , dx2q p x , dx2qp, dx1q 1B px1q1B px2q 1B pxmq p x . . . x , dx2 q p x , dx1 q p, dx2 q pmqp, B2 B1 Bmq
m 1 1 1 2 m 1 m 1 2
p x1
due to the cocycle condition. This equation holds for every next neighbor transposition and hence by iteration for all transpositions. Since every permutation has a decomposition into transpositions, the equation holds for all permutations, therefore pmq p, q is a symmetric measure.
29
1. Point Processes In the previous subsection the Papangelou kernel was obtained as the disintegration of the reduced Campbell measure permitting the interpretation that p, q is the conditional intensity measure conditioned on the conguration . Thus in the following plays the role of a boundary condition, for which the point process is constructed. Let M pX q and Z pmq pq be the mass of the iterated kernel pmq as well as pq the possibly innite limit of the series, Z pmq pq : pmq p, X
. . . X q,
pq :
m 0
is called integrable if pq V for each M pX q, and in this case the point process P given by P pq : 1 pq
Z pmq pq . m!
(1.11)
m 0
1 m!
Xm
(1.12)
is well-dened. By [Zes09]: Proposition 1.29. Let : M pX q Mf be a nite kernel satisfying the cocycle condition and assume to be integrable. Then for every boundary conguration M pX q, P is a solution of the partial integration formula CP phq hpx, x q p , dxqP pd q.
Denition 1.30 (Papangelou process for kernel ). P is the (nite) Papangelou process for the symmetric and integrable kernel . Proof of proposition 1.29. By the denition of the Campbell measure CP phq hpx, q pdxqP pd q
m Xm j 1
1 p1q m ! m0
hpxj , x1
. . . x q
m
pmq
, dx1 , . . . , dxm ,
30
1.2. The Campbell Measure of a Point Process then rstly by the symmetry of pmq and secondly by integration and the denition of the iterated kernels 1 p1q pm 1q! m0 hpxm , x1
Xm
. . . x q
m
1 p1q pm hpxm , x . . . x q 1q! X X m0 p x . . . x , dxm q pm1q , dx1 , . . . , dxm hpx, xqp , dxqP pd q.
1 m 1
The aim is to extend the construction of nite Papangelou processes for nite kernels to kernels which are -nite. The strategy is to construct the process locally and then to glue these locally dened processes together. For a bounded and measurable B let the -algebra of the events inside B be EB B I : B I B, B I B0 pX q and dene the restriction to B for EB -measurable as B p, q :
m 0
1 m!
Bm
Note that the mass of in B is cut. The denitions of the normalisation pmq constants ZB pq and B pq in equation (1.11) carry over directly. is called locally integrable if for each bounded B and M pX q, B pq is nite. A construction similar to the Poisson process can be used in case of additional measurability conditions on B , that is if B cannot see what happens outside B . Theorem 1.31 (Papangelou processes with independent increments). Let the measurable mapping : M pX q MpX q be locally integrable and satisfy the cocycle condition. If in addition B dened in equation (1.13)
31
1. Point Processes is EB -measurable, then there exists a point process P on X which is independent of the boundary conguration M pX q and for which is a Papangelou kernel, i.e. P satises the partial integration formula, CP phq hpx, x q p, dxqP pd q.
Proof. Two main steps have to be done: Firstly P has to be constructed and secondly the partial integration formula for P has to be shown. For the rst part let pBn qn0 be a locally nite partition of X of bounded sets, i.e. for each bounded set B B0 pX q, B Bn $ r only for nitely many n. Then is constructed according to equation (1.13), on each Bn a point process Pn
Pn pq :
Bn pq
Bn p, q.
m Bn
. . . x qpmq
m m
c , dx1 , . . . , dxm B n
m Bn
px1
. . . x qpmq
0, dx1 , . . . , dxm
may be dropped. is independent of and the superscript of Pn Thus for each n a point process Pn on M pBn q is constructed. Let
N :
n 0
M pBn q
the product space. By the Daniell-Kolmogorov extension theorem, there exists a probability measure P on the product space N with nite-dimensional distributions given by the corresponding product of the Pn s. Finally map P via N M pX q, pnqn1 n
n 0
32
1.2. The Campbell Measure of a Point Process to obtain a probability measure on M pX q, which also will be denoted by P with abuse of notation. It remains to show that is the Papangelou kernel for P . First of all, let h be of the form hpx, q g pxqpq with supp g Bj for some j 1 and being EB0 ...Bm -measurable for some m j . Then CP phq
V
n 0
h x, g pxq
k 0
since supp g able,
V
Bj
k 0
Bj
g pxq
m k 0
k j pdxqP pdq,
for which P can now be replaced by P1 Pm . Finally the application of partial integration for Pj yields
Bj
m k 0 m k 0 m k 0
Bj
Bj
k , dx P0 pd0 q Pm pdm q
33
1. Point Processes Particularly in the step of the application of the partial integration formula the measurability condition on B , and hence the independence simplied a lot. For general , that is B not necessarily EB -measurable, ner instruments are necessary. In fact, proposition 1.31 is a special case of proposition 1.33. In [Zes09] the theorem is only given for the boundary conguration 0, here we drop this restriction. Furthermore an additional assumption on the normalisation constants seems to be required in contrast to [Zes09]. The main schedule, rstly to construct the global process and secondly to show the partial integration formula, stays the same. The means of theorem 1.31 have to be rened: a Markov construction together with the theorem of Ionescu Tulcea [Kal02, thm. 6.17] yields the rst part, for the second further assumptions are necessary. Theorem 1.32 (Ionescu Tulcea). For any measurable spaces pSn , Sn q and probability kernels n from S1 Sn1 to Sn , n N, there exist some d random elements n in Sn , n N, such that p1 , . . . , n q 1 n for all n. A kernel is said to satisfy the Feller condition if for every increasing sequence pBn qn of bounded sets which exhausts X , pBn ,
q p, q
vaguely as n V. A further condition needs to be discussed: If p y , q is absolutely continuous with respect to p, q outside {y } for every boundary conguration M pX q, 1{y}c pxq p y , dxq 1{y}c pxqf py, xq p, dxq, then for bounded, measurable B and x B , B p x , q 1 m! (1.14)
m 0
Bm
px1
. . . x B x q
m c
pmq
B c
x, dx1, . . . , dxm
34
m 0
1 m!
Bm
px1
. . . x B xq
m c
Theorem 1.33 (General Papangelou processes). Assume that the measurable mapping : M pX q MpX q satises the cocycle condition and the Feller condition, and is locally integrable. Let M pX q be a given boundary conguration. If furthermore p y , q is absolutely continuous with respect to p, q outside {y } as in equation (1.14), and for each B B0 pX q, x B c the normalisation constants satisfy B pq B p x q, then there exists a point process P on X for which is a Papangelou kernel, i.e. P satises the partial integration formula CP phq hpx, x q p, dxqP pd q.
Proof. As in the proof of proposition 1.31 let pBn qn0 be a locally nite partition of X of bounded sets. Then the following nite point processes exist by equation (1.12)
Pn pq
n pq
(1.16) Successively on each Bn a nite point process will be constructed with the boundary condition in the regions B0 , . . . , Bn1 replaced by a realisation of the corresponding, already constructed, nite point processes on B0 , . . . , Bn1 . Let Q0 pd0 q : P0 pd0q,
m 0
1 m!
m Bn
px1
. . . n qpmq
m
c , dx1 , . . . , dxm . B n
35
1. Point Processes the dependence on suppressed for the moment. Furthermore denote by m the sum 0 . . . m and by pmq the restriction of to the complement of B0 . . . Bm . For p0 , . . . , m1 q M pB0 q M pBm1 q let
pm1q m1 Qm p0 , . . . , m1 ; dm q : Pm pdmq,
(1.17)
i.e. Qm is a probability kernel M pB0 q M pBm1 q M pBm q. Since by equation (1.16) any contribution of inside Bm is cut, in the denition (1.17) pm1q may be replaced by pmq . The choice pm1q is consistent with the denition of Q0 . By the theorem of Ionescu Tulcea there exists a probability measure P on N such that its nite-dimensional distributions Pp 0,...,nq are given by
pn1q n1 Pp pdnq P1p0q pd1qP0pd0q. 0,...,nq pd0 , . . . , dn q Pn
For simplicity P again is identied with its image under the mapping N
MpX q,
pnqn0
n 0
n .
Therefore the point process P exists, considered either as a point process on N or M pX q, respectively. The partial integration formula remains to be shown. In choosing again h to be of the form hpx, q g pxqpq with supp g Bj for some j 1 and being EB0 ...Bm -measurable for some m j , the rst lines of the proof of the partial integration in the proof of proposition 1.31 can be followed, and then continued by CP phq
pm1q m1 g pxq m j pdxqPm pdmq
Bj
P1p q pd1qP0pd0q,
0
36
1.2. The Campbell Measure of a Point Process for which the partial integration formula for nite Papangelou processes can be applied for the j -th kernel
Bj
Because of x Bj 1 , by equation (1.15) and the equality of the normalisation constants, and the integrations can be exchanged such that rstly with respect to x is integrated,
Bj
g pxq m x
pm1q m1
m k j 1
f px, k q p j , dxq
For the last line observe that by assumption f px, j 1 q p j , dxq p j 1 , dxq. Indeed, the EB -measurability in proposition 1.31 implies Feller and absolute continuity condition, and therefore proposition 1.31 is a special case of proposition 1.33.
37
2. Deviation Principles
One of the main techniques used in the later parts is the principle of large deviations. Basically, the situation is the following: given a sequence pYk qk1 of identically distributed, uncorrelated random variables with nite second moment, the weak law of large numbers states that the average of the rst 1 Sn tends to the expectation EY1 weakly as n V. Clearly n variables n 1 the probability that the average n Sn stays away from EY1 , tends to 0. Such 1 events {| n Sn EY1 | } are called rare events, untypical events or large deviations. The basic question is: What is the probability of a rare event and at which speed does it vanish? Cram ers theorem (see e.g. Dembo and Zeitouni [DZ98, thm. 2.2.3] or Deuschel and Stroock [DS00, thm. 1.2.6]) states that this probability behaves like exp n inf {I pxq : |x EY1 | } , where I is a non-negative function called rate function. Several observations can be made: Firstly, the 1 probability of this rare event decays exponentially fast, i.e. n Sn may deviate with only exponentially small probability. Secondly, inf {I pxq : x R} 0 1 and moreover, the inmum is in fact a minimum. The limit of the n Sn s occurs as the minimum of I , which is at the same time a zero of I . Finally the theorem is not restricted to only random variables; for random vectors Cram ers theorem is also valid and allows extensions to projective limits (which are not trivial indeed). Particularly the second observation, the determination of a weak limit as the minimiser of an optimisation problem, will be important: the weak limit is exactly the minimiser of I . Conversely, determining the minimiser 1 Sn of I means to nd a weak limit. Moreover, if a condition on the n is present, the large deviation principle leads to an optimisation problem with constraints.
39
2. Deviation Principles In sections 4.5, 7.2 and 7.3, large deviation principles are a basic tool for the derivation of limit theorems. The large deviations for Poisson processes at high intensity, see [GW95], and required adaptations are discussed in section 2.2. For the discussion of the asymptotic behaviour of the model to be introduced in chapter 3 large and small deviations for Brownian bridges are given in subsection 2.3.
40
BpY q closed:
lim inf
n
V |an | log Pn G
1 |an| log Pn F
yinf I py q G yinf I py q. F
(2.1) (2.2)
lim sup
n
nx P euS eunx eunx E euS exp n ux log puq , where puq : log E euY is the logarithmic moment generating function.
P Sn
n n 1
By the lower semicontinuity, see e.g. Deuschel and Stroock [DS00, lemma 2.1.1], the rate function can be shown to be unique in case of existence once the speed is xed. Therefore I is said to govern the large deviations of pPnqn. The main job is to determine the rate function I . Consider Pn to be the law of Sn as in the example at the beginning, then by the Markov inequality
Optimising the exponent on the rhs. with respect to u yields a candidate for the rate function I governing the large deviations of pSn qn . Indeed, basically Cram ers theorem states that the convex conjugate of is the rate function, and the speed can be chosen to be an n. Furthermore I can be shown to have exactly one minimiser given by EY1 , which is simultaneously the unique zero. Roughly speaking, Sn satisfying a large deviation principle with rate function I means P Sn A en inf xA I pxq . Such a statement remains true in greater generality for measures on a space Y . Let Y be the dual of Y , then for a probability measure P let P be its logarithmic moment generating function P : Y I pxq pxq : sup{ux log puq : u R}.
P puq : log
where u, y : upy q.
exp u, y P pdy q,
41
2. Deviation Principles Denition 2.3 (Fenchel-Legendre transform). For a sequence pPn qn of probability measures and an increasing sequence pan qn of positive real numbers let 1 puq : lim Pn pu{an q nV an The Fenchel-Legendre transform of is the convex conjugate of , py q : sup{ u, y
puq : u Y }.
In the following we need to assume that is well-dened and nite in an open set containing 0, G ateaux-dierentiable and lower semicontinuous. The rate function then is exactly , see e.g. [DZ98, thm. 4.5.27]. Theorem 2.4 (G artner-Ellis). Let pPn qn be an exponentially tight sequence of probability measures. If exists in a neighborhood of 0, is G ateauxdierentiable and lower semicontinuous, then pPn qn satises a large deviation principle with good rate function .
42
2.2. Large Deviations for Poisson Processes at increasing Intensity Theorem 2.5 (Large deviation principle for Poisson processes at high in tensity). As r V, P satises a large deviation principle on r r MpX q with speed r and rate function I p ; q : MpX q r0, Vs, I p; q
pf log f
f 1q
if 3 , f : otherwise
d d , f
log f
f 1 L1p q .
The function I p ; q is called relative entropy with respect to and agrees with . Because of some necessary comments on that result, the main points of the proof are demonstrated Sketch of Proof. The moment generating function for the Poisson process was given in section 1.1.3, and therefore passing to the limit yields pf q 1 ef d (2.3)
for any continuous f with bounded support. Moreover, is G ateauxdierentiable in a neighborhood of f with dpf qrg s : d pf dt
tgq|t0
g ef d.
X
Therefore P satises a large deviation principle with rate funcr r tion , which can be identied as the relative entropy with respect to by solving the variational principle.
Remark 2.6. Instead of the intensity measure r for an increasing factor r a n sequence pn qn of intensity measures with n is sucient for the limit in equation (2.3) to exist, which is exactly the situation in section 4.5. Particularly the case X N is important for sections 4.5, 7.2 and 7.3, where the intensity measure is a nite measure of the form
j 1
zj j j
(2.4)
43
2. Deviation Principles for certain parameters z p0, 1s and 0. There even stronger results hold and are required: In section 4.5, z 1 and 2, the test functions are allowed to have an unbounded support, but stay bounded; and in chapter 7, 1 and z 1, the test functions are allowed to grow linearly. These stronger results have to be justied. In the rst case, pj q j for some 2, for f and g bounded the following two estimates hold pf q dpf qrg s
j 1
j ef pj q 1 j g pj q ef pj q
ec pNq V
1 1
j 1
for constants c1 |f | and c2 |g |. The rst estimate ensures that the domain of contains an open neighborhood of 0, the second ensures the dierentiability. Corollary 2.7 (Large Deviations for weak topology). The large deviation principle of the family of Poisson processes P r , r 0, on N with given by equation (2.4) for 2 and z 1 in theorem 2.5 holds true on MpNq equipped with the topology of weak convergence. For the second case, pj q then pf q dpf qrg s
zj j ,
c2 ec pNq V
let |f pj q| c1 p1 j q and |g pj q| c2 p1 j q,
j 1
z j f pj q e 1 j zj j g pj q ef pj q
j 1
j 1
c2
j 1
For suciently small c1 , the rhs. of the rst equation converges, hence the domain of contains an open neighborhood of 0, and the second estimate yields the dierentiability of in the domain of its convergence. Let the -topology be the topology on MpNq generated by these at most linearly growning functions.
