Classical and non-classical limit theorems for the Ising model on random graphs
Citation for published version (APA): Prioriello, M. L. (2016). Classical and non-classical limit theorems for the Ising model on random graphs. Technische Universiteit Eindhoven.
Document status and date: Published: 27/01/2016
Document Version: Publisher’s PDF, also known as Version of Record (includes final page, issue and volume numbers)
Please check the document version of this publication:
• A submitted manuscript is the version of the article upon submission and before peer-review. There can be important differences between the submitted version and the official published version of record. People interested in the research are advised to contact the author for the final version of the publication, or visit the DOI to the publisher's website. • The final author version and the galley proof are versions of the publication after peer review. • The final published version features the final layout of the paper including the volume, issue and page numbers. Link to publication
General rights Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.
• Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain • You may freely distribute the URL identifying the publication in the public portal.
If the publication is distributed under the terms of Article 25fa of the Dutch Copyright Act, indicated by the “Taverne” license above, please follow below link for the End User Agreement: www.tue.nl/taverne
Take down policy If you believe that this document breaches copyright please contact us at: [email protected] providing details and we will investigate your claim.
Download date: 25. Sep. 2021 CLASSICALANDNON-CLASSICALLIMITTHEOREMS FORTHEISINGMODELONRANDOMGRAPHS
Maria Luisa Prioriello This work was financially supported by The Netherlands Organization for Scientific Research (NWO) through the VICI grant of Remco van der Hofstad.
© Maria Luisa Prioriello, 2016.
Classical and non-classical limit theorems for the Ising model on random graphs by M. L. Prioriello - Eindhoven University of Technology, 2016.
A catalogue record is available from the Eindhoven University of Technology Library. ISBN: 978-90-386-4004-4
Cover design: Caroline Ashwood, “Copper tree”.
Printed by Gildeprint Drukkerijen, Enschede. Classical and non-classical limit theorems for the Ising model on random graphs
proefschrift
ter verkrijging van de graad van doctor aan de Technische Universiteit Eindhoven, op gezag van de rector magnificus, prof.dr.ir. F.P.T. Baaijens, voor een commissie aangewezen door het College voor Promoties in het openbaar te verdedigen op woensdag 27 januari 2016 om 16.00 uur
door
Maria Luisa Prioriello
geboren te Isernia, Italië
iii Dit proefschrift is goedgekeurd door de promotoren en de samenstelling van de promotiecommissie is als volgt:
voorzitter: prof. dr. ir. B. Koren 1e promotor: prof. dr. R. W. van der Hofstad 2e promotor: prof. C. Giardinà (Università di Modena e Reggio Emilia) copromotor: prof. C. Giberti (Università di Modena e Reggio Emilia) leden: prof. dr. A. C. D. van Enter (University of Groningen) prof. dr. J. S. H. van Leeuwaarden dr. F. R. Nardi prof. S. Polidoro (Università di Modena e Reggio Emilia)
Het onderzoek dat in dit proefschrift wordt beschreven is uitgevoerd in overeen- stemming met de TU/e Gedragscode Wetenschapsbeoefening.
iv ACKNOWLEDGMENTS
This thesis is the result of four years of research carried out between the Universities of Modena and Eindhoven. There are many people who deserve my gratitude for having supported me along this path. First of all, my special thanks go to my three supervisors. Thanks for having been such an edifying example of excellent researchers, thanks for the wonderful synergy among you that allowed me to work and grow up in a comfortable envi- ronment and thanks to your perseverance that made this double doctoral degree project possible. I want to thank Claudio for his patience and kindness: when I had a question, the door of your office was always open. Thank you for having helped me to pursue my research, putting beside your strong mathematical rigor a good dose of humor. Thank you Cristian, for having driven me to accomplish my PhD with great positivity and pragmatism. I admire your eagerness and passion for the research. Remco, thank you for having given me the opportunity to work with you. You are one of the most brilliant mind that I ever met, always smiling and polite, with a contagious enthusiasm for mathematics. I am indebted to each of you. I would like to express my gratitude to Aernout van Enter, Johan van Leeuwaar- den, Francesca Nardi and Sergio Polidoro for agreeing to be part of my doctoral committee, for reading my thesis and for their useful comments. I am also grateful to my master thesis supervisor Silvana De Lillo for having encouraged me to enroll the PhD program and for having continued our pleasant collaboration during these years. I am indebted with my co-author Sander Dommers for his significant contribute on the subjects of Chapter 6 and for the fruitful discussions. I would like to thank my delightful housemates Alessandro G. e Melania. Thanks for making the house such a cheerful and loving place. I enjoyed a lot the amazing time spent together and the great harmony among us. I could not meet better housemates than you are. A big thank you also to my housemates in Bologna, especially to the people of Casa Molise for being such a family, for the great feasts and the endless laughs. I wish to thank Alessandro Z. e Carlo, the people who welcomed me here, who have always supported me giving me thousands of advices and suggestions on any matter. I would like to show my gratitude to all my colleagues of the Stochastics section for the awesome atmosphere and their great company. In particular, thanks to the Italian group: Enrico, Alessandro G., Lorenzo, Marta, Gianmarco, Alessandro Z., Fabio and Carlo for the wonderful time we have had together enjoying dinners, barbecues and trips.
v I also had a pleasant time in the office with Angelos, Jaron, Rick, Serban and Thembile. Thanks for tolerating my chitchat and my video calls. In particular, thanks to Jaron for the useful conversations and for the style tips about my thesis. Many thanks also to my colleagues in Modena, especially Carlos and Gioia, for their cheerful company during our travels and conferences. A kind thought is devoted to my graduate fellows Giacomo e Irma for carrying forward our friendship despite the distance. Thanks for coming to visit me and for showing always your support. I am immensely grateful to my best friends Fiorella, Giovanna and Noemi. Each one of you knows how much I owe to you and how much you are indispensable to me. There is no need to list all the reasons for which I should like to thank you. One of my biggest thanks is for my family. Thanks to my parents Mario and Patrizia, my brother Antonio e my sister Serena to be always present in my life. You continuously encourage and support me, expressing your interest and curiosity and making sure that everything goes well with your help. Without your love I would never have arrived this far. It is wonderful to see you so proud of me. I would also to express my gratitude to zia Angela and zio Pietro for their affection and care. Finally, my last acknowledgment is dedicated to you, Dino. Thanks for being so close to me with your love, especially when we are far away. Thanks for your constant reassurance, for your sweet carefulness and for managing always to get a smile out of me. You are my strength and I am extremely lucky to have you by my side.
vi CONTENTS
1 Introduction and motivations 1 2 Ising model on random graphs 5 2.1 Random graph models ...... 5 2.1.1 Locally tree-like random graph ...... 5 2.1.2 Configuration model ...... 7 2.1.3 Generalized random graph ...... 8
2.1.4 Local tree-likeness of CMN (d) and GRGN (w) ...... 9 2.1.5 Phase transition in random graphs ...... 9 2.1.6 Why these graphs? ...... 10 2.1.7 Other random graph models ...... 10 2.2 Ising model ...... 11 2.3 Measures and thermodynamic quantities for the Ising model on random graphs ...... 12 2.4 Phase transition and critical behavior ...... 15 2.5 Preliminaries ...... 16 2.5.1 State of the art ...... 17 2.5.2 Correlation inequalities ...... 17 2.5.3 Thermodynamic limits for the random quenched law . . . 18 2.5.4 Critical exponents for the random quenched law . . . . . 20 2.6 Other spin models ...... 21 2.6.1 The inhomogeneous Curie–Weiss model ...... 21 2.6.2 The Potts model ...... 22 2.6.3 The XY model ...... 22 2.7 Overview ...... 23 2.7.1 Discussion on the measures ...... 23 2.7.2 Overview of results ...... 25 2.7.3 CLT proof strategy ...... 26 2.7.4 Differences between random quenched and averaged quenched setting ...... 28 2.7.5 Properties of annealing ...... 28 2.7.6 Universality classes for power-law random graphs . . . . . 29 2.7.7 Violation of CLT at βc ...... 29 2.7.8 Low-temperature region ...... 30 2.8 Organization of the thesis ...... 30 3 Random quenched measure 33 3.1 LLN and CLT for locally tree-like random graphs ...... 33 3.2 Proofs ...... 34 3.2.1 Exponential convergence ...... 34 3.2.2 Random Quenched SLLN: Proof of Theorem 3.1.1 . . . . 35 3.2.3 Random Quenched CLT: Proof of Theorem 3.1.2 . . . . . 36
vii Contents
4 Averaged quenched measure 39 4.1 Thermodynamic limits and LLN for locally tree-like random graphs 39
4.2 CLT for CMN (2) ...... 40
4.3 CLT for CMN (1, 2) ...... 41 4.4 Proofs for locally tree-like random graphs ...... 41 4.4.1 Averaged quenched thermodynamic limits: Proof of Theo- rem 4.1.1 ...... 41 4.4.2 Averaged quenched WLLN: Proof of Theorem 4.1.2 . . . . 42
4.5 Proofs for CMN (2) ...... 43 4.5.1 Partition functions for the one-dimensional Ising model . 43 4.5.2 Quenched pressure ...... 44 4.5.3 Cumulant generating functions ...... 46 4.5.4 Averaged quenched CLT: Proof of Theorem 4.2.1 . . . . . 47
4.6 Proofs for CMN (1, 2) ...... 51 4.6.1 Quenched pressure ...... 52 4.6.2 Averaged quenched CLT: Proof of Theorem: 4.3.1 . . . . . 54 5 Annealed measure: uniqueness regime 65
5.1 Thermodynamic limits, SLLN and CLT for GRGN (w) ...... 65
5.2 Thermodynamic limits, SLLN and CLT for CMN (2) ...... 67
5.3 Thermodynamic limits, SLLN and CLT for CMN (1, 2) ...... 68
5.4 Proofs for GRGN (w) ...... 70 5.4.1 Annealed thermodynamic limits: Proof of Theorem 5.1.1 . 71 5.4.2 Annealed SLLN: Proof of Theorem 5.1.2 ...... 76 5.4.3 Annealed CLT: Proof of Theorem 5.1.3 ...... 76
5.5 Proofs for CMN (2) ...... 77 5.5.1 Annealed thermodynamic limits: Proof of Theorem 5.2.1 . 77 5.5.2 Annealed SLLN: Proof of Theorem 5.2.2 ...... 78 5.5.3 Annealed CLT: Proof of Theorem 5.2.3 ...... 78
5.6 Proofs for CMN (1, 2) ...... 83 5.6.1 Annealed partition function ...... 84 5.6.2 Annealed thermodynamic limits: Proof of Theorem 5.3.1 . 99 5.6.3 Annealed SLLN: Proof of Theorem 5.3.3 ...... 100 5.6.4 Annealed CLT: Proof of Theorem 5.3.4 ...... 100 5.6.5 Annealed pressure (way 2): Proof of Theorem 5.3.2 . . . . 101 6 Annealed measure: at criticality 107 6.1 Weight sequences conditions ...... 107
6.2 Critical exponents for GRGN (w) ...... 108
6.3 Non-classical limit theorem for GRGN (w) ...... 109 6.4 Proofs of Theorem 6.2.1: Annealed critical exponents ...... 111 6.4.1 Magnetization: critical exponents β and δ ...... 112 6.4.2 Susceptibility: critical exponents γ and γ0 ...... 120 6.5 Proofs of Theorem 6.3.1: Non-classical limit theorem ...... 129 6.5.1 Rewrite of the moment generating function ...... 130 6.5.2 Convergence for E[W 4] < ∞ ...... 131
viii Contents
6.5.3 Convergence for τ ∈ (3, 5) ...... 135 6.5.4 Proof of Theorem 6.3.1 ...... 139 6.5.5 Scaling window ...... 144 7 Zero temperature recursion 147 7.1 Analytical solution ...... 147 7.1.1 Fixed point at zero temperature and zero external field . 148 7.1.2 Fixed point at zero temperature and positive external field 149 7.1.3 Fixed point approaching zero temperature ...... 151 7.2 Numerical analysis ...... 153 7.2.1 Zero temperature and zero external field case ...... 154 7.2.2 Zero temperature and positive external field case . . . . . 155 7.2.3 Non-zero temperature case ...... 159
Bibliography 161
ix
INTRODUCTIONANDMOTIVATIONS 1
The Ising model, named after Ernst Ising, is a mathematical model introduced to understand ferromagnetism. The model was suggested to Ising for his Ph.D thesis [41] by his adviser Lenz in the 1920s [57, 58, 59]. The idea is to represent the interaction between atoms of a magnetic material. Each atom has a magnetic moment, typically represented by an arrow, which is allowed to point either up or down and flips between the two orientations. The interactions among atoms decay rapidly with distance, so in general, only the influence on neighboring elements is taken into account. In addition an external magnetic field, which tends to align the moments of the atoms in its direction, may be present. The classical mathematical model lives on a lattice and a particle is assigned to each lattice site. The magnetic moment (spin) of a particle is restricted to take only two values: +1 or −1.
Ising phase transition. An interesting feature of the Ising model is the presence of a phase transition, which in general occurs when a small change in parameters, such as temperature or pressure, causes a qualitative change in the state of the system. Phase transitions are common in nature and in everyday life, for example, a liquid becomes a gas above a fixed temperature, or freezes when the temperature goes down. In ferromagnetic materials the phase transition can appear when, below a certain temperature, the system becomes magnetized because a fraction of atoms larger than 1/2 are magnets pointing in a particular direction, while this magnetization is lost at sufficiently high temperatures. In his study, Ising discovered that the one-dimensional model does not exhibit a phase transition and he asserted, erroneously, that there was no phase transition in any dimension.
Ising models and cooperative phenomena. The interactions between the spin random variables are crucially determined by the spatial structure where the spin variables are sitting. Besides regular lattices, in recent years much attention has been devoted to the setting in which the spins variables are placed on the vertices of random graphs. These models can be studied as a paradigmatic model for cooperative phenomena living on complex networks. The behavior of the whole system can be understood as the consequence of the cooperation between units of the system. This turns out to be useful for analyzing problems in fields as diverse as biology, economy, neuroscience, social science, opinion formation. We can think, for example, at the mapping between the dynamics of the Ising model and some processes on brain networks [30, 40, 61, 63]. Our brain consists of approximately 1010 neurons connected in a very complex network, called
1 Introduction and motivations neural network. Neurons communicate with other neurons via electric signals. At a fixed time, a neuron either can or cannot generate an electric signal, i.e. we say that it has two states: it is active or it is at rest. It turns out that this system can be described by an Ising model, where the external field is replaced by the average of emitting electric signal for each cell. An other instance is the dynamic of the opinion formation in a social system [24, 25, 33, 31, 32]. We consider a population of N individuals and we associate at each individual a variable which can take only two values. This variable represents the opinion, being in favor or against, about a certain topic, i.e. vote yes/no in a referendum. The individuals may change their opinions after the interaction with others. The strength of this interaction could be positive or negative, favors agreement or disagreement and it could be stronger among particular individuals. We can also take an additional external force that favors one of the two opinions into account, like the influence of media, government, existing culture. This kind of model has been implemented to study phenomena of interaction between two different cultures [11, 10]. In this case the whole population is divided into two homogeneous groups with interactions that are stronger and imitative between the individuals of each group, whereas they could be even counter-imitative between individuals belonging to different groups. Processes like this one do not live on lattices, but more likely they take place on networks that could be formed, for example, by a set of people (the vertices of the network) connected by links that represent friendships, or some other form of social interaction, between individuals. Networks of this type are called social networks.
Random graphs as network models. In the real world a large number of examples of complex networks exist [55, 56]. Beyond the social networks, we can find information networks (e.g., World Wide Web, citation networks between academic papers), technological networks (e.g., Internet, transportation networks such as airline routes, road and rail networks) and biological networks (e.g., protein interaction network, neural networks). Many of these real-world networks share features. First of all, these networks are large and they grow in size as time proceeds. Nevertheless, a small-world behavior appears, meaning that most pairs of vertices are connected by relatively short chains of links. Another interesting characteristic is that the networks are scale free, meaning that the majority of vertices have “few” connections, whereas the number of vertices with high connections decays slowly. We model networks using random graphs. We choose random graph models that possess particular properties of interest, such as specified degree distributions, to describe a given network. A simple example of a random graph is obtained by fixing a number of vertices and adding connections between them with a certain probability. Alternatively we can get a random graph growing in size adding vertices successively and attaching or removing connections. There exist
2 Introduction and motivations different ways to build a random graph, each of which leads to a graph with different properties [3, 43, 64, 65].
Ising models on random graphs. Ising models on random graphs are also an ideal probabilistic model to study dependent random variables in the presence of two sources of randomness. Indeed they possess a dependence structure between the Ising random variables and an extra level of spatial disorder given by the random graph. Taking this double randomness into account in various ways, three different measures arise. For a given assignment of the graph the Random Quenched measure coincides with the random Boltzmann–Gibbs distribution on that random graph. This means that we just look at the evolution of the dynamic of the spin towards an equilibrium state whereas the environment is frozen. If we think of the example of opinion formation, the network of the social interactions is fixed and the individuals just change their opinion according to that of their neighbors. The Averaged Quenched and the Annealed measures are obtained by averaging over all possible random graphs in two different ways. These two measures arise in the presence of a dynamical evolution of the environment. In particular, in the annealed setting the dynamics of the graph is much faster than the dynamics of the spins, i.e. the friendship network changes much more quickly than the opinion of the individuals. Another interesting feature that emerges from the study of the Ising model on random graphs is the prominent role of the inhomogeneity of the graph models. This inhomogeneity, that concerns the variable numbers of connections of vertices, is not present in the analysis on lattice and it has to be taken in account in our investigations.
Contribution of this thesis. The aim of the thesis is to study asymptotic theorems for spin variables of the Ising model on random graphs. In particular, we look at the behavior of the sum of the spins under the three different measures mentioned in the previous paragraph. Asymptotic theorems are relevant to understand properties of the system when its size becomes large, as for the models under consideration. We highlight that the study of these theorems in this setting is particularly challenging because we are dealing with dependent random variables. Our analysis is inspired by the results proved by Ellis for the Ising model on Zd (see [26]). The first step is to prove the law of large numbers. We will show the convergence of the average of the sum of the spins to their expected value. The rate of convergence will be different for the different measures. Then, by analyzing the fluctuations around the expected value, we prove the central limit theorem. For the random quenched setting, the result holds for a generic class of random graphs. In the averaged quenched and annealed case, instead, we have to limit our analysis to particular random graph models in
3 Introduction and motivations which explicit calculations are possible. Indeed, in these two settings we have to take into account the contribution of the graph and we can not replicate the general strategy used with the random quenched measure. We emphasize that we restrict these theorems to the set of the parameters such that there exists a unique Gibbs measure, i.e., the uniqueness regime, outside the critical region. Finally we analyze what happens at the critical point. Here the central limit theorem breaks down and a different scaling of the total spin is needed to obtain a non-classical limiting distribution.
Organization. This thesis is structured in the following way. In the next chapter we introduce the spin models, the random graphs, the main definitions and we provide a general overview on this work. Chapters 3, 4 and 5 are organized according to the three different measures and they collect results proved in [37, 36]. For each measure we present the asymptotic theorems for the random graphs under consideration outside the critical region. Chapter 6, based on [19], concerns the study at criticality: we show results about critical exponents and a non-classical limit theorem. Chapter 7 investigates a recursion equation that arises for the Ising model on random trees. For a more extensive description of the organization of the thesis see Section 2.8.
4 ISINGMODELONRANDOMGRAPHS 2
2.1 Random graph models
In this section we present the spatial structures, i. e., the random graphs, on which our spin model is defined. We first introduce the general class of locally tree-like random graphs, where the local structure around any vertex is given by a random tree. Then we define two specific models, namely the configuration model and the generalized random graph.
We denote by GN = (VN , EN ) a random graph with vertex set VN = [N] and edge set EN ⊂ VN × VN and by GN the set of all possible graphs with N vertices. Here and in the rest of this thesis [N] = {1, ... , N} will be used to denote the vertex set of GN .
2.1.1 Locally tree-like random graph
The class of models that we consider here is the same as defined in [16, 20].
Namely, we study random graph sequences (GN )N≥1 that are assumed to be locally like a homogeneous random tree, uniformly sparse and whose degree dis- tribution has a strongly finite mean. In order to formally state these assumptions, we need to introduce some notation.
Let the integer-valued random variable D have distribution P = (pk)k≥1 , i.e., P (D = k) = pk, for k = 1, 2, ... . We define the size-biased law ρ = (ρk)k≥0 of D by (k + 1) p ρ = k+1 , (2.1.1) k E [D] where the expected value of D is supposed to be finite. Let K be a random variable with P (K = k) = ρk and let ν be the average value of K, i.e.,
X E [D (D − 1)] ν := kρ = , (2.1.2) k E [D] k≥0 that will play an important role in what follows. The random rooted tree T (D, ρ, `) is a branching process with ` generations, where the root offspring has distribution D and the vertices in each next genera- tion have offspring that are independent and identically distributed (i.i.d.) with distribution ρ. We write that an event A holds almost surely (a.s.) if P (A) = 1. The ball
Bi (r) of radius r around vertex i of a graph GN is defined as the graph induced by the vertices at graph distance at most r from vertex i. For two rooted trees
5 Ising model on random graphs
∼ T1 and T2, we write that T1 = T2, when there exists a bijective map from the vertices of T1 to those of T2 that preserves the adjacency relations.
