{"id":"51d2b251-d9a7-4aba-9fda-a672e18595cb","arxiv_id":"1908.09363","paper_version":3,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Adaptive Langevin dynamics converges exponentially in L2 to its invariant measure with a spectral gap at least proportional to min(gamma, gamma^-1, gamma*epsilon^2, (gamma*epsilon^2)^-1), yielding a central limit theorem.","lead":"This paper proves that Adaptive Langevin dynamics, a sampling method for Bayesian inference with noisy gradients, converges to its target distribution exponentially fast, and it quantifies how the rate depends on the method's parameters. The result gives theoretical support for a widely used stochastic-gradient MCMC scheme and yields a central limit theorem for time averages.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The continuous-time theorem appears sound; the load-bearing soft spot is the state-independent white-noise assumption that connects the result to stochastic-gradient Bayesian sampling.","rationale":"The paper's central theorem is a rigorous statement about a specific continuous-time SDE, and the proof appears internally consistent. I checked the structure of the hypocoercivity argument: the coercivity estimate in Proposition 2.2, the choice of a_{epsilon,gamma}, the treatment of the epsilon > 1 regime via the rescaled regularization operator, and the determinant positivity condition all combine to yield the stated rate min(gamma, 1/gamma, gamma epsilon^2, 1/(gamma epsilon^2)). The constants C and lambda are uniform because a_{epsilon,gamma} <= a < 1. The CLT in Corollary 2 follows from the bounded inverse of the generator. The weakest link is the assumption that gradient noise is additive white noise with state-independent variance. This assumption is explicit in the model (1.2) and is necessary for the invariant measure (1.3); it is also explicitly invoked in Appendix C.2 for the numerical integrator. In the Bayesian subsampling application, minibatch gradient noise is state-dependent, so the subsampled algorithm does not exactly sample the prescribed posterior. The paper is transparent about this limitation, and the MNIST experiment is best read as an empirical illustration rather than a proof of exact sampling. Because the theorem itself is not invalidated, the verdict should remain ACCEPT/UNCHANGED.","tokens_in":27949,"tokens_out":31567,"duration_ms":304537,"concrete_test":"Run the ODABADO minibatch-subsampling scheme on the double-well model with a finite synthetic dataset, using the same batch-size construction as (1.6). Estimate the empirical conditional covariance of the minibatch gradient residual at two widely separated q values (e.g., near the two wells) from many independent minibatches. If the covariance differs by more than a few percent, or if a normality test on the residual rejects Gaussianity, the state-independent white-noise assumption is violated and the invariant measure of the subsampled scheme cannot be expected to equal pi. This would settle whether the CLT claimed for time averages applies to the advertised stochastic-gradient setting.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 2.1 establishes exponential L2 convergence for the SDE (2.4), whose generator decomposition (2.1)-(2.6) and invariant measure (1.3) depend on the gradient perturbation being an additive, state-independent Ito noise process sigma_G dW_G. I found no internal inconsistency in the hypocoercivity proof: the modified regularization operator for epsilon > 1, the determinant condition (2.24), and the eigenvalue estimates in Section 2.4 are coherent and the scaling matches the Galerkin numerics in Section 4.1. The real soft spot is the bridge to the motivating application. In stochastic-gradient MCMC, the minibatch residual R_{G,k} = grad U(q_k) - grad-hat U(q_k) from (1.6) has state-dependent covariance proportional to the empirical covariance of the per-data-point log-likelihood gradients, which varies with q. Appendix C.2 explicitly assumes R_{G,k} is Gaussian and independent of q_k. When that fails, the ODABADO scheme does not exactly sample the posterior (1.3), and the CLT of Corollary 2 does not transfer to the subsampled algorithm. This is a genuine limitation of the paper's practical reach, but it is acknowledged in the text and does not invalidate the theorem for the idealized model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes the Adaptive Langevin dynamics SDE, a sous-thermostatted Langevin system with an additional gradient-noise term of unknown, constant intensity. After rescaling the thermostat variable, the authors decompose the generator and prove (Theorem 2.1) exponential convergence in L^2(pi) for the semigroup, with a rate lower bound proportional to min(gamma, 1/gamma, gamma epsilon^2, 1/(gamma epsilon^2)). They then infer invertibility of the generator on the mean-zero subspace and derive a central limit theorem for ergodic averages with bounds on the asymptotic variance. A companion result (Proposition 3.1) shows that, for observables depending only on (q,p), the asymptotic variance converges to that of standard underdamped Langevin dynamics as epsilon -> infinity. The theoretical scaling is tested numerically by Galerkin computation of the spectral gap, by simulation of asymptotic variances, and by a Bayesian logistic-regression example on MNIST using a minibatch gradient estimator.","tokens_in":28054,"tokens_out":30472,"duration_ms":270480,"significance":"If the results are correct, this is a substantial contribution: it gives the first quantitative hypocoercivity estimate for Adaptive Langevin dynamics with explicit dependence on the friction and thermostat-mass parameters, and it upgrades earlier qualitative results (e.g., Herzog's Lyapunov-function analysis) to an L^2 rate and a CLT. The proof is self-contained and analytic, with no fitted parameters; the derived scaling is concrete and falsifiable, and the Galerkin numerics in Section 4.1 support the predicted exponents. The Bayesian-sampling application illustrates the intended use, although the bridge from the idealized model to the minibatch setting requires a modeling assumption that is not exactly satisfied.","major_comments":[{"comment":"The statement of Proposition 2.2 fixes the regularization parameter as a_{\\varepsilon,\\gamma} = a \\min(\\gamma, \\gamma^{-1}, \\gamma\\varepsilon^2, (\\gamma\\varepsilon^2)^{-1}), but the proof in the paragraph \"For \\varepsilon > 1\" instead chooses a_{\\varepsilon,\\gamma} = a \\min(\\gamma\\varepsilon, (\\gamma\\varepsilon)^{-1}). These two choices differ by a factor \\varepsilon in the regime \\varepsilon > 1: for \\gamma\\varepsilon < 1 the stated formula gives a\\gamma while the proof uses a\\gamma\\varepsilon, and for \\gamma\\varepsilon > 1 it gives a/(\\gamma\\varepsilon^2) while the proof uses a/(\\gamma\\varepsilon). With the stated formula, the determinant condition (2.24) yields a lower bound of order \\gamma/\\varepsilon (respectively 1/(\\gamma\\varepsilon^3)) instead of \\gamma (respectively 1/(\\gamma\\varepsilon^2)) in these regimes, so the uniform constant \\tilde{\\lambda} in Proposition 2.2 cannot exist. The proposition statement must be corrected, for example by giving the piecewise definition of a_{\\varepsilon,\\gamma} that is actually used in the proof, so that the statement and proof are consistent.","section":"Section 2.4, Proposition 2.2 and equation (2.13)"}],"minor_comments":[{"comment":"In the last momentum update of the ODABADO algorithm, the text reads p_{k+1} = \\alpha(\\zeta_{k+1}, \\Delta t/2) p_k + G(\\sigma_A, \\zeta_{k+1}, \\Delta t/2) R_{k+1/2}; for consistency with the symmetric splitting (C.12), the argument should be \\hat{p}_{k+1/2} (the value after the gradient-noise substep), not p_k.","section":"Appendix C.2, ODABADO algorithm"},{"comment":"The numerical CLT demonstration for Bayesian logistic regression uses a minibatch gradient estimator, whose residual noise is state-dependent and only approximately Gaussian, whereas the ODABADO integrator is derived under the assumption in Appendix C.2 that the residual is Gaussian and independent of q_k. The text should state explicitly in Section 4.3 that the experiment is an empirical check under this modeling assumption and is not a direct consequence of Corollary 2 for the actual subsampled process.","section":"Section 4.3 and Appendix C.2"},{"comment":"The word \"Collorary\" appears in place of \"Corollary\" in the headings/callouts for Corollary 1 and Corollary 2; also \"trajecotry\" appears in Section 4.2 and \"different\" appears in Section 4.2.","section":"Section 3, Corollaries 1 and 2"},{"comment":"The section title \"Pathwise ergodicity and functional central limit theorem\" promises a functional CLT, but only the ordinary CLT is stated in Corollary 2; either add the functional CLT or change the title.","section":"Section 3 title"},{"comment":"In the solvability condition (3.5), the notation \\Pi_0\\phi is ambiguous: it should be made clear that the integral is a partial average over (q,p), which vanishes because \\phi is restricted to functions of (q,p) only in Proposition 3.1; otherwise it can be mistaken for the global projection defined in (2.8).","section":"Equation (3.5)"}],"recommendation":"major_revision","confidential_remarks":"The main theorem appears substantively sound and the hypocoercivity argument is convincing; the discrepancy between Proposition 2.2 and its proof is localized and fixable, but it is a load-bearing technical point. The MNIST application would benefit from a clearer caveat about the state-dependence of the minibatch noise."