[racket] matrix-solve and approximation errors

From: Konrad Hinsen (konrad.hinsen at fastmail.net)
Date: Wed Apr 16 15:18:10 EDT 2014

Neil Toronto writes:

 > The design decision is to determine how much the solvers should automate 
 > detecting and responding to floating-point error. I have no idea what 
 > extent popular linear algebra libraries go to, nor whether we should try 
 > to do better.

Any linear-algebra library meant for "serious" work does not do any
non-essential work "just" for error checking. Performance is king.
Users are supposed to know the pitfalls and take care to avoid them.
Unfortunately, many don't.

I'd say that the best compromise is a two-layer library: a low-level
layer with "raw" algorithms and good performance, and a layer on top
of that with more user-friendliness. However, I haven't seen any such
design, so developing it may well represent some serious work.


Posted on the users mailing list.