The Free Energy of an Enriched Continuous Random Energy Model in the Weak Correlation Regime

Alexander Alban; Fu-Hsuan Ho; Justin Ko

The Free Energy of an Enriched Continuous Random Energy Model in the Weak Correlation Regime

T0 review · 0 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read For every continuous random energy model with covariance bounded by the identity and a finite one-sided slope at 1, the enriched free energy converges to a universal one-parameter variational formula independent of the covariance.

desk verdict A careful, honest HJ proof of an enriched CREM variational formula that does not change the known q=0 answer but extends the method to a non-product configuration space; worth refereeing despite a restrictive derivative assumption and some harmless-looking typos. read the letter →

arxiv 2508.17313 v1 pith:WCPNADJH submitted 2025-08-24 math.PR math.AP

Alexander Alban , Fu-Hsuan Ho , Justin Ko This is my paper · ORCID

classification math.PRmath.AP MSC 60K3582B44

keywords continuousrandomenergymodelenrichedRuelleprobabilitycascadeweakcorrelationregimeHopfformulaHamilton-Jacobiequationuniversalityoffree

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proves that, for every continuous random energy model (CREM) whose covariance $A$ satisfies $A(x)\le x$ and has a finite left derivative at $1$, the limiting free energy of a suitably enriched model is given by a single one-parameter variational formula independent of $A$. Because the covariance drops out entirely, the result confirms that in the weak-correlation regime the detailed shape of correlations is irrelevant to the thermodynamic limit. The proof passes through a Hamilton–Jacobi equation and a Hopf formula, yielding a new route to the CREM free energy and recovering the classical formula at zero enrichment. The paper also exhibits a non-convex covariance for which the naive extension of the formula fails, showing why the weak-correlation restriction is essential.

What carries the argument

The engine is an enriched Hamiltonian built from the CREM field $H^A_N$ together with a Ruelle cascade: normalized weights $v_\alpha$ and Gaussian fields $Y_q$ indexed by spin and cascade level, so the model interpolates between the CREM and the Ruelle probability cascade. The enrichment gives a complete overlap structure and makes the free energy tractable as a finite-dimensional Hamilton–Jacobi equation with monotone nonlinearity. The explicit initial condition $\Psi$ is computed by writing the partition function as a nested branching-random-walk expectation and applying the asymptotic shape theorem for branching random walks. A Gaussian comparison lemma and a convex, Lipschitz lower envelope of $A$ sandwich the free energy between the identity-covariance case and the convex case, and the Hopf formula for the Hamilton–Jacobi equation collapses to the one-parameter supremum in Theorem 1.3.

What would settle it

Take $A(x)=x$, $q(u)=u$, $t=1$, and simulate the enriched partition function $-(1/N)\mathbb{E}\ln\sum_{\alpha}v_\alpha\sum_{\sigma}\exp(H_N)$ for increasing $N$; if the finite-$N$ values do not approach $\sup_{\lambda\in[0,1)}(\lambda+\int_{1-\lambda}^{1}u\,du-\ln2/(1-\lambda))$, then Theorem 1.3 is false.

Watch

Extended reading notes

Core claim

The central claim is Theorem 1.3: for all $t>0$ and all $q\in Q^1$, the enriched free energy converges to $f(t,q)=\sup_{\lambda\in[0,1)}\bigl(\lambda t+\int_{1-\lambda}^{1}q(u)\,du-\frac{\ln 2}{1-\lambda}\bigr)$, for every covariance $A$ with $A(x)\le x$ and finite left derivative at $1$. At $q=0$ this reads $f(t,0)=-\ln2$ for $t\le\ln2$ and $f(t,0)=t-2\sqrt{t\ln2}$ for $t>\ln2$. The theorem is proved by showing the free energy is the viscosity solution of an infinite-dimensional Hamilton–Jacobi equation with initial condition $\Psi(q)=-\ln2+\int_0^1(q(u)-\ln2/u^2)_+\,du$, then solving the equation by a Hopf variational formula. A corollary is that the limit is universal in the weak-correlation regime: the only remnant of the enrichment is the integral of $q$ over an interval of length $\lambda$ at the right endpoint.

