Professional Documents
Culture Documents
J. Frédéric Bonnans
INRIA Saclay and CMAP, Ecole Polytechnique
Z T
Min `(t, yt, ẏt)dt; y0, yT given.
y 0
1
Strong solutions in the classical calculus of variations
Strong solution: in a L∞ neighborhood for y only.
The Weierstrass necessary condition (1879) is that the excess function
2
Orientation
• Various extensions
3
Sources
4
Controlled system
5
L1 expansion of trajectories
Pontryaguine linearization
6
Proof (dimension 1)
We recall that
7
Second order expansion
ζ̇t = f¯y (t)ζt + 21 f¯yy (t)(zt)2 + [fy (ut, ȳt) − f¯y (t)]zt.
8
Proof Set w := ȳ + z + ζ − y. Then
9
Expansion of cost
10
Expansion of cost
RT
+ 0 [Hy [p̄t](ut, ȳt) − Hy [p̄t](ūt, ȳt)]ztdt
+ 12 φ00(ȳ0, ȳT )(z0, zT )2 + O(ku − ūk31)
11
We rewrite the crossed term !
12
New expansion
RT
J(u) − J(ū) = 0
[H[p̄t](ut, ȳt) − H[p̄t](ūt, ȳt)]dt
1 T
+ 2 0 Hyy [p̄t](ūt, ȳt)(zt)2dt
R
RT
+ 0 [Hyu[p̄t](ūt, ȳt)(ut − ūt, zt)dt
+ 21 φ00(ȳ0, ȳT )(z0, zT )2 + O(ku − ūk32)
13
Large deviation set
p
Let u ∈ U1. We set c(u) := ku − ūk1 et
Tchebycheff inequality:
T T
|ut − ūt|
Z Z
meas(B(u)) = 1{|ut−ūt|≥c(u)}dt ≤ dt = c(u).
0 0 c(u)
14
Decomposition Principle:
Z
J(u) − J(ū0) = [H[p̄t](ut, ȳt) − H[p̄t](ūt, ȳt)]dt
B(u)
Z T
Ω[p̄](v, z) = H 00[p̄t](ūt, ȳt)(vt, zt)dt + φ00(ȳ0, ȳT )(z0, zT )2.
0
15
Necessary conditions
• Pontryaguin’s Principe
16
Analysis of sufficient conditions I
But
0
RT
φ (ȳt)δy = [p̄δy]T0 = 0 p̄t(f (u, y) − f (ū, ȳ) − fy (ū, ȳ)δy)
RT RT
= 0 δH + 0 p̄t(f (u, y) − f (u, ȳ) − fy (u, ȳ)δy
17
Analysis of sufficient conditions II
Lemma 3. Let ū be a strong solution and uk be bounded in L∞. If
lim inf k J(uk ) ≤ J(ū), and y k → ȳ uniformly, then
Z T
k
H[p̄t](ut , ȳt) − H[p̄t](ūt, ȳt) dt → 0
0
18
Analysis of sufficient conditions III
19
Constrained problems
T
Z
Min `(ut, yt)dt + φ(y0, yT );
0 (4)
(i) ẏt = f (ut, yt); t ∈ (0, T );
(ii) Φ(y0, yT ) ∈ K,
with Φ = (Φ1, Φ2) and
K = {0}RnΦ1 × RnΦ2 .
20
Optimality conditions
Costate equation
Minimum principle
21
Second order conditions I
Critical cone in Uq
(v, z) ∈ Uq × Yq ;
Cq := żt = f 0(ūt, ȳt)(vt, zt) sur (0, T ); z0 = 0. .
0
Φ (ȳT )(zT ) ∈ TK (Φ(ȳT ))
22
Second order necessary conditions II
23
Sufficient condition
24
Various extensions ?
• Stochastic control
25
Control of elliptic equations ?
−∆y + ϕ(u, y) = 0 in Ω
26
Control of elliptic equations I
27
Second order necessary optimality conditions
Z
Q(v) := H 00[p̄](ū, ȳ)(v, z)2 ≥ 0, for all v ∈ C(ū),
Ω
with z solution of linearized equation
28
Challenge
29
The End !
30