Gibbs Measures and Phase Transitions

1 Specifications of random fields

1.1 Preliminaries

Definition 1.1 Juxtaposition
#

Let \(E\) and \(S\) be sets. Let \(\Delta \in \mathcal{P}(S)\), and let \(\omega \in E^S\). We define

\begin{align} \operatorname{Juxt}_{\omega } : E^\Delta & \to E^S\\ \zeta & \mapsto \delta \mapsto \begin{cases} \zeta _\delta & \delta \in \Delta \\ \omega _\delta & \delta \notin \Delta \end{cases}\end{align}

to be the juxtaposition of \(\zeta \) and \(\omega \) (for each \(\zeta \in E^\Delta \)).

Definition 1.2 Cylinder events
#

Let \((E,\mathcal{E})\) be a measurable space, and let \(S\) be a set. Then,

\begin{align} \mathcal{F}:\mathcal{P}(S)& \to \left\{ \text{sigma algebras on } E^S \right\} \\ \Delta & \mapsto \sigma (\{ \text{proj}_\delta :E^S\to E\mid \delta \in \Delta \} ) \end{align}

defines the cylinder events in \(\Delta \) (for each \(\Delta \in \mathcal{P}(S)\)), where each \(\text{proj}_\delta \) is the coordinate projection at coordinate \(\delta \).

Definition 1.3 Kernel
#

Let \((X,\mathcal X)\) and \((Y,\mathcal{Y})\) be measurable spaces. Then,

\[ \text{Ker}_{\mathcal{Y},\mathcal X}:=\left\{ \pi :\mathcal X\times Y\to [0,\infty ]~ \middle \vert ~ \forall y\in Y,\pi (\cdot \mid y)\in \mathfrak {M}(X,\mathcal X);~ \forall A\in \mathcal X,\pi (A\mid \cdot )\text{ is }\mathcal{Y}\text{-measurable}\right\} \]

defines the set of kernels from \(\mathcal{Y}\) to \(\mathcal X\), where \(\mathfrak {M}(X,\mathcal X)\) is the space of measures on \(X\).

Definition 1.4 Markov kernel
#

Let \((X,\mathcal X)\) and \((Y,\mathcal{Y})\) be measurable spaces. We say that \(\pi \in \text{Ker}_{\mathcal{Y},\mathcal X}\) is a Markov kernel iff \(\pi (X\mid \cdot )=1\).

Let \((X, \mathcal X)\) be a measurable space, let \(\mathcal B\) be a sub \(\sigma \)-algebra of \(\mathcal X\). Let \(\pi \in \text{Ker}_{\mathcal B, \mathcal X}\).

Definition 1.5 Proper kernel
#

\(\pi \) is proper iff \(\pi (A\cap B\mid x)=\pi (A\mid x)\cdot \mathbf{1}_B(x)\) for all \(A\in \mathcal X\), \(B\in \mathcal B\) and \(x\in X\).

Lemma 1.6 Lebesgue integral characterisation of proper kernels

If \(\pi \) is proper, then

\[ \int f(x) g(x)\ \pi (dx\mid x_0) = g(x_0)\int f(x)\ \pi (dx\mid x_0) \]

for all \(x_0 \in X\) and functions \(f, g : X \to [0, \infty ]\) such that \(f\) is \(\mathcal X\)-measurable, \(g\) is \(\mathcal B\)-measurable.

Proof

Standard machine.

Lemma 1.7 Integral characterisation of proper kernels

If \(\pi \) is a proper Markov kernel, then

\[ \int f(x) g(x)\ \pi (dx\mid x_0) = g(x_0)\int f(x)\ \pi (dx\mid x_0) \]

for all \(x_0 \in X\) and functions \(f, g : X \to \mathbb R\) such that \(f\) is bounded \(\mathcal X\)-measurable and \(g\) is bounded \(\mathcal B\)-measurable.

Proof

Standard machine.

Definition 1.8 Conditional expectation kernel

Let \(\mu \in \mathfrak {M}(X,\mathcal X)\). Then, \(\pi \in \text{Ker}_{\mathcal B,\mathcal X}\) is a conditional expectation kernel for \(\mu \) if \(\mu (A\mid \mathcal B)=\pi (A\mid \cdot )\) \(\mu \)-a.e.

Lemma 1.9 Lebesgue integral characterisation of proper conditional expectation kernels

If \(\pi \in \text{Ker}_{\mathcal B, \mathcal X}\) is a conditional expectation kernel for \(\mu \), then

\[ \mu [f\mid \mathcal B] = \int f(x)\ \pi (\partial x\mid \cdot )\ \mu \text{-a.e.} \]

for all \(\mathcal X\)-measurable functions \(f : X \to [0, \infty ]\).

