Will GMRES always converge after one iteration if the initial guess vector is an eigenvector?

286 Views Asked by At

Will the GMRES iterative method for finding solutions to systems of equations, $Ax=b$, always converge after one iteration if the initial guess vector $b$ ($q_1 = b / \|b\|$) is an eigenvector of $A$? I think this is true, but I can't seem to make a strong statement about it, but what I have so far is.


Given an eigenvector of $A$, $b$, at the first iteration, well even before the first iteration, of the Arnoldi iteration we have $q_1 = b / \|b\|$. If we then take the Ritz value, $\theta_1 = q_1^T A q_1$, at this first iteration we find that it is equivalent to the Rayleigh quotient at an eigenvector of $A$, which produces the correspodning eigenvalue of $b$, $\lambda_b$.

Next, in GMRES we attempt to find $x_n \in K_n$ where $K_n$, the $n$th Krylov subspace, is produced from the vectors $q_j$ of the Arnoldi iteration. This potential solution vector $x_n$ can be written in terms of these $q_j$ as $x_n = Q_n y_n$ (where $y_n \in \Re^n)$.

At this first iteration we know $Q_n$ contains only $q_1$ which is a normalized eigenvector of $A$. Thus we have $x_1 = q_1 y_1$. If we allow $y_1 = 1 / \lambda_b$ then $x_n$ is a solution to $Ax=b$. (We allow $y_1$ to equal this because a solution to $Ax=b$ when $b$ is an eigenvector of $A$ is $x = \frac{1}{\lambda} b$).


I think this makes sense, to me at least, but again it seems kind of weak. Specifically it doesn't address how long Arnoldi is ran before GMRES commences. I want to say $0$ iterations (so that $Q_1$ only contains $q_1$), but I'm not sure if that is necessary to enforce. It also hinges on that fact that we can "arbitrarily" select the value of $y_1$, though, if $b$ is the eigenvector shouldn't $y_1$ be selected as this value to minimize the residual?


Couldn't a separate, much simpler, proof be made stating that the solution vector lies in the $K_1$ subspace because $x = \frac 1 \lambda b$ lies in $K_1 = \operatorname{span} \lbrace b \rbrace$?