Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Global Convergence of Adaptive Sensing for Principal Eigenvector Estimation

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Adaptive compressed PCA in two measurements per sample provably tracks the top eigenvector, at a rate that exceeds the fully observed minimax by one factor of d.

desk verdict A solid upper-bound analysis buried under unsupported lower-bound claims; fix the abstract and tighten A.1.4 and this is publishable. read the letter →

arxiv 2505.10882 v2 pith:AO4DEP7J submitted 2025-05-16 cs.LG stat.ML

Alex Saad-Falcon , Brighton Ancelin , Justin Romberg This is my paper · ORCID
classification cs.LGstat.ML MSC 62H2568W27
keywords principalcomponentanalysisOja'salgorithmadaptivesensingcompressedmeasurementseigenvectorestimationsubspacetrackingstreamingPCAconvergence
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper analyzes a compressed variant of Oja's algorithm in which each sample is reduced to two scalar measurements: one along the current eigenvector estimate and one in a random orthogonal direction. It proves that this adaptive sensing scheme converges globally for Gaussian data, first taking $O(\lambda_1\lambda_2 d^2/\Delta^2)$ iterations to reach constant alignment and then decaying as $O(\lambda_1\lambda_2 d^2/(\Delta^2 t))$ in expected sine-squared error. The abstract additionally claims a matching information-theoretic lower bound, so the extra factor of $d$ compared with fully observed PCA would be the fundamental cost of compression, with non-adaptive two-measurement schemes one more power of $d$ worse. The result matters wherever acquisition hardware limits each sample to a few inner products, since it provides the first convergence guarantee for adaptive compressed subspace tracking in the noisy case.

What carries the argument

The proof is carried by a scalar recurrence on the squared cosine alignment $c^2 = (\bar u^T u)^2$. Writing $X=1+\eta g^2$ and $Y=\eta gh$ with $g=v^Tu$, $h=v^Tb$, and $z=\bar u^Tb$, the normalized update gives $c_{t+1}^2 = (cX+zY)^2/(X^2+Y^2)$. The analysis lower-bounds this by a self-term and a cross-term, where the cross-term inequality converts the joint Gaussian structure of $(g,h)$ into a deterministic bound involving $\eta$, $\Delta$, $c$, and $z$. Introducing the auxiliary step size $\hat\eta = \Delta\eta/(d-1)$ and the scale $S$ defined above collapses the bound to $E[c_{t+1}^2|c] \ge c^2 + 2\hat\eta c^2(1-c^2) - Sc^2\hat\eta^2$, which is then solved separately in the warmup region $c^2\le 0.5$ and the local region $c^2\ge 0.5$.

What would settle it

Choose any dimension $d$, eigengap $\Delta$, step size $\eta$, and current alignment $c$, then numerically average $2czXY/(X^2+Y^2)$ over many draws of $v\sim N(0,\Sigma)$ and random unit $b$ orthogonal to $u$. If the average is ever below $2\eta\Delta c^2 E[z^2]/(1+3\eta(a^2+b^2))$ with $a^2,b^2$ as defined in Appendix A.1.1, the key cross-term bound is false and the theorem's recurrence does not follow. A second, global check: simulate Algorithm 1 with $\lambda_1=2$, $\lambda_2=1$, $d=10$ and the Theorem 2 step-size schedule; the expected sine-squared error must stay within the stated $O(\lambda_1\lambda_2 d^2/(\Delta^2 t))$ envelope after warmup.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central contribution is an upper-bound theorem for Algorithm 1: starting from a uniformly random unit vector and drawing $v_t \sim N(0,\Sigma)$ with eigengap $\Delta = \lambda_1-\lambda_2>0$, a constant-step-size warmup of $O(S\log d)$ iterations reaches expected sine-squared error at most $0.5$, and a subsequent step-size schedule $\hat\eta_t = 2/(4S+(t-t_0))$ yields $E[1-(\bar u^T u_t)^2] \le O(\lambda_1\lambda_2 d^2/(\Delta^2(t-t_0)))$, where $S = \lambda_1\lambda_2 d^2/\Delta^2 + 13\lambda_1 d/\Delta$. This is the first noisy-setting convergence guarantee for adaptive compressed subspace tracking. The abstract additionally states an information-theoretic lower bound $\Omega(\lambda_1\lambda_2 d^2/(\Delta^2 t))$ and a stronger non-adaptive lower bound $\Omega(\lambda_2^2 d^3/(\Delta^2 t))$; the proof text establishes the upper-bound direction, while the lower-bound statements appear as claims in the abstract and introduction.

Load-bearing premise

The entire convergence proof rests on a single expectation inequality for the cross-term involving the random orthogonal measurement direction; if that inequality does not hold for all parameter choices, the warmup and local-phase recurrences collapse.

Editorial extensions

If this is right

  • Compression costs exactly one power of $d$: the proved rate $O(\lambda_1\lambda_2 d^2/(\Delta^2 t))$ is one dimension factor above the fully observed minimax rate $\Theta(\lambda_1\lambda_2 d/(\Delta^2 t))$.
  • A practical two-measurement protocol suffices: one inner product along the current estimate and one in a random orthogonal direction balance exploitation and exploration, with no full $d$-dimensional sample ever required.
  • The guarantee handles nonzero trailing eigenvalues ($\lambda_2>0$), so it applies to noisy covariance streams rather than only to rank-one noiseless data.
  • If the stated lower bounds are correct, non-adaptive two-measurement compressed PCA is strictly worse by one more power of $d$, separating full observation, adaptive compression, and non-adaptive compression in dimension dependence.
  • With a constant step size the recurrence has a fixed point $x^* = S\hat\eta/2$, so the algorithm can track a slowly drifting leading eigenvector with a predictable steady-state misalignment.

Reading between the lines

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

  • A subgaussian version of the guarantee is not a free corollary: the cross-term bound relies on joint normality of $g$ and $h$ and on Isserlis' theorem, so extending to bounded or subgaussian samples would require a new argument for that step.
  • The fixed-point analysis suggests an explicit tuning rule for tracking, $\hat\eta \approx \sqrt{V/S}$ for motion scale $V$, with steady-state error about $V+\sqrt{VS}$; the paper derives these quantities but does not present them as a tuning prescription.
  • If the advertised lower bounds receive full proofs, a practical consequence follows: adaptive sensing is preferable to any non-adaptive two-measurement scheme at large $d$, since the gap between the two rates grows as $d^3/d^2 = d$ in the convergence denominator.
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 an adaptive compressed variant of Oja's algorithm in which each iteration uses two scalar measurements: one along the current estimate and one in a random orthogonal direction. The authors claim a two-phase convergence guarantee for expected sine-squared error to the leading eigenvector: a warmup phase of O((λ1λ2d^2/Δ^2 + λ1d/Δ) log d) iterations followed by a local convergence phase of O((λ1λ2d^2/Δ^2 + λ1d/Δ)/(t−t0)). The abstract further claims a matching information-theoretic lower bound Ω(λ1λ2d^2/(Δ^2t)) for adaptive compressed PCA and a non-adaptive lower bound Ω(λ2^2d^3/(Δ^2t)). The paper derives the upper bound via a recurrence on the cosine alignment, with the proof split into self-term and cross-term bounds in the appendix. It also includes a short empirical evaluation and an extension to a non-stationary setting with a moving eigenvector.

Significance. If the upper bound were rigorously established, the algorithm and its analysis would be a useful addition to the compressed-PCA literature, providing a concrete adaptive sensing scheme with a polynomial-time guarantee and a proof technique that is elementary. The claimed lower bounds, if proven, would be a significant separation result across fully-observed, adaptive-compressed, and non-adaptive-compressed PCA. However, the manuscript does not contain proofs of either lower bound, and the abstract's optimality claims are contradicted by the paper's own statement that lower bounds are left to future work. The correctness of the upper bound is also in question because the central cross-term bound in the appendix is not justified. The empirical section is illustrative but does not compensate for the missing theory.

