{"id":"6113f502-bc6e-402c-bffc-94e2b80f0ab4","arxiv_id":"2608.13487","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The lazy hit-and-run walk mixes from an M-warm start on any isotropic convex body in O(n^2 ψ_n^{-2} log^3(M/ε)) total-variation distance, matching the ball walk up to logarithmic factors.","lead":"An isotropic convex body is sampled by the lazy hit-and-run walk in O(n^2 ψ_n^{-2} log^3(M/ε)) steps from any M-warm start, matching the ball walk up to logarithms. The result removes a polynomial warmness/accuracy cost from a 2026 Chen-Eldan bound and answers their open question.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Corollary 7.2's log factor is inverted: the proof derives a conductance bound that is too strong, and Theorem 7.3 relies on the reciprocal version to get (80).","rationale":"The reader accepted the paper with moderate confidence and identified Klartag's guided localization as the main external risk. I agree that the black-box import in Lemma 5.1 is a genuine risk, and I do not dispute the high-level architecture. However, an internal algebraic error is more concrete and directly load-bearing: the proof of Corollary 7.2 reverses the direction of the log factor, making the stated conductance lower bound too strong. The final theorem's stated mixing bound (80) is exactly what the corrected, weaker conductance bound yields, so a one-line fix rescues the main result. Because the manuscript as submitted contains a false inequality at the conductance step, the appropriate verdict is conditional acceptance pending that correction rather than unconditional acceptance.","tokens_in":29375,"tokens_out":40900,"duration_ms":380581,"concrete_test":"Recompute the second inequality in the proof of Corollary 7.2: for p ≥ s/2, check that 1+log(1/p) ≤ 1+log(2/s), hence p/(1+log(1/p)) ≥ p/(1+log(2/s)). Replace γ_s by c(ψ_n/√n)/(1+log(2/s)), apply Proposition 7.1, and verify that Φ_s ≳ ψ_n/(n(1+log(2/s))). Then insert this corrected conductance into Lemma 2.3 with s = ε/(2M) and confirm that the resulting bound is exactly (80), namely τ ≤ C n^2/ψ_n^2 (1+log(4M/ε))^2 log(2M/ε).","verdict_should_be":"CONDITIONAL","load_bearing_attack":"In the proof of Corollary 7.2, the text states that for p = min{μ(E), 1-μ(E)} ≥ s/2, P_{r_K}(E) ≳ (ψ_n/√n) p/(1+log(1/p)) ≥ (ψ_n/√n)(1+log(2/s))p. The second inequality is reversed: since p ≥ s/2 implies 1/p ≤ 2/s, one has 1+log(1/p) ≤ 1+log(2/s), so p/(1+log(1/p)) ≥ p/(1+log(2/s)), not ≥ (1+log(2/s))p. The correct γ_s is c(ψ_n/√n)/(1+log(2/s)), and Proposition 7.1 then gives Φ_s ≳ (ψ_n/n)/(1+log(2/s)), rather than the printed (ψ_n/n)(1+log(2/s)). This is load-bearing because it is the step that converts the isoperimetric estimate into the conductance bound used in Theorem 7.3. The final theorem's bound (80) matches the corrected reciprocal version, so the headline claim is repairable, but the manuscript as written proves a false intermediate inequality.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proves that the lazy hit-and-run walk on an isotropic convex body K mixes from any M-warm start to total-variation distance ε in O(n^2 ψ_n^{-2} log^3(M/ε)) steps, where ψ_n^{-1} is the KLS constant. The proof goes through s-conductance: it establishes a local-radius weighted isoperimetric inequality via Klartag's guided localization and a one-dimensional concave-weight isoperimetric theorem, then converts this into a conductance bound using an intrinsic local-radius metric and Lovász's overlap lemma. Appendices supply proofs of one-dimensional log-concave estimates, properties of the relaxed weighted perimeter, properties of the local conductance radius, and a Euclidean specialization of Klartag's localization theorem. The claimed bound improves on Chen--Eldan's polynomial dependence on M/ε and removes their condition n ≳ log(M/ε), thereby resolving the open question they posed.","tokens_in":29599,"tokens_out":15299,"duration_ms":131918,"significance":"The main theorem, if correct, is a significant advance in geometric random walks: it places hit-and-run on the same asymptotic footing as the ball walk up to logarithmic factors and improves the warm-start mixing time from polynomial to polylogarithmic dependence on M/ε. The paper is careful and detailed: the one-dimensional isoperimetric theorem (Section 4), the intrinsic geometry and overlap argument (Section 6), and the perimeter and localization appendices are all developed at length, and the external black boxes (Klartag's guided localization, the Lovász overlap lemma, Berwald's inequality, Cheeger--Poincaré equivalence) are identified clearly. The central mechanism—guided localization plus weighted Poincaré averaging—is novel in this context. However, there is a genuine reversed-inequality error in Corollary 7.2 that invalidates the corollary as printed; the final theorem appears repairable because its stated bound matches the corrected reciprocal form. The paper should not be accepted before this error is fixed.","major_comments":[{"comment":"The displayed inequality in the proof of Corollary 7.2 is reversed. Since p ≥ s/2 implies 1+log(1/p) ≤ 1+log(2/s), we have p/(1+log(1/p)) ≥ p/(1+log(2/s)), not p/(1+log(1/p)) ≥ (1+log(2/s))p. The asserted inequality is false for small s: for example, with s = 10^{-4} and p = 1/2 the left-hand side is about 0.045 while the right-hand side is about 5. The valid conclusion is γ_s = c ψ_n/(√n(1+log(2/s))), and Proposition 7.1 then yields Φ_s(P_HR^L) ≳ ψ_n/(n(1+log(2/s))), which is the reciprocal of the printed (79). This step is load-bearing because it is exactly the conversion of the isoperimetric estimate into the conductance bound used in Theorem 7.3. The printed Corollary 7.2 is therefore false as stated, even though the final theorem's bound (80) is consistent with the corrected reciprocal form.","section":"§7, Corollary 7.2"},{"comment":"The same reversed inequality appears in the technical overview. Inequality (3) states Φ_s(P_HR^L) ≳ (ψ_n/n)(1+log(2/s)); if this were true it would imply the stronger mixing-time bound ~ n^2 ψ_n^{-2} log(2M/ε)/(1+log(2/s))^2, not the theorem's stated n^2 ψ_n^{-2} log^3(8M/ε). The correct reduction, consistent with Theorem 7.3, is Φ_s ≳ (ψ_n/n)/(1+log(2/s)). The overview and the corollary must be corrected together so that the paper's stated reduction, its intermediate conductance bound, and the final theorem are mutually consistent.","section":"§1.2, Eq. (3)"}],"minor_comments":[{"comment":"The explicit constants c0 = 1/(64 C_loc) and α0 = 1/1000 are used throughout the proof; stating them in the lemma statement would help the reader and make the subsequent numerical checks in Proposition 7.1 easier to follow.","section":"§6, Lemma 6.5"},{"comment":"The derivation of the conditional balance property (50) relies on the statement that every measurable saturated set has zero f-integral, imported from [Kla17, Lemma 4.6]. Since this is the only point at which the balance property enters, a few more sentences explaining that import would increase confidence in this black-box step.","section":"Appendix D, proof of Lemma 5.1"},{"comment":"The provided text contains numerous spacing and line-break artifacts (for example, 'hit-and-run[ Smi84,Lov99,LV06a]; bothextend' and inconsistent spaces in expressions like 'λ(x,r )'). These should be cleaned up in the journal version.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The reversed-inequality error in Corollary 7.2 is localized and easily repairable, and I see no evidence that the main theorem is wrong once (79) is replaced by its reciprocal. However, because the error affects a load-bearing step and the printed corollary is false as stated, major revision is appropriate. The paper also depends heavily on Klartag's guided localization and on several concurrent or very recent references (Letwin, Chen--Klartag); the editor may wish to verify that those references are publicly available and stable before final acceptance. The AI-use disclosure in the acknowledgments is transparent and, in my view, not a problem."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth your time: this is the first hit-and-run mixing bound with n^2 psi^{-2} and polylog(M/epsilon), matching the ball walk. The proof architecture is genuinely nice: local-radius weighted perimeter, a one-dimensional concave-weight isoperimetric inequality, lifted by Klartag guided localization, then averaged with Poincare and Berwald. Sections 4 and 5 are careful, and I found no local error there. The high-level approach is believable and the literature is engaged honestly.\n\nBut the manuscript as written has a load-bearing sign error in Corollary 7.2. The proof derives P_{r_K}(E) >~ (psi_n/sqrt(n)) p/(1+log(1/p)) and then claims this is >= (psi_n/sqrt(n))(1+log(2/s)) p when p >= s/2. That inequality is backwards: p >= s/2 gives 1+log(1/p) <= 1+log(2/s), so the correct lower bound is p/(1+log(2/s)), not p(1+log(2/s)). Consequently the conductance should be Phi_s >~ (psi_n/n)/(1+log(2/s)), not the printed (psi_n/n)(1+log(2/s)).\n\nThis matters for two reasons. First, the wrong direction is exactly the step that converts the isoperimetric estimate into the conductance bound. Second, the theorem proof says \"substituting (79)\" to get (80), but substituting the printed (79) gives roughly n^2 psi^{-2} log(2M/epsilon)/log^2(4M/epsilon), the reciprocal of (80). So there is an internal inconsistency as written. The stress-test note is correct.