Pith. sign in

REVIEW 3 major objections 5 minor 63 references

Entropy Adjusted Graphical Lasso for Sparse Precision Matrix Estimation

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Lower sparse-matrix errors with entropy-adjusted Graphical Lasso.

desk verdict EAGL is GLasso with a rescaled covariance and penalty, so the entropy-adjustment claim does not survive the paper's own algebra. read the letter →

arxiv 2501.05308 v1 pith:SWQGBTJV submitted 2025-01-09 stat.ME stat.CO

Vahe Avagyan This is my paper · ORCID
classification stat.MEstat.CO MSC 62H1262J0762F12
keywords ElasticNetEntropyadjustmentGaussianGraphicalModelsLassoMaximumLikelihoodEstimationprecisionmatrixlog-determinantpenalty
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 introduces the Entropy Adjusted Graphical Lasso (EAGL), a precision-matrix estimator that appends a log-determinant term, interpreted as the entropy of the multivariate Gaussian, to the Graphical Lasso objective. The claim is that Lasso-induced sparsity increases uncertainty by shrinking eigenvalues, and the log-determinant penalty counteracts that shrinkage, yielding better-conditioned estimates without sacrificing sparsity. Across seven simulated precision-matrix structures and two real applications (prostate-cancer classification and S&P 500 minimum-variance portfolios), the paper reports that EAGL lowers Kullback-Leibler and matrix losses relative to GLasso and is at least competitive with elastic-net, ridge, and SCAD alternatives. If the empirical claims hold, practitioners get a target-free, computationally cheap variant of GLasso that improves both matrix accuracy and graph-selection performance.

What carries the argument

The central object is the additive penalty $\gamma(1-\alpha)\log\det(\Omega^{-1})$ in Eq. (7). Because $\log\det(\Omega^{-1}) = -\log\det(\Omega)$, this term has exactly the same functional form as the log-determinant already present in the Gaussian log-likelihood; substituting it turns the EAGL objective into Eq. (16), a GLasso objective with likelihood weight $1+\gamma(1-\alpha)$ and Lasso weight $\gamma\alpha$. The paper uses this rewritten form for its convergence proof and interprets the extra term as an entropy correction that prevents the eigenvalue shrinkage caused by Lasso. This is the mechanism the paper identifies as the source of its improved numerical performance.

What would settle it

Solve EAGL for a small $p$ with fixed $\gamma$ and $\alpha$ and compare the estimate entrywise with $(1+\gamma(1-\alpha))$ times the GLasso solution at penalty $\gamma\alpha$; if the two matrices agree to numerical precision and share the same sparsity pattern, the entropy adjustment is a rescaling of GLasso, not a new penalty mechanism.

Watch

Extended reading notes

Core claim

The central claim is that uncertainty should be penalized directly when estimating a sparse precision matrix. EAGL solves the problem in Eq. (7): minimize $-\log\det(\Omega) + \operatorname{trace}(\Omega S) + \gamma[\alpha\|\Omega\|_1 + (1-\alpha)\log\det(\Omega^{-1})]$, where the added log-determinant term equals Gaussian entropy up to additive constants and is intended to keep the eigenvalues of the estimated precision matrix from shrinking as fast as they do under plain GLasso. The paper reports that, averaged over 100 replications for seven models, EAGL outperforms GLasso on reverse Kullback-Leibler loss, relative trace error, Frobenius, spectral, and $\ell_1$ losses for all models, and delivers the best out-of-sample portfolio risk and Sharpe ratio in the S&P 500 application. A convergence rate in the Frobenius norm is established under fixed-$p$ asymptotics.

Load-bearing premise

The load-bearing premise is that adding $\log\det(\Omega^{-1})$ as a separate penalty genuinely adjusts entropy rather than duplicating the log-determinant term already present in the Gaussian likelihood; if the duplication is the whole story, the proposed estimator reduces to a rescaled GLasso and the claimed new mechanism collapses.

Editorial extensions

If this is right

  • If EAGL's reported gains hold, practitioners can obtain lower statistical losses than GLasso without choosing a target matrix, using only the scalar $\alpha$ (fixed at 0.5 in the paper).
  • The estimator remains fully sparse because the Lasso term is retained, so graph-selection tools built on GLasso can be applied directly to EAGL output.
  • In the paper's prostate-cancer LDA experiment, EAGL's misclassification rate (0.128) is lower than GLasso's (0.168) and close to the targeted T-GEN, without requiring target-matrix selection.
  • In the S&P 500 application, the minimum-variance portfolio built on EAGL attains the lowest out-of-sample risk (0.034) and highest Sharpe ratio (0.438) among all compared estimators.
  • The established Frobenius-norm convergence rate is of the same order as standard GLasso rates, so the entropy adjustment does not sacrifice asymptotic accuracy.

Reading between the lines

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

  • Since $\log\det(\Omega^{-1}) = -\log\det(\Omega)$, the EAGL objective is algebraically a rescaled GLasso objective (Eq. 16); the EAGL solution is therefore $(1+\gamma(1-\alpha))$ times the GLasso solution for the same sample covariance with penalty $\gamma\alpha$. This is our inference from the paper's own rewrite, not a claim the author draws.
  • Because the rescaling factor is positive, EAGL and GLasso at the corresponding penalty share the exact same sparsity pattern; any reported differences in graph-selection metrics reflect different effective tuning levels rather than a different selection mechanism.
  • The eigenvalue-preservation effect shown in Figure 1 follows from multiplying the GLasso solution by a factor greater than one, which inflates all eigenvalues; a direct test would compare EAGL against GLasso after setting the effective penalty and covariance scale to the same values.
  • The same algebraic reduction applies to any 'entropy-adjusted' version of another estimator (for instance the suggested entropy-adjusted Graphical Ridge), so the framework is a rescaling recipe rather than a new penalty family.
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.

Referee Report

3 major / 5 minor

Summary. The paper proposes an 'Entropy Adjusted Graphical Lasso' (EAGL) estimator for sparse precision matrices. The proposal adds a log-determinant penalty term, interpreted as an entropy adjustment, to the standard Graphical Lasso objective. The authors report extensive simulation comparisons across seven precision-matrix models, two real-data applications, and an asymptotic convergence-rate result. The central claim is that EAGL is a new estimator that improves on GLasso and other existing methods by simultaneously promoting sparsity and reducing uncertainty.

Significance. If the proposed method were genuinely new, the entropy-adjustment idea could be of interest for precision-matrix estimation, and the empirical study is broad: seven simulation models, several loss functions, sparsity-pattern metrics, two real-data applications, and both CV and BIC tuning. However, the central methodological claim is not supported: the 'additional' log-determinant penalty is algebraically identical to the log-determinant term already present in the Gaussian negative log-likelihood. As a result, EAGL is a reparameterization of GLasso rather than a new estimator, and the interpretation of the numerical improvements as an entropy effect is not valid.

major comments (3)
  1. [Section 2.2, Eq. (7)] The proposed objective collapses into the standard GLasso objective. Since log det(Omega^{-1}) = -log det(Omega), Eq. (7) equals -(1 + gamma(1-alpha)) log det(Omega) + trace(Omega S) + gamma alpha ||Omega||_1. The factor 1 + gamma(1-alpha) is a positive constant, so the argmin is identical to that of Eq. (2) with penalty parameter gamma alpha / (1 + gamma(1-alpha)). Thus EAGL is not a new estimator; it is GLasso with a reparameterized tuning parameter. Appendix D's Eq. (16) displays exactly this equivalent form but the paper does not acknowledge that this is simply GLasso. Consequently, the central 'entropy adjustment' claim, the eigenvalue comparison in Figure 1, and the interpretation of the simulation gains as arising from an additional penalty are unsupported; the gains, where present, are due to a smaller effective Lasso penalty.
  2. [Section 3.5 and Table 2] The text and the summary table directly contradict each other on the uMCC results. Section 3.5 states that GLasso performs better than EAGL in terms of uMCC for models 4, 5 and 6, while Table 2 reports EAGL (checkmark) as better for all seven models. The underlying tables support Table 2: for Model 5 the uMCC values are 0.677 (EAGL) vs 0.669 (GLasso), for Model 6 they are 0.776 vs 0.708, and for Model 7 they are 0.815 vs 0.744. The narrative in Section 3.5 needs to be reconciled with the reported numbers before the comparison claims can be assessed.
  3. [Appendix D, Eqs. (15)-(20)] The asymptotic proof contains substantive gaps. The rate in Eq. (15) is stated under 'standard asymptotics (i.e., assuming that p remains fixed)' while the bound contains (p+s) log p, which is not a meaningful divergence rate when p is fixed. More importantly, the derivation of Eq. (20) from Eq. (18) is not justified: the coefficient 1/4 appears without proper accounting for the factor in the Taylor expansion lower bound, and Eq. (19) writes lambda_min(Omega + tau Delta)^{-1} in a way that is not clearly defined. In addition, Eq. (21) bounds trace(Sigma Delta) by ||Sigma||_2 ||Delta||_2 and then treats this as a lower bound, but trace(Sigma Delta) can be negative; a lower bound would require -||Sigma||_2 ||Delta||_2. These issues mean the claimed convergence rate in Eq. (15) is not established by the presented argument.
minor comments (5)
  1. [Section 2.1, after Eq. (5)] The text says that for alpha = 0 the Graphical Elastic-Net reduces to 'the Rope approach (4)', but the Rope estimator is defined in Eq. (3); Eq. (4) is the targeted Ridge estimator.
  2. [Section 4.1 and Table 3] There is a typo: 'T-Gridg' should be 'T-GRidge', and 'missclassification' should be 'misclassification' throughout the section and table caption.
  3. [Section 3.5] The placeholder lines 'Tables 5 - 11 about here' and 'Figure 2 about here' should be replaced by the actual tables and figure in the final version.
  4. [Section 2.2, Eq. (7)] The notation H_C(X) is used in Eq. (7) before it is defined; the definition H_C(X) = log det(Omega^{-1}) should appear immediately before or with Eq. (7), rather than only in the following sentence and Appendix A.
  5. [Section 3.2, Model 4] The description 'generated using Matlab commands prandsym with a parameter 0.5' would benefit from a precise citation or a full description of the construction, since the resulting sparsity level and eigenvalue properties are important for reproducibility.

Circularity Check

2 steps flagged · score 8.0 of 10

EAGL's 'additional log-determinant penalty' is the same log-det term as the Gaussian likelihood; Eq. (16) shows the estimator reduces to GLasso with a rescaled covariance and penalty.

  1. self definitional [Section 2.2, Eq. (7); Appendix A; Appendix D, Eq. (16)]
    "The proposed approach is an augmented version of the GLasso method (2) with an additional HC(X) = log det(Omega^-1) = log(det(Sigma)) penalty term. ... Consider the following optimization problem: bOmega_EAGL = arg min_Omega - log det(Omega) + trace(Omega S) + gamma { alpha ||Omega||_1 + (1 - alpha) log det(Omega^-1) }, (7)"

    The 'additional' penalty is the negative of the log-determinant term already in the Gaussian negative log-likelihood: log det(Omega^-1) = -log det(Omega). Substituting into (7) gives the objective in (16): -(1+(1-alpha)gamma) log det(Omega) + trace(Omega S) + gamma alpha ||Omega||_1. Since 1+(1-alpha)gamma is a positive constant independent of Omega, dividing by it leaves the argmin unchanged, yielding -log det(Omega) + (1/C) trace(Omega S) + (gamma alpha / C) ||Omega||_1, which is exactly the GLasso objective (2) with sample covariance S/C and penalty gamma alpha / C. Hence no independent entropy penalty exists; the claimed extra term is the likelihood term itself, and the method is GLasso on a rescaled covariance.

  2. renaming known result [Appendix D, Eq. (16) and surrounding proof]
    "First, we rewrite the optimization problem of our proposed methodology as: bOmega_EAGL = arg min_Omega - (1 + (1 - alpha) gamma) log det(Omega) + trace(Omega S) + gamma alpha ||Omega||_1. (16) ... The proof of (15) is motivated by Rothman et al. (2008)."

