ott.neural.networks.icnn.ICNN.potential_value_fn

ICNN.potential_value_fn(params, other_potential_value_fn=None)

Return a function giving the value of the potential.

Applies the module if is_potential is True, otherwise constructs the value of the potential from the gradient with

\[g(y) = -f(\nabla_y g(y)) + y^T \nabla_y g(y)\]

where \(\nabla_y g(y)\) is detached for the envelope theorem [Bertsekas, 1971, Danskin, 1967] to give the appropriate first derivatives of this construction.

Parameters:

• params (FrozenDict[str, Array]) – parameters of the module

• other_potential_value_fn (Optional[Callable[[Array], Array]]) – function giving the value of the other potential. Only needed when is_potential is False.

Return type:

Callable[[Array], Array]

Returns:

A function that can be evaluated to obtain a potential value, or a linear interpolation of a potential.