Moscow, 1956: the blind topologist Lev Pontryagin turns to control and states the law that decides an optimal steer — the optimal control extremises a Hamiltonian pointwise, with a costate running backward. On the minimum-time double integrator it forces bang–bang: the control saturates at ±1 and switches at most once. This panel is runnable: it drives (x₀,0) to the origin, lands it in exactly T* = 2√|x₀|, and catches any “gentle interior” control as slower.
source L. S. Pontryagin, V. G. Boltyanskii, R. V. Gamkrelidze & E. F. Mishchenko, The Mathematical Theory of Optimal Processes (1961; principle announced 1956). Room: THE NUMBER. Rendered, not quoted.
Lev Semyonovich Pontryagin (1908–1988) was blinded at 14 by an exploded stove; his mother read mathematics aloud to him. He became one of the century’s great topologists (Pontryagin duality, characteristic classes), then in the 1950s turned to optimal control and, with his students, announced the Maximum Principle in 1956.
He sits in THE NUMBER among the Soviet mathematicians; the Principle ties tangentially to Kantorovich’s optimisation and the transportation problem next door.
For minimum time on x′=v, v′=u with |u|≤1, form H = 1 + pₓ·v + pₔ·u. The costate runs by p′ = −∂H/∂x: so pₓ′ = 0 (pₓ is constant) and pₔ′ = −pₓ (pₔ is affine in t).
Because H is linear in u, the optimum sits at a vertex: minimising this cost–Hamiltonian pointwise gives u* = −sign(pₔ). (Pontryagin’s original statement instead maximises H = −1 + pₓv + pₔu, with the cost entering with a minus — that is the source of the name “Maximum Principle”; the control it selects is identical.) An affine pₔ has at most one zero — hence at most one switch. That is the whole shape of the answer, read off before any trajectory is drawn.
The Principle is the necessary condition at the heart of optimal control. From the calculus of variations it passes through Pontryagin, then meets Bellman’s dynamic programming from the other side; together they give LQR, the Kalman filter’s dual, and modern trajectory optimisation.
Every rocket ascent, robot arm and reentry profile that asks “least time / least fuel” is answered in this language. Bang–bang is why a minimum-time actuator slams to a limit and holds.
The plant is the double integrator: state (x,v), dynamics x′=v, v′=u, control bounded |u|≤1. Start at rest, (x₀,0); drive to the origin (0,0) in least time.
x₀>0 needs u=−1 first (brake the fall), then u=+1; x₀<0 is the mirror. The magnitude of u is fixed by the Principle at the boundary |u|=1.
The phase plane (v across, x up). The switching curve x = −½v|v| in red; the bang–bang trajectory in blue slides along an arc until it meets the curve, switches once, and rides the curve into the origin.
u* = −sign(pₔ), |u*| = — · switch at t₁ = — · sign changes of pₔ: —
min time T* = 2√|x₀| = — · simulated landing time = — · final (x,v) = —
| control | |u| | lands? | time | vs T* |
|---|
The bang–bang law (|u|=1) is simulated against interior controls (|u|<1). Time-optimal must be the saturated one: |u*|=1 and time = T*. Any interior control is strictly slower — the witness holds only while that is true.
Set every |u|=½ and the engine still lands the state — but in 2√(2|x₀|) > T*. Gentleness costs time; the boundary is where least-time lives.
“A gentle interior control (|u|<1) is the elegant time-optimum.” Cut. H linear in u forces the max to a vertex; interior u is strictly slower.
“Optimal control needs many switches to be smooth.” Cut. pₔ is affine, so it has at most one zero — at most one switch.
“You must integrate every candidate path to find the fastest.” Cut. The Maximum Principle reads u*=−sign(pₔ) off the costate; the switching curve gives T*=2√|x₀| in closed form.
The red move: claim a smooth interior control is optimal — force the engine to use |u|=½ instead of the boundary. The state still reaches the origin, but slower than T*.
Unsaturate the control and the landing time exceeds T* while |u*|≠1: the witness “time-optimal control is bang–bang” turns red. The attack is real and it is caught.