acceptodds
Under review as a conference paper at ICLR 2027

A Unified Framework for Bayesian Data Assimilation with Generative Models and Observation Interpolants

Abstract

Bayesian data assimilation combines model forecasts with noisy observations, but sampling high-dimensional, non-Gaussian posteriors remains challenging. We introduce an observation-interpolant framework that turns pretrained stochastic interpolant, flow matching, and diffusion models into posterior samplers without retraining. Conditioning the interpolant path on observations yields a shared likelihood-score correction to the drift or velocity, unifying stochastic and deterministic posterior sampling. The resulting SDEs and ODEs sample the exact posterior when the intermediate likelihood score is known. For practical computation, we approximate this score using a closed-form Gaussian surrogate with a bias-corrected mean and covariance inflated by the model’s source covariance. Jacobian-free and ensemble-shared approximations make the method tractable in high dimensions. We evaluate the framework on linear–Gaussian dynamics, stochastic two-dimensional Navier–Stokes, and urban airflow with up to degrees of freedom.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Data assimilation, posterior sampling, stochastic interpolants, flow matching, diffusion models, stochastic differential equations

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_dcLZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a unified framework for data assimilation using pretrained prior. Assuming that we can sample from climatology, we can pretrain a density model such as diffusion model, continuous normalizing flow with flow matching, or stochastic interpolant. The authors propose a unified posterior sampler (equation 11 / Theorem 4.1 in the paper) for SI, DM, and FM. The proposed method generally outperform other work with the same framework, but the 1000-member ensemble kalman filter equipped with the true solver performs significantly better.

Critical strengths
  • Unified framework is presented beautifully. The writing is clear and educational (e.g., Figure 1 and Section 3) so that it presents a new perspective for a reader with less experience in the SDE-based view of density models.
  • In the Navier-Stokes experiment, the authors show experimental results candidly that 1000-member EnKF performs significantly better than the guidance methods. It helps to position the work clearly against classical baselines.
Critical weaknesses
  • It was hard to assess the core novelty of this paper. The observation-based guidance of a trained density model such as diffusion model or flow matching is already established (e.g., “Score-Based Generative Modeling through Stochastic Differential Equations”), the unification of diffusion model and flow matching as stochastic interpolant is already established (“Stochastic Interpolants: A Unifying Framework for Flows and Diffusions”), and sample generation under linear observation operator has a few precedents (“FIG: Flow with Interpolant Guidance for Linear Inverse Problems” and ”Tweedie Moment Projected Diffusions for Inverse Problems”).
  • Perhaps the approximations presented in Sections 4.2 and 4.3 may be original, but I could not assess how novel and significant these are.
Questions for authors
  • Which aspects of the paper do authors regard as the core contribution of this work? Is it the unified posterior sampler (Theorem 4.1, equation (11))? If that is the case, this seems to me a reinterpretation of existing ideas, which kind of already exists in the stochastic interpolant paper, so the novelty/originality is not very high.
  • Are the approximations in Sections 4.2 and 4.3 new and important? If so, could you discuss the differences with neighboring methods such as FIG and TMPD?
1 comment
AuthorsOpenReview ↗

Official response

Thank you for the kind words on the presentation and for the direct questions about novelty, which helped us sharpen what the paper claims. Please see the general response at the top for an overview of all changes.

"Which aspects of the paper do authors regard as the core contribution of this work? Is it the unified posterior sampler (Theorem 4.1)?"

Not Theorem 4.1 on its own, which the revision presents as a consolidation of known identities. The core contribution is the framework as a whole: existing works each cover part of it, and ours brings the parts together and adds two that are missing.

  1. One statement of the guided sampler covering Gaussian and point-mass sources and both SDE and ODE samplers (stochastic interpolants, flow matching, and diffusion models), with an explicit guidance weight.
  2. A derivation that separates the exact from the approximate. The score identity and the likelihood moments are exact. The Gaussian closure, where the equivalence with TMPD and MMPS arises (Proposition 4.4), is a separate and replaceable step.
  3. Covariance approximations (Jacobian-free and ensemble-shared) that avoid the per-particle Jacobian products of TMPD, MMPS, DAISI and FA-APF.
  4. An EnKF-type correction for the filtering posterior, that balances the efficiency of EnKF with the non-Gaussian and nonlinear update that the guidance provides, as an alternative to particle weights.

The table in the general response (point 1) sets this against the closest works.

"Are the approximations in Sections 4.2 and 4.3 new and important? If so, could you discuss the differences with neighboring methods such as FIG and TMPD?"

Yes on both counts. These approximations (now all in Section 4.2) are the main practical contribution.

Why they are new. The observation interpolant gives the exact mean and covariance of the intermediate likelihood for general schedules, directly from the learned drift or velocity. Our two variants approximate that exact covariance: the Jacobian-free one drops the Jacobian term, and the shared one evaluates the Jacobian once for the whole ensemble. Because the formulation is unified, both apply unchanged to the SI, FM and diffusion samplers. Neither appears in TMPD, MMPS, DAISI or FA-APF.

Why they matter. Methods of this type usually need one derivative computation of the network per ensemble member per step. Our variants need none, or one for the whole ensemble, which is what makes the field cases affordable.

  • TMPD uses the same Gaussian likelihood (Proposition 4.4) and differs in how the covariance is evaluated: it needs Jacobian products for every member at every step. On the same prior, the two agree on the linear–Gaussian test and in Navier–Stokes super-resolution from the true initial condition. On Navier–Stokes, TMPD takes about 270 s per step against 8–14 s (Jacobian-free) and 45–134 s (shared) for ours. Under sparse sensors and in the urban case TMPD diverged, likely because of its row-sum covariance, while our samplers remained stable.
  • FIG also interpolates the measurement, but it uses a normalised corrector without a covariance and treats xτ≈x1x_\tau\approx x_1. It is competitive at 32232^2 super-resolution only. Under sparse sensors it over-smooths (RMSE ×10−1\times10^{-1} 7.23 at 5%, against 1.73 for our SI-SDE sampler with the shared covariance), and it does not recover from the poor initial condition (8.69 and 8.53, against 4.25 and 1.09).

How much the covariance matters. Under sparse sensors the shared covariance reduces the SI-SDE error by a factor of 2–4. From the poor initial condition the Jacobian-free covariance is the most accurate, at 5–9 times lower cost. Which variant is best is therefore problem-dependent, as the conclusion states, and the refresh-rate ablation (Appendix J.2) quantifies the trade-off between accuracy and cost.

On the 1000-member EnKF in your summary: it has the lowest errors in three of the four true-initial-condition scenarios, while at 32232^2 super-resolution our samplers are ahead (RMSE ×10−1\times10^{-1} 0.63–0.69 against 0.94), and it needs the true solver at about 100 s per step. We also added two tests with a reference posterior and an experiment from a poor initial condition. The general response summarises the findings, including where the classical EnKF remains as good or better.

Reviewer_tGUhOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a unified posterior sampling method using generative models, which places observation guidance under generative models based on stochastic interpolants, flow matching and diffusion models under a single framework. To perform guidance under linear observations, the paper introduces the observation interpolant, enabling a closed-form expression for the likelihood score. On experiments on an 2D linear system, Navier-Stokes and urban airflow, the proposed method yields improved RMSE and CRPS scores compared to previous generative-based DA algorithms including FlowDAS, SDA and D-Flow SGLD.

Critical strengths

The writing is generally easy to follow and from a practitioner's point of view, the way that the paper is presented is useful, with explicit formulas laid out already. Hence, if one has a trained generative forecast model, then they can simply use the formulas that are laid out to perform sequential data assimilation with minimal work required from the practitioner. The only concern / question I have here is whether SURGE-style particle correction is used in the proposed approach, as without it, it will sample from a biased distribution p(xn∣xn−1,yn)p(x_n \mid x_{n-1}, y_n) vs the true filtering distribution p(xn∣y1,…,yn)p(x_n \mid y_1, \dots, y_n), even when using the closed-form likelihood score in section 4.1. Hence, some clarification here is necessary.

Critical weaknesses

Limited novelty: From my read of the paper, the biggest weakness is the lack of new contributions. The main contributions of the proposed filtering algorithm are stated as:

  1. Unified perspective of posterior sampling using stochastic interpolants (technically, this refers in particular to Follmer-style interpolant proposed in [Chen et al., 2024]), flow matching and diffusion.
  2. A formula for the exact likelihood score estimation.

