Abstract
Let \(\mathbf {X}=(X_{jk})_{j,k=1}^n\) denote a Hermitian random matrix with entries \(X_{jk}\), which are independent for \(1\le j\le k\le n\). We consider the rate of convergence of the empirical spectral distribution function of the matrix \(\mathbf {X}\) to the semi-circular law assuming that \(\mathbf{E}X_{jk}=0\), \(\mathbf{E}X_{jk}^2=1\) and that
and
By means of a recursion argument it is shown that the Kolmogorov distance between the expected spectral distribution of the Wigner matrix \(\mathbf {W}=\frac{1}{\sqrt{n}}\mathbf {X}\) and the semicircular law is of order \(O(n^{-1})\).
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
Consider a family \(\mathbf {X} = \{X_{jk}\}\), \(1 \le j \le k \le n\), of independent real random variables defined on some probability space , for any \(n\ge 1\). Assume that \(X_{jk} = X_{kj}\), for \(1 \le k < j \le n\), and introduce the symmetric matrices
The matrix \(\mathbf {W}\) has a random spectrum \(\{\lambda _1,\dots ,\lambda _n\}\) and an associated spectral distribution function \(\mathcal {F}_{n}(x) = \frac{1}{n}\ \mathrm{card}\,\{j \le n: \lambda _j \le x\}, x \in {\mathbb {R}}\). Averaging over the random values \(X_{ij}(\omega )\), define the expected (non-random) empirical distribution functions \( F_{n}(x) = \mathbf{E}\,\mathcal {F}_{n}(x)\). Let \(G(x)\) denote the semi-circular distribution function with density \(g(x)=G'(x)=\frac{1}{2\pi }\sqrt{4-x^2}\mathbb {I}_{[-2,2]}(x)\), where \(\mathbb {I}_{[a,b]}(x)\) denotes the indicator-function of the interval \([a,b]\). The rate of convergence to the semi-circular law has been studied by several authors. We proved in [13] that the Kolmogorov distance between \(\mathcal {F}_n(x)\) and the distribution function \(G(x)\), \(\Delta _n^*:=\sup _x|\mathcal {F}_n(x)-G(x)|\) is of order \(O_P(n^{-\frac{1}{2}})\) (i.e. \(n^{\frac{1}{2}}\Delta _n^*\) is bounded in probability). Bai et al. [1, 2] and Girko [8] showed that \(\Delta _n:=\sup _x| F_n(x)-G(x)|=O(n^{-\frac{1}{2}})\). Bobkov, Götze and Tikhomirov [4] proved that \(\Delta _n\) and \(\mathbf{E}\Delta _n^*\) have order \(O(n^{-\frac{2}{3}})\) assuming a Poincaré inequality for the distribution of the matrix elements. For the Gaussian Unitary Ensemble respectively for the Gaussian Orthogonal Ensemble, see [12] respectively [21], it has been shown that \(\Delta _n=O(n^{-1})\). Denote by \(\gamma _{n1}\le \cdots \le \gamma _{nn}\), the quantiles of \(G\), i.e. \(G(\gamma _{nj})=\frac{j}{n}\), and introduce the notation \({\text {llog}}_n:=\log \log n\). Erdös et al. [6, 7] showed, for matrices with elements \(X_{jk}\) which have a uniformly sub exponential decay, i.e.
for some \(\varkappa >0\), \(A>0\) and for any \(t\ge 1\), the following result
for \(n\) large enough. It is straightforward to check that this bound implies that
From the last inequality it follows that \(\mathbf{E}\Delta _n^*\le Cn^{-1}(\log n)^{C {\text {llog}}_n}\). Similar results were obtained in [20], Theorem 32], assuming additionally that the distributions of the entries of matrices have vanishing third moment.
In this paper we derive the optimal bound for the rate of convergence of the expected spectral distribution to the semi-circular law. Using arguments similar to those used in [17] we provide a self-contained proof based on recursion methods developed in the papers of Götze and Tikhomirov [9, 13] and [22]. It follows from the results of Gustavsson [14] that the best possible bound in the Gaussian case for the rate of convergence in probability is \(O(n^{-1}\sqrt{\log n})\). The best possible bound for \(\Delta _n\) is of order \(O(n^{-1})\). For Gaussian matrices such bounds were obtained in [12] and [21]. Our setup includes the case that the distributions of \(X_{jk}=X_{jk}^{(n)}\) may depend on \(n\). In the following we shall investigate the rate of convergence of expected spectral distribution function \(F_n(x)=\mathbf{E}\mathcal {F}_n(x)\) to the semi-circular distribution function by estimating the quantity \(\Delta _n\). The main result of this paper is the following
Theorem 1.1
Let \(\mathbf{E}X_{jk}=0\), \(\mathbf{E}X_{jk}^2=1\). Assume that
Assume as well that there exists a constant \(D_0\) such that for all \(n\ge 1\)
Then, there exists a positive constant \(C=C(D_0,\mu _4)\) depending on \(D_0\) and \(\mu _4\) only such that
Corollary 1.1
Let \(\mathbf{E}X_{jk}=0\), \(\mathbf{E}X_{jk}^2=1\). Assume that
Then, there exists a positive constant \(C=C(\mu _8)\) depending on \(\mu _8\) only such that
Remark 1.2
Note that the bound (1.6) in Theorem 1.1 and the bound (1.8) in the Corollary 1.1 are not improvable and coincide with the corresponding bounds in the Gaussian case.
We state here as well the results for the Stieltjes transform of the expected spectral distribution of the matrix \(\mathbf {W}\). Let \(\mathbf {R}\) denote the resolvent matrix of the matrix \(\mathbf {W}\),
Here and in what follows \(\mathbf {I}\) denotes the unit matrix of corresponding dimension. For any distribution function \(F(x)\) we define the Stieltjes transform \(s_F(z)\), for \(z=u+iv\) with \(v>0\), via formula
Denote by \(m_n(z)\) the Stieltjes transform of the distribution function \(\mathcal {F}_n(x)\). It is a well-known fact that
By \(s(z)\) we denote the Stieltjes transform of the semi-circular law,
The Stieltjes transform of the semi-circular distribution satisfies the equation
(see, for example, [13], equality (4.20)]).
Introduce for \(z=u+iv\) and a positive constant \(A_0>0\)
For any \(0<\varepsilon <\frac{1}{2}\), and \(A_0>0\), define a region \(\mathbb {G}=\mathbb {G}(A_0,n,\varepsilon )\subset \mathbb {C}_+\), by
Let \(a>0\) be a positive number such that
We prove the following result.
Theorem 1.3
Let \(\frac{1}{2}>\varepsilon >0\) be a sequence of positive numbers in (1.11) such that
Assuming the conditions of Theorem 1.1, there exists a positive constant \(C=C(D_0,A_0,\mu _4)\) depending on \(D\), \(A_0\) and \(\mu _4\) only, such that, for \(z\in \mathbb {G}\)
1.1 Sketch of the proof
1. We start with an estimate of the Kolmogorov-distance to the Wigner distribution via an integral over the difference of the corresponding Stieltjes transforms along a contour in the upper half-plane using a smoothing inequality (2.1) and Cauchy’s formula developed by the authors in [11]. The resulting bound (2.3) involves an integral over a segment at a fixed distance, say \(V=4\), from the real axis and a segment \(u+i A_0 n^{-1}(2-\left| u\right| )^{-\frac{1}{2}}, \; |u|\le 2\) at a distance of order \(n^{-1}\) but avoiding to come close to the endpoints \(\pm 2\) of the support. These segments are part of the boundary of an \(n\)-dependent region \(\mathbb {G}\) where bounds of Stieltjes transforms are needed. Since the Stieltjes-transform and the diagonal elements \(R_{jj}(z)\) of the resolvent of the Wigner-matrix \(\mathbf {W}\) are uniformly bounded on the segment with \(\mathrm{Im}\,z=V\) by \(1{/}V\) (see Sect. 3.1) proving a bound of order \(O(n^{-1})\) for the latter segment near the \(x\)-axis is the essential problem.
2. In order to investigate this crucial part of the error we start with the 2nd resolvent or self-consistency equation for the expected Stieltjes transform resp. the quantities \(R_{jj}(z)\) of \(\mathbf {W}\) (see (3.2) below) based on the difference of the resolvent of \(\mathbf {W}^{(j)}\) (\(j\)th row and column removed) and \(\mathbf {W}\). For the equivalent representation for the difference of Stieltjes transforms (see (3.3)) we have to show an error bound of order \(O((nv)^{-\frac{3}{2}}|z^2-4|^{-\frac{1}{4}})\) for \(z\in \mathbb {G}\). To prove this bound we use a recursive version of this representations as in (5.32). Obviously bounds for \(\mathbf{E}|R_{jj}|^p\) for \(z=u+iv\) close to the real line are needed for the sufficiently large \(p\), which follow once the error terms \(\varepsilon _j\) are small in the region \(\mathbb {G}\). But proving that \(\mathbf{E}|\varepsilon _j|^p\) is small requires in turn again bounds of \(\mathbf{E}|R_{jj}|^{2p}\).
An approach suggested recently in [17] turns out to be very fruitful in dealing with this recursion problem. Assuming that \(\mathbf{E}|R_{jj}|^{2p}\le C_0^{2p}\) for some \(z=u+iv\), we can show that \(\mathbf{E}|R_{jj}|^p\le C_0^p\) with \(z=u+iv/s_0\) with some fixed scale factor \(s_0>1\). This allows us to prove by induction a bound of type \(\mathbf{E}|R_{jj}|^q\le C_0^q\) for some fixed \(q\) (independent of \(n\)) and \(z=u+iv\) with \(v\ge Cn^{-1}\) starting with \(\mathbf{E}|R_{jj}|^p\le C_0^p\) for \(p=s_0^q\) and \(z=u+iv\) for fixed \(v=4\), say. The latter assumption can be easily verified.
Note that one of the errors, that is \(\varepsilon _{j2}\), in (3.2) is a quadratic form in independent random variables. Thus, in case that \(W\) has entries with exponential or even sub-Gaussian tails, inequalities for quadratic forms of independent random variables, like [11], Lemma 3.8] or [17], Proposition A.1], could be applied.
Assuming eight moments in Corollary 1.1 or four moments and a truncation condition in Theorem 1.1 only, we can’t use these strong tail estimates for quadratic forms anymore. Our solution is an recursive application of Burkholder’s inequality for the \(p\)th moment resulting in a bound involving moments of order \(p/2\) of another quadratic form in independent variables in each step. This is the crucial part of the moment recursion for \(R_{jj}\) described above. Details of this procedure are described in Sects. 5.1 and 5.2.
3. In Sect. 6 we prove a bound for the error \(\mathbf{E}\Lambda _n:=\mathbf{E}(m_n(z)-s(z))\) of the form \( {n^{-1} v^{-\frac{3}{4}}}+(n v)^{-\frac{3}{2}}|z^2-4|^{-\frac{1}{4}}\) which suffices to prove the rate \(O(n^{-1})\) in Theorem 1.1. Here we use a series of martingale-type decompositions to evaluate the expectation \(\mathbf{E}m_n(z)\) combined with the bound \(\mathbf{E}|\Lambda _n|^2 \le C (nv)^{-2}\) of Lemma 7.24 in the Appendix which is again based on a recursive inequality for \(\mathbf{E}|\Lambda _n|^2\) in (7.71). A direct application of this bound to estimate the error terms \(\varepsilon _{j4}\) would result in a less precise bound of order \(O(n^{-1}\log n)\) in Theorem 1.1. Bounds of such type will be shown for the Kolmogorov distance of the random spectral distribution to Wigner’s law in a separate paper. For the expectation we provide sharper bounds in Sect. 6.2 involving \(m'_n(z)\).
4. The necessary auxiliary bounds for all these steps are collected in the Appendix.
2 Bounds for the Kolmogorov distance of spectral distributions via Stieltjes transforms
To bound the error \(\Delta _n\) we shall use an approach developed in previous work of the authors, see [13].
We modify the bound of the Kolmogorov distance between an arbitrary distribution function and the semi-circular distribution function via their Stieltjes transforms obtained in [13], Lemma 2.1]. For \(x\in [-2,2]\) define \(\gamma (x):=2-|x|\). Given \(\frac{1}{2}>\varepsilon >0\) introduce the interval \(\mathbb {J}_{\varepsilon }=\{x\in [-2,2]:\, \gamma (x)\ge \varepsilon \}\) and \(\mathbb {J}'_{\varepsilon }=\mathbb {J}_{\varepsilon /2}\). For a distribution function \(F\) denote by \(S_F(z)\) its Stieltjes transform.
Proposition 2.1
Let \(v>0\) and \(a>0\) and \(\frac{1}{2}>\varepsilon >0\) be positive numbers such that
and
If \(G\) denotes the distribution function of the standard semi-circular law, and \(F\) is any distribution function, there exist some absolute constants \(C_1\) and \(C_2\) such that
Remark 2.2
For any \(x\in \mathbb {J}_{\varepsilon }\) we have \(\gamma =\gamma (x)\ge \varepsilon \) and according to condition (2.2), \(\frac{av}{\sqrt{\gamma }}\le \frac{\varepsilon }{2}\).
For a proof of this Proposition see [11], Proposition 2.1].
Lemma 2.1
Under the conditions of Proposition 2.1, for any \(V>v\) and \(0<v\le \frac{\varepsilon ^{3/2}}{2a}\) and \(v'=v/\sqrt{\gamma }, \gamma = 2-|x|\), \(x\in {\mathbb {J}}'_{\varepsilon }\) as above, the following inequality holds
Proof
Let \(x\in \mathbb {J}'_{\varepsilon }\) be fixed. Let \(\gamma =\gamma (x)\). Put \(z=u+iv'\). Since \(v'=\frac{v}{\sqrt{\gamma }}\le \frac{\varepsilon }{2a}\), see (2.2), we may assume without loss of generality that \(v'\le 4\) for \(x\in \mathbb {J}'_{\varepsilon }\). Since the functions \(S_F(z)\) and \(S_G(z)\) are analytic in the upper half-plane, it is enough to use Cauchy’s theorem. We can write for \(x\in \mathbb {J}'_{\varepsilon }\)
By Cauchy’s integral formula, we have
Denote by \(\xi \text { (resp. }\eta )\) a random variable with distribution function \(F(x)\) (resp. \(G(x)\)). Then we have
for any \(v'\le u\le V\). Similarly,
These inequalities imply that
which completes the proof. \(\square \)
Combining the results of Proposition 2.1 and Lemma 2.1, we get
Corollary 2.2
Under the conditions of Proposition 2.1 the following inequality holds
where \(v'=\frac{v_0}{\sqrt{\gamma }}\) with \(\gamma =2-|x|\) and \(C_1,C_2 >0\) denote absolute constants.
3 Proof of Theorem 1.1
Proof
We shall apply Corollary 2.2 to prove the Theorem 1.1. We choose \(V=4\) and \(v_0\) as defined in (1.10) and use the quantity \(\varepsilon =(2av_0)^{\frac{2}{3}}\).
3.1 Estimation of the first integral in (2.3) for \(V=4\)
Denote by \(\mathbb {T}=\{1,\ldots ,n\}\). In the following we shall systematically use for any \(n\times n \) matrix \(\mathbf {W}\) together with its resolvent \(\mathbf {R}\), its Stieltjes transform \(m_n\) etc. the corresponding quantities \(\mathbf {W}^{(\mathbb {A})}\), its resolvent \(\mathbf {R}^{(\mathbb {A})}\) and its Stieltjes transform \(m_n^{(\mathbb {A})}\) for the corresponding sub matrix with entries \(X_{jk}, j, k \not \in \mathbb {A}\), \(\mathbb {A} \subset \mathbb {T}=\{1,\ldots ,n\}\). Let \(\mathbb {T}_{\mathbb {A}}=\mathbb {T}{\setminus }\mathbb {A}\). Observe that
By \(\mathfrak {M}^{(\mathbb {A})}\) we denote the \(\sigma \)-algebra generated by \(X_{lk}\) with \(l,k\in \mathbb {T}_{\mathbb {A}}\). If \(\mathbb {A}=\emptyset \) we shall omit the set \(\mathbb {A}\) as exponent index.
We shall use the representation
(see, for example, [13], equality (4.6)]). We may rewrite it as follows
where \(\varepsilon _j:=\varepsilon _{j1}+\varepsilon _{j2}+ \varepsilon _{j3}+ \varepsilon _{j4}\) with
Let
Summing equality (3.2) in \(j=1,\ldots ,n\) and solving with respect \(\Lambda _n\), we get
where
Obvious bounds like \(|z+s(z)|\ge 1\), \(|\lambda _j-z|^{-1}\le v^{-1}\), \(\max \{|R_{jj}^{(\mathbb {J})}(z)|,\,|m_n^{(\mathbb {J})}(z)|\}\le v^{-1}\), imply that for \(V=4\) and for any \(\mathbb {J}\subset \mathbb {T}\),
and therefore,
Using equality (3.3), we may write
We use that \(\mathbf{E}\{\varepsilon _{j\nu }|\mathfrak {M}^{(j)}\}=0\), for \(\nu =1,2,3\) and obtain
Thus, according to inequalities (3.4), Lemmas 7.8, 7.9, 7.10, 7.12 in the Appendix and Eq. (1.9), we obtain
where \(C\) depends on \(\mu _4\) only. For \(\nu =4\), Lemma 7.12 in the Appendix, inequality (3.4) and relation (1.9) yield
with some absolute constant \(C\). Furthermore, applying the Cauchy–Schwartz inequality and inequality (3.4) and relation (1.9), we get
We may rewrite the representation (3.2) using \(\Lambda _n=m_n(z)-s(z)\) and (1.9) as (compare (3.1))
Applying representations (3.3) and (3.7) together with (3.4) and \(|R_{jj}|\le \frac{1}{4}\), we obtain
Combining inequalities (3.6) and (3.8), we get
Applying now Lemmas 7.8, 7.9, 7.10 and 7.12, we get
Inequality (3.5) and (3.10) together imply
Consider now the integral
for \(V=4\). Using inequality (3.11), we have
Finally, we note that
Therefore,
3.2 Estimation of the second integral in (2.3)
To finish the proof of Theorem 1.1 we need to bound the second integral in (2.3) for \(z\in \mathbb {G}\) and \(v_0=A_0n^{-1}\), where \(\varepsilon =(2av_0)^{\frac{2}{3}}\) is defined in such a way that condition (2.2) holds. We shall use the results of Theorem 1.3. According to these results we have, for \(z\in \mathbb {G}\),
We have
After integrating we get
Inequalities (3.13) and (3.15) complete the proof of Theorem 1.1. Thus Theorem 1.1 is proved. \(\square \)
4 The proof of Corollary 1.1
To prove the Corollary 1.1 we consider truncated random variables \(\widehat{X}_{jl}\) defined by
Let \(\widehat{\mathcal {F}}_n(x)\) denote the empirical spectral distribution function of the matrix \(\widehat{\mathbf {W}}=\frac{1}{\sqrt{n}}(\widehat{X}_{jl})\).
Lemma 4.1
Assuming the conditions of Theorem 1.1 there exists a constant \(C>0\) depending on \(\mu _8\) only such that
Proof
We shall use the rank inequality of Bai. See [3], Theorem A.43, p. 503]. According this inequality
Observing that the rank of a matrix is not larger then numbers of its non-zero entries, we may write
Thus, the Lemma is proved. \(\square \)
Note that in the bound of the first integral in (2.3) we used the condition (1.4) only. We shall compare the Stieltjes transform of the matrix \(\widehat{\mathbf {W}}\) and the matrix obtained from \(\widehat{\mathbf {W}}\) by centralizing and normalizing its entries. Introduce \(\widetilde{X}_{jk}=\widehat{X}_{jk}-\mathbf{E}\widehat{X}_{jk}\) and \(\widetilde{\mathbf {W}}=\frac{1}{\sqrt{n}}(\widetilde{X}_{jk})_{j,k=1}^n\). We normalize the r.v.’s \(\widetilde{X}_{jk}\). Let \(\sigma _{jk}^2=\mathbf{E}|\widetilde{X}_{jk}|^2\). We define the r.v.’s \(\breve{X}_{jk}=\sigma _{jk}^{-1}\widetilde{X}_{jk}\). Finally, let \(\breve{m}_n(z)\) (resp. \(\widehat{m}_n(z)\), \(\widetilde{m}_n(z)\)) denote the Stieltjes transform of the empirical spectral distribution function of the matrix \(\breve{\mathbf {W}}= \frac{1}{\sqrt{n}}(\breve{X}_{jk})_{j,k=1}^n\) (resp. \(\widehat{\mathbf {W}}\), \(\widetilde{\mathbf {W}}\)).
Remark 4.1
Note that
for some absolute constant \(D_1\). That means that the matrix \(\breve{\mathbf {W}}\) satisfies the conditions of Theorem 1.3.
Lemma 4.2
There exists some absolute constant \(C\) depending on \(\mu _8\) such that
Proof
Note that
Therefore,
Using the simple inequalities \(|\mathrm{Tr}\,\mathbf {A}\mathbf {B}|\le \Vert \mathbf {A}\Vert _2\Vert \mathbf {B}\Vert _2\) and \(\Vert \mathbf {A}\mathbf {B}\Vert _2\le \Vert \mathbf {A}\Vert \Vert \mathbf {B}\Vert _2\), we get
Furthermore, we note that,
and
Since
therefore
Applying Lemma 7.6 inequality (7.11) in the Appendix we get
Using now inequalities (4.6) and (4.7), we obtain
According Remark 4.1, we may apply Corollary 5.14 in Sect. 5 with \(q=1\) to prove the claim. Thus, Lemma 4.2 is proved. \(\square \)
Lemma 4.3
For some absolute constant \(C>0\) we have
Proof
Similar to (4.3), we write
This yields
Furthermore, we note that, by definition (4.1) and condition (1.7), we have
Applying (Lemma 7.6 inequality (7.11) in the Appendix) and inequality (4.9), we obtain using \(\Vert \widehat{\mathbf {R}}\Vert \le v^{-1}\),
By Lemma 4.2,
for some constant \(C\) depending on \(\mu _8\) and \(A_0\). According to Corollary 5.14 in Sect. 5.2 with \(q=1\)
with a constant \(C\) depending on \(\mu _4\), \(D_0\). Using these inequalities, we get
Thus Lemma 4.3 is proved. \(\square \)
Corollary 4.4
Assuming the conditions of Corollary 1.1, we have for \(z\in \mathbb {G}\),
Proof
The proof immediately follows from the inequality
Lemmas 4.2 and 4.3 and Theorem 1.3. \(\square \)
The proof of Corollary 1.1 follows now from Lemma 4.1, Corollary 2.2, inequality (3.13) and inequality
5 Resolvent matrices and quadratic forms
The crucial problem in the proof of Theorem 1.3 is the following bound for any \(z\in \mathbb {G}\)
for \(j=1,\ldots ,n\) and some absolute constant \(C>0\). To prove this bound we use an approach similar to the proof of Lemma 3.4 in [17]. In order to arrive at our goal we need additional bounds of quadratic forms of type
To prove this bound we recurrently use Rosenthal’s and Burkholder’s inequalities.
5.1 The key lemma
In this Section we provide auxiliary lemmas needed for the proof of Theorem 1.1.
For any \(\mathbb {J}\subset \mathbb {T}\) introduce \(\mathbb {T}_{\mathbb {J}}=\mathbb {T}{\setminus }\mathbb {J}\). We introduce the quantity, for some \(\mathbb {J}\subset \mathbb {T}\),
By Lemma 7.6, inequality (7.12) in the Appendix, we have
Furthermore, introduce the quantities
where, \(a^{(\mathbb {J},k,0)}_{qr}\) are defined recursively via
Using these notations we have
Lemma 5.1
Under the conditions of Theorem 1.1 we have
Moreover,
Proof
We apply Hölder’s inequality and obtain
Summing in \(q\) and \(r\), (5.5) and (5.6) follow. \(\square \)
Corollary 5.2
Under the conditions of Theorem 1.1 we have
and
Proof
By definition of \(a_{qr}^{(\mathbb {J},k,0)}\), see (5.3), applying (Lemma 7.6 inequality (7.11) in the Appendix), we get
and by definition (5.3),
The general case follows now by induction in \(\nu \), Lemma 5.1, and Lemma 7.6 inequality (7.12) in the Appendix. \(\square \)
Corollary 5.3
Under the conditions of Theorem 1.1 we have
Proof
The result immediately follows from the definition of \(a_{rr}^{(k,\nu )}\) and Corollary 5.2.
\(\square \)
In what follows we shall use the notations
Let \(s_0\) denote some fixed number (for instance \(s_0=2^8\)). Let \(A_1\) be a constant (to be chosen later) and \(0<v_1\le 4\) a constant such that \(v_0=A_0n^{-1}\le v_1\) for all \(n\ge 1\).
Lemma 5.4
Assuming the conditions of Theorem 1.1 and for \(p\le A_1(nv)^{\frac{1}{4}}\)
we have for \(v\ge v_1/s_0\) and \(p\le A_1(nv)^{\frac{1}{4}}\), and \(k\in \mathbb {T}_{\mathbb {J}}\)
Proof
Using the representation (5.4) and the triangle inequality, we get
Let \(\mathfrak {M}^{(\mathbb {A})}\) denote the \(\sigma \)-algebra generated by r.v.’s \(X_{j,l}\) for \(j,l\in \mathbb {T}_{\mathbb {A}}\), for any set \(\mathbb {A}\). Conditioning on \(\mathfrak {M}^{(\mathbb {J},k)}\) (\(\mathbb {A}=\mathbb {J}\cup \{k\}\)) and applying Rosenthal’s inequality (see Lemma 7.1), we get
where \(C_1\) denotes the absolute constant in Rosenthal’s inequality. By Remark 4.1, we get
Analogously conditioning on \(\mathfrak {M}^{(\mathbb {J},k)}\) and applying Burkholder’s inequality (see Lemma 7.3), we get
where \(C_2\) denotes the absolute constant in Burkholder’s inequality. Conditioning again on \(\mathfrak {M}^{(\mathbb {J},k)}\) and applying Rosenthal’s inequality, we obtain
Combining inequalities (5.16) and (5.17), we get
Using Remark 4.1, this implies
Using the definition (5.2) of \(Q^{(\mathbb {J},k)}_{\nu 1}\) and the definition (5.3) of coefficients \( a^{(\mathbb {J},k,\nu +1)}_{rr}\), it is straightforward to check that
Combining (5.15), (5.19) and (5.20), we get by (5.13)
Applying now Lemma 5.1 and Corollaries 5.2 and 5.3, we obtain
where \(C_3=3C_1C_2\). Applying Cauchy–Schwartz inequality, we may rewrite the last inequality in the form
Introduce the notation
We rewrite the inequality (5.23) using \(\Gamma _p(z)\) and the notations of (5.10) as follows
Note that
where \(\Psi ^{(\mathbb {J})}=\mathrm {Im}\;\!m_n^{(\mathbb {J})}(z)+\frac{1}{nv}\). Furthermore,
Without loss of generality we may assume \(p=2^L\) and \(\nu =0,\ldots ,L\). We may write
By induction we get
It is straightforward to check that
Note that, for \(l\ge 1\),
Applying this relation, we get
Note that for \(l\ge 0\), \(\frac{l}{2^l}\le \frac{1}{2}\) and recall that \(p=2^L\). Using this observation, we get
This implies that for \(\frac{4v}{C_3^2p^2}\le \frac{1}{2}\),
Furthermore, by definition of \(T_{\nu ,p}\), we have
Applying Corollary 5.2 and Hölder’s inequality, we get
By condition (5.11), we have
Using this inequality, we get,
Applying relation (5.26), we obtain
Without loss of generality we may assume that \(C_3\ge 2(C_0s_0)\). Then we get
Analogously we get
Combining inequalities (5.25), (5.27), (5.28), (5.29), we finally arrive at
Thus, Lemma 5.4 is proved. \(\square \)
5.2 Diagonal entries of the resolvent matrix
We shall use the representation, for any \(j\in \mathbb {T}_{\mathbb {J}}\),
(see, for example, [13], equality (4.6)]). Recall the following relations (compare (3.2), (3.7))
or
where \(\varepsilon _j^{(\mathbb {J})}:=\varepsilon _{j1}^{(\mathbb {J})}+\varepsilon _{j2}^{(\mathbb {J})}+\varepsilon _{j3}^{(\mathbb {J})}+ \varepsilon _{j4}^{(\mathbb {J})}\) with
Since \(|s(z)|\le 1\), the representation (5.32) yields, for any \(p\ge 1\),
Applying the Cauchy–Schwartz inequality, we get
We shall investigate now the behavior of \(\mathbf{E}|\varepsilon _j^{(\mathbb {J})}|^{2p}\) and \(\mathbf{E}|\Lambda _n^{(\mathbb {J})}|^{2p} \). First we note,
Lemma 5.5
Assuming the conditions of Theorem 1.1 we have, for any \(p\ge 1\), and for any \(z=u+iv\in \mathbb {C}_+\),
Proof
The proof follows immediately from the definition of \(\varepsilon _{j1}\) and condition (1.4). \(\square \)
Lemma 5.6
Assuming the conditions of Theorem 1.3 we have, for any \(p\ge 1\), and for any \(z=u+iv\in \mathbb {C}_+\),
Proof
For a proof of this Lemma see [13], Lemma 4.1]. \(\square \)
Let \(A_1>0\) and \(0\le v_1\le 4\) be a fixed.
Lemma 5.7
Assuming the conditions of Theorem 1.1, and assuming for all \(\mathbb {J}\subset T\) with \(|\mathbb {J}|\le L\) and all \(l\in \mathbb {T}_{\mathbb {J}}\)
we have, for all \(v\ge v_1/s_0\), and for all \(\mathbb {J}\subset \mathbb {T}\) with \(|\mathbb {J}|\le L-1\),
Proof
Recall that \(s_0=2^4\) and note that if \(p\le A_1(nv)^{\frac{1}{4}}\) for \(v\ge v_1/s_0\) then \(q=2p\le A_1(nv)^{\frac{1}{4}}\) for \(v\ge v_1\). Let \(v':=vs_0\). If \(v\ge v_1/s_0\) then \(v'\ge v_1\). We have
We apply now Rosenthal’s inequality for the moments of sums of independent random variables and get
According to inequality (5.37) we may apply Lemma 7.7 and condition (5.36) for \(q=2p\). We get, for \(v\ge v_1/s_0\),
We use as well that by the conditions of Theorem 1.1, \( \mathbf{E}|X_{jl}|^{4p}\le D_0^{4p-4}n^{p-1}\mu _4\), and by Jensen’s inequality, \((\frac{1}{n}\sum _{l\in \mathbb {T}_j}|R_{ll}^{(j)}|^{2})^{p}\le \frac{1}{n}\sum _{l\in \mathbb {T}_j}|R_{ll}^{(j)}|^{2p}\). Thus, Lemma 5.7 is proved. \(\square \)
Lemma 5.8
Assuming the conditions of Theorem 1.1, condition (5.36), for \(v\ge v_1\) and \(p\le A_1(nv)^{\frac{1}{4}}\), we have, for any \(v\ge v_1/s_0\) and \(p\le A_1(nv)^{\frac{1}{4}}\),
Proof
We apply Burkholder’s inequality for quadratic forms. See Lemma 7.3 in the Appendix. We obtain
Using now the quantity \(Q_0^{(\mathbb {J},j)}\) for the first term and Rosenthal’s inequality and condition (1.4) for the second term, we obtain with Lemma 7.6, inequality (7.12), in the Appendix and \(C_3=C_1C_2\)
By Lemma 7.7 and condition (5.36), we get
Applying now Lemma 5.4, we get the claim. Thus, Lemma 5.8 is proved. \(\square \)
Recall that
Lemma 5.9
Assuming the conditions of Theorem 1.1, we have, for \(z=u+iv\) with \(u\in [-2,2]\),
Proof
See e. g. [17], inequality (2.10)]. For completeness we include short proof here. Obviously
Denote by
First we assume that \(|z+2s(z)+\Lambda _n^{(\mathbb {J})}|> \sqrt{|\widetilde{T}_n(z)|}\). Then the claim of Lemma 5.9 holds. In the case \(|z+2s(z)+\Lambda _n^{(\mathbb {J})}|\le \sqrt{|\widetilde{T}_n(z)|}\), we assume \(|\Lambda _n^{(\mathbb {J})}|>2\sqrt{|\widetilde{T}_n(z)|}\). Otherwise the Lemma is proved. Under these assumptions we have
On the other hand
We take here the branch of \(\sqrt{z}\) such that \(\mathrm {Im}\;\!\sqrt{z}\ge 0\). Note that, for \(z\in [-2,2]\) we have \(\mathrm {Re}\;\!\{z^2-4\}\le 0\). This implies that
Inequalities (5.39), (5.40) and (5.41) together imply
The last inequality and Eq. (5.38) complete the proof of Lemma 5.9. \(\square \)
Lemma 5.10
Assuming the conditions of Theorem 1.1 and condition (5.36), we obtain, for \(|\mathbb {J}|\le Cn^{\frac{1}{2}}\)
Proof
By Lemma 5.8, we have
Furthermore,
Lemmas 5.5–5.8 together with Lemma 7.7 imply
Thus, Lemma 5.10 is proved. \(\square \)
Lemma 5.11
Assuming the conditions of Theorem 1.1 and condition (5.36), there exists an absolute constant \(C_4\) such that, for \(p\le A_1(nv)^{\frac{1}{4}}\) and \(v\ge v_1/s_0\), we have, uniformly in \(\mathbb {J}\subset \mathbb {T}\) such that \(|\mathbb {J}|\le Cn^{\frac{1}{2}}\),
Proof
We start from the obvious inequality, using \(|s(z)|\le 1\),
Furthermore, applying Lemma 5.10, we get
By definition,
Inequalities (5.45) and (5.44) together imply
Let \(C'{=}s_0\max \{9,C_3^{\frac{1}{2}}, C_0^{\frac{3}{2}}C_1^{\frac{1}{2}}, 3C_3C_0^{\frac{1}{2}}, C_3C_0^{\frac{3}{4}}\}\). Using Lemma 7.4 with \(x=(A_p^{(\mathbb {J})})^{\frac{1}{4}}\), \(t=4\), \(r=1\), we get
For \(p\le A_1(nv)^{\frac{1}{4}}\), we get, for \(z\in \mathbb {G}\),
where \(C_4\) is some absolute constant. We may take \(C_4=2C'\). \(\square \)
Corollary 5.12
Assuming the conditions of Theorem 1.1 and condition (5.36), we have , for \(v\ge v_1/s_0\), and for any \(\mathbb {J}\subset \mathbb {T}\) such that \(|\mathbb {J}|\le \sqrt{n}\)
where
Proof
Without loss of generality we may assume that \(C_0>1\). The bound (5.46) follows now from Lemmas 5.10 and 5.11. \(\square \)
Lemma 5.13
Assuming the conditions of Theorem 1.1 and condition (5.36) for \(\mathbb {J}\subset \mathbb {T}\) such that \(|\mathbb {J}|\le L\le \sqrt{n}\), there exist positive constant \(A_0, C_0, A_1\) depending on \(\mu _4, D_0\) only , such that we have, for \(p\le A_1(nv)^{\frac{1}{4}}\) and \(v\ge v_1/s_0\) uniformly in \(\mathbb {J}\) and \(v_1\)
with \(|\mathbb {J}|\le L-1\).
Proof
According to inequality (5.35), we have
Applying condition (5.36), we get
Combining results of Lemmas 5.5–5.8 and Corollary 5.12, we obtain
We may rewrite the last inequality as follows
where
Note that for
and \(C_0\ge 25\), we obtain that
Thus Lemma 5.13 is proved. \(\square \)
Corollary 5.14
Assuming the conditions of Theorem 1.1, we have, for \(p\le 8\) and \(v\ge v_0= A_0n^{-1}\) there exist a constant \(C_0>0\) depending on \(\mu _4\) and \(D_0\) only such that for all \(1\le j\le n\) and all \(z\in \mathbb {G}\)
and
Proof
Let \(L=[-\log _{s_0}v_0]+1\). Note that \(s_0^{-L}\le v_0\) and \(A_1\frac{n^{\frac{1}{4}}}{s_0^{\frac{L}{4}}}\ge A_1(nv_0)^{\frac{1}{4}}\). We may choose \(C_0=25\) and \(A_0, A_1\) such that (5.47) holds and
Then, for \(v=1\), and for any \(p\ge 1\), for any set \(\mathbb {J}\subset \mathbb {T}\) such that \(|\mathbb {J}|\le L\)
By Lemma 5.13, inequality (5.50) holds for \(v\ge 1/s_0\) and for \(p\le A_1n^{\frac{1}{4}}/s_0^{\frac{1}{4}}\) and for \(\mathbb {J}\subset \mathbb {T}\) such that \(|\mathbb {J}|\le L-1\). After repeated application of Lemma 5.13 (with (5.50) as assumption valid for \(v \ge 1/s_0\)) we arrive at the conclusion that the inequality (5.50) holds for \(v\ge 1/s_0^{2}\), \(p\le A_1n^{\frac{1}{4}}/s_0^{\frac{1}{2}}\) and all \(\mathbb {J}\subset \mathbb {T}\) such that \(|\mathbb {J}|\le L-2\). Continuing this iteration inequality (5.50) finally holds for \(v\ge A_0n^{-1}\), \(p\le 8\) and \(\mathbb {J}=\emptyset \).
The proof of inequality of (5.49) is similar: We have by (1.9)
Furthermore, using that \(|m_n'(z)|\le |\frac{1}{n}\mathrm{Tr}\,\mathbf {R}^2|\le \frac{1}{n}\sum _{j=1}^n\mathbf{E}|R_{jj}|^2\) and Lemma 7.6, inequality (7.11) in the Appendix, we get
By integration, this implies that (see the proof of Lemma 7.7)
Inequality (5.51) and the Cauchy–Schwartz inequality together imply
Applying inequality (5.52), we obtain
Using Corollary 5.12, we get, for \(v\ge 1/s_0\)
Thus inequality (5.49) holds for \(v\ge 1/s_0\) as well. Repeating this argument inductively with \(A_0,A_1,C_)\) satisfying (5.47) for the regions \(v \ge s_0^{-\nu }\), for \(\nu =1,\ldots ,L\) and \(z \in \mathbb {G}\), we get the claim. Thus, Corollary 5.14 is proved. \(\square \)
6 Proof of Theorem 1.3
We return now to the representation (5.32) which implies that
We may continue the last equality as follows
where
Note that the definition of \(\varepsilon _{j4}\) in (5.32) and equality (7.34) together imply
Thus we may rewrite (6.2) as
Denote by
6.1 Estimation of \(\mathfrak {T}\)
We represent \(\mathfrak {T}\)
where
6.1.1 Estimation of \(\mathfrak {T}_{1}\)
We may decompose \(\mathfrak {T}_{1}\) as
where
It is easy to see that, by conditional expectation
Applying the Cauchy–Schwartz inequality, for \(\nu =1,2,3\), we get
Applying the Cauchy–Schwartz inequality again, we get
Inequalities (6.8), (6.9), Corollaries 7.23 and 5.14 together imply
By Lemmas 7.14 and 7.5 we have for \(\nu =1\)
By Corollary 7.17, inequality (7.31) with \(\alpha =0\) and \(\beta =4\) in the Appendix we have for \(\nu =2,3\)
with some constant \(C>0\) depending on \(\mu _4\) and \(D_0\) only. Using that, by Lemma 7.13, for \(z\in \mathbb {G}\),
we get from (6.10), (6.11) and (6.12) that for \(z\in \mathbb {G}\),
with some constant \(C>0\) depending on \(\mu _4\) and \(D_0\) only.
6.1.2 Estimation of \(\mathfrak {T}_{2}\)
Using the representation (5.32), we write
Furthermore we note that
where
We now decompose \(\mathfrak {T}_{2}\) as follows
where
Applying the Cauchy–Schwartz inequality, we obtain
In what follows we denote by \(C\) a generic constant depending on \(\mu _4\) and \(D_0\) only. We note that
Using Lemmas 7.14, 7.15, 7.16, we get
This inequality and Lemma 7.5 together imply
Inequality (6.18) and Corollary 5.14 together imply
Applying Lemma 7.13 for \(z\in \mathbb {G}\), we get
By Hölder’s inequality, we have
Using now Lemmas 7.22, 7.13 and Corollary 5.14, we may write, for \(z\in \mathbb {G}\),
We continue now with \(\mathfrak {T}_{21}\). We represent it in the form
where
Furthermore, using the representation
(compare with (3.2)), we bound \(H_2\) in the following way
where
Using inequality (7.42) in the Appendix and Hölder inequality, we get, for \(\nu =1,2,3\)
Applying Corollary 7.17 with \(\beta =\frac{4}{3}\) and \(\alpha =\frac{8}{3}\), we obtain, for \(z\in \mathbb {G}\), and for \(\nu =1,2,3\)
This yields together with Corollary 5.14 and inequality (6.23)
Consider now \(H_1\). Using the equality
and
we represent it in the form
where
In order to apply conditional independence, we write
where
It is straightforward to check that
Using equality (6.3) for \(m_n'(z)\) and the corresponding relation for \({m_n^{(j)}}'(z)\), we may write
where
Using Lemmas 7.5, 7.18, 7.26, and Corollary 5.14, it is straightforward to check that
Applying inequality (7.42), we may write
Conditioning on \(\mathfrak {M}^{(j)}\) and applying Lemma 7.15, Lemma 7.5, inequality (7.8), Corollary 5.14 and equality (6.25), we get
By Lemma 7.24, we get
Similar we get
We rewrite now the equations (6.2) and (6.4) as follows, using the remainder term \(\mathfrak {T}_3\), which is bounded by means of inequalities (6.14), (6.19), (6.20), (6.24).
where
Note that
In (6.31) we estimate now the remaining quantity
6.2 Estimation of \(\mathfrak {T}_4\)
Using that \(\Lambda _n=m_n(z)-s(z)\) we rewrite \(\mathfrak {T}_4\) as
where
6.2.1 Estimation of \(\mathfrak {T}_{42}\)
First we investigate \(m_n'(z)\). The following equality holds
where
Using equality (7.34), we may write
Denote by
Using these notation we may write
Solving this equation with respect to \(m_n'(z)\) we obtain
where
Note that for the semi-circular law
Applying this relation we rewrite equality (6.35) as
Using the last equality, we may represent \(\mathfrak {T}_{42}\) now as follows
where
Recall that, by (6.33),
Applying the Cauchy–Schwartz inequality, we get for \(z\in \mathbb {G}\),
Using Corollary 7.23 and Corollary 5.14, we get
6.2.2 Estimation of \(\mathfrak {T}_{422}\)
We represent now \(\mathfrak {T}_{421}\) in the form
where
At first we investigate \(\mathfrak {T}_{53}\). Note that, by Lemma 7.26,
Therefore, for \(z\in \mathbb {G}\), using Lemma 7.5, inequality (7.9), and Lemma 7.13,
Furthermore, we consider the quantity \(\mathfrak {T}_{5\nu }\), for \(\nu =2,3\). Applying the Cauchy–Schwartz inequality and inequality (7.42) in the Appendix as well, we get
By Lemma 7.25 together with Lemma 7.5 in the Appendix, we obtain
This implies that
Inequalities (6.39) and (6.40) yield
Combining (6.38) and (6.41), we get, for \(z\in \mathbb {G}\),
6.2.3 Estimation of \(\mathfrak {T}_{43}\)
Recall that
Applying equality (6.36), we obtain
where
Applying the Cauchy–Schwartz inequality, we get
By definition of \(D_1\) and Lemmas 7.18 and 7.24 , we get
Applying now Corollary 5.14 and Lemma 7.22, we get
For \(z\in \mathbb {G}\) this yields
Applying again the Cauchy–Schwartz inequality, we get for \(\mathfrak {T}_{432}\) accordingly
By Lemma 7.24, we have
By definition of \(D_2\),
Applying Lemmas 7.25 with \(\nu =2,3\), and 7.26, we get
Inequalities (6.43) and (6.44) together imply, for \(z\in \mathbb {G}\),
Finally we observe that
and, therefore
For \(z\in \mathbb {G}\) we may rewrite it
Combining now relations (6.31), (6.26), (6.24), (6.39), (6.41), (6.45), we get for \(z\in \mathbb {G}\),
The last inequality completes the proof of Theorem 1.3.
References
Bai, Z.D., Miao, B., Tsay, J.: Convergence rates of the spectral distributions of large Wigner matrices. Int. Math. J. 1, 65–90 (2002)
Bai, Z.D.: Convergence rate of expected spectral distributions of large random matrices. I. Wigner matrices. Ann. Probab. 21(2), 625–648 (1993)
Bai, Z.D., Silverstein, J.W.: Spectral Analysis of Large Dimensional Random Matrices. Springer, Berlin (2010)
Bobkov, S., Götze, F., Tikhomirov, A.N.: On concentration of empirical measures and convergence to the semi-circle law. J. Theor. Probab. 23, 792–823 (2010)
Burkholder, D.L.: Distribution function inequalities for martingales. Ann. Probab. 1, 19–42 (1973)
Erdös, L., Yau, H.-T., Yin, J.: Rigidity of eigenvalues of generalized Wigner matrices. Adv. Math. 229(3), 1435–1515 (2012). Preprint, arXiv:1007.4652
Erdös, L., Yau, H.-T., Yin, J.: Bulk universality for generalized Wigner matrices. Probab. Theory Related Fields 154(1–2), 341–407 (2012). Preprint, arXiv:1001.3453
Girko, V.L.: Extended proof of the statement: convergence rate of expected spectral functions of symmetric random matrices \(\Sigma _n\) is equal \(O(n^{-\frac{1}{2}})\) and the method of critical steepest descent. Random Oper. Stochastic Equ. 10, 253–300 (2002)
Götze, F., Tikhomirov, A.N.: The rate of convergence of spectra of sample covariance matrices. Teor. Veroyatn. Primen. 54, 196–206 (2009)
Götze, F., Tikhomirov, A.N.: Rate of convergence to the semi-circular law. Probab. Theory Relat. Fields 127(2), 228–276 (2003)
Götze, F., Tikhomirov, A.N.: On the rate of convergence to the semi-circular law. Preprint, arXiv:1109.0611v3
Götze, F., Tikhomirov, A.N.: The rate of convergence for spectra of GUE and LUE matrix ensembles. Central Eur. J. Math. 3, 666–704 (2005)
Götze, F., Tikhomirov, A.N.: Rate of convergence to the semi-circular law. Probab. Theory Relat. Fields 127, 228–276 (2003)
Gustavsson, J.: Gaussian fluctuations of eigenvalues in the GUE. Ann. I. H. Poincare PR 41, 151–178 (2005)
Hitczenko, P.: Best constant in the martingale version of Rosenthal’s inequality. Ann. Probab. 18(4), 1656–1668 (1990)
Johnson, W.B., Schechtman, G., Zinn, J.: Best constants in moment inequalities for linear combinations of independent and exchangeable random variables. Ann. Probab. 13(1), 234–253 (1985)
Cacciapuoti, C., Maltsev, A., Schlein, B.: Optimal Bounds on the Stieltjes Transform of Wigner Matrices. J. Math. Phys. 54(4), 043302 (2013). Preprint, arXiv:1311.0326
Horn, R.A.: Horn, ChR: Matrix Analysis. Cambridge University Press, Cambridge (1985)
Rosenthal, H.P.: On the subspaces of Lp \((p\ge 2)\) spanned by sequences of independent random variables. Isr. J. Math. 8, 273–303 (1970)
Tao, T., Vu, V.: Random matrices: universality of local eigenvalues statistics. Acta Math. 206, 127–204 (2011)
Timushev, D.A., Tikhomirov, A.N., Kholopov, A.A.: On the accuracy of the approximation of the GOE spectrum by the semi-circular law (Russian). Teor. Veroyatn. Primen. 52(1), 180–185 (2007) [translation in Theory Probab. Appl. 52(1), 171–177 (2008)]
Tikhomirov, A.N.: On the rate of convergence of the expected spectral distribution function of a Wigner matrix to the semi-circular law. Siberian Adv. Math. 19, 211–223 (2009)
Author information
Authors and Affiliations
Corresponding author
Additional information
Research supported by SFB 701 Spectral Structures and Topological Methods in Mathematics University of Bielefeld. A. Tikhomirovs research was supported by grants RFBR N 14-01- 00500 and by Program of Fundamental Research, Ural Division of RAS, Project N 12-P-1-1013.
Appendix
Appendix
1.1 Rosenthal’s and Burkholder’s inequalities
In this subsection we state the Rosenthal and Burkholder inequalities, starting with Rosenthal’s inequality. Let \(\xi _1,\ldots ,\xi _n\) be independent random variables with \(\mathbf{E}\xi _j=0\), \(\mathbf{E}\xi _j^2=1\) and for \(p\ge 1\) \(\mathbf{E}|\xi _j|^p\le \mu _p\) for \(j=1,\ldots ,n\).
Lemma 7.1
(Rosenthal’s inequality) There exists an absolute constant \(C_1\) such that
Proof
For a proof see [19], Theorem 3] and [16], inequality (A)]. \(\square \)
Let \(\xi _1,\ldots \xi _n\) be a martingale-difference with respect to the \(\sigma \)-algebras \(\mathfrak {M}_j=\sigma (\xi _1,\ldots ,\xi _{j-1})\). Assume that \(\mathbf{E}\xi _j^2=1\) and \(\mathbf{E}|\xi _j|^p<\infty \).
Lemma 7.2
(Burkholder’s inequality) There exist an absolute constant \(C_2\) such that
Proof
For a proof of this inequality see [5], Theorem 3.2] and [15], Theorem 4.1]. \(\square \)
We rewrite the Burkholder inequality for quadratic forms in independent random variables. Let \(\zeta _1,\ldots ,\zeta _n\) be independent random variables such that \(\mathbf{E}\zeta _j=0\), \(\mathbf{E}|\zeta _j|^2=1\) and \(\mathbf{E}|\zeta _j|^p\le \mu _p\). Let \(a_{ij}=a_{ji}\) for all \(i,j=1,\ldots n\). Consider the quadratic form
Lemma 7.3
There exists an absolute constant \(C_2\) such that
Proof
Introduce the random variables, for \(j=2,\ldots ,n\),
It is straightforward to check that
and that \(\xi _j\) are \(\mathfrak {M}_{j}\) measurable. That means that \(\xi _1,\ldots ,\xi _n\) are martingale-differences. We may write
Applying now Lemma 7.2 and using
we get the claim. Thus, Lemma 7.3 is proved. \(\square \)
We prove as well the following simple Lemma
Lemma 7.4
Let \(t>r\ge 1\) and \(a,b>0\). Any \(x>0\) satisfying the inequality
is explicitly bounded as follows
Proof
First assume that \(x\le a^{\frac{1}{t}}\). Then inequality (7.3) holds. If \(x\ge a^{\frac{1}{t}}\), then according to inequality (7.2)
or
Using that for any \(\alpha >0\) and \(a>0,b>0\)
we get the claim. \(\square \)
In what follows we prove several lemmas about the resolvent matrices. Recall the equation, for \(j\in \mathbb {T}_{\mathbb {J}}\), (compare with (3.2))
where
Summing these equations for \(j\in \mathbb {T}_{\mathbb {J}}\), we get
where
Note that
where
Equalities (7.4) and (7.5) together imply
Solving this with respect to \(\Lambda _n^{(\mathbb {J})}\), we get
Lemma 7.5
Assuming the conditions of Theorem 1.1, there exists an absolute constant \(c_0>0\) such that for \(\mathbb {J}\subset \mathbb {T}\),
moreover, for any \(z\in \mathbb {G}\)
Proof
Firstly note that \(\mathrm {Im}\;\!(z+s(z))\ge 0\) and \(\mathrm {Im}\;\!m_n^{(\mathbb {J})}(z)\ge 0\). Therefore
Furthermore,
Note that for \(z\in \mathbb {G}\)
Therefore,
Thus Lemma 7.5 is proved. \(\square \)
Lemma 7.6
For any \(z=u+iv\) with \(v>0\) and for any \(\mathbb {J}\subset \mathbb {T}\), we have
For any \(l\in \mathbb {T}_{\mathbb {J}}\)
and
Moreover, for any \(\mathbb {J}\subset T\) and for any \(l\in \mathbb {T}_{\mathbb {J}}\) we have
and, for any \(p\ge 1\)
Finally,
and
We have as well
Proof
For \(l\in \mathbb {T}_{\mathbb {J}}\) let us denote by \(\lambda ^{(\mathbb {J})}_l\) for \(l\in \mathbb {T}_{\mathbb {J}}\) eigenvalues of matrix \(\mathbf {W}^{(\mathbb {J})}\). Then we may write (compare (7.20))
Note that, for any \(x\in \mathbb {R}^1\)
We may write
and
So, inequality (7.11) is proved. Let denote now by \(\mathbf {u}_l^{(\mathbb {J})}=(u^{(\mathbb {J})}_{lk})_{k\in \mathbb {T}_{\mathbb {J}}}\) the eigenvector of matrix \(\mathbf {W}^{(\mathbb {J})}\) corresponding to the eigenvalue \(\lambda ^{(\mathbb {J})}_l\). Using this notation we may write
It is straightforward to check the following inequality
Thus, inequality (7.12) is proved. Similar we get
This proves inequality (7.13). To prove inequality (7.14) we observe that
This inequality implies
Applying now inequality (7.12), we get
This leads (using \(|R^{(\mathbb {J})}_{ll}|\le v^{-1}\) ) to the following bound
Thus inequality (7.14) is proved. Furthermore, applying inequality (7.23), we may write
Applying (7.12), this inequality yields
The last inequality proves inequality (7.15). Note that
Thus, inequality (7.16) is proved. To finish we note that
Applying inequality (7.13), we get
To prove inequality (7.18), we note
This inequality implies
Applying inequality (7.11), we get the claim. Thus, Lemma 7.6 is proved. \(\square \)
Lemma 7.7
For any \(s\ge 1\), and for any \(z=u+iv\) and for any \(\mathbb {J}\subset \mathbb {T}\), and \(j\in \mathbb {T}_{\mathbb {J}}\),
Proof
See [17], Lemma 3.4]. For the readers convenience we include the short argument here. Note that, for any \(j\in \mathbb {T}_{\mathbb {J}}\),
Furthermore,
and
From here it follows that
We may write now
The last inequality yields the claim. Thus Lemma 7.7 follows. \(\square \)
Lemma 7.8
Assuming the conditions of Theorem 1.1, we get
Proof
The proof follows immediately from the definition of \(\varepsilon _{j1}\) and conditions of Theorem 1.1. \(\square \)
1.1.1 Some auxiliary bounds for resolvent matrices for \(\mathrm {Im}\;\!z=4\)
We need the bound for the \(\varepsilon _{j\nu }\), and \(\eta _{j}\) for \(V=4\).
Lemma 7.9
Assuming the conditions of Theorem 1.1, we get
Proof
Conditioning on \(\mathfrak {M}^{(j)}\), we get
Applying now Lemma 7.6, inequality (7.12), we get with \(\mathrm {Im}\;\!R^{(\mathbb {J})}\le \frac{1}{4}\),
Thus Lemma 7.9 is proved. \(\square \)
Lemma 7.10
Assuming the conditions of Theorem 1.1, we get
Proof
Conditioning on \(\mathfrak {M}^{(j)}\), we obtain as above
Thus Lemma 7.10 is proved. \(\square \)
Lemma 7.11
Assuming the conditions of Theorem 1.1, we get
Proof
The proof is similar to proof of Lemma 7.9. We need to use that \(|[(R^{(j)})^2]_{kl}|\le V^{-2}=\frac{1}{16}\) and \(\sum _{l\in \mathbb {T}_j}|[(R^{(j)})^2]_{kl}|^2\le V^{-4}\). \(\square \)
Lemma 7.12
Assuming the conditions of Theorem 1.1, we get
k,
Proof
The result follows immediately from the bound
See for instance [10], Lemma 3.3]. \(\square \)
1.2 Some auxiliary bounds for resolvent matrices for \(z\in \mathbb {G}\)
Introduce now the region
In the next lemma we state some useful inequalities for the region \(\mathbb {G}\).
Lemma 7.13
For any \(z\in \mathbb {G}\) we have
Proof
We observe that
This inequality proves the Lemma. \(\square \)
Lemma 7.14
Assuming the conditions of Theorem 1.1, there exists an absolute constant \(C>0\) such that for any \(j=1,\ldots ,n\),
Proof
The result follows immediately from the definition of \(\varepsilon _{j1}\). \(\square \)
Lemma 7.15
Assuming the conditions of Theorem 1.1, there exists an absolute constant \(C>0\) such that for any \(j=1,\ldots ,n\),
and
Proof
Note that r.v.’s \(X_{jl}\), for \(l\in \mathbb {T}_j\) are independent of \(\mathfrak {M}^{(j)}\) and that for \(l,k\in \mathbb {T}_j\) \(R^{(j)}_{lk}\) are measurable with respect to \(\mathfrak {M}^{(j)}\). This implies that \(\varepsilon _{j2}\) is a quadratic form with coefficients \(R^{(j)}_{lk}\) independent of \(X_{jl}\). Thus its variance and fourth moment are easily available.
Here we use the notation \(|\mathbf {A}|^2=\mathbf {A}\mathbf {A}^*\) for any matrix \(\mathbf {A}\). Applying Lemma 7.6, inequality (7.11), we get equality (7.26).
