{"id":"22d3859d-1d86-4a65-bf7d-a8c77ee1b25a","arxiv_id":"2412.12019","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A size-invariant graph neural network, trained on small DMRG-simulated Rydberg Ising arrays, predicts atom positions on larger arrays from spin correlators, backed by a bijection theorem between correlations and interactions.","lead":"Researchers trained a graph neural network on small simulated Rydberg-atom arrays and showed it can infer atom positions on larger arrays from measured spin correlations. The paper also proves a theorem that, in the transverse-field Ising model, the full set of two-spin correlations uniquely determines all interaction strengths, supporting the learning approach.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Appendix A proves a full-correlation-matrix bijection, but the paper's componentwise claim that NN/NNN correlators determine the corresponding couplings, and the GNN's actual local-input map, are not consequences of that proof.","rationale":"Appendix A's HK-Henderson argument appears sound as a full-matrix statement: positivity of the ground state in the Z basis, uniqueness, and the variational contradiction give injectivity of J -> c. My concern is not with the proof itself but with the use made of it. The paper's Section II B and Conclusions import a componentwise conclusion that does not follow. The central in-principle claim in the abstract is about determining Hamiltonian parameters from correlation functions; the actual protocol uses a strict subset of correlators. A global injective map can have non-injective projections, so the theorem cannot be cited as the theoretical foundation for the local-feature GNN. The reader's weakest assumption focused on finite-snapshot data not being exactly J-representable; that is a valid practical limitation, but I regard it as less load-bearing because the paper explicitly trains on snapshot data and reports finite-snapshot errors, whereas the local-vs-global gap concerns the infinite-precision, exact-correlator limit and the paper's stated 'in principle' claim. The GNN results on simulated data are strong and would survive even if this gap were left open, so the verdict stays CONDITIONAL rather than REJECT: the authors should either prove or numerically support the restricted injectivity, or soften the componentwise statements in Section II B. The proposed exact-diagonalization search is a direct, cheap falsification test: finding a collision would require substantial revision of the theoretical narrative; not finding one would turn the issue into a documentation/rigor problem rather than a demonstrated failure.","tokens_in":27793,"tokens_out":14005,"duration_ms":139369,"concrete_test":"For a small lattice (e.g., 4x4, using exact diagonalization) at fixed Omega and delta, compute the ground-state NN and NNN correlator features for many random displacement configurations around the nominal positions. Then, for a set of target configurations, run a multi-start nonlinear least-squares search for a distinct displacement configuration whose local correlator vector matches the target to numerical precision (e.g., with scipy.optimize.least_squares). If two distinct configurations with the same local feature vector are found, the componentwise/local bijection asserted in Section II B is false and the GNN's input representation is non-unique. If no collisions emerge across many targets and random starts, the restricted map is empirically injective on the physical manifold, reducing the concern to a proof gap rather than a demonstrated failure.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Appendix A establishes injectivity of the map J -> c on the full matrix of all pairwise ground-state correlators for fixed Omega and delta, and the proof is coherent for that statement. The paper then overreads it. Section II B says that 'any two-body correlation function ... is uniquely mapped to a corresponding Hamiltonian interaction term' and that NN correlation functions suffice to determine NN couplings. Injectivity of a full-matrix map does not imply injectivity of a coordinate projection: two different interaction matrices with identical NN/NNN correlators but different NN couplings are not excluded by the theorem. The GNN input in Section II C 2 consists only of local magnetizations and NN/NNN correlators across an Omega-history, not the full c matrix. Thus the theorem does not guarantee that the map the GNN learns is single-valued. If two distinct displacement configurations of the physical array produce the same local feature vector, then 'local correlation measurements suffice in principle' is false even in the infinite-snapshot, exact-correlator limit, and the learned predictions would be ambiguous. The numerical results on the sampled disorder distribution are evidence but do not settle this: collisions may simply be rare or absent on that distribution. The finite-snapshot J-representability issue noted by the reader is secondary: even with exactly J-representable inputs, the local feature map is not covered by the theorem.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a graph neural network (GNN) approach to learning the interaction parameters of a transverse-field Ising model (TFIM) realized by Rydberg atom arrays. The GNN is trained on DMRG-generated observables — local magnetizations and nearest-neighbor/next-nearest-neighbor spin correlators — for small square arrays (4x4 to 6x6) and is reported to extrapolate to larger and rectangular arrays up to 9x9 with small prediction errors. The authors also prove a theorem in Appendix A establishing a bijection between the full set of two-point ground-state correlation functions and the interaction matrix for fixed nonzero transverse field and fixed detuning. The manuscript claims that this theorem provides a theoretical foundation for the learning algorithm and, in particular, that nearest-neighbor correlation functions suffice to determine nearest-neighbor couplings and atomic displacements. A companion numerical study examines training scenarios, snapshot-size effects, comparison with a multilayer perceptron, and the scaling of errors with the number of projective measurements.","tokens_in":27972,"tokens_out":10623,"duration_ms":92997,"significance":"If the claims held as stated, this would be a significant contribution: it would provide a scalable and fast Hamiltonian-learning protocol for Rydberg simulators, with a concrete path toward feedback control of optical tweezers. The numerical work is extensive and carefully executed, and the authors make the data and code available on Zenodo and GitHub, which is a clear strength. The full-matrix bijection theorem is a genuine formal result and is proved by a coherent Perron-Frobenius/variational argument. However, the paper's central theoretical justification overreaches: the theorem does not support the componentwise or NN-only uniqueness claims used to motivate the GNN's local input representation. The practical transferability across system sizes is demonstrated empirically but is not proven; the authors partly acknowledge this in the conclusions, but the abstract and Sec. II B present the theorem as a stronger foundation than it is. With appropriate reframing, the numerical study and the full-matrix theorem remain valuable, but the current form needs substantial revision.","major_comments":[{"comment":"The theorem in Appendix A proves that the map from the full interaction matrix J to the full matrix of two-point correlators c (all pairs i<j) is injective for fixed Omega != 0 and fixed delta. The manuscript repeatedly makes stronger componentwise statements, e.g., \"any two-body correlation function ci,j ... is uniquely mapped to a corresponding Hamiltonian interaction term Ji,j\" and \"NN correlation functions suffice to uniquely determine NN couplings\" (Sec. II B, restated in the first paragraph of Appendix A). These do not follow from the theorem: injectivity of the full-matrix map does not imply injectivity of a coordinate projection, and two different interaction matrices with identical NN/NNN correlators but different NN couplings are not excluded. Since the GNN input (Sec. II C 2) consists of local magnetizations and NN/NNN correlators across an Omega-history, not the full c matrix, the theorem does not guarantee single-valuedness of the map the GNN learns. The numerical results on the sampled disorder distribution are evidence but do not rule out collisions. Please either prove the restricted injectivity (or provide a citation that does) or weaken the theoretical claims to match what the theorem actually establishes.","section":"II B and Appendix A"},{"comment":"The bijection theorem is conditional on the correlation matrix being exactly J-representable, i.e., the exact ground-state correlator of a TFIM at fixed Omega and delta. In the snapshot protocol that the paper targets, correlation functions are estimated from finitely many projective measurements and are generically not exactly J-representable. The theorem therefore does not strictly apply to the experimental scenario, and the sentence in Sec. II B that \"This result serves as a theoretical foundation for reconstructing Hamiltonians from experimentally measured correlation functions\" overstates the proof's reach. The snapshot results should be explicitly framed as an empirical robustness study whose relationship to the theorem is heuristic, not a direct consequence.","section":"III C and Appendix A"}],"minor_comments":[{"comment":"References