Definition 2.1.1 (Local convergence to random trees). Let PN denote the law induced on the ball Bi (t) in GN centered at a uniformly chosen vertex i ∈ [N] = {1, ..., N}. We say that the graph sequence (GN )N≥1 is locally tree- like with asymptotic degree distributed as D when, for any rooted tree T with t generations, a.s., ∼ ∼ lim PN (Bi (t) = T ) = P (T (D, ρ, t) = T ) . (2.1.3) N→∞ This property implies, in particular, that the degree of a uniformly chosen vertex in GN is asymptotically distributed as D.
Definition 2.1.2 (Uniform sparsity). We say that the graph sequence (GN )N≥1 is uniformly sparse when a.s., 1 X 1 lim lim sup di {di≥`} = 0, (2.1.4) `→∞ N→∞ N i∈[N] where di is the degree of vertex i in GN and 1A denotes the indicator of the event A. A consequence of local convergence and uniform sparsity is that the number of edges per vertex converges a.s. to E[D]/2 as N goes to infinity. Definition 2.1.3 (Strongly finite mean degree distribution). We say that the degree distribution P = (pk)k≥1 has strongly finite mean when there exist constants τ > 2 and c > 0 such that ∞ X −(τ−1) pi ≤ ck . (2.1.5) i=k In Chapter 6 we focus our attention on graphs where the degree distribution satisfies a power law, defined in the following definition.
Definition 2.1.4 (Power laws). We say that the distribution P = (pk)k≥1 obeys a power law with exponent τ when there exist constants Cp > cp > 0 such that, for all k = 1, 2, ... −(τ−1) X −(τ−1) cpk ≤ pi ≤ Cpk . (2.1.6) i≥k Power law distributions describe settings where nodes of the graph have a highly variable number of neighbours. The effect of this inhomogeneity will be relevant in particular in the analysis at the critical point, as we will see in the following.
Two specific examples of random graph models belonging to the class defined by some of these properties will be considered in detail in the following subsections: the configuration model and the generalized random graph under certain regularity conditions.
6 2.1 Random graph models
2.1.2 Configuration model
The configuration model is a multigraph, that is, a graph possibly having self- loops and multiple edges between pairs of vertices, with fixed degrees. Fix an integer N and consider a sequence of integers d = (di)i∈[N]. The aim is to construct an undirected multigraph with N vertices, where vertex j has degree dj. We assume that dj ≥ 1 for all j ∈ [N] and we define the total degree X `N := di. (2.1.7) i∈[N]
We assume `N to be even in order to be able to construct the graph. Assuming that initially dj half-edges are attached to each vertex j ∈ [N], one way of obtaining a multigraph with the given degree sequence is to pair the half-edges belonging to the different vertices in a uniform way. Two half-edges together form an edge, thus creating the edges in the graph. To construct the multigraph with degree sequence d, the half-edges are numbered in an arbitrary order from 1 to `N . Then we start by randomly connecting the first half-edge with one of the `N − 1 remaining half-edges. Once paired, two half-edges form a single edge of the multigraph. We continue the procedure of randomly choosing and pairing the half-edges until all half-edges are connected, and call the resulting graph the configuration model with degree sequence d, abbreviated as CMN (d). We will consider, in particular, the following models:
(1) The 2-regular random graph, denoted by CMN (2), which is the configura- tion model with di = 2 for all i ∈ [N].
(2) The configuration model with di ∈ {1, 2} for all i ∈ [N], denoted by
CMN (1, 2), in which, for a given p ∈ [0, 1], we have N − bpNc vertices with degree 1 and bpNc vertices with degree 2.
Remark 2.1.1. The configuration model CMN (1, 2) can be implemented by assigning to each vertex degree 2 with probability p or degree 1 with probability 1 − p, conditioned to having bpNc vertices of degree 2. Another configuration model would be obtained by considering the independent Bernoulli assignment to each vertex. This yields a random graph that has bpNc vertices of degree 2 only on average.
The degree sequence of the configuration model CMN (d) is often assumed to satisfy a regularity condition, which is expressed as follows. Denoting the degree of a uniformly chosen vertex UN ∈ [N] by DN = dUN , we assume that the following property is satisfied:
Condition 2.1.1 (Degree regularity). There exists a random variable D with finite second moment such that, as N → ∞,
D (a) DN −→ D,
7 Ising model on random graphs
(b) E[DN ] → E[D] < ∞,
2 2 (c) E[DN ] → E[D ] < ∞, where −→D denotes convergence in distribution. We call D the asymptotic degree. Further, we assume that P(D ≥ 1) = 1.
As we will see in Section 2.1.4, the properties stated in Condition 2.1.1 are crucial to prove that the configuration models handled in this thesis belong to the general class of locally tree-like random graphs. Moreover, the asymptotic degree D is central to detect the existence of a phase transition.
2.1.3 Generalized random graph
In the generalized random graph, each vertex i ∈ [N] receives a weight wi. Given the weights, edges are present independently, but the occupation probabilities for different edges are not identical, but are rather moderated by the weights of the vertices. For a given sequence of weights w = (wi)i∈[N], the graph is denoted by GRGN (w). We call Iij the Bernoulli indicator that the edge between vertex i and vertex j is present and pij = P (Iij = 1) is equal to
wiwj pij = , (2.1.8) lN + wiwj
PN where lN is the total weight of all vertices given by lN = i=1 wi. The degree sequence of the generalized random graph GRGN (w) is often assumed to satisfy a regularity condition, which is expressed as follows. Denoting WN = wUN the weight of an uniformly chosen vertex UN ∈ [N], we introduce the following property:
Condition 2.1.2 (Weight regularity). There exists a random variable W with finite second moment such that, as N → ∞,
D (a) WN −→ W ,
(b) E[WN ] → E[W ] < ∞,
2 2 (c) E[WN ] → E[W ] < ∞, where −→D denotes convergence in distribution. We call W the asymptotic weight. Further, we assume that E[W ] > 0.
In the following, we will consider deterministic sequences of weights that satisfy Condition 2.1.2. In many cases, one could also work with weights w = (wi)i∈[N] that are i.i.d. random variables. When the weights are themselves random variables, they introduce a double randomness in the random graphs: firstly there is the randomness introduced by the weights, and secondly there is the
8 2.1 Random graph models randomness introduced by the edge occupation statuses, which are conditionally independent given the weights. The generalized random graph is an example of an inhomogeneous random graph, a large class of models in which the edge probabilities are not identical, but moderated by certain vertex weights. Other instances of inhomogeneous random graphs are studied in the literature (see [4]). For example we can consider the Chung-Lu random graph [7], where the edge probabilities are given by p(CL) = wiwj ∧ 1 or the Norrow-Reittu model [60], with a Poisson number of ij lN edges in between any two vertices i and j, with parameter equal to wiwj /lN . In [64] it is proved that, under suitable hypotheses, these models are asymptotically equivalent to a generalized random graph, therefore the analysis performed in this thesis is still true in these other cases.
2.1.4 Local tree-likeness of CMN (d) and GRGN (w)
Referring to [64] we remark that CMN (d) and GRGN (w) are locally tree-like graphs uniformly sparse when properties (a) and (b) of the Conditions 2.1.1 and 2.1.2 hold. When also property (c) holds, the graphs have strongly finite mean. It is straightforward to prove that the degree sequences of CMN (2) and CMN (1, 2) satisfy Condition 2.1.1. For example, for CMN (1, 2), we have
P(DN = 2) = 1 − P(DN = 1) = bpNc/N → p, so that the limiting degree D has distribution P(D = 2) = 1 − P(D = 1) = p.
In the case of the GRGN (w) with asymptotic weight W the asymptotic degree D is a mixed Poisson P oi(W ), i.e. W k P(D = k) = E e−W , (2.1.9) k! as shown in [64].
2.1.5 Phase transition in random graphs
Random graphs exhibit a phase transition in the size of the connected components. The existence of a phase transition in the graphs depends on the asymptotic degree D: it is known [4, 42, 51, 52] that when ν > 1 a giant component exists. As we will observe in the following, the phase transition for the Ising model on random graphs is strictly related to the graphs phase transition (see Theorems 2.5.3 (iii) and 5.1.1 (iii) below).
For GRGN (w) the degree distribution is D = P oi (W ) and we have that ν = E[W 2]/E[W ], because E[D] = E[W ] and E [D (D − 1)] = E[W 2]. Thus, depending on W , a giant component for GRGN (w) may exist.
On the other hand, the graph phase transition does not exist for CMN (2) and CMN (1, 2), since ν ≤ 1 in these cases. In particular, ν = 1 for CMN (2) and ν = 2p/(1 + p) < 1 for CMN (1, 2). Note that CMN (2) is critical for the existence of the giant component.
9 Ising model on random graphs
2.1.6 Why these graphs?
The need to work on these particular examples of graphs, i. e., the configuration model and the generalized random graph, arises from the fact that general tools, that we use in the proofs of random quenched theorems, are not available in the averaged quenched and annealed setting. Under these two measures, as we will see in the following, the fluctuations due to the spatial structure become relevant in the study of the central limit theorem. Therefore, due to the lack of general formalism to treat the generic class of locally tree-like graphs, we analyze these specific random graphs as benchmarks, since they allow for explicit computations. We choose the configuration model with vertex degree 2 and the configuration model with vertex degrees 1 and 2 because in these cases only one-dimensional structures, like cycles and lines, appear. Then the partition function of the Ising model on these graphs can be expressed in term of the well-known partition function of the one-dimensional nearest-neighbor Ising model. The choice of the generalized random graph is related to the fact that in this model the edges are independent. This implies that in the annealed setting the Ising model on the generalized random graph is reduced to an inhomogeneous Curie–Weiss model, as we explain in Section 5.4 below.