Proof

Standard machine.

Lemma 1.10 Integral characterisation of proper conditional expectation kernels

If \(\pi \in \text{Ker}_{\mathcal B, \mathcal X}\) is a conditional expectation kernel for \(\mu \), then

\[ \mu (f\mid \mathcal B) = \int f(x)\ \pi (\partial x\mid \cdot )\ \mu \text{-a.e.} \]

for all bounded \(\mathcal X\)-measurable functions \(f : X \to \mathbb R\).

Proof

Standard machine.

Lemma 1.11 Characterisation of proper conditional expectation kernels, Remark 1.20

Let \(\mu \in \mathfrak {M}(X,\mathcal X)\) be a finite measure and let \(\pi \in \text{Ker}_{\mathcal B,\mathcal X}\) be a proper kernel. Then,

\[ \pi \text{ is a conditional expectation kernel for }\mu \iff \mu \pi = \mu \]
Proof

By the characterisation of conditional expectation,

\[ \pi \text{ is a conditional expectation kernel for }\mu \iff \forall A \in \mathcal X, \forall B \in \mathcal B, \mu (A \cap B) = \int _B \pi (A|\cdot )\ \partial \mu \]

By properness of \(\pi \),

\[ \int _B \pi (A|\cdot )\ \partial \mu = \mu \pi (A \cap B) \]

Hence

\begin{align} \pi \text{ is a cond. exp. kernel with respect to }\mu & \iff \forall A \in \mathcal X, \forall B \in \mathcal B, \mu (A \cap B) = \mu \pi (A \cap B) \\ & \iff \forall A \in \mathcal X, \mu (A) = \mu \pi (A) \\ & \iff \mu = \mu \pi \end{align}

1.2 Prescribing conditional probabilities

Definition 1.12 Specification
#

A specification is a family of kernels \(\gamma : \operatorname{Finset}S \to \operatorname{Ker}_{\mathcal{F}_{S\setminus \Lambda }, \mathcal{E}^S}\) which is consistent, in the sense that

\[ \forall \Lambda _1, \Lambda _2 \in \operatorname{Finset}(S), \Lambda _1 \subseteq \Lambda _2 \implies \gamma _{\Lambda _1} \circ _k \gamma _{\Lambda _2} = \gamma _{\Lambda _2} \]

All specifications will be with parameter set \(S\) and state space \((E, \mathcal{E})\) in this chapter.

Definition 1.13 Independent specification
#

A specification \(\gamma \) is independent iff

\[ \forall \Lambda _1, \Lambda _2 \in \operatorname{Finset}(S), \gamma _{\Lambda _1} \circ _k \gamma _{\Lambda _2} = \gamma _{\Lambda _1\cup \Lambda _2} \]
Definition 1.14 Markov specification
#

A specification \(\gamma \) is a Markov specification iff \(\gamma _\Lambda \) is a probability kernel for every \(\Lambda \in \operatorname{Finset}(S)\).

Definition 1.15 Proper specification
#

A specification \(\gamma \) is proper iff the kernel \(\gamma _\Lambda \) is proper for every \(\Lambda \in \operatorname{Finset}(S)\).

Definition 1.16 Gibbs measures

Given a specification \(\gamma \), a Gibbs measures specified by \(\gamma \) is a measure \(\nu \in \mathfrak {M}(E^S, \mathcal{E}^S)\) such that \(\gamma _\Lambda (A|\cdot )\) is a conditional expectation kernel for \(\nu \) for all \(A \in \mathcal E^S\) and \(\Lambda \in \operatorname{Finset}(S)\).

Lemma 1.17 Characterisation of Gibbs measures, Remark 1.24

Let \(\gamma \) be a proper specification with parameter set \(S\) and state space \((E, \mathcal{E})\), and let \(\nu \in \mathfrak {P}(E^S, \mathcal{E}^S)\). TFAE:

  1. \(\nu \in \mathcal{G}(\gamma )\).

  2. \(\gamma _\Lambda \circ _m\nu = \nu \) for all \(\Lambda \in \operatorname{Finset}(S)\).

  3. \(\gamma _\Lambda \circ _m\nu = \nu \) frequently as \(\Lambda \to S\).

