Pith. sign in

REVIEW 3 major objections 4 minor 53 references

Two distributions can be distinguished by the midpoint-conditional mean displacement; a sample-split studentized alignment test on this field is asymptotically normal under the null and consistent under alternatives with positive signal.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 07:06 UTC pith:3FMPCOKZ

load-bearing objection A well-designed sample-split test whose identification rests on an unproved self-cited theorem—worth refereeing, but the zero-flow condition needs to be proved or independently verified. the 3 major comments →

arxiv 2607.21542 v1 pith:3FMPCOKZ submitted 2026-07-23 cs.LG stat.ML

Zero-Flow Two-Sample Tests

Yakun Wang , Leyang Wang , Song Liu , Taiji Suzuki This is my paper
classification cs.LG stat.ML MSC 62G1062G2062H15
keywords two-sample testingzero-flow discrepancymidpoint velocitywitness learningsample splittingsign-flip calibrationflow matchinghypothesis testing
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper is trying to establish that the zero-flow criterion — the midpoint-conditional mean displacement between paired samples — is both a valid population measure of distributional difference and a workable basis for a practical two-sample test. It defines the Zero-Flow Discrepancy, proves it equals zero exactly when the two distributions are identical, and then builds ZF2ST, a sample-splitting test that learns a vector witness on a training split and evaluates its alignment with held-out displacements. The central claim is that the held-out studentized statistic is asymptotically standard normal under the null for any fixed witness, and that the test is consistent whenever the learned witness has positive average alignment with the true midpoint velocity, so flexible neural-network witnesses can be used without breaking type-I error control. A sympathetic reader would care because this offers a way to convert the representational power of learned flow fields into statistically calibrated tests for distribution shift, model criticism, and dataset comparison.

Core claim

On its own terms, the paper's central discovery is that the vector field of expected midpoint displacement—the zero-flow velocity v(m)=E[Y−X | (X+Y)/2=m]—vanishes identically if and only if the two source distributions coincide. The paper defines the Zero-Flow Discrepancy as the expected squared norm of this field, proves a scale-invariant dual representation in which ZFD is the supremum over witness fields of squared alignment divided by witness norm, and then constructs ZF2ST. The test learns a witness field on a training split, forms fixed pairs on a held-out split, and studentizes the mean of per-pair alignment scores. Conditional on the training data, the witness is fixed, so under the

What carries the argument

The central object is the midpoint velocity field v(m)=E[Y−X | (X+Y)/2=m], the expected displacement of paired samples given their midpoint; the zero-flow criterion says this field vanishes everywhere exactly when P=Q. The test evaluates a learned vector field u (the witness) through the signed alignment T(P,Q;u)=E[⟨u(M),D⟩], which is zero under the null for any fixed u. By splitting the data, learning u on a training fold and evaluating on an independent test fold, the per-pair alignment scores are i.i.d. conditionally on u, so the studentized mean obeys a standard normal CLT. The power-optimized witness objective maximizes the ratio of alignment to its standard deviation, and sign-flip ran

Load-bearing premise

The test's validity rests entirely on the zero-flow theorem—that a zero conditional mean displacement at every midpoint forces the two distributions to be equal—which this paper imports from an earlier preprint and does not prove here; if that theorem fails for some distributions, the null centering and consistency claims would fail with it.

