pith. sign in

arxiv: 2506.04170 · v1 · pith:S4CLDA3Inew · submitted 2025-06-04 · 🪐 quant-ph · cond-mat.stat-mech· cs.LG· hep-lat· hep-th

Estimation of the reduced density matrix and entanglement entropies using autoregressive networks

Piotr Bia{\l}as , Piotr Korcyl , Tomasz Stebel , Dawid Zapolski This is my paper

Pith reviewed 2026-05-22 00:17 UTC · model grok-4.3

classification 🪐 quant-ph cond-mat.stat-mechcs.LGhep-lathep-th
keywords autoregressive neural networksreduced density matrixentanglement entropyquantum spin chainsTrotter-Suzuki mappingIsing modelMonte Carlo simulations
0
0 comments X

The pith

Autoregressive networks estimate all reduced density matrix elements for quantum spin chains from a single training run.

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

The paper demonstrates that autoregressive neural networks can learn the conditional probabilities of classical spin configurations obtained from the Trotter-Suzuki mapping of quantum spin chains. This learned distribution lets the networks directly supply the matrix elements of a reduced density matrix for a small subsystem. For the Ising chain the authors extract the continuum-limit von Neumann and Rényi entanglement entropies of intervals containing up to five spins. A reader would care because the same trained network supplies every needed element, removing the need to retrain for each matrix entry.

Core claim

A hierarchy of autoregressive neural networks that estimate conditional probabilities of consecutive spins in the classical two-dimensional configuration is shown to evaluate every element of the reduced density matrix for a chosen subsystem after only one training, for fixed time discretization and lattice volume; this is used to compute the continuum limit of the ground-state von Neumann and Rényi bipartite entanglement entropies of an interval of up to five spins in the Ising chain.

What carries the argument

Hierarchy of autoregressive neural networks estimating conditional probabilities of consecutive spins to evaluate elements of the reduced density matrix directly.

If this is right

  • All matrix elements needed for entanglement entropy can be obtained without retraining the networks for each element.
  • The method applies directly to other spin-chain models, including those with defects.
  • Entanglement entropies of thermal states at nonzero temperature can be estimated with the same architecture.
  • Continuum limits of the entropies become accessible for small subsystems once the networks are trained at fixed discretization.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The single-training property could make entanglement calculations feasible for larger subsystems or higher-dimensional lattices where repeated trainings would be costly.
  • One could test whether a network trained at one volume still produces accurate reduced-density-matrix elements when the subsystem size is increased without retraining.
  • The same conditional-probability representation might be used to extract other subsystem observables, such as local correlation functions, beyond entanglement entropy.

Load-bearing premise

The autoregressive networks accurately reproduce the conditional probabilities of the classical spin configurations that correspond to the quantum ground state under the Trotter-Suzuki mapping.

What would settle it

Exact diagonalization of the reduced density matrix for the Ising chain at the same discretization and volume yields matrix elements that differ substantially from those estimated by the trained networks.

Figures

Figures reproduced from arXiv: 2506.04170 by Dawid Zapolski, Piotr Bia{\l}as, Piotr Korcyl, Tomasz Stebel.

Figure 1
Figure 1. Figure 1: The structure of the spin configurations generated [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Natural logarithm of the density matrix elements in [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Eigenvalues of reduced density matrix {λi}i=0,...,2l−1 obtained for L = 32, l = 5, k = 8, ∆τ = 0.2. They were plotted in terms of ordinal numbers (from biggest to smallest). B. Entanglement Entropies In this section, we present results for von Neumann and R´enyi entropies. They are obtained from eigenvalues of the reduced density matrix {λi}i=0,...,2 l−1 as: S(A) = 2 Xl−1 i=0 λi log λi , (25) Sn(A) = 1 1 −… view at source ↗
Figure 5
Figure 5. Figure 5: Von Neumann entropy extrapolated to k = ∞ as a function of (∆τ ) 2 for l = 2. In [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: von Neumann (n = 1) and R´enyi entropies (n ≥ 2) as functions of subsystem size l for spin chain with L = 32 spins at the zero temperature (points). The dashed line corresponds to the conformal field theory prediction [21], see Eq. (28). Solid lines represent fits of equation (29) to our results for 3 ≤ l ≤ 5. for entanglement entropy:    y∆τ1 (k) = S + a1(∆τ1) 2 + b∆τ1 exp(−c∆τ1 k) y∆τ2 (k) = S +… view at source ↗
read the original abstract

We present an application of autoregressive neural networks to Monte Carlo simulations of quantum spin chains using the correspondence with classical two-dimensional spin systems. We use a hierarchy of neural networks capable of estimating conditional probabilities of consecutive spins to evaluate elements of reduced density matrices directly. Using the Ising chain as an example, we calculate the continuum limit of the ground state's von Neumann and R\'enyi bipartite entanglement entropies of an interval built of up to 5 spins. We demonstrate that our architecture is able to estimate all the needed matrix elements with just a single training for a fixed time discretization and lattice volume. Our method can be applied to other types of spin chains, possibly with defects, as well as to estimating entanglement entropies of thermal states at non-zero temperature.

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 / 2 minor

Summary. The manuscript applies autoregressive neural networks to Monte Carlo sampling of quantum spin chains via the Trotter-Suzuki mapping to classical 2D Ising systems. A hierarchy of networks is used to compute conditional probabilities that directly yield elements of the reduced density matrix; the method is demonstrated on the transverse-field Ising chain by extracting the continuum-limit von Neumann and Rényi entanglement entropies for subsystems of up to five sites, with the key assertion that a single training suffices for all required matrix elements at fixed discretization and volume.

Significance. If the numerical accuracy of the marginals and conditionals is established, the single-training feature would constitute a practical advance for entanglement calculations in quantum Monte Carlo, reducing computational cost relative to repeated trainings or full wave-function storage and extending naturally to thermal states or chains with defects.

major comments (2)
  1. [§4 and abstract] The central claim that a single training suffices for all RDM elements (abstract and §4) rests on the assumption that the learned conditional probabilities p(s_i | s_<i) remain accurate when marginalizing over the traced-out spins. No quantitative error analysis, bias estimates, or validation against exact marginals for varied subsystem boundary conditions is provided; any systematic deviation here directly affects the RDM eigenvalues and the extrapolated entropies.
  2. [§5] In the Ising-chain demonstration (§5), the reported continuum-limit entropies lack error bars, convergence tests with respect to network width/depth or Monte Carlo statistics, and direct comparisons to exact diagonalization results for small intervals; without these benchmarks the extraction of continuum values cannot be assessed for reliability.
minor comments (2)
  1. [§3] The description of how the hierarchy of autoregressive networks assembles the full set of RDM matrix elements would benefit from an explicit algorithmic outline or pseudocode.
  2. [§2] A few typographical inconsistencies appear in the notation for the conditional probabilities and the definition of the Rényi index.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments on our manuscript. We address each major comment below and describe the changes we will make in the revised version.

read point-by-point responses

  1. Referee: [§4 and abstract] The central claim that a single training suffices for all RDM elements (abstract and §4) rests on the assumption that the learned conditional probabilities p(s_i | s_<i) remain accurate when marginalizing over the traced-out spins. No quantitative error analysis, bias estimates, or validation against exact marginals for varied subsystem boundary conditions is provided; any systematic deviation here directly affects the RDM eigenvalues and the extrapolated entropies.

