Introduction
The model and the question
Let G = ( V , E ) G=(V,E) G = ( V , E ) be a finite subgraph of the nearest-neighbour graph of Z 2 \mathbb{Z}^2 Z 2 , let b ∈ [ 0 , ∞ ) E b\in[0,\infty)^E b ∈ [ 0 , ∞ ) E be edge strengths, and let J = ( J 1 , J 2 , J 3 ) ∈ [ 0 , ∞ ) 3 \mathbf{J}=(J^1,J^2,J^3)\in[0,\infty)^3 J = ( J 1 , J 2 , J 3 ) ∈ [ 0 , ∞ ) 3 . The anisotropic Heisenberg model is the probability measure on ( S 2 ) V (S^2)^V ( S 2 ) V with density
1 Z exp ( ∑ { x , y } ∈ E b x y ∑ i = 1 3 J i σ x i σ y i ) (1) \frac{1}{Z}\exp\left(\sum_{\{x,y\}\in E}b_{xy}\sum_{i=1}^{3}J^i\sigma_x^i\sigma_y^i\right) \tag*{(1)} Z 1 exp { x , y } ∈ E ∑ b x y i = 1 ∑ 3 J i σ x i σ y i (1)
with respect to the product of uniform probability measures on S 2 S^2 S 2 ; we write ⟨ ⋅ ⟩ G , b , J \langle\cdot\rangle_{G,b,\mathbf{J}} ⟨ ⋅ ⟩ G , b , J for its expectation, and ⟨ ⋅ ⟩ β , λ \langle\cdot\rangle_{\beta,\lambda} ⟨ ⋅ ⟩ β , λ when b ≡ β b\equiv\beta b ≡ β and J = ( 1 , 1 , λ ) \mathbf{J}=(1,1,\lambda) J = ( 1 , 1 , λ ) (the λ \lambda λ -model below). Except in Theorem 1.5 and Section 7, all boundary conditions in the statements of results are free. For J = ( 1 , 1 , 1 ) \mathbf{J}=(1,1,1) J = ( 1 , 1 , 1 ) this is the classical Heisenberg model, and Polyakov [24 ] predicted exponential decay of correlations at every β \beta β . Proofs have recently been posted [21 , 5 ] . Write ⟨ ⋅ ⟩ G , b ( n ) \langle\cdot\rangle^{(n)}_{G,b} ⟨ ⋅ ⟩ G , b ( n ) for the S n − 1 S^{n-1} S n − 1 -valued model with density proportional to exp ( ∑ { x , y } ∈ E b x y σ x ⋅ σ y ) \exp(\sum_{\{x,y\}\in E} b_{xy}\sigma_x\cdot\sigma_y) exp ( ∑ { x , y } ∈ E b x y σ x ⋅ σ y ) . In the form uniform in subgraphs and strengths, for every n ≥ 3 n\ge3 n ≥ 3 and β < ∞ \beta<\infty β < ∞ there are A , m > 0 A,m>0 A , m > 0 with
0 ≤ ⟨ σ x ⋅ σ y ⟩ G , b ( n ) ≤ A e − m ∥ x − y ∥ 2 (2) 0\le\langle\sigma_x\cdot\sigma_y\rangle^{(n)}_{G,b}\le Ae^{-m\lVert x-y\rVert_2} \tag*{(2)} 0 ≤ ⟨ σ x ⋅ σ y ⟩ G , b ( n ) ≤ A e − m ∥ x − y ∥ 2 (2)
for every finite subgraph G G G of Z 2 \mathbb{Z}^2 Z 2 and every b ∈ [ 0 , β ] E b\in[0,\beta]^E b ∈ [ 0 , β ] E [21 Theorem 1.1] . Ding, Huang and Yang prove exponential decay of the free-boundary two-point function in the boxes B L B_L B L of side length L L L centred at the origin, for points of B L / 2 B_{L/2} B L /2 , and of the one-point function under constant boundary conditions, for every n ≥ 3 n\ge3 n ≥ 3 [5 Theorem 1.1] . Their annulus crossing bound [5 Theorem 5.2] , combined with the sign-cluster argument of § 2 –§ 3 , also gives (1.2) in the uniform form (Remark 5.3).
The λ \lambda λ -model has b ≡ β b\equiv\beta b ≡ β and J = ( 1 , 1 , λ ) \mathbf{J}=(1,1,\lambda) J = ( 1 , 1 , λ ) . For λ > 1 \lambda>1 λ > 1 it has an Ising-type ordered phase at low temperature [20 , 18 , 10 ] . For λ < 1 \lambda<1 λ < 1 , Bricmont, Lebowitz and Pfister conjectured a Berezinskii–Kosterlitz–Thouless (BKT) phase at low temperature [3 ] , which Dunlop proved for small λ \lambda λ [7 ] and Garban and de Montgolfier proved for every λ ∈ ( − 1 , 1 ) \lambda\in(-1,1) λ ∈ ( − 1 , 1 ) [11 Theorem 1.2] : there is β ˉ ( λ ) \bar{\beta}(\lambda) β ˉ ( λ ) such that for β ≥ β ˉ ( λ ) \beta\ge\bar{\beta}(\lambda) β ≥ β ˉ ( λ ) the in-plane two-point function in free or Dirichlet infinite-volume limits, in particular in free-boundary limits, decays at most polynomially. The conjectured phase diagram [11 Figure 2] has a curve λ c ( T ) \lambda_c(T) λ c ( T ) separating the BKT phase from an exponential phase and approaching λ = 1 \lambda=1 λ = 1 as T → 0 T\to0 T → 0 , with renormalization-group prediction λ c ( T ) ≈ 1 − e − c / T \lambda_c(T)\approx1-e^{-c/T} λ c ( T ) ≈ 1 − e − c / T [17 , 14 , 23 ] . For the hemisphere model, in which the spins are restricted to the upper half of S 2 S^2 S 2 , Garban and de Montgolfier proved that β ˉ ( λ ) → ∞ \bar{\beta}(\lambda)\to\infty β ˉ ( λ ) → ∞ as λ → 1 \lambda\to1 λ → 1 , with exponential decay whenever exp ( C β ) < 1 / ( 1 − λ ) \exp(C\beta)<1/(1-\lambda) exp ( C β ) < 1/ ( 1 − λ ) [11 Theorem 1.7] . For the full sphere, whether β ˉ ( λ ) → ∞ \bar{\beta}(\lambda)\to\infty β ˉ ( λ ) → ∞ as λ ↑ 1 \lambda\uparrow1 λ ↑ 1 has been open, and it is at least as hard as Polyakov’s prediction: by a correlation inequality of Dunlop [6 ] the in-plane correlation is nonincreasing in λ \lambda λ [11 Proposition A.1] , so exponential decay at some λ < 1 \lambda<1 λ < 1 gives it at λ = 1 \lambda=1 λ = 1 .
Results near isotropy
Theorem 1.1. For every β ∈ ( 0 , ∞ ) \beta\in(0,\infty) β ∈ ( 0 , ∞ ) there are δ = δ ( β ) > 0 \delta=\delta(\beta)>0 δ = δ ( β ) > 0 and A , m > 0 A,m>0 A , m > 0 such that, for every finite subgraph G G G of Z 2 \mathbb{Z}^2 Z 2 , every b ∈ [ 0 , β ] E b\in[0,\beta]^E b ∈ [ 0 , β ] E , every J ∈ [ 1 − δ , 1 + δ ] 3 \mathbf{J}\in[1-\delta,1+\delta]^3 J ∈ [ 1 − δ , 1 + δ ] 3 , every i ∈ { 1 , 2 , 3 } i\in\{1,2,3\} i ∈ { 1 , 2 , 3 } and all x , y ∈ V x,y\in V x , y ∈ V ,
0 ≤ ⟨ σ x i σ y i ⟩ G , b , J ≤ A e − m ∥ x − y ∥ 2 . 0\le\langle\sigma_x^i\sigma_y^i\rangle_{G,b,\mathbf{J}}\le Ae^{-m\lVert x-y\rVert_2}. 0 ≤ ⟨ σ x i σ y i ⟩ G , b , J ≤ A e − m ∥ x − y ∥ 2 . One can take A = A 0 : = 1 / ( 4 ( 1 − 2 − 1 / 25 ) ) A=A_0:=1/(4(1-2^{-1/25})) A = A 0 := 1/ ( 4 ( 1 − 2 − 1/25 )) , an absolute constant.
The off-diagonal functions ⟨ σ x i σ y j ⟩ \langle\sigma_x^i\sigma_y^j\rangle ⟨ σ x i σ y j ⟩ , i ≠ j i\ne j i = j , vanish, because (1.1) is invariant under the sign change of the i i i th coordinate at every site, so Theorem 1.1 bounds every spin two-point function. The bounds are uniform in the volume, so they pass to free-boundary limits (§ 6 ). Applied with b = β ′ ≤ β b=\beta'\le\beta b = β ′ ≤ β , Theorem 1.1 gives the following.
Corollary 1.2. If ∣ 1 − λ ∣ ≤ δ ( β ) |1-\lambda|\le\delta(\beta) ∣1 − λ ∣ ≤ δ ( β ) , the λ \lambda λ -model has exponentially decaying two-point functions at every inverse temperature β ′ ∈ ( 0 , β ] \beta'\in(0,\beta] β ′ ∈ ( 0 , β ] , uniformly in finite free volumes and in free-boundary limits. Consequently every threshold β ˉ ( λ ) \bar{\beta}(\lambda) β ˉ ( λ ) as in [11 Theorem 1.2] satisfies β ˉ ( λ ) → ∞ \bar{\beta}(\lambda)\to\infty β ˉ ( λ ) → ∞ as λ ↑ 1 \lambda\uparrow1 λ ↑ 1 ; and if 0 ≤ λ − 1 ≤ δ ( β ) 0\le\lambda-1\le\delta(\beta) 0 ≤ λ − 1 ≤ δ ( β ) , there is no long-range order in free-boundary limits at any β ′ ≤ β \beta'\le\beta β ′ ≤ β .
The same argument applies to other small anisotropies for which the planar energies in the proof stay nonnegative combinations of cosines (§ 2 ); for single-ion terms this is the alignment condition below.
Theorem 1.3. There is δ ′ ( β ) ∈ ( 0 , δ ( β ) ] \delta'(\beta)\in(0,\delta(\beta)] δ ′ ( β ) ∈ ( 0 , δ ( β )] , given in § 5 , such that the following hold, with the constants A , m A,m A , m of Theorem 1.1 , which do not depend on n n n .
(a) The conclusion of Theorem 1.1 holds for the density exp ( ∑ e b e ∑ i J i σ x i σ y i + ∑ x ∑ i c x , i ( σ x i ) 2 ) \exp\left(\sum_{e} b_e\sum_i J^{i}\sigma_x^i\sigma_y^i+\sum_x\sum_i c_{x,i}(\sigma_x^i)^2\right) exp ( ∑ e b e ∑ i J i σ x i σ y i + ∑ x ∑ i c x , i ( σ x i ) 2 ) with J ∈ [ 1 − δ ′ , 1 + δ ′ ] 3 \mathbf{J}\in[1-\delta',1+\delta']^3 J ∈ [ 1 − δ ′ , 1 + δ ′ ] 3 and c x , i ∈ [ − β δ ′ , β δ ′ ] c_{x,i}\in[-\beta\delta',\beta\delta'] c x , i ∈ [ − β δ ′ , β δ ′ ] , provided the anisotropies are aligned: some permutation π \pi π of { 1 , 2 , 3 } \{1,2,3\} { 1 , 2 , 3 } satisfies J π 1 ≥ J π 2 ≥ J π 3 J^{\pi1}\ge J^{\pi2}\ge J^{\pi3} J π 1 ≥ J π 2 ≥ J π 3 and c x , π 1 ≥ c x , π 2 ≥ c x , π 3 c_{x,\pi1}\ge c_{x,\pi2}\ge c_{x,\pi3} c x , π 1 ≥ c x , π 2 ≥ c x , π 3 for every x x x .
(b) For every n ≥ 3 n\ge3 n ≥ 3 , the conclusion holds for the S n − 1 S^{n-1} S n − 1 -valued model with density proportional to exp ( ∑ e b e ∑ i = 1 n J i σ x i σ y i ) \exp\left(\sum_e b_e\sum_{i=1}^n J^i\sigma_x^i\sigma_y^i\right) exp ( ∑ e b e ∑ i = 1 n J i σ x i σ y i ) and J ∈ [ 1 − δ , 1 + δ ] n \mathbf{J}\in[1-\delta,1+\delta]^n J ∈ [ 1 − δ , 1 + δ ] n , with δ = δ ( β ) \delta=\delta(\beta) δ = δ ( β ) of Theorem 1.1.
Part (a) covers the confining potential H ~ ε = − ∑ ⟨ σ i , σ j ⟩ + ε ∑ ( σ i z ) 2 \widetilde H_\varepsilon=-\sum\langle\sigma_i,\sigma_j\rangle+\varepsilon\sum(\sigma_i^z)^2 H ε = − ∑ ⟨ σ i , σ j ⟩ + ε ∑ ( σ i z ) 2 of [11 Definition 1.8(4)] for ∣ ε ∣ ≤ δ ′ ( β ) |\varepsilon|\le\delta'(\beta) ∣ ε ∣ ≤ δ ′ ( β ) , for which Garban and de Montgolfier state a BKT phase at large β \beta β when ε > 0 \varepsilon>0 ε > 0 , with an outline of proof [11 Theorem 1.9] . It also covers the λ \lambda λ -model with an easy-plane single-ion term when λ ≤ 1 \lambda\le1 λ ≤ 1 , or an easy-axis one when λ ≥ 1 \lambda\ge1 λ ≥ 1 . Finally, with free boundary conditions the in-plane correlation ⟨ σ x 1 σ y 1 ⟩ β , λ \langle\sigma_x^1\sigma_y^1\rangle_{\beta,\lambda} ⟨ σ x 1 σ y 1 ⟩ β , λ is nonincreasing and the out-of-plane correlation ⟨ σ x 3 σ y 3 ⟩ β , λ \langle\sigma_x^3\sigma_y^3\rangle_{\beta,\lambda} ⟨ σ x 3 σ y 3 ⟩ β , λ nondecreasing in λ ∈ [ 0 , ∞ ) \lambda\in[0,\infty) λ ∈ [ 0 , ∞ ) [11 Proposition A.1] . Combined with (1.2), this bounds the weakly coupled components at every λ \lambda λ and β \beta β .
Corollary 1.4. For every β > 0 \beta>0 β > 0 and with free boundary conditions on finite boxes of Z 2 \mathbb{Z}^2 Z 2 : if λ ∈ [ 0 , 1 ] \lambda\in[0,1] λ ∈ [ 0 , 1 ] then ⟨ σ x 3 σ y 3 ⟩ β , λ ≤ A 3 e − m ∥ x − y ∥ 2 \langle\sigma_x^3\sigma_y^3\rangle_{\beta,\lambda}\le\frac{A}{3}e^{-m\|x-y\|_2} ⟨ σ x 3 σ y 3 ⟩ β , λ ≤ 3 A e − m ∥ x − y ∥ 2 , and if λ ≥ 1 \lambda\ge1 λ ≥ 1 then ⟨ σ x 1 σ y 1 ⟩ β , λ ≤ A 3 e − m ∥ x − y ∥ 2 \langle\sigma_x^1\sigma_y^1\rangle_{\beta,\lambda}\le\frac{A}{3}e^{-m\|x-y\|_2} ⟨ σ x 1 σ y 1 ⟩ β , λ ≤ 3 A e − m ∥ x − y ∥ 2 , with A , m A,m A , m from (1.2) for n = 3 n=3 n = 3 ; for instance A = 3 A 0 A=3A_0 A = 3 A 0 and m = m 0 ( β ) m=m_0(\beta) m = m 0 ( β ) of Remark 5.3.
Rates
How fast must δ ( β ) \delta(\beta) δ ( β ) shrink as β → ∞ \beta\to\infty β → ∞ ? In the renormalization-group picture the anisotropy is negligible on short scales, where the model looks isotropic: its energy on a correlated block of side ℓ \ell ℓ is of order β δ ℓ 2 \beta\delta\ell^2 β δ ℓ 2 . It takes over at the scale ξ ( β ) \xi(\beta) ξ ( β ) of the isotropic correlation length once β δ ξ ( β ) 2 \beta\delta\xi(\beta)^2 β δ ξ ( β ) 2 is of order one; the crossover is at β δ ξ ( β ) 2 ≍ 1 \beta\delta\xi(\beta)^2\asymp1 β δ ξ ( β ) 2 ≍ 1 . For n = 3 n=3 n = 3 the predicted correlation length is ξ ≈ e 2 π β / β \xi\approx e^{2\pi\beta}/\beta ξ ≈ e 2 π β / β , so the crossover is at β ≈ ( 4 π ) − 1 ∣ log δ ∣ \beta\approx(4\pi)^{-1}|\log\delta| β ≈ ( 4 π ) − 1 ∣ log δ ∣ ; for the λ \lambda λ -model this is the predicted BKT threshold β B K T ( λ ) ≈ ( 4 π ) − 1 log ( 1 / ( 1 − λ ) ) \beta_{\mathrm{BKT}}(\lambda)\approx(4\pi)^{-1}\log(1/(1-\lambda)) β BKT ( λ ) ≈ ( 4 π ) − 1 log ( 1/ ( 1 − λ )) [14 , 15 ] . For n = 4 n=4 n = 4 , [22 Theorem 1.1] gives ξ ≍ e π β / β \xi\asymp e^{\pi\beta}/\sqrt{\beta} ξ ≍ e π β / β for large β \beta β , so the crossover is at β ≈ ( 2 π ) − 1 ∣ log δ ∣ \beta\approx(2\pi)^{-1}|\log\delta| β ≈ ( 2 π ) − 1 ∣ log δ ∣ . For n ≥ 4 n\ge4 n ≥ 4 we reach this O ( 4 ) O(4) O ( 4 ) threshold; for n = 4 n=4 n = 4 it is the crossover, and for n ≥ 5 n\ge5 n ≥ 5 it lies below the O ( n ) O(n) O ( n ) prediction.
Theorem 1.5. Assume the estimates (I1)–(I3) of Section 7.2, taken from [22 ] : a stiffness bound that fixes the coarse scale, a deterministic profile construction, and a bad-site estimate conditional on the outside rings. There are absolute constants A ∗ , C Ξ < ∞ A_*,C_\Xi<\infty A ∗ , C Ξ < ∞ such that, with
Ξ ( β ) : = exp ( π β + C Ξ ( 1 + β ) log ( 2 + β ) ) , \Xi(\beta):=\exp\left(\pi\beta+C_\Xi\sqrt{(1+\beta)\log(2+\beta)}\right), Ξ ( β ) := exp ( π β + C Ξ ( 1 + β ) log ( 2 + β ) ) , for every ε ∈ ( 0 , 1 ) \varepsilon\in(0,1) ε ∈ ( 0 , 1 ) there is δ 0 ( ε ) > 0 \delta_0(\varepsilon)>0 δ 0 ( ε ) > 0 with the following property. Let δ ∈ ( 0 , δ 0 ( ε ) ) \delta\in(0,\delta_0(\varepsilon)) δ ∈ ( 0 , δ 0 ( ε )) , 0 < β ≤ ( 1 − ε ) ∣ log δ ∣ / ( 2 π ) 0<\beta\le(1-\varepsilon)|\log\delta|/(2\pi) 0 < β ≤ ( 1 − ε ) ∣ log δ ∣/ ( 2 π ) and n ≥ 4 n\ge4 n ≥ 4 . Let V ⊂ Z 2 V\subset\mathbb{Z}^2 V ⊂ Z 2 be finite, E E E a set of nearest-neighbour edges between sites of V V V , P ⊂ V P\subset V P ⊂ V a set of pinned sites with arbitrary values σ ˉ ∈ ( S n − 1 ) P \bar\sigma\in(S^{n-1})^P σ ˉ ∈ ( S n − 1 ) P , and, for each e ∈ E e\in E e ∈ E , b e ∈ [ − β , β ] b_e\in[-\beta,\beta] b e ∈ [ − β , β ] and m e ∈ R n \mathbf{m}_e\in\mathbb{R}^n m e ∈ R n with ∣ m e k − b e ∣ ≤ β δ |m_e^k-b_e|\le\beta\delta ∣ m e k − b e ∣ ≤ β δ for all k k k ; for instance m e = b e J \mathbf{m}_e=b_e\mathbf{J} m e = b e J with b e ∈ [ 0 , β ] b_e\in[0,\beta] b e ∈ [ 0 , β ] and J ∈ [ 1 − δ , 1 + δ ] n \mathbf{J}\in[1-\delta,1+\delta]^n J ∈ [ 1 − δ , 1 + δ ] n . Let ⟨ ⋅ ⟩ \langle\cdot\rangle ⟨ ⋅ ⟩ denote expectation under the probability measure on ( S n − 1 ) V ∖ P (S^{n-1})^{V\setminus P} ( S n − 1 ) V ∖ P with density proportional to exp ( ∑ { x , y } ∈ E ∑ k m x y k σ x k σ y k ) \exp\left(\sum_{\{x,y\}\in E}\sum_k m_{xy}^k\sigma_x^k\sigma_y^k\right) exp ( ∑ { x , y } ∈ E ∑ k m x y k σ x k σ y k ) with respect to the product of uniform measures, where σ p = σ ˉ p \sigma_p=\bar\sigma_p σ p = σ ˉ p for p ∈ P p\in P p ∈ P . Then, for all i , j ∈ { 1 , … , n } i,j\in\{1,\ldots,n\} i , j ∈ { 1 , … , n } and x , y ∈ V ∖ P x,y\in V\setminus P x , y ∈ V ∖ P ,
⟨ σ x i ⟩ ≤ A ∗ exp ( − dist ∞ ( x , P ) / Ξ ( β ) ) , ⟨ σ x i σ y j ⟩ ≤ A ∗ exp ( − min { ∥ x − y ∥ ∞ , dist ∞ ( x , P ) } / Ξ ( β ) ) . \begin{aligned}
\langle\sigma_x^i\rangle&\le A_*\exp\left(-\operatorname{dist}_\infty(x,P)/\Xi(\beta)\right),\\
\langle\sigma_x^i\sigma_y^j\rangle&\le A_*\exp\left(-\min\{\|x-y\|_\infty,\operatorname{dist}_\infty(x,P)\}/\Xi(\beta)\right).
\end{aligned} ⟨ σ x i ⟩ ⟨ σ x i σ y j ⟩ ≤ A ∗ exp ( − dist ∞ ( x , P ) /Ξ ( β ) ) , ≤ A ∗ exp ( − min { ∥ x − y ∥ ∞ , dist ∞ ( x , P )} /Ξ ( β ) ) . with dist ∞ ( x , ∅ ) = ∞ \operatorname{dist}_\infty(x,\varnothing)=\infty dist ∞ ( x , ∅ ) = ∞ . With free boundary conditions (P = ∅ P=\varnothing P = ∅ ) this gives ∣ ⟨ σ x i σ y j ⟩ ∣ ≤ A ∗ e − ∥ x − y ∥ ∞ / Ξ ( β ) |\langle\sigma_x^i\sigma_y^j\rangle|\le A_*e^{-\|x-y\|_\infty/\Xi(\beta)} ∣ ⟨ σ x i σ y j ⟩ ∣ ≤ A ∗ e − ∥ x − y ∥ ∞ /Ξ ( β ) , and ⟨ σ x i σ y j ⟩ = 0 \langle\sigma_x^i\sigma_y^j\rangle=0 ⟨ σ x i σ y j ⟩ = 0 for i ≠ j i\ne j i = j . On square tori ( Z / L Z ) 2 (\mathbb{Z}/L\mathbb{Z})^2 ( Z / L Z ) 2 with such couplings on all nearest-neighbour edges, ∣ ⟨ σ x i σ y j ⟩ ∣ ≤ e − d T ( x , y ) / Ξ ( β ) |\langle\sigma_x^i\sigma_y^j\rangle|\le e^{-d_T(x,y)/\Xi(\beta)} ∣ ⟨ σ x i σ y j ⟩ ∣ ≤ e − d T ( x , y ) /Ξ ( β ) , with d T d_T d T the cyclic ℓ ∞ \ell^\infty ℓ ∞ distance. Every infinite-volume Gibbs measure μ \mu μ for such couplings on the edges of Z 2 \mathbb{Z}^2 Z 2 has μ ( σ x i ) = 0 \mu(\sigma_x^i)=0 μ ( σ x i ) = 0 and ∣ μ ( σ x i σ y j ) ∣ ≤ A ∗ e − ∥ x − y ∥ ∞ / Ξ ( β ) |\mu(\sigma_x^i\sigma_y^j)|\le A_*e^{-\|x-y\|_\infty/\Xi(\beta)} ∣ μ ( σ x i σ y j ) ∣ ≤ A ∗ e − ∥ x − y ∥ ∞ /Ξ ( β ) .
For n = 4 n = 4 n = 4 , Theorem 7.1 gives the same bounds for every odd observable of one spin and every anisotropy matrix M e M_e M e with ∥ M e − b e I ∥ o p ≤ β δ \lVert M_e-b_e I\rVert_{\mathrm{op}} \le\beta\delta ∥ M e − b e I ∥ op ≤ β δ . The condition on β \beta β can be relaxed to β ≤ ( 2 π ) − 1 ( ∣ log δ ∣ − K ∣ log δ ∣ log ∣ log δ ∣ ) \beta\le(2\pi)^{-1}\left(\lvert\log\delta\rvert-K\sqrt{\lvert\log\delta\rvert}\log\lvert\log\delta\rvert\right) β ≤ ( 2 π ) − 1 ( ∣ log δ ∣ − K ∣ log δ ∣ log ∣ log δ ∣ ) for small δ \delta δ (Lemma 7.9 ). The proof uses the condition only in the form β δ Ξ ( β ) 2 ≤ c 1 \beta\delta\Xi(\beta)^2 \le c_1 β δ Ξ ( β ) 2 ≤ c 1 : the crossover condition with ξ \xi ξ replaced by Ξ ( β ) \Xi(\beta) Ξ ( β ) , which differs from e π β / β e^{\pi\beta}/\sqrt{\beta} e π β / β by a factor e O ( β log β ) e^{O(\sqrt{\beta}\log\beta)} e O ( β l o g β ) . Ξ ( β ) \Xi(\beta) Ξ ( β ) has the form of the length scale of [22 (1.7), (3.1)] , and 1 / Ξ ( β ) 1/\Xi(\beta) 1/Ξ ( β ) that of the lower bound on the O ( 4 ) O(4) O ( 4 ) mass in [22 Corollary 1.2] . So, to leading order, the threshold of Theorem 1.5 is the crossover. It applies, for example, to the easy-axis model J = ( 1 + δ , 1 , 1 , 1 ) J=(1+\delta,1,1,1) J = ( 1 + δ , 1 , 1 , 1 ) , for which the renormalization-group picture predicts Ising-type order at larger β \beta β : no Gibbs measure is magnetized at β ≤ ( 1 − ε ) ∣ log δ ∣ / ( 2 π ) \beta\le(1-\varepsilon)\lvert\log\delta\rvert/(2\pi) β ≤ ( 1 − ε ) ∣ log δ ∣ / ( 2 π ) . For n ≥ 5 n \ge5 n ≥ 5 the threshold is that of O ( 4 ) O(4) O ( 4 ) , below the prediction ( n − 2 ) ∣ log δ ∣ / ( 4 π ) (n-2)\lvert\log\delta\rvert/(4\pi) ( n − 2 ) ∣ log δ ∣ / ( 4 π ) . Theorem 1.5 holds under every boundary condition and for signed couplings, but only for odd observables: the analogous one-point statement for even observables with vanishing S 3 S^3 S 3 -average, such as ( σ x ) 2 − 1 4 (\sigma_x)^2-\frac{1}{4} ( σ x ) 2 − 4 1 for n = 4 n=4 n = 4 , fails already on two sites (Remark 7.10).
