pith. sign in

arxiv: 2607.01895 · v1 · pith:EIUZYMRGnew · submitted 2026-07-02 · 💻 cs.LG · math.OC· math.ST· stat.ML· stat.TH

Regularized Variational and Spectral Log-Density-Ratio Estimation in the Gaussian Location Model

Francis Bach (SIERRA) This is my paper

Pith reviewed 2026-07-03 17:43 UTC · model grok-4.3

classification 💻 cs.LG math.OCmath.STstat.MLstat.TH
keywords log-density-ratio estimationGaussian location modelvariational estimatorspectral estimatorridge regularizationhigh-dimensional asymptoticsdensity ratio estimation
0
0 comments X

The pith

In the Gaussian location model, ridge-regularized variational log-density-ratio estimation has lower asymptotic risk than the spectral alternative when observations greatly exceed dimension, but the spectral estimator is preferable with few

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

The paper compares two ridge-regularized estimators for log-density ratios under the Gaussian location model q ~ N(0,I), p ~ N(Δ,I). It derives deterministic high-dimensional equivalents for their population risks as dimension and sample sizes grow with fixed ratios. The variational estimator is obtained by minimizing an empirical KL objective with L2 penalty, while the spectral estimator solves a continuum of ridge least-squares problems. Asymptotics show that the variational form wins for large observation counts because it exploits the model specification more fully, whereas the spectral form has lower variance when data is scarce. Experiments sweep aspect ratios and regularization to confirm the risk crossover.

Core claim

Under the analyzed high-dimensional regimes with fixed observation-to-dimension ratios, the well-specified variational estimator attains smaller population risk than the spectral estimator when the number of observations is large, while the spectral estimator is favored with fewer observations due to its covariance-based construction having lower variance; the limits are characterized via CGMT for the variational case and resolvent equivalents for the spectral case.

What carries the argument

High-dimensional deterministic asymptotic equivalents for the population risks of the two ridge-regularized estimators, obtained respectively from the convex-Gaussian-min-max theorem and from resolvents of weighted sums of two independent Gaussian sample covariances.

If this is right

  • For density-ratio tasks in high dimensions, the variational estimator should be selected when sample size is large relative to dimension.
  • The spectral estimator should be preferred in the small-sample regime to exploit its lower variance.
  • A nuclear-norm penalty can be added to either estimator to induce feature learning, though its analysis remains partial.

Where Pith is reading between the lines

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

  • The risk comparison may extend to other well-specified parametric families beyond the Gaussian location model.
  • The derived asymptotic formulas could be used to optimize the regularization parameter as a function of the aspect ratio without cross-validation.
  • Similar variational-versus-spectral trade-offs may appear in other ratio estimation problems such as importance sampling or mutual-information estimation.

Load-bearing premise

The observations are drawn exactly from the Gaussian location model with identical covariance matrices for the two distributions.

What would settle it

A finite-sample simulation in the Gaussian location model that measures empirical risks for both estimators across a sweep of observation-to-dimension ratios and checks whether the risk ordering reverses at the ratio predicted by the asymptotic formulas.

Figures

Figures reproduced from arXiv: 2607.01895 by Francis Bach (SIERRA).

Figure 1
Figure 1. Figure 1: Fixed-signal first-order expansion of the unregul [PITH_FULL_IMAGE:figures/full_fig_p021_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Second-order separating curves in the weak-signa [PITH_FULL_IMAGE:figures/full_fig_p023_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Regimes of feasibility for unregularized estimat [PITH_FULL_IMAGE:figures/full_fig_p024_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of two estimators for a given [PITH_FULL_IMAGE:figures/full_fig_p024_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of two estimators for a given [PITH_FULL_IMAGE:figures/full_fig_p025_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of estimators, limits, and empirical p [PITH_FULL_IMAGE:figures/full_fig_p027_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of estimators when varying αp (αq = 0.1 fixed, with αp swept over [0, 1]). Left: variational, right: spectral. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of estimators when varying αq (αp = 0.05 fixed, with αq swept over [0, 1], only shown partially for the variational estimator since the performance is out of range). Left: variational, right: spectral. 8 Conclusion Summary. This paper compared ridge-regularized variational and spectral density-ratio estimation in a simple Gaussian location model under proportional high-dimensional asymptotics m … view at source ↗
Figure 9
Figure 9. Figure 9: Performance of the estimator with nuclear penalty [PITH_FULL_IMAGE:figures/full_fig_p060_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Comparison of two spectral estimators for a given [PITH_FULL_IMAGE:figures/full_fig_p060_10.png] view at source ↗
read the original abstract

We study ridge-regularized log-density-ratio estimation in the Gaussian location model with a common covariance matrix. By affine invariance, the model is written as q $\sim$ N(0, I), p $\sim$ N($\Delta$, I), with linear features, where $\Delta$ is a mean vector. The variational estimator is the empirical Kullback-Leibler (KL) log-normalized fit with a squared L2-penalty on its nonconstant coefficient, and the spectral estimator recently introduced in [1] replaces a single variational problem by a continuum of ridge-regularized least-squares problems. We derive high-dimensional deterministic asymptotic equivalents when the numbers of observations and dimension tend to infinity with fixed ratios. The regularized variational limit is characterized by a scalar entropy minimization problem derived from the convex-Gaussian-min-max theorem (CGMT), while the regularized spectral limit follows from deterministic equivalents for resolvents of weighted sums of two independent Gaussian sample covariance matrices. We use these formulas to compare population risks, with experiments focused on fixed-signal aspect-ratio sweeps and optimized regularization. Our conclusion is that with many observations, under the criteria and asymptotic regimes analyzed here, the well-specified variational estimator has the smaller risk, while with fewer observations, the spectral estimator is favored because its covariance-based construction has lower variance. We also study how a nuclear penalty can be used and partially analyzed to perform feature learning.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 2 minor

Summary. The manuscript derives high-dimensional asymptotic equivalents for the population risks of ridge-regularized variational and spectral log-density-ratio estimators in the Gaussian location model (q ~ N(0,I), p ~ N(Δ,I) with common covariance). The variational limit is characterized via the convex-Gaussian-min-max theorem (CGMT) as a scalar entropy minimization problem, while the spectral limit follows from deterministic equivalents for resolvents of weighted sums of two independent Gaussian sample covariance matrices. These formulas are used to compare risks, with the conclusion that the well-specified variational estimator has smaller risk with many observations while the spectral estimator is favored with fewer observations due to lower variance from its covariance-based construction. Experiments focus on fixed-signal aspect-ratio sweeps with optimized regularization, and a nuclear penalty is partially analyzed for feature learning.

Significance. If the derivations hold, the work provides precise asymptotic risk characterizations that enable direct comparison of the two estimators in the high-dimensional regime with fixed observation-to-dimension ratios. The explicit use of CGMT for the variational case and resolvent deterministic equivalents for the spectral case is a strength, as these are standard tools that yield exact limits under the stated model. The scoping of the conclusion to the analyzed criteria, regimes, and Gaussian location model is clearly stated upfront.

minor comments (2)
  1. [Abstract] Abstract: the phrase 'optimized regularization' is used for the risk comparisons but the manuscript does not specify whether the regularization strength is chosen by oracle knowledge of the population risk or by a data-driven procedure; this affects how the comparison should be interpreted in practice.
  2. [Model setup paragraph] The high-dimensional limit is taken with fixed ratios of observations to dimension, but the notation for the limiting ratios (e.g., n_p/d, n_q/d) is not introduced until later sections; adding it to the model-setup paragraph would improve readability.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the careful reading, positive assessment of the derivations and scoping, and recommendation of minor revision. No major comments appear in the provided report.

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper derives high-dimensional asymptotic equivalents for the variational estimator via the external convex-Gaussian-min-max theorem (CGMT) and for the spectral estimator via resolvent deterministic equivalents of weighted sample covariances. These are standard external mathematical tools. The subsequent population-risk comparison occurs inside the stated Gaussian location model and fixed-ratio asymptotic regime, but does not reduce any claim to a fitted parameter, self-definition, or self-citation chain by construction. Model assumptions are declared explicitly at the outset; no load-bearing step collapses to the paper's own inputs.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claims rest on the exact Gaussian location model, the high-dimensional proportional limit, and the validity of the CGMT and resolvent deterministic equivalents; no new entities are postulated.

free parameters (1)
  • ridge regularization strength
    The paper studies ridge-regularized estimators and optimizes the penalty parameter; this is a tunable hyperparameter whose value affects the reported risk comparison.
axioms (2)
  • domain assumption Data exactly follows the two-Gaussian location model with identical covariance matrices
    Invoked in the model definition and used to obtain affine invariance.
  • domain assumption High-dimensional limit with n, d → ∞ at fixed aspect ratios
    Required for the deterministic equivalents derived via CGMT and resolvent theory.