Load-bearing premise

The proof requires the covariance function to have a finite one-sided slope at its right endpoint, because that slope is used to construct a convex path lying below $A$; without it, the lower-bound sandwich and the theorem are not established.

Editorial extensions

If this is right

The limiting free energy is the same for every covariance $A$ with $A(x)\le x$ and finite left derivative at $1$, so within this regime the model is universal.
At zero enrichment, the formula reproduces the classical CREM transition: $f(t,0)=-\ln2$ for $t\le\ln2$ and $t-2\sqrt{t\ln2}$ for $t>\ln2$.
Because the initial condition $\Psi$ is explicit, the variational formula can be solved in closed form, reducing the limit to a one-parameter supremum over $\lambda$.
Outside the weak-correlation regime, the naive Hopf extension fails: for a simple non-convex covariance, it is strictly larger than the true limiting free energy.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the pure-REM inequality proposed in Section 6.1 can be proved directly, the finite left-derivative hypothesis in Theorem 1.3 could be removed, extending universality to every covariance with $A(x)\le x$; the paper identifies this inequality as the only obstruction.
The explicit gap in the non-convex example makes this enriched model a natural testbed for conjectures about non-convex multi-species mean-field models, because both the true free energy and the naive variational bound are computable.
The universal formula depends on $q$ only through $\int_{1-\lambda}^{1}q(u)\,du$, so for many purposes a step-function or even constant enrichment path is as informative as a general path, which may simplify numerical checks.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The paper deserves a serious referee. The headline result is not the free energy itself — for q=0 in the weak-correlation regime it reproduces known CREM values — but the route to it: an enriched model that puts the CREM into the Hamilton–Jacobi framework, an explicit initial condition, and a clean Hopf-formula proof that the limit is independent of the covariance A. The non-convex counterexample in Section 6.2 is a useful sanity check: it shows that a naive variational extension overestimates the true free energy. That is a real contribution, not a restatement.

What I like most is the bookkeeping. Theorem 1.3 is stated with the finite-left-derivative hypothesis, the proof uses that hypothesis exactly where Lemma 5.4 needs it, and the outlook (Section 6.1) honestly explains why removing it is not a trivial continuity argument. The initial-condition computation in Section 4 is the most original technical part, and the use of Biggins' shape theorem is appropriate. The dependence on external HJ theory is standard; the authors say what they import and what they prove.

Soft spots, in proportion. The finite-left-derivative assumption is genuinely restrictive; the paper does not remove it, and the REM-based route sketched in Section 6.1 is left open. That is fine, but it means Theorem 1.3 is not the last word on the weak-correlation regime. The non-convex analysis is confined to a two-piece linear A; it illustrates the failure mode but does not give a general theory. The stress-test note found apparent coefficient inconsistencies in the displayed computations of Proposition 5.6 and 5.8 (e.g., 1/M versus 1/(M+1) factors). They do not appear load-bearing — the inequalities point the right way and the infinite-M limit washes them out — but a referee should ask the authors to fix them. The self-citations are benign: one is an alternative proof of a range-filling fact, the other restates the known CREM formula.

Overall: the central argument holds up, the limitations are labeled, and the method is genuinely new even if the q=0 answer is not. This is the kind of paper that should go to a competent referee with a request to check the finite-dimensional approximation details and the typos. I would cite the enriched-model construction and the non-convex counterexample in my own work on HJ methods for disordered systems.

