Pith. sign in

REVIEW 3 major objections 5 minor 37 references

High Dimensional Sparse Canonical Correlation Analysis for Elliptical Symmetric Distributions

T0 review · 3 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

Pith's one-line read Spatial-sign covariance keeps sparse CCA consistent under heavy tails.

desk verdict The core idea is reasonable but the main theorem's key error bound assumes r^{-1} is sub-Gaussian, which fails for every distribution simulated, leaving the theoretical guarantee and the experiments mismatched. read the letter →

arxiv 2504.13018 v1 pith:53NM5X6T submitted 2025-04-17 stat.ME

classification stat.ME MSC 62H2062H1262F35
keywords canonicalcorrelationanalysisspatial-signcovarianceellipticaldistributionshigh-dimensionalstatisticsrobustestimationsparsityheavytails
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

This paper claims that a sparse canonical correlation analysis built on the spatial-sign covariance matrix, rather than the sample covariance, remains consistent for elliptical symmetric distributions with heavy tails. Classical CCA requires finite second moments and breaks down in high dimensions; the proposed SSCCA replaces the sample covariance with a sign-based scatter matrix and adds an $\ell^1$ penalty. If the main theorem is right, the leading canonical correlation and the direction vectors are recovered with high probability under a mild sparsity condition, with a convergence rate that differs from prior work only by a $p^{-1/2}$ bias term coming from the sign approximation. The practical upshot, supported by simulations, is that SSCCA matches standard methods under normality and beats them under $t$ or mixture distributions.

What carries the argument

The carrying object is the spatial-sign covariance matrix $S = E[U(X-\mu)U(X-\mu)^\top]$ with $U(x)=x/\|x\|$. For elliptical symmetric $X=\mu+r\Gamma u$, the population covariance obeys $\Sigma \approx \operatorname{tr}(\Sigma) S$ as dimension grows, and because scaling a linear combination does not change correlation, the unknown trace can be ignored; the method solves problem (3) with $p\widehat{S}$ in place of the sample covariance. The proof then chains a uniform error bound for $p\widehat{S}-\Sigma$ through lower and upper bounds on the penalized objective at a normalized truth, producing an approximate maximizer on an $\ell^1$ ball, and converts the canonical-correlation gap assumption into cosine-angle closeness. The $\ell^1$ penalties enforce sparsity while the sign estimator absorbs heavy tails.

What would settle it

Simulate elliptical data with the same covariance structure and check the empirical distribution of $\|p\widehat{S}-\Sigma\|_\infty$ against the Lemma 1 bound; a violation at the claimed probability, or an inconsistent estimate of the leading direction as $n,p$ grow along paths where $\tau_1^2(\sqrt{\log p}/n+p^{-1/2})\to 0$, would refute Theorem 1.

Watch

Extended reading notes

Core claim

The paper's central claim is Theorem 1: for data from an elliptical symmetric distribution, solving the $\ell^1$-penalized problem with $p\widehat{S}$ in place of the sample covariance produces a local maximizer whose cosine-squared angle with the true leading canonical direction tends to 1, and whose estimated canonical correlation approaches $\rho_{1,*}$, provided $\tau_1^2(\sqrt{\log p}/n + p^{-1/2}) \to 0$. In other words, heavy tails need not be transformed away or modeled explicitly: the spatial-sign transformation alone stabilizes the cross-covariance enough for sparse CCA to be consistent. The paper also highlights the price of this robustness: a $p^{-1/2}$ approximation bias appears in the rate because $p\widehat{S}$ approximates $\Sigma$ only as the dimension grows.

Load-bearing premise

The theorem rests on Lemma 1's uniform error bound for $p\widehat{S}-\Sigma$, whose proof is not given in this paper but cited from a companion preprint; if that bound, or the underlying approximation $\Sigma\approx\operatorname{tr}(\Sigma)S$ with its $p^{-1/2}$ bias, fails for the actual data-generating process, the consistency claim collapses.

Editorial extensions

If this is right

  • If Theorem 1 is correct, SSCCA gives a consistency guarantee for sparse CCA without requiring Gaussian or sub-Gaussian tails, so high-dimensional genomics or financial data can be analyzed directly under an elliptical model.
  • The convergence rate matches earlier sparse CCA rates whenever $p\log p/n \to \infty$, so in the usual high-dimensional regime the sign approximation bias disappears and no asymptotic efficiency is lost.
  • Simulations indicate that under heavy-tailed $t$ and mixture-normal distributions, SSCCA has lower estimation error and prediction loss than Kendall-tau and sample-covariance sparse CCA, and it selects sparser and more accurate support.
  • Because the method reuses the same convex optimization algorithm as existing sparse CCA solvers, the robustness improvement can be obtained without new computational machinery.
  • In the nutrimouse application, SSCCA achieves a higher out-of-sample canonical correlation while selecting fewer genes and fatty acids, suggesting a stronger per-variable predictive signal.

Reading between the lines

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

  • A natural extension would replace the $\ell^1$ penalty with group or fused penalties; the spatial-sign formulation is agnostic to the penalty choice, so the same proof template likely carries over when the sparsity assumption is replaced by a group-sparsity norm.
  • The $p^{-1/2}$ bias term implies SSCCA may be systematically less accurate when $p$ is small relative to $n$; a simulation scan with fixed $p$ and increasing $n$ could test where the sign approximation starts to hurt.
  • Because the spatial-sign covariance depends only on directions, the method may tolerate missing or clipped values that preserve sign, a robustness property worth testing explicitly but not claimed in the paper.
  • The consistency statement is for a local maximizer; an editorial guess is that the same proof route could be tightened to a global maximizer under stronger concavity or initialization conditions, though the paper does not address this.
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

3 major / 5 minor

Summary. The paper proposes SSCCA, a sparse CCA method that replaces sample covariance matrices with the spatial-sign covariance matrix p-hat-S under an elliptical symmetric model X = μ + rΓu. The method solves the penalized problem (3) with L1 penalties. The main theoretical result, Theorem 1, asserts that under the error bound of Lemma 1 and sparsity/gap assumptions, a local maximizer of (3) is consistent for the leading canonical correlation and canonical directions, with a rate involving sqrt(log p/n) + p^{-1/2}. Simulations compare SSCCA with KSCCA and SCCA under normal, t_3, and mixture-normal distributions and report superior performance of SSCCA under heavy tails; a nutrimouse data application is also presented.

Significance. If the theoretical claims were fully established, the paper would offer a robust, heavy-tailed-friendly sparse CCA method with a natural use of spatial signs, and the numerical comparisons would be of practical interest. The core idea is reasonable and the appendix proof of Theorem 1 is coherent conditional on Lemma 1. However, the theoretical foundation is incomplete: Lemma 1 is not proved and is cited from an unpublished preprint, and its stated sub-Gaussian condition on r^{-1} appears to exclude all three simulation distributions. The abstract's claim of an 'optimal rate' is also unsupported by any lower-bound or minimax analysis. These gaps are load-bearing for the paper's central consistency claim and for the claimed empirical support.

