Residuation–Galois Calculus for Exact Safety Preimages of Monotone Max–Plus Neural Networks
A proof-only calculus is developed for exact set abstraction and backward safety certification of monotone neural operators. A Galois insertion between concrete sets and interval boxes yields the best correct interval transformer. For every coordinatewise monotone map F, this transformer maps [ℓ,u] exactly to [F(ℓ),F(u)]; hence layerwise interval propagation is hull-exact for networks with nonnegative weights and isotone activations. The analysis is sharpened for max–plus layers T(x) =W ⊗x⊕b. Their upper safety preimages are either empty or principal ideals generated by the residualW\y, where (W\y)i = inf j:Wji>−∞(yj−Wji). Reverse residual propagation through a depth-L network computes the greatest input vector satisfying a prescribed upper output bound. Consequently, the largest weighted ℓ∞ radius around a nominal input is obtained in closed form, without optimization, branching, sampling, or relaxation. Soundness, maximality, compositionality, exactness, homogeneous collapse, and target-bound stability are proved. Four open problems concern signed architectures, two-sided class margins, residuated transformer attention, and completeness beyond boxes.
Volume & Issue
Vol. Volume 6 / Iss. Issue 1