Notes
1 Our proposed estimator will also use cross-fitting (Chernozhukov et al. Citation2018), but we suppress this notation here for conciseness.
2 Replication files for all experiments are available at github.com/swager/amlinear, in the folder debiased_single_index_experiments. For convex optimization, we use the R package CVXR (Fu et al. Citation2017).
3 In fact, as a consequence of the convergence of to
, which we will establish below, this criterion will imply that
.
4 The parts of condition (vi) involving are used only to control the first term in our error expansion, which we treated above using a variation on the argument used in Hirshberg and Wager (Citation2018).