Corrected all four adaptive function-to-scalar DoF branches (with/without smoothing and unpenalized covariates) to use n * lambda1 with the unnormalized Gram matrix. Adaptive DoF and BIC values can change.
Included the coordinate-weight factors in the adaptive penalty-gradient derivative in functional and linear grouped DoF routines. This matters for nonunit var_weights, including the default overlapping spline weights.
Aligned function-to-scalar documentation with RSS/(2n) and (lambda2/2) ||D psi||^2, and corrected the displayed linear-loss scale. Solver updates and fitted paths at fixed parameters are unchanged.
Added functional adaptive DoF regression tests against central finite differences and an independent KKT derivative, including nonunit weights, smoothing and unpenalized covariates.
Fixed missing response forwarding in the adaptive Lasso degrees-of-freedom calculation and the internal grouped-penalty dispatcher.
Used the fixed-weight degrees-of-freedom branch for Group Lasso.
Passed unpenalized covariates to the Lasso degrees-of-freedom calculation and corrected the augmented penalty-matrix dimensions for Group Lasso.
Matched the adaptive correction to the solver’s normalized least-squares loss, including the sample-size factor, and used the partial least-squares pilot when unpenalized covariates are present.
Corrected the ordering of coefficient-surface slices returned in
f2fSP()$sp.fun.path.
Added regression tests using finite-difference divergence checks and coefficient-to-surface reconstruction checks.
Updated the documentation for f2sSP() and
f2sSP_cv() to describe the sparsity (lambda1)
and smoothness (lambda2) regularization parameters more
clearly.
Standardized the names of the sparsity-parameter arguments and
returned components to use the lambda1 terminology
consistently.
Revised the examples and the documentation of the returned model-selection quantities for fits with and without a smoothness penalty.