Pith. sign in

REVIEW 2 major objections 3 minor 2 cited by

A squared Wasserstein distance to a standard Gaussian, maximized over whitened projections, provably recovers independent components from linear mixtures.

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 ↗

A squared-Wasserstein-to-Gaussian contrast is proven to be maximized exactly at independent components, yielding a new ICA algorithm (OT-ICA) that improves separation on heterogeneous continuous mixtures.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection A genuinely new Wasserstein contrast for ICA with a clean proof under smooth densities, but the paper oversells its distribution-free and empirical claims. the 2 major comments →

arxiv 2607.14081 v1 pith:SU4BX3DF submitted 2026-07-15 cs.LG stat.ML

Linear Independent Component Analysis via Optimal Transport

Ashutosh Jha , Michel Besserve , Simon Buchholz This is my paper
classification cs.LG stat.ML MSC 62H2549Q22
keywords Independent Component AnalysisOptimal TransportWasserstein distanceNon-GaussianityContrast functionBlind source separationOrthogonal group optimizationEEG artifact removal
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 reading

Linear ICA recovers hidden independent sources from mixed signals, but classical algorithms rely on approximate non-Gaussianity proxies that can vanish or stall on many distributions. This paper claims that the squared optimal-transport cost W2² between a projected signal and a standard Gaussian is an exact contrast: under mild conditions it is strictly maximized only when the projection isolates one independent source. The authors prove this via a coupling argument and regularity theory, then turn it into OT-ICA, a gradient algorithm that sorts samples instead of estimating densities. On synthetic mixtures, OT-ICA outperforms proxy-based methods on continuous and heterogeneous source types, and it handles applied EEG and price-discovery tasks without distributional assumptions.

Core claim

The paper establishes Theorem 3: for centered independent unit-variance sources with at most one Gaussian and smooth strictly positive densities, every normalized direction α that mixes at least two sources satisfies W2²(α·Z,Γ) < Σ α_i² W2²(Z_i,Γ), so the W2²-to-Gaussian contrast is maximized only at pure source directions (Corollary 1). This transfers the information-geometric equivalence between independence and non-Gaussianity to a metric computable by quantile sorting. The same inequality, summed over rows of an orthogonal matrix, gives Corollary 2, so joint symmetric optimization of all components is theoretically justified. OT-ICA implements this with Riemannian gradient ascent on the

What carries the argument

The central object is the squared L2-Wasserstein distance W2² between a one-dimensional projection and the standard Gaussian Γ, computed by matching sorted samples to analytic Gaussian quantile targets. The proof engine is the inequality W2²(α·Z,Γ) ≤ Σ α_i² W2²(Z_i,Γ), obtained by coupling the mixture to a Gaussian through shared normal noise; strictness follows from Brenier's theorem and Caffarelli regularity, which force equality to imply linear transport maps, contradicting non-Gaussianity. Summing this inequality over the rows of an orthogonal matrix yields the joint-contrast bound that licenses symmetric optimization.

Load-bearing premise

The strict inequality that powers the result is proven only for sources with smooth, strictly positive densities; several experimental distributions (uniform, Beta(0.5,0.5), Laplace, discrete) violate this, so the theoretical guarantee does not cover those regimes.

What would settle it

Take two independent non-smooth sources, e.g., uniform on [-√3,√3] and Laplace, and compute W2²(α·Z,Γ) for a grid of unit-norm α with α₁²+α₂²=1. If the maximum occurs at a mixed α rather than at e₁ or e₂, or if W2²(α·Z,Γ) equals Σ α_i² W2²(Z_i,Γ) for some mixed α, then Theorem 3's strictness claim is false for that class.

