Introduction
How large must a periodic box be for the large-N N N Gaussian approximation to describe local spin measurements? In the nearest-neighbour O ( N ) O(N) O ( N ) model the spin length is fixed at N \sqrt{N} N and the box has side length M M M . Convergence as N → ∞ N \to\infty N → ∞ at fixed M M M gives no error estimate for a box that grows with N N N or for a state obtained by taking volume first.
For the isotropic periodic model at zero field, Aru and Korzhenkova identify the local law obtained by taking N → ∞ N \to\infty N → ∞ before M → ∞ M \to\infty M → ∞ [1 Theorems 4.1–4.2] . At the critical temperature of the limiting spherical theory, it is a scaled massless Gaussian free field (GFF). Below that temperature it also contains an independent Gaussian shift shared by all sites. They ask whether the limits can be interchanged [1 Section 1.2] . See also [22 Theorem 5.1.2 and Section 1.5, question 8] . We control this approximation uniformly in volume at and below the limiting critical temperature, allowing positive directional couplings.
For the first p ≤ N p \le N p ≤ N spin coordinates on a fixed spatial window, Theorem 2.1 bounds the unnormalized 1-Wasserstein distance to the infinite-lattice Gaussian law for sufficiently large M , N M,N M , N by C u p ( N − 1 / 2 + s M ) C u\sqrt{p}\left(N^{-1/2}+s_M\right) C u p ( N − 1/2 + s M ) , uniformly over the compact parameter ranges in § 2 . The spectral shift s M s_M s M is determined by the finite-volume spherical constraint (4) . For fixed p p p , the approximation error tends to zero along every joint path. At fixed β \beta β , J J J , the joint limit is the same Gaussian law obtained from any choice of periodic accumulation states when volume is taken first (Corollary 2.5 ).
When p p p grows with M M M and N N N , two lower bounds limit the approximation. The one-site sphere marginal detects the fixed spin length [14 , 19 ] . On a window containing a site and its d d d positive-direction neighbours (a nearest-neighbour star), a test based on spin differences detects the finite-volume error. The conditions p = o ( N ) p=o(N) p = o ( N ) and p s M → 0 \sqrt{p}\,s_M\to0 p s M → 0 are necessary and sufficient on such windows and sufficient on every fixed window (Theorem 2.3 ). For periodic accumulation states the criterion reduces to p = o ( N ) p=o(N) p = o ( N ) on any nonempty fixed window (Corollary 2.4).
A sharper estimate holds in the ordered regime away from the critical point. In d ≥ 5 d\ge5 d ≥ 5 , if β − β c , J ≥ κ > 0 \beta-\beta_{c,J}\ge\kappa>0 β − β c , J ≥ κ > 0 , the upper bound improves to C U , κ ( p / N + p M − d ) C_{U,\kappa}(p/N+\sqrt{p}\,M^{-d}) C U , κ ( p / N + p M − d ) . For fixed p p p , matching lower bounds give the sharp order N − 1 + M − d N^{-1}+M^{-d} N − 1 + M − d on star-containing windows and N − 1 N^{-1} N − 1 in periodic accumulation states. Theorem 2.8 also improves the fixed-p p p dependence on N N N in d = 3 , 4 d=3,4 d = 3 , 4 .
The relation between spin and spherical models is classical [3 , 32 , 27 ] . Kac and Thompson proved agreement of limiting free energies at noncritical temperatures [20 ] . Shcherbina’s radial Hamiltonian compares the free energies per site and spin component with error of order ( log N ) / N (\log N)/N ( log N ) / N , uniformly in volume and at every temperature [29 Theorem 1, Eqs. (2.3)–(2.4) and Section 3] . Proposition 4.2 adapts that comparison to obtain a volume-uniform O ( N − 1 ) O(N^{-1}) O ( N − 1 ) error.
Above the limiting critical temperature, Kupiainen obtained volume-uniform 1 / N 1/N 1/ N expansions for the free energy and correlation functions of the isotropic nearest-neighbour model [24 ] . Shcherbina stated joint convergence of fixed-order zero-field correlations in that regime [29 Section 4, Remark 1] . Ye and Zhu identify the critical and ordered Gaussian fields as stationary laws of their infinite-volume limiting dynamics [34 Theorem 1.5] . At small inverse temperature, 0 < κ < 1 / ( 32 d ) 0<\kappa<1/(32d) 0 < κ < 1/ ( 32 d ) in their notation, they bound the W 2 W_2 W 2 error for fixed-coordinate marginals of the spin laws and prove weak convergence to the same one-coordinate law in either order of limits [34 Theorems 1.6, 5.15–5.16 and Corollary 5.17] .
In a uniform field, Pearce and Thompson identified the limiting one-point functions with those of the mean spherical model by differentiating the limiting free energy. Their two-point comparison assumes the required interchanges of limits and derivatives [28 ] . In the field h ( 1 , … , 1 ) h(1,\ldots,1) h ( 1 , … , 1 ) , Shcherbina proved joint convergence of fixed-order coordinate correlations, with the field removed only after M , N → ∞ M,N\to\infty M , N → ∞ [29 Theorem 2] . At low temperature this order of limits selects a magnetization direction. The periodic zero-field states considered here retain O ( N ) O(N) O ( N ) invariance.
Following the reflection-positivity approach suggested in [1 Section 1.2] , we combine Gaussian domination with the pressure comparison to control the local Gram matrix. The conditional Haar projection turns that estimate into control of coordinate laws. The orbit representation and the corresponding Gaussian mixture are classical [13 Example 2.1 and Proposition 2.3] . In the ordered regime a sphere Ward identity and a second-order Gaussian mixture estimate give the improved rate. Section 3 outlines the argument. Appendix A controls the finite-volume spectral shift.
Model and main results
Let d ≥ 3 d\ge3 d ≥ 3 , J = ( J 1 , … , J d ) J=(J_1,\ldots,J_d) J = ( J 1 , … , J d ) with J i > 0 J_i>0 J i > 0 , and j = ∑ i J i j=\sum_i J_i j = ∑ i J i . On Λ M = ( Z / M Z ) d \Lambda_M=(\mathbb{Z}/M\mathbb{Z})^d Λ M = ( Z / M Z ) d , M ≥ 4 M\ge4 M ≥ 4 , each S x ∈ R N S_x\in\mathbb{R}^N S x ∈ R N has length N \sqrt{N} N . Let σ N \sigma_N σ N be uniform probability on that sphere. The Gibbs law is
1 Z M , N , β , J exp { β ∑ x ∈ Λ M ∑ i = 1 d J i S x ⋅ S x + e i } ∏ x σ N ( d S x ) . (1) \frac{1}{Z_{M,N,\beta,J}}\exp\left\{\beta\sum_{x\in\Lambda_M}\sum_{i=1}^{d}J_i S_x\cdot S_{x+e_i}\right\}\prod_x\sigma_N(\mathrm{d}S_x). \tag*{(1)} Z M , N , β , J 1 exp { β x ∈ Λ M ∑ i = 1 ∑ d J i S x ⋅ S x + e i } x ∏ σ N ( d S x ) . (1)
Each unoriented edge is counted once, and the external field is zero. Define
( L J , M f ) ( x ) = ∑ i J i [ 2 f ( x ) − f ( x + e i ) − f ( x − e i ) ] , λ J ( k ) = 2 ∑ i J i ( 1 − cos k i ) . (L_{J,M}f)(x)=\sum_i J_i[2f(x)-f(x+e_i)-f(x-e_i)],\qquad\lambda_J(k)=2\sum_i J_i(1-\cos k_i). ( L J , M f ) ( x ) = i ∑ J i [ 2 f ( x ) − f ( x + e i ) − f ( x − e i )] , λ J ( k ) = 2 i ∑ J i ( 1 − cos k i ) .
Write K M = ( 2 π / M ) { 0 , … , M − 1 } d \mathcal{K}_M=(2\pi/M)\{0,\ldots,M-1\}^d K M = ( 2 π / M ) { 0 , … , M − 1 } d , with frequencies modulo 2 π 2\pi 2 π , and V = M d V=M^d V = M d . Define
G J ( x , y ) = ∫ [ − π , π ] d e i k ⋅ ( x − y ) λ J ( k ) d k ( 2 π ) d , β c , J = G J ( 0 , 0 ) . G_J(x,y)=\int_{[-\pi,\pi]^d}\frac{e^{ik\cdot(x-y)}}{\lambda_J(k)}\frac{dk}{(2\pi)^d},\qquad\beta_{c,J}=G_J(0,0). G J ( x , y ) = ∫ [ − π , π ] d λ J ( k ) e ik ⋅ ( x − y ) ( 2 π ) d d k , β c , J = G J ( 0 , 0 ) .
These integrals are finite for d ≥ 3 d\ge3 d ≥ 3 . Here β c , J \beta_{c,J} β c , J is the critical inverse temperature of the infinite-volume mean spherical model. All temperature comparisons below use this spherical threshold. The transition point of the fixed-N N N spin model is a separate quantity.
Let G J , M 0 G^0_{J,M} G J , M 0 be the inverse of L J , M L_{J,M} L J , M on zero-sum functions, extended as zero on constants. For a real finitely supported zero-sum source a a a , set
a ^ ( k ) = ∑ x a x e i k ⋅ x , Q M 0 ( a ) = 1 V ∑ k ≠ 0 ∣ a ^ ( k ) ∣ 2 λ J ( k ) , Q ∞ ( a ) = ∫ ∣ a ^ ( k ) ∣ 2 λ J ( k ) d k ( 2 π ) d . \widehat{a}(k)=\sum_x a_xe^{ik\cdot x},\qquad Q^0_M(a)=\frac{1}{V}\sum_{k\ne0}\frac{|\widehat{a}(k)|^2}{\lambda_J(k)},\qquad Q_\infty(a)=\int\frac{|\widehat{a}(k)|^2}{\lambda_J(k)}\frac{dk}{(2\pi)^d}. a ( k ) = x ∑ a x e ik ⋅ x , Q M 0 ( a ) = V 1 k = 0 ∑ λ J ( k ) ∣ a ( k ) ∣ 2 , Q ∞ ( a ) = ∫ λ J ( k ) ∣ a ( k ) ∣ 2 ( 2 π ) d d k .
For M M M large enough that the support embeds injectively, we view each finitely supported source a a a as a zero-sum source on Λ M \Lambda_M Λ M by periodic identification.
Fix a nonempty set of distinct sites U = { x 1 , … , x ℓ } ⊂ Z d \mathcal{U}=\{x_1,\ldots,x_\ell\}\subset\mathbb{Z}^d U = { x 1 , … , x ℓ } ⊂ Z d . The law of the first p p p spin coordinates at these sites is ν M , N , β , J U , p \nu^{\mathcal{U},p}_{M,N,\beta,J} ν M , N , β , J U , p , on R ℓ × p \mathbb{R}^{\ell\times p} R ℓ × p . All Wasserstein distances use the unnormalized Frobenius norm (no division by p \sqrt{p} p ; see § 2.1 ):
W 1 ( μ , ν ) = inf { E ∥ X − Y ∥ F : L ( X ) = μ , L ( Y ) = ν } . W_1(\mu,\nu)=\inf\{\mathbb{E}\|X-Y\|_F:\mathcal{L}(X)=\mu,\mathcal{L}(Y)=\nu\}. W 1 ( μ , ν ) = inf { E ∥ X − Y ∥ F : L ( X ) = μ , L ( Y ) = ν } .
For β ≥ β c , J \beta\ge\beta_{c,J} β ≥ β c , J , let γ β , J U , p \gamma^{\mathcal{U},p}_{\beta,J} γ β , J U , p have p p p independent centred Gaussian columns with covariance
Σ β , J ( x , y ) = G J ( x , y ) β + 1 − β c , J β , x , y ∈ U . (2) \Sigma_{\beta,J}(x,y)=\frac{G_J(x,y)}{\beta}+1-\frac{\beta_{c,J}}{\beta},\qquad x,y\in\mathcal{U}. \tag*{(2)} Σ β , J ( x , y ) = β G J ( x , y ) + 1 − β β c , J , x , y ∈ U . (2)
Its diagonal is one. Each column is β − 1 / 2 \beta^{-1/2} β − 1/2 times an anisotropic GFF plus an independent Gaussian constant of variance 1 − β c , J / β 1-\beta_{c,J}/\beta 1 − β c , J / β . The spatially constant term belongs to the limiting ordered-state covariance.
Uniform statements fix d , U d,\mathcal{U} d , U , 0 < j 0 ≤ j 1 < ∞ 0<j_0\le j_1<\infty 0 < j 0 ≤ j 1 < ∞ , and 0 < b < B < ∞ 0<b<B<\infty 0 < b < B < ∞ , and range over
J ∈ [ j 0 , j 1 ] d , β ∈ [ b , B ] , β ≥ β c , J . (3) J\in[j_0,j_1]^d,\qquad\beta\in[b,B],\qquad\beta\ge\beta_{c,J}. \tag*{(3)} J ∈ [ j 0 , j 1 ] d , β ∈ [ b , B ] , β ≥ β c , J . (3)
We write f ≍ g f\asymp g f ≍ g for two-sided bounds with constants uniform over (3) . Set
τ = β − β c , J , a d ( M ) = { M , d = 3 , log M , d = 4 , 1 , d ≥ 5 , t d ( M ) = a d ( M ) V . \tau=\beta-\beta_{c,J},\qquad a_d(M)=
\begin{cases}
M,&d=3,\\
\log M,&d=4,\\
1,&d\ge5,
\end{cases}
\qquad t_d(M)=\sqrt{\frac{a_d(M)}{V}}. τ = β − β c , J , a d ( M ) = ⎩ ⎨ ⎧ M , log M , 1 , d = 3 , d = 4 , d ≥ 5 , t d ( M ) = V a d ( M ) .
The positive shift s M = s M , β , J > 0 s_M=s_{M,\beta,J}>0 s M = s M , β , J > 0 solves
1 V ∑ k ∈ K M 1 λ J ( k ) + s M = β . (4) \frac{1}{V}\sum_{k\in\mathcal{K}_M}\frac{1}{\lambda_J(k)+s_M}=\beta. \tag*{(4)} V 1 k ∈ K M ∑ λ J ( k ) + s M 1 = β . (4)
The left side is strictly decreasing in s M s_M s M , diverges as s M ↓ 0 s_M\downarrow0 s M ↓ 0 and tends to zero as s M → ∞ s_M\to\infty s M → ∞ , so the solution exists and is unique. Following common usage for the spherical model we call s M s_M s M the mass, although it is the square of a mass in the convention L + m 2 I L + m^2 I L + m 2 I . The equation fixes an auxiliary finite-volume Gaussian covariance β − 1 ( L J , M + s M I ) − 1 \beta^{-1}(L_{J,M} + s_M I)^{-1} β − 1 ( L J , M + s M I ) − 1 with unit diagonal. This auxiliary Gaussian is the zero-field mean spherical model on Λ M \Lambda_M Λ M : the spherical constraint is imposed on the mean squared spin in the finite box, as in the comparison model of [29 ] . The comparison target in the theorems below remains the infinite-lattice covariance (2) restricted to U \mathcal{U} U .
Theorem 2.1 (Uniform local joint approximation). There are C , M 0 < ∞ C, M_0 < \infty C , M 0 < ∞ , depending only on the fixed data above, such that for every integer M ≥ M 0 M \ge M_0 M ≥ M 0 , N ≥ max ( 4 , ℓ ) N \ge\max(4,\ell) N ≥ max ( 4 , ℓ ) , and 1 ≤ p ≤ N 1 \le p \le N 1 ≤ p ≤ N ,
W 1 ( ν M , N , β , J U , p , γ β , J U , p ) ≤ C p [ N − 1 / 2 + s M ] . (5) W_1\left(\nu^{\mathcal{U},p}_{M,N,\beta,J},\gamma^{\mathcal{U},p}_{\beta,J}\right) \le C\sqrt{p}\left[N^{-1/2}+s_M\right]. \tag*{(5)} W 1 ( ν M , N , β , J U , p , γ β , J U , p ) ≤ C p [ N − 1/2 + s M ] . (5) Here M 0 M_0 M 0 is chosen so that U \mathcal{U} U embeds injectively in the torus.
Proposition 2.2 (Mass orders). The mass satisfies
s M ≍ 1 V { τ + t d ( M ) } s_M \asymp\frac{1}{\sqrt{V\{\tau+t_d(M)\}}} s M ≍ V { τ + t d ( M )} 1 uniformly on the parameter ranges of Theorem 2.1. At criticality its order is
ρ d ( M ) = { M − 2 , d = 3 , M − 2 / log M , d = 4 , M − d / 2 , d ≥ 5. \rho_d(M)=
\begin{cases}
M^{-2}, & d=3,\\
M^{-2}/\sqrt{\log M}, & d=4,\\
M^{-d/2}, & d\ge5.
\end{cases} ρ d ( M ) = ⎩ ⎨ ⎧ M − 2 , M − 2 / log M , M − d /2 , d = 3 , d = 4 , d ≥ 5. If τ ≥ κ > 0 \tau\ge\kappa> 0 τ ≥ κ > 0 , it is bounded by C κ M − d C_\kappa M^{-d} C κ M − d .
It is proved in Appendix § 9 as Proposition A.4.
Theorem 2.3 (Finite-spin lower bounds and exact coordinate criterion). For U 0 = { 0 , e 1 , … , e d } \mathcal{U}_0=\{0,e_1,\ldots,e_d\} U 0 = { 0 , e 1 , … , e d } , all integers M ≥ 4 M\ge4 M ≥ 4 , N ≥ 1 N\ge1 N ≥ 1 and 1 ≤ p ≤ N 1\le p\le N 1 ≤ p ≤ N , under (3),
W 1 ( ν M , N , β , J U 0 , p , γ β , J U 0 , p ) ≥ c p / M d . W_1\left(\nu^{\mathcal{U}_0,p}_{M,N,\beta,J},\gamma^{\mathcal{U}_0,p}_{\beta,J}\right)\ge c\sqrt{p}/M^d. W 1 ( ν M , N , β , J U 0 , p , γ β , J U 0 , p ) ≥ c p / M d . For N ≥ 4 N\ge4 N ≥ 4 , the same distance also satisfies
W 1 ( ν M , N , β , J U 0 , p , γ β , J U 0 , p ) ≥ c p s M − C p / N . W_1\left(\nu^{\mathcal{U}_0,p}_{M,N,\beta,J},\gamma^{\mathcal{U}_0,p}_{\beta,J}\right)\ge c\sqrt{p}s_M-C\sqrt{p/N}. W 1 ( ν M , N , β , J U 0 , p , γ β , J U 0 , p ) ≥ c p s M − C p / N . The constants are uniform on the compact parameter ranges. For the star or any fixed set containing a translated star, along M n , N n → ∞ M_n,N_n\to\infty M n , N n → ∞ with parameters satisfying (3),
W 1 ( ν M n , N n , β n , J n U , p n , γ β n , J n U , p n ) → 0 ⟺ p n = o ( N n ) and p n s M n , β n , J n → 0. W_1\left(\nu^{\mathcal{U},p_n}_{M_n,N_n,\beta_n,J_n},\gamma^{\mathcal{U},p_n}_{\beta_n,J_n}\right)\to0
\quad\Longleftrightarrow\quad
p_n=o(N_n)\ \text{and}\ \sqrt{p_n}\,s_{M_n,\beta_n,J_n}\to0. W 1 ( ν M n , N n , β n , J n U , p n , γ β n , J n U , p n ) → 0 ⟺ p n = o ( N n ) and p n s M n , β n , J n → 0. At criticality the volume conditions are p = o ( M 4 ) p=o(M^4) p = o ( M 4 ) for d = 3 d=3 d = 3 , p = o ( M 4 log M ) p=o(M^4\log M) p = o ( M 4 log M ) for d = 4 d=4 d = 4 , and p = o ( M d ) p=o(M^d) p = o ( M d ) for d ≥ 5 d\ge5 d ≥ 5 . Under β − β c , J ≥ κ > 0 \beta-\beta_{c,J}\ge\kappa>0 β − β c , J ≥ κ > 0 , the condition is p = o ( M 2 d ) p=o(M^{2d}) p = o ( M 2 d ) . Each of these conditions is accompanied by p = o ( N ) p=o(N) p = o ( N ) .
Corollary 2.4 (Periodic accumulation states). For each N N N , choose any local accumulation state μ N , β , J \mu_{N,\beta,J} μ N , β , J of the periodic laws as M → ∞ M\to\infty M → ∞ , that is, any local weak subsequential limit of the periodized torus laws. Uniformly over all such choices and under (3),
W 1 ( μ N , β , J , γ β , J ) ≤ C p / N , N ≥ 1 , 1 ≤ p ≤ N . W_1\left(\mu_{N,\beta,J},\gamma_{\beta,J}\right)\le C\sqrt{p/N},\qquad N\ge1,\quad1\le p\le N. W 1 ( μ N , β , J , γ β , J ) ≤ C p / N , N ≥ 1 , 1 ≤ p ≤ N . For any sequence 1 ≤ p N ≤ N 1\le p_N\le N 1 ≤ p N ≤ N (and any parameter sequence in (3)), the displayed Wasserstein distance tends to zero as N → ∞ N\to\infty N → ∞ if and only if p N = o ( N ) p_N=o(N) p N = o ( N ) .
Corollary 2.5 (Interchange of limits). Fix U \mathcal{U} U and p p p , and write W 1 = W 1 ( ν M , N , β , J U , p , γ β , J U , p ) W_1 = W_1(\nu_{M,N,\beta,J}^{\mathcal{U},p}, \gamma_{\beta,J}^{\mathcal{U},p}) W 1 = W 1 ( ν M , N , β , J U , p , γ β , J U , p ) . Uniformly over (3) :
(i) sup W 1 → 0 \sup W_1 \to0 sup W 1 → 0 as min ( M , N ) → ∞ \min(M,N) \to\infty min ( M , N ) → ∞ , the supremum being over (3) ; in particular W 1 → 0 W_1 \to0 W 1 → 0 along every joint path;
(ii) lim M → ∞ lim sup N → ∞ W 1 = 0 \lim_{M\to\infty}\limsup_{N\to\infty} W_1 = 0 lim M → ∞ lim sup N → ∞ W 1 = 0 ;
(iii) lim N → ∞ sup W 1 ( μ N , β , J U , p , γ β , J U , p ) = 0 \lim_{N\to\infty}\sup W_1(\mu_{N,\beta,J}^{\mathcal{U},p},\gamma_{\beta,J}^{\mathcal{U},p}) = 0 lim N → ∞ sup W 1 ( μ N , β , J U , p , γ β , J U , p ) = 0 , the supremum being over all choices of periodic accumulation states.
Proof. For fixed p p p , the bound of Theorem 2.1 is uniform in M ≥ M 0 M \ge M_0 M ≥ M 0 and N ≥ max ( 4 , ℓ ) N \ge\max(4,\ell) N ≥ max ( 4 , ℓ ) , and s M ≤ C ρ d ( M ) → 0 s_M \le C\rho_d(M) \to0 s M ≤ C ρ d ( M ) → 0 uniformly by Proposition A.4 . This gives (i), and also (ii), since lim sup N → ∞ W 1 ≤ C p s M \limsup_{N\to\infty} W_1 \le C\sqrt{p}\,s_M lim sup N → ∞ W 1 ≤ C p s M . Corollary 2.4 gives (iii).
Remark 2.6 (State and observation scope). All spatial windows are fixed, and the periodic accumulation states retain global O ( N ) O(N) O ( N ) invariance. Fixed-N N N uniqueness, directional pure states, arbitrary boundary conditions and growing spatial windows are outside the theorem. For β < β c , J \beta< \beta_{c,J} β < β c , J see § 8 .
Coordinate scales
Table 1 collects the mass orders of Proposition 2.2 and the resulting ranges of Theorem 2.3 .
Table 1. Mass orders and necessary and sufficient coordinate ranges for the finite-spin laws on fixed star-containing windows. All four ranges are sufficient on any fixed spatial window.
The criterion also covers temperatures varying through the crossover, where τ ≍ t d ( M ) \tau\asymp t_d(M) τ ≍ t d ( M ) . In that case the volume condition is p / [ V { τ + t d ( M ) } ] → 0 \sqrt{p}/[V\{\tau+t_d(M)\}] \to0 p / [ V { τ + t d ( M )}] → 0 . Each finite-( M , N , β ) (M,N,\beta) ( M , N , β ) law is compared with the target γ β , J U , p \gamma_{\beta,J}^{\mathcal{U},p} γ β , J U , p at the same β \beta β ; the variance of the Gaussian constant tends to zero when τ → 0 \tau\to0 τ → 0 . The mass orders and this interpolation are proved in Appendix § 9 .
In the unnormalized Frobenius metric, W 1 W_1 W 1 controls every 1-Lipschitz function of the whole p p p -column block, for example the Euclidean norm of the observed part of one spin, which feels the hard length constraint. Dividing by p \sqrt{p} p would give W 1 / p ≤ C ( N − 1 / 2 + s M ) W_1/\sqrt{p} \le C(N^{-1/2}+s_M) W 1 / p ≤ C ( N − 1/2 + s M ) , in which p p p no longer appears; for p ≍ c N p\asymp cN p ≍ c N this normalized distance is O ( N − 1 / 2 ) O(N^{-1/2}) O ( N − 1/2 ) in accumulation states, whereas the unnormalized distance stays bounded below by a positive constant (Proposition 6.1 ).
Remark 2.7 (Auxiliary Gaussian reference). Let γ M U , p \gamma_M^{\mathcal{U},p} γ M U , p have p p p independent centred Gaussian columns with covariance Σ M = β − 1 ( L J , M + s M I ) − 1 \Sigma_M=\beta^{-1}(L_{J,M}+s_MI)^{-1} Σ M = β − 1 ( L J , M + s M I ) − 1 restricted to U \mathcal{U} U . For isotropic couplings and fixed p p p , it is also the limit in law of ν M , N , β , J U , p \nu_{M,N,\beta,J}^{\mathcal{U},p} ν M , N , β , J U , p as N → ∞ N\to\infty N → ∞ at fixed M M M [1 Theorem 4.1] . Appendix A.3 shows that, on every fixed window U \mathcal{U} U , for every p ≥ 1 p\ge1 p ≥ 1 and every M M M large enough that U \mathcal{U} U embeds, with C U C_{\mathcal{U}} C U independent of p p p , M M M and the parameters in (3) ,
W 1 ( γ M U , p , γ β , J U , p ) ≤ C U p s M . (6) W_1(\gamma_M^{\mathcal{U},p},\gamma_{\beta,J}^{\mathcal{U},p}) \le C_{\mathcal{U}}\sqrt{p}\,s_M. \tag*{(6)} W 1 ( γ M U , p , γ β , J U , p ) ≤ C U p s M . (6)
For the star U 0 U_0 U 0 and every p ≥ 1 p \ge1 p ≥ 1 ,
W 1 ( γ M U 0 , p , γ β , J U 0 , p ) ≥ p s M 4 π j . (7) W_1\left(\gamma_M^{U_0,p},\gamma_{\beta,J}^{U_0,p}\right) \ge\frac{\sqrt{p}\,s_M}{4\sqrt{\pi}\,j}. \tag*{(7)} W 1 ( γ M U 0 , p , γ β , J U 0 , p ) ≥ 4 π j p s M . (7)
The lower bound also holds on any fixed window containing a translated star. Thus on such windows the volume term p s M \sqrt{p}\,s_M p s M in Theorem 2.1 is of the order of the distance between the auxiliary reference γ M U , p \gamma_M^{U,p} γ M U , p and the target. By the triangle inequality, Theorem 2.1 and (6) give
W 1 ( ν M , N , β , J U , p , γ M U , p ) ≤ C p ( N − 1 / 2 + s M ) ; W_1\left(\nu_{M,N,\beta,J}^{U,p},\gamma_M^{U,p}\right) \le C\sqrt{p}\left(N^{-1/2}+s_M\right); W 1 ( ν M , N , β , J U , p , γ M U , p ) ≤ C p ( N − 1/2 + s M ) ;
whether s M s_M s M can be removed is discussed in § 8 .
Stronger ordered estimates
The general bound has an N − 1 / 2 N^{-1/2} N − 1/2 spin-dimensional contribution for fixed p p p . At a positive ordered gap this can be improved. The following theorem keeps the dependence on p p p explicit.
Theorem 2.8 (Stronger ordered rate with explicit coordinate dependence). Fix κ > 0 \kappa> 0 κ > 0 . Under (3) and β − β c , J ≥ κ \beta-\beta_{c,J} \ge\kappa β − β c , J ≥ κ , for sufficiently large integer M M M , N ≥ max ( 4 , ℓ ) N \ge\max(4,\ell) N ≥ max ( 4 , ℓ ) and 1 ≤ p ≤ N 1 \le p \le N 1 ≤ p ≤ N ,
W 1 ( ν M , N , β , J U , p , γ β , J U , p ) ≤ C U , κ [ min { p N , p ψ d ( N ) } + p M − d ] , W_1\left(\nu_{M,N,\beta,J}^{U,p},\gamma_{\beta,J}^{U,p}\right) \le C_{U,\kappa}\left[\min\left\{\sqrt{\frac{p}{N}},p\psi_d(N)\right\}+\sqrt{p}\,M^{-d}\right], W 1 ( ν M , N , β , J U , p , γ β , J U , p ) ≤ C U , κ [ min { N p , p ψ d ( N ) } + p M − d ] , where, with L N = log ( e N ) L_N=\log(eN) L N = log ( e N ) ,
ψ d ( N ) = { ( L N / N ) 3 / 4 , d = 3 , L N / N , d = 4 , N − 1 , d ≥ 5. \psi_d(N)=
\begin{cases}
(L_N/N)^{3/4}, & d=3,\\
L_N/N, & d=4,\\
N^{-1}, & d\ge5.
\end{cases} ψ d ( N ) = ⎩ ⎨ ⎧ ( L N / N ) 3/4 , L N / N , N − 1 , d = 3 , d = 4 , d ≥ 5. For every choice of periodic accumulation states, the same bound holds with the volume term omitted. The constant is independent of p p p , N N N , M M M , β \beta β , J J J within these ranges. For each fixed p p p , this gives C U , p , κ { ψ d ( N ) + M − d } C_{U,p,\kappa}\{\psi_d(N)+M^{-d}\} C U , p , κ { ψ d ( N ) + M − d } . For d ≥ 5 d \ge5 d ≥ 5 , the bound simplifies to
C U , κ ( p N + p M − d ) . C_{U,\kappa}\left(\frac{p}{N}+\sqrt{p}\,M^{-d}\right). C U , κ ( N p + p M − d ) . In these dimensions, for fixed p p p and a window containing a translated nearest-neighbour star,
c U , p , κ ( N − 1 + M − d ) ≤ W 1 ( ν M , N , β , J U , p , γ β , J U , p ) ≤ C U , p , κ ( N − 1 + M − d ) . c_{U,p,\kappa}\left(N^{-1}+M^{-d}\right) \le W_1\left(\nu_{M,N,\beta,J}^{U,p},\gamma_{\beta,J}^{U,p}\right) \le C_{U,p,\kappa}\left(N^{-1}+M^{-d}\right). c U , p , κ ( N − 1 + M − d ) ≤ W 1 ( ν M , N , β , J U , p , γ β , J U , p ) ≤ C U , p , κ ( N − 1 + M − d ) . For d ≥ 5 d \ge5 d ≥ 5 , on every nonempty fixed window and for every choice of periodic accumulation states, the fixed-p p p distance is bounded above and below by positive constant multiples of N − 1 N^{-1} N − 1 .