Proof

1 is equivalent to 2 by Lemma 1.9. 2 trivially implies 3. Now, 3 implies 2 because for each \(\Lambda \) there exists some \(\Lambda ' \supseteq \Lambda \) such that \(\gamma _{\Lambda '}\circ _k\nu = \nu \). Then

\[ \nu \gamma _\Lambda = \nu \gamma _{\Lambda '}\gamma _\Lambda = \nu \gamma _{\Lambda '} = \nu \]

1.3 \(\lambda \)-specifications

Let \(S\) be a set, \((E, \mathcal{E})\) be a measurable space and \(\nu \) a measure on \(E\).

Definition 1.18 Product probability measure
#

Let \(I\) be a set. Suppose for each \(i\in I\) that \((\Omega _i,\mathcal B_i,P_i)\) is a probability space. Then, \(P:=\bigotimes _{i\in I}P_i\) is a well-defined product probability measure on \(\prod _{i\in I}\Omega _i\).

Definition 1.19 Independent Specification with Single Spin Distribution (ISSSD)
#

The Independent Specification with Single Spin Distribution \(\nu \) is

\begin{align} \operatorname{ISSSD}:\mathfrak {P}(E, \mathcal{E})& \to \operatorname{Finset}(S)\to \mathcal{E}^S\times E^S\to \overline{\mathbb {R}_{\geq 0}}\\ \nu & \mapsto \Lambda \mapsto (A\mid \omega )\mapsto \left(\nu ^\Lambda \left(\operatorname{Juxt}_{\omega }^{-1}(A)\right)\right) \end{align}

defines the Independent Specification with Single Spin Distribution with \(\nu \) (for each \(\nu \in \mathfrak {P}(E, \mathcal{E})\)), where \(\nu ^\Lambda \) is the usual product measure.

Lemma 1.20 Independence of ISSSDs

\(\operatorname{ISSSD}(\nu )\) is independent.

Proof

Immediate.

Definition 1.21 ISSSDs are specifications
#

\(\operatorname{ISSSD}(\nu )\) is a specification.

Lemma 1.22 ISSSDs are proper specifications

\(\operatorname{ISSSD}(\nu )\) is a proper specification.

Proof

We already know it’s a specification. Properness is immediate.

Lemma 1.23 Uniqueness of a Gibbs measure specified by an ISSSD

There is at most one Gibbs measure specified by \(\operatorname{ISSSD}(\nu )\).

Proof

See book.

Lemma 1.24 Existence of a Gibbs measure specified by an ISSSD

The product measure \(\nu ^S\) is a Gibbs measure specified by \(\operatorname{ISSSD}(\nu )\).

Proof

Immediate.

Definition 1.25 Modifier

A modifier of \(\gamma \) is a family

\[ \rho : \operatorname{Finset}(S) \to \Omega \to [0, \infty [ \]

such that the corresponding family of kernels \(\rho \gamma \) is a specification.

Lemma 1.26 Modifier of a modifier

Modifying a specification \(\gamma \) by \(\rho _1\) then \(\rho _2\) is the same as modifying it by their product.

Proof

TODO

Lemma 1.27 A modifier of a proper specification is proper

If \(\gamma \) is a specification and \(\rho \) a modifier of \(\gamma \), then \(\rho \gamma \) is a proper specification.

Proof

For all \(\Lambda \in \operatorname{Finset}(S)\), \(A \in \mathcal E^S\), \(B \in \mathcal{F}_{S\setminus \Lambda }\), \(\eta : S \to E\), we want to prove

\[ (\rho \gamma )_\Lambda (A \cap B | \eta ) = 1_B(\eta ) (\rho \gamma )_\Lambda (A B | \eta ) \]

Expanding out, this is equivalent to

\[ \int _{\zeta \in A \cap B} \rho _\Lambda (\zeta )\ d(\gamma _\Lambda (\eta )) = 1_B(\eta ) \int _{\zeta \in A} \rho _\Lambda (\zeta )\ d(\gamma _\Lambda (\eta )) \]

which is true by Lemma 1.6 with \(f = 1_A\rho _\Lambda \), \(g = 1_B\).

Lemma 1.28 Every specification is a modification of some ISSSD, Remark 1.28.5

If \(E\) is countable, \(\nu \) is the counting measure and \(\gamma \) is any specification, then

\[ \rho _\Lambda (\eta ) = \gamma _\Lambda (\{ \sigma _\Lambda = \eta _\Lambda \} | \eta ) \]

is a modifier from \(\operatorname{ISSSD}(\nu )\) to \(\gamma \).

