acceptodds
Desk rejected at ICLR 2027

Making Gradient History Count: Spectrally Calibrated Monte Carlo

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Bayesian inference and uncertainty quantification rely on Monte Carlo estimates, whose accuracy can remain limited by correlated samples under a finite gradient budget. We ask how gradients already computed during sampling can reduce this error. Spectrally calibrated score repulsion Monte Carlo (SC-SRMC) uses the running average of target scores, the gradients of the log density, to adjust acceptance in the Metropolis adjusted Langevin algorithm (MALA). An exact Gaussian recursion identifies the largest score covariance eigenvalue as the scale for increasing tilt strength. A radial cap bounds the resulting tilt. Under uniform stability and regularity conditions, we bound the score history over finite horizons, with explicit terms for transition correlations. An exact projection transfers the bound to the score component of observable error. The method preserves MALA's rule for generating candidate states and requires no extra online gradients or second derivatives. At matched online gradient budgets, one configuration selected on five development targets gives lower observed mean squared error of coordinate means across 11 Gaussian and quartic validation targets, including six new geometries. The geometric mean reduction relative to diagonally preconditioned MALA is 5.65 fold at the same step size and 6.59 fold with separately tuned step sizes. The error decomposition ties parameter choice to both the observable and target geometry.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Markov chain Monte Carlo, Bayesian inference, Metropolis-adjusted Langevin algorithm, Gradient reuse, Score repulsion, Spectral calibration

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_aEKpOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Spectrally Calibrated Score Repulsion Monte Carlo (SC-SRMC), a sampling method that modifies the Metropolis–Hastings acceptance probability using a tilt derived from score information, where the score denotes the gradient of the log target density. The approach extends Score Repellent Monte Carlo (SRMC; Hu et al., 2026) by incorporating the history of past scores into future acceptance decisions. Concretely, score information is accumulated over the trajectory and used to construct a tilt vector, while the proposal distribution itself remains the standard MALA proposal. A key design choice is the construction of the tilt vector. The authors choose it to align with the accumulated score history and introduce a schedule that adaptively controls its magnitude, including radial clipping when the norm becomes too large. As a result, the tilt can be viewed as a scaled and truncated version of the running sum of past scores. Beyond the algorithmic proposal, the paper provides theoretical analysis and empirical results on unimodal target distributions. In these experiments, SC-SRMC yields lower observed mean squared error for estimating coordinate-wise means than preconditioned MALA.

Critical strengths

The present paper turns SRMC into a stable and cheap MALA modification. SC-SRMC shows lower error of the estimated means compared to preconditioned MALA.

Critical weaknesses

The scope of the empirical and theoretical evaluation appears limited. The current evaluation does not make the contribution feel fully convincing. On the theory side, the most concrete and interpretable theoretical calibration result is derived in the Gaussian setting, while the more general theory relies on fairly strong abstract assumptions and may be harder to assess in practically relevant non-Gaussian settings. On the empirical side, the strongest empirical claims are supported primarily on Gaussian/quartic unimodal targets, and the broader posterior benchmark section seems more mixed and less conclusive.

This is at odds with how the method is motivated, namely as a general-purpose sampling approach for Bayesian inference and uncertainty quantification. In its current form, the paper does not demonstrate relevance on more challenging target distributions that better reflect modern MCMC use cases. In particular, it is unclear how the method behaves on difficult unimodal distributions with strong scale heterogeneity or heavy tails, such as Neal's funnel, or on multimodal targets such as Gaussian mixtures.

As a result, the practical significance of the method remains somewhat unclear: the paper shows gains on relatively simple unimodal problems, but does not yet establish whether those gains carry over to settings where improved sampling methods are most needed.

Minor points

  • The notation π\pi is used both for the target distribution and for its density. This is potentially confusing, especially when switching between expressions such as the pointwise density π(x)\pi(x) and expectations written in the form π(f)\pi(f). Using separate notation for the measure and its density would improve clarity.

  • The formatting of lemmas, theorems, propositions, and related statements is unusual and makes the structure harder to parse. These results are presented more like inline paragraphs than clearly delimited theorem environments, which can make it difficult to identify where a statement begins and ends. Standard theorem-style formatting would improve readability.

