Filters

Lecture 6: Digital Signal What?

21/10/2004
Summary:
  • Spirals become spirals
  • The eigenvalue is the Fourier Transform of the filter kernel at that frequency
  • Complex-rate spirals are also eigenvectors (this is one way of looking at the z and Laplace transforms)
  • Deconvolution of a finite causal filter (FIR/moving average(/all-zero)) is a recursive/feedback/autoregressive(/all-pole) signal network
  • Inverse of linear system is linear
  • Inverse of shift invariant system is shift invariant
  • Inverse of LSI system is LSI
    • An autoregressive filter is LSI, having an infinite filter kernel (IIR)
  • If you have some filters, you can do two things -
    • Series (cascade)
    • Parallel (addition of output = addition of impulse response)
  • FIR filters in cascade - impulse responses convolve
  • FIR filters in parallel - impulse responses add up
  • Autoregressive filter in cascade - impulse response of inverse system convolve
  • Autoregressive filters in parallel - give a somewhat complicated autoregressive filter
  • The most general filter implementable on a computer is a cascade of FIR and autoregressive filter
    • Called autoregressive-moving average (ARMA), (or pole-zero, or rational) filters
    • Implemented as direct form I (cascaded MA-AR) or direct form II (cascaded AR-MA) signal network on a computer
  • Polynomial multiplication is convolution of coefficients
  • Autoregressive inverse filter of an FIR filter can be thought of as a reciprocal of a polynomial function
    • The filter kernel coefficients are the Taylor series expansion of this "rational function". Turns out we never actually have to evaluate the Taylor series, because the autoregressive system can directly be implemented as a feedback network
  • All the complicated signal-network manipulations are equivalent to simple algebraic manipulations of rational functions (another way of looking at the z and Laplace transforms)