{"id":"b802a355-a6d2-4a82-bea5-7cbb441876d2","arxiv_id":"2607.18032","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"partial","parameter_count":0,"one_line_summary":"The zero-temperature Parisi measure for the SK spin glass is absolutely continuous with smooth density and support [0,1), and the positive-temperature endpoint q_β → 1 at rate Θ(β^−2).","lead":"This paper proves that in the Sherrington-Kirkpatrick model of disordered magnets at temperature zero, the Parisi order parameter is a smooth function with support covering every overlap value from 0 to 1. It also pins down how fast the positive-temperature endpoint tends to 1 as temperature drops.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Concern: the zero-temperature full-support result depends on the unverified sign bound D=u_xxx≤0 (Lemma 3.4), which drives H-monotonicity across jumps and hence the gap-exclusion theorem; the Lean formalization treats it as an axiom.","rationale":"The reader's weakest-assumption analysis matches my own reading: the most delicate and most load-bearing step in the zero-temperature argument is the sign control of the third spatial derivative of the Parisi PDE solution, which is used at every jump of γ to preserve monotonicity of Q_xx/Q (Prop. 3.7 Step 3). The formal verification is conditional on exactly this input, so the machine-checked part of the paper does not independently certify the key analytic estimate. I do not see a specific flaw in the proof of Lemma 3.4; the maximum-principle strategy is standard and the appendix gives the details. The concern is one of verification, not a demonstrated error. For that reason, the appropriate verdict remains CONDITIONAL: the paper should either machine-verify Lemma 3.4 or subject it to independent expert scrutiny. My proposed concrete test — extending the Lean formalization to prove (3.20) — is the single check most likely to settle whether the trust boundary is benign. I therefore agree with the reader's CONDITIONAL verdict and do not propose a change.","tokens_in":38035,"tokens_out":39048,"duration_ms":322341,"concrete_test":"In the existing Lean 4 repository (https://github.com/hbchen-math/FRSB_zero_temp_SK), replace the axiom that encodes Eq. (3.20) (D(t,x)=u_xxx(t,x)≤0 on [0,1)×(0,∞)) with a theorem proved from the already-formalized PDE, regularity lemmas, and the maximum-principle argument in Appendix B.3, and run the verifier. If the proof compiles without new axioms, the load-bearing premise is machine-verified; if it requires additional unproven statements, the paper's trust boundary is wider than the seven stated inputs and the CONDITIONAL verdict should be retained or strengthened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Theorem 1.3 rests on Proposition 4.2 (arbitrary-gap crossing), whose Step 3 requires H(t,B)=Q_xx/Q to be strictly increasing on any gap (a,b) where γ=m>0. That strict increase is proved in Proposition 3.7 (Step 5) via the heat-kernel total-positivity argument, but the entire monotonicity of H is built inductively on the sign bound D=u_xxx≤0 of Lemma 3.4, Eq. (3.20). At each jump of γ, the preservation identity (3.35) gives (H_+)_x = (H_-)_x + nonnegative terms − δ_t D(t,x); the conclusion (H_+)_x≥0 on (0,∞) uses −δ_t D≥0. If D≤0 failed at any jump, H could fail to be nondecreasing, the strict-increase input to (4.14) would be lost, and the residual construction (τ≤cz) in Step 4 would not repair the positivity of I'(t). The paper's own Lean 4 formalization explicitly treats Eq. (3.20) (with six other analytic inputs) as an assumption; the maximum-principle proof in Appendix B.3 is not machine-checked. This is an honest, clearly flagged trust boundary rather than a discovered inconsistency, but it is load-bearing: the full-support and smooth-density conclusions of Theorem 1.3 stand or fall on this analytic fact.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proves structural results for the Parisi measure of the zero-field Sherrington–Kirkpatrick model. At positive temperature, Theorem 1.2 gives quantitative endpoint control: the atom c_β at q_β satisfies c_β > 1/3 and 1−q_β lies between (3−c_β)/(2β^2) and 2/β^2, so q_β→1 as β→∞. At zero temperature, Theorem 1.3 asserts that the unique minimizer γ_⋆ of the Parisi functional is absolutely continuous on [0,1) with a smooth density and that its Stieltjes measure has full support [0,1). The proof combines variational consistency identities, finite-cascade Cole–Hopf inequalities, a zero-temperature crossing argument excluding internal and terminal support gaps, and a regularity bootstrapping argument. The paper also includes a conditional Lean 4 formalization of Theorems 1.2–1.3 that treats seven analytic inputs as assumptions, one