In long-horizon optimal control, you'd expect the optimal trajectory to look different at every time — responding to changing conditions, adapting to constraints. The turnpike property says otherwise: for most of the time horizon, the optimal trajectory stays near a single steady state, the “turnpike.” Departures happen only near the beginning (reaching the turnpike) and the end (meeting the terminal condition).
The authors (arXiv:2603.23303) prove exponential turnpike theorems for nonlinear deterministic mean-field optimal control problems. “Exponential” means the trajectory approaches the turnpike at an exponential rate and departs from it at an exponential rate near the end — the time spent away from the steady state is proportional to log(T), not T. Almost all of a long planning horizon is spent at the optimal steady state.
Mean-field problems describe populations — each agent is small, but the aggregate behavior matters. The turnpike property for these problems means that optimal population management, over long horizons, reduces to finding the right steady state and paying a logarithmic cost to reach and leave it.
The through-claim: long-horizon optimization is mostly steady-state optimization in disguise. The dynamic problem — with its initial conditions, terminal constraints, and time-varying objectives — collapses to a static problem for all but a logarithmic fraction of the time horizon. The complexity of the planning problem doesn't grow linearly with the horizon; it saturates.