What would settle it

Search for two distinct distributions P and Q with finite second moments for which E[Y−X | (X+Y)/2 = m] = 0 holds for every midpoint m. A single such pair would make the zero-flow discrepancy zero despite P≠Q, breaking the identification that the test depends on; the search could begin with atomic distributions or distributions whose characteristic functions vanish on a set.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If ZF2ST is correct, two-sample testing can be done with flexible neural-network witnesses without risk of biased null calibration, provided witness learning is confined to a separate split.
  • The zero-flow discrepancy is a valid population divergence: it is nonnegative, symmetric, and zero exactly when the distributions are equal, so it can be reported as an interpretable measure of mismatch.
  • The test is consistent for any fixed alternative with positive signal-to-noise ratio, so power approaches one as the evaluation sample grows whenever the learned witness is positively aligned with the true midpoint velocity.
  • The scale-invariant dual form implies only the direction of the witness matters for the population discrepancy, not its magnitude, which justifies direction-focused training objectives.
  • The paired sign-flip procedure gives finite-sample type-I error control under the null without retraining or re-evaluating the witness.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The zero-flow criterion, if it holds broadly, suggests a deeper connection between midpoint conditional moments and distributional equality; this might yield other discrepancy measures based on higher-order conditional moments or conditional characteristic functions.
  • One could extend the sample-splitting scheme beyond midpoints to time-dependent flow velocities, potentially giving a family of flow-based tests that trade off local sensitivity and sample complexity.
  • The paper's own experiments suggest a limitation: for highly localized within-mode changes, random cross-pairing can dilute the signal; a variant that adaptively selects informative pairs or uses multiple pairings during evaluation might recover power at the cost of more complex dependence structure.
  • Because the null centering does not depend on witness accuracy, the same framework could be used for distributional comparison in high-dimensional settings where the witness is a deep feature representation, without needing a separate calibration dataset beyond the split.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes a two-sample test based on the zero-flow criterion. For independent X~P and Y~Q, with midpoint M=(X+Y)/2 and displacement D=Y-X, the population midpoint velocity v(m)=E[D|M=m] is used to define the Zero-Flow Discrepancy ZFD(P,Q)=E||v(M)||^2. The paper claims that ZFD is a valid discrepancy (zero iff P=Q) and develops ZF2ST, a sample-split testing procedure: a witness field u is learned on one half of the data and evaluated on the other half through a studentized alignment statistic T(P,Q;u)=E<u(M),D>. Conditional on the training split, the held-out alignment scores are i.i.d., yielding an asymptotic normal null distribution; under the alternative, power is characterized by a signal-to-noise ratio. A paired sign-flip calibration is also proposed and shown to control type-I error exactly. Experiments on Gaussian-mixture, Blob, and MNIST benchmarks compare two witness-learning variants (regression and max-SNR) against MMD and classifier-based baselines, with ZF2ST performing well on structured high-dimensional alternatives.

Significance. If the zero-flow criterion is valid, ZF2ST provides an interpretable vector-valued witness for two-sample testing and benefits from a clean sample-splitting construction that decouples witness learning from null calibration. The exact sign-flip calibration is a useful finite-sample guarantee, and the SNR-based power analysis gives practical guidance for witness learning. The experiments are reasonably comprehensive and suggest that the method can outperform strong baselines on structured alternatives. However, the paper's core identification claim rests on an unproved, self-cited theorem, and the consistency result is conditional on an oracle property of the learned witness. These gaps currently limit the theoretical foundation of the method.

major comments (3)
  1. [Section 2, Proposition 2.1; Section 3, Proposition 3.1] The identification of ZFD (Prop 3.1) depends entirely on the zero-flow criterion (Prop 2.1), which is stated without proof and cited to a self-published preprint (Wang et al., 2026). The proof in Appendix A.1 simply invokes this theorem. The converse direction—v≡0 implies P=Q—is nontrivial and is what guarantees power against every alternative. If the criterion fails for some distributions, ZFD can vanish for P≠Q and the consistency claim in Prop 4.2 becomes empty. Please provide a self-contained proof of Prop 2.1 (or a detailed verification) in an appendix; as written, the paper does not establish the central identification claim.
  2. [Section 4.3, Proposition 4.2] The consistency result is conditional on the learned witness u_tr having T(P,Q;u_tr)>0. This condition is not guaranteed by the proposed learning rules (8) and (9) and is not a property of the alternative alone. The statement 'under any fixed alternative H1 for which T...=c′>0' is an oracle conditional power analysis, not a consistency theorem for the procedure. To substantiate the claim that ZF2ST is consistent, the authors need to prove that with high probability the learned witness yields positive alignment when P≠Q (e.g., under universal approximation and bounded optimization error), or explicitly label Prop 4.2 as an oracle result. Section 7 also lists convergence of witness learning as future work, confirming this gap.
  3. [Abstract and Section 1] The abstract and contributions state 'We prove the validity of ZFD.' Given that Proposition 3.1's identification is imported from the unproved Proposition 2.1, this statement overstates what is established. Please rephrase to attribute the zero-flow criterion to Wang et al. (2026) and make the conditional nature of the theoretical results transparent.
minor comments (4)
  1. [Section 4.2, Eq. (5)] The statistic bZZF is a t-statistic. The text states that 'for sufficiently large Nte, Theorem 4.1 gives the asymptotic right-tail p-value.' It should be clarified that this Gaussian calibration is an approximation and that the experiments correctly use the exact sign-flip calibration.
  2. [Section 4.3, Eq. (7)] The power approximation (7) is derived informally by replacing bσ_te with σ in the signal term. While the direct consistency proof in Appendix A.4 is rigorous, (7) should be clearly labeled as an approximation, not an asymptotic equality.
  3. [Section 5, Figure 1] In the HDGM dimension experiment, only one sample size (n=3000) is used. It would be informative to report results at a smaller sample size as dimension increases, since the paper's motivation is high-dimensional alternatives where power is typically limited.
  4. [References] Wang et al. (2026) is cited as an arXiv preprint. Since Proposition 2.1 is load-bearing, the authors should ensure the preprint is publicly available and, ideally, include a version identifier or DOI.

Circularity Check

1 steps flagged

ZFD identification is imported from the same authors' zero-flow theorem, which is load-bearing but not proved here; the rest of the derivation is self-contained.

specific steps
  1. uniqueness imported from authors [Section 2 'Zero-flow criterion' (Proposition 2.1); Section 3.1 Proposition 3.1; Appendix A.1]
    "Recently, a key property of the population FM velocity at the midpoint was established by Wang et al. (2026): the midpoint velocity vanishes if and only if the two endpoint distributions coincide. ... Proposition 2.1 (Zero-flow condition; Theorem 3.1 in Wang et al. (2026)). Let X∼P and Y∼Q be independent random variables with finite second moments. Then v(z; 1/2)=0, ∀z if and only if P=Q."

    The central identification claim, Proposition 3.1(1), is proved in Appendix A.1 only by 'By Proposition 2.1, this is equivalent to P=Q.' The nontrivial direction (v=0 implies P=Q) is not derived in this manuscript; it is quoted as Theorem 3.1 of the authors' own preprint Wang et al. (2026). This converse is load-bearing for ZFD's validity and for Proposition 4.2's consistency guarantee: if it fails, T(P,Q;u)=0 for all u even when P≠Q, so the test has no power. Since the cited theorem is neither machine-checked nor independently reproduced here, the paper's validity claim is not self-contained at this step.

full rationale

The statistical core is otherwise non-circular. Proposition 4.1's dual representation follows from the tower identity and Cauchy-Schwarz; Theorem 4.1 is a conditional CLT plus Slutsky argument for a fixed held-out witness; Proposition C.1's sign-flip validity follows from exchangeability under H0; and Proposition 4.2 restates the CLT divergence of the SNR term rather than fitting any parameter to the evaluation data. The null centering under H0 only needs the trivial direction (P=Q implies v=0), which is elementary. The genuine circularity concern is the importation of the iff zero-flow theorem from the authors' own prior work, which is the sole support for identification and for the alternative-side power guarantee. That is a load-bearing self-citation rather than an independently verified external result, but the paper retains substantial independent content (sample splitting, sign-flip calibration, experiments). Hence a moderate score of 4.

Axiom & Free-Parameter Ledger

1 free parameters · 5 axioms · 0 invented entities

The central claim rests on an external self-cited zero-flow theorem, a sample-splitting independence assumption, and an unverified positive-alignment condition for the learned witness. The only hand-chosen number is the ZF-SNR regularization lambda. No new physical entities are introduced.

