1: \begin{abstract}

2: We provide convergence rates for Krylov subspace solutions

3: to the trust-region and cubic-regularized (nonconvex) quadratic problems. 

4: Such solutions may be efficiently computed by the Lanczos method and 

5: have long been used in practice. We prove error bounds of the 

6: form $1/t^2$ and $e^{-4t/\sqrt{\kappa}}$, where $\kappa$ is a  

7: condition number for the problem, and $t$ is the Krylov subspace order 

8: (number of Lanczos iterations). We also

9: provide lower bounds showing that our analysis is sharp.

10: \end{abstract}

11: