Pith. sign in

REVIEW 4 major objections 6 minor 85 references

Spectral distribution of sparse Gaussian Ensembles of Real Asymmetric Matrices

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Average spectra of sparse real asymmetric matrices are fixed by a single complexity parameter $Y$, so ensembles with different sparsity structures share identical spectral densities whenever $Y$ matches, with real Ginibre statistics as…

desk verdict Useful framework extension, but the central evolution equation is derived under constraints that force the matrix to be normal, which excludes the target sparse real-asymmetric ensembles. read the letter →

arxiv 2507.21002 v2 pith:6UIVQZ3Y submitted 2025-07-28 cond-mat.dis-nn cond-mat.stat-mechmath-phmath.MPquant-ph

Ratul Dutta , Pragya Shukla This is my paper · ORCID
classification cond-mat.dis-nncond-mat.stat-mechmath-phmath.MPquant-ph MSC 60B2015B52 PACS 05.40.-a
keywords randommatrixtheoryrealasymmetricmatricescomplexityparameterspectraldensityGinibreensemblesparseensemblesnon-Hermitianstatisticsuniversality
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

The paper aims to show that a very large family of random real asymmetric matrices — including sparse and banded ones whose entries have wildly different sizes — share one universal spectral story: their average eigenvalue density is governed by a single 'complexity parameter' $Y$ assembled from all the entry variances, not by the detailed matrix structure. From a multiparametric Gaussian ensemble density the authors derive diffusion-type evolution equations for the density of real eigenvalues and of complex-conjugate pairs, in which $Y-Y_0$ is the only evolution time. The steady state reached as $Y\to\infty$ reproduces the real Ginibre ensemble — uniform density inside a radius $\sqrt{N}$, exponential decay outside — and the Poisson-to-Ginibre crossover is predicted to be identical in shape for ensembles with different sparsity profiles but equal $Y-Y_0$. This matters because synaptic matrices and neural-network weight matrices are real, asymmetric, and sparse, so a single-parameter description would let one predict their spectral stability without knowing every entry distribution.

What carries the argument

The load-bearing object is the complexity-parameter mapping of eq. (7): $Y=t_1=(2/M)\sum_{m,n}q_{mn;1}\ln(y_{mn}/|\gamma-2y_{mn}|)+c_0$, a single functional of all entry variances chosen so that $\partial\rho_1/\partial Y = L\rho_1$, where $L=\sum_{k,l}\partial^2/\partial H_{kl}^2 + \gamma\sum_{k,l}(\partial/\partial H_{kl})H_{kl}$ is the diffusion generator in matrix space. This reduction collapses the many-parameter ensemble dynamics into a one-parameter evolution, and carrying it through the eigenvalue joint density yields the evolution equations (17) for the real-eigenvalue density and (33)–(36) for the complex-pair density. The real-axis equation reduces to a Dyson–Pastur form once the repulsion integral $s_1(e)/e$ is kept only in leading order, giving Gaussian then semicircle solutions; the complex-plane equation separates in $r$ and $\theta$ and, after dropping the real–complex repulsion terms $J_c$ and $J_d$, adopts the complex-matrix solutions of [68] built from confluent hypergeometric functions, with uniform limit inside $\sqrt{N}$.

What would settle it

Measure the ensemble-averaged complex-plane density near the real axis for a small real-asymmetric sparse ensemble ($N\sim 10$–$20$, where the fraction of real eigenvalues is largest) and compare it with the complex-matrix solution of [68] that the paper adopts; a density excess near the real axis growing with the fraction of real eigenvalues would show the dropped terms $J_c$ and $J_d$ are not negligible. A second check: build two ensembles with identical $Y-Y_0$ but different variance profiles and compare their two-point correlation $R_{2z}$ — the paper's claim fixes only one-point densities, so any mismatch delineates the extent of the asserted universality.

Watch

Extended reading notes

Core claim

The paper's central claim is that the spectral joint probability density of a multiparametric Gaussian ensemble of real asymmetric matrices, with ensemble density $\rho(H;y,x)=\exp[-\sum_{k,l}(y_{kl}H_{kl}^{2}+x_{kl}H_{kl}H_{lk})]$, evolves in the complexity parameter $Y$ through $\partial P_1/\partial Y = (L_e + L_z)P_1$ (eq. (10)), so that the ensemble-averaged densities of real eigenvalues and of complex-conjugate pairs depend on the ensemble only through $Y-Y_0$. In the equilibrium limit the complex-plane density becomes constant inside a circle of radius $\sqrt{N}$, decays exponentially beyond it, and is independent of angle — the real Ginibre statistics — while the projected real-axis density runs from a Gaussian (small $Y-Y_0$) to a semicircle (large $Y-Y_0$). Numerical diagonalization of three sparse ensembles with different variance profiles (constant, power-law, and exponential off-diagonal decay) shows the density curves collapse when plotted against $Y$, confirming single-parameter control of the spectrum.

