Pith. sign in

REVIEW 3 major objections 4 minor 48 references

Fluctuations of the largest eigenvalues of transformed spiked Wigner matrices

T0 review · 3 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read The largest eigenvalue of a spiked Wigner matrix whose entries are transformed entrywise still shows the Baik–Ben Arous–Péché phase transition, with the signal-to-noise ratio replaced by an effective one.

desk verdict A genuine fluctuation-level BBP theorem for transformed spiked Wigner matrices, with a coherent proof strategy and two technical gaps that look repairable. read the letter →

arxiv 2502.04720 v3 pith:SJDY2C4K submitted 2025-02-07 math.PR

classification math.PR MSC 60B2015B52
keywords spikedWignermatrixentrywisetransformBBPphasetransitioneffectivesignal-to-noiseratioGOETracy-WidomdistributioneigenvaluerigiditylocallawGreenfunctioncomparison
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 proves that applying a smooth, centered function entrywise to a spiked Wigner matrix does not change the fluctuation law of the largest eigenvalue; it only replaces the signal-to-noise ratio by an effective one. For the transformed matrix the top eigenvalue has a BBP phase transition governed by $\lambda_e = \lambda(\mathbb{E}[f'(\sqrt{N}W_{12})])^2$: when $\lambda_e>1$, the fluctuation of $\mu_1(\widehat{M})$ is $N^{-1/2}$-scale Gaussian with variance $2(\lambda_e-1)/\lambda_e$; when $\lambda_e<1$, the fluctuation is $N^{-2/3}$-scale GOE Tracy–Widom. The paper also proves rigidity bounds: the top eigenvalue stays within $N^{-1/2+\epsilon}$ of $\sqrt{\lambda_e}+1/\sqrt{\lambda_e}$ in the supercritical case and within $N^{-2/3+\epsilon}$ of $2$ in the subcritical case, with overwhelming probability. This matters because entrywise transformations are used to boost the signal-to-noise ratio before PCA, and the result gives the exact fluctuation laws needed for tests based on the largest eigenvalue.

What carries the argument

The central object is the interpolating matrix $H(t)=V(t)+\sqrt{\lambda_e}\,xx^T+(\lambda/2)\mathbb{E}[f''(\sqrt{N}W_{12})]\sqrt{N}\,x^2(x^2)^T$, where $V(t)$ is a Wigner-type matrix whose variance profile rescales along a path from a Wigner matrix $V(0)$ to the actual noise $V(1)$, and $x^2=(x_1^2,\ldots,x_N^2)$ is the entrywise square of the spike. The resolvent $G(t,z)=(H(t)-zI)^{-1}$ satisfies an entrywise local law around the Stieltjes transform $m_{\mathrm{sc}}(z)$ of the semicircle law, and this local law makes the Green function comparison theorem applicable: expectations of smooth functionals of the resolvent are nearly $t$-independent, so the fluctuation of the largest eigenvalue of $H(1)$ equals that of $H(0)$. The single effective parameter $\lambda_e=\lambda(\mathbb{E}[f'(\sqrt{N}W_{12})])^2$ decides which regime applies.

What would settle it

Rerun the authors' numerical experiment near the supercritical threshold, for example $N=4096$ with Gaussian noise, Rademacher spike, $f(x)=(x^2+3x-1)/\sqrt{11}$, and $\lambda=2.5$ so that $\lambda_e\approx 1.294$; if the empirical distribution of $N^{1/2}(\mu_1(\widehat{M})-(\sqrt{\lambda_e}+1/\sqrt{\lambda_e}))$ has a nonzero mean or variance differing from $2(\lambda_e-1)/\lambda_e$ by more than sampling error, Theorem 2.5 is false. A more targeted check of the proof's load-bearing step is to verify that the largest eigenvalue of the interpolating matrix $H(t)$ changes by $o(N^{-1/2})$ as $t$ runs from $0$ to $1$.

