{"id":"2912ca92-1924-4411-91fd-1b3c864bba23","arxiv_id":"2607.13383","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A Hager-Zhang conjugate gradient method for set optimization is proposed and proved globally convergent without finite-generator or regularity assumptions.","lead":"This paper extends the Hager-Zhang conjugate gradient method to set optimization, where the objective is a finite collection of vector functions and the ordering cone need not be finitely generated. It proves the method's direction is a descent direction and claims global convergence, with numerical tests against two earlier conjugate-gradient variants.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.3's convergence proof is deferred to [12] and omits the key boundedness-of-{||d_k||} step; the central global-convergence claim is not substantiated.","rationale":"The reader's verdict (CONDITIONAL) is reasonable. My load-bearing concern differs from the reader's stated weakest assumption: the bounded-below condition in Theorem 3.1 is overstrong but likely relaxable to the bounded level set in Assumption 4.1, so it does not threaten the central theorem. The genuinely load-bearing issue is the absence of the decisive contradiction in Theorem 4.3: the paper proves ||d_k|| -> inf under the contrary assumption but never proves the needed upper bound, instead deferring to [12]. This is a gap in the central claim itself, not a fixable presentation slip in a supporting theorem. Since the gap is fillable in principle, CONDITIONAL remains the appropriate verdict rather than REJECT or UNVERDICTED; my read does not change the reader's verdict.","tokens_in":23925,"tokens_out":32121,"duration_ms":294870,"concrete_test":"Write out a complete proof of Theorem 4.3. In particular, follow [12, Theorem 2] and supply the lemma that, under the contrary assumption ||u_k||>=l_b, the norms {||d_k||} are bounded. Check each ingredient against the present assumptions, verifying that Lemma 4.3's bound beta_k <= M||x_k-x_{k-1}|| together with the Wolfe conditions and Assumptions 4.1–4.3 forces this boundedness. If such a lemma cannot be derived from the stated assumptions and Lemmas 4.1–4.4, Theorem 4.3 is unproved and the headline claim of global convergence is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—Theorem 4.3, liminf_k ||u_k||=0—is not proved in the manuscript. Its proof is one sentence: 'by contradiction exactly as in [12, Theorem 2].' This deferral is not routine because the setting differs from [12] (set-valued finite-cardinality objective, non-finitely generated cone, Drummond–Svaiter scalarization, active-set-dependent beta_k). The in-text results are insufficient to complete the contradiction: Lemma 4.4 shows under the contrary assumption ||u_k||>=l_b that sum 1/||d_k||^2 < inf, hence ||d_k|| -> inf. The missing half is a proof that {||d_k||} is nevertheless bounded, which in [12, Theorem 2] is a dedicated argument. The paper's conclusion even says 'it has been reported that {||d_k||} is bounded (Theorem 4.3)'—circular, since Theorem 4.3 is the assertion being proved. Without boundedness of {||d_k||}, no contradiction is derived, and global convergence does not follow. The reader's flagged bounded-below assumption in Theorem 3.1 is real but repairable; the omitted final step is the load-bearing gap.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a nonlinear Hager-Zhang conjugate gradient method for unconstrained set optimization problems of the form F(x) = {f^1(x), ..., f^p(x)} with the lower set preorder, where K is a closed convex pointed solid cone. The authors introduce Wolfe conditions based on the Drummond-Svaiter scalarization, define an HZ conjugate parameter, prove a sufficient descent property and a Zoutendijk-type condition, and claim global convergence in the sense liminf_k ||u_k|| = 0. Numerical experiments compare the method with PRP and HS variants on several test problems, including a nonpolyhedral cone. The main convergence theorem is not proved in the manuscript but deferred to a prior vector-optimization result.","tokens_in":1531,"tokens_out":1655,"duration_ms":187679,"significance":"If the central convergence claim were fully established, the paper would be a useful contribution: it extends Hager-Zhang conjugate gradient methods from vector optimization to set optimization, and it does so without requiring a finite generator of the ordering cone or regularity at a solution. The explicit algorithm, the descent theorem, and the Zoutendijk-type theorem are valuable pieces, and the numerical experiments include a nonpolyhedral cone, which is a nice test of the advertised generality. However, the paper's main global-convergence result, Theorem 4.3, is not