regularization#
Backend regularization terms for the fitting objective min_x ½‖ω⊙(S(x)−y)‖² + ρ(x).
A Regularizer is an additive objective term ρ(X) folded into the local Gauss-Newton model
(t3toolbox.backend.optimizers.LocalModel) and Problem.objective, so it composes with
every optimizer, sampling kind, geometry, and representation with no changes to any of them. This
layer is check-free – a raw-.data user constructs and attaches a regularizer directly, exactly as
a frontend user does (design record + the razor check: dev/regularization_design.md §5a).
The interface is geometry-agnostic: each method receives a GeometryOps and leans only on its
primitives (point_norm_sq / point_tangent / project / inner), so the SAME regularizer
works on manifold or corewise. Extending to a new regularizer (e.g. the Grasedyck-Kramer inverse-unfolding
-singular-value weighting) is a new subclass with the same four methods – Problem / LocalModel /
the optimizers are untouched.
Classes#
Protocol for a quadratic regularizer |
|
Identity (Tikhonov) regularization |