{"id":"7f89a72c-1422-4d15-9e72-4b087d249fd4","arxiv_id":"2501.05308","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":1.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"The proposed EAGL estimator adds a log-determinant penalty that duplicates the likelihood's own log-det term, making it equivalent to a scaled Graphical Lasso.","lead":"This paper introduces an 'entropy adjusted' Graphical Lasso that adds a log-determinant penalty to precision matrix estimation. Because that penalty is algebraically the same as a term already in the likelihood, the proposed estimator is just a rescaled Graphical Lasso.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'entropy' penalty in Eq. (7) is just the log-likelihood term: substituting log det(Omega^{-1}) = -log det(Omega) reduces EAGL to GLasso with a rescaled penalty, so the central claim of a new estimator collapses.","rationale":"The reader's weakest assumption is exactly the load-bearing flaw: the log-determinant term in Eq. (7) is not independent of the Gaussian log-likelihood. The algebraic identity log det(Omega^{-1}) = -log det(Omega) turns the EAGL objective into a scaled GLasso objective, so the argmin is identical to GLasso with a recalibrated penalty. This invalidates the claim of a new entropy-adjusted estimator and the interpretation of the numerical results as evidence for entropy adjustment. The paper's own Appendix D derives Eq. (16), which exhibits the collapse, yet treats it as a technical reformulation rather than a demonstration that the method reduces to GLasso. No other concern is more load-bearing: the fixed-p asymptotic rate, lack of code, and table inconsistencies are secondary. The reader's REJECT verdict remains appropriate.","tokens_in":28149,"tokens_out":2758,"duration_ms":29116,"concrete_test":"Implement Eq. (7) and Eq. (2) directly. For any fixed S, gamma, and alpha in (0,1), compute the EAGL solution and the GLasso solution with penalty lambda = gamma*alpha / (1 + gamma*(1-alpha)). If the two estimated precision matrices are identical (up to numerical tolerance) for every S, gamma, and alpha, the concern is confirmed. A stronger check: rerun the Model 1 simulation of Table 5 using glassoFast for both EAGL and GLasso with lambda as above, and verify that the EAGL column matches the GLasso column exactly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that log det(Omega^{-1}) in Eq. (7) is an independent entropy-adjustment penalty. But log det(Omega^{-1}) = -log det(Omega), which is already the log-likelihood term. Substituting into Eq. (7) gives the objective -(1 + gamma(1-alpha)) log det(Omega) + trace(Omega S) + gamma alpha ||Omega||_1. Since 1 + gamma(1-alpha) > 0, dividing by this constant leaves the argmin unchanged, yielding exactly the GLasso objective (2) with penalty gamma alpha / (1 + gamma(1-alpha)). Thus EAGL is not a new estimator; it is GLasso with a reparameterized tuning parameter. Appendix D confirms this by rewriting the problem as Eq. (16), but the paper does not acknowledge that this is simply GLasso. Consequently, the empirical improvements, eigenvalue behavior, and the claimed 'entropy adjustment' are all artifacts of using a smaller effective Lasso penalty, not of any additional penalty term. The methodological novelty and the entropy interpretation are unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":28289,"tokens_out":3535,"duration_ms":36888,"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":[{"comment":"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.","section":"Section 2.2, Eq. (7)"},{"comment":"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.","section":"Section 3.5 and Table 2"},{"comment":"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.","section":"Appendix D, Eqs. (15)-(20)"}],"minor_comments":[{"comment":"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.","section":"Section 2.1, after Eq. (5)"},{"comment":"There is a typo: 'T-Gridg' should be 'T-GRidge', and 'missclassification' should be 'misclassification' throughout the section and table caption.","section":"Section 4.1 and Table 3"},{"comment":"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.","section":"Section 3.5"},{"comment":"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.","section":"Section 2.2, Eq. (7)"},{"comment":"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.","section":"Section 3.2, Model 4"}],"recommendation":"reject","confidential_remarks":"The core problem is mathematical rather than empirical: the proposed penalty term is a scalar multiple of the likelihood's log-determinant term, so EAGL is a reparameterization of GLasso. This cannot be repaired by additional experiments or by restating the asymptotic argument, because the claimed new estimator does not exist as such. The paper would need a genuinely different penalty or a different interpretation to be viable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the entropy adjustment is not an adjustment. Substitute log det(Omega^{-1}) = -log det(Omega) into (7) and the objective becomes -(1+gamma(1-alpha))log det(Omega) + trace(Omega S) + gamma alpha ||Omega||_1. Divide by the positive constant and you get the GLasso objective with sample covariance S/(1+gamma(1-alpha)) and penalty gamma alpha/(1+gamma(1-alpha)). That is a reparameterized GLasso, not a new estimator. The paper's own Appendix D rewrites the objective as (16) but never notes the equivalence. So the central claim of a distinct entropy-adjusted estimator is unsupported.