A control barrier function certifies safety — it provides a mathematical guarantee that a system's state will remain within a safe set, regardless of disturbances, as long as the control law satisfies certain conditions. The function defines a boundary, and the control keeps the system inside. The question: can such a function always be found for a given safe set?
No. The researchers extend Brockett's 1983 theorem — which identified topological obstructions to continuous stabilizing control — to control barrier functions. The result: some safe sets, by their topological structure, cannot admit control barrier functions. The geometry of the set itself prevents the existence of the safety certificate, no matter how clever the control design.
Note (2026-11-12): checked against the paper's full text (de Sa and Ames). The paper's necessary conditions are not about the shape of the safe set alone. For a compact safe set whose Euler characteristic is not zero, a control barrier function together with a continuous controller implies that for every small enough perturbing vector field there is a state and an input that produce it (the system's dynamics must reach every small perturbation), in a form analogous to Brockett's test. So the obstruction comes from the set's topology (the Euler characteristic, not a general account of holes or the fundamental group) combined with a property of the dynamics, and it concerns a continuous safe controller as well as the barrier function. The paper says kinematic nonholonomic systems fail its necessary conditions (its worked example for barrier functions is a kinematic underactuated satellite on SO(3), where the dynamics cannot produce certain perturbing vector fields; the nonholonomic integrator is the paper's recalled example of Brockett's own condition); the cause is the dynamics failing the condition, not a safe set with holes. The essay's 'depends on the shape of what you protect, not on the dynamics' and the 'safe but not certifiably safe' framing are the essay's own.
The obstruction is topological, meaning it depends on the shape of the safe set (its holes, connectivity, fundamental group) rather than on the specific dynamics. A safe set with the wrong topology cannot be certified safe by a CBF even if the system can actually stay safe inside it. The barrier function might exist as a concept but not as a smooth function on that topology.
The application to nonholonomic systems is concrete. A car that cannot move sideways (nonholonomic constraint) has a state space with topological features that classical smooth barrier functions cannot accommodate. The constraint creates holes in the control authority that no smooth CBF can paper over.
The structural insight: safety certification and actual safety are different problems. A system can be safe (the controller keeps it in the set) without being certifiably safe (no CBF exists). The gap between “is safe” and “can prove it's safe” is topological — it depends on the shape of what you're trying to protect, not on how hard you try to protect it.
(arXiv:2603.18422)