{"id":"2fd6ba77-bf49-476d-96bb-74615db2ff3e","arxiv_id":"2411.12479","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A graph-based square-root sparse regression estimator is proposed with sigma-free tuning, finite-sample bounds, asymptotic normality, and selection consistency.","lead":"This paper proposes a new sparse regression method that combines square-root loss with graph-structured penalties, so tuning no longer needs the unknown noise level. Simulation and real-data tests show it often beats lasso, square-root lasso, and graph-based square-loss rivals.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Assumption 2.2 is load-bearing but is violated by the paper's own Example 3, so Theorem 2.1 does not cover the advertised regime.","rationale":"The reader's weakest-assumption analysis identified Assumption 2.2, and my reading agrees: it is the one assumption that, if removed, makes the main finite-sample guarantee (Theorem 2.1) and the high-dimensional selection guarantee (Theorem 2.3) unsupported. The concern is sharpened by the observation that the paper's own Example 3 violates the assumption: the AR(1)-style graph gives node 15 (active) an edge to node 16 (inactive), so the flagship simulation lives outside the theorem's hypotheses. This is not a fatal flaw of the estimator—the numerical results may still be valid—but it is a real gap between theory and the advertised regime. Other issues raised by the reader, such as the non-Gaussian noise claims in the abstract/Remark 2.4 and the suspicious equivalence in Lemma 2.2(c), are secondary: they are scope or periphery concerns, whereas Assumption 2.2 sits directly inside the proof of the paper's central theoretical result. The conditional verdict remains appropriate: the paper should either prove the bounds under a weaker graph-support alignment condition, move the sensitivity analysis into the main text, or explicitly state that the theorem requires the support to be aligned with the graph.","tokens_in":30594,"tokens_out":9282,"duration_ms":97471,"concrete_test":"Re-examine the proof of Theorem 2.1 at the point where Assumption 2.2 is invoked (supplementary Section B, Basic Inequality). Construct the minimal violating configuration: n=200, p=2, X_1 and X_2 with correlation ρ=0.9, β*=(1,0), graph edge (1,2), Gaussian noise σ=1, λ chosen as in Lemma 2.3. Run GSRE and record false-positive probability and ∥X(β̂−β*)∥/√n over 1000 replications. If the Theorem 2.1-type bound is systematically exceeded (or β̂_2≠0 with non-negligible probability) while the same setup with the edge removed satisfies the bound, then Assumption 2.2 is doing essential work and the theorem cannot be extended to the advertised real-data regime without an explicit correction or a much weaker assumption.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Assumption 2.2 (Section 2.4) is the load-bearing condition for Theorem 2.1 and Theorem 2.3: every graph neighborhood of an active predictor must lie inside the true support. If an active node is adjacent to an inactive node, the node-wise penalty in (1) assigns mass to the inactive coordinate through the active node's group V(i), so the basic inequality in the proof can no longer force that coordinate to zero; the displayed bounds in Theorem 2.1 are not derived (and need not hold). The paper's own Example 3 violates the assumption: with AR(1) design and β* supported on {1,...,15}, node 15 is adjacent to node 16, which is inactive, so Assumption 2.2 fails for i=15; yet GSRE-o/GSRE is reported as best. Hence the main theoretical guarantee does not cover the paper's flagship simulation, and the claimed 'reasonable' justification in Remark 2.5 is not supported by an experiment in the main text (sensitivity is relegated to the supplement). This is a correctness/coverage gap in the central claim, not a disagreement with the estimator's empirical usefulness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes graph-based square-root estimation (GSRE), a sparse linear regression estimator that combines the square-root loss with a node-wise graphical penalty. The authors show that the method reduces to square-root lasso and group square-root lasso for special graphs, provide finite-sample oracle-type bounds, asymptotic normality, and model-selection consistency under a σ-free tuning parameter, and present an ADMM algorithm. Numerical experiments on simulated and real data compare GSRE with lasso, adaptive lasso, elastic net, square-root lasso, and graph-based least-squares regression. The main theoretical contribution is that the tuning parameter does not depend on the unknown noise level σ, and the main practical claim is improved estimation, prediction, and variable selection under heavy-tailed and non-Gaussian noise.","tokens_in":30816,"tokens_out":6006,"duration_ms":59739,"significance":"If the theoretical claims are correct, the paper makes a useful extension of square-root lasso to graph-structured penalties, and the σ-free tuning construction is a genuine practical advantage in high-dimensional settings. The paper also ships code and reports extensive simulations across several noise distributions and predictor graphs, which strengthens the empirical contribution. However, the two issues raised below — the violation of the key graph-support assumption by the paper's own flagship example, and an incorrect equivalence claim for complete graphs — need to be addressed before the theoretical and generality claims can be taken at face value.","major_comments":[{"comment":"Example 3 of Section 4.1 violates Assumption 2.2, which is required by Theorem 2.1 and Theorem 2.3. In Example 3 the predictors follow an AR(1) model with Σ_{ij} = 0.5^{|i-j|}, so the graph is a path with edges between consecutive nodes. With β* supported on {1,...,15}, node 15 has neighborhood {14,15,16}, and node 16 is inactive. Hence Assumption 2.2 fails for i=15, and the finite-sample bounds and model-selection consistency results do not cover this simulation, which is one of the paper's main advertised successes. The main text only states that a sensitivity analysis appears in the supplementary materials, which does not repair the coverage gap. Please either extend the theory to allow active nodes adjacent to inactive nodes, or explicitly state that the theoretical guarantees apply only when Assumption 2.2 holds and classify Example 3 as outside the theoretical scope.","section":"Section 2.4 (Assumption 2.2) and Section 4.1 (Example 3)"},{"comment":"The claimed equivalence in Lemma 2.2(c) is not correct. For a complete graph, every neighborhood N_i equals [p], so the graph norm ∥β∥_{G,τ} equals (min_i τ_i) ∥β∥_2: putting all mass in the node with the smallest τ_i gives the upper bound, and the triangle inequality gives the lower bound. Since ∥β∥_{G,τ} is positively homogeneous while the squared ℓ2 penalty in (7) is not, the two optimization problems cannot share the same solutions. This invalidates the assertion that GSRE reduces to the squared-ℓ2-penalized problem (7) when the predictor graph is complete. Please correct or remove this part, and revisit the 'generality' discussion in Remark 2.2 accordingly.","section":"Lemma 2.2(c)"}],"minor_comments":[{"comment":"The claim that 'subset recovery can be guaranteed with no noise, i.e., σ = 0' is problematic because the event A in (8) is defined through the ratio V = max_i √n ∥X_{N_i}^⊤ ε∥₂ / (τ_i ∥ε∥₂), which is undefined when ε = 0. The σ = 0 case should be treated separately or the statement qualified.","section":"Remark 2.6"},{"comment":"The augmented-Lagrangian step size is denoted by σ, which clashes with the noise-level σ used throughout the rest of the paper. Please use a different symbol, e.g., ρ, to avoid confusion.","section":"Section 3"},{"comment":"There is a typo in the definition of the Matthews correlation coefficient: 'Mattews' should be 'Matthews'. Also, the caption of Figure 3 refers to 'bodyfa2' while the text uses 'bodyfat2'.","section":"Section 4.1"},{"comment":"The notation '≲' used in Theorem 2.1 is not defined in the paper. Please specify whether it hides absolute constants, and clarify the dependence of the constants on the model parameters.","section":"Notation and Section 2.4"}],"recommendation":"major_revision","confidential_remarks":"The paper's own Example 3 violates Assumption 2.2, which is a central condition of the main theorems. This is a substantive coverage gap rather than a presentation issue; the authors should either weaken the assumption or clearly delimit the theoretical claims. The error in Lemma 2.2(c) is also a mathematical statement that should be corrected. I recommend major revision rather than rejection because the core estimator and the σ-free tuning construction appear potentially useful, and the remaining theoretical results follow standard oracle-inequality techniques."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: GSRE is a genuinely new estimator — the first square-root-loss variant of Yu and Liu's node-wise graph penalty — with sigma-free tuning, sensible theory, and strong simulations across noise types. Worth taking seriously. But the main finite-sample theorem is proven under an assumption that the paper's own flagship example violates, and Lemma 2.2(c) looks wrong.\n\nWhat's new and good: the combination is natural and the paper is the first to present it. The special-case reductions to square-root lasso and group square-root lasso are correctly derived. The sigma-free tuning bound (Lemma 2.3) is a real advantage over the original SRIG, whose tuning depends on σ. The ADMM algorithm with the Woodbury identity is standard but sound. The simulations are extensive: three predictor structures, four noise types, p=100, n=20/40/60, and GSRE generally wins on L2 error, prediction, and MCC. The two real-data examples also favor it.\n\nSoft spots, in order of seriousness. First, Assumption 2.2 — every active predictor's neighborhood must lie inside the true support — is load-bearing for Theorem 2.1 and Theorem 2.3. Example 3 violates it: with AR(1) design and β* supported on {1,...,15}, node 15 is adjacent to inactive node 16. So the main bound doesn't cover the paper's most impressive simulation. The sensitivity study is relegated to the supplement, and Remark 2.5's \"reasonable\" doesn't help. This is a coverage gap, not a fatal flaw, but it should be fixed or clearly acknowledged. Second, Lemma 2.2(c): for a complete graph, the node-wise norm is a constant times ||β||_2, not ||β||_2^2. Claiming the GSRE problem has the same nonzero solutions as a squared-L2-penalized problem is at best trivial (both are dense) and at worst false. This needs correction. Third, the non-Gaussian claim: the finite-sample bounds rely on Gaussian noise via Lemma 2.3; Remark 2.4 only sketches sub-exponential extensions. The simulations do show robustness, but the theoretical statement should be toned down or backed with a real proof.\n\nVerdict: this deserves peer review, not a desk reject. I'd send it to a capable referee with instructions to push on Assumption 2.2 and Lemma 2.2(c). The estimator itself is useful and the empirical story is convincing; the theory needs tightening.","headline":"GSRE is a genuinely new estimator with strong empirical results, but the main theorem's central assumption is violated by its own Example 3, and the complete-graph claim in Lemma 2.2(c) is dubious.","tokens_in":31329,"tokens_out":6500,"would_cite":true,"duration_ms":60185,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62J07","62F12","62J05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Graph-based square-root estimation makes sparse regression tuning independent of the error standard deviation, with finite-sample bounds, asymptotic normality, and sign consistency.","keywords":["sparse linear regression","square-root loss","graphical structure among predictors","node-wise graph penalty","sigma-free tuning","alternating direction method of multipliers","oracle property","high-dimensional statistics"],"falsifier":"Simulate a design with $p=100$, $s^*=15$, one true predictor connected by an edge to a zero predictor, fit GSRE with the correct graph, and check whether the prediction-error bound $\\|X(\\hat\\beta-\\beta^*)\\|_2/\\sqrt{n}\\lesssim \\sigma\\lambda\\sqrt{s^*}/(n\\kappa)$ holds across 50 replications; a systematic violation would show Assumption 2.2 is doing the work.","tokens_in":30409,"feed_emoji":"📊","tokens_out":7597,"duration_ms":69924,"temperature":0.7,"pith_summary":"This paper proposes the graph-based square-root estimator (GSRE) for sparse linear regression, which replaces the usual squared loss with the square-root loss $\\|y-X\\beta\\|_2/\\sqrt{n}$ and penalizes coefficients with a node-wise graph norm built from each predictor's graph neighborhood. The central aim is to show that this estimator can recover the true sparse signal with finite-sample error bounds, asymptotic normality, and model-selection consistency while choosing the tuning parameter $\\lambda$ without knowing or estimating the noise standard deviation $\\sigma$. If the claims hold, a practitioner can run sparse regression on correlated predictors under non-Gaussian noise and still get near-oracle estimation and selection. The paper also shows that square-root lasso and group square-root lasso are special cases of the model, and that a complete predictor graph links it to a squared-$\\ell_2$-penalized problem.","feed_headline":"Graph square-root estimator recovers sparse signals without sigma","feed_subtitle":"Graph structure sets the penalty, so estimation and variable selection need no noise-variance estimate.","key_machinery":"The key object is the node-wise graph norm $\\|\\beta\\|_{G,\\tau} = \\min_{\\sum_i V^{(i)}=\\beta,\\ \\mathrm{supp}(V^{(i)})\\subseteq N_i}\\sum_i \\tau_i\\|V^{(i)}\\|_2$, which lets the graph structure enter as overlapping groups instead of edge pairs. It is coupled with the square-root loss $\\|y-X\\beta\\|_2/\\sqrt{n}$, whose homogeneity makes the subgradient condition depend on the ratio $X_{N_i}^\\top(y-X\\beta)/\\|y-X\\beta\\|_2$ rather than on $\\sigma$. The ratio statistic $V=\\max_i \\sqrt{n}\\|X_{N_i}^\\top\\epsilon\\|_2/(\\tau_i\\|\\epsilon\\|_2)$ converts the choice of $\\lambda$ into a deviation bound on a maximum of Gaussian quadratic forms, removing $\\sigma$ from the tuning. The theoretical engine is then a compatibility-type condition on the design adapted to overlapping neighborhoods, plus the neighborhood-containment assumption.","core_discovery":"On a design with known predictor graph $G$ and weights $\\tau_i$, GSRE minimizes $\\|y-X\\beta\\|_2/\\sqrt{n} + (\\lambda/n)\\|\\beta\\|_{G,\\tau}$ with $\\|\\beta\\|_{G,\\tau}$ built from overlapping neighborhood groups $N_i$. The paper's main result, Theorem 2.1, states that under the assumption that each true predictor's neighborhood lies inside the true support and a compatibility condition on $X$, every optimal solution obeys $\\|X(\\hat\\beta-\\beta^*)\\|_2/\\sqrt{n} \\lesssim \\sigma\\lambda\\sqrt{s^*}/(n\\kappa)$, with corresponding bounds for the graph norm and Euclidean error, all on an event that holds with high probability. The tuning parameter is $\\sigma$-free: $\\lambda$ is set from a maximum over neighborhoods of the ratio $\\sqrt{n}\\|X_{N_i}^\\top\\epsilon\\|_2/(\\tau_i\\|\\epsilon\\|_2)$, so no error-variance estimate enters. For fixed $p$, Theorem 2.2 gives $\\sqrt{n}(\\hat\\beta_{I^*}-\\beta^*_{I^*}) \\xrightarrow{d} N(0,\\sigma^2 M_{I^*,I^*}^{-1})$, and in high dimensions Theorem 2.3 gives sign consistency under an irrepresentable-type condition.","pith_inferences":["Beyond the paper: the neighborhood-containment condition is the real design constraint; if a useful predictor is adjacent to a pure noise predictor, the node-wise penalty will tend to bring that noise variable into the support, so the method's practical safety depends on graph estimation quality.","Beyond the paper: the paper only tests sensitivity to graph misspecification in the supplement; an explicit extension would derive bounds under partial neighborhood contamination, allowing a fraction of $N_i \\cap I^*_c$ nodes.","Beyond the paper: the $\\sigma$-free tuning mechanism suggests a natural plug-in rule for adaptive variants of GSRE with nonconvex penalties, where the same ratio statistic could set local penalty scales without cross-validation.","Beyond the paper: on complete graphs the nonzero solutions coincide with a square-root-loss plus squared-$\\ell_2$ penalty, so the graph penalty is doing ridge-like shrinkage; a user should expect edge density to change selection behavior."],"forward_implications":["A user can calibrate $\\lambda$ from $n$, the graph degrees, and a confidence level $\\alpha$ alone; no preliminary estimate of the noise variance is needed.","The