Questions for authors
  • Can SC-SRMC be extended to heavy-tailed and multimodal targets?
  • How does this affect the scaling and clipping of the tilt?
5 comments
AuthorsOpenReview ↗

Response to W1: Broader evidence and a precise practical scope

Thank you for recognizing the value of a stable, inexpensive MALA modification and for asking us to establish its practical scope more clearly. We have prepared substantial additions that directly address the three parts of this concern: interpretable theory beyond Gaussian targets, difficult target geometries, and comparison with adaptive NUTS including adaptation costs.

The theory now separates mechanism, general conditions, and estimation guarantees. The Gaussian calculation explains a concrete design choice: a large tilt can magnify early errors, and spectral warmup controls this amplification. The general result is not obtained by assuming that a posterior is Gaussian. Its key identity is DS(u)=−Cov⁡πu(s)DS(u)=-\operatorname{Cov}_{\pi_u}(s), which describes how the mean score changes with the tilt. Theorem 1 then accounts for correlations between transitions through a Poisson equation. In accessible terms, this separates systematic feedback from correlated sampling noise.

The prepared revision makes each level of the analysis explicit:

ResultWhat it establishes
Exact Gaussian analysisThe origin of early amplification and a spectral condition preventing it
Theorem 1A finite horizon score history bound under stated covariance and kernel regularity conditions
Proposition 1Verification of those score history conditions for a non-Gaussian von Mises target using wrapped random walk Metropolis on the torus
New Corollary 1Conditional mean square consistency of ordinary sample means, including burnin, under additional residual conditions

The last result uses the finite radial cap and the existing second moment history bound. It does not require the fourth moment rate used for the sharper residual bound. We state the extra observable conditions separately so that the reader can identify exactly what is proved.

The new experiments directly test the requested geometries. We added rotated Student targets in dimension 10 with degrees of freedom 3 and 5 and scatter condition number 100. We also added unequal Gaussian mixtures in dimensions 2 and 10, with moderate or large separation and both dispersed and single mode initializations. The complete study has 48 configurations, each with 64 chains, evaluated at three gradient budgets up to 20,000. It reports tail events, mode occupancy and actual crossings between component cores as well as means. These choices test whether a mean improvement corresponds to useful exploration. The results distinguish admissibility from mixing: the Student and mixture tilts are normalizable, while the harder Student cases do not establish a stable mean advantage and widely separated mixtures still require attention to movement between modes. We address the centered funnel analytically in Q1 and explain its implications for clipping in Q2.

The practical comparison includes the cost of adaptation. We additionally evaluated 32 chains on two public data posteriors with 50,000 score calls per chain, including setup. Shared MALA and SC setup is amortized across the 32 chains, while NUTS adaptation is charged per chain. NUTS continues from its adaptation endpoint. The table gives absolute coordinate mean MSE for the fixed reference SC configuration.

PosteriorpMALASC-SRMCNUTS
Diabetes Gaussian linear posterior5.05e-65.87e-72.81e-6
Digits2.69e-22.79e-26.14e-4

Diabetes provides a concrete posterior mean benefit after accounting for setup, with approximately 8.6 times lower MSE than pMALA and 4.8 times lower MSE than NUTS. Digits identifies a geometry in which NUTS is more effective. This usual NUTS pipeline comparison was added after an initialization stress test and is reported alongside that original arm. Second moments and predictive quantities are reported separately, so the mean result is not presented as a guarantee for uncertainty quantification.

Together, these additions support a specific practical contribution: reusing already computed gradients can materially improve finite budget mean estimation on suitable targets, with explicit checks of the target, observable and mixing behavior. Section 5 and Appendix J of the prepared revision contain the full results. Thank you for prompting this more informative assessment of where the method is useful.

AuthorsOpenReview ↗

Response to W2: Separate notation for the target measure and density

Thank you for pointing out this source of ambiguity. We have implemented the distinction throughout the prepared revision so that density evaluations and expectations have visibly different notation.