    This rewrite is the paper's own reduction of EAGL to a GLasso-type problem. After Eq. (16), the proof of the Frobenius-norm rate (15) is 'motivated by Rothman et al. (2008)' and proceeds by standard GLasso/sparse log-det inequalities. Thus the theoretical result is inherited from GLasso theory with a rescaled input, not from a new entropy penalty. Presenting EAGL as a new entropy-adjusted estimator while its objective is, by construction, GLasso with modified S and penalty is renaming a known estimator family rather than deriving a first-principles entropy adjustment.

full rationale

Score is 8 rather than 0-2 because the central methodological claim, not just a peripheral citation, reduces by definition. The identity log det(Omega^-1) = -log det(Omega) makes the 'additional penalty' in Eq. (7) identical to the log-det term already present in the Gaussian negative log-likelihood in Eq. (2); the paper itself writes the reduced objective as Eq. (16). This is not a matter of interpretation: EAGL equals GLasso applied to S/(1+(1-alpha)gamma) with Lasso tuning gamma alpha/(1+(1-alpha)gamma), so the entropy term is absorbed into the likelihood. The simulations and real-data benchmarks are externally evaluated and the estimator may perform well, but they cannot rescue the stated novelty, because EAGL is not a new penalty family; it is a reparameterized/renamed GLasso with a scaled covariance. No load-bearing self-citation was found; the circularity is definitional, not citational.

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

The estimator depends on two tuning parameters, alpha and gamma, both chosen by hand or cross-validation. The core modeling axioms are multivariate normality and the entropy interpretation of log det(Omega^{-1}), the latter of which is algebraically redundant with the Gaussian likelihood. No new physical or conceptual entities are introduced.

free parameters (2)
  • alpha = 0.5
    Combination weight between Lasso and log-det penalties; fixed by hand at 0.5 following Bernardini et al. and Kovacs et al., not tuned.
  • gamma = selected by 5-fold CV or BIC
    Tuning parameter for the penalty; chosen by cross-validation in the main results, a data-driven free parameter.
assumptions (5)
  • domain assumption Data are i.i.d. multivariate normal with covariance Sigma = Omega^{-1}.
    Invoked in Section 2.1 to justify the Gaussian log-likelihood and entropy interpretation.
  • domain assumption log det(Omega^{-1}) is an appropriate measure of uncertainty or entropy.
    Appendix A derives H(X) = C + (1/2) log det(Omega^{-1}) for Gaussian X; this motivates the penalty but is also the source of redundancy with the likelihood.
  • standard math Eigenvalue bounds A1 and A2: lambda_min(Omega) >= lambda > 0 and lambda_max(Omega) <= lambda_bar.
    Assumptions in Appendix D used to establish the Frobenius-norm convergence rate.
  • domain assumption Fixed-p asymptotic regime with p fixed and n going to infinity.
    Appendix D states standard asymptotics; the claimed rate is not a high-dimensional result despite the paper's simulation setting.
  • standard math Convexity of the EAGL objective.
    Needed for the argument that a positive lower bound on a sphere implies the minimizer is inside; objective is convex for alpha > 0.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Entropy Adjusted Graphical Lasso for Sparse Precision Matrix Estimation." pith.science (2026). https://pith.science/paper/SWQGBTJV

@misc{pith2026250105308,
  author       = {Pith},
  title        = {Pith review of: Entropy Adjusted Graphical Lasso for Sparse Precision Matrix Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SWQGBTJV}},
  note         = {Machine review of arXiv:2501.05308}
}
read the original abstract

The estimation of a precision matrix is a crucial problem in various research fields, particularly when working with high dimensional data. In such settings, the most common approach is to use the penalized maximum likelihood. The literature typically employs Lasso, Ridge and Elastic Net norms, which effectively shrink the entries of the estimated precision matrix. Although these shrinkage approaches provide well-conditioned precision matrix estimates, they do not explicitly address the uncertainty associated with these estimated matrices. In fact, as the matrix becomes sparser, the precision matrix imposes fewer restrictions, leading to greater variability in the distribution, and thus, to higher entropy. In this paper, we introduce an entropy-adjusted extension of widely used Graphical Lasso using an additional log-determinant penalty term. The objective of the proposed technique is to impose sparsity on the precision matrix estimate and adjust the uncertainty through the log-determinant term. The advantage of the proposed method compared to the existing ones in the literature is evaluated through comprehensive numerical analyses, including both simulated and real-world datasets. The results demonstrate its benefits compared to existing approaches in the literature, with respect to several evaluation metrics.

Figures

Figures reproduced from arXiv: 2501.05308 by the authors.

