Backward differentiation formulae with nonnegative coefficients for solving initial value problems (Q1202765)

A numerical method for treating initial value problems in ordinary differential equations involves linear combinations of derivative values at a number of discrete points. The conditions which are imposed to achieve high algebraic accuracy of the method equate sums of subsets of the coefficients in each combination to fixed constants. To minimize error resulting from finite precision arithmetic, it is desirable that each sum of absolute values of these coefficients is minimal, and this is achieved if all coefficients are nonnegative. The author determines that only a subset of order 1 backward-differentiation methods (BDF), and only subsets of orders 2 and 3 second-derivative BDF methods have this property. Using several methods of the final type, the author illustrates that methods with nonnegative coefficients tend to have smaller arithmetic truncation errors for a linear constant-coefficient problem. In order to control local error as a means to controlling global error, most software for treating such problems is designed so that error due to mathematical truncation is the dominating component of error. Accordingly, for many problems, the most effective (stable) algorithms are likely to be those for which coefficients are restricted primarily by considerations for making the mathematical truncation error small, but significant. If some freedom remains in choosing the coefficients, attempts to minimize errors due to numerical imprecision can be of additional advantage for some problems.