Automatic reasoning about numerical stability of rational expressions

摘要

While numerical (e.g. Fortran) code generation from computer algebra is nowadays relatively easy to do, the expressions as they are produced via computer algebra typically benefit from non-trivial reformulation for efficiency and numerical stability. To assist in automatic "expert reformulation", we desire good automated tools to assess the stability of a particular form of an expression. In this paper, we discuss an approach to proofs of numerical stability (with respect to roundoff error) of rational expressions. The proof technique is based upon the ability to propagate properties such as sign, exact representability, or a certain kind of numerical stability, to floating point results from properties of their antecedents.

rn

The qualitative approach to numerical stability (inspired by [12]) lends itself to implementation as a backwards-chaining theorem prover. While it is not a replacement for alternative forms of stability analysis, it can sometimes discover stability and explain it straightforwardly.