Intersymbol interference equalization by universal likelihood
Kelly, Owen Ernest
Johnson, Don H.
Doctor of Philosophy
Codelength based inference is used to decode binary symbols distorted by an inter-symbol interference (ISI) channel with additive noise. The transmitted signals are antipodal waveforms constructed from $\pm$1 valued signature sequences. The receiver adapts to the unknown channel by training on a known preamble sequence. The training sequence is quantized to a small alphabet and the resulting discrete valued sequence is used to train a context-tree model of the combined data-source and ISI channel. In additive white Gaussian noise, the context tree equalizer has a probability of error slightly larger than an adaptive decision feedback equalizer (DFE). The difference corresponds to approximately 1 dB loss of signal-to-noise ratio. However, a context tree equalizer performs equally well in the presence of heavy-tailed noise for which the adaptive DFE essentially fails. An empirical classification result of Gutman is extended to Markov processes of alphabet size two and unknown Markov order.
Statistics; Electronics; Electrical engineering