44
2.3. Deviations for Brownian Motions and Brownian Bridges Corollary 2.8 (Large Deviations for -topology). The large deviation principle of the family of Poisson processes P r , r 0, on N with given by equation (2.4) for 1 and z 1 in theorem 2.5 holds true on MpNq equipped with the topology of -convergence.
W where I py q y
2 2
Schilders theorem is a particular case of a large deviation principle for general centered Gaussian measures on separable, real Banach spaces [DS00, thm. 3.4.12] Denition 2.10 (Wiener quadruple). pE, H, S, P q is a Wiener quadruple if i) E is a separable, real Banach space, ii) H is a separable, real Hilbert space,
45
for all E , where S : E H is the adjoint map of S . Theorem 2.11 (Large Deviations of centered Gaussian processes). If P is a centered Gaussian measure on the separable, real Banach space E , then there exist a separable, real Hilbert space H and a continuous, linear injection S : H E such that pE, H, S, P q is a Wiener quadruple. Moreover, if pE, H, S, P q is any Wiener quadruple, then S is a compact map, satises S
1 2
2 H
x
E
2 E
pdxq
1 2
and the family ppqqn1 satises a large deviation principle with rate function 1 S 1 x 2 if x S pH q H pxq 2 V if x E \S pH q For the the Wiener measure W T on r0, T s, E {x C r0, T s, Rd : xp0q 0}, H is the space of absolutely continuous functions whose derivatives L2 -norm is bounded and S is given by Sh ptq S is computed to be S ptq from which the covariance follows as Qp, I q : S , S I
t 0
hpsqds.
T
pdsq,
T 0
T 0
s t I pdsqpdtq.
(2.5)
46
2.3. Deviations for Brownian Motions and Brownian Bridges In case of the Brownian bridge on r0, T s, E gets the additional condition xpT q 0, H gets the condition that the integral over r0, T s vanishes, and the kernel for the covariance in equation (2.5) is replaced by s t st T. 1 Particularly S is still the derivative. Therefore, if B pBt qtr01,s is a Brownian bridge, then the probability of the event {suptr0,1s |Bt | L} vanishes exponentially fast as L V. More precisely 1 lim log P sup |Bt | L 2, LV L2 tr0,1s and Brownian motions and bridges are very unlikely to leave a large region at least once. On the other hand both processes are very likely to leave very small regions. The precise statements about the small ball probabilities are given by Li and Shao [LS01, thm. 6.3]. Theorem 2.12 (Small deviations of Brownian bridges; Shao, Li). Let denote by pBt qtr0,1s a Brownian bridge. Then
lim 2 log Pp Bt
V q 8 .
The deviation results are going to be used in the section 3.1 to get further insight into the behaviour of the Brownian loop measure, which is going to be constructed there.
47
Part II.
This chapter is devoted to the construction of the ideal Bose gas. The initial point is the already mentioned work of Ginibre [Gin71], where inter alia a Feynman-Kac representation of a Bose gas is derived. His results restricted to the non-interacting case are interpreted in terms of point processes, which leads to a Poisson process on a space of loops. For the construction of the Poisson process the construction of the measurable space of composite loops X, B pX q together with a ring of bounded sets B0 pX q and a locally nite measure is sucient. The space of loops is constructed in denition 3.1, followed by the Borel- -algebra and the ring of bounded Borel sets in denition 3.3. Subsequently the intensity measure is constructed. is shown to be invariant under a group of translations isomorphic to Rd and therefore permits a disintegration with respect to the Lebesque measure. Two important consequences are its local boundedness, lemma 3.6, and the absence of atoms, lemma 3.7. The section concludes with an application of the deviations for Brownian bridges of section 2.3 giving a deeper insight into the behaviour of . Particularly the weight of for long loops in a xed region drops exponentially fast, lemma 3.9, and the weight for short loops is exponentially close to the volume of that region, lemma 3.10.
51
j 1
Xj
The image of a j -loop x in Rd represents j simultaneously moving particles starting at xpk q, k 0, . . . , j 1 and changing its positions during a time interval of length . x rk, pk 1q s is the trace of a single particle or elementary component. Each of the spaces of j -loops is endowed with the Borel topology B pXj q, and X is endowed with the corresponding disjoint union topology, that is the
52
3.1. The Loop Space and the Brownian Loop Measure nest topology such that the canonical injections Xj Let B pX q denote this topology on X .
X are continuous.
where Bj
j 1 Bj ,
BpXj q
The pre-image of an open set of any canonical injection Xj X is always open and further sets cannot be added to X keeping the injections continuous. B pX q is much ner than the product topology, which is generated by the canonical projections X Xj . The latter only admits sets of the form of lemma 3.2, where all but a nite number of Bj s is allowed to dier from Xj . However, the -algebras generated by both topologies agree due to the countability of the index set. Let B0 pRd q be the ring of bounded Borel sets of Rd , which is a partially ordered set when endowed with the inclusion B0 pRd q, . Denition 3.3 (Bounded sets). For B0 pRd q dene the set of bounded sets of X to be B0
B0pX q
Therefore a loop x is contained in some region , whenever the image of the loop is fully contained in , for which x is written; a set of loops is bounded, whenever there exists some bounded region , which contains these loops. Clearly, if 1 , 2 B0 pRd q are two disjoint bounded regions, then X1 X2 X1 2 without equality in general, since loops may start in one region and cross the other one. For s Rd let Ts : X X, Ts x x s
be the shift of a loop x by s and T {Ts }sRd . T is a translation group acting on the space of loops X , each Ts shifting loops as a whole by s Rd .
53
3. Construction of the ideal Bose Gas Thus the space of loops and a ring of bounded sets B pX q is constructed. A Poisson process on pMpX q, B pX q, B0 pX qq is dened by its intensity measure, whose construction is the next aim. Let be the density of the centered normal distribution on Rd with covariance matrix I and consider on pRd qj the measure j pdaq pa1 a0 q . . . paj 1 aj 2 q pa0 aj 1 qda0 . . . daj 1 . With abuse of notation let T be the corresponding translation group on
Lemma 3.4 (Disintegration of ). j permits a disintegration a0 pda1 , . . . , daj 1 qda0 , (3.1) f pa0 ; a1 . . . , aj 1 q j,
a0 is the distribution of a random walk bridge of length j starting where j, at a0 and having normally distributed steps. Proof. j pf q f pa0 , . . . , aj 1 q pa1 a0 q . . . paj 1 aj 2 q
p2 qd p2j qd{2
pa0 aj1qda0 . . . daj1 pa1 a0q2 pa0 aj1q2 f pa0 ; a1 . . . , aj 1 q exp 2 2 pa2 a1q . . . paj1 aj2qda1 . . . daj1da0 a pda1 , . . . , daj 1 qda0 f pa0 ; a1 . . . , aj 1 q j,
0
a0 the probability measure of a random walk starting at a0 with with j, normally distributed steps conditioned on returning at the j -th step to the starting point. In choosing f only depending on a0 , one gets the prefactor due to j convolutions of normal distributions.
54
a2 a1 a5
a5
a2 a1
a0
a3
a0
a3
a4
a4
(a) points in Rd
Figure 3.1.: Construction of a 6-loop Therefore j acts in the following way: The rst component is weighted according to a multiple, which depends on j , of the Lebesgue measure on Rd , and the remaining ones are weighted according to a random walk a0 . Figure 3.1(a) bridge of j steps given by the probability measure j, shows such an arrangement of some points. In the next step these points are connected in the corresponding order in a way such that the resulting j measure is translation invariant. Hence j , which acts on Rd , is lifted to a measure j on the space of j -loops Xj such that the translation invariance remains true for j . More precisely, let p : Xj pRd qj be the projection x xp0q, xp q, . . . , xppj 1q q and j be a T -invariant measure on Xj such that j p1 j . Figure 3.1(b) shows the projection of the linear interpolation of the points of gure 3.1(a) in Rd . A further possibility is to choose Brownian bridges instead of the linear interpolation. Any T -invariant choice of j satises automatically a representation analogue to equation (3.1), therefore to each
55
j pf q p2j qd{2
0 f pxqa j, pdxqda0 .
a0 the disii) j is called the random walk loop measure on Xj and j, tribution of a random walk bridge of j steps starting at a0 obtained from the disintegration of j , j pf q p2j qd{2 a0 pdxqda0 . f pxq j,
In the sequel the constructions can be carried out with j as well as with its bared version. However, most of the results do not depend on the choice, particularly the ones of chapter 4. Whenever dierences occur or only one of these measures is used, an explicit hint will be given. Each of these measures j will be, up to a constant, the intensity measure of a Poisson process on X , and the superposition of these will lead to a Poisson process on X with intensity z :
j 1
where the parameter z p0, 1s is the fugacity. The latter Poisson process is well-dened, if the intensity measure z is -nite.
zj j , j
(3.2)
56
3.1. The Loop Space and the Brownian Loop Measure Lemma 3.6. For any z measure on X .
Proof. Basically the -niteness is directly concluded from the disintegration lemma 3.4, that is for every j N j 1 s1 p2j qd{2
holds with s : X Rd being the projection on the initial point of a loop, s : x xp0q and denoting the Lebesgue measure on Rd . Hence z s1
p2 qd{2g1 pzq,
d 2
(3.3)
j 1
The claim follows from the niteness of g on r0, 1s for every 1. Observe that for 0 1 the series g is only nite on r0, 1q without the right boundary. Furthermore g is strictly increasing and continuous whenever it is nite. A further property of z which follows directly from T -invariance or from the disintegration is the absence of atoms. Since any locally nite measure on Rd with atoms cannot be translation invariant. Lemma 3.7. z has no xed atoms. Lemma 3.6 gives a bound from above for z pX q for bounded , which in fact is really crude since for every bounded region, loops x Xj for very large j hardly stay in and hence do not contribute signicantly. The next lemma shows that nevertheless the bound of lemma 3.6 asymptotically is the best one for sequences of cubes.
57
3. Construction of the ideal Bose Gas Lemma 3.8 (Large cube asymptotics). For the sequence of cubes given by k rk, k sd , j pXk q d{2 lim kV |k | p2j q . Proof. Because of Lemma 3.6 1X pxqj pdxq p2j qd{2 ||, and therefore the limit is bounded from above by the correct value. If k k I k r 2 , 2 s and b : X {x X : xp0q 0} shifts a loop to the origin, b : x x sx, then clearly Xk and 1Xk pxqj pdxq which tends to the desired quantity. In lemma 3.8 loops are xed to a certain length and a statement about increasing regions the loops live in was derived. The next aim is to x a cube with given side length and to get results on the behaviour when varying j . This includes varying j at xed inverse temperature as well as varying at xed length j . Lemma 3.9 (Long loop asymptotics). Fix k rk, k sd . Then the contribution of j pXk q as j V can be estimated by lim sup
j
{x X : sx Ik , bx XI }
k
1 log j 1 log j
2k 2j
d d
j pXk q j pXk q
d , 8k 2 d 8p1 q2 k 2
2
(3.5) (3.6)
lim inf
j
2k 2j
for evey
p0, 1q.
58
2c
2k
Figure 3.2.: Loops in k with their starting points marked. The black loops start inside the small cube and stay completely inside the large one. The grey loops either do not start inside the small cube or leave the large one. Lemma 3.9 gives estimates on the contribution of j pXk q as j V. Two eects can be seen: rstly long loops hardly stay in small regions, which yields the decay on the exponential scale, and secondly the disintegration, from which followed that the mass of loops which start in k is p2j qd{2 times the volume of k . In both cases j may be replaced by , i.e. the estimates also hold as the inverse temperature increases. Proof. By theorem 2.12, 1 log 0 j, j V j lim where C
2 2
x : sup
0 t j
|xiptq| k, i 1, . . . , d C,
(3.7)
59
x : xptq k , t r0, j s
0 j,
x : xptq k , t r0, j s
and therefore j pXk q can be estimated from above by j pXk q p2k qd 0 j, x : xptq k , t r0, j s
and the estimate (3.5) holds. For the lower bound divide k into two parts: a centered inner cube I of side length 2c 2k , where c is chosen later, and an outer part, see gure 3.2. On the outer part we forget about the contribution of the loops and only estimate the contribution in I . Let s : X Rd , sx xp0q be the projection of a loop on its starting point. j pXk q j pXk
Finally choose
p0, 1q and c k to obtain the estimate (3.6). Instead of c k the choice c k for appropriate is also possible with rk, ksd.
(3.8) (3.9)
2k 2j
d d
j pXk q j pXk q
2k 2j
for each
p0, 1q.
60
1
|xiptq| k, i 1, . . . , d
0 t j
0 j,
x : sup
|xiptq| k, i 1, . . . , d
the arguments agree with the ones of the proof of lemma 3.9 with the small deviations replaced by the large deviations of Brownian bridges. Finally we collect the results of this section and dene the Bose gas.
Denition 3.11 (Ideal Bose Gas). The ideal Bose gas with fugacity z z on M pX q for X given in denition 3.1.
Figure 3.3(a) shows a realisation of P z and gure 3.3(b) a realisation of P z . Since z is T -invariant, P z inherits this invariance with each Ts now
61
3. Construction of the ideal Bose Gas shifting complete congurations of loops M pX q. P z is indeed a simple Poisson process since by lemma 3.7 its intensity measure z does not have xed atoms, loops in a conguration occur at most once P z -a.s. zj d On xed cubes k rk, k s , P z realises j -loops with intensity j j , which is close to p2j qd{2 by lemma 3.10 for small j and close to 0 for large j by lemma 3.9.
62
Once C is identied, its structure needs to be clearied. Clearly C is a convex set. If C contains exactly one element, the local characteristic determines this element uniquely. Otherwise a phase transition is said to occur. Due to the convexity, whenever a subset of C is given, further elements can be obtained by convex combinations, hence are the barycentre of that combination. The basic question which follows is if there exists a subset C , such that every P C can be represented uniquely as the barycentre of this subset of extremal points under a certain probability measure. Let pk qk B0 pRd q be an increasing sequence of bounded sets which exhausts Rd , such as an increasing sequence of centered cubes. Furthermore
63
EV :
B0
be the tail- -algebra of E. Since B0 pRd q is directed from above, i.e. for any two bounded sets there exists another bounded set which contains the former two, the intersection over the -algebras for all bounded sets may be restricted to the countable family pk qk while keeping equality. Due to this fact and monotonicity of pEk qk , for every P C p q the limit P p|EV q lim k p , q
k
p q
Rd
(4.3)
exists P -a.s. As outlined in Dynkin [Dyn78], a -algebra EV is sucient for a class M of probability measures, if there exists a probability kernel Q, such that for every P M , P conditioned on EV is given by Q, P p|EV q Q pq P -a.s. (4.4) If even Q M , then EV , following Dynkin [Dyn78], is called H-sucient. Furthermore, if Q is the weak limit of a sequence pQk qk , then the latter sequence is called asymptotically H-sucient. In the given situation for an increasing sequence pk qk of bounded regions, pk qk is indeed an asymptotically H-sucient statistics for the set C p q of stochastic elds. Furthermore there exists a subset of extremal points of C , such that every point process P can be written as the barycentre of this set of extremal points under a probability measure, and the extremal point are exactly those, for which the probability measure is just a Dirac measure. Hence the programme is the following: the rst major step is to determine the limiting kernel Q. Set CV p q the set of limits
k
lim k pk ,
q,
which is, as a measurable space, called the Martin-Dynkin boundary of . The essential part of the Martin-Dynkin boundary will be the set of those
64
CV C , for which the limits Q are P -a.s. constant, Q pAq P pAq P -a.s.pq.