major comments (3)
  1. [Section 2, Lemma 1 and proof of Theorem 1] The consistency result in Theorem 1 is entirely conditional on Lemma 1, but Lemma 1 is not proved in this paper; the paper only cites Lu and Feng (2025), an unpublished preprint. Because Theorem 1's proof begins 'By Lemma 1', the main theoretical contribution cannot be verified from the manuscript. The paper should either include a complete proof of Lemma 1 or clearly state that Theorem 1 is conditional on an external result and provide the full argument in an accessible form.
  2. [Section 2, Lemma 1, Assumption (ii)] The requirement that r^{-1} is sub-Gaussian is violated by all three distributions simulated in Section 3. Under model (2), the radial density of a standard multivariate normal, a scaled t_3, and the 10/0.8 mixture normal all behave as f_r(ε) ≈ C ε^{p-1} near zero, so P(r^{-1} > t) = P(r < 1/t) ≈ C' t^{-p}, which is a polynomial tail rather than a sub-Gaussian tail. Therefore Lemma 1, and with it Theorem 1, cannot be applied to the normal, t_3, or mixture-normal cases used in the simulations. If the proof of Lemma 1 actually only requires the moment conditions E(|r|^{-k}) ≤ ζ{E(|r|^{-1})}^k for k=2,3,4, the paper should state the weaker condition; as written, the theoretical guarantee and the empirical scope are mismatched, so the claimed robustness under heavy tails is not supported by the theory.
  3. [Abstract and Section 1] The paper claims the estimator converges at an 'optimal rate' but provides no minimax lower bound or any comparison with known optimal rates for sparse CCA. The rate in Theorem 1 contains an additional p^{-1/2} bias term relative to the rates in Mai and Zhang (2019) and Gao et al. (2017), and the condition τ_1^2(√(log p/n) + p^{-1/2}) → 0 is not shown to be minimax optimal. The 'optimal' claim should be removed or replaced by a precise statement about the rate achieved under the stated assumptions.
minor comments (5)
  1. [Appendix, Lemma 2] The inequality direction in Lemma 2 is reversed in both statements: it states max_j cos^2(...) ≤ 1 - 2δ/γ, but the proof shows that each cos^2(...) ≥ 1 - 2δ/γ, and the second bound should also be '≥' with the factor 4κδ/γ. The proof uses the correct direction, so this is a typographical error; please correct the lemma statements.
  2. [Section 2, Lemma 1 statement] The probability bound in Lemma 1 is written as C√(log p + log(α^{-1/2}) / n + C/√p) with ambiguous parentheses; the intended bound appears to be C√({log p + log(1/α)}/n) + C/√p. Please clarify the notation.
  3. [After Theorem 1] The consistency condition is written as τ_1^2 (√(log p)/n + p^{-1/2}) → 0, which is a typo: the first term should be √(log p / n) (or equivalently √(log p)/√n). Please correct the expression.
  4. [Section 3, definitions of FPR and FNR] The definitions of FPR_g and FNR_g appear swapped: the formula for FPR_g is actually the false negative rate among nonzero coefficients, and the formula for FNR_g is the false positive rate among zero coefficients. Please correct the labels or the formulas.
  5. [References] The reference 'Gonz´alez et al. (2008)' is listed twice with identical details; remove the duplicate.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the consistency theorem is a genuine consequence of a stated covariance-approximation lemma, and the theoretical argument does not fit parameters to the target quantities.

full rationale

The paper's central claim is that solving the spatial-sign covariance objective (3) yields directions close to the population CCA directions under Assumptions 1-3 and a covariance approximation error bound. The proof of Theorem 1 (Appendix, after Lemma 2) is a standard optimization argument: it takes the entrywise bound ||p\hat S - Sigma||_inf <= eps from Lemma 1, lower-bounds the penalized objective at normalized true directions, and converts objective-value closeness into angle closeness via Lemma 2 (from Mai and Zhang 2019). Nothing is fitted to the simulated or real outcomes; the tuning lambda is set theoretically as C1*eps*tau1, and the estimated canonical correlation is not an input to the proof. The main external ingredient, Lemma 1, is imported from the same group's Lu and Feng (2025) preprint rather than proved here; this is a missing-proof/completeness concern and a mild load-bearing self-citation, not a circularity, because the lemma is a parameter-free covariance estimation bound whose assumptions do not include the target CCA consistency claim. (A scope issue is that the sub-Gaussian condition on r^{-1} in Lemma 1 appears violated by the normal, t_3, and mixture-normal simulation models; that is a correctness/matching-assumptions concern, not a circularity concern.) The derivation does not reduce to its inputs by construction, so the non-circular verdict is warranted.

