The proofs were generated through extensive interaction with the AI models Claude (Opus 5.5) and ChatGPT (Sol 6.1), based on the problem formulation and guidance of the authors.
Introduction
The thermodynamic limit of a disordered system is taken as the system size N → ∞ N \to\infty N → ∞ at fixed intensive parameters. Establishing its existence provides a basis for the analysis of limiting thermodynamic observables. For mean-field spin glasses, including the Sherrington–Kirkpatrick model [47 ] , Guerra and Toninelli [31 ] proved convergence of the statistical pressure by interpolation. Subsequent work established the corresponding replica-symmetry-breaking predictions: the Parisi variational formula [44 ] follows from Guerra’s interpolation bounds [30 ] and Talagrand’s matching result [50 ] , while Panchenko [41 ] proved that the Ghirlanda–Guerra identities [28 ] imply ultrametricity. These results address distinct questions: existence of the limit, its value, and the limiting overlap structure.
For associative neural networks, the rigorous theory remains less complete. The Hopfield model [33 ] is an Ising system whose interaction matrix is a sum of rank-one pattern matrices; its statistical mechanics at extensive load, where the number K K K of stored patterns grows proportionally to the number N N N of neurons, was developed by Amit, Gutfreund and Sompolinsky [4 , 5 , 6 ] using the replica method at the replica symmetric level of description. The Hubbard–Stratonovich transformation represents the model as a bipartite system of N N N Ising spins and K K K Gaussian “hidden” spins, also interpretable as a hybrid Boltzmann machine [11 , 12 , 7 ] . Blocking yields a vector spin glass whose covariance is the product of two partial traces of the overlap matrix. This covariance is not convex, so neither the Guerra–Toninelli argument nor the classical theory of multi-species and vector spin glasses [42 , 43 , 8 ] applies directly. Configurations may also correlate with individual patterns, producing signal (Mattis) magnetizations [34 ] .
Rigorous contributions to the extensively loaded Hopfield model include retrieval Gibbs states [18 , 19 ] , self-averaging [16 ] , concentration-based replica-symmetric equations [45 ] , and results at high temperature or small load where the replica-symmetric picture can be justified [45 , 17 , 49 , 51 ] . For Gaussian patterns, the annealed region β ≤ ( 1 + α ) − 1 \beta\le(1+\sqrt{\alpha})^{-1} β ≤ ( 1 + α ) − 1 is understood [14 , 13 ] . Agliari, Barra, Bianco, Fachechi and Pallara Agliari, Barra, Bianco, Fachechi and Pallara [2 ] obtained a hard–soft interpolation identity and proved convergence under replica-symmetric concentration assumptions. Concentration around an expected statistical pressure, however, does not by itself establish that the expectation converges, and convergence throughout the parameter space has remained open for binary patterns; see the discussion in [2 ] .
For Gaussian patterns, Camilli, Contucci and Mingione Camilli, Contucci and Mingione [20 ] have announced convergence of the quenched statistical pressure at every α , β > 0 \alpha,\beta> 0 α , β > 0 , without an external field or additional pattern, together with an explicit variational principle for the limit. Our approach also treats a uniform external field and an additional binary pattern, and extends to Rademacher patterns. It does not determine the limit explicitly (see Remark 2.9 ).
We prove existence of the thermodynamic limit of the quenched statistical pressure of the Hopfield model with independent symmetric Rademacher patterns, at every load α ∈ ( 0 , ∞ ) \alpha\in(0,\infty) α ∈ ( 0 , ∞ ) , every finite inverse temperature and every uniform external field, with a limit that depends on ( K ( N ) ) (K(N)) ( K ( N )) only through α \alpha α (Theorems 2.1 and 2.2 ). For binary patterns, convergence was previously known only in restricted regions of the ( α , β ) (\alpha,\beta) ( α , β ) plane, typically at high temperature or small load where the replica-symmetric picture can be justified [17 , 49 , 51 ] , or conditionally on hypotheses on the Gibbs measure, such as self-averaging of the Edwards–Anderson order parameter [45 ] or replica-symmetric concentration of the overlaps [2 ] .
The proof applies to all α > 0 \alpha> 0 α > 0 , β < ∞ \beta< \infty β < ∞ and uniform fields, without replica symmetry or concentration of overlaps or Mattis magnetizations. It also requires no non-condensation assumption, that is, no hypothesis that typical configurations have vanishing overlaps with the patterns. This distinction matters for disorder comparison: the direct replacement of all binary patterns by Gaussian ones (Lemma 5.2 ) yields only the inequality A R ≥ A G A^{R} \ge A^{G} A R ≥ A G , because its error is controlled only for bounded hidden coordinates and a vanishing truncation error is available only on the Gaussian side. No binary-to-Gaussian universality is assumed or proved; the two limiting pressures are related by inequalities (Corollary 2.6).
The pressure also concentrates around its mean with exponentially small deviation probabilities and converges almost surely (Theorem 2.3 ). The almost-sure statement holds for any coupling of the pattern matrices across volumes. Convexity in β \beta β then transfers convergence from the pressure to its derivative: the internal energy and the entropy converge at every differentiability point of the limiting pressure, hence outside an at most countable set of inverse temperatures, and one-sided bounds hold at the remaining points (Theorem 2.4 ). Letting β → ∞ \beta\to\infty β → ∞ gives convergence of the ground-state energy (Corollary 9.3).
The passage from Gaussian to binary disorder requires an additional comparison. The compact vector-spin theorem of Ho and Chen, Ho and Issa [32 , 23 ] (Theorem 3.6) requires the interaction to be a Gaussian process with a prescribed covariance. Binary disorder is not of this form, so the theorem cannot be applied to the binary model itself, and replacing all binary patterns by Gaussian ones gives only a one-sided bound. For the upper bound, we retain the patterns whose hidden coordinates may be large and replace the others. This finite-pattern extraction reduces the binary model to Gaussian bulks in which finitely many binary patterns are retained, and to which the compact theorem applies through Theorem 4.8 .
The retained patterns are not fixed in advance. After a radial cutoff ∥ z ∥ 2 ≤ ϱ N \lVert z\rVert^2 \le\varrho N ∥ z ∥ 2 ≤ ϱN of the hidden spins, we partition the hidden integral into sectors indexed by the set S S S of hidden coordinates larger than δ N \delta\sqrt{N} δ N , with ∣ S ∣ ≤ 1 + ⌊ ϱ / δ 2 ⌋ \lvert S\rvert\le1+\lfloor\varrho/\delta^2\rfloor ∣ S ∣ ≤ 1 + ⌊ ϱ / δ 2 ⌋ . On each sector the patterns in S S S , associated with potentially macroscopic hidden coordinates, are retained as binary and all others are replaced. The proof does not require identifying condensed equilibrium states; Remark 6.3 explains the relation between large hidden coordinates and Mattis overlaps.
The extraction uses three estimates. The first, a collective replacement (Lemma 6.2 ), replaces all small-coordinate columns in a single interpolation under one common Gibbs measure and sums the fourth moments before bounding them. The resulting error is O ( β 2 ϱ δ 2 ) O(\beta^2\varrho\delta^2) O ( β 2 ϱ δ 2 ) , uniformly in N N N , K K K , S S S and h h h , whereas an entry-by-entry replacement would diverge (Remark 6.3). The second estimate controls sector selection (§ 6.3 ). The dominant sector depends on the disorder, but the number of sectors is polynomial in N N N and the sector pressures are convex in the disorder with Lipschitz constants O ( N − 1 / 2 ) O(N^{-1/2}) O ( N − 1/2 ) ; convex concentration on the discrete cube bounds the cost of selecting the dominant sector inside the quenched expectation by O ( log N / N ) O(\sqrt{\log N/N}) O ( log N / N ) . The third estimate is a lower comparison (Lemma 5.2 ), in which an arbitrary finite set of binary patterns is retained and the rest are replaced after a truncation of the hidden spins at level N 1 / 6 N^{1/6} N 1/6 , the Gaussian side being controlled by Proposition 4.2. Taking N → ∞ N\to\infty N → ∞ at fixed δ \delta δ , and then δ ↓ 0 \delta\downarrow0 δ ↓ 0 , gives A R = sup r p r A^{R}=\sup_{r}p_{r} A R = sup r p r without any uniformity in r r r . Replacement methods [22 ] , spin-glass universality [21 ] , bipartite comparison [26 ] and mixed Boolean–Gaussian models [3 ] provide background, but no unrestricted universality statement is invoked.
Both comparisons rely on the Gaussian analysis, which also proves convergence for Gaussian patterns. At h = 0 h=0 h = 0 , Theorem 2.1 for Gaussian patterns provides an alternative proof of the existence result announced in [20 ] . Its first ingredient is a uniform truncation (Proposition 4.2). The hidden spins are unbounded, while Theorem 3.6 requires compact spins; truncating them at level L L L changes the pressure by at most C c u t L − 2 C_{\mathrm{cut}}L^{-2} C cut L − 2 , uniformly in the volume, in a compact load interval and in an arbitrary Ising potential. The bound follows from the rotational invariance of the Gaussian pattern matrix and requires no assumption on the Gibbs measure. It allows the limits N → ∞ N\to\infty N → ∞ and L → ∞ L\to\infty L → ∞ to be exchanged, and its uniformity in the potential permits comparisons with arbitrary extensive potentials in § 5 –§ 6 .
The second ingredient extends the Gaussian result to bulks whose visible sites carry finitely many types of fields, with arbitrary proportions, uniformly on compact sets of fields (Theorem 4.8 ). The mixed models require this extension, since their types are the sign patterns of the retained binary columns, and a scalar uniform-field result would not suffice. The same analysis treats a uniform field and an additional binary Mattis pattern (Theorem 2.5 ), with explicit concentration bounds (Proposition 8.1 ).
The representation A R = sup r p r A^{R}=\sup_{r}p_{r} A R = sup r p r leaves the optimal number of retained patterns undetermined, including whether one suffices. By Corollary 2.6, one suffices at zero field exactly when the binary pressure equals the Gaussian pressure with one additional binary pattern. The number of retained patterns is not a replica-symmetry-breaking depth. The argument does not prove that the remainder in the hard–soft interpolation of [2 ] vanishes. Broken-replica interpolation methods [30 , 1 ] provide related background, but no identification with their variational functionals is needed below.
The proof begins with the external convergence result in Theorem 3.6 . Together with Proposition 4.2, it yields Theorem 4.8 and hence Theorem 2.1 for Gaussian patterns (§ 4 ). Theorem 4.8 then gives the mixed-model limits (Proposition 5.1 ), which, combined with Lemmas 5.2 and 6.2 and the sector selection, give Theorem 2.2 and Theorem 2.1 for binary patterns (Sections 5–7). Concentration yields sample convergence (§ 8 ). Convexity gives the results on energy, entropy and ground states (§ 9 ), and gauge invariance yields the reduction for an additional pattern (§ 10 ). § 2 introduces the models and states the main results, § 3 collects the preliminary tools, and § 11 discusses the results.
Models and main results
We write Σ N : = { − 1 , 1 } N \Sigma_N := \{-1,1\}^{N} Σ N := { − 1 , 1 } N , m ( σ ) : = N − 1 ∑ i σ i m(\sigma) := N^{-1}\sum_i \sigma_i m ( σ ) := N − 1 ∑ i σ i , and γ \gamma γ for the standard Gaussian measure on R \mathbb{R} R ; Y Y Y denotes a standard Gaussian random variable or vector. For a matrix M M M , ∥ M ∥ op \lVert M\rVert_{\operatorname{op}} ∥ M ∥ op and ∥ M ∥ H S \lVert M\rVert_{\mathrm{HS}} ∥ M ∥ HS are the operator and Hilbert–Schmidt norms. The expectation E \mathbb{E} E is over all disorder variables present; subscripts indicate a restricted expectation. The symbol ω \omega ω denotes the Gibbs expectation associated with the partition function under consideration. Constants C C C may change between occurrences; their parameter dependence is indicated when relevant. All logarithms are natural.
For a real N × K N \times K N × K matrix J J J with columns J μ ∈ R N J^\mu\in\mathbb{R}^{N} J μ ∈ R N let
H N ( σ ; J ) : = − 1 2 N ∑ μ = 1 K ( J μ ⋅ σ ) 2 = − 1 2 N ∥ J T σ ∥ 2 , σ ∈ Σ N . (1) H_N(\sigma;J):=-\frac{1}{2N}\sum_{\mu=1}^{K}(J^\mu\cdot\sigma)^2=-\frac{1}{2N}\lVert J^{T}\sigma\rVert^2,\qquad\sigma\in\Sigma_N. \tag*{(1)} H N ( σ ; J ) := − 2 N 1 μ = 1 ∑ K ( J μ ⋅ σ ) 2 = − 2 N 1 ∥ J T σ ∥ 2 , σ ∈ Σ N . (1)
The diagonal terms i = j i=j i = j are retained; see Remark 2.8. For β ≥ 0 \beta\geq0 β ≥ 0 and a potential V : Σ N → R V:\Sigma_N\to\mathbb{R} V : Σ N → R put
f N ( J ) : = 1 N log ∑ σ ∈ Σ N e − β H N ( σ ; J ) + V ( σ ) , c V : = log 2 + 1 N max Σ N ∣ V ∣ , (2) f_N(J):=\frac{1}{N}\log\sum_{\sigma\in\Sigma_N}e^{-\beta H_N(\sigma;J)+V(\sigma)},\qquad c_V:=\log2+\frac{1}{N}\max_{\Sigma_N}|V|, \tag*{(2)} f N ( J ) := N 1 log σ ∈ Σ N ∑ e − β H N ( σ ; J ) + V ( σ ) , c V := log 2 + N 1 Σ N max ∣ V ∣ , (2)
suppressing the dependence on β \beta β .
We consider two pattern laws. Let G = ( g i μ ) G=(g_i^\mu) G = ( g i μ ) be an N × K N\times K N × K matrix of independent standard Gaussian variables and ε = ( ε i μ ) \varepsilon=(\varepsilon_i^\mu) ε = ( ε i μ ) an N × K N\times K N × K matrix of independent symmetric Rademacher variables. Throughout, ∙ ∈ { G , R } \bullet\in\{G,R\} ∙ ∈ { G , R } labels the pattern law, with J G : = G J^G:=G J G := G and J R : = ε J^R:=\varepsilon J R := ε . For h ∈ R h\in\mathbb{R} h ∈ R let
Z N , K ∙ ( β , h ) : = ∑ σ ∈ Σ N e − β H N ( σ ; J ∙ ) + h N m ( σ ) , A ^ N , K ∙ : = 1 N log Z N , K ∙ , A N , K ∙ : = E A ^ N , K ∙ . (3) Z^\bullet_{N,K}(\beta,h):=\sum_{\sigma\in\Sigma_N}e^{-\beta H_N(\sigma;J^\bullet)+hNm(\sigma)},\qquad\widehat{A}^\bullet_{N,K}:=\frac{1}{N}\log Z^\bullet_{N,K},\qquad A^\bullet_{N,K}:=\mathbb{E}\widehat{A}^\bullet_{N,K}. \tag*{(3)} Z N , K ∙ ( β , h ) := σ ∈ Σ N ∑ e − β H N ( σ ; J ∙ ) + h N m ( σ ) , A N , K ∙ := N 1 log Z N , K ∙ , A N , K ∙ := E A N , K ∙ . (3)
Thus A ^ N , K ∙ ( β , h ) = f N h N m ( J ∙ ) \widehat{A}^\bullet_{N,K}(\beta,h)=f_N^{hNm}(J^\bullet) A N , K ∙ ( β , h ) = f N h N m ( J ∙ ) . Following the convention of [31 ] , A ^ N , K ∙ \widehat{A}^\bullet_{N,K} A N , K ∙ is the (random) statistical pressure and A N , K ∙ A^\bullet_{N,K} A N , K ∙ the quenched statistical pressure . For β > 0 \beta>0 β > 0 the free energy is
F N , K ∙ : = − ( β N ) − 1 log Z N , K ∙ , F^\bullet_{N,K}:=-(\beta N)^{-1}\log Z^\bullet_{N,K}, F N , K ∙ := − ( β N ) − 1 log Z N , K ∙ ,
so that A ^ N , K ∙ = − β F N , K ∙ \widehat{A}^\bullet_{N,K}=-\beta F^\bullet_{N,K} A N , K ∙ = − β F N , K ∙ ; the statistical pressure is defined for all β ≥ 0 \beta\geq0 β ≥ 0 , and all results are stated for it. The field h h h is dimensionless and conjugate to m m m ; at h = 0 h=0 h = 0 we recover the ordinary Hopfield model. We write Z N , K ∙ ( h ) Z^\bullet_{N,K}(h) Z N , K ∙ ( h ) , A N , K ∙ ( h ) A^\bullet_{N,K}(h) A N , K ∙ ( h ) when β \beta β is fixed and Z N , K ∙ ( β ) Z^\bullet_{N,K}(\beta) Z N , K ∙ ( β ) , A N , K ∙ ( β ) A^\bullet_{N,K}(\beta) A N , K ∙ ( β ) when h = 0 h=0 h = 0 , and set Z N , 0 ( β , h ) : = ( 2 cosh h ) N Z_{N,0}(\beta,h):=(2\cosh h)^N Z N , 0 ( β , h ) := ( 2 cosh h ) N . Limits are denoted A G A^G A G and A R A^R A R ; their equality is neither assumed nor proved.
For an integer r ≥ 1 r\geq1 r ≥ 1 and K ≥ r K\geq r K ≥ r , let ζ 1 , … , ζ r ∈ Σ N \zeta^1,\ldots,\zeta^r\in\Sigma_N ζ 1 , … , ζ r ∈ Σ N be independent patterns with independent symmetric Rademacher entries, and let G ′ G' G ′ be an independent standard Gaussian N × ( K − r ) N\times(K-r) N × ( K − r ) matrix. The mixed model has pattern matrix $(\zeta^1,\ldots,\zeta^r,G'):
Z N , K ( r ) ( β , h ) : = ∑ σ exp { β 2 N ∥ G ′ T σ ∥ 2 + β 2 N ∑ a = 1 r ( ζ a ⋅ σ ) 2 + h N m ( σ ) } , p N , K ( r ) : = 1 N E log Z N , K ( r ) . (4) Z^{(r)}_{N,K}(\beta,h):=\sum_{\sigma}\exp\left\{\frac{\beta}{2N}\lVert G'^{T}\sigma\rVert^2+\frac{\beta}{2N}\sum_{a=1}^{r}(\zeta^a\cdot\sigma)^2+hNm(\sigma)\right\},\qquad p^{(r)}_{N,K}:=\frac{1}{N}\mathbb{E}\log Z^{(r)}_{N,K}. \tag*{(4)} Z N , K ( r ) ( β , h ) := σ ∑ exp { 2 N β ∥ G ′ T σ ∥ 2 + 2 N β a = 1 ∑ r ( ζ a ⋅ σ ) 2 + h N m ( σ ) } , p N , K ( r ) := N 1 E log Z N , K ( r ) . (4)
These are auxiliary comparison models for binary patterns. For fixed r r r , the sequence p N , K ( N ) ( r ) p^{(r)}_{N,K(N)} p N , K ( N ) ( r ) is considered for N ≥ N r N \ge N_r N ≥ N r , with N r N_r N r chosen so that K ( N ) ≥ r K(N) \ge r K ( N ) ≥ r for all N ≥ N r N \ge N_r N ≥ N r . Such an N r N_r N r exists when K ( N ) / N → α > 0 K(N)/N \to\alpha> 0 K ( N ) / N → α > 0 ; the terms with K ( N ) < r K(N) < r K ( N ) < r are left undefined.
Let η ∈ Σ N \eta\in\Sigma_N η ∈ Σ N be random, with arbitrary law (which may depend on N N N and need not have independent coordinates), independent of J ∙ \mathbf{J}^{\bullet} J ∙ . The model with the additional (Mattis) pattern η \eta η is
H N M ( σ ) : = H N ( σ ; J ∙ ) − ( η ⋅ σ ) 2 2 N , A ^ N , K ∙ , M ( β ) : = 1 N log ∑ σ ∈ Σ N e − β H N M ( σ ) , A N , K ∙ , M : = E A ^ N , K ∙ , M . (5) H_N^M(\sigma) := H_N(\sigma;\mathbf{J}^{\bullet})-\frac{(\eta\cdot\sigma)^2}{2N}, \qquad\widehat{A}^{\bullet,M}_{N,K}(\beta):=\frac{1}{N}\log\sum_{\sigma\in\Sigma_N}e^{-\beta H_N^M(\sigma)}, \qquad A^{\bullet,M}_{N,K}:=\mathbb{E}\widehat{A}^{\bullet,M}_{N,K}. \tag*{(5)} H N M ( σ ) := H N ( σ ; J ∙ ) − 2 N ( η ⋅ σ ) 2 , A N , K ∙ , M ( β ) := N 1 log σ ∈ Σ N ∑ e − β H N M ( σ ) , A N , K ∙ , M := E A N , K ∙ , M . (5)
There is no additional uniform field in this definition. Models of this type, with symmetric binary η \eta η , are considered in [2 , 3 ] . We call η \eta η an additional pattern : no condensation, i.e. no nonzero equilibrium overlap with η \eta η , is asserted below.
Main results
Theorem 2.1 (Thermodynamic limit). Let ∙ ∈ { G , R } \bullet\in\{G,R\} ∙ ∈ { G , R } , α ∈ ( 0 , ∞ ) \alpha\in(0,\infty) α ∈ ( 0 , ∞ ) and K ( N ) / N → α K(N)/N \to\alpha K ( N ) / N → α . For every β ≥ 0 \beta\ge0 β ≥ 0 and h ∈ R h \in\mathbb{R} h ∈ R the limit
A ∙ ( α , β , h ) : = lim N → ∞ A N , K ( N ) ∙ ( β , h ) A^{\bullet}(\alpha,\beta,h):=\lim_{N\to\infty}A^{\bullet}_{N,K(N)}(\beta,h) A ∙ ( α , β , h ) := N → ∞ lim A N , K ( N ) ∙ ( β , h ) exists, is finite and depends on ( K ( N ) ) (K(N)) ( K ( N )) only through α \alpha α . The convergence is locally uniform in ( β , h ) ∈ [ 0 , ∞ ) × R (\beta,h)\in[0,\infty)\times\mathbb{R} ( β , h ) ∈ [ 0 , ∞ ) × R . We write A ∙ ( α , β ) : = A ∙ ( α , β , 0 ) A^{\bullet}(\alpha,\beta):=A^{\bullet}(\alpha,\beta,0) A ∙ ( α , β ) := A ∙ ( α , β , 0 ) .
Theorem 2.2 (Binary patterns through finitely many retained patterns). Let α ∈ ( 0 , ∞ ) \alpha\in(0,\infty) α ∈ ( 0 , ∞ ) and K ( N ) / N → α K(N)/N\to\alpha K ( N ) / N → α . For every fixed r ≥ 1 r\ge1 r ≥ 1 , K ( N ) ≥ r K(N)\ge r K ( N ) ≥ r for all sufficiently large N N N . For every β ≥ 0 \beta\ge0 β ≥ 0 and h ∈ R h\in\mathbb{R} h ∈ R the limit p r ( α , β , h ) : = lim N p N , K ( N ) ( r ) ( β , h ) p_r(\alpha,\beta,h):=\lim_N p^{(r)}_{N,K(N)}(\beta,h) p r ( α , β , h ) := lim N p N , K ( N ) ( r ) ( β , h ) , taken along these sufficiently large N N N , exists, is finite and depends on ( K ( N ) ) (K(N)) ( K ( N )) only through α \alpha α . The convergence is locally uniform on [ 0 , ∞ ) × R [0,\infty)\times\mathbb{R} [ 0 , ∞ ) × R , including compact sets that meet β = 0 \beta=0 β = 0 . The sequence r ↦ p r ( α , β , h ) r\mapsto p_r(\alpha,\beta,h) r ↦ p r ( α , β , h ) is nondecreasing and bounded, and
A R ( α , β , h ) = sup r ≥ 1 p r ( α , β , h ) = lim r → ∞ p r ( α , β , h ) . (6) A^R(\alpha,\beta,h)=\sup_{r\ge1}p_r(\alpha,\beta,h)=\lim_{r\to\infty}p_r(\alpha,\beta,h). \tag*{(6)} A R ( α , β , h ) = r ≥ 1 sup p r ( α , β , h ) = r → ∞ lim p r ( α , β , h ) . (6) The number r r r of retained patterns is fixed while taking the thermodynamic limit and may increase afterwards. Formula (6) does not assert that r = 1 r=1 r = 1 suffices.
Theorem 2.3 (Concentration and almost-sure limit).
(a) Exponential concentration. Let ∙ ∈ { G , R } \bullet\in\{G,R\} ∙ ∈ { G , R } , β > 0 \beta>0 β > 0 , α + > 0 \alpha_+>0 α + > 0 , c ˉ > 0 \bar{c}>0 c ˉ > 0 and t > 0 t>0 t > 0 . There are c > 0 c>0 c > 0 and C < ∞ C<\infty C < ∞ such that for all N , K ≥ 1 N,K\ge1 N , K ≥ 1 with K ≤ α + N K\le\alpha_+N K ≤ α + N and all deterministic V : Σ N → R V:\Sigma_N\to\mathbb{R} V : Σ N → R with c V ≤ c ˉ cV\le\bar{c} c V ≤ c ˉ ,
P ( ∣ f N V ( J ∙ ) − E f N V ( J ∙ ) ∣ > t ) ≤ C e − c N . \mathbb{P}\left(\left|f_N^V(\mathbf{J}^{\bullet})-\mathbb{E}f_N^V(\mathbf{J}^{\bullet})\right|>t\right)\le Ce^{-cN}. P ( f N V ( J ∙ ) − E f N V ( J ∙ ) > t ) ≤ C e − c N . Explicit bounds are given in Proposition 8.1 .
(b) Almost-sure limit. Let ∙ ∈ { G , R } \bullet\in\{G,R\} ∙ ∈ { G , R } and K ( N ) / N → α ∈ ( 0 , ∞ ) K(N)/N\to\alpha\in(0,\infty) K ( N ) / N → α ∈ ( 0 , ∞ ) , and let the matrices J N , K ( N ) ∙ \mathbf{J}^{\bullet}_{N,K(N)} J N , K ( N ) ∙ , N ≥ 1 N\ge1 N ≥ 1 , be realized on a common probability space with arbitrary dependence between volumes, and let α + : = sup N K ( N ) / N \alpha_+:=\sup_N K(N)/N α + := sup N K ( N ) / N . Then, for every deterministic V : Σ N → R V:\Sigma_N\to\mathbb{R} V : Σ N → R ,
Var f N V ( J N , K ( N ) ∙ ) ≤ 2 β 2 N − 2 E ∥ J N , K ( N ) ∙ ∥ o p 2 ≤ C α + , β / N . \operatorname{Var}f_N^V(\mathbf{J}^{\bullet}_{N,K(N)})\le2\beta^2N^{-2}\mathbb{E}\left\|\mathbf{J}^{\bullet}_{N,K(N)}\right\|_{\mathrm{op}}^2\le C_{\alpha_+,\beta}/N. Var f N V ( J N , K ( N ) ∙ ) ≤ 2 β 2 N − 2 E J N , K ( N ) ∙ op 2 ≤ C α + , β / N . There is an event of probability one on which A ^ N , K ( N ) ∙ ( β , h ) → A ∙ ( α , β , h ) \widehat{A}^{\bullet}_{N,K(N)}(\beta,h)\to A^{\bullet}(\alpha,\beta,h) A N , K ( N ) ∙ ( β , h ) → A ∙ ( α , β , h ) for every ( β , h ) ∈ [ 0 , ∞ ) × R (\beta,h)\in[0,\infty)\times\mathbb{R} ( β , h ) ∈ [ 0 , ∞ ) × R , locally uniformly. For fixed ( β , h ) (\beta,h) ( β , h ) the convergence also holds in L 2 L^2 L 2 , and for every compact C ⊂ [ 0 , ∞ ) × R C\subset[0,\infty)\times\mathbb{R} C ⊂ [ 0 , ∞ ) × R ,
sup C ∣ A ^ N , K ( N ) ∙ − A ∙ ∣ → 0 in L 1 . \sup_C\left|\widehat{A}^{\bullet}_{N,K(N)}-A^{\bullet}\right|\to0\quad\text{in }L^1. C sup A N , K ( N ) ∙ − A ∙ → 0 in L 1 .
Let ω β ( σ ) : = e − β H N ( σ ; J ∙ ) / Z N , K ∙ ( β ) \omega_{\beta}(\sigma):=e^{-\beta H_N(\sigma;J^{\bullet})}/Z^{\bullet}_{N,K}(\beta) ω β ( σ ) := e − β H N ( σ ; J ∙ ) / Z N , K ∙ ( β ) . The internal energy and entropy per spin and their quenched averages are
u N ∙ : = 1 N ω β ( H N ) , s N ∙ : = − 1 N ∑ σ ω β ( σ ) log ω β ( σ ) , E N ∙ : = E u N ∙ , S N ∙ : = E s N ∙ , (7) u_N^{\bullet}:=\frac{1}{N}\omega_{\beta}(H_N),\quad s_N^{\bullet}:=-\frac{1}{N}\sum_{\sigma}\omega_{\beta}(\sigma)\log\omega_{\beta}(\sigma),\quad E_N^{\bullet}:=\mathbb{E}u_N^{\bullet},\quad S_N^{\bullet}:=\mathbb{E}s_N^{\bullet}, \tag*{(7)} u N ∙ := N 1 ω β ( H N ) , s N ∙ := − N 1 σ ∑ ω β ( σ ) log ω β ( σ ) , E N ∙ := E u N ∙ , S N ∙ := E s N ∙ , (7)
with K = K ( N ) K=K(N) K = K ( N ) . The same quantities for (5) are denoted u N ∙ , M u_N^{\bullet,M} u N ∙ , M , s N ∙ , M s_N^{\bullet,M} s N ∙ , M , E N ∙ , M E_N^{\bullet,M} E N ∙ , M , S N ∙ , M S_N^{\bullet,M} S N ∙ , M . With the sign convention of (1) , u N ∙ ≤ 0 u_N^{\bullet}\le0 u N ∙ ≤ 0 .
Theorem 2.4 (Internal energy and entropy). Let ∙ ∈ { G , R } \bullet\in\{G,R\} ∙ ∈ { G , R } , h = 0 h=0 h = 0 , α > 0 \alpha>0 α > 0 and K ( N ) / N → α K(N)/N\to\alpha K ( N ) / N → α .
(i) A N , K ( N ) ∙ ( ⋅ ) A_{N,K(N)}^{\bullet}(\cdot) A N , K ( N ) ∙ ( ⋅ ) is convex and C 1 C^1 C 1 on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) , with ∂ β A N , K ( N ) ∙ = − E N ∙ \partial_{\beta}A_{N,K(N)}^{\bullet}=-E_N^{\bullet} ∂ β A N , K ( N ) ∙ = − E N ∙ and A N , K ( N ) ∙ = − β E N ∙ + S N ∙ A_{N,K(N)}^{\bullet}=-\beta E_N^{\bullet}+S_N^{\bullet} A N , K ( N ) ∙ = − β E N ∙ + S N ∙ .
(ii) A ∙ ( α , ⋅ ) A^{\bullet}(\alpha,\cdot) A ∙ ( α , ⋅ ) is convex, and the set N α ∙ \mathcal{N}_{\alpha}^{\bullet} N α ∙ of β > 0 \beta>0 β > 0 where it is not differentiable is at most countable. For every β ∈ ( 0 , ∞ ) ∖ N α ∙ \beta\in(0,\infty)\setminus\mathcal{N}_{\alpha}^{\bullet} β ∈ ( 0 , ∞ ) ∖ N α ∙ the limits E ∙ ( α , β ) : = lim N E N ∙ ( β ) E^{\bullet}(\alpha,\beta):=\lim_{N}E_N^{\bullet}(\beta) E ∙ ( α , β ) := lim N E N ∙ ( β ) and S ∙ ( α , β ) : = lim N S N ∙ ( β ) S^{\bullet}(\alpha,\beta):=\lim_{N}S_N^{\bullet}(\beta) S ∙ ( α , β ) := lim N S N ∙ ( β ) exist and
E ∙ = − ∂ β A ∙ , A ∙ = − β E ∙ + S ∙ . (8) E^{\bullet}=-\partial_{\beta}A^{\bullet},\qquad A^{\bullet}=-\beta E^{\bullet}+S^{\bullet}. \tag*{(8)} E ∙ = − ∂ β A ∙ , A ∙ = − β E ∙ + S ∙ . (8) On an event of probability one, u N ∙ ( β ) → E ∙ ( α , β ) u_N^{\bullet}(\beta)\to E^{\bullet}(\alpha,\beta) u N ∙ ( β ) → E ∙ ( α , β ) and s N ∙ ( β ) → S ∙ ( α , β ) s_N^{\bullet}(\beta)\to S^{\bullet}(\alpha,\beta) s N ∙ ( β ) → S ∙ ( α , β ) simultaneously for all β ∈ ( 0 , ∞ ) ∖ N α ∙ \beta\in(0,\infty)\setminus\mathcal{N}_{\alpha}^{\bullet} β ∈ ( 0 , ∞ ) ∖ N α ∙ ; for each such β \beta β the convergence also holds in L p L^p L p for every 1 ≤ p < ∞ 1\le p<\infty 1 ≤ p < ∞ .
(iii) For every β > 0 \beta>0 β > 0 , the limit points of E N ∙ ( β ) E_N^{\bullet}(\beta) E N ∙ ( β ) , and almost surely those of u N ∙ ( β ) u_N^{\bullet}(\beta) u N ∙ ( β ) , lie in [ − ∂ β + A ∙ ( α , β ) , − ∂ β − A ∙ ( α , β ) ] [-\partial_{\beta}^{+}A^{\bullet}(\alpha,\beta),-\partial_{\beta}^{-}A^{\bullet}(\alpha,\beta)] [ − ∂ β + A ∙ ( α , β ) , − ∂ β − A ∙ ( α , β )] ; correspondingly those of S N ∙ ( β ) S_N^{\bullet}(\beta) S N ∙ ( β ) and s N ∙ ( β ) s_N^{\bullet}(\beta) s N ∙ ( β ) lie in [ A ∙ − β ∂ β + A ∙ , A ∙ − β ∂ β − A ∙ ] [A^{\bullet}-\beta\partial_{\beta}^{+}A^{\bullet},A^{\bullet}-\beta\partial_{\beta}^{-}A^{\bullet}] [ A ∙ − β ∂ β + A ∙ , A ∙ − β ∂ β − A ∙ ] .
(iv) On ( 0 , ∞ ) ∖ N α ∙ (0,\infty)\setminus\mathcal{N}_{\alpha}^{\bullet} ( 0 , ∞ ) ∖ N α ∙ , E ∙ ( α , ⋅ ) E^{\bullet}(\alpha,\cdot) E ∙ ( α , ⋅ ) is nonincreasing, E ∙ ≤ − α / 2 E^{\bullet}\le-\alpha/2 E ∙ ≤ − α /2 and 0 ≤ S ∙ ≤ log 2 0\le S^{\bullet}\le\log2 0 ≤ S ∙ ≤ log 2 . Moreover E G ≥ − 1 2 ( 1 + α ) 2 E^{G}\ge-\frac{1}{2}(1+\sqrt{\alpha})^2 E G ≥ − 2 1 ( 1 + α ) 2 .
The same holds for the additional-pattern models, with A ∙ , M A^{\bullet,M} A ∙ , M , E ∙ , M E^{\bullet,M} E ∙ , M , S ∙ , M S^{\bullet,M} S ∙ , M and the corresponding exceptional set N α ∙ , M \mathcal{N}_{\alpha}^{\bullet,M} N α ∙ , M , except for the bounds on the energy in (iv).
Theorem 2.5 (Additional binary pattern). Let ∙ ∈ { G , R } \bullet\in\{G,R\} ∙ ∈ { G , R } , α > 0 \alpha>0 α > 0 , K ( N ) / N → α K(N)/N\to\alpha K ( N ) / N → α and β ≥ 0 \beta\ge0 β ≥ 0 . The limit A ∙ , M ( α , β ) : = lim N A N , K ( N ) ∙ , M ( β ) A^{\bullet,M}(\alpha,\beta):=\lim_N A_{N,K(N)}^{\bullet,M}(\beta) A ∙ , M ( α , β ) := lim N A N , K ( N ) ∙ , M ( β ) exists, does not depend on the law of η \eta η , and
A ∙ , M ( α , β ) = max ∣ v ∣ ≤ 1 { A ∙ ( α , β , β v ) − β v 2 2 } = sup v ∈ R { A ∙ ( α , β , β v ) − β v 2 2 } . (9) A^{\bullet,M}(\alpha,\beta)=\max_{|v|\le1}\left\{A^{\bullet}(\alpha,\beta,\beta v)-\frac{\beta v^2}{2}\right\}=\sup_{v\in\mathbb{R}}\left\{A^{\bullet}(\alpha,\beta,\beta v)-\frac{\beta v^2}{2}\right\}. \tag*{(9)} A ∙ , M ( α , β ) = ∣ v ∣ ≤ 1 max { A ∙ ( α , β , β v ) − 2 β v 2 } = v ∈ R sup { A ∙ ( α , β , β v ) − 2 β v 2 } . (9) The random pressures A ^ N , K ( N ) ∙ , M \widehat{A}_{N,K(N)}^{\bullet,M} A N , K ( N ) ∙ , M converge to A ∙ , M A^{\bullet,M} A ∙ , M in L 2 L^2 L 2 and almost surely, locally uniformly in β \beta β , for any joint realization of ( J N , K ( N ) ∙ , η N ) N ≥ 1 (J_{N,K(N)}^{\bullet},\eta_N)_{N\ge1} ( J N , K ( N ) ∙ , η N ) N ≥ 1 . For binary patterns, moreover,
A R , M ( α , β ) = A R ( α , β ) ; equivalently A R ( α , β , β v ) ≤ A R ( α , β ) + β v 2 2 ( ∣ v ∣ ≤ 1 ) . (10) A^{R,M}(\alpha,\beta)=A^R(\alpha,\beta);\qquad\text{equivalently}\qquad A^R(\alpha,\beta,\beta v)\le A^R(\alpha,\beta)+\frac{\beta v^2}{2}\quad(|v|\le1). \tag*{(10)} A R , M ( α , β ) = A R ( α , β ) ; equivalently A R ( α , β , β v ) ≤ A R ( α , β ) + 2 β v 2 ( ∣ v ∣ ≤ 1 ) . (10) The parameter v v v in (9) is a variational parameter for the Mattis magnetization, i.e. the overlap with η \eta η ; the formula does not by itself establish a nonzero equilibrium overlap or retrieval. Identity (10) reflects that an additional binary pattern is one more binary pattern; it has no analogue for the Gaussian bulk, where it would compare a binary pattern with Gaussian ones. This asymmetry is quantified by the following consequence of Theorems 2.2 and 2.5.
Corollary 2.6 (Comparison of the two pattern laws). For all α > 0 \alpha>0 α > 0 , β ≥ 0 \beta\ge0 β ≥ 0 and h ∈ R h\in\mathbb{R} h ∈ R ,
A G ( α , β , h ) ≤ p 1 ( α , β , h ) ≤ A R ( α , β , h ) , p 1 ( α , β , 0 ) = A G , M ( α , β ) . A^G(\alpha,\beta,h)\le p_1(\alpha,\beta,h)\le A^R(\alpha,\beta,h),\qquad p_1(\alpha,\beta,0)=A^{G,M}(\alpha,\beta). A G ( α , β , h ) ≤ p 1 ( α , β , h ) ≤ A R ( α , β , h ) , p 1 ( α , β , 0 ) = A G , M ( α , β ) . In particular A G ( α , β ) ≤ A G , M ( α , β ) ≤ A R ( α , β ) = A R , M ( α , β ) A^G(\alpha,\beta)\le A^{G,M}(\alpha,\beta)\le A^R(\alpha,\beta)=A^{R,M}(\alpha,\beta) A G ( α , β ) ≤ A G , M ( α , β ) ≤ A R ( α , β ) = A R , M ( α , β ) , and sup r p r ( α , β , 0 ) = p 1 ( α , β , 0 ) \sup_r p_r(\alpha,\beta,0)=p_1(\alpha,\beta,0) sup r p r ( α , β , 0 ) = p 1 ( α , β , 0 ) if and only if A R ( α , β ) = A G , M ( α , β ) A^R(\alpha,\beta)=A^{G,M}(\alpha,\beta) A R ( α , β ) = A G , M ( α , β ) .
Thus “one retained pattern suffices” at zero field is equivalent to the statement that the binary Hopfield pressure equals that of a single binary pattern added to a Gaussian bulk, in line with the standard heuristic in which non-condensed patterns act as Gaussian noise [6 , 3 ] . The identification p 1 = A G , M p_1=A^{G,M} p 1 = A G , M is restricted to h = 0 h=0 h = 0 : Theorem 2.5 is formulated without a uniform field, and for h ≠ 0 h\ne0 h = 0 the gauge transformation of Lemma 10.1, which turns the additional pattern into a Curie–Weiss term, turns the uniform field into the pattern-dependent field h ∑ i η i σ i h\sum_i\eta_i\sigma_i h ∑ i η i σ i .
Proposition 2.7 (Properties of the limits). Let ∙ ∈ { G , R } \bullet\in\{G,R\} ∙ ∈ { G , R } .
(a) h ↦ A ∙ ( α , β , h ) h\mapsto A^\bullet(\alpha,\beta,h) h ↦ A ∙ ( α , β , h ) is convex, even and 1 1 1 -Lipschitz; β ↦ A ∙ ( α , β , h ) \beta\mapsto A^\bullet(\alpha,\beta,h) β ↦ A ∙ ( α , β , h ) is convex and nondecreasing; α ↦ A ∙ ( α , β , h ) − β α / 2 \alpha\mapsto A^\bullet(\alpha,\beta,h)-\beta\alpha/2 α ↦ A ∙ ( α , β , h ) − β α /2 is nondecreasing, and so is α ↦ A ∙ ( α , β , h ) \alpha\mapsto A^\bullet(\alpha,\beta,h) α ↦ A ∙ ( α , β , h ) .
(b) A ∙ ( α , β , h ) ≥ log ( 2 cosh h ) + β α / 2 A^\bullet(\alpha,\beta,h)\ge\log(2\cosh h)+\beta\alpha/2 A ∙ ( α , β , h ) ≥ log ( 2 cosh h ) + β α /2 , and, if β < 1 \beta<1 β < 1 , A ∙ ( α , β , h ) ≤ log ( 2 cosh h ) − α log ( 1 − β ) / 2 A^\bullet(\alpha,\beta,h)\le\log(2\cosh h)-\alpha\log(1-\beta)/2 A ∙ ( α , β , h ) ≤ log ( 2 cosh h ) − α log ( 1 − β ) /2 .
(c) A G ( α , β , h ) ≤ log ( 2 cosh h ) + β ( 1 + α ) 2 / 2 A^G(\alpha,\beta,h)\le\log(2\cosh h)+\beta(1+\sqrt{\alpha})^2/2 A G ( α , β , h ) ≤ log ( 2 cosh h ) + β ( 1 + α ) 2 /2 , and for 0 < α 1 ≤ α 2 0<\alpha_1\le\alpha_2 0 < α 1 ≤ α 2 , A G ( α 2 , β , h ) − A G ( α 1 , β , h ) ≤ β ( 1 − α 1 / α 2 ) ( 1 + α 2 ) 2 / 2 A^G(\alpha_2,\beta,h)-A^G(\alpha_1,\beta,h)\le\beta(1-\alpha_1/\alpha_2)(1+\sqrt{\alpha_2})^2/2 A G ( α 2 , β , h ) − A G ( α 1 , β , h ) ≤ β ( 1 − α 1 / α 2 ) ( 1 + α 2 ) 2 /2 .
Remark 2.8 (Diagonal terms). Removing the terms i = j i=j i = j subtracts T N : = β ∥ J ∙ ∥ H S 2 / ( 2 N 2 ) T_N:=\beta\lVert J^\bullet\rVert_{\mathrm{HS}}^2/(2N^2) T N := β ∥ J ∙ ∥ HS 2 / ( 2 N 2 ) from A ^ N , K ∙ \widehat A^\bullet_{N,K} A N , K ∙ , T N + β / ( 2 N ) T_N+\beta/(2N) T N + β / ( 2 N ) from A ^ N , K ∙ , M \widehat A^{\bullet,M}_{N,K} A N , K ∙ , M , and T N ( r ) : = β ( N r + ∥ G ′ ∥ H S 2 ) / ( 2 N 2 ) T_N^{(r)}:=\beta(Nr+\lVert G'\rVert_{\mathrm{HS}}^2)/(2N^2) T N ( r ) := β ( N r + ∥ G ′ ∥ HS 2 ) / ( 2 N 2 ) from the mixed pressure N − 1 log Z N , K ( r ) N^{-1}\log Z^{(r)}_{N,K} N − 1 log Z N , K ( r ) , where E T N ( r ) = β K / ( 2 N ) \mathbb{E}T_N^{(r)}=\beta K/(2N) E T N ( r ) = β K / ( 2 N ) and T N ( r ) → β α / 2 T_N^{(r)}\to\beta\alpha/2 T N ( r ) → β α /2 almost surely, as for T N T_N T N below; equivalently, it adds ∥ J ∙ ∥ H S 2 / ( 2 N 2 ) \lVert J^\bullet\rVert_{\mathrm{HS}}^2/(2N^2) ∥ J ∙ ∥ HS 2 / ( 2 N 2 ) , respectively ∥ J ∙ ∥ H S 2 / ( 2 N 2 ) + 1 / ( 2 N ) \lVert J^\bullet\rVert_{\mathrm{HS}}^2/(2N^2)+1/(2N) ∥ J ∙ ∥ HS 2 / ( 2 N 2 ) + 1/ ( 2 N ) , to the energy per spin. For binary patterns T N = β K / ( 2 N ) T_N=\beta K/(2N) T N = β K / ( 2 N ) is deterministic. For Gaussian patterns E T N = β K / ( 2 N ) \mathbb{E}T_N=\beta K/(2N) E T N = β K / ( 2 N ) and
E ( T N − E T N ) 4 = ( β / 2 N 2 ) 4 [ 60 N K + 12 N K ( N K − 1 ) ] ≤ C ( β , α + ) N − 4 , \mathbb{E}(T_N-\mathbb{E}T_N)^4=(\beta/2N^2)^4[60NK+12NK(NK-1)]\le C(\beta,\alpha_+)N^{-4}, E ( T N − E T N ) 4 = ( β /2 N 2 ) 4 [ 60 N K + 12 N K ( N K − 1 )] ≤ C ( β , α + ) N − 4 ,
with α + : = sup N K ( N ) / N \alpha_+:=\sup_N K(N)/N α + := sup N K ( N ) / N , since T N T_N T N is a sum of N K NK N K i.i.d. variables with E ( Y 2 − 1 ) 4 = 60 \mathbb{E}(Y^2-1)^4=60 E ( Y 2 − 1 ) 4 = 60 ; hence T N → β α / 2 T_N\to\beta\alpha/2 T N → β α /2 almost surely and in every L p L^p L p , 1 ≤ p < ∞ 1\le p<\infty 1 ≤ p < ∞ . Write off for quantities without diagonal terms.
(i) Transformation formulas.
A ∙ , o f f = A ∙ − β α 2 , p r o f f = p r − β α 2 , A ∙ , M , o f f = A ∙ , M − β α 2 , g ∙ , o f f = g ∙ − α 2 , A^{\bullet,\mathrm{off}}=A^\bullet-\frac{\beta\alpha}{2},\qquad p_r^{\mathrm{off}}=p_r-\frac{\beta\alpha}{2},\qquad A^{\bullet,M,\mathrm{off}}=A^{\bullet,M}-\frac{\beta\alpha}{2},\qquad g^{\bullet,\mathrm{off}}=g^\bullet-\frac{\alpha}{2}, A ∙ , off = A ∙ − 2 β α , p r off = p r − 2 β α , A ∙ , M , off = A ∙ , M − 2 β α , g ∙ , off = g ∙ − 2 α ,
E ∙ , o f f = E ∙ + α 2 , E ∙ , M , o f f = E ∙ , M + α 2 , S ∙ , o f f = S ∙ , S ∙ , M , o f f = S ∙ , M , E^{\bullet,\mathrm{off}}=E^\bullet+\frac{\alpha}{2},\qquad E^{\bullet,M,\mathrm{off}}=E^{\bullet,M}+\frac{\alpha}{2},\qquad S^{\bullet,\mathrm{off}}=S^\bullet,\qquad S^{\bullet,M,\mathrm{off}}=S^{\bullet,M}, E ∙ , off = E ∙ + 2 α , E ∙ , M , off = E ∙ , M + 2 α , S ∙ , off = S ∙ , S ∙ , M , off = S ∙ , M ,
where − g ∙ -g^\bullet − g ∙ is the ground-state energy of Corollary 9.3.
(ii) Transfer of conclusions. Existence and sequence independence of the limits, local uniformity, almost-sure, L p L^p L p (1 ≤ p < ∞ 1\le p<\infty 1 ≤ p < ∞ ) and self-averaging statements in Theorems 2.1 –2.5 and Corollary 9.3 hold for the off-diagonal quantities. The shifts are linear in β \beta β , so differentiability points, the exceptional sets, the identities (2.8), the one-sided bounds of Theorem 2.4 (iii), the formulas (6) , (9) and (10) (with off-diagonal quantities on both sides) and the inequalities of Corollary 2.6 are unchanged. Exponential concentration (Theorem 2.3 (a)) also persists: for binary patterns the shift is deterministic, and for Gaussian patterns T N − E T N T_N-\mathbb{E}T_N T N − E T N satisfies Bernstein’s inequality for sums of N K NK N K independent sub-exponential variables [15 ] . The explicit constants of Proposition 8.1 (a) are not claimed for the off-diagonal Gaussian model.
(iii) Transformed inequalities. The bounds of Proposition 2.7 and Theorem 2.4 (iv) become
log ( 2 cosh h ) ≤ A ∙ , o f f ( α , β , h ) ≤ log ( 2 cosh h ) − α 2 [ log ( 1 − β ) + β ] ( β < 1 ) , \log(2\cosh h)\le A^{\bullet,\mathrm{off}}(\alpha,\beta,h)\le\log(2\cosh h)-\frac{\alpha}{2}\left[\log(1-\beta)+\beta\right]\qquad(\beta<1), log ( 2 cosh h ) ≤ A ∙ , off ( α , β , h ) ≤ log ( 2 cosh h ) − 2 α [ log ( 1 − β ) + β ] ( β < 1 ) ,
A G , o f f ( α , β , h ) ≤ log ( 2 cosh h ) + β 2 ( 1 + 2 α ) , − 1 2 − α ≤ E G , o f f ≤ 0 , E R , o f f ≤ 0 , (11) A^{G,\mathrm{off}}(\alpha,\beta,h)\le\log(2\cosh h)+\frac{\beta}{2}(1+2\sqrt{\alpha}),\qquad-\frac{1}{2}-\sqrt{\alpha}\le E^{G,\mathrm{off}}\le0,\qquad E^{R,\mathrm{off}}\le0, \tag*{(11)} A G , off ( α , β , h ) ≤ log ( 2 cosh h ) + 2 β ( 1 + 2 α ) , − 2 1 − α ≤ E G , off ≤ 0 , E R , off ≤ 0 , (11)
and, for 0 < α 1 ≤ α 2 0 < \alpha_1 \le\alpha_2 0 < α 1 ≤ α 2 ,
0 ≤ A G , o f f ( α 2 , β , h ) − A G , o f f ( α 1 , β , h ) ≤ β 2 [ ( 1 − α 1 α 2 ) ( 1 + α 2 ) 2 − ( α 2 − α 1 ) ] . 0 \le A^{G,\mathrm{off}}(\alpha_2,\beta,h)-A^{G,\mathrm{off}}(\alpha_1,\beta,h) \le\frac{\beta}{2}\left[\left(1-\frac{\alpha_1}{\alpha_2}\right)(1+\sqrt{\alpha_2})^2-(\alpha_2-\alpha_1)\right]. 0 ≤ A G , off ( α 2 , β , h ) − A G , off ( α 1 , β , h ) ≤ 2 β [ ( 1 − α 2 α 1 ) ( 1 + α 2 ) 2 − ( α 2 − α 1 ) ] .
The properties in h h h , convexity and monotonicity in β \beta β (since E ∙ , o f f ≤ 0 E^{\bullet,\mathrm{off}} \le0 E ∙ , off ≤ 0 ), monotonicity of E ∙ , o f f E^{\bullet,\mathrm{off}} E ∙ , off and the bounds 0 ≤ S ≤ log 2 0 \le S \le\log2 0 ≤ S ≤ log 2 are unchanged, and α ↦ A ∙ , o f f ( α , β , h ) \alpha\mapsto A^{\bullet,\mathrm{off}}(\alpha,\beta,h) α ↦ A ∙ , off ( α , β , h ) is nondecreasing by Proposition 2.7 (a).
Preliminaries
We collect the concentration, convexity and matrix estimates used below. In (T1)–(T3), Y Y Y denotes a standard Gaussian vector in R n \mathbb{R}^n R n .
(T1) Gaussian Poincaré inequality [15 Thm. 3.20] : Var φ ( Y ) ≤ E ∥ ∇ φ ( Y ) ∥ 2 \operatorname{Var}\varphi(Y) \le\mathbb{E}\lVert\nabla\varphi(Y)\rVert^2 Var φ ( Y ) ≤ E ∥ ∇ φ ( Y ) ∥ 2 for φ ∈ C 1 \varphi\in C^1 φ ∈ C 1 with φ ( Y ) , ∇ φ ( Y ) ∈ L 2 \varphi(Y),\nabla\varphi(Y) \in L^2 φ ( Y ) , ∇ φ ( Y ) ∈ L 2 , and, by approximation, for Lipschitz φ \varphi φ .
(T2) Gaussian concentration (Tsirelson–Ibragimov–Sudakov) [15 Thm. 5.6] : if φ \varphi φ is L L L -Lipschitz, then P ( φ ( Y ) − E φ ( Y ) ≥ t ) ≤ e − t 2 / ( 2 L 2 ) \mathbb{P}(\varphi(Y)-\mathbb{E}\varphi(Y)\ge t)\le e^{-t^2/(2L^2)} P ( φ ( Y ) − E φ ( Y ) ≥ t ) ≤ e − t 2 / ( 2 L 2 ) and P ( ∣ φ ( Y ) − E φ ( Y ) ∣ ≥ t ) ≤ 2 e − t 2 / ( 2 L 2 ) \mathbb{P}(|\varphi(Y)-\mathbb{E}\varphi(Y)|\ge t)\le2e^{-t^2/(2L^2)} P ( ∣ φ ( Y ) − E φ ( Y ) ∣ ≥ t ) ≤ 2 e − t 2 / ( 2 L 2 ) .
(T3) McShane extension [35 ] : if φ : B → R \varphi:B\to\mathbb{R} φ : B → R is L L L -Lipschitz on B ⊂ R n B\subset\mathbb{R}^n B ⊂ R n , then x ↦ inf y ∈ B { φ ( y ) + L ∥ x − y ∥ } x\mapsto\inf_{y\in B}\{\varphi(y)+L\lVert x-y\rVert\} x ↦ inf y ∈ B { φ ( y ) + L ∥ x − y ∥} is L L L -Lipschitz on R n \mathbb{R}^n R n and equals φ \varphi φ on B B B .
(T4) Borel–Cantelli : if ∑ N P ( E N ) < ∞ \sum_N\mathbb{P}(E_N)<\infty ∑ N P ( E N ) < ∞ , then almost surely only finitely many E N E_N E N occur (no independence is needed).
(T5) Convex functions [46 Thms. 10.1, 10.8, 24.1] : a finite convex function on an open interval has nondecreasing one-sided derivatives f − ≤ f + f^- \le f^+ f − ≤ f + , with f + ( x ) ≤ f − ( y ) f^+(x)\le f^-(y) f + ( x ) ≤ f − ( y ) for x < y x<y x < y , hence is differentiable outside a countable set; and convex functions converging pointwise to a finite limit on a dense subset converge everywhere, locally uniformly, to a finite convex function.
The next lemma transfers pressure convergence to convergence of the derivative in § 9 . Applied to differentiation in β \beta β , it gives the internal-energy limit and one-sided bounds at nondifferentiability points.
Lemma 3.1. Let f N , f f_N,f f N , f be convex on an open interval I I I , with f N f_N f N differentiable and f N → f f_N\to f f N → f pointwise. Then f − ( x ) ≤ lim inf N f N ′ ( x ) ≤ lim sup N f N ′ ( x ) ≤ f + ( x ) f^-(x)\le\liminf_N f'_N(x)\le\limsup_N f'_N(x)\le f^+(x) f − ( x ) ≤ lim inf N f N ′ ( x ) ≤ lim sup N f N ′ ( x ) ≤ f + ( x ) for every x ∈ I x\in I x ∈ I .
Proof. For y < x < z y<x<z y < x < z in I I I , convexity gives f N ( x ) − f N ( y ) x − y ≤ f N ′ ( x ) ≤ f N ( z ) − f N ( x ) z − x \frac{f_N(x)-f_N(y)}{x-y}\le f'_N(x)\le\frac{f_N(z)-f_N(x)}{z-x} x − y f N ( x ) − f N ( y ) ≤ f N ′ ( x ) ≤ z − x f N ( z ) − f N ( x ) . Let N → ∞ N\to\infty N → ∞ , then y ↑ x y\uparrow x y ↑ x and z ↓ x z\downarrow x z ↓ x . □
Random matrices
The statistical pressure is controlled by the operator norm of the pattern matrix, since 0 ≤ − H N ( σ ; J ) ≤ 1 2 ∥ J ∥ o p 2 0\le-H_N(\sigma;J)\le\frac12\lVert J\rVert_{\mathrm{op}}^2 0 ≤ − H N ( σ ; J ) ≤ 2 1 ∥ J ∥ op 2 . We first collect the moment and tail bounds for Gaussian matrices, which are used in the uniform truncation estimate (§ 4.1 ) and in the concentration bounds for Gaussian patterns.
Lemma 3.2 (Gaussian matrices). For N ≥ 1 N\ge1 N ≥ 1 , K ≥ 0 K\ge0 K ≥ 0 : E ∥ G ∥ o p ≤ N + K \mathbb{E}\lVert G\rVert_{\mathrm{op}}\le\sqrt{N}+\sqrt{K} E ∥ G ∥ op ≤ N + K , E ∥ G ∥ o p 2 ≤ ( N + K ) 2 + 1 ≤ 3 ( N + K ) \mathbb{E}\lVert G\rVert_{\mathrm{op}}^2\le(\sqrt{N}+\sqrt{K})^2+1\le3(N+K) E ∥ G ∥ op 2 ≤ ( N + K ) 2 + 1 ≤ 3 ( N + K ) and E ∥ G ∥ o p 4 ≤ 8 ( N + K ) 4 + 128 ≤ 160 ( N + K ) 2 \mathbb{E}\lVert G\rVert_{\mathrm{op}}^4\le8(\sqrt{N}+\sqrt{K})^4+128\le160(N+K)^2 E ∥ G ∥ op 4 ≤ 8 ( N + K ) 4 + 128 ≤ 160 ( N + K ) 2 . Moreover P ( ∥ G ∥ o p > 2 ( N + K ) ) ≤ e − ( N + K ) 2 / 2 \mathbb{P}(\lVert G\rVert_{\mathrm{op}}>2(\sqrt{N}+\sqrt{K}))\le e^{-(\sqrt{N}+\sqrt{K})^2/2} P (∥ G ∥ op > 2 ( N + K )) ≤ e − ( N + K ) 2 /2 .
Proof. The first bound is the Slepian–Gordon inequality [25 Thm. II.13], [52, Thm. 7.3.1] . Since G ↦ ∥ G ∥ op G \mapsto\lVert G\rVert_{\operatorname{op}} G ↦ ∥ G ∥ op is 1-Lipschitz for the Hilbert–Schmidt norm, Z : = ∥ G ∥ op − E ∥ G ∥ op Z := \lVert G\rVert_{\operatorname{op}}-\mathbb{E}\lVert G\rVert_{\operatorname{op}} Z := ∥ G ∥ op − E ∥ G ∥ op satisfies Var Z ≤ 1 \operatorname{Var} Z \le1 Var Z ≤ 1 by (T1) and P ( ∣ Z ∣ > t ) ≤ 2 e − t 2 / 2 \mathbb{P}(\lvert Z\rvert> t) \le2e^{-t^2/2} P (∣ Z ∣ > t ) ≤ 2 e − t 2 /2 by (T2), whence E Z 4 ≤ 16 \mathbb{E}Z^4 \le16 E Z 4 ≤ 16 and the tail bound. Use ( x + y ) 4 ≤ 8 x 4 + 8 y 4 (x+y)^4 \le8x^4+8y^4 ( x + y ) 4 ≤ 8 x 4 + 8 y 4 , ( N + K ) 2 ≤ 2 ( N + K ) (\sqrt{N}+\sqrt{K})^2 \le2(N+K) ( N + K ) 2 ≤ 2 ( N + K ) and N + K ≥ 1 N+K \ge1 N + K ≥ 1 . □
For binary and mixed pattern matrices, a net argument gives the following bound without rotational invariance. It applies to any matrix with independent Gaussian or Rademacher entries; the constant c ⋆ c_\star c ⋆ below fixes the radius of the high-probability event used in the radial cutoff (Lemma 6.1 ) and in the concentration bounds for binary patterns.
Lemma 3.3 (Sub-Gaussian matrices). Let J J J be an N × K N \times K N × K matrix with independent entries, each standard Gaussian or symmetric Rademacher. Then
P ( ∥ J ∥ op > s ) ≤ min { 1 , 2 ⋅ 9 N + K e − s 2 / 8 } , E ∥ J ∥ op q ≤ C q ( N + K ) q / 2 , 1 ≤ q < ∞ . (12) \mathbb{P}(\lVert J\rVert_{\operatorname{op}}>s) \le\min\{1,2\cdot9^{N+K}e^{-s^2/8}\}, \qquad\mathbb{E}\lVert J\rVert_{\operatorname{op}}^q \le C_q(N+K)^{q/2}, \qquad1\le q<\infty. \tag*{(12)} P (∥ J ∥ op > s ) ≤ min { 1 , 2 ⋅ 9 N + K e − s 2 /8 } , E ∥ J ∥ op q ≤ C q ( N + K ) q /2 , 1 ≤ q < ∞. (12) In particular, if K ≤ α + N K\le\alpha_+N K ≤ α + N and c ⋆ : = ( 8 { 1 + ( 1 + α + ) log 9 } ) 1 / 2 c_\star:= \left(8\{1+(1+\alpha_+)\log9\}\right)^{1/2} c ⋆ := ( 8 { 1 + ( 1 + α + ) log 9 } ) 1/2 (as in Proposition 8.1 (b)), then P ( ∥ J ∥ op > c ⋆ N ) ≤ 2 e − N \mathbb{P}(\lVert J\rVert_{\operatorname{op}}>c_\star\sqrt{N})\le2e^{-N} P (∥ J ∥ op > c ⋆ N ) ≤ 2 e − N for every N N N .
Proof. For unit vectors u , v u,v u , v , the bilinear form u T J v = ∑ i , μ u i v μ J i μ u^{T}Jv=\sum_{i,\mu}u_iv_\mu J_i^\mu u T J v = ∑ i , μ u i v μ J i μ is centered sub-Gaussian with P ( ∣ u T J v ∣ > t ) ≤ 2 e − t 2 / 2 \mathbb{P}(\lvert u^{T}Jv\rvert>t)\le2e^{-t^2/2} P (∣ u T J v ∣ > t ) ≤ 2 e − t 2 /2 . For 1 / 4 1/4 1/4 -nets N 1 , N 2 \mathcal{N}_1,\mathcal{N}_2 N 1 , N 2 of the unit spheres of R N \mathbb{R}^N R N and R K \mathbb{R}^K R K , of cardinalities at most 9 N 9^N 9 N and 9 K 9^K 9 K , one has ∥ J ∥ op ≤ 2 max u ∈ N 1 , v ∈ N 2 u T J v \lVert J\rVert_{\operatorname{op}}\le2\max_{u\in\mathcal{N}_1,v\in\mathcal{N}_2}u^{T}Jv ∥ J ∥ op ≤ 2 max u ∈ N 1 , v ∈ N 2 u T J v [52 Ex. 4.4.3] . A union bound gives the tail estimate, and integrating it gives the moment bound. For the last claim, 2 ⋅ 9 N + K e − c ⋆ 2 N / 8 ≤ 2 exp { ( N + K ) log 9 − N − ( 1 + α + ) N log 9 } ≤ 2 e − N 2\cdot9^{N+K}e^{-c_\star^2N/8}\le2\exp\{(N+K)\log9-N-(1+\alpha_+)N\log9\}\le2e^{-N} 2 ⋅ 9 N + K e − c ⋆ 2 N /8 ≤ 2 exp {( N + K ) log 9 − N − ( 1 + α + ) N log 9 } ≤ 2 e − N . □
Elementary properties of the statistical pressure
We next record the analytic properties of f N V f_N^V f N V in the disorder and the parameters. Lemma 3.4 shows that f N V f_N^V f N V is convex in the disorder, with its gradient controlled by the operator norm. These properties underlie the concentration estimates for both pattern laws. Lemma 3.5 collects the monotonicity, convexity and Lipschitz properties in β \beta β and h h h , the a priori bounds on the quenched pressure, and the annealed bound, which is used for Proposition 2.7 and for the equicontinuity arguments that upgrade pointwise to locally uniform convergence.
Lemma 3.4. Let V : Σ N → R V:\Sigma_N\to\mathbb{R} V : Σ N → R . The function J ↦ f N V ( J ) J\mapsto f_N^V(J) J ↦ f N V ( J ) on R N × K \mathbb{R}^{N\times K} R N × K is smooth and convex, ∇ f N V ( 0 ) = 0 \nabla f_N^V(0)=0 ∇ f N V ( 0 ) = 0 , and
∥ ∇ f N V ( J ) ∥ H S ≤ β N ∥ J ∥ op , ∣ f N V ( J ) ∣ ≤ c V + β 2 N ∥ J ∥ op 2 . (13) \lVert\nabla f_N^V(J)\rVert_{\mathrm{HS}}\le\frac{\beta}{N}\lVert J\rVert_{\operatorname{op}}, \qquad\lvert f_N^V(J)\rvert\le c_V+\frac{\beta}{2N}\lVert J\rVert_{\operatorname{op}}^2. \tag*{(13)} ∥ ∇ f N V ( J ) ∥ HS ≤ N β ∥ J ∥ op , ∣ f N V ( J )∣ ≤ c V + 2 N β ∥ J ∥ op 2 . (13) Proof. Each exponent β 2 N ∥ J T σ ∥ 2 + V ( σ ) \frac{\beta}{2N}\lVert J^T\sigma\rVert^2+V(\sigma) 2 N β ∥ J T σ ∥ 2 + V ( σ ) is convex in J J J , and log-sum-exp preserves convexity. With ω \omega ω the Gibbs measure, ∇ f N V = β N 2 ω ( σ ( J T σ ) T ) \nabla f_N^V=\frac{\beta}{N^2}\omega(\sigma(J^T\sigma)^T) ∇ f N V = N 2 β ω ( σ ( J T σ ) T ) , which vanishes at J = 0 J=0 J = 0 , and Jensen’s inequality for ∥ ⋅ ∥ H S 2 \lVert\cdot\rVert_{\mathrm{HS}}^2 ∥ ⋅ ∥ HS 2 gives ∥ ∇ f N V ∥ H S 2 ≤ β 2 N 4 ω ( N ∥ J T σ ∥ 2 ) ≤ β 2 N 2 ∥ J ∥ op 2 \lVert\nabla f_N^V\rVert_{\mathrm{HS}}^2\le\frac{\beta^2}{N^4}\omega(N\lVert J^T\sigma\rVert^2)\le\frac{\beta^2}{N^2}\lVert J\rVert_{\operatorname{op}}^2 ∥ ∇ f N V ∥ HS 2 ≤ N 4 β 2 ω ( N ∥ J T σ ∥ 2 ) ≤ N 2 β 2 ∥ J ∥ op 2 . The second bound follows from 0 ≤ − H N ( σ ; J ) ≤ 1 2 ∥ J ∥ op 2 0\le-H_N(\sigma;J)\le\frac12\lVert J\rVert_{\operatorname{op}}^2 0 ≤ − H N ( σ ; J ) ≤ 2 1 ∥ J ∥ op 2 . □
Lemma 3.5. Let J J J be a real N × K N\times K N × K matrix and A ^ ( β , h ) : = f N h N m ( J ) \widehat{A}(\beta,h):=f_N^{hN m}(J) A ( β , h ) := f N h N m ( J ) . Pointwise in J J J , log ( 2 cosh h ) ≤ A ^ ≤ log ( 2 cosh h ) + β 2 N ∥ J ∥ op 2 \log(2\cosh h)\le\widehat{A}\le\log(2\cosh h)+\frac{\beta}{2N}\lVert J\rVert_{\operatorname{op}}^2 log ( 2 cosh h ) ≤ A ≤ log ( 2 cosh h ) + 2 N β ∥ J ∥ op 2 and ∣ A ^ ( β , h ) − A ^ ( β , h ′ ) ∣ ≤ ∣ h − h ′ ∣ \lvert\widehat{A}(\beta,h)-\widehat{A}(\beta,h')\rvert\le\lvert h-h'\rvert ∣ A ( β , h ) − A ( β , h ′ )∣ ≤ ∣ h − h ′ ∣ ; A ^ \widehat{A} A is convex and even in h h h , and convex and nondecreasing in β \beta β , with 0 ≤ ∂ β A ^ ≤ ∥ J ∥ op 2 / ( 2 N ) 0\le\partial_\beta\widehat{A}\le\lVert J\rVert_{\operatorname{op}}^2/(2N) 0 ≤ ∂ β A ≤ ∥ J ∥ op 2 / ( 2 N ) . For ∙ ∈ { G , R } \bullet\in\{G,R\} ∙ ∈ { G , R } ,
log ( 2 cosh h ) + β K 2 N ≤ A N , K ∙ ( h ) ≤ log ( 2 cosh h ) + β 2 N E ∥ J ∙ ∥ op 2 , (14) \log(2\cosh h)+\frac{\beta K}{2N}\le A^\bullet_{N,K}(h)\le\log(2\cosh h)+\frac{\beta}{2N}\mathbb{E}\lVert J^\bullet\rVert_{\operatorname{op}}^2, \tag*{(14)} log ( 2 cosh h ) + 2 N β K ≤ A N , K ∙ ( h ) ≤ log ( 2 cosh h ) + 2 N β E ∥ J ∙ ∥ op 2 , (14) with 1 N E ∥ G ∥ op 2 ≤ ( 1 + K / N ) 2 + 1 N \frac{1}{N}\mathbb{E}\lVert G\rVert_{\operatorname{op}}^2\le(1+\sqrt{K/N})^2+\frac{1}{N} N 1 E ∥ G ∥ op 2 ≤ ( 1 + K / N ) 2 + N 1 , and A N , K ∙ ( h ) ≤ log ( 2 cosh h ) − K 2 N log ( 1 − β ) A^\bullet_{N,K}(h)\le\log(2\cosh h)-\frac{K}{2N}\log(1-\beta) A N , K ∙ ( h ) ≤ log ( 2 cosh h ) − 2 N K log ( 1 − β ) if β < 1 \beta<1 β < 1 .
Proof. The pointwise bounds follow from 0 ≤ − H N ≤ 1 2 ∥ J ∥ o p 2 0 \le-H_N \le\frac{1}{2}\lVert J\rVert_{\mathrm{op}}^2 0 ≤ − H N ≤ 2 1 ∥ J ∥ op 2 . With ω \omega ω the Gibbs measure, ∂ h A ^ = ω ( m ) ∈ [ − 1 , 1 ] \partial_h\widehat{A}=\omega(m)\in[-1,1] ∂ h A = ω ( m ) ∈ [ − 1 , 1 ] , ∂ h 2 A ^ = N Var ω ( m ) ≥ 0 \partial_h^2\widehat{A}=N\operatorname{Var}_{\omega}(m)\ge0 ∂ h 2 A = N Var ω ( m ) ≥ 0 , ∂ β A ^ = − N − 1 ω ( H N ) \partial_\beta\widehat{A}=-N^{-1}\omega(H_N) ∂ β A = − N − 1 ω ( H N ) and ∂ β 2 A ^ = N − 1 Var ω ( H N ) ≥ 0 \partial_\beta^2\widehat{A}=N^{-1}\operatorname{Var}_{\omega}(H_N)\ge0 ∂ β 2 A = N − 1 Var ω ( H N ) ≥ 0 ; evenness follows from σ ↦ − σ \sigma\mapsto-\sigma σ ↦ − σ . The upper bound in (14) is the pointwise one, and the bound on E ∥ G ∥ o p 2 \mathbb{E}\lVert G\rVert_{\mathrm{op}}^2 E ∥ G ∥ op 2 is Lemma 3.2. For the lower bound, with ν h ( ± 1 ) : = e ± h / ( 2 cosh h ) \nu_h(\pm1):=e^{\pm h}/(2\cosh h) ν h ( ± 1 ) := e ± h / ( 2 cosh h ) , Jensen’s inequality gives A ^ ( β , h ) ≥ log ( 2 cosh h ) + β 2 N 2 ν h ⊗ N ( ∥ J T σ ∥ 2 ) \widehat{A}(\beta,h)\ge\log(2\cosh h)+\frac{\beta}{2N^2}\nu_h^{\otimes N}(\lVert J^{T}\sigma\rVert^2) A ( β , h ) ≥ log ( 2 cosh h ) + 2 N 2 β ν h ⊗ N (∥ J T σ ∥ 2 ) , and E ∥ ( J ∘ ) T σ ∥ 2 = K N \mathbb{E}\lVert(J^\circ)^T\sigma\rVert^2=KN E ∥( J ∘ ) T σ ∥ 2 = K N for every σ \sigma σ , since the entries are independent, centered and of unit variance. Finally, for fixed σ \sigma σ the variables X μ : = N − 1 / 2 ( J ∘ ) μ ⋅ σ X_\mu:=N^{-1/2}(J^\circ)^\mu\cdot\sigma X μ := N − 1/2 ( J ∘ ) μ ⋅ σ are independent and satisfy E e λ X μ ≤ e λ 2 / 2 \mathbb{E}e^{\lambda X_\mu}\le e^{\lambda^2/2} E e λ X μ ≤ e λ 2 /2 (with equality in the Gaussian case, and since cosh ( λ / N ) N ≤ e λ 2 / 2 \cosh(\lambda/\sqrt{N})^N\le e^{\lambda^2/2} cosh ( λ / N ) N ≤ e λ 2 /2 in the binary case). For β < 1 \beta<1 β < 1 , with Y Y Y an independent standard Gaussian, E e β X μ / 2 = E e β Y X μ ≤ E e β Y 2 / 2 = ( 1 − β ) − 1 / 2 \mathbb{E}e^{\beta X_\mu/2}=\mathbb{E}e^{\sqrt{\beta}Y X_\mu}\le\mathbb{E}e^{\beta Y^2/2}=(1-\beta)^{-1/2} E e β X μ /2 = E e β Y X μ ≤ E e β Y 2 /2 = ( 1 − β ) − 1/2 . Hence E Z N , K ( h ) ≤ ( 2 cosh h ) N ( 1 − β ) − K / 2 \mathbb{E}Z_{N,K}(h)\le(2\cosh h)^N(1-\beta)^{-K/2} E Z N , K ( h ) ≤ ( 2 cosh h ) N ( 1 − β ) − K /2 , and E log Z ≤ log E Z \mathbb{E}\log Z\le\log\mathbb{E}Z E log Z ≤ log E Z .
Vector spin glasses with compact spins
For Gaussian patterns, the required external convergence theorem concerns vector spin glasses with compactly supported spins and possibly non-convex covariance. After blocking, the Hubbard–Stratonovich representation of the Hopfield model takes this form, except that the hidden spins are unbounded; the truncation estimate of § 4.1 removes this obstruction. We state the result in the form used in § 4.2 .
For S , T ∈ R D × n S,T\in\mathbb{R}^{D\times n} S , T ∈ R D × n with columns S j , T j S_j,T_j S j , T j , let R n ( S , T ) : = n − 1 ∑ j = 1 n S j T j T R_n(S,T):=n^{-1}\sum_{j=1}^n S_jT_j^T R n ( S , T ) := n − 1 ∑ j = 1 n S j T j T , and let B D : = { x ∈ R D : ∥ x ∥ ≤ 1 } B_D:=\{x\in\mathbb{R}^D:\lVert x\rVert\le1\} B D := { x ∈ R D : ∥ x ∥ ≤ 1 } . We use [23 Theorem 1.1] with the auxiliary external parameter set to zero. Existence of the limit was first proved in [32 Theorem 1.1] for any compactly supported μ \mu μ , without the restriction to the unit ball used below; the limit is identified in [23 ] .
Theorem 3.6 (Ho [32 ] ; Chen–Ho–Issa [23 ] ). Let μ \mu μ be a probability measure on R D \mathbb{R}^D R D with μ ( B D ) = 1 \mu(B_D)=1 μ ( B D ) = 1 , and let ψ : R D × D → R \psi:\mathbb{R}^{D\times D}\to\mathbb{R} ψ : R D × D → R be given by an everywhere absolutely convergent power series with ψ ( 0 ) = 0 \psi(0)=0 ψ ( 0 ) = 0 . Assume that for each n n n there is a centered Gaussian process ( X n ( S ) ) S ∈ R D × n (X_n(S))_{S\in\mathbb{R}^{D\times n}} ( X n ( S ) ) S ∈ R D × n with E X n ( S ) X n ( T ) = n ψ ( R n ( S , T ) ) \mathbb{E}X_n(S)X_n(T)=n\psi(R_n(S,T)) E X n ( S ) X n ( T ) = n ψ ( R n ( S , T )) . Then for every t ≥ 0 t\ge0 t ≥ 0 the limit of
Φ n ( t ) : = 1 n E log ∫ exp { 2 t X n ( S ) − t n ψ ( R n ( S , S ) ) } μ ⊗ n ( d S ) (15) \Phi_n(t):=\frac{1}{n}\mathbb{E}\log\int\exp\left\{\sqrt{2t}X_n(S)-tn\psi(R_n(S,S))\right\}\,\mu^{\otimes n}(dS) \tag*{(15)} Φ n ( t ) := n 1 E log ∫ exp { 2 t X n ( S ) − t n ψ ( R n ( S , S )) } μ ⊗ n ( d S ) (15) exists and is finite. Convexity of ψ \psi ψ is not assumed.
In [23 ] the theorem concerns − Φ n ( t ) -\Phi_n(t) − Φ n ( t ) , which the authors call the free energy; they also give a variational formula for the limit. Here we use only its existence, at t = 1 2 t=\frac{1}{2} t = 2 1 .
Concentration on the discrete cube
For binary patterns, concentration on the discrete cube replaces the Gaussian inequalities (T1)–(T2). The relevant functions are convex in the disorder (Lemma 3.4), so three dimension-free estimates suffice. Lemma 3.7 bounds the variance and gives L 2 L^2 L 2 self-averaging for binary patterns. Lemma 3.8 bounds the upper tail and the expected maximum of finitely many, possibly dependent, convex functions; it controls the disorder-dependent choice of the retained patterns in the sector selection (§ 6.3 ). Lemma 3.9, based on Talagrand’s convex-distance inequality, gives two-sided exponential concentration, from which almost-sure convergence follows.
Let X X X be uniform on { − 1 , 1 } d \{-1,1\}^d { − 1 , 1 } d , and let x ( i ) x^{(i)} x ( i ) denote x x x with its i i i -th coordinate flipped.
Lemma 3.7 (Variance of convex functions). If φ : R d → R \varphi:\mathbb{R}^d\to\mathbb{R} φ : R d → R is convex and differentiable, then Var ( φ ( X ) ) ≤ 2 E ∥ ∇ φ ( X ) ∥ 2 \operatorname{Var}(\varphi(X))\le2\mathbb{E}\lVert\nabla\varphi(X)\rVert^2 Var ( φ ( X )) ≤ 2 E ∥ ∇ φ ( X ) ∥ 2 .
Proof. By the Efron–Stein inequality and invariance of the law of X X X under coordinate flips, Var φ ≤ 1 4 E ∑ i ( φ − φ ( i ) ) 2 = 1 2 E ∑ i ( φ − φ ( i ) ) + 2 \operatorname{Var}\varphi\le\frac14\mathbb{E}\sum_i(\varphi-\varphi^{(i)})^2=\frac12\mathbb{E}\sum_i(\varphi-\varphi^{(i)})_+^2 Var φ ≤ 4 1 E ∑ i ( φ − φ ( i ) ) 2 = 2 1 E ∑ i ( φ − φ ( i ) ) + 2 , where φ ( i ) : = φ ( X ( i ) ) \varphi^{(i)}:=\varphi(X^{(i)}) φ ( i ) := φ ( X ( i ) ) . Convexity gives φ ( x ( i ) ) ≥ φ ( x ) − 2 x i ∂ i φ ( x ) \varphi(x^{(i)})\ge\varphi(x)-2x_i\partial_i\varphi(x) φ ( x ( i ) ) ≥ φ ( x ) − 2 x i ∂ i φ ( x ) , so ( φ − φ ( i ) ) + ≤ 2 ∣ ∂ i φ ∣ (\varphi-\varphi^{(i)})_+\le2|\partial_i\varphi| ( φ − φ ( i ) ) + ≤ 2∣ ∂ i φ ∣ . □
Lemma 3.8 (Convex upper concentration). If φ : R d → R \varphi:\mathbb{R}^{d}\to\mathbb{R} φ : R d → R is differentiable and convex with ∥ ∇ φ ∥ ≤ L \lVert\nabla\varphi\rVert\le L ∥ ∇ φ ∥ ≤ L everywhere, then for $\lambda\ge 0
log E e λ ( φ ( X ) − E φ ( X ) ) ≤ 2 L 2 λ 2 . \log\mathbb{E}e^{\lambda(\varphi(X)-\mathbb{E}\varphi(X))}\le2L^{2}\lambda^{2}. log E e λ ( φ ( X ) − E φ ( X )) ≤ 2 L 2 λ 2 . Consequently, for m m m such functions φ 1 , … , φ m \varphi_{1},\ldots,\varphi_{m} φ 1 , … , φ m , not necessarily independent,
E max j ≤ m φ j ( X ) ≤ max j E φ j ( X ) + 2 2 L log ( m + 1 ) . \mathbb{E}\max_{j\le m}\varphi_{j}(X)\le\max_{j}\mathbb{E}\varphi_{j}(X)+2\sqrt{2}L\sqrt{\log(m+1)}. E j ≤ m max φ j ( X ) ≤ j max E φ j ( X ) + 2 2 L log ( m + 1 ) .
Proof. For a two-point uniform law, Jensen’s inequality gives Ent ( e g ) ≤ Cov ( e g , g ) \operatorname{Ent}(e^{g})\le\operatorname{Cov}(e^{g},g) Ent ( e g ) ≤ Cov ( e g , g ) . For g = λ φ g=\lambda\varphi g = λ φ with values λ a ≥ λ b \lambda a\ge\lambda b λa ≥ λb this covariance is 1 4 ( e λ a − e λ b ) λ ( a − b ) ≤ 1 4 λ 2 e λ a ( a − b ) 2 \frac{1}{4}(e^{\lambda a}-e^{\lambda b})\lambda(a-b)\le\frac{1}{4}\lambda^{2}e^{\lambda a}(a-b)^{2} 4 1 ( e λa − e λb ) λ ( a − b ) ≤ 4 1 λ 2 e λa ( a − b ) 2 , which equals λ 2 2 \frac{\lambda^{2}}{2} 2 λ 2 times the conditional expectation of e λ φ ( φ − φ ( i ) ) + 2 e^{\lambda\varphi}(\varphi-\varphi^{(i)})_{+}^{2} e λ φ ( φ − φ ( i ) ) + 2 . Tensorization of entropy [15 ] gives Ent ( e λ φ ) ≤ λ 2 2 E e λ φ ∑ i ( φ − φ ( i ) ) + 2 ≤ 2 λ 2 L 2 E e λ φ \operatorname{Ent}(e^{\lambda\varphi})\le\frac{\lambda^{2}}{2}\mathbb{E}e^{\lambda\varphi}\sum_{i}(\varphi-\varphi^{(i)})_{+}^{2}\le2\lambda^{2}L^{2}\mathbb{E}e^{\lambda\varphi} Ent ( e λ φ ) ≤ 2 λ 2 E e λ φ ∑ i ( φ − φ ( i ) ) + 2 ≤ 2 λ 2 L 2 E e λ φ , by the convexity bound of Lemma 3.7. With Λ ( λ ) : = log E e λ ( φ − E φ ) \Lambda(\lambda):=\log\mathbb{E}e^{\lambda(\varphi-\mathbb{E}\varphi)} Λ ( λ ) := log E e λ ( φ − E φ ) this reads ( λ Λ ′ − Λ ) / λ 2 ≤ 2 L 2 (\lambda\Lambda'-\Lambda)/\lambda^{2}\le2L^{2} ( λ Λ ′ − Λ ) / λ 2 ≤ 2 L 2 ; integrating ( Λ / λ ) ′ (\Lambda/\lambda)' ( Λ/ λ ) ′ from 0 0 0 proves the first claim. For the second, E max j ( φ j − E φ j ) ≤ λ − 1 log ∑ j E e λ ( φ j − E φ j ) ≤ λ − 1 log m + 2 L 2 λ \mathbb{E}\max_{j}(\varphi_{j}-\mathbb{E}\varphi_{j})\le\lambda^{-1}\log\sum_{j}\mathbb{E}e^{\lambda(\varphi_{j}-\mathbb{E}\varphi_{j})}\le\lambda^{-1}\log m+2L^{2}\lambda E max j ( φ j − E φ j ) ≤ λ − 1 log ∑ j E e λ ( φ j − E φ j ) ≤ λ − 1 log m + 2 L 2 λ ; optimize in λ \lambda λ .
Lemma 3.9 (Two-sided convex concentration). If φ : R d → R \varphi:\mathbb{R}^{d}\to\mathbb{R} φ : R d → R is convex and L L L -Lipschitz, L > 0 L>0 L > 0 , and m φ m_{\varphi} m φ is a median of φ ( X ) \varphi(X) φ ( X ) , then for t > 0 t>0 t > 0
P ( ∣ φ ( X ) − m φ ∣ > t ) ≤ 4 exp { − t 2 / ( 16 L 2 ) } , ∣ E φ ( X ) − m φ ∣ ≤ 8 π L . (16) \mathbb{P}(|\varphi(X)-m_{\varphi}|>t)\le4\exp\{-t^{2}/(16L^{2})\},\qquad|\mathbb{E}\varphi(X)-m_{\varphi}|\le8\sqrt{\pi}L. \tag*{(16)} P ( ∣ φ ( X ) − m φ ∣ > t ) ≤ 4 exp { − t 2 / ( 16 L 2 )} , ∣ E φ ( X ) − m φ ∣ ≤ 8 π L . (16)
Proof. For B ⊆ { − 1 , 1 } d B\subseteq\{-1,1\}^{d} B ⊆ { − 1 , 1 } d let d T ( x , B ) : = sup { inf y ∈ B ∑ i a i 1 { x i ≠ y i } : a i ≥ 0 , ∑ i a i 2 ≤ 1 } d_{T}(x,B):=\sup\{\inf_{y\in B}\sum_{i}a_{i}\mathbf{1}_{\{x_{i}\ne y_{i}\}}:a_{i}\ge0,\ \sum_{i}a_{i}^{2}\le1\} d T ( x , B ) := sup { inf y ∈ B ∑ i a i 1 { x i = y i } : a i ≥ 0 , ∑ i a i 2 ≤ 1 } . Talagrand’s convex-distance inequality [48 ] states P ( B ) P ( d T ( X , B ) ≥ u ) ≤ e − u 2 / 4 \mathbb{P}(B)\mathbb{P}(d_{T}(X,B)\ge u)\le e^{-u^{2}/4} P ( B ) P ( d T ( X , B ) ≥ u ) ≤ e − u 2 /4 . Every subgradient v x v_{x} v x of φ \varphi φ at x x x has norm at most L L L , and φ ( x ) − φ ( y ) ≤ v x ⋅ ( x − y ) ≤ 2 ∑ i ∣ v x , i ∣ 1 { x i ≠ y i } \varphi(x)-\varphi(y)\le v_{x}\cdot(x-y)\le2\sum_{i}|v_{x,i}|\mathbf{1}_{\{x_{i}\ne y_{i}\}} φ ( x ) − φ ( y ) ≤ v x ⋅ ( x − y ) ≤ 2 ∑ i ∣ v x , i ∣ 1 { x i = y i } ; hence, if φ ( x ) − φ ( y ) ≥ t \varphi(x)-\varphi(y)\ge t φ ( x ) − φ ( y ) ≥ t for every y ∈ B y\in B y ∈ B , then d T ( x , B ) ≥ t / ( 2 L ) d_{T}(x,B)\ge t/(2L) d T ( x , B ) ≥ t / ( 2 L ) . Apply the inequality with B = { φ ≤ m φ } B=\{\varphi\le m_{\varphi}\} B = { φ ≤ m φ } for the upper tail, and with B = { φ ≤ m φ − t } B=\{\varphi\le m_{\varphi}-t\} B = { φ ≤ m φ − t } , using P ( φ ≥ m φ ) ≥ 1 2 \mathbb{P}(\varphi\ge m_{\varphi})\ge\frac{1}{2} P ( φ ≥ m φ ) ≥ 2 1 , for the lower tail. Each tail is at most 2 e − t 2 / ( 16 L 2 ) 2e^{-t^{2}/(16L^{2})} 2 e − t 2 / ( 16 L 2 ) , and integrating gives E ∣ φ − m φ ∣ ≤ 8 π L \mathbb{E}|\varphi-m_{\varphi}|\le8\sqrt{\pi}L E ∣ φ − m φ ∣ ≤ 8 π L .
Gaussian bulks with finitely many site types
We first control the unbounded Gaussian hidden spins, then apply the compact-spin theorem to a block representation. Field, volume and load comparisons extend the resulting limit beyond the block sequences.
Uniform truncation of the hidden spins
Let V : Σ N → R V:\Sigma_{N}\to\mathbb{R} V : Σ N → R be arbitrary. Write ω G V \omega_{G}^{V} ω G V for the Gibbs measure on Σ N \Sigma_{N} Σ N with weights e − β H N ( σ ; G ) + V ( σ ) e^{-\beta H_{N}(\sigma;G)+V(\sigma)} e − β H N ( σ ; G ) + V ( σ ) , let P V ( d G , σ ) : = γ ⊗ N K ( d G ) ω G V ( σ ) P_{V}(dG,\sigma):=\gamma^{\otimes NK}(dG)\omega_{G}^{V}(\sigma) P V ( d G , σ ) := γ ⊗ N K ( d G ) ω G V ( σ ) be the joint law of disorder and spins, E G ω V \mathbb{E}_{G}^{\omega_{V}} E G ω V the corresponding expectation, and set a : = β / N G T σ ∈ R K a:=\sqrt{\beta/N}G^{T}\sigma\in\mathbb{R}^{K} a := β / N G T σ ∈ R K . Since ∫ R K e x ⋅ z γ ⊗ K ( d z ) = e ∥ x ∥ 2 / 2 \int_{\mathbb{R}^{K}}e^{x\cdot z}\gamma^{\otimes K}(dz)=e^{\lVert x\rVert^{2}/2} ∫ R K e x ⋅ z γ ⊗ K ( d z ) = e ∥ x ∥ 2 /2 and ∥ a ∥ 2 / 2 = − β H N ( σ ; G ) \lVert a\rVert^{2}/2=-\beta H_{N}(\sigma;G) ∥ a ∥ 2 /2 = − β H N ( σ ; G ) (Hubbard–Stratonovich),
Z V : = ∑ σ e − β H N ( σ ; G ) + V ( σ ) = ∑ σ e V ( σ ) ∫ R K e β / N σ T G z γ ⊗ K ( d z ) . (17) Z^{V}:=\sum_{\sigma}e^{-\beta H_{N}(\sigma;G)+V(\sigma)}
=\sum_{\sigma}e^{V(\sigma)}\int_{\mathbb{R}^{K}}e^{\sqrt{\beta/N}\,\sigma^{T}Gz}\gamma^{\otimes K}(dz). \tag*{(17)} Z V := σ ∑ e − β H N ( σ ; G ) + V ( σ ) = σ ∑ e V ( σ ) ∫ R K e β / N σ T G z γ ⊗ K ( d z ) . (17)
For L > 0 L>0 L > 0 let Z V , [ L ] Z^{V,[L]} Z V , [ L ] be the same expression with the integral over [ − L , L ] K [-L,L]^{K} [ − L , L ] K , without renormalizing the Gaussian measure. If V V V is random and independent of G G G , all statements below hold conditionally on V V V , and for the unconditional expectations whenever E max σ ∣ V ( σ ) ∣ < ∞ \mathbb{E}\max_{\sigma}|V(\sigma)|<\infty E max σ ∣ V ( σ ) ∣ < ∞ .
Lemma 4.1 (Rotational invariance). Let N , K ≥ 1 N,K\ge1 N , K ≥ 1 . Under P V P^{V} P V the law of a a a is O ( K ) O(K) O ( K ) -invariant, hence exchangeable, and E G ω V ( a 1 4 ) = 3 K ( K + 2 ) E G ω V ( ∥ a ∥ 4 ) ≤ 3 β 2 K ( K + 2 ) E ∥ G ∥ o p 4 \mathbb{E}_{G}^{\omega_{V}}(a_{1}^{4})=\frac{3}{K(K+2)}\mathbb{E}_{G}^{\omega_{V}}(\lVert a\rVert^{4})\le\frac{3\beta^{2}}{K(K+2)}\mathbb{E}\lVert G\rVert_{\mathrm{op}}^{4} E G ω V ( a 1 4 ) = K ( K + 2 ) 3 E G ω V (∥ a ∥ 4 ) ≤ K ( K + 2 ) 3 β 2 E ∥ G ∥ op 4 . If K ≥ α − N K\ge\alpha_{-}N K ≥ α − N , then E G ω V ( a 1 4 ) ≤ M 4 : = 480 β 2 ( 1 + α − − 1 ) 2 \mathbb{E}_{G}^{\omega_{V}}(a_{1}^{4})\le M_{4}:=480\beta^{2}(1+\alpha_{-}^{-1})^{2} E G ω V ( a 1 4 ) ≤ M 4 := 480 β 2 ( 1 + α − − 1 ) 2 , uniformly in N , K , V N,K,V N , K , V .
Proof. For O ∈ O ( K ) O \in O(K) O ∈ O ( K ) , the map G ↦ G O G \mapsto GO G ↦ GO preserves γ ⊗ N K \gamma^{\otimes NK} γ ⊗ N K (the rows of G G G are i.i.d. standard Gaussian in R K \mathbb{R}^{K} R K ) and H N H_N H N ; since V V V depends only on σ \sigma σ , it preserves ω G V \omega_G^V ω G V and hence P V P^V P V , while a ↦ O T a a \mapsto O^T a a ↦ O T a . If O O O is Haar-distributed and independent, a 1 a_1 a 1 has the law of a ⋅ U a \cdot U a ⋅ U with U = O e 1 U = Oe_1 U = O e 1 uniform on the sphere and independent of a a a , and E U 1 4 = 3 / ( K ( K + 2 ) ) \mathbb{E}U_1^4 = 3/(K(K+2)) E U 1 4 = 3/ ( K ( K + 2 )) . Finally ∥ a ∥ 2 ≤ β ∥ G ∥ o p 2 \lVert a\rVert^2 \le\beta\lVert G\rVert_{\mathrm{op}}^2 ∥ a ∥ 2 ≤ β ∥ G ∥ op 2 , and Lemma 3.2 gives
3 β 2 K 2 160 ( N + K ) 2 ≤ M 4 . \frac{3\beta^2}{K^2}160(N+K)^2 \le M_4. K 2 3 β 2 160 ( N + K ) 2 ≤ M 4 .
