{"id":"c0d73189-9466-4b79-ae6e-e0ae2864ed0f","arxiv_id":"2602.14616","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"New Hit-and-Run MCMC variants that incorporate gradient and curvature—LHR, smHR, smLHR—sample linearly constrained densities more efficiently than several existing samplers, with convergence guarantees.","lead":"This paper introduces new MCMC samplers for probability distributions that live inside a constrained region (a polytope). The samplers use gradient and curvature information while guaranteeing every proposed point stays inside the region, and their experiments show efficiency gains on hard constrained problems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The convergence proof's clipping construction (Eq. 8) chooses the largest positive ratio instead of the smallest, so the drifted base point can leave the polytope, invalidating Theorem 3.2 as written.","rationale":"The central claim is that combining higher-order information with the HR proposal mechanism improves sampling efficiency. A necessary condition is that the samplers converge to the target, which the paper tries to establish via Thm. 3.2. The reader's weakest assumption identifies exactly the flaw in Eq. (8): the clipping uses 'largest' instead of 'smallest' positive ratio, breaking the proof. This is load-bearing because if the base point is infeasible, the EHR proposal is not defined on P and the chain may not have π as its stationary distribution. However, this is likely a typo rather than a fundamental flaw: the paper's own wording says 'closest constraint', and the code may well implement the correct min. The empirical evaluation is extensive and shows plausible gains; the transposed Cholesky issue that the reader also notes is secondary because the algorithm is still a valid MH proposal even if it is not exactly smMALA. Given that the theoretical claim needs a fix but the empirical contribution appears solid, the conditional acceptance verdict remains appropriate. I agree with the reader that this is the weakest assumption and no stronger objection is warranted.","tokens_in":40645,"tokens_out":5397,"duration_ms":55092,"concrete_test":"Inspect the public code in the blackjax fork (https://github.com/ripaul/blackjax) for the LHR drift-clipping implementation. In the file implementing the proposal, check whether the code computes the step size as `min` or `max` over the positive entries of the ratio vector from Eq. (8). Then run a minimal experiment in 2D: target a standard Gaussian truncated to the unit square [0,1]^2, with gradient at a point near a corner pointing outside; compare the acceptance rate and the L1 marginal error of LHR with the printed max formula versus the intended min formula. If the code uses max, the base point will be infeasible and the sampler will show a clear bias or error; if it uses min, the theoretical proof can be repaired by changing one word in the text.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The convergence guarantee (Thm. 3.2) depends on the proposal's base point x + ε̂∇logφ(x) staying inside P. In §3.1, Eq. (8) defines a vector κ of ratios (b_i − a_iᵀx)/(a_iᵀ∇logφ(x)) and then states 'we choose κ as the largest positive entry from κ'. The feasible step along a ray is the smallest positive ratio—the first constraint hit. If the largest is used, κ/2 can exceed the distance to the nearest boundary, making x + ε̂∇logφ(x) infeasible. The proof of Thm. 3.2 explicitly asserts 'x + κ∇logφ(x) ∈ P' and uses convexity to conclude the clipped point is in P. That assertion is false when κ is the largest ratio. The text's own phrase 'closest constraint' indicates the intended κ is the smallest positive ratio, suggesting a typo, but as printed the proof's load-bearing premise is wrong. This threatens the theoretical validity of the proposed samplers, though the empirical results may still stand if the code implements the intended min.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes three Hit-and-Run-style MCMC samplers for densities constrained to polytopes: LHR, smHR, and smLHR. LHR adds a clipped log-density gradient drift to the standard HR proposal; smHR uses an elliptical HR proposal with a curvature-based metric; smLHR combines both. The authors derive the elliptical HR proposal density, state a total-variation convergence theorem under positivity/continuity assumptions, and evaluate the methods on 2240 synthetic constrained problems and two 13C-MFA posterior sampling problems. They report that the new samplers often outperform RWMH, MALA, smMALA, Dikin walk, MAPLA, and HR, and they provide code.","tokens_in":40966,"tokens_out":5614,"duration_ms":56633,"significance":"If the results hold, the paper offers practically useful samplers for Bayesian inverse problems with polytope constraints. The benchmark is unusually broad, the real-world 13C-MFA examples are nontrivial, and the code is publicly available. The derivation of the