REVIEW 4 minor 25 references
This paper proves a sharp, mean-dependent moderate-deviation exponent for Gaussian maxima and derives the critical Sherrington–Kirkpatrick free-energy variance from entropy.
Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →
T0 review · deepseek-v4-flash
2026-08-01 07:38 UTC pith:UORSLCDA
load-bearing objection Theorem 1 is a clean, likely correct resolution of the Ding–Eldan–Zhai question; the SK half is a credible alternative proof of Du–Huang, with its only real fragility sitting in two external critical-window estimates.
Moderate Deviations for Gaussian Maxima and an Entropy Proof of Critical SK Free Energy Fluctuations
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
The paper claims two sharp results. First, for a centered Gaussian vector with individual variances at most 1 and expected maximum E m_N, if E m_N is at least α√log N and E m_N + κ√log N ≤ √(2 log N), then P(max ≥ E m_N + κ√log N) ≤ N^{−κ²/(2−α²)+o(1)}. The exponent comes from optimizing the ratio (t²−2)/(t−α)² at t=2/α, and an equicorrelated Gaussian field shows no better exponent is possible. Second, the free energy of the Sherrington–Kirkpatrick model at β_c=1/√2 has variance (1/6) log N + O(1). The upper bound is an entropy calculation: Gaussian convexity bounds the variance by a tilted entropy, which is identified with the relative entropy of a planted Gaussian synchronization model; th
What carries the argument
The load-bearing identity is the concavity of the Gaussian quantile of the maximum's distribution: the map t → Φ^{-1}(P(max X_i ≤ t)) is concave, allowing the quantile at the moderate-deviation level to be interpolated from the median and a high deterministic level t√log N. A union bound plus two-sided Gaussian tail bounds supplies the quantile at the high level, and optimizing h(t)=(t²−2)/(t−α)² over t>√2 at t=2/α yields the exponent 2/(2−α²). For the spin-glass upper bound, the machinery is a chain: Gaussian convexity of the log-partition function bounds variance by an entropy; a diagonal/off-diagonal split reduces it to the off-diagonal entropy; that entropy is identified with the relativ
Load-bearing premise
The spin-glass upper bound rests on a quoted sharp estimate at the critical random-graph window—two fixed vertices must be connected with probability at most order N^{−2/3}—and on an information-percolation comparison, and if either degraded by a polynomial factor the O(1) transfer from the critical window would fail and the variance constant would not close.
What would settle it
Compute the two-point connection probability P(1↔2) in the critical random graph with edge probability 1/N + A N^{−4/3}; the upper-bound chain needs this to be O(N^{−2/3}). If simulation or rigorous estimate showed a slower decay, the entropy derivative bound would exceed O(N^{1/3}) and the variance constant would not close. For the moderate-deviation half, the equicorrelated example already saturates the exponent, so the falsifying observation would be any Gaussian vector satisfying the hypotheses with a strictly larger tail than N^{−κ²/(2−α²)+o(1)}.
If this is right
- For every Gaussian vector satisfying the stated scale conditions, the moderate-deviation tail is at most N^{−κ²/(2−α²)+o(1)}, and the equicorrelated field shows this exponent is best possible.
- The Sherrington–Kirkpatrick free-energy variance at β_c=1/√2 is (1/6) log N + O(1), confirming the coefficient predicted by the critical-window cutoff through an independent entropy route.
- The upper-bound chain shows the variance at criticality is governed by an entropy derivative of size O(N^{1/3}) over the critical window [1−N^{−1/3},1]; the O(1) cost of moving to the critical point relies on two-point connection probabilities O(N^{−2/3}) in the critical random graph.
- The lower bound exhibits a matching N^{1/2}e^{−N/2} inverse second moment for the normalized SK partition function, which is what produces the 1/6 coefficient when combined with the entropy upper bound.
Where Pith is reading between the lines
- If the entropy route extends to the near-critical window, it should recover the full divergence profile −(1/2) log(1−β²/β_c²) as β approaches β_c, not just the critical constant.
