{"id":"ce7dbd94-7364-49ad-90bf-1f1e808dd331","arxiv_id":"2608.09496","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Autoregressive PQMC uses an RNN as the guiding wavefunction in projective quantum Monte Carlo and achieves lower energies and lower variance than unguided PQMC on 1D Hermitian and non-Hermitian Ising chains.","lead":"This paper combines two existing quantum simulation tools, projective quantum Monte Carlo and recurrent neural networks, so that the neural network guides the simulation toward the ground state. The result is a numerical method that is more accurate than the plain version of the Monte Carlo algorithm on the one-dimensional models tested, including a non-Hermitian chain.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"NH self-learning fixed point solves H_eta, not H: Appendix C.4's real-imaginary decomposition converges to the Hermitian ground state, so the 'regardless ... non-Hermitian' claim is not established.","rationale":"The Hermitian core of the paper is convincing: Table I and Figs. 5-7 show that autoregressive PQMC accurately reproduces the exact JWB/ED energies for N=20 to N=150, with relative errors of order 1e-5 to 1e-6 and a clear improvement over SPQMC. The central claim, however, explicitly extends 'regardless of system size or whether the Hamiltonian is Hermitian or non-Hermitian,' and the non-Hermitian extension is the paper's claimed first. The weakest load-bearing premise is therefore the correctness of the NH method, and here the paper's own Appendix C.4 supplies a concrete internal inconsistency: the walker distribution in the self-learning loop is f = psi_T psi(tau), and training the RNN to f drives the guide to the ground state of the dynamical Hamiltonian. When the dynamical Hamiltonian is H_eta rather than H, the fixed point is psi_0(H_eta), so the 'sufficiently good approximation to the ground state of H' condition cannot be satisfied at convergence; the resulting energy would equal the Hermitian H_eta energy plus a vanishing <D> term, not E_0(H). This is not merely an unproven assumption but a predicted failure of the algorithm as described, which a stint-count scan can test directly. The reader's weakest_assumption identified the self-learning loop convergence and unbiasedness generally; I agree with that diagnosis and sharpen it to a specific NH failure mode that the reader did not spell out. The verdict remains CONDITIONAL because the Hermitian contribution is solid and the NH issue is addressable by reporting xi, disclosing which sign-problem mitigation is actually used, and benchmarking against ED for nonzero xi over many stints. If the drift test confirms the fixed-point analysis, the NH section should be revised to remove the 'regardless ... non-Hermitian' claim or to restrict it to the regime where the guide is not allowed to re-train toward H_eta.","tokens_in":20682,"tokens_out":16073,"duration_ms":167159,"concrete_test":"Restrict to the exact algorithm of Sec. III.A with the real-imaginary decomposition (Appendix C.4), with N=10, eta=1.6, xi=0.4, and report xi explicitly. Run for 5, 10, 20, and 50 stints, plotting energy per spin against exact diagonalization of Eq. (13) and against E_0(H_eta)/N from Eq. (E1). If the energy drifts from ED toward E_0(H_eta)/N as stints increase (or equals E_0(H_eta)/N at 50 stints), the NH claim is not established; if it remains at ED for all stint counts, the fixed-point analysis above is incomplete and the paper should explain the stabilizing mechanism.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In the real-imaginary decomposition (Appendix C.4), the PQMC dynamics are generated entirely by the stoquastic Hermitian Hamiltonian H_eta = -J Sum sigma_z sigma_z - eta Sum sigma_x (Eq. C7); the term i xi D never enters the walker dynamics. In importance-sampled PQMC the walkers sample f(x) = psi_T(x) psi(x,tau) (Eq. 6), and the self-learning loop (Sec. III.A) trains the RNN to this distribution, so the stable fixed point satisfies psi_T proportional to psi_0(H_eta), the ground state of H_eta, not of H = H_eta + i xi D. The paper's condition that this is 'exact when the guiding wavefunction ... is a sufficiently good approximation to the ground state of H itself' is therefore not met by the algorithm at convergence. At that fixed point the local-energy average is E_0(H_eta) + i xi <psi_0(H_eta)|D|psi_0(H_eta)> / <psi_0(H_eta)|psi_0(H_eta)>; by the uniform symmetry of H_eta, <D>=0, so the reported energy equals the Hermitian H_eta ground-state energy independent of xi. The observed agreement with ED/series may reflect the guide being initialized near psi_0(H) and the loop being stopped after only s=5-10 stints before the drift to psi_0(H_eta) completes; the paper claims these stints are sufficient for convergence, but it does not report the imaginary part xi used in any benchmark, so the NH results cannot be reproduced or checked. This is the load-bearing weakness of the abstract's 'regardless ... non-Hermitian' claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an autoregressive projective quantum Monte Carlo (PQMC) framework in which a recurrent neural network (RNN) is alternately used as an importance-sampling guide for PQMC and trained on the resulting walker distribution via KL-divergence minimization. The method is benchmarked on a 1D staggered transverse-field Ising model against exact Jordan-Wigner-Bogoliubov (JWB) diagonalization and exact diagonalization (ED), with reported relative errors around 10^-5 to 10^-6 and comparisons against variational RNNs. The paper then claims an extension to a PT-symmetric non-Hermitian Ising chain with complex staggered fields, using a gauge rotation, a similarity transformation, or a real-imaginary decomposition to address the sign problem. The Hermitian energy benchmarks are clean and internally consistent, but the non-Hermitian section contains load-bearing technical errors in the stochastic formulation and in the reported estimator.","tokens_in":20982,"tokens_out":14517,"duration_ms":159416,"significance":"If the claims were correct, the Hermitian part would be a useful demonstration that RNN-guided continuous-time PQMC can reduce the variance and bias of standard PQMC at polynomial cost, and the non-Hermitian extension would be a significant novelty. The manuscript includes an exact analytical solution for the Hermitian benchmark, detailed hyperparameters, and systematic comparisons across system sizes, which are strengths. However, the non-Hermitian extension is a central advertised contribution, and the algorithm as described does not simulate the stated non-Hermitian Hamiltonian; several supporting claims in Appendix C are internally inconsistent. The significance of the paper in its current form is therefore substantially reduced.","major_comments":[{"comment":"Equation (12) defines the local energy as E_loc(x) = E_p(x) - g \\sum_i \\psi_T(\\tilde{x}_i)/\\psi_T(x_i), without the factor (-1)^i that appears in the staggered Hamiltonian Eq. (9). As written, this is the local energy of the uniform-field TFIM H' = -J\\sum \\sigma^z\\sigma^z - g\\sum \\sigma^x, not of H in Eq. (9). Because H' is unitarily equivalent to H, the energy benchmarks in Table I and Figs. 5-7 still agree with E_0(H), but the RNN is then trained on the ground state of H' rather than on U|\\Omega'>, so the claimed fidelity to the staggered ground state and the stated motivation of learning two inequivalent sublattices are not realized. The authors should either include the missing (-1)^i factor and rerun, or explicitly state that the benchmark is the unitarily rotated uniform model and define fidelity relative to that model.","section":"Sec. IV.C, Eq. (12)"},{"comment":"The real-imaginary decomposition used for the non-Hermitian results generates walkers exclusively from H_eta = -J\\sum \\sigma^z\\sigma^z - \\eta\\sum \\sigma^x; the term i\\xi D never enters the stochastic dynamics. The self-learning loop therefore converges to a guide proportional to \\psi_0(H_eta), and the mixed estimator on that ensemble equals E_0(H_eta) plus i\\xi\\langle D\\rangle_{H_eta}, not the ground-state energy of H = H_eta + i\\xi D. For PBCs \\langle D\\rangle = 0 by one-site translation symmetry; for OBCs the real part is still E_0(H_eta) because \\langle D\\rangle is real. Hence Figs. 9-10 do not demonstrate PQMC for the non-Hermitian Hamiltonian Eq. (13), and the abstract's \"regardless ... non-Hermitian\" claim is unsupported. In addition, the manuscript never reports the value of \\xi used in any simulation, so the non-Hermitian results cannot be reproduced or checked.","section":"Appendix C.4, Eqs. (C6)-(C7)"},{"comment":"The gauge-rotation argument is incorrect for generic complex g. For a spin flip at site j, the two off-diagonal matrix elements acquire opposite phases, so after the rotation they are of the form -g e^{i\\delta} and -g e^{-i\\delta}. Both can be real and non-positive only if arg(g) is a multiple of \\pi/2; for g = \\eta + i\\xi with \\eta,\\xi \\neq 0 the transformed pair is complex, and the sign problem is not removed. The text itself notes that one direction gives -|g|e^{\\pm 2i\\phi}, which is complex, and the subsequent assertion that the Hamiltonian \"becomes stoquastic\" is therefore not established. This invalidates the stoquasticization claim made in Sec. III.B and Sec. V.A.","section":"Appendix C.1, Eq. (C1)"},{"comment":"The central claim that the self-learning loop is unbiased and that s=5 stints suffice is not supported by a fixed-point analysis. Appendix B only shows energy-per-stint curves for two models, and in the non-Hermitian case those curves are generated by the H_eta walk (see the major comment on Appendix C.4), so they cannot validate convergence to the ground state of Eq. (13). The authors should either prove or numerically demonstrate that the KL-fitted normalized autoregressive distribution converges to \\psi_0 rather than to a biased fixed point, and