1: \documentclass[12pt]{article}
2: \usepackage{amsfonts}
3: %\usepackage{amsfonts,psfig}
4: \usepackage{amsfonts}
5: \usepackage{graphicx}
6: \usepackage{amsmath}
7: \usepackage{amssymb}
8: \usepackage{amsthm}
9: %\usepackage{natbib}
10: %\usepackage{showkeys}
11: %\textwidth=160mm \oddsidemargin=0mm \topmargin=0mm \headheight=0mm
12: %\headsep=0mm
13: % \textheight=230mm
14:
15: \def\LT{{\mathbb{LT}}}
16:
17: \input pre.tex
18:
19: \def\mucr{{\mu_{cr}}}
20: \def\wtY{{\widetilde{Y}}}
21:
22: \DeclareMathAlphabet{\mathpzc}{OT1}{pzc}{m}{it}
23: \newcommand{\T}{\mathbf{T}}
24: \newcommand{\ds}{\,\nu(d\sigma)}
25: \newcommand{\x}{{{\bf x}}}
26: \newcommand{\y}{{{\bf y}}}
27: \newcommand{\z}{{{\bf z}}}
28: \newcommand{\X}{\mathbf{X}}
29: \newcommand{\Y}{\mathbf{Y}}
30: \newcommand{\U}{\mathbf{U}}
31: %\newcommand{\Z}{\mathbf{Z}}
32: \newcommand{\ed}{\mathrm{d}}
33: %\newcommand{\const}{{\rm const}}
34: \newcommand{\diam}{\mathrm{diam}\, }
35: \newcommand{\ovl}[1]{\overline{#1}}
36: %\newcommand{\wt}[1]{\widetilde{#1}}
37: \newcommand{\wh}[1]{\widehat{#1}}
38: \newcommand{\rr}{\ovl{r}}
39: \newcommand{\und}[1]{\ensuremath{\underline{#1}}}
40: \newcommand{\vv}[1]{\marginpar{\scriptsize #1}}
41: \newcommand{\reff}[1]{(\ref{#1})}
42: \newcommand{\pr}[1]{\vspace{.5cm}\textit{Proof of Theorem} \ref{#1}}
43:
44: \numberwithin{equation}{section}
45: %\newtheorem{theorem}{Theorem}
46: %\newtheorem{lemma}{Lemma}[section]
47: %\newtheorem{stat}{Statement}[section]
48: %\newtheorem{lemma}{Lemma}
49: %\newtheorem{definition}{Definition}
50:
51: \theoremstyle{remark}
52: \newtheorem{rem}{Remark}[section]
53:
54: \theoremstyle{definition}
55: \newtheorem{df}{Definition}[section]
56: \newtheorem{cond}{Condition}[section]
57: %\newtheorem{pr}{Proof of Theorem \ref{ }}
58:
59: %\newcommand{\R}{{\mathbb R}}
60: %\newcommand{\Z}{{\mathbb Z}}
61: %\newcommand{\N}{{\mathbb N}}
62: %\renewcommand{\P}{{\sf P}}
63: %\newcommand{\E}{{\sf E}}
64: \newcommand{\V}{{\sf V}}
65: %\renewcommand{\L}{{\mathfrak L}}
66:
67: %\newcommand{\eps}{\varepsilon}
68:
69: \title{Phase transition for the Ising model on the Critical Lorentzian triangulation}
70: \author{Maxim~Krikun%
71: \thanks{Current affiliation: Google Ireland, Dublin, Ireland.}
72: \thanks{Partially supported by BQR 2007 UMR7502 IAEM 0039.}$^{~,1}$%
73: \and Anatoly~Yambartsev\thanks{Partially supported by the ``Rede
74: Matem\'{a}tica Brasil-Fran\c{c}a, CNPq (306092/2007-7)
75: and CNPR ``Edital Universal 2006'' (471925/2006-3).}$^{~,2}$
76: }
77:
78: \begin{document}
79: \maketitle {\footnotesize
80:
81: \noindent $^1$
82: Institut \'Elie Cartan Nancy (IECN), Nancy-Universit\'e, CNRS, INRIA,
83: Boulevard des Aiguillettes B.P. 239 F-54506 Vand{\oe}uvre-l\`es-Nancy, France. \\
84: E-mail: krikun@iecn.u-nancy.fr\\
85: {current affiliation: Google Ireland, Dublin, Ireland}
86:
87: \noindent $^2$ Department of Statistics, Institute of Mathematics
88: and Statistics, University of S\~ao Paulo, Rua do Mat\~ao 1010,
89: CEP 05508--090, S\~ao Paulo SP, Brazil.\\
90: E-mail: yambar@ime.usp.br}
91:
92:
93: \begin{abstract}
94: Ising model without external field on an infinite Lorentzian triangulation
95: sampled from the uniform distribution is considered. We prove
96: uniqueness of the Gibbs measure in the high temperature region
97: and coexistence of at least two Gibbs measures at low temperature.
98: The proofs are based on the disagreement percolation method and
99: on a variant of Peierls method.
100: The critical temperature is shown to be constant a.s.
101:
102:
103: \smallskip
104: \noindent {\it Keywords:} \/Lorentzian triangulation, Ising model,
105: dynamical triangulation, quantum gravity
106:
107: \noindent AMS 2000 Subject Classifications: 82B20, 82B26, 60J80
108:
109: \end{abstract}
110:
111:
112: \section{Introduction}
113:
114: Triangulations, and planar graphs in general, appear in physics
115: in the context of 2-dimensional quantum gravity as a model for
116: the discretized time-space. Perhaps the best understood it the
117: model of Euclidean Dynamical Triangulations, which can be viewed
118: as a way of constructing a random graph by gluing together a
119: large number of equilateral triangles in all possible ways,
120: with only topological conditions imposed on such gluing.
121: Putting a spin system on such a random graph can be interpreted
122: as a {\em coupling of gravity with matter}, and was an object of
123: persistent interest in physics since the successful application
124: of matrix integral methods to the Ising model on random lattice
125: by Kazakov \cite{kazakov}.
126:
127: More recently, a model of Casual Dynamical Triangulations was
128: introduced (see \cite{ambjorn} for an overview).
129: The distinguishing feature of this
130: model is its lack of isotropy --- the triangulation now has a
131: distinguished time-like direction, giving it a partial order
132: structure similar to Minkowski space, and imposing some
133: non-topological restrictions on the way elementary triangles
134: are glued.
135: This last fact destroys the connection between the model and
136: matrix integrals, in particular the analysis of the Ising model
137: requires completely different methods (see e.g.~\cite{benedetti}).
138:
139: From a mathematical perspective, we deal here with nothing but
140: a spin system on a random graph.
141: Random graphs, arising from the CDT approach,
142: were considered in~\cite{myz} under the name of Lorentzian models.
143: In the present paper we consider the Ising model on such graphs.
144: %
145: When defining the model we pursue the formal Gibbsian approach \cite{georgii};
146: namely, given a realization of an infinite triangulation,
147: we consider probability measures on the set of spin configurations
148: that correspond to a certain formal Hamiltonian.
149:
150: Our setting is drastically different from e.g.~\cite{kazakov} and~\cite{benedetti}
151: in that we do not consider ``simultaneous randomness'', when both
152: the triangulations and spin configurations are included into
153: one Hamiltonian. Instead we first sample an infinite triangulation
154: from some natural ``uniform'' measure, and then run an Ising model
155: on it, thus the resulting semi-direct product measure is ``quenched''.
156:
157: A modest goal of this paper is to establish a phase transition
158: for the Ising model in the above described ``quenched'' setting
159: (the ``annealed'' version of the problem is surely interesting,
160: but is also more technically challenging, so we don't attempt
161: it for the moment).
162: In Section~3 we use a variant of Peierls method to prove
163: non-uniqueness of the Gibbs measure at low temperature.
164: Quite surprisingly, proving the uniqueness at high temperature
165: is not easy -- the difficulty consists in presence of vertices
166: of arbitrarily large degree, which does not allow for immediate
167: application of uniqueness criteria such as e.g.~\cite{weitz}.
168: We resort instead to the method of disagreement percolation~\cite{berg},
169: and use the idea of ``ungluing'', borrowed from the paper~\cite{bassalygo},
170: to get rid of vertices of very high degree.
171:
172: Finally, in Section~5 we show that the critical temperature
173: is in fact non-random and coincides for a.e. random Lorentzian triangulation.
174: Section~2 below contains the main definitions
175: and summarizes some of the results of~\cite{myz}.