- The analytic optimization used for the moderate-deviation exponent is not Gaussian-specific in shape; a similar tail-sharpening pattern might hold for any process whose upper tail at moderate deviations matches the Gaussian and whose expected maximum sits at the same scale.
- The critical random graph's N^{−2/3} two-point connection probability is what sets the N^{−1/3} critical window; replacing the graph comparison by another percolation model would probe whether the 1/6 coefficient is universal across mean-field spin glasses.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper contains two main results. Theorem 1 gives a moderate-deviation upper bound for the maximum of an N-dimensional centered Gaussian vector with variances at most one: under the assumptions E max >= alpha sqrt(log N) and E max + kappa sqrt(log N) <= sqrt(2 log N), the probability that the maximum exceeds its mean by kappa sqrt(log N) decays at most as N^{-kappa^2/(2-alpha^2)+o(1)}. The exponent is shown to be sharp by an equicorrelated Gaussian field. Theorem 3 establishes Var(F_N(beta_c)) = (1/6) log N + O(1) for the Sherrington--Kirkpatrick free energy at beta_c = 1/sqrt(2). The proof is based on an entropy/tilting argument: the upper bound identifies the variance with a Kullback--Leibler divergence of a Gaussian synchronization model and controls its derivative via the I--MMSE formula, information percolation, and critical random-graph susceptibility; the lower bound combines Gaussian convexity at a negative replica parameter with spherical inverse moments and GOE eigenvalue identities.
Significance. Both theorems are substantial. Theorem 1 answers the question of Ding--Eldan--Zhai with a sharp, explicit exponent and a short proof. Theorem 3 gives a genuinely different route to a recently proved critical SK result: even though Du--Huang already established the asymptotic, the entropy perspective is of independent interest and likely transferable to related models. I checked the key internal computations -- the t* = 2/alpha optimization, the exact second-moment calculation at lambda_0, the diagonal split B_N = beta^2/2 + Btilde_N, the KL representation, and the spherical inverse-moment/contour chain -- and found no internal inconsistency. The argument is not circular: the SK upper bound is anchored at the critical-window endpoint lambda_0 = 1 - N^{-1/3} and the lower bound does not assume the target variance. The main external dependence is [1, Thm 3.6] and [15, Cor 5.3]; the application in Lemma 11 is correct provided those results have exactly the quoted normalization and uniformity.
minor comments (4)
- [Section 2, Eq. (11) and Lemma 3] The symbol b_N in Eq. (11) is undefined; it should be m_N. The same typo appears in Lemma 3 when invoking (11).
- [Theorem 1] The assumption 'Emax + kappa sqrt(log N) <= sqrt(2) log N' should read 'Emax + kappa sqrt(log N) <= sqrt(2 log N)', as in the abstract. If taken literally as (sqrt(2)) log N, the statement is false and inconsistent with the proof, which requires the shifted maximum to be below t sqrt(log N) for a fixed t > sqrt(2).
- [Lemma 11, Eq. (45)] The closing O(1) bound in Lemma 11 is load-bearing and rests entirely on the exact forms of [1, Thm 3.6] and [15, Cor 5.3]. For transparency, please state explicitly the version of [15, Cor 5.3] used, namely the uniform estimate E sum_C |C|^2 = O_A(N^{4/3}) for p = 1/N + A N^{-4/3}, and the exact information-percolation inequality used from [1]. I do not regard this as a gap, but as written the reader cannot verify the absence of extra polynomial factors without consulting the cited papers.
- [Lemma 15 / Proposition 4] The soft-edge input from [19] is standard, but the statement 'joint soft edge convergence ... up to a fixed positive deterministic scaling constant' is terse. Since the lower bound requires a positive-probability event involving the first and fourth eigenvalues, it would help to spell out the limiting joint law and the choice of continuity points q_-, q_+, L.
