Critical and near-critical influence bounds
for ferromagnetic Ising models

Abstract

For a ferromagnetic Ising model on a graph of maximum degree $\Delta\geq 3$, we prove a bound of order $\sqrt{n}$ on every row of the influence matrix at the tree uniqueness threshold. The estimate is uniform in the degree, the external fields, and all pinnings. More generally, if the couplings are bounded by $\beta$ and $\varepsilon=((\Delta-1)\tanh\beta-1)_{+}$, the bound is $C(\sqrt{n}+n\varepsilon)$. The proof combines a pointwise cavity bound with a positive-series magnetization tilt and the field comparison theorem of Ding, Song and Sun. The critical estimate removes the logarithm in recent general graphical bounds for the ferromagnetic case. As a consequence, zero-field single-site Glauber dynamics mixes in polynomial time throughout the supercritical window $\varepsilon=O(\sqrt{\log n/n})$, with the polynomial degree depending on the window size.

1 Introduction and statements

At the uniqueness threshold of the regular tree, Ising correlations need not have a uniformly summable tree bound. A finite graph nevertheless constrains how large their total influence can be. The expected critical scale is $\sqrt{n}$, and this scale governs spectral independence and its applications to Glauber dynamics. Chen, Chen, Yin and Zhang [1] studied this endpoint and retained the square-root graphical bound as a conjecture after retracting a coupling argument in their third version. Recent work of Bencs and coauthors [2] gives an $O_{\Delta}(\sqrt{n}\log n)$ bound on regular graphs through non-backtracking walks, and the sharp square-root order when a fixed spectral gap is available. Here we obtain the exact square-root order for ferromagnets, together with a finite-size estimate above the threshold.

Let $G=(V,E)$ be a finite simple graph, $|V|=n\geq 1$, with maximum degree at most an integer $\Delta\geq 3$. For edge couplings $J_{e}\geq 0$ and fields $a\in\mathbb{R}^{V}$, write

$$\mu_{a}(\sigma)=\frac{1}{Z_{a}}\exp\left\{\sum_{uv\in E}J_{uv}\sigma_{u}\sigma_{v}+\sum_{v\in V}a_{v}\sigma_{v}\right\},\qquad\sigma\in\{-1,1\}^{V}.$$

We use the source-first influence convention

$$\Psi_{a}(u,v)=\mu_{a}(\sigma_{v}=1\mid\sigma_{u}=1)-\mu_{a}(\sigma_{v}=1\mid\sigma_{u}=-1),\qquad\Psi_{a}(u,u)=1.$$ (1)

All entries are nonnegative. Thus $\|\Psi_{a}\|_{\infty}$ is the maximum row sum. If $D_{a}=\operatorname{diag}(\operatorname{Var}_{\mu_{a}}\sigma_{u})$, then $\Psi_{a}=D_{a}^{-1}\operatorname{Cov}_{\mu_{a}}(\sigma)$ is similar to the symmetric matrix $D_{a}^{-1/2}\operatorname{Cov}_{\mu_{a}}(\sigma)D_{a}^{-1/2}$. Our row bounds in particular bound its largest eigenvalue. They therefore give spectral independence, with the unit diagonal included in the convention.

Theorem 1.1.

Suppose $0\leq J_{e}\leq\beta_{c}$, where $\beta_{c}=\operatorname{arctanh}(1/(\Delta-1))$. For every finite field vector and every pinning leaving $k\geq 1$ vertices, the influence matrix of the conditional measure satisfies

$$\|\Psi\|_{\infty}\leq 4\sqrt{6e}\,\sqrt{k}<17\sqrt{k}.$$

The exponent $1/2$ is sharp: the critical examples in [1, Corollary 5.3] have largest influence eigenvalue of this order. The constant here is independent of $\Delta$.

Theorem 1.2.

