Pith. sign in

REVIEW 4 major objections 6 minor 16 references

This paper claims that a profiled mutual-information criterion can learn grouped componentwise transformations inside ICA, restoring identifiability and preserving first-order inference when data are nonlinearly scaled.

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 06:24 UTC pith:5BLAZ5RM

load-bearing objection Interesting idea, but the central identifiability theorem is false as stated (identity mixing is a counterexample) and the proofs are all in a missing supplement, so the submission is not ready. the 4 major comments →

arxiv 2607.21901 v1 pith:5BLAZ5RM submitted 2026-07-24 stat.ME

AdaptICA: Data-Adaptive Transformation Learning for Independent Component Analysis

Lida Jalili , Jingyu Liu , Vince D. Calhoun , Li-Hsiang Lin This is my paper
classification stat.ME MSC 62H2562F12
keywords independent component analysisBox-Cox transformationprofiled mutual informationnonlinear distortionidentifiabilityoracle propertyEEG mu-band powerblind source separation
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 proposes AdaptICA, a two-stage method that treats standard linear ICA as a special case. It claims that by profiling out the demixing matrix and selecting grouped Box-Cox transformations to minimize residual mutual information, the true transformation is identifiable and its estimator is root-n consistent and asymptotically normal. It further claims an oracle property: using the estimated transformation in place of the true one does not change the limiting distribution of the final FastICA demixing directions. If correct, this extends ICA to data whose measurement scale has been altered by nonlinear preprocessing, such as band-specific EEG power, without invalidating standard inference. The method includes the identity transformation as a candidate, so it reduces to ordinary ICA when no scale adjustment helps.

Core claim

AdaptICA's central claim is that transformation and demixing can be learned jointly without losing ICA's guarantees. At the population level, the profiled criterion g(λ)=inf_W MI(WT_λ(x)) is minimized uniquely at the true λ0; the paper claims this follows from combining invertibility of the Box-Cox family with classical ICA identifiability (Theorem 1, proof in the supplement). The sample minimizer bλ_n is claimed to be consistent and root-n asymptotically normal with sandwich covariance (Theorem 2), and Stage 2 FastICA applied at bλ_n is claimed to have the same limiting covariance as if λ0 were known (Theorem 3). A stronger almost-sure joint consistency for both transformation and demixing

What carries the argument

The load-bearing object is the profiled mutual-information criterion g(λ)=inf_W MI(WT_λ(x)) and its sample proxy bĝ_n(λ). For each candidate transformation parameter, the data are transformed, optimally demixed, and residual dependence is measured; the transformation is chosen by minimizing that proxy. This profile structure separates transformation selection from the final demixing step and is what permits the oracle argument: bλ_n converges fast enough that Stage 2 behaves as if the true transformation were used.

Load-bearing premise

The theory stands on two premises not fully proven in the submitted text: the sample profiled dependence measure converges uniformly to the population mutual-information criterion over the compact parameter space, and the population criterion is strictly positive for every wrong transformation (Equation 9); the proof of the strict positivity is deferred to an appendix that is not included, and the regularity conditions of that appendix are not stated.

What would settle it

Construct a two-component counterexample: a mixing matrix A0 and a source vector with at most one Gaussian component such that, for some non-identity Box-Cox parameter λ≠λ0, the transformed vector T_λ(A0 s) can be linearly demixed into independent components, making g(λ)=0 and violating Equation 9. Alternatively, inspect the supplement's Appendix B: if the strict positivity in (9) cannot be derived from the stated assumptions, the identifiability result collapses.

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

If this is right

  • If Theorem 1 holds, misspecified transformations leave measurable dependence that no linear demixing can remove, so the transformation parameter is uniquely identified by the independence objective alone.
  • If Theorems 2-3 hold, practitioners can apply AdaptICA and then use standard FastICA asymptotic covariance formulas without adding variance inflation for the preliminary transformation step.
  • Because the identity transformation is in the candidate set, AdaptICA contains standard ICA as a degenerate case and selects it when the original scale is already adequate, as the MEG application shows.
  • In the EEG mu-band power application, the selected regional Box-Cox transforms reduce mutual information among recovered components and yield more spatially focal hand-imagery components.
  • If Theorem 4 holds, downstream plug-in constructions based on the recovered sources inherit almost-sure stability along the entire sample path.

