Fractional graph coloring for functional compression with side information

We describe a rational approach to reduce the computational and communication complexities of lossless pointto-point compression for computation with side information. The traditional method relies on building a characteristic graph with vertices representing the source symbols and with edges that assign a source symbol to a collection of independent sets to be distinguished for the exact recovery of the function. Our approach uses fractional coloring for a b-fold coloring of characteristic graphs to provide a linear programming relaxation to the traditional coloring method and achieves coding at a finegrained granularity. We derive the fundamental lower bound for compression, given by the fractional characteristic graph entropy, through generalizing the notion of Körner’s graph entropy. We demonstrate the coding gains of fractional coloring over traditional coloring via a computation example. We conjecture that the integrality gap between fractional coloring and traditional coloring approaches the smallest b that attains the fractional chromatic number to losslessly represent the independent sets for a given characteristic graph, up to a linear scaling which is a function of the fractional chromatic number.