Pith. sign in

REVIEW 4 major objections 5 minor 68 references

Supervised and unsupervised learning of the many-body critical phase, phase transitions, and critical exponents in disordered quantum systems

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Neural networks trained on raw eigenvalue spacings can find the many-body critical phase and its subphases, and reproduce known correlation-length exponents in the single-particle limit.

desk verdict Useful single-particle scaling result wrapped around a circular many-body story; the subphase claims need independent validation before they can be taken seriously. read the letter →

arxiv 2501.03981 v2 pith:AVEFFDPR submitted 2025-01-07 cond-mat.dis-nn cond-mat.stat-mechphysics.data-an

Aamna Ahmed , Nilanjan Roy This is my paper · ORCID
classification cond-mat.dis-nncond-mat.stat-mechphysics.data-an
keywords many-bodycriticalphasesupervisedlearningunsupervisedprincipalcomponentanalysiseigenvaluespacingseigenvectorprobabilitydensitiesinverseparticipationratiofinite-sizescaling
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper asks whether a neural network can detect a disputed third phase—the many-body critical (MBC) phase—in a finite interacting quasiperiodic system using nearly unprocessed inputs. It claims that fully connected networks trained only on eigenvalue spacings or eigenvector probability densities separate MBC from the ergodic and many-body-localized phases, and that the network outputs locate phase boundaries consistently with conventional diagnostics. It further claims that the same outputs reveal two MBC subphases split near $\lambda=2$ and a crossover inside the MBL region near $\mu=1$. In the non-interacting limit, it claims that PCA entropy of eigenvector probability densities marks single-particle transitions, and that a binary classifier fed inverse-participation-ratio components recovers the correlation-length exponents $\nu=1$, $0.4$, and $0.76$ found by scaling the IPR itself. A sympathetic reader would care because it suggests that generic, cheap spectral data can stand in for tailored order parameters in poorly understood disordered phases.

What carries the argument

The central machinery is a set of fully connected neural networks that learn phase structure directly from raw spectral data: eigenvalue spacings $\{E_{i+1}-E_i\}$ and eigenvector probability densities $|\psi_i|^2$ drawn from a narrow energy window. Binary classifiers expose the untrained MBC phase through output ambiguity near $0.5$, while multiclass classifiers with softmax outputs provide phase probabilities averaged over disorder samples; a confusion heuristic—a class whose probability stays broadly high across parameter space is genuine, while one that peaks only near training points is an artifact—drives the subphase claims. For unsupervised detection, PCA of eigenvector probability densities yields a PCA entropy $S_{PCA}=-\sum_i p_i\log_2 p_i$ whose numerical derivative peaks at transition points. For critical exponents, the paper scales the output probability of a binary classifier trained on IPR components using the ansatz $P=f(|\tau-\tau_c|N^{1/\nu})$ and a cost function that measures the scatter of the collapsed curve.

What would settle it

Train the same three-class network on eigenvalue spacings from the claimed MBC region with the MBC labels replaced by randomly shuffled ME/MBL labels at the same parameter points; if accuracy stays near 99% and the MBC probability remains broad, the network is tracking label structure rather than a physical phase. Alternatively, extrapolate the subthermal volume-law entanglement coefficient to larger N across the MBC window: if the window shrinks with N, the detected phase is a finite-size artifact.

Watch

Extended reading notes

Core claim

The paper's central claim is that the many-body critical phase in the interacting extended Aubry-André-Harper model can be separately detected by supervised neural networks fed with unprocessed eigenvalue spacings and eigenvector probability densities, without needing entanglement spectra or other sophisticated inputs. Trained on labels taken from the known phase diagram, binary classifiers give ambiguous output probability near $0.5$ precisely in the MBC region, and three-class classifiers assign probability near $1$ in each of the ME, MBC, and MBL phases. The resulting boundaries for the ME–MBC, ME–MBL, and MBC–MBL transitions are consistent with conventional measures. When the network is asked to split the phase diagram at $\lambda=2$, it returns broad, training-point-independent probabilities for two MBC subphases but not for two ME subphases, which the authors interpret as a genuine crossover in the multifractal structure of many-body eigenstates. In the single-particle limit, PCA entropy jumps at the known transition points, and finite-size scaling of the classifier probability trained on IPR components gives correlation-length exponents identical to those from direct IPR scaling.

Load-bearing premise

The training labels assume that the phase diagram of a prior finite-size study is correct—that the many-body critical phase exists and sits where that study placed it—so the network's high accuracy is a consistency check on those labels rather than an independent discovery of the phase.