Proposition 4.2 (Uniform truncation). Let 0 < α − ≤ α + 0 < \alpha_- \le\alpha_+ 0 < α − ≤ α + and C c u t : = α + ( 32 / e + 6 M 4 ) C_{\mathrm{cut}} := \alpha_+(32/e+6M_4) C cut := α + ( 32/ e + 6 M 4 ) . For all L ≥ 4 L \ge4 L ≥ 4 , N , K ≥ 1 N,K \ge1 N , K ≥ 1 with K / N ∈ [ α − , α + ] K/N \in[\alpha_-,\alpha_+] K / N ∈ [ α − , α + ] , and all V V V ,
0 ≤ 1 N E log Z V − 1 N E log Z V , [ L ] ≤ C c u t L 2 . (18) 0 \le\frac{1}{N}\mathbb{E}\log Z^V-\frac{1}{N}\mathbb{E}\log Z^{V,[L]} \le\frac{C_{\mathrm{cut}}}{L^2}. \tag*{(18)} 0 ≤ N 1 E log Z V − N 1 E log Z V , [ L ] ≤ L 2 C cut . (18)
Proof. Let π L ( v ) : = P ( ∣ Y + v ∣ ≤ L ) \pi_L(v) := \mathbb{P}(|Y+v|\le L) π L ( v ) := P ( ∣ Y + v ∣ ≤ L ) and f L : = − log π L ≥ 0 f_L := -\log\pi_L \ge0 f L := − log π L ≥ 0 . Since ∫ − L L e x z γ ( d z ) = e x 2 / 2 π L ( x ) \int_{-L}^{L}e^{xz}\gamma(dz)=e^{x^2/2}\pi_L(x) ∫ − L L e x z γ ( d z ) = e x 2 /2 π L ( x ) ,
Z G V , [ L ] Z G V = ω G V ( exp { − ∑ μ ≤ K f L ( a μ ) } ) ∈ ( 0 , 1 ] . \frac{Z_G^{V,[L]}}{Z_G^V}=\omega_G^V\left(\exp\left\{-\sum_{\mu\le K}f_L(a_\mu)\right\}\right)\in(0,1]. Z G V Z G V , [ L ] = ω G V exp ⎩ ⎨ ⎧ − μ ≤ K ∑ f L ( a μ ) ⎭ ⎬ ⎫ ∈ ( 0 , 1 ] .
The lower bound follows. For the upper bound, Jensen’s inequality and exchangeability give
1 N E log ( Z V Z V , [ L ] ) ≤ K N E ω G V ( f L ( a 1 ) ) . \frac{1}{N}\mathbb{E}\log\left(\frac{Z^V}{Z^{V,[L]}}\right)\le\frac{K}{N}\mathbb{E}\omega_G^V(f_L(a_1)). N 1 E log ( Z V , [ L ] Z V ) ≤ N K E ω G V ( f L ( a 1 )) .
If ∣ v ∣ ≤ L / 2 |v|\le L/2 ∣ v ∣ ≤ L /2 , then 1 − π L ( v ) ≤ P ( ∣ Y ∣ > L / 2 ) ≤ 2 e − L 2 / 8 ≤ 2 e − 2 1-\pi_L(v)\le\mathbb{P}(|Y|>L/2)\le2e^{-L^2/8}\le2e^{-2} 1 − π L ( v ) ≤ P ( ∣ Y ∣ > L /2 ) ≤ 2 e − L 2 /8 ≤ 2 e − 2 , and − log ( 1 − x ) ≤ 2 x -\log(1-x)\le2x − log ( 1 − x ) ≤ 2 x on [ 0 , 1 2 ] [0,\frac{1}{2}] [ 0 , 2 1 ] gives f L ( v ) ≤ 4 e − L 2 / 8 f_L(v)\le4e^{-L^2/8} f L ( v ) ≤ 4 e − L 2 /8 . For all v v v , π L ( v ) ≥ ∫ − 1 1 e − ( z − v ) 2 / 2 2 π d z ≥ 2 2 π e − ( ∣ v ∣ + 1 ) 2 / 2 \pi_L(v)\ge\int_{-1}^{1}\frac{e^{-(z-v)^2/2}}{\sqrt{2\pi}}\,dz\ge\frac{2}{\sqrt{2\pi}}e^{-(|v|+1)^2/2} π L ( v ) ≥ ∫ − 1 1 2 π e − ( z − v ) 2 /2 d z ≥ 2 π 2 e − ( ∣ v ∣ + 1 ) 2 /2 , so f L ( v ) ≤ v 2 + 2 f_L(v)\le v^2+2 f L ( v ) ≤ v 2 + 2 . By Markov’s inequality and Lemma 4.1 ,
E ω G V ( f L ( a 1 ) ) ≤ 4 e − L 2 / 8 + E ω G V [ ( a 1 2 + 2 ) 1 { ∣ a 1 ∣ > L / 2 } ] ≤ 4 e − L 2 / 8 + ( 4 L 2 + 32 L 4 ) M 4 ≤ 32 / e + 6 M 4 L 2 , \mathbb{E}\omega_G^V(f_L(a_1))\le4e^{-L^2/8}+\mathbb{E}\omega_G^V[(a_1^2+2)\mathbf{1}_{\{|a_1|>L/2\}}]\le4e^{-L^2/8}+\left(\frac{4}{L^2}+\frac{32}{L^4}\right)M_4\le\frac{32/e+6M_4}{L^2}, E ω G V ( f L ( a 1 )) ≤ 4 e − L 2 /8 + E ω G V [( a 1 2 + 2 ) 1 { ∣ a 1 ∣ > L /2 } ] ≤ 4 e − L 2 /8 + ( L 2 4 + L 4 32 ) M 4 ≤ L 2 32/ e + 6 M 4 ,
using e − x ≤ ( e x ) − 1 e^{-x}\le(ex)^{-1} e − x ≤ ( e x ) − 1 and L ≥ 4 L\ge4 L ≥ 4 .
Site fields and blocking
For a field vector h ⃗ = ( h 1 , … , h N ) ∈ R N \vec{h}=(h_1,\ldots,h_N)\in\mathbb{R}^N h = ( h 1 , … , h N ) ∈ R N let
B N , K ( h ⃗ ) : = 1 N E log ∑ σ ∈ Σ N exp { β 2 N ∥ G T σ ∥ 2 + ∑ i = 1 N h i σ i } = E f N V h ⃗ ( G ) , V h ⃗ ( σ ) : = ∑ i h i σ i . (19) B_{N,K}(\vec{h}):=\frac{1}{N}\mathbb{E}\log\sum_{\sigma\in\Sigma_N}\exp\left\{\frac{\beta}{2N}\lVert G^T\sigma\rVert^2+\sum_{i=1}^{N}h_i\sigma_i\right\}=\mathbb{E}f_N^{V_{\vec{h}}}(G),\qquad V_{\vec{h}}(\sigma):=\sum_i h_i\sigma_i. \tag*{(19)} B N , K ( h ) := N 1 E log σ ∈ Σ N ∑ exp { 2 N β ∥ G T σ ∥ 2 + i = 1 ∑ N h i σ i } = E f N V h ( G ) , V h ( σ ) := i ∑ h i σ i . (19)
Fix D ≥ 1 D\ge1 D ≥ 1 and fields H = ( H 1 , … , H D ) ∈ R D H=(H_1,\ldots,H_D)\in\mathbb{R}^D H = ( H 1 , … , H D ) ∈ R D . A type assignment is a map t : { 1 , … , N } → { 1 , … , D } t:\{1,\ldots,N\}\to\{1,\ldots,D\} t : { 1 , … , N } → { 1 , … , D } , and H t : = ( H t 1 , … , H t N ) H_t:=(H_{t_1},\ldots,H_{t_N}) H t := ( H t 1 , … , H t N ) . With D = 1 D=1 D = 1 and H 1 = h H_1=h H 1 = h , B N , K ( H t ) = A N , K G ( β , h ) B_{N,K}(H_t)=A_{N,K}^G(\beta,h) B N , K ( H t ) = A N , K G ( β , h ) .
Fix a , b ∈ N a,b\in\mathbb{N} a , b ∈ N , D : = a + b D:=a+b D := a + b , a map τ : { 1 , … , b } → { 1 , … , D } \tau:\{1,\ldots,b\}\to\{1,\ldots,D\} τ : { 1 , … , b } → { 1 , … , D } , and let N = b n N=bn N = bn , K = a n K=an K = an . Relabel σ i = σ j , u \sigma_i=\sigma_{j,u} σ i = σ j , u (i = ( j − 1 ) b + u i=(j-1)b+u i = ( j − 1 ) b + u , u ≤ b u\le b u ≤ b ), z μ = z k , v z_\mu=z_{k,v} z μ = z k , v (μ = ( k − 1 ) a + v \mu=(k-1)a+v μ = ( k − 1 ) a + v , v ≤ a v\le a v ≤ a ) and g ( j , u ) , ( k , v ) : = g j i g_{(j,u),(k,v)}:=g_{j}^{i} g ( j , u ) , ( k , v ) := g j i , and give site ( j , u ) (j,u) ( j , u ) the type τ ( u ) \tau(u) τ ( u ) ; we call this the block assignment . For S ∈ R D × n S\in\mathbb{R}^{D\times n} S ∈ R D × n with columns S j = ( x j , 1 , … , x j , b , y j , 1 , … , y j , a ) S_j=(x_{j,1},\ldots,x_{j,b},y_{j,1},\ldots,y_{j,a}) S j = ( x j , 1 , … , x j , b , y j , 1 , … , y j , a ) define
X n ( S ) : = β b n ∑ j , k ≤ n ∑ u ≤ b , v ≤ a g ( j , u ) , ( k , v ) x j , u y k , v , Ψ ( R ) : = β b ( ∑ u ≤ b R u u ) ( ∑ v = b + 1 D R v v ) . (20) X_n(S):=\sqrt{\frac{\beta}{bn}}\sum_{j,k\le n}\sum_{u\le b,\,v\le a}g_{(j,u),(k,v)}x_{j,u}y_{k,v},\qquad\Psi(R):=\frac{\beta}{b}\left(\sum_{u\le b}R_{uu}\right)\left(\sum_{v=b+1}^{D}R_{vv}\right). \tag*{(20)} X n ( S ) := bn β j , k ≤ n ∑ u ≤ b , v ≤ a ∑ g ( j , u ) , ( k , v ) x j , u y k , v , Ψ ( R ) := b β ( u ≤ b ∑ R uu ) ( v = b + 1 ∑ D R v v ) . (20)
Then X n X_n X n is a centered Gaussian process on all of R D × n \mathbb{R}^{D\times n} R D × n with E X n ( S ) X n ( T ) = n Ψ ( R n ( S , T ) ) \mathbb{E}X_n(S)X_n(T)=n\Psi(R_n(S,T)) E X n ( S ) X n ( T ) = n Ψ ( R n ( S , T )) , since
E X n ( S ) X n ( T ) = β b n ( ∑ j , u x j , u x j , u ′ ) ( ∑ k , v y k , v y k , v ′ ) . \mathbb{E}X_n(S)X_n(T)=\frac{\beta}{bn}\left(\sum_{j,u}x_{j,u}x'_{j,u}\right)\left(\sum_{k,v}y_{k,v}y'_{k,v}\right). E X n ( S ) X n ( T ) = bn β ( j , u ∑ x j , u x j , u ′ ) k , v ∑ y k , v y k , v ′ .
The polynomial Ψ \Psi Ψ is not convex: on diag ( s I b , t I a ) \operatorname{diag}(sI_b,tI_a) diag ( s I b , t I a ) it equals β a s t \beta a st β a s t . Let ν h \nu_h ν h be as in Lemma 3.5 and
λ L ( d z ) : = 1 [ − L , L ] ( z ) e β z 2 / 2 d L γ ( d z ) , d L : = ∫ − L L e β z 2 / 2 γ ( d z ) , μ L : = ( ⨂ u ≤ b ν H τ ( u ) ) ⊗ λ L ⊗ a , \lambda_L(dz):=\frac{\mathbf{1}_{[-L,L]}(z)e^{\beta z^2/2}}{d_L}\gamma(dz),\qquad d_L:=\int_{-L}^{L}e^{\beta z^2/2}\gamma(dz),\qquad\mu_L:=\left(\bigotimes_{u\le b}\nu_{H_{\tau(u)}}\right)\otimes\lambda_L^{\otimes a}, λ L ( d z ) := d L 1 [ − L , L ] ( z ) e β z 2 /2 γ ( d z ) , d L := ∫ − L L e β z 2 /2 γ ( d z ) , μ L := ( u ≤ b ⨂ ν H τ ( u ) ) ⊗ λ L ⊗ a ,
a probability measure on R D \mathbb{R}^D R D supported in the ball of radius c : = b + a L 2 c:=\sqrt{b+aL^2} c := b + a L 2 , for every β ≥ 0 \beta\ge0 β ≥ 0 . (Without truncation, e β z 2 / 2 γ ( d z ) e^{\beta z^2/2}\gamma(dz) e β z 2 /2 γ ( d z ) has infinite mass for β ≥ 1 \beta\ge1 β ≥ 1 .) Let Z [ L ] Z^{[L]} Z [ L ] denote the truncated partition function (17) with V = V H t V=V_{H_t} V = V H t for the block assignment.
Lemma 4.3. For the block assignment,
Z [ L ] = ∏ u ≤ b ( 2 cosh H τ ( u ) ) n d L a n ∫ exp { X n ( S ) − n 2 Ψ ( R n ( S , S ) ) } μ L ⊗ n ( d S ) . (21) Z^{[L]}=\prod_{u\le b}(2\cosh H_{\tau(u)})^n d_L^{an}\int\exp\left\{X_n(S)-\frac{n}{2}\Psi(R_n(S,S))\right\}\mu_L^{\otimes n}(dS). \tag*{(21)} Z [ L ] = u ≤ b ∏ ( 2 cosh H τ ( u ) ) n d L an ∫ exp { X n ( S ) − 2 n Ψ ( R n ( S , S )) } μ L ⊗ n ( d S ) . (21)
Proof. When ( σ , z ) (\sigma,z) ( σ , z ) is arranged into S S S as above, X n ( S ) = β / N σ T G z X_n(S)=\sqrt{\beta/N}\,\sigma^{T}Gz X n ( S ) = β / N σ T G z . If all x j , u ∈ { ± 1 } x_{j,u}\in\{\pm1\} x j , u ∈ { ± 1 } , as holds μ L ⊗ n \mu_L^{\otimes n} μ L ⊗ n -a.s., then ∑ u R n ( S , S ) u u = b \sum_u R_n(S,S)_{uu}=b ∑ u R n ( S , S ) uu = b , so that n 2 Ψ ( R n ( S , S ) ) = β 2 ∑ k , v z k , v 2 \frac{n}{2}\Psi(R_n(S,S))=\frac{\beta}{2}\sum_{k,v}z_{k,v}^2 2 n Ψ ( R n ( S , S )) = 2 β ∑ k , v z k , v 2 . The claim follows from 1 [ − L , L ] ( z ) γ ( d z ) = d L e − β z 2 / 2 λ L ( d z ) \mathbf{1}_{[-L,L]}(z)\gamma(dz)=d_L e^{-\beta z^2/2}\lambda_L(dz) 1 [ − L , L ] ( z ) γ ( d z ) = d L e − β z 2 /2 λ L ( d z ) and ∑ s = ± 1 e h s φ ( s ) = 2 cosh h ν h ( φ ) \sum_{s=\pm1}e^{hs}\varphi(s)=2\cosh h\,\nu_h(\varphi) ∑ s = ± 1 e h s φ ( s ) = 2 cosh h ν h ( φ ) , applied at every visible site. □
Proposition 4.4. For the block assignment and all a , b ∈ N a,b\in\mathbb{N} a , b ∈ N , τ \tau τ , H H H , β ≥ 0 \beta\ge0 β ≥ 0 , the sequence B b n , a n ( H t ) B_{bn,an}(H_t) B bn , an ( H t ) converges as n → ∞ n\to\infty n → ∞ .
Proof. Fix L ≥ 4 L\ge4 L ≥ 4 . Let μ ~ \widetilde{\mu} μ be the image of μ L \mu_L μ L under x ↦ x / c x\mapsto x/c x ↦ x / c , Ψ ~ ( R ) : = Ψ ( c 2 R ) \widetilde{\Psi}(R):=\Psi(c^2R) Ψ ( R ) := Ψ ( c 2 R ) and X ~ n ( S ) : = X n ( c S ) \widetilde{X}_n(S):=X_n(cS) X n ( S ) := X n ( c S ) . Then μ ~ ( B D ) = 1 \widetilde{\mu}(B_D)=1 μ ( B D ) = 1 , Ψ ~ \widetilde{\Psi} Ψ is a polynomial with Ψ ~ ( 0 ) = 0 \widetilde{\Psi}(0)=0 Ψ ( 0 ) = 0 , and X ~ n \widetilde{X}_n X n is a centered Gaussian process with covariance n Ψ ~ ( R n ) n\widetilde{\Psi}(R_n) n Ψ ( R n ) on all of R D × n \mathbb{R}^{D\times n} R D × n . Since X n ( S ) − n 2 Ψ ( R n ( S , S ) ) = X ~ n ( S / c ) − n 2 Ψ ~ ( R n ( S / c , S / c ) ) X_n(S)-\frac{n}{2}\Psi(R_n(S,S))=\widetilde{X}_n(S/c)-\frac{n}{2}\widetilde{\Psi}(R_n(S/c,S/c)) X n ( S ) − 2 n Ψ ( R n ( S , S )) = X n ( S / c ) − 2 n Ψ ( R n ( S / c , S / c )) and t = 1 2 t=\frac{1}{2} t = 2 1 gives 2 t = 1 \sqrt{2t}=1 2 t = 1 , (21) yields
1 b n E log Z [ L ] = 1 b ∑ u ≤ b log ( 2 cosh H τ ( u ) ) + a b log d L + 1 b Φ n ( 1 / 2 ) , \frac{1}{bn}\mathbb{E}\log Z^{[L]}=\frac{1}{b}\sum_{u\le b}\log(2\cosh H_{\tau(u)})+\frac{a}{b}\log d_L+\frac{1}{b}\Phi_n(1/2), bn 1 E log Z [ L ] = b 1 u ≤ b ∑ log ( 2 cosh H τ ( u ) ) + b a log d L + b 1 Φ n ( 1/2 ) ,
with Φ n \Phi_n Φ n as in (15) for ( μ ~ , Ψ ~ , X ~ n ) (\widetilde{\mu},\widetilde{\Psi},\widetilde{X}_n) ( μ , Ψ , X n ) . By Theorem 3.6 the limit ℓ L \ell_L ℓ L of the left-hand side exists. By Proposition 4.2 with α ± = a / b \alpha_\pm=a/b α ± = a / b , ℓ L ≤ lim inf n B b n , a n ≤ lim sup n B b n , a n ≤ ℓ L + C c u t L − 2 \ell_L\le\liminf_n B_{bn,an}\le\limsup_n B_{bn,an}\le\ell_L+C_{\mathrm{cut}}L^{-2} ℓ L ≤ lim inf n B bn , an ≤ lim sup n B bn , an ≤ ℓ L + C cut L − 2 . Let L → ∞ L\to\infty L → ∞ . Theorem 3.6 is applied at fixed L L L ; the exchange of limits relies only on (18) . □
Comparison of fields, volumes and loads
Lemma 4.5 (Fields). For a permutation π \pi π of { 1 , … , N } \{1,\ldots,N\} { 1 , … , N } , B N , K ( h ⃗ ∘ π ) = B N , K ( h ⃗ ) B_{N,K}(\vec{h}\circ\pi)=B_{N,K}(\vec{h}) B N , K ( h ∘ π ) = B N , K ( h ) . For h ⃗ , h ⃗ ′ ∈ R N \vec{h},\vec{h}^{\,\prime}\in\mathbb{R}^{N} h , h ′ ∈ R N , pointwise in G G G , ∣ f N V h ⃗ ( G ) − f N V h ⃗ ′ ( G ) ∣ ≤ N − 1 ∑ i ∣ h i − h i ′ ∣ |f_N^{V_{\vec{h}}}(G)-f_N^{V_{\vec{h}^{\,\prime}}}(G)|\le N^{-1}\sum_i|h_i-h_i^{\prime}| ∣ f N V h ( G ) − f N V h ′ ( G ) ∣ ≤ N − 1 ∑ i ∣ h i − h i ′ ∣ .
Proof. The rows of G G G are i.i.d., so permuting them leaves the law of G G G unchanged; relabelling σ \sigma σ accordingly proves the first claim. The exponents differ by at most ∑ i ∣ h i − h i ′ ∣ \sum_i|h_i-h_i^{\prime}| ∑ i ∣ h i − h i ′ ∣ uniformly in σ \sigma σ . □
Lemma 4.6 (Loads). Let V : Σ N → R V:\Sigma_N\to\mathbb{R} V : Σ N → R and 0 ≤ K − ≤ K + 0\le K_-\le K_+ 0 ≤ K − ≤ K + , K + ≥ 1 K_+\ge1 K + ≥ 1 , and let G ± G_\pm G ± be standard Gaussian N × K ± N\times K_\pm N × K ± matrices. Then 0 ≤ E f N V ( G + ) − E f N V ( G − ) ≤ β 2 K + − K − K + ( ( 1 + K + / N ) 2 + 1 N ) 0\le\mathbb{E}f_N^V(G_+)-\mathbb{E}f_N^V(G_-)\le\frac{\beta}{2}\frac{K_+-K_-}{K_+}\left((1+\sqrt{K_+/N})^2+\frac{1}{N}\right) 0 ≤ E f N V ( G + ) − E f N V ( G − ) ≤ 2 β K + K + − K − ( ( 1 + K + / N ) 2 + N 1 ) . In particular, if α − N ≤ K + ≤ α + N \alpha_-N\le K_+\le\alpha_+N α − N ≤ K + ≤ α + N , the difference is at most c l o a d ( K + − K − ) / N c_{\mathrm{load}}(K_+-K_-)/N c load ( K + − K − ) / N with c l o a d : = β [ ( 1 + α + ) 2 + 1 ] / ( 2 α − ) c_{\mathrm{load}}:=\beta[(1+\sqrt{\alpha_+})^2+1]/(2\alpha_-) c load := β [( 1 + α + ) 2 + 1 ] / ( 2 α − ) .
Proof. Realize G − G_- G − as the first K − K_- K − columns of G + G_+ G + and let Δ : = β 2 N ∑ u > K − ( g u ⋅ σ ) 2 ≥ 0 \Delta:=\frac{\beta}{2N}\sum_{u>K_-}(g^u\cdot\sigma)^2\ge0 Δ := 2 N β ∑ u > K − ( g u ⋅ σ ) 2 ≥ 0 . With ω + \omega_+ ω + the Gibbs measure for G + G_+ G + , Jensen’s inequality gives 0 ≤ log Z + − log Z − = − log ω + ( e − Δ ) ≤ ω + ( Δ ) 0\le\log Z_+-\log Z_-=-\log\omega_+(e^{-\Delta})\le\omega_+(\Delta) 0 ≤ log Z + − log Z − = − log ω + ( e − Δ ) ≤ ω + ( Δ ) . By column exchangeability (Lemma 4.1 ), E ω + ( Δ ) = K + − K − K + β 2 N E ω + ∥ G + T σ ∥ 2 ≤ K + − K − K + β 2 N E ∥ G + ∥ o p 2 \mathbb{E}\omega_+(\Delta)=\frac{K_+-K_-}{K_+}\frac{\beta}{2N}\mathbb{E}\omega_+\|G_+^T\sigma\|^2\le\frac{K_+-K_-}{K_+}\frac{\beta}{2N}\mathbb{E}\|G_+\|_{\mathrm{op}}^2 E ω + ( Δ ) = K + K + − K − 2 N β E ω + ∥ G + T σ ∥ 2 ≤ K + K + − K − 2 N β E ∥ G + ∥ op 2 ; divide by N N N and apply Lemma 3.2 . □
Lemma 4.7 (Volumes). Let d 0 ∈ N d_0\in\mathbb{N} d 0 ∈ N , α + > 0 \alpha_+>0 α + > 0 . There is C p a d ( β , α + , d 0 ) C_{\mathrm{pad}}(\beta,\alpha_+,d_0) C pad ( β , α + , d 0 ) such that for 0 ≤ d ≤ d 0 0\le d\le d_0 0 ≤ d ≤ d 0 , K ≤ α + N K\le\alpha_+N K ≤ α + N , N ′ : = N + d N':=N+d N ′ := N + d , h ⃗ ∈ R N \vec{h}\in\mathbb{R}^{N} h ∈ R N and h ⃗ ′ ∈ R d \vec{h}^{\,\prime}\in\mathbb{R}^{d} h ′ ∈ R d ,
∣ N ′ B N ′ , K ( h ⃗ , h ⃗ ′ ) − N B N , K ( h ⃗ ) − ∑ l ≤ d log ( 2 cosh h l ′ ) ∣ ≤ C p a d ( d + N d ) . \left|N'B_{N',K}(\vec{h},\vec{h}^{\,\prime})-NB_{N,K}(\vec{h})-\sum_{l\le d}\log(2\cosh h_l^{\prime})\right|\le C_{\mathrm{pad}}(d+\sqrt{Nd}). N ′ B N ′ , K ( h , h ′ ) − N B N , K ( h ) − l ≤ d ∑ log ( 2 cosh h l ′ ) ≤ C pad ( d + N d ) . Consequently ∣ B N ′ , K ( h ⃗ , h ⃗ ′ ) − B N , K ( h ⃗ ) ∣ ≤ C ( d + N d ) / N + 2 d ( log 2 + ∥ h ⃗ ∥ ∞ + ∥ h ⃗ ′ ∥ ∞ + C ) / N |B_{N',K}(\vec{h},\vec{h}^{\,\prime})-B_{N,K}(\vec{h})|\le C(d+\sqrt{Nd})/N+2d(\log2+\|\vec{h}\|_\infty+\|\vec{h}^{\,\prime}\|_\infty+C)/N ∣ B N ′ , K ( h , h ′ ) − B N , K ( h ) ∣ ≤ C ( d + N d ) / N + 2 d ( log 2 + ∥ h ∥ ∞ + ∥ h ′ ∥ ∞ + C ) / N , with C = C ( β , α + , d 0 ) C=C(\beta,\alpha_+,d_0) C = C ( β , α + , d 0 ) .
Proof. Write the N ′ × K N' \times K N ′ × K matrix as ( G 1 G 2 ) \begin{pmatrix}G_1\\G_2\end{pmatrix} ( G 1 G 2 ) , G 1 G_1 G 1 of size N × K N \times K N × K , and let w : = G 1 T σ w := G_1^T\sigma w := G 1 T σ , w ′ : = G 2 T τ w' := G_2^T\tau w ′ := G 2 T τ for σ ∈ Σ N \sigma\in\Sigma_N σ ∈ Σ N , τ ∈ Σ d \tau\in\Sigma_d τ ∈ Σ d . The exponents differ by − β d 2 N N ′ ∥ w ∥ 2 + β N ′ w ⋅ w ′ + β 2 N ′ ∥ w ′ ∥ 2 -\frac{\beta d}{2NN'}\lVert w\rVert^2+\frac{\beta}{N'}w\cdot w'+\frac{\beta}{2N'}\lVert w'\rVert^2 − 2 N N ′ β d ∥ w ∥ 2 + N ′ β w ⋅ w ′ + 2 N ′ β ∥ w ′ ∥ 2 , which is bounded in absolute value, uniformly in ( σ , τ ) (\sigma,\tau) ( σ , τ ) , by
Δ : = β d 2 N N ′ ∥ G 1 ∥ o p 2 + β N ′ N d ∥ G 1 ∥ o p ∥ G 2 ∥ o p + β d 2 N ′ ∥ G 2 ∥ o p 2 . \Delta:= \frac{\beta d}{2N N'}\lVert G_1\rVert_{\mathrm{op}}^2+\frac{\beta}{N'}\sqrt{N d}\lVert G_1\rVert_{\mathrm{op}}\lVert G_2\rVert_{\mathrm{op}}+\frac{\beta d}{2N'}\lVert G_2\rVert_{\mathrm{op}}^2. Δ := 2 N N ′ β d ∥ G 1 ∥ op 2 + N ′ β N d ∥ G 1 ∥ op ∥ G 2 ∥ op + 2 N ′ β d ∥ G 2 ∥ op 2 .
The sum over τ \tau τ of e ∑ l h l ′ τ l e^{\sum_l h'_l\tau_l} e ∑ l h l ′ τ l is ∏ l 2 cosh h l ′ \prod_l 2\cosh h'_l ∏ l 2 cosh h l ′ , so the difference in the first claim is at most E Δ \mathbb{E}\Delta E Δ . Since G 2 G_2 G 2 has K K K columns, E ∥ G 2 ∥ o p 2 ≤ 3 ( d 0 + α + ) N \mathbb{E}\lVert G_2\rVert_{\mathrm{op}}^2 \le3(d_0+\alpha_+)N E ∥ G 2 ∥ op 2 ≤ 3 ( d 0 + α + ) N , and E ∥ G 1 ∥ o p 2 ≤ 3 ( 1 + α + ) N \mathbb{E}\lVert G_1\rVert_{\mathrm{op}}^2 \le3(1+\alpha_+)N E ∥ G 1 ∥ op 2 ≤ 3 ( 1 + α + ) N ; Cauchy–Schwarz gives E Δ ≤ C p a d ( d + N d ) \mathbb{E}\Delta\le C_{\mathrm{pad}}(d+\sqrt{N d}) E Δ ≤ C pad ( d + N d ) . For the second claim write B N ′ , K − B N , K = 1 N ′ [ N ′ B N ′ , K − N B N , K − ∑ l log ( 2 cosh h l ′ ) ] + 1 N ′ [ ∑ l log ( 2 cosh h l ′ ) − d B N , K ] B_{N',K}-B_{N,K}=\frac{1}{N'}[N'B_{N',K}-NB_{N,K}-\sum_l\log(2\cosh h'_l)]+\frac{1}{N'}[\sum_l\log(2\cosh h'_l)-dB_{N,K}] B N ′ , K − B N , K = N ′ 1 [ N ′ B N ′ , K − N B N , K − ∑ l log ( 2 cosh h l ′ )] + N ′ 1 [ ∑ l log ( 2 cosh h l ′ ) − d B N , K ] and use ∣ B N , K ( h ⃗ ) ∣ ≤ log 2 + ∥ h ⃗ ∥ ∞ + 3 2 β ( 1 + α + ) \lvert B_{N,K}(\vec h)\rvert\le\log2+\lVert\vec h\rVert_\infty+\frac{3}{2}\beta(1+\alpha_+) ∣ B N , K ( h )∣ ≤ log 2 + ∥ h ∥ ∞ + 2 3 β ( 1 + α + ) , by (13) and Lemma 3.2 . Only bounded fields are used below. □
The preceding comparisons remove the divisibility restrictions in the block construction and allow arbitrary limiting proportions of site types.
