I have provided a patch that implements the function gsl_root_fsolver_set_with_values(), which takes the function values as argument rather than evaluating the (potentially costly) function internally, see here: https://savannah.gnu.org/bugs/index.php?66576

Hello,

Root Bracketing Algorithms (bisection, brent, and falsepos) expect that the sign of f(a) and f(b) be different so that the endpoints do straddle f(x)=0. gsl_root_fsolver_set returns GSL_SUCCESS when the supplied interval is good and throws GSL_ERROR ("endpoints do not straddle y=0", GSL_EINVAL); otherwise.

I've a particular case where I've to split up my complete domain into subdomains to avoid singularities. My root is an angle and I've to avoid 0, 90, 180, and 360 degrees. So I split my domain into 4 quadrants; only one of which will be a valid interval. Now I could either test each interval before calling gsl_root_fsolver_set or redefine gsl's error handler to not "abort()". Option 1 is a performance issue for me because f(x) is a costly function. On the one hand gsl_root_fsolver_set expects the user to check beforehand that the end points do straddle y=0. But it checks again internally the validity of the same condition. Option 2 is what I do now.

If gsl_root_fsolver_set returns GSL_EINVAL instead of throwing an error it would give the user an option to handle the error (without redefining gsl's error handler) and also avoid checking the "endpoints do straddle y=0" condition twice.

Regards,
Navi