2.1.7 Other random graph models
We present in this subsection some other examples of random graphs. We are not going to consider them in this thesis, but they should be interesting for future works. One of the simplest random graph model is the Erdős–Rényi random graph, ER [29, 38], denoted by GN (p). This graph has N vertices and each pair of vertices is independently connected with a probability p. If we choose p = c/N the degree of a random vertex in the graph has distribution Bin(N − 1, c/N) that, for N → ∞, converges to a P oiss(c). Despite the simplicity of this model, it is fascinating because of the presence of a phase transition when p varies. The preferential attachment model,[1, 5], is an instance of a graph growing in time. Vertices having a fixed number of edges are sequentially added to the graph. Their edges are attached to a receiving vertex with a probability proportional to the degree of the receiving vertex at that time. Thus, vertices that already have a high degree are more likely to attract edges of new vertices. While the Erdős–Rényi random graph and the preferential attachment model are locally tree-like random graphs, we introduce now a model that does not belong to this class: the random intersection graph,[62, 44]. In this model, we assign to each vertex v a random subset Sv of a given set S. Then two vertices v and w are connected by an edge only if their assigned sets Sv and Sw intersect. This graphs are natural models for social networks. For example, we can consider
10 2.2 Ising model users of a social network as vertices of the graph and social groups as elements of S. Then two users are joined by an edge if they share at least one group.
2.2 Ising model
We start introducing the Ising model on a subset with N sites of the lattice Zd. To each site we assign a spin variable σi which only takes the values +1 (spin-up) or −1 (spin-down). A spin configuration of the system σ = (σ1, σ2, ..., σN ) takes values on the space of spin configurations ΩN , i.e., the set of all spin sequences N N σ. Therefore ΩN = {1, −1} and we have 2 possible configurations. We introduce the Hamiltonian of the system that describes the total energy of a configuration as
N X X HN (σ) = − Ji,jσiσj − h σi. (2.2.1) i,j i=1 ki−jk=1
The first sum is over all pairs of variables that are connected by exactly one bond on the lattice, i.e., we have only nearest-neighbor interactions. If for all (i, j) the coupling constants Ji,j satisfy Ji,j ≥ 0 then the model is ferromagnetic, otherwise if Ji,j < 0 ∀ i, j the model describes an antiferromagnet. For a ferromagnet, if the spins of two nearest neighbors in the configuration have the same orientation (both either up or down), then this pair contributes an energy −Ji,j. If they have opposite orientations, they contribute an energy Ji,j. So, a configuration in which most nearest neighbors have the same orientation has a lower energy level. In this thesis we assume that Ji,j = 1 for each couple (i, j) of nearest neighbors. The parameter h correspond to the presence of an external magnetic field that tends to align the spin variables in its direction. The probability of a particular configuration σ is given by the Boltzmann–Gibbs measure e−βHN (σ) µN,β,h(σ) = , (2.2.2) ZN (β, h) where β = 1/T is the inverse temperature and ZN (β, h) is the partition function given by X −βHN (σ) ZN (β, h) = e . (2.2.3)
σ∈ΩN From the partition function we can derive most of the information about the system. The negative sign in the exponential of the Boltzmann–Gibbs measure gives a higher probability to states with lower energy. A small value of β (which corresponds to a high temperature) tends to make all configurations more or less equally likely, while a large value of β (corresponding to a low temperature) tends to accentuate the probabilities of the lowest-energy states.
11 Ising model on random graphs
We can also define the Ising model on a graph G with N vertices. We denote by V (G) the vertex set of G and by E(G) its edge set. Assigning a spin variable σi to each vertex and considering interactions only between vertices that are connected by one edge, we have that the Hamiltonian is given by X X HN (σ) = − Ji,jσiσj − h σi. (2.2.4) (i,j)∈E(G) i ∈V (G)
Then we can proceed as above defining the partition function and the Boltzmann– Gibbs measure.
2.3 Measures and thermodynamic quantities for the Ising model on random graphs
Now we define the Ising model on finite random graphs. In this setting two sources of randomness, i.e., the Ising random variables and the spatial disorder, appear. Then, to describe the system, we have to take into account this double randomness in different ways. We can first consider a fixed realization of the graph. In this case the probability of a spin configuration is given again by the Boltzmann–Gibbs distribution with the sum of the nearest neighbors taken on the edge set. Next we can evaluate a dynamical evolution of the graphs. So the probability of a spin configuration is now obtained by averaging over all possible random graphs with N vertices and different edge sets. As we will see, we can take this average in two different ways. In order to give the definition of the three measures that arise, we denote by
QN the law of the graphs with N vertices belonging to GN , for any N ∈ N. Definition 2.3.1 (Measures and expectations). For the spin variables σ = N (σ1, ..., σN ) taking values on the space of spin configurations ΩN = {−1, 1} we consider the following measures:
(i) Random quenched measure. For a given realization GN ∈ GN the ran- dom quenched measure coincides with the random Boltzmann–Gibbs distri- bution h i exp β P σ σ + B P σ (i,j)∈EN i j i∈[N] i µGN (σ) = , (2.3.1) ZGN (β, B) where " # X X X ZGN (β, B) = exp β σiσj + B σi (2.3.2) σ∈ΩN (i,j)∈EN i∈[N]
is the partition function. Here β ≥ 0 is the inverse temperature and B ∈ R is the uniform external magnetic field.
12 2.3 Measures and thermodynamic quantities for the Ising model on random graphs
(ii) Averaged quenched measure. This law is obtained by averaging the random Boltzmann–Gibbs distribution over all possible random graphs h i exp β P σ σ + B P σ (i,j)∈EN i j i∈[N] i PN (σ) = QN (µGN (σ)) = QN . ZGN (β, B) (2.3.3)
(iii) Annealed measure. This law is obtained by averaging separately the partition function in the Boltzmann–Gibbs distribution h i Q exp β P σ σ + B P σ N (i,j)∈EN i j i∈[N] i PeN (σ) = . (2.3.4) QN (ZGN (β, B))
Note that, in the previous definitions, the inverse temperature β does not multiply the external field. This turns out to be technically convenient and does not change the results, because we are only looking at systems at equilibrium, and hence this would just be a reparametrization.
In the following, with a slight abuse of notation we use the same symbol to denote both a measure and the corresponding expectation. Namely, for any function k : ΩN −→ R, we let X µGN (k) = k(σ)µGN (σ) , σ∈ΩN denote the expectation of k with respect to µGN , and similarly for the measures QN , PN and PeN . Moreover we remark that all the measures defined above depend sensitively on the two parameters (β, B). However, for the sake of notation, we drop the dependence of the measures on these parameters. We denote the partition function by ZN instead of ZGN and sometimes we use Varµ(X) to denote the variance of a random variable X with law µ.
Considering these three measures we can compute the expectation of any function that depends on the spins and on the graphs. In particular, we next present some expectations, called thermodynamic quantities, that are particularly useful to describe the behavior of system. These quantities are functions of the inverse temperature and of the external field.
Definition 2.3.2 (Thermodynamic quantities). For a given N ∈ N we introduce the following thermodynamic quantities at finite volume:
(i) The random quenched pressure: 1 ψN (β, B) := log ZN (β, B) . (2.3.5) N
13 Ising model on random graphs
(ii) The averaged quenched pressure: 1 ψ (β, B) := QN (log ZN (β, B)) . (2.3.6) N N
(iii) The annealed pressure: 1 ψeN (β, B) := log (QN (ZN (β, B))) . (2.3.7) N
(iv) The random quenched, averaged quenched and annealed magnetization, respectively: SN MN (β, B) := µ , (2.3.8) GN N SN M N (β, B) := PN , (2.3.9) N SN MfN (β, B) := PeN , (2.3.10) N where the total spin is defined as X SN := σi . (2.3.11) i∈[N]
(v) The random quenched susceptibility:
∂ χN (β, B) := MN (β, B). (2.3.12) ∂B
The averaged quenched and annealed susceptibility, respectively χN (β, B) and χeN (β, B), are defined in the same way as the random quenched sus- ceptibility with MN replaced by M N and MfN .
In the following we focus our attention on the thermodynamic limit of these quantities, i.e. their limits as N → ∞. This limit is particularly interesting because critical phenomena, as well as phase transition, may appear.
Remark 2.3.1 (Susceptibility and total spin variance). We remark that in the random quenched setting, the susceptibility coincides with the variance of the sum of the spins rescaled by the square root of N. Indeed, a simple calculation gives
14 2.4 Phase transition and critical behavior
∂ χ (β, B) = M (β, B) N ∂B N P P 2 β σiσj +B σi 1 P P (i,j)∈EN i∈[N] N σ∈Ω i∈[N] σi e = N ZN P P β σiσj +B σi 2 √1 P P (i,j)∈EN i∈[N] ! σ∈Ω i∈[N] σie − N N ZN 2 2 SN SN = µG √ − µG √ N N N N SN = Varµ √ . (2.3.13) GN N SN A similar results is true in the annealed case: χN (β, B) = Var √ . e PeN N In the averaged quenched setting the situation is different. Indeed, accordingly to Definition 2.3.2, the averaged quenched susceptibility is ∂ SN χ (β, B) = M N (β, B) = QN Varµ √ = QN (χN (β, B)) . N ∂B GN N (2.3.14) On the other hand, using the law of total variance, the variance with respect the averaged quenched measure is SN SN SN VarP √ = QN Varµ √ + VarQ µG √ , (2.3.15) N N GN N N N N i.e., we have the sum of two contributions:√ the first is given by the average over graphs of the conditional variance of SN / N (the conditioning is given by the graph realization);√ the second contribution is the variance of conditional mean of SN / N. We can also rewrite the expression in (2.3.15) using the thermodynamic quantities in Definition 2.3.2: SN √ VarP √ = χ (β, B) + VarQ NMN (β, B) . (2.3.16) N N N N
2.4 Phase transition and critical behavior
A phase transition occurs when a small change in a parameter, such as the temperature, causes a qualitative change in a quantity that describes the state of the system. In our case this quantity is called spontaneous magnetization. Suppose we put a piece of magnetic material, like a piece of iron, in a magnetic field at a constant temperature. This field forces the spins of the material to align with its direction, inducing an amount of magnetization into the material. When
15 Ising model on random graphs we turn off the external field we see that for high temperature the material returns to be unmagnetized. Instead, for low temperature, an amount of magnetization, called spontaneous magnetization, is kept. The temperature that separates these two states is called critical temperature.