to \"Sec. A\" should be to \"Appendix A\" (two occurrences).","section":"II C 4"},{"comment":"The Hadamard operator H_Hd should be defined explicitly as a tensor product of single-site Hadamard gates; the current wording \"the involutive Hadamard operator defined in the Z-basis\" is ambiguous.","section":"Eq. (4)"},{"comment":"The heading appears as \"DA T A A V AILABILITY\", which is a typographical error.","section":"Data availability"},{"comment":"The phrase \"neural network correlations\" should likely be \"neural-network correlators\" or \"spin-spin correlations\".","section":"Fig. 1 caption"},{"comment":"The sentence \"We prove a theorem establishing a bijective correspondence between the correlation functions and the interaction parameters\" should specify \"the full set of two-point correlation functions\" to avoid implying a componentwise bijection.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The numerical study is thorough and the code/data availability is a strength. The main risk is the gap between the proven full-matrix theorem and the componentwise/NN-only claims that are used to justify the GNN's local input. If the authors can either prove the restricted injectivity (possibly using the Omega-history and the 1/R^6 parametrization) or substantially qualify the theoretical claims, the paper would be publishable. The snapshot-data portion is valuable as an empirical study, but the theorem should not be presented as covering it directly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: credible ML-plus-theory paper with a clean proof that is slightly oversold. The Appendix A bijection between J and the full matrix of two-point correlators for the TFIM at fixed nonzero Ω and δ is correctly argued via Perron-Frobenius and the variational principle; I checked the steps and found no flaw. That is a genuine, citable contribution. The numerical part is also solidly done: DMRG ground states for 4x4 to 6x6, PNA GNN extrapolating to 9x9 and rectangular shapes, snapshot-size scaling, and a fair MLP comparison. The code and data are mostly available (GNN code on GitHub, datasets on Zenodo; the DMRG producer code is on request only, which is a minor knock).\n\nThe soft spot is the gap between the theorem and the learning setup. The theorem proves injectivity of J -> full c-matrix. It does not prove that the NN and NNN correlators alone, or the local feature vector the GNN actually consumes, determine the corresponding couplings. Section II B claims the NN correlators suffice for NN couplings; that is a coordinate projection claim the proof does not establish. The GNN inputs are local magnetizations and NN/NNN correlators across an Ω-history, so the map it learns is not guaranteed single-valued by the theorem. The numerical success on the sampled disorder distribution is evidence that collisions are absent or rare there, but it is not a proof. The finite-snapshot issue (estimated correlators are not exactly J-representable) is secondary; even with exact inputs the local-map gap remains. I'd like the authors to test for collision/ambiguity directly on their disorder distribution, and to state the theorem in its exact full-matrix form.\n\nAlso, the 'minimal set of measurements' phrase is not quantified against any baseline; that is a minor overreach. No experimental data yet, so the practical calibration claims are forward-looking.\n\nOverall: worth serious peer review. The theorem is right, the ML study is careful, and the overclaim is fixable in revision. I'd cite the theorem for sure when working on Rydberg Hamiltonian learning.","headline":"Solid Rydberg-TFIM Hamiltonian learning paper with a correct full-matrix bijection proof, but the componentwise/local-input claims go beyond the proof and need an explicit test.","tokens_in":28627,"tokens_out":1827,"would_cite":true,"duration_ms":16882,"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 authors prove a bijection between spin-spin correlations and interaction parameters in the transverse-field Ising model, and show a graph neural network learns it on small arrays and extrapolates to larger ones.","keywords":["Hamiltonian learning","Rydberg atom arrays","graph neural networks","transverse-field Ising model","spin-spin correlations","bijection theorem","DMRG","extrapolation"],"falsifier":"Train the same GNN on the same $4\\times4$, $5\\times5$, and $6\\times6$ exact data and test it on a Rydberg array whose true atomic positions are independently calibrated by a different method; if predicted nearest-neighbor distances deviate by more than the claimed roughly 10% at 10,000 snapshots, the transferability claim would be refuted. A more direct test of the theorem: prepare two different disorder configurations whose estimated correlators agree within shot noise, and show that the bijection-based inversion assigns them the same interactions.","tokens_in":27495,"feed_emoji":"⚛️","tokens_out":6007,"duration_ms":51973,"temperature":0.7,"pith_summary":"This paper tries to establish that the interaction parameters of a Rydberg-atom transverse-field Ising simulator can be read off from measured spin-spin correlations, and that a graph neural network can learn that map on small simulated systems and carry it to larger and differently shaped arrays. The load-bearing theoretical result is a bijection theorem: for fixed nonzero transverse field and fixed detuning, the ground-state correlation functions $c_{i,j}=\\langle \\hat{S}^z_i \\hat{S}^z_j\\rangle$ determine the interaction matrix $J_{i,j}$ uniquely, and vice versa. If this is right, local correlation measurements in principle suffice to reconstruct the Hamiltonian, and the remaining imprecision is a statistical estimation problem rather than a fundamental one. The authors demonstrate numerically that a size-invariant GNN trained on $4\\times4$ through $6\\times6$ arrays extrapolates to $9\\times9$ and rectangular shapes with only a small error increase, and that combining $z$- and $x$-basis correlators from about 10,000 projective snapshots keeps average distance errors near 10%.","feed_headline":"Correlations uniquely fix Rydberg Hamiltonian couplings","feed_subtitle":"A graph neural net trained on small arrays predicts atomic positions on larger ones from spin-correlation snapshots.","key_machinery":"The theorem is the central object. It generalizes the Hohenberg-Kohn bijection between density and potential, and the Henderson uniqueness theorem for pair correlations, to the Ising interaction matrix. Its proof rests on a lemma: with $\\Omega\\neq0$ the TFIM ground state is unique and has nonvanishing amplitude on every classical spin configuration, so that any supposed equality of two different interaction matrices would have to hold on an overcomplete set of $2^N$ equations; summing over configurations forces the couplings to coincide. The GNN machinery that carries this into practice is a graph neural network with PNA layers and shared task networks that predict a distance from pairs of node embeddings and edge correlators, making the architecture invariant to system size.","core_discovery":"The paper's central claim is the Hohenberg-Kohn-Henderson theorem for the transverse-field Ising model: for the Hamiltonian $\\hat H = \\Omega\\sum_i \\hat\\sigma^x_i + \\delta\\sum_i \\hat\\sigma^z_i + \\sum_{i<j} J_{i,j}\\hat\\sigma^z_i\\hat\\sigma^z_j$ with fixed $\\Omega\\neq0$ and fixed $\\delta$, there is a bijection between the set of $J$-representable correlation functions $c_{i,j}=\\langle \\hat S^z_i \\hat S^z_j\\rangle$ and the interaction parameters $J_{i,j}$. The proof uses stoquasticity and irreducibility to show the ground state is unique and has nonzero amplitude on every $\\hat\\sigma^z$ basis configuration, then applies the variational principle to rule out two different $J$ matrices sharing the same correlators. On the machine-learning side, the claim is that a principal-neighborhood-aggregation graph neural network, with size-independent task networks, learns the local map from magnetization and nearest-neighbor and next-nearest-neighbor correlators to atomic distances on small DMRG-generated ground states and extrapolates to larger and rectangular arrays.","pith_inferences":["If the bijection holds beyond the TFIM as the authors conjecture, the same correlator-to-coupling inversion could be used for other stoquastic or irreducible Hamiltonians, turning Hamiltonian learning into a regression on local observables.","A practical consequence the paper leaves implicit: the learned map could be run in reverse as a diagnostic of optical-tweezer positioning error, and the same snapshot budget could be allocated adaptively across $\\Omega$ values near criticality to reduce uncertainty.","A testable extension would be to add disorder in $\\Omega$ or $\\delta$ and check whether the GNN can separate homogeneous-field errors from positional disorder, which the current protocol does not distinguish.","The near-critical $\\Omega$ sweet spot suggests active-learning strategies: choose measurement fields where correlator variance is largest rather than using a fixed history."],"forward_implications":["With exact correlators, measuring only nearest-neighbor spin correlations should uniquely determine all nearest-neighbor couplings, hence the relative atomic distances.","A GNN trained on small square arrays extrapolates to larger square and rectangular arrays, so training data can be generated classically at small sizes and applied to experimental systems beyond DMRG's reach.","With snapshot data, combining $z$- and $x$-basis correlators from roughly 10,000 projective measurements brings mean absolute errors on distances near 10%.","Graph preprocessing gives an advantage over a direct multi-layer perceptron, especially where edge effects matter.","Including more system sizes and staying near the quantum critical region in the $\\Omega$ history improves extrapolation and prediction error."],"supporting_citations":[{"why":"Supplies the density-to-potential bijection whose logic the theorem generalizes to Ising interactions.","marker":"[72]"},{"why":"The fluid pair-correlation uniqueness theorem the authors call their result a quantum generalization of.","marker":"[74]"},{"why":"Gives the stoquasticity property used to prove the ground state is unique and has all basis amplitudes nonzero.","marker":"[75]"},{"why":"Gives the irreducibility property used in the same lemma.","marker":"[76]"},{"why":"Introduces the density matrix renormalization group used to compute ground states for training data.","marker":"[68]"},{"why":"Provides the density-matrix algorithm for MPS convergence underpinning the DMRG calculations.","marker":"[69]"},{"why":"Defines the principal neighborhood aggregation scheme that makes the GNN layer size-invariant.","marker":"[77]"},{"why":"Supplies the Rydberg van der Waals coefficient $C_6$ and platform parameters defining the simulated Hamiltonian.","marker":"[64]"},{"why":"Demonstrates that a neural network can learn a Hohenberg-Kohn map, motivating the same approach for correlators.","marker":"[97]"},{"why":"Provides the perfect-sampling method used to draw bit-string snapshots from the MPS for snapshot datasets.","marker":"[86]"}],"fun_headline_variants":["GNN learns Rydberg array couplings from correlations","Neural net unlocks Rydberg Hamiltonian from spin snapshots","Correlation snapshots map Rydberg interactions exactly","Bijection theorem powers Rydberg Hamiltonian learning","Graph nets predict Rydberg interactions on larger arrays"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole scheme assumes that the correlations measured in the experiment are exactly the ground-state correlations of a transverse-field Ising model at the same fixed field and detuning used in training, and that what the network learns on small arrays still holds on larger and rectangular ones.","fun_headline_variants_meta":{"raw":{"variants":["GNN learns Rydberg array couplings from correlations","Neural net unlocks Rydberg Hamiltonian from spin snapshots","Correlation snapshots map Rydberg interactions exactly","Bijection theorem powers Rydberg Hamiltonian learning","Graph nets predict Rydberg interactions on larger arrays"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000706,"raw_usage":{"total_tokens":3231,"prompt_tokens":1041,"completion_tokens":2190,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":657,"completion_tokens_details":{"reasoning_tokens":2113}},"tokens_in":657,"tokens_out":2190,"duration_ms":12225,"temperature":1.0,"reasoning_tokens":2113,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:22:31.199876+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same GNN on the same $4\\times4$, $5\\times5$, and $6\\times6$ exact data and test it on a Rydberg array whose true atomic positions are independently calibrated by a different method; if predicted nearest-neighbor distances deviate by more than the claimed roughly 10% at 10,000 snapshots, the transferability claim would be refuted. A more direct test of the theorem: prepare two different disorder configurations whose estimated correlators agree within shot noise, and show that the bijection-based inversion assigns them the same interactions.","supporting_citations":[{"cited_title":"Wietek and A","cited_arxiv_id":null,"evidence_quote":"The fluid pair-correlation uniqueness theorem the authors call their result a quantum generalization of."},{"cited_title":"Scholl, M","cited_arxiv_id":null,"evidence_quote":"Gives the irreducibility property used in the same lemma."},{"cited_title":"Valenti, G","cited_arxiv_id":null,"evidence_quote":"Introduces the density matrix renormalization group used to compute ground states for training data."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the density-matrix algorithm for MPS convergence underpinning the DMRG calculations."},{"cited_title":"Henriet, L","cited_arxiv_id":null,"evidence_quote":"Defines the principal neighborhood aggregation scheme that makes the GNN layer size-invariant."},{"cited_title":"Henderson, A uniqueness theorem for fluid pair cor- relation functions, Phys","cited_arxiv_id":null,"evidence_quote":"Provides the perfect-sampling method used to draw bit-string snapshots from the MPS for snapshot datasets."}],"review_version":1}