pith. sign in

arxiv: 2605.17111 · v1 · pith:HGYWUMHDnew · submitted 2026-05-16 · 📊 stat.ME · cs.IT· eess.SP· math.IT

Symmetry-Aware Convex Shrinkage for High-Dimensional Covariance Estimation

Mitchell A. Thornton This is my paper

Pith reviewed 2026-05-20 14:50 UTC · model grok-4.3

classification 📊 stat.ME cs.ITeess.SPmath.IT
keywords high-dimensional covariance estimationshrinkage estimatorssymmetry groupsLedoit-Wolf shrinkagedata-adaptive selectionReynolds projectionheld-out validationfinite symmetry groups
0
0 comments X

The pith

Selecting a symmetry group from held-out data and projecting the sample covariance onto it produces a shrinkage target that dominates Ledoit-Wolf under a sufficient-match condition.

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

The paper introduces shrinkage estimators for high-dimensional covariance matrices that replace the usual identity target with the Reynolds projection of the sample covariance under a symmetry group chosen by held-out negative log-likelihood. The method forms a convex combination between the sample covariance and this structured target, with the combination weight also tuned on held-out data. It generalizes the Ledoit-Wolf estimator by allowing structured targets and generalizes fixed group-symmetric estimators by making group selection automatic. A reader would care because covariance estimation underlies many high-dimensional tasks in finance, climate modeling, and genomics, and the approach supplies concrete conditions under which exploiting symmetry reduces error. The authors prove a quantitative sufficient-match condition for dominance in Frobenius mean-squared error together with regret and oracle bounds on the adaptive steps.

Core claim

The central claim is that a two-tier procedure selecting a finite symmetry group from a candidate library via held-out negative log-likelihood, then forming a convex combination of the sample covariance with its Reynolds projection under that group, satisfies a quantitative sufficient-match condition under which the estimator dominates Ledoit-Wolf shrinkage in Frobenius mean-squared error, while also admitting a finite-sample regret bound for the held-out calibration of the convex weight and an oracle inequality for the data-driven group selection.

What carries the argument

The Reynolds projection of the sample covariance onto the matrices invariant under a data-selected finite symmetry group, serving as the structured target in an adaptively weighted convex shrinkage estimator.

Load-bearing premise

The held-out data used for group selection and weight calibration is independent of the training sample and representative of the same distribution so that negative log-likelihood reliably identifies a sufficiently matching symmetry group.

What would settle it

A synthetic covariance matrix with no symmetry match where the held-out procedure still selects a group yet the resulting estimator shows higher Frobenius mean-squared error than Ledoit-Wolf shrinkage.

Figures

Figures reproduced from arXiv: 2605.17111 by Mitchell A. Thornton.