From a formal point of view, when M(β, B) := limN→∞ MN (β, B) exists, where
MN (β, B) is any of the magnetizations defined in (2.3.8), (2.3.9) or (2.3.10), criticality manifests itself in the behavior of the spontaneous magnetization + defined as M(β, 0 ) = limB↓0 M(β, B). In fact, the critical inverse temperature is defined as + βc := inf{β > 0 : M(β, 0 ) > 0}, (2.4.1) and thus, depending on the setting, we can obtain the random quenched, averaged rq aq an quenched and annealed critical point denoted by βc , βc and βc respectively. When, for one of the three, 0 < βc < ∞, we say that the system undergoes a phase transition at β = βc. Such transition in mathematical terms is a singularity that occurs when a discontinuity in one of the derivatives of the pressure appears. Note that βc can only exist in the thermodynamic limit, because the finite volume pressure is an analytic function on β and B. Moreover, due to the Z2 symmetry of the Ising model, the magnetization of a finite graph always satisfies + MN (β, 0 ) = 0. The behavior around the critical temperature can be quantified in terms of critical exponents. In particular, we introduce the critical exponents that describe the behavior of the magnetization and susceptibility at the critical inverse temperature βc and at zero external magnetic field. Definition 2.4.1 (Critical exponents). The critical exponents β, δ, γ, γ0 are defined by
+ β M(β, 0 ) (β − βc) , for β & βc; (2.4.2) 1/δ M(βc, B) B , for B & 0; (2.4.3) + −γ χ(β, 0 ) (βc − β) , for β % βc; (2.4.4) + −γ0 χ(β, 0 ) (β − βc) , for β & βc, (2.4.5) where we write f(x) g(x) if the ratio f(x)/g(x) is bounded away from 0 and infinity for the specified limit.
Remark 2.4.1. We emphasize that there is a difference between the symbol β for the inverse temperature and the bold symbol β for the critical exponent in (2.4.2). Both notations are standard in the literature, so we decided to follow both of them and distinguish them by the font style.
2.5 Preliminaries
In this section we provide some theorems that are essential to prove our results. In particular, we present two correlation inequalities and a theorem that states
16 2.5 Preliminaries the existence of the limit of the thermodynamical quantities, under the random quenched measure, achieved by Dembo and Montanari first [16] and further by Dommers, Giardinà and van der Hofstad [20]. Moreover, we recall a theorem by Dommers, Giardinà and van der Hofstad [21] that specifies the value of critical exponents for locally tree-like random graph in the random quenched setting. We will compare this result with the annealed critical exponents studied below in Chapter 6. We begin this section by giving first a quick overview on the state of the art for the Ising model on random graphs.
2.5.1 State of the art
The literature on the Ising model is enormous. Since the original study of Ising [41] a massive number of research papers have been published on the subject. We focus our overview on the class of Ising models on random graphs. Ising models on random graphs and complex networks have been first studied with the tools of theoretical physics [22, 47, 23] and more recently they have been the subject of rigorous mathematical studies. A first mathematical approach to the study of the Ising model on random regular graphs was provided by Gerschcnfeld and Montanari in [35]. Subsequently, De Sanctis and Guerra [15] investigated the Ising model on the Erdős–Rényi random graph. They studied the limit of the free energy density at high temperature and at zero temperature. In [16], Dembo and Montanari analyzed the ferromagnetic Ising model for any temperature on random graphs with a finite-variance degree distribution. Later, Dommers, Giardinà and van der Hofstad [20], extended these results to the case where the degree distribution has infinite variance, but strongly finite mean. They computed the thermodynamic limit of the pressure and of various physical quantities, such as the magnetization, the susceptibility and the internal energy. In [18] Dembo, Montanari and Sun generalize this further to more general locally tree-like graphs. An analysis of the critical behavior is reported in [21] by Dommers, Giardinà and van der Hofstad. They studied the critical behavior of the Ising model on power-law random graphs by computing critical exponents for the magnetization and the susceptibility. Regarding the analysis of asymptotic theorems for the sum of the spin, results on lattices or on the complete graph can be found in [26, 28, 27].
2.5.2 Correlation inequalities
We introduce two classical correlation inequalities: the Griffiths, Kelly, Sherman inequality (GKS) [45] and the Griffiths, Hurst, Sherman inequality (GHS) [39]. We underline that these inequalities hold in the random quenched setting.
Theorem 2.5.1 (GKS inequality). Consider two Ising model with measures µ 0 0 0 and µ on graphs GN = (VN , EN ) and GN = (VN , EN ), with inverse temperature
17 Ising model on random graphs
0 0 0 0 β and β and external fields B and B respectively. If EN ⊆ EN , β ≤ β and 0 0 ≤ B ≤ B , then for any U ⊆ V N, ! ! Y 0 Y 0 ≤ µ σi ≤ µ σi . (2.5.1) i∈U i∈U Theorem 2.5.2 (GHS inequality). Let β ≥ 0, then for all B ≥ 0,
∂2 M (β, B) ≤ 0. (2.5.2) ∂B2 N i.e. the magnetization is a concave function of the positive external field. Remark 2.5.1. The GHS inequality also holds in the averaged quenched setting. Indeed, using Definition 2.3.2 and Theorem 2.5.2, it results ∂2 ∂2 S ∂2 M (β, B) = Q µ N = Q (M (β, B)) ∂B2 N ∂B2 N GN N ∂B2 N N ∂2 = Q M (β, B) ≤ 0. (2.5.3) N ∂B2 N These results are particularly useful to study properties of the magnetization and, as we will see, the GHS inequality is crucial to prove the central limit theorem in the random quenched case. Due to the lack of this instrument in the annealed setting, we can not follow the Ellis’ strategy to prove the central limit theorem for the general class of locally tree-like random graphs.
2.5.3 Thermodynamic limits for the random quenched law
In the next theorem we collect some results, taken from [16, 20], that guarantee the existence of the thermodynamic limit of the quantities previously defined in the random quenched setting. In order to state the existence of the limit magnetization, we define the set
rq rq U := {(β, B) : β ≥ 0, B 6= 0 or 0 < β < βc , B = 0} (2.5.4)
rq where βc is the quenched critical value which is given in the next theorem. Theorem 2.5.3 (Thermodynamic limits for the random quenched law [16, 20]).
Assume that the random graph sequence (GN )N≥1 is locally tree-like, uniformly sparse, and with asymptotic degree distribution D with strongly finite mean, then the following conclusions hold: (i) For all 0 ≤ β < ∞ and B ∈ R, the quenched pressure exists almost surely in the thermodynamic limit N → ∞ and is given by
ψ(β, B) := lim ψN (β, B). (2.5.5) N→∞ Moreover, ψ(β, B) is a non-random quantity.
18 2.5 Preliminaries
(ii) For all (β, B) ∈ U rq, the random quenched magnetization per vertex exists almost surely in the limit N → ∞ and is given by
M(β, B) := lim MN (β, B). (2.5.6) N→∞ ∂ The limit value M(β, B) equals M(β, B) = ∂B ψ(β, B) for B 6= 0, rq whereas M(β, B) = 0 in the region 0 < β < βc , B = 0. (iii) The critical inverse temperature is given by
rq βc = atanh (1/ν) , (2.5.7) where ν is defined in (2.1.2). (iv) For all (β, B) ∈ U rq, the thermodynamic limit of susceptibility exists almost surely and is given by ∂2 χ(β, B) := lim χN (β, B) = ψ(β, B). (2.5.8) N→∞ ∂B2
Let us remark that, since ν ≤ 1 for both CMN (2) and CMN (1, 2), from (2.5.7) rq it follows that βc = ∞, which means that there is no quenched phase transition in these models.
In the next proposition we also provide an explicit expression for the random quenched pressure and magnetization in the thermodynamic limit. Proposition 2.5.4 (Explicit expression for the thermodynamic quantities [16,
20]). Assume that the random graph sequence (GN )N≥1 is locally tree-like, uniformly sparse, and with asymptotic degree distribution D with strongly finite mean, then for all 0 ≤ β < ∞ and B > 0 the thermodynamic limit of the pressure is given by E[D] E[D] ψ(β, B) = log cosh(β) − E [log (1 + tanh(β) tanh(h ) tanh(h ))] 2 2 1 2 L h B Y + E log e (1 + tanh(β) tanh(hi)) i=1 L −B Y i + e (1 + tanh(β) tanh(hi)) , (2.5.9) i=1 where (ii) L has distribution D;
∗ ∗ (ii) (hi)i≥1 are i.i.d. copies of the fixed point h = h (β, B) of the distributional recursion
Kt (t+1) d X (t) h = B + atanh(tanh(β) tanh(hi )), (2.5.10) i=1
19 Ising model on random graphs
(0) where h ≡ B, (Kt)t≥0 are i.i.d. random variables with distribution ρ and (t) (t) (hi )i≥1 are i.i.d. copies of h independent of Kt;
(iii) L and (hi)i≥1 are independent. Moreover, taking the derivatives of (2.5.9) with to respect the external field B, we obtain the explicit expression for the thermodynamic limit of the magnetization, i.e.
L h X i M(β, B) = E tanh B + atanh(tanh(β) tanh(hi)) . (2.5.11) i=1
2.5.4 Critical exponents for the random quenched law
We next present some critical exponents in the random quenched setting for the Ising model on random graphs with degree distribution satisfying a power law with exponent τ . The values of the critical exponents for magnetization and susceptibility, for different values of τ, are stated in the following theorem:
Theorem 2.5.5 (Critical exponents for the random quenched law [21]). Assume that the random graph sequence (GN )N≥1 is locally tree-like, uniformly sparse, and with asymptotic degree distribution D that obeys E[K3] < ∞ or a power law with exponent τ ∈ (3, 5]. Then the critical exponents exist and satisfy
τ ∈ (3, 5) E[K3] < ∞ β 1/(τ − 3) 1/2 δ τ − 2 3 γ 1 1 γ0 ≥ 1 ≥ 1
For the boundary case τ = 5 there are the following logarithmic corrections for β = 1/2 and δ = 3:
1/2 + β − βc rq Mf(β, 0 ) for β & βc , (2.5.12) log 1/(β − βc) B 1/3 Mf(βrq, B) for B & 0. (2.5.13) c log(1/B)
Note that the critical exponents β, δ and γ take the classical mean-field values for E[K3] < ∞, but they are different for τ ∈ (3, 5).
20 2.6 Other spin models
2.6 Other spin models
Beyond the Ising model, other spin models are interesting to model processes on networks, see [23] for an overview. We now present three such models: the Curie–Weiss model, the Potts model and the XY model.
2.6.1 The inhomogeneous Curie–Weiss model
The Curie–Weiss model is an Ising model that allows interaction between each couple of spins. That means that the spatial distance does not play a role anymore. We next define the inhomogeneous Curie–Weiss model. This is a generalization of the classical Curie–Weiss model in which the strength of the ferromagnetic interaction between spins is not spatially uniform. As the standard Curie–Weiss model, it is defined on the complete graph with vertex set [N] := {1, ... , N}.
