{"id":"ed7afeb9-1e91-46a2-b735-4c6bcca10ea9","arxiv_id":"2411.11415","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The low-temperature ratio of the log-Sobolev constant to temperature converges to the Polyak-Lojasiewicz constant of the potential.","lead":"A new theorem shows that for functions with one minimum, the low-temperature limit of the log-Sobolev constant, which controls how fast sampling converges, exactly equals the Polyak-Lojasiewicz constant, which controls how fast gradient descent converges. This is an exact bridge between optimization and sampling theory.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: the main theorem is internally coherent, and the unique-minimizer limitation is necessary and explicitly acknowledged.","rationale":"The manuscript's central claim is supported by a constructive lower bound and a matching upper bound. I rechecked the small/large scalings in Theorem 12: with r0 = A sqrt(t), E0 and E1 vanish as t -> 0 for fixed A, E2 L0 t = O(1/A^2), and E2 t^2 = O(t^2 + t/A^2), so the defective-LSI constants have C_t = CPL t + o(t) plus O(t/A^2) and D_t = o(1) + O(1/A^2). Since Lemma 8 gives CP(mu_t) = O(t), Lemma 3 turns this into CLS(mu_t) <= CPL t + O(t^2 + t/A^2) + (o(1) + O(1/A^2)) O(t), yielding limsup <= CPL + O(1/A^2), then A -> infinity. The lower bound CPL <= liminf is a straightforward localization argument. Minor typographical issues (the denominator 2t/L1 instead of 2L1 t in Step 4; a nabla f / nabla g typo in a Step 3 display) do not change any estimate. The unique-minimizer assumption is necessary by the paper's own squared-distance counterexample and is honestly acknowledged; the open multiple-minimizer formula is outside the claimed theorem. Thus the ACCEPT verdict stands, subject only to the usual expectation that the cited improved tightening result [Wan24] is correct.","tokens_in":17775,"tokens_out":33676,"duration_ms":331227,"concrete_test":"Independently verify Lemma 3 by re-deriving it from the standard Rothaus lemma or by computing exact CLS and CP for a finite reversible Markov chain (e.g., a two- or three-point Laplacian) and checking CLS <= C + (D/2) CP for a (C,D)-defective LSI with C < CLS; if the inequality fails, recompute the Step 5 conclusion of Theorem 12 with the standard tightening to see whether the claimed exact constant survives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"After a line-by-line check of the lower and upper bounds, I did not find a load-bearing flaw. Theorem 11's lower bound is elementary and robust; Theorem 12's upper bound is delicate but coherent: the Gaussian-LSI comparison around the minimizer, the small/large scale split with r0 = A sqrt(t), the tail control via quadratic growth plus PL, and the final use of the improved tightening Lemma 3 all line up, with the stated o(1) and O(1/A^2) errors. The unique-minimizer assumption is the main structural restriction; the paper proves it is necessary (Section 1 counterexample) and explicitly leaves the multiple-minimizer case as an open question. The one dependency I would flag for verification is Lemma 3 ([Wan24, Prop. 5]), which is imported rather than proved: the exact constant in Theorem 1 relies on removing the extra CP term from the standard Rothaus tightening. This is a verification dependency, not an identified error; nothing in the present proof contradicts it.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proves that for a C^2 function f on R^d with a unique global minimizer and with Δf ≤ L(1+||∇f||^2), the low-temperature ('ballistic') limit C_bLS(f)=lim_{t→0+} C_LS(μ_t)/t of the log-Sobolev constant of μ_t ∝ e^{-f/t} equals the Polyak–Łojasiewicz constant C_PL(f) (Theorem 1). The lower bound (Theorem 11) is obtained by testing the LSI against smooth measures supported near arbitrary points; the upper bound (Theorem 12) uses a Gaussian LSI comparison near the minimizer, a small/large scale split with radius r0=A√t, tail control via quadratic growth and the PL inequality, and an improved Rothaus tightening (Lemma 3). The paper also proves an exact formula for the corresponding Poincaré constant, C_bP(f)=1/λ_min(∇^2f(x*)) (Theorem 2), and shows through a distance-to-convex-set example that the uniqueness assumption is necessary; the multiple-minimizer case is left as open problem (1.6).","tokens_in":17862,"tokens_out":14353,"duration_ms":135935,"significance":"If correct, these results give a clean and quantitatively sharp bridge between optimization and sampling: the normalized low-temperature log-Sobolev constant captures the global PL constant, while