{"id":"b6d74622-bc55-4329-976b-ec6a066e4f80","arxiv_id":"2412.16710","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":2,"one_line_summary":"For diffusions on Riemannian manifolds with boundary, the paper establishes a quantitative space-time divergence lemma and proves that randomized Hamiltonian Monte Carlo and Langevin dynamics achieve the optimal square-root reduction in relaxation time.","lead":"The paper proves sharp quantitative bounds on how much faster certain non-reversible Markov chains, such as randomized Hamiltonian Monte Carlo and Langevin dynamics, can converge to equilibrium when sampling from probability measures on curved spaces with boundaries. These bounds show the acceleration is at most a square root and that the studied algorithms actually achieve the optimal square-root speedup on convex domains.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The proof of Theorem 5 constructs g whose weak second time-derivative contains a Dirac mass, so the asserted regularity g∈H^{2,2}_D(µ) in (14) is not established.","rationale":"The reader identified Assumption 1(i), local convexity, as the load-bearing premise. Local convexity is indeed essential for dropping the boundary term in the Reilly formula, and I do not dispute that. However, the more immediate and concrete obstacle in the proof as written is the regularity of the functions g constructed in the high-mode cases of Theorem 5. The explicit cutoff φ_k is only C¹, making w_k=u_k−v˙_k have a jump in its first derivative. Such a function is not in H²([0,T]), so g_k is not in H^{2,2}_D(µ). This is not a matter of insufficient smoothness assumptions on M or U; it is an internal inconsistency in the construction relative to the theorem's stated regularity. The gap is likely fixable by a smoothing argument or by restating Theorem 5 with the weaker regularity that the proof actually establishes (g(t,·)∈Dom(L), ∇g, ∂_t∇g, ∇²g∈L²), which is all that Theorem 15 uses. Because the main optimality claims rest on the divergence lemma, the paper should not be accepted without correcting this point. I therefore recommend keeping the CONDITIONAL verdict, but for a different reason than the reader's weakest-assumption choice.","tokens_in":31834,"tokens_out":35848,"duration_ms":297609,"concrete_test":"Take α_k=3/T and u_k(t)=e^{−αt}−e^{−α(T−t)} in Case 3, and compute the left and right derivatives of the constructed w_k at t=α_k^{-1}. The difference is −2α_k²∫_0^{α_k^{-1}}u_k(s)ds, which is strictly positive, so w_k'' contains a Dirac mass at t=α_k^{-1}. This directly shows g_k=(1/α_k²)w_k e_k is not in H^{2,2}_D(µ), falsifying the claimed regularity (14) of Theorem 5 for the constructed solution.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"In Cases 3 and 4 of the proof of Theorem 5 (high modes), the authors set φ_k(t)=(α_k t−1)^2 1_{[0,α_k^{-1}]}(t), which is only C¹ and piecewise C², and define v_k by v_k(t)=φ_k(t)∫_0^t u_k(s)ds on [0,T/2] (and symmetrically on [T/2,T]), with w_k=u_k−v˙_k. Then g_k=(1/α_k²)w_k e_k. The function w_k is continuous, but w'_k has a jump at t=α_k^{-1}: the left derivative is u'_k−2α_k²∫_0^{α_k^{-1}}u_k(s)ds while the right derivative is u'_k, and the integral of u_k=e^{−αt}∓e^{−α(T−t)} over [0,α_k^{-1}] is strictly positive. Hence the distributional second derivative w_k'' contains a nonzero Dirac mass, so w_k∉H²([0,T]). Consequently g_k∉H^{2,2}_D(µ), contradicting the regularity required in (14). Moreover, the proof never bounds ∂_t²g; the constant c1(T) in (19) is derived from estimates on ∂_t h, ∇h, ∂_t∇g and ∇²g, i.e. the reduced space-time gradient of X, not the full Hessian ∇∇g appearing in (18). Thus the central divergence lemma is not proved as stated, although the subsequent space-time Poincaré inequality in Theorem 15 may only require the reduced norm.