However, if my understanding of the contribution above is correct, I would argue that both are known in the general diffusion literature. For the first point, this simply follows from computing the posterior score as a sum of prior score and likelihood score and use the relation between score and velocity (also the drift b, but that is simply a linear combination of the score and noise) (see e.g. Appendix A in [Andrae et al. 2025]), and I would say that these tricks are already quite well-known in the literature.

For the second point, the proposed closed-form guidance term appears identical to MMPS [Rozet et al. 2024] under some change of variables. In particular, MMPS is used now in some more recent diffusion-based DA literature, e.g. the aforementioned work [Andrae et al. 2025], or in APPA [Andry et al. 2025] and FA-APF [Savary et al. 2026]. In particular, I believe FA-APF is very similar to the proposed approach in the paper (generative dynamics + MMPS-style guidance + SURGE-style correction), and therefore the literature review needs better positioning.

[Andrae et al. 2025] Andrae, Martin, et al. "DAISI: Data Assimilation with Inverse Sampling using Stochastic Interpolants." arXiv preprint arXiv:2512.00252 (2025).

[Rozet et al. 2024] Rozet, François, et al. "Learning diffusion priors from observations by expectation maximization." Advances in Neural Information Processing Systems 37 (2024): 87647-87682.

[Andry et al. 2025] Andry, Gérôme, et al. "APPA: Bending weather dynamics with latent diffusion models for global data assimilation." arXiv preprint arXiv:2504.18720 (2025).

[Savary et al. 2026] Savary, Thomas, François Rozet, and Gilles Louppe. "Training-Free Bayesian Filtering with Generative Emulators." arXiv preprint arXiv:2605.20028 (2026).

Framework cannot be extended beyond linear observations: As also explained in the Limitations section, the main contribution hinges on the linearity of the observation operator. This severely limits the applicability of the method, as most realistic observation operators are nonlinear. In addition, the setting of using linear observation operators is where traditional filtering (e.g. EnKF, LETKF) already performs strongly. The experiments on Navier-Stokes show an EnKF baseline (presumably stochastic EnKF) with 64 ensemble members, which perform slightly worse compared to the proposed method. However, using a more state-of-the-art variant such as LETKF with well-tuned inflation and localization parameters should produce stronger results in this setting. Hence, by focussing on linear observation operators, it is difficult to see where the diffusion/flow/interpolant-based approach wins. Also note that the advantage of MMPS (mentioned above) is that it extends naturally to nonlinear operators and that is what is used in the aforementioned recent filtering approaches such as DAISI and FA-APF.

Questions for authors
  • For the proposed method, do you also use SURGE-style particle correction, similar to the FlowDAS and SDA baselines? As mentioned earlier, without the correction, theoretically the method will target a biased distribution p(xn∣xn−1,yn)p(x_n \mid x_{n-1}, y_n) instead of the true filtering distribution p(xn∣y1,…,yn)p(x_n \mid y_1, \ldots, y_n). This is also the problem with FlowDAS without the SURGE-enhancement.
  • 15 assimilation steps to assess the filtering quality for NS is quite short. Have the authors considered running the filter for longer? I am suspicious since the RMSE/CRPS plots in Fig 11-12 have not converged, indicating that the filter is still in warm-up stage.
  • What is the initialisation used for the filter in the experiments? In the plots in Fig 11-12, the RMSE/CRPS increases, indicating that the initialisation is already very strong and the filtered results diverges from the ground truth. Ideally, we would like to see the opposite, where the initialisation is poor and the filter works to correct the deviation from the truth.
  • I have a minor gripe with the use of the word "stochastic interpolant" in this work. Stochastic interpolants, as introduced in [Albergo et al. JMLR (2025)] refers to the general idea of how to convert a stochastic interpolant xt=αtx0+βtx1+γtϵx_t = \alpha_t x_0 + \beta_t x_1 + \gamma_t \epsilon into a generative ODE/SDE with matching marginal distributions pt(xt)p_t(x_t). In particular, it is a framework to show how flow matching and diffusion models can be interpreted under a unified perspective. The "stochastic interpolant" that is used in this work is something much narrower, and refers to the Follmer-process interpolant used in [Chen et al. ICML (2024)], which takes a point measure as the reference, i.e. p0(x0)=δx(x0)p_0(x_0) = \delta_{x}(x_0). In my opinion, this creates confusion as the paper treats stochastic interpolants and flow matching / diffusion models as different frameworks, when the purpose of stochastic interpolants was to unify these perspectives. Hence, I think it would be much better to call the "stochastic-interpolant" used in this work a "Follmer-process interpolant" (or something along this line).
2 comments
AuthorsOpenReview ↗

Official response (1/2)

Thank you for the thorough review and for the pointers to DAISI, APPA and FA-APF, which we had missed. Your main concerns are novelty, nonlinear observations and the classical baselines. We answer each after your first question. Please see the general response at the top for an overview of all changes.

"whether SURGE-style particle correction is used in the proposed approach, as without it, it will sample from a biased distribution"

No. Our samplers used no particle correction, and SURGE was applied only to the FlowDAS and SDA baselines. You are right that the method then targets p(xn∣xn−1,yn)p(x_n\mid x_{n-1},y_n) per member and not the filtering distribution. The revision states this (Sections 3 and 4.4, Appendix D) and addresses it in two ways.

  • A forecast-residual correction (Eq. 18). We chose an EnKF-type step over importance weights because weight-based resampling degenerates in high dimensions and often collapses small ensembles. The correction strikes a balance: the guided sampler keeps the nonlinear, non-Gaussian part of each member's update, and the EnKF-type step adds the cross-member part efficiently and without weights. On the multistep linear–Gaussian test with the shared covariance, it brings the KL from 6⋅10−26\cdot10^{-2} to 6.5⋅10−56.5\cdot10^{-5} for SI-SDE (EnKF: 5.5⋅10−55.5\cdot10^{-5}), and on Lorenz 96 it brings our samplers on par with the EnKF. On Navier–Stokes it helps from a poor initial condition but raises the errors from the true one, likely because a linear, Gaussian update degrades the non-Gaussian samples of an ensemble that already tracks the truth (Section 5.3 and the limitations).
  • SURGE around our own SDE samplers. On the multistep test it reaches a KL of about 1.3⋅10−41.3\cdot10^{-4} around SI-SDE. With 64 particles it raises the error on Lorenz 96 and, with one exception, on Navier–Stokes from the poor initial condition (Tables 4 and 7). On Navier–Stokes its ensembles collapse. Our experiments thus confirm that reweighting-based methods degenerate with small ensembles in high dimensions.

"Limited novelty ... both are known in the general diffusion literature"

The individual ingredients are known, and the paper now says so. DAISI's Appendix A.4 gives the guided drift for Gaussian-source interpolants, FA-APF's Appendix E derives it for an interpolant started at the previous state, without testing it, and Proposition 4.4 states the equivalence with MMPS and TMPD. Section 2 cites MMPS, TMPD, APPA, DAISI and FA-APF.

What these works do not provide is the unification, which we see as the first contribution. Each is developed and tested for one model class: DAISI for Gaussian-source interpolants and FA-APF for diffusion models. Both remark that other model classes fit their framework, but neither gives a single guided formulation for them. Our formulation covers stochastic interpolants, flow matching and diffusion models, with Gaussian and point-mass sources and with SDE and ODE samplers, in a single statement: one posterior drift, one guidance weight and one observation-interpolant likelihood, implemented as one algorithm. A method or approximation developed for one model class then carries over to the others, and the three samplers can be compared on equal terms, which the paper does on common test cases.

Within this formulation we add two pieces that none of these works has.

  • Covariance approximations that avoid the per-particle Jacobian products, a main cost of MMPS-type guidance. On Navier–Stokes, on the same prior and hardware, the per-step cost goes from about 270 s for TMPD to 8–14 s (Jacobian-free) or 45–134 s (shared).
  • An EnKF-type correction for the filtering posterior that needs no particle weights and runs with 64 members. FA-APF relies on tempered weights, and DAISI leaves a principled correction to future work.