Editorial extensions

If this is right

  • If the MBC detection is correct, the interacting EAAH model has interior structure near $\lambda=2$ that level statistics alone cannot resolve.
  • If network outputs locate phase boundaries consistently, machine-classified boundaries could serve as a cheap first pass over parameter space before expensive entanglement or Fock-space diagnostics are run.
  • If IPR-component-trained classifiers reproduce $\nu=1$, $0.4$, and $0.76$ for the three single-particle transitions, classifier probabilities inherit the scaling content of the input data and can act as a proxy observable for scaling collapse.
  • The transfer test in Appendix C, if reliable, indicates that eigenvalue-spacing patterns carry phase information even in models where gap ratios are inconclusive.
  • For the many-body system, the same scaling approach does not yet yield standard $\tau_c$ and $\nu$ at the available sizes, so critical exponents in the MBC region remain an open problem.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • One testable extension is to train the multiclass network on labels from one model and apply it to a second model whose phase boundaries are known independently; the paper's Appendix C does one such transfer, but the MBC-training-label dependence on the first model's phase diagram remains untested.
  • The subphase claims rest on a heuristic about broad versus peak-like output probabilities; comparing the classifier probability surface against a null model trained on randomly split versions of the same phase would formalize that heuristic.
  • If the MBC-I and MBC-II split near $\lambda=2$ is real, it may signal a change in the multifractal spectrum of many-body eigenstates rather than a thermodynamic transition, which could be tested by computing the full multifractal spectrum on each side at larger system sizes.
  • The PCA-entropy route might generalize to Floquet or open systems where exact eigenstate probability densities are replaced by steady-state density matrices, although the paper does not discuss that extension.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper applies supervised fully connected neural networks and unsupervised PCA to eigenvalue spacings and eigenvector probability densities of the interacting extended Aubry-André-Harper (EAAH) model, aiming to detect the many-body critical (MBC) phase, locate ME/MBC/MBL transitions, and identify subphases (MBC-I/II, MBL-I/II). In the single-particle limit, the authors use PCA entropy to locate delocalized-critical-localized transitions and perform finite-size scaling of a binary classifier trained on inverse participation ratio (IPR) components, comparing the extracted exponents with those from scaling of the IPR itself. The paper also tests a network trained on EAAH level spacings on an interacting flat-band diamond model where level statistics are inconclusive. The central conceptual claim is that ML can "support the existence of the MBC phase" and reveal genuine substructure, while the concrete quantitative result is that the single-particle correlation-length exponent ν extracted from classifier outputs matches ν from IPR scaling for three transitions.

Significance. If the label-free aspects of the claim could be established, the paper would provide a useful demonstration that raw eigenvalue spacings and eigenvector probability densities carry enough information for machine-learning phase classification in a quasiperiodic interacting system. The single-particle finite-size scaling results in Fig. 9 and Table III are internally consistent and give a concrete, falsifiable demonstration that a classifier output can reproduce the correlation-length exponent ν obtained from IPR scaling. The PCA-entropy diagnostics in Section IV B locate known single-particle transitions with reasonable accuracy, and the cross-model transfer experiment in Appendix C is intriguing. However, the most novel physical claims—independent detection of the MBC phase and identification of MBC and MBL subphases—are not supported by the present analysis because the supervised labels are taken directly from the phase diagram of Ref. [28] and the subphase criterion is an unquantified broadness heuristic. The paper is therefore best viewed as a promising methodology demonstration whose headline physics conclusions require additional label-free controls.

