Sparse Vector Linear Prediction with Optimal Structures

A modification of a classical Vector Linear Prediction (VLP) technique is proposed in this paper, enabling significant reduction in complexity. The proposed sparse VLP technique (sVLP) is based on predictors with reduced number of nonzero elements. For a given input vector process, a design procedure for obtaining optimal sparse predictor structures and matrix elements is described. The effectiveness of the sVLP is evaluated on interframe predictive coding of Line Spectrum Frequencies (LSF) and compared to the classical VLP based on Switched-Adaptive Interframe Prediction scheme. The loss of the prediction gain due to sparse structures is calculated using various design parameters. Simulation results prove that a 6-fold reduction in complexity of prediction can be achieved causing only insignificant loss of the prediction gain and coder performance.