acceptodds
Under review as a conference paper at ICLR 2027

Not All Observations Are Equal: Guiding Derivative-Free Ensembles via Uncertainty Contraction

Abstract

Ensemble Kalman Guidance (EnKG) is attractive for high-dimensional inverse problems, especially with black-box physical forward models where repeated gradient computation is difficult or impractical, as it only requires forward evaluations. However, its empirical performance can be hindered by the finite ensemble size when the parameter dimension is large, resulting in unreliable or biased estimates. To address this issue, we propose Uncertainty-Contraction Ensemble Guidance (UCEG), a monitor-guided adaptive sampling framework that improves the quality of the ensemble correction without increasing the ensemble size. UCEG treats guidance construction as an observation-selection problem and uses update activity as a practical proxy for local uncertainty contraction in the finite-ensemble correction, retaining informative observations while suppressing uninformative ones. Experiments on challenging inverse problems including full waveform inversion (FWI) and Navier–Stokes show that UCEG can outperform the corresponding full-observation correction using fewer observations, with faster inference when observation subsampling reduces evaluation cost.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Diffusion Models, Ensemble Kalman Guidance, PDE-Constrained Inference

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 3 reviews, above 27% of the others.
Reviewer_bw2wOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Uncertainty-Contraction Ensemble Guidance (UCEG), an adaptive observation selection method for Ensemble Kalman Guidance (EnKG) in black-box inverse problems. The method uses historical ensemble update activity as a proxy for uncertainty contraction to select informative observations at each correction step. Experiments on FWI, Navier–Stokes, and Darcy flow show improved reconstruction accuracy over EnKG, with substantial computational savings in FWI.

Critical strengths

The paper addresses a meaningful problem: improving the reliability and efficiency of finite-ensemble guidance in high-dimensional inverse problems. The proposed method is relatively simple and achieves promising results across multiple PDE benchmarks, particularly in FWI.

Critical weaknesses
  1. Limited methodological novelty. Observation selection based on information content or uncertainty reduction has been studied in ensemble data assimilation [1, 2], while observation subsampling has also been explored in ensemble Kalman inversion [3]. The theoretical connection between expected mean-update energy and covariance contraction also follows from standard linear-Gaussian conditioning. The paper does not sufficiently clarify its novelty relative to these existing lines of work.
  2. Insufficient justification and validation of the update-history monitor. The theoretical result connects expected mean-update energy to covariance contraction under a linear-Gaussian model, but does not directly justify using realized historical particle updates to predict the informativeness of future observations. Although the authors provide random-monitor ablations and empirical covariance-contraction analysis, the additional reconstruction improvements from the proposed monitor are relatively modest. More importantly, it remains unclear whether update history is a better selection criterion than simpler alternatives, such as ensemble variance or observation residuals.

[1] Migliorini, S. (2013). Information-based data selection for ensemble data assimilation. Quarterly Journal of the Royal Meteorological Society, 139(677), 2033–2054. https://doi.org/10.1002/qj.2104

[2] Shen, Z., Zhong, Q., & Chen, Z. (2022). Parameter estimation using adaptive observations toward maximum total variance reduction with ensemble adjustment Kalman filter. Frontiers in Climate, 4, 850386. https://doi.org/10.3389/fclim.2022.850386

[3] Hanu, M., Latz, J., & Schillings, C. (2023). Subsampling in ensemble Kalman inversion. Inverse Problems, 39(9), 094002. https://doi.org/10.1088/1361-6420/ace64b

Questions for authors
  1. Could the authors clarify the key methodological differences from existing information-based or uncertainty-based observation selection methods in ensemble data assimilation?
  2. Could the authors compare the update-history monitor with simpler adaptive selection criteria, such as ensemble variance or observation residuals, under the same observation budget?
  3. Is there empirical evidence that historical update activity is correlated with the actual uncertainty contraction induced by the selected observations at subsequent correction steps?
  4. Are the relatively small improvements over the random-monitor baseline statistically significant across test cases?
5 comments
AuthorsOpenReview ↗

Official Rebuttal to Reviewer bw2w (1/3)

Weakness 1 and Question 1: Methodological Novelty and Related Work

We thank the reviewer for highlighting the connections to information-based observation selection and ensemble Kalman subsampling. These studies share our motivation of improving ensemble-based inference through informative observation selection and uncertainty reduction. Our work builds on Ensemble Kalman Guidance (EnKG) and focuses on developing an efficient, adaptive observation-selection mechanism for finite-ensemble diffusion guidance.

