{"id":"fa32e137-a4f9-4741-9c4f-ed65ec4fb2b1","arxiv_id":"2509.15410","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A unified Φ-Sobolev proof transfers Poincaré and log-Sobolev inequalities from mixture components to their mixture or joint distribution, with explicit constants and Markov chain Monte Carlo applications.","lead":"This paper gives general conditions under which a mixture or joint of distributions inherits a Poincaré or log-Sobolev inequality from its components. It obtains these conditions through a unified Φ-Sobolev framework and applies them to Langevin, proximal, and Hamiltonian Monte Carlo samplers.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Uniform score condition is the load-bearing premise; the two-scale proof is sound, but exact-HMC applicability and the abstract's MCMC scope are not established.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing premise: the uniform two-scale score condition (Var)/(MGF). My independent reading of the proof of Theorem 3.2 confirms that the cross term T'_3 is controlled only through pointwise use of this condition; no averaged or relaxed substitute appears in the argument. The theorem itself is not contradicted: for the t^2 and t log t instantiations, the decomposition, the use of Proposition 5.4, and the constants xi = beta + alpha*beta*Lbar^2 all check out. The concern is rather about scope: the paper's own applications verify the score condition in the Gaussian or LSI-plus-Lipschitz cases, but the exact HMC application explicitly leaves Var unverified outside the Gaussian case, and the promised appendix treatment of a bounded-variation alternative is absent. These are support gaps rather than mathematical contradictions, so the appropriate disposition remains conditional acceptance. The concrete test would determine whether the exact-HMC gap is real by checking whether the uniform score criterion actually holds for a simple non-Gaussian target.","tokens_in":20128,"tokens_out":34405,"duration_ms":294072,"concrete_test":"Implement exact HMC for a one-dimensional non-Gaussian strongly convex target, e.g. V(x) = mu*x^2/2 + epsilon*cos(x) with epsilon > 0 small and mu > 0, and estimate S = sup_{y, u: ||u||=1} E_{x|y}[<u, nabla_y log p_{X|Y=y}(x)>^2] by simulating Hamiltonian flow for many y. Compare S against the Gaussian epsilon = 0 value, which is finite and proportional to c_ULA^2 / eta. If S is finite, Section 4.1.3's use of Theorem 3.2 can go through; if S is unbounded or grows with epsilon, the uniform score criterion is the obstruction and the exact-HMC claim needs an additional regularity assumption.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The transfer theorem is internally coherent; I could not identify an algebraic error in the proof of Theorem 3.2. The load-bearing premise is the uniform two-scale score condition (Var)/(MGF): the bound on T'_3 in Section 5.2 applies the score inequality pointwise in y and then integrates, so if the score bound fails on a set of y of positive rho-measure, the cross term has no control and the claimed constant xi has no basis. This is a limitation of the hypotheses, not a contradiction. The ULA application and the proximal forward step satisfy the criterion by Gaussianity; the proximal backward step satisfies it via LSI plus a Lipschitz condition on nabla_2 G (Section 3.1, case 3). The exact HMC application, however, is exactly where the criterion is left unverified: Section 4.1.3 states that Var 'is hard to check in general' and only computes the Gaussian case. Section 3.1 also promises an appendix treatment of a beta-bounded variation alternative that the appendix does not provide. The abstract's 'variety of Markov chains' therefore overstates the verified scope, even though the main theorem is sound.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops two-scale sufficient conditions under which a joint distribution ν(x,y)=P_{X|Y=y}(x)ρ(y) or its mixture μ(x)=∫P_{X|Y=y}(x)dρ(y) satisfies a Poincaré or log-Sobolev inequality, given that the mixing measure ρ and each conditional P_{X|Y=y} satisfy the relevant inequality. The conditions are a variance bound (Var) on the conditional score for the Poincaré case and a sub-Gaussian moment bound (MGF) on the same score for the log-Sobolev case. The proofs are carried out in the framework of Φ-Sobolev inequalities, using a duality inequality for Φ-entropies and Young/Cauchy-Schwarz estimates. The paper then applies the two-scale criteria to the unadjusted Langevin algorithm, the proximal sampler, and exact Hamiltonian Monte Carlo, deriving explicit recursions for the isoperimetric constants and, in some cases, their biased or unbiased limits.","tokens_in":20474,"tokens_out":14749,"duration_ms":114812,"significance":"If the results hold, the paper provides a unified and elementary proof of previously known two-scale transfer results (Otto-Reznikoff, Mou et al., Ge et al.) and adds a new sub-Gaussian criterion for log-Sobolev transfer. The constants in Theorems 3.1 and 3.2 are explicit and parameter-free, and the recursions for ULA and the proximal sampler reproduce known rates with a concise argument. The proof of Proposition 5.4 and its instantiations are carefully written and appear algebraically sound. However, the advertised scope is broader than what is actually verified: the exact HMC application is only carried out in the Gaussian case, and one auxiliary sufficient condition is promised to be deferred to an appendix that does not contain it. These issues do not invalidate the central transfer theorem, but they limit the paper's claims of covering a 'variety of Markov chains'.","major_comments":[{"comment":"The manuscript states that Var holds under a β-bounded variation condition and says 'We include details about this in the appendix.' The appendix (Section A) contains only Definitions A.1 and A.2 and Hoeffding's lemma; it does not discuss any β-bounded variation condition. This explicit promise to the reader is unfulfilled and should be either implemented or removed.","section":"Section 3.1, Var case 3"},{"comment":"The abstract claims that the proposed criteria 'are satisfied by a variety of Markov chains, and consequently allows us to characterise the evolution of these functional inequalities for iterates generated by simulating these Markov chains.' The exact HMC application in Section 4.1.3, however, states that Var 'is hard to check in general' and only verifies the Gaussian case. The abstract therefore overstates the verified scope; it should be qualified to the ULA and proximal sampler applications, which are fully verified under their stated conditions.","section":"Section 4.1.3 / Abstract"},{"comment":"The claim that σ-sub-Gaussianity in the sense of Definition A.2 implies Var with Lbar = 2σ is not correct. From the defining inequality E exp(λ⟨u,Z−EZ⟩) ≤ exp(λ²σ²/2), differentiating twice at λ=0 gives E[⟨u,Z−EZ⟩²] ≤ σ², so the correct implication is Lbar = σ. The factor 2σ should be corrected.","section":"Section 3.1, MGF case 2"},{"comment":"The proofs of Theorems 3.1 and 3.2 additionally assume that Φ is of Legendre type, that 1/Φ'' is concave, and that the range of Φ' is R, whereas the theorem statements only say 'Let Φ : S → R be a twice differentiable convex function' and 'specified later.' Although the two instantiated choices satisfy the extra conditions, the statements should either include the extra hypotheses or explicitly note that the proofs are only carried out for the two displayed choices of Φ. The phrase 'specified later' is confusing.","section":"Theorems 3.1-3.2 / Section 5"}],"minor_comments":[{"comment":"There are numerous typographical errors, including 'phenomemon' in the Introduction, 'time-homoegeneous' in Section 4.1.3, 'A clean error analysis' in the Abstract, and 'Principled' where 'principal' is likely intended. The manuscript would benefit from a careful proofreading pass.","section":"Throughout"},{"comment":"The first sentence of the proof of Theorem 3.2 says 'We would like to show that J^Φ_ν[ψ] ≤ ...', but the theorem concerns the mixture measure μ; it should read J^Φ_μ[ψ].","section":"Section 5.2"},{"comment":"After the proximal sampler recursion, the convergence condition is stated as 'β < η^3' in the text, but the preceding line and the subsequent analysis use the condition β/η < 1, i.e., β < η. The appearance of η^3 appears to be a typographical error.","section":"Section 4.1.2"}],"recommendation":"minor_revision","confidential_remarks":"The central transfer theorem is sound and the Φ-entropy presentation is elegant, but the paper's advertised applications are broader than what is verified. The missing appendix and the unverified exact HMC claim should be addressed before publication. The novelty is incremental but the unification is clean; no concerns about citation practices or originality beyond the scope overstatement."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is a genuinely useful paper, with a clear gap between what it proves and what it advertises.\n\nThe new thing is the unified Phi-Sobolev proof of two-scale PI/LSI transfer for Euclidean mixtures. Theorem 3.2 is the load-bearing result: if the mixing measure and each conditional satisfy PI/LSI, and the conditional score satisfies a uniform variance or MGF bound, the mixture satisfies the same inequality with explicit constant beta + alpha*beta*Lbar^2. The proof via Phi-duality is elementary and clean; it recovers the constants of Otto-Reznikoff and Mou/Ge without Markov semigroup machinery. The ULA and proximal sampler recursions are worked out in detail, and the Gaussian HMC analysis gives a closed-form recursion. That is a real contribution, and I checked the algebra in the main proof; it holds up.\n\nNow the soft spots. First, the theorem statements are under-specified: they say Phi is twice differentiable convex and 'specified later,' but the proofs require Legendre type, 1/Phi'' concave, and range of Phi' equal to R. A reader can infer this, but the statements as written are formally weaker than the proofs. Fixable, but it matters.\n\nSecond, the abstract says the criteria hold for 'a variety of Markov chains.' The verified applications are ULA, the proximal sampler, and HMC for a Gaussian target. The exact HMC section explicitly admits that the Var condition is 'hard to check in general' and only computes the Gaussian case. So the scope is narrower than claimed. That is an honest limitation, not a contradiction.\n\nThird, Section 3.1 promises an appendix treatment of a beta-bounded variation alternative to the Lipschitz condition; the appendix does not contain it. Minor, but it signals the draft is unfinished.\n\nThe load-bearing premise is the uniform score bound (Var/MGF). If it fails on a set of y of positive rho-measure, the cross term in T'_3 is uncontrolled. This is a real restriction; the applications verify it under Gaussianity or Lipschitz/strong-convexity. The paper does not hide it, but users should not expect the transfer to hold without it.\n\nNet: this is a competent, honest paper with a useful unification and explicit constants. It needs revision to align statements with hypotheses, trim the abstract, and either provide or remove the appendix promise. It deserves a serious referee.","headline":"A genuinely useful two-scale PI/LSI transfer with explicit constants; the core proof is sound, but the exact-HMC scope is left unverified and the theorem hypotheses are stated looser than the proofs require.","tokens_in":20885,"tokens_out":2141,"would_cite":true,"duration_ms":19208,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60E15","60J22","26D10"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves explicit two-scale criteria under which mixtures and joint distributions inherit Poincaré and log-Sobolev inequalities, and shows how the constants propagate through MCMC algorithms.","keywords":["Poincaré inequality","log-Sobolev inequality","Φ-Sobolev inequality","two-scale criteria","mixture distributions","Markov chain Monte Carlo","unadjusted Langevin algorithm","proximal sampler"],"falsifier":"Take scalar Gaussians $\\rho=\\mathcal{N}(0,\\sigma_1^2)$ and $P_{X|Y=y}=\\mathcal{N}(a y,\\sigma_2^2)$. Then $\\alpha=\\sigma_1^2$, $\\beta=\\sigma_2^2$, $\\bar L^2=a^2/\\sigma_2^2$, and the mixture is $\\mathcal{N}(0,a^2\\sigma_1^2+\\sigma_2^2)$, whose exact LSI constant is $a^2\\sigma_1^2+\\sigma_2^2=\\xi$. Checking this identity across parameter ranges tests the constants; any parameter regime where the optimal mixture constant exceeds $\\xi$ while the assumptions hold would falsify the theorem.","tokens_in":19914,"feed_emoji":"🎲","tokens_out":14378,"duration_ms":114160,"temperature":0.7,"pith_summary":"This paper establishes two-scale criteria under which a mixture or joint distribution inherits a Poincaré or log-Sobolev inequality from its constituent parts. If the mixing measure satisfies the inequality with constant $\\alpha$, every conditional distribution satisfies it with constant $\\beta$, and the conditional score is uniformly controlled in the sense of a variance or moment-generating condition with constant $\\bar L$, then the mixture satisfies the same class of inequality with explicit constant $\\beta+\\alpha\\beta\\bar L^2$, and the joint distribution satisfies an explicit but more intricate constant. The proof is unified through $\\Phi$-Sobolev inequalities, so the variance and entropy cases come from one mechanism rather than separate arguments. The interest for MCMC is that the constants propagate through recursions for the unadjusted Langevin algorithm, the proximal sampler, and exact Hamiltonian Monte Carlo, yielding explicit rates at which the isoperimetric quality of iterates converges.","feed_headline":"One score bound transfers mixing constants to mixtures","feed_subtitle":"New explicit two-scale criteria unify prior results and drive MCMC iterate recursions.","key_machinery":"The machinery is the $\\Phi$-entropy $J_\\pi^\\Phi[f]=\\mathbb{E}_\\pi[\\Phi(f)]-\\Phi(\\mathbb{E}_\\pi[f])$ and the associated $\\Phi$-Sobolev inequality $J_\\pi^\\Phi[f]\\le \\frac{\\gamma}{2}\\mathbb{E}_\\pi[\\Phi''(f)\\Vert\\nabla f\\Vert^2]$. The proof decomposes the $\\Phi$-entropy of a joint or mixture into a conditional term and a marginal term, then bounds the marginal term by estimating the gradient of the conditional expectation with the identity $\\nabla_y\\log p_{X|Y=y}(x)=\\mathbb{E}_{X|Y=y}[\\nabla_2G(x,y)]-\\nabla_2G(x,y)$ and a variational duality result for $\\Phi$-entropies. The Var and MGF conditions are exactly the two estimates needed to control the cross term that couples the conditional score with the test function; choosing $\\Phi(t)=t^2$ or $\\Phi(t)=t\\log t$ makes the general argument specialize to Poincaré or log-Sobolev.","core_discovery":"The paper's central claim is Theorem 3.2: for a mixture $\\mu$ with mixing measure $\\rho$ and conditional distributions $P_{X|Y=y}$, if $\\rho$ satisfies a $\\Phi$-Sobolev inequality with constant $\\alpha$, every $P_{X|Y=y}$ satisfies the same $\\Phi$-Sobolev inequality with constant $\\beta$, and the conditional score $\\nabla_y\\log p_{X|Y=y}(x)$ satisfies either the variance bound $\\mathbb{E}_{X|Y=y}[\\langle u,\\nabla_y\\log p_{X|Y=y}(X)\\rangle^2]\\le \\bar L^2\\|u\\|^2$ or its moment-generating analogue, then $\\mu$ satisfies the corresponding Poincaré or log-Sobolev inequality with constant $\\xi=\\beta+\\alpha\\beta\\bar L^2$. The joint-distribution version (Theorem 3.1) carries the same information with the more complicated constant $\\zeta$ obtained by optimizing an auxiliary parameter. Choosing $\\Phi(t)=t^2$ gives the Poincaré case and $\\Phi(t)=t\\log t$ gives the log-Sobolev case, so both results are corollaries of one $\\Phi$-Sobolev argument. In the applications the score conditions reduce to bounded-score, bounded-score-variance, or Lipschitz-Hessian assumptions, which the paper verifies for the Gaussian conditionals arising in ULA, the proximal sampler, and exact HMC.","pith_inferences":["Averaged, rather than uniform, control of the score would likely suffice: the proof only needs the cross term controlled in expectation over $\\rho$, so a mean-version of Var or MGF might replace the uniform condition at the cost of a different constant.","The linear-Gaussian calculation suggests the mixture bound $\\xi=\\beta+\\alpha\\beta\\bar L^2$ is sharp in the linear case; a natural extension is to ask whether the same transfer constant is optimal for the joint theorem and for non-Gaussian conditionals near the Gaussian regime.","A practical diagnostic for MCMC use is to track the pointwise Fisher information $\\mathbb{E}_{X|Y=y}\\Vert\\nabla_y\\log p_{X|Y=y}(X)\\Vert^2$ along a chain; the theory predicts the usable PI/LSI constant degrades with its supremum, so when that supremum grows the theorem's guarantee becomes vacuous even if each conditional is individually well behaved.","The same $\\Phi$-entropy decomposition might transfer two-scale criteria to other inequalities in the Beckner family, but the direct duality route fails there because both duality terms are nonzero; a new estimate for the second term would be needed."],"forward_implications":["For the unadjusted Langevin algorithm, if $c=\\sup_y\\|I_d-\\eta\\nabla^2 V_\\star(y)\\|_{\\mathrm{op}}<1$, the PI/LSI constants of the iterates follow $\\alpha^{(k+1)}=2\\eta+c^2\\alpha^{(k)}$ and converge to the biased limit $2\\eta/(1-c^2)$; for $\\eta\\le 1/\\lambda$ this recovers the known bound $\\alpha^{(\\infty)}\\ge 2/\\mu$.","For the proximal sampler, when the conditional constant