- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 1
Read review
Paper summary
CoreFlow is a generative model for matrix-valued data learned from limited and possibly incomplete training samples. It first learns shared orthonormal row and column bases using alternating masked fitting and low-rank imputation, then fits a flow-matching CNF to the vectorized cores. Its main bound decomposes the Wasserstein error into subspace, core-modeling, and residual terms. Experiments on four real and four synthetic datasets report better performance than ambient-space diffusion and flow baselines, particularly with limited data and substantial missingness, as well as faster training.
Critical strengths
S1. Strong robustness to missingness where the held-out evaluation can be verified. ERSSTv5 uses a temporal split and a freeze-then-test pipeline. At 40% missingness, CoreFlow substantially outperforms MissFlow and MissDiff and remains stable across three training seeds.
S2. The ablations include appropriate controls. SMG-Core combines the learned subspaces with a Gaussian core, while PCA-Flow combines the same CNF with a flattened PCA representation. CoreFlow clearly outperforms SMG-Core on Solar and Solar 2.
Critical weaknesses
W1. The stated held-out protocol and several reported values cannot be verified from the released artifacts.**
-
Evaluation reference. Appendix C describes held-out sets for Solar, Solar 2, and LSPF, and §5 says simulations use complete held-out matrices. The released evaluators instead use the training matrices; only ERSSTv5 includes a held-out script and separate evaluation tensors. The artifacts therefore do not establish which reference sets produced the main Solar and LSPF results or permit a memorization check.
-
Configurations and LSPF. Table 7 says that
configs/papercontains every setting, but it omits LSPF at 20% missingness and four of twelve simulation settings;configs/baselineandresults/paper, both listed in the README, are absent. The README and report script also say that learning rates were selected to match the paper's values. Tables 1 and 23 report 2.70±0.19 for CoreFlow on LSPF at 0% missingness, while the script records 2.88±0.19 and the Table 11 runs average 2.846. At 2.88, PCA-Flow has the lower SVRelL2 mean (2.74), although the intervals overlap. -
Other discrepancies. The paper describes five independent sets of 500 samples, whereas the code splits one generated batch into 20 subsets. MMD is reported but not implemented in the repository, and the LSPF 40% configuration uses a 72×72 crop rather than the size stated in Appendix C.3.
W2. The source of the gains is only partly isolated.
- PCA-Flow. Quantitative results cover only LSPF and complete-data simulations; ERSSTv5 appears only qualitatively. PCA-Flow matches CoreFlow on LSPF SVRelL2, while CoreFlow is better on the other metrics. Because PPCA and EM-PCA support missing values, an analogous impute-and-refit baseline could test whether matrix structure matters under missingness.
- Simulation comparison. The 576-dimensional PCA subspace minimizes training reconstruction error among subspaces of that size, yet CoreFlow is reported to be 2–5.5 times better. The gap therefore requires an explanation beyond retained energy, such as layout, normalization, or tuning.
- SMG-Core. On Blobs, the true core is Gaussian, the sample size exceeds the core dimension, and the recovered subspaces are exact to the reported precision. Table 3 nevertheless reports 195 for SMG-Core versus 0.43 for CoreFlow, and SMG-Core improves by roughly an order of magnitude when its assumed rank exceeds the true rank. Without the prior settings or code, this comparison is difficult to interpret.
- The ambient baselines use weak missing-data handling; an impute-then-generate or EM-based baseline would be more informative.
Questions for authors
- Please release the non-ERSSTv5 held-out data and evaluator, original configurations, and Solar nearest-training-neighbor distances.
- What produced the Solar block entries in Table 12 and the Solar 2 heterogeneous value of 0.59, given the archived value of 0.94?
- How does PCA-Flow with EM-PCA perform at 20% and 40% missingness? Please also report its quantitative ERSSTv5 results.
- Why does Table 5 show non-negligible 25th singular values for data whose generator implies rank at most 24?
Minor:
- Please document mean-fill initialization, the use of centering despite the uncentered statement of Proposition 4.3, core standardization, and how the additional Table 11 seeds were set despite a hard-coded seed.
- Relevant missing work includes MisGAN, Ambient Diffusion, as well as Davis–Kahan perturbation theory for Proposition 4.3.
3 comments
Response to Reviewer jmMe [Part 2/2]
[Part 2/2] of our response
4. Completion followed by PCA-Flow, and ERSSTv5 (W2; Q3).
Thank you for this suggestion! We agree that the key question is whether the representation helps beyond the completion step. Following your advice, we implemented two PCA-Flow ablations to examine this:
- PCA-Flow (Stage I) uses the same completed training matrices as CoreFlow, then replaces the matrix cores with PCA scores. This tests the representation used for generation.
- PCA-Flow (iterative PCA) alternates PCA fitting with reconstruction of missing entries, keeping observed entries fixed. It then trains the flow on PCA scores. This uses iterative PCA imputation.
Appendix D.8 now describes the implementations. Table 24 gives the full results. The LSPF spectral errors are shown below (SVRelL2 ×100; lower is better). Missingness CoreFlow PCA-Flow (iterative PCA) 20% 4.11 ± 0.07 4.66 ± 0.03 40% 7.12 ± 0.09 10.12 ± 0.11 With the same Stage-I completion, neither representation is best on every metric. At 20% missingness, Stage-I + PCA-Flow has lower spectral error (3.20 versus 4.11). CoreFlow has lower error in the mean Frobenius norm (0.17 versus 0.46).
For ERSSTv5, we now report quantitative results in Table 24. CoreFlow and PCA-Flow have spectral errors of 21.26 ± 0.46 and 23.46 ± 1.16, respectively. The table also reports the moment discrepancies and MMD.
PCA-Flow is a representation ablation we developed within our framework. It tests replacing matrix cores with PCA scores and can also give competitive results.
We keep CoreFlow as our main formulation because it separates shared matrix structure from variation between samples. The bases capture the shared row and column spaces, while the core models sample-specific variation. Our completion, compression, and generation use the same bases, without an additional PCA fit. Also, our CoreFlow comes from theoretical interpretations as proved in Section 4.
5. PCA reconstruction and generation (W2).
PCA minimizes training reconstruction error at a fixed dimension. Generation also depends on how well the flow learns the distribution of those coordinates. So lower reconstruction error does not guarantee lower generation error. Appendix D.8 now makes this more clear.
6. SMG-Core on Blobs (W2).
Thank you for catching this! Your concern was correct. The relatively large error came from covariance scaling. The executed sampler fitted an empirical Gaussian to the Stage-I cores. It was divided by the compact-SVD dimension minus one, rather than the training sample count minus one.
Now, we used matched inference seeds 0, 1, and 2 for SMG-Core and CoreFlow, and updated all related tables. Appendix C.7 now describes the corrected implementation. On complete Blobs, SMG-Core has lower spectral error: 0.28 ± 0.04 versus 1.34 ± 0.06 for CoreFlow (Table 3, ×100 units).
The corrected comparisons still show notably lower spectral errors for CoreFlow under missingness (Tables 2–3). We wish to note that SMG-Core already uses the same Stage-I subspaces learned by CoreFlow, including those estimated from incomplete training data. It therefore benefits from CoreFlow’s shared row and column structure before fitting its Gaussian core model.
7. The 25th singular value (Q4).
Thank you for pointing this out. The rank of 24 refers to the simulated data before preprocessing. Table 5 measures the data after preprocessing.
For Waves and Crosshatch, we normalize each matrix to [0, 1] before training. This also shifts all its entries by a constant. The shift differs across matrices, so subtracting the mean training matrix does not fully remove it. That’s why the 25th singular value is non-zero in Waves and Crosshatch.
We checked this by removing the constant component for the spectral calculation only. The 25th-to-first singular-value ratio then falls to 8.64 × 10⁻⁸ for Waves and 7.09 × 10⁻⁸ for Crosshatch, both close to zero. Appendix C.2 (“Rank selection and spectral concentration”), in the paragraph before Table 5, now explains this cause and reports the check.
8. Implementation details and related work (Minor).
Thank you for pointing these out. Section 3 now explains how we initialize missing entries with observed-value means. Appendix C describes centering before the initialization SVD. The supplementary code README now documents core standardization with one mean and one population standard deviation of the training cores. The extra runs in revised Table 9 use training seeds 42, 123, and 2026, as listed in ablations/configs.yaml. Also, we have added MisGAN and Ambient Diffusion to Related Work.
Thank you again for these careful checks! We hope the added experiments and clarifications address your concerns. We are happy to discuss any remaining questions!
Final Follow-up on Our Rebuttal
Dear Reviewer jmMe,
As the discussion period ends today, we would really appreciate it if you could take a look at our rebuttal and revised manuscript. We have put considerable efforts into additional experiments and corrections to address your concerns.
We would be very grateful for any final feedback and would appreciate the last opportunity to clarify any remaining concerns before the discussion closes.
Thank you so much for your time and efforts!
Best, Authors
Response to Reviewer jmMe [Part 1/2]
We’re sorry for the confusion. We mistakenly submitted the code from our previous OpenReview submission. That version did not include the subsequent revisions and experiments. We have now updated the paper and supplementary code on OpenReview supplementary materials. We take responsibility for this mistake and sincerely appreciate your careful checks.
1. Evaluation references and supplementary materials (W1; Q1).
Section 5 and Appendix C now specify the evaluation references:
For Solar and Solar 2, we deliberately use a few-sample setting with 300 training images. All methods use these images as the reference in the main comparisons. This measures fit to the training data. To fully address your concerns, we also added a separate Solar experiment with 240 training images and 60 held-out test images. Table 9 reports three independent training runs on this split at 0% and 40% missingness.
The supplementary code now includes the evaluation reference tensors in
data/, the current CoreFlow settings inconfigs.yaml, and the evaluator. The README links to the larger LSPF and simulation training datasets. Configurations for the independent-training study are inablations/. Appendix C.5 gives the training settings.Unless stated otherwise in the paper, our latest results use 500 generated matrices with each of sampling seeds 0, 1, and 2. We calculate each batch’s metrics separately and report their mean and sample SD (Appendix C). Also, our latest code in supplementary materials now includes the MMD implementation in metrics.py.
The revised LSPF evaluation uses matrices before any PCA reconstruction across samples. We split them into 8,260 training and 500 test matrices before fitting. The complete-data spectral error is 10.34 ± 0.02 under this protocol (Table 1, ×100 units). We also corrected the LSPF size description in Appendix C.3. At 20% and 40% missingness, the 9 × 9 patch construction uses a 72 × 72 crop. Every method is evaluated against reference matrices on that same region. Complete-data evaluation uses the full 80 × 80 fields.
Our latest paper’s results use 500 generated matrices with each of sampling seeds 0, 1, and 2. We calculate each batch’s metrics separately and report their mean and sample SD (Appendix C). Also, the supplementary code includes the MMD implementation in
metrics.py.2. Solar nearest-neighbor check (W1; Q1).
Following your suggestion, Appendix D.10 (Table 26) now reports the Solar nearest-training-neighbor distances. On average, generated images are farther from their nearest training image than training images are from their closest other training image. We found no exact copies among the 1,500 generated images per condition. This means CoreFlow is not simply copying and pasting training images under this extremely few-data training setting. Table 9 separately evaluates generation against held-out Solar images.
3. MMD and the LSPF crop (W1).
Appendix C.1 defines the MMD estimator. For the revised evaluations, the RBF bandwidth is computed only from true reference matrices and shared across methods within each dataset and evaluated region. Our latest code in supplementary materials now includes the MMD implementation in metrics.py.
Each LSPF field is originally 80 × 80. With complete training data, we evaluate the whole field. In the missing-data experiments, CoreFlow applied a patchification with 9 × 9 patch size before training. Since 80 is not divisible by 9, preprocessing keeps the upper-left 72 × 72 region, which fits eight patches along each side. We evaluate all methods on this same region at both 20% and 40% missingness. Appendix C.3 now explains this more clearly.