{"id":"9d9cb64c-b46e-462e-bfdc-8fb8cfb1a837","arxiv_id":"1909.01422","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A successive continuation method using Fischer-Burmeister complementarity functions locates KKT points for equality and inequality constrained optimization, including from infeasible initial guesses.","lead":"This paper extends a numerical trick, called successive continuation, to find optima of problems with both equality and inequality constraints, even when the starting guess breaks the constraints. It replaces inequality conditions with smooth relaxed equations and then walks them to the true solution, with tests on pendulum optimal control and boundary-value problems.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Global reachability of each continuation stage is assumed, not guaranteed; the paper's own examples require manual intervention or lucky branch alignment.","rationale":"The paper's central claim is carefully hedged with 'may be found' and 'found to be possible,' and the analytical finite-dimensional example plus numerical demonstrations provide real support for the existence of successful staged searches. However, the method's practical validity rests on the ability of each continuation stage to drive its target parameter to the desired value within the chosen domain, and to pass through singular points on zero-level surfaces when they occur. The authors themselves concede in Section 7 that this is only tacitly assumed and that their examples show failures. The failure in Section 3's U+{/-} case is explicit: the secondary branch is negatively aligned, the pseudo-arclength algorithm cannot proceed, and the authors simply switch to another initial point. Similarly, the FP2 path in Section 6.1 terminates at a singular point on G_int=0 before nu_p2 reaches 0, requiring a larger domain and opposite-direction continuation. Lemma A.6 provides a local branch on the zero surface but says nothing about the alignment of tangents needed for automatic bypass, which Figure 3 shows is essential. This is not a claim of internal inconsistency; the lemmas are local and the examples are honest about their difficulties. Rather, the concern is that the advertised capability, especially from infeasible initial guesses, is not backed by any algorithmically checkable condition for success. The reader's weakest assumption correctly identifies this reachability requirement, and the paper's own text confirms it is not guaranteed. I therefore see no reason to change the conditional verdict; the concern is real but already accounted for in the reader's assessment. Other potential objections, such as the nonsmoothness of the Fischer-Burmeister function or the unreleased modified COCO core, are either addressed by the authors or secondary to the reachability issue.","tokens_in":23776,"tokens_out":7741,"duration_ms":78787,"concrete_test":"Implement the Lemma A.6 branch onto the G_int=0 surface at the singular point encountered in the FP2 continuation of Section 6.1, and continue until nu_p2=0; if the branch terminates before the target or requires leaving the zero surface again, the tacit reachability assumption is not satisfied even with explicit branch switching, and the advertised generality of the method is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 7 concedes: 'It has been tacitly assumed that each successive stage of continuation is able to drive the appropriate continuation parameters to their desired values, preferably monotonically,' and that examples show this may fail. This assumption is load-bearing for the central claim of a successful search from infeasible initial guesses, because every stage of the proposed algorithm depends on reaching a target value of nu or kappa within the chosen computational domain. The finite-dimensional example in Section 3 provides a concrete failure: starting from u0 in U+{/-}, continuation terminates at the singular point (7/5, 7/5) on G2=0, where the secondary branch is negatively aligned, so the pseudo-arclength algorithm cannot bypass it and the authors instead restart from a different infeasible point. In Section 6.1, starting from FP2, the stage driving nu_p2 to 0 ends at a singular point on G_int=0 where Newton's method fails; the authors note that success would require a larger computational domain and continuation in the other direction. Lemma A.6 establishes only a local one-dimensional manifold on the zero-level surface of G_k; it does not ensure that the tangent at the singular point is positively aligned with the incoming branch, which Figure 3 identifies as the condition for automatic bypass. Thus the theoretical lemmas do not provide criteria under which the required stages are realizable, and the success of the method remains contingent on the uncontrolled branch geometry of the specific problem instance.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper generalizes the successive continuation paradigm of Kernévez and Doedel to constrained optimization problems with simultaneous equality and inequality constraints. The authors augment the KKT conditions with the Fischer-Burmeister complementarity function, define a family of restricted continuation problems, and use staged continuation to drive relaxation parameters to their limiting values. They prove several local lemmas (Appendix A), implement the method by extending the COCO software package, and demonstrate it on a finite-dimensional example, a two-point boundary-value problem with an integral inequality constraint, and an optimal control problem. The central claim is that optima can be found from easily initialized stages, without seeding nonzero Lagrange multipliers in the first stage, and starting from points that may violate the inequality constraints.","tokens_in":24069,"tokens_out":3352,"duration_ms":35224,"significance":"If the central claim is fully established, the paper offers a genuinely useful extension of a known continuation-based optimization technique, with potential applications to boundary-value and optimal control problems. The authors provide analytic and numerical evidence for the finite-dimensional motivating example, a working software implementation, and explicit discussion of the method's limitations, which is commendable. The local lemmas in Appendix A are plausible and provide a useful framework for understanding branch points and secondary branches. However, the paper's headline claim about a successful search from infeasible initial guesses is not fully supported by the theoretical analysis, because the global reachability of each continuation stage is assumed rather than proved, and the numerical examples themselves reveal multiple failures that require manual intervention or fortuitous branch alignment.","major_comments":[{"comment":"The paper's own concluding section admits that the success of the method depends on an unproved assumption: 'It has been tacitly assumed that each successive stage of continuation is able to drive the appropriate continuation parameters to their desired values, preferably monotonically. But the examples showed that this may not be possible within a given computational domain, or may only be possible by occasionally bearing in a direction away from the desired values.' This is a load-bearing assumption for the abstract's claim that 'a successful search for optima is found to be possible also from an infeasible initial solution guess,' because every stage of the algorithm requires reaching a target value of a component of ν or κ. The finite-dimensional example in Section 3 explicitly fails for the initial point u0 in U+/-: continuation terminates at the singular point (7/5, 7/5) on G2=0, the secondary branch is negatively aligned, and the authors restart from a different infeasible point. Similarly, in Section 6.1 the run starting from FP2 fails to reach νp2=0, terminating at a singular point on G_int=0, and the authors note that success would require a larger computational domain and continuation in the other direction. These failures are not isolated artifacts: they concern the central mechanism by which the method is supposed to reach a KKT point.","section":"Section 7 and Sections 3, 6.1"},{"comment":"Lemma A.6 establishes only a local one-dimensional manifold on the zero-level surface of G_k, with σ_k and σ_P nonzero for η1 close to 0. It does not provide any criterion ensuring that the secondary branch can be followed to η1=1, nor that the tangent directions at the singular point are positively aligned so that the pseudo-arclength algorithm can automatically bypass the singularity. The text in Section 4 states 'Suppose that no element of G_Z equals 0 along this manifold for η1∈[0,1]' and 'Lemma A.6 allows for the possibility of branch switching,' but these suppositions are not derived from the lemmas. The example in Section 3 shows exactly the opposite: at the singular point on G2=0 the secondary branch is negatively aligned and the pseudo-arclength algorithm cannot bypass it. Without a global reachability or alignment condition, the proposed procedure remains a heuristic with local justification, not a rigorous guarantee of successful search.","section":"Appendix A, Lemma A.6 and Section 4, paragraph 5"},{"comment":"The paper describes a 'rigorous framework' in the conclusions, but the lemmas in Appendix A are proofs by sketch relying on 'generically' and continuity arguments (e.g., Lemma A.3), and they do not quantify how far the secondary branches can be continued. Moreover, the algorithm requires the user to choose the initial point u0, the index sets I and P, the continuation order, and the computational domain; the paper itself states 'We are not able to propose a systematic selection algorithm beyond the principles outlined above.' While user choices are common in continuation methods, the combination of an unproved reachability assumption and the acknowledged need for manual switching or restarts in several examples means the manuscript overstates the degree to which it delivers a reliable search method. The claims should either be weakened to describe the method as a set of locally justified heuristics, or supplemented with additional conditions under which the staged continuation is guaranteed to reach a KKT point.","section":"Section 4, 'Initialization' bullet and Section 7, first paragraph"}],"minor_comments":[{"comment":"The singularity of the Fischer-Burmeister function at (0,0) is identified, and the paper correctly assumes that no active inequality constraints are present at the initial point in the lemmas. The notation κ0 and the role of the set P would benefit from being summarized in a table for readability, but this is not essential.","section":"Section 2, Eq. (2.7)-(2.10)"},{"comment":"The numerical results for the finite-dimensional example are presented through multiple figures with color-coded markers. In the printed text, the colors are referenced (e.g., 'red dots,' 'blue dots'), but a reader without access to the color figures may find it difficult to follow the description. Consider referencing markers by shape or adding a short legend in each figure.","section":"Section 3, numerical results paragraphs"},{"comment":"The text says 'two fold points in the value of κint are encountered on the way to 0 in the final continuation run' but does not explain how the continuation handles these folds. Since the paper's main algorithm relies on the ability to drive κ to zero, a brief remark about the handling of folds would be useful.","section":"Section 6.1, first case"},{"comment":"The paper frequently refers to reference [18] for implementation details of adjoint construction and for previous observations about non-monotone continuation. Since [18] is the authors' own work, the dependence is understandable, but the manuscript should make clear which specific claims rely on [18] and which are new in this paper.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest about its limitations, which is a strength, but the central claim in the abstract is stronger than what the theoretical and numerical evidence support. The global reachability issue appears to be a genuine gap: the lemmas are local, and the examples show failures that require manual intervention. If the authors can reframe the contributions as a locally justified extension with heuristic search properties, or add sufficient conditions that guarantee each stage reaches its target, the manuscript could become acceptable. In its current form, the gap is load-bearing and requires major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is the short version. This is a real extension of the Kernévez–Doedel successive continuation idea to inequality constraints, and the numerical demonstrations support the claim that infeasible starts can work. The central caveat is that the paper's own examples show the staged scheme can fail, and Section 7 concedes the reachability assumption. That assumption is load-bearing, so the headline claim should be read as \"can succeed\" rather than \"is guaranteed.\"\n\nWhat is new: the Fischer–Burmeister complementarity function inside successive continuation, the branch-point lemmas for inequality constraints, and the demonstration of infeasible-start searches for KKT points in BVP-constrained and optimal control problems. The finite-dimensional example is worked analytically and matches the continuation results, which is real evidence. The appendix lemmas give a local picture: branch points exist and secondary branches are locally parameterized by eta_1. That is useful.\n\nThe soft spots are real but not fatal. The lemmas are sketches. They rely on \"generically\" and continuity arguments, and they do not provide checkable conditions for positive branch alignment or for driving nu and kappa to their target values. The stress-test note is right: Lemma A.6 only gives a local manifold on the G_k = 0 surface, not the alignment needed for the pseudo-arclength algorithm to bypass the singular point. The paper itself reports cases where continuation terminates at singular points or domain boundaries and manual restart is needed. That is not fatal for a methods paper, but it means the algorithm is a heuristic with theoretical scaffolding, not a provable method. Also, the modified COCO core is not released, so the BVP and optimal-control results are not immediately reproducible. That should be fixed before publication.\n\nThe citation pattern is fine. Self-citation to the authors' earlier COCO work is appropriate because that is the implementation the paper builds on; there is no circular fitting.\n\nWho this is for: people working on numerical continuation for constrained optimization, especially in COCO, will get a clear framework and worked examples. I would send this to a serious referee. The right outcome is probably major revision, not desk rejection, with referees asking for clearer assumptions and code or data release. The core idea holds up.","headline":"A genuine extension of the Kernévez–Doedel successive continuation framework to inequality constraints, with honest numerical demonstrations and a load-bearing reachability assumption that the paper itself concedes.","tokens_in":24549,"tokens_out":2427,"would_cite":true,"duration_ms":25385,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49K15","49K27","49M05","49M29","90C33","45J05","34B99"],"pacs":[],"model":"deepseek-v4-flash","headline":"Staged continuation reaches constrained optima from infeasible starts","keywords":["constrained optimization","complementarity conditions","Karush-Kuhn-Tucker conditions","successive continuation","boundary-value problems","optimal control","Fischer-Burmeister function","parameter continuation"],"falsifier":"A concrete check: take a smooth constrained problem with a unique KKT point and try every admissible initial point and index order; if every staged path either stalls at a negatively aligned singular point or leaves the computational domain before $\\nu=1$ and $\\kappa=0$ are reached, the claimed search-from-infeasible-start property fails. The paper's own motivating example shows such a stall for one initial choice ($(3,1)$ stalling on $G_2=0$), so the question is whether any problem exhibits this stall for all choices.","tokens_in":23574,"feed_emoji":"🎯","tokens_out":7523,"duration_ms":69349,"temperature":0.7,"pith_summary":"This paper claims that local optima of an objective subject to both equality and inequality constraints can be found by a sequence of easy-to-initialize continuation runs, without ever needing a starting guess for the Lagrange multipliers. The trick is to replace the complementary slackness conditions of the Karush-Kuhn-Tucker theory by relaxed equations built from a complementarity function, then drive the relaxation parameters to zero through successive stages of continuation. The paper proves lemmas showing how branch points in these stages allow nonzero multipliers to appear from a zero-multiplier start, and it shows that the initial guess may even violate the inequality constraints. If true, this removes the main practical barrier--finding an adequate initial guess for the multipliers--for optimization along families of boundary-value problems and in optimal control.","feed_headline":"Staged continuation reaches constrained optima from infeasible starts","feed_subtitle":"Complementarity relaxations let a sequence of easy runs locate KKT points without multiplier guesses.","key_machinery":"The engine of the method is the augmented system $F_{\\mathrm{aug}}=0$ built from (i) the equality constraints $\\Phi(u)=0$ and objective monitor $\\Psi(u)-\\mu=0$, (ii) the adjoint condition $(D\\Phi(u))^*\\lambda+(D\\Psi(u))^*\\eta+(DG(u))^*\\sigma=0$ with normalization $\\eta-\\nu=0$, and (iii) the relaxed complementary conditions $\\chi(\\sigma_i,-G_i(u))-\\kappa_i=0$ for each inequality. The complementarity function $\\chi(a,b)=\\sqrt{a^2+b^2}-a-b$ vanishes exactly when $a,b\\ge 0$ and $ab=0$, so the equations enforce complementary slackness in the limit $\\kappa=0$; its nonsmooth zero contour at the origin is what creates the singular points where branch switching to nonzero multipliers can occur. The continuation parameters $\\nu$ and $\\kappa$ are frozen or released stage by stage, and the lemmas show when the corresponding solution sets are one-dimensional manifolds, when a stationary point of the objective is a branch point, and when a secondary branch proceeds to the desired terminal values.","core_discovery":"The central claim is that a generalized successive continuation paradigm solves constrained optimization problems with simultaneous equality and inequality constraints by seeking roots of an augmented system that includes the original equations, adjoint conditions linear and homogeneous in the Lagrange multipliers, and relaxed complementarity conditions of the form $\\chi(\\sigma_i,-G_i(u))=\\kappa_i$, with $\\chi(a,b)=\\sqrt{a^2+b^2}-a-b$. Starting from a solution of the equality constraints with all multipliers set to zero and with $\\kappa_i=0$ for inactive inequalities and $\\kappa_i>0$ for violated ones, the algorithm follows one-dimensional solution manifolds, detects fold points or singular points, switches branches to manifolds on which multipliers become nonzero, and finally drives the relaxation parameters $\\kappa_i$ (and normalization parameters $\\nu$) to their KKT limits. The paper's lemmas establish, under rank conditions, that the necessary branch points exist and that the terminal point of the sequence satisfies the KKT conditions. Consequently, the search can begin from a potentially infeasible initial guess and requires no nonzero multiplier initialization.","pith_inferences":["A natural test is to replace the nonsmooth complementarity function by a smooth approximant and add a final continuation stage that drives the smoothing parameter to zero; this could eliminate many of the singular points that force manual branch switching, an extension the paper itself raises.","Because the order in which components of $\\nu$ and $\\kappa$ are driven to their limits is chosen by the user, different orders may converge to different local optima; a systematic ordering rule could be benchmarked on multi-extremum problems.","The framework is stated for finite-dimensional inequality constraints, so a meaningful extension is to discretize infinite-dimensional inequality constraints in a way consistent with the adjoint discretization and test whether the staged convergence