Pith. sign in

REVIEW 3 major objections 4 minor 56 references

Hub Detection in Gaussian Graphical Models

T0 review · 3 major / 4 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 Hub detection reduces to a spectral decomposition of the precision matrix; IPC-HD recovers the hub set with high probability without estimating the graphical model.

desk verdict The spectral hub-detection idea is real and the simulations are convincing, but the main theorem proves recovery only for some unspecified threshold, not for the rules actually used. read the letter →

arxiv 2505.23707 v1 pith:GMSQ3ZA4 submitted 2025-05-29 stat.ME

José Á. Sánchez Gómez , Weibin Mo , Junlong Zhao , Yufeng Liu This is my paper
classification stat.ME MSC 62H2562H1262F12
keywords hubdetectionGaussiangraphicalmodelprecisionmatrixspectraldecompositionprincipalcomponentanalysislow-rankapproximationcovarianceestimationhigh-dimensionalstatistics
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 sets out to establish that hub detection in a Gaussian graphical model can be solved directly from the spectral decomposition of a covariance estimate, bypassing graph estimation. The authors define hubs as variables with large weighted degree in the precision matrix, and they prove that such hubs force the precision matrix to have a low effective rank, with the squared entries of the leading eigenvectors concentrated on the hub variables. On this basis they propose IPC-HD, a method that thresholds per-variable influence measures built from the top eigenvectors. If the central claim is right, hub detection becomes a fast eigenvalue computation with provable recovery, and it no longer depends on the accuracy of a full graph estimate.

What carries the argument

The machinery is the influence measure $\omega_k^{(s)}=\sum_{i=1}^s v_{ik}^2$, the diagonal of the projection matrix onto the top $s$ eigenvectors of the precision matrix. It acts as an eigenvector-based surrogate for weighted degree, and the paper shows that, under the assumed eigenvalue gap, the hub set is exactly the set of variables whose influence measures dominate the rest by a factor of order $\tau_p$. The estimator combines this with the regularized eigenvalue-ratio statistic $\delta_\rho(i)=(\hat{\gamma}_{p-i}+\rho)/(\hat{\gamma}_{p-i+1}+\rho)$ to locate the spike index $s$, then thresholds the estimated influence measures.

What would settle it

Run the simulation design of Section 5.1 but force the tail eigenvalues so that $(\lambda_{s+1}/\lambda_1)^2$ is the same order as $\min_{h\in H}\omega_h^{(s)}$, violating condition (6); if IPC-HD still separates hubs from non-hubs at the sample size predicted by Theorem 2, the stated condition is not necessary, while a failure would show the condition is the one doing the work.

Watch

Extended reading notes

Core claim

The central claim is that a hub set $H$, defined by weighted degree in the precision matrix $\Theta$, is also the set of variables whose squared coordinates in the leading $s$ eigenvectors are large. Under Assumption 1 and the influential signal condition $(\lambda_{s+1}/\lambda_1)^2 \le c \min_{h\in H}\omega_h^{(s)}$, the paper proves in Theorem 1 that $H$ is a $(\tau_p,c')$-influential set for the influence measure $\omega_k^{(s)}=\sum_{i=1}^s v_{ik}^2$, so the hub set can be recovered from the top eigenspace alone. Theorem 2 then gives the operational guarantee: for a covariance estimator satisfying the spectral-norm concentration bound of Assumption 4, IPC-HD returns exactly the hub set (with data-driven $s$) or a superset of it (with over-estimated $s$) with probability at least $1-2\Delta$, provided the estimation error satisfies $E(n,p,\Delta) \le c p^{-(1-\beta)}$.

Load-bearing premise

The load-bearing premise is that the precision matrix has one clean eigenvalue jump: the top $s$ eigenvalues are well separated from the rest, and the squared rank-$s$ approximation error $(\lambda_{s+1}/\lambda_1)^2$ is small compared with the weakest hub's influence $\min_{h\in H}\omega_h^{(s)}$; if that gap is not clean, the top sample eigenvectors can mix hubs with non-hubs, and the threshold $\kappa$ has no theoretical support.

