{"id":"3a9f7053-2c3a-41a7-ad12-a59ce63f9cb0","arxiv_id":"2607.05991","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A relative inexact proximal ALM with a tailored semismooth Newton solver solves sparse spectral-risk optimization faster than ADMM while matching stationarity and sparsity on synthetic and real data.","lead":"This paper develops a relative inexact proximal augmented Lagrangian method with a semismooth Newton subproblem solver for sparse spectral risk measure optimization. It matters because average-loss training can miss rare severe failures in high-stakes decisions, and the method aims to make SRM-based sparse learning computationally practical.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The dual/Moreau reduction plus practical SSN termination may not rigorously inherit the relative-inexact PALM global convergence guarantees under the paper's standing assumptions on spectral weights and the regularizer.","rationale":"The reader's weakest_assumption correctly isolates the load-bearing hinge: whether relative inexactness plus the dual/Moreau reductions preserve the global guarantees so that practical SSN termination still yields the claimed stationarity and sparsity. No stronger internal inconsistency is visible from the abstract and method claims; the algorithmic construction (PAVA Jacobians, dual reduction) is a legitimate specialization of known tools, and the empirical runtime edge is presented as comparable-quality rather than asymptotically superior. The concern is therefore one of inheritance and verification rather than outright contradiction. Because the manuscript is algorithmic math.OC without machine-checked proofs or shipped code, the CONDITIONAL verdict already reflects the appropriate contingency on full theory and baseline checks; the concrete residual audit above would settle the inheritance question without requiring a change of category. Novelty and significance assessments remain unchanged.","tokens_in":2035,"tokens_out":562,"duration_ms":22397,"concrete_test":"On the synthetic instances of Section 5, extract the dual residual and the Moreau-envelope gradient residual at every outer iteration where SSN is terminated; check whether these residuals satisfy the paper's stated relative inexactness criterion (the one used in the convergence theorem for the proximal ALM). If the criterion is violated on more than a few iterations while the reported stationarity measure is still met, the inheritance claim fails and the runtime advantage cannot be attributed to a convergent relative-inexact scheme.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on a relative-inexact proximal ALM whose subproblems are reduced via a dual reformulation and Moreau-envelope structure, then solved by SSN with PAVA-evaluated generalized Jacobians. For the reported stationarity residuals and sparsity to be theoretically supported (rather than merely observed), the relative inexactness criteria must remain valid after the dual/Moreau reduction and under the concrete termination used by the SSN solver. The standing assumptions on spectral weights, loss, and regularizer are invoked to justify this inheritance, yet the reduction changes the dual-variable geometry and the precise residual that must be controlled. If the relative criterion is only stated for the original primal ALM and is not re-derived (or shown equivalent) for the reduced dual subproblems, practical Newton termination can produce the tabulated residuals without satisfying the conditions that guarantee global convergence or the claimed quality relative to ADMM. This is the single point where the theory-to-practice link is least secure.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript develops a relative inexact proximal augmented Lagrangian method whose subproblems are solved by a semismooth Newton (SSN) solver for sparse spectral-risk-measure (SRM) optimization. By dual reformulation and Moreau-envelope structure the ALM subproblems are reduced to structured dual-variable problems; explicit generalized Jacobians of the dual residual maps are characterized and evaluated by a tailored pool-adjacent-violators algorithm. Global convergence of the outer relative-inexact proximal ALM is claimed under standing assumptions on spectral weights, loss, and regularizer. Numerical experiments on synthetic and real instances report lower runtimes than an ADMM baseline at comparable stationarity residuals and solution sparsity.","tokens_in":2243,"tokens_out":1047,"duration_ms":27588,"significance":"SRM-based empirical risk minimization is a practically relevant nonsmooth, nonseparable problem class in risk-sensitive learning. A dual/Moreau reduction that yields tractable SSN subproblems, together with explicit generalized-Jacobian formulae and a specialized PAVA evaluation, constitutes a concrete algorithmic contribution. If the relative-inexactness theory carries through the reduction and the reported speed-ups hold under fair baselines, the method would be a useful addition to the toolkit for spectral-risk optimization. The paper supplies comparative numerics and structured Jacobian characterizations rather than free-parameter fitting; those are genuine strengths.","major_comments":[{"comment":"The central theory-to-practice link is the inheritance of the relative inexact proximal-ALM convergence guarantees after the dual/Moreau reduction. The relative residual criteria appear to be stated for the original primal ALM; after dual reformulation the controlled residual and dual-variable geometry change. The manuscript needs an explicit argument (or equivalence lemma) showing that the concrete SSN termination used in the dual subproblems still satisfies the relative inexactness conditions that