For n = 3 n=3 n = 3 we have a weaker rate, from the explicit constants of [21 ] ; the argument behind Theorem 1.5 does not reach n = 3 n=3 n = 3 (Remark 7.11).
Theorem 1.6. There is an absolute constant C < ∞ C < \infty C < ∞ such that δ ( β ) \delta(\beta) δ ( β ) in Theorem 1.1 can be taken with δ ( β ) ≥ exp ( − C ( 1 + β 2 ) ) \delta(\beta) \ge\exp(-C(1+\beta^2)) δ ( β ) ≥ exp ( − C ( 1 + β 2 )) . Hence there is an absolute c > 0 c>0 c > 0 such that, for λ \lambda λ close to 1 1 1 , the λ \lambda λ -model has exponentially decaying correlations at every β ≤ c ∣ log ∣ 1 − λ ∣ ∣ 1 / 2 \beta\le c\lvert\log\lvert1-\lambda\rvert\rvert^{1/2} β ≤ c ∣ log ∣ 1 − λ ∣ ∣ 1/2 .
The exponent 1 2 \frac{1}{2} 2 1 comes from the factor β 2 \beta^2 β 2 in the rotation estimate of [21 Proposition 2.2] (§ 4 ); the prediction is of order ∣ log ∣ 1 − λ ∣ ∣ \lvert\log\lvert1-\lambda\rvert\rvert ∣ log ∣ 1 − λ ∣∣ .
Sources
Theorems 1.1 and 1.3 and Corollaries 1.2 and 1.4 rest on [5 Theorem 5.2] and the loop representation of [5 §§2–3] , used through Proposition 3.1 with non-constant couplings (Remark 3.2); the sign-cluster arguments of Campbell and Chayes [4 ] and of [21 §4] are written out in § 2 and § 3 , and these results use no statement of [21 ] . Corollaries 1.2 and 1.4 also use [11 Theorem 1.2 and Proposition A.1] . Theorem 1.6 uses [21 Proposition 3.1] in place of Proposition 3.1 (§ 4 ) and the explicit constants in its proof, namely [21 Lemma 2.1, Proposition 2.2, Lemma 3.2, (3.10), (3.16)–(3.20)] . Theorem 1.5 uses [22 (3.54) and (3.58)] and the construction of [22 §3.10.1] , in the form stated in § 7.2 ; in [22 ] , (3.58) is proved from the stiffness bound in (3.54) and [22 (3.10), (3.11), (3.13), (3.14)] , that is, from the stiffness iteration of [22 §§3.1–3.8] . None of [5 ] , [21 ] and [22 ] has been refereed.
Idea of the proofs
The proofs of Theorems 1.1, 1.3 and 1.6 share one scheme, which goes back to Campbell and Chayes [4 ] and is carried out for the isotropic model in [21 §4] : a sign-cluster representation turns a crossing bound for a single isotropic square annulus { N ≤ ∥ v ∥ ∞ ≤ 2 N } \{N \le\lVert v\rVert_\infty\le2N\} { N ≤ ∥ v ∥ ∞ ≤ 2 N } into exponential decay. The signs of one spin component form a ferromagnetic Ising model given the amplitudes, open bonds join equal signs, pinning annular layers to the pole increases crossing probabilities, and a sum over coarse paths at one fixed scale N N N gives a positive mass. This part uses only positive association, and its isotropy enters through a Ginibre-type inequality for the planar components [21 Lemma 4.2] . The annulus bound has two suppliers. For Theorems 1.1 and 1.3 it is [5 Theorem 5.2] . Ding, Huang and Yang prove exponential decay through loop soups and Lupu’s isomorphism [19 ] ; their key estimate bounds the probability that an annulus is crossed by a loop cluster, uniformly over a class of boundary conditions, and its proof uses spin waves on S 1 S^1 S 1 instead of rotations of S 2 S^2 S 2 about two non-commuting axes [5 Remark 1.5] . For Theorem 1.6 it is [21 Proposition 3.1] : if the outer layer of the annulus is pinned to e 1 e_1 e 1 and its inner layer to e 1 e_1 e 1 rotated by an angle t t t about an axis orthogonal to e 1 e_1 e 1 , the partition function Z ( t ) Z(t) Z ( t ) satisfies min t Z ( t ) / max t Z ( t ) ≥ 1 − C β / log N \min_t Z(t)/\max_t Z(t) \ge1-C_\beta/\log N min t Z ( t ) / max t Z ( t ) ≥ 1 − C β / log N . This step uses rotations about two non-commuting axes, and it fails for anisotropic couplings; this is why we use the isotropic input only on one finite annulus.
Figure 1. A schematic phase diagram of the λ \lambda λ -model (n = 3 n = 3 n = 3 ), not to scale. Along λ = 1 \lambda= 1 λ = 1 correlations decay exponentially at every β \beta β [21 , 5 ] . Theorems 1.1 and 1.6 give exponential decay in the grey region around λ = 1 \lambda= 1 λ = 1 , whose half-width at height β \beta β is at least e − C ( 1 + β 2 ) e^{-C(1+\beta^2)} e − C ( 1 + β 2 ) (drawn much wider than it is); it contains every point below it at the same λ \lambda λ , because the estimates are uniform in strengths up to β \beta β . The hatched region is the BKT phase of [11 Theorem 1.2] , whose boundary β ˉ ( λ ) \bar{\beta}(\lambda) β ˉ ( λ ) must therefore stay above the grey region and tend to infinity as λ ↑ 1 \lambda\uparrow1 λ ↑ 1 . The dotted region is the Ising phase for λ > 1 \lambda> 1 λ > 1 . The dashed curves sketch the conjectured transition lines β ≈ ( 4 π ) − 1 log ∣ 1 − λ ∣ \beta\approx(4\pi)^{-1} \log|1-\lambda| β ≈ ( 4 π ) − 1 log ∣1 − λ ∣ , which the renormalization-group computation gives on both sides [14 , 15 ] ; the drawn shapes of the hatched and dotted boundaries are illustrative only.
For anisotropic couplings we keep the sign-cluster part and use the isotropic input on one annulus at one scale N N N only. When both layers are pinned to the pole, the bound of [5 Theorem 5.2] is a bound on the probability that open sign-cluster bonds cross the isotropic annulus (Proposition 3.1 ). On that finite graph the anisotropic and isotropic Hamiltonians differ by at most β δ ∣ E N ∣ \beta\delta|E_N| β δ ∣ E N ∣ , so the two partition functions with the same pins agree up to a factor e ± β δ ∣ E N ∣ e^{\pm\beta\delta|E_N|} e ± β δ ∣ E N ∣ . Choosing N N N first, as a function of β \beta β , and then δ ≈ ( β N 2 ) − 1 \delta\approx(\beta N^2)^{-1} δ ≈ ( β N 2 ) − 1 transfers the crossing bound (Lemma 3.3 ). For the second part, the planar energy J l cos θ x cos θ y + J k sin θ x sin θ y J^l\cos\theta_x\cos\theta_y + J^k\sin\theta_x\sin\theta_y J l cos θ x cos θ y + J k sin θ x sin θ y is a combination with nonnegative coefficients of cos ( θ x − θ y ) \cos(\theta_x-\theta_y) cos ( θ x − θ y ) and cos ( θ x + θ y ) \cos(\theta_x+\theta_y) cos ( θ x + θ y ) . Ginibre’s doubling argument handles both terms in the same way (Lemma 2.5), so amplitude association, pinning and the crossing identity go through; § 2 writes these steps out. The scale N ( β ) N(\beta) N ( β ) is where the price is paid: δ ( β ) \delta(\beta) δ ( β ) is of order ( β N ( β ) 2 ) − 1 (\beta N(\beta)^2)^{-1} ( β N ( β ) 2 ) − 1 . The scale given by [5 ] is not explicit. With the input of [21 Proposition 3.1] instead, log N ( β ) \log N(\beta) log N ( β ) is proportional to C β C_\beta C β , and Proposition 4.1 shows that C β C_\beta C β is at most quadratic in β \beta β ; the quadratic term comes from the coefficient K r o t ( β ) = 6 β 2 + 4 β 3 / 2 + β K_{\rm rot}(\beta)=6\beta^2+4\beta^{3/2}+\beta K rot ( β ) = 6 β 2 + 4 β 3/2 + β of the rotation estimate [21 Proposition 2.2] . This gives Theorem 1.6.
Theorem 1.5 is proved differently, following the proof of decay for charged observables of the isotropic O ( 4 ) O(4) O ( 4 ) model in [22 §§3.9–3.10] : orientation variables on coarse cells of side s ( β ) = e π β + O ( β log β ) s(\beta)=e^{\pi\beta+O(\sqrt{\beta\log\beta})} s ( β ) = e π β + O ( β l o g β ) , a test for bad coarse vertices, at which rotating the orientation can change the cell actions by more than a fixed amount, and a disagreement coupling in the manner of van den
Figure 2. How the results are proved. A solid arrow means that its source is used in the proof of its target; only the main dependencies are drawn. Box styles give provenance; grey labels give sources and where each item is proved. The adapted steps are listed in § 1.5 .
Berg and Maes [2 ] . The anisotropic and isotropic actions of one coarse cell differ by at most 8 β δ s 2 8\beta\delta s^{2} 8 β δ s 2 (Lemma 7.2). As long as β δ s 2 \beta\delta s^{2} β δ s 2 is small, two things survive. The isotropic bad-site estimate [22 (3.58)] , which holds conditionally on everything outside a collar, transfers to the anisotropic law through a bounded density ratio on the collar (Lemma 7.3). At good vertices the one-site conditional law of the orientation stays close to Haar measure. The averaging over common left rotations by all of S 3 S^{3} S 3 in [22 ] needs isotropy; we replace it by the symmetry σ ↦ − σ \sigma\mapsto-\sigma σ ↦ − σ , which every bilinear interaction keeps and which suffices for odd observables (Lemmas 7.4–7.5 ). The condition that β δ s 2 \beta\delta s^{2} β δ s 2 be small is the crossover condition with ξ = s \xi=s ξ = s .
Adapted results. Lemmas 2.6 and 2.7 and Proposition 2.8 adapt [4 Lemma 2, Corollaries 1 and 3] and [21 Lemmas 4.3, 4.4 and Proposition 4.5] , with Lemma 2.5 in place of [21 Lemma 4.2] ; Lemma 2.9 is [21 (4.9)] ; Lemma 3.4 and Proposition 3.5 adapt [21 Propositions 4.6 and 4.7] ; Proposition 3.1 translates [5 Theorem 5.2] into a bound on sign-cluster crossings; Proposition 4.1 extracts the β \beta β -dependence of the constant in [21 Proposition 3.1] from its proof; Theorem 1.3(b) uses the conditional projection of [1 Proposition 2.4] in the block form of [21 Lemma 4.8] ; Section 7 adapts [22 §§3.9–3.10] , with the credits given there.
§ 2 sets up sign clusters for anisotropic couplings; § 3 proves Theorem 1.1 , § 4 Theorem 1.6, § 5 Theorem 1.3 , § 6 the corollaries, and Section 7 Theorem 1.5 .
Sign clusters for anisotropic couplings
This section writes out, for anisotropic couplings, the sign-cluster representation of Campbell and Chayes [4 ] , which builds on Wolff’s embedding [26 ] and the Edwards–Sokal coupling [8 ] , and the association and pinning arguments of [21 §§4.1–4.2] . Fix J ∈ ( 0 , ∞ ) 3 \mathbf{J} \in(0,\infty)^3 J ∈ ( 0 , ∞ ) 3 and a coordinate i i i . By permuting coordinates (which preserves the reference measure) we may assume i = 1 i=1 i = 1 and J 2 ≥ J 3 J^2 \ge J^3 J 2 ≥ J 3 .
Lemma 2.1 (decomposition). Write, as in [21 (4.1)] ,
σ x = ( r x τ x , q x cos θ x , q x sin θ x ) , q x = ( 1 − r x 2 ) 1 / 2 . \sigma_x = (r_x\tau_x,\ q_x\cos\theta_x,\ q_x\sin\theta_x), \qquad q_x = (1-r_x^2)^{1/2}. σ x = ( r x τ x , q x cos θ x , q x sin θ x ) , q x = ( 1 − r x 2 ) 1/2 . Under the reference measure r x , τ x , θ x r_x,\tau_x,\theta_x r x , τ x , θ x are independent and uniform on [ 0 , 1 ] [0,1] [ 0 , 1 ] , { ± 1 } \{\pm1\} { ± 1 } and R / 2 π Z \mathbb{R}/2\pi\mathbb{Z} R /2 π Z . The weight of an edge is b x y ( J 1 r x r y τ x τ y + q x q y T x y ) b_{xy}(J^1r_xr_y\tau_x\tau_y+q_xq_yT_{xy}) b x y ( J 1 r x r y τ x τ y + q x q y T x y ) with
T x y = J 2 cos θ x cos θ y + J 3 sin θ x sin θ y = α cos ( θ x − θ y ) + γ cos ( θ x + θ y ) , (3) T_{xy}=J^2\cos\theta_x\cos\theta_y+J^3\sin\theta_x\sin\theta_y=\alpha\cos(\theta_x-\theta_y)+\gamma\cos(\theta_x+\theta_y), \tag*{(3)} T x y = J 2 cos θ x cos θ y + J 3 sin θ x sin θ y = α cos ( θ x − θ y ) + γ cos ( θ x + θ y ) , (3) where α = 1 2 ( J 2 + J 3 ) > 0 \alpha=\frac{1}{2}(J^2+J^3)>0 α = 2 1 ( J 2 + J 3 ) > 0 , γ = 1 2 ( J 2 − J 3 ) ≥ 0 \gamma=\frac{1}{2}(J^2-J^3)\ge0 γ = 2 1 ( J 2 − J 3 ) ≥ 0 . Conditionally on r r r , the signs form a ferromagnetic Ising system with couplings K x y = b x y J 1 r x r y K_{xy}=b_{xy}J^1r_xr_y K x y = b x y J 1 r x r y and the angles an independent planar system with couplings L x y = b x y q x q y L_{xy}=b_{xy}q_xq_y L x y = b x y q x q y and edge energies T x y T_{xy} T x y ; the amplitude law has density ρ ( r ) ∝ Z I ( K ( r ) ) Z P ( L ( r ) ) \rho(r)\propto Z_I(K(r))Z_P(L(r)) ρ ( r ) ∝ Z I ( K ( r )) Z P ( L ( r )) on [ 0 , 1 ] V [0,1]^V [ 0 , 1 ] V , which is positive and continuous.
Proof. By Archimedes’ theorem the first coordinate r τ r\tau r τ of a uniform point of S 2 S^2 S 2 is uniform on [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] and independent of the azimuth θ \theta θ . The rest is algebra.
Lemma 2.2 (bonds and the connection identity). Given the spins , sample independent bonds with P ( η e = 1 ∣ σ ) = p e 1 { τ x = τ y } \mathbb{P}(\eta_e=1\mid\sigma)=p_e\mathbf{1}_{\{\tau_x=\tau_y\}} P ( η e = 1 ∣ σ ) = p e 1 { τ x = τ y } , p e = 1 − e − 2 K e p_e=1-e^{-2K_e} p e = 1 − e − 2 K e . Conditionally on ( r , θ , η ) (r,\theta,\eta) ( r , θ , η ) , the signs are independent fair signs, constant on open clusters; conditionally on r r r , the bonds are independent of θ \theta θ and have law proportional to 2 k ( η ) ∏ e p e η e ( 1 − p e ) 1 − η e 2^{k(\eta)}\prod_e p_e^{\eta_e}(1-p_e)^{1-\eta_e} 2 k ( η ) ∏ e p e η e ( 1 − p e ) 1 − η e , where k ( η ) k(\eta) k ( η ) is the number of open clusters, singletons included. Consequently
⟨ σ x 1 σ y 1 ⟩ G , b , J = E [ r x r y 1 { x ↔ y } ] ∈ [ 0 , P ( x ↔ y ) ] . (4) \langle\sigma_x^1\sigma_y^1\rangle_{G,b,\mathbf{J}}=\mathbb{E}\left[r_xr_y\mathbf{1}_{\{x\leftrightarrow y\}}\right]\in[0,\mathbb{P}(x\leftrightarrow y)]. \tag*{(4)} ⟨ σ x 1 σ y 1 ⟩ G , b , J = E [ r x r y 1 { x ↔ y } ] ∈ [ 0 , P ( x ↔ y )] . (4)
Proof. This is the Edwards–Sokal coupling [8 ] for the embedded Ising model [26 ] ; see [4 Corollary 4, (7)] and [21 (4.4)–(4.6)] . With p = 1 − e − 2 K p=1-e^{-2K} p = 1 − e − 2 K one has e K τ τ ′ = e K [ ( 1 − p ) + p 1 { τ = τ ′ } ] e^{K\tau\tau'}=e^K[(1-p)+p\mathbf{1}_{\{\tau=\tau'\}}] e K τ τ ′ = e K [( 1 − p ) + p 1 { τ = τ ′ } ] . Expanding the product over edges, the joint density of ( r , τ , θ , η ) (r,\tau,\theta,\eta) ( r , τ , θ , η ) is proportional to ∏ e e K e p e η e ( 1 − p e ) 1 − η e ∏ e : η e = 1 1 { τ x = τ y } ⋅ exp ( ∑ e L e T e ( θ ) ) \prod_e e^{K_e}p_e^{\eta_e}(1-p_e)^{1-\eta_e}\prod_{e:\eta_e=1}\mathbf{1}_{\{\tau_x=\tau_y\}}\cdot\exp(\sum_e L_eT_e(\theta)) ∏ e e K e p e η e ( 1 − p e ) 1 − η e ∏ e : η e = 1 1 { τ x = τ y } ⋅ exp ( ∑ e L e T e ( θ )) . Its dependence on τ \tau τ given ( r , θ , η ) (r,\theta,\eta) ( r , θ , η ) is the indicator that τ \tau τ is constant on open clusters, summing over η \eta η recovers the Gibbs density, and the conditional law of η \eta η given σ \sigma σ is the stated kernel. Summing over τ \tau τ gives the law of η \eta η given r r r . Hence E [ τ x τ y ∣ r , θ , η ] = 1 { x ↔ y } \mathbb{E}[\tau_x\tau_y\mid r,\theta,\eta]=\mathbf{1}_{\{x\leftrightarrow y\}} E [ τ x τ y ∣ r , θ , η ] = 1 { x ↔ y } and ⟨ σ x 1 σ y 1 ⟩ = E [ r x r y τ x τ y ] \langle\sigma_x^1\sigma_y^1\rangle=\mathbb{E}[r_xr_y\tau_x\tau_y] ⟨ σ x 1 σ y 1 ⟩ = E [ r x r y τ x τ y ] .
Lemma 2.3 (FKG inequality). Let X = ∏ k = 1 m X k X=\prod_{k=1}^m X_k X = ∏ k = 1 m X k , where each X k X_k X k is a compact interval with Lebesgue measure or a finite chain with counting measure, and let w > 0 w>0 w > 0 be continuous and log-supermodular: w ( u ∨ v ) w ( u ∧ v ) ≥ w ( u ) w ( v ) w(u\vee v)w(u\wedge v)\ge w(u)w(v) w ( u ∨ v ) w ( u ∧ v ) ≥ w ( u ) w ( v ) . Then the law μ ∝ w \mu\propto w μ ∝ w is associated: Cov ( F , H ) ≥ 0 \operatorname{Cov}(F,H)\ge0 Cov ( F , H ) ≥ 0 for bounded measurable coordinatewise increasing F , H F,H F , H . Moreover the conditional law of ( x 1 , … , x m − 1 ) (x_1,\ldots,x_{m-1}) ( x 1 , … , x m − 1 ) given x m x_m x m is stochastically increasing in x m x_m x m .
Proof. This is the FKG inequality [9 , 25 ] ; we give the standard induction, as in [21 Lemma 4.1] . For m = 1 m=1 m = 1 , with X ′ X' X ′ an independent copy, 2 Cov ( F , H ) = E [ ( F ( X ) − F ( X ′ ) ) ( H ( X ) − H ( X ′ ) ) ] ≥ 0 2\operatorname{Cov}(F,H)=\mathbb{E}[(F(X)-F(X'))(H(X)-H(X'))]\geq0 2 Cov ( F , H ) = E [( F ( X ) − F ( X ′ )) ( H ( X ) − H ( X ′ ))] ≥ 0 , since both differences have the sign of X − X ′ X-X' X − X ′ . For m > 1 m>1 m > 1 write x = ( u , t ) x=(u,t) x = ( u , t ) . The conditional density w ( ⋅ , t ) w(\cdot,t) w ( ⋅ , t ) is positive, continuous and log-supermodular in u u u , so the conditional law μ t \mu_t μ t is associated by induction. For t ′ ≥ t t'\geq t t ′ ≥ t and u ≤ v u\leq v u ≤ v , log-supermodularity applied to ( u , t ′ ) (u,t') ( u , t ′ ) and ( v , t ) (v,t) ( v , t ) gives w ( v , t ′ ) w ( u , t ) ≥ w ( u , t ′ ) w ( v , t ) w(v,t')w(u,t)\geq w(u,t')w(v,t) w ( v , t ′ ) w ( u , t ) ≥ w ( u , t ′ ) w ( v , t ) , so g ( u ) : = w ( u , t ′ ) / w ( u , t ) g(u):=w(u,t')/w(u,t) g ( u ) := w ( u , t ′ ) / w ( u , t ) is increasing and bounded. Then E t ′ [ F ] = E t [ F g ] / E t [ g ] ≥ E t [ F ] \mathbb{E}_{t'}[F]=\mathbb{E}_t[Fg]/\mathbb{E}_t[g]\geq\mathbb{E}_t[F] E t ′ [ F ] = E t [ F g ] / E t [ g ] ≥ E t [ F ] for increasing F F F , by association of μ t \mu_t μ t ; this is the second assertion, and it makes t ↦ E [ F ∣ t ] t\mapsto\mathbb{E}[F\mid t] t ↦ E [ F ∣ t ] increasing. Finally Cov ( F , H ) = E [ Cov ( F , H ∣ t ) ] + Cov ( E [ F ∣ t ] , E [ H ∣ t ] ) ≥ 0 \operatorname{Cov}(F,H)=\mathbb{E}[\operatorname{Cov}(F,H\mid t)]+\operatorname{Cov}(\mathbb{E}[F\mid t],\mathbb{E}[H\mid t])\geq0 Cov ( F , H ) = E [ Cov ( F , H ∣ t )] + Cov ( E [ F ∣ t ] , E [ H ∣ t ]) ≥ 0 , the first term by induction and the second by the case m = 1 m=1 m = 1 .
Lemma 2.4. (Griffiths’ inequalities). For an Ising model on a finite graph with free boundary conditions and couplings K e ≥ 0 K_e\geq0 K e ≥ 0 , every ⟨ τ x τ y ⟩ ≥ 0 \langle\tau_x\tau_y\rangle\geq0 ⟨ τ x τ y ⟩ ≥ 0 and every Cov ( τ x τ y , τ u τ v ) ≥ 0 \operatorname{Cov}(\tau_x\tau_y,\tau_u\tau_v)\geq0 Cov ( τ x τ y , τ u τ v ) ≥ 0 .
Proof. These are Griffiths’ inequalities in the form of Kelly and Sherman [16 ] and Ginibre [13 ] ; we give Ginibre’s doubling proof, as in [21 Lemma 4.2] . The means are nonnegative because, after expanding the exponential, the reference moments of sign monomials are 0 0 0 or 1 1 1 . For the covariances take two independent copies d , d ′ d,d' d , d ′ and put u = ( d + d ′ ) / 2 u=(d+d')/2 u = ( d + d ′ ) /2 , v = ( d − d ′ ) / 2 v=(d-d')/2 v = ( d − d ′ ) /2 at each site. The pair ( u , v ) (u,v) ( u , v ) takes the values ( ± 1 , 0 ) (\pm1,0) ( ± 1 , 0 ) and ( 0 , ± 1 ) (0,\pm1) ( 0 , ± 1 ) with probability 1 / 4 1/4 1/4 each, so every monomial in ( u , v ) (u,v) ( u , v ) at one site has nonnegative expectation. Since d x d y + d x ′ d y ′ = 2 ( u x u y + v x v y ) d_xd_y+d'_xd'_y=2(u_xu_y+v_xv_y) d x d y + d x ′ d y ′ = 2 ( u x u y + v x v y ) and d x d y − d x ′ d y ′ = 2 ( u x v y + v x u y ) d_xd_y-d'_xd'_y=2(u_xv_y+v_xu_y) d x d y − d x ′ d y ′ = 2 ( u x v y + v x u y ) , the quantity 2 Cov ( τ x τ y , τ u τ v ) 2\operatorname{Cov}(\tau_x\tau_y,\tau_u\tau_v) 2 Cov ( τ x τ y , τ u τ v ) is a ratio whose numerator, after expanding exp ( 2 ∑ e K e ( u x u y + v x v y ) ) \exp(2\sum_e K_e(u_xu_y+v_xv_y)) exp ( 2 ∑ e K e ( u x u y + v x v y )) , is a sum of nonnegative coefficients times products over sites of nonnegative moments.
Lemma 2.5. (planar Ginibre inequality). Consider a planar system on a finite graph with density proportional to exp ( ∑ a L a f a ( θ ) ) \exp(\sum_a L_af_a(\theta)) exp ( ∑ a L a f a ( θ )) , where L a ≥ 0 L_a\geq0 L a ≥ 0 and each f a f_a f a is cos ( n a ⋅ θ ) \cos(n_a\cdot\theta) cos ( n a ⋅ θ ) for an integer vector n a n_a n a . Then every ⟨ f a ⟩ ≥ 0 \langle f_a\rangle\geq0 ⟨ f a ⟩ ≥ 0 and every Cov ( f a , f b ) ≥ 0 \operatorname{Cov}(f_a,f_b)\geq0 Cov ( f a , f b ) ≥ 0 . In particular this holds for the edge energies (3) with L e ≥ 0 L_e\geq0 L e ≥ 0 , and also with additional site terms M x cos ( 2 θ x ) M_x\cos(2\theta_x) M x cos ( 2 θ x ) , M x ≥ 0 M_x\geq0 M x ≥ 0 .