Theorem 4.8 (Gaussian bulk with finitely many site types). Let D ≥ 1 D\ge1 D ≥ 1 and let π = ( π 1 , … , π D ) \pi=(\pi_1,\ldots,\pi_D) π = ( π 1 , … , π D ) satisfy π j ≥ 0 \pi_j\ge0 π j ≥ 0 and ∑ j π j = 1 \sum_j\pi_j=1 ∑ j π j = 1 . For each N N N let t N t^N t N be a type assignment with N − 1 # { i : t i N = j } → π j N^{-1}\#\{i:t_i^N=j\}\to\pi_j N − 1 # { i : t i N = j } → π j for every j j j , and let K ( N ) / N → α ∈ ( 0 , ∞ ) K(N)/N\to\alpha\in(0,\infty) K ( N ) / N → α ∈ ( 0 , ∞ ) . Then for every β ≥ 0 \beta\ge0 β ≥ 0 and H ∈ R D H\in\mathbb{R}^D H ∈ R D the limit
b ( α , β ; H , π ) : = lim N → ∞ B N , K ( N ) ( H t N ) \mathfrak{b}(\alpha,\beta;H,\pi):=\lim_{N\to\infty}B_{N,K(N)}(Ht^N) b ( α , β ; H , π ) := N → ∞ lim B N , K ( N ) ( H t N ) exists, is finite, and depends on ( t N ) (t^N) ( t N ) and ( K ( N ) ) (K(N)) ( K ( N )) only through π \pi π and α \alpha α . The convergence is uniform on compact subsets of ( β , H ) ∈ [ 0 , ∞ ) × R D (\beta,H)\in[0,\infty)\times\mathbb{R}^D ( β , H ) ∈ [ 0 , ∞ ) × R D .
Proof. Step 1: block sequences. Assume first that π \pi π is rational, and let ρ = a / b \rho=a/b ρ = a / b be rational. Multiplying a a a and b b b by a common factor we may assume b π j ∈ { 0 , 1 , 2 , … } b\pi_j\in\{0,1,2,\ldots\} b π j ∈ { 0 , 1 , 2 , … } for every j j j , and we choose τ \tau τ with # τ − 1 ( j ) = b π j \#\tau^{-1}(j)=b\pi_j # τ − 1 ( j ) = b π j . By Proposition 4.4 , ℓ ρ : = lim n B b n , a n ( H t ) \ell_\rho:=\lim_n B_{bn,an}(H t) ℓ ρ := lim n B bn , an ( H t ) exists for the block assignment.
Step 2: load ⌊ ρ N ⌋ \lfloor\rho N\rfloor ⌊ ρN ⌋ . Given N N N , let n : = ⌊ N / b ⌋ n:=\lfloor N/b\rfloor n := ⌊ N / b ⌋ , N ′ : = b n N':=bn N ′ := bn and d : = N − N ′ < b d:=N-N'<b d := N − N ′ < b . Let n j : = # { i ≤ N : t i N = j } n_j:=\#\{i\le N:t_i^N=j\} n j := # { i ≤ N : t i N = j } , so that n j = N π j + o ( N ) n_j=N\pi_j+o(N) n j = N π j + o ( N ) , while the block assignment on N ′ N' N ′ sites has exactly n b π j = N π j + O ( 1 ) nb\pi_j=N\pi_j+O(1) nb π j = N π j + O ( 1 ) sites of type j j j . By the first part of Lemma 4.5 we may permute the sites so that, for every j j j , min ( n j , n b π j ) \min(n_j,nb\pi_j) min ( n j , nb π j ) of the first N ′ N' N ′ sites carry type j j j at positions where the block assignment has type j j j ; the remaining first N ′ N' N ′ sites, at most m N : = ∑ j ∣ n j − n b π j ∣ = o ( N ) m_N:=\sum_j\lvert n_j-nb\pi_j\rvert=o(N) m N := ∑ j ∣ n j − nb π j ∣ = o ( N ) of them, have the wrong type. By the second part of Lemma 4.5 , retyping them changes B B B by at most 2 max j ∣ H j ∣ m N / N → 0 2\max_j\lvert H_j\rvert m_N/N\to0 2 max j ∣ H j ∣ m N / N → 0 . By Lemma 4.7 , removing the last d d d sites changes B B B by O ( N − 1 / 2 ) O(N^{-1/2}) O ( N − 1/2 ) . Finally ∣ ⌊ ρ N ⌋ − a n ∣ ≤ ρ b + 1 \lvert\lfloor\rho N\rfloor-an\rvert\le\rho b+1 ∣⌊ ρN ⌋ − an ∣ ≤ ρ b + 1 , and Lemma 4.6 changes the load at cost c l o a d ( ρ b + 1 ) / N ′ c_{\mathrm{load}}(\rho b+1)/N' c load ( ρ b + 1 ) / N ′ . Hence B N , ⌊ ρ N ⌋ ( H t N ) → ℓ ρ B_{N,\lfloor\rho N\rfloor}(Ht^N)\to\ell_\rho B N , ⌊ ρN ⌋ ( H t N ) → ℓ ρ ; in particular ℓ ρ \ell_\rho ℓ ρ depends only on ρ \rho ρ , π \pi π and H H H .
Step 3: all loads. Let c I c_I c I be c l o a d c_{\mathrm{load}} c load with α − = α / 2 \alpha_-=\alpha/2 α − = α /2 , α + = 2 α \alpha_+=2\alpha α + = 2 α , and let ρ ∈ Q ∩ [ 3 4 α , 3 2 α ] \rho\in\mathbb{Q}\cap[\frac{3}{4}\alpha,\frac{3}{2}\alpha] ρ ∈ Q ∩ [ 4 3 α , 2 3 α ] . For large N N N , both K ( N ) K(N) K ( N ) and ⌊ ρ N ⌋ \lfloor\rho N\rfloor ⌊ ρN ⌋ lie in [ α 2 N , 2 α N ] [\frac{\alpha}{2}N,2\alpha N] [ 2 α N , 2 α N ] , and Lemma 4.6 gives ∣ B N , K ( N ) − B N , ⌊ ρ N ⌋ ∣ ≤ c I ∣ K ( N ) − ⌊ ρ N ⌋ ∣ / N \lvert B_{N,K(N)}-B_{N,\lfloor\rho N\rfloor}\rvert\le c_I\lvert K(N)-\lfloor\rho N\rfloor\rvert/N ∣ B N , K ( N ) − B N , ⌊ ρN ⌋ ∣ ≤ c I ∣ K ( N ) − ⌊ ρN ⌋∣ / N . Thus lim sup N B N , K ( N ) − lim inf N B N , K ( N ) ≤ 2 c I ∣ α − ρ ∣ \limsup_N B_{N,K(N)}-\liminf_N B_{N,K(N)}\le2c_I\lvert\alpha-\rho\rvert lim sup N B N , K ( N ) − lim inf N B N , K ( N ) ≤ 2 c I ∣ α − ρ ∣ ; let ρ → α \rho\to\alpha ρ → α . The limit is finite, since ∣ B N , K ( h ⃗ ) ∣ ≤ log 2 + ∥ h ⃗ ∥ ∞ + 3 2 β ( 1 + K / N ) \lvert B_{N,K}(\vec h)\rvert\le\log2+\lVert\vec h\rVert_\infty+\frac{3}{2}\beta(1+K/N) ∣ B N , K ( h )∣ ≤ log 2 + ∥ h ∥ ∞ + 2 3 β ( 1 + K / N ) by (13) and Lemma 3.2 , and two sequences with the same α \alpha α and π \pi π have the same limit by the same comparisons.
Step 4: arbitrary proportions. Steps 1–3 prove the claim for rational π \pi π . For general π \pi π , let π ′ \pi' π ′ be a rational probability vector. Changing the types of at most N 2 ∥ π − π ′ ∥ 1 + o ( N ) \frac{N}{2}\lVert\pi-\pi'\rVert_1+o(N) 2 N ∥ π − π ′ ∥ 1 + o ( N ) sites turns t N t^N t N into an assignment t ′ N t'^N t ′ N with proportions converging to π ′ \pi' π ′ , and by Lemma 4.5 each change costs at most 2 max j ∣ H j ∣ / N 2\max_j\lvert H_j\rvert/N 2 max j ∣ H j ∣ / N . Hence
lim sup N ∣ B N , K ( N ) ( H t N ) − b ( α , β ; H , π ′ ) ∣ ≤ max j ∣ H j ∣ ∥ π − π ′ ∥ 1 . \limsup_N\left\lvert B_{N,K(N)}(Ht^N)-\mathfrak{b}(\alpha,\beta;H,\pi')\right\rvert\le\max_j\lvert H_j\rvert\lVert\pi-\pi'\rVert_1. N lim sup B N , K ( N ) ( H t N ) − b ( α , β ; H , π ′ ) ≤ j max ∣ H j ∣ ∥ π − π ′ ∥ 1 .
For the extension, put M H : = max j ∣ H j ∣ M_H := \max_j |H_j| M H := max j ∣ H j ∣ . For rational probability vectors π ′ \pi' π ′ and π ′ ′ \pi'' π ′′ , the same retyping estimate, applied to balanced assignments and then passed to the limit, gives
∣ b ( α , β ; H , π ′ ) − b ( α , β ; H , π ′ ′ ) ∣ ≤ M H ∥ π ′ − π ′ ′ ∥ 1 . \left|b(\alpha,\beta;H,\pi')-b(\alpha,\beta;H,\pi'')\right|\le M_H\|\pi'-\pi''\|_1. ∣ b ( α , β ; H , π ′ ) − b ( α , β ; H , π ′′ ) ∣ ≤ M H ∥ π ′ − π ′′ ∥ 1 .
Consequently, for any sequence of rational probability vectors π ( k ) → π \pi^{(k)}\to\pi π ( k ) → π , the sequence b ( α , β ; H , π ( k ) ) b(\alpha,\beta;H,\pi^{(k)}) b ( α , β ; H , π ( k ) ) is Cauchy. Its limit is independent of the approximating sequence, by the same inequality; denote it by b ( α , β ; H , π ) b(\alpha,\beta;H,\pi) b ( α , β ; H , π ) . Given ϵ > 0 \epsilon>0 ϵ > 0 , choose a rational π ′ \pi' π ′ such that both M H ∥ π − π ′ ∥ 1 < ϵ M_H\|\pi-\pi'\|_1<\epsilon M H ∥ π − π ′ ∥ 1 < ϵ and ∣ b ( α , β ; H , π ′ ) − b ( α , β ; H , π ) ∣ < ϵ \left|b(\alpha,\beta;H,\pi')-b(\alpha,\beta;H,\pi)\right|<\epsilon ∣ b ( α , β ; H , π ′ ) − b ( α , β ; H , π ) ∣ < ϵ . The preceding limsup bound then yields
lim sup N ∣ B N , K ( N ) ( H t N ) − b ( α , β ; H , π ) ∣ ≤ 2 ϵ . \limsup_{N}\left|B_{N,K(N)}(H_{t_N})-b(\alpha,\beta;H,\pi)\right|\le2\epsilon. N lim sup B N , K ( N ) ( H t N ) − b ( α , β ; H , π ) ≤ 2 ϵ .
Let ϵ ↓ 0 \epsilon\downarrow0 ϵ ↓ 0 . This proves existence of the limit and its independence of the sequence of type assignments. Approximating any two probability vectors by rational ones also gives
∣ b ( α , β ; H , π ) − b ( α , β ; H , π ′ ) ∣ ≤ M H ∥ π − π ′ ∥ 1 \left|b(\alpha,\beta;H,\pi)-b(\alpha,\beta;H,\pi')\right|\le M_H\|\pi-\pi'\|_1 ∣ b ( α , β ; H , π ) − b ( α , β ; H , π ′ ) ∣ ≤ M H ∥ π − π ′ ∥ 1
for arbitrary π \pi π and π ′ \pi' π ′ .
Step 5: uniformity. By Lemma 4.5 , ∣ B N , K ( H t ) − B N , K ( H t ′ ) ∣ ≤ max j ∣ H j − H j ′ ∣ \left|B_{N,K}(H_t)-B_{N,K}(H'_t)\right|\le\max_j|H_j-H'_j| ∣ B N , K ( H t ) − B N , K ( H t ′ ) ∣ ≤ max j ∣ H j − H j ′ ∣ , and ∣ ∂ β B N , K ∣ ≤ E ∥ G ∥ o p 2 / ( 2 N ) ≤ 1 2 ( 1 + K / N ) \left|\partial_\beta B_{N,K}\right|\le\mathbb{E}\|G\|_{\mathrm{op}}^2/(2N)\le\frac{1}{2}(1+K/N) ∣ ∂ β B N , K ∣ ≤ E ∥ G ∥ op 2 / ( 2 N ) ≤ 2 1 ( 1 + K / N ) by Lemma 3.5. The functions ( β , H ) ↦ B N , K ( N ) ( H t N ) (\beta,H)\mapsto B_{N,K(N)}(H_{t_N}) ( β , H ) ↦ B N , K ( N ) ( H t N ) are therefore equi-Lipschitz on compact sets, and pointwise convergence implies uniform convergence on compact sets. The derivative bound is uniform down to β = 0 \beta=0 β = 0 . At that endpoint,
B N , K ( N ) ( H t N ) ∣ β = 0 = ∑ j = 1 D # { i : t i N = j } N log ( 2 cosh H j ) ⟶ ∑ j = 1 D π j log ( 2 cosh H j ) . B_{N,K(N)}(H_{t_N})\big|_{\beta=0}
=\sum_{j=1}^{D}\frac{\#\{i:t_i^N=j\}}{N}\log(2\cosh H_j)
\longrightarrow\sum_{j=1}^{D}\pi_j\log(2\cosh H_j). B N , K ( N ) ( H t N ) β = 0 = j = 1 ∑ D N # { i : t i N = j } log ( 2 cosh H j ) ⟶ j = 1 ∑ D π j log ( 2 cosh H j ) .
Thus the finite nets may be chosen in the closed parameter space [ 0 , ∞ ) × R D [0,\infty)\times\mathbb{R}^D [ 0 , ∞ ) × R D . The uniformity statement includes compact sets meeting β = 0 \beta=0 β = 0 , and, with D = 1 D=1 D = 1 , gives the corresponding assertion of Theorem 2.1 .
Proof of Theorem 2.1 for Gaussian patterns. Take D = 1 D=1 D = 1 , π 1 = 1 \pi_1=1 π 1 = 1 and H 1 = h H_1=h H 1 = h in Theorem 4.8 .
Binary patterns: mixed models and the lower bound
The finite-type Gaussian theorem yields the limits of mixed models with finitely many retained binary patterns. A truncated disorder comparison then gives the lower bound for the binary pressure.
For a real random variable Q Q Q under a probability measure ω \omega ω we write κ k , ω ( Q ) \kappa_{k,\omega}(Q) κ k , ω ( Q ) for its k k k -th cumulant; in particular
κ 4 , ω ( Q ) = ω Q 4 − 4 ω Q 3 ω Q − 3 ( ω Q 2 ) 2 + 12 ω Q 2 ( ω Q ) 2 − 6 ( ω Q ) 4 . \kappa_{4,\omega}(Q)=\omega Q^4-4\omega Q^3\omega Q-3(\omega Q^2)^2+12\omega Q^2(\omega Q)^2-6(\omega Q)^4. κ 4 , ω ( Q ) = ω Q 4 − 4 ω Q 3 ω Q − 3 ( ω Q 2 ) 2 + 12 ω Q 2 ( ω Q ) 2 − 6 ( ω Q ) 4 .
and Hölder’s inequality implies ∣ κ 4 , ω ( Q ) ∣ ≤ 26 ω ∣ Q ∣ 4 \left|\kappa_{4,\omega}(Q)\right|\le26\omega|Q|^4 ∣ κ 4 , ω ( Q ) ∣ ≤ 26 ω ∣ Q ∣ 4 .
Mixed-model limits
For r ≥ 1 r\ge1 r ≥ 1 , u ∈ [ − 1 , 1 ] r u\in[-1,1]^r u ∈ [ − 1 , 1 ] r , β ≥ 0 \beta\ge0 β ≥ 0 and h ∈ R h\in\mathbb{R} h ∈ R , consider D = 2 r D=2^r D = 2 r types indexed by s ∈ { − 1 , 1 } r s\in\{-1,1\}^r s ∈ { − 1 , 1 } r , with proportions π s = 2 − r \pi_s=2^{-r} π s = 2 − r and fields H s ( u ; β , h ) : = β u ⋅ s + h H_s(u;\beta,h):=\beta u\cdot s+h H s ( u ; β , h ) := β u ⋅ s + h , and let
b r ( u ; α , β , h ) : = b ( α , β ; H ( u ; β , h ) , ( 2 − r ) s ) b_r(u;\alpha,\beta,h):=b\left(\alpha,\beta;H(u;\beta,h),(2^{-r})_s\right) b r ( u ; α , β , h ) := b ( α , β ; H ( u ; β , h ) , ( 2 − r ) s )
be the corresponding limit of Theorem 4.8 .
Proposition 5.1. Let K ( N ) / N → α ∈ ( 0 , ∞ ) K(N)/N \to\alpha\in(0,\infty) K ( N ) / N → α ∈ ( 0 , ∞ ) and fix r ≥ 1 r \ge1 r ≥ 1 . For all sufficiently large N N N , K ( N ) ≥ r K(N) \ge r K ( N ) ≥ r , and the resulting tail of the sequence p N , K ( N ) ( r ) ( β , h ) p_{N,K(N)}^{(r)}(\beta,h) p N , K ( N ) ( r ) ( β , h ) converges, uniformly on compact subsets of ( β , h ) ∈ [ 0 , ∞ ) × R (\beta,h) \in[0,\infty) \times\mathbb{R} ( β , h ) ∈ [ 0 , ∞ ) × R , to
p r ( α , β , h ) = sup u ∈ [ − 1 , 1 ] r { b r ( u ; α , β , h ) − β 2 ∥ u ∥ 2 } , (22) p_r(\alpha,\beta,h)=\sup_{u\in[-1,1]^r}\left\{b_r(u;\alpha,\beta,h)-\frac{\beta}{2}\lVert u\rVert^2\right\}, \tag*{(22)} p r ( α , β , h ) = u ∈ [ − 1 , 1 ] r sup { b r ( u ; α , β , h ) − 2 β ∥ u ∥ 2 } , (22) which depends on ( K ( N ) ) (K(N)) ( K ( N )) only through α \alpha α . Moreover log ( 2 cosh h ) ≤ p r ( α , β , h ) ≤ log 2 + ∣ h ∣ + C β ( 1 + α ) \log(2\cosh h) \le p_r(\alpha,\beta,h) \le\log2+|h|+C\beta(1+\alpha) log ( 2 cosh h ) ≤ p r ( α , β , h ) ≤ log 2 + ∣ h ∣ + C β ( 1 + α ) , with C C C independent of r r r .
Proof. Write K = K ( N ) K=K(N) K = K ( N ) and m ⃗ ( σ ) : = N − 1 ( ζ a ⋅ σ ) a ≤ r ∈ [ − 1 , 1 ] r \vec m(\sigma):=N^{-1}(\zeta^a\cdot\sigma)_{a\le r}\in[-1,1]^r m ( σ ) := N − 1 ( ζ a ⋅ σ ) a ≤ r ∈ [ − 1 , 1 ] r , which takes at most ( N + 1 ) r (N+1)^r ( N + 1 ) r values. For u ∈ [ − 1 , 1 ] r u\in[-1,1]^r u ∈ [ − 1 , 1 ] r let
b ^ N ( u ) : = 1 N log ∑ σ exp { β 2 N ∥ G T σ ∥ 2 + ∑ i ( β ∑ a = 1 r u a ζ i a + h ) σ i } . \widehat b_N(u):=\frac{1}{N}\log\sum_{\sigma}\exp\left\{\frac{\beta}{2N}\lVert G^T\sigma\rVert^2+\sum_i\left(\beta\sum_{a=1}^r u_a\zeta_i^a+h\right)\sigma_i\right\}. b N ( u ) := N 1 log σ ∑ exp { 2 N β ∥ G T σ ∥ 2 + i ∑ ( β a = 1 ∑ r u a ζ i a + h ) σ i } .
Since β 2 N ∑ a ( ζ a ⋅ σ ) 2 = β N 2 ∥ m ⃗ ( σ ) ∥ 2 \frac{\beta}{2N}\sum_a(\zeta^a\cdot\sigma)^2=\frac{\beta N}{2}\lVert\vec m(\sigma)\rVert^2 2 N β ∑ a ( ζ a ⋅ σ ) 2 = 2 β N ∥ m ( σ ) ∥ 2 , decomposing according to the values of m ⃗ \vec m m gives, pointwise,
0 ≤ 1 N log Z N , K ( r ) ( β , h ) − sup u ∈ [ − 1 , 1 ] r { b ^ N ( u ) − β 2 ∥ u ∥ 2 } ≤ r log ( N + 1 ) N . (23) 0\le\frac{1}{N}\log Z_{N,K}^{(r)}(\beta,h)-\sup_{u\in[-1,1]^r}\left\{\widehat b_N(u)-\frac{\beta}{2}\lVert u\rVert^2\right\}\le\frac{r\log(N+1)}{N}. \tag*{(23)} 0 ≤ N 1 log Z N , K ( r ) ( β , h ) − u ∈ [ − 1 , 1 ] r sup { b N ( u ) − 2 β ∥ u ∥ 2 } ≤ N r log ( N + 1 ) . (23)
The lower bound uses 1 2 ∥ m ⃗ ∥ 2 ≥ u ⋅ m ⃗ − 1 2 ∥ u ∥ 2 \frac{1}{2}\lVert\vec m\rVert^2\ge u\cdot\vec m-\frac{1}{2}\lVert u\rVert^2 2 1 ∥ m ∥ 2 ≥ u ⋅ m − 2 1 ∥ u ∥ 2 ; the upper bound takes u = m ⃗ u=\vec m u = m in each constrained sum.
The rows ( ζ i 1 , … , ζ i r ) ∈ { − 1 , 1 } r (\zeta_i^1,\ldots,\zeta_i^r)\in\{-1,1\}^r ( ζ i 1 , … , ζ i r ) ∈ { − 1 , 1 } r define a random type assignment with counts n s n_s n s ; site i i i carries the field H s ( u ; β , h ) H_s(u;\beta,h) H s ( u ; β , h ) of its type. Conditionally on ζ \zeta ζ , E G ′ b ^ N ( u ) = B N , K − r ( H ( u ) t ) \mathbb{E}_{G'}\widehat b_N(u)=B_{N,K-r}(H(u)_{\mathfrak t}) E G ′ b N ( u ) = B N , K − r ( H ( u ) t ) in the notation of (4.3), and by Lemma 4.5 it depends on ζ \zeta ζ only through the counts n s n_s n s . Let t ˉ N \bar{\mathfrak t}_N t ˉ N be a deterministic type assignment with counts n ˉ s \bar n_s n ˉ s , ∣ n ˉ s − N 2 − r ∣ ≤ 1 |\bar n_s-N2^{-r}|\le1 ∣ n ˉ s − N 2 − r ∣ ≤ 1 . After a permutation, the two assignments differ at 1 2 ∑ s ∣ n s − n ˉ s ∣ \frac{1}{2}\sum_s|n_s-\bar n_s| 2 1 ∑ s ∣ n s − n ˉ s ∣ sites, where the fields differ by at most 2 β ∥ u ∥ 1 ≤ 2 β r 2\beta\lVert u\rVert_1\le2\beta r 2 β ∥ u ∥ 1 ≤ 2 β r . By Lemma 4.5,
E sup u ∣ E G ′ b ^ N ( u ) − B N , K − r ( H ( u ) t ˉ N ) ∣ ≤ β r N ∑ s E ∣ n s − n ˉ s ∣ = O r , β ( N − 1 / 2 ) . \mathbb{E}\sup_u\left|\mathbb{E}_{G'}\widehat b_N(u)-B_{N,K-r}(H(u)_{\bar{\mathfrak t}_N})\right|
\le\frac{\beta r}{N}\sum_s\mathbb{E}|n_s-\bar n_s|
=O_{r,\beta}(N^{-1/2}). E u sup E G ′ b N ( u ) − B N , K − r ( H ( u ) t ˉ N ) ≤ N β r s ∑ E ∣ n s − n ˉ s ∣ = O r , β ( N − 1/2 ) .
Theorem 4.8, applied with ( K − r ) / N → α (K-r)/N\to\alpha ( K − r ) / N → α and the compact set of fields { H ( u ; β , h ) : u ∈ [ − 1 , 1 ] r } \{H(u;\beta,h):u\in[-1,1]^r\} { H ( u ; β , h ) : u ∈ [ − 1 , 1 ] r } , gives B N , K − r ( H ( u ) t ˉ N ) → b r ( u ) B_{N,K-r}(H(u)_{\bar{\mathfrak t}_N})\to b_r(u) B N , K − r ( H ( u ) t ˉ N ) → b r ( u ) uniformly in u u u . Conditionally on ζ \zeta ζ , (T1) and (3.2) give Var G ′ b ^ N ( u ) ≤ β 2 N − 2 E ∥ G ′ ∥ o p 2 = O ( N − 1 ) \operatorname{Var}_{G'}\widehat b_N(u)\le\beta^2N^{-2}\mathbb{E}\lVert G'\rVert_{\mathrm{op}}^2=O(N^{-1}) Var G ′ b N ( u ) ≤ β 2 N − 2 E ∥ G ′ ∥ op 2 = O ( N − 1 ) , uniformly in u u u and ζ \zeta ζ . Finally ∂ u a b ^ N = β N − 1 ∑ i ζ i a ω ( σ i ) ∈ [ − β , β ] \partial_{u_a}\widehat b_N=\beta N^{-1}\sum_i\zeta_i^a\omega(\sigma_i)\in[-\beta,\beta] ∂ u a b N = β N − 1 ∑ i ζ i a ω ( σ i ) ∈ [ − β , β ] , so b ^ N \widehat b_N b N and its limit are β \beta β -Lipschitz for the ℓ 1 \ell^1 ℓ 1 -norm. Fix θ > 0 \theta>0 θ > 0 and choose a finite set U θ ⊂ [ − 1 , 1 ] r U_\theta\subset[-1,1]^r U θ ⊂ [ − 1 , 1 ] r such that every u u u is within ℓ 1 \ell^1 ℓ 1 -distance θ \theta θ of some v ∈ U θ v\in U_\theta v ∈ U θ . Its cardinality m θ m_\theta m θ is independent of N N N . Put X N ( v ) : = b ^ N ( v ) − E G ′ b ^ N ( v ) X_N(v):=\widehat b_N(v)-\mathbb{E}_{G'}\widehat b_N(v) X N ( v ) := b N ( v ) − E G ′ b N ( v ) , where the second term is conditional on ζ \zeta ζ . The conditional variance bound and Cauchy–Schwarz give
E max v ∈ U θ ∣ X N ( v ) ∣ ≤ ( ∑ v ∈ U θ E ∣ X N ( v ) ∣ 2 ) 1 / 2 ≤ m θ C / N . \mathbb{E}\max_{v\in U_\theta}|X_N(v)|
\le\left(\sum_{v\in U_\theta}\mathbb{E}|X_N(v)|^2\right)^{1/2}
\le\sqrt{m_\theta C/N}. E v ∈ U θ max ∣ X N ( v ) ∣ ≤ ( v ∈ U θ ∑ E ∣ X N ( v ) ∣ 2 ) 1/2 ≤ m θ C / N .
For u u u and a corresponding net point v v v , the two Lipschitz bounds give an error at most 2 β θ 2\beta\theta 2 β θ . Therefore
E sup u ∣ b ^ N ( u ) − b r ( u ) ∣ ≤ 2 β θ + m θ C / N + E sup u ∣ E G ′ b ^ N ( u ) − b r ( u ) ∣ . \mathbb{E}\sup_u|\widehat b_N(u)-b_r(u)|
\le2\beta\theta+\sqrt{m_\theta C/N}
+\mathbb{E}\sup_u|\mathbb{E}_{G'}\widehat b_N(u)-b_r(u)|. E u sup ∣ b N ( u ) − b r ( u ) ∣ ≤ 2 β θ + m θ C / N + E u sup ∣ E G ′ b N ( u ) − b r ( u ) ∣.
The last term tends to zero by the comparison of type counts and the uniform convergence for balanced types proved above. First taking N → ∞ N\to\infty N → ∞ at fixed θ \theta θ gives a limsup at most $2\beta\theta$; then θ ↓ 0 \theta\downarrow0 θ ↓ 0 proves E sup u ∣ b ^ N ( u ) − b r ( u ) ∣ → 0 \mathbb{E}\sup_u|\widehat b_N(u)-b_r(u)|\to0 E sup u ∣ b N ( u ) − b r ( u ) ∣ → 0 . By (5.2), ∣ p N , K ( r ) − sup u { b r ( u ) − β 2 ∥ u ∥ 2 } ∣ ≤ r log ( N + 1 ) / N + E sup u ∣ b ^ N − b r ∣ → 0 |p_{N,K}^{(r)}-\sup_u\{b_r(u)-\frac{\beta}{2}\lVert u\rVert^2\}|\le r\log(N+1)/N+\mathbb{E}\sup_u|\widehat b_N-b_r|\to0 ∣ p N , K ( r ) − sup u { b r ( u ) − 2 β ∥ u ∥ 2 } ∣ ≤ r log ( N + 1 ) / N + E sup u ∣ b N − b r ∣ → 0 , which proves (5.1) at fixed ( β , h ) (\beta,h) ( β , h ) . The functions $p_{N,K}^{(r)} are 1-Lipschitz in h h h and satisfy ∣ ∂ β p N , K ( r ) ∣ ≤ E ∥ ( ζ , G ′ ) ∥ o p 2 / ( 2 N ) ≤ C ( 1 + K / N ) |\partial_{\beta}p_{N,K}^{(r)}| \le\mathbb{E}\|(\zeta,G')\|_{\mathrm{op}}^{2}/(2N) \le C(1+K/N) ∣ ∂ β p N , K ( r ) ∣ ≤ E ∥ ( ζ , G ′ ) ∥ op 2 / ( 2 N ) ≤ C ( 1 + K / N ) by Lemmas 3.5 and 3.3. In particular, on the tail N ≥ N r N \ge N_r N ≥ N r , these functions have a common Lipschitz constant L L L for the metric ∣ β − β ′ ∣ + ∣ h − h ′ ∣ |\beta-\beta'|+|h-h'| ∣ β − β ′ ∣ + ∣ h − h ′ ∣ , and their pointwise limit inherits that bound. If C ⊂ [ 0 , ∞ ) × R C \subset[0,\infty)\times\mathbb{R} C ⊂ [ 0 , ∞ ) × R is compact and T θ T_\theta T θ is a finite θ \theta θ -net in C C C , then
sup ( β , h ) ∈ C ∣ p N , K ( N ) ( r ) ( β , h ) − p r ( α , β , h ) ∣ ≤ 2 L θ + max ( β , h ) ∈ T θ ∣ p N , K ( N ) ( r ) ( β , h ) − p r ( α , β , h ) ∣ . \sup_{(\beta,h)\in C}\left|p_{N,K(N)}^{(r)}(\beta,h)-p_r(\alpha,\beta,h)\right| \le2L\theta+\max_{(\beta,h)\in T_\theta}\left|p_{N,K(N)}^{(r)}(\beta,h)-p_r(\alpha,\beta,h)\right|. ( β , h ) ∈ C sup p N , K ( N ) ( r ) ( β , h ) − p r ( α , β , h ) ≤ 2 L θ + ( β , h ) ∈ T θ max p N , K ( N ) ( r ) ( β , h ) − p r ( α , β , h ) .
For each fixed net, the maximum tends to zero by pointwise convergence; letting θ ↓ 0 \theta\downarrow0 θ ↓ 0 proves local uniformity. This includes compact sets meeting β = 0 \beta=0 β = 0 , because the derivative bound is uniform there and p N , K ( N ) ( r ) ( 0 , h ) = p r ( α , 0 , h ) = log ( 2 cosh h ) p_{N,K(N)}^{(r)}(0,h)=p_r(\alpha,0,h)=\log(2\cosh h) p N , K ( N ) ( r ) ( 0 , h ) = p r ( α , 0 , h ) = log ( 2 cosh h ) . The bounds follow from Lemma 3.5, since the full mixed matrix has independent unit sub-Gaussian entries and Lemma 3.3 is uniform in r ≤ K r\le K r ≤ K .
Lower comparison
The mixed-model limits provide comparison values for the binary pressure. We first establish a replacement estimate uniform in the potential generated by the retained patterns.
