{"id":"4a225e11-eac3-4219-80eb-31c4aa31b8eb","arxiv_id":"2507.11253","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper proves no-gap pointbased second-order characterizations of tilt-stable local minimizers for composite optimization with convex parabolically regular functions, under metric subregularity constraint qualification.","lead":"A new mathematical method gives second-order conditions that tell when a small tilt of an optimization problem keeps a solution a stable minimizer, for a wide class of composite optimization problems. The work unifies previously scattered results for nonlinear, conic, and spectral-norm programs under one weak constraint qualification.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.3's singleton-multiplier hypothesis is an extra nondegeneracy-like restriction, not a consequence of tilt stability; a simple degenerate example satisfies all standing assumptions but violates it, so the advertised no-gap characterization is not established.","rationale":"The paper develops a substantial variational-analysis framework, and the neighborhood characterization of Theorem 4.1 is plausibly correct. The main advertised contribution, however, is the pointbased no-gap pair. Looking at the exact statements, Theorem 4.3 differs from Theorem 4.2 by more than the strict/non-strict inequality: it also assumes a sequence of multipliers with singleton multiplier sets near x̄. This is precisely a nondegeneracy-type requirement, since uniqueness of multipliers at nearby points is not implied by MSCQ or tilt stability. The 2-D example in the attack is not pathological: it is a tilt-stable problem with linearly dependent active constraint gradients, exactly the degenerate situation the paper claims to cover. Since the singleton assumption is not satisfied, Theorem 4.3 simply does not apply, and the no-gap conclusion is not established for this setting. A secondary concern is that Assumption 2.3 is verified only for the spectral norm; however, the structural gap in the necessary condition alone is sufficient to justify conditional acceptance with mandatory revision. The reader's weakest_assumption pointed to Assumption 2.3 and to the singleton condition; I focus on the latter, hence partial agreement. I am not recommending rejection because the individual theorems may be correct under their stated hypotheses; they must be reframed as a conditional necessary condition, with the abstract and Section 5 revised accordingly.","tokens_in":30399,"tokens_out":16066,"duration_ms":203168,"concrete_test":"Compute the singleton-multiplier condition for the displayed example with μ=(1/2,1/2). Concretely, enumerate all sequences (x_k, μ_k) → (0, μ) with μ_k ∈ ∂g(F(x_k)); each has x_{k,1}=0 eventually and Λ(x_k, ∇F^T μ_k) is a line segment, so no such sequence is singleton. A second check: if the authors maintain the assumption is automatic under tilt stability, they should supply a derivation from the strong metric regularity of ∂f at (x̄,0); without such a derivation the no-gap claim should be withdrawn or restricted.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central 'no-gap' claim is undercut by the singleton-multiplier hypothesis in Theorem 4.3. That theorem requires that, for every relevant multiplier μ and direction v, there exist sequences (x_k, μ_k) → (x̄, μ) with μ_k ∈ ∂g(F(x_k)) and Λ(x_k, ∇F(x_k)^T μ_k) a singleton. This is neither derived from tilt stability nor implied by the standing assumptions, and it fails in a simple tilt-stable degenerate example. Let n=2, f0(x) = -x1 + (1/2)x2^2, F(x) = (x1, x1), and g = δ_{R_-^2}. Then (1) is minimize -x1 + (1/2)x2^2 subject to x1 ≤ 0, with a duplicated constraint; x̄=0 is tilt-stable with modulus 1. The multiplier set is the segment μ1+μ2=1, μ≥0, and the critical cone is span{(0,1)}, so every interior multiplier is relevant. For an interior μ, any sequence with μ_k ∈ ∂g(F(x_k)) and μ_k→μ must have x_{k,1}=0 eventually; then ∂g(F(x_k)) = R_+^2 and Λ(x_k, ∇F(x_k)^T μ_k) is the whole segment {ν≥0 : ν1+ν2 = μ_{k,1}+μ_{k,2}}, never a singleton. If x_{k,1}<0 then μ_k=0, which cannot approach μ. Thus the extra hypothesis fails in a problem satisfying MSCQ and Assumptions 2.1–2.3 (the projection onto R_-^2 has W=I in ∂B Prox, so Assumption 2.3 holds). Consequently Theorems 4.2 and 4.3 are not a no-gap pair in the sense defined in the paper, and the 'without nondegeneracy' claim is not justified as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies tilt stability of local minimizers for composite optimization problems of the form minimize f0(x) + g(F(x)), where g is l.s.c. convex and parabolically regular, F and f0 are smooth, and no nondegeneracy condition is imposed. The authors introduce a second-order variational function (SOVF) Γ_g attached to the proximal mapping of g, derive its domain and differential properties under a new Assumption 2.3, and then use these tools to prove a neighborhood