ObjectNotation in the revision
Target probability measureπ\pi
Target densityp(x)p(x)
Scores(x)=∇log⁡p(x)s(x)=\nabla\log p(x)
Target expectation of an observableπ(f)=Eπ[f(X)]\pi(f)=\mathbb E_{\pi}[f(X)]
Tilted probability measureπu\pi_u

Section 2 introduces these definitions before the algorithm. The MALA acceptance probability now uses the density ratio p(y)/p(x)p(y)/p(x), while integrals and expectations use π\pi or πu\pi_u. The same convention is used in the Gaussian derivation, the compact example, and the discussion of normalizability. This also makes the distinction between the original target and a tilted target easier to follow.

We appreciate this helpful suggestion. It improves the presentation at precisely the point where the algorithm switches between evaluating a density and reasoning about its associated probability distribution.

AuthorsOpenReview ↗

Response to W3: Standard theorem environments and clearer statement boundaries

Thank you for this practical suggestion. We have replaced the paragraph style statements with standard, explicitly delimited theorem environments in the prepared revision. Lemmas, theorems, propositions, corollaries and assumptions now have distinct labels and numbering, with statement boundaries separated from the surrounding explanation.

In particular, the covariance drift lemma, the finite horizon theorem, the compact verification proposition, and the new consistency corollary are presented as separate mathematical statements. The assumptions in the proof appendix are also clearly delimited. Explanatory paragraphs then describe the meaning of the result and direct readers to the relevant proof.

We have additionally included a compact table that pairs each level of the theory with its required conditions and conclusion. This helps readers distinguish the exact Gaussian calculation from the controlled kernel theorem and the ordinary mean consistency result without reconstructing that structure from the prose.

Thank you for helping us make the mathematical contribution easier to navigate. These changes are included in Sections 3 and 4 and the proof appendices of the prepared revision.

Reviewer_pWGTOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The method at the core of the paper is interesting. It uses previous score evaluations of the target density to define an exponentially tilt of the original target, tilted to favor states for which the running average of previous scores have low inner product with the current score. Since the average score should vanish at equilibrium this tilted distribution should fluctuate around the true target. The tilted distribution enters the algorithm only through the acceptance probability to avoid an intractable (or at least expensive) proposal distribution. This is a nice idea. Unfortunately the ideas I find most compelling are already in the Hu 2026 paper cited in the current manuscript. Relative to Hu 2026, the current manuscript introduces a reasonable warmup strategy and some non-asymptotic theory (that rests heavily on it's assumptions), among other useful, but not critical (in my view), extensions. Though they do not hide the relation to Hu 2026 at all, the authors could be more clear about the value their paper adds specifically in relation to Hu 2026. Because they haven't done that, it's possible I am not giving them enough credit. In general I found the manuscript a little difficult to read and understand in places, though I don't think it's missing any essential details and the authors are very upfront about limitations. The computational examples are below the bar in my opinion.

Critical strengths

Relative to Hu 2026, there seems to be some nice explanation for how the method works (or fails) and certainly some theortical extensions.

Critical weaknesses

The ideas I find most compelling are already in the Hu 2026 paper cited in the current manuscript. Though they do not hide the relation to Hu 2026 at all, the authors could be more clear about the value their paper adds specifically in relation to Hu 2026. Because they haven't done that, it's possible I am not giving them enough credit. In general I found the manuscript a little difficult to read and understand in places, though I don't think it's missing any essential details and the authors are very upfront about limitations. The computational examples are below the bar in my opinion.

Questions for authors

Can you please provide a detailed comparison with Hu 2026 and a list of contributions specific to the new manuscript?

2 comments
AuthorsOpenReview ↗

Response to W1: Clarifying the contribution and strengthening the evaluation

Thank you for recognizing the interest of the mechanism and the theoretical extensions. We appreciate your request to make the contribution beyond Hu et al. explicit and to strengthen the evidence. We have prepared a revision organized around one concrete question: how can score feedback be calibrated and understood when the sampling budget is finite?

The contribution relative to Hu et al.

Our revision explicitly attributes the score history, exponential tilt, fixed proposal Metropolis wrapper, rational warmup form, and score projection to Hu et al. Their work also establishes asymptotic convergence results. Building on that foundation, our contribution is a finite horizon analysis and a specific calibration that can be implemented with the gradients already computed by MALA.

The Gaussian calculation makes the practical issue transparent. A stronger history correction can reduce asymptotic variance while amplifying early fluctuations. We derive the exact recursion and identify a sufficient spectral condition that prevents amplification at every individual step in that model. This is a result for exact tilted Gaussian draws. Theorem 1 separately analyzes correlated transitions, retaining the terms caused by the changing kernel. Thus the Gaussian calculation supplies an interpretable design principle, and the general bound states the additional stability and regularity conditions needed for actual transitions.

We also added Corollary 1 for the ordinary sample mean, including estimates after burnin. Under the stated score and residual Poisson conditions, the capped schedule gives conditional mean square consistency with MSE of order 1/n1/n. This connects the controlled history to the reported estimator. This supplies an explicit finite horizon connection for the present capped schedule.

Direct evidence against the closest prior algorithms.

We implemented Hu Algorithms 2 and 4 and a guardrail variant. Each method receives an independent pilot with the same score evaluation budget. Selection uses jumping distance and acceptance, without consulting the true mean. Hu variants can select their history exponent from 0.6,0.8,1\\{0.6,0.8,1\\}, while SC-SRMC retains the ordinary running average. Each evaluation uses 64 chains and 6,000 transitions.

The table reports SC-SRMC divided by the corresponding Hu method's coordinate mean MSE. Values below one favor SC-SRMC. Brackets give descriptive paired 95% bootstrap intervals.

TargetSC-SRMC / Hu Algorithm 2SC-SRMC / Hu Algorithm 4
Gaussian, condition number 1000.337 [0.244, 0.452]0.664 [0.476, 0.911]
Quartic, quadratic condition number 250.237 [0.183, 0.307]0.195 [0.145, 0.262]
Separated Gaussian mixture1.24 [1.10, 1.42]2.32 [1.66, 3.38]

These comparisons show substantial gains on the Gaussian and quartic and identify the dependence on the target. They use supplied synthetic geometry and match score counts. Algorithm 4 additionally uses Hessian vector products, whose counts and measured runtime are reported separately. Our cached score implementation uses no additional online gradients or Hessian vector products. The complete comparisons and 88 calibration ablations are in Appendices J.3 and J.4 of the prepared revision. They separate the added calibration's contribution from the complete method's gain over pMALA.

Stronger evaluation and clearer presentation.

The new appendix also supplies absolute MSE with a bias and variance decomposition, longer trajectories, an independent logistic reference, Student targets, mixtures with mode diagnostics, and two public posterior comparisons that include setup costs. This makes the evidence assessable by observable and computational budget.

We have rewritten the introduction around the contributions above, added the detailed Hu comparison in Appendix I, separated the assumptions and conclusions in a main text table, and placed mathematical statements in standard theorem environments. The algorithm is now presented as explicit steps, with target measure and density denoted separately. Thank you again for encouraging a more precise comparison. These changes make both the new theoretical results and their practical evidence substantially easier to locate and evaluate.

AuthorsOpenReview ↗

Response to the Question: Detailed comparison with Hu et al. and our specific contributions

Thank you for asking for a detailed comparison. We have added Appendix I to the prepared revision and rewritten the contribution list in the introduction. The comparison below uses the ICML 2026 version of Hu et al., with page numbers referring to that paper's own numbering.

What is inherited and what is added.

TopicHu et al. (2026)Contribution of this manuscript
History and tilted targetRunning score history and exponential tilt, Eqs. (1) and (2), p. 3These are the shared starting point. Our extension specifies how to calibrate the tilt over a finite run.
Metropolis constructionAlgorithm 2 permits a fixed proposal. Algorithm 4 uses the tilted score in its MALA proposal, pp. 14 and 15We instantiate Algorithm 2 with ordinary MALA and reuse its cached scores. This retains the base proposal and online gradient count.
Warmup and protectionRational warmup and an exponent guardrail, Appendix D.1.4We set the warmup scale using the largest score covariance eigenvalue and apply a radial cap at every step. The rational form is inherited.
Gaussian analysisCovariance driven feedback and asymptotic Gaussian mean results, p. 6 and Appendix C.10.2We derive an exact finite step recursion and a sufficient spectral condition preventing individual step amplification under exact tilted draws.
General convergence theoryAlmost sure convergence and a central limit theorem under controlled Markov noise assumptions, Theorem 3.3We provide an explicit conditional finite horizon history bound with transition correlation terms. The revision adds conditional ordinary mean consistency under additional observable assumptions.
Observable decompositionScore projection and residual covariance in the asymptotic analysis, Appendix C.10.1We use the inherited projection in exact trajectory identities, including after burnin, and bound the residual under additional conditions.

