{"id":"c05fdb09-a3a6-4060-a0fc-d122a1aade8d","arxiv_id":"2607.25872","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"By scaling a universal fermionic neural network and certifying it with fixed-phase diffusion Monte Carlo, the paper shows the variational energy descending toward the exact ground state, with the DMC gap collapsing to ~10^-6 (quantum dot) and ~2e-4 (2D electron gas) of the total energy.","lead":"Fermi Sets, a neural-network wavefunction that can in principle represent any fermionic state, is scaled up and combined with fixed-phase diffusion Monte Carlo: as the network grows, its raw energy drops and the DMC correction shrinks toward zero, signaling approach to the exact ground state. The protocol is demonstrated on electrons in a magnetic-field quantum dot and in a strongly correlated two-dimensional electron gas.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Phase error unmeasured: vanishing ΔE certifies amplitude only, so abstract's 'convergence to ground state' overreaches.","rationale":"The reader's weakest assumption—that ΔE → 0 implies phase convergence—is indeed the most load-bearing point. Eq. (8) is correct, but the paper's conclusions in the abstract and Discussion rely on an unprobed phase error. The paper is transparent about this in Sec. IIB, which is a credit, but the strong 'indicating convergence' language overstates the certificate. The reader's CONDITIONAL verdict appropriately demands either softening the claim or providing additional evidence for phase accuracy. My proposed test would directly assess the phase error for the dot, the only system with an independent benchmark. The secondary time-step issue for the dot is a concrete technical omission but likely smaller in impact; it reinforces the need for qualification. I find no fundamental error in the methods or the presented scaling trends, so a verdict of ACCEPT/REJECT is not warranted. UNCHANGED keeps the reader's CONDITIONAL, which matches the evidence.","tokens_in":13767,"tokens_out":7839,"duration_ms":77359,"concrete_test":"For the quantum dot at B = 8 T and B = 4 T, run fixed-phase DMC using the phase of the highest-available Landau-level ED ground state (e.g., 8–10 LL) as the trial phase, and compare the resulting fixed-phase energy E_FP^ED with the Fermi-Sets-based E_DMC from Table I. If E_FP^ED is significantly lower (beyond statistical error), the learned phase is suboptimal and the convergence claim fails; if the two agree, the network phase has converged to the ED reference phase. Additionally, repeat the Appendix H time-step scan for the dot's largest network to verify that E_DMC(Δτ) is flat at Δτ = 0.001.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that ΔE → 0 indicates convergence to the exact ground state rests on Eq. (8): E_VMC − E_0 = ΔE + (E_FP − E_0), with both terms nonnegative. The measured ΔE = E_VMC − E_DMC captures only the amplitude error; the phase error E_FP − E_0 is explicitly not probed by ΔE, as acknowledged in Sec. IIB. A vanishing gap is necessary but not sufficient. The paper's indirect evidence for phase convergence is (i) monotone descent of E_DMC across network sizes and (ii) agreement with truncated Landau-level ED, which is itself an upper bound, not the continuum E_0. For the UEG, no independent reference is available. Thus the abstract's wording 'indicating convergence to the ground state' strengthens the result beyond what the diagnostic can certify. A secondary but concrete gap: the assertion in Sec. IIB that residual time-step and population-control biases are 'quantified in Appendix H' is not supported for the quantum dot—Appendix H checks only the 2D UEG. If the dot's DMC energies carry a time-step bias, the reported ΔE values and their collapse could be systematically offset.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage computational framework for continuum fermionic systems: a variational Monte Carlo optimization of the complex-valued Fermi Sets neural-network wavefunction, followed by fixed-phase diffusion Monte Carlo (FPDMC) on the optimized state. The central technical object is the gap ΔE = E_VMC − E_DMC, which by Eq. (8) equals the residual amplitude error for the fixed learned phase. The authors apply the framework to a four-electron parabolic quantum dot at two magnetic fields (B = 4, 8 T) and to a 16-electron 2D uniform electron gas at r_s = 30. In both cases, as the network parameter count is increased, E_VMC and E_DMC descend monotonically, and ΔE shrinks to the order of the statistical uncertainty (0.24–0.25×10⁻⁴ H* for the dot, 0.567(7)×10⁻⁵ Ha/N for the UEG). The paper interprets this collapse as indicating convergence to the exact ground state and uses it to support the claim that Fermi Sets is an asymptotically exact, universal fermion solver.","tokens_in":13977,"tokens_out":5838,"duration_ms":53479,"significance":"If the convergence claim could be fully supported, the framework would be a significant advance: it would provide a systematically improvable continuum solver that, unlike fixed-node DMC, remains valid for time-reversal-broken systems. The paper's strengths are real: the decomposition in Eq. (8) is clean and correct; the DMC projection is a genuinely independent check in the sense that E_DMC is a measured projection result, not a fit to the ED or Tanatar–Ceperley benchmarks; the statistical analysis is careful, with reblocked error bars (Appendix F) and a time-step extrapolation for the UEG (Appendix H). The paper is also honest in Sec. IIB in stating that the phase error E_FP − E_0 is not probed by ΔE. However, the abstract and conclusions go beyond that caveat by asserting that the vanishing gap indicates convergence to the true ground state, which is not logically implied by the diagnostic. The missing quantum-dot time-step check is a secondary but concrete gap. With appropriate revision, the paper would be a useful contribution to the neural-quantum-state and DMC literatures.","major_comments":[{"comment":"The central claim that the collapse of ΔE = E_VMC − E_DMC 'indicates convergence to the ground state' is stronger than what Eq. (8) supports. A vanishing amplitude gap is necessary but not sufficient, since E_FP − E_0 is not bounded by ΔE. The paper explicitly acknowledges this in Sec. IIB, yet the abstract and Sec. IIIA phrase the result as evidence that 'the network finds the true ground state.' The indirect evidence for phase convergence—monotone descent of E_DMC and agreement with truncated Landau-level ED—is suggestive but not conclusive, especially because the ED reference is itself an upper bound on the continuum ground state and no independent reference exists for the UEG. Please rephrase the convergence claim as convergence within the learned phase manifold, and either present the phase-convergence evidence separately with an explicit statement of its indirectness or add a phase","section":"Abstract; Sec. IIB, Eq. (8); Sec. IIIA"},{"comment":"The sentence in Sec. IIB stating that residual time-step and population-control biases are 'quantified in Appendix H and verified to lie below our statistical resolution' is not supported for the quantum dot. Appendix H presents a time-step scan only for the 2D UEG; no analogous scan is reported for the dot. Since the dot's largest-network ΔE values (0.24–0.25×10⁻⁴ H*, Table I) are comparable to the quoted DMC error bars, an unquantified time-step bias could systematically offset the reported ΔE collapse. Please either provide a quantum-dot time-step extrapolation or explicitly restrict the bias claim to the UEG and state what assumption is being made for the dot.","section":"Sec. IIB, Appendix H, Table I"}],"minor_comments":[{"comment":"Typo: 'illustratses' should be 'illustrates'.","section":"Fig. 2 caption"},{"comment":"The unit 'H*' (effective Hartree) is used without definition; please define it at first occurrence.","section":"Table I caption"},{"comment":"The phrase 'in Hartree atomic units reads' is awkward; suggest 'reads, in Hartree atomic units,' or equivalent.","section":"Sec. IIB, Eq. (6)"},{"comment":"The text refers to 'exact diagonalization,' but the calculation is performed in a truncated multi-Landau-level basis (7-LL or 8-LL). This is stated in Table II but could be misread in the main text; please use 'truncated ED' or 'ED in a Landau-level basis' consistently.","section":"Sec. IIIA, Appendix D"},{"comment":"The citation 'Ref. [14,17]' should be formatted as 'Refs. [14,17]' for consistency.","section":"Appendix E, Eq. (E2)"},{"comment":"The sentence 'the steady descent of E_DMC is direct evidence that the learned phase improves systematically with capacity' could be read as stronger than warranted; consider 'consistent with' rather than 'direct evidence,' given the caveat in Sec. IIB.","section":"Sec. IIIB"}],"recommendation":"major_revision","confidential_remarks":"The paper contains a solid and interesting empirical study, but the headline claim overstates what the ΔE diagnostic can certify. The main fix is a careful rewording of the convergence claim and either adding the quantum-dot time-step scan or caveating its absence. I do not see a fatal technical error in the method or the data, so rejection would be disproportionate. The current version is, however, not acceptable as-is because the core assertion in the abstract and conclusions is not logically supported by the presented diagnostic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read: this is the first time complex-valued neural trial states have been pushed through fixed-phase DMC, and the capacity-scaling story is credible. The Fermi Sets architecture itself isn't new, but the combination—using ΔE = E_VMC − E_DMC as a diagnostic while growing the network—is a genuinely useful way to benchmark neural wavefunctions. The quantum dot numbers are impressive: ΔE drops to around 0.25e−4 H* at the largest network, DMC lands below the truncated ED benchmark, and the network learns the L sectors to better than 99.99%. For the UEG, energies descend monotonically with capacity and beat the Tanatar–Ceperley reference. The error bars are handled carefully: reblocked uncertainties, independent VMC/DMC runs, and a time-step scan for the UEG. The decomposition in Eq. (8) is exactly right, and the text in Sec. IIB is honest that a vanishing gap is necessary, not sufficient.\n\nSoft spots: the abstract and Discussion overclaim. A vanishing ΔE only certifies that the amplitude has saturated the learned phase manifold; it does not test the phase error. The indirect evidence for phase convergence—monotone DMC descent and agreement with truncated Landau-level ED, which is itself an upper bound—does not license an unconditional 'convergence to the ground state.' The UEG leg is N=16 only, with no twist averaging or finite-size extrapolation; the authors acknowledge this, but it means the strongest thermodynamic phrasing is not yet supported. Also, no code or data are released, which matters for a heavily numerical claim. And the paper says residual time-step and population-control biases are quantified in Appendix H, but Appendix H only scans the UEG, not the quantum dot. That is a concrete fix, not a fatal flaw. The self-cited universality theorem [12] is load-bearing for interpretation, but the empirical scaling behavior does not depend on it.\n\nWho this is for: people working on neural quantum states and continuum QMC. It deserves a serious referee and is worth citing for the FPDMC-plus-neural-wavefunction combination. I would want the authors to either release artifacts or soften the convergence language to 'convergence of the amplitude within the learned phase,' and to add a dot time-step scan. With those revisions, the paper holds up.","headline":"Solid numerical study; the fixed-phase DMC assessment is a real step forward, but the abstract's 'convergence to ground state' overstates what ΔE can certify.","tokens_in":14628,"tokens_out":1718,"would_cite":true,"duration_ms":16487,"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 scaling up the Fermi Sets neural-network ansatz drives the variational energy toward the exact ground state, and that the residual gap measured by fixed-phase diffusion Monte Carlo collapses to statistical zero, certif","keywords":["Fermi Sets","neural-network wavefunction","fixed-phase diffusion Monte Carlo","variational Monte Carlo","quantum dot","two-dimensional electron gas","composite fermion","ground-state convergence"],"falsifier":"Run the same two-stage protocol on a small system whose exact ground-state energy is known from a much larger exact-diagonalization basis than the paper used, and check whether E_DMC − E0 → 0 while ΔE → 0. A case with ΔE statistically zero but E_DMC clearly above E0 would show the gap metric is not a certificate. A complementary test: deliberately fix a wrong phase (for example, a wrong angular-momentum sector in the quantum dot) but optimize the amplitude; a small ΔE with E_FP well above E0 would demonstrate the same failure directly.","tokens_in":13551,"feed_emoji":"⚛️","tokens_out":5987,"duration_ms":52156,"temperature":0.7,"pith_summary":"This paper tries to establish that Fermi Sets—a neural-network ansatz that provably can represent any continuous antisymmetric fermionic wavefunction—can be scaled to numerically exact interacting ground states. The authors add a fixed-phase diffusion Monte Carlo projection and use the energy drop ΔE = E_VMC − E_DMC as a quality metric. In two very different systems, a magnetic quantum dot and a strongly correlated 2D electron gas, ΔE shrinks monotonically as the network grows, reaching statistical zero in the dot and near-zero in the gas. A sympathetic reader would take this as practical evidence that the ansatz is asymptotically exact and that its own convergence can be certified without knowing the exact answer. The paper is careful that ΔE only bounds amplitude error; the phase error is not probed.","feed_headline":"Fermionic neural networks converge to exact ground states as they grow","feed_subtitle":"Fixed-phase projection shows the wavefunction's residual energy gap shrinking to zero as the network grows.","key_machinery":"The central object is the Fermi Sets wavefunction, a complex-valued ansatz written as a sum of a small number of Slater determinants weighted by symmetric many-body functions Ω_k(R), with learnable complex orbitals; this form can approximate any continuous antisymmetric function and carries nontrivial phase structure. The companion mechanism is fixed-phase diffusion Monte Carlo: the learned phase Φθ is held fixed while only the nonnegative amplitude evolves under an effective Hamiltonian, yielding the fixed-phase energy E_FP. The identity E_VMC − E0 = ΔE + (E_FP − E0) is the machine that turns DMC into a metric: a statistically vanishing ΔE signals that the variational state has saturated al","core_discovery":"On its own terms: Fermi Sets is a universal approximator of continuous antisymmetric wavefunctions, and this paper demonstrates empirically that the ansatz can be systematically scaled. Combining it with fixed-phase diffusion Monte Carlo yields the chain E0 ≤ E_FP ≤ E_VMC and the exact identity E_VMC − E0 = ΔE + (E_FP − E0), where both terms are nonnegative. In the magnetic quantum dot, the measured ΔE falls from about 5.6×10⁻⁴ to 0.25×10⁻⁴ effective Hartree as the network grows to 84k parameters, while the variational energy matches and sometimes beats truncated exact diagonalization. In the fully spin-polarized 2D electron gas at N=16, r_s=30, ΔE falls monotonically to about 0.57×10⁻⁵ Hart","pith_inferences":["If the scaling behavior holds beyond the two testbeds, the ΔE metric could serve as a practical stopping criterion for neural-network variational Monte Carlo, flagging when additional capacity no longer improves the amplitude.","A natural extension the paper leaves implicit: combining the ΔE diagnostic with twist averaging and finite-size extrapolation would carry certified accuracy to the thermodynamic limit, e.g., sharper location of the Wigner-crystallization boundary in the 2D electron gas.","Since the certificate only covers amplitude error, a user who adopts this method as a black-box convergence test would need an independent phase check—such as comparing E_DMC against a more complete exact-diagonalization basis or measuring a known symmetry—before claiming true ground-state convergence.","The paper's logic suggests a sharper test: in systems where the exact phase is known by symmetry, the same two-stage protocol should show both ΔE→0 and E_FP→E0; failure of the second would separate phase convergence from amplitude convergence."],"forward_implications":["The residual gap ΔE between VMC and fixed-phase DMC energies provides a quantitative, system-agnostic measure of a neural wavefunction's residual amplitude error.","Scaling Fermi Sets yields near-exact energies in a time-reversal-broken system, including a composite-fermion state at filling 3/7 that a Slater-Jastrow ansatz cannot describe even qualitatively.","In the strongly correlated 2D electron gas, the Fermi Sets DMC energy improves on the traditional Slater-Jastrow benchmark and provides a variational upper bound on the ground-state energy at that system size.","The framework is stated to apply without modification to fractional quantum Hall liquids and fractional Chern insulators in moiré materials.","The monotone descent of E_DMC with network capacity is presented as direct evidence that the learned phase improves systematically as the network grows."],"fun_headline_variants":["Fermi sets scale to exact ground states, DMC gap shrinks to zero","Neural net wavefunctions converge: DMC confirms vanishing energy gap","Scaling Fermi nets: DMC shows residual gap collapsing to zero","Quantum dot and jellium: Fermi sets approach exact ground states","As Fermi nets grow, DMC gap vanishes—ground state reached"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the learned phase is converging to the exact phase; the paper's ΔE metric certifies only that the amplitude has saturated, and the authors state a vanishing gap is necessary but not sufficient for the exact ground state.","fun_headline_variants_meta":{"raw":{"variants":["Fermi sets scale to exact ground states, DMC gap shrinks to zero","Neural net wavefunctions converge: DMC confirms vanishing energy gap","Scaling Fermi nets: DMC shows residual gap collapsing to zero","Quantum dot and jellium: Fermi sets approach exact ground states","As Fermi nets grow, DMC gap vanishes—ground state reached"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000325,"raw_usage":{"total_tokens":1617,"prompt_tokens":665,"completion_tokens":952,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":409,"completion_tokens_details":{"reasoning_tokens":867}},"tokens_in":409,"tokens_out":952,"duration_ms":8571,"temperature":1.0,"reasoning_tokens":867,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T01:15:26.535885+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same two-stage protocol on a small system whose exact ground-state energy is known from a much larger exact-diagonalization basis than the paper used, and check whether E_DMC − E0 → 0 while ΔE → 0. A case with ΔE statistically zero but E_DMC clearly above E0 would show the gap metric is not a certificate. A complementary test: deliberately fix a wrong phase (for example, a wrong angular-momentum sector in the quantum dot) but optimize the amplitude; a small ΔE with E_FP well above E0 would demonstrate the same failure directly.","supporting_citations":[],"review_version":1}