Abstract
In this paper, we adopt the eigenvector empirical spectral distribution (VESD) to investigate the limiting behavior of eigenvectors of a large dimensional Wigner matrix \(\mathbf {W}_n.\) In particular, we derive the optimal bound for the rate of convergence of the expected VESD of \(\mathbf{W}_n\) to the semicircle law, which is of order \(O(n^{-1/2})\) under the assumption of having finite 10th moment. We further show that the convergence rates in probability and almost surely of the VESD are \(O(n^{-1/4})\) and \(O(n^{-1/6}),\) respectively, under finite eighth moment condition. Numerical studies demonstrate that the convergence rate does not depend on the choice of unit vector involved in the VESD function, and the best possible bound for the rate of convergence of the VESD is of order \(O(n^{-1/2}).\)
Similar content being viewed by others
Change history
12 September 2019
Unfortunately, due to a technical error, the articles published in issues 60:2 and 60:3 received incorrect pagination. Please find here the corrected Tables of Contents. We apologize to the authors of the articles and the readers.
References
Anderson TW (1963) Asymptotic theory for principal component analysis. Ann Math Stat 34:122–148
Bai ZD (1993) Convergence rate of expected spectral distributions of large random matrices. I. Wigner matrices. Ann Probab 21(2):625–648
Bai ZD, Pan GM (2012) Limiting behavior of eigenvectors of large Wigner matrices. J Stat Phys 146(3):519–549
Bai ZD, Silverstein JW (1998) No eigenvalues outside the support of the limiting spectral distribution of large-dimensional sample covariance matrices. Ann Probab 26(1):316–345
Bai ZD, Silverstein JW (2004) CLT for linear spectral statistics of large-dimensional sample covariance matrices. Ann Probab 32(1A):553–605
Bai ZD, Silverstein JW (2010) Spectral analysis of large dimensional random matrices. Springer series in statistics, vol 20, 2nd edn. Springer, New York
Bai ZD, Yao JF (2005) On the convergence of the spectral empirical process of Wigner matrices. Bernoulli 11(6):1059–1092
Bai ZD, Miao BQ, Tsay JS (1997) A note on the convergence rate of the spectral distributions of large random matrices. Stat Probab Lett 34(1):95–101
Bai ZD, Miao BQ, Pan GM (2007) On asymptotics of eigenvectors of large sample covariance matrix. Ann Probab 35(4):1532–1572
Bai ZD, Hu J, Pan GM, Zhou W (2011) A note on rate of convergence in probability to semicircular law. Electron J Probab 16(88):2439–2451
Bao ZG, Pan GM, Zhou W (2013) Central limit theorem for partial linear eigenvalue statistics of Wigner matrices. J Stat Phys 150(1):88–129
Croux C, Filzmoser P, Fritz H (2013) Robust sparse principal component analysis. Technometrics 55(2):202–214
Erdős L, Schlein B, Yau HT (2009) Local semicircle law and complete delocalization for Wigner random matrices. Commun Math Phys 287(2):641–655
Erdős L, Yau HT, Yin J (2012) Bulk universality for generalized Wigner matrices. Probab Theory Relat Fields 154(1–2):341–407
Götze F, Tikhomirov A (2003) Rate of convergence to the semi-circular law. Probab Theory Relat Fields 127(2):228–276
Götze F, Tikhomirov A (2005) The rate of convergence for spectra of GUE and LUE matrix ensembles. Cent Eur J Math 3(4):666–704
Götze F, Tikhomirov A (2016) Optimal bounds for convergence of expected spectral distribution to the semi-circular law. Probab Theory Relat Fields 165:163–233
Guo WX, Liu XH, Zhang SL (2016) The principal correlation components estimator and its optimality. Stat Pap 57(3):755–779
Heewon P, Sadanori K (2015) Principal component selection via adaptive regularization method and generalized information criterion. Stat Pap 1–14. doi:10.1007/s00362-015-0691-1
Johnstone IM (2001) On the distribution of the largest eigenvalue in principal component analysis. Ann Stat 29(2):295–327
Johnstone IM, Lu AY (2009) On consistency and sparsity for principal component analysis in high dimensions. J Am Stat Assoc 104(486):682–693
Knowles A, Yinm J (2013) Eigenvector distribution of Wigner matrices. Probab Theory Relat Fields 155(3–4):543–582
Li WM (2014) Local expectations of the population spectral distribution of a high-dimensional covariance matrix. Stat Pap 55(2):563–573
Ma ZM (2013) Sparse principal component analysis and iterative thresholding. Ann Stat 41(2):772–801
Paul D (2007) Asymptotics of sample eigenstructure for a large dimensional spiked covariance model. Stat Sin 17(4):1617–1642
Paulo CR, Ara TL (2009) Analysis of an European Union election using principal component analysis. Stat Sin 50(4):895–904
Shen HP, Huang JHZ (2008) Sparse principal component analysis via regularized low rank matrix approximation. J Multivar Anal 99(6):1015–1034
Silverstein JW (1989) On the eigenvectors of large-dimensional sample covariance matrices. J Multivar Anal 30(1):1–16
Silverstein JW (1990) Weak convergence of random functions defined by the eigenvectors of sample covariance matrices. Ann Probab 18(3):1174–1194
Silverstein JW, Bai ZD (1995) On the empirical distribution of eigenvalues of a class of large-dimensional random matrices. J Multivar Anal 54(2):175–192
Tao T, Vu V (2012) Random matrices: universal properties of eigenvectors. Random Matrices Theory Appl 1(1):1150001, 27
Xia NN, Bai ZD (2015) Functional CLT of eigenvectors for large sample covariance matrices. Stat Pap 56(1):23–60
Xia NN, Qin YL, Bai ZD (2013) Convergence rates of eigenvector empirical spectral distribution of large dimensional sample covariance matrix. Ann Stat 41(5):2572–2607
Zou H, Hastie T, Tibshirani R (2006) Sparse principal component analysis. J Comput Graph Stat 15(2):265–286
Acknowledgements
The research was supported by NSFC 11501348, Shanghai Pujiang Program 15PJ1402300, IRTSHUFE and the State Key Program in the Major Research Plan of NSFC 91546202. The research was partially supported by National Natural Science Foundation of China (Grant No. 11571067).
Author information
Authors and Affiliations
Corresponding author
Appendices
Appendix 1
Lemma 8
Under the conditions of Theorem 1, when \(z\in {\mathcal D},\) there exists a constant C, such that for any \(l\ge 1,\)
where \(\Delta =\Vert EF^{\mathbf{W}_n}-F\Vert .\)
Proof
For Stieltjes transform \(s_n(z),\) integration by parts yields
Combined with the fact \(|s(z)|\le 1,\) implies that \(|Es_n(z)|\le C(1+\Delta /v).\) Under finite eighth moment assumption, we have \(E|X_{jk}|^{4l}\le Cn^{l-2}.\) Thus by Lemma 16, it follows that
\(\square \)
Lemma 9
Under the conditions of Theorem 1, for all \(z\in {\mathcal D},\) there exists a large constant \(C>0,\) such that for any \(t>0\) and for all \(k=1,\ldots ,n,\) we have
Proof
Recall that
Define
First if we suppose that for any fixed \(t>0,\) there exists a large constant \(C>0,\) such that the following equation holds,
Then from
we have
where the last inequality comes from the facts that \(|tr(\mathbf{A}_k^{-1}(z)-\mathbf{A}^{-1}(z))|\le v^{-1}\) and \(|z^{-1}|\le v^{-1}.\) If (26) is true, for \(v\ge O(n^{-1/2}),\) we can choose n large enough such that \(2|z^{-1}\beta _n|(nv)^{-1}\le 1/2.\) Then \(|z^{-1}\gamma _k|\le 2|z^{-1}\beta _n|,\) which implies that \(P(|z^{-1}\gamma _k|>C)=o(n^{-t})\) holds.
Similarly, consider \(z^{-1}\alpha _k,\) if \(|z^{-1}\gamma _k||\xi _k|\le 1/2\) holds, then we have
That is, for any \(p\ge 1,\)
Since \(|X_{jk}|\le \varepsilon _n n^{1/4},\) choose \(\eta =\varepsilon _nn^{-1/4}\) in Lemma 20, for \(p\ge \log n,\)
For any fixed \(t>0,\) when n is large enough so that \(\log \varepsilon _n^{-1}>t+1,\) it can be shown that
This shows that Lemma 9 is true if we can prove Eq. (26).
From (8.1.13) and (8.1.18) in Bai and Silverstein (2010), we can see that
which implies that \(|z^{-1}b_n|\le 1.\) Further from
we obtain
if \(|s_n(z)-Es_n(z)|\le 1/2.\) Therefore, by Lemma 19, for any \(l> \max \{1,\,2t\}\) and \(v\ge O(n^{-1/2}),\) we have
This completes the proof. \(\square \)
Lemma 10
Under the conditions of Theorem 1, for any \(l\ge 1\) and \(z\in {\mathcal D},\) there exists a constant \(C>0,\) such that
Proof
By (27), (26), (18) and Lemmas 19 and 18, it follows that
\(\square \)
Lemma 11
Under the conditions of Theorem 1, for \(z\in {\mathcal D},\) we have
Proof
Write \(\mathbf{x}_n^*\mathbf{A}^{-1}(z)\mathbf{x}_n-E\mathbf{x}_n^*\mathbf{A}^{-1}(z)\mathbf{x}_n\) as the sum of a martingale difference sequence. Denote \(E_k(\cdot )\) as the conditional expectation given \(\{X_{ij},\,k<i,\,j\le n\}.\)
By (14) and (12), we further decompose \(\phi _{n1}\) as
Similar to \(\phi _{n1},\) we obtain
Notice that \(\phi _{nij},\,i,\,j=1,\,2\) are martingale difference sequence, for any \(l\ge 1,\) we conclude from Lemmas 13(b) and 16 that
As in dealing with \(E|\phi _{n11}|^{2l},\) similarly we obtain
Since \(\alpha _k=\gamma _k-z^{-1}\alpha _k\gamma _k\xi _k\) and \((E_{k-1}-E_k)\gamma _k=0,\) we have
Again, by Lemmas 13(b) and 9, it follows that
Now, we turn to term \(\phi _{n22}.\) Substitute \(\alpha _k=\gamma _k-z^{-1}\alpha _k\gamma _k\xi _k\) into \(\phi _{n22},\) and note that
We further decompose \(\phi _{n22}\) as
By Lemma 16, it is easy to verify that
Therefore, by Lemma 13(b), it leads to
Combined the above results together, we obtain
Note that
For \(z\in {\mathcal D},\) that is \(v\ge C_0n^{-1/2},\) we can choose \(C_0\) large enough, such that \(\dfrac{C}{n^lv^{2l}}\le 2/3.\) Therefore,
Next, we use induction to treat the term \(E(\mathfrak {I}(s_n^H(z)))^l\) for \(l>1.\)
When \(1/2<l<1,\) applying Lemma 13(a) to \(E|\phi _{nij}|^{2l},\) we have
This shows that for \(1<p<2,\)
Thus the lemma follows for \(1<l<2\) by substituting the result of \(E(\mathfrak {I}(s_n^H(z)))^p\) into (28). Therefore, the lemma holds by using induction and (28). This completes the proof of lemma 11. \(\square \)
Lemma 12
Under the conditions in Theorem 1, for any fixed \(t>0,\)
Proof
For any fixed \(t>0\) and \(x>0,\) by Lemma 17, it follows that
Note that
Therefore, we have
\(\square \)
Appendix 2
Lemma 13
(Burkholder Inequalities) Let \(\{X_k\}\) be a complex martingale difference sequence with respect to the increasing \(\sigma \)-field \(\{\mathcal {F}_k\},\) and let \(\text{ E }_k\) denote the conditional expectation with respect to \(\mathcal {F}_k.\) Then we have
(a) for \(p>1,\)
(b) for \(p \ge 2,\)
where \(K_p\) is a constant which depends upon p only.
Lemma 14
(Lemma 2.6 of Silverstein and Bai 1995) Let \(z\in \mathbb {C}^+\) with \(v=\mathfrak {I}z,\, \mathbf {A}\) and \(\mathbf {B}\) Hermitian and \(\mathbf {r}\in \mathbb {C}^n.\) Then
Lemma 15
((1.15) in Bai and Silverstein 2004) Let \(\mathbf {Y}=(Y_1,\ldots ,Y_n),\) where \(Y_i\)’s are i.i.d. complex random variables with mean zero and variance 1. Let \(\mathbf {A}=(a_{ij})_{n\times n}\) and \(\mathbf {B}=(b_{ij})_{n\times n}\) be complex matrices. Then the following identity holds:
Lemma 16
(Lemma 2.7 of Bai and Silverstein 1998) For \(\mathbf {X}=(X_1,\ldots ,X_n)^{\prime }\) with i.i.d. standardized real or complex entries such that \(\text {E}X_i=0\) and \(\text {E}|X_i|^2=1,\) and for \(\mathbf {C}\) an \(n\times n\) complex matrix, we have, for any \(p\ge 2,\)
where \(K_p\) is a constant which depends upon p only.