Watch

Extended reading notes

Core claim

The discovery is a universality statement: the fluctuations of $\mu_1(\widehat{M})$ coincide with those of a rank-2 spiked Wigner matrix carrying one effective spike of size $\sqrt{\lambda_e}$, not with those of the naive first-order approximation. The Taylor remainder that the first-order expansion discards—the fluctuating part of $f'$ and the second-derivative term—is actually the same size as the fluctuation being studied, so the true noise term is a Wigner-type matrix and the spike is effectively rank-2. An interpolation path $V(t)$ connects this Wigner-type noise to a true Wigner matrix, and a Green function comparison argument, powered by a local law for the resolvent, shows that smooth functionals of the resolvent do not change along the path. The endpoint comparison yields the Gaussian limit in the supercritical case through known finite-rank deformation results, and the GOE Tracy–Widom limit in the subcritical case through eigenvalue sticking together with edge universality.

Load-bearing premise

The proof relies on a technical local-law estimate for the resolvent of the interpolating Wigner-type matrix in the exact scaling window, and on applying a rigidity result originally stated for two orthogonal spikes to the non-orthogonal pair $x$ and $x^2$.

Editorial extensions

If this is right

  • Above the threshold $\lambda_e>1$, the top eigenvalue of the transformed matrix is asymptotically Gaussian with $N^{1/2}(\mu_1(\widehat{M})-(\sqrt{\lambda_e}+1/\sqrt{\lambda_e})) \Rightarrow \mathcal{N}(0,2(\lambda_e-1)/\lambda_e)$, giving usable p-values for transformed PCA.
  • Below the threshold $\lambda_e<1$, the top eigenvalue follows GOE Tracy–Widom on the $N^{-2/3}$ scale, so no test based only on $\mu_1$ can distinguish a subcritical spike from the noise edge asymptotically.
  • With overwhelming probability the largest eigenvalue is rigid at the optimal scale: $O(N^{-1/2+\epsilon})$ above threshold and $O(N^{-2/3+\epsilon})$ below, so the fluctuation laws describe typical behavior, not just limits.
  • The transition is governed entirely by $\lambda_e$: any two transforms $f$ with the same $\mathbb{E}[f'(\sqrt{N}W_{12})]$ give the same limiting fluctuation of the largest eigenvalue.

Reading between the lines

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

  • Assuming the authors' conjectured extension to rank-$k$ spikes, each supercritical effective spike $\lambda_e^{(i)}>1$ would contribute a Gaussian fluctuation of variance $2(\lambda_e^{(i)}-1)/\lambda_e^{(i)}$, while the remaining eigenvalues would follow GOE statistics; this would make multi-signal transformed PCA amenable to the same tests.
  • The appendix's treatment of $\lambda_e=0$ suggests a hierarchy of BBP transitions indexed by the first non-vanishing derivative of $f$: with $\lambda\sim \lambda_0 N^{(1-1/k_f)/2}$, the effective SNR is set by $\mathbb{E}[f^{(k_f)}(\sqrt{N}W_{12})]$. A natural simulation check is to take an even transform such as $f(x)=x^2$ on Gaussian noise and test the predicted shifted Gaussian limit.
  • Since the fluctuation law coincides with that of the raw spiked Wigner model at both scales, the largest eigenvalue alone carries no information about whether an entrywise transform was applied; distinguishing transformed from raw data would require the full spectrum or eigenvector statistics.
  • The subcritical eigenvalue-sticking mechanism implies a strong statistical indistinguishability statement: even with the optimal transform, a spike below threshold cannot move the top eigenvalue's law away from the pure-noise edge at the Tracy–Widom scale.
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

3 major / 4 minor

Summary. The paper studies the largest eigenvalue of a transformed spiked Wigner matrix fM with entries fM_ij = N^{-1/2} f(√N M_ij), where M = W + √λ xx^T. The main result, Theorem 2.5, is a BBP-type fluctuation theorem: if the effective SNR λe = λ(E[f'(√N W_12)])^2 is larger than 1, then N^{1/2}(μ_1(fM) - (√λe + 1/√λe)) converges to a centered Gaussian with variance 2(λe-1)/λe, while if λe < 1, then N^{2/3}(μ_1(fM)-2) converges to the GOE Tracy-Widom distribution. Theorem 2.6 provides matching rigidity estimates with overwhelming probability. The proof strategy is to Taylor-expand f, approximate fM by a rank-2 spiked Wigner-type matrix H, interpolate through a family V(t) of Wigner-type matrices, import the local law for general Wigner-type matrices from [2], and then apply Green function comparison to reduce the fluctuation problem to known results for spiked Wigner matrices [48] and for the Wigner edge [33]. An appendix treats the case λe = 0 with a larger SNR scaling.