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proves a quantitative space-time divergence lemma on product domains [0,T] x M for weighted Riemannian manifolds with boundary satisfying a local convexity condition, a lower Bakry-Emery curvature bound, and a Poincare inequality. The lemma asserts that every zero-mean f on [0,T] x M decomposes as f = d_t h - L g with explicit constants c0(T) = 2T^2 + 43/m and c1(T) = 290 + 991/(mT^2) + 43 max(1/m, T^2/pi^2) rho, with h and g satisfying Dirichlet conditions in time and g(t,.) in Dom(L). The authors then derive a space-time Poincare inequality and use the second-order lift framework to show that randomized Hamiltonian Monte Carlo with refreshments and Langevin dynamics with reflection are C-optimal lifts of overdamped Langevin diffusion on this class of manifolds, with relaxation time of order 1/sqrt(m) (hence order d for a domain of diameter d). The proof rests on a generalized Reilly formula for weighted manifolds with boundary, a spectral decomposition into high/low and symmetric/antisymmetric modes, and explicit estimates for each mode.","tokens_in":32131,"tokens_out":9903,"duration_ms":88700,"significance":"If the theorem were proved as stated, this would be a significant contribution: the constants are explicit, there is no fitting or post-hoc selection, and the extension to curved state spaces with reflecting boundary is new. The generalized Reilly formula and the careful tracking of constants are valuable, and the optimality results are structurally sound given the divergence lemma and the published lift framework [25]. However, the central regularity and full-Hessian bounds in Theorem 5 have a genuine gap, so the main result is currently conditional on a repair. The subsequent optimality results are convincing only after Theorem 5 is corrected or weakened in the way suggested below.","major_comments":[{"comment":"The functions g_k constructed in Cases 3 and 4 do not have the regularity asserted in (14). For t in [0,T/2], w_k(t) = u_k(t) - v_k'(t), where v_k(t) = phi_k(t) integral_0^t u_k(s) ds and phi_k(t) = (alpha_k t - 1)^2 1_{[0, alpha_k^{-1}]}(t). Since phi_k is only C^1 and identically zero for t >= alpha_k^{-1}, w_k is continuous and piecewise C^1, but w_k' has a jump at t = alpha_k^{-1}: the left limit is u_k'(alpha_k^{-1}) - 2 alpha_k^2 integral_0^{alpha_k^{-1}} u_k(s) ds and the right limit is u_k'(alpha_k^{-1}), and the integral is strictly positive for u_k(t) = e^{-alpha_k t} - e^{-alpha_k(T-t)}. Hence the distributional second derivative w_k'' contains a nonzero Dirac mass at alpha_k^{-1}, so g_k = alpha_k^{-2} w_k e_k is not in H^2([0,T]) and g is not in H^{2,2}_D(mu). Thus (14) is not proved for the constructed g. The theorem needs either a genuinely H^2 construction (for example, smoother cutoffs) or an explicit weakening of the regularity class.","section":"Section 3, proof of Theorem 5, Cases 3 and 4"},{"comment":"The proof controls ||d_t h||, ||grad h||, ||d_t grad g|| and the spatial Hessian ||grad^2 g|| via Corollary 4; these are the components of the reduced space-time gradient of X = (-h, grad g). However, (18) bounds ||grad grad g||^2_{L^2(mu)} with grad = (d_t, grad), which includes ||d_t^2 g||^2. No estimate for ||d_t^2 g|| is provided anywhere in Cases 1-4 or the H^perp case, and by the previous comment the constructed high-mode g has d_t^2 g containing a Dirac mass, so the L^2 norm is infinite. Therefore the asserted bound (18) with the constant c1(T) from (19) is not established. Since Theorem 15 uses Theorem 5 through Lemma 14, this gap is load-bearing; if the applications only require the reduced norm, Theorem 5 should be restated accordingly and the proof should say so explicitly.","section":"Section 3, Eq. (18) and Eq. (19)"}],"minor_comments":[{"comment":"The heading 'Let f in Hl,s' should read 'Let f in Hh,s', since the symmetric low modes Hl,s were already treated in Case 2 and the text of Case 4 concerns symmetric high modes.","section":"Section 3, Case 4 heading"},{"comment":"The proof is deferred entirely to [25, Lemma 22] with the remark that it works in the Riemannian setting. Because the present setting includes Riemannian manifolds with boundary and the lemma is used to transfer the divergence lemma to the tangent bundle, please include at least a sketch verifying (35) and (36), in particular the treatment of the boundary terms.","section":"Lemma 14"},{"comment":"The norm notation 'L^2(mu), H^{-1}(kappa)' and 'L^2(mu, H^{-1}(kappa))' is inconsistent and contains misplaced parentheses; it should be L^2(lambda tensor mu; H^{-1}(kappa)).","section":"Theorem 