toms748_solve produces -nan when evaluating a smooth function in a fixed interval

[Ola Skavhaug <ola@…>]

In my algorithm that is finding the point where two gamma gamma distributions crosses, the toms748 solver produces a -nan value for the evaluation point. The behaviour occurs quite rarely, and only for extremely specific values of the governing parameters. I have not checked if the problem is cross platform.

I see that the solver evaluates the same point twice, and suspect that the subsequent calculation of the change of the evaluation point somehow divides by 0.

A minimal example is enclosed.

[Ola Skavhaug <ola@…>]

Minimal example, compile with: g++ -std=c++11 -fPIC -O2 -o test test.cpp

[anonymous]

can confirm, same bug, ubuntu 14.04 and gcc

[John Maddock]

Thanks, I tried to reproduce with msvc but no luck, I'll look into GCC later.

[John Maddock]

There's a typo/bug in the root bracketing code, fixed in ​https://github.com/boostorg/math/commit/8a29241cc3a59fe04b10ab6937d8eefe07dd85fb

Note this will not be in the next release which already in beta.