major comments (4)
  1. [III C, Fig. 2] The three-class classifier is trained on labels taken directly from the phase diagram of Ref. [28] (ME: μ=0.3, λ=1, 2.5; MBC: μ=2, λ=0.3, 1, 2.5, 3; MBL: μ=0.5, 2, λ=4.5), so the reported 99% and 95% accuracies and the transition locations in Fig. 2 measure label consistency, not an independent confirmation of the MBC phase. The binary classifier's P≈0.5 plateau in the MBC region could also arise generically from a network trained only on two separated training clusters, without implying a distinct third phase. To support the Section III C claim that "this analysis supports the existence of the MBC phase," the authors need a label-free control, such as unsupervised clustering of the same feature vectors or a network trained on two artificially split clusters and tested on the same parameter region, with a quantitative comparison.
  2. [III D, Appendix D] The four-class subphase inference relies on an unquantified heuristic: a "broad probability distribution across the phase" is taken to indicate a genuine subphase, while a probability peaked only near training points is taken to indicate an artificial split. No null model, threshold, confidence interval, or alternative explanation (for example, finite-size interpolation near the training points) is provided. Since the phase is divided at arbitrary values (λ=2 for MBC-I/II, μ=1 for MBL-I/II) and the same heuristic is used for both the proposed subphases and the MBL crossover, the subphase and crossover claims are not currently supported. Please add a statistical test, such as label permutation, bootstrap resampling, or a comparison with a synthetic single-phase model, to show that broad probability is not a generic feature of four-class softmax outputs.
  3. [V, Table III] The abstract's phrase "identical critical exponents" overstates the results: in Table III the correlation-length exponent ν agrees between IPR scaling and classifier scaling, but the exponent w differs substantially for the delocalized-to-localized transition (0.42 vs 0.03) and for the critical-to-localized transition (0.24 vs 0). Furthermore, the paper itself states that the many-body scaling collapse is not satisfactory, with τc and ν deviating from the values reported in Ref. [28]. The abstract and conclusion should make clear that the demonstrated agreement is for the single-particle correlation-length exponent ν only, that w differs by construction or by measurement, and that the many-body exponent extraction remains an open problem.
  4. [IV C, Fig. 8] The many-body PCA results are presented as an effective phase indicator, but the transition locations are only described as "roughly" μ≈1 and λ≈2.7, no error bars or finite-size extrapolation are given, and the text explicitly acknowledges that the MBC-to-MBL transition is not clearly defined. This is acceptable as a preliminary observation, but the main text should consistently label these PCA-based boundaries as indicative rather than as established transition points, particularly since the preprocessing (rearranging normalized PDs) was chosen after observing the data structure in Fig. 7.
minor comments (5)
  1. [Appendix B] There is a typo in "criitcal-to-localized" in the sentence describing Fig. 11(c); it should read "critical-to-localized."
  2. [Fig. 7 caption] The caption says "The first principle component plotted against the second"; the correct term is "principal component." The same correction should be applied in the running text where the misspelling appears.
  3. [II B] The energy window ε∈[0.53,0.55] is stated without justification. A sentence explaining why this window was chosen (for example, proximity to infinite-temperature states or avoidance of spectral edges) would improve reproducibility.
  4. [V] The text reports that the cost function is minimized for specific (ν, w) pairs, but the cost landscape is not shown for all transitions. Showing CQ as a function of the scaling parameters, or at least stating the grid resolution and search range, would allow readers to assess how well the minimum is determined.
  5. [General] No code or data availability statement is provided. The exact diagonalization and TensorFlow-based training procedures described in Sections II and III would be much easier to reproduce if the scripts and generated datasets were released.

Circularity Check

2 steps flagged · score 6.0 of 10

The many-body MBC and subphase claims reduce to training labels from the very phase diagram being questioned plus a self-defined broadness criterion; the single-particle PCA and scaling-collapse content is independent.

  1. fitted input called prediction [Section III C, 'Three-class classification' (training-data paragraph)]
    "Utilizing the known phase diagram, we train the network on datasets deeply representative of each phase and then test it on the unseen data. The training data sets for each class are generated as follows: 1250 disorder realizations for eigenvalue spacings and 10000 for the probability densities corresponding to eigenstates. For the ME phase, we use µ = 0.3 and λ = 1, 2.5; for the MBC phase, µ = 2 and λ = 0.3, 1, 2.5, 3; and for the MBL phase, µ = 0.5, 2 and λ = 4.5."

    The paper opens by saying it is 'questioning the existence' of the MBC phase, but the three-class network is trained with MBC labels taken directly from the phase diagram of Ref. [28], the very source whose MBC claim is under test. The reported 99%/95% accuracy therefore certifies only that the labelled MBC region is separable in eigenvalue spacings and eigenvector probability densities; it does not establish that a distinct MBC phase exists. The later statement 'This analysis supports the existence of the MBC phase' is an interpolation of the input labels: the network maps test points inside the labelled MBC region to the MBC class by construction. Any 'detection' of the MBC phase along the µ and λ paths is thus the fitted label structure renamed as a prediction.

  2. self definitional [Section III D, 'Four-class classification' (genuine vs artificial subphase criterion); also applied in Appendix D]
    "If a phase is artificially split, the probability of the non-existent subphase will remain high only near the training data points. Conversely, for a genuine subphase, the neural network output would exhibit a broad probability distribution across the phase."

    What counts as a 'genuine subphase' is defined by the broadness of the very network output used as evidence. The phase diagram is first partitioned at an arbitrary cut (λ = 2 for MBC-I/II, µ = 1 for MBL-I/II), the network is trained to separate those invented classes, and then the observation that the output is broad is presented as confirmation that the split is physical. No null model, threshold, or independent order parameter is supplied, so the conclusion 'the MBC phase can be meaningfully divided into MBC-I and MBC-II subphases' is the defining heuristic restated as a finding. The MBL-I/II crossover claim in Appendix D uses the identical logic.

full rationale

The central many-body claim is substantially circular. The supervised classifiers are trained on phase labels taken from Ref. [28] (ME, MBC, and MBL training points), while the stated goal is to question the existence of the MBC phase; the high test accuracy and the MBC probability plateaus are therefore consequences of the training labels, not independent evidence for the phase. The four-class subphase analysis goes further: it defines genuineness as broad network probability after training on a partition of the phase diagram, so MBC-I/II and MBL-I/II are classified as real by the same criterion that constitutes them. These are real circular reductions, not speculative ones, and they affect the paper's headline claims about the many-body phase structure. However, not all content is circular. The single-particle PCA-entropy analysis is parameter-free and benchmarked against known AAH/EAAH transitions; the finite-size scaling of a binary classifier trained on IPR components reproduces ν from IPR scaling, which is a consistency check on the network rather than an independent derivation but is not a fitted-input renaming of the phase diagram. The many-body scaling collapse is explicitly reported as failing, and the binary-classifier 'confusion' signal is at least not trained on MBC labels. On balance, the paper contains independent material outside the circular many-body classification, so an intermediate score is appropriate; the central MBC and subphase claims are the parts that reduce by construction.

