{"id":"b4c0af0d-43f4-4cc7-8f12-b6a8f577bbe1","arxiv_id":"2608.06669","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A gauge-invariant Fock-commutator preconditioner makes Jacobian-free Newton-Krylov solvers converge robustly for coupled cluster equations in arbitrary orbital gauges.","lead":"This paper presents a more robust way to solve the central equations of coupled cluster theory, a widely used quantum chemistry method, by making the solver independent of the orbital coordinate system. It removes the need for manual level-shifting tricks, which could make high-throughput chemistry calculations more reliable.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Efficiency claim omits inner A_F solves, which carry O(N^5)-per-application cost and may dominate wall time in non-canonical gauges.","rationale":"The reader's conditional verdict already identifies the unproven invertibility of A_F and the unmeasured cost of inner preconditioner solves as the weakest assumption. My stress-test concurs and sharpens the point: Algorithm 1's inner A_F^{-1} applications are not counted in the paper's residual-evaluation metric, yet each application has the same asymptotic scaling as a residual evaluation. This directly undermines the central 'efficient' claim, independent of the small benchmark set. The nonsingularity issue is also real, but the numerical experiments on stretched H2 suggest the method can still proceed in practice, so the cost-accounting flaw is the more load-bearing concern. Since the reader's verdict already conditions on these issues, my analysis does not move the verdict, but it provides a concrete test that would settle the matter.","tokens_in":15475,"tokens_out":5038,"duration_ms":57250,"concrete_test":"Instrument the public ARCC.jl implementation to log, for the ethane cc-pVTZ Ra-gauge PNK run in Fig. 5, the cumulative number of inner GMRES iterations used for A_F^{-1} solves and the wall time spent in those solves, then recompute the comparison of PNK vs SFP+DIIS using wall time instead of residual evaluations. If the inner A_F solves account for more than 10% of total wall time, the residual-evaluation metric hides a dominant cost and the efficiency claim requires revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 1 (Appendix F) applies A_F^{-1} at lines 3 and 6, and Table I states that this inversion is performed by GMRES. However, Section III counts only residual evaluations as the efficiency metric. In a non-canonical gauge A_F is not diagonal (Eq. C4); each application of A_F to a trial vector is a four-index Fock contraction of cost O(N_v^3 N_o^2 + N_v^2 N_o^3), which is the same polynomial scaling as a residual term and is not negligible. An inner GMRES solve requiring even tens of iterations adds a multiplicative cost that the reported residual-evaluation counts ignore. Figures 4 and 5 compare PNK against SFP+DIIS solely on this metric, so the claimed wall-time advantage is not established. Additionally, A_F is singular whenever F has degenerate occupied-virtual pairs (e.g., stretched H2 in the small-gap limit), and no conditioning analysis or regularization of the inner solves is provided; a near-singular A_F would make the inner GMRES stagnate, directly threatening the 'robust' claim. The paper should either report wall-clock timings that include inner solves or justify that A_F^{-1} is inexpensive (e.g., diagonal dominance in the targeted gauges).","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a method for solving coupled-cluster (CC) amplitude equations called PNK: a Jacobian-free Newton-Krylov solver preconditioned by the matrix A_F with entries [A_F]_{\\mu\\nu} = \\langle\\Phi_\\mu|[F, X_\\nu]|\\Phi_0\\rangle. This preconditioner is gauge-invariant by construction, generalizing the diagonal energy-denominator preconditioner that is only valid in the canonical gauge. The authors derive working equations for CCD in an arbitrary gauge, give an explicit algorithm (Algorithm 1, Appendix F) that applies A_F^{-1} by inner GMRES, and compare PNK against fixed-point, level-shifted fixed-point plus DIIS, and unpreconditioned Newton-Krylov variants on two molecular systems: stretched H2 at CCD/cc-pVTZ and ethane in several gauges at CCD/6-31G and CCD/cc-pVTZ. The efficiency metric used throughout is the number of residual evaluations, justified by the assertion that residual evaluation dominates the computational cost. The paper claims that PNK is robust and efficient across arbitrary gauges without system-dependent level shifting, and suggests it as a promising standard solver for CC equations.","tokens_in":15745,"tokens_out":2467,"duration_ms":28505,"significance":"If the central claims hold, the paper would make a useful contribution to CC solver technology: replacing a gauge-dependent, ad hoc level-shifted fixed-point iteration with a preconditioned Newton-Krylov method is conceptually attractive for local-correlation and non-canonical-gauge formulations, where the standard energy-denominator preconditioner fails. The derivation of the gauge-invariant preconditioner from the commutator form of the Jacobian is standard but cleanly presented, and the authors provide explicit working equations and publicly available Julia code. They also disclose, in Appendix D, that the level-shift values used for the SFP baseline are optimized post hoc, which is honest and actually strengthens the comparison: PNK beats an SFP baseline with a shift chosen after the fact. However, the numerical evidence is limited to two small molecules at CCD level, and the efficiency claim is measured only in residual evaluations while the proposed algorithm incurs additional inner GMRES solves for A_F^{-1}, whose cost is not included.","major_comments":[{"comment":"The efficiency claim is not established by the reported metric. Section III states that convergence is assessed by residual evaluations because residual construction is the dominant cost, but Algorithm 1 applies A_F^{-1} by GMRES at lines 3 and 6, and in a non-canonical gauge A_F is not diagonal (see Eq. C4). Each application of A_F to a trial vector is a four-index Fock contraction with the same polynomial scaling as a residual term, and an inner GMRES solve multiplies the per-outer-iteration cost by the number of inner iterations. Figures 4 and 5 compare PNK against SFP+DIIS and NK only in residual evaluations, so the claimed wall-time advantage is not demonstrated. The authors should either report wall-clock timings that include the inner preconditioner solves or justify that A_F^{-1} is inexpensive in the gauges tested (e.g., by diagonal dominance or by a direct factorization).","section":"Section III and Algorithm 1 (Appendix F)"},{"comment":"The robustness claim requires an analysis of A_F, not just an assumption that A_F^{-1} exists. The preconditioner A_F is singular whenever F has degenerate occupied-virtual pairs, as in stretched H2 at the small-gap limit discussed in Section III.A, and the paper does not provide a conditioning estimate or a regularization strategy for the inner solves. If A_F is near-singular, the inner GMRES for A_F^{-1} can stagnate, and the 'robust' claim would fail exactly in the challenging regime the paper emphasizes. Please add a conditioning analysis, report smallest singular values for the test cases, or document a practical regularization (e.g., Tikhonov or truncated SVD) and verify that the outer Newton iterations remain robust under it.","section":"Section II, Eq. (12), and Algorithm 1"},{"comment":"The numerical support is too narrow for the breadth of the abstract and conclusion. The tests are CCD on H2 and ethane only; the abstract promises 'a range of molecular systems' and the conclusion extrapolates to 'next-generation CC solvers' and local correlation methods. There are no calculations in a genuinely local-correlation setting, no larger basis sets beyond cc-pVTZ, and no demonstration that applying A_F^{-1} remains competitive at larger problem sizes. At minimum, the authors should temper the wording to match the evidence, and ideally add at least one larger or more strongly correlated system with wall-clock timings.","section":"Section III and Conclusion"}],"minor_comments":[{"comment":"There are numerous typographical errors that should be corrected: 'approache' and 'imporatant' in the introduction, 'Relieable' for 'Reliable', 'perconditioner' in Section II, and 'ad hoc' frequently missing the space before 'chosen'. A careful proofreading pass is needed.","section":"Introduction and throughout"},{"comment":"The Wick-expansion terms in Eq. (C2) are printed as a long list of nearly identical-looking terms; the notation with repeated expression blocks is almost unreadable. It would be much clearer to present the final result (C3) and move the derivation details to a table or a supplementary derivation.","section":"Appendix C, after Eq. (C2)"},{"comment":"Table I classifies INK as having no preconditioner and lists 'None' for its preconditioner inversion, but the INK method still solves the linear system A_F \\delta t = -r(t) by GMRES without a preconditioner. The distinction between 'preconditioner' and 'preconditioner inversion' is confusing; please clarify in the text whether the INK GMRES applies A_F directly or uses a different preconditioner.","section":"Table I and Section II"},{"comment":"The algorithm box does not define the finite-difference step \\delta used at line 6, nor does the main text state how it is chosen. Since the Jacobian-free approximation is a numerical parameter that affects convergence, please give the formula or value used in the experiments.","section":"Appendix F, Algorithm 1"},{"comment":"Appendix D states that the optimal shift was chosen by minimizing the spectral radius, but the text does not explain how the spectral radius was computed (e.g., by building the Jacobian explicitly or by a power iteration). A brief sentence on this would make the comparison reproducible.","section":"Appendix D and Figure 6"}],"recommendation":"major_revision","confidential_remarks":"The core idea is sound and the paper is honest about the optimized baseline, but the efficiency claim currently rests on an incomplete metric. I would ask for wall-clock timings including inner solves, or a convincing complexity argument that the inner solves are cheap, before publication. The conditioning issue of A_F is also important: without it, the 'robust' claim is not fully supported. The paper fits the journal's scope, but the breadth of the claims should match the two-molecule CCD evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper gives a clean way to precondition Newton-Krylov solvers for coupled cluster equations in non-canonical gauges. The idea is to replace the diagonal energy denominator with the Fock-commutator matrix A_F and apply A_F^{-1} through inner GMRES. In the canonical basis A_F reduces to the usual denominator, so this is a genuine generalization of Yang et al. and of the AO-gauge fixed-point workaround. The derivation is standard but clearly laid out, and the algorithm is specified in enough detail to reproduce. On the small test set, H2 at stretched bond lengths and ethane in MO, AO, and random gauges, the robustness claim holds: FP diverges in AO and random gauges, while their Newton-type methods converge.\n\nThe main soft spot is the efficiency metric. They count residual evaluations only, and the inner A_F solves are not counted. Each application of A_F to a trial vector is the same O(N^5) scaling as a residual term, and an inner GMRES solve taking even a few dozen iterations multiplies the cost. So the claimed wall-time advantage over SFP+DIIS is not established. The paper should report actual timings or justify that A_F is cheap to invert in the targeted gauges. Related to this, A_F can become singular when F has degenerate occupied-virtual pairs, as in stretched H2; there is no conditioning analysis or regularization of the inner solves. That is a real risk for the 'robust' label.\n\nA second soft spot is the scope: two molecules at CCD level is thin for a claim that this is a new standard. The numerical results are consistent with the theory, but they are not broad enough to support that language.\n\nWhat I liked: the authors are honest about the baseline. They optimize the level shift post-hoc and explicitly say that is impractical, which is more than many papers do. The derivation of A_F from first-order perturbation theory is right, and the reduction to the canonical denominator is a nice sanity check.\n\nThis is for people working on local-correlation and embedding CC, where non-canonical gauges are the rule. It deserves peer review. The referee should ask for wall-clock timings that include inner solves, a conditioning discussion for A_F, and at least one larger or local-correlation example. I would not desk-reject it.","headline":"A sensible gauge-invariant preconditioner for Newton-Krylov CC solvers, but the efficiency claim overstates what the evidence supports.","tokens_in":16275,"tokens_out":2953,"would_cite":true,"duration_ms":28971,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that a gauge-invariant Fock-commutator preconditioner, applied by Krylov solves inside a Newton iteration, makes coupled-cluster solvers robust in any orbital gauge and removes level-shift tuning.","keywords":["coupled-cluster equations","preconditioned Newton–Krylov method","gauge-invariant preconditioner","Fock commutator","fixed-point iteration","DIIS acceleration","local correlation methods"],"falsifier":"A concrete experiment would be to form $A_F$ explicitly for a moderately sized molecule in several random gauges and track its smallest singular value along the PNK solution path; if any gauge drives that singular value to zero or makes the inner GMRES solves require more residual evaluations than the outer iteration saves, the robustness and efficiency claims for PNK would