How we train numerical programmers

["E. Robert Tisdale" <edwin at netwood dot net>]

There is a problem with
the way we train numerical programmers.
Most mathematicians, scientists and engineers
are amateur application programmers
who write code for their own use
and do little or no testing before
they attempt to use their programs.
Most professional application programmers,
on the other hand, are computer scientists
who write code for other people to use
and they do extensive testing before
the application is placed into service.

Numerical class libraries should be designed
by experts in numerical analysis,
computer architecture and computer arithmetic
so that neither amateur or professional
application programmers are obliged
to muck about with the implementation
of the numerical class libraries that they use.

Part of the problem is that we don't teach
numerical programmers to distinguish
between the role of the class library developer
and the role of the application programmer.
We teach numerical programmers
the details of numerical algorithms
that are implemented as library functions
as if we expected all of them
to become numerical class library developers
when, in fact, most of them will actually
become numerical application programmers.
For the most part, they would be better off
if we would teach them how to use, evaluate
and select a numerical class library.
We need more numerical application programmers
NOT more numerical class library developers.