A Model for Understanding Numerical Stability

We present a model of roundoff error analysis that combines simplicity with predictive power. Though not considering all sources of roundoff within an algorithm, the model is related to a recursive roundoff error analysis and therefore capable of correctly predicting stability or instability of an algorithm. By means of nontrivial examples, such as the componentwise backward stability analysis of Gaussian elimination with a single iterative refinement step, we demonstrate that the model even yields quantitative backward error bounds that show all the known problem-dependent terms (with the exception of dimension-dependent constants, which are the weak spot of any a priori analysis). The model can serve as a convenient tool for teaching or as a heuristic device to discover stability results before entering a further, detailed analysis.