Many drivers (GLE-Langevin, HMC, some surrogates) benefit from or require gradients.

Supplying a gradient

• Implement grad(self, x) on your Objective subclass (returns ndarray).

• Or wrap a Python callable + grad callable with the binding helpers.

• When no gradient is available, finite differences are used (n_p+1 cost per gradient evaluation) and charged against the budget.

See the GLE and HMC tutorials in anneal for end-to-end usage.