fail.","tokens_in":15298,"feed_emoji":"⚛️","tokens_out":15741,"duration_ms":132563,"temperature":0.7,"pith_summary":"The paper sets out to show that the coupled-cluster (CC) equations can be solved robustly in any orbital gauge by treating them as a nonlinear root-finding problem and preconditioning the Newton update with a gauge-invariant Fock-commutator operator $A_F$, rather than with the diagonal energy denominators used in canonical fixed-point iterations. The proposed preconditioned Newton–Krylov (PNK) framework replaces heuristic level-shifting and direct-inversion-of-the-iterative-subspace (DIIS) tuning with Krylov-based inversion of $A_F$, and the paper presents CCD-level (coupled-cluster doubles) tests on hydrogen dissociation and ethane in molecular-orbital, atomic-orbital, and randomly rotated gauges. In these tests, the method converges where fixed-point iterations diverge, and it reaches the target residual with fewer residual evaluations than carefully optimized fixed-point solvers with level-shift and DIIS. If this claim holds, it removes a major source of empirical tuning from coupled-cluster calculations and extends Newton-type solvers to local-correlation and embedding formulations that use non-canonical gauges.","feed_headline":"Gauge-invariant preconditioner tames coupled-cluster convergence","feed_subtitle":"The Fock-commutator preconditioner makes Newton–Krylov converge in any orbital gauge, no level-shift tuning","key_machinery":"The load-bearing object is the gauge-invariant preconditioner $A_F$, defined by $[A_F]_{\\mu\\nu} = \\langle\\Phi_\\mu|[F, X_\\nu]|\\Phi_0\\rangle$, the commutator of the Fock operator with an excitation operator taken between the reference determinant and an excited determinant. It is the exact first-order Fock contribution to the Jacobian of the CC residual, so it acts as the correct Newton preconditioner in any orbital gauge; the canonical-gauge diagonal energy denominator is its special case. The algorithm makes $A_F^{-1}$ practical through an inner GMRES solve, then uses a Jacobian-free Newton–Krylov iteration with an adaptive forcing term for the outer linear system, so the full Jacobian is never formed.","core_discovery":"The paper's central claim is that the Newton step for the coupled-cluster equations is governed, in any gauge, by the linear system $A_F \\delta t = -r(t)$, where $[A_F]_{\\mu\\nu} = \\langle\\Phi_\\mu|[F, X_\\nu]|\\Phi_0\\rangle$ is the Fock-commutator approximation to the Jacobian. In the canonical molecular-orbital gauge this matrix reduces to the diagonal energy denominator $\\Delta\\varepsilon_\\mu$, so standard fixed-point iteration is recovered as a quasi-Newton method; in a non-canonical gauge the full commutator keeps the off-diagonal Fock blocks that a diagonal approximation discards. The paper then constructs PNK by applying $A_F^{-1}$ with GMRES and using that preconditioned residual inside a Jacobian-free Newton–Krylov outer loop with an adaptive forcing term. Its reported numerical results at the CCD level show robust convergence in small-gap and gauge-transformed settings, with PNK outperforming fixed-point-plus-DIIS solvers in residual-evaluation count on the systems tested.","pith_inferences":["Editorial inference: the paper's efficiency metric counts outer residual evaluations and not the inner $A_F^{-1}$ GMRES iterations, so a wall-time comparison on larger systems is the natural next test; if inner solves dominate, the reported advantage could erode.","Editorial inference: the derivation of $A_F$ uses only the Fock structure of the Hamiltonian, so the same preconditioner should carry over to CCSD, higher-order truncations, and coupled-cluster response equations, though only CCD is demonstrated here.","Editorial inference: a useful diagnostic suggested by the paper's $2\\times2$ example is to compare the spread of diagonal energy denominators with the spectral spread of the full Fock-commutator operator, predicting when the diagonal-Fock approximation will fail before running a calculation."],"forward_implications":["A CC solver can replace the diagonal energy-denominator preconditioner with $A_F$ applied via GMRES and keep the same Newton–Krylov outer loop, eliminating the gauge-dependent approximation in every orbital representation.","Local-correlation methods that work in atomic-orbital or other non-canonical gauges no longer need the ad hoc diagonal-Fock fix, since the same preconditioner is valid in those gauges.","The small-gap regime of stretched $\\mathrm{H}_2$, where the fixed-point Jacobian becomes repulsive, is handled by the inexact Newton step without an optimized level shift, on