{"id":"ab7d353c-fc7a-407d-ad0c-017781e8a5f7","arxiv_id":"2607.25252","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Two proximal gradient methods, with and without line search, are shown to find stationary points of non-smooth set-valued optimization problems at an O(1/√k) rate.","lead":"This paper designs two proximal-gradient algorithms for optimization problems where the objective is a set of functions rather than a single number, ordered by a cone, and proves they converge to stationary points. The line-search variant is claimed to converge without the regularity assumptions that earlier steepest-descent methods for set optimization required.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.1's stationarity equivalence swaps ∀v∃j into ∃j∀v; a simple linear example has Θ_l(0)=0 yet fails Definition 3.4, so the claimed link to weak minimality is not established as stated.","rationale":"The reader's weakest assumption (Assumption 5.1) is a standard level-boundedness condition and is not the main risk. A more fundamental correctness risk is the stationarity characterization itself: the proof of Theorem 3.1 incorrectly converts ∀v∃j into ∃j∀v. The linear counterexample above lies inside the paper's setting (finite cardinality, C=R_+^2, f_j linear, g_j=0) and shows both Proposition 3.2 and Theorem 3.1 are false as stated. The algorithms may still converge to Θ_l=0, and the convergence proofs may be salvageable by redefining stationarity as no common descent direction; so a conditional accept with mandatory correction, rather than outright reject, is appropriate. I credit the paper for a plausible algorithmic framework and do not dispute the O(1/√k) complexity argument conditional on a corrected stationarity statement. But as written, the central claim linking Θ_l=0 to stationarity/weak minimality is not sound, which is why the reader's verdict should remain conditional pending this fix.","tokens_in":29924,"tokens_out":37412,"duration_ms":352231,"concrete_test":"Run the following analytical check. Set C=R_+^2 and H(x)={(x1,x2),(1-x1,-1-x2)} at x0=0. (1) Verify H(0) is its own minimal set and P_0={(1,2)}. (2) Verify local weak minimality by cases: for ||x||∞<1/2, 0∈H(x)+int(C) forces x∈-int(C) or x1>1,x2<-1; u=(1,-1)∈H(x)+int(C) forces x∈int(C) or x1<1,x2<-1; the constraints cannot hold simultaneously for small x. (3) Compute Θ_l(0) with l>0: φ(w)=max(w1,w2), so max(φ(v),φ(-v))=max(|v1|,|v2|), giving Θ_l(0)=0 at v=0. (4) Show Definition 3.4 fails for a=(1,2): j=1 fails at v=(-1,-1), j=2 fails at v=(1,1). This contradicts Theorem 3.1 and pinpoints the quantifier swap as the cause.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Proposition 3.2 passes from local weak minimality to '∀a ∀v ∃j: D_v h_{a_j}(y)∉-int(C)' and then asserts '∃j ∀v', a quantifier reversal. The reversal is not harmless. Let C=R_+^2, H(x)={h1(x),h2(x)} with h1(x)=(x1,x2), h2(x)=(1-x1,-1-x2), y=0. H(0)={(0,0),(1,-1)}; both elements are minimal, so P_0={(1,2)}. For any ||x||∞<1/2, H(x)={x,(1-x1,-1-x2)} fails to dominate H(0): to cover 0 one needs x in the negative quadrant, but then (1-x1,-1-x2) has second component < -1/2, so u=(1,-1) is uncovered; the other cases leave 0 uncovered. Hence 0 is locally weakly minimal. Yet Definition 3.4 fails: for a=(1,2), j=1 fails at v=(-1,-1), j=2 fails at v=(1,1); no single index blocks all directions. Meanwhile Θ_l(0)=min_v [max(φ(v),φ(-v))+l/2||v||²]=0 because max(φ(v),φ(-v))=||v||∞, attained at v=0. Thus v_l(0)=0 and Θ_l(0)=0, contradicting Theorem 3.1's equivalence with Definition 3.4. The real stationarity condition is the weaker F(v)≥0 ∀v (no common descent), but the paper's Definition 3.4 and Proposition 3.2 are false as written. Since the algorithms are defined through Θ_l and the claimed connection to weak minimality rests on Theorem 3.1, this is a load-bearing gap, not a typo.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the unconstrained set-valued optimization problem (SOP_l) with objective H(x)={h_1(x),...,h_p(x)}, each h_j=f_j+g_j, under the lower set-less order induced by a closed convex pointed solid cone. It introduces a stationarity notion based on a scalarization of the minimal set and a partition set of active indices, then proposes two proximal gradient methods: Algorithm 1 with an Armijo-type line search and Algorithm 2 with a fixed step size and adjusted proximal parameter. The paper claims global convergence of both methods to stationary points under an image-space level-boundedness assumption (Assumption 5.1) and an O(1/sqrt(k)) best-iterate complexity rate, followed by numerical experiments on constructed set-valued test problems with robust uncertainty terms.","tokens_in":30390,"tokens_out":16735,"duration_ms":157717,"significance":"If the theoretical claims are correct, this would be a valuable contribution: it appears to be the first proximal gradient framework for non-smooth set-valued optimization problems with finitely many image points, it removes the regularity assumptions required by the steepest descent method of Bouza et al., and it provides the first complexity analysis for this class. The numerical implementation is reasonably extensive and demonstrates practical viability. However, the central optimality characterization is false as stated: the quantifier order in Definition 3.4 is interchanged, which breaks the claimed equivalence with weak minimality and therefore undermines the interpretation of the algorithms' stopping criteria. The paper also contains serious proof errors in the line-search existence argument and an unsupported rate claim. These issues are repairable, but the current version is not publishable without major revision.","major_comments":[{"comment":"The quantifier order in Definition 3.4 is invalid. From Proposition 3.1 applied to the vector subproblem (VOP_a) one obtains for all a in P_y and all v in R^n that there exists j with D_v h_{a_j}(y) not in -int(C). Definition 3.4 instead requires for all a that there exists j such that this holds for all v. This is strictly stronger. The proof of Theorem 3.1(ii)=>(i) actually derives max_j phi(D_v h_{a_j}(y)) >= 0 for all a,v, which is the correct forall-v-exists-j condition, not the stated definition. The difference is material: take C=R_+^2, H(x)={(x_1,x_2),(1-x_1,-1-x_2)}, y=0. The point y is locally weakly minimal, yet no single index j blocks both v=(-1,-1) and v=(1,1); meanwhile Theta_l(0)=0 and v_l(0)=0 because max(phi(v),phi(-v))=||v||_infty. Thus Theorem 3.1 and Proposition 3.2 are false as written, and the claimed connection between the algorithms' stationarity test and weak mi","section":"Section 3 (Definition 3.4, Proposition 3.2, Theorem 3.1)"},{"comment":"The proof of Theorem 4.1 is logically misdirected. After assuming that no non-singleton interval satisfies the Armijo condition (17), the proof must exhibit a sequence alpha_k -> 0 for which (17) fails for some index j. Instead, the proof writes the Armijo inequality as holding and then derives a contradiction from (19). This does not establish the existence of a step-size, so the well-definedness of Step 4 of Algorithm 1 is not proved. The existence statement is true and can be proved by a direct expansion using convexity of g and the fact that the vector (1-alpha)D + small terms lies in -int(C) for D in -int(C), but the proof as presented is invalid.","section":"Section 4.1 (Theorem 4.1 and Algorithm 1)"},{"comment":"The claimed O(1/sqrt(k)) complexity does not follow from the displayed results. Theorems 5.3 and 5.4 only bound |Theta_l(x^k)| < C/k for some best iterate k. Since Theta_l(x^k) = F_k(v_l^k) + (l/2)||v_l^k||^2 with F_k(v_l^k) <= 0, a bound on |Theta| does not by itself bound ||v_l^k||. The missing step is the strong convexity of v -> F_k(v)+(l/2)||v||^2: evaluating the strong convexity inequality at v=0 yields (l/2)||v_l^k||^2 <= -Theta_l(x^k) = |Theta_l(x^k)|. This step should be stated explicitly; without it, the abstract's O(1/sqrt(k)) claim is unsupported.","section":"Section 5 (Theorems 5.3, 5.4 and following paragraph)"}],"minor_comments":[{"comment":"The function phi_l is introduced as phi_l(a,v; y) but later referred to as phi_l(v;a,y); make the argument order consistent throughout.","section":"Equation (11) and surrounding