Two arguments against modeling probabilities by size of propositions

To my surprise, there are quite a few people here at ANU who believe that probabilities of various kinds can be modeled in terms of relative size of propositions: something has probability 1 if it is true in all (or 100%) of the relevant worlds, probability 0 if it is true in none (or 0%), and probability 0.5 if it is true in half of the worlds (or 50%). I also find it surprisingly hard to explain why I think that's wrong. Here are two arguments I've come up with so far (apart from obvious worries about making sense of these fractions in infinite and proper-class cases).

First, an argument from relativity. Take objective chance, the kind of thing quantum physics talks about. In the actual world, the chance of a radium atom decaying within 1600 years is roughly 0.5. But this is a contingent fact. In other worlds the chance is 0.9, or 0.1. On the probabilities-by-sizes account, the relative size of the proposition will therefore differ depending on which world is actual. But why would it? In fact, it seems not unreasonable to assume that for any world w that is possible relative to the actual world, there is a corresponding world w' that is possible relative to worlds where the chances are different such that w and w' agree on all relevant matters of particular fact, e.g. on when individual radium atoms decay. After all, our laws do not rule out that all radium atoms decay within a millisecond, or only after a billion years; they only declare this extremely improbable. So any possible decay pattern that is nomologically possible relative to this world is also possible relative to the other worlds. But then it can't be the case that within the worlds accessible from here, 50% of all radium atoms decay within 1600 years whereas within the worlds accessible from elsewhere, only 10% do so.

Likewise for credence. It often happens that you and I assign different credence to some proposition, say, to whether the latest US elections were rigged. According to the probabilities-by-sizes account, this means that such worlds are more common in your belief space than in mine (or vice versa). But again, couldn't we assign different credence to this proposition even though there is no proposition that I rule out and you do not (or vice versa)? If there is no such proposition, there is presumably a 1-1 map between the worlds in your space and the worlds in mine which maps rigged worlds to rigged worlds and non-rigged worlds to non-rigged worlds. And then it's hard to see how the fraction of rigged worlds could be higher in your space than in mine (or vice versa).

Second, an argument from probability distributions. Let's again begin with objective single-case chance. Physics tells us how the probabilities are distributed, and this distribution is almost never uniform: the chance that a radium atom decays within between 1000 and 2000 years is much higher than that it decays between 2000 and 3000 years, and that again is much higher than a decay between 3000 and 4000 years, and so on, without end. So there are lots of possibilities that have an extremely low chance of coming true, whereas there are comparatively few possibilities that have a comparatively high chance of coming true. So chance doesn't correspond to proportion among possible cases. This is particularly obvious if we assume that the total number of nomologically possible worlds is finite. Suppose for example that there are exactly 1 billion times at which an atom can decay. Then physics tells us that a handful of these have a high chance of truth, whereas the overwhelming majority is very improbable.

Again, this carries over to credence. Our credences are also probability distributions, which often assign higher probability to more specific propositions. To take a dummy example, suppose I give credence 0.5 to some very specific proposition p, true at only one world w. In case p is false, I am pretty much undecided between various options: ~p & q & r, ~p & ~q & r, ~p & q & ~r, ~p & ~q & ~r; all get about the same credence of 0.125. But now there must be at least 5 worlds in my belief space, and p isn't true at half of them. But still I give it credence 0.5.

I suppose one could get out of these problems by using indistinguishable worlds and adjusting their numbers in such a way that the probabilities-by-sizes account becomes true: whenever p has probability 0.5, it is true in half the relevant worlds. If only one world comes into question, simply add in sufficiently many perfect duplicates of it. The 1-1 maps considered above then turn out to be 1-many or many-1, where the manys are indistinguishable duplicates. But that's cheating, and we learn nothing at all about probability from the fact that it can be interpreted in this way.

Comments

# on 22 December 2006, 00:04

That's good - I've discussed these things with some of the people at ANU and never been totally convinced by their point. I can see why it might be appealing to say that probability corresponds to size of the set of possibilities (whether possible worlds or whatever you've got in the subjective case), but I think your point about uniform distributions is right. There's no reason for things to have uniform distributions (as we see from Bertrand's paradox) and in fact, we know that there can be no uniform distribution either over the natural numbers, or even over the set of all real numbers (though there can be a uniform distribution over any bounded interval of real numbers). The only way to make probabilities always work out as uniform distributions seems to be just to stipulate that this is how you'll measure size. Or else hope that the size works out in such a way that it's possible to have a uniform distribution.