Remark 2.9 (The one-site lower bound). The fixed-p p p lower bound of order N − 1 N^{-1} N − 1 in Theorem 2.8 comes from a one-site norm test that uses only the uniform sphere marginal of a single spin and holds for every β \beta β . So for d ≥ 5 d \ge5 d ≥ 5 the theorem shows that, apart from the volume term M − d M^{-d} M − d , the interaction does not worsen the one-site rate N − 1 N^{-1} N − 1 . At criticality the test gives only the lower bound
W 1 ≥ χ ˉ p ( N χ ˉ N − 1 ) = χ ˉ p { 1 4 N + O ( N − 2 ) } , W_1 \ge\bar{\chi}_p\left(\frac{\sqrt{N}}{\bar{\chi}_N}-1\right)=\bar{\chi}_p\left\{\frac{1}{4N}+O\left(N^{-2}\right)\right\}, W 1 ≥ χ ˉ p ( χ ˉ N N − 1 ) = χ ˉ p { 4 N 1 + O ( N − 2 ) } ,
of order p / N \sqrt{p}/N p / N , where χ ˉ q \bar{\chi}_q χ ˉ q is the mean norm of a standard q q q -dimensional Gaussian vector; the critical spin-dimensional rate N − 1 / 2 N^{-1/2} N − 1/2 of Theorem 2.1 is therefore not known to be sharp.
Outline of the proofs
The pressure derivative is the J J J -weighted sum of the nearest-neighbour spin correlations per component. Gaussian domination bounds the single-coordinate variances of spin differences by their massless Gaussian values. Let δ M , N \delta_{M,N} δ M , N be the J J J -weighted sum of the nearest-neighbour variance deficits. The corresponding deficit of the finite-volume mean spherical model is s M − 1 / ( β V ) s_M - 1/(\beta V) s M − 1/ ( β V ) . Because the spin and comparison laws both have unit site variances, the difference between the deficits is determined by the difference of the pressure derivatives. This accounts for the spectral shift in the error.
§ 4 compares the two pressures with error O ( N − 1 ) O(N^{-1}) O ( N − 1 ) , uniformly in M M M . An upper secant over an interval of length h h h bounds the derivative difference from above. The pressure comparison contributes an error of order 1 / ( N h ) 1/(Nh) 1/ ( N h ) and the curvature of the comparison pressure contributes an error of order h h h . Taking h = N − 1 / 2 h=N^{-1/2} h = N − 1/2 gives
δ M , N ≤ C N − 1 / 2 + s M . (8) \delta_{M,N} \le CN^{-1/2} + s_M. \tag*{(8)} δ M , N ≤ C N − 1/2 + s M . (8)
The variance deficit for a general zero-sum source is a positive semidefinite quadratic form. Writing a spin difference as a sum along a nearest-neighbour path extends the bound to each pair in the window. Gaussian domination also controls the upper tail of the squared norm of each difference. The spin length fixes the diagonal of the local Gram matrix, so the Green-function comparison is needed only for dipoles. Their zero sum gives a quadrature error of O ( M − d ) O(M^{-d}) O ( M − d ) , which can be absorbed into s M s_M s M (Appendix § 9 ).
By rotation invariance the observed block has the law of A 1 / 2 Q N A^{1/2}Q_N A 1/2 Q N , where A A A is the local Gram matrix and Q N Q_N Q N is an independent, scaled block of a Haar orthogonal matrix. Replacing Q N Q_N Q N by a standard Gaussian matrix gives the comparison even for singular A A A . The pressure secant, squared-norm fluctuations and Haar approximation all contribute to the O ( p / N ) O(\sqrt{p/N}) O ( p / N ) error.
For the lower bounds, the test compares the norm of the observed coordinates of a nearest-neighbour difference with the norm of those coordinates at a single site. The single-site term calibrates the test and its Lipschitz constant stays bounded as p p p grows. The J J J -weighted difference between the target and spin variances is δ M , N + 1 / ( β V ) \delta_{M,N} + 1/(\beta V) δ M , N + 1/ ( β V ) , so it is at least 1 / ( β V ) 1/(\beta V) 1/ ( β V ) . A left pressure secant also bounds it below by s M − C / N s_M-C/\sqrt{N} s M − C / N for N ≥ 4 N \ge4 N ≥ 4 , even if its left endpoint lies below the critical inverse temperature. The direction witnessing the weighted shortfall can depend on the parameters. A star contains every direction that may be needed (§ 6 ).
When β − β c , J ≥ κ > 0 \beta-\beta_{c,J}\ge\kappa>0 β − β c , J ≥ κ > 0 , a sphere Ward identity sharpens the deficit estimate. The covariance expansion cancels a term equal to the deficit times the excess mean squared magnetization above its finite-volume massless reference value. A one-sided fluctuation estimate and spectral interpolation then give the improved deficit bound (Proposition 7.2 ).
The other gain comes from expanding Gaussian expectations around the mean local Gram matrix. When this mean is bounded away from singularity, the linear term averages to zero and the mixture error is bounded by C p C_p C p times the second moment of the Gram fluctuation. When N < 2 ( V + 1 ) N<2(V+1) N < 2 ( V + 1 ) , the sharper deficit estimate controls these fluctuations. When N ≥ 2 ( V + 1 ) N\ge2(V+1) N ≥ 2 ( V + 1 ) , the full Gram density controls the variance and a second Ward identity controls the covariance bias. The split keeps the volume term at p / V \sqrt{p}/V p / V . Using the deficit estimate alone would leave a p / V p/V p / V term when N N N is much larger than V V V (§ 7 ).
Gaussian domination and the pressure comparison
Proposition 4.1 gives Gaussian domination for every M ≥ 4 M\ge4 M ≥ 4 , including odd M M M ; Proposition 4.2 compares the pressure with the Gaussian comparison pressure; and Lemma 4.3 turns this comparison into the deficit bound (24).
Gaussian domination
The following directional-coupling form of the transfer-operator argument of Fröhlich, Simon and Spencer [15 pp. 83–84] applies to every integer M ≥ 4 M \ge4 M ≥ 4 .
Proposition 4.1 (Vector-source bound). Let β > 0 \beta> 0 β > 0 , and let M ≥ 4 M \ge4 M ≥ 4 , even or odd, and N ≥ 1 N \ge1 N ≥ 1 be integers. For b x ∈ R N \mathbf{b}_x \in\mathbb{R}^N b x ∈ R N with ∑ x b x = 0 \sum_x \mathbf{b}_x = 0 ∑ x b x = 0 , the Gibbs law (1) satisfies
E exp { ∑ x b x ⋅ S x } ≤ exp { 1 2 β ∑ α = 1 N ( b α , G J , M 0 b α ) } . (9) \mathbb{E}\exp\left\{\sum_x \mathbf{b}_x \cdot S_x\right\} \le\exp\left\{\frac{1}{2\beta}\sum_{\alpha=1}^{N}\left(\mathbf{b}^{\alpha},G_{J,M}^{0}\mathbf{b}^{\alpha}\right)\right\}. \tag*{(9)} E exp { x ∑ b x ⋅ S x } ≤ exp { 2 β 1 α = 1 ∑ N ( b α , G J , M 0 b α ) } . (9) In particular, for a real zero-sum source a a a and v ∈ R N v \in\mathbb{R}^N v ∈ R N ,
E exp { v ⋅ ∑ x a x S x } ≤ exp { ∣ v ∣ 2 Q M 0 ( a ) / ( 2 β ) } . (10) \mathbb{E}\exp\left\{v\cdot\sum_x a_xS_x\right\}\leq\exp\{|v|^2Q_M^0(a)/(2\beta)\}. \tag*{(10)} E exp { v ⋅ x ∑ a x S x } ≤ exp { ∣ v ∣ 2 Q M 0 ( a ) / ( 2 β )} . (10)
Proof of Proposition 4.1. For arbitrary bond vectors h x , i h_{x,i} h x , i , define
Z ( h ) = ∫ exp { − β 2 ∑ x , i J i ∣ S x − S x + e i − h x , i ∣ 2 } ∏ x σ N ( d S x ) . \mathcal{Z}(h)=\int\exp\left\{-\frac{\beta}{2}\sum_{x,i}J_i\left|S_x-S_{x+e_i}-h_{x,i}\right|^2\right\}\prod_x\sigma_N(dS_x). Z ( h ) = ∫ exp { − 2 β x , i ∑ J i ∣ S x − S x + e i − h x , i ∣ 2 } x ∏ σ N ( d S x ) .
We first show Z ( h ) ≤ Z ( 0 ) \mathcal{Z}(h)\leq\mathcal{Z}(0) Z ( h ) ≤ Z ( 0 ) . Temporarily replace the prior by the same smooth nonnegative compactly supported probability density at each site. Fix direction q q q . A transverse layer has variable u ∈ R N M d − 1 u\in\mathbb{R}^{NM^{d-1}} u ∈ R N M d − 1 . Let F i F_i F i denote multiplication by the product of the layer priors and transverse bond factors; this product is a bounded, nonnegative, integrable function. Collect the source between layers i i i and i + 1 i+1 i + 1 into a i a_i a i . On L 2 ( R N M d − 1 ) L^2(\mathbb{R}^{NM^{d-1}}) L 2 ( R N M d − 1 ) , set
T a ( u , v ) = e − β J q ∣ u − v − a ∣ 2 / 2 , U a f ( u ) = f ( u − a ) . T_a(u,v)=e^{-\beta J_q|u-v-a|^2/2},\qquad U_af(u)=f(u-a). T a ( u , v ) = e − β J q ∣ u − v − a ∣ 2 /2 , U a f ( u ) = f ( u − a ) .
Gaussian convolution T 0 T_0 T 0 is bounded and positive by its Fourier multiplier; translation U a U_a U a is unitary and commutes with T 0 T_0 T 0 and its square root. Thus T a = T 0 1 / 2 U a T 0 1 / 2 T_a=T_0^{1/2}U_aT_0^{1/2} T a = T 0 1/2 U a T 0 1/2 . The positive operators B i = T 0 1 / 2 F i T 0 1 / 2 B_i=T_0^{1/2}F_iT_0^{1/2} B i = T 0 1/2 F i T 0 1/2 are trace class: F i 1 / 2 T 0 1 / 2 F_i^{1/2}T_0^{1/2} F i 1/2 T 0 1/2 is Hilbert–Schmidt, since F i F_i F i is integrable and the Gaussian half-convolution kernel is square integrable. Cyclicity and the integral kernels give
Z ( h ) = Tr ( ∏ i = 1 M B i U a i ) ≤ ∏ i = 1 M ∥ B i U a i ∥ S M = ∏ i = 1 M [ Tr ( B i M ) ] 1 / M , \mathcal{Z}(h)=\operatorname{Tr}\left(\prod_{i=1}^{M}B_iU_{a_i}\right)\leq\prod_{i=1}^{M}\|B_iU_{a_i}\|_{S_M}=\prod_{i=1}^{M}\left[\operatorname{Tr}(B_i^M)\right]^{1/M}, Z ( h ) = Tr ( i = 1 ∏ M B i U a i ) ≤ i = 1 ∏ M ∥ B i U a i ∥ S M = i = 1 ∏ M [ Tr ( B i M ) ] 1/ M ,
by Schatten Hölder with exponent M M M . Each trace is the partition function obtained by setting all direction-q q q sources to zero and copying layer i i i ’s transverse sources to all layers. Apply this operation in directions 1 , … , d 1,\ldots,d 1 , … , d . Copying never reintroduces a source in a direction already set to zero. Every final branch has source zero, and the geometric weights sum to one. This proves Z ( h ) ≤ Z ( 0 ) \mathcal{Z}(h)\leq\mathcal{Z}(0) Z ( h ) ≤ Z ( 0 ) . The argument uses homogeneous directional couplings and identical priors; the weights J i J_i J i may differ and M M M may be odd.
Choose smooth priors supported in one compact set and weakly converging to the sphere probability. At fixed M , N , h M,N,h M , N , h , the continuous integrand is bounded there, so product weak convergence restores the sphere inequality. Because the spin length is fixed, the law normalized by Z ( 0 ) \mathcal{Z}(0) Z ( 0 ) coincides with the Gibbs law (1). Finally put f = β − 1 G J , M 0 b f=\beta^{-1}G^0_{J,M}b f = β − 1 G J , M 0 b coordinatewise and h x , i = f x − f x + e i h_{x,i}=f_x-f_{x+e_i} h x , i = f x − f x + e i . Summation by parts gives
β ∑ x , i J i h x , i ⋅ ( S x − S x + e i ) = ∑ x b x ⋅ S x , β 2 ∑ x , i J i ∣ h x , i ∣ 2 = 1 2 β ∑ α ⟨ b α , G J , M 0 b α ⟩ . \beta\sum_{x,i}J_i h_{x,i}\cdot(S_x-S_{x+e_i})=\sum_x b_x\cdot S_x,\qquad\frac{\beta}{2}\sum_{x,i}J_i|h_{x,i}|^2=\frac{1}{2\beta}\sum_\alpha\langle b^\alpha,G^0_{J,M}b^\alpha\rangle. β x , i ∑ J i h x , i ⋅ ( S x − S x + e i ) = x ∑ b x ⋅ S x , 2 β x , i ∑ J i ∣ h x , i ∣ 2 = 2 β 1 α ∑ ⟨ b α , G J , M 0 b α ⟩ .
Expanding the shifted partition proves (9) ; take b x = a x v b_x=a_xv b x = a x v to obtain (10) .
Uniform pressure comparison
Write P M , N = ( N V ) − 1 log Z M , N , β , J P_{M,N}=(NV)^{-1}\log Z_{M,N,\beta,J} P M , N = ( N V ) − 1 log Z M , N , β , J , and define the Gaussian comparison pressure by
P M c o m p ( β ) = β j + β s M 2 − 1 2 − 1 2 log β − 1 2 V log det ( L J , M + s M I ) . (11) P^{\mathrm{comp}}_M(\beta)=\beta j+\frac{\beta s_M}{2}-\frac{1}{2}-\frac{1}{2}\log\beta-\frac{1}{2V}\log\det(L_{J,M}+s_MI). \tag*{(11)} P M comp ( β ) = β j + 2 β s M − 2 1 − 2 1 log β − 2 V 1 log det ( L J , M + s M I ) . (11)
For this comparison, (4) defines s M s_M s M for every β > 0 \beta>0 β > 0 . The next proposition is therefore available on both sides of the limiting critical inverse temperature, as needed for the pressure secants. The sphere prior is normalized to a probability measure. Keeping this normalization is essential: the radial ( log N ) / N (\log N)/N ( log N ) / N terms in the proof cancel, leaving an O ( N − 1 ) O(N^{-1}) O ( N − 1 ) pressure error. The construction below follows Shcherbina’s auxiliary radial Hamiltonian [29 Section 3] ; tracking the exact polar Jacobian in this way improves her bound of order ( log N ) / N (\log N)/N ( log N ) / N to O ( N − 1 ) O(N^{-1}) O ( N − 1 ) with explicit constants.
Proposition 4.2 (Uniform pressure comparison). For every B ′ > 0 B'>0 B ′ > 0 , all 0 < β ≤ B ′ 0<\beta\le B' 0 < β ≤ B ′ , all integers M ≥ 4 M\ge4 M ≥ 4 , and all N ≥ 4 N\ge4 N ≥ 4 ,
∣ P M , N ( β ) − P M c o m p ( β ) ∣ ≤ ε N ( B ′ , J ) : = 8 ( 1 + 4 B ′ j ) + 3 / 2 N . (12) \left|P_{M,N}(\beta)-P^{\mathrm{comp}}_M(\beta)\right|\le\varepsilon_N(B',J):=\frac{8(1+4B'j)+3/2}{N}. \tag*{(12)} ∣ P M , N ( β ) − P M comp ( β ) ∣ ≤ ε N ( B ′ , J ) := N 8 ( 1 + 4 B ′ j ) + 3/2 . (12) In particular this estimate is uniform in volume, and is uniform for J J J in any positive compact set.
Proof. Put H = L J , M + s M I \mathcal{H}=L_{J,M}+s_MI H = L J , M + s M I , z = 2 j + s M z=2j+s_M z = 2 j + s M , Δ = s M + 4 j \Delta=s_M+4j Δ = s M + 4 j , and t = β Δ t=\beta\Delta t = β Δ . The absolute row sum of H \mathcal{H} H is Δ \Delta Δ . Since 0 ≤ λ J ≤ 4 j 0\le\lambda_J\le4j 0 ≤ λ J ≤ 4 j , the mass equation gives
1 ≤ t ≤ T : = 1 + 4 B ′ j . (13) 1\le t\le T:=1+4B'j. \tag*{(13)} 1 ≤ t ≤ T := 1 + 4 B ′ j . (13)
For X x = ρ x S x X_x=\rho_xS_x X x = ρ x S x , with ρ x > 0 \rho_x>0 ρ x > 0 and ∣ S x ∣ 2 = N |S_x|^2=N ∣ S x ∣ 2 = N , the exact polar Jacobian relative to the probability prior is
d X x = c N ρ x N − 1 d ρ x σ N ( d S x ) , c N = 2 π N / 2 N N / 2 Γ ( N / 2 ) . (14) dX_x=c_N\rho_x^{N-1}\,d\rho_x\,\sigma_N(dS_x),\qquad c_N=\frac{2\pi^{N/2}N^{N/2}}{\Gamma(N/2)}. \tag*{(14)} d X x = c N ρ x N − 1 d ρ x σ N ( d S x ) , c N = Γ ( N /2 ) 2 π N /2 N N /2 . (14)
The Gaussian Hamiltonian and partition function are
H G ( X ) = 1 2 ∑ x , y H x y X x ⋅ X y , Z G = ( 2 π / β ) N V / 2 det ( H ) − N / 2 . H_G(X)=\frac{1}{2}\sum_{x,y}\mathcal{H}_{xy}X_x\cdot X_y,\qquad Z_G=(2\pi/\beta)^{NV/2}\det(\mathcal{H})^{-N/2}. H G ( X ) = 2 1 x , y ∑ H x y X x ⋅ X y , Z G = ( 2 π / β ) N V /2 det ( H ) − N /2 .
Under its normalized law the N N N coordinate fields are independent, with covariance Σ G = β − 1 H − 1 \Sigma^G=\beta^{-1}\mathcal{H}^{-1} Σ G = β − 1 H − 1 ; its restriction to a window U U U is the covariance Σ M \Sigma_M Σ M of Remark 2.7. Translation invariance and (4) imply Σ x x G = 1 \Sigma^G_{xx}=1 Σ xx G = 1 , hence ∣ Σ x y G ∣ ≤ 1 |\Sigma^G_{xy}|\le1 ∣ Σ x y G ∣ ≤ 1 .
For angular configurations set
q x = 1 N ∑ y H x y S x ⋅ S y , ∣ q x ∣ ≤ Δ , q_x=\frac{1}{N}\sum_y\mathcal{H}_{xy}S_x\cdot S_y,\qquad|q_x|\le\Delta, q x = N 1 y ∑ H x y S x ⋅ S y , ∣ q x ∣ ≤ Δ ,
I N ( q ) = ∫ 0 ∞ ρ N − 1 e − β N [ q ( ρ − 1 ) + Δ ( ρ − 1 ) 2 ] d ρ , Φ N ( q ) = β − 1 log I N ( q ) . I_N(q)=\int_0^\infty\rho^{N-1}e^{-\beta N[q(\rho-1)+\Delta(\rho-1)^2]}\,d\rho,\qquad\Phi_N(q)=\beta^{-1}\log I_N(q). I N ( q ) = ∫ 0 ∞ ρ N − 1 e − β N [ q ( ρ − 1 ) + Δ ( ρ − 1 ) 2 ] d ρ , Φ N ( q ) = β − 1 log I N ( q ) .
Consider the auxiliary Hamiltonian
Γ ( X ) = H G ( S ) + N ∑ x q x ( ρ x − 1 ) + Δ N ∑ x ( ρ x − 1 ) 2 + ∑ x Φ N ( q x ) . \Gamma(X)=H_G(S)+N\sum_x q_x(\rho_x-1)+\Delta N\sum_x(\rho_x-1)^2+\sum_x\Phi_N(q_x). Γ ( X ) = H G ( S ) + N x ∑ q x ( ρ x − 1 ) + Δ N x ∑ ( ρ x − 1 ) 2 + x ∑ Φ N ( q x ) .
The radial integrals cancel exactly, so, with Z A = ∫ e − β H G ( S ) ∏ x σ N ( d S x ) Z_A=\int e^{-\beta H_G(S)}\prod_x\sigma_N(dS_x) Z A = ∫ e − β H G ( S ) ∏ x σ N ( d S x ) ,
Z Γ = c N V Z A , Z M , N = e β z N V / 2 Z A . (15) Z_\Gamma=c_N^V Z_A,\qquad Z_{M,N}=e^{\beta zNV/2}Z_A. \tag*{(15)} Z Γ = c N V Z A , Z M , N = e β z N V /2 Z A . (15)
We retain both scalar prefactors. Write ξ = ρ − 1 \xi=\rho-1 ξ = ρ − 1 and c = 1 − 1 / N − β q c=1-1/N-\beta q c = 1 − 1/ N − β q . For ∣ ξ ∣ ≤ N − 1 / 2 ≤ 1 / 2 |\xi|\le N^{-1/2}\le1/2 ∣ ξ ∣ ≤ N − 1/2 ≤ 1/2 , log ( 1 + ξ ) ≥ ξ − ξ 2 \log(1+\xi)\ge\xi-\xi^2 log ( 1 + ξ ) ≥ ξ − ξ 2 , and consequently the exponent in I N ( q ) I_N(q) I N ( q ) is at least N c ξ − N ( 1 + t ) ξ 2 Nc\xi-N(1+t)\xi^2 N c ξ − N ( 1 + t ) ξ 2 . Integration over this symmetric interval and cosh ( N c ξ ) ≥ 1 \cosh(Nc\xi)\ge1 cosh ( N c ξ ) ≥ 1 give
log I N ( q ) ≥ − 1 2 log N + log 2 − 1 − t . (16) \log I_N(q)\ge-\frac{1}{2}\log N+\log2-1-t. \tag*{(16)} log I N ( q ) ≥ − 2 1 log N + log 2 − 1 − t . (16)
The inequality log ρ ≤ ρ − 1 \log\rho\le\rho-1 log ρ ≤ ρ − 1 , completion of the square, and extension of the integral to the whole real line give
log I N ( q ) ≤ N 4 t ( 1 − 1 / N − β q ) 2 + 1 2 log π t N . (17) \log I_N(q)\le\frac{N}{4t}(1-1/N-\beta q)^2+\frac{1}{2}\log\frac{\pi}{tN}. \tag*{(17)} log I N ( q ) ≤ 4 t N ( 1 − 1/ N − β q ) 2 + 2 1 log tN π . (17)
These inequalities hold for every real q q q .
We next bound the Gaussian expectation of R = Γ − H G \mathcal{R}=\Gamma-H_G R = Γ − H G using only the diagonal covariance normalization. Since N ρ x 2 N\rho_x^2 N ρ x 2 has a chi-square distribution with N N N degrees of freedom,
E G ρ x 2 = 1 , E G ( ρ x 2 − 1 ) 2 = 2 / N , E G ( ρ x − 1 ) 2 ≤ 2 / N . \mathbb{E}_G\rho_x^2=1,\qquad\mathbb{E}_G(\rho_x^2-1)^2=2/N,\qquad\mathbb{E}_G(\rho_x-1)^2\le2/N. E G ρ x 2 = 1 , E G ( ρ x 2 − 1 ) 2 = 2/ N , E G ( ρ x − 1 ) 2 ≤ 2/ N .
Wick’s formula gives E G ρ x 2 ρ y 2 = 1 + 2 ( Σ x y G ) 2 / N ≤ 1 + 2 / N \mathbb{E}_G\rho_x^2\rho_y^2=1+2(\Sigma^G_{xy})^2/N\le1+2/N E G ρ x 2 ρ y 2 = 1 + 2 ( Σ x y G ) 2 / N ≤ 1 + 2/ N . Also
E G ρ x ρ y = 1 − 1 2 E G ( ρ x − ρ y ) 2 ≥ 1 − 4 / N , E G ( ρ x ρ y − 1 ) 2 ≤ 10 / N . \mathbb{E}_G\rho_x\rho_y=1-\frac{1}{2}\mathbb{E}_G(\rho_x-\rho_y)^2\ge1-4/N,\qquad\mathbb{E}_G(\rho_x\rho_y-1)^2\le10/N. E G ρ x ρ y = 1 − 2 1 E G ( ρ x − ρ y ) 2 ≥ 1 − 4/ N , E G ( ρ x ρ y − 1 ) 2 ≤ 10/ N .
Put ζ x = N − 1 X x ⋅ ( H X ) x \zeta_x=N^{-1}X_x\cdot(\mathcal{H}X)_x ζ x = N − 1 X x ⋅ ( H X ) x . Using H Σ G = β − 1 I \mathcal{H}\Sigma^G=\beta^{-1}I H Σ G = β − 1 I , independent coordinate summation and Wick’s formula give
E G ( β ζ x ) = 1 , Var G ( β ζ x ) = β H x x + 1 N ≤ t + 1 N . \mathbb{E}_G(\beta\zeta_x)=1,\qquad\operatorname{Var}_G(\beta\zeta_x)=\frac{\beta H_{xx}+1}{N}\le\frac{t+1}{N}. E G ( β ζ x ) = 1 , Var G ( β ζ x ) = N β H xx + 1 ≤ N t + 1 .
The identity
q x − ζ x = ∑ y H x y S x ⋅ S y N ( 1 − ρ x ρ y ) q_x-\zeta_x=\sum_y H_{xy}\frac{S_x\cdot S_y}{N}(1-\rho_x\rho_y) q x − ζ x = y ∑ H x y N S x ⋅ S y ( 1 − ρ x ρ y )
and weighted Cauchy–Schwarz therefore imply
E G { β 2 ( q x − ζ x ) 2 } ≤ 10 t 2 N , E G ( 1 − 1 / N − β q x ) 2 ≤ 2 t + 4 + 20 t 2 N . \mathbb{E}_G\{\beta^2(q_x-\zeta_x)^2\}\le\frac{10t^2}{N},\qquad\mathbb{E}_G(1-1/N-\beta q_x)^2\le\frac{2t+4+20t^2}{N}. E G { β 2 ( q x − ζ x ) 2 } ≤ N 10 t 2 , E G ( 1 − 1/ N − β q x ) 2 ≤ N 2 t + 4 + 20 t 2 .
For the second bound one applies ( u + v ) 2 ≤ 2 u 2 + 2 v 2 (u+v)^2\le2u^2+2v^2 ( u + v ) 2 ≤ 2 u 2 + 2 v 2 to u = 1 − 1 / N − β ζ x u=1-1/N-\beta\zeta_x u = 1 − 1/ N − β ζ x and v = β ( ζ x − q x ) v=\beta(\zeta_x-q_x) v = β ( ζ x − q x ) . Expansion gives
R = Δ N ∑ x ( ρ x − 1 ) 2 − 1 2 ∑ x , y H x y ( ρ x − 1 ) ( ρ y − 1 ) S x ⋅ S y + ∑ x Φ N ( q x ) . \mathcal{R}=\Delta N\sum_x(\rho_x-1)^2-\frac{1}{2}\sum_{x,y}H_{xy}(\rho_x-1)(\rho_y-1)S_x\cdot S_y+\sum_x\Phi_N(q_x). R = Δ N x ∑ ( ρ x − 1 ) 2 − 2 1 x , y ∑ H x y ( ρ x − 1 ) ( ρ y − 1 ) S x ⋅ S y + x ∑ Φ N ( q x ) .
The absolute value of the cross term is at most Δ N ∑ x ( ρ x − 1 ) 2 / 2 \Delta N\sum_x(\rho_x-1)^2/2 Δ N ∑ x ( ρ x − 1 ) 2 /2 . Thus R ≥ ∑ x Φ N ( q x ) \mathcal{R}\ge\sum_x\Phi_N(q_x) R ≥ ∑ x Φ N ( q x ) pointwise and, by (17),
β V E G R ≤ 3 t + 2 t + 4 + 20 t 2 4 t + 1 2 log π t N = 8 t + 1 2 + 1 t + 1 2 log π t N . \frac{\beta}{V}\mathbb{E}_G\mathcal{R}\le3t+\frac{2t+4+20t^2}{4t}+\frac{1}{2}\log\frac{\pi}{tN}=8t+\frac{1}{2}+\frac{1}{t}+\frac{1}{2}\log\frac{\pi}{tN}. V β E G R ≤ 3 t + 4 t 2 t + 4 + 20 t 2 + 2 1 log tN π = 8 t + 2 1 + t 1 + 2 1 log tN π .
The Gibbs variational inequalities are − β E G R ≤ log ( Z Γ / Z G ) ≤ − β E Γ R -\beta E_{\mathrm{G}}\mathcal{R}\le\log(Z_{\Gamma}/Z_{\mathrm{G}})\le-\beta E_{\Gamma}\mathcal{R} − β E G R ≤ log ( Z Γ / Z G ) ≤ − β E Γ R . The lower scalar bound controls the latter expectation; hence
1 2 log N − 8 t − 1 2 − 1 t − 1 2 log ( π / t ) N ≤ 1 N V log Z Γ Z G ≤ 1 2 log N + 1 + t − log 2 N . (18) \frac{\frac{1}{2}\log N-8t-\frac{1}{2}-\frac{1}{t}-\frac{1}{2}\log(\pi/t)}{N}\le\frac{1}{NV}\log\frac{Z_{\Gamma}}{Z_{\mathrm{G}}}\le\frac{\frac{1}{2}\log N+1+t-\log2}{N}. \tag*{(18)} N 2 1 log N − 8 t − 2 1 − t 1 − 2 1 log ( π / t ) ≤ N V 1 log Z G Z Γ ≤ N 2 1 log N + 1 + t − log 2 . (18)
All the expectations exist: the angular variables range over a compact set and the auxiliary radial Hamiltonian has a positive quadratic coefficient.
Stirling’s inequalities give
log c N = N 2 log ( 2 π e ) + 1 2 log ( N / π ) − θ N , 0 < θ N < 1 6 N . \log c_N=\frac{N}{2}\log(2\pi e)+\frac{1}{2}\log(N/\pi)-\theta_N,\qquad0<\theta_N<\frac{1}{6N}. log c N = 2 N log ( 2 π e ) + 2 1 log ( N / π ) − θ N , 0 < θ N < 6 N 1 .
Combining this with (15) yields the exact normalization identity
P M , N − P M c o m p = 1 N V log Z Γ Z G − log ( N / π ) 2 N + θ N N . (19) P_{M,N}-P_M^{\mathrm{comp}}=\frac{1}{NV}\log\frac{Z_{\Gamma}}{Z_{\mathrm{G}}}-\frac{\log(N/\pi)}{2N}+\frac{\theta_N}{N}. \tag*{(19)} P M , N − P M comp = N V 1 log Z G Z Γ − 2 N log ( N / π ) + N θ N . (19)
The log N \log N log N terms cancel on both sides of (18) . The remaining lower bound is
− 8 t − 1 / 2 − 1 / t + ( log t ) / 2 + θ N N ≥ − 8 T + 3 / 2 N . \frac{-8t-1/2-1/t+(\log t)/2+\theta_N}{N}\ge-\frac{8T+3/2}{N}. N − 8 t − 1/2 − 1/ t + ( log t ) /2 + θ N ≥ − N 8 T + 3/2 .
The upper bound is at most ( T + 1 ) / N (T+1)/N ( T + 1 ) / N , because θ N < 1 / 24 < log 2 − 1 2 log π \theta_N<1/24<\log2-\frac{1}{2}\log\pi θ N < 1/24 < log 2 − 2 1 log π for N ≥ 4 N\ge4 N ≥ 4 . This proves (12) .
Pressure derivatives and the energy deficit
By Proposition 4.1, the Gibbs measure satisfies the hypothesis (20) of the next lemma.
