quote:
Really? So proving that earth is not a flat disk mounted on the backs of turtles makes it more likely that the earth is actually being pushed around its orbit by butterflies?
Yes, it does. Of course, increasing a probability that is already vanishingly small by an infinitesimally small amount isn't going to matter very much in practical terms, but why do you think it has no effect?
quote:
I don't believe you've thought this principle through.
In the original formulation, let the prior probabilities of X and Y be p(X) and p(Y), and the probability of all other possibilities be p(Z). Given these definitions, p(X) + p(Y) + p(Z) = 1. After encountering the new data (call it D), p(X|D) = 0 and by Bayes' theorem
p(Y|D) = p(D|Y)p(Y) / (p(D|Y)p(Y) + p(D|Z)p(Z)).
I stipulated that the new data does not distinguish between Y and Z, which I take to mean that Y and Z and equally likely to yield the observed data if they are true, i.e. p(D|Y) = p(D|Z). Given that, the above expression for p(Y|D) reduces to p(Y) / (p(Y) + p(Z)). Our relative certainty about Y (that is, our estimate of the probability that Y is true after seeing the new data, relative to our starting estimate) is given by
p(Y|D) / p(Y) = 1 / (p(Y) + p(Z)) = 1 / (1 - p(X)),
which was what I claimed.