text"},{"comment":"The lemma assumes lower semi-continuity of h_j in its proof, but this assumption is not stated in the lemma. Please add it to the hypotheses.","section":"Lemma 3.1"},{"comment":"The 'if and only if' claim for weak minimality is stated without proof. Either prove it or indicate precisely where it follows from Lemma 3.1.","section":"Remark 3.1"},{"comment":"The notation h_{a_k,j} is used before it is defined. Define the indexing map (k,j) -> a_{k,j} explicitly.","section":"Algorithm 1, Step 4"},{"comment":"Several entries in the performance table appear misaligned, making it hard to associate min/mean/max columns with Algo 1 and Algo 2. Please reformat.","section":"Table 2"},{"comment":"Theorems 5.3 and 5.4 provide best-iterate bounds; the text should explicitly state that the O(1/sqrt(k)) rate is a best-iterate rate, not a rate for the last iterate.","section":"End of Section 5"}],"recommendation":"major_revision","confidential_remarks":"The central algorithm may well be salvageable: the quantity Theta_l is actually aligned with the correct stationarity condition (forall a, forall v, exists j), and the convergence and complexity analyses can likely be repaired once Definition 3.4 and the proofs of Theorem 3.1 and Theorem 4.1 are corrected. I therefore recommend major revision rather than rejection. The numerical section is largely independent of the faulty equivalence, but the authors should verify that their implementation's stopping criterion indeed targets the corrected stationarity notion."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: the stress-test note is correct, and it hits a load-bearing part of the paper. Definition 3.4 defines stationarity as, for each partition element a, ∃j such that ∀v, D_v h_{a_j}(y)∉-int(C). But the proof of Proposition 3.2 only establishes the weaker ∀v∃j, and the proof of Theorem 3.1 (the direction (ii)=> (i)) also gives ∀v∃j, not ∃j∀v. The affine example in the stress test is clean: H(x)={(x1,x2),(1−x1,−1−x2)} at 0 is locally weakly minimal, fails Definition 3.4, yet Θ_l(0)=0. So Theorem 3.1 and Proposition 3.2 are false as written. The good news: redefining stationarity as ∀a∀v∃j fixes the connection, and the algorithms’ stopping criterion Θ_l=0 converges exactly to that corrected condition. This looks like a genuine error, not a shallow one, but it is also a repairable one.\n\nWhat is new and useful: this is the first proximal-gradient treatment of set-valued maps of finite cardinality with nonsmooth composite terms; the Θ_l/v_l subproblem is a natural descent-direction scheme; the claimed removal of Bouza et al.’s regularity assumption is interesting; and the O(1/√k) best-iterate complexity rate is plausible. The numerical experiments are extensive, though there is no code or data.\n\nSoft spots, in order:\n1. The quantifier error above. It is load-bearing and must be fixed before the results can be trusted.\n2. Theorem 4.1’s proof has sign/inequality problems: the contradiction assumption seems to assume the Armijo condition holds when it should fail, and the subsequent ⪯_C chain goes in the wrong direction. The existence of an Armijo step-size is likely recoverable by a standard argument, but the written proof is not correct.\n3. The O(1/√k) conclusion is asserted from a best-iterate bound on Θ_l; the short step |Θ_l|≥(l/2)||v_l||² that gives ||v_l||=O(1/√k) is not written down.\n4. The introduction claims there is no proximal gradient method for the vector problem (VOPC), while citing [2], [4], and [27] that do exactly that. That overclaim should be removed.\n5. Assumption 5.1 is doing real work (image-space lower boundedness). It is an assumption, not a consequence, but it is a standard kind of level-boundedness and not itself a red flag.\n\nRecommendation: send to peer review. The core idea is sound, the flaws are specific and repairable, and the topic is timely for the set-optimization community. I would not cite the current version in my own work, but I would engage with a revision.","headline":"The proximal-gradient framework for finite-cardinality set optimization is worth engaging, but the stationarity characterization has a quantifier-reversal error that breaks the claimed link to weak minimality as stated.","tokens_in":30865,"tokens_out":11866,"would_cite":false,"duration_ms":115517,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C47","90C29","49J53","90C46"],"pacs":[],"model":"deepseek-v4-flash","headline":"Set optimization with finite-cardinality set-valued maps is tackled by two proximal gradient methods—one with an Armijo-type line search, one without—that converge globally to stationary points under a mild image-space boundedness assumptio","keywords":["set optimization","lower set-less ordering","proximal gradient method","Armijo line search","stationarity condition","set-valued maps of finite cardinality","global convergence","O(1/√k) convergence rate"],"falsifier":"Run Algorithm 1 on a smooth set optimization problem (g_j=0) whose solution point is non-regular—e.g., a problem where the active minimal-set indices change at the optimum—and check whether ∥v_l^k∥ tends to 0 from a starting point in its domain. If the iterates consistently fail to converge or converge to a non-stationary point, the claim that regularity can be dropped would be falsified.","tokens_in":29832,"feed_emoji":"🎯","tokens_out":7336,"duration_ms":66330,"temperature":0.7,"pith_summary":"This paper introduces the first proximal gradient methods for unconstrained set-valued optimization, where the objective map returns a finite set of vectors ordered by a cone. The authors establish a stationarity condition for these problems—a point is stationary precisely when a certain proximal subproblem over the partition set of active indices returns the zero direction—and build two algorithms around it: one with an Armijo-type line search and one with a fixed unit step-size. They prove both methods converge globally to stationary points under a mild image-space boundedness assumption, and they show that the line-search variant converges without the regularity hypotheses that earlier steepest descent methods required. They also derive a best-iterate O(1/√k) convergence rate for both algorithms and demonstrate the methods on uncertain set optimization test problems. A sympathetic reader would care because this extends composite non-smooth optimization to a setting that models uncertainty and set comparisons, and it supplies the missing convergence and complexity analysis for a whole class of first-order set optimization methods.","feed_headline":"Converge without regularity: proximal gradients for set optimization","feed_subtitle":"Two methods handle non-smooth, uncertain set-valued objectives; the line-search one drops an old regularity assumption.","key_machinery":"The load-bearing object is the stationarity function Θ_l and its minimizer v_l. For a point ȳ, the partition set P̄_ȳ collects, for each enumeration of the minimal set M(H(ȳ),C), one choice of index for each minimal value; the scalarizing function φ(w)=sup_{z∈Z} zᵀw converts cone inequalities into real inequalities. Θ_l(ȳ) is the minimum over partition elements a and directions v of max_j φ(∇f_{a_j}(ȳ)ᵀv + g_{a_j}(ȳ+v) – g_{a_j}(ȳ)) + (l/2)||v||². This single function does three jobs: it characterizes stationarity (Θ_l=0 iff stationary), it produces a C-descent direction v_l(ȳ) at non-stationary points, and its value bounds the per-iteration decrease of the scalar gap ξ∘H, which is what tele","core_discovery":"The central claim is that Algorithms 1 and 2 generate iterates whose stationarity measure Θ_l(x^k) tends to zero (equivalently, the proximal descent directions v_l^k vanish) for any starting point in dom(H), under Assumption 5.1: every ⪯ℓ-decreasing sequence of images has a bounded lower bound set. The stationarity measure is defined as the minimum over the partition set P_{x^k} and directions v of the maximum, over the active minimal-set indices, of the scalarized composite proximal term plus (l/2)||v||². The line-search variant achieves this without requiring the regularity of the solution point that the steepest descent method needed; the no-line-search variant instead requires Lipschitz","pith_inferences":["The partition-set reduction is a