Figure 1
Figure 1. Gaussian Graphical models corresponding to each precision matrix model. [PITH_FULL_IMAGE:figures/full_fig_p035_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

63 extracted references · 42 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type url volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.all := #1 'mid.sentence := ...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in " " * FUNCTION format....

  3. [3]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in " " * FUNCTION format....

  4. [4]

    , " * write output.state after.block = add.period write newline

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.all := #1 'mid.sentence := #2 '...

  5. [5]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in " " * FUNCTION format....

  6. [6]

    Anderson, T., W. (2003). An Introduction to Multivariate Statistical Analysis. Wiley-Interscience

    work page 2003

  7. [7]

    Avagyan, V. (2021). D-trace estimation of a precision matrix with eigenvalue control. Communications in Statistics-Simulation and Computation , 50(4):1231--1247

    work page 2021

  8. [8]

    M., and Nogales, F

    Avagyan, V., Alonso, A. M., and Nogales, F. J. (2018). D-trace estimation of a precision matrix using adaptive lasso penalties. Advances in Data Analysis and Classification , 12(2):425--447

    work page 2018

Show all 63 references
  1. [9]

    and Mei, X

    Avagyan, V. and Mei, X. (2022). Precision matrix estimation under data contamination with an application to minimum variance portfolio selection. Communications in Statistics-Simulation and Computation , 51(4):1381--1400

    2022

  2. [10]

    Banerjee, O., El Ghaoui, L., d'Aspremont, A., and Natsoulis, G. (2006). Convex optimization techniques for fitting sparse gaussian graphical models. Proceedings of the 23^ rd International Conference on Machine Learning

    2006

  3. [11]

    Bekker, A., Kheyri, A., and Arashi, M. (2023). A computational note on the graphical ridge in high-dimension. arXiv preprint arXiv:2312.15781

    2023 arXiv

  4. [12]

    Bernardini, D., Paterlini, S., and Taufer, E. (2022). New estimation approaches for graphical models with elastic net penalty. Econometrics and Statistics

    2022

  5. [13]

    and Levina, E

    Bickel, P., J. and Levina, E. (2008). Regularized estimation of large covariance matrices. The Annals of Statistics , 36(1):199--227

    2008

  6. [14]

    and Tibshirani, R

    Bien, J. and Tibshirani, R. J. (2011). Sparse estimation of a covariance matrix. Biometrika , 98(4):807--820

    2011

  7. [15]

    Bishop, C. M. and Nasrabadi, N. M. (2006). Pattern recognition and machine learning . Springer

    2006

  8. [16]

    H., Ong, C

    Brodersen, K. H., Ong, C. S., Stephan, K. E., and Buhmann, J. M. (2010). The balanced accuracy and its posterior distribution. In 2010 20th international conference on pattern recognition

    2010

  9. [17]

    Cai, T., Liu, W., and Luo, X. (2011). A constrained _1 minimization approach to sparse precision matrix estimation. Journal of the American Statistical Association , 106(494):594--607

    2011

  10. [18]

    T., Li, H., Liu, W., and Xie, J

    Cai, T. T., Li, H., Liu, W., and Xie, J. (2013). Covariate-adjusted precision matrix estimation with an application in genetical genomics. Biometrika , 100(1):139--156

    2013

  11. [19]

    and Jurman, G

    Chicco, D. and Jurman, G. (2020). The advantages of the matthews correlation coefficient (mcc) over f1 score and accuracy in binary classification evaluation. BMC Genomics , 21(1):1--13

    2020

  12. [20]

    Choi, Y.-G., Lim, J., and Choi, S. (2019). High-dimensional markowitz portfolio optimization problem: empirical comparison of covariance matrix estimators. Journal of Statistical Computation and Simulation , 89(7):1278--1300

    2019

  13. [21]

    J., and Uppal, R

    DeMiguel, V., Garlappi, L., Nogales, F. J., and Uppal, R. (2009a). A generalized approach to portfolio optimization: Improving performance by constraining portfolio norms. Management Science , 55(5):798--812

    2009

  14. [22]

    DeMiguel, V., Garlappi, L., and Uppal, R. (2009b). Optimal versus naive diversification: How inefficient is the 1/n portfolio strategy? The Review of Financial studies , 22(5):1915--1953

    2009

  15. [23]

    Dempster, A. (1972). Covariance selection. Biometrics , 28(1):157--175

    1972

  16. [24]