Suppose $0\leq\beta<\infty$, $0\leq J_{e}\leq\beta$, and put $\varepsilon=((\Delta-1)\tanh\beta-1)_{+}$. For every finite field vector and every pinning leaving $k\geq 1$ vertices,

$$\|\Psi\|_{\infty}\leq\min\{k,\,512e(\sqrt{k}+k\varepsilon)\}.$$ (2)

In particular, if $(\Delta-1)\tanh\beta\leq 1+A/\sqrt{n}$ for some fixed $A\geq 0$, the bound is $512e(1+A)\sqrt{k}$ after every pinning.

Here is the corresponding dynamical consequence. Let $T_{\rm mix}$ denote worst-state total-variation mixing time at distance $1/4$ for discrete-time, random-scan heat-bath dynamics.

Corollary 1.3.

In zero external field, under the hypotheses of Theorem 1.2, suppose $\varepsilon\leq 1/64$ and put $A=\sqrt{n}\,\varepsilon$ and $C=512e$. There is a numerical constant $K$ such that

$$T_{\rm mix}\leq Kn^{\,2+2/(\Delta-2)}\log(en)\exp\{5C(A+A^{2}/2)\}.$$

Consequently, for each fixed $B\geq 0$, the window $\varepsilon\leq B\sqrt{\log n/n}$ has polynomial mixing time for all sufficiently large $n$ depending on $B$. More explicitly, the bound is $O(n^{\,2+2/(\Delta-2)+5CB^{2}}\log(en))$, with a constant depending on $B$.

Polynomial mixing in a sparse critical window is an explicit question in the introduction of Galanis, Štefankovič and Vigoda [5]. Corollary 1.3 gives the ferromagnetic Ising case. The proof does not yield the optimal mixing exponent.

The proof uses a small uniform positive field as a probe of zero-field susceptibility. A cavity inequality from the magnetization argument of Carlson, Davies, Kolla and Perkins [3] controls the magnetization at each vertex by $Ch^{1/3}$ at criticality. Positivity of all terms in a spin expansion then gives $h\chi_{v}\leq m_{v}(h)\mathbb{E}_{0}e^{hM}$, where $M$ is total magnetization. Choosing $h$ of order $n^{-3/4}$ gives the square-root bound on each susceptibility row. Ding, Song and Sun’s comparison theorem [4] transfers it to arbitrary fields. Above criticality the same argument uses $m_{v}(h)\leq C(\sqrt{\varepsilon}+h^{1/3})$ and a smaller probe field.

2 A pointwise cavity estimate

We first work in a uniform field $h>0$. Denote the magnetization at $u$ in a graph $H$ by $m_{u}^{H}$, and suppress $h$ in the notation. The Griffiths correlation inequalities imply that spin-product expectations increase when nonnegative edges are added. These are the classical inequalities used in [3, Section 4].

Lemma 2.1.

Suppose every edge has coupling $\beta$, and let $t=\tanh\beta$ and $f_{\beta}(x)=\operatorname{arctanh}(t\tanh x)$. For each oriented edge put $L_{u\to v}=\operatorname{arctanh}(m_{u}^{G-uv})$. Then

$$L_{u\to v} \leq h+\sum_{w\in N(u)\setminus\{v\}}f_{\beta}(L_{w\to u}),$$ (3)

$$\operatorname{arctanh}(m_{u}^{G}) \leq h+\sum_{w\in N(u)}f_{\beta}(L_{w\to u}).$$ (4)

Proof.

For an edge $uv$ of $H$, let $A=m_{u}^{H-uv}$, $B=m_{v}^{H-uv}$, and $C=\mathbb{E}_{H-uv,h}[\sigma_{u}\sigma_{v}]$. Adding the edge gives the exact identity $m_{u}^{H}=(A+tB)/(1+tC)$. Since $C\geq AB$ and $A+tB\geq 0$,

$$\operatorname{arctanh}(m_{u}^{H})\leq\operatorname{arctanh}A+\operatorname{arctanh}(tB).$$

Remove the edges incident to $u$ one at a time. At the removal of $uw$, monotonicity bounds the neighbor magnetization by $m_{w}^{G-uw}$. The final isolated vertex has magnetization $\tanh h$. Starting with $G-uv$ or $G$ proves the two inequalities. ∎

Proposition 2.2.

Uniformly over the graph, its size, and $\Delta\geq 3$,

1. if $J_{e}\leq\beta_{c}$, then $m_{u}(h)\leq 4h^{1/3}$ for every $u$ and $h>0$;

2. if $J_{e}\leq\beta$ and $0\leq\varepsilon=(\Delta-1)\tanh\beta-1\leq 1/64$, then $m_{u}(h)\leq 16(\sqrt{\varepsilon}+h^{1/3})$ for every $u$ and $h>0$.

Proof.

By coupling monotonicity it suffices to use uniform coupling at its upper bound. Isolated vertices satisfy the assertions directly. Otherwise put $d=\Delta-1$ and $X=\max_{u\to v}L_{u\to v}$, a finite attained maximum. At $\beta=\beta_{c}$, Lemma 2.1 gives $F_{d}(X)\leq h$, where

$$F_{d}(x)=x-d\operatorname{arctanh}(\tanh x/d),\qquad F_{d}^{\prime}(x)=\frac{(1-d^{-2})\tanh^{2}x}{1-d^{-2}\tanh^{2}x}.$$

The function is strictly increasing on $(0,\infty)$, and $F_{d}(x)\geq x^{3}/16$ for $0\leq x\leq 1$, because $\tanh x\geq x/2$ there and $d\geq 2$. For $h\leq 1/16$, the solution $x_{h}$ of $F_{d}(x_{h})=h$ satisfies $X\leq x_{h}\leq(16h)^{1/3}$. Since $f_{\beta_{c}}(x)\leq x/d$,

$$m_{u}(h)\leq h+\Delta f_{\beta_{c}}(x_{h})=x_{h}+f_{\beta_{c}}(x_{h})\leq\tfrac{3}{2}(16h)^{1/3}<4h^{1/3}.$$

For $h>1/16$ the bound follows from $m_{u}\leq 1$.

For the second assertion put $t=(1+\varepsilon)/d<5/8$ and $F_{\beta}(x)=x-d\operatorname{arctanh}(t\tanh x)$. Again $F_{\beta}(X)\leq h$. Comparison in the parameter $t$ gives

$$0\leq F_{d}(x)-F_{\beta}(x)\leq\frac{\varepsilon\tanh x}{1-t^{2}}\leq 2\varepsilon x.$$ (5)

Also

$$F_{\beta}^{\prime}(x)=\frac{-\varepsilon+(1+\varepsilon-t^{2})\tanh^{2}x}{1-t^{2}\tanh^{2}x}>0\qquad(x\geq 1),$$

and $F_{\beta}(1)\geq 1/16-2\varepsilon\geq 1/32$. Thus $h\leq 1/64$ forces $X<1$, even though $F_{\beta}$ need not be increasing near zero. Using (5), $X^{3}/16\leq h+2\varepsilon X$. If $X\geq 8\sqrt{\varepsilon}$, absorption gives $X\leq(32h)^{1/3}$; otherwise $X<8\sqrt{\varepsilon}$. In either case $X\leq 8\sqrt{\varepsilon}+4h^{1/3}$. Since $f_{\beta}(x)\leq tx$ and $\Delta t<2$, (4) yields $m_{u}(h)\leq 16\sqrt{\varepsilon}+9h^{1/3}$. For larger $h$, the stated bound follows from $m_{u}\leq 1$. ∎

3 Positive-series tilting

At zero external field put $M=\sum_{u\in V}\sigma_{u}$, $Z(h)=\mathbb{E}_{0}e^{hM}$, and $\chi_{v}=\sum_{u}\mathbb{E}_{0}[\sigma_{v}\sigma_{u}]$. Global spin-flip symmetry gives

$$m_{v}(h)Z(h)=\mathbb{E}_{0}[\sigma_{v}\sinh(hM)]\geq h\chi_{v}.$$ (6)

To verify the inequality, expand the hyperbolic sine. Every term $\mathbb{E}_{0}[\sigma_{v}M^{2j+1}]$ is nonnegative by the first Griffiths inequality after expanding the power and canceling squares. The series is absolutely convergent since $|M|\leq n$.

