{"id":"89068e01-82ea-44aa-b9ab-36cc03353b03","arxiv_id":"2607.09599","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"Nonlinear imaginary-time evolution (NITE) extends ITE to general costs via state-dependent generators, equals variational QNGD, and converges exponentially to local minima.","lead":"The paper generalizes imaginary-time evolution into a nonlinear version (NITE) that works for any differentiable cost on quantum states, not just energy. This gives a geometry-aware optimizer that converges faster than plain gradient descent on variance minimization, quantum sensing, and excited-state tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The reader correctly isolates the local character of Theorem 1 as the weakest point, yet that locality is standard for Riemannian gradient flows and does not undermine the claim as stated. The proofs in the supplement (Riesz representation for H(ϕ), chain-rule derivation of the QNGD update, and the PL argument) are complete and free of algebraic gaps. The three numerical examples, especially the full-space controls that remove ansatz expressivity, supply independent empirical corroboration of the predicted linear rate. Consequently no adjustment to the ACCEPT verdict is warranted.","tokens_in":14977,"tokens_out":373,"duration_ms":66930,"concrete_test":"Independently recompute the minimal eigenvalue of the Fubini–Study Hessian of the variance cost (Eq. 12) at a known eigenstate of the 6-qubit H3+ Hamiltonian and verify that the observed late-time slope of log|Var(t)| in the full-space RK45 trajectories of Fig. S1 matches 2λ* within 5 %.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (local exponential convergence of NITE under a positive-definite Fubini–Study Hessian, plus exact equivalence of its variational form to QNGD) is supported by a standard Riemannian-gradient-flow argument that is correctly specialized to the product Fubini–Study metric. The local Polyak–Łojasiewicz step after Eq. (S6) is the usual consequence of strong geodesic convexity on a small convex neighborhood; the paper already flags the locality of the guarantee and recovers the classical spectral-gap rate for ordinary ITE as a special case (Remark 1). No internal inconsistency or hidden assumption that would invalidate the theorem under its stated hypotheses was found.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript generalizes normalized imaginary-time evolution from energy expectation values to arbitrary C^{2} cost functions on pure states (and products of pure states). The resulting Nonlinear Imaginary-Time Evolution (NITE) is the Fubini–Study gradient flow of the cost; its generator is the state-dependent Hermitian operator obtained from the Hilbert–Schmidt gradient of the cost (Eqs. (5)–(6) and multi-state analogues (8)–(9)). The authors prove local exponential convergence to non-degenerate local minima under a positive-definiteness assumption on the Fubini–Study Hessian (Theorem 1 and the manifold version Theorem 2 in the supplement), recover the classical spectral-gap rate of ordinary ITE as a special case (Remark 1), and show that the variational discretization of NITE is exactly quantum natural gradient descent (single-state and multi-state forms, supplement §§III–IV). Numerical illustrations on variance minimization, quantum Fisher information maximization, and penalty-based simultaneous ground/excited-state preparation demonstrate faster and more robust convergence than Euclidean gradient descent, both with hardware-efficient ansätze and with full-space parameterization.","tokens_in":15158,"tokens_out":838,"duration_ms":8747,"significance":"If the claims hold, the work supplies a clean geometric unification of imaginary-time evolution and quantum natural gradient for objectives that are not simple expectation values—precisely the setting of many modern variational subroutines (variance minimization, QFI maximization, penalty methods). The local-convergence theorem is a standard but carefully specialized Riemannian Polyak–Łojasiewicz argument that recovers the known spectral-gap rate, and the exact equivalence of the variational form to QNGD is useful for near-term implementations. The multi-state extension and the explicit applications give a practical route to low-energy subspace preparation and metrology-oriented state preparation. Strengths include the transparent continuous-time derivation, the reduction proofs, the local-rate guarantee with an explicit Hessian hypothesis, and the side-by-side numerics (means ± std over 10 seeds, full-space checks that isolate ansatz expressivity).","major_comments":[],"minor_comments":[{"comment":"Fig. 2 caption and main-text discussion of the multi-state experiment should state the concrete numerical values of the penalty coefficients λ0, λ1, λp (or the general λk, λij) used in the runs; they are free parameters of the cost (15)–(16) and affect the landscape.","section":null},{"comment":"In the main text after Eq. (5), a one-sentence pointer that the non-uniqueness of H(ϕ) is resolved by the Hilbert–Schmidt definition (and that this choice recovers ordinary ITE) would help readers who do not immediately consult the supplement.","section":null},{"comment":"Supplement Fig. S4 reports an empirical rate comparison with 2λ*; a brief remark in the main text that the observed slopes are consistent with the local-rate prediction of Theorem 1 would strengthen the link between theory and numerics.","section":null},{"comment":"A few typographical inconsistencies appear (e.g., “EffcientSU2”, “withrespecttothe”, “andpo-tential”); a light copy-edit pass would remove them.","section":null},{"comment":"The outlook mentions finite-temperature free-energy functionals and quantum-machine-learning losses; a short clarifying sentence that the same Hilbert–Schmidt construction applies once the cost is written as a function of the purified state (or of the density operator) would make the claimed generality more concrete.","section":null}],"recommendation":"accept","confidential_remarks":"The manuscript is a clean, self-contained contribution that sits comfortably in the quantum-algorithms / variational-methods literature. No novelty or citation-pattern concerns; the geometric framing is standard but usefully extended. Suitable for a general quant-ph or quantum-information journal without further major work."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing worth knowing is that this paper takes the known Fubini–Study gradient-flow picture of ordinary imaginary-time evolution and writes the same flow for any smooth pure-state cost. The resulting state-dependent generator H(ϕ) is defined cleanly via the Hilbert–Schmidt gradient, reduces exactly to the usual Hamiltonian when the cost is an expectation value, and yields a multi-state system of coupled nonlinear equations when the cost depends on several pure states. The local exponential convergence theorem (supplement Theorems 1–2) is standard Riemannian geometry applied carefully: positive-definite Fubini–Study Hessian implies a local Polyak–Łojasiewicz inequality, hence E(ϕ(t))−E* decays as e^{-μt} for μ < 2λ_min. They recover the classical spectral-gap rate as a special case, which is a nice consistency check.\n\nWhat is new is the multi-state formulation, the exact identification of the variational discretization with summed-metric QNGD, and the demonstration that the same optimizer works for variance minimization, QFI maximization, and penalty-based excited-state preparation. The continuous-time derivation and the reduction proofs in the supplement are solid. Numerics (means ± std over 10 seeds, both ansatz and full-space RK45) show faster and more robust convergence than plain gradient descent on the three subroutines; the full-space runs confirm the linear rate once expressivity is removed.\n\nSoft spots are real but modest. The guarantee is purely local and requires the Hessian to stay positive-definite on a geodesically convex neighborhood that traps the flow—standard for these arguments and already flagged by the authors. Penalty coefficients and learning rate remain free parameters, as usual. No code is released, which is a practical inconvenience rather than a soundness issue. Citation pattern is appropriate; Stokes et al. 2020 is properly credited as the geometric starting point.\n\nThis is for people who already run variational quantum algorithms or tensor-network / neural-state optimizers and want a geometry-aware method that is not limited to energy. It does not open a new foundational problem, but it is a clean, usable technical contribution. I would send it to peer review without hesitation; a serious referee will find the math and the controls ready for scrutiny.","headline":"Clean geometric generalization of ITE to arbitrary pure-state costs, with a correct local exponential-rate theorem and exact multi-state QNGD equivalence; useful and ready for referees.","tokens_in":15765,"tokens_out":568,"would_cite":true,"duration_ms":12543,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Nonlinear imaginary-time evolution optimizes general pure-state costs and converges locally exponentially.","keywords":["imaginary-time evolution","Fubini-Study metric","quantum natural gradient","variational quantum algorithms","state preparation","quantum Fisher information","variance minimization"],"falsifier":"On any of the three example tasks, replace the Fubini-Study metric by the Euclidean metric (ordinary gradient descent) or deliberately choose an initial state whose trajectory leaves the positive-Hessian neighborhood; if the observed convergence rate remains exponential and identical to the NITE rate, the local-exponential claim fails.","tokens_in":15878,"feed_emoji":"⏱️","tokens_out":882,"duration_ms":8696,"temperature":0.7,"pith_summary":"Standard imaginary-time evolution finds ground states by flowing downhill in energy under the Fubini-Study geometry of pure states. This paper generalizes that flow to any smooth cost defined on pure states (or several pure states at once). The resulting nonlinear imaginary-time evolution (NITE) is the Fubini-Study gradient flow of the chosen cost; its effective generator is state-dependent, so the differential equation is nonlinear. A hardware-efficient variational version of NITE is exactly quantum natural gradient descent on the same cost. The authors prove local exponential convergence to non-degenerate minima, apply the method to variance minimization, quantum Fisher-information maximization, and penalty-based excited-state preparation, and show that it converges faster and more reliably than ordinary gradient descent.","feed_headline":"Nonlinear imaginary-time flow beats gradient descent on quantum costs","feed_subtitle":"General pure-state objectives get local exponential convergence and a hardware-ready variational form.","key_machinery":"Nonlinear Imaginary-Time Evolution (NITE): the ordinary differential equation d|ϕ>/dt = −(H(ϕ)−⟨H(ϕ)⟩)|ϕ>, where the state-dependent Hermitian generator H(ϕ) is the Hilbert-Schmidt gradient of the cost evaluated at the pure-state density matrix; multi-state NITE couples several such equations.","core_discovery":"Any twice-differentiable cost on pure quantum states (or on a tuple of them) defines a nonlinear imaginary-time evolution given by the Fubini-Study gradient flow. That continuous flow decreases the cost monotonically and converges locally exponentially to any non-degenerate local minimum; its first-order variational discretization is precisely quantum natural gradient descent for that cost.","pith_inferences":["Because the exponential rate is controlled by the smallest positive eigenvalue of the Fubini-Study Hessian, one can diagnose and regularize near-degenerate directions before training begins.","The multi-state metric sum that appears in the variational update is formally identical to the metric of a product-state manifold, suggesting a direct lift to multi-parameter quantum sensing or multi-objective variational algorithms.","If the free-energy functional of a purified thermal state is used as the cost, the same NITE construction yields a geometry-aware imaginary-time flow for finite-temperature variational manifolds."],"forward_implications":["Any pure-state variational subroutine whose cost is not a simple expectation value can replace ordinary gradient steps by NITE/QNGD steps and inherit local exponential convergence.","Multi-state NITE supplies a single coupled flow that simultaneously prepares an orthonormal low-energy subspace rather than one eigenstate at a time.","The same construction applies immediately to matrix-product states, neural quantum states, or other variational manifolds once the Fubini-Study metric on that manifold is available.","When the cost is a nonlinear composite of expectation values (least-squares or cross-entropy loss), NITE becomes a geometry-aware training rule for variational quantum classifiers."],"fun_headline_variants":["Generalized NITE: Fubini-Study flow for any pure-state quantum cost","Nonlinear ITE yields local exponential convergence on quantum costs","Variational NITE matches quantum natural gradient beyond ground states","NITE turns general pure-state objectives into exponentially convergent flows","Hardware-efficient NITE outperforms gradient descent on quantum subroutines"],"cache_read_input_tokens":128,"weakest_assumption_plain":"Near the target minimum the Fubini-Study Hessian must stay strictly positive-definite inside a whole geodesically convex neighborhood, and every trajectory that starts nearby must remain inside that neighborhood.","fun_headline_variants_meta":{"raw":{"variants":["Generalized NITE: Fubini-Study flow for any pure-state quantum cost","Nonlinear ITE yields local exponential convergence on quantum costs","Variational NITE matches quantum natural gradient beyond ground states","NITE turns general pure-state objectives into exponentially convergent flows","Hardware-efficient NITE outperforms gradient descent on quantum subroutines"]},"model":"grok-4.5","effort":"low","cost_usd":0.003918,"raw_usage":{"total_tokens":1173,"prompt_tokens":684,"num_sources_used":0,"completion_tokens":90,"cost_in_usd_ticks":39180000,"prompt_tokens_details":{"text_tokens":684,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":399,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":684,"tokens_out":90,"duration_ms":4905,"temperature":1.0,"reasoning_tokens":399,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T01:50:45.735240+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On any of the three example tasks, replace the Fubini-Study metric by the Euclidean metric (ordinary gradient descent) or deliberately choose an initial state whose trajectory leaves the positive-Hessian neighborhood; if the observed convergence rate remains exponential and identical to the NITE rate, the local-exponential claim fails.","supporting_citations":[],"review_version":1}