Large deviations for the maximum of the generalized TAP free energy

Jeanne Boursier

REVIEW 3 major objections 4 minor 27 references

Large deviations for the maximum of the generalized TAP free energy

T0 review · 3 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read The supersymmetric state-counting formula of spin-glass physics is really a large-deviation rate: it governs the probability, not the average number, of high-free-energy TAP states.

desk verdict A substantial and honest paper that identifies the SUSY complexity with a large-deviation exponent, but the main theorem is explicitly conditional on an unproven strict Plefka inequality and should go to a serious referee rather than be accepted as-is. read the letter →

arxiv 2607.29330 v1 pith:H4VGENEC submitted 2026-07-31 math.PR math-phmath.MP

Jeanne Boursier This is my paper · ORCID

classification math.PRmath-phmath.MP MSC 60F1082B44

keywords TAPfreeenergylargedeviationsspinglasssupersymmetriccomplexityParisiformulasphericalcodeKac–RicePlefkacondition

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 aims to pin down what the physicists' supersymmetric TAP complexity formula actually computes, a question that has lingered since the late 1970s. It proves that, for regular free-energy levels above equilibrium and under a strict stability condition, the probability that the TAP landscape contains a maximum at level f decays as exp(-N Σ(f)), where Σ(f) is the Legendre transform of a constrained Parisi value. The same exponent is shown to be the annealed exponent of a restricted class of supersymmetric critical points that are isolated local maxima; because these states form a one-sided spherical code, their annealed count doubles as a lower bound on the probability of existence. The result resolves the controversy: the supersymmetric formula is a large-deviation rate, not the ordinary annealed complexity, correcting earlier conjectures. A reader should care because it gives a concrete probabilistic meaning to formulas physicists use to count metastable states.

What carries the argument

The load-bearing objects are ε-SUSY states: critical points of the generalized TAP free energy at level f whose self-overlap, free energy, and empirical cavity-field law lie within ε of those prescribed by the constrained Parisi minimizer and its associated diffusion. Two inequalities do the work: the strict Plefka condition (1.8) gives a uniform negative Hessian upper bound around each retained state, and the strict positivity of the Parisi obstacle on (0, q_f) excludes competitors at intermediate overlaps. Distinct retained states therefore have pairwise overlap below δ; radially normalized, they form a one-sided spherical code, and a classical coding bound caps their number at exp(o(N)) f

What would settle it

Evaluate the strict Plefka factor 1 − ξ''(q_f) E[(∂xx Φ_{ζ_f}(q_f, X_{q_f}))^2] at the first positive contact point of the constrained Parisi minimizer for a concrete mixture, for instance ξ(x) = x^2 + x^4. If it equals zero, the lower-bound construction fails and the rate identity (1.9) would need modification. A numerical check would simulate the TAP landscape at moderate N and compare the empirical rate of existence of maxima near level f with Σ(f) computed from the constrained Parisi problem.

Watch

Extended reading notes

Core claim

The paper's central claim: the physicists' supersymmetric complexity formula — a Legendre transform in the bottom-atom mass of the Parisi order parameter — is the large-deviation exponent for the existence of TAP maximum states, not the ordinary annealed complexity. For regular f ∈ (f_eq, f_1), under the strict Plefka condition, (1/N) log P(E_{N,ε}(f)) → Σ(f) = inf_θ [Λ(θ) − θ f] in the relevant double limit, with Λ the constrained Parisi value. The lower bound retains only ε-SUSY states, whose pairwise overlaps are forced below any fixed δ; as a one-sided spherical code, their annealed count becomes an existence-probability lower bound. The upper bound is a new interpolation comparison for

Load-bearing premise

The argument assumes that the constrained Parisi minimizer satisfies the strict Plefka inequality at its first positive contact point; only the non-strict inequality is automatic, and if strictness ever fails the construction of isolated SUSY states and the Hessian gap collapses.

