[teambayleaf]

> Kelly criterion doesn't work. Your link says it maximizes log wealth. If potential wealth is unbounded, you will still take bets that are positive E(log utility).

Actually Kelly criterion works well, since it would limit your exposure to the game.

In this situation, Pascal as a Kelly-better would bet 12 deniers out of 10 livres (= 0.05% of his wealth), which is a penny.

[ikeboy]

How are you getting those numbers?

If you're trying to maximize log wealth, someone promising an insane amount of wealth for a trivial investment is a positive value for log wealth. The amount may need to be slightly more insane than it would be if you were just maximizing wealth, but with massive numbers such as used in Pascal's mugging this is easy enough.

[teambayleaf]

> If you're trying to maximize log wealth, someone promising an insane amount of wealth for a trivial investment is a positive value for log wealth.

Nah it does not work in such way.

Mathematically speaking, even if the mugger can propose an infinite large return in this situation, a kelly better would not bet more than 0.1% (= p = 1 / 1000) of his wealth.

It's very sound mathematics and nothing mystifying.

Edit: A Kerry better thinks in term of a long running sequence of bets. If you all of your wealth just because the game is favorable, you'll eventually lose all of your capital with probability one (in other words, betting everything minimizes the expected value of your wealth in long term). The natural conclusion here is that you need to bet a fraction of your wealth to maximize your long term wealth. But how much? Kelly criterion answers this queation formally. Read the Wikipedia article (or better, read Kelly 1956. It's a good paper) for how it handles the question.

[ikeboy]

The link says it's trying to maximize log wealth. If it's not actually doing that, then sure, it can limit loss. What exactly is being optimized?

[teambayleaf]

> The link says it's trying to maximize log wealth

Yes it is maximizing log wealth.

Think this way: If you give all of your fortune to the mugger, your capital will end up being 1) 0 livres(99.9%) or 2) 20000 livres (0.1%). So the expected log wealth is:

  log(0) * 0.999 + log(20000) * 0.001 = -inf

On the other hand, if you bet 10% of your wealth, you will end up having 1) 9 livres (99.9%) or 2) 2009 livres (0.1%)

  log(9) * 0.999 + log(2009) * 0.001 = 2.3

So you will prefer to bet 10% over 100%. The math does not bring you to "bet all of your fortune!" even if the odds is 1:100000000.

[ikeboy]

Yes, maximizing the log implies you shouldn't bet everything. But you should still bet 99.99%, if the odds are a googol to 1.

[teambayleaf]

> Yes, maximizing the log implies you shouldn't bet everything. But you should still bet 99.99%, if the odds are a googol to 1.

Not quite. If the odds are infinitely favorable, the log wealth is maximized when you bet 0.1% of your wealth. Any fractions other than that produce inferior results. This might be counter-intuitive but actually can be easily proven by basic calculus.

If you still doubt it, you can just compute it to be sure! For example, if odds is indeed 1:googol (1:10^100), the log wealth for betting 99.99% is -6.67, less than betting 0.1% which produces 2.52.