Scalable Incremental Robustness Analysis of Neural Network Feedback Systems

This paper addresses the scalability challenge in verifying neural network feedback systems. Traditional semidefinite programming certificates scale with total neuron count, while small-gain tests are scalable but conservative. The authors develop a unified framework that combines structured decomposition techniques to enable efficient incremental robust stability and performance analysis for feedback interconnections involving high-dimensional neural networks and unmodeled dynamics. The approach balances computational tractability with verification accuracy for NN-augmented control systems.