The PASTd (Projection Approximation Subspace Tracking with Deflation) algorithm was designed to estimate efficiently the signal subspace of a possibly non-stationary vector process [4]. In this paper we analyse the numerical stability of this algorithm. After a brief summary of the PASTd algorithm we investigate asymptotically the error propagation from one recursion step to the following. It will be shown that the error propagation in the PASTd algorithm is stable. To underline our investigations, we show results of a fixed point implementation of the algorithm for image compression based on the Karhunen-Loeve (KL) transform.