I’ve been thinking lately about formalizations of the Correspondence Principle - the idea that new theories should reproduce old theories, at least in the places where the old theories work. Special relativity reduces to Galilean relativity at low speed/energy, general relativity reduces to Newtonian gravity when the fields are weak, quantum mechanics should reproduce classical mechanics at large scale, etc. More conceptually, it’s the idea that flowers are “real”: any model which does a sufficiently-good job of predicting the world around me should have some kind of structure in it corresponding to my notion of a flower (though it may not be ontologically basic).

I want theorems telling me when my models are of the right kind, and sufficiently-well-nailed-down, that I can expect some kind of correspondence along these lines to apply for any future theory (assuming the future theory has sufficiently strong predictive power). Ideally, this would allow us to take a model and extract the “generalizable” information from it, construct some universal representation of certain components of the model which lets us find the “corresponding” components in any other model with sufficiently-good predictive performance on the same data. (This problem is one example of a potential application.)

This post introduces one such theorem. We’ll show that a (particular type of) correspondence theorem applies in exactly those cases where our model is able to make deterministic predictions about the relationship between some observable variables.

I've tried to keep the hairiest of the math in the appendices and the "key ideas from the math" section. If you want the key ideas with minimal math, then skip those sections.

The Claim

Here’s the setup. We have two random variables,  and , and we have enough data that we can perfectly figure out their joint distribution . We don’t necessarily know what physical process generates these variables - i.e. we don’t know the true underlying generative model. You can imagine some scientific experiment where we can collect lots of independent samples of , then brute-force count how often each  pair occurs in order to estimate , but we don’t have any way to experimentally probe the “internals” of the experimental setup. In other words, it's the traditional setup from frequentist probability.

We’ll assume that the class of “theories” we’re interested in are generative models - i.e. programs which take in some independent uniform variables and spit out . For our purposes, we’ll assume these programs are represented as systems of circuits/causal DAGs/Bayes nets - see here for what that sort of representation looks like. (Note that this representation is Turing complete, so we’re not imposing any significant restriction.)

As long as  and  are not independent, any generative model which reproduces the distribution  must have some variable(s) upstream of both  and  - some common upstream cause(s) which account for the relationship between the variables. The claim I’d like to make is: any upstream variable(s) capable of accounting for the relationship between  and  must contain some “minimal” information. Formally: we can construct some variable  (defined by the distributions ) such that  and  are independent given :

… and, for any other  which induces conditional independence of  and  is a (possibly stochastic) function of : for all  such that 

... we have

.

Conceptually:  is the “minimal” information about the relationship between  and . It must be “included in” any variable  capable of accounting for the relationship between those two variables.

Visualization in terms of Bayes nets. Our main goal is to prove the existence of , and construct it from . We’ll see that this can only be done for some distributions.

That’s the claim I’d like to make. However, that claim is too strong - not all distributions  have such a . Indeed, in some sense “most” do not, though we can work around that to a large extent. The next section will explain exactly when we can construct such a , and the appendix gives the proof. Specifically, we can construct  in exactly those cases where the relationship between  and  is independent variation subject to a deterministic constraint. Formally: we can construct  exactly when  can be represented as

… for some deterministic functions  (here  is the indicator function). In that case, we can choose .

Key Ideas From The Math

We have to be able to calculate our “universal” upstream  from any  which induces independence of . In particular,  and  are independent given either  or  itself - so we can choose either of those to be . Combine that with the original requirements, and we have:

  •  independent of  given 
  •  independent of  given 
  •  independent of  given 

This three-way conditional independence requirement is the main constraint which determines which distributions  have a .

One easy way to satisfy this three-way conditional independence is when , and  are all just completely independent. Another is when all three are deterministically equal - i.e.  implies  implies , etc. And of course we can rename variable values while maintaining three-way conditional independence - e.g. we could rename  values from  to . So all three variables isomorphic - rather than equal - also works: , with  and  invertible.

We could also combine these two possibilities: , and  could each have two components, where one component is independent and the other component is fully determined by either of the other variables. For instance, we could have , with , and  all completely independent, but , and  all deterministically isomorphic. Again, we can also rename variable values, so our variables don’t need to literally be written with two components.

… and that turns out to be the most general possibility.

Visually: if we lay out the distribution  in a matrix (with rows corresponding to  values and columns corresponding to  values), then we can arrange that matrix to look like this:

Here the dark blocks are nonzero values; everything else is zero. Within the dark blocks, we have  - i.e. the dark blocks are each rank-1 submatrices. Intuitively, our distribution  consists of a deterministic relationship (the choice of which block we’re in) plus independent variation (choice of values within the block). This generalizes in the obvious way to the full distribution : the full distribution consists of non-overlapping 3D blocks of nonzeros, with the three variables independent within each block.

For the full proof that these are the only distributions for which we can construct , see the appendix.

Note that we do still have a degree of freedom in choosing  - we can add or remove independent components. The obvious choice is to remove all the independent components from , and just keep the deterministic component. Visually, we take  to “choose a block” in the graphic above. (In fact, we could have just specified upfront that  is a deterministic function of any possible , rather than the more general stochastic function, but I usually prefer to use the more general version just in case the problem has some implicit embedded game with mixed equilibria - the stochastic function allows  to “randomize its strategy”. In this case, it turns out that the generality doesn’t buy us anything, and that is itself useful to know; it means there’s probably no diagonalization shenanigans going on.)

The appendix proves that this choice of  indeed satisfies all of our original conditions.