The paper also tests the point-mass-source case, which FA-APF derives but does not test.

(Continued in part 2/2.)

AuthorsOpenReview ↗

Official response (2/2)

(Continued from part 1/2.)

"Framework cannot be extended beyond linear observations ... the advantage of MMPS is that it extends naturally to nonlinear operators"

Section 4.3 now extends the observation interpolant to differentiable operators by linearisation at the posterior mean, as DAISI does with MMPS. We do not claim it as new. We test it on Lorenz 96 with arctan observations (Section 5.2). The corrected samplers are on par with the EnKF, and our Jacobian-free SI-SDE sampler with the correction has the lowest kkNN-KL (3.02 against 3.50) and a slightly higher sliced W2W_2 (1.21 against 1.10), while Guided FM and D-Flow SGLD degrade.

"a more state-of-the-art variant such as LETKF with well-tuned inflation and localization parameters should produce stronger results"

Your reading is correct: the 64-member EnKF from the true initial condition is a stochastic EnKF without localisation or inflation, and we did not run an LETKF. With all the new test cases, the rebuttal period did not allow us to rerun every case with an LETKF. We will gladly include such runs in a camera-ready version. We would like to explain how we see the role of the classical filters in this paper.

The purpose of the paper is to present methods that use generative models for data assimilation. We therefore test them first on benchmark problems where a reference exists and where classical filters are known to do well: the linear–Gaussian test, Lorenz 96 and stochastic Navier–Stokes. Papers in this line of work often omit classical filters altogether. We include them because we agree with you that they are the right yardstick, and we added a localised stochastic EnKF in two settings.

  • Lorenz 96 (localisation and inflation tuned): it is on par with our corrected samplers.
  • Navier–Stokes from the poor initial condition (radius 20): RMSE ×10−1\times10^{-1} 4.82 and 1.48, against 4.25 and 1.09 for our best sampler. The EnKF is better calibrated and cheaper (6 s against 14–19 s per step), and without localisation it diverges at 16216^2.

At the same ensemble size of 64, our best sampler is clearly ahead of the EnKF in several cases (super-resolution, sparse sensors and recovery from the poor initial condition) and on par in the others.

Generative, data-driven methods also open up problems that classical filters cannot reach: those where the forward model is too expensive to run in an ensemble, or is not available at all. The urban case demonstrates this. Its forward model is a large-eddy simulation, and our samplers assimilate sparse observations there in 20–84 s per step without calling the solver. The trained prior also models a single horizontal slice of the three-dimensional flow directly. No numerical model exists for such a slice on its own, so a classical filter would have to run the full three-dimensional simulation.

"15 assimilation steps ... is quite short" and "What is the initialisation used ... we would like to see the opposite, where the initialisation is poor"

The original runs start from the true state, which is why the errors rise towards their stationary level, as you suspected. We added the experiment you describe (Section 5.3, Appendix J.3): every method starts from the history of a different trajectory and assimilates 35 steps. The errors now decrease over the run (Figures 11–12), and our FM-ODE sampler with the Jacobian-free covariance recovers best in RMSE and CRPS, ahead of the localised EnKF and well ahead of the generative baselines. This covers the two super-resolution scenarios only. Longer horizons are also covered by Lorenz 96 (50 assimilation steps) and the urban case (49 updates).

"it would be much better to call the 'stochastic-interpolant' used in this work a 'Follmer-process interpolant'"

Thank you for pointing this put. Section 3 now says that stochastic interpolants generalise diffusion and flow matching, that we use the Föllmer-process construction of Chen et al. (2024), and that our results carry over to the original formulation. We keep "SI" as shorthand for this construction and say so explicitly. We hope this is acceptable with the additional text in the manuscript.

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

The paper proposes an approach to Bayesian data assimilation using pretrained flow-based generative models as forecast priors. To sample from the posterior without retraining, the authors add a likelihood-score correction to the drift (SDE) or velocity (ODE) of the sampler. To compute this score in practice, the authors interpolate the observation with the same schedule as the state and approximate its likelihood by a Gaussian, with a bias-corrected mean and an inflated covariance computed from the prior model's score.

Critical strengths
  • The paper presents useful theoretical results, including a single posterior sampler for stochastic interpolants, flow matching and diffusion (Theorem 4.1), in addition to a closed-form approximation of the likelihood score for SI.
  • The experimental setup is carefully designed and described, with a clear evaluation of the cost-accuracy tradeoff of the evaluated methods.
Critical weaknesses
  • W1. The main contribution is overstated Section 4.1 (lines 203–212) says that interpolating the observation "removes" the intractability of the likelihood score, and that the exact score can "then" be replaced with the interpolated score. In truth, the interpolated observation makes nothing tractable: given x0x_0, the interpolated observation yˉτ=ατHx0+βτy\bar{y}_\tau = \alpha_\tau H x_0 + \beta_\tau y (Eq. 12) is a fixed rescaling and shift of yy, which means ∇xτlog⁡p(yˉτ∣xτ,x0)=∇xτlog⁡p(y∣xτ,x0)\nabla_{x_\tau} \log p(\bar{y}_\tau \mid x_\tau, x_0) = \nabla_{x_\tau} \log p(y \mid x_\tau, x_0). Tractability comes only from the Gaussian approximation in Eq. 16. For FM and diffusion (ατ=0\alpha_\tau = 0), dividing yˉτ\bar{y}_\tau by βτ\beta_\tau, the moments in Lemmas 4.2–4.3 give a Gaussian likelihood for yy with mean HE[x1∣xτ,x0]H \mathbb{E}[x_1 \mid x_\tau, x_0] and covariance R+HCov[x1∣xτ,x0]H⊤R + H \mathrm{Cov}[x_1 \mid x_\tau, x_0] H^\top. This is the first- and second-order Tweedie surrogate of TMPD [1] and MMPS [2]. The authors do not cite either paper, which makes their contribution seem more substantial than it is. I suggest the authors fix this presentation throughout the paper as follows:

    1. State that ∇xτlog⁡p(yˉτ∣xτ,x0)=∇xτlog⁡p(y∣xτ,x0)\nabla_{x_\tau} \log p(\bar{y}_\tau \mid x_\tau, x_0) = \nabla_{x_\tau} \log p(y \mid x_\tau, x_0), so the rescaling is exact and gives no new approximation.
    2. Present the interpolation as a way to compute the exact mean and covariance of the likelihood.
    3. State that tractability comes from the Gaussian approximation in Eq. 16, and compare it with the earlier Gaussian approximations (DPS [3], ΠGDM [4], TMPD [1]).
  • W2. The experimental results do not establish the superiority of the method. The method is useful only if it does better where Gaussian methods fail, i.e. with sparse observations and non-Gaussian posteriors. The current empirical results fail to establish a gain in performance in this setting. Specifically:

    • (a) No test of a non-Gaussian posterior. The only comparison to an exact posterior is the 2D linear-Gaussian case (Section 5.1, Fig. 3). The paper says the surrogate "is exact in the linear–Gaussian regime" (line 242), so this test cannot show how well the approximation works for non-Gaussian posteriors.
    • (b) With sparse sensors, the baselines are better. From Table 1 (lines 398–415): with 5% sensors, SDA+SURGE beats the proposed method on RMSE, and with 1/64 sensors, both D-Flow SGLD and SDA+SURGE beat it on RMSE, and D-Flow SGLD also beats it on CRPS.
    • (c) The classical baseline is better, and should be tested more extensively. The EnKF with 1000 members has the lowest RMSE and CRPS in 3 of the 4 Navier–Stokes scenarios, and costs 102–122 s per step, compared with 41–165 s for the shared variant of the proposed method. The authors accept that it "attains the lowest sparse-scenario errors" (line 376), but attribute this to its access to the "true solver" (line 377). I suggest the authors test this explanation by applying the EnKF update to the same learned prior as the generative methods.
  • W3. Tips to improve the paper's presentation:

    • The presentation of the tables is non-standard and confusing, despite the clarifications in the captions. It would be better to follow the standard presentation: add arrows directly to the column headers, bold only the best value, and underline the second-best value.
    • The same symbol has two meanings. x0x_0 is the previous physical state (Section 3), but X0∼N(0,I)X_0 \sim \mathcal{N}(0, I) in Eq. 5 is the noise source of FM. Also, μˉ\bar{\mu} and Σˉ\bar{\Sigma} denote the likelihood moments in Eq. 14 but the interpolant moments in Eq. 48 (Appendix G).
    • Typos: "generatibe" (conclusion, line 471); "All runs are performed an NVIDIA" is missing "on" (line 329).
    • CRPS is the main metric for the ensemble, but it is defined only in Appendix E. A short definition in the main text would help the reader.
Questions for authors
  1. For FM and diffusion (ατ=0\alpha_\tau = 0), are Lemmas 4.2–4.3 with Eq. 16 equivalent to the Gaussian likelihood of TMPD [1] and MMPS [2]? If yes, what is new beyond the SI case?
  2. Can you apply the EnKF update to the same 64 forecasts from your trained model? Can you also run the 64-member EnKF with localisation and inflation? This would separate your correction step from the "true solver" advantage (lines 376–377).
  3. Can you test a small nonlinear system with a reference posterior, for example Lorenz-96 [5] with a long particle filter as reference?