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

If this is right

  • An ICA algorithm needs no density model or higher-order cumulant proxy; a sort-and-optimize procedure suffices.
  • The contrast is nonzero for every non-Gaussian distribution, avoiding the zero-gradient failure modes of logcosh and cumulant-based proxies.
  • Joint recovery of all components is theoretically justified by the summed row inequality, so symmetric orthogonal-group optimization is sound rather than heuristic.
  • Applied tasks such as EEG artifact removal and price-discovery identification become solvable without distributional assumptions, with separation quality limited mainly by sample size and dimension.
  • On heterogeneous mixtures where proxy contrasts fail, the paper reports substantially lower Amari error than proxy-based methods, with the margin widening as dimension grows.

Where Pith is reading between the lines

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

  • Because the contrast is defined by sorting rather than density estimation, one could test whether it inherits favorable sample-complexity scaling; the paper's scaling experiments already show a noise ceiling around d=40 at N=10,000, and a finite-sample theory could make that ceiling quantitative.
  • The proof's smooth-density assumption suggests strictness may fail for atomic or nonsmooth sources; a concrete test would be to check the inequality on two-point or uniform sources, which the paper reports are exactly where the solver stalls.
  • The same coupling argument could be adapted to other reference distributions, potentially yielding a family of exact non-Gaussianity contrasts and extending the approach to nonlinear or causal ICA settings.
  • The W2² contrast supplies a distribution-free scalar measure of 'how non-Gaussian is this projection', which could serve as a normality test statistic or as a regularizer in latent-variable models beyond ICA.
Share X Bluesky LinkedIn Reddit HN

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

2 major / 3 minor

Summary. The paper proposes OT-ICA, an ICA method that measures non-Gaussianity by the squared 2-Wasserstein distance W_2^2 of a projected signal to a standard Gaussian. The central theoretical claim (Theorem 3, Appendix B) is that, for centered independent unit-variance sources with at most one Gaussian and smooth strictly positive densities, any proper mixture α·Z satisfies a strict subadditivity bound, W_2^2(α·Z,Γ) < Σ_i α_i^2 W_2^2(Z_i,Γ), so the contrast is maximized only at independent components (Corollary 1). The paper then describes a Riemannian gradient-ascent algorithm on O(d), with analytical Gaussian targets and Gaussian dithering, and reports synthetic benchmarks, EEG artifact removal, and a price-discovery application. The proof's weak upper bound is obtained by an explicit coupling; strictness is shown by a parallel-gradient/Hessian argument. The code is released.

Significance. If Theorem 3 is correct under its stated assumptions, the paper contributes a principled contrast that avoids known proxy failures such as zero negentropy and vanishing curvature, and the explicit-coupling proof is elegant. The release of code and the synthetic benchmarks with known ground truth are positive features. However, the presentation overclaims both the scope of the theory and the empirical superiority: the theorem requires smooth strictly positive source densities, while several benchmark distributions (Uniform, Beta(0.5,0.5), Laplace, discrete counts) violate this, and Table 3 shows FastICA beating OT-ICA on several pure-source regimes. The central idea is sound and the gap is likely fixable, but the manuscript in its current form needs substantial revision of the claims and an analysis of the dithered objective actually optimized.

