This paper is a low-level conference presentation tutorial on a computational approach to estimating condition numbers for, or analyzing the sensitivity of, a computational process. The basis of the approach is the standard one of recomputing the results for nearby data, though the definition of “nearby” is perhaps too broad to detect major instabilities.
Computational problems are subdivided into three categories. The sector of interest is a subclass of a small change in inputs leading to a large variation in outputs, characterized by the fact that the “change is unacceptable; it does not correctly model reality and indicates an error.” The interpretation is that the program needs to be revisited. It is important to note that this type of numerical instability may be inherent in the model or in the algorithm used, not necessarily in its implementation.
The principal achievement described in the paper is an automated sensitivity analyzer that runs in conjunction with an Excel spreadsheet to generate variations on input data, and therefore allows the user to analyze the behavior. Unfortunately, it does not appear from the description that it is possible to select particular inputs for analysis, which appears to make it difficult to analyze sensitivity to a particular input, for example.