Paper 2026/1087

Low-Norm Nullstellensatz Hypothesis for the AND Code is False

Abstract

The recent work [Devadas-Hopkins-Kalai-Kothari-Lombardi-Mathialagan, STOC 2026] proposed a low-norm Nullstellensatz hypothesis for the "AND code": every polynomial $f$ vanishing on the "AND-code ideal'' should admit a Nullstellensatz decomposition over the local AND constraints whose total coefficient \(\ell_1\)-norm is only polynomially larger than the \(\ell_1\)-norm of $f$. We give a counterexample to this conjecture by proving an exponential lower bound on the total coefficient \(\ell_1\)-norm. The core idea of the proof was discovered by ChatGPT 5.5 Pro, and we verified and reorganized the proof to improve its exposition. The proof constructs a dual linear functional, whose analysis leverages the rank of the quadratic forms to bound Fourier correlations. The counterexample can also be extended to give the first \(\ell_1\)-norm lower bound for Nullstellensatz refutations over the \(\{\pm1\}\)-basis. Previously, \(\ell_1\)-norm lower bounds for Nullstellensatz refutations were known only over the \(\{0,1\}\)-basis, due to Potechin and Zhang [ICALP 2024]. We believe this is of independent interest to proof complexity.