{"id":"19e3b22c-755e-4c89-8c66-da49729617e6","arxiv_id":"2607.11239","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":5.5,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"A globally convergent augmented-Lagrangian method for constrained nonlinear least-squares with structured SR1 Hessian approximation and direct handling of linear constraints via gradient projection.","lead":"The paper gives a trust-region augmented-Lagrangian solver for nonlinear least-squares with mixed linear and nonlinear constraints, using a structured SR1 Hessian of the AL and gradient projection on the linear set. It proves global first-order convergence and shows competitive runtimes versus IPOPT and Percival on 79 test instances.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reduced-gradient gap already flagged by the reader.","rationale":"The central claim is the standard AL global-convergence result specialized to structured NLS (Theorem 9) plus the empirical observation that Hybrid-SR1 TRAULLS is competitive with IPOPT in CPU time. The only place where the proved statement and the shipped code diverge is exactly the reduced-gradient versus projected-gradient distinction already identified by the reader; the paper is transparent about the required sign condition. No hidden assumption in the Cauchy-point analysis, the structured secant equation, or the hybrid switching rule undermines the argument under the stated hypotheses. Dense Hessian storage is acknowledged as a scalability limit, not a correctness flaw. Consequently the reader’s ACCEPT / HIGH-confidence assessment stands; no verdict adjustment is warranted.","tokens_in":28415,"tokens_out":492,"duration_ms":6222,"concrete_test":"On the 79-instance suite, after each accepted outer iterate record both ∥P_Ω[x_K−∇Φ_K]−x_K∥ and ∥P_T(x_K)[∇Φ_K]∥ together with the signs of the bound multipliers recovered from the normal-cone residual. If the two norms differ by more than a small multiple of machine epsilon on any successful run, or if any bound multiplier has the wrong sign at a declared solution, the numerical claims rest on a weaker criticality certificate than the theorems.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader’s weakest_assumption correctly isolates the only material gap between the proved statement (Theorem 9 / Lemma 8 under the projected-gradient residual (2.6)) and the code that produces the numerical claims (reduced-gradient test (2.10)). The paper itself states the precise condition under which the two measures coincide (correct sign of bound multipliers at every outer iterate; end of §3). Under that condition the theorems apply to the iterates the implementation produces; without it only the tangential component is controlled. No further load-bearing inconsistency appears in the global-convergence argument (Assumptions 1–9, Lemmas 1–8, Theorems 4–5 and 9) or in the experimental design. The hybrid-SR1 competitiveness claim is an empirical observation on a fixed 79-instance set and is not overstated.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes TRAULLS, an augmented Lagrangian algorithm for nonlinear least-squares problems with mixed nonlinear and linear (equality/bound) constraints. Nonlinear constraints are penalized via the AL function (2.1), while linear constraints are retained and handled by a trust-region gradient-projection inner solver (Cauchy point along the projected path plus projected CG minor iterations). The AL Hessian is approximated by a Gauss–Newton term plus a structured SR1 correction of the second-order residual/constraint contribution (secant equation (2.26), update (2.27)), with an optional hybrid switch (2.29). Global convergence of the inner process (Theorem 5) and of the outer AL iterates (Theorem 9 / Lemma 8) is established under standard assumptions (twice continuous differentiability, full-rank active linear constraints, compact level sets, controlled Hessian growth, and rank of C(x*)N* ≥ nc). Numerical experiments on 79 instances compare Hessian variants and show the hybrid-SR1 version competitive with IPOPT in CPU time and superior to Percival.","tokens_in":28709,"tokens_out":903,"duration_ms":8826,"significance":"The work cleanly combines two mature lines—AL methods with gradient-projection subproblem solves and structured quasi-Newton updates for least-squares—into a solver that treats general nonlinear equalities/inequalities together with linear equalities and bounds. The structured secant equation for the AL second-order term and the polyhedral adaptation of the Cauchy-point path are concrete technical contributions. Global-convergence proofs follow the Conn–Gould–Toint / Conn–Gould–Sartenaer–Toint templates after careful reformulation for polyhedra, and a public Julia implementation (TRAULLS) with reproducible benchmarks is provided. The hybrid-SR1 variant is shown to be competitive with a state-of-the-art interior-point solver on a standard test set while exploiting first-order structure only. These strengths make the paper a useful addition to the constrained NLS literature.","major_comments":[{"comment":"End of §3: the proved outer convergence (Theorem 9 / Lemma 8) is stated for the projected-gradient residual (2.6), while the implementation and all numerical claims use the reduced-gradient test (2.10). The authors correctly note that the two coincide only when bound multipliers have the correct sign at every outer iterate; without that condition only the tangential component is controlled and a limit point need not be critical. This is the sole material gap between the theorems and the code. A short clarification—either an explicit additional assumption that the sign condition holds along the generated sequence, a practical safeguard that restores the projected residual when signs are wrong, or a statement that the numerical claims are conditional on the observed sign pattern—would close the gap without altering the rest of the argument.","section":null}],"minor_comments":[{"comment":"§2.5 / Figure 1: the hierarchy of outer / inner / minor iterations is clear, but a one-sentence reminder that the outer index K and inner index k are independent (and that B0 is reset to zero at each outer iteration) would help readers who jump between sections.","section":null},{"comment":"§4.1: the performance profiles are informative; adding the absolute number of failures (already mentioned in the text) to the figure captions would make the profiles self-contained.","section":null},{"comment":"Appendix A, Algorithm 4: the termination test “|A(ti)| = n−m” is correct under Assumption 4, but a brief remark that the trust-region bounds guarantee eventual full activation would remove any ambiguity for readers unfamiliar with the polyhedral path.","section":null},{"comment":"Several minor typos appear (e.g., “nonlinar”, “Compuations”, “Transcations”, “Reasearch”); a careful proof-reading pass would clean them.","section":null},{"comment":"References [37] and [26] both concern NL2SOL-related work; ensuring the year and author list of [37] match the published version would avoid confusion.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The reduced-gradient gap is already acknowledged by the authors and is the only load-bearing issue; it is easily fixed by a short clarifying paragraph. The paper is otherwise solid, well-written, and suitable for a computational optimization journal. No concerns about novelty disclosure or citation pattern."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a competent methods paper that does exactly what it claims. The new piece is the concrete combination: AL outer loop that keeps linear equalities and bounds, polyhedral gradient-projection inner solver (Cauchy + projected CG minor steps), and a structured SR1 correction of the second-order term of the AL Hessian with a hybrid switch. That package is not LANCELOT, not Percival, and not the SQP structured methods they cite. They ship Julia code (TRAULLS) and prove global convergence under standard assumptions by adapting the Conn–Gould–Toint / SBMIN template to polyhedra and the usual AL outer theory.\n\nWhat works: the algorithm description is clear, the structured secant equation for the AL second-order term is cleanly derived, the hybrid switch is tested fairly against GN and plain BFGS/SR1, and the comparison with IPOPT and Percival is honest. Hybrid SR1 is competitive with IPOPT on CPU time on their 79-instance set and clearly better than Percival; they do not oversell it. Dense storage of the structured correction is acknowledged as the large-scale limit. Citations are appropriate and not padded.\n\nThe only material soft spot is the one the reader already flagged: theorems use the true projected-gradient residual; the code uses the reduced-gradient norm. The paper itself states that these coincide only when bound multipliers have the correct sign at every outer iterate. Under that condition the theorems cover the iterates they produce; without it only the tangential component is controlled. That is a known practical gap, not a contradiction of the math. Hyperparameters are fixed once and reported; they do not prop up the theorems.