{"id":"ebd75bbd-be34-4e50-9ead-4c2b3ab52f12","arxiv_id":"2506.11227","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An optimized finite-temperature resonating valence bond state captures measured spin and dopant correlations of doped Fermi-Hubbard simulators on square and triangular lattices.","lead":"This paper fits a simple quantum state, a resonating valence bond spin liquid, to snapshots from ultracold-atom Fermi-Hubbard simulators by tuning a few parameters with gradient descent. The optimized state reproduces most measured spin and dopant correlations, suggesting simple states can describe finite-temperature doped Hubbard physics.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The doping dependence of the variational parameters is not identifiable from the flat loss landscape, so the physical spin-liquid interpretation is not supported.","rationale":"I focused on parameter non-identifiability because it targets the physical-insights conclusion that is advertised as central, and it is directly exposed by the paper's own benchmark: even in the ideal case of fitting the ansatz to data generated from the same ansatz, the exact parameters cannot be recovered. A conditional verdict remains appropriate: the correlation-capture workflow is partly validated by held-out correlations such as the d=2 spin correlator, but the parameter interpretation should not be accepted without the identifiability check described above. I also checked the abstract's claim that three-point correlations are captured \"even when they are not specifically used in the optimization\"; for the triangular lattice, CDZZ is explicitly included in the objective for the Ref. [18] dataset, so that particular overstatement is a wording issue rather than the most load-bearing scientific concern. I therefore partially agree with the reader's weakest-assumption statement: the reader includes ansatz expressivity and h/JMF=0, whereas my main concern is the identifiability of the fitted parameters within the ansatz.","tokens_in":13016,"tokens_out":8107,"duration_ms":104665,"concrete_test":"For each doping value in the square-lattice datasets of Refs. [12] and [25], run a large ensemble (e.g., 100) of optimizations from uniformly sampled random initial conditions, and record every final parameter set whose objective L is within the Monte Carlo statistical uncertainty (or within 10%) of the best L. Plot the union of these low-loss parameter sets in the (phi, TMF/JMF) plane as a function of doping. If the low-loss sets for adjacent dopings overlap substantially, then the doping trend of TMF/JMF in Fig. 5B is not identifiable from the data, and the physical interpretation drawn from that trend should be withdrawn or explicitly qualified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's second central claim is that the optimized mean-field parameters have physical meaning: their doping dependence \"can be used to obtain physical insights\" into a possible underlying spin liquid (Sec. VII). This requires that the experiment data determine the parameters. The paper's own benchmarks show they do not. Fig. 1A displays an arc-shaped region of near-identical objective values, and the benchmark in Sec. IV (Fig. 2) demonstrates that even when the reference data are generated from the very same ansatz, gradient descent cannot recover the target parameters: the final parameters are \"significantly spread out\" despite reaching L on the order of 10^-5. The authors acknowledge only that a low-loss region is identified. Consequently, for each experimental doping, the reported optimized point (phi, TMF/JMF) is partly an artifact of initialization and the optimization trajectory. The doping-dependent trend of TMF/JMF in Fig. 5B is therefore not a reliable physical observable unless it is shown to be shared by the entire low-loss manifold. This does not undermine the correlation-capture result, but it removes the evidential basis for the claimed interpretation of the parameters as revealing a possible quantum spin liquid.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a method to optimize the variational parameters of finite-temperature Gutzwiller-projected resonating-valence-bond (RVB) states against correlation functions measured in cold-atom Fermi-Hubbard quantum simulators. The ansatz is a staggered-flux (SF) or staggered-flux-plus-Néel (SFN) mean-field state with at most three parameters: the flux φ, the effective temperature TMF/JMF, and the Néel field h/JMF. The authors benchmark the gradient-descent optimizer on self-generated Monte Carlo data and on quantum Monte Carlo data for the Heisenberg model, then apply it to square-lattice and triangular-lattice experimental datasets. They report that the optimized RVB states reproduce two- and three-point correlation functions reasonably in the hole-doped regime, including a held-out distance on the square lattice, and they interpret the doping dependence of the optimized parameters as evidence for a possible underlying spin-liquid description.","tokens_in":13249,"tokens_out":5795,"duration_ms":72785,"significance":"If the correlation-capture claim survives scrutiny, this is a valuable demonstration that a very small number of physically motivated parameters can describe local correlations of the doped Fermi-Hubbard model at finite temperature, providing an interpretable alternative to high-dimensional variational states. The paper also contains honest benchmarks and an explicit discussion of the flat optimization landscape. The main strength is the comparison against QMC Heisenberg data and the held-out distance d=2 for the square lattice, which give independent support beyond simple fitting. The main weakness is that the optimized parameters are not identifiable from the flat loss landscape, which undermines the second central claim that the doping dependence of the parameters can be directly interpreted as physical insight.","major_comments":[{"comment":"The benchmark on self-generated data shows that the optimizer cannot recover the true target parameters even when the reference data come from the same ansatz: the final parameters are 'significantly spread out' while the objective reaches values of order 10^-5. This confirms that the loss landscape contains a large low-loss manifold. Since the experimental analysis in Section V and the physical interpretation in Section VII rely on point estimates (φ, TMF/JMF) and their doping dependence, the paper must demonstrate that the reported doping trends are shared by the full low-loss manifold, rather than being artifacts of initialization and trajectory. I would like to see, for each doping, the distribution of optimized parameters over many runs together with a statement of which functionals of the parameters (if any) are stable across the manifold.","section":"Section IV, Fig. 2"},{"comment":"The objective function is an equally weighted sum of squared deviations over all observables and distances, with no treatment of experimental uncertainties or Monte Carlo sampling errors. In a flat landscape small changes in the objective can move the optimum substantially, so the reported optimized parameters are not well defined unless the sensitivity to the weighting scheme is quantified. For example, if one reweights CZ(d=1) relative to CZ(d=sqrt2) or CD(d=sqrt2), do the doping trends in Fig. 5B survive? Without such a sensitivity analysis, the 'physically meaningful' interpretation of the parameters is not supported.","section":"Section V, Eq. (8)"},{"comment":"The abstract states that finite-temperature data from Fermi-Hubbard quantum simulators 'can be well captured by RVB states,' and the introduction and conclusion repeat this broad claim. However, Section VI reports that for the triangular lattice at large particle doping the ansatz fails to reproduce the sign change in CZ(d=1) and also misses details of CDZZ in the intermediate particle-doped regime. The claim should be qualified to specify the regime in which the ansatz is successful, namely hole doping on both lattices and, on the square lattice, the investigated doping ranges. As written, the central claim overstates the evidence in the paper itself.","section":"Abstract and Section VI"},{"comment":"The paper justifies setting h/JMF=0 for the experimental analyses by stating that at T/J≈0.65–0.8 the optimized h/TMF is very small based on the QMC benchmark in Section IV. This is not quantitatively demonstrated: Fig. 3B shows a monotonic decrease of h/TMF with T/J, but at T/J=0.7 the value appears to be a sizable fraction of the range shown. Please provide the actual optimized h/TMF values from the benchmark for the relevant T/J window, or explicitly state the threshold below which h/JMF is deemed negligible. Without this, the reduction to 'only two variational parameters' for the experimental fits is not fully justified.","section":"Section V, setting h/JMF=0"}],"minor_comments":[{"comment":"There are several typographical issues, including inconsistent spacing in 'R VB', nonstandard accents such as 'N´ eel' and 'Ans¨ atze', and occasional grammatical slips. These should be corrected in a final revision.","section":"Throughout"},{"comment":"The caption says the SF state captures CZ(d) for all distances considered 'even if we do not consider a given distance in the optimization as for d=2 and [25].' For the [25] dataset, both d=1 and d=sqrt2 are used in the optimization, so the held-out claim applies only to d=2 for the [12] dataset. Please clarify the wording.","section":"Section V, Fig. 4 caption"},{"comment":"The comparison to the non-optimized π-flux state would benefit from a precise definition of that state and from a statement of whether its parameters are fixed to φ=π and TMF/JMF=T/J in all datasets, or only for the Chiu et al. data.","section":"Section V, around Eq. (16)"},{"comment":"The paper does not provide details on the number of Monte Carlo samples, the Metropolis proposal, the step size for finite-difference gradients, or the stopping criterion. Since the method is a central contribution, these numerical details should be reported for reproducibility.","section":"Section III, Eqs. (8)-(10)"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest about its limitations, and the correlation-capture result is likely salvageable through a more careful treatment of parameter identifiability. The flat landscape is acknowledged in the text, so the physical-parameter interpretation is the load-bearing point that needs additional work rather than a fatal flaw. I would encourage the editor to request a revision that either demonstrates that the doping trends are robust across the low-loss manifold or substantially weakens the interpretive claims. The manuscript is a better fit for a journal that values interpretable variational ansätze for quantum simulator data, and it could become a useful methodological reference after revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Something worth reading here. The new thing is practical: a gradient-descent pipeline that optimizes the two or three mean-field parameters of a finite-temperature Gutzwiller-projected RVB ansatz against correlation functions from quantum simulator snapshots. That is a real step beyond the fixed pi-flux comparisons in Chiu et al. The benchmarks are decent: they get the loss down to about 1e-5 on self-generated data, reproduce QMC Heisenberg correlations, and the held-out CZ(d=2) on the square lattice plus the triangular CDZZ behavior are genuinely not fit, so the correlation-capture claim mostly holds. They also beat the unoptimized ansatz on the charge sector, which is nice.\n\nThe soft spot is exactly the one in the stress-test: the parameter values are not identifiable. The arc-shaped loss landscape in Fig. 1A and the benchmark in Fig. 2 show this openly—even when the data is generated from the same ansatz, the optimizer lands on significantly different (phi, TMF/JMF) with equally small loss. Given that, the doping trend of TMF/JMF in Fig. 5B is partly an artifact of initialization, and the abstract's statement that the parameter analysis can be used to obtain physical insights and reveal a spin liquid goes beyond what the data determines. That is not a fatal flaw for the correlation-capture part, but it does mean the paper's second headline claim is unsupported as written. The authors could fix it by reporting the whole low-loss manifold for each doping, or by showing the trend is shared across runs and initializations—they only report the mean with a standard deviation, which understates the flatness.\n\nOther minor issues: no code or data shipped, so the MC implementation is not independently checkable; a few details of the sampling and gradient estimator are terse; and the ansatz by construction excludes doublon-hole pairs, which limits it at larger doping and in the particle-doped triangular regime, where they themselves note the mismatch.\n\nWho is this for? People working on interpreting quantum gas microscope data with variational spin-liquid states. It deserves a serious referee round, not a desk reject. The revision should address identifiability and release the code.","headline":"Useful optimization pipeline for RVB states against quantum simulator data, but the doping-dependent parameter trends are underdetermined by the flat loss landscape.","tokens_in":13777,"tokens_out":2498,"would_cite":true,"duration_ms":28476,"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":"This paper claims that finite-temperature resonating valence bond states, with only two optimized mean-field parameters, reproduce measured two- and three-point correlations in doped Fermi-Hubbard quantum simulators.","keywords":["resonating valence bond states","Gutzwiller projection","Fermi-Hubbard model","quantum gas microscope","variational Monte Carlo","spin correlations","doping dependence","quantum spin liquid"],"falsifier":"The claim would be refuted if a high-precision experiment and an unbiased numerical simulation of the Hubbard model agreed on a correlation—for instance a four-point correlator, or the sign of $C_Z$ at $d=\\sqrt2$ for large hole doping on the square lattice—that no choice of $\\phi$ and $T_{\\mathrm{MF}}/J_{\\mathrm{MF}}$ reproduces.","tokens_in":12820,"feed_emoji":"⚛️","tokens_out":7764,"duration_ms":77290,"temperature":0.7,"pith_summary":"This paper claims that finite-temperature resonating valence bond (RVB) states—a simple family of spin-liquid wavefunctions in which double occupancies are projected out—can describe the local correlations seen in ultracold-atom Fermi-Hubbard simulators. The authors optimize just two mean-field parameters, the staggered flux $\\phi$ and the ratio $T_{\\mathrm{MF}}/J_{\\mathrm{MF}}$, by gradient descent against experimental spin-spin and dopant correlation functions on square and triangular lattices. They find that the optimized states match two-point and three-point correlations that were not included in the optimization, and that the fitted parameters vary with doping in a physically sensible way. If the claim holds, thousands of parameters in numerical variational states are unnecessary for capturing the local physics of these doped antiferromagnets.","feed_headline":"Two-parameter spin-liquid state matches Hubbard simulator data","feed_subtitle":"Optimized RVB states reproduce two- and three-point correlations on square and triangular lattices.","key_machinery":"The machinery is a finite-temperature RVB ansatz built from a quadratic mean-field Hamiltonian with staggered flux $\\phi$ and a Néel field $h$; the thermal density matrix of this Hamiltonian is Gutzwiller-projected by restricting Monte Carlo configurations to singly occupied sites. Expectation values are estimated via Metropolis sampling over pairs of real-space and mean-field basis states, and gradients of the objective $L$ are computed from the same samples using finite differences. This lets the authors run gradient descent over the two-parameter landscape, whose arc-shaped valley explains why many parameter sets give nearly identical correlations. The key work of the machinery is to turn experimental snapshots into a handful of interpretable parameters while exposing which experimental correlations the ansatz cannot represent.","core_discovery":"The central discovery is that a Gutzwiller-projected finite-temperature mean-field state, with the Néel field set to zero and only $\\phi$ and $T_{\\mathrm{MF}}/J_{\\mathrm{MF}}$ varied, captures the measured two- and three-point correlations of Fermi-Hubbard quantum simulators. Using gradient descent with a Metropolis-sampled objective, the authors reproduce spin-spin correlations $C_Z$ at distances not used in the fit, the dopant-dopant function $\\tilde g^{(2)}(d=\\sqrt2)$, and the connected dopant-spin-spin correlator $C_{DZZ}$ on the triangular lattice. The optimized ratio $T_{\\mathrm{MF}}/J_{\\mathrm{MF}}$ rises with doping up to about 25% and then falls, consistent with a crossover from Heisenberg to free-fermion physics. The paper also reports where the ansatz fails: the sign change of $C_Z(d=\\sqrt2)$ at large square-lattice doping and the ferromagnetic correlations on the particle-doped triangular lattice are not captured.","pith_inferences":["The flat, arc-shaped optimization landscape means individual optimized parameters are not uniquely determined; physical conclusions should be drawn from the matched correlation functions rather than from the specific values of $\\phi$ and $T_{\\mathrm{MF}}/J_{\\mathrm{MF}}$.","The specific failures—the square-lattice $d=\\sqrt2$ sign change and the particle-doped triangular lattice—suggest that physics outside the no-doublon spin-liquid picture, such as doublon-hole fluctuations or different magnetic order, becomes relevant in those regimes.","The same optimization pipeline could be applied directly to other lattice geometries and to data from Rydberg tweezer arrays or $t$–$J$ models, searching for hidden RVB signatures.","A natural extension is to optimize against the full distribution of experimental snapshots via the Kullback-Leibler divergence, rather than a small set of correlation functions, which would test whether the ansatz captures correlations beyond those studied here."],"forward_implications":["The optimized RVB states describe hole-doped correlations on the triangular lattice across the full measured doping range, including the three-point correlator $C_{DZZ}$.","A fixed $\\pi$-flux RVB state with unoptimized parameters performs