Editorial extensions

If this is right

  • Hub detection requires only a consistent covariance estimator and its eigen-decomposition, so the method avoids the expense and tuning of full graph estimation.
  • No sparsity of the precision matrix is needed; the structural requirement is a separation rate $\tau_p=\Omega(p^\beta)$ between hub and non-hub weighted degrees, with stronger separation giving faster convergence.
  • With the sample covariance, exact recovery holds once $n \ge c\max\{p^{3-2\beta}, p^{2-2\beta}\log\Delta^{-1}\}$; with screening, masking, or thresholding, the sample size can drop to order $\log p$ under appropriate structure.
  • Over-estimating the number of spikes $s$ still guarantees that the detected set contains the true hubs, making the method usable as a screening tool even when $s$ is hard to identify.

Reading between the lines

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

  • The same eigenvector-concentration mechanism could plausibly be exported to non-Gaussian graphical models such as Ising or copula models, where the precision analogue is less direct but the spectral signature of hubs may survive; this is an extension the paper does not attempt.
  • Because the influential signal condition (6) is assumed rather than derived, a natural testable refinement would be to exhibit generative hub models that imply it from primitive degree parameters, telling practitioners when the eigenvalue gap can be verified before running IPC-HD.
  • The superset guarantee for over-estimated $s$ suggests a two-stage workflow: use IPC-HD with a deliberately large $s$ to shortlist candidate hubs, then run a full graph estimator on the shortlist; the paper's simulations support the screening version but stop short of this explicit recommendation.
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 / 4 minor

Summary. The paper develops a method, IPC-HD, for directly detecting hub variables in a Gaussian graphical model without first estimating the graph. Hubs are defined through weighted degree, the squared ℓ2 norm of each column of the precision matrix Θ. Under Assumption 1, a finite hub set with growing separation rate τp is shown to produce spiked leading eigenvalues of Θ (Proposition 1). Under an additional influential signal condition (Eq. (6)), Theorem 1 shows that the hub set coincides with an influential set based on the squared entries of the top s eigenvectors. The proposed algorithm estimates the covariance, chooses s either by a data-driven eigenvalue-ratio rule or by over-estimation ŝ=⌊p/5⌋, computes influence scores, and thresholds them. Theorem 2 and Corollaries 1–4 provide high-probability recovery guarantees for sample, screened, masked, and thresholded covariance estimators. Simulations compare IPC-HD with GLASSO, HGL, HWGL, and raw inverse correlation; a prostate cancer gene expression application detects five hub genes.

Significance. The spectral link between hub presence and eigenvector concentration is a fresh idea in this literature, and bypassing full graph estimation is a genuine practical advantage. The paper is careful to state assumptions and provides convergence rates that improve as the hub separation strengthens; it also covers several covariance estimators and makes code available. The main value is conditional: Theorem 2 guarantees recovery only for some unspecified threshold κ and regularizer ρ, while the implemented rules are heuristic, and the core proofs are in the supplement. The hub definition is built on weighted degree, so part of the spectral connection is definitional, but the nontrivial concentration claim in Theorem 1 is a real result if the proof is correct.