established as written, and the line-search existence theorem relies on assumptions that are not part of the convergence framework. These gaps are load-bearing, so the contribution is currently conditional.","major_comments":[{"comment":"The proof of the main claim, liminf_k ||u_k||=0, is a single sentence: 'by contradiction exactly as in [12, Theorem 2].' The material prepared in Lemmas 4.3 and 4.4 only shows, under the contrary assumption ||u_k|| >= l_b, that sum 1/||d_k||^2 < infinity and sum ||r_k - r_{k-1}||^2 < infinity. It does not establish the boundedness of {||d_k||}, which in [12, Theorem 2] is a separate and essential step needed to obtain a contradiction. Without that bound, no contradiction follows. The conclusion section even states that '{||d_k||} is bounded (Theorem 4.3)', but Theorem 4.3 is the assertion being proved, not a result about {||d_k||}. This is a load-bearing omission: the central convergence claim is unsubstantiated in this manuscript.","section":"Section 4, Theorem 4.3"},{"comment":"Theorem 3.1 assumes there exists a bounded set B with B <=_l F(x) for all x (stated as x in R^m, probably R^n). This is a global bounded-below assumption, stronger than Assumptions 4.1 and 4.3, and it is not included in the convergence theorem or in the well-definedness discussion. The well-definedness paragraph invokes Theorem 3.1 to assert existence of a Wolfe step at every iteration, but under the assumptions used for convergence this existence is not established. Additionally, the proof contains an invalid inference: from f_j(x+alpha_k d) - ... notin -K for all k, the text concludes the limit is not in -K. A limit of points outside a closed set need not lie outside that set; the later contradiction is obtained by showing the limit lies in -int K, so the argument can be repaired, but as written it is not rigorous.","section":"Section 3, Theorem 3.1 and well-definedness of Algorithm 1"},{"comment":"The first displayed inequality in the proof of Lemma 4.4, sum 1/||d_k||^2 <= (1/l_b) sum ||u_k||^4/||d_k||^2, is false in general. For example, if l_b = 0.1 and ||u_k|| = l_b, the right-hand side is 0.1/||d_k||^2, which is smaller than 1/||d_k||^2. The intended estimate can be repaired by using 1/||d_k||^2 <= ||u_k||^4/(l_b^4 ||d_k||^2), yielding the later constant 1/l_b^4. As written, the proof of the first part of (38) is invalid, though repairable.","section":"Section 4, Lemma 4.4, Eq. (38)"}],"minor_comments":[{"comment":"The statement says 'for all x in R^m'; this should be 'x in R^n'.","section":"Section 3, Theorem 3.1"},{"comment":"In the proof, 'hat S := S - J(F(x0))' mixes a set with a scalar; the intended quantity is J(S) - J(F(x0)). The subsequent division by sigma-1 is also confusing; the convergence of the negative series follows from bounded-belowness of its partial sums, and the argument should be rewritten.","section":"Section 4, Theorem 4.2"},{"comment":"Continuity of the mapping d_zeta is deferred to a proof in [5]. Since this continuity is used to ensure boundedness of u_k, a short self-contained argument would improve readability.","section":"Section 4, Lemma 4.2"},{"comment":"The proof writes 'sigma max_j phi(...) < max_j phi(...)' though the Wolfe condition gives '>='. The strict inequality is not needed; use '<=' throughout.","section":"Section 4, Theorem 4.2"},{"comment":"The text contains the typo 'Hazer-Zhang'; it should be 'Hager-Zhang'.","section":"Section 3"}],"recommendation":"major_revision","confidential_remarks":"The main gap appears repairable: the authors need to supply the missing boundedness-of-directions argument adapted from [12, Theorem 2] to the active-index-dependent beta_k and the set-valued objective. If that proof is provided and the line-search existence theorem is reconciled with the convergence assumptions, the paper may become publishable. I also note the heavy reliance on the authors' own prior work [11,23]; this is defensible in context, but the novelty of the present contribution should be stated more crisply."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper extends the Hager-Zhang conjugate gradient method to set optimization with finite-cardinality set-valued objectives, dropping finite-generation and regularity assumptions. The main novelty is the beta_k in (18) and its descent argument; the descent theorem (4.1) and the Zoutendijk condition (4.2) are actually proved in the text, and the numerics cover a non-finitely generated cone. That is real work and deserves a serious referee. But the central global-convergence claim is not proved. Theorem 4.3 is one sentence: 'by contradiction exactly as in [12, Theorem 2].' The contradiction requires a boundedness result for {||d_k||} that is not in the paper. Lemma 4.4 only gives sum 1/||d_k||^2 < infinity, which is consistent with ||d_k|| going to infinity. The text even says 'it has been reported that {||d_k||} is bounded (Theorem 4.3)'—circular, since Theorem 4.3 is the assertion being proved. This is not a routine deferral: the setting differs from [12] in the finite-cardinality set-valued objective, the active-set-dependent beta_k, and the non-finitely generated cone. The missing step is load-bearing.\n\nThere is also a smaller issue in Theorem 3.1. The proof infers from a_k not in -K and a_k -> a that a is not in -K, but -K is closed; the limit can lie in -K. That step needs repair. The theorem also assumes a global bounded-below set B, which is stronger than the usual bounded level set and not implied by Assumption 4.1. The parameter ordering is stated inconsistently (rho in (sigma,1) versus sigma in (rho,1) used later). These are fixable.\n\nWhat I like: the descent proof is self-contained; the method avoids regularity at the optimal point and finite-generation of the cone; the numerical comparison includes a genuinely non-finitely generated cone; and the authors are honest about cases where HZ does not win. There is no code or data, so the experiments are a sanity check rather than strong evidence.\n\nBottom line: for people working on conjugate gradient methods in vector/set optimization, this is a relevant contribution if the missing proof is supplied. I would send it to peer review with a request for major revision; I would not accept it in its current form.","headline":"Useful extension of HZ-CG to set optimization, with genuinely proved descent results, but the main global-convergence theorem is deferred to another paper and misses a load-bearing boundedness step; warrants major revision rather than acceptance.","tokens_in":24757,"tokens_out":3718,"would_cite":false,"duration_ms":37406,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49J53","90C29","90C47"],"pacs":[],"model":"deepseek-v4-flash","headline":"An HZ-type conjugate gradient method converges to stationarity for set optimization with finite-cardinality set-valued objectives, without requiring a finitely generated ordering cone or regularity at the solution.","keywords":["set optimization","lower set preorder","conjugate gradient method","HZ parameter","line-search conditions","scalarization function","global convergence","stationary point"],"falsifier":"Run Algorithm 1 on the infinite-cone test problem of Example 5.1 from a different initial point in [−14, −7], tracking ||u_k|| to machine precision; if a positive lower bound on ||u_k|| (or a limit cycle with ||u_k|| ≥ ε) appears, the liminf-zero conclusion is false. As a smaller check, the scalar-vector instance F(x) = {(0,x)} with K = R_+^2 and descent direction d = −1 satisfies the sufficient-decrease inequality for every α > 0, so no curvature step exists unless the global bounded-below assumption is in force.","tokens_in":23762,"feed_emoji":"🧮","tokens_out":11201,"duration_ms":110511,"temperature":0.7,"pith_summary":"The paper studies unconstrained set optimization where the objective is a set F(x) = {f^1(x), ..., f^p(x)} and the comparison is the lower-set order. It introduces an HZ-type nonlinear conjugate gradient method and proves that, under boundedness and Lipschitz assumptions, the algorithm is well-defined and produces a sequence whose stationarity gaps satisfy liminf ||u_k|| = 0. The advance over earlier set and vector conjugate gradient methods is that neither the ordering cone need be finitely generated nor must a regularity condition hold at the optimal point. The proof supplies the missing step-length existence theorem, a sufficient-descent inequality, and a summability condition that powers the convergence argument.","feed_headline":"Conjugate gradient method converges for set optimization","feed_subtitle":"An HZ-type update needs no finitely generated ordering cone or regularity at the optimum, reaching stationarity asymptotically.","key_machinery":"The scalarization functional φ(y) = sup{w^T y : w ∈ C} with C = {w ∈ K* : w^T e = 1} turns lower-set order inequalities into scalar inequalities. At each iterate, minimizing V_x(a,d) = max_j φ(∇f^{a_j}(x)^T d) + 1/2 ||d||^2 over the finite set of active index partitions yields the pair (a_k, u_k); the norm ||u_k|| measures nonstationarity. The HZ parameter β_k, built from differences of φ along the previous direction and safeguarded by μ > 1/2, guarantees a sufficient descent condition with constant 1 − 1/(2μ), which powers the summability result and the contradiction proof of convergence.","core_discovery":"The paper's central claim is that the HZ conjugate-gradient update, with the scalar parameter