Circularity Check
No significant circularity: the derivations are self-contained and rely on independent external theorems, not on their own conclusions.
full rationale
I walked both derivation chains and found no step in which a claimed prediction reduces by construction to an input, nor any load-bearing self-citation. For Theorem 1, the proof uses Ehrhard's concavity, Gaussian concentration around a median, and standard two-sided Mills bounds; the quantile interpolation and the optimization over t are explicit and do not assume the desired tail exponent. For the SK upper bound, the variance is only bounded by the entropy quantity B_N via convexity, and B_N is then computed directly at lambda0 by exact second-moment identities and propagated via the I-MMSE derivative, the Abbe--Boix-Adsera information-percolation inequality, and the Janson--Spencer critical-window susceptibility estimate. These are external, parameter-free results with stated general assumptions; no fitted value from the present paper enters them. For the lower bound, Gaussian convexity at s=-2 reduces the problem to an inverse moment and the already-proved entropy bound; using the entropy upper bound inside the lower bound is a legitimate reuse of an independent lemma, not an assumption of the target variance. The spherical inverse moment is obtained through Haar averaging and GOE soft-edge/dimension-shift identities. I also checked the acknowledgements and references: the paper does not cite itself, and the cited external theorems are not author-uniqueness claims or ansatz-based rescaling imported from the author's own prior work. The only fragility noted by a skeptical reader is the exact normalization of [1, Thm 3.6] and [15, Cor 5.3]; if either had hidden polynomial losses the O(1) transfer in Lemma 11 would fail. But that is a dependence on external inputs, not circularity, and the manuscript itself does not assert those theorems or claim they are proved here.
Axiom & Free-Parameter Ledger
free parameters (1)
- λ_0 (critical-window endpoint 1 − N^{−1/3}) =
1 − N^{-1/3}
axioms (9)
- standard math Ehrhard's concavity: Φ^{-1}(P(max_{i≤N} X_i ≤ t)) is concave in t (Lemma 1, cited to [12]).
- standard math Gaussian concentration around the median: P(|max X_i − m_N| ≥ u) ≤ 2e^{−u²/2} (Lemma 2, cited to [25]).
- domain assumption Wei-Kuo Chen's Gaussian convexity: for convex Ψ, K_Ψ(s) = (1/s) log E e^{sΨ(G)} is convex on R (Theorem 4, cited to [5]).
- standard math I-MMSE formula: (d/dγ) I(U; Y_γ) = ½ E‖U − E[U|Y_γ]‖² (Theorem 5, [14]).
- domain assumption Abbe–Boix-Adserà information-percolation bound: I₂(θ_u;θ_v|Y) ≤ P(u↔v in bond percolation with p_e = I₂(θ_i;θ_j|Y_e)) (Theorem 6, [1, Thm 3.6]).
- domain assumption Janson–Spencer critical-window component-size moments: at p = 1/N + O(N^{−4/3}), E Σ_C |C|² = O(N^{4/3}) ([15, Cor 5.3]).
- domain assumption β-ensemble soft-edge / stochastic Airy spectrum universality: the top eigenvalues of GOE centered at 2 and scaled N^{2/3} converge jointly ([19]).
- standard math Mehta integral formula C_{n,a} = (2/a)^{n(n+1)/4} (2π)^{n/2} Π Γ(1+j/2)/Γ(3/2) (eq. (74), [13]).
- standard math Bromwich/contour inversion for the spherical partition function Z^{sph}_N = (Γ(N/2)/(2πi(N/2)^{N/2−1})) ∫ e^{Nz/2} det(zI−W)^{−1/2} dz (Lemma 14, following [3]).
read the original abstract
In this paper, we study two problems concerning Gaussian maxima. First, let $(X_1,\ldots,X_N)$ be a centered Gaussian vector with $\operatorname{Var}(X_i)\leq 1$. Suppose that, for fixed $\alpha\in(0,\sqrt 2)$ and $\kappa>0$, $\mathbb{E}\max_iX_i\geq\alpha\sqrt{\log N}$ and $\mathbb{E}\max_iX_i+\kappa\sqrt{\log N}\leq\sqrt{2\log N}$. We prove that $\mathbb{P}\left(\max_iX_i\geq \mathbb{E}\max_iX_i+\kappa\sqrt{\log N}\right) \leq N^{-\kappa^2/(2-\alpha^2)+o(1)}$. This answers a question of Ding, Eldan and Zhai. The exponent is sharp, as witnessed by an equicorrelated Gaussian field. Second, for the Sherrington--Kirkpatrick model at the critical inverse temperature $\beta_c=1/\sqrt2$, we prove $\operatorname{Var}\bigl(F_N(\beta_c)\bigr)=\frac16\log N+O(1)$. Our argument establishes the variance asymptotics at the critical temperature from an entropy perspective, via a route distinct from that of Du and Huang. For the upper bound, we express the variance as an entropy under exponential tilting and identify this entropy with the Kullback--Leibler divergence of a Gaussian synchronization model. Its derivative is then bounded using the I-MMSE formula, information percolation, and estimates for the susceptibility of the critical Erd\H{o}s--R\'enyi random graph. For the lower bound, we combine Gaussian convexity applied at the replica parameter with an estimate for inverse moments on the sphere and an identity relating GOE eigenvalue densities in consecutive dimensions.
Reference graph
Works this paper leans on
-
[1]
EmmanuelAbbeandEnricBoix-Adser‘a.Aninformation-percolationboundforspinsynchronizationongeneralgraphs. Ann. Appl. Probab., 30(3):1066–1090, 2020
2020
-
[2]
Free-energy fluctuations and chaos in the sherrington-kirkpatrick model.Phys
Timo Aspelmeier. Free-energy fluctuations and chaos in the sherrington-kirkpatrick model.Phys. Rev. Lett., 100(11):117205, 2008
2008
-
[3]
Fluctuations of the free energy of the spherical sherrington–kirkpatrick model.J
Jinho Baik and Ji Oon Lee. Fluctuations of the free energy of the spherical sherrington–kirkpatrick model.J. Stat. Phys., 165(2):185–224, 2016
2016
-
[4]
An error bound in the Sudakov–Fernique inequality.arXiv preprint arXiv:math/0510424, 2005
Sourav Chatterjee. An error bound in the Sudakov–Fernique inequality.arXiv preprint arXiv:math/0510424, 2005
Pith/arXiv arXiv 2005
-
[5]
A Gaussian convexity for logarithmic moment generating functions with applications in spin glasses
Wei-Kuo Chen. A Gaussian convexity for logarithmic moment generating functions with applications in spin glasses. Ann. Inst. Henri Poincaré Probab. Stat., 62(1), 2026
2026
-
[6]
Order of fluctuations of the free energy in the SK model at critical temperature
Wei-Kuo Chen and Wai-Kit Lam. Order of fluctuations of the free energy in the SK model at critical temperature. ALEA Lat. Am. J. Probab. Math. Stat., 16(1):809–816, 2019
2019
-
[7]
Gaussian approximations and multiplier bootstrap for maxima of sums of high-dimensional random vectors.Ann
Victor Chernozhukov, Denis Chetverikov, and Kengo Kato. Gaussian approximations and multiplier bootstrap for maxima of sums of high-dimensional random vectors.Ann. Statist., 41(6):2786–2819, 2013
2013
-
[8]
Comparison and anti-concentration bounds for maxima of Gaussian random vectors.Probab
Victor Chernozhukov, Denis Chetverikov, and Kengo Kato. Comparison and anti-concentration bounds for maxima of Gaussian random vectors.Probab. Theory Related Fields, 162(1–2):47–70, 2015
2015
-
[9]
Partha S Dey and Taegu Kang. Fluctuations for the sherrington–kirkpatrick spin glass model near the critical tem- perature.arXiv preprint arXiv:2603.05636, 2026
arXiv 2026
-
[10]
On multiple peaks and moderate deviations for the supremum of a Gaussian field.Ann
Jian Ding, Ronen Eldan, and Alex Zhai. On multiple peaks and moderate deviations for the supremum of a Gaussian field.Ann. Probab., 43(6):3468–3493, 2015
