A Randomized Algorithm for Preconditioner Selection

The task of choosing a preconditioner $\boldsymbol{M}$ to use when solving a linear system $\boldsymbol{Ax}=\boldsymbol{b}$ with iterative methods is difficult. For instance, even if one has access to a collection $\boldsymbol{M}_1,\boldsymbol{M}_2,\ldots,\boldsymbol{M}_n$ of candidate preconditioners, it is currently unclear how to practically choose the $\boldsymbol{M}_i$ which minimizes the number of iterations of an iterative algorithm to achieve a suitable approximation to $\boldsymbol{x}$. This paper makes progress on this sub-problem by showing that the preconditioner stability $\|\boldsymbol{I}-\boldsymbol{M}^{-1}\boldsymbol{A}\|_\mathsf{F}$, known to forecast preconditioner quality, can be computed in the time it takes to run a constant number of iterations of conjugate gradients through use of sketching methods. This is in spite of folklore which suggests the quantity is impractical to compute, and a proof we give that ensures the quantity could not possibly be approximated in a useful amount of time by a deterministic algorithm. Using our estimator, we provide a method which can provably select the minimal stability preconditioner among $n$ candidates using floating point operations commensurate with running on the order of $n\log n$ steps of the conjugate gradients algorithm. Our method can also advise the practitioner to use no preconditioner at all if none of the candidates appears useful. The algorithm is extremely easy to implement and trivially parallelizable. In one of our experiments, we use our preconditioner selection algorithm to create to the best of our knowledge the first preconditioned method for kernel regression reported to never use more iterations than the non-preconditioned analog in standard tests.