    Authors: We agree that a quantitative error analysis strengthens the central claim. The autoregressive decomposition factorizes the joint distribution exactly, so marginalization over traced spins is performed by summing the learned conditionals; any inaccuracy in the conditionals would propagate to the RDM. In the revision we will add an appendix with direct comparisons of the learned conditionals and resulting marginals against exact results on small lattices, including bias estimates and tests for the range of boundary conditions that arise when varying subsystem size. These additions will quantify the accuracy of the single-training procedure. revision: yes

  2. Referee: [§5] In the Ising-chain demonstration (§5), the reported continuum-limit entropies lack error bars, convergence tests with respect to network width/depth or Monte Carlo statistics, and direct comparisons to exact diagonalization results for small intervals; without these benchmarks the extraction of continuum values cannot be assessed for reliability.

    Authors: We accept that error bars, convergence tests, and benchmarks are required to assess reliability. In the revised manuscript we will report statistical error bars on the extracted entropies, include convergence plots with respect to network width, depth, and Monte Carlo sample size, and add direct comparisons to exact diagonalization for intervals of one and two spins. For the larger intervals we will discuss consistency with the small-interval benchmarks and with known scaling behavior of the Ising chain. revision: yes

Circularity Check

0 steps flagged

No circularity; standard Trotter-Suzuki mapping plus autoregressive sampling yields RDM elements directly

full rationale

The derivation trains autoregressive networks on full classical 2D configurations obtained from the Trotter-Suzuki mapping of the quantum Ising chain. Conditional probabilities p(s_i | s_<i) are learned once and then used to compute the marginals required for any reduced-density-matrix element by summing over traced spins. This step is a direct probabilistic marginalization, not a redefinition or a fitted parameter renamed as a prediction. No equation reduces the reported von Neumann or Rényi entropies to the training loss by construction, and the single-training claim is an empirical demonstration of the autoregressive factorization rather than a self-referential identity. The approach therefore remains self-contained against external Monte Carlo benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim depends on the standard Trotter-Suzuki mapping of quantum spin chains to classical two-dimensional systems and on the assumption that autoregressive networks can faithfully represent the resulting probability distribution after training.

free parameters (1)
  • Autoregressive network parameters
    Weights and biases are fitted during training on Monte Carlo samples to match conditional spin probabilities.
axioms (1)
  • domain assumption Quantum spin chains in imaginary time can be mapped to classical two-dimensional Ising-like systems whose configurations are sampled by Monte Carlo.
    Invoked to justify the use of classical Monte Carlo data as training input for the neural networks.

pith-pipeline@v0.9.0 · 5681 in / 1264 out tokens · 44060 ms · 2026-05-22T00:17:32.584672+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Sampling two-dimensional spin systems with transformers

    cond-mat.dis-nn 2026-04 unverdicted novelty 7.0

    Transformer networks sample up to 180x180 2D Ising systems and 64x64 Edwards-Anderson systems by generating spin groups with probability approximations, yielding ~20x higher effective sample size than prior neural sam...

  2. Variational Autoregressive Networks with probability priors

    cs.LG 2026-05 unverdicted novelty 5.0

    Incorporating probability priors into variational autoregressive networks reduces training burden and enables larger system sizes for sampling in the Ising and Edwards-Anderson models.

Reference graph

Works this paper leans on

39 extracted references · 39 canonical work pages · cited by 2 Pith papers · 1 internal anchor

  1. [1]

    Given two product states of subsystem A, |µA⟩ and |νA⟩, we fix the classical spins in the first and last row (denoted as numbers 1 to 6 in Figure 1)

    work page

  2. [2]

    The probability of spins in part B depends on spins A

    We draw conditionally the spins from part B (num- bers 7-11 in Figure 1) using the first autoregres- sive network from HAN. The probability of spins in part B depends on spins A. Spins from part B are copied to the first and last row of the configu- ration

    work page

  3. [3]

    Once the upper and lower rows are fixed, the inte- rior of the configuration is filled. We make use of the Hammersley-Clifford theorem [19, 20], which assures that once a closed loop of spins is fixed, the probability of loop-interior spins does not depend on the exterior. This is the key observation for the HAN algorithm [18], which is applied here. We u...

    work page

  4. [4]