the Poincaré constant only sees the Hessian at the minimizer. The paper's proofs are largely self-contained, the main hypotheses are explicit, and the authors honestly delineate the necessity of the unique-minimizer assumption and the open general formula. The non-asymptotic bounds in Remarks 9 and 13 and in Proposition 14 are useful byproducts. The only external input that carries the exact constant is the improved tightening lemma [Wan24, Prop. 5]; if that lemma is correct as quoted, the main argument is coherent and the result is likely to become a standard reference.","major_comments":[],"minor_comments":[{"comment":"In the KL-decomposition displayed at the start of Step 1, the term '-1/t' inside the integral appears to be spurious, and the quadratic term should be written as (1/2)||x||^2_{Σ_t^{-1}} to match the density of N(0,Σ_t) under the notation fixed at the end of the introduction; the subsequent cancellation with the Gaussian LSI is correct once this notation is adjusted.","section":"Section 4, Step 1"},{"comment":"The denominator in the displayed bound for (1/t^2)∫||∇f||^2 e^{-g} should be 1 - 2L_1 t, not 1 - 2t/L_1; the stated condition t < 1/(2L_1) is consistent with the former, and the subsequent order estimates are unaffected.","section":"Section 4, Step 4"},{"comment":"Please add a precise pointer to [Wan24, Prop. 5] and, if space permits, a short proof sketch, because this lemma is the one imported ingredient that fixes the exact leading-order constant in Theorem 12.","section":"Section 2.1, Lemma 3"},{"comment":"There are several OCR/encoding artifacts in the text (for example, the inserted '/suppress' in Polyak–Łojasiewicz and the malformed '{ f /greaterorequalslant0 }'); these should be cleaned before the final version.","section":"Throughout"}],"recommendation":"minor_revision","confidential_remarks":"I concur with the reader's assessment. The paper is suitable for publication after minor typographical corrections; the main theorem is significant and the proof structure is coherent. The only point I could not verify from the manuscript is the external improved-tightening lemma [Wan24, Prop. 5], but this is a standard type of citation and not, on the present evidence, a reason to delay acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Chewi and Stromme prove that for f in C^2 with a unique global minimizer and Delta f <= L(1 + ||grad f||^2), the low-temperature limit of C_LS(mu_t)/t exists iff the Polyak-Lojasiewicz constant is finite, and equals C_PL(f). That is a genuinely new, exact bridge: the sampling constant that governs KL contraction for Langevin dynamics at low temperature is literally the optimization constant that governs gradient flow convergence. The paper also gives an exact formula for the ballistic Poincare constant as 1/lambda_min(Hess f(x*)).\n\nThe proofs are solid. The lower bound (Theorem 11) is an elementary testing argument: localize a test measure near any point and let t go to zero. The upper bound (Theorem 12) is more delicate: Gaussian LSI comparison near the minimizer, a small/large scale split with r = A sqrt(t), control of the tail via quadratic growth and PL, and an integration by parts step. I checked the main line and the error terms all behave as claimed. The key import is Lemma 3, an improved Rothaus tightening from Wang 2024, which avoids an extra factor of the Poincare constant. The paper does not prove that lemma, so the exact constant depends on an external result, but it is cited and the rest of the paper does not contradict it. I did not find a load-bearing flaw.\n\nThe unique minimizer assumption is necessary, not an artifact: for f equal to squared distance to a convex set with interior, C_PL is finite but the Gibbs measures converge to uniform on the set, so the ballistic LS constant is infinite. The multiple-minimizer case is left open, and the paper says so honestly. My only real complaints are minor: a typo in a displayed equation in Step 3 of Theorem 12 (missing an inner product term) and a loose use of 'exists' when the limit can be infinite. Neither affects the conclusions. The comparison with concurrent work by Chen and Sridharan is fair and factual.