Pith. sign in

REVIEW 1 major objections 5 minor 44 references

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

T0 review · 1 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read Maximizing the squared 2-Wasserstein distance of recovered components to the standard Gaussian recovers independent sources up to sign and permutation, and identifies the causal order in linear non-Gaussian acyclic models.

desk verdict A solid, well-written extension of W2-to-Gaussian ICA with a real but patchable gap between the whitening theory and the implementation. read the letter →

arxiv 2607.12832 v2 pith:6I7QXJQF submitted 2026-07-14 stat.ML stat.ME

Félix Laplante , Christophe Ambroise , Pierre Humbert This is my paper · ORCID
classification stat.MLstat.ME MSC 62H1262G0562G20
keywords IndependentComponentAnalysisCausalDiscoveryLiNGAMWassersteindistanceOptimaltransportNon-GaussianityDistribution-freeestimationorderidentification
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper establishes that a fixed, tuning-free non-Gaussianity score—the squared 2-Wasserstein distance to the standard Gaussian—solves both linear independent component analysis and causal order recovery in linear non-Gaussian acyclic models. At the population level, maximizers of the score over orthogonal rotations are exactly the true unmixing matrices up to signed permutation, and maximizers over variable orders are exactly the true causal orders. The score needs no density estimation, no choice of contrast nonlinearity, and only finite second moments; empirically it is computed exactly from order statistics and its global maximizers are consistent. This gives a distribution-free, parameter-free alternative to moment-based and information-theoretic ICA and LiNGAM methods.

What carries the argument

The load-bearing inequality is a strict subadditivity property of the squared 2-Wasserstein distance to the standard Gaussian (Lemma 10): any unit-norm linear combination that mixes two or more independent standardized non-Gaussian sources has strictly smaller squared 2-Wasserstein distance to the Gaussian than the weighted sum of the individual source distances. For causal orders, Lemma 26 shows that sequential least-squares residuals under any candidate order form an orthogonal rotation of the standardized independent noises, so the same separation principle transfers from ICA rotations to candidate causal orders.

What would settle it

For two independent standardized non-Gaussian variables S1 and S2, compute the difference W2(S1,N(0,1))^2 + W2(S2,N(0,1))^2 - W2((S1+S2)/√2,N(0,1))^2 at high precision over a grid of distributions; if this difference is ever zero or negative for a non-Gaussian pair, Lemma 10 and hence Theorem 11 would be invalidated.

Watch

Extended reading notes

Core claim

The paper proves that the total squared 2-Wasserstein distance of the recovered components to the standard Gaussian is a valid non-Gaussianity contrast with exact oracle identifiability: its maximizers over the orthogonal group are precisely the true unmixing matrices, up to signed permutation (Theorem 11). The same strict-subadditivity principle, applied to standardized sequential least-squares residuals, characterizes the set of true causal orders of a linear non-Gaussian structural equation model (Theorem 28). The paper also establishes distribution-free uniform convergence of the empirical objective at an n^{-1/4} rate under a finite-moment condition, and gives exact and greedy algorithm

Load-bearing premise

The causal identification chain requires the data to follow an exactly linear, acyclic structural equation model with independent noises and no latent confounding; if any of these fails, the sequential residuals are no longer an orthogonal mixture of independent noises and the separation argument collapses.

