 |
 |
> > IMSL carried over the ENTIRE content of > > its Fortran 77 library to Fortran 90. So did Numerical > > Recipes. > > Are Numerical Recipes any good yet? I haven't looked for about a > decade, but the last time I had correspondence with Bill Press he > couldn't understand that the fundamental mathematical foundation of > MEDFIT (an L1 fitting routine) was flawed: It depended on derivatives > exactly and only where they don't exist! Then he couldn't understand > that it was a Bad Thing that the (flawed) mathematics was converted into > an algorithm incorrectly, or that the (inherently incorrect) algorithm > was coded incorrectly. Lots of other areas were defective. Have they > been repaired? I haven't followed this particular problem (which you have mentioned in various places many times), but certainly later editions of NR have corrected errors in earlier versions. I also appreciate the fact that NR were one of the earliest F90 evangelists. I have never used MEDFIT, but have used a dozen or so other NR routines (and in addition several more which are used internally by these). I have never found that they don't live up to what they claim to be. (Of course, they don't claim to be NAG-quality routines, especially with regard to error handling. The idea is that one has the discussion of an algorithm and its implementation in source code in one place. Moreover, this source code is readable for the novice. (Normal users don't have access to the NAG source code at all, of course.)) They should be thought of as jumping-off points for one's own implementation. (I have also used the NAG routines when they were the right tool for the job.)