major comments (3)
  1. [Theorem 2; Algorithm 1; Sections 5.2, 6] The central recovery guarantee is existential in the tuning parameters, not a guarantee for the procedure as run. Theorem 2 asserts that there exist κ,ρ>0 such that the data-driven estimator equals H and the over-estimated estimator contains H with high probability, but Algorithm 1 requires κ as an input and gives no admissible interval; Section 5.2 selects hubs by a two-standard-deviation rule and Section 6 uses κ=2ŝ/p, and neither rule is connected to the constants or rates in Theorem 2. The 1.5 multiplier in the data-driven ŝ rule is likewise heuristic. This is load-bearing because, for β<1, the finite-sample gap between min_{h∈H}ω_h and max_{k∉H}ω_k can be small, so a mis-calibrated threshold can fall on the wrong side. The paper should either prove high-probability recovery for a data-dependent threshold (or give an explicit valid range for κ and ρ), or explicitly frame Theorem 2 as an oracle-threshold result and provide a separate analysis or calibration for the thresholds actually used.
  2. [Assumption 3; Eq. (6); Theorem 1] Assumption 3 and Eq. (6) are assumed rather than derived from the hub definition. The abstract claims that hub existence implies a low effective rank structure, but Theorem 1 requires the influential signal condition (λ_{s+1}/λ_1)^2 ≤ c min_{h∈H} ω_h, which is not shown to follow from Assumptions 1–2. The requirement of a unique spike index s with bounded eigenvalue ratios λ_1/λ_s and λ_{s+1}/λ_p is also restrictive. Please state checkable sufficient conditions under which Eq. (6) holds for a (τ_p,c)-hub set, and verify that the simulation design in Section 5 satisfies them; otherwise the theory covers a narrower class of problems than the narrative suggests.
  3. [Sections 2–4; Supplementary Materials] The proofs of Proposition 1, Theorem 1, and Theorem 2 are referenced only to a supplementary file that is not included in the arXiv submission. Because the main claims rest on these proofs and because Theorem 1 uses Eq. (6) in a way that is not derived in the main text, the supplement should be included in the revision so that the proof logic can be checked, particularly the steps connecting the hub definition to eigenvector concentration.
minor comments (4)
  1. [Algorithm 1; Section 3] The eigenvector indexing is confusing: v_p corresponds to the largest covariance eigenvalue, while the influence scores sum v_{ik}^2 over i=1..ŝ, i.e., the smallest-covariance eigenvectors. Please add a sentence in or after Algorithm 1 so that readers do not invert the order.
  2. [Section 2.2] The statement that exp{2·I(X_k; X_{P\{k}})} is asymptotically equivalent to α_k is made without stating the exact conditions or giving a precise reference to the supplement; please state the result or point explicitly to the supplementary section.
  3. [Section 5.2] The comparison uses different connectivity measures for different methods (discrete degree for GLASSO/HGL/HWGL, weighted degree for raw inverse correlation, and influence scores for IPC-HD), so the common two-standard-deviation threshold is not strictly comparable across methods; please discuss this limitation.
  4. [Section 5.3; Section 6] There is a double negative in the sentence 'except in the case that breaks the block-diagonal assumption T <200 is not satisfied,' and the notation for p_H/p_NH is inconsistent (ph vs. p_H). Also, in Section 6 the text says δ(3)>>1.5·max δ(i) but the figure does not show the actual ratio values; please report them.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: hub-to-spectral results are derived from explicit assumptions and no self-citation chain is load-bearing.

full rationale

The derivation chain is not circular. Hubs are defined through the weighted column norm alpha_k = ||Theta_.k||^2 (Eq. 1), and influence through diagonal entries of the projection P_s (Eq. 4). Proposition 1 and Theorem 1 then prove, rather than assume, that under Assumption 1 and the additional influential-signal condition (6) the hub set is recovered as an influential set; condition (6) is an explicit separation/approximation assumption, not the conclusion, and the proof steps are nontrivial matrix-perturbation arguments. No parameter is fitted to data and then reported as a prediction: the threshold kappa and regularizer rho in Theorem 2 are existential, and the paper's implemented heuristics (2-SD rule, kappa = 2*hat s/p) are not used as inputs to the theorem; this is a tuning gap without being circular. The paper also does not rely on self-citations: the covariance-concentration and masked-estimator results it invokes are external works. For these reasons there is no step in which a 'prediction' reduces by construction to an input or to a self-citation.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The method does not invent new physical entities. Its theoretical contribution is purchased with a specific definition of hubs (weighted column norm), a growing-separation assumption, a unique-spike assumption, and strong covariance-concentration conditions. The practical threshold and regularizer are free tuning parameters not covered by the existence statement of Theorem 2.