The Martin-Dynkin boundary technique has its origin in the works of Dynkin [Dyn71a, Dyn71b] about general Markov processes. The extension to specications was studied intensely by Preston [Pre79], F ollmer [F ol75] and nally the statistical interpretation with various applications, including F ollmers work, is outlined in Dynkin [Dyn78]. As a consequence, a characterisation of Poisson processes by their local specications was given by Nguyen and Zessin [NZ77]. The aim is to examine the role of E, its tail- -algebra EV and its eect on the corresponding set C of stochastic elds. They are precisely dened in the rst section 4.1. The specications associated to P z and hence the corresponding ensembles are obtained through dierent ways of counting the loops inside a bounded region: in particular {F } associated to the microcanonical loop ensemble counting loops according to each type, {G } associated to the canonical ensemble counting loops without discrimination and {E } yielding the grand canonical loop ensemble. Of special interest will be {H } associated to the canonical ensemle of elementary components counting the elementary components. This is a biased version of the canonical loop ensemble in which every loop gets an additional weight according to its length, but, as will be seen, the behaviour is fundamentally dierent. Its importance is due to the fact that it describes the canonical ensemble of an ideal Bose gas. Starting with P z for a xed z in equation (4.1), the main task is to determine the possible limits Q in equation (4.4), which is done for various loop ensembles in sections 4.2 4.4. By identifying their Laplace functionals, these limits are identied in propositions 4.5 and 4.8 for the microcanonical and the canonical loop ensemble as mixed Poisson processes P Y and P z for the W , respectively, and in proposition 4.10 as the Poisson process P grand canonical ensemble. Therefore only in the latter ensemble no phase transition occurs, and the extremal points are the corresponding Poisson processes, respectively.
65
4. Limit theorems and Extremal Measures In section 4.5 the canonical ensemble is the main subject. Proposition 4.16 identies the limits Q as mixed Poisson processes P Z by means of a principle of large deviations. Consequently, theorem 4.17 shows that the essential part of the Martin-Dynkin boundary of the canonical ensemle consists of the Poisson processes P z with z r0, 1s for d 3 and z r0, 1q for d 1, 2. A major observation is the fact, that the particle density of P Z is always bounded from above by a critical density, which is given explicitly and agrees with the one given in physics literature, see e.g. [Hua87].
d MpX q : Ip, q is a measure, ii) dE E : I p , E q is E -measurable. An E-specication { } is a collection of probability kernels on M pX q E such that i) dA E : p , Aq is E -measurable, ii) dA E : p , Aq 1A , iii) d M pX q : p, M pX qq {0, 1}, iv) d I : I I .
i)
A stochastic eld with respect to the E-specication is a probability measure P on M pX q such that its conditional expectations given the -algebras in the family E are given by the corresponding kernel, P
|E pq p, q
P-a.s.
66
4.1. The Construction of Martin-Dynkin Boundaries Let C C p q denote the set of those stochastic elds. If C contains more than one element, then P is not uniquely dened by the specication and one says that a phase transition occurs. Fix an increasing sequence pk qk in B0 pRd q exhausting Rd and satisfying k int k1 for any k . Furthermore denote by CV CV p q be the set of all limits lim k pk , q (4.5)
kV for sequences pk qk M pX q. CV does not depend on the choice of the family pk qk . Since X is polish, so M pX q and the set of probability measures on M pX q are, and since CV is complete, it is polish when endowed with the induced Borel eld CV . The Martin-Dynkin boundary associated to is the measurable space pCV , CV q. Finally, let Q for any M pX q be the limit
Q : lim k p,
k
q.
(4.6)
Denition 4.1 (Essential part of the Martin-Dynkin boundary). The essential part of the Martin-Dynkin boundary is the set of those P CV C , for which the limit Q is P-almost surely constant, i.e. Q pAq PpAq for P-a.a. . (4.7)
pX,j qj ,
(4.8)
67
4. Limit theorems and Extremal Measures where X,j : Xj X is the set of j -loops which are fully contained in . n is indeed an almost surely nite measure under P z is locally z , since P nite and hence X,j V almost surely for any bounded region . From the denition immediately follows that n nI for each conguration and bounded regions I . Therefore spatial increments can be dened, that is for , I B0 pRd q with I nI , : M pX q M f pNq, nI , : nI
n .
The family of increments denes the outside events and the family of outside events E {E } , E
nI ,
: I
B0pRdq, I, M f pNq
(4.9)
which is the smallest -algebra, such that the increments of the region are measurable. In keeping the terminology of Preston, the stochastic elds corresponding to E form the grand canonical loop ensemble. Adding more detailed information about the interior leads to the family F {F } , F E n : M (4.10) f pNq , which is associated to the microcanonical loop ensemble. For a conguration M pX q let c n pNq be the total number of loops inside and G
E
k
:k
(4.11)
then G {G } denes the canonical loop ensemble. n passes its monotonicity and measurability properties on to c . Finally, much interest lies in what happens if we give dierent weights to loops of dierent lengths, in particular we consider the counting variable N : M pX q N, N
j 1
jn pj q,
(4.12)
68
4.2. The Microcanonical Loop Ensemble which counts the number of elementary components of the loops inside . It is clear that N fullls the same monotonicity and measurablility properties of the increments as c . Let H
E
k
:k
(4.13)
and call the corresponding ensemble H {H } canonical ensemble. In the following sections specications with respect to these decreasing families and their limit points are going to be discussed: In section 4.2 the microcanonical loop ensemble F, in section 4.3 the canonical loop ensemble G, in section 4.4 the grand canonical loop ensemble E and nally in section 4.5 the canonical ensemble H.
z, pj qpj q , pj q! j
z jpj q j pX qpj q j pj q pj q! j
since z pX q z, pNq. Let z, denote the normalisation of the nite measure z, , then the of the probability measures j, , j 1 for some -convolution P z , Mf pNq is dened as
P, :
pj q j j, ,
(4.15)
which represents the superposition of loops of a given length j according to the number pj q. The P z, -combination of that convolution is z , P
M f N
p q
P z, p qP, .
(4.16)
z , is given by a two step mechanism: At rst choose a Accordingly, P composition M f pNq dening the number of loops in some bounded region and then realise a conguration according to this composition. An eect is that the fugacity z does only aect the choice of the composition and not P, . These probability measures are closely related to the ideal Bose gas restricted to bounded sets , P z , . Lemma 4.3. P z, pA|n
q P, pAq.
70
4.2. The Microcanonical Loop Ensemble Proof. Since exactly K j pj q loops are contained in and if they are ordered in increasing length, P z, A {n
}
exp
z pXq
1
1{n }px . . . x qz,pdx1q z,pdxnq 1 1A exp z pXq K ! 1{n }px . . . x qz,pdx1q z,pdxN q z jpj q j pX qpj q 1A exp z pXq j pj q pj q! j 1{n }px . . . x q z, pdx1 , . . . , dxN q j p j q p z j pX q j q exp z pXq j pj q pj q! j P, A {n } ,
n 1 K 1 N
n 0
1 n!
1A
Finally, setting A M pX q the normalisation constant is obtained and using the fact that P, pn q 1, the assertion follows. z , Corollary 4.4. P
P , .
z
p q p q
P z, p qP, pq
M f N
P z, p qP , p|n
q P , pq
z
71
4. Limit theorems and Extremal Measures That way a new representation of P z is found. For c , dene on X M pX q be the restriction of on X
F p, q P z ,
MpX q let pq
pq pq
n
Pn
z,
F is a probability kernel. F { F } is indeed an Fand observe that specication, which follows from the conditioning procedure of the Poisson F F process. By denition, P z C p q, hence the set of stochastic elds C p q associated to F is not empty. Let pk qk be the sequence of cubes of lemma 3.8, FV k Fk be the tail-eld, and P C p F q. Then for L1 pPq, F Pp|FV q lim p , q k k
F p, q exist P-a.s. in and are by conTherefore the limits Q limk k struction contained in the Martin-Dynkin boundary CV p F q. Dene the j -loop density of some conguration in k as
P-a.s..
(4.17)
Yj,k pq
let Yj be its limit as k V provided that the limit exists and write Y pYj qj . Let M be the set of all those MpX q, such that Yj exists for each j N and is nite. Note that instead of the volume of k the volume of Xk is used to dene the density. However, it has been shown in lemma 3.8 that, asymptotically, their volume is the same up to the constant p2j qd{2 . For notationally purpose we denote the convex y -combination by y :
j 1
nk pj q ; j pXk q
(4.18)
for any sequence y pyj qj of non-negative real numbers. These preparations lead to the limits
yj j
(4.19)
72
4.2. The Microcanonical Loop Ensemble Proposition 4.5. Let f : X R be non-negative and measurable with bounded support, M and Y pq exppf q 1 convergent. Then for
F Pp|FV q lim p , q PY pq k k
any P C , L1 pPq
P-a.s.
(4.20)
Proof. At rst existence and equality of the following limit is shown, lim LF
p, q pf q LQ pf q exp
j 1
Yj pqj 1 exppf q
. (4.21)
F M pX q : lim p, k k
q exists
Let f : X R be non-negative and measurable with bounded support and such that exppf q 1 d $ 0, then there exists k0 such that supp f k for k k0 . Provided N , LF p, q pf q
exp exp
pf q
F p, d q k
k
exp
pf q
n
Pkk
pd q
j
f px1q . . . f pxn
exp
f pxq j, pdxq
k
j
f 1
nk j
pq q
pq q
.
j
f 1 j pX q
k
j Xk
n j k j X k
73
4. Limit theorems and Extremal Measures supp f supp exppf q 1 yields the last line. Since the lhs converges by assumption, so the rhs does. Therefore N M . Vice versa, if M , the rhs converges and so the lhs does, hence M N and (4.21) is shown. Immediately follows that Q is a Poisson process with intensity measure Y pq , which is the claim. In case of divergence of the series, LQ pf q 0 whenever f $ 0, and there is no suitable limit for Q . Thus it follows that the only possible limits for Q are Poisson processes. For FV -measurable and P C proposition 4.5 implies P f pQ q and therefore P Y pq pQ
Pp P Y f Q q
Qq 1
P-a.s.
Particularly, Yj Yj pq P-a.s. for each j . Let F {P CV C : Q P P -a.s.} be the essential part of the Martin-Dynkin boundary associated to F. For a state P C dene a probability measure V P on F as V P pAq PpQ hence by conditioning Ppq PpQ pqq P pqV P pdP q
Aq,
can be written as a Cox process. Vice versa, any probability measure V on F induces a state P C . This argumentation in combination with proposition 4.5 results into the theorem Theorem 4.6. The essential part of the Martin-Dynkin boundary of F consists of all Poisson processes with intensity measure y for non-negative sequences y pyj qj such that y is a -nite measure on X , F
{P y : y -nite}.
74
4.3. The Canonical Loop Ensemble Proof. Let y be -nite. As already seen, P y sition 4.5 and its proof Q For arbitrary P F , P pqV P pdP q Ppq Q pq
y
P y
P y -a.s.
P-a.s.
This implies V P
P
The essential part of the Martin-Dynkin boundary therefore consists of Poisson processes with arbitrary intensities of loops of each kind, where the only restriction is the -niteness of the intensity measure y .
{ M f pNq :
P z, p q
P z, pBk q
Bk
75
P z, pBk q P z , pc
kq.
Since c is the sum of independent, Poisson distributed random variables, c is Poisson distributed itself with the given intensity. From the decomposition of P z , in Corollary 4.4 follows P z , p|c
cq
Bc
P z, p q
1
Bc
P z, p qP, pq
(4.22)
for any measurable function on X , which again emphasises the two step mechanism: At rst choose a composition according to some law and then realise the loops according to the given composition. G } given by Clearly G {
G p, q P z ,
pq
c
(4.23)
G is a G-specication with P z contained in C p q. If pk qk is the sequence of cubes of lemma 3.8, GV k Gk the tail-eld and P CV p G q, then for L1 pPq, G Pp|GV q lim p , q P-a.s. (4.24) k k
klim G p, q V
k
(4.25)
exist P-a.s. in and are by construction contained in the Martin-Dynkin boundary CV p G q in case of existence. Let the loop density of a conguration in k be Wk pq ck , z pXk q (4.26)
and let W be its limit as k V provided that the limit exists. Let M be the set of all those M pX q, such that W exists.
76
4.3. The Canonical Loop Ensemble Proposition 4.8. Let f : X R be non-negative and measurable with bounded support and W pq V. Then for any P C , L1 pPq
G Pp|GV q lim p , q P W z pq k k
P-a.s..
(4.27)
ck
z Xk
c k z X k
W pqz 1 exppf q
that is that Q is a Poisson process with intensity measure W pqz . Similar to the microcanonical case, if W pq is not nite, LQ pf q 0 whenever f $ 0, and there is no suitable limit for Q . Furthermore, the possible limits Q are Poisson processes. Since this implies for GV -measurable , Pp f pQ qq Pp P W z pf pQ qqq one gets P P-a.s. W pqz pQ Q q 1
Particularly W W pq P-a.s. holds. Let G {P CV C |Q P P -a.s.} be the essential part of the Martin-Dynkin boundary of G . For P C dene a probability measure V P on G as V P pAq PpQ Aq,
77
4. Limit theorems and Extremal Measures for that reason Ppq PpQ pqq
G
P pqV P pdP q
is a mixed Poisson process. Vice versa, any probability measure V on G induces a P C . All this can be put together: Theorem 4.9. The essential part of the Martin-Dynkin boundary of G consists of all Poisson processes with intensity measure wz for any positive real number w, G {P wz |w 0}. Proof. If w is a positive real number, wz is a -nite measure on X . Since G P P wz C p q, and by proposition 4.8 Q P wz wz -a.s. For arbitrary F P we have
G
P-a.s.
This implies V P
P
pq
(4.28)
E Pp|EV q lim p , q k
P-a.s..
(4.29)
78
klim E p, q V
k
(4.30)
exist P-a.s. in and are by construction contained in the Martin-Dynkin boundary CV p E q. Proposition 4.10. Let f be non-negative and measurable with bounded support. Then LQ pf q lim LE p, q pf q exists, is non-degenerate and
k
LQ pf q exp
z 1 exppf q
(4.31)
Proof. The proof of the corresponding microcanonical loop ensemble applies j with Yj zj . This means that the Poisson process with intensity measure z is the only limit, hence there is no phase transition. We obtain Theorem 4.11. The essential part of the Martin-Dynkin boundary of E consists of the Poisson process with intensity measure z .
j 1
jn pj q.
