betaFeedback

WorksMathematics

Tracy-Widom law for sample covariance matrices under quadratic form concentration

Abstract

We study sample covariance matrices formed from vectors qi=Σ1/2ui∈RMq_i = Σ^{1/2}u_i \in \mathbb{R}^M, 1≤i≤N1 \leq i \leq N, where u1,…,uNu_1, \ldots, u_N are independent, centered, isotropic random vectors whose coordinates may not be independent. We assume that the random vectors qiq_i satisfy a concentration bound for their quadratic forms qiTAqiq_i^T A q_i. Under this assumption and a regular right edge condition on ΣΣ, we establish edge universality and obtain the Tracy-Widom law for the largest eigenvalue of the sample covariance matrix after appropriate centering and scaling. In particular, for log-concave ensembles, our result removes the unconditionality assumption imposed in existing edge universality results~\cite{BX}, resolving the question that originally motivated our work. The concentration assumption also allows us to extend the result beyond the log-concave class.

Introduction

Background and matrix model.

Let S=(q1,…,qN)∈RM×NS=(q_1,\ldots,q_N)\in\mathbb{R}^{M\times N} be a data matrix whose columns are independent observations, not necessarily identically distributed. We assume that each column is centered with common covariance Σ\Sigma. Thus

Eqi=0,EqiqiT=Σ.(1)\mathbb{E}q_i=0,\qquad\mathbb{E}q_iq_i^T=\Sigma. \tag*{(1)}

We are interested in the normalized covariance matrix formed by the data matrix

H≡HN:=1N∑i=1NqiqiT=XXT,X=(x1,…,xN):=N−1/2S.(2)H\equiv H_N:=\frac{1}{N}\sum_{i=1}^{N}q_iq_i^T=XX^T,\qquad X=(x_1,\ldots,x_N):=N^{-1/2}S. \tag*{(2)}

Throughout the paper, we consider the proportional regime M/N≡y→y∞∈(0,∞)∖{1}M/N\equiv y\to y_\infty\in(0,\infty)\setminus\{1\} as NN tends to infinity.

The main motivation for (2) arises from estimating the population covariance matrix by the sample covariance matrix in high-dimensional statistics and big data applications. In the low dimensional case, under some moment conditions, the sample covariance matrix HH is a consistent estimator of the population covariance matrix Σ\Sigma, i.e., ∥H−Σ∥op→0\lVert H-\Sigma\rVert_{\mathrm{op}}\to0 as N→∞N\to\infty provided that MM is fixed. However, this is no longer the case in the high dimensional regime when MM is proportional to NN. There is an extensive literature discussing this effect. In the case when each column has identity population covariance and independent entries, the Marchenko–Pastur (MP) law [19] shows that the empirical spectral distribution converges weakly to a deterministic limit supported on [λ−∞,λ+∞][\lambda^\infty_-,\lambda^\infty_+] (and possibly {0}\{0\}), where λ±∞=(1±y∞)2\lambda^\infty_\pm=(1\pm\sqrt{y_\infty})^2. The right edge λ+∞\lambda^\infty_+ is the threshold for the largest eigenvalue λ1(H)\lambda_1(H), and λ1(H)→(1+y∞)2\lambda_1(H)\to(1+\sqrt{y_\infty})^2 almost surely under a finite fourth moment condition [24]. Moreover, our primary focus is on the fluctuations of λ1(H)\lambda_1(H) around the edge, which have been shown to asymptotically follow the Tracy–Widom law with proper centering and scaling, for independent Gaussian entries [12, 13] and for independent entries with general distributions [22]. In fact, based on the classical three step strategy [10], the largest eigenvalue of many matrix models is proved to exhibit the Tracy–Widom fluctuations. For instance, we refer the reader to [9, 17, 15, 16, 14, 1, 21, 3, 4, 6, 20, 23].

For a population covariance Σ\Sigma, the MP law is replaced by the deformed MP law. Its regular right edge and the square root decay rate depend on Σ\Sigma and M/NM/N, see [14].