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The punchline: for math.PR, Theorem 2.1 is the real thing. The proof is a complete Dolbeault–Mouhot–Schmeiser hypocoercivity argument, and the modified regularization operator for ε > 1 is the right fix to recover the γε² term. The rate min(γ, γ⁻¹, γε², (γε²)⁻¹) and the resulting CLT are new; prior work had exponential relaxation without these parameter dependencies. This is a genuine advance, and the paper deserves a serious referee.\n\nWhat is done well: the operator decomposition (2.1)–(2.6) is clear; Lemmas 2.4–2.10 supply the estimates; the determinant condition (2.24) correctly guides the choice of a_{ε,γ}; the sharpness checks with simple functions and the Galerkin numerics match the theorem’s scaling. The large-ε asymptotic variance expansion in Proposition 3.1 is a useful by-product. The MNIST experiment is reasonable evidence that the CLT holds in a discretized setting, though discretization bias is treated as negligible.\n\nThe soft spot, in proportion: the theorem is for the idealized SDE (2.4), where the gradient perturbation is additive, state-independent white noise — exactly what the invariant measure (1.3) and generator decomposition require. The paper’s motivating application is stochastic-gradient Bayesian sampling, where the minibatch residual has state-dependent covariance. Appendix C.2 explicitly assumes R_{G,k} is Gaussian and independent of q_k. When that fails, the ODABADO scheme does not exactly sample the posterior, and Corollary 2’s CLT does not automatically transfer. The authors acknowledge this, but it is the main gap between theorem and practice. Two smaller points: the lower bound 1/(γε²) is only demonstrated numerically, not proved sharp, and the ε → 0 singular limit is explicitly scoped out in Remark 3.2. Both are honest, stated limitations.\n\nThe citation pattern is fine. The paper builds on the authors’ own earlier constructions [16,22,33], but the convergence theorem is not assumed there; citing them is appropriate. No code or data are shipped, but the integrators are specified well enough to reimplement.\n\nOverall: the math is coherent, the central theorem holds, and the main limitation is the idealized noise model. This paper is for people working on MCMC theory, especially stochastic-gradient and thermostatted samplers; they should read it. I would cite it if working in this area. Recommendation: send to peer review; it deserves careful refereeing, with the key question being how far the result can be extended to state-dependent minibatch noise.","headline":"Solid hypocoercivity proof for Adaptive Langevin with an honest parameter-dependent rate; the main gap is the idealized state-independent noise model that limits how far the theorem reaches into stochastic-gradient MCMC.","tokens_in":28724,"tokens_out":1834,"would_cite":true,"duration_ms":19723,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60J70","35B40","46N30","35Q84","65C30"],"pacs":[],"model":"deepseek-v4-flash","headline":"Adaptive Langevin sampling converges exponentially fast to its target measure.","keywords":["Adaptive Langevin dynamics","hypocoercivity","exponential convergence","central limit theorem","stochastic gradient noise","Bayesian posterior sampling","spectral gap","thermostat"],"falsifier":"On the quadratic potential $U(q)=q^2/2$, compute the spectral gap of a finite spectral discretization of $-L_{\\mathrm{AdL}}$ along a curve with $\\gamma\\varepsilon^2$ fixed and $\\varepsilon\\to0$; Theorem 2.1 requires the gap to stay at least a fixed fraction of $\\gamma\\varepsilon^2$, so a gap decaying faster than that bound would falsify the rate. Separately, replacing the additive gradient noise in a simulation by state-dependent noise of the same mean and comparing the empirical stationary distribution with $e^{-\\beta U}$ would test whether the invariant-measure assumption carries over.","tokens_in":27612,"feed_emoji":"🎲","tokens_out":9440,"duration_ms":91557,"temperature":0.7,"pith_summary":"Adaptive Langevin dynamics (ALD) is a sampling scheme that promotes the friction coefficient to a dynamical variable updated by a feedback control law, so that unknown noise corrupting the gradient of a target potential does not destroy the invariant measure. This paper establishes that, for smooth target measures with a spectral-gap inequality on the position marginal and gradient noise of additive white-noise type, the law of ALD converges exponentially fast in $L^2(\\pi)$ to the target measure. The exponential rate is explicitly bounded below by a constant times $\\min(\\gamma, \\gamma^{-1}, \\gamma\\varepsilon^2, (\\gamma\\varepsilon^2)^{-1})$, where $\\gamma$ is the total noise intensity and $\\varepsilon=\\sqrt{\\nu}$ is the square root of the thermostat mass. A central limit theorem for single-trajectory time averages follows, with asymptotic variance bounded by the reciprocal of that rate. The result matters because ALD is used in Bayesian inference with stochastic minibatch gradients, where previously no quantitative $L^2$ convergence rate was available.","feed_headline":"Adaptive Langevin sampling converges exponentially fast","feed_subtitle":"Explicit rate bounds in noise and thermostat mass yield a central limit theorem for time averages.","key_machinery":"The argument is carried by a modified norm $\\mathcal{H}(\\phi)=\\tfrac12\\|\\phi\\|^2_{L^2(\\pi)}+a_{\\varepsilon,\\gamma}\\langle A_\\varepsilon\\phi,\\phi\\rangle_{L^2(\\pi)}$, where $A_\\varepsilon$ is a regularizing operator constructed from the antisymmetric part of the generator, that is, the Hamiltonian part plus the scaled feedback-control part. The definition of $A_\\varepsilon$ switches at $\\varepsilon=1$: for $\\varepsilon\\le 1$ it regularizes the Hamiltonian transport, while for $\\varepsilon>1$ the generator is first multiplied by $\\varepsilon$ and the feedback part is regularized instead. This switch is what makes the partial coercivity estimate hold with the correct $\\varepsilon$-dependent rate. The resulting coercivity of $-L_{\\mathrm{AdL}}$ in the modified scalar product gives the exponential decay via a differential-inequality argument, and the same coercivity bounds the inverse generator, which is the input to the central limit theorem.","core_discovery":"The central claim is Theorem 2.1: for any $\\varepsilon,\\gamma>0$, the semigroup $e^{tL_{\\mathrm{AdL}}}$ contracts in $L^2(\\pi)$ toward $\\int\\phi\\,d\\pi$ with rate at least $\\lambda\\min(\\gamma,\\gamma^{-1},\\gamma\\varepsilon^2,(\\gamma\\varepsilon^2)^{-1})$. The proof exhibits a modified norm in which the generator is coercive, and then applies a differential-inequality estimate. The invertibility of the generator on mean-zero functions follows, and with it a functional central limit theorem: $\\sqrt{t}(\\hat\\phi_t-\\mathbb{E}_\\pi\\phi)$ converges to a centered normal distribution with variance $2\\langle -L_{\\mathrm{AdL}}^{-1}\\Pi_0\\phi,\\Pi_0\\phi\\rangle_{L^2(\\pi)}$, bounded above by $O(\\max(\\gamma,\\gamma^{-1},\\gamma\\varepsilon^2,(\\gamma\\varepsilon^2)^{-1}))$. In the large-thermostat-mass limit $\\varepsilon\\to\\infty$, the asymptotic variance approaches the underdamped Langevin variance at order $1/\\varepsilon$, so the added thermostat does not degrade efficiency in that regime. Numerical experiments on a double-well model and on Bayesian logistic regression for handwritten-digit classification confirm the predicted scalings and the normal limit law.","pith_inferences":["In the actual minibatch setting the gradient residual is state-dependent and only approximately normal, so the theorem does not by itself certify the Bayesian posterior; a natural test would be to measure the bias introduced by a state-dependent $\\sigma_G^2(q)$ and compare with the white-noise idealization.","The $\\varepsilon\\to0$ singular limit, where the feedback dynamics dominates on level sets of $\\Phi=\\xi^2+|p|^2-(2n/\\beta)\\ln|p|^2$, suggests a homogenized slow description that could yield sharp asymptotic variance predictions; the paper does not carry this out.","The same modified-norm machinery should transfer to other thermostatted sampling schemes whose generator splits into symmetric and antisymmetric parts, giving explicit rates in terms of their parameters.","A quantitative check of the theorem's boundary: increasing minibatch size lowers gradient-noise variance; the theory predicts the rate improves as $\\gamma$ decreases toward order one, which is testable on a fixed data set."],"forward_implications":["Users of ALD can estimate posterior expectations from one long trajectory and attach asymptotically valid normal error bars, provided the gradient noise is approximately additive and white.","The rate formula identifies the dangerous regimes: mixing degrades as $\\gamma\\to0$ or $\\gamma\\to\\infty$, and as $\\nu=\\varepsilon^2\\to\\infty$ or $\\nu\\to0$, with the precise power law $\\min(\\gamma,\\gamma^{-1},\\gamma\\varepsilon^2,(\\gamma\\varepsilon^2)^{-1})$.","The inverse-generator bound of order $\\max(\\gamma,\\gamma^{-1},\\gamma\\varepsilon^2,(\\gamma\\varepsilon^2)^{-1})$ tells practitioners how long to integrate: of order that maximum will bring the variance of a time average to order one.","As $\\varepsilon\\to\\infty$, ALD inherits