acceptodds
Under review as a conference paper at ICLR 2027

FREESIA: COVARIANCE-AWARE POSTERIOR TRANSPORT FOR EXPRESSIVE AND SCALABLE DATA ASSIMILATION

Abstract

Data assimilation aims to infer the state of complex dynamical systems based on observational data. However, accurate inference of the multimodal posteriors induced by nonlinear or non-injective observation operators remains a key challenge under high-dimensional and sparse observation conditions. Ensemble filters scale to high dimensions but are confined by restrictive distributional assumptions, while training-free generative filters (e.g., EnSF, EnFF) alleviate this limitation but may introduce structural errors and hinder information propagation under sparse observations. To address these issues, we propose a training-free, asymptotically exact posterior transport method. Firstly, a covariance-aware posterior transport scheme is designed, which embeds the forecast cross-covariance into flow-based transport and accurately recovers unobserved states while preserving the non-Gaussian posterior structure. Furthermore, the method combines a tractable observation-adaptive proposal with posterior correction, ensuring accurate approximation of the nonlinear posterior distribution. Finally, we establish the corresponding posterior flow theory, from which the asymptotic exactness of the proposed method relative to finite-ensemble surrogates and the Wasserstein error bound are derived. Experiments on Double-Well, Lorenz–96, and Kolmogorov flow show that the proposed method captures complex posterior structure and remains accurate under sparse, nonlinear, and non-injective observations. In the sparse non-injective setting, it reduces RMSE by 56% relative to the best baseline.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Data Assimilation, Nonlinear Filtering, Posterior Transport, Flow Matching, Ensemble Methods, Training-Free

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.20 over 5 reviews, above 49% of the others.
Reviewer_61oVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a training-free data assimilation method using sparse observations to correct model predictions. It constructs a Gaussian mixture distribution from model predictions and measures the covariance between observed and hidden state variables to update states that are not need to recover. This method first uses a linear observation model to sample candidate states, then uses the true nonlinear likelihood to correct their weights. These weighted samples are used to estimate a velocity of the flow matching from Gaussian distribution to posterior distribution.

The experiments on Double-Well, Lorenz–96, and Kolmogorov flow prove the method's ability in capturing multiple posterior modes and improves the estimation of unobserved states and a significant improve from baselines.

Critical strengths
  1. The method combines covariance-aware transport with nonlinear likelihood correction. By utilizing this, the method allows sparse observations to update unobserved states while preserving complex posterior distributions, while previous approaches like Ensemble Kalman Filter, rely mainly on mean and covariance information and may fail to capture multiple posterior modes.

  2. The paper provides detailed proofs and derivations to support the proposed method.

Critical weaknesses
  1. The paper are not clearly presented. The main figures provide limited information and seems have some errors, this need to double check before submission.

  2. The dynamical model is also important for data assimilation, but the paper does not clearly show its prediction accuracy before the assimilation update. The EnKF results of Kolmogorov flow in Figure 2 appear noisy, and this may indicates implementation issues whether the update improves the model predictions or instead degrades them. Several baselines perform unexpectedly poorly. The large performance gaps need further explanation and validation to establish a fair comparison.

Questions for authors

How large is the prediction error when running the dynamical model without data assimilation?

If both the ground-truth data and the dynamical model predictions are generated using the same numerical solver, how are differences between their trajectories are introduced in experiments?

Can FREESIA also work when the governing equations are unknown and predictions come from learned / black-box dynamical models?

3 comments

Response to Reviewer 6IoV

We thank the reviewer for the constructive comments and suggestions regarding the numerical experiments and the applicability of FREESIA. We provide point-by-point responses below.

Comment 1.

How large is the prediction error when running the dynamical model without data assimilation?

Response:

We additionally evaluated the prediction error without data assimilation in the 1,000-dimensional Lorenz-96 experiment under stride-4 observations.

Without assimilation, the RMSE is 3.7248. Both the classical ensemble filters and FREESIA substantially reduce this prediction error by incorporating observational information. For example, EnKF-MDA achieves RMSE 2.056, while FREESIA further reduces it to 1.575, demonstrating the effectiveness of data assimilation in correcting model predictions.

Comment 2.

If both the ground-truth data and the dynamical model predictions are generated using the same numerical solver, how are differences between their trajectories are introduced in experiments?

Response:

The ground-truth trajectory and forecast ensemble are initialized from different states. Even when propagated using the same governing equations and numerical solver, different initial conditions lead to distinct trajectories, particularly in chaotic systems such as Lorenz-96 and Kolmogorov flow.

In Lorenz-96, both the truth and initial forecast ensembles are sampled from the post-spin-up chaotic regime. The truth uses a time step of 0.01, while the forecast uses 0.02.

For Kolmogorov flow, the truth and forecast ensembles are similarly initialized from the developed-flow regime after spin-up. These different initial states naturally lead to diverging trajectories during model propagation.

Comment 3.

Can FREESIA also work when the governing equations are unknown and predictions come from learned / black-box dynamical models?

Response:

Yes. FREESIA only requires a forecast ensemble generated by the prediction step and an evaluable observation operator. The forecast ensemble is used to construct the Gaussian-mixture prior surrogate, from which the posterior transport is derived.

Since the analysis step does not require derivatives or an explicit formulation of the dynamical model, the forecast ensemble can also be generated by learned or black-box prediction models. Therefore, FREESIA can be directly integrated with data-driven forecasting frameworks.

Thanks for the responses and additional experiments. While the problem in figure 2 and results of Kolmogorov flow is still not solved. To best of my understanding, the prediction errors are introduced by using different setting in numerical solver, and the EnKF results in Kolmogorov flow indicates that data assimilation may even disrupt the underlying dynamics rather than improve the predictions.

We thank the reviewer for the follow-up comments.

Comment 1.

To best of my understanding, the prediction errors are introduced by using different setting in numerical solver.

Response:

In our experiments, the truth and forecast ensembles are initialized from different states in the post-spin-up dynamical regime. In chaotic systems, these initial-state differences naturally lead to diverging trajectories, even when the same dynamical model and numerical solver are used.

While Lorenz-96 also uses different numerical time steps for the truth and forecast, initial-state uncertainty is itself sufficient to generate prediction errors and motivate data assimilation.

Comment 2.

The EnKF results in Kolmogorov flow indicates that data assimilation may even disrupt the underlying dynamics rather than improve the predictions.