Lemma 4.3 (Pressure derivatives and finite-volume deficit). Suppose the Gibbs measure satisfies the full-vector Gaussian domination
E exp { v ⋅ ∑ x a x S x } ≤ exp { 1 2 K M ( a ) ∣ v ∣ 2 } , K M ( a ) = β − 1 Q M 0 ( a ) , Q M 0 ( a ) = ( a , G J , M 0 a ) , (20) \mathbb{E}\exp\left\{v\cdot\sum_x a_xS_x\right\}\le\exp\left\{\frac{1}{2}K_M(a)|v|^2\right\},\qquad K_M(a)=\beta^{-1}Q_M^0(a),\quad Q_M^0(a)=(a,G_{J,M}^0a), \tag*{(20)} E exp { v ⋅ x ∑ a x S x } ≤ exp { 2 1 K M ( a ) ∣ v ∣ 2 } , K M ( a ) = β − 1 Q M 0 ( a ) , Q M 0 ( a ) = ( a , G J , M 0 a ) , (20) for every zero-sum real source a a a and every v ∈ R N v\in\mathbb{R}^N v ∈ R N . Let
D ( a , a ) = K M ( a ) − E ∣ ∑ x a x S x ∣ 2 N , b i = δ 0 − δ e i , δ M , N = ∑ i J i D ( b i , b i ) . D(a,a)=K_M(a)-\mathbb{E}\frac{\left|\sum_xa_xS_x\right|^2}{N},\qquad b_i=\delta_0-\delta_{e_i},\qquad\delta_{M,N}=\sum_iJ_iD(b_i,b_i). D ( a , a ) = K M ( a ) − E N ∣ ∑ x a x S x ∣ 2 , b i = δ 0 − δ e i , δ M , N = i ∑ J i D ( b i , b i ) . Then
( P M c o m p ) ′ = j + s M / 2 − 1 2 β , 0 ≤ ( P M c o m p ) ′ ′ ≤ 1 2 β 2 , (21) (P_M^{\mathrm{comp}})'=j+s_M/2-\frac{1}{2\beta},\qquad0\le(P_M^{\mathrm{comp}})''\le\frac{1}{2\beta^2}, \tag*{(21)} ( P M comp ) ′ = j + s M /2 − 2 β 1 , 0 ≤ ( P M comp ) ′′ ≤ 2 β 2 1 , (21) δ M , N = 2 [ P M , N ′ − ( P M c o m p ) ′ ] + s M − 1 β V . (22) \delta_{M,N}=2\left[P'_{M,N}-(P_M^{\mathrm{comp}})'\right]+s_M-\frac{1}{\beta V}. \tag*{(22)} δ M , N = 2 [ P M , N ′ − ( P M comp ) ′ ] + s M − β V 1 . (22) For 0 < h ≤ B ′ − β 0<h\le B'-\beta 0 < h ≤ B ′ − β and N ≥ 4 N\ge4 N ≥ 4 ,
0 ≤ δ M , N ≤ 4 ε N ( B ′ , J ) h + h 2 β 2 + s M − 1 β V . (23) 0\le\delta_{M,N}\le\frac{4\varepsilon_N(B',J)}{h}+\frac{h}{2\beta^2}+s_M-\frac{1}{\beta V}. \tag*{(23)} 0 ≤ δ M , N ≤ h 4 ε N ( B ′ , J ) + 2 β 2 h + s M − β V 1 . (23) In particular, on J ∈ [ j 0 , j 1 ] d J\in[j_0,j_1]^d J ∈ [ j 0 , j 1 ] d and β ∈ [ b , B ] \beta\in[b,B] β ∈ [ b , B ] , with 0 < j 0 ≤ j 1 0<j_0\le j_1 0 < j 0 ≤ j 1 and 0 < b ≤ B < ∞ 0<b\le B<\infty 0 < b ≤ B < ∞ , one may take
δ M , N ≤ c ∗ N − 1 / 2 + s M , c ∗ = 2 2 max ( 1 , b − 1 ) K P + , K P + = 8 ( 1 + 4 ( B + 1 ) d j 1 ) + 3 / 2. (24) \delta_{M,N}\le c_*N^{-1/2}+s_M,\qquad c_*=2\sqrt{2}\max(1,b^{-1})\sqrt{K_P^+},\qquad K_P^+=8(1+4(B+1)dj_1)+3/2. \tag*{(24)} δ M , N ≤ c ∗ N − 1/2 + s M , c ∗ = 2 2 max ( 1 , b − 1 ) K P + , K P + = 8 ( 1 + 4 ( B + 1 ) d j 1 ) + 3/2. (24) These assertions hold also for β < β c , J \beta<\beta_{c,J} β < β c , J .
Proof. The partial derivative of (11) with respect to its mass is zero by (4). Moreover
s M ′ ( β ) = − [ 1 V ∑ k ( λ J ( k ) + s M ) − 2 ] − 1 . s'_M(\beta)=-\left[\frac{1}{V}\sum_k\left(\lambda_J(k)+s_M\right)^{-2}\right]^{-1}. s M ′ ( β ) = − [ V 1 k ∑ ( λ J ( k ) + s M ) − 2 ] − 1 .
The bracket is at least β 2 \beta^2 β 2 by Jensen, proving (21) . Global O ( N ) O(N) O ( N ) symmetry and the second derivative at v = 0 v=0 v = 0 in (20) give D ( a , a ) ≥ 0 D(a,a)\ge0 D ( a , a ) ≥ 0 . The Fourier identity ∑ i J i ∣ b ^ i ( k ) ∣ 2 = λ J ( k ) \sum_i J_i\lvert\hat b_i(k)\rvert^2=\lambda_J(k) ∑ i J i ∣ b ^ i ( k ) ∣ 2 = λ J ( k ) gives
∑ i J i Q M 0 ( b i ) = 1 − V − 1 . \sum_i J_iQ_M^0(b_i)=1-V^{-1}. i ∑ J i Q M 0 ( b i ) = 1 − V − 1 .
Writing u M , N = P M , N ′ = ∑ i J i E ( S 0 ⋅ S e i ) / N u_{M,N}=P'_{M,N}=\sum_i J_i\mathbb{E}(S_0\cdot S_{e_i})/N u M , N = P M , N ′ = ∑ i J i E ( S 0 ⋅ S e i ) / N and using ∣ S x ∣ 2 = N \lvert S_x\rvert^2=N ∣ S x ∣ 2 = N , we obtain δ M , N = ( 1 − V − 1 ) / β − 2 ( j − u M , N ) \delta_{M,N}=(1-V^{-1})/\beta-2(j-u_{M,N}) δ M , N = ( 1 − V − 1 ) / β − 2 ( j − u M , N ) . This is (22) . Convexity of P M , N P_{M,N} P M , N , the pressure error, and the curvature bound give
2 [ P M , N ′ − ( P M c o m p ) ′ ] ≤ 4 ε N h + h 2 β 2 , 2\left[P'_{M,N}-\left(P_M^{\mathrm{comp}}\right)'\right]\le\frac{4\varepsilon_N}{h}+\frac{h}{2\beta^2}, 2 [ P M , N ′ − ( P M comp ) ′ ] ≤ h 4 ε N + 2 β 2 h ,
which proves (23) . For the uniform estimate use B ′ = B + 1 B'=B+1 B ′ = B + 1 and ε N ≤ K P + / N \varepsilon_N\le K_P^+/N ε N ≤ K P + / N . If K P + / N ≤ 1 / ( 8 b 2 ) K_P^+/N\le1/(8b^2) K P + / N ≤ 1/ ( 8 b 2 ) , take h = 2 2 b K P + / N ≤ 1 h=2\sqrt{2}b\sqrt{K_P^+/N}\le1 h = 2 2 b K P + / N ≤ 1 . The two secant terms then sum to at most ( 2 2 / b ) K P + / N (2\sqrt{2}/b)\sqrt{K_P^+/N} ( 2 2 / b ) K P + / N . Otherwise δ M , N ≤ ∑ i J i K M ( b i ) ≤ 1 / b \delta_{M,N}\le\sum_i J_iK_M(b_i)\le1/b δ M , N ≤ ∑ i J i K M ( b i ) ≤ 1/ b and 1 / b < 2 2 K P + / N 1/b<2\sqrt{2}\sqrt{K_P^+/N} 1/ b < 2 2 K P + / N . Both cases give (24) .
The upper bound
Full-vector domination, paths, and the Gram matrix
In this subsection the field may be the above Gibbs field, or any field on Λ M \Lambda_M Λ M having translation invariance, simultaneous O ( N ) O(N) O ( N ) invariance, the hard constraint ∣ S x ∣ 2 = N \lvert S_x\rvert^2=N ∣ S x ∣ 2 = N , and (20) . The same statements hold on Z d \mathbb{Z}^d Z d for finitely supported zero-sum sources, with K ∞ ( a ) = β − 1 Q ∞ ( a ) K_\infty(a)=\beta^{-1}Q_\infty(a) K ∞ ( a ) = β − 1 Q ∞ ( a ) in place of K M ( a ) K_M(a) K M ( a ) , D ∞ ( a , a ) = K ∞ ( a ) − E ∣ ∑ x a x S x ∣ 2 / N D_\infty(a,a)=K_\infty(a)-\mathbb{E}\lvert\sum_xa_xS_x\rvert^2/N D ∞ ( a , a ) = K ∞ ( a ) − E ∣ ∑ x a x S x ∣ 2 / N in place of D D D , and δ ∞ , N = ∑ i J i D ∞ ( b i , b i ) \delta_{\infty,N}=\sum_iJ_iD_\infty(b_i,b_i) δ ∞ , N = ∑ i J i D ∞ ( b i , b i ) in place of δ M , N \delta_{M,N} δ M , N . The path argument uses this lattice geometry and translation invariance.
Lemma 5.1 (Quantitative saturation of the full norm). Under these assumptions D D D is positive semidefinite on zero-sum sources. Let ( T z a ) x = a x − z (T_za)_x=a_{x-z} ( T z a ) x = a x − z denote translation by z ∈ Z d z\in\mathbb{Z}^d z ∈ Z d . For any path decomposition a = ∑ q α q T z q b i q a=\sum_q\alpha_qT_{z_q}b_{i_q} a = ∑ q α q T z q b i q ,
D ( a , a ) ≤ δ M , N ( ∑ q ∣ α q ∣ J i q ) 2 . (25) D(a,a)\le\delta_{M,N}\left(\sum_q\frac{\lvert\alpha_q\rvert}{\sqrt{J_{i_q}}}\right)^2. \tag*{(25)} D ( a , a ) ≤ δ M , N ( q ∑ J i q ∣ α q ∣ ) 2 . (25) For Y a = ∑ x a x S x Y_a=\sum_xa_xS_x Y a = ∑ x a x S x , U a = ∣ Y a ∣ 2 / N U_a=\lvert Y_a\rvert^2/N U a = ∣ Y a ∣ 2 / N and every N ≥ 1 N\ge1 N ≥ 1 ,
E ∣ U a − K M ( a ) ∣ ≤ D ( a , a ) + c χ K M ( a ) / N , c χ = 2 π + 4. (26) \mathbb{E}\lvert U_a-K_M(a)\rvert\le D(a,a)+c_\chi K_M(a)/\sqrt{N},\qquad c_\chi=2\sqrt{\pi}+4. \tag*{(26)} E ∣ U a − K M ( a )∣ ≤ D ( a , a ) + c χ K M ( a ) / N , c χ = 2 π + 4. (26) When K M ( a ) > 0 K_M(a)>0 K M ( a ) > 0 , the sharper upper-tail bound is
P { U a ≥ K M ( a ) ( 1 + s ) } ≤ exp { − N [ s − log ( 1 + s ) ] / 2 } , s > 0. (27) \mathbb{P}\{U_a\ge K_M(a)(1+s)\}\le\exp\{-N[s-\log(1+s)]/2\},\qquad s>0. \tag*{(27)} P { U a ≥ K M ( a ) ( 1 + s )} ≤ exp { − N [ s − log ( 1 + s )] /2 } , s > 0. (27)
Proof. Rotation invariance gives E Y a = 0 \mathbb{E}Y_a=0 E Y a = 0 and E ( Y a 1 ) 2 = E ∣ Y a ∣ 2 / N \mathbb{E}(Y_a^1)^2=\mathbb{E}\lvert Y_a\rvert^2/N E ( Y a 1 ) 2 = E ∣ Y a ∣ 2 / N . Taking the second derivative at zero in (20) proves D ( a , a ) ≥ 0 D(a,a)\ge0 D ( a , a ) ≥ 0 . Since D D D is a difference of quadratic forms, its polarization satisfies ∣ D ( a , b ) ∣ ≤ D ( a , a ) D ( b , b ) \lvert D(a,b)\rvert\le\sqrt{D(a,a)D(b,b)} ∣ D ( a , b )∣ ≤ D ( a , a ) D ( b , b ) . Translation invariance and positivity of the J i J_i J i give D ( T z b i , T z b i ) = D ( b i , b i ) ≤ δ M , N / J i D(T_zb_i,T_zb_i)=D(b_i,b_i)\le\delta_{M,N}/J_i D ( T z b i , T z b i ) = D ( b i , b i ) ≤ δ M , N / J i . The triangle inequality for this positive semidefinite seminorm gives (25) . Let g ∼ N ( 0 , I N ) g \sim\mathcal{N}(0,I_N) g ∼ N ( 0 , I N ) be independent. Tonelli’s theorem and (20) give
E e t U a / 2 ≤ E g E e t / N g ⋅ Y a ≤ ( 1 − t K M ( a ) N ) − N / 2 , 0 ≤ t < N K M ( a ) . (28) \mathbb{E}e^{tU_a/2}\leq\mathbb{E}_g\mathbb{E}e^{\sqrt{t/N}\,g\cdot Y_a}\leq\left(1-\frac{tK_M(a)}{N}\right)^{-N/2},\qquad0\leq t<\frac{N}{K_M(a)}. \tag*{(28)} E e t U a /2 ≤ E g E e t / N g ⋅ Y a ≤ ( 1 − N t K M ( a ) ) − N /2 , 0 ≤ t < K M ( a ) N . (28)
If K M ( a ) = 0 K_M(a)=0 K M ( a ) = 0 , the second-moment bound makes Y a = 0 Y_a=0 Y a = 0 almost surely, and (26) is immediate. Otherwise put K = K M ( a ) K=K_M(a) K = K M ( a ) . Optimizing the Chernoff bound in (28) at t = ( N / K ) s / ( 1 + s ) t=(N/K)s/(1+s) t = ( N / K ) s / ( 1 + s ) proves (27) . For an explicit integrated bound, the inequality − log ( 1 − z ) − z ≤ z 2 / [ 2 ( 1 − z ) ] -\log(1-z)-z\leq z^2/[2(1-z)] − log ( 1 − z ) − z ≤ z 2 / [ 2 ( 1 − z )] , 0 ≤ z < 1 0\leq z<1 0 ≤ z < 1 , gives
log E e θ ( U a − K ) ≤ v θ 2 2 ( 1 − c θ ) , v = 2 K 2 / N , c = 2 K / N , 0 ≤ θ < 1 / c . \log\mathbb{E}e^{\theta(U_a-K)}\leq\frac{v\theta^2}{2(1-c\theta)},\qquad v=2K^2/N,\qquad c=2K/N,\qquad0\leq\theta<1/c. log E e θ ( U a − K ) ≤ 2 ( 1 − c θ ) v θ 2 , v = 2 K 2 / N , c = 2 K / N , 0 ≤ θ < 1/ c .
For x > 0 x>0 x > 0 use θ = 2 x / v / ( 1 + c 2 x / v ) \theta=\sqrt{2x/v}/(1+c\sqrt{2x/v}) θ = 2 x / v / ( 1 + c 2 x / v ) . The resulting Chernoff bound is
P { U a − K ≥ 2 K x / N + 2 K x / N } ≤ e − x . \mathbb{P}\{U_a-K\geq2K\sqrt{x/N}+2Kx/N\}\leq e^{-x}. P { U a − K ≥ 2 K x / N + 2 K x / N } ≤ e − x .
Integration after this increasing change of threshold yields
E ( U a − K ) + ≤ K ∫ 0 ∞ e − x ( 1 N x + 2 N ) d x = K ( π N + 2 N ) . \mathbb{E}(U_a-K)_+\leq K\int_0^\infty e^{-x}\left(\frac{1}{\sqrt{Nx}}+\frac{2}{N}\right)\,dx=K\left(\frac{\sqrt{\pi}}{\sqrt{N}}+\frac{2}{N}\right). E ( U a − K ) + ≤ K ∫ 0 ∞ e − x ( N x 1 + N 2 ) d x = K ( N π + N 2 ) .
Finally E ( K − U a ) + − E ( U a − K ) + = D ( a , a ) \mathbb{E}(K-U_a)_+-\mathbb{E}(U_a-K)_+=D(a,a) E ( K − U a ) + − E ( U a − K ) + = D ( a , a ) . Adding the two positive parts and using N − 1 ≤ N − 1 / 2 N^{-1}\leq N^{-1/2} N − 1 ≤ N − 1/2 proves (26) .
Lemma 5.2 (Gram error and a positive target). Fix distinct sites U = { x 1 , … , x ℓ } \mathcal{U}=\{x_1,\ldots,x_\ell\} U = { x 1 , … , x ℓ } , injectively embedded in the torus, and put A i j = S x i ⋅ S x j / N A_{ij}=S_{x_i}\cdot S_{x_j}/N A ij = S x i ⋅ S x j / N and a i j = δ x i − δ x j a_{ij}=\delta_{x_i}-\delta_{x_j} a ij = δ x i − δ x j . For each pair choose a fixed nearest-neighbour path and let h i j ( J ) h_{ij}(J) h ij ( J ) be the sum of J q − 1 / 2 J_q^{-1/2} J q − 1/2 over its steps. Suppose a deterministic matrix Σ \Sigma Σ satisfies
Σ i i = 1 , Σ ⪰ m I ℓ ( m > 0 ) , Σ i j = 1 − 1 2 K ∞ ( a i j ) . \Sigma_{ii}=1,\qquad\Sigma\succeq mI_\ell\quad(m>0),\qquad\Sigma_{ij}=1-\frac{1}{2}K_\infty(a_{ij}). Σ ii = 1 , Σ ⪰ m I ℓ ( m > 0 ) , Σ ij = 1 − 2 1 K ∞ ( a ij ) . Define
H U , J = 1 2 ∑ i , j h i j ( J ) 2 , R U , M = c λ 2 ∑ i , j K M ( a i j ) , ϵ U ( M ) = 1 2 ∑ i , j ∣ K M ( a i j ) − K ∞ ( a i j ) ∣ . H_{\mathcal{U},J}=\frac{1}{2}\sum_{i,j}h_{ij}(J)^2,\qquad R_{\mathcal{U},M}=\frac{c_\lambda}{2}\sum_{i,j}K_M(a_{ij}),\qquad\epsilon_{\mathcal{U}}(M)=\frac{1}{2}\sum_{i,j}|K_M(a_{ij})-K_\infty(a_{ij})|. H U , J = 2 1 i , j ∑ h ij ( J ) 2 , R U , M = 2 c λ i , j ∑ K M ( a ij ) , ϵ U ( M ) = 2 1 i , j ∑ ∣ K M ( a ij ) − K ∞ ( a ij ) ∣. Then
E ∥ A − Σ ∥ F ≤ H U , J δ M , N + R U , M N − 1 / 2 + ϵ U ( M ) . (29) \mathbb{E}\|A-\Sigma\|_F\leq H_{\mathcal{U},J}\delta_{M,N}+R_{\mathcal{U},M}N^{-1/2}+\epsilon_{\mathcal{U}}(M). \tag*{(29)} E ∥ A − Σ ∥ F ≤ H U , J δ M , N + R U , M N − 1/2 + ϵ U ( M ) . (29) For the critical or ordered lattice target
Σ i j = G J ( x i , x j ) β + 1 − β c , J β , β ≥ β c , J , β c , J = G J ( 0 , 0 ) , \Sigma_{ij}=\frac{G_J(x_i,x_j)}{\beta}+1-\frac{\beta_{c,J}}{\beta},\qquad\beta\geq\beta_{c,J},\qquad\beta_{c,J}=G_J(0,0), Σ ij = β G J ( x i , x j ) + 1 − β β c , J , β ≥ β c , J , β c , J = G J ( 0 , 0 ) , one has K ∞ ( a ) = β − 1 ( a , G J a ) K_\infty(a)=\beta^{-1}(a,G_Ja) K ∞ ( a ) = β − 1 ( a , G J a ) and
Σ ⪰ ( 4 β j ) − 1 I ℓ . (30) \Sigma\succeq(4\beta j)^{-1}I_\ell. \tag*{(30)} Σ ⪰ ( 4 β j ) − 1 I ℓ . (30)
Proof. The full Gram matrix has A ⪰ 0 A\succeq0 A ⪰ 0 , A i i = 1 A_{ii}=1 A ii = 1 , and tr A = ℓ \operatorname{tr}A=\ell tr A = ℓ . Hard length gives A i j = 1 − U a i j / 2 A_{ij}=1-U_{a_{ij}}/2 A ij = 1 − U a ij /2 . Apply (25) and (26) entrywise, then use ∥ B ∥ F ≤ ∑ i , j ∣ B i j ∣ \|B\|_F\leq\sum_{i,j}|B_{ij}| ∥ B ∥ F ≤ ∑ i , j ∣ B ij ∣ to prove (29) . For the specified lattice target the dipole relation follows from translation invariance of G J G_J G J . If z ∈ R ℓ z\in\mathbb{R}^{\ell} z ∈ R ℓ and a = ∑ i z i δ x i a=\sum_i z_i\delta_{x_i} a = ∑ i z i δ x i , Parseval and λ J ≤ 4 j \lambda_J\leq4j λ J ≤ 4 j almost everywhere give
z T Σ z = 1 β ∫ ∣ a ^ ( k ) ∣ 2 λ J ( k ) d k ( 2 π ) d + ( 1 − β c , J β ) ( ∑ i z i ) 2 ≥ ∣ z ∣ 2 4 β j . z^{\mathsf{T}}\Sigma z=\frac{1}{\beta}\int\frac{|\widehat{a}(k)|^2}{\lambda_J(k)}\frac{dk}{(2\pi)^d}+\left(1-\frac{\beta_{c,J}}{\beta}\right)\left(\sum_i z_i\right)^2\geq\frac{|z|^2}{4\beta j}. z T Σ z = β 1 ∫ λ J ( k ) ∣ a ( k ) ∣ 2 ( 2 π ) d d k + ( 1 − β β c , J ) ( i ∑ z i ) 2 ≥ 4 β j ∣ z ∣ 2 .
The integrals exist for d ≥ 3 d \ge3 d ≥ 3 . For uniform constants on J ∈ [ j 0 , j 1 ] d J \in[j_0,j_1]^d J ∈ [ j 0 , j 1 ] d , β ∈ [ b , B ] \beta\in[b,B] β ∈ [ b , B ] , let n i j n_{ij} n ij be the chosen path’s step count. Then h i j ≤ n i j / j 0 h_{ij} \le n_{ij}/\sqrt{j_0} h ij ≤ n ij / j 0 , and Fourier path decomposition gives ∣ a ^ i j ( k ) ∣ ≤ h i j λ J ( k ) |\widehat{a}_{ij}(k)| \le h_{ij}\sqrt{\lambda_J(k)} ∣ a ij ( k ) ∣ ≤ h ij λ J ( k ) . Consequently
H U , J ≤ H ∗ : = 1 2 j 0 ∑ i , j n i j 2 , R U , M ≤ R ∗ : = c χ H ∗ / b . H_{U,J} \le H_* := \frac{1}{2j_0}\sum_{i,j} n_{ij}^{2}, \qquad R_{U,M} \le R_* := c_\chi H_*/b. H U , J ≤ H ∗ := 2 j 0 1 i , j ∑ n ij 2 , R U , M ≤ R ∗ := c χ H ∗ / b .
For the critical or ordered lattice target one may take m = m ∗ : = ( 4 B d j 1 ) − 1 m=m_* := (4Bd j_1)^{-1} m = m ∗ := ( 4 B d j 1 ) − 1 by (30). The remaining deterministic term is a fixed finite collection of zero-sum Green errors, estimated in Lemma A.2.
Haar projection with all admissible coordinate counts
Proposition 5.3 (Quantitative rotational projection). Let X ∈ R ℓ × N X \in\mathbb{R}^{\ell\times N} X ∈ R ℓ × N be invariant in law under right multiplication by every O ∈ O ( N ) O \in O(N) O ∈ O ( N ) , and let A = X X T / N A=XX^{\mathsf{T}}/N A = X X T / N satisfy A i i = 1 A_{ii}=1 A ii = 1 almost surely. If N ≥ ℓ N \ge\ell N ≥ ℓ , 1 ≤ p ≤ N 1 \le p \le N 1 ≤ p ≤ N , and Σ ⪰ m I ℓ \Sigma\succeq mI_\ell Σ ⪰ m I ℓ is deterministic for some m > 0 m>0 m > 0 , then the law ν U , p \nu_{U,p} ν U , p of the first p p p columns satisfies
W 1 ( ν U , p , γ Σ ⊗ p ) ≤ p [ κ ℓ N + m − 1 / 2 E ∥ A − Σ ∥ F ] , κ ℓ = ( 2 + 4 ) ℓ 3 / 2 ℓ + 1 . (31) W_1\left(\nu_{U,p},\gamma_{\Sigma}^{\otimes p}\right) \le\sqrt{p}\left[\frac{\kappa_\ell}{\sqrt{N}}+m^{-1/2}\mathbb{E}\lVert A-\Sigma\rVert_F\right], \qquad\kappa_\ell=(\sqrt{2}+4)\ell^{3/2}\sqrt{\ell+1}. \tag*{(31)} W 1 ( ν U , p , γ Σ ⊗ p ) ≤ p [ N κ ℓ + m − 1/2 E ∥ A − Σ ∥ F ] , κ ℓ = ( 2 + 4 ) ℓ 3/2 ℓ + 1 . (31) Here W 1 W_1 W 1 uses the unnormalized Frobenius norm and γ Σ ⊗ p \gamma_{\Sigma}^{\otimes p} γ Σ ⊗ p has p p p independent N ( 0 , Σ ) \mathcal{N}(0,\Sigma) N ( 0 , Σ ) columns. The random Gram matrix A A A may be singular.
Proof. Let P = [ I ℓ 0 ] P=[I_\ell\ 0] P = [ I ℓ 0 ] and let E p : R p → R N E_p:\mathbb{R}^p\to\mathbb{R}^N E p : R p → R N embed the first p p p coordinates. Every deterministic X X X has the same row Gram matrix as N A 1 / 2 P \sqrt{N}A^{1/2}P N A 1/2 P , hence belongs to its right orthogonal orbit. For a rank-deficient X X X , the isometry between the two row spaces extends to an orthogonal map on R N \mathbb{R}^N R N , so this assertion still holds. If H H H is an independent Haar-distributed random matrix in O ( N ) O(N) O ( N ) , rotational invariance and Haar averaging therefore give the exact kernel representation
X E p = l a w A 1 / 2 Q N , Q N = N P H E p , Q N independent of A . XE_p \mathrel{\stackrel{\mathrm{law}}{=}} A^{1/2}Q_N, \qquad Q_N=\sqrt{N}PHE_p,\qquad Q_N\text{ independent of }A. X E p = law A 1/2 Q N , Q N = N P H E p , Q N independent of A .
This follows by averaging each deterministic orbit.
Independently of A A A , take G ∈ R N × ℓ G\in\mathbb{R}^{N\times\ell} G ∈ R N × ℓ with independent standard normal entries, and set
B = G T G / N , Z = G T E p , T = G ( G T G ) − 1 / 2 . B=G^{\mathsf{T}}G/N,\qquad Z=G^{\mathsf{T}}E_p,\qquad T=G(G^{\mathsf{T}}G)^{-1/2}. B = G T G / N , Z = G T E p , T = G ( G T G ) − 1/2 .
Since N ≥ ℓ N\ge\ell N ≥ ℓ , G G G has full column rank almost surely. The law of T T T is left O ( N ) O(N) O ( N ) invariant on { T : T T T = I ℓ } \{T:T^{\mathsf{T}}T=I_\ell\} { T : T T T = I ℓ } , so it is the Haar Stiefel law. Thus we may realize the preceding kernel as
Q N = N T T E p = B − 1 / 2 Z . Q_N=\sqrt{N}T^{\mathsf{T}}E_p=B^{-1/2}Z. Q N = N T T E p = B − 1/2 Z .
The entries of Z Z Z remain independent standard normals for every 1 ≤ p ≤ N 1\le p\le N 1 ≤ p ≤ N ; B B B and Z Z Z are in general dependent. The entire coupling ( B , Z , Q N ) (B,Z,Q_N) ( B , Z , Q N ) is independent of A A A . Direct second-moment computations yield
E ∥ B − I ℓ ∥ F 2 = ℓ ( ℓ + 1 ) N , E ∥ Z ∥ F 2 = E ∥ Q N ∥ F 2 = ℓ p . \mathbb{E}\lVert B-I_\ell\rVert_F^2=\frac{\ell(\ell+1)}{N},\qquad\mathbb{E}\lVert Z\rVert_F^2=\mathbb{E}\lVert Q_N\rVert_F^2=\ell p. E ∥ B − I ℓ ∥ F 2 = N ℓ ( ℓ + 1 ) , E ∥ Z ∥ F 2 = E ∥ Q N ∥ F 2 = ℓ p .
On E = { ∥ B − I ℓ ∥ o p ≤ 1 / 2 } \mathcal{E}=\{\lVert B-I_\ell\rVert_{\mathrm{op}}\le1/2\} E = {∥ B − I ℓ ∥ op ≤ 1/2 } , spectral calculus gives ∥ B − 1 / 2 − I ∥ o p ≤ 2 ∥ B − I ∥ o p \lVert B^{-1/2}-I\rVert_{\mathrm{op}}\le\sqrt{2}\lVert B-I\rVert_{\mathrm{op}} ∥ B − 1/2 − I ∥ op ≤ 2 ∥ B − I ∥ op . Cauchy–Schwarz then bounds the good-event expectation by
E [ ∥ Q N − Z ∥ F ; E ] ≤ 2 ℓ ( ℓ + 1 ) / N ℓ p . \mathbb{E}[\lVert Q_N-Z\rVert_F;\mathcal{E}]\le\sqrt{2}\sqrt{\ell(\ell+1)/N}\sqrt{\ell p}. E [∥ Q N − Z ∥ F ; E ] ≤ 2 ℓ ( ℓ + 1 ) / N ℓ p .
Markov gives P ( E c ) ≤ 4 ℓ ( ℓ + 1 ) / N \mathbb{P}(E^c) \le4\ell(\ell+ 1)/N P ( E c ) ≤ 4 ℓ ( ℓ + 1 ) / N . On its complement we use only the second moments:
E [ ∥ Q N − Z ∥ F ; E c ] ≤ ( E ∥ Q N ∥ F 2 + E ∥ Z ∥ F 2 ) P ( E c ) ≤ 4 ℓ p ℓ ( ℓ + 1 ) / N . \mathbb{E}[\lVert Q_N-Z\rVert_F;E^c] \le\left(\sqrt{\mathbb{E}\lVert Q_N\rVert_F^2}+\sqrt{\mathbb{E}\lVert Z\rVert_F^2}\right)\sqrt{\mathbb{P}(E^c)} \le4\sqrt{\ell p\ell(\ell+1)/N}. E [∥ Q N − Z ∥ F ; E c ] ≤ ( E ∥ Q N ∥ F 2 + E ∥ Z ∥ F 2 ) P ( E c ) ≤ 4 ℓ p ℓ ( ℓ + 1 ) / N .