    Fan, J., Feng, J., and Wu, Y. (2009). Network exploration via the adaptive lasso and scad penalties. The Annals of Applied Statistics , 3(2):521--541

    2009

  17. [25]

    Fan, J., Liao, Y., and Liu, H. (2016). An overview of the estimation of large covariance and precision matrices. The Econometrics Journal , 19(1):C1--C32

    2016

  18. [26]

    Friedman, J., Hastie, T., and Tibshirani, R. (2008). Sparse inverse covariance estimation with the graphical lasso. Biostatistics , 9(3):432--441

    2008

  19. [27]

    and Xu, Y

    Goto, S. and Xu, Y. (2015). Improving mean variance optimization through sparse hedging restrictions. Journal of Financial and Quantitative Analysis , 50(6):1415--1441

    2015

  20. [28]

    Haff, L. (1980). Empirical bayes estimation of the multivariate normal covariance matrix. The Annals of Statistics , pages 586--597

    1980

  21. [29]

    Huang, S., Li, J., Sun, L., Ye, J., Fleisher, A., Wu, T., Chen, K., and Reiman, E. (2010). Learning brain connectivity of alzheimer's disease by sparse inverse covariance estimation. NeuroImage , 50:935--949

    2010

  22. [30]

    and Stein, C

    James, W. and Stein, C. (1992). Estimation with quadratic loss. In Breakthroughs in statistics: Foundations and basic theory , pages 443--460. Springer

    1992

  23. [31]

    Kov \'a cs, S., Ruckstuhl, T., Obrist, H., and B \"u hlmann, P. (2021). Graphical elastic net and target matrices: Fast algorithms and software for sparse precision matrix estimation. arXiv preprint arXiv:2101.02148

    2021 arXiv

  24. [32]

    Kuismin, M., Kemppainen, J., and Sillanp \"a \"a , M. (2017). Precision matrix estimation with rope. Journal of Computational and Graphical Statistics , 26(3):682--694

    2017

  25. [33]

    Lauritzen, S. (1996). Graphical Models. Clarendon Press. Oxford

    1996

  26. [34]

    and Wolf, M

    Ledoit, O. and Wolf, M. (2004). Honey, i shrunk the sample covariance matrix. The Journal of Portfolio Management , 30(4):110--119

    2004

  27. [35]

    and Luo, X

    Liu, W. and Luo, X. (2015). Fast and adaptive sparse precision matrix estimation in high dimensions. Journal of Multivariate Analysis , 135:153--162

    2015

  28. [36]

    P., and Piantadosi, S

    Liu, Z., Lin, S., Deng, N., McGovern, D. P., and Piantadosi, S. (2016). Sparse inverse covariance estimation with _0 penalty for network construction with omics data. Journal of Computational Biology , 23(3):192--202

    2016

  29. [37]

    and Hero, A

    Marjanovic, G. and Hero, A. O. (2015). _0 sparse inverse covariance estimation. IEEE Transactions on Signal Processing , 63(12):3218--3231

    2015

  30. [38]

    and Solo, V

    Marjanovic, G. and Solo, V. (2014). On _q optimization and sparse inverse covariance selection. IEEE transactions on signal processing , 62(7):1644--1654

    2014

  31. [39]

    Markowitz, H. (1952). Portfolio selection. Journal of Finance , 7:77--91

    1952

  32. [40]

    Matthews, B. W. (1975). Comparison of the predicted and observed secondary structure of t4 phage lysozyme. Biochimica et Biophysica Acta , 405:442--451

    1975

  33. [41]

    McLachlan, S. (2004). Discriminant Analysis and Statistical Pattern Recognition. Willey Interscience

    2004

  34. [42]

    and B\" u hlmann, P

    Meinshausen, N. and B\" u hlmann, P. (2006). High-dimensional graphs and variable selection with the lasso. The Annals of Statistics , 34(2):1436--1462

    2006

  35. [43]

    F., Bilgrau, A

    Peeters, C. F., Bilgrau, A. E., and van Wieringen, W. N. (2022). rags2ridges: A one-stop-shop for graphical modeling of high-dimensional precision matrices. Journal of Statistical Software , 102(4):1--32

    2022

  36. [44]