Editorial extensions

If this is right

The probability that the TAP landscape has a maximum at level f decays at rate Σ(f); the familiar Legendre formula is a large-deviation rate, not the mean state count.
Because the annealed and quenched exponents can differ by the conditional expected number of states given that at least one exists, the supersymmetric computation should be read as counting isolated local maxima with negative-definite Hessians, which legitimizes dropping the absolute value of the determinant.
Only the non-strict Plefka inequality is automatic for the constrained minimizer; the strict version is a technical assumption the author expects can be removed (Remark 1.2).
The same estimates give a lower bound for upper deviations of the ordinary free energy at regular levels (Proposition 3.12).
Iterating the band argument above TAP ancestors at successive contact points would give a constructive proof of the Parisi formula for the Ising model under a technical strict-stability assumption.

Reading between the lines

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

Other supersymmetric calculations in disordered systems — for example earlier Kac–Rice counts of metastable states — may likewise be the rare-event rates of the extremal landscape rather than equilibrium complexity; the spherical counterexamples in Appendix C support re-reading them this way.
The spherical-code mechanism is generic: any class of critical points whose annealed first moment is computable and whose mutual overlaps are uniformly bounded can be converted into an existence lower bound, a template that could transfer to other Gaussian landscapes such as Gaussian fields on the sphere or the hypercube.
The rate identity in pure even p-spin models, where strict convexity of ξ fails and the perturbation reduction is applied, is a natural test case for whether the technical assumptions are actually necessary.

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.

Desk Editor's Note

read the letter

First thing you should know: this is a real result, but the central theorem is a conditional theorem. The strict Plefka condition (1.8) is load-bearing, and the paper says so itself in Remark 1.2. If equality holds for some model — and Appendix C shows these quantities can vanish in spherical analogues — the lower bound in Section 3.7 loses its spectral gap and the spherical-code argument collapses. That doesn't make the paper bad; it makes it honest about what remains.

What's genuinely new is the identification of the supersymmetric formula with the large-deviation exponent for existence of TAP maxima, not the annealed complexity. That is a clean conceptual correction to the physics conjectures, and the spherical counterexamples in Appendix C are a nice contribution in their own right. The Guerra-type interpolation for the upper bound is new and likely reusable, and adapting the Huang–Sellke lower-bound strategy to the Ising TAP landscape is a nontrivial piece of work. The consistency check with Talagrand's fractional-moment formula gives an independent sanity check that the rate function is the right one.

The paper also does the right thing by displaying the strict Plefka assumption in the theorem statement and conceding that only the non-strict inequality is automatic. Appendix B contains real work: the conditional Hessian estimates, the free-convolution gap, and the determinant lower tail are substantive. The spherical-code step is a neat trick.

Soft spots, in proportion. First and main: the strict Plefka condition is assumed, not proved. The stress-test note is right that equality is not a purely hypothetical pathology; Appendix C shows examples where the analogous quantity vanishes at an RSB transition. Second: pure even p-spin is excluded by the strict convexity assumption on ξ, and the perturbation in Section 3.2 is only sketched — no uniformity argument for γ↓0 is given. Third: Appendix A is quoted without proof; that is fine for standard results, but the first-variation formulas there are load-bearing, and the paper's own note understates their role. These are real gaps, but they are the kind a referee can work with, not signs of a hollow argument.

Who this is for: people doing spin-glass landscape complexity or large deviations for TAP free energies. It deserves a serious referee, not a desk reject. My recommendation: send it out, and put the strict-Plefka issue front and center in the review. The paper is a conditional proof of a plausible and important statement, and the referee's job is to decide whether the condition can be removed or is generic.

Referee Report

3 major / 4 minor

Summary. The paper studies the large deviations of the maximum of the generalized TAP free energy for the Ising mixed p-spin model. The upper bound is obtained by a new Guerra-type interpolation comparing the TAP maximum with an additive Ruelle probability cascade, yielding the expected maximum and its exponential moments. The lower bound follows the Huang--Sellke strategy adapted to the Ising TAP landscape: one computes the annealed complexity of certain 'SUSY states', shows with a local Hessian estimate that isolated SUSY states form a spherical code, and converts the first moment into an existence probability. The main result, Theorem 1, states that at regular levels f in (f_eq,f_1), under the strict Plefka condition (1.8), the probability that a TAP critical point exists at level f has rate Σ(f)=inf_θ[Λ(θ)-θf], and the corresponding exponential moment converges to Λ(θ_f). The paper also contains a transfer argument (Proposition 3.12) giving a lower bound for upper deviations of the free energy, and a discussion of a constructive route to the Parisi formula.