References

[1] B. Boys, M. Girolami, J. Pidstrigach, S. Reich, A. Mosca, and O. D. Akyildiz. Tweedie Moment Projected Diffusions for Inverse Problems. Transactions on Machine Learning Research (TMLR), 2024.

[2] F. Rozet, G. Andry, F. Lanusse, and G. Louppe. Learning Diffusion Priors from Observations by Expectation Maximization. Advances in Neural Information Processing Systems (NeurIPS), 2024.

[3] H. Chung, J. Kim, M. T. Mccann, M. L. Klasky, and J. C. Ye. Diffusion Posterior Sampling for General Noisy Inverse Problems. International Conference on Learning Representations (ICLR), 2023.

[4] J. Song, A. Vahdat, M. Mardani, and J. Kautz. Pseudoinverse-Guided Diffusion Models for Inverse Problems. International Conference on Learning Representations (ICLR), 2023.

[5] E. N. Lorenz. Predictability: A problem partly solved. Proc. Seminar on Predictability, vol. 1, pp. 1–18, ECMWF, 1996.

4 comments
AuthorsOpenReview ↗

Official response (1/2)

Thank you for the constructive review. We followed the three steps you proposed under W1 and ran the experiments you asked for in Q2 and Q3. Please see the general response at the top for an overview of all changes.

W1: "The main contribution is overstated"

We rewrote the presentation along your three points.

  1. Section 4.1 states that ∇xτlog⁡p(yˉτ∣xτ,x0)=∇xτlog⁡p(y∣xτ,x0)\nabla_{x_\tau}\log p(\bar y_\tau\mid x_\tau,x_0)=\nabla_{x_\tau}\log p(y\mid x_\tau,x_0), so the rescaling is exact and gives no new approximation.
  2. The interpolation is presented as a way to obtain the exact mean and covariance of the likelihood, and the section is retitled accordingly.
  3. Section 4.2 states that tractability comes from the Gaussian approximation, and Section 2 positions it against DPS, ΠGDM, TMPD and MMPS.

The text you quote ("removes" the intractability) is gone, and TMPD and MMPS are now cited and discussed. Figure 1 keeps "inflated covariance" and "bias-corrected mean" as labels for the exact likelihood moments, which are given by the Tweedie covariance and mean.

With the overstatement removed, we believe the role of the interpolation is clearer. It offers an alternative view of guidance: the observation is transported along the same path as the state, which gives the exact likelihood moments for general schedules, directly from a learned drift or velocity. The Gaussian approximation then enters as one explicit and replaceable step.

Q1: "are Lemmas 4.2–4.3 with Eq. 16 equivalent to the Gaussian likelihood of TMPD and MMPS? If yes, what is new beyond the SI case?"

Yes. Proposition 4.4 shows the equivalence for any ατ\alpha_\tau and βτ≠0\beta_\tau\neq0, so also for SI. This Gaussian step is the only place where the method coincides with TMPD and MMPS. Beyond the SI case, three things are new.

  • One guided sampler stated for stochastic interpolants, flow matching and diffusion models, where existing works are each developed for one model class.
  • Covariance approximations that avoid the per-particle Jacobian, which is what makes the method affordable in the field cases.
  • An EnKF-type correction for the filtering posterior.

None of TMPD, MMPS, DAISI or FA-APF has the last two (general response, point 1).

W2: "The method is useful only if it does better where Gaussian methods fail"

This is a fair test, and the answer is mixed.

  • Sparse sensors, your criterion. Our SI-SDE sampler with the shared covariance has lower RMSE than the 64-member EnKF, with the true solver or the trained prior (1.73 against 2.77–3.00 at 5%, and 3.85 against 6.21–6.51 at 1.563%).
  • Super-resolution. Our shared samplers have the best RMSE and CRPS among the generative methods, with TMPD on par. They are ahead of the 64-member EnKF and, at 32232^2, of the 1000-member EnKF.
  • Poor initial condition. Our best sampler is ahead of a localised EnKF with the true solver in RMSE and CRPS.
  • No gain. Against the generative baselines at the sparsest density, and on Lorenz 96, where the EnKF is as good.

W2(a) / Q3: "No test of a non-Gaussian posterior ... Lorenz-96 with a long particle filter as reference?"

We added stochastic Lorenz 96 with a 4⋅1064\cdot10^6-particle reference and linear and arctan observations (Section 5.2, Table 4). With the correction, our samplers are on par with the 64-member EnKF (kkNN-KL 4.1–4.4 against 4.26 for linear and 3.0–3.5 against 3.50 for arctan observations) and ahead of the generative baselines. The posterior is only mildly non-Gaussian, so this validates the samplers but shows no gain over the EnKF.

(Continued in part 2/2.)

AuthorsOpenReview ↗

Official response (2/2)

(Continued from part 1/2.)

W2(b): "With sparse sensors, the baselines are better."

At 1.563% this is correct: D-Flow SGLD and SDA + SURGE have lower RMSE (2.31 and 2.97 against 3.85), with D-Flow SGLD at about 700 s per step against 42 s for ours. At 5%, SDA + SURGE is slightly lower (1.62±0.161.62\pm0.16 against 1.73±0.201.73\pm0.20) and our SI-SDE sampler with the shared covariance has the best CRPS. Section 5.3 now states this.

Looking across all test cases, however, these baselines do well in a few settings only, while our samplers perform well throughout. The two that lead under sparse sensors are well behind in super-resolution (RMSE ×10−1\times10^{-1} 3.70 and 1.58 for SDA + SURGE and 2.46 and 2.01 for D-Flow SGLD, against 2.02 and 0.63 for ours) and do not recover from the poor initial condition (at least 9.85 and 8.80, against 4.25 and 1.09). SDA + SURGE also collapses the ensemble, and D-Flow SGLD degrades on Lorenz 96 with arctan observations. Our samplers are best or close to the best in super-resolution, from the poor initial condition and in the urban case, and competitive at 5% sensors.

W2(c) / Q2: "Can you apply the EnKF update to the same 64 forecasts from your trained model? Can you also run the 64-member EnKF with localisation and inflation?"

We ran the first, and the second in part.

  • EnKF on the trained prior. Under sparse sensors it matches the EnKF with the true solver (RMSE ×10−1\times10^{-1} 2.86–3.00 and 6.21–6.51, against 2.77 and 6.22), so at 64 members the true solver is not what the EnKF gains from. In super-resolution it is well behind our samplers (4.3–4.5 and 1.4, against 2.0–2.8 and 0.63–0.69), and on Lorenz 96 it is on par with our corrected samplers.
  • The 1000-member EnKF has the lowest errors in three of the four true-initial-condition scenarios. At 32232^2 our samplers are ahead (RMSE 0.63–0.69 against 0.94), with the Jacobian-free samplers at 7–13 s per step against 102 s. The shared samplers cost 78–165 s there.
  • Localisation and inflation. On Lorenz 96 both are used and tuned. On Navier–Stokes we ran a localised EnKF (radius 20, no inflation) from the poor initial condition only (Table 7): RMSE ×10−1\times10^{-1} 4.82 and 1.48, against 4.25 and 1.09 for our best sampler, with better calibration for the EnKF at 6 s per step.