characterization (Theorem 4.1), a pointbased sufficient condition (Theorem 4.2), and a pointbased necessary condition (Theorem 4.3) for tilt stability with modulus κ under MSCQ. The paper claims these two pointbased results form a no-gap characterization without nondegeneracy. Appendix A verifies Assumption 2.3 explicitly for the matrix spectral norm function.","tokens_in":30882,"tokens_out":6283,"duration_ms":79140,"significance":"If the advertised no-gap claim held as stated, this would be a substantial contribution: it would extend the pointbased tilt-stability theory, previously available for nonlinear programming and special conic cases, to a broad class of nonpolyhedral composite problems under only metric subregularity. The construction of the SOVF is novel, and the explicit verification for the spectral norm in Appendix A is a concrete and useful calculation. The paper also correctly identifies the relationship of its sufficient condition to the earlier NLP result of Gfrerer and Mordukhovich [9]. However, as discussed below, the necessary pointbased theorem carries an extra singleton-multiplier hypothesis that is not implied by tilt stability, so the central no-gap claim is not established in the generality announced in the abstract and conclusions.","major_comments":[{"comment":"The necessary condition in Theorem 4.3 is not a no-gap counterpart of Theorem 4.2 because it assumes, in addition to the hypotheses of Theorem 4.1 and Assumption 2.3, the existence of approximating sequences (x_k, μ_k) → (x̄, μ) with μ_k ∈ ∂g(F(x_k)) for which Λ(x_k, ∇F(x_k)^T μ_k) is a singleton. This is a nondegeneracy-type restriction; it is neither derived from tilt stability nor implied by MSCQ and Assumptions 2.1–2.3. A concrete tilt-stable problem satisfying all the standing assumptions but violating this condition is the following: n=2, f0(x) = −x1 + (1/2)x2^2, F(x) = (x1, x1), and g = δ_{R_-^2}. Then (1) is minimize −x1 + (1/2)x2^2 subject to x1 ≤ 0, x̄ = 0 is tilt-stable with modulus 1, the multiplier set is the segment {μ ≥ 0 : μ1 + μ2 = 1}, and every interior multiplier is relevant for the critical direction (0,1). For any sequence with μ_k ∈ ∂g(F(x_k)) converging to an interior μ, we must have x_{k,1} = 0 eventually, in which case Λ(x_k, ∇F(x_k)^T μ_k) is the whole segment {ν ≥ 0 : ν1 + ν2 = μ_{k,1} + μ_{k,2}} and is never a singleton. Thus the extra hypothesis fails, and the advertised no-gap characterization is not established without additional nondegeneracy-like assumptions.","section":"§4, Theorem 4.3"},{"comment":"Assumption 2.3 is load-bearing: it is used for the domain representation of the SOVF in Proposition 3.2, for the equality form of Γ_g in Proposition 3.3, and for the limiting inequalities in Theorems 4.2 and 4.3. The paper verifies this assumption only for the spectral norm function in Appendix A, while Section 3 and the concluding section state that the framework covers the nuclear norm and indicators of standard cones such as the positive semidefinite cone. Since the verification for those cases is not supplied, the claim that the general results apply to these important classes is unsupported as it stands. The authors should either verify Assumption 2.3 for those cases or explicitly restrict the scope of the main theorems.","section":"§3, Assumption 2.3 and Section 5"},{"comment":"Several central results are quoted from companion preprints rather than proved in the paper: Proposition 3.1 is cited from [33, Proposition 3.3], the proof of Proposition 3.2 invokes [33, Lemma 4.2] and [28, Lemma 5.2], and Theorem 3.1 uses [28, Lemma 5.2] and [33, Proposition 3.2]. Because these lemmas are essential to the domain representation and to the equality form of the SOVF, the manuscript is not self-contained at exactly the points where its main novelty resides. The authors should include proofs of these auxiliary results or at least state them as theorems with full proofs in the appendix.","section":"§3, Propositions 3.1–3.3 and Theorem 3.1"}],"minor_comments":[{"comment":"The statement of Theorem 4.3 is grammatically unclear: the clause 'suppose that for any nonzero critical direction w ... and ∇F(¯x)v ∈ dom Γg(F(¯x), µ), and there exist (xk, µk) → (¯x, µ)...' does not make clear the quantification over v and whether the sequence condition is required for every such v or only for the v used in the contradiction. This should be rewritten with explicit quantifiers.","section":"§4, Theorem 4.3"},{"comment":"There is a typo in Definition 2.2: 'satisfies the the metric subregularity constraint qualification' should read 'satisfies the metric subregularity constraint qualification.'","section":"§2, Definition 2.2"},{"comment":"The quantifier structure in Assumption 2.3 is difficult to parse: the phrase 'we have y = W ¯z, where ¯z satisfies the equalities' should specify whether ¯z depends on y and on the chosen V, and whether the displayed minimum must be attained