Diffusion models have emerged as powerful generative priors for high-dimensional inverse problems, offering expressive representations of complex solution distributions. EnKG makes these priors particularly attractive for scientific inverse problems by enabling derivative-free, measurement-informed guidance through empirical ensemble cross-covariances, even with expensive black-box forward models. However, with a limited ensemble, incorporating a large number of observation channels can introduce noisy correlations and reduce the reliability of the guidance correction. We address this limitation by adapting the observations used in each EnKG correction. Our theoretical analysis establishes a direct relationship between expected mean-update energy and covariance contraction under local linear-Gaussian conditioning. This connection motivates using realized ensemble update activity as an informative signal for uncertainty reduction.

The key methodological contribution of UCEG is an efficient, history-driven observation-selection mechanism that operates throughout the reverse diffusion trajectory. At each correction step, UCEG records the realized particle updates, aggregates their spatial activity, and projects the resulting monitor onto candidate observation units. These scores determine the observations used in subsequent EnKG corrections, allowing observation allocation to adapt continuously as inference progresses. Importantly, the selection signal is obtained directly from previous ensemble corrections, without additional forward-model evaluations for candidate scoring.

This mechanism differs from the three cited approaches. Migliorini (2013) identifies informative observation components through ensemble-based information analysis, while Shen et al. (2022) select observations by evaluating prospective total variance reduction using the ensemble adjustment Kalman filter. Hanu et al. (2023) study randomized data subsampling in ensemble Kalman inversion. In contrast, UCEG exploits the relationship between update activity and uncertainty contraction to construct an adaptive selection mechanism directly from the evolving history of diffusion-based EnKG corrections. This integrates observation selection into the guidance process rather than treating it as a separate information-evaluation procedure. We have expanded the Related Work section to acknowledge these studies and clarify UCEG's distinct update-history-guided observation selection mechanism.

We further validate this mechanism through observation-budget-matched experiments. In FWI, history-guided selection achieves greater mean relative covariance contraction than randomized-monitor selection (20.119% vs. 19.607%) and reduces the final covariance trace from 549.19 to 521.15, with improvements in 9 of 10 test cases. More details can be found in our Appendix E. We will also show additional comparisons with residual-based and variance-based selection in the below responses which demonstrate the computational efficiency of reusing correction history. These results establish the practical value of the proposed update-activity-driven mechanism for efficient uncertainty reduction in diffusion-based inverse inference.

Weakness 2 and Question 3: Theoretical Justification and Empirical Validation of the Update-History Monitor

We thank the reviewer for highlighting the connection between Proposition 1 and the practical update-history monitor. We have extended our theoretical analysis to establish a direct relationship between EnKG particle-update activity and uncertainty contraction, with the complete derivation and proof provided in the new Appendix H.

Theoretical Connection.

Proposition 1 establishes that expected mean-update energy equals covariance contraction under local linear--Gaussian conditioning. Building on the correction structure of EnKG, we further analyze how this relationship extends to the practical particle-update activity used by UCEG.

AuthorsOpenReview ↗

Official Rebuttal to Reviewer bw2w (2/3)

Specifically, the EnKG correction uses the empirical state--observation cross-covariance Bi,I=Cxx(i)Ai,I⊤B_{i,\mathcal I}=C_{xx}^{(i)}A_{i,\mathcal I}^{\top} and particle-dependent observation residuals. Let Hi(r;yI)H_i(r;y_{\mathcal I}) denote the resulting update activity at spatial location rr, and define H~i(r;yI)=Hi(r;yI)/γi(yI)\widetilde H_i(r;y_{\mathcal I})=H_i(r;y_{\mathcal I})/\gamma_i(y_{\mathcal I}), where γi\gamma_i is the scalar correction step size. Using Jensen's inequality and variance decomposition, we establish bounds on the expected squared activity in terms of the ensemble cross-covariance and predictive observation covariance. Combining these bounds with Proposition 1 yields

\lambda_{\min}^2 U_i(r;\mathcal I) &\leq \mathbb E_{y_{\mathcal I}} \left[\widetilde H_i(r;y_{\mathcal I})^2\right] \\\\ &\leq 2\lambda_{\max}^2 U_i(r;\mathcal I), \end{aligned}$$ where $U_i(r;\mathcal I)=\operatorname{tr}(L_r\Delta C_{xx}^{(i)}(\mathcal I)L_r^\top)$ is the projected covariance contraction, and $\lambda_{\min}$ and $\lambda_{\max}$ are the minimum and maximum eigenvalues of the positive-definite predictive observation covariance $S_{i,\mathcal I}$. We establish a mathematical connection between the practical EnKG update activity and the uncertainty-contraction utility in Proposition 1, with the proof provided in Appendix H. We thank the reviewer for motivating this analysis, which strengthens the theoretical foundation of our proposed history-adaptive observation selection for diffusion-based inverse inference. **Empirical Validation.** We further evaluate whether historical update activity identifies observations that produce stronger uncertainty contraction in subsequent EnKG corrections. We conduct a paired full-trajectory FWI experiment across 10 test cases. Both methods use the same initial ensemble, noise realization, correction schedule, observation budget, shot subsets, and uncertainty-extra receivers. The only difference is that UCEG selects the monitor-controlled receivers using historical update activity, whereas the randomized-monitor control replaces these selections with random receivers. We measure uncertainty using the empirical covariance trace of the physical-space ensemble. The relative contraction at correction step $k$ is $$\rho_k=\frac{T_k^{-}-T_k^{+}}{T_k^{-}}\times100\\%,$$ where $T_k^{-}$ and $T_k^{+}$ denote the covariance traces immediately before and after the correction. We average the relative contraction across correction steps within each case and then across the 10 test cases. UCEG increases mean relative covariance contraction from 19.607% to 20.119%, corresponding to a paired improvement of 0.512 percentage points (95% CI: $[0.283,0.719]$). It also reduces the final covariance trace from 549.19 to 521.15, with a paired difference of $-28.03$ (95% CI: $[-43.43,-12.24]$). History-guided selection achieves stronger contraction and lower final covariance in 9 of the 10 paired cases. The new theoretical analysis establishes the connection between update activity and covariance contraction, while the controlled FWI experiments demonstrate that historical update-guided observation selection achieves stronger uncertainty reduction throughout diffusion-based EnKG inference. ## Question 2: Comparison with Residual-Based and Variance-Based Selection We thank the reviewer for suggesting comparisons with simpler adaptive selection criteria. We added residual-based and variance-based baselines under the same correction observation budget, keeping all other components of UCEG unchanged. **Experimental Setup.** We evaluate the three selection criteria on FWI across 10 test cases. Residual-based selection uses the discrepancy between predicted and observed measurements, while variance-based selection uses the variance of ensemble-predicted observations. In contrast, UCEG constructs its selection scores from the realized history of EnKG particle corrections. All methods maintain the same correction observation budget and inference configuration. We measure the empirical uncertainty contraction induced by each selection strategy along the FWI inference trajectories, using the mean relative covariance contraction and final covariance trace defined in our preceding response. | Selection criterion | Mean relative contraction $\uparrow$ | Final covariance trace $\downarrow$ | |---|---|---| | Randomized monitor | 19.607% | 549.19 | | Residual-based | 19.690% | 548.26 | | Variance-based | 19.842% | 539.07 | | UCEG update-history | **20.119%** | **521.15** |
AuthorsOpenReview ↗

Official Rebuttal to Reviewer bw2w (3/3)

Uncertainty-Contraction Results.

UCEG achieves the highest mean relative contraction and the lowest final covariance trace among all four selection strategies. Compared with variance-based selection, UCEG increases mean relative contraction from 19.842% to 20.119% and reduces the final covariance trace from 539.07 to 521.15. The residual-based criterion also improves over randomized selection, but produces smaller uncertainty reduction than the variance-based and update-history approaches.

Reconstruction Accuracy and Computational Efficiency.

We further compare the three adaptive selection criteria under the 4-random + 4-adaptively-selected-shot setting across 10 FWI test cases. All methods use the same shot and receiver budgets, ensemble size, and correction schedule. We report reconstruction accuracy and inference runtime in the following table.

MethodRel. ℓ2↓\ell_2 \downarrowPSNR ↑\uparrowSSIM ↑\uparrowTime (min/case) ↓\downarrow
Residual-based0.02088±0.005120.02088\pm0.0051215.716±1.77915.716\pm1.7790.4925±0.07880.4925\pm0.078848.46±1.3448.46\pm1.34
Variance-based0.02099±0.004980.02099\pm0.0049815.662±1.77215.662\pm1.7720.4936±0.07730.4936\pm0.077349.13±2.0049.13\pm2.00
UCEG0.02081±0.004800.02081\pm0.0048015.724±1.74015.724\pm1.7400.4932±0.07650.4932\pm0.076521.95±0.3321.95\pm0.33

UCEG achieves the lowest mean relative ℓ2\ell_2 error and highest mean PSNR, with comparable SSIM across the three methods. More importantly, UCEG requires only 21.9521.95 minutes per case, achieving 2.21×2.21\times and 2.24×2.24\times speedups over residual-based and variance-based selection, respectively. These efficiency gains arise because residual-based and variance-based scoring require additional forward evaluations, whereas UCEG reuses the particle corrections already generated during EnKG guidance. The results demonstrate that update-history-guided selection achieves competitive reconstruction accuracy with substantially lower computational cost.