Editorial extensions

If this is right

  • The unmixing matrix in linear ICA is identifiable by maximizing a fixed, nonparametric objective with no tuning parameters, under the standard assumption of at most one Gaussian source.
  • Causal orders in linear non-Gaussian acyclic models are identifiable by maximizing the same Wasserstein score over sequential least-squares residuals, with consistency of global maximizers.
  • The empirical objective converges uniformly over the orthogonal group with rate n^{-1/4} up to log factors, under only finite moments of order greater than four.
  • Exact optimization over causal orders is feasible via dynamic programming over subsets, and a greedy variant recovers a causal order in the oracle setting when all noises share the same Wasserstein non-Gaussianity.
  • The one-dimensional Wasserstein distance to the Gaussian can be computed exactly and efficiently from order statistics, requiring only sorting of each recovered component.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The identification argument is specific to exactly linear, acyclic systems with independent noises; extending the same contrast to nonlinear mechanisms or latent confounders would require a different residual representation and likely new separation results.
  • The strict subadditivity principle may be reusable with reference distributions other than the Gaussian, but the exact form of Lemma 10 would need to be re-derived for each reference.
  • The separation margin Γ(δ) between true and false solutions could serve as a data-dependent confidence measure for the recovered unmixing or causal order, but the paper does not explore this operational use.
  • The greedy order-search variant is expected to be sensitive to heterogeneous non-Gaussianity across noises, as the paper demonstrates in a two-variable counterexample; this suggests the exhaustive objective is preferable when non-Gaussianity levels differ.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

1 major / 5 minor

Summary. The paper proposes using the squared 2-Wasserstein distance to the standard Gaussian as a non-Gaussianity measure for linear ICA and for causal order recovery in linear non-Gaussian acyclic models (LiNGAM). The main population-level results are exact identifiability of the ICA unmixing matrix up to signed permutation (Theorem 11) and of the set of causal orders (Theorem 28), both derived from a strict subadditivity property (Lemma 10) whose proof proceeds via Brenier optimal transport maps and a Pexider equation. The empirical side defines plug-in estimators, states uniform convergence and consistency results for global maximizers of the empirical ICA objective (Theorem 15, Corollary 19) and of the causal order objective (Theorems 30 and 36, Corollary 32), and develops a Picard-style L-BFGS solver, an exhaustive dynamic program, and a greedy order search. Experiments compare favorably with FastICA, DirectLiNGAM, ICA-LiNGAM, and DAGMA, and reproducible open-source implementations are provided.

Significance. If the results hold, the paper makes a substantial contribution: it shows that a fixed, parameter-free optimal-transport contrast yields exact identification of ICA and causal orders under only finite second moments plus the standard non-Gaussianity assumption. The strict-subadditivity proof is elegant and appears sound, and the order-statistics plug-in gives an exact and cheaply computable empirical criterion. The causal extension through sequential residualization is nontrivial and gives exact order recovery, not merely an ICA-based heuristic. The paper is also transparent: detailed proofs are included, the local nature of the ICA solver is acknowledged, a closely related simultaneous preprint is disclosed, and reproducible code is provided. The main weakness is that the consistency theorems are stated for a fixed whitened distribution, whereas the implemented pipeline uses estimated whitening; this gap is real but appears patchable.

major comments (1)
  1. [§3.3, Theorems 15/19, §5.1 and §6.1] The consistency theory is developed for the objective pF_n evaluated on an i.i.d. sample from a fixed whitened distribution P (Model (1), Assumption 6). Theorem 15 bounds E[sup_W |pF_n(W)-F(W)|] for this oracle-whitened sample, and Corollary 19 converts that bound into consistency of global maximizers. The implemented pipeline (Algorithm 1; §6.1) instead subtracts the sample mean and applies an estimated whitening map to raw observations before computing pF_n. The transformed sample is not i.i.d. from P, and no result in the paper bounds sup_W |pF_n(W)-F(W)| or ρ(W_n) for this plug-in estimator. Because Corollary 19's conclusion is the basis for the abstract's consistency claim, the empirical half of the central claim is not covered as stated. This is patchable, for example by proving Lipschitz continuity of the objective in the whitening matrix and combining it with covariance concentra
minor comments (5)
  1. [§3.3 / §5.1] The notation should distinguish the raw sample from the whitened sample. Define \hat X_i = \hat\Sigma^{-1/2}(X_i-\bar X) and state that Theorems 15/19 assume known whitening; this would make the gap explicit.
  2. [§4.2 / Theorem 28] The identification result for causal orders relies on exact linearity, acyclicity, independent noises, and no latent confounding (Model (3), Assumption 21). The conclusion states this, but a short reminder near Theorem 28 would prevent overreading of the term 'causal inference'.
  3. [Figure 4] The y-axes include negative values although disorder is nonnegative; the error bars are the source. Consider one-sided error bars or adjusting the axis range to avoid visual artifacts.
  4. [§5.2.3 / Theorem 40] The greedy guarantee requires the extra equal-Wasserstein-non-Gaussianity assumption (Eq. (5)). This is clearly stated, but the main text could emphasize that the exhaustive objective does not need this assumption.
  5. [Appendix C.2] The Pexider step in Lemma 10 is terse: after fixing z0, the reduced equation holds almost everywhere on R^2, and the cited affine-representation results are for a.e. equations. A one-sentence justification that the standard results apply to this a.e. version would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the identification results are derived from optimal-transport inequalities and are not reduced to fitted inputs or self-citations.

