MLSE offers adaptive non-linear receiver-sided equalization for channels with memory. New areas of application considered are directly modulated DFB laser, D(B/Q)PSK, and duo-binary transmission systems. Reduced computational complexity is a new research area.I. MLSE BACKGROUND An overview of FFE, DFE and MLSE [1] (maximum likelihood sequence estimator) receiver-sided equalization technologies will be presented. The use of MLSE to compensate for deterministic non-linear distortions in channels exhibiting memory [2], [3], and the importance of the Viterbi algorithm (VA) [4] will be described. The branch metrics used by the VA can be implemented using non-parametric histogram approximations to the actual PDFs, or parametric analytic closed form expressions [9]. The chromatic dispersion (CD) compensation benefits of MLSE have been experimentally demonstrated with two commercial realizations of an MLSE processor [5], [6]. A reach of 150 km has been shown for NRZ transmission using a 4-state MLSE at the receiver [5]. Training sequences can be used to establish the branch metrics in the MLSE, but in practice this adds complexity to the system and reduces the transmission capacity. An MLSE processor with automated acquisition has been developed [5], and a theoretical investigation into the design of a MLSE with blind acquisition capability has been reported [7]. The dynamic polarization mode dispersion (PMD) compensation capability has also been demonstrated by experimental results, which showed that an MLSE update rate of 2 kHz is suitable to compensate for a polarization rotation rate of up to 0.25 rad/ms with first order PMD [8].
II. NEW TRENDS IN MLSE DESIGN
A. Applications requiring an increased number of statesThe ability of a MLSE to compensate for CD is related to the number of states that are used. The increasing capability of electronic processors will therefore lead to the development of MLSE processors with an increasing compensation capability. However this places increasing demands on the clock recovery, which is expected to increase in sophistication as the number of states in MLSE processors increases in the future.
B. Signal statisticsThe best performance is obtained from the MLSE when the branch metrics are closely matched to the real statistical noise distribution of the signal samples. The histogrambased approach to constructing the branch metrics, used in the Viterbi trellis, requires a significant time to accumulate the data, particularly for the less probable events in the tails of the histograms. The inherent quantization also introduces performance limitations. However the branch metrics do then approach the complex measured non-central statistical noise distributions.The time delay involved in implementing this approach has driven the effort to simplify the branch metrics using analytic functions [9], [10]. A range of analytic approximations has been examined from Gaussian, to simple and non-central chi-squared functions, to the more complex Karhunen-Loeve expansion. Unfortunately these functio...