improve on relative tolerance for constructing tables
The existing approach is not robust with respect to function or derivative values near zero, and can fail in double precision for different compilers, build types, and number of data points tested. We need an approach to testing that our functions do that right job that is more resilient to the kinds of arbitrarily curved functions users might want to use, e.g. from Boltzmann inversion.
#3 Updated by Erik Lindahl about 1 month ago
- Subject changed from improve on relative tolerance for testing implementation of tabulated interaction tables to improve on relative tolerance for constructing tables
This is not mainly related to testing, but how to construct the tables, since we presently use relative tolerances to decide the table density. Once we have a robust solution there, we should just apply the same one during testing.