Significance. If the technical gaps described below are repaired, this is a valuable contribution: it upgrades the convergence of the top eigenvalue of transformed spiked Wigner matrices, established in [46] and related work, to the fluctuation scale, and it identifies the effective SNR λe as a completely deterministic functional of the model with no fitted parameters. The subcritical fluctuation result is especially nontrivial because the naive Taylor approximation (1.3) has errors comparable to or larger than the target fluctuation. The numerical experiments in Section 5 support the main theorem for two concrete choices of f. The paper is carefully structured and gives substantial detail in the appendices; the main caveat is that two load-bearing inputs are imported from the literature without fully verifying their hypotheses in the precise N-, t-, and spike-dependent regime used here.

major comments (3)
  1. [Lemma B.3; Propositions 3.2 and 4.2] The local laws for V(t) are imported from [2] by asserting parameter identifications: Lemma B.3 states that Theorem 1.7 of [2] applies with κ(z)=Θ(1) and ρ(z)=Θ(η) via equations (1.17), (1.23) and (4.5f) of [2], and Proposition 4.2 similarly cites ρ(z)=Θ(√κ+η) via (4.5d). The variance profile S_ij(t) = 1 + C_V^1 t√N x_i x_j + C_V^2 t N x_i^2 x_j^2 in (3.4)-(3.5) is N-dependent, t-dependent and spike-dependent, and the paper does not verify the hypotheses behind those cited parameter formulas, nor the well-posedness of the quadratic vector equation (3.10) in this regime. This is load-bearing because the anisotropic local law (B.9) with unit vectors x and x^2 is used in the resolvent estimates (B.12), (B.14), and again in (C.2)-(C.4). In addition, the identification κ(z)=Θ(1) is not self-evident: in Proposition 3.2, κ = τ - (√λe + 1/√λe) can be as small as N^{-1/2+ε}, so if κ in [2] denotes a different quantity the notation should be disentangled. The authors should either verify the parameter regime directly or state precisely which hypotheses of [2] are satisfied by the profile (3.5).
  2. [Appendix B.1, proof of Theorem 2.6] Proposition B.2, quoted from Theorem 2.7 of [26], is stated for a Wigner matrix with two orthogonal spikes, ⟨x,y⟩=0, and with 0<λ_2<1<λ_1. It is applied to H-∆+D with spikes x and y=x^2, but x and x^2 are not orthogonal: Assumption 2.3 only gives ∑_i x_i^3 = O(N^{-1+ε}), not zero. Since Theorem 2.6 (supercritical rigidity) is used in Proposition 3.3 to ensure that only the top eigenvalue can contribute to the window [E,E_+], this is a load-bearing step in the Green function comparison. The authors should write out the repair, for example by passing to an orthogonal eigenbasis of A_N whose top two eigenvalues are θ_1 = √λe + O(N^{-1+ε}) and θ_2 = O(N^{-1/2+ε}) as in Proposition 3.5, and then verifying that Proposition B.2 applies to that basis with the required λ_1, λ_2. As written, the proof is conditional on this non-orthogonality being harmless.
  3. [Proposition 4.1, final paragraph] The proof of Proposition 4.1 shows that for ξ ∈ [N^{-3/4}, N^{-1/2+ε}] the assumption µ_1(H) ∈ [µ_1(V)+ξ, µ_1(V)+ξ+η] leads to a contradiction, and then states: 'Finally, adapting the strategy of the proof of Theorem 2.6 in the supercritical case, we can prove that |µ_1(H)-µ_1(V)| ≺ N^{-1/2}.' This final sentence is not a proof: in the subcritical case there is no outlier eigenvalue, and the supercritical rigidity proof invoked Proposition B.2, which is unavailable here. The displayed argument does not cover gaps larger than N^{-1/2+ε}, so the stated bound µ_1(H)-µ_1(V) ≤ N^{-3/4} does not follow from the preceding estimates alone. Please provide the missing argument, for example by extending the local-law estimate to ξ up to a constant using interlacing, or by a different eigenvalue-sticking argument.
minor comments (4)
  1. [Section 4, first paragraph] The sentence 'The detailed proofs for the results in Section 3 can be found in Appendix B' should refer to Section 4 and Appendix C; as written it is a typographical error.
  2. [Section 5.2 and Figure 3 caption] The text says the subcritical simulation uses λ = 0.1 with λe ≈ 0.350, while the caption of Figure 3(b) states λ = 0.15; please reconcile these values.
  3. [Assumption 2.3] The conditions 'P_i x_i = O(N^ε)' and 'P_i x_i^3 = O(N^{-1+ε})' should be written with explicit summation symbols (e.g., ∑_i x_i and ∑_i x_i^3) to avoid confusion with powers of a single coordinate.
  4. [Proof of Proposition 3.4] The event Ω_ε is defined as max_{i,j}|G_ij - m_sc δ_ij| < N^{-1/2+7ε} for all x ∈ [E_1,E_2], but Proposition 3.2 is stated for a fixed z = x+iη. A lattice argument or a uniform-in-τ version of the local law should be cited or sketched to justify passage from pointwise to uniform control on the interval [E_1,E_2].