\n\nThis paper is for anyone working on Langevin sampling, low-temperature asymptotics, or the PL condition. It deserves a serious referee. I would send it to peer review.","headline":"Exact low-temperature bridge between log-Sobolev and PL constants, with a solid proof and an honest discussion of the necessary unique-minimizer assumption.","tokens_in":18453,"tokens_out":2225,"would_cite":true,"duration_ms":20513,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60J60","60E15"],"pacs":[],"model":"deepseek-v4-flash","headline":"The low-temperature limit of the log-Sobolev constant equals the Polyak–Łojasiewicz constant.","keywords":["ballistic log-Sobolev constant","Polyak-Lojasiewicz constant","low-temperature Gibbs measures","Langevin dynamics","log-Sobolev inequality","Poincare constant","gradient flow","low-temperature limit"],"falsifier":"Compute the ballistic log-Sobolev constant for $f(x)=\\frac{\\alpha}{2}\\,d(x,K)^2$ with a convex set $K$ of nonempty interior: $C_{\\mathsf{PL}}(f)$ is finite, yet $\\mu_t$ converges to the uniform measure on $K$, making $C_{\\mathsf{bLS}}(f)$ infinite and showing that removing the unique-minimizer assumption breaks the theorem.","tokens_in":17509,"feed_emoji":"📉","tokens_out":11728,"duration_ms":99190,"temperature":0.7,"pith_summary":"This paper proves an exact identity linking two previously separate rates: the speed at which the Langevin diffusion converges to the Gibbs measure proportional to $e^{-f/t}$ at low temperature, measured by the log-Sobolev constant $C_{\\mathsf{LS}}(\\mu_t)$, and the speed at which gradient flow converges for $f$, measured by the Polyak–Łojasiewicz constant $C_{\\mathsf{PL}}(f)$. Under the assumptions that $f$ is twice continuously differentiable, has a unique global minimizer, and satisfies $\\Delta f \\le L(1+\\|\\nabla f\\|^2)$, the paper shows that the limit $\\lim_{t\\to 0^+} C_{\\mathsf{LS}}(\\mu_t)/t$ exists exactly when $C_{\\mathsf{PL}}(f)$ is finite, and the two numbers are equal. The same argument yields the companion formula $\\lim_{t\\to 0^+} C_{\\mathsf{P}}(\\mu_t)/t = 1/\\lambda_{\\min}(\\nabla^2 f(x^\\star))$ for the Poincaré constant. A sympathetic reader would care because the analogy between sampling and optimization becomes a theorem with an exact constant, and the Polyak–Łojasiewicz constant acquires a new meaning as a low-temperature spectral quantity of Gibbs measures.","feed_headline":"Low-temperature sampling rate equals the PL constant","feed_subtitle":"For Gibbs measures exp(-f/t), the log-Sobolev constant divided by t converges to the Polyak-Lojasiewicz constant of f.","key_machinery":"The machinery views the log-Sobolev inequality as a Polyak–Łojasiewicz inequality for the Kullback–Leibler functional on the space of probability measures, then proves matching lower and upper bounds. The lower bound localizes: test measures concentrated near any point $x$ force the log-Sobolev inequality to imply the pointwise gradient inequality defining $C_{\\mathsf{PL}}(f)$. The upper bound compares $\\mu_t$ with a Gaussian of covariance $t[\\nabla^2 f(x^\\star)]^{-1}$, splits space into a small ball around the minimizer and its complement, and uses quadratic growth of PL functions to control the tail. The final constant is assembled by first proving a defective log-Sobolev inequality and then applying an improved tightening lemma that converts it into a full log-Sobolev inequality without losing a factor of two. The companion Poincaré result uses a Lyapunov-function argument to show $C_{\\mathsf{P}}(\\mu_t)=O(t)$, with the exact prefactor identified by a Gaussian rescaling near the minimizer.","core_discovery":"The central claim is an equality of constants. Define $C_{\\mathsf{PL}}(f)$ as the least $C$ such that $f(x)-f^\\star \\le \\frac{C}{2}\\|\\nabla f(x)\\|^2$ for all $x$; this is the constant that controls uniform exponential convergence of gradient flow. For $\\mu_t \\propto e^{-f/t}$, let $C_{\\mathsf{LS}}(\\mu_t)$ be the smallest constant such that $\\mathrm{KL}(\\nu\\|\\mu_t)\\le \\frac{C}{2}\\mathrm{FI}(\\nu\\|\\mu_t)$ for all smooth compactly supported $\\nu$. Theorem 1 states that when $f\\in C^2(\\mathbb{R}^d)$ has a unique global minimizer and $\\Delta f\\le L(1+\\|\\nabla f\\|^2)$, the ballistic log-Sobolev constant $C_{\\mathsf{bLS}}(f):=\\lim_{t\\to0^+} C_{\\mathsf{LS}}(\\mu_t)/t$ exists if and only if $C_{\\mathsf{PL}}(f)<\\infty$, and in that case $C_{\\mathsf{bLS}}(f)=C_{\\mathsf{PL}}(f)$. Theorem 2 states that $C_{\\mathsf{bP}}(f):=\\lim_{t\\to0^+} C_{\\mathsf{P}}(\\mu_t)/t$ equals $1/\\lambda_{\\min}(\\nabla^2 f(x^\\star))$. The authors also show the uniqueness assumption is not removable: for $f(x)=\\frac{\\alpha}{2}d(x,K)^2$ with a convex set $K$ of nonempty interior, $C_{\\mathsf{PL}}(f)$ is finite but the Gibbs measures converge to the uniform measure on $K$, so the ballistic log-Sobolev constant is infinite.","pith_inferences":["A testable extension is to use numerical estimates of the