Referee Report

0 major / 5 minor

Summary. The paper studies the continuous random energy model (CREM) with covariance A, enriched by a Ruelle-cascade term that couples the binary-tree spins to cascade weights. It defines an enriched free energy F_N(t,q) for step functions q, proves a Lipschitz extension to Q1 (Proposition 1.2), and computes the initial condition Ψ(q) explicitly (Theorem 1.7). The main result, Theorem 1.3, states that under A(x)≤x and a finite left derivative at 1, the limit f(t,q) is the one-parameter variational formula sup_{λ∈[0,1)} [λt + ∫_{1-λ}^1 q(u)du - ln2/(1-λ)], independent of A. The proof uses a Gaussian comparison sandwich (Lemmas 5.3–5.4), viscosity supersolution and subsolution arguments (Propositions 5.6 and 5.8), the external Hamilton–Jacobi framework of [13,14], and an explicit computation of the convex dual Ψ* (Proposition 5.9). Section 6 discusses the finite-left-derivative hypothesis as an open problem and constructs a two-speed covariance A_{θ,c} for which the naive variational extension strictly overestimates the true CREM free energy (Proposition 6.4).

Significance. If the main theorem is correct, it gives a new Hamilton–Jacobi proof of the known CREM free energy, establishes universality of the limit over all weak-correlation covariances, and provides a tractable enriched model interpolating between the CREM and the Ruelle probability cascade. The paper is careful about its hypotheses: the finite-left-derivative condition is used exactly where stated, in Lemma 5.4, and the limitations of the non-convex extension are honestly recorded. The explicit initial condition Ψ, the closed-form dual Ψ*, and the counterexample outside the weak-correlation regime are concrete contributions. The proof is detailed and the reliance on external results is clearly flagged; there are no fitted parameters and the central claim is falsifiable.

minor comments (5)

[Section 5, Eq. (5.26), (5.29)] The normalization and index ranges in these displays do not match the definition of H^A_M in Proposition 5.6. The sums should run over m=0,...,M with factor 1/(M+1); as written, the displayed equality followed by the Jensen step is not algebraically correct. The intended inequalities are clear and local, but the displayed proof should be corrected.
[Section 5, paragraph before Eq. (5.12)] The sequence q_i^M is defined for i=0,...,M+1, but C^(M) has M+1 coordinates q_0,...,q_M; the upper index should be M.
[Section 5.1, proof of Proposition 5.6] The sentence after (5.27) says one can substitute q'−q_N by q' in (5.28); since (5.28) already contains q'−q_N, the intended statement is that the cone structure lets one replace q'−q_N by an arbitrary element of C^(M)_≤. Please rephrase for clarity.
[Section 4, proof of Theorem 1.7] In the lower-bound part, the text cites '(4.29) follows from the fact that z_M(σ)∼N(0,N)' but equation (4.29) has not yet appeared at that point; the cross-reference should be corrected.
[Section 5.3, proof of Proposition 5.9] In the display following (5.47), the term 'u∗(y2)' should presumably be 'u∗(y)^2'; the surrounding computation indicates a typographical error.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the enriched free energy is derived from the model definition via external Hamilton–Jacobi theory; the two self-citations are auxiliary and not load-bearing.

full rationale

The central claim of Theorem 1.3 is not circular. The enriched free energy FN(t,q) is defined directly in (1.8) from the Gaussian CREM process, the Ruelle cascade weights, and the enrichment process; the target formula (1.11) is not used as an input in the definition. The proof proceeds by (i) extending FN to Q1 via Proposition 1.2, (ii) computing the initial condition Ψ(q) in Theorem 1.7 from the recursive branching-random-walk representation, (iii) identifying the limiting free energy as the viscosity solution of a Hamilton–Jacobi equation using external results [13, 14], and (iv) solving the resulting Hopf formula with the explicit Legendre transform of Ψ from Proposition 5.9. Each step has independent analytical content, and the known CREM formula in (1.14) is used only as a benchmark or consistency check, not as a premise. The self-citations are not load-bearing: [1] in Remark 4.3 gives an alternative proof of a range-filling fact already supplied by Biggins' theorem [3], and [21] restates the known CREM free energy formula used for comparison in Section 1.2, not for the derivation of Theorem 1.3. There are no fitted constants, no parameter is calibrated to data or to the target formula, and the regularity hypothesis in Lemma 5.4 is honestly stated and used exactly where the paper indicates (Remark 5.5). The negative example in Section 6.2 is computed from the model and compared with the separate known CREM formula, so it is also not circular. Overall, the derivation is self-contained modulo standard external PDE theory and classical REM/CREM results.

