GaussNewtonModel.jacobian

t3toolbox.fitting.GaussNewtonModel.jacobian(p)

def jacobian(
        self,
        p:  t3m.T3Tangent,
) -> NDArray:  # J p = 𝒥(Π p); apply/entries: shape W+C; probe: len=d, elm_shape=W+C+(Ni,)

The linearized forward J p = 𝒥(Π p) (the Gauss-Newton Jacobian-vector product).

ONE forward sweep – no transpose 𝒥ᵀ, no gauge re-projection of the output, no tangent assembly. The general forward primitive: it gives the predicted residual r + J p (for trust-region / line-search predicted reduction) and, via gn_quadratic(), the Gauss-Newton quadratic form. The result lives in the sample space (a scalar per sample for apply / entries; one vector per mode for probe).

Parameters:

p (T3Tangent)

Return type:

NDArray