Analytic Continuation by Feature Learning

Ce Wang; Jingping Xu; Yaping Yang; Zhe Zhao

REVIEW 5 major objections 5 minor 33 references

Analytic Continuation by Feature Learning

T0 review · 5 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a neural network which first learns a compact latent feature representation of spectral functions and then maps imaginary-time Green's functions into that feature space reconstructs real-frequency spectra with at…

desk verdict A genuinely useful neural-network architecture for analytic continuation with a solid accuracy claim on synthetic benchmarks, but the advertised robustness trade-off rests on unvalidated approximations and needs a noisy-input check before it can be believed. read the letter →

arxiv 2411.17728 v1 pith:5LWQ3MXQ submitted 2024-11-22 cond-mat.str-el cs.LGeess.SPphysics.comp-phstat.ML

Zhe Zhao , Jingping Xu , Ce Wang , Yaping Yang This is my paper · ORCID

classification cond-mat.str-elcs.LGeess.SPphysics.comp-phstat.ML

keywords analyticcontinuationspectralfunctionneuralnetworkfeaturelearningrobustnessanalysissingularvaluedecompositionill-posedinverseproblemmaximumentropymethod

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

Analytic continuation, recovering a real-frequency spectral function from imaginary-time data, is an ill-posed inverse problem where standard methods like the maximum entropy method over-smooth sharp or multi-peak features. This paper proposes FL-net, which trains an autoencoder on spectral functions to find a low-dimensional latent feature vector, then trains a network to map the Green's function into that latent space before decoding back to the spectrum. On synthetic Gaussian and Lorentzian multi-peak datasets, FL-net attains roughly 20% lower prediction loss than MEM, a previously published neural network, and a direct-mapping baseline with comparable parameters. The paper also derives an SVD-based sensitivity measure showing that increasing the latent dimension lowers loss but makes the network more sensitive to input noise.

What carries the argument

FL-net is a two-encoder, one-decoder architecture: an autoencoder trained on spectral functions defines a latent feature vector h; a second encoder maps the imaginary-time Green's function G(iωn) to h, and a decoder maps h back to the spectral function A(Ω). The robustness machinery is the SVD of the Jacobian M = ∂A/∂G, whose singular values τj and left singular vectors uj yield Dj = ln(Σ_i u_{ij}^2/A_i), the log of the inverse Fisher information along mode j, and Sj = 2 ln τj + Dj, a mode-wise sensitivity that lets the authors compare networks of different hidden dimension.

What would settle it

Train FL-net and the comparison methods on one of the paper's synthetic Gaussian datasets, then evaluate on spectra with substantially different line shapes (e.g., asymmetric or strongly overlapping peaks) or on Green's functions with added noise of known variance; if the loss advantage over MEM and the previous neural network falls below 20% or reverses, the central claim fails outside the training distribution.

Watch

Extended reading notes

Core claim

The paper's central claim is that the mapping from imaginary-time Green's functions to spectral functions becomes both more accurate and more interpretable when it is factored through a learned latent feature space rather than performed directly. For a single Gaussian peak, FL-net's two latent variables are numerically equivalent to the peak's mean and variance, verified through the Jacobian of the encoder. For multi-peak spectra, the latent dimension that minimizes test loss is around three times the peak number minus one, the number of independent parameters in the synthetic data. The robustness analysis defines a per-mode noise sensitivity Sj = 2 ln τj + Dj from the singular value decomposition of the input–output Jacobian, and shows that larger hidden dimensions produce larger Sj, i.e., higher noise sensitivity, while lowering the normalized prediction loss.

Load-bearing premise

The reported accuracy gain assumes the test spectra come from the same synthetic families used for training, sums of Gaussian or Lorentzian peaks generated without input noise, so the 20% improvement and the robustness tradeoff are established only for those families.

Editorial extensions

If this is right

FL-net reproduces sharp and multi-peak spectra that MEM smooths out, improving loss by roughly 20% on the tested synthetic families.
The learned latent features provide a compact description of the spectrum; for single-peak Gaussians they recover exactly the mean and variance.
The SVD sensitivity Sj offers a quantitative robustness metric that can be reported alongside prediction loss for any trained continuation network.
The findings imply a tradeoff: raising hidden dimension improves accuracy but increases noise sensitivity, so model selection should balance both.
The method extends to Lorentzian spectra, where it also achieves the lowest loss among the methods compared.

