Paper 1, Section II, H

Coding and Cryptography
Part II, 2009

(i) State and prove Gibbs' inequality.

(ii) A casino offers me the following game: I choose strictly positive numbers a1,,ana_{1}, \ldots, a_{n} with j=1naj=1\sum_{j=1}^{n} a_{j}=1. I give the casino my entire fortune ff and roll an nn-sided die. With probability pjp_{j} the casino returns uj1ajfu_{j}^{-1} a_{j} f for j=1,2,,nj=1,2, \ldots, n. If I intend to play the game many times (staking my entire fortune each time) explain carefully why I should choose a1,,ana_{1}, \ldots, a_{n} to maximise j=1npjlog(uj1aj)\sum_{j=1}^{n} p_{j} \log \left(u_{j}^{-1} a_{j}\right).

[You should assume n2n \geqslant 2 and uj,pj>0u_{j}, p_{j}>0 for each j.j . ]

(iii) Determine the appropriate a1,,ana_{1}, \ldots, a_{n}. Let i=1nui=U\sum_{i=1}^{n} u_{i}=U. Show that, if U<1U<1, then, in the long run with high probability, my fortune increases. Show that, if U>1U>1, the casino can choose u1,,unu_{1}, \ldots, u_{n} in such a way that, in the long run with high probability, my fortune decreases. Is it true that, if U>1U>1, any choice of u1,,unu_{1}, \ldots, u_{n} will ensure that, in the long run with high probability, my fortune decreases? Why?