W3: presentation

  • Tables. All results tables now have arrows, bold for the best value and underline for the second best.
  • Typos. Both are fixed.
  • Symbols. XτX_\tau is the sampler process and x0x_0 the conditioning state, so X0∼N(0,I)X_0\sim\mathcal N(0,I) and x0x_0 are different objects. We kept this notation to avoid a further symbol and now state the distinction right after Eq. 5. The second clash is removed: μˉ\bar\mu and Σˉ\bar\Sigma denote only the likelihood moments.
  • CRPS. With the new experiments Section 5 is at the page limit, so the definition stays in Appendix F, which Section 5 points to. We hope this is acceptable.
Reviewer_GFyCOpenReview ↗

Thank you for your response; my concern in W2 remains

Thank you for the detailed response and the extensive new experiments. The revision addresses W1 and W3 well, but my main concern in W2 remains. The new results still do not show a gain where Gaussian methods should fail, and I believe this is because none of the test cases contains such a setting. Concretely, in Lorenz 96 the transition noise is Gaussian and the posterior is only mildly non-Gaussian, as you note, and without the correction your samplers are behind the 64-member EnKF, and in Navier–Stokes there is no reference posterior, and the EnKF is untuned in the true-initial-condition runs.

As a result, the experiments cannot tell whether the method gains over Gaussian methods in strongly non-Gaussian settings, or whether it is limited there. Both parts of the method also rely on Gaussian assumptions: the guidance uses a Gaussian approximation of the likelihood (Eq. 15), and the correction is an EnKF step. So a limitation is plausible.

I would suggest one of the following:

  1. A small test where the EnKF is known to fail, with a long particle filter as reference. For example, Lorenz 63 or Lorenz 96 with long intervals between observations, or with observations that give a two-peaked posterior (e.g. y=x2y = x^2 or ∣x∣|x|). Comparing against a tuned EnKF there would show directly whether the method helps where Gaussian methods fail.
  2. If this is not feasible, please state clearly in the limitations that the method has not been shown to improve on Gaussian methods for strongly non-Gaussian posteriors, and soften the claim in the conclusion that "only our samplers perform well in all of them".
AuthorsOpenReview ↗

Follow-up response

Thank you for the constructive follow-up. We agree that no test case contained a setting where Gaussian methods are truly expected to fail, and we followed your first suggestion.

New test: Lorenz 96 with h(x)=∣x∣h(x)=|x|. The system is the same as in the other Lorenz 96 cases, with every 4th site observed. The sign of an observed site is now not observed, so its posterior can have two peaks. The reference is a long particle filter with the exact optimal proposal, which draws both signs with their exact masses.

Results. All settings are taken unchanged from the arctan case, for every method. The table gives the time-averaged distance to the reference with 64 members and M=25M=25 sampler steps, as mean ± standard deviation over five test trajectories (lower is better).

MethodSliced W2W_2kNN-KL
Ours, SI-SDE, shared + corr.1.25 ± 0.133.46 ± 0.25
Ours, DM-SDE, shared + corr.1.24 ± 0.083.67 ± 0.19
Ours, FM-ODE, shared + corr.1.22 ± 0.063.52 ± 0.17
Ours, SI-SDE, Jac-free + corr.1.30 ± 0.083.19 ± 0.19
Ours, DM-SDE, Jac-free + corr.1.26 ± 0.063.60 ± 0.18
Ours, FM-ODE, Jac-free + corr.1.32 ± 0.053.49 ± 0.13
EnKF (localised)1.29 ± 0.034.13 ± 0.19
Particle filter2.25 ± 0.195.51 ± 0.42
FlowDAS1.64 ± 0.093.83 ± 0.23
FlowDAS + SURGE2.76 ± 0.136.69 ± 0.27
Guided FM (FIG)1.66 ± 0.073.79 ± 0.15
D-Flow SGLD1.89 ± 0.065.52 ± 0.12
SDA1.95 ± 0.035.87 ± 0.20
SDA + SURGE3.09 ± 0.197.69 ± 0.46
Reference, 64 draws0.71 ± 0.012.28 ± 0.05

Conclusion. With the correction, all six variants of our sampler have a lower kNN-KL than the EnKF, by 11–23%, and the shared variants also have a lower sliced W2W_2. They are the only generative methods at or ahead of the EnKF in both metrics. The other generative baselines are 27–51% behind it in sliced W2W_2, and the 64-member particle filter degenerates. The results change little for MM up to 250. The gain in sliced W2W_2 is within one standard deviation, so we see this as a clear but moderate improvement in a two-peaked setting.

Given the time constraints, we did not have time to add this to the paper before the end of the Paper Discussion Phase. We will make sure to add it and its results to the camera-ready version, together with visualisations of the posterior marginals that illustrate the multimodality.

Reviewer_4zVjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper uses pretrained Stochastic Interpolants (SI), Flow Matching (FM), or Diffusion Models (DM) as the prior p(x1∣x0)p(x_1\mid x_0) and sample one-step posterior for data assimilation. The paper shows that wll of them use the same likelihood guidance term. They differ only in the diffusion term in the weight. To approximate the likelihood score, the paper interpolates the observations and fits a Gaussian for it. The mean and covariance come from the prior and the Jacobian. The paper also tries to reduce cost by offering a Jacobian-free version and another version that shares one Jacobian across the ensembel. Experiments cover a 2D linear-Gaussian case, a Navier-Stokes case, and a 2D slice of urban LES.

Critical strengths
  • Clearly writings. The proofs are self contained and easy to check, and the single weight links the guidance strength to the diffusion of the sampler.
  • No separate guidance step size to tune (this is great!). Only the damping λ\lambda is tuned, though results seems still depends on the approximation of covariance. The method give best numerical results when given dense observations. The Jacobian-free version can be 10x cheaper than the shared-Jacobian version.
  • The experiemnts are broad. Many generative baselines plus EnKF and particle filters.
Critical weaknesses
  1. The likelihood appriximation matches existing moment-base guidance methods. I find two points here:

    • Exact level. For fixed x0x_0 and βτ≠0\beta_\tau\neq0, yˉτ\bar y_\tau is an invertible affine map of yy. So p(yˉτ∣xτ,x0)=βτ−Nyp(y∣xτ,x0)p(\bar y_\tau\mid x_\tau,x_0)=\beta_\tau^{-N_y}p(y\mid x_\tau,x_0), and both have the same score in xτx_\tau. Interpolating the observation does not change the exact score.
    • Gaussian level. From Eq. 22, x^1:=E[x1∣xτ,x0]=(xτ−μτs−ατx0)/βτ\hat x_1:=\mathbb E[x_1\mid x_\tau,x_0]=(x_\tau-\mu^{\mathrm s}_\tau-\alpha_\tau x_0)/\beta_\tau and Cov(x1∣xτ,x0)=Στs/βτ2\mathrm{Cov}(x_1\mid x_\tau,x_0)=\Sigma^{\mathrm s}_\tau/\beta_\tau^2. Putting them into Eq. 16, the βτ\beta_\tau factors cancel and we get (∇xτx^1)⊤H⊤(R+H,Cov(x1∣xτ,x0)H⊤)−1(y−Hx^1)(\nabla_{x_\tau}\hat x_1)^\top H^\top\big(R+H\\,\mathrm{Cov}(x_1\mid x_\tau,x_0)H^\top\big)^{-1}(y-H\hat x_1). This is just an exact rewrite of Eq 16, with the covariance held fixed as in the paper.

    To me, this is the same formula as TMPD and MMPS. They use the Tweedie mean and covariance and they also both hold the covariance fixed when taking the gradient. I believe these work should be discussed somewhere in this manuscript. So, the "bias-corrected mean" and the "covariance inflation" are actually the Tweedie mean and covariance. So, the gains beyond the naive likelihood N(Hxτ,βτ2R)\mathcal N(Hx_\tau,\beta_\tau^2R) in section 4.2 comes from these moments, rather than from the interpolation. This conflicts with 'the method differs from methods that use the original likelihood at a single estimated terminal state' statement in the section 2, because the math matchs.. So, the theorem 4.1 about 'the first unified posterior-sampling result' also seems too strong to me... It would be great if authors can explain what the obs interpolant adds beyond applying known moment based guidance work to more other model types/styles.