Definition 2.6.1 (Inhomogeneous Curie–Weiss model). Let σ = {σi}i∈[N] ∈ {−1, 1}N be spin variables. The inhomogeneous Curie–Weiss model, denoted by
CWN (J), is defined by the Boltzmann–Gibbs measure
eHN (σ) µN (σ) = (2.6.1) ZN where the Hamiltonian is 1 X X H (σ) = J (β)σ σ + B σ (2.6.2) N 2 ij i j i i,j∈[N] i∈[N] and ZN is the normalizing partition function. Here β is the inverse temperature, B is the external magnetic field and J = {Jij (β)}i,j∈[N] are the spin couplings.
In the above, the interactions Ji,j (β) might be arbitrary functions of the inverse temperature (in particular no translation invariance is required), provided that the thermodynamic limit is well-defined, i.e., the following limit defining the pressure exists and is finite, 1 ψ(β, B) := lim log ZN (β, B) . (2.6.3) N→∞ N
Since the coupling constants J = {Jij (β)}i,j∈[N] are different for different edges, we speak of an inhomogeneous Curie–Weiss model. In Chapter 5 it is shown that the inhomogeneous Curie–Weiss model arises in the study of the annealed Ising model with network of interactions given by the generalized random graph, also called the rank-1 inhomogeneous random graph. In that case, each vertex i ∈ [N] receives a weight wi and we will take J = {Jij (β)}i,j∈[N] such that wiwj X Jij (β) = β, where `N = wk. (2.6.4) `N k∈[n]
21 Ising model on random graphs
In the case where wi ≡ 1, our model reduces to the homogeneous Curie–Weiss model. We call the coupling constants in (2.6.4) the rank-1 inhomogeneous Curie–Weiss model.
2.6.2 The Potts model
The Potts model is a generalization of the Ising model. It is introduced on the random graph GN as follows: Definition 2.6.2 (Potts model). Fix an integer q ≥ 2 and assign to each vertex i ∈ [N] a Potts spin variable σi ∈ [q]. Then the Potts model on GN is defined by the Boltzmann–Gibbs measure h i exp β P δ + B P δ i,j∈EN σi,σj i∈[N] σi,1 µN (σ) = , (2.6.5) ZN (β, B) where δa,b is the Kronecker delta i.e. δa,b = 0, 1 if a 6= b and a = b, respectively. Note that the Ising model is equivalent to a Potts model with q = 2, (with σi ∈ {−1, 1} replaced with σi ∈ {1, 2}).
The Potts model has been intensively studied in statistical mechanics because of its importance in the theory of phase transitions and critical phenomena. Moreover, this model is closely related to the graph coloring problem, where the aim is to know if the graph can be colored with q colors in such a way that no two neighboring vertices have the same color. We next mention some results for the Potts model on random graphs. In [9], Contucci, Dommers, Giardinà, and Starr analyze the antiferromagnetic Potts model on the Erdős–Rényi random graph. They show the existence of the thermodynamic limit of the pressure and they identify a phase transition. Dembo, Montanari and Sun [18] study the existence of the free energy density for a Potts model on uniformly sparse graph sequences converging locally to a random tree. Subsequently, Dembo, Montanari, Sly and Sun in [17] provide an explicit formula for the limiting free energy density for ferromagnetic Potts models on uniformly sparse graph sequences converging locally to regular trees of even degree, covering all temperature regimes.
2.6.3 The XY model
Unlike the Ising and Potts models with discrete spins, the XY model is described by continuous spin variables. Indeed, in the XY model, the classical spins can rotate in the interval [0, 2π]. The definition for the model on a random graph
GN is the following:
22 2.7 Overview
Definition 2.6.3 (XY model). Assign to each vertex i ∈ [N] a spin variable
σi ∈ [0, 2π]. Then the XY model on GN is defined by the Boltzmann–Gibbs measure h i exp β P cos(σ − σ ) + B P cos(σ ) i,j∈EN i j i∈[N] i µN (σ) = . (2.6.6) ZN (β, B) There are few studies of the XY model on complex networks. We can find an overview of them in [23].
2.7 Overview
In this section we discuss about the main topics of this thesis. We provide a wide analysis on the three measures and their properties. We present our principal results and we give an idea of the strategy of the proofs. Furthermore an examination for the scenario at the critical point and in the low-temperature region is performed.
2.7.1 Discussion on the measures
To understand the difference between the quenched and annealed settings, it is convenient to think of a microscopic dynamics yielding the equilibrium state. For instance, one could imagine that the spins are subject to a Glauber dynamics with a reversible Boltzmann–Gibbs distribution and the graph also has its own dynamical evolution approaching the graph’s stationary distribution. In general, these two dynamics are intertwined and both concur to determine the equilibrium state, i.e., the asymptotic value of an ergodic dynamical time average. The quenched and annealed state arise as follows: (a) In the quenched state, the changes of the graph happen on a time-scale that is infinitely longer than the time-scale over which the changes of the spin variables occur. Thus in the quenched state the graph viewed by the evolving spins is frozen. One distinguishes between the random quenched measure, i.e., the random Boltzmann–Gibbs distribution of a given real- ization of the graph, and the averaged quenched measure, i.e., the average of the Boltzmann–Gibbs distribution over the graph ensemble. Several thermodynamic observables (e.g., the free energy per particle, the internal energy per particle, etc.) are self-averaging, and therefore the random quenched values and their averaged quenched expectations do coincide in the thermodynamic limit. In the study of the fluctuations of the properly rescaled magnetization one finds a Gaussian limiting law. Interestingly, the asymptotic variances of the random quenched and averaged quenched state might be different due to local Gaussian fluctuations of graph properties. (b) In the annealed state, the random environment evolves much faster than the spin variables. As a consequence, the environment seen by the spins
23 Ising model on random graphs
includes all possible arrangements of the random graph. The annealed measure (defined in (2.3.4)) is given by the stationary reversible measure of a Glauber spin dynamics in which the transition from a configuration σ to another configuration σ0 occurs with probability
0 E[e−βH(σ )] ∧ 1, (2.7.1) E[e−βH(σ)]
where H is the Hamiltonian and E[·] represents the average over the graph ensemble. The above dynamics corresponds to an extremely fast random graph dynamics in which we do not even observe the graph at any time, but merely see it averaged over the random graph distribution. This is equivalent to an effective Glauber dynamics with (annealed) Hamiltonian equal to 1 Han(σ) = − log(E[e−βH(σ)]). (2.7.2) β Thus, by construction, the annealed measure is necessarily non-random and its normalization includes the average over the graphs ensemble. In this thesis, we will study annealed central limit theorems for the ferromagnetic Ising model on random graphs, in order to deduce what the effect of annealing on the macroscopic properties of the Ising model is. The definition of the annealed measure in the context of Ising models on random graphs is thus different than in other classes of problems with disorder. Usually, this can be interpreted in terms of the dynamics of the environment and the process on it being equally fast. In the context of random walks in random environment [8], annealing is rather similar to what we here have called the averaged quenched measure. This is due to the lack of a partition function for random walks in random environment. In disordered systems (such as spin glasses [50, 12]), annealed disorder is usually considered to be easier to deal with mathematically, since the average on the disorder and the thermal average are treated on the same footing. This is true whenever the edges of the graph are independent, due to the form of the Hamiltonian that allows a factorization of expectations w.r.t. the bond variables. If instead the edge distribution in the graph does not have a product structure, then the annealed case can actually be more difficult than the quenched case. Indeed, whereas the random-quenched case is dominated by the typical realization of the graph (often having the local structure of a random tree), in the annealed case (as in the averaged quenched case) rare graph samples having a very low energy actually give a contribution that can not be ignored. This is due to the fact that the Ising model gives rise to exponential functionals on the random graph, and expectations of exponential functionals tend to be dominated by rare events in which the exponential functional is larger than it would be under the quenched law. Deriving such statement rigorously requires a deep understanding of the large deviation properties of random graphs, a highly
24 2.7 Overview
interesting but also challenging topic. In this thesis, we consider graph ensembles of both types, i.e., random graphs with independent edges (the generalized random graph) or depen- dent edges (the configuration model).
2.7.2 Overview of results
The primary purpose of this thesis is to study asymptotic results for the total spin SN under the three different measures introduced in Section 2.3. First of all, we prove the existence of the law of large numbers. Namely, for all settings, we prove that the total spin normalized by the volume N converges to the magnetization of the model. Depending on the measure involved, the convergence can be in probability or almost surely. The scenario for the central limit theorem is more complex. For each measure we work in the region of the parameters (β, B) such that there exists a unique Gibbs measure. For the general class of the locally tree-like random graphs, working under√ the random quenched measure, we obtain that the total spin SN rescaled by N converges to a Gaussian random variable with variance equal to the susceptibility of the system. In order to prove this, we mimic the strategy used by Ellis for the Ising model on Zd in [26], getting the same result. On the 2-regular configuration model, for all three measures the central limit theorems show the convergence to the same Gaussian random variable. Indeed, because the degree is not allowed to fluctuate, there is no contribution from the graph and, as a result, the limiting variance coincides with the random quenched susceptibility.
The situation is different for the CMN (1, 2). Here the fluctuations of the degree of the graph become relevant providing a different central limit theorem for each of the three measures. In the random quenched setting the variance will still coincide with the random quenched susceptibility, while it will be larger in the averaged quenched setting and different in the annealed one. The Ising model on the generalized random graph under the annealed measure is reduced to an inhomogeneous Curie–Weiss model. Then we can again follow Ellis’ strategy for the central limit theorem, obtaining as limiting variance of the Gaussian the annealed susceptibility of the model. In this setting we discover that the model presents an annealed critical temperature which is larger than the quenched one. So we are not able to compare the two different variances of the central limit theorems, i.e. the susceptibilities under the random quenched and annealed measures. Thanks to this link between the Ising model on the generalized random graph and the Curie–Weiss model, we can go further and see what happens at the critical point. Adapting the strategy in [21] to our case, we prove that the annealed critical exponents match those of quenched setting and, following [26], we show a non-classical limit theorem: rescaling in an appropriate way the total spin we obtain the convergences to a non-Gaussian random variable.
25 Ising model on random graphs
The role of the inhomogeneity of the graphs, expressed in terms of the power law distribution of the degrees, is crucial for the study at the critical point. Indeed, the analysis of the critical exponents and the non-classical limit theorem depends significantly on the value of the exponent τ of the power law. On the contrary, this inhomogeneity does not have any influence when we investigate the scenario in the uniqueness regime.
