Pith. sign in

REVIEW 4 major objections 3 minor 36 references

Ground States of the Mean-Field Spin Glass with 3-Spin Couplings

T0 review · 4 major / 3 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read This paper reports that ground states of the mean-field 3-spin Ising spin glass have finite-size corrections, distributions, and dilution behavior matching the random energy model (REM), making p=2 (SK) the exceptional case.

desk verdict A plausible and genuinely new numerical result—p=3 ground-state corrections look REM-like, unlike SK—but the unquantified heuristic bias and single-dilution evidence base keep it from being airtight. read the letter →

arxiv 2501.13205 v1 pith:R36WUDDH submitted 2025-01-22 cond-mat.dis-nn quant-ph

Stefan Boettcher , Ginger E. Lau (Emory U) This is my paper · ORCID
classification cond-mat.dis-nnquant-ph MSC 82B4482D3060G70
keywords mean-fieldspinglassp-spinmodelgroundstateenergyfinite-sizecorrectionsrandomextremaloptimizationGumbeldistributionSherrington-Kirkpatrick
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

The paper tries to establish that, at zero temperature, the mean-field spin glass with three-spin couplings (all triplets of N Ising spins coupled) is governed by the same finite-size corrections as the random energy model, the p=∞ limit. Using an extremal-optimization heuristic on GPUs to reach N=256, the authors find the average ground-state energy density approaches its thermodynamic limit with a ln N/N correction, the width of its distribution decays as 1/N, and the distribution is close to Gumbel. Bond dilution does not change the leading corrections, unlike the Sherrington-Kirkpatrick (p=2) model where the correction exponent varies with dilution. If true, all p-spin models with p≥3 behave REM-like at T=0, and SK's anomalous corrections (like the $N^{-5}$/6 width) are the exception rather than the rule.

What carries the argument

The load-bearing machinery is the extremal optimization (EO) heuristic, run in a GPU-vectorized, bit-packed form, which produces approximate ground states up to N=256; its reliability is judged by extrapolation plots to the exact thermodynamic limit. The analytic anchor is the exactly known ground-state energy density ⟨e0⟩_{p=3,∞} = −0.8132... from 1-step replica symmetry breaking calculations, which the fits are forced to reproduce. The finite-size correction ansatz C_N ~ ln N / N, borrowed from REM, is what discriminates between competing correction forms, and the overlap gap condition for p≥3 explains why the search is harder than for SK.

What would settle it

An exact or rigorously certified computation of ground states for, say, N=128 or 256 (or a demonstration that the EO systematic error itself decays as ln N/N) would settle the claim: if the true energies depart from the ln N/N extrapolation, or if the 1/N width scaling breaks down, the REM-like conclusion fails.

Watch

Extended reading notes

Core claim

The central claim is that the p=3 mean-field spin glass is REM-like in its zero-temperature ground-state statistics. Concretely: fixing the thermodynamic-limit energy to the exactly known value from 1-step replica calculations, the ensemble-averaged ground-state energy density follows ⟨e0⟩_N ~ ⟨e0⟩_∞ + A ln N / N, rather than a power-law; a power-law fit with exponent ω≈4/5 would miss the known limit by about 20σ. The standard deviation of the ground-state energy distribution scales as 1/N (with possible undetectable logarithmic factors), and the rescaled distribution is a generalized Gumbel with m≈0.87, close to the pure Gumbel of REM and far from SK's m≈5. The paper further claims that dilution to 25% bond density leaves these corrections essentially unchanged once energies are rescaled by 1/√q, in stark contrast to SK. The conclusion drawn is that p=2 is the special model, and all p≥3 share REM's ground-state corrections.

Load-bearing premise

The results stand on the assumption that the configurations the extremal-optimization heuristic returns are close enough to true ground states that the heuristic's systematic error decays faster than the measured ln N/N and 1/N corrections; there is no certificate of optimality.