\n\nThe good news: the final theorem bound (80) matches the corrected reciprocal conductance version, and the rest of the argument does not seem to depend on the erroneous direction. This looks eminently repairable: flip the inequality in Corollary 7.2, state the corollary with the reciprocal log, and the proof of Theorem 7.3 works. I have not found a second issue of the same magnitude.\n\nAs written, the paper should not be accepted. The main claim is probably right, but the middle of the proof is currently false, and an editor or referee would need the author to confirm the fix. It is significant enough to deserve referee time in a revised form. I would bring it to a reading group to trace the corrected conductance, and I would cite the final theorem once the correction is in place.","headline":"The main theorem is probably repairable, but the paper as written contains a load-bearing inverted log factor in Corollary 7.2, and the proof of Theorem 7.3 does not actually follow from the stated conductance bound.","tokens_in":30175,"tokens_out":6832,"would_cite":true,"duration_ms":62323,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68W20","52A40","60J05"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that the lazy hit-and-run walk on any isotropic convex body mixes from an M-warm start in O(n^2 ψ_n^{-2} log^3(M/ε)) steps, matching the ball walk up to logarithmic factors.","keywords":["hit-and-run walk","mixing time","convex body sampling","KLS constant","conductance","localization","isoperimetric inequality","warm start"],"falsifier":"Compute the relaxed weighted perimeter P_{r_K}(A) for a half-space cut in a cube and compare it with the claimed lower bound of order $n^{{-1/2}}$ p/(1 + log(1/p)); a smaller value, or a measured mixing time on a cube that grows faster than $n^{2}$ $log^{3}$(M/ε) for fixed ε and M, would refute the theorem.","tokens_in":29155,"feed_emoji":"🎲","tokens_out":12923,"duration_ms":104099,"temperature":0.7,"pith_summary":"The paper claims that the lazy hit-and-run walk—a Markov chain that, from a point inside a convex body, samples its next point uniformly along a random chord—mixes in O($n^{2}$ $ψ_n^{{-2}}$ $log^{3}$(M/ε)) steps from any M-warm starting distribution on an n-dimensional isotropic convex body, where $ψ_n^{{-1}}$ is the KLS constant. Up to logarithmic factors this matches the best-known warm-start mixing time of the ball walk, and it replaces the polynomial dependence on the warmness M and accuracy ε in the previous hit-and-run bound with a polylogarithmic one. The result matters because hit-and-run needs no step-size parameter, so a parameter-free sampler becomes asymptotically as fast as the state-of-the-art tuned walk. The proof works through conductance: every cut of small stationary mass has a large weighted perimeter when the boundary is weighted by the local radius r_K(x), and the walk's one-step transition laws overlap uniformly in the metric that this same radius defines.","feed_headline":"Hit-and-run walk now mixes as fast as the ball walk","feed_subtitle":"A warm start on any convex body reaches uniform in polylogarithmic dependence on warmness and accuracy.","key_machinery":"The carrying mechanism is a one-dimensional concave-weight isoperimetric inequality (Theorem 4.2) lifted to the body by a guided localization that preserves the mass of the cut on each needle. The central objects are the local conductance radius r_K(x), the largest radius r for which the ball B(x,r) is at least 63/64 contained in K, and the relaxed weighted perimeter P_w(E), the minimum weighted total variation of Lipschitz approximations to the indicator of E under weight w. On a single needle, the inequality bounds the weighted perimeter of a set of mass p below by a constant times (m_I/σ_I) p/(1 + log(1/p)), where m_I is the average of the concave weight and σ_I is the conditional standard deviation of the guiding coordinate. Averaging over needles uses the law of total variance, the spectral-gap inequality, and the reverse Hölder bound for concave weights to show that the average of m_I/σ_I is at least a constant times E[r_K]/√CPI(μ), and hence about ψ_n/√n. The same local radius defines the intrinsic metric d_{r_K} in which one-step overlap is uniform, and a weighted coarea inequality converts the weighted perimeter bound into an ergodic-flow, and therefore conductance, bound.","core_discovery":"The central claim is that for every isotropic convex body K, every M-warm start, and every ε<1/2, the lazy hit-and-run walk has τ_mix(ε, μ_init, μ; P^L_HR) ≲ $n^{2}$ $ψ_n^{{-2}}$ (1 + log(4M/ε))^2 log(2M/ε), which simplifies to O($n^{2}$ $ψ_n^{{-2}}$ $log^{3}$(8M/ε)). The paper derives this from an s-conductance lower bound Φ_s(P^L_HR) ≳ ψ_n $n^{{-1}}$(1 + log(2/s)) that holds for every small-set threshold s<1/4. The geometric estimate underneath is a local-radius weighted isoperimetric inequality: for every measurable A with p = min{μ(A), 1-μ(A)} ≤ 1/2, the relaxed weighted