Assumptions & free parameters 0 free parameters · 5 assumptions · 1 invented entities

No free parameters are fitted to data: t, q, and A are inputs to the theorem. The axioms are the model definition, external theorems from the Hamilton-Jacobi and branching random walk literature, and the explicit weak-correlation hypothesis. The enrichment is a mathematical construction, not a physical entity with independent evidence.

assumptions (5)

domain assumption Existence and overlap structure of the CREM Gaussian process for right-continuous increasing A with A(0)=0 and A(1)=1.
This is the model definition in (1.1) and (1.2); the central claim is about this model.
standard math Ruelle cascade weights (v_α) are well defined and satisfy the recursive expectation formula of Proposition 2.2 from [19, Theorem 5.25].
Invoked in Lemma 2.4 to reduce the enriched free energy to nested branching random walk expectations.
standard math The Hamilton-Jacobi viscosity solution framework of [13,14] applies to the finite-dimensional cone C^≤ and to Q2, including uniqueness and comparison for monotone nonlinearities.
Used in Section 5 to pass from viscosity subsolutions and supersolutions to the Hopf formulas (5.12) and (5.13).
standard math Biggins' shape theorem characterizes the range of the branching random walk by the ball of radius sqrt(2 ln2).
Used in Section 4, equation (4.38), to prove the upper bound for the initial condition Ψ.
domain assumption The covariance A satisfies A(x)≤x and has a finite left derivative at 1.
This is the stated scope of Theorem 1.3 and is used in Lemma 5.4 to construct a convex lower envelope.

invented entities (1)

Ruelle cascade enrichment (v_α, Y_q) interpolating between CREM and Ruelle probability cascade
purpose: To make the Hamilton-Jacobi approach applicable to the tree-structured CREM configuration space.
A well-defined mathematical construction introduced in (1.3)-(1.8). It has no falsifiable empirical handle outside the model; its role is structural.

how reviews work

0 comments

Cite this review

Pith. "Pith review of The Free Energy of an Enriched Continuous Random Energy Model in the Weak Correlation Regime." pith.science (2026). https://pith.science/paper/WCPNADJH

@misc{pith2026250817313,
  author       = {Pith},
  title        = {Pith review of: The Free Energy of an Enriched Continuous Random Energy Model in the Weak Correlation Regime},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WCPNADJH}},
  note         = {Machine review of arXiv:2508.17313}
}

abstract

We revisit the proof of the limiting free energy of the continuous random energy model (CREM) using the Hamilton--Jacobi approach for mean-field disordered systems. To achieve this, we introduce an enriched model that interpolates between the CREM and the Ruelle probability cascade. We focus on the weak correlation regime, where the CREM's covariance function $A$ is bounded above by the identity function. In the weak correlation regime, we show that the free energy is given by the Hopf formula. The resulting expression is independent of $A$, confirming that in this regime the free energy does not depend on the precise form of the covariance function. Outside of the weak correlation regime, the Hamilton--Jacobi framework no longer applies. Moreover, we provide an example where a formal application of the associated variational principle fails to yield the correct free energy.