Response:

We have added comprehensive final-time comparisons of all seven methods under both observation regimes in Figures 18 and 19 of the revised manuscript.

Under 1% monotone observations, EnKF-MDA and LETKF successfully recover the dominant vorticity structures, achieving RMSEs of 0.148 and 0.218, respectively. EnKF also recovers parts of the flow, with an RMSE of 0.377.

The 2% non-injective observation experiment is considerably more challenging because the absolute-arctangent operator introduces sign ambiguity in the observed velocity components. This ambiguity poses particular difficulties for covariance-based linear updates.

Moreover, vorticity involves spatial derivatives of the velocity field, which amplify local estimation errors. The displayed vorticity fields characterize the accuracy of the final analysis rather than the accuracy of subsequent forecasts. The latter is a separate quantity from the reconstruction accuracy presented in these figures.

We have also revised the method illustrations and added a complete algorithm overview in Figures 1 and 2 to improve the clarity of the presentation.

Reviewer_FTwNOpenReview ↗
Rating
2
Confidence
2(−2)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces FREESIA, which combines covariance information, conditional sampling, and importance sampling for training-free data assimilation. The experimental results are promising, but the methodological novelty and the effects of certain computational approximations require further clarification.

Critical strengths

• The method performs well under sparse, non-injective observations, and the Double-Well experiment directly evaluates posterior approximation quality.

• The ablation experiments support the effectiveness of cross-covariance and likelihood correction.

Critical weaknesses

• The method uses established techniques, including Gaussian mixture models, covariance-based updates, and importance sampling. Its new contributions relative to existing methods need to be clarified.

• The high-dimensional experiments use a diagonal approximation when calculating component selection probabilities, ignoring some correlations. Appendix A.11 shows that likelihood correction cannot eliminate the resulting bias, but an empirical assessment is missing.

• Figure 13 shows highly concentrated importance weights over most of the flow, while the main experiments use only five candidate samples per estimation. Further validation is needed to determine whether this sample count is sufficient.

• The absence of experimental comparisons with IEnSF and EnSF-LR limits the assessment of the method’s advantages.

Questions for authors
  1. What is the main new contribution compared with existing methods?
  2. How would the results and computation time change without the diagonal approximation?
  3. Do the high-dimensional results remain stable as the number of candidate samples increases?
1 comment

Response to Reviewer FTwN

We thank the reviewer for the careful assessment of our manuscript and the constructive suggestions. We have clarified the methodological contributions and added experiments to evaluate the component-evidence approximation and sampling sensitivity.

Comment 1.

What is the main new contribution compared with existing methods?

Response:

The main contribution of FREESIA is a covariance-aware posterior transport framework that combines an analytically tractable Gaussian bridge with consistent nonlinear posterior conditioning.

By selecting a Gaussian source distribution with the same covariance structure as the forecast surrogate, the conditional endpoint distribution p(Z1∣Zt)p(Z_1\mid Z_t) admits closed-form Gaussian-mixture expressions. The conditional means and covariance are obtained without repeated high-dimensional covariance inversions, while preserving the cross-coordinate correlations required for sparse-observation information transfer.

For nonlinear observations, we construct component-wise Kalman-style endpoint proposals and use exact nonlinear likelihood ratios to consistently estimate E[Z1∣Zt,y]\mathbb{E}[Z_1\mid Z_t,y]. This conditional expectation directly determines the posterior-flow velocity. Theorem 2 establishes the correctness of the proposal-and-correction identity, and Theorem 3 provides surrogate-relative convergence guarantees.

This construction differs from existing training-free generative methods. EnSF uses heuristic posterior-score guidance, while EnFF employs approximate observation guidance in its practical localized implementation. EnSF-LR introduces an additional empirical covariance-regression step to recover unobserved states. IEnSF derives a posterior-score expression under a GMM prior but approximates the conditional expectation of the likelihood score through evaluation at a representative point and iteratively refines this approximation using reverse-SDE sampling.

FREESIA instead estimates the posterior endpoint expectation through a covariance-aware proposal and nonlinear likelihood correction, providing a direct and mathematically consistent route to the surrogate posterior.

We also expanded the numerical comparisons with IEnSF, BPF, and EnGMF-DR on Double-Well, and EnSF-LR and EnGMF-DR on Lorenz-96.

On Double-Well, FREESIA achieves RMSE 0.120 and Wasserstein-1 error 0.048, compared with 0.132 and 0.176 for IEnSF. On Lorenz-96 stride-4, FREESIA achieves RMSE 1.575, compared with 1.676 for EnSF-LR and 2.836 for EnGMF-DR.

Comment 2.

How would the results and computation time change without the diagonal approximation?

Response:

We added a direct comparison between full and diagonal component evidence on dense Lorenz-96 with γ=0.3\gamma=0.3, T=20T=20, and NIS=5N_{\mathrm{IS}}=5.

Full evidence achieves RMSE 0.404, compared with 0.458 under diagonal evidence. The diagonal approximation reduces mean analysis time by 12.8×.

The approximation is applied only to the component-evidence calculation. Both variants retain the localized cross-covariance structure in the conditional endpoint proposals.

For sparse Lorenz-96 with γ=0\gamma=0, all mixture components coincide, so the full and diagonal evidence formulations yield identical normalized component probabilities.

These results demonstrate that diagonal evidence provides a substantial computational speedup while retaining competitive filtering accuracy.

Comment 3.

Do the high-dimensional results remain stable as the number of candidate samples increases?

Response:

We expanded the Lorenz-96 stride-4 sensitivity analysis over T,NIS∈1,5,10,20,50T,N_{\mathrm{IS}}\in\\{1,5,10,20,50\\}.

At T=20T=20, varying NISN_{\mathrm{IS}} from 1 to 50 produces RMSEs between approximately 1.56 and 1.65, indicating stable filtering accuracy across a broad range of candidate counts. Increasing the number of flow steps produces a more pronounced improvement than increasing the candidate count.

The effectiveness of small candidate budgets comes from the observation-adaptive Kalman proposal, which already incorporates observational information through local linearization and forecast cross-covariances. Progressive flow updates then transport the ensemble toward observation-compatible regions, while nonlinear likelihood-ratio correction provides a consistent estimator of the posterior conditional endpoint mean as NISN_{\mathrm{IS}} increases.

The results demonstrate that FREESIA achieves accurate high-dimensional filtering with a small number of endpoint candidates, providing a practical balance between computational cost and posterior approximation accuracy.

Reviewer_q38bOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes FREESIA, a training-free data assimilation method. It uses a flow to transport samples to the posterior of a Gaussian-mixture surrogate built from the forecast ensemble, combining locally linearized Kalman proposals with likelihood-ratio correction. It reports lower RMSE than classical and generative filters on Double-Well, Lorenz-96 and Kolmogorov flow, especially under sparse non-injective observations. The derivation is thorough and the Kolmogorov-flow result is striking. However, the motivation does not state what the closest methods leave unsolved, the method is hard to follow, and IEnSF and EnSF-LR are not compared. The Double-Well statistics also look inconsistent with independent runs, and the scalability and exactness claims need qualifying. I will raise my score if these points are addressed.

Critical strengths
  1. The topic is interesting and practical: filtering from sparse, nonlinear observations matters for weather and climate prediction, and a training-free filter needs no offline data.
  2. The technical motivation is clear: under sparse observations, the likelihood gradient acts only on observed coordinates.
  3. The method is derived and justified thoroughly, with explicit error terms.
  4. The Kolmogorov-flow result is striking: with 2% |arctan(10x)| observations, RMSE falls to about 0.08 while the plotted baselines stay above 0.9.
Critical weaknesses
  1. Motivation: the paper does not say what IEnSF and EnSF-LR leave unsolved, or which real observing system needs non-injective operators. Please state both.
  2. Readability: Figure 1 shows only one particle’s update and no covariance structure. Section 3 never says that Eq. 3.9 is a Kalman update, and the variant actually run appears only in the appendix. Please add a full overview figure and explain the method in terms of known updates.
  3. Baselines: keeping classical filters is fine, but the closest recent methods are missing. IEnSF reports the same setting as stride-2, and its curves give a total RMSE of about 0.68 against 0.615 here, a much smaller lead than the 38% in Table 2. Please compare directly with IEnSF and EnSF-LR, and add a particle filter and a Gaussian-mixture filter.
  4. Table 7 reports RMSE 0.14417 ± 0.00035 over 10 independent trajectories, but a re-simulation gives a standard deviation of about 0.047 even for the exact filter. The 3-run pilot in Figure 5 also nearly equals Table 7.
  5. The scalability result uses a diagonal variant that removes cross-covariance, the high-dimensional runs use γ of at most 0.3, and Theorem 3’s rate does not appear to reach the reported regime. Please qualify these claims.
Questions for authors
  1. What exactly does FREESIA add beyond IEnSF and EnSF-LR, and which real observing system requires non-injective operators?
  2. How does Section 3 relate to known Kalman updates, and what does the flow add over a one-shot update (T = 1, N_IS = 1)? Can you add a full overview figure that includes the variant actually run?
  3. How does FREESIA compare directly with IEnSF and EnSF-LR on their published Lorenz-96 settings, and with a particle filter and a Gaussian-mixture filter?
  4. In the Double-Well experiment, how were seeds and truth trajectories handled? Why are the standard deviations in Table 7 so small, and why does the 3-run pilot in Figure 5 nearly match Table 7?
  5. How much accuracy remains at large scale when cross-covariance is retained, what does the mixture contribute at γ ≤ 0.3, and does Theorem 3’s rate apply to the reported runs?
4 comments

Response to Reviewer q38b

We thank the reviewer for the detailed comments and constructive suggestions. We have clarified the methodological contributions, improved the presentation, and expanded the numerical comparisons in the revised manuscript.

Comment 1.

What exactly does FREESIA add beyond IEnSF and EnSF-LR, and which real observing system requires non-injective operators?

Response:

FREESIA provides a covariance-aware posterior transport framework with two main advantages over existing training-free generative filters.

First, by choosing the Gaussian source distribution to share the forecast-surrogate covariance, Z0∼N(0,Σf)Z_0 \sim \mathcal{N}(0,\Sigma^f), we obtain the closed-form conditional endpoint distribution p(Z1∣Zt)p(Z_1\mid Z_t) in Eq. (3.6). Both the conditional mean and covariance admit explicit expressions involving scalar coefficients of Σf\Sigma^f, avoiding repeated high-dimensional covariance inversions during Gaussian bridge conditioning. This construction also preserves the forecast cross-covariances required for sparse-observation information transfer.

Second, FREESIA estimates the posterior conditional endpoint mean E[Z1∣Zt,y]\mathbb{E}[Z_1\mid Z_t,y] using observation-adaptive Kalman proposals and nonlinear likelihood-ratio correction. Theorem 2 establishes that this proposal-and-correction construction consistently recovers the conditional expectation under the full-evidence formulation.

In comparison, IEnSF approximates the conditional expectation of the likelihood score by evaluating it at a representative conditional mean. This pointwise approximation does not explicitly integrate the variation of the nonlinear likelihood score over the full conditional distribution. IEnSF further employs iterative posterior-score refinement through repeated reverse-SDE sampling. FREESIA instead directly estimates the required conditional endpoint expectation and constructs the posterior transport through a deterministic flow.

EnSF-LR addresses sparse observations by combining EnSF with a covariance-regression update. Its posterior-score guidance inherits the heuristic likelihood approximation of EnSF, while the regression step provides an empirical correction for unobserved variables. FREESIA incorporates the cross-covariance structure directly into the conditional endpoint distribution and derives the observation-conditioned update from a well-defined surrogate posterior target.

We also added a practical example of non-injective observations in the Introduction. Satellite scatterometer measurements of ocean-surface radar backscatter can correspond to multiple plausible wind directions, requiring ambiguity resolution during wind-vector retrieval.

Comment 2.

How does Section 3 relate to known Kalman updates, and what does the flow add over a one-shot update (T = 1, N_IS = 1)? Can you add a full overview figure that includes the variant actually run?

Response:

We clarified after Eqs. (3.9)–(3.10) that the component-wise proposal is an extended-Kalman-style conditioning step. FREESIA uses this proposal to construct observation-adaptive conditional endpoints and progressively transports particles toward the nonlinear surrogate posterior.

We added a direct one-shot importance-sampling comparison using the same Gaussian forecast surrogate, localized covariance, Kalman proposal, and nonlinear likelihood weights.

On Lorenz-96 stride-4, one-shot sampling with 100, 500, and 2000 candidates gives RMSEs of 5.162, 5.168, and 5.143, respectively, compared with 1.575 for FREESIA. All one-shot configurations suffer severe weight degeneracy at the first analysis, leaving only one distinct resampled particle.

