Comparison of various digital filter structures with respect to roundoff noise

For the error estimation by simulation a matrix description of time invariant linear discrete-time systems is used, that yields a simulation algorithm which is independent of the individual filter structure to be simulated. Various quantization and overflow characteristics for fixed and floating point arithmetic with variable wordlengths can be utilized with the aid of a modular simulation concept. The implementation of the simulation system is described. A couple of standard analyses like limit-cycle-analysis, computation of frequency and impulse response and roundoff-noise-analysis for deterministic and random input signals can be performed. An investigation of different filter structures realizing the same transfer function with respect to roundoff-noise behaviour for fixed and floating point arithmetic yields the result, that filter structures which are insensitive against coefficient quantization, exhibit a superior signal to noise ratio.