truncated at zero and a line search satisfying sufficient-decrease and curvature conditions, solves the lower-set preorder set optimization problem even when the ordering cone is not finitely generated and no regularity is imposed at the optimum. Under bounded level-set and Lipschitz assumptions, the generated sequence satisfies liminf_{k→∞} ||u_k|| = 0, meaning a subsequence of iterates becomes stationary. Because vector optimization is the p = 1 special case, the result simultaneously covers unconstrained vector optimization with no finite-generation or regularity restrictions.","pith_inferences":["The sufficient-descent proof is largely structural, so a natural testable extension is to port other conjugate-gradient parameter families to the same set-optimization setting without finite generation or regularity, using the same scalarization machinery.","The line-search existence theorem relies on a global bounded-below set B, which is stronger than a bounded level set; a concrete extension is to replace B by a one-dimensional boundedness condition along each descent ray and check whether the maximal-step argument still yields a line-search step.","The asymptotic convergence (liminf ||u_k|| = 0) does not identify a particular weak minimal point; a future refinement could combine the method with a restart or continuation strategy to force full convergence to a weak minimum rather than merely a stationary point.","Because the scalarization depends on the chosen interior point e, the practical behavior of the algorithm is tied to e; comparing different choices of e on the infinite-cone example may reveal a selection criterion that improves the rate of convergence."],"forward_implications":["At every nonstationary iterate, the HZ-generated direction is a K-descent direction, so the method always finds a way to lower the set objective.","A step length satisfying the strong line-search conditions (sufficient decrease and curvature) exists along a K-descent direction when the set-valued image is bounded below by a fixed bounded set in the lower-set order.","The directions and step lengths satisfy a summability condition, which is the engine for asymptotic convergence to stationarity.","Since p = 1 recovers vector optimization, the same guarantee applies to unconstrained vector optimization without finite generation of the ordering cone or regularity at the optimum.","On the paper's test problems, the HZ variant reaches the stopping tolerance with fewer iterations and less runtime than the compared conjugate-gradient variants in most cases, with the paper noting some instances where it does not."],"fun_headline_variants":["HZ conjugate gradient converges for set optimization without extra conditions","Set optimization: HZ method converges with no regularity or generator","Hager-Zhang CG converges for set optimization without cone generator","Set optimization: HZ method needs no finite generator or regularity","HZ method attains stationary point in set optimization without regularity"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The proof that the line search always succeeds assumes there exists a fixed bounded set lying below every value F(x) in the lower-set order; without this global bounded-below condition, the algorithm may be undefined.","fun_headline_variants_meta":{"raw":{"variants":["HZ conjugate gradient converges for set optimization without extra conditions","Set optimization: HZ method converges with no regularity or generator","Hager-Zhang CG converges for set optimization without cone generator","Set optimization: HZ method needs no finite generator or regularity","HZ method attains stationary point in set optimization without regularity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00085,"raw_usage":{"total_tokens":3525,"prompt_tokens":726,"completion_tokens":2799,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":470,"completion_tokens_details":{"reasoning_tokens":2714}},"tokens_in":470,"tokens_out":2799,"duration_ms":19885,"temperature":1.0,"reasoning_tokens":2714,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T05:20:03.280265+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 1 on the infinite-cone test problem of Example 5.1 from a different initial point in [−14, −7], tracking ||u_k|| to machine precision; if a positive lower bound on ||u_k|| (or a limit cycle with ||u_k|| ≥ ε) appears, the liminf-zero conclusion is false. As a smaller check, the scalar-vector instance F(x) = {(0,x)} with K = R_+^2 and descent direction d = −1 satisfies the sufficient-decrease inequality for every α > 0, so no curvature step exists unless the global bounded-below assumption is in force.","supporting_citations":[],"review_version":1}