{"id":"0c3872a7-1645-4d72-98fc-43bc61854b30","arxiv_id":"2511.23064","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An exact line search based on bisection makes Newton's method converge reliably in staggered phase-field fracture, resolving benchmark cases that previously failed.","lead":"This paper shows that adding a simple, exact line search step to Newton's method fixes common convergence failures in computer simulations of brittle fracture. It matters because the fix is small, open-source, and makes hard 3D crack simulations that previously stalled run to completion.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 2's printed termination test omits the absolute value, so the implemented line search is not guaranteed to be exact; Proposition 2's proof therefore does not cover the method as specified.","rationale":"The reader's CONDITIONAL verdict is appropriate, but the most load-bearing issue is slightly more elementary than the one highlighted by the reader. The reader correctly notes that Proposition 2 requires exact minimization and that the non-covered cases (reduced-space active-set, star-convex γ*>0) rest on empirical sampling. I agree with that assessment and with the paper's own transparency about those gaps. However, the manuscript's Algorithm 2, as printed, fails the exactness requirement even in the convex regime because its convergence criterion does not take an absolute value. This means the published description of the method does not actually implement the exact line search used in the proof, unless the PETSc source contains an implicit absolute value that the pseudocode omits. That is a concrete, checkable correctness issue in the central argument, not merely a scope limitation. The numerical results and the availability of the PETSc contribution are real supporting evidence, and a missing absolute value is easily fixed, so the verdict should remain CONDITIONAL rather than moving to ACCEPT or REJECT. The reader's weakest assumption overlaps with this concern because it flags exact minimization as a fragile assumption, but the specific mechanism is different, hence 'partial' agreement.","tokens_in":30685,"tokens_out":19742,"duration_ms":204384,"concrete_test":"Inspect the PETSc SNESLineSearchBisection source or run Algorithm 2 verbatim on a strictly convex quadratic with φ'(0)<0 and the unique minimizer at λ=0.8. With the printed termination test, the algorithm returns λ=0.5; with the corrected |φ'| test, it continues and returns λ≈0.8. Additionally, compare the λ* returned by the algorithm for a representative set of Newton iterates in one benchmark against a high-accuracy Brent/golden-section minimizer; any discrepancy beyond the stated atol/ltol confirms that the line search is not exact as published.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central result, Proposition 2, assumes an exact line search that minimizes φ(λ) over [0,1] (Section 3.2, equations (50)-(53)). But Algorithm 2 in Section 3.1 uses the termination test `φ'(λ_l)/||Δw||² ≤ atol` with atol = 1e-12, not `|φ'(λ_l)|/||Δw||² ≤ atol`. For any λ to the left of the root of φ', φ' is negative and is therefore automatically ≤ any positive atol. Consequently, if the first bisection point λ=0.5 lies left of the true minimizer, the algorithm exits immediately at λ=0.5 without bracketing the root. The returned step is not the exact minimizer of (30), so the key inequality (50) and the subsequent residual-norm convergence proof do not apply to the algorithm as written. This is a concrete failure of the exactness assumption even in the strictly convex subproblems that the paper claims to cover. If the PETSc implementation actually uses `abs(φ')`, then the pseudocode is incorrect and should be corrected; if it follows the pseudocode literally, the global convergence guarantee is not valid for the shipped solver. This issue is independent of the acknowledged proof gaps for the reduced-space active-set strategy and the star-convex model with γ* > 0 (Remark 4, Section 4.2), which further limit the 'exact line search is all you need' claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses iterative non-convergence of Newton solvers inside staggered (alternate minimization) schemes for phase-field brittle fracture. It proposes to globalize each Newton solve by an 'exact' line search that minimizes the restricted energy φ(λ) along the Newton direction, using a bisection algorithm on the directional derivative φ'(λ). Two theoretical results are stated: Proposition 1 gives linear convergence of the bisection root for strictly monotone φ', and Proposition 2 gives convergence of the residual norm to zero for Newton's method with exact line search under strict convexity, coercivity, and uniform spectral bounds on the Hessian. The numerical part shows that the proposed line search resolves the reported convergence failures on an 'obstacle course' of five benchmarks, with several strain-energy decompositions and two irreversibility treatments, and demonstrates the method on a demanding 3D Brazilian test. The paper explicitly acknowledges that the convergence proof does not cover the reduced-space active-set strategy and the star-convex model with γ*>0, for which robustness is established empirically.","tokens_in":31067,"tokens_out":9310,"duration_ms":89385,"significance":"If the correctness gap identified below is closed, this is a practically valuable contribution: the bisection-based exact line search is simple, is contributed to PETSc, and the implementation is publicly available. The proof of Proposition 2 in Appendix D is internally consistent under its stated assumptions, and Proposition 1 is a standard bisection result. The numerical study is unusually broad for this area, covering multiple splits, two irreversibility strategies, and a large 3D problem, and the paper is honest about the cases not covered by the theory. These strengths make the manuscript worth serious consideration. However, the link between the proved statement and the algorithm as printed is broken by a concrete detail in the termination test of Algorithm 2, and this must be resolved before the central claim can be accepted.","major_comments":[{"comment":"Algorithm 2 uses the termination test φ'(λ_l)/||Δw||² ≤ atol, not |φ'(λ_l)|/||Δw||² ≤ atol. Since φ'(0)<0 for a Newton direction, φ'(λ_l) is negative for every λ_l to the left of the true root, so any such λ_l automatically satisfies φ'(λ_l)/||Δw||² ≤ atol for positive atol. If the first bisection midpoint lies to the left of the minimizer, the algorithm exits immediately at that midpoint without bracketing the root. The returned step is then not the minimizer of (30), and the key inequality (50) — and with it the residual-norm convergence proof of Proposition 2 — does not apply to the algorithm as written. The pseudocode must be corrected to use |φ'(λ_l)|, or, if the PETSc implementation already uses the absolute value, the pseudocode and the text in Section 3.1 and Appendix F.2 must be corrected to match the implementation. As printed, the global convergence guarantee does not cover th","section":"Section 3.1, Algorithm 2"},{"comment":"Even after inserting the absolute value, Proposition 2 assumes an exact minimizer of φ over [0,1], whereas Algorithm 2 stops at finite tolerances atol and ltol. With fixed tolerances the returned λ* is only an approximate minimizer. The paper should state explicitly that the theorem is for the idealized exact line search, and clarify the relation to the finite-tolerance implementation (e.g., a limiting argument with tolerances tending to zero, or a separate convergence result for an inexact line search satisfying a sufficient-decrease condition). As written, the proof of (50) and the subsequent residual-norm bound apply only to the idealized algorithm, not directly to the stopping criteria actually used in the numerical experiments.","section":"Section 3.2 / Appendix D, Eq. (50)"},{"comment":"Remark 4 states that if the mechanical and damage subproblems are strictly convex, alternate minimization with exact line search is 'guaranteed' to converge to a critical point. However, Proposition 2 requires not only strict convexity but also the uniform spectral bounds (34) on the Hessian for all iterates. Strict convexity alone does not imply a positive uniform lower bound on the smallest eigenvalue over unbounded sets. The remark should explicitly carry over the spectral-bound assumption; otherwise the guarantee is stronger than the proof supports. This is a clarity issue in the main claim, though the proof itself is sound under the stated assumptions.","section":"Remark 4 / Eq. (34)"}],"minor_comments":[{"comment":"The citation '[1, Theorem 1]' for convergence of alternate minimization appears to point to the 2000 Bourdin–Francfort–Marigo paper; the relevant convergence theorem is more commonly associated with Bourdin's 2007 paper (reference [27]) or Burke–Ortner–Süli (reference [28]). Please verify and correct the citation.","section":"Remark 4 / References"},{"comment":"The initialization 'λ_0 = λ_left_l = 0, λ_l = λ_right_l = 1' overloads λ_l as both the current iterate and the right bracket. The pseudocode is understandable but could be clearer, e.g. by using λ_mid for the evaluated point.","section":"Section 3.1, Algorithm 2"},{"comment":"The description of SNESLineSearchCP uses the same non-absolute termination test φ'(λ_l)/||Δw||² ≤ atol as Algorithm 2. If the shipped PETSc implementation does not take an absolute value, the efficiency comparison in Section 4.3 may be affected by the same early-exit behavior; please clarify.","section":"Appendix F.2"},{"comment":"The empirical convexity check by sampling 100 points per Newton iteration is a reasonable sanity check, but the wording 'could not find any evidence of non-convexity up to numerical tolerances' should be kept clearly separate from a proof. The current wording is mostly careful, but a sentence explicitly stating that this check is not a substitute for the missing proof would help avoid overstatement.","section":"Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The decisive issue is the missing absolute value in the termination test of Algorithm 2. Before acceptance, I would ask the authors to confirm whether the PETSc implementation uses |φ'(λ)|; if it does not, the global-convergence theorem does not cover the shipped solver. If it does, the pseudocode and the Appendix F.2 description need correction. The rest of the technical content is solid and the numerical study is convincing; this is a fixable but load-bearing defect."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this is a genuinely useful paper, and it should go to review. The authors don't hide what they can't prove, the benchmarks are extensive, and the code is public. But the pseudocode for their bisection line search (Algorithm 2) omits the absolute value in the termination test, and as written that makes their own convergence proof inapplicable to the method they specify. That needs fixing before publication.