still reaches KKT points.","The same complementarity relaxation could be applied to regularize bang-bang optimal control problems, treating the singular limit as the terminal continuation stage; the paper mentions such regularization as a possible combination."],"forward_implications":["For a problem with $d$ unknown design or control variables, the staged construction reaches a KKT point in $d+1$ continuation runs (or $d$ runs when the first stage uses branch switching), independent of the number of constraints.","Zero Lagrange multipliers and an infeasible initial point are enough to start the search; the method does not require a feasible starting guess.","Inequality constraints add new opportunities for branch switching, so they can help rather than merely obstruct the continuation search.","The formulation is compatible with staged-construction continuation software that generates adjoints automatically, so it applies to two-point boundary-value problems and optimal control problems with integral inequality constraints.","If a secondary branch is negatively aligned at a singular point, the standard predictor-corrector continuation algorithm may fail there; the paper notes that manual switching or an alternative initial choice and ordering of stages is then needed."],"supporting_citations":[{"why":"Provides the original successive continuation method for equality-constrained optimization that this paper generalizes to simultaneous equality and inequality constraints.","marker":"[16]"},{"why":"Supplies the generalized Karush-Kuhn-Tucker conditions in Banach spaces that the paper's optimality formulation rests on.","marker":"[1]"},{"why":"Provides the two-point boundary-value problem example used as a numerical test and an earlier demonstration of continuation-based optimization.","marker":"[9]"},{"why":"Establishes the staged construction of adjoints for integro-differential boundary-value problems that the present software extension builds upon.","marker":"[18]"},{"why":"Supplies background on complementarity problems that motivates converting complementary slackness into nonlinear equations.","marker":"[3]"},{"why":"Names the continuation software platform whose staged construction paradigm is matched by the proposed formulation and used for the reported numerical results.","marker":"[20]"}],"fun_headline_variants":["Continuation finds optima without initial multipliers","Infeasible start? Continuation still reaches KKT points","Relaxed complementarity enables continuation to optima","Staged continuation handles equality and inequality","No multiplier seeding needed for constrained optima"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Each continuation stage must actually be able to drive its target parameter ($\\nu$ components to 1, $\\kappa$ components to 0) within the chosen computational domain, which may require successfully bypassing singular points; the paper does not prove this reachability for all problems.","fun_headline_variants_meta":{"raw":{"variants":["Continuation finds optima without initial multipliers","Infeasible start? Continuation still reaches KKT points","Relaxed complementarity enables continuation to optima","Staged continuation handles equality and inequality","No multiplier seeding needed for constrained optima"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000223,"raw_usage":{"total_tokens":1460,"prompt_tokens":950,"completion_tokens":510,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":437}},"tokens_in":566,"tokens_out":510,"duration_ms":5113,"temperature":1.0,"reasoning_tokens":437,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:17:47.655891+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete check: take a smooth constrained problem with a unique KKT point and try every admissible initial point and index order; if every staged path either stalls at a negatively aligned singular point or leaves the computational domain before $\\nu=1$ and $\\kappa=0$ are reached, the claimed search-from-infeasible-start property fails. The paper's own motivating example shows such a stall for one initial choice ($(3,1)$ stalling on $G_2=0$), so the question is whether any problem exhibits this stall for all choices.","supporting_citations":[{"cited_title":"Kern ´evez and E","cited_arxiv_id":null,"evidence_quote":"Provides the original successive continuation method for equality-constrained optimization that this paper generalizes to simultaneous equality and inequality constraints."},{"cited_title":"Li and H","cited_arxiv_id":null,"evidence_quote":"Establishes the staged construction of adjoints for integro-differential boundary-value problems that the present software extension builds upon."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies background on complementarity problems that motivates converting complementary slackness into nonlinear equations."},{"cited_title":"Schilder, H","cited_arxiv_id":null,"evidence_quote":"Names the continuation software platform whose staged construction paradigm is matched by the proposed formulation and used for the reported numerical results."}],"review_version":1}