Contraction Certification from Streaming Data: Wasserstein Robustness and Compositional Stability for Interconnected Nonlinear System
arXiv:2607.11982
2026
Dynamics
2 ideas extracted · analyzed Aug 30, 2026
What the math gives to ML
The paper provides two transferable mechanisms: an online Wasserstein-robust contraction certificate that discounts a nominal margin according to distributional shift estimated from recent residual kurtosis, and a compositional small-gain certificate for interconnected subsystems that requires only local contraction estimates. For neural networks, these mechanisms can become adaptive step-size or trust-region controllers under nonstationary and heavy-tailed gradients, and stability constraints for modular networks with cross-module attention or message passing. Both transfers make sharp predictions: the robust certificate should turn negative during sufficiently heavy-tailed shift, while compositional stability should fail near the coupling threshold \(\gamma=\sqrt{\beta_A\beta_B}\).
Ideas from this paper
✗ Failed on benchmark
2026
Partition a neural network into independently trained or independently monitored modules and constrain their cross-module interaction gain using a compositional contraction certificate. This enables stable deep modular MLPs, graph blocks, or recurrent modules without estimating the full network Jacobian, while providing an explicit coupling threshold for when the architecture loses contraction.
Useful8/10
Difficulty6/10
Novelty7/10
△ Mechanism confirmed, baseline not beaten
2026
Treat one optimizer update as a stochastic dynamical map and estimate its local contraction margin from recent parameter-update or gradient residuals. Reduce the usable margin, and therefore the learning rate or trust-region radius, by a Wasserstein/heavy-tail penalty based on online excess kurtosis so distribution shifts cause graceful step-size shrinkage rather than sudden divergence.
Useful8/10
Difficulty5/10
Novelty7/10