\n\nThis is for people who need a specialized constrained NLS solver or who work on AL / structured quasi-Newton methods. It is not field-redefining, but it is solid subfield engineering with reproducible code and careful analysis. I would send it to referees without hesitation. Worth reading if you care about this niche; I would cite the algorithm and the hybrid-SR1 comparison if I were working on related solvers.","headline":"Solid specialized AL solver for constrained NLS: careful combination of known pieces, honest numerics, one known gap between theorem and code.","tokens_in":29344,"tokens_out":500,"would_cite":true,"duration_ms":5550,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C30","65K05","90C55"],"pacs":[],"model":"grok-4.5","headline":"An augmented Lagrangian method solves constrained nonlinear least-squares by keeping linear constraints explicit and approximating the penalty Hessian with a structured SR1 update.","keywords":["constrained nonlinear least-squares","augmented Lagrangian","structured quasi-Newton","SR1 update","gradient projection","trust region","hybrid Hessian approximation"],"falsifier":"Run the published Julia implementation on a problem whose active-bound multipliers change sign near a limit point and check whether the reduced-gradient residual vanishes while the true projected-gradient residual stays bounded away from zero.","tokens_in":29283,"feed_emoji":"📐","tokens_out":608,"duration_ms":6016,"temperature":0.7,"pith_summary":"Many fitting and parameter-estimation problems are nonlinear least-squares with a mixture of nonlinear constraints and linear equalities or bounds. This paper builds a solver that folds only the nonlinear constraints into an augmented Lagrangian while leaving the linear set untouched, so each outer step reduces to a linearly constrained subproblem that can be attacked by gradient projection and projected conjugate gradients. The Hessian of that subproblem is never formed exactly; instead a Gauss–Newton block is corrected by a structured SR1 update that respects the sum-of-squares structure of both residuals and constraints. Under standard smoothness and rank assumptions the outer sequence is shown to produce first-order critical points. Numerical tests on seventy-nine problems show that a hybrid version of the structured update is competitive in wall-clock time with a leading interior-point solver while remaining matrix-free.","feed_headline":"Constrained least-squares solved by structured AL steps","feed_subtitle":"Hybrid SR1 Hessian keeps linear constraints exact and matches interior-point speed","key_machinery":"The structured secant equation for the second-order remainder of the augmented Lagrangian, Bk+1 sk = (Jk+1 − Jk)⊤ rk+1 + (Ck+1 − Ck)⊤ λ̄k+1, which is realized by an SR1 update and optionally switched with pure Gauss–Newton via a residual-ratio test.","core_discovery":"The paper establishes that an augmented-Lagrangian outer loop combined with trust-region gradient-projection inner iterations and a hybrid structured-SR1 Hessian approximation converges globally to first-order KKT points of a nonlinear least-squares problem subject to mixed nonlinear and linear constraints, and that the resulting algorithm is practically competitive on a standard test set.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["AL outer loop with gradient-projection solves constrained NLS","Hybrid SR1 Hessian keeps linear constraints exact in AL steps","Structured AL algorithm converges for mixed-constraint least-squares","Gradient projection on linear constraints inside augmented Lagrangian","AL method with structured Hessian competitive for constrained NLS"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The code stops when the reduced gradient is small, yet the global-convergence theorem is proved for the true projected gradient; the two coincide only when active-bound multipliers keep the correct sign at every outer iterate.","fun_headline_variants_meta":{"raw":{"variants":["AL outer loop with gradient-projection solves constrained NLS","Hybrid SR1 Hessian keeps linear constraints exact in AL steps","Structured AL algorithm converges for mixed-constraint least-squares","Gradient projection on linear constraints inside augmented Lagrangian","AL method with structured Hessian competitive for constrained NLS"]},"model":"grok-4.5","effort":"low","cost_usd":0.004222,"raw_usage":{"total_tokens":1163,"prompt_tokens":593,"num_sources_used":0,"completion_tokens":60,"cost_in_usd_ticks":42220000,"prompt_tokens_details":{"text_tokens":593,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":510,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":593,"tokens_out":60,"duration_ms":5571,"temperature":1.0,"reasoning_tokens":510,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T06:04:17.424877+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the published Julia implementation on a problem whose active-bound multipliers change sign near a limit point and check whether the reduced-gradient residual vanishes while the true projected-gradient residual stays bounded away from zero.","supporting_citations":[],"review_version":1}