elliptical HR density and the overall MH construction are sound. However, the theoretical convergence proof as written contains a concrete error in the clipping construction, and the empirical tuning protocol uses the ground-truth L1 error on the test problems. These are load-bearing issues for the central claims, though both appear fixable.","major_comments":[{"comment":"Eq. (8) defines a vector of ratios and then states 'we choose κ as the largest positive entry.' The largest positive ratio is not the first constraint hit along the drift direction; the feasible step is the smallest positive ratio. With the largest κ, the asserted inclusion x+κ∇logφ(x)∈P is false in general, so the proof's use of convexity to conclude x+ε̂∇logφ(x)∈P collapses. The surrounding text says 'closest constraint,' which suggests the intended definition is the smallest positive entry. This is a typo-level fix, but as printed the proof of Thm. 3.2 is invalid. Please correct Eq. (8) to use the minimum and ensure the code implements the minimum.","section":"§3.1, Eq. (8); Appendix A, Proof of Thm. 3.2"},{"comment":"The experimental protocol tunes the step size (and for manifold samplers the metric and δ) for each algorithm on each problem by grid search against the ground-truth L1 error computed from separate large MCMC simulations, and then reports the best L1 and minESS for that tuned setting. This is an information leak: the test metric is used as the selection criterion. The reported relative performances are therefore optimistic upper bounds and may not reflect what a user would obtain without access to ground truth. I recommend reporting results under a data-independent tuning rule (e.g., maximum acceptance rate, or a short pilot run) or showing sensitivity to the tuning criterion.","section":"§5.1, Evaluation"},{"comment":"The text states that the smHR/smLHR metric is s.p.d. 'by construction in Eq. (14) and for δ large enough.' Eq. (14) takes the element-wise square root of the absolute values of H^T H and adds δI; this matrix is not guaranteed to be positive definite for arbitrary δ. The paper's fallback to the diagonal when Cholesky fails changes the proposal covariance discontinuously, and the convergence theorem's assumption that Σ is s.p.d. is then not ensured by the stated construction. Please state explicitly that the metric used in practice is the diagonal fallback or impose a condition on δ that guarantees positive definiteness.","section":"§3.2.1; §5.1"},{"comment":"The paper motivates LHR/smLHR by the fact that HR with a χ_d step distribution becomes a Gaussian proposal as γmax→∞, and then claims that LHR simplifies to MALA in this limit. However, the experiments use a half-normal step distribution moment-matched to χ_d, not the χ_d itself. For the implemented samplers the unconstrained-limit proposal is not exactly Gaussian, so the 'recovers MALA' statement is only approximate for the benchmarked algorithms. Please clarify this distinction between the theoretical limit (χ_d) and the implementation (half-normal).","section":"§3.1, unconstrained limit"}],"minor_comments":[{"comment":"The same symbol κ is used for the vector of ratios and for the scalar clipping length. Use a different symbol for the vector to avoid confusion.","section":"Eq. (8)"},{"comment":"The proof header says 'Proof of Thm. 3.1' but the result is Lemma 3.1; similarly 'Thm. A.1' appears where 'Lemma A.1' is meant. Also Assumption A2 states p is positive on R_+, while Lemma A.1 says 'positive and continuous on R'; please align the domains.","section":"Appendix A"},{"comment":"The captions of Fig. 5 and Fig. 6 refer to 'Eq.(14) Eq.(15)' but the text introduces only the squared Hessian metric (Eq. (13)) and the scaled squared Hessian metric (Eq. (14)). It seems Eq. (15) should be Eq. (13), or an equation number is missing.","section":"Figure 5 and text"},{"comment":"The δ-parametrization sets δ=λ^{-2} and fixes ε=1. In the convergence proof, the assumption 'for δ large enough' is stated, but in the λ-parametrization, λ large corresponds to δ small. Please make the dependence of the theoretical condition on λ explicit.","section":"§3.2.1, δ-parametrization"}],"recommendation":"major_revision","confidential_remarks":"The core idea is sound and the empirical study is unusually thorough for this niche. The Eq. (8) clipping error is almost certainly a typo and should be easy to fix, but it currently invalidates the stated proof of Theorem 3.2. The tuning protocol is a more subtle issue: reporting best L1 after peeking at ground truth makes the performance numbers hard to interpret. I would be willing to see a revision that fixes Eq. (8), clarifies the metric positive-definiteness condition, and adds a data-independent tuning comparison."