79
4. Limit theorems and Extremal Measures Hence, under P z , N has a compound Poisson distribution whenever z 1 for d 3 and z 1 for d 1, 2. However, the nature of the sub- -algebras does not allow a direct computation of the limits like in the propositions 4.5, H as a conditioned 4.8 and 4.10. Similar to the loop ensembles we dene Poisson process, represent it as a convex combination of P , but instead of these computations we show a large deviation principle for the mixing measure. If the latter measure converges to a suitable limiting probability H will measure, then, since the microcanonical weak limits are known, converge as well. From now on x d 3, z 1 and write instead of 1 . Remark 4.18 below comments on what diers in the cases z 1 and d 1, 2. At rst we derive the representation in terms of P, . Lemma 4.12. With CM { M f pNq : compositions with rst moment M and with N M , it follows pq 1CM p qP pd q j pj q M } being the set of MpX q a xed conguration
CM
P p qP,
pq
. (4.32)
Proof. This can be seen from disintegration of conditional expectations like in the beginning of section 4.2. If we now condition P on the event {N M } on the lhs of equation (4.32), this turns into P conditioned on CM on the rhs. Dene
H p, q P ,
pq
N
P pd |CN q,
(4.33) (4.34)
P,
pq
H } is even an which is indeed a probability kernel on X M pX q, H { H-specication. Like in the previous sections, let pk qk be the sequence of cubes of lemma 3.8. Before we turn to the analysis of the Martin-Dynkin
80
4.5. The Canonical Ensemble of Elementary Components boundary of H , we derive a large deviation principle for P p |CN q. This one can be shown in using a large deviation principle for P p q. Since the deviation is done for xed , we write Mk instead of Nk and think of k it as an increasing parameter in k such that |M k | converges to some nite limit as k V. Large deviation principle for P k . Recall from lemma 3.8 that the intensity measure k grows asymptotically like the volume of k , and let : lim
k
V |k |
p2 qd{2
1
j 1
j . d 2
(4.35)
satises a large deviation principle with speed |k | and good rate function I : MpNq r0, Vs given by the relative entropy with respect to , I p; q pf log f
|k |
f 1q
if 3 , f : otherwise
d d , f
log f
f 1 L1p q ,
1 k |k | log P
inf I p; q G inf I p; q. F
(4.36)
(4.37)
Large deviation principle for P k p |CMk q. The conditioned Poisson process is absolutely continuous with respect to the unconditioned process,
81
4. Limit theorems and Extremal Measures where the density is an indicator function times a normalisation constant. That way the LDP for P k transforms into some LDP for P k p |CMk q. P k exppCMk q k p |CMk q P
exp
C pq P p q,
Mk k
if A . otherwise
As known in large deviation theory, the rate function for P k p |CMk q will be the rate function for P k plus a functional of the form A for a suitable set A, see i.e. [DS00]. Because of poor continuity properties of these functionals A additional care has to be taken. Let Du : MpNq : jpj q u
be the set of measures on N with rst moment u representing the densities of the loops of the dierent kinds. Observe that in the weak topology Du is neither upper nor lower semicontinuous. But if its upper or lower semicontinuous regularisations are not innite for every MpNq, one may deduce the lower and upper large deviation bound, respectively, as we will do in the sequel. Lemma 4.13. The upper and lower semicontinuous regularisations usc Du and lsc of with respect to the weak topology are D u Du usc Du pq V, lsc Du pq
V
0
if jpj q u . otherwise
(4.38)
lsc Proof. First note that usc A int A and A cl A . But cl Du { MpNq : jpj q u}, hence we get the lower semicontinuous regularisation c qc r and the upper of Du . By the same argument we get int Du pcl Du semicontinuous regularisation.
82
4.5. The Canonical Ensemble of Elementary Components Upper large deviation bound of the partition function. Lemma 2.1.7] we get the upper bound as lim sup
k
Applying [DS00,
|k | log P
exppCMk q
(4.39)
Since Du is not lower semicontinuous, it is replaced by its lower semicontinuous regularisation on the rhs. We solve the variational problem on the rhs. of equation (4.39), which is a minimisation problem with a constraint. Proposition 4.14. Let zu be the solution of
p2 qd{2gd{2pzq u u,
(4.40)
where u : p2 qd{2 gd{2 p1q and gd{2 is given in equation (3.4). Then the minimiser of
M N
inf I
p q
lsc D
(4.41)
is given by p2 qd{2
1d{2 j j zu
j .
j 1
Proof. The minimisation of I lsc Du is equivalent to the minimisation of I under the constraint jpj q u. For the moment, assume u u and minimise I given jpj q for any u. By the Euler-Lagrange method of conditional minimisation, I pq jpj q log z
j 1
j 1
j log j log
pj q pj q
1 pNq 1 pNq,
j 1
pj q z j pj q
j 1
log z j pj q
83
I p q
j 1
j z pj q pNq
j 1
j p1 z q pj q
follows. Since necessarily z 1 and z is an increasing function of , equation (4.41) holds. Now let u u , so there is no solution of the equation p2 qd{2 gd{2 pz q u. Let u0 u p2 qd{2 gd{2 p1q be the surplus mass. Dene and 0 pnq u , then clearly for all n n n
j 1
j pnq pj q
j 1
j pj q u0
u
0 pnq u n log pnq
pNq pNq
Lower large deviation bound of the partition function. By lemma 4.13, the upper semicontinuous regularisation usc Du of Du is not nite, and the analogue argument for the lower bound does not apply. The reason is the sparseness of Du in the weak topology which even holds for the blow ups of D of the form D { MpNq : | Du jpj q u| } for any 0. u u Otherwise this could have been used for some kind of Boltzmann principle, see e.g. [RZ93].
84
4.5. The Canonical Ensemble of Elementary Components However, the 2-parameter sets Dm,s : MpNq :
j m
jpj q s ,
(4.42)
0m 1
Dm,s
cl Ds, V,
V kV
1 k exp |k | log P
m,s
1{Dm,s L}
(4.43)
we get for any m and by [DS00, Lemma 2.1.8] a lower bound lim inf
k
1 k exppDm,s q |k | log P
M inf I D pNq
m,s
(4.44)
for the system restricted to the rst m components. Therefore we get the lower bound for the original problem as m V and 0. Consider now the family of minimisation problems on the rhs of equation (4.44). Here we have to link the two parameters m and s. Since 1 j m j d{2 is strictly decreasing to 0, there exists m0 N such that for any Proposition 4.15. Let 0 and m N be such that sm, : u p2 qd{2 jm jd1{2 0 and zpm,q be the solution of p2 qd{2 jm jzdj{2 sm,. Then the inmum of I Dm,sm, on MpNq is attained at with m m0 , u p2 qd{2
1 j m j d{2
0.
proposition 4.14.
85
4. Limit theorems and Extremal Measures Proof. The rst part is similar to the previous proof where the minimiser is given in equation 4.45. To see the second part, assume for the moment u u , then sm, is not exactly the m-th partial sum of the series of p2 qd{2gd{2p1q, but close to it. Observe that zpm,q 1 for each m m0 and pzpm,q qmm0 is an decreasing sequence for any 0. Indeed, from
p2 qd{2
j m
d{2 j
sm, p2 qd{2
j zp m,q j m
1 and
1 pm 1qd{2
j d{2
p2 qd{2
Hence zpm,q 1 for any 0 as m V. For u u these arguments apply as well. Let now u u , x 0 such that u u and m0 be even large enough, such that sm, 0. Then rstly zpm,q 1 for each m m0 follows since
m m0
yields the decrease. Finally the sequence pzpm,q qm can not be bounded away j zp m,q from 1 for any 0 since otherwise would diverge. j m j d{2
pm 1qd{2
m1 zp m,q
p2 qd{2
j zp m,q j m
Next we show that pzpm,q qmm0 is an increasing sequence in m and tends to zu . Since sm1, sm,
jd 2
1
j m
jd 2
p2 qd{2
1
j m
jd 2
{ .
zpm1,q needs to be bigger than zpm,q . Since pzpm,q qm is bounded from above by 1, the sequence converges and the only limit can be zu since
z m1
86
V.
weakly. In particular, the case u u causes the diculties in propositions 4.14 and 4.15. See also remark 4.19. Martin-Dynkin boundary. Back to Martin-Dynkin boundary technique, we interpret the boundary condition M pX q as a random element and write capital letters instead of small ones to emphasise the dependence on Nk . Let U be the limiting particle density, U pq limkV |k | , in case of existence of the limit and put U pq V if the limit does not exist. For each conguration with U pq V there exists Z Z pq such that
Since the minimiser of the minimisation problem was unique, the conditioned Poisson process is asymptotically degenerate and lim P : (4.46) k kV |k | CMk zu
p2 qd{2gd{2pZ q U u.
(4.47)
The considerations on large deviations lead to the desired weak convergence and we obtain Proposition 4.16. Let f : X R be non-negative and measurable with bounded support, P C and M . Then for any P C , L1 pPq
H Pp|HV q lim p , q P Z pq k k
P-a.s.
as k V. Now we can use the results of section 4.2 to deduce that the measures converge H lim p, q P Z pq . k
k
|k |
CNk
p q
87
4. Limit theorems and Extremal Measures Again the reasoning of the previous sections applies. Since for HV measurable , Pp f pQ qq Pp P Z pf pQ qqq holds, we get
P -a.s.} be the essential part of the Martin-Dynkin boundary associated to H, then we deduce Theorem 4.17. The essential part of the Martin-Dynkin boundary of H consists of all Poisson processes with intensity measure z for z r0, 1s and d 3, H {P z |0 z 1}.
H Proof. P z C p q, and Q H trary P we have
P Z pq pQ
P
adjusted. The Lagrange multiplier z , which occurs during the minimisation procedure using , will be, given z I , some z related to z via z z I z . The discussion is carried out in more detail in subsection 7.4. In fact, since gd{2 p1q diverges for d 1, 2, the minimisation problems in proposition 4.14 and 4.15 simplify since no mass can get lost. Remark 4.19. We constructed the minimiser in proposition 4.14 and 4.15 A for the weak topology on MpNq, which means that the loop densities | , k| the loop densities represented as measures on N, converge as k V for arbitrary A N. However, for the particle densities dierent behaviours occur. For a low particle density u u the total mass is conserved, hence
P . Remark 4.18. For d 1, 2, we necessarily start with the intensity measure z I for some z I 1. By corollary 2.8, the large deviation principle remains valid with respect to the -topology with the relative entropy properly
This implies V P
88
4.5. The Canonical Ensemble of Elementary Components we get convergence for any A N and any particle is contained in some nite loop, whereas for u u some mass is moved to innity and lost. Therefore for the limits Q the particle density U is P-a.s. bounded for any P C p H q, hence P-a.s. the limit (4.46) even holds on N endowed with the -topology. This phenomenon of bounded particle density shows that a condensation eect is present, but does not occur with positive probability. Remark 4.20. One may collect the surplus mass at an exterior point, say V, by replacing N endowed with the vague topology by its Alexandrov compactication N {V}. Still P k p |CMk q converges weakly to the same deterministic limit. At low density u u the particle densities stay the same, but at u u the surplus mass u0 reaches V. On the contrary, by the proof of proposition 4.14, there is no excess of loops. Hence one may dene a density of particles contained in innitely long loops, but no density of innite loops.
89
resulting in a family {P x }xX , see proposition 1.20. This construction permits the interpretation that P x is P conditioned on the event {{x} 0}. Further analysis can be carried out in case of P obeying additional invariance properties, particularly translation invariance. Indeed, as shown in chapter 1, P z is invariant under the translation group T {Ts }sRd , i.e. P z is invariant under the translations Ts : X X , x s x for each s Rd . Under such a condition a result of Mecke [Mec67] is extended: Let s : X Rd , x xp0q be the projection of a loop to its starting point, g : Rd V be a non-negative, measurable and sz -integrable function, then the Palm distribution can be obtained by the g -weighted average over all loops of a conguration which start in the support of g , and then average with respect to the point process. Hence the Palm distribution takes the form P 0 pAq s pg q1 g puq1A pTu qspduqP pdq. (5.2)
91
N
0,
j 12
1 j 2 I b0
1 d j d 1
1 1
N
N
0, g {
1
0,
j I 12 gd 2 z
d{2 pzq I
pq
2 p2q k
k!
k d k 1 j d 1 1 n
Table 5.1.: Geometric Properties Furthermore, in [Mec67] from the stationarity of P the independence of the particular choice of g is shown. In chapter 4 the point processes which are a stochastic eld for a given specication were characterised. They are given as a mixture of extremal elements from the Martin-Dynkin boundary, its essential part. These are exactly the ergodic point processes, and for those additional results with an important interpretation can be given. In equation (5.2) replace g step by step by the indicators of nice, convex sets k In sections 5.2 5.5 dierent properties related to the typical loop are considered. In contrast to chapter 4, results may dier depending on whether a loop is Brownian or a random walk loop. Table 5.1 shows some of the results for typical loops. The rst property to explore is the typical barycentre in section 5.2, which turns out to be normally distributed in any case. In the Brownian bridge case, the typical j -loop barycentre turns out to be normally distributed gd{2 pz q with covariance matrix j 12 I , proposition 5.4 and g1d{2 pz q I for the typical
92
barycentre, corollary 5.5, where I is the identity matrix. The covariance of the typical j -loop barycentre in the random walk bridge case agrees with the one in the Brownian bridge case up to an additional correction factor 1j 2 , proposition 5.8, and turns out to be closely related to the computation of the barycentre of a given set of points in Rd , which is an important task in multivariate statistics [And84]. In particular the variance of the barycentre of the random walk j -loop is always smaller than the variance of the barycentre of the Brownian j -loop, but they agree asymptotically. In section 5.3 the location of the typical random walk loop at the discrete times 0, , 2, . . . is considered. The mean euclidean distance between succeeding points as well as the mean euclidean length of a typical loop is determined. As one expects, the mean length of a step of a j -loop turns out to be shorter than a corresponding step of an unconditioned random walk, with the correction factor being 1 1{j , see corollary 5.17. Even more holds: two succeeding steps dene a triangle with associated area or 2-volume and more generally, k steps dene a simplex with the associated k -volume. Similar to the 1-volume case, the k -volume of a j -loop is smaller than the k -volume of an unconditioned random walk, now by a factor 1 k {j . The following section 5.4 again considers the set of vertices in R2 given by the random walk j -loop. The interest lies in the number of extremal points of this set and its asymptotic behaviour as j V. These extremal points are exactly the vertices of the convex hull of the given set, and therefore the vertices of a polytope. In general independent points, uniformly distributed in a domain of a special shape, have been considered by various authors. Computations for normally distributed points can be found in R enyi and Sulanke [RS63]. Here, however, the dependence is the main diculty, but we use fruitfully a close connection to the event for random walks to stay positive. Section 5.5 is concerned with percolation. A typical conguration consists of innitely many loops, of which some may overlap. Clusters are built from overlapping loops and the basic question is about the size of the typical cluster which is the cluster which contains the typical loop. We show in
93
5. Geometric Aspects of the ideal Bose Gas proposition 5.28 that for suciently small z , there is P z -a.s. no unbounded cluster, and moreover in corollary 5.29 that the diameter of the typical cluster has at least a nite forth moment.