Lemma 5.2. Let V : Σ N → R V:\Sigma_N\to\mathbb{R} V : Σ N → R be deterministic. Let J J J be an N × K N\times K N × K matrix with independent entries, n R n_R n R of them symmetric Rademacher and the others standard Gaussian, let G G G be a standard Gaussian N × K N\times K N × K matrix, and let K / N ∈ [ α − , α + ] K/N\in[\alpha_-,\alpha_+] K / N ∈ [ α − , α + ] . Then for every L ≥ 4 L\ge4 L ≥ 4 ,
E f N V ( J ) ≥ E f N V ( G ) − C c u t L 2 − 13 3 β 2 n R L 4 N 3 . \mathbb{E}f_N^V(J)\ge\mathbb{E}f_N^V(G)-\frac{C_{\mathrm{cut}}}{L^2}-\frac{13}{3}\frac{\beta^2 n_RL^4}{N^3}. E f N V ( J ) ≥ E f N V ( G ) − L 2 C cut − 3 13 N 3 β 2 n R L 4 . In particular, for J = ε J=\varepsilon J = ε and N ≥ 4 6 N\ge4^6 N ≥ 4 6 , E f N V ( ε ) ≥ E f N V ( G ) − ( C c u t + 13 3 β 2 α + ) N − 1 / 3 \mathbb{E}f_N^V(\varepsilon)\ge\mathbb{E}f_N^V(G)-(C_{\mathrm{cut}}+\frac{13}{3}\beta^2\alpha_+)N^{-1/3} E f N V ( ε ) ≥ E f N V ( G ) − ( C cut + 3 13 β 2 α + ) N − 1/3 , uniformly in V V V . If V V V is random and independent of J J J and G G G , both bounds hold conditionally on V V V , and hence for the unconditional expectations whenever E max σ ∣ V ( σ ) ∣ < ∞ \mathbb{E}\max_\sigma|V(\sigma)|<\infty E max σ ∣ V ( σ ) ∣ < ∞ .
Proof. Let f N V , [ L ] ( J ) : = N − 1 log ∑ σ e V ( σ ) ∫ [ − L , L ] K e β N − 1 / 2 σ T J z ⊗ K ( d z ) f_N^{V,[L]}(J):=N^{-1}\log\sum_\sigma e^{V(\sigma)}\int_{[-L,L]^K}e^{\beta N^{-1/2}\sigma^TJz}\otimes^K(dz) f N V , [ L ] ( J ) := N − 1 log ∑ σ e V ( σ ) ∫ [ − L , L ] K e β N − 1/2 σ T J z ⊗ K ( d z ) , so that f N V ≥ f N V , [ L ] f_N^V\ge f_N^{V,[L]} f N V ≥ f N V , [ L ] pointwise by (17) . As a function of a single entry x = J i μ x=J_i^\mu x = J i μ , the k k k -th derivative of f N V , [ L ] f_N^{V,[L]} f N V , [ L ] is N − 1 ( β / N ) k / 2 κ k , ω ( σ i z μ ) N^{-1}(\beta/N)^{k/2}\kappa_{k,\omega}(\sigma_i z_\mu) N − 1 ( β / N ) k /2 κ k , ω ( σ i z μ ) , with ω \omega ω the truncated Gibbs measure on Σ N × [ − L , L ] K \Sigma_N\times[-L,L]^K Σ N × [ − L , L ] K ; in particular the fourth derivative is bounded by 26 β 2 L 4 / N 3 26\beta^2L^4/N^3 26 β 2 L 4 / N 3 . The first three moments of a symmetric Rademacher and a standard Gaussian variable agree, and their fourth moments add up to 1 + 3 = 4 1+3=4 1 + 3 = 4 . Replacing the n R n_R n R Rademacher entries one at a time by Gaussian ones and using a third-order Taylor expansion with fourth-order remainder therefore gives ∣ E f N V , [ L ] ( J ) − E f N V , [ L ] ( G ) ∣ ≤ 26 ⋅ 4 24 β 2 L 4 N 3 n R = 13 3 β 2 n R L 4 N 3 |\mathbb{E}f_N^{V,[L]}(J)-\mathbb{E}f_N^{V,[L]}(G)|\le\frac{26\cdot4}{24}\frac{\beta^2L^4}{N^3}n_R=\frac{13}{3}\frac{\beta^2n_RL^4}{N^3} ∣ E f N V , [ L ] ( J ) − E f N V , [ L ] ( G ) ∣ ≤ 24 26 ⋅ 4 N 3 β 2 L 4 n R = 3 13 N 3 β 2 n R L 4 . Proposition 4.2 gives E f N V , [ L ] ( G ) ≥ E f N V ( G ) − C c u t L − 2 \mathbb{E}f_N^{V,[L]}(G)\ge\mathbb{E}f_N^V(G)-C_{\mathrm{cut}}L^{-2} E f N V , [ L ] ( G ) ≥ E f N V ( G ) − C cut L − 2 . For the last claim take n R = N K n_R=NK n R = N K and L = N 1 / 6 L=N^{1/6} L = N 1/6 . □
For each fixed r r r , let ζ 1 , … , ζ r \zeta^1,\ldots,\zeta^r ζ 1 , … , ζ r be the first r r r columns of ε \varepsilon ε , condition on them, and use V ( σ ) = β 2 N ∑ a = 1 r ( ζ a ⋅ σ ) 2 + h N m ( σ ) V(\sigma)=\frac{\beta}{2N}\sum_{a=1}^r(\zeta^a\cdot\sigma)^2+hNm(\sigma) V ( σ ) = 2 N β ∑ a = 1 r ( ζ a ⋅ σ ) 2 + h N m ( σ ) with the remaining K − r K-r K − r columns as bulk. Since max σ ∣ V ( σ ) ∣ ≤ N ( β r / 2 + ∣ h ∣ ) \max_\sigma|V(\sigma)|\le N(\beta r/2+|h|) max σ ∣ V ( σ ) ∣ ≤ N ( β r /2 + ∣ h ∣ ) , the integrability condition of Lemma 5.2 holds. Then A N , K R ( h ) A_{N,K}^{R}(h) A N , K R ( h ) and p N , K ( r ) ( h ) p_{N,K}^{(r)}(h) p N , K ( r ) ( h ) are the binary- and Gaussian-bulk pressures with this potential, and Lemma 5.2 and Proposition 5.1 yield
lim inf N A N , K ( N ) R ( h ) ≥ p r ( α , β , h ) for every r ≥ 1 , \liminf_N A_{N,K(N)}^{R}(h)\ge p_r(\alpha,\beta,h)\qquad\text{for every }r\ge1, N lim inf A N , K ( N ) R ( h ) ≥ p r ( α , β , h ) for every r ≥ 1 ,
lim inf N A N , K ( N ) R ( h ) ≥ A ∗ : = sup r ≥ 1 p r ( α , β , h ) . (24) \liminf_N A_{N,K(N)}^{R}(h)\ge A_*:=\sup_{r\ge1}p_r(\alpha,\beta,h). \tag*{(24)} N lim inf A N , K ( N ) R ( h ) ≥ A ∗ := r ≥ 1 sup p r ( α , β , h ) . (24)
The estimate is independent of the size of the potential, which may be extensive and depend on N N N , and requires no Gibbs-overlap condition. With V = h N m V=hNm V = h N m and no retained column, Lemma 5.2 also gives A N , K R ( h ) ≥ A N , K G ( h ) − C N − 1 / 3 A_{N,K}^{R}(h)\ge A_{N,K}^{G}(h)-CN^{-1/3} A N , K R ( h ) ≥ A N , K G ( h ) − C N − 1/3 .
Binary patterns: radial cutoff and sector decomposition
For the upper bound, the retained patterns must be selected from the hidden coordinates. A radial cutoff bounds their number, and a sector decomposition permits comparison with the mixed models.
A radial hidden-spin cutoff
Return to the hidden representation of all K K K binary columns:
Z N , K R ( h ) = ∑ σ ∫ γ ⊗ K ( d z ) exp { β N σ T ε z + h N m ( σ ) } . Z_{N,K}^{R}(h)=\sum_{\sigma}\int\gamma^{\otimes K}(dz)\exp\left\{\sqrt{\frac{\beta}{N}}\sigma^{T}\varepsilon z+hNm(\sigma)\right\}. Z N , K R ( h ) = σ ∑ ∫ γ ⊗ K ( d z ) exp { N β σ T ε z + h N m ( σ ) } .
Write Z N , K R , r a d Z_{N,K}^{R,\mathrm{rad}} Z N , K R , rad for the restriction to ∥ z ∥ 2 ≤ ϱ N \lVert z\rVert^2\le\varrho N ∥ z ∥ 2 ≤ ϱN .
Lemma 6.1. Let K ≤ α + N K\le\alpha_+N K ≤ α + N . There is ϱ < ∞ \varrho<\infty ϱ < ∞ , depending only on β \beta β and α + \alpha_+ α + , such that
0 ≤ A N , K R ( h ) − N − 1 E log Z N , K R , r a d ( h ) ≤ − 1 N log ( 1 − e − N / 2 ) + 2 e − N ( β K + C K N ) , 0\le A_{N,K}^{R}(h)-N^{-1}\mathbb{E}\log Z_{N,K}^{R,\mathrm{rad}}(h)\le-\frac{1}{N}\log(1-e^{-N/2})+2e^{-N}\left(\beta K+\frac{CK}{N}\right), 0 ≤ A N , K R ( h ) − N − 1 E log Z N , K R , rad ( h ) ≤ − N 1 log ( 1 − e − N /2 ) + 2 e − N ( β K + N C K ) , with C = C ( β , α + ) C=C(\beta,\alpha_+) C = C ( β , α + ) , uniformly in h h h .
Proof. Let c ⋆ c_\star c ⋆ be as in Lemma 3.3, so that the event { ∥ ε ∥ o p ≤ c ⋆ N } \{\lVert\varepsilon\rVert_{\mathrm{op}}\le c_\star\sqrt{N}\} {∥ ε ∥ op ≤ c ⋆ N } has probability at least 1 − 2 e − N 1-2e^{-N} 1 − 2 e − N , and take ϱ > β c ⋆ + α + + 1 \sqrt{\varrho}>\sqrt{\beta}c_\star+\sqrt{\alpha_+}+1 ϱ > β c ⋆ + α + + 1 . Conditionally on σ \sigma σ , the hidden law is N ( a , I K ) N(a,I_K) N ( a , I K ) , with a = β / N ε T σ a=\sqrt{\beta/N}\,\varepsilon^T\sigma a = β / N ε T σ and ∥ a ∥ ≤ β ∥ ε ∥ o p \lVert a\rVert\le\sqrt{\beta}\lVert\varepsilon\rVert_{\mathrm{op}} ∥ a ∥ ≤ β ∥ ε ∥ op , and Z N , K R , r a d / Z R = ω ( P Y ( ∥ Y + a ∥ 2 ≤ ϱ N ) ) Z_{N,K}^{R,\mathrm{rad}}/Z^R=\omega(\mathbb{P}_Y(\lVert Y+a\rVert^2\le\varrho N)) Z N , K R , rad / Z R = ω ( P Y (∥ Y + a ∥ 2 ≤ ϱN )) with ω \omega ω the Ising Gibbs measure. On the good event, Gaussian norm concentration gives, uniformly in σ \sigma σ , P ( ∥ Y + a ∥ > ϱ N ) ≤ P ( ∥ Y ∥ − E ∥ Y ∥ > N ) ≤ e − N / 2 \mathbb{P}(\lVert Y+a\rVert>\sqrt{\varrho N})\le\mathbb{P}(\lVert Y\rVert-\mathbb{E}\lVert Y\rVert>\sqrt{N})\le e^{-N/2} P (∥ Y + a ∥ > ϱN ) ≤ P (∥ Y ∥ − E ∥ Y ∥ > N ) ≤ e − N /2 , so the normalized loss is at most − N − 1 log ( 1 − e − N / 2 ) -N^{-1}\log(1-e^{-N/2}) − N − 1 log ( 1 − e − N /2 ) .
On the exceptional event, choose d : = ( ϱ / ( 2 α + ) ) 1 / 2 d:=(\varrho/(2\alpha_+))^{1/2} d := ( ϱ / ( 2 α + ) ) 1/2 , so that K d 2 ≤ ϱ N / 2 Kd^2\le\varrho N/2 K d 2 ≤ ϱN /2 . Integrating the shifted Gaussian density on [ − d , d ] K [-d,d]^K [ − d , d ] K and using ∥ z − a ∥ 2 / 2 ≤ ∥ a ∥ 2 + ∥ z ∥ 2 \lVert z-a\rVert^2/2\le\lVert a\rVert^2+\lVert z\rVert^2 ∥ z − a ∥ 2 /2 ≤ ∥ a ∥ 2 + ∥ z ∥ 2 gives P ( ∥ Y + a ∥ ≤ ϱ N ) ≥ exp { − ∥ a ∥ 2 − C K } \mathbb{P}(\lVert Y+a\rVert\le\sqrt{\varrho N})\ge\exp\{-\lVert a\rVert^2-CK\} P (∥ Y + a ∥ ≤ ϱN ) ≥ exp { − ∥ a ∥ 2 − C K } with C = d 2 + ( log ( 2 π / ( 2 d ) ) ) + C=d^2+(\log(\sqrt{2\pi}/(2d)))_+ C = d 2 + ( log ( 2 π / ( 2 d )) ) + . Thus the normalized loss is at most ( β ∥ ε ∥ o p 2 + C K ) / N ≤ β K + C K / N (\beta\lVert\varepsilon\rVert_{\mathrm{op}}^2+CK)/N\le\beta K+CK/N ( β ∥ ε ∥ op 2 + C K ) / N ≤ β K + C K / N , since ∥ ε ∥ o p 2 ≤ ∥ ε ∥ H S 2 = N K \lVert\varepsilon\rVert_{\mathrm{op}}^2\le\lVert\varepsilon\rVert_{\mathrm{HS}}^2=NK ∥ ε ∥ op 2 ≤ ∥ ε ∥ HS 2 = N K .
Finite sectors
Fix ϱ \varrho ϱ as in Lemma 6.1 and δ > 0 \delta>0 δ > 0 . Column 1 is always kept binary; this reference column only keeps the index of the mixed model at least one. For S ⊆ { 1 , … , K } S\subseteq\{1,\ldots,K\} S ⊆ { 1 , … , K } with 1 ∈ S 1\in S 1 ∈ S let
D S : = { z ∈ R K : ∥ z ∥ 2 ≤ ϱ N , ∣ z μ ∣ > δ N ( μ ∈ S ∖ { 1 } ) , ∣ z μ ∣ ≤ δ N ( μ ∉ S ) } . D_S:=\left\{z\in\mathbb{R}^K:\ \lVert z\rVert^2\le\varrho N,\ |z_\mu|>\delta\sqrt{N}\ (\mu\in S\setminus\{1\}),\ |z_\mu|\le\delta\sqrt{N}\ (\mu\notin S)\right\}. D S := { z ∈ R K : ∥ z ∥ 2 ≤ ϱN , ∣ z μ ∣ > δ N ( μ ∈ S ∖ { 1 }) , ∣ z μ ∣ ≤ δ N ( μ ∈ / S ) } .
These domains partition the radial ball, since each z z z determines S = { 1 } ∪ { μ ≠ 1 : ∣ z μ ∣ > δ N } S=\{1\}\cup\{\mu\ne1:|z_\mu|>\delta\sqrt{N}\} S = { 1 } ∪ { μ = 1 : ∣ z μ ∣ > δ N } . Only sets with
∣ S ∣ ≤ M δ : = 1 + ⌊ ϱ δ 2 ⌋ |S|\le M_\delta:=1+\left\lfloor\frac{\varrho}{\delta^2}\right\rfloor ∣ S ∣ ≤ M δ := 1 + ⌊ δ 2 ϱ ⌋
can have positive measure, because ( ∣ S ∣ − 1 ) δ 2 N < ∥ z ∥ 2 ≤ ϱ N (|S|-1)\delta^2N<\lVert z\rVert^2\le\varrho N ( ∣ S ∣ − 1 ) δ 2 N < ∥ z ∥ 2 ≤ ϱN on D S D_S D S . Let S N \mathcal{S}_N S N be the family of sets S S S with γ ⊗ K ( D S ) > 0 \gamma^{\otimes K}(D_S)>0 γ ⊗ K ( D S ) > 0 ; then
∣ S N ∣ ≤ M δ ( K + 1 ) M δ , log ∣ S N ∣ = O ϱ , δ , α + ( log N ) . (25) |\mathcal{S}_N|\le M_\delta(K+1)^{M_\delta},\qquad\log|\mathcal{S}_N|=O_{\varrho,\delta,\alpha_+}(\log N). \tag*{(25)} ∣ S N ∣ ≤ M δ ( K + 1 ) M δ , log ∣ S N ∣ = O ϱ , δ , α + ( log N ) . (25)
For a real N × K N \times K N × K matrix J J J and S ∈ S N S \in S_N S ∈ S N let
Φ S ( J ) : = 1 N log Z S ( J ) , Z S ( J ) : = ∑ σ e h N m ( σ ) ∫ D S e β / N σ T J z γ ⊗ K ( d z ) , \Phi_S(J):=\frac{1}{N}\log Z_S(J),\qquad Z_S(J):=\sum_{\sigma}e^{hN m(\sigma)}\int_{\mathcal{D}_S}e^{\sqrt{\beta/N}\,\sigma^T Jz_{\gamma}\otimes K}(dz), Φ S ( J ) := N 1 log Z S ( J ) , Z S ( J ) := σ ∑ e h N m ( σ ) ∫ D S e β / N σ T J z γ ⊗ K ( d z ) ,
which is finite and smooth. The domains depend only on the integration variables and on ϱ \varrho ϱ , δ \delta δ , N N N , not on the disorder; in particular the disorder keeps its product law in every sector.
Lemma 6.2 (Small-coordinate comparison). Keep J μ = ε μ J^\mu=\varepsilon^\mu J μ = ε μ for μ ∈ S \mu\in S μ ∈ S and replace all other columns by independent standard Gaussian columns G S c G_{S^c} G S c . Then
∣ E Φ S ( ε ) − E Φ S ( ε S , G S c ) ∣ ≤ q δ : = 13 3 β 2 δ 2 e 4 β δ , (26) \left|\mathbb{E}\Phi_S(\varepsilon)-\mathbb{E}\Phi_S(\varepsilon_S,G_{S^c})\right|\le q_\delta:=\frac{13}{3}\beta^2\delta^2e^{4\sqrt{\beta}\,\delta}, \tag*{(26)} ∣ E Φ S ( ε ) − E Φ S ( ε S , G S c ) ∣ ≤ q δ := 3 13 β 2 δ 2 e 4 β δ , (26) uniformly in N N N , K K K , S S S , h h h .
Proof. For μ ∉ S \mu\notin S μ ∈ / S interpolate J i μ ( t ) = t ε i μ + 1 − t g i μ J_i^\mu(t)=\sqrt{t}\,\varepsilon_i^\mu+\sqrt{1-t}\,g_i^\mu J i μ ( t ) = t ε i μ + 1 − t g i μ , and let ω t \omega_t ω t be the corresponding Gibbs measure on Σ N × D S \Sigma_N\times\mathcal{D}_S Σ N × D S . Set Q = σ i z μ Q=\sigma_i z_\mu Q = σ i z μ and let f ( x ) = ω t ( x ) ( Q ) f(x)=\omega_t^{(x)}(Q) f ( x ) = ω t ( x ) ( Q ) as a function of the binary entry ε i μ = x \varepsilon_i^\mu=x ε i μ = x , all other variables being fixed. For a symmetric sign s s s ,
∣ E s [ s f ( s ) ] − E s f ′ ( s ) ∣ ≤ 2 3 sup ∣ x ∣ ≤ 1 ∣ f ′ ′ ′ ( x ) ∣ ; (27) \left|\mathbb{E}_s[sf(s)]-\mathbb{E}_s f'(s)\right|\le\frac{2}{3}\sup_{|x|\le1}|f'''(x)|; \tag*{(27)} ∣ E s [ s f ( s )] − E s f ′ ( s ) ∣ ≤ 3 2 ∣ x ∣ ≤ 1 sup ∣ f ′′′ ( x ) ∣ ; (27)
indeed, expand f f f to second order and f ′ f' f ′ to first order at zero: the leading terms agree, and the remainders are bounded by ∥ f ′ ′ ′ ∥ ∞ / 6 \|f'''\|_\infty/6 ∥ f ′′′ ∥ ∞ /6 and ∥ f ′ ′ ′ ∥ ∞ / 2 \|f'''\|_\infty/2 ∥ f ′′′ ∥ ∞ /2 .
Now f ′ ′ ′ ( x ) = ( β t / N ) 3 / 2 κ 4 , ω t ( x ) ( Q ) f'''(x)=(\beta t/N)^{3/2}\kappa_{4,\omega_t^{(x)}}(Q) f ′′′ ( x ) = ( β t / N ) 3/2 κ 4 , ω t ( x ) ( Q ) , so ∣ f ′ ′ ′ ( x ) ∣ ≤ 26 ( β t / N ) 3 / 2 ω t ( x ) ( z μ 4 ) |f'''(x)|\le26(\beta t/N)^{3/2}\omega_t^{(x)}(z_\mu^4) ∣ f ′′′ ( x ) ∣ ≤ 26 ( β t / N ) 3/2 ω t ( x ) ( z μ 4 ) . Changing x ∈ [ − 1 , 1 ] x\in[-1,1] x ∈ [ − 1 , 1 ] to either actual sign changes the exponent by at most 2 β t δ 2\sqrt{\beta t}\,\delta 2 β t δ , because ∣ z μ ∣ ≤ δ N |z_\mu|\le\delta\sqrt{N} ∣ z μ ∣ ≤ δ N on the whole sector; hence the normalized Gibbs densities differ by a factor at most e 4 β δ e^{4\sqrt{\beta}\,\delta} e 4 β δ , and for both values s ∈ { − 1 , 1 } s\in\{-1,1\} s ∈ { − 1 , 1 } , sup ∣ x ∣ ≤ 1 ω t ( x ) ( z μ 4 ) ≤ e 4 β δ ω t ( s ) ( z μ 4 ) \sup_{|x|\le1}\omega_t^{(x)}(z_\mu^4)\le e^{4\sqrt{\beta}\,\delta}\omega_t^{(s)}(z_\mu^4) sup ∣ x ∣ ≤ 1 ω t ( x ) ( z μ 4 ) ≤ e 4 β δ ω t ( s ) ( z μ 4 ) .
Let q S ( t ) = E Φ S ( J ( t ) ) q_S(t)=\mathbb{E}\Phi_S(J(t)) q S ( t ) = E Φ S ( J ( t )) . For 0 < t < 1 0<t<1 0 < t < 1 ,
q S ′ ( t ) = β 2 N 3 / 2 ∑ i , μ ∉ S { E ε i μ ω t ( Q ) t − E g i μ ω t ( Q ) 1 − t } . q_S'(t)=\frac{\sqrt{\beta}}{2N^{3/2}}\sum_{i,\,\mu\notin S}\left\{\frac{\mathbb{E}\varepsilon_i^\mu\omega_t(Q)}{\sqrt{t}}-\frac{\mathbb{E}g_i^\mu\omega_t(Q)}{\sqrt{1-t}}\right\}. q S ′ ( t ) = 2 N 3/2 β i , μ ∈ / S ∑ { t E ε i μ ω t ( Q ) − 1 − t E g i μ ω t ( Q ) } .
Gaussian integration by parts gives E g i μ ω t ( Q ) = β ( 1 − t ) / N E [ ω t ( Q 2 ) − ω t ( Q ) 2 ] \mathbb{E}g_i^\mu\omega_t(Q)=\sqrt{\beta(1-t)/N}\,\mathbb{E}[\omega_t(Q^2)-\omega_t(Q)^2] E g i μ ω t ( Q ) = β ( 1 − t ) / N E [ ω t ( Q 2 ) − ω t ( Q ) 2 ] , while E f ′ ( ε i μ ) = β t / N E [ ω t ( Q 2 ) − ω t ( Q ) 2 ] \mathbb{E}f'(\varepsilon_i^\mu)=\sqrt{\beta t/N}\,\mathbb{E}[\omega_t(Q^2)-\omega_t(Q)^2] E f ′ ( ε i μ ) = β t / N E [ ω t ( Q 2 ) − ω t ( Q ) 2 ] . The variance terms cancel, and (27) gives
∣ q S ′ ( t ) ∣ ≤ 26 3 β 2 t N 3 e 4 β δ ∑ i = 1 N ∑ μ ∉ S E ω t ( z μ 4 ) . |q_S'(t)|\le\frac{26}{3}\frac{\beta^2t}{N^3}e^{4\sqrt{\beta}\,\delta}\sum_{i=1}^{N}\sum_{\mu\notin S}\mathbb{E}\omega_t(z_\mu^4). ∣ q S ′ ( t ) ∣ ≤ 3 26 N 3 β 2 t e 4 β δ i = 1 ∑ N μ ∈ / S ∑ E ω t ( z μ 4 ) .
All terms in this sum use the same interpolating Gibbs measure, and on D S \mathcal{D}_S D S , ∑ μ ∉ S z μ 4 ≤ δ 2 N ∑ μ ∉ S z μ 2 ≤ δ 2 ϱ N 2 \sum_{\mu\notin S}z_\mu^4\le\delta^2N\sum_{\mu\notin S}z_\mu^2\le\delta^2\varrho N^2 ∑ μ ∈ / S z μ 4 ≤ δ 2 N ∑ μ ∈ / S z μ 2 ≤ δ 2 ϱ N 2 . Hence ∣ q S ′ ( t ) ∣ ≤ 26 3 β 2 t ϱ δ 2 e 4 β δ |q_S'(t)|\le\frac{26}{3}\beta^2t\varrho\delta^2e^{4\sqrt{\beta}\,\delta} ∣ q S ′ ( t ) ∣ ≤ 3 26 β 2 tϱ δ 2 e 4 β δ , and integration in t t t gives (26) . The differentiations are justified as follows. By (29) below, Φ S \Phi_S Φ S is β ϱ / N \sqrt{\beta\varrho/N} β ϱ / N -Lipschitz in the Hilbert–Schmidt norm, so q S q_S q S is continuous on [ 0 , 1 ] [0,1] [ 0 , 1 ] . On [ t 0 , t 1 ] ⊂ ( 0 , 1 ) [t_0,t_1]\subset(0,1) [ t 0 , t 1 ] ⊂ ( 0 , 1 ) , ∣ ∂ t Φ S ( J ( t ) ) ∣ ≤ β ϱ / N ( ∥ ε ∥ H S + ∥ G ∥ H S ) / ( 2 t ( 1 − t ) ) |\partial_t\Phi_S(J(t))|\le\sqrt{\beta\varrho/N}(\|\varepsilon\|_{\mathrm{HS}}+\|G\|_{\mathrm{HS}})/(2\sqrt{t(1-t)}) ∣ ∂ t Φ S ( J ( t )) ∣ ≤ β ϱ / N ( ∥ ε ∥ HS + ∥ G ∥ HS ) / ( 2 t ( 1 − t ) ) , an integrable bound that permits differentiation under the expectation. After the cancellation, ∣ q S ′ ∣ |q_S'| ∣ q S ′ ∣ is bounded on ( 0 , 1 ) (0,1) ( 0 , 1 ) , so q S ( 1 ) − q S ( 0 ) = ∫ 0 1 q S ′ ( t ) d t q_S(1)-q_S(0)=\int_0^1q_S'(t)\,dt q S ( 1 ) − q S ( 0 ) = ∫ 0 1 q S ′ ( t ) d t . □
Dropping the sector constraints at the Gaussian endpoint increases the partition function. Since the binary columns are i.i.d., the Gaussian columns are i.i.d. and γ ⊗ K \gamma^{\otimes K} γ ⊗ K is permutation invariant, relabelling the columns identifies the unconstrained model with binary columns indexed by S S S with the mixed model (4) with r = ∣ S ∣ r=|S| r = ∣ S ∣ . The uniform field is unchanged; the reference column is not a field source. Therefore
E Φ S ( ε ) ≤ p N , K ( ∣ S ∣ ) ( h ) + q δ . (28) \mathbb{E}\Phi_S(\varepsilon)\le p_{N,K}^{(|S|)}(h)+q_\delta. \tag*{(28)} E Φ S ( ε ) ≤ p N , K ( ∣ S ∣ ) ( h ) + q δ . (28)
Remark 6.3 (Why the replacement is collective). Under the joint Gibbs measure ω j o i n t \omega_{\mathrm{joint}} ω joint of ( σ , z ) (\sigma,z) ( σ , z ) on Σ N × R K \Sigma_N \times\mathbb{R}^K Σ N × R K , without truncation, conditionally on σ \sigma σ one has z μ = β N m μ + Y μ z_\mu=\sqrt{\beta N}\,m_\mu+Y_\mu z μ = β N m μ + Y μ with m μ : = N − 1 ε μ ⋅ σ m_\mu:=N^{-1}\varepsilon^\mu\cdot\sigma m μ := N − 1 ε μ ⋅ σ and ( Y μ ) (Y_\mu) ( Y μ ) independent standard Gaussian. A large hidden coordinate does not by itself imply a large overlap, but for fixed β , δ > 0 \beta,\delta>0 β , δ > 0 ,
ω j o i n t ( ∃ μ : ∣ z μ ∣ > δ N , ∣ m μ ∣ ≤ δ 2 β ) ≤ 2 K e − δ 2 N / 8 . \omega_{\mathrm{joint}}\left(\exists\mu:\ |z_\mu|>\delta\sqrt{N},\ |m_\mu|\leq\frac{\delta}{2\sqrt{\beta}}\right)\leq2K e^{-\delta^2N/8}. ω joint ( ∃ μ : ∣ z μ ∣ > δ N , ∣ m μ ∣ ≤ 2 β δ ) ≤ 2 K e − δ 2 N /8 .
since the event forces ∣ Y μ ∣ > δ N / 2 |Y_\mu|>\delta\sqrt{N}/2 ∣ Y μ ∣ > δ N /2 for some μ \mu μ . In this sense the sector S S S retains the patterns associated with potentially macroscopic overlaps; their number is at most M δ M_\delta M δ , and the proof uses neither their identity nor any statement about condensed equilibrium states. For the columns outside S S S only ∣ z μ ∣ ≤ δ N |z_\mu|\leq\delta\sqrt{N} ∣ z μ ∣ ≤ δ N is known. Using this bound entry by entry, as in Lemma 5.2 , would cost 13 4 β 2 δ 4 / N \frac{13}{4}\beta^2\delta^4/N 4 13 β 2 δ 4 / N per entry and 13 4 β 2 δ 4 K \frac{13}{4}\beta^2\delta^4K 4 13 β 2 δ 4 K in total, which diverges. The proof of Lemma 6.2 keeps one interpolating Gibbs measure for all replaced entries and sums the fourth moments before bounding them; the radial cutoff then makes ∑ μ ∉ S z μ 4 \sum_{\mu\notin S}z_\mu^4 ∑ μ ∈ / S z μ 4 of order δ 2 N 2 \delta^2N^2 δ 2 N 2 instead of δ 4 N 3 \delta^4N^3 δ 4 N 3 .