To wrap up this section, let’s write out the shape of  mathematically. Each value of  has nonzero probability in only one “block”, so we can create a function  which picks out the block for each  value. Likewise for  can be nonzero only if  and  are in the same block: . The chance of landing in a particular block is  (which is equal to , since  and  must always be equal). Within the block (i.e. conditional on ),  and  are independent. Put all that together, and we get

Since  picks out the block, and  is the block choice, 

Applicability

At first glance, it looks like the conditions required to use this theorem are pretty restrictive. But there’s an easy trick to generalize it somewhat.

Suppose we have a distribution  which does not fully satisfy the requirements, but it does have a deterministic constraint:  with probability 1. Then we can’t apply our correspondence theorem to  and , but we can apply it to  and . We can construct  satisfying this:

In particular, we can choose .

So: whenever we can identify a deterministic constraint in the world, we can apply this correspondence theorem. In other words, deterministic constraints in one model should have corresponding structures in other models, to the extent that all models match the environment.

Note that “deterministic constraints” is exactly the content of most models in the hard sciences. For instance, equations in physics are almost always deterministic constraints on the trajectory of world-states. To the extent that these equations match the real world, we should see corresponding constraints in future theories/models which also match the real world.

We can push this further.

We haven’t talked about approximations here, but let’s assume that the theorem generalizes to approximations in the obvious way - i.e. approximately deterministic constraints give approximate correspondence, with independencies replaced by approximate independencies. Once we have approximations, we can construct deterministic constraints via statistical identification.

Here’s what that means. Consider something like the ideal gas law, . Observing a few specific particles will not tell us the pressure or temperature of the gas. Individual particles have only partial information about the high-level variables, so we don’t have a correspondence theorem. However, if we bucket together a whole bunch of particles, then we can get a very precise estimate of temperature and pressure - in stats terminology,  and  can be “identified” from our many independent particles. We can then have deterministic relationships between the identified variables, like . Then we can apply the correspondence theorem.

Similarly, we could identify variables using multiple “bunches of particles” - or, more generally, multiple meaurements/multiple lines of evidence. For instance, maybe we can use a whole bunch of measurements to determine the gravitational constant. If we do this again with another set of measurements, we expect to get the same number. That’s a deterministic constraint: gravitational constant calculated from one set of measurements should equal gravitational constant calculated from another. To the extent that this constraint matches the real world, our correspondence theorem says the gravitational constant should correspond to something in any future theory.

Conclusion

In some ways, this theorem leaves a lot to be desired. In particular, it assumes access to the “true distribution”, which I’d really like to get rid of. I expect there are other correspondence theorems to be found, with very different formulations, some of which would be better in that regard. For instance, for maximum entropy models, there’s an easy correspondence theorem which says “given an old model and a new model, either the constraints of the old model are satisfied by the new model, or we can construct a third model which strictly outperforms the new model”. (That theorem is unimpressive in other ways, however - it doesn’t really say anything about the “internal structure” of the models.)

Aside from direct use as a correspondence theorem, “independent variation subject to deterministic constraints” is an interesting thing to see pop out. It’s a common pattern in the sciences - it describes far-apart low-level variables in any abstraction where the high-level model is deterministic. I also think it’s equivalent to zero distributed information, which I previously predicted ought to show up quite often when looking at information relevant far away in a system. On top of that, humans often have a (mistaken) intuition that information works like sets - information comes in discrete chunks, and a variable either contains a chunk or it doesn’t. This leads to some common mistakes when studying information theory. For systems in which variation is independent subject to deterministic constraints, the “information is like sets” intuition is correct, and I suspect that it’s the only case where that picture works in general (for some formulation of “in general”).

It seems like there’s something fundamental about this pattern, and this correspondence theorem is another hint.

Appendix: Proofs

Even having guessed the answer, I found these proofs surprisingly difficult. There’s probably some way to set things up so it’s more obvious what to do, but I’m not sure what it is, other than that someone will tell me it has something to do with category theory.

The problem: we want  for which

and for all  such that , we have

.

We’ll show that:

  • This is only possible when  for some 
  • When  has that form, the choice  works.

Correspondence Only If Independent Variation Subject to Deterministic Constraint

We'll start off the same as earlier: we want  (in English:  independent of  given ) for any  satisfying . Well, there’s two choices for  which definitely satisfy : namely,  and  themselves. So:

… and by assumption, . So, any two of the variables , and  are independent given the third.

Let’s write out  in two different ways, using two of our conditional independencies:

Note that , so we can cancel those terms out and find

if . Note that, since our three-way independence condition is symmetric in the variables, we can also switch around the variables to get:

  •  when 
  •  when 

These are quite strong. Pick any two  values,  and , for which any  value has both  and . (Terminology: we’ll say that these two  values “overlap” on , i.e. either value occurs with nonzero probability when . We'll also say that values of two different variables overlap when they have nonzero joint probability, e.g.  overlaps  iff ). Then

Furthermore, since  for all , any  with  will also have . In other words, if two  values overlap on any  value, then they also overlap on all  values which overlap either  value. That, in turn, means

… for any value  on which  overlap. Finally, this also means that two  values which overlap on any  value overlap on all  values which overlap either  value: ( and ) for any  implies that  on all  with  , and vice versa.

That last condition is especially useful, since it means that overlap is transitive: if  and  overlap, and  and  overlap, then  and  overlap, and all three overlap on the same  values.

That finally lets us construct our functions  and . Since we have transitivity of overlap, we can use overlap as an equivalence relation, and let  choose the equivalence class into which  falls.  chooses the equivalence class of  values with which  overlaps (note that there can only be one, since this class contains all the  values which overlap with this  value). This is the “choice of block” from our earlier visual.

The rest follows (relatively) easily. From the definition of overlap,  whenever . Our independence conditions from earlier give