free parameters (4)
  • threshold kappa = user-specified; kappa=2s-hat/p in the Section 6 application, 2 SD rule in simulations
    Controls which variables are declared hubs; Theorem 2 only asserts existence of some kappa, so the practical rule is an unanalyzed tuning choice.
  • regularizer rho = user-specified
    Enters the eigenvalue-ratio statistic delta_rho(i) = (gamma_{p-i}+rho)/(gamma_{p-i+1}+rho); no selection criterion is given.
  • screening size T = user-specified, T<=n
    Defines the retained submatrix for the screened covariance estimator in Corollary 2; the signal block size is assumed known or chosen by the user.
  • thresholding parameter xi = user-specified
    Sets the masked or thresholded covariance estimator in Corollary 4; no data-driven choice is provided.
assumptions (6)
  • domain assumption Data are i.i.d. Gaussian, X~N_p(mu,Sigma), and Theta=Sigma^{-1} is positive definite
    Assumed at the start of Section 2; the graphical model and all spectral arguments require normality.
  • ad hoc to paper Theta contains a (tau_p,c)-hub set with |H|=r finite and tau_p=Omega(p^beta) for some beta in (0,1]
    Assumptions 1 and 2 define the exact target structure; separation growing with p is what makes recovery possible.
  • ad hoc to paper There is a unique spike index s with lambda_1/lambda_s<=c, lambda_{s+1}/lambda_p<=c, and the influential signal condition (6) holds
    Assumption 3; without a clean eigengap the leading eigenvectors need not align with the hub set.
  • domain assumption The covariance estimator satisfies the spectral-norm concentration Assumption 4 with E(n,p,Delta)<=c p^{-(1-beta)}
    This is the engine of Theorem 2 and the sample-size scalings in Corollaries 1-4; for n much smaller than p it is only available under additional structure.
  • domain assumption For the screened estimator, Theta and Sigma are approximately block-diagonal as in (8), with small off-block norms and a detectable signal block
    Used in Section 4.1 and Corollary 2 to justify reducing dimension from p to T before IPC-HD.
  • domain assumption For masked and thresholded estimators, the mask bias is small (Assumption 5) or the covariance is sparse (Assumption 6)
    Needed for Corollaries 3 and 4 to get spectral-norm concentration at near-logarithmic sample sizes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hub Detection in Gaussian Graphical Models." pith.science (2026). https://pith.science/paper/GMSQ3ZA4

@misc{pith2026250523707,
  author       = {Pith},
  title        = {Pith review of: Hub Detection in Gaussian Graphical Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GMSQ3ZA4}},
  note         = {Machine review of arXiv:2505.23707}
}
read the original abstract

Graphical models are popular tools for exploring relationships among a set of variables. The Gaussian graphical model (GGM) is an important class of graphical models, where the conditional dependence among variables is represented by nodes and edges in a graph. In many real applications, we are interested in detecting hubs in graphical models, which refer to nodes with a significant higher degree of connectivity compared to non-hub nodes. A typical strategy for hub detection consists of estimating the graphical model, and then using the estimated graph to identify hubs. Despite its simplicity, the success of this strategy relies on the accuracy of the estimated graph. In this paper, we directly target on the estimation of hubs, without the need of estimating the graph. We establish a novel connection between the presence of hubs in a graphical model, and the spectral decomposition of the underlying covariance matrix. Based on this connection, we propose the method of inverse principal components for hub detection (IPC-HD). Both consistency and convergence rates are established for IPC-HD. Our simulation study demonstrates the superior performance and fast computation of the proposed method compared to existing methods in the literature in terms of hub detection. Our application to a prostate cancer gene expression dataset detects several hub genes with close connections to tumor development.

Figures

Figures reproduced from arXiv: 2505.23707 by the authors.

Figure 1
Figure 1. Plots for the illustrative example Θ (100) 1 . Left panel: visualization of the precision matrix Θ (100) 1 . The 25-th variable is highly connected to the other variables. Center panel: plot of the ordered eigenvalues of Θ (100) 1 . The first eigenvalue is separated from the rest. Right panel: coordinates of the first eigenvector of Θ (100) 1 . The red line corresponds to the 25-th coordinate, which has a large magn… view at source ↗
Figure 2
Figure 2. Plots for the illustrative example Θ (100) 2 . Left panel: visualization of the precision matrix Θ (100) 2 . The 85-th variable has large weighted connections to the other variables. Center panel: plot of the ordered eigenvalues of Θ (100) 2 . The first eigenvalue is visibly separated from the rest. Right panel: coordinates of the first eigenvector of Θ (100) 2 . The red line highlights the 85-th coordinate, which h… view at source ↗
Figure 3
Figure 3. Comparison of the simulated true positive rate of hub detection (TPR) across [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Comparison of the simulated true positive rate of hub detection (TPR) across [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: From the plots of ordered eigenvalues and consecutive eigenvalue ratios, we [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]
Figure 6
Figure 6. Figure 6: Outputs of the IPC-HD and 2-step weighted GLASSO methods. Left panel: [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 56 canonical work pages

  1. [1]

