1: \begin{abstract}

2: We establish a near-optimality guarantee for the full orthogonalization method (FOM), showing that the \emph{overall} convergence of FOM is nearly as good as GMRES. 

3: In particular, we prove that at every iteration $k$, there exists an iteration $j\leq k$ for which the FOM residual norm at iteration $j$ is no more than $\sqrt{k+1}$ times larger than the GMRES residual norm at iteration $k$.

4: This bound is sharp, and it has implications for algorithms for approximating the action of a matrix function on a vector.

5: \end{abstract}

6: