In Code/Numerics/Optimizer/BFGSOpt.h, the gradient-convergence check
computed
double term = std::max(funcVal * gradScale, 1.0);
...
test /= term;
if (test < gradTol) return 0;
When funcVal (the current energy) is negative, funcVal * gradScale is
negative and std::max clamps the denominator to 1.0. The convergence
test therefore divides the gradient norm by 1 instead of by the
intended |E| * gradScale, which over-tightens the criterion by a factor
of |funcVal * gradScale| whenever |funcVal * gradScale| > 1.
Negative energies are a normal mid-minimization state for force fields
that include stabilizing terms (MMFF94, UFF with charges, AMBER-style
potentials), so this affects realistic workloads: extra BFGS iterations
or, occasionally, hitting MAXITS and returning the "too many iterations"
status when convergence would otherwise have been reached.
The fix is to use |funcVal| in the denominator, matching the pattern
used three lines below ('std::max(fabs(pos[i]), 1.0)') and matching the
intended interpretation as a magnitude.
A new test case 'testBFGSOptimizationNegativeEnergy' in
testOptimizer.cpp minimizes a 2D quadratic whose value is always
negative along the convergence path and verifies the optimizer reaches
the analytic minimum.
git blame attributes the original line to commit e08e0d16d (Nov 2015),
when the optimizer was restructured; the surrounding code does use
absolute values, so this reads as an oversight rather than an
intentional choice.