satisfies $\\beta<\\eta$, the LSI constants converge to $\\alpha^{(\\infty)}=(1/\\beta-1/\\eta)^{-1}$, and for a $\\mu$-strongly convex target this equals $1/\\mu$, the constant predicted by strong convexity.","For exact HMC with a Gaussian target, the LSI constants of the iterates converge to $1/\\lambda_{\\min}(M)$, consistent with the unbiased nature of exact HMC.","When the conditionals do not depend on $y$, the product measure satisfies $\\Phi\\mathrm{SI}(\\max\\{\\alpha,\\beta\\})$ and the convolution satisfies $\\Phi\\mathrm{SI}(\\alpha+\\beta)$; the product constant is attained by Gaussian factors.","Because PI/LSI control variance and concentration of functionals, these recursions translate directly into an error analysis for plug-in estimators that use $N$ parallel MCMC chains."],"supporting_citations":[{"why":"Supplies the $\\Phi$-entropy and $\\Phi$-Sobolev framework and the variational proposition that anchors the unified proof.","marker":"Chafaï, 2004"},{"why":"Prior log-Sobolev two-scale criterion with Lipschitz Hessian assumptions that the MGF condition generalizes.","marker":"Otto and Reznikoff, 2007"},{"why":"Prior Poincaré two-scale criterion via a bounded conditional score; one of the cases verifying the Var condition.","marker":"Mou et al., 2019"},{"why":"Prior Poincaré two-scale criterion using bounded score variance and pointwise Fisher information; another case verifying Var.","marker":"Ge et al., 2020"},{"why":"Supplies the Herbst, Bakry-Émery, and Holley-Stroock tools used to verify Var and MGF in the applications.","marker":"Bakry et al., 2014"},{"why":"Provides the ULA isoperimetry result recovered as a special case of the paper's recursion.","marker":"Vempala and Wibisono, 2023"},{"why":"Supplies the conditional Poincaré constant for exact HMC used in the HMC application.","marker":"Chen and Vempala, 2022"},{"why":"Gives the Lipschitz-transport log-Sobolev bound used to verify the conditional LSI in one proximal-sampler setting.","marker":"Brigati and Pedrotti, 2024"}],"fun_headline_variants":["Two-scale score bounds give mixture Poincaré and log-Sobolev","Unified Φ-Sobolev criterion transfers constants to mixtures","Score variance condition gives mixture Poincaré and log-Sobolev","Two-scale criteria: one bound for Poincaré and log-Sobolev"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every conditional density's sensitivity to its conditioning point is uniformly controlled—its variance, or its exponential tail, is bounded by the same constant for every possible conditioning value—and if this uniformity fails anywhere with positive weight, the proof's cross term is uncontrolled.","fun_headline_variants_meta":{"raw":{"variants":["Two-scale score bounds give mixture Poincaré and log-Sobolev","Unified Φ-Sobolev criterion transfers constants to mixtures","Score variance condition gives mixture Poincaré and log-Sobolev","Two-scale criteria: one bound for Poincaré and log-Sobolev"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000698,"raw_usage":{"total_tokens":3173,"prompt_tokens":982,"completion_tokens":2191,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":2111}},"tokens_in":598,"tokens_out":2191,"duration_ms":384559,"temperature":1.0,"reasoning_tokens":2111,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:51:00.711426+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take scalar Gaussians $\\rho=\\mathcal{N}(0,\\sigma_1^2)$ and $P_{X|Y=y}=\\mathcal{N}(a y,\\sigma_2^2)$. Then $\\alpha=\\sigma_1^2$, $\\beta=\\sigma_2^2$, $\\bar L^2=a^2/\\sigma_2^2$, and the mixture is $\\mathcal{N}(0,a^2\\sigma_1^2+\\sigma_2^2)$, whose exact LSI constant is $a^2\\sigma_1^2+\\sigma_2^2=\\xi$. Checking this identity across parameter ranges tests the constants; any parameter regime where the optimal mixture constant exceeds $\\xi$ while the assumptions hold would falsify the theorem.","supporting_citations":[{"cited_title":"Reznikoff","cited_arxiv_id":null,"evidence_quote":"Prior log-Sobolev two-scale criterion with Lipschitz Hessian assumptions that the MGF condition generalizes."},{"cited_title":"Analysis and geometry of M arkov diffusion operators , volume 348 of Grundlehren der mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences]","cited_arxiv_id":null,"evidence_quote":"Supplies the Herbst, Bakry-Émery, and Holley-Stroock tools used to verify Var and MGF in the applications."}],"review_version":2}