@misc{ boys2024tweedie, title={Tweedie Moment Projected Diffusions for Inverse Problems}, author={Benjamin Boys and Mark Girolami and Jakiw Pidstrigach and Sebastian Reich and Alan Mosca and Omer Deniz Akyildiz}, year={2024}, }

@inproceedings{ rozet2024learning, title={Learning Diffusion Priors from Observations by Expectation Maximization}, author={Fran{\c{c}}ois Rozet and G{'e}r{^o}me Andry and Francois Lanusse and Gilles Louppe}, booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems}, year={2024}, }

  1. The experiments can not separate the effect of the sampler and the sparse-senser results seem mixed. The three samplers in theory target the same posterior but in practice, they can differ a lot. In table 1, the DM-SDE and FM-ODE use the same FM prior so they are close. The big gap is between SI-SDE and the two FM-based samplers. But the SI and FM priors differ in many ways, e.g., network depth, training target, schedule, source distribution, etc. To me, it is hard to say how much of the gap comes from which. Besides, the sparse-sensor results are mixed compared to baselines. The SDA+SURGE has lower RMSE at 5% and 1/641/64 while D-Flow SGLD has the best 1/641/64 RMSE and CRPS, but it costs about 17×17\times more than SI-SDE in terms of s/step... I notice there are only 5 NS trajectories used, and there is no spread across trajectories. As for the urban case, the table 2 claims 4-5 held-out trajs, but I notice there is one held-out trajectory claimed in Appendix I. It would be great if the results can consider adding corresponding parts discussed above.
  2. The exact one-step sampling cannot give a exact filter. The paper shows the exact sampling in theorem 4.1, which is p(un∣yn,un−1)p(u^n\mid y^n,u^{n-1}). This indeed holds in its own setting. But the experiments run Algo 1, each member is drawn from uin∼p(un∣yn,uin−1)u_i^n\sim p(u^n\mid y^n,u_i^{n-1}), and all previous members have equal weight. The weights p(yn∣un−1)p(y^n\mid u^{n-1}) are missing, so the ensemble cannot follow the filtering posterior p(un∣y1:n)p(u^n\mid y^{1:n}), even though we have exact every step. Let's consider a scalar example. Let un−1∣y1:n−1∼N(m,P)u^{n-1}\mid y^{1:n-1}\sim\mathcal N(m,P), un=un−1+ξu^n=u^{n-1}+\xi and yn=un+ηy^n=u^n+\eta, with ξ∼N(0,Q)\xi\sim\mathcal N(0,Q) and η∼N(0,R)\eta\sim\mathcal N(0,R). The procedure use the gain Q/(Q+R)Q/(Q+R), but the KF uses (P+Q)/(P+Q+R)(P+Q)/(P+Q+R). When P=Q=R=1P=Q=R=1, this procedure gives N(m+yn2,34)\mathcal N\big(\tfrac{m+y^n}{2},\tfrac34\big), however, the true posterior is N(m+2yn3,23)\mathcal N\big(\tfrac{m+2y^n}{3},\tfrac23\big). Here in this example, the ensemble gives too little weight to the observation and has too much spread. They agree only when P is 0. The limitations section talked about this, but the paper presents the method as Bayesian DA and the only test against an exact posterior is one step with H=R=I and a Gaussian prior, where the Gaussian approx is anyway exact. I would recommend the authors to state what distribution is the sequential ensemble following and when the gap is small (e.g., a multi-step linear Gaussian test against the KF).
Questions for authors
  1. The derivation in Weakness 1 show that Eq 16 is equivalent to the Tweedie-moment Gaussian likelihood, i.e., N(Hx^1,R+H,Cov(x1∣xτ,x0)H⊤)\mathcal N(H\hat x_1,R+H\\,\mathrm{Cov}(x_1\mid x_\tau,x_0)H^\top), evaluted at y. Can authors confirm if this is correct? If yes, what does the observation interpolant add beyond applying TMPD or MMPS style guidance to more generative model types/styles? Could you consider compare the method to their guidance with the same prior?
  2. In the sparse setting in table 1, how much of the gap between SI-SDE and FMs comes from the sampler its own, and how much from the different priors (e.g., network arch, schedule, source, objective, etc) and from the Jacobian refresh rate? Could you consider add the per-traj spread for table 1 and table 2 and report how many urban trajs are used?
  3. In the case that each one-step sample is exact, what is the distribution that the sequential ensemble of Algo. 1 following? Can you please perform a multi-step linear Gaussian test?
4 comments
AuthorsOpenReview ↗

Official response (1/2)

Thank you for the detailed review. The derivation in Weakness 1 and the scalar example in Weakness 3 are both correct and have shaped the revision. Please see the general response at the top for an overview of all changes.

W1 / Q1: "Eq 16 is equivalent to the Tweedie-moment Gaussian likelihood ... Can authors confirm if this is correct?"

Yes, at both levels (your Eq. 16 is Eq. 15 in the revision). Section 4.1 now states that yˉτ\bar y_\tau is an invertible affine map of yy with the same likelihood score, so the interpolation introduces no approximation. The new Proposition 4.4 follows the argument you outline: the factors of βτ\beta_\tau cancel and the score is the TMPD/MMPS score with the covariance held fixed. We removed the Section 2 sentence you quote and no longer call the result the "first unified posterior-sampling result".

We still regard two things as contributions: the single statement of the guided sampler across SI, FM and diffusion models, and the exact likelihood moments for general schedules that come before the Gaussian closure (Lemmas 4.2 and 4.3). The "bias-corrected mean" and "covariance inflation" name these moments. As you say, they are given by the Tweedie mean and covariance, and with the Gaussian closure the guidance equals that of TMPD and MMPS, which Proposition 4.4 now states.

"What does the observation interpolant add beyond applying TMPD or MMPS style guidance to more generative model types?"

It offers an alternative viewpoint on guidance: the observation is transported along the same path as the state, so the likelihood is treated in the interpolant's own variables. This adds no new approximation. It gives a derivation in which the score identity and the likelihood moments are exact for general schedules, in terms of a learned drift or velocity, with the Gaussian closure as a separate and replaceable step. The equivalence holds only at that step. The framework also adds two pieces that none of the related works has: covariance approximations that avoid the per-particle Jacobian, and an EnKF-type correction for the filtering posterior (general response, point 1).

"Could you consider compare the method to their guidance with the same prior?"

We added TMPD on the same FM prior, with the same diffusion and guidance weight as our DM-SDE sampler (Appendix G). We chose TMPD over MMPS because the MMPS paper focuses on training a diffusion prior from noisy and sparse data, not on adding guidance to an already trained model, and the two share the same likelihood and differ only in the linear solve. TMPD has no hyperparameters to tune, and for the field cases we use its row-sum covariance, as the original paper recommends.

Ours (DM-SDE, shared)TMPD
Multistep linear–Gaussian, KL ×10−3\times10^{-3}62.261.9
Navier–Stokes, true IC, RMSE ×10−1\times10^{-1} (16216^2 / 32232^2)2.16 / 0.642.22 / 0.69
Navier–Stokes, poor IC, RMSE ×10−1\times10^{-1} (16216^2 / 32232^2)7.03 / 1.659.07 / 3.62 (one 32232^2 run diverged)
Navier–Stokes, sparse sensorsstablediverged (all runs at 5%, some at 1.563%)
Lorenz 96, linear, kkNN-KL6.927.76
Urbanstablediverged (all runs)
Cost, Navier–Stokes poor IC (s/step)79–134about 270

The two agree on the multistep test and in super-resolution from the true initial condition, as Proposition 4.4 predicts. Elsewhere TMPD is behind or unstable in our implementation. In the field cases this is likely due to its row-sum approximation of the covariance, although we have not isolated the cause. For completeness, our SI-SDE sampler with the shared covariance also diverged on two of the five poor-initial-condition trajectories at 32232^2.

(Continued in part 2/2.)

AuthorsOpenReview ↗

Official response (2/2)

(Continued from part 1/2.)

W2 / Q2: "how much of the gap between SI-SDE and FMs comes from the sampler its own, and how much from the different priors ... and from the Jacobian refresh rate?"

We can separate two of the three factors (Section 5.3, Appendix J.2).

  • Sampler. DM-SDE and FM-ODE share one prior and stay close from the true initial condition (Tables 1 and 5), so the sampler contributes little there. From the poor initial condition FM-ODE recovers faster.
  • Refresh rate. Refreshing the Jacobian every step nearly closes the gap at 5% sensors (Table 6): RMSE ×10−1\times10^{-1} 1.33 for SI-SDE against 1.47–1.61 for the FM-based samplers at k=1k=1, compared with 2.25 against 4.39–5.26 at k=10k=10. At 1.563% a gap remains (2.89 against 4.55–4.87).
  • Prior. The EnKF on the trained priors lets us compare them directly. With the same analysis step and no guidance, the two priors perform alike in all four true-initial-condition scenarios and on Lorenz 96, within one standard deviation in each case. Under sparse sensors, for example, the RMSE ×10−1\times10^{-1} is 3.00±0.463.00\pm0.46 and 6.51±0.786.51\pm0.78 for SI against 2.86±0.432.86\pm0.43 and 6.21±0.656.21\pm0.65 for FM (Table 5), so FM is if anything slightly ahead. With guidance, SI-SDE is far ahead in the same scenarios (1.73 against 3.42–3.83 at 5%). This indicates that the gap does not come from the forecast quality of the priors, and that it arises in how the guidance acts on each model.

The results do depend on the covariance approximation, as you note. Which variant is better is problem-dependent, and the conclusion now says so.

"the sparse-sensor results are mixed compared to baselines"

They are. At 5%, SDA + SURGE has slightly lower RMSE (1.62±0.161.62\pm0.16 against 1.73±0.201.73\pm0.20 for our SI-SDE with the shared covariance), and ours has the best CRPS. At 1.563%, D-Flow SGLD and SDA + SURGE are ahead on RMSE (2.31 and 2.97 against 3.85), with D-Flow SGLD at about 700 s per step against 42 s. Section 5.3 now states this. Here, it is important to note that our samplers generally perform well across all cases, while the baselines perform well on individual cases and not throughout.

"only 5 NS trajectories used, and there is no spread across trajectories ... report how many urban trajs are used"

Both cases use five held-out trajectories, and all tables now report the standard deviation over them. The "one trajectory" in the appendix was a leftover: the field figures show one trajectory and the error curves are means over the five.

W3 / Q3: "what is the distribution that the sequential ensemble of Algo. 1 following? Can you please perform a multi-step linear Gaussian test?"

Each member is drawn from its own one-step posterior, so the ensemble follows the equal-weight mixture of these posteriors, while the filtering posterior weights them by p(yn∣uin−1)p(y^n\mid u^{n-1}_i). As in your example, the per-member update applies the gain KQK_Q of the transition noise instead of the Kalman gain. The gap vanishes when all members share the same forecast mean, is small when the transition noise dominates the spread of the forecast means, and does not shrink with more members or sampler steps (Sections 3 and 4.4, Appendix D).

The new 20-cycle linear–Gaussian test (Section 5.1) confirms your analysis: uncorrected, our samplers level off at a KL of 6⋅10−26\cdot10^{-2} or more, whatever the number of sampler steps. With the new forecast-residual correction (Eq. 18) and the shared covariance, SI-SDE reaches 6.5⋅10−56.5\cdot10^{-5}, against 5.5⋅10−55.5\cdot10^{-5} for the EnKF. Existing remedies struggle more in our tests: weight-based resampling and SURGE often collapse the ensemble with 64 members. Our approach balances the efficiency of an EnKF-type update with the non-Gaussian and nonlinear update that the guidance provides. While the correction improves results in many of the cases, it is still not perfect. With the Jacobian-free covariance it raises the KL on this test. On Navier–Stokes it helps from a poor initial condition but raises the errors from the true one, likely because a linear, Gaussian update degrades the non-Gaussian samples of an ensemble that already tracks the truth (Section 5.3).

Thank you for the detailed response and the revised paper. Your effort is really appreciated!

Your rebuttal addressed some of my concerns, however, some still remains:

  1. You mentioned 'the method matches TMPD and MMPS only where the closure is exact', however, the guidance (Eq 15) seems still always equal to the TMPD/MMPS. The MMPS, which always used for DA (at least to me...), differs from the proposed solver only in the linear solve, and the main pratical claim, to me, is the cheaper covariance. So, I still think it should be a key baseline, and it is missing...
  2. The Eq 18 is derived for Gaussian. NS and urban do not satisfy... The guidance makes the errors bigger in all cases from the true IC, and diverges in the urban one. You mentioned something like 'Gaussian update breaks non-Gaussian samples' but is it possible that the forecast means are close, from the true IC, so Ktextens−KQK{text{ens}}-KQ becomes near zero?
  3. From the results, the best variant is always changing, e.g., shared under the sparse obs; jacb-free in the poor IC, and it seems hard to pick the one ahead of time... So the claim 'Only our samplers perform well in all...' seems too strong to me.

Overall, the paper is more clear and honest compared to its initial version. Though some wordings are corrected and the new prior+EnKF shows that the guidance can work, well, the core approximation is somewhat known, the new correction shows a mixed results, and the key (personally speaking...) baseline (compare to MMPS) is still missing. At this stage, I may want to maintain my original assessment.

AuthorsOpenReview ↗

Follow-up response

Thank you very much for taking the time to read the revision so carefully. Your remaining points are fair, and we address them in order.

1. MMPS as a baseline. You are right that MMPS itself is not among the baselines, and we will make sure to add it to the camera-ready version.

We chose TMPD for the revision because MMPS and TMPD use the same Gaussian likelihood, and TMPD was developed for guiding an already trained model, which is our setting. MMPS was introduced as the sampler inside a scheme that trains a prior from noisy and sparse data rather than guiding an already trained one.

On the multistep test and Lorenz 96, TMPD is run with the full covariance and an exact solve, which is the solution that the conjugate-gradient solve of MMPS converges to. There it agrees with our sampler (KL 61.9⋅10−361.9\cdot10^{-3} against 62.2⋅10−362.2\cdot10^{-3} on the multistep test).

We would also like to clarify that our variants differ from both in more than the linear solve. TMPD and MMPS need Jacobian products of the network for every member at every step. The Jacobian-free variant needs none, and the shared variant evaluates them once for the whole ensemble.

2. The correction. Eq. 18 uses only the mean and covariance of the forecast, as the EnKF does (Appendix D). It is exact for a Gaussian forecast and an approximation otherwise.

Thank you for the suggestion on the true initial condition. We find it very plausible and will add it to the paper. When the forecast means are close, Kens−KQK_{\mathrm{ens}}-K_Q should be near zero. With a finite ensemble, however, it is the difference of two estimated gains, so its sampling error remains and is added to samples that were already accurate. The correction is therefore optional, and meant for ensembles whose forecast means are spread out. That is where it helps:

  • Poor initial condition (Navier–Stokes): the corrected FM-ODE sampler has the lowest error at 16216^2.
  • Lorenz 96 with h(x)=∣x∣h(x)=|x| (new, two-peaked posterior): without the correction our samplers are 25–32% behind the EnKF in sliced W2W_2. With it, all six variants are on par with or ahead of the EnKF in sliced W2W_2 and have a lower kNN-KL, by 11–23% over five trajectories. The correction thus improves the samplers even though the posterior is strongly non-Gaussian. None of the generative baselines reaches the EnKF in sliced W2W_2 (27–51% behind) and the majority are behind in kNN-KL. The table is in our follow-up to Reviewer GFyC.

3. Choice of variant. We agree, and will reword the claim to: ours is the only method with a variant that performs well in every case. Section 4.2 gives a starting rule: the Jacobian-free covariance for dense observations and the shared one for sparse sensors. It holds in the sparse-sensor cases, where the shared covariance is 2–4 times more accurate. It is not a complete rule, however. From a poor initial condition the Jacobian-free variant is the better choice, likely because the approximated Jacobian is unreliable when the ensemble is far from the truth.

We see it as a strength of the framework that it provides variants for a wide range of scenarios. Making the best choice automatically (online) is left for future work, and we will state both the rule and this limitation in the paper.

We hope this addresses your remaining concerns, and we are grateful for a discussion that has clearly improved the paper.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response to all reviewers (1/2)

We thank the four reviewers for their careful and constructive reading. The reviews share three main points: the overlap with Tweedie-moment guidance (TMPD, MMPS), the gap between the one-step and the filtering posterior, and the limited evidence beyond the linear–Gaussian setting. We have revised the paper on all three.