Question 4: Statistical Significance of the Monitor Improvement

We thank the reviewer for suggesting a statistical analysis of the reconstruction improvements. We conduct paired comparisons between UCEG and randomized-monitor controls on FWI and Darcy flow under matched observation budgets. For each test case, we define Δℓ2=ℓ2randomized−ℓ2UCEG\Delta\ell_2=\ell_2^{\mathrm{randomized}}-\ell_2^{\mathrm{UCEG}} and compute 95% paired bootstrap confidence intervals using test cases as the statistical units.

Experimental Setup.

For FWI, we evaluate the 4-random + 4-selected-shot configuration across 10 test cases. The randomized-monitor control replaces only the monitor-selected receivers while retaining the five uncertainty-extra receivers and all other inference components. For Darcy flow, we evaluate the same monitor-replacement strategy across 20 test cases at ensemble sizes N=64N=64 and N=128N=128, using the same 20% observation budget.

TaskRandomized monitorUCEG95% CI (Δℓ2\Delta\ell_2)Wins
FWI (4+4)0.021950.02081[−0.000151,0.002502][-0.000151,0.002502]7/10
Darcy (N=64N=64)0.216130.21348[0.00163,0.00385][0.00163,0.00385]19/20
Darcy (N=128N=128)0.185920.18327[0.00164,0.00385][0.00164,0.00385]19/20

Statistical Analysis.

On FWI, UCEG reduces mean relative ℓ2\ell_2 error by 5.19%, achieving better reconstruction accuracy in 7 of the 10 paired cases. The mean paired improvement is 0.0011400.001140, with a 95% CI of [−0.000151,0.002502][-0.000151,0.002502] and a two-sided permutation pp-value of 0.1430.143.

On Darcy flow, UCEG achieves consistent improvements across both ensemble sizes, reducing mean reconstruction error by 1.23% at N=64N=64 and 1.43% at N=128N=128. At each ensemble size, UCEG outperforms randomized-monitor selection in 19 of 20 cases, with both paired confidence intervals strictly positive.

The paired analysis shows lower mean reconstruction error across the evaluated settings, with statistically significant improvements on Darcy flow and favorable results in the majority of FWI cases.

Reviewer_rmqSOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

For a physical model whose gradient evaluation is impractical, evaluating its gradients with ensemble such as ensemble kalman guidance (EnKG) is an appealing option. In EnKG, the performance could be degraded if some of the observations are redundant or too weakly informative. To smartly sub-sample observations to improve assimilation performance, the authors propose UCEG, a method to choose important observations from raw observations and filter out uninformative ones. The model selects observations that achieve the best total covariance contraction (Eq.6 is the one-step definition). Experiments in FWI, Navier-Stokes, and Darcy flow show improved performance over the standard EnKG.

Critical strengths
  • The empirical performance seems strong compared to reference methods.
Critical weaknesses
  • The definition of the observation “utility” as the total covariance contraction is very similar to that of Bishop+01. See page 427, Section 4 (b): “Specifically, Eq. (17b) together with the tI+v=96t_{I+v} = 96 h ensemble perturbations were used to find which of the 4096 grid points should be observed at tI+1=48t_{I+1} = 48 h in order to maximize the trace of the signal covariance matrix, S(tI+v∣HI+1q)\textbf{S}(t_{I+v}|H^q_{I+1}), associated with a single vorticity observation at the qqth grid point”. So, the idea of measuring the usefulness of each observation via covariance contraction is presented in the original ETKF paper but not cited.
  • Given that the central contribution was already there from 25 years ago, I feel the contribution of this paper to be incremental even if it is appropriately cited.
  • What is “energy” in the introduction and section 3.3? It is not defined in the paper and I could not infer its meaning. Is it the left hand side of Eq.8?
  • (Minor weakness) many “Eq. equation” appearing in the paper.
  • (Minor weakness) There are two entries for EnKG: the arXiv version and the TMLR version.
Questions for authors
  • Is your definition of utility (Eq.6) same or different from Bishop+01? If it’s the same, can you appropriately cite the paper and clarify what the core originality of this paper is? If it’s not the same, can you discuss the differences in detail?
4 comments
AuthorsOpenReview ↗

Official Rebuttal to Reviewer rmqS (1/2)

Weakness 1 and Question: Methodological Novelty Relative to Bishop et al. (2001)

We thank the reviewer for highlighting Bishop et al. (2001). We clarify how Eq. (6) relates to existing ensemble-based observation selection and highlight UCEG's theoretical and methodological contributions.

Relationship to Prior Work.

Bishop et al. (2001) establish that signal covariance equals the reduction in error covariance at the analysis time, and use ETKF to prospectively evaluate candidate observing networks. Our Eq.~(6) adopts the same covariance-reduction utility under local linear--Gaussian conditioning, but considers the one-step uncertainty contraction induced by selected observations during EnKG inference. The key difference lies in how observations are selected, rather than in the utility itself.

