arxiv: 2604.05970 · v1 · submitted 2026-04-07 · ✦ hep-th

Recognition: no theorem link

Holographic entanglement entropy, Wilson loops, and neural networks

Veselin G. Filev

Authors on Pith no claims yet

Pith reviewed 2026-05-10 19:29 UTC · model grok-4.3

classification ✦ hep-th

keywords neural networksholographic inverse problementanglement entropyWilson loopsbulk reconstructionRyu-Takayanagi formulaAdS/CFT

0 comments

The pith

Neural networks reconstruct bulk metric functions from boundary entanglement entropy and Wilson loop data.

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

The paper shows that artificial neural networks can solve the holographic inverse problem by treating the Ryu-Takayanagi area functional as a differentiable loss that drives reconstruction of the bulk metric. Validation on the AdS-Schwarzschild background recovers the blackening factor to 1.7 percent accuracy. In finite-density models such as Gubser-Rocha, entanglement entropy alone fixes only the spatial component, but adding Wilson loop data, which couples to the timelike metric, removes the degeneracy. A three-network variational scheme that minimizes the combined area and Nambu-Goto losses then recovers both functions to sub-0.2 percent accuracy without closed-form derivative relations. This matters because it supplies a general numerical route to bulk geometry when analytic inversion is unavailable.

Core claim

A variational neural-network method that minimizes the combined Ryu-Takayanagi area and Nambu-Goto actions reconstructs both the spatial and timelike metric functions in finite-density backgrounds to sub-0.2 percent accuracy without requiring closed-form derivative relations, thereby providing a flexible framework for integrating multiple holographic observables.

What carries the argument

The three-network variational optimizer that back-propagates through the holographic area functionals for entanglement entropy and Wilson loops to learn the metric components.

If this is right

Single-observable degeneracies in the inverse problem can be lifted by adding a second observable that probes a different metric component.
The method operates without analytic expressions for the derivatives of the loss functions.
High accuracy on known backgrounds supports use on geometries that lack closed-form solutions.
The framework extends naturally to other holographic observables that can be expressed as area functionals.

Where Pith is reading between the lines

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

The same variational approach could incorporate additional probes such as conductivity or higher-point correlators to constrain more metric components.
Numerical reconstruction of this kind may allow quantitative comparison between holographic models and lattice or experimental data in finite-density systems.
Once trained on families of backgrounds, the networks might serve as fast surrogates for exploring parameter space in holographic condensed-matter models.

Load-bearing premise

The network optimization must reach the global minimum that matches the true bulk geometry rather than a training artifact.

What would settle it

Apply the trained network to the known AdS-Schwarzschild black hole and check whether the output blackening factor agrees with the analytic expression to within 2 percent.

read the original abstract

We apply artificial neural networks to the holographic inverse problem, reconstructing bulk geometry from boundary entanglement entropy by using the Ryu--Takayanagi area functional as a differentiable loss. Validated on the AdS-Schwarzschild background, this approach recovers the blackening factor to 1.7% accuracy. For finite-density backgrounds like the Gubser--Rocha model, we demonstrate that strip entanglement entropy determines only the spatial metric. We resolve this exact one-function degeneracy by incorporating holographic Wilson loop data, which couples to the timelike metric. We present a semi-analytical inversion combining Bilson's and Hashimoto's formulas, alongside a general three-network variational method minimizing the combined area and Nambu--Goto actions. The neural network achieves sub-0.2% accuracy for both metric functions without closed-form derivative relations, establishing a flexible framework for integrating multiple holographic observables.

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.

Desk Editor's Note private letter to a colleague

The paper trains neural nets to recover bulk metrics from entanglement entropy plus Wilson loops by minimizing the Ryu-Takayanagi and Nambu-Goto functionals directly, with decent accuracy on test backgrounds but no proof the optimizer reaches the true solution.

read the letter

The main point is that a neural network can reconstruct the metric functions in holographic models by treating the Ryu-Takayanagi area as a differentiable loss, and that Wilson loop data lifts the exact degeneracy between spatial and timelike components that appears in finite-density cases. On AdS-Schwarzschild it recovers the blackening factor to roughly 1.7 percent, and on the Gubser-Rocha model the combined loss gives sub-0.2 percent accuracy for both metric pieces without needing closed-form derivatives. They also give a semi-analytic route using Bilson and Hashimoto formulas as a check. This combination of differentiable holographic losses inside a network plus the Wilson-loop fix is not in the earlier literature they cite, so the technical step is new. The three-network variational setup looks workable for integrating multiple observables when analytic inversion is unavailable. The approach is useful for people who already work with holographic entanglement entropy and want a numerical tool to extract bulk geometry from boundary data. The soft spots are on the numerical side. The abstract reports concrete accuracies but gives no training protocols, convergence diagnostics, or checks against multiple random seeds. Because the loss depends on minimal surfaces and string embeddings, it is non-convex, so nothing rules out other metric profiles that match the same boundary observables to within tolerance. The tests on known solutions show the method can succeed, yet they do not establish that the optimizer consistently finds the global minimum or that the recovered functions are unique. A reader who needs a black-box inverter for strongly coupled models will still find the framework worth trying, but anyone planning to rely on the output for new physics will want stronger evidence that the numerics are stable. The paper deserves a serious referee because the method is coherent and the reported numbers are concrete enough to evaluate. I would send it to review rather than desk reject, with the expectation that referees will press on the optimization details and possible non-uniqueness.

Referee Report

3 major / 1 minor

Summary. The manuscript proposes using neural networks to solve the holographic inverse problem of reconstructing bulk metric functions from boundary data. It employs the Ryu-Takayanagi area functional as a differentiable loss for entanglement entropy and the Nambu-Goto action for Wilson loops, validates recovery of the blackening factor to 1.7% accuracy on AdS-Schwarzschild, resolves a one-function degeneracy in the Gubser-Rocha model by combining observables, and reports sub-0.2% accuracy for both metric components via a three-network variational method without requiring closed-form derivative relations.

Significance. If the numerical reconstructions prove robust, the work supplies a flexible, observable-agnostic framework for holographic metric inversion that can incorporate multiple boundary quantities. The explicit demonstration of degeneracy resolution between entanglement entropy and Wilson loops, together with the semi-analytic Bilson-Hashimoto comparison, is a concrete strength that could extend to backgrounds lacking analytic inversions.

major comments (3)

[Abstract] Abstract: the reported accuracies (1.7% on AdS-Schwarzschild and sub-0.2% on Gubser-Rocha) are presented without any description of network architecture, training hyperparameters, optimization convergence diagnostics, number of independent runs, or statistical error bars. These details are load-bearing for the central claim that the method recovers the true metric functions.
[Three-network variational method] Three-network variational method: the combined area plus Nambu-Goto loss is a non-convex functional of the metric. The manuscript supplies no uniqueness argument, no multiple-random-initialization tests, and no convergence certificate showing that the optimizer reaches the physically correct global minimum rather than a different metric profile that reproduces the same boundary data to within numerical tolerance. This directly affects the claim that both metric functions are recovered.
[Gubser-Rocha model] Gubser-Rocha discussion: the statement that strip entanglement entropy constrains only the spatial metric component is asserted but not accompanied by an explicit demonstration (e.g., a plot or derivative showing vanishing sensitivity to the timelike component). Without this, the necessity of adding Wilson-loop data to lift the degeneracy remains incompletely justified.

minor comments (1)

[Notation] The notation distinguishing the two metric functions (spatial vs. timelike) could be made more explicit in the equations defining the loss, to avoid potential reader confusion when comparing the semi-analytic and neural-network results.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the careful reading of our manuscript and the constructive comments. We address each major comment in turn below and have revised the manuscript accordingly to strengthen the presentation and address the concerns raised.

read point-by-point responses

Referee: [Abstract] Abstract: the reported accuracies (1.7% on AdS-Schwarzschild and sub-0.2% on Gubser-Rocha) are presented without any description of network architecture, training hyperparameters, optimization convergence diagnostics, number of independent runs, or statistical error bars. These details are load-bearing for the central claim that the method recovers the true metric functions.

Authors: We agree that the abstract reports numerical accuracies without sufficient methodological context. In the revised manuscript we will update the abstract to include a concise description of the neural-network approach and explicitly direct readers to the main text for full details on architecture, training hyperparameters, convergence diagnostics, number of independent runs, and statistical error bars. revision: yes
Referee: [Three-network variational method] Three-network variational method: the combined area plus Nambu-Goto loss is a non-convex functional of the metric. The manuscript supplies no uniqueness argument, no multiple-random-initialization tests, and no convergence certificate showing that the optimizer reaches the physically correct global minimum rather than a different metric profile that reproduces the same boundary data to within numerical tolerance. This directly affects the claim that both metric functions are recovered.

Authors: The referee correctly notes that the combined loss is non-convex and that the manuscript lacks explicit multiple-initialization tests or convergence diagnostics. While the current results show consistent recovery on known backgrounds, we will add in the revision results from multiple random initializations together with convergence metrics. These empirical checks will support that the optimizer reaches the physically correct metric; a general analytic uniqueness proof lies outside the scope of the present numerical study but is not required for the concrete demonstrations we provide. revision: yes
Referee: [Gubser-Rocha model] Gubser-Rocha discussion: the statement that strip entanglement entropy constrains only the spatial metric component is asserted but not accompanied by an explicit demonstration (e.g., a plot or derivative showing vanishing sensitivity to the timelike component). Without this, the necessity of adding Wilson-loop data to lift the degeneracy remains incompletely justified.

Authors: We acknowledge that an explicit demonstration would strengthen the justification for the degeneracy. In the revised manuscript we will include a plot (or numerical derivative) that quantifies the sensitivity of the strip entanglement entropy to variations in the timelike metric component, confirming its negligible effect and thereby clarifying why Wilson-loop data is required to resolve the one-function degeneracy. revision: yes

Circularity Check

0 steps flagged

No circularity: NN optimization uses standard RT/Nambu-Goto losses validated on external known solutions

full rationale

The derivation reconstructs bulk metric functions by minimizing the Ryu-Takayanagi area functional and Nambu-Goto action as differentiable losses against boundary entanglement entropy and Wilson-loop data. Validation occurs on independent known backgrounds (AdS-Schwarzschild, Gubser-Rocha) where the recovered functions are compared to exact analytic expressions, achieving sub-0.2% accuracy without closed-form derivative relations. The semi-analytic inversion combines Bilson and Hashimoto formulas (external references) with a three-network variational method. No step reduces by the paper's own equations to a quantity defined in terms of a fitted parameter, self-citation chain, or ansatz smuggled from prior author work. The central claim rests on numerical convergence to known external benchmarks rather than internal self-definition.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on two standard domain assumptions from holography and introduces no new free parameters or invented entities; the neural-network weights are part of the numerical method rather than physical inputs.

axioms (2)

domain assumption The Ryu-Takayanagi formula equates boundary entanglement entropy to the area of a bulk minimal surface.
Invoked throughout the abstract as the differentiable loss for the inverse problem.
domain assumption The Nambu-Goto action gives the expectation value of holographic Wilson loops.
Used to couple the timelike metric component and resolve the degeneracy.

pith-pipeline@v0.9.0 · 5439 in / 1561 out tokens · 54472 ms · 2026-05-10T19:29:32.619533+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

24 extracted references · 22 canonical work pages · 1 internal anchor

[1]

Holographic Derivation of Entanglement Entropy from AdS/CFT

S. Ryu and T. Takayanagi, “Holographic derivation of entanglement entropy from AdS/CFT,” Phys. Rev. Lett.96(2006) 181602, [hep-th/0603001]

work page Pith review arXiv 2006
[2]

Aspects of Holographic Entanglement Entropy

S. Ryu and T. Takayanagi, “Aspects of holographic entanglement entropy,” JHEP08(2006) 045, [hep-th/0605073]

work page Pith review arXiv 2006
[3]

Holographic flavour and neural networks,

V. G. Filev, “Holographic flavour and neural networks,” JHEP11(2025) 031, [arXiv:2506.20115]

work page arXiv 2025
[4]

Deep Learning and Holographic QCD,

K. Hashimoto, S. Sugishita, A. Tanaka and A. Tomiya, “Deep Learning and Holographic QCD,” Phys. Rev. D98(2018) 106014, [arXiv:1809.10536]

work page arXiv 2018
[5]

Dual geometry of entanglement entropy via deep learning,

C. Park, C. Hwang, K. Cho and S.-J. Kim, “Dual geometry of entanglement entropy via deep learning,” Phys. Rev. D106(2022) 106017, [arXiv:2205.04445]

work page arXiv 2022
[6]

Holographic reconstruction of black hole spacetime: machine learning and entanglement entropy,

D. Ahn, Y. Jeong, D. Kim and K.-Y. Yun, “Holographic reconstruction of black hole spacetime: machine learning and entanglement entropy,” JHEP01(2025) 025, [arXiv:2406.07395]

work page arXiv 2025
[7]

Aspects of holographic entanglement entropy using PINNs,

P. Deb and A. Sanghavi, “Aspects of holographic entanglement entropy using PINNs,” [arXiv:2509.25311]

work page arXiv
[8]