pith-pipeline@v0.9.1-grok · 5791 in / 1500 out tokens · 26608 ms · 2026-07-03T17:43:34.602771+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

63 extracted references · 4 canonical work pages · 2 internal anchors

  1. [1]

    A Spectral Framework for Closed-Form Relative Density Estimation

    Francis Bach. A spectral framework for closed-form rela tive density estimation, 2026. arXiv:2605.10668. (cited on pages 1, 2, 3, 5, 6, 7, 26, 28, and 29) 59 Figure 9: Performance of the estimator with nuclear penalty , when varying α p (α q = 0. 1 fixed, with α p swept over [0, 1]). Figure 10: Comparison of two spectral estimators for a given s = 1 when r...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  2. [2]

    Direct importance estimation for covariate shif t adaptation

    Masashi Sugiyama, Taiji Suzuki, Shinichi Nakajima, His ashi Kashima, Paul von B¨ unau, and Motoaki Kawanabe. Direct importance estimation for covariate shif t adaptation. Annals of the Institute of Statistical Mathematics , 60(4):699–746, 2008. (cited on pages 1 and 3)

    2008

  3. [3]

    A least-squares approach to direct im- portance estimation

    Takafumi Kanamori, Shohei Hido, and Masashi Sugiyama. A least-squares approach to direct im- portance estimation. Journal of Machine Learning Research , 10:1391–1445, 2009. (cited on pages 1 and 3)

    2009

  4. [4]

    Density Ratio Estimation in Machine Learning

    Masashi Sugiyama, Taiji Suzuki, and Takafumi Kanamori. Density Ratio Estimation in Machine Learning. Cambridge University Press, 2012. (cited on pages 1 and 3)

    2012

  5. [5]

    Wainwright, and Michael I

    XuanLong Nguyen, Martin J. Wainwright, and Michael I. Jo rdan. Estimating divergence functionals and the likelihood ratio by convex risk minimization. IEEE Transactions on Information Theory , 56(11):5847–5861, 2010. (cited on pages 1, 3, 4, 7, and 29)

    2010

  6. [6]

    Donsker and S

    Monroe D. Donsker and S. R. Srinivasa Varadhan. Asymptot ic evaluation of certain Markov process expectations for large time, I. Communications on Pure and Applied Mathematics , 28(1):1–47, 1975. (cited on pages 1, 3, 7, and 29)

    1975

  7. [7]

    Inferences for case-control and semiparametr ic two-sample density ratio models

    Jing Qin. Inferences for case-control and semiparametr ic two-sample density ratio models. Biometrika, 85(3):619–630, 1998. (cited on page 1)

    1998

  8. [8]

    Cand` es and Pragya Sur

    Emmanuel J. Cand` es and Pragya Sur. The phase transition for the existence of the maximum likelihood estimate in high-dimensional logistic regression. The Annals of Statistics , 48(1):27–42, 2020. (cited on pages 1 and 3)

    2020

  9. [9]

    Fi nite-sample performance of the maximum likelihood estimator in logistic regression, 2024

    Hugo Chardon, Matthieu Lerasle, and Jaouad Mourtada. Fi nite-sample performance of the maximum likelihood estimator in logistic regression, 2024. arXiv: 2411.02137. (cited on pages 1 and 3)

    work page arXiv 2024

  10. [10]

    Hoerl and Robert W

    Arthur E. Hoerl and Robert W. Kennard. Ridge regression : Biased estimation for nonorthogonal problems. Technometrics, 12(1):55–67, 1970. (cited on page 1)

    1970

  11. [11]

    Smola, Arthur Gretton, Kar sten M

    Jiayuan Huang, Alexander J. Smola, Arthur Gretton, Kar sten M. Borgwardt, and Bernhard Sch¨ olkopf. Correcting sample selection bias by unlabeled data. In Advances in Neural Information Processing Systems, 2007. (cited on page 3)

    2007

  12. [12]

    Linking losses for density ratio and class-probability estimation

    Aditya Krishna Menon and Cheng Soon Ong. Linking losses for density ratio and class-probability estimation. In International Conference on Machine Learning , 2016. (cited on page 3)

    2016

  13. [13]

    Cand` es

    Pragya Sur and Emmanuel J. Cand` es. A modern maximum-li kelihood theory for high-dimensional logistic regression. Proceedings of the National Academy of Sciences , 116(29):14516–14525, 2019. (cited on page 3)

    2019

  14. [14]

    The i mpact of regularization on high-dimensional logistic regression

    Fariborz Salehi, Ehsan Abbasi, and Babak Hassibi. The i mpact of regularization on high-dimensional logistic regression. In Advances in Neural Information Processing Systems , 2019. (cited on pages 3, 13, and 28)

    2019

  15. [15]

    Silverstein

    Zhidong Bai and Jack W. Silverstein. Spectral Analysis of Large Dimensional Random Matrices . Springer, 2nd edition, 2010. (cited on pages 3, 14, 28, and 29) 61

    2010

  16. [16]

    Dete rministic equivalents for certain functionals of large random matrices

    Walid Hachem, Philippe Loubaton, and Jamal Najim. Dete rministic equivalents for certain functionals of large random matrices. The Annals of Applied Probability , 17(3):875–930, 2007. (cited on pages 3 and 28)

    2007

  17. [17]

    High-dimensional asy mptotics of prediction: Ridge regression and classification

    Edgar Dobriban and Stefan Wager. High-dimensional asy mptotics of prediction: Ridge regression and classification. The Annals of Statistics , 46(1):247–279, 2018. (cited on page 3)

    2018

  18. [18]

    The interpolation p hase transition in neural networks: Memo- rization and generalization under lazy training

    Andrea Montanari and Yiqiao Zhong. The interpolation p hase transition in neural networks: Memo- rization and generalization under lazy training. The Annals of Statistics , 50(5):2816–2847, 2022. (cited on pages 3, 14, and 28)

    2022

  19. [19]

    High-dimensional analysis of double des cent for linear regression with random projec- tions

    Francis Bach. High-dimensional analysis of double des cent for linear regression with random projec- tions. SIAM Journal on Mathematics of Data Science , 6(1):26–50, 2024. (cited on pages 3, 14, 15, 16, 28, 42, and 43)

    2024

  20. [20]

    Peter McCullagh and John A. Nelder. Generalized Linear Models . Chapman & Hall, 1989. (cited on page 4)

    1989

  21. [21]

    Convex Optimization

    Stephen Boyd and Lieven Vandenberghe. Convex Optimization . Cambridge University Press, 2004. (cited on page 4)

    2004

  22. [22]

    Jorge Nocedal and Stephen J. Wright. Numerical Optimization. Springer, 2nd edition, 2006. (cited on pages 4 and 13)

    2006

  23. [23]

    Minimization of ϕ -divergences on sets of signed measures

    Michel Broniatowski and Amor Keziou. Minimization of ϕ -divergences on sets of signed measures. Studia Scientiarum Mathematicarum Hungarica , 43(4):403–442, 2006. (cited on page 5)

    2006

  24. [24]

    Orthogonal Polynomials: Computation and Approximation

    Walter Gautschi. Orthogonal Polynomials: Computation and Approximation . Oxford University Press,

  25. [25]

    (cited on pages 6 and 13)

  26. [26]

    Existence of direct density ratio estimators, 2025

    Erika Banzato, Mathias Drton, Kian Saraf-Poor, and Hon gjian Shi. Existence of direct density ratio estimators, 2025. arXiv:2502.12738. (cited on page 9)

    work page arXiv 2025

  27. [27]

    The implicit bias of gradient descent on separable data

    Daniel Soudry, Elad Hoffer, Mor Shpigel Nacson, Suriya Gu nasekar, and Nathan Srebro. The implicit bias of gradient descent on separable data. Journal of Machine Learning Research , 19(70):1–57, 2018. (cited on page 9)

    2018

  28. [28]

    Lee, Daniel Soudry, and Nath an Srebro

    Suriya Gunasekar, Jason D. Lee, Daniel Soudry, and Nath an Srebro. Implicit bias of gradient descent on linear convolutional networks. In Advances in Neural Information Processing Systems , 2018. (cited on page 9)

    2018

  29. [29]

    van der Vaart

    Aad W. van der Vaart. Asymptotic Statistics . Cambridge University Press, 1998. (cited on pages 10, 19, and 20)

    1998

  30. [30]

    High-Dimensional Probability: An Introduction with Applic ations in Data Science

    Roman Vershynin. High-Dimensional Probability: An Introduction with Applic ations in Data Science . Cambridge University Press, 2018. (cited on page 10)

    2018

  31. [31]

    The Gaussian min-max theorem in the Presence of Convexity

    Christos Thrampoulidis, Samet Oymak, and Babak Hassib i. The Gaussian min-max theorem in the presence of convexity, 2014. arXiv:1408.4837. (cited on pages 12 and 37) 62

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  32. [32]

    Regularized linear regression: A precise analysis of the estimation error

    Christos Thrampoulidis, Samet Oymak, and Babak Hassib i. Regularized linear regression: A precise analysis of the estimation error. In Conference on Learning Theory , 2015. (cited on page 12)

    2015

  33. [33]

    Corless, Gaston H

    Robert M. Corless, Gaston H. Gonnet, David E. G. Hare, Da vid J. Jeffrey, and Donald E. Knuth. On the Lambert W function. Advances in Computational Mathematics , 5:329–359, 1996. (cited on page 13)

    1996

  34. [34]

    Trace Ideals and Their Applications

    Barry Simon. Trace Ideals and Their Applications . American Mathematical Society, 2005. (cited on page 18)

    2005

  35. [35]

    Richard P. Brent. Algorithms for Minimization Without Derivatives . Prentice Hall, 1973. (cited on page 26)

    1973

  36. [36]

    Convex multi-task feature learning

    Andreas Argyriou, Theodoros Evgeniou, and Massimilia no Pontil. Convex multi-task feature learning. Machine Learning, 73(3):243–272, 2008. (cited on pages 28, 45, and 47)

    2008

  37. [37]

    Sinan G¨ unt¨ urk

    Ingrid Daubechies, Ronald DeVore, Massimo Fornasier, and C. Sinan G¨ unt¨ urk. Iteratively reweighted least squares minimization for sparse recovery. Communications on Pure and Applied Mathematics , 63(1):1–38, 2010. (cited on page 29)

    2010

  38. [38]

    Optimization with sparsity-inducing penalties

    Francis Bach, Rodolphe Jenatton, Julien Mairal, and Gu illaume Obozinski. Optimization with sparsity-inducing penalties. Foundations and Trends in Machine Learning , 4(1):1–106, 2012. (cited on page 29)

    2012

  39. [39]

    Ajanki, L´ aszl´ o Erd˝ os, and Torben H

    Oskari H. Ajanki, L´ aszl´ o Erd˝ os, and Torben H. Kr¨ uger. Stability of the matrix Dyson equation and random matrices with correlations. Probability Theory and Related Fields , 173(1–2):293–373, 2019. (cited on pages 29, 49, 51, and 53)

    2019

  40. [40]

    Mutual information neural esti mation

    Mohamed Ishmael Belghazi, Aristide Baratin, Sai Rajes hwar, Sherjil Ozair, Yoshua Bengio, Aaron Courville, and Devon Hjelm. Mutual information neural esti mation. In International Conference on Machine Learning, 2018. (cited on page 29)

    2018

  41. [41]

    Alemi, and George Tucker

    Ben Poole, Sherjil Ozair, Aaron van den Oord, Alexander A. Alemi, and George Tucker. On variational bounds of mutual information. In International Conference on Machine Learning, 2019. (cited on page 29)

    2019

  42. [42]

    Random features for larg e-scale kernel machines

    Ali Rahimi and Benjamin Recht. Random features for larg e-scale kernel machines. In Advances in Neural Information Processing Systems , 2007. (cited on page 29)

    2007

  43. [43]

    Christopher K. I. Williams and Matthias Seeger. Using t he Nystr¨ om method to speed up kernel machines. In Advances in Neural Information Processing Systems , 2001. (cited on page 29)

    2001

  44. [44]

    A rando m matrix approach to neural networks

    Cosme Louart, Zhenyu Liao, and Romain Couillet. A rando m matrix approach to neural networks. The Annals of Applied Probability , 28(2):1190–1248, 2018. (cited on page 29)

    2018

  45. [45]

    The generalization erro r of random features regression: Precise asymptotics and the double descent curve

    Song Mei and Andrea Montanari. The generalization erro r of random features regression: Precise asymptotics and the double descent curve. Communications on Pure and Applied Mathematics , 75(4):667–766, 2022. (cited on page 29)

    2022

  46. [46]

    Hong Hu and Yue M. Lu. Universality laws for high-dimens ional learning with random features. IEEE Transactions on Information Theory , 69(3):1932–1964, 2023. (cited on page 29) 63

    1932

  47. [47]

    Gener- alisation error in learning with random features and the hid den manifold model

    Federica Gerace, Bruno Loureiro, Florent Krzakala, Ma rc M´ ezard, and Lenka Zdeborov´ a. Gener- alisation error in learning with random features and the hid den manifold model. In International Conference on Machine Learning , 2020. (cited on page 29)

    2020

  48. [48]

    Learning curves of generic features mapsfor realistic datasets with a teacher-student model

    Bruno Loureiro, C´ edric Gerbelot, Hugo Cui, Sebastian Goldt, Florent Krzakala, Marc M´ ezard, and Lenka Zdeborov´ a. Learning curves of generic features mapsfor realistic datasets with a teacher-student model. In Advances in Neural Information Processing Systems , 2021. (cited on page 29)

    2021

  49. [49]

    Fluctuations, bias, variance & ensemble of learners: Exact asymptotics fo r convex losses in high-dimension

    Bruno Loureiro, Cedric Gerbelot, Maria Refinetti, Gabr iele Sicuro, and Florent Krzakala. Fluctuations, bias, variance & ensemble of learners: Exact asymptotics fo r convex losses in high-dimension. In International Conference on Machine Learning , 2022. (cited on page 29)

    2022

  50. [50]

    Johnstone

    Iain M. Johnstone. On the distribution of the largest ei genvalue in principal components analysis. The Annals of Statistics , 29(2):295–327, 2001. (cited on page 29)

    2001

  51. [51]

    Ph ase transition of the largest eigenvalue for nonnull complex sample covariance matrices

    Jinho Baik, G´ erard Ben Arous, and Sandrine P´ ech´ e. Ph ase transition of the largest eigenvalue for nonnull complex sample covariance matrices. The Annals of Probability , 33(5):1643–1697, 2005. (cited on page 29)

    2005

  52. [52]

    Precise error analysis of regularized M- estimators in high dimensions

    Christos Thrampoulidis, Ehsan Abbasi, and Babak Hassi bi. Precise error analysis of regularized M- estimators in high dimensions. IEEE Transactions on Information Theory , 64(8):5592–5628, 2018. (cited on page 39)

    2018

  53. [53]

    On the numerical evaluation of Fred holm determinants

    Folkmar Bornemann. On the numerical evaluation of Fred holm determinants. Mathematics of Com- putation, 79(270):871–915, 2010. (cited on page 44)

    2010

  54. [54]

    Perturbation Theory for Linear Operators

    Tosio Kato. Perturbation Theory for Linear Operators . Springer, 1995. (cited on page 46)

    1995

  55. [55]

    Isotropic local laws for sample covariance and generalized Wigner matrices

    Alex Bloemendal, L´ aszl´ o Erd˝ os, Antti Knowles, Horng-Tzer Yau, and Jun Yin. Isotropic local laws for sample covariance and generalized Wigner matrices. Electronic Journal of Probability , 19(33):1–53,

  56. [56]

    (cited on pages 49 and 53)

  57. [57]

    Charles M. Stein. Estimation of the mean of a multivaria te normal distribution. The Annals of Statistics, 9(6):1135–1151, 1981. (cited on page 49)

    1981

  58. [58]

    Khorunzhy, Boris A

    Alexei M. Khorunzhy, Boris A. Khoruzhenko, and Leonid A . Pastur. Asymptotic properties of large random matrices with independent entries. Journal of Mathematical Physics , 37(10):5033–5060, 1996. (cited on page 49)

    1996

  59. [59]

    Ajanki, L´ aszl´ o Erd˝ os, and Torben H

    Oskari H. Ajanki, L´ aszl´ o Erd˝ os, and Torben H. Kr¨ uger. Quadratic Vector Equations on Complex Upper Half-Plane . Memoirs of the American Mathematical Society. American Ma thematical Society,

  60. [60]

    (cited on pages 51 and 53)

  61. [61]

    Kr¨ uger

    Johannes Alt, L´ aszl´ o Erd˝ os, and Torben H. Kr¨ uger. The Dyson equation with linear self-energy: Spectral bands, edges and cusps. Documenta Mathematica , 25:1421–1539, 2020. (cited on pages 51 and 53)

    2020

  62. [62]

    Fixed point equations and nonlinear eig envalue problems in ordered Banach spaces

    Herbert Amann. Fixed point equations and nonlinear eig envalue problems in ordered Banach spaces. SIAM Review , 18(4):620–709, 1976. (cited on page 59)

    1976

  63. [63]

    Positive Definite Matrices

    Rajendra Bhatia. Positive Definite Matrices . Princeton University Press, 2007. (cited on page 59) 64

    2007