At criticality Proposition 2.2 gives

$$\log Z(h)=\int_{0}^{h}\sum_{u}m_{u}(s)\,ds\leq 3nh^{4/3}.$$

Combining this with (6), then choosing $h=(6n)^{-3/4}$, proves

$$\max_{v}\chi_{v}\leq 4h^{-2/3}e^{3nh^{4/3}}=4\sqrt{6e}\sqrt{n}.$$ (7)

This is a pointwise row bound; an estimate on average magnetization alone would not supply the same conclusion.

If $\beta\leq\beta_{c}$, (7) already gives (2) with $\varepsilon=0$. For $0\leq\varepsilon=(\Delta-1)\tanh\beta-1\leq 1/64$, the second response estimate gives

$$\log Z(h) \leq 16n\sqrt{\varepsilon}\,h+12nh^{4/3},$$ (8)

$$\chi_{v} \leq 16(\sqrt{\varepsilon}/h+h^{-2/3})e^{16n\sqrt{\varepsilon}h+12nh^{4/3}}.$$ (9)

Let $a=n^{1/4}\sqrt{\varepsilon}$, $b=1+a$, and $h=n^{-3/4}/(16b)$. The exponent in (9) is at most $a/b+b^{-1}=1$. Its prefactor divided by $16\sqrt{n}$ is $16a(1+a)+16^{2/3}(1+a)^{2/3}$. Using $2a\leq 1+a^{2}$ and $(1+a)^{2}\leq 2(1+a^{2})$ bounds this by $32(1+a^{2})$. Hence

$$\max_{v}\chi_{v}\leq 512e(\sqrt{n}+n\varepsilon).$$ (10)

For $\varepsilon>1/64$, the trivial row bound $n$ is already smaller than $512e\,n\varepsilon$.

4 Fields and pinnings

We record the exact use of the field comparison, since normalized influence requires more than a covariance comparison. Theorem 1.1 of Ding, Song and Sun [4] states that, for a ferromagnetic Ising model and fields $g$ and $b\geq 0$,

$$m_{v}(g+b)-m_{v}(g-b)\leq m_{v}(b)-m_{v}(-b).$$

It allows infinite pinning fields provided $\min\{|g_{w}|,b_{w}\}<\infty$ at each vertex. For $u\neq v$, take $b_{u}=+\infty$, $b_{w}=0$ off $u$, $g_{u}=0$, and $g_{w}=a_{w}$ off $u$. The field at a pinned spin is irrelevant. Thus the left side is $2\Psi_{a}(u,v)$, while zero-field spin-flip symmetry makes the right side $2\mathbb{E}_{0}[\sigma_{u}\sigma_{v}]$. Ferromagnetic monotonicity, including the diagonal convention, gives

$$0\leq\Psi_{a}(u,v)\leq\mathbb{E}_{0}[\sigma_{u}\sigma_{v}].$$ (11)

Summing (11) and using (7) or (10) proves the bounds with $k=n$.

After pinning vertices $S$ to a configuration $\tau$, the model on $W=V\setminus S$ has the same couplings on $G[W]$ and fields $a^{\prime}_{w}=a_{w}+\sum_{s\in S:sw\in E}J_{sw}\tau_{s}$. Its maximum degree is at most $\Delta$. Applying the already field-uniform bounds to this graph of size $k$ proves both theorems. Finally, $\varepsilon\leq A/\sqrt{n}$ implies $k\varepsilon\leq A\sqrt{k}$ because $k\leq n$.