Editorial extensions

If this is right

  • If correct, the zero-temperature ground-state properties of all p-spin models with p≥3 are universal and coincide with REM's, so p=3 data can serve as a benchmark for testing heuristics.
  • The overlap gap condition makes local search hard, but the observed ln N/N corrections imply that even for p=3, near-optima extrapolate reliably, extending the usefulness of extrapolation-based evaluation of heuristics.
  • The dilution independence gives a sharper distinction between p=2 and p≥3: any theory of SK's q-dependent correction exponent (ω(q)) must explain why p=3 has no such dependence.
  • The measured local-field pseudo-gap with P(0) ~ N^−0.6 provides a concrete signature for future mean-field or replica calculations to reproduce.

Reading between the lines

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

  • Editorial extension: a natural testable next step is to run the same extrapolation analysis for p=4 and p=5 to see whether the same ln N/N and Gumbel forms hold, which would corroborate the p≥3 universality without relying on the p→∞ limit.
  • Editorial extension: the claimed REM-like behavior at T=0 might connect to the finite-temperature random first-order transition; one could check whether the finite-size corrections cross over to SK-like behavior near the transition temperature.
  • Editorial extension: the heuristic's slower convergence for p=3 than for SK, despite similar system sizes, could be repurposed as a practical benchmark for the computational hardness implied by the overlap gap condition.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 3 minor

Summary. The paper studies ground-state properties of the fully connected p=3 Ising spin glass with all-to-all triplet bonds, using a GPU-vectorized extremal optimization (EO) heuristic for system sizes N=16 to 256, for the complete model and for a bond-diluted version with 25% of bonds present. From tabulated mean ground-state energies and standard deviations, the authors argue that the finite-size correction to the mean energy is C_N = ln N/N (Eq. 4), that the standardized ground-state energy distribution is close to a Gumbel form with m=0.87(5) (Eq. 5), that the standard deviation decays as 1/N, and that the local-field distribution develops a pseudogap. The extrapolation is anchored to the exactly known 1RSB thermodynamic limit <e_0>_∞ = -0.8132, which is used to reject a power-law correction that extrapolates to -0.8151. The results are contrasted with SK (p=2) and used to propose that all p-spin models with p>=3 share REM-like ground-state corrections.

Significance. If the central claim is correct, the paper establishes a sharp T=0 distinction between p=2 and p>=3 spin glasses and predicts a qualitative contrast in dilution dependence, which would be an important step toward a unified picture of finite-size corrections in mean-field spin glasses. The paper's strengths are its extensive tabulated data, its use of the exactly known thermodynamic limit as a model-selection constraint, and its explicit comparison of candidate correction forms. The central weakness is that all reported states are approximate optima of a heuristic whose systematic error is not quantified, and the exact-limit anchor constrains only the sum of the true correction and the heuristic bias. The stress-test concern therefore lands: the REM-like conclusion is plausible and well presented, but it is not yet certified.

