W. Sprößig In this paper, we propose a new adaptive method for frequency-domain identification problem of discrete LTI systems. It is based on a dictionary that is consisting of normalized reproducing kernels. We prove that the singular values of the matrix generated by this dictionary converge to zero rapidly; this makes it quite efficient in representing the original systems with only a few elements. For different systems, it results in different selected sequences from the dictionary, that is, its adaptivity. Meanwhile, the stability of results is automatically guaranteed according to the structure of the dictionary. Two examples are presented to illustrate the idea.