major comments (2)
  1. [Abstract and §3 (Table 3)] The abstract and conclusion state that OT-ICA outperforms proxy-based methods 'for different distributions of the latent variables' and 'on all continuous and heterogeneous distributions.' Table 3 contradicts this: on Pure Uniform at d=30, OT-ICA gives the clip ceiling 1.500 while FastICA gives 0.179; on Pure Laplace and Pure Student-t at d=30, FastICA is also slightly better (0.261 vs 0.280 and 0.207 vs 0.219). The performance claim should be restricted to the mixed-source configurations where the advantage actually holds, and the pure-source results should be discussed honestly.
  2. [Appendix B (Theorem 3) and Algorithm 1] Theorem 3 assumes every source has a smooth and strictly positive density, and the strictness proof differentiates the optimal transport maps T_i via Caffarelli regularity. The experimental section includes Uniform[-√3,√3], Beta(0.5,0.5), Laplace, and discrete count sources, which violate these assumptions (discontinuous density, unbounded density, non-C^2 map, atomic distributions). Moreover, Algorithm 1 optimizes a Gaussian-dithered objective (line 12), not the plain W_2^2 contrast addressed by the theorem. The paper should either prove the strict inequality for the smoothed/dithered family (or as a limit of smooth approximations), or explicitly mark the non-smooth/discrete results as empirical and remove the 'without distributional assumptions' claim from the Abstract and Conclusion.
minor comments (3)
  1. [Section 5 / Table 3] The conclusion says 'On purely discrete sources, all ICA methods fail as d increases (E>0.5).' At d=10, JADE has E=0.443, which is below the stated failure threshold of 0.5. Please rephrase to 'as d becomes large' or qualify the d=10 case.
  2. [Abstract] The phrase 'without distributional assumptions' in the abstract is ambiguous. The algorithm is nonparametric, but the theoretical guarantee in Theorem 3 explicitly assumes smooth strictly positive densities. Clarify that the distribution-free statement applies to the algorithmic contrast, not to the theorem.
  3. [Appendix C.1, Eq. (25)] The analytical Gaussian target T_i depends on the batch size N. If the batch size changes during optimization, the target must be recomputed. This is not stated in Algorithm 1; please clarify the dependency.

Circularity Check

0 steps flagged

No significant circularity: Theorem 3 and its corollaries are derived from first principles, and the paper's self-citations are not load-bearing.

full rationale

The central claim is Theorem 3, which bounds W_2^2(alpha·Z, Γ) by sum_i alpha_i^2 W_2^2(Z_i, Γ) and proves the bound is strict for proper mixtures. The proof is self-contained: it constructs an explicit coupling using optimal transport maps T_i pushing Γ to Z_i, applies Brenier's theorem and Caffarelli regularity, and derives the strict inequality by showing equality would force every T_i to be linear, contradicting non-Gaussianity. No fitted parameter is renamed as a prediction; no quantity in the theorem is defined in terms of the contrast it is supposed to justify. Corollary 1 and Corollary 2 follow directly from the theorem's inequality by summing rows of an orthogonal matrix. The synthetic benchmarks use known ground-truth mixing matrices and report Amari errors against them, so the empirical section is not circular. Self-citations (Buchholz et al. 2022, Liang et al. 2023) appear only in the concluding future-work discussion, not in the proof of Theorem 3 or in the derivation of the algorithm. The mismatch between the theorem's smooth strictly positive density assumption and some experimental distributions (Uniform, Beta(0.5,0.5), Laplace, discrete counts) is a real scope/overclaim issue, and the paper explicitly acknowledges failure on discrete sources and the need for Gaussian dithering; but this is an assumption-coverage limitation, not circularity. The proof does not assume the conclusion it is trying to establish, and no external result is replaced by a self-citation chain. Thus the derivation chain is independent and no circular step is present.

Axiom & Free-Parameter Ledger

1 free parameters · 5 axioms · 0 invented entities

The theoretical result is parameter-free; the only hand-chosen number is the dithering variance. The main assumptions are standard ICA identifiability plus smooth positive densities; the latter is silently violated in several experiments.

free parameters (1)
  • Gaussian dithering variance σ_dither = 0.01
    Chosen by hand to smooth discrete empirical CDFs for differentiability (Appendix C.3). It affects the optimized objective and optimization landscape, though not the population-level contrast.
axioms (5)
  • domain assumption At most one source is Gaussian (Comon identifiability)
    Theorem 3 and Corollary 1 assume this to rule out Gaussian mixtures; inherited from Theorem 1.
  • domain assumption Sources are independent, centered, and unit variance
    Used throughout Theorem 3 and in the whitening reduction to the orthogonal group.
  • domain assumption Each source density is smooth and strictly positive
    Required in Appendix B Step 2 for Caffarelli regularity of the transport maps. Not satisfied by Uniform, Laplace, or discrete sources tested in Section 3.
  • standard math Caffarelli regularity of optimal 1D transport maps
    Used to differentiate the optimal maps T_i in the equality analysis; standard for smooth positive densities.
  • standard math Whitening is lossless and reduces the search to O(d)
    Standard preprocessing; holds in the population setting with exact covariance.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Linear Independent Component Analysis via Optimal Transport." pith.science (2026). https://pith.science/paper/SU4BX3DF