176:
177: We thank E.~Pechersky for numerous useful discussions during the preparation
178: of this paper.
179:
180:
181: \section{Definitions and Main Results}
182:
183: Now we define rooted infinite Lorentzian triangulations in a
184: cylinder ${\sf C}=S^1\times [0,\infty)$.
185:
186: \begin{definition}
187: Consider a connected graph $G$ embedded in a cylinder ${\sf C}$.
188: A face is a connected component of ${\sf C} \setminus G$. The face
189: is a triangle if its boundary meets precisely three edges of the
190: graph. An embedded triangulation $T$ is such a graph $G$ together
191: with a subset of the triangular faces of $G$. Let the support
192: $S(T)\subset {\sf C}$ be the union of $G$ and the triangular faces
193: in $T$. Two embedded triangulations $T$ and $T'$ are considered
194: equivalent if there is a homeomorphism of $S(T)$ and $S(T')$ that
195: corresponds $T$ and $T'$.
196: \end{definition}
197: For convenience, we usually abbreviate ``equivalence class of
198: embedded triangulations'' to ``triangulation''. This should not
199: cause much confusion. We suppose that the number of the vertices
200: of $G$ is finite or countable.
201:
202: \begin{definition}
203: A triangulation $T$ of ${\sf C}$ is called Lorentzian if the
204: following conditions hold: each triangular face of $T$ belongs to
205: some strip $S^1\times[j,j+1], j=0,1,\dots,$ and has all vertices
206: and exactly one edge on the boundary $(S^1\times\{j\})\cup
207: (S^1\times\{j+1\})$ of the strip $S^1\times[j,j+1]$; and the
208: number of edges on $S^1\times\{j\}$ is positive and finite for any
209: $j=0,1, \dots.$ \end{definition} In this paper we will consider
210: only the case when the number of edges on the first level $S^1
211: \times \{0\}$ equal to 1. This is not restriction, only it gives
212: formulas more clean.
213:
214: \begin{definition}
215: A triangulation $T$ is called rooted if it has a root. The root in
216: the triangulation $T$ consists of a triangle $t$ of $T$, called
217: the root face, with an ordering on its vertices $(x,y,z)$. The
218: vertex $x$ is the root vertex and the directed edge $(x,y)$ is the
219: root edge. The $x$ and $(x,y)$ belong to $S^1\times \{0\}$.
220: \end{definition}
221:
222: Note that this definition also means that the homeomorphism in the
223: definition of the equivalence class respects the root vertex and
224: the root edge. For convenience, we usually abbreviate
225: ``equivalence class of embedded rooted Lorentzian triangulations''
226: to ``Lorentzian triangulation'' or LT.
227:
228: In the same way we also can define a Lorentzian triangulation of a
229: cylinder ${\sf C}_N=S^1\times [0,N].$ Let $\LT_N$ and
230: $\LT_{\infty}$ denote the set of Lorentzian triangulations with
231: support ${\sf C}_N$ and ${\sf C}$ correspondingly.
232:
233: \subsection{ Gibbs and Uniform Lorentzian triangulations.}
234: Let $\LT_N$ be the set of all Lorentzian triangulations with only
235: one (rooted) edge on the root boundary and with $N$ slices. The
236: number of edges on the upper boundary $S^1\times \{N\}$ is not
237: fixed. Introduce a Gibbs measure on the (countable) set $\LT_N$:
238: \begin{equation}
239: {\sf P} _{N,\mu}(T)= Z_{[0,N]}^{-1} \exp(-\mu F(T)), \label{eq.Gm}%
240: \end{equation}
241: where $F(T)$ denotes the number of triangles in a triangulation
242: $T$ and $Z_{[0,N]}$ is the partition function:%
243: \[
244: Z_{[0,N]}=\sum_{T\in\LT_N}\exp(-\mu F(T)).
245: \]
246: The measure on the set of infinite triangulations $\LT_\infty$ is
247: then defined as a weak limit %when $N\to\infty$,
248: \[ \P_\mu := \lim_{N\to\infty} \P_{N,\mu}. \]
249: It was shown in
250: \cite{myz} that this limit exists for all $\mu\ge \mucr := \ln2$.
251: \begin{theorem}[\cite{myz}]\label{Tmyz}
252: Let $k_n$ be the number of vertices at $n$-th level in a triangulation $T$
253: for each $n\ge0$.
254: \begin{itemize}
255: \item For $\mu>\ln2$ under the limiting measure $\P_\mu$
256: the sequence $\{ k_n \}$ is a positive recurrent Markov chain.
257: \item For $\mu = \mucr = \ln 2$ the
258: sequence $\{k_n\}$ is distributed as the branching process $\xi_n$
259: with geometric offspring distribution with parameter $1/2$,
260: conditioned to non-extinction at infinity.
261: \end{itemize} \end{theorem}
262: Below we briefly sketch the proof of the second part of Theorem~\ref{Tmyz},
263: a deeper investigation of related ideas will appear in~\cite{syz}.
264:
265: Given a triangulation $T\in\LT_N$,
266: define the subgraph $\tau \subset T$ by taking,
267: for each vertex $v \in T$, the leftmost edge going from $v$ downwards
268: (see \figref{lt-tree}).
269: The graph thus obtained is a spanning forest of $T$,
270: and moreover, if one associates with each vertex of $\tau$ it's height in $T$
271: then $T$ can be completely reconstructed knowing $\tau$.
272: %
273: We call $\tau$ the \defined{tree parametrization} of $T$.
274:
275: \putfigurescale{lt-tree}{Tree parametrization}
276:
277: For every vertex $u\in\tau$ denote by $\delta_u$ it's out-degree,
278: i.e.~the number of edges of $\tau$ going from $u$ upwards.
279: Comparing the out-degrees in $\tau$ to the number of vertical
280: edges in $T$, and comparing the latter to the total number of
281: triangles, it is not hard to obtain the identity
282: \begin{equation}\label{eq.sumdeg}
283: \sum_{u \in \tau \bs S^1\times\{N\}} (\delta_u + 1) = F(T)
284: \end{equation}
285: (the sum on the left runs over all vertices of $\tau$ except for the $N$-th level).
286: Thus under the measure $\P_{N,\mucr}$ the probability of a forest $\tau$
287: is proportional to
288: \begin{equation}\label{eq.emucr}
289: e^{-\mucr F(T) }
290: = \prod_{u \in \tau \bs S^1\times\{N\}} \B(\f12\B)^{\delta_u+1}
291: \end{equation}
292: which is exactly the probability to observe $\tau$ as a
293: realization of a branching process with offspring distribution
294: ${\rm Geom}(1/2)$. After normalization we'll obtain, on the left
295: in \eqref{emucr}, the probability $\P_{N,\mu}(\tau)$ as defined by
296: \eqref{Gm}, an on the right the {\em conditional} probability to
297: see $\tau$ as a realization of the branching process $\xi$ given
298: $\xi_N>0$.
299: %
300: So quite naturally when $N\to\infty$ the
301: distribution of $\tau$ converges to the Galton-Watson tree,
302: conditioned to non-extinction at infinity.
303:
304: In particular it follows from Theorem~\ref{Tmyz} that
305: \begin{equation} \label{GW}
306: {\sf P}_\mucr (k_n=m) = Pr(\xi_n = m \mid \xi_\infty >0 ) = m
307: Pr(\xi_n = m).
308: \end{equation}
309:
310: \begin{rem}\label{note1}
311: The last equality in ($\ref{GW}$) means that the measure $P_\mucr$
312: on triangulations can be considered as a $Q$-process defined by
313: Athreya and Ney \cite{AN} for a critical Galton-Watson branching
314: process. Such a process is exactly a critical Galton-Watson tree
315: conditioned to survive forever. \end{rem}
316:
317: We will also note for further use that the offspring generating function
318: of the branching process $\xi$,
319: \[ \psi(s) = \sum_{k\ge 0} \B(\f12\B)^{k+1} s^k = \f{s}{2-s}, \]
320: and the generating function for $\xi_n$ (with initial condition $\xi_0=1$) is
321: \[ \psi_n(s) = \f{ n-(n-1)s }{ n+1 - ns }. \]
322:
323:
324:
325:
326:
327:
328:
329:
330:
331: \subsection{Ising model on Uniform Infinite Lorentzian triangulation
332: -- quenched case.}
333:
334: Let $T$ be some fixed Lorentzian triangulation, $T\in \LT_\infty.$
335: Let $T_N$ be the projection of $T$ on the cylinder ${\sf C}_N.$
336: We associate with every vertex $v$ a spin
337: %$\sigma_v$,
338: %with two possible values
339: $\sigma_v\in \{-1, 1\}$.
340: Let $\Sigma(T)$ and $\Sigma_N(T)$ denote the set of of all spin
341: configurations on $T$ and $T_N$, respectively.
342: %and $\Sigma_N(T)$ be the set of configurations on the finite triangulation $T_N.$
343: The Ising model on $T$ is defined by a formal Hamiltonian
344: %, let $\sigma$ be a configuration
345: \begin{equation} \label{Hm} H(\sigma) = \sum _{
346: \langle v, v' \rangle \in V} \sigma_v \sigma_{v'} \end{equation}
347: where $\langle v, v' \rangle $ means that vertices $v, v'$ are
348: neighbors, i.e.~are connected by an edge in $T$.
349: %-- there exists edge from $T$ which connect them.
350: Let $\partial T_N$ be the set of vertices of $T$ that lie on the
351: circle $S^1 \times \{N+1\}.$ Fix some configuration on the
352: boundary $\partial T_N$ and denote it $\partial \sigma.$ The Gibbs
353: distribution with boundary condition $\partial \sigma$ is
354: defined by the following. Let $V(T_N)$ be the set of all vertices
355: in $T_N,$ then the energy of configuration $\sigma \in \Sigma
356: _N(T)$ is \begin{equation} \label{Hb} H_N (\sigma | \partial
357: \sigma) = \sum_{ \langle v,v' \rangle:\ v,v'\in V(T_N) } \sigma_v
358: \sigma_{v'} + \sum _{ \langle v,v' \rangle: v\in V(T_N), v'\in
359: \partial T_N } \sigma_v \sigma_{v'} \end{equation} which defines
360: the probability
361: \begin{equation} \label{GN} P^T_{N, \partial \sigma} ( \sigma ) =
362: \frac{\exp\{ - \beta H_N (\sigma | \partial \sigma) \} }{ Z_{N,
363: \partial \sigma} (T)}
364: \end{equation} where $$ Z_{N, \partial \sigma} (T) = \sum _{\sigma \in
365: \Sigma_N(T)} \exp\{ - \beta H_N (\sigma | \partial \sigma) \}. $$
366: When $N\to\infty$, for any sequence of boundary conditions $\d \sigma$,
367: a limit (at least along some subsequence) of measures $P^T_{N, \d\sigma}$
368: exists by compactness.
369: Such a limit is a probability measure on $\Sigma(T)$ with a natural
370: $\sigma$-algebra, which we refer to as a Gibbs measure.
371:
372: In general, it is well known that at least one Gibbs measure exists for
373: the Ising model on any locally finite graph and for any value of
374: the parameter $\beta$ (see, e.g., \cite{GM} page 71).
375: %A more substantial question is whether such a measure is unique or not.
376: It is also known that the existence of more than one Gibbs measure
377: is increasing in $\beta$,
378: i.e.~there exists a critical value $\beta_c \in [0,\infty]$ such
379: that there is a unique Gibbs measure when $\beta > \beta_c$,
380: and multiple Gibbs measures when $\beta < \beta_c$ (see
381: \cite{haggstrom} for an overview of relations between percolation
382: and Ising model on general graphs).
383:
384: Thus when considering the Ising model on Lorentzian triangulations
385: it is natural to ask whether the critical temperature is finite
386: (different from both $0$ and $\infty$), and whether it depends on the
387: triangulation. In the following two sections we show that the critical
388: temperature is a.s.~bounded both from $0$ and $\infty$.
389: In the last section we prove that the critical temperature obeys
390: a zero-one law and is therefore a.s.~constant.
391:
392: \section{Phase transition at low temperature}
393: In this section we prove the following theorem.
394: \begin{theorem} \label{Th1} There
395: exists a $\beta_0$ such that for all $\beta \in (\beta_0, \infty)$
396: there exist at least two Gibbs measures for ${\sf P}_\mucr$-a.e.
397: $T$.
398: \end{theorem}
399:
400: We remind first the classical Peierls method for the Ising model
401: on $\Z^2$. Let $\Lambda \subset \Z^2$ be a large square box centered at
402: the origin, and let $P^-_\Lambda$ be the distribution for the
403: spins in $\Lambda$ under the condition that all of the spins
404: outside of $\Lambda$ are negative. When the size of the box tends
405: to infinity, $P^-_\Lambda$ converges to some probability measure
406: $P^-$, which we call Gibbs measure with negative boundary
407: conditions. The measure $P^+$ is defined similarly
408: taking positive boundary conditions.
409:
410: Due to an obvious symmetry between $P^-$ and $P^+$ we have
411: \[ P^-(\sigma_0=+1) = P^+(\sigma_0=-1), \]
412: therefore if the Gibbs measure is unique then necessarily $P^-=P^+$ and
413: \begin{equation}\label{eq.pleq}
414: P^-(\sigma_0=+1) = P^-(\sigma_0=-1) = 1/2.
415: \end{equation}
416: Thus our goal will be to disprove \eqref{pleq} for large enough $\beta$.
417:
418: Consider some configuration $\sigma$ such that $\sigma_0=+1$ and
419: $\sigma_i=-1$ for all $i\notin \Lambda$.
420: Define the \defined{cluster} $K_+\subset \Z^2$
421: as the maximal connected subgraph of $\Z^2$,
422: containing the origin, such that $\sigma_i = +1$ for all $i \in K_+$.
423: Let $\gamma$ be the contour in the dual graph $(\Z^2)^*$,
424: corresponding to the outer boundary of $K_+$,
425: and let $\sigma'$ be the configuration obtained from $\sigma$
426: by inverting all the spins inside $\gamma$.
427: Note that $\sigma'_0 = -1$.
428: Let $C$ be the set of all contours in $(\Z^2)^*$ surrounding the origin,
429: and consider the application
430: \begin{equation}
431: \begin{array}{rrcl}
432: {\rm \bf inv}: & \Sigma |_{\sigma_0=+1} & \rightarrow & \Sigma |_{\sigma_0=-1} \times C \\
433: & \sigma & \rightarrow & (\sigma', \gamma).
434: \end{array}
435: \end{equation}
436: Clearly ${\rm \bf inv}$ is injective: given $(\sigma',\gamma)$ one can
437: easily reconstruct $\sigma$ by inverting the spins inside
438: $\gamma$. Also the following property holds:
439: \begin{equation} \label{eq.plsig}
440: P^-_\Lambda ( \sigma ) = P^-_\Lambda ( \sigma' ) \cdot e^{-2\beta |\gamma|},
441: \end{equation}
442: where $|\gamma|$ denotes the length of $\gamma$. Indeed, for every
443: edge $\langle i,j \rangle$ traversed by $\gamma$ we have $\sigma_i
444: \neq \sigma_j$ by construction, but after inversion $\sigma'_i =
445: \sigma'_j$, so the contribution of the pair $\langle i,j \rangle$
446: to the Hamiltonian is increased by $2$. On the other hand, for all
447: other pairs $\langle i,j \rangle$ the contribution to the
448: Hamiltonian doesn't change. Informally we can write $H(\sigma) -
449: H(\sigma') = 2|\gamma|$, which is equivalent to~\eqref{plsig}.
450:
451: Now taking a sum over all $\sigma: \sigma_0=+1$ in \eqref{plsig}
452: we get
453: \begin{eqnarray}
454: P^-_{\Lambda}(\sigma_0=+1)
455: &=& \sum_{\sigma: \sigma_0 = +1} P^-_\Lambda(\sigma) \nonumber\\
456: &\le& \sum_{\sigma': \sigma'_0=-1, \gamma \in C}
457: P^-_\Lambda(\sigma') \cdot e^{-2\beta|\gamma|}
458: \nonumber\\
459: &=& P^-_\Lambda(\sigma_0=-1) \cdot \sum_{n\ge 1} e^{-2\beta n} \# C_n,
460: \end{eqnarray}
461: where $C_n$ is the set of contours $\gamma\in C$ of length exactly $n$.
462:
463: But $\# C_n \le n 3^n$: indeed, every contour $\gamma \in C_n$ intersects the
464: $x$-axis somewhere between $0$ and $n$.
465: Starting from this intersection point there are at most $3^n$ distinct
466: self-avoiding paths of length $n$ in $(\Z^2)^*$,
467: only a few of them really belonging to $C_n$.
468: %
469: Therefore we have the inequality
470: \begin{equation}\label{eq.plsum}
471: P^-_\Lambda(\sigma_0=+1)
472: \le
473: P^-_\Lambda(\sigma_0=-1) \cdot \sum_{n\ge 1} e^{-2\beta n} n 3^n,
474: \end{equation}
475: and by taking $\beta$ large enough the sum in the right-hand side
476: of~\eqref{plsum} can be made strictly less than one. Taking the
477: limit $\Lambda \to \Z^2$ we establish
478: \[ P^-(\sigma_0=+1) < P^-(\sigma_0=-1). \]
479: Thus $P^- \neq P^+$ and the Gibbs measure for the Ising model in
480: $\Z^2$ at inverse temperature $\beta$ is not unique.
481:
482: \putfigurescale{z2-lt--contours}{Peierls' method on $\Z^2$ (left)
483: and on a random Lorentzian triangulation (right).
484: Dashed line represents the inversion contour $\gamma$.}
485:
486:
487: \noindent Next we will slightly generalize the above classical
488: argument.
489: \begin{lemma} \label{GenP} Let $G$ be an infinite planar graph
490: and $G^*$ it's planar dual. Let $v_0$ be a vertex of $G$ and
491: denote by $C_n(v_0)$ be the set of contours of length $n$ in
492: $G^*$, separating $v_0$ from the infinite part of the graph.
493:
494: If the following sum is finite,
495: \begin{equation}\label{eq.sfin}
496: \sum_{n\ge 1} \# C_n(v_0) e^{-2\beta n} < \infty,
497: \end{equation}
498: then the Gibbs measure for the Ising model on $G$ at inverse
499: temperature $\beta$ is not unique.
500: \end{lemma}
501:
502: \begin{rem} \label{note2} Strictly speaking, in the above Lemma
503: we should also require the graph $G$ to be one-ended;
504: i.e.~for any finite subgraph $B\subset G$
505: the complement $G\backslash B$ must contain exactly one
506: infinite connected component.
507: The reason for this is that if the graph $G$ fails to be one-ended
508: (e.g.~a doubly-infinite cylinder)
509: then the outer boundary of the cluster $K_+$
510: may happen to have multiple connected components,
511: which slightly complicates the argument.
512: We don't consider this case in detail since
513: Lorentzian triangulations are one-ended by construction.
514: \end{rem}
515:
516: \proof Assume that \eqref{sfin} holds. Then for some large $N$
517: \[ \sum_{n\ge N} \#C_n(v_0) e^{-2\beta n} < 1. \]
518: Also, for every $n$ the set $C_n(v_0)$ is finite,
519: so there exists some large $R$ such that the ball $B_R(v_0)$ contains
520: all of the $C_n(v_0)$, $n=1\ldots N$.
521:
522: Let now $P^+$ and $P^-$ be the Gibbs measures for the Ising model on $G$,
523: constructed
524: %as a weak limit over some exhaustion
525: with positive and negative boundary conditions respectively,
526: and let us compare the events
527: \[ A_+ := ( \sigma_v = +1 \text{ for all } v \in B_R(v_0) ) \]
528: and
529: \[ A_- := ( \sigma_v = -1 \text{ for all } v \in B_R(v_0) ). \]
530: Proceeding as in the case of $\Z^2$ above, we can show that
531: \[ P^-(A_+) \le P^-(A_-) \cdot \sum_{n\ge 1} e^{-2\beta n} \# C_n( B_R(v_0)), \]
532: where $C_n( B_R(v_0))$ now is the set of contours of length $n$ in $G^*$
533: that surround the whole ball $B_R(v_0)$.
534: But by construction we have $C_n(B_R(v_0))=\emptyset$ for $n\le N$
535: and $C_n(B_R(v_0)) \subset C_n(v_0)$ for all other $n$,
536: %because every contour that surrounds the ball $B_R(v_0)$ also surrounds the vertex $v_0$.
537: therefore
538: \[ \sum_{n\ge 1} e^{-2\beta n} \# C_n( B_R(v_0))
539: \le
540: \sum_{n\ge N} e^{-2\beta n} \# C_n(v_0) < 1
541: \]
542: and
543: \[ P^-(A_+) < P^-(A_-). \]
544: Since by symmetry $P^+(A_-) = P^-(A_+)$, it follows that $P^-\neq
545: P^+$. \qed
546:
547: \begin{lemma} \label{l2} Let $T$ be a random Lorentzian triangulation,
548: and let $v_0$ be the root vertex of $T$. There exists $\beta_0$
549: such that for every $\beta > \beta_0$ \begin{equation}
550: \label{finit} \sum_{n\ge 1} \# C_n(v_0) e^{-2 \beta n} < \infty
551: \quad \mbox{$\P_\mucr$-a.e.},
552: \end{equation} where $\# C_n(v_0)$ is defined as in Lemma~\ref{GenP}.
553: \end{lemma}
554:
555: \proof In order to prove (\ref{finit}) it will be sufficient to
556: show that the expectation (with respect to the measure $\P_\mucr$)
557: of the sum is finite
558: \begin{equation}\label{mcont} {\sf E}_\mucr \sum_{n=0}^\infty \# C_n(0)
559: e^{-2\beta n} < \infty.
560: \end{equation}
561:
562: First of all we choose in any Lorentzian triangulation $T$ a vertical
563: path $\gamma_\infty = \gamma_\infty(T) = (v_0, v_1, \dots )$
564: starting at the root $v_0$ and such that $v_i\in S^1\times \{i\}$.
565: Let $C_{R,n}(v_0) \subset C_n(v_0)$ be the set of contours of
566: length $n$ which surround $v_0$ and intersect $\gamma_\infty$ at
567: height $R$. Note that any such contour does not exit from the
568: strip $[R-n, R+n]$.
569: %
570: Let also $S_{R,n}$ be the number of particles in the tree
571: parametrization of $T$ at height $R-n$ which have nonempty offspring in
572: the generation located at height $R+n$. Since every contour from
573: $C_{R,n}$, in order to surround $v_0$, must cross each of the
574: $S_{R,n}$ corresponding subtrees, we have
575: \[ (S_{R,n} > n ) \,\Rightarrow\, (C_{R,n}=\emptyset). \]
576: On the other hand
577: \[ C_{R,n} \le 2^n \]
578: since the contours $C_{R,n}$ live on the dual graph $T^*$, which
579: has all vertices of degree $3$, thus there are at most $2^n$
580: self-avoiding paths with a fixed starting point (which is in our
581: case the intersection with $\gamma_\infty$).
582:
583: Therefore we have
584: \begin{equation} \label{est} {\sf E} _\mucr
585: \sum_{n=0}^\infty \# C_n(v_0) e^{-2\beta n} \le \sum_{n=0} ^\infty
586: e^{-2\beta n} 2^n \sum_{R=1} ^\infty {\sf P} _\mucr ( S_{R,n} \le
587: n )
588: \end{equation}
589: and for this sum to be finite it's enough to show that sum over
590: $R$ has polynomial order in $n$.
591: %
592: First let us estimate
593: $$ \sum_{R=1}
594: ^\infty {\sf P} _\mucr ( S_{R,n} \le n ) \le n^5 + \sum_{R\ge n^5}
595: ^\infty {\sf P} _\mucr ( S_{R,n} \le n ).
596: $$
597: and write for some $\epsilon \in (0,1/4)$
598: \begin{eqnarray} \P _\mucr ( S_{R,n} \le n ) &=& \sum _{k\le R^
599: \epsilon } {\sf P}_\mucr ( S_{R,n} \le n \mid k_{R-n } = k) {\sf
600: P} _\mucr ( k_{R-n} =k ) \nonumber\\
601: & + & \sum _{k> R^{\epsilon} } {\sf P} _\mucr ( S_{R,n} \le n \mid
602: k_{R-n } = k ) {\sf P} _\mucr ( k_{R-n} =k ). \label{eq.sparts}
603: \end{eqnarray}
604: The first sum above satisfies the inequality
605: \begin{equation} \label{est3} \sum _{k\le R^ \epsilon } {\sf P} _\mucr
606: ( k_{R-n} =k ) = {\sf P} _\mucr ( k_{R-n} \le R^ \epsilon ) \le
607: \Bigl( \frac{ R^\epsilon}{ R-n} \Bigr) ^2.
608: \end{equation}
609: Indeed, thanks the representation (\ref{GW}) the generating
610: function $\phi_n(s)$ of $k_n$ is related to the generating function
611: $\psi_n(s)$ of the branching process $\xi_n$ described above by
612: the following equation
613: \begin{equation} \label{gen.func} \phi_n(s) =
614: s\psi_n'(s). \end{equation}
615: Since $\psi_n(s)$ can be calculated explicitly
616: \begin{equation} \label{gen.GW}
617: \psi_n(s) = \frac{ n - (n-1) s}{ n+1 - ns},
618: \end{equation}
619: from (\ref{gen.func}) and (\ref{gen.GW}) we obtain
620: \begin{equation}
621: \phi_n(s) = \frac{ s}{ (1+ n - ns)^ 2},
622: \end{equation}
623: \begin{eqnarray} \nonumber
624: {\sf P} _\mucr ( k_n = k ) &=& \frac{1}{k!} \frac{d^k}{ds^k} \phi_n(s) \Bigl|_{s=0} \\
625: %&=& n^{k-1} \frac{ k(n+1) + ns}{ (1+n-ns) ^{k+2}} \Bigl|_{s=0} =
626: &=& \f{k n^{k-1}}{(n+1)^k} \le \f{k}{n^2},
627: %\frac{k}{ n^2 (1+ 1/n)^{k+1} }
628: \label{prob1}
629: \end{eqnarray}
630: and
631: \[
632: \sum_{k\le R^\epsilon} \P _\mucr ( k_{R-n}=k ) \le \sum_{k\le
633: R^\epsilon} \f{k}{(R-n)^2} \le \f{ R^{2\epsilon}}{(R-n)^2},
634: \]
635: which proves the estimate (\ref{est3}).
636: Since $\epsilon < 1/4$ there exists a constant $c_1$ such that
637: \begin{equation} \sum_{R\ge n^5} \frac{ R^{2\epsilon} } {(R-n)^2}
638: < c_1. \end{equation}
639: The second sum in \eqref{sparts} can be estimated by the
640: probability $\P _\mucr ( S_{R,n} \le n \mid k_{R-n} = R^\epsilon).$
641: To do this, let us remind some properties of conditioned or size-biased
642: Galton-Watson trees (see e.g.~\cite{peres}, \cite{geiger} or \cite{AN}).
643: If $\tau$ is a Galton-Watson tree of a critical branching process,
644: conditioned to non-extinction at infinity, then conditionally on
645: $k_{R-n}$ the subtrees of $\tau$, originating on the $(R-n)$-th level,
646: are distributed as follows: one subtree, chosen uniformly at random,
647: has the same distribution as $\tau$ (i.e.~is infinite),
648: while the remaining subtrees are regular Galton-Watson trees
649: corresponding to the original branching process (and are finite a.s.).
650:
651: The probability for a particle of the branching process $\xi$
652: to survive up to time $2n$ equals $1/(2n+1)$,
653: thus conditionally on $k_{R-n}$, $S_{R,n}$ is stochastically
654: minorated by the binomial distribution
655: with parameters $(k_{R-n}, \f{1}{2n+1})$. %success probability $1/2n.$
656: Using Hoeffding's inequality for binomial distribution we obtain
657: \begin{equation} \P _\mucr ( S_{R,n} \le n \mid k_{R-n} =
658: R^\epsilon ) < \exp \Bigl( -2 \frac{ (R^\epsilon/(2n+1) - n)^2
659: }{R^\epsilon } \Bigr) \end{equation} and the sum over $R>n^5$ is
660: bounded by some absolute constant $c_2$
661: \begin{equation} \sum_{R>
662: n^5} \exp \Bigl( -2 \frac{ (R^\epsilon/(2n+1) - n)^2 }{R^\epsilon
663: } \Bigr) \le e^2 \sum_{R> n^5} \exp \Bigl( - \frac{ 2R^\epsilon}
664: {(2n+1)^2 } \Bigr) < c_2. \end{equation}
665: Thus we have proved that
666: \begin{equation} \label{est2} {\sf E} _\mucr \sum_{n=0}^\infty C_n(0)
667: e^{-2\beta n} \le \sum_{n=0} ^\infty e^{-2\beta n} 2^n (n^5 + c_1
668: + c_2)
669: \end{equation} %where $c = c_1 + c_2.$
670: and so for all $\beta > \f{\ln 2}{2}$ the sum (\ref{finit}) is
671: finite. \qed
672:
673: \noindent
674: {\it Proof of Theorem~\ref{Th1}.}
675: Theorem~\ref{Th1} follows immediately from Lemmas~\ref{GenP} and~\ref{l2}.
676:
677:
678:
679:
680:
681:
682: % part 2
683:
684: \section{Uniqueness for high temperature}
685:
686: In this section we prove the following theorem
687: \begin{theorem}\label{T2}
688: There exists a small enough $\beta_h$ such that for every $\beta \in [0, \beta_h)$
689: for $\P_{\mucr}$-a.e.~Lorentzian triangulation $T$
690: the Gibbs measure for the Ising model on $T$ is
691: unique.
692: \end{theorem}
693:
694: The proof goes as follows. First, for a fixed triangulation $T$ we
695: use disagreement percolation to reduce the problem of uniqueness
696: of a Gibbs measure to the problem of existence of an infinite open
697: cluster under some specific site-percolation model on $T$. Then we
698: consider the joint distribution of a triangulation with this
699: percolation model on it (i.e.~with randomly open/closed vertices),
700: where the triangulation $T$ is distributed according to
701: $\P_{\mucr}$, and conditionally on $T$ the state of each vertex in
702: $T$ is chosen independently (but with vertex-dependent
703: probabilities). We show that the probability for such a random
704: triangulation to contain an infinite open cluster is $0$,
705: therefore this probability is $0$ for
706: $\P_{\mucr}$-a.e.~triangulation as well, and the theorem follows.
707:
708: \subsection{Disagreement percolation}
709:
710: The following result of Van den Berg and Maes \cite{berg}
711: provides a sufficient condition for the uniqueness of the Gibbs measure,
712: expressed in terms of some percolation-type problem.
713: \begin{theorem} %[VdB-M]
714: Let $G$ be a countably infinite locally finite graph with the vertex set $V$,
715: and let $S$ be a finite spin space.
716: %
717: Let $Y$ be a specification of a Markov field on $G$,
718: i.e.~for every finite $A\subset V$,
719: $Y_A(\cdot, \eta)$ is the conditional distribution of the spins on $A$,
720: given that the spins outside of $A$ coincide with the configuration $\eta \in S^V$.
721:
722: Define for every $v\in V$
723: \[ p_v := \max_{\eta, \eta'\in S^V} \, \rho_{\rm var}( Y_{\{v\}}(\cdot, \eta), Y_{\{v\}}(\cdot, \eta' )), \]
724: where $\rho_{\rm var}$ denotes the distance in variation. Let
725: $P^{p,G}$ be the product measure under which each vertex $v$ is
726: open with probability $p_v$ and closed with probability $1-p_v.$
727:
728: If under site-percolation on $G$, one has
729: \[
730: P^{p,G} \{ \mbox{there is an infinite open path}\} = 0,
731: \]
732: then the specification $Y$ admits at most one Gibbs measure.
733: \end{theorem}
734:
735: In the case of the Ising model, the only spins that affect the
736: distribution of the spin at a given vertex $v$ are those at
737: vertices $u$ connected to $v$ by an edge (denoted $u\sim v$). Let
738: $d_v$ be the degree of $v$ and let $S_v = \sum_{u\sim v}
739: \sigma_u$. The conditional distribution of $\sigma_v$, given the
740: values of spins $\sigma_u$ for all $u \sim v$, only depends on
741: $S_v$ and is given by
742: \[
743: P^T\{ \sigma_v = +1 | S_v \} = \f{ e^{-\beta S_v}}{ e^{-\beta
744: S_v} + e^{\beta S_v}}, \qquad P^T\{ \sigma_v = -1 | S_v \} = \f{
745: e^{\beta S_v}}{ e^{-\beta S_v} + e^{\beta S_v}}.
746: \]
747: The distance in variation between two such conditional distributions is maximized
748: by taking two extreme values of $S_v$, namely $S_v = d_v$ and $S_v = -d_v$,
749: corresponding to all-plus and all-minus boundary conditions.
750: Therefore we have
751: \begin{eqnarray*}
752: p_v &=&
753: \f12 \B(
754: \B| \f{e^{-\beta d_v}}{ e^{-\beta d_v} + e^{\beta d_v}} -
755: \f{e^{\beta d_v}}{ e^{-\beta d_v} + e^{\beta d_v}} \B| \\
756: &&{}
757: +
758: \B| \f{e^{-\beta d_v}}{ e^{-\beta d_v} + e^{\beta d_v}} -
759: \f{e^{\beta d_v}}{ e^{-\beta d_v} + e^{\beta d_v}} \B|
760: \B)
761: = \tanh(\beta d_v)
762: \end{eqnarray*}
763: %
764: and we need to prove the following
765: \begin{lemma}\label{L4}
766: Let $T$ be an infinite random Lorentzian triangulation
767: sampled from the measure $\P_{\mucr}$,
768: and let each vertex $v\in T$ be open with probability
769: $p_v = \tanh(\beta d_v)$ and closed otherwise.