Sector selection
Every Φ S \Phi_S Φ S is convex in J ∈ R N × K J\in\mathbb{R}^{N\times K} J ∈ R N × K , as a normalized log-Laplace transform of linear functions of J J J , and the radial bound gives the global Lipschitz estimate
∥ ∇ Φ S ( J ) ∥ H S ≤ L N s e c : = β ϱ / N . (29) \|\nabla\Phi_S(J)\|_{\mathrm{HS}}\leq L_N^{\mathrm{sec}}:=\sqrt{\beta\varrho/N}. \tag*{(29)} ∥∇ Φ S ( J ) ∥ HS ≤ L N sec := β ϱ / N . (29)
Indeed, ∂ Φ S / ∂ J i μ = N − 1 β / N ω ( σ i z μ ) \partial\Phi_S/\partial J_i^\mu=N^{-1}\sqrt{\beta/N}\,\omega(\sigma_i z_\mu) ∂ Φ S / ∂ J i μ = N − 1 β / N ω ( σ i z μ ) , so ∥ ∇ Φ S ∥ H S 2 ≤ β N − 3 ∑ i , μ ω ( z μ 2 ) = β N − 2 ω ∥ z ∥ 2 ≤ β ϱ / N \|\nabla\Phi_S\|_{\mathrm{HS}}^2\leq\beta N^{-3}\sum_{i,\mu}\omega(z_\mu^2)=\beta N^{-2}\omega\|z\|^2\leq\beta\varrho/N ∥∇ Φ S ∥ HS 2 ≤ β N − 3 ∑ i , μ ω ( z μ 2 ) = β N − 2 ω ∥ z ∥ 2 ≤ β ϱ / N . Since Z R , r a d = ∑ S ∈ S N Z S ( ε ) Z^{R,\mathrm{rad}}=\sum_{S\in\mathcal{S}_N}Z_S(\varepsilon) Z R , rad = ∑ S ∈ S N Z S ( ε ) , Lemma 3.8 gives
1 N E log Z R , r a d ≤ E max S Φ S ( ε ) + log ∣ S N ∣ N ≤ max S E Φ S ( ε ) + 2 2 L N s e c log ( ∣ S N ∣ + 1 ) + log ∣ S N ∣ N . \frac{1}{N}\mathbb{E}\log Z^{R,\mathrm{rad}}\leq\mathbb{E}\max_S\Phi_S(\varepsilon)+\frac{\log|\mathcal{S}_N|}{N}\leq\max_S\mathbb{E}\Phi_S(\varepsilon)+2\sqrt{2}L_N^{\mathrm{sec}}\sqrt{\log(|\mathcal{S}_N|+1)}+\frac{\log|\mathcal{S}_N|}{N}. N 1 E log Z R , rad ≤ E S max Φ S ( ε ) + N log ∣ S N ∣ ≤ S max E Φ S ( ε ) + 2 2 L N sec log ( ∣ S N ∣ + 1 ) + N log ∣ S N ∣ .
Combining {(citep)}(6.4), {(citep)}(6.1) and Lemma 6.1 , we obtain, for each fixed δ > 0 \delta>0 δ > 0 ,
A N , K R ( h ) ≤ max 1 ≤ r ≤ min ( K , M δ ) p N , K ( r ) ( h ) + q δ + O ϱ , δ , β , α + ( log N / N ) . (30) A^R_{N,K}(h)\leq\max_{1\leq r\leq\min(K,M_\delta)}p^{(r)}_{N,K}(h)+q_\delta+O_{\varrho,\delta,\beta,\alpha_+}\left(\sqrt{\log N/N}\right). \tag*{(30)} A N , K R ( h ) ≤ 1 ≤ r ≤ m i n ( K , M δ ) max p N , K ( r ) ( h ) + q δ + O ϱ , δ , β , α + ( log N / N ) . (30)
uniformly in h h h .
Binary patterns: conclusion of the proof
Proof of Theorems 2.1 for binary patterns and 2.2. At β = 0 \beta=0 β = 0 all quantities equal log ( 2 cosh h ) \log(2\cosh h) log ( 2 cosh h ) ; let β > 0 \beta>0 β > 0 . Fix α , β , h \alpha,\beta,h α , β , h , let α + : = sup N K ( N ) / N \alpha_+:=\sup_N K(N)/N α + := sup N K ( N ) / N and choose ϱ \varrho ϱ by Lemma 6.1 . For every fixed δ > 0 \delta>0 δ > 0 , M δ M_\delta M δ is finite, and K ( N ) ≥ M δ K(N)\geq M_\delta K ( N ) ≥ M δ for all large N N N since α > 0 \alpha>0 α > 0 . Proposition 5.1 , applied to each of the finitely many r ≤ M δ r\leq M_\delta r ≤ M δ , and {(citep)}(6.6) imply
lim sup N A N , K ( N ) R ( h ) ≤ max 1 ≤ r ≤ M δ p r ( α , β , h ) + q δ ≤ A ∗ + q δ . \limsup_N A^R_{N,K(N)}(h)\leq\max_{1\leq r\leq M_\delta}p_r(\alpha,\beta,h)+q_\delta\leq A_*+q_\delta. N lim sup A N , K ( N ) R ( h ) ≤ 1 ≤ r ≤ M δ max p r ( α , β , h ) + q δ ≤ A ∗ + q δ .
Together with {(citep)}(5.3) and q δ → 0 q_\delta\to0 q δ → 0 as δ ↓ 0 \delta\downarrow0 δ ↓ 0 ,
A ∗ ≤ lim inf N A N , K ( N ) R ( h ) ≤ lim sup N A N , K ( N ) R ( h ) ≤ A ∗ , A_*\leq\liminf_N A^R_{N,K(N)}(h)\leq\limsup_N A^R_{N,K(N)}(h)\leq A_*, A ∗ ≤ N lim inf A N , K ( N ) R ( h ) ≤ N lim sup A N , K ( N ) R ( h ) ≤ A ∗ ,
which proves {(citep)}(2.6). The order of limits is: fix ϱ \varrho ϱ and δ \delta δ , let N → ∞ N\to\infty N → ∞ , then δ ↓ 0 \delta\downarrow0 δ ↓ 0 . Convergence need not be uniform in r r r , and M δ M_\delta M δ need not stay bounded as δ ↓ 0 \delta\downarrow0 δ ↓ 0 . Sequence independence follows from that of every p r p_r p r (Proposition 5.1 ); monotonicity in r r r is Remark 5.3 , and boundedness is in Proposition 5.1 , so sup r p r = lim r p r \sup_r p_r=\lim_r p_r sup r p r = lim r p r . Finally, A N , K R A^R_{N,K} A N , K R is 1-Lipschitz in h h h and ∣ ∂ β A N , K R ∣ ≤ E ∥ ε ∥ o p 2 / ( 2 N ) ≤ C ( 1 + α + ) |\partial_\beta A^R_{N,K}|\leq\mathbb{E}\|\varepsilon\|_{\mathrm{op}}^2/(2N)\leq C(1+\alpha_+) ∣ ∂ β A N , K R ∣ ≤ E ∥ ε ∥ op 2 / ( 2 N ) ≤ C ( 1 + α + ) by Lemmas 3.5 and 3.3. These bounds hold for all β ≥ 0 \beta\geq0 β ≥ 0 , with a common Lipschitz constant L L L for ∣ β − β ′ ∣ + ∣ h − h ′ ∣ |\beta-\beta'|+|h-h'| ∣ β − β ′ ∣ + ∣ h − h ′ ∣ ; passing to the pointwise limit gives the same bound for A R A^{R} A R . Let C ⊂ [ 0 , ∞ ) × R \mathcal{C} \subset[0,\infty) \times\mathbb{R} C ⊂ [ 0 , ∞ ) × R be compact, and choose a finite θ \theta θ -net T θ ⊂ C T_{\theta} \subset\mathcal{C} T θ ⊂ C . For every parameter pair, comparison with a net point gives
sup ( β , h ) ∈ C ∣ A N , K ( N ) R ( β , h ) − A R ( α , β , h ) ∣ ≤ 2 L θ + max ( β , h ) ∈ T θ ∣ A N , K ( N ) R ( β , h ) − A R ( α , β , h ) ∣ . \sup_{(\beta,h)\in\mathcal{C}}\left|A^{R}_{N,K(N)}(\beta,h)-A^{R}(\alpha,\beta,h)\right| \le2L\theta+\max_{(\beta,h)\in T_{\theta}}\left|A^{R}_{N,K(N)}(\beta,h)-A^{R}(\alpha,\beta,h)\right|. ( β , h ) ∈ C sup A N , K ( N ) R ( β , h ) − A R ( α , β , h ) ≤ 2 L θ + ( β , h ) ∈ T θ max A N , K ( N ) R ( β , h ) − A R ( α , β , h ) .
At fixed θ \theta θ , the finite maximum tends to zero. Taking the limsup in N N N and then letting θ ↓ 0 \theta\downarrow0 θ ↓ 0 proves local uniformity. At β = 0 \beta=0 β = 0 , both pressures equal log ( 2 cosh h ) \log(2\cosh h) log ( 2 cosh h ) , and the same Lipschitz bounds hold up to the boundary. This proves the uniformity assertions of Theorems 2.1 and 2.2. □
Concentration and the almost-sure thermodynamic limit
The convergence of the quenched means has now been established for both pattern laws. We control sample fluctuations to obtain almost-sure convergence, including local uniformity in the parameters.
Proposition 8.1 (Explicit concentration bounds). Let β > 0 \beta>0 β > 0 , α + > 0 \alpha_{+}>0 α + > 0 , N , K ≥ 1 N,K\ge1 N , K ≥ 1 with K ≤ α + N K\le\alpha_{+}N K ≤ α + N , and let V : Σ N → R V:\Sigma_{N}\to\mathbb{R} V : Σ N → R be deterministic.
(a) Gaussian patterns. Let δ N : = e − N / 4 ( 12 c V 2 + 120 β 2 ( 1 + α + ) 2 + 24 β 2 α + ( 1 + α + ) N ) 1 / 2 \delta_{N}:=e^{-N/4}\left(12c_{V}^{2}+120\beta^{2}(1+\alpha_{+})^{2}+24\beta^{2}\alpha_{+}(1+\alpha_{+})N\right)^{1/2} δ N := e − N /4 ( 12 c V 2 + 120 β 2 ( 1 + α + ) 2 + 24 β 2 α + ( 1 + α + ) N ) 1/2 . For every t > 0 t>0 t > 0 ,
P ( ∣ f N V ( G ) − E f N V ( G ) ∣ ≥ t + δ N ) ≤ 2 exp { − N t 2 8 β 2 ( 1 + α + ) 2 } + e − N / 2 . (31) \mathbb{P}\left(\left|f_{N}^{V}(G)-\mathbb{E}f_{N}^{V}(G)\right|\ge t+\delta_{N}\right)\le2\exp\left\{-\frac{Nt^{2}}{8\beta^{2}(1+\sqrt{\alpha_{+}})^{2}}\right\}+e^{-N/2}. \tag*{(31)} P ( f N V ( G ) − E f N V ( G ) ≥ t + δ N ) ≤ 2 exp { − 8 β 2 ( 1 + α + ) 2 N t 2 } + e − N /2 . (31) (b) Rademacher patterns. Let c ⋆ : = ( 8 { 1 + ( 1 + α + ) log 9 } ) 1 / 2 c_{\star}:=\left(8\{1+(1+\alpha_{+})\log9\}\right)^{1/2} c ⋆ := ( 8 { 1 + ( 1 + α + ) log 9 } ) 1/2 , L N : = β c ⋆ / N L_{N}:=\beta c_{\star}/\sqrt{N} L N := β c ⋆ / N and d N : = ( 4 c V + β K ) e − N d_{N}:=(4c_{V}+\beta K)e^{-N} d N := ( 4 c V + β K ) e − N . For every t > 0 t>0 t > 0 ,
P ( ∣ f N V ( ε ) − E f N V ( ε ) ∣ > t + d N + 8 π L N ) ≤ 2 e − N + 4 exp { − N t 2 16 β 2 c ⋆ 2 } . (32) \mathbb{P}\left(\left|f_{N}^{V}(\varepsilon)-\mathbb{E}f_{N}^{V}(\varepsilon)\right|>t+d_{N}+8\sqrt{\pi}L_{N}\right)\le2e^{-N}+4\exp\left\{-\frac{Nt^{2}}{16\beta^{2}c_{\star}^{2}}\right\}. \tag*{(32)} P ( f N V ( ε ) − E f N V ( ε ) > t + d N + 8 π L N ) ≤ 2 e − N + 4 exp { − 16 β 2 c ⋆ 2 N t 2 } . (32)
Proof of Proposition 8.1(a). Let f : = f N V ( G ) f:=f_{N}^{V}(G) f := f N V ( G ) , R : = 2 ( N + K ) R:=2(\sqrt{N}+\sqrt{K}) R := 2 ( N + K ) , Λ : = β R / N \Lambda:=\beta R/N Λ := β R / N and B : = { ∥ G ∥ op ≤ R } \mathcal{B}:=\{\lVert G\rVert_{\operatorname{op}}\le R\} B := {∥ G ∥ op ≤ R } . Since B \mathcal{B} B is convex, (13) and the mean value theorem show that f f f is Λ \Lambda Λ -Lipschitz on B \mathcal{B} B for the Hilbert–Schmidt norm. Let g g g be its McShane extension (T3). As Λ 2 ≤ 4 β 2 ( 1 + α + ) 2 / N \Lambda^{2}\le4\beta^{2}(1+\sqrt{\alpha_{+}})^{2}/N Λ 2 ≤ 4 β 2 ( 1 + α + ) 2 / N , (T2) gives P ( ∣ g − E g ∣ ≥ t ) ≤ 2 exp { − N t 2 / ( 8 β 2 ( 1 + α + ) 2 ) } \mathbb{P}(|g-\mathbb{E}g|\ge t)\le2\exp\{-Nt^{2}/(8\beta^{2}(1+\sqrt{\alpha_{+}})^{2})\} P ( ∣ g − E g ∣ ≥ t ) ≤ 2 exp { − N t 2 / ( 8 β 2 ( 1 + α + ) 2 )} . By Lemma 3.2, P ( B c ) ≤ e − ( N + K ) 2 / 2 ≤ e − N / 2 \mathbb{P}(\mathcal{B}^{c})\le e^{-(\sqrt{N}+\sqrt{K})^{2}/2}\le e^{-N/2} P ( B c ) ≤ e − ( N + K ) 2 /2 ≤ e − N /2 . Since g ( 0 ) = f ( 0 ) g(0)=f(0) g ( 0 ) = f ( 0 ) and ∣ f ( 0 ) ∣ ≤ c V |f(0)|\le c_{V} ∣ f ( 0 ) ∣ ≤ c V , we have ∣ g ∣ ≤ c V + Λ ∥ G ∥ H S |g|\le c_{V}+\Lambda\lVert G\rVert_{\mathrm{HS}} ∣ g ∣ ≤ c V + Λ ∥ G ∥ HS , so, by (13) ,
E ( f − g ) 2 ≤ 12 c V 2 + 3 β 2 4 N 2 E ∥ G ∥ op 4 + 3 Λ 2 E ∥ G ∥ H S 2 ≤ 12 c V 2 + 120 β 2 ( 1 + α + ) 2 + 24 β 2 α + ( 1 + α + ) N , \mathbb{E}(f-g)^{2}\le12c_{V}^{2}+\frac{3\beta^{2}}{4N^{2}}\mathbb{E}\lVert G\rVert_{\operatorname{op}}^{4}+3\Lambda^{2}\mathbb{E}\lVert G\rVert_{\mathrm{HS}}^{2}\le12c_{V}^{2}+120\beta^{2}(1+\alpha_{+})^{2}+24\beta^{2}\alpha_{+}(1+\alpha_{+})N, E ( f − g ) 2 ≤ 12 c V 2 + 4 N 2 3 β 2 E ∥ G ∥ op 4 + 3 Λ 2 E ∥ G ∥ HS 2 ≤ 12 c V 2 + 120 β 2 ( 1 + α + ) 2 + 24 β 2 α + ( 1 + α + ) N ,
using E ∥ G ∥ H S 2 = N K \mathbb{E}\lVert G\rVert_{\mathrm{HS}}^{2}=NK E ∥ G ∥ HS 2 = N K and Λ 2 ≤ 8 β 2 ( N + K ) / N 2 \Lambda^{2}\le8\beta^{2}(N+K)/N^{2} Λ 2 ≤ 8 β 2 ( N + K ) / N 2 . As f = g f=g f = g on B \mathcal{B} B , Cauchy–Schwarz gives ∣ E f − E g ∣ ≤ ( E ( f − g ) 2 ) 1 / 2 P ( B c ) 1 / 2 ≤ δ N |\mathbb{E}f-\mathbb{E}g|\le(\mathbb{E}(f-g)^{2})^{1/2}\mathbb{P}(\mathcal{B}^{c})^{1/2}\le\delta_{N} ∣ E f − E g ∣ ≤ ( E ( f − g ) 2 ) 1/2 P ( B c ) 1/2 ≤ δ N . Hence { ∣ f − E f ∣ ≥ t + δ N } ⊂ B c ∪ { ∣ g − E g ∣ ≥ t } \{|f-\mathbb{E}f|\ge t+\delta_{N}\}\subset\mathcal{B}^{c}\cup\{|g-\mathbb{E}g|\ge t\} { ∣ f − E f ∣ ≥ t + δ N } ⊂ B c ∪ { ∣ g − E g ∣ ≥ t } , which proves (31) . □
The bound has the form exp ( − c N t 2 ) \exp(-cNt^{2}) exp ( − c N t 2 ) of the concentration theorem in [31 ] . The term e − N / 2 e^{-N/2} e − N /2 and the exponentially small shift δ N \delta_{N} δ N account for the unbounded gradient of the statistical pressure.
Proof of Proposition 8.1(b). Let f : = f N V f:=f_{N}^{V} f := f N V on R N × K \mathbb{R}^{N\times K} R N × K and K N : = { J : ∥ J ∥ op ≤ c ⋆ N } \mathcal{K}_{N}:=\{J:\lVert J\rVert_{\operatorname{op}}\le c_{\star}\sqrt{N}\} K N := { J : ∥ J ∥ op ≤ c ⋆ N } , a compact convex set on which (13) bounds the gradient by L N = β c ⋆ / N L_{N}=\beta c_{\star}/\sqrt{N} L N = β c ⋆ / N . Define the convex extension by supporting planes
f ~ ( J ) : = sup W ∈ K N { f ( W ) + ⟨ ∇ f ( W ) , J − W ⟩ H S } . \widetilde{f}(J):=\sup_{W\in\mathcal{K}_{N}}\{f(W)+\langle\nabla f(W),J-W\rangle_{\mathrm{HS}}\}. f ( J ) := W ∈ K N sup { f ( W ) + ⟨ ∇ f ( W ) , J − W ⟩ HS } .
It is finite, convex and L N L_{N} L N -Lipschitz; convexity of f f f implies f ~ ≤ f \widetilde{f}\le f f ≤ f everywhere, with equality on K N \mathcal{K}_{N} K N . Since 0 ∈ K N 0\in\mathcal{K}_{N} 0 ∈ K N and ∇ f ( 0 ) = 0 \nabla f(0)=0 ∇ f ( 0 ) = 0 (Lemma 3.4), f ~ ≥ f ( 0 ) ≥ log 2 − N − 1 max ∣ V ∣ \widetilde{f}\ge f(0)\ge\log2-N^{-1}\max|V| f ≥ f ( 0 ) ≥ log 2 − N − 1 max ∣ V ∣ . On the binary cube ∥ ε ∥ o p 2 ≤ N K \|\varepsilon\|_{\mathrm{op}}^2 \le NK ∥ ε ∥ op 2 ≤ N K , so f ( ε ) ≤ c V + β K / 2 f(\varepsilon) \le c_V+\beta K/2 f ( ε ) ≤ c V + β K /2 by (13) , and 0 ≤ f ( ε ) − f ~ ( ε ) ≤ 2 c V + β K / 2 0 \le f(\varepsilon)-\widetilde f(\varepsilon) \le2c_V+\beta K/2 0 ≤ f ( ε ) − f ( ε ) ≤ 2 c V + β K /2 . By Lemma 3.3, P ( ε ∉ K N ) ≤ 2 e − N \mathbb{P}(\varepsilon\notin K_N)\le2e^{-N} P ( ε ∈ / K N ) ≤ 2 e − N , hence 0 ≤ E f ( ε ) − E f ~ ( ε ) ≤ d N 0\le\mathbb{E}f(\varepsilon)-\mathbb{E}\widetilde f(\varepsilon)\le d_N 0 ≤ E f ( ε ) − E f ( ε ) ≤ d N . If m f ∼ m_f^{\sim} m f ∼ is a median of f ~ ( ε ) \widetilde f(\varepsilon) f ( ε ) , Lemma 3.9 gives ∣ E f ~ − m f ∼ ∣ ≤ 8 π L N |\mathbb{E}\widetilde f-m_f^{\sim}|\le8\sqrt{\pi}L_N ∣ E f − m f ∼ ∣ ≤ 8 π L N and
P ( ∣ f − E f ∣ > t + d N + 8 π L N ) ≤ P ( ε ∉ K N ) + P ( ∣ f ~ − m f ∼ ∣ > t ) ≤ 2 e − N + 4 e − t 2 / ( 16 L N 2 ) . \mathbb{P}\left(|f-\mathbb{E}f|>t+d_N+8\sqrt{\pi}L_N\right)\le\mathbb{P}(\varepsilon\notin K_N)+\mathbb{P}\left(|\widetilde f-m_f^{\sim}|>t\right)\le2e^{-N}+4e^{-t^2/(16L_N^2)}. P ( ∣ f − E f ∣ > t + d N + 8 π L N ) ≤ P ( ε ∈ / K N ) + P ( ∣ f − m f ∼ ∣ > t ) ≤ 2 e − N + 4 e − t 2 / ( 16 L N 2 ) .
which is (32) .
Proof of Theorem 2.3. (a) In the Gaussian case, c V ≤ c ‾ c_V\le\overline c c V ≤ c gives δ N ≤ e − N / 4 ( 12 c ‾ 2 + 120 β 2 ( 1 + α + ) 2 + 24 β 2 α + ( 1 + α + ) N ) 1 / 2 → 0 \delta_N\le e^{-N/4}(12\overline c^2+120\beta^2(1+\alpha_+)^2+24\beta^2\alpha_+(1+\alpha_+)N)^{1/2}\to0 δ N ≤ e − N /4 ( 12 c 2 + 120 β 2 ( 1 + α + ) 2 + 24 β 2 α + ( 1 + α + ) N ) 1/2 → 0 ; in the binary case d N + 8 π L N ≤ ( 4 c ‾ + β α + N ) e − N + 8 π β c ⋆ N − 1 / 2 → 0 d_N+8\sqrt{\pi}L_N\le(4\overline c+\beta\alpha_+N)e^{-N}+8\sqrt{\pi\beta c_\star}N^{-1/2}\to0 d N + 8 π L N ≤ ( 4 c + β α + N ) e − N + 8 π β c ⋆ N − 1/2 → 0 . For N ≥ N 0 N\ge N_0 N ≥ N 0 these shifts are at most t / 2 t/2 t /2 , and Proposition 8.1 with t / 2 t/2 t /2 in place of t t t bounds the probability by 2 e − c N + e − N / 2 2e^{-cN}+e^{-N/2} 2 e − c N + e − N /2 , respectively 2 e − N + 4 e − c N 2e^{-N}+4e^{-cN} 2 e − N + 4 e − c N , with c > 0 c>0 c > 0 depending on β , α + , t \beta,\alpha_+,t β , α + , t . For N < N 0 N<N_0 N < N 0 the probability is at most 1 ≤ e c N 0 e − c N 1\le e^{cN_0}e^{-cN} 1 ≤ e c N 0 e − c N .
(b) If β = 0 \beta=0 β = 0 , then A ^ N , K ∙ ( h ) = log ( 2 cosh h ) \widehat A^\bullet_{N,K}(h)=\log(2\cosh h) A N , K ∙ ( h ) = log ( 2 cosh h ) is deterministic; the finitely many N N N with K ( N ) = 0 K(N)=0 K ( N ) = 0 are also deterministic. Let α + : = sup N K ( N ) / N \alpha_+:=\sup_N K(N)/N α + := sup N K ( N ) / N . By (13) , (T1) and Lemma 3.2 in the Gaussian case, and by Lemmas 3.7 and 3.3 in the binary case, Var f N V ( J ) ≤ 2 β 2 N − 2 E ∥ J ∥ o p 2 ≤ C α + , β / N \operatorname{Var}f_N^V(\mathcal{J})\le2\beta^2N^{-2}\mathbb{E}\|\mathcal{J}\|_{\mathrm{op}}^2\le C_{\alpha_+,\beta}/N Var f N V ( J ) ≤ 2 β 2 N − 2 E ∥ J ∥ op 2 ≤ C α + , β / N for every V V V ; with V = h N m V=hN\mathfrak{m} V = h N m and Theorem 2.1 this gives L 2 L^2 L 2 convergence.
For each rational pair ( β , h ) (\beta,h) ( β , h ) with β > 0 \beta>0 β > 0 and each t > 0 t>0 t > 0 , the right-hand sides of (31) and (32) with V = h N m V=hN\mathfrak{m} V = h N m are summable in N N N , and δ N , d N , L N → 0 \delta_N,d_N,L_N\to0 δ N , d N , L N → 0 . By (T4) and Theorem 2.1, A ^ N , K ( N ) ∙ ( β , h ) → A ∙ ( α , β , h ) \widehat A^\bullet_{N,K(N)}(\beta,h)\to A^\bullet(\alpha,\beta,h) A N , K ( N ) ∙ ( β , h ) → A ∙ ( α , β , h ) almost surely, simultaneously for all rational pairs. By Lemmas 3.2 and 3.3 and (T4), almost surely ∥ J N , K ( N ) ∙ ∥ o p ≤ c N \|\mathcal{J}^{\bullet}_{N,K(N)}\|_{\mathrm{op}}\le c\sqrt N ∥ J N , K ( N ) ∙ ∥ op ≤ c N for all large N N N , with c : = max { 2 ( 1 + α + ) , c ⋆ } c:=\max\{2(1+\sqrt{\alpha_+}),c_\star\} c := max { 2 ( 1 + α + ) , c ⋆ } . On this event, by Lemma 3.5, ∣ ∂ h A ^ N , K ∙ ∣ ≤ 1 |\partial_h\widehat A^\bullet_{N,K}|\le1 ∣ ∂ h A N , K ∙ ∣ ≤ 1 and 0 ≤ ∂ β A ^ N , K ∙ ≤ c 2 / 2 0\le\partial_\beta\widehat A^\bullet_{N,K}\le c^2/2 0 ≤ ∂ β A N , K ∙ ≤ c 2 /2 for all parameters and all large N N N , while the limits are locally Lipschitz by the same bounds on the means. Approximation by finite rational nets gives convergence locally uniformly on [ 0 , ∞ ) × R [0,\infty)\times\mathbb{R} [ 0 , ∞ ) × R . No independence between volumes has been used. Finally, on a compact C ⊂ [ 0 , ∞ ) × R C\subset[0,\infty)\times\mathbb{R} C ⊂ [ 0 , ∞ ) × R , sup C ∣ A ^ N , K ∙ − A ∙ ∣ ≤ C C ( 1 + ∥ J ∥ o p 2 / N ) \sup_C|\widehat A^\bullet_{N,K}-A^\bullet|\le C_C(1+\|\mathcal{J}\|_{\mathrm{op}}^2/N) sup C ∣ A N , K ∙ − A ∙ ∣ ≤ C C ( 1 + ∥ J ∥ op 2 / N ) by Lemma 3.5, a family bounded in L 2 L^2 L 2 ; uniform integrability and almost-sure convergence give L 1 L^1 L 1 convergence.
Internal energy, entropy and ground states
Convexity transfers the pressure limits to internal energy and entropy at differentiability points. The entropy bound then controls the zero-temperature limit.
Proof of Theorem 2.4. Write A ^ N : = A ^ N , K ( N ) ∙ ( ⋅ , 0 ) \widehat A_N:=\widehat A^\bullet_{N,K(N)}(\mathord\cdot,0) A N := A N , K ( N ) ∙ ( ⋅ , 0 ) , A N : = E A ^ N A_N:=\mathbb{E}\widehat A_N A N := E A N , A : = A ∙ ( α , ⋅ ) A:=A^\bullet(\alpha,\mathord\cdot) A := A ∙ ( α , ⋅ ) , u N : = u N ∙ u_N:=u_N^\bullet u N := u N ∙ , etc., and C N : = ∥ J ∥ o p 2 / ( 2 N ) C_N:=\|\mathcal{J}\|_{\mathrm{op}}^2/(2N) C N := ∥ J ∥ op 2 / ( 2 N ) , so that sup N E C N q < ∞ \sup_N\mathbb{E}C_N^q<\infty sup N E C N q < ∞ for every q < ∞ q<\infty q < ∞ by Lemmas 3.2 and 3.3.
(i) For fixed disorder, β ↦ A ^ N ( β ) \beta\mapsto\widehat A_N(\beta) β ↦ A N ( β ) is smooth and convex with ∂ β A ^ N = − u N \partial_\beta\widehat A_N=-u_N ∂ β A N = − u N and ∂ β 2 A ^ N = N − 1 Var ω β ( H N ) ≥ 0 \partial_\beta^2\widehat A_N=N^{-1}\operatorname{Var}_{\omega_\beta}(H_N)\ge0 ∂ β 2 A N = N − 1 Var ω β ( H N ) ≥ 0 (Lemma 3.5). Since ∣ u N ∣ ≤ C N |u_N|\le C_N ∣ u N ∣ ≤ C N is integrable, dominated convergence gives ∂ β A N = − E N \partial_\beta A_N=-E_N ∂ β A N = − E N , continuous in β \beta β . From log ω β = − β H N − log Z N \log\omega_\beta=-\beta H_N-\log Z_N log ω β = − β H N − log Z N ,
(9.1)
s N = A ^ N + β u N , 0 ≤ s N ≤ log 2 , s_N=\widehat A_N+\beta u_N,\qquad0\le s_N\le\log2, s N = A N + β u N , 0 ≤ s N ≤ log 2 ,
the bounds because N s N Ns_N N s N is the Shannon entropy of a probability measure on 2 N 2^N 2 N points; taking expectations gives A N = − β E N + S N A_N=-\beta E_N+S_N A N = − β E N + S N .
(ii)–(iii) A A A is convex as a pointwise limit of convex functions, hence differentiable outside a countable set ((T5)). Lemma 3.1 with f N = A N f_N=A_N f N = A N gives the one-sided bounds in (iii) for E N E_N E N and E N ( β ) → − ∂ β A ( β ) E_N(\beta)\to-\partial_\beta A(\beta) E N ( β ) → − ∂ β A ( β ) at differentiability points; then S N = A N + β E N → A + β E ‾ = : S S_N=A_N+\beta E_N\to A+\beta\overline E=:S S N = A N + β E N → A + β E =: S , and the bounds for S N S_N S N follow from those for E N E_N E N . On the event of Theorem 2.3(b), A ^ N → A \widehat A_N\to A A N → A on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , and Lemma 3.1 with f N = A ^ N f_N=\widehat A_N f N = A N gives the same statements for u N u_N u N simultaneously at all β > 0 ; x r e f ‘ ( 9.1 ) ‘ t r a n s f e r s t h e m t o \beta>0; {xref}`(9.1)` transfers them to β > 0 ; x r e f ‘ ( 9.1 ) ‘ t r an s f er s t h e m t o s_N. F o r $ 1 ≤ p < ∞ . For \$1\le p<\infty . F or $1 ≤ p < ∞ , the bound ∣ u N ∣ ≤ C N |u_N|\le C_N ∣ u N ∣ ≤ C N makes ( ∣ u N ∣ p ) (|u_N|^p) ( ∣ u N ∣ p ) uniformly integrable, so almost-sure convergence implies L p L^p L p convergence; s N s_N s N is bounded. (iv) E = − ∂ β A E=-\partial_{\beta}A E = − ∂ β A is nonincreasing since ∂ β A \partial_{\beta}A ∂ β A is nondecreasing. By convexity, − E N ( β ) ≥ ∂ β A N ( 0 ) = 1 2 N 2 E 2 − N ∑ σ ∈ Σ N ∥ ( J ∙ ) T σ ∥ 2 = K ( N ) 2 N -E_N(\beta)\geq\partial_{\beta}A_N(0)=\frac{1}{2N^2}\mathbb{E}2^{-N}\sum_{\sigma\in\Sigma_N}\lVert(J^{\bullet})^{T}\sigma\rVert^2=\frac{K(N)}{2N} − E N ( β ) ≥ ∂ β A N ( 0 ) = 2 N 2 1 E 2 − N ∑ σ ∈ Σ N ∥( J ∙ ) T σ ∥ 2 = 2 N K ( N ) , since 2 − N ∑ σ ∈ Σ N ( J μ ⋅ σ ) 2 = ∥ J μ ∥ 2 2^{-N}\sum_{\sigma\in\Sigma_N}(J^{\mu}\cdot\sigma)^2=\lVert J^{\mu}\rVert^2 2 − N ∑ σ ∈ Σ N ( J μ ⋅ σ ) 2 = ∥ J μ ∥ 2 has expectation N N N ; so E ≤ − α / 2 E\leq-\alpha/2 E ≤ − α /2 . In the Gaussian case E N ≥ − E ∥ G ∥ o p 2 / ( 2 N ) E_N\geq-\mathbb{E}\lVert G\rVert_{\mathrm{op}}^2/(2N) E N ≥ − E ∥ G ∥ op 2 / ( 2 N ) gives E G ≥ − 1 2 ( 1 + α ) 2 E^G\geq-\frac{1}{2}(1+\sqrt{\alpha})^2 E G ≥ − 2 1 ( 1 + α ) 2 by Lemma 3.2 . The bounds on S S S follow from (9.1).