for every V in the union or only for some. Clarifying this would improve readability.","section":"§2, Assumption 2.3"},{"comment":"The sentence claiming that Γ_f is 'generalized quadratic with its domain being a linear subspace' for locally Lipschitz C2-cone reducible convex functions should be justified or marked as a consequence of the cited companion results, since it relies on the range equality proved under Assumption 2.3.","section":"§3, after Proposition 3.3"},{"comment":"In the proof of Theorem 4.1, the final step cites [4, Theorem 3.3] without stating which hypotheses of that theorem are verified in the present setting; adding a sentence identifying the verified hypotheses would help the reader.","section":"§4, Theorem 4.1"}],"recommendation":"major_revision","confidential_remarks":"The core sufficiency argument in Theorem 4.2 appears plausible and the SOVF construction is interesting, but the paper's main advertised achievement, the no-gap pointbased characterization without nondegeneracy, is not currently established because Theorem 4.3 carries a singleton-multiplier hypothesis that is not a consequence of tilt stability. The heavy reliance on two companion preprints ([28] and [33]) for key lemmas is also a concern for the reviewing process; the editor may wish to ask the authors to include those proofs or to confirm that the preprints are publicly available and stable. The paper fits the scope of the journal and the results are likely repairable by restating the necessary condition under its extra assumption and softening the no-gap claims accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a substantial paper, but the headline 'no-gap without nondegeneracy' does not survive contact with Theorem 4.3. The necessary condition contains an extra hypothesis—existence of multiplier sequences with singleton multiplier sets—that is not a consequence of tilt stability and fails in a perfectly ordinary degenerate example. The paper is still worth engaging, but it needs a serious revision.\n\nWhat is genuinely new: the second-order variational function Γ_g, the unified treatment under MSCQ, and the sufficient condition in Theorem 4.2. The neighborhood characterization in Theorem 4.1 is consistent with the known theory, and the spectral norm verification in Appendix A is detailed and convincing. The authors are honest that Assumption 2.3 is only verified there, even though the claimed framework includes PSD cone indicators and the nuclear norm.\n\nThe soft spot is load-bearing. Take n=2, f0(x) = -x1 + ½x2², F(x) = (x1,x1), g = indicator of R_-². The problem is min -x1 + ½x2² s.t. x1 ≤ 0, with the constraint duplicated. At x̄=0 the multiplier set is the segment μ1+μ2=1, μ≥0, and the critical cone is span{(0,1)}. This point is tilt-stable with modulus 1, and all standing assumptions (MSCQ, parabolic regularity, Assumption 2.3) hold. But for any interior multiplier, any sequence μ_k ∈ ∂g(F(x_k)) converging to μ must have x_{k,1}=0 eventually, and then Λ(x_k, ∇F(x_k)^T μ_k) is the whole segment, never a singleton. So the hypothesis of Theorem 4.3 fails exactly in a degenerate but tilt-stable example. The paper therefore does not actually prove a pointbased no-gap characterization 'without nondegeneracy'; it proves a sufficient condition plus a necessary condition that requires a nondegeneracy-like property at nearby points.\n\nMinor issues: key lemmas are sourced from companion preprints ([28], [33]), so the paper is not self-contained, and Assumption 2.3 is only checked for the spectral norm. These are less serious than the Theorem 4.3 gap.\n\nWho this is for: researchers in variational analysis and stability in conic/composite optimization. The sufficient-condition side and the SOVF concept are worth citing. I would send it to a serious referee, but the referee should be asked to address the necessary-condition gap and to reframe the claims accordingly.","headline":"The SOVF machinery is a genuine advance, but the advertised no-gap pointbased characterization is not established because Theorem 4.3 relies on a hidden singleton-multiplier condition that fails in a simple degenerate tilt-stable example.","tokens_in":31339,"tokens_out":5732,"would_cite":true,"duration_ms":68434,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49J52","49J53","90C31"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that tilt stability of local minimizers in composite optimization is characterized by a pointbased inequality involving a new second-order variational function, with no gap between the sufficient and necessary conditions.","keywords":["tilt stability","composite optimization","second-order variational function","metric subregularity constraint qualification","parabolic regularity","generalized differentiation","spectral norm","no-gap characterizations"],"falsifier":"Compute for $g$ equal to the nuclear norm on $3\\times 3$ matrices, or the indicator of the positive semidefinite cone, at a point where two singular values or eigenvalues coincide, and check whether Proposition 3.2's equality $\\mathrm{dom}\\,\\Gamma_g = \\mathrm{aff}\\{d : dg(x)(d)=\\langle u,d\\rangle\\}$ holds; if it fails for any such $g$, the necessity theorem 4.3 loses its domain condition, and the no-gap claim would be false as stated.","tokens_in":30242,"feed_emoji":"📐","tokens_out":6037,"duration_ms":66854,"temperature":0.7,"pith_summary":"Tilt stability asks whether a local minimizer survives small linear perturbations with a unique nearby minimizer that moves Lipschitz-continuously; it underpins error bounds and convergence analysis for optimization algorithms. This paper targets composite problems of the form $\\min f_0(x)+g(F(x))$, where $g$ is convex and nonsmooth but parabolically regular, and aims for conditions stated entirely at the candidate point. Its central claim is that, under the metric subregularity constraint qualification and a technical range-condition on the proximal mapping of $g$, tilt stability with modulus $\\kappa$ is equivalent to a strict (sufficient) or nonstrict (necessary) inequality involving the Hessian of the Lagrangian plus a new 'second-order variational function' $\\Gamma_g$ applied to the directional derivative of $F$. An earlier gap, where necessary conditions required nondegeneracy assumptions, is closed, and the whole apparatus is constructively verified for the matrix spectral norm.","feed_headline":"Tilt-stability gap closed by one second-order function","feed_subtitle":"Under a weak constraint qualification, tilt-stable minimizers are now pinned down by one second-order function.","key_machinery":"The load-bearing object is the second-order variational function $\\Gamma_f(x,u)(v) := \\min \\langle v, d-v\\rangle$ over $d$ with $v=Vd$ for $V$ in the generalized Jacobian $J\\,\\mathrm{Prox}_f(x+u)$, with value $\\infty$ outside the union of the ranges of those matrices. It supplies the second-order curvature term that replaces the second subderivative $d^2g$ when one wants conditions at a single point. Under Assumption 2.3 the domain of $\\Gamma_f$ is a linear subspace equal to the affine span of the critical directions, and the function is generalized quadratic; this domain identity is what converts neighborhood conditions into pointbased no-gap conditions. The companion machinery is the generalized Jacobian and B-subdifferential of the proximal map, together with the metric subregularity constraint qualification (MSCQ), which permits nonunique Lagrange multipliers.","core_discovery":"On the paper's own terms, the discovery is a no-gap pointbased second-order characterization: for problem (1) with parabolically regular convex $g$ satisfying Assumptions 2.1–2.3 and MSCQ, a feasible point $\\bar x$ is a tilt-stable local minimizer with modulus $\\kappa$ exactly when, for every nonzero critical direction $v$ and every relevant Lagrange multiplier $\\mu$, $\\langle v,\\nabla^2_{xx} L(\\bar x,\\mu)v\\rangle + \\Gamma_g(F(\\bar x),\\mu)(\\nabla F(\\bar x)v) \\ge \\frac{1}{\\kappa}\\|v\\|^2$, with the strict version as the sufficient condition. Here $\\Gamma_g$ is the second-order variational function, defined through the generalized Jacobian of the proximal mapping $\\mathrm{Prox}_g$. The paper presents Theorem 4.2 (sufficiency) and Theorem 4.3 (necessity) as the matching pair, with Theorem 4.1 giving the neighborhood form. The result covers nonpolyhedral problems such as conic and spectral-norm programs without the nondegeneracy assumptions that earlier pointbased results required.","pith_inferences":["Because Assumption 2.3 is verified only for the spectral norm, a natural next step is to test it for the nuclear norm and for indicators of the positive semidefinite and second-order cones; if it holds there, the no-gap characterization extends unchanged, and if not, a modified second-order variational function would be needed.","The second-order variational function can be read as a measure of nonpolyhedrality, vanishing for polyhedral $g$ and becoming a quadratic form for spectral functions, so the sharpness of the inequality could be used to classify which composite problems exhibit a gap between necessary and sufficient second-order conditions.","A testable extension would be to use the pointbased inequality to certify uniform quadratic growth or strong metric subregularity of the subdifferential, connecting tilt stability directly to rates of convergence of first-order methods."],"forward_implications":["For any composite problem satisfying Assumptions 2.1–2.3, tilt stability with modulus $\\kappa$ can be certified or disproved by checking one inequality at the candidate point, without enumerating points nearby.","When $g$ is the matrix spectral norm, all assumptions, including Assumption 2.3, are explicitly verified, yielding explicit no-gap second-order conditions for spectral-norm composite programs.","The