# What is the expected root mean square determinant of an $n\times n$ matrix?

The expected mean determinant of random nxn matrices of 0s and 1s is 0. What is the expected root mean square determinant? e.g. $\frac{\sqrt{3}}{2\sqrt{2}}$ for a $2\times 2$

-
Actually rotating a matrix by $90$ degrees does not always change its sign. Works if the size is $2$ or $3$, but not if the size is $1$, $4$, $5$ (or even $0$), in general not for sizes $0$ or $1$ modulo $4$. –  Marc van Leeuwen Feb 6 '12 at 10:19
$\frac38$ is the mean square determinant, not the root mean square determinant for $2\times2$. –  joriki Feb 6 '12 at 23:29

As you say, the expected mean determinant is zero:

$$\def\expect#1{\left\langle#1\right\rangle} \begin{eqnarray} \expect{\det A} &=& \expect{\sum_\pi\sigma(\pi)\prod_iA_{i\pi(i)}} \\ &=& \sum_\pi\sigma(\pi)\expect{\prod_iA_{i\pi(i)}} \\ &=& \sum_\pi\sigma(\pi)\expect{\prod_iA_{ii}} \\ &=& \expect{\prod_iA_{ii}}\sum_\pi\sigma(\pi) \\ &=& 0\;. \end{eqnarray}$$

The expected mean squared determinant of an $n\times n$ matrix can be calculated similarly:

$$\begin{eqnarray} \expect{(\det A)^2} &=& \expect{\left(\sum_\pi\sigma(\pi)\prod_iA_{i\pi(i)}\right)^2} \\ &=& \expect{\left(\sum_\pi\sigma(\pi)\prod_iA_{i\pi(i)}\right)\left(\sum_\rho\sigma(\rho)\prod_iA_{i\rho(i)}\right)} \\ &=& \expect{\sum_{\pi,\rho}\sigma(\pi)\sigma(\rho)\prod_iA_{i\pi(i)}\prod_iA_{i\rho(i)}} \\ &=& \expect{\sum_{\pi,\rho}\sigma(\pi^{-1}\rho)\prod_iA_{ii}\prod_iA_{i(\pi^{-1}\rho)(i)}} \\ &=& n!\expect{\sum_{\pi}\sigma(\pi)\prod_iA_{ii}\prod_iA_{i\pi(i)}}\;. \\ &=& n!\sum_{\pi}\sigma(\pi)\expect{\prod_iA_{ii}\prod_iA_{i\pi(i)}}\;. \end{eqnarray}$$

The expectation value in the last line has $n$ factors of $1/2$ from the diagonal elements and another factor of $1/2$ for each $i$ that is not a fixed point of $\pi$. Thus we have

$$\begin{eqnarray} \expect{(\det A)^2} &=& \frac{n!}{2^n}\sum_\pi\sigma(\pi)M_{i\pi(i)} \\ &=& \frac{n!}{2^n}\det M \end{eqnarray}$$

with

$$M_{ij}=\begin{cases}\vphantom{\frac12}1&i=j\;,\\\frac12&i\ne j\;.\end{cases}$$

This matrix has $n-1$ eigenvalues of $1/2$ and one eigenvalue of $(n+1)/2$, and so

$$\begin{eqnarray} \expect{(\det A)^2} &=& \frac{n!}{2^n}\frac{n+1}{2^n} \\ &=& \frac{(n+1)!}{2^{2n}} \end{eqnarray}$$

and

$$\sqrt{\expect{(\det A)^2}}=\frac{\sqrt{(n+1)!}}{2^n}\;,$$

in agreement with your result for $n=2$.

See also OEIS sequence A055137 for more on the connection to fixed points of permutations.

The calculation doesn't rely on the distribution of the matrix elements up to the point where we plug in the factors of $1/2$; evaluating the result for matrix elements uniformly chosen as $\pm1$ instead and noting that in this case any unpaired matrix element causes the expectation value to vanish, we obtain a mean squared determinant of $n!$ for a random matrix of $\pm1$s.

-
That the mean determinant is zero is perhaps even easier to see by symmetry. Let $A'$ be the matrix $A$ with the first two rows swapped. Then $\det A' = - \det A$. On the other hand, $A$ and $A'$ are identically distributed, so $\mathbb{E} \det A = \mathbb{E} \det A' = - \mathbb{E} \det A$. –  Nate Eldredge Feb 7 '12 at 23:40
@Nate: Good point. As you can still see from Marc's comment, the question had originally contained a claim of proof by symmetry that turned out to be wrong. Yours is right, of course. My purpose in that first part wasn't so much to give the most elegant proof of the result for the mean determinant but to introduce the relabeling technique in a simpler setting with a known result before applying it to the question. –  joriki Feb 8 '12 at 4:58