@misc{pith2026260714081,
  author       = {Pith},
  title        = {Pith review of: Linear Independent Component Analysis via Optimal Transport},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SU4BX3DF}},
  note         = {Machine review of arXiv:2607.14081}
}
Share X Bluesky LinkedIn Reddit HN
abstract

Linear Independent Component Analysis (ICA) recovers jointly independent source signals from their linear mixtures. To achieve this, classical ICA algorithms attempt to maximize non-Gaussianity, measured by negentropy, which is linked to independence by information theory. Because exact negentropy optimization is intractable, they rely on proxy contrast functions, such as fourth-order cumulants, and parametric log-likelihoods. We propose instead to measure non-Gaussianity using the squared Wasserstein distance $W_2^2$ to a standard Gaussian. We prove that the Wasserstein distance between a standard normal distribution and linear projections of the data is maximized when the projection recovers an independent component. Based on this observation, we propose the OT-ICA algorithm which finds this projection by gradient-based optimization. Empirical evaluation on simulated data shows that OT-ICA outperforms proxy-based methods for different distributions of the latent variables. Application to EEG artifact removal and econometric price discovery confirm OT-ICA can be used for applied ICA tasks without distributional assumptions.

Figures

Figures reproduced from arXiv: 2607.14081 by Ashutosh Jha, Michel Besserve, Simon Buchholz.

Figure 1
Figure 1. Figure 1: OT-ICA convergence on four standard continuous source types. Estimated latent components are smoothed with kernel den￾sity estimation [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Mean Amari error (↓ better, clipped at 1.5) across five mixture regimes and d ∈ {10, 20, 30} (N = 10,000, 10 trials, 95% CI). Dotted line marks E = 0.3, the good-separation threshold. conductor [Hyvärinen et al., 2001]: ocular blink artifacts mix with neural signals before reaching scalp electrodes. Blink artifacts are impulsive and super-Gaussian; their high W2 2 to Gaussian separates them from neural com… view at source ↗
Figure 3
Figure 3. Figure 3: shows OT-ICA concentrating the blink artifact into a single component with excess kurtosis substantially above the remaining four. Zeroing that component and inverting the total unmixing matrix produces the RMS reduction quan￾tified in the panel title, without tuning a density model. 5 CONCLUSION OT-ICA replaces proxy contrast functions with the W2 2 met￾ric computed by quantile sorting; Theorem 3 guarante… view at source ↗
Figure 4
Figure 4. Figure 4: Geometric sketch of optimization on Orthogonal Group O(d). The Euclidean gradient G is projected onto the tangent space to yield the Riemannian gradient ∇O(d)B. A retraction maps the estimate back to the orthogonal surface. 0 1 2 3 Projected Value y = w > x 0.0 0.2 0.4 0.6 0.8 1.0 C u m u l a t i v e P r o b a b i l i t y Fw ( y ) » N(0; ¾ 2 dither) Discrete Empirical CDF (Staircase) Dithered CDF (Continuo… view at source ↗
Figure 5
Figure 5. Figure 5: The non-differentiable staircase of a discrete empirical CDF (gray) is convolved with a narrow Gaussian kernel via dithering, yielding a continuous, differentiable curve (black) [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: W2 2 and logcosh optimization landscapes for two-source Poisson and Binomial mixtures across parameterizations. The W2 2 landscape retains a non-zero signal throughout (confirming the contrast is present), but the step-function geometry of discrete CDFs creates flat plateaus where the gradient evaluates to near-zero, stalling gradient-based solvers. The logcosh landscape degenerates to near-zero for standa… view at source ↗
Figure 7
Figure 7. Figure 7: Amari error comparison for an engineered zero-negentropy distribution across Low and High compute regimes. FastICA’s error scales with dimension as the proxy gradient vanishes, crossing the good separation threshold (E = 0.3). OT-ICA maintains good separation till higher dimensions depending on compute regime. E.1.2 The Vanishing Curvature Condition FastICA utilizes a Newton fixed-point iteration. Defining… view at source ↗
Figure 8
Figure 8. Figure 8: The Vanishing Curvature failure mode. Top: A trimodal distribution engineered such that negative central curvature and positive outer curvature (side peaks at b = ±1.44) cancel (E[sg(s) − g ′ (s)] = 0) while maintaining unit variance. Bottom: Amari error comparison on mixtures containing this distribution. FastICA’s Newton solver diverges rapidly as dimension increases across both compute regimes. OT-ICA m… view at source ↗
Figure 9
Figure 9. Figure 9: Amari error across dimensions for Laplacian sources at fixed N = 10,000 samples. Both OT-ICA and FastICA hit the same curse-of-dimensionality noise ceiling at d = 40, a finite-sample sparsity effect shared by all empirical ICA methods at this regime. While OT-ICA maintains error rates comparable to FastICA (E < 0.3) in lower dimensions, both algorithms exhibit a loss in unmixing quality at d = 40. This lim… view at source ↗
Figure 10
Figure 10. Figure 10: Total execution time (in seconds) for FastICA and OT-ICA across dimensions. G PRICE DISCOVERY APPLICATION: DATA GENERATING PROCESS AND IS DEFINITION G.1 DATA GENERATING PROCESS We simulate a three-market VECM following the non-Gaussian information-share framework of Zema and Cordoni [2025]. The three markets share a single common efficient price, with long-run impact vector ψ = (1, 1, 1)⊤ (each market mov… view at source ↗

