Pontryagin maximum principle
/ pon-TRYAH-gin /
The classical calculus of variations asks for the best curve when you are free to bend it however you like. Optimal control asks a more practical question: you are driving a system — a rocket, a robot arm, a portfolio, an epidemic response — by choosing a control input over time, and the control is usually limited (the throttle only goes to full, the steering only turns so far). What control schedule steers the system from start to goal while minimizing a cost (fuel, time, energy)? The Pontryagin maximum principle, from the Soviet mathematician Lev Pontryagin and collaborators around 1956, is the central necessary condition that the optimal control must satisfy.
The setup has a state x(t) evolving by x' = f(x, u) under your control u(t), and a cost to minimize. Pontryagin introduces a companion vector of 'costate' or adjoint variables p(t) — the dynamic analogue of Lagrange multipliers — and a Hamiltonian H = p . f(x, u) - (running cost). The principle says three things hold along the optimal path: the state evolves by x' = dH/dp, the costate evolves backward by p' = -dH/dx, and crucially, at each instant the optimal control u(t) is the one that maximizes H over the allowed control set. That third condition is the 'maximum' in the name, and it is what handles the constraints that the smooth Euler-Lagrange equation cannot.
The decisive advance over classical variations is exactly this handling of bounded controls. When the best policy slams the control to a boundary (full thrust, then off — so-called bang-bang control) the Euler-Lagrange equation, which assumes you can vary freely in every direction, breaks down; the maximum principle does not, because it maximizes over the constraint set rather than setting a derivative to zero. It is the foundation of modern control engineering, aerospace trajectory optimization, robotics, and economics (where the costate is a shadow price). Two honest cautions: like Euler-Lagrange, the maximum principle is a necessary condition, not a guarantee of optimality — candidates still must be checked; and the costate equations run backward in time from terminal conditions, so the resulting two-point boundary-value problem is genuinely harder to solve than a plain initial-value problem.
To stop a frictionless cart at the origin in least time using a bounded force |u| <= 1, the maximum principle gives bang-bang control: push full one way, then full the other, switching exactly once. Euler-Lagrange cannot produce this because the optimum sits on the control boundary, not at an interior stationary point.
With bounded controls the optimum can sit on the boundary (bang-bang); the maximum principle captures it, Euler-Lagrange does not.
Despite the name, the principle is a necessary condition, not a proof of a maximum or even a minimum of cost. The 'maximum' refers only to maximizing the Hamiltonian over the control at each instant; optimality of the whole trajectory still needs separate verification.