free parameters (1)
  • lambda (ZF-SNR regularization) = 1e-3
    Regularization constant in the max-SNR objective (eq. 9); set to 1e-3 for all benchmarks without tuning. It controls bias-variance of the learned witness but is not central to the validity claim.
axioms (5)
  • domain assumption Zero-flow criterion (Prop 2.1): for independent X~P, Y~Q with finite second moments, E[Y-X | (X+Y)/2 = m] = 0 for all m iff P=Q.
    Imported from the self-cited preprint Wang et al. (2026), Theorem 3.1, not proved here. It is the basis of ZFD's identification (Prop 3.1) and the null centering in Theorem 4.1.
  • domain assumption The learned witness field u_tr is fixed and independent of the evaluation split.
    Sample-splitting protocol; conditional on training data, the witness is deterministic, making held-out scores i.i.d. (Theorem 4.1).
  • domain assumption T(P,Q;u_tr) > 0 for the learned witness under the alternative.
    Required for the consistency claim (Prop 4.2). The paper does not prove that regression or max-SNR witnesses achieve positive alignment; it is an empirical property.
  • standard math Finite second moments of P and Q.
    Assumed in the definition of ZFD and the zero-flow criterion.
  • domain assumption The evaluation pairs are one-to-one fixed pairs, giving independence; using all cross-pairs would introduce U-statistic dependence.
    Stated in Section C.1; needed for the CLT and sign-flip exchangeability.

pith-pipeline@v1.3.0-alltime-deepseek · 15149 in / 14556 out tokens · 138565 ms · 2026-08-01T07:06:46.015319+00:00 · methodology

0 comments
read the original abstract

We propose a new approach to two-sample testing for deciding whether two sets of samples are drawn from the same distribution. The test is built on a statistical discrepancy based on the zero-flow criterion, termed zero-flow discrepancy (ZFD). We prove the validity of ZFD and propose a practical testing procedure, termed the zero-flow two-sample test (ZF2ST). The key idea is to learn how samples from the two distributions are locally misaligned and use the resulting directional pattern as evidence of distributional difference. By separating witness learning from hypothesis evaluation, ZF2ST can use flexible neural networks while maintaining valid statistical calibration. We develop both regression-based and power-maximized approaches for learning the witness. Experiments on synthetic and image datasets demonstrate that ZF2ST can achieve strong testing power for structured distributional changes while maintaining well-calibrated type-I error.

Figures

Figures reproduced from arXiv: 2607.21542 by Leyang Wang, Song Liu, Taiji Suzuki, Yakun Wang.

Figure 1
Figure 1. Figure 1: Results on the HDGM benchmark for α = 0.05 (black line). Left: average test power (a) and type-I error (b) as the number of samples per distribution n increases, with d = 10. Right: average test power (c) and type-I error (d) as the dimension d increases, with n = 3000. Shaded regions show one standard error across independent training trials. (a) Blob benchmarks. (b) MNIST rare-contamination benchmark [P… view at source ↗
Figure 2
Figure 2. Figure 2: Results on the Blob and MNIST benchmarks for [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Fixed and dynamic pairing ablation on the Blob benchmarks. [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of synthetic datasets. Left: Blob-S/D; Right: The first two dimensions of HDGM-S/D with (cj ) 9 j=1 = (−.020, −.022, −.024, −.026, 0, .020, .022, .024, .026), where the grid points are ordered row by row. HDGM-S/D. HDGM is an equally weighted two-component Gaussian mixture with means µ0 = 0, µ1 = 0.5 × 1d. Distribution P has identity covariance in both components. Under HDGM-D, the covariance… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

53 extracted references · 7 linked inside Pith

  1. [1]

    Advances in Neural Information Processing Systems , volume=

    A permutation-free kernel two-sample test , author=. Advances in Neural Information Processing Systems , volume=

  2. [2]

    International conference on machine learning , pages=

    Learning deep kernels for non-parametric two-sample tests , author=. International conference on machine learning , pages=. 2020 , organization=

    2020

  3. [3]

    arXiv preprint arXiv:1610.06545 , year=

    Revisiting classifier two-sample tests , author=. arXiv preprint arXiv:1610.06545 , year=

    Pith/arXiv arXiv

  4. [4]

    Neural networks , volume=

    Least-squares two-sample test , author=. Neural networks , volume=. 2011 , publisher=

    2011

  5. [5]

    The journal of machine learning research , volume=

    A kernel two-sample test , author=. The journal of machine learning research , volume=. 2012 , publisher=

    2012

  6. [6]

    Journal of Machine Learning Research , volume=

    Estimation of non-normalized statistical models by score matching , author=. Journal of Machine Learning Research , volume=

  7. [7]

    arXiv preprint arXiv:2602.00797 , year=

    Zero-Flow Encoders , author=. arXiv preprint arXiv:2602.00797 , year=

    Pith/arXiv arXiv

  8. [8]

    arXiv preprint arXiv:2512.05150 , year=

    TwinFlow: Realizing One-step Generation on Large Models with Self-adversarial Flows , author=. arXiv preprint arXiv:2512.05150 , year=

    arXiv

  9. [9]

    arXiv preprint arXiv:2602.04770 , year=

    Generative Modeling via Drifting , author=. arXiv preprint arXiv:2602.04770 , year=

    Pith/arXiv arXiv

  10. [10]

    2000 , publisher=

    Asymptotic statistics , author=. 2000 , publisher=

    2000

  11. [11]

    2006 , publisher=

    All of nonparametric statistics , author=. 2006 , publisher=

    2006

  12. [12]

    Foundations of Computational mathematics , volume=

    Optimal rates for the regularized least-squares algorithm , author=. Foundations of Computational mathematics , volume=. 2007 , publisher=

    2007

  13. [13]

    Journal of Machine Learning Research , volume=

    Adaptive approximation and generalization of deep neural network with intrinsic dimensionality , author=. Journal of Machine Learning Research , volume=

  14. [14]

    Nonparametric regression on low-dimensional manifolds using deep

    Chen, Minshuo and Jiang, Haoming and Liao, Wenjing and Zhao, Tuo , journal=. Nonparametric regression on low-dimensional manifolds using deep. 2022 , publisher=

    2022

  15. [15]

    Nonparametric regression using deep neural networks with

    Schmidt-Hieber, Johannes , journal=. Nonparametric regression using deep neural networks with

  16. [16]

    Adaptivity of deep

    Suzuki, Taiji , booktitle=. Adaptivity of deep

  17. [17]

    Deep learning is adaptive to intrinsic dimensionality of model smoothness in anisotropic

    Suzuki, Taiji and Nitanda, Atsushi , booktitle=. Deep learning is adaptive to intrinsic dimensionality of model smoothness in anisotropic

  18. [18]

    Dimension-agnostic inference using cross

    Kim, Ilmun and Ramdas, Aaditya , journal=. Dimension-agnostic inference using cross

  19. [19]

    The Econometrics Journal , volume=

    Double/debiased machine learning for treatment and structural parameters , author=. The Econometrics Journal , volume=

  20. [20]

    Journal of Econometrics , volume=

    Consistent model specification tests , author=. Journal of Econometrics , volume=

  21. [21]

    Econometrica , pages=

    A consistent conditional moment test of functional form , author=. Econometrica , pages=

  22. [22]

    Annals of Statistics , volume=

    Classification accuracy as a proxy for two-sample testing , author=. Annals of Statistics , volume=

  23. [23]

    A kernelized

    Liu, Qiang and Lee, Jason and Jordan, Michael , booktitle=. A kernelized

  24. [24]

    International Conference on Artificial Intelligence and Statistics , pages=

    A witness two-sample test , author=. International Conference on Artificial Intelligence and Statistics , pages=. 2022 , organization=

    2022

  25. [25]

    arXiv preprint arXiv:1611.04488 , year=

    Generative models and model criticism via optimized maximum mean discrepancy , author=. arXiv preprint arXiv:1611.04488 , year=

    Pith/arXiv arXiv

  26. [26]

    2019 , publisher=

    High-dimensional statistics: A non-asymptotic viewpoint , author=. 2019 , publisher=

    2019

  27. [27]

    2002 , publisher=

    A distribution-free theory of nonparametric regression , author=. 2002 , publisher=

    2002

  28. [28]

    goodness of fit

    Asymptotic theory of certain" goodness of fit" criteria based on stochastic processes , author=. The annals of mathematical statistics , pages=. 1952 , publisher=

    1952

  29. [29]