Progressive transport substantially improves filtering accuracy while preserving ensemble diversity. Instead of relying on one global importance-resampling update, FREESIA repeatedly constructs local observation-adaptive proposals along the flow, allowing the ensemble to approach the posterior progressively.

We have added a full overview of the FREESIA filtering framework in Figure 1. Figure 2 illustrates the three stages of the particle-level update: observation-adaptive component selection, covariance-aware endpoint sampling, and nonlinear likelihood correction. Algorithm 1 in Appendix A.10 provides the complete ensemble procedure, including covariance construction, flow integration, and the structured-covariance implementation.

Comment 3.

How does FREESIA compare directly with IEnSF and EnSF-LR on their published Lorenz-96 settings, and with a particle filter and a Gaussian-mixture filter?

Response:

We added IEnSF, BPF, and EnGMF-DR to Double-Well. Their RMSEs are 0.132, 0.150, and 0.273, respectively, compared with 0.120 for FREESIA. The corresponding Wasserstein-1 errors are 0.176, 0.086, and 0.273, compared with 0.048 for FREESIA.

For Lorenz-96, we additionally implemented EnSF-LR and EnGMF-DR. At stride-2, their RMSEs are 0.942 and 1.067, compared with 0.615 for FREESIA. At stride-4, they achieve 1.676 and 2.836, compared with 1.575 for FREESIA.

For EnSF-LR, we used the retuned EnSF configuration and added cyclic Gaspari-Cohn localization to the observed-to-unobserved covariance regression, as documented in Appendix B.2.

We also implemented IEnSF for Lorenz-96 and investigated its reverse-SDE integration under the corresponding nonlinear observation settings. The implementation exhibited numerical stiffness and rapid growth during the early reverse-time integration steps, which prevented stable filtering statistics. We therefore report the IEnSF results obtained on Double-Well, together with the EnSF-LR and EnGMF-DR comparisons on Lorenz-96.

Comment 4.

In the Double-Well experiment, how were seeds and truth trajectories handled? Why are the standard deviations in Table 7 so small, and why does the 3-run pilot in Figure 5 nearly match Table 7?

Response:

The unusually small standard deviations in the previous submission arose from the initialization protocol. The true initial states were concentrated near the potential well at x=−1x=-1, resulting in limited variability across independently generated trajectories. This also explains the similarity between the three-run pilot and the original ten-run statistics.

We changed the initialization to X0∼N(0,0.52)X_0\sim\mathcal{N}(0,0.5^2), consistent with the initial filtering distribution, and re-evaluated all methods on ten matched independent truth trajectories.

The revised FREESIA RMSE is 0.12007 ± 0.03112, with the standard deviation computed across trajectories. We also regenerated the baseline comparisons and parameter-sensitivity experiments under the revised initialization.

The updated experiments provide a more representative evaluation of Double-Well dynamics and posterior multimodality.

Comment 5.

How much accuracy remains at large scale when cross-covariance is retained, what does the mixture contribute at γ ≤ 0.3, and does Theorem 3's rate apply to the reported runs?

Response:

The Kolmogorov-flow experiments retain structured cross-coordinate covariances at 32,768 dimensions. With only 2% non-injective observations, FREESIA achieves RMSE 0.422, compared with 0.970 for the strongest evaluated baseline. The corresponding CRPS decreases from 0.569 to 0.209.

The mixture separation parameter γ\gamma controls the contribution of individual forecast ensemble members to the GMM component means. In Double-Well, we use γ=0.9\gamma=0.9 to preserve separated posterior modes and their probability masses. For dense Lorenz-96 and Kolmogorov flow, we use γ=0.3\gamma=0.3 and γ=0.2\gamma=0.2, respectively, maintaining mixture structure while moderating component separation in high dimensions.

For sparse Lorenz-96, we set γ=0\gamma=0, reducing the surrogate to a single Gaussian. FREESIA still achieves RMSE 0.615 under stride-2 observations and 1.575 under stride-4 observations, demonstrating the effectiveness of covariance-aware conditioning and progressive transport independently of mixture separation.

Theorem 3 establishes a one-step Wasserstein-1 error bound that separates finite endpoint sampling, flow discretization, and forecast-surrogate discrepancy:

\leq C_\delta N_{\mathrm{IS}}^{-\delta/(1+\delta)} +C_2T^{-1}+\varepsilon_{\mathrm{sur}}(y).$$ Under the theorem's regularity and moment conditions, this establishes asymptotic convergence to the surrogate posterior as the candidate count and flow resolution increase. Appendix A.11 further quantifies the additional error associated with diagonal component evidence. Together, the theoretical analysis characterizes the approximation errors, while the experiments demonstrate accurate filtering with small ensembles and finite flow and sampling budgets.

I appreciate the new EnSF-LR comparison and the rerun of the Double-Well experiment. My main concern remains the comparison with IEnSF on Lorenz-96 stride-2. The authors explain why their implementation did not yield stable results, but the response does not reconcile this outcome with IEnSF’s published result in an apparently comparable setting. That result suggests a materially smaller improvement than the 38% highlighted in Table 2.

We thank the reviewer for the follow-up and for acknowledging the additional experiments.

Comment.

My main concern remains the comparison with IEnSF on Lorenz-96 stride-2. The authors explain why their implementation did not yield stable results, but the response does not reconcile this outcome with IEnSF’s published result in an apparently comparable setting. That result suggests a materially smaller improvement than the 38% highlighted in Table 2.

Response:

We acknowledge the competitive Lorenz-96 results reported in IEnSF. As discussed in our previous response, we made several attempts to reproduce IEnSF in our experimental setting but were unable to obtain stable filtering results.

The original 38% improvement was calculated relative to EnKF-MDA (RMSE 0.987 versus 0.615). With the inclusion of EnSF-LR in the revised Table 2, the improvement over the strongest evaluated baseline is approximately 35% (0.942 versus 0.615).

These percentages refer to the baselines evaluated under our experimental protocol. We agree that the published IEnSF result suggests a smaller gap, although a direct comparison under matched conditions remains necessary to quantify this difference.

We appreciate your suggestions concerning the additional baselines, Double-Well statistics, and presentation of the method. We have also added a complete overview of the filtering framework in Figure 1 of the revised manuscript.