Where Pith is reading between the lines

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

  • The profiled criterion could be reused to select richer transformation families (e.g., monotone splines), since the theory only requires uniform convergence of the sample proxy and strict identifiability of the population minimizer.
  • The oracle property implies that uncertainty in the transformation is negligible at first order, so any observed benefit in the EEG application comes from the transformation itself rather than from extra free parameters; a testable extension would compare second-order inference with and without the estimated transformation.
  • A practical risk the paper does not fully address: using the same dependence measure both to estimate λ and to select the number of groups K may favor increasingly flexible transformations in finite samples; cross-checking with a held-out dependence measure could be turned into a formal selection rule.
  • The theory assumes at most one Gaussian source and a fixed grouping; extending to time-dependent or multi-subject settings would likely require auxiliary identifiability conditions, as the authors note for temporal ICA.

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

4 major / 6 minor

Summary. The paper proposes AdaptICA, a two-stage procedure for linear ICA under componentwise nonlinear distortion. Stage 1 selects grouped Box-Cox transformation parameters by minimizing a profiled empirical dependence criterion after an inner ICA fit; Stage 2 applies FastICA at the selected transformation. The paper claims four theoretical results: identifiability of the transformation parameter (Theorem 1), consistency and asymptotic normality of the Stage 1 estimator (Theorem 2), an oracle property for the Stage 2 demixing estimator (Theorem 3), and joint almost-sure consistency (Theorem 4). Supporting evidence consists of simulations against several ICA baselines and two real-data analyses (motor-imagery EEG and MEG).

Significance. If the theoretical results were correct, AdaptICA would provide a useful extension of linear ICA to nonlinearly scaled data while preserving standard ICA as a special case. The framework is well motivated by the EEG power-feature application, and the inclusion of the identity transformation plus the use of distance correlation as an out-of-criterion check in the EEG analysis are sensible design choices. However, the theoretical contribution is not currently established: all proofs and key regularity conditions are deferred to an unavailable supplement, and the central identifiability theorem is false under the stated assumptions. The paper also does not provide code or reproducible scripts. The idea is promising, but the manuscript in its present form does not support its main claims.

major comments (4)
  1. [Section 3, Theorem 1, Eq. (9)] The claimed strict positivity inf_{lambda != lambda0} g(lambda) > 0 is false under the assumptions of the paper. Model (1) allows A0 = I_T or, more generally, any generalized permutation matrix A0 = PD. In that case A0 s has independent components, and for any lambda the vector T_lambda(x) = T_lambda(T^{-1}_{lambda0}(PDs)) is a componentwise transformation of independent variables, hence has independent components. Taking W = I_T gives MI(WT_lambda(x)) = 0, so g(lambda) = 0 for all lambda in Lambda. Thus lambda0 is not unique and Eq. (9) collapses. This is not a missing proof but a concrete counterexample. A non-permutation / irreducibility condition on A0 (and on the Box-Cox family relative to the source distributions) must be added and proved before Theorems 2-4 can be meaningful.
  2. [Section 3, Theorems 1-4] All four theorems depend on unstated conditions in 'Appendix A of Supplemental Material', and all proofs are deferred to Appendices B-E of a supplement that is not included in the arXiv submission. Theorem 4 further references 'Assumptions??-??', which are not defined anywhere in the submitted text. Consequently, the reader cannot verify the uniform convergence, smoothness, identifiability, or functional-regularity conditions on which the results rest. For a serious statistics journal, the assumptions must appear in the manuscript or the supplement must be included with the submission.
  3. [Section 2, before Eq. (7)] The 'essential requirement' that the empirical criterion ghat_n(lambda) converges uniformly to the population profiled mutual information g(lambda) over Lambda is asserted but never stated as a formal assumption or proven. This is especially delicate because ghat_n(lambda) is defined through an inner ICA fit and an unspecified 'dependence measure' dDep_n; different ICA algorithms and dependence estimators will produce different finite-sample criteria. Uniform convergence of such a data-dependent, non-smooth objective is a nontrivial high-level condition and is load-bearing for the consistency and asymptotic normality in Theorem 2. It must be stated explicitly and justified.
  4. [Section 3, Theorem 3, Eq. (12)] The oracle property claims that replacing lambda0 by the sqrt(n)-consistent estimator lambdahat_n does not change the limiting covariance of the Stage 2 FastICA demixing directions. This requires the FastICA functional to be sufficiently regular (e.g., Hadamard differentiable in the relevant distributional perturbation) and the error from estimation of lambda to be asymptotically negligible uniformly over the neighborhood of lambda0. None of these conditions is stated or proved in the submitted text. Given that the proof is in an unavailable appendix, the claimed oracle property is unsupported.
minor comments (6)
  1. [Throughout] The paper alternates between the names 'AdaptICA' and 'TB-ICA' (e.g., Section 6 and parts of the Introduction). The nomenclature should be unified.
  2. [Throughout] Several unresolved cross-references remain: 'Section??' in Example 2, 'Section??' in Example 3, 'Supplemental Figure??', 'Supplemental Appendix??', and 'Assumptions??-??' in Theorem 4. These must be fixed before submission.
  3. [Section 3, Theorem 2] The matrices H_lambda and V_lambda in Eq. (11) are not defined precisely. In particular, 'the asymptotic covariance of the score associated with the criterion ghat_n(lambda)' is ambiguous because ghat_n is not a standard likelihood or estimating-function objective and depends on the inner ICA fit.
  4. [Section 5.1] The reported 38.8% reduction in mutual information is partly mechanical, since MI is the criterion minimized in Stage 1. The paper does report a 43.5% reduction in distance correlation, which is a better independent check, but the manuscript should clearly state that the MI reduction is expected by construction and emphasize the distance-correlation result.
  5. [Section 4] The tables report mean and standard deviation only. The paper does not disclose how the empirical MI is computed in simulations, which is important because different MI estimators can give substantially different values.
  6. [Section 5.1] The latent dimension r appears in the real-data analysis (br = 10) but is not formally defined in the model or treated theoretically. The paper should state whether r is assumed known and how its data-adaptive selection affects the asymptotic results.