    Peng, W., Wang, P., Zhou, N., and Zhu, J. (2009). Partial correlation estipation by joint sparse regression models. Journal of the American Statistical Association , 104(486):735--746

    2009

  37. [45]

    Ravikumar, P., Wainwright, M., Raskutti, G., and Yu, B. (2011). High-dimensional covariance estimation by minimizing _1 -penalized log-determinant divergence. Electronic Journal of Statistics , 5:935--980

    2011

  38. [46]

    Rothman, A., Bickel, P., Levina, E., and Zhu, J. (2008). Sparse permutation invariant covariance estimation. Electronic Journal of Statistics , 2:494--515

    2008

  39. [47]

    Rothman, A. J. (2012). Positive definite estimators of large covariance matrices. Biometrika , 99(2):733--740

    2012

  40. [48]

    Shannon, C. E. (1948). A mathematical theory of communication. The Bell System Technical Journal , 27(3):379--423

    1948

  41. [49]

    G., Ross, K., Jackson, D

    Singh, D., Febbo, P. G., Ross, K., Jackson, D. G., Manola, J., Ladd, C., Tamayo, P., Renshaw, A. A., D'Amico, A. V., Richie, J. P., et al. (2002). Gene expression correlates of clinical prostate cancer behavior. Cancer Cell , 1(2):203--209

    2002

  42. [50]

    Stevens, G. V. G. (1998). On the inverse of the covariance matrix in portfolio analysis. The Journal of Finance , 53(5):1821--1827

    1998

  43. [51]

    Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society Series B: Statistical Methodology , 58(1):267--288

    1996

  44. [52]

    C., Florquin, R., Delhaye, M., Kornreich, C., Williams, D

    Till, A. C., Florquin, R., Delhaye, M., Kornreich, C., Williams, D. R., and Briganti, G. (2023). A network perspective on abnormal child behavior in primary school students. Psychological reports , 126(4):1933--1953

    2023

  45. [53]

    van Wieringen, W. N. and Peeters, C. F. (2016). Ridge estimation of inverse covariance matrices from high-dimensional data. Computational Statistics & Data Analysis , 103:284--303

    2016

  46. [54]

    and Roeder, K

    Wasserman, L. and Roeder, K. (2009). High dimensional variable selection. The Annals of Statistics , 37(5A):2178

    2009

  47. [55]

    Witten, D. M. and Tibshirani, R. (2009). Covariance-regularized regression and classification for high dimensional problems. Journal of the Royal Statistical Society Series B: Statistical Methodology , 71(3):615--636

    2009

  48. [56]

    and Li, J

    Yin, J. and Li, J. (2013). Adjusting for high-dimensional covariates in sparse precision matrix estimation by _1 -penalization. Journal of Multivariate Analysis , 116:365--381

    2013

  49. [57]

    Yuan, M. (2010). High dimensional inverse covariance matrix estimation via linear programming. Journal of Machine Learning Research , 11:2261--2286

    2010

  50. [58]

    and Lin, Y

    Yuan, M. and Lin, Y. (2007). Model selection and estimation in the gaussian graphical model. Biometrika , 94(1):19--35

    2007

  51. [59]

    and Zou, H

    Zhang, T. and Zou, H. (2014). Sparse precision matrix estimation via lasso penalized d-trace loss. Biometrika , 88:1--18

    2014

  52. [60]

    and Hastie, T

    Zou, H. and Hastie, T. (2005). Regularization and variable selection via the elastic net. Journal of the Royal Statistical Society Series B: Statistical Methodology , 67(2):301--320

    2005

  53. [61]

    @esa ( ) , n @biblabelnum##1 ##1

    \@ifclassloaded aguplus natbib The aguplus class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command natbib from the document \@ifclassloaded nlinproc natbib The nlinproc class already includes natbib cod...

  54. [62]

    @stdbsttrue NAT@ctr \@lbibitem[ NAT@ctr ] \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 [ @natanchorstart #2\@extra@b@citeb \@biblabel @num @natanchorend] @ifc...

  55. [63]

    [ ,l WCHmc&('`D, K #^Ø ! msǥ sL ZL5f

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifundefined NAT@sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifundefined bib@heading @heading NAT@ctr thebibliography [1] @ \@biblabel NAT@ctr \@bib...

    2000

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.