Robust sieve estimators for functional canonical correlation analysis

Abstract

In this paper, we propose robust estimators for the first canonical correlation and directions of random elements on Hilbert separable spaces by combining sieves and robust association measures, leading to Fisher-consistent estimators for appropriate choices of the association measure. Under regularity conditions, the resulting estimators are consistent. The robust procedure allows us to construct detection rules to identify possible influential observations. The finite sample performance is illustrated through a simulation study in which contaminated data is included. The benefits of considering robust estimators are also illustrated on a real data set where the detection methods reveal the presence of influential observations for the first canonical directions that would be missed otherwise.