2.7.3 CLT proof strategy
The proofs of the central limit theorems for the three different settings are different (in fact they require the control of the fluctuations of SN with respect to different ensembles), but are based on the same main idea [26]. This idea consists in using the moment generating function of the random variables S −E(S ) N √ N , where E is the average in the chosen measure, and in showing that N with respect to the same measure these moment generating functions converge to those of Gaussian random variables. Whenever possible, this step requires the computation of the variances of the limiting Gaussian variables. This could be achieved by considering the scaled cumulant generating functions of SN , given by 1 c (t) = log µ [exp (tS )] (2.7.3) N N GN N in the random quenched setting, 1 c (t) = log P [exp (tS )] (2.7.4) N N N N in the averaged quenched setting and 1 cN (t) = log PeN [exp (tSN )] (2.7.5) e N in the annealed setting. By taking the second derivative of (2.7.3), (2.7.4) SN and (2.7.5) and evaluating it in zero one obtains, respectively, Varµ √ , GN N SN SN VarP √ and Var √ . The crucial argument in the proof of the theo- N N PeN N rems is to show the existence of these variances in the limit N → ∞. This in 00 turn is a consequence of the existence of the limit of the sequences (cN (t))N≥1, 00 00 (cN (t))N≥1 and (ecN (t))N≥1 for t = tN = o(1). While the existence of the limit c(t) := limN→∞ cN (t) can be established for the Ising model on locally tree-like random graphs as a simple consequence of the existence of the random quenched pressure, the existence of the limit c(t) := limN→∞ cN (t) is more challenging. In particular, it requires detailed knowledge not only of the typical local structure of the graph around a random vertex, but also of the fluctuations around that structure. Moreover, the general argument of [26] relies on concavity of the first derivatives of the cumulant generating functions. This can be achieved 0 for cN (t), i.e., in the random quenched setting, thanks to the GHS inequality, which holds for the ferromagnetic Boltzmann–Gibbs measures µGN . On the
26 2.7 Overview other hand, under the averaged quenched measure, the first derivative of the cumulant generating functions can not be expressed in term of the averaged quenched magnetization to exploit the GHS inequality, and we are not able 0 to prove the concavity of cN (t). Thus in the averaged quenched setting we focus on two specific models, i.e. the CMN (2) and CMN (1, 2) random graphs, which allow for explicit computations of the relevant quantities. In fact in these cases, since the typical structure in the graphs are cycles (for CMN (2)) and lines and cycles (for CMN (1, 2)), the averaged quenched pressure of the Ising model on these graphs can be expressed in terms of the Ising model pressure ψd=1(β, B) of the one-dimensional nearest-neighbor Ising model. It turns out however that whereas the averaged quenched pressure of the regular random d=1 graph CMN (2) exactly equals ψ (β, B), in the case of CMN (1, 2) the pressure is more involved. Indeed, besides ψd=1(β, B), a new term appears that depends (N) on a set of random variables (p` )`≥1 whose value depends on the realization (N) of the random graph. More precisely, Np` is the number of lines of length ` in the graph (the cycles give a vanishing contribution in the thermodynamic limit).
Then, in order to prove the CLT for CMN (1, 2), it is of pivotal importance to (N) control the fluctuations of the random variables p` in the thermodynamic limit. This result is obtained in [14], where it is proven that the joint limit law of the number of connected components in a graph with vertices of degrees 1 and 2 is Gaussian. Relying on this result, we can complete our proof of the averaged quenched CLT for CMN (1, 2). In the annealed setting we can again rewrite the cumulant generating function in terms of the pressure, since c˜N (t) = ψeN (β, B + 1) − ψeN (β, B). We will show by an explicit computation that the annealed pressure of GRGN (w) coincides with that of an inhomogeneous Curie–Weiss model. From this fact, the thermo- dynamic limit of the annealed pressure, magnetization and susceptibility can be obtained. This again relies on the GHS inequality that is valid also for this inhomogeneous ferromagnetic system. Thus, for the generalized random graph the annealed CLT can be proven in a similar way as for the random quenched measure. On the other hand, the proofs of the CLT for the configuration models do not follow from the abstract argument based on the GHS inequality, since GHS is not available in the general annealed context. Because of that, we have 00 to explicitly control the limit (c˜N (tN ))N≥1 throughout the computation of the annealed pressure. It is relatively simple to accomplish this task in the case of the regular CMN (2) graph. The fluctuating degree of CMN (1, 2) makes the 00 computation of the pressure and of the limit (c˜N (tN ))N≥1 much more compli- cated. While the cycles give a vanishing contribution to the thermodynamic limit, the distribution of the length of the lines has to be carefully analyzed and its Gaussian fluctuations appear in the CLT for the total spin.
27 Ising model on random graphs
2.7.4 Differences between random quenched and averaged quenched setting
Here we explain the distinction between the random quenched and the averaged quenched CLTs. As we said before, in the central limit theorem for the total spin in the random quenched setting, the variance of the Gaussian random variable SN coincides with the random quenched susceptibility χ. The variance VarP √ , √ N N expressing the fluctuations of SN / N with respect to the averaged quenched measure, has instead two contributions. The first is given√ by the average over random graphs of the conditional variance of SN / N with respect to the Boltzmann–Gibbs measure (the conditioning is given by the graph realization);√ the second contribution is given by the variance of conditional mean of SN / N (see (2.3.15)). If a CLT with respect to the averaged quenched measure holds, then the thermodynamic limit of the first term in the right-hand side of (2.3.15) equals the magnetic susceptibility χ, which is a self-averaging quantity. It is clear from (2.3.15) that one expects a different variance in the CLT in the averaged quenched case whenever the thermodynamic limit of the second term on the right-hand side is different from zero. The analysis that we perform on the configuration models under consideration leads us to conjecture that when the vertex degrees are not all equal, one has that lim Var µ √SN is a strictly positive number. This means N→∞ QN GN N that when there are non-vanishing fluctuations of the rescaled total spin with respect to the graph measure, those will be determined by the fluctuations of the degrees distribution of the graph. On the contrary, we predict that when the degrees are fixed, such as in the r-regular random graph, the two limiting variances are equals, thus yielding no distinction between random and averaged quenched central limit theorems. We prove this in the case where r = 2.
2.7.5 Properties of annealing
Working with the annealed measure some interesting phenomena emerge. First of all, in the presence of a ferromagnetic phase transition, annealing can an rq change the critical temperature, meaning that βc < βc . We proved this for the rank-1 inhomogeneous graph. For the configuration models with vertex degrees an rq at most two that we have analyzed, it holds βc = βc = ∞. We conjecture that in the general case when there is a positive proportion of vertices of degree at least 3 and ν > 1 (so that there exists a giant component), an annealed positive critical temperature exists. We believe that this annealed critical temperature is strictly larger than the quenched critical temperature whenever the vertex degrees fluctuate and a positive proportion of the vertices have at least degree three. Furthermore, the annealed state satisfies a central limit theorem for the rescaled magnetization, as the quenched state does. Unfortunately, we can only
28 2.7 Overview prove this for certain random graph sequences, but we believe this to be true in general. The variance of the annealed CLT and the variance of the quenched CLT are different whenever the degrees are allowed to fluctuate. We showed this in the case of the generalized random graph, where they can not be ordered because the quenched and annealed critical temperatures are different and the quenched and annealed susceptibilities diverge at the critical point. For CMN (1, 2) having zero critical temperature and fluctuating degrees the variances are also different, however we have not been able to compare them.
From the analysis of the CMN (2), we see that both the annealed critical tem- perature and the annealed variance are the same of their quenched counterparts. We conjecture this behavior to occur for all random regular graphs.
2.7.6 Universality classes for power-law random graphs
Universality is a key concept in the theory of phase transitions, with application to a large variety of physical systems. Informally, universality means that in the thermodynamic limit different systems show common properties close to criticality. The theory based on the renormalization group suggests that systems fall into universality classes, defined by the values of their critical exponents describing the nature of the singularities of measurable thermodynamic quantities at the critical point. In our study we investigate universality for spin system on power-law random graphs, i.e., graph sequences where the fraction of nodes that have k neighbors is proportional to k−τ for some τ > 1, displaying phase transitions. In [21] the quenched critical exponents have been rigorously analyzed for a large class of random graph models. More precisely in [21] it is proved that the critical exponent, describing the behavior of magnetization and susceptibility near the critical point, take the same values as the mean-field Curie–Weiss model whenever the degree distribution has a finite fourth moment. For power law random graphs it is proved that for τ > 5 the model is in the mean-field universality class, whereas the critical exponent are different from the mean-field values for τ ∈ (3, 5). In Chapter 6 we provide the analysis of the critical behavior but in the annealed setting. Our results are fully compatible with the universality conjecture. The annealed critical temperature is different (actually higher) than the quenched critical temperature, but the set of annealed critical exponents are the same as the quenched critical exponents.
2.7.7 Violation of CLT at βc
In the case of the random quenched CLT, the variance is given by the sus- rq ceptibility of the model. If B = 0 and β = βc the variance diverges because the susceptibility becomes infinite at the critical temperature. So, for these parameters, the CLT breaks down and a different scaling of the total spin SN is
29 Ising model on random graphs needed to obtain a non-trivial limiting distribution. For example, in [26], it is 3/4 shown that for the Curie–Weiss model, at B = 0, β = βc, the quantity SN /N converges in distribution to the random variable X with density proportional to exp −x4/12. We conjecture a similar behavior for the Ising model on locally tree-like random graphs with finite fourth moment of the degree distribution. Indeed, as said before, those graphs have been shown to be in the same univer- sality class as the Curie–Weiss model, i.e., their critical exponents agree with the mean-field critical exponents of the Curie–Weiss model [21]. We investigate this problem in detail for the generalized random graphs under the annealed measure in Chapter 6. We prove that the scaling with N 3/4 is also correct when E[W 4] < ∞, but different for τ ∈ (3, 5). Furthermore we show that when E[W 4] = ∞, different asymptotic distributions arise in the scaling limit. We characterize them for a weight deterministic sequence in which the weights follows a precise power-law. Such a sequence is rather generic in the sense that it produces an asymptotic weight that is also power-law distributed. The analysis shows that the fluctuations of the total spin decrease as the exponent τ becomes smaller and the distribution seen in the scaling limit has tails proportional to τ−1 e−Cx .