770: Then for small enough $\beta$ the probability
771: that there is an infinite open path in $T$ is zero.
772: \end{lemma}
773:
774: \subsection{Elementary approach to non-percolation}
775:
776: Imagine for a moment that the graph $G$ that we deal with in
777: Lemma~\ref{L4} has uniformly bounded degrees. If with positive
778: probability there exists an infinite open percolation cluster,
779: then with positive probability this cluster will also be connected
780: to the origin. Consider the event
781: \[
782: A_n = \left\{{
783: \mbox{there exists a self-avoiding open path of length $n$}
784: \atop\mbox{starting at the origin.}
785: }\right\}
786: \]
787: The number of paths of length $n$, starting at the origin, is at most $D^n$,
788: where $D$ is the maximal degree of a vertex in our graph.
789: Obviously the same estimate is valid for the number of self-avoiding paths.
790: The probability for a vertex to be open is bounded by $\tanh(\beta D) \le \beta D$,
791: therefore we can estimate
792: \[ P^{p,G}(A_n) \le D^n \tanh(\beta D)^n \le \beta^n D^{2n}. \]
793: Taking $\beta < 1/{D^2}$, we get $P^{p,G} (A_n)\to 0$ as
794: $n\to\infty$, and therefore the probability that the origin
795: belongs to an infinite open cluster is zero.
796:
797: Unfortunately, the above argument can't be applied directly in the context of Lemma~\ref{L4}:
798: in an infinite random Lorentzian triangulation with probability one one will
799: encounter regions consisting of vertices of arbitrarily large
800: degrees, and these regions can be arbitrarily large as well.
801: On the other hand, we know that in a Lorentzian triangulation the degree
802: of a typical vertex is not too large (in some sense, which we're not going to make
803: precise in this paper, it can be approximated
804: by a sum of two independent geometric distributions plus a constant),
805: %namely the tail of it's distribution is decreasing exponentially.
806: therefore one can hope to show that the vertices of very large degree are so rare
807: that they don't mess up the picture.
808:
809: Before proceeding further, let us remark that in the above argument
810: one can consider, instead of the set of all the self-avoiding paths,
811: only the paths that are both self-avoiding and \defined{locally geodesic}.
812: We say that a path $\gamma$ in a graph $G$ is locally geodesic
813: if whenever two vertices, belonging to $\gamma$, are connected by and edge in $G$,
814: this edge also belongs to $\gamma$
815: (in other words, a locally geodesic path avoids making unnecessary detours).
816: %
817: Clearly, if under site percolation on $G$ there exists an infinite
818: open self-avoiding path starting at $v_0$, there exists also an
819: infinite self-avoiding locally geodesic path. It is also easy to
820: see that if $\gamma$ is a locally geodesic path, then every vertex
821: $v\in\gamma$ has at most two neighbors belonging to $\gamma$. We
822: will use this observation in the next section.
823:
824:
825: \subsection{Generalization to random triangulations}
826:
827: Fix an infinite Lorentzian triangulation $T$ and let $\gamma$ be a
828: finite self-avoiding, locally geodesic path in $T$, starting at
829: the root vertex (which we denote by $v_0$). Let $\Gamma\subset T$
830: be a \defined{$1$-neighborhood} of $\gamma$, i.e.~a
831: sub-triangulation consisting of the path $\gamma$ together with
832: all triangles, adjacent to $\gamma$. Such a sub-triangulation
833: enjoys the \defined{rigidity} property: if $\Gamma$ can be
834: embedded into
835: some Lorentzian triangulation, then this embedding is unique%
836: \footnote{
837: we require that the embedding maps $v_0$ to the root vertex of $T$,
838: and sends horizontal/vertical edges of $\Gamma$ to edges of the same type in $T$
839: }.
840: %
841: We define the event ``$T$ contains $\Gamma$'' (denoted
842: $\Gamma\subset T$) to be the subset of triangulations in
843: $\LT_\infty$ for which such an embedding exists.
844:
845: \begin{lemma}\label{L2B}
846: Let $\Gamma$ be a $1$-neighborhood of a locally geodesic path $\gamma$,
847: let $\gamma$ have length $n$ and let the vertices of $\gamma$ have degrees
848: $d_0, d_1, \ldots, d_n$.
849:
850: The probability that a random Lorentzian triangulation $T$
851: contains $\Gamma$ is bounded by
852: \[
853: %\P\{ \Gamma \subset T, \mbox{$\gamma$ is open} \}
854: \P_\mucr \{ \Gamma \subset T \}
855: \le C^n \prod_{i=0}^n \f{1}{d_i^9}
856: \]
857: for some absolute constant $C$.
858: \end{lemma}
859:
860: We can use this lemma to obtain estimates on percolation probability as follows.
861: Given the degrees $d_0, \ldots, d_n$ one can completely specify $\Gamma$
862: using the following collection of numbers
863: \begin{itemize}
864: \item for each vertex we need to specify
865: how it's degree $d_j$ is split into $d_j^{(up)}$ and $d_j^{(dn)}$,
866: i.e.~edges going up and down from the vertex
867: --- this gives one integer parameter in the range $[1, d_j]$;
868: \item also we need to specify the edge along which the path $\gamma$ entered the vertex
869: and the edge along which it left; this adds two more parameters.
870: \end{itemize}
871: The total number of meaningful combinations for the above
872: parameters clearly does not exceed $\prod d_j^3$. Now, if $T$ is a
873: $\P_{\mucr}$-distributed random triangulation with
874: site-percolation on it, by Lemma~\ref{L2B} the probability for $T$
875: to contain an open path from the origin with vertex degrees $d_0,
876: \ldots, d_n$ is bounded by
877: \[
878: \prod_j \f{ C }{ d_j^9 } \times \prod_j \tanh(\beta d_j) \times \prod_j d_j^3
879: \le (\beta C)^n \prod \f{1}{d_j^5}.
880: \]
881: Summing over all possible values of $d_0, \ldots, d_n$ we obtain the estimate
882: \[
883: \sum_{d_0,\ldots, d_n} (\beta C)^n \prod \f{1}{d_j^5}
884: \le \B( \beta C \sum_{d\ge 4} \f{1}{d^5} \B)^n
885: \]
886: for the probability to have an open path of length $n$ in a random
887: triangulation. For $\beta$ sufficiently small, the last expression
888: tends to $0$ as $n\to \infty$, and the probability to have an
889: infinite open path from the origin is zero.
890:
891: Note that here we are estimating the annealed probability
892: (semi-direct product $\P_\mucr\times P^{p,T}$) of the existence of
893: open path. Clearly, if this annealed probability is zero, then
894: also for $\P_\mucr$-a.e.~triangulation $T$ the
895: $P^{p,T}$-probability of an infinite open path is zero as well.
896:
897:
898: Therefore it remains to prove Lemma~\ref{L2B}.
899:
900: \subsection{Elementary perturbations}
901: Consider a finite Lorentzian triangulation $T \in \LT_N$
902: containing an internal vertex $v$ of very high degree, say~$d_v
903: \ge 100$. Let $d_v^{(up)}$ and $d_v^{(dn)}$ be the degree
904: contributed to $d_v$ by edges going upwards and downwards from
905: $v$, so that $d_v = d_v^{(up)} + d_v^{(dn)} + 2$. Let $O(v)$ be a
906: $1$-neighborhood of $v$, i.e.~all the triangles adjacent to $v$,
907: and let $T'$ be the complement to $O(v)$ in $T$.
908:
909: Let us estimate $\P_{N,\mu}(T | T')$, i.e.~the probability to see
910: $T$ as a realization of a random Lorentzian triangulation,
911: conditionally on the event that this triangulation contains $T'$.
912: For this purpose consider a modification of the $1$-neighborhood
913: of $v$, which consists in adding a pair of triangles as shown on
914: \figref{ins1}.
915: %
916: \putfigure{ins1}{Inserting a pair of triangles at a high-degree
917: vertex $v$. The position to insert new triangles is determined by
918: the choice of two neighboring vertices on the layers below and
919: above $v$.}
920: %
921: There are $d_v^{(up)} \times d_v^{(dn)}$ to distinct ways to make such a modification,
922: and every time the Hamiltonian of the Gibbs measure~\eqref{Gm}
923: is increased by $2\mu$ (since there are $2$ triangles added)%.