For the additional pattern the same proof applies, using the convexity of β ↦ A ^ N ∙ , M ( β ) \beta\mapsto\widehat{A}^{\bullet,M}_N(\beta) β ↦ A N ∙ , M ( β ) (a log-Laplace transform), Theorem 2.5 and ∣ u N ∙ , M ∣ ≤ C N + 1 2 \lvert u_N^{\bullet,M}\rvert\leq C_N+\frac{1}{2} ∣ u N ∙ , M ∣ ≤ C N + 2 1 .
Remark 9.1 (Endpoints and possible transitions). At β = 0 \beta=0 β = 0 , u N ∙ ( 0 ) = − ∥ J ∙ ∥ H S 2 / ( 2 N 2 ) u_N^{\bullet}(0)=-\lVert J^{\bullet}\rVert_{\mathrm{HS}}^2/(2N^2) u N ∙ ( 0 ) = − ∥ J ∙ ∥ HS 2 / ( 2 N 2 ) and s N ∙ ( 0 ) = log 2 s_N^{\bullet}(0)=\log2 s N ∙ ( 0 ) = log 2 ; their limits are − α / 2 -\alpha/2 − α /2 (almost surely in the Gaussian case, by the moment bound in Remark 2.8) and log 2 \log2 log 2 . If A ∙ ( α , ⋅ ) A^{\bullet}(\alpha,\cdot) A ∙ ( α , ⋅ ) is not differentiable at some β > 0 \beta>0 β > 0 , the internal energy may oscillate within the interval of Theorem 2.4 (iii); such points would correspond to first-order phase transitions. Differentiability everywhere, or a unique limiting energy at a transition, would require a separate proof.
Remark 9.2 (Dimensionless field versus physical field). The uniform field h h h in (3) is not multiplied by β \beta β . At fixed h h h , with u N i n t = N − 1 ω ( H N ) u_N^{\mathrm{int}}=N^{-1}\omega(H_N) u N int = N − 1 ω ( H N ) and m ^ N = ω ( m ) \widehat{m}_N=\omega(m) m N = ω ( m ) , one has s N = A ^ N ∙ ( β , h ) + β u N i n t − h m ^ N s_N=\widehat{A}^{\bullet}_N(\beta,h)+\beta u_N^{\mathrm{int}}-h\widehat{m}_N s N = A N ∙ ( β , h ) + β u N int − h m N , so that S = A ∙ + β E − h ∂ h A ∙ S=A^{\bullet}+\beta E-h\partial_hA^{\bullet} S = A ∙ + β E − h ∂ h A ∙ wherever both partial derivatives exist. To keep the identity S = A + β E S=A+\beta E S = A + β E in a physical field B B B , set h = β B h=\beta B h = β B and include the field in the Hamiltonian, H N B = H N − B N m H_N^B=H_N-BNm H N B = H N − B N m . The statistical pressure A B ∙ ( α , β ) : = A ∙ ( α , β , β B ) A_B^{\bullet}(\alpha,\beta):=A^{\bullet}(\alpha,\beta,\beta B) A B ∙ ( α , β ) := A ∙ ( α , β , β B ) is convex in β \beta β , and the proof of Theorem 2.4 applies verbatim, using ∣ N − 1 ω ( H N B ) ∣ ≤ C N + ∣ B ∣ \lvert N^{-1}\omega(H_N^B)\rvert\leq C_N+\lvert B\rvert ∣ N − 1 ω ( H N B )∣ ≤ C N + ∣ B ∣ : E B = − ∂ β A B ∙ E_B=-\partial_{\beta}A_B^{\bullet} E B = − ∂ β A B ∙ and S B = A B ∙ + β E B S_B=A_B^{\bullet}+\beta E_B S B = A B ∙ + β E B at every differentiability point of A B ∙ A_B^{\bullet} A B ∙ .
Corollary 9.3 (Ground states). Let ∙ ∈ { G , R } \bullet\in\{G,R\} ∙ ∈ { G , R } and define the maximal interaction gain
g ^ N ∙ : = − 1 N min σ ∈ Σ N H N ( σ ; J ∙ ) = 1 2 N 2 max σ ∈ Σ N ∥ ( J ∙ ) T σ ∥ 2 , g N ∙ : = E g ^ N ∙ , \widehat{g}_N^{\bullet}:=-\frac{1}{N}\min_{\sigma\in\Sigma_N}H_N(\sigma;J^{\bullet})=\frac{1}{2N^2}\max_{\sigma\in\Sigma_N}\lVert(J^{\bullet})^{T}\sigma\rVert^2,\qquad g_N^{\bullet}:=\mathbb{E}\widehat{g}_N^{\bullet}, g N ∙ := − N 1 σ ∈ Σ N min H N ( σ ; J ∙ ) = 2 N 2 1 σ ∈ Σ N max ∥( J ∙ ) T σ ∥ 2 , g N ∙ := E g N ∙ , with K = K ( N ) K=K(N) K = K ( N ) , and let C N : = ∥ J ∙ ∥ o p 2 / ( 2 N ) C_N:=\lVert J^{\bullet}\rVert_{\mathrm{op}}^2/(2N) C N := ∥ J ∙ ∥ op 2 / ( 2 N ) . Then g ∙ ( α ) : = lim β → ∞ A ∙ ( α , β ) / β g^{\bullet}(\alpha):=\lim_{\beta\to\infty}A^{\bullet}(\alpha,\beta)/\beta g ∙ ( α ) := lim β → ∞ A ∙ ( α , β ) / β exists and is finite, g ^ N ∙ → g ∙ ( α ) \widehat{g}_N^{\bullet}\to g^{\bullet}(\alpha) g N ∙ → g ∙ ( α ) , and g ^ N G → g ∙ ( α ) \widehat{g}_N^G\to g^{\bullet}(\alpha) g N G → g ∙ ( α ) almost surely and in L 1 L^1 L 1 . The ground-state energy density is − g ∙ ( α ) -g^{\bullet}(\alpha) − g ∙ ( α ) , and g G ( α ) ≤ g R ( α ) g^G(\alpha)\leq g^R(\alpha) g G ( α ) ≤ g R ( α ) .
Proof. Fix α \alpha α and write A ( β ) : = A ∙ ( α , β ) A(\beta):=A^{\bullet}(\alpha,\beta) A ( β ) := A ∙ ( α , β ) . We first establish the zero-temperature limit, then use it to bound the finite-volume ground-state gains. Since A A A is convex and A ( 0 ) = log 2 A(0)=\log2 A ( 0 ) = log 2 , the secant slopes
q ( β ) : = A ( β ) − log 2 β , β > 0 , q(\beta):=\frac{A(\beta)-\log2}{\beta},\qquad\beta>0, q ( β ) := β A ( β ) − log 2 , β > 0 ,
are nondecreasing. Indeed, for 0 < s < t 0<s<t 0 < s < t , convexity gives A ( s ) ≤ ( 1 − s / t ) A ( 0 ) + ( s / t ) A ( t ) A(s)\leq(1-s/t)A(0)+(s/t)A(t) A ( s ) ≤ ( 1 − s / t ) A ( 0 ) + ( s / t ) A ( t ) , hence q ( s ) ≤ q ( t ) q(s)\leq q(t) q ( s ) ≤ q ( t ) . Lemma 3.5 and the matrix moment bounds give 0 ≤ q ( β ) ≤ C ( 1 + α ) 0\leq q(\beta)\leq C(1+\alpha) 0 ≤ q ( β ) ≤ C ( 1 + α ) for all β > 0 \beta>0 β > 0 . Thus q ( β ) q(\beta) q ( β ) has a finite limit, denoted g ∙ ( α ) g^{\bullet}(\alpha) g ∙ ( α ) , and
A ( β ) β = q ( β ) + log 2 β ⟶ g ∙ ( α ) . \frac{A(\beta)}{\beta}=q(\beta)+\frac{\log2}{\beta}\longrightarrow g^{\bullet}(\alpha). β A ( β ) = q ( β ) + β log 2 ⟶ g ∙ ( α ) .
This limit depends on the sequence ( K ( N ) ) (K(N)) ( K ( N )) only through α \alpha α , because A A A does. For every β > 0 \beta>0 β > 0 , the finite-volume entropy bound gives
β g ^ N ∙ ≤ A ^ N ∙ ( β ) ≤ β g ^ N ∙ + log 2. \beta\widehat{g}_N^{\bullet}\leq\widehat{A}_N^{\bullet}(\beta)\leq\beta\widehat{g}_N^{\bullet}+\log2. β g N ∙ ≤ A N ∙ ( β ) ≤ β g N ∙ + log 2.
Taking expectations and then the lower and upper limits in N N N yields
A ( β ) − log 2 β ≤ lim inf N g N ∙ ≤ lim sup N g N ∙ ≤ A ( β ) β . \frac{A(\beta)-\log2}{\beta}\leq\liminf_{N}g_N^{\bullet}\leq\limsup_{N}g_N^{\bullet}\leq\frac{A(\beta)}{\beta}. β A ( β ) − log 2 ≤ N lim inf g N ∙ ≤ N lim sup g N ∙ ≤ β A ( β ) .
Both endpoints tend to g ∙ ( α ) g^{\bullet}(\alpha) g ∙ ( α ) as β → ∞ \beta\to\infty β → ∞ , proving g N ∙ → g ∙ ( α ) g_N^{\bullet}\to g^{\bullet}(\alpha) g N ∙ → g ∙ ( α ) . For almost-sure convergence, let Ω 0 \Omega_0 Ω 0 be the event of probability one in Theorem 2.3 (b). Fix a realization in Ω 0 \Omega_0 Ω 0 . For every positive integer j j j , apply the same finite-volume inequality at inverse temperature β = j \beta=j β = j and use sample-pressure convergence to obtain
A ( j ) − log 2 j ≤ lim inf N g ^ N ≤ lim sup N g ^ N ≤ A ( j ) j . \frac{A(j)-\log2}{j}\le\liminf_{N}\widehat{g}_N\le\limsup_{N}\widehat{g}_N\le\frac{A(j)}{j}. j A ( j ) − log 2 ≤ N lim inf g N ≤ N lim sup g N ≤ j A ( j ) .
These bounds hold simultaneously for all j j j on Ω 0 \Omega_0 Ω 0 . Letting j → ∞ j\to\infty j → ∞ forces both the liminf and the limsup to equal g ∘ ( α ) g^\circ(\alpha) g ∘ ( α ) , proving almost-sure convergence. Finally, 0 ≤ g ^ N ≤ C N 0\le\widehat{g}_N\le C_N 0 ≤ g N ≤ C N and sup N E C N 2 < ∞ \sup_N\mathbb{E}C_N^2<\infty sup N E C N 2 < ∞ give uniform integrability, hence convergence in L 1 L^1 L 1 . The comparison g G ≤ g R g^G\le g^R g G ≤ g R follows by dividing A G ≤ A R A^G\le A^R A G ≤ A R (Corollary 2.6) by β \beta β and letting β → ∞ \beta\to\infty β → ∞ . □
Proof of Proposition 2.7 . (a) The properties in h h h and β \beta β follow from Lemma 3.5 by passing to the limit. Let ξ \xi ξ be a new pattern, independent of J ∙ J^\bullet J ∙ and with the same law as its columns, and let ω K \omega_K ω K be the Gibbs measure with K K K patterns. Jensen’s inequality gives log Z N , K + 1 ∙ − log Z N , K ∙ = log ω K ( e β ( ξ ⋅ σ ) 2 / ( 2 N ) ) ≥ β 2 N ω K ( ( ξ ⋅ σ ) 2 ) \log Z^\bullet_{N,K+1}-\log Z^\bullet_{N,K}=\log\omega_K(e^{\beta(\xi\cdot\sigma)^2/(2N)})\ge\frac{\beta}{2N}\omega_K((\xi\cdot\sigma)^2) log Z N , K + 1 ∙ − log Z N , K ∙ = log ω K ( e β ( ξ ⋅ σ ) 2 / ( 2 N ) ) ≥ 2 N β ω K (( ξ ⋅ σ ) 2 ) , and E ξ ( ξ ⋅ σ ) 2 = N \mathbb{E}_{\xi}(\xi\cdot\sigma)^2=N E ξ ( ξ ⋅ σ ) 2 = N for every σ \sigma σ . Hence A N , K + 1 ∙ − A N , K ∙ ≥ β 2 N A^\bullet_{N,K+1}-A^\bullet_{N,K}\ge\frac{\beta}{2N} A N , K + 1 ∙ − A N , K ∙ ≥ 2 N β , so K ↦ A N , K ∙ − β K 2 N K\mapsto A^\bullet_{N,K}-\frac{\beta K}{2N} K ↦ A N , K ∙ − 2 N β K is nondecreasing. For α 1 ≤ α 2 \alpha_1\le\alpha_2 α 1 ≤ α 2 and K i ( N ) : = ⌊ α i N ⌋ K_i(N):=\lfloor\alpha_iN\rfloor K i ( N ) := ⌊ α i N ⌋ , Theorem 2.1 gives monotonicity of α ↦ A ∙ − β α / 2 \alpha\mapsto A^\bullet-\beta\alpha/2 α ↦ A ∙ − β α /2 . (b) follows from (14) and the annealed bound in Lemma 3.5. (c) The upper bound follows from (14) , and by Lemma 4.6 , A N , K 2 G − A N , K 1 G ≤ β 2 K 2 − K 1 K 2 [ ( 1 + K 2 / N ) 2 + 1 N ] A^G_{N,K_2}-A^G_{N,K_1}\le\frac{\beta}{2}\frac{K_2-K_1}{K_2}\left[(1+\sqrt{K_2/N})^2+\frac{1}{N}\right] A N , K 2 G − A N , K 1 G ≤ 2 β K 2 K 2 − K 1 [ ( 1 + K 2 / N ) 2 + N 1 ] , which converges to the stated bound. □
The additional binary pattern
We finally use the uniform-field limit to treat an additional Mattis interaction. Gauge invariance removes the dependence on the law of the additional pattern, and a finite-magnetization decomposition reduces the problem to a scalar maximization.
For a real N × K N\times K N × K matrix J J J let A ^ N C W ( J ) : = f N β N m 2 / 2 ( J ) \widehat{A}^{\mathrm{CW}}_N(J):=f_N^{\beta Nm^2/2}(J) A N CW ( J ) := f N β N m 2 /2 ( J ) , the statistical pressure with an additional Curie–Weiss term; here c V = log 2 + β / 2 c_V=\log2+\beta/2 c V = log 2 + β /2 .
Lemma 10.1 (Gauge invariance). Let ∙ ∈ { G , R } \bullet\in\{G,R\} ∙ ∈ { G , R } , D η : = diag ( η ) D_\eta:=\operatorname{diag}(\eta) D η := diag ( η ) and J ~ : = D η J ∙ \widetilde{J}:=D_\eta J^\bullet J := D η J ∙ . Then A ^ N , K ∙ , M ( β ) = A ^ N C W ( J ~ ) \widehat{A}^{\bullet,M}_{N,K}(\beta)=\widehat{A}^{\mathrm{CW}}_N(\widetilde{J}) A N , K ∙ , M ( β ) = A N CW ( J ) , and J ~ \widetilde{J} J has the law of J ∙ J^\bullet J ∙ and is independent of η \eta η . In particular A N , K ∙ = E A ^ N C W ( J ∙ ) A^\bullet_{N,K}=\mathbb{E}\widehat{A}^{\mathrm{CW}}_N(J^\bullet) A N , K ∙ = E A N CW ( J ∙ ) does not depend on the law of η \eta η .
Proof. Under the bijection σ ↦ D η σ \sigma\mapsto D_\eta\sigma σ ↦ D η σ of Σ N \Sigma_N Σ N , ( J ∙ ) T σ = J ~ T ( D η σ ) (J^\bullet)^T\sigma=\widetilde{J}^{T}(D_\eta\sigma) ( J ∙ ) T σ = J T ( D η σ ) and η ⋅ σ = N m ( D η σ ) \eta\cdot\sigma=Nm(D_\eta\sigma) η ⋅ σ = N m ( D η σ ) . Given η \eta η , the entries η i ( J ∙ ) i μ \eta_i(J^\bullet)^\mu_i η i ( J ∙ ) i μ are independent with the same symmetric law as those of J ∙ J^\bullet J ∙ , so the conditional law of J ~ \widetilde{J} J does not depend on η \eta η . □
Lemma 10.2. For every real N × K N\times K N × K matrix J J J ,
0 ≤ A ^ N C W ( J ) − max ∣ v ∣ ≤ 1 { f N β v N m ( J ) − β v 2 2 } ≤ log ( N + 1 ) N . 0\le\widehat{A}^{\mathrm{CW}}_N(J)-\max_{\lvert v\rvert\le1}\left\{f_N^{\beta vNm}(J)-\frac{\beta v^2}{2}\right\}\le\frac{\log(N+1)}{N}. 0 ≤ A N CW ( J ) − ∣ v ∣ ≤ 1 max { f N β v N m ( J ) − 2 β v 2 } ≤ N log ( N + 1 ) .
Proof. For ∣ v ∣ ≤ 1 \lvert v\rvert\le1 ∣ v ∣ ≤ 1 , m 2 / 2 ≥ v m − v 2 / 2 m^2/2\ge vm-v^2/2 m 2 /2 ≥ v m − v 2 /2 gives the lower bound. Conversely, for each of the N + 1 N+1 N + 1 values x x x of m ( σ ) m(\sigma) m ( σ ) , ∑ m ( σ ) = x e − β H N + β N x 2 / 2 ≤ e − β N x 2 / 2 ∑ σ e − β H N + β x N m ( σ ) \sum_{m(\sigma)=x}e^{-\beta H_N+\beta Nx^2/2}\le e^{-\beta Nx^2/2}\sum_\sigma e^{-\beta H_N+\beta xNm(\sigma)} ∑ m ( σ ) = x e − β H N + β N x 2 /2 ≤ e − β N x 2 /2 ∑ σ e − β H N + β x N m ( σ ) , since each term on the left appears on the right multiplied by e β x N m ( σ ) − β N x 2 = 1 e^{\beta xNm(\sigma)-\beta Nx^2}=1 e β x N m ( σ ) − β N x 2 = 1 and the remaining terms are positive. □
Proof of Theorem 2.5 . If β = 0 \beta=0 β = 0 all quantities equal log 2 \log2 log 2 ; let β > 0 \beta>0 β > 0 and write K = K ( N ) K=K(N) K = K ( N ) . Since f N β v N m ( J ∙ ) = A ^ N , K ∙ ( β , β v ) f_N^{\beta vNm}(J^\bullet)=\widehat{A}^\bullet_{N,K}(\beta,\beta v) f N β v N m ( J ∙ ) = A N , K ∙ ( β , β v ) and ∣ max φ 1 − max φ 2 ∣ ≤ sup ∣ φ 1 − φ 2 ∣ \lvert\max\varphi_1-\max\varphi_2\rvert\le\sup\lvert\varphi_1-\varphi_2\rvert ∣ max φ 1 − max φ 2 ∣ ≤ sup ∣ φ 1 − φ 2 ∣ , Lemmas 10.1 and 10.2 give
∣ A ^ N , K ∙ , M − max ∣ v ∣ ≤ 1 { A ^ ∙ ( α , β , β v ) − β v 2 2 } ∣ ≤ log ( N + 1 ) N + E sup ∣ h ∣ ≤ β ∣ A ^ N , K ∙ ( β , h ) − A ∙ ( α , β , h ) ∣ → 0 \left|\widehat{A}^{\bullet,M}_{N,K}-\max_{\lvert v\rvert\le1}\left\{\widehat{A}^\bullet(\alpha,\beta,\beta v)-\frac{\beta v^2}{2}\right\}\right|\le\frac{\log(N+1)}{N}+\mathbb{E}\sup_{\lvert h\rvert\le\beta}\left|\widehat{A}_{N,K}^\bullet(\beta,h)-A^\bullet(\alpha,\beta,h)\right|\to0 A N , K ∙ , M − ∣ v ∣ ≤ 1 max { A ∙ ( α , β , β v ) − 2 β v 2 } ≤ N log ( N + 1 ) + E ∣ h ∣ ≤ β sup A N , K ∙ ( β , h ) − A ∙ ( α , β , h ) → 0
by Theorem 2.3 (b). For ∣ v ∣ > 1 |v| > 1 ∣ v ∣ > 1 and s = sgn v s = \operatorname{sgn} v s = sgn v , the 1-Lipschitz property of A ∙ A^\bullet A ∙ in h h h gives A ∙ ( α , β , β v ) − β v 2 / 2 ≤ A ∙ ( α , β , β s ) − β / 2 − β ( ∣ v ∣ − 1 ) 2 / 2 A^\bullet(\alpha,\beta,\beta v) - \beta v^2/2 \le A^\bullet(\alpha,\beta,\beta s) - \beta/2 - \beta(|v| - 1)^2/2 A ∙ ( α , β , β v ) − β v 2 /2 ≤ A ∙ ( α , β , β s ) − β /2 − β ( ∣ v ∣ − 1 ) 2 /2 , so the supremum over R \mathbb{R} R equals the maximum over [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] .
For the sample statements, apply Lemmas 10.1 and 10.2 with J ~ N : = D η N J N , K ( N ) ∙ \widetilde{J}_N := D_{\eta_N}J^\bullet_{N,K(N)} J N := D η N J N , K ( N ) ∙ . The sequence ( J ~ N ) (\widetilde{J}_N) ( J N ) has the marginal laws of J ∙ J^\bullet J ∙ at every volume, so Theorem 2.3 (b), which does not require independence between volumes, applies to it: almost surely f N h N m ( J ~ N ) → A ∙ ( α , β , h ) f_N^{hNm}(\widetilde{J}_N) \to A^\bullet(\alpha,\beta,h) f N h N m ( J N ) → A ∙ ( α , β , h ) (with β \beta β the inverse temperature in f f f ) locally uniformly in ( β , h ) (\beta,h) ( β , h ) , which permits passage to the maximum in Lemma 10.2 simultaneously on compact β \beta β -intervals. For L 2 L^2 L 2 convergence, Lemma 3.4 with V = β N m 2 / 2 V = \beta Nm^2/2 V = β N m 2 /2 , together with (T1) or Lemma 3.7, gives Var A ^ N C W ( J ∙ ) ≤ 2 β 2 N − 2 E ∥ J ∙ ∥ o p 2 = O ( N − 1 ) \operatorname{Var}\widehat{A}^{\mathrm{CW}}_N(J^\bullet) \le2\beta^2N^{-2}\mathbb{E}\|J^\bullet\|_{\mathrm{op}}^2 = O(N^{-1}) Var A N CW ( J ∙ ) ≤ 2 β 2 N − 2 E ∥ J ∙ ∥ op 2 = O ( N − 1 ) , and Lemma 10.1 identifies the law of A ^ N C W ( J ~ N ) \widehat{A}^{\mathrm{CW}}_N(\widetilde{J}_N) A N CW ( J N ) with that of A ^ N , K ∙ , M \widehat{A}^{\bullet,M}_{N,K} A N , K ∙ , M .
Finally, for binary patterns choose η \eta η with independent symmetric Rademacher entries. The model (5) is then the binary Hopfield model with K ( N ) + 1 K(N)+1 K ( N ) + 1 independent patterns, and since ( K ( N ) + 1 ) / N → α (K(N)+1)/N \to\alpha ( K ( N ) + 1 ) / N → α , Theorem 2.1 gives A R , M ( α , β ) = A R ( α , β ) A^{R,M}(\alpha,\beta) = A^R(\alpha,\beta) A R , M ( α , β ) = A R ( α , β ) . As the value v = 0 v = 0 v = 0 in (9) gives A R , M ≥ A R A^{R,M} \ge A^R A R , M ≥ A R , the identity is equivalent to the stated inequality. □
Proof of Corollary 2.6. Adding the nonnegative term β 2 N ( ζ 1 ⋅ σ ) 2 \frac{\beta}{2N}(\zeta^1\cdot\sigma)^2 2 N β ( ζ 1 ⋅ σ ) 2 to the exponent shows p N , K ( 1 ) ( h ) ≥ p N , K − 1 ( 1 ) ( h ) p^{(1)}_{N,K}(h) \ge p^{(1)}_{N,K-1}(h) p N , K ( 1 ) ( h ) ≥ p N , K − 1 ( 1 ) ( h ) , hence A G ≤ p 1 A^G \le p_1 A G ≤ p 1 by Theorem 2.1 ; p 1 ≤ A R p_1 \le A^R p 1 ≤ A R is (24) . At h = 0 h = 0 h = 0 , p N , K ( 1 ) p^{(1)}_{N,K} p N , K ( 1 ) is the quenched pressure of the Gaussian model with K − 1 K-1 K − 1 patterns and the additional pattern η = ζ 1 \eta= \zeta^1 η = ζ 1 , so Theorem 2.5 with ( K ( N ) − 1 ) / N → α (K(N)-1)/N \to\alpha ( K ( N ) − 1 ) / N → α gives p 1 ( α , β , 0 ) = A G , M ( α , β ) p_1(\alpha,\beta,0) = A^{G,M}(\alpha,\beta) p 1 ( α , β , 0 ) = A G , M ( α , β ) . The remaining assertions follow from (6) and (10) . □
Discussion
For mean-field spin glasses, the Guerra–Toninelli interpolation [31 ] established existence of the thermodynamic limit before the proofs of the Parisi formula [44 , 30 , 50 ] and of ultrametricity [41 ] . For the Hopfield model, the lack of joint convexity of the bipartite covariance requires a different argument. For Gaussian and binary patterns, at every positive load and finite inverse temperature, we have proved convergence and self-averaging of the statistical pressure (also in a uniform field and with an additional binary pattern), convergence of the ground-state energy, and convergence of the internal energy and entropy at every differentiability point of the limiting pressure; all limits depend on ( K ( N ) ) (K(N)) ( K ( N )) only through α \alpha α . We do not establish the phase diagram of Amit, Gutfreund and Sompolinsky [5 , 6 ] : the retrieval thresholds, the location and order of transitions, and the concentration of order parameters such as Mattis magnetizations and overlaps are not addressed. The limiting pressure is not evaluated explicitly. For Gaussian patterns, the representation below expresses the limit implicitly, through the Hamilton–Jacobi variational formula of [23 ] and a limit in the truncation level, and an explicit variational principle has been announced in [20 ] ; we do not prove that the two coincide. For binary patterns, no formula in terms of conventional order parameters is obtained here; equation (6) expresses A R A^R A R through the Gaussian-bulk limits p r p_r p r .
Combining the proof of Proposition 4.4 with the identification of the limit in [23 Theorem 1.1] gives an implicit variational representation. For rational ρ = a / b \rho= a/b ρ = a / b and every L ≥ 4 L \ge4 L ≥ 4 the limit ℓ L \ell_L ℓ L of the truncated statistical pressure of the Gaussian model is
ℓ L = log ( 2 cosh h ) + a b log d L − 1 b min ν ∈ P ( supp μ ~ ) { Ent ( ν ∣ μ ~ ) + f ν ( 1 2 , 0 ) } . \ell_L = \log(2\cosh h) + \frac{a}{b}\log d_L - \frac{1}{b}\min_{\nu\in\mathcal{P}(\operatorname{supp}\widetilde{\mu})}\left\{\operatorname{Ent}(\nu\mid\widetilde{\mu}) + f^\nu\left(\frac{1}{2},0\right)\right\}. ℓ L = log ( 2 cosh h ) + b a log d L − b 1 ν ∈ P ( supp μ ) min { Ent ( ν ∣ μ ) + f ν ( 2 1 , 0 ) } .
where f ν f^\nu f ν is the viscosity solution of a Hamilton–Jacobi equation on the space of monotone paths with nonlinearity Ψ ~ \widetilde{\Psi} Ψ and a ν \nu ν -dependent initial condition (the sign reflects that [23 ] state the result for − Φ ν -\Phi_\nu − Φ ν ). By Proposition 4.2, A G ( ρ , β , h ) A^G(\rho,\beta,h) A G ( ρ , β , h ) is the limit of these values as L → ∞ L \to\infty L → ∞ , with error at most C c u t L − 2 C_{\mathrm{cut}}L^{-2} C cut L − 2 uniformly in h h h . The same applies to the Gaussian-bulk limits b ( ρ , β ; H , π ) b(\rho,\beta;H,\pi) b ( ρ , β ; H , π ) , with log ( 2 cosh h ) \log(2\cosh h) log ( 2 cosh h ) replaced by b − 1 ∑ u ≤ b log ( 2 cosh H τ ( u ) ) b^{-1}\sum_{u\le b}\log(2\cosh H_{\tau(u)}) b − 1 ∑ u ≤ b log ( 2 cosh H τ ( u ) ) , and, through (22) , to every p r p_r p r ; irrational loads follow by the load continuity of Lemma 4.6 . We do not determine whether this representation reduces to the replica-symmetric expressions of [6 , 11 , 2 ] in the appropriate region, or to the variational principle announced in [20 ] . The Hamilton–Jacobi structure follows the approach introduced by Guerra, in which mean-field spin glasses are studied through partial differential equations from analytical mechanics [29 , 27 , 9 , 10 ] , developed for exact computations by Mourrat and collaborators [36 , 38 , 39 , 40 ] .
The comparison between pattern laws reflects the distinct roles of Gaussianity in the proof. The Gaussian argument uses Gaussianity in three places: the blocked interaction is a Gaussian process, as required by Theorem 3.6 ; the concentration estimates use {(T1)–(T2)}; and the truncation estimate uses rotational invariance. For binary patterns, Theorem 3.6 and rotational invariance are used only for the Gaussian bulk of the mixed models, the binary disorder being handled by finite-pattern extraction and a Lindeberg-type comparison, and Gaussian concentration is replaced by convex concentration on the discrete cube. At zero field, Corollary 2.6 shows A G ≤ A G , M ≤ A R A^{G}\le A^{G,M}\le A^{R} A G ≤ A G , M ≤ A R . Whether the supremum in (6) reduces to a fixed number of retained patterns, in particular to one, remains open; at zero field this is the question whether A R = A G , M A^{R}=A^{G,M} A R = A G , M , i.e. whether binary non-condensed patterns act on the pressure as Gaussian noise. Identification with an explicit hard–soft functional as in [2 ] would require control of the interpolation remainder and of the order of its extrema. These identifications are not required for the existence theorem. The finite-pattern extraction argument may also apply to other bipartite architectures.
Acknowledgements
A.B. acknowledges financial support from Sapienza University of Rome, project number RM12519999AB8CA9 (“Neural Networks and Learning Machines: asymptotic behaviors on structured datasets”) and project number RP1261A106890761 (“Neural Exponential Engine 4 Data Localization & Extraction”). A.B. is a member of GNFM-INdAM, and F.D. is a member of GNAMPA-INdAM.
The proofs presented in this paper were generated through extensive interaction with the AI models Claude (Opus 5.5) and ChatGPT (Sol 6.1), and the authors do not claim credit for their derivation. The authors formulated the problem, guided the research by posing questions, and suggested various approaches to its solution. This write-up is made available as a service to the research community and is not intended for journal publication in its current form. All statements and proofs have been checked by the authors, who take responsibility for the content of the paper.
References
Adriano Barra: Dipartimento di Scienze di Base e Applicazioni all’Ingegneria, Sapienza Università di Roma, Italy; GNFM, Istituto Nazionale d’Alta Matematica, Sezione di Roma, Italy; INFN, Istituto Nazionale di Fisica Nucleare, Sezione di Roma, Italy; CNR, Nanotec, Salento Unit, Lecce, Italy
Email address: adriano.barra@uniroma1.it
Fabrizio Durante: Dipartimento di Matematica e Fisica “Ennio De Giorgi”, Università del Salento, Lecce, Italy
Email address: fabrizio.durante@unisalento.it