pith. sign in

arxiv: 2606.22713 · v1 · pith:TYOAVPNEnew · submitted 2026-06-21 · 🪐 quant-ph · physics.comp-ph

An interpretable closed form for entanglement entropy from bitstrings, guided by a graph neural network

Anas Saleh This is my paper

Pith reviewed 2026-06-26 09:52 UTC · model grok-4.3

classification 🪐 quant-ph physics.comp-ph
keywords entanglement entropybitstring distributionsclosed-form expressiongraph neural networkRydberg atom arraysvon Neumann entropydensity-matrix renormalization group
0
0 comments X

The pith

A six-term linear formula built from boundary correlators approximates bipartite entanglement entropy to within 0.024 nats.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper seeks a simple readable expression for the bipartite von Neumann entropy that works directly from the bitstring distributions measured in Rydberg-atom experiments. A graph neural network localizes the needed information to the two-point correlators across the bipartition cut, after which an exhaustive search isolates a six-term linear combination of those scalars. The resulting closed form reaches 0.024 nats mean absolute error and, when applied without retraining, beats or matches the original network on five of six out-of-distribution test sets. An independent density-matrix renormalization-group calculation up to one hundred atoms shows that the functional form survives size extrapolation once coefficients are refit per size, with two slopes obeying clean inverse-size laws.

Core claim

The bipartite von Neumann entropy constrained by bitstring distributions admits a six-term linear closed form in bitstring-derivable scalars. The form is localized by a graph neural network to the two-point correlators on the bipartition boundary and reaches 0.024 nats mean absolute error. When refit per system size the expression remains accurate to 25-50 mnat up to 100 atoms, with two slopes obeying inverse-size laws that allow label-free use at 40-80 mnat error.

What carries the argument

Six-term linear closed form in boundary two-point correlators, selected via graph neural network localization followed by exhaustive search.

Load-bearing premise

Entanglement entropy is accurately captured by a linear combination of only the two-point correlators on the bipartition boundary, with coefficients that remain stable enough to refit per system size.

What would settle it

A direct comparison on DMRG or exact data for system sizes around 100 atoms showing that refitted coefficients yield mean absolute error above 50 mnat on held-out bitstrings would disprove the scaling claim.

Figures

Figures reproduced from arXiv: 2606.22713 by Anas Saleh.

Figure 1
Figure 1. Figure 1: Headline summary. (a) MAE on the in-distribution eval pool and the six OOD pools for the base GNN ([5], sky blue), the fine-tuned GNN (bluish-green, hatched), the six-feature closed form Eq. 1 (orange), and the classical mutual information I (vermilion), log scale. The fine-tuned column is the published checkpoint fine-tuned in [5] for the Ny = 7, 8 ladder and applied unchanged; it is hatched because its s… view at source ↗
Figure 2
Figure 2. Figure 2: Eq. 1 residuals on the evaluation set. Left: residual versus predicted Sˆ vN—the spread widens with the prediction. Middle: residual by system size Ny—the interquartile range and tails grow with Ny. Right: quantile–quantile plot, confirming heavy tails. The dispersion concentrates at higher entropy, consistent with corr(|r|, S)= 0.47. 8.2 Size scaling Eq. 1’s slopes are not scale-invariant. Refit within 2×… view at source ↗
Figure 3
Figure 3. Figure 3: Shot-noise stress test on a random 5,000-row sample of the evaluation pool: MAE versus the number of measurement shots S for Eq. 1, XGBoost, and the mutual-information baseline I. 9 Independent validation at scale: DMRG to N = 100 Every result so far rests on exact-diagonalisation labels, which end at N = 20. This section takes the closed form to that scale against an independent numerical method, at five … view at source ↗
Figure 4
Figure 4. Figure 4: Eq. 1 slopes refit per size on the balanced half cut, ED (N = 8–20, circles) and DMRG (N = 24– 100, squares), with 95% bootstrap intervals; solid/dashed curves are the best and runner-up parametric laws by AICc (weighted R2 in the legends). The anchor slope b1 grows and curves downward; the two correlator-kernel slopes b2, b3 decay as 1/N (an inverse square root fits comparably); b4, b5 show no resolv￾able… view at source ↗
Figure 5
Figure 5. Figure 5: Held-out per-hook Ridge-probe R2 for SvN, I, and the six features. Peaks at the readout (R2 ≥0.996 for SvN, ≥0.92 for every feature). many correlated directions, so ablating the low-dimensional probe span removes one copy and the rest fills in. (iii) The gradient SVD (§10.3) shows the dominant prediction direction tracks I, an Eq. 1 feature— so the prediction does move along these quantities. The correct s… view at source ↗
Figure 6
Figure 6. Figure 6: 2-D t-SNE of the GNN post-readout state (1,200 graphs), coloured by each of the six Eq. 1 features (cross-validated kNN R2 in each title). The embedding separates cleanly by the information￾theoretic anchor and the extremal correlators (R2 0.80–0.92) and more weakly by the two correlator-shape sums (≈0.60), showing the features are geometrically primary in the representation. 5 10 15 20 singular value inde… view at source ↗
Figure 7
Figure 7. Figure 7: Gradient-SVD of the readout, recomputed on [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
read the original abstract

The empirical bitstring distribution is the most accessible observable on Rydberg-atom arrays, but the bipartite von~Neumann entropy it constrains is far costlier to obtain. We present a six-term linear closed form for the entropy, built on bitstring-derivable physics scalars, and characterize its accuracy, portability, scaling behaviour, and calibration cost. The feature set is selected with guidance from a trained graph neural network: probing the network localizes its entropy prediction to the two-point correlators on the bipartition boundary, and an exhaustive ground-truth search restricted to those boundary correlators isolates the form. It reaches $0.024$~nats mean absolute error in distribution: $6.4$ times the network's error, but in a form a human can read and apply without retraining. Fit once and applied unchanged, it has lower error than the base network on five of six out-of-distribution pools and ties the sixth. An independent density-matrix renormalization-group study to one hundred atoms -- five times the reach of exact diagonalization -- settles the size-extrapolation question: coefficients frozen at small size fail at scale, but the failure is structured. Refit per size the form holds to $25$--$50$~mnat (cross-validated); two of its six slopes follow clean inverse-size laws, one a downward curving growth, and the others are trendless; the fitted laws deploy the form label-free at roughly $40$--$80$~mnat. The result fixes a label-budget rule: at large sizes, a few dozen labels recalibrate the closed form to match a fine-tuned in-distribution ensemble on the same features, while nonlinear ML models pull ahead only given large labelled datasets.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 3 minor

Summary. The manuscript claims that a graph neural network trained on bitstring data from Rydberg arrays localizes its entanglement-entropy predictions to two-point correlators on the bipartition boundary; an exhaustive search over those features then isolates a six-term linear closed form whose coefficients, when refit per system size, achieve 0.024 nats in-distribution MAE and 25–50 mnat error up to 100 atoms (five times the exact-diagonalization limit) according to DMRG validation, while remaining portable across out-of-distribution pools when the functional form is held fixed.

Significance. If the empirical result holds, the work supplies a compact, human-readable formula that converts readily measured bitstrings into entropy estimates without retraining a network, together with explicit scaling laws for two of the six coefficients. The GNN-guided feature discovery and the DMRG size-extrapolation protocol constitute reproducible, falsifiable contributions that could be adopted by experimental groups working with Rydberg arrays.