    Abasi, M., Bazi, Z., Mohammadi-Yeganeh, S., Soleimani, M., Haghpanah, V., Zargami, N., and Ghanbarian, H. (2016). 7SK small nuclear rna transcription level down-regulates in human tumors and stem cells.Medical Oncology, 33(11):128/1–5

    work page 2016

  2. [2]

    Abbe, E. (2017). Community detection and stochastic block models: recent develop- ments. The Journal of Machine Learning Research, 18(1):6446–6531

    work page 2017

  3. [3]

    and Wit, E

    Abegaz, F. and Wit, E. (2013). Sparse time series chain graphical models for recon- structing genetic networks.Biostatistics, 14(3):586–599

    work page 2013

  4. [4]

    and Silverstein, J

    Baik, J. and Silverstein, J. W. (2006). Eigenvalues of large sample covariance matrices of spiked population models.Journal of Multivariate Analysis, 97(6):1382–1408. 20

    work page 2006

  5. [5]

    Bi, D., Ning, H., Liu, S., Que, X., and Ding, K. (2015). Gene expression patterns com- bined with network analysis identify hub genes associated with bladder cancer.Compu- tational Biology and Chemistry, 56(1):71–83

    work page 2015

  6. [6]

    Bickel, P. J. and Levina, E. (2008a). Covariance regularization by thresholding.The Annals of Statistics, 36(6):2577–2604

    work page 2008

  7. [7]

    Regularizedestimationoflargecovariancematrices

    Bickel, P.J.andLevina, E.(2008b). Regularizedestimationoflargecovariancematrices. The Annals of Statistics, 36(1):199–227

    work page 2008

  8. [8]

    Cai, T., Liu, W., and Luo, X. (2011). A constrained l-1 minimization approach to sparse precision matrix estimation. Journal of the American Statistical Association, 106(494):594–607

    work page 2011

Show all 56 references
  1. [9]

    T., Han, X., and Pan, G

    Cai, T. T., Han, X., and Pan, G. (2020). Limiting laws for divergent spiked eigenvalues and largest nonspiked eigenvalue of sample covariance matrices.The Annals of Statistics, 48(3):1255–1280

    2020

  2. [10]

    T., Liu, W., and Zhou, H

    Cai, T. T., Liu, W., and Zhou, H. H. (2016). Estimating sparse precision matrix: Op- timal rates of convergence and adaptive estimation.The Annals of Statistics, 44(2):455– 488

    2016

  3. [11]

    T., Zhang, C.-H., and Zhou, H

    Cai, T. T., Zhang, C.-H., and Zhou, H. H. (2010). Optimal rates of convergence for covariance matrix estimation.The Annals of Statistics, 38(4):2118–2144

    2010

  4. [12]

    M., Izumi, K., Zheng, Y., Gordetsky, J., Yao, J

    Canacci, A. M., Izumi, K., Zheng, Y., Gordetsky, J., Yao, J. L., and Miyamoto, H. (2011). Expression of semenogelins I and II and its prognostic significance in human prostate cancer. The Prostate, 71(10):1108–1114

    2011

  5. [13]

    Y., Gittens, A., and Tropp, J

    Chen, R. Y., Gittens, A., and Tropp, J. A. (2012a). The masked sample covariance es- timator: an analysis using matrix concentration inequalities.Information and Inference: A Journal of the IMA, 1(1):2–20

    2012

  6. [14]

    Y., Gittens, A., and Tropp, J

    Chen, R. Y., Gittens, A., and Tropp, J. A. (2012b).The masked sample covariance estimator: An analysis via the matrix Laplace transform. California Institute of Tech- nology, Pasadena, CA. ACM Report

    2012

  7. [15]

    D., Li, H., and Yang, Y

    Chen, X., Lee, J. D., Li, H., and Yang, Y. (2022). Distributed estimation for principal componentanalysis: Anenlargedeigenspaceanalysis. Journal of the American Statistical Association, 117(540):1775–1786

    2022

  8. [16]

    Chen, Y., Chi, Y., Fan, J., and Ma, C. (2021). Spectral methods for data science: A statistical perspective. Foundations and Trends in Machine Learning, 14(5):566–806

    2021

  9. [17]

