Fast subspace tracking by a novel information criterion

A new approach to tracking the principal subspace of a vector sequence is developed. This approach is based on a novel non-quadratic cost function referred to as novel information criterion (NIC). The NIC algorithm is guaranteed to converge globally to an arbitrary set of orthonormal base vectors of the desired principal subspace. It is also a fast algorithm as it has a computational complexity of O(M(2)r) flops per update in a batch mode implementation or O(Mr)flops per update in a recursive least-squares implementation, where M is the vector dimension and r the predetermined dimension of the principal subspace. The NIC algorithm is more general and more robust than the PAST algorithm by Yang.