Figure 1
Figure 1. Figure 1: S&P 500 daily returns, 2015–2019, M = 55 stocks, 47 rolling windows of Ntrain = 252 days. Panel (A): held-out negative log-likelihood per day under each estimator. Panel (B): per￾window shrinkage intensity at the BMG-selected group: αˆ ∗ MSE from the closed-form plug-in (20) (blue), αˆ ∗ NLL from the K = 5-fold cross-validation (22) (orange), and the leading-order asymptotic prediction α¯ ∗ NLL from Propos… view at source ↗
Figure 2
Figure 2. Figure 2: S&P 500 daily returns, 2019–2024, M = 55 stocks, 59 rolling windows of Ntrain = 252 days. Panels (A) and (B) mirror the corresponding panels of [PITH_FULL_IMAGE:figures/full_fig_p030_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: BMG group selection across the 59 rolling windows for the 2019–2024 CRSP panel. [PITH_FULL_IMAGE:figures/full_fig_p031_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: NOAA OISST sea-surface temperature anomalies, midocean patch (30 [PITH_FULL_IMAGE:figures/full_fig_p033_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: NOAA OISST sea-surface temperature anomalies, gulfstream patch (38 [PITH_FULL_IMAGE:figures/full_fig_p034_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: BMG group selection across the 74 rolling windows for each OISST region. Panel (A) is [PITH_FULL_IMAGE:figures/full_fig_p035_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: TCGA-BRCA gene expression, M = 100 genes drawn from five MSigDB Hallmark pathways, 50 random subsample splits with Ntrain = 50 and Ntest = 200. Panel (A): held-out NLL per sample under each estimator. The Sample covariance and the Shah projection at the trivial group are rank-deficient at N < M and produce held-out NLL of order 1012 on every split; both are omitted from this panel for visual range reasons.… view at source ↗
Figure 8
Figure 8. Figure 8: RadioML 2018.A held-out NLL by training size [PITH_FULL_IMAGE:figures/full_fig_p040_8.png] view at source ↗
Figure 8
Figure 8. Figure 8: (continued) RadioML 2018.A held-out NLL, modulation classes AM-DSB-SC, BPSK, FM. [PITH_FULL_IMAGE:figures/full_fig_p041_8.png] view at source ↗
Figure 8
Figure 8. Figure 8: (continued) RadioML 2018.A held-out NLL, modulation classes GMSK, OQPSK, QPSK. [PITH_FULL_IMAGE:figures/full_fig_p042_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: RadioML 2018.A BMG selection composition by ( [PITH_FULL_IMAGE:figures/full_fig_p046_9.png] view at source ↗
Figure 9
Figure 9. Figure 9: (continued) RadioML 2018.A BMG selection composition, modulation classes AM-DSB-SC, [PITH_FULL_IMAGE:figures/full_fig_p047_9.png] view at source ↗
Figure 9
Figure 9. Figure 9: (continued) RadioML 2018.A BMG selection composition, modulation classes GMSK, [PITH_FULL_IMAGE:figures/full_fig_p048_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Galaxy10 DECaLS held-out NLL by training size [PITH_FULL_IMAGE:figures/full_fig_p050_10.png] view at source ↗
Figure 10
Figure 10. Figure 10: (continued) Galaxy10 DECaLS held-out NLL, classes 4–6. Rows and axes as in Figure [PITH_FULL_IMAGE:figures/full_fig_p051_10.png] view at source ↗
Figure 10
Figure 10. Figure 10: (continued) Galaxy10 DECaLS held-out NLL, classes 7–9. Rows and axes as in Figure [PITH_FULL_IMAGE:figures/full_fig_p052_10.png] view at source ↗
Figure 10
Figure 10. Figure 10: (continued) Galaxy10 DECaLS held-out NLL, class 10. Rows and axes as in Figure [PITH_FULL_IMAGE:figures/full_fig_p053_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Galaxy10 DECaLS BMG selection composition by ( [PITH_FULL_IMAGE:figures/full_fig_p056_11.png] view at source ↗
Figure 11
Figure 11. Figure 11: (continued) Galaxy10 DECaLS BMG selection composition, classes 4–6. Rows and axes [PITH_FULL_IMAGE:figures/full_fig_p057_11.png] view at source ↗
Figure 11
Figure 11. Figure 11: (continued) Galaxy10 DECaLS BMG selection composition, classes 7–9. Rows and axes [PITH_FULL_IMAGE:figures/full_fig_p058_11.png] view at source ↗
Figure 11
Figure 11. Figure 11: (continued) Galaxy10 DECaLS BMG selection composition, class 10. Rows and axes as [PITH_FULL_IMAGE:figures/full_fig_p059_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: CIFAR-10 image-patch covariances at 16 × 16 grayscale resolution (M = 256), Ntrain = 4,000 and Ntest = 1,000 per class. Per-class held-out NLL margin over LW 2004 (negative = AD has lower NLL); per-class αˆ ∗ NLL and αˆ ∗ MSE shrinkage intensities at the BMG-selected group. AD-NLL-BMG is preferred on all ten classes; AD-MSE-BMG is at parity with LW. Principal results. Across the ten classes, AD-NLL-BMG do… view at source ↗
Figure 13
Figure 13. Figure 13: BMG group selection across the ten CIFAR-10 classes. The procedure selects [PITH_FULL_IMAGE:figures/full_fig_p062_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Distribution-shift behavior of the AD-NLL-BMG-vs-LW per- class held-out NLL margin [PITH_FULL_IMAGE:figures/full_fig_p065_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: CIFAR-10 protocol sweep, class 0 airplane at [PITH_FULL_IMAGE:figures/full_fig_p077_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: CIFAR-10.1 protocol sweep, class 0 airplane at [PITH_FULL_IMAGE:figures/full_fig_p079_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: RadioML 2018.A protocol sweep, BPSK class at 18 dB SNR, [PITH_FULL_IMAGE:figures/full_fig_p082_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: NOAA OISST midocean region, the protocol sweep. Left: all six estimators on the [PITH_FULL_IMAGE:figures/full_fig_p085_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: NOAA OISST gulfstream region, the protocol sweep. Same two-panel layout as Figure [PITH_FULL_IMAGE:figures/full_fig_p086_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Phase diagram for the AD shrinkage family in ( [PITH_FULL_IMAGE:figures/full_fig_p091_20.png] view at source ↗
read the original abstract