spectral gap of discretized Gibbs measures at small $t$ as an estimator of $C_{\\mathsf{PL}}(f)$; the paper does not address dimension dependence or discretization error for such an estimator.","The identity suggests a finite-temperature notion of PL constant, $C_{\\mathsf{PL},t}(f):=C_{\\mathsf{LS}}(\\mu_t)/t$, whose limit is the usual PL constant; studying its approach to the limit could inform annealing schedules, but this direction is not in the paper.","For functions whose minimizer set has positive dimension, the failure of the equality indicates that the ballistic limit should depend on the shape of the minimizer set rather than only on $C_{\\mathsf{PL}}(f)$; the paper does not propose a general formula for that case."],"forward_implications":["If $f$ satisfies the assumptions, then $C_{\\mathsf{LS}}(\\mu_t)=t\\,C_{\\mathsf{PL}}(f)+o(t)$, so the exponential rate of convergence of the Langevin dynamics in Kullback–Leibler divergence is dictated, in the low-temperature limit, by the same constant that governs gradient flow.","To leading order, the Poincaré constant of the same Gibbs measures is $C_{\\mathsf{P}}(\\mu_t)\\sim t/\\lambda_{\\min}(\\nabla^2 f(x^\\star))$; this ballistic limit sees only the local Hessian at the minimizer, not the global landscape.","The Polyak–Łojasiewicz constant can be read off from spectral data of Gibbs measures, giving a new characterization of the PL condition in terms of sampling.","The paper's non-asymptotic estimates imply $C_{\\mathsf{LS}}(\\mu_t)\\le C_{\\mathsf{PL}}(f)t + \\text{lower-order terms}$, an improvement by a factor of $t$ over known constant-order bounds for unique-minimizer PL landscapes.","When the global minimizer is not unique, the exact equality fails; the paper leaves the refined conjecture $C_{\\mathsf{LS}}(\\mu_t)-C_{\\mathsf{LS}}(\\mu_0)\\sim t\\,C_{\\mathsf{PL}}(f)$ as an open problem."],"supporting_citations":[{"why":"Supplies the definitions of log-Sobolev and Poincaré constants, the Gaussian log-Sobolev inequality, and the Lyapunov criterion for Poincaré inequalities.","marker":"[BGL14]"},{"why":"Introduces the Polyak condition and shows it implies linear convergence of gradient methods.","marker":"[Pol64]"},{"why":"Introduces the Lojasiewicz inequality in greater generality, one of the two historical sources of the PL condition.","marker":"[Lo63]"},{"why":"Provides the gradient-flow argument proving that PL functions satisfy quadratic growth, used to control tails and the Hessian at the minimizer.","marker":"[OV00]"},{"why":"Supplies the improved tightening lemma that turns the defective log-Sobolev inequality into a full one with the exact leading constant.","marker":"[Wan24]"},{"why":"Supplies the Wasserstein convergence tools used to extract the lower bound for the ballistic Poincaré constant.","marker":"[Vil03]"}],"fun_headline_variants":["Ballistic log-Sobolev limit equals PL constant","Log-Sobolev low-T limit matches Polyak-Łojasiewicz","Gibbs sampling: log-Sobolev limit yields PL constant","Exact equality: ballistic LS constant = PL constant","Low-temperature Langevin: log-Sobolev limit is PL constant"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the function has exactly one global minimum and that its curvature does not grow faster than a constant times $1+\\|\\nabla f\\|^2$; if the set of minima has any interior, the Gibbs measures spread out instead of concentrating, and the equality can fail.","fun_headline_variants_meta":{"raw":{"variants":["Ballistic log-Sobolev limit equals PL constant","Log-Sobolev low-T limit matches Polyak-Łojasiewicz","Gibbs sampling: log-Sobolev limit yields PL constant","Exact equality: ballistic LS constant = PL constant","Low-temperature Langevin: log-Sobolev limit is PL constant"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000733,"raw_usage":{"total_tokens":3358,"prompt_tokens":1102,"completion_tokens":2256,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":718,"completion_tokens_details":{"reasoning_tokens":2167}},"tokens_in":718,"tokens_out":2256,"duration_ms":18174,"temperature":1.0,"reasoning_tokens":2167,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:32:07.225579+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the ballistic log-Sobolev constant for $f(x)=\\frac{\\alpha}{2}\\,d(x,K)^2$ with a convex set $K$ of nonempty interior: $C_{\\mathsf{PL}}(f)$ is finite, yet $\\mu_t$ converges to the uniform measure on $K$, making $C_{\\mathsf{bLS}}(f)$ infinite and showing that removing the unique-minimizer assumption breaks the theorem.","supporting_citations":[],"review_version":1}