line search backtracking algorithm


Does anyone here know how the line search backtracking algorithm implemented in Lammps compares to the linmin.cpp from Numerical Recipes (NR)? Just trying to know if the motivation for the line backtracking in Lammps was performance-related (faster than the NR one) or you just could not use the NR one due to copyright issues…

I will do the comparison on my own anyways but wanted to gather some initial

thoughts from the people who implemented it.


The one we used is in the Shewchuk paper referenced in src/min_linesearch.cpp.
We do avoid NR when possible due to the copyright restrictions.


Thanks. I knew the reference but was wondering about the choice per se. Was interested because in methods like hyperdynamics the lowest eigenvector of the Hessian has to be computed by an iterative procedure that avoids explicit calculation of the Hessian. If hyperdynamics is at some point to be implemented and CG employed within the eigenvector calculation, a fast linesearch method is a must to deal with big systems.