High-Dimensional Gaussian Graphical Model Neighborhood Selection with the Lasso

⚠️ Curated Limitations & Negative Results

• Neighborhood Stability Requirement and Pseudo-Likelihood Asymmetry: Exact asymptotic neighborhood recovery (Theorem 1 and Corollary 1) requires the Neighborhood Stability condition |S_a(b)| < delta < 1 for all b not in ne_a (Assumption 6, Equation 5; Proposition 3 proves inconsistency when |S_a(b)| > 1), and because nodewise Lasso regressions are solved separately for each node a in Gamma(n), finite-sample neighborhood estimates can be asymmetric (a in ne_hat_b^lambda != b in ne_hat_a^lambda) prior to AND/OR edge aggregation (Equations 7-8) and do not directly produce a positive-definite precision matrix estimate.
• Gaussian Conditional Independence Assumption: The exact equivalence between zero linear regression coefficient theta_b^a = 0, zero precision entry K_{ab} = 0, and conditional independence X_a _|_ X_b | X_{Gamma(n)\{a,b}} relies on multivariate normality X ~ N(mu, Sigma) (Section 2, pp. 3-4); for non-Gaussian distributions, theta_b^a = 0 characterizes zero partial correlation (linear projection) rather than general conditional independence.
• Primary-Source Typographical Notes in Meinshausen & Buhlmann (2006, arXiv:math/0608017v1): Verified five self-correcting typographical details during full-text inspection and numerical verification: (1) Section 2.3 (p. 7, line 23) and Appendix (p. 19, Proof of Lemma A.3, line 16) write pi_{ab} = theta_b^a / (Var(X_a | X_{Gamma(n)\{a}}) Var(X_b | X_{Gamma(n)\{b}}))^{1/2}, whereas since theta_b^a = -K_{ab}/K_{aa} and Var(X_a | X_{Gamma(n)\{a}}) = 1/K_{aa} (p. 19, line 12), the exact partial correlation is pi_{ab} = -K_{ab}/sqrt(K_{aa} K_{bb}) = theta_b^a sqrt(Var(X_b | X_{Gamma(n)\{b}}) / Var(X_a | X_{Gamma(n)\{a}})) = sign(theta_b^a) sqrt(theta_b^a theta_a^b) (under Assumption 2 where conditional variances are bounded in [v^2, 1], both ratios are bounded away from 0 and infty, leaving Assumption 5 and Lemma A.3 unchanged); (2) Section 2.3 (p. 7, lines 11-14) introduces theta^{b, A} and then writes ||theta^{a, ne_b \ {a}}||_1 <= vartheta in Assumption 4 (used symmetrically as ||theta^{b, ne_a \ {b}}||_1 <= vartheta in Equation A.7 on p. 17); (3) Appendix (p. 22, Proof of Proposition 1, line 17) writes 0 < delta_theta < 1/2(s - tau_max) for 0 < delta_theta < (1/2)(s - tau_max), where L_{delta_theta} - L_0 = -2(s - tau_max) delta_theta + 2 delta_theta^2 < 0 holds on (0, s - tau_max) and achieves its minimum at delta_theta* = (1/2)(s - tau_max); (4) Appendix (p. 24, Proof of Theorem 1, line 18) writes <X_a - X theta_hat_k^{a, ne_a, lambda}, V_b> with an extraneous coordinate subscript k on the vector theta_hat^{a, ne_a, lambda}; and (5) Appendix (p. 27, Proof of Theorem 3, line 14) defines sigma_hat_a^2 = 4 n^{-1} ||X_a||_2^2 inside the proof (absorbing the factor 4 from R^2 <= 4 n^{-1} ||X_a||_2^2), whereas Equation (9) on p. 12 defines sigma_hat_a^2 = n^{-1} <X_a, X_a> and writes the factor 2 explicitly in lambda(alpha) = (2 sigma_hat_a / sqrt(n)) Phi_tilde^{-1}(alpha / (2 p(n)^2)).
• All 28 pages of Meinshausen & Buhlmann (2006, Annals of Statistics 34(3):1436-1462, arXiv:math/0608017v1) were read. Constructed numerical checks in research/discovery/high-dimensional-gaussian-graphical-model-lasso-neighborhood-selection-boundaries.py verify (1) exact Gaussian precision-matrix neighborhood identity theta_b^a = -K_{ab}/K_{aa} (error 1.11e-16), conditional variance 1/K_{aa} (error 1.11e-16), and partial correlation identity (error 1.67e-16); (2) Proposition 1 prediction-oracle inconsistency risk reduction L_{delta*} - L_0 = -0.0162 < 0 at delta* = 0.5(s - tau_max); (3) Neighborhood Stability S_0(3) = 0.70 < 1 (exact zero false positive) vs unstable S_0(3) = 1.20 > 1 (forced false positive theta_hat_3 = 0.0282 > 0, Proposition 3) and tree stability S_tree = 0.75 < 1; and (4) AND/OR edge aggregation and Theorem 3 finite-sample connectivity-component FWER control (empirical FWER 0.025 <= alpha = 0.10).