Debiased distributed PCA under high dimensional spiked model

Junpeng Zhu; Siyu Wang; Weiming Li; Yanqing Yin; Zeng Li

REVIEW 4 major objections 5 minor 23 references

Debiased distributed PCA under high dimensional spiked model

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Distributed PCA can be made consistent in high-dimensional spiked models by debiasing each machine's eigenvector estimates before aggregating, and for sparse eigenvectors this consistency holds even with a fixed number of machines.

desk verdict A genuinely new debiased distributed PCA with a real but bounded sparsity claim; worth serious referee time, but the sparse consistency result overstates its scope. read the letter →

arxiv 2505.22015 v1 pith:UDTVNLX6 submitted 2025-05-28 stat.ME

Weiming Li , Zeng Li , Siyu Wang , Yanqing Yin , Junpeng Zhu This is my paper

classification stat.ME MSC 62H2560B20

keywords distributedPCAhigh-dimensionalspikedmodeleigenvectorbiascorrectionrandommatrixtheorysparseeigenvectorsfinitesixthmomentcorrelationmajorityvoting

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

Under a high-dimensional spiked model—a covariance structure with a few separated leading eigenvalues—sample eigenvectors deviate from their population counterparts by a persistent angle, so averaging local PCA results across machines inherits a bias unless the number of machines diverges. The paper proposes to remove that bias locally: each machine rescales its sample eigenvectors by a correction factor derived from its own sample eigenvalues before the estimates are aggregated. The central claim is that this debiased aggregation is consistent as the number of machines tends to infinity under only a finite sixth moment and without symmetric innovations, conditions much weaker than those required by existing distributed PCA methods. When the leading eigenvectors are sparse, the paper further claims consistency for any fixed number of machines, which is the regime where the bias problem is most severe.

What carries the argument

The load-bearing object is the correction factor $\theta_i^{(\ell)}$ of equation (2), a function of the local sample eigenvalues that quantifies the asymptotic shrinkage of sample eigenvector entries relative to population entries; it enters the identity $|\hat u_{ij}/\theta_i|=|u_{ij}|+o_p(1)$, which the algorithm uses to debias each machine. The second mechanism is the decomposition of each local projection matrix into a signal part $(u_i^\top\hat u_i)^2 u_i u_i^\top$ and a residual $R_n$, with the property that residual averages vanish as $m\to\infty$. The third mechanism is the signal-identification step: threshold indicators on the rescaled entries are sent to a central node, which takes a majority vote to decide which coordinates are strong; strong coordinates are then estimated from the debiased local projections, while weak coordinates are filled in from the aggregated matrix and a sparsity criterion decides whether to keep a sparse support. These three pieces together carry the consistency argument.

What would settle it

Simulate a spiked model whose first eigenvector has entries decaying as $u_j \propto j^{-\alpha}$ with small $\alpha$, so that no fixed threshold separates strong from weak entries, and run the algorithm with several thresholds $t$ at large $N$ and small $m$; if the estimated support $\hat A(t)$ continues to misclassify coordinates and the error $\rho$ fails to vanish while an oracle estimator with the true support is consistent, the core separation premise fails.

Watch

Extended reading notes

Core claim

The paper's central discovery is an entrywise bias-correcting relation: under the spiked model, $|\hat u_{ij}/\theta_i|=|u_{ij}|+o_p(1)$, where $\hat u_{ij}$ is a local sample eigenvector entry, $u_{ij}$ is the corresponding population entry, and $\theta_i$ is a correction factor computed from local sample eigenvalues. Dividing by $\theta_i$ before aggregation converts biased local projections into approximately unbiased rank-one estimates of $u_i u_i^\top$. The paper proves that the residual terms in the projection-matrix decomposition cancel in the cross-machine average, yielding consistency as $m\to\infty$ (Theorem 4), that the majority-vote signal-identification step recovers the strong coordinates for any $m$ (Theorem 3), and that for sparse eigenvectors the final estimator is consistent even for fixed $m$ (Theorem 5). The same guarantees are extended to the sample correlation matrix.

Load-bearing premise

Everything rests on the assumption that a single fixed threshold can separate strong from weak eigenvector entries—strong entries uniformly above $t_0$ and weak entries uniformly tending to zero—so that the majority vote across machines classifies coordinates correctly.