major comments (3)
  1. [Abstract and §1] The abstract claims a matching information-theoretic lower bound Ω(λ1λ2d^2/(Δ^2t)) for adaptive compressed PCA and a non-adaptive lower bound Ω(λ2^2d^3/(Δ^2t)), and states that the d^2 factor 'cannot be improved.' The manuscript, however, contains no theorem, lemma, or proof for either lower bound; Section 1 explicitly says 'We leave the development of an information theoretic lower bound for general adaptive measurements to future work.' The claimed optimality and the 'three powers of d' separation are therefore unsupported by any result in the paper. This is a load-bearing overclaim that must be removed or substantiated.
  2. [§4.2 and §A.1.4 (Eq. (2))] The cross-term lower bound, which is essential for the recurrence (2), is not established. The derivation replaces E_b[z·u^TΣb/(1+3η(E[g^2]+E[h^2]))] by E_b[z·u^TΣb]/(1+3η(a^2+b^2)) with a^2 = Δc^2+λ2 and b^2 = Δz^2+λ2, thereby moving a denominator that depends on z outside the expectation without justification. This is not valid because the numerator z·u^TΣb is not pointwise nonnegative. In addition, the step from the indicator-weighted Taylor bounds to the displayed integral inequality is asserted without proof, and the initial reduction of XY/(X^2+Y^2) to ηgh/(1+ηg^2+ηh^2) reverses direction when gh<0. Since both the warmup and local convergence phases use Eq. (2), Theorem 2 is not proven as written.
  3. [Theorem 2 vs. Abstract/Theorem 1] The formal rate in Theorem 2 is governed by S = λ1λ2d^2/Δ^2 + 13λ1d/Δ, giving warmup t0 = (4S+1)log(d/2) and local error O(S/(t−t0)). The abstract and informal Theorem 1 instead state O(λ1λ2d^2/(Δ^2t)). The additional 13λ1d/Δ term is not dominated by λ1λ2d^2/Δ^2 in all parameter regimes (e.g., when λ2d < Δ), so the advertised minimax-style rate does not follow from the formal theorem. The paper should either prove the stronger rate or adjust the stated claims.
minor comments (5)
  1. [§1] The word 'comopressive' should be 'compressive'.
  2. [§2.2] The minimax lower bound is displayed as 'E[1−(¯u^T u_t)^2] ≥ O(λ1λ2d/(Δ^2t))'; using big-O notation in a lower bound is incorrect and should be Ω or a constant multiple.
  3. [§A.1.5] The symbol 'd^2' is used for z^2 = (¯u^T b)^2, colliding with the ambient dimension d; the line '0≤d^2≤1' is confusing and, if read literally as the dimension squared, false. Use a different symbol, e.g., z^2.
  4. [§A.1.4] The display 'E[ηgh/(1+ητ)]' appears to contain a typo; the denominator should be 1+τ, matching the preceding and following expressions.
  5. [§5] The statement that the d^2 dependence is 'likely optimal' is speculative because no lower bound is proven in the paper; this should be phrased as a conjecture or removed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the upper-bound derivation is self-contained and parameter-free, and the asserted matching lower bound is missing rather than circular.

full rationale

The paper's central upper bound is a parameter-free analysis of a fixed algorithm under a stated Gaussian model. The step sizes η0 and ηt are explicit functions of the assumed spectral inputs λ1, λ2, Δ and the dimension d; they are not fitted to the target sine-squared error. The recurrence in Eq. (2) is obtained by algebraic inequalities from conditional expectations, and the final rate E[1−(ū·u_t)^2] ≤ C1/(4S+t−t0)+C2/(4S+t−t0)^2 follows by unrolling that recurrence with K=2 and T=4S. No displayed equation in the proof is equivalent to the conclusion by construction, and no fitted parameter is renamed as a prediction. The adaptive measurement scheme is adopted from [15] and the imputed Oja update from [6], but these citations only define the algorithm under analysis; they do not supply the convergence theorem. The only self-citation to the present authors' prior work, [19], appears in a related-work discussion of array processing applications and is not load-bearing. The abstract's matching lower-bound claim is not a circular step: it is an unsupported assertion, since the formal section contains only the upper bound (Theorem 2), and Section 6 even states 'Future work should explore whether the dependence on the ambient dimension we have derived is optimal. We postulate that with adaptive sensing, a linear dependence in the ambient dimension may be possible.' That contradiction is a correctness/completeness concern, not circularity. Similarly, the questionable treatment of b² outside the z-expectation in Appendices A.1.4–A.1.5 is a technical validity issue, not a reduction of the conclusion to the assumptions. The paper is therefore self-contained against its external benchmark result [24] for the upper-bound direction.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

