7
$\begingroup$

I wonder if there is a paper that can point out how to compute the determinant of a $d \times d$ autoregressive correlation matrix of the form

$$R = \begin{pmatrix} 1 & r & \cdots & r^{d-1}\\ r & 1 & \cdots & r^{d-2}\\ \vdots & \vdots & \ddots & \vdots\\ r^{d-1} & r^{d-2} & \cdots & 1 \end{pmatrix}$$

$\endgroup$
1

3 Answers 3

7
$\begingroup$

The determinant of this matrix is $(1-r^2)^{d-1}$. This can be proven by induction on $d$ using the block matrix identity $\det \pmatrix{A & B\cr C & D} = \det(A) \det(D - C A^{-1} B)$, where $A$ is the top left $(d-1) \times (d-1)$ submatrix. Note that $$A \pmatrix{0\cr \ldots\cr 0\cr r\cr} = B$$ so $D - C A^{-1} B = 1 - r^2$.

$\endgroup$
1
  • $\begingroup$ This is a correct answer I misread the OP $\endgroup$
    – Henry.L
    Jul 19, 2017 at 21:57
1
$\begingroup$

The method of Dodgson's evaluation of determinants offers a simple inductive (recursive) proof.

If we denote your determinant by $R_d$, the above technique leads to $$R_d=\frac{R_{d-1}^2-0^2}{R_{d-2}}=\frac{(1-r^2)^{2d-4}}{(1-r^2)^{d-3}}=(1-r^2)^{d-1},$$ since the top right-most (and bottom left-most) matrices of size $(d-1)\times(d-1)$ have vanishing determinant.

$\endgroup$
0
$\begingroup$

A thorough discussion is contained in

Finch, P. D. "On the covariance determinants of moving-average and autoregressive models." Biometrika 47.1/2 (1960): 194-196. JSTOR

But it is more common to write it in form of product of conditional densities $f(x_2\mid x_1)f(x_3\mid x_1,x_2)\cdots $ and consider each update sequentially to discuss the behavior of the covariance matrix.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.