To code, or not to code, at the racetrack: Kelly betting and single-letter codes
share
For a gambler with side information, Kelly betting gives the optimal log growth rate of the gambler's fortune, which is closely related to the mutual information between the correct winner and the noisy side information. We show conditions under which optimal Kelly betting can be implemented using single-letter codes. We show that single-letter coding is optimal for a wide variety of systems; for example, all systems with diagonal reward matrices admit optimal single-letter codes. We also show that important classes of systems do not admit optimal single-letter codes for Kelly betting, such as when the side information is passed through a Z channel. Our results are important to situations where the computational complexity of the gambler is constrained, and may lead to new insights into the fitness value of information for biological systems.
READ FULL TEXT