Note that it is necessary to limit the number of data vectors , , for which the kernel matrix is calculated. Contrary to standard linear RLS, for which the correlation matrices have fixed sizes depending on the (fixed) dimension of the input vectors , the size of the kernel matrix in an online scenario depends on the number of observations .
In [6], a kernel RLS algorithm is designed that limits the matrix sizes by means of a sparsification procedure, which maps the samples to a (limited) dictionary. It allows both to reduce the order of the feature space (which prevents overfitting) and to keep the complexity of the algorithm bounded. In our approach these two measures are obtained by two different mechanisms. On one hand, the regularization against overfitting is done by penalizing the solutions, as in (9). On the other hand, the complexity of the algorithm is reduced by considering only the observations in a window with fixed length. The advantage of the latter approach is that it is able to track time variations without any extra computational burden.