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 32 canonical work pages

[1]

A. Alban. The Influence of Correlations on Random Energy Models . PhD thesis, Johannes Gutenberg-Universit¨ at, 2025.https://openscience.ub.uni-mainz.de/handle/20.500. 12030/12378

work page 2025
[2]

Bates and Y

E. Bates and Y. Sohn. Balanced multi-species spin glasses, July 2025. http://arxiv.org/ abs/2507.06522

arXiv 2025
[3]

J. D. Biggins. The asymptotic shape of the branching random walk. Advances in Applied Probability, 10(1):62–84, 1978

work page 1978
[4]

Bovier and I

A. Bovier and I. Kurkova. Derrida’s generalised random energy models. I. Models with finitely many hierarchies. Ann. Inst. H. Poincar´ e Probab. Statist., 40(4):439–480, 2004

work page 2004
[5]

Bovier and I

A. Bovier and I. Kurkova. Derrida’s generalized random energy models. II. Models with continuous hierarchies. Ann. Inst. H. Poincar´ e Probab. Statist., 40(4):481–495, 2004

work page 2004
[6]

Capocaccia, M

D. Capocaccia, M. Cassandro, and P. Picco. On the existence of thermodynamics for the generalized random energy model. J. Stat. Phys. , 46(3):493–505, 1987

work page 1987
[7]

H.-B. Chen. Hamilton-Jacobi equations for nonsymmetric matrix inference. Ann. Appl. Probab., 32(4):2540–2567, 2022

work page 2022
[8]

H.-B. Chen. On free energy of non-convex multi-species spin glasses, Nov. 2024. http: //arxiv.org/abs/2411.13342

arXiv 2024

Show all 37 references

[9]

Chen and J.-C

H.-B. Chen and J.-C. Mourrat. On the free energy of vector spin glasses with non-convex interactions. Probab. Math. Phys., 6(1):1–80, 2025

2025
[10]

Chen, J.-C

H.-B. Chen, J.-C. Mourrat, and J. Xia. Statistical inference of finite-rank tensors. Ann. H. Lebesgue, 5:1161–1189, 2022

2022
[11]

Chen and J

H.-B. Chen and J. Xia. Hamilton-Jacobi equations for inference of matrix tensor products. Ann. Inst. H. Poincar´ e Probab. Stat., 58(2):755–793, 2022

2022
[12]

Chen and J

H.-B. Chen and J. Xia. Free energy of multi-layer generalized linear models. Comm. Math. Phys., 400(3):1861–1913, 2023

1913
[13]

Chen and J

H.-B. Chen and J. Xia. Hamilton-Jacobi equations with monotone nonlinearities on convex cones, June 2024. http://arxiv.org/abs/2206.12537

2024 arXiv
[14]

Chen and J

H.-B. Chen and J. Xia. Hamilton–Jacobi equations from mean-field spin glasses. Probab. Theory Relat. Fields, 192(3):803–873, Aug. 2025. 40

2025
[15]

B. Derrida. Random-energy model: An exactly solvable model of disordered systems. Phys. Rev. B, 24(5):2613, 1981

1981
[16]

B. Derrida. A generalization of the random energy model which includes correlations between energies. J. Phys. Lett. (France), 46(9):401–407, 1985

1985
[17]

Derrida and E

B. Derrida and E. Gardner. Solution of the generalised random energy model. J. Phys. C , 19:2253–2274, 1986

1986
[18]

Dominguez and J.-C

T. Dominguez and J.-C. Mourrat. Mutual information for the sparse stochastic block model. The Annals of Probability , 52(2):434–501, 2024

2024
[19]

Dominguez and J.-C

T. Dominguez and J.-C. Mourrat. Statistical Mechanics of Mean-Field Disordered Systems: A Hamilton-Jacobi Approach, volume 32 of Zur. Lect. Adv. Math. EMS Press, 2024

2024
[20]

F. Guerra. Broken replica symmetry bounds in the mean field spin glass model. Commu- nications in mathematical physics , 233(1):1–12, 2003

2003
[21]

F.-H. Ho. Efficient approximation of the CREM Gibbs measure and the hardness threshold. J. Stat. Phys. , 192(3):Paper No. 36, 33, 2025

2025
[22]

V. Issa. A Hopf-like formula for mean-field spin glass models, Oct. 2024. http://arxiv. org/abs/2410.08754

2024 arXiv
[23]