major comments (4)
  1. [Eq. (3)-(4), Fig. 1, Table I, Fig. 5] The inference that C_N = ln N/N rests on extrapolating data that are explicitly labeled as approximate. Table I calls all entries 'approximate optima', Fig. 5 refers to the 'presumed ground state energy', and the adaptive stopping rule (use twice as many flips as were needed for the latest new optimum) provides no lower bound on the true ground-state energy. At N=256 the measured deviation from -0.8132 is about 0.0135 in energy density; a heuristic bias of a few times 10^-3 at that size would change the fitted amplitude A in <e0>_N = e_∞ + A ln N/N by tens of percent. Anchoring to the exact 1RSB limit and rejecting the power-law fit that extrapolates to -0.8151 are genuinely useful consistency checks, but they constrain only the combined effect of true corrections and bias and cannot separate the two. The authors should add an independent optimality benchmark, for example exact branch-and-bound or a second heuristic with a different landscape bias for smaller N, and report how any residual bias scales with N relative to ln N/N.
  2. [Table I, Fig. 1] The claim of no dilution dependence is supported by only one diluted value, q=0.25. The collapse of this single series onto the q=1 data after the 1/sqrt(q) rescaling from Ref. [26] is suggestive, but it cannot establish the abstract's stronger assertion that no variation with q is found, nor the broader conclusion that the absence of q-dependence is a general property of p>=3. The SK inset in Fig. 1 shows many q values, whereas the p=3 panel shows only q=1 and q=0.25. At least a few dilution values spanning different regimes are needed before claiming a qualitative contrast with SK, especially since the diluted data are produced by the same heuristic and may share the same systematic bias.
  3. [Fig. 3 and accompanying text] The variance scaling claim is internally inconsistent with the stated REM comparison. The text says the observed behavior 'matches more closely with REM (which has sigma ~ ln N/N)' than with SK, but Fig. 3 is presented as showing 'simply a series in 1/N'. These are parametrically different leading behaviors, and a logarithmic factor cannot be dismissed as hard to discern when the claimed distinction between candidate forms is the main message. The authors should fit sigma(e0) N against ln N, or a two-term form with both 1/N and ln N/N, and state whether the data actually favor a log-free 1/N law or a REM-like ln N/N law. This affects the central claim that p>=3 has exactly the same corrections as REM.
  4. [Eq. (5), Fig. 2] The Gumbel conclusion is based on pooling all sizes after standardizing by sigma and on a four-parameter fit, with fitted m=0.87(5). The parameter m is about 2.6 standard errors below the pure Gumbel value m=1, so the evidence for an exact Gumbel form is weaker than the abstract implies. The paper should report a goodness-of-fit measure and examine how the fitted parameters vary with N, since the pooled fit is dominated by the many small-N samples. Without such tests, 'consistent with a Gumbel distribution' is a reasonable summary, but it is not a quantitative confirmation of the REM-like distributional claim.
minor comments (3)
  1. [Paragraph discussing Fig. 2] There is a typo: 'there its very little variation' should read 'there is very little variation'.
  2. [Fig. 4 and Eq. (7)] The local-field pseudogap fit in Eq. (7) reports alpha ~ 0.6, A ~ 0.24, and B ~ 0.06 without uncertainties or a goodness-of-fit statement; please add fit details or label the values as preliminary.
  3. [Section IV, implementation paragraph] The GPU implementation is deferred to a future paper via Ref. [31]; since the numerical data are the main product of the manuscript, a short description of the vectorized update rule and the bit-packed bond representation would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the finite-size correction form is selected against an independent exact replica value, and self-citations are methodological only.

full rationale

The paper's central quantitative claim is an empirical characterization of ground-state energies produced by an extremal optimization heuristic. The choice of the ln N/N correction in Eq. (4) is not derived from an assumption that already contains it; instead, the paper rejects a power-law form because it extrapolates to -0.8151(1), some 20 sigma away from the exactly known value <e0>_infty = -0.8132 obtained from 1-step replica calculations [24,25], which are external to this work and not authored by the present authors. The ln N/N form is then adopted because, with the independent limit fixed, it yields a stable fit. The Gumbel distribution statement is descriptive: the fit returns m = 0.87(5) and is compared with, not imposed by, the REM value m = 1. The comparison with SK dilution behavior uses prior published data [18] as external benchmark. Self-citations appear only for the EO heuristic itself [29,30], for a vectorized implementation note [31], and for arguments that extrapolation plots can assess heuristic scalability [32,33]; none of these supplies the p=3 REM-like conclusion. The possible bias of the uncertified EO heuristic is a correctness risk, not a circularity, because the conclusion is not built into the input data by construction.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The ledger is dominated by fitted correction amplitudes and descriptive distribution parameters. The only external input is the replica thermodynamic limit, plus a heuristic-quality assumption that is never verified. No new physical entities are introduced. The dilution claim rests on one q value and an external rescaling theorem.