Editorial extensions

If this is right

Distributed PCA implementations can drop the symmetric-innovation and sub-Gaussian assumptions; finite sixth moments suffice for consistency of the debiased estimator as $m\to\infty$.
For sparse leading eigenvectors, a fixed number of machines is enough: the estimation error $\rho$ converges to zero as the local sample sizes grow, removing the usual requirement that the machine count diverge.
The correction step uses only each machine's own eigenvalues and eigenvectors and communicates only threshold votes and rescaled projections, so the communication cost stays comparable to existing averaging methods.
The same debiased aggregation applies to the sample correlation matrix, so the method can be used after standardization without new theory.

Reading between the lines

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

Editorial inference: the threshold-separation structure of Assumption 4, not the moment condition, is what makes signal identification work; datasets whose eigenvector entries decay smoothly rather than in two separated groups could violate the majority-vote assumption even when all moment conditions hold.
Editorial inference: the same entrywise identity suggests a natural way to build debiased distributed sparse PCA—use $\hat A(t)$ as an initial support estimate and then run a local refinement step on the strong coordinates, which may improve downstream sparse loading recovery beyond the paper's current estimator.
Editorial inference: because the proof mechanism is a resolvent-plus-contour-integration decomposition of projection matrices, the debiasing strategy should extend to other spectral functionals of the spiked model, such as subspace distances or linear statistics of eigenvectors, not only to the top-$K$ projection.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Two things to know upfront. First, this paper is a serious contender: it fixes a real bias problem in distributed PCA when the number of machines m is small, and the proposed debiasing-and-sparsity-detection scheme is genuinely new, not a repackaging of Fan et al. (2019) or Li et al. (2025). Second, its headline claim about sparse eigenvectors is narrower than it reads. The sparse consistency result (Theorem 5) rests on Assumption 4, which demands a fixed threshold t0 separating strong from weak entries; since the eigenvector has unit norm, the strong support can have at most 1/t0^2 coordinates. So the result covers eigenvectors with O(1) nonzeros, not the growing-support case where entries decay as s^{-1/2}. The simulations only use supports of size 4 and 6, so the practical advantage for genuinely high-dimensional sparse vectors is not demonstrated. That is a scope restriction, not a fatal flaw.

What the paper does well: the local debiasing step using the Yin-Zhou (2024) relation is a natural and plausible fix, and the majority-vote for signal identification is a reasonable way to aggregate local information. The main consistency theorem (Theorem 1) improves on Fan et al. (2019) by requiring only a finite sixth moment and dropping the symmetric-innovation assumption — a real generalization. The simulations agree: the method reduces error substantially when m is small, and it also handles exponential data, which is more than many competitors can claim.

Where I would push back: all proofs are deferred to a supplementary file that is not in the arXiv version, so the central claims cannot be verified from the preprint. The tuning parameters are set by hand (t = 0.1 in simulation, t = 0.005 on real data) with no guidance on how to choose them. No code or standard errors are provided. None of this is disqualifying for a preprint, but it matters for reproducibility.

The debiasing factor comes from the authors' own prior work; that is fine because the relation is cited and not fitted to the current problem. The circularity concern in the reader's report does not land once you read the algorithm.

My bottom line: I would send this to a serious referee. The method is likely correct and fills a real gap, and the theory is testable. The referee should ask for the supplementary proofs, a clear statement of the sparsity regime, and code. Whoever reviews it should push on Theorem 5's interpretation.

Referee Report

4 major / 5 minor

Summary. This paper studies horizontal distributed PCA under a high-dimensional spiked covariance model, where local sample eigenvectors are biased. It proposes a three-stage procedure: local eigendecomposition with a debiasing factor θ; identification of 'strong signal' coordinates by thresholding |\u005cu005cu005cscap |\u005cu005cu005chat u_{ij}/θ| and majority voting across machines; and recovery that combines a debiased rank-one aggregation with sparsity and orthogonality refinements. The main theoretical claims are consistency of the estimated eigenspace as m→∞ under only a finite sixth moment and without symmetric innovations (Theorems 1–4), and consistency for fixed m when the eigenvectors are sparse (Theorem 5); extensions to correlation matrices are also stated. Simulation experiments and a MiniBooNE data analysis report smaller Frobenius errors than the methods of Fan et al. (2019) and Li et al. (2025).