Adding the bounds gives
E ∥ Q N − Z ∥ F ≤ ( 2 + 4 ) ℓ p ℓ ( ℓ + 1 ) / N . (32) \mathbb{E}\lVert Q_N-Z\rVert_F \le(\sqrt{2}+4)\sqrt{\ell p\ell(\ell+1)/N}. \tag*{(32)} E ∥ Q N − Z ∥ F ≤ ( 2 + 4 ) ℓ p ℓ ( ℓ + 1 ) / N . (32)
The square-root comparison uses only the lower bound on Σ \Sigma Σ :
∥ A 1 / 2 − Σ 1 / 2 ∥ F ≤ m − 1 / 2 ∥ A − Σ ∥ F . (33) \lVert A^{1/2}-\Sigma^{1/2}\rVert_F \le m^{-1/2}\lVert A-\Sigma\rVert_F. \tag*{(33)} ∥ A 1/2 − Σ 1/2 ∥ F ≤ m − 1/2 ∥ A − Σ ∥ F . (33)
Indeed, for S = A 1 / 2 − Σ 1 / 2 S=A^{1/2}-\Sigma^{1/2} S = A 1/2 − Σ 1/2 , A − Σ = A 1 / 2 S + S Σ 1 / 2 A-\Sigma=A^{1/2}S+S\Sigma^{1/2} A − Σ = A 1/2 S + S Σ 1/2 . In the left eigenbasis of A 1 / 2 A^{1/2} A 1/2 and the right eigenbasis of Σ 1 / 2 \Sigma^{1/2} Σ 1/2 the entries are related by the factors a i + c j ≥ m \sqrt{a_i}+\sqrt{c_j}\ge\sqrt{m} a i + c j ≥ m . The Frobenius norm is invariant under these two changes of basis, proving (33) even for singular A A A . Finally ∥ A 1 / 2 ∥ o p ≤ ℓ \lVert A^{1/2}\rVert_{\mathrm{op}}\le\sqrt{\ell} ∥ A 1/2 ∥ op ≤ ℓ , and independence of A A A and Z Z Z implies
E Z ∥ ( A 1 / 2 − Σ 1 / 2 ) Z ∥ F ≤ p ∥ A 1 / 2 − Σ 1 / 2 ∥ F . \mathbb{E}Z\lVert(A^{1/2}-\Sigma^{1/2})Z\rVert_F\le\sqrt{p}\lVert A^{1/2}-\Sigma^{1/2}\rVert_F. E Z ∥( A 1/2 − Σ 1/2 ) Z ∥ F ≤ p ∥ A 1/2 − Σ 1/2 ∥ F .
Couple A 1 / 2 Q N A^{1/2}Q_N A 1/2 Q N to Σ 1 / 2 Z \Sigma^{1/2}Z Σ 1/2 Z and combine the last inequalities with (32) to obtain (31).
Corollary 5.4 (Quantitative saturation criterion). Under the assumptions of Lemma 5.2, for N ≥ ℓ N\ge\ell N ≥ ℓ and 1 ≤ p ≤ N 1\le p\le N 1 ≤ p ≤ N ,
W 1 ( ν U , p , γ ⊗ p ) ≤ p [ κ ℓ + m − 1 / 2 R U , M N + m − 1 / 2 H U , J δ M , N + m − 1 / 2 ϵ U ( M ) ] . (34) W_1(\nu^{U,p},\gamma^{\otimes p})\le\sqrt{p}\left[\frac{\kappa_\ell+m^{-1/2}R_{U,M}}{\sqrt{N}}+m^{-1/2}H_{U,J}\delta_{M,N}+m^{-1/2}\epsilon_U(M)\right]. \tag*{(34)} W 1 ( ν U , p , γ ⊗ p ) ≤ p [ N κ ℓ + m − 1/2 R U , M + m − 1/2 H U , J δ M , N + m − 1/2 ϵ U ( M ) ] . (34) For an infinite-lattice field satisfying the same assumptions with K ∞ K_\infty K ∞ , the last term is absent, δ M , N \delta_{M,N} δ M , N is replaced by δ ∞ , N \delta_{\infty,N} δ ∞ , N , and R U , M R_{U,M} R U , M is replaced by R U , ∞ = ( c χ / 2 ) ∑ i , j K ∞ ( a i j ) R_{U,\infty}=(c_\chi/2)\sum_{i,j}K_\infty(a_{ij}) R U , ∞ = ( c χ /2 ) ∑ i , j K ∞ ( a ij ) . All displayed dependence on p p p is the factor p \sqrt{p} p .
Proof. Insert (29) into (31). For the infinite field the proof of the Gram bound uses the reference form directly, so the deterministic reference error is zero. For the critical or ordered lattice target, the bounds in Lemma 5.2 make these constants uniform over (3).
Proof of the uniform approximation
Proof of Theorem 2.1. For the full-spin Gram matrix A u v = S x u ⋅ S x v / N A_{uv}=S_{x_u}\cdot S_{x_v}/N A uv = S x u ⋅ S x v / N , hard length gives A u u = 1 A_{uu}=1 A uu = 1 and
A u v = 1 − 1 2 N ∣ S x u − S x v ∣ 2 . A_{uv}=1-\frac{1}{2N}\lvert S_{x_u}-S_{x_v}\rvert^2. A uv = 1 − 2 N 1 ∣ S x u − S x v ∣ 2 .
In the torus Gibbs model, combining (24) with the sharp zero-sum quadrature estimate Q M 0 ( a ) − Q ∞ ( a ) = O ( M − d ) Q_M^0(a)-Q_\infty(a)=O(M^{-d}) Q M 0 ( a ) − Q ∞ ( a ) = O ( M − d ) in Lemma A.2 gives
E ∥ A − Σ β , J ∥ F ≤ C U ( N − 1 / 2 + s M + M − d ) \mathbb{E}\lVert A-\Sigma_{\beta,J}\rVert_F\le C_U(N^{-1/2}+s_M+M^{-d}) E ∥ A − Σ β , J ∥ F ≤ C U ( N − 1/2 + s M + M − d )
by Lemma 5.2. By (30), Σ β , J ⪰ ( 4 β j ) − 1 I ℓ \Sigma_{\beta,J}\succeq(4\beta j)^{-1}I_\ell Σ β , J ⪰ ( 4 β j ) − 1 I ℓ . Proposition 5.3, valid even when A A A is singular, therefore gives (this is Corollary 5.4)
W 1 ≤ C U p ( N − 1 / 2 + s M + M − d ) . W_1\le C_U\sqrt{p}(N^{-1/2}+s_M+M^{-d}). W 1 ≤ C U p ( N − 1/2 + s M + M − d ) .
The mass equation gives s M ≥ 1 / ( β V ) s_M\ge1/(\beta V) s M ≥ 1/ ( β V ) , so M − d ≤ B s M M^{-d}\le Bs_M M − d ≤ B s M on β ≤ B \beta\le B β ≤ B ; this proves (5).
Lower bounds and the coordinate criterion
The lower bounds below concern the actual finite-spin laws. The argument uses radial symmetry and both the infrared and pressure comparisons.
The coordinate-growth obstruction
The coordinate-growth obstruction below comes from the one-site sphere marginal. Related sharp one-dimensional Gaussian asymptotics for Euclidean-ball marginals are studied in [9 ] .
Proposition 6.1 (Necessary coordinate-growth condition). Let N → ∞ N \to\infty N → ∞ , 1 ≤ p N ≤ N 1 \le p_N \le N 1 ≤ p N ≤ N , and let the site set be fixed and nonempty. For any globally O ( N ) O(N) O ( N ) invariant spin laws with ∣ S x ∣ 2 = N \lvert S_x\rvert^2 = N ∣ S x ∣ 2 = N , if p N / N → c ∈ ( 0 , 1 ] p_N/N \to c \in(0,1] p N / N → c ∈ ( 0 , 1 ] , then comparison to a centred Gaussian block with independent columns and unit diagonal covariance satisfies
lim inf N W 1 ( ν N U , p N , γ N U , p N ) ≥ 1 − 1 − c π > 0. (35) \liminf_{N} W_1\left(\nu_{N}^{\mathcal{U},p_N},\gamma_{N}^{\mathcal{U},p_N}\right) \ge\frac{1-\sqrt{1-c}}{\sqrt{\pi}} > 0. \tag*{(35)} N lim inf W 1 ( ν N U , p N , γ N U , p N ) ≥ π 1 − 1 − c > 0. (35) Thus p N = o ( N ) p_N = o(N) p N = o ( N ) is necessary for convergence in the unnormalized Euclidean W 1 W_1 W 1 metric.
Proof. It suffices to project onto one site, a 1-Lipschitz map. Let Θ \Theta Θ be the norm of its first p = p N p=p_N p = p N spin coordinates, and T T T the norm of a standard p p p -dimensional Gaussian vector. The Gaussian representation of uniform directions gives
Θ 2 = l a w N A p A p + B N − p , A p ∼ χ p 2 , B N − p ∼ χ N − p 2 independently , \Theta^2 \mathrel{\stackrel{\mathrm{law}}{=}} \frac{N A_p}{A_p+B_{N-p}}, \qquad A_p\sim\chi_p^2,\quad B_{N-p}\sim\chi_{N-p}^2\quad\text{independently}, Θ 2 = law A p + B N − p N A p , A p ∼ χ p 2 , B N − p ∼ χ N − p 2 independently ,
where B 0 = 0 B_0=0 B 0 = 0 . Thus E Θ 2 = p \mathbb{E}\Theta^2=p E Θ 2 = p and Var ( Θ 2 ) = 2 p ( N − p ) / ( N + 2 ) \operatorname{Var}(\Theta^2)=2p(N-p)/(N+2) Var ( Θ 2 ) = 2 p ( N − p ) / ( N + 2 ) , whence
E ( Θ − p ) 2 ≤ E ( Θ 2 − p ) 2 / p = 2 ( N − p ) / ( N + 2 ) ≤ 2. \mathbb{E}(\Theta-\sqrt{p})^2\le\mathbb{E}(\Theta^2-p)^2/p=2(N-p)/(N+2)\le2. E ( Θ − p ) 2 ≤ E ( Θ 2 − p ) 2 / p = 2 ( N − p ) / ( N + 2 ) ≤ 2.
If 0 < c < 1 0<c<1 0 < c < 1 , the two chi-square central limit theorems and expansion of N A p / ( A p + B N − p ) \sqrt{N A_p/(A_p+B_{N-p})} N A p / ( A p + B N − p ) at ( p , N − p ) (p,N-p) ( p , N − p ) give
Θ − p = 1 − p / N 2 p ( A p − p ) − p / N 2 p ( B N − p − ( N − p ) ) + o P ( 1 ) ⟹ N ( 0 , ( 1 − c ) / 2 ) . \Theta-\sqrt{p}=\frac{1-p/N}{2\sqrt{p}}(A_p-p)-\frac{p/N}{2\sqrt{p}}\left(B_{N-p}-(N-p)\right)+o_{\mathbb{P}}(1)\Longrightarrow\mathcal{N}(0,(1-c)/2). Θ − p = 2 p 1 − p / N ( A p − p ) − 2 p p / N ( B N − p − ( N − p ) ) + o P ( 1 ) ⟹ N ( 0 , ( 1 − c ) /2 ) .
Likewise T − p ⟹ N ( 0 , 1 / 2 ) T-\sqrt{p}\Longrightarrow\mathcal{N}(0,1/2) T − p ⟹ N ( 0 , 1/2 ) and E ( T − p ) 2 ≤ Var ( T 2 ) / p = 2 \mathbb{E}(T-\sqrt{p})^2\le\operatorname{Var}(T^2)/p=2 E ( T − p ) 2 ≤ Var ( T 2 ) / p = 2 . These second-moment bounds make the absolute centred norms uniformly integrable, so their expectations tend respectively to 1 − c / π \sqrt{1-c}/\sqrt{\pi} 1 − c / π and 1 / π 1/\sqrt{\pi} 1/ π . If c = 1 c=1 c = 1 , the bound for Θ − p \Theta-\sqrt{p} Θ − p tends to zero in L 2 L^2 L 2 , giving the same formula. The function u ↦ ∣ u ∣ − p u\mapsto\lvert u\rvert-\sqrt{p} u ↦ ∣ u ∣ − p is 1-Lipschitz. Its expectation difference proves (35) . If p N / N p_N/N p N / N does not tend to zero, a subsequence has a positive limiting ratio, which rules out convergence. □
Star lower bounds and the coordinate criterion
Proposition 6.2 (Finite-spin star lower bounds). Under (3), for every integer M ≥ 4 M\ge4 M ≥ 4 , N ≥ 1 N\ge1 N ≥ 1 and 1 ≤ p ≤ N 1\le p\le N 1 ≤ p ≤ N , with U 0 = { 0 , e 1 , … , e d } \mathcal{U}_0=\{0,e_1,\ldots,e_d\} U 0 = { 0 , e 1 , … , e d } , write
w = W 1 ( ν M , N , β , J U 0 , p , γ β , J U 0 , p ) . w=W_1\left(\nu_{M,N,\beta,J}^{\mathcal{U}_0,p},\gamma_{\beta,J}^{\mathcal{U}_0,p}\right). w = W 1 ( ν M , N , β , J U 0 , p , γ β , J U 0 , p ) . Then
w ≥ p 4 π ( 1 + 2 ) β j V . (36) w\ge\frac{\sqrt{p}}{4\sqrt{\pi}(1+\sqrt{2})\beta jV}. \tag*{(36)} w ≥ 4 π ( 1 + 2 ) β j V p . (36)
For N ≥ 4 N \ge4 N ≥ 4 , let
K P − = 8 ( 1 + 4 B d j 1 ) + 3 2 , C E = ( 8 K P − + 1 ) / b . K_P^- = 8(1+4Bd_{j1})+\frac{3}{2}, \qquad C_E = (8K_P^-+1)/b. K P − = 8 ( 1 + 4 B d j 1 ) + 2 3 , C E = ( 8 K P − + 1 ) / b .
We also have
w ≥ p 4 π ( 1 + 2 ) j max { 1 β V , s M − C E N } . (37) w \ge\frac{\sqrt{p}}{4\sqrt{\pi}(1+\sqrt{2})j}\max\left\{\frac{1}{\beta V},s_M-\frac{C_E}{\sqrt{N}}\right\}. \tag*{(37)} w ≥ 4 π ( 1 + 2 ) j p max { β V 1 , s M − N C E } . (37)
In particular w ≥ c p s M − C p / N w \ge c\sqrt{p}\,s_M-C\sqrt{p/N} w ≥ c p s M − C p / N uniformly on the compact parameter ranges. Both lower bounds hold for any fixed window containing a translated star. If N − 1 / 2 ≤ s M / ( 2 C E ) N^{-1/2}\le s_M/(2C_E) N − 1/2 ≤ s M / ( 2 C E ) , the mass order p s M \sqrt{p}\,s_M p s M is matched at finite spin dimension.
Proof. Let v i = E ( S 0 1 − S e i 1 ) 2 v_i=\mathbb{E}(S_0^1-S_{e_i}^1)^2 v i = E ( S 0 1 − S e i 1 ) 2 and let v i ∞ v_i^\infty v i ∞ be the corresponding single-column target variance. Hard length and the unit target diagonal give 0 ≤ v i , v i ∞ ≤ 4 0\le v_i,v_i^\infty\le4 0 ≤ v i , v i ∞ ≤ 4 . Set
ω = ∑ i J i ( v i ∞ − v i ) . \omega=\sum_i J_i(v_i^\infty-v_i). ω = i ∑ J i ( v i ∞ − v i ) .
The target constant mode cancels from each dipole. The exact Fourier identities and the spin energy derivative give
∑ i J i v i ∞ = 1 β , ∑ i J i v i = 2 ( j − P M , N ′ ) , ω = s M + 2 [ P M , N ′ − ( P M c o m p ) ′ ] . \sum_i J_i v_i^\infty=\frac{1}{\beta}, \qquad\sum_i J_i v_i=2(j-P'_{M,N}), \qquad\omega=s_M+2[P'_{M,N}-(P_M^{\mathrm{comp}})']. i ∑ J i v i ∞ = β 1 , i ∑ J i v i = 2 ( j − P M , N ′ ) , ω = s M + 2 [ P M , N ′ − ( P M comp ) ′ ] .
The last equality uses Lemma 4.3. Gaussian domination bounds v i ≤ Q M 0 ( δ 0 − δ e i ) / β v_i\le Q_M^0(\delta_0-\delta_{e_i})/\beta v i ≤ Q M 0 ( δ 0 − δ e i ) / β ; hence
ω ≥ 1 β V > 0 , \omega\ge\frac{1}{\beta V}>0, ω ≥ β V 1 > 0 ,
because ∑ i J i Q M 0 ( δ 0 − δ e i ) = 1 − V − 1 \sum_iJ_iQ_M^0(\delta_0-\delta_{e_i})=1-V^{-1} ∑ i J i Q M 0 ( δ 0 − δ e i ) = 1 − V − 1 ; by (22), ω = δ M , N + 1 / ( β V ) \omega=\delta_{M,N}+1/(\beta V) ω = δ M , N + 1/ ( β V ) .
We first turn ω \omega ω into a Wasserstein lower bound. Write χ ˉ q = E χ q \bar{\chi}_q=\mathbb{E}\chi_q χ ˉ q = E χ q and d N = N / χ ˉ N ≥ 1 d_N=\sqrt{N}/\bar{\chi}_N\ge1 d N = N / χ ˉ N ≥ 1 , where χ q \chi_q χ q is the norm of a standard Gaussian vector in R q \mathbb{R}^q R q . For any rotationally invariant Y ∈ R N Y\in\mathbb{R}^N Y ∈ R N , Haar averaging gives the exact identity
E ∣ Y ∣ p = χ ˉ p χ ˉ N E ∣ Y ∣ . \mathbb{E}|Y|^p=\frac{\bar{\chi}_p}{\bar{\chi}_N}\mathbb{E}|Y|. E ∣ Y ∣ p = χ ˉ N χ ˉ p E ∣ Y ∣.
Indeed, a standard Gaussian radius is independent of its uniform direction, so the expected norm of the first p p p coordinates of a uniform unit vector is χ ˉ p / χ ˉ N \bar{\chi}_p/\bar{\chi}_N χ ˉ p / χ ˉ N . Averaging a fixed vector of any radius proves the identity, including a possible atom at zero and N = 1 N=1 N = 1 . The one-site spin marginal is exactly uniform on the sphere of radius N \sqrt{N} N . Thus
E ∣ S 0 [ p ] ∣ = d N χ ˉ p , E ∣ ( S 0 − S e i ) [ p ] ∣ ≤ d N χ ˉ p v i , \mathbb{E}|S_0^{[p]}|=d_N\bar{\chi}_p,\qquad\mathbb{E}|(S_0-S_{e_i})^{[p]}|\le d_N\bar{\chi}_p\sqrt{v_i}, E ∣ S 0 [ p ] ∣ = d N χ ˉ p , E ∣ ( S 0 − S e i ) [ p ] ∣ ≤ d N χ ˉ p v i ,
where the inequality follows from Jensen and E ∣ S 0 − S e i ∣ 2 = N v i \mathbb{E}|S_0-S_{e_i}|^2=Nv_i E ∣ S 0 − S e i ∣ 2 = N v i .
Choose a direction with v i ∞ − v i ≥ ω / j > 0 v_i^\infty-v_i\ge\omega/j>0 v i ∞ − v i ≥ ω / j > 0 . On the star block z = ( z 0 , z e 1 , … , z e d ) z=(z_0,z_{e_1},\ldots,z_{e_d}) z = ( z 0 , z e 1 , … , z e d ) , consider
F i ( z ) = v i ∞ ∣ z 0 ∣ − ∣ z 0 − z e i ∣ . F_i(z)=\sqrt{v_i^\infty}\,|z_0|-|z_0-z_{e_i}|. F i ( z ) = v i ∞ ∣ z 0 ∣ − ∣ z 0 − z e i ∣.
Its Lipschitz constant for the Frobenius norm is at most 2 + 2 2+\sqrt{2} 2 + 2 . Its expectation under the target is zero: the two expected norms are χ ˉ p \bar{\chi}_p χ ˉ p and χ ˉ p v i ∞ \bar{\chi}_p\sqrt{v_i^\infty} χ ˉ p v i ∞ , respectively. Under the spin law,
E F i ≥ d N χ ˉ p ( v i ∞ − v i ) ≥ χ ˉ p ( v i ∞ − v i ) 4 . \mathbb{E}F_i\ge d_N\bar{\chi}_p\left(\sqrt{v_i^\infty}-\sqrt{v_i}\right)\ge\frac{\bar{\chi}_p(v_i^\infty-v_i)}{4}. E F i ≥ d N χ ˉ p ( v i ∞ − v i ) ≥ 4 χ ˉ p ( v i ∞ − v i ) .
Here v i ∞ + v i ≤ 4 \sqrt{v_i^\infty}+\sqrt{v_i}\le4 v i ∞ + v i ≤ 4 . The elementary inequality ∣ g ∣ ≥ p − 1 / 2 ∑ α = 1 p ∣ g α ∣ |g|\ge p^{-1/2}\sum_{\alpha=1}^{p}|g_\alpha| ∣ g ∣ ≥ p − 1/2 ∑ α = 1 p ∣ g α ∣ gives χ ˉ p ≥ 2 p / π \bar{\chi}_p\ge\sqrt{2p/\pi} χ ˉ p ≥ 2 p / π . Kantorovich duality therefore yields
w ≥ p ω 4 π ( 1 + 2 ) j . w\ge\frac{\sqrt{p}\,\omega}{4\sqrt{\pi}(1+\sqrt{2})j}. w ≥ 4 π ( 1 + 2 ) j p ω .
Combining this with ω ≥ 1 / ( β V ) \omega\ge1/(\beta V) ω ≥ 1/ ( β V ) proves (36) for every finite spin dimension.
For the mass lower bound take N ≥ 4 N\ge4 N ≥ 4 and h = b / ( 2 N ) h=b/(2\sqrt{N}) h = b / ( 2 N ) . The interval [ β − h , β ] [\beta-h,\beta] [ β − h , β ] lies in [ b / 2 , B ] [b/2,B] [ b /2 , B ] . Proposition 4.2 applies on this interval even when its left end is below β c , J \beta_{c,J} β c , J , and gives a pressure error at most K P − / N K_{\!P}^{-}/N K P − / N . Convexity of P M , N P_{M,N} P M , N and the curvature bound from Lemma 4.3 imply
P M , N ′ ( β ) ≥ P M , N ( β ) − P M , N ( β − h ) h ≥ ( P M c o m p ) ′ ( β ) − 2 K P − N h − h b 2 . P'_{M,N}(\beta)\ge\frac{P_{M,N}(\beta)-P_{M,N}(\beta-h)}{h}
\ge(P_M^{\mathrm{comp}})'(\beta)-\frac{2K_{\!P}^{-}}{Nh}-\frac{h}{b^2}. P M , N ′ ( β ) ≥ h P M , N ( β ) − P M , N ( β − h ) ≥ ( P M comp ) ′ ( β ) − N h 2 K P − − b 2 h .
The last term follows by integrating ( P M c o m p ) ′ ′ ≤ 2 / b 2 (P_M^{\mathrm{comp}})''\le2/b^2 ( P M comp ) ′′ ≤ 2/ b 2 to bound the difference between the derivative at β \beta β and the secant on [ β − h , β ] [\beta-h,\beta] [ β − h , β ] . Substitution of h h h gives
ω = s M + 2 [ P M , N ′ − ( P M c o m p ) ′ ] ≥ s M − 8 K P − + 1 b N . \omega=s_M+2[P'_{M,N}-(P_M^{\mathrm{comp}})']\ge s_M-\frac{8K_{\!P}^{-}+1}{b\sqrt{N}}. ω = s M + 2 [ P M , N ′ − ( P M comp ) ′ ] ≥ s M − b N 8 K P − + 1 .
Together with ω ≥ 1 / ( β V ) \omega\ge1/(\beta V) ω ≥ 1/ ( β V ) and the norm-test bound, this proves (37) . Projection onto a contained star is 1-Lipschitz; translation invariance supplies the claim for translated stars.
Proof of Theorem 2.3. The two finite-spin lower estimates are (36) , since β ≤ B \beta\le B β ≤ B and j ≤ d j 1 j\le d j_1 j ≤ d j 1 on (3), and the consequence w ≥ c p s M − C p / N w\ge c\sqrt{p}\,s_M-C\sqrt{p/N} w ≥ c p s M − C p / N of (37) stated in Proposition 6.2 . If the joint distance tends to zero, projection onto one site and the sphere obstruction of Proposition 6.1 force p n = o ( N n ) p_n=o(N_n) p n = o ( N n ) . The mass lower bound then gives
c p n s M n , β n , J n ≤ W 1 + C p n / N n ⟶ 0. c\sqrt{p_n}\,s_{M_n,\beta_n,J_n}\le W_1+C\sqrt{p_n/N_n}\longrightarrow0. c p n s M n , β n , J n ≤ W 1 + C p n / N n ⟶ 0.
Conversely, the two conditions make the upper bound in Theorem 2.1 tend to zero. A projection onto a translated star is 1-Lipschitz, so the same necessity holds on every fixed window containing such a star. By the uniform mass comparison in Proposition A.4 , the volume condition is equivalent to
p n = o ( [ V n { β n − β c , J n + t d ( M n ) } ] 2 ) . p_n=o\left([V_n\{\beta_n-\beta_{c,J_n}+t_d(M_n)\}]^2\right). p n = o ( [ V n { β n − β c , J n + t d ( M n )} ] 2 ) .
This gives the displayed conditions at criticality and under a uniform positive ordered gap.
Periodic accumulation states
Proof of Corollary 2.4. At fixed N N N , extend each torus configuration periodically to Z d \mathbb{Z}^d Z d . The product of the compact spin spheres on this countable site set is a compact metrizable space, so these laws have local weak accumulation points. Translation and global rotation invariance pass to the limit. These accumulation states satisfy the Dobrushin–Lanford–Ruelle equations of the infinite-lattice Gibbs specification. For a fixed finite region, test the finite-torus conditional identity with a continuous function of the spins in that region and a continuous exterior cylinder function. For all sufficiently large M M M , the region, its interaction boundary and the test supports embed without identifications in the torus. The conditional kernel then agrees with that on Z d \mathbb{Z}^d Z d and is continuous in the boundary spins. The identity passes to the weak limit; cylinder approximation and the monotone-class theorem extend it to bounded measurable tests. Each coordinate has second moment one, so the local first moments are uniformly integrable and the finite-volume bound passes to W 1 W_1 W 1 . For N ≥ max ( 4 , ℓ ) N \ge\max(4,\ell) N ≥ max ( 4 , ℓ ) , send M → ∞ M \to\infty M → ∞ in Theorem 2.1 . For N < max ( 4 , ℓ ) N < \max(4,\ell) N < max ( 4 , ℓ ) and a unit-diagonal target, the independent coupling bound is W 1 ≤ 2 ℓ p W_1 \le2\sqrt{\ell p} W 1 ≤ 2 ℓ p ; enlarging the uniform constant by 2 ℓ max ( 4 , ℓ ) 2\sqrt{\ell\max(4,\ell)} 2 ℓ max ( 4 , ℓ ) extends the C 1 p / N C_1\sqrt{p/N} C 1 p / N statement to all N ≥ 1 N \ge1 N ≥ 1 , p ≤ N p \le N p ≤ N . If p N / N p_N/N p N / N fails to tend to zero, it has a subsequence along which p N / N → c ∈ ( 0 , 1 ] p_N/N \to c \in(0,1] p N / N → c ∈ ( 0 , 1 ] . On that subsequence Proposition 6.1 gives lim inf W 1 ≥ ( 1 − 1 − c ) / π > 0 \liminf W_1 \ge(1-\sqrt{1-c})/\sqrt{\pi} > 0 lim inf W 1 ≥ ( 1 − 1 − c ) / π > 0 .
The ordered regime
Throughout this section the parameters lie in the compact ranges (3) . We additionally fix κ > 0 \kappa> 0 κ > 0 and assume β − β c , J ≥ κ \beta-\beta_{c,J} \ge\kappa β − β c , J ≥ κ . Put L N = log ( e N ) L_N=\log(eN) L N = log ( e N ) , and let ψ d ( N ) \psi_d(N) ψ d ( N ) be as in Theorem 2.8 . Constants are uniform in M , N , β , J M,N,\beta,J M , N , β , J ; constants in the local estimates may also depend on the fixed window. Sections 7.1–7.3 prove the deficit bound of Proposition 7.2 ; Section 7.4 gives the mixture lemma and the estimate in terms of δ M , N \delta_{M,N} δ M , N ; Section 7.5 proves Proposition 7.5 ; and Section 7.6 proves Theorem 2.8 .
One-sided quadratic fluctuations
For a zero-sum source a a a on the torus, write
U a = 1 N ∣ ∑ x a x S x ∣ 2 , K a = β − 1 Q M 0 ( a ) , D a = K a − E U a . U_a=\frac{1}{N}\left|\sum_x a_xS_x\right|^2,\qquad K_a=\beta^{-1}Q_M^0(a),\qquad D_a=K_a-\mathbb{E}U_a. U a = N 1 x ∑ a x S x 2 , K a = β − 1 Q M 0 ( a ) , D a = K a − E U a .
The source may have support depending on M M M . For a real variable F F F , let F + = max ( F , 0 ) F^+=\max(F,0) F + = max ( F , 0 ) and F − = max ( − F , 0 ) F^-=\max(-F,0) F − = max ( − F , 0 ) . The event estimate below keeps the mean deficit D a D_a D a linear while allowing the event to depend on a second source. Through the layer-cake argument it controls the products F b x y F a z F_{b_{xy}}F_{a_z} F b x y F a z used in the Ward identity argument of Section 7.3.
Lemma 7.1 (Quadratic fluctuations and their upper tail). For N ≥ 2 N \ge2 N ≥ 2 and F a = K a − U a F_a=K_a-U_a F a = K a − U a ,
E F a 2 ≤ 4 e 1 / 2 K a D a + 8 e 1 / 2 K a 2 / N , (38) \mathbb{E}F_a^2\le4e^{1/2}K_aD_a+8e^{1/2}K_a^2/N, \tag*{(38)} E F a 2 ≤ 4 e 1/2 K a D a + 8 e 1/2 K a 2 / N , (38) E ( F a − ) 2 ≤ C K a 2 / N . (39) \mathbb{E}(F_a^-)^2\le CK_a^2/N. \tag*{(39)} E ( F a − ) 2 ≤ C K a 2 / N . (39) In particular E ( F a + ) 2 ≤ C ( K a D a + K a 2 / N ) \mathbb{E}(F_a^+)^2\le C(K_aD_a+K_a^2/N) E ( F a + ) 2 ≤ C ( K a D a + K a 2 / N ) . For every event E E E ,
E [ F a + 1 E ] ≤ e D a + C K a P ( E ) / N . (40) \mathbb{E}[F_a^+\mathbf{1}_E]\le eD_a+CK_a\sqrt{\mathbb{P}(E)/N}. \tag*{(40)} E [ F a + 1 E ] ≤ e D a + C K a P ( E ) / N . (40) If K a , K c ≤ K ∗ K_a,K_c\le K_\ast K a , K c ≤ K ∗ for two zero-sum sources, then
− E ( F a F c ) ≤ C K ∗ { L N / N ( D a + D c ) + N − 1 } . (41) -\mathbb{E}(F_aF_c)\le CK_\ast\left\{\sqrt{L_N/N(D_a+D_c)}+N^{-1}\right\}. \tag*{(41)} − E ( F a F c ) ≤ C K ∗ { L N / N ( D a + D c ) + N − 1 } . (41)
Proof. If K a = 0 K_a=0 K a = 0 , domination forces U a = 0 U_a=0 U a = 0 almost surely. Otherwise Gaussian linearization of (20) gives
E e t U a / ( 2 K a ) ≤ ( 1 − t / N ) − N / 2 , 0 ≤ t < N . \mathbb{E}e^{tU_a/(2K_a)}\le(1-t/N)^{-N/2},\qquad0\le t<N. E e t U a / ( 2 K a ) ≤ ( 1 − t / N ) − N /2 , 0 ≤ t < N .