Reading between the lines

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

The latent-feature factorization can be read as learning a data-driven default model, analogous to MEM's prior but fitted from the training set; a natural test is whether its advantage persists on spectra with line shapes absent from training.
The SVD sensitivity analysis suggests a practical regularizer: truncating or penalizing high-Sj modes of the Jacobian could improve robustness without retraining, though the paper does not propose this.
The paper's synthetic benchmarks use noiseless Green's functions; whether the 20% advantage holds for noisy quantum Monte Carlo data with unknown inverse temperature is an open empirical question the paper leaves implicit.
The architecture's principle, learn features on the output distribution and then map inputs to those features, could transfer to other ill-posed inverse problems beyond analytic continuation.

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.

Desk Editor's Note

read the letter

The paper is worth a serious look. The core idea—learn a low-dimensional latent feature space from spectral functions, then map G to h and h to A—is simple and effective. On synthetic Gaussian and Lorentzian benchmarks it beats MEM, a fully connected network, and the Fournier network by a clear margin, and the loss curves in Fig. 4(a) support the main claim. The Jacobian check for the single-peak case is a nice touch, showing that the learned latent variables are locally equivalent to μ and σ. Code is public, which helps reproducibility. This is a legitimate step beyond the prior neural-network work it cites, and the subfield could plausibly adopt it if the results hold up on real QMC data.

The soft spots are real but mostly fixable. First, all benchmarks are synthetic and noiseless; the inverse temperature β is never specified, so the Matsubara grid is underdefined. That matters because noise is the whole point of analytic continuation. The abstract's "at least 20%" is also a bit loose—the figure shows roughly 20% against the Fournier network, not necessarily against every baseline. Second, the robustness analysis is the weakest part. The SVD sensitivity Sj in Eq. (15) depends on two assumptions in Appendix D: that the left singular vectors are approximately orthogonal under the 1/Ai weighting, and that the noise coefficients cj are equal across modes. Neither is justified numerically. The stress-test note lands: for multi-peak spectra with small Ai between peaks, the weighted cross terms can be large, so Eq. D5 may not approximate Eq. D1. The paper never checks the predicted sensitivity ordering against actual injected noise in G. Until that test is done, the claim that higher hidden dimension reduces robustness is not established.

That said, the primary accuracy claim does not rest on the robustness analysis, and it looks solid on the synthetic benchmarks. The paper would benefit from adding noisy-input experiments, stating β explicitly, and either validating the SVD metric or softening the robustness conclusion. For a serious referee, I'd send it out. It's a competent contribution to an important practical problem, and the weaknesses are addressable in revision rather than fatal. I would bring it to a reading group if your group works on analytic continuation or ML for quantum many-body problems.

Referee Report

5 major / 5 minor

Summary. The paper proposes FL-net, a neural network for analytic continuation that first maps the imaginary-time Green's function G to a low-dimensional latent feature h learned by an autoencoder on spectral functions, and then decodes h to the spectral function A. The authors report that FL-net achieves lower normalized prediction loss than MEM, Fournier's network, D-net, and a fully connected network on synthetic multi-peak Gaussian and Lorentzian spectra, with the abstract claiming an improvement of at least 20%. They also introduce an SVD-based sensitivity metric S_j to argue that increasing the hidden dimension n_h reduces test loss but increases noise sensitivity. The manuscript includes implementation details, comparisons with baselines, and a code repository.

Significance. If the accuracy and robustness claims hold, FL-net would be a useful addition to the toolkit for analytic continuation, with the feature-learning idea being intuitive and the SVD-based sensitivity analysis a novel attempt to quantify robustness in a problem where noise sensitivity is central. The paper provides a reproducible code repository and compares against several established baselines, including MEM and the Fournier network. However, the current evidence is limited to noiseless synthetic spectra drawn from the same family as the training data, the inverse temperature is never specified, and the robustness analysis rests on unvalidated approximations, so the significance is conditional on substantial additional validation.

major comments (5)

[Section II and Section III, Eq. (1)] The inverse temperature β is never specified for the synthetic datasets, although the Matsubara frequencies ω_n=(2n+1)π/β and the kernel in Eq. (1) depend on it. Without β the benchmark is not reproducible, and it is unknown whether the reported gains persist at physically relevant temperatures. Please state the value of β used and, ideally, test at least one other β to show the method is not tuned to a particular temperature.
[Abstract and Section III, Fig. 4] The abstract's claim of 'an improvement of at least 20%' is not quantitatively supported by the reported evidence. Fig. 4(a) shows approximately 20% loss reduction relative to the Fournier network, but with no error bars, no multiple-seed statistics, and no systematic numeric table for the Gaussian datasets; the only numeric table (Table I) is for Lorentzian spectra. Please provide mean and standard deviation over retraining seeds for MEM, FC, Fournier, D-net, and FL-net on each dataset and state the exact margins.
[Section III and Table I] The benchmarks use only noiseless Green's functions generated from the same synthetic families used for training, yet the motivation of the paper is analytic continuation of noisy quantum Monte Carlo data. The claimed superiority over MEM and previous networks is therefore demonstrated only in an idealized setting. Please add experiments with injected noise on G at inference and, if possible, out-of-distribution spectral shapes, so that the robustness discussion is connected to actual noisy inputs.
[Section IV and Appendix D, Eqs. (D1)-(D7)] The robustness analysis is built on two unvalidated assumptions: that the left singular vectors u_j are approximately orthogonal under the 1/A_i weighting so that cross terms vanish in Eq. (D5), and that the projection coefficients c_j are equal across modes. These assumptions are not checked numerically, and no experiment injects Gaussian noise into G to compare the actual KL divergence with the S_j ordering. Please quantify the weighted inner products Σ_i u_{ij} u_{ik}/A_i for the trained matrices and test whether the S_j ranking predicts the measured error degradation under realistic noise.
[Section IV, Fig. 6] The claim that increasing n_h lowers loss but decreases robustness is based on averaged S_j values with no error bars and no statistical test. Fig. 6 plots curves for different n_h but does not report variability across the dataset or across retraining seeds, and the averaging procedure over modes and samples is described only qualitatively. Please provide error bars and a clear scalar summary, such as mean or maximum S_j with uncertainty, to make the trade-off quantitative.

minor comments (5)

[Table I] The entry 'Dent' appears to be a typo for 'D-net', and the table caption says 'Table reftab', which should be resolved to a proper reference.
[Reference [27]] The citation '[27] Appendix, see appendix...' is unhelpful as a reference; please point to the specific appendix sections where the encoder/decoder architectures, dataset generation, and sensitivity analysis are described.
[Section II, Eq. (14)] Equation (14) calls D_j the logarithm of the Fisher information along the u_j direction, but Fisher information is not defined in the manuscript; please add a definition or rephrase to avoid overloading the term.
[Figure 5] Some panels in Fig. 5 lack axis labels or units; in particular, the horizontal axis in panel (b) is labeled only as 'mode range', which is not self-explanatory. Please add clear labels and a description in the caption.
[Section VI] The acknowledgment section writes 'ACKNOWLEDGEMENT' and 'Grant Nos.' in a nonstandard way; please make the wording consistent with the journal style.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the accuracy and robustness claims are empirical benchmarks and a Jacobian-based sensitivity analysis, not derivations that reduce to their inputs.

full rationale

The central accuracy claim (FL-net achieves at least 20% lower loss) is established by held-out test loss comparisons against external baselines (MEM, Fournier's net, D-net, FC net) on synthetic Gaussian and Lorentzian data, with loss measured directly (Fig. 4 and Table I). This is supervised learning evaluated on a test set, not a quantity fitted and then renamed a prediction. The robustness claim in Section IV is an approximate analytical sensitivity measure S_j derived from the trained network's Jacobian via SVD, with the simplifications explicitly stated in Appendix D (neglect of cross terms in Eq. D5 and equal coefficients c_j); these are stated assumptions whose validity is an empirical question, not a circular reduction. Reference [23] is a self-citation by a co-author, but it is cited only as general background that neural networks can outperform MEM and is not load-bearing for any FL-net result. No equation in the paper is defined in terms of the quantity it is used to predict, and no fitted parameter is presented as a prediction. Therefore there is no significant circularity.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The paper's central claim rests mainly on a synthetic benchmark and on assumptions in the robustness analysis. No new physical entities are introduced; the listed free parameters are network and data-generation choices rather than fitted physics constants.

free parameters (2)

hidden dimension nh = 256
Chosen after scanning test loss in Fig. 3; the authors report lower loss but worse robustness at larger values.
inverse temperature beta = not stated
Defines the Matsubara grid in Eq. (1) and Appendix B; without a value, the training data cannot be exactly reproduced.

assumptions (3)

domain assumption The Fredholm kernel G(iωn) = ∫ A(Ω)/(iωn-Ω)dΩ correctly describes the data generation for all test cases.
All synthetic Green's functions are produced from Eq. 1; if the physical kernel differs, for example due to finite bandwidth or a different Matsubara grid, the benchmark may not represent real QMC inputs.
ad hoc to paper Spectral functions of interest are drawn from the same family as the training spectra: sums of Gaussian peaks (Eq. 8) and Lorentzian peaks (Eq. C1) with uniform parameters.
No real experimental or QMC spectra are used; generalization beyond this synthetic family is assumed but not demonstrated.
ad hoc to paper In the robustness analysis, left singular vectors u_j are approximately orthogonal under the 1/A_i weighting, so DKL cross terms can be dropped, and the projection coefficients c_j are equal.
Appendix D, Eqs. D4-D6; these unvalidated assumptions define the Sj metric that underlies the claim that higher hidden dimension reduces robustness.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Analytic Continuation by Feature Learning." pith.science (2026). https://pith.science/paper/5LWQ3MXQ

@misc{pith2026241117728,
  author       = {Pith},
  title        = {Pith review of: Analytic Continuation by Feature Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5LWQ3MXQ}},
  note         = {Machine review of arXiv:2411.17728}
}

abstract

Analytic continuation aims to reconstruct real-time spectral functions from imaginary-time Green's functions; however, this process is notoriously ill-posed and challenging to solve. We propose a novel neural network architecture, named the Feature Learning Network (FL-net), to enhance the prediction accuracy of spectral functions, achieving an improvement of at least $20\%$ over traditional methods, such as the Maximum Entropy Method (MEM), and previous neural network approaches. Furthermore, we develop an analytical method to evaluate the robustness of the proposed network. Using this method, we demonstrate that increasing the hidden dimensionality of FL-net, while leading to lower loss, results in decreased robustness. Overall, our model provides valuable insights into effectively addressing the complex challenges associated with analytic continuation.

Figures

Figures reproduced from arXiv: 2411.17728 by the authors.

**Figure 2.** FIG. 2: FL-net: The network structure consists of two en [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 4.** FIG. 4: (a) Relative loss comparison across datasets with varying numbers of peaks for different networks: fully connected [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗

Figures from the paper (6 more)

**Figure 3.** Figure 3: FIG. 3: The red curve corresponds to the rapid decrease in [PITH_FULL_IMAGE:figures/full_fig_p003_3.png]

**Figure 5.** Figure 5: FIG. 5: (a) Prediction and singular value: The upper plot shows the true spectral results in gray and the predicted results [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]

**Figure 6.** Figure 6: FIG. 6: The red, yellow, green, and blue lines represent the [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

**Figure 8.** Figure 8: FIG. 8: Learning process of [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]

**Figure 7.** Figure 7: FIG. 7: Left: Encoder; Right: Decoder. The detailed struc [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

**Figure 9.** Figure 9: FIG. 9: The gray line represents the true spectrum, while the [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 28 canonical work pages

[1]

E. Gull, A. J. Millis, A. I. Lichtenstein, A. N. Rubtsov, M. Troyer, and P. Werner, Reviews of Modern Physics 83, 349 (2010)

work page 2010
[2]

Georges, G

A. Georges, G. Kotliar, W. Krauth, and M. J. Rozenberg, Reviews of Modern Physics 68, 13 (1996)

1996
[3]

Jarrell and J

M. Jarrell and J. Gubernatis, Physics Reports 269, 133 (1996)

1996
[4]

Schollwoeck, Annals of Physics 326, 96 (2010)

U. Schollwoeck, Annals of Physics 326, 96 (2010)

work page 2010
[5]

J. E. Gubernatis, M. Jarrell, R. N. Silver, and D. S. Sivia, Phys. Rev. B 44, 6011 (1991)

1991
[6]

G. D. Mahan, Many-Particle Physics , 3rd ed., Physics of Solids and Liquids (Springer New York, NY, 2000)

work page 2000
[7]

S. I. Kabanikhin, Inverse and Ill-posed Problems (De Gruyter, Berlin, Boston, 2011)

work page 2011
[8]

R. A. Willoughby, SIAM Rev. 21, 266–267 (1979)

work page 1979

Show all 33 references

[9]

R. K. Bryan, European Biophysics Journal 18, 165 (1990)

1990
[10]

Fei, C.-N

J. Fei, C.-N. Yeh, and E. Gull, Phys. Rev. Lett. 126, 056402 (2021)

2021
[11]

Kiss, Phys

A. Kiss, Phys. Rev. B 100, 214417 (2019)

2019
[12]

H. J. Vidberg and J. W. Serene, Journal of Low Temper- ature Physics 29, 179 (1977)

1977
[13]

A. Weh, J. Otsuki, H. Schnait, H. G. Evertz, U. Eckern, A. I. Lichtenstein, and L. Chioncel, Phys. Rev. Res. 2, 043263 (2020)