This is a case of Ginibre’s inequalities [13 ] ; in this form it also follows from Dunlop’s inequality as restated in [11 Theorem A.2 and (A.1)–(A.2)] . We include the short proof, which follows [21 proof of Lemma 4.2] .
Proof. Expanding the exponential, products of cosines of integer combinations are nonnegative combinations of such cosines, whose Haar integrals are nonnegative; this gives ⟨ f a ⟩ ≥ 0 \langle f_a\rangle\geq0 ⟨ f a ⟩ ≥ 0 . For covariances, take two independent copies θ , θ ′ \theta,\theta' θ , θ ′ and substitute θ = φ + ψ \theta=\varphi+\psi θ = φ + ψ , θ ′ = φ − ψ \theta'=\varphi-\psi θ ′ = φ − ψ ; the map ( φ , ψ ) ↦ ( φ + ψ , φ − ψ ) (\varphi,\psi)\mapsto(\varphi+\psi,\varphi-\psi) ( φ , ψ ) ↦ ( φ + ψ , φ − ψ ) is a surjective homomorphism of tori, so it preserves Haar probability [21 proof of Lemma 4.2] . For every integer vector n n n ,
cos ( n ⋅ θ ) + cos ( n ⋅ θ ′ ) = 2 cos ( n ⋅ φ ) cos ( n ⋅ ψ ) , cos ( n ⋅ θ ) − cos ( n ⋅ θ ′ ) = − 2 sin ( n ⋅ φ ) sin ( n ⋅ ψ ) . \cos(n\cdot\theta)+\cos(n\cdot\theta')=2\cos(n\cdot\varphi)\cos(n\cdot\psi),\qquad
\cos(n\cdot\theta)-\cos(n\cdot\theta')=-2\sin(n\cdot\varphi)\sin(n\cdot\psi). cos ( n ⋅ θ ) + cos ( n ⋅ θ ′ ) = 2 cos ( n ⋅ φ ) cos ( n ⋅ ψ ) , cos ( n ⋅ θ ) − cos ( n ⋅ θ ′ ) = − 2 sin ( n ⋅ φ ) sin ( n ⋅ ψ ) .
So the doubled density is exp ( 2 ∑ a L a cos ( n a ⋅ φ ) cos ( n a ⋅ ψ ) ) \exp(2\sum_a L_a\cos(n_a\cdot\varphi)\cos(n_a\cdot\psi)) exp ( 2 ∑ a L a cos ( n a ⋅ φ ) cos ( n a ⋅ ψ )) and ( f a − f a ′ ) ( f b − f b ′ ) = 4 sin ( n a ⋅ φ ) sin ( n b ⋅ φ ) sin ( n a ⋅ ψ ) sin ( n b ⋅ ψ ) (f_a-f'_a)(f_b-f'_b)=4\sin(n_a\cdot\varphi)\sin(n_b\cdot\varphi)\sin(n_a\cdot\psi)\sin(n_b\cdot\psi) ( f a − f a ′ ) ( f b − f b ′ ) = 4 sin ( n a ⋅ φ ) sin ( n b ⋅ φ ) sin ( n a ⋅ ψ ) sin ( n b ⋅ ψ ) . Expanding the exponential, 2 Cov ( f a , f b ) = E [ ( f a − f a ′ ) ( f b − f b ′ ) ] 2\operatorname{Cov}(f_a,f_b)=\mathbb{E}[(f_a-f'_a)(f_b-f'_b)] 2 Cov ( f a , f b ) = E [( f a − f a ′ ) ( f b − f b ′ )] is a sum of nonnegative coefficients times ( ∫ F d φ ) 2 (\int F\,d\varphi)^2 ( ∫ F d φ ) 2 for real trigonometric polynomials F F F . Absolute convergence on the finite torus justifies the expansions. For (3) write L e T e = L e α cos ( n e − ⋅ θ ) + L e γ cos ( n e + ⋅ θ ) L_eT_e=L_e\alpha\cos(n_e^-\cdot\theta)+L_e\gamma\cos(n_e^+\cdot\theta) L e T e = L e α cos ( n e − ⋅ θ ) + L e γ cos ( n e + ⋅ θ ) with n e ± = e x ± e y n_e^\pm=e_x\pm e_y n e ± = e x ± e y , and use bilinearity of the covariance.
Lemma 2.6. (amplitude association). The amplitude law with density ρ \rho ρ is associated.
Proof. As in [4 Lemma 2] and [21 Lemma 4.3] . For a partition function Z ( c ) = ∫ exp ( ∑ a c a f a ) Z(c)=\int\exp(\sum_a c_af_a) Z ( c ) = ∫ exp ( ∑ a c a f a ) one has ∂ 2 log Z / ∂ c a ∂ c b = Cov ( f a , f b ) \partial^2\log Z/\partial c_a\partial c_b=\operatorname{Cov}(f_a,f_b) ∂ 2 log Z / ∂ c a ∂ c b = Cov ( f a , f b ) , so for distinct sites x , y x,y x , y ,
∂ x ∂ y log Z ( c ( r ) ) = ∑ a ⟨ f a ⟩ ∂ x ∂ y c a + ∑ a , b Cov ( f a , f b ) ∂ x c a ∂ y c b . \partial_x\partial_y\log Z(c(r))=\sum_a\langle f_a\rangle\,\partial_x\partial_y c_a+\sum_{a,b}\operatorname{Cov}(f_a,f_b)\,\partial_xc_a\,\partial_yc_b. ∂ x ∂ y log Z ( c ( r )) = a ∑ ⟨ f a ⟩ ∂ x ∂ y c a + a , b ∑ Cov ( f a , f b ) ∂ x c a ∂ y c b .
For Z I Z_I Z I , the coefficients K x y = b x y J 1 r x r y K_{xy}=b_{xy}J^1r_xr_y K x y = b x y J 1 r x r y are nondecreasing with ∂ x ∂ y K x y = b x y J 1 ≥ 0 \partial_x\partial_yK_{xy}=b_{xy}J^1\ge0 ∂ x ∂ y K x y = b x y J 1 ≥ 0 , and Lemma 2.4 gives nonnegative means and covariances. For Z P Z_P Z P , the coefficients L x y = b x y q x q y L_{xy}=b_{xy}q_xq_y L x y = b x y q x q y are nonincreasing, so products of two first derivatives are nonnegative, ∂ x ∂ y L x y = b x y r x r y / ( q x q y ) ≥ 0 \partial_x\partial_yL_{xy}=b_{xy}r_xr_y/(q_xq_y)\ge0 ∂ x ∂ y L x y = b x y r x r y / ( q x q y ) ≥ 0 , and Lemma 2.5 gives nonnegative means and covariances. So log ρ \log\rho log ρ is C 2 C^2 C 2 on [ 0 , 1 ) V [0,1)^V [ 0 , 1 ) V with nonnegative mixed derivatives, hence supermodular there; by continuity ρ \rho ρ is log-supermodular on [ 0 , 1 ] V [0,1]^V [ 0 , 1 ] V , and Lemma 2.3 applies.
Lemma 2.7 (joint association). The joint law of ( r , η ) (r,\eta) ( r , η ) is associated.
Proof. As in [4 Corollary 1] and [21 Lemma 4.4] . Given r r r , the bond law of Lemma 2.2 has weight 2 k ( η ) ∏ e p e η e ( 1 − p e ) 1 − η e 2^{k(\eta)}\prod_e p_e^{\eta_e}(1-p_e)^{1-\eta_e} 2 k ( η ) ∏ e p e η e ( 1 − p e ) 1 − η e , and k k k is supermodular (∣ V ∣ |V| ∣ V ∣ minus the rank of a graphic matroid), so this weight is log-supermodular on { 0 , 1 } E \{0,1\}^E { 0 , 1 } E (edges with p e = 0 p_e=0 p e = 0 are closed almost surely; restrict to the others). By Lemma 2.3 the conditional bond law is associated. Raising p e p_e p e tilts it by a factor increasing in η e \eta_e η e , so by association the bond law is stochastically increasing in each p e p_e p e , and p e = 1 − exp ( − 2 b e J 1 r x r y ) p_e=1-\exp(-2b_eJ^1r_xr_y) p e = 1 − exp ( − 2 b e J 1 r x r y ) is nondecreasing in r r r since J 1 > 0 J^1>0 J 1 > 0 . For bounded increasing F ( r , η ) F(r,\eta) F ( r , η ) the map r ↦ E [ F ( r , η ) ∣ r ] r\mapsto\mathbb{E}[F(r,\eta)\mid r] r ↦ E [ F ( r , η ) ∣ r ] is therefore increasing, and Cov ( F , H ) = E [ Cov ( F , H ∣ r ) ] + Cov ( E [ F ∣ r ] , E [ H ∣ r ] ) ≥ 0 \operatorname{Cov}(F,H)=\mathbb{E}[\operatorname{Cov}(F,H\mid r)]+\operatorname{Cov}(\mathbb{E}[F\mid r],\mathbb{E}[H\mid r])\ge0 Cov ( F , H ) = E [ Cov ( F , H ∣ r )] + Cov ( E [ F ∣ r ] , E [ H ∣ r ]) ≥ 0 by conditional association and Lemma 2.6.
Proposition 2.8 (pinning). Pinning any set B B B of spins to s ∗ = e 1 s_\ast=e_1 s ∗ = e 1 , with the same interactions and with bonds sampled by the rule of Lemma 2.2 , cannot decrease the probability of an increasing bond event.
Proof. This is [21 Proposition 4.5] . For 0 < ε < 1 0<\varepsilon<1 0 < ε < 1 let C ε = { σ x 1 ≥ 1 − ε for all x ∈ B } = { r x ≥ 1 − ε , τ x = 1 for all x ∈ B } C_\varepsilon=\{\sigma_x^1\ge1-\varepsilon\text{ for all }x\in B\}=\{r_x\ge1-\varepsilon,\tau_x=1\text{ for all }x\in B\} C ε = { σ x 1 ≥ 1 − ε for all x ∈ B } = { r x ≥ 1 − ε , τ x = 1 for all x ∈ B } . By Lemma 2.2 , P ( C ε ∣ r , θ , η ) = 1 { r x ≥ 1 − ε ∀ x ∈ B } 2 − k B ( η ) = : T ε ( r , η ) \mathbb{P}(C_\varepsilon\mid r,\theta,\eta)=\mathbf{1}_{\{r_x\ge1-\varepsilon\ \forall x\in B\}}2^{-k_B(\eta)}=:T_\varepsilon(r,\eta) P ( C ε ∣ r , θ , η ) = 1 { r x ≥ 1 − ε ∀ x ∈ B } 2 − k B ( η ) =: T ε ( r , η ) , where k B k_B k B is the number of open clusters meeting B B B . Opening a bond cannot increase k B k_B k B , so T ε T_\varepsilon T ε is increasing, and for an increasing bond event A A A Lemma 2.7 gives P ( A ∣ C ε ) = E [ 1 A T ε ] / E [ T ε ] ≥ P ( A ) \mathbb{P}(A\mid C_\varepsilon)=\mathbb{E}[\mathbf{1}_A T_\varepsilon]/\mathbb{E}[T_\varepsilon]\ge\mathbb{P}(A) P ( A ∣ C ε ) = E [ 1 A T ε ] / E [ T ε ] ≥ P ( A ) . As ε ↓ 0 \varepsilon\downarrow0 ε ↓ 0 the law of σ \sigma σ given C ε C_\varepsilon C ε converges weakly to the pinned law, and the bond kernel P ( η e = 1 ∣ σ ) = 1 − exp ( − 2 b e J 1 max { σ x 1 σ y 1 , 0 } ) \mathbb{P}(\eta_e=1\mid\sigma)=1-\exp(-2b_eJ^1\max\{\sigma_x^1\sigma_y^1,0\}) P ( η e = 1 ∣ σ ) = 1 − exp ( − 2 b e J 1 max { σ x 1 σ y 1 , 0 }) is continuous in σ \sigma σ ; so P ( A ∣ C ε ) \mathbb{P}(A\mid C_\varepsilon) P ( A ∣ C ε ) converges to the pinned probability of A A A .
For an integer N ≥ 2 N\ge2 N ≥ 2 and z ∈ Z 2 z\in\mathbb{Z}^2 z ∈ Z 2 let D N ( z ) = { v : N ≤ ∥ v − z ∥ ∞ ≤ 2 N } D_N(z)=\{v:N\le\lVert v-z\rVert_\infty\le2N\} D N ( z ) = { v : N ≤ ∥ v − z ∥ ∞ ≤ 2 N } , with inner layer { ∥ v − z ∥ ∞ = N } \{\lVert v-z\rVert_\infty=N\} {∥ v − z ∥ ∞ = N } and outer layer { ∥ v − z ∥ ∞ = 2 N } \{\lVert v-z\rVert_\infty=2N\} {∥ v − z ∥ ∞ = 2 N } . An annulus graph is a subgraph H H H of the nearest-neighbour graph on D N ( z ) D_N(z) D N ( z ) with strengths b ∈ [ 0 , β ) E ( H ) b\in[0,\beta)^{E(H)} b ∈ [ 0 , β ) E ( H ) ; its pinned measure fixes every vertex of H H H on either layer at s ∗ s_\ast s ∗ and keeps all edge interactions, edges between pinned vertices included, and its bonds are sampled by the rule of Lemma 2.2 . The annulus is crossed if an open path in H H H joins the two layers.
Lemma 2.9 (crossing identity). For an annulus graph and any J ∈ ( 0 , ∞ ) 3 \mathbf{J}\in(0,\infty)^3 J ∈ ( 0 , ∞ ) 3 , let Z J ( 0 ) Z_{\mathbf{J}}(0) Z J ( 0 ) be its partition function with both layers pinned to s ∗ s_\ast s ∗ , and Z J ( π ) Z_{\mathbf{J}}(\pi) Z J ( π ) the same with the inner layer pinned to − s ∗ -s_\ast − s ∗ . Then
P J p i n n e d ( the annulus is crossed ) = 1 − Z J ( π ) Z J ( 0 ) . (5) \mathbb{P}_{\mathbf{J}}^{\mathrm{pinned}}(\text{the annulus is crossed})=1-\frac{Z_{\mathbf{J}}(\pi)}{Z_{\mathbf{J}}(0)}. \tag*{(5)} P J pinned ( the annulus is crossed ) = 1 − Z J ( 0 ) Z J ( π ) . (5)
Proof. This is [21 (4.9)] . Pinned sites have r = 1 r=1 r = 1 and q = 0 q=0 q = 0 for either sign, so their planar factors and all couplings K , L K,L K , L are the same in Z J ( 0 ) Z_{\mathbf{J}}(0) Z J ( 0 ) and Z J ( π ) Z_{\mathbf{J}}(\pi) Z J ( π ) . Expand every Ising factor as in the proof of Lemma 2.2 and sum over the unpinned signs: for each η \eta η the sum equals 2 k 0 ( η ) − ∣ Γ ∣ 2^{k_0(\eta)-|\Gamma|} 2 k 0 ( η ) − ∣Γ∣ times the indicator that every open cluster carries a single prescribed sign, where Γ \Gamma Γ is the set of unpinned vertices and k 0 k_0 k 0 the number of open clusters avoiding the layers. With all pins + + + the indicator is 1 1 1 ; with the inner layer − - − it is the indicator that no open cluster meets both layers. All other factors coincide at every ( r I , θ I ) (r_I,\theta_I) ( r I , θ I ) . If a layer has no vertex, both sides vanish.
The annulus input and the proof of Theorem 1.1
Let ε N ( β ) \varepsilon_N(\beta) ε N ( β ) be the supremum, over z ∈ Z 2 z \in\mathbb{Z}^2 z ∈ Z 2 , annulus graphs H H H on D N ( z ) D_N(z) D N ( z ) and strengths in [ 0 , β ] E ( H ) [0,\beta]^{E(H)} [ 0 , β ] E ( H ) , of the pinned crossing probability of the isotropic model J = ( 1 , 1 , 1 ) \mathbf{J}=(1,1,1) J = ( 1 , 1 , 1 ) . By Lemma 2.9, 1 − ε N ( β ) 1-\varepsilon_N(\beta) 1 − ε N ( β ) is the infimum of Z ( 1 , 1 , 1 ) ( π ) / Z ( 1 , 1 , 1 ) ( 0 ) Z_{(1,1,1)}(\pi)/Z_{(1,1,1)}(0) Z ( 1 , 1 , 1 ) ( π ) / Z ( 1 , 1 , 1 ) ( 0 ) .
Proposition 3.1 (the isotropic annulus input). Let β > 0 \beta>0 β > 0 , β 0 = max { 1 , β / 2 } \beta_0=\max\{1,\beta/2\} β 0 = max { 1 , β /2 } , and let M 0 = M 0 ( β 0 , 3 , 4 − 26 ) M_0=M_0(\beta_0,3,4^{-26}) M 0 = M 0 ( β 0 , 3 , 4 − 26 ) be the constant of [5, Theorem 5.2} for three spin components and their parameter c 2 = 4 − 26 c_2=4^{-26} c 2 = 4 − 26 . Then ε N ( β ) ≤ 4 − 26 \varepsilon_N(\beta)\leq4^{-26} ε N ( β ) ≤ 4 − 26 for every integer N ≥ max { 2 , M 0 } N\geq\max\{2,M_0\} N ≥ max { 2 , M 0 } .
Proof. Translate so that z = 0 z=0 z = 0 , and fix an annulus graph H H H with strengths b b b .
Step 1: the full annulus with positive strengths suffices. Let F F F be the full nearest-neighbour graph on D N D_N D N , and put b ~ e = b e \widetilde b_e=b_e b e = b e on E ( H ) E(H) E ( H ) and b ~ e = 0 \widetilde b_e=0 b e = 0 on E ( F ) ∖ E ( H ) E(F)\setminus E(H) E ( F ) ∖ E ( H ) . In the pinned model on ( F , b ~ ) (F,\widetilde b) ( F , b ) , bonds on edges with b ~ e = 0 \widetilde b_e=0 b e = 0 are closed almost surely, unpinned vertices outside V ( H ) V(H) V ( H ) are independent uniform spins, and pinned ones outside V ( H ) V(H) V ( H ) only contribute constants; so the law of ( σ ∣ V ( H ) , η ∣ E ( H ) ) (\sigma|_{V(H)},\eta|_{E(H)}) ( σ ∣ V ( H ) , η ∣ E ( H ) ) is that of the pinned model on ( H , b ) (H,b) ( H , b ) , and the crossing events agree. The crossing probability is a ratio of integrals over the compact space of unpinned spins of jointly continuous bounded functions of ( b ~ , σ ) (\widetilde b,\sigma) ( b , σ ) (the bond kernel is 1 − exp ( − 2 b ~ e max { σ u x σ v x , 0 } ) 1-\exp(-2\widetilde b_e\max\{\sigma^x_u\sigma^x_v,0\}) 1 − exp ( − 2 b e max { σ u x σ v x , 0 }) ), so it is continuous in b ~ ∈ [ 0 , β ] E ( F ) \widetilde b\in[0,\beta]^{E(F)} b ∈ [ 0 , β ] E ( F ) . It therefore suffices to bound it for b ~ ∈ ( 0 , β ] E ( F ) \widetilde b\in(0,\beta]^{E(F)} b ∈ ( 0 , β ] E ( F ) .
Step 2: dictionary. For b ~ ∈ ( 0 , β ] E ( F ) \widetilde b\in(0,\beta]^{E(F)} b ∈ ( 0 , β ] E ( F ) put J ^ e : = b ~ e / β ∈ ( 0 , 1 ] \widehat J_e:=\widetilde b_e/\beta\in(0,1] J e := b e / β ∈ ( 0 , 1 ] (the couplings of [5 ] ) and β v : = β / 2 ≤ β 0 \beta_v:=\beta/2\leq\beta_0 β v := β /2 ≤ β 0 for every v ∈ D N v\in D_N v ∈ D N . The spin measure [5 (2.3)] on F F F with couplings J ^ e \widehat J_e J e , local time condition ( β v ) (\beta_v) ( β v ) and spin boundary condition e 1 e_1 e 1 on both layers has weight
exp ( ∑ e 2 J ^ e β u β v σ u ⋅ σ v ) = exp ( ∑ e b ~ e σ u ⋅ σ v ) ; \exp\left(\sum_e 2\widehat J_e\sqrt{\beta_u\beta_v}\,\sigma_u\cdot\sigma_v\right)
=\exp\left(\sum_e\widetilde b_e\sigma_u\cdot\sigma_v\right); exp ( e ∑ 2 J e β u β v σ u ⋅ σ v ) = exp ( e ∑ b e σ u ⋅ σ v ) ;
it is our pinned isotropic measure.
Step 3: the loop representation. The spin boundary condition has first coordinate + 1 +1 + 1 and the other coordinates 0 0 0 on both layers. By [5 (2.15)] and the proof of [5, Theorem 1.1, where the three-component version is used, the loop measure coupled to it is ν F , β ξ \nu^{\xi}_{F,\beta} ν F , β ξ with the local time boundary condition ξ \xi ξ that has x x x -local time β / 2 \beta/2 β /2 and y y y - and z z z -local times 0 0 0 on the layers, wired in every copy; since the x x x - and y y y -local times sum to β v \beta_v β v and the z z z -local time vanishes, this is an x y xy x y boundary condition in the sense of [5 §2.2.2] . Under the coupling the field is φ v x = 2 β v σ v 1 \varphi_v^x=\sqrt{2\beta_v}\sigma_v^1 φ v x = 2 β v σ v 1 [5 (1.6), (2.1)] . After conditioning on the z z z -copy the x x x - and y y y -copies form the loop representation of an X Y XY X Y model [5 Remark 3.7] , and [5 Lemma 3.2] gives: conditionally on the signed fields, the cable indicators are independent with
P ( ω ^ u v x = 1 ∣ φ ) = 1 { φ u x φ v x > 0 } ( 1 − e − 2 J ^ u v ∣ φ u x φ v x ∣ ) = 1 { τ u = τ v } ( 1 − e − 2 b ~ u v r u r v ) . \mathbb{P}(\widehat\omega_{uv}^x=1\mid\varphi)=\mathbf{1}\{\varphi_u^x\varphi_v^x>0\}\left(1-e^{-2\widehat J_{uv}|\varphi_u^x\varphi_v^x|}\right)=\mathbf{1}\{\tau_u=\tau_v\}\left(1-e^{-2\widetilde b_{uv}r_ur_v}\right). P ( ω uv x = 1 ∣ φ ) = 1 { φ u x φ v x > 0 } ( 1 − e − 2 J uv ∣ φ u x φ v x ∣ ) = 1 { τ u = τ v } ( 1 − e − 2 b uv r u r v ) .
the bond kernel of Lemma 2.2 at J 1 = 1 J^1=1 J 1 = 1 ; and connections by positive x x x -local time through complete cables are open-bond connections. So the pinned crossing probability is the ν F , β ξ \nu^\xi_{F,\beta} ν F , β ξ probability that the two layers are joined by a path of positive x x x -local time. Such a path contains a sub-path from its last visit to the inner layer to the next visit to the outer layer, and since N ≥ 2 N\geq2 N ≥ 2 this sub-path uses no edge between two layer vertices; so neither the cables between boundary vertices nor the boundary wiring, which [5 ] exclude from connections, matter.
Step 4. With three components, local times at most β 0 \beta_0 β 0 , couplings J ^ ∈ ( 0 , 1 ] E ( F ) \widehat J\in(0,1]^{E(F)} J ∈ ( 0 , 1 ] E ( F ) and the x y xy x y boundary condition ξ \xi ξ on the boundary layers of N 2 , N = D N N_{2,N}=D_N N 2 , N = D N , [5 Theorem 5.2] with M = N ≥ M 0 M=N\geq M_0 M = N ≥ M 0 bounds the probability of an x y xy x y -crossing by 4 − 26 4^{-26} 4 − 26 , and an x x x -crossing is an x y xy x y -crossing. □ \square □
Remark 3.2. (1) The bound of [5 Theorem 5.2] is stated with M 0 M_0 M 0 depending only on β 0 \beta_0 β 0 , the number of components and c 2 c_2 c 2 , for couplings normalized by max e J ^ e ≤ 1 \max_e\widehat J_e\leq1 max e J e ≤ 1 [5 §2.1] ; the applications in [5 ] take J ^ ≡ 1 \widehat J\equiv1 J ≡ 1 . We checked that in [5 §§2–5] the couplings enter through exact representations and through the upper bound 2 J ^ e β u β v ≤ 2 β 0 2\widehat J_e\sqrt{\beta_u\beta_v}\leq2\beta_0 2 J e β u β v ≤ 2 β 0 , and that no lower bound on J ^ e \widehat J_e J e is used; a few displays are written for J ^ ≡ 1 \hat{J} \equiv1 J ^ ≡ 1 , and with J ^ e \hat{J}_e J ^ e restored they hold for all J ^ e ≤ 1 \hat{J}_e \le1 J ^ e ≤ 1 . The cable of an edge e e e is to be read with length 1 / J ^ e 1/\hat{J}_e 1/ J ^ e , the reading under which [5 Lemma 3.2] holds as stated. Step 1 removes the need for zero couplings. We use the statements of [5 ] without re-proving them. (2)
We read the annulus Λ M , 2 M \Lambda_{M,2M} Λ M , 2 M of [5 Theorem 5.2] , which carries the boundary condition on its two boundary layers, as D M D_M D M , as in its application [5 (5.4)] . Under any reading with the pinned layers at sup-radii a ∈ [ N − 1 , N + 1 ] a \in[N-1,N+1] a ∈ [ N − 1 , N + 1 ] and b ′ ∈ [ 2 N , 2 N + 1 ] b' \in[2N,2N+1] b ′ ∈ [ 2 N , 2 N + 1 ] , the arguments below run with D N D_N D N replaced by { a ≤ ∥ v − z ∥ ∞ ≤ b ′ } \{a \le\lVert v-z\rVert_\infty\le b'\} { a ≤ ∥ v − z ∥ ∞ ≤ b ′ } and N ≥ 3 N \ge3 N ≥ 3 , at the price of replacing 82 by 100 in (6) and 124 by 150 in Section 5.
