A model for understanding numerical stability

^** Email: bornemann{at}ma.tum.de We present a model of roundoff error analysis that combinessimplicity with predictive power. Though not considering allsources of roundoff within an algorithm, the model is relatedto a recursive roundoff error analysis and therefore is capableof correctly predicting stability or instability of an algorithm.By means of nontrivial examples, such as the componentwise backwardstability analysis of Gaussian elimination with a single iterativerefinement step, we demonstrate that the model even yields quantitativebackward error bounds that show all the known problem-dependentterms with the exception of dimension-dependent constants. Themodel can serve as a convenient tool for teaching or as a heuristicdevice to discover stability results before entering a furtherdetailed analysis.