    At the next step, ×-spins are drawn using yet an- other network and they depend only on their sur- rounding spins (16 in total for L = 8). Since the functional dependencies on the outside spins are the same for each remaining square of the configura- tion, we can use the same network, run in parallel, to fix ” ×”-spins in each square. The last network fro...

    work page

  5. [5]

    Zero temperature and continuum extrapolations In Figure 4 we show von Neumann entanglement en- tropy as functions of k for several values of ∆ τ (points plotted with different colors). The corrections coming from finite k decay exponentially with k [11] hence the extrapolation k → ∞ is performed by fitting a function of the form y∆τ(k) = a∆τ + b∆τ exp(−c∆...

    work page

  6. [6]

    The dashed line corresponds to the conformal field theory prediction [21], see Eq

    as functions of subsystem size l for spin chain with L = 32 spins at the zero temperature (points). The dashed line corresponds to the conformal field theory prediction [21], see Eq. (28). Solid lines represent fits of equation (29) to our results for 3 ≤ l ≤ 5. for entanglement entropy:    y∆τ1(k) = S + a1(∆τ1)2 + b∆τ1 exp(−c∆τ1 k) y∆τ2(k) = S +...

    work page

  7. [7]

    We keep the size of the whole system as L = 32

    Ground state entanglement We present the von Neumann and R´ enyi entanglement entropies as functions of l, the size of the subsystem A. We keep the size of the whole system as L = 32. One expects that lim n→1 Sn(A) = S(A); therefore, we denote a value of the von Neumann entropy as n = 1. In Figure 6 we show S(l) and Sn(l) for n = 2, 3, 4, 9 (points) - not...

    work page 2024

  8. [8]

    Quan- tum source of entropy for black holes,

    L. Bombelli, R. K. Koul, J. Lee, and R. D. Sorkin, “Quan- tum source of entropy for black holes,” Phys. Rev. D , vol. 34, pp. 373–383, Jul 1986. 9

    work page 1986

  9. [9]

    Entropy and area,

    M. Srednicki, “Entropy and area,” Phys. Rev. Lett. , vol. 71, pp. 666–669, Aug 1993

    work page 1993

  10. [10]

    Entanglement entropy and quantum field theory,

    P. Calabrese and J. Cardy, “Entanglement entropy and quantum field theory,” Journal of Statistical Mechanics: Theory and Experiment, vol. 2004, p. 06002, June 2004

    work page 2004

  11. [11]

    Colloquium: Area laws for the entanglement entropy,

    J. Eisert, M. Cramer, and M. B. Plenio, “Colloquium: Area laws for the entanglement entropy,” Rev. Mod. Phys., vol. 82, pp. 277–306, Feb 2010

    work page 2010

  12. [12]

    The density-matrix renormalization group in the age of matrix product states,

    U. Schollw¨ ock, “The density-matrix renormalization group in the age of matrix product states,” Annals of Physics, vol. 326, pp. 96–192, Jan. 2011

    work page 2011

  13. [13]

    Measuring renyi entanglement entropy in quan- tum monte carlo simulations,

    M. B. Hastings, I. Gonz´ alez, A. B. Kallin, and R. G. Melko, “Measuring renyi entanglement entropy in quan- tum monte carlo simulations,” Phys. Rev. Lett., vol. 104, p. 157201, Apr 2010

    work page 2010

  14. [14]

    Quantum Monte Carlo calculation of entanglement Renyi entropies for generic quantum systems,

    S. Humeniuk and T. Roscilde, “Quantum Monte Carlo calculation of entanglement Renyi entropies for generic quantum systems,” Phys. Rev. B , vol. 86, p. 235116, 2012

    work page 2012

  15. [15]

    Entanglement entropy from nonequilib- rium work,

    J. D’Emidio, “Entanglement entropy from nonequilib- rium work,” Phys. Rev. Lett. , vol. 124, p. 110602, Mar 2020

    work page 2020

  16. [16]

    Measuring r´ enyi entanglement entropy with high efficiency and precision in quantum monte carlo simulations,

    J. Zhao, B.-B. Chen, Y.-C. Wang, Z. Yan, M. Cheng, and Z. Y. Meng, “Measuring r´ enyi entanglement entropy with high efficiency and precision in quantum monte carlo simulations,” npj Quantum Materials , vol. 7, June 2022

    work page 2022

  17. [17]

    Entanglement entropy from non-equilibrium Monte Carlo simulations,

    A. Bulgarelli and M. Panero, “Entanglement entropy from non-equilibrium Monte Carlo simulations,” JHEP, vol. 06, p. 030, 2023

    work page 2023

  18. [18]

    R´ enyi entanglement entropy of a spin chain with generative neu- ral networks,

    P. Bia las, P. Korcyl, T. Stebel, and D. Zapolski, “R´ enyi entanglement entropy of a spin chain with generative neu- ral networks,” Phys. Rev. E , vol. 110, no. 4, p. 044116, 2024

    work page 2024

  19. [19]

    Flow-Based Sampling for Entanglement Entropy and the Machine Learning of Defects,

    A. Bulgarelli, E. Cellini, K. Jansen, S. K¨ uhn, A. Nada, S. Nakajima, K. A. Nicoli, and M. Panero, “Flow-Based Sampling for Entanglement Entropy and the Machine Learning of Defects,” Phys. Rev. Lett. , vol. 134, no. 15, p. 151601, 2025

    work page 2025

  20. [20]

    Asymptotically unbiased es- timation of physical observables with neural samplers,

    K. A. Nicoli, S. Nakajima, N. Strodthoff, W. Samek, K.- R. M¨ uller, and P. Kessel, “Asymptotically unbiased es- timation of physical observables with neural samplers,” Phys. Rev. E , vol. 101, p. 023304, Feb. 2020

    work page 2020

  21. [21]

    Es- timation of thermodynamic observables in lattice field theories with deep generative models,

    K. A. Nicoli, C. J. Anders, L. Funcke, T. Hartung, K. Jansen, P. Kessel, S. Nakajima, and P. Stornati, “Es- timation of thermodynamic observables in lattice field theories with deep generative models,” Phys. Rev. Lett., vol. 126, p. 032001, Jan 2021

    work page 2021

  22. [22]

    Order and disorder in gauge systems and magnets,

    E. Fradkin and L. Susskind, “Order and disorder in gauge systems and magnets,” Phys. Rev. D , vol. 17, pp. 2637– 2658, May 1978

    work page 1978

  23. [23]

    MADE: Masked Autoencoder for Distribution Estimation

    M. Germain, K. Gregor, I. Murray, and H. Larochelle, “MADE: Masked Autoencoder for Distribution Estima- tion,” arXiv e-prints, p. arXiv:1502.03509, Feb. 2015

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  24. [24]

    Solving Statistical Mechanics Using Variational Autoregressive Networks,

    D. Wu, L. Wang, and P. Zhang, “Solving Statistical Mechanics Using Variational Autoregressive Networks,” Phys. Rev. Lett., vol. 122, p. 080602, Mar. 2019

    work page 2019

  25. [25]

    Hierarchical autore- gressive neural networks for statistical systems,

    P. Bia las, P. Korcyl, and T. Stebel, “Hierarchical autore- gressive neural networks for statistical systems,” Com- put. Phys. Commun. , vol. 281, p. 108502, 2022

    work page 2022

  26. [26]

    Markov fields on finite graphs and lattices,

    P. C. J. M. Hammersley, “Markov fields on finite graphs and lattices,” 1971

    work page 1971

  27. [27]

    Markov random fields in statistics,

    P. Clifford, “Markov random fields in statistics,” in Dis- order in Physical Systems. A Volume in Honour of John M. Hammersley, Clarendon Press, 1990

    work page 1990

  28. [28]