Lemma 3.3 (annulus transfer). If J ∈ [ 1 − δ , 1 + δ ] 3 \mathbf{J} \in[1-\delta,1+\delta]^3 J ∈ [ 1 − δ , 1 + δ ] 3 with δ ≤ 1 2 \delta\le\frac{1}{2} δ ≤ 2 1 , then for every annulus graph on D N ( z ) D_N(z) D N ( z ) , N ≥ 2 N \ge2 N ≥ 2 , with strengths in [ 0 , β ] [0,\beta] [ 0 , β ] ,
P J p i n n e d ( c r o s s i n g ) ≤ ε N ( β ) + 82 β δ N 2 . \mathbb{P}_{\mathbf{J}}^{\mathrm{pinned}}(\mathrm{crossing}) \le\varepsilon_N(\beta)+82\beta\delta N^2. P J pinned ( crossing ) ≤ ε N ( β ) + 82 β δ N 2 .
Proof. For both pin choices the Hamiltonians differ pointwise by
∣ ∑ e b e ∑ i ( J e i − 1 ) σ x i σ y i ∣ ≤ β δ ∑ e ∑ i ∣ σ x i ∣ ∣ σ y i ∣ ≤ β δ ∣ E N ∣ , \left|\sum_e b_e\sum_i(J_e^i-1)\sigma_x^i\sigma_y^i\right|
\le\beta\delta\sum_e\sum_i|\sigma_x^i||\sigma_y^i|
\le\beta\delta|E_N|, e ∑ b e i ∑ ( J e i − 1 ) σ x i σ y i ≤ β δ e ∑ i ∑ ∣ σ x i ∣∣ σ y i ∣ ≤ β δ ∣ E N ∣ ,
by Cauchy–Schwarz, including edges between pinned sites, where E N E_N E N is the edge set of the annulus graph. Hence Z J / Z ( 1 , 1 , 1 ) ∈ [ e − β δ ∣ E N ∣ , e β δ ∣ E N ∣ ] Z_{\mathbf{J}}/Z_{(1,1,1)} \in[e^{-\beta\delta|E_N|},e^{\beta\delta|E_N|}] Z J / Z ( 1 , 1 , 1 ) ∈ [ e − β δ ∣ E N ∣ , e β δ ∣ E N ∣ ] and Z J ( π ) / Z J ( 0 ) ≥ e − 2 β δ ∣ E N ∣ ( 1 − ε N ( β ) ) Z_{\mathbf{J}}(\pi)/Z_{\mathbf{J}}(0) \ge e^{-2\beta\delta|E_N|}(1-\varepsilon_N(\beta)) Z J ( π ) / Z J ( 0 ) ≥ e − 2 β δ ∣ E N ∣ ( 1 − ε N ( β )) . With (5) , 1 − e − x ( 1 − a ) ≤ x + a 1-e^{-x}(1-a) \le x+a 1 − e − x ( 1 − a ) ≤ x + a and ∣ E N ∣ ≤ 2 ∣ D N ∣ ≤ 2 ( 4 N + 1 ) 2 ≤ 41 N 2 |E_N| \le2|D_N| \le2(4N+1)^2 \le41N^2 ∣ E N ∣ ≤ 2∣ D N ∣ ≤ 2 ( 4 N + 1 ) 2 ≤ 41 N 2 , the claim follows.
For N ≥ 2 N \ge2 N ≥ 2 put
δ N ( β ) = min { 1 2 , 4 − 26 82 β N 2 } , p : = 2 ⋅ 4 − 26 = 2 − 51 . (6) \delta_N(\beta)=\min\left\{\frac{1}{2},\frac{4^{-26}}{82\beta N^2}\right\},\qquad p:=2\cdot4^{-26}=2^{-51}. \tag*{(6)} δ N ( β ) = min { 2 1 , 82 β N 2 4 − 26 } , p := 2 ⋅ 4 − 26 = 2 − 51 . (6)
If ε N ( β ) ≤ 4 − 26 \varepsilon_N(\beta)\le4^{-26} ε N ( β ) ≤ 4 − 26 and J ∈ [ 1 − δ N ( β ) , 1 + δ N ( β ) ] 3 \mathbf{J}\in[1-\delta_N(\beta),1+\delta_N(\beta)]^3 J ∈ [ 1 − δ N ( β ) , 1 + δ N ( β ) ] 3 , Lemma 3.3 bounds every pinned crossing probability by p p p .
Lemma 3.4 (independent crossings). Let N ≥ 2 N\ge2 N ≥ 2 with ε N ( β ) ≤ 4 − 26 \varepsilon_N(\beta)\le4^{-26} ε N ( β ) ≤ 4 − 26 , let J ∈ [ 1 − δ N ( β ) , 1 + δ N ( β ) ] 3 \mathbf{J}\in[1-\delta_N(\beta),1+\delta_N(\beta)]^3 J ∈ [ 1 − δ N ( β ) , 1 + δ N ( β ) ] 3 with J 2 ≥ J 3 J^2\ge J^3 J 2 ≥ J 3 , G G G finite and b ∈ [ 0 , β ] E b\in[0,\beta]^E b ∈ [ 0 , β ] E , and let z 1 , … , z k ∈ Z 2 z_1,\ldots,z_k\in\mathbb{Z}^2 z 1 , … , z k ∈ Z 2 have pairwise disjoint boxes { ∥ v − z j ∥ ∞ ≤ 2 N } \{\lVert v-z_j\rVert_\infty\le2N\} {∥ v − z j ∥ ∞ ≤ 2 N } . Under the free law, the probability that every annulus D N ( z j ) D_N(z_j) D N ( z j ) is crossed by an open path of G G G inside D N ( z j ) D_N(z_j) D N ( z j ) is at most p k p^k p k .
Proof. This follows [21 Proposition 4.6] . The intersection of the crossing events is increasing. Pin to s ∗ s_* s ∗ every vertex of G G G on a layer of one of the annuli; by Proposition 2.8 this does not decrease the probability. Let I j I_j I j be the set of vertices of G G G with N < ∥ v − z j ∥ ∞ < 2 N N<\lVert v-z_j\rVert_\infty<2N N < ∥ v − z j ∥ ∞ < 2 N . Along an edge ∥ v − z j ∥ ∞ \lVert v-z_j\rVert_\infty ∥ v − z j ∥ ∞ changes by at most 1, so every edge of G G G with an endpoint in I j I_j I j has its other endpoint in I j I_j I j or on a layer of annulus j j j ; the I j I_j I j are disjoint, and every edge meeting the remaining unpinned vertices has no endpoint in ⋃ j I j \bigcup_j I_j ⋃ j I j . With the pins fixed, the Gibbs density therefore factorizes: the σ ∣ I j \sigma|_{I_j} σ ∣ I j are independent, and σ ∣ I j \sigma|_{I_j} σ ∣ I j has the pinned law of the annulus graph H j = G ∩ D N ( z j ) H_j=G\cap D_N(z_j) H j = G ∩ D N ( z j ) . The crossing of annulus j j j depends only on the bonds of E ( H j ) E(H_j) E ( H j ) , whose conditional law given σ ∣ D N ( z j ) \sigma|_{D_N(z_j)} σ ∣ D N ( z j ) depends only on σ ∣ D N ( z j ) \sigma|_{D_N(z_j)} σ ∣ D N ( z j ) , and the sets E ( H j ) E(H_j) E ( H j ) are disjoint. So the crossing events are independent, each of probability at most p p p .
Proposition 3.5 (connections). Let N ≥ 2 N\ge2 N ≥ 2 with ε N ( β ) ≤ 4 − 26 \varepsilon_N(\beta)\le4^{-26} ε N ( β ) ≤ 4 − 26 , and put A 0 = 1 / ( 4 ( 1 − 2 − 1 / 25 ) ) A_0=1/(4(1-2^{-1/25})) A 0 = 1/ ( 4 ( 1 − 2 − 1/25 )) and m 0 = ( log 2 ) / ( 25 2 N ) m_0=(\log2)/(25\sqrt{2}N) m 0 = ( log 2 ) / ( 25 2 N ) . For all G G G , b ∈ [ 0 , β ] E b\in[0,\beta]^E b ∈ [ 0 , β ] E and J ∈ [ 1 − δ N ( β ) , 1 + δ N ( β ) ] 3 \mathbf{J}\in[1-\delta_N(\beta),1+\delta_N(\beta)]^3 J ∈ [ 1 − δ N ( β ) , 1 + δ N ( β ) ] 3 with J 2 ≥ J 3 J^2\ge J^3 J 2 ≥ J 3 , the bond law satisfies P ( x ↔ y ) ≤ A 0 e − m 0 ∥ x − y ∥ 2 \mathbb{P}(x\leftrightarrow y)\le A_0e^{-m_0\lVert x-y\rVert_2} P ( x ↔ y ) ≤ A 0 e − m 0 ∥ x − y ∥ 2 .
Proof. This follows [21 Proposition 4.7] ; the lattice-animal sum in the proof of [5 Theorem 5.1] has the same shape. Translate so that x = 0 x=0 x = 0 and let d = ∥ y ∥ ∞ d=\|y\|_\infty d = ∥ y ∥ ∞ . If d ≤ 4 N d\leq4N d ≤ 4 N , then ∥ y ∥ 2 ≤ 4 2 N \|y\|_2\leq4\sqrt{2}N ∥ y ∥ 2 ≤ 4 2 N and A 0 e − m 0 ∥ y ∥ 2 ≥ A 0 2 − 4 / 25 ≥ 1 A_0e^{-m_0\|y\|_2}\geq A_0 2^{-4/25}\geq1 A 0 e − m 0 ∥ y ∥ 2 ≥ A 0 2 − 4/25 ≥ 1 . Let d > 4 N d>4N d > 4 N . Partition Z 2 \mathbb{Z}^2 Z 2 into the boxes Q w = w + { 0 , … , N − 1 } 2 Q_w=w+\{0,\ldots,N-1\}^2 Q w = w + { 0 , … , N − 1 } 2 , w ∈ N Z 2 w\in N\mathbb{Z}^2 w ∈ N Z 2 . An open path from 0 0 0 to y y y visits a sequence of boxes, consecutive ones adjacent, and loop erasure gives a simple coarse path w 0 , … , w k w^0,\ldots,w^k w 0 , … , w k of visited boxes with Q w 0 ∋ 0 Q_{w^0}\ni0 Q w 0 ∋ 0 , Q w k ∋ y Q_{w^k}\ni y Q w k ∋ y and k ≥ d / N − 1 k\geq d/N-1 k ≥ d / N − 1 . For each j j j take a visited vertex v ∈ Q w j v\in Q_{w^j} v ∈ Q w j , so ∥ v − w j ∥ ∞ < N \|v-w^j\|_\infty<N ∥ v − w j ∥ ∞ < N ; one of 0 , y 0,y 0 , y is at sup-distance more than 2 N 2N 2 N from w j w^j w j , since otherwise d ≤ 4 N d\leq4N d ≤ 4 N . Along the open path from v v v to that endpoint, ∥ ⋅ − w j ∥ ∞ \|\cdot-w^j\|_\infty ∥ ⋅ − w j ∥ ∞ moves by unit steps from below N N N to above 2 N 2N 2 N , and the part between the last visit of the inner layer before the first visit of the outer layer stays in D N ( w j ) D_N(w^j) D N ( w j ) : the annulus D N ( w j ) D_N(w^j) D N ( w j ) is crossed. One of the 25 residue classes of w j / N w^j/N w j / N modulo 5 contains at least ( k + 1 ) / 25 (k+1)/25 ( k + 1 ) /25 of the distinct w j w^j w j ; two distinct centres in one class differ by at least 5 N 5N 5 N in some coordinate, so their closed 2 N 2N 2 N -boxes are disjoint. Choosing such a class deterministically for each coarse path, Lemma 3.4 bounds the probability that a given coarse path occurs by p ( k + 1 ) / 25 p^{(k+1)/25} p ( k + 1 ) /25 . There are at most 4 k 4^k 4 k simple coarse paths of length k k k from Q w 0 Q_{w^0} Q w 0 . With α = 4 p 1 / 25 = 2 − 1 / 25 \alpha=4p^{1/25}=2^{-1/25} α = 4 p 1/25 = 2 − 1/25 ,
P ( 0 ↔ y ) ≤ ∑ k ≥ d / N − 1 4 k p ( k + 1 ) / 25 = p 1 / 25 ∑ k ≥ d / N − 1 α k ≤ p 1 / 25 α d / N − 1 1 − α = A 0 2 − d / ( 25 N ) ≤ A 0 e − m 0 ∥ y ∥ 2 , \mathbb{P}(0\leftrightarrow y)\leq\sum_{k\geq d/N-1}4^kp^{(k+1)/25}=p^{1/25}\sum_{k\geq d/N-1}\alpha^k\leq\frac{p^{1/25}\alpha^{d/N-1}}{1-\alpha}=A_0 2^{-d/(25N)}\leq A_0e^{-m_0\|y\|_2}, P ( 0 ↔ y ) ≤ k ≥ d / N − 1 ∑ 4 k p ( k + 1 ) /25 = p 1/25 k ≥ d / N − 1 ∑ α k ≤ 1 − α p 1/25 α d / N − 1 = A 0 2 − d / ( 25 N ) ≤ A 0 e − m 0 ∥ y ∥ 2 ,
using p 1 / 25 / α = 1 4 p^{1/25}/\alpha=\frac14 p 1/25 / α = 4 1 and d ≥ ∥ y ∥ 2 / 2 d\geq\|y\|_2/\sqrt{2} d ≥ ∥ y ∥ 2 / 2 .
Proof of Theorem 1.1 . Let N ( β ) = max { 2 , ⌈ M 0 ⌉ } N(\beta)=\max\{2,\lceil M_0\rceil\} N ( β ) = max { 2 , ⌈ M 0 ⌉} with M 0 M_0 M 0 as in Proposition 3.1 , so that ε N ( β ) ( β ) ≤ 4 − 26 \varepsilon_{N(\beta)}(\beta)\leq4^{-26} ε N ( β ) ( β ) ≤ 4 − 26 , and put δ ( β ) = δ N ( β ) ( β ) \delta(\beta)=\delta_{N(\beta)}(\beta) δ ( β ) = δ N ( β ) ( β ) , A = A 0 A=A_0 A = A 0 , m = m 0 m=m_0 m = m 0 . For i = 1 i=1 i = 1 , relabel coordinates 2, 3 so that J 2 ≥ J 3 J^2\geq J^3 J 2 ≥ J 3 ; this preserves the reference measure, the box [ 1 − δ , 1 + δ ] 3 [1-\delta,1+\delta]^3 [ 1 − δ , 1 + δ ] 3 , σ 1 \sigma^1 σ 1 and the bond rule. By (2.2) and Proposition 3.5 , 0 ≤ ⟨ σ x 1 σ y 1 ⟩ ≤ P ( x ↔ y ) ≤ A 0 e − m 0 ∥ x − y ∥ 2 0\leq\langle\sigma_x^1\sigma_y^1\rangle\leq\mathbb{P}(x\leftrightarrow y)\leq A_0e^{-m_0\|x-y\|_2} 0 ≤ ⟨ σ x 1 σ y 1 ⟩ ≤ P ( x ↔ y ) ≤ A 0 e − m 0 ∥ x − y ∥ 2 . For i = 2 , 3 i=2,3 i = 2 , 3 , permute coordinates; the box is permutation invariant. The cap δ ≤ 1 2 \delta\leq\frac12 δ ≤ 2 1 keeps J i > 0 J^i>0 J i > 0 , as Lemmas 2.6–2.7 need.
The only dependence on β \beta β is through N ( β ) N(\beta) N ( β ) , which [5 ] do not make explicit. A shift of the scale by one, as in Remark 3.2(2), changes only constants.
The dependence on beta: proof of Theorem 1.6
Let C β C_\beta C β be the constant of [21 Proposition 3.1] : for the isotropic model, Z ( 1 , 1 , 1 ) ( π ) / Z ( 1 , 1 , 1 ) ( 0 ) ≥ 1 − C β / log N Z_{(1,1,1)}(\pi)/Z_{(1,1,1)}(0)\geq1-C_\beta/\log N Z ( 1 , 1 , 1 ) ( π ) / Z ( 1 , 1 , 1 ) ( 0 ) ≥ 1 − C β / log N for every N ≥ 2 N\geq2 N ≥ 2 , every annulus graph and all strengths in [ 0 , β ] [0,\beta] [ 0 , β ] . By Lemma 2.9, ε N ( β ) ≤ C β / log N \varepsilon_N(\beta)\leq C_\beta/\log N ε N ( β ) ≤ C β / log N .
Proposition 4.1. There is κ < ∞ \kappa<\infty κ < ∞ , depending only on the fixed test profiles of [21 §3] , such that C β ≤ κ max { β , β 2 } C_\beta\leq\kappa\max\{\beta,\beta^2\} C β ≤ κ max { β , β 2 } .
Proof. We follow the proof of [21 Proposition 3.1] and record where β \beta β enters. The Gibbs input is the gradient estimate ∣ ⟨ X ( ( d f ) g ) ⟩ ∣ 2 ≤ β ∥ d f ∥ 2 2 \lvert\langle X((df)g)\rangle\rvert^2\leq\beta\|df\|_2^2 ∣⟨ X (( df ) g )⟩ ∣ 2 ≤ β ∥ df ∥ 2 2 and R ( h , h ~ ) ≤ β ∥ h ∥ 2 2 R(h,\tilde h)\leq\beta\|h\|_2^2 R ( h , h ~ ) ≤ β ∥ h ∥ 2 2 [21 Lemma 2.1] , which is linear in β \beta β ; the remaining factors of β \beta β come from the pointwise bound b e ≤ β b_e\leq\beta b e ≤ β on local coefficients. The rotation estimate [21 Proposition 2.2] states
R ( Q , Q ‾ ) ≤ K r o t ( β ) ∥ T ∥ H S 2 + β ∑ i , j min { ∥ f i ∥ ∞ 2 ∥ d f j ∥ 2 2 , ∥ f j ∥ ∞ 2 ∥ d f i ∥ 2 2 } , K r o t ( β ) = 6 β 2 + 4 β 3 / 2 + β . R(Q,\overline{Q})\leq K_{\mathrm{rot}}(\beta)\|T\|_{HS}^2+\beta\sum_{i,j}\min\{\|f_i\|_\infty^2\|df_j\|_2^2,\|f_j\|_\infty^2\|df_i\|_2^2\},\qquad K_{\mathrm{rot}}(\beta)=6\beta^2+4\beta^{3/2}+\beta. R ( Q , Q ) ≤ K rot ( β ) ∥ T ∥ H S 2 + β i , j ∑ min { ∥ f i ∥ ∞ 2 ∥ d f j ∥ 2 2 , ∥ f j ∥ ∞ 2 ∥ d f i ∥ 2 2 } , K rot ( β ) = 6 β 2 + 4 β 3/2 + β .
For the test family of [21 Lemma 3.2] , whose constants depend only on the fixed profiles, Q = − i λ N h ~ Q=-i\lambda_N\tilde h Q = − i λ N h ~ with λ N ≥ 1 2 ∑ r ≤ N r − 1 ≥ 1 2 log N \lambda_N\geq\frac12\sum_{r\leq N}r^{-1}\geq\frac12\log N λ N ≥ 2 1 ∑ r ≤ N r − 1 ≥ 2 1 log N , and the bracket in [21 (3.10)] is at most C ( 1 + log N ) C(1+\log N) C ( 1 + log N ) . Hence, uniformly in the twist, R t ( h ~ , h ~ ) ≤ 4 C K r o t ( β ) ( 1 + log N ) / ( log N ) 2 R_t(\tilde h,\tilde h)\leq4CK_{\mathrm{rot}}(\beta)(1+\log N)/(\log N)^2 R t ( h ~ , h ~ ) ≤ 4 C K rot ( β ) ( 1 + log N ) / ( log N ) 2 , which is the bound [21 (3.16)] . with an explicit β \beta β -dependence. The transfer from h ~ \widetilde h h to h h h [21 (3.17)–(3.19)] adds terms bounded by C ( β 3 / 2 / N + β 2 / N 2 + β / N 2 ) C(\beta^{3/2}/N+\beta^2/N^2+\beta/N^2) C ( β 3/2 / N + β 2 / N 2 + β / N 2 ) , and the integration step that ends the proof is free of β \beta β . Since log N / N ≤ 1 \log N/N\le1 log N / N ≤ 1 and β 3 / 2 ≤ max { β , β 2 } \beta^{3/2}\le\max\{\beta,\beta^2\} β 3/2 ≤ max { β , β 2 } , this bounds the quantity estimated in [21 (3.20)] by κ ′ max { β , β 2 } / log N \kappa'\max\{\beta,\beta^2\}/\log N κ ′ max { β , β 2 } / log N , and C β = κ ′ 2 max { β , β 2 } C_\beta=\frac{\kappa'}{2}\max\{\beta,\beta^2\} C β = 2 κ ′ max { β , β 2 } works.
Proof of Theorem 1.6. Take N = max { 2 , ⌈ exp ( 4 26 C β ) ⌉ } N=\max\{2,\lceil\exp(4^{26}C_\beta)\rceil\} N = max { 2 , ⌈ exp ( 4 26 C β )⌉} , so that ε N ( β ) ≤ C β / log N ≤ 4 − 26 \varepsilon_N(\beta)\le C_\beta/\log N\le4^{-26} ε N ( β ) ≤ C β / log N ≤ 4 − 26 . The proof of Theorem 1.1 then runs with this N N N in place of N ( β ) N(\beta) N ( β ) and gives Theorem 1.1 with δ ( β ) = δ N ( β ) \delta(\beta)=\delta_N(\beta) δ ( β ) = δ N ( β ) . By Proposition 4.1 and log N ≤ 4 26 C β + log 2 \log N\le4^{26}C_\beta+\log2 log N ≤ 4 26 C β + log 2 ,
log 1 δ ( β ) ≤ max { log 2 , 2 ⋅ 4 26 κ max { β , β 2 } + log ( 82 ⋅ 4 26 β ) + 2 log 2 } ≤ C ( 1 + β 2 ) . \log\frac{1}{\delta(\beta)}\le\max\left\{\log2,\;2\cdot4^{26}\kappa\max\{\beta,\beta^2\}+\log(82\cdot4^{26}\beta)+2\log2\right\}\le C(1+\beta^2). log δ ( β ) 1 ≤ max { log 2 , 2 ⋅ 4 26 κ max { β , β 2 } + log ( 82 ⋅ 4 26 β ) + 2 log 2 } ≤ C ( 1 + β 2 ) .
Given λ \lambda λ with ∣ 1 − λ ∣ ≤ e − 2 C |1-\lambda|\le e^{-2C} ∣1 − λ ∣ ≤ e − 2 C , put L = log ( 1 / ∣ 1 − λ ∣ ) ≥ 2 C L=\log(1/|1-\lambda|)\ge2C L = log ( 1/∣1 − λ ∣ ) ≥ 2 C and β = ( L / C − 1 ) 1 / 2 \beta=(L/C-1)^{1/2} β = ( L / C − 1 ) 1/2 , so that C ( 1 + β 2 ) = L C(1+\beta^2)=L C ( 1 + β 2 ) = L and β ≥ ( L / ( 2 C ) ) 1 / 2 \beta\ge(L/(2C))^{1/2} β ≥ ( L / ( 2 C ) ) 1/2 . Then ∣ 1 − λ ∣ ≤ δ ( β ) |1-\lambda|\le\delta(\beta) ∣1 − λ ∣ ≤ δ ( β ) , and Corollary 1.2 applies at every β ′ ≤ ( 2 C ) − 1 / 2 ∣ log ∣ 1 − λ ∣ ∣ 1 / 2 \beta'\le(2C)^{-1/2}|\log|1-\lambda||^{1/2} β ′ ≤ ( 2 C ) − 1/2 ∣ log ∣1 − λ ∣ ∣ 1/2 .
Remark 4.2. The exponent 1 2 \frac{1}{2} 2 1 is lost in the β 2 \beta^2 β 2 part of K r o t K_{\mathrm{rot}} K rot , which comes from four terms in which a local coefficient of size β \beta β meets a current variance of size β \beta β and two products of local coefficients; the physical prediction corresponds to C β C_\beta C β of order β \beta β .
Extensions: proof of Theorem 1.3
Proof of Theorem 1.3(a). Choose the sign component i i i and order the other two coordinates ( j , k ) (j,k) ( j , k ) by π \pi π , so that J j ≥ J k J^j\ge J^k J j ≥ J k and c x , j ≥ c x , k c_{x,j}\ge c_{x,k} c x , j ≥ c x , k . The single-ion terms become c x , i r x 2 + 1 2 q x 2 ( c x , j + c x , k ) + M x cos ( 2 θ x ) c_{x,i}r_x^2+\frac{1}{2}q_x^2(c_{x,j}+c_{x,k})+M_x\cos(2\theta_x) c x , i r x 2 + 2 1 q x 2 ( c x , j + c x , k ) + M x cos ( 2 θ x ) with M x = 1 2 q x 2 ( c x , j − c x , k ) ≥ 0 M_x=\frac{1}{2}q_x^2(c_{x,j}-c_{x,k})\ge0 M x = 2 1 q x 2 ( c x , j − c x , k ) ≥ 0 . The first two are functions of r x r_x r x alone: they multiply ρ \rho ρ by one-site factors, which do not change mixed derivatives of log ρ \log\rho log ρ , and they do not involve τ \tau τ . The planar system now has the edge energies (3) and the site terms M x cos ( 2 θ x ) M_x\cos(2\theta_x) M x cos ( 2 θ x ) , covered by Lemma 2.5. In the proof of Lemma 2.6, every coefficient of the planar system (L e L_e L e and M x M_x M x ) is nonincreasing in the relevant amplitude and M x M_x M x has no mixed derivative, so all terms remain nonnegative. Pinning, the crossing identity and the separation of annuli are unaffected by one-site terms. In Lemma 3.3 the Hamiltonian changes by at most β δ ′ ( ∣ E N ∣ + ∣ V N ∣ ) ≤ 62 β δ ′ N 2 \beta\delta'(|E_N|+|V_N|)\le62\beta\delta'N^2 β δ ′ ( ∣ E N ∣ + ∣ V N ∣ ) ≤ 62 β δ ′ N 2 , since ∣ ∑ i c x , i ( σ x i ) 2 ∣ ≤ max i ∣ c x , i ∣ ≤ β δ ′ \left|\sum_i c_{x,i}(\sigma_x^i)^2\right|\le\max_i|c_{x,i}|\le\beta\delta' ∑ i c x , i ( σ x i ) 2 ≤ max i ∣ c x , i ∣ ≤ β δ ′ and ∣ V N ∣ ≤ ( 4 N + 1 ) 2 ≤ 21 N 2 |V_N|\le(4N+1)^2\le21N^2 ∣ V N ∣ ≤ ( 4 N + 1 ) 2 ≤ 21 N 2 ; so the argument of Proposition 3.5 runs with
δ ′ ( β ) = min { 1 2 , 4 − 26 124 β N 2 } ≤ δ ( β ) , \delta'(\beta)=\min\left\{\frac{1}{2},\frac{4^{-26}}{124\beta N^2}\right\}\le\delta(\beta), δ ′ ( β ) = min { 2 1 , 124 β N 2 4 − 26 } ≤ δ ( β ) ,
for the same N = N ( β ) N=N(\beta) N = N ( β ) , and gives the same A 0 , m 0 A_0,m_0 A 0 , m 0 . For H ~ ε \widetilde H_\varepsilon H ε of [11 Definition 1.8(4)] , with weight exp ( − β ′ H ~ ε ) \exp(-\beta'\widetilde H_\varepsilon) exp ( − β ′ H ε ) at β ′ ≤ β \beta'\le\beta β ′ ≤ β , we have J = ( 1 , 1 , 1 ) \mathbf J=(1,1,1) J = ( 1 , 1 , 1 ) and c x , 3 = − β ′ ε c_{x,3}=-\beta'\varepsilon c x , 3 = − β ′ ε , c x , 1 = c x , 2 = 0 c_{x,1}=c_{x,2}=0 c x , 1 = c x , 2 = 0 , which is aligned for either sign of ε \varepsilon ε , with ∣ c x , 3 ∣ ≤ β δ ′ |c_{x,3}|\le\beta\delta' ∣ c x , 3 ∣ ≤ β δ ′ when ∣ ε ∣ ≤ δ ′ ( β ) |\varepsilon|\le\delta'(\beta) ∣ ε ∣ ≤ δ ′ ( β ) .
Remark 5.1. The proof of part (a) for a fixed component i i i uses only that, for the complementary pair ( j , k ) (j,k) ( j , k ) , J j − J k J^j-J^k J j − J k and all c x , j − c x , k c_{x,j}-c_{x,k} c x , j − c x , k have one weak sign; asking this for all three i i i is equivalent to alignment. For example, if λ < 1 \lambda<1 λ < 1 is combined with an easy-axis single-ion term, the bound for ⟨ σ x 3 σ y 3 ⟩ \langle\sigma_x^3\sigma_y^3\rangle ⟨ σ x 3 σ y 3 ⟩ still holds, but for i = 1 , 2 i=1,2 i = 1 , 2 the terms cos ( θ x + θ y ) \cos(\theta_x+\theta_y) cos ( θ x + θ y ) and cos ( 2 θ x ) \cos(2\theta_x) cos ( 2 θ x ) enter with opposite signs. Then the covariance inequality of Lemma 2.5 fails, and the amplitude law itself need not be log-supermodular: on a path with three sites, with λ = 0.98 \lambda=0.98 λ = 0.98 , c x , 3 = 0.08 c_{x,3}=0.08 c x , 3 = 0.08 at every site, b e = 4 b_e=4 b e = 4 and r = ( 0.8 , 0.9 , 0.5 ) \mathbf r=(0.8,0.9,0.5) r = ( 0.8 , 0.9 , 0.5 ) , numerical quadrature gives ∂ r 1 ∂ r 3 log ρ ≈ − 1.2 × 10 − 5 \partial_{r_1}\partial_{r_3}\log\rho\approx-1.2\times10^{-5} ∂ r 1 ∂ r 3 log ρ ≈ − 1.2 × 1 0 − 5 , while the aligned control (c x , 3 = − 0.08 c_{x,3}=-0.08 c x , 3 = − 0.08 ) gives ≈ + 3.9 × 10 − 4 \approx+3.9\times10^{-4} ≈ + 3.9 × 1 0 − 4 . So this route does not extend to such cases.
Lemma 5.2 (conditional projection). Let 2 ≤ k < n 2 \le k < n 2 ≤ k < n and write σ ∈ S n − 1 \sigma\in S^{n-1} σ ∈ S n − 1 as σ = ( ρ u , v ) \sigma= (\rho u,v) σ = ( ρ u , v ) with u ∈ S k − 1 u \in S^{k-1} u ∈ S k − 1 , v ∈ R n − k v \in\mathbb{R}^{n-k} v ∈ R n − k and ρ = ( 1 − ∣ v ∣ 2 ) 1 / 2 \rho= (1-|v|^2)^{1/2} ρ = ( 1 − ∣ v ∣ 2 ) 1/2 . Under the uniform measure on S n − 1 S^{n-1} S n − 1 , ρ > 0 \rho> 0 ρ > 0 almost surely, and u u u is uniform on S k − 1 S^{k-1} S k − 1 and independent of v v v .
Proof. This is the block form [21 Lemma 4.8] of [1 Proposition 2.4] . Let U ∈ R k U \in\mathbb{R}^k U ∈ R k and W ∈ R n − k W \in\mathbb{R}^{n-k} W ∈ R n − k be independent standard Gaussian vectors; ( U , W ) / ∣ ( U , W ) ∣ (U,W)/|(U,W)| ( U , W ) /∣ ( U , W ) ∣ is uniform on S n − 1 S^{n-1} S n − 1 . By rotation invariance U / ∣ U ∣ U/|U| U /∣ U ∣ is uniform on S k − 1 S^{k-1} S k − 1 and independent of ( ∣ U ∣ , W ) (|U|,W) ( ∣ U ∣ , W ) , while v = W / ∣ ( U , W ) ∣ v = W/|(U,W)| v = W /∣ ( U , W ) ∣ and ρ = ∣ U ∣ / ∣ ( U , W ) ∣ \rho= |U|/|(U,W)| ρ = ∣ U ∣/∣ ( U , W ) ∣ are functions of ( ∣ U ∣ , W ) (|U|,W) ( ∣ U ∣ , W ) .
Proof of Theorem 1.3 (b). Permute coordinates so that the component of interest is the first, and apply Lemma 5.2 with k = 3 k=3 k = 3 at each site. The interaction is ∑ e b e [ ∑ i ≤ 3 J i ρ x ρ y u x i u y i + ∑ i > 3 J i v x i v y i ] \sum_e b_e[\sum_{i\le3}J^i\rho_x\rho_yu_x^iu_y^i+\sum_{i>3}J^iv_x^iv_y^i] ∑ e b e [ ∑ i ≤ 3 J i ρ x ρ y u x i u y i + ∑ i > 3 J i v x i v y i ] ; given v v v the second sum is constant, so the law of u u u is the three-component model with anisotropy ( J 1 , J 2 , J 3 ) (J^1,J^2,J^3) ( J 1 , J 2 , J 3 ) and strengths b e ρ x ρ y ∈ [ 0 , β ] b_e\rho_x\rho_y \in[0,\beta] b e ρ x ρ y ∈ [ 0 , β ] . Theorem 1.1 gives 0 ≤ ⟨ u x 1 u y 1 ⟩ v ≤ A 0 e − m 0 ∥ x − y ∥ 2 0 \le\langle u_x^1u_y^1\rangle_v \le A_0e^{-m_0\|x-y\|^2} 0 ≤ ⟨ u x 1 u y 1 ⟩ v ≤ A 0 e − m 0 ∥ x − y ∥ 2 , and ⟨ σ x 1 σ y 1 ⟩ = E [ ρ x ρ y ⟨ u x 1 u y 1 ⟩ v ] \langle\sigma_x^1\sigma_y^1\rangle= \mathbb{E}[\rho_x\rho_y\langle u_x^1u_y^1\rangle_v] ⟨ σ x 1 σ y 1 ⟩ = E [ ρ x ρ y ⟨ u x 1 u y 1 ⟩ v ] .
Remark 5.3. At J = ( 1 , … , 1 ) \mathbf{J}=(1,\ldots,1) J = ( 1 , … , 1 ) the isotropic law is invariant under permutations of coordinates, so Theorem 1.3 (b) gives 0 ≤ ⟨ σ x ⋅ σ y ⟩ G , b ( n ) = n ⟨ σ x 1 σ y 1 ⟩ ≤ n A 0 e − m 0 ∥ x − y ∥ 2 0 \le\langle\sigma_x\cdot\sigma_y\rangle_{G,b}^{(n)} = n\langle\sigma_x^1\sigma_y^1\rangle\le nA_0e^{-m_0\|x-y\|^2} 0 ≤ ⟨ σ x ⋅ σ y ⟩ G , b ( n ) = n ⟨ σ x 1 σ y 1 ⟩ ≤ n A 0 e − m 0 ∥ x − y ∥ 2 for every n ≥ 3 n \ge3 n ≥ 3 , finite G G G and b ∈ [ 0 , β ] E b \in[0,\beta]^E b ∈ [ 0 , β ] E , with m 0 = m 0 ( β ) = ( log 2 ) / ( 25 2 N ( β ) ) m_0=m_0(\beta)=(\log2)/(25\sqrt{2N(\beta)}) m 0 = m 0 ( β ) = ( log 2 ) / ( 25 2 N ( β ) ) . This is the bound (2) with A = n A 0 A=nA_0 A = n A 0 and m = m 0 m=m_0 m = m 0 , obtained from [5 Theorem 5.2] through Proposition 3.1 .
The corollaries
Proof of Corollary 1.2. The λ \lambda λ -model at β ′ ≤ β \beta' \le\beta β ′ ≤ β has b ≡ β ′ ∈ [ 0 , β ] b \equiv\beta' \in[0,\beta] b ≡ β ′ ∈ [ 0 , β ] and J = ( 1 , 1 , λ ) \mathbf{J}=(1,1,\lambda) J = ( 1 , 1 , λ ) , so Theorem 1.1 applies on every free box when ∣ 1 − λ ∣ ≤ δ ( β ) |1-\lambda| \le\delta(\beta) ∣1 − λ ∣ ≤ δ ( β ) . The bound is uniform in the box, so it holds for the limit superior along boxes and, since σ ↦ σ x i σ x i \sigma\mapsto\sigma_x^i\sigma_x^i σ ↦ σ x i σ x i is a bounded continuous local function, for every subsequential local weak limit. If β ˉ ( λ ) \bar{\beta}(\lambda) β ˉ ( λ ) is any threshold as in [11 Theorem 1.2] , which concerns free or Dirichlet limits, the free-boundary part of its conclusion is incompatible with exponential decay, so β ˉ ( λ ) > β \bar{\beta}(\lambda)>\beta β ˉ ( λ ) > β whenever 0 < 1 − λ ≤ δ ( β ) 0<1-\lambda\le\delta(\beta) 0 < 1 − λ ≤ δ ( β ) . For λ > 1 \lambda>1 λ > 1 , exponential decay of ⟨ σ x 3 σ y 3 ⟩ \langle\sigma_x^3\sigma_y^3\rangle ⟨ σ x 3 σ y 3 ⟩ in free-boundary limits excludes long-range order there.
Proof of Corollary 1.4 . By [11 Proposition A.1] , with free boundary conditions, λ ↦ ⟨ σ x 1 σ y 1 ⟩ β , λ \lambda\mapsto\langle\sigma_x^1\sigma_y^1\rangle_{\beta,\lambda} λ ↦ ⟨ σ x 1 σ y 1 ⟩ β , λ is nonincreasing and λ ↦ ⟨ σ x 3 σ y 3 ⟩ β , λ \lambda\mapsto\langle\sigma_x^3\sigma_y^3\rangle_{\beta,\lambda} λ ↦ ⟨ σ x 3 σ y 3 ⟩ β , λ is nondecreasing on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) ; this rests on Dunlop’s inequality [6 Theorem 1] . At λ = 1 \lambda=1 λ = 1 , ⟨ σ x i σ y i ⟩ = 1 3 ⟨ σ x ⋅ σ y ⟩ \langle\sigma_x^i\sigma_y^i\rangle=\frac{1}{3}\langle\sigma_x\cdot\sigma_y\rangle ⟨ σ x i σ y i ⟩ = 3 1 ⟨ σ x ⋅ σ y ⟩ , which (2) bounds with b ≡ β b\equiv\beta b ≡ β ; by Remark 5.3 one may take A = 3 A 0 A=3A_0 A = 3 A 0 , m = m 0 m=m_0 m = m 0 .
The physical rate for $n \ge 4: proof of Theorem 1.5
This section adapts the proof of decay for charged observables of the isotropic O ( 4 ) O(4) O ( 4 ) model in [22 §§3.9–3.10] to anisotropic couplings. The mechanism there is kept: coarse orientation variables, a test for bad coarse vertices, a disagreement exploration following van den Berg and Maes [2 ] , and averaging over a symmetry. New here are the use of the isotropic bad-site estimate only as an event bound at one scale (Lemmas 7.2 –7.3), the replacement of the averaging over common left rotations by S 3 S^3 S 3 with averaging over { ± 1 } \{\pm1\} { ± 1 } (Lemmas 7.4–7.5 ), and the bookkeeping of the threshold (§ 7.5 ). Throughout, ω n \omega_n ω n is the uniform probability measure on S n − 1 S^{n-1} S n − 1 .
Systems and the result for n = 4 n = 4 n = 4
A system S = ( V , E , P , σ ˉ , M ) \mathcal{S} = (V, E, P, \bar{\sigma}, M) S = ( V , E , P , σ ˉ , M ) consists of a finite V ⊂ Z 2 V \subset\mathbb{Z}^2 V ⊂ Z 2 , a set E E E of nearest-neighbour edges with both endpoints in V V V , each with a fixed orientation e = ( x → y ) e = (x \to y) e = ( x → y ) , a set P ⊂ V P \subset V P ⊂ V of pinned sites with values σ ˉ ∈ ( S n − 1 ) P \bar{\sigma} \in(S^{n-1})^P σ ˉ ∈ ( S n − 1 ) P , and real n × n n \times n n × n matrices M e M_e M e , e ∈ E e \in E e ∈ E . Its Gibbs measure is the probability measure on ( S n − 1 ) V ∖ P (S^{n-1})^{V\setminus P} ( S n − 1 ) V ∖ P given by
⟨ ⋅ ⟩ S ∝ exp ( ∑ e = ( x → y ) ∈ E σ x T M e σ y ) ∏ z ∈ V ∖ P ω n ( d σ z ) , σ p : = σ ˉ p ( p ∈ P ) . \langle\cdot\rangle_{\mathcal{S}} \propto\exp\left(\sum_{e=(x\to y)\in E}\sigma_x^{\mathsf{T}}M_e\sigma_y\right)\prod_{z\in V\setminus P}\omega_n(d\sigma_z), \qquad\sigma_p := \bar{\sigma}_p\ (p\in P). ⟨ ⋅ ⟩ S ∝ exp e = ( x → y ) ∈ E ∑ σ x T M e σ y z ∈ V ∖ P ∏ ω n ( d σ z ) , σ p := σ ˉ p ( p ∈ P ) .
The system is ( β , θ ) (\beta,\theta) ( β , θ ) -admissible if for every e e e there is b e ∈ [ − β , β ] b_e \in[-\beta,\beta] b e ∈ [ − β , β ] with ∥ M e − b e I ∥ o p ≤ θ \lVert M_e-b_eI\rVert_{\mathrm{op}}\leq\theta ∥ M e − b e I ∥ op ≤ θ . A torus system is defined in the same way on ( Z / L 1 Z ) × ( Z / L 2 Z ) (\mathbb{Z}/L_1\mathbb{Z})\times(\mathbb{Z}/L_2\mathbb{Z}) ( Z / L 1 Z ) × ( Z / L 2 Z ) with all nearest-neighbour edges and P = ∅ P=\varnothing P = ∅ ; d T d_T d T is the cyclic maximum-norm distance. An observable on A A A is a bounded measurable function O O O of σ A \sigma_A σ A ; it is odd if O ( − σ A ) = − O ( σ A ) O(-\sigma_A)=-O(\sigma_A) O ( − σ A ) = − O ( σ A ) . The systems of Theorem 1.5 have M e = diag ( m e 1 , … , m e n ) M_e=\operatorname{diag}(m_e^1,\ldots,m_e^n) M e = diag ( m e 1 , … , m e n ) with ∣ m e k − b e ∣ ≤ β δ \lvert m_e^k-b_e\rvert\leq\beta\delta ∣ m e k − b e ∣ ≤ β δ , so they are ( β , β δ ) (\beta,\beta\delta) ( β , β δ ) -admissible.
Theorem 7.1. Assume (I1)–(I3) of § 7.2 . There are absolute constants c 1 > 0 c_1>0 c 1 > 0 and A ∗ , C Ξ < ∞ A_*,C_{\Xi}<\infty A ∗ , C Ξ < ∞ such that, with Ξ ( β ) \Xi(\beta) Ξ ( β ) as in Theorem 1.5 , the following hold whenever β > 0 \beta>0 β > 0 , θ ≥ 0 \theta\geq0 θ ≥ 0 and θ Ξ ( β ) 2 ≤ c 1 \theta\Xi(\beta)^2\leq c_1 θ Ξ ( β ) 2 ≤ c 1 . For every ( β , θ ) (\beta,\theta) ( β , θ ) -admissible system with n = 4 n=4 n = 4 , every odd observable O O O and every observable O ′ O' O ′ on S 3 S^3 S 3 :
(a) ∣ ⟨ O ( σ x ) ⟩ ∣ ≤ A ∗ ∥ O ∥ ∞ exp ( − dist ∞ ( x , P ) / Ξ ( β ) ) \lvert\langle O(\sigma_x)\rangle\rvert\leq A_*\lVert O\rVert_\infty\exp(-\operatorname{dist}_\infty(x,P)/\Xi(\beta)) ∣⟨ O ( σ x )⟩∣ ≤ A ∗ ∥ O ∥ ∞ exp ( − dist ∞ ( x , P ) /Ξ ( β )) for x ∈ V ∖ P x\in V\setminus P x ∈ V ∖ P ; the left side vanishes if P = ∅ P=\varnothing P = ∅ ;
(b) for x , y ∈ V ∖ P x,y\in V\setminus P x , y ∈ V ∖ P ,
∣ ⟨ O ( σ x ) O ′ ( σ y ) ⟩ ∣ ≤ A ∗ ∥ O ∥ ∞ ∥ O ′ ∥ ∞ exp ( − min { ∥ x − y ∥ ∞ , dist ∞ ( x , P ) } / Ξ ( β ) ) ; \lvert\langle O(\sigma_x)O'(\sigma_y)\rangle\rvert\leq A_*\lVert O\rVert_\infty\lVert O'\rVert_\infty\exp\left(-\min\{\lVert x-y\rVert_\infty,\operatorname{dist}_\infty(x,P)\}/\Xi(\beta)\right); ∣⟨ O ( σ x ) O ′ ( σ y )⟩∣ ≤ A ∗ ∥ O ∥ ∞ ∥ O ′ ∥ ∞ exp ( − min {∥ x − y ∥ ∞ , dist ∞ ( x , P )} /Ξ ( β ) ) ; (c) on a torus system with ( β , θ ) (\beta,\theta) ( β , θ ) -admissible couplings, the bound of (b) holds with min { ∥ x − y ∥ ∞ , dist ∞ ( x , P ) } \min\{\lVert x-y\rVert_\infty,\operatorname{dist}_\infty(x,P)\} min {∥ x − y ∥ ∞ , dist ∞ ( x , P )} replaced by
r T ( x , y ) : = min { d T ( x , y ) , ⌊ min { L 1 , L 2 } − 1 2 ⌋ } . r_T(x,y):=\min\left\{d_T(x,y),\left\lfloor\frac{\min\{L_1,L_2\}-1}{2}\right\rfloor\right\}. r T ( x , y ) := min { d T ( x , y ) , ⌊ 2 min { L 1 , L 2 } − 1 ⌋ } .
The coarse structure of the O ( 4 ) O(4) O ( 4 ) proof and the inputs
Identify R 4 \mathbb{R}^4 R 4 with the quaternions H \mathbb{H} H by sending the standard basis to ( 1 , i , j , k ) (1,i,j,k) ( 1 , i , j , k ) . Then S 3 S^3 S 3 is the group of unit quaternions, ω 4 \omega_4 ω 4 is its Haar measure, σ ⋅ σ ′ = Re ( σ ′ σ ˉ ) \sigma\cdot\sigma'=\operatorname{Re}(\sigma'\bar{\sigma}) σ ⋅ σ ′ = Re ( σ ′ σ ˉ ) for unit σ , σ ′ \sigma,\sigma' σ , σ ′ , and left multiplication by the central element − 1 -1 − 1 is σ ↦ − σ \sigma\mapsto-\sigma σ ↦ − σ .
Fix n = 4 n=4 n = 4 , β > 0 \beta>0 β > 0 , θ ≥ 0 \theta\geq0 θ ≥ 0 and a ( β , θ ) (\beta,\theta) ( β , θ ) -admissible system on a pinned square : V = Q ( S ) = { 0 , … , S } 2 V=Q(S)=\{0,\ldots,S\}^2 V = Q ( S ) = { 0 , … , S } 2 , E E E all nearest-neighbour edges of Q ( S ) Q(S) Q ( S ) , P = ∂ Q ( S ) P=\partial Q(S) P = ∂ Q ( S ) (the sites with a coordinate in { 0 , S } \{0,S\} { 0 , S } ) and arbitrary σ ˉ \bar{\sigma} σ ˉ ; general systems are reduced to this case in § 7.6 . Its reference isotropic system has the same ( V , E , P , σ ˉ ) (V,E,P,\bar{\sigma}) ( V , E , P , σ ˉ ) and M e M_e M e replaced by b e I b_eI b e I , that is, edge energies b e σ x ⋅ σ y = b e Re ( σ y σ ˉ x ) b_e\sigma_x\cdot\sigma_y=b_e\operatorname{Re}(\sigma_y\bar{\sigma}_x) b e σ x ⋅ σ y = b e Re ( σ y σ ˉ x ) . This is the class of [22 §3.1] with untwisted bonds (U e = 1 U_e=1 U e = 1 ) and signed coefficients ∣ b e ∣ ≤ b : = β \lvert b_e\rvert\leq b:=\beta ∣ b e ∣ ≤ b := β . The arguments of [22 §3] depend on fixed choices (exponents D D D , C 0 C_0 C 0 , and the collars and profile families of [22 §§3.9–3.10] ), and its constants depend only on them, not on the lattice size, the couplings or the pins [22 p. 13] . We keep these choices. For B B B above the threshold of [22 §3] we use:
(I1) [22 (3.54)] : for ∣ b e ∣ ≤ b ≤ B \lvert b_e\rvert\leq b\leq B ∣ b e ∣ ≤ b ≤ B there is an integer R ∗ R_* R ∗ , depending only on b b b , B B B and the fixed choices, such that pinned squares of side R ∗ R_* R ∗ in the class have constant-field stiffness H R ∗ ≤ B − D I H_{R_*}\leq B^{-D}I H R ∗ ≤ B − D I and log R ∗ ≤ π b + C D B log B \log R_*\leq\pi b+C_D\sqrt{B}\log B log R ∗ ≤ π b + C D B log B , where C D C_D C D depends only on the fixed choices. The coarse scale is s : = R ∗ M ∗ s:=R_*M_* s := R ∗ M ∗ with M ∗ = ⌈ B D + C 0 ⌉ M_*=\lceil B^{D+C_0}\rceil M ∗ = ⌈ B D + C 0 ⌉ [22 §3.9] . Only the bound on log R ∗ \log R_* log R ∗ is used directly below (Lemma 7.8 ); the stiffness bound enters through (13).
Fixed choices. Let ε ∗ : = 1 / 32 \varepsilon_* := 1/32 ε ∗ := 1/32 and q : = 1 − e − ε ∗ ≤ 1 / 32 q := 1-e^{-\varepsilon_*} \le1/32 q := 1 − e − ε ∗ ≤ 1/32 . For an interior coarse vertex X X X (below), coll ( X ) \operatorname{coll}(X) coll ( X ) is the fixed collar of [22 §3.9] : a set of at most K c o l l K_{\mathrm{coll}} K coll coarse cells containing the four cells at X X X (the star of X X X ), each corner of which is at grid distance at most r c o l l r_{\mathrm{coll}} r coll from X X X . We read it as a block of cells without holes, so that the profiles of (I2), which are the identity on and outside the outer collar boundary, are the identity outside the set I ( X ) I(X) I ( X ) defined below; otherwise replace coll ( X ) \operatorname{coll}(X) coll ( X ) by the set of cells whose rings meet the support of those profiles, and nothing below changes. Put C c o l l : = ( 2 r c o l l + 1 ) 2 C_{\mathrm{coll}} := (2r_{\mathrm{coll}}+1)^2 C coll := ( 2 r coll + 1 ) 2 and w : = r c o l l + 2 w := r_{\mathrm{coll}}+2 w := r coll + 2 . Choose B 1 ≥ e B_1 \ge e B 1 ≥ e at least the threshold of [22 ] and so large that ζ 1 / 128 ( B ) ≤ 1 2 ⋅ 32 − C c o l l \zeta_{1/128}(B) \le\frac{1}{2}\cdot32^{-C_{\mathrm{coll}}} ζ 1/128 ( B ) ≤ 2 1 ⋅ 3 2 − C coll for all B ≥ B 1 B \ge B_1 B ≥ B 1 , which (I3) below allows. Put B : = max { β , B 1 } B := \max\{\beta,B_1\} B := max { β , B 1 } and let s s s be the scale of (I1) with b = β b=\beta b = β .
Tiling and coarse graph. If S ≥ s S \ge s S ≥ s , split [ 0 , S ] [0,S] [ 0 , S ] into consecutive integer intervals 0 = x 0 < ⋯ < x k = S 0=x_0<\cdots<x_k=S 0 = x 0 < ⋯ < x k = S with lengths in [ s , 2 s ] [s,2s] [ s , 2 s ] . The coarse vertices are the points ( x i , x j ) (x_i,x_j) ( x i , x j ) , and the cells are the rectangles [ x i , x i + 1 ] × [ x j , x j + 1 ] ∩ Z 2 [x_i,x_{i+1}]\times[x_j,x_{j+1}]\cap\mathbb{Z}^2 [ x i , x i + 1 ] × [ x j , x j + 1 ] ∩ Z 2 . A cell c c c has ring ∂ c \partial c ∂ c (its boundary sites) and interior c ∘ c^\circ c ∘ . A coarse vertex is pinned if i i i or j j j lies in { 0 , k } \{0,k\} { 0 , k } , and free otherwise; d G d_G d G is the grid maximum-norm distance between coarse vertices. Let G \mathcal{G} G be the graph on the free coarse vertices with d G d_G d G -adjacency (at most 8 neighbours); the free corners of a cell form a clique of G \mathcal{G} G . The ring sites are R = ⋃ c ∂ c R=\bigcup_c\partial c R = ⋃ c ∂ c ; the pinned ones form ∂ Q ( S ) \partial Q(S) ∂ Q ( S ) and the others form R f R_f R f . Let W : = { X free : d G ( X , pinned ) ≤ w } W:=\{X\ \text{free}:d_G(X,\text{pinned})\le w\} W := { X free : d G ( X , pinned ) ≤ w } (the seed layer ) and U : = { X free : d G ( X , pinned ) > w } U:=\{X\ \text{free}:d_G(X,\text{pinned})>w\} U := { X free : d G ( X , pinned ) > w } (the tested vertices ). For X ∈ U X\in U X ∈ U let I ( X ) I(X) I ( X ) be the set of ring sites that lie on ∂ c \partial c ∂ c for some c ∈ coll ( X ) c\in\operatorname{coll}(X) c ∈ coll ( X ) and on no ∂ c ′ \partial c' ∂ c ′ with c ′ ∉ coll ( X ) c'\notin\operatorname{coll}(X) c ′ ∈ / coll ( X ) . Then:
(F3) if d G ( X , Y ) > 2 r c o l l d_G(X,Y)>2r_{\mathrm{coll}} d G ( X , Y ) > 2 r coll then coll ( X ) ∩ coll ( Y ) = ∅ \operatorname{coll}(X)\cap\operatorname{coll}(Y)=\varnothing coll ( X ) ∩ coll ( Y ) = ∅ ; distinct coarse vertices whose indices agree modulo 2 r c o l l + 1 2r_{\mathrm{coll}}+1 2 r coll + 1 are at d G d_G d G -distance at least 2 r c o l l + 1 2r_{\mathrm{coll}}+1 2 r coll + 1 , so U U U splits into C c o l l C_{\mathrm{coll}} C coll classes, each with pairwise disjoint collars.
