Convergent Gate Elimination and Constructive Circuit Lower Bounds

Abstract

Towards better understanding of gate elimination, the only method known that can prove complexity lower bounds for explicit functions against unrestricted Boolean circuits, this work contributes: (1) formalizing circuit simplifications as a convergent term graph rewriting system and (2) giving a simple and constructive proof of a classical lower bound using this system. First, we show that circuit simplification is a convergent term graph rewriting system over the DeMorgan and $\{\land, \lor, \oplus\}$ bases. We define local rewriting rules from Boolean identities such that every simplification sequence yields an identical final result (up to circuit isomorphism or bisimulation). Convergence enables rigorous reasoning about structural properties of simplified circuits without dependence on the order of simplification. Then, we show that there is \emph{no similar} convergent formalization of circuit simplification over the $U_2$ and $B_2$ bases. Then, we use our simplification system to give a constructive circuit lower bound, generalizing Schnorr's classical result that the XOR function requires $3(n - 1)$ gates to compute in the DeMorgan basis. A constructive lower bound $f \not\in C$ gives an algorithm (called a "refuter") that efficiently finds counter-examples for every $C$-circuit trying to compute the function $f$. Chen, Jin, Santhanam, and Williams showed that constructivity plays a central role in many longstanding open problems about complexity theory (FOCS 2021), so it is natural to ask for constructive circuit lower bounds from gate elimination arguments. This demonstrates how using convergent simplification can lead to shorter and more modular proofs of circuit lower bounds. Furthermore, until this work, no constructive lower bound had been proved via gate elimination.