Resolution with Counting: Lower Bounds for Proofs of Membership in the Complement of a Linear Map Image of the Boolean Cube

02/16/2022
โˆ™
by   Fedor Part, et al.
โˆ™
0
โˆ™

We propose a new approach to proving lower bounds for sizes of dag-like proofs in the proof system Res(lin_๐”ฝ_p), where ๐”ฝ_p is a finite field of prime order pโ‰ฅ 5. An exponential lower bound on sizes of arbitrary dag-like refutations in proof systems Res(lin_๐”ฝ) has previously been proven in (Part, Tzameret, ITCS'20) in case ๐”ฝ is a field of characteristic 0 for an instance, which is not CNF: for the binary value principle x_1+2x_2+โ€ฆ+2^n-1x_n = -1. The proof of this lower bound substantially uses peculiarities of characteristic 0 regime and does not give a clue on how to prove lower bounds neither for finite fields nor for CNFs. Aiming at constructing a bridge between lower bounds for the binary value principle and CNF lower bounds we initiate the development of methods for proving dag-like Res(lin_๐”ฝ_p) lower bounds for tautologies of the form bโˆ‰ A({0,1}^n), where A is a linear map. The negation of such a tautology can be represented in the language of Res(lin_๐”ฝ_p) as a system of linear equations Aยท x = b unsatisfiable over the boolean assignments. Instances of this form are in some ways simpler than CNFs, this makes analysis of their Res(lin_๐”ฝ_p) refutations more approachable and might be aided by tools from linear algebra and additive combinatorics. We identify hardness criterions for instances of the form Aยท x = b using notions of an error correcting code and what we call (s, r)-robustness, a combinatorial, algebraic property of linear systems Aยท x = b, which we introduce. We prove two lower bounds for fragments of Res(lin_๐”ฝ_p) that capture two complementary aspects of Res(lin_๐”ฝ_p) refutations and constitute a combinatorial toolbox for approaching general dag-like Res(lin_๐”ฝ_p) refutations.

READ FULL TEXT

Please sign up or login with your details

Forgot password? Click here to reset

Sign in with Google

ร—

Use your Google Account to sign in to DeepAI

ร—

Consider DeepAI Pro