The width w w w of the seed layer is chosen so that tested collars contain no pinned site, as [22 §3.9] requires.
Cell actions [22 §§3.1, 3.3] . For an edge e e e with both endpoints in a cell c c c let κ c , e = 1 2 \kappa_{c,e}=\frac{1}{2} κ c , e = 2 1 if both endpoints lie on one side of ∂ c \partial c ∂ c and κ c , e = 1 \kappa_{c,e}=1 κ c , e = 1 otherwise. Every edge that does not join two pinned sites has ∑ c κ c , e = 1 \sum_c\kappa_{c,e}=1 ∑ c κ c , e = 1 . For ∙ ∈ { a n , i s o } \bullet\in\{\mathrm{an},\mathrm{iso}\} ∙ ∈ { an , iso } (anisotropic and isotropic) let
A c a n ( σ c ) = ∑ e ∈ c ∘ κ c , e σ x M e σ y , A c i s o ( σ c ) = ∑ e ∈ c ∘ κ c , e b e σ x ⋅ σ y , A_c^{\mathrm{an}}(\sigma_c)=\sum_{e\in c^\circ}\kappa_{c,e}\sigma_xM_e\sigma_y,\qquad A_c^{\mathrm{iso}}(\sigma_c)=\sum_{e\in c^\circ}\kappa_{c,e}b_e\,\sigma_x\cdot\sigma_y, A c an ( σ c ) = e ∈ c ∘ ∑ κ c , e σ x M e σ y , A c iso ( σ c ) = e ∈ c ∘ ∑ κ c , e b e σ x ⋅ σ y ,
F c ∙ ( ϱ ) = − log ∫ e A c ∙ ( ϱ , ℓ ) ∏ z ∈ c ∘ ω 4 ( d ℓ z ) . F_c^\bullet(\varrho)=-\log\int e^{A_c^\bullet(\varrho,\ell)}\prod_{z\in c^\circ}\omega_4(d\ell_z). F c ∙ ( ϱ ) = − log ∫ e A c ∙ ( ϱ , ℓ ) z ∈ c ∘ ∏ ω 4 ( d ℓ z ) .
F ∙ = ∑ c F c ∙ F^\bullet=\sum_cF_c^\bullet F ∙ = ∑ c F c ∙ , and let ν ∙ ( d ω ) ∝ e − F ∙ ( ω ) ∏ z ∈ R f ω 4 ( d ω z ) \nu^\bullet(d\omega)\propto e^{-F^\bullet(\omega)}\prod_{z\in R_f}\omega_4(d\omega_z) ν ∙ ( d ω ) ∝ e − F ∙ ( ω ) ∏ z ∈ R f ω 4 ( d ω z ) be the ring law, with ω = σ ˉ \omega=\bar\sigma ω = σ ˉ on ∂ Q ( S ) \partial Q(S) ∂ Q ( S ) . Given the rings, the cell interiors are independent with laws μ c ∙ ( d ℓ ∣ ϱ ) ∝ e A c ∙ ( ϱ , ℓ ) d ℓ \mu_c^\bullet(d\ell\mid\varrho)\propto e^{A_c^\bullet(\varrho,\ell)}d\ell μ c ∙ ( d ℓ ∣ ϱ ) ∝ e A c ∙ ( ϱ , ℓ ) d ℓ , since a site of c ∘ c^\circ c ∘ has all its neighbours in c c c .
Coarse orientations [22 §3.10] . Let g ∈ ( S 3 ) { f r e e c o a r s e v e r t i c e s } g\in(S^3)^{\{\mathrm{free\ coarse\ vertices}\}} g ∈ ( S 3 ) { free coarse vertices } , with g X : = 1 g_X:=1 g X := 1 at pinned X X X . On a coarse edge X → Y X\to Y X → Y (a side of a cell, positively oriented) put Φ ( g ; t ) = g X exp [ η ( t ) ℓ ( g X − 1 g Y ) ] \Phi(g;t)=g_X\exp[\eta(t)\ell(g_X^{-1}g_Y)] Φ ( g ; t ) = g X exp [ η ( t ) ℓ ( g X − 1 g Y )] , t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] , where η \eta η is a fixed smooth function, constant near the endpoints, with η ( 0 ) = 0 \eta(0)=0 η ( 0 ) = 0 and η ( 1 ) = 1 \eta(1)=1 η ( 1 ) = 1 , and ℓ \ell ℓ is a fixed measurable logarithm of norm at most π \pi π with ℓ ( 1 ) = 0 \ell(1)=0 ℓ ( 1 ) = 0 . Write ω = Φ ( g ) ω ~ \omega=\Phi(g)\widetilde\omega ω = Φ ( g ) ω for the sitewise left multiplication ω z = Φ ( g ; t z ) ω ~ z \omega_z=\Phi(g;t_z)\widetilde\omega_z ω z = Φ ( g ; t z ) ω z of the free ring sites, t z t_z t z being the position of z z z on its coarse edge. For fixed g g g this preserves Haar measure, and Φ ( g ) ω ~ ∣ ∂ c \Phi(g)\widetilde\omega|_{\partial c} Φ ( g ) ω ∣ ∂ c depends on g g g only through the corners of c c c . Let the pair ( g , ω ~ ) (g,\widetilde\omega) ( g , ω ) have density proportional to e − F a n ( Φ ( g ) ω ~ ) e^{-F^{\mathrm{an}}(\Phi(g)\widetilde\omega)} e − F an ( Φ ( g ) ω ) with respect to Haar measure.
The change of variables ( g , ω ~ ) ↦ ( g , Φ ( g ) ω ~ ) (g,\widetilde{\omega})\mapsto(g,\Phi(g)\widetilde{\omega}) ( g , ω ) ↦ ( g , Φ ( g ) ω ) has unit Jacobian, so ( g , ω ) (g,\omega) ( g , ω ) has density proportional to e − F a n ( ω ) e^{-F^{\mathrm{an}}(\omega)} e − F an ( ω ) : g g g is *Haar distributed and independent of ω \omega ω , and ω ∼ ν a n \omega\sim\nu^{\mathrm{an}} ω ∼ ν an . Let ν ω ~ \nu_{\widetilde{\omega}} ν ω be the conditional law of g g g given ω ~ \widetilde{\omega} ω ; its density is proportional to ∏ c ψ c ( g ) \prod_c\psi_c(g) ∏ c ψ c ( g ) with ψ c ( g ) = exp ( − F c a n ( Φ ( g ) ω ~ ∣ ∂ c ) ) \psi_c(g)=\exp(-F_c^{\mathrm{an}}(\Phi(g)\widetilde{\omega}|_{\partial c})) ψ c ( g ) = exp ( − F c an ( Φ ( g ) ω ∣ ∂ c )) .
Bad vertices [22 §3.10] , with the anisotropic action. A vertex X ∈ U X\in U X ∈ U is bad for ω ~ \widetilde{\omega} ω if there are assignments g ′ , g ′ ′ g',g'' g ′ , g ′′ differing only at X X X , and logarithms of norm at most π \pi π satisfying the endpoint equations on the edges at X X X , with the same data on the other edges, such that the interpolations Φ ′ , Φ ′ ′ \Phi',\Phi'' Φ ′ , Φ ′′ they define satisfy
∣ ∑ c ∈ star ( X ) [ F c a n ( Φ ′ ( g ′ ) ω ~ ) − F c a n ( Φ ′ ′ ( g ′ ′ ) ω ~ ) ] ∣ > ε ∗ ; \left|\sum_{c\in\operatorname{star}(X)}\left[F_c^{\mathrm{an}}(\Phi'(g')\widetilde{\omega})-F_c^{\mathrm{an}}(\Phi''(g'')\widetilde{\omega})\right]\right|>\varepsilon_*; c ∈ star ( X ) ∑ [ F c an ( Φ ′ ( g ′ ) ω ) − F c an ( Φ ′′ ( g ′′ ) ω ) ] > ε ∗ ;
otherwise X X X is good. At a good vertex, with the other orientations fixed, ∑ c ∋ X F c a n ( Φ ( g ) ω ~ ) \sum_{c\ni X}F_c^{\mathrm{an}}(\Phi(g)\widetilde{\omega}) ∑ c ∋ X F c an ( Φ ( g ) ω ) varies by at most ε ∗ \varepsilon_* ε ∗ as g X g_X g X ranges over S 3 S^3 S 3 . The set of valid parameters is compact and the star action is continuous in them and in ω ~ \widetilde{\omega} ω , so { X bad } \{X\text{ bad}\} { X bad } is open, hence measurable.
(I2) [22 §3.10.1] : for every X ∈ U X\in U X ∈ U there are a compact parameter set Λ X \Lambda_X Λ X , a finite union of coordinate rectangles, and ring rotations ϕ λ \phi_\lambda ϕ λ , λ ∈ Λ X \lambda\in\Lambda_X λ ∈ Λ X (sitewise left multiplications, smooth in λ \lambda λ , equal to the identity outside I ( X ) I(X) I ( X ) ), such that for every g g g , all g ′ , g ′ ′ g',g'' g ′ , g ′′ differing only at X X X and all admissible logarithms there are λ ′ , λ ′ ′ ∈ Λ X \lambda',\lambda''\in\Lambda_X λ ′ , λ ′′ ∈ Λ X (both extensions are built in one chart; Λ X \Lambda_X Λ X may carry a label j ∈ { 1 , 2 } j\in\{1,2\} j ∈ { 1 , 2 } for the two trials) with ϕ λ ′ = ϕ λ ′ ′ \phi_{\lambda'}=\phi_{\lambda''} ϕ λ ′ = ϕ λ ′′ on ∂ c \partial c ∂ c for every c ∉ star ( X ) c\notin\operatorname{star}(X) c ∈ / star ( X ) , and ϕ λ ′ = Φ ′ ( g ′ ) Φ ( g ) − 1 \phi_{\lambda'}=\Phi'(g')\Phi(g)^{-1} ϕ λ ′ = Φ ′ ( g ′ ) Φ ( g ) − 1 , ϕ λ ′ ′ = Φ ′ ′ ( g ′ ′ ) Φ ( g ) − 1 \phi_{\lambda''}=\Phi''(g'')\Phi(g)^{-1} ϕ λ ′′ = Φ ′′ ( g ′′ ) Φ ( g ) − 1 on the rings of the star. This is a deterministic construction; its constants do not depend on the lattice size, the couplings or the configurations.
(I3) [22 (3.58) and pp. 26–28] : for the reference isotropic system, every X ∈ U X\in U X ∈ U and every ε > 0 \varepsilon>0 ε > 0 , with R X : = R f ∖ I ( X ) R_X:=R_f\setminus I(X) R X := R f ∖ I ( X ) ,
ν i s o ( sup λ ∈ Λ X ∣ F i s o ( ϕ λ ω ) − F i s o ( ω ) ∣ > ε ∣ ω R X ) ≤ ζ ε ( B ) \nu^{\mathrm{iso}}\left(\sup_{\lambda\in\Lambda_X}\left|F^{\mathrm{iso}}(\phi_\lambda\omega)-F^{\mathrm{iso}}(\omega)\right|>\varepsilon\mathrel{\big|}\omega_{R_X}\right)\le\zeta_\varepsilon(B) ν iso ( λ ∈ Λ X sup F iso ( ϕ λ ω ) − F iso ( ω ) > ε ω R X ) ≤ ζ ε ( B )
for every configuration ω R X \omega_{R_X} ω R X , where ζ ε ( B ) → 0 \zeta_\varepsilon(B)\to0 ζ ε ( B ) → 0 as B → ∞ B\to\infty B → ∞ and ζ ε ( B ) \zeta_\varepsilon(B) ζ ε ( B ) depends only on ε \varepsilon ε , B B B and the fixed choices (this includes the fixed factor from the finitely many parameter rectangles [22 p. 28] ).
In [22 ] , (3.58) is stated for a collar without pins, uniformly under conditioning on all rings outside the collar, and is then applied to the whole bad-site test [22 p. 28] ; by (F1) tested collars have no pins. Given ω R X \omega_{R_X} ω R X , the conditional law of ω I ( X ) \omega_{I(X)} ω I ( X ) involves only the cells of coll ( X ) \operatorname{coll}(X) coll ( X ) , so the estimate does not depend on the ambient domain. In [22 ] the proof of (3.58) uses the stiffness bound in (3.54) and [22 (3.10), (3.11), (3.13), (3.14)] , that is, the iteration of [22 §§3.1–3.8] ; we assume (I3) as stated. Apart from the group structure of S 3 S^3 S 3 used for the coarse orientations, (I1) and (I3) are the only places where the dimension two and the value n = 4 n=4 n = 4 enter.
Transfer of the bad-site estimate
Lemma 7.2 (cell transfer). For every cell c c c with sides a c , b c ≤ 2 s a_c,b_c\le2s a c , b c ≤ 2 s and every ring configuration ρ \rho ρ ,
∣ F c a n ( ρ ) − F c i s o ( ρ ) ∣ ≤ 2 a c b c θ ≤ 8 θ s 2 . \left|F_c^{\mathrm{an}}(\rho)-F_c^{\mathrm{iso}}(\rho)\right|\le2a_cb_c\theta\le8\theta s^2. F c an ( ρ ) − F c iso ( ρ ) ≤ 2 a c b c θ ≤ 8 θ s 2 . For X ∈ U X\in U X ∈ U and every sitewise left multiplication ϕ \phi ϕ of the ring that is the identity outside I ( X ) I(X) I ( X ) ,
∣ [ F a n ( ϕ ω ) − F a n ( ω ) ] − [ F i s o ( ϕ ω ) − F i s o ( ω ) ] ∣ ≤ 16 K c o l l θ s 2 . \left|\left[F^{\mathrm{an}}(\phi\omega)-F^{\mathrm{an}}(\omega)\right]-\left[F^{\mathrm{iso}}(\phi\omega)-F^{\mathrm{iso}}(\omega)\right]\right|\le16K_{\mathrm{coll}}\theta s^2. [ F an ( ϕ ω ) − F an ( ω ) ] − [ F iso ( ϕ ω ) − F iso ( ω ) ] ≤ 16 K coll θ s 2 . Proof. For unit vectors ∣ σ T ( M e − b e I ) σ ′ ∣ ≤ θ \left|\sigma^{\mathsf T}(M_e-b_eI)\sigma'\right|\le\theta σ T ( M e − b e I ) σ ′ ≤ θ , so ∣ A c a n − A c i s o ∣ ≤ θ ∑ e ⊂ c κ c , e \left|A_c^{\mathrm{an}}-A_c^{\mathrm{iso}}\right|\le\theta\sum_{e\subset c}\kappa_{c,e} A c an − A c iso ≤ θ ∑ e ⊂ c κ c , e . A rectangle with sides a , b a,b a , b has a ( b + 1 ) + b ( a + 1 ) a(b+1)+b(a+1) a ( b + 1 ) + b ( a + 1 ) edges, of which 2 ( a + b ) 2(a+b) 2 ( a + b ) lie on its boundary and have weight 1 2 \frac{1}{2} 2 1 , so ∑ e κ c , e = 2 a b ≤ 8 s 2 \sum_e\kappa_{c,e}=2ab\le8s^2 ∑ e κ c , e = 2 ab ≤ 8 s 2 . Since e − ∥ Δ ∥ ∞ ∫ e A c i s o ≤ ∫ e A c a n ≤ e ∥ Δ ∥ ∞ ∫ e A c i s o e^{-\|\Delta\|_\infty}\int e^{A_c^{\mathrm{iso}}}\le\int e^{A_c^{\mathrm{an}}}\le e^{\|\Delta\|_\infty}\int e^{A_c^{\mathrm{iso}}} e − ∥Δ ∥ ∞ ∫ e A c iso ≤ ∫ e A c an ≤ e ∥Δ ∥ ∞ ∫ e A c iso with Δ = A c a n − A c i s o \Delta=A_c^{\mathrm{an}}-A_c^{\mathrm{iso}} Δ = A c an − A c iso , the first claim follows. For the second, F ∙ ( ϕ ω ) − F ∙ ( ω ) = ∑ c [ F c ∙ ( ϕ ω ) − F c ∙ ( ω ) ] F^\bullet(\phi\omega)-F^\bullet(\omega)=\sum_c[F_c^\bullet(\phi\omega)-F_c^\bullet(\omega)] F ∙ ( ϕ ω ) − F ∙ ( ω ) = ∑ c [ F c ∙ ( ϕ ω ) − F c ∙ ( ω )] over the cells whose rings meet I ( X ) I(X) I ( X ) ; these lie in coll ( X ) \operatorname{coll}(X) coll ( X ) , so there are at most K c o l l K_{\mathrm{coll}} K coll of them, and each contributes at most 2 ⋅ 8 θ s 2 2\cdot8\theta s^2 2 ⋅ 8 θ s 2 .
Lemma 7.3 (bad-site bound). Assume 16 K c o l l θ s 2 ≤ ε ∗ / 4 16K_{\mathrm{coll}}\theta s^2 \le\varepsilon_*/4 16 K coll θ s 2 ≤ ε ∗ /4 . There are events G X G_X G X , X ∈ U X \in U X ∈ U , of the ring variables such that: (a) if X X X is bad for ω ~ \widetilde{\omega} ω , then Φ ( g ) ω ~ ∈ G X \Phi(g)\widetilde{\omega} \in G_X Φ ( g ) ω ∈ G X for every g g g ; (b) G X G_X G X depends only on ω \omega ω restricted to ⋃ c ∈ coll ( X ) ∂ c \bigcup_{c\in\operatorname{coll}(X)}\partial c ⋃ c ∈ coll ( X ) ∂ c ; (c) for every finite S ⊂ U S \subset U S ⊂ U ,
ν a n ( ⋂ X ∈ S G X ) ≤ ζ a n ⌈ ∣ S ∣ / C c o l l ⌉ , ζ a n : = e ε ∗ / 4 ζ ε ∗ / 4 ( B ) ≤ 32 − C c o l l . \nu^{\mathrm{an}}\left(\bigcap_{X\in S}G_X\right)\leq\zeta_{\mathrm{an}}^{\lceil|S|/C_{\mathrm{coll}}\rceil},\qquad\zeta_{\mathrm{an}}:=e^{\varepsilon_*/4}\zeta_{\varepsilon_*/4}(B)\leq32^{-C_{\mathrm{coll}}}. ν an ( X ∈ S ⋂ G X ) ≤ ζ an ⌈ ∣ S ∣/ C coll ⌉ , ζ an := e ε ∗ /4 ζ ε ∗ /4 ( B ) ≤ 3 2 − C coll .
Proof. Step 1 (deterministic). Let g g g be arbitrary, ω = Φ ( g ) ω ~ \omega=\Phi(g)\widetilde{\omega} ω = Φ ( g ) ω , and let g ′ , g ′ ′ g',g'' g ′ , g ′′ and the logarithms be as in the definition of a bad vertex. With λ ′ , λ ′ ′ \lambda',\lambda'' λ ′ , λ ′′ from (I2), for every cell-additive F = ∑ c F c F=\sum_c F_c F = ∑ c F c ,
F ( ϕ λ ′ ω ) − F ( ϕ λ ′ ′ ω ) = ∑ c ∈ star ( X ) [ F c ( Φ ′ ( g ′ ) ω ~ ) − F c ( Φ ′ ′ ( g ′ ′ ) ω ~ ) ] , F(\phi_{\lambda'}\omega)-F(\phi_{\lambda''}\omega)=\sum_{c\in\operatorname{star}(X)}\left[F_c(\Phi'(g')\widetilde{\omega})-F_c(\Phi''(g'')\widetilde{\omega})\right], F ( ϕ λ ′ ω ) − F ( ϕ λ ′′ ω ) = c ∈ star ( X ) ∑ [ F c ( Φ ′ ( g ′ ) ω ) − F c ( Φ ′′ ( g ′′ ) ω ) ] ,
since the two rotations agree on the rings of the cells outside the star, as [22 p. 28] notes. Let G X : = { ω : sup λ ∈ Λ X ∣ F a n ( ϕ λ ω ) − F a n ( ω ) ∣ > ε ∗ / 2 } G_X:=\{\omega:\sup_{\lambda\in\Lambda_X}|F^{\mathrm{an}}(\phi_\lambda\omega)-F^{\mathrm{an}}(\omega)|>\varepsilon_*/2\} G X := { ω : sup λ ∈ Λ X ∣ F an ( ϕ λ ω ) − F an ( ω ) ∣ > ε ∗ /2 } . If X X X is bad, the left side exceeds ε ∗ \varepsilon_* ε ∗ for some choice, so one of the two displacements exceeds ε ∗ / 2 \varepsilon_*/2 ε ∗ /2 ; this is (a). Since ϕ λ \phi_\lambda ϕ λ is the identity outside I ( X ) I(X) I ( X ) , only cells of coll ( X ) \operatorname{coll}(X) coll ( X ) change, which is (b). G X G_X G X is open, since Λ X \Lambda_X Λ X is a finite union of compact rectangles and the displacement is jointly continuous.
Step 2 (transfer). By Lemma 7.2 , G X ⊂ G X i s o : = { ω : sup λ ∈ Λ X ∣ F i s o ( ϕ λ ω ) − F i s o ( ω ) ∣ > ε ∗ / 4 } G_X\subset G_X^{\mathrm{iso}}:=\{\omega:\sup_{\lambda\in\Lambda_X}|F^{\mathrm{iso}}(\phi_\lambda\omega)-F^{\mathrm{iso}}(\omega)|>\varepsilon_*/4\} G X ⊂ G X iso := { ω : sup λ ∈ Λ X ∣ F iso ( ϕ λ ω ) − F iso ( ω ) ∣ > ε ∗ /4 } .
Step 3 (isotropic bound). By (I3) with ε = ε ∗ / 4 \varepsilon=\varepsilon_*/4 ε = ε ∗ /4 , ν i s o ( G X i s o ∣ ω R X ) ≤ ζ ε ∗ / 4 ( B ) \nu^{\mathrm{iso}}(G_X^{\mathrm{iso}}\mid\omega_{R_X})\leq\zeta_{\varepsilon_*/4}(B) ν iso ( G X iso ∣ ω R X ) ≤ ζ ε ∗ /4 ( B ) for every ω R X \omega_{R_X} ω R X , where R X = R f ∖ I ( X ) R_X=R_f\setminus I(X) R X = R f ∖ I ( X ) .
Step 4 (conditional density). Fix ω R X \omega_{R_X} ω R X . For ∙ ∈ { a n , i s o } \bullet\in\{\mathrm{an},\mathrm{iso}\} ∙ ∈ { an , iso } the conditional ring law is ν ∙ ( d ω I ∣ ω R X ) ∝ exp ( − ∑ c ∈ coll ( X ) F c ∙ ( ω ) ) ∏ z ∈ I ω 4 ( d ω z ) \nu^\bullet(d\omega_I\mid\omega_{R_X})\propto\exp(-\sum_{c\in\operatorname{coll}(X)}F_c^\bullet(\omega))\prod_{z\in I}\omega_4(d\omega_z) ν ∙ ( d ω I ∣ ω R X ) ∝ exp ( − ∑ c ∈ coll ( X ) F c ∙ ( ω )) ∏ z ∈ I ω 4 ( d ω z ) , I = I ( X ) I=I(X) I = I ( X ) , since the other cells do not involve ω I \omega_I ω I . By Lemma 7.2 the Radon–Nikodym derivative of ν a n ( ⋅ ∣ ω R X ) \nu^{\mathrm{an}}(\cdot\mid\omega_{R_X}) ν an ( ⋅ ∣ ω R X ) with respect to ν i s o ( ⋅ ∣ ω R X ) \nu^{\mathrm{iso}}(\cdot\mid\omega_{R_X}) ν iso ( ⋅ ∣ ω R X ) is at most exp ( 2 ∑ c ∈ coll ( X ) ∥ F c a n − F c i s o ∥ ∞ ) ≤ e 16 K c o l l θ s 2 ≤ e ε ∗ / 4 \exp(2\sum_{c\in\operatorname{coll}(X)}\|F_c^{\mathrm{an}}-F_c^{\mathrm{iso}}\|_\infty)\leq e^{16K_{\mathrm{coll}}\theta s^2}\leq e^{\varepsilon_*/4} exp ( 2 ∑ c ∈ coll ( X ) ∥ F c an − F c iso ∥ ∞ ) ≤ e 16 K coll θ s 2 ≤ e ε ∗ /4 . Hence ν a n ( G X ∣ ω R X ) ≤ e ε ∗ / 4 ν i s o ( G X i s o ∣ ω R X ) ≤ ζ a n \nu^{\mathrm{an}}(G_X\mid\omega_{R_X})\leq e^{\varepsilon_*/4}\nu^{\mathrm{iso}}(G_X^{\mathrm{iso}}\mid\omega_{R_X})\leq\zeta_{\mathrm{an}} ν an ( G X ∣ ω R X ) ≤ e ε ∗ /4 ν iso ( G X iso ∣ ω R X ) ≤ ζ an . The anisotropic law of ω R X \omega_{R_X} ω R X never enters.