    2009 , publisher=

    Optimal transport: old and new , author=. 2009 , publisher=

    2009

  30. [30]

    2008 , publisher=

    Introduction to empirical processes and semiparametric inference , author=. 2008 , publisher=

    2008

  31. [31]

    1993 , publisher=

    Efficient and adaptive estimation for semiparametric models , author=. 1993 , publisher=

    1993

  32. [32]

    Advances in Neural Information Processing Systems , volume=

    Interpretable distribution features with maximum testing power , author=. Advances in Neural Information Processing Systems , volume=

  33. [33]

    International conference on machine learning , pages=

    A kernel test of goodness of fit , author=. International conference on machine learning , pages=. 2016 , organization=

    2016

  34. [34]

    IEEE Transactions on Information Theory , volume=

    Classification logit two-sample testing by neural networks for differentiating near manifold densities , author=. IEEE Transactions on Information Theory , volume=. 2022 , publisher=

    2022

  35. [35]

    arXiv preprint arXiv:2209.03003 , year=

    Flow straight and fast: Learning to generate and transfer data with rectified flow , author=. arXiv preprint arXiv:2209.03003 , year=

    Pith/arXiv arXiv

  36. [36]

    2013 , publisher=

    Permutation tests: a practical guide to resampling methods for testing hypotheses , author=. 2013 , publisher=

    2013

  37. [37]

    Behaviormetrika , volume=

    The bootstrap method for assessing statistical accuracy , author=. Behaviormetrika , volume=. 1985 , publisher=

    1985

  38. [38]

    Conference on computer vision and pattern recognition , year=

    Masked autoencoders are scalable vision learners , author=. Conference on computer vision and pattern recognition , year=

  39. [39]

    International Conference on Machine Learning , year=

    A simple framework for contrastive learning of visual representations , author=. International Conference on Machine Learning , year=

  40. [40]

    International Conference on Learning Representations , year=

    Score-Based Generative Modeling through Stochastic Differential Equations , author=. International Conference on Learning Representations , year=

  41. [41]

    International Conference on Learning Representations , year=

    Flow Matching for Generative Modeling , author=. International Conference on Learning Representations , year=

  42. [42]

    2009 , publisher=

    Approximation theorems of mathematical statistics , author=. 2009 , publisher=

    2009

  43. [43]

    Journal of Machine Learning Research , volume=

    Stochastic interpolants: A unifying framework for flows and diffusions , author=. Journal of Machine Learning Research , volume=

  44. [44]

    Exact and asymptotically robust permutation tests , author=

  45. [45]

    Econometrica , volume=

    Randomization tests under an approximate symmetry assumption , author=. Econometrica , volume=. 2017 , publisher=

    2017

  46. [46]

    Test , volume=

    Exact testing with random permutations , author=. Test , volume=. 2018 , publisher=

    2018

  47. [47]

    arXiv preprint arXiv:1603.05766 , year=

    Permutation P-values should never be zero: calculating exact P-values when permutations are randomly drawn , author=. arXiv preprint arXiv:1603.05766 , year=

    Pith/arXiv arXiv

  48. [48]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    On the decreasing power of kernel and distance based nonparametric hypothesis tests in high dimensions , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  49. [49]

    Proceedings of the IEEE , volume=

    Gradient-based learning applied to document recognition , author=. Proceedings of the IEEE , volume=. 1998 , publisher=

    1998

  50. [50]

    Advances in Neural Information Processing Systems , volume=

    Failing loudly: An empirical study of methods for detecting dataset shift , author=. Advances in Neural Information Processing Systems , volume=

  51. [51]

    arXiv preprint arXiv:2605.29920 , year=

    Midpoint Generative Models , author=. arXiv preprint arXiv:2605.29920 , year=

    Pith/arXiv arXiv

  52. [52]

    arXiv preprint arXiv:2511.08552 , year=

    Fmmi: Flow matching mutual information estimation , author=. arXiv preprint arXiv:2511.08552 , year=

    arXiv

  53. [53]

    Advances in Neural Information Processing Systems , volume=

    Missing data imputation by reducing mutual information with rectified flows , author=. Advances in Neural Information Processing Systems , volume=