the sampling efficiency of underdamped Langevin dynamics, since the asymptotic variance converges to the Langevin value at rate $1/\\varepsilon$."],"supporting_citations":[{"why":"Supplies the hypocoercivity framework and the regularization-operator construction that the proof adapts to ALD.","marker":"[8, 9]"},{"why":"Provides a recent Lyapunov-based exponential relaxation result for the same dynamics, whose quantitative spectral-gap scaling this paper refines.","marker":"[13]"},{"why":"Introduced adaptive Langevin dynamics as a sampling method for systems with noisy gradient forces.","marker":"[16]"},{"why":"Analyzes adaptive thermostats for noisy gradient systems and supplies the splitting integrators used in the numerical experiments.","marker":"[22]"},{"why":"Makes the connection between thermostatted dynamics and Bayesian posterior sampling with stochastic gradients.","marker":"[5]"},{"why":"Establishes the stochastic-gradient Langevin baseline that motivates the noisy-gradient sampling setting.","marker":"[37]"},{"why":"Gives the functional central limit theorem for Markov processes from a bounded inverse generator, used to obtain Corollary 2.","marker":"[2]"},{"why":"Provides resolvent estimates for underdamped Langevin dynamics used in the large-$\\varepsilon$ expansion of the asymptotic variance.","marker":"[18]"}],"fun_headline_variants":["Adaptive Langevin: exponential mixing with explicit rate","Thermostat noise tamed: exponential convergence proven","Explicit hypocoercivity rate for adaptive Langevin sampling","Adaptive Langevin speeds to equilibrium: quantified rate","Central limit theorem for adaptive Langevin time averages"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that gradient corruption is an additive white-noise term with constant covariance matrix $\\sigma_G^2 I$, independent of position; the invariant measure and the whole generator decomposition are built on that form, and real minibatch gradient noise only approximates it.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive Langevin: exponential mixing with explicit rate","Thermostat noise tamed: exponential convergence proven","Explicit hypocoercivity rate for adaptive Langevin sampling","Adaptive Langevin speeds to equilibrium: quantified rate","Central limit theorem for adaptive Langevin time averages"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000181,"raw_usage":{"total_tokens":1303,"prompt_tokens":939,"completion_tokens":364,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":287}},"tokens_in":555,"tokens_out":364,"duration_ms":4163,"temperature":1.0,"reasoning_tokens":287,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:15:07.575391+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the quadratic potential $U(q)=q^2/2$, compute the spectral gap of a finite spectral discretization of $-L_{\\mathrm{AdL}}$ along a curve with $\\gamma\\varepsilon^2$ fixed and $\\varepsilon\\to0$; Theorem 2.1 requires the gap to stay at least a fixed fraction of $\\gamma\\varepsilon^2$, so a gap decaying faster than that bound would falsify the rate. Separately, replacing the additive gradient noise in a simulation by state-dependent noise of the same mean and comparing the empirical stationary distribution with $e^{-\\beta U}$ would test whether the invariant-measure assumption carries over.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides a recent Lyapunov-based exponential relaxation result for the same dynamics, whose quantitative spectral-gap scaling this paper refines."},{"cited_title":"Jones and B","cited_arxiv_id":null,"evidence_quote":"Introduced adaptive Langevin dynamics as a sampling method for systems with noisy gradient forces."},{"cited_title":"Leimkuhler and X","cited_arxiv_id":null,"evidence_quote":"Analyzes adaptive thermostats for noisy gradient systems and supplies the splitting integrators used in the numerical experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Makes the connection between thermostatted dynamics and Bayesian posterior sampling with stochastic gradients."},{"cited_title":"Welling and Y","cited_arxiv_id":null,"evidence_quote":"Establishes the stochastic-gradient Langevin baseline that motivates the noisy-gradient sampling setting."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the functional central limit theorem for Markov processes from a bounded inverse generator, used to obtain Corollary 2."},{"cited_title":"Kopec , Weak backward error analysis for Langevin process , BIT Numerical Mathematics, 55 (2015), pp","cited_arxiv_id":null,"evidence_quote":"Provides resolvent estimates for underdamped Langevin dynamics used in the large-$\\varepsilon$ expansion of the asymptotic variance."}],"review_version":1}