Lemma 17
((5.1.16) of Bai and Silverstein 2010) Suppose that the entries of the matrix \(\mathbf {W}_n=\dfrac{1}{\sqrt{n}}(X_{ij})\) are independent (but depend on n) and satisfy
-
(1)
\(\text{ E }X_{jk}=0,\)
-
(2)
\(\text {E}|X_{jk}|^2\le \sigma ^2,\)
-
(3)
\(\text{ E }|X_{jk}|^l\le b(\delta _n\sqrt{n})^{l-3}\) for all \(l\ge 3,\)
where \(\delta _n\rightarrow 0\) and \(b>0.\) Then, for fixed \(\varepsilon >0\) and \(x>0,\)
Lemma 18
(Theorem 8.6 in Bai and Silverstein 2010) Under assumptions in Theorem 1, we have
Lemma 19
(Lemma 8.20 in Bai and Silverstein 2010) If \(|z|<C,\,v\ge O(n^{-1/2})\) and \(l \ge 1,\) then
where \(\Delta =\Vert \text {E}F^{W_n}-F\Vert .\)
Lemma 20
(Lemma 9.1 in Bai and Silverstein 2010) Suppose that \(X_i,\,i=1,\ldots , n,\) are independent, with \(\text{ E }X_i=0,\,\text{ E }|X_i|^2=1,\,\sup \text{ E }|X_i|^4=\nu < \infty \) and \(|X_i|\le \eta \sqrt{n}\) with \(\eta >0.\) Assume that \(\mathbf {A}\) is a complex matrix. Then for any given p such that \(2\le p \le b \log (n\nu ^{-1}\eta ^4 )\) and \(b>1,\) we have
where \(\mathbf {\alpha }=(X_1,\ldots , X_n)^T.\)
Truncation part
We shall prove that the underlying random variables satisfy \(X_{ii}=0\) and for \(i\ne j,\,|X_{ij}|\le \varepsilon _nn^{1/4},\, \text {E}X_{ij}=0\) and \(Var(X_{ij})=1,\) where \(\varepsilon _n\) is a sequence decreasing to 0 and \(\varepsilon _nn^{1/4}\) increasing to infinity.
1.1 Truncation
Under finite 10th moment condition, we can choose \(\varepsilon _n\) decreasing to 0 and \(\varepsilon _nn^{1/4}\) increasing to infinity as \(n\rightarrow \infty ,\) such that
Let \(\widehat{\mathbf {X}}_n\) denote the truncated matrix whose entry on the ith row and jth column is \(X_{ij}\text {I}(|X_{ij}|\le \varepsilon _nn^{1/4}),\, i=1,\ldots ,n,\,j=1,\ldots ,n.\) Define \(\widehat{\mathbf {W}}_n=\dfrac{1}{\sqrt{n}}\widehat{\mathbf {X}}_n.\)
Step 1(a). Truncation for Theorem1.
Step 1(b). Truncation for Theorems2and3.
1.2 Removing diagonal entries
Based on step 1, we can assume that \(\mathbf {W}_n=\dfrac{1}{\sqrt{n}}(X_{ij}\text {I}(|X_{ij}|\le \varepsilon _nn^{1/4}))_{i,j=1}^n.\) Let \(\mathbf {W}_n^0=\dfrac{1}{\sqrt{n}}(X_{ij}\text {I}(|X_{ij}|\le \varepsilon _nn^{1/4})(1-\delta _{ij}))_{i,j=1}^n.\) Here \(\delta _{ij}=1\) or 0 according to \(i=j,\) or \(i\ne j.\)
Step 2(a). Removing diagonal entries for Theorem1.
Step 2(b). Removing diagonal entries for Theorems2and3.
1.3 Centralization
Suppose that \(z\in {\mathcal D}\) and denote \(\widehat{\mathbf {W}}_n=\dfrac{1}{\sqrt{n}}(\widehat{X}_{ij}),\) where
Centralization for Theorem1.
Centralization for Theorems2and3.
1.4 Rescaling
The rescaling procedures for the three theorems are exactly the same, and only eighth moment is required. Thus we treat them uniformly. Write \(\mathbf {W}_n=\dfrac{1}{\sqrt{n}}(\widehat{X}_{ij}),\) where \(\widehat{X}_{ij}=X_{ij}\text {I}(|X_{ij}|\le \varepsilon _nn^{1/4})-\text {E}X_{ij}\text {I}(|X_{ij}|\le \varepsilon _nn^{1/4}).\) And \(\widetilde{\mathbf {W}}_n=\dfrac{1}{\sqrt{n}}(Y_{ij}),\) where \(Y_{ij}=X_{ij}/\sigma _1\) and \(\sigma _1^2=\text {E}|X_{12}\text {I}(|X_{12}|\le \varepsilon _nn^{1/4})-\text {E}X_{12}\text {I}(|X_{12}|\le \varepsilon _nn^{1/4})|^2.\) Notice that \(\sigma _1\le 1,\) and \(\sigma _1\) tends to 1 as n goes to \(\infty .\) For \(z\in {\mathcal D},\) we obtain
Rights and permissions
About this article
Cite this article
Xia, N., Bai, Z. Convergence rate of eigenvector empirical spectral distribution of large Wigner matrices. Stat Papers 60, 983–1015 (2019). https://doi.org/10.1007/s00362-016-0860-x
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00362-016-0860-x