error bounds and selection consistency extend to non-Gaussian noise settings such as $t$, Laplace, and uniform errors, matching the experiments in the paper.","For fixed dimension, coefficient estimates on the true support are asymptotically normal with the oracle covariance matrix $\\sigma^2 M_{I^*,I^*}^{-1}$, enabling standard inferential readouts.","Under a graph-adapted irrepresentable condition, the estimator recovers the exact sign pattern with probability tending to one even when $p$ grows with $n$.","Because square-root lasso and group square-root lasso are special cases of GSRE, any algorithmic or theoretical improvement for GSRE transfers to those established estimators."],"supporting_citations":[{"why":"introduced square-root lasso and the pivotal sigma-free recovery idea that the loss function builds on.","marker":"Belloni et al., 2011"},{"why":"defined the node-wise graph norm and the SRIG model from which the penalty term is taken.","marker":"Yu and Liu, 2016"},{"why":"developed group square-root lasso, a special case of GSRE under disconnected complete subgraphs, and provided the comparison rates for the bounds in Theorem 2.1.","marker":"Bunea et al., 2013"},{"why":"supplied the restricted eigenvalue idea that Assumption 2.3 relaxes as the compatibility condition on the design.","marker":"Bickel et al., 2009"},{"why":"provided the irrepresentable condition that Assumption 2.10 adapts for the sign-consistency result of Theorem 2.3.","marker":"Zhao and Yu, 2006"},{"why":"graphical lasso estimates the sparse precision matrix used to construct the predictor graph when G is unknown.","marker":"Friedman et al., 2008"},{"why":"the HBIC criterion used in experiments to select the tuning parameter lambda from solution paths.","marker":"Wang et al., 2013"}],"fun_headline_variants":["No sigma needed: graph-based sparse regression","Graph square-root estimator: no sigma, no problem","Square-root loss + graph penalty = regression without sigma","Sigma-free sparse regression via graph square-root"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is Assumption 2.2: every predictor that is truly useful must have all of its graph neighbors also truly useful; if a signal variable is connected to a noise variable, the finite-sample bounds and selection guarantees in the paper do not apply.","fun_headline_variants_meta":{"raw":{"variants":["No sigma needed: graph-based sparse regression","Graph square-root estimator: no sigma, no problem","Square-root loss + graph penalty = regression without sigma","Sigma-free sparse regression via graph square-root"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000875,"raw_usage":{"total_tokens":3826,"prompt_tokens":1024,"completion_tokens":2802,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":640,"completion_tokens_details":{"reasoning_tokens":2743}},"tokens_in":640,"tokens_out":2802,"duration_ms":18181,"temperature":1.0,"reasoning_tokens":2743,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:28:17.959883+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a design with $p=100$, $s^*=15$, one true predictor connected by an edge to a zero predictor, fit GSRE with the correct graph, and check whether the prediction-error bound $\\|X(\\hat\\beta-\\beta^*)\\|_2/\\sqrt{n}\\lesssim \\sigma\\lambda\\sqrt{s^*}/(n\\kappa)$ holds across 50 replications; a systematic violation would show Assumption 2.2 is doing the work.","supporting_citations":[{"cited_title":"Chernozhukov, and L","cited_arxiv_id":null,"evidence_quote":"introduced square-root lasso and the pivotal sigma-free recovery idea that the loss function builds on."},{"cited_title":"Lederer, and Y","cited_arxiv_id":null,"evidence_quote":"developed group square-root lasso, a special case of GSRE under disconnected complete subgraphs, and provided the comparison rates for the bounds in Theorem 2.1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplied the restricted eigenvalue idea that Assumption 2.3 relaxes as the compatibility condition on the design."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provided the irrepresentable condition that Assumption 2.10 adapts for the sign-consistency result of Theorem 2.3."}],"review_version":1}