Minimal generating sets from infinite rank constraints in the multi-block case

Determine which of the infinitely many rank constraints arising from linear combinations of block output matrices M_i yield a minimal—or even finite—generating set for the ideal J^{\mathbf{A}} of the ReLU pattern variety when multiple activation blocks are present.

Background

For multiple activation blocks, the paper shows that many determinantal constraints can be derived from linear combinations of block outputs, leading to infinitely many potential invariants.

Identifying a minimal finite set of generators would enable a complete implicit description of the output space in the multi-block setting and clarify which constraints are essential.

References

Moreover, in the case of multiple blocks, it remains to determine which of the infinitely many rank conditions yield a minimal (or even finite) generating set for the ideal.

Constraining the outputs of ReLU neural networks (2508.03867 - Alexandr et al., 5 Aug 2025) in Conclusion and future work (Section 9)