V. Issa. Uniqueness of semi-concave weak solutions for Hamilton-Jacobi equations, Feb
[24]

V. Issa. Weak-Strong Uniqueness Principle for Hamilton-Jacobi Equations, Oct. 2024. http://arxiv.org/abs/2410.00628

2024 arXiv
[25]

J.-C. Mourrat. Hamilton-Jacobi equations for finite-rank matrix inference. Ann. Appl. Probab., 30(5):2234–2260, 2020

2020
[26]

J.-C. Mourrat. Hamilton–Jacobi equations for mean-field disordered systems. Ann. H. Lebesgue, 4:453–484, 2021

2021
[27]

J.-C. Mourrat. Nonconvex interactions in mean-field spin glasses. Probab. Math. Phys. , 2(2):281–339, 2021

2021
[28]

J.-C. Mourrat. The Parisi formula is a Hamilton-Jacobi equation in Wasserstein space. Canad. J. Math. , 74(3):607–629, 2022

2022
[29]

J.-C. Mourrat. Free energy upper bound for mean-field vector spin glasses. Ann. Inst. H. Poincar´ e Probab. Stat., 59(3):1143–1182, 2023

2023
[30]

J.-C. Mourrat. Un-inverting the Parisi formula, Sept. 2023. http://arxiv.org/abs/2308. 10715

2023
[31]

Mourrat and D

J.-C. Mourrat and D. Panchenko. Extending the Parisi formula along a Hamilton-Jacobi equation. Electron. J. Probab., 25:1–17, 2020

2020
[32]

Panchenko

D. Panchenko. The Parisi ultrametricity conjecture. Ann. Math. (2), 177(1):383–393, 2013

2013
[33]

Panchenko

D. Panchenko. The Sherrington-Kirkpatrick Model. Springer Monographs in Mathematics. Springer, New York, 2013

2013
[34]

G. Parisi. Infinite number of order parameters for spin-glasses. Phys. Rev. Lett. , 43(23):1754, 1979. 41

1979
[35]

Talagrand

M. Talagrand. The Parisi formula. Ann. of Math. (2) , 163(1):221–263, 2006

2006
[36]

H. Whitney. Functions differentiable on the boundaries of regions. Ann. Math., 35(3):482– 485, 1934. 42

1934
[2024]

http://arxiv.org/abs/2402.07665

arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [1]

A. Alban. The Influence of Correlations on Random Energy Models . PhD thesis, Johannes Gutenberg-Universit¨ at, 2025.https://openscience.ub.uni-mainz.de/handle/20.500. 12030/12378

work page 2025

[2] [2]

Bates and Y

E. Bates and Y. Sohn. Balanced multi-species spin glasses, July 2025. http://arxiv.org/ abs/2507.06522

arXiv 2025

[3] [3]

J. D. Biggins. The asymptotic shape of the branching random walk. Advances in Applied Probability, 10(1):62–84, 1978

work page 1978

[4] [4]

Bovier and I

A. Bovier and I. Kurkova. Derrida’s generalised random energy models. I. Models with finitely many hierarchies. Ann. Inst. H. Poincar´ e Probab. Statist., 40(4):439–480, 2004

work page 2004

[5] [5]

Bovier and I

A. Bovier and I. Kurkova. Derrida’s generalized random energy models. II. Models with continuous hierarchies. Ann. Inst. H. Poincar´ e Probab. Statist., 40(4):481–495, 2004

work page 2004

[6] [6]

Capocaccia, M

D. Capocaccia, M. Cassandro, and P. Picco. On the existence of thermodynamics for the generalized random energy model. J. Stat. Phys. , 46(3):493–505, 1987

work page 1987

[7] [7]

H.-B. Chen. Hamilton-Jacobi equations for nonsymmetric matrix inference. Ann. Appl. Probab., 32(4):2540–2567, 2022

work page 2022

[8] [8]

H.-B. Chen. On free energy of non-convex multi-species spin glasses, Nov. 2024. http: //arxiv.org/abs/2411.13342

arXiv 2024