Assumptions & free parameters 0 free parameters · 9 assumptions · 0 invented entities

The method introduces no new free parameters or invented entities. The axioms are the standard structural assumptions for high-dimensional sparse CCA plus the elliptical distributional model and the spatial-sign covariance approximation. The most fragile entry is the unproved Lemma 1, which is borrowed from a self-cited preprint and carries the central error bound.

assumptions (9)
  • domain assumption X = mu + r Gamma u with u uniform on the unit sphere and E(r^2)=p (elliptical symmetric distribution)
    Section 2, Eq. (2). The entire method and theory are restricted to this distributional family, which includes the heavy-tailed cases targeted by the paper.
  • domain assumption Sigma approximately tr(Sigma) S, with error tending to zero as the dimension grows
    Section 2, after Eq. (3), citing Lemma 6 in Lu and Feng (2025). This justifies replacing Sigma by p hat S in the CCA objective. It is an asymptotic approximation that depends on bounded eigenvalues of Sigma.
  • domain assumption Eigenvalues of Sigma satisfy kappa^{-1/2} <= lambda_min(Sigma) <= lambda_max(Sigma) <= kappa^{1/2}
    Lemma 1, Assumption (i). This bounded-eigenvalue condition is used to control the concentration of the spatial-sign covariance estimator.
  • domain assumption r^{-1} is sub-Gaussian and satisfies moment bounds E(|r|^{-1}) <= zeta and E(|r|^{-k})/(E(|r|^{-1}))^k <= zeta for k=2,3,4
    Lemma 1, Assumption (ii). This ensures ||X-mu|| is bounded away from zero so that the spatial sign function U(x)=x/||x|| is well behaved.
  • domain assumption The condition number of Sigma_j is bounded: lambda_max(Sigma_j)/lambda_min(Sigma_j) <= kappa for j=1,2
    Assumption 1 in Section 2. Used in the proof of Theorem 1 to translate angle bounds from the transformed space to the original space.
  • domain assumption The gap between the leading and second canonical correlations satisfies rho_1^* - max_{k>=2} rho_k^* > gamma
    Assumption 2 in Section 2. Ensures identifiability of the leading pair of canonical directions, a common requirement in sparse CCA analysis.
  • domain assumption The true canonical vectors satisfy max(||w1,*||_1, ||w2,*||_1) <= tau_1
    Assumption 3 in Section 2. Bounds the L1 norm of the true directions, which is needed for the L1-penalized estimator and the rate in Theorem 1.
  • ad hoc to paper Lemma 1 error bound ||p hat S - Sigma||_inf <= C( sqrt(log p + log(1/alpha))/n + 1/sqrt(p) ) holds
    Section 2, Lemma 1. The consistency theorem depends on this bound, but the proof is not included in the paper and is only attributed to the authors' preprint Lu and Feng (2025).
  • standard math Lemma 2 from Mai and Zhang (2019) relating approximate canonical correlation to the angle between estimated and true directions
    Appendix, Lemma 2. This is a published result from a peer-reviewed paper and is used as a black box in the proof of Theorem 1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of High Dimensional Sparse Canonical Correlation Analysis for Elliptical Symmetric Distributions." pith.science (2026). https://pith.science/paper/53NM5X6T

@misc{pith2026250413018,
  author       = {Pith},
  title        = {Pith review of: High Dimensional Sparse Canonical Correlation Analysis for Elliptical Symmetric Distributions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/53NM5X6T}},
  note         = {Machine review of arXiv:2504.13018}
}
read the original abstract

This paper proposes a robust high-dimensional sparse canonical correlation analysis (CCA) method for investigating linear relationships between two high-dimensional random vectors, focusing on elliptical symmetric distributions. Traditional CCA methods, based on sample covariance matrices, struggle in high-dimensional settings, particularly when data exhibit heavy-tailed distributions. To address this, we introduce the spatial-sign covariance matrix as a robust estimator, combined with a sparsity-inducing penalty to efficiently estimate canonical correlations. Theoretical analysis shows that our method is consistent and robust under mild conditions, converging at an optimal rate even in the presence of heavy tails. Simulation studies demonstrate that our approach outperforms existing sparse CCA methods, particularly under heavy-tailed distributions. A real-world application further confirms the method's robustness and efficiency in practice. Our work provides a novel solution for high-dimensional canonical correlation analysis, offering significant advantages over traditional methods in terms of both stability and performance.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 30 canonical work pages