Migliorini (2013) develops information-based observation selection to retain informative components that can be effectively assimilated by a finite ensemble. Shen et al. (2022) select adaptive observations according to their expected total variance reduction in parameter estimation using EAKF. Hanu et al. (2023) investigate randomized data subsampling in ensemble Kalman inversion to improve computational efficiency. These studies establish the importance of selective observation assimilation, while adopting different mechanisms for evaluating or allocating observations. We have expanded Related Work to acknowledge these studies and clarify UCEG's update-history-guided selection.

Methodological Novelty.

Our work advances uncertainty-guided observation selection within diffusion-based inverse inference, where EnKG enables derivative-free guidance through empirical ensemble cross-covariances. With a limited ensemble, the reliability of the guidance correction depends on which observation channels are incorporated into each update. We address this challenge by developing an efficient observation-selection mechanism that adapts to the evolving ensemble throughout reverse diffusion.

Theoretical Connection.

Proposition 1 establishes that expected mean-update energy equals covariance contraction under local linear--Gaussian conditioning. In our new Appendix H, we extend this analysis to the practical particle-update activity used by EnKG.

Specifically, the EnKG correction uses the empirical state--observation cross-covariance Bi,I=Cxx(i)Ai,I⊤B_{i,\mathcal I}=C_{xx}^{(i)}A_{i,\mathcal I}^{\top} and particle-dependent observation residuals. Let Hi(r;yI)H_i(r;y_{\mathcal I}) denote the resulting particle-update activity at spatial location rr, and define H~i=Hi/γi\widetilde H_i=H_i/\gamma_i, where γi\gamma_i is the scalar correction step size shared across the ensemble. Using Jensen's inequality and variance decomposition, we establish

\lambda_{\min}^2 U_i(r;\mathcal I) &\leq \mathbb E_{y_{\mathcal I}} [\widetilde H_i(r;y_{\mathcal I})^2] \\\\ &\leq 2\lambda_{\max}^2 U_i(r;\mathcal I), \end{aligned}$$ where $U_i(r;\mathcal I)$ is the projected covariance contraction, and $\lambda_{\min}$ and $\lambda_{\max}$ are the extreme eigenvalues of the positive-definite predictive observation covariance. This establishes a mathematical connection between the practical EnKG particle-update activity and the uncertainty-contraction utility in Proposition 1. **Adaptive Selection Mechanism.** Based on this connection, UCEG uses the realized particle-update activity from an EnKG correction to guide observation selection at the next correction step. The spatial activity is projected onto candidate observation units to construct selection scores, which determine the observations used in the subsequent subset-conditioned correction. Repeating this process throughout reverse diffusion allows the observation allocation to adapt to the evolving ensemble. Unlike prospective covariance-based observation selection, UCEG reuses the particle corrections already generated by EnKG, without additional forward evaluations for candidate scoring. We have also revised Section 3.3 to better explain the connection between Proposition 1, the new theoretical analysis in Appendix H, and the practical update-activity monitor used by UCEG.
AuthorsOpenReview ↗

Official Rebuttal to Reviewer rmqS (2/2)

Empirical Evidence.

We further validate whether the proposed update-history monitor leads to stronger uncertainty contraction during EnKG inference. We conduct a paired full-trajectory FWI experiment across 10 test cases, where UCEG and the randomized-monitor control share the same initial ensembles, noise realizations, correction schedules, and observation budgets. The only difference is that the monitor-selected receivers in UCEG are replaced by randomly selected receivers in the control, while all other components remain unchanged.

We measure uncertainty using the trace of the empirical ensemble covariance in physical space. At each correction step kk, we define the relative covariance contraction as

ρk=Tk−−Tk+Tk−×100\rho_k = \frac{T_k^{-}-T_k^{+}}{T_k^{-}}\times 100\\%,

where Tk−T_k^{-} and Tk+T_k^{+} denote the ensemble covariance traces immediately before and after the EnKG correction, respectively. A larger ρk\rho_k indicates a greater proportional reduction in ensemble uncertainty. We first average the relative contraction across correction steps within each test case, and then average across the 10 cases. We also report the final covariance trace to measure the remaining ensemble uncertainty.

Selection methodMean relative contraction ↑\uparrowFinal covariance trace ↓\downarrow
Randomized monitor19.607%549.19
UCEG update-history20.119%521.15

Compared with the randomized-monitor control, UCEG increases mean relative covariance contraction by 0.512 percentage points, with a paired 95% bootstrap confidence interval of [0.283,0.719][0.283,0.719] percentage points. It also reduces the final covariance trace by 28.03, with a paired 95% confidence interval of [−43.43,−12.24][-43.43,-12.24] for the UCEG-minus-control difference. UCEG achieves stronger contraction and lower final covariance in 9 of the 10 paired cases.

These results demonstrate that historical particle-update activity provides an effective signal for selecting observations that induce stronger uncertainty contraction in subsequent EnKG corrections.

Weakness 2: Definition of Update Energy

We thank the reviewer for pointing this out. By ``update energy,'' we mean the squared norm of the ensemble mean update, ∥Δxˉi(I)∥22\|\Delta\bar{x}_i(\mathcal I)\|_2^2. Its expectation is exactly the left-hand side of Eq.~(8):

&\mathbb{E}_{y_{\mathcal I}} [\lVert\Delta\bar{x}_i(\mathcal I)\rVert_2^2] \\\\ &=\operatorname{tr}(\Delta C_{xx}^{(i)}(\mathcal I)). \end{aligned}$$ For a specific spatial location $r$, we use the projected update energy $\|L_r\Delta\bar{x}_i(\mathcal I)\|_2^2$. We have clarified these definitions in the Introduction and Section 3.3. ## Minor Weakness 3: Equation References We thank the reviewer for identifying these formatting errors. We have corrected the duplicated "Eq. equation" references and standardized equation citations throughout the revised manuscript. ## Minor Weakness 4: Duplicate EnKG References We thank the reviewer for pointing out the duplicate entries. We have removed the arXiv entry and retained the published TMLR version of EnKG (Zheng et al., 2025), updating the corresponding citations throughout the manuscript. **References** [1] Migliorini, S. (2013). Information-based data selection for ensemble data assimilation. [2] Shen, Z., Zhong, Q., & Chen, Z. (2022). Parameter estimation using adaptive observations toward maximum total variance reduction with ensemble adjustment Kalman filter. [3] Hanu, M., Latz, J., & Schillings, C. (2023). Subsampling in ensemble Kalman inversion. [4] Bishop, C. H., Etherton, B. J., & Majumdar, S. J. (2001). Adaptive sampling with the ensemble transform Kalman filter. Part I: Theoretical aspects.

Thank you for your detailed response. However, my main concern of limited novelty remains, and I would like to maintain my score.

AuthorsOpenReview ↗

Official Rebuttal to Reviewer rmqS

We thank the reviewer for considering our response. We recognize that the concern regarding novelty remains, and we would like to clarify precisely what the manuscript does and does not claim.

We do not claim the covariance-contraction utility in Eq. (6), nor the identity relating the second moment of the mean update to the reduction in error covariance; both were established by Bishop et al. (2001). In the revised manuscript, this identity is presented as Lemma 1 and attributed accordingly.

Our contributions are twofold. First, we address a different problem. Bishop et al. (2001) decide where future observations should be made, before any data are collected. UCEG operates when all observations are already available, and selects which of them enter each finite-ensemble correction as the ensemble evolves along the reverse-diffusion trajectory. Under exact Gaussian conditioning, the covariance-contraction utility is monotone in the observation set and is therefore maximized by using every observation. The classical criterion thus cannot account for our central finding that a selected subset outperforms full-observation EnKG (Tables 1–3); this gain arises from finite-ensemble error, which lies outside the idealized model underlying Eq. (6).

Second, we introduce a new selection mechanism. Rather than scoring candidates prospectively, UCEG reuses the particle corrections that EnKG has already computed. The new Proposition 1 shows that the expected particle-level update activity is bounded above and below by the covariance contraction, extending the classical mean-level identity to the quantity that the sampler actually records. At matched observation budgets, replacing only the monitor-selected units with randomly chosen ones increases reconstruction error on all three PDE tasks (Tables 1, 9 and 10) and reduces covariance contraction in 9 of 10 paired FWI cases (Table 8).

We appreciate that the reviewer may weigh these contributions differently, and we hope this clarifies where our contribution lies relative to Bishop et al. (2001).

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

This paper proposes UCEG, a diffusion-based derivative-free algorithm for solving inverse problems. It improves EnKG by selecting a subset of informative observation units to perform the ensemble update, which could be computationally more efficient and bring some regularization effect. The authors test UCEG on three PDE-based inverse problems and find consistent gain over EnKG with less computational cost.

Critical strengths
  1. The proposed UCEG achieves better reconstruction accuracy across three PDE inverse problems. In FWI, it also substantially reduce runtime.
  2. The idea of selecting observation units based on the informativeness is well-motivated with theoretical intuition.
Critical weaknesses
  • The paper introduces several additional practical knobs on top of EnKG. These knobs seem important to the empirical performance but their constructions are poorly explained. Unnecessary abstraction and equations make Sec 3.4 hard to understand.
    • Since mi=Hˉi−1m_i=\bar{H}_{i-1} throughout the paper, introducing operator M is unnecessary abstraction.
    • It's unclear how uncertainty extra budget is implemented. What is the intuition behind this design of ViV_i? Why is it different from mim_i used in Eq 13?
    • Does effective uniform ratio also apply to extra budget?
    • Is the normalization in Eq 12 necessary? The scaling factor will be normalized by πi\pi_i anyways?
Questions for authors
  • How is wq(r)w_q(r) exactly constructed for each task? What's the intuition behind the design? The projection kernel seems central to this method but the authors provide limited information or experiments about this critical design.
  • Table 4 shows consistently better performance as the uniform ratio decreases, but where is the zero baseline? The current results do not support that a non-zero uniform ratio is beneficial.
4 comments
AuthorsOpenReview ↗

Official Rebuttal to Reviewer AyTK (1/2)

Weakness 1: Practical Formulation and Normalization

We thank the reviewer for these constructive suggestions and apologize for the unclear presentation of Section 3.4. We have revised this section to clarify how particle-update activity is mapped to observation scores and used for subsequent EnKG corrections, while removing unnecessary notation.

Clarifying the Selection Mechanism.

The key idea is to use the particle-update activity from an EnKG correction to guide observation selection at the next correction step. Proposition 1 establishes the relationship between expected mean-update energy and covariance contraction under local linear--Gaussian conditioning. Our new Appendix H further establishes a mathematical connection between the practical particle-update activity and the theoretical covariance-contraction utility.

Specifically, the particle corrections provide a spatial update-activity map Hi−1(r)H_{i-1}(r). Since candidate observations are associated with different spatial regions, we use the projection kernel ωq(r)\omega_q(r) to map this activity into an observation-level score:

ai(q)=∑r∈Ωxωq(r)Hi−1(r)∑r∈Ωxωq(r)+ϵ0.a_i(q)=\frac{\sum_{r\in\Omega_x}\omega_q(r)H_{i-1}(r)}{\sum_{r\in\Omega_x}\omega_q(r)+\epsilon_0}.

For spatial measurements, this projection evaluates the activity associated with the measurement location, while for FWI it incorporates source--receiver acquisition geometry. The resulting scores are normalized into sampling probabilities, πi(q)=ai(q)/∑q~∈Qai(q~)\pi_i(q)=a_i(q)/\sum_{\tilde q\in\mathcal Q}a_i(\tilde q), and used to select observations for the next EnKG correction. Repeating this process along the reverse diffusion trajectory enables observation selection to adapt to the evolving ensemble corrections.

Simplifying the Notation and Normalization.

We agree that introducing the operator M\mathcal M is unnecessary since the identity monitor is used throughout our experiments. We remove this operator and directly use Hi−1(r)H_{i-1}(r) in the projection in our revised paper.

We also thank the reviewer for pointing out the redundant normalization in previous Eq. (12). Since this normalization scales the entire spatial activity map by a common factor, the projected observation scores are scaled by the same factor, which cancels during the final normalization into sampling probabilities. Therefore, this additional normalization is unnecessary for observation selection, and we have removed it from the main formulation.

We hope that these revisions clarify the connection between the theoretical motivation, practical particle-update activity, and subsequent observation selection while simplifying Section 3.4.

Weakness 2: Uncertainty Extra Budget and Observation Allocation

We thank the reviewer for pointing out the need to clarify the uncertainty extra budget and its relationship to the main observation-selection mechanism.

Uncertainty Extra Budget.

The uncertainty extra budget introduces additional observations based on the current ensemble uncertainty. Specifically, we compute the spatial uncertainty map

\qquad \hat{x}_{N,i}^{(j)}=\phi_i(x_i'^{(j)}).$$ and project it onto candidate observation units using the same projection in Eq 12, replacing the update-activity map with $V_i$. The additional observations are then sampled from the remaining candidates outside the main observation budget. In our FWI experiments, the uncertainty extra budget is set to $5\%$ and activated during the final three EnKG corrections. The two selection criteria provide complementary information. Update activity reflects the spatial changes induced by the preceding EnKG correction, while ensemble uncertainty identifies regions where the current particles still exhibit substantial disagreement. The uncertainty extra budget therefore allows UCEG to explore uncertain regions that may exhibit limited update activity. **Relationship to the Effective Uniform Ratio.** The effective uniform ratio applies only to the main observation budget. Within this budget, it determines the proportion allocated to uniform coverage, while the remaining observations are selected according to the update-activity scores. Uniform coverage helps prevent repeated selection of a limited set of observations, particularly during early corrections, while maintaining broader coverage of the observation space. The uncertainty extra budget is specified separately and selects additional observations using the uncertainty scores. It is not divided according to the effective uniform ratio. We have clarified these distinctions in Section 3.4 to make the observation-selection procedure and the role of each budget explicit.
AuthorsOpenReview ↗

Official Rebuttal to Reviewer AyTK (2/2)

Weakness 3: Projection Kernel Construction and Validation

We thank the reviewer for highlighting the importance of the projection kernel and apologize for the insufficient explanation of its construction. We have clarified its design intuition and task-specific implementation in the revised manuscript.