2020
[14]

Gunnarsson, M

O. Gunnarsson, M. W. Haverkort, and G. Sangiovanni, Phys. Rev. B 82, 165125 (2010)

2010
[15]

P. C. Hansen, SIAM Rev. 34, 561 (1992)

1992
[16]

Ghanem and E

K. Ghanem and E. Koch, Phys. Rev. B 107, 085129 (2023)

2023
[17]

Skilling, Classic maximum entropy, in Maximum En- tropy and Bayesian Methods: Cambridge, England, 1988 , edited by J

J. Skilling, Classic maximum entropy, in Maximum En- tropy and Bayesian Methods: Cambridge, England, 1988 , edited by J. Skilling (Springer Netherlands, Dordrecht,

1988
[18]

Bergeron and A.-M

D. Bergeron and A.-M. S. Tremblay, Phys. Rev. E 94, 023303 (2016)

2016
[19]

Ghanem and E

K. Ghanem and E. Koch, Phys. Rev. B 108, L201107 (2023)

2023
[20]

Rumetshofer, D

M. Rumetshofer, D. Bauernfeind, and W. von der Linden, Phys. Rev. B 100, 075137 (2019)

2019
[21]

Reymbaut, D

A. Reymbaut, D. Bergeron, and A.-M. S. Tremblay, Phys. Rev. B 92, 060509 (2015)

2015
[22]

Fournier, L

R. Fournier, L. Wang, O. V. Yazyev, and Q. Wu, Phys. Rev. Lett. 124, 056401 (2020)

2020
[23]

J. Yao, C. Wang, Z. Yao, and H. Zhai, Machine Learning: Science and Technology 3, 025010 (2022)

2022
[24]

Huang and Y.-f

D. Huang and Y.-f. Yang, Phys. Rev. B 105, 075112 (2022)

2022
[25]

Hendry and A

D. Hendry and A. E. Feiguin, Phys. Rev. B 100, 245123 (2019)

2019
[26]

A. S. Rakin, Z. He, L. Yang, Y. Wang, L. Wang, and D. Fan, ArXiv abs/1905.13074 (2019)

2019 arXiv
[27]

Appendix, see appendix for detailed information on en- coder/decoder architectures, training dataset generation, model comparisons, FL-net performance on Lorentzian spectra, and noise sensitivity analysis
[28]

D. P. Kingma and J. Ba, CoRR abs/1412.6980 (2014)

2014 arXiv
[29]

Glorot, A

X. Glorot, A. Bordes, and Y. Bengio, in Interna- tional Conference on Artificial Intelligence and Statistics (2011)

2011
[30]

Vaswani, N

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, in Proceedings of the 31st International Conference on Neural Information Processing Systems , NIPS’17 (Cur- ran Associates Inc., Red Hook, NY, USA, 2017) p. 6000–6010

2017
[31]

Jiang, X

W. Jiang, X. Li, H. Hu, Q. Lu, and B. Liu, IEEE Access 9, 69700 (2021)

2021
[32]

Clevert, T

D.-A. Clevert, T. Unterthiner, and S. Hochreiter, arXiv: Learning (2015)

2015
[33]

GitHub repository: https://github.com/Order-inz/ Analytic-Continuation-by-Feature-Learning

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [1]

E. Gull, A. J. Millis, A. I. Lichtenstein, A. N. Rubtsov, M. Troyer, and P. Werner, Reviews of Modern Physics 83, 349 (2010)

work page 2010

[2] [2]

Georges, G

A. Georges, G. Kotliar, W. Krauth, and M. J. Rozenberg, Reviews of Modern Physics 68, 13 (1996)

1996

[3] [3]

Jarrell and J

M. Jarrell and J. Gubernatis, Physics Reports 269, 133 (1996)

1996

[4] [4]

Schollwoeck, Annals of Physics 326, 96 (2010)

U. Schollwoeck, Annals of Physics 326, 96 (2010)

work page 2010

[5] [5]

J. E. Gubernatis, M. Jarrell, R. N. Silver, and D. S. Sivia, Phys. Rev. B 44, 6011 (1991)

1991

[6] [6]

G. D. Mahan, Many-Particle Physics , 3rd ed., Physics of Solids and Liquids (Springer New York, NY, 2000)

work page 2000

[7] [7]

S. I. Kabanikhin, Inverse and Ill-posed Problems (De Gruyter, Berlin, Boston, 2011)

work page 2011

[8] [8]

R. A. Willoughby, SIAM Rev. 21, 266–267 (1979)

work page 1979