The specific contributions and why they matter.

  1. A quantitative explanation of early instability. For a Gaussian precision eigenvalue qq, the history multiplier with fixed strength α\alpha is 1−(1+αq)/n1-(1+\alpha q)/n. Its magnitude can exceed one early in the run. For the warmup schedule, the multiplier becomes 1−1/n−q/(C+n/αˉ)1-1/n-q/(C+n/\bar\alpha). We show that C≥λmax⁡(Q)C\geq\lambda_{\max}(Q) places it in [−1,1][-1,1] at every step. This explains how the most sensitive direction determines a useful warmup scale. The calculation assumes exact tilted draws, and is not presented as a general MALA stability theorem.

  2. A finite budget connection between feedback and estimation. Theorem 1 bounds the history while keeping the effect of correlated transitions explicit. Section 4 then shows exactly how history enters an observable's error. For an estimate retaining states b+1b+1 through nn, the score contribution is Bf(nθn−bθb)/(n−b)B_f(n\theta_n-b\theta_b)/(n-b). In accessible terms, this identifies the part of the estimation error that score feedback directly controls. Corollary 1 now establishes conditional mean square consistency for the ordinary estimator under the stated residual conditions, including fixed burnin and a discarded fraction tending to a value below one. This supplements Hu's asymptotic theory with bounds for the present capped schedule and estimator.

  3. An explicit calibration with direct empirical checks. The spectral scale and radial cap give a concrete implementation using the ordinary MALA proposal. The acceptance calculation reuses cached scores and adds no online gradient or Hessian vector evaluation. We now test this choice directly against Hu Algorithms 2 and 4 and the guardrail variant, alongside an 88 cell component study. The comparison allows Hu's history exponent to vary and selects parameters using independent pilots. Synthetic geometry is supplied equally and score budgets are matched, with additional Hessian vector costs and runtime reported separately. This protocol evaluates the incremental calibration, whose benefit depends on the target, separately from the original comparison with pMALA.

Thank you again for prompting this clarification. The revised positioning is a specific extension of SRMC: an interpretable finite horizon calibration, quantitative analysis of the resulting history and estimator, and direct evidence on where the calibration helps. The inherited framework and the new results are now distinguished at the point where each is introduced.

Reviewer_BjTROpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SC-SRMC, a Metropolis-adjusted Langevin algorithm (MALA) variant that reuses the running average of target scores to adjust the acceptance rule. The target is exponentially tilted by e−u⊤s(x)e^{-u^\top s(x)}, with the tilt uu set from the score history and calibrated by the largest eigenvalue of the score covariance; a radial cap bounds the tilt. The paper provides an exact Gaussian recursion, a finite-horizon bound on the score history under uniform stability and regularity conditions, and an exact projection that separates an observable’s error into a score component and a residual. Experiments use 11 Gaussian and quartic validation targets (five development, six new) and 15 posterior benchmarks, comparing against diagonally/densely preconditioned MALA, HMC, and NUTS. The headline result is a 5.65× geometric-mean reduction in coordinate-mean MSE relative to diagonal pMALA at the same step size, and 6.59× with separately tuned step sizes.

Critical strengths
  1. A clear mechanistic idea. Reusing cached scores to adjust acceptance, rather than adding gradients, is an appealing way to exploit information that MALA already computes.
  2. The exact Gaussian analysis. The recursion in Eq. (13) and the spectral warmup condition in Eq. (15) give a clean explanation of why strong tilts can amplify early errors, and the warmup is designed to prevent that. This is the most convincing part of the paper.
  3. The finite-horizon theory. The bound in Theorem 1 explicitly tracks the tilt strength, the score covariance, and the transition correlations; it is a nontrivial extension of stochastic-approximation and Poisson-equation arguments to a time-varying tilted kernel.
Critical weaknesses

1.The estimator targets a tilted distribution, and the bias is not quantified. For a fixed tilt uu, the chain preserves πu\pi_u, not π\pi; the plain sample mean is therefore biased at finite nn. The method relies on un→0u_n \to 0 for consistency, but the paper does not state or prove a consistency theorem for the sample mean under the adaptive schedule, nor does it report a bias–variance decomposition. The abstract’s “variance reduction” is misleading: this is a bias–variance tradeoff, and the finite-sample MSE gains may come from bias. The authors should report bias, variance, and MSE separately, and compare against an importance-weighted estimator that corrects for the tilt. 2. The theory does not establish an MSE improvement. Theorem 1 bounds the score history θn\theta_n, not the estimation error; the residual rf,nr_{f,n} is controlled only by a separate bound (Eq. 19) under additional assumptions. The constants depend on the tilted kernel, and the paper itself notes that the bound “does not imply that increasing α\alpha always reduces the history error.” There is no theorem showing that SC-SRMC has lower MSE than MALA, or even that its MSE decreases with α\alpha. The gap between the theoretical contribution and the empirical claim is large. 3. The largest empirical gains are surprising and may be reference-dependent. The reference configuration gives MSE ratios of 0.116 (quartic) and 0.045 (synthetic logistic) relative to pMALA (Appendix Table 2), i.e., 9–22× reductions from a modest change in acceptance. The logistic target’s mean is estimated by importance sampling, not analytically; the paper’s own Appendix Tables 11–12 discuss the sensitivity of the reference. Without absolute MSE values and an independent reference, it is hard to tell whether these ratios reflect a real gain or a poorly tuned pMALA/reference. The paper should report absolute MSE, effective sample size, and autocorrelation, and validate the logistic reference with an independent method.

Questions for authors

see weaknesses

5 comments
AuthorsOpenReview ↗

Response to W1: Consistency, bias and variance, and importance weighting

Thank you for identifying the three distinctions that matter here: consistency of the ordinary mean, its finite sample bias, and correction of a fixed tilt. We have addressed each in the prepared revision, with a new consistency result and direct experimental controls.

Consistency of the estimator actually used. Corollary 1 in Section 4 concerns the ordinary mean computed after discarding initial samples. Conditional on the preliminary calibration output, it assumes Theorem 1's uniform covariance and score Poisson conditions, the additional residual regularity and Poisson conditions in Appendix B.1, and finite second moments for the initial prefix. For the implemented capped schedule and deterministic burnin bnb_n with bn/n→β<1b_n/n\to\beta<1, it proves

E[(f^bn:n−π(f))2∣G]≤Kf2(n+bn)2(n−bn)2=O(n−1).\mathbb E[(\widehat f_{b_n:n}-\pi(f))^2\mid\mathcal G]\le K_f^2\frac{(\sqrt n+\sqrt{b_n})^2}{(n-b_n)^2}=O(n^{-1}).

Here G\mathcal G fixes the preliminary calibration and KfK_f is finite under these assumptions. Thus the actual sample mean converges in conditional mean square, including fixed burnin and the quarter trajectory burnin used in the experiments. The score history continues through discarded samples, as implemented. Appendix B.3 gives the full proof. The finite cap lets this result use the existing second moment history bound without an additional fourth moment rate. This explicitly controls estimation error, rather than relying on the statement that the tilt tends to zero.

Measured bias and variance. We added 64 independently initialized replicate chains per method and target, evaluated at 6,000, 12,000 and 24,000 transitions. The following terminal results report coordinate averages. “Ref” and “Sel” denote the reference and selected SC-SRMC settings, respectively.

TargetMethodMSEVarianceCorrected squared bias
GaussianpMALA2.524e-32.496e-32.753e-5
GaussianSC Ref4.399e-44.365e-43.338e-6
GaussianSC Sel3.245e-43.221e-42.459e-6
QuarticpMALA4.649e-44.693e-4-4.431e-6
QuarticSC Ref4.127e-54.170e-5-4.234e-7
QuarticSC Sel2.170e-52.176e-5-5.979e-8
LogisticpMALA1.971e-51.981e-5-1.044e-7
LogisticSC Ref5.051e-72.240e-72.811e-7
LogisticSC Sel9.818e-71.348e-78.470e-7

The correction subtracts variance divided by 64 from the squared error of the average replicate estimate, giving the exact empirical identity MSE = variance + corrected squared bias. Negative estimates reflect finite replication, not negative population bias. Logistic uses the independent NUTS reference, with reference uncertainty analyzed separately in Appendix J.2.