Circularity Check

2 steps flagged

EEG 'MI reduction' is the Stage-1 objective restated; Theorem 1's in-text identifiability argument also assumes the strict-positivity conclusion rather than proving it.

specific steps
  1. fitted input called prediction [Section 2 (Algorithm 1; paragraph after Eq. 8) and Section 5.1 (EEG results)]
    "Because bgn is the same profiled mutual-information proxy used in (7), selecting K requires no additional criterion. ... In practice, we suggest using FastICA in both Stage 1 and Stage 2; under this choice, Stage 2 is in principle computationally redundant because Stage 1 already provides a demixing estimate at the selected λ̂. ... For the selected specification (br, bK) = (10,7), AdaptICA reduced the mutual information among the recovered components from 0.07668 to 0.04696, corresponding to a 38.8% reduction."

    Stage 1's objective bgn is a sample proxy for profiled mutual information, and the same criterion is used to select K (and, in the EEG analysis, the latent dimension r). With FastICA used in both stages, Stage 2 is explicitly 'computationally redundant', so the final components whose MI is reported are the same ones used in Stage 1's minimization. The 38.8% MI reduction is therefore the minimized training criterion restated as an empirical success, not an independent verification. The paper partially mitigates this by separately reporting distance correlation, which is 'not used in estimation or selection', and by physiological checks; those independent criteria weaken the circularity, but the headline MI-specific claim remains by construction.

  2. other [Section 3, paragraph after Theorem 1]
    "The role of invertibility is crucial here: because Tλ is one-to-one on the support, the only way to achieve mutual information zero after demixing is to undo the correct componentwise distortion, up to the usual ICA permutation and scaling."

    This sentence is offered as the justification for Theorem 1's strict-positivity claim, inf_{λ≠λ0} g(λ)>0 in (9). But 'the only way to achieve MI zero is to undo the correct componentwise distortion' is simply a restatement of that identifiability conclusion. Invertibility alone cannot deliver it: an invertible componentwise function of independent coordinates preserves independence, so if A0 is a generalized permutation (e.g., A0=I), g(λ)=0 for every λ and (9) fails. As presented, the uniqueness result is assumed in its own proof sketch; the formal proof is deferred to an appendix that is not included in the submitted text.

full rationale

The AdaptICA derivation is not a self-citation or ansatz-smuggling chain: the cited ICA, FastICA, and Box-Cox literature is standard external material, and the two self-citations in the paper are not load-bearing. The clearest circularity is in the application evaluation: the same profiled MI proxy minimized in Stage 1, and used for group/dimension selection, is reported as the MI reduction in the EEG analysis; with FastICA in both stages, Stage 2 reuses the Stage-1 demixing estimate, so the 38.8% reduction is the fitted objective restated. The distance-correlation and physiological checks are not used in estimation and therefore provide independent support, which keeps the score at 6 rather than higher. Separately, the in-text justification of Theorem 1 ('only way to achieve MI zero...') is a restatement of the strict-positivity conclusion, and the paper's own model (A0 merely invertible) admits identity/generalized-permutation counterexamples; that is a question-begging proof sketch rather than a demonstrated non-circular derivation. Appendices A-B are not included, so the full proofs cannot be audited; their absence is a missing-support flag, which I weight as a correctness risk rather than as additional circularity beyond the steps above.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 0 invented entities