full rationale

The paper's central derivation chain is self-contained. The population ICA identification (Theorem 11) follows from the strict subadditivity lemma (Lemma 10), which is proven from Brenier's theorem and a Pexider-equation argument rather than assumed; the two-variable precedent in Johnson and Samworth (2005) is a reference for completeness, with a full proof supplied. No parameter is fitted and then renamed as a prediction: the objective F(W) is defined directly as a sum of squared 2-Wasserstein distances to the standard Gaussian, and its maximizers are proven to be exactly the signed-permuted unmixing matrices under Assumptions 6 and 7. The causal-order identification (Theorem 28) likewise follows from the Gram-Schmidt structure of sequential least-squares residuals (Lemma 26), the same strictness lemma, and the exogeneity characterization (Lemma 39), all within the stated linear acyclic model. The empirical uniform-convergence results (Theorem 15, Corollary 19) are proven from an external max-sliced Wasserstein concentration bound (Han et al. 2024) and the explicit order-statistics computation (Lemma 38), with no fitted parameters entering the argument. The greedy causal-order result (Theorem 40) is conditional on an explicit equal-non-Gaussianity assumption and is accompanied by a demonstrated failure mode under heterogeneous noise (Remark 41), so it is not a hidden self-justifying premise. Self-citations are limited to the authors' software packages and a transparency note about the independent simultaneous work of Jha et al. (2026); none carries a load-bearing mathematical claim. The one substantive gap noted by a skeptical reader—that the implemented estimator uses an estimated whitening map while the consistency theorem is stated for a fixed whitened population distribution—is a real correctness/consistency concern about applying Theorem 15 to the full plug-in pipeline, but it is not circular: no equation identifies the estimated-whitening objective with the oracle-whitened objective by construction, and no fitted parameter is being passed off as a prediction. The paper also explicitly acknowledges its modeling limitations (linearity, no hidden confounding, local optimizers). Accordingly, no circular step can be exhibited, and the appropriate score is 0.

Assumptions & free parameters 0 free parameters · 10 assumptions · 0 invented entities

The central claims rest on standard optimal-transport theorems, standard distributional assumptions for ICA/LiNGAM, and one ad hoc equal-non-Gaussianity assumption for the greedy variant. No parameters are fitted to data and no new entities are postulated.