the paper's evidence.","In the ethane tests, PNK reaches the $10^{-8}$ residual threshold in fewer residual evaluations than fixed-point with level-shift and DIIS, supporting the paper's efficiency claim when residual evaluations dominate the cost."],"supporting_citations":[{"why":"It supplies the atomic-orbital gauge formulation and the diagonal-Fock workaround that the new preconditioner generalizes.","marker":"[25]"},{"why":"It supplies the earlier demonstration that energy-denominator-preconditioned Newton–Krylov beats optimized fixed-point solvers in the canonical gauge, the starting point generalized here.","marker":"[68]"},{"why":"It supplies the Jacobian-free Newton–Krylov methodology, including finite-difference Jacobian–vector products used in the outer iteration.","marker":"[70]"},{"why":"It supplies the GMRES algorithm used for the Jacobian inversion and for applying the preconditioner without forming it.","marker":"[71]"},{"why":"It supplies the adaptive forcing term used to stop the inner GMRES solves.","marker":"[77]"},{"why":"It defines level-shifting, the heuristic stabilization mechanism whose need the PNK framework claims to remove.","marker":"[10]"},{"why":"It introduces DIIS acceleration for coupled-cluster equations, the baseline against which PNK is compared.","marker":"[13]"},{"why":"It supplies the coupled-cluster working equations and the perturbative derivation of the fixed-point energy-denominator update that the paper reformulates.","marker":"[74]"}],"fun_headline_variants":["Gauge-invariant preconditioner speeds coupled-cluster solver","Newton-Krylov with Fock commutator tames CC convergence","No level-shift tuning: new coupled-cluster solver converges robustly","Coupled-cluster solver that works in any orbital gauge"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the Fock-commutator matrix $A_F$ is nonsingular enough that GMRES can reliably apply $A_F^{-1}$ in any gauge, and that the cost of those inner solves is small enough to preserve the efficiency counted in residual evaluations.","fun_headline_variants_meta":{"raw":{"variants":["Gauge-invariant preconditioner speeds coupled-cluster solver","Newton-Krylov with Fock commutator tames CC convergence","No level-shift tuning: new coupled-cluster solver converges robustly","Coupled-cluster solver that works in any orbital gauge"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000239,"raw_usage":{"total_tokens":1517,"prompt_tokens":954,"completion_tokens":563,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":570,"completion_tokens_details":{"reasoning_tokens":491}},"tokens_in":570,"tokens_out":563,"duration_ms":5883,"temperature":1.0,"reasoning_tokens":491,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:49:03.874839+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete experiment would be to form $A_F$ explicitly for a moderately sized molecule in several random gauges and track its smallest singular value along the PNK solution path; if any gauge drives that singular value to zero or makes the inner GMRES solves require more residual evaluations than the outer iteration saves, the robustness and efficiency claims for PNK would fail.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the atomic-orbital gauge formulation and the diagonal-Fock workaround that the new preconditioner generalizes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the earlier demonstration that energy-denominator-preconditioned Newton–Krylov beats optimized fixed-point solvers in the canonical gauge, the starting point generalized here."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the Jacobian-free Newton–Krylov methodology, including finite-difference Jacobian–vector products used in the outer iteration."},{"cited_title":"Saad and M","cited_arxiv_id":null,"evidence_quote":"It supplies the GMRES algorithm used for the Jacobian inversion and for applying the preconditioner without forming it."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the adaptive forcing term used to stop the inner GMRES solves."},{"cited_title":"level–shifting","cited_arxiv_id":null,"evidence_quote":"It defines level-shifting, the heuristic stabilization mechanism whose need the PNK framework claims to remove."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It introduces DIIS acceleration for coupled-cluster equations, the baseline against which PNK is compared."},{"cited_title":"Shavitt and R","cited_arxiv_id":null,"evidence_quote":"It supplies the coupled-cluster working equations and the perturbative derivation of the fixed-point energy-denominator update that the paper reformulates."}],"review_version":1}