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is genuinely new: wrapping HR's feasibility-guaranteeing proposal around gradient and curvature information, with the clipped drift and an elliptical direction distribution. The derivation of the EHR proposal density for state-dependent metrics is correct, and the two Hessian-based metrics are a reasonable addition. The empirical study is the strongest part — 2240 synthetic problems plus two 13C-MFA cases, with code released. That alone makes this worth serious attention.\n\nNow the soft spots, in order of importance. The stress-test concern lands. Eq. (8) says to take the largest positive ratio from the vector defining intersections along the gradient, but the feasible step is the smallest positive ratio — the first constraint hit. The text's own 'closest constraint' phrasing says the authors meant the smallest. As printed, the proof of Thm. 3.2 asserts x + κ∇logφ(x) ∈ P, which is false if κ is the largest ratio. This is probably a typo, but it is load-bearing in the convergence argument. Fixing it to the intended minimum should repair the proof.\n\nThe second issue is the claimed reduction to smMALA in the unconstrained limit. For LHR, yes: HR with χd step becomes Gaussian, so LHR becomes MALA. For smHR/smLHR, no. With Cholesky G^{-1} = LL^T and direction v = L^T u, the effective covariance is L^T L, not LL^T. So the proposal does not simplify to smMALA. That claim needs to be corrected or the Cholesky orientation needs changing.\n\nA milder caveat: the benchmark grid-searches hyperparameters per problem and reports the best metric variant. That likely overstates the typical gain. It is common in this literature, but the authors should be more explicit about the selection procedure.\n\nNone of this invalidates the empirical method. The MH filter keeps the chain correct as long as the base point of the proposal is feasible; the clipping typo, once fixed to minimum, restores that. The performance results on the non-stationary 13C-MFA problem are striking. This paper deserves peer review, and the theoretical claims are repairable. I would engage with it.","headline":"Real algorithmic contribution with a large benchmark, but the convergence proof has a fixable typo (largest vs smallest ratio) and the smMALA reduction claim is not mathematically right; worth refereeing.","tokens_in":41433,"tokens_out":2625,"would_cite":true,"duration_ms":28730,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65C05","60J22","62F15"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that gradient and curvature information can be grafted onto the Hit-&-Run proposal mechanism—while preserving its guarantee of feasible proposals—to make MCMC sampling of non-uniform densities on linearly constrained domai","keywords":["Markov chain Monte Carlo","Hit-and-Run","constrained sampling","polytope","Langevin dynamics","Riemannian manifold","Bayesian inverse problems","metabolic flux analysis"],"falsifier":"Take a 2D triangle, place x near one slanted side, and set the log-density gradient pointing toward the opposite vertex. Compute κ from Eq. (8) as the largest positive ratio (b_j - a_j^T x)/(a_j^T ∇logφ(x)) and check whether x+(κ/2)∇logφ(x) lies in the triangle. If it falls outside, the feasibility premise of Theorem 3.2 is violated for an explicit state, and the proof of convergence for the printed LHR algorithm would not apply.","tokens_in":40556,"feed_emoji":"🎲","tokens_out":7970,"duration_ms":68127,"temperature":0.7,"pith_summary":"The paper's central claim is that the reason gradient-based samplers struggle on constrained domains is not the gradient itself, but the fact that unconstrained proposals frequently leave the feasible region. It therefore fuses first- and second-order information—the log-density's gradient and curvature—into the Hit-&-Run proposal, which by construction only proposes points inside the polytope. The result is three samplers (LHR, smHR, and smLHR) that are proven to converge to the target density under mild assumptions. In a benchmark of 2,240 problems plus two metabolic-flux inverse problems, the new samplers often beat existing constrained samplers, and the curvature-aware versions were the only ones to converge on the harder non-stationary problem. A sympathetic reader would care because Bayesian inverse problems in science routinely combine expensive likelihoods with linear constraints, where every rejected proposal is wasted compute.","feed_headline":"Gradient-guided Hit-and-Run samples constrained densities faster","feed_subtitle":"All proposals stay feasible, which saves compute in Bayesian inverse problems with linear constraints.","key_machinery":"The central object is the elliptical Hit-&-Run (EHR) proposal: it draws a uniform direction u on the sphere, maps it through the Cholesky factor of a covariance to obtain v=L^T u, truncates the step length at the first intersection with the polytope boundary, and produces a feasible proposal. Lemma 3.1 derives the EHR proposal density in closed form. Gradient information enters by shifting the proposal base point to x+ε̂∇logφ(x) (LHR), where ε̂ is a