Re: [Help-gsl] use of recurrence relation while computing mean, variance etc.

[Juan Pablo Amorocho D.]

Hi Brian and Awhan,

just a quick follow-up on the recurrence formula of the variance. Awhan, if
you are interested in a deeper understanding of this issue and have access
to the following book, I would recommend you take a look at Higham,
N.*Accuracy and Stability of Numerical Algorithms
*. Second Edition. SIAM, 2002. Section 1.9 is called "Computing the Sample
Variance" .

Cheers,
-- Juan

2011/8/7 Brian Hawkins <address@hidden>

> Hi Awhan,
>
> I didn't write the code, but I think it's safe to say that the
> recurrence relations and use of long double is indeed motivated by
> numerical concerns.  Running sums are a common cause of overflow.
> Variables inside a loop have local scope but otherwise behave
> similarly as other variables.  IMO this coding style is cleaner, and
> it also makes parallelism simpler to implement.
>
> Regards,
> Brian
>
> > Date: Sun, 7 Aug 2011 04:22:49 +0530
> > From: Awhan Patnaik <address@hidden>
> > To: address@hidden
> > Subject: [Help-gsl] use of recurrence relation while computing mean,
> >        variance etc.
> > Message-ID:
> >        <CAM+ON=
> address@hidden>
> > Content-Type: text/plain; charset=ISO-8859-1
> >
> > hello all,
> >
> > i have 3 questions to bother you with.
> >
> > 1) what is the motivation behind using the recurrence relation for the
> > computation of the mean ? seems to me that a division by the number of
> > elements *after* the for loop will result in fewer division operations
> > than the current implementation. i have noticed the use of recurrence
> > type computation of variance as well.
> >
> > 2) many functions have a return type of double but the quantity of
> > interest that is to be returned is declared as a long double inside
> > the function body. why is this done? is it that conversion from a long
> > double to double results in less loss of precision?
> >
> > 3) declaration of variables inside the for loop body e.g. delta in the
> > following snippet from variance_source.c
> >
> > /* find the sum of the squares */
> >  for (i = 0; i < n; i++)
> >    {
> >      const long double delta = (data[i * stride] - mean);
> >      variance += (delta * delta - variance) / (i + 1);
> >    }
> >
> > does the compiler keep creating a local delta each time control enters
> > the loop or is it created once but treated as a local variable and
> > valid only in the scope of the for loop?
>
> _______________________________________________
> Help-gsl mailing list
> address@hidden
> https://lists.gnu.org/mailman/listinfo/help-gsl
>