Home Online Mathematical Communication Computational Complexity Natural Proofs is Not the Barrier You Think It Is

Natural Proofs is Not the Barrier You Think It Is

Computational Complexity 2024-09-11

If there's a position where I differ from most other complexity theorists it's that I don't believe that natural proofs present a significant barrier to proving circuit results. I wrote about this before but I buried the lead and nobody noticed.

Let's review natural proofs. I'll give a very high-level description. Li-Yang Tan gives a good technical description or you could read the original Razborov-Rudich paper. A natural proof to prove lower bounds against a circuit class \(\cal C\) consists of a collection \(C_n\) of Boolean functions on \(n\) inputs such that

  1. No polynomial-size circuit family from \(\cal C\) can compute an element of \(C_n\) for large enough \(n\). 
  2. \(C_n\) is a large fraction of all the function on \(n\) inputs.
  3. A subset of \(C_n\) is constructive--given the truth-table of a function, you can determine whether it sits in the subset in time polynomial in the length of the truth-table. Note: This is a different than the usual notion of "constructive proof". 
The natural proof theorem states that if all three conditions hold than you can break pseudorandom generators and one-way functions.
My problem is with the third property, constructivity. I haven't seen good reasons why a proof should be constructive. When I saw Rudich give an early talk on the paper, he both had to change the definition of constructivity (allowing subsets instead of requiring an algorithm for \(C_n\) itself) and needed to give heavily modified proofs of old theorems to make them constructive. Nothing natural about it. Compare this to the often maligned relativization where most proofs in complexity relativize without any changes.
Even Razborov and Rudich acknowledge they don't have a good argument for constructivity.
We do not have any formal evidence for constructivity, but from experience it is plausible to say that we do not yet understand the mathematics of \(C_n\) outside exponential time (as a function of \(n\)) well enough to use them effectively in a combinatorial style proof.
Let's call a proof semi-natural if conditions (1) and (2) hold. If you have a semi-natural proof you get the following implication. 
Constructivity \(\Rightarrow\) One-way Functions Fail
In other words, you still get the lower bound, just with the caveat that if an algorithm exists for the property than an algorithm exists to break a one-way function. You still get the lower bound, but you are not breaking a one-way function, just showing that recognizing the proofs would be as hard as breaking one-way functions. An algorithm begets another algorithm. You don't have to determine constructivity either way to get the lower bound. 
Even if they aren't a great barrier to circuit lower bounds, natural proofs can be an interesting, if badly named, concept in their own right. For example the Carmosino-Impagliazzo-Kabanets-Kolokolova paper Learning Algorithms from Natural Proofs.
So if I don't believe in the barrier, why are circuit lower bounds hard? In recent years, we've seen the surprising power of neural nets which roughly correspond to the complexity class TC\(^0\), and we simply don't know how to prove lower bounds against powerful computational models. Blame our limited ability to understand computation, not a natural proof barrier that really isn't there.