Lazy Jacobians

A LazyJacobian represents a Jacobian by its matrix-free operators instead of a dense or sparse matrix, which is useful when the Jacobian is large but its Jacobian-vector and vector-Jacobian products are cheap.

LazyJacobian(; jvp = (v) -> ..., jtvp = (w) -> ..., symmetric = false)

• jvp(v) computes the Jacobian-vector product J * v.
• jtvp(w) computes the vector-Jacobian product J' * w.
• when symmetric = true, the matrix is symmetric and only one of the two operators needs to be supplied.

LazyJacobian overloads adjoint, transpose and *, so it can be used wherever a matrix is expected, including inside the derivative functions of NonconvexCore and any ChainRules-based outer differentiation.