assumptions (10)
  • standard math Brenier's theorem on existence and uniqueness of quadratic optimal transport maps (Theorem 5).
    Used in Lemma 10 to identify the optimal coupling and derive the strict inequality from equality conditions.
  • standard math Measurable solutions of the Pexider equation are affine (De Bruijn, Aczel-Dhombres, Reem).
    Used in Lemma 10 to force two transport maps to be affine, making the source distributions Gaussian.
  • standard math Max-sliced Wasserstein concentration bound of Han et al. (2024, Theorem 3.5).
    External concentration result used in Theorem 15 for uniform convergence of the empirical objective.
  • standard math Sub-Gaussian covariance concentration theorem of Vershynin (2026).
    Used in Theorem 36 to quantify the error between empirical and oracle least-squares residuals.
  • domain assumption Assumptions 6 and 7: sources are independent, centered, variance-one, and at most one is Gaussian.
    Standard ICA identifiability conditions; required for Theorems 11 and 19.
  • domain assumption Assumptions 21 and 22: structural noises are independent, have positive variance, and at most one is Gaussian.
    Standard LiNGAM identifiability conditions; required for Theorems 28, 30, and 36.
  • domain assumption Assumption 14: finite p-th moment for some p > 4.
    Used in Theorem 15 to apply the max-sliced Wasserstein concentration bound.
  • domain assumption Assumption 34: whitened observations are sub-Gaussian.
    Used in Theorem 36 to obtain a quantitative convergence rate for causal order estimation.
  • ad hoc to paper Equation (5): all structural noises have the same W2 non-Gaussianity score c.
    An additional equal-non-Gaussianity condition introduced only to prove correctness of the greedy algorithm (Theorem 40); not part of standard LiNGAM assumptions.
  • domain assumption Model (3): X = B* X + eps with B* acyclic, linear, and no latent confounders.
    The causal identification argument relies on residuals being linear functions of independent noises; any hidden confounder or nonlinearity breaks the orthogonal-representation lemma.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Contrast-Free ICA and Causal Inference via Wasserstein Distances to the Gaussian." pith.science (2026). https://pith.science/paper/6I7QXJQF

@misc{pith2026260712832,
  author       = {Pith},
  title        = {Pith review of: Contrast-Free ICA and Causal Inference via Wasserstein Distances to the Gaussian},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6I7QXJQF}},
  note         = {Machine review of arXiv:2607.12832}
}
abstract

We study the squared $2$-Wasserstein distance to the standard Gaussian as a non-Gaussianity criterion and use it for linear Independent Component Analysis (ICA) and causal discovery in Linear Non-Gaussian Acyclic Models (LiNGAM). Unlike commonly used parametric contrasts and approximations of information-theoretic quantities, this criterion requires no distributional regularity beyond finite second moments, involves neither approximation nor tuning parameters, and can be computed exactly and efficiently from empirical order statistics. Our analysis relies on a strict subadditivity property of the $2$-Wasserstein distance to the Gaussian. At the population level, we prove exact identification of the ICA unmixing matrix, up to signed permutation, and give an analogous characterization of causal orders through sequential least-squares residuals. We then define empirical plug-in estimators and prove distribution-free uniform convergence under finite-moment assumptions, before detailing three practical solvers: a Picard-style orthogonal optimizer for ICA, an exhaustive dynamic program for causal order search, and a greedy order search variant. Empirically, we demonstrate competitive performance for both tasks and provide open-source implementations for source separation and causal discovery.

Figures

Figures reproduced from arXiv: 2607.12832 by the authors.

Figure 1
Figure 1. ICA recovery for Laplace, uniform, centered-exponential, and calibrated uniform– [PITH_FULL_IMAGE:figures/full_fig_p021_1.png] view at source ↗
Figure 2
Figure 2. ICA recovery as each source is replaced by [PITH_FULL_IMAGE:figures/full_fig_p021_2.png] view at source ↗
Figure 3
Figure 3. ICA fitting time with variance-one uniform sources. Left: varying sample size [PITH_FULL_IMAGE:figures/full_fig_p022_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Causal-order recovery on ER2, ER4, SF2, and SF4 graphs. Top: varying sample [PITH_FULL_IMAGE:figures/full_fig_p023_4.png]
Figure 5
Figure 5. Figure 5: Causal-order recovery under increasingly Gaussian structural noise on ER2, ER4, [PITH_FULL_IMAGE:figures/full_fig_p024_5.png]
Figure 6
Figure 6. Figure 6: Performance on ERk graphs as density varies, with n “ 1,000 and d “ 8. Left: mean disorder. Center: mean SHD. Right: mean directed-edge F1 score. Points show means over 20 runs, and capped error bars show one standard deviation. 250 500 1000 2000 4000 n (sample size), …
Figure 7
Figure 7. Figure 7: Runtime on ER2 graphs. Left: varying sample size with [PITH_FULL_IMAGE:figures/full_fig_p025_7.png]
Figure 8
Figure 8. Figure 8: Wasserstein and log-cosh objectives along rotations of a true unmixing direc [PITH_FULL_IMAGE:figures/full_fig_p030_8.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 3 linked inside Pith

