{"id":"d575364c-5e8d-421f-956e-5c2ae9afdbc9","arxiv_id":"2507.22396","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CLuP±Hop approximates Hopfield ground state free energies to within about 0.3% using simple gradient descent, backed by the author's fully lifted random duality theory.","lead":"This paper presents CLuP±Hop, a gradient-based algorithm that numerically computes near-ground-state free energies of positive and negative Hopfield models, reaching about 1.77 and 0.33 for n in the thousands. It argues these problems are typically easy and reports a structural difference between the two models in near-optimal configurations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim rests on an imported strong-duality theorem (Thm. 1) whose hypotheses are not checked for the entropy-constrained set X(rx,r̄x); the Table 1 'theory' column is a self-generated benchmark.","rationale":"The reader's weakest assumption correctly identifies the dependency on fl RDT Theorem 1, and I agree that this is the load-bearing point. The paper's internal consistency is good: equations (25)–(26) reduce the CLuP objective to ξ(rx,r̄x), and the 3-spl dynamics curves match the simulations at moderate n. But those successes do not validate the 6-spl thermodynamic limits used as ground truth in Table 1, because the same theorem generates both the prediction and the interpretation of the simulation. The CLuP algorithm itself is a heuristic gradient descent with manual parameters κ, t0, c; even if the output is reproducible, the claim that it 'achieves' the ground state free energy is only as strong as the imported theorem. No independent rigorous lower/upper bounds are cited for the −Hop limit, and for +Hop the predicted 1.77842 lies inside the known interval [1.7632, 1.7832] without being singled out by an external argument. Thus the CONDITIONAL verdict is appropriate; the concern does not move the verdict, but it should remain conditional until Theorem 1 is verified for this domain or an independent benchmark is supplied.","tokens_in":25007,"tokens_out":10619,"duration_ms":127425,"concrete_test":"Attempt a self-contained derivation of Theorem 1 (or at least its 3-spl specialization used for the algorithm dynamics) restricted to X(rx,r̄x), starting from definitions (6)–(8) and the rigorous bounds in [70,71], and independently recompute the r=6 entries of Tables 3–4 from the stationarity equations (46). If the derivation cannot be completed without an unproven 'trivial' extension, or if the recomputed f±,6 differ from 1.77842/0.32807 by more than 2×10⁻⁴, the central claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is Theorem 1 (§3), imported from [74,78] with the proof given as 'Follows immediately from [74,78] after trivial cosmetic changes in the definition of set X'. The set used here is X(rx,r̄x) = {x : ||x||_2 = rx, x_i^2 ≤ 1/n, (1/n)Σ log(1−nx_i^2) = r̄x} (Eq. 14), a nonconvex, entropy-constrained domain. Theorem 2 then converts the strong-duality equality (41) into the variational characterization ξ = f_chop(∞) = −ψ_rd(..., 49), and the 6-spl numbers in Tables 3–4 are compared against CLuP±Hop in Table 1. Nothing in the paper verifies that the 'complete sfl RDT frame' hypotheses hold for this particular X, nor that the stationarity system (46) selects the correct solution. If the imported strong-duality equality is not exact for this domain, the algorithm may simply be approaching a value predicted by a flawed duality formula; the agreement in Table 1 would then be internally consistent but not evidence about true ground state free energies. The absence of independent code and the self-cited nature of the theorem make the numerical theory column unverifiable as a standalone benchmark.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the problem of approximating ground-state free energies of positive and negative Hopfield models, i.e., maximizing x^T G^T G x or -x^T G^T G x over binary {±1/√n}^n vectors with Gaussian G. It introduces a CLuP±Hop gradient-descent algorithm, reports finite-n simulations (n up to 8000) reaching ξ≈1.77 for +Hop and ≈0.33 (magnitude) for -Hop, and compares these with thermodynamic-limit values ≈1.7784 and ≈0.3281 computed from the author's fully lifted random duality theory (fl RDT) at the sixth lifting level. The paper also studies overlap distributions, reports a qualitative difference between +Hop and -Hop near-optimal configurations, and concludes that these problems are 'typically easy.'","tokens_in":25303,"tokens_out":9628,"duration_ms":112469,"significance":"If the imported fl RDT characterization is correct and the simulations are reproducible, the paper gives a concrete algorithmic heuristic that appears to find near-ground-state configurations of random Hopfield models at moderate n, and the overlap cdfs in Figures 8-10 are testable predictions. The strengths are the breadth of the empirical study—finite-size convergence, concentration histograms, landscape evaluation, and comparison with SK—and the explicit falsifiable numbers in Tables 3-5. However, the main theoretical benchmark is not established in this paper: the strong-duality theorem is imported from the author's own preprints, and the reported agreement is between the algorithm and a self-generated prediction. The significance is therefore conditional on independent verification of the theory column.","major_comments":[{"comment":"The proof of Theorem 1 is one sentence: 'Follows immediately from [74, 78] after trivial cosmetic changes in the definition of set X.' The set X(rx,r̄x) in Eq. (14) is nonconvex and carries the entropy constraint (1/n)Σ log(1 - n x_i^2) = r̄x; no verification is given that the strong sfl RDT duality (33) holds for this domain, nor that the stationary point of system (46) selected by the numerics is the correct global extremum. Because Eq. (41) identifies f_chop(∞) with -ψ_rd and Table 1's theory column is computed from this identification, the agreement reported in Table 1 is currently an agreement between the algorithm and a self-generated prediction. This is load-bearing for the central claim that CLuP±Hop practically achieves the ground state free energies. I would need either a proof/check of the theorem's hypotheses for X(rx,r̄x) (with a precise reference to the result in [74,78] being invoked), or an independent benchmark—e.g., exact brute-force values for small n, rigorous bounds, or results from another group—against which both the algorithm and the 6-spl numbers are tested.","section":"§3, Theorem 1 (Eq. 33)"},{"comment":"The 6-spl numbers 1.77842 and 0.32807 are quoted to five significant digits, but the paper does not report the numerical procedure used to solve the stationarity system (46), any error bars, or a check that the solution is stable under changes of discretization or initialization. The finite-n gaps in Table 1 (e.g., +Hop: 1.7735 at n=8000 vs 1.7784; -Hop: 0.3330 vs 0.3281) are then impossible to interpret as purely finite-size effects, because the theory column itself has no stated precision. Providing code (or at least a detailed solver description plus a sensitivity analysis) and an independent validation of the 6-spl values would materially strengthen the paper.","section":"§3.2.1 and Tables 3–4"},{"comment":"The conclusion that ±Hop ground state problems are 'typically easy' is stronger than the evidence presented. The paper demonstrates empirically that a particular heuristic without restarts comes close to the claimed limits for n up to 8000, but it provides no runtime scaling analysis, no statement about typical-case polynomial time, and no overlap-gap-property style evidence; the 'no local optima' observation in Figure 7 is a numerical evaluation of the 3-spl RDT surrogate objective at selected t0x, not of the original optimization landscape. I recommend either supplying a formal typical-case statement or tempering the claim to 'empirically easy for the tested system sizes.'","section":"§4 and abstract"}],"minor_comments":[{"comment":"The spectral lower bound is written as '2√2/π' but then as '√(8/π)'; since 2√(2/π)=√(8/π) and 2√2/π≈0.900, one of these expressions is a typo. Please correct.","section":"§2.1"},{"comment":"The negative-model free energy f^-_sq(∞) is defined as a negative quantity in Eq. (8), but Table 1 and the abstract report '0.33' and '0.3281' as positive magnitudes. Please state the sign convention explicitly so that the negative-model free energy is not presented with the wrong sign.","section":"§2, Eqs. (7)-(8) vs Table 1"},{"comment":"The quantity ξhat is defined with an expectation E_G, but in the simulations it is an empirical average over algorithm outputs for fixed G; the notation should distinguish the random quantity from its expectation.","section":"§2, Eq. (12)"},{"comment":"The sentence 'Table 4 to +Hop model' should read 'Table 4 to -Hop model.'","section":"§3.2.2"},{"comment":"The term '3-spf RDT' should be '3-spl RDT.'","section":"Conclusion"},{"comment":"The expression 's max_{x∈Y, ∥y∥=...}' contains a typo; the maximization over Y should be over y.","section":"§3, Eq. (30)"}],"recommendation":"major_revision","confidential_remarks":"The central theoretical input is an imported theorem from the author's own preprints [74,78], and the numerical benchmark in Table 1 is generated by the same framework. This is not in itself disqualifying, but it makes independent verification particularly important. I recommend that the editor solicit a referee with expertise in random duality/interpolation methods and encourage the author to release code and small-n exact benchmarks. The 'typically easy' claim will attract attention and should be supported or qualified before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a real algorithmic result, not a toy. CLuP±Hop is a direct adaptation of the author's CLuP-SK, and it appears to find near-optimal configurations for both positive and negative Hopfield models at n in the thousands. The overlap finding—near-optimal +Hop configurations close to each other, −Hop configurations far apart—is new and genuinely interesting. The convergence and concentration plots are convincing as far as they go. Credit where earned: the paper ships a concrete algorithm, plain gradient descent with no restarts, and the simulations line up well with the theory's dynamics. The no-local-optima landscape observation is useful and empirically supported.\n\nThe soft spot is the benchmark. Table 1 compares CLuP±Hop's output to f values computed from the author's own fl RDT framework, via Theorem 1 imported from [78] with the proof given as 'Follows immediately.' That theorem converts a strong-duality equality into the variational formula used for the numbers, and Theorem 2 inherits it. Nothing in this paper checks that the 'complete sfl RDT frame' hypotheses hold for the entropy-constrained set X(rx,r̄x), nor that the stationarity system selects the correct solution. So the agreement in Table 1 is internal: algorithm versus self-generated prediction. That doesn't mean the numbers are wrong, but it means the central claim—'typically easy, free energies ≈1.7784/≈0.3281'—is only as solid as an unproven imported duality plus a 6-level numerical solution. There is no independent benchmark: no code, no small-n exact comparisons, no external rigorous bounds, no other group's results. The absence of code is minor on its own, but combined with the self-cited theorem it makes the theory column unverifiable as a standalone benchmark.\n\nI'd push back a little on the reader's conditional verdict being overly harsh. Within the author's program this is a coherent extension, and the algorithm's empirical performance is the kind of thing that should be reproducible. But the stress-test concern lands: the claimed 'ground state free energy' is not independently established. The paper would be much stronger with an external comparison or a proof sketch of Theorem 1 for this particular X.\n\nWho this is for: people working on spin-glass algorithms, Hopfield models, and random quadratic optimization. It deserves a serious referee, not a desk reject. A referee should ask for code/data and for verification of the imported theorem's hypotheses before the 'typically easy' conclusion is accepted.","headline":"A genuine algorithmic extension to Hopfield models, but the numbers it advertises are compared only against the author's own unproven fl RDT framework; worth refereeing with demands for independent verification.","tokens_in":25817,"tokens_out":2446,"would_cite":false,"duration_ms":30655,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["82B44","82D30"],"pacs":["75.10.Nr"],"model":"deepseek-v4-flash","headline":"This paper claims that the CLuP±Hop algorithm computes near-ground-state free energies of random positive and negative Hopfield models, reaching about 1.77 and 0.33 for n in the low thousands, against thermodynamic limits of roughly…","keywords":["Hopfield models","ground state free energy","CLuP algorithm","fully lifted random duality theory","Sherrington–Kirkpatrick model","semidefinite quadratic programming","binary optimization","overlap distributions"],"falsifier":"Run an independent, certified exhaustive search at $n=200$ (or branch-and-bound at $n=400$) on the same random $G$ used in the paper, and check whether the true +Hop maximum (resp. −Hop minimum) matches CLuP±Hop's output within the paper's reported finite-$n$ gap; a systematic shortfall would falsify the near-optimality claim.","tokens_in":24774,"feed_emoji":"🎯","tokens_out":11634,"duration_ms":120483,"temperature":0.7,"pith_summary":"The paper studies the ground-state free energy of random positive and negative Hopfield models, which are equivalent to maximizing $x^T A x$ over binary vectors $x \\in \\{\\pm 1/\\sqrt{n}\\}^n$ with $A = G^T G$ (positive) or $A = -G^T G$ (negative) for a standard-normal random matrix $G$. It introduces the Controlled Loosening-up algorithms CLuP+Hop and CLuP−Hop, which run gradient descent on a log-barrier objective while steadily increasing a loosening parameter $t_{0x}$, and reports that for $n$ as small as a few thousand they reach ground-state free energies of approximately $1.77$ and $0.33$. These values closely approach the thermodynamic limits $\\approx 1.7784$ and $\\approx 0.3281$ computed by fully lifted random duality theory at the sixth lifting level, and the simulations track the theory's predicted dynamics already at the third level. If the claim stands, computing near-ground-state energies of Hopfield models is typically easy even though the same quadratic-programming problem is worst-case hard, and the sixth-level analysis exposes a structural contrast: near-optimal configurations of the positive model are typically close to one another, while those of the negative model are typically nearly orthogonal.","feed_headline":"Hopfield ground states computed within 0.5% at n = 8,000","feed_subtitle":"The same CLuP-style procedure reaches ~1.77 and ~0.33, approaching thermodynamic limits ≈1.7784 and ≈0.3281.","key_machinery":"The load-bearing object is the CLuP±Hop objective, a log-barrier function $\\bar f^{\\pm}_{b,x}(x;\\bar t_{0x}) = -\\bar t_{0x}\\|x\\|^2 - \\log(-(x^T M^{\\pm} x - \\kappa)) - \\frac{1}{n}\\sum_i \\log(1-nx_i^2)$ with $M^{+} = 4I - \\frac{1}{n}G^TG$ and $M^{-} = \\frac{1}{n}G^TG$, whose gradient steps keep iterates inside the cube $\\{\\pm 1/\\sqrt{n}\\}^n$ while the growing $\\bar t_{0x}$ gradually loosens the auxiliary term so the procedure sharpens toward the ground state. The analytical engine is fully lifted random duality theory: Theorem 1 (imported from [78]) converts the bilinearly indexed random maximization into a deterministic lifted random dual $\\bar{\\psi}_{rd}$, and Theorem 2 identifies the ground-state value $\\xi(r_x,\\bar r_x)=f_{\\rm chop}(\\infty)$ with the negative of that dual, evaluated at the stationary point of $p,q,c,\\gamma,\\nu,\\gamma_{sq}$ for lifting level $r$. The lifting level controls accuracy: level 3 already reproduces the algorithm's whole trajectory, and level 6 yields the limiting free energies and the overlap cumulative distribution functions.","core_discovery":"The central claim is that the CLuP±Hop update $$$x^{{(t+1)}}$ \\gets \\operatorname{gradbar}\\big(\\bar $f^{{\\pm}}$_{b,x}(x; \\bar t_{0x}^{(t)}); $x^{{(t)}}$, \\bar t_{0x}^{(t)}\\big), \\qquad \\bar t_{0x}^{(t+1)} \\gets $c^{{(t)}}$ \\bar t_{0x}^{(t)},$$ with $c^{(t)}=1.1$, finds near-ground-state configurations of both Hopfield models: plain gradient descent on the barrier objective $\\bar f^{+}_{b,x} = -\\bar t_{0x}\\|x\\|^2 - \\log(-(x^T(4I - \\frac{1}{n}G^T G)x - \\kappa)) - \\frac{1}{n}\\sum_i \\log(1 - n x_i^2)$ (and the analogous $0I$ form for −Hop) reaches $\\hat\\xi \\approx 1.7704$–$1.7735$ for +Hop and $\\approx 0.3355$–$0.3330$ for −Hop at $n = 2000$–$8000$. The paper derives these dynamics from fully lifted random duality theory: Theorem 2 expresses the ground-state value $\\xi(r_x,\\bar r_x) = f_{\\rm chop}(\\infty)$ as the negative of a lifted random dual whose stationary conditions are solved at lifting levels $r=2,\\dots,6$, giving thermodynamic limits $f^{+,6}_{sq}(\\infty)\\approx 1.77842$ and $f^{-,6}_{sq}(\\infty)\\approx -0.32807$ (the −Hop value reported as a positive minimum). It then reads the overlap structure off the same sixth-level solution: the Gibbs-measure overlap distributions show +Hop near-optimal configurations typically close to each other and −Hop configurations typically almost orthogonal, with the Sherrington–Kirkpatrick overlap behavior resembling +Hop rather than −Hop.","pith_inferences":["Because the thermodynamic targets come from the paper's own fully lifted random duality theory, the numerical agreement is currently a self-consistency check; an independent derivation or proof of Theorem 1's limits would turn the close agreement into a genuine validation.","The +Hop/−Hop overlap dichotomy suggests a practical rule of thumb the paper does not systematically test: for −Hop, many well-separated restarts matter more than careful local refinement, while for +Hop a single good run suffices.","The observed no-local-optima landscape suggests CLuP-style annealing may transfer to other bilinearly indexed random-process models; testing it on $p$-spin or planted analogues at matched $n$ would reveal whether the favorable landscape is a generic feature.","The finite-$n$ gaps ($1.7784 - \\hat\\xi_+$, $0.3281 - \\hat\\xi_-$) appear to shrink steadily with $n$; fitting their decay rate would let practitioners predict the dimension needed for any target accuracy, an extrapolation the paper leaves implicit."],"forward_implications":["For $n=2000$–$8000$, restart-free CLuP±Hop reaches $\\hat\\xi \\approx 1.7704$–$1.7735$ (+Hop) and $\\approx 0.3355$–$0.3330$ (−Hop), approaching the thermodynamic limits $\\approx 1.7784$ and $\\approx 0.3281$ closely enough to call the near-ground-state problem typically easy.","The approximation factor for random ±Hop instances can be pushed toward 1, so the worst-case NP-hardness of indefinite quadratic programming does not manifest on typical instances.","Restarting and retuning the factor $c^{(t)}$ (e.g., $c^{(t)} = \\mathrm{Unif}[1,1.3]$) further cuts finite-$n$ error, improving −Hop at $n=500$ from $0.3430$ to $0.3358$, so plain descent is not the ceiling.","The sixth-level overlap distributions say that +Hop near-optimal configurations are typically close to each other and −Hop configurations are typically almost orthogonal; the Sherrington–Kirkpatrick model's overlap distribution resembles +Hop.","The same lifting progression gives $f^{(7)}_{csk}(\\infty) \\approx 0.76319$ for the SK model at the seventh level, supporting the existing predictions $\\approx 0.76321 \\pm 0.00003$ and $\\approx 0.76317$ for its true ground-state free energy."],"supporting_citations":[{"why":"Introduced CLuP-SK for the non-planted SK model; the algorithm and annealing schedule this paper adapts to ±Hop.","marker":"[82]"},{"why":"Supplies fully lifted random duality theory and the imported Theorem 1 that characterizes CLuP±Hop typical dynamics.","marker":"[78]"},{"why":"Prior fl RDT study of Hopfield models, giving the third-level ground-state limits and the overlap/lifting machinery reused here.","marker":"[80]"},{"why":"Stationarized fully lifted interpolation frame (complete sfl RDT) that Theorem 1 invokes for its proof.","marker":"[74]"},{"why":"Earlier bounds and iterative procedures for ±Hop ground-state energies, the baselines the new algorithm outperforms.","marker":"[70]"},{"why":"Bilinearly indexed random process comparisons used to connect the CLuP±Hop models to the fl RDT machinery.","marker":"[77]"},{"why":"IAMP result showing the SK ground-state energy is typically computable; motivation for expecting no computational gap.","marker":"[46]"},{"why":"Worst-case inapproximability of quadratic programming; the contrast that makes typical ease interesting.","marker":"[7]"}],"fun_headline_variants":["CLuP±Hop hits 1.77 and 0.33 for Hopfield ground states","Hopfield ground states within 0.5% via CLuP","CLuP makes Hopfield ground states typically easy","CLuP±Hop approaches thermodynamic limits for Hopfield","Near-exact Hopfield free energies with CLuP"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed agreement stands on an imported theorem that characterizes the thermodynamic limits; the theorem's proof is not given here but deferred to earlier work, so if that theorem or its numerical solution is inaccurate, the algorithm is being compared with a self-generated target.","fun_headline_variants_meta":{"raw":{"variants":["CLuP±Hop hits 1.77 and 0.33 for Hopfield ground states","Hopfield ground states within 0.5% via CLuP","CLuP makes Hopfield ground states typically easy","CLuP±Hop approaches thermodynamic limits for Hopfield","Near-exact Hopfield free energies with CLuP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000393,"raw_usage":{"total_tokens":2265,"prompt_tokens":1349,"completion_tokens":916,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":965,"completion_tokens_details":{"reasoning_tokens":824}},"tokens_in":965,"tokens_out":916,"duration_ms":9457,"temperature":1.0,"reasoning_tokens":824,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T11:43:17.511641+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run an independent, certified exhaustive search at $n=200$ (or branch-and-bound at $n=400$) on the same random $G$ used in the paper, and check whether the true +Hop maximum (resp. −Hop minimum) matches CLuP±Hop's output within the paper's reported finite-$n$ gap; a systematic shortfall would falsify the near-optimality claim.","supporting_citations":[{"cited_title":"A CLuP algorithm to practically achieve $\\sim 0.76$ SK--model ground state free energy","cited_arxiv_id":"2507.09247","evidence_quote":"Introduced CLuP-SK for the non-planted SK model; the algorithm and annealing schedule this paper adapts to ±Hop."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies fully lifted random duality theory and the imported Theorem 1 that characterizes CLuP±Hop typical dynamics."},{"cited_title":"Studying Hopfield models via fully lifted random duality theory","cited_arxiv_id":"2312.00071","evidence_quote":"Prior fl RDT study of Hopfield models, giving the third-level ground-state limits and the overlap/lifting machinery reused here."},{"cited_title":"Bounding ground state energy of Hopfield models","cited_arxiv_id":"1306.3764","evidence_quote":"Earlier bounds and iterative procedures for ±Hop ground-state energies, the baselines the new algorithm outperforms."},{"cited_title":"Montanari","cited_arxiv_id":null,"evidence_quote":"IAMP result showing the SK ground-state energy is typically computable; motivation for expecting no computational gap."}],"review_version":1}