Load-bearing premise

The complex-plane density rests on three approximations the paper states explicitly: the real–complex repulsion terms $J_c$ and $J_d$ are dropped, pair correlations are factorized as products of one-point densities, and the complex density is taken as nearly angle-independent; if any of these fails, the predicted complex-plane density is the complex-matrix one rather than the real-asymmetric one.

Editorial extensions

If this is right

  • For any Gaussian real-asymmetric ensemble — sparse, banded, or full — the average real and complex spectral densities are fixed by the single scalar $Y-Y_0$, so computing $Y$ from the entry variances predicts the spectrum without large-scale diagonalization.
  • The real-eigenvalue density crosses from a Gaussian to a semicircle as $Y-Y_0$ grows, while the complex-pair density flattens to uniformity inside radius $\sqrt{N}$ with exponential decay outside, recovering the real Ginibre steady state in the $Y\to\infty$ limit.
  • Ensembles with equal $Y-Y_0$ and equal initial statistics share identical spectral densities even away from equilibrium, so the Poisson-to-Ginibre crossover has a universal shape across sparsity structures, not just a universal endpoint.
  • The same complexity-parameter evolution extends to correlated, non-zero-mean ensembles of the form (1), as the numerical multivariate case indicates, although the explicit $Y$ for that case is not yet derived.

Reading between the lines

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

  • The universality established here concerns one-point densities; whether two-point correlations (the level-repulsion profile) also collapse under $Y$ is untested, and a mismatch there would bound the universality claim rather than break it.
  • Because the complex-plane equation is inherited from the complex-matrix case by dropping the real–complex repulsion, the regime where real eigenvalues are abundant — small $N$ or strongly banded matrices — is the natural place to look for deviations, and the paper's numerics do not resolve that regime.
  • For neural-network practice the paper suggests a cheap diagnostic it does not itself propose: track the complexity parameter $Y$ of the weight matrix during training, since the spectral edge (and hence stability of the silent state) is governed by it.
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 / 6 minor

Summary. The paper claims to extend the complexity-parameter (Y) framework to sparse Gaussian ensembles of real asymmetric matrices. It derives (with details deferred to supplemental material) an evolution equation for the spectral joint probability density function, solves approximate equations for the real-eigenvalue density and the complex-eigenvalue density, and confronts these with exact diagonalization of four ensembles (BE, PE, EE, ME). The authors conclude that the spectral statistics depend on the ensemble only through Y, revealing a universality across sparse structures.

Significance. The target class of sparse real asymmetric matrices is highly relevant to neuroscience and machine learning, and a one-parameter description of their spectral densities would be a valuable theoretical tool. The paper's numerical data (Fig. 1) show a suggestive collapse of the real-to-complex eigenvalue ratio as a function of Y. However, the central derivation is not verifiable from the main text, the key eigenvector conditions are incompatible with non-normal matrices, and the quantitative comparisons rely on per-ensemble fits. As a result, the paper does not provide a trustworthy derivation or a falsifiable predictive test. The absence of machine-checked proofs or reproducible code further limits the assessment to the analytic arguments and the reported fits.

major comments (4)
  1. [Section II.B, Eqs. (10)-(12)] The derivation of the spectral JPDF uses U V = I together with U U^T = I and V V^T = I. For E = U H V with U V = I, these conditions force V = U^T and H = U^T E U, hence H H^T = H^T H, i.e., H is normal. The target ensembles BE, PE, EE, and ME consist of generic non-normal real matrices, so this derivation cannot be applied to them. The paper's own remarks in Section II.B, noting that different eigenvector conditions in [46] and [68] lead to different eigenvalue dynamics, indicate that the normalization is consequential rather than a harmless gauge choice. Therefore Eq. (10) is not established for the class of matrices under study, and the solutions in Sections III and IV inherit this gap.
  2. [Section VI, Eqs. (43)-(44) and Table I] The 'agreement with theory' is largely a fitting exercise. The initial density R1(r; Y0) is obtained by numerical fit, the coefficients q_1nu and chi in Eqs. (43)-(44) are determined by numerical fits for each ensemble and b value, and the real-line densities are fitted with Gaussian parameters C1 and q that vary with b. No cross-prediction is made: ensembles at the same (Y-Y0) are not compared with one another using common parameters. The collapse in Fig. 1 concerns a single scalar ratio, not the full spectral density. Thus the numerical results provide only weak, non-falsifiable support for the complexity-parameter universality claim.
  3. [Section IV.B, after Eq. (42)] The neglect of the real-complex repulsion terms J_c and J_d is asserted without quantitative justification; the statement that they are 'relatively smaller' than I_c and I_d is not demonstrated. For small b, a substantial fraction of eigenvalues are real (Fig. 1), so the real-complex interaction may be non-negligible. Since the complex-plane solution is then taken verbatim from the complex-matrix case [68], the derived R1z is not shown to be the density of the real-asymmetric ensemble. The weak theta-dependence of R1z is also inferred from the same numerical data that are subsequently fitted, making the argument circular.
  4. [Sections II.B and IV.B; Supplemental Material [83]] The central derivation of Eq. (10) and the detailed solution of the complex-plane equation are deferred to supplemental material [83], which is not included for review. Given that the validity of Eq. (10) is contested above, the omission of this material prevents verification of the load-bearing steps, including the response of eigenvalues and eigenvectors to matrix perturbations. For a paper whose quantitative predictions rest on this equation, the main text should contain enough detail for the derivation to be checked.
minor comments (6)
  1. [Section V] The paragraph beginning 'Based on the analysis discussed in previous section...' is an exact duplicate of the preceding paragraph and should be removed.
  2. [Tables] Both the real-line fit table and the complex-plane fit table are labeled 'TABLE I'; they should be renumbered as Tables I and II, and the in-text references updated accordingly.
  3. [Throughout] There are several typos: 'exhibitory' should be 'excitatory' (Abstract), 'assymetric' appears in Section VII, and reference [73] is garbled ('Paslur L A 1972 Th. Math. Phyx. IO 67' should be 'Pastur L A 1972 Theor. Math. Phys. 10 67').
  4. [Fig. 5 caption] The caption says 'The other details of the figure are same as in figure 2' but also states the figure is for EE; the figure is actually for the multivariate (ME) case, so the caption should be corrected.
  5. [Table I (complex-plane fits)] The fit labels 'fita1, fitb1, fitn' are not defined in the main text; the functional forms should be defined or explicitly referenced to the supplemental material.
  6. [Introduction / Notation] The notation warning that subscripts change meaning is helpful, but the paper would benefit from a consolidated notation table to reduce ambiguity in equations such as (7), (21), and (43).

Circularity Check

3 steps flagged · score 6.0 of 10

Spectral-density 'predictions' reduce to fitted initial densities and fitted mode coefficients; the complex-plane result is also imported from same-author prior work.

  1. fitted input called prediction [Section VI, complex-plane analysis, around Eqs. (43)-(44) and Table I]
    "As our formulation is based on a prior knowledge of initial density R 1(r, θ;Y0), we determine it by a numerical fit. The latter gives, for the initial ensemble at b=b 0 = 1/N, R 1(r, θ;Y0)≡R 1(r;Y 0) =A r −1/2 J1/2(Br) e −Cr 2 with A, B, Cas constants (the values for each ensemble given in table I). ... with only first few terms (withν≤2) retained in the P ν and the constantsq 1ν ofU 1ν determined by numerical fits; the fitting parameters in each case are given in table I."

    Equation (43) is an eigenmode expansion of the initial density; the coefficients q_{1ν} are not derived from Y0 or from the complexity parameter, but fitted to the same radial density curves that the equation is claimed to predict. Since the initial A, B, C are also fitted at b=1/N and q_{1ν} are refitted for each b, the 'agreement' between eq. (43) and Figs. 2-4 is a multi-parameter curve fit, not an independent prediction of the complexity-parameter evolution.

  2. fitted input called prediction [Section III.B and Section VI, real-line fits; Eqs. (24)-(27) and Table I]
    "As an example, we determine R 1(e, Y) from eq.(24) for a Gaussian initial level-density R1(e;Y 0) = 1√ 2πσ2 e− x2 2σ2 ; (this is the case analyzed later in our numerical analysis too discussed in section VI). ... the numerically fitted functions; the latter in each case agrees with our theoretical prediction i.e a Gaussian for small Y−Y 0 and a semicircle for large Y−Y 0."