2.7.8 Low-temperature region
When the graph sequence (GN )N≥1 is such that there exists a finite critical inverse temperature βc, then the law of large numbers and the central limit theorem do not apply to the low temperature region corresponding to B = 0, d β > βc. For the Ising model on Z [54, 48] or on the complete graph [26], in the low temperature region the law of large numbers for the empirical sum of the spin breaks down because, in the thermodynamic limit, the Boltzmann–Gibbs measure 1 + − becomes a mixture of two pure states µ = 2 (µ + µ ). As a consequence, the empirical magnetization is distributed like the sum of two Dirac deltas at the ∗ ∗ symmetric values ±M (β), with M (β) = limB&0 M(β, B), whereas the CLT with respect to the Boltzmann–Gibbs measure breaks down. However, by using general properties of ferromagnetic systems (e.g. GKS inequalities) it is possible to prove a CLT with respect to the measure µ+, respectively µ− [54, 48]. For the Ising model on random graphs it is believed that a similar picture applies. For instance, for the Ising model on regular random graphs it has been proved in [53] that the low-temperature measure is a convex combination of + and − states and therefore, by appealing to the general results of [54] we conclude that a CLT holds in the pure phase.
2.8 Organization of the thesis
This thesis is organized as follows. In each chapter we present the theorems and the proofs concerning a particular measure.
30 2.8 Organization of the thesis
In Chapter 3, in the general context of locally tree-like random graphs, we present our results with respect to the random quenched measure. We establish the rate at which the law of large numbers for the total spin is reached, and we prove a central limit theorem for the (centered) total spin rescaled by the square root of the volume. As in the case of the Ising model on the lattice or on the complete graph [26], the variance appearing in the CLT result is given by the spin susceptibility. In Chapter 4 we obtain asymptotic results with respect to the averaged quenched measure. While the law of large numbers can be easily formulated for the entire class of locally tree-like random graphs, the scaling to a normal random variable turns out to be much more challenging. Thus, we restrict ourselves to the configuration models with degrees at most 2 and consider the two simplest cases. In particular, we will consider the configuration model with all vertices having degree two, i.e. the 2-regular random graph, and the case where a fraction of the vertices has degrees one and the remaining fraction has degree two. In both cases we obtain the central limit theorem by an explicit computation. In the first case, we find that the CLT with respect to the averaged quenched measure is the same as the CLT with respect to the random quenched measure. In particular the variance of the Gaussian law is given by the susceptibility of the Ising model in one dimension. In the second case, we prove that the asymptotic variance of the rescaled total spin is larger than the susceptibility. This difference originates from the fluctuations of the connected component sizes of the configuration model with degrees 1 and 2, that also follows a Gaussian law in the limit of very large graphs. In Chapter 5 we deal with the annealed measure in the uniqueness regime outside the critical region. First we analyze the generalized random graph model for which we compute the pressure and magnetization in the thermodynamic limit, identify the critical temperature and then prove the law of large number and the central limit theorem. All of these results rely on the fact that the Ising model on the generalized random graph in the annealed setting turns into an inhomogeneous Curie–Weiss model. Then we continue by studying the thermodynamic quantities and the central limit theorems for the 2-regular configuration model and for the model with vertices of degrees 1 and 2. In the former case we show that the variance of the limiting normal variable is the susceptibility of the one-dimensional Ising model. In the latter case, which is much more difficult, the varying degrees of the vertices affect the pressure and the limiting distribution. Finally in Chapter 6 we analyze what happens at the critical point. As mentioned before, at the critical point the CLT breaks down and a different scaling of the total spin is needed to obtain a non-trivial limiting distribution. We investigate this problem for the Ising model on generalized random graph in the annealed setting. We show that the critical exponents for this model match those of an Ising model on locally tree-like random graphs in the random quenched setting. The second aim is to prove a non-classical limit theorem at
31 Ising model on random graphs the critical point. It results that we need different scalings for the total spin according if the degree distribution satisfies a prescribed power law or it has a finite fourth moment In Chapter 7 we introduce a speculative analysis on the distribution of the fixed point of the stochastic recursion equation for the cavity field of the Ising model on a locally tree-like random graph. We provide some analytical results in the zero temperature limit equipped with numerical simulation.
32 RANDOMQUENCHEDMEASURE 3
In this chapter we present our results in the random quenched setting on the general class of the locally tree-like random graphs. We study the asymptotic behavior of the total spin SN under different scalings, i.e. we prove the law of large numbers and the central limit theorem. A full control for the quenched pressure and the main thermodynamical ob- servables has been achieved in [16, 20], see Theorem 2.5.3. We leverage on the knowledge about the quenched pressure in the proofs of both random quenched asymptotic theorems for the total spin. Furthermore, a key role is played by the GHS inequality in the proof of the central limit theorem. Thanks to these ingredients, it turns out that our proofs can be performed along the lines of Ellis’ proofs in the Zd case [26, Section V.7].
3.1 LLN and CLT for locally tree-like random graphs
Our first result in the random quenched setting is the strong law of large numbers for SN /N. In [20] it is proved that this sum normalized by N converges almost surely to a number that is the magnetization of the model as given in (2.5.6). As a preliminary step we will prove here a similar result with a different approach based on large deviation theory, which leads to exponentially fast convergence in probability. Exponential convergence is defined as follows:
Definition 3.1.1 (Exponential convergence). We say that a sequence of random variables XN with laws µN converges in probability exponentially fast to a exp constant x0 w.r.t. µN , and write XN −→ x0, if for any ε > 0 there exists a number L = L(ε) > 0 such that
−NL µN (|XN − x0| ≥ ε) ≤ e for all sufficiently large N.
Then we can present the statement for the strong law of large numbers:
Theorem 3.1.1 (Random quenched SLLN). Let (GN )N≥1 be a sequence of random graphs that are locally tree-like, uniformly sparse, and with asymptotic degree distribution with strongly finite mean. Then, for all (β, B) ∈ U qu,
S exp N −→ M w.r.t. µ , as N → ∞, N GN where M = M(β, B) is defined in (2.5.6). √ To see fluctuations of the total spin, one needs to rescale SN − NMN by N. To prove the CLT, we restrict to the uniqueness regime U qu of parameters (β, B)
33 Random quenched measure such that there exists a unique Gibbs measure [53]. As in the case of the Ising model on the lattice or on the complete graph [26], the variance appearing in the CLT result is given by the spin susceptibility:
Theorem 3.1.2 (Random quenched CLT). Let (GN )N≥1 be a sequence of random graphs that are locally tree-like, uniformly sparse, and with asymptotic degree distribution with strongly finite mean. Then, for all (β, B) ∈ U qu,
SN − NMN D √ −→ N (0, χ) , w.r.t. µG , as N → ∞, N N where χ = χ(β, B) is defined in (2.5.8) and N (0, χ) denotes a centered Gaussian random variable with variance χ.
The remainder of the chapter is devoted to the proofs of these two theorems.
3.2 Proofs
In order to prove the strong law of large lumbers in the random quenched setting, we present a preliminary theorem that guarantees exponential convergence under general hypotheses.
3.2.1 Exponential convergence
Theorem 3.2.1 (Exponential convergence and cumulant generating functions).
Let W = (Wn)n≥1 be a sequence of random vectors which are defined on prob- D ability spaces {(Ωn, Fn, Pn)}n≥1 and which take values in R . We define the cumulant generating functions as
1 D cn(t) = log En [exp(ht, Wni)] , n = 1, 2, ... , t ∈ R , (3.2.1) an where (an)n≥1 is a sequence of positive real numbers tending to infinity, En denotes expectation with respect to Pn, and h−, −i is the Euclidean inner product on RD. We assume that the following hypotheses hold:
D (a) Each function cn(t) is finite for all t ∈ R ;
D (b) c(t) = limn→∞ cn(t) exists for all t ∈ R and is finite. Then the following statements are equivalent: exp (1) Wn/an −→ z0;
(2) c(t) is differentiable at t = 0 and ∇c(0) = z0.
34 3.2 Proofs
See [26, Theorem II.6.3] for a proof based on a large deviation argument.
Now we are ready to prove the strong law of large numbers using the cumulant generating function. As we will see, the existence of the limiting cumulant generating function of the total spin with respect to the random quenched measure will be a direct consequence of the existence of the thermodynamic limit of the quenched pressure.
3.2.2 Random Quenched SLLN: Proof of Theorem 3.1.1
According to Theorem 3.2.1, exponential convergence can be obtained by proving the existence of the limit of the random quenched cumulant generating function cN (t) defined in (2.7.3), and proving the differentiability of the limiting function in t = 0. We have 1 c (t) = log µ [exp (tS )] N N GN N P P exp β σiσj + (B + t) σi 1 X (i,j)∈eN i∈[N] = log N ZN (β, B) σ∈ΩN 1 Z (β, B + t) = log N , (3.2.2) N ZN (β, B) and recalling the definition (2.3.5), the function cN (t) can be rewritten as a difference of random quenched pressures as
cN (t) = ψN (β, B + t) − ψN (β, B). (3.2.3)
The existence of the limit
c(t) := lim cN (t) = ψ(β, B + t) − ψ(β, B) a.s. (3.2.4) N→∞ is then obtained from the existence of the pressure in the thermodynamic limit, as stated in Theorem 2.5.3. Moreover, it results that c(t) is finite. From the differentiability of the infinite-volume random pressure with respect to B (see Theorem 2.5.3), we also obtain ∂ ∂ c0(t) = [ψ(β, B + t) − ψ(β, B)] = [ψ(β, B + t)] , ∂t ∂B and hence ∂ c0(0) = [ψ(β, B)] = M(β, B). ∂B Thus, by Theorem 3.2.1, we obtain the exponential convergence.
Note that, as a consequence of Borel–Cantelli Lemma, the exponential conver- gence in probability implies the convergence almost surely (see [26, Theorem II.6.4]).
35 Random quenched measure
Below, we show the central limit theorem proving the convergence of the moment generating function. In order to do this, we again use the cumulant generating function and its derivatives together with the GHS inequality.
3.2.3 Random Quenched CLT: Proof of Theorem 3.1.2
We give the proof for B ≥ 0 only, the case B < 0 is handled similarly. The strategy of the proof is to show that w.r.t. the random quenched measure the moment generating function of the random variable
SN − NMN (β, B) VN = √ (3.2.5) N converges to the moment generating function of a Gaussian random variable with variance χ(β, B) given in (2.5.8), i.e., 1 2 lim µG (exp (tVN )) = exp χ(β, B)t for all t ∈ [0, α) , (3.2.6) N→∞ N 2 and some α > 0. This can be done by expressing µGN (exp (tVN )) in terms of the second derivative of the cumulant generating function cN (t). A simple computation shows that 0 1 µGN (SN exp (tSN )) SN cN (t) = = µGN (β,B+t) , (3.2.7) N µGN (exp (tSN )) N where, in order to stress the dependence on the magnetic field, we have used the symbol µGN (β,B+t) (·) to denote the µGN -average in the presence of the field B + t. Thus,