Technical Summary: When, Why, and How Diffusion Posterior Samplers Fail: A Finite-Sample Lens

Problem Statement

Diffusion models have become a dominant approach for solving imaging inverse problems by serving as powerful statistical priors. In Bayesian inverse problems, the goal is to sample from the posterior distribution $p(x|y)$, which combines a learned prior $p_{pr}(x)$ with a likelihood $p(y|x)$ derived from a measurement model $y = A(x) + \eta$.

While score-based diffusion methods theoretically allow for exact posterior sampling by decomposing the posterior score into a prior score and a likelihood score, the likelihood score is generally intractable. It requires evaluating a high-dimensional integral over the denoising distribution $p_{0|t}(x_0|x_t)$. To achieve computational tractability, existing "plug-and-play" or zero-shot methods (e.g., DPS [16], $\Pi$GDM [17], TMPD [18]) employ moment-matching approximations. These methods approximate the denoising distribution with either a Dirac delta (first-moment) or a Gaussian (second-moment) distribution.

Although these approximations often yield high-quality reconstructions, they suffer from sporadic and unexplained failures, such as hallucinating modes inconsistent with measurements or misweighting posterior modes. The root causes of these failures, particularly how approximation errors at intermediate timesteps propagate to the final posterior, remain poorly understood. Crucially, it is unclear whether these failures stem from nonlinear measurement models, multimodal posteriors, or the approximations themselves.

Methodology: The Finite-Sample Lens

To rigorously analyze the behavior of these samplers, the authors introduce a finite-sample perspective on posterior diffusion sampling. Instead of assuming a continuous prior distribution learned from data, they assume the prior $p_{pr}$ is a discrete empirical measure composed of $N$ i.i.d. samples $\{x^{(i)}\}_{i=1}^N$.

Key Theoretical Insights

1. Analytical Tractability: In this regime, the forward diffusion process acting on a discrete prior yields a Gaussian mixture marginal $p_t(x_t)$. Consequently, the denoising distribution $p_{0|t}(x_0|x_t)$ becomes a known discrete measure (a weighted sum of Dirac deltas).
2. Exact Likelihood and Posterior: Because the denoiser is a discrete measure, the marginal likelihood $p_{y|t}(y|x_t)$ and the posterior $p_{t|y}(x_t|y)$ can be computed analytically at any timestep $t$, even for nonlinear measurement operators $A(x)$. This provides a "ground truth" surrogate for the posterior evolution that is inaccessible in standard continuous settings.
3. Convergence: As $N \to \infty$, the finite-sample posterior converges to the true posterior at a Monte Carlo rate of $O(N^{-1/2})$. The authors demonstrate that while large $N$ is required for accuracy near $t=0$, moderate $N$ suffices to accurately characterize posterior behavior at intermediate timesteps.

Diagnostic Framework

The authors use this finite-sample regime as a diagnostic lens to compare existing moment-matching approximations against the analytically computed finite-sample posterior. They evaluate:

• Dirac Approximations (DPS): Which replace the denoiser with a point mass at the mean.
• Gaussian Approximations: Which replace the denoiser with a Gaussian, utilizing either a heuristic covariance inflation (Song et al. [17]) or the exact denoiser covariance via Tweedie's formula (Boys et al. [18]).

Key Contributions and Findings

1. Failure Modes Are Not Limited to Nonlinearity or Multimodality

Contrary to the common belief that these methods fail primarily due to nonlinear measurement operators or multimodal posteriors, the authors demonstrate that failures occur even under linear measurement models, Gaussian noise, and unimodal posteriors, provided the prior is multimodal.

2. Specific Failure Mechanisms of Popular Approximations

Using the finite-sample lens, the paper identifies distinct failure modes for different approximation strategies:

• Gaussian Approximations (Second-Moment):

  • Overestimation of Spread: These methods tend to overestimate the posterior covariance at intermediate timesteps when the prior is multimodal.
  • Consequences: This leads to "hallucination" where the sampler generates samples from prior modes that are inconsistent with the measurement (i.e., the likelihood is low for those modes). The sampler effectively behaves like unconditional generation in these regions because the prior score dominates the inaccurate likelihood score.
  • Timing: They tend to "commit" to prior modes too early in the sampling process, resulting in inaccurate relative weighting of posterior modes.

• Dirac Approximations (First-Moment, e.g., DPS):

  • Misweighting via Hyperparameter $\zeta$: The choice of the scalar hyperparameter $\zeta$ (which scales the likelihood gradient) critically determines the balance between the prior and the likelihood.
  • Consequences:
    • If $\zeta$ is too small, the sampler may fail to suppress prior modes inconsistent with the measurement.
    • If $\zeta$ is too large, the sampler may over-weight the likelihood, leading to "mode collapse" or the generation of measurement-consistent modes that are implausible under the prior.
    • Even with optimal tuning, Dirac approximations can fail to capture the correct posterior variance and may entirely miss posterior modes or hallucinate prior modes that are inconsistent with the measurement.

3. The Critical Role of Intermediate Timesteps

The paper emphasizes that the quality of the final sample ($t=0$) is often a poor indicator of the method's reliability. The distribution at intermediate timesteps is where errors propagate.

• Existing methods often explore regions of low probability under the unconditional prior $p_t(x_t)$.
• In these regions, the learned score function (trained via score matching) is not explicitly controlled, leading to unpredictable behavior.
• Errors in the intermediate posterior spread (variance) directly cause downstream failures in the final reconstruction, such as incorrect mode weighting or hallucination.

Results

The authors validate their framework through experiments on synthetic 1D and 2D problems with discrete, Gaussian, and Gaussian mixture priors, and both linear and nonlinear (quadratic, cubic, sine) forward models.

• Convergence: The finite-sample posterior converges to the true posterior as $N$ increases, validating its use as a ground truth proxy.
• Visual Evidence: Heatmaps of the posterior density across timesteps reveal that:
  • Gaussian approximations produce "fat tails" and sample from incorrect modes even when the true posterior is unimodal.
  • Dirac approximations often fail to capture the correct mean at intermediate times and exhibit incorrect variance, leading to sampling from prior atoms that the measurement should have ruled out.
• Hyperparameter Sensitivity: Tuning $\zeta$ in DPS can mitigate some errors (e.g., correcting mode weights) but cannot universally prevent hallucination or mode collapse, as the fundamental issue is the inability of a single scalar to correct the structural mismatch of the approximation across all timesteps.

Significance and Claims

The paper claims to provide the first finite-sample perspective specifically designed to understand posterior sampling (as opposed to unconditional sampling). Its significance lies in:

1. Diagnostic Capability: It offers a "drop-in" diagnostic tool that is agnostic to the type of likelihood approximation or forward model. It allows researchers to evaluate the accuracy and failure modes of existing and future samplers without needing an analytical ground truth for complex real-world problems.
2. Redefining Failure: It challenges the assumption that failures are solely due to problem complexity (nonlinearity/multimodality). Instead, it attributes failures to the inaccurate propagation of posterior spread at intermediate timesteps, even in simple linear settings with multimodal priors.
3. Practical Implications: The findings suggest that high-quality final reconstructions can mask significant errors in the uncertainty quantification (the posterior distribution) at intermediate steps. This is critical for applications requiring reliable uncertainty estimates, such as medical imaging (e.g., distinguishing a tumor from noise) or autonomous driving, where "hallucinated" modes could lead to dangerous misinterpretations.

The authors conclude that understanding the intermediate dynamics of the posterior is essential for diagnosing and improving diffusion-based inverse problem solvers, and their finite-sample lens provides a rigorous framework for this analysis.