2015
-
[11]
Hang Du and Brice Huang. Fluctuations of the Sherrington–Kirkpatrick free energy at critical temperature.arXiv preprint arXiv:2607.02172, 2026
Pith/arXiv arXiv 2026
-
[12]
Symétrisation dans l’espace de Gauss.Math
Antoine Ehrhard. Symétrisation dans l’espace de Gauss.Math. Scand., 53(2):281–301, 1983
1983
-
[13]
Forrester.Log-Gases and Random Matrices, volume 34 ofLondon Mathematical Society Monographs Series
Peter J. Forrester.Log-Gases and Random Matrices, volume 34 ofLondon Mathematical Society Monographs Series. Princeton University Press, Princeton, NJ, 2010
2010
-
[14]
Mutual information and minimum mean-square error in gaussian channels.IEEE Trans
Dongning Guo, Shlomo Shamai, and Sergio Verd’u. Mutual information and minimum mean-square error in gaussian channels.IEEE Trans. Inform. Theory, 51(4):1261–1282, 2005
2005
-
[15]
A point process describing the component sizes in the critical window of the random graph evolution.Combin
Svante Janson and Joel Spencer. A point process describing the component sizes in the critical window of the random graph evolution.Combin. Probab. Comput., 16(4):631–658, 2007
2007
-
[16]
Virasoro.Spin Glass Theory and Beyond: An Introduction to the Replica Method and Its Applications, volume 9 ofWorld Scientific Lecture Notes in Physics
Marc Mézard, Giorgio Parisi, and Miguel A. Virasoro.Spin Glass Theory and Beyond: An Introduction to the Replica Method and Its Applications, volume 9 ofWorld Scientific Lecture Notes in Physics. World Scientific Publishing Co., Inc., Teaneck, NJ, 1987
1987
-
[17]
Springer Monographs in Mathematics
Dmitry Panchenko.The Sherrington–Kirkpatrick Model. Springer Monographs in Mathematics. Springer, New York, NY, 2013
2013
-
[18]
Phase diagram and large deviations in the free energy of mean-field spin-glasses
Giorgio Parisi and Tommaso Rizzo. Phase diagram and large deviations in the free energy of mean-field spin-glasses. Phys. Rev. B, 79(13):134205, 2009
2009
-
[19]
Ramírez, Brian Rider, and Bálint Virág
José A. Ramírez, Brian Rider, and Bálint Virág. Beta ensembles, stochastic airy spectrum, and a diffusion.J. Amer. Math. Soc., 24(4):919–944, 2011
2011
-
[20]
Adrien Schertzer. The order of free energy fluctuations in the critical Sherrington–Kirkpatrick model revisited.arXiv preprint arXiv:2606.21360, 2026
Pith/arXiv arXiv 2026
-
[21]
Solvable model of a spin-glass.Phys
David Sherrington and Scott Kirkpatrick. Solvable model of a spin-glass.Phys. Rev. Lett., 35(26):1792–1796, 1975
1975
-
[22]
Michel Talagrand.Spin Glasses: A Challenge for Mathematicians: Cavity and Mean Field Models, volume 46 of Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics. Springer, Berlin, Heidelberg, 2003
2003
-
[23]
Volume I: Basic Examples, volume 54 ofErgebnisse der Math- ematik und ihrer Grenzgebiete
Michel Talagrand.Mean Field Models for Spin Glasses. Volume I: Basic Examples, volume 54 ofErgebnisse der Math- ematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics. Springer, Berlin, Heidelberg, 2011
2011
-
[24]
Volume II: Advanced Replica-Symmetry and Low Temperature, volume55ofErgebnisse der Mathematik und ihrer Grenzgebiete
MichelTalagrand.Mean Field Models for Spin Glasses. Volume II: Advanced Replica-Symmetry and Low Temperature, volume55ofErgebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics. Springer, Berlin, Heidelberg, 2011
2011
-
[25]
Cambridge University Press, Cambridge, 2018
Roman Vershynin.High-dimensional probability, volume 47 ofCambridge Series in Statistical and Probabilistic Math- ematics. Cambridge University Press, Cambridge, 2018. An introduction with applications in data science, With a foreword by Sara van de Geer. School of Mathematical Sciences, Peking University Email address:ymchenmath@math.pku.edu.cn
2018
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.