with the family {P x }x being the Palm kernel. The Palm distribution P x of P at x is interpreted as P conditioned on the event {x} 0 that there is at least one point at the site x. In case of P being a stationary point process an independence of the Palm distribution P x of x should be expected. Assume X to be an Abelian group with the translations T {Tx }xX acting measurably on X . A point process P on X is stationary if P pTx Aq P pAq for every x X . In this case, by Mecke [Mec67], the Palm distribution may be dened alternatively as P0 pAq 1 pg q 1A pTx qg pxqpdxqP pdq
for any non-negative, measurable function g : X R with g d V. The stationarity ensures that this denition does not depend on the choice
94
5.1. Palm Distributions and Stationarity of g and therefore is well-dened. If X Rd the usual choice for g is the indicator of the unit cube. The Palm kernel dened in equation (5.3) agrees with the denition in equation (5.1) in the following sense [DVJ08b, thm.13.2.III], Theorem 5.1. Let P be a stationary point process. Then the Palm kernels {P x }xX can be chosen such that for any x X , P x pAq P0 pTx Aq. Therefore P x can be obtained from P0 by shifting the origin towards x and the subscript is allowed to become a superscript. Due to the denition of the family of Palm kernels as a Radon-Nikod ym derivative, {P x }xX has to be chosen appropriately on -null sets. The next step is to relax the assumption that X is an Abelian group, but assume that a group of translations T acts measurably on X . Since by the discussion in section 1.1.3 the T -invariance of a Poisson process is reected by the T -invariance of its intensity measure and vice versa, by the disintegration lemma 3.4 for z , P z is invariant under the translation group T {Ts }sRd . x Especially for P z the disintegration means the Palm distribution P z is the Poisson process P z conditioned on the occurrence of a xed loop x X . In attempting to use the T -invariance, the condition on the occurrence of some loop starting at s Rd seems more suitable. By construction the loop measures j were required to satisfy a disintegration j pf q p2 qd{2 f px0 qs j, pdx0 qds f ps x0 q0 j, pdx0 qds
Rd
X{s}
p2 qd{2
Rd X{0}
d with s j, being the distribution of the j -loop starting at s R . Necessarily 0 s j, agrees with j, Ts . Consider again the disintegration of the Campbell measure CP and suppose further that the function h, which is integrated
95
5. Geometric Aspects of the ideal Bose Gas with respect to CP , depends on x only via the starting point sx : xp0q, then CP z phq
x hpsx, qP z pdqz pdxq
j 1
zj j p2j qd{2 1
g
1 d 2
{ pz q
s pdqds, hps, qP z
s pdq is the weighted convex combination of the Psx0 pdq0 s, where P z z j, s pdq P z j 1d{2 zj
sx0 P pdq0 z j, pdx0 q.
j 1
x Instead of P z conditioned on the particular z being the Poisson process P s is P loop x being contained in the conguration, P z conditioned on some z d loop starting at s R . Thus we lost information due to the averaging, but s of the position. gained the independence of the averaged Palm kernels P z
Denition 5.2 (Typical loop). The typical loop of P z is the loop starting 0 . at the origin under P z Therefore we get an analogue of equation (5.1) for the T -invariant Poisson process P z , which reads as 0 pq g1d{2 pz q P z p sxq1F psxqpdxqP z pdq,
where F is the unit cube in Rd . Here the starting point of every loop starting in F is moved towards the origin, where is evaluated. By now the typical loop of P is the loop which starts at the origin with 0 respect to the Palm distribution P . On the other hand, a loop consists of possibly several elementary constituents and a typical particle could be of
96
5.1. Palm Distributions and Stationarity another interest than a typical loop. If we introduce the symmetrisation t of a loop x Xj as t : x
tpxq
j 1
k 0
xpk q
and appropriately continued for M pX q, we get the symmetrised Point process tP z . Note that simple congurations stay simple if and only if for any two distinct loops x, y and any k the k -time shift of x is dierent from y . Furthermore the projection of the loop of a conguration into Rd does not change under symmetrisation. Let sk : Xj Rd be the projection on the starting point of the k -th particle, sk : x for 0 k
skx xpk q
j 1 k 0 sk
j.
holds.
p sxq1F psxqpdxqtP z pdq pt sxq1F psxqtpdxqP z pdq pt stxq1F pstxqpdxqP z pdq pt sk xq1F psk xqpdxqP z pdq
k
97
sm .
In subsection 5.2.2 these points are given by the locations of a random walk loop, a j -loop brings along j points. The denition of the barycentre carries over to a set given by a measurable function f on some interval r0, T s, S : 1 T f ptqdt.
Of special interest in subsection 5.2.1 is the barycentre of the typical Brownian loop.
N let
b : Xj
Rd ,
1 j
j 0
xpsqds
assign to each loop its barycentre and assume b acting on X . Furthermore continue b on M pX q such that for M pX q, b M pRd q is the point conguration of the barycentres of the loops of . For M pX q with pX{0} q 0 let b0 be the barycentre of the loop starting at the origin, b0 : bx if x with sx 0. Recall s : X Rd being the projection on the starting point. Before we turn to the typical loop, we compute the distribution of the barycentre of the 0 typical j -loop, i.e. the distribution of b0 under P j at inverse temperature 0.
98
5.2. The Barycentre Proposition 5.4 (Typical j -loop barycentre). Let j N be a positive in0 teger and 0. Then under P j , the typical j -loop barycentre is b0
N
0,
j I . 12
In fact there is no big surprise that the barycentre of a Brownian bridge starting and ending at the origin is normally distributed with its mean at the origin. Proof. Let T : j and pBt q0tT be a Brownian motion. Then with At pxq xptq, see e.g. Revuz, Yor [RY91, prop. 1.3.7], At and b
d
t d Bt BT T
1 T s BT 1 T 1 T Bs ds Bs ds BT ds . T 0 T 0 T T 0 2 Therefore b is normally distributed and it suces to compute expectation and covariance matrix. But the expectation vanishes since the Brownian motion is a centered process. Therefore the covariance matrix remains, for which it suces to compute the diagonal elements because of the independence of the components. Hence for i 1, . . . , d bi
2
1 T 1 T2
T 0 T 0
i Bs ds i ds Bs
i BT 2 2
i BT T
T 0
i ds Bs
i BT 4
Starting from the last one we calculate the three expectations. Clearly i 2 T . Because of EB i B i s for s T , the expectation of the E BT T s second summand is E
i BT T T 0 i Bs ds
1 T
T 0
sds
T . 2
99
5. Geometric Aspects of the ideal Bose Gas For the rst summand let It : square,
2 IT t i 0 Bs ds
I02 2
T 0
T 0
It dIt
2
hence by Fubini 1 E T2
2 i Bs ds
T22 T . 3
T 0 0
i i EBs Bt dsdt
Summation leads to the desired result. Alternative proof. Let T : j and 0 j, be the Brownian bridge measure on Xj from denition 3.5. Then, see e.g. Revuz, Yor [RY91, prop. 1.3.7], for i 1, . . . , d, st . (5.4) xi psqxi ptq0 T pdxq s t T Since 0 T is the probability measure of a Gaussian process and bx 1 T
T 0
xptqdt,
b is normally distributed and it suces to compute expectation and covariance matrix. But the expectation vanishes since the Brownian bridge is a centered process and the covariance matrix remains. Since the components of the Brownian bridge are independent, it is sucient to compute the diagonal elements. For i 1, . . . , d,
pbxqi 20 j, pdxq
1 T
T 0
xi psqds
0 j, pdxq.
t i 0 x
(5.5)
Partial integration of the square of the inner integral It pxi q : leads to IT pxi q2
psqds
I0pxiq2 2
T 0
It dIt
2
T 0 0
xi psqdsxi ptqdt.
100
5.2. The Barycentre Applying Funbini and equation (5.4) continues equation (5.5) as 2 pbxqi 20 j, pdxq T2
T 0 T 0 0 0 t t
T22
T . 12
Corollary 5.5 (Typical Barycentre). Let the fucacity satisfy 0 z 1 if 0 d 3 and strictly less than 1 if d 1, 2. Then under P z is the typical barycentre b0
N
0,
gd{2 pz q I . g1d{2 pz q
This means that at fugacity z 1 still each typical loop has a welldened barycentre as long as the dimension is at least 3. Only in the low dimensions 1 and 2, when the particle number is not integrable anyways, the barycentre has no well-dened distribution. Proposition 5.6 (Expected sample variance of the typical j -loop). The expected sample variance of the typical j -loop is j 12 I . Proof. 1 T
T 0
1 T
pbxqi
dt0 j, pdxq
101
5. Geometric Aspects of the ideal Bose Gas Starting from the last integral we calculate the three quantities, 1 T 2 T
T 0 T 0
pbxqi
dt0 j, pdxq
xi ptqpbxqi 0 j, pdxqdt
1 T
T 0
xi ptq2 0 j, pdxqdt
1 T 2 T2 2 T2 1 T 1 T
T 0
T T dt 12 12
T 0
T 0 T 0 T 0 T 0 0
t t
t2 dt T t2 dt T
T 12
R ,
d
1 j
j 1
m 0
xpm q
102
5.2. The Barycentre assign to each loop its barycentre and assume b acting on X . Furthermore continue b on M pX q such that for M pX q, b M pRd q is the point conguration of the barycentres of the loops of . For M pX q with pX{0} q 0 let b0 be the barycentre of the typical loop, b0 : bx if x with sx 0.
First of all, for an overview, a few sums are collected, which will occur in the sequel. They are given without proof. Lemma 5.7. Let j iq iiq iiiq
N. Then
j 1
m 0 j 1
m 0 j 1
m 0
1 pj 1qpj 1q. 6
At rst the distribution of the barycentre of the typical j -loop, i.e. the 0 distribution of b0 under P j at inverse temperature 0 is computed, and later the distribution of b of the typical loop is focused. Proposition 5.8 (Typical j -loop barycentre). Let j N be a positive in0 teger and 0. Then under P j the typical j -loop barycentre is b0
N
0,
j 12
1 j2
I .
Proof. Let pRm qm0...N be a random walk with independent, N p0, I qdistributed steps. Then the distribution of the m-th step, xpm q, and Rm m Rj j
103
5. Geometric Aspects of the ideal Bose Gas are equal and the appropriate representation of the barycentre is b
d
1 j
j 1
1 j
m 0 j 1
Rm
1 j
j 1
m 0
m Rj j 1 Rj . j 2
Rm 1
m 0
Therefore b is normally distributed. The expectation of b vanishes since the family pRm qm is centered. We compute the covariance matrix. In analogy to the continuous case for i 1, . . . , d, bi
2
1 j 1 j2
j 1
m 0 j 1
i Rm 1 2 i Rm
i 1 Rj j 2
(5.6)
j 1
m 0
i 1 Rj j j
m 0
Rm 1
1 j
i Rj 4
(5.7)
and we calculate the three expectations starting from the last one. Clearly i 2 j . Because of ERi Ri m for m j , E Rj j m
j 1
m 0
i i ERm Rj
1q j pj 2
j 1
and nally
j 1
m 0
i Rm
j 1
m 0
1q j pj 2
i E Rm
2
j 1
j 1 i i Rn ERm
m 0n m 1
m 1
mpj 1 mq
5.2. The Barycentre Therefore putting the expectation on equation (5.7) leads to E bi
2
j 1 1 j j 12
1
1 j
1 2j
1 3
2 j
1 2
1 j
1 4
1 j
1 . j
Alternative proof. The covariances of each component i random walk bridge are given by mn 0 xi pm qxi pn q T pdxq m n j Since bx 1 j
j 1
1 . . . , d of the
. (5.8)
m 0
xpm q,
b is normally distributed and it suces to compute expectation and covariance matrix. But the expectation vanishes since the random walk bridge is a centered process and the covariance matrix remains, again the computation of the diagonal elements is sucient,
2 0 bi T
pdxq
j 1
1 j
j 1
m 0
xi pm q
0 T pdxq.
(5.9)
pdxq
1 j2
m 0
2 0 xi pm q T pdxq j 1 m 1 m 0n 0
j22
j2
j 1
0 xi pm qxi pn q T pdxq
m 0
m2 j
2 j2
j 1 m 1 m 0n 0
n 1
m j
105
j2 6
j 1
m 0
1 1 2 j 1 1 j2 3 j j 1 7 2 1 1 j 4 j 3j 3j 2 1 1 1 2 j 1 j12 1 1 j 6 j 3 j 7 2 1 1 3 3j 3j j 1 1 1 1 12 j j 1
m2 j
j2
j 1
m 0
mpm 1q 1
m j
Therefore the results of proposition 5.8 and proposition 5.4 agree up to the correction term 1 j 2 . As j increases, both variances grow by the same rate. Moreover, the variance of the barycentre of a random walk loop is always smaller than the corresponding variance of the random walk bridge. Directly the distribution of the typical barycentre follows: Corollary 5.9 (Typical Barycentre). Let the fugacity satisfy 0 0 d 3 and strictly less than 1 if d 1, 2. Then under P z ,
z 1 if
b0
N
0,
g1d{2 pz q
j 1
zj 1 1 2 I . d { 2 j j
Proposition 5.10 (Expected sample variance of the typical j -loop). The expected sample variance of the typical j -loop is 12 pj 1qI .
106
m 0
x pm q pbxq
i
0 T pdxq
j 1
m 0
xi pm q2 2xi pm qpbxqi
j 1
m 0 j 1
x pm q
i
2 0 T
pdxq
j 1
pbxqi
m 1 m j
0 T pdxq
m 0
pj 1qpj 1q 6
j 1
m 0
pbxqi
i
2 0 T
pdxq 12 pj 1qpj 1q
m 0
x pm qpbxq
i 0 T
pdxq
j 1
m 0
m 1
m j
pj 1qpj 1q, 6
i 0 T
pdxq
j 1
j
mn n 1 m2 j m j .
n 0 m 1 n 0
mn j
j 1 n m
mn j
2
m 0
107
5.3. k -Volumes
The basic question addressed in this section is What is the length of a typical loop? More precise, the question could be How many steps does a typical loop have? Another way of thinking about that is the question for expected euclidean length of a typical loop, for a j -loop this is j times the expected length of one step, say the rst one. If x is the loop of starting at the origin, this is the expected distance between the two successive points xp0q 0 and xp q. In taking two steps one gets the three points xp0q, xp q and xp2 q forming a triangle which has a certain area. This way the 1-volume and the 2-volume of one and two successive steps are dened, respectively. The generalisation to general k is straightforward. We start with independent vectors before passing to the dependent case and nally to more general rotational invariant situation.
Rm Rm Rm
. . .
sin m 0 cos m 0
Rm cos m d2 m r0, q for j 1 are the polar where m r 0 , 2 q is the azimuth angle, 0 j angles and Rm : |Ym | is the length of Ym . Because of rotational invariance
of the k -volume, we may choose k d2
1 1 k d3 . . . dk1 0.