Proof

For all \(\Lambda \in \operatorname{Finset}(S)\), \(A\) measurable, \(\eta : S \to E\), we have

\begin{align} (\rho \operatorname{ISSSD}(\nu ))_\Lambda (A|\eta ) & = \int _\zeta \rho _\Lambda (\zeta ) \operatorname{ISSSD}(\nu )(d\zeta |\eta ) \\ & = \int _\zeta \gamma _\Lambda (\{ \sigma _\Lambda = \eta _\Lambda \} | \eta ) \operatorname{ISSSD}(\nu )(d\zeta |\eta ) \\ & = \gamma _\Lambda (A|\eta ) \end{align}
Proposition 1.29 Characterisation of modifiers, Proposition 1.30.1
#

If \(\rho \) is a family of measurable densities and \(\gamma \) is a proper specification, then TFAE

  1. \(\rho \) is a modifier of \(\gamma \)

  2. For all \(\Lambda _1, \Lambda _2\) with \(\Lambda _1 \subseteq \Lambda _2\) and all \(\eta : S \to E\), we have

    \[ \rho _{\Lambda _2} = \rho _{\Lambda _1}\cdot (\gamma _{\Lambda _1} \rho _{\Lambda _2}) \quad \gamma _{\Lambda _2}(\cdot |\eta )\text{-a.e.} \]
Proof
  • (\(\implies \)) \(\rho _{\Lambda _2} = \rho _{\Lambda _1}\cdot (\gamma _{\Lambda _1} \rho _{\Lambda _2}) \quad \gamma _{\Lambda _2}(\cdot |\eta )\text{-a.e.}\)

    • \(\implies \rho _{\Lambda _2}\gamma _{\Lambda _2} = \)

Proposition 1.30 Characterisation of modifiers of independent specifications, Proposition 1.30.2

If \(\rho \) is a family of measurable densities and \(\gamma \) is a proper independent specification, then TFAE

  1. \(\rho \) is a modifier of \(\gamma \)

  2. For all \(\Lambda _1, \Lambda _2\) with \(\Lambda _1 \subseteq \Lambda _2\), \(\eta : S \to E\) and \(\gamma _{\Lambda _2 \setminus \Lambda _1}(\cdot |\alpha )\)-almost all \(\eta _2 : S \to E\), we have

    \[ \rho _{\Lambda _2}(\zeta _1)\rho _{\Lambda _1}(\zeta _2) = \rho _{\Lambda _2}(\zeta _2) \rho _{\Lambda _1}(\zeta _1) \]

    for \(\gamma _{\Lambda _1}(\cdot |\eta _2) \times \gamma _{\Lambda _2}(\cdot |\eta _2)\)-almost all \((\zeta _1, \zeta _2)\).

Proof
Definition 1.31 Premodifier, Definition 1.31
#

A family of measurable functions \(h_\Lambda : (S \to E) \to [0, \infty [\) is a premodifier if

\[ h_{\Lambda _2}(\zeta )h_{\Lambda _1}(\eta ) = h_{\Lambda _1}(\zeta )h_{\Lambda _2}(\eta ) \]

for all \(\Lambda _1 \subseteq \Lambda _2\) and all \(\zeta , \eta : S \to E\) such that \(\zeta _{\Lambda _1^c} = \eta _{\Lambda _1^c}\).

Lemma 1.32 Modifiers are premodifiers

If \(\rho \) is a modifier of \(\operatorname{ISSSD}(\nu ^S)\), then it is a premodifier if any of the following conditions hold:

  1. \(E\) is countable and \(\nu \) is equivalent to the counting measure.

  2. \(E\) is a second countable Borel space.

  3. \(\nu \) is everywhere dense.

  4. For all \(\Lambda _1 \subseteq \Lambda _2\) and all \(\eta : S \to E\), \(\zeta \mapsto \rho _{\Lambda _1}(\zeta \eta _{\Lambda _1^c})\) is continuous on \(E^{\Lambda _1}\).

Proof
  1. Use Proposition ??.

  2. Omitted.

  3. Omitted.

  4. Omitted.

Lemma 1.33 Premodifiers give rise to modifiers, Remark 1.32

If \(h\) is a premodifier and \(\nu \) is such that \(0 {\lt} \nu _\Lambda h_\Lambda {\lt} \infty \) for all \(\Lambda \), then

\[ \rho _\Lambda := \frac{h_\Lambda }{\operatorname{ISSSD}(\nu )_\Lambda h_\Lambda } \]

is a modifier of \(\operatorname{ISSSD}(\nu )\).

Proof

TODO