Significance. If the claims hold, this is a substantial contribution: it clarifies the long-standing question of what the supersymmetric complexity counts, identifying the Legendre transform of the constrained Parisi value with the large-deviation exponent for the existence of TAP maxima rather than with the annealed count. The upper-bound Guerra-type interpolation is new and potentially useful. The lower-bound strategy is well structured and connects recent ideas of Huang--Sellke and Subag with the Ising TAP setting. The paper is careful to state its main theorem conditionally, and the appendices contain detailed computations, including verifiable spherical counterexamples. The central limitation is that the strict Plefka condition (1.8) is assumed rather than proved for the relevant constrained minimizer, and the reduction from nondegenerate mixtures to pure even p-spin models is only sketched. Conditional on these points being resolved, the paper is significant and publishable.

major comments (3)

[§1.3, Theorem 1 and Remark 1.2; §3.7 and Proposition B.1] The strict Plefka condition (1.8) is assumed, not established. Remark 1.2 explicitly concedes that only the non-strict inequality is automatic and that strictness is merely 'technical'. The lower-bound construction depends crucially on strictness: the spectral gap in Proposition B.1, the uniform negative Hessian bound in Definition 3.5(ii), and therefore the spherical-code exclusion in Step 2 of the proof of Proposition 3.3 all collapse if equality holds at q_f. Appendix C shows that the analogous inequalities are sensitive to the model, so equality is not a vacuous possibility. The paper needs either a proof of (1.8) under natural conditions, a demonstration that it holds for a nontrivial family of models, or an explicitly formulated version of the theorem that avoids the reliance on strictness. As written, the main result is conditional on a load-bearing unproved spectral-gap hypothesi
[§3.2, 'For the Kac–Rice arguments ... pure case'] Assumption 1.1 requires strict convexity of ξ on [-1,1], which excludes pure even p-spin mixtures such as ξ(x)=β^2 x^p with p even. The perturbation adding an independent even ℓ-spin Hamiltonian is described in a single paragraph: 'Continuity of the Parisi and TAP quantities preserves the strict Plefka inequality for small γ. ... Letting γ↓0 after N→∞ proves the pure case.' No argument is given that the constrained minimizer ζ_f, the first contact point q_f, the rate function Σ(f), or the strict Plefka inequality pass through the γ→0 limit uniformly. Since the abstract claims the result for Ising mixed p-spin models generally, this is a load-bearing gap. The proof of the pure even p case must be supplied in full or the theorem must be restricted to mixtures satisfying Assumption 1.1 with a separate treatment of the limiting case.
[§1.3, Theorem 1 vs. Abstract and Introduction] The theorem is correctly stated with the strict Plefka condition, but the abstract and the introductory discussion present the identification of the supersymmetric formula with the large-deviation exponent without this caveat. Given that the strict Plefka condition is not verified for any model in the paper, the unconditional presentation overstates the result. The authors should either establish the condition in a meaningful class of examples or qualify the abstract and introduction accordingly.

minor comments (4)

[Appendix A] Appendix A explicitly states that all results are 'quoted without proof'. This is acceptable if these are standard and the references are complete, but since several of these identities (e.g., (A.17), (A.20), Lemma A.4) are used in load-bearing parts of the paper, the authors should state more precisely which of them are proved in the cited literature and which are assembled from [7]. The current phrasing may leave the reader uncertain about the provenance of the key variational identities.
[§3.2, pure case perturbation] The paragraph begins 'For γ>0, the joint law of (∇H^γ_N(m), H^γ_N(m)) is nondegenerate when q_m>0.' If the added ℓ is chosen as 2 (when p is even and p≠2), this is fine, but if p=2 and one adds an even ℓ>2, strict convexity still fails at 0; the paragraph should specify the choice explicitly. Also, 'Letting γ↓0 after N→∞ proves the pure case' is a strong statement that needs the missing uniformity mentioned above.
[Notation, §1.7] The notation uses both ˆζ and ˆζζ in the same paragraph; the latter appears in the definition of the cumulant (1.11) with a wide hat, while other places use ˆζ. This is analogous but not identical, and could confuse the reader. Please unify.
[Throughout] The paper contains several forward references to appendices that are not yet available in the version under review; this is normal, but the appendix statements should be checked for consistency with the main text, especially the signs in the Hessian estimates and the contact equations.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the TAP large-deviation rate is independently derived and cross-checked against Talagrand's LDP; the main caveat is an explicitly assumed strict Plefka condition.