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the fluctuation theorem is proved by reducing to external local laws and spiked-Wigner fluctuation results; the effective SNR is a deterministic functional, not a fitted input.

full rationale

The derivation chain is self-contained as a proof, even though it relies on cited theorems. The effective SNR is defined deterministically in (2.1) as lambda_e = lambda(E[f'(sqrt(N) W12)])^2, with no parameter fitted to the eigenvalue data being predicted. The approximation step, Proposition 3.1, expands f entrywise and proves mu_1(fM) - mu_1(H) = O(N^{-1+epsilon}) using Weyl's inequality and norm bounds; the matrix H is defined by the Taylor expansion of f, not by the desired limiting law. The supercritical proof then constructs an interpolation H(t) with V(1)=V and V(0) a Wigner matrix, imports the Wigner-type local law from the external reference [2] (Ajanki-Erdos-Kruger), proves a Green function comparison (Proposition 3.4) by Stein's method, and reads off the final Gaussian fluctuation from the external rank-2 spiked Wigner theorem [48]. Proposition 3.5 independently computes the effective spike eigenvalue theta_1 = sqrt(lambda_e) + O(N^{-1+epsilon}), so the final variance 2(lambda_e-1)/lambda_e is derived, not assumed. The subcritical proof similarly compares H with V, imports the edge local law from [2] and terminal Tracy-Widom law from [33], again with no fitted parameter. The self-citations ([32], [33]) are to independent published theorems: [33] supplies Wigner edge universality and [32] supplies a standard Poisson-kernel cutoff lemma for Green function comparison; neither assumes the present result, and the paper reproduces the relevant arguments rather than importing the conclusion as a black-box uniqueness statement. The remaining concerns flagged in the paper are technical correctness risks, not circularity: the local law from [2] is cited with terse verification of the kappa/rho regime, and Proposition B.2 is stated for orthogonal spikes but applied to x and x^2; these are potential gaps in the proof, but they do not make any equation reduce to its own input. Appendix D explicitly offers only an idea for the lambda_e=0 scaling and is not presented as a theorem. No circular step was found.

Assumptions & free parameters 0 free parameters · 6 assumptions · 0 invented entities

No parameter is fitted to the simulation data; the free-parameter list is empty. The central result depends on external theorems on local laws, rigidity, outlier fluctuations, and edge universality, plus explicit model assumptions. These are standard input theorems rather than axioms generated by the paper.