Set X = ( U a − K a ) / ( 2 K a ) ≥ − 1 / 2 X=(U_a-K_a)/(2K_a)\ge-1/2 X = ( U a − K a ) / ( 2 K a ) ≥ − 1/2 . At t = 1 t=1 t = 1 the bound is E e X ≤ e 1 / [ 4 ( N − 1 ) ] ≤ 1 + 1 / N \mathbb{E}e^X\le e^{1/[4(N-1)]}\le1+1/N E e X ≤ e 1/ [ 4 ( N − 1 )] ≤ 1 + 1/ N . The integral remainder in the exponential gives e x − 1 − x ≥ e − 1 / 2 x 2 / 2 e^x-1-x\ge e^{-1/2}x^2/2 e x − 1 − x ≥ e − 1/2 x 2 /2 for x ≥ − 1 / 2 x\ge-1/2 x ≥ − 1/2 . Using E X = − D a / ( 2 K a ) \mathbb{E}X=-D_a/(2K_a) E X = − D a / ( 2 K a ) proves (38). For (39), take t = N / 2 t=\sqrt{N}/2 t = N /2 . Then
log E e t X ≤ − t 2 − N 2 log ( 1 − t / N ) ≤ t 2 4 N ( 1 − t / N ) ≤ 1 4 . (42) \log\mathbb{E}e^{tX}\le-\frac{t}{2}-\frac{N}{2}\log(1-t/N)\le\frac{t^2}{4N(1-t/N)}\le\frac{1}{4}. \tag*{(42)} log E e tX ≤ − 2 t − 2 N log ( 1 − t / N ) ≤ 4 N ( 1 − t / N ) t 2 ≤ 4 1 . (42)
Consequently P ( U a − K a > u ) ≤ e 1 / 4 e − N u / ( 4 K a ) \mathbb{P}(U_a-K_a>u)\le e^{1/4}e^{-\sqrt{N}u/(4K_a)} P ( U a − K a > u ) ≤ e 1/4 e − N u / ( 4 K a ) . Integrating 2 u 2u 2 u times this tail over u > 0 u>0 u > 0 proves (39) . This uses domination for a single source only.
For (40) , first suppose N ≥ 16 N\ge16 N ≥ 16 and K a > 0 K_a>0 K a > 0 . Put f = F a / K a f=F_a/K_a f = F a / K a and d a = D a / K a d_a=D_a/K_a d a = D a / K a . For 0 ≤ λ ≤ N 0\le\lambda\le\sqrt{N} 0 ≤ λ ≤ N , Gaussian linearization and the logarithmic remainder bound give
E e − λ f ≤ e H N ( λ ) , H N ( λ ) = − λ − N 2 log ( 1 − 2 λ / N ) ≤ λ 2 N ( 1 − 2 λ / N ) ≤ 2 λ 2 N ≤ 2. \mathbb{E}e^{-\lambda f}\le e^{H_N(\lambda)},\qquad H_N(\lambda)=-\lambda-\frac{N}{2}\log(1-2\lambda/N)\le\frac{\lambda^2}{N(1-2\lambda/N)}\le\frac{2\lambda^2}{N}\le2. E e − λ f ≤ e H N ( λ ) , H N ( λ ) = − λ − 2 N log ( 1 − 2 λ / N ) ≤ N ( 1 − 2 λ / N ) λ 2 ≤ N 2 λ 2 ≤ 2.
The convex remainder e − λ f − 1 + λ f e^{-\lambda f}-1+\lambda f e − λ f − 1 + λ f is nonnegative for every real f f f , and hence
E { e − λ f − 1 + λ f } ≤ ( e 2 − 1 ) λ 2 / N + λ d a . \mathbb{E}\{e^{-\lambda f}-1+\lambda f\}\le(e^2-1)\lambda^2/N+\lambda d_a. E { e − λ f − 1 + λ f } ≤ ( e 2 − 1 ) λ 2 / N + λ d a .
For y ≥ 1 y\ge1 y ≥ 1 , e − y − 1 + y ≥ e − 1 y e^{-y}-1+y\ge e^{-1}y e − y − 1 + y ≥ e − 1 y ; indeed the ratio of the left side to y y y is increasing there and equals e − 1 e^{-1} e − 1 at one. For s ≥ N − 1 / 2 s\ge N^{-1/2} s ≥ N − 1/2 take λ = 1 / s \lambda=1/s λ = 1/ s and restrict the nonnegative remainder to { f > s } \{f>s\} { f > s } . This proves
E [ f + 1 { f > s } ] ≤ e d a + C / ( N s ) . \mathbb{E}[f^+\mathbf{1}_{\{f>s\}}]\le e d_a+C/(Ns). E [ f + 1 { f > s } ] ≤ e d a + C / ( N s ) .
If q = P ( E ) > 0 q=\mathbb{P}(E)>0 q = P ( E ) > 0 , choose s = ( N q ) − 1 / 2 s=(Nq)^{-1/2} s = ( N q ) − 1/2 and split according to f ≤ s f\le s f ≤ s and f > s f>s f > s . Then
E [ f + 1 E ] ≤ s q + e d a + C / ( N s ) ≤ e d a + C q / N . \mathbb{E}[f^+\mathbf{1}_E]\le sq+e d_a+C/(Ns)\le e d_a+C\sqrt{q/N}. E [ f + 1 E ] ≤ s q + e d a + C / ( N s ) ≤ e d a + C q / N .
For q = 0 q=0 q = 0 the claim is immediate. For 2 ≤ N ≤ 15 2\le N\le15 2 ≤ N ≤ 15 , use 0 ≤ f + ≤ 1 0\le f^+\le1 0 ≤ f + ≤ 1 and q ≤ 15 q / N q\le\sqrt{15}\sqrt{q/N} q ≤ 15 q / N . The case K a = 0 K_a=0 K a = 0 was already treated by domination. Multiplication by K a K_a K a proves (40) .
For (41) , the Chernoff tail bound in the proof of Lemma 5.1 gives, with
T a = 2 K a 3 L N / N + 6 K a L N / N ≤ C K a L N / N , T_a=2K_a\sqrt{3L_N/N}+6K_aL_N/N\le CK_a\sqrt{L_N/N}, T a = 2 K a 3 L N / N + 6 K a L N / N ≤ C K a L N / N ,
P ( F a − > T a ) ≤ e − 3 L N \mathbb{P}(F_a^->T_a)\le e^{-3L_N} P ( F a − > T a ) ≤ e − 3 L N . By (39) and Cauchy–Schwarz, E [ F a − 1 { F a − > T a } ] ≤ C K a N − 2 \mathbb{E}[F_a^-\mathbf{1}_{\{F_a^->T_a\}}]\le CK_aN^{-2} E [ F a − 1 { F a − > T a } ] ≤ C K a N − 2 . Since 0 ≤ F c + ≤ K c 0\le F_c^+\le K_c 0 ≤ F c + ≤ K c and the upper-tail bound used for (39) gives q a ( u ) : = P ( F a − > u ) ≤ e 1 / 4 e − N u / ( 4 K a ) q_a(u):=\mathbb{P}(F_a^->u)\le e^{1/4}e^{-\sqrt{N}u/(4K_a)} q a ( u ) := P ( F a − > u ) ≤ e 1/4 e − N u / ( 4 K a ) , we have
∫ 0 ∞ q a ( u ) d u ≤ C K a / N . \int_0^\infty\sqrt{q_a(u)}\,du\le CK_a/\sqrt{N}. ∫ 0 ∞ q a ( u ) d u ≤ C K a / N .
Use the layer-cake identity and (40) on the possibly correlated event { F a − > u } \{F_a^->u\} { F a − > u } :
E ( F a − F c + ) ≤ ∫ 0 T a E [ F c + 1 { F a − > u } ] d u + K c E [ F a − 1 { F a − > T a } ] ≤ e D c T a + C K c N ∫ 0 T a q a ( u ) d u + C K a K c N − 2 ≤ C K a D c L N / N + C K a K c / N . \begin{aligned}
\mathbb{E}(F_a^-F_c^+)&\le\int_0^{T_a}\mathbb{E}[F_c^+\mathbf{1}_{\{F_a^->u\}}]\,du+K_c\mathbb{E}[F_a^-\mathbf{1}_{\{F_a^->T_a\}}]\\
&\le eD_cT_a+\frac{CK_c}{\sqrt{N}}\int_0^{T_a}\sqrt{q_a(u)}\,du+CK_aK_cN^{-2}\\
&\le CK_aD_c\sqrt{L_N/N}+CK_aK_c/N.
\end{aligned} E ( F a − F c + ) ≤ ∫ 0 T a E [ F c + 1 { F a − > u } ] d u + K c E [ F a − 1 { F a − > T a } ] ≤ e D c T a + N C K c ∫ 0 T a q a ( u ) d u + C K a K c N − 2 ≤ C K a D c L N / N + C K a K c / N .
Interchange a , c a,c a , c and discard the two nonnegative products in F a F c = F a + F c + + F a − F c − − F a + F c − − F a − F c + F_aF_c=F_a^+F_c^++F_a^-F_c^--F_a^+F_c^--F_a^-F_c^+ F a F c = F a + F c + + F a − F c − − F a + F c − − F a − F c + . This proves (41) , including the cases of zero reference forms by continuity or domination.
Magnetization and spectral interpolation
Define
S ‾ = V − 1 ∑ x S x , R = ∣ S ‾ ∣ 2 / N , r = E R , t x = S x ⋅ S ‾ / N , \overline{S}=V^{-1}\sum_x S_x,\qquad R=|\overline{S}|^2/N,\qquad r=\mathbb{E}R,\qquad t_x=S_x\cdot\overline{S}/N, S = V − 1 x ∑ S x , R = ∣ S ∣ 2 / N , r = E R , t x = S x ⋅ S / N ,
a x = δ x − V − 1 1 , K 0 = 1 β V ∑ k ≠ 0 λ J ( k ) − 1 , r 0 = 1 − K 0 , η = r − r 0 . a_x=\delta_x-V^{-1}\mathbf{1},\qquad K_0=\frac{1}{\beta V}\sum_{k\ne0}\lambda_J(k)^{-1},\qquad r_0=1-K_0,\qquad\eta=r-r_0. a x = δ x − V − 1 1 , K 0 = β V 1 k = 0 ∑ λ J ( k ) − 1 , r 0 = 1 − K 0 , η = r − r 0 .
Here K a x = K 0 K_{ax}=K_0 K a x = K 0 , and hard length implies
U a x = 1 + R − 2 t x , V − 1 ∑ x U a x = 1 − R . U_{ax}=1+R-2t_x,\qquad V^{-1}\sum_x U_{ax}=1-R. U a x = 1 + R − 2 t x , V − 1 x ∑ U a x = 1 − R .
Translation invariance gives E t x = r \mathbb{E}t_x=r E t x = r and D a x = η ≥ 0 D_{ax}=\eta\ge0 D a x = η ≥ 0 . The diagonal quadrature estimate in Lemma A.2 yields K 0 = β c , J / β + O ( M 2 − d ) K_0=\beta_{c,J}/\beta+O(M^{2-d}) K 0 = β c , J / β + O ( M 2 − d ) . Thus, after increasing a threshold depending on κ \kappa κ ,
r 0 ≥ κ / ( 2 B ) > 0. (43) r_0\ge\kappa/(2B)>0. \tag*{(43)} r 0 ≥ κ / ( 2 B ) > 0. (43)
Let c ( z ) = E ( S 0 ⋅ S z ) / N c(z)=\mathbb{E}(S_0\cdot S_z)/N c ( z ) = E ( S 0 ⋅ S z ) / N and use the Fourier convention c ^ ( k ) = ∑ z ∈ Λ M c ( z ) e − i k ⋅ z \widehat{c}(k)=\sum_{z\in\Lambda_M}c(z)e^{-ik\cdot z} c ( k ) = ∑ z ∈ Λ M c ( z ) e − ik ⋅ z , so that V − 1 ∑ k c ^ ( k ) = c ( 0 ) = 1 V^{-1}\sum_k\widehat{c}(k)=c(0)=1 V − 1 ∑ k c ( k ) = c ( 0 ) = 1 . These Fourier eigenvalues are nonnegative. Gaussian domination gives, for k ≠ 0 k\ne0 k = 0 ,
d ( k ) = 1 β λ J ( k ) − c ^ ( k ) ≥ 0 , d ( k ) ≤ 1 β λ J ( k ) . d(k)=\frac{1}{\beta\lambda_J(k)}-\widehat{c}(k)\ge0,\qquad d(k)\le\frac{1}{\beta\lambda_J(k)}. d ( k ) = β λ J ( k ) 1 − c ( k ) ≥ 0 , d ( k ) ≤ β λ J ( k ) 1 .
With δ = δ M , N \delta=\delta_{M,N} δ = δ M , N from Lemma 4.3 ,
η = 1 V ∑ k ≠ 0 d ( k ) , δ = 1 V ∑ k ≠ 0 λ J ( k ) d ( k ) . (44) \eta=\frac{1}{V}\sum_{k\ne0}d(k),\qquad\delta=\frac{1}{V}\sum_{k\ne0}\lambda_J(k)d(k). \tag*{(44)} η = V 1 k = 0 ∑ d ( k ) , δ = V 1 k = 0 ∑ λ J ( k ) d ( k ) . (44)
For centred representatives k ∈ [ − π , π ] d k\in[-\pi,\pi]^d k ∈ [ − π , π ] d , λ J ( k ) ≍ ∣ k ∣ 2 \lambda_J(k)\asymp|k|^2 λ J ( k ) ≍ ∣ k ∣ 2 uniformly by the bounds on J J J . For 0 < u ≤ 1 0<u\le1 0 < u ≤ 1 , counting shells ∣ k ∣ ≍ n / M |k|\asymp n/M ∣ k ∣ ≍ n / M therefore gives
1 V ∑ 0 < λ J ( k ) ≤ u λ J ( k ) − 1 ≤ C u ( d − 2 ) / 2 . \frac{1}{V}\sum_{0<\lambda_J(k)\le u}\lambda_J(k)^{-1}\le C u^{(d-2)/2}. V 1 0 < λ J ( k ) ≤ u ∑ λ J ( k ) − 1 ≤ C u ( d − 2 ) /2 .
Indeed the shell sum is at most C M 2 − d ∑ n ≤ C M u n d − 3 CM^{2-d}\sum_{n\le CM\sqrt{u}}n^{d-3} C M 2 − d ∑ n ≤ C M u n d − 3 . If the low-frequency set is empty the estimate is immediate; otherwise M u M\sqrt{u} M u is bounded below and the displayed power follows. Splitting (44) at u u u gives η ≤ C u ( d − 2 ) / 2 + δ / u \eta\le Cu^{(d-2)/2}+\delta/u η ≤ C u ( d − 2 ) /2 + δ / u . For 0 < δ ≤ 1 0<\delta\le1 0 < δ ≤ 1 , take u = δ 2 / d u=\delta^{2/d} u = δ 2/ d . For δ = 0 \delta=0 δ = 0 , (44) gives η = 0 \eta=0 η = 0 ; larger δ \delta δ are covered by η ≤ K 0 ≤ C \eta\le K_0\le C η ≤ K 0 ≤ C . Hence
η ≤ C δ ( d − 2 ) / d . (45) \eta\le C\delta^{(d-2)/d}. \tag*{(45)} η ≤ C δ ( d − 2 ) / d . (45)
A Ward identity and the deficit bound
Let J x y = J i J_{xy}=J_i J x y = J i when y = x ± e i y=x\pm e_i y = x ± e i , and put
h x = ∑ y ∼ x J x y S y , e x = ∑ y ∼ x J x y ( 1 − S x ⋅ S y / N ) , k 0 = ( 1 − V − 1 ) / β . h_x=\sum_{y\sim x}J_{xy}S_y,\qquad e_x=\sum_{y\sim x}J_{xy}(1-S_x\cdot S_y/N),\qquad k_0=(1-V^{-1})/\beta. h x = y ∼ x ∑ J x y S y , e x = y ∼ x ∑ J x y ( 1 − S x ⋅ S y / N ) , k 0 = ( 1 − V − 1 ) / β .
In particular S x ⋅ h x / N = 2 j − e x S_x\cdot h_x/N=2j-e_x S x ⋅ h x / N = 2 j − e x and E e x = k 0 − δ \mathbb{E}e_x=k_0-\delta E e x = k 0 − δ . On the sphere at x x x , let P x = I − S x S x T / N P_x=I-S_xS_x^{\mathsf T}/N P x = I − S x S x T / N . The tangent vector field P x S ‾ P_x\overline{S} P x S has divergence ( N − 1 ) / V − ( N − 1 ) t x (N-1)/V-(N-1)t_x ( N − 1 ) / V − ( N − 1 ) t x . Integration by parts under the Gibbs density gives
0 = ( N − 1 ) ( V − 1 − r ) + β E { S ‾ ⋅ h x − t x S x ⋅ h x } . 0=(N-1)(V^{-1}-r)+\beta\mathbb{E}\{ \overline{S}\cdot h_x-t_xS_x\cdot h_x\}. 0 = ( N − 1 ) ( V − 1 − r ) + β E { S ⋅ h x − t x S x ⋅ h x } .
Translation invariance and ∑ x h x = 2 j ∑ x S x \sum_x h_x=2j\sum_x S_x ∑ x h x = 2 j ∑ x S x turn the last expectation into N E ( t x e x ) N\mathbb{E}(t_xe_x) N E ( t x e x ) . Therefore
β r δ = β Cov ( e x , t x ) + r N + 1 − r − 1 / N V . (46) \beta r\delta=\beta\operatorname{Cov}(e_x,t_x)+\frac{r}{N}+\frac{1-r-1/N}{V}. \tag*{(46)} β r δ = β Cov ( e x , t x ) + N r + V 1 − r − 1/ N . (46)
Set F x = K 0 − U a x F_x=K_0-U_{ax} F x = K 0 − U a x , F ‾ = V − 1 ∑ z F z = R − r 0 \overline{F}=V^{-1}\sum_zF_z=R-r_0 F = V − 1 ∑ z F z = R − r 0 , and f x = k 0 − e x f_x=k_0-e_x f x = k 0 − e x . Then
t x = r 0 + 1 2 ( F x + F ‾ ) , E f x = δ , E F x = E F ‾ = η , t_x=r_0+\frac{1}{2}(F_x+\overline{F}),\qquad\mathbb{E}f_x=\delta,\qquad\mathbb{E}F_x=\mathbb{E}\overline{F}=\eta, t x = r 0 + 2 1 ( F x + F ) , E f x = δ , E F x = E F = η ,
so that
Cov ( e x , t x ) = δ η − 1 2 E { f x ( F x + F ‾ ) } . (47) \operatorname{Cov}(e_x,t_x)=\delta\eta-\frac{1}{2}\mathbb{E}\{f_x(F_x+\overline{F})\}. \tag*{(47)} Cov ( e x , t x ) = δ η − 2 1 E { f x ( F x + F )} . (47)
For b x y = δ x − δ y b_{xy}=\delta_x-\delta_y b x y = δ x − δ y , f x = 1 2 ∑ y ∼ x J x y F b x y f_x=\frac{1}{2}\sum_{y\sim x}J_{xy}F_{b_{xy}} f x = 2 1 ∑ y ∼ x J x y F b x y . The reference equality 1 2 ∑ y ∼ x J x y K b x y = k 0 \frac{1}{2}\sum_{y\sim x}J_{xy}K_{b_{xy}}=k_0 2 1 ∑ y ∼ x J x y K b x y = k 0 follows from the Fourier trace of L J , M G J , M 0 L_{J,M}G^0_{J,M} L J , M G J , M 0 . Likewise 1 2 ∑ y ∼ x J x y D b x y = δ \frac{1}{2}\sum_{y\sim x}J_{xy}D_{b_{xy}}=\delta 2 1 ∑ y ∼ x J x y D b x y = δ . All source reference forms are bounded uniformly. Apply (41) separately to each pair b x y , a z b_{xy},a_z b x y , a z , and then sum with the bond weights and, when needed, the weights 1 / V 1/V 1/ V . Since D a z = η D_{a_z}=\eta D a z = η , this gives, for G = F x G=F_x G = F x or G = F ‾ G=\overline{F} G = F ,
E ( f x G ) ≥ − C { L N N ( δ + η ) + N − 1 } . \mathbb{E}(f_xG)\geq-C\left\{\sqrt{\frac{L_N}{N}(\delta+\eta)}+N^{-1}\right\}. E ( f x G ) ≥ − C { N L N ( δ + η ) + N − 1 } .
Inserting this into (46) and then (45) cancels δ η \delta\eta δ η against the same part of r δ = ( r 0 + η ) δ r\delta=(r_0+\eta)\delta r δ = ( r 0 + η ) δ :
r 0 δ ≤ C { L N N ( δ + η ) + N − 1 + V − 1 } . (48) r_0\delta\leq C\left\{\sqrt{\frac{L_N}{N}(\delta+\eta)}+N^{-1}+V^{-1}\right\}. \tag*{(48)} r 0 δ ≤ C { N L N ( δ + η ) + N − 1 + V − 1 } . (48)
Here 0 ≤ r ≤ 1 0\leq r\leq1 0 ≤ r ≤ 1 and β ≥ b \beta\geq b β ≥ b .
Proposition 7.2 (Pointwise ordered deficit). For N ≥ 4 N\geq4 N ≥ 4 and sufficiently large M M M , under β − β c , J ≥ κ \beta-\beta_{c,J}\geq\kappa β − β c , J ≥ κ ,
δ M , N ≤ C κ { ( L N N ) d / 4 + N − 1 + M − d } ≤ C κ { ψ d ( N ) + M − d } . (49) \delta_{M,N}\leq C_\kappa\left\{\left(\frac{L_N}{N}\right)^{d/4}+N^{-1}+M^{-d}\right\}\leq C_\kappa\{\psi_d(N)+M^{-d}\}. \tag*{(49)} δ M , N ≤ C κ { ( N L N ) d /4 + N − 1 + M − d } ≤ C κ { ψ d ( N ) + M − d } . (49)
Proof. For sufficiently large N N N , absorb the C L N / N δ C\sqrt{L_N/N}\,\delta C L N / N δ term using (42). Combine (44) and (47). Young’s inequality, with any sufficiently small fixed ϵ > 0 \epsilon>0 ϵ > 0 , gives
C L N N δ ( d − 2 ) / d ≤ ϵ δ + C ϵ ( L N N ) d / 4 . C\sqrt{\frac{L_N}{N}}\delta^{(d-2)/d}\leq\epsilon\delta+C_\epsilon\left(\frac{L_N}{N}\right)^{d/4}. C N L N δ ( d − 2 ) / d ≤ ϵδ + C ϵ ( N L N ) d /4 .
Absorb the linear term. The finitely many smaller N ≥ 4 N\geq4 N ≥ 4 are covered by 0 ≤ δ ≤ 1 / b 0\leq\delta\leq1/b 0 ≤ δ ≤ 1/ b . For d ≥ 5 d\geq5 d ≥ 5 , the ratio N ( L N / N ) d / 4 = N 1 − d / 4 L N d / 4 N(L_N/N)^{d/4}=N^{1-d/4}L_N^{d/4} N ( L N / N ) d /4 = N 1 − d /4 L N d /4 is bounded on N ≥ 4 N\geq4 N ≥ 4 ; the cases d = 3 , 4 d=3,4 d = 3 , 4 follow directly. This proves (48).
Gaussian mixtures and an estimate in terms of the deficit
Lemma 7.3 (Gaussian mixtures with explicit coordinate dependence). Fix ℓ \ell ℓ and m > 0 m>0 m > 0 . Let A A A be a random positive semidefinite ℓ × ℓ \ell\times\ell ℓ × ℓ matrix of trace ℓ \ell ℓ , and let A ‾ = E A ⪰ m I \overline{A}=\mathbb{E}A\succeq mI A = E A ⪰ m I . For an independent standard Gaussian matrix Z ∈ R ℓ × p Z\in\mathbb{R}^{\ell\times p} Z ∈ R ℓ × p ,
W 1 ( L ( A 1 / 2 Z ) , L ( A ‾ 1 / 2 Z ) ) ≤ C ℓ , m , p E ∥ A − A ‾ ∥ F 2 , p ≥ 1. W_1\bigl(\mathcal{L}(A^{1/2}Z),\mathcal{L}(\overline{A}^{1/2}Z)\bigr)\leq C_{\ell,m,p}\mathbb{E}\lVert A-\overline{A}\rVert_F^2,\qquad p\geq1. W 1 ( L ( A 1/2 Z ) , L ( A 1/2 Z ) ) ≤ C ℓ , m , p E ∥ A − A ∥ F 2 , p ≥ 1. If Σ ⪰ m I \Sigma\succeq mI Σ ⪰ m I also has trace ℓ \ell ℓ , then
W 1 ( L ( A ‾ 1 / 2 Z ) , L ( Σ 1 / 2 Z ) ) ≤ C m p ∥ A ‾ − Σ ∥ F . W_1\bigl(\mathcal{L}(\overline{A}^{1/2}Z),\mathcal{L}(\Sigma^{1/2}Z)\bigr)\leq C_m\sqrt{p}\lVert\overline{A}-\Sigma\rVert_F. W 1 ( L ( A 1/2 Z ) , L ( Σ 1/2 Z ) ) ≤ C m p ∥ A − Σ ∥ F . Proof. For a 1-Lipschitz function g g g with g ( 0 ) = 0 g(0)=0 g ( 0 ) = 0 , set F g ( Ξ ) = E g ( Ξ 1 / 2 Z ) \mathcal{F}_g(\Xi)=\mathbb{E}g(\Xi^{1/2}Z) F g ( Ξ ) = E g ( Ξ 1/2 Z ) . First suppose g g g is smooth. On m I / 2 ⪯ Ξ ⪯ ℓ I mI/2\preceq\Xi\preceq\ell I m I /2 ⪯ Ξ ⪯ ℓ I , write X = Ξ 1 / 2 Z X=\Xi^{1/2}Z X = Ξ 1/2 Z and, for a symmetric direction H H H , let
L H ( X ) = − p 2 tr ( Ξ − 1 H ) + 1 2 ∑ α = 1 p X α T Ξ − 1 H Ξ − 1 X α . L_H(X)=-\frac{p}{2}\operatorname{tr}(\Xi^{-1}H)+\frac{1}{2}\sum_{\alpha=1}^{p}X_\alpha^{\mathsf{T}}\Xi^{-1}H\Xi^{-1}X_\alpha. L H ( X ) = − 2 p tr ( Ξ − 1 H ) + 2 1 α = 1 ∑ p X α T Ξ − 1 H Ξ − 1 X α . Differentiating the Gaussian density and integrating by parts once gives
D F g ( Ξ ) [ H ] = 1 2 E ∑ α ∇ α g ( X ) ⋅ H Ξ − 1 X α . (50) D\mathcal{F}_g(\Xi)[H]=\frac{1}{2}\mathbb{E}\sum_\alpha\nabla_\alpha g(X)\cdot H\Xi^{-1}X_\alpha. \tag*{(50)} D F g ( Ξ ) [ H ] = 2 1 E α ∑ ∇ α g ( X ) ⋅ H Ξ − 1 X α . (50)
Since ∥ ∇ g ∥ F ≤ 1 \lVert\nabla g\rVert_F \le1 ∥ ∇ g ∥ F ≤ 1 , this is bounded by C m p ∥ H ∥ F C_m\sqrt{p}\,\lVert H\rVert_F C m p ∥ H ∥ F . Differentiating this expression with X X X as the integration variable gives
D 2 F g ( Ξ ) [ H , H ] = 1 2 E ∑ α ∇ α g ( X ) ⋅ { H Ξ − 1 X α L H ( X ) − H Ξ − 1 H Ξ − 1 X α } . D^2F_g(\Xi)[H,H]=\frac{1}{2}\mathbb{E}\sum_\alpha\nabla_\alpha g(X)\cdot\left\{H\Xi^{-1}X_\alpha L_H(X)-H\Xi^{-1}H\Xi^{-1}X_\alpha\right\}. D 2 F g ( Ξ ) [ H , H ] = 2 1 E α ∑ ∇ α g ( X ) ⋅ { H Ξ − 1 X α L H ( X ) − H Ξ − 1 H Ξ − 1 X α } .
Here E L H = 0 \mathbb{E}L_H=0 E L H = 0 and E L H 2 = ( p / 2 ) tr ( Ξ − 1 H Ξ − 1 H ) \mathbb{E}L_H^2=(p/2)\operatorname{tr}(\Xi^{-1}H\Xi^{-1}H) E L H 2 = ( p /2 ) tr ( Ξ − 1 H Ξ − 1 H ) . Also E ∥ H Ξ − 1 X ∥ F 2 = p tr ( H Ξ − 1 H ) \mathbb{E}\lVert H\Xi^{-1}X\rVert_F^2=p\operatorname{tr}(H\Xi^{-1}H) E ∥ H Ξ − 1 X ∥ F 2 = p tr ( H Ξ − 1 H ) . Cauchy–Schwarz bounds the first term by C m p ∥ H ∥ F 2 C_mp\lVert H\rVert_F^2 C m p ∥ H ∥ F 2 , and the second by C m p ∥ H ∥ F 2 C_m\sqrt{p}\,\lVert H\rVert_F^2 C m p ∥ H ∥ F 2 . Thus ∣ D 2 F g ( Ξ ) [ H , H ] ∣ ≤ C m p ∥ H ∥ F 2 \lvert D^2F_g(\Xi)[H,H]\rvert\le C_mp\lVert H\rVert_F^2 ∣ D 2 F g ( Ξ ) [ H , H ]∣ ≤ C m p ∥ H ∥ F 2 . Smooth mollification and Gaussian moment bounds extend these formulas and estimates to every Lipschitz g g g .
Put H = A − A ˉ H=A-\bar A H = A − A ˉ . If ∥ H ∥ o p ≤ m / 2 \lVert H\rVert_{\mathrm{op}}\le m/2 ∥ H ∥ op ≤ m /2 , Taylor’s theorem along A ˉ + t H \bar A+tH A ˉ + t H gives a remainder of at most C m p ∥ H ∥ F 2 C_mp\lVert H\rVert_F^2 C m p ∥ H ∥ F 2 . Otherwise ∥ H ∥ F > m / 2 \lVert H\rVert_F>m/2 ∥ H ∥ F > m /2 , while ∣ F g ( A ) ∣ , ∣ F g ( A ˉ ) ∣ ≤ ℓ p \lvert F_g(A)\rvert,\lvert F_g(\bar A)\rvert\le\sqrt{\ell p} ∣ F g ( A )∣ , ∣ F g ( A ˉ )∣ ≤ ℓ p and ∣ D F g ( A ˉ ) [ H ] ∣ ≤ C m p ∥ H ∥ F \lvert DF_g(\bar A)[H]\rvert\le C_m\sqrt{p}\,\lVert H\rVert_F ∣ D F g ( A ˉ ) [ H ]∣ ≤ C m p ∥ H ∥ F . These bounds give the same quadratic remainder with a constant depending only on ℓ , m \ell,m ℓ , m . Taking expectations eliminates the linear term because E H = 0 \mathbb{E}H=0 E H = 0 . Kantorovich duality proves the first claim.
