acceptodds
Under review as a conference paper at ICLR 2027

Diffusion-Based Posterior Sampling: A Feynman-Kac Analysis of Bias and Stability

Abstract

Diffusion-based posterior samplers draw from measurement- or reward-conditioned posteriors using pretrained diffusion priors. Though widely used for inverse problems, their output is biased even with exact prior scores. First, we quantify this bias for the DPS algorithm, the base method for a family of samplers, by introducing a surrogate curve of probability densities which interpolates between the true posterior and the standard Gaussian. The ratio of sampler to surrogate densities solves a parabolic PDE, and its Feynman–Kac formula yields an explicit representation of this ratio as the expectation of a path functional. This representation reveals which posterior regions are over- or under-sampled. This framework also provides a design principle for bias reduction that includes STSL, a related sampler. Second, in hard-constraint problems (e.g. the low measurement noise regime) these samplers are known to be numerically unstable. We find that this instability of discretized DPS arises from an effective guidance update whose strength scales inversely with step size, inevitably violating the forward-Euler stability criterion near the constraint. This finding motivates implicit (backward-Euler) guidance updates. Finally, we show that the standard practice of early guidance-stopping, used to avoid oscillations, introduces a secondary bias which we characterize through our framework.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
diffusion posterior sampling, training-free guidance, Feynman-Kac, sampling bias, score-based generative models, numerical stability, inverse problems

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_r9uTOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors consider Posterior Diffusion Sampling (PDS) and aim to quantify the effects of approximating the intractable conditional score when using this method to solve inverse problems. Using the Feynman-Kac formula, the authors analyze the bias introduced by the DPS approximation to understand which samples are under- or over-sampled. They also quantify the effects of heuristics such as early guidance-stopping and trajectory-dependent step size, which practitioners use to prevent instability in the numerical implementation of DPS.

Critical strengths

I think the paper's objective of quantifying the limitations of the heuristics behind DPS is important and deserves attention. An interesting result is the precise quantification of the discrepancy between the DPS-SDE and the true posterior via the Keynman-Kac formula (Theorem 1), which explains when DPS over- or under-samples the posterior and points to principled ways of designing paths with smaller bias (Theorem 2).

Critical weaknesses

I believe there is an issue with Theorem 2. Its proof uses Lemma 3 and 4: the former requires hh to be smooth, while the latter uses D2RyD^2 R_y. However, in Theorem 2, the reward Ry(x)=−∥A(x)−y∥2R_y (x) = -\|A(x)-y\|_2, which is not sufficiently smooth. If the authors can solve this issue, or convincingly explain why it does not arise, and the other reviewers have not identified any other technical issues with the paper, I'm willing to raise my scores.

As minor points regarding the presentation, I believe Theorem 2 deserves a more extensive discussion, which is currently missing. Also, the paper does not have a conclusion section.

Questions for authors

See weaknesses.

2 comments

Thank you for the careful reading and for identifying this gap. You are right: the smoothness assumptions of Lemmas 3 and 4 do not directly cover the norm reward. We have revised Theorem 2 to include a Feynman–Kac formula accounting for its nonsmoothness.

1. Smoothness of the reward (Lemmas 3 and 4).

The reward is not C2C^2: the Hessian of the norm is singular on the constraint, so the lemmas do not apply directly. In the revised Theorem 2, the reaction coefficient uses the absolutely continuous part of the Hessian, while an additional Stieltjes integral accounts for the singular contribution:

ζ∫0T−t∗ηt∗+s,dΛs.\zeta \int_{0}^{T-t_{\ast}} \eta_{t_{\ast}+s}\\,d\Lambda_{s}.

Here ζ\zeta is the DPS step-size hyperparameter, ηt\eta_t is the reciprocal of the discretization step size, and Λ\Lambda is a continuous nondecreasing process arising from the generalized Itô formula for the norm of the Tweedie residual along the OU path. For a scalar measurement, Λ\Lambda is the local time of the residual at zero, as in Tanaka’s formula:

∣rτ∣=∣r0∣+∫0τsgn⁡(rs),drs+Λτ.\lvert r_\tau\rvert = \lvert r_0\rvert + \int_0^\tau\operatorname{sgn}(r_s)\\,dr_s + \Lambda_\tau.

