Académique Documents
Professionnel Documents
Culture Documents
Multipliers
Martin J. Gander
Mechanics
Archimedes
From Lagrangian Mechanics to Optimal Varignon
Bernoulli
Lagrange
Martin J. Gander Multiplier
martin.gander@unige.ch Constraints
Multiplier Method
Optimization
Optimal Control
University of Geneva Hamiltonian
Maximum Principle
Pontryagin
Conclusion
Lagrange
Archimedes’ Law of the Lever Multipliers
Martin J. Gander
Archimedes (287-212 BC): “Two bodies are in equilibrium
Mechanics
if their weights are inversely proportional to their arm length” Archimedes
Varignon
Bernoulli
Bernoulli’s Rule
Lever
Double Lever
Varignon example
Lagrange
Multiplier
Constraints
Multiplier Method
Optimization
Optimal Control
Hamiltonian
Maximum Principle
Pontryagin
Adjoint
PDE Constraint
Optimization
Lions
Adjoint
Conclusion
Lagrange
Archimedes’ Law of the Lever Multipliers
Martin J. Gander
Archimedes (287-212 BC): “Two bodies are in equilibrium
Mechanics
if their weights are inversely proportional to their arm length” Archimedes
Varignon
Bernoulli
Bernoulli’s Rule
Lever
Double Lever
Varignon example
Lagrange
Multiplier
Constraints
Multiplier Method
Optimization
Optimal Control
Conclusion
2) Redistribute weights to reach symmetric configuration Lagrange
Multipliers
Martin J. Gander
Mechanics
Archimedes
Varignon
Bernoulli
Bernoulli’s Rule
Lever
Double Lever
Varignon example
Lagrange
Multiplier
Constraints
Multiplier Method
Optimization
Optimal Control
Hamiltonian
Maximum Principle
Pontryagin
Adjoint
PDE Constraint
Optimization
Lions
Adjoint
Conclusion
Lagrange
Pierre Varignon Multipliers
Bernoulli’s Rule
Hundreds of results illustrated in 64 plates of figures: Lever
Double Lever
Varignon example
Lagrange
Multiplier
Constraints
Multiplier Method
Optimization
Optimal Control
Hamiltonian
Maximum Principle
Pontryagin
Adjoint
PDE Constraint
Optimization
Lions
Adjoint
Conclusion
Lagrange
Pierre Varignon Multipliers
Martin J. Gander
Mechanics
Archimedes
Varignon
Bernoulli
Bernoulli’s Rule
Lever
Double Lever
Varignon example
Lagrange
Multiplier
Constraints
Multiplier Method
Optimization
Optimal Control
Hamiltonian
Maximum Principle
Pontryagin
Adjoint
PDE Constraint
Optimization
Lions
Adjoint
Conclusion
Lagrange
Johann Bernoulli Multipliers
to Varignon: Bernoulli
Bernoulli’s Rule
“Votre projet d’une nouvelle mechanique fourmille d’un Lever
Double Lever
grand nombre d’exemples, dont quelques uns à en juger par Varignon example
Martin J. Gander
Mechanics
Archimedes
Varignon
Bernoulli
Bernoulli’s Rule
Lever
Double Lever
Varignon example
Lagrange
Multiplier
Constraints
Multiplier Method
Optimization
Optimal Control
Hamiltonian
Maximum Principle
Pontryagin
“In every equilibrium of arbitrary forces, no matter how they
Adjoint
are applied, and in which directions they act the ones on the PDE Constraint
Optimization
others, either indirectly or directly, the sum of the positive Lions
Adjoint
energies will be equal to the sum of the negative energies Conclusion
taken positively.”
Varignon gives however the wrong date 1717 for the letter of Bernoulli,
which was later copied by Lagrange.
Lagrange
Lagrange Explains Bernoulli’s Rule Multipliers
Bernoulli’s Rule
Lever
Double Lever
Varignon example
Lagrange
Multiplier
Constraints
Multiplier Method
Optimization
Optimal Control
Assume the lever is in equilibrium: Archimedes implies Hamiltonian
Maximum Principle
Pontryagin
P b
Pa = Qb =⇒ = Adjoint
Q a PDE Constraint
Optimization
Lions
Bernoulli’s idea: apply a “virtual velocity” during an Adjoint
Conclusion
infinitely small time interval to displace the lever. Then
P dq
a ∝ dp and b ∝ dq =⇒ =− =⇒ Pdp + Qdq = 0
Q dp
Lagrange
So Forget Archimedes ! Multipliers
Mechanics
Archimedes
Varignon
Bernoulli
Bernoulli’s Rule
Lever
Double Lever
Varignon example
Lagrange
Multiplier
Constraints
Multiplier Method
we get Optimization
Optimal Control
Pdp + Qdq = 0. Hamiltonian
Maximum Principle
Pontryagin
Now if dp = dx for an arbitrary dx, then dq = − ba dx, and Adjoint
hence PDE Constraint
Optimization
Lions
b Adjoint
Martin J. Gander
Mechanics
Archimedes
Varignon
Bernoulli
Bernoulli’s Rule
Lever
Double Lever
Varignon example
Lagrange
Multiplier
Constraints
Multiplier Method
Optimization
Applying the rule of Bernoulli, we get Optimal Control
Hamiltonian
Maximum Principle
Pontryagin
Pdp + Qdq + Rdr = 0 (∗) Adjoint
PDE Constraint
Martin J. Gander
To use Mechanics
Archimedes
Varignon
Bernoulli
Pdp + Qdq + Rdr = 0 (∗)
Bernoulli’s Rule
Lever
q Lagrange
Multiplier
p = (x −a)2 + (y −b)2 + (z −c)2 Constraints
Multiplier Method
Optimization
Optimal Control
1 Hamiltonian
dp = ·((x−a)dx+(y −b)dy +(z−c)dz) Maximum Principle
p Pontryagin
Adjoint
and similarly q, dq, and r , dr . PDE Constraint
Optimization
Inserted into (∗) we get for equilibrium Lions
Adjoint
y −b y −g y −m
with X = P x−a x−f x−l
p +Q q +R r , Y =P p +Q q +R r
and Z = P z−c z−h z−n
p +Q q +R r .
Lagrange
Same System with Constraint Multipliers
Martin J. Gander
Mechanics
Motion (dx, dy , dz) Archimedes
Varignon
Bernoulli’s Rule
∂L ∂L ∂L Lever
Multiply (1) by λ = −Z / ∂L
∂z and add Adjoint
X + λ ∂L ∂L
Z + λ ∂L
∂x · dx + Y + λ ∂y · dy = 0 , ∂z = 0 . Conclusion
Martin J. Gander
Physical interpretation of λ by Lagrange:
Mechanics
λ( ∂L ∂L ∂L
∂x , ∂y , ∂z ) represents the force that holds the particle on Archimedes
Varignon
the surface L = 0 Bernoulli
Bernoulli’s Rule
Lever
Double Lever
“Équation générale” for ALL problems of equilibria: Varignon example
Lagrange
Multiplier
Constraints
Multiplier Method
Optimization
Optimal Control
◮ “Méthode très-simple” in Section IV of the first edition Hamiltonian
Maximum Principle
from 1788. Pontryagin
Adjoint
◮ In the second edition from 1811, Lagrange baptizes the PDE Constraint
Optimization
method Lions
Adjoint
Conclusion
Lagrange
Lagrange Multiplier in Optimization Multipliers
Lagrange
Multiplier
f (x) −→ max, g(x) = 0, Constraints
Multiplier Method
Optimization
with f : Rn → R and g : Rn → Rm , m < n. Optimal Control
Hamiltonian
Maximum Principle
Elimination of the constraints: Pontryagin
Adjoint
x = (y, u), y ∈ Rm , u ∈ Rn−m PDE Constraint
Optimization
Lions
Adjoint
g(x) = 0 and implicit function theorem =⇒ y = y(u)
Conclusion
f (y(u), u) −→ max .