template: any vector optimization method equipped with a scalar stationarity measure could be lifted to set optimization by minimizing that measure over the partition set at each iterate, suggesting that accelerated or inexact proximal variants may inherit the same or better rates.","The image-space boundedness Assumption 5.1 is likely to hold whenever ξ∘H is coercive on the sublevel set L; practitioners could verify convergence by checking coercivity of the scalarized objective, a condition easier to test than the assumption's stated form.","The authors' observation that step-size collapses below 10^-15 in the line-search variant points to a concrete next test: replacing the Armijo rule with a non-monotone or spectral step could avoid stagnation and preserve the convergence guarantee.","The best-iterate rate suggests that recording the smallest observed |Θ_l(x^k)|, rather than the last iterate, is the right way to use these algorithms in practice; an averaging scheme might convert the bound to a last-iterate rate under convexity."],"forward_implications":["When all non-smooth terms vanish (g_j=0), the line-search algorithm reduces to the steepest descent method for set optimization studied in the literature, and the new convergence proof removes the regularity assumption on the solution point that the earlier analysis imposed.","For p=1 the problem reduces to composite vector optimization with an arbitrary solid cone; the paper thereby supplies the first proximal gradient method for that vector setting, covering objectives not handled by descent methods relying on gradient Lipschitz continuity.","Both methods achieve a best-iterate O(1/√k) rate on the stationarity measure, giving a first complexity benchmark for first-order set optimization methods.","The robust minimax test problems used in the numerical section show the methods can be applied directly to uncertain set-valued optimization, with the image sets shrinking to an optimal set in the lower set-less order.","Because the partition set is finite, the subproblem is a finite min-max, so existing scalar convex optimization solvers can be used inside each iteration."],"fun_headline_variants":["Proximal gradients tame finite set optimization without regularity","Set optimization: line search beats old regularity requirement","Finite set-valued maps: proximal gradient methods converge globally","O(1/√k) convergence for set optimization with proximal gradients","Line search proximal gradient drops regularity assumption in set optimization"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The convergence and complexity theorems all depend on Assumption 5.1: every decreasing sequence of image sets below the starting image must have some bounded set lying below all of them—a kind of image-space lower boundedness that the paper assumes rather than derives.","fun_headline_variants_meta":{"raw":{"variants":["Proximal gradients tame finite set optimization without regularity","Set optimization: line search beats old regularity requirement","Finite set-valued maps: proximal gradient methods converge globally","O(1/√k) convergence for set optimization with proximal gradients","Line search proximal gradient drops regularity assumption in set optimization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000496,"raw_usage":{"total_tokens":2294,"prompt_tokens":791,"completion_tokens":1503,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":535,"completion_tokens_details":{"reasoning_tokens":1423}},"tokens_in":535,"tokens_out":1503,"duration_ms":12018,"temperature":1.0,"reasoning_tokens":1423,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T02:59:34.936341+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 1 on a smooth set optimization problem (g_j=0) whose solution point is non-regular—e.g., a problem where the active minimal-set indices change at the optimum—and check whether ∥v_l^k∥ tends to 0 from a starting point in its domain. If the iterates consistently fail to converge or converge to a non-stationary point, the claim that regularity can be dropped would be falsified.","supporting_citations":[],"review_version":1}