Counterfactuals, thick and thin

post by Nisan · 2018-07-31T15:43:59.187Z · score: 29 (12 votes) · LW · GW · 11 comments

Summary: There's a "thin" concept of counterfactual that's easy to formalize and a "thick" concept that's harder to formalize.

Suppose you're trying to guess the outcome of a coinflip. You guess heads, and the coin lands tails. Now you can ask how the coin would have landed if you had guessed tails. The obvious answer is that it would still have landed tails. One way to think about this is that we have two variables, your guess and the coin , that are independent in some sense; so we can counterfactually vary while keeping constant.

But consider the variable XOR . If we change to tails and keep the same, we conclude that if we had guessed tails, the coin would have landed heads!

Now this is clearly silly. In real life, we have a causal model of the world that tells us that the first counterfactual is correct. But we don't have anything like that for logical uncertainty; the best we have is logical induction, which just give us a joint distribution. Given a joint distribution over , there's no reason to prefer holding constant rather than holding XOR constant. I want a thin concept of counterfactuals that includes both choices. Here are a few definitions, in increasing generality:

1. Given independent discrete random variables and , such that is uniform, a thin counterfactual is a choice of permutation of for every .

2. Given a joint distribution over and , a thin counterfactual is a random variable independent of and an isomorphism of probability spaces that commutes with the projection to .

3. Given a probability space and a probability kernel , a thin counterfactual is a probability space and a kernel such that .

There are often multiple choices of thin counterfactual. When we say that one of the thin counterfactuals is more natural or better than the others, we are using a thick concept of counterfactuals. Pearl's concept of counterfactuals is a thick one. No one has yet formalized a thick concept of counterfactuals in the setting of logical uncertainty.

11 comments

Comments sorted by top scores.

comment by Dacyn · 2018-07-31T18:37:57.886Z · score: 7 (4 votes) · LW(p) · GW(p)

The question "how would the coin have landed if I had guessed tails?" seems to me like a reasonably well-defined physical question about how accurately you can flip a coin without having the result be affected by random noise such as someone saying "heads" or "tails" (as well as quantum fluctuations). It's not clear to me what the answer to this question is, though I would guess that the coin's counterfactual probability of landing heads is somewhere strictly between 0% and 50%.

comment by Nisan · 2018-08-01T05:46:46.037Z · score: 3 (2 votes) · LW(p) · GW(p)

Oh, interesting. Would your interpretation be different if the guess occurred well after the coinflip (but before we get to see the coinflip)?

comment by Dacyn · 2018-08-01T15:58:29.931Z · score: 5 (3 votes) · LW(p) · GW(p)

Sure, in that case there is a 0% counterfactual chance of heads, your words aren't going to flip the coin.

comment by Nisan · 2018-08-02T22:29:32.624Z · score: 4 (2 votes) · LW(p) · GW(p)

Ok. I think that's the way I should have written it, then.

comment by shminux · 2018-08-01T03:22:27.044Z · score: 3 (2 votes) · LW(p) · GW(p)

I agree that is is a well-defined question, though not easily answered without knowing how guessing physically affects flipping the coin, reading the results (humans are notoriously prone to making mistakes like that) and so on. But I suspect that Nisan is asking something else, though I am not quite sure what. The post says

In real life, we have a causal model of the world that tells us that the first counterfactual is correct. But we don't have anything like that for logical uncertainty; the best we have is logical induction, which just give us a joint distribution.

I am not sure how physical uncertainty is different from logical uncertainty, maybe there are some standard examples there that could help the uninitiated like myself.

comment by Vaniver · 2018-08-01T05:05:46.816Z · score: 4 (2 votes) · LW(p) · GW(p)

If we have an ordering over logical sentences, such that we can look at two sentences and determine at most one of (A is simpler than B), (B is simpler than A), then it seems natural to privilege the counterfactual that keeps the simpler term constant (and likely that this ordering is such that you never have to choose between counterfactuals at the same level of simplicity).

This doesn't fully solve the problem--now I have a concept of thickness that's predicated on an ordering, and the ordering is (in some sense) arbitrary for the reasons noted elsewhere (I could define B = A xor C as the ground term, which makes A = B xor C now a composite term). But it seems (to me) like the important thing is being able to build a model that doesn't allow cyclical behavior at all. Afterwards, one can check to see whether or not the ordering matters (and if so, try to figure out the criteria that make for a good ordering), or view it as arbitrary in approximately the way that axiom sets are arbitrary.

comment by Chris_Leong · 2018-08-01T03:59:53.545Z · score: 3 (2 votes) · LW(p) · GW(p)

Thanks for this post, it's really helpful. I would really like to understand the maths in this post, is there anywhere which describes this in more detail? In particular, I can't follow:

  • Why are probabilities being permutated?
  • What kind of kernel are you referring to?
comment by Nisan · 2018-08-01T06:03:21.138Z · score: 5 (3 votes) · LW(p) · GW(p)

The definition involving the permutation is a generalization of the example earlier in the post: is the identity and swaps heads and tails. And . In general, if you observe and , then the counterfactual statement is that if you had observed , then you would have also observed .

I just learned about probability kernels thanks to user Diffractor [LW · GW]. I might be using them wrong.

comment by GuySrinivasan · 2018-07-31T16:46:43.987Z · score: 2 (1 votes) · LW(p) · GW(p)

I don't know enough math to understand whether you've covered this in your examples, but here's my intuition in the form of typing without a lot of reflection or editing okay disclaimer over:

If we have two variables, A and C, and we're considering A, C, and (A xor C), it sounds to me like we've privileged things arbitrarily in some sense... relabeling them A, B, and C it's clear that we could have pivoted to consider any two of them the "base" variables and the third the "xor'd" variable, so there should be no preferred counterfactual. It's a loopy cause, a causal diagram that's not a DAG. Which doesn't show up IRL. Like going back in time to kill grandpa.

But we often pretend they occur by abstracting time and saying steady-state is a thing (or steady-states, and we're looking at the map of transitions) and then we get loops and start studying feedback and whatnot. But if you unpacked any of those loops you'd get a very-very-repetitive DAG that looks a lot like the initial diagram copied over and over with one-way arrows from copy to copy.

Seems like there are three options to deal with {A,B,C}. They are isomorphic to each other, so in some sense we shouldn't be able to say which counterfactuals to use. We could:

  • do our modeling relative to a specified imposed ordering of all variables, which seems really hard, or
  • somehow calculate all possible results and average over permutations, which seems either factorially harder or much easier depending on Math!, or
  • assume there is hidden structure, that A, B, and C are abstractions atop a real DAG, and use a not-known-to-me mathematics of loopy causation to define something other than counterfactuals atop the variables, calling counterfactuals over A, B, C a sort of type error.
comment by Chris_Leong · 2018-08-20T13:36:52.997Z · score: 2 (1 votes) · LW(p) · GW(p)

I'm not following what you're saying about loopy causation. How are you constructing this graph?

comment by Nisan · 2018-08-01T01:57:40.074Z · score: 2 (1 votes) · LW(p) · GW(p)

That sounds about right to me. I think people have taken stabs at looking for causality-like structure in logic, but they haven't found anything useful.