Step 5 (iterated conditioning). By (F3), choose S ′ ⊂ S S'\subset S S ′ ⊂ S with ∣ S ′ ∣ ≥ ⌈ ∣ S ∣ / C c o l l ⌉ |S'|\geq\lceil|S|/C_{\mathrm{coll}}\rceil ∣ S ′ ∣ ≥ ⌈ ∣ S ∣/ C coll ⌉ and pairwise disjoint collars, and enumerate it as X 1 , … , X m X_1,\ldots,X_m X 1 , … , X m . For j < m j<m j < m , G X j G_{X_j} G X j depends on the rings of coll ( X j ) \operatorname{coll}(X_j) coll ( X j ) , which avoid I ( X m ) I(X_m) I ( X m ) , because a site of I ( X m ) I(X_m) I ( X m ) lies only on rings of cells of coll ( X m ) \operatorname{coll}(X_m) coll ( X m ) . So
ν a n ( ⋂ j ≤ m G X j ) = E a n [ ∏ j < m 1 G X j ν a n ( G X m ∣ ω R X m ) ] ≤ ζ a n ν a n ( ⋂ j < m G X j ) , \nu^{\mathrm{an}}\left(\bigcap_{j\leq m}G_{X_j}\right)=\mathbb{E}^{\mathrm{an}}\left[\prod_{j<m}\mathbf{1}_{G_{X_j}}\nu^{\mathrm{an}}(G_{X_m}\mid\omega_{R_{X_m}})\right]\leq\zeta_{\mathrm{an}}\nu^{\mathrm{an}}\left(\bigcap_{j<m}G_{X_j}\right), ν an ( j ≤ m ⋂ G X j ) = E an [ j < m ∏ 1 G X j ν an ( G X m ∣ ω R X m ) ] ≤ ζ an ν an ( j < m ⋂ G X j ) ,
and induction gives ζ a n m \zeta_{\mathrm{an}}^m ζ an m . Finally ζ a n ≤ e 1 / 128 ⋅ 1 2 ⋅ 32 − C c o l l ≤ 32 − C c o l l \zeta_{\mathrm{an}}\leq e^{1/128}\cdot\frac12\cdot32^{-C_{\mathrm{coll}}}\leq32^{-C_{\mathrm{coll}}} ζ an ≤ e 1/128 ⋅ 2 1 ⋅ 3 2 − C coll ≤ 3 2 − C coll . □
Odd observables and the coupling
Lemma 7.4 (the symmetry σ ↦ − σ \sigma\mapsto-\sigma σ ↦ − σ ). (a) A c a n ( − ϱ , − ι ) = A c a n ( ϱ , ι ) A_c^{\mathrm{an}}(-\varrho,-\iota)=A_c^{\mathrm{an}}(\varrho,\iota) A c an ( − ϱ , − ι ) = A c an ( ϱ , ι ) ; hence F c a n ( − ϱ ) = F c a n ( ϱ ) F_c^{\mathrm{an}}(-\varrho)=F_c^{\mathrm{an}}(\varrho) F c an ( − ϱ ) = F c an ( ϱ ) , and μ c a n ( ⋅ ∣ − ϱ ) \mu_c^{\mathrm{an}}(\cdot\mid-\varrho) μ c an ( ⋅ ∣ − ϱ ) is the image of μ c a n ( ⋅ ∣ ϱ ) \mu_c^{\mathrm{an}}(\cdot\mid\varrho) μ c an ( ⋅ ∣ ϱ ) under ι ↦ − ι \iota\mapsto-\iota ι ↦ − ι . (b) Let T T T negate g g g at every free coarse vertex. On every coarse edge with both endpoints free, Φ ( T g ; t ) = − Φ ( g ; t ) \Phi(Tg;t)=-\Phi(g;t) Φ ( T g ; t ) = − Φ ( g ; t ) ; hence for every cell c c c with four free corners Φ ( T g ) ω ~ ∣ ∂ c = − Φ ( g ) ω ~ ∣ ∂ c \Phi(Tg)\widetilde{\omega}|_{\partial c}=-\Phi(g)\widetilde{\omega}|_{\partial c} Φ ( T g ) ω ∣ ∂ c = − Φ ( g ) ω ∣ ∂ c and ψ c ( T g ) = ψ c ( g ) \psi_c(Tg)=\psi_c(g) ψ c ( T g ) = ψ c ( g ) . (c) The law ν − 1 : = T # ν ~ \nu_{-1}:=T_{\#}\widetilde{\nu} ν − 1 := T # ν has density proportional to ∏ c ψ c ( T g ) \prod_c\psi_c(Tg) ∏ c ψ c ( T g ) , whose factor for every cell with four free corners, in particular by (F2) for every cell with a corner in U U U , equals that of ν ~ \widetilde{\nu} ν .
Proof. (a) Each term ( − σ ) T M c ( − σ ′ ) = σ T M c σ ′ (-\sigma)^{\mathsf T}M_c(-\sigma')=\sigma^{\mathsf T}M_c\sigma' ( − σ ) T M c ( − σ ′ ) = σ T M c σ ′ is bilinear, and ω 4 \omega_4 ω 4 is invariant under σ ↦ − σ \sigma\mapsto-\sigma σ ↦ − σ . (b) Since − 1 -1 − 1 is central, ( − g X ) − 1 ( − g Y ) = g X − 1 g Y (-g_X)^{-1}(-g_Y)=g_X^{-1}g_Y ( − g X ) − 1 ( − g Y ) = g X − 1 g Y , so Φ ( T g ; t ) = ( − g X ) exp [ η ( t ) ℓ ( g X − 1 g Y ) ] = − Φ ( g ; t ) \Phi(Tg;t)=(-g_X)\exp[\eta(t)\ell(g_X^{-1}g_Y)]=-\Phi(g;t) Φ ( T g ; t ) = ( − g X ) exp [ η ( t ) ℓ ( g X − 1 g Y )] = − Φ ( g ; t ) ; then use (a). (c) T T T is a Haar-preserving involution. □
This is the only place where the structure of the anisotropy is used beyond Lemma 7.2 : the interaction is even. The common left multiplication by a general k ∈ S 3 k\in S^3 k ∈ S 3 , which [22 ] averages over, does not preserve σ T M c σ ′ \sigma^{\mathsf T}M_c\sigma' σ T M c σ ′ unless M c M_c M c commutes with left multiplication by k k k .
Lemma 7.5 (odd observables). Let O O O be an odd observable on A ⊂ Q ( S ) ∖ ∂ Q ( S ) A \subset Q(S) \setminus\partial Q(S) A ⊂ Q ( S ) ∖ ∂ Q ( S ) , let C A C_A C A be the set of cells whose rectangle meets A A A , and let A ∗ A^* A ∗ be the set of corners of cells in C A C_A C A , so ∣ A ∗ ∣ ≤ 16 ∣ A ∣ |A^*| \le16|A| ∣ A ∗ ∣ ≤ 16∣ A ∣ . Put O ♯ ( ω ) : = ∫ O ( σ A ) ∏ c ∈ C A μ c ε ( d ι c ∣ ω ∣ ∂ c ) O^\sharp(\omega) := \int O(\sigma_A) \prod_{c \in C_A} \mu_c^{\varepsilon}(d\iota_c \mid\omega|_{\partial c}) O ♯ ( ω ) := ∫ O ( σ A ) ∏ c ∈ C A μ c ε ( d ι c ∣ ω ∣ ∂ c ) , where σ A \sigma_A σ A is read off ( ω , ι ) (\omega,\iota) ( ω , ι ) , and O ∗ ( g ) : = O ♯ ( Φ ( g ) ω ~ ) O^*(g) := O^\sharp(\Phi(g)\widetilde\omega) O ∗ ( g ) := O ♯ ( Φ ( g ) ω ) . Then (a) ⟨ O ⟩ = E ω ~ [ E ν ω ~ O ∗ ] \langle O\rangle= \mathbb{E}_{\widetilde\omega}[\mathbb{E}_{\nu_{\widetilde\omega}}O^*] ⟨ O ⟩ = E ω [ E ν ω O ∗ ] , where ω ~ \widetilde\omega ω has its marginal law under the joint law; (b) ∣ O ∗ ∣ ≤ ∥ O ∥ ∞ |O^*| \le\|O\|_\infty ∣ O ∗ ∣ ≤ ∥ O ∥ ∞ and O ∗ O^* O ∗ depends on g g g only through g A ∗ g^{A^*} g A ∗ ; (c) if every cell of C A C_A C A has four free corners, then O ∗ ( T g ) = − O ∗ ( g ) O^*(Tg) = -O^*(g) O ∗ ( T g ) = − O ∗ ( g ) .
Proof. (a) The Gibbs law is the ring law ν a n \nu^{\mathrm{an}} ν an followed by the independent interior kernels, so ⟨ O ⟩ = E ν a n [ O ♯ ] \langle O\rangle= \mathbb{E}_{\nu^{\mathrm{an}}}[O^\sharp] ⟨ O ⟩ = E ν an [ O ♯ ] , and Φ ( g ) ω ~ ∼ ν a n \Phi(g)\widetilde\omega\sim\nu^{\mathrm{an}} Φ ( g ) ω ∼ ν an under the joint law. (b) is clear. (c) By Lemma 7.4(a) and the substitution ι ↦ − ι \iota\mapsto-\iota ι ↦ − ι , O ♯ ( − ω ) = − O ♯ ( ω ) O^\sharp(-\omega) = -O^\sharp(\omega) O ♯ ( − ω ) = − O ♯ ( ω ) ; by Lemma 7.4(b), Φ ( T g ) ω ~ = − Φ ( g ) ω ~ \Phi(Tg)\widetilde\omega= -\Phi(g)\widetilde\omega Φ ( T g ) ω = − Φ ( g ) ω on the rings of the cells of C A C_A C A .
Lemma 7.6 (disagreement coupling). Let G \mathcal{G} G be a finite graph with vertex set U ⊔ W U \sqcup W U ⊔ W , let ( Y , ϖ ) (Y,\varpi) ( Y , ϖ ) be a compact probability space, and let C \mathcal{C} C be a finite family of cliques of G \mathcal{G} G . For j = 1 , 2 j=1,2 j = 1 , 2 let μ j ( d g ) = Z j − 1 ∏ c ∈ C ψ c j ( g c ) ϖ ⊗ ( U ⊔ W ) ( d g ) \mu^j(dg) = Z_j^{-1}\prod_{c\in\mathcal{C}}\psi_c^j(g_c)\varpi^{\otimes(U\sqcup W)}(dg) μ j ( d g ) = Z j − 1 ∏ c ∈ C ψ c j ( g c ) ϖ ⊗ ( U ⊔ W ) ( d g ) , with each ψ c j \psi_c^j ψ c j measurable and bounded above and below by positive constants, and with ψ c 1 = ψ c 2 \psi_c^1=\psi_c^2 ψ c 1 = ψ c 2 whenever c ∩ U ≠ ∅ c\cap U\ne\varnothing c ∩ U = ∅ . Let p ∈ [ 0 , 1 ] U p\in[0,1]^U p ∈ [ 0 , 1 ] U satisfy μ j ( g v ∈ ⋅ ∣ g r e s t ) ≥ ( 1 − p v ) ϖ \mu^j(g_v\in\cdot\mid g_{\mathrm{rest}})\ge(1-p_v)\varpi μ j ( g v ∈ ⋅ ∣ g rest ) ≥ ( 1 − p v ) ϖ for every v ∈ U v\in U v ∈ U , every j j j and every configuration of the rest. Then for every A ′ ⊂ U A'\subset U A ′ ⊂ U and every bounded measurable H H H of g A ′ g_{A'} g A ′ ,
∣ μ 1 ( H ) − μ 2 ( H ) ∣ ≤ 2 ∥ H ∥ ∞ P p ( A ′ ↔ W ) , \left|\mu^1(H)-\mu^2(H)\right|\le2\|H\|_\infty\mathbb{P}_p(A'\leftrightarrow W), μ 1 ( H ) − μ 2 ( H ) ≤ 2∥ H ∥ ∞ P p ( A ′ ↔ W ) ,
where under P p \mathbb{P}_p P p the vertices of U U U are independently open with probabilities p v p_v p v , and { A ′ ↔ W } \{A'\leftrightarrow W\} { A ′ ↔ W } is the event that there are distinct open v 0 , … , v m − 1 ∈ U v_0,\ldots,v_{m-1}\in U v 0 , … , v m − 1 ∈ U , m ≥ 1 m\ge1 m ≥ 1 , with v 0 ∈ A ′ v_0\in A' v 0 ∈ A ′ , v i − 1 v_{i-1} v i − 1 adjacent to v i v_i v i , and v m − 1 v_{m-1} v m − 1 adjacent to W W W .
Proof. The mechanism is that of van den Berg and Maes [2 Theorem 1] , which is stated for finite single-site spaces; the continuous version with adaptive exploration is [22 §3.10.2] . We give the argument. Let ( U v ) v ∈ U (U_v)_{v\in U} ( U v ) v ∈ U be independent and uniform on [ 0 , 1 ] [0,1] [ 0 , 1 ] . Draw X W 1 X_W^1 X W 1 and X W 2 X_W^2 X W 2 independently from the W W W -marginals of μ 1 \mu^1 μ 1 and μ 2 \mu^2 μ 2 , and declare W W W revealed and marked. While some unrevealed vertex of U U U is adjacent to a marked vertex, let v v v be the first such vertex in a fixed enumeration and let m j : = μ j ( g v ∈ ⋅ ∣ g r e v e a l e d = X r e v e a l e d j ) m^j:=\mu^j(g_v\in\cdot\mid g_{\mathrm{revealed}}=X_{\mathrm{revealed}}^j) m j := μ j ( g v ∈ ⋅ ∣ g revealed = X revealed j ) ; as an average of one-site conditionals, m j ≥ ( 1 − p v ) ϖ m^j\ge(1-p_v)\varpi m j ≥ ( 1 − p v ) ϖ . If U v ≥ p v U_v\ge p_v U v ≥ p v , draw y ∼ ϖ y\sim\varpi y ∼ ϖ and set X v 1 = X v 2 = y X_v^1=X_v^2=y X v 1 = X v 2 = y . Otherwise draw X v j X_v^j X v j independently from ( m j − ( 1 − p v ) ϖ ) / p v (m^j-(1-p_v)\varpi)/p_v ( m j − ( 1 − p v ) ϖ ) / p v and mark v v v . Reveal v v v . When this stops, let Q Q Q be the set of unrevealed vertices. Every neighbour of Q Q Q is revealed and unmarked, so X 1 = X 2 X^1=X^2 X 1 = X 2 on the boundary of Q Q Q . Since the factors are cliques, the conditional law of g Q g_Q g Q given the revealed values is proportional to ∏ c ∩ Q ≠ ∅ ψ c j \prod_{c\cap Q\ne\varnothing}\psi_c^j ∏ c ∩ Q = ∅ ψ c j and depends only on the values on the boundary of Q Q Q ; these factors meet U U U , so they agree for j = 1 , 2 j=1,2 j = 1 , 2 , and we draw X Q 1 = X Q 2 X_Q^1=X_Q^2 X Q 1 = X Q 2 from the common kernel.
Marginals. Fix j j j and a bounded H H H , and let M t M_t M t be the μ j \mu^j μ j -conditional expectation of H H H given the values revealed after t t t steps. Given the history, the next vertex is determined, U v U_v U v is independent, and X v j X_v^j X v j has law ( 1 − p v ) ϖ + p v ( m j − ( 1 − p v ) ϖ ) / p v = m j (1-p_v)\varpi+p_v(m^j-(1-p_v)\varpi)/p_v=m^j ( 1 − p v ) ϖ + p v ( m j − ( 1 − p v ) ϖ ) / p v = m j ; so ( M t ) (M_t) ( M t ) is a martingale, the final draw has the conditional law, and E H ( X j ) = E M 0 = μ j ( H ) \mathbb{E}H(X^j)=\mathbb{E}M_0=\mu^j(H) E H ( X j ) = E M 0 = μ j ( H ) . Disagreement. If X a 1 ≠ X a 2 X_a^1\ne X_a^2 X a 1 = X a 2 for some a ∈ A ′ a\in A' a ∈ A ′ , then a a a is marked: each marked vertex of U U U was revealed while adjacent to an earlier marked vertex or to W W W , and following these predecessors gives distinct marked v 0 = a , v 1 , … v_0=a,v_1,\ldots v 0 = a , v 1 , … ending next to W W W . Marked vertices have U v < p v U_v<p_v U v < p v , so { X A ′ 1 ≠ X A ′ 2 } \{X_{A'}^1\ne X_{A'}^2\} { X A ′ 1 = X A ′ 2 } is contained in an event of ( U v ) (U_v) ( U v ) of probability at most P p ( A ′ ↔ W ) \mathbb{P}_p(A'\leftrightarrow W) P p ( A ′ ↔ W ) .
Proposition 7.7 (pinned decay). Assume 16 K c o l θ s 2 ≤ ε ∗ / 4 16K_{\mathrm{col}}\theta s^2\le\varepsilon_*/4 16 K col θ s 2 ≤ ε ∗ /4 . For every odd observable O O O on A ⊂ Q ( S ) ∖ ∂ Q ( S ) A\subset Q(S)\setminus\partial Q(S) A ⊂ Q ( S ) ∖ ∂ Q ( S ) , with d : = dist ∞ ( A , ∂ Q ( S ) ) d:=\operatorname{dist}_\infty(A,\partial Q(S)) d := dist ∞ ( A , ∂ Q ( S )) ,
∣ ⟨ O ⟩ ∣ ≤ 2 3 + w ∣ A ∣ ∥ O ∥ ∞ exp ( − ( log 2 ) d 2 s ) . |\langle O\rangle|\le2^{3+w}|A|\|O\|_\infty\exp\left(-\frac{(\log2)d}{2s}\right). ∣ ⟨ O ⟩ ∣ ≤ 2 3 + w ∣ A ∣∥ O ∥ ∞ exp ( − 2 s ( log 2 ) d ) .
Proof. If d ≤ 2 s ( w + 2 ) d \le2s(w+2) d ≤ 2 s ( w + 2 ) , which includes the case S < s S < s S < s , then ∣ ⟨ O ⟩ ∣ ≤ ∥ O ∥ ∞ ≤ 2 w + 2 e − ( log 2 ) d / ( 2 s ) ∥ O ∥ ∞ \lvert\langle O\rangle\rvert\le\lVert O\rVert_\infty\le2^{w+2}e^{-(\log2)d/(2s)}\lVert O\rVert_\infty ∣⟨ O ⟩∣ ≤ ∥ O ∥ ∞ ≤ 2 w + 2 e − ( l o g 2 ) d / ( 2 s ) ∥ O ∥ ∞ . Let d > 2 s ( w + 2 ) d > 2s(w+2) d > 2 s ( w + 2 ) ; then S ≥ 2 d > s S \ge2d > s S ≥ 2 d > s and the tiling exists.
Geometry. A corner X X X of a cell of C A C_A C A is within sup-distance 2 s 2s 2 s of A A A , so dist ∞ ( X , ∂ Q ( S ) ) ≥ d − 2 s \operatorname{dist}_\infty(X,\partial Q(S)) \ge d-2s dist ∞ ( X , ∂ Q ( S )) ≥ d − 2 s ; each grid step moves at most 2 s 2s 2 s , so d G ( X , p i n n e d ) ≥ d / ( 2 s ) − 1 > w + 1 d_G(X,\mathrm{pinned}) \ge d/(2s)-1 > w+1 d G ( X , pinned ) ≥ d / ( 2 s ) − 1 > w + 1 . Hence A ∗ ⊂ U A^* \subset U A ∗ ⊂ U and every cell of C A C_A C A has four free corners.
Oddness. Fix ω ~ \widetilde\omega ω and write ν = ν ω ~ \nu=\nu_{\widetilde\omega} ν = ν ω . By Lemma 7.5 7.5 7.5 (c), E ν O ∗ = − E ν [ O ∗ ∘ T ] = − E ν − 1 O ∗ \mathbb{E}_{\nu}O^*=-\mathbb{E}_{\nu}[O^*\circ T]=-\mathbb{E}_{\nu_{-1}}O^* E ν O ∗ = − E ν [ O ∗ ∘ T ] = − E ν − 1 O ∗ , so ∣ E ν O ∗ ∣ = 1 2 ∣ E ν O ∗ − E ν − 1 O ∗ ∣ \lvert\mathbb{E}_{\nu}O^*\rvert=\frac{1}{2}\lvert\mathbb{E}_{\nu}O^*-\mathbb{E}_{\nu_{-1}}O^*\rvert ∣ E ν O ∗ ∣ = 2 1 ∣ E ν O ∗ − E ν − 1 O ∗ ∣ .
Coupling. Apply Lemma 7.6 7.6 7.6 with G , U , W G,U,W G , U , W as above, C C C the cells restricted to their free corners, ϖ = ω 4 \varpi=\omega_4 ϖ = ω 4 , μ 1 = ν \mu^1=\nu μ 1 = ν , μ 2 = ν − 1 \mu^2=\nu_{-1} μ 2 = ν − 1 , ψ c 1 = ψ c \psi_c^1=\psi_c ψ c 1 = ψ c , ψ c 2 = ψ c ∘ T \psi_c^2=\psi_c\circ T ψ c 2 = ψ c ∘ T , A ′ = A ∗ A'=A^* A ′ = A ∗ and H = O ∗ H=O^* H = O ∗ . The factors agree on every cell meeting U U U (Lemma 7.4 7.4 7.4 (c)) and are bounded above and below. Put p v = q p_v=q p v = q if v v v is good and p v = 1 p_v=1 p v = 1 if v v v is bad. At v ∈ U v\in U v ∈ U the one-site conditional density of ν \nu ν is proportional to ∏ c ∋ v ψ c \prod_{c\ni v}\psi_c ∏ c ∋ v ψ c as a function of g v g_v g v ; if v v v is good its exponent varies by at most ε ∗ \varepsilon_* ε ∗ , so the density is at least e − ε ∗ = 1 − q e^{-\varepsilon_*}=1-q e − ε ∗ = 1 − q , and the same holds for ν − 1 \nu_{-1} ν − 1 , whose factors at v v v coincide with those of ν \nu ν . Hence ∣ E ν O ∗ ∣ ≤ ∥ O ∥ ∞ P p ( ω ~ ) ( A ∗ ↔ W ) \lvert\mathbb{E}_{\nu}O^*\rvert\le\lVert O\rVert_\infty\mathbb{P}_{p(\widetilde\omega)}(A^*\leftrightarrow W) ∣ E ν O ∗ ∣ ≤ ∥ O ∥ ∞ P p ( ω ) ( A ∗ ↔ W ) .
Average over ω ~ \widetilde\omega ω . By Lemma 7.5 7.5 7.5 (a) and a union bound, ∣ ⟨ O ⟩ ∣ ≤ ∥ O ∥ ∞ E ω ~ [ ∑ π ∏ v ∈ π p v ( ω ~ ) ] \lvert\langle O\rangle\rvert\le\lVert O\rVert_\infty\mathbb{E}_{\widetilde\omega}\left[\sum_{\pi}\prod_{v\in\pi}p_v(\widetilde\omega)\right] ∣⟨ O ⟩∣ ≤ ∥ O ∥ ∞ E ω [ ∑ π ∏ v ∈ π p v ( ω ) ] , summed over sequences π = ( v 0 , … , v m − 1 ) \pi=(v_0,\ldots,v_{m-1}) π = ( v 0 , … , v m − 1 ) as in Lemma 7.6 7.6 7.6 . Here ∏ v ∈ π p v = ∏ v ∈ π ( q + ( 1 − q ) 1 v b a d ) \prod_{v\in\pi}p_v=\prod_{v\in\pi}(q+(1-q)\mathbf{1}_{v\ \mathrm{bad}}) ∏ v ∈ π p v = ∏ v ∈ π ( q + ( 1 − q ) 1 v bad ) , which by Lemma 7.3 7.3 7.3 (a) is at most ∏ v ( q + ( 1 − q ) 1 G v ( Φ ( g ) ω ~ ) ) \prod_v(q+(1-q)\mathbf{1}_{G_v(\Phi(g)\widetilde\omega)}) ∏ v ( q + ( 1 − q ) 1 G v ( Φ ( g ) ω ) ) for every g g g . Taking g g g jointly distributed with ω ~ \widetilde\omega ω , so that Φ ( g ) ω ~ ∼ ν a n \Phi(g)\widetilde\omega\sim\nu^{\mathrm{an}} Φ ( g ) ω ∼ ν an , expanding the product and using Lemma 7.3 7.3 7.3 (c) and ζ a n ≤ 1 \zeta_{\mathrm{an}}\le1 ζ an ≤ 1 ,
E ω ~ [ ∏ v ∈ π p v ] ≤ ∑ S ⊂ π q m − ∣ S ∣ ( 1 − q ) ∣ S ∣ ν a n ( ⋂ v ∈ S G v ) ≤ ( q + ( 1 − q ) ζ a n 1 / C c o l l ) m ≤ 16 − m . \mathbb{E}_{\widetilde\omega}\left[\prod_{v\in\pi}p_v\right]\le\sum_{S\subset\pi}q^{m-\lvert S\rvert}(1-q)^{\lvert S\rvert}\nu^{\mathrm{an}}\left(\bigcap_{v\in S}G_v\right)\le\left(q+(1-q)\zeta_{\mathrm{an}}^{1/C_{\mathrm{coll}}}\right)^m\le16^{-m}. E ω [ v ∈ π ∏ p v ] ≤ S ⊂ π ∑ q m − ∣ S ∣ ( 1 − q ) ∣ S ∣ ν an ( v ∈ S ⋂ G v ) ≤ ( q + ( 1 − q ) ζ an 1/ C coll ) m ≤ 1 6 − m .
since q ≤ 1 / 32 q\le1/32 q ≤ 1/32 and ζ a n 1 / C c o l l ≤ 1 / 32 \zeta_{\mathrm{an}}^{1/C_{\mathrm{coll}}}\le1/32 ζ an 1/ C coll ≤ 1/32 . This is the analogue of [22 (3.60)] .
Path count. Every vertex of W W W is within d G d_G d G -distance w w w of a pinned vertex, so every such π \pi π starting at v 0 ∈ A ∗ v_0\in A^* v 0 ∈ A ∗ has m ≥ d G ( v 0 , p i n n e d ) − w ≥ d / ( 2 s ) − 1 − w = : k 0 m\ge d_G(v_0,\mathrm{pinned})-w\ge d/(2s)-1-w=:k_0 m ≥ d G ( v 0 , pinned ) − w ≥ d / ( 2 s ) − 1 − w =: k 0 . There are at most 8 m 8^m 8 m such π \pi π for each start, so the sum is at most ∣ A ∗ ∣ ∑ m ≥ ⌈ k 0 ⌉ 8 m 16 − m ≤ ∣ A ∗ ∣ 2 − k 0 / 4 ≤ 4 ∣ A ∣ 2 1 + w 2 − d / ( 2 s ) \lvert A^*\rvert\sum_{m\ge\lceil k_0\rceil}8^m16^{-m}\le\lvert A^*\rvert2^{-k_0}/4\le4\lvert A\rvert2^{1+w^2-d/(2s)} ∣ A ∗ ∣ ∑ m ≥ ⌈ k 0 ⌉ 8 m 1 6 − m ≤ ∣ A ∗ ∣ 2 − k 0 /4 ≤ 4 ∣ A ∣ 2 1 + w 2 − d / ( 2 s ) .
With θ = 0 \theta=0 θ = 0 this is the decay of [22 §3.10] restricted to odd observables, with the seed-layer width fixed as w w w .