The framework assumes a parametric componentwise transformation model, correct grouping, and unstated regularity conditions that make the empirical criterion converge uniformly. The proofs of the main theorems are deferred to a missing supplement, so the key theoretical premises are not independently checkable in this submission.

free parameters (3)
  • Box-Cox transformation parameters λ_k (k=1..K) = EEG: (0.115, -0.509, -0.771, -1.000, 0.258, 0.873, 0.055); simulations: true values 0.5, (0.5,1.0,1.5), etc.
    Estimated by minimizing the profiled MI criterion (7); central to the claim that AdaptICA recovers the true transformation.
  • Number of transformation groups K = K=7 in EEG; selected by bK = argmin_K bg_n(λ̂(K))
    Data-adaptively selected using the same criterion; affects the transformation model dimension.
  • Number of latent components r = r=10 in EEG
    Selected by the same profiled criterion; determines the demixing dimension.
axioms (6)
  • domain assumption The observed data follow the transformation-based ICA model x = T^{-1}_{λ0}(A0 s) with invertible A0 and mutually independent source components (at most one Gaussian).
    Stated in Section 2, model (1); this is the exact data-generating assumption that the theory relies on.
  • domain assumption The grouped Box-Cox family (2) contains the true transformation T_{λ0}.
    The theory and simulations assume correct specification; Example 3 only partially relaxes this.
  • ad hoc to paper The empirical dependence measure bg_n(λ) converges uniformly to the population profiled MI g(λ) over the compact set Λ, and g is locally quadratic with positive definite Hessian at λ0.
    Stated in Section 2 ('The essential requirement for the theory is that bg_n converges uniformly to g') and Theorem 2; no proof or checkable conditions are provided in the main text.
  • ad hoc to paper Regularity conditions in Appendix A of the Supplemental Material (bounded support, smoothness of densities, etc.) hold.
    Invoked by Theorems 1-4; the appendix is not included in the submission, so the conditions are unverifiable.
  • ad hoc to paper The FastICA functional is sufficiently regular so that plugging in √n-consistent λ̂ does not change the limiting covariance of the demixing estimator.
    Used for Theorem 3 oracle property; proof deferred to Appendix D.
  • domain assumption Anatomical grouping of channels is a known and correct partition.
    The EEG application groups 22 channels into 7 regions; if the grouping is wrong, the transformation model is misspecified.

pith-pipeline@v1.3.0-alltime-deepseek · 15536 in / 13688 out tokens · 129567 ms · 2026-08-01T06:24:36.219807+00:00 · methodology

0 comments
read the original abstract

Independent component analysis (ICA) is widely used to recover latent structure from signal and imaging data, but standard ICA assumes that the observed measurement scale preserves a linear mixing structure. This assumption may fail for features produced through nonlinear preprocessing, such as band-specific power in motor-imagery EEG. We propose AdaptICA, an adaptive transformation-based framework that jointly learns grouped componentwise transformations and the demixing structure using a profiled mutual-information criterion. Because the transformation and demixing parameters may compensate for one another, their joint estimation introduces new identifiability and asymptotic challenges. We establish identifiability, consistency, and asymptotic normality of the transformation estimator, together with joint strong consistency of the transformation and demixing estimators. AdaptICA selects the transformation structure data-adaptively and includes the identity transformation as a candidate, thereby reducing to standard ICA when no scale adjustment is needed. Extensive simulations support the theoretical results. Applications demonstrate that AdaptICA can recover more independent and interpretable sources when transformation is beneficial while retaining standard ICA when the original measurement scale is adequate.

Figures

Figures reproduced from arXiv: 2607.21901 by Jingyu Liu, Lida Jalili, Li-Hsiang Lin, Vince D. Calhoun.

Figure 1
Figure 1. Figure 1: Representative marginal distributions of the cleaned EEG signals on three mea [PITH_FULL_IMAGE:figures/full_fig_p025_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Recover component comparisons of standard ICA and AdaptICA for right-hand, [PITH_FULL_IMAGE:figures/full_fig_p028_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Quantitative comparison of spatial concentration for the components selected for [PITH_FULL_IMAGE:figures/full_fig_p029_3.png] 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

16 extracted references · 6 canonical work pages

  1. [9]

    Aapo Hyv¨ arinen and Hiroshi Morioka

    doi: 10.1109/72.761722. Aapo Hyv¨ arinen and Hiroshi Morioka. Unsupervised feature extraction by time-contrastive learning and nonlinear ICA.Neural Computation, 28(5):1285–1318,

    doi:10.1109/72.761722

  2. [14]

    Ricardo Vig´ ario, Jaakko Sarela, Veikko Jousm¨ aki, Matti H¨ am¨ al¨ ainen, and Erkki Oja

    doi: 10.1109/78.790661. Ricardo Vig´ ario, Jaakko Sarela, Veikko Jousm¨ aki, Matti H¨ am¨ al¨ ainen, and Erkki Oja. In- dependent component approach to the analysis of EEG and MEG recordings.IEEE Transactions on Biomedical Engineering, 47(5):589–593,

    work page doi:10.1109/78.790661

  3. [15]

    Joni Virta, Klaus Nordhausen, and Hannu Oja

    doi: 10.1109/10.841330. Joni Virta, Klaus Nordhausen, and Hannu Oja. Asymptotic properties of the JADE estima- tor.Journal of Multivariate Analysis, 145:168–182,

    work page doi:10.1109/10.841330

  4. [1993]

    Stefan Harmeling, Andreas Ziehe, Motoaki Kawanabe, and Klaus-Robert M¨ uller

    doi: 10.1103/RevModPhys.65.413. Stefan Harmeling, Andreas Ziehe, Motoaki Kawanabe, and Klaus-Robert M¨ uller. Kernel- based nonlinear blind source separation. InAdvances in Neural Information Processing Systems, volume 15, pages 601–608,

    doi:10.1103/revmodphys.65.413

  5. [1994]

    Annette J

    doi: 10.1016/0165-1684(94)90029-9. Annette J. Dobson and Adrian Barnett.An Introduction to Generalized Linear Models. Texts in Statistical Science. Chapman and Hall/CRC, 4 edition,

    doi:10.1016/0165-1684(94)90029-9

  6. [1995]

    Berlin Brain-Computer Interface

    doi: 10.1162/neco.1995.7.6.1129. Berlin Brain-Computer Interface. BCI Competition IV Dataset 2a

    doi:10.1162/neco.1995.7.6.1129 1995

  7. [1999]

    Jean-Fran¸ cois Cardoso

    doi: 10.1162/089976699300016863. Jean-Fran¸ cois Cardoso. Dependence, correlation and gaussianity in independent component analysis.Journal of Machine Learning Research, 4:1177–1203,

    doi:10.1162/089976699300016863

  8. [2000]

    Aapo Hyv¨ arinen, Juha Karhunen, and Erkki Oja.Independent Component Analysis

    doi: 10.1016/S0893-6080(00)00026-5. Aapo Hyv¨ arinen, Juha Karhunen, and Erkki Oja.Independent Component Analysis. John Wiley & Sons,

    doi:10.1016/s0893-6080(00)00026-5

  9. [2001]

    doi: 10.1002/hbm.1048. Vince D. Calhoun, Jingyu Liu, and T¨ ulay Adali. A review of group ica for fmri data and ica for joint inference of imaging, genetic, and erp data. 45(1 Suppl):S163–S172,

    doi:10.1002/hbm.1048

  10. [2006]

    32 Aiyou Chen and Peter J

    doi: 10.1214/009053606000000939. 32 Aiyou Chen and Peter J. Bickel. Consistent independent component analysis and prewhiten- ing.IEEE Transactions on Signal Processing, 56(9):4437–4448,

    work page doi:10.1214/009053606000000939

  11. [2008]

    2008.921144

    doi: 10.1109/TSP. 2008.921144. Pierre Comon. Independent component analysis, a new concept?Signal Processing, 36(3): 287–314,

    arXiv 2008

  12. [2011]

    doi: 10.1214/11-AOS921. Risto J. Ilmoniemi, Matti S. H¨ am¨ al¨ ainen, and Drazenko Kicic. Brain signals: MEG and EEG.Comprehensive Biomedical Physics, 6:1–11,

    work page doi:10.1214/11-aos921

  13. [2012]

    34 Jaakko S¨ arel¨ a and Harri Valpola

    doi: 10.1016/j.sigpro.2012.02.008. 34 Jaakko S¨ arel¨ a and Harri Valpola. Denoising source separation.Journal of Machine Learning Research, 6(9):233–272,

    work page doi:10.1016/j.sigpro.2012.02.008 2012

  14. [2016]

    doi: 10.1016/j.jmva.2015.10.006. 35

    work page doi:10.1016/j.jmva.2015.10.006 2015

  15. [2017]

    MNE Contributors

    doi: 10.1080/01621459.2016.1180989. MNE Contributors. MNE Sample Dataset. Open Science Framework,https://osf.io/ 86qa2/,

    arXiv 2016

  16. [2025]

    Aiyou Chen and Peter J

    URL https://arxiv.org/abs/2511.12732. Aiyou Chen and Peter J. Bickel. Efficient independent component analysis.The Annals of Statistics, 34(6):2824–2855,

    Pith/arXiv arXiv