acceptodds
Under review as a conference paper at ICLR 2027

Improving score-based sampling via affine post-processing

Abstract

Sampling from a probability density with access only to its log-likelihood is a foundational problem in statistics, machine learning, and the physical sciences. This problem is studied in the data-free score-based diffusion literature, which samples the target density via estimating scores of densities produced by convolution with normal random variables and implementing a reverse SDE. In this framework, score estimators are constructed via Markov chain Monte Carlo, importance sampling, or rejection sampling. The estimators often operate far beyond the regime where mixing time and error control can be established and have large bias and variance when estimating the underlying score. In our method we propose post-processing: splitting a fixed computation budget across multiple queries, querying the estimator at different points, and fitting a single pooled estimate via linear combination to minimize bias and variance at a target location. The final value is an affine map of a vector of multiple score estimates. We show this reduces score estimation error when compared to the computation-matched base estimator queried at a single point. We demonstrate improved sample quality over other score estimation methods matched for computation cost on difficult target densities such as log-concave densities with large condition number, multimodal densities, and high-dimensional densities.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Sampling, Diffusion Models, Markov Chain Monte Carlo

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_9LDwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies score estimation via Monte Carlo when the target distribution is specified by an unnormalized density e−f(x)e^{-f(x)}, so that the score is estimated by sampling rather than by training a neural network. The authors propose an affine post-processing rule that combines score estimates at multiple signal levels aia_i to reduce the bias and variance of the base estimator. Experiments on Gaussian, Gaussian-mixture, and several other targets are conducted to verify the effectiveness of the proposed method.

Critical strengths
  1. Since the score s(u,a)s(u,a) is a smooth function with respect to the signal level aa, it is reasonable to use score estimation at different levels to construct a more accurate estimator.
  2. The optimal affine post-processor is analyzed theoretically, and an interesting observation is given: any unbiased estimator with non-zero variance is not optimal among affine post-processors.
Critical weaknesses
  1. Equation (24) approximates the score as ∑kϕk(a)βk(u)\sum_k \phi_k(a)\beta_k(u). The authors should explain more explicitly why such a decomposition exists. I understand that s(u,a)s(u,a) is a smooth function of aa, but it is not immediately clear why it can be expanded in a basis {ϕk(a)}\{\phi_k(a)\} independent of uu, considering that s(u,a)s(u,a) is a vector. A brief justification would help.

  2. The paper uses a fixed set of design levels aa, but the main text does not discuss how they are chosen or how sensitive the method is to this choice. The authors should clarify how to choose them in the main text and discuss the impact of this choice.

  3. The paper provides a closed-form Wiener oracle and shows that unbiased estimation is not optimal in general. However, the practical Legendre ridge-regression post-processor is not accompanied by a theoretical guarantee that it reduces variance or MSE, nor by a finite-sample error bound. The current theory applies to the oracle rule, not to the practical one. The authors should provide such guarantees or related analyses clearly.

  4. The experimental targets include many artificial distributions such as banana, ellipsoidal shell, and Gaussian mixtures. However, all targets are synthetic. It would be better to include at least one real-world application, such as a molecular energy function, to demonstrate practicality in high-dimensional settings.

  5. Theorems 1 and 2 and Corollaries 1 and 2 would benefit from additional discussion around them.

Questions for authors

see weakness.

Reviewer_xq85OpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

Diffusion based samplers provide benefits when sampling from unnormalised probability distributions by targeting intermediate probability distributions instead of the final target directly. However, they require estimating the score which leveraging Tweddies formula is provided by an expectation wrt the posterior distribution. A lot of available methods incur in large bias and variance when estimating the score. This paper proposes a meta algorithm that can be used on top of a diffusion sampler to reduce the bias and variance of the expectation. For a fixed position they propose to split the budget and obtain estimators across different noise levels and use all of them to fit a kernel ridge regression. They study an oracle setting which requires access to the score as well as a practical implementation by using Legendre polynomials. They finally show the effectiveness of the method, showing how it improves over the base sampler.

Critical strengths
  • Multiple noise levels: I think the idea of leveraging multiple noise levels is really interesting, specially, because for lower noise levels the posterior distribution becomes increasingly strongly log-concave and therefore the expectation is easier to estimate. Therefore, as the authors mention instead of spending all the budget for a single noise levels which might lead to a minimal increase in the ESS, leveraging different levels we can reduce the bias and variance substantially.

  • Flexibility: I really like the fact that the method is flexible and can be added on top of your favourite diffusion sampler.

  • Extensive experiments: the experiments cover a wide range of settings where diffusion samplers typically struggle and show the benefits of the proposed method. Also the experimental analysis of when the method stops providing substantial improvements , such as when the mode separation of two anysotropic Gaussians is very large, is very insightful.

Critical weaknesses
  • Extra computations. While the method allows for bias and variance reduction it requires extra computations to fit the regression coefficients. A more detailed discussion of the extra cost would be helpful.

  • Although the method is well grounded, there are no theoretical analysis. I appreciate the authors discussing this in the last section. I understand that this is outside the scope of the paper, but will definitely make the paper more complete.

Questions for authors
  1. [Clarification] Could you provide some more details regarding the extra computational cost of the method?
  2. [Clarification] Would it be possible to provide a mathematical intuition/explanation of the cases were the method improves upon the base estimator and when it does not?
Reviewer_hCJdOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies sampling from an unnormalized density via score-based diffusion, where the noised score must be estimated by Monte Carlo. Instead of spending the whole estimation budget on a single pointwise estimate at the required noise level, the authors propose to split it across several noise levels and combine the resulting estimates, together with the two exactly known endpoint scores, through a fixed affine map. They derive the optimal weights in closed form, show that pure on-policy estimation is optimal only under a condition that unbiased estimators with nonzero variance violate, and give a practical rule based on constrained ridge regression in a Legendre basis, which they show is equivalent to a small quadratic program on the weights. The method is evaluated at matched compute on synthetic targets, including log-concave densities with large condition number, multimodal densities, and high-dimensional densities.

Critical strengths

The core idea is simple, general and useful. It works on any score estimator and costs almost nothing extra and The empirical gains are substantial

Critical weaknesses

The theory does not rigorously prove the main claim that splitting a fixed budget across several signal levels and pooling the estimates yields lower score error than spending the whole budget on a single on-policy estimate.

Questions for authors

-Q1 I'm not sure the degree-5 fit can handle the Gaussian case at large κ\kappa. The exact score curve has a sharp bend of width about 1/κ1/\sqrt\kappa near a=0a=0, and six nodes spaced uniformly in aa won't see it. Figure 2 only covers κ=10\kappa=10. Can you show it for κ=100\kappa=100 and 10001000?

-Q2 Both IS and RDMC are biased in practice, but the derivation assumes they aren't. How much of that bias gets passed through the weights?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.