underwrite global convergence and the claimed stationarity quality versus ADMM. Without that bridge, the tabulated residuals and sparsity are only empirical observations, not consequences of the theory.","section":"Theory: relative-inexact PALM and dual/Moreau subproblem reduction"},{"comment":"Standing assumptions on the spectral weight vector (monotonicity/ordering) and on the regularizer are invoked to justify both the dual reduction and the generalized-Jacobian formulae. It should be stated precisely which of these assumptions are used for (i) well-posedness of the dual residual map, (ii) the PAVA-based Jacobian evaluation, and (iii) the relative-inexact outer convergence. If any assumption is stronger than what is needed for the numerics (e.g., strictly decreasing weights), that gap should be flagged so that the scope of the guarantees is clear.","section":"Standing assumptions; Jacobian characterizations"},{"comment":"The ADMM baseline is the sole external comparator. For the claim of systematically lower running times at comparable residuals to be load-bearing, the ADMM implementation must be described at the same level of detail as the proposed method (penalty update schedule, termination tolerances, warm-start policy, and whether the same dual/Moreau structure is exploited). Otherwise the runtime advantage may reflect implementation disparity rather than algorithmic superiority. A short fairness paragraph or an additional first-order baseline would strengthen the numerical claim.","section":"Numerical experiments / ADMM baseline"}],"minor_comments":[{"comment":"Notation for the spectral weight vector and the associated sorting operator should be fixed once and used consistently; occasional switches between w and lambda (or similar) make the dual residual maps harder to track.","section":"Notation / preliminaries"},{"comment":"The tailored PAVA procedure is central to Jacobian evaluation; a short complexity statement (or reference to the classical O(n) bound under the paper's ordering assumptions) would help readers assess per-iteration cost.","section":"Jacobian evaluation / PAVA"},{"comment":"Figures and tables reporting residuals and sparsity should state the precise stationarity measure (primal residual, dual residual, or KKT residual) and the sparsity threshold used, so that 'comparable' is reproducible.","section":"Numerical results tables/figures"},{"comment":"A brief remark on how the free ALM penalty and relative-tolerance sequences are chosen in practice (fixed schedule vs adaptive) would aid reproducibility.","section":"Algorithm parameters"}],"recommendation":"major_revision","confidential_remarks":"The skeptic's concern about inheritance of relative inexactness after the dual/Moreau reduction is the single load-bearing theoretical gap; if the authors supply a short equivalence or residual-control argument, the paper is close to minor-revision territory. Scope fits a computational optimization journal; novelty is algorithmic rather than conceptual. No citation-pattern or authorship concerns noted from the materials provided."},"author_rebuttal":{"model":"grok-4.5","summary":"We thank the referee for a careful and constructive report. The three major comments correctly identify the places where the theory-to-algorithm link, the scope of standing assumptions, and the fairness of the ADMM comparison need to be made fully explicit. We will revise the manuscript accordingly: (i) add an equivalence argument relating the dual SSN residual to the relative inexactness criterion of the outer proximal ALM; (ii) itemize precisely which assumptions are used for dual well-posedness, PAVA Jacobian evaluation, and outer convergence, and flag any gap relative to the numerical setting; (iii) expand the ADMM baseline description to the same level of implementation detail as the proposed method and add a short fairness paragraph. None of these revisions alter the algorithmic design or the reported numerical conclusions; they strengthen the supporting arguments.","responses":[{"response":"We agree that an explicit bridge is required and that the current draft does not state it with sufficient clarity. The outer relative-inexact proximal ALM is formulated in the primal variables; after the dual/Moreau reduction the subproblem residual is measured on a dual residual map whose geometry differs from the primal residual. Because the Moreau envelope of the dual spectral-risk term is continuously differentiable with Lipschitz gradient, and because the primal–dual pair is linked by the proximal mapping of that envelope, the dual residual norm controls the primal residual that appears in the relative inexactness criterion. We will insert a short equivalence lemma (after the dual reduction) that makes this control quantitative: any dual residual small enough relative to the current dual step size implies the relative residual bound required by the outer convergence theorem. The concrete SSN termination used in the implementation will then be shown to satisfy that dual residual bound. With the lemma in place, global convergence of the outer iteration and the stationarity quality of the returned points become consequences of the theory rather than purely empirical observations. The tabulated residuals will be re-interpreted in light of this guarantee.","revision_made":"yes","referee_comment":"The central theory-to-practice link is the inheritance of the relative inexact proximal-ALM convergence guarantees after the dual/Moreau reduction. The relative residual criteria appear to be stated for the original primal ALM; after dual reformulation the controlled residual and dual-variable geometry change. The manuscript needs an explicit argument (or equivalence lemma) showing that the concrete SSN termination used in the dual subproblems still satisfies the relative inexactness conditions that underwrite global convergence and the claimed stationarity quality versus ADMM. Without that bridge, the tabulated residuals and sparsity are only empirical observations, not consequences of the theory."