No fitted constants; the proof is a parameter-free derivation conditional on the Gaussian covariance model. The main uncharged inputs are Gaussianity and known spectral parameters for step sizing.

assumptions (3)
  • domain assumption Samples v_t are i.i.d. Gaussian N(0,Sigma) with fixed covariance Sigma having eigengap Delta > 0.
    The proof's distributional calculations (A.1.1-A.1.4, Isserlis theorem) require Gaussianity; the paper acknowledges general subgaussian is future work.
  • domain assumption The algorithm has access to S = lambda1*lambda2*d^2/Delta^2 + 13*d*lambda1/Delta to set the step size schedule in Theorem 2.
    Warmup and local step sizes are defined in terms of S; without these values the schedule cannot be implemented.
  • standard math Standard Isserlis and Jensen-type inequalities are applied without proof.
    Used in A.1.3 and A.1.4 for joint normal moments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Global Convergence of Adaptive Sensing for Principal Eigenvector Estimation." pith.science (2026). https://pith.science/paper/AO4DEP7J

@misc{pith2026250510882,
  author       = {Pith},
  title        = {Pith review of: Global Convergence of Adaptive Sensing for Principal Eigenvector Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AO4DEP7J}},
  note         = {Machine review of arXiv:2505.10882}
}
abstract

Principal component analysis classically requires full $d$-dimensional samples, yet in various applications hardware limits acquisition to a few scalar measurements per sample. We analyze a compressed variant of Oja's algorithm for estimating the principal eigenvector of the data covariance matrix using only two adaptive measurements per sample. At each iteration, we observe one measurement along the current estimate and one in a random orthogonal direction. We prove that after $t$ iterations, the expected sine-squared error to the true eigenvector is $\mathcal{O}(\lambda_1\lambda_2 d^2 / (\Delta^2 t))$, where $d$ is the ambient dimension, $\lambda_1, \lambda_2$ are the leading eigenvalues, and $\Delta = \lambda_1 - \lambda_2$ is the eigengap. We complement this with a matching information-theoretic lower bound of $\Omega(\lambda_1\lambda_2 d^2 / (\Delta^2 t))$ -- the first for compressed eigenvector estimation -- proving that the $d^2$ factor, an additional factor of $d$ compared to the fully-observed minimax rate $\Theta(\lambda_1\lambda_2 d / (\Delta^2 t))$, is the fundamental cost of compression and cannot be improved. In contrast, any non-adaptive scheme with two measurements per iteration suffers $\Omega(\lambda_2^2 d^3 / (\Delta^2 t))$, an additional power of $d$. This separates fully-observed, adaptive-compressed, and non-adaptive-compressed PCA across three powers of $d$. Our analysis handles the noisy setting where the covariance has nonzero trailing eigenvalues, providing the first convergence guarantee for adaptive compressed subspace tracking beyond the noiseless case.

Figures

Figures reproduced from arXiv: 2505.10882 by the authors.

Figure 1
Figure 1. Empirical results of applying Algorithm 1 compared to the convergence bound in Theorem 2. [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. A comparison between the predicted fixed point (Equation 8) and the empirically observed [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. The Convergence and Error Analysis of Coordinate Descent Methods with Compression for Full Configuration Interaction

    math.NA 2026-07 conditional novelty 6.0 of 10

    Compressed CDFCI converges linearly to a restricted eigenproblem whose eigenvalue error is O(τ²) under spectral-gap and exponential-decay assumptions.