  1. [1]

    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

  2. [2]

    Colloquium Mathematicum , volume=

    On almost additive functions , author=. Colloquium Mathematicum , volume=. 1966 , organization=

    1966

  3. [3]

    Natural image statistics: A probabilistic approach to early computational vision , pages=

    Independent component analysis , author=. Natural image statistics: A probabilistic approach to early computational vision , pages=. 2001 , publisher=

    2001

  4. [4]

    Max-sliced Wasserstein concentration and uniform ratio bounds of empirical measures on

    Han, Ruiyu and Rush, Cynthia and Wiesel, Johannes , journal=. Max-sliced Wasserstein concentration and uniform ratio bounds of empirical measures on

  5. [5]

    Annual review of statistics and its application , volume=

    Statistical aspects of Wasserstein distances , author=. Annual review of statistics and its application , volume=. 2019 , publisher=

    2019

  6. [6]

    International conference on algorithmic learning theory , pages=

    Measuring statistical dependence with Hilbert-Schmidt norms , author=. International conference on algorithmic learning theory , pages=. 2005 , organization=

    2005

  7. [7]

    Mathematical programming , volume=

    On the limited memory BFGS method for large scale optimization , author=. Mathematical programming , volume=. 1989 , publisher=

    1989

  8. [8]

    Journal of artificial intelligence research , volume=

    Searching for Bayesian network structures in the space of restricted acyclic partially directed graphs , author=. Journal of artificial intelligence research , volume=

Show all 44 references
  1. [9]

    Advances in Neural Information Processing Systems , volume=

    Beware of the simulated dag! causal discovery benchmarks may be easy to game , author=. Advances in Neural Information Processing Systems , volume=

  2. [10]

    Bernoulli , volume=

    Central limit theorem and convergence to stable laws in Mallows distance , author=. Bernoulli , volume=. 2005 , publisher=

    2005

  3. [11]

    Reviews of modern physics , volume=

    Statistical mechanics of complex networks , author=. Reviews of modern physics , volume=. 2002 , publisher=

    2002

  4. [12]

    Transactions of the American Mathematical Society , year=

    On the evolution of random graphs , author=. Transactions of the American Mathematical Society , year=

  5. [13]

    European Conference on Machine Learning , pages=

    Independent subspace analysis on innovations , author=. European Conference on Machine Learning , pages=. 2005 , organization=

    2005

  6. [14]

    IEEE transactions on Neural Networks , volume=

    Fast and robust fixed-point algorithms for independent component analysis , author=. IEEE transactions on Neural Networks , volume=. 1999 , publisher=

    1999

  7. [15]

    IEE Proceedings F (Radar and Signal Processing) , volume=

    Blind beamforming for non-Gaussian signals , author=. IEE Proceedings F (Radar and Signal Processing) , volume=. 1993 , publisher=

    1993

  8. [16]

    Proceedings of the 35th International Conference on Machine Learning , volume=

    Mutual information neural estimation , author=. Proceedings of the 35th International Conference on Machine Learning , volume=. 2018 , publisher=

    2018

  9. [17]

    arXiv preprint arXiv:2607.14081 , year=

    Linear independent component analysis via optimal transport , author=. arXiv preprint arXiv:2607.14081 , year=

    arXiv

  10. [18]

    Scikit-learn: Machine learning in

    Pedregosa, Fabian and Varoquaux, Ga. Scikit-learn: Machine learning in. Journal of Machine Learning Research , volume=

  11. [19]

    Neural computation , volume=

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

    1995

  12. [20]

    Signal processing , volume=

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

    1994

  13. [21]

    1989 , publisher=

    Functional equations in several variables , author=. 1989 , publisher=

    1989

  14. [22]