    The Gaussian shape is inserted as the assumed initial condition, and Table I Fit 1 refits C1 and q at every b value (e.g. C1=0.317, q=0.315 at b=1/N; different values at 1.5, 2.5, 15 and N). The theoretical solution (27) is not used to predict those parameters from Y−Y0; instead the same Gaussian functional form is fitted to each dataset and then labelled as agreement with theory. This is the fitted initial condition renamed as a prediction.

1 more flagged steps
  1. other [Section III.B, after Eq. (20), truncation of I_ez]
    "Based on insights given by our numerical analysis for three prototypical cases of the ensemble (2) (discussed later in section VII), R 1z(r, θ) does not vary significantly with θ. This along with cos(mθ) present in both s m and l m implies a faster oscillation of integrand as m increases."

    The weak θ-dependence of R1z used to truncate the series to m=1 is inferred from the same numerical data that later serve as the target of the theoretical fit. The simplified evolution equation (22) is therefore calibrated on the data it is then said to predict. This does not make the full derivation circular, but it weakens the numerical confirmation of the complexity-parameter universality.

full rationale

The paper does derive an evolution equation (10) for the spectral JPDF from the ensemble density, so the central formalism is not circular by construction. However, the numerical validation presented as 'agreement with our theoretical prediction' is substantially circular: the complex-plane density is expanded in modes whose coefficients q_{1ν} are fitted to the data, and the initial density is also obtained by a numerical fit; the real-line Gaussian shape is inserted as the initial condition and then refitted at every b. The weak θ-dependence that justifies truncating the series is itself inferred from the same data. These are fitted inputs called predictions. We do not count the reliance on [68] as circular because it is a published prior derivation by the same group, though it means the complex-plane solution is not independently re-derived here. Separately, the orthogonality condition U U^T=I, V V^T=I imposed in Section II.B would force normal matrices, which is a serious correctness concern for non-normal sparse real-asymmetric ensembles, but it is not a circularity of the type scored here. Overall, the core equations are independent, but the paper's claimed numerical confirmations of the spectral-density predictions reduce in part to curve fitting, giving a score of 6.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The central claims rest on the complexity parameter framework developed in prior work by the same group, on the neglect of the real-complex repulsion for the complex-plane density, and on the mean-field factorization of two-point correlations. The numerical evidence for the universality claim requires fitting the initial density and several coefficients to the same data.

free parameters (5)
  • gamma = 1
    Arbitrary parameter in the generator T, set to 1 to reach the Ginibre steady state.
  • Initial complex density constants A, B, C = A=1.06, B=0.73, C=0.294 for BE (Table I)
    The initial density R1(r;Y0)=A r^{-1/2} J_{1/2}(B r) e^{-C r^2} is fitted to numerical data at b0=1/N.
  • Fitting coefficients q_1nu = For BE: q_10=1.151, q_11=-4.333, q_12=108.132 (Table I)
    The coefficients in the theoretical expressions (43) and (44) are determined by numerical fits to the same data.
  • chi = 0.088
    The decay rate chi in the complex-plane solution is fixed at 0.088 for all fitted cases.
  • Real-density Gaussian parameters C1, q = For BE at b=1/N: C1=0.317, q=0.315 (Table I)
    The predicted Gaussian form f1 is fitted to the real eigenvalue density for each b value separately.
assumptions (5)
  • domain assumption The ensemble density in eq. (2) admits a complexity parameter mapping in which the evolution is governed by a single parameter t1 = Y
    Inherited from [46], [68], [69]; the present paper reviews it and applies it without re-deriving the mapping in the main text.
  • domain assumption The real asymmetric matrix is diagonalizable with U U^T = I and V V^T = I, with no exceptional points
    Stated in Section II.B as the basis for the JPDF diffusion equation.
  • domain assumption Two-point spectral correlations factorize, R2 ~ R1 R1, for distances beyond the mean level spacing
    Used in Section III.B to evaluate Iez and in Section IV.B for the complex-plane integrals.
  • ad hoc to paper R1z(r,theta) has weak theta-dependence, so the m >= 1 terms in eq. (21) are negligible
    Justified by the authors' numerical observations for three ensembles, using the same data later fitted to the derived formulas.
  • ad hoc to paper The real-complex eigenvalue repulsion terms J_c and J_d are negligible compared with I_c and I_d
    Stated in Section IV.B without a quantitative bound; this reduces the real-asymmetric equation to the complex case [68].
invented entities (1)
  • Complexity parameter Y and complexity constants t_alpha
    purpose: Single reparametrization of the ensemble variances claimed to govern spectral statistics
    The parameter is constructed from the ensemble variances; the only supporting evidence is the paper's own collapse of one observable (L/M) and fits of spectral densities with the same data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Spectral distribution of sparse Gaussian Ensembles of Real Asymmetric Matrices." pith.science (2026). https://pith.science/paper/6UIVQZ3Y

@misc{pith2026250721002,
  author       = {Pith},
  title        = {Pith review of: Spectral distribution of sparse Gaussian Ensembles of Real Asymmetric Matrices},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6UIVQZ3Y}},
  note         = {Machine review of arXiv:2507.21002}
}
read the original abstract

Theoretical analysis of biological and artificial neural networks e.g. modelling of synaptic or weight matrices necessitate consideration of the generic real-asymmetric matrix ensembles, those with varying order of matrix elements e.g. a sparse structure or a banded structure. We pursue the complexity parameter approach to analyze the spectral statistics of the multiparametric Gaussian ensembles of real asymmetric matrices and derive the ensemble averaged spectral densities for real as well as complex eigenvalues. Considerations of the matrix elements with arbitrary choice of mean and variances render us the freedom to model the desired sparsity in the ensemble. Our formulation provides a common mathematical formulation of the spectral statistics for a wide range of sparse real-asymmetric ensembles and also reveals, thereby, a deep rooted universality among them.

Figures

Figures reproduced from arXiv: 2507.21002 by the authors.

Figure 1
Figure 1. FIG. 1 [PITH_FULL_IMAGE:figures/full_fig_p031_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2 [PITH_FULL_IMAGE:figures/full_fig_p035_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3 [PITH_FULL_IMAGE:figures/full_fig_p036_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: FIG. 4 [PITH_FULL_IMAGE:figures/full_fig_p037_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5 [PITH_FULL_IMAGE:figures/full_fig_p038_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

85 extracted references · 69 canonical work pages

  1. [46]

    J.T.Chalker and Z.J.Wang, Phys. Rev. Lett. 79, 1797, (1997)

    work page 1997

  2. [68]

    however focussed on the spectral density analysis only for the complex matrices. With real random matrices appearing more often among complex systems studies, a derivation of their spectral density is desirable too.With present study focused on the real asymmetric matrices, we briefly review, in sections II.A and II.B, the complexity parameter formulation...

  3. [83]

    Gleit and A

    A. Gleit and A. J. LAZAR, Journal of Functional Analysis 22, 354 (1976)

    work page 1976

  4. [1]

    Crisanti, H

    H.J.Sommers, A. Crisanti, H. Sompolinsky and Y. Stein, Phys. Rev. Lett., 60, 1895 (1988)

    1988

  5. [2]

    For small (Y−Y 0), therefore,⟨R 1(r, θ, Y)⟩θ remains almost of the same form as the initial density except for a change of coefficients

    and also in supplemental material [83], this leads to ⟨R1(r, θ, Y)⟩θ = ∞X ν=0 U1ν e−νϕ ≈ ⟨R1(r, θ, Y0)⟩θ −ϕ ∞X ν=1 ν U1ν (43) 22 withU 10 =⟨R 1(r, θ, Y0)⟩θ andU 1ν ≡q 1ν F1 νχ+ 1, 3 2 ,− γr 2 2 forν >0 (see details in [83]). For small (Y−Y 0), therefore,⟨R 1(r, θ, Y)⟩θ remains almost of the same form as the initial density except for a change of coefficie...

  6. [3]

    Ranjan and L

    K. Ranjan and L. F. Abbott, Phys. Rev. Lett., 97, 188104, (2006)

    2006

  7. [4]

    D.R.Nelson and N.M.Shnerb, 58, 1383, (1998)

    1998

  8. [5]

    Ahmadian, F

    Y. Ahmadian, F. Mumarola and K. D. Miller, Phys. Rev. E 91, 012820, (2015)

    work page 2015

Show all 85 references
  1. [6]

    A. Amir, N. Hatano and D. R. Nelson, Phys. Rev. E 93, 042310, (2016)

    2016

  2. [7]

    Granziol, S

    D. Granziol, S. Zohren and S. Roberts, Learning Rates as a Function of Batch Size: A Random Matrix Theory Approach to Neural Network Training, Journal of Machine Learning Research 23 (2022) 1-65

    2022

  3. [8]

    Pastur, On random matrices arising in deep neural networks: Gaussian case, Pure Appl

    L. Pastur, On random matrices arising in deep neural networks: Gaussian case, Pure Appl. Funct. Anal. (2020), in press, arXiv:2001.06188

    2020 arXiv

  4. [9]

    Pennington and P

    J. Pennington and P. Worah, Nonlinear random matrix theory for deep learning, Journal of Statistical Mechanics: Theory and Experiment, 2019(12):124005, dec 2019. doi: 10.1088/1742- 5468/ab3bc3

    2019 doi

  5. [10]

    Pennington and Y

    J. Pennington and Y. Bahri, Geometry of Neural Network Loss Surfaces via Random Matrix Theory, Proceedings of the 34 th International Conference on Machine Learning, Sydney, Australia, PMLR 70, 2017

    2017

  6. [11]

    Thamm, M

    M. Thamm, M. Staats and B. Rosenow, Random matrix analysis of deep neural network weight matrices, Physical Review E, 106(5):054124, 2022

    2022

  7. [12]

    Ghorbani, S

    B. Ghorbani, S. Krishnan, and Y. Xiao, An investigation into neural net optimization via Hessian eigenvalue density. arXiv preprint arXiv:1901.10159, 2019

    1901 arXiv

  8. [13]

    A. K. Lampinen and S. Ganguli, An analytic theory of generalization dynamics and transfer learning in deep linear networks. preprint arXiv:1809.10374, 2018

    2018 arXiv

  9. [14]

    Louart, Z

    C. Louart, Z. Liao, and R. Couillet. A random matrix approach to neural networks. The Annals of Applied Probability, 28(2):1190 – 1248, 2018. doi: 10.1214/ 17-AAP1328. 27

    2018

  10. [15]

    Pennington, S

    J. Pennington, S. Schoenholz and S. Ganguli, The emergence of spectral universality in deep networks, Proc. Mach. Learn. Res. 84 (2018) 1924–1932, arXiv:1802.09979

    2018 arXiv

  11. [16]

    Seroussi and Z

    I. Seroussi and Z. Ringel. Separation of scales and a thermodynamic description of feature learning in some cnns. preprint arXiv:2112.15383, 2021

    2021 arXiv

  12. [17]

    preprint arXiv:2006.07721, 2020

    Diego Granziol, Beyond random matrix theory for deep networks. preprint arXiv:2006.07721, 2020

    2006 arXiv

  13. [18]

    N. P. Baskerville, D. Granziol, and J. P. Keating. Applicability of Random Matrix Theory in Deep Learning. preprint arXiv:2102.06740

    arXiv

  14. [19]

    Martin, T (Serena) Peng, and M

    C H. Martin, T (Serena) Peng, and M. W. Mahoney. Predicting trends in the quality of state- of-the-art neural networks without access to training or testing data. Nature Communications, 12(1):4122, 2021. doi: 10.1038/s41467-021-24025-8

    2021 doi

  15. [20]

    Implicit self-regularization in deep neural networks: Evi- dence from random matrix theory and implications for learning

    C H Martin and M W Mahoney. Implicit self-regularization in deep neural networks: Evi- dence from random matrix theory and implications for learning. Journal of Machine Learning Research, 22(165):1–73, 2021

    2021

  16. [21]

    Okuma, K

    N. Okuma, K. Kawabata, K. Shiozaki, and M. Sato, Phys.Rev. Lett.124, 086801 (2020)

    2020

  17. [22]

    D. S. Borgnia, A. J. Kruchkov, and R.-J. Slager, , Phys. Rev.Lett.124, 056802 (2020)

    2020

  18. [23]

    M. S. Rudner and L. S. Levitov, Phys. Rev. Lett.102,065703 (2009)

    2009

  19. [24]

    Yao and Z

    S. Yao and Z. Wang, Phys. Rev. Lett.121,086803 (2018)

    2018

  20. [25]

    Esaki, M

    K. Esaki, M. Sato, K. Hasebe, and M. Kohmoto, Phys. Rev. B84, 205128 (2011)

    2011

  21. [26]

    Y. C. Hu and T. L. Hughes, Phys. Rev. B84, 153101 (2011)

    2011

  22. [27]

    Schomerus, Opt

    H. Schomerus, Opt. Lett.38, 1912 (2013)

    2013

  23. [28]

    Z. Gong, Y. Ashida, K. Kawabata, K. Takasan, S. Hi-gashikawa, and M. Ueda, Phys. Rev. X8, 031079 (2018)

    2018

  24. [29]

    N.Moiseyev, Non-Hermitian Quantum Mechanics (Cambridge University Press, 2011)

    2011

  25. [30]

    Ashida, Z

    Y. Ashida, Z. Gong, and M. Ueda, Advances in Physics 69, 249 (2020)

    2020

  26. [31]

    Zabalo, M

    A. Zabalo, M. J. Gullans, J. H. Wilson, R. Vasseur,A. W. W. Ludwig, S. Gopalakrishnan, D. A. Huse, and J. H. Pixley, Phys. Rev.Lett. 128, 050602 (2022)

    2022

  27. [32]

    Skinner, J

    B. Skinner, J. Ruhman, and A. Nahum, Phys. Rev. X9, 031009 (2019)

    2019

  28. [33]

    Feinberg and A

    J. Feinberg and A. Zee, Phys. Rev. E59, 6433 (1999)

    1999

  29. [34]

    A. C. Potter and R. Vasseur, arXiv e-prints, arXiv:2111.08018 (2021), arXiv:2111.08018 28

    2021 arXiv

  30. [35]

    Huang and B

    Y. Huang and B. I. Shklovskii, Phys. Rev. B 101, 014204 (2020)

    2020

  31. [36]

    L. G. Molinari, Journal of Physics A: Mathematical and Theoretical 42, 265204 (2009)

    2009

  32. [37]

    G. L. Celardo, M. Angeli, and R. Kaiser, arXiv preprint arXiv:1702.04506 (2017)

    2017 arXiv

  33. [38]

    Kawabata and S

    K. Kawabata and S. Ryu, Phys. Rev. Lett.126, 166801(2021)

    2021

  34. [39]

    C. E. Maximo, N. A. Moreira, R. Kaiser, and R. Bachelard, Phys. Rev. A100, 063845 (2019)

    2019

  35. [40]

    Cottier, A

    F. Cottier, A. Cipris, R. Bachelard, and R. Kaiser, Phys. Rev. Lett.123, 083401 (2019)

    2019

  36. [41]

    Hatano and D.R.Nelson, Phys

    N. Hatano and D.R.Nelson, Phys. Rev. Lett. 77, 570 (1996); K.B. Efetov, Phys Rev. B 56 9630 (1997); I.Y. Glodsheild and B.A. Khoruzhenko, Phys. Rev. Lett., 80, 2897 (1998); C.Mudry, B.D.Simons and A.Altland, Phys. Rev. Lett., 80, 4257 (1998)

    1996

  37. [42]

    A. Guo, G. J. Salamo, D. Duchesne, R. Morandotti, M. Volatier-Ravat, V. Aimez, G. A. Siviloglou, and D. N.Christodoulides, Phys. Rev. Lett.103,093902 (2009)

    2009

  38. [43]

    M. L. Mehta, Random Matrices (Academic, New York, 1991)

    1991

  39. [44]

    Haake, Quantum Signatures of Chaos (Springer-Verlag, Berlin, 1991)

    F. Haake, Quantum Signatures of Chaos (Springer-Verlag, Berlin, 1991)

    1991

  40. [45]

    J.T.Chalker and B.Mehlig, Phys. Rev. Lett. 81, 3367, (1998)

    1998

  41. [47]

    Ginibre, J

    J. Ginibre, J. Math. Phys. 6 440, (1965)

    1965

  42. [48]

    Shukla, Phys

    P. Shukla, Phys. Rev. Lett. 87, 19, 194102, (2001)

    2001

  43. [49]

    Nils Lehmann and H-j Sommers, Phys. Rev. Lett. 67, 941, (1991)

    1991

  44. [50]

    B, 88, 359, (1992)

    F.Haake et al., Z.Phys. B, 88, 359, (1992)

    1992

  45. [51]

    Bender and S

    C.M. Bender and S. Boettcher, Phys. Rev. Lett. 80, 5243, (1998)

    1998

  46. [52]

    Bender, N

    C.M. Bender, N. Hassanpour, D.W.Hook, S.P.Klevansky, C. Sunderhauf and Z. Wen, Phys. Rev. A 95, (2017)

    2017

  47. [53]

    Mostafazadeh, J

    A. Mostafazadeh, J. Math. Phys. 43, 205, (2002); 43, 2814, (2002); 43, 3944, (2002)

    2002

  48. [54]

    Bender, D.C

    C.M. Bender, D.C. Brody and H.F.Jones, Phys. Rev. Lett. 89, (2002)

    2002

  49. [55]

    Mudute-Ndumbe and M

    E-M Graefe, S. Mudute-Ndumbe and M. Taylor, J. Phys. A: Math. Theo. 48, 38FT02, (2015)

    2015

  50. [56]

    karr, Phys

    Y.N.Joglekar and W.A. karr, Phys. Rev. E, 83, (2011)

    2011

  51. [57]