    Danaher, P., Wang, P., and Witten, D. M. (2014). The joint graphical lasso for inverse covariance estimation across multiple classes. Journal of the Royal Statistical Society Series B: Statistical Methodology, 76(2):373–397

    2014

  10. [18]

    El Karoui, N. et al. (2008). Operator norm consistent estimation of large-dimensional sparse covariance matrices.The Annals of Statistics, 36(6):2717–2756. 21

    2008

  11. [19]

    D., Hellyer, P

    Fagerholm, E. D., Hellyer, P. J., Scott, G., Leech, R., and Sharp, D. J. (2015). Discon- nection of network hubs and cognitive impairment after traumatic brain injury.Brain, 138(6):1696–1709

    2015

  12. [20]

    Fan, J., Liu, H., and Wang, W. (2018). Large covariance estimation through elliptical factor models. The Annals of Statistics, 46(4):1383–1414

    2018

  13. [21]

    Fan, J., Wang, K., Zhong, Y., and Zhu, Z. (2021). Robust high dimensional factor models with applications to statistical machine learning.Statistical Science, 36(2):303– 327

    2021

  14. [22]

    Fan, X., Wang, L., Guo, Y., Xiong, X., Zhu, L., and Fang, K. (2016). Inhibition of prostate cancer growth using doxorubicin assisted by ultrasound-targeted nanobubble destruction. International Journal of Nanomedicine, 11(1):3585–3596

    2016

  15. [23]

    Feng, Q., Hannig, J., and Marron, J. S. (2016). A note on automatic data transfor- mation. Stat, 5(1):82–87

    2016

  16. [24]

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

    2008

  17. [25]

    and Bengtsson, T

    Furrer, R. and Bengtsson, T. (2007). Estimation of high-dimensional prior and pos- terior covariance matrices in kalman filter variants.Journal of Multivariate Analysis, 98(2):227–255

    2007

  18. [26]

    Q., Wu, Y., and Xu, H

    Gao, X., Pu, D. Q., Wu, Y., and Xu, H. (2012). Tuning parameter selection for pe- nalized likelihood estimation of Gaussian graphical model.Statistica Sinica, 22(3):1123– 1146

    2012

  19. [27]

    and Rajaratnam, B

    Hero, A. and Rajaratnam, B. (2012). Hub discovery in partial correlation graphs. IEEE Transactions on Information Theory, 58(9):6064–6078

    2012

  20. [28]

    Hotelling, H. (1933). Analysis of a complex of statistical variables into principal com- ponents. Journal of Educational Psychology, 24(6):417–441

    1933

  21. [29]

    Jia, B., Xu, S., Xiao, G., Lamba, V., and Liang, F. (2017). Learning gene regulatory networks from next generation sequencing data.Biometrics, 73(4):1221–1230

    2017

  22. [30]

    I., Ghahramani, Z., Jaakkola, T

    Jordan, M. I., Ghahramani, Z., Jaakkola, T. S., and Saul, L. K. (1999). An introduc- tion to variational methods for graphical models.Machine Learning, 37(2):183–233

    1999

  23. [31]

    and Lounici, K

    Koltchinskii, V. and Lounici, K. (2017). Concentration inequalities and moment bounds for sample covariance operators.Bernoulli, 23(1):110–133

    2017

  24. [32]

    Kuang, Z., Geng, S., and Page, D. (2017). A screening rule for l1-regularized ising model estimation. In Guyon, I., Luxburg, U. V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., and Garnett, R., editors,Advances in Neural Information Processing Systems, volume 30. Cur...

    2017

  25. [33]

    Lauritzen, S. L. (1996).Graphical models, volume 17. Clarendon Press. 22

    1996

  26. [34]

    and Vershynin, R

    Levina, E. and Vershynin, R. (2012). Partial estimation of covariance matrices.Prob- ability Theory and Related Fields, 153(3):405–419

    2012

  27. [35]

    Liang, F., Song, Q., and Qiu, P. (2015). An equivalent measure of partial correlation coefficients for high-dimensional gaussian graphical models. Journal of the American Statistical Association, 110(511):1248–1265

    2015

  28. [36]

    Liang, F., Xue, J., and Jia, B. (2022). Markov neighborhood regression for high- dimensional inference. Journal of the American Statistical Association, 117(539):1200– 1214

    2022

  29. [37]