108
5.3. k -Volumes This rotation causes Yk to direct to the north pole and xes the remaining vectors such that the k -volume of the k vectors admits a simpler representation. Lemma 5.11. The k -volumes volk satisfy the recursion volk pY1 , . . . , Yk q R1 k
d 2 j d k
Proof. Due to the choice of the angles, k j is the angle between the line through Y1 and the origin and the plane given by {0, Y2 , . . . , Yk } and hence the calculation is standard. Conditioned on the vectors Y2 , . . . , Yk this directly leads to Lemma 5.12. Let n denote the volume of the n-dimensional unit sphere. Then the expected k -volumes satisfy the recursion 2 dk1 volk1 Y2 , . . . , Yk . E volk Y1 , . . . , Yk Y2 , . . . , Yk k dk2 Proof. Because of the independence of the length of a vector and its direction, E volk Y1 , . . . , Yk Y2 , . . . , Yk
1 d
d 2 j d k
volk1
j 1
Y2 , . . . , Y k
sin 1 j
j 1
d k 1
ER1 k
1 1 sinj 1 j d1 dd2
The sines result from the transformation to polar coordinates and lemma d 5.11. Because of ER1 2 and the integrals over the products yield d1 d1 dk2 and dk1 , respectively, we may continue to deduce E volk Y1 , . . . , Yk Y2 , . . . , Yk
volk1pY2, . . . , Yk q
2 dk1 . k dk2
109
5. Geometric Aspects of the ideal Bose Gas Corollary 5.13. With n denoting the volume of the n-dimensional unit sphere for the k -volume of k independent, normally distributed vectors Y1 , . . . Yk holds E volk Y1 , . . . , Yk In particular we get E vol1 Y1 q p2k !
k 2
dk1 . d1
E vol2 Y1 , Y2
d d1 d1 d1 2
E vold Y1 , . . . , Yd
q p2d !
1 d 2
d 2
1 . d1
110
1k
.. . . . . 1k
kk
..
. kk
with k diagonal blocks of size d d on each row and column. Particularly, if is symmetric and positive denite, also is. In this case we may nd symmetric and positive denite matrices and . such that T and T and the relation between and is the same as between and . Now complete X1 , . . . , Xk with d k unit vectors Xk1 , . . . , Xd which are orthogonal among themselves and to X1 , . . . , Xk . Then volk pX1 , . . . , Xk q 1 det X1 , . . . , Xd k!
Proposition 5.15. Let X1 , . . . Xk normally distributed random vectors with geometric covariance . Then E volk pX1 , . . . , Xk q det E volk pY1 , . . . , Yk q where Y1 , . . . , Yk are i.i.d. normally distributed. Proof. By the denition of the volume and with x denoting the kd-vector obtained from joining x1 , . . . xk , E volk pX q
detpx1 , . . . , xd q exp
1 xT 1 x 2
dx.
111
In fact, py qj j 1 y1 . . . jk yk is just a linear combination of vectors and therefore by linearity of the determinant in each component det py q1 , . . . , pyqk , xk1, . . . , xd det det y1, . . . , yk , xx1, . . . xd
k.
x m
m 0, . . . , k
112
5.3. k -Volumes be the visited points of the rst k steps including the starting point. Then, again dy dy1 dyk , Ef p1 , . . . , k q p2j q 2
d
pyj1qdy1 dyj1 p2j q f py1, . . . yk q py1q py2 y1q pyk yk1qpjkq pyk qdy1 . . . dyj1 p2j q f py1 , . . . yk q p2 q p2pj kq q 1 exp y T 1 y dy 2
d 2 d 2 kd 2 d 2
and we may identify the inverse of the geometric covariance as the k k matrix 2 1 0 1 . . . . . . 1 .. . 2 1 1 0 1 1 jk with the remaining elements being 0. Lemma 5.16. The determinant of 1 is det 1
k
jk
j k j . k
Proof. This can be seen by induction in applying successively Laplace expansion from the lower right corner or from the normalisation in the calculation above. Finally put proposition 5.15 and corollary 5.13 together to obtain the result
113
a2 a1 a5
a0
a3
a4
Figure 5.1.: Simplex dened by the rst two steps of a 6-loop Corollary 5.17. The k -volumes of the rst k steps of j -loops for k E volk p1 , . . . , k Proof. Due to proposition 5.15 E volk p1 , . . . , k q 2
k
j are
q p2 q
k!
k 2
k dk1 . j d1
k 1 E volk pY1 , . . . , Yk q, j
k 2
p2 q
k!
k dk1 j d1
since Y1 , . . . , Yk are independent and normally distributed. Therefore the k -volumes of the rst k steps of a j -loop are up to a factor depending on j the k -volumes of a random walk with independent steps.
114
5.3. k -Volumes For j large this is expected to be close to the independent case and the corollary shows exactly the dierence. Dene the k -volume of a j -loop as the sum of the k -volumes when starting at each of steps xp0q, xp q, . . . , x pj 1q , which is by symmetry j times the k -volume of the rst k steps. Suppose furthermore volk to be continued on M pX q such that volk measures the k -volume of the loop starting at the origin. Corollary 5.18 (Expected k -volume of the typical loop). The expected k -volume of the typical loop is
0 P z
pvolk q g
1d{2 pz q j k
zj j d{2
p2 q 1 k!
k j
k 2
dk1 d1
rd1 pdrq C , i.e. if the radius of a sphere has distribution , its expected volume is nite. In this case 1 Ef pZ q f pr, qrd1 pdrqd d C for any rotational invariant random vector Z with radial distribution . The results then take the form
115
5. Geometric Aspects of the ideal Bose Gas Lemma 5.19. Let Z1 , . . . , Zk be k independent and rotationally invariant random vectors with radial distribution . Then the k -volume of the simplex spanned by the origin and the ZiI s satises the recursion E volk Z1 , . . . , Zk Z2 , . . . , Zk d1 dk1 V volk1 k
d d k 2
Z2 , . . . , Z k .
This is directly obtained from the proof of lemma 5.12 in replacing ER1 by V . Therefore it is no surprise that the results only dier by a factor. From the recursion one obtains an explicit result, Corollary 5.20. Let Z1 , . . . , Zk be k independent and rotationally invariant random vectors with radial distribution . Then the k -volume of the simplex spanned by the origin and the ZiI s is exactly E volk Z1 , . . . , Zk
V k!
d1 d
dk1 . d1
116
5.4. Convex Hulls in R2 Unfortunately the relation between edges and vertices fails in higher dimensions, but the idea how to identify a line between two points of V as an edge of conv V in two dimensions can still be used to identify the appropriate part of a hyperplane as a face of the polytope conv V . We address this questions of higher dimensions at the end of this section and keep on considering d 2. For v, w V let vw be the line dened by and rvws be the line segment between v and w. Introduce the indicator vw : 1 0 if rvws is an edge of conv V . otherwise
Then the total number of edges is half of the sum vw over all pairs v, w pV q : 1 2
V,
v,w V v w
vw .
Assume to act on X through the relation V {xpk q : k 0, . . . , j 1} if x Xj and continue on M pX q such that to a starting point xp0q of a loop x the mark pV q is attached. Let 0 be the corresponding value of the typical loop. To get the expected number of edges of V , one has to compute the probability that vw is an edge. There is a strong connection between this probability and the probability pn that a random walk bridge of length n stays non-negative. Lemma 5.21 (Vertices of the convex hull of j -loops). Let x Xj , then
0 P j
p0q j
j 1
n 1
pn pj n .
Proof. For the moment x v xp0q and w xpk q for some k {1, . . . , j 1}. Then the loop resolves into the two independent bridges from v to w and from w to v , respectively. Therefore it suces to compute the expectation
117
a2 a1 a5
a0
a3
a4
{xpk q : k 0, . . . , j 1}, j
j 1
p0q
1 2
i,k 0,...j 1 i k
xpi qxpk q
n 1
pn pj n .
Lemma 5.22 (Probability of positivity of a random walk bridge). The probability of a random walk bridge of length n 1 is pn 1 n .
118
5.4. Convex Hulls in R2 Proof. The argument is standard in random polymers and relies on the fact that for a bridge Z pZk qk0,...n , pn
for any k . Since the latter is the probability that Z has its minimum at k , the claim follows. The combination of these results yields an explicit expression for the number of vertices of the convex hull of a random walk j -loop. Denote by 1 h2 pmq four times the partial sum of the harmonic series h2 pmq 4 m n1 n . Factor and index are motivated in the discussion of the higher dimensions below. Theorem 5.23 (Expected number of vertices for the typical j -loop in two dimensions). Let j N. Then the number of vertices of the convex hull of a j -loop is 1 0 P j p0 q h2 pj 1q. 2 Proof. By lemma 5.21,
0 P j p0 q j j 1
n 1
with pk
j n 1 . Because of j k n 1 j n , 0 P j p0 q j 1
pn pj n
n 1
1 n
j 1
n 1
1 jn
1 h2 pj 1q. 2
Thus the expected number of edges and vertices, respectively, grows like the logarithm of the length of the loop, which is faster than the result R enyi and Sulanke obtained in [RS63] for independently distributed points. They showed that the expected number of extremal points in the independent case grows like the square root of the logarithm of the number of points. Consequently one gets the expected number of vertices of the typical loop,
119
5. Geometric Aspects of the ideal Bose Gas Theorem 5.24 (Expected number of vertices for the typical loop in two dimensions). zj 0 P p q 2 h2 pj 1q. 0 z j 1d{2 j 1 Remark 5.25. These arguments apply in a similar manner to dimensions d 2, where edges have to be replaced by facets of the polytope conv V . Facets are dened by d vertices v1 , . . . , vk for which we write in the style of dimension two rv1 , . . . , vk s. In between these points are now d bridges instead of two in lemma 5.21, which is generalised straight forward. Lemma 5.26 (Facets of the convex hull of j -loops in higher dimensions). Let d 2 and x Xj , then
0 P j
p0q
2j d
pd1q j n1 pd2q
n1 1
n2 1
j n1 ... nd1 1 nd 1
1 n1
1 . n
d
Proof. As in the proof of lemma 5.21, x a starting point and subdivide the j steps into exactly d parts n1 , . . . , nd 1 with
j pd 1q n2 j n1 pd 2q . . . .. . nd j n1 . . . nd1 1.
n1 For xed n1 , . . . , nd , the probability that the d random walk bridges stay completely on one side of the hyperplane dened by d vertices is 2 1 n1 1 n .
d
Sum over all these partitions n1 , . . . , nd , and since each of the j points may occur as a starting point and each that way each partition is counted d times due to cyclic permutation, the claim follows.
120
5.4. Convex Hulls in R2 Unfortunately there seems to be no nice explicit formula apart from a generalisation given in lemma 5.26 with d 1 iterated sums, or, equivalently, a sum over all integer partitions of j consisting of d positive integers. But there is a possibility to obtain a recursion in d. Setting h1 : 2, then as in the proof of theorem 5.23,
0 P j
p0q
1 2
j 1
n1 1
1 j h1 n1 j n1
1 2
j 1
n1 1
h1
1 n1
1 1 j h2 pj 1q. n 2
1
For d 3 one additional bridge is inserted, and since only the orthogonal component matter, this case is obtained from the case d 2 as follows: Fix one bridge, which has length say n1 , then the remaining bridges have total length j n1 and
0 P j
p0q
2j 3 j 3
j 2 j n1 1 n1 1 n2 1 j 2
1 1 1 n1 n2 j n1 n 2
j n1 1 n2 1
n1 1 j 2
j 3
1 1 pj n1q n1 j n1
1 1 n2 j n1 n2
n1 1 j 2
1 1 h2 pj n1 1q n1 j n1 h2 pj n1 1q 1 n1 1 1 j 3 h3 pj 2q. n
1
1 3
n1 1
By continuing this recursion, the following result holds Proposition 5.27 (Facets of the convex hull of j -loops). Let x Xj be a random walk loop in Rd , then the expected number of facets of the convex hull satises the recursion in d
0 P j p0 q
1 hd j pd 1q , d
h1
2.
121
122
Figure 5.3.: Conguration of loops with clusters, each loop is shown with the corresponding disc size distribution). Two loops x, y of a conguration M pX q interact whenever there exist loops x0 , . . . , xn with x0 x, xn y and xk xk1 $ r. This interaction denes an equivalence relation on and thus we get clusters as connected components of interacting loops. Of particular interest will be the typical loop and the cluster the typical loop is contained in, which will be called the typical cluster. If this cluster is unbounded with positive probability, we say that percolation occurs, that is
0 P z ptypical cluster unboundedq 0.
We use the results of Gou er e [Gou08] to show that for suciently low fugacity the ideal Bose gas admits no loop percolation. Theorem 5.28. There exists z0 loop percolation.
admits no
Proof. For a loop x Xj let cx be the pair centre and radius of the smallest
123
5. Geometric Aspects of the ideal Bose Gas disc which contains x and has center xp0q, c:X
R2 R ,
Assume c to be continued on X . Furthermore continue c on M pX q by c x cx . Thus cP z is a Poisson process which realises circles with random radii, and if cP z admits no percolation, so P z does. Hence the job is to check whether the expected volume of a typical disc of cP z is nite for some z 0. Let B pBt qtr0,1s be a 2-dimensional Brownian motion and Yt : Bt tB1 . Then Y pYt qtr0,1s is a 2-dimensional Brownian bridge. Let M1 : suptr0,1s |Yt |, then by standard estimates
2 M1
if x Xj .
2 8
t 0,1
r s
sup
sup |Yt |
t 0,1 2
t 0,1
r s
1 1 sup Bt tB1 1 2 Bt .
r s
sup
1 1 Bt tB1 1 sup Bt
2 2 Bt tB1 2
t 0,1
r s
1 4 B1
t 0,1
r s
2 V follows. If M By Doobs L2 -inequality EM1 j is the corresponding 2 j EM 2 . maximum of a Brownian bridge on r0, j s, then by scaling EMj 1 Let r : R2 R R , py, q
be the projection of c on the second component, i.e. the radius of the disc and for a conguration M pR2 R q with p{0} R q 0 let r be the radius of the disc centered at the origin. Then
0 2 1 cP z pr q g12{2 pz q
2 EM1
g2 pz q
j 1 zj j 1
2 EMj 12{2
zj
1 pz q g2
g1 pz q EM12 g pzq
2
j 1
zj 2 j EM1 j2
124
EM12g1pzq.
Very interesting is that this criterion for non-percolation is independent of . This is due to the fact, that in two dimensions the loss of area of the spheres due to decreasing exactly compensates the gain of intensity or conversely, thinning compensates growing. This proof allows a stronger version, namely the typical cluster is not only bounded almost surely, its diameter D : sup sup |xpsq y ptq|
x,y S s,t
is integrable. The corresponding theorem is stated in the already mentioned paper of Gou er e [Gou08].
2 Proof. Because of EMj vious proof we get
0 2 cP q g2pzq1 z pr
2 1 2 EM1 g1 2
2 Similar to the previous proof we get EM2 V in using Doobs Lpinequality for p 2 . Therefore the rhs. of the equation above is nite for at least 4.
g2 pz q
zj 2 EMj j2
{ pz q .
125
Part III.
129
6. The P olya Sum Processes from that large set of colours, a reward is given to that colour introducing the P olya property. After the N -th draw they obtain a random measure on the set of colours of total mass N , which, if being normalised by N , converges as N V to a limiting random probability measure. The nite dimensional distributions of this limiting random probability measure are shown to be Dirichlet distributed. Both constructions are similar in their spirit. However, the latter construction of Blackwell and MacQueen is more general, as for a measure on the set of colours having atoms, the black ball may, with positive probability, introduce a ball of a colour already drawn. In the following sections the P olya sum process, a point process using these conditional constructions, is going to be constructed and some properties are determined. The basic measure on the set of colours is allowed to be -nite, but innite. In this chapter we rstly compute the Laplace funtionals of the P olya sum process in using the partial integration formula and derive dierent representations thereof. From that follows that the P olya sum process is innitely divisible. Furthermore we compute its Palm distribution.
m 0
1 m!
