Fair gambling games should be… fair, right?!
Félix Locas, here at UQAM, introduced me to “fair but unfavorable” gambling games of the following type:
- you pay to draw a random positive gain such that ;
- you can play for as long as you want.
Precisely, we have a sequence of independent and identically distributed random variables representing your gain (minus ) at the th game. Thus and .
The surprising fact is that there exists such “fair” gambling games for which the total gain tends to in probability. More precisely (and even worse), there are games for which
for all small . (See W. Feller, Note on the law of large number and “fair” games, 1945.)
The game, as I just described it, can still be favorable to you. In fact, playing it for long enough, you can be certain to make as much money as you want. You just have to know when to stop: for all , there almost surely exists an indice such that .
But here’s the catch: You can’t (generally) play it for long enough. You could be ruined before you make up for previous losses.
Let’s therefore adjust the model to make it more realistic. The player has only a finite amount of money. She can play for as long as she is not broke, and can stop at any time.
Let be a winning goal for the player. Consider the ruin time and the stopping time . If the player keeps playing until she is broke or until , then her expected gains are
(Note that both and are almost surely defined unless is constant.)
Suppose that and are of finite variance. Then for all winning goal , there exists a starting amount such that the expected gain is positive.
Sketch of proof.
Note that and that
is of order . Letting , we find that .
Are there games such that for most amounts ?
Put another way: are there games for which no strategy yield a positive expected gain?
Problems with unbiaised consistent estimators
In close relationship with the total gains is the Monte Carlo estimator . Here, is an unbiaised and consistent estimate of .
Although the strange behavior of Feller’s game (that quite fast in probability) is not as striking in this case – consistency obliges – problems may still arise in pratical situations.
If is of infinite variance, then for all
Conversly, if is of finite variance, then for all and small
Proof. See the next post.
In practice, the convergence of the estimator is often assessed using an empirical criterion. The estimate is taken to be sufficiently good when the criterion is first satistied. For instance, we may stop at the first indice such that the estimated variance of is less than some fixed (this is sometimes eyeballed). The expected estimate of is then
For all , there exists a game such that
How bad can things get when the variance of is finite? What is the probability that a significant error is made?