assumptions (6)
  • standard math Local law for Wigner-type matrices with variance profile (Theorem 1.7 and 1.13 in [2])
    Used in Lemma B.3 and Proposition 4.2 to control the resolvent of V(t) and obtain the Green function comparison estimates.
  • standard math Outlier fluctuation theorem for finite-rank deformations of Wigner matrices (Theorem 1.3 in [48])
    Used in Proposition B.5 to give the Gaussian law for the largest eigenvalue of the rank-2 spiked Wigner target matrix in the supercritical case.
  • standard math Rigidity for spiked Wigner matrices with orthogonal spikes (Theorem 2.7 in [26])
    Used in Proposition B.2 to obtain the N^{-1/2} rigidity of the rank-2 spiked Wigner target; the paper applies it to non-orthogonal spikes without comment.
  • standard math Edge universality for Wigner matrices (Theorem 1.2 in [33])
    Used in the subcritical case to identify the GOE Tracy-Widom limit of V(0).
  • domain assumption Assumption 2.3 on the spike: max |x_i| = O(N^{-1/2+ε}), Σ x_i = O(N^ε), Σ x_i^3 = O(N^{-1+ε})
    Used throughout for the delocalization of the spike, e.g., in Proposition B.1, B.3, and the power counting.
  • domain assumption Assumption 2.4 on f: C^3, polynomial growth, E[f]=0, E[f^2]=1, E[f'] ≥ 0
    Controls the Taylor expansion and ensures the effective SNR λe is well-defined.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fluctuations of the largest eigenvalues of transformed spiked Wigner matrices." pith.science (2026). https://pith.science/paper/SJDY2C4K

@misc{pith2026250204720,
  author       = {Pith},
  title        = {Pith review of: Fluctuations of the largest eigenvalues of transformed spiked Wigner matrices},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SJDY2C4K}},
  note         = {Machine review of arXiv:2502.04720}
}
read the original abstract

We consider a spiked random matrix model obtained by applying a function entrywise to a signal-plus-noise symmetric data matrix. We prove that the largest eigenvalue of this model, which we call a transformed spiked Wigner matrix, exhibits Baik-Ben Arous--P\'ech\'e (BBP) type phase transition. We show that the law of the fluctuation converges to the Gaussian distribution when the effective signal-to-noise ratio (SNR) is above the critical number, and to the GOE Tracy-Widom distribution when the effective SNR is below the critical number. We provide precise formulas for the limiting distributions and also concentration estimates for the largest eigenvalues, both in the supercritical and the subcritical regimes.

Figures

Figures reproduced from arXiv: 2502.04720 by the authors.

Figure 1
Figure 1. The noise density and the entrywise transformation used in the numerical experiment [PITH_FULL_IMAGE:figures/full_fig_p014_1.png] view at source ↗
Figure 2
Figure 2. The sampled distribution of µ1(Mf) for non-Gaussian noise with N = 1024, λ = 0.8(left) and λ = 0.1(right), after shifting and rescaling. The lines in figure 2(a) plots the Gaussian distribution introduced in Theorem 2.5, and the line in figure 2(b) plots the GOE Tracy–Widom distribution. 5.2 Spiked Gaussian Wigner Matrix with Entrywise Transformation We next consider a spiked Gaussian Wigner matrix, entrywise transf… view at source ↗
Figure 3
Figure 3. The sampled distribution of µ1(Mf) for Gaussian noise with N = 1024, λ = 2.5 (left) and λ = 0.15 (right), after shifting and rescaling. The line in figure 3(a) plots the Gaussian distribution introduced in Theorem 2.5, and the line in figure 3(b) plots the GOE Tracy–Widom distribution. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 46 canonical work pages

  1. [46]

    Optimality and sub- optimality of PCA I: Spiked random matrix models

    Amelia Perry, Alexander S Wein, Afonso S Bandeira, and Ankur Moitra. Optimality and sub- optimality of PCA I: Spiked random matrix models. The Annals of Statistics , 46(5):2416–2451, 2018

  2. [2]

    Universality for general Wigner-type matrices

    Oskari H Ajanki, L´ aszl´ o Erd˝ os, and Torben Kr¨ uger. Universality for general Wigner-type matrices. Probability Theory and Related Fields , 169:667–727, 2017. 40

  3. [48]

    On finite rank deformations of Wigner matrices II: Delocalized perturbations

    David Renfrew and Alexander Soshnikov. On finite rank deformations of Wigner matrices II: Delocalized perturbations. Random Matrices: Theory and Applications , 2(01):1250015, 2013. 44

  4. [33]

    A necessary and sufficient condition for edge universality of Wigner matrices

    Ji Oon Lee and Jun Yin. A necessary and sufficient condition for edge universality of Wigner matrices. Duke Mathematical Journal , 163(1):117–173, 2014

  5. [26]

    The isotropic semicircle law and deformation of Wigner matrices

    Antti Knowles and Jun Yin. The isotropic semicircle law and deformation of Wigner matrices. Communications on Pure and Applied Mathematics , 66(11):1663–1749, 2013

  6. [1]

    Community detection and stochastic block models: recent developments

    Emmanuel Abbe. Community detection and stochastic block models: recent developments. The Journal of Machine Learning Research , 18(1):6446–6531, 2017

  7. [3]

    Learning in the presence of low-dimensional structure: a spiked random matrix perspective

    Jimmy Ba, Murat A Erdogdu, Taiji Suzuki, Zhichao Wang, and Denny Wu. Learning in the presence of low-dimensional structure: a spiked random matrix perspective. Advances in Neural Information Processing Systems, 36, 2024

  8. [4]

    High- dimensional asymptotics of feature learning: How one gradient step improves the representation

    Jimmy Ba, Murat A Erdogdu, Taiji Suzuki, Zhichao Wang, Denny Wu, and Greg Yang. High- dimensional asymptotics of feature learning: How one gradient step improves the representation. Advances in Neural Information Processing Systems , 35:37932–37946, 2022

Show all 48 references
  1. [5]

    Phase transition of the largest eigenvalue for nonnull complex sample covariance matrices

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

  2. [6]

    The eigenvalues and eigenvectors of finite, low rank perturbations of large random matrices

    Florent Benaych-Georges and Raj Rao Nadakuditi. The eigenvalues and eigenvectors of finite, low rank perturbations of large random matrices. Advances in Mathematics , 227(1):494–521, 2011

  3. [7]

    On the principal components of sample covariance matrices

    Alex Bloemendal, Antti Knowles, Horng-Tzer Yau, and Jun Yin. On the principal components of sample covariance matrices. Probability theory and related fields , 164(1):459–552, 2016

  4. [8]

    Detection of a sparse submatrix of a high-dimensional noisy matrix

    Cristina Butucea, Yuri I Ingster, et al. Detection of a sparse submatrix of a high-dimensional noisy matrix. Bernoulli, 19(5B):2652–2688, 2013

  5. [9]

    The largest eigenvalues of finite rank deformation of large Wigner matrices: Convergence and nonuniversality of the fluctuations

    Mireille Capitaine, Catherine Donati-Martin, and Delphine F´ eral. The largest eigenvalues of finite rank deformation of large Wigner matrices: Convergence and nonuniversality of the fluctuations. The Annals of Probability , 37(1):1–47, 2009

  6. [10]

    Nonconvex optimization meets low-rank matrix factorization: An overview

    Yuejie Chi, Yue M Lu, and Yuxin Chen. Nonconvex optimization meets low-rank matrix factorization: An overview. IEEE Transactions on Signal Processing , 67(20):5239–5269, 2019

  7. [11]

    Weak detection of signal in the spiked wigner model

    Hye Won Chung and Ji Oon Lee. Weak detection of signal in the spiked wigner model. In International Conference on Machine Learning , pages 1233–1241. PMLR, 2019

  8. [12]

    Weak detection in the spiked wigner model.IEEE Transactions on Information Theory , 68(11):7427–7453, 2022

    Hye Won Chung and Ji Oon Lee. Weak detection in the spiked wigner model.IEEE Transactions on Information Theory , 68(11):7427–7453, 2022

  9. [13]

    Asymptotic normality of log likelihood ratio and fundamental limit of the weak detection for spiked Wigner matrices

    Hye Won Chung, Jiho Lee, and Ji Oon Lee. Asymptotic normality of log likelihood ratio and fundamental limit of the weak detection for spiked Wigner matrices. Bernoulli, 31(3):2276–2301, 2025

  10. [14]

    Asymptotics of feature learning in two-layer networks after one gradient-step

    Hugo Cui, Luca Pesce, Yatin Dandi, Florent Krzakala, Yue Lu, Lenka Zdeborova, and Bruno Loureiro. Asymptotics of feature learning in two-layer networks after one gradient-step. In International Conference on Machine Learning . PMLR, 2024

  11. [15]

    Neural networks can learn represen- tations with gradient descent

    Alexandru Damian, Jason Lee, and Mahdi Soltanolkotabi. Neural networks can learn represen- tations with gradient descent. In Conference on Learning Theory, pages 5413–5452. PMLR, 2022. 41

  12. [16]

    A random matrix theory perspective on the spectrum of learned features and asymptotic generalization capabilities

    Yatin Dandi, Luca Pesce, Hugo Cui, Florent Krzakala, Yue Lu, and Bruno Loureiro. A random matrix theory perspective on the spectrum of learned features and asymptotic generalization capabilities. In The 28th International Conference on Artificial Intelligence and Statistics (A...

  13. [17]

    Mutual information for symmetric rank-one matrix estimation: A proof of the replica formula

    Mohamad Dia, Nicolas Macris, Florent Krzakala, Thibault Lesieur, Lenka Zdeborov´ a, et al. Mutual information for symmetric rank-one matrix estimation: A proof of the replica formula. Advances in Neural Information Processing Systems , 29, 2016

  14. [18]

    Fundamental limits of detection in the spiked Wigner model

    Ahmed El Alaoui, Florent Krzakala, and Michael Jordan. Fundamental limits of detection in the spiked Wigner model. Annals of Statistics , 48(2):863–885, 2020

  15. [19]

    Spectral properties of elementwise-transformed spiked matrices

    Michael J Feldman. Spectral properties of elementwise-transformed spiked matrices. SIAM Journal on Mathematics of Data Science , 7(2):542–571, 2025

  16. [20]

    The largest eigenvalue of rank one deformation of large Wigner matrices

    Delphine F´ eral and Sandrine P´ ech´ e. The largest eigenvalue of rank one deformation of large Wigner matrices. Communications in Mathematical Physics , 272:185–228, 2007

  17. [21]

    Spectral phase transitions in non-linear wigner spiked models

    Alice Guionnet, Justin Ko, Florent Krzakala, Pierre Mergny, and Lenka Zdeborov´ a. Spectral phase transitions in non-linear wigner spiked models. arXiv:2310.14055, 2023

  18. [22]

    On the distribution of the largest eigenvalue in principal components analysis

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

  19. [23]

    Testing in high-dimensional spiked models

    Iain M Johnstone and Alexei Onatski. Testing in high-dimensional spiked models. The Annals of Statistics , 48(3):1231–1254, 2020

  20. [24]

    Detection of signal in the spiked rectangular models

    Ji Hyung Jung, Hye Won Chung, and Ji Oon Lee. Detection of signal in the spiked rectangular models. In International Conference on Machine Learning , pages 5158–5167. PMLR, 2021

  21. [25]

    Detection problems in the spiked random matrix models

    Ji Hyung Jung, Hye Won Chung, and Ji Oon Lee. Detection problems in the spiked random matrix models. IEEE Transactions on Information Theory , 70(10):7194–7231, 2024

  22. [27]

    The outliers of a deformed Wigner matrix

    Antti Knowles and Jun Yin. The outliers of a deformed Wigner matrix. The Annals of Probability, 42(5):1980–2031, 2014

  23. [28]

    Determining the number of components in a factor model from limited noisy data

    Shira Kritchman and Boaz Nadler. Determining the number of components in a factor model from limited noisy data. Chemometrics and Intelligent Laboratory Systems , 94(1):19–32, 2008

  24. [29]

    Mutual information in rank-one matrix estimation

    Florent Krzakala, Jiaming Xu, and Lenka Zdeborov´ a. Mutual information in rank-one matrix estimation. In 2016 IEEE Information Theory Workshop (ITW) , pages 71–75. IEEE, 2016

  25. [30]

    Spqr: controlling q-ensemble independence with spiked random model for reinforcement learning

    Dohyeok Lee, Seungyub Han, Taehyun Cho, and Jungwoo Lee. Spqr: controlling q-ensemble independence with spiked random model for reinforcement learning. Advances in Neural Information Processing Systems, 36:65224–65251, 2023. 42

  26. [31]

    Demys- tifying disagreement-on-the-line in high dimensions

    Donghwan Lee, Behrad Moniri, Xinmeng Huang, Edgar Dobriban, and Hamed Hassani. Demys- tifying disagreement-on-the-line in high dimensions. In International Conference on Machine Learning, pages 19053–19093. PMLR, 2023

  27. [32]

    Local law and Tracy–Widom limit for sparse random matrices

    Ji Oon Lee and Kevin Schnelli. Local law and Tracy–Widom limit for sparse random matrices. Probability Theory and Related Fields , 171:543–616, 2018

  28. [34]

    Fundamental limits of symmetric low-rank matrix estimation

    Marc Lelarge and L´ eo Miolane. Fundamental limits of symmetric low-rank matrix estimation. Probability Theory and Related Fields , 173(3-4):859–929, 2019

  29. [35]

    MMSE of probabilistic low-rank matrix estimation: Universality with respect to the output channel

    Thibault Lesieur, Florent Krzakala, and Lenka Zdeborov´ a. MMSE of probabilistic low-rank matrix estimation: Universality with respect to the output channel. In 2015 53rd Annual Allerton Conference on Communication, Control, and Computing (Allerton) , pages 680–687. IEEE, 2015

  30. [36]

    Fundamental limits of non-linear low-rank matrix estimation

    Pierre Mergny, Justin Ko, Florent Krzakala, and Lenka Zdeborov´ a. Fundamental limits of non-linear low-rank matrix estimation. In The Thirty Seventh Annual Conference on Learning Theory, pages 3873–3873. PMLR, 2024

  31. [37]

    Precise error rates for computationally efficient testing

    Ankur Moitra and Alexander S Wein. Precise error rates for computationally efficient testing. The Annals of Statistics , 53(2):854–878, 2025

  32. [38]

    Approximate message passing with spectral initialization for generalized linear models

    Marco Mondelli and Ramji Venkataramanan. Approximate message passing with spectral initialization for generalized linear models. In International Conference on Artificial Intelligence and Statistics , pages 397–405. PMLR, 2021

  33. [39]

    Signal-plus-noise decomposition of nonlinear spiked random matrix models

    Behrad Moniri and Hamed Hassani. Signal-plus-noise decomposition of nonlinear spiked random matrix models. arXiv:2405.18274, 2024

  34. [40]

    A theory of non-linear feature learning with one gradient step in two-layer neural networks

    Behrad Moniri, Donghwan Lee, Hamed Hassani, and Edgar Dobriban. A theory of non-linear feature learning with one gradient step in two-layer neural networks. InInternational Conference on Machine Learning , pages 36106–36159. PMLR, 2024

  35. [41]

    On the limitation of spectral methods: From the gaussian hidden clique problem to rank-one perturbations of gaussian tensors

    Andrea Montanari, Daniel Reichman, and Ofer Zeitouni. On the limitation of spectral methods: From the gaussian hidden clique problem to rank-one perturbations of gaussian tensors. In Advances in Neural Information Processing Systems , pages 217–225, 2015

  36. [42]

    Gradient-based feature learning under structured data

    Alireza Mousavi-Hosseini, Denny Wu, Taiji Suzuki, and Murat A Erdogdu. Gradient-based feature learning under structured data. Advances in Neural Information Processing Systems , 36:71449–71485, 2023

  37. [43]

    Testing hypotheses about the number of factors in large factor models

    Alexei Onatski. Testing hypotheses about the number of factors in large factor models. Econometrica, 77(5):1447–1479, 2009. 43

  38. [44]

    The largest eigenvalue of small rank perturbations of Hermitian random matrices

    Sandrine P´ ech´ e. The largest eigenvalue of small rank perturbations of Hermitian random matrices. Probability Theory and Related Fields , 134:127–173, 2006

  39. [45]

    Nonlinear random matrix theory for deep learning

    Jeffrey Pennington and Pratik Worah. Nonlinear random matrix theory for deep learning. Advances in neural information processing systems , 30, 2017

  40. [47]

    On finite rank deformations of Wigner matrices

    Alessandro Pizzo, David Renfrew, and Alexander Soshnikov. On finite rank deformations of Wigner matrices. Annales de l’IHP Probabilit´ es et Statistiques, 49(1):64–94, 2013

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.