should state an explicit stopping criterion for the number of stints s.","section":"Sec. III.A and Appendix B"}],"minor_comments":[{"comment":"The text says that PQMC produces \"an updated walker distribution approximating |\\psi_0|^2\", but the importance-sampled distribution in Eq. (6) is f(x,\\tau) = \\psi_T(x)\\psi(x,\\tau), which equals \\psi_T\\psi_0 only in the \\tau\\to\\infty limit; the training target should be stated consistently with Eq. (6).","section":"Sec. III.A"},{"comment":"Equation (14) writes the local energy with the exact right ground-state wavefunction \\Psi_{0,R}; in the algorithm the guide is always the current RNN ansatz, so the conditions under which the biorthogonal mixed estimator with an approximate guide equals the non-Hermitian ground-state energy should be stated explicitly.","section":"Sec. V.B, Eq. (14)"},{"comment":"The caption contains \"doted line\" (should be \"dotted line\") and \"spontaneously PT-unbroken regime\" (should be \"PT-unbroken regime\").","section":"Fig. 9 caption"},{"comment":"The hyperparameter table does not list the value of \\xi (or the complex field g = \\eta + i\\xi) used for the non-Hermitian simulations; without this value the results in Figs. 9-10 are not reproducible.","section":"Appendix D, Table II"},{"comment":"In the sentence \"By the bipartite structure ... the Hamiltonian becomes stoquastic in the later case\", the phrase should be \"in the latter case\", and the sentence is internally contradictory because the preceding expression includes the complex factor e^{\\pm 2i\\phi}.","section":"Appendix C.1"}],"recommendation":"reject","confidential_remarks":"The Hermitian benchmark is clean and the exact-solution cross-checks are a genuine strength, but the non-Hermitian extension contains fundamental errors: the gauge rotation cannot render the off-diagonal elements real for generic complex g, and the real-imaginary decomposition samples a Hermitian Hamiltonian whose ground state has no dependence on \\xi. These are not local presentation issues; they invalidate the paper's central 'regardless ... non-Hermitian' claim. A revision would require a substantially new non-Hermitian algorithm and new benchmarks, which is beyond the scope of a normal revision. I therefore recommend rejection rather than major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The Hermitian half of this paper is a solid contribution; the non-Hermitian half does not support the abstract's 'regardless' claim. The genuinely new piece is using an autoregressive RNN as the guiding wavefunction inside self-learning PQMC. That combination is new, and the Hermitian benchmarks are convincing: the staggered Ising chain has an exact JWB solution, the ED and JWB agree, and the RNN-guided PQMC reaches relative errors around 1e-5 to 1e-6 with lower variance than unguided PQMC. The comparison against variational RNNs is also honest and shows a real improvement on these 1D models. That part is worth taking seriously.\n\nThe soft spots are in the non-Hermitian extension. The paper describes a gauge rotation and a real-imaginary decomposition in Appendix C.4. The decomposition is explicitly 'sampling the ground state of the manifestly stoquastic Hermitian Hamiltonian H_eta' and then evaluating ⟨iξD⟩ as a mixed estimator. That means the walkers never feel the iξD term. At the self-learning fixed point the RNN guide is trained to the H_eta ground state, not the H ground state, and by symmetry ⟨D⟩ vanishes on that state. The paper says this is exact when the guide is a good approximation to the ground state of H itself, but the guide is being trained on H_eta walkers, so that condition is circular. If the main NH results in Fig. 9 use the gauge rotation instead, that method is sound in principle, but the paper never states which method produced Fig. 9, and it never reports the value of ξ used. Without ξ the NH benchmarks are not reproducible, and the 'regardless ... non-Hermitian' claim is unsupported. This is the load-bearing weakness.\n\nThe convergence of the self-learning loop is asserted rather than proved; Appendix B only shows energy per stint for two models. That is a minor issue for the Hermitian part because the numbers are clean, but it matters for the NH part because the loop is doing something different from what the prose claims. No code or data is shipped; 'available upon reasonable request' is below current reproducibility norms.\n\nWho is this for? Anyone working on neural-guided QMC or on the non-Hermitian many-body problem will get value from the Hermitian method and from the discussion of why the NH extension is subtle. It deserves serious refereeing: the Hermitian method is a real step forward and the NH claim needs referee time to either fix or properly scope. I would not desk-reject. My recommendation: send to review, and ask the authors to clarify which NH method produced the figures, report ξ, and either repair the C.4 estimator or remove the 'regardless of non-Hermitian' language from the abstract.","headline":"Solid Hermitian RNN-PQMC method, but the non-Hermitian claim is undermined by the real-imaginary decomposition and unreported ξ.","tokens_in":21612,"tokens_out":5940,"would_cite":true,"duration_ms":57577,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Autoregressive PQMC: an RNN guide sharpens ground-state energies to about $10^{-6}$ relative error, for both Hermitian and PT-symmetric non-Hermitian Ising chains.","keywords":["projective quantum Monte Carlo","recurrent neural network wave functions","autoregressive sampling","self-learning PQMC","non-Hermitian quantum systems","transverse-field Ising chain","sign problem","neural quantum states"],"falsifier":"Run the same self-learning loop on a stoquastic Hamiltonian with a known exact ground state that a one-layer RNN cannot represent exactly, such as a frustrated two-dimensional Ising model, and check whether the mixed-estimator energy converges to the exact value as stints grow; a capacity-dependent plateau would indicate that the loop's fixed point is biased by the ansatz. For the non-Hermitian model, monitor the imaginary part of the local energy after the gauge rotation: persistent nonzero values would signal that the walk is not sampling the true right ground state.","tokens_in":20348,"feed_emoji":"⚛️","tokens_out":12056,"duration_ms":131206,"temperature":0.7,"pith_summary":"The paper introduces autoregressive projective quantum Monte Carlo (PQMC), where a recurrent neural network supplies the guiding wavefunction and is retrained on the walker distribution every few projection steps. It claims this guided version is substantially more accurate than standard unguided PQMC while retaining polynomial cost, for both Hermitian and non-Hermitian Hamiltonians. On the staggered transverse-field Ising chain the method reaches relative energy errors around $10^{-5}$ to $10^{-6}$ for systems of 50--150 spins and beats a variational RNN by roughly two orders of magnitude. A sympathetic reader should care because PQMC is a leading method for sign-problem-free Hamiltonians, and a reliable self-learning guide removes its main practical bottleneck: the need for a good hand-built or separately optimized trial state.","feed_headline":"Guided quantum Monte Carlo hits ground states to 1e-6 error","feed_subtitle":"An autoregressive neural guide trained on walkers beats unguided PQMC and variational RNNs on Ising chains.","key_machinery":"The load-bearing mechanism is the self-learning loop, not any single network design. A recurrent neural network parameterizes the guiding wavefunction through the autoregressive chain rule, so every spin configuration gets an explicit normalized probability and can be sampled exactly, in sequence, without a sampling step over hidden units. Each PQMC stint uses that RNN for importance sampling and continuous-time imaginary-time propagation; the walker configurations are then fed back as training data for a maximum-likelihood (relative-entropy) update of the network parameters. The autoregressive factorization makes this fit cheap and stable, and the continuous-time propagation removes finite-time-step bias. For the non-Hermitian model, a local unitary gauge rotation turns the complex off-diagonal terms into a stoquastic Hamiltonian, and the local energy is evaluated with a biorthogonal mixed estimator that needs only the right ground state.","core_discovery":"The paper's central claim is that a recurrent neural network can act as a self-improving guiding wavefunction inside projective quantum Monte Carlo (PQMC), and that this combination is substantially more accurate than the unguided version at polynomial cost. At each stint, the RNN's autoregressive wavefunction $\\psi_{\\theta_s}(x)$ is used for importance sampling; the walkers generated by imaginary-time projection are then used to retrain the RNN by maximizing the likelihood of the walker configurations, equivalently minimizing the relative entropy between the network distribution and the walker distribution. The paper shows that after about five stints the mixed-estimator energy stabilises at the exact ground-state energy of the Hermitian staggered transverse-field Ising chain, with relative errors of $2.3\\times10^{-6}$, $5.9\\times10^{-6}$, and $1.2\\times10^{-5}$ at $N=50$, $100$, and $150$ spins, and that the same loop works for a PT-symmetric non-Hermitian Ising chain after a unitary gauge rotation makes the Hamiltonian sign-problem-free. The method outperforms variational RNNs and unguided PQMC in both accuracy and variance.","pith_inferences":["If the self-learning loop converges for the same reason in other stoquastic models, the recipe should transfer to higher-dimensional and frustrated lattices, provided the RNN ansatz can represent the relevant correlations; the paper benchmarks only one-dimensional chains.","The gauge-rotation trick is model-specific; applying the method to a non-Hermitian Hamiltonian without