The additional contribution is a Stieltjes integral against Λ\Lambda, so the weight is well defined. The appendix proves the formula by smoothing the norm, applying Lemmas 3 and 4 to the smoothed reward, and passing to the limit. After guidance stops, the remaining step uses only the OU semigroup and does not require a C2C^2 reward. Under the other stated assumptions, the revised theorem covers any C2C^2 measurement operator A:Rd→RL\mathcal{A}:\mathbb{R}^d\to\mathbb{R}^L whose Jacobian has full rank on the constraint manifold.

2. Discussion of Theorem 2.

The revised discussion interprets the formula in two stages. When guidance stops, the Feynman–Kac weight characterizes the bias accumulated relative to the annealed surrogate. Generation then continues without guidance, smoothing both the reward tilt and this bias. The stopping time and guidance strength can therefore be tuned to approximate a reward tilt at a chosen temperature, while the formula characterizes the remaining bias. We also connect this interpretation to Huang et al., who use early guidance stopping to recover variance comparable to the reward tilt in a Gaussian example.

3. Conclusion.

We added a Conclusion section summarizing the Feynman–Kac characterization of DPS bias, ELG as an alternative tilt with a simpler reaction term and empirical improvements (Section 4), and the tradeoff introduced by early guidance stopping (Theorem 2). It also identifies a direction for future work: evaluating ELG at scale, where the conditional covariance Σt\Sigma_t must be estimated from the learned denoiser.

The authors have solved my concern about the proof of Theorem 2. I raised my score accordingly.

Reviewer_LpY8OpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper focusses on guidance in score-based diffusion models using Diffusion-based Posterior Sampling (DPS). It is known that DPS is a coarse approximation to the true guidance term, and therefore diffusion models guided with DPS do not converge to the true posterior/tilted distribution. This paper analyses the difference between what DPS converges to and the true posterior.

Given a prior distribution ρt\rho_t and a function hth_t the normalised the density ρtht/Zt\rho_t h_t/Z_t satisfies a PDE including a reaction term which accounts for hh being non-harmonic. This would correspond to an SDE with a killing rate. At the start and end times ρtht/Zt\rho_t h_t/Z_t has the correct density i.e. the Gaussian and the posterior/reward tilted density that one wants to sample from. If instead you take the SDE without the killing rate, this converges to another density which isn’t the correct one. Effectively, this is what the DPS algorithm does, and using this setup the paper shows the difference between the true posterior and the DPS approximate posterior is given by the reaction term.

Critical strengths

Overall, I like this paper. I think the idea is interesting and I learned something reading it. I found the analysis of the difference between the true distribution and the distribution to which DPS converges (and indeed a more general class of approximate guidance terms) useful and novel. For the most part, I think the paper is well-written and mostly clear; though there are some places where I think it could be slightly improved. The paper seems polished and I didn’t spot any mistakes, in both the maths and the spelling.

Critical weaknesses
  1. Conclusions: The paper provides a nice analysis of some existing DPS type guidance algorithms and the bias that arises. However, no concrete suggestions of how to use this to improve DPS type algorithms are given. The paper does suggest directions i.e. by choosing paths with a smaller reaction term via a drift term ∇U\nabla U, and shows that the existing STSL provides paths with smaller reaction terms. However, no new suggestions are provided i.e. no suggestions of what UU should be.
  2. Presentation: As mentioned above, overall I found the presentation decent. However, I think in some places in the text more explanation could help. Personally, I found Section 3: the introduction of the surrogate path, algorithm path and algorithm SDE quite confusing and needed to read it a couple of times. I think adding more motivation for why the algorithm path is introduced would help.
  3. Theorem 2: There’s no discussion of Theorem 2 afterwards. Therefore, it is not fully clear what the consequences are or how to interpret it. For example, could it be used to inform a choice of early stopping time? More discussion on this would be useful.
  4. Experiments: I liked the examples included. However, there are no real experiments. In principle, I find this fine as the main contribution is the analysis which I find reasonably complete without further experiments. However, e.g. experiments showing that reducing the reaction term leads to better samples could strengthen the paper.
  5. Hard constraints: In general the paper considers rewards of form exp⁡(R(x))\exp(R(x)) and not hard-constraints. However, hard constraints are mentioned in the abstract. I do not think this method is applicable in this case, since, for example two different Dirac distributions will be mutually singular, and therefore no Radon-Nikodym derivative will exist. I think this should be clarified.