15 proof"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the stress-test concern lands. The paper's advertised main theorem is not proved as stated, but the defect is localized to the high-mode construction in Cases 3-4 and to the interpretation of the Hessian norm in (18). I recommend major revision rather than rejection: if the authors either construct genuinely H^2 high-mode functions or explicitly restate Theorem 5 with the reduced space-time gradient norm that the applications actually use, the rest of the argument, including the generalized Reilly formula, the space-time Poincare inequality, and the optimality corollaries, appears in good shape."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper extends the Cao–Lu–Wang space-time divergence lemma to Riemannian manifolds with boundary and uses it to prove C-optimality of reflected RHMC and Langevin lifts. That extension is real: the generalized Reilly formula with boundary terms is proved, the constants in the divergence lemma are explicit, and the bound on relaxation time O(1/√m) matches the conjectured order. This is a substantive contribution to hypocoercivity and sampling.\n\nThe main soft spot is in Theorem 5, and it is not cosmetic. In Cases 3 and 4, the construction uses φ_k(t) = (α_k t − 1)² 1_{[0,α_k^{-1}]}(t), which is only C¹. The corresponding w_k = u_k − v̇_k has a derivative jump at t = α_k^{-1}; the second distributional derivative of w_k contains a Dirac mass (the integral of u_k over [0,α_k^{-1}] is strictly positive), so w_k ∉ H²([0,T]). Hence the asserted regularity g ∈ H^{2,2}_D(µ) in (14) is not achieved. Moreover, the proof never bounds ||∂_t² g||, which is part of the full Hessian appearing in (18); the constant c1(T) is derived from the reduced gradient of X = (−h, ∇g), i.e. ∂_t h, ∇h, ∂_t ∇g, and spatial ∇²g. So (18) is not established either. The space-time Poincaré inequality in Theorem 15 only needs the reduced norm, so the main applications may survive a re-statement, but the central lemma as written is false.\n\nMinor items: Case 4 is mislabeled H_{l,s} instead of H_{h,s}; Lemma 14 is imported from the authors' [25], which is acceptable since that paper is published; Assumption 2 (discrete spectrum) is acknowledged as relaxable; and the well-posedness of reflected Langevin dynamics on manifolds is only supported by Euclidean references, which is a gap in scope.\n\nThis is repairable: smooth the cut-off or explicitly weaken the regularity and the second bound to what the proof actually yields. Given the importance of the application and the care elsewhere, I would send this to a serious referee. It is not ready as is, but it is worth engaging with.","headline":"Strong extension of the divergence lemma with a real regularity gap in the high-mode construction; main results likely salvageable but the central theorem overclaims.","tokens_in":32726,"tokens_out":7686,"would_cite":false,"duration_ms":59675,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60J25","60J35","58J65","58J05","35H10"],"pacs":[],"model":"deepseek-v4-flash","headline":"On locally convex manifolds, non-reversible lifts achieve the optimal square-root speedup of relaxation time with explicit constants.","keywords":["non-reversible lifts","space-time divergence lemma","hypocoercivity","Langevin dynamics","randomized Hamiltonian Monte Carlo","Riemannian manifolds with boundary","local convexity","Reilly formula"],"falsifier":"Take the flat interval $M=[0,d]$ with uniform measure and the low symmetric mode $f(t,x)=\\cos(2\\pi t/T)\\sin(\\pi x/d)$; solve the one-dimensional equations $f=\\partial_t h-Lg$ explicitly with $h(0)=h(T)=0$ and $g(0)=g(T)=0$, and check whether the inequalities (17) and (18) hold with the stated $c_0(T)$ and $c_1(T)$. A contradiction would falsify the divergence lemma; equality would indicate the constants are sharp in this geometry.","tokens_in":31578,"feed_emoji":"⚡","tokens_out":7502,"duration_ms":60735,"temperature":0.7,"pith_summary":"This paper proves that on a locally convex Riemannian manifold with boundary and a lower, not necessarily positive, curvature bound, every zero-mean function on the time-space cylinder $[0,T]\\times M$ decomposes as $f=\\partial_t h-Lg$, with explicit quantitative bounds on $h$ and $g$ and explicit constants $c_0(T)$ and $c_1(T)$. The constants have the order conjectured for the Euclidean case: $c_0(T)=2T^2+43/m$ and $c_1(T)=290+991/(mT^2)+43\\max(1/m,T^2/\\pi^2)\\rho$. From this divergence lemma the paper derives space-time Poincar\\'e inequalities and exponential decay for randomized Hamiltonian Monte Carlo and Langevin dynamics with reflection at the boundary. The upshot is that these non-reversible lifts are $C$-optimal: they achieve the square-root reduction of relaxation time up to a constant factor, so on a convex domain of diameter $d$ they relax in time of order $d$ rather than $d^2$.","feed_headline":"Sampling on convex domains speeds up from quadratic to linear time","feed_subtitle":"Reflected HMC and Langevin samplers relax in time order d—the square-root-optimal rate for any lift of a reversible diffusion.","key_machinery":"The load-bearing object is the generalized Reilly formula, an integration-by-parts identity for the weighted Bochner formula whose boundary terms involve the scalar second fundamental form $h$. Under local convexity $h(v,v)\\le 0$ on $\\partial M$, the boundary terms drop, and the formula yields the Hessian bound $\\|\\nabla^2u\\|^2_{L^2(\\mu)}\\le \\|Lu\\|^2_{L^2(\\mu)}+\\rho\\|\\nabla u\\|^2_{L^2(\\mu)}$ for functions with Neumann boundary conditions. That bound is what makes the space-time divergence lemma quantitative: solving $(\\partial_t^2+L)u=f$ and controlling second derivatives requires controlling the Hessian in terms of $L$ alone. The second ingredient is a spectral decomposition of $L$, with the space of space-time harmonic right-hand sides split into high and low modes and symmetric and antisymmetric parts; the low symmetric modes are the delicate case that controls the order of $c_1(T)$.","core_discovery":"The central discovery is Theorem 5: under the paper's Assumptions 1 and 2, the space-time divergence equation $f=\\partial_t h-Lg$ is solvable for any mean-zero $f$, with $h$ and $g$ vanishing at the time endpoints and $g(t,\\cdot)$ in the domain of $L$ for almost every time $t$, and with the explicit constants $c_0(T)=2T^2+43/m$ and $c_1(T)=290+991/(mT^2)+43\\max(1/m,T^2/\\pi^2)\\rho$ in the bounds (17) and (18). The proof splits the space of right-hand sides into space-time harmonic functions and their orthogonal complement, then treats symmetric and antisymmetric, high- and low-mode components separately. The paper feeds this lemma into a space-time Poincar\\'e inequality for the lifted dynamics, yielding exponential decay of time-averages with rate $\\nu=\\gamma/(\\gamma^2C_0(T)+C_1(T))$, and hence relaxation-time bounds for randomized Riemannian Hamiltonian Monte Carlo and Riemannian Langevin dynamics with specular reflection. With the choice $T=\\pi/\\sqrt{m}$ and refresh rate or friction $\\gamma$ proportional to $\\sqrt{m}$, these bounds give $C$-optimality of both lifts whenever the curvature parameter $\\rho$ is at most a constant multiple of $m$, and in particular a relaxation time of order $d$ for a convex domain of diameter $d$.","pith_inferences":["The paper states that the constants in the divergence lemma are surely not optimal, so a direct numerical check of the lemma on simple convex domains such as an interval, disk, or ball could reveal how far the values 43, 290, and 991 can be improved.","Since the authors note that the discrete-spectrum assumption can be relaxed, the same divergence lemma should extend to non-compact settings where Assumption 2 fails, with spectral projections replaced by functional calculus.","Remark 2 extends the Reilly formula to manifolds with corners, so the relaxation-time bounds should carry over to convex polytopes in $\\mathbb{R}^d$, a case relevant to volume computation and constrained sampling.","The divergence lemma is stated for continuous-time generators, but the same $h,g$ construction should transfer to discretized or numerically integrated versions of HMC, where reflection at the boundary and refreshment schedules would need separate analysis."],"forward_implications":["On a locally convex domain of diameter $d$ in Euclidean space, reflected Langevin dynamics