Formalizing Wigner’s Semicircle Law

5 Weak Convergence

Lemma 5.0.1 R-3-1 : lem:weak_convergence
#

Let \(k \in \mathbb {N}\) and \(\epsilon {\gt} 0\). Then for any \(b {\gt} 4\),

\[ \underset {n \rightarrow \infty }{\lim \sup } \mathbb {P} \biggl( \int _{|x| {\gt} b} |x|^k \mu _{\mathbf{X}_n}(dx) {\gt} \epsilon \biggl) = 0. \]
Lemma 5.0.2 R-3-2 : lem:R-3-2
\[ \mathbb {P} \biggl( \int _{|x| {\gt} b} |x|^k \mu _{\mathbf{X}_n}(dx) {\gt} \epsilon \biggl) \leq \frac{1}{\epsilon } \mathbb {E} \biggl( \int _{|x| {\gt} b} |x|^k \mu _{\mathbf{X}_n}(dx) \biggl). \]
Proof

This follows from Markov’s inequality.

Definition 5.0.3 R-3-3 : def:R-3-3
#

We define \(\nu \) to be the random measure \(\nu (dx) = |x|^k \mu _{\mathbf{X}_n}(dx)\).

Lemma 5.0.4 R-3-4 : lem:R-3-4
\[ \int _{|x| {\gt} b} |x|^k \mu _{\mathbf{X}_n}(dx) = \nu \{ x : |x|^k {\gt} b^k\} . \]
Proof
\[ \int _{|x| {\gt} b} |x|^k \mu _{\mathbf{X}_n}(dx) = \nu \{ x : |x| {\gt} b\} = \nu \{ x : |x|^k {\gt} b^k\} . \]
Lemma 5.0.5 R-3-5 : lem:R-3-5
\[ \int _{|x| {\gt} b} |x|^k \mu _{\mathbf{X}_n}(dx) \leq \frac{1}{b^k} \int |x|^{2k} \mu _{\mathbf{X}_n}(dx). \]
Proof

First, applying Markov’s inequality on Lemma 5.0.4 gives

\[ \int _{|x| {\gt} b} |x|^k \mu _{\mathbf{X}_n}(dx) \leq \frac{1}{b^k} \int |x|^k \nu (dx). \]

Next, ‘substituting’ \(\nu (dx)\) with \(|x|^k \mu _{\mathbf{X}_n}(dx)\) gives

\[ \int _{|x| {\gt} b} |x|^k \mu _{\mathbf{X}_n}(dx) \leq \frac{1}{b^k} \int |x|^{2k} \mu _{\mathbf{X}_n}(dx). \]
Lemma 5.0.6 R-3-6 : lem:R-3-6
\[ \mathbb {P} \biggl( \int _{|x| {\gt} b} |x|^k \mu _{\mathbf{X}_n}(dx) {\gt} \epsilon \biggl) \leq \frac{1}{\epsilon b^k} \mathbb {E} \biggl( \int |x|^{2k} \mu _{\mathbf{X}_n}(dx) \biggl). \]
Proof

Substituting the term inside the expectation of Lemma 5.0.2 with the expression acquired from Lemma 5.0.5 gives

\[ \mathbb {P} \biggl( \int _{|x| {\gt} b} |x|^k \mu _{\mathbf{X}_n}(dx) {\gt} \epsilon \biggl) \leq \frac{1}{\epsilon } \mathbb {E} \biggl( \int _{|x| {\gt} b} |x|^k \mu _{\mathbf{X}_n}(dx) \biggl) \leq \frac{1}{\epsilon b^k} \mathbb {E} \biggl( \int |x|^{2k} \mu _{\mathbf{X}_n}(dx) \biggl). \]
Lemma 5.0.7 R-3-7 : lem:R-3-7
\[ \mathbb {P} \biggl( \int _{|x| {\gt} b} |x|^k \mu _{\mathbf{X}_n}(dx) {\gt} \epsilon \biggl) \leq \frac{1}{\epsilon b^k} \cdot \frac{1}{n} \mathbb {E} \mathrm{Tr} (\mathbf{X}_n^k). \]
Proof