noticeably worse in the charge sector, so parameter optimization is essential to capturing the measured $\\tilde g^{(2)}(d=\\sqrt2)$.","The doping dependence of $T_{\\mathrm{MF}}/J_{\\mathrm{MF}}$ can be read as a physical crossover, with Heisenberg-like behavior near half-filling and free-fermion-like behavior at high doping.","Because only two parameters are used, the ansatz offers an interpretable benchmark against which more complex variational states, such as those containing doublon-hole pairs, can be compared."],"supporting_citations":[{"why":"Introduces resonating valence bond states as a description of doped Mott insulators, the physical picture the ansatz embodies.","marker":"[2]"},{"why":"Supplies the square-lattice experimental spin and dopant correlations used in the optimization, as well as the fixed pi-flux baseline the optimized state outperforms.","marker":"[12]"},{"why":"Develops finite-temperature RVB descriptions of Fermi-Hubbard simulators that the present optimization extends.","marker":"[13]"},{"why":"Provides the triangular-lattice connected dopant-spin-spin correlations C_DZZ used as reference data.","marker":"[18]"},{"why":"Gives the mean-field construction with staggered flux from which the trial Hamiltonian is built.","marker":"[19]"},{"why":"Supplies the variational Monte Carlo sampling techniques used to evaluate projected thermal expectation values.","marker":"[20]"},{"why":"Provides quantum Monte Carlo Heisenberg-model data used to benchmark the optimization at finite temperature.","marker":"[23]"},{"why":"Provides square- and triangular-lattice experimental spin correlations from a second dataset, used for optimization and for distances not included in the objective.","marker":"[25]"}],"fun_headline_variants":["ML-optimized RVB state matches Hubbard simulator data","Two-parameter spin liquid predicts unseen Hubbard correlations","Gutzwiller ansatz captures three-point correlations in simulators"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument stands on the assumption that the real Fermi-Hubbard state at U/t of roughly 8–9 and T/J near 0.65–0.8 is well represented by a Gutzwiller-projected spin-liquid ansatz with no double occupancies and with the Néel field set to zero.","fun_headline_variants_meta":{"raw":{"variants":["ML-optimized RVB state matches Hubbard simulator data","Two-parameter spin liquid predicts unseen Hubbard correlations","Gutzwiller ansatz captures three-point correlations in simulators"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000529,"raw_usage":{"total_tokens":2575,"prompt_tokens":994,"completion_tokens":1581,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":1528}},"tokens_in":610,"tokens_out":1581,"duration_ms":17693,"temperature":1.0,"reasoning_tokens":1528,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:12:12.088254+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The claim would be refuted if a high-precision experiment and an unbiased numerical simulation of the Hubbard model agreed on a correlation—for instance a four-point correlator, or the sign of $C_Z$ at $d=\\sqrt2$ for large hole doping on the square lattice—that no choice of $\\phi$ and $T_{\\mathrm{MF}}/J_{\\mathrm{MF}}$ reproduces.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces resonating valence bond states as a description of doped Mott insulators, the physical picture the ansatz embodies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the square-lattice experimental spin and dopant correlations used in the optimization, as well as the fixed pi-flux baseline the optimized state outperforms."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the triangular-lattice connected dopant-spin-spin correlations C_DZZ used as reference data."},{"cited_title":"Classifying Snapshots of the Doped Hubbard Model with Machine Learning","cited_arxiv_id":"1811.12425","evidence_quote":"Gives the mean-field construction with staggered flux from which the trial Hamiltonian is built."},{"cited_title":"Koepsell, D","cited_arxiv_id":null,"evidence_quote":"Supplies the variational Monte Carlo sampling techniques used to evaluate projected thermal expectation values."},{"cited_title":"Semeghini, H","cited_arxiv_id":null,"evidence_quote":"Provides quantum Monte Carlo Heisenberg-model data used to benchmark the optimization at finite temperature."},{"cited_title":"Wen, Physical Review B 65, 165113 (2002)","cited_arxiv_id":null,"evidence_quote":"Provides square- and triangular-lattice experimental spin correlations from a second dataset, used for optimization and for distances not included in the objective."}],"review_version":1}