Lagrange
Lagrange Multiplier in Optimization Multipliers
Martin J. Gander
Necessary condition for a local maximum of f (y(u), u):
Mechanics
Archimedes
df ∂f ∂y ∂f Varignon
= + =0 Bernoulli
du ∂y ∂u ∂u Bernoulli’s Rule
Lever
Double Lever
∂y
For ∂u , implicitly differentiate constraint g(y(u), u) = 0 Varignon example
Lagrange
Multiplier
∂f ∂y ∂f Constraints
+ = 0, n − m equations Multiplier Method
∂y ∂u ∂u Optimization
Optimal Control
∂g ∂y ∂g Hamiltonian
+ = 0, m(n − m) equations Maximum Principle
∂y ∂u ∂u Pontryagin
Adjoint
g = 0, m equations PDE Constraint
Optimization
Lions
n + m(n − m) equations for the n unknowns in y and u Adjoint
∂y Conclusion
combined, and the m(n − m) unknowns in the Jacobian ∂u
∂y
=⇒ A very big system because of ∂u .
Lagrange
Gaussian Elimination of Lagrange Multipliers
Martin J. Gander
∂f ∂y ∂f Mechanics
+ = 0, n − m equations Archimedes
∂y ∂u ∂u Varignon
Bernoulli
∂g ∂y ∂g
+ = 0, m(n − m) equations Bernoulli’s Rule
∂y ∂u ∂u Lever
Double Lever
g = 0, m equations Varignon example
−1 Lagrange
∂g Multiplier
Multiply 2nd equation by λ := − ∂f
∂y ∂y and add to 1st: Constraints
Multiplier Method
Optimization
∂f ∂g Optimal Control
+λ = 0, n − m equations Hamiltonian
∂u ∂u Maximum Principle
Pontryagin
∂f ∂g
+λ = 0, m equations Adjoint
∂y ∂y PDE Constraint
Optimization
Lions
g = 0, m equations Adjoint
Conclusion
for n unknowns in y and u combined, plus m in λ.
Lagrange: get this directly by differentiating
L(u, y, λ) := f (y, u) + λg (y, u)
Lagrange
Hestenes’ Optimal Control Problem 1950 Multipliers
Martin J. Gander
Mechanics
Archimedes
Varignon
Bernoulli
Bernoulli’s Rule
Lever
Z T Double Lever
Varignon example
f (y, u)dt −→ max, Lagrange
0 Multiplier
ẏ = g(y, u), Constraints
Multiplier Method
y(0) = y0 , Optimization
Optimal Control
Hamiltonian
y(T ) = yT . Maximum Principle
Pontryagin
0 0 Conclusion
Z T Z T Mechanics
Archimedes
L(y, u, λ) := f (y, u)dt + λ (g(y, u) − ẏ) dt, Varignon
Bernoulli
0 0
Bernoulli’s Rule
we obtain Lever
Double Lever
Varignon example
d
RT ∂f
RT ∂g
dε L(y + εz, u, λ)|ε=0 = 0 zdt + λ z − ż dt Lagrange
0 = +λ Lions
∂u ∂u Adjoint
Conclusion
ẏ = g(y, u) y(0) = y0 , y(T ) = yT
∂f ∂g
−λ̇ = +λ
∂y ∂y
Lagrange
Hamiltonian Structure Multipliers
Mechanics
H(y, u, λ) := f (y, u) + λg(y, u), Archimedes
Varignon
Bernoulli
trajectories! Pontryagin
Adjoint
PDE Constraint
Optimization
This was already discovered by Carathéodory 1926: Lions
Adjoint
Conclusion
Lagrange
Necessary Condition from the Hamiltonian Multipliers
Z T Z T Mechanics
Bernoulli’s Rule