Our first motivation was to remove the additional assumption in proving edge universality for a log-concave ensemble in [5]. Inspired by [2, 22, 11, 18], we further consider the random vectors exhibiting a concentration of quadratic form

∣qiTAqi−Tr⁡(ΣA)∣≺∥A∥HS,(3)\left|q_i^T A q_i-\operatorname{Tr}(\Sigma A)\right|\prec\|A\|_{\mathrm{HS}}, \tag*{(3)}

uniformly over deterministic matrices A∈CM×MA\in\mathbb{C}^{M\times M}. This condition (3) is also achieved by Gaussian vectors and log-concave vectors. We refer to [11, 18] for more examples. Our theorem uses the moment form of the concentration given in (4) for convenience.

Main results.

We denote the largest eigenvalue of HH by λ1(H)\lambda_1(H). Our main result shows that the fluctuation of the largest eigenvalue around a regular right edge is universal across columns with common covariance Σ\Sigma satisfying the following assumption. The Gaussian reference matrix has the same population covariance Σ\Sigma. Its largest eigenvalue is known to follow the Tracy–Widom distribution after proper centering and scaling [14].

Throughout the paper, our basic assumption on the random vectors is as follows.

Assumption 1.1. We make the following three assumptions on the vectors qiq_i.

(i) (On qq) We assume that the random vectors q1,…,qNq_1,\ldots,q_N are independent, but not necessarily identically distributed, and satisfy

Eqi=0,EqiqiT=Σ,1≤i≤N.\mathbb{E}q_i=0,\qquad\mathbb{E}q_iq_i^T=\Sigma,\qquad1\le i\le N.

(ii) (On yy) We assume that M/N≡y→y∞∈(0,∞)∖{1}M/N\equiv y\to y^\infty\in(0,\infty)\setminus\{1\}, as N→∞N\to\infty.

(iii) (Concentration of quadratic forms) For every fixed p≥2p\ge2, there exists a constant Cp>0C_p>0, uniform in ii, MM and NN, such that, for every deterministic symmetric matrix A∈CM×MA\in\mathbb{C}^{M\times M},

∥qiTAqi−Tr⁡(ΣA)∥p≤Cp∥A∥HS,1≤i≤N.(4)\left\|q_i^T A q_i-\operatorname{Tr}(\Sigma A)\right\|_p\le C_p\|A\|_{\mathrm{HS}},\qquad1\le i\le N. \tag*{(4)}

where ∥A∥HS=(Tr⁡A∗A)1/2\|A\|_{\mathrm{HS}}=(\operatorname{Tr} A^*A)^{1/2}.

In particular, we do not assume that the fourth or higher moments agree between different columns. Throughout the paper, our estimates hold uniformly in the column index, with constants depending on the common concentration constants CpC_p.

We next state the assumption on the population covariance. Let σ1≥⋯≥σM\sigma_1\ge\cdots\ge\sigma_M be the eigenvalues of Σ\Sigma and set

f(m):=−1m+1N∑α=1Mσα1+σαm.(5)f(m):=-\frac{1}{m}+\frac{1}{N}\sum_{\alpha=1}^{M}\frac{\sigma_\alpha}{1+\sigma_\alpha m}. \tag*{(5)}

There is a unique m+∈(−σ1−1,0)m_+ \in(-\sigma_1^{-1},0) satisfying f′(m+)=0f'(m_+)=0. Define the right edge and its scaling parameter by

λ+:=f(m+),γ+:=(2f′′(m+))1/3.(6)\lambda_+ := f(m_+), \qquad\gamma_+ := \left(\frac{2}{f''(m_+)}\right)^{1/3}. \tag*{(6)}

Assumption 1.2. (On Σ\Sigma) The matrix Σ=ΣN\Sigma=\Sigma_N is deterministic, real symmetric and positive definite. For some constant τ>0\tau>0, independent of NN, we assume

τIM≤Σ≤τ−1IM,1+σ1m+≥τ.(7)\tau I_M \le\Sigma\le\tau^{-1}I_M, \qquad1+\sigma_1m_+ \ge\tau. \tag*{(7)}

The second condition in (7) was first introduced in [8], and is the regularity condition at the rightmost edge in the anisotropic local law in [14]. Together with the bounds on Σ\Sigma and yy, this condition ensures a uniformly regular right edge separated from the other edges.

