A study of bounds on limit cycles in digital filters (Q1908845)

Due to quantization errors, the state of a realization of a stable linear multivariable digital filter may tend to a limit cycle described by \[ x(n)= Ax(n- 1)+ u(n),\tag{1} \] where \(n\in \mathbb{N}\), \(x(0)= x_0\), \(A\) is stable, \(u(n)\) is T-periodic and is related to the equantization errors (and possibly to constant inputs), \(x(n)\) is T-periodic. After using the periodicity condition in (1), bounds for \(\max|x(n)|_2\) using the induced 2-norm or for \(|x(n)|\) (for all \(n\)) replacing the elements of the matrices or vectors involved by their absolute value, are given. Taking the discrete Fourier series representation of both sides in (1), allows to find the bound \[ |x|_{rms}\leq \overset l= {T}{\underset l= {1}\max} \Biggl|\Biggl(Id- \exp\Biggl(- i {2\pi l\over T}\Biggr) A\Biggr)^{- 1} \Biggr|_{i2} |u|_{rms}, \] where \(|x|^2_{rms}= {1\over T} \sum^{j= T}_{j= 1} x^t(j) x(j)\). A numerical comparison (when looking at the components of the vectors) in the 2-dimensional case shows that the last bound is the best and it improves previous results.