However, maybe there is a way to make sense of some of the symmetry arguments that Alan Hájek likes to make, in a less direct way than what you're considering here.

# on 22 December 2006, 02:19

Here's one of the points I would make in response to the argument from relativity of objective chances if I believed in the probability-as-sizes acount:

"Suppose that E is the result of a genuinely indeterministic process to which the actual laws of world w assign Prob(E, h, t) = 50% given the history h of w up through time t. Then according to my account, 50% of the worlds that share the laws and history of w up through t are worlds in which E occur, and 50% are worlds in which E fails to occur.

"Now take any two such worlds w* and w**.

"Assumption: History and laws up through a time t completely fix the chances of outcomes at t.

"Corollary: Prob(E, h, t) = 50 % at w* and w** as well.

"So according to the probability-of-sizes account: 50% of the worlds that share the laws and history of both w* and w** up through t are worlds in which E occur, and 50% are worlds in which E fails to occur.

"Generalizing: There is no relativity of objective chances between w* and w**, if they agree in laws and history up through t.

"Thus: Either you are denying the above (very plausible) assumption, or something is wrong with your argument.

"Diagnosis: It is right to say that the laws and history up through a time t do not determine the frequency of a chancy outcome E in accessible worlds at times //later than t// in those worlds. But these aren't the frequencies that probability-as-sizes account is interested in. For us, what determines the chances of E is the frequency of outcome E across accessible worlds //at precisely that time t// in those worlds. The frequency of E at times //later than t// in accessible worlds is not determined by their laws and history, but so what? Those aren't the frequencies that fix the chances of E at t in accessible worlds."

(Apologies in advance if this doesn't make sense...)

# on 22 December 2006, 04:50

hey Alex, I've tried to get around this response with the mappings idea. I don't assume that there is a relativity of objective chances among worlds that agree in laws and history. Here is the argument in somewhat more detail:

Consider a world w where the chance of radium atom A decaying within 10 years is 0.5, whereas at @ it is 0.1. Nevermind that w has different laws. Look at the worlds nomologically accessible from w. Partition them by A's lifetime. That is, divide the possibilities into a class throughout which the lifetime is 1 millisecond, one where it is 10 years, one where it is 1000 years, and so on. All these classes are nonempty, because (I assume) the laws of w only assign very low probability to the atom decaying after 1000 years; they don't rule it out completely.

Do the same with @, the actual world. Here, too, all lifetimes are possible. So again you get an equivalence class for 1 millisecond, another for 10 years, another for 1000 years, and so on. However, by assumption the probability that A's lifetime is 1000 years -- or, if you're afraid of zeros, between 900 and 1000 years -- is a lot higher in @ than in w. By the probability-as-sizes account, this entails (silly tricks with infinity aside) that there are more worlds in the "1000 years" class for @ than in the "1000 years" class for w. But there is no reason to believe this.

Of course, there are many "1000 years" worlds for @. They differ in the outcome of other probabilistic events, for example, in where exactly A is located at the beginning of year 956. But by the same reasoning as above, we can assume that all these possibilities are possible relative to w as well. If in @ there is a positive chance of the atom living 1000 years and being located at x in 956, then at w there is also a positive (though different) chance of it living 1000 years and being located at x in 956. In general, for any world in @'s "1000 years" class, there is a corresponding world in w's "1000 years" class. So the one class cannot have more worlds than the other.

# on 22 December 2006, 05:08

BTW, these arguments all look way too technical. Really, my opinion isn't that the probabilities-as-sizes proposal faces various technical problems. Rather, I think it is obviously and intuitively false. Specific propositions (few worlds) often have high probability, and unspecific propositions very low probability. For example, I'm pretty sure that Canberra is around 16000 km from Berlin, but there are a lot more ways for this to be false than true, even if we restrict the ways to ways I cannot definitely rule out. Again, I'm pretty sure there is no God, but there are lots and lots of worlds with Gods none of which I can rule out for certain.

The principle might sound attractive because probability is somehow related to frequency, and obviously if we define P(x) as the relative frequency of x within a suitable domain, and take this domain as the basis for measuring the size of a proposition, then P(x) = 0.5 iff x obtains in 50% of the relevant cases. But at least the probabilities I care about (chance and credence) cannot meaningfully be defined as frequencies.

Hm. Maybe the probabilities-as-sizes account just *is* naive frequentism (without using limits for infinite cases)?

Add a comment

Please leave these fields blank (spam trap):

No HTML please.
You can edit this comment until 30 minutes after posting.