\n\nWhat's actually new: the systematic application of exact bisection line search to the two subproblems of staggered phase-field fracture, including the constraint-aware variant for active-set and penalty formulations, plus the 3D Brazilian test with contact and the star-convex model. The convergence result (Proposition 2) is classical in spirit—greedy Newton is not new, see Shea & Schmidt—but the adaptation to phase-field subproblems and the empirical demonstration of robustness across many energy splits is a real contribution. The paper is honest: Remark 4 and Section 4.2 explicitly say the reduced-space active-set and star-convex gamma*>0 cases are not covered by the proof, and they rely on a 100-point sampling of phi' instead. Good.\n\nThe soft spots:\n\n1. Algorithm 2. The exit condition is `phi'(lambda_l)/||Delta w||^2 <= atol`, not `|phi'(lambda_l)|/...`. For any lambda left of the minimizer, phi' is negative and hence trivially <= positive atol. So if the first bisection point 0.5 lies left of the root, the loop exits at 0.5 without bracketing. That means the printed algorithm does not perform exact line search in general, and the key inequality (50) in the proof does not hold for the algorithm as specified. If the PETSc code actually uses abs, then the pseudocode is wrong; if it follows the pseudocode, the shipped solver isn't covered by Proposition 2. Either way it's a concrete fix, but it's load-bearing because the entire theoretical claim rests on exactness.\n\n2. The title. 'Exact line search is all you need' overstates. The proof covers the strictly convex subproblems; the active-set and star-convex cases are empirical. The authors say this themselves, but the title and conclusions lean harder than the results.\n\n3. Minor: the novelty boundary with the authors' own prior work [76] is not clearly drawn. Some readers will want a sentence on what's new relative to that paper.\n\nThe math is otherwise fine—Appendix D's proof is consistent under its stated assumptions, and the numerical study goes well beyond what's typical. This deserves a serious referee, but the reviewer should request the Algorithm 2 fix and a title/scope adjustment.","headline":"Solid, honest, and useful paper on exact line search for phase-field fracture; but the printed bisection algorithm has a sign bug in its termination test that breaks the exactness assumption, and the title oversells the unproved cases.","tokens_in":31532,"tokens_out":2952,"would_cite":true,"duration_ms":26702,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["74R10","65K05","65H10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Alternate minimization for phase-field brittle fracture can be made globally convergent by equipping each Newton solve with an exact line search that minimizes the energy along the search direction.","keywords":["phase-field fracture","brittle fracture","alternate minimization","Newton's method","exact line search","bisection","non-convex minimization","iterative convergence"],"falsifier":"Run the bisection line search on a strictly convex subproblem instance and record the Hessian eigenvalue range along the Newton iterates: a violation of (34), or a Newton direction along which φ' is not monotone (detectable by sampling φ' at 100 points and finding a sign-change pattern inconsistent with a single root), would invalidate the global-convergence guarantee. Concretely, the residual norm failing to decrease to zero on such an instance—or the bisection bracketing losing the root because φ' changes sign more than once—would refute Proposition 2's scope.","tokens_in":30565,"feed_emoji":"","tokens_out":4358,"duration_ms":37908,"temperature":0.7,"pith_summary":"The paper targets a practical breakdown: in staggered (alternate minimization) solution of variational phase-field brittle fracture, the per-subproblem Newton solvers often fail to converge, stalling the whole computation. It claims the failure is a step-length problem, not a direction problem: full Newton steps overshoot the one-dimensional minimum of the energy along the Newton direction. The proposed fix is an exact line search that locates this minimum by bisection on the directional derivative, which is monotone when the subproblem energy is strictly convex. Under strict convexity, coercivity, and uniform spectral bounds on the Hessian, Newton with this line search is proven to drive the residual norm to zero, so the outer alternate-minimization loop is guaranteed to reach a critical point. The same recipe is shown numerically to resolve all tested benchmark failures, including the star-convex energy decomposition for which the theoretical proof does not apply, and a demanding 3D Brazilian test with contact.","feed_headline":"Exact line search guarantees Newton convergence in fracture solvers","feed_subtitle":"Bisection-based step lengths turn stalled staggered phase-field computations into convergent ones, 3D included.","key_machinery":"The load-bearing mechanism is the exact line search implemented by bisection on the directional derivative φ'(λ)=R_w(w+λΔw)·Δw. Along a Newton direction of a strictly convex subproblem energy, φ is strictly convex, so φ' is monotone and has at most one root in (0,1], which is the energy minimizer. Bisection converges linearly with bracket error |λ_l−λ⋆|≤2^{-l}. The global-convergence proof then couples this with the eigenvalue bounds ρ_L≤λ_min(K_ww) and ρ_U≥λ_max(K_ww): an upper bound on the energy decrease per iteration, E(w_{k+1})−E(w_k) ≤ −(1/2)(ρ_L/ρ_U²)||R_w(w_k)||², telescopes into a finite sum of squared residual norms, forcing the residual to zero.","core_discovery":"The central discovery is that Newton's method with an exact line search—minimizing the energy along the Newton update direction—converges globally for the strictly convex displacement and damage subproblems that arise in phase-field brittle fracture. The paper proves that the sequence of residual norms generated by this 'greedy' Newton converges to zero whenever the Hessian eigenvalues stay within a uniform interval [ρ_L, ρ_U]; strict convexity then gives convergence to the unique subproblem minimizer. Consequently, if both subproblems are strictly convex, alternate minimization with exact line search is guaranteed to converge to a critical point of the total energy. For cases outside the pr","pith_inferences":["The proof's ingredients — strict convexity, coercivity, uniform Hessian spectrum — are generic, so the same bisection line search should globalize Newton's method for any strictly convex finite-element minimization, not just phase-field fracture; a direct test on a different convex nonlinear PDE problem would verify that.","The paper shows numerical reliability for the star-convex γ⋆>0 case but replaces the proof with a 100-point monotonicity check of φ'; a sharper investigation would seek whether non-convexity can actually occur and, if so, whether the bisection still converges or needs a bracketing fix.","The claim 'exact line search is all you need' implicitly argues that globalization can be handled by the line search alone, without trust regions or inertia corrections; comparing this method head-to-head against such monolithic globalization strategies on the same tests would test that stronger claim.","Because the line search minimizes along each Newton direction, the method's energy-monotonicity may also help with second-order stability checks (assessing whether the found critical point is a local minimizer), a connection the paper notes but does not exploit."],"forward_implications":["With the exact line search active, every benchmark in the obstacle course converges — including the volumetric–deviatoric, spectral, no-tension, Drucker–Prager-like, and star-convex splits — where plain Newton previously stalled in the mechanical subproblem.","When irreversibility is enforced by the reduced-space active-set method, applying the line search to the mechanical subproblem alone is sufficient for convergence in all tested cases; the damage subproblem remains linear and needs no line search.","When irreversibility is enforced by penalization, the damage subproblem becomes strongly nonlinear and the exact line search must also be applied there for several cases; with it, all cases converge.","In the tested cases the bisection line search needs fewer total Newton iterations than backtracking or secant line searches, at the cost of more residual evaluations per iteration, which becomes cheaper relative to the linear solve as problems grow.","The approach computed the 3D Brazilian test with the star-convex model and contact constraints, producing a branched crack pattern within a single load step — a case that previously could not be computed."],"fun_headline_variants":["Exact line search ensures Newton convergence in fracture","Bisection line search guarantees fracture solver convergence","Why Newton fails in phase-field fracture without line search","Fracture computations converge when Newton uses exact line search","A simple fix: exact line search for Newton in fracture"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The proof rests on assuming that, for every Newton iterate, the Hessian eigenvalues stay inside a fixed interval [ρ_L,ρ_U] and that the energy along the search direction is strictly convex so its derivative is monotone; for the star-convex and reduced-space cases only numerical sampling supports these assumptions.","fun_headline_variants_meta":{"raw":{"variants":["Exact line search ensures Newton convergence in fracture","Bisection line search guarantees fracture solver convergence","Why Newton fails in phase-field fracture without line search","Fracture computations converge when Newton uses exact line search","A simple fix: exact line search for Newton in fracture"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000432,"raw_usage":{"total_tokens":2037,"prompt_tokens":739,"completion_tokens":1298,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":483,"completion_tokens_details":{"reasoning_tokens":1222}},"tokens_in":483,"tokens_out":1298,"duration_ms":10629,"temperature":1.0,"reasoning_tokens":1222,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T19:36:00.110353+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the bisection line search on a strictly convex subproblem instance and record the Hessian eigenvalue range along the Newton iterates: a violation of (34), or a Newton direction along which φ' is not monotone (detectable by sampling φ' at 100 points and finding a sign-change pattern inconsistent with a single root), would invalidate the global-convergence guarantee. Concretely, the residual norm failing to decrease to zero on such an instance—or the bisection bracketing losing the root because φ' changes sign more than once—would refute Proposition 2's scope.","supporting_citations":[],"review_version":1}