Theorem 1.3 (Universality of largest eigenvalue). Suppose that Assumptions 1.1 and 1.2 hold for the independent, not necessarily identically distributed vectors q1,…,qNq_1,\ldots,q_N. Let XwX^w have independent N(0,N−1)N(0,N^{-1}) entries and set W:=Σ1/2Xw(Xw)TΣ1/2W:=\Sigma^{1/2}X^w(X^w)^{T}\Sigma^{1/2}. Then, for every x∈Rx\in\mathbb{R},

lim⁡N→∞{PH(γ+N2/3(λ1(H)−λ+)≤x)−PW(γ+N2/3(λ1(W)−λ+)≤x)}=0.\lim_{N\to\infty}\left\{\mathbb{P}^{H}\left(\gamma_+N^{2/3}(\lambda_1(H)-\lambda_+)\le x\right)-\mathbb{P}^{W}\left(\gamma_+N^{2/3}(\lambda_1(W)-\lambda_+)\le x\right)\right\}=0.

It is known that the largest eigenvalue of the Gaussian sample covariance matrix with a regular right edge has the Tracy–Widom limit [14 Corollary 3.19(i)]. Combining this result with Theorem 1.3, we have the following corollary.

Corollary 1.4. Under the same conditions as Theorem 1.3, we have

γ+N2/3(λ1(H)−λ+)⟶TW1\gamma_+N^{2/3}(\lambda_1(H)-\lambda_+) \longrightarrow\mathrm{TW}_1

in distribution, where TW1\mathrm{TW}_1 denotes the Tracy–Widom distribution.

Remark 1.5. When Σ=IM\Sigma=I_M, we have

m+=−11+y,λ+=(1+y)2,γ+=y1/6(1+y)4/3.m_+=-\frac{1}{1+\sqrt{y}}, \qquad\lambda_+=(1+\sqrt{y})^2, \qquad\gamma_+=\frac{y^{1/6}}{(1+\sqrt{y})^{4/3}}.

The normalization in Corollary 1.4 is therefore

Nλ1(H)−(M+N)2(M+N)(1/M+1/N)1/3.\frac{N\lambda_1(H)-(\sqrt{M}+\sqrt{N})^2}{(\sqrt{M}+\sqrt{N})(1/\sqrt{M}+1/\sqrt{N})^{1/3}}.

The formulation above does not require the empirical eigenvalue distribution of ΣN\Sigma_N to converge.

Proof strategy.

Traditional proofs of edge universality follow the three step strategy developed by the Erdős–Yau school, including local law, dynamical approach, and Green function comparison. We refer to the monograph [10] for a standard structure of the proof. The recent results of [11, 18] provide the local law and spectral rigidity for the present population model under quadratic form concentration.

Following the approach of [22, 5], we consider a small Gaussian perturbation. Let XwX^w be an M×NM \times N matrix with independent N(0,N−1)N(0,N^{-1}) entries, independent of S=(q1,…,qN)S=(q_1,\ldots,q_N). For 0≤t≤τ/20\le t\le\tau/2, set

References

  1. [1]J. Alt, L. Erdős, T. Krüger, and D. Schröder. Correlated random matrices: band rigidity and edge universality. Annals of Probability 48 (2020), 963–1001.DOI
  2. [2]Z. D. Bai and W. Zhou. Large sample covariance matrices without independence structures in columns. Statistica Sinica 18 (2008), 425–442.
  3. [3]Z. Bao, G. Pan, and W. Zhou. Tracy–Widom law for the extreme eigenvalues of sample correlation matrices. Electronic Journal of Probability 17 (2012), no. 88, 1–32. arXiv:1110.5208.arxiv.org/abs/1110.5208
  4. [4]Z. Bao, G. Pan, and W. Zhou. Universality for the largest eigenvalue of sample covariance matrices with general population. Annals of Statistics 43 (2015), 382–421. arXiv:1304.5690.DOI
  5. [5]Z. Bao and X. Xu. Extreme eigenvalues of log-concave ensemble. Annales de l’Institut Henri Poincaré, Probabilités et Statistiques 61 (2025), no. 1, 155–184. arXiv:2212.11634. Numbered references to results in this paper use the arXiv version.arxiv.org/abs/2212.11634
  6. [6]X. Ding and F. Yang. A necessary and sufficient condition for edge universality at the largest singular values of covariance matrices. Annals of Applied Probability 28 (2018), 1679–1738.DOI
  7. [7]X. Ding and F. Yang. Tracy–Widom distribution for heterogeneous Gram matrices with applications in signal detection. IEEE Transactions on Information Theory 68 (2022), 6682–6715. arXiv:2008.04166.arxiv.org/abs/2008.04166
  8. [8]N. El Karoui. Tracy–Widom limit for the largest eigenvalue of a large class of complex sample covariance matrices. Annals of Probability 35 (2007), no. 2, 663–714. doi:10.1214/009117906000000917.DOI
  9. [9]L. Erdős, B. Schlein, and H.-T. Yau. Rigidity of eigenvalues of generalized Wigner matrices. Advances in Mathematics 229 (2012), 1435–1515.arxiv.org/abs/1007.4652
  10. [10]L. Erdős and H.-T. Yau. A Dynamical Approach to Random Matrix Theory. Courant Lecture Notes in Mathematics 28. Courant Institute of Mathematical Sciences, New York: American Mathematical Society, Providence, RI, 2017. doi:10.1090/cln/028.DOI
  11. [11]Z. Fan, R. Ma, E. Paquette, and Z. Wang. Anisotropic local law for non-separable sample covariance matrices. Preprint (2026). arXiv:2602.17960v1.arxiv.org/abs/2602.17960
  12. [12]K. Johansson. Shape fluctuations and random matrices. Communications in Mathematical Physics 209 (2000), 437–476. arXiv:math/9903134.DOI
  13. [13]I. M. Johnstone. On the distribution of the largest eigenvalue in principal components analysis. Annals of Statistics 29 (2001), 295–327. doi:10.1214/aos/1009210544.DOI
  14. [14]A. Knowles and J. Yin. Anisotropic local laws for random matrices. Probability Theory and Related Fields 169 (2017), 257–352. arXiv:1410.3516.arxiv.org/abs/1410.3516
  15. [15]B. Landon and H.-T. Yau. Edge statistics of Dyson Brownian motion. Preprint (2017). arXiv:1712.03881.arxiv.org/abs/1712.03881
  16. [16]J. O. Lee and K. Schnelli. Tracy–Widom distribution for the largest eigenvalue of real sample covariance matrices with general population. Annals of Applied Probability 26 (2016), 3786–3839. arXiv:1409.4979.arxiv.org/abs/1409.4979
  17. [17]J. O. Lee and J. Yin. A necessary and sufficient condition for edge universality of Wigner matrices. Duke Mathematical Journal 163 (2014), 117–173.DOI
  18. [18]R. Ma and T. Misiakiewicz. The anisotropic local law for sample covariance matrices under quadratic-form concentration. Preprint (2026). arXiv:2609.09440v1.arxiv.org/abs/2609.09440
  19. [19]V. A. Marchenko and L. A. Pastur. Distribution of eigenvalues for some sets of random matrices. Mathematics of the USSR-Sbornik 1 (1967), 457–483. doi:10.1070/SM1967v001n04ABEH001994.
  20. [20]S. Péché. Universality results for the largest eigenvalues of some sample covariance matrix ensembles. Probability Theory and Related Fields 143 (2009), 481–516. arXiv:0705.1701.arxiv.org/abs/0705.1701
  21. [21]N. S. Pillai and J. Yin. Edge universality of correlation matrices. Annals of Statistics 40 (2012), 1737–1763. arXiv:1112.2381.DOI
  22. [22]N. S. Pillai and J. Yin. Universality of covariance matrices. Annals of Applied Probability 24 (2014), 935–1001. arXiv:1110.2501.DOI
  23. [23]A. Soshnikov. A note on universality of the distribution of the largest eigenvalues in certain sample covariance matrices. Journal of Statistical Physics 108 (2002), 1033–1056. arXiv:math/0104113.
  24. [24]Y. Q. Yin, Z. D. Bai, and P. R. Krishniah. On the limit of the largest eigenvalue of the large dimensional sample covariance matrix. Probability Theory and Related Fields 78 (1988), 509–521. doi:10.1007/BF00353874.DOI

Paper details

vlnc:W_3v6pucwj
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC0-1.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

GPT-5.6, GPT-6

View statement
Looking for related papers…
Contents