Abstract
The problem of measuring conditional dependence between two random phenomenaarises when a third one (a confounder) has a potential influence on the amountof information between them. A typical issue in this challenging problem is theinversion of ill-conditioned autocorrelation matrices. This paper presents anovel measure of conditional dependence based on the use of incomplete unbiasedstatistics of degree two, which allows to re-interpret independence asuncorrelatedness on a finite-dimensional feature space. This formulationenables to prune data according to observations of the confounder itself, thusavoiding matrix inversions altogether. The proposed approach is articulated asan extension of the Hilbert-Schmidt independence criterion, which becomesexpressible through kernels that operate on 4-tuples of data.