Paper 2024/1531

FLI: Folding Lookup Instances

Abstract

We introduce two folding schemes for lookup instances: FLI and FLI+SOS. Both use a PIOP to check that a matrix has elementary basis vectors as rows, with FLI+SOS adding a twist based on Lasso’s SOS-decomposability. FLI takes two lookup instances $\{\mathbf{a}_1\}, \{\mathbf{a}_2\}\subseteq\mathbf{t}$, and expresses them as matrix equations $M_i\cdot\mathbf{t}^\mathsf{T}=\mathbf{a}_i^\mathsf{T}$ for $i=1,2$, where each matrix $M_i\in\mathbb{F}^{m\times N}$ has rows which are elementary basis vectors in $\mathbb{F}^N$. Matrices that satisfy this condition are said to be in $R_{\mathsf{elem}}$. Then, a folding scheme for $R_{\mathsf{elem}}$ into a relaxed relation is used, which combines the matrices $M_1, M_2$ as $M_1+\alpha M_2$ for a random $\alpha\in\mathbb{F}$. Finally, the lookup equations are combined as $(M_1+\alpha M_2)\cdot \mathbf{t}^{\mathsf{T}} = (\mathbf{a}_1+\alpha \mathbf{a}_2)^\mathsf{T}$. In FLI, only the property that a matrix is in $R_{\mathsf{elem}}$ is folded, and this makes the FLI folding step the cheapest among existing solutions. The price to pay is in the cost for proving accumulated instances. FLI+SOS builds upon FLI to enable folding of large SOS-decomposable tables. This is achieved through a variation of Lasso's approach to SOS-decomposability, which fits FLI naturally. For comparison, we describe (for the first time to our knowledge) straightforward variations of Protostar and Proofs for Deep Thought that also benefit from SOS-decomposability. We see that for many reasonable parameter choices, and especially those arising from lookup-based zkVMs, FLI+SOS can concretely be the cheapest folding solution.