full rationale

I walked the derivation chain and found no step in which a claimed prediction reduces by construction to a fitted input or to a definitional identity. The rate function Σ(f) is not fitted to the event E_{N,ε}(f): it is defined through the constrained Parisi value Λ(θ), and the upper bound derives limsup E exp{θ max F_TAP} ≤ Λ(θ) via a Guerra-type comparison (Prop. 2.7), while the lower bound computes the annealed SUSY critical-point count by Kac–Rice (Prop. 3.6) and then converts this count into an existence probability through the spherical-code bound (Lemma 3.11). These are independent derivations that meet at the same variational object. The paper is also benchmarked externally: Remark 1.3 notes that Talagrand's large-deviation principle [26, Section 7] identifies the same -Σ as the upper-tail rate of N^{-1} log Z_N, so the main exponent is not being asserted solely from the paper's own construction. The strict Plefka condition (1.8) is an explicit hypothesis of Theorem 1, not an output of the proof; Remark 1.2 concedes that only the non-strict inequality is automatic and that strictness is technical. That is a limitation, and the pure even-p perturbation in §3.2 is only sketched, but a missing or conditional spectral-gap proof is a correctness risk rather than circularity. The paper does rely on the author's prior work [7] for several technical identities (e.g., Auffinger–Chen formulas, left-edge formula in Proposition B.1, deformed-GOE determinant comparison), and Appendix A quotes standard statements without proof. This self-citation is real and load-bearing in the technical sense, but it does not make the central claim equivalent to its inputs: the spherical-code argument and the existence-probability conversion are new and independent, and the final rate matches an external benchmark. I therefore find no exhibited circular reduction and assign a low score.

Assumptions & free parameters 2 free parameters · 3 assumptions · 1 invented entities

The paper introduces no extra free parameters beyond the standard Parisi variational problem. The main assumptions are the strict Plefka condition and strict convexity of xi, both explicitly flagged as technical. The paper's proposed constructive proof of the Parisi formula is conditional on an additional unproven strict stability assumption.

free parameters (2)

regular level f = any f in (f_eq, f_1)
Chosen by the user of the theorem rather than fitted; it parameterizes the rate function, not a fitted constant.
theta_f (minimizer) = the unique minimizer of Lambda(theta) - theta f
Defined as the argmin of the rate function; it is not a free parameter but a derived quantity.

assumptions (3)

domain assumption stricte Plefka condition (1.8)
Assumed in Theorem 1 and Proposition 3.3; only the non-strict version is automatic. The author states it is technical.
domain assumption strict convexity of xi on [-1,1] (Assumption 1.1)
Used in the Slepian comparison in Lemma 2.4 and in the covariance comparison in Proposition 3.9. Pure even p-spin models fail it; a perturbation is only sketched.
domain assumption Technical strict stability assumption for the Parisi formula sketch
Explicitly mentioned in Section 1.3 for the constructive proof of the Parisi formula, which is only a claim of a future program.

invented entities (1)

SUSY states (Definition 3.4)
purpose: Selected critical points with negative definite Hessian used in the lower bound
This is a modeling construct, not a new physical entity.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Large deviations for the maximum of the generalized TAP free energy." pith.science (2026). https://pith.science/paper/H4VGENEC

@misc{pith2026260729330,
  author       = {Pith},
  title        = {Pith review of: Large deviations for the maximum of the generalized TAP free energy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H4VGENEC}},
  note         = {Machine review of arXiv:2607.29330}
}

abstract

We study large deviations of the maximum of the generalized TAP free energy introduced by Chen, Panchenko, and Subag for the Ising mixed $p$-spin model. The upper bound uses a new Guerra-type interpolation that may also be useful for other spin-glass models. The lower bound follows the strategy of Huang and Sellke. We show that supersymmetric critical points have no competitor at positive overlap on their own sphere. As such, they form a spherical code, which turns the annealed count into a lower bound on the probability of existence. This identifies the supersymmetric formula proposed by physicists with the large-deviation exponent for the existence of TAP maxima, rather than with the ordinary annealed complexity. The resulting rate function is the Legendre transform of a constrained Parisi value in the bottom-atom mass. The same argument, iterated in bands above TAP ancestors, would give a constructive proof of the Parisi formula for the Ising model under a technical strict stability assumption.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 3 linked inside Pith

[1]

Annibale, A

A. Annibale, A. Cavagna, I. Giardina, G. Parisi, and E. Trevigne,The role of the Becchi–Rouet–Stora–Tyutin su- persymmetry in the calculation of the complexity for the Sherrington–Kirkpatrick model, J. Phys. A: Math. Gen.36 (2003), no. 43, 10937–10953

2003
[2]

Auffinger, G

A. Auffinger, G. Ben Arous, and J. Černý,Random matrices and complexity of spin glasses, Comm. Pure Appl. Math. 66(2013), no. 2, 165–201

2013
[3]

Auffinger and W.-K

A. Auffinger and W.-K. Chen,The Parisi formula has a unique minimizer, Comm. Math. Phys.335(2015), no. 3, 1429–1444

2015
[4]

Probab.45(2017), no

,Parisi formula for the ground state energy in the mixedp-spin model, Ann. Probab.45(2017), no. 6B, 4617–4631

2017
[5]

Belius, F

D. Belius, F. Concetti, and G. Genovese,On the determinant in Bray–Moore’s TAP complexity formula, arXiv:2401.08529, 2024

arXiv 2024
[6]

Biane,On the free convolution with a semi-circular distribution, Indiana Univ

P. Biane,On the free convolution with a semi-circular distribution, Indiana Univ. Math. J.46(1997), no. 3, 705–718

1997
[7]

Boursier,The Legendre structure of the TAP complexity for the Ising spin glass, arXiv:2604.20660, 2026

J. Boursier,The Legendre structure of the TAP complexity for the Ising spin glass, arXiv:2604.20660, 2026

arXiv 2026
[8]

A. J. Bray and M. A. Moore,Metastable states in spin glasses, J. Phys. C: Solid State Phys.13(1980), no. 19, L469–L476

1980

Show all 27 references

[9]

Capitaine, C

M. Capitaine, C. Donati-Martin, D. Féral, and M. Février,Free convolution with a semicircular distribution and eigenvalues of spiked deformations of Wigner matrices, Electron. J. Probab.16(2011), no. 64, 1750–1792

2011
[10]

Cavagna, I

A. Cavagna, I. Giardina, G. Parisi, and M. Mézard,On the formal equivalence of the TAP and thermodynamic methods in the SK model, J. Phys. A: Math. Gen.36(2003), no. 5, 1175–1194

2003
[11]

H.-B. Chen, A. Guionnet, J. Ko, B. Lacroix-A-Chez-Toine, and J.-C. Mourrat,One-sided large deviations for the ground-state energy of spin glasses, arXiv:2603.06368, 2026

2026
[12]

W.-K. Chen, D. Panchenko, and E. Subag,The generalized TAP free energy II, Comm. Math. Phys.381(2021), no. 1, 257–291

2021
[13]

Pure Appl

,The generalized TAP free energy, Comm. Pure Appl. Math.76(2023), no. 7, 1329–1415

2023
[14]

Crisanti, L

A. Crisanti, L. Leuzzi, G. Parisi, and T. Rizzo,Complexity in the Sherrington–Kirkpatrick model in the annealed approximation, Phys. Rev. B68(2003), no. 17, 174401

2003
[15]

Huang and M

B. Huang and M. Sellke,A constructive proof of the spherical Parisi formula, arXiv:2311.15495, 2023

2023 arXiv
[16]

Jekel, J

D. Jekel, J. S. Sandhu, and J. Shi,Potential hessian ascent: The Sherrington–Kirkpatrick model, Proceedings of the 2025 Annual ACM–SIAM Symposium on Discrete Algorithms (SODA), SIAM, 2025, pp. 5307–5387. 50 JEANNE BOURSIER

2025
[17]

Knowles and J

A. Knowles and J. Yin,Anisotropic local laws for random matrices, Probab. Theory Related Fields169(2017), no. 1–2, 257–352

2017
[18]

Panchenko,The Sherrington–Kirkpatrick model, Springer Monographs in Mathematics, Springer, New York, 2013

D. Panchenko,The Sherrington–Kirkpatrick model, Springer Monographs in Mathematics, Springer, New York, 2013

2013
[19]

Parisi and T

G. Parisi and T. Rizzo,On supersymmetry breaking in the computation of the complexity, J. Phys. A: Math. Gen.37 (2004), no. 33, 7979–7992

2004
[20]

R. A. Rankin,The closest packing of spherical caps inndimensions, Proc. Glasgow Math. Assoc.2(1955), 139–144

1955
[21]

Subag,The complexity of sphericalp-spin models—a second moment approach, Ann

E. Subag,The complexity of sphericalp-spin models—a second moment approach, Ann. Probab.45(2017), no. 5, 3385–3450

2017
[22]

Pure Appl

,Following the ground states of full-RSB spherical spin glasses, Comm. Pure Appl. Math.74(2021), no. 5, 1021–1044

2021
[23]

J.173(2024), no

,Free energy landscapes in spherical spin glasses, Duke Math. J.173(2024), no. 7, 1291–1357

2024
[24]

Subag and O

E. Subag and O. Zeitouni,The extremal process of critical points of the purep-spin spherical spin glass model, Probab. Theory Related Fields168(2017), no. 3–4, 773–820

2017
[25]

,Concentration of the complexity of spherical purep-spin models at arbitrary energies, J. Math. Phys.62 (2021), no. 12, 123301

2021
[26]

Talagrand,Large deviations, Guerra’s and A.S.S

M. Talagrand,Large deviations, Guerra’s and A.S.S. schemes, and the Parisi hypothesis, J. Stat. Phys.126(2007), no. 4–5, 837–894

2007
[27]

D. J. Thouless, P. W. Anderson, and R. G. Palmer,Solution of ‘solvable model of a spin glass’, Philos. Mag.35 (1977), no. 3, 593–601. Department of Mathematics, Massachusetts Institute of Technology, Cambridge, MA, USA Email address:boursier@mit.edu

1977

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [1]

Annibale, A

A. Annibale, A. Cavagna, I. Giardina, G. Parisi, and E. Trevigne,The role of the Becchi–Rouet–Stora–Tyutin su- persymmetry in the calculation of the complexity for the Sherrington–Kirkpatrick model, J. Phys. A: Math. Gen.36 (2003), no. 43, 10937–10953

2003

[2] [2]

Auffinger, G

A. Auffinger, G. Ben Arous, and J. Černý,Random matrices and complexity of spin glasses, Comm. Pure Appl. Math. 66(2013), no. 2, 165–201

2013

[3] [3]

Auffinger and W.-K

A. Auffinger and W.-K. Chen,The Parisi formula has a unique minimizer, Comm. Math. Phys.335(2015), no. 3, 1429–1444

2015

[4] [4]

Probab.45(2017), no

,Parisi formula for the ground state energy in the mixedp-spin model, Ann. Probab.45(2017), no. 6B, 4617–4631

2017

[5] [5]

Belius, F

D. Belius, F. Concetti, and G. Genovese,On the determinant in Bray–Moore’s TAP complexity formula, arXiv:2401.08529, 2024

arXiv 2024

[6] [6]

Biane,On the free convolution with a semi-circular distribution, Indiana Univ

P. Biane,On the free convolution with a semi-circular distribution, Indiana Univ. Math. J.46(1997), no. 3, 705–718

1997

[7] [7]

Boursier,The Legendre structure of the TAP complexity for the Ising spin glass, arXiv:2604.20660, 2026

J. Boursier,The Legendre structure of the TAP complexity for the Ising spin glass, arXiv:2604.20660, 2026

arXiv 2026

[8] [8]

A. J. Bray and M. A. Moore,Metastable states in spin glasses, J. Phys. C: Solid State Phys.13(1980), no. 19, L469–L476

1980