Significance. If the omitted proofs are correct, the paper makes a useful contribution: it removes the symmetric-innovation and sub-Gaussian assumptions for the consistency of distributed eigenspace estimation, provides a debiasing step that visibly improves small-m performance, and gives a clear algorithm with reproducible simulation designs. The paper is also commendable for comparing against two relevant baselines, treating both covariance and correlation matrices, and including a real-data experiment. The main caveat is that the fixed-m sparse result is much narrower than the abstract suggests, and the central proof machinery is not available in the preprint for verification.

major comments (4)

[§4.3 (Theorem 5) and §4.2 (Assumption 4)] Theorem 5, combined with Assumption 4, only covers sparse eigenvectors with bounded support. Because min_{j in A_i} |u_{ij}| > t_0 and ||u_i||=1, the strong support satisfies |A_i| ≤ 1/t_0^2, independent of p. In the standard growing-support sparse regime, the nonzero entries are of order s^{-1/2} -> 0, so no fixed threshold t can separate them from weak entries; the majority vote in Algorithm 1-2 will then misclassify a non-vanishing fraction of the signal, and the sparsity refinement in Algorithm 1-3 will zero out a non-vanishing fraction of the eigenvector norm. The simulations in Section 5 use only fixed supports of size 4 and 6 (plus a dense weak component in the mixed case), so the claimed advantage 'when the leading eigenvectors are sparse' is not demonstrated for growing-support sparsity. The abstract and Theorem 5 should be restricted to the bounded-support case, or the algorithm should be extended to handle support size growing with p.
[§3.2 and §5 (threshold t and criterion (5))] The signal identification step depends on a single user-specified threshold t, and the sparsity refinement uses the hand-set constant 2 m^{-1/4} p^{-1/2} in (5). The paper uses t=0.1 in all simulations and t=0.005 in the real-data experiment, but gives no data-dependent selection rule and no sensitivity analysis. Since Theorems 3–5 require Assumption 4 with t_0>t, the practical validity of the consistency claims is contingent on an unvalidated tuning choice. Please provide a sensitivity study over t and, if possible, an automatic or theoretically guided choice of t and of the constant in (5).
[§4 (Theorems 1–6; Supplementary Material)] All proofs of the main theorems are deferred to the Supplementary Material, which is not included in the submission. The central claims—in particular the finite-sixth-moment consistency of Theorem 1 and the fixed-m sparse consistency of Theorem 5—therefore cannot be verified from the preprint as it stands. The supplementary material must be provided, and at least the key steps (e.g., the decomposition (7) and the control of the averaged residual terms) need to be visible to the reader.
[§4.1 (Theorem 1) and §3.3 (Algorithm 1-3)] There is a gap between the theorem stated in the main text and the estimator whose consistency is claimed. Theorem 1 concerns the average of the raw projection matrices \u005cu005cu005chat u_i^{(\u005cuel)} \u005cu005cu005chat u_i^{(\u005cuel)\u005cutop}, whereas Algorithm 1-3 uses the debiased rank-one matrices [\u005ctheta_i^{(\u005cuel)}]^{-2} \u005cu005cu005chat u_i^{(\u005cuel)} \u005cu005cu005chat u_i^{(\u005cuel)\u005cutop} and then applies thresholding and sparsity refinement. The consistency of \u005ccheck U_K in Theorem 4 does not follow from Theorem 1 alone unless additional arguments (e.g., uniform control of \u005ctheta_i^{(\u005cuel)} and of the threshold indicators) are supplied. Please state explicitly how Theorem 1, or a modified version of it, implies Theorems 4–5, or add the missing intermediate theorem.

minor comments (5)

[§3.3] The sparsity condition is written as u_{i1}^\utop u_{i1} > 1 - 2 m^{-1/4} p^{-1/2} in (5) but as \u005c|u_{i1}\u005c|^2 \u005cgeq 1 - 2/(m^{1/4} p^{1/2}) in Algorithm 1-3; please make the strict/weak inequality and the notation for the threshold consistent.
[§4.3, Theorem 6] The displayed expression contains 'U_KUk', which should be 'U_K U_K^\utop'.
[§2.3 and §4.2] The symbol \u005casymp is used informally in the definition of strong signals; since Assumption 4 builds on this distinction, please define the notation precisely before the assumptions.
[§1 and §4.1] The statement that existing distributed PCA methods require sub-Gaussianity and symmetric innovations is not reconciled with He et al. (2024), which is cited as working 'without moment constraints'; please clarify the exact comparison class of assumptions.
[§6, Figure 6] Figure 6 reports only the average AR over 1000 replications; adding boxplots or standard errors would make the claimed superiority over the baselines more convincing.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the debiasing factor is imported from a prior random-matrix theorem, but the distributed aggregation and sparse-recovery consistency are derived, not fitted to the target.

full rationale

The only self-referential element is the reliance on Yin and Zhou (2024) for the entrywise eigenvector relation (1) and the debiasing factor theta_i in (2). This is a self-citation (Yanqing Yin is a co-author), and it is load-bearing for the bias correction. However, it is used as an external mathematical theorem about spiked-model eigenvectors, with assumptions (spiked eigenvalues, p/n tending to c) that do not include the paper's consistency conclusions; the paper also cites Mestre, Paul, Benaych-Georges and Nadakuditi for the same phenomenon. The aggregation step is proved from a signal/residual decomposition rather than assumed, the signal-identification step is a majority vote with a user-chosen threshold t, and no parameter is fitted to the target eigenspace and then reported as a prediction. The fixed-m sparse consistency in Theorem 5 is genuinely qualified by Assumption 4: because min strong entries exceed a fixed t0 and ||u_i||=1, the strong support has size at most 1/t0^2, so the theorem covers bounded-support sparse vectors rather than growing-support sparsity. This is a scope limitation, not a circularity. Overall the derivation chain is self-contained apart from a standard prior-result citation.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The estimator's guarantees rest on the spiked covariance model, the high-dimensional scaling, a strong/weak signal separation on the true eigenvectors, and the bias formula of Yin and Zhou (2024). The algorithm has two hand-set parameters (t and the constant 2 in the sparsity criterion).

free parameters (2)

threshold t = 0.1 in simulation, 0.005 in real data
User-specified threshold in signal identification (Algorithm 1-2). The theory requires t0 > t, but no data-driven choice is given.
sparsity threshold constant = 2
Constant in the sparsity criterion (5): ||u_i1||^2 >= 1 - 2/(m^{1/4} p^{1/2}). Chosen by hand, no sensitivity analysis.

assumptions (5)

domain assumption Data model x_i = Sigma^{1/2} w_i with i.i.d. w_ij, E w_ij=0, E w_ij^2=1, finite sixth moment (Assumption 1).
Defines the data generating process; all theorems rely on it.
domain assumption High-dimensional scaling p/n_l -> c_l in (0,infty), max c_l < c0, m = O(p) (Assumption 2).
Establishes the asymptotic regime for random matrix results.
domain assumption Spiked population model with well-separated spikes and condition (6) (Assumption 3).
Ensures local sample eigenvectors carry biased but recoverable information about population eigenvectors.
domain assumption Strong/weak signal separation: min_{j in A_i} |u_ij| > t0, max_{j not in A_i} |u_ij| -> 0 (Assumption 4).
Needed for the signal identification step (Algorithm 1-2) to classify coordinates correctly.
standard math Asymptotic relation (1)-(3) from Yin and Zhou (2024): |u_hat_ij / theta_i| = |u_ij| + o_p(1).
Basis for bias correction; quoted as a known theorem.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Debiased distributed PCA under high dimensional spiked model." pith.science (2026). https://pith.science/paper/UDTVNLX6

@misc{pith2026250522015,
  author       = {Pith},
  title        = {Pith review of: Debiased distributed PCA under high dimensional spiked model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UDTVNLX6}},
  note         = {Machine review of arXiv:2505.22015}
}

read the original abstract

We study distributed principal component analysis (PCA) in high-dimensional settings under the spiked model. In such regimes, sample eigenvectors can deviate significantly from population ones, introducing a persistent bias. Existing distributed PCA methods are sensitive to this bias, particularly when the number of machines is small. Their consistency typically relies on the number of machines tending to infinity. We propose a debiased distributed PCA algorithm that corrects the local bias before aggregation and incorporates a sparsity-detection step to adaptively handle sparse and non-sparse eigenvectors. Theoretically, we establish the consistency of our estimator under much weaker conditions compared to existing literature. In particular, our approach does not require symmetric innovations and only assumes a finite sixth moment. Furthermore, our method generally achieves smaller estimation error, especially when the number of machines is small. Empirically, extensive simulations and real data experiments demonstrate that our method consistently outperforms existing distributed PCA approaches. The advantage is especially prominent when the leading eigenvectors are sparse or the number of machines is limited. Our method and theoretical analysis are also applicable to the sample correlation matrix.

Figures

Figures reproduced from arXiv: 2505.22015 by the authors.

**Figure 2.** (a) Box plots of |uˆ ⊤ 1 u1| (m = 1) under the spiked model with K = 1, Σ = diag{2, 1, 1, . . . , 1} and the non-spiked model with Σ = Ip. The orange horizontal line represents asymptotic limit of [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

**Figure 3.** Diagram of the composite recovery region [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Estimation error under the sparse case. 10 20 30 40 50 60 70 80 90 100 m 0.05 0.1 0.15 0.2 (a) zij ∼ N (0, 1). 10 20 30 40 50 60 70 80 90 100 m 0.05 0.1 0.15 0.2 (b) zij ∼ Exp(1) − 1 [PITH_FULL_IMAGE:figures/full_fig_p025_4.png]

**Figure 5.** Figure 5: Estimation error under the mixed case. 6 Real data analysis. In this section, we apply the distributed PCA algorithms to the MiniBooNE dataset which is collected from tabular data benchmark (https://huggingface.co/datasets/ inria-soda/tabular-benchmark). The MiniBooNE …

**Figure 6.** Figure 6: Average AR based on 1000 replications for different number of machines [PITH_FULL_IMAGE:figures/full_fig_p027_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

23 extracted references · 21 canonical work pages

[1]

Anderson, T. W. (1963). Asymptotic theory for principal component analysis.Ann. Math. Statist., 34:122–148

work page 1963
[2]

D., Miao, B

Bai, Z. D., Miao, B. Q., and Pan, G. M. (2007). On asymptotics of eigenvectors of large sample covariance matrix.Ann. Probab., 35(4):1532–1572

work page 2007
[3]

Bai, Z. D. and Yao, J. F. (2008). Central limit theorems for eigenvalues in a spiked population model.Ann. Inst. Henri Poincar´ e Probab. Stat., 44(3):447–474. 27

work page 2008
[4]

Bai, Z. D. and Yao, J. F. (2012). On sample eigenvalues in a generalized spiked population model.J. Multivariate Anal., 106:167–177

work page 2012
[5]

and Nadakuditi, R

Benaych-Georges, F. and Nadakuditi, R. R. (2011). The eigenvalues and eigenvectors of finite, low rank perturbations of large random matrices.Adv. Math., 227(1):494–521

work page 2011
[6]

and Moonen, M

Bertrand, A. and Moonen, M. (2014). Distributed adaptive estimation of covariance ma- trix eigenvectors in wireless sensor networks with application to distributed pca.Signal Processing, 104:120–135

work page 2014
[7]

Cheung, Y.-m., Lou, J., and Yu, F. (2021). Vertical federated principal component analysis on feature-wise distributed data. InWeb Information Systems Engineering–WISE 2021: 22nd International Conference on Web Information Systems Engineering, WISE 2021,

work page 2021
[8]

Springer

Melbourne, VIC, Australia, October 26–29, 2021, Proceedings, Part I 22, pages 173–188. Springer

work page 2021

Show all 23 references

[9]

and Liao, Z

Couillet, R. and Liao, Z. (2022).Random matrix methods for machine learning. Cambridge University Press

2022
[10]

Fan, J., Wang, D., Wang, K., and Zhu, Z. (2019). Distributed estimation of principal eigenspaces.Ann. Statist., 47(6):3009–3031

2019
[11]

He, Y., Liu, Z., and Wang, Y. (2024). Distributed learning for principal eigenspaces without moment constraints.Journal of Computational and Graphical Statistics, pages 1–22

2024
[12]

Johnstone, I. M. (2001). On the distribution of the largest eigenvalue in principal compo- nents analysis.Ann. Statist., 29(2):295–327

2001
[13]

Kargupta, H., Huang, W., Sivakumar, K., and Johnson, E. (2001). Distributed cluster- 28 ing using collective principal component analysis.Knowledge and Information Systems, 3:422–448

2001
[14]

Li, Z., Qin, K., He, Y., Zhou, W., and Zhang, X. (2024). Knowledge transfer across multiple principal component analysis studies.arXiv preprint arXiv:2403.07431

2024 arXiv
[15]

Li, Z., Zhang, X., and Zhou, W. (2025). Two-round distributed principal component analysis: Closing the statistical efficiency gap.arXiv preprint arXiv:2503.03123

2025 arXiv
[16]

Mestre, X. (2008). Improved estimation of eigenvalues and eigenvectors of covariance ma- trices using their sample estimates.IEEE Trans. Inform. Theory, 54(11):5113–5129

2008
[17]

Paul, D. (2007). Asymptotics of sample eigenstructure for a large dimensional spiked covariance model.Statist. Sinica, 17(4):1617–1642

2007
[18]

Qu, Y., Ostrouchov, G., Samatova, N., and Geist, A. (2002). Principal component analysis for dimension reduction in massive distributed data sets.Proceedings of IEEE Interna- tional Conference on Data Mining (ICDM), 1318:1788

2002
[19]

P., Yang, H.-J., Zhu, J., Liu, Y., Stancu, I., and McGregor, G

Roe, B. P., Yang, H.-J., Zhu, J., Liu, Y., Stancu, I., and McGregor, G. (2005). Boosted decision trees as an alternative to artificial neural networks for particle identification. Nuclear Instruments and Methods in Physics Research Section A: Accelerators, Spec- trometers, Det...

2005
[20]

Schizas, I. D. and Aduroja, A. (2015). A distributed framework for dimensionality reduction and denoising.IEEE Transactions on Signal Processing, 63(23):6379–6394

2015
[21]

Shen, S., Lu, J., and Lin, X. (2023). Fadi: Fast distributed principal component analysis with high accuracy for large-scale federated data.arXiv preprint arXiv:2306.06857. 29

2023 arXiv
[22]

and Zhou, W

Yin, Y. and Zhou, W. (2024). High-dimensional pca revisited: Insights from general spiked models and data normalization effects.arXiv preprint arXiv:2408.13848

2024
[23]

Zhang, Z., Zheng, S., Pan, G., and Zhong, P.-S. (2022). Asymptotic independence of spiked eigenvalues and linear spectral statistics for large sample covariance matrices. Ann. Statist., 50(4):2205–2230. 30

2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Anderson, T. W. (1963). Asymptotic theory for principal component analysis.Ann. Math. Statist., 34:122–148

work page 1963

[2] [2]

D., Miao, B

Bai, Z. D., Miao, B. Q., and Pan, G. M. (2007). On asymptotics of eigenvectors of large sample covariance matrix.Ann. Probab., 35(4):1532–1572

work page 2007

[3] [3]

Bai, Z. D. and Yao, J. F. (2008). Central limit theorems for eigenvalues in a spiked population model.Ann. Inst. Henri Poincar´ e Probab. Stat., 44(3):447–474. 27

work page 2008

[4] [4]

Bai, Z. D. and Yao, J. F. (2012). On sample eigenvalues in a generalized spiked population model.J. Multivariate Anal., 106:167–177

work page 2012

[5] [5]

and Nadakuditi, R

Benaych-Georges, F. and Nadakuditi, R. R. (2011). The eigenvalues and eigenvectors of finite, low rank perturbations of large random matrices.Adv. Math., 227(1):494–521

work page 2011

[6] [6]

and Moonen, M

Bertrand, A. and Moonen, M. (2014). Distributed adaptive estimation of covariance ma- trix eigenvectors in wireless sensor networks with application to distributed pca.Signal Processing, 104:120–135

work page 2014

[7] [7]

Cheung, Y.-m., Lou, J., and Yu, F. (2021). Vertical federated principal component analysis on feature-wise distributed data. InWeb Information Systems Engineering–WISE 2021: 22nd International Conference on Web Information Systems Engineering, WISE 2021,

work page 2021

[8] [8]

Springer

Melbourne, VIC, Australia, October 26–29, 2021, Proceedings, Part I 22, pages 173–188. Springer

work page 2021