2
$\begingroup$

My current problem involves having an exact (symbolic) inverse of a scaled AR(1) matrix for $n$-dimension. (I don't know what this matrix would be called in general; I'm sure it is used often.) This is used as a smoothing prior on a function sampled on a uniform grid. For a $1$-dimensional function, the matrix is

$$C = \rho \begin{bmatrix} 1 & \alpha & \alpha^2 & \cdots & \alpha^n \\\\ \alpha & 1 & \alpha & \cdots & \alpha^{n-1}\\\\ \alpha^2 & \alpha & 1 & \\\\ \vdots & \vdots & & \ddots & \\\\ \alpha^n & \alpha^{n-1} & & & 1\end{bmatrix}$$

and I know the inverse, which is

$$C^{-1} = \frac{1}{\rho(1-\alpha^2)} \begin{bmatrix} 1 & -\alpha & & & 0\\\\ -\alpha & 1+\alpha^2 & - \alpha \\\\ & -\alpha & 1+\alpha^2 &\ddots \\\\ & &\ddots & \ddots & -\alpha \\\\ 0 & & & -\alpha & 1\end{bmatrix}$$

(This can be also found in Kac, M., Murdock, W., and Szegö, G. (1953). On the eigenvalues of certain Hermitian forms. J. Rational Mech. Anal, 2:767–800.)

I would imagine that this can be generalized to higher dimensional case where the $\alpha$ now spreads in each direction. This would allow my uniformly sampled $n$-dimensional function to be smooth. Sort of having the form

\begin{equation} C_{(i,j),(k,l)} = \rho \alpha_x^{|i-k|} \alpha_y^{|j-l|}, \end{equation}

but as a giant matrix for flattened function (the vec operation; representing the $n$-dimensional function as a vector with some ordering). Can anybody recommend a book on such symbolic matrix inversions that would have this?

$\endgroup$
1
  • $\begingroup$ @RodrigodeAzevedo, this is not normal: editing a 7 years old post, inactive, only to perform minor changes! Do you realize that this bumps it to the main page, for nothing? $\endgroup$ – Alex M. Jun 11 '18 at 17:42
2
$\begingroup$

Your $C_{(i,j)(k,l)}$ is the Kronecker product of two $C$ matrices (with different constant $\alpha_x$ and $\alpha_y$. And $(A \otimes B)^{-1}=A^{-1}\otimes B^{-1}$.

$\endgroup$
3
  • $\begingroup$ If $C_{(i,j),(k,l)}=\rho \alpha_x^{∣i−j∣} \alpha_x^{∣k−l∣}$, then I would agree, but the indices are a bit mixed. Does it still hold? $\endgroup$ – Memming May 24 '11 at 19:16
  • $\begingroup$ I think so, unless I'm totally misunderstanding your notation. The pair $(i,j)$ forms the row index of the large matrix, doesn't it? $\endgroup$ – Federico Poloni May 24 '11 at 21:11
  • $\begingroup$ Correct. Well, $(i,j)$ for column, $(k,l)$ for row, but that's not important. The Kronecker product of $C_{i,k}$ and $C_{j,l}$ is correct. Thanks. $\endgroup$ – Memming May 24 '11 at 22:04

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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