major comments (2)
  1. [§3] §3 (feature-selection protocol): the exhaustive search is restricted a priori to the boundary two-point correlators highlighted by the GNN; the manuscript does not report the performance of an unrestricted search over all bitstring-derived scalars, leaving open whether the six-term form is the globally simplest or merely the simplest within the GNN-localized subset.
  2. [§5.2] §5.2 (DMRG extrapolation): the claim that the form 'holds' at large size rests on per-size refitting of the six coefficients; the paper shows that frozen coefficients produce structured growth in error, but does not quantify how many additional labels are required to keep the refitted error below the in-distribution network baseline as N→∞.
minor comments (3)
  1. [Table 1] Table 1: the reported MAE values for the closed form versus the GNN should include the number of independent training runs or cross-validation folds used to obtain the quoted uncertainties.
  2. [Eq. (3)] Eq. (3): the six-term expression is written with numerical coefficients; the manuscript should also display the symbolic form with the six physics scalars named explicitly before numerical fitting.
  3. [Figure 4] Figure 4: axis labels on the coefficient-versus-size plots are too small for print; the inverse-size and curving-growth trends would be clearer with larger fonts and an inset showing the functional fits.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive report and the recommendation for minor revision. We address the two major comments below, indicating where revisions will be made to the manuscript.

read point-by-point responses

  1. Referee: [§3] §3 (feature-selection protocol): the exhaustive search is restricted a priori to the boundary two-point correlators highlighted by the GNN; the manuscript does not report the performance of an unrestricted search over all bitstring-derived scalars, leaving open whether the six-term form is the globally simplest or merely the simplest within the GNN-localized subset.

    Authors: The restriction of the exhaustive search to the GNN-localized boundary correlators is intentional and central to the method: the GNN serves to identify physically relevant features before the search, yielding an interpretable form with clear physical motivation. An unrestricted search over the full space of bitstring-derived scalars was not performed, as the combinatorial number of candidate features renders it computationally prohibitive and would defeat the purpose of GNN-guided discovery. We will revise §3 to explicitly state this rationale, note the limitation that the six-term form is the simplest within the GNN-highlighted subset, and add a brief discussion of why boundary two-point correlators are expected to dominate on physical grounds. revision: yes

  2. Referee: [§5.2] §5.2 (DMRG extrapolation): the claim that the form 'holds' at large size rests on per-size refitting of the six coefficients; the paper shows that frozen coefficients produce structured growth in error, but does not quantify how many additional labels are required to keep the refitted error below the in-distribution network baseline as N→∞.

    Authors: The manuscript already reports that a few dozen labels suffice for recalibration up to N=100 to reach errors comparable to the in-distribution network baseline. We agree, however, that an explicit functional dependence of the required label count on N as N→∞ is not provided. We will revise §5.2 to restate the observed label budget more clearly in comparison to the network baseline and to note that determining the asymptotic scaling would require additional DMRG data at still larger sizes, which lies outside the present scope. revision: partial

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper's derivation is an explicit empirical procedure: GNN localization followed by exhaustive search over boundary two-point correlators, then linear regression to obtain a six-term form. This is presented as a fitted, interpretable approximation rather than a first-principles derivation. The manuscript directly validates portability via OOD tests (fixed coefficients outperform or tie the network on five of six pools) and size extrapolation via independent DMRG up to 100 atoms, distinguishing frozen vs. refit coefficients and reporting structured scaling. No load-bearing step reduces by construction to its inputs, no self-citation chain is invoked for uniqueness, and no ansatz is smuggled; the result is self-contained against the reported benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on the empirical observation that boundary two-point correlators suffice for a linear approximation; this is not derived from first principles but isolated by search. The six coefficients are free parameters fitted to data. No new physical entities are postulated.

free parameters (1)
  • six linear coefficients
    Determined by regression on small-system data; must be refit per system size for the form to hold at large N.
axioms (2)
  • domain assumption Entanglement entropy is linearly related to the selected boundary two-point correlators
    Invoked when the exhaustive search is restricted to those features and a linear model is fitted.
  • domain assumption GNN attention localizes the relevant physics to the bipartition boundary
    Used to justify restricting the search space to boundary correlators.

pith-pipeline@v0.9.1-grok · 5841 in / 1756 out tokens · 24404 ms · 2026-06-26T09:52:20.763517+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

23 extracted references · 1 canonical work pages

  1. [1]

    A. L. Shaw et al. Benchmarking highly entangled states on a 60-atom analogue quantum simulator. Nature, 628:71, 2024

    2024

  2. [2]

    Bernien et al

    H. Bernien et al. Probing many-body dynamics on a 51-atom quantum simulator.Nature, 551:579, 2017

    2017

  3. [3]

    Ebadi et al

    S. Ebadi et al. Quantum phases of matter on a 256-atom programmable quantum simulator.Nature, 595:227, 2021

    2021

  4. [4]

    Scholl et al

    P. Scholl et al. Quantum simulation of 2D antiferromagnets with hundreds of Rydberg atoms.Nature, 595:233, 2021

    2021

  5. [5]

    A. Saleh. Predicting the von Neumann entanglement entropy using a graph neural network.Mach. Learn.: Sci. Technol., 6:035034, 2025

    2025

  6. [6]

    A. Saleh. Predicting the von Neumann entanglement entropy using a graph neural network. Master’s thesis, University of Iowa, 2025. DOI:10.25820/etd.008062. Extended version of [5]; additionally trains and evaluates the GNN on transverse-field Ising configurations

    work page doi:10.25820/etd.008062 2025

  7. [7]

    Kaufman et al

    A. Kaufman et al. Improved entanglement entropy estimates from filtered bitstring probabilities. Phys. Rev. A, 112:032430, 2025

    2025

  8. [8]

    Islam et al

    R. Islam et al. Measuring entanglement entropy in a quantum many-body system.Nature, 528:77, 2015

    2015

  9. [9]

    Brydges et al

    T. Brydges et al. Probing Rényi entanglement entropy via randomized measurements.Science, 364:260, 2019. 21

    2019

  10. [10]

    Huang, R

    H.-Y. Huang, R. Kueng, and J. Preskill. Predicting many properties of a quantum system from very few measurements.Nat. Phys., 16:1050, 2020

    2020

  11. [11]

    A. S. Holevo. Bounds for the quantity of information transmitted by a quantum communication channel.Probl. Inf. Transm., 9:177, 1973

    1973

  12. [12]

    M. M. Wilde.Quantum Information Theory. Cambridge University Press, 2nd edition, 2017

    2017

  13. [13]

    M. M. Wolf, F. Verstraete, M. B. Hastings, and J. I. Cirac. Area laws in quantum systems: mutual information and correlations.Phys. Rev. Lett., 100:070502, 2008

    2008

  14. [14]

    Carleo and M

    G. Carleo and M. Troyer. Solving the quantum many-body problem with artificial neural networks. Science, 355:602, 2017

    2017

  15. [15]

    Carrasquilla and R

    J. Carrasquilla and R. G. Melko. Machine learning phases of matter.Nat. Phys., 13:431, 2017

    2017

  16. [16]

    Carleo et al

    G. Carleo et al. Machine learning and the physical sciences.Rev. Mod. Phys., 91:045002, 2019

    2019

  17. [17]

    Udrescu and M

    S.-M. Udrescu and M. Tegmark. AI Feynman: a physics-inspired method for symbolic regression. Sci. Adv., 6:eaay2631, 2020

    2020

  18. [18]

    Cranmer et al

    M. Cranmer et al. Discovering symbolic models from deep learning with inductive biases. InAdvances in Neural Information Processing Systems (NeurIPS), 2020

    2020

  19. [19]

    Geiger, H

    A. Geiger, H. Lu, T. Icard, and C. Potts. Causal abstractions of neural networks. InAdvances in Neural Information Processing Systems (NeurIPS), 2021

    2021

  20. [20]

    Jain and B

    S. Jain and B. C. Wallace. Attention is not explanation. InProceedings of NAACL-HLT, 2019

    2019

  21. [21]

    Wiegreffe and Y

    S. Wiegreffe and Y. Pinter. Attention is not not explanation. InProceedings of EMNLP-IJCNLP, 2019

    2019

  22. [22]

    M. Cranmer. Interpretable machine learning for science with PySR and SymbolicRegression.jl.arXiv preprint arXiv:2305.01582, 2023

    Pith/arXiv arXiv 2023

  23. [23]

    Hauschild and F

    J. Hauschild and F. Pollmann. Efficient numerical simulations with tensor networks: Tensor Network Python (TeNPy).SciPost Phys. Lect. Notes, page 5, 2018. 22

    2018