{"id":"5852b882-ea10-45c6-a2ee-2bbccd4fcf2d","arxiv_id":"2607.12832","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"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.","lead":"This paper shows that a single, tuning-free score based on optimal transport can solve two problems at once: recovering hidden independent sources from linear mixtures, and finding the causal order of variables in linear non-Gaussian systems. It proves the score identifies both the true unmixing and the causal order, and backs it with efficient solvers and experiments.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sample-whitening of the ICA pipeline is not covered by the consistency theorems: theory assumes a fixed whitened P, implementation uses estimated whitening.","rationale":"I read the population-level identifiability proofs carefully. Theorem 11 follows from the strict subadditivity Lemma 10, whose Pexider-equation argument is sound; Theorem 28 likewise follows from Lemma 26 and Proposition 27. Lemma 26 holds because sequential OLS residuals are Gram-Schmidt vectors, hence orthogonal, and standardization yields an orthogonal matrix. I find no internal inconsistency in the oracle claims. The main empirical claim, however, is proved only for i.i.d. draws from the already-whitened distribution P. The actual software whitens with an estimated covariance, which breaks the i.i.d. assumption, and no theorem covers this two-step estimator. This is the most load-bearing gap because the paper's advertised contribution includes empirical consistency for plug-in estimators, and the experiments use the sample-whitened pipeline. The reader's rationale also flagged this gap, though the stated weakest_assumption was about model misspecification, so I mark partial agreement. I do not think the gap warrants rejection—it is likely repairable—so I keep the CONDITIONAL verdict.","tokens_in":31662,"tokens_out":17535,"duration_ms":183289,"concrete_test":"Independently derive the uniform bound for sup_W |pF_n^{whiten}(W)-F(W)| where pF_n^{whiten} uses X_hat_i = Sigma_hat^{-1/2}(Z_i - Z_bar), starting from Theorem 15 and adding a term for ||Sigma_hat^{-1/2} - Sigma^{-1/2}||. If the augmented bound is o_p(1) under Assumption 14, the gap closes; if it requires extra assumptions not stated in the paper, the practical ICA consistency claim remains unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 15 and Corollary 19 prove uniform convergence and consistency for the empirical objective pF_n defined on an i.i.d. sample from a fixed whitened distribution P (Model (1), Assumption 6). The implemented estimator in Section 3.3 and Section 6.1 instead applies sample mean subtraction and an estimated whitening map (e.g., empirical covariance eigendecomposition) to raw observations. The resulting transformed sample is not i.i.d. from P, so the stated uniform-convergence result does not directly apply to the actual plug-in estimator. No theorem, remark, or appendix quantifies the effect of the estimated whitening on sup_W |pF_n(W)-F(W)| or on the recovery error rho(W). The abstract explicitly claims consistency of empirical plug-in estimators, but the only consistency result for ICA is for the oracle-whitened objective. This is the least secure condition in the empirical half of the central claim: without an additional argument—e.g., Lipschitz continuity of the objective in the whitening matrix plus covariance concentration—the consistency of the method as deployed is unproven. The gap is patchable, but it is real and should be stated or closed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":31939,"tokens_out":20498,"duration_ms":204954,"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":[{"comment":"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","section":"§3.3, Theorems 15/19, §5.1 and §6.1"}],"minor_comments":[{"comment":"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.","section":"§3.3 / §5.1"},{"comment":"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'.","section":"§4.2 / Theorem 28"},{"comment":"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.","section":"Figure 4"},{"comment":"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.","section":"§5.2.3 / Theorem 40"},{"comment":"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.","section":"Appendix C.2"}],"recommendation":"major_revision","confidential_remarks":"The core mathematical claims appear sound and the paper is unusually transparent. The single substantive issue is the estimated-whitening gap in the empirical ICA pipeline; I view it as patchable rather than fatal. The disclosure of the simultaneous work by Jha et al. is appropriate and does not affect my assessment."