    Rotter, arXiv:1011.0645, (2010)

    I. Rotter, arXiv:1011.0645, (2010)

    2010 arXiv

  52. [58]

    Rotter, J

    I. Rotter, J. Phys. A: Math. Theo., 42, 153001, (2009)

    2009

  53. [59]

    Sommers, Phys

    Y.V.Fyodorov, B.A.Khoruzhenko and H.-J. Sommers, Phys. Rev. Lett., 79, 557, (1997). 29

    1997

  54. [60]

    Sommers, J

    Y.V.Fyodorov and H.-J. Sommers, J. Math. Physics (N.Y.), 38, 1918, (1997)

    1997

  55. [61]

    Feinberg and A.Zee, Phys

    J. Feinberg and A.Zee, Phys. Rev. E 59, 6433, (1999)

    1999

  56. [62]

    Y.V.Fyodorov and B.A.Khoruzhenko, Ann. Inst. Henri Poincare (Physique Theorique), 68, 449, (1998)

    1998

  57. [63]

    Hamazaki, K

    R. Hamazaki, K. Kawabata, N. Kura and M. Ueda, Phys. Rev. Res., 2, 023286, (2020)

    2020

  58. [64]

    L. Sa, P. Ribciro and T. Prosen, Phys. Rev. X, 10, 021019, (2020)

    2020

  59. [65]

    Bohigas and M

    O. Bohigas and M. P. Pato, J. Phys. A: Math. Theor. 46, 115001, (2013)

    2013

  60. [66]

    Khaymovich, Phys

    G.De Tomasi and I. Khaymovich, Phys. Rev. B, 106, 094204 (2022)

    2022

  61. [67]

    Suthar, Y-C Wang, Y-P Huang, H.H.Jens and J-H You, Phys

    K. Suthar, Y-C Wang, Y-P Huang, H.H.Jens and J-H You, Phys. Rev. B 106, 064208 (2022)

    2022

  62. [69]

    A. M. Garcia-Garicia, S. M. Nishigaki and J.J.M. Verbaarschot, Phys. Rev. E 66, 016132, (2002)

    2002

  63. [70]

    Anasari and P

    M.G. Anasari and P. Shukla, J. Phys. A: Math. Theor. (IOP), 57, 095005, (2024)

    2024

  64. [71]

    Anasari and P

    M.G. Anasari and P. Shukla, J. Phys. A: Math. Theor. (IOP) 57 455001, (2024)

    2024

  65. [72]

    A. M. Mambuca, C. Cammarota and I. Neri, Phys. Rev. E 105, 014305 (2022)

    2022

  66. [73]

    Neri and F

    I. Neri and F. Lucas Metz, Phys. Rev. Research, 2, 033313 (2020)

    2020

  67. [74]

    Shukla, Int

    P. Shukla, Int. Jou. of Mod.Phys B (WSPC), 26, 12300008, (2012)

    2012

  68. [75]

    Paslur L A 1972 Th. Math. Phyx. IO 67

    1972

  69. [76]

    Kumar and A.Pandey, Annals of Physics

    S. Kumar and A.Pandey, Annals of Physics

  70. [77]

    Pandey, Phase Transitions (Taylor and Francis), 77, 835 (2004)

    A. Pandey, Phase Transitions (Taylor and Francis), 77, 835 (2004)

    2004

  71. [78]

    Shukla, Phys

    P. Shukla, Phys. Rev. B, 98, 054206 (2018)

    2018

  72. [79]

    Neri and T

    F.L.Metz, I. Neri and T. Rogers, J. Phys. A: Math. Theor. 52, 434003, (2019)

    2019

  73. [80]

    I. I. Arkhipov and F. Minganti, Phys. Rev. A, 107, 012202 (2023)

    2023

  74. [81]

    A.Pandey and P.Shukla, J. Phys. A, (1991)

    1991

  75. [82]

    Temme and E.J.M

    N.M. Temme and E.J.M. Veling, Indagationes Mathematicae (Elsevier) 33, 1221, (2022)

    2022

  76. [84]

    Megginson,An introduction to Banach Space theory, Springer, N.Y

    R.E. Megginson,An introduction to Banach Space theory, Springer, N.Y. (1998)

    1998

  77. [85]

    Supplemental material. 30 FIG. 1.Ratio of the real eigenvalues and complex conjugate pairs:The figure displays the ratioL/MofLreal eigenvalues andMcomplex conjugate pairs for manybvalues and fixed N= 1024 for three ensembles withN=L+ 2M. As clear from the figure, the number of...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.