What I'd like to know is why some letters have higher probability than others? Is that because some letters are more likely on the elliptic curve vs not? If you plug sequences into vanitygen it will give different difficulties depending on letter, eg.
11... > 1A...
1Q... > 1D...
1R... > 1Q... and so on.
I made a table of the first three chars after 1 to help me in a project with fast lookup of difficulty calculation. There are broad classes of probabilities depending on what letters are involved. I guess there is some non 1-1 mapping from 2^160 possible addresses and base 58 representations.
In ECDSA, you start with a large integer as your private key then multiply ("multiplication" here must be understood as a group operator not as arithmetic multiplication) the base point G of your elliptic curve to get the corresponding public key.
Therefore your public key is a point (x,y) on the elliptic curve.
In bitcoin, your address is some hash of you public key.
As a result, if you start from a random string, the likelihood of a collision between said string and a "valid" address depends on how many integers n satisfy the condition "n*G is a point on the curve AND this point hashes to a matching string".
Therefore a brute force algorithm (like vanitygen) might display varying difficulty as a function of the chosen string.