Mark proves that if is an AC0 boolean circuit (with NOT gates and with AND gates and OR gates of unbounded fan-in) of depth and size , and if is any -wise independent distribution with , then
that is, “fools” the circuit into thinking that is the uniform distribution over . Plausibly, this might be true even for .
Nothing was known for depth 3 or more, and the depth-2 case was settled only recently by Bazzi, with a proof that, as you may remember, has been significantly simplified by Razborov about six months ago.
Mark’s proof relies on approximating via low-degree polynomials. The point is that if is an -variate (real valued) polynomial of degree , and is a -wise independent distribution ranging over , then
Now if we could show that approximates both under and under , in the sense that , and also , then we would be done.
The Razborov-Smolenski lower bound technique gives a probabilistic construction of a polynomial such that for every input one has a high probability that . In particular, one get one polynomial such that both
Unfortunately this is not sufficient, because the polynomial might be very large at a few points, and so even if agrees with with high probability there is no guarantee that the average of is close to the average of .
Using a result of Linial, Mansour and Nisan (developed in the context of learning theory), one can construct a different kind of low-degree approximating polynomial , which is such that
The Linial-Mansour-Nisan approximation, however, says nothing about the relation between and under the distribution .
Using ideas of Bazzi’s, however, if we had a single polynomial such that properties (1), (2) and (3) are satisfied simultaneously, then we could construct another low-degree polynomial such that , and also , giving us that is fooled by .
As far as I understand, Mark constructs a polynomial satisfying properties (1), (2) and (3) by starting from the Razborov-Smolenski polynomial , and then observing that the indicator function of the points on which is itself a boolean function admitting a Linial-Mansour-Nisan approximation . Defining , we have that has all the required properties, because multiplying by “zeroes out” the points on which is excessively large.
I have been interested in this problem for some time because of a connection with the complexity of 3SAT on random instances.
As I have discussed in the past, if we take a random instance of 3SAT constructed by randomly picking, say, 10n clauses over n variables, then there is an extremely high probability that the formula is unsatisfiable. Certifying that such formulas are unsatisfiable, however, appears to be hard.
Here what we want is a certification algorithm that, on input a boolean formula , outputs either unsatisfiable or don’t know; if the algorithm outputs unsatisfiable then the formula is definitely guaranteed to be unsatisfiable; and the algorithm outputs unsatisfiable on at least, say, a
inverse polynomial constant fraction of inputs. (Thanks for Albert Atserias for pointing out that the problem is trivial if one requires only inverse polynomial refutation probability.)
The existence of such an algorithm for random 3SAT with variables and clauses has been ruled out in a variety of bounded models of computation. (And of course the lower bounds extend to higher number of clauses than .) For example, we know that, except with very small probability, no sub-exponential size “tree-like resolution” proof of unsatisfiability can exist for formulas from this distribution, and we know that when a back-tracking based algorithm finishes without having found a satisfying assignment, its computation defines a tree-like resolution proof. Hence no back-tracking based algorithm can run in sub-exponential time. We also know lower bounds for algorithms based on convex relaxations of Max 3SAT, including algorithms based on Lasserre semidefinite programming relaxations, which are about the most powerful convex programming relaxations that we know how to construct.
Notably, I don’t know of a lower bound showing that no such refutation algorithm can be designed in AC0, even though AC0 is a class for which we are usually able to prove lower bounds. (As far as I know, the question remains open.)
All the above lower bounds, by the way, apply also to random instances of the 3XOR problem with, say, n variables and 10n equations. (3XOR is like 3SAT but every “clause” is a linear equation mod 2 involving three variables.) 3XOR is of course solvable in polynomial time via Gaussian elimination, but in many simplified models of computation it seems to capture the hardness of 3SAT.
So what about refuting random 3XOR in AC0? Consider the following two distributions of instances:
- Distribution contains random instances with variables, clauses, and uniformly chosen right-hand-side
- Distribution contains instances with a random left-hand side involving variables and clauses, and a right-hand side chosen to be consistent with a random assignment.
Now the left-hand side has the same distribution in and , and for, most left-hand-sides, the distribution of the right-hand side in in -wise independent, while the right-hand side in is always uniform.
From Mark’s result it follows that and are indistinguishable by AC0 circuits, and so no AC0 circuit can refute random instances from distribution . (Otherwise it would incorrectly output unsatisfiable with positive probability given samples from , while all samples from are satisfiable.)