A method to find and verify contraction metrics

A new workflow uses a learner and a verifier to establish stability guarantees for continuous-time autonomous systems.

A new workflow connects the learning of a metric with numerical verification. This process provides quantitative dynamical guarantees for continuous-time autonomous systems on compact domains. It uses a learner to propose a metric and a verifier to check it.

Previously, learning a metric and proving it works were separate tasks. A low training loss on sampled states does not prove a contraction inequality holds across a whole domain. Also, a verifier with a conservative margin might reject a valid metric.

The method uses counterexample-guided inductive synthesis, or CEGIS. In this loop, a learner proposes a metric using polynomial or neural representations. A verifier then checks a matrix inequality or provides diagnostic points to help retraining.

The study analyzes why scaling and fixed-margin training can fail to certify valid metrics. It shows that the scaling construction needs a scale-closed family. The experimental fixed-floor class used in the study generally lacks this property.

The researchers used a physical swing-model to test the method. Verified constant metrics were able to certify the polynomial benchmark and the smallest swing-model domain. For these specific results, neural state dependence was not required.

The work provides computable polynomial bounds in any dimension. It also uses outward-rounded verification for neural, trigonometric, and polynomial representations. This approach allows for the conversion of verified slack into improved contraction rates.

The implementation records parameters, bounds, and outcomes. This data is necessary to audit successful certificates. The study also quantifies how much model error can be tolerated using certified residual and Jacobian bounds.

Key facts

  • The workflow uses counterexample-guided inductive synthesis to alternate between generating candidates and verifying them.
  • Verified slack can be converted into explicit improvements for the contraction rate.
  • The method provides ways to quantify model-error tolerances using certified residual and Jacobian bounds.
  • The study identifies that scaling and fixed-margin training can leave valid metrics uncertified.

All sections

Edition