Regularized Linear Function Spaces (\(\ell_2-\ell_2\))
These estimators use linear learners and critics constrained to Euclidean balls. Their optimistic projected-gradient updates are the \(\ell_2\) analogue of the entropy updates in Sparse Linear Function Spaces (\ell_1-\ell_1). The critic objectives are linear: the current implementations do not subtract a quadratic critic covariance term.
For each class, returned coefficients and critic weights average feasible
iterates. The fitted duality_gap_ evaluates the exact constrained learner
best response. Consequently, it handles both active ball boundaries and
singular empirical covariance matrices. Moment violations are reported in the
same \(\ell_2\) geometry as the critic ball.
One-stage estimators
Let
Estimator 1 - coefficient-L2 penalty
The coefficient-regularized class solves
The critic best response is the unit direction of \(m(\alpha)\), with the
zero vector used when the moment is zero. The learner update is projected onto
the radius-B Euclidean ball.
|
Linear NPIV with coefficient-L2 regularization. |
Estimator 2 - empirical-L2 penalty
The fitted-value ridge class replaces the coefficient norm by
Its learner best response is a convex trust-region problem. The gap calculation uses the unconstrained solution when feasible and otherwise solves for the boundary multiplier; it does not approximate the response using a single eigenvalue bound.
Linear NPIV with empirical-L2 learner regularization. |
Nested estimators
Let \(\mathbb E_p\) and \(\mathbb E_q\) be the normalized empirical means used for the two bridge moments, and define
W is the observation-level multiplier on the first learner in the second
bridge moment; it is not a sample weight. W=None uses one. If
subsetted=False, both stage means use all observations. If
subsetted=True, subset_ind1 is a required nonempty binary mask and an
omitted subset_ind2 is its complement. Two explicit nonempty masks may
overlap or leave observations unused. The ridge matrices \(Q_A,Q_B\)
always use the full sample.
Estimator 3 - coefficient-L2 penalties
The simultaneous coefficient-regularized game is
Both learner and critic updates are optimistic gradient steps followed by Euclidean projection onto their respective balls. The reported gap solves the two isotropic learner best responses exactly.
|
Nested linear NPIV with coefficient-L2 penalties. |
Estimator 4 - empirical-L2 penalties
The nested fitted-value ridge class solves
The two gap components are exact Euclidean trust-region minima. An eigendecomposition of each symmetric learner covariance supports both full-rank and singular cases without changing the estimator’s moment game.
|
Nested linear NPIV with empirical-L2 learner penalties. |