Mixed-precision iterative refinement for low-rank Lyapunov equations
We develop a mixed-precision iterative refinement framework for solving low-rank Lyapunov matrix equations $AX + XA^T + W =0$, where $W=LL^T$ or $W=LSL^T$. Via rounding error analysis of the algorithms we derive sufficient conditions for the attainable normwise residuals in different precision settings and show how the algorithmic parameters should be chosen. These conditions are independent of the choice of inner solver, provided that the prescribed residual accuracy is attained in the inner solves. Using the sign-function Newton iteration as the solver, we demonstrate that reduced precisions, such as half precision with unit roundoff $u_s$, can be used for Lyapunov equations with condition number of order $u_s^{-1}$ without compromising the attainable solution quality. The numerical experiments validate our error analysis, and we establish cost models to quantify the expected speedup for different low-to-high precision timing ratios. This work provides an algorithmic framework towards exploiting native low-precision hardware to accelerate low-rank Lyapunov solvers without sacrificing accuracy.