Approximate polymorphisms

Abstract: For a function $g\colon{0,1}^m\to{0,1}$, a function $f\colon {0,1}^n\to{0,1}$ is called a $g$-polymorphism if their actions commute: $f(g(\mathsf{row}1(Z)),\ldots,g(\mathsf{row}_n(Z))) = g(f(\mathsf{col}_1(Z)),\ldots,f(\mathsf{col}_m(Z)))$ for all $Z\in{0,1}^{n\times m}$. The function $f$ is called an approximate polymorphism if this equality holds with probability close to $1$, when $Z$ is sampled uniformly. We study the structure of exact polymorphisms as well as approximate polymorphisms. Our results include: - We prove that an approximate polymorphism $f$ must be close to an exact polymorphism; - We give a characterization of exact polymorphisms, showing that besides trivial cases, only the functions $g = \mathsf{AND}, \mathsf{XOR}, \mathsf{OR}, \mathsf{NXOR}$ admit non-trivial exact polymorphisms. We also study the approximate polymorphism problem in the list-decoding regime (i.e., when the probability equality holds is not close to $1$, but is bounded away from some value). We show that if $f(x \land y) = f(x) \land f(y)$ with probability larger than $s\land \approx 0.815$ then $f$ correlates with some low-degree character, and $s_\land$ is the optimal threshold for this property. Our result generalize the classical linearity testing result of Blum, Luby and Rubinfeld, that in this language showed that the approximate polymorphisms of $g = \mathsf{XOR}$ are close to XOR's, as well as a recent result of Filmus, Lifshitz, Minzer and Mossel, showing that the approximate polymorphisms of AND can only be close to AND functions.