We develop a class of data-adaptive shrinkage estimators for high-dimensional covariance estimation in which the shrinkage target is a Reynolds projection of the sample covariance under a finite symmetry group selected from a candidate library by held-out predictive performance. The class generalizes the convex shrinkage estimator of Ledoit and Wolf by replacing the scalar-identity target with a structured target derived from a symmetry group when one is available, and generalizes the group-symmetric maximum-likelihood estimator of Shah and Chandrasekaran by combining structural targeting with adaptive convex shrinkage and by selecting the group from data rather than treating it as prespecified. A two-tier procedure performs the group selection: a universal per-candidate evaluation based on held-out negative log-likelihood, optionally preceded by a domain-specific step that constructs the candidate library from structural priors. We establish a finite-sample regret bound for the held-out calibration of the convex combination weight, an oracle inequality for the data-driven group selection, and a quantitative sufficient-match condition under which the proposed estimator dominates Ledoit-Wolf shrinkage in Frobenius mean-squared error. The procedure is illustrated on six real-data problems spanning finance (S&P~500 daily returns), climate (NOAA OISST sea-surface temperature anomalies), genomics (TCGA-BRCA gene expression), radio signal processing (RadioML 2018.A), astronomical imaging (Galaxy10 DECaLS), and natural image patches (CIFAR-10 with a CIFAR-10.1 distribution-shift companion). An empirical comparison is also made against the Bayesian permutation-symmetry estimator of Chojecki and colleagues. Outside the few-shot regime, where structural priors carry the most information per observation, Ledoit-Wolf shrinkage remains the appropriate baseline.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper develops symmetry-aware convex shrinkage estimators for high-dimensional covariance estimation. It selects a symmetry group from a candidate library using held-out negative log-likelihood (optionally guided by domain priors), then forms a convex combination of the sample covariance and the Reynolds projection under the selected group. The approach generalizes Ledoit-Wolf shrinkage and prespecified group-symmetric MLE. It establishes a finite-sample regret bound for calibration of the convex weight, an oracle inequality for the data-driven group selection, and a quantitative sufficient-match condition under which the estimator dominates Ledoit-Wolf in Frobenius mean-squared error. The procedure is illustrated on six real datasets spanning finance, climate, genomics, radio signals, astronomy, and natural images, with comparisons to Ledoit-Wolf and a Bayesian permutation-symmetry estimator.