The Gaussian and quartic gains are predominantly variance reductions. Logistic illustrates the tradeoff explicitly: selected SC has lower variance than reference SC but higher bias and MSE, while both have much lower mean MSE than pMALA. The revised abstract and Section 5 accordingly describe finite budget mean MSE gains, rather than an unconditional variance reduction claim.

Importance weighting control. We also ran 64 chains for each of two frozen tilt magnitudes on all three targets. For fixed uu, the density ratio is proportional to wu(x)=exp⁡(u⊤s(x))w_u(x)=\exp(u^\top s(x)), so we compare the plain mean and self normalized importance sampling (SNIS) on the same path. At the larger standardized tilt magnitude, the 24,000 transition results are:

TargetSNIS MSE / plain MSE
Gaussian0.799
Quartic0.0658
Logistic0.0678

Appendix J.5 reports both magnitudes and weight diagnostics. This is a valid fixed tilt correction with finite sample ratio bias. The evolving adaptive path is not distributed according to a single fixed tilted law, so these weights are not claimed to correct that path exactly. Thank you for prompting a clearer separation of the consistency theorem, finite sample tradeoff, and weighting control.

AuthorsOpenReview ↗

Response to W2: Connecting score history to observable MSE

Thank you for carefully distinguishing a bound on score history from a bound on the observable being estimated. Your comment helped us make this connection explicit and separate what is proved from what is measured. The prepared revision now contains an ordinary mean consistency corollary, a table of the assumptions behind each guarantee, and empirical MSE comparisons with uncertainty.

Why score control can improve a mean estimate. The score is the gradient of the log density. Its target expectation is zero, so a nonzero running average of scores records a particular component of sampling error. Equation (17) separates an observable's estimation error exactly into the part explained by this score average and the remaining part:

fˉn−π(f)=Bfθn+rˉf,n.\bar f_n-\pi(f)=B_f\theta_n+\bar r_{f,n}.

The coefficient BfB_f is the population linear regression coefficient of the observable on the score. This identity holds on every realized trajectory. It explains which component the acceptance adjustment acts on and why the observable matters.

For a Gaussian target with precision matrix QQ, this connection is especially strong: the coordinate mean has no residual, and Xˉn−μ=−Q−1θn\bar X_n-\mu=-Q^{-1}\theta_n. In the exact tilted draw model, the recursion therefore gives an exact mean error analysis. Along an eigendirection with precision qq, Equation (14) proves

n,E(Xˉn,q−μq)2⟶1q(1+2αq).n\\,\mathbb E(\bar X_{n,q}-\mu_q)^2\longrightarrow\frac{1}{q(1+2\alpha q)}.

The corresponding value for independent target draws is 1/q1/q. Thus the model proves an asymptotic reduction by the factor 1/(1+2αq)1/(1+2\alpha q), while the finite step recursion identifies the early amplification that spectral warmup prevents. This statement concerns exact tilted draws, not a comparison theorem for the MALA transition kernel.

The new observable guarantee. Corollary 1 now proves conditional L2L^2 consistency of the ordinary mean with MSE O(n−1)O(n^{-1}), including fixed burnin or a deterministic discarded fraction below one. It combines Theorem 1 with the explicitly stated residual Poisson and regularity conditions in Appendix B.1. The proof uses the actual finite radial cap and does not require the extra fourth moment rate needed for the sharper residual bound in Equation (19). Appendix B.3 supplies the derivation and handles the finite initial steps. Table 1 presents these assumptions alongside their conclusions.

This closes the missing implication from history control to consistency under stated conditions. The comparison constants still depend on the transition family, so the general theorem is not presented as proving dominance over MALA or monotone improvement with tilt strength. The empirical claim is specifically about finite budget coordinate mean MSE.

Directly measuring that empirical claim. The new experiment uses 64 chains per configuration with paired randomized starts and matched online score counts. At 24,000 transitions, the selected SC-SRMC configuration gives:

TargetSC MSE / pMALA MSEPaired 95% interval
Rotated Gaussian0.129[0.111, 0.149]
Quartic0.0467[0.0365, 0.0597]
Synthetic logistic0.0498[0.0439, 0.0567]

These are descriptive whole chain bootstrap intervals conditional on the preliminary runs and reference. Gaussian and quartic means are analytic. Logistic uses an independent NUTS reference with a separate uncertainty analysis in Appendix J.2. Section 5 and Table 2 also report absolute MSE and its bias and variance decomposition, so the gain is evaluated against the target expectation rather than inferred from a small score average or high acceptance rate.

We appreciate this suggestion because it sharpens the paper's contribution: exact analysis explains the Gaussian mechanism, controlled kernel theory establishes history bounds, the new corollary connects these bounds to ordinary mean consistency under additional assumptions, and replicated experiments quantify the finite budget advantage for the reported means.

AuthorsOpenReview ↗

Response to W3: Independent reference validation, absolute errors and diagnostics

Thank you for requesting an independent check of the strongest empirical gains. We reconstructed the identical synthetic logistic target, verified its stored data hash, and computed a separate dense metric NUTS reference. Recomputing the original trajectories against this reference preserves the large reduction in coordinate mean MSE.

Independent reference for the original claim. The reference uses four chains with 1,200 adaptation steps per chain and a cumulative integration budget of 480,000 per chain. Its diagnostics are:

DiagnosticValue
Maximum rank R^\widehat R1.0000473
Minimum bulk ESS500,938
Minimum tail ESS211,672
Evaluation divergences0
Mean coordinate MCSE squared2.92945e-8

The MCSE squared is below 1% of the smallest original SC MSE at 6,000 transitions, meeting the precision threshold specified for this check. Reference computation is accounted for separately from the evaluated sampler budgets. “Ref” and “Sel” below denote the reference and selected SC-SRMC settings.

MethodOriginal IS reference MSEIndependent NUTS reference MSE
pMALA6.887e-56.834e-5
SC Ref3.078e-62.997e-6
SC Sel3.452e-63.189e-6

These are the same archived 24-chain trajectories at 6,000 transitions, with only the reference changed. For reference SC, the new MSE ratio to pMALA is 0.04386 with paired 95% interval [0.02988, 0.06357]. A sensitivity analysis that perturbs the shared reference for all methods together gives [0.03006, 0.06437]. Thus the roughly 22 fold gain survives a different reference method and explicit allowance for its estimated Monte Carlo uncertainty. These diagnostics quantify Monte Carlo precision rather than bound unknown systematic reference bias.

Absolute errors in new, longer runs. A separate experiment uses 64 independently initialized chains per method, at 6,000, 12,000 and 24,000 transitions. Terminal absolute coordinate mean MSE is:

TargetpMALASC RefSC Sel
Rotated Gaussian2.524e-34.399e-43.245e-4
Quartic4.649e-44.127e-52.170e-5
Synthetic logistic1.971e-55.051e-79.818e-7

The Gaussian and quartic means are analytic. For logistic, the primary reference MCSE squared is about 5.8% of the smaller 24,000 transition SC MSE, so the earlier 1% precision statement is restricted to the archived 6,000 transition comparison. Appendix J.2 reports a longer NUTS continuation as a separate sensitivity analysis, with MCSE squared 7.26e-9, about 1.36% of the corresponding reference SC MSE. It retains both SC configurations and the same qualitative ordering.

ESS, autocorrelation and baseline calibration. Appendix J.1 now includes rank diagnostics, bulk and tail ESS, and autocorrelation for all three targets. For the first logistic coordinate, using all 64 chains with 18,000 retained states each:

MethodBulk ESSTail ESSLag 10 autocorrelation
pMALA173,938300,4750.0483
SC Ref204,164330,590-0.00421
SC Sel210,409342,486-0.0286

For adaptive SC these are descriptive diagnostics, while replicated error against the reference remains the accuracy measure. Their improvement need not equal the MSE gain, because controlling accumulated scores can change the error of a mean estimator without an equally large change in a local autocorrelation statistic.

The original 11 target study also distinguishes a common proposal step comparison from separate step selection for both methods. The geometric mean MSE reductions are 5.65 fold and 6.59 fold, respectively, with setup and selection costs recorded in Appendix G.5. These results evaluate the complete algorithm relative to pMALA. Together with the absolute errors and independent reference check, they make the gain directly inspectable. Thank you for motivating this stronger validation.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.