which gives along an optimal trajectory ẏ = g(y, u) Lever
Double Lever
Z T Varignon example
Bernoulli’s Rule
Lever
Double Lever
Varignon example
Lagrange
Multiplier
Constraints
Multiplier Method
Optimization
Optimal Control
Hamiltonian
Maximum Principle
Pontryagin
Adjoint
PDE Constraint
Optimization
Lions
Adjoint
Conclusion
Lagrange
Discovery of Pontryagin Multipliers
Bernoulli’s Rule
Lever
Double Lever
Varignon example
Lagrange
Multiplier
Constraints
Multiplier Method
Optimization
Optimal Control
Hamiltonian
Maximum Principle
Pontryagin
Adjoint
PDE Constraint
Optimization
Lions
Adjoint
Conclusion
Lagrange
Discovery of Pontryagin Multipliers
Bernoulli’s Rule
◮ ÿ = ±M (Feldbaum 1949) Lever
◮ ÿ = ±u, |u| ≤ M (Feldbaum 1953) Double Lever
Varignon example
Lagrange
Multiplier
Constraints
Multiplier Method
Optimization
Optimal Control
Hamiltonian
Maximum Principle
Pontryagin
Adjoint
PDE Constraint
Optimization
Lions
Adjoint
Conclusion
Lagrange
Discovery of Pontryagin Multipliers
Bernoulli’s Rule
◮ ÿ = ±M (Feldbaum 1949) Lever
◮ ÿ = ±u, |u| ≤ M (Feldbaum 1953) Double Lever
Varignon example
◮ Pontryagin discovers the Hamiltonian formulation Lagrange
without Lagrange multipliers Multiplier
Constraints
Multiplier Method
Optimization
Optimal Control
Hamiltonian
Maximum Principle
(Gamkrelidze (1999) “Discovery of the Maximum Principle”) Pontryagin
Adjoint
PDE Constraint
Optimization
Lions
Adjoint
Conclusion
Lagrange
Discovery of Pontryagin Multipliers
Bernoulli’s Rule
◮ ÿ = ±M (Feldbaum 1949) Lever
◮ ÿ = ±u, |u| ≤ M (Feldbaum 1953) Double Lever
Varignon example
◮ Pontryagin discovers the Hamiltonian formulation Lagrange
without Lagrange multipliers Multiplier
Constraints
Multiplier Method
Optimization
Optimal Control
Hamiltonian
Maximum Principle
(Gamkrelidze (1999) “Discovery of the Maximum Principle”) Pontryagin
Adjoint
PDE Constraint
Optimization
Lions
Adjoint
Conclusion
Lagrange
Historical Discovery of the Maxmimum Principle Multipliers
Martin J. Gander
Boltyanski, Gamkrelidze and Pontryagin (1956):
Mechanics
“This fact appears in many cases as a general principle, which we Archimedes
call the maximum principle” Varignon
Bernoulli
Bernoulli’s Rule
Lever
Double Lever
Varignon example
Lagrange
Multiplier
Constraints
Multiplier Method
Optimization
Optimal Control
Hamiltonian
Maximum Principle
Pontryagin
Adjoint
PDE Constraint
Optimization
Lions
Adjoint
Conclusion
Pontryagin (1959): “ In the case that Ω is an open set [. . . ],
the variational problem formulated here turns out to be a special
case of the problem of Lagrange.”
