Owner import path¶
jaxstro.numerics.autodiff
Purpose¶
Named wrappers expose JVP, VJP, Hessian-vector, Gauss-Newton, and empirical Fisher products without owning a scientific derivative-validity policy.
Public records and callables¶
jvp, vjp, jacobian_vector_product, vector_jacobian_product, hvp,
gauss_newton_product, and empirical_fisher_product.
Shape and dtype expectations¶
Inputs and tangents must follow the wrapped function’s PyTree shapes. Products preserve JAX array dtypes; mixed or integer differentiation follows JAX rules.
JAX transforms and AD classification¶
These helpers compose JAX transformations. Their derivatives are only as scientifically meaningful as the selected function and branch.
Failure behavior¶
JAX tracing, shape, or dtype errors propagate. The module does not replace non-finite products or certify derivative meaning.
Contract and evidence links¶
See Autodiff products and Validation.
Canonical import example¶
from jaxstro.numerics.autodiff import hvp