Consider as an example the following Var(1) process: $y_t=Gy_{t-1}+\varepsilon_{t+1}$, where $\varepsilon_{t+1}$ is white noise and the matrix $G$ has all eigenvalues strictly smaller than 1. Then by standard results the process is stationary, and since $\sum_iG^i$ is absolutely summable, it can be represented as a MA($\infty$): $y_t=\sum_{i=0}^{+\infty}G^i \varepsilon_{t-i}$.

Now, all the books I checked (e.g. Lutkepohl proposition C.10, Brockwell Davis chapter 11, Hamilton prop 10.2) state that, without any further condition beyond absolute summability, we get that the autocovariance is: $\Gamma(k)=\sum_i^{+\infty} G^{k+i}(G')^{i}$

However, consider the matrix $G=0.9\left(\begin{array}{ccc}0 & 1 & 0.5 \\ 0 & 0 & 0.5\\ 1& 0 & 0\end{array}\right)$

The eigenvalues of $G$ have all modulus smaller than 1 (they are 0.9, 0.9(0.5$\pm$0.5i)), hence the process is stationary and can be represented as a moving average.

Yet, the expression $\Gamma(k)=\sum_i^{+\infty} G^{k+i}(G')^{i}$ does not converge, because $GG'$ has an eigenvalue bigger than 1 (barely, around 1.05).

As a further evidence, for example for $k=0$ we would have, if it converged:


since its a geometric series. But this cannot be the limit because it has negative terms (calculation approximate because done with the pc): $(I-GG')^{-1}=\left(\begin{array}{ccc}-15 & -4 & 0 \\ -4 & 0.25 & 0\\ 0& 0 & 5\end{array}\right)$

So it seems that some further condition is needed to ensure that the process has a finite variance, or at least that the variance can be calculated with that sum. Indeed, this makes sense, because in the one dimensional case $\sum_i g^i$ and $\sum_i g^{2i}$ both converge iff $|g|<1$, but in the matrix case, in general there is no guarantee that the eigenvalues of a product $GG'$ are not bigger than the product of eigenvalues (as the example above shows).

Am I right, or is there something I am missing? Do someone know a reference where some of these conditions are discussed? For example, if we assume $G$ symmetric then the matrix to be summed is $G^2$, and the eigenvalues of $G^2$ will all be smaller than one if and only if the eigenvalues of $G$ are, so in this case it works.


1 Answer 1


Because matrix multiplication is not commutative in general, it is not true that $$ \sum_{i=0}^\infty G^i (G')^i = \sum_{i=0}^\infty (GG')^i. $$
This would only hold if $G$ is normal. So your formula for $\Gamma_k$ does not simplify to a geometric series in general.

For non-normal $G$ you can solve for $\Gamma_0$ using the alternative method based on vectorization of the matrix equation $\Gamma_0 = G\Gamma_0 G' + \Sigma$.


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.