of which is the third-derivative sign bound used in the monotonicity argument.","tokens_in":38243,"tokens_out":5795,"duration_ms":67132,"significance":"If the analytic hypotheses are fully verified, the result is significant: it identifies the zero-temperature Parisi measure for the SK model as a smooth, atomless measure filling the entire interval [0,1), thereby proving the no-overlap-gap condition used in algorithmic and ultrametric interpretations. The paper is also praiseworthy for its explicit disclosure of the formalization's trust boundary, for the absence of fitted or calibrated parameters, and for the transparent chain of propositions leading to the gap-exclusion theorem. The main caveat is that the central zero-temperature conclusion depends on a sign bound for u_xxx that is not machine-checked and whose proof in Appendix B.3 is compressed.","major_comments":[{"comment":"The assertion D(t,x)=u_xxx(t,x)≤0 on [0,1)×(0,∞) is load-bearing. It is used in Step 3 of Proposition 3.7 through Eq. (3.35) to preserve monotonicity of H=Q_xx/Q across jumps of γ, and the strict monotonicity of H is then an input to the crossing identity (4.14) in Proposition 4.2. If D≤0 failed at a jump, the preservation identity (3.35) would not imply (H_+)_x≥0, and the implication Γ''(t)=0 ⇒ Γ'''(t)>0 would no longer follow; internal or terminal gaps could survive. The proof in Appendix B.3 invokes a maximum principle 'in the form used in [10, proof of Lemma 8]' without stating the growth and boundary hypotheses precisely, and the accompanying Lean 4 repository explicitly treats Eq. (3.20) as one of seven assumptions rather than a machine-checked theorem. This is a fixable but load-bearing gap: the manuscript should either supply a complete self-contained proof of D≤0 or extend the f","section":"Lemma 3.4, Eq. (3.20); Props. 3.7 and 4.2"},{"comment":"The paper's own footnote states that the Lean formalization 'treats seven explicitly identified analytic inputs as assumptions' and is 'not an assumption-free verification of the entire paper.' This is honest, but the main text later says the manuscript is self-contained apart from Theorem 1.1 and well-established results. The reader should be told explicitly, in the introduction or at the statement of Theorem 1.3, that the zero-temperature conclusion is conditional on the unformalized analytic inputs, especially the sign bound in Lemma 3.4. This is not a mathematical inconsistency, but the presentation currently understates the verification debt.","section":"Sec. 3.1-3.2 and formalization footnote"}],"minor_comments":[{"comment":"The notation u_λ := u_{λ,γ∧λ} = u_{h_λ,γ∧λ} is confusing: the first subscript appears to be a parameter but is then identified with a terminal datum. Define the two-argument notation explicitly before Eq. (3.19).","section":"Eq. (3.19) and surrounding text"},{"comment":"The maximum-principle step for the third derivative cites [10, proof of Lemma 8] without stating the exact function class, growth conditions, or treatment of the unbounded half-line. Since this is the main unverified step, the proof should be expanded.","section":"Appendix B.3"},{"comment":"There are several typographical and formatting issues: the title contains broken spacing ('INTER V AL', 'SUPPOR T'), and Appendix C contains 'aiXiv' instead of 'arXiv'. These do not affect the mathematics but should be corrected.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically interesting and the argument is coherent conditional on the stated analytic inputs. My main reservation is the unverified third-derivative sign bound, which is not merely a technical aside: it drives the monotonicity of H and hence the gap-exclusion theorem. If the author can provide a complete proof of Lemma 3.4 or machine-check it, I would support acceptance; at present the verification gap is too large for the zero-temperature claim to be considered established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Chen's paper is the first to give a complete zero-temperature structural description of the SK Parisi measure: smooth density, full support [0,1), plus quantitative β^{-2} endpoint bounds at positive temperature. That is a real step beyond Auffinger–Chen–Zeng's infinitely-many-points result and Lopatto's positive-temperature interval theorem. The endpoint argument in Section 2 is genuinely clever — the moment identities plus the monotone transformed density give clean bounds on c_β and q_β without fitting anything.