Order-2 Tightness of Block-Sparse SOS Relaxations for One-Layer ReLU Network Verification with a Matching Input-Sharing Graph

Azuma, Kim, and Yamashita formulated the verification problem for one-layer ReLU networks as a quadratically constrained quadratic program and established tight semidefinite relaxations for the edgeless case and for one-unit settings. In this work, we represent the sharing pattern of undecided ReLUs over a box input set through an input-sharing graph and focus on the case where this graph is a matching. We then derive an explicit, checkable sufficient condition for the tightness of the order-$2$ block-sparse SOS relaxation associated with the connected-component decomposition of this graph. Under the matching assumption, the global problem decomposes into isolated-vertex blocks and single-edge blocks. The key difficulty, which is absent from the edgeless case, is establishing tightness for a two-unit edge block. For regular rank-one edges, we show that the convex hull of each two-unit local set can be described exactly by two reduced one-unit hulls coupled through a common shared scalar. Combining the one-unit tightness result of Azuma et al. with Farkas’ lemma and affine elimination, we obtain a local order-$2$ certificate for each edge block. Isolated-vertex blocks reduce to one-unit problems over box input sets and are therefore handled at the same order. We prove that, when the input-sharing graph is a matching and every edge satisfies the regular rank-one condition, the order-$2$ block-sparse SOS relaxation is tight. This extends the tight sparse relaxation result for the edgeless case to the first sparse setting with a nontrivial two-unit interaction.

Article

Download

View PDF