GoogleCodeExporter
commented
9 years ago If you accept my arguments that the code must be changed, here is a better
method for doing it than using sum(exp(log(dij) − bic[j] + min_bic)). It is
mathematically equivalent, just as numerically stable, and avoids an unncessary
log. The translation from C to Java should be straightforward.
//assumptions:
// * branch length vectors 'b[i]' of length 'nedges' have been stored
// for all models 'i'
// * BIC values have already been stored into an array 'bic'
// * bic_min has already been determined
double denom = 0.0;
for (uint i = 0; i < nmodels; i++)
{
/* Note that we don't need to worry about underflow here, as at least
one of the "bic[i] - bic_min" values will be zero, and any terms
that underflow to zero when exponentiated would make a negligible
contribution to the sum anyway.
*/
bic_like[i] = exp_safe(-0.5*(bic[i] - bic_min));
denom += bic_like[i];
}
for (uint i = 0; i < nmodels; i++)
{
double sum = 0.0;
for (j = 0; j < nmodels; j++)
if (j != i)
sum += euclideanDistance(b[i], b[j], nedges)*bic_like[j];
dt[i] = sum/denom;
}
Original comment by dave.swo...@gmail.com on 17 Mar 2014 at 4:16
Original issue reported on code.google.com by
dave.swo...@gmail.comon 17 Mar 2014 at 3:07Attachments: