dq.gradient.ForwardAutograd

ForwardAutograd()

Forward-mode automatic differentiation.

Enables support for forward-mode automatic differentiation (like jax.jvp or jax.jacfwd).

Note

This is the most efficient when the function has more outputs than inputs. For instance, it's the preferred method when simulating a Lindbladian parameterized with a few values and computing the Jacobian of a function returning the expectation values of many observables (or the same observable at many different times).

Warning

This cannot be backward-mode autodifferentiated (e.g. using jax.jacrev). Try using dq.gradient.CheckpointAutograd if that is something you need.

Warning

By default jax.grad uses reverse mode. Use jax.jacfwd to compute the gradient in forward mode.

Note

For Diffrax-based methods, this falls back to the diffrax.ForwardMode option.