  1. [1]

    Agniel, D. and T. Cai (2017). Analysis of multiple diverse phenotypes via semiparametric canonical correlation analysis. Biometrics\/ 73 , 1254–1265

  2. [2]

    Croux, and P

    Alfons, A., C. Croux, and P. Filzmoser (2017). Robust maximum association estimators. Journal of the American Statistical Association\/ 112\/ (517), 436--445

  3. [3]

    Branco, J. A., C. Croux, P. Filzmoser, and M. R. Oliveira (2005). Robust canonical correlations: A comparative study. Computational Statistics\/ 20\/ (2), 203--229

  4. [4]

    Chen, M., C. Gao, Z. Ren, and H. Zhou (2017). Sparse cca via precision adjusted iterative thresholding. In Proceedings of International Congress of Chinese Mathematicians 2016

  5. [5]

    Chen, X. and H. Liu (2011). An efficient optimization algorithm for structured sparse cca, with applications to eqtl mapping. Statistical Biosciences\/ 4 , 3–26

  6. [6]

    Chi, E. C., G. I. Allen, H. Zhou, O. Kohannim, K. Lange, and P. M. Thompson (2013). Imaging genetics via sparse canonical correlation analysis. In 2013 IEEE 10th International Symposium on Biomedical Imaging , New York, pp.\ 740–743. Institute of Electrical and Electronics Engineers

  7. [7]

    and M.-L

    Cruz Cano, R. and M.-L. T. Lee (2014). Fast regularized canonical correlation analysis. Computational Statistics & Data Analysis\/ 70 , 88–100

  8. [8]

    Feng, L. (2024). Spatial sign based principal component analysis for high dimensional data. arXiv preprint arXiv:2409.13267\/

Show all 37 references
  1. [9]

    Gao, C., Z. Ma, Z. Ren, and H. H. Zhou (2015). Minimax estimation in sparse canonical correlation analysis. Annals of Statistics\/ 43 , 2168–2197

  2. [10]

    Ma, and H

    Gao, C., Z. Ma, and H. H. Zhou (2017). Sparse cca: Adaptive estimation and computational barriers. Annals of Statistics\/ 45 , 2074–2101

  3. [11]

    D \'e jean, P

    Gonz \'a lez, I., S. D \'e jean, P. G. Martin, and A. Baccini (2008). Cca: An r package to extend canonical correlation analysis. Journal of Statistical Software\/ 23 , 1--14

  4. [12]

    Déjean, P

    González, I., S. Déjean, P. G. Martin, and A. Baccini (2008). Cca: An r package to extend canonical correlation analysis. Journal of Statistical Software\/ 23 , 1–14

  5. [13]

    Guo, Y., X. Ding, C. Liu, and J.-H. Xue (2016). Sufficient canonical correlation analysis. IEEE Transactions on Image Processing\/ 25 , 2610–2619

  6. [14]

    Hardoon, D. R., S. Szedmak, and J. Shawe-Taylor (2004). Canonical correlation analysis: An overview with application to learning methods. Neural Computation\/ 16 , 2639–2664

  7. [15]

    Hotelling, H. (1936). Relations between two sets of variates. Biometrika\/ 28 , 321–377

  8. [16]

    Lopuhaa, H. P. (1989). On the relation between s-estimators and m-estimators of multivariate location and covariance. Annals of Statistics\/ 17\/ (4), 1662--1683

  9. [17]

    Lu, Z. and L. Feng (2025). Robust sparse precision matrix estimation and its application. arXiv preprint arXiv:2503.03575\/

  10. [18]

    Mai, Q. and X. Zhang (2019). An iterative penalized least squares approach to sparse canonical correlation analysis. Biometrics\/ 75 , 734–744

  11. [19]

    Martin, P. G., H. Guillou, F. Lasserre, S. D \'e jean, A. Lan, J.-M. Pascussi, M. SanCristobal, P. Legrand, P. Besse, and T. Pineau (2007). Novel aspects of ppar -mediated regulation of lipid and xenobiotic metabolism revealed through a nutrigenomic study. Hepatology\/ 45\/ (3...