free parameters (5)
  • FSC amplitude for ln N/N correction = not reported (fit in Fig. 1)
    Coefficient A in Eq. (3) with C_N = ln N/N; determined by least-squares fit to the tabulated average energies after fixing the known infinite-size value.
  • Quadratic FSC coefficient = not reported
    Coefficient of C_N^2 in the dashed fit in Fig. 1, included to match data for N >= 32.
  • Rejected power-law exponent = omega ≈ 4/5 (within 1%)
    A free power-law fit gives an exponent near 4/5 but extrapolates to -0.8151, about 20 sigma from the known -0.8132; this rejection motivates the ln N/N form.
  • Gumbel distribution parameters = m=0.87(5), A=0.50(4), u=0.33(5), v=1.27(4)
    Four parameters fit to the combined distribution in Fig. 2; m close to 1 is offered as evidence for Gumbel/REM statistics.
  • P(h) pseudogap parameters = alpha≈0.6, A≈0.24, B≈0.06
    Eq. (7) fits the small-field behavior of the local-field distribution; peripheral to the central finite-size scaling claim.
assumptions (4)
  • domain assumption The 1-step replica calculation gives the exact thermodynamic limit <e0>_infinity = -0.8132 for p=3 (Refs [24,25]).
    This constant is fixed in Eq. (3) for the ln N/N extrapolation and is the basis for rejecting the power-law fit. If this value were wrong, the correction-form selection would collapse.
  • standard math The p-spin Hamiltonian normalization in Eq. (2), J0 = sqrt(2/(N^{p-1} p!)), defines the energy scale and the ensemble.
    All energy densities, the exact limit, and the 1/sqrt(q) dilution rescaling refer to this normalization; it is standard and not derived in the paper.
  • ad hoc to paper The EO heuristic optima have systematic error negligible compared with ln N/N at the sizes studied.
    No exact ground-state comparisons or convergence certificates are provided; Table I calls the results approximate optima and Fig. 5 calls the energy a presumed ground state energy.
  • domain assumption The Carmona-Hu 1/sqrt(q) rescaling applies to the p=3 diluted model.
    It is used to collapse the q=0.25 data onto the q=1 curve in Fig. 1, and only one dilution value is measured.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Ground States of the Mean-Field Spin Glass with 3-Spin Couplings." pith.science (2026). https://pith.science/paper/R36WUDDH

@misc{pith2026250113205,
  author       = {Pith},
  title        = {Pith review of: Ground States of the Mean-Field Spin Glass with 3-Spin Couplings},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R36WUDDH}},
  note         = {Machine review of arXiv:2501.13205}
}
abstract

We use heuristic optimization methods in extensive computations to determine with low systematic error ground state configurations of the mean-field $p$-spin glass model with $p=3$. Here, all possible triplets in a system of $N$ Ising spins are connected with a bond. This model has been of recent interest, since it exhibits the ``overlap gap condition'', which should make it prohibitive to find ground states asymptotically with local search methods when compared, for instance, with the $p=2$ case better-known as the Sherrington-Kirkpatrick model (SK). Indeed, it proves more costly to find good approximations for $p=3$ than for SK, even for our heuristic. Compared to SK, the ground-state behavior for $p=3$ is quite distinct also in other ways. For SK, finite-size corrections for large system sizes $N\to\infty$ of both, the ensemble average over ground state energy densities and the width of their distribution, vary anomalously with non-integer exponents. In the $p=3$ case here, the energy density and its distribution appear to scale with $\ln N/N$ and $1/N$ corrections, respectively. The distribution itself is consistent with a Gumbel form. Even more stark is the contrast for the bond-diluted case, where SK has shown previously a notable variation of the anomalous corrections exponent with the bond density, while for $p=3$ no such variation is found here. Hence, for the 3-spin model, all measured corrections scale the same as for the random energy model (REM), corresponding to $p=\infty$. This would suggest that all $p$-spin models with $p\geq3$ exhibit the same ground-state corrections as in REM.

Figures

Figures reproduced from arXiv: 2501.13205 by the authors.

Figure 1
Figure 1. Extrapolation plot for the average ground state en [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Distribution of ground state energy density fluctu [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Plot of the distribution P(h) of local fields h im￾pinging on spins by their neighbors in the ground state con￾figurations of the 3-spin model, plotted for different system sizes. (Since P(h) is symmetric in h, only values for h ≥ 0 are shown.) Although significant finite-size effects are appar￾ent, there is only a small variation between the two largest sizes, suggesting that the shape for P(h) at N = 256 is close … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 30 canonical work pages

  1. [26]

    Boettcher, Journal of Statistical Mechanics: Theory and Experiment 2010, P07002 (2010)

    S. Boettcher, Journal of Statistical Mechanics: Theory and Experiment 2010, P07002 (2010)

    work page 2010

  2. [1]

    S. F. Edwards and P. W. Anderson, J. Phys. F5, 965 (1975)

    work page 1975

  3. [2]

    Sherrington and S

    D. Sherrington and S. Kirkpatrick, Phys. Rev. Lett.35, 1792 (1975)

    1975

  4. [3]

    properties, we have first arranged the energies⟨e0⟩N in an extrapolation plot according to some presumed form of the finite-size corrections (FSC), ⟨e0⟩N ∼ ⟨e0⟩∞ + ACN +

    Note the much more significant variation in corrections withq in SK compared to the 3-spin model. properties, we have first arranged the energies⟨e0⟩N in an extrapolation plot according to some presumed form of the finite-size corrections (FSC), ⟨e0⟩N ∼ ⟨e0⟩∞ + ACN + . . . , (N → ∞) , (3) where CN is the expected correction that should vanish for N → ∞. I...

    work page

  5. [4]

    overlap gap condition

    Its behavior is quite similar to that observed for SK at finite N [28]. The distribution is somewhat broader here but similarly develops a pseudo-gap, i.e.,P (h) → 0 4 0 1 2 3 4 5 6 7 h 0 0.1 0.2 P(h) SK N=16 N=32 N=64 N=128 N=256 16 32 64 128 256 N 0.01 0.1 P(0) Figure 4. Plot of the distribution P (h) of local fields h im- pinging on spins by their neig...

    work page

  6. [5]

    for the ground state energy density atN = ∞, which is unacceptable, since it is some 20σ from the exactly-known value of ⟨e0⟩p=3 ∞ = −0.8132

    However, the extrapolation predicts a fitted value of −0.8151(1) . . .for the ground state energy density atN = ∞, which is unacceptable, since it is some 20σ from the exactly-known value of ⟨e0⟩p=3 ∞ = −0.8132 . . ., based on 1-step replica calcula- tions [24, 25]. Fixing ⟨e0⟩p=3 ∞ in the power-law fit does not produce a stable result anymore, even when ...

    work page

  7. [6]

    Charbonneau, E

    P. Charbonneau, E. Marinari, and M. Mezard, eds.,Spin glass theory and far beyond (World Scientific, Singapore, 2023)

    work page 2023

  8. [7]

    Mézard and A

    M. Mézard and A. Montanari, Constraint Satisfaction Networks in Physics and Computation (Oxford Univer- sity Press, Oxford, 2006)

    work page 2006

Show all 36 references
  1. [8]

    Kochenberger, J.-K

    G. Kochenberger, J.-K. Hao, F. Glover, M. Lewis, Z. Lü, H. Wang, and Y. Wang, Journal of Combinatorial Opti- mization 28, 58 (2014)

    2014

  2. [9]

    S.Boettcher,PhysicalReview Research 1, 033142 (2019)

    2019

  3. [10]

    C. C. McGeoch and C. Wang, inProceedings of the ACM International Conference on Computing Frontiers , CF ’13 (ACM, New York, NY, USA, 2013) pp. 23:1–23:11

    2013

  4. [11]

    Albash, T

    T. Albash, T. Ronnow, M. Troyer, and D. Lidar, The Eu- ropean Physical Journal Special Topics224, 111 (2015)

    2015

  5. [12]