    Liu, H., Han, F., Yuan, M., Lafferty, J., and Wasserman, L. (2012). High-dimensional semiparametric Gaussian copula graphical models.The Annals of Statistics, 40(4):2293– 2326

    2012

  30. [38]

    Luo, S., Song, R., and Witten, D. (2014). Sure screening for gaussian graphical models. arXiv preprint arXiv:1407.7819

    2014 arXiv

  31. [39]

    and Hastie, T

    Mazumder, R. and Hastie, T. (2012). Exact covariance thresholding into connected components for large-scale graphical lasso.The Journal of Machine Learning Research, 13(1):781–794

    2012

  32. [40]

    McGillivray, A., Khalili, A., and Stephens, D. A. (2020). Estimating sparse networks with hubs.Journal of Multivariate Analysis, 179(1):104655/1–20

    2020

  33. [41]

    and Bühlmann, P

    Meinshausen, N. and Bühlmann, P. (2006). High-dimensional graphs and variable selection with the lasso.The Annals of Statistics, 34(3):1436–1462

    2006

  34. [42]

    Mohan, K., London, P., Fazel, M., Witten, D., and Lee, S.-I. (2014). Node-based learn- ing of multiple gaussian graphical models.The Journal of Machine Learning Research, 15(1):445–488

    2014

  35. [43]

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

    2007

  36. [44]

    and Zhou, X.-H

    Qiu, Y. and Zhou, X.-H. (2020). Estimating c-level partial correlation graphs with application to brain imaging.Biostatistics, 21(4):641–658

    2020

  37. [45]

    J., and Lafferty, J

    Ravikumar, P., Wainwright, M. J., and Lafferty, J. D. (2010). High-dimensional Ising model selection using l1-regularized logistic regression.The Annals of Statistics, 38(3):1287–1319

    2010

  38. [46]

    Ren, Z., Sun, T., Zhang, C.-H., and Zhou, H. H. (2015). Asymptotic normality and optimalities in estimation of large gaussian graphical models.The Annals of Statistics, 43(3):991–1026

    2015

  39. [47]

    V., Alves, L

    Ribeiro, H. V., Alves, L. G. A., Martins, A. F., Lenzi, E. K., and Perc, M. (2018). The dynamical structure of political corruption networks.Journal of Complex Networks, 6(6):989–1003. 23

    2018

  40. [48]

    Shen, D., Shen, H., Zhu, H., and Marron, J. (2016). The statistics and mathematics of high dimension low sample size asymptotics.Statistica Sinica, 26(4):1747–1770

    2016

  41. [49]

    M., London, P., Mohan, K., Lee, S.-I., Fazel, M., and Witten, D

    Tan, K. M., London, P., Mohan, K., Lee, S.-I., Fazel, M., and Witten, D. (2014). Learning graphical models with hubs. The Journal of Machine Learning Research, 15(95):3297–3331

    2014

  42. [50]

    and Ravikumar, P

    Tandon, R. and Ravikumar, P. (2014). Learning graphs with a few hubs.Proceedings of Machine Learning Research, 32(1):602–610

    2014

  43. [51]

    Wainwright, M. J. (2019).High-dimensional statistics: A non-asymptotic viewpoint, volume 48. Cambridge University Press

    2019

  44. [52]

    and Fan, J

    Wang, W. and Fan, J. (2017). Asymptotics of empirical eigenstructure for high di- mensional spiked covariance.The Annals of Statistics, 45(3):1342–1374

    2017

  45. [53]

    M., Friedman, J

    Witten, D. M., Friedman, J. H., and Simon, N. (2011). New insights and faster computations for the graphical lasso.Journal of Computational and Graphical Statistics, 20(4):892–900

    2011

  46. [54]

    I., and Liu, Z

    Yang, E., Ravikumar, P., Allen, G. I., and Liu, Z. (2013). On Poisson graphical models. Advances in Neural Information Processing Systems, 26(1):1718–1726

    2013

  47. [55]

    and Lin, Y

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

    2007

  48. [56]

    and Duan, Z.-H

    Zhao, H. and Duan, Z.-H. (2019). Cancer genetic network inference using Gaussian graphical models. Bioinformatics and Biology Insights, 13(1):1–9. 24

    2019

Pith tools

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