a known stoquastic rotation would test whether the sign-problem mitigation, rather than the RNN guide, is the limiting ingredient.","The biorthogonal mixed estimator uses only the right ground state; the same framework could be extended to left-ground-state observables or to excited states by symmetry projection, which the paper does not explore."],"forward_implications":["For a Hermitian staggered transverse-field Ising chain with up to 150 spins, the guided method's relative energy error is about two orders of magnitude smaller than a variational RNN's and remains below $10^{-5}$.","The number of stints needed for convergence stays around five to ten across system sizes, about four times fewer than the RBM-guided self-learning PQMC baseline.","The same RNN-guided pipeline applies to PT-symmetric non-Hermitian Ising chains after a gauge rotation, so non-Hermitian ground-state energies can be obtained without a separate variational optimization.","The computational cost remains polynomial in system size, with wall-clock times comparable to unguided PQMC for the system sizes tested.","Because the mixed estimator is not variational, the method's accuracy is tied to the quality of the guide; the more expressive the RNN, the smaller the fluctuations around the true ground state."],"supporting_citations":[{"why":"Introduces self-learning PQMC guided by restricted Boltzmann machines, the baseline this paper extends to recurrent neural networks.","marker":"[14]"},{"why":"Supplies the variational RNN benchmark and the non-Hermitian Ising chain setup.","marker":"[20]"},{"why":"Provides the recurrent neural network wavefunction ansatz and autoregressive sampling machinery.","marker":"[21]"},{"why":"Supplies the high-order series expansion benchmark for the non-Hermitian model.","marker":"[52]"},{"why":"Gives the projective quantum Monte Carlo formalism including branching and the propagator treatment.","marker":"[9]"},{"why":"Provides the continuous-time PQMC algorithm used to eliminate finite-time-step bias.","marker":"[33]"},{"why":"Supplies the exact solution of the Ising chain used as the Hermitian benchmark.","marker":"[50]"}],"fun_headline_variants":["RNN-guided PQMC hits 1e-6 error on Hermitian and non-Hermitian chains","Autoregressive PQMC with RNN guide beats variational ansatze","Neural-guided PQMC: from Hermitian to non-Hermitian with 1e-6 accuracy","Self-training RNN walks quantum Monte Carlo to 1e-6 error"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that repeatedly retraining the RNN on the current PQMC walker distribution converges to the true ground state rather than to a biased fixed point, and that five to ten stints are enough; the paper supports this with stint-by-stint energy plots for two models, not with a proof.","fun_headline_variants_meta":{"raw":{"variants":["RNN-guided PQMC hits 1e-6 error on Hermitian and non-Hermitian chains","Autoregressive PQMC with RNN guide beats variational ansatze","Neural-guided PQMC: from Hermitian to non-Hermitian with 1e-6 accuracy","Self-training RNN walks quantum Monte Carlo to 1e-6 error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000764,"raw_usage":{"total_tokens":3382,"prompt_tokens":928,"completion_tokens":2454,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":2360}},"tokens_in":544,"tokens_out":2454,"duration_ms":17102,"temperature":1.0,"reasoning_tokens":2360,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:21:42.205115+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same self-learning loop on a stoquastic Hamiltonian with a known exact ground state that a one-layer RNN cannot represent exactly, such as a frustrated two-dimensional Ising model, and check whether the mixed-estimator energy converges to the exact value as stints grow; a capacity-dependent plateau would indicate that the loop's fixed point is biased by the ansatz. For the non-Hermitian model, monitor the imaginary part of the local energy after the gauge rotation: persistent nonzero values would signal that the walk is not sampling the true right ground state.","supporting_citations":[{"cited_title":"Kr¨ amer, J","cited_arxiv_id":null,"evidence_quote":"Introduces self-learning PQMC guided by restricted Boltzmann machines, the baseline this paper extends to recurrent neural networks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the variational RNN benchmark and the non-Hermitian Ising chain setup."},{"cited_title":"Isozaki, Many-body schr¨ odinger equation, Springer, Tokyo https://doi.org/10.1007/978-981-99-3704-2 (2004)","cited_arxiv_id":null,"evidence_quote":"Provides the recurrent neural network wavefunction ansatz and autoregressive sampling machinery."},{"cited_title":"Feldbacher, K","cited_arxiv_id":null,"evidence_quote":"Provides the continuous-time PQMC algorithm used to eliminate finite-time-step bias."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the exact solution of the Ising chain used as the Hermitian benchmark."}],"review_version":1}