\n\nThe zero-temperature machinery is also well organized. The gap-exclusion argument via the crossing theorem is intricate but laid out honestly: each step names its inputs, and the Cole–Hopf inequalities are reproved in an appendix so the paper is mostly self-contained. The biggest soft spot is exactly what the stress test flags: the strict monotonicity of H=Q_xx/Q and the gap crossing rest on the sign bound D=u_xxx≤0 (Lemma 3.4, Eq. (3.20)). The maximum-principle proof in Appendix B.3 is plausible but not machine-checked, and the Lean formalization treats it as one of seven axioms. If D≤0 fails at some jump, the monotonicity preservation (3.35) breaks and the whole gap-exclusion argument collapses. That is not a discovered contradiction — it is a clearly flagged trust boundary — but it is load-bearing.\n\nI also note the paper leans on Lopatto's Theorem 1.1 without reproof. That is acceptable given the provenance disclosure, but it does mean the positive-temperature characterization is an external input. The authorship note is unusual and refreshingly direct.\n\nWho gets value from this: anyone working on mean-field spin glasses or the Parisi structure, and people tracking the no-overlap-gap assumptions behind optimization algorithms. It deserves a serious referee — the potential payoff is high and the risk is concentrated in one identifiable analytic lemma. I would send it out, with instructions to scrutinize Lemma 3.4 and the seven Lean assumptions. If those hold, the paper is a major result; if not, the zero-temperature theorem is unproven. Conditional accept is the right posture.","headline":"If Lemma 3.4's sign bound holds, this settles zero-temperature FRSB for SK; the proof is transparent and the trust boundaries are honest, but the load-bearing third-derivative assumption needs independent verification before the result is accepted.","tokens_in":38891,"tokens_out":901,"would_cite":true,"duration_ms":13138,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60K35","82B44"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that the zero-temperature Parisi measure of the SK spin glass is smooth with support [0,1), so all overlap values are realized by near-ground states; it also quantifies how the positive-temperature support endpoint approach","keywords":["Sherrington-Kirkpatrick model","Parisi measure","full replica symmetry breaking","zero temperature","overlap gap","variational problem","absolute continuity","support"],"falsifier":"Run a high-accuracy numerical solve of the zero-temperature Parisi PDE for a simple admissible two-step γ, such as γ=0 on [0,t0) and γ=m on [t0,1), and check the sign of u_xxx at time t0 for positive x; any positive value would violate Lemma 3.4, the monotonicity of H across the jump, and therefore the gap-exclusion proof. Alternatively, producing an admissible γ with a strict support gap that still satisfies the first-order optimality conditions would directly falsify Theorem 1.3.","tokens_in":37767,"feed_emoji":"❄️","tokens_out":7313,"duration_ms":74396,"temperature":0.7,"pith_summary":"The paper's goal is to pin down the exact structure of the Parisi variational object for the Sherrington–Kirkpatrick spin glass at zero temperature: it claims that the unique minimizer is a smooth, atomless, strictly increasing function whose Stieltjes measure has support the whole half-open interval [0,1). If true, this means that at zero temperature every overlap value in [0,1) is approximated by pairs of near-ground-state configurations, so the model realizes infinite-step replica symmetry breaking with no overlap gaps. The paper also sharpens the positive-temperature picture by showing that the support endpoint qβ approaches 1 at a rate no faster than a constant times β^{-2} and no slower than (3−cβ)/(2β²), with the endpoint atom weight cβ > 1/3. A sympathetic reader should care because the shape of this measure is the order parameter of the model: it controls the ultrametric structure of ground states and is exactly the input that optimization algorithms based on message passing assume to be strictly increasing.","feed_headline":"At zero temperature, spin-glass overlaps cover the whole interval","feed_subtitle":"The Parisi measure is smooth and gapless, validating an algorithmic assumption at zero temperature.","key_machinery":"The carrying object is the zero-temperature Parisi minimizer γ⋆ and its Stieltjes measure ν⋆ = dγ⋆. The Parisi PDE, ∂_t u + ½(u_xx + γ(t) u_x²) = 0 with terminal condition u(1,x)=|x|, turns a choice of γ into a convex function u, and the functional P(γ)=u(0,0)−½∫₀¹ t γ(t)dt selects γ⋆. The technical engine is the transformed density Q(t,x)=ρ_t(x)e^{−γ(t)u(t,x)} and the ratio H(t,x)=Q_xx/Q. The proof shows H(t,·) is nondecreasing and strictly increasing inside every gap, using the third-derivative sign bound u_xxx ≤ 0 and strict total positivity of the heat kernel; this monotonicity, combined with Cole–Hopf inequalities for the slope-space curvature quantities K