Consider a proof consisting of n steps. Suppose that it is possible to check the correctness of each of these steps, but that n is so large that it is impractical to check all of them. Instead, start checking the steps at random and stop if you find an incorrect step (in which case the theorem is not proved) or if you reach m correct steps. What can you conclude in the latter eventuality?
Let C denote the event that the proof is corect and let A denote the event that you have found m randomly selected steps correct. By Bayes’s Theorem:
where
is the conditional
(or posterior) probability of C given A; is the conditional
probability of A given C; is the unconditional
(or prior) probability of C; and is the conditional
probability of A given the complement of C. As , the posterior probability simplifies to:
A low Bayesian trick is to eliminate the
prior probability by considering the so-called Bayes factor, defined as the
ratio of the prior to posterior odds in favor of C, which in this case reduces
to . Let the random
variable K be the number of incorrect steps in the proof given that the proof
is incorrect. Then:
Consider two simple cases. In Case 1:
That
is, conditional on the proof being incorrect, there is only a single
error. In this case, so that, for example,
to achieve a Bayes factor of 0.05 (i.e., posterior odds 20 times prior odds)
for a proof of n = 100 steps, you
have to sample 95 of the steps. Not
much help. In Case 2:
That
is, all possible numbers of incorrect steps are equally likely. My guess is that there is a simple
expression for , but I came up against a Limit to Laziness in trying to find
it, so I resorted to computation. Here
is the result for n = 100:
m
10
0.082
20
0.038
30
0.023
40
0.015
50
0.010
60
0.007
70
0.004
80
0.002
90
0.001
In
this case, to achieve a Bayes factor of 0.05, you only have to check that
around 15 randomly chosen steps are correct.
Result: It is harder to check a correct theorem of a
good mathematician than of an error-prone mathematician.