Scale and threshold
Lemma 7.8 (scale). There is an absolute C Ξ C_{\Xi} C Ξ such that ξ s : = 2 s / log 2 ≤ Ξ ( β ) \xi_s:=2s/\log2\le\Xi(\beta) ξ s := 2 s / log 2 ≤ Ξ ( β ) for all β > 0 \beta>0 β > 0 . Consequently e − ( log 2 ) d / ( 2 s ) ≤ e − d / Ξ ( β ) e^{-(\log2)d/(2s)}\le e^{-d/\Xi(\beta)} e − ( l o g 2 ) d / ( 2 s ) ≤ e − d /Ξ ( β ) , and θ Ξ ( β ) 2 ≤ c 1 : = 1 / ( 256 K c o l l ) \theta\Xi(\beta)^2\le c_1:=1/(256K_{\mathrm{coll}}) θ Ξ ( β ) 2 ≤ c 1 := 1/ ( 256 K coll ) implies 16 K c o l l θ s 2 ≤ ε ∗ / 4 16K_{\mathrm{coll}}\theta s^2\le\varepsilon_*/4 16 K coll θ s 2 ≤ ε ∗ /4 .
Proof. By (I1) and M ∗ ≤ 2 B D + C 0 M^*\le2B^{D+C_0} M ∗ ≤ 2 B D + C 0 , log s ≤ π β + C D B log B + ( D + C 0 ) log B + log 2 \log s\le\pi\beta+C_D\sqrt{B\log B}+(D+C_0)\log B+\log2 log s ≤ π β + C D B log B + ( D + C 0 ) log B + log 2 . If β ≥ B 1 \beta\ge B_1 β ≥ B 1 , then B = β B=\beta B = β , log β ≤ β log β ≤ ( 1 + β ) log ( 2 + β ) \log\beta\le\sqrt{\beta\log\beta}\le\sqrt{(1+\beta)\log(2+\beta)} log β ≤ β log β ≤ ( 1 + β ) log ( 2 + β ) , and log s ≤ π β + ( C D + D + C 0 ) ( 1 + β ) log ( 2 + β ) + log 2 \log s\le\pi\beta+(C_D+D+C_0)\sqrt{(1+\beta)\log(2+\beta)}+\log2 log s ≤ π β + ( C D + D + C 0 ) ( 1 + β ) log ( 2 + β ) + log 2 . If β < B 1 \beta<B_1 β < B 1 , then log s ≤ π β + K 1 \log s\le\pi\beta+K_1 log s ≤ π β + K 1 with K 1 : = C D B 1 log B 1 + ( D + C 0 ) log B 1 + log 2 K_1:=C_D\sqrt{B_1\log B_1}+(D+C_0)\log B_1+\log2 K 1 := C D B 1 log B 1 + ( D + C 0 ) log B 1 + log 2 . Since ( 1 + β ) log ( 2 + β ) ≥ log 2 \sqrt{(1+\beta)\log(2+\beta)}\ge\sqrt{\log2} ( 1 + β ) log ( 2 + β ) ≥ log 2 , the choice C Ξ : = C D + D + C 0 + ( K 1 + log ( 4 / log 2 ) ) / log 2 C_{\Xi}:=C_D+D+C_0+(K_1+\log(4/\log2))/\sqrt{\log2} C Ξ := C D + D + C 0 + ( K 1 + log ( 4/ log 2 )) / log 2 gives log ξ s ≤ log Ξ ( β ) \log\xi_s\le\log\Xi(\beta) log ξ s ≤ log Ξ ( β ) in both cases. Finally s 2 = ( log 2 / 2 ) 2 ξ s 2 ≤ Ξ 2 / 8 s^2=(\log2/2)^2\xi_s^2\le\Xi^2/8 s 2 = ( log 2/2 ) 2 ξ s 2 ≤ Ξ 2 /8 , so 16 K c o l l θ s 2 ≤ 2 K c o l l θ Ξ 2 ≤ 1 / 128 16K_{\mathrm{coll}}\theta s^2\le2K_{\mathrm{coll}}\theta\Xi^2\le1/128 16 K coll θ s 2 ≤ 2 K coll θ Ξ 2 ≤ 1/128 .
Lemma 7.9 (threshold). Put L : = ∣ log δ ∣ L:=\lvert\log\delta\rvert L := ∣ log δ ∣ ; the function f ( β ) : = β Ξ ( β ) 2 f(\beta):=\beta\Xi(\beta)^2 f ( β ) := β Ξ ( β ) 2 is increasing. (a) For every ε ∈ ( 0 , 1 ) \varepsilon\in(0,1) ε ∈ ( 0 , 1 ) there is δ 0 ( ε ) ∈ ( 0 , 1 ) \delta_0(\varepsilon)\in(0,1) δ 0 ( ε ) ∈ ( 0 , 1 ) such that 0 < δ < δ 0 ( ε ) 0<\delta<\delta_0(\varepsilon) 0 < δ < δ 0 ( ε ) and 0 < β ≤ ( 1 − ε ) L / ( 2 π ) 0<\beta\le(1-\varepsilon)L/(2\pi) 0 < β ≤ ( 1 − ε ) L / ( 2 π ) imply β δ Ξ ( β ) 2 ≤ c 1 \beta\delta\Xi(\beta)^2\le c_1 β δ Ξ ( β ) 2 ≤ c 1 . (b) With K : = 3 C Ξ + 2 K:=3C_{\Xi}+2 K := 3 C Ξ + 2 there is δ 1 > 0 \delta_1>0 δ 1 > 0 such that 0 < δ < δ 1 0<\delta<\delta_1 0 < δ < δ 1 and 0 < β ≤ ( L − K L log L ) / ( 2 π ) 0<\beta\le(L-K\sqrt{L\log L})/(2\pi) 0 < β ≤ ( L − K L log L ) / ( 2 π ) imply β δ Ξ ( β ) 2 ≤ c 1 \beta\delta\Xi(\beta)^2\le c_1 β δ Ξ ( β ) 2 ≤ c 1 .
Proof. (a) With β δ : = ( 1 − ε ) L / ( 2 π ) ≤ L \beta_\delta:=(1-\varepsilon)L/(2\pi)\le L β δ := ( 1 − ε ) L / ( 2 π ) ≤ L ,
δ f ( β ) ≤ δ f ( β δ ) = β δ exp ( − L + ( 1 − ε ) L + 2 C Ξ ( 1 + β δ ) log ( 2 + β δ ) ) \delta f(\beta)\le\delta f(\beta_\delta)=\beta_\delta\exp\left(-L+(1-\varepsilon)L+2C_{\Xi}\sqrt{(1+\beta_\delta)\log(2+\beta_\delta)}\right) δ f ( β ) ≤ δ f ( β δ ) = β δ exp ( − L + ( 1 − ε ) L + 2 C Ξ ( 1 + β δ ) log ( 2 + β δ ) ) ≤ L exp ( − ε L + 2 C Ξ ( 1 + L ) log ( 2 + L ) ) , (7) \le L\exp\left(-\varepsilon L+2C_{\Xi}\sqrt{(1+L)\log(2+L)}\right), \tag*{(7)} ≤ L exp ( − ε L + 2 C Ξ ( 1 + L ) log ( 2 + L ) ) , (7) @page 20/23 starts_mid=1 ends_mid=1
which tends to 0 as L → ∞ L \to\infty L → ∞ . (b) For L ≥ 4 L \ge4 L ≥ 4 we have ( 1 + L ) log ( 2 + L ) ≤ 1.875 L log L (1+L)\log(2+L) \le1.875L\log L ( 1 + L ) log ( 2 + L ) ≤ 1.875 L log L , hence
2 C Ξ ( 1 + L ) log ( 2 + L ) ≤ 2.74 C Ξ L log L 2C_{\Xi}\sqrt{(1+L)\log(2+L)} \le2.74C_{\Xi}\sqrt{L\log L} 2 C Ξ ( 1 + L ) log ( 2 + L ) ≤ 2.74 C Ξ L log L
and
δ f ( β ) ≤ L exp ( − ( K − 2.74 C Ξ ) L log L ) ≤ L e − 2 L log L , \delta f(\beta) \le L\exp\left(-(K-2.74C_{\Xi})\sqrt{L\log L}\right) \le Le^{-2\sqrt{L\log L}}, δ f ( β ) ≤ L exp ( − ( K − 2.74 C Ξ ) L log L ) ≤ L e − 2 L l o g L ,
which tends to 0.
General systems; proofs of Theorems 7.1 and 1.5
Box conditioning. Let S \mathcal{S} S be ( β , θ ) (\beta,\theta) ( β , θ ) -admissible, x ∈ V ∖ P x \in V \setminus P x ∈ V ∖ P , and r ≥ 1 r \ge1 r ≥ 1 an integer with dist ∞ ( x , P ) ≥ r \operatorname{dist}_{\infty}(x,P) \ge r dist ∞ ( x , P ) ≥ r . Put Λ ∘ : = { z : ∥ z − x ∥ ∞ < r } \Lambda^{\circ} := \{z : \|z-x\|_{\infty} < r\} Λ ∘ := { z : ∥ z − x ∥ ∞ < r } , R : = { z : ∥ z − x ∥ ∞ = r } R := \{z : \|z-x\|_{\infty} = r\} R := { z : ∥ z − x ∥ ∞ = r } and Λ : = Λ ∘ ∪ R \Lambda:= \Lambda^{\circ} \cup R Λ := Λ ∘ ∪ R , a translate of Q ( 2 r ) Q(2r) Q ( 2 r ) . Add the sites of Λ ∖ V \Lambda\setminus V Λ ∖ V with independent ω n \omega_n ω n -distributed spins, and the missing edges of Λ \Lambda Λ with b e = 0 b_e=0 b e = 0 and M e = 0 M_e=0 M e = 0 ; the system stays admissible and the law of the original spins does not change. By the Markov property of nearest-neighbour Gibbs measures (see [12 Chapter 2] ), given everything outside Λ ∘ \Lambda^{\circ} Λ ∘ the spins in Λ ∘ \Lambda^{\circ} Λ ∘ have the law ⟨ ⋅ ⟩ Λ , σ ∣ R \langle\cdot\rangle_{\Lambda,\sigma|_R} ⟨ ⋅ ⟩ Λ , σ ∣ R of the pinned-square system on Λ \Lambda Λ with ring values σ ∣ R \sigma|_R σ ∣ R (edges inside R R R contribute constants). So for bounded F F F measurable with respect to the spins outside Λ ∘ \Lambda^{\circ} Λ ∘ and a function h h h of σ x \sigma_x σ x ,
⟨ F h ( σ x ) ⟩ = ⟨ F ⟨ h ( σ x ) ⟩ Λ , σ ∣ R ⟩ . (8) \langle Fh(\sigma_x)\rangle=\left\langle F\langle h(\sigma_x)\rangle_{\Lambda,\sigma|_R}\right\rangle. \tag*{(8)} ⟨ F h ( σ x )⟩ = ⟨ F ⟨ h ( σ x ) ⟩ Λ , σ ∣ R ⟩ . (8)
The case n = 4 n=4 n = 4 . If θ Ξ ( β ) 2 ≤ c 1 \theta\Xi(\beta)^2 \le c_1 θ Ξ ( β ) 2 ≤ c 1 , Lemma 7.8 and Proposition 7.7 with A = { x } A=\{x\} A = { x } and d = r d=r d = r give, for every ring configuration ϱ \varrho ϱ and every odd O O O ,
∣ ⟨ O ( σ x ) ⟩ Λ , ϱ ∣ ≤ 2 3 + w ∥ O ∥ ∞ e − r / Ξ ( β ) . (9) \left|\langle O(\sigma_x)\rangle_{\Lambda,\varrho}\right|\le2^{3+w}\|O\|_{\infty}e^{-r/\Xi(\beta)}. \tag*{(9)} ∣ ⟨ O ( σ x ) ⟩ Λ , ϱ ∣ ≤ 2 3 + w ∥ O ∥ ∞ e − r /Ξ ( β ) . (9)
Put A ∗ : = 2 3 + w A_* := 2^{3+w} A ∗ := 2 3 + w , which also covers the case d ≤ 2 s ( w + 2 ) d \le2s(w+2) d ≤ 2 s ( w + 2 ) of Proposition 7.7 .
The case n ≥ 5 n \ge5 n ≥ 5 , diagonal M e M_e M e . Fix i i i and a set W i ∋ i W_i \ni i W i ∋ i of four coordinates, and write σ = ( ρ u , v ) \sigma=(\rho u,v) σ = ( ρ u , v ) with u ∈ S 3 ⊂ R W i u \in S^3 \subset\mathbb{R}^{W_i} u ∈ S 3 ⊂ R W i , v ∈ R W i c v \in\mathbb{R}^{W_i^c} v ∈ R W i c and ρ = ( 1 − ∣ v ∣ 2 ) 1 / 2 \rho=(1-|v|^2)^{1/2} ρ = ( 1 − ∣ v ∣ 2 ) 1/2 . By Lemma 5.2 with k = 4 k=4 k = 4 , ρ > 0 \rho>0 ρ > 0 almost surely and under ω n \omega_n ω n the direction u u u is uniform on S 3 S^3 S 3 and independent of v v v . For diagonal M e M_e M e ,
σ z T M e σ z ′ = ρ z ρ z ′ ∑ k ∈ W i m k k u z u z ′ + ∑ k ∉ W i m k k v z k v z ′ k , \sigma_z^{\mathsf{T}}M_e\sigma_{z'}=\rho_z\rho_{z'}\sum_{k\in W_i}m_k^ku_zu_{z'}+\sum_{k\notin W_i}m_k^kv_z^kv_{z'}^k, σ z T M e σ z ′ = ρ z ρ z ′ k ∈ W i ∑ m k k u z u z ′ + k ∈ / W i ∑ m k k v z k v z ′ k ,
with no cross term. Conditionally on v v v in Λ ∘ \Lambda^{\circ} Λ ∘ and on the ring values, u u u is a pinned-square system with n = 4 n=4 n = 4 , couplings M e ′ = ρ z ρ z ′ diag ( m k k ) k ∈ W i M'_e=\rho_z\rho_{z'}\operatorname{diag}(m_k^k)_{k\in W_i} M e ′ = ρ z ρ z ′ diag ( m k k ) k ∈ W i , b e ′ = ρ z ρ z ′ b e b'_e=\rho_z\rho_{z'}b_e b e ′ = ρ z ρ z ′ b e , and pins u p u_p u p (arbitrary where ρ p = 0 \rho_p=0 ρ p = 0 , in which case the couplings to p p p vanish). It is ( β , θ ) (\beta,\theta) ( β , θ ) -admissible, since ∣ b e ′ ∣ ≤ β |b'_e|\le\beta ∣ b e ′ ∣ ≤ β and ∥ M e ′ − b e ′ I ∥ ≤ ρ z ρ z ′ θ ≤ θ \|M'_e-b'_eI\|\le\rho_z\rho_{z'}\theta\le\theta ∥ M e ′ − b e ′ I ∥ ≤ ρ z ρ z ′ θ ≤ θ . By (9) applied to u i u^i u i ,
∣ ⟨ σ x i ⟩ Λ , ϱ ∣ = ∣ ⟨ ρ x ⟨ u x i ∣ v ⟩ ⟩ Λ , ϱ ∣ ≤ A ∗ e − r / Ξ ( β ) . \left|\langle\sigma_x^i\rangle_{\Lambda,\varrho}\right|=\left|\left\langle\rho_x\left\langle u_x^i\mid v\right\rangle\right\rangle_{\Lambda,\varrho}\right|\le A_*e^{-r/\Xi(\beta)}. ⟨ σ x i ⟩ Λ , ϱ = ⟨ ρ x ⟨ u x i ∣ v ⟩ ⟩ Λ , ϱ ≤ A ∗ e − r /Ξ ( β ) .
Proof of Theorem 7.1 . (a) If P = ∅ P=\varnothing P = ∅ , the symmetry σ ↦ − σ \sigma\mapsto-\sigma σ ↦ − σ gives ⟨ O ( σ x ) ⟩ = 0 \langle O(\sigma_x)\rangle=0 ⟨ O ( σ x )⟩ = 0 ; otherwise take r = dist ∞ ( x , P ) r=\operatorname{dist}_{\infty}(x,P) r = dist ∞ ( x , P ) and F = 1 F=1 F = 1 in (8) and use (9) . (b) Take r = min { ∥ x − y ∥ ∞ , dist ∞ ( x , P ) } r=\min\{\|x-y\|_{\infty},\operatorname{dist}_{\infty}(x,P)\} r = min { ∥ x − y ∥ ∞ , dist ∞ ( x , P )} ; then y ∉ Λ ∘ y\notin\Lambda^{\circ} y ∈ / Λ ∘ , and F = O ′ ( σ y ) F=O'(\sigma_y) F = O ′ ( σ y ) is measurable outside Λ ∘ \Lambda^{\circ} Λ ∘ . (c) Let d = d T ( x , y ) d=d_T(x,y) d = d T ( x , y ) and r = min { d , ⌊ ( min { L 1 , L 2 } − 1 ) / 2 ⌋ } r=\min\{d,\lfloor(\min\{L_1,L_2\}-1)/2\rfloor\} r = min { d , ⌊( min { L 1 , L 2 } − 1 ) /2 ⌋} ; if r = 0 r=0 r = 0 the bound is trivial since A ∗ ≥ 1 A_*\ge1 A ∗ ≥ 1 . Otherwise 2 r + 1 ≤ min { L 1 , L 2 } 2r+1\le\min\{L_1,L_2\} 2 r + 1 ≤ min { L 1 , L 2 } , so the box Λ \Lambda Λ around x x x embeds in the torus without wrapping, y ∉ Λ ∘ y\notin\Lambda^{\circ} y ∈ / Λ ∘ , and the argument of (b) applies with the Markov property on the torus.
Proof of Theorem 1.5 . The system is ( β , β δ ) (\beta,\beta\delta) ( β , β δ ) -admissible, and by Lemma 7.9 (a), with δ 0 ( ε ) \delta_0(\varepsilon) δ 0 ( ε ) from there, β δ Ξ ( β ) 2 ≤ c 1 \beta\delta\Xi(\beta)^2\le c_1 β δ Ξ ( β ) 2 ≤ c 1 . For n = 4 n=4 n = 4 apply Theorem 7.1 with O = σ i O=\sigma^i O = σ i and O ′ = σ j O'=\sigma^j O ′ = σ j . For n ≥ 5 n\ge5 n ≥ 5 , (8) and the bound on ⟨ σ x i ⟩ Λ , ϱ \langle\sigma_x^i\rangle_{\Lambda,\varrho} ⟨ σ x i ⟩ Λ , ϱ above give (i) and (ii) in the same way. For P = ∅ P=\varnothing P = ∅ and diagonal M e M_e M e , flipping σ i \sigma^i σ i at every site preserves the measure, so ⟨ σ x i σ y j ⟩ = 0 \langle\sigma_x^i\sigma_y^j\rangle=0 ⟨ σ x i σ y j ⟩ = 0 for i ≠ j i\ne j i = j . On a square torus d T ( x , y ) ≤ ⌊ L / 2 ⌋ ≤ ⌊ ( L − 1 ) / 2 ⌋ + 1 d_T(x,y)\le\lfloor L/2\rfloor\le\lfloor(L-1)/2\rfloor+1 d T ( x , y ) ≤ ⌊ L /2 ⌋ ≤ ⌊( L − 1 ) /2 ⌋ + 1 , so r ≥ d T ( x , y ) − 1 r\ge d_T(x,y)-1 r ≥ d T ( x , y ) − 1 and e − r / Ξ ≤ e ⋅ e − d T ( x , y ) / Ξ e^{-r/\Xi}\le e\cdot e^{-d_T(x,y)/\Xi} e − r /Ξ ≤ e ⋅ e − d T ( x , y ) /Ξ since Ξ ≥ 1 \Xi\ge1 Ξ ≥ 1 . For an infinite-volume Gibbs measure μ \mu μ and every r ≥ 1 r\ge1 r ≥ 1 , the DLR equations give μ ( ⋅ ∣ F Z n ∖ Λ ∘ ) = ⟨ ⋅ ⟩ Λ , σ ∣ R \mu(\cdot\mid\mathcal{F}_{\mathbb{Z}^n\setminus\Lambda^{\circ}})=\langle\cdot\rangle_{\Lambda,\sigma|_R} μ ( ⋅ ∣ F Z n ∖ Λ ∘ ) = ⟨ ⋅ ⟩ Λ , σ ∣ R μ \mu μ -almost surely; so ∣ μ ( σ x i ) ∣ ≤ A ∗ e − r / Ξ |\mu(\sigma_x^i)|\le A_*e^{-r/\Xi} ∣ μ ( σ x i ) ∣ ≤ A ∗ e − r /Ξ for every r r r , hence μ ( σ x i ) = 0 \mu(\sigma_x^i)=0 μ ( σ x i ) = 0 , and with r = ∥ x − y ∥ ∞ r=\|x-y\|_{\infty} r = ∥ x − y ∥ ∞ the argument of (ii) gives the two-point bound. The sharper condition on β \beta β stated after Theorem 1.5 is Lemma 7.9 (b).
For n = 4 n=4 n = 4 the same argument shows that every Gibbs measure of a ( β , θ ) (\beta,\theta) ( β , θ ) -admissible specification with θ Ξ ( β ) 2 ≤ c 1 \theta\Xi(\beta)^2\le c_1 θ Ξ ( β ) 2 ≤ c 1 gives mean zero to every odd observable of one spin.
What does not transfer
Remark 7.10 (only odd observables transfer). For the isotropic model, [22 §3.10] gives decay for every bounded observable whose average under common left multiplication by S 3 S^3 S 3 vanishes. For anisotropic couplings this fails already on two sites. Take the free two-site law on S 3 × S 3 S^3\times S^3 S 3 × S 3 proportional to exp ( b u ⋅ v + b t u 1 v 1 ) \exp(b\,u\cdot v+bt\,u^1v^1) exp ( b u ⋅ v + b t u 1 v 1 ) . Under the isotropic law (t = 0 t=0 t = 0 ), E [ v ∣ u ] = c ( b ) u \mathbb{E}[v\mid u]=c(b)u E [ v ∣ u ] = c ( b ) u with c ( b ) = ⟨ u ⋅ v ⟩ = I 2 ( b ) / I 1 ( b ) > 0 c(b)=\langle u\cdot v\rangle=I_2(b)/I_1(b)>0 c ( b ) = ⟨ u ⋅ v ⟩ = I 2 ( b ) / I 1 ( b ) > 0 for b > 0 b>0 b > 0 , where I 1 , I 2 I_1,I_2 I 1 , I 2 are modified Bessel functions of the first kind, and E [ ( u 1 ) 4 ] = 1 8 \mathbb{E}[(u^1)^4]=\frac{1}{8} E [( u 1 ) 4 ] = 8 1 , E [ ( u 1 ) 2 ] = 1 4 \mathbb{E}[(u^1)^2]=\frac{1}{4} E [( u 1 ) 2 ] = 4 1 . Hence
d d t ⟨ ( u 1 ) 2 ⟩ ∣ t = 0 = b Cov ( ( u 1 ) 2 , u 1 v 1 ) = b ( c 8 − 1 4 ⋅ c 4 ) = b c ( b ) 16 ≠ 0. \left.\frac{d}{dt}\langle(u^1)^2\rangle\right|_{t=0}
=b\operatorname{Cov}((u^1)^2,u^1v^1)
=b\left(\frac{c}{8}-\frac{1}{4}\cdot\frac{c}{4}\right)
=\frac{bc(b)}{16}\ne0. d t d ⟨( u 1 ) 2 ⟩ t = 0 = b Cov (( u 1 ) 2 , u 1 v 1 ) = b ( 8 c − 4 1 ⋅ 4 c ) = 16 b c ( b ) = 0.
So the observable ( u 1 ) 2 − 1 4 (u^1)^2-\frac{1}{4} ( u 1 ) 2 − 4 1 , whose average over common left rotations vanishes, has nonzero mean for small t ≠ 0 t\ne0 t = 0 , with no pins at all; at b = 1 b=1 b = 1 the derivative is 0.01501 … 0.01501\ldots 0.01501 … , which a direct quadrature of the two-site integral confirms. Putting b e = 0 b_e=0 b e = 0 on every other edge embeds this example in pinned squares of every size, within the class of Theorem 7.1. Hence the restriction to odd observables cannot be replaced by a vanishing S 3 S^3 S 3 -average: even observables, such as bond energies or ( σ x 1 ) 2 (\sigma_x^1)^2 ( σ x 1 ) 2 , are not covered, and neither is uniqueness of the Gibbs measure beyond the vanishing of odd one-point functions. In the proof, oddness enters only in Lemma 7.5(c) and in the oddness step of Proposition 7.7, and evenness of the interaction only in Lemma 7.4(a). An odd single-site term, such as a field h ∑ x σ x 1 h\sum_x\sigma_x^1 h ∑ x σ x 1 with ∣ h ∣ ≤ β δ |h|\le\beta\delta ∣ h ∣ ≤ β δ , satisfies Lemma 7.2 but not Lemma 7.4, and with it the conclusion of Theorem 7.1(a) fails.
Remark 7.11 (scope). Apart from the group structure of S 3 S^3 S 3 used for the coarse orientations, the value n = 4 n=4 n = 4 and the dimension two enter only through (I1) and (I3), and the reduction from n ≥ 5 n\ge5 n ≥ 5 to n = 4 n=4 n = 4 lowers n n n , so it cannot reach n = 3 n=3 n = 3 . For n = 3 n=3 n = 3 the predicted correlation length e 2 π β / β e^{2\pi\beta}/\beta e 2 π β / β is much larger than Ξ ( β ) \Xi(\beta) Ξ ( β ) , so an argument at the scale Ξ ( β ) \Xi(\beta) Ξ ( β ) cannot reach the O ( 3 ) O(3) O ( 3 ) crossover. For n ≥ 5 n\ge5 n ≥ 5 Theorem 1.5 gives the O ( 4 ) O(4) O ( 4 ) threshold, below the prediction ( n − 2 ) ∣ log δ ∣ / ( 4 π ) (n-2)|\log\delta|/(4\pi) ( n − 2 ) ∣ log δ ∣/ ( 4 π ) . Not covered are: truncated correlations of even observables; the regime above the threshold, where no matching order statement is proved; non-diagonal anisotropy for n ≥ 5 n\ge5 n ≥ 5 ; single-site anisotropy (odd single-site terms are excluded by Remark 7.10; even ones would add O ( β δ s 2 ) O(\beta\delta s^2) O ( β δ s 2 ) per cell in Lemma 7.2, but this is not carried out here); and rectangular tori beyond the scale min { L 1 , L 2 } / 2 \min\{L_1,L_2\}/2 min { L 1 , L 2 } /2 . The constants c 1 , A ∗ , C Ξ , δ 0 ( ε ) c_1,A_*,C_\Xi,\delta_0(\varepsilon) c 1 , A ∗ , C Ξ , δ 0 ( ε ) are not numerical; they depend on the fixed choices of [22 ] .