    Quantum Spin Chain, Toeplitz Determinants and the Fisher—Hartwig Conjec- ture,

    B. Q. Jin and V. E. Korepin, “Quantum Spin Chain, Toeplitz Determinants and the Fisher—Hartwig Conjec- ture,” Journal of Statistical Physics , vol. 116, pp. 79–95, Aug. 2004

    work page 2004

  29. [29]

    En- tanglement in quantum critical phenomena,

    G. Vidal, J. I. Latorre, E. Rico, and A. Kitaev, “En- tanglement in quantum critical phenomena,” Phys. Rev. Lett., vol. 90, p. 227902, Jun 2003

    work page 2003

  30. [30]

    Entanglement entropy and conformal field theory,

    P. Calabrese and J. Cardy, “Entanglement entropy and conformal field theory,”Journal of Physics A: Mathemat- ical and Theoretical, vol. 42, p. 504005, dec 2009

    work page 2009

  31. [31]

    Form factors of branch-point twist fields in quantum integrable models and entanglement entropy,

    J. L. Cardy, O. A. Castro-Alvaredo, and B. Doyon, “Form factors of branch-point twist fields in quantum integrable models and entanglement entropy,” J. Statist. Phys. , vol. 130, pp. 129–168, 2008

    work page 2008

  32. [32]

    Finite-size scaling of half- chain entanglement entropy in the one-dimensional trans- verse field ising model and the xx model,

    H. Tian, T. He, and X. Wu, “Finite-size scaling of half- chain entanglement entropy in the one-dimensional trans- verse field ising model and the xx model,” Phys. Rev. B, vol. 111, p. 104437, Mar 2025

    work page 2025

  33. [33]

    En- tanglement, combinatorics and finite-size effects in spin chains,

    B. Nienhuis, M. Campostrini, and P. Calabrese, “En- tanglement, combinatorics and finite-size effects in spin chains,” Journal of Statistical Mechanics: Theory and Experiment, vol. 2009, p. 02063, Feb. 2009

    work page 2009

  34. [34]

    Unusual Corrections to Scal- ing in Entanglement Entropy,

    J. Cardy and P. Calabrese, “Unusual Corrections to Scal- ing in Entanglement Entropy,” J. Stat. Mech., vol. 1004, p. P04023, 2010

    work page 2010

  35. [35]

    Spectral properties of the critical (1+1)-dimensional Abelian-Higgs model,

    T. Chanda, M. Dalmonte, M. Lewenstein, J. Zakrzewski, and L. Tagliacozzo, “Spectral properties of the critical (1+1)-dimensional Abelian-Higgs model,” Phys. Rev. B , vol. 109, no. 4, p. 045103, 2024

    work page 2024

  36. [36]

    Entan- glement entropy in quantum impurity systems and sys- tems with boundaries,

    I. Affleck, N. Laflorencie, and E. S. Sørensen, “Entan- glement entropy in quantum impurity systems and sys- tems with boundaries,” Journal of Physics A Mathemat- ical General, vol. 42, p. 504009, Dec. 2009

    work page 2009

  37. [37]

    Entanglement Entropy in the Ising Model with Topological Defects,

    A. Roy and H. Saleur, “Entanglement Entropy in the Ising Model with Topological Defects,” Phys. Rev. Lett., vol. 128, p. 090603, Mar. 2022

    work page 2022

  38. [38]

    Hierarchical autoregressive neural net- works in three-dimensional statistical system,

    P. Bia las, V. Chahar, P. Korcyl, T. Stebel, M. Winiarski, and D. Zapolski, “Hierarchical autoregressive neural net- works in three-dimensional statistical system,” 2025

    work page 2025

  39. [39]

    Metropolized independent sampling with comparisons to rejection sampling and importance sam- pling,

    J. S. Liu, “Metropolized independent sampling with comparisons to rejection sampling and importance sam- pling,” Statistics and Computing , vol. 6, no. 2, pp. 113– 119, 1996

    work page 1996