924:
925: Therefore we can estimate
926: \[ \P_{N,\mu} (T \,|\, T' ) \le \f{e^{2\mu}}{ d_v^{(up)} d_v^{(dn)}}
927: \le \f{ e^{2\mu}}{ (d_v-2)/2}. \]
928:
929: A better estimate can be obtained by applying $k$
930: elementary perturbations at $v$ simultaneously.
931: Such a set of perturbations is specified by a sequence of pairs
932: $\{(u^{(up)}_i, u^{(dn)}_i)\}_{i=1}^k$ of vertices
933: lying on the layers above and below $v$ (\figref{ins2}).
934:
935: \putfigure{ins2}{Multiple insertions at the same vertex.
936: Note that one of the edges is used twice for insertions,
937: and the corresponding vertex has two dots.}
938:
939: In order to assure that all the perturbations can be applied unambiguously,
940: we require that the vertices $u_1^{(up)}, \ldots u_k^{(up)}$ to be
941: ordered from left to right, as well as $u_1^{(dn)}, \ldots u_k^{(dn)}$.
942: We may however allow some vertices to coincide -- otherwise it would
943: be impossible to apply $n$ perturbations at once in a vertex whose
944: up-degree or down-degree are not large enough.
945:
946: If $d_v$ is large, one of $d_v^{(up)}$ or $d_v^{(dn)}$ is at least
947: $(d_v -2)/2$, and therefore the number of distinct ways to choose,
948: say, $k=10$ modifications is at least
949: \[ { (d_v -2)/2 \choose 10 } = C_{10} d_v^{10} + O(d_v ^9). \]
950:
951: Now consider a path $\gamma \subset \Gamma$ with vertices of degrees $d_0, \ldots, d_n$.
952: Let $W_\Gamma$ be the set of possible modifications we can make to $\Gamma$,
953: by inserting exactly $10$ pairs of triangles at vertices of high degree
954: (say, $d_i \ge 100$), and leaving all other vertices intact.
955: Then for some constant $\widetilde C_{10}$
956: \[ \# W_\Gamma \ge \prod \widetilde C_{10} d_j^{10}. \]
957:
958: Let $\LT_{\Gamma,N} \subset \LT_N$ be the set of Lorentzian triangulations that contain~$\Gamma$.
959: Clearly, each modification $w\in W_\Gamma$ can be applied to every $T \in \LT_{\Gamma,N}$,
960: producing a modified triangulation which we denote by $w(T)$.
961: Consider the application of modifications $w$ to a triangulation $T$
962: as a mapping
963: \begin{equation}\label{eq.app}
964: \begin{array}{rccl}
965: {\rm \bf app}: & \LT_{\Gamma,N} \times W_\Gamma & \rightarrow & \LT_N \\
966: & (T, w) & \rightarrow & w(T).
967: \end{array}
968: \end{equation}
969: %\[ {\rm app} : \LT_\Gamma \times W_\Gamma \to \LT \]
970: %where $\LT$ is the set of all Lorentzian triangulations.
971:
972: Since any modification to $T \in \LT_{\Gamma,N}$
973: %affects at most $n+1$ lowest layers, and
974: adds at most $20 n$ new triangles,
975: we have, for all $(T,w) \in \LT_\Gamma \times W_\Gamma$
976: \[ \P_{N,\mu}( w(T) ) \ge e^{-20n \mu} \P_{N,\mu}(T). \]
977: Taking a sum over all $(T,w)$ we get
978: \begin{eqnarray}\label{eq.twsum}
979: \sum_{(T,w)\in \LT_{\Gamma,N} \times W_\Gamma} \P_{N,\mu}(w(T))
980: &\ge& e^{-20n \mu} \# W_\Gamma \sum_{T\in \LT_{\Gamma,N}}
981: \P_{N,\mu}(T) \\ && {} = e^{-20n \mu} \# W_\Gamma
982: \P_{N,\mu}(\Gamma). \nonumber
983: \end{eqnarray}
984:
985: \subsection{Overcounting via random reconstruction}
986:
987: In order to get an estimate on $\P_{N,\mu}(\Gamma)$ from~\eqref{twsum},
988: it would be helpful if the application~\eqref{app} was an injection.
989: But this is not easy to prove (if at all true!),
990: so instead we will estimate the \defined{overcounting}
991: \[
992: {\rm over}_\Gamma(T')
993: := \#\{ (T,w) \in \LT_{\Gamma,N} \times W_\Gamma \,|\, w(T) = T' \}.
994: \]
995:
996: \begin{lemma}\label{L2C}
997: Given a path $\gamma$ with $1$-neighborhood $\Gamma$,
998: with vertex degrees $d_0, \ldots, d_n$, we have for any $T'\in \LT_N$
999: \begin{equation}\label{eq.overt}
1000: {\rm over}_\Gamma(T') \le \prod 12 (d_j + 10).
1001: \end{equation}
1002: \end{lemma}
1003:
1004: \proof
1005: First note that each elementary perturbation (i.e.~insertion of two adjacent
1006: triangles) can be undone by collapsing the newly inserted horizontal edge.
1007: Therefore in order to reconstruct the pair $(T,w)$, knowing $T'=w(T)$,
1008: it will be sufficient to identify the edges of $T'$ that were added by $w$.
1009:
1010: It's not clear whether we can perform such a reconstruction deterministically,
1011: but we certainly can achieve it {\em with some probability} with the help
1012: of the following randomized algorithm:
1013: \begin{itemize}
1014: \item{} the algorithm starts from the root vertex
1015: and at each step moves to an adjacent
1016: vertex, chosen uniformly at random;
1017: \item{} it also modifies the triangulation along his way as following
1018: \item{} each time arriving to a new vertex
1019: it should decide whether to contract a sequence of $10$ horizontal edges
1020: (at least one of these edges must be adjacent to the current location),
1021: or do nothing.
1022: %since the edges added by modifications $w\in W_\Gamma$ always come in packs of $10$,
1023: This gives $12 = 11 + 1$ possibilities,
1024: among which one is chosen uniformly at random.
1025: \item{} after $n$ steps, the path traversed by the algorithm is declared to
1026: be the (conjectured) path $\gamma$, and if it really is,
1027: and if the resulting triangulation coincides with the original triangulation $t$,
1028: the reconstruction is considered successful.
1029: \end{itemize}
1030: Most of the time this algorithm will fail,
1031: but with some probability it may actually reconstruct both $T$ and $w$.
1032: Let us now estimate this probability.
1033:
1034: First note that the degree of $v_0$ in $T'$ can be larger than the degree of $v_0$
1035: in $T$ because of the modifications made to the second vertex of $\gamma$,
1036: but in any case it will not exceed $d_0 + 10$, since each elementary
1037: modification in $v_1$ increases the degree of $v_0$ at most by $1$.
1038: Therefore, the algorithm will guess the correct direction from $v_0$
1039: with probability at least $1/(d_0 + 10)$.
1040:
1041: Assuming the direction it took on the first step was correct,
1042: there is exactly $1$ right choice out of $12$ on the next step
1043: (collapse some $10$ nearby horizontal edges,
1044: which can be chosen in $11$ different ways, or do nothing),
1045: so the probability not to fail is $1/12$.
1046:
1047: Assuming again that
1048: the modifications to the second vertex $v_1$ were undone correctly,
1049: there is at least one edge that will lead
1050: to the next vertex $v_2$.
1051: Since the path $\gamma$ is locally geodesic,
1052: at most two neighbors of $v_1$ belong to $\gamma$,
1053: namely these are $v_0$ and $v_2$.
1054: Modifications at $v_2$ add at most $10$ to the degree of $v_1$,
1055: so with probability at least $1/(d_1 + 10)$
1056: the algorithm will take the correct direction again.
1057:
1058: Proceeding this way, we see that the probability to win is at least
1059: \begin{equation}\label{eq.rec}
1060: \prod_{j=0}^n \f{1}{12 (d_j+10)}.
1061: \end{equation}
1062:
1063: Now for every $T'$ and for every $(T,w) \in \LT_{\Gamma,N} \times W_\Gamma$
1064: such that $w(T)=T'$ the probability to correctly reconstruct
1065: $(T,w)$ from $T'$ is at least \eqref{rec},
1066: therefore
1067: \[ {\rm over}_\Gamma(T') \le \prod 12 (d_j+10). \]
1068: This finishes the proof of Lemma~\ref{L2C}
1069: \qed
1070:
1071: Finally, combining the inequalities \eqref{twsum} and \eqref{overt} we get
1072: \begin{eqnarray*}
1073: e^{-20n \mu} \# W_\Gamma \P_{N,\mu}(\Gamma) &\le& \sum_{(T,w)\in
1074: \LT_{\Gamma,N} \times W_\Gamma} \P_{N,\mu}(w(T))
1075: \\
1076: &\le&
1077: \sum_{T' \in \LT_N} {\rm over}_\Gamma(T') \P_{N,\mu}(T')
1078: \\
1079: &\le&
1080: \prod_{j=0}^n 12 (d_j+10) {\sum_{T' \in \LT} \P_{N,\mu}(T') }
1081: \\
1082: &\le&
1083: \prod_{j=0}^n 12 (d_j+10),
1084: \end{eqnarray*}
1085: and
1086: \[
1087: \P_{N,\mu}(\Gamma)
1088: \le
1089: \f{ e^{20n \mu} \prod\limits_{j=0}^n 12 (d_j+10) }
1090: { \prod\limits_{j=0}^n \widetilde C_{10} d_j^{10} }
1091: \le
1092: \prod_{j=0}^n \f{e^{20\mu}{\widetilde C} }{ d_j^9 }.
1093: \]
1094: Since the last estimate is uniform in $N$,
1095: it holds also for the limiting measure~$\P_\mu$,
1096: and setting $\mu=\mucr$ we obtain the statement of Lemma~\ref{L2B}.
1097: This also finishes the proof of Theorem~\ref{T2}.
1098:
1099: %This finishes the proof of Lemma~\ref{L2B}, and thus of Theorem~\ref{T2}.
1100:
1101: % end part 2
1102:
1103: \section{Critical temperature is constant a.s.}
1104: Consider the critical temperature $\beta_c(G)$ of the Ising model on a graph $G$
1105: as a function of $G$. In the above two sections we have shown that
1106: when $T$ is a $P_{\mucr}$-random Lorentzian triangulation,
1107: we have $\beta_c(T) \in [\beta_h, \beta_0]$ a.s.
1108: In this section we show that in fact
1109: \begin{theorem}\label{T3}
1110: $\beta_c(T)$ is constant $\P_\mucr$-a.s.
1111: \end{theorem}
1112: %
1113: The proof relies on two lemmas.
1114: \begin{lemma}\label{Lsimplex}
1115: Let $G$, $G'$ be two locally finite infinite graphs
1116: that differ only by a finite subgraph,
1117: i.e.~there exist finite subgraphs $H\subset G$, $H'\subset G'$
1118: such that $G\bs H$ is isomorphic to $G'\bs H'$.
1119: Then $\beta_c(G) = \beta_c(G')$.
1120: \end{lemma}
1121: \proof
1122: This statement is widely known and follows from Theorem~7.33~in~\cite{georgii}.
1123:
1124: More exactly,
1125: consider the Ising model on $G$ at some fixed inverse temperature $\beta\in(0,\infty)$.
1126: Let $Y$ be the corresponding specification (collection of conditional distributions)
1127: and denote by $\calG(Y)$ the set of Gibbs measures, satisfying this specification.
1128: It is an easy fact that $\calG(Y)$ is a convex set in the linear space of
1129: finite measures on the set of configurations.
1130:
1131: Any specification $Y$ on $G$ can be naturally restricted
1132: to a specification $\wtY$ on~$G\bs H$.
1133: Namely,
1134: %by integrating over all possible spin configurations in $H$.
1135: for each finite $A \subset G\bs H$ and each $\eta \in \{-1,1\}^{G\bs H}$ put
1136: %%%\[ \wtY_A(\cdot, \eta) =
1137: %%% \P\left\{
1138: %%% \sigma_i =
1139: %%% \begin{cases}
1140: %%% \eta_i & i\in (G\bs H) \bs A\cr
1141: %%% \cdot_i& i\in A\cr
1142: %%% \text{anything} & i\in H
1143: %%% \end{cases}
1144: %%% \right\}
1145: %%%\]
1146: \begin{eqnarray}
1147: \wtY_A(\cdot, \eta) &=&
1148: \P\{ \sigma_i = (\cdot)_i\,\forall i \in A;\,\,
1149: \sigma_i = \eta_i\,\forall i\in (G\bs H)\bs A
1150: \}.
1151: \nonumber\\
1152: &=& \sum_{\wt\eta \in \{-1,+1\}^H} Y_{A\union H}( (\cdot)\union \wt\eta \,|\, \eta )
1153: \label{eq.wyah}
1154: \end{eqnarray}
1155: (the last sum runs over all spin configurations in $H$).
1156: Informally, the restricted specification $\wtY$ can be interpreted as
1157: an Ising model on $G$, but with the spins in $H$ being hidden from the
1158: observer.
1159:
1160: %The last probability here can be expressed in terms of the specification $Y$
1161: %as a finite weighted sum of conditional probabilities.
1162: %
1163: Let now $Y'$ be the specification for the Ising model on $G'$,
1164: and let $\wtY'$ be an analogous restriction of $Y'$ to $G'\bs H'$.
1165: Since $G\bs H$ and $G'\bs H'$ are isomorphic, the specifications
1166: $\wtY$ and $\wtY'$ are defined over the same graph,
1167: and it follows from \eqref{wyah} that they are \defined{equivalent}
1168: in that there exists a constant $c>1$ such that
1169: \[ \f1c < \f{ \wtY_A(\cdot \,|\, \eta) }{ \wtY'_A (\cdot \,|\, \eta) } < c \]
1170: for all possible values of $A$, $\eta$ and $(\cdot)$.
1171: Theorem 7.33 in \cite{GM} then implies that
1172: $\calG(\wtY)$ is affinely isomorphic to $\calG(\wtY')$.
1173: In particular $\calG(\wtY)$ consists of a single element
1174: -- a unique Gibbs measure -- if and only if $\calG(\wtY')$ does.
1175:
1176: Since any Gibbs measure on spin configurations in $G\bs H$,
1177: satisfying $\wtY$, can be uniquely extended to a Gibbs measure on $G$, satisfying $Y$,
1178: and the same holds for $G'$, $H'$, $Y'$, the lemma follows.
1179: \eop
1180:
1181: \begin{lemma}\label{Lspinal}
1182: A Galton-Watson tree for a critical branching process,
1183: conditioned to non-extinction at infinity,
1184: can be described as following
1185: \begin{itemize}
1186: \item{} it contains a single infinite path (a \defined{spine}), $\{v_0, v_1, \ldots\}$,
1187: starting at the root vertex;
1188: \item{} at each vertex $v_i$ of the spine a pair of finite trees $(L_i, R_i)$
1189: is attached, one of each side of the spine;
1190: \item{} and the pairs $(L_i, R_i)$ are independent and identically distributed.
1191: \end{itemize}
1192: \end{lemma}
1193:
1194: \begin{rem} In fact, the law of $(L_i, R_i)$ can be explicitly expressed
1195: in terms of the original branching process, but for our purposes
1196: it's enough to know that the trees are i.i.d.
1197: \end{rem}
1198:
1199: \proof See e.g.~\cite{peres} or \cite{geiger}.
1200: \eop
1201:
1202: \proofof{Theorem~\ref{T3}}
1203: Consider now the tree parametrization $\tau$ of a $\P_\mucr$-distributed
1204: infinite Lorentzian triangulation $T$.
1205: By Lemma~\ref{Lspinal}, $\tau$ is encoded by
1206: a sequence of pairs of finite trees $(L_i,R_i)_{i\in\Z}$.
1207: Clearly, this is an exchangeable sequence.
1208: Let $\sigma_{i j}$ be a permutation of the $i$-th and $j$-th pairs;
1209: then $\sigma_{i j}\tau$ is a tree-parametrization of some other
1210: infinite Lorentzian triangulation $\sigma_{i j}T$,
1211: which only differs from $T$ by a finite subgraph.
1212: By Lemma~\ref{Lsimplex} $\beta_c(T) = \beta_c(\sigma_{i j} T)$,
1213: and Hewitt-Savage zero-one law applies to $\beta_c$, considered
1214: as a function of the sequence $(L_i,R_i)_{i\in\Z}$.
1215: Thus $\beta_c(T)$ is constant a.s.
1216: \eop
1217:
1218:
1219:
1220: \bibliographystyle{amsplain}
1221: \bibliography{lt}
1222: \end{document}
1223: