Wednesday, 27 February 2008 which I pacify the fussy NAG routine

As usual, this is of interest to probably no-one but me, but I have just found the bug in my program.
I am trying to call NAG routine G05MZF which generates random integers from a probability distribution supplied by you. I have found routines give unexpected answers if you don't pass them variables of the correct type, so I was careful about this. The error I was getting at run-time was this:

At line 106 of file /scratch/zohair/FLL3A21DF/fll3a21df/source/g/g05mzfn.f
Fortran runtime error: Missing initial left parenthesis in format

I have had errors like this before which have been due to the reference vector (which contains the CDF) being too small, but this wasn't the case here. Instead, it turned out to be angry about being passed what it considered to be a non-normalized PDF. For example, passing a PDF which was 307 entries long with each entry set to 1.0D0/307.0D0 (ie. a flat distribution) worked fine. However, setting each entry to the actual value, 0.32573290D-02 was not accurately normalised enough, and it threw a hissy fit.

So the moral of the story is:
If you think you've normalized your PDF, the NAG library knows better. Expressed in code, this is:
P = P/sum(P)

In other news, I find it incredible how easy it is to destroy my code. Simply compiling program.f90 to the output file program.f90 will replace my code with executable junk. Surely they should have idiot-proofed this?


Edit: 20th March 2008

Actually there's more to it than this. Firstly, the bizarre and unhelpful error is due to using gcc version 4.1.2 rather than the recommended version 4.2 which I don't yet have. If you use version 4.3, the error is more helpful, and gives you the value of IFAIL and results in a NAG hard failure telling you the deviation of the probability sum from unity. There are other issues too, but using the right compiler is helpful!

No comments: