Hello, I'm Frank Ferraro, a Ph.D. student in computer science at Johns Hopkins University. I've been using GSL for a research project, and believe I've found an issue with the bfgs2 implementation.
The smallest example I could come up with is at https://github.com/fmof/maxent-mwe. The (convex) function is the (negative) log-likelihood of a maximum entropy model. Please see README.md for steps-to-reproduce <https://github.com/fmof/maxent-mwe#compile>, a more detailed overview <https://github.com/fmof/maxent-mwe#what-the-program-does> of the program, the optimizable objective function, and actual vs. expected output <https://github.com/fmof/maxent-mwe#output-actual-vs-expected>. I've tested this with GSL 1.16, with multiple levels of optimization (including -O0). My machine is: * x86_64 Linux (Ubuntu 14.04) * GCC 4.8.2 (Ubuntu 4.8.2-19ubuntu1) * GNU make 3.81 * 16GB of RAM, 3MB cache * Intel(R) Core(TM) i5-3320M CPU @ 2.60GHz Thank you, Frank Ferraro http://cs.jhu.edu/~ferraro/
