4

Let $A$ be an $n\times n$ real symmetric random matrix. Then $A$ has a spectral decomposition $A=Q\Lambda Q'$ with $Q$ orthogonal and $\Lambda$ diagonal. Am trying to show that we can choose $Q$ and $\Lambda$ to be random matrices, i.e. to have measurable entries.

Let $M_n$ denote the space $n\times n$ real matrices, and let $S_n$ denote the subspace of symmetric matrices in $M_n$. I endow $M_n$ with the Frobenius norm and the corresponding metric topology. I will use the following principle of measurable choice proven here:


Theorem. Let $X,Y$ be complete separable metric spaces and $E$ a closed $\sigma$-compact subset of $X\times Y$. Then $\pi_1(E)$ is a Borel set in $X$ and there exists a Borel function $\varphi:\pi_1(E)\to Y$ whose graph is contained in $E$. (Here $\pi_1$ denotes the projection of $X\times Y$ on $X$).


Let $X=M_n$, $Y=M_n\times M_n$ and

$$E=\bigg\{\big(A,Q,\Lambda\big)\in M_n\times M_n\times M_n: A \text{ is symmetric}, U \text{ is orthogonal},\Lambda \text{ is diagonal}\text{ and } A=Q\Lambda Q' \bigg\}$$

By identification of $M_n$ and $M_n\times M_n$ with $\mathbb{R}^{n^2}$ and $\mathbb R^{2n^2}$ we see that $X,Y$ are separable complete metric spaces, and by identification of $M_n\times M_n\times M_n$ with $\mathbb R^{3n^2}$ we see that $E$ is $\sigma$-compact. Using the sequential characterization of a closed set in a metric space and continuity of matrix multiplication we see that $E$ is closed. By the spectral theorem we have $\pi_1(E)=S_n$.

Therefore the above theorem provides us with a Borel function $\varphi:S_n\to M_n\times M_n$ whose graph is contained in $E$, i.e. such that $\varphi_1(A)$ is orthogonal, $\varphi_2(A)$ is diagonal and $A=\varphi_1(A)\varphi_2(A)\varphi_1(A)'$ for all $A\in S_n$. Since projections are continuous, $\varphi_1,\varphi_2:S_n\to M_n$ are also Borel measurable.

$A$ being a random matrix means that each entry of $A$ is a random variable on some measure space $(\Omega,\mathcal F)$. By identification of the measure spaces $(M_n,\mathcal B(M_n))$ and $(\mathbb{R}^{n^2},\mathcal B(\mathbb{R}^{n^2}))$ we see that $A:\Omega\to M_n$ is Borel measurable. By assumption $A$ take values in the closed subset $S_n$ of $M_n$, and so in fact $A:\Omega\to S_n$ is Borel measurable.

We conclude that the composition maps $\omega\mapsto Q(\omega):= \varphi_1(A(\omega))$ and $\omega\mapsto \Lambda(\omega):=\varphi_2(A(\omega))$ from $\Omega$ to $M_n$ are both Borel measurable and such that $A(\omega)=Q(\omega)\Lambda(\omega)Q'(\omega)$ for all $\omega\in\Omega$.


Is this proof legitimate? Thank you for your help.

Alphie
  • 4,740
  • I don’t see any mistakes. As I don’t deal with abstract measure theory everyday, there are a few things that I would like to prove to be entirely sure, such as “a matrix-valued function is Borel measurable iff all its entries are borel measurable”, or “the Borel $\sigma$-algebra of a Borel subset $A\subseteq B$ is the restriction (as a $\sigma$-algebra) of the Borel $\sigma$-algebra of $B$ to $A$“. I guess they are true and easy to prove, but they are not entirely obvious to me. Thank you for the reference to the principle of measurable choice, it is very nice to know for applications. – Lorenzo Pompili Aug 13 '21 at 21:24
  • @LorenzoPompili Thank you for your comment and feedback. The first point is proven here : https://math.stackexchange.com/a/837758/522332. The second point is proven here: https://math.stackexchange.com/a/2927656/522332. – Alphie Aug 14 '21 at 14:34
  • Perfect! These were the only thing I was not sure about. Thank you for that :) – Lorenzo Pompili Aug 14 '21 at 15:43

0 Answers0