Summary I enjoyed this paper, and did not find any major weaknesses. Although I think addressing the aforementioned weaknesses would improve the paper, I do not think they need to be addressed in order for the paper to be published. Therefore I give it a weak accept score. If the authors can clarify the hard constraints and add discussion of Theorem 2, I would be willing to increase.

Questions for authors

Could the authors provide suggestions on how to choose UU as in section 4?

How should Theorem 2 be interpreted?

2 comments

Thank you for the constructive review and precise questions.

1. Concrete suggestions (Weakness 1, first question).

Section 4 now places expected log-likelihood guidance (ELG), htELG=exp⁡(E[Ry(X0)∣Xt])h_t^{\mathrm{ELG}}=\exp(\mathbb{E}[R_y(X_0)\mid X_t]), between the exact canonical tilt and DPS. ELG retains the conditional fluctuations that DPS discards by evaluating the reward at the conditional mean. For quadratic rewards, ELG is exactly computable from the conditional mean and covariance, without assuming a Gaussian conditional law.

Relative to DPS, a tilt hth_t induces the correction potential ϑUh=2(Ry(x^t)−log⁡ht)\vartheta U_h=2(R_y(\hat{x}_t)-\log h_t). For a linear-Gaussian reward, the ELG potential is proportional to tr⁡(AΣtA⊤)\operatorname{tr}(A\Sigma_tA^\top), the total conditional variance of the measurement AX0AX_0. When A⊤A=cIA^\top A=cI, as in denoising, this reduces to the STSL potential tr⁡Σt\operatorname{tr}\Sigma_t, with a time scaling prescribed by the surrogate path. Implemented STSL uses a tuned constant coefficient instead; for general AA, its isotropic penalty also differs from ELG’s measurement-weighted correction.

ELG’s reaction term has no Hessian contribution, suggesting a possible reduction in bias. We do not prove such a reduction, but observe improved agreement with the posterior in our two-dimensional experiment.

2. Experiments (Weakness 4).

A new appendix section evaluates ELG on the problem in Figure 2. The variance of the Feynman–Kac weights decomposes into the χ2\chi^2-divergence, which measures sampling bias, and an additional contribution from path noise. In this experiment, ELG reduces both: the χ2\chi^2-divergence falls from 6868 to 0.860.86, and the effective sample size increases from 0.0140.014\\% to 4242\\%. Reweighting ELG samples recovers the posterior, while unweighted ELG remains biased. These results demonstrate bias reduction in a single two-dimensional example with exact conditional moments.

3. Theorem 2 (Weakness 3, second question).

We added a discussion and extended the theorem to general C2C^2 measurement operators A:Rd→RL\mathcal{A}:\mathbb{R}^d\to\mathbb{R}^L, subject to the stated rank condition. The discussion interprets generation in two stages: the Feynman–Kac weight characterizes the bias accumulated when guidance stops, and subsequent guidance-free generation smooths both the reward tilt and this bias. Earlier stopping also bounds the additional nonnegative weight contribution associated with paths that linger near the constraint. The theorem makes this tradeoff explicit without prescribing a stopping time; the stopping time and guidance strength can be tuned in practice to approximate a reward tilt at a desired temperature.

4. Hard constraints (Weakness 5).

We agree that, for a genuinely hard-constrained posterior, the two laws are mutually singular. Our results concern large but finite constraint strengths. We show that the DPS discretization introduces an implicit annealing factor, driving the sampler toward a low-temperature target with a constraint inverse temperature of order 10410^4 near the end of sampling. Our contribution is to identify this factor and its consequences, rather than to advocate DPS for hard constraints. Such strong constraint enforcement may be undesirable in image applications, where practitioners often turn guidance off entirely near the data manifold.

I thank the authors for their response and additions to the paper. This clears up my concerns, therefore I increase my score to an accept.

Reviewer_LNfAOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper analyzes the bias of the diffusion posterior sampling (DPS) algorithm through a comparison with the reversed score-based diffusion. An novel integral formulation is then given for this bias based on a Feynman-Kac analysis, comparing the reverse SDE for DPS with a "surrogate path" that exactly samples the posterior at some fixed T>0T>0. This is then related to the STSL algorithm of Rout et al., interpreting an additional drift as a change of surrogate trajectory.

Critical strengths

None

Critical weaknesses
  • The analysis does not seem to hold if the initial distribution is supported on a low dimensional manifold. The authors argue that oscillations occur due to a violation of the "forward Euler stability criterion", however [P22] offers a much simpler and complementary explanation: the score simply explodes as t goes to 0, and would be a scaled projection [LSHH26]. In view of this, it is also unclear as to why the oscillation will be tangent to the manifold. Is it because of the tilting term? This needs some careful discussion and Figure 3 does not support this claim, affects contribution (ii).
  • Appendix C.2.1: Eq. (36) has an error caused by the preceding O(∥v∥2)O(\|v\|^2) term, the expansion is missing the ⟨x−p,correction⟩\langle x-p, \text{correction} \rangle which is O(∥v∥2)O(\|v\|^2). As such, Eq. (40) is missing a curvature term, and the authors can check that it is incorrect using the unit sphere. See [LSHH26] for what it should look like. This might cause problems in Section 5.

[P22] Pidstrigach, J. (2022). Score-based generative models detect manifolds. Advances in Neural Information Processing Systems, 35, 35852-35865.

[LSHH26] Li, X., Shen, Z., Hsieh, Y. P., & He, N. (2026, April). When scores learn geometry: Rate separations under the manifold hypothesis. In International Conference on Learning Representations (Vol. 2026, pp. 40627-40660).

Questions for authors
  • How much of the bias actually comes from the inexactness of the DPS starting from Gaussians rather than the true distribution + noise?
  • How difficult is it to compute ω\omega in Theorem 1? How do you compute Σt\Sigma_t?
6 comments

Thank you for your review and supporting references. We have corrected our short-time expansion of ∇x^t(x)\nabla\hat{x}_t(x). This correction does not affect our conclusion, since the additional curvature term vanishes as xx approaches the manifold. We have also added an example and further explanation to clarify the instability of DPS. As detailed below, the oscillations we identify arise from the guidance term; the singularity of the score as t→0t\to 0 is a separate effect.

Weakness 1: Oscillations

The instability studied in Section 5 originates from the guidance term, rather than from the divergence of the score as t→0t\to 0 [P22].

To isolate this effect, we added Figure 8, “On a circle, DPS oscillates in the tangential direction,” to the appendix section “Empirical Evidence of Instabilities.” This example uses data uniformly distributed on a circle in the plane, for which the score is available analytically. Unguided denoising converges to the circle without oscillations, whereas guided DPS oscillates tangentially. For the circle, this tangential motion is captured directly by the angular coordinate.

Our MNIST experiment also isolates this effect (appendix “Empirical Evidence of Instabilities,” paragraph “Conditional Guidance for MNIST digits,” cited in the caption of Figure 3). With the same learned score, switching guidance off for the last 100 steps removes the observed oscillations: the alignment of consecutive steps stays near 00, whereas with guidance it drops to −1-1, indicating a period-two oscillation.