    Rosenberg, M

    G. Rosenberg, M. Vazifeh, B. Woods, and E. Haber, Computational Optimization and Applications 65, 845 (2016)

    2016

  6. [13]

    H. Goto, K. Tatsumura, and A. R. Dixon, Science Ad- vances 5, eaav2372 (2019)

    2019

  7. [14]

    H. Goto, K. Endo, M. Suzuki, Y. Sakai, T. Kanao, Y. Hamakawa, R. Hidaka, M. Yamasaki, and K. Tat- sumura, Science Advances7, eabe7953 (2021)

    2021

  8. [15]

    Aramon, G

    M. Aramon, G. Rosenberg, E. Valiante, T. Miyazawa, H. Tamura, and H. G. Katzgraber, Frontiers in Physics 7, 10.3389/fphy.2019.00048 (2019)

    2019

  9. [16]

    Gamarnik, Proceedings of the National Academy of Sciences 118, 10.1073/pnas.2108492118 (2021)

    D. Gamarnik, Proceedings of the National Academy of Sciences 118, 10.1073/pnas.2108492118 (2021)

    2021 doi

  10. [17]

    A. E. Alaoui and A. Montanari, arxiv:2009.11481 (2020)

    2020 arXiv

  11. [18]

    Derrida, Phys

    B. Derrida, Phys. Rev. Lett.45, 79 (1980)

    1980

  12. [19]

    Derrida, Phys

    B. Derrida, Phys. Rev. B24, 2613 (1981)

    1981

  13. [20]

    M. R. Garey and D. S. Johnson, Computers and In- tractability: A Guide to the Theory of NP-Completeness (W. H. Freeman, New York, 1979)

    1979

  14. [21]

    Boettcher, Physical Review Letters 124, 177202 (2020)

    S. Boettcher, Physical Review Letters 124, 177202 (2020)

    2020

  15. [22]

    Wang, Physical Review B106, 134209 (2022)

    W. Wang, Physical Review B106, 134209 (2022)

    2022

  16. [23]

    Gardner, Nuclear Physics B257, 747 (1985)

    E. Gardner, Nuclear Physics B257, 747 (1985)

    1985

  17. [24]

    Boettcher, The European Physical Journal B46, 501 (2005)

    S. Boettcher, The European Physical Journal B46, 501 (2005)

    2005

  18. [25]

    T.Aspelmeier, A.Billoire, E.Marinari,andM.A.Moore, Journal of Physics A: Mathematical and Theoretical41, 324008 (2008)

    2008

  19. [27]

    Montanari and F

    A. Montanari and F. Ricci-Tersenghi, The European Physical Journal B - Condensed Matter33, 339 (2003)

    2003

  20. [28]

    Yeo and M

    J. Yeo and M. A. Moore, Physical Review E101, 032127 (2020)

    2020

  21. [29]

    Carmona and Y

    P. Carmona and Y. Hu, Annales de l’Institut Henri Poincare (B) Probability and Statistics42, 215 (2006)

    2006

  22. [30]

    Parisi and T

    G. Parisi and T. Rizzo, Phys. Rev. Lett. 101, 117205 (2008)

    2008

  23. [31]

    Boettcher, H

    S. Boettcher, H. G. Katzgraber, and D. Sherrington, Journal of Physics A: Mathematical and Theoretical41, 324007 (2008)

    2008

  24. [32]

    Boettcher and A

    S. Boettcher and A. G. Percus, Physical Review Letters 86, 5211 (2001)

    2001

  25. [33]

    Boettcher and A

    S. Boettcher and A. Percus, Artificial Intelligence119, 275 (2000)

    2000

  26. [34]

    The details of this implementation will be discussed else- where

  27. [35]

    Boettcher, Nature Machine Intelligence5, 24 (2022)

    S. Boettcher, Nature Machine Intelligence5, 24 (2022)

    2022

  28. [36]

    Boettcher, Nature Communications14, 5658 (2023)

    S. Boettcher, Nature Communications14, 5658 (2023)

    2023

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.