An inhomogeneous uncertainty principle for digital low-pass filters (Q2502054)

By a filter the authors mean a bounded, \(2\pi\) periodic, complex-valued function \(h\). A filter \(h\) is a (half-band) low pass filter if \(\lim_{\xi\to \pi} h(\xi)=0\) while \(\lim_{\xi\to 0} h(\xi)=1\). A filter is said to be an implementable filter if it is absolutely continuous with derivative in \(L^2[0, 2\pi)\). A filter is interpolatory if \(h(\xi)+h(\xi+\pi)=1\). The ideal (half-band) low pass filter \(I\) is defined to be the \(2\pi \)-periodic extension of the characteristic function of \([-\pi/2,\pi/2)\). The ideal filter is not implementable and, in practice, one often substitutes a nonideal approximant. Here an attempt is made to quantify the nonideality in terms of the cost functional \(U(h)=\| h-I\| \| h'\| \) where \(\| \cdot\| \) denotes the \(L^2\)-norm on \([0,2\pi)\). The main result here is that \(U(h)>1/2 +| h(-\pi/2)-1/2| ^2+| h(\pi/2)-1/2| ^2\), that this bound is sharp but never achieved over the class of implementable filters, but that one can come arbitrarily close to \(U(h)=1/2\) in this class. Precise estimates are given for specific classes of filters including Butterworth filters, for which the minimum cost is approximately \(0.5075\) and the \textit{I. Daubechies} filters [Ten lectures on wavelets. CBMS-NSF Regional Conference Series in Applied Mathematics. 61. PA: SIAM, Society for Industrial and Applied Mathematics (1992; Zbl 0776.42018)] whose uncertainty products are lower bounded by \(0.5135\). The practical matters of construction of cost minimizers among filters of a given length, and of asymptotic minimizers that are near optimal for a given length, are also addressed.