  12. [20]

    Oja, H. (2010). Multivariate nonparametric methods with R: an approach based on spatial signs and ranks . Springer Science & Business Media

  13. [21]

    Ollila, E. and A. Breloy (2022). Regularized tapered sample covariance matrix. IEEE Transactions on Signal Processing\/ 70 , 2306--2320

  14. [22]

    Tritchler, and J

    Parkhomenko, E., D. Tritchler, and J. Beyene (2009). Sparse canonical correlation analysis with application to genomic data integration. Statistical Applications in Genetics and Molecular Biology\/ 8 , 1–34

  15. [23]

    Raninen, E. and E. Ollila (2021). Bias adjusted sign covariance matrix. IEEE Signal Processing Letters\/ 29 , 339--343

  16. [24]

    Raninen, E., D. E. Tyler, and E. Ollila (2021). Linear pooling of sample covariance matrices. IEEE Transactions on Signal Processing\/ 70 , 659--672

  17. [25]

    Tibshirani, and J

    Reid, S., R. Tibshirani, and J. Friedman (2016). A study of error variance estimation in lasso regression. Statistica Sinica\/ 26 , 35–67

  18. [26]

    Rousseeuw, P. J. (1984). Least median of squares regression. Journal of the American Statistical Association\/ 79\/ (388), 871--880

  19. [27]

    Safo, S. E., S. Li, and Q. Long (2018). Integrative analysis of transcriptomic and metabolomic data via sparse canonical correlation analysis with incorporation of biological information. Biometrics\/ 74 , 300–312

  20. [28]

    Tan, K. M., Z. Wang, H. Liu, and T. Zhang (2018). Sparse generalized eigenvalue problem: Optimal statistical rates via truncated rayleigh flow. Journal of the Royal Statistical Society Series B: Statistical Methodology\/ 80\/ (5), 1057--1086

  21. [29]

    Croux, A

    Taskinen, S., C. Croux, A. Kankainen, E. Ollila, and H. Oja (2006). Influence functions and efficiencies of the canonical correlation and vector estimates based on scatter and shape matrices. Journal of Multivariate Analysis\/ 97\/ (2), 359--384

  22. [30]

    Tibshirani, R. J. and J. Taylor (2012). Degrees of freedom in lasso problems. Annals of Statistics\/ 40 , 1198–1232

  23. [31]

    Tyler, D. E. (1987). A distribution-free m-estimator of multivariate scatter. Annals of Statistics\/ 15\/ (1), 234--251

  24. [32]

    Ollila, V

    Visuri, S., E. Ollila, V. Koivunen, J. Möttönen, and H. Oja (2003). Affine equivariant multivariate rank methods. Journal of Statistical Planning and Inference\/ 114\/ (1–2), 161--185

  25. [33]

    Wilms, I. and C. Croux (2015). Sparse canonical correlation analysis from a predictive point of view. Biometrical Journal\/ 57 , 834–851

  26. [34]

    Witten, D. M., R. J. Tibshirani, and T. Hastie (2009). A penalized matrix decomposition, with applications to sparse principal components and canonical correlation analysis. Biostatistics\/ 10 , 515–534

  27. [35]

    Yoon, G., R. J. Carroll, and I. Gaynanova (2020). Sparse semiparametric canonical correlation analysis for data of mixed types. Biometrika\/ 107\/ (3), 609 -- 625

  28. [36]

    Zoh, R. S., B. Mallick, I. Ivanov, V. Baladandayuthapani, G. Manyam, R. S. Chapkin, J. W. Lampe, and R. J. Carroll (2016). Pcan: Probabilistic correlation analysis of two non - normal data sets. Biometrics\/ 72 , 1358–1368

  29. [37]

    Zou, H. and L. Xue (2018). A selective overview of sparse principal component analysis. Proceedings of the IEEE\/ 106\/ (8), 1311--1320

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.