For the second, couple the two Gaussian matrices with the same Z Z Z . The square-root estimate from Proposition 5.3 gives ∥ A ˉ 1 / 2 − Σ 1 / 2 ∥ F ≤ C m ∥ A ˉ − Σ ∥ F \lVert\bar A^{1/2}-\Sigma^{1/2}\rVert_F\le C_m\lVert\bar A-\Sigma\rVert_F ∥ A ˉ 1/2 − Σ 1/2 ∥ F ≤ C m ∥ A ˉ − Σ ∥ F ; the Gaussian coupling therefore has expected distance at most C m p ∥ A ˉ − Σ ∥ F C_m\sqrt{p}\lVert\bar A-\Sigma\rVert_F C m p ∥ A ˉ − Σ ∥ F .
Proposition 7.4 (Local comparison with explicit coordinate dependence). Fix U \mathcal{U} U . Under the assumptions of this section, for sufficiently large M M M , N ≥ max ( 4 , ℓ ) N\ge\max(4,\ell) N ≥ max ( 4 , ℓ ) and 1 ≤ p ≤ N 1\le p\le N 1 ≤ p ≤ N , the distance W 1 = W 1 ( ν M , N , β , J U , γ β , J U ) W_1=W_1(\nu_{M,N,\beta,J}^{\mathcal{U}},\gamma_{\beta,J}^{\mathcal{U}}) W 1 = W 1 ( ν M , N , β , J U , γ β , J U ) satisfies
W 1 ≤ C U { p ( δ M , N + N − 1 ) + p V } . (51) W_1\le C_{\mathcal{U}}\left\{p(\delta_{M,N}+N^{-1})+\frac{\sqrt{p}}{V}\right\}. \tag*{(51)} W 1 ≤ C U { p ( δ M , N + N − 1 ) + V p } . (51) The constant is independent of p p p . In particular, for each fixed p p p the bound is C U , p ( δ M , N + N − 1 + M − d ) C_{\mathcal{U},p}(\delta_{M,N}+N^{-1}+M^{-d}) C U , p ( δ M , N + N − 1 + M − d ) .
Proof. Let A i j = S x i ⋅ S x j / N A_{ij}=S_{x_i}\cdot S_{x_j}/N A ij = S x i ⋅ S x j / N and Σ = Σ β , J ∣ U \Sigma=\Sigma_{\beta,J}\lvert_{\mathcal{U}} Σ = Σ β , J ∣ U . They are positive semidefinite with trace ℓ \ell ℓ , and Σ ⪰ m I \Sigma\succeq mI Σ ⪰ m I by Lemma 5.2 . For a i j = δ x i − δ x j a_{ij}=\delta_{x_i}-\delta_{x_j} a ij = δ x i − δ x j , hard length and the unit target diagonal give
E ( A i j − Σ i j ) = 1 2 D a i j − 1 2 ( K a i j − K ∞ ( a i j ) ) . \mathbb{E}(A_{ij}-\Sigma_{ij})=\frac{1}{2}D_{a_{ij}}-\frac{1}{2}(K_{a_{ij}}-K_\infty(a_{ij})). E ( A ij − Σ ij ) = 2 1 D a ij − 2 1 ( K a ij − K ∞ ( a ij )) . Apply (38) to a i j a_{ij} a ij , the fixed-path deficit estimate from Lemma 5.1 , and the dipole quadrature estimate from Lemma A.2 . Summing over the fixed number of entries yields
∥ E A − Σ ∥ F ≤ C U ( δ + M − d ) , E ∥ A − Σ ∥ F 2 ≤ C U ( δ + N − 1 + M − 2 d ) . (52) \lVert\mathbb{E}A-\Sigma\rVert_F\le C_{\mathcal{U}}(\delta+M^{-d}),\qquad\mathbb{E}\lVert A-\Sigma\rVert_F^2\le C_{\mathcal{U}}(\delta+N^{-1}+M^{-2d}). \tag*{(52)} ∥ E A − Σ ∥ F ≤ C U ( δ + M − d ) , E ∥ A − Σ ∥ F 2 ≤ C U ( δ + N − 1 + M − 2 d ) . (52) Put A ˉ = E A \bar A=\mathbb{E}A A ˉ = E A . For a = a i j a=a_{ij} a = a ij , write U a = ∣ ∑ x a x S x ∣ 2 / N U_a=\lvert\sum_xa_xS_x\rvert^2/N U a = ∣ ∑ x a x S x ∣ 2 / N and K a = Q 0 M ( a ) / β K_a=Q_0^M(a)/\beta K a = Q 0 M ( a ) / β . The centred variance satisfies
Var ( A i j ) = 1 4 Var ( U a ) ≤ 1 4 E ( U a − K a ) 2 ≤ C U ( δ + N − 1 ) \operatorname{Var}(A_{ij})=\frac{1}{4}\operatorname{Var}(U_a)\le\frac{1}{4}\mathbb{E}(U_a-K_a)^2\le C_{\mathcal{U}}(\delta+N^{-1}) Var ( A ij ) = 4 1 Var ( U a ) ≤ 4 1 E ( U a − K a ) 2 ≤ C U ( δ + N − 1 ) by (38) and the fixed-path bound for D a D_a D a . With (52) , the fixed-path estimates therefore give
∥ A ˉ − Σ ∥ F ≤ C U ( δ M , N + V − 1 ) , E ∥ A − A ˉ ∥ F 2 ≤ C U ( δ M , N + N − 1 ) . \lVert\bar A-\Sigma\rVert_F\le C_{\mathcal{U}}(\delta_{M,N}+V^{-1}),\qquad\mathbb{E}\lVert A-\bar A\rVert_F^2\le C_{\mathcal{U}}(\delta_{M,N}+N^{-1}). ∥ A ˉ − Σ ∥ F ≤ C U ( δ M , N + V − 1 ) , E ∥ A − A ˉ ∥ F 2 ≤ C U ( δ M , N + N − 1 ) . If δ \delta δ is below a sufficiently small fixed threshold and M M M is large enough, A ˉ ⪰ m I / 2 \bar A\succeq mI/2 A ˉ ⪰ m I /2 . For an independent standard Gaussian matrix Z ∈ R ℓ × p Z\in\mathbb{R}^{\ell\times p} Z ∈ R ℓ × p , Lemma 7.3 applied with lower eigenvalue bound m / 2 m/2 m /2 then gives
W 1 ( L ( A 1 / 2 Z ) , L ( Σ 1 / 2 Z ) ) ≤ C ℓ , m { p E ∥ A − A ˉ ∥ F 2 + p ∥ A ˉ − Σ ∥ F } . W_1\bigl(\mathcal{L}(A^{1/2}Z),\mathcal{L}(\Sigma^{1/2}Z)\bigr)\le C_{\ell,m}\left\{p\mathbb{E}\lVert A-\bar A\rVert_F^2+\sqrt{p}\,\lVert\bar A-\Sigma\rVert_F\right\}. W 1 ( L ( A 1/2 Z ) , L ( Σ 1/2 Z ) ) ≤ C ℓ , m { p E ∥ A − A ˉ ∥ F 2 + p ∥ A ˉ − Σ ∥ F } .
Finally use the orbit kernel in Proposition 5.3 : X E p = l a w A 1 / 2 Q N XE_p \mathrel{\overset{\mathrm{law}}{=}} A^{1/2}Q_N X E p = law A 1/2 Q N , with Q N = N P H E p Q_N = \sqrt{N}PHE_p Q N = N P H E p independent of A A A . Theorem 4.5 of Chatterjee and Meckes [8 p. 273] , applied to the ℓ p \ell_p ℓ p matrices N E i j \sqrt{N}E_{ij} N E ij , gives
W 1 ( L ( Q N ) , L ( Z ) ) ≤ 2 ℓ p N − 1 . (53) W_1(\mathcal{L}(Q_N),\mathcal{L}(Z)) \le\frac{\sqrt{2}\,\ell_p}{N-1}. \tag*{(53)} W 1 ( L ( Q N ) , L ( Z )) ≤ N − 1 2 ℓ p . (53)
Their Hilbert–Schmidt inner products are N δ i i ′ δ j j ′ N\delta_{ii'}\delta_{jj'} N δ i i ′ δ j j ′ , as required by that theorem; the Euclidean norm of the vectorized block is the Frobenius norm used here. The theorem does not require ℓ p ≤ N \ell_p \le N ℓ p ≤ N , so this bound covers every p ≤ N p \le N p ≤ N . Couple Q N , Z Q_N,Z Q N , Z independently of A A A and use ∥ A 1 / 2 ∥ o p ≤ ℓ \|A^{1/2}\|_{\mathrm{op}} \le\sqrt{\ell} ∥ A 1/2 ∥ op ≤ ℓ . Combining this Haar error with the mixture bound and the two fixed-path estimates proves (49) when δ \delta δ is below the threshold. If δ \delta δ is at or above this threshold, W 1 ≤ 2 ℓ p W_1 \le2\sqrt{\ell_p} W 1 ≤ 2 ℓ p is bounded by C p δ C p\delta C p δ , so the same estimate holds. This also covers a singular A ˉ \bar{A} A ˉ . For fixed p p p , (49) is the bound C U , p ( δ M , N + N − 1 + M − d ) C_{U,p}(\delta_{M,N}+N^{-1}+M^{-d}) C U , p ( δ M , N + N − 1 + M − d ) , since V − 1 = M − d V^{-1}=M^{-d} V − 1 = M − d .
Spin dimension larger than volume
This subsection proves the following estimate.
Proposition 7.5 (Ordered estimate for spin dimension larger than volume). Fix κ > 0 \kappa> 0 κ > 0 . Under (3) and β − β c , J ≥ κ \beta-\beta_{c,J} \ge\kappa β − β c , J ≥ κ , there are C U , κ C_{U,\kappa} C U , κ and M 0 M_0 M 0 such that, for every integer M ≥ M 0 M \ge M_0 M ≥ M 0 , V = M d V=M^d V = M d , N > V + 1 N>V+1 N > V + 1 and 1 ≤ p ≤ N 1 \le p \le N 1 ≤ p ≤ N ,
W 1 ( ν M , N , β , J U , p , γ β , J U , p ) ≤ C U , κ { p V + p N − V − 1 } . W_1(\nu^{U,p}_{M,N,\beta,J},\gamma^{U,p}_{\beta,J}) \le C_{U,\kappa}\left\{\frac{\sqrt{p}}{V}+\frac{p}{N-V-1}\right\}. W 1 ( ν M , N , β , J U , p , γ β , J U , p ) ≤ C U , κ { V p + N − V − 1 p } . The constant is independent of p , M , N , β , J p,M,N,\beta,J p , M , N , β , J in these ranges. If N ≥ 2 ( V + 1 ) N \ge2(V+1) N ≥ 2 ( V + 1 ) , the bound is C U , κ ( p / V + p / N ) C_{U,\kappa}(\sqrt{p}/V+p/N) C U , κ ( p / V + p / N ) . For a window containing a translated nearest-neighbour star, if N ≥ 2 ( V + 1 ) N \ge2(V+1) N ≥ 2 ( V + 1 ) and V p ≤ N V\sqrt{p} \le N V p ≤ N , then
c p / V ≤ W 1 ( ν M , N , β , J U , p , γ β , J U , p ) ≤ C U , κ p / V . c\sqrt{p}/V \le W_1(\nu^{U,p}_{M,N,\beta,J},\gamma^{U,p}_{\beta,J}) \le C_{U,\kappa}\sqrt{p}/V. c p / V ≤ W 1 ( ν M , N , β , J U , p , γ β , J U , p ) ≤ C U , κ p / V . For star-containing windows, the total error is of order V − 1 V^{-1} V − 1 for each fixed p p p when N ≥ 2 ( V + 1 ) N \ge2(V+1) N ≥ 2 ( V + 1 ) .
The density of the full Gram matrix is a standard matrix-distribution formula; for this spin model see Tang, Khoo and Ying [33 Theorem 4] . In their notation, n = N n=N n = N , k = V k=V k = V , the unit spin is S x / N S_x/\sqrt{N} S x / N , and the symmetric interaction matrix has entries J x y / 2 J_{xy}/2 J x y /2 on neighbouring pairs, so its double sum counts each edge once. Combining this density with a Brascamp–Lieb-type variance bound [5 ] in the form of [26 arXiv v2, Theorem 1.1] gives the variance bounds below without a factor depending on the volume.
We keep S ˉ , R , t x , K 0 , r 0 \bar{S},R,t_x,K_0,r_0 S ˉ , R , t x , K 0 , r 0 from § 7.2 and h x , e x h_x,e_x h x , e x from § 7.3 , and write δ = δ M , N = ( 1 − V − 1 ) / β − E e x ≥ 0 \delta=\delta_{M,N}=(1-V^{-1})/\beta-\mathbb{E}e_x \ge0 δ = δ M , N = ( 1 − V − 1 ) / β − E e x ≥ 0 . By (42), r 0 ≥ r ∗ : = κ / ( 2 B ) > 0 r_0 \ge r_*:=\kappa/(2B)>0 r 0 ≥ r ∗ := κ / ( 2 B ) > 0 for M ≥ M 0 M \ge M_0 M ≥ M 0 . Set a c u t = r 0 / 4 a_{\mathrm{cut}}=r_0/4 a cut = r 0 /4 . The good-event longitudinal energy is
G = 1 V ∑ { x , y } J x y E [ 1 { t x , t y ≥ a c u t } ( t x − t y ) 2 t x t y ] . \mathcal{G}=\frac{1}{V}\sum_{\{x,y\}}J_{xy}\mathbb{E}\left[1_{\{t_x,t_y\ge a_{\mathrm{cut}}\}}\frac{(t_x-t_y)^2}{t_xt_y}\right]. G = V 1 { x , y } ∑ J x y E [ 1 { t x , t y ≥ a cut } t x t y ( t x − t y ) 2 ] .
The sum is over unoriented edges, each once.
Lemma 7.6. If N > V + 1 N>V+1 N > V + 1 , then
Var ( t x − t y ) ≤ C N − V − 1 \operatorname{Var}(t_x-t_y)\le\frac{C}{N-V-1} Var ( t x − t y ) ≤ N − V − 1 C for all sites x , y x,y x , y , with an absolute constant. Moreover, Var ( A x y ) ≤ C / ( N − V − 1 ) \operatorname{Var}(A_{xy})\le C/(N-V-1) Var ( A x y ) ≤ C / ( N − V − 1 ) for A x y = S x ⋅ S y / N A_{xy}=S_x\cdot S_y/N A x y = S x ⋅ S y / N .
Proof. Under the product sphere prior, the full Gram matrix A x y = S x ⋅ S y / N A_{xy}=S_x\cdot S_y/N A x y = S x ⋅ S y / N has density c N , V ( det A ) α c_{N,V}(\det A)^\alpha c N , V ( det A ) α on the open convex set A = { A ≻ 0 : diag A = 1 } \mathcal{A}=\{A\succ0:\operatorname{diag} A=1\} A = { A ≻ 0 : diag A = 1 } , with α = ( N − V − 1 ) / 2 \alpha=(N-V-1)/2 α = ( N − V − 1 ) /2 , relative to off-diagonal Lebesgue measure. This is also the density in Tang, Khoo and Ying [33 Theorem 4, Eq. (2.2)] at β = 0 \beta=0 β = 0 . Inductively, suppose v − 1 v-1 v − 1 independent unit sphere vectors have Gram matrix A v − 1 A_{v-1} A v − 1 . The coordinates z z z of the next vector on their orthonormal span have density proportional to ( 1 − ∣ z ∣ 2 ) ( N − v − 1 ) / 2 (1-|z|^2)^{(N-v-1)/2} ( 1 − ∣ z ∣ 2 ) ( N − v − 1 ) /2 . The change of variables from z z z to its inner products u u u with the preceding vectors has Jacobian ( det A v − 1 ) 1 / 2 (\det A_{v-1})^{1/2} ( det A v − 1 ) 1/2 . The Schur complement gives
det A v = det A v − 1 ( 1 − u T A v − 1 − 1 u ) . \det A_v=\det A_{v-1}(1-u^{\mathsf T}A_{v-1}^{-1}u). det A v = det A v − 1 ( 1 − u T A v − 1 − 1 u ) .
After including the reciprocal Jacobian and the inductive density ( det A v − 1 ) ( N − v ) / 2 (\det A_{v-1})^{(N-v)/2} ( det A v − 1 ) ( N − v ) /2 , every power of det A v − 1 \det A_{v-1} det A v − 1 cancels, leaving ( det A v ) ( N − v − 1 ) / 2 (\det A_v)^{(N-v-1)/2} ( det A v ) ( N − v − 1 ) /2 . Starting with v = 1 v=1 v = 1 proves the formula. The full-rank assertion holds almost surely for N ≥ V N\ge V N ≥ V .
The Gibbs tilt is affine in A A A . Thus its negative log density is, up to a constant,
f ( A ) = − α log det A − β N ∑ { z , w } J z w A z w ( A ∈ A ) , f(A)=-\alpha\log\det A-\beta N\sum_{\{z,w\}}J_{zw}A_{zw}\qquad(A\in\mathcal{A}), f ( A ) = − α log det A − β N { z , w } ∑ J z w A z w ( A ∈ A ) ,
and is + ∞ +\infty + ∞ outside A \mathcal{A} A . For symmetric tangent matrices H H H with zero diagonal,
D 2 f ( A ) [ H , H ] = α tr ( A − 1 H A − 1 H ) > 0 ( H ≠ 0 ) . D^2f(A)[H,H]=\alpha\operatorname{tr}(A^{-1}HA^{-1}H)>0\qquad(H\ne0). D 2 f ( A ) [ H , H ] = α tr ( A − 1 H A − 1 H ) > 0 ( H = 0 ) .
Use Frobenius-orthonormal coordinates on this tangent space. Changing from off-diagonal coordinates affects only the density’s constant. For any symmetric L L L , the inverse-Hessian quadratic form of the restricted gradient of tr ( L A ) \operatorname{tr}(LA) tr ( L A ) is bounded by
sup diag H = 0 { 2 tr ( L H ) − α tr ( A − 1 H A − 1 H ) } ≤ tr ( A L A L ) α . \sup_{\operatorname{diag}H=0}\left\{2\operatorname{tr}(LH)-\alpha\operatorname{tr}(A^{-1}HA^{-1}H)\right\}\le\frac{\operatorname{tr}(ALAL)}{\alpha}. diag H = 0 sup { 2 tr ( L H ) − α tr ( A − 1 H A − 1 H ) } ≤ α tr ( A L A L ) .
The last inequality enlarges the supremum to all symmetric H H H and completes the square. Take b = δ x − δ y b=\delta_x-\delta_y b = δ x − δ y , 1 = ( 1 , … , 1 ) T \mathbf{1}=(1,\ldots,1)^{\mathsf T} 1 = ( 1 , … , 1 ) T and L = ( b 1 T + 1 b T ) / ( 2 V ) L=(b\mathbf{1}^{\mathsf T}+\mathbf{1}b^{\mathsf T})/(2V) L = ( b 1 T + 1 b T ) / ( 2 V ) . Then tr ( L A ) = t x − t y = : Δ \operatorname{tr}(LA)=t_x-t_y=:\Delta tr ( L A ) = t x − t y =: Δ , and, with U b = b T A b = ∣ S x − S y ∣ 2 / N U_b=b^{\mathsf T}Ab=|S_x-S_y|^2/N U b = b T A b = ∣ S x − S y ∣ 2 / N ,
tr ( A L A L ) = U b R + Δ 2 2 ≤ 4. \operatorname{tr}(ALAL)=\frac{U_bR+\Delta^2}{2}\le4. tr ( A L A L ) = 2 U b R + Δ 2 ≤ 4.
Indeed R ≤ 1 R\le1 R ≤ 1 , U b ≤ 4 U_b\le4 U b ≤ 4 , and Δ 2 ≤ U b R \Delta^2\le U_bR Δ 2 ≤ U b R . The log-concave linear-variance bound in [26 arXiv v2, Theorem 1.1] now applies directly: its inverse-Hessian quantity is at most 4 / α 4/\alpha 4/ α almost surely, so the event in that theorem has probability one. This yields Var Δ ≤ C / α \operatorname{Var}\Delta\le C/\alpha Var Δ ≤ C / α . A zero restricted gradient instead gives a constant observable and zero variance. The theorem allows a convex potential equal to + ∞ +\infty + ∞ outside its open domain, as here. For x ≠ y x\ne y x = y , the functional A x y A_{xy} A x y has L = ( δ x δ y T + δ y δ x T ) / 2 L=(\delta_x\delta_y^{\mathsf T}+\delta_y\delta_x^{\mathsf T})/2 L = ( δ x δ y T + δ y δ x T ) /2 . Then tr ( A L A L ) = ( 1 + A x y 2 ) / 2 ≤ 1 \operatorname{tr}(ALAL)=(1+A_{xy}^2)/2\le1 tr ( A L A L ) = ( 1 + A x y 2 ) /2 ≤ 1 , so the same inequality gives the claimed entry variance. Diagonal entries are constant. □
Lemma 7.7. For M ≥ M 0 M\ge M_0 M ≥ M 0 and N ≥ 4 N\ge4 N ≥ 4 , the actual Gibbs law satisfies, uniformly in the ordered compact ranges,
∣ δ − G ∣ ≤ C κ ( N − 1 + V − 1 ) . |\delta-\mathcal{G}|\le C_\kappa(N^{-1}+V^{-1}). ∣ δ − G ∣ ≤ C κ ( N − 1 + V − 1 ) .
Proof. Let U x = ∣ S x − S ‾ ∣ 2 / N U_x=|S_x-\overline{S}|^2/N U x = ∣ S x − S ∣ 2 / N . Gaussian domination for the zero-sum source δ x − V − 1 1 \delta_x-V^{-1}\mathbf{1} δ x − V − 1 1 and Gaussian linearization give
E e z N U x ≤ ( 1 − 2 z K 0 ) − N / 2 , 0 < 2 z K 0 < 1. \mathbb{E}e^{zNU_x}\le(1-2zK_0)^{-N/2},\qquad0<2zK_0<1. E e z N U x ≤ ( 1 − 2 z K 0 ) − N /2 , 0 < 2 z K 0 < 1.
Hard length implies U x − K 0 = R + r 0 − 2 t x U_x-K_0=R+r_0-2t_x U x − K 0 = R + r 0 − 2 t x . Hence t x < a c u t t_x<a_{\mathrm{cut}} t x < a cut implies U x − K 0 > r 0 / 2 U_x-K_0>r_0/2 U x − K 0 > r 0 /2 . By (42), K 0 ≤ 1 K_0\leq1 K 0 ≤ 1 and r 0 ≥ r ∗ r_0\geq r_* r 0 ≥ r ∗ , so a sufficiently small fixed z > 0 z>0 z > 0 in the preceding inequality yields
P ( t x < a c u t ) ≤ C κ e − c κ N . (54) \mathbb{P}(t_x<a_{\mathrm{cut}})\leq C_\kappa e^{-c_\kappa N}. \tag*{(54)} P ( t x < a cut ) ≤ C κ e − c κ N . (54)
This one-site bound is uniform in the volume. Since V − 1 ∑ { x , y } J x y = j V^{-1}\sum_{\{x,y\}}J_{xy}=j V − 1 ∑ { x , y } J x y = j , it also gives
1 V ∑ { x , y } J x y P ( t x < a c u t or t y < a c u t ) ≤ 2 j sup z P ( t z < a c u t ) ≤ C κ e − c κ N . \frac{1}{V}\sum_{\{x,y\}}J_{xy}\mathbb{P}(t_x<a_{\mathrm{cut}}\text{ or }t_y<a_{\mathrm{cut}})\leq2j\sup_z\mathbb{P}(t_z<a_{\mathrm{cut}})\leq C_\kappa e^{-c_\kappa N}. V 1 { x , y } ∑ J x y P ( t x < a cut or t y < a cut ) ≤ 2 j z sup P ( t z < a cut ) ≤ C κ e − c κ N .
Choose a smooth function φ \varphi φ on [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] equal to 1 / t 1/t 1/ t for t ≥ a c u t t\geq a_{\mathrm{cut}} t ≥ a cut and zero for t ≤ a c u t / 2 t\leq a_{\mathrm{cut}}/2 t ≤ a cut /2 . Its value and derivative have bounds depending only on r ∗ r_* r ∗ . At site x x x use the tangent field φ ( t x ) P x S \varphi(t_x)P_xS φ ( t x ) P x S , where P x = I − S x S x T / N P_x=I-S_xS_x^{\mathsf T}/N P x = I − S x S x T / N . Its divergence is
( N − 1 ) φ ( t x ) ( V − 1 − t x ) + φ ′ ( t x ) ( R − t x 2 ) . (N-1)\varphi(t_x)(V^{-1}-t_x)+\varphi'(t_x)(R-t_x^2). ( N − 1 ) φ ( t x ) ( V − 1 − t x ) + φ ′ ( t x ) ( R − t x 2 ) .
The logarithmic Gibbs derivative along the field is β N φ ( t x ) { t x e x − ( L J , M t ) x } \beta N\varphi(t_x)\{t_xe_x-(L_J,Mt)_x\} β N φ ( t x ) { t x e x − ( L J , M t ) x } . Sphere integration by parts, followed by a volume average, gives
0 = ( N − 1 ) E [ φ ( t x ) ( V − 1 − t x ) ] + E [ φ ′ ( t x ) ( R − t x 2 ) ] + β N E [ φ ( t x ) t x e x ] − β N V ∑ { x , y } J x y E [ ( φ ( t x ) − φ ( t y ) ) ( t x − t y ) ] . 0=(N-1)\mathbb{E}[\varphi(t_x)(V^{-1}-t_x)]+\mathbb{E}[\varphi'(t_x)(R-t_x^2)]+\beta N\mathbb{E}[\varphi(t_x)t_xe_x]
-\frac{\beta N}{V}\sum_{\{x,y\}}J_{xy}\mathbb{E}[(\varphi(t_x)-\varphi(t_y))(t_x-t_y)]. 0 = ( N − 1 ) E [ φ ( t x ) ( V − 1 − t x )] + E [ φ ′ ( t x ) ( R − t x 2 )] + β N E [ φ ( t x ) t x e x ] − V β N { x , y } ∑ J x y E [( φ ( t x ) − φ ( t y )) ( t x − t y )] .
On good sites φ = 1 / t \varphi=1/t φ = 1/ t ; on good edges the product in the last line is − ( t x − t y ) 2 / ( t x t y ) -(t_x-t_y)^2/(t_xt_y) − ( t x − t y ) 2 / ( t x t y ) . All bad-site and bad-edge replacements cost O κ ( N e − c κ N ) O_\kappa(Ne^{-c_\kappa N}) O κ ( N e − c κ N ) , by (54) , ∣ t x ∣ ≤ 1 |t_x|\leq1 ∣ t x ∣ ≤ 1 , R ≤ 1 R\leq1 R ≤ 1 , 0 ≤ e x ≤ 4 j 0\leq e_x\leq4j 0 ≤ e x ≤ 4 j , and bounded φ , φ ′ \varphi,\varphi' φ , φ ′ . Each volume-averaged site has only 2 d 2d 2 d neighbours. Substituting E e x = ( 1 − V − 1 ) / β − δ \mathbb{E}e_x=(1-V^{-1})/\beta-\delta E e x = ( 1 − V − 1 ) / β − δ and dividing by N N N yields
β δ = 2 N − 1 V + N − 1 N V E [ 1 { t x ≥ a c u t } t x ] − 1 N E [ R 1 { t x ≥ a c u t } t x 2 ] + β G + O κ ( e − c κ N ) . \beta\delta=\frac{2}{N}-\frac{1}{V}+\frac{N-1}{NV}\mathbb{E}\left[\frac{\mathbf{1}_{\{t_x\geq a_{\mathrm{cut}}\}}}{t_x}\right]-\frac{1}{N}\mathbb{E}\left[\frac{R\mathbf{1}_{\{t_x\geq a_{\mathrm{cut}}\}}}{t_x^2}\right]+\beta\mathcal{G}+O_\kappa(e^{-c_\kappa N}). β δ = N 2 − V 1 + N V N − 1 E [ t x 1 { t x ≥ a cut } ] − N 1 E [ t x 2 R 1 { t x ≥ a cut } ] + β G + O κ ( e − c κ N ) .
The two expectations are bounded by a c u t − 1 a_{\mathrm{cut}}^{-1} a cut − 1 and a c u t − 2 a_{\mathrm{cut}}^{-2} a cut − 2 , and β \beta β is bounded away from zero. This proves the lemma. □
Proof of Proposition 7.5. Write q = N − V − 1 > 0 q=N-V-1>0 q = N − V − 1 > 0 . Translation invariance gives E ( t x − t y ) = 0 \mathbb{E}(t_x-t_y)=0 E ( t x − t y ) = 0 . Lemma 7.6 , t x t y ≥ a c u t 2 t_xt_y\geq a_{\mathrm{cut}}^2 t x t y ≥ a cut 2 on good edges, and V − 1 ∑ { x , y } J x y = j V^{-1}\sum_{\{x,y\}}J_{xy}=j V − 1 ∑ { x , y } J x y = j imply G ≤ C κ / q \mathcal{G}\leq C_\kappa/q G ≤ C κ / q . Lemma 7.7 and N − 1 ≤ q − 1 N^{-1}\leq q^{-1} N − 1 ≤ q − 1 give
δ M , N ≤ C κ ( q − 1 + V − 1 ) . \delta_{M,N}\leq C_\kappa(q^{-1}+V^{-1}). δ M , N ≤ C κ ( q − 1 + V − 1 ) .
Now restrict the full Gram matrix to U \mathcal{U} U , retaining the notation A A A , and put A ‾ = E A \overline{A}=\mathbb{E}A A = E A , Σ = Σ β , J ∣ U \Sigma=\Sigma_{\beta,J}|_{\mathcal{U}} Σ = Σ β , J ∣ U . The fixed-path deficit estimate and dipole quadrature, as in (52) , give
∥ A ‾ − Σ ∥ F ≤ C U , κ ( q − 1 + V − 1 ) . \|\overline{A}-\Sigma\|_F\leq C_{\mathcal{U},\kappa}(q^{-1}+V^{-1}). ∥ A − Σ ∥ F ≤ C U , κ ( q − 1 + V − 1 ) .
Lemma 7.6 bounds the variance of each entry, so
E ∥ A − A ‾ ∥ F 2 ≤ C U / q . \mathbb{E}\|A-\overline{A}\|_F^2\leq C_{\mathcal{U}}/q. E ∥ A − A ∥ F 2 ≤ C U / q .
By Lemma 5.2 , Σ ⪰ m I \Sigma\succeq mI Σ ⪰ m I uniformly for a fixed m > 0 m>0 m > 0 . Choose M 0 M_0 M 0 and a fixed q 0 q_0 q 0 large enough that, if q ≥ q 0 q\geq q_0 q ≥ q 0 , the bias bound implies A ‾ ⪰ m I / 2 \overline{A}\succeq mI/2 A ⪰ m I /2 . Lemma 7.3 then gives a mixture error C p / q C p/q C p / q and a covariance bias C p ( q − 1 + V − 1 ) C\sqrt{p}(q^{-1}+V^{-1}) C p ( q − 1 + V − 1 ) . The Haar orbit kernel in Proposition 5.3 and (53) give an additional error C p / N Cp/N C p / N . Since p ≥ 1 p \ge1 p ≥ 1 and N − 1 ≤ q − 1 N^{-1} \le q^{-1} N − 1 ≤ q − 1 , their sum is the asserted bound. If 0 < q < q 0 0 < q < q_0 0 < q < q 0 , both local laws have second moment ℓ p \ell_p ℓ p ; independent coupling gives W 1 ≤ 2 ℓ p W_1 \le2\sqrt{\ell_p} W 1 ≤ 2 ℓ p . This is bounded by C p / q Cp/q C p / q after enlarging C C C by a constant independent of p p p . Taking M 0 M_0 M 0 larger also ensures V ≥ ℓ V \ge\ell V ≥ ℓ and N ≥ 4 N \ge4 N ≥ 4 .
