Abstract. Deterministic sampling can be used for nonlinear propagation of the statistics of signal processing models. Unlike Monte Carlo methods, random generators are not utilized in any stage. The samples are instead calculated deterministically. Our novel approach generalizes the deterministic sampling technique for propagating covariance in the unscented Kalman filter by introducing generic excitation matrices describing small discrete canonical ensembles. The approximation lies in how well the available statistical information is encoded in the discrete ensemble, not how each sample is propagated. The application and performance of deterministic sampling are illustrated for a typical step response analysis of an electrical device modeled with an uncertain digital filter. [20,12,11], which is widely utilized in the signal processing community. There are very few other examples. In the most general context presented here, DS utilizes small and thus highly effective ensembles of samples of models devised for particular purposes. This work targets parameterized models h(q, t) = g(q, t) x(t), which are identified from calibration measurements and thus have dependent parameters q. Such a model is usually nonlinear in parameters and describes the linear response of a system with impulse response g(q, t) to an excitation x(t). The convolution ( ) will here be evaluated with classical linear digital filters [6].Once the samples are found, the procedure of DS is identical to that of RS-the model is evaluated for all samples of the ensemble, followed by calculation of the desired statistics. For instance, assume h(q) depends on one parameter q with mean q and variance δ 2 q , where · denotes statistical expectation. The mean h and the variance δ 2 h of the model can