discussion (0)

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

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Contrast-Free ICA and Causal Inference via Wasserstein Distances to the Gaussian

    stat.ML 2026-07 conditional novelty 6.0

    Squared 2-Wasserstein distance to the Gaussian identifies ICA unmixing matrices and LiNGAM causal orders via a strict inequality on independent non-Gaussian sources.

  2. Contrast-Free ICA and Causal Inference via Wasserstein Distances to the Gaussian

    stat.ML 2026-07 conditional novelty 6.0

    Maximizing the squared 2-Wasserstein distance of recovered components to the standard Gaussian identifies ICA unmixing matrices and LiNGAM causal orders, with distribution-free consistency.

Reference graph

Works this paper leans on

30 extracted references · cited by 1 Pith paper

  1. [1]

    Neural Computation , volume=

    Independent Component Analysis Using an Extended Infomax Algorithm for Mixed Subgaussian and Supergaussian Sources , author=. Neural Computation , volume=. 1999 , publisher=

    1999

  2. [2]

    Faster independent component analysis by preconditioning with

    Ablin, Pierre and Cardoso, Jean-Fran. Faster independent component analysis by preconditioning with. IEEE Transactions on Signal Processing , volume=. 2018 , publisher=

    2018

  3. [3]

    Proceedings of the 34th International Conference on Machine Learning , pages =

    Martin Arjovsky and Soumith Chintala and L. Proceedings of the 34th International Conference on Machine Learning , pages =. 2017 , editor =

    2017

  4. [4]

    Neural Computation , volume=

    An information-maximization approach to blind separation and blind deconvolution , author=. Neural Computation , volume=. 1995 , publisher=

    1995

  5. [5]

    Blind beamforming for non-

    Cardoso, Jean-Fran. Blind beamforming for non-. 1993 , publisher=

    1993

  6. [6]

    Tests of goodness of fit based on the

    del Barrio, Eustasio and Cuesta-Albertos, Juan A and Matr. Tests of goodness of fit based on the. Annals of Statistics , volume=. 1999 , publisher=

    1999

  7. [7]

    Blind separation of sources, part

    Jutten, Christian and Herault, Jean , journal=. Blind separation of sources, part. 1985 , publisher=