\n\nCredit where due: the numerical study is thorough—seven models, a broad set of losses and graph metrics, two real-data applications, and honest limitations. The simulations are reported in enough detail to be reproducible in principle, and the computational-time comparison is useful. The paper is clearly written and the entropy derivation in Appendix A is correct.\n\nSoft spots. The reader's exact claim that EAGL equals (1+gamma(1-alpha)) times the GLasso solution is slightly off; the correct statement is GLasso on a scaled covariance matrix with a scaled penalty. The conclusion is the same, but the precise equivalence matters. Section 3.5 says GLasso wins uMCC on models 4,5,6 while Table 2 marks EAGL as best on all seven; the tables actually show EAGL slightly ahead on those models, so the text is wrong. The asymptotic proof assumes p fixed, which sits oddly with a high-dimensional claim; the rate (p+s)log p/n is standard but not the paper's engine. No code or data are shipped, though the R packages are named.\n\nThe empirical gains, the eigenvalue plots, and the 'entropy' interpretation are all artifacts of using a smaller effective penalty on a rescaled likelihood. One could reframe this as a short note on the equivalence between entropy-adjusted and standard GLasso, but the current framing does not hold. I would not send it to a referee; the algebra is elementary and the novelty claim fails immediately. A serious editor should desk-reject unless the authors want to rewrite it as a reparameterization note.","headline":"EAGL is GLasso with a rescaled covariance and penalty, so the entropy-adjustment claim does not survive the paper's own algebra.","tokens_in":28898,"tokens_out":6705,"would_cite":false,"duration_ms":59051,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H12","62J07","62F12"],"pacs":[],"model":"deepseek-v4-flash","headline":"Lower sparse-matrix errors with entropy-adjusted Graphical Lasso.","keywords":["Elastic Net","Entropy adjustment","Gaussian Graphical Models","Lasso","Maximum Likelihood Estimation","precision matrix estimation","Graphical Lasso","log-determinant penalty"],"falsifier":"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.","tokens_in":27847,"feed_emoji":"📉","tokens_out":16295,"duration_ms":137751,"temperature":0.7,"pith_summary":"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.","feed_headline":"Lower sparse-matrix errors with entropy-adjusted Graphical Lasso","feed_subtitle":"The new estimator beats GLasso, elastic-net, ridge, and SCAD on most losses and portfolio risk.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Defines the convex optimization formulation for sparse inverse covariance estimation that EAGL extends.","marker":"Banerjee et al., 2006"},{"why":"Introduces the GLasso estimator as penalized maximum likelihood for Gaussian graphical models.","marker":"Yuan and Lin, 2007"},{"why":"Provides the fast GLasso algorithm and the eigenvalue-shrinkage observation EAGL is designed to counter.","marker":"Friedman et al., 2008"},{"why":"Proposes Graphical Elastic Net, whose convex Lasso-Ridge combination inspires EAGL's penalty structure.","marker":"Bernardini et al., 2022"},{"why":"Introduces the targeted elastic-net estimator and supplies simulation models and CV tuning used in the numerical study.","marker":"Kovács et al., 2021"},{"why":"Defines targeted Graphical Ridge estimators that serve as comparison methods.","marker":"van Wieringen and Peeters, 2016"},{"why":"Introduces ROPE and the scalar target matrix choice used for ridge-based comparators.","marker":"Kuismin et al., 2017"},{"why":"Provides the proof template for EAGL's Frobenius-norm convergence rate.","marker":"Rothman et al., 2008"},{"why":"Yields the elementwise bound on sample covariance deviation that drives the rate proof.","marker":"Bickel and Levina, 2008"},{"why":"Supplies the shrinkage covariance estimator benchmark in the portfolio application.","marker":"Ledoit and Wolf, 2004"}],"fun_headline_variants":["Entropy-adjusted Graphical Lasso beats GLasso on losses","EAGL: Graphical Lasso with entropy penalty for precision","Uncertainty-aware Graphical Lasso improves matrix estimates","Entropy-adjusted GLasso wins on portfolio risk and losses","New precision estimator: EAGL outperforms GLasso in tests"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Entropy-adjusted Graphical Lasso beats GLasso on losses","EAGL: Graphical Lasso with entropy penalty for precision","Uncertainty-aware Graphical Lasso improves matrix estimates","Entropy-adjusted GLasso wins on portfolio risk and losses","New precision estimator: EAGL outperforms GLasso in tests"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000152,"raw_usage":{"total_tokens":1200,"prompt_tokens":939,"completion_tokens":261,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":178}},"tokens_in":555,"tokens_out":261,"duration_ms":3052,"temperature":1.0,"reasoning_tokens":178,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:13:50.160983+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the convex optimization formulation for sparse inverse covariance estimation that EAGL extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Proposes Graphical Elastic Net, whose convex Lasso-Ridge combination inspires EAGL's penalty structure."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines targeted Graphical Ridge estimators that serve as comparison methods."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces ROPE and the scalar target matrix choice used for ridge-based comparators."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the proof template for EAGL's Frobenius-norm convergence rate."},{"cited_title":"and Levina, E","cited_arxiv_id":null,"evidence_quote":"Yields the elementwise bound on sample covariance deviation that drives the rate proof."}],"review_version":1}