Assumptions & free parameters 6 free parameters · 4 assumptions · 2 invented entities

The list shows what the central claims pull from prior literature or from hand choices. The labels (and hence the phase and subphase 'discoveries') come from Ref [28]'s phase diagram; the exponents are fit outputs; and the subphase boundaries are hand-selected. The paper itself supplies no parameter-free derivation and no independent falsifiable handle for the new substructure.

free parameters (6)
  • Correlation-length exponent nu (delocalized-localized, mu=0) = 1
    Extracted by minimizing the collapse cost function CQ for both IPR and classifier probability (Fig. 9a-b, Table III).
  • Correlation-length exponent nu (delocalized-critical, lambda=0.5) = 0.4
    Extracted by minimizing CQ (Fig. 9c-d, Table III).
  • Correlation-length exponent nu (critical-localized, mu=1.5) = 0.76
    Extracted by minimizing CQ (Fig. 9e-f, Table III).
  • Scaling exponent w (IPR and classifier) = IPR: 0.42, 0.08, 0.24; classifier: 0.03, 0.06, 0.0
    Fitted from the same collapse; the discrepancy between IPR and classifier w values is acknowledged in the text and weakens the abstract's 'identical critical exponents' phrasing.
  • Eigenvector energy window for many-body PDs = epsilon in [0.53, 0.55]
    Hand-chosen window for extracting eigenvector PD samples (Section II B); results could depend on this choice.
  • Subphase split for four-class tests = lambda=2 (MBC-I/II) and mu=1 (MBL-I/II)
    Hand-chosen partition points for the four-class classifiers; the subphase conclusions are defined relative to these splits (Section III D, Appendix D).
assumptions (4)
  • domain assumption The phase diagram of the interacting EAAH model from Ref [28] (ME, MBC, MBL regions) is correct for N=14 and is used to label all supervised training data.
    Section III C: 'Utilizing the known phase diagram, we train the network on datasets deeply representative of each phase.' The paper claims to question the MBC phase, but the labels presuppose its existence and boundaries.
  • domain assumption The scaling ansatz IPR = N^{-w} f(|tau - tau_c| N^{1/nu}) and its analogue for the classifier probability hold.
    Section V, Eq. 13; standard finite-size scaling, but the validity of the ansatz for the ML probability is assumed, and the many-body case fails.
  • ad hoc to paper The 'broad probability across parameter space vs probability peaked only at training points' criterion identifies genuine subphases.
    Section III D: this criterion has no theoretical or independent empirical backing; it is introduced in this paper to interpret four-class outputs.
  • domain assumption The binary classifier output near 0.5 on untrained data indicates a distinct unseen phase (confusion signature).
    Section III B and used to claim MBC detection; this heuristic is standard in the ML-for-physics literature but is not a proof of a distinct phase.
invented entities (2)
  • MBC-I and MBC-II subphases within the many-body critical phase, separated near lambda=2
    purpose: To partition the MBC phase into two distinct regimes.
    The only evidence is the four-class NN probability broadness (Section III D, Figs. 4a-d), which is circular because the classifier was trained to distinguish the split. No independent physical observable shows a clear transition; the paper itself frames it as a 'possibility of a crossover.'
  • MBL-I and MBL-II subphases or crossover within the MBL phase, separated near mu=1
    purpose: To partition the MBL phase into two distinct regimes.
    Appendix D, Fig. 13: similarly based on trained four-class NN probability broadness; no independent corroboration is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Supervised and unsupervised learning of the many-body critical phase, phase transitions, and critical exponents in disordered quantum systems." pith.science (2026). https://pith.science/paper/AVEFFDPR

@misc{pith2026250103981,
  author       = {Pith},
  title        = {Pith review of: Supervised and unsupervised learning of the many-body critical phase, phase transitions, and critical exponents in disordered quantum systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AVEFFDPR}},
  note         = {Machine review of arXiv:2501.03981}
}
read the original abstract

In this work, we begin by questioning the existence of a new kind of nonergodic extended phase, namely, the many-body critical (MBC) phase in finite systems of an interacting quasiperiodic system. We find that this phase can be separately detected from the other phases such as the many-body ergodic (ME) and many-body localized (MBL) phases in the model through supervised neural networks made for both binary and multi-class classification tasks, utilizing, rather un-preprocessed, eigenvalue spacings and eigenvector probability densities as input features. Moreover, the output of our trained neural networks can also indicate the critical points separating ME, MBC and MBL phases, which are consistent with the same obtained from other conventional methods. We also employ unsupervised learning techniques, particularly principal component analysis (PCA) of eigenvector probability densities to investigate how this framework, without any training, captures the, rather unknown, many-body phases (ME, MBL and MBC) and single particle phases (delocalized, localized and critical) of the interacting and non-interacting systems, respectively. Our findings reveal that PCA entropy serves as an effective indicator (order parameter) for detecting phase transitions in the single-particle systems. Moreover, this method proves applicable to many-body systems when the data undergoes a suitable pre-processing. Interestingly, when it comes to extraction of critical (correlation length) exponents through a finite size-scaling, we find that for single-particle systems, scaling collapse of neural network outputs is obtained using components of inverse participation ratio (IPR) as input data. Remarkably, we observe identical critical exponents as obtained from scaling collapse of the IPR directly for different single-particle phase transitions.

Figures

Figures reproduced from arXiv: 2501.03981 by the authors.

Figure 1
Figure 1. FIG. 1. Neural network predictions for the binary classifier are shown for (a, d) increasing hopping amplitude [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. Neural network predictions for the 3 [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3. Neural network probabilities for the 4-class classifier, [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: FIG. 4. Neural network probabilities for the 4-class classifier, [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5. (a) PCA-entropy [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6. For the extended Harper model, (a) PCA-entropy [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 8
Figure 8. Figure 8: FIG. 8. The difference between the first explained variance ra [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 7
Figure 7. Figure 7: FIG. 7. The first principle component plotted against the [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 9
Figure 9. Figure 9: FIG. 9. Finite-size scaling of the IPR and binary classifier [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: FIG. 10. Gap ratio (a), (c), (e) and MIPR (b), (d), (f) [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: FIG. 11. Neural network probability for the 3-class classifier: (a) as a function of hopping amplitude [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]
Figure 12
Figure 12. Figure 12: FIG. 12. For the system described by the ABF Hamiltonian in [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

68 extracted references · 49 canonical work pages

  1. [28]

    D. A. Abanin, E. Altman, I. Bloch, and M. Serbyn, Col- loquium: Many-body localization, thermalization, and entanglement, Reviews of Modern Physics 91, 021001 (2019)

    2019

  2. [1]

    The Rectified Linear Unit (ReLU) [51]: ReLU: f (xi) = ( 0, x i < 0, xi, x i ≥ 0, (2) 4 (a) (b) (c) (d) (e) (f) FIG. 1. Neural network predictions for the binary classifier are shown for (a, d) increasing hopping amplitude µ at λ = 0.5, and for increasing disorder strength λ at fixed (b, e) µ = 0.5 and (c, f) µ = 1.8. Panels (a–c) utilize eigenvalue spacin...

    work page

  3. [2]

    These computations can result in both positive and negative values propagating through the network

    The Exponential Linear Unit (ELU) [52]: ELU: f (xi) = ( α(exi − 1), x i ≤ 0, xi, x i > 0, (3) Although the initial inputs to the network are strictly positive, the use of activation functions in the hidden layers is essential due to the intermediate computations involving weights and biases. These computations can result in both positive and negative valu...

    work page

  4. [3]

    Importantly, the eigenvalues of the covariance matrix are non-negative and sum up to the total variance of the original dataset

    Explained Variance and Dimensionality The maximum number of principal components is de- termined by the rank of the data matrix X, given by min(n, m). Importantly, the eigenvalues of the covariance matrix are non-negative and sum up to the total variance of the original dataset. The principal components are or- dered by the magnitude of variance they capt...

    work page

  5. [4]

    PCA-Entropy The eigenvalues e1, e2, . . . , ek of the covariance matrix can be normalized to compute the explained variance ra- tios (EVR): pi = eiP j ej , (8) where pi represents the fraction of variance captured by the ith principal component. The PCA entropy is then defined as [56]: SP CA= − kX i=1 pi log2 pi. (9) This entropy is rescaled between 0 and...

    work page

  6. [5]

    (11) Here, λ represents any disorder parameter

    Transition Region and Numerical Derivative To analyze the transition region of the PCA entropy, its numerical derivative is computed using the symmetric difference quotient: δ ˜SPCA δλ = ˜SP CA(λ + ∆λ) − ˜SP CA(λ − ∆λ) 2∆λ . (11) Here, λ represents any disorder parameter. This ap- proach identifies regions of significant change in PCA en- tropy, providing...

    work page

  7. [6]

    Schindler, N

    F. Schindler, N. Regnault, and T. Neupert, Probing many-body localization with neural networks, Phys. Rev. B 95, 245134 (2017)

    work page 2017

  8. [7]

    S. J. Wetzel, Unsupervised learning of phase transitions: From principal component analysis to variational autoen- coders, Phys. Rev. E 96, 022140 (2017)

    2017

Show all 68 references
  1. [8]

    W. Hu, R. R. P. Singh, and R. T. Scalettar, Discovering phases, phase transitions, and crossovers through unsu- pervised machine learning: A critical examination, Phys. Rev. E 95, 062122 (2017)

    2017

  2. [9]

    ˇCadeˇ z, B

    T. ˇCadeˇ z, B. Dietz, D. Rosa, A. Andreanov, K. Slevin, and T. Ohtsuki, Machine learning wave functions to iden- tify fractal phases, Phys. Rev. B 108, 184202 (2023)

    2023

  3. [10]

    S. Lu, X. Gao, and L.-M. Duan, Efficient representation of topologically ordered states with restricted boltzmann machines, Phys. Rev. B 99, 155136 (2019)

    2019

  4. [11]

    Carleo, Y

    G. Carleo, Y. Nomura, and M. Imada, Constructing ex- act representations of quantum many-body systems with deep neural networks, Nature Communications 9 (2018)

    2018

  5. [12]

    Torlai, G

    G. Torlai, G. Mazzola, J. Carrasquilla, M. Troyer, R. Melko, and G. Carleo, Neural-network quantum state tomography, Nature Physics 14, 447 (2018)

    2018

  6. [13]

    Cai and J

    Z. Cai and J. Liu, Approximating quantum many-body wave functions using artificial neural networks, Phys. Rev. B 97, 035116 (2018)

    2018

  7. [14]

    Gardas, M

    B. Gardas, M. M. Rams, and J. Dziarmaga, Quantum neural networks to simulate many-body quantum sys- tems, Physical Review B 98 (2018)

    2018

  8. [15]

    Carrasquilla and G

    J. Carrasquilla and G. Torlai, How To Use Neural Networks To Investigate Quantum Many-Body Physics, 17 PRX Quantum 2 (2021)

    2021

  9. [16]

    Schmitt and M

    M. Schmitt and M. Heyl, Quantum Many-Body Dynam- ics in Two Dimensions with Artificial Neural Networks, Physical Review Letters 125 (2020)

    2020

  10. [17]

    Venderley, V

    J. Venderley, V. Khemani, and E.-A. Kim, Machine learning out-of-equilibrium phases of matter, Phys. Rev. Lett. 120, 257204 (2018)

    2018

  11. [18]

    H. Cao, D. G. Angelakis, and D. Leykam, Unsupervised learning of quantum many-body scars using intrinsic di- mension, Machine Learning: Science and Technology 5, 025049 (2024)

    2024

  12. [19]

    Ch’ng, J

    K. Ch’ng, J. Carrasquilla, R. G. Melko, and E. Khatami, Machine learning phases of strongly correlated fermions, Phys. Rev. X 7, 031038 (2017)

    2017

  13. [20]

    Broecker, J

    P. Broecker, J. Carrasquilla, R. G. Melko, and S. Trebst, Machine learning quantum phases of matter beyond the fermion sign problem, Scientific Reports 7, 8823 (2017)

    2017

  14. [21]

    Carleo and M

    G. Carleo and M. Troyer, Solving the quan- tum many-body problem with artificial neu- ral networks, Science 355, 602 (2017), https://www.science.org/doi/pdf/10.1126/science.aag2302

    2017 doi

  15. [22]

    Zhang and E.-A

    Y. Zhang and E.-A. Kim, Quantum loop topography for machine learning, Phys. Rev. Lett. 118, 216401 (2017)

    2017

  16. [23]

    Ahmed, A

    A. Ahmed, A. Nelson, A. Raina, and A. Sharma, Phase classification in the long-range harper model using ma- chine learning, Phys. Rev. B 108, 155128 (2023)

    2023

  17. [24]

    Muzzi, R

    C. Muzzi, R. S. Cortes, D. S. Bhakuni, A. Jeli´ c, A. Gam- bassi, M. Dalmonte, and R. Verdel, Principal component analysis of absorbing state phase transitions, Phys. Rev. E 110, 064121 (2024)

    2024

  18. [25]

    Vanoni and V

    C. Vanoni and V. Vitale, Analysis of localization transi- tions using nonparametric unsupervised learning, Phys. Rev. B 110, 024204 (2024)

    2024

  19. [26]

    R. K. Panda, R. Verdel, A. Rodriguez, H. Sun, G. Bian- coni, and M. Dalmonte, Non-parametric learning critical behavior in Ising partition functions: PCA entropy and intrinsic dimension, SciPost Phys. Core 6, 086 (2023)

    2023

  20. [27]

    L. c. v. Frk, P. Bal´ aˇ z, E. Archemashvili, and M.ˇZonda, Unsupervised machine learning phase classification for the falicov-kimball model, Phys. Rev. B 111, 205116 (2025)

    2025

  21. [29]

    Alet and N

    F. Alet and N. Laflorencie, Many-body localization: An introduction and selected topics, Comptes Rendus Physique 19, 498 (2018)

    2018

  22. [30]

    Sierant, M

    P. Sierant, M. Lewenstein, A. Scardicchio, L. Vidmar, and J. Zakrzewski, Many-body localization in the age of classical computing*, Reports on Progress in Physics 88, 026502 (2025)

    2025

  23. [31]

    J. M. Deutsch, Eigenstate thermalization hypothesis, Re- ports on Progress in Physics 81, 082001 (2018)

    2018

  24. [32]

    X. Li, S. Ganeshan, J. Pixley, and S. Das Sarma, Many- body localization and quantum nonergodicity in a model with a single-particle mobility edge, Physical review let- ters 115, 186601 (2015)

    2015

  25. [33]

    Y. Wang, C. Cheng, X.-J. Liu, and D. Yu, Many-body critical phase: Extended and nonthermal, Phys. Rev. Lett. 126, 080602 (2021)

    2021

  26. [34]

    Rispoli, A

    M. Rispoli, A. Lukin, R. Schittko, S. Kim, M. E. Tai, J. L´ eonard, and M. Greiner, Quantum critical behaviour at the many-body localization transition, Nature 573, 385 (2019)

    2019

  27. [35]

    N. Roy, J. Sutradhar, and S. Banerjee, Diagnostics of nonergodic extended states and many body localization proximity effect through real-space and fock-space exci- tations, Physical Review B 107, 115155 (2023)

    2023

  28. [36]

    N. Roy, S. Mukerjee, and S. Banerjee, Many-body critical phase in a quasiperiodic chain and dynamical widom lines in fock space properties (2025), arXiv:2505.02234 [cond- mat.dis-nn]

    2025

  29. [37]

    S. Bera, T. Martynec, H. Schomerus, F. Heidrich- Meisner, and J. H. Bardarson, One-particle density ma- trix characterization of many-body localization, Annalen der Physik 529, 1600356 (2017)

    2017

  30. [38]

    Serbyn, A

    M. Serbyn, A. A. Michailidis, D. A. Abanin, and Z. Papi´ c, Power-law entanglement spectrum in many-body local- ized phases, Phys. Rev. Lett. 117, 160601 (2016)

    2016

  31. [39]

    Ahmed, N

    A. Ahmed, N. Roy, and A. Sharma, Interplay of many- body interactions and quasiperiodic disorder in the all-bands-flat diamond chain, Physical Review B 107, 245110 (2023)

    2023

  32. [40]

    Y.-T. Hsu, X. Li, D.-L. Deng, and S. Das Sarma, Machine learning many-body localization: Search for the elusive nonergodic metal, Phys. Rev. Lett. 121, 245701 (2018)

    2018

  33. [41]

    Rao, Machine learning the many-body localization transition in random spin systems, Journal of Physics: Condensed Matter 30, 395902 (2018)

    W.-J. Rao, Machine learning the many-body localization transition in random spin systems, Journal of Physics: Condensed Matter 30, 395902 (2018)

    2018

  34. [42]

    S. Wold, K. Esbensen, and P. Geladi, Principal compo- nent analysis, Chemometrics and Intelligent Laboratory Systems 2, 37 (1987), proceedings of the Multivariate Statistical Workshop for Geologists and Geochemists

    1987

  35. [43]

    Hatsugai and M

    Y. Hatsugai and M. Kohmoto, Energy spectrum and the quantum hall effect on the square lattice with next- nearest-neighbor hopping, Phys. Rev. B 42, 8282 (1990)

    1990

  36. [44]

    J. H. Han, D. J. Thouless, H. Hiramoto, and M. Kohmoto, Critical and bicritical properties of harper’s equation with next-nearest-neighbor coupling, Phys. Rev. B 50, 11365 (1994)

    1994

  37. [45]

    Takada, K

    Y. Takada, K. Ino, and M. Yamanaka, Statistics of spectra for critical quantum chaos in one-dimensional quasiperiodic systems, Phys. Rev. E 70, 066203 (2004)

    2004

  38. [46]

    F. Liu, S. Ghosh, and Y. D. Chong, Localization and adiabatic pumping in a generalized aubry-andr´ e-harper model, Phys. Rev. B 91, 014108 (2015)

    2015

  39. [47]

    Modugno, Exponential localization in one- dimensional quasi-periodic optical lattices, New Journal of Physics 11, 033023 (2009)

    M. Modugno, Exponential localization in one- dimensional quasi-periodic optical lattices, New Journal of Physics 11, 033023 (2009)

    2009

  40. [48]

    Aubry and G

    S. Aubry and G. Andr´ e, Analyticity breaking and Ander- son localization in incommensurate lattices, Ann. Israel Phys. Soc. 3, 18 (1980)