    1985

  8. [8]

    2001 , publisher=

    Independent Component Analysis , author=. 2001 , publisher=

    2001

  9. [9]

    Advances in neural information processing systems , pages=

    A new learning algorithm for blind signal separation , author=. Advances in neural information processing systems , pages=

  10. [10]

    2003 , publisher=

    Topics in Optimal Transportation , author=. 2003 , publisher=

    2003

  11. [11]

    Entropy , volume=

    Independent Component Analysis in the Light of Information Geometry , author=. Entropy , volume=. 2022 , publisher=

    2022

  12. [12]

    Communications on pure and applied mathematics , volume=

    Polar factorization and monotone rearrangement of vector-valued functions , author=. Communications on pure and applied mathematics , volume=. 1991 , publisher=

    1991

  13. [13]

    Journal of the American Mathematical Society , volume=

    The regularity of mappings with a convex potential , author=. Journal of the American Mathematical Society , volume=

  14. [14]

    2008 , publisher=

    Optimization Algorithms on Matrix Manifolds , author=. 2008 , publisher=

    2008

  15. [15]

    IEEE Transactions on Communication Technology , volume=

    Dither signals and their effect on quantization noise , author=. IEEE Transactions on Communication Technology , volume=. 1964 , publisher=

    1964

  16. [16]

    The Journal of Finance , volume=

    One Security, Many Markets: Determining the Contributions to Price Discovery , author=. The Journal of Finance , volume=. 1995 , publisher=

    1995

  17. [17]

    Journal of Financial Markets , volume=

    Price discovery and common factor models , author=. Journal of Financial Markets , volume=. 2002 , publisher=

    2002

  18. [18]

    Econometrica: Journal of the Econometric Society , volume=

    Co-integration and error correction: Representation, estimation, and testing , author=. Econometrica: Journal of the Econometric Society , volume=. 1987 , publisher=

    1987

  19. [19]

    Econometrica: Journal of the Econometric Society , volume=

    Estimation and hypothesis testing of cointegration vectors in Gaussian vector autoregressive models , author=. Econometrica: Journal of the Econometric Society , volume=. 1991 , publisher=

    1991

  20. [20]

    Available at SSRN 5234231 , year=

    A unifying non-Gaussian information share approach to price discovery , author=. Available at SSRN 5234231 , year=

  21. [21]

    Journal of Machine Learning Research , volume=

    A linear non-Gaussian acyclic model for causal discovery , author=. Journal of Machine Learning Research , volume=

  22. [22]

    Advances in Neural Information Processing Systems , volume=

    Function Classes for Identifiable Nonlinear Independent Component Analysis , author=. Advances in Neural Information Processing Systems , volume=

  23. [23]

    Advances in Neural Information Processing Systems , volume=

    Causal Component Analysis , author=. Advances in Neural Information Processing Systems , volume=

  24. [24]

    Monge, Gaspard , journal=. M

  25. [25]

    On the translocation of masses , author=. C. R. (Doklady) Acad. Sci. URSS (N.S.) , volume=

  26. [26]

    Foundations and Trends

    Computational Optimal Transport: With Applications to Data Science , author=. Foundations and Trends. 2019 , publisher=

    2019

  27. [27]

    Signal processing , volume=

    Independent component analysis, a new concept? , author=. Signal processing , volume=. 1994 , publisher=

    1994

  28. [28]

    Probability Theory and Related Fields , volume=

    On the rate of convergence in Wasserstein distance of the empirical measure , author=. Probability Theory and Related Fields , volume=. 2015 , publisher=

    2015

  29. [29]

    Naval Research Logistics Quarterly , volume=

    The Hungarian method for the assignment problem , author=. Naval Research Logistics Quarterly , volume=. 1955 , publisher=

    1955

  30. [30]

    Advances in Neural Information Processing Systems , volume=

    Sinkhorn Distances: Lightspeed Computation of Optimal Transport , author=. Advances in Neural Information Processing Systems , volume=

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.