The probability of a probability

The Shores of the Dirac Sea has a somewhat head-scratching puzzle about probabilities:

Let us say that someone gives you a lopsided bet. Say that with probability \(\) one gets heads, and with probability \(\) one gets tails, and you have to pick heads or tails. You only know the outcome of the first event. Let's say after the first toss it came out heads. What is the probability that \(\)?

At first glance, the question makes no sense. The outcome of a bet is determined by the probability, so what is it with asking about the probability of the probability? Maybe rephrasing the question might make it a bit more clearer:

Given that the probability of a head is \(\), the probability of a head is \(\). Given that a single toss ended up a head, what was the probability that \(\)?

The first sentence might seem blatantly redundant, and you might be confused about what's different with the second sentence. If it doesn't make sense to you, thinking a lot harder is one way to figure out the difference, but another way is to learn about conditional probabilities. For the more mathematically inclined, expressing the problem in formulas makes the difference even more apparent, not to mention explaining the apparent redundancy:



Given this reformulation, all you have to do is to take advantage of the properties of conditional probabilities:


Embarrassingly, my first answer ended up being \(\):


I confused \(\) with \(\), which led me to the obvious-looking and yet seemingly wrong answer. (It's because \(\) is such an obvious guess that it seems so wrong, given my experience with conditional probabilities.) The denominator and numerator have to be computed separately, but I foolishly didn't do so. Since the probability of a head being exactly \(\) is \(\), in which case the probability of a head is \(\), the probability of \(\) and a head simultaneously is \(\). (Not being a pure mathematician, I'll live with the bit of fudging here.) And since \(\):


So the probability of the probability of a head coming up being larger than 0.5, given that a single toss turned out to be a head, ends up being 0.75. This probability is much more in line with what I would expect when working with conditional probabilities.

The next time you come across a brand new sort of coin that no one has ever seen before, and a single toss ends up being a head, you can confidently say that it's much more likely to be biased towards heads. Of course in real life, most of the coins you see will be like every other coin you've seen before, and all of them probably turned out to be fair, so all of this doesn't apply. On the other hand, there are other sorts of practical problems you can apply the same sort of reasoning to.

5 Replies to “The probability of a probability”

  1. You're absolutely right. On the other hand, if nothing at all is known about the probability distribution, we may as well assume it to be uniform. Otherwise, I smash into another "probability distribution of probability distribution" problem whose solution is not apparent to me, and I'm not sure there can be one since the answer obviously depends on the distribution.

    Fortunately for me, the puzzle at the Shores of the Dirac Sea says that we might as well assume a uniform distribution. :D

  2. There's yet another way of thinking about this problem (and about probability-of-probability kinds of questions in general) which makes it clearer:

    For the coin flip, substitute drawing with replacement of a white or black colored ball from a (well-shaken, idealized). The parameter r then reflects the number of white and black balls in the urn. Your question about "probability of probability" then becomes: how likely is it that the urn from which we are drawing has more white balls than black ones? The correct answer is the intuitive one: we can only sensibly answer the question when we know more about the urn. Otherwise we must make assumptions out of thin air, and the number we quote will reflect those assumptions more than anything else. We can then assume, appealing to principles such as symmetry, exchangeability and "least bias", that any urn, with any combination of black/white balls, is just as probable as any other, given our complete lack of information.

    (The neat thing about the 'urn' is that it is easy to imagine nested urns and multiple draws if we need to model probability-of-probability at any depth.)

  3. This question is totally stupid. There are problems with probability of a probability which makes sense but this problem makes no sense. Let me explain. You have a distribution with two discrete values; one is r and the other is 1-r. These values refer to an event which can be described by a random variable X. This whole together describes a random process if you have some samples to evaluate. Lets have x_i a sample of X and this sample is known. Does it give any idea about the distribution of an event? definitely it does not. x_i might even be coming from a Gaussian distribution. You need to ask it to the god ))

Comments are closed.