Lagrange
Early Work with PDE Constraints Multipliers
Martin J. Gander
Egorov (1962,1963): Some problems in the theory of
optimal control, Optimal control in Banach spaces Mechanics
Archimedes
Varignon
Minimum time control problem for the parabolic equation Bernoulli
Bernoulli’s Rule
∂y
∂t + Ay + b(u)y = f +u on Ω × (0, T ) Lever
Double Lever
Lagrange
Multiplier
with initial condition y (t0 ; u) = y0 and target y (t1 ; u) = yT . Constraints
Multiplier Method
Optimization
Optimal Control
Hamiltonian
Maximum Principle
Pontryagin
Adjoint
PDE Constraint
Optimization
Lions
Adjoint
Conclusion
Lagrange
Early Work with PDE Constraints Multipliers
Martin J. Gander
Egorov (1962,1963): Some problems in the theory of
optimal control, Optimal control in Banach spaces Mechanics
Archimedes
Varignon
Minimum time control problem for the parabolic equation Bernoulli
Bernoulli’s Rule
∂y
∂t + Ay + b(u)y = f +u on Ω × (0, T ) Lever
Double Lever
Lagrange
Multiplier
with initial condition y (t0 ; u) = y0 and target y (t1 ; u) = yT . Constraints
Multiplier Method
Optimization
J.-L. Lions: “Le travail de Yu. V. Egorov contient une étude Optimal Control
Hamiltonian
détaillée de ce problème, mais nous n’avons pas pu comprendre Maximum Principle
Pontryagin
tous les points des démonstrations de cet auteur, les résultats Adjoint
étant très probablement tous corrects.” PDE Constraint
Optimization
Lions
Adjoint
Conclusion
Lagrange
Early Work with PDE Constraints Multipliers
Martin J. Gander
Egorov (1962,1963): Some problems in the theory of
optimal control, Optimal control in Banach spaces Mechanics
Archimedes
Varignon
Minimum time control problem for the parabolic equation Bernoulli
Bernoulli’s Rule
∂y
∂t + Ay + b(u)y = f +u on Ω × (0, T ) Lever
Double Lever
Lagrange
Multiplier
with initial condition y (t0 ; u) = y0 and target y (t1 ; u) = yT . Constraints
Multiplier Method
Optimization
J.-L. Lions: “Le travail de Yu. V. Egorov contient une étude Optimal Control
Hamiltonian
détaillée de ce problème, mais nous n’avons pas pu comprendre Maximum Principle
Pontryagin
tous les points des démonstrations de cet auteur, les résultats Adjoint
étant très probablement tous corrects.” PDE Constraint
Optimization
Lions
Fattorini (1964): Time-optimal control of solutions of Adjoint
Conclusion
operational differential equations (proof of the “bang-bang”
property, no maximum principle)
Friedman (1967): Optimal control for parabolic equations
Lagrange
Jacques-Louis Lions Multipliers
Martin J. Gander
R. M. Temam (SIAM News, July 10, 2001):
Mechanics
“A new adventure began for Lions in the early Archimedes
Varignon
1960s, when he met (in spirit) another of his Bernoulli
Bernoulli’s Rule
intellectual mentors, John von Neumann. By then, Lever
using computers built from his early designs, von Double Lever
Varignon example
Neumann was developing numerical methods for Lagrange
Multiplier
the solution of PDEs from fluid mechanics and Constraints
Lagrange
closed convex subset U, and PDE constraint Multiplier
Constraints
Martin J. Gander
Defining p(v ) ∈ V by A∗ p(v ) = C ∗ Λ(Cy (v ) − zd ), we get
Mechanics
Archimedes
(C ∗ Λ(Cy (u) − zd ), y (v ) − y (u))V + (Nu, v − u)U Varignon
Bernoulli
= (Λ−1 ∗
U B p(u) + Nu, v − u)U ≥ 0 Constraints
Multiplier Method
Optimization
Optimal Control
where B ∗ : V → U ′ is the adjoint of B, ΛU : U → U ′ is the Hamiltonian
Maximum Principle
canonical isomorphism from U to U ′ . Pontryagin
Adjoint
Hence the adjoint p(v ) permits elimination of y (u), and PDE Constraint
Optimization
Lions
(Λ−1 ∗ −1 ∗
U B p(u) + Nu, u)U = inf (ΛU B p(u) + Nu, v )U
Adjoint
v ∈Uad Conclusion
Martin J. Gander
Bernoulli’s Rule
analyzed using “virtual velocities” Lever
Double Lever
Varignon example
2. The Lagrange multiplier is just a multiplier from
Lagrange
Gaussian elimination Multiplier
Constraints
Multiplier Method
3. Using Lagrange multipliers, one can find the adjoint Optimization
Optimal Control
equation in optimal control Hamiltonian
Maximum Principle
Pontryagin
Conclusion
“Constrained Optimization: from Lagrangian Mechanics to
Optimal Control and PDE Constraints”, Gander, Kwok,
Wanner, 2014.