{"id":"81c5b1c4-83c8-4fc0-8bd7-4a1fb789244f","arxiv_id":"2606.09021","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Introduces the sparse convexification hierarchy K^(s) = conv{v in K : ||v||_0 <= s} and proves an oracle inequality for a penalized estimator adapting to the best sparse convex approximation under sub-Gaussian assumptions.","lead":"The paper constructs a sparse convexification hierarchy K^(s) from a general symmetric convex body K and proposes a penalized least-squares estimator over this hierarchy for high-dimensional regression. A smart generalist might read it to see how general convex constraints can be handled with adaptive sparse approximations beyond fixed penalties like the Lasso.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"No significant objection identified","rationale":"Reader correctly flagged the symmetry and sub-Gaussian assumptions as the explicit hypotheses; the statistical claim is the standard width-based oracle inequality and no internal gap is visible from the stated result.","tokens_in":1700,"tokens_out":250,"duration_ms":17462,"concrete_test":"Verify that the oracle inequality in the main theorem continues to hold when the penalty is set exactly to the Gaussian width of K^(s) (or a constant multiple) and the design is isotropic sub-Gaussian; check whether the excess risk bound remains O(σ w(K^(s))/√n) without further restricted-eigenvalue assumptions on K^(s).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is an oracle inequality for a penalized least-squares estimator over the hierarchy K^(s), yielding a rate controlled by σ and the Gaussian width of K^(s) under explicit sign-symmetry, permutation-invariance of K, and sub-Gaussian design/noise. These conditions are standard for controlling empirical processes via Gaussian widths in random-design constrained regression; the abstract states the result holds under precisely those assumptions with no additional hidden steps indicated.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper studies high-dimensional linear regression under a general symmetric convex constraint. It defines the sparse convexification hierarchy K^{(s)} = conv{v in K : ||v||_0 <= s} for a sign-symmetric and permutation-invariant convex body K, proposes a penalized least-squares estimator over this hierarchy, and proves an oracle inequality under standard sub-Gaussian assumptions on the random design and noise. The estimator adapts to the best sparse convex approximation of the target; for an s-sparse target the squared-error rate is governed by the noise level σ and the Gaussian width of K^{(s)}. The framework applies to symmetric norm balls, is implementable via oracle access to the Minkowski functional of K, and recovers consistency for the constrained Lasso as a special case.","tokens_in":1805,"tokens_out":473,"duration_ms":21135,"significance":"If the oracle inequality holds, the work supplies a general mechanism for high-dimensional constrained regression that adapts to sparsity without committing to a specific penalty, by leveraging the geometry of the convex body via its sparse convexification and Gaussian width. This unifies results across different symmetric convex constraints and recovers the constrained Lasso. The explicit invocation of standard sub-Gaussian tail bounds and the implementation remark via Minkowski functional are strengths; the result is parameter-free in the sense that the rate depends only on intrinsic geometric quantities of K^{(s)} rather than on fitted tuning parameters beyond the hierarchy level s.","major_comments":[],"minor_comments":[{"comment":"The abstract states that the estimator 'searches over this hierarchy' but does not give the explicit form of the penalized objective (e.g., whether the penalty is the Minkowski functional of K^{(s)} or an indicator constraint); adding the precise optimization problem in the introduction would clarify the method.","section":null},{"comment":"The Gaussian width appears as the controlling quantity in the rate; a brief parenthetical reminder of its definition (supremum of the expected supremum of the Gaussian process over the set) would help readers outside the empirical-process literature.","section":null},{"comment":"The claim that the method 'can be implemented using oracle access to the Minkowski functional of K' is stated without an algorithmic sketch; a short paragraph outlining how the convex program is solved would strengthen the practical contribution.","section":null}],"recommendation":"minor_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the detailed summary, positive significance assessment, and recommendation of minor revision. No major comments appear in the report, so we have no specific points to address or revise.","responses":[],"tokens_in":1296,"tokens_out":48,"duration_ms":8990,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main contribution is the explicit construction of K^(s) as the convex hull of the s-sparse elements inside a general sign-symmetric, permutation-invariant convex body K, together with a penalized least-squares estimator that searches over this hierarchy and adapts to the best sparse convex approximation of the target.\n\nIt does a clean job of unifying several constrained regression problems under one template instead of deriving separate theory for each norm ball. The special case that recovers the constrained Lasso serves as a useful consistency check. The oracle inequality is stated under the usual sub-Gaussian assumptions on design and noise, with the rate controlled by sigma and the Gaussian width of K^(s); that is the expected geometric quantity in this literature.\n\nThe soft spots are limited. The bound still requires bounding the Gaussian width of the new set K^(s), so the geometric work is not removed for a concrete K. The method needs oracle access to the Minkowski functional of K, which is theoretically convenient but restricts immediate use to bodies where that oracle is cheap. No circularity appears in the stated claim.\n\nThis is for people working on high-dimensional constrained estimation who want a broader framework. A reader already comfortable with Gaussian widths and oracle inequalities will get the most out of it. The paper deserves a serious referee because the setup is precise and the central result is stated without hidden steps.","headline":"The paper gives a general sparse convexification hierarchy for arbitrary symmetric convex constraints and proves a standard oracle inequality for the penalized estimator over it.","tokens_in":2259,"tokens_out":344,"would_cite":false,"duration_ms":17240,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A penalized least-squares estimator over the sparse convexification of a convex body adapts to the best s-sparse approximation under sub-Gaussian assumptions.","keywords":["high-dimensional regression","sparse convexification","oracle inequality","convex constraints","constrained Lasso","Gaussian width","sub-Gaussian noise","penalized least squares"],"falsifier":"A sequence of instances with an s-sparse target in K where the estimator's squared error remains larger than any fixed multiple of σ times the Gaussian width of K^(s) as the sample size grows would falsify the claimed oracle inequality.","tokens_in":2598,"feed_emoji":"","tokens_out":673,"duration_ms":17171,"temperature":0.7,"pith_summary":"The paper introduces sparse convexification for high-dimensional linear regression under a general symmetric convex constraint K. It defines the hierarchy K^(s) as the convex hull of all at-most-s-sparse vectors inside K and proposes a penalized estimator that searches over this hierarchy. The estimator is proved to satisfy an oracle inequality that automatically adapts to the closest sparse convex approximation of the unknown target. For an exactly s-sparse target the resulting squared-error bound is controlled by the noise level and the Gaussian width of K^(s). The construction covers arbitrary sign-symmetric permutation-invariant convex bodies and recovers the constrained Lasso as a special case.","feed_headline":"Sparse convexification adapts high-dim regression estimator to s-sparse targets","feed_subtitle":"Penalized least-squares over K^(s) yields squared-error rate set by noise level and Gaussian width of the sparse convexification.","key_machinery":"The sparse convexification hierarchy K^(s) = conv{v ∈ K : ||v||_0 ≤ s}, which turns an arbitrary convex constraint into a searchable family of sparse approximations.","core_discovery":"Under standard sub-Gaussian assumptions on the random design and noise, the penalized least-squares estimator over the sparse convexification hierarchy adapts to the best sparse convex approximation of the target; for an s-sparse target this yields a squared-error rate governed by the noise level σ and the Gaussian width of K^(s).","pith_inferences":["The same hierarchy construction could be applied to other convex bodies that arise in structured signal recovery beyond norm balls.","Efficient projection oracles onto K^(s) for concrete K would turn the theoretical estimator into a practical algorithm.","Gaussian-width calculations for specific sparse convexifications might yield new explicit rates for problems currently handled only by generic Lasso or group-Lasso penalties."],"forward_implications":["The approach applies to any sign-symmetric permutation-invariant convex body, including all symmetric norm balls.","Implementation requires only oracle access to the Minkowski functional of the original set K.","The special case of the ℓ1 ball recovers a consistency result for the constrained Lasso.","The rate depends on the Gaussian width of K^(s) rather than on ambient dimension or explicit sparsity penalties."],"fun_headline_variants":["Sparse convexification adapts high-dim regression to sparse sets","Penalized estimator adapts to best sparse convex approximation","K^(s) sets adaptive squared-error rates for sparse regression","Sparse convexification yields consistency for constrained Lasso"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The convex body K must be sign-symmetric and permutation-invariant, and both the design matrix and the noise must satisfy sub-Gaussian tail bounds.","fun_headline_variants_meta":{"raw":{"variants":["Sparse convexification adapts high-dim regression to sparse sets","Penalized estimator adapts to best sparse convex approximation","K^(s) sets adaptive squared-error rates for sparse regression","Sparse convexification yields consistency for constrained Lasso"]},"model":"grok-4.3","cost_usd":0.006687,"raw_usage":{"total_tokens":3093,"prompt_tokens":621,"num_sources_used":0,"completion_tokens":61,"cost_in_usd_ticks":66874500,"prompt_tokens_details":{"text_tokens":621,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2411,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":621,"tokens_out":61,"duration_ms":13878,"temperature":1.0,"reasoning_tokens":2411,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-27T14:56:48.842435+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A sequence of instances with an s-sparse target in K where the estimator's squared error remains larger than any fixed multiple of σ times the Gaussian width of K^(s) as the sample size grows would falsify the claimed oracle inequality.","supporting_citations":[],"review_version":1}