Bm
px1
. . . x qpmq
m
B c , dx1 , . . . , dxm .
Since the mapping B p, q is EB -measurable, proposition 1.31 applies and the P olya sum process is constructed as a point process with independent increments.
130
6.1. The Denition of the P olya Sum Process Denition 6.1 (P olya sum process). The P olya sum process Sz, for pz, q is the point process constructed from the Papangelou kernel in equation (6.1), explicitly for B B0 pX q and EB -measurable, non-negative Sz,,B pq : p1 z qpB q B px1
. . . x
m 0
1 m!
px1
m 1
. . . x q
m
The construction of the P olya sum process reveals the relation to the P olya urn: If the point x is drawn in one step, an additional weight of unit size is given to that point in the next and the following draws. The parameter z controls the total number of draws and ensures its niteness. Note that the choice of non-unit weights for drawn points can be reached by adjusting z and appropriately. For example choose X N, z p0, 1q and the counting measure. Then the P olya sum process realises at each n N a geometrically distributed number of points independently of the other sites. If the counting measure is replaced by an integer multiple of the counting measure, the geometric distribution is replaced by the corresponding negative binomial distribution. In general the integer multiple of the counting measure can be replaced by any positive multiple, and therefore the number of point at each site is in a generalised sense negative binomially distributed with a non-integral parameter. The fundamental property of the P olya sum process is that it solves the partial integration formula CSz, phq hpx, x qz pdxqSz, pdq,
which diers from the formula of the Poisson process only in the additional summand in the kernel. In fact it will turn out that Sz, shares many important properties with the Poisson process, such as complete randomness and innite divisibility.
131
p q u
.
m 0
z eu
pB qrms
m!
ze exp pB q log 1 1z
By the independence property this is extended to linear combinations and by monotone convergence to general continuous f with bounded support. Besides the complete randomness, from proposition 6.2 follows that the z P olya sum process is innitely divisible. Setting : 1 z results LSz, pf q exp
log 1
1 ef pxq pdxq ,
therefore Sz, is a gamma process-Poisson-mixture. The L evy-Khinchinrepresentation of the gamma process then yields Corollary 6.3 (Gamma-Poisson representation of the P olya sum process). LSz, pf q exp
p0,Vq
1 exp
s 1 ef pxq
pdsqpdxq
132
1 s e ds s The Gamma-Poisson representation expresses the representation of the negative binomial distribution as a Poisson distribution with gamma distributed intensity. An important second representation is the L evy-Khintchin representation of the P olya sum process, which is obtained by expanding the logarithm in proposition 6.2. pdsq Corollary 6.4 (L evy-Khintchin representation of the P olya sum process). LSz, pf q exp zj 1 ejf pxq pdxq . j
j 1
j 1 zj j 1
z j ejf pxq j
j 1
zj j
1 ejf pxq .
The L evy-Khintchin representation relates the P olya sum process with compound Poisson processes. That is, the P olya sum process Sz, can be recovered as the image of the Poisson process P z on X N with intensity measure zj j z : j
j 1
px,j q
px,j q
px,j q
jx .
(6.2)
133
6. The P olya Sum Processes Proposition 6.5. Let P z be the Poisson process on X N with intensity measure z for given z p0, 1q and MpX q. Then the P olya sum process Sz, for the pair pz, q is the image of P under the mapping MpX Nq z MpX q, px,j q px,j q px,j q jx .
The L evy-Khintchin representation of the P olya sum process Sz, in connection with the last remark about the niteness of z allows the immediate , which is given as the image of Sz, computation of the support process Sz, under the mapping . For simplicity the result is restricted to diuse . In analogy to the measures , which ensures the innite divisibility of Sz, Poisson process, this property is lost if has atoms (for the Poisson process this statement is trivial). Corollary 6.6 (Laplace functional of support process). Let be a diuse measure
pf q exp LSz,
logp1 zq.
pf q exp LSz,
X j 1
In case of d a with d being the diuse part and a being a non-vanishing atomic part of , both parts need to be treated separately. While for d the corollary above applies, a leads to a binomial part with term p1 zqp{x}q 1 p1 zqp{x}q ef pxq . This treatment is necessary since there is no possibility to distinguish if during the successive placement of the points a point is placed at an atom x of because of a parent at x or just by chance.
zj 1 ef pxq pdxq . j
134
z 1 z
j 1
z j Sz, x
Proof. Iterated application of the partial integration formula yields CSz, phq hpx, qpdxqSz, pdq
N j 1
zj
V
zN
zj
j 1
1B M pX q , i.e.
The immediate consequence is that the typical point has a geometrically distributed total mass whenever it is not an atom of .
135
is a Poisson Next we show in using the partial integration formula that Sz, process with intensity measure logp1 z q.
Proposition 6.8 (Partial integration formula for support process). Let MpX q be a diuse measure and z p0, 1q. Then
phq CSz,
Proof.
logp1 zq pdxq
phq CSz,
1X 0 pq
1X 0 p x q
Since the conguration x contains a point, the indicator equals one and
z
136
6.3. Disintegration and Partial Integration In the numerator of the second summand p x q because of the integration with respect to . Furthermore pxq 0 -a.s. in the denominator of the rst integrand. Therefore inductively follows
N j 1
zj
zN
V zj
j 1
h x, p x q pdxqSz, pdq
h x, p x q hpx, x q
logp1 zq pdxqSz,pdq
137
j 1
(7.1)
The P olya sum process can be recovered form P z as the image of the
139
7. Limit Theorems for Conditioned P olya Sum Processes mapping given in equation (6.2), M pX
Nq MpX q,
px,j q
px,j q
px,j q
jx .
(7.2)
Since the basic structure of the intensity measure z is very much in the spirit of the intensity measure of the Bose gas for d 0, large parts of the discussion are closely related to those of chapter 4. The object of interest is the P olya sum process conditioned on some tail -eld. Following the lines of chapter 4, the methods of the canonical loop ensemble in section 4.3 and the methods of the canonical ensemble of elementary components in section 4.5 apply up to minor modications to the Poisson process P z . Since a priori the image of the thermodynamic limits are not necessarily of P olya type, this has to be checked. Conditioning on the number of loops is analogue to conditioning on the number of turrets as well as conditioning on the number of elementary components accords conditioning on the number of building bricks.
pB q,
where B : B counts the support of a conguration , i.e. counts was shown to be a Poisson process the number of turrets in B . Since Sz, with intensity measure logp1 z q, the programme of section 4.3 can be adopted directly. If the local specication G is given by
G B p, q : Sz, |GB pq Sz, p
B q|B B ,
c
140
pdx1 q pdxn q (7.3) follow. Particularly equation (7.3) means that in B exactly n B towers of geometric size each are distributed independently. Besides Sz, , any P olya sum process Sz,m with 0 m V has the local specication G . Particularly C t : C p G q is not empty. Let pBk qk be an increasing sequence of bounded sets which exhausts X , GV : k GBk the tail- -eld, then for P-integrable , P C t ,
G P |GV pq lim B p, q k k
1z zpB q
n B n i ,...,i 1 n 1
pi1 x1
q,
t , as well as W the number of which is by construction an element of CV k turrets in Bk normalised by its volume
Wk :
Bk . pBk q
If W is the limit of Wk in case of existence, then from the results of section 4.3 can be deduced Proposition 7.1. Let f : X R be non-negative and measurable with bounded support and W pq V. Then for any P C t and L1 pPq
G Pp|GV q lim B p , q Sz,W pq n n
P-a.s.
(7.4)
141
7. Limit Theorems for Conditioned P olya Sum Processes Proof. As mentioned, the results for the Poisson process P z are going to be applied. Therefore with abuse of notation let B : M pX Nq N {V} the mapping which counts the number of points in B N and G the corresponding decreasing family of -algebras. Then by proposition 4.8 P z |GV pq P W pqz pq, which has the correct structure such that under the mapping (7.2) the result is a P olya sum process. By the reasoning of section 4.3, the extremal points of the Martin-Dynkin boundary C t are exactly those, for which W is almost surely constant and therefore
{Sz,w : 0 w V}.
142
7.2. The Brick Ensemble limit as the minimiser of a functional. Due to conditioning, two steps are necessary, rstly for the process without condition, given in corollary 2.8, and then with the correct condition. Let HB be the -algebra generated by the -algebra of the outside events EB and pB q, where B measures the total mass of a conguration , i.e. counts the number of bricks in B . Now follow the programme of section 4.5. If the local specication H is given by
H p, q : Sz, |HB pq Sz, p B
B q|B B ,
c
then immediately
H B p, q
where n B . Equation (7.5) reects the construction of the P olya sum process by means of conditional intensities. Apart from Sz I , , any P olya sum process Sz, with 0 z 1 has the local specication H . Particularly C b : C p H q is not empty and Martin-Dynkin boundary technique may be applied. Let pBk qk be an increasing sequence of bounded sets which exhausts X , HV : k HBk the tail- -eld, then for P-integrable , P C b , P |HV pq lim Bk p, q
k
P-a.s.
q,
b , as well as U the number of which is by construction an element of CV k bricks in Bk normalised by its volume
Uk :
Bk . pBk q
If U is the limit of Uk in case of existence, then from the results of section 4.5, particularly in connection with remark 4.18 can be deduced
143
7. Limit Theorems for Conditioned P olya Sum Processes Proposition 7.3. Let f : X R be non-negative and measurable with bounded support and U pq V. Then for any P C b , L1 pPq
G Pp|HV q lim B p , q SZ,pq n n
P-a.s.
(7.6)
In fact, Z depends on the conguration as well as on the choice of z I in the very beginning. The reparametrisation carried out in propositions 7.12 and 7.13 allows Z to be less than one in any case. Proof. Let H denote the specication obtained from the Poisson process P z I , B : P zI |HB with abuse of notation of the family H. Then by proposition 7.13 below for any P C p H q, Pp|HV q P Z with Z being the solution of
j 1
Zj
U.
Since U exists and is nite P-a.s., so Z does. Finally observe, that the P olya sum process SZ, is the image of P under the mapping (7.2). Z Remark 7.4. The discussion of minimiser of the rate function simplies since here there is no critical value present and hence no condensation eects occur. Theorem 7.5 (Martin-Dynkin boundary P olya sum process). The tail -eld HV is H-sucient for the family C b and the essential part of the Martin-Dynkin boundary consists exactly of the family b
{Sz, : 0 z 1}.
144
pB q pB q, B q|B B , B B ,
c
hence the number of turrets in B as well as their total height is known. Let the local specication I be given by
I B p, q : Sz, |IB pq Sz, p
then
I B p, q
1 pB qm
1
Bn k1 ... km n m
1
pdxm q pdx1 q, (7.7) where n B and m B . Here exactly m B towers have to be built from n B bricks, where each tower contains at least one brick. Therefore the local specication can be constructed by rst choosing the sites for the m towers and afterwards by placing each of the n m remaining bricks independently and uniformly at the given sites. At least the P olya sum process Sz I , has the local specication I , therefore C tb C p I q is not empty. Let pBk qk be an increasing sequence of bounded sets which exhausts X , IV : k IBk the tail- -eld, then for P-integrable , P C tb ,
m
1
k ,...km
mn m
nm k
P |IV pq lim I p, q
k
P-a.s.
q,
145
Proposition 7.6. Let f : X R be non-negative and measurable with bounded support and U pq V. Then for any P C tb and L1 pPq
G Pp|FV q lim B p , q SZ,W pq n n
P-a.s.
(7.8)
where W 1Z Z
U
W logp1 Z q V
P |IV pq.
Proof. The important part is to determine P zI , again with abuse of notation, conditioned on the tail- -algebra
|IB
with abuse of notation of the family I. Then by proposition 7.12 below for any P C p I q, Pp|IV q P W Z with W and Z being the solution of the pair of equations W
j 1
Zj
U,
j 1
Zj j
V.
146
7.4. Large Deviations Since the basic disintegration still holds true, only the limit of the mixing measure has to be determined, which is analogue to the procedure in section 4.5. Theorem 7.7 (Martin-Dynkin boundary P olya sum process). The taileld IV is H-sucient for the family C bt , and its extremal points are given exactly by all P olya sum processes for the pairs pz, wq, bt
{Sz,w : 0 z V, 0 w V}.
Proof. Follow the lines of the proof of canonical ensemble. The very rst step was the determination of the microcanonical limits, then the densities with large deviations.
j 1
z I pBk {j }qj
j 1
z Ij j j
then by the discussion following theorem 2.5 yielding corollary 2.8, P k satises a large deviation principle in the -topology, where the rate function is the relative entropy with respect to , I p; q pf log f
f 1q
if 3 , f : otherwise
d d , f
log f
f 1 L1p q ,
Next the general ensemble is considered in detail while the dierences occuring for the brick ensemble are mentioned. Let Du,v : { MpNq : Du : { MpNq : pj q v, jpj q u} jpj q u}
147
7. Limit Theorems for Conditioned P olya Sum Processes denote the measures on N with total mass v and rst moment u, and rst moment u, respectively. Then Du,v as well as Du are -closed, but not -open. Recall the setting A pq : hence Lemma 7.8 (Semicontinuous Regularisations of Du,v ). The upper and lsc lower semicontinuous regularisations usc Du,v and Du,v of Du,v with respect to -topology are usc Du,v pq V, lsc Du,v pq Du,v (7.9)
A , A
Lemma 7.9 (Semicontinuous Regularisations of Du ). The upper and lower lsc semicontinuous regularisations usc Du and Du of Du with respect to topology are usc Du pq V, lsc Du pq Du (7.10)
Both results are consequences of the fact that whenever a sequence of measures in MpNq converges with respect to the -topology, their total mass and their rst moment need to converge as well. From these two lemmas for each of the ensembles the upper bound of the particular large deviation principle follows directly, lim sup
k
V
k
u,v
(7.11) (7.12)
lim sup
Because of the previous lemmas, the superscript lsc can be dropped. Before we study the lower bound, the minimisation problem is solved for the general ensemble. Much work has already been done in section 4.5.