Reference graph

Works this paper leans on

24 extracted references · 22 canonical work pages · cited by 1 Pith paper

  1. [1]

    Simplified neuron model as a principal component analyzer.Journal of Mathematical Biology, 15(3):267–273, November 1982

    Erkki Oja. Simplified neuron model as a principal component analyzer.Journal of Mathematical Biology, 15(3):267–273, November 1982

    work page 1982

  2. [2]

    The Fast Convergence of Incremental PCA

    Akshay Balsubramani, Sanjoy Dasgupta, and Yoav Freund. The Fast Convergence of Incremen- tal PCA, January 2015. arXiv:1501.03796 [cs]

    work page Pith review arXiv 2015

  3. [3]

    First Efficient Convergence for Streaming k-PCA: a Global, Gap-Free, and Near-Optimal Rate

    Zeyuan Allen-Zhu and Yuanzhi Li. First Efficient Convergence for Streaming k-PCA: a Global, Gap-Free, and Near-Optimal Rate, April 2017. arXiv:1607.07837 [cs, math, stat]

    work page Pith review arXiv 2017

  4. [4]

    Streaming k-PCA: Efficient guarantees for Oja’s algorithm, beyond rank-one updates, February 2021

    De Huang, Jonathan Niles-Weed, and Rachel Ward. Streaming k-PCA: Efficient guarantees for Oja’s algorithm, beyond rank-one updates, February 2021. arXiv:2102.03646 [cs, math]

    arXiv 2021

  5. [5]

    Global Convergence of a Grassmannian Gradient Descent Algorithm for Subspace Estimation

    Dejiao Zhang and Laura Balzano. Global Convergence of a Grassmannian Gradient Descent Algorithm for Subspace Estimation, June 2016. arXiv:1506.07405 [cs, math, stat]

    work page Pith review arXiv 2016

  6. [6]

    On the equivalence of oja’s algorithm and grouse

    Laura Balzano. On the equivalence of oja’s algorithm and grouse. In Gustau Camps-Valls, Francisco J. R. Ruiz, and Isabel Valera, editors,Proceedings of The 25th International Confer- ence on Artificial Intelligence and Statistics, volume 151 ofProceedings of Machine Learning Research, pages 7014–7030. PMLR, 28–30 Mar 2022

    work page 2022

  7. [7]

    Compressed Sensing: Applica- tions in Radar and Communications.The Scientific World Journal, 2016:1–2, 2016

    Sandra Costanzo, Álvaro Rocha, and Marco Donald Migliore. Compressed Sensing: Applica- tions in Radar and Communications.The Scientific World Journal, 2016:1–2, 2016

    work page 2016

  8. [8]

    Genomic prediction using subsampling.BMC Bioinformatics, 18(1):191, December 2017

    Alencar Xavier, Shizhong Xu, William Muir, and Katy Martin Rainey. Genomic prediction using subsampling.BMC Bioinformatics, 18(1):191, December 2017

    work page 2017

Show all 24 references
  1. [9]

    Compressed sensing MRI: a review of the clinical literature.The British Journal of Radiology, 88(1056):20150487, December 2015

    Oren N Jaspan, Roman Fleysher, and Michael L Lipton. Compressed sensing MRI: a review of the clinical literature.The British Journal of Radiology, 88(1056):20150487, December 2015

    2015

  2. [10]

    Efficient subsampling for high-dimensional data, November 2024

    Vasilis Chasiotis, Lin Wang, and Dimitris Karlis. Efficient subsampling for high-dimensional data, November 2024. arXiv:2411.06298 [stat]

    2024 arXiv

  3. [11]

    Blind Compressive Sensing Framework for Collabo- rative Filtering, May 2015

    Anupriya Gogna and Angshul Majumdar. Blind Compressive Sensing Framework for Collabo- rative Filtering, May 2015. arXiv:1505.01621 [cs]

    2015 arXiv

  4. [12]

    Laura Balzano, Yuejie Chi, and Yue M. Lu. Streaming PCA and Subspace Tracking: The Missing Data Case, June 2018. arXiv:1806.04609 [cs, math, stat]

    2018 arXiv

  5. [13]