results recover the earlier nonlinear-programming characterization of [9] when $g$ is the indicator of the positive orthant, since $\\Gamma_g$ vanishes in the polyhedral case.","The criteria give a way to compute or bound the tilt-stability modulus $\\kappa$ from data at the point, which is the input needed for generalized Newton and other second-order algorithms."],"supporting_citations":[{"why":"Introduces tilt stability and proves the basic second-order subdifferential characterization that this paper extends to composite problems.","marker":"[29]"},{"why":"Supplies the nonlinear-programming no-gap characterization under MSCQ that the polyhedral case of Theorem 4.2 reduces to.","marker":"[9]"},{"why":"Provides the subgradient graphical-derivative characterization of tilt stability used in the proofs of Theorems 4.1–4.3.","marker":"[4]"},{"why":"Gives the twice epi-differentiability results for composite functions that ground Assumption 2.1.","marker":"[18]"},{"why":"Develops the variational analysis of composite models, including the subdifferential formula $\\partial(g\\circ F)=\\nabla F^T\\partial g(F)$ used in Lemma 4.1.","marker":"[17]"},{"why":"Supplies the relationship between second subderivatives and the proximal mapping that underlies the domain representation of the second-order variational function.","marker":"[33]"},{"why":"Provides the projection and proximal formulas for the spectral norm used in Appendix A to verify Assumption 2.3.","marker":"[3]"},{"why":"Companion work whose Lemma 5.2 and related results are invoked for graphical-derivative domain representations.","marker":"[28]"}],"fun_headline_variants":["No-gap second-order test for tilt stability","Tilt stability now characterized by a single second-order inequality","Exact second-order criterion for tilt-stable minimizers","Weak constraint qualification enables full tilt-stability description","Pointwise second-order rule settles tilt-stability question"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pointbased no-gap argument rests on the assumption that the generalized Jacobian of the proximal map of $g$ has one common range and that a certain minimization over its matrices has the explicit Moore–Penrose form (Assumption 2.3), which the paper verifies only for the spectral norm.","fun_headline_variants_meta":{"raw":{"variants":["No-gap second-order test for tilt stability","Tilt stability now characterized by a single second-order inequality","Exact second-order criterion for tilt-stable minimizers","Weak constraint qualification enables full tilt-stability description","Pointwise second-order rule settles tilt-stability question"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000219,"raw_usage":{"total_tokens":1406,"prompt_tokens":874,"completion_tokens":532,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":490,"completion_tokens_details":{"reasoning_tokens":456}},"tokens_in":490,"tokens_out":532,"duration_ms":6373,"temperature":1.0,"reasoning_tokens":456,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:12:16.118602+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute for $g$ equal to the nuclear norm on $3\\times 3$ matrices, or the indicator of the positive semidefinite cone, at a point where two singular values or eigenvalues coincide, and check whether Proposition 3.2's equality $\\mathrm{dom}\\,\\Gamma_g = \\mathrm{aff}\\{d : dg(x)(d)=\\langle u,d\\rangle\\}$ holds; if it fails for any such $g$, the necessity theorem 4.3 loses its domain condition, and the no-gap claim would be false as stated.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces tilt stability and proves the basic second-order subdifferential characterization that this paper extends to composite problems."},{"cited_title":"Gfrerer and B","cited_arxiv_id":null,"evidence_quote":"Supplies the nonlinear-programming no-gap characterization under MSCQ that the polyhedral case of Theorem 4.2 reduces to."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the subgradient graphical-derivative characterization of tilt stability used in the proofs of Theorems 4.1–4.3."},{"cited_title":"Mohammadi and M","cited_arxiv_id":null,"evidence_quote":"Gives the twice epi-differentiability results for composite functions that ground Assumption 2.1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Develops the variational analysis of composite models, including the subdifferential formula $\\partial(g\\circ F)=\\nabla F^T\\partial g(F)$ used in Lemma 4.1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the projection and proximal formulas for the spectral norm used in Appendix A to verify Assumption 2.3."},{"cited_title":"Characterizations of Strong Variational Sufficiency in General Models of Composite Optimization","cited_arxiv_id":"2507.09522","evidence_quote":"Companion work whose Lemma 5.2 and related results are invoked for graphical-derivative domain representations."}],"review_version":1}