If N ≥ 2 ( V + 1 ) N \ge2(V+1) N ≥ 2 ( V + 1 ) , then q ≥ N / 2 q \ge N/2 q ≥ N /2 and the upper bound is C ( p / V + p / N ) C(\sqrt{p}/V + p/N) C ( p / V + p / N ) . On a star-containing window, the first lower bound in Proposition 6.2, namely c p / V c\sqrt{p}/V c p / V , holds throughout this regime. The additional condition V p ≤ N V\sqrt{p} \le N V p ≤ N therefore yields a matching two-sided p / V \sqrt{p}/V p / V order. For fixed p p p and N ≥ 2 ( V + 1 ) N \ge2(V+1) N ≥ 2 ( V + 1 ) , p / N ≤ p / ( 2 V ) p/N \le p/(2V) p / N ≤ p / ( 2 V ) , which gives the two-sided V − 1 V^{-1} V − 1 order on star-containing windows with constants depending on p p p .
For V > N V > N V > N , the full Gram law is supported on the nonconvex rank-at-most-N N N set. At N = V N = V N = V and N = V + 1 N = V + 1 N = V + 1 its density exists but the determinant exponent is nonpositive. The positive-Hessian argument applies to none of these ranges, and the restricted estimate alone gives no improved volume-first or accumulation-state rate.
Proof of the ordered estimate
Proof of Theorem 2.8. Proposition 7.4 gives (51) . A direct substitution of (49) would introduce p / V p/V p / V . To retain p / V \sqrt{p}/V p / V , split into two ranges.
If N < 2 ( V + 1 ) N < 2(V+1) N < 2 ( V + 1 ) , then N ≥ 4 N \ge4 N ≥ 4 implies V > N / 4 V > N/4 V > N /4 . The volume part of the deficit in (49) can therefore be absorbed into C / N C/N C / N , and (51) gives C U , κ { p ψ d ( N ) + p / V } C_{U,\kappa}\{p\psi_d(N) + \sqrt{p}/V\} C U , κ { p ψ d ( N ) + p / V } . If N ≥ 2 ( V + 1 ) N \ge2(V+1) N ≥ 2 ( V + 1 ) , Proposition 7.5 gives the sharper bound C U , κ ( p / N + p / V ) C_{U,\kappa}(p/N + \sqrt{p}/V) C U , κ ( p / N + p / V ) , since N − V − 1 ≥ N / 2 N - V - 1 \ge N/2 N − V − 1 ≥ N /2 . These ranges cover all N , V N,V N , V . Taking the smaller of this bound and Theorem 2.1, using s M ≤ C κ / V s_M \le C_\kappa/V s M ≤ C κ / V (Proposition 2.2), proves the asserted upper bound. At each fixed N , p N,p N , p , local second moments equal ℓ p \ell_p ℓ p ; uniform integrability passes both estimates to every periodic accumulation state. For d ≥ 5 d \ge5 d ≥ 5 , p / N ≤ p / N p/N \le\sqrt{p}/N p / N ≤ p / N , which gives the simplified rate.
For the lower bounds, put χ ˉ q = E χ q \bar{\chi}_q = \mathbb{E}\chi_q χ ˉ q = E χ q , where χ q \chi_q χ q is the norm of a standard q q q -dimensional Gaussian vector. At any one site the spin is uniform on the sphere of radius N \sqrt{N} N , by rotation invariance. The 1-Lipschitz test given by the norm of its first p p p coordinates therefore gives, on every nonempty window,
W 1 ≥ χ ˉ p ( N χ ˉ N − 1 ) . W_1 \ge\bar{\chi}_p\left(\frac{\sqrt{N}}{\bar{\chi}_N}-1\right). W 1 ≥ χ ˉ p ( χ ˉ N N − 1 ) .
Here χ ˉ N = 2 Γ ( ( N + 1 ) / 2 ) / Γ ( N / 2 ) \bar{\chi}_N = \sqrt{2}\,\Gamma((N+1)/2)/\Gamma(N/2) χ ˉ N = 2 Γ (( N + 1 ) /2 ) /Γ ( N /2 ) , and the gamma-ratio expansion gives N / χ ˉ N − 1 = 1 / ( 4 N ) + O ( N − 2 ) \sqrt{N}/\bar{\chi}_N - 1 = 1/(4N) + O(N^{-2}) N / χ ˉ N − 1 = 1/ ( 4 N ) + O ( N − 2 ) . For each fixed p p p this yields a positive lower bound of order N − 1 N^{-1} N − 1 . Strict Jensen gives χ ˉ N < N \bar{\chi}_N < \sqrt{N} χ ˉ N < N for every finite N N N , so the finitely many smaller admissible N N N can be included in a lower bound c p / N c_p/N c p / N . On star-containing windows, Theorem 2.3 also gives c p M − d c_pM^{-d} c p M − d ; the maximum of these two lower bounds controls N − 1 + M − d N^{-1} + M^{-d} N − 1 + M − d . Every periodic accumulation state has the same one-site sphere marginal. The same norm test thus proves its stated N − 1 N^{-1} N − 1 lower bound, while the upper bound has already passed to the limit.
Discussion
The local law estimate bounds the error in Gaussian predictions. For F F F Lipschitz on the ℓ × p \ell\times p ℓ × p coordinate block in the Frobenius norm, Theorem 2.1 gives
∣ E ν M , N , β , J u , p F − E γ β , J u , p F ∣ ≤ C Lip ( F ) p ( N − 1 / 2 + s M ) . \left|\mathbb{E}_{\nu_{M,N,\beta,J}^{u,p}} F-\mathbb{E}_{\gamma_{\beta,J}^{u,p}}F\right| \le C\operatorname{Lip}(F)\sqrt{p}\left(N^{-1/2}+s_M\right). E ν M , N , β , J u , p F − E γ β , J u , p F ≤ C Lip ( F ) p ( N − 1/2 + s M ) .
For example, take distinct sites x , y ∈ U x,y\in\mathcal{U} x , y ∈ U and the two observables F ± ( X ) = cos ( t ( X x , 1 ± X y , 1 ) ) F_{\pm}(X)=\cos(t(X_{x,1}\pm X_{y,1})) F ± ( X ) = cos ( t ( X x , 1 ± X y , 1 )) with t ∈ R t\in\mathbb{R} t ∈ R . Both have Lipschitz constant at most 2 ∣ t ∣ \sqrt{2}\,|t| 2 ∣ t ∣ , so with p = 1 p=1 p = 1 each prediction has error at most C 2 ∣ t ∣ ( N − 1 / 2 + s M ) C\sqrt{2}\,|t|(N^{-1/2}+s_M) C 2 ∣ t ∣ ( N − 1/2 + s M ) . Write q = 1 − β c , J / β q=1-\beta_{c,J}/\beta q = 1 − β c , J / β for the variance of the Gaussian constant mode. Under the Gaussian law, the expectations of F ± F_{\pm} F ± are exp ( − t 2 v ± / 2 ) \exp(-t^2v_{\pm}/2) exp ( − t 2 v ± /2 ) with
v − = 2 β ( β c , J − G J ( x , y ) ) , v + = 2 β ( β c , J + G J ( x , y ) ) + 4 q . v_-=\frac{2}{\beta}\bigl(\beta_{c,J}-G_J(x,y)\bigr),\qquad v_+=\frac{2}{\beta}\bigl(\beta_{c,J}+G_J(x,y)\bigr)+4q. v − = β 2 ( β c , J − G J ( x , y ) ) , v + = β 2 ( β c , J + G J ( x , y ) ) + 4 q .
The constant mode cancels in the spin difference. In the spin sum it multiplies the GFF prediction by exp ( − 2 t 2 q ) \exp(-2t^2q) exp ( − 2 t 2 q ) . The same observable error bound in periodic accumulation states has no s M s_M s M term, by Corollary 2.4.
A fixed number of coordinates of a uniformly oriented vector of length N q \sqrt{Nq} N q converge as N → ∞ N\to\infty N → ∞ to independent Gaussians of variance q q q [14 ] . In the mean spherical reference the constant mode is Gaussian with variance 1 / ( β V s M ) 1/(\beta V s_M) 1/ ( β V s M ) , which tends to q q q when β > β c , J \beta>\beta_{c,J} β > β c , J is fixed (Proposition A.4 ). The same Gaussian zero-mode mechanism occurs in the continuum mean spherical model with a cutoff [11 Appendix B] . In the thermodynamic limit of the isotropic periodic Berlin–Kac model below its critical temperature, the independent constant takes the values ± q \pm\sqrt{q} ± q with equal probability [1 Theorem 1.1] . It produces the factor cos ( 2 t q ) \cos(2t\sqrt{q}) cos ( 2 t q ) in the spin sum. For q > 0 q>0 q > 0 and sufficiently small nonzero t t t , the sum distinguishes the two limiting laws despite their equal zero-mode variance. The difference between zero-field spherical and mean spherical averages below criticality was already noted in [21 ] .
For fixed U , p \mathcal{U},p U , p , every fixed-order local mixed moment also converges to its Gaussian counterpart. Global O ( N ) O(N) O ( N ) invariance gives
E ∣ S x α ∣ 2 r = N r ( 2 r − 1 ) ! ! ∏ j = 0 r − 1 ( N + 2 j ) ≤ ( 2 r − 1 ) ! ! , r ∈ N . \mathbb{E}|S_x^\alpha|^{2r}=\frac{N^r(2r-1)!!}{\prod_{j=0}^{r-1}(N+2j)}\le(2r-1)!!,\qquad r\in\mathbb{N}. E ∣ S x α ∣ 2 r = ∏ j = 0 r − 1 ( N + 2 j ) N r ( 2 r − 1 )!! ≤ ( 2 r − 1 )!! , r ∈ N .
This identity also holds in every periodic accumulation state. Hölder’s inequality, applied with higher-order moments, gives uniform integrability of each fixed-degree monomial on the coordinate block. The Gaussian coordinates have the same moment bounds. Truncation and Corollary 2.5 therefore give, for every fixed polynomial P P P on R ℓ × p \mathbb{R}^{\ell\times p} R ℓ × p ,
sup ( β , J ) satisfying ( 3 ) ∣ E M , N , β , J P − E γ β , J P ∣ ⟶ 0 as min ( M , N ) → ∞ . \sup_{\substack{(\beta,J)\\\text{satisfying }(3)}}\left|\mathbb{E}_{M,N,\beta,J}P-\mathbb{E}_{\gamma_{\beta,J}}P\right|\longrightarrow0\qquad\text{as }\min(M,N)\to\infty. ( β , J ) satisfying ( 3 ) sup E M , N , β , J P − E γ β , J P ⟶ 0 as min ( M , N ) → ∞.
The corresponding convergence for periodic accumulation states holds as N → ∞ N\to\infty N → ∞ , uniformly over the same parameter range and all choices of states. The Gaussian expectations are given by Wick’s rule at the same ( β , J ) (\beta,J) ( β , J ) .
For a fixed number of coordinates, the general error bound separates the choice of box size from the spin-dimensional approximation. At criticality in d = 3 d=3 d = 3 , s M ≍ M − 2 s_M\asymp M^{-2} s M ≍ M − 2 (Proposition A.4 ), so M / N 1 / 4 → ∞ M/N^{1/4}\to\infty M / N 1/4 → ∞ makes the finite-volume term smaller than the N − 1 / 2 N^{-1/2} N − 1/2 term. For d ≥ 5 d\ge5 d ≥ 5 at a fixed positive ordered gap, Theorem 2.8 gives C U , p , κ Lip ( F ) ( N − 1 + M − d ) C_{\mathcal{U},p,\kappa}\operatorname{Lip}(F)(N^{-1}+M^{-d}) C U , p , κ Lip ( F ) ( N − 1 + M − d ) for fixed p p p . A volume M d ≥ N M^d\ge N M d ≥ N keeps the finite-volume term no larger than the spin-dimensional term. The matching lower bound on star-containing windows is for the W 1 W_1 W 1 distance, which takes the supremum over unit Lipschitz observables. The coordinate-growth criterion is exact, while the matching rate concerns fixed p p p . The optimal rate as a function of growing p p p remains open.
Infrared bound and order. The full-lattice extension of (2) saturates the Fröhlich–Simon–Spencer infrared bound [15 ] , which Proposition 4.1 gives in the present normalization. Its spectral measure has density β − 1 λ J ( k ) − 1 \beta^{-1}\lambda_J(k)^{-1} β − 1 λ J ( k ) − 1 at nonzero frequencies and an atom of weight 1 − β c , J / β 1-\beta_{c,J}/\beta 1 − β c , J / β at zero. The unit diagonal is the sum rule. The massless part carries weight β c , J / β \beta_{c,J}/\beta β c , J / β and leaves 1 − β c , J / β 1-\beta_{c,J}/\beta 1 − β c , J / β for the zero mode. For β < β c , J \beta<\beta_{c,J} β < β c , J , the massless part alone would exceed the sum rule. The target (2) is therefore restricted to β ≥ β c , J \beta\ge\beta_{c,J} β ≥ β c , J .
For β > β c , J \beta>\beta_{c,J} β > β c , J the same bound forces long-range order at every fixed N N N [15 ] . By § 7.2 , E ∣ S ‾ ∣ 2 / N ≥ 1 − β c , J / β − O ( M 2 − d ) \mathbb{E}|\overline{S}|^2/N\ge1-\beta_{c,J}/\beta-O(M^{2-d}) E ∣ S ∣ 2 / N ≥ 1 − β c , J / β − O ( M 2 − d ) for the mean spin S ‾ = V − 1 ∑ x S x \overline{S}=V^{-1}\sum_x S_x S = V − 1 ∑ x S x . This is a bound on the atom at zero of the spectral measure of the normalized spin covariance E ( S 0 ⋅ S x ) / N \mathbb{E}(S_0\cdot S_x)/N E ( S 0 ⋅ S x ) / N . Along any periodic accumulation sequence at fixed N N N , these covariances converge, so the spectral measures converge weakly. The closed-set bound for { 0 } \{0\} { 0 } then gives an atom of weight at least 1 − β c , J / β 1-\beta_{c,J}/\beta 1 − β c , J / β in the limit.
The two-dimensional continuum linear sigma model has a soft constraint and a massive Gaussian limit. Delgadino and Smith obtain an N − 1 / 2 N^{-1/2} N − 1/2 estimate for periodic accumulation states in a weighted Wasserstein distance with cost averaged over components [12 Theorem 3.9] . Korzhenkova and Sepúlveda study angular freezing and entropic repulsion in a two-dimensional vector GFF conditioned by a pointwise lower bound on its norm [23 ] . Inverse-size rates for fixed marginals also occur for continuous mean-field Gibbs measures [25 ] . For lattice spins at fixed N N N , low-temperature results include correlation expansions in states selected by a field [2 , 17 ] , a transverse GFF limit on finite graphs with a pinned spin [18 Theorem 2.5] , and estimates uniform over boundary conditions [16 Theorem 1.8 and Proposition 1.10] .
A natural open problem is whether W 1 ( ν M , N , β , J U 0 , p , γ M U 0 , p ) ≤ C p / N W_1(\nu_{M,N,\beta,J}^{U_0,p},\gamma_M^{U_0,p})\le C\sqrt{p/N} W 1 ( ν M , N , β , J U 0 , p , γ M U 0 , p ) ≤ C p / N uniformly in M M M , that is, whether the volume term disappears when the spin law is compared with the mean spherical reference γ M \gamma_M γ M of Remark 2.7 instead of the target. For N ≥ 4 N\ge4 N ≥ 4 , the J J J -weighted sums of nearest-neighbour difference variances agree to within C N − 1 / 2 CN^{-1/2} C N − 1/2 . Let v i M v_i^M v i M be the variance of one coordinate of X 0 − X e i X_0-X_{e_i} X 0 − X e i for X ∼ γ M U 0 , p X\sim\gamma_M^{U_0,p} X ∼ γ M U 0 , p . Then ∑ i J i { Q M 0 ( δ 0 − δ e i ) / β − v i M } = s M − 1 / ( β V ) \sum_i J_i\{Q_M^0(\delta_0-\delta_{e_i})/\beta-v_i^M\}=s_M-1/(\beta V) ∑ i J i { Q M 0 ( δ 0 − δ e i ) / β − v i M } = s M − 1/ ( β V ) , with Q M 0 Q_M^0 Q M 0 as defined in § 2 (see Appendix § 9 .§ 3 ), while Lemma 4.3 and the left secant in the proof of Proposition 6.2 give ∣ δ M , N − ( s M − 1 / ( β V ) ) ∣ ≤ C N − 1 / 2 |\delta_{M,N}-(s_M-1/(\beta V))|\le CN^{-1/2} ∣ δ M , N − ( s M − 1/ ( β V )) ∣ ≤ C N − 1/2 for N ≥ 4 N\ge4 N ≥ 4 . Gaussian domination, however, bounds the spin dipole variances by the massless form, and no sign is available relative to Σ M \Sigma_M Σ M . For β < β c , J \beta<\beta_{c,J} β < β c , J , Proposition 4.2 and both pressure secants remain valid. For each s > 0 s>0 s > 0 , Riemann sums give g M ( s ) → g ∞ ( s ) = ∫ T d ( λ J ( k ) + s ) − 1 d μ ( k ) g_M(s)\to g_\infty(s)=\int_{\mathbb{T}^d}(\lambda_J(k)+s)^{-1}\,\mathrm{d}\mu(k) g M ( s ) → g ∞ ( s ) = ∫ T d ( λ J ( k ) + s ) − 1 d μ ( k ) , using the notation of Appendix § 9 . The function g ∞ g_\infty g ∞ decreases continuously from β c , J \beta_{c,J} β c , J to zero. Evaluating at positive shifts on either side of the unique solution s ∞ s_\infty s ∞ of g ∞ ( s ∞ ) = β g_\infty(s_\infty)=\beta g ∞ ( s ∞ ) = β brackets s M s_M s M by monotonicity, so s M → s ∞ > 0 s_M\to s_\infty>0 s M → s ∞ > 0 . The energy deficit of Lemma 4.3 then satisfies δ M , N ≥ s M − 1 / ( β V ) − C N − 1 / 2 \delta_{M,N}\ge s_M-1/(\beta V)-CN^{-1/2} δ M , N ≥ s M − 1/ ( β V ) − C N − 1/2 ; for fixed J J J and 0 < β < β c , J 0<\beta<\beta_{c,J} 0 < β < β c , J it is therefore bounded away from zero as M , N → ∞ M,N\to\infty M , N → ∞ , so the saturation argument gives no convergence; moreover, the target γ β , J \gamma_{\beta,J} γ β , J is not defined there. The known results in that regime are recalled in § 1 . Finally, the bounds may be improvable at criticality and in ordered dimensions three and four; growing windows and states selected by a field or by boundary conditions are outside the present estimates.
Spectral estimates
The quadrature and mass bounds below give the volume term in Theorem 2.1 and Proposition 2.2. The final subsection compares the auxiliary Gaussian reference with the target, as stated in Remark 2.7.
Fix d ≥ 3 d\ge3 d ≥ 3 and 0 < j 0 ≤ j 1 < ∞ 0<j_0\le j_1<\infty 0 < j 0 ≤ j 1 < ∞ . Throughout this section, J ∈ [ j 0 , j 1 ] d J\in[j_0,j_1]^d J ∈ [ j 0 , j 1 ] d , j = ∑ i J i j=\sum_iJ_i j = ∑ i J i , M ≥ 4 M\ge4 M ≥ 4 is an integer, and V = M d V=M^d V = M d . Frequencies belong to T d = R d / ( 2 π Z ) d \mathbb{T}^d=\mathbb{R}^d/(2\pi\mathbb{Z})^d T d = R d / ( 2 π Z ) d , with normalized measure d μ ( k ) = d k / ( 2 π ) d \mathrm{d}\mu(k)=\mathrm{d}k/(2\pi)^d d μ ( k ) = d k / ( 2 π ) d , and
K M = ( 2 π / M ) { 0 , … , M − 1 } d ( m o d 2 π ) , λ J ( k ) = 2 ∑ i = 1 d J i ( 1 − cos k i ) . \mathcal{K}_M=(2\pi/M)\{0,\ldots,M-1\}^d\pmod{2\pi},\qquad\lambda_J(k)=2\sum_{i=1}^dJ_i(1-\cos k_i). K M = ( 2 π / M ) { 0 , … , M − 1 } d ( mod 2 π ) , λ J ( k ) = 2 i = 1 ∑ d J i ( 1 − cos k i ) .
Write ρ ( k ) \rho(k) ρ ( k ) for the distance from k k k to zero on this torus. Constants in the spectral estimates depend only on d d d , j 0 j_0 j 0 , j 1 j_1 j 1 and, where present, a fixed source or fixed local set. In mass and Gaussian comparisons we also fix 0 < b ≤ B < ∞ 0 < b \le B < \infty 0 < b ≤ B < ∞ and restrict to β ∈ [ b , B ] \beta\in[b,B] β ∈ [ b , B ] , β ≥ β c , J \beta\ge\beta_{c,J} β ≥ β c , J . All stated uniform limits have these compact parameter ranges.
For a finitely supported real source a a a with ∑ x a x = 0 \sum_x a_x=0 ∑ x a x = 0 , set
a ^ ( k ) = ∑ x a x e i k ⋅ x . \widehat{a}(k)=\sum_x a_x e^{ik\cdot x}. a ( k ) = x ∑ a x e ik ⋅ x .
Q M 0 ( a ) = 1 V ∑ k ≠ 0 ∣ a ^ ( k ) ∣ 2 λ J ( k ) , Q ∞ ( a ) = ∫ T d ∣ a ^ ( k ) ∣ 2 λ J ( k ) d μ ( k ) . Q_M^0(a)=\frac{1}{V}\sum_{k\ne0}\frac{|\widehat{a}(k)|^2}{\lambda_J(k)},\qquad Q_\infty(a)=\int_{\mathbb{T}^d}\frac{|\widehat{a}(k)|^2}{\lambda_J(k)}\,d\mu(k). Q M 0 ( a ) = V 1 k = 0 ∑ λ J ( k ) ∣ a ( k ) ∣ 2 , Q ∞ ( a ) = ∫ T d λ J ( k ) ∣ a ( k ) ∣ 2 d μ ( k ) .
where grid sums always range over K M \mathcal{K}_M K M . Also define
β c , J = ∫ λ J ( k ) − 1 d μ ( k ) , \beta_{c,J}=\int\lambda_J(k)^{-1}\,d\mu(k), β c , J = ∫ λ J ( k ) − 1 d μ ( k ) ,
g M ( s ) = 1 V ∑ k 1 λ J ( k ) + s , g M 0 ( s ) = 1 V ∑ k ≠ 0 1 λ J ( k ) + s . g_M(s)=\frac{1}{V}\sum_k\frac{1}{\lambda_J(k)+s},\qquad g_M^0(s)=\frac{1}{V}\sum_{k\ne0}\frac{1}{\lambda_J(k)+s}. g M ( s ) = V 1 k ∑ λ J ( k ) + s 1 , g M 0 ( s ) = V 1 k = 0 ∑ λ J ( k ) + s 1 .
The additive mass s M > 0 s_M>0 s M > 0 is the unique solution of g M ( s M ) = β g_M(s_M)=\beta g M ( s M ) = β : g M g_M g M is strictly decreasing, tends to infinity at zero because of the zero mode, and tends to zero at infinity. Thus G J , M , s = ( L J , M + s I ) − 1 G_{J,M,s}=(L_{J,M}+sI)^{-1} G J , M , s = ( L J , M + s I ) − 1 ; throughout, s s s denotes the additive parameter itself.
Smooth Fourier quadrature and the zero-sum gain
Lemma A.1 (Fourier aliasing). Let L > d L>d L > d be an integer and F ∈ C L ( T d ) F\in C^L(\mathbb{T}^d) F ∈ C L ( T d ) . With normalized L 1 L^1 L 1 norms,
1 V ∑ k ∈ K M F ( k ) − ∫ F d μ = ∑ q ∈ Z d ∖ { 0 } F ^ ( M q ) , \frac{1}{V}\sum_{k\in\mathcal{K}_M}F(k)-\int F\,d\mu=\sum_{q\in\mathbb{Z}^d\setminus\{0\}}\widehat{F}(Mq), V 1 k ∈ K M ∑ F ( k ) − ∫ F d μ = q ∈ Z d ∖ { 0 } ∑ F ( M q ) , ∣ 1 V ∑ k F ( k ) − ∫ F d μ ∣ ≤ C d , L M − L max i ∥ ∂ i L F ∥ 1 . (55) \left|\frac{1}{V}\sum_k F(k)-\int F\,d\mu\right|\le C_{d,L}M^{-L}\max_i\|\partial_i^L F\|_1. \tag*{(55)} V 1 k ∑ F ( k ) − ∫ F d μ ≤ C d , L M − L i max ∥ ∂ i L F ∥ 1 . (55) where F ^ ( n ) = ∫ F ( k ) e − i n ⋅ k d μ ( k ) \widehat{F}(n)=\int F(k)e^{-in\cdot k}\,d\mu(k) F ( n ) = ∫ F ( k ) e − in ⋅ k d μ ( k ) . The series on the right is absolutely convergent.
Proof. For n ≠ 0 n\ne0 n = 0 , select a coordinate i i i with ∣ n i ∣ = ∣ n ∣ ∞ |n_i|=|n|_\infty ∣ n i ∣ = ∣ n ∣ ∞ . Integrating by parts L L L times in that coordinate, with all boundary terms cancelling by periodicity, gives
∣ F ^ ( n ) ∣ ≤ ∣ n ∣ ∞ − L max i ∥ ∂ i L F ∥ 1 . |\widehat{F}(n)|\le|n|_\infty^{-L}\max_i\|\partial_i^L F\|_1. ∣ F ( n ) ∣ ≤ ∣ n ∣ ∞ − L i max ∥ ∂ i L F ∥ 1 . The shell ∣ n ∣ ∞ = m |n|_\infty=m ∣ n ∣ ∞ = m has at most C d m d − 1 C_d m^{d-1} C d m d − 1 points, so ∑ n ≠ 0 ∣ n ∣ ∞ − L < ∞ \sum_{n\ne0}|n|_\infty^{-L}<\infty ∑ n = 0 ∣ n ∣ ∞ − L < ∞ . Hence the Fourier series of F F F converges absolutely and uniformly, and its sum is F F F by uniqueness of Fourier coefficients of continuous functions.
We may now average that series over the finite grid. The geometric sum
1 V ∑ m ∈ { 0 , … , M − 1 } d e 2 π i n ⋅ m / M = 1 { n ∈ M Z d } \frac{1}{V}\sum_{m\in\{0,\ldots,M-1\}^d}e^{2\pi in\cdot m/M}=\mathbf{1}_{\{n\in M\mathbb{Z}^d\}} V 1 m ∈ { 0 , … , M − 1 } d ∑ e 2 π in ⋅ m / M = 1 { n ∈ M Z d } gives the identity in (A.1); the coefficient bound and the convergent shell sum give its estimate. □ \square □
Lemma A.2 (Sharp zero-sum and diagonal quadrature). For each fixed finitely supported real zero-sum source a a a ,
∣ Q M 0 ( a ) − Q ∞ ( a ) ∣ ≤ C a M − d . (56) \lvert Q_M^0(a)-Q_\infty(a)\rvert\le C_a M^{-d}. \tag*{(56)} ∣ Q M 0 ( a ) − Q ∞ ( a )∣ ≤ C a M − d . (56) Moreover, with b M = g M 0 ( 0 ) − β c , J b_M=g_M^0(0)-\beta_{c,J} b M = g M 0 ( 0 ) − β c , J ,
∣ b M ∣ ≤ C M 2 − d . (57) \lvert b_M\rvert\le C M^{2-d}. \tag*{(57)} ∣ b M ∣ ≤ C M 2 − d . (57) Both statements hold for every integer M ≥ 4 M \ge4 M ≥ 4 , uniformly over the positive compact coupling range.
Proof. For principal representatives in [ − π , π ] d [-\pi,\pi]^d [ − π , π ] d ,
4 j 0 π 2 ρ ( k ) 2 ≤ λ J ( k ) ≤ j 1 ρ ( k ) 2 . (58) \frac{4j_0}{\pi^2}\rho(k)^2 \le\lambda_J(k) \le j_1\rho(k)^2. \tag*{(58)} π 2 4 j 0 ρ ( k ) 2 ≤ λ J ( k ) ≤ j 1 ρ ( k ) 2 . (58)
The bounds follow from 1 − cos t ≥ 2 t 2 / π 2 1-\cos t \ge2t^2/\pi^2 1 − cos t ≥ 2 t 2 / π 2 and 1 − cos t ≤ t 2 / 2 1-\cos t \le t^2/2 1 − cos t ≤ t 2 /2 . For any fixed derivative order, the trigonometric formula gives ∣ ∂ α λ J ( k ) ∣ ≤ C α ρ 2 − ∣ α ∣ \lvert\partial^\alpha\lambda_J(k)\rvert\le C_\alpha\rho^{2-\lvert\alpha\rvert} ∣ ∂ α λ J ( k )∣ ≤ C α ρ 2 − ∣ α ∣ when 0 < ρ ≤ 1 0<\rho\le1 0 < ρ ≤ 1 . The cases of order zero and one use the quadratic and linear vanishing, respectively; at higher orders the derivatives are bounded. Differentiating λ J λ J − 1 = 1 \lambda_J\lambda_J^{-1}=1 λ J λ J − 1 = 1 yields
∂ α ( λ J − 1 ) = − λ J − 1 ∑ 0 < η ≤ α ( α η ) ( ∂ η λ J ) ∂ α − η ( λ J − 1 ) . \partial^\alpha(\lambda_J^{-1})=-\lambda_J^{-1}\sum_{0<\eta\le\alpha}\binom{\alpha}{\eta}(\partial^\eta\lambda_J)\partial^{\alpha-\eta}(\lambda_J^{-1}). ∂ α ( λ J − 1 ) = − λ J − 1 0 < η ≤ α ∑ ( η α ) ( ∂ η λ J ) ∂ α − η ( λ J − 1 ) .
Induction and (A.4) therefore give
∣ ∂ α ( λ J − 1 ) ( k ) ∣ ≤ C α ρ − 2 − ∣ α ∣ . (59) \lvert\partial^\alpha(\lambda_J^{-1})(k)\rvert\le C_\alpha\rho^{-2-\lvert\alpha\rvert}. \tag*{(59)} ∣ ∂ α ( λ J − 1 ) ( k )∣ ≤ C α ρ − 2 − ∣ α ∣ . (59)
Away from zero, all these derivatives are uniformly bounded.
Choose a fixed lattice point x ∗ x_* x ∗ and multiply a ^ \widehat a a by e − i k ⋅ x ∗ e^{-ik\cdot x_*} e − ik ⋅ x ∗ , which does not change its absolute square. The zero-sum property gives
e − i k ⋅ x ∗ a ^ ( k ) = ∑ x a x ( e i k ⋅ ( x − x ∗ ) − 1 ) = O a ( ρ ) . e^{-ik\cdot x_*}\widehat a(k)=\sum_x a_x\left(e^{ik\cdot(x-x_*)}-1\right)=O_a(\rho). e − ik ⋅ x ∗ a ( k ) = x ∑ a x ( e ik ⋅ ( x − x ∗ ) − 1 ) = O a ( ρ ) .
