Enum InputLocationDerivative 

pub enum InputLocationDerivative<'a> {
    Radial {
        centers: ArrayView2<'a, f64>,
        radial_kind: &'a RadialScalarKind,
    },
    Jet(ArrayView3<'a, f64>),
}

Carrier for the ∂Φ/∂t chain-rule input, dispatched on basis kind by LatentCoordValues::design_gradient_wrt_t_dispatch.

• InputLocationDerivative::Radial is the radial-kernel path: the caller supplies the radial kernel family together with the center coordinates, and the chain rule ∂Φ/∂t = q(r) · (t − c) is applied internally. This covers every isotropic radial basis — Duchon (any nullspace order), Matérn (every supported half-integer ν), and anything else whose pointwise gradient is radial. Helpers: crate::basis::duchon_radial_first_derivative_nd, crate::basis::matern_radial_first_derivative_nd.
• InputLocationDerivative::Jet is the pre-computed jet path: the caller has already assembled a closed-form (N, K, d) tensor for a basis whose chain rule is not a simple radial scalar times a unit vector. Sphere kernels carry the tangent-direction times K'(cos γ); periodic-cyclic B-splines carry the closed-form cardinal derivative; tensor-product B-splines carry the product-rule mix. Helpers: crate::basis::sphere_first_derivative_nd, crate::basis::periodic_bspline_first_derivative_nd, crate::basis::bspline_tensor_first_derivative.

The dispatch is an enum rather than a trait because each path’s arguments differ structurally (radial bases reuse scalar radial kernels shared with the kernel-shape chain machinery; jet bases ship the full tensor). All chain rules are analytic and closed-form; no autodiff, no finite differences.

Variants

Radial

Radial-kernel chain rule. The chain rule (t − c)/r is reconstructed internally from the finite q = φ'(r)/r scalar and the center coordinates.

Fields

centers: ArrayView2<'a, f64>

radial_kind: &'a RadialScalarKind

Jet(ArrayView3<'a, f64>)

Pre-computed analytic (n_obs, n_centers, latent_dim) jet.