Projection Design.

Our projection follows a simple intuition: observations associated with regions exhibiting stronger particle-update activity are expected to provide more useful information for subsequent EnKG corrections. For spatial measurements, nearby observation units are more closely associated with the corresponding state regions. We therefore use spatial proximity and acquisition geometry to construct observation-selection scores.

Specifically, given the update-activity map Hi−1(r)H_{i-1}(r), we compute

ai(q)=∑r∈Ωxωq(r)Hi−1(r)∑r∈Ωxωq(r)+ϵ0,a_i(q)=\frac{\sum_{r\in\Omega_x}\omega_q(r)H_{i-1}(r)}{\sum_{r\in\Omega_x}\omega_q(r)+\epsilon_0},

where ωq(r)\omega_q(r) describes the spatial association between observation unit qq and state location rr.

For FWI, we construct the projection using source--receiver acquisition geometry. For receiver qq, we define

\frac{1}{\\|r-r_q\\|+d_0} \left( \frac{1}{|\mathcal S|} \sum_{s\in\mathcal S} \frac{1}{\\|r-r_s\\|+d_0} \right),$$ where $r_q$ and $r_s$ denote the receiver and source positions, $\mathcal S$ is the selected source set, and $d_0>0$ prevents singularities. The two terms represent receiver proximity and average source illumination, respectively. Source scores are constructed analogously using source proximity and receiver illumination. For Navier--Stokes and Darcy flow, the state and observation grids have the same dimensions, so we directly assign the update activity at each grid location to its corresponding observation unit. **Methodological Contribution.** We adopt these simple, geometry-informed projections to efficiently convert spatial update activity into observation-selection scores without additional forward evaluations. The main contribution of UCEG is the use of realized EnKG particle corrections to guide adaptive observation selection throughout reverse diffusion. Our new Appendix H further establishes the mathematical connection between practical particle-update activity and the theoretical covariance-contraction utility. The projection framework also allows more detailed kernels incorporating physical sensitivities, wave propagation, or spatial correlations to be explored in future work. ## Weakness 4: Missing Zero-Uniform-Ratio Baseline We thank the reviewer for pointing out the missing zero-uniform-ratio baseline. We have added the 0% setting to Table 4 and updated Section 4.3 to include the new results and corresponding discussion, while keeping all other experimental configurations unchanged, including the separate 5% uncertainty extra budget. | Uniform ratio | 0% | 25% | 50% | 90% | |---|---|---|---|---| | Rel. $\ell_2 \downarrow$ | $0.0235\pm0.0069$ | **$0.0213\pm0.0040$** | $0.0218\pm0.0040$ | $0.0220\pm0.0041$ | | PSNR $\uparrow$ | $14.783\pm2.135$ | **$15.432\pm1.412$** | $15.259\pm1.462$ | $15.151\pm1.435$ | | SSIM $\uparrow$ | **$0.484\pm0.064$** | $0.481\pm0.061$ | $0.479\pm0.060$ | $0.477\pm0.062$ | The additional baseline shows that a 25% uniform ratio achieves the lowest relative $\ell_2$ error and highest PSNR, reducing relative $\ell_2$ error by approximately 9.4% compared with the zero-uniform setting. The results indicate that moderate uniform coverage complements history-guided selection by maintaining spatial coverage while preserving sufficient budget for adaptive allocation. Increasing the uniform ratio beyond 25% reduces the monitor-controlled budget and leads to slightly worse relative $\ell_2$ error and PSNR. Although the zero-uniform setting achieves the highest SSIM, its advantage is small relative to the reported variability. Overall, the new ablation supports a moderate uniform-coverage allocation rather than either removing explicit coverage or relying predominantly on uniform selection.

Thanks for the additional experiments and paper revisions. It adds a lot of clarity and addresses my questions. While the overall work feels somewhat incremental engineering from a methodological standpoint, I believe the empirical gains are useful. So I will maintain my score.

AuthorsOpenReview ↗

Official Response to Reviewer AyTK

We sincerely thank the reviewer for reviewing our revisions and for the constructive comments, which have substantially improved the clarity of Section 3.4 and the ablation study. We have also moved the theoretical link between the selection signal and uncertainty contraction into the main text (the new Proposition 1), which we hope further clarifies the methodological basis of UCEG. We are grateful for the reviewer's support of this work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.