Optimization Of Distributed Quantizers Using An Alternating Information Bottleneck Approach

This paper considers a scenario with distributed nodes receiving statistically dependent signals. The nodes have to quantize these signals and forward them over capacity-limited links to a common receiver. As a joint vector quantization is assumed to be infeasible, individual scalar quantization is performed. However, these scalar quantizers have to be designed jointly. The optimization is based on the alternating Information Bottleneck method. The quantizers are optimized successively, keeping all other quantizers fixed and using their statistics as side-information. Following this approach, we fulfill given rate constraints and preserve more relevant information than using independently optimized scalar quantizers.