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis paper is worth your time if you work on ICA or LiNGAM. The core idea is simple: use the squared 2-Wasserstein distance to the standard Gaussian as a non-Gaussianity score, and exploit a strict subadditivity property to show it identifies unmixing matrices and causal orders. The population results are the cleanest part. Lemma 10 extends Johnson–Samworth to d dimensions, and Theorems 11 and 28 follow from it without hand-waving. I checked the proofs of the strictness argument and the causal characterization; they are careful. The exact empirical computation via order statistics (Lemma 38) is a nice practical touch, and the experiments come with reproducible code.\n\nWhat's genuinely new relative to Jha et al. (2026) is the empirical side: uniform convergence and consistency of maximizers, plus the causal-order extension with exact and greedy solvers. The paper is transparent about the simultaneous work, which I appreciate.\n\nThe main soft spot is real. The consistency theorems (Theorem 15, Corollary 19, and the causal analogues) are stated for a fixed, already-whitened population P. The implementation and Section 3.3, however, use sample-mean subtraction and an estimated whitening map. The resulting transformed sample is not i.i.d. from P, so the stated uniform convergence does not directly apply to the deployed estimator. The gap is patchable — a Lipschitz-in-whitening argument plus covariance concentration should do it — but it is unproven, and the abstract's claim of consistency for 'empirical plug-in estimators' overstates what the theorems cover. The authors should either close this gap or restate the guarantees for the actual pipeline.\n\nTwo minor points. The greedy solver's exactness guarantee (Theorem 40) rests on the equal-Wasserstein-non-Gaussianity assumption, which is clearly nonstandard; the paper acknowledges this and Remark 41 gives a concrete failure mode, so it is a limitation rather than a flaw. And the causal identification chain inherits the usual LiNGAM assumptions (linearity, acyclicity, no confounding), which the conclusion states plainly.\n\nVerdict: the core mathematics holds up, the contribution over prior work is material, and the flaw in the empirical half is real but not fatal. I would send it to peer review. A good referee should push the authors on the whitening gap before it is ready for publication.\n\nBest,","headline":"A solid, well-written extension of W2-to-Gaussian ICA with a real but patchable gap between the whitening theory and the implementation.","tokens_in":32404,"tokens_out":2702,"would_cite":true,"duration_ms":26776,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H12","62G05","62G20"],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["Independent Component Analysis","Causal Discovery","LiNGAM","Wasserstein distance","Optimal transport","Non-Gaussianity","Distribution-free estimation","Causal order identification"],"falsifier":"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.","tokens_in":31548,"feed_emoji":"📊","tokens_out":4437,"duration_ms":47725,"temperature":0.7,"pith_summary":"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.","feed_headline":"One distance to the Gaussian solves ICA and causal order","feed_subtitle":"Squared 2-Wasserstein non-Gaussianity recovers unmixing and causal order with no tuning parameters.","key_machinery":"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.","core_discovery":"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","pith_inferences":["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."],"forward_implications":["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."],"fun_headline_variants":["One Wasserstein distance cracks ICA and causal order","Squared 2-Wasserstein to Gaussian: new key for ICA and LiNGAM","One distance metric: exact ICA and causal discovery sans tuning","Wasserstein non-Gaussianity: exact unmixing and causal order","No-param contrast: Wasserstein to Gaussian solves ICA and causality"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["One Wasserstein distance cracks ICA and causal order","Squared 2-Wasserstein to Gaussian: new key for ICA and LiNGAM","One distance metric: exact ICA and causal discovery sans tuning","Wasserstein non-Gaussianity: exact unmixing and causal order","No-param contrast: Wasserstein to Gaussian solves ICA and causality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000189,"raw_usage":{"total_tokens":1158,"prompt_tokens":718,"completion_tokens":440,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":462,"completion_tokens_details":{"reasoning_tokens":349}},"tokens_in":462,"tokens_out":440,"duration_ms":4753,"temperature":1.0,"reasoning_tokens":349,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T06:17:21.381543+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[],"review_version":2}