    Neural Networks , volume=

    Independent component analysis: algorithms and applications , author=. Neural Networks , volume=. 2000 , publisher=

    2000

  15. [23]

    Journal of Artificial Intelligence Research , volume=

    Learning optimal Bayesian networks: A shortest path perspective , author=. Journal of Artificial Intelligence Research , volume=

  16. [24]

    Biometrika , volume=

    On causal discovery with an equal-variance assumption , author=. Biometrika , volume=. 2019 , publisher=

    2019

  17. [25]

    IEEE Transactions on Signal Processing , volume=

    Faster independent component analysis by preconditioning with Hessian approximations , author=. IEEE Transactions on Signal Processing , volume=. 2018 , publisher=

    2018

  18. [26]

    Journal of Machine Learning Research , volume=

    Kernel independent component analysis , author=. Journal of Machine Learning Research , volume=

  19. [27]

    Aequationes mathematicae , volume=

    Remarks on the Cauchy functional equation and variations of it , author=. Aequationes mathematicae , volume=. 2017 , publisher=

    2017

  20. [28]

    Journal of the American Statistical Association , volume=

    Independent component analysis via distance covariance , author=. Journal of the American Statistical Association , volume=. 2017 , publisher=

    2017

  21. [29]

    arXiv preprint arXiv:2211.15639 , year=

    Distribution-free joint independence testing and robust independent component analysis using optimal transport , author=. arXiv preprint arXiv:2211.15639 , year=

    arXiv

  22. [30]

    arXiv preprint arXiv:2302.12693 , year=

    Wasserstein projection pursuit of non-Gaussian signals , author=. arXiv preprint arXiv:2302.12693 , year=

    arXiv

  23. [31]

    Journal of Machine Learning Research , volume=

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

  24. [32]

    Journal of Machine Learning Research , volume=

    Shimizu, Shohei and Inazumi, Takanori and Sogawa, Yasuhiro and Hyv. Journal of Machine Learning Research , volume=

  25. [33]

    Journal of the American statistical association , volume=

    The adaptive lasso and its oracle properties , author=. Journal of the American statistical association , volume=. 2006 , publisher=

    2006

  26. [34]

    Proceedings of the Twenty-Second Conference on Uncertainty in Artificial Intelligence , pages=

    A simple approach for finding the globally optimal Bayesian network structure , author=. Proceedings of the Twenty-Second Conference on Uncertainty in Artificial Intelligence , pages=

  27. [35]

    Journal of Machine Learning Research , volume=

    High-dimensional learning of linear causal networks via inverse covariance estimation , author=. Journal of Machine Learning Research , volume=

  28. [36]

    Biometrika , volume=

    Identifiability of Gaussian structural equation models with equal error variances , author=. Biometrika , volume=. 2014 , publisher=

    2014

  29. [37]

    and Xing, Eric P

    Zheng, Xun and Aragam, Bryon and Ravikumar, Pradeep K. and Xing, Eric P. , booktitle=

  30. [38]

    Bello, Kevin and Aragam, Bryon and Ravikumar, Pradeep , booktitle=

  31. [39]

    2009 , publisher=

    Optimal Transport: Old and New , author=. 2009 , publisher=

    2009

  32. [40]

    2026 , publisher=

    High-Dimensional Probability: An Introduction with Applications in Data Science , author=. 2026 , publisher=

    2026

  33. [41]

    Journal of Machine Learning Research , volume=

    Identifiability of additive noise models using conditional variances , author=. Journal of Machine Learning Research , volume=

  34. [42]

    Foundations and Trends in Machine Learning , volume=

    Computational optimal transport , author=. Foundations and Trends in Machine Learning , volume=

  35. [43]

    Science , volume=

    Causal protein-signaling networks derived from multiparameter single-cell data , author=. Science , volume=. 2005 , publisher=

    2005

  36. [44]

    Signal processing , volume=

    Blind separation of sources, part I: An adaptive algorithm based on neuromimetic architecture , author=. Signal processing , volume=. 1991 , publisher=

    1991

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.