Optimal use of a black-box learner in semiparametric estimation

What the math gives to ML

The paper identifies a sharp way to prevent black-box nuisance errors from contaminating a target coefficient: its proposed estimator has error rate n^{-1/2}+delta_{a,mu} delta_{a,pi}+delta_s^2, eliminating the imbalance-sensitive term max(delta_{a,mu},delta_{a,pi}) delta_s. The transferable asset is adversarial conditional-moment calibration of inference weights produced by arbitrary learners. Neural networks can provide both nuisance estimates and adversarial test functions, with a constrained weight-editing layer applied after cross-fitting. The most direct experiment is partial-linear estimation with deliberately imbalanced nuisance difficulty, comparing DML against calibrated neural residualization for coefficient bias and confidence-interval coverage.