Miller's algorithm is a device for computing the most rapidly decreasing solution of a second-order linear difference equation. In this paper strict upper bounds are given for the errors in the values yielded by the algorithm, and general conclusions are drawn concerning the accuracy of the process.