    Convergence of a grassmannian gradient descent algorithm for subspace estimation from undersampled data.CoRR, abs/1610.00199, 2016

    Dejiao Zhang and Laura Balzano. Convergence of a grassmannian gradient descent algorithm for subspace estimation from undersampled data.CoRR, abs/1610.00199, 2016

    2016 arXiv

  6. [14]

    Subspace Learning with Partial Information, May 2016

    Alon Gonen, Dan Rosenbaum, Yonina Eldar, and Shai Shalev-Shwartz. Subspace Learning with Partial Information, May 2016. arXiv:1402.4844 [cs, stat]

    2016 arXiv

  7. [15]

    Enhanced online subspace estimation via adaptive sensing

    Greg Ongie, David Hong, Dejiao Zhang, and Laura Balzano. Enhanced online subspace estimation via adaptive sensing. In2017 51st Asilomar Conference on Signals, Systems, and Computers, pages 993–997, Pacific Grove, CA, October 2017. IEEE

    2017

  8. [16]

    Online Estimation of Coherent Subspaces with Adaptive Sampling

    Greg Ongie, David Hong, Dejiao Zhang, and Laura Balzano. Online Estimation of Coherent Subspaces with Adaptive Sampling. In2018 IEEE Statistical Signal Processing Workshop (SSP), pages 841–845, Freiburg im Breisgau, Germany, June 2018. IEEE

    2018

  9. [17]

    Van Trees.Optimum Array Processing: Part IV of Detection, Estimation, and Modulation Theory

    Harry L. Van Trees.Optimum Array Processing: Part IV of Detection, Estimation, and Modulation Theory. Wiley, 1 edition, March 2002

    2002

  10. [18]

    Independent Component Analysis

    Hannu Oja and Klaus Nordhausen. Independent Component Analysis. In Abdel H. El-Shaarawi and Walter W. Piegorsch, editors,Encyclopedia of Environmetrics. Wiley, 1 edition, August 2012

    2012

  11. [19]

    Subspace Tracking with Dynamical Models on the Grassmannian

    Alex Saad-Falcon, Brighton Ancelin, and Justin Romberg. Subspace Tracking with Dynamical Models on the Grassmannian. In2024 IEEE 13rd Sensor Array and Multichannel Signal Processing Workshop (SAM), pages 1–5, Corvallis, OR, USA, July 2024. IEEE. 10

    2024

  12. [20]

    Convergence of Stochastic Gradient Descent for PCA, January 2016

    Ohad Shamir. Convergence of Stochastic Gradient Descent for PCA, January 2016. arXiv:1509.09002 [cs]

    2016 arXiv

  13. [21]

    Kakade, Praneeth Netrapalli, and Aaron Sidford

    Prateek Jain, Chi Jin, Sham M. Kakade, Praneeth Netrapalli, and Aaron Sidford. Streaming PCA: Matching Matrix Bernstein and Near-Optimal Finite Sample Guarantees for Oja’s Algorithm, March 2016. arXiv:1602.06929 [cs, stat]

    2016 arXiv

  14. [22]

    On the Optimality of the Oja’s Algorithm for Online PCA.Statistics and Computing, 33(3):62, June 2023

    Xin Liang. On the Optimality of the Oja’s Algorithm for Online PCA.Statistics and Computing, 33(3):62, June 2023. arXiv:2104.00512 [cs, math]

    2023 arXiv

  15. [23]

    Global Convergence of Stochastic Gradient Descent for Some Non-convex Matrix Problems, February 2015

    Christopher De Sa, Kunle Olukotun, and Christopher Ré. Global Convergence of Stochastic Gradient Descent for Some Non-convex Matrix Problems, February 2015. arXiv:1411.1134 [cs, math, stat]

    2015 arXiv

  16. [24]

    velocity

    Chris Junchi Li, Mengdi Wang, Han Liu, and Tong Zhang. Near-optimal stochastic approxima- tion for online principal component estimation.Mathematical Programming, 167(1):75–97, January 2018. 11 A Technical Appendices and Supplementary Material A.1 Supporting Lemmas A.1.1 Varia...

    2018

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.