Significance. If the derivations hold, the finite-sample regret bound for the shrinkage weight and the oracle inequality for group selection constitute clear strengths, providing non-asymptotic guarantees that go beyond typical asymptotic analyses in high-dimensional covariance estimation. The sufficient-match condition offers a concrete route to dominance over Ledoit-Wolf when structural symmetry is present. The real-data illustrations across diverse domains demonstrate practical utility, though the strength of the dominance claim depends on how well the theory transfers to the non-Gaussian settings in the examples.

major comments (2)
  1. [the section deriving the sufficient-match condition] The section deriving the sufficient-match condition for Frobenius MSE dominance over Ledoit-Wolf: the condition is stated quantitatively, yet the group selection step minimizes held-out Gaussian negative log-likelihood. In the six real-data examples (daily returns, sea-surface temperatures, gene expression, radio signals, astronomical images, natural patches), the distributions are plausibly non-Gaussian, so the selected group need not satisfy the sufficient-match condition; this directly affects whether the dominance guarantee applies to the reported empirical comparisons.
  2. [the section stating the oracle inequality for group selection] The oracle inequality for data-driven group selection: it is presented as holding when the held-out NLL identifies a sufficiently matching group, but the paper does not provide a quantitative bound on the probability that NLL selection fails to recover a group satisfying the sufficient-match condition under non-Gaussianity or mild distribution shift (as in the CIFAR-10.1 companion). This is load-bearing for extending the theoretical claims to the real-data regime outside the few-shot setting.
minor comments (2)
  1. [notation and preliminaries] The definition of the Reynolds projection and its relation to the symmetry group should be stated explicitly with a short example in the notation section to aid readers without prior exposure to group-theoretic covariance models.
  2. [empirical illustrations] In the empirical section, the tables or figures comparing estimators should report the specific symmetry group selected for each dataset alongside the performance metrics to allow direct assessment of whether the sufficient-match condition is plausibly met.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive feedback. We appreciate the positive assessment of the finite-sample regret bound, oracle inequality, and sufficient-match condition. We address each major comment below with clarifications and proposed revisions.

read point-by-point responses

  1. Referee: The section deriving the sufficient-match condition for Frobenius MSE dominance over Ledoit-Wolf: the condition is stated quantitatively, yet the group selection step minimizes held-out Gaussian negative log-likelihood. In the six real-data examples (daily returns, sea-surface temperatures, gene expression, radio signals, astronomical images, natural patches), the distributions are plausibly non-Gaussian, so the selected group need not satisfy the sufficient-match condition; this directly affects whether the dominance guarantee applies to the reported empirical comparisons.

    Authors: We agree that the sufficient-match condition and associated dominance result are derived under modeling assumptions (including that the held-out NLL serves as a suitable surrogate for Frobenius risk) that are most directly justified in Gaussian or sub-Gaussian regimes. The real-data examples are presented as empirical illustrations rather than as formal verification of the dominance theorem. In the revision we will add an explicit caveat in the theoretical section and in the discussion of the experiments clarifying the scope of the guarantee. We will also insert a short controlled simulation study under non-Gaussian noise to illustrate when the selected group continues to yield improvement even if the quantitative sufficient-match condition is only approximately satisfied. revision: partial

  2. Referee: The oracle inequality for data-driven group selection: it is presented as holding when the held-out NLL identifies a sufficiently matching group, but the paper does not provide a quantitative bound on the probability that NLL selection fails to recover a group satisfying the sufficient-match condition under non-Gaussianity or mild distribution shift (as in the CIFAR-10.1 companion). This is load-bearing for extending the theoretical claims to the real-data regime outside the few-shot setting.

    Authors: The oracle inequality already bounds the excess risk of the data-driven selector relative to the oracle group without requiring that the selected group satisfy the sufficient-match condition with high probability. Introducing a separate high-probability bound on the event that NLL fails to recover a matching group would necessitate stronger assumptions on the degree of non-Gaussianity or distribution shift; such assumptions would limit rather than broaden the result. We therefore view the current oracle inequality as the appropriate non-asymptotic statement and will add a short clarifying remark distinguishing the roles of the oracle inequality and the sufficient-match condition. The real-data comparisons remain empirical demonstrations of practical utility. revision: no

Circularity Check

0 steps flagged

No significant circularity; derivation relies on independent held-out evaluation and mathematical bounds

full rationale

The paper's central results consist of a finite-sample regret bound on held-out calibration of the convex weight, an oracle inequality for data-driven group selection via held-out negative log-likelihood, and a quantitative sufficient-match condition for Frobenius dominance over Ledoit-Wolf. These are derived as mathematical statements that bound the procedure's risk relative to an oracle or under an explicit condition; they do not reduce by construction to the fitted values or selection criterion. The held-out NLL serves as an independent performance measure separate from the training sample, and the dominance claim is conditional rather than unconditional. No self-definitional steps, fitted inputs renamed as predictions, or load-bearing self-citations appear in the derivation chain.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claims rest on the availability of a candidate library containing at least one sufficiently matching symmetry group and on the representativeness of held-out data for selection; no new physical entities are postulated.

free parameters (1)
  • convex shrinkage weight
    Data-adaptive weight balancing the sample covariance against the symmetry-projected target, calibrated by held-out negative log-likelihood.
axioms (1)
  • domain assumption A finite symmetry group from the candidate library adequately represents the data structure.
    Invoked for the Reynolds projection to yield a useful target and for the sufficient-match condition to guarantee dominance over Ledoit-Wolf.

pith-pipeline@v0.9.0 · 5839 in / 1489 out tokens · 78802 ms · 2026-05-20T14:50:11.208852+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages

  1. [1]

    Bickel and Elizaveta Levina

    Peter J. Bickel and Elizaveta Levina. Covariance regularization by thresholding.Annals of Statistics, 36(6):2577–2604, 2008a. doi: 10.1214/08-AOS600. Peter J. Bickel and Elizaveta Levina. Regularized estimation of large covariance matrices.Annals of Statistics, 36(1):199–227, 2008b. doi: 10.1214/009053607000000758. Tony Cai and Weidong Liu. Adaptive thres...

    work page doi:10.1214/08-aos600

  2. [2]

    96 Tony T

    doi: 10.1198/jasa.2011.tm10560. 96 Tony T. Cai and Harrison H. Zhou. Optimal rates of convergence for sparse covariance matrix estimation.Annals of Statistics, 40(5):2389–2420,

    work page doi:10.1198/jasa.2011.tm10560 2011

  3. [3]

    Center for Research in Security Prices

    doi: 10.1214/12-AOS998. Center for Research in Security Prices. CRSP US Stock Database. Center for Research in Security Prices, Booth School of Business, The University of Chicago. Distributed via Wharton Research Data Services (WRDS),https://wrds-www.wharton.upenn.edu/.,

    work page doi:10.1214/12-aos998

  4. [4]

    Adam Chojecki, Pawe l Morgen, and Bartosz Ko lodziejek

    doi: 10.1109/TSP.2010.2053029. Adam Chojecki, Pawe l Morgen, and Bartosz Ko lodziejek. Learning permutation symmetry of a Gaussian vector with gips in R.Journal of Statistical Software, 112(7):1–38,

    work page doi:10.1109/tsp.2010.2053029 2010

  5. [5]

    doi: 10.18637/jss.v112.i07. A. P. Dempster. Covariance selection.Biometrics, 28(1):157–175,

    work page doi:10.18637/jss.v112.i07

  6. [6]

    doi: 10.1038/s41587-020-0546-8. Gene H. Golub and Charles F. Van Loan.Matrix Computations. Johns Hopkins University Press, 4th edition,

    work page doi:10.1038/s41587-020-0546-8

  7. [7]

    doi: 10.1214/22-AOS2174. L. R. Haff. Empirical Bayes estimation of the multivariate normal covariance matrix.The Annals of Statistics, 8(3):586–597,

    work page doi:10.1214/22-aos2174

  8. [8]

    doi: 10.1111/j.1467-9868.2008.00666.x. Roger A. Horn and Charles R. Johnson.Matrix Analysis. Cambridge University Press, 2nd edition,

    work page doi:10.1111/j.1467-9868.2008.00666.x 2008

  9. [9]