Furthermore, direct calculations show that
Here again we used Lemma 7.6, inequality (7.11). Thus Lemma 7.15 is proved. \(\square \)
Lemma 7.16
Assuming the conditions of Theorem 1.1, there exists an absolute constant \(C>0\) such that for any \(j=1,\ldots ,n\),
and
Proof
The first inequality is easy. To prove the second, we apply Rosenthal’s inequality. We obtain
Using \(|X_{jl}|\le Cn^{\frac{1}{4}}\) we get \(\mu _8\le Cn\mu _4\) and the claim. Thus Lemma 7.16 is proved. \(\square \)
Corollary 7.17
Assuming the conditions of Theorem 1.1, there exists an absolute constant \(C>0\), depending on \(\mu _4\) and \(D_0\) only, such that for any \(j=1,\ldots ,n\), \(\nu =1,2,3\) \(z\in \mathbb {G}\), and \(0\le \alpha \le \frac{1}{2}A_1(nv)^{\frac{1}{4}}\) and \(\beta \ge 1\)
and for \(\beta \ge 2\)
We have as well, for \(\nu =2,3\), and \(\beta \ge 4\)
Proof
For \(\nu =1\), by Lemma 7.5, we have
Applying now Corollary 5.14 and Lemma 7.13, we get the claim. The proof of the second inequality for \(\nu =1\) is similar. For \(\nu =2\) we apply Lemmas 7.15 and 7.5, inequality (7.26) and obtain, using inequality (7.8),
Similar, using Lemma 7.15, inequality (7.27), and inequality (7.8) we get
Applying Corollary 5.14, we get the claim. For \(\nu =3\), we apply Lemma 7.16, inequalities (7.28) and (7.29) and Lemma 7.5. We get
and
Using now the Cauchy–Schwartz inequality and Corollary 5.14, we get the claim. \(\square \)
Lemma 7.18
Assuming the conditions of Theorem 1.1, there exists an absolute constant \(C>0\) such that for any \(j=1,\ldots ,n\),
Proof
This inequality follows from
which may be obtained using the Schur complement formula. See, for instance [10], Lemma 3.3]. \(\square \)
Corollary 7.19
Assuming the conditions of Theorem 1.1, there exists an absolute constant \(C>0\) such that for any \(j=1,\ldots ,n\),
Proof
By definition of \(\varepsilon _j\) (see (3.2)), we have
Applying now Lemmas 7.14, 7.15 (inequality (7.27)), 7.16 (inequality (7.29)) and 7.18 and taking expectation where needed (7.35) follows. Thus the Corollary is proved. \(\square \)
Introduce quantities
Lemma 7.20
Assuming the conditions of Theorem 1.1, we have, for \(z\in \mathbb {G}\), for any \(j=1,\ldots ,n\)
Proof
Direct calculation shows
Applying Lemma 7.6, the inequality (7.15) and the inequality \(\mu _8\le Cn\mu _4\), we get
Thus, inequality (7.37) is proved. Consider now the \(8\)th moment of \(\eta _{j3}\). Applying Rosenthal’s inequality, we get
Using now Lemma 7.6, inequalities (7.14) and (7.15), and that \(\mu _{16}\le Cn^3\mu _4\), we obtain
Thus, inequality (7.38) is proved. \(\square \)
Lemma 7.21
Assuming the conditions of Theorem 1.1, we have, for \(z\in \mathbb {G}\), for any \(j=1,\ldots ,n\)
Proof
Direct calculation shows that
Applying inequalities (7.16) and (7.17) of Lemma 7.6, we get
Furthermore, we have
Note that
Using this relations, we get
Applying now Lemma 7.6, we obtain
Thus, the Lemma is proved. \(\square \)
Lemma 7.22
Assuming the conditions of Theorem 1.1, we have
Proof
Using the representations (7.34)–(7.36) we have
Applying the Cauchy–Schwartz inequality, we get
Using Corollary 5.14, we may write
Observe that,
Therefore, by Lemmas 7.18, 7.5 and 7.13, for \(z\in \mathbb {G}\),
Using (7.42) we get by definition of \(\eta _{j1}\), Lemma 7.6, and inequality (7.11)
Furthermore, applying inequality (7.42) again, we obtain
Conditioning with respect to \(\mathfrak {M}^{(j)}\) and applying Lemma 7.20, we obtain
Using Lemma 7.5, inequality (7.8), together with Corollary 5.14 we get
Applying inequality (7.42) and conditioning with respect to \(\mathfrak {M}^{(j)}\) and applying Lemma 7.21, we get
Using that \(|z+m_n^{(j)}(z)+s(z)|\ge \mathrm {Im}\;\!m_n^{(j)}(z)\) together with Lemma 7.6, we arrive at
Summarizing we may write now, for \(z\in \mathbb {G}\),
For \(z\in \mathbb {G}\), see (7.24) and Lemma 7.13, this inequality may be simplified by means of the following bounds (with \(v_0= A_0n^{-1}\))
Using these relation, we obtain
Thus Lemma 7.22 is proved. \(\square \)
Corollary 7.23
Assuming the conditions of Theorem 1.1, we have
Proof
The result follows immediately from Lemma 7.22 and Jensen’s inequality. \(\square \)
Lemma 7.24
Assuming the conditions of Theorem 1.1, we have, for \(z\in \mathbb {G}\),
Proof
We write
where
First we observe that by (7.34)
Hence \(|z+m_n(z)+s(z)|\ge \mathrm {Im}\;\!m_n(z)\) and Jensen’s inequality yields
Furthermore, we represent \(T_{n1}\) as follows
where
Using these notations we may write
Applying the Cauchy–Schwartz inequality twice and using the definition of \(\varepsilon _{j1}\) (see (3.2)), we get by Lemma 7.5
By Rosenthal’s inequality, we have, for \(z\in \mathbb {G}\)
Using (3.2) we rewrite \(T_{n12}\), obtaining
By the Cauchy–Schwartz inequality, using the definition of \(\varepsilon _j\) (see representation (3.2)), we obtain
For \(\nu =1\), we have
Applying the Cauchy–Schwartz inequality twice and Lemma 7.5, we arrive at
Observe that
The last inequality, inequality (7.49), Corollary 5.14 and Lemma 7.13 together imply
Furthermore, for \(\nu =4\), by Lemma 7.18 we have
Applying the Cauchy–Schwartz inequality and Lemma 7.5, we get
Similar to inequality (7.51), applying Lemma 7.13, inequality (7.50) and Corollary 5.14, we get
By Hölder’s inequality, we have for \(\nu =2,3\),
Note that for \(\nu =2,3\), r.v. \(X_{jj}\) doesn’t depend on \(\varepsilon _{j\nu }\) and on \(\sigma \)-algebra \(\mathfrak {M}^{(j)}\). Using inequality (7.42) for \(z\in \mathbb {G}\), we get
Applying now Lemmas 7.15, 7.16 and 7.13, arrive at
Inequalities (7.51), (7.52), (7.53) together imply
Consider now the quantity
for \(\nu =2,3\). We represent it as follows
where
By the representation (5.32), which is similar to (3.2) we have
Using inequality (7.42), we may write, for \(z\in \mathbb {G}\)
Applying the Cauchy–Schwartz inequality and the inequality \(ab\le \frac{1}{2}(a^2+b^2)\), we get
Using Corollary 7.17 with \(\alpha =2\), we arrive at
In order to estimate \(Y_{\nu 1}\) we introduce now the quantity
Recall that
Note that
We use that (see Lemma 5.5 in [11])
Note that
This yields
We represent \(Y_{\nu 1}\) in the form
where
First, note that by conditional independence
Furthermore, applying Hölder’s inequality, we get
Using Corollary 7.17 with \(\alpha =4\) and Lemmas 7.22 and 7.5, we obtain
For \(z\in \mathbb {G}\) we may rewrite this bound using Lemma 7.13
Furthermore, note that
This inequality together with (7.59) implies that
where
By representation (3.2), we have
This implies that
where
Applying Hölder inequality, we get
By Lemma 7.21, inequality (7.39) and Corollary 5.14, we have
Furthermore, applying Hölder inequality again, we get
The last inequality, Corollaries 5.14, 7.17, Lemma 7.22 together imply
Inequalities (7.64) and (7.66) together imply
To bound \(Z_{\nu 2}\) we first apply inequality (7.42) and obtain
Furthermore, similar to bound of \(Z_{\nu 4}\)—inequality (7.62)—we may write
where
Applying inequality (7.63) and Cauchy–Schwartz inequality, we get
Lemmas 7.15, 7.16, 7.22, inequality (7.39) and Corollary 5.14 together imply
Applying now Hölder’s inequality, we get
The last inequality together with Lemmas 7.22, 7.20, 7.21 and Corollary 5.14, 7.17 implies
Combining now inequalities (7.46), (7.47), (7.54), (7.55), (7.60), (7.61), (7.67), (7.69), (7.70), we get
Applying Lemma 7.4 with \(t=2\), \(r=1\) completes the proof of Lemma 7.24. \(\square \)
We relabel \(\eta _{j2},\,\eta _{j3}\) and introduce the following quantity
Lemma 7.25
Assuming the conditions of Theorem 1.1, we have, for \(\nu =2,3\),
Proof
We recall that by \(C\) we denote the generic constant depending on \(\mu _4\) and \(D_0\) only. By definition of \(\beta _{j\nu }\) for \(\nu =2,3\), conditioning on \(\mathfrak {M}^{(j)}\), we get
Applying Lemma 7.6, we get the claim. Thus Lemma 7.25 is proved. \(\square \)
Lemma 7.26
Assuming the conditions of Theorem 1.1, we have, for \(j=1,\ldots ,n\),
Proof
Let \(\mathcal {F}_n^{(j)}(x)\) denote empirical spectral distribution function of matrix \(\mathbf {W}^{(j)}\). According to interlacing eigenvalues Theorem (see [18], Theorem 4.38]) we have
Furthermore, we represent
Integrating by parts, we get the claim.
Thus Lemma 7.26 is proved. \(\square \)
Rights and permissions
About this article
Cite this article
Götze, F., Tikhomirov, A. Optimal bounds for convergence of expected spectral distributions to the semi-circular law. Probab. Theory Relat. Fields 165, 163–233 (2016). https://doi.org/10.1007/s00440-015-0629-5
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00440-015-0629-5