Re: [Help-gsl] Bessel functions and continued fractions

[Jonny Taylor]

Thanks for your reply Brian.

>> Is there anybody where who knows the ins and outs of the Bessel
>> function implementation in GSL? Under certain circumstances my code
>> is spending over 90% of its time in gsl_sf_bessel_J_CF1. This occurs
>> for large x<1000 and n<50. The time taken to evaluate the continued
>> fraction grows with x.
> 
> Can you send a test case where it's really bad for adding to the test
> suite (specfunc/test_bessel.c), I appreciate that it's a performance
> problem rather than an accuracy problem but it's easiest to keep track
> of it if it's in there.  Are the results actually accurate? Usually
> these excessive numbers of iterations are a sign of poor convergence.
The following is demonstrates the problem:
    /*  Testcase demonstrating long calculation time:
        Time spent in gsl_sf_bessel_J_CF1 for large x<1000 and n<5
        grows in proportion to x */
    TEST_SF(s, gsl_sf_bessel_Jn_e, (45, 900.0, &r),     0.02562434700634278108, 
   TEST_TOL0, GSL_SUCCESS);

The results do appear to be accurate (the reference value here is taken from 
Mathematica) in spite of the large number of iterations.

>> Is there a reason that CF2 is not used in GSL? I don't know about
>> the reasoning behind the implementation and the regimes where each
>> of the various evaluation methods are used. Is there a reason for
>> not using CF2 in the appropriate circumstances, or is it just that
>> nobody has felt the need to implement it up to now?
> 
> I do not know about the choices behind it, the original author was
> address@hidden  If you can make a CF2 patch which is faster and
> passes the test suite that would be the way forward.
Unfortunately it turns out I wasn't reading things properly. Although CF2 
converges faster in some cases, it cannot on its own be used to calculate Jn: I 
was misunderstanding what I was reading. Sadly there does not appear to be an 
obvious way of speeding things up.

Jonny