Asymptotic behavior of block floating-point digital filters (Q1306179)

In the theory of digital filters topics of interest are (among others): stability and the existence of limit cycle oscillations. In the paper it is assumed that the filters under investigation are represented in the standard matrix form, i.e. \(x(n+1)= Ax(n)\), \(A\) is an \(n\times n\) matrix. It is known that for a system with \(A\) being a Schur matrix (i.e., one with all eigenvalues located inside the unit circle), the stability is insured. The stability means that for each initial condition the limiting point is always the zero state. However, if the filter is implemented in Block-Floating Point BFP architecture, where quantization and normalization occur after each iteration, the formulated fact may be not true. Three cases are possible: R1 the response becomes unbounded \((l_\infty =\infty)\); R2 the response is bounded, and R3 the response converges to zero. Some theorems insuring the fulfilment of the R2 and R3 cases are proved. These theorems provide sufficient conditions which in practice are easily satisfied. The overall conclusion is that a filter in the BFP architecture is free of limit cycles if a sufficient state mantissa wordlength is available. Some numerical examples illustrate the presented theory. The paper possesses a high quality by the clear way to explain and present the ideas underlying the theory.