},{"response":"The referee’s request for a precise inventory is well taken. We will add a short subsection (or a clearly labeled paragraph at the beginning of the theory section) that lists each standing assumption and tags it with the roles (i)–(iii). Concretely: monotonic nonincreasing spectral weights are used for (i) dual residual well-posedness (via the dual of the ordered weighted average) and for (ii) correctness of the pool-adjacent-violators evaluation of the generalized Jacobian; they are not required for (iii) outer relative-inexact convergence once the dual residual is well-defined. Strict decrease is used only to guarantee uniqueness of certain active-set partitions inside the PAVA routine and is therefore stronger than needed for the outer theory; the numerical instances satisfy it, but we will explicitly flag that the outer guarantees remain valid under mere monotonicity. The regularizer assumptions (convexity, properness, and the existence of a computable proximal mapping) are used for (i) and (iii) but not for the PAVA formulae themselves. This itemization will make the scope of every claim transparent and will prevent readers from over-interpreting the strength of the hypotheses.","revision_made":"yes","referee_comment":"Standing assumptions on the spectral weight vector (monotonicity/ordering) and on the regularizer are invoked to justify both the dual reduction and the generalized-Jacobian formulae. It should be stated precisely which of these assumptions are used for (i) well-posedness of the dual residual map, (ii) the PAVA-based Jacobian evaluation, and (iii) the relative-inexact outer convergence. If any assumption is stronger than what is needed for the numerics (e.g., strictly decreasing weights), that gap should be flagged so that the scope of the guarantees is clear."},{"response":"We accept the criticism. The current draft describes the ADMM baseline only at a high level, which leaves open the possibility that the observed speed-ups are partly implementation artifacts. In the revision we will expand the experimental section with a dedicated fairness paragraph that records, for ADMM: the precise penalty-parameter update schedule (including the initial value and the multiplicative factor), the absolute and relative termination tolerances used for both the outer ADMM iteration and any inner linear-system solves, the warm-start policy for the dual and slack variables, and whether the same dual/Moreau reduction of the spectral-risk term is exploited inside the ADMM subproblems. Both codes will be confirmed to share the same linear-algebra backend and the same sparsity-aware data structures. We will also note that the dual residual maps used for termination are comparable, so that the stationarity residuals reported in the tables are measured on an equal footing. If space permits we will add a simple first-order proximal-gradient baseline on the dual formulation as a further sanity check; otherwise the expanded ADMM description alone will make the runtime comparison load-bearing.","revision_made":"yes","referee_comment":"The ADMM baseline is the sole external comparator. For the claim of systematically lower running times at comparable residuals to be load-bearing, the ADMM implementation must be described at the same level of detail as the proposed method (penalty update schedule, termination tolerances, warm-start policy, and whether the same dual/Moreau structure is exploited). Otherwise the runtime advantage may reflect implementation disparity rather than algorithmic superiority. A short fairness paragraph or an additional first-order baseline would strengthen the numerical claim."}],"tokens_in":1812,"tokens_out":1309,"duration_ms":27025,"standing_objections":[]},"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a carefully engineered specialized solver for sparse spectral risk minimization, not new risk theory. They take relative inexact proximal ALM, reduce the subproblems via dual/Moreau structure, and equip the SSN step with explicit generalized Jacobians evaluated by a tailored PAVA. On the instances they report, it beats their ADMM baseline on runtime at comparable residuals and sparsity.\n\nWhat is actually new is the integration for this problem class. Dual reformulation plus Moreau envelope to get structured dual-variable subproblems, plus the PAVA-based Jacobian evaluation for the sorting-induced structure, is a legitimate algorithmic contribution. The pieces exist separately (ALM, SSN, PAVA for isotonic problems, SRM/CVaR optimization), but packaging them this way for sparse SRM and writing down the Jacobians is real work. Numerical claims are comparative and look clean enough for a methods paper: lower runtimes, similar stationarity and sparsity. Circularity burden is low—this is methods work against an external baseline, not fitting free constants to define the target.\n\nSoft spots are proportionate. The stress-test concern is real but not automatically fatal: after dual/Moreau reduction the residual geometry changes, so the relative inexactness criteria need to be shown to carry over (or to be equivalent) under the standing assumptions on spectral weights, loss, and regularizer. If the theory only states the criterion for the original primal ALM and does not re-derive it for the reduced dual subproblems, the global-convergence claim is thinner than the abstract suggests. That is a referee-check item, not a reason to dismiss the paper. Other soft spots: free ALM/proximal parameters and inexactness tolerances; only ADMM as baseline; no shipped code in what we have. None of these sink the contribution if the Jacobian characterizations and the inheritance argument check out.