5 Glauber dynamics above the threshold

We apply the entropy localization theorem in [1, Theorem 3.21]. In its normalization an interaction matrix $J\succeq 0$ gives density proportional to $\exp(\frac{1}{2}\sigma^{\mathsf{T}}J\sigma+a\cdot\sigma)$. If $\mathfrak{c}(s)$ bounds the largest influence eigenvalue at interaction $sJ$ for every field, the approximate tensorization constant is at most

$$\exp\left\{\|J\|_{2}\int_{0}^{1}\mathfrak{c}(s)\,ds\right\}.$$

Approximate tensorization with constant $K_{\rm AT}$ means that the entropy of every nonnegative function is bounded by $K_{\rm AT}$ times the sum of its expected single-site conditional entropies.

Let $W$ be the weighted adjacency matrix of the given couplings. The matrix $J=W+\Delta\beta I$ is positive semidefinite and $\|J\|_{2}\leq 2\Delta\beta$; its diagonal shift changes the energy by a constant. Using the temperature bound $q=s\beta$, the localization exponent is at most $2\Delta\int_{0}^{\beta}\mathfrak{c}(q)\,dq$, now writing $\mathfrak{c}(q)$ for the bound at temperature upper bound $q$. Put $d=\Delta-1$ and $b_{c}=\operatorname{arctanh}(1/d)$. For $q<b_{c}$, the subcritical spectral bound from [1, Lemma 3.23] and the critical cap give

$$\mathfrak{c}(q)\leq\min\left\{17\sqrt{n},\,\frac{\Delta}{d(1-d\tanh q)}\right\}.$$ (12)

This applies to nonuniform couplings bounded by $q$ as well. Indeed, (11) bounds their field-dependent influence entries by their zero-field correlations, GKS bounds those by the uniform coupling-$q$ correlations, and the Perron eigenvalue is increasing under entrywise comparison of nonnegative matrices. Only a spectral bound is used in (12).

The exact identity

$$\frac{\Delta}{d(1-d\tanh q)}=\frac{1}{d}+\frac{2}{(\Delta-2)(e^{2(b_{c}-q)}-1)}$$

and $e^{x}-1\geq x$ allow integration up to $b_{c}(1-n^{-1/2})$; use $17\sqrt{n}$ on the remaining interval. Since $\Delta b_{c}\leq 3$,

$$2\Delta\int_{0}^{b_{c}}\mathfrak{c}(q)\,dq\leq 105+\frac{\Delta}{\Delta-2}\log n.$$ (13)

For $q\in[b_{c},\beta]$, Theorem 1.2 gives $\mathfrak{c}(q)\leq C[\sqrt{n}+n(d\tanh q-1)]$. With $r=d\tanh q-1\leq 1/64$,

$$\frac{dq}{dr}=\frac{1}{d(1-\tanh^{2}q)}\leq\frac{64}{39d},\qquad 2\Delta\frac{dq}{dr}<5.$$

It follows that

$$2\Delta\int_{b_{c}}^{\beta}\mathfrak{c}(q)\,dq\leq 5C(\sqrt{n}\,\varepsilon+n\varepsilon^{2}/2).$$

If $\beta<b_{c}$, simply omit this last integral and retain (13). Thus

$$K_{\rm AT}\leq e^{105}n^{\Delta/(\Delta-2)}e^{5C(A+A^{2}/2)}.$$

In zero field, $\log(1/\mu_{\min})\leq n(\log 2+\Delta\beta)\leq 4n$ under $\varepsilon\leq 1/64$. Single-site entropy contraction followed by Pinsker’s inequality therefore bounds $T_{\rm mix}$ by a numerical constant times $nK_{\rm AT}\log(en)$. This proves the first assertion of Corollary 1.3. For fixed $B$, the last assertion follows from $A\leq B\sqrt{\log n}$ and $A+A^{2}/2\leq 1/2+A^{2}\leq 1/2+B^{2}\log n$. The restriction $\varepsilon\leq 1/64$ then holds for all sufficiently large $n$.