perimeter P_{r_K}(A) is at least a constant times ψ_n $n^{{-1/2}}$ p/(1 + log(1/p)). This all-scale perimeter estimate avoids discarding a boundary layer, and the construction keeps the mass of the cut equal on every localization needle, so that both the fixed-core loss and the probabilistic cut-survival loss of the previous approach disappear.","pith_inferences":["If the guided-localization machinery extends to general log-concave measures, the same local-radius weighting could give polylog-warm mixing bounds for non-uniform targets; this is an extension the paper does not claim.","The logarithmic factors come from cut transitions near the boundary, where r_K is small; a sharper one-dimensional estimate for endpoint transitions would plausibly remove a log, and this could be checked numerically on one-dimensional log-concave weights.","Cold starts are untouched: a point mass is not M-warm for any finite M, so converting this theorem into a full sampling algorithm still requires a separate warm-start construction or a cold-start argument."],"forward_implications":["Warm-start hit-and-run now carries the same O~(n^2/ψ_n^2) mixing dependence as the ball walk, so the parameter-free chord sampler loses nothing asymptotically to a step-size-tuned walk.","With the current best KLS estimate ψ_n^{-1}=O(log^{1/4} n), the bound becomes O(n^2 log^{1/2} n log^3(M/ε)) for every isotropic convex body.","The previous restriction n ≳ log(M/ε) and the polynomial (M/ε)^{11} cost disappear; mixing now depends only polylogarithmically on warmness and accuracy.","Because the conductance lower bound holds uniformly over all small-set thresholds s, the result covers the full range of warm-start accuracies, not just large cuts."],"supporting_citations":[{"why":"Supplies the guided localization theorem with a common 1-Lipschitz guiding function and conditionally balanced needles that the proof's averaging step requires.","marker":"[Kla17]"},{"why":"Provides the prior hit-and-run mixing bound with polynomial M/ε dependence and poses the open question about polylog warm-start mixing that this paper resolves.","marker":"[CE26]"},{"why":"Establishes the hit-and-run mixing framework, the step quantile definition, and the one-step overlap lemma used in the local-overlap argument.","marker":"[Lov99]"},{"why":"Introduces the local conductance radius r_K and its concavity, Lipschitz, and average-scale properties, and gives the corner hit-and-run analysis that the weighted-perimeter argument refines.","marker":"[LV06a]"},{"why":"Sets the ball-walk warm-start mixing benchmark that the theorem matches.","marker":"[KLS97]"},{"why":"Provides the s-conductance definition and the standard lemma converting a conductance lower bound into a warm-start mixing time.","marker":"[LS93]"},{"why":"Supplies the reverse Hölder inequality for concave weights on convex bodies, used to bound the second moment of the averaged needle weights.","marker":"[Ber47]"},{"why":"Provides one-dimensional log-concave density and quantile estimates used in the one-dimensional isoperimetric lemma.","marker":"[LV07]"},{"why":"Defines the Cheeger-type constant ψ_n and the associated conjecture, giving the meaning of the KLS constant throughout the bound.","marker":"[KLS95]"}],"fun_headline_variants":["Hit-and-run matches ball walk mixing speed","Polylog warm-start mixing for hit-and-run","Warm-start hit-and-run matches ball walk","Hit-and-run: polylog mixing from any warm start"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof rests on a balance-preserving localization theorem that decomposes the body into one-dimensional pieces on which every measurable cut keeps exactly its global mass; if that theorem is not valid for arbitrary measurable sets, the averaging step that produces the weighted isoperimetric inequality collapses.","fun_headline_variants_meta":{"raw":{"variants":["Hit-and-run matches ball walk mixing speed","Polylog warm-start mixing for hit-and-run","Warm-start hit-and-run matches ball walk","Hit-and-run: polylog mixing from any warm start"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001224,"raw_usage":{"total_tokens":5045,"prompt_tokens":973,"completion_tokens":4072,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":4010}},"tokens_in":589,"tokens_out":4072,"duration_ms":521955,"temperature":1.0,"reasoning_tokens":4010,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:57:25.091697+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the relaxed weighted perimeter P_{r_K}(A) for a half-space cut in a cube and compare it with the claimed lower bound of order $n^{{-1/2}}$ p/(1 + log(1/p)); a smaller value, or a measured mixing time on a cube that grows faster than $n^{2}$ $log^{3}$(M/ε) for fixed ε and M, would refute the theorem.","supporting_citations":[],"review_version":1}