betaFeedback

WorksMathematics

Nonparametric Instrumental Kernel Regression: The Well- and Misspecified Setting Under General Link Conditions

Abstract

We study nonparametric instrumental variable regression when the outcome takes values in a separable Hilbert space Y\mathbb{Y}, for instance a curve, a density or the kernel mean embedding of a distribution. The structural function solves an ill-posed operator equation whose conditional expectation operator need not be injective, and whose solutions need not belong to the hypothesis space. We analyse a two-stage kernel estimator that extends kernel instrumental variable regression to Y\mathbb{Y}-valued outcomes and allows spectral regularisation in both stages, including Landweber iteration and Showalter's method in the second stage. We show that the second stage is a regularisation method in the Hilbert scale generated by the inverse of the kernel integral operator. This viewpoint identifies the solution towards which the estimator converges: the minimum-norm solution in the reproducing kernel Hilbert space in the well-specified case and, in the misspecified case, the unique solution in the range of the continuous extension of the projection operator associated to the well-specified setting. Under source conditions and link conditions expressed through general index functions, we derive non-asymptotic bounds in a scale of interpolation norms that hold with high probability in terms of the first-stage error, with or without identification and in the well- and misspecified settings. For polynomial link conditions, we recover the rate of scalar kernel instrumental regression and extend it to targets outside the hypothesis space; for exponential link conditions, we obtain logarithmic rates in the severely ill-posed regime.

Paper details

vlnc:W_akstsn4s
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
arXiv nonexclusive license
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…