    Journal of Climate , author =

    doi: 10.1175/JCLI-D-20-0166.1. Gordon James and Adalbert Kerber.The Representation Theory of the Symmetric Group, volume 16 ofEncyclopedia of Mathematics and Its Applications. Addison-Wesley,

    work page doi:10.1175/jcli-d-20-0166.1

  10. [10]

    Olivier Ledoit and Michael Wolf

    doi: 10.1016/S0047-259X(03) 00096-4. Olivier Ledoit and Michael Wolf. Nonlinear shrinkage estimation of large-dimensional covariance matrices.Annals of Statistics, 40(2):1024–1060,

    work page doi:10.1016/s0047-259x(03

  11. [11]

    Olivier Ledoit and Michael Wolf

    doi: 10.1214/12-AOS989. Olivier Ledoit and Michael Wolf. Analytical nonlinear shrinkage of large-dimensional covariance matrices.Annals of Statistics, 48(5):3043–3065,

    work page doi:10.1214/12-aos989

  12. [12]

    Karim Lounici

    doi: 10.1016/j.cels.2015.12.004. Karim Lounici. High-dimensional covariance matrix estimation with missing observations.Bernoulli, 20(3):1029–1058,

    work page doi:10.1016/j.cels.2015.12.004 2015

  13. [13]

    Timothy J

    doi: 10.1070/SM1967v001n04ABEH001994. Timothy J. O’Shea, Tamoghna Roy, and T. Charles Clancy. Over-the-air deep learning based radio signal classification.IEEE Journal of Selected Topics in Signal Processing, 12(1):168–179,

    work page doi:10.1070/sm1967v001n04abeh001994

  14. [14]

    Mohsen Pourahmadi.High-Dimensional Covariance Estimation: With High-Dimensional Data

    doi: 10.1109/JSTSP.2018.2797022. Mohsen Pourahmadi.High-Dimensional Covariance Estimation: With High-Dimensional Data. Wiley,

    work page doi:10.1109/jstsp.2018.2797022 2018

  15. [15]

    98 Juliane Sch¨ afer and Korbinian Strimmer

    doi: 10.1175/2007JCLI1824.1. 98 Juliane Sch¨ afer and Korbinian Strimmer. A shrinkage approach to large-scale covariance matrix estimation and implications for functional genomics.Statistical Applications in Genetics and Molecular Biology, 4(1),

    work page doi:10.1175/2007jcli1824.1

  16. [16]

    Ilya Soloveychik, Dmitry Trushin, and Ami Wiesel

    doi: 10.1214/12-EJS723. Ilya Soloveychik, Dmitry Trushin, and Ami Wiesel. Group symmetric robust covariance estimation. IEEE Transactions on Signal Processing, 64(1):244–257,

    work page doi:10.1214/12-ejs723

  17. [17]

    Charles Stein

    doi: 10.1109/TSP.2015.2486739. Charles Stein. Inadmissibility of the usual estimator for the mean of a multivariate normal distribution. InProceedings of the Third Berkeley Symposium on Mathematical Statistics and Probability, volume 1, pages 197–206. University of California Press,

    work page doi:10.1109/tsp.2015.2486739 2015

  18. [18]

    Charles M

    doi: 10.1007/BF01085007. Charles M. Stein. Estimation of the mean of a multivariate normal distribution.The Annals of Statistics, 9(6):1135–1151,

    work page doi:10.1007/bf01085007

  19. [19]

    Mitchell A

    doi: 10.1038/nature11412. Mitchell A. Thornton. Algebraic diversity for high-dimensional covariance estimation, 2026a. Mitchell A. Thornton. Algebraic diversity for spectral estimation, 2026b. Roman Vershynin.High-Dimensional Probability: An Introduction with Applications in Data Science, volume 47 ofCambridge Series in Statistical and Probabilistic Mathe...

    work page doi:10.1038/nature11412