Substituting the term inside the expectation of Lemma 5.0.6 with the expression acquired in Lemma ?? gives

\[ \mathbb {P} \biggl( \int _{|x| {\gt} b} |x|^k \mu _{\mathbf{X}_n}(dx) {\gt} \epsilon \biggl) \leq \frac{1}{\epsilon b^k} \mathbb {E} \biggl( \int |x|^{2k} \mu _{\mathbf{X}_n}(dx) \biggl) = \frac{1}{\epsilon b^k} \cdot \frac{1}{n} \mathbb {E} \mathrm{Tr} (\mathbf{X}_n^k). \]
Lemma 5.0.8 Catalan Number bound
#

Let \(C_{k}\) be the Catalan number. Then, for all \(k \in \mathbb {N}\), we have:

\[ C_{k} \leq 4^{k} \]
Lemma 5.0.9 Bound for \(\lim \sup _{n\to \infty }\frac{1}{n}\mathbb {E} \operatorname {Tr} (\mathbf{X}_{n}^{k})\)

Let \(\{ Y_{ij}\} _{1 \leq i \leq j}\) be independent random variables, with \(\{ Y_{ii}\} _{i\geq 1}\) identically distributed and \(\{ Y_{ij}\} _{1 \leq i {\lt} j}\) identically distributed. Suppose that \(r_k = \max \{ \mathbb {E}(|Y_{11}|^k),\mathbb {E}(|Y_{12}|^k)\} {\lt} \infty \) for each \(k\in \mathbb {N}\). Suppose further than \(\mathbb {E}(Y_{ij})=0\) for all \(i,j\). If \(i{\gt}j\), define \(Y_{ij} \equiv Y_{ji}\), and let \(\mathbf{Y}_n\) be the \(n\times n\) matrix with \([\mathbf{Y}_n]_{ij} = Y_{ij}\) for \(1\le i,j\le n\). Let \(\mathbf{X}_n = n^{-1/2}\mathbf{Y}_n\) be the corresponding Wigner matrix. Then, we have:

\[ \lim \sup _{n\to \infty } \frac{1}{n}\mathbb {E}\operatorname {Tr} (\mathbf{X}_{n}^{k}) \leq 4^{k} \]
Proof

From proposition 4.1.69, we know that \(\lim _{n\to \infty } \frac{1}{n}\mathbb {E}\operatorname {Tr}(\mathbf{X}_{n}^{k}) \leq C_{k}\) (odd and even cases of \(k\)). We further know that since the limit exists, \(\lim \sup _{n\to \infty } \frac{1}{n}\mathbb {E}\operatorname {Tr}(\mathbf{X}_{n}^{k}) = \lim _{n\to \infty } \frac{1}{n}\mathbb {E}\operatorname {Tr}(\mathbf{X}_{n}^{k})\). Using lemma 5.0.8, we have:

\[ \lim \sup _{n\to \infty }\mathbb {E}\frac{1}{n}\operatorname {Tr}(\mathbf{X}_{n}^{k}) \leq C_{k} \leq 4^{k} \]

as required.

Lemma 5.0.10 New bound for \(\lim \sup _{n\to \infty } \mathbb {P}(\int _{|x| {\gt} b}|x|^{k} \mu \mathbf{x}_{n}(dx))\)

let \(k \in \mathbb {N}\) and \(\epsilon {\gt} 0\). then for any \(b {\gt} 4\),

\[ \lim \sup _{n\to \infty } \mathbb {P}(\int _{|x| {\gt} b}|x|^{k} \mu \mathbf{x}_{n}(dx)) \leq \frac{1}{\epsilon }(\frac{4}{b})^{k} \]
Lemma 5.0.11 Increasing Sequence \(|x|^{k}\)
#

for \(x {\gt} b {\gt} 4 {\gt} 1 \in \mathbb {R}\), we have \(k \mapsto |x|^{k}\) is increasing with \(k\).

Lemma 5.0.12 Increasing Sequence \(\mathbb {P}(\int _{|x| {\gt} b}|x|^{k}\mu \mathbf{x}_{n}(dx))\)
#

for \(x {\gt} b {\gt} 4 {\gt} 1 \in \mathbb {R}\), we have \(k \mapsto \mathbb {P} (\int _{|x| {\gt} b}|x|^{k}\mu \mathbf{x}_{n} (dx))\) is increasing with \(k\).

Lemma 5.0.13 Increasing Sequence \(\lim \sup _{n\to \infty }\mathbb {P}(\int _{|x| {\gt} b}|x|^{k}\mu \mathbf{x}_{n}(dx))\)
#

for \(x {\gt} b {\gt} 4 {\gt} 1 \in \mathbb {R}\), we have \(k \mapsto \lim \sup _{n\to \infty }\mathbb {P} (\int _{|x| {\gt} b}|x|^{k}\mu \mathbf{x}_{n} (dx))\) is increasing with \(k\).

Lemma 5.0.14 Nonnegative Sequence \(\lim \sup _{n\to \infty }\mathbb {P}(\int _{|x| {\gt} b}|x|^{k}\mu \mathbf{x}_{n}(dx))\)
#

for \(x {\gt} b {\gt} 4 {\gt} 1 \in \mathbb {R}\), we have \(k \mapsto \lim \sup _{n\to \infty }\mathbb {P} (\int _{|x| {\gt} b}|x|^{k}\mu \mathbf{x}_{n} (dx)) \geq 0\) for all \(k \in \mathbb {N}\).

Lemma 5.0.15 Decreasing Sequence \(\frac{1}{\epsilon }(\frac{4}{b})^{k}\)
#

for \(x {\gt} b {\gt} 4 {\gt} 1 \in \mathbb {R}\), we have \(k \mapsto \frac{1}{\epsilon }(\frac{4}{b})^{k}\) is decreasing to \(0\) as \(k\to \infty \).

Lemma 5.0.16 Limit of Sequence \(\lim \sup _{n\to \infty }\mathbb {P}(\int _{|x| {\gt} b}|x|^{k}\mu \mathbf{x}_{n}(dx))\)

for \(x {\gt} b {\gt} 4 {\gt} 1 \in \mathbb {R}\), the sequence \(k \mapsto \lim \sup _{n\to \infty }\mathbb {P} (\int _{|x| {\gt} b}|x|^{k}\mu \mathbf{x}_{n} (dx))\) has:

\[ \lim _{k\to \infty } \lim \sup _{n\to \infty }\mathbb {P}(\int _{|x| {\gt} b}|x|^{k}\mu \mathbf{x}_{n}(dx)) = 0 \]
Proof

We know that for all \(k \in \mathbb {N}\), \(\lim \sup _{n\to \infty }\mathbb {P}(\int _{|x| {\gt} b}|x|^{k}\mu \mathbf{x}_{n}(dx)) \geq 0\) (lemma 5.0.14). We also see that the sequence is bounded by \(\frac{1}{\epsilon } (\frac{4}{b})^{k}\) (lemma 5.0.10), which is decreasing to \(0\) as \(k\to \infty \) (lemma 5.0.15). Via squeeze theorem (where we use the zero sequence as another bound), we have:

\[ \lim _{k\to \infty } \lim \sup _{n\to \infty }\mathbb {P}(\int _{|x| {\gt} b}|x|^{k}\mu \mathbf{x}_{n}(dx)) = 0 \]

as required.

Lemma 5.0.17 All terms of Sequence are Zero conditions
#

for any sequence \((a_{k})_{k=1}^{\infty } \in \mathbb {R}\), if \(a_{k}\) has:

  1. for all \(k \in \mathbb {N}\), \(a_{k} \geq 0\) (nonnegative sequence)

  2. for all \(k \in \mathbb {N}\), we have \(a_{k+1} \geq a_{k}\) (strictly increasing sequence)

  3. \(\lim _{k\to \infty } a_{k} = 0\)

    then for all \(k \in \mathbb {N}\), we have \(a_{k} = 0\)

Lemma 5.0.18 Lemma 4.7 from [ 1 ]

let \(k \in \mathbb {N}\) and \(\epsilon {\gt} 0\). Then for any \(b {\gt} 4\),

\[ \lim \sup _{n\to \infty }\mathbb {P}(\int _{|x| {\gt} b}|x|^{k} \mu \mathbf{x}_{n}(dx)) = 0 \]
Proof

Consider sequence \((a_{k})_{k =1}^{\infty }\) defined as \(a_{k} = \lim \sup _{n\to \infty }\mathbb {P}(\int _{|x| {\gt} b}|x|^{k} \mu \mathbf{x}_{n}(dx))\). Then, we use the conditions from lemma 5.0.17:

  1. From lemma 5.0.14, we know that \(a_{k} \geq 0\) for all \(k \in \mathbb {N}\), so the sequence is nonnegative.

  2. From lemma 5.0.13, we know that \(a_{k+1} \geq a_{k}\) for all \(k \in \mathbb {N}\), so the sequence is strictly increasing.

  3. From lemma 5.0.16, we know that \(\lim _{k\to \infty } a_{k} = 0\), so the sequence converges to zero.

Therefore the sequence \(a_{k} = \lim \sup _{n\to \infty }\mathbb {P}(\int _{|x| {\gt} b}|x|^{k} \mu \mathbf{x}_{n}(dx)) = 0\) for all \(k \in \mathbb {N}\), as required.

Lemma 5.0.19 Function from Weierstrass Approx Theorem
#

Fix a bounded, continuous function \(f \in C_{b}(\mathbb {R})\), fix \(\epsilon {\gt} 0\), fix \(b {\gt} 4\). Then, there exists a polynomial \(P_{\epsilon }\) such that:

\[ \sup _{|x| \leq b} |f(x) - P_{\epsilon }(x)| {\lt} \frac{\epsilon }{6} \]
Lemma 5.0.20 Estimates from Triangle Inequality

Let \(f \in C_{b}(\mathbb {R})\), fix \(\epsilon {\gt} 0\), and \(b {\gt} 4\), and let \(P_{\epsilon }\) be the polynomial from lemma 5.0.19. Then, we have:

\[ \left|\int f d\mu _{\mathbf{x}_{n}} - \int f d\sigma _{1}\right| \leq \left|\int f d\mu _{\mathbf{x}_{n}} - \int P_{\epsilon } d\mu _{\mathbf{x}_{n}}\right| + \left|\int P_{\epsilon } d\mu _{\mathbf{x}_{n}} - \int P_{\epsilon } d\sigma _{1}\right| + \left|\int P_{\epsilon } d\sigma _1 - \int f d\sigma _1\right| \]
Lemma 5.0.21

If \(\left|\int f\, d\mu _{X_n} - \int f\, d\sigma _1\right|{\gt}\epsilon \), then \(\left|\int f\, d\mu _{X_n} - \int P_\epsilon \, d\mu _{X_n}\right|{\gt}\epsilon /3, \left|\int P_\epsilon \, d\mu _{X_n} - \int P_\epsilon \, d\sigma _1\right| {\gt}\epsilon /3,\) or \(\left|\int P_\epsilon \, d\sigma _1 - \int f\, d\sigma _1\right|{\gt}\epsilon /3\).

Proof

Pigeonhole.

Lemma 5.0.22

\(\mathbb {P}\left( \left|\int f\, d\mu _{X_n} - \int f\, d\sigma _1\right|{\gt}\epsilon \right) \le \mathbb {P}\left(\left|\int f\, d\mu _{X_n} - \int P_\epsilon \, d\mu _{X_n}\right|{\gt}\epsilon /3\right) + \mathbb {P}\left(\left|\int P_\epsilon \, d\mu _{X_n} - \int P_\epsilon \, d\sigma _1\right| {\gt}\epsilon /3\right) + \mathbb {P}\left(\left|\int P_\epsilon \, d\sigma _1 - \int f\, d\sigma _1\right|{\gt}\epsilon /3\right)\)

Proof

Triangle equality: use MeasureTheory.lintegral_edist_triangle. (Not sure if this is needed)

Lemma 5.0.23

\(\mathbb {P}\left(\left|\int P_\epsilon \, d\sigma _1 - \int f\, d\sigma _1\right|{\gt}\epsilon /3\right)\) is identically zero.

Proof

By construction, \(|P_\epsilon -f|{\lt}\epsilon /6\) on \([-b,b]\), which includes the support \([-2,2]\) of \(\sigma _1\).

Lemma 5.0.24
\[ \left|\int (f-P_\epsilon )\, d\mu _{X_n}\right| \le \int _{|x|\le b} |f(x)-P_\epsilon (x)|\, \mu _{X_n}(dx) + \int _{|x|{\gt}b} |f(x)-P_\epsilon (x)|\, \mu _{X_n}(dx). \]
Proof

Break up the integral.

Lemma 5.0.25

If \(\left|\int f\, d\mu _{X_n} - \int P_\epsilon \, d\mu _{X_n}\right|{\gt}\epsilon /3\), then \(\int |f-P_\epsilon |\mathbb {1}_{|x|\le b}\, d\mu _{X_n}{\gt}\epsilon /6\) or \(\int |f-P_\epsilon |\mathbb {1}_{|x|{\gt} b}\, d\mu _{X_n}{\gt}\epsilon /6\).

Proof

Pigeonhole.

Lemma 5.0.26
\[ \mathbb {P}\left(\left|\int f\, d\mu _{X_n} - \int P_\epsilon \, d\mu _{X_n}\right|{\gt}\epsilon /3\right) \le \mathbb {P}\left(\int |f-P_\epsilon |\mathbb {1}_{|x|\le b}\, d\mu _{X_n}{\gt}\epsilon /6\right) + \mathbb {P}\left(\int |f-P_\epsilon |\mathbb {1}_{|x|{\gt} b}\, d\mu _{X_n}{\gt}\epsilon /6\right). \]
Proof

Use previous two lemmas.

Lemma 5.0.27

\( \limsup _{n\to \infty } \left(\mathbb {P}\left(\int |f-P_\epsilon |\mathbb {1}_{|x|\le b}\, d\mu _{\mathrm{X}_n}{\gt}\epsilon /6\right) \right) = 0\)

\begin{align*} \mathbb {P}\left( \left|\int f\, d\mu _{\mathrm{X}_n} - \int f\, d\sigma _1\right|{\gt}\epsilon \right) & \le \mathbb {P}\left(\int |f-P_\epsilon |\mathbb {1}_{|x|{\gt} b}\, d\mu _{\mathrm{X}_n}{\gt}\epsilon /6\right) \\ & + \mathbb {P}\left(\left|\int P_\epsilon \, d\mu _{\mathrm{X}_n} - \int P_\epsilon \, d\sigma _1\right| {\gt}\epsilon /3\right). \end{align*}
Lemma 5.0.29

\( \limsup _{n \to \infty } \left(\mathbb {P}\left(\left|\int P_\epsilon \, d\mu _{\mathrm{X}_n} - \int P_\epsilon \, d\sigma _1\right| {\gt}\epsilon /3\right) \right) = 0 \)

\(\mathbb {P}\left( \left|\int f\, d\mu _{\mathrm{X}_n} - \int f\, d\sigma _1\right|{\gt}\epsilon \right) \leq \mathbb {P}\left(\int |f-P_\epsilon |\mathbb {1}_{|x|{\gt} b}\, d\mu _{\mathrm{X}_n}{\gt}\epsilon /6\right)\)

Lemma 5.0.31
#

\( \mathbb {P}\left(\int |f-P_\epsilon | \mathbb {1}_{|x|\ge b}\, d\mu _{\mathrm{X}_n} {\gt} \epsilon /6\right) \le \mathbb {P}\left(\int c|x|^k\mathbb {1}_{|x|\ge b}\, \mu _{\mathrm{X}_n}(dx) {\gt} \epsilon /6\right) \)

Proof

Let \(k =\)deg\(P_\epsilon \), and since \(f\) is bounded, \(|f(x) - P_{\epsilon }(x)| \leq \| f\| _{\infty } + |P_\epsilon (x)|\). Also note it is on interval \(|x| {\gt} b\), which completes the proof.

Lemma 5.0.32

\(\limsup _{n \to \infty }\mathbb {P}\left(\int c|x|^k\mathbb {1}_{|x|\ge b}\, \mu _{\mathrm{X}_n}(dx) {\gt} \epsilon /6\right) = 0 \)

Lemma 5.0.33

\(\limsup _{n \to \infty } \mathbb {P}\left(\int |f-P_\epsilon | \mathbb {1}_{|x|\ge b}\, d\mu _{\mathrm{X}_n} {\gt} \epsilon /6\right) = 0\)

Proposition 5.0.34

Let \(\mathrm{X}_n=n^{-1/2}\mathrm{Y}_n\) be a sequence of Wigner matrices, with entries satisfying \(\mathbb {E}(Y_{ij})=0\) for all \(i,j\) and \(\mathbb {E}(Y_{12}^2)=t\). Then the empirical law of eigenvalues \(\mu _{\mathrm{X}_n}\) converges in probability to \(\sigma _t\) as \(n\to \infty \). Precisely: for any \(f\in C_b(\mathbb {R})\) (continuous bounded functions) and each \(\epsilon {\gt}0\),

\[ \lim _{n\to \infty } \mathbb {P}\left(\left|\int f\, d\mu _{\mathrm{X}_n} - \int f\, d\sigma _t\right|{\gt}\epsilon \right)=0. \]