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

This work proposes FREESIA, a training-free ensemble filtering method. In each filtering cycle, it builds a Gaussian-mixture surrogate from the forecast ensemble, where the component means are shrunk ensemble members and all components share the covariance. Particles then move along a straight interpolantion path. The velocity is computed from the conditional endpoint mean, estimated by importance sampling. The proposal is a linearized Kalman update and the weights use the true nonlinear likelihood. This work also provides the bound of one-step W1W_1 error. Experiments include double-well, lorenz-96 and kolmogorov flow with sparse, non-linear and non-injective observations.

Critical strengths
  1. The method is clean. The gaussian bridge gives closed-form weights and endpoint distributions.
  2. On double-well, there is a grid reference posterior, and the FREESIA recovers both modes and their masses. This directly proves the posetrior-shape claim in low dimension.
  3. The gains in high dim are large, also against KF. The appendix includes std, run-time tables and ablation and nESS plots.
Critical weaknesses
  1. It's unclear to me where does the gain come from. The high-dim results under non-injective observations are good, but I cannot tell how much of the gaims comes from the multi-step flow and how much from the likelihood re-weighting.
    • Both sparse Lorenz-96 settings use γ=0\gamma=0, so the surrogate reduces to a single Gaussian. The paper says sparse monotone observations isolate covariance-mediated information transfer (line 389). However, EnKF-MDA also uses a localized covariance. So the gain over EnKF-MDA could come from the flow, the linearlized proposal, the reweighting, or the different localization and inflation. They are not separated.
    • In Figure 15, T=20T=20 with NIS=1N_{\mathrm{IS}}=1 already gives RMSE very close to the default on Kolmogorov 1% arctan. With NIS=1N_{\mathrm{IS}}=1, the weight in Eq. 3.12 is always 1, so the reweighting actually does nothing. Then, in Lorenz-96 stride-4 experiment, using uniform weights (no-likelihood-correction in the figure 11) can raise RMSE significantly, from 1.575 to 2.969 as reported.
    • Because Z0⊥(Z1,Y)Z_0 \perp (Z_1, Y), at t=0 (in Eq. 3.5), the endpoint distribution is already the full surrogate posterior. So, a simple sampler can be done in one step: use the same kalman proposal and weights, then resample the endpoints, instead of taking the weighted mean in Eq. 3.13. I would suggest authors to compare the flow with this simpler sampler. In summary, my main confusion is that the experiments show a large accuracy gain for FREESIA, as a whole thing. But the paper leaves it open how much the proposed transport adds beyond a simpler update with the same surrogate and proposal...
  2. Limited comparison with close baselines and at similar cost. The IEnSF and EnSF-LR are two close covariance-aware score-based filters, but they are not compared. Ensemble Gaussian-mixture kernel filters use a similar surrogate, but also not compared. I can somehow understand why IEnSF and EnSF-LR were left out as authors explained in the appendix. However, EnSF and EnFF do not explicitly use the covariance in their guidance, so FREESIA can be expected to beat them under sparse observations. About the cost, one FREESIA analysis step is about 6-12x slower than EnKF-MDA. Figure 12 sweeps T for FREESIA but each baseline is only shown at one setting. The equal-time claim is only for EnSF. I also do not find the guidance settings or search ranges for EnSF or EnFF. In Lorenz-96 stride-4 table (table 5), the RMSE of observed coordinate of EnFF-F2P drops by half when sigma increases from 0.01 to 0.05. I'm not sure if the guidance was re-tuned for different noise levels.
  3. The diagonal-evidence error is not shown in the main mixture experiments. The Lorenz and Kolmogorov use diagnoal evidence approximation with positive γ\gamma. Authors give a bound for this case in the appendix, but it adds an extra term εdiag\varepsilon_{\mathrm{diag}}, which is zero only when γ=0\gamma=0, and the paper does not estimate it. So, it's a bit hard to tell how much this approximation costs in the main experiments. The abstract says "relative to finite-ensemble surrogate", but its first claim on "asymptotically exact posterior transport method and the intro sounds stronger than what is shown for the high-dim algorithm.
Questions for authors
  1. Could you compare the multi-step flow with one-step sampling at t=0 (Kalman proposal, likelihood weights, then resampling) with the same surrogate and similar cost? e.g., on Double-well? Besides, how will the RMSE and CRPS change if only the weights are set to uniform at T=20 and NIS=5N_{\mathrm{IS}}=5, on kolmogorov 2% ∣arctan⁡∣|\arctan|?
  2. Could you give something like a accuracy-cost curve for EnKF-MDA, and the guidance settings and search ranges for EnSF and EnFF? Were they re-tuned for different noise levels?
  3. About the approximation and the importance weights. a. In dense Lorenz–96 or Kolmogorov flow, how much does diagonal evidence change the component probabilities and final estimates compared with full evidence? b. In stride-4 Lorenz–96, What are the values of λmax⁡(PΣfP⊤)\lambda_{\max}(P\Sigma^fP^\top) (and so κ\kappa) over cycles do you observe? c. Do the weights in Figure 11 concentrate on one sample, as shown in Fig 13? If so, why does this help compared with uniform weighting?
6 comments

Q1. Contributions of flow transport, the Kalman proposal, and likelihood correction

We added a direct one-shot importance-sampling comparison using the same Gaussian forecast surrogate, localized covariance, Kalman proposal, and nonlinear likelihood weights.

On Lorenz–96 stride-4, one-shot sampling with 100, 500, and 2000 candidates gives RMSEs of 5.162, 5.168, and 5.143, compared with 1.575 for FREESIA. All one-shot configurations suffer severe weight degeneracy at the first analysis, leaving only one distinct resampled particle. This highlights the benefit of progressive transport over one-shot global resampling.

The strong performance at NIS=1N_{\mathrm{IS}}=1 comes from the observation-adaptive Kalman proposal. Even with a single candidate, the proposal already incorporates observational information through local linearization and forecast cross-covariances. Repeated flow updates then progressively transport the ensemble. For example, T=50,NIS=1T=50,N_{\mathrm{IS}}=1 achieves RMSE 1.542 on Lorenz–96 stride-4.

Multiple proposals and likelihood-ratio correction serve a complementary purpose: they correct nonlinear proposal bias and ensure consistent estimation of the posterior conditional endpoint mean as NISN_{\mathrm{IS}} increases.

This also explains the uniform-weight ablation. Replacing the corrected weighted mean of five candidates by their unweighted mean increases RMSE from 1.575 to 2.969 at T=20T=20. This update differs from drawing one observation-adaptive endpoint per flow step.

Thus, the Kalman proposal provides efficient finite-sample conditioning, progressive flow transport preserves ensemble diversity, and importance correction enables asymptotically accurate nonlinear posterior transport.

Q2. Additional baselines and hyperparameter tuning

We added EnSF-LR and EnGMF-DR to the Lorenz–96 comparison. At stride-2, their RMSEs are 0.942 and 1.067, compared with 0.615 for FREESIA. At stride-4, they achieve 1.676 and 2.836, compared with 1.575.

The EnSF-LR implementation uses localized observed-to-unobserved covariance regression, extending the original configuration with cyclic Gaspari–Cohn localization.

We also added IEnSF, BPF, and EnGMF-DR to Double-Well. Their RMSEs are 0.132, 0.150, and 0.273, respectively, compared with 0.120 for FREESIA.

We also made substantial efforts to reproduce IEnSF on Lorenz–96. We developed an implementation ourselves in collaboration with colleagues and separately tested another implementation generated using GPT-6 Astra. Both implementations encountered apparent numerical stiffness and rapid growth during the early stages of reverse-SDE integration, preventing reliable filtering results. We therefore did not include these unstable results in the comparison.

For baseline tuning, we used method-specific hyperparameter searches. For EnSF, we considered ϵα∈[10−3,1]\epsilon_\alpha\in[10^{-3},1] and ϵβ∈[10−3,0.3]\epsilon_\beta\in[10^{-3},0.3]. For EnFF, we searched σmin⁡∈[10−5,1]\sigma_{\min}\in[10^{-5},1] and guidance scales in [10−3,10][10^{-3},10].

Hyperparameters were first coarsely screened using a small number of trajectories, followed by finer selection based on average performance over 10 different validation trajectories. This procedure was applied separately to each observation regime and noise level. The selected values are reported in Appendix B.

Q3. Diagonal evidence and importance-weight behavior

We added a direct comparison between full and diagonal component evidence on dense Lorenz–96 with γ=0.3\gamma=0.3.

Full evidence achieves RMSE 0.404, compared with 0.458 using diagonal evidence. The diagonal approximation reduces mean analysis time by 12.8×, while both versions retain the structured covariance in the conditional endpoint proposals.

For sparse Lorenz–96, γ=0\gamma=0, so all GMM components coincide. Their normalized component probabilities are identical under full and diagonal evidence, and the corresponding approximation error vanishes.

We also extended the importance-weight diagnostics across flow time and assimilation cycles. The weights are concentrated during early and intermediate flow stages, but become more balanced near t=1t=1 as the conditional endpoint distribution contracts. The concentration does not systematically worsen across assimilation cycles.

Combined with the TT-NISN_{\mathrm{IS}} sensitivity experiments, these results show that the observation-adaptive proposal and progressive transport remain effective with small candidate budgets.

Thanks for the detailed reply and the new experiments. Some of my concerns are now resolved.

The one-shot test helps. With about the same proposal budget, one-shot sampling collapses to one particle (RMSE about 5.165.16 vs. 1.5751.575), so the multi-step update clearly matters. The new baselines, the tuning ranges, and the full vs. diagonal evidence test also answer my questions.

My main concern is still the likelihood correction, and your new results make it sharper. With NIS=1N_{\mathrm{IS}}=1, the weight is always 1, so the correction does nothing. On stride-4 at T=50T=50, NIS=1N_{\mathrm{IS}}=1 gets 1.5421.542, while NIS=5N_{\mathrm{IS}}=5 is about 1.581.58 in Fig. 12. At T=20T=20, NISN_{\mathrm{IS}} from 1 to 50 stays between about 1.561.56 and 1.651.65. I agree that the uniform-weight run in Fig. 11 is a different estimator. But then Fig. 11 doesn't really show the effect of the correction, even though the paper uses it that way.

To be fair, these results are all for monotone arctan observations and RMSE. The correction is meant for non-injective, multimodal cases, so it may matter more there. That's why I think the key test is the one from my first question: on Kolmogorov 2% ∣arctan⁡∣|\arctan| (or dense Lorenz–96 ∣arctan⁡∣|\arctan|), how does NIS=1N_{\mathrm{IS}}=1 compare with the default? If the default is clearly better, that would support the correction where it counts. If not, the claims about it should be toned down.

One smaller point: an accuracy–cost curve for EnKF-MDA is still missing. It matters less for the Kolmogorov ∣arctan⁡∣|\arctan| case, where the gap is large, but it would help for settings like stride-4.

For now I'll keep my score. Thanks again for the work on the rebuttal.

We thank the reviewer for the follow-up comments and for recognizing the improvements made in our revision. We would like to further clarify the role of likelihood correction and provide the additional accuracy-cost comparison.

Comment 1.

My main concern is still the likelihood correction, and your new results make it sharper. With N_IS = 1, the weight is always 1, so the correction does nothing. On stride-4 at T = 50, N_IS = 1 gets 1.542, while N_IS = 5 is about 1.58 in Fig. 12. At T = 20, N_IS from 1 to 50 stays between about 1.56 and 1.65. I agree that the uniform-weight run in Fig. 11 is a different estimator. But then Fig. 11 doesn't really show the effect of the correction, even though the paper uses it that way.

To be fair, these results are all for monotone arctan observations and RMSE. The correction is meant for non-injective, multimodal cases, so it may matter more there. That's why I think the key test is the one from my first question: on Kolmogorov 2% |arctan| (or dense Lorenz-96 |arctan|), how does N_IS = 1 compare with the default? If the default is clearly better, that would support the correction where it counts. If not, the claims about it should be toned down.

Response:

The primary purpose of importance sampling in FREESIA is to ensure the accuracy of posterior distribution approximation. As established in Theorems 2 and 3, likelihood-ratio correction provides a consistent estimator of the posterior conditional endpoint expectation and ensures convergence of the full-evidence transport to the surrogate posterior under the stated assumptions, with the approximation error characterized in Wasserstein-1 distance.

While RMSE and marginal CRPS evaluate state estimation and probabilistic prediction accuracy, improvements in posterior distribution approximation do not necessarily translate into improvements in these two metrics.

The case NIS=1N_{\mathrm{IS}}=1 represents a computationally efficient specialization of FREESIA, in which the conditional endpoint expectation is approximated by a single sample drawn from the locally linearized conditional distribution. Its strong empirical performance demonstrates the effectiveness of our observation-adaptive proposal and progressive transport.

The benefit of increasing the number of importance samples can be directly evaluated in the non-injective Double-Well experiment, where a numerical reference posterior is available. As shown in Figure 5 of Appendix B.1, with T=20T=20 fixed:

MetricN_IS = 1N_IS = 5N_IS = 10
RMSE0.1290.1160.114
CRPS0.0820.0780.077
Wasserstein-10.0710.0510.049
JS divergence0.0440.0300.029
Mode-mass error0.0400.0300.029

Increasing N_IS from 1 to 10 reduces the Wasserstein-1 distance, JS divergence, and mode-mass error by approximately 31%, 34%, and 28%, respectively, while the improvements in RMSE and CRPS are comparatively modest.

Moreover, Figure 5 illustrates that these metrics can exhibit different trends:

  • At T = 5, increasing N_IS from 1 to 5 reduces JS divergence from 0.036 to 0.034, while RMSE increases from 0.117 to 0.128 and CRPS increases from 0.076 to 0.099.
  • At T = 10, increasing N_IS from 1 to 5 reduces Wasserstein-1 distance from 0.055 to 0.050 and JS divergence from 0.035 to 0.027, even though CRPS increases from 0.079 to 0.081.

These results illustrate that RMSE and CRPS alone cannot fully characterize the improvement in posterior approximation arising from likelihood correction.

Since a numerical reference posterior is unavailable for the high-dimensional benchmarks, we focus the direct evaluation of posterior approximation quality on the Double-Well experiment rather than conducting another comparison based solely on RMSE and CRPS.

FREESIA therefore provides a flexible framework that combines efficient finite-sample state estimation with theoretically consistent nonlinear posterior transport.

Reviewer_4pCuOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

FREESIA is a training-free posterior transport method for sequential data assimilation. At each assimilation step, it constructs a Gaussian-mixture forecast surrogate from the current ensemble and uses its covariance structure to propagate observational information to unobserved variables. Locally linearized observation models provide tractable, observation-adaptive endpoint proposals, while likelihood-ratio reweighting corrects for nonlinear observations. The paper establishes surrogate-relative asymptotic exactness and error bounds separating finite endpoint sampling, flow discretization, and forecast-surrogate error, with additional analysis of the diagonal component-evidence approximation. Experiments on Double-Well, Lorenz–96, and Kolmogorov flow evaluate posterior structure, sparse-observation recovery, filtering accuracy, and computational cost.

Critical strengths
  1. A principled construction with explicit error analysis. The method uses local linearization to construct proposals while retaining the nonlinear likelihood in the correction. The theoretical analysis distinguishes convergence to the surrogate posterior from approximation of the underlying Bayesian posterior and further analyzes forecast-surrogate and component-evidence errors. This makes the scope of the exactness claim substantially clearer than a posterior-sampling claim supported only by reconstruction accuracy.

  2. Strong accuracy improvements, particularly under sparse nonlinear observations. FREESIA outperforms established covariance-based filters and training-free generative baselines in several challenging settings. On Kolmogorov flow with 2% absolute-arctangent observations, RMSE decreases from the best evaluated baseline’s 0.970 to 0.422. Observed/unobserved error decompositions and ablations support the contributions of cross-covariance propagation and nonlinear likelihood correction. The Double-Well experiment additionally compares against a grid-resolved posterior, providing evidence of improved distributional fidelity rather than only improved point estimates.

Critical weaknesses
  1. The dimension-dependent accuracy of the practical approximation remains uncertain. The error analysis identifies relevant contributions but does not establish that they are small in the reported high-dimensional, small-ensemble settings. The constants and forecast-surrogate discrepancy can depend on dimension, covariance structure, and observation informativeness. The diagonal component-evidence approximation also introduces an additional error in the general implementation. Although the million-dimensional experiment demonstrates computational scaling, it uses a fully observed, diagonal-covariance, single-Gaussian specialization and therefore does not establish comparable scaling for the covariance-aware sparse-observation regime central to the paper.

  2. The practical accuracy–cost advantage is not fully established. The paper appropriately reports runtime and accuracy–cost sweeps, but the main FREESIA configurations require substantially more analysis time than EnKF-MDA and other baselines. The additional cost may be justified by the accuracy improvements, especially for nonlinear sparse observations, but comparisons that also increase baseline ensemble sizes or iteration budgets would make this assessment more conclusive. Closely related covariance-aware methods, particularly IEnSF and EnSF-LR, are absent from the numerical comparison, although the authors explain the implementation difficulties.

Questions for authors
  1. High-dimensional approximation error. Which terms or constants in the error bounds are expected to deteriorate with state or observation dimension, and under what structural assumptions might this dependence remain manageable? Could controlled experiments at tractable dimensions quantify the forecast-surrogate and diagonal component-evidence errors separately?

  2. Scope of scalability. How do accuracy, runtime, and memory scale when sparse observations and cross-coordinate covariances are retained as dimension increases? Please distinguish the scalability demonstrated by FREESIA-Diag from that of the general covariance-aware implementation.

  3. Comparison at matched computational budgets. Do the accuracy gains persist when EnKF-MDA or other strong baselines are allowed larger ensembles or additional analysis iterations within FREESIA’s wall-clock budget? This would help assess whether the additional online computation provides a favorable practical tradeoff.

2 comments

Response to Reviewer 4pCu

We thank the reviewer for the careful assessment of our manuscript and the constructive suggestions. We have revised the manuscript and conducted additional numerical experiments to address the questions below.

Comment 1.

High-dimensional approximation error. Which terms or constants in the error bounds are expected to deteriorate with state or observation dimension, and under what structural assumptions might this dependence remain manageable? Could controlled experiments at tractable dimensions quantify the forecast-surrogate and diagonal component-evidence errors separately?

Response:

The error analysis separates three contributions: finite endpoint sampling, flow discretization, and forecast-surrogate discrepancy. The dimension-dependent constants are mainly influenced by the observation likelihood, conditional covariance structure, and the parameter κ in Theorem 3, which depends on the observation Jacobian, forecast covariance, and observation-noise covariance.

Covariance localization restricts the interactions between state coordinates, while observation-adaptive Kalman proposals concentrate endpoint sampling around observation-compatible regions. These mechanisms help reduce the effective dimension of the conditioning problem and improve sampling efficiency in high-dimensional filtering.

We further added a full-versus-diagonal component-evidence experiment on dense Lorenz-96 with γ = 0.3. Full evidence achieves RMSE 0.404, compared with 0.458 for diagonal evidence, while the latter reduces analysis time by 12.8×. The experiment directly quantifies the effect of the component-evidence approximation on filtering accuracy.

For sparse Lorenz-96 with γ = 0, all GMM components coincide, and the diagonal component-evidence approximation is exact. The forecast-surrogate contribution is separately characterized by the surrogate discrepancy term in Eq. (4.1) and the corresponding Wasserstein error analysis.

Comment 2.

Scope of scalability. How do accuracy, runtime, and memory scale when sparse observations and cross-coordinate covariances are retained as dimension increases? Please distinguish the scalability demonstrated by FREESIA-Diag from that of the general covariance-aware implementation.

Response:

We evaluate two complementary scalability regimes.

FREESIA-Diag uses diagonal covariance and a single-Gaussian surrogate, achieving O(MTd) computational complexity and O(Md) memory. The corresponding experiments extend to one million dimensions.

Structured FREESIA retains cross-coordinate covariance information in the conditional endpoint proposals. The implementation exploits block-structured covariance operations and observation-space linear solves, avoiding dense state-space covariance storage. The complete implementation and computational complexity are described in Appendix A.10.

The 32,768-dimensional Kolmogorov-flow experiments evaluate this implementation under sparse nonlinear observations. With only 2% non-injective observations, FREESIA achieves RMSE 0.422 and CRPS 0.209, compared with 0.970 and 0.569 for the strongest evaluated baselines on the respective metrics.

These experiments demonstrate efficient dimensional scaling in the diagonal specialization and accurate covariance-aware filtering at PDE scale.

Comment 3.

Comparison at matched computational budgets. Do the accuracy gains persist when EnKF-MDA or other strong baselines are allowed larger ensembles or additional analysis iterations within FREESIA's wall-clock budget? This would help assess whether the additional online computation provides a favorable practical tradeoff.

Response:

We conducted an additional accuracy-cost comparison on Lorenz-96 stride-4 by increasing the number of EnKF-MDA analysis substeps from the original 20 to 30, 50, 100, 160, and 250, while keeping the ensemble size and other experimental settings unchanged.

Increasing the number of substeps improves EnKF-MDA's RMSE from 2.056 at 20 substeps to 1.855 at 100 substeps and 1.828 at 160 substeps.

At approximately matched computational cost, EnKF-MDA with 160 substeps requires 251.5 ms per analysis and achieves RMSE 1.828, while FREESIA requires approximately 251 ms and achieves RMSE 1.575. FREESIA therefore reduces RMSE by 13.8% at nearly identical analysis time.

FREESIA also achieves lower CRPS (0.584 versus 0.836) and more accurate unobserved-state estimation (RMSE 1.805 versus 2.104), demonstrating the benefit of covariance-aware posterior transport beyond repeated Kalman updates.

Further increasing EnKF-MDA to 250 substeps requires 385.8 ms per analysis, while its RMSE improves only slightly to 1.816. The improvement becomes progressively smaller as the number of analysis substeps increases.

We additionally expanded the baseline comparisons with IEnSF, BPF, and EnGMF-DR on Double-Well, and EnSF-LR and EnGMF-DR on Lorenz-96.

Together with the FREESIA accuracy-cost sweeps, these results demonstrate that FREESIA maintains a substantial accuracy advantage over EnKF-MDA under comparable computational budgets, particularly in the recovery of unobserved states.

Comments on the paper

1 comment

Summary of Revisions and Discussion

Dear Area Chair,

We would like to thank you and all five reviewers for their time and constructive feedback. The discussions have helped us identify several aspects of the paper that required further clarification and experimental validation. We would like to summarize the main changes we have made during the discussion period.

1. Clarification of the method and its contributions.

We have expanded the discussion of closely related methods in the Introduction, clarified the connection between our endpoint proposal and Kalman conditioning in Section 3, and added a complete overview of the filtering framework. We have also revised the existing method illustration.

Following the discussions with the reviewers, we now distinguish more explicitly between the roles of covariance-aware proposals, progressive transport, and nonlinear likelihood correction in both the theoretical analysis and the interpretation of the numerical results.

2. Additional baseline comparisons and experimental validation.

We have added IEnSF, BPF, and EnGMF-DR to the Double-Well comparison, as well as EnSF-LR and EnGMF-DR to the Lorenz-96 experiments. Following Reviewer q38b's observation regarding the Double-Well statistics, we revised the initialization protocol and repeated the comparisons using ten independent truth trajectories.

We also added comprehensive final-time comparisons of all seven methods under both Kolmogorov-flow observation regimes to provide a clearer picture of the reconstruction accuracy of the different filters.

3. Further investigation of the algorithmic components and computational cost.

To examine the contribution of progressive transport, we compared FREESIA with one-shot importance sampling using the same surrogate and proposal construction. The one-shot configurations suffered severe weight degeneracy and achieved RMSEs of approximately 5.16, compared with 1.575 for FREESIA.

We evaluated the diagonal component-evidence approximation and found that it provides a 12.8-fold speedup while retaining competitive accuracy. We also extended the EnKF-MDA comparison by increasing its number of analysis substeps. At nearly identical analysis times, FREESIA achieves an RMSE of 1.575 compared with 1.828 for EnKF-MDA.

4. Connection between posterior approximation theory and numerical results.

An important point raised by Reviewer nvqK was the role of likelihood correction when the observation-adaptive proposal already performs well with a single candidate. We therefore revised the manuscript to clarify the efficiency of the proposal and the complementary role of importance sampling in consistent posterior approximation.

In the Double-Well experiment, we further examined the effect of increasing the candidate count using a numerical reference posterior. With the number of flow steps fixed, increasing the candidate count from 1 to 10 reduces Wasserstein-1 distance, JS divergence, and mode-mass error by approximately 31%, 35%, and 28%, respectively. The additional trajectory-level paired statistics support these improvements.

These results complement our theoretical analysis by demonstrating that increasing the candidate budget can improve posterior distribution approximation even when changes in RMSE and CRPS are comparatively modest.

The discussions have resolved several of the reviewers' original concerns. The main remaining baseline question raised by Reviewers nvqK and q38b concerns a direct comparison with IEnSF on Lorenz-96. Despite our reproduction efforts, we have not obtained stable results under our experimental protocol, although we have included IEnSF in the Double-Well comparison.

We appreciate the reviewers' suggestions and believe that the resulting revisions provide a clearer presentation of the method, a more comprehensive experimental evaluation, and a stronger connection between the theoretical results and numerical evidence.

Thank you again for your time and consideration.

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.