Learning the Inverse Ryu–Takayanagi Formula with Transformers,

D. Kim, “Learning the Inverse Ryu–Takayanagi Formula with Transformers,” [arXiv:2511.06387]

work page arXiv
[9]

Bulk metric reconstruction from entanglement entropy,

N. Jokela, M. Liimatainen, M. Sarkkinen and L. Tzou, “Bulk metric reconstruction from entanglement entropy,” JHEP10(2025) 079, [arXiv:2504.07016]

work page arXiv 2025
[10]

Inverse problem of correlation functions in holography,

B.-W. Fan and R.-Q. Yang, “Inverse problem of correlation functions in holography,” JHEP 10(2024) 228, [arXiv:2310.10419]

work page arXiv 2024
[11]

Application of solving inverse scattering problem in holographic bulk reconstruction,

B.-W. Fan and R.-Q. Yang, “Application of solving inverse scattering problem in holographic bulk reconstruction,” JHEP03(2026) 044, [arXiv:2511.12886]

work page arXiv 2026
[12]

Building bulk from Wilson loops,

K. Hashimoto, “Building bulk from Wilson loops,” PTEP2021(2021) 023B04, [arXiv:2008.10883]

work page arXiv 2021
[13]

Peculiar properties of a charged dilatonic black hole in AdS5,

S. S. Gubser and F. D. Rocha, “Peculiar properties of a charged dilatonic black hole in AdS5,” Phys. Rev. D81(2010) 046001, [arXiv:0911.2898]. – 29 –

work page arXiv 2010
[14]

Inability of linear axion holographic Gubser–Rocha model to capture all the transport anomalies of strange metals,

W. Li and S. Liu, “Inability of linear axion holographic Gubser–Rocha model to capture all the transport anomalies of strange metals,” Phys. Rev. B108(2023) 235104, [arXiv:2307.04433]

work page arXiv 2023
[15]

Wilson loops in large N field theories

J. M. Maldacena, “Wilson loops in largeNfield theories,” Phys. Rev. Lett.80(1998) 4859, [hep-th/9803002]

work page Pith review arXiv 1998
[16]

Macroscopic strings as heavy quarks: Large-N gauge theory and anti-de Sitter supergravity

S.-J. Rey and J.-T. Yee, “Macroscopic strings as heavy quarks in largeNgauge theory and anti-de Sitter supergravity,” Eur. Phys. J. C22(2001) 379, [hep-th/9803001]

work page Pith review arXiv 2001
[17]

One-loop effective action of the holographic antisymmetric Wilson loop,

A. Faraggi, W. Mueck and L. A. Pando Zayas, “One-loop effective action of the holographic antisymmetric Wilson loop,” Phys. Rev. D85(2012) 106015, [arXiv:1112.5028]

work page arXiv 2012
[18]

Universal properties of the Langevin diffusion coefficients,

D. Giataganas and H. Soltanpanahi, “Universal properties of the Langevin diffusion coefficients,” Phys. Rev. D89(2014) 026011, [arXiv:1310.6725]

work page arXiv 2014
[19]

Extracting spacetimes using the AdS/CFT conjecture,

S. Bilson, “Extracting spacetimes using the AdS/CFT conjecture,” JHEP08(2008) 073, [arXiv:0807.3695]

work page arXiv 2008
[20]

Extracting Spacetimes using the AdS/CFT Conjecture: Part II,

S. Bilson, “Extracting Spacetimes using the AdS/CFT Conjecture: Part II,” JHEP02 (2011) 050, [arXiv:1012.1812]

work page arXiv 2011
[21]

A Covariant Holographic Entanglement Entropy Proposal

V. E. Hubeny, M. Rangamani and T. Takayanagi, “A covariant holographic entanglement entropy proposal,” JHEP07(2007) 062, [arXiv:0705.0016]

work page Pith review arXiv 2007
[22]

PyTorch: An Imperative Style, High-Performance Deep Learning Library,

A. Paszkeet al., “PyTorch: An Imperative Style, High-Performance Deep Learning Library,” NeurIPS (2019)

2019
[23]

Adam: A Method for Stochastic Optimization

D. P. Kingma and J. Ba, “Adam: A Method for Stochastic Optimization,” ICLR (2015), [arXiv:1412.6980]

work page internal anchor Pith review Pith/arXiv arXiv 2015
[24]

Get Physics Done (GPD),

Physical Superintelligence PBC, “Get Physics Done (GPD),” v1.1.0 (2026), https://github.com/psi-oss/get-physics-done. – 30 –

2026