    1980

  41. [49]

    D. J. Thouless, Bandwidths for a quasiperiodic tight- binding model, Phys. Rev. B 28, 4272 (1983)

    1983

  42. [50]

    Evers and A

    F. Evers and A. D. Mirlin, Anderson transitions, Rev. Mod. Phys. 80, 1355 (2008)

    2008

  43. [51]

    Chang, K

    I. Chang, K. Ikezawa, and M. Kohmoto, Multifractal properties of the wave functions of the square-lattice tight-binding model with next-nearest-neighbor hopping in a magnetic field, Physical Review B 55, 12971 (1997)

    1997

  44. [52]

    Aditya and N

    S. Aditya and N. Roy, Family-vicsek dynamical scaling and kardar-parisi-zhang-like superdiffusive growth of sur- face roughness in a driven one-dimensional quasiperiodic model, Phys. Rev. B 109, 035164 (2024)

    2024

  45. [53]

    S. R. Dubey, S. K. Singh, and B. B. Chaudhuri, Activa- tion functions in deep learning: A comprehensive survey and benchmark (2021). 18

    2021

  46. [54]

    I. J. Goodfellow, Y. Bengio, and A. Courville, Deep Learning (MIT Press, Cambridge, MA, USA, 2016) http://www.deeplearningbook.org

    2016

  47. [55]

    Fukushima, Cognitron: A self-organizing mul- tilayered neural network - Biological Cybernetics, https://link.springer.com/article/10.1007/BF00342633 (1975)

    K. Fukushima, Cognitron: A self-organizing mul- tilayered neural network - Biological Cybernetics, https://link.springer.com/article/10.1007/BF00342633 (1975)

    1975 doi

  48. [56]

    A. F. Agarap, Deep learning using rectified linear units (relu) (2019), arXiv:1803.08375 [cs.NE]

    2019 arXiv

  49. [57]

    Clevert, T

    D.-A. Clevert, T. Unterthiner, and S. Hochreiter, Fast and accurate deep network learning by exponential linear units (elus) (2016), arXiv:1511.07289 [cs.LG]

    2016 arXiv

  50. [58]

    Feng and S

    J. Feng and S. Lu, Performance Analysis of Various Ac- tivation Functions in Artificial Neural Networks, Journal of Physics: Conference Series 1237, 022030 (2019)

    2019

  51. [59]

    Goodfellow, Y

    I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning (MIT Press, 2016) http://www. deeplearningbook.org

    2016

  52. [60]

    Bottou, F

    L. Bottou, F. E. Curtis, and J. Nocedal, Optimization methods for large-scale machine learning, SIAM Review 60, 223 (2018), https://doi.org/10.1137/16M1080173

    2018 doi

  53. [61]

    C. E. Shannon, A mathematical theory of communica- tion, The Bell system technical journal 27, 379 (1948)

    1948

  54. [62]

    Sahoo, A

    A. Sahoo, A. Saha, and D. Rakshit, Stark localization near aubry-andr´ e criticality, Phys. Rev. B 111, 024205 (2025)

    2025

  55. [63]

    ˇSuntajs, J

    J. ˇSuntajs, J. Bonˇ ca, T. c. v. Prosen, and L. Vidmar, Ergodicity breaking transition in finite disordered spin chains, Phys. Rev. B 102, 064207 (2020)

    2020

  56. [64]

    Bauer, T.-M

    J. Bauer, T.-M. Chang, and J. L. Skinner, Correla- tion length and inverse-participation-ratio exponents and multifractal structure for anderson localization, Phys. Rev. B 42, 8121 (1990)

    1990

  57. [65]

    Th´ eveniaut and F

    H. Th´ eveniaut and F. Alet, Neural network setups for a precise detection of the many-body localization tran- sition: Finite-size scaling and limitations, Phys. Rev. B 100, 224202 (2019)

    2019

  58. [66]

    Abadi, A

    M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. Corrado, A. Davis, J. Dean, M. Devin, S. Ghemawat, I. Goodfellow, A. Harp, G. Irving, M. Is- ard, Y. Jia, R. Jozefowicz, L. Kaiser, M. Kudlur, J. Lev- enberg, D. Man´ e, R. Monga, S. Moore, D. Murray, C. Olah, M...

    2015

  59. [67]

    Y. Y. Atas, E. Bogomolny, O. Giraud, and G. Roux, Distribution of the ratio of consecutive level spacings in random matrix ensembles, Phys. Rev. Lett. 110, 084101 (2013)

    2013

  60. [68]

    T. L. M. Lezama, S. Bera, H. Schomerus, F. Heidrich- Meisner, and J. H. Bardarson, One-particle density ma- trix occupation spectrum of many-body localized states after a global quench, Phys. Rev. B 96, 060202 (2017)

    2017

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.