Section, table, figure and equation numbers refer to the revised PDF unless they appear inside a quoted reviewer comment. Changes in the revised PDF are coloured by the reviewer who prompted them:

  • Blue (4zVj): the one-step versus filtering posterior, the forecast-residual correction (Section 4.4, Appendix D), the multistep linear–Gaussian test, and the equivalence with TMPD and MMPS (Proposition 4.4).
  • Orange (GFyC): the Lorenz 96 experiment, TMPD and the EnKF on the trained prior as baselines, and notation and metric clarifications.
  • Green (tGUh): nonlinear observations (Section 4.3), Navier–Stokes from a poor initial condition, the localised EnKF, and the related-work discussion in Section 2.
  • Purple (dcLZ): the reframing of the contribution in the abstract, introduction and contribution list.

Several comments overlap between reviewers. Such a change carries one colour only, so a passage in one colour may also address other reviewers.

1. Relation to TMPD and MMPS

The reviewers are right. The interpolated observation is an invertible affine map of the observation, so its likelihood score is the original one, and under the Gaussian approximation our guidance equals that of TMPD and MMPS. Section 4.1 now states that the interpolation introduces no approximation and serves to give the exact likelihood moments. Section 4.2 states that tractability comes from the Gaussian approximation, and the new Proposition 4.4 proves the equivalence. Section 2 positions the method against DPS, ΠGDM, TMPD, MMPS, APPA, DAISI, FA-APF and FIG. The claim "first unified posterior-sampling result" is removed. The "bias-corrected mean" and "inflated covariance" name the exact likelihood moments, which are given by the Tweedie mean and covariance.

What is new. Each of the closest works covers part of the picture. TMPD and MMPS give the Gaussian likelihood for static inverse problems. DAISI derives the guided drift for Gaussian-source interpolants. FA-APF adds particle weights for filtering and derives the point-mass-source case in an appendix, without testing it. None covers all of it, and two pieces are in none. Our framework brings these together and adds the missing pieces:

  1. One statement covering Gaussian and point-mass sources and both SDE and ODE samplers, with a single guidance weight. The paper also gives a side-by-side evaluation of the three samplers on common test cases and experiments with moment-based guidance for a point-mass-source interpolant, neither of which we have seen reported.
  2. A derivation that separates the exact from the approximate. The score identity and the first two likelihood moments are exact, for general schedules and directly in terms of a learned drift or velocity. The Gaussian closure is one explicit choice made afterwards. The equivalence with TMPD and MMPS arises only at that last step, so the closure can be replaced without touching the rest.
  3. Covariance approximations that none of TMPD, MMPS, DAISI or FA-APF has. All four need Jacobian products of the denoiser for every particle at every step. Our Jacobian-free variant needs no Jacobian, and the shared variant evaluates it once for the whole ensemble. This addresses the cost and memory overhead of differentiating the denoiser, which FA-APF notes as a practical limitation.
  4. An EnKF-type correction for the filtering posterior, which needs no importance weights and runs with 64 members. FA-APF uses tempered particle weights, which its authors note introduce a bias, and DAISI leaves a principled correction to future work.
TMPD / MMPSDAISIFA-APFOurs
Generative modeldiffusionstochastic interpolantdiffusion (interpolant derived, not tested)stochastic interpolant, flow matching and diffusion, all tested
SourceGaussianGaussianGaussian (point mass derived, not tested)Gaussian and point mass, both tested
Guidance covarianceJacobian products per particleJacobian products per particleJacobian products per particleJacobian-free, or one Jacobian per ensemble
Filtering posteriornot addressed (static problems)no correction (future work)tempered particle weightsEnKF-type correction
Nonlinear observationsTMPD: linear onlylinearisedlinearisedlinearised, tested

(Continued in part 2/2.)

AuthorsOpenReview ↗

General response to all reviewers (2/2)

(Continued from part 1/2.)

Comparison with TMPD. We ran TMPD on the same prior, following the original paper: for the field cases we use its row-sum covariance (DTMPD). In Navier–Stokes super-resolution from the true initial condition it matches our DM-SDE sampler, as Proposition 4.4 predicts. On the same GPUs (poor-initial-condition runs) it takes about 270 s per step, against 8–14 s (Jacobian-free) and 45–134 s (shared) for ours. In our implementation it was unstable under sparse sensors and in the urban case.

2. One-step versus filtering posterior

Theorem 4.1 concerns each member's one-step posterior, and the paper now says "one-step" where the result is stated and used. The new Section 4.4 and Appendix D state what the sequential ensemble follows: an equal-weight mixture of one-step posteriors, whereas the filtering posterior weights them by the predictive likelihoods. We add an EnKF-type forecast-residual correction (Eq. 18) that supplies the missing cross-member part of the update. We chose an EnKF-type correction over importance weights because weight-based resampling often collapses the ensemble when the ensemble is small. Our experiments confirm this: with 64 members, SURGE collapses the ensemble on Navier–Stokes and raises the error on Lorenz 96.

3. New experiments

  • A multistep linear–Gaussian test against the Kalman filter (Section 5.1).
  • Stochastic Lorenz 96 against a 4⋅1064\cdot10^6-particle filter, with linear and arctan observations (Section 5.2).
  • Navier–Stokes from a poor initial condition over 35 steps (Section 5.3).
  • New baselines: TMPD, an EnKF analysis on the same trained prior, a localised EnKF, and SURGE around our own samplers.
  • Standard deviations, arrows, and bold/underline ranking in all results tables, and five urban trajectories, stated consistently.

4. Main findings

  • Multistep. Uncorrected, our samplers level off at a KL of 6⋅10−26\cdot10^{-2} or more, however many steps are used: the bias Reviewer 4zVj described. With the shared covariance, the correction brings it to 6.56.5–12⋅10−512\cdot10^{-5}, close to the EnKF (5.5⋅10−55.5\cdot10^{-5}). With the Jacobian-free covariance it raises the KL.
  • Lorenz 96. The corrected samplers are on par with the 64-member EnKF, and so is an EnKF analysis on the same trained prior. The posterior is only mildly non-Gaussian, so this validates the samplers but shows no gain over the EnKF.
  • Navier–Stokes. In super-resolution from the true initial condition our shared samplers have the best RMSE and CRPS, with TMPD on par, and are ahead of the other generative baselines and the 64-member EnKF. At 32232^2 they are also ahead of the 1000-member EnKF. Under sparse sensors the picture is mixed. From a poor initial condition our FM-ODE sampler recovers best in RMSE and CRPS, ahead of a localised EnKF with the true solver. The correction helps every sampler from the poor initial condition at 16216^2 but only two of six at 32232^2. From the true initial condition it raises the errors of every sampler. The correction is a linear, Gaussian update, so it helps an ensemble that is far from the truth, but it likely degrades the non-Gaussian samples of an ensemble that already tracks it.
  • Urban. TMPD and the correction diverged in every run. For the correction, a likely cause is that its localised update does not mask the building cells, so it perturbs cells that contain no flow and pushes the state too far away from the training distribution. Our uncorrected samplers are stable, and the best of them is best or tied in every column: DM-SDE on velocity and SI-SDE on temperature. D-Flow SGLD is equally accurate at about 360 s per step against 20–84 s.

The conclusion and limitations reflect these findings, including where the method does not help.

5. Scope of the new experiments

  • Related methods. TMPD is run on the same prior in all four test cases. It stands in for MMPS, which shares its likelihood and differs only in the linear solve.
  • Classical filters. We use a stochastic EnKF, localised and tuned on Lorenz 96 and localised for Navier–Stokes from the poor initial condition, together with an EnKF on the trained prior and a 1000-member reference. An LETKF could not be run within the rebuttal period, and we will gladly add it in a camera-ready version.
  • SI and FM priors. The priors are trained separately. The refresh-rate ablation separates part of the gap between the samplers, and the EnKF on the trained priors indicates that their forecast quality is similar.
  • Poor initial condition. This experiment covers the two super-resolution scenarios, with every sampler variant and baseline.
  • Hardware. All runs are complete. To finish within the rebuttal period, some were moved to different hardware. Their timings are marked †^{\dagger} and are indicative only.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.