All positive-order derivatives of this finite sum are bounded. For n ( k ) = ∣ a ^ ( k ) ∣ 2 n(k)=\lvert\widehat a(k)\rvert^2 n ( k ) = ∣ a ( k ) ∣ 2 , the product rule consequently gives n = O a ( ρ 2 ) n=O_a(\rho^2) n = O a ( ρ 2 ) , ∂ i n = O a ( ρ ) \partial_i n=O_a(\rho) ∂ i n = O a ( ρ ) , and bounded derivatives of every higher fixed order. In particular, ∣ ∂ α n ∣ ≤ C a , α ρ 2 − ∣ α ∣ \lvert\partial^\alpha n\rvert\le C_{a,\alpha}\rho^{2-\lvert\alpha\rvert} ∣ ∂ α n ∣ ≤ C a , α ρ 2 − ∣ α ∣ on 0 < ρ ≤ 1 0<\rho\le1 0 < ρ ≤ 1 . Together with (A.5), this proves
f a ( k ) = ∣ a ^ ( k ) ∣ 2 λ J ( k ) , ∣ ∂ α f a ( k ) ∣ ≤ C a , α ρ − ∣ α ∣ . (60) f_a(k)=\frac{\lvert\widehat a(k)\rvert^2}{\lambda_J(k)},\qquad\lvert\partial^\alpha f_a(k)\rvert\le C_{a,\alpha}\rho^{-\lvert\alpha\rvert}. \tag*{(60)} f a ( k ) = λ J ( k ) ∣ a ( k ) ∣ 2 , ∣ ∂ α f a ( k )∣ ≤ C a , α ρ − ∣ α ∣ . (60)
Set f a ( 0 ) = 0 f_a(0)=0 f a ( 0 ) = 0 to implement removal of the zero grid mode.
Fix χ ∈ C ∞ ( R d ) \chi\in C^\infty(\mathbb{R}^d) χ ∈ C ∞ ( R d ) with 0 ≤ χ ≤ 1 0\le\chi\le1 0 ≤ χ ≤ 1 , χ = 1 \chi=1 χ = 1 on ∣ u ∣ ≤ 1 \lvert u\rvert\le1 ∣ u ∣ ≤ 1 , and χ = 0 \chi=0 χ = 0 on ∣ u ∣ ≥ 2 \lvert u\rvert\ge2 ∣ u ∣ ≥ 2 . The function χ ( M k ) \chi(Mk) χ ( M k ) , defined in the principal cell and then periodically extended, has support strictly inside that cell. Put F M = ( 1 − χ ( M k ) ) f a ( k ) F_M=(1-\chi(Mk))f_a(k) F M = ( 1 − χ ( M k )) f a ( k ) and F M ( 0 ) = 0 F_M(0)=0 F M ( 0 ) = 0 . It is smooth and periodic. Every nonzero grid point has torus distance at least 2 π / M > 2 / M 2\pi/M>2/M 2 π / M > 2/ M , so its grid average is exactly Q M 0 ( a ) Q_M^0(a) Q M 0 ( a ) . Its integral differs from Q ∞ ( a ) Q_\infty(a) Q ∞ ( a ) by at most C a M − d C_aM^{-d} C a M − d . For an integer L > d L>d L > d ,
max i ∥ ∂ i L F M ∥ 1 ≤ C a M L − d . \max_i\|\partial_i^L F_M\|_1\le C_aM^{L-d}. i max ∥ ∂ i L F M ∥ 1 ≤ C a M L − d .
Indeed, outside the transition annulus this follows by integrating ρ − L \rho^{-L} ρ − L from 1 / M 1/M 1/ M to a fixed radius; on the annulus each Leibniz term is O a ( M L ) O_a(M^L) O a ( M L ) and the volume is O ( M − d ) O(M^{-d}) O ( M − d ) . The derivatives vanish inside the inner ball. Applying Lemma A.1 proves (A.2).
For the diagonal estimate use instead F M = ( 1 − χ ( M k ) ) λ J ( k ) − 1 F_M=(1-\chi(Mk))\lambda_J(k)^{-1} F M = ( 1 − χ ( M k )) λ J ( k ) − 1 . Its grid average is g M 0 ( 0 ) g_M^0(0) g M 0 ( 0 ) , the removed integral is O ( M 2 − d ) O(M^{2-d}) O ( M 2 − d ) , and (A.5) gives
max i ∥ ∂ i L F M ∥ 1 ≤ C M L + 2 − d . \max_i \|\partial_i^L F_M\|_1 \leq C M^{L+2-d}. i max ∥ ∂ i L F M ∥ 1 ≤ C M L + 2 − d .
The same aliasing estimate proves (A.3). In both applications the Fourier formula is applied to the smooth cutoff function, whose derivative norms depend on M M M .
The order in (A.2) is sharp for this class of fixed sources. For b i = δ 0 − δ e i b_i=\delta_0-\delta_{e_i} b i = δ 0 − δ e i , ∑ i J i ∣ b ^ i ( k ) ∣ 2 = λ J ( k ) \sum_i J_i|\widehat b_i(k)|^2=\lambda_J(k) ∑ i J i ∣ b i ( k ) ∣ 2 = λ J ( k ) , and hence
∑ i J i { Q M 0 ( b i ) − Q ∞ ( b i ) } = − M − d , (61) \sum_i J_i\{Q_M^0(b_i)-Q_\infty(b_i)\}=-M^{-d}, \tag*{(61)} i ∑ J i { Q M 0 ( b i ) − Q ∞ ( b i )} = − M − d , (61)
max i ∣ Q M 0 ( b i ) − Q ∞ ( b i ) ∣ ≥ M − d j . \max_i|Q_M^0(b_i)-Q_\infty(b_i)|\geq\frac{M^{-d}}{j}. i max ∣ Q M 0 ( b i ) − Q ∞ ( b i ) ∣ ≥ j M − d .
The individual directional errors may have either sign.
The mass equation and the crossover
Define
H M ( s ) = 1 V ∑ k ≠ 0 1 λ J ( k ) ( λ J ( k ) + s ) . H_M(s)=\frac{1}{V}\sum_{k\ne0}\frac{1}{\lambda_J(k)(\lambda_J(k)+s)}. H M ( s ) = V 1 k = 0 ∑ λ J ( k ) ( λ J ( k ) + s ) 1 .
a d ( M ) = { M , d = 3 , log M , d = 4 , 1 , d ≥ 5 , t d ( M ) = a d ( M ) V , ρ d ( M ) = ( V a d ( M ) ) − 1 / 2 . a_d(M)=
\begin{cases}
M, & d=3,\\
\log M, & d=4,\\
1, & d\geq5,
\end{cases}
\qquad
t_d(M)=\sqrt{\frac{a_d(M)}{V}},
\qquad
\rho_d(M)=(V a_d(M))^{-1/2}. a d ( M ) = ⎩ ⎨ ⎧ M , log M , 1 , d = 3 , d = 4 , d ≥ 5 , t d ( M ) = V a d ( M ) , ρ d ( M ) = ( V a d ( M ) ) − 1/2 .
Lemma A.3 (Centred-cell bound and resolvent shells). For β ≥ β c , J \beta\geq\beta_{c,J} β ≥ β c , J ,
1 β V ≤ s M ≤ η M : = 2 j ( 1 − sin ( π / M ) π / M ) ≤ j π 2 3 M 2 . (62) \frac{1}{\beta V}\leq s_M\leq\eta_M:=2j\left(1-\frac{\sin(\pi/M)}{\pi/M}\right)\leq\frac{j\pi^2}{3M^2}. \tag*{(62)} β V 1 ≤ s M ≤ η M := 2 j ( 1 − π / M sin ( π / M ) ) ≤ 3 M 2 j π 2 . (62) For each fixed C 0 < ∞ C_0<\infty C 0 < ∞ , uniformly for 0 ≤ s ≤ C 0 M − 2 0\leq s\leq C_0M^{-2} 0 ≤ s ≤ C 0 M − 2 ,
c C 0 a d ( M ) ≤ H M ( s ) ≤ C a d ( M ) . (63) c_{C_0}a_d(M)\leq H_M(s)\leq C a_d(M). \tag*{(63)} c C 0 a d ( M ) ≤ H M ( s ) ≤ C a d ( M ) . (63)
Proof. The lower bound in (A.8) follows immediately from β = g M ( s M ) ≥ 1 / ( V s M ) \beta=g_M(s_M)\geq1/(V s_M) β = g M ( s M ) ≥ 1/ ( V s M ) . To obtain the upper bound, partition the Fourier torus into cells centred at the points of K M K_M K M , each with coordinate half-width π / M \pi/M π / M . Put c M = sin ( π / M ) / ( π / M ) c_M=\sin(\pi/M)/(\pi/M) c M = sin ( π / M ) / ( π / M ) . The cell average of cos ( k i + u i ) \cos(k_i+u_i) cos ( k i + u i ) is c M cos k i c_M\cos k_i c M cos k i , so the average of λ J \lambda_J λ J over the cell centred at k k k is exactly
c M λ J ( k ) + 2 j ( 1 − c M ) ≤ λ J ( k ) + η M . c_M\lambda_J(k)+2j(1-c_M)\leq\lambda_J(k)+\eta_M. c M λ J ( k ) + 2 j ( 1 − c M ) ≤ λ J ( k ) + η M .
Jensen’s inequality for the convex function x ↦ 1 / x x\mapsto1/x x ↦ 1/ x therefore gives the cell average of λ J − 1 \lambda_J^{-1} λ J − 1 at least ( λ J ( k ) + η M ) − 1 (\lambda_J(k)+\eta_M)^{-1} ( λ J ( k ) + η M ) − 1 . At the cell containing zero, apply Jensen first to ( λ J + ϵ ) − 1 (\lambda_J+\epsilon)^{-1} ( λ J + ϵ ) − 1 and then let ϵ ↓ 0 \epsilon\downarrow0 ϵ ↓ 0 ; the integral is finite since d > 2 d>2 d > 2 and (A.4) holds. Summing the cells proves β c , J ≥ g M ( η M ) \beta_{c,J}\geq g_M(\eta_M) β c , J ≥ g M ( η M ) . Strict decrease of g M g_M g M , together with g M ( s M ) = β ≥ β c , J g_M(s_M)=\beta\geq\beta_{c,J} g M ( s M ) = β ≥ β c , J , implies s M ≤ η M s_M\leq\eta_M s M ≤ η M . Finally, 1 − sin u / u ≤ u 2 / 6 1-\sin u/u\leq u^2/6 1 − sin u / u ≤ u 2 /6 yields the last bound in (A.8).
Choose centred integer representatives k = 2 π q / M k=2\pi q/M k = 2 π q / M with ∣ q i ∣ ≤ M / 2 |q_i|\leq M/2 ∣ q i ∣ ≤ M /2 . By (A.4) and the bound O ( n d − 1 ) O(n^{d-1}) O ( n d − 1 ) for the shell ∣ q ∣ ∞ = n |q|_\infty=n ∣ q ∣ ∞ = n ,
H M ( 0 ) ≤ C M 4 − d ∑ n = 1 ⌊ M / 2 ⌋ n d − 5 ≤ C a d ( M ) . H_M(0)\leq C M^{4-d}\sum_{n=1}^{\lfloor M/2\rfloor} n^{d-5}\leq C a_d(M). H M ( 0 ) ≤ C M 4 − d n = 1 ∑ ⌊ M /2 ⌋ n d − 5 ≤ C a d ( M ) .
For the reverse bound in d = 3 d=3 d = 3 , the single frequency q = e 1 q=e_1 q = e 1 gives H M ( 0 ) ≥ c M H_M(0)\ge cM H M ( 0 ) ≥ c M . In d = 4 d=4 d = 4 , use the full shells 1 ≤ n ≤ ⌊ M / 4 ⌋ 1\le n\le\lfloor M/4\rfloor 1 ≤ n ≤ ⌊ M /4 ⌋ : they have ( 2 n + 1 ) 4 − ( 2 n − 1 ) 4 ≥ c n 3 (2n+1)^4-(2n-1)^4\ge cn^3 ( 2 n + 1 ) 4 − ( 2 n − 1 ) 4 ≥ c n 3 points, and λ J ( 2 π q / M ) ≤ C n 2 / M 2 \lambda_J(2\pi q/M)\le Cn^2/M^2 λ J ( 2 π q / M ) ≤ C n 2 / M 2 there. Their contributions sum to at least c log M c\log M c log M , after adjusting the constant for the finitely many small values of M M M . For d ≥ 5 d\ge5 d ≥ 5 , the global bound λ J ≤ 4 j \lambda_J\le4j λ J ≤ 4 j gives H M ( 0 ) ≥ ( 1 − V − 1 ) / ( 4 j ) 2 ≥ c H_M(0)\ge(1-V^{-1})/(4j)^2\ge c H M ( 0 ) ≥ ( 1 − V − 1 ) / ( 4 j ) 2 ≥ c . Thus H M ( 0 ) ≍ a d ( M ) H_M(0)\asymp a_d(M) H M ( 0 ) ≍ a d ( M ) .
Every nonzero eigenvalue is at least c M − 2 cM^{-2} c M − 2 . For s ≤ C 0 M − 2 s\le C_0M^{-2} s ≤ C 0 M − 2 , the multiplier λ J ( k ) / ( λ J ( k ) + s ) \lambda_J(k)/(\lambda_J(k)+s) λ J ( k ) / ( λ J ( k ) + s ) lies between a positive constant depending on C 0 C_0 C 0 and one. This proves (A.9).
Proposition A.4 (Uniform critical-to-ordered crossover). With τ = β − β c , J ≥ 0 \tau=\beta-\beta_{c,J}\ge0 τ = β − β c , J ≥ 0 , there are uniform positive constants c , C c,C c , C such that
c V { τ + t d ( M ) } ≤ s M ≤ C V { τ + t d ( M ) } . (64) \frac{c}{V\{\tau+t_d(M)\}}\le s_M\le\frac{C}{V\{\tau+t_d(M)\}}. \tag*{(64)} V { τ + t d ( M )} c ≤ s M ≤ V { τ + t d ( M )} C . (64) At criticality this gives
s M ≍ ρ d ( M ) = { M − 2 , d = 3 , M − 2 ( log M ) − 1 / 2 , d = 4 , M − d / 2 , d ≥ 5. (65) s_M\asymp\rho_d(M)=
\begin{cases}
M^{-2}, & d=3,\\
M^{-2}(\log M)^{-1/2}, & d=4,\\
M^{-d/2}, & d\ge5.
\end{cases} \tag*{(65)} s M ≍ ρ d ( M ) = ⎩ ⎨ ⎧ M − 2 , M − 2 ( log M ) − 1/2 , M − d /2 , d = 3 , d = 4 , d ≥ 5. (65) If τ ≥ κ > 0 \tau\ge\kappa>0 τ ≥ κ > 0 , then, uniformly on that parameter set,
1 β V ≤ s M ≤ 2 κ V ( M ≥ M 0 ( κ ) ) , V s M = 1 τ + O κ ( M 2 − d ) . (66) \frac{1}{\beta V}\le s_M\le\frac{2}{\kappa V}\quad(M\ge M_0(\kappa)),\qquad Vs_M=\frac{1}{\tau}+O_\kappa(M^{2-d}). \tag*{(66)} β V 1 ≤ s M ≤ κV 2 ( M ≥ M 0 ( κ )) , V s M = τ 1 + O κ ( M 2 − d ) . (66) Proof. The exact resolvent identity g M 0 ( s ) = g M 0 ( 0 ) − s H M ( s ) g_M^0(s)=g_M^0(0)-sH_M(s) g M 0 ( s ) = g M 0 ( 0 ) − s H M ( s ) rewrites the mass equation as
1 V s M = τ − b M + s M H M ( s M ) , \frac{1}{Vs_M}=\tau-b_M+s_MH_M(s_M), V s M 1 = τ − b M + s M H M ( s M ) , 1 = V s M ( τ − b M ) + V s M 2 H M ( s M ) . (67) 1=Vs_M(\tau-b_M)+Vs_M^2H_M(s_M). \tag*{(67)} 1 = V s M ( τ − b M ) + V s M 2 H M ( s M ) . (67) At τ = 0 \tau=0 τ = 0 , the following calculation uses only the bounds on J J J , even when β c , J \beta_{c,J} β c , J lies outside [ b , B ] [b,B] [ b , B ] . Put z = s M / ρ d ( M ) z=s_M/\rho_d(M) z = s M / ρ d ( M ) . By Lemma A.3, H M ( s M ) ≍ a d ( M ) H_M(s_M)\asymp a_d(M) H M ( s M ) ≍ a d ( M ) ; by (A.3), ∣ V s M b M ∣ ≤ C z M 2 ρ d ( M ) |V s_Mb_M|\le CzM^2\rho_d(M) ∣ V s M b M ∣ ≤ C z M 2 ρ d ( M ) . Here
M 2 ρ d ( M ) = { 1 , d = 3 , ( log M ) − 1 / 2 , d = 4 , M 2 − d / 2 , d ≥ 5 , M^2\rho_d(M)=
\begin{cases}
1, & d=3,\\
(\log M)^{-1/2}, & d=4,\\
M^{2-d/2}, & d\ge5,
\end{cases} M 2 ρ d ( M ) = ⎩ ⎨ ⎧ 1 , ( log M ) − 1/2 , M 2 − d /2 , d = 3 , d = 4 , d ≥ 5 , is uniformly bounded. The critical mass identity consequently implies
c z 2 − C z ≤ 1 ≤ C z 2 + C z . cz^2-Cz\le1\le Cz^2+Cz. c z 2 − C z ≤ 1 ≤ C z 2 + C z . The first inequality bounds z z z above by the positive root of a fixed quadratic; the second bounds it below by a positive root. This proves (A.11).
Write t = t d ( M ) t=t_d(M) t = t d ( M ) , so ρ d ( M ) = 1 / ( V t ) \rho_d(M)=1/(Vt) ρ d ( M ) = 1/ ( V t ) . In all dimensions, (A.3) implies ∣ b M ∣ ≤ C b t |b_M|\le C_b t ∣ b M ∣ ≤ C b t . Since s M s_M s M decreases as β \beta β increases, the critical estimate gives s M ≤ C ρ d ( M ) s_M\le C\rho_d(M) s M ≤ C ρ d ( M ) throughout τ ≥ 0 \tau\ge0 τ ≥ 0 . Let L 0 = max ( 1 , 2 C b ) L_0=\max(1,2C_b) L 0 = max ( 1 , 2 C b ) . If τ ≥ L 0 t \tau\ge L_0t τ ≥ L 0 t , then τ − b M ≥ τ / 2 \tau-b_M\ge\tau/2 τ − b M ≥ τ /2 , so (A.13) gives
s M ≤ 2 V τ ≤ 4 V ( τ + t ) . s_M\le\frac{2}{V\tau}\le\frac{4}{V(\tau+t)}. s M ≤ V τ 2 ≤ V ( τ + t ) 4 . If τ < L 0 t \tau<L_0t τ < L 0 t , the critical upper bound gives s M ≤ C / ( V t ) ≤ C ( L 0 + 1 ) / [ V ( τ + t ) ] s_M\le C/(Vt)\le C(L_0+1)/[V(\tau+t)] s M ≤ C / ( V t ) ≤ C ( L 0 + 1 ) / [ V ( τ + t )] . This proves the upper bound in (A.10).
For the lower bound, set u = V s M ( τ + t ) u=Vs_M(\tau+t) u = V s M ( τ + t ) . Using H M ( s M ) ≤ C a d ( M ) H_M(s_M)\le C a_d(M) H M ( s M ) ≤ C a d ( M ) in the second identity of (A.13) gives
1 ≤ u τ + C b t τ + t + C u 2 a d ( M ) / V ( τ + t ) 2 ≤ C ′ ( u + u 2 ) . 1\le u\frac{\tau+C_b t}{\tau+t}+C u^2\frac{a_d(M)/V}{(\tau+t)^2}\le C'(u+u^2). 1 ≤ u τ + t τ + C b t + C u 2 ( τ + t ) 2 a d ( M ) / V ≤ C ′ ( u + u 2 ) .
Thus u u u is bounded below by a uniform positive constant. Both sides of (A.10) hold also when τ \tau τ varies with M M M .
Finally, if τ ≥ κ \tau\ge\kappa τ ≥ κ , choose M 0 M_0 M 0 so that ∣ b M ∣ ≤ κ / 2 |b_M|\le\kappa/2 ∣ b M ∣ ≤ κ /2 . The first mass identity implies s M ≤ 2 / ( κ V ) s_M\le2/(\kappa V) s M ≤ 2/ ( κV ) , and the zero mode gives the stated lower bound. The shell estimate now yields
0 ≤ s M H M ( s M ) ≤ C κ V − 1 a d ( M ) ≤ C κ M 2 − d . 0\le s_M H_M(s_M)\le C_\kappa V^{-1}a_d(M)\le C_\kappa M^{2-d}. 0 ≤ s M H M ( s M ) ≤ C κ V − 1 a d ( M ) ≤ C κ M 2 − d .
Consequently 1 / ( V s M ) = τ + O κ ( M 2 − d ) 1/(Vs_M)=\tau+O_\kappa(M^{2-d}) 1/ ( V s M ) = τ + O κ ( M 2 − d ) , with the denominator bounded below by κ / 2 \kappa/2 κ /2 . Taking its reciprocal proves (A.12).
The finite-size behaviour of the spherical-model mass is classical; see [4 ] . Related finite-size analyses of spherical, large-N N N and ϕ 4 \phi^4 ϕ 4 models include [6 , 7 , 10 , 30 , 31 ] .
The auxiliary Gaussian reference
Let U \mathcal{U} U be a fixed set of distinct lattice points and ℓ = ∣ U ∣ \ell=|\mathcal{U}| ℓ = ∣ U ∣ , with M M M large enough that it embeds injectively in the torus. Define the covariance matrices restricted to this set by
Σ M ( x , y ) = β − 1 G J , M , s M ( x , y ) , Σ ∞ ( x , y ) = β − 1 G J ( x , y ) + 1 − β c , J / β . \Sigma_M(x,y)=\beta^{-1}G_{J,M,s_M}(x,y),\qquad\Sigma_\infty(x,y)=\beta^{-1}G_J(x,y)+1-\beta_{c,J}/\beta. Σ M ( x , y ) = β − 1 G J , M , s M ( x , y ) , Σ ∞ ( x , y ) = β − 1 G J ( x , y ) + 1 − β c , J / β .
Here Σ ∞ = Σ β , J \Sigma_\infty=\Sigma_{\beta,J} Σ ∞ = Σ β , J is the target covariance (2) , and both matrices have diagonal one. As in Remark 2.7, denote by γ M U , p \gamma_M^{\mathcal{U},p} γ M U , p and γ β , J U , p \gamma_{\beta,J}^{\mathcal{U},p} γ β , J U , p the laws of p p p independent centred Gaussian columns with these covariances. The metric below is W 1 W_1 W 1 for the Euclidean norm on R ℓ × p \mathbb{R}^{\ell\times p} R ℓ × p . We now prove the bounds stated in Remark 2.7.
Proof. For a fixed zero-sum source let Q M , s ( a ) = V − 1 ∑ k ≠ 0 ∣ a ^ ( k ) ∣ 2 / ( λ J ( k ) + s ) Q_{M,s}(a)=V^{-1}\sum_{k\ne0}|\widehat{a}(k)|^2/(\lambda_J(k)+s) Q M , s ( a ) = V − 1 ∑ k = 0 ∣ a ( k ) ∣ 2 / ( λ J ( k ) + s ) . By (A.4) and the first-order vanishing of a ^ \widehat{a} a , ∣ a ^ ( k ) ∣ 2 ≤ C a λ J ( k ) |\widehat{a}(k)|^2\le C_a\lambda_J(k) ∣ a ( k ) ∣ 2 ≤ C a λ J ( k ) on the torus. Consequently
0 ≤ Q M 0 ( a ) − Q M , s M ( a ) = s M V ∑ k ≠ 0 ∣ a ^ ( k ) ∣ 2 λ J ( k ) ( λ J ( k ) + s M ) ≤ C a s M g M 0 ( 0 ) ≤ C a s M , 0\le Q_M^0(a)-Q_{M,s_M}(a)=\frac{s_M}{V}\sum_{k\ne0}\frac{|\widehat{a}(k)|^2}{\lambda_J(k)(\lambda_J(k)+s_M)}\le C_a s_M g_M^0(0)\le C_a s_M, 0 ≤ Q M 0 ( a ) − Q M , s M ( a ) = V s M k = 0 ∑ λ J ( k ) ( λ J ( k ) + s M ) ∣ a ( k ) ∣ 2 ≤ C a s M g M 0 ( 0 ) ≤ C a s M ,
where (A.3) bounds g M 0 ( 0 ) g_M^0(0) g M 0 ( 0 ) uniformly. For each dipole a = δ x − δ y a=\delta_x-\delta_y a = δ x − δ y , its variance under Σ M \Sigma_M Σ M is Q M , s M ( a ) / β Q_{M,s_M}(a)/\beta Q M , s M ( a ) / β , and under Σ ∞ \Sigma_\infty Σ ∞ it is Q ∞ ( a ) / β Q_\infty(a)/\beta Q ∞ ( a ) / β because the constant mode cancels. The diagonal entries are one, so Lemma A.2 gives
∥ Σ M − Σ ∞ ∥ F ≤ C U ( s M + M − d ) ≤ C U s M . \|\Sigma_M-\Sigma_\infty\|_F\le C_{\mathcal{U}}(s_M+M^{-d})\le C_{\mathcal{U}}s_M. ∥ Σ M − Σ ∞ ∥ F ≤ C U ( s M + M − d ) ≤ C U s M .
The last inequality uses M − d ≤ β s M ≤ B s M M^{-d}\le\beta s_M\le B s_M M − d ≤ β s M ≤ B s M from (A.8).
By (30) , Σ ∞ ⪰ m I \Sigma_\infty\succeq mI Σ ∞ ⪰ m I with m = ( 4 β j ) − 1 m=(4\beta j)^{-1} m = ( 4 β j ) − 1 . The square-root estimate (33) , applied with A = Σ M A=\Sigma_M A = Σ M and Σ = Σ ∞ \Sigma=\Sigma_\infty Σ = Σ ∞ , gives
∥ Σ M 1 / 2 − Σ ∞ 1 / 2 ∥ F ≤ m − 1 / 2 ∥ Σ M − Σ ∞ ∥ F . \|\Sigma_M^{1/2}-\Sigma_\infty^{1/2}\|_F\le m^{-1/2}\|\Sigma_M-\Sigma_\infty\|_F. ∥ Σ M 1/2 − Σ ∞ 1/2 ∥ F ≤ m − 1/2 ∥ Σ M − Σ ∞ ∥ F .
Set E Σ = Σ M 1 / 2 − Σ ∞ 1 / 2 E_\Sigma=\Sigma_M^{1/2}-\Sigma_\infty^{1/2} E Σ = Σ M 1/2 − Σ ∞ 1/2 . Coupling the two Gaussian blocks by the same ℓ × p \ell\times p ℓ × p standard Gaussian matrix Z Z Z gives
W 1 ≤ E ∥ E Σ Z ∥ F ≤ p ∥ E Σ ∥ F , W_1\le\mathbb{E}\|E_\Sigma Z\|_F\le\sqrt{p}\,\|E_\Sigma\|_F, W 1 ≤ E ∥ E Σ Z ∥ F ≤ p ∥ E Σ ∥ F ,
which proves (6).
For the lower bound let v i M v_i^M v i M , v i ∞ v_i^\infty v i ∞ be the single-column variances of X 0 − X e i X_0-X_{e_i} X 0 − X e i under the two laws. Since ∑ i J i ∣ b ^ i ∣ 2 = λ J \sum_i J_i|\hat b_i|^2=\lambda_J ∑ i J i ∣ b ^ i ∣ 2 = λ J , the mass equation gives the exact identities
∑ i J i v i M = 1 β V ∑ k λ J ( k ) λ J ( k ) + s M = 1 β − s M , ∑ i J i v i ∞ = 1 β . \sum_i J_i v_i^M=\frac{1}{\beta V}\sum_k\frac{\lambda_J(k)}{\lambda_J(k)+s_M}=\frac{1}{\beta}-s_M,\qquad\sum_i J_i v_i^\infty=\frac{1}{\beta}. i ∑ J i v i M = β V 1 k ∑ λ J ( k ) + s M λ J ( k ) = β 1 − s M , i ∑ J i v i ∞ = β 1 .
Thus
∑ i J i ( v i ∞ − v i M ) = s M , v i ∞ − v i M ≥ s M / j for some i . (68) \sum_i J_i(v_i^\infty-v_i^M)=s_M,\qquad v_i^\infty-v_i^M\ge s_M/j\quad\text{for some }i. \tag*{(68)} i ∑ J i ( v i ∞ − v i M ) = s M , v i ∞ − v i M ≥ s M / j for some i . (68)
The second assertion uses only the weighted average; individual differences may have either sign. The selected direction may depend on M M M , β \beta β , J J J , and the star contains all possible directions. Together with ∑ i J i Q M 0 ( b i ) = 1 − V − 1 \sum_i J_iQ_M^0(b_i)=1-V^{-1} ∑ i J i Q M 0 ( b i ) = 1 − V − 1 from the proof of Lemma 4.3 , the first identity also gives the Gaussian deficit ∑ i J i { Q M 0 ( b i ) / β − v i M } = s M − 1 / ( β V ) \sum_i J_i\{Q_M^0(b_i)/\beta-v_i^M\}=s_M-1/(\beta V) ∑ i J i { Q M 0 ( b i ) / β − v i M } = s M − 1/ ( β V ) used in § 3 and § 8 .
For centred isotropic Gaussian laws in R p \mathbb{R}^p R p , let G G G be a standard Gaussian vector and write χ p = ∣ G ∣ \chi_p=|G| χ p = ∣ G ∣ . Then
W 1 ( N ( 0 , v I p ) , N ( 0 , w I p ) ) = E χ p ∣ v − w ∣ . W_1\left(\mathcal{N}(0,vI_p),\mathcal{N}(0,wI_p)\right)=\mathbb{E}\chi_p\left|\sqrt{v}-\sqrt{w}\right|. W 1 ( N ( 0 , v I p ) , N ( 0 , w I p ) ) = E χ p v − w .
Coupling the two laws with the same G G G gives the upper bound, and the 1-Lipschitz test ∣ x ∣ |x| ∣ x ∣ gives equality. Pointwise ∣ G ∣ ≥ p − 1 / 2 ∑ α = 1 p ∣ G α ∣ |G|\ge p^{-1/2}\sum_{\alpha=1}^p|G_\alpha| ∣ G ∣ ≥ p − 1/2 ∑ α = 1 p ∣ G α ∣ , so
E χ p ≥ 2 π p . \mathbb{E}\chi_p\ge\sqrt{\frac{2}{\pi}}\sqrt{p}. E χ p ≥ π 2 p .
The two laws of the dipole vector ( X 0 α − X e α α ) α = 1 p (X_0^\alpha-X_{e_\alpha}^\alpha)_{\alpha=1}^p ( X 0 α − X e α α ) α = 1 p have covariances v i M I p v_i^M I_p v i M I p and v i ∞ I p v_i^\infty I_p v i ∞ I p , since their Gaussian columns are independent. Both dipole variances are at most four because the site variances are one. Their standard deviations therefore sum to at most four, and (A.14) yields vector distance at least 2 / π p s M / ( 4 j ) \sqrt{2/\pi}\sqrt{p}\,s_M/(4j) 2/ π p s M / ( 4 j ) . The map from the entire block to this dipole vector has Lipschitz constant 2 \sqrt{2} 2 . Contraction of W 1 W_1 W 1 under that map proves (7) . Projection proves the assertion for a larger local set.
Statements and Declarations
Acknowledgements. We acknowledge Gewu Intelligence Lab for providing the collaborative research environment in which this project was developed. Large language models (OpenAI Codex and Anthropic Claude) assisted with literature retrieval, proof development and checking, and editing. The authors checked the resulting derivations and take full responsibility for the scientific content and conclusions of this paper.
Competing interests. The authors declare no competing interests.
Data availability. No datasets were generated or analysed in this study.