\n\nWho it is for: people who actually solve nonsmooth risk-sensitive ERM problems, or who work on second-order methods for structured nonsmooth convex programs. A serious referee in math.OC or computational optimization should see it. I would send it to peer review; the contribution is accept-shaped if the relative-inexact theory and Jacobian characterizations hold. Engage—worth a careful pass on convergence inheritance after reduction and on fairness of the ADMM comparison.","headline":"Solid specialized ALM–SSN solver for sparse spectral-risk ERM; real within-subfield computational progress, not a foundational advance.","tokens_in":2896,"tokens_out":572,"would_cite":false,"duration_ms":20322,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C25","90C06","90C30","65K05"],"pacs":[],"model":"grok-4.5","headline":"A relative inexact proximal ALM with semismooth Newton solves sparse spectral-risk optimization faster than ADMM","keywords":["spectral risk measure","augmented Lagrangian method","semismooth Newton","Moreau envelope","pool-adjacent-violators","sparse optimization","relative inexactness","empirical risk minimization"],"falsifier":"On a suite of sparse spectral-risk instances with known spectral weights, run the method with the stated relative inexactness tolerances and compare final stationarity residual, support cardinality, and wall-clock time against ADMM; if residuals are systematically worse or times are not lower while solutions remain sparse, the central claim fails.","tokens_in":2889,"feed_emoji":"⚡","tokens_out":604,"duration_ms":9422,"temperature":0.7,"pith_summary":"Average-loss training can hide rare but severe failures. Spectral risk measures fix that by weighting ordered losses, yet the sorting step makes the objective nonsmooth and nonseparable, so standard solvers struggle. This paper builds a relative inexact proximal augmented Lagrangian method whose subproblems are solved by a semismooth Newton iteration. A dual reformulation and Moreau-envelope structure collapse each subproblem onto a structured dual variable whose generalized Jacobian can be evaluated by a tailored pool-adjacent-violators algorithm. Under the paper's standing assumptions the method converges, and on synthetic and real instances it matches the stationarity and sparsity quality of an ADMM baseline while finishing in less wall-clock time. The practical payoff is a reliable first-order-to-Newton pipeline for high-stakes sparse learning that respects worst-case tails without sacrificing speed.","feed_headline":"Semismooth Newton ALM beats ADMM on sparse spectral-risk problems","feed_subtitle":"Dual Moreau reformulation plus pool-adjacent-violators Jacobians cut runtime while matching stationarity and sparsity.","key_machinery":"The dual residual map of the Moreau-envelope reformulation of the proximal ALM subproblem; its generalized Jacobian is assembled by a pool-adjacent-violators algorithm that enforces the ordered spectral weights, turning each Newton step into a structured linear solve.","core_discovery":"A relative inexact proximal augmented Lagrangian method, with subproblems solved by a semismooth Newton solver that exploits a dual reformulation and Moreau-envelope structure, solves sparse spectral-risk optimization problems to stationarity and sparsity quality comparable to ADMM while attaining lower running times. Explicit generalized Jacobians of the dual residual map are obtained via a tailored pool-adjacent-violators algorithm, making each Newton step inexpensive.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Semismooth Newton ALM cuts runtime vs ADMM for sparse SRMs","Dual Moreau reformulation speeds sparse spectral-risk ALM","PAV-tailored Jacobians cut Newton cost in spectral risk opt","Inexact proximal ALM yields lower times than ADMM on SRMs","Dual reform and Moreau structure simplify SRM ALM subproblems"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That the relative inexactness criteria together with the dual and Moreau reductions still guarantee global convergence of the proximal ALM under the paper's assumptions on spectral weights, loss, and regularizer, so that practical Newton termination yields the claimed stationarity and sparsity.","fun_headline_variants_meta":{"raw":{"variants":["Semismooth Newton ALM cuts runtime vs ADMM for sparse SRMs","Dual Moreau reformulation speeds sparse spectral-risk ALM","PAV-tailored Jacobians cut Newton cost in spectral risk opt","Inexact proximal ALM yields lower times than ADMM on SRMs","Dual reform and Moreau structure simplify SRM ALM subproblems"]},"model":"grok-4.5","cost_usd":0.0125,"raw_usage":{"total_tokens":2604,"prompt_tokens":733,"num_sources_used":0,"completion_tokens":96,"cost_in_usd_ticks":125000000,"prompt_tokens_details":{"text_tokens":733,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1775,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":733,"tokens_out":96,"duration_ms":48889,"temperature":1.0,"reasoning_tokens":1775,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-08T19:05:31.199207+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a suite of sparse spectral-risk instances with known spectral weights, run the method with the stated relative inexactness tolerances and compare final stationarity residual, support cardinality, and wall-clock time against ADMM; if residuals are systematically worse or times are not lower while solutions remain sparse, the central claim fails.","supporting_citations":[],"review_version":1}