and J, yields the crossing imp","core_discovery":"On its own terms, the paper establishes two structural theorems. Theorem 1.3 states that for the zero-temperature variational problem with terminal datum |x|, the unique minimizer γ⋆ satisfies γ⋆(0)=0 and its Stieltjes measure dγ⋆ equals ρ∞(t)dt for a nonnegative smooth function ρ∞ ∈ C^∞([0,1)), with support [0,1). Hence the Parisi measure has no atom and no singular continuous part, and every overlap in [0,1) lies in the support. Theorem 1.2 states that at every β > 1 the positive-temperature Parisi measure has an endpoint atom of weight cβ and support [0,qβ], with cβ > 1/3 and (3−cβ)/(2β²) < 1−qβ < 2/β², so qβ → 1 as β → ∞. The proof takes the known positive-temperature interval-and-atom s","pith_inferences":["The third-derivative sign condition u_xxx ≤ 0 is doing more work than any other indivisible estimate: it is the step that keeps H monotone across jumps. A natural testable extension is whether the same condition holds for mixed even p-spin models; if it fails, the full-support theorem may fail there even when the Parisi formula is valid.","Because the no-overlap-gap assumption now holds at zero temperature, the algorithmic question shifts from whether the assumption holds to whether the constant C(ε) in the running time C(ε)N² can be made polynomial in 1/ε; the present proof does not address that rate.","The endpoint estimates suggest a boundary-layer picture: as β grows, the positive-temperature density on [0,qβ] plus an atom at qβ reshapes, after scaling by β, into a smooth full-interval Stieltjes measure. One could try to recover ρ∞ as the limit of a free-boundary problem in which the endpoint atom melts into the continuum."],"forward_implications":["Every overlap u in [0,1) is realized, up to arbitrarily small error, by two near-ground-state spin configurations with probability exponentially close to 1; the support statement plus the known overlap-to-ground-state correspondence gives this directly.","The zero-temperature no-overlap-gap assumption used by El Alaoui, Montanari, and Sellke holds for the SK model: because γ⋆ is strictly increasing, their message-passing algorithm reaches energy within any fixed ε of the optimum.","Since γ⋆ is smooth and strictly increasing, the zero-temperature Parisi measure is neither atomic nor singular continuous; it represents infinite-step replica symmetry breaking.","The positive-temperature support endpoint qβ converges to 1 at rate β^{-2} (up to constants), while the endpoint atom weight stays bounded below by 1/3, so the endpoint layer is a quantitative boundary effect.","The zero-temperature variational problem cannot see an atom at the right endpoint t=1, so any closed-interval convention with a mass at 1 requires an external limiting prescription; the positive-temperature boundary layer is one such prescription."],"fun_headline_variants":["Zero temp spin glass: overlaps span entire interval","Full RSB at zero temp: Parisi measure smooth and gapless","At T=0, spin-glass overlaps fill [0,1) completely","Zero-temperature SK: Parisi density supported on [0,1)"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is the strict monotonicity of H=Q_xx/Q on (0,∞), which the proof obtains from the third-derivative sign bound u_xxx ≤ 0 on [0,1)×(0,∞); the paper itself notes that its machine-checked formalization treats this analytic sign bound as an assumption, so if that sign ever fails the crossing implication inside gaps collapses.","fun_headline_variants_meta":{"raw":{"variants":["Zero temp spin glass: overlaps span entire interval","Full RSB at zero temp: Parisi measure smooth and gapless","At T=0, spin-glass overlaps fill [0,1) completely","Zero-temperature SK: Parisi density supported on [0,1)"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000203,"raw_usage":{"total_tokens":1183,"prompt_tokens":665,"completion_tokens":518,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":409,"completion_tokens_details":{"reasoning_tokens":444}},"tokens_in":409,"tokens_out":518,"duration_ms":5162,"temperature":1.0,"reasoning_tokens":444,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T16:17:15.730039+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a high-accuracy numerical solve of the zero-temperature Parisi PDE for a simple admissible two-step γ, such as γ=0 on [0,t0) and γ=m on [t0,1), and check the sign of u_xxx at time t0 for positive x; any positive value would violate Lemma 3.4, the monotonicity of H across the jump, and therefore the gap-exclusion proof. Alternatively, producing an admissible γ with a strict support gap that still satisfies the first-order optimality conditions would directly falsify Theorem 1.3.","supporting_citations":[],"review_version":1}