148
7.4. Large Deviations Proposition 7.10 (Minimiser of I Du,v ). Let 0 v wu,v be the solution of the system w zj j
w V and zu,v ,
(7.13)
j 1
v, lsc D
j 1
zj
u.
u,v
is given by
j zu,v . j
wu,v Proof. I pq
j 1
(7.14)
j 1
jpj q log z j
j 1
pj q log pj q
j 1
1 pNq
j 1
j 1
j 1
1 pNq,
u,v which has a unique minimiser on MpNq, j 1 wu,v j j with zu,v , wu,v being the solution of equation system (7.13). The uniqueness of the solution of the equation system (7.13) can be seen in the following way: Let
zj
then we have to show that f 1 pu, v q contains exactly one element, which is pzu,v , wu,v q. Consider both components of f separately. Then u f1 pz, wq w 1z z , v
(7.15) (7.16)
f2pz, wq w logp1 zq
149
7. Limit Theorems for Conditioned P olya Sum Processes implicitly dene two functions. Express z in terms of y : 1{w, then z1 py q uy , 1 uy z2 py q 1 evy
are two strictly increasing functions p0, Vq p0, 1q, for which the limits agree as y 0 and y V, respectively. Because of u v they intersect at exactly one point, which implies f 1 pu, v q {pzu,v , wu,v q}. The initially given parameter z I is contained in zu,v . In a similar fashion the minimisation problem for the brick ensemble is solved, and due to the missing condition on the number of towers, w drops out. Therefore Proposition 7.11 (Minimiser of I
D
zj
j 1
u.
u
lsc D
j 1
is given by
j zu . j
(7.18)
Note that in both cases weak and vague topology are too weak for Du,v and Du to be closed. In these cases the closure would contain at least measures with rst moment less or equal to u, which has a negative eect on the minimisation problems, particularly if u is larger than the rst moment of . As already seen in section 4.5, the Boltzmann principle helps to get the lower bound of the partition function, as will be the case here. The blowups of Du,v and Du in the -topology are now easier to handle. For 0
150
j 1
j 1
jpj q u
j 1
and D the blow ups. Since each of the complements is clearly closed, Du,v u are open, and since the conditions
V kV pBk q V kV
log P k exp
D D
u,v
1{D
u
u,v
L}
V,
(7.19) (7.20)
1{D L}
V,
hold by the non-negativity of A , by [DS00, Lemma 2.1.8] the large deviation lower bounds on the normalisation constants are given by lim inf 1 q inf log P I Du,v k exppDu,v kV pBk q MpNq 1 q inf log P lim inf I Du k exppDu kV pBk q MpNq (7.21) (7.22)
for every 0. Hence, for 0, the lower bound is obtained. In general are the existence of a minimiser can not be expected since Du,v and Du open. Nevertheless we construct a sequence which converges as 0.
). Let 0 v u V. For Proposition 7.12 (Minimiser of I Du,v suciently small 0 there exists a pair pzu,v, , wu,v, q p0, 1q p0, Vq such that the inmum of I Du,v on MpNq is attained at
wu,v,
j 1
j zu,v, . j
(7.23)
151
in -topology, where is given by equation (7.14) and zu,v and wu,v by equation (7.13). Proof. In the proof of proposition 7.10 we showed that for xed u v , the minimiser of I on Du,v was given by pz, wq w zj j
j 1
with pz, wq f 1 pv, uq and f given in equation (7.15). Here we have to minimise with respect to the parameters pz, wq f 1 pv , v q pu , u q , and therefore consider I as a mapping on p0, 1q p0, Vq. f is not only injective, but also continuous and maps open sets to open sets. Fix 0 such that 2 u v . Then the pre-image A : f 1 pv , v q pu , u q
p0, 1q p0, Vq
is open. If A can be shown to be bounded, then I has a minimiser z pq, wpq cl A . Therefore consider z pq, w pq : f 1 pv , u q, then z pq is an upper bound for z with pz, wq A as well as w pq is a lower bound for w with pz, wq A since as few turrets as possible have to be built with as many bricks as possible. Similarly z pq, w pq : f 1 pv , u q yields the corresponding lower and upper bound. Hence there exist parameters z pq and wpq for which I is minimal. As 0, we have to show that z pq, wpq pzu,v , wu,v q. Indeed, since f 1 is continuous, z pq zu,v and w pq wu,v .
). Let 0 u V. Then there Proposition 7.13 (Minimiser of I Du exists zu, r0, Vq such that the inmum of I Du on MpNq is attained
152
0
j 1
j zu, . j
(7.24)
zu and
lim
153
8. Concluding Remarks
In this thesis we studied two models: Firstly a Poisson process on a space of composite loops, and secondly the P olya sum process. For both we identied the structure of the process conditioned on various invariant -algebras. Since these -algebras were obtained as tail- -algebras of decreasing families of -algebras, the conditioned stochastic elds was obtained by a thermodynamic limit. The discussion of the Poisson process included geometric properties of the typical loop. Furthermore we discussed the fundamental Laplace functional as well as the Palm distribution for the P olya sum process. The most delicate part in the discussion of the limit theorems for the Bose gas was the canonical ensemble of elementary components where nally we obtained the limits by a large deviation argument. For the rst moment strange phenomenon occurred: When we computed the pointwise limits Q , we had to pay attention to the conguration . If the particle density U exceeded some critical value u , then we showed that U u Q -a.s. Hence some density got lost during the limiting procedure, but if we considered to be distributed according to some of the possible limits, then U is always bounded by u . An open problem is to collect the mass which was moved to longer and longer loops, such that the limit does not swallow the surplus mass. Furthermore working with a Poisson process means to work in the noninteracting case. More generally, the question about the point process including interaction between the elementary components of a single as well as of two dierent loops and conditioned on invariant -algebras remains open. The properties and the construction of the P olya sum process, especially
155
8. Concluding Remarks compared to those of the Poisson process, were very surprising. Both processes are completely random and innitely divisible. Moreover both processes can be constructed by partitioning the whole space and constructing on each set of the partition a nite point process independently of the other regions. Indeed, in general a Papangelou process can not be constructed in that way. Furthermore their Palm distributions agree in spirit: While the Palm distribution of the Poisson process is the process itself with an added point, the Palm distribution of the P olya sum process gets an extra point with geometrically distributed weight. A fundamental question arises: What is the structure of similar processes? The P olya sum process opens a vast eld of questions which we could not treat due to a lack of time. Connections to the work of Kingman should be revealed, for which the Gamma-Poisson representation could be an initial point. The Gamma-Poisson representation includes the fact that negative binomially distributed random variables can be represented as Poisson distributed with a gamma distributed intensity. More generally, the underlying gamma process may be replaced by any random measure. So how does this change aect the point process and its Papangelou property? The L evy-Khinchin representation of the Laplace functional established a fruitful connection to the Bose gas. Limit theorems for conditioned stochastic elds could be discussed analogously to those of the Bose gas. Finally, but not the nal question, increasing the parameter z means to increase the mean size of the turrets. By a neat normalisation, is there a suitable limit if z 1 and how is it characterised?
156
Bibliography
[AK07] S. Adams and W. K onig. Large deviations for many Brownian bridges with symmetrised initial-terminal condition. J. Probab. Theory Relat. Fields, 142:79124, 2007. T. W. Anderson. An Introduction to Multivariate Statistical Analysis. Wiley Series in Probability and Mathematical Statistics. John Wiley and Sons, 2 edition, 1984.
[And84]
[BCMP05] G. Benfatto, M. Cassandro, I. Merola, and E. Presutti. Limit theorems for statistics of combinatorial partitions with applications to mean eld Bose gas. J. of Math. Phys., 46:033303 033341, 2005. [BM73] [BU09] [DS00] [DVJ08a] D. Blackwell and J. B. MacQueen. Ferguson distributions via P olya urn schemes. Ann. Statist., 1:3535, 1973. V. Betz and D. Ueltschi. Spatial random permutations and innite cycles. Comm. Math. Phys., 285(2):469501, 2009. J.-D. Deuschel and D. W. Stroock. Large Deviations:. AMS Chelsea Publishing, 2 edition, 2000. D. Daley and D. Vere-Jones. An Introduction to the Theory of Point Processes, volume I: An Introduction to the Theory of Point Processes. Springer, 2 edition, 2008.
[DVJ08b] D. Daley and D. Vere-Jones. An Introduction to the Theory of Point Processes, volume II: General Theory and Structure. Springer, 2 edition, 2008.
157
Bibliography [Dyn71a] [Dyn71b] [Dyn78] [DZ98] [Fey48] [Fey53a] [Fey53b] [Fey90] [Fic91a] [Fic91b] [F ol75] [Gin71] E. B. Dynkin. Entrance and exit spaces for a Markov process. Actes Congres Intern. Math. 1970, 2:50712, 1971. E. B. Dynkin. The initial and nal behaviour of trajectories of a Markov process. Russian Math. Surveys, 26(4):16585, 1971. E. B. Dynkin. Sucient statistics and extreme points. Ann. Probab., 6(5):70530, 1978. A. Dembo and O. Zeitouni. Large Deviations: Techniques and Applications. Springer, 2 edition, 1998. R. P. Feynman. Space-Time Approach to Non-Relativistic Quantum Mechanics. Rev. Mod. Phys., 20(2):36787, 1948. R. P. Feynman. Atomic Theory of Helium Near Asolute Zero. Phys. Rev., 91(6):13018, 1953. R. P. Feynman. Atomic Theory of the Transition in Helium. Phys. Rev., 91(6):12911301, 1953. R. P. Feynman. Statistical mechanics: A set of lectures. Addison-Weslay, 13 edition, 1990. K. H. Fichtner. On the Position Distribution of the Ideal Bose Gas. Math. Nach., 151:5967, 1991. K. H. Fichtner. Random Permutations of countable sets. Probab. Theory Relat. Fields, 89:3560, 1991. H. F ollmer. Phase transition and Martin Boundary. Seminaire de probabilites (Strasbourg), 9:30517, 1975. J. Ginibre. Les Houches Summer School Theoret. Phys, chapter Some Applications of functional Integration in Statistical Mechanics. Statist. Mech. and Quantum Field Theory, pages 327427. Gordon and Breach, 1971.
158
Bibliography [Gou08] J.-B. Gou er e. Subcritical Regimes in the Poisson Boolean Model of Continuum Percolation. Ann. Probab., 36(4):1209 1220, 2008. M. Z. Guo and L. M. Wu. Several large deviation estimations for the Poisson point processes. Adv. in Math. (China), 24(4):313 319, 1995. P. Hall. On Continuum Percolation. Ann. Probab., 13(4):1250 1266, 1985. F. M. Hoppe. P olya-like urns and the Ewens sampling formula. J. Math. Biol., 20(1):914, 1984. K. Huang. Introduction to Statistical Physics. Wiley, 2 edition, 1987. O. Kallenberg. On Conditional Intensities of Point Processes. Z. Wahrscheinlichkeitstheorie verw. Geb., 41:20520, 1978. O. Kallenberg. Random Measures. Akademie-Verlag Berlin, 1983. O. Kallenberg. Foundations of Modern probability. Springer, 2 edition, 2002. J. F. C. Kingman. Completely Random Measures. Pacic J. Math., 21(1):5978, 1967. J. F. C. Kingman. Random partitions in population genetics. Proc. Roy. Soc. London Ser. A, 361(1704):120, 1978. J. F. C. Kingman. The representation of partition structures. J. London Math. Soc. (2), 18(2):37480, 1978. G. Kummer and K. Matthes. Verallgemeinerung eines Satzes von Sliwnjak. Rev. Roumaine Math. Pures Appl., 15:84570, 1970.
[GW95]
[Hal85] [Hop84] [Hua87] [Kal78] [Kal83] [Kal02] [Kin67] [Kin78a] [Kin78b] [KM70]
159
Bibliography [KMM74] J. Kerstan, K. Matthes, and J. Mecke. Unbgegrenzt teilbare Punktprozesse. Akademie-Verlag Berlin, 1974. [LS01] W. V. Li and Q.-M. Shao. Stochastic Processes: Theory and Methods, volume 19, chapter Gaussian Processes: Inequalities, Small Ball Probabilities and Applications, pages 53397. NorthHolland,Amsterdam, 2001. J. Mecke. Station are zuf allige Mae auf lokalkompakten Abelschen Gruppen. Z. Wahrscheinlichkeitstheorie verw. Geb., 9:3658, 1967. R. Meester and R. Roy. Continuum percolation. Cambridge Tracts in Mathematics 119. Cambridge University Press, 1996.
[Mec67]
[MR96]
[MWM79] K. Matthes, W. Warmuth, and J. Mecke. Bemerkungen zu einer Arbeit von Nguyen Xuan Xanh and Hans Zessin. Math. Machr., 88:11727, 1979. [NZ79] X. X. Nguyen and H. Zessin. Ergodic Theorems for Spatial Processes. Z. Wahrscheinlichkeitstheorie verw. Gebiete, 48:133 58, 1979. X. X. Nguyen and H. Zessin. Martin-Dynkin boundary of mixed Poisson processes. Z. Wahrscheinlichkeitstheorie verw. Gebiete, 37(3):191200, 1976/77. C. Preston. Canonical and Microcanonical Gibbs States. Z. Wahrscheinlichkeitstheorie verw. Gebiete, 46:12558, 1979. A. R enyi and R. Sulanke. Uber die konvexe H ulle von n zuf allig gew ahlten Punkten. Z. Wahrscheinlichkeitstheorie verw. Gebiete, 2:7584, 1963. D. Revuz and M. Yor. Continuous Martingales and Brownian Motion. A Series of Comprehensive Studies in Mathematics 293. Springer-Verlag, 1991.
[NZ77]
[Pre79] [RS63]
[RY91]
160
Bibliography [RZ93] S. Rlly and H. Zessin. The Equivalence of Equilibrium Principles in Statistical Mechanics and some Applications to Large Particle Systems. Expo. Math., 11:385405, 1993. M. Stroppel. Locally Compact Groups. EMS textbooks in Mathematics. EMS, 2006. A. S ut o. Percolation transition in the Bose Gas. J. Phys. A: Math. Gen., 26:4689710, 1993. A. S ut o. Percolation transition in the Bose Gas: II. J. Phys. A: Math. Gen., 35:69957002, 2002. N. Tsilevich. Distribution of cycle lengths of innite permutatons. J. Math. Sci. (N.Y.), 87(6):407281, 1997. D. Ueltschi. Feynman cycles in the Bose Gas:. J. Math. Phys., 47(12):123303, 15pp., 2006. D. Ueltschi. Relation between Feynman cycles and o-diagonal long-range order. Phys. Rev. Lett., 97(17):170601, 4pp., 2006. D. Ueltschi. The Model of interacting spatial permutations and its relation to the Bose gas. page arXiv: 0712.2443v3, 2008. A. Vershik and A. Schmidt. Limit measures arising in the asymptotic theory of symmetric groups I. Theor. Prob. Appl., 22:7985, 1977. A. M. Vershik and A. A. Schmidt. Limit measures arising in the asymptotic theory of symmetric groups II. Theor. Prob. Appl., 23:3649, 1978. W. von Waldenfels. Charakteristische Funktionale zuf alliger Mae. Z. Wahrscheinlichkeitstheorie verw. Gebiete, 10:27983, 1968.
[VS78]
[vW68]
161
Bibliography [Weg77] H. Wegmann. Characterization of Palm Distributions and Innitely Divisible rnadom measures. Z. Wahrscheinlichkeitstheorie verw. Gebiete, 39:25762, 1977. H. Zessin. Der Papangelou Prozess. Journal of Contemporary Mathematical Analysis, 44(1):3644, 2009.
[Zes09]
162
This thesis considers on the one hand the construction of point processes via conditional intensities, motivated by the partial Integration of the Campbell measure of a point process. Under certain assumptions on the intensity the existence of such a point process is shown. A fundamental example turns out to be the Plya sum process, whose conditional intensity is a generalisation of the Plya urn dynamics. A Cox process representation for that point process is shown. A further process considered is a Poisson process of Gaussian loops, which represents a noninteracting particle system derived from the discussion of indistinguishable particles. Both processes are used to dene particle systems locally, for which thermodynamic limits are determined.
ISBN 978-3-86956-029-8