with critical friction and randomized HMC with critical refresh rate have relaxation time of order $d$, whereas the underlying reversible diffusion has Poincar\\'e constant of order $d^2$.","The general lower bound for second-order lifts, which says no lift can relax more than a square root faster than the reversible process, is attained up to a constant factor by these two dynamics.","The explicit constants convert the space-time Poincar\\'e inequality into non-asymptotic exponential decay of time-averages, so the relaxation-time bounds are directly usable in quantitative sampling estimates.","The results cover Riemannian manifolds with possibly empty boundary and with a lower bound $\\rho\\ge 0$ on $\\mathrm{Ric}+\\nabla^2U$; when $\\rho$ is of order $m$, the optimality constants remain bounded.","For the class of potentials satisfying $\\rho\\le cm$, randomized HMC and Langevin dynamics with reflection are $C$-optimal lifts with explicit constants $C$ linear in $\\sqrt{1+c/9}$."],"supporting_citations":[{"why":"Supplies the variational space-time Poincar\\'e method that the paper adapts to Riemannian manifolds with boundary.","marker":"[4]"},{"why":"Proves the earlier quantitative divergence lemma in Euclidean space whose constant order this paper extends to manifolds with boundary.","marker":"[18]"},{"why":"Develops the second-order lift framework and the square-root lower bound on relaxation time that defines $C$-optimality.","marker":"[25]"},{"why":"Provides the generalized Reilly formula for drifting Laplacians that yields the boundary terms used in Corollary 4.","marker":"[49]"},{"why":"The classical Reilly identity that the paper's generalized formula extends to weighted manifolds.","marker":"[50]"},{"why":"Shows an alternative averaging lemma without the discrete-spectrum assumption, used to motivate that Assumption 2 is technical.","marker":"[16]"}],"fun_headline_variants":["Sampling on convex domains now linear time, not quadratic","Reflected HMC and Langevin samplers hit optimal speedup","Square-root speedup proven for reflected Riemannian samplers","Convex domain sampling: relaxation time drops to order d","Optimal non-reversible lifts: relaxation time linear in diameter"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument collapses if the boundary is not locally convex: it needs the scalar second fundamental form $h(v,v)\\le 0$ on $\\partial M$ so that the boundary terms in the Reilly formula can be discarded, and without that no Hessian bound as in Corollary 4 is available.","fun_headline_variants_meta":{"raw":{"variants":["Sampling on convex domains now linear time, not quadratic","Reflected HMC and Langevin samplers hit optimal speedup","Square-root speedup proven for reflected Riemannian samplers","Convex domain sampling: relaxation time drops to order d","Optimal non-reversible lifts: relaxation time linear in diameter"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00064,"raw_usage":{"total_tokens":3036,"prompt_tokens":1125,"completion_tokens":1911,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":741,"completion_tokens_details":{"reasoning_tokens":1825}},"tokens_in":741,"tokens_out":1911,"duration_ms":15544,"temperature":1.0,"reasoning_tokens":1825,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T10:20:54.063729+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the flat interval $M=[0,d]$ with uniform measure and the low symmetric mode $f(t,x)=\\cos(2\\pi t/T)\\sin(\\pi x/d)$; solve the one-dimensional equations $f=\\partial_t h-Lg$ explicitly with $h(0)=h(T)=0$ and $g(0)=g(T)=0$, and check whether the inequalities (17) and (18) hold with the stated $c_0(T)$ and $c_1(T)$. A contradiction would falsify the divergence lemma; equality would indicate the constants are sharp in this geometry.","supporting_citations":[{"cited_title":"Eberle and F","cited_arxiv_id":null,"evidence_quote":"Develops the second-order lift framework and the square-root lower bound on relaxation time that defines $C$-optimality."},{"cited_title":"Ma and S.-H","cited_arxiv_id":null,"evidence_quote":"Provides the generalized Reilly formula for drifting Laplacians that yields the boundary terms used in Corollary 4."}],"review_version":1}