clipped step size chosen so the base point stays feasible; curvature enters by setting the covariance to ε²G(x)^{-1} (smHR), and smLHR does both. The fact that the proposal density is positive and continuous on the whole polytope is what carries","core_discovery":"On the paper's own terms, the discovery is that a Gaussian proposal can be decomposed into a direction and a magnitude, and Hit-&-Run already samples exactly this way; therefore gradient and curvature information can be injected into a feasible-direction proposal without giving up the feasibility guarantee. The Langevin Hit-&-Run (LHR) clips the gradient drift so the proposal base point remains inside the polytope and then runs Hit-&-Run from that point. The simplified manifold versions replace the spherical direction with an ellipsoidal direction derived from a metric tensor, giving the elliptical Hit-&-Run proposal; smLHR combines the clipped natural-gradient drift with that ellipsoidal di","pith_inferences":["If the clipping definition is corrected to measure the nearest constraint in the gradient direction, the feasibility proof becomes airtight; testing both definitions on a simple triangle would separate a proof fix from a substantive design choice.","The same direction/magnitude decomposition could be applied to other proposal mechanisms, such as Hamiltonian trajectories, by replacing the Gaussian magnitude with the truncated step distribution, yielding feasible curvature-informed proposals beyond the Langevin family.","Because the Hessian-based metric is local and automatic differentiation supplies it cheaply, these samplers could be combined with adaptive metric estimation to remove manual δ tuning, making them plug-in samplers for black-box constrained densities.","The benchmark's scale axis suggests a practical selection rule: use a curvature-aware sampler when the high-probability region approaches a boundary, and a plain feasible sampler when the distribution is nearly uniform."],"forward_implications":["If the central claim holds, constrained Bayesian inverse problems with expensive likelihood evaluations can be sampled with fewer wasted proposals, because every proposed state lies inside the feasible set.","The curvature-aware samplers smHR and smLHR are the only methods tested that converge within budget on the 9-dimensional non-stationary metabolic-flux problem, suggesting that curvature information is necessary for strongly correlated, linearly constrained targets.","The paper's δ-parametrization, which controls step size through the metric's regularization rather than an explicit scale, makes the second-order methods easier to tune, though the paper leaves the reason for its empirical success unexplained.","When the target is close to uniform, the gain from gradient information shrinks and the plain feasible sampler remains competitive, so the methods do not lose much in the regime where boundary constraints dominate."],"fun_headline_variants":["Higher-order Hit-and-Run for constrained densities","Gradient-guided Hit-and-Run: feasible proposals, faster sampling","Elliptical Hit-and-Run: curvature-aware constrained MCMC","Hit-and-Run with gradients and curvature for constrained domains","Feasible drift: higher-order Hit-and-Run for polytopes"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The convergence proof for LHR and smLHR assumes the clipped gradient step x+ε̂∇logφ(x) stays inside the polytope, which requires κ to be the distance to the nearest constraint in the gradient direction; as printed, Eq. (8) selects the largest positive ratio rather than the smallest, so the proposal base point can leave the feasible region.","fun_headline_variants_meta":{"raw":{"variants":["Higher-order Hit-and-Run for constrained densities","Gradient-guided Hit-and-Run: feasible proposals, faster sampling","Elliptical Hit-and-Run: curvature-aware constrained MCMC","Hit-and-Run with gradients and curvature for constrained domains","Feasible drift: higher-order Hit-and-Run for polytopes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000234,"raw_usage":{"total_tokens":1292,"prompt_tokens":659,"completion_tokens":633,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":403,"completion_tokens_details":{"reasoning_tokens":546}},"tokens_in":403,"tokens_out":633,"duration_ms":6479,"temperature":1.0,"reasoning_tokens":546,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T23:08:28.768096+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a 2D triangle, place x near one slanted side, and set the log-density gradient pointing toward the opposite vertex. Compute κ from Eq. (8) as the largest positive ratio (b_j - a_j^T x)/(a_j^T ∇logφ(x)) and check whether x+(κ/2)∇logφ(x) lies in the triangle. If it falls outside, the feasibility premise of Theorem 3.2 is violated for an explicit state, and the proof of convergence for the printed LHR algorithm would not apply.","supporting_citations":[],"review_version":1}