Paper 2024/1210

More Optimizations to Sum-Check Proving

Abstract

Many fast SNARKs apply the sum-check protocol to an $n$-variate polynomial of the form $g(x) = \text{eq}(w,x) \cdot p(x)$, where $p$ is a product of multilinear polynomials, $w \in \mathbb{F}^n$ is a random vector, and $\text{eq}$ is the multilinear extension of the equality function. In this setting, we describe an optimization to the sum-check prover that substantially reduces the cost coming from the $\text{eq}(w, x)$ factor. Our work further improves on a prior optimization by Gruen (ePrint 2023), and in the small-field case, can be combined with additional optimizations by Bagad, Domb, and Thaler (ePrint 2024), and Dao and Thaler (ePrint 2024). Over large prime-order fields, our optimization eliminates roughly $2^{n + 1}$ field multiplications compared to a standard linear-time implementation of the prover, and roughly $2^{n-1}$ field multiplications when considered on top of Gruen's optimization. These savings are about a 25% (respectively 10%) end-to-end prover speedup in common use cases, and potentially even larger when working over binary tower fields.