Robust properties of padé approximants obtained by network learning

The paper deals with a problem to construct a Padé approximant in a noisy environment. To some function identification problems with single input variable, the standard least mean square (LMS) method and a method using an arithmetic network with learning capability are applied. The Padé formulas obtained by the two methods are compared in terms of the pole-zero distribution of resultant Padé formulas and robustness, and it is shown from some numerical results that a Padé approximant by the learning network has poles and zeros in a considerably broad region of the complex plane, and as a consequence, has much more robustness in the domain outside of the training data than that by the LMS.