Section 5 now explains the mechanism. DPS discretizes

\left( Y_t + 2\nabla\log\rho_{T-t}(Y_t) \underbrace{-2\zeta\eta_{T-t}G_{T-t}(Y_t)}_{\text{guidance}} \right)dt +\sqrt{2}\\,dB_t,$$ where $$G_t(x):=\nabla_x\left\\|\mathcal{A}(\hat{x}_t(x))-y\right\\|,$$ $\zeta$ is the DPS step-size hyperparameter, and $\eta_t$ is an annealing schedule. The first two drift terms, together with the Brownian noise, define the reverse OU denoising process. Its DDPM discretization uses the score only through the bounded Tweedie mean, so the component normal to the data manifold contracts. The guidance update subtracts $2\zeta G_{t_i}(x_i)$. Indeed, the DPS schedule $$\zeta_i=\frac{\zeta}{\\|y-\mathcal{A}(\hat{x}_{t_i}(x_i))\\|}$$ cancels the residual norm in the gradient of the squared residual, while $\eta_{t_i}=1/\Delta t_i$ cancels the time-step factor. Thus, the guidance step does not shrink with the residual and can overshoot near the constraint. The oscillation is tangential because $G_t$ contains the Jacobian $\nabla\hat{x}_t$, which is tangent-valued to leading order near the manifold (see the limit below). Our exact-score circle example demonstrates this directly: with the same random seeds, denoising alone reaches the circle, while guidance produces oscillations in the tangential coordinate. **Weakness 2: Equations 36 and 40** Thank you for identifying this omission. Our previous expansion dropped the pairing of $x-p$ with the second fundamental form. The corrected limit, now given in Appendix C.2.1, is $$\lim_{t\to 0}\nabla\hat{x}_t(x) = (I-S_{x-p})^{-1}P_{T_p\mathcal{M}}, \qquad p=P_{\mathcal{M}}(x).$$ where $S_{x-p}$ is the shape-operator term, as in [LSHH26]. On the unit sphere, this gives $P_{T_p\mathcal{M}}/\\|x\\|$. For fixed $p$, $S_{x-p}$ is linear in the normal displacement $x-p$ and satisfies $$\\|S_{x-p}\\|\leq\kappa\\|x-p\\|,$$ where $\kappa$ bounds the curvature of $\mathcal{M}$. Expanding the inverse, the appendix now states$$\begin{aligned} \nabla\hat{x}_t(x) ={}& \underbrace{P_{T_p\mathcal{M}}}_{\text{leading order}} + \underbrace{S_{x-p}P_{T_p\mathcal{M}}}_{O(\\|x-p\\|)} \\\\ &+ \underbrace{S_{x-p}^2(I-S_{x-p})^{-1}P_{T_p\mathcal{M}}}_{O(\\|x-p\\|^2)} + \underbrace{R_t(x)}_{O(t)}. \end{aligned}$$ Thus, $P_{T_p\mathcal{M}}$ is the leading term as $t\to 0$ and $x$ approaches $\mathcal{M}$. Curvature enters at first order in the distance to the manifold, so this correction vanishes on approach to it. The corrected limit remains tangent-valued, preserving the conclusions of Section 5.

Questions

Gaussian start. We added a clarification at the beginning of the discussion of Theorem 1. DPS starts from γ=N(0,I)\gamma=\mathcal{N}(0,I) rather than the terminal surrogate law. Theorem 1 accounts for this through the boundary factor μ→TDPS/γ\overrightarrow{\mu}^{\mathrm{DPS}}_T/\gamma. Under the stated assumptions, exponential convergence of the OU semigroup gives

∥μ→TDPSγ−1∥C1(K)=O(e−T)for every compact K⊂Rd.\Vert \frac{\overrightarrow{\mu}^{\mathrm{DPS}}_T}{\gamma}-1 \Vert_{C^1(K)} = O(e^{-T}) \qquad \text{for every compact } K\subset\mathbb{R}^d.

For the 1000-step schedule, e−T≈6×10−3e^{-T}\approx 6\times10^{-3}. The bias that does not vanish as TT increases therefore comes from the path integral of the reaction term cDPSc_{\mathrm{DPS}}.

Computing 1/ω1/\omega. For Figure 2, we compute 1/ω1/\omega using a finite-difference solver for the linear PDE satisfied by the density ratio (appendix “Experimental Details for Figure 2”). In higher dimensions, a spatial grid becomes infeasible. The Feynman–Kac representation instead permits Monte Carlo estimation over OU paths or sequential Monte Carlo with path reweighting.

The accuracy of these methods is limited by the effective sample size, which decreases as the variance of the path weights grows. We therefore do not claim that 1/ω1/\omega can be computed efficiently in general. Rather, the formula reveals where and why DPS over- or under-samples the posterior. In the quadratic measurement setting, the expression for cDPSc_{\mathrm{DPS}} explains the pronounced over-sampling of regions closer to the constraint.

I thank the authors for correcting this issue.

  1. After the authors' clarification, I still don't see the significance of this oscillation result. It seems to be that the guidance term is large in all directions, is this correct? As in the revised manuscript, lines 466-478, this appears to be a simple observation that gradient descent with large step sizes exhibits oscillations. Why does this relate to instability of DPS? I also could not understand the preceding paragraphs in the revision, there are several broken clauses.
    • It seems that this is used to motivate the implicit method proposed in Appendix J. However no experiments are done on larger experiments, probably due to the computational infeasibility.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.