Worst-case shape optimization
for the Dirichlet energy
Giuseppe Buttazzo
Dipartimento di Matematica
Università di Pisa
http://cvgmt.sns.it
“A Mathematical Tribute to Ennio De Giorgi”Pisa, September 19–23, 2016
Joint work
Worst-case shape optimization for the Dirich-let energy
José Carlos Bellido (Castilla La Mancha, Spain)Giuseppe Buttazzo (Pisa, Italy)Bozhidar Velichkov (Grenoble, France)
To appear on Nonlinear Analysis TMA
available at:http://cvgmt.sns.it
http://arxiv.org
4
Optimization problems:
min{F (u) : u ∈ X
}.
Problems in the calculus of variations are ofthis type, u is the state variable varying in aspace of functions and F is usually a func-tional expressed in an integral form. In thisformulation we only observe the system with-out intervening.
Example:
min{ ∫
Ωj(x, u,∇u) dx : u ∈ H10(Ω)
}.
5
Optimal control problems:
min{F (u, v) : u ∈ X, v ∈ Y, u ∈ A(v)
}.
The class A(v) is usually given through a dif-ferential equation (elliptic PDE in our case),
called state equation, v is the control vari-
able and represents the way we can operate
in the system.
Example:
minu∈H10(Ω), v∈L2(Ω)
{ ∫Ω|u−u0|2+αv2 dx : −∆u = f+v
}.
6
Let us emphasize the presence of given datain the optimal control problem, that we de-note by f . Then the problem is written as
min{F (u, v) : u ∈ A(v, f)
}.
For instance, the example above describesthe vertical displacement of a membrane fixedon ∂Ω under the action of the exterior loadf and of an extra load v that we may add.
Consider the case when the data f are onlyknown only up to some degree of uncer-tainty; nevertheless, we still want to find anoptimal solution in some sense.
7
A possibility (we do not deal with) is to as-sume that f is known with a probability P ;in this case the average cost functional canbe optimized and we are in the interestingframework of stochastic optimization.
We want on the contrary consider the worstcase for f ; more precisely, we assume thatthe data can be perturbed as f + g with‖g‖Lp ≤ δ and we optimize the worst casecost
Fwc(u, v) = sup‖g‖Lp≤δ
{F (u, v) : u ∈ A(v, f+g)
}.
8
We are interested in the case when the con-trol is the domain Ω; we are then in theframework of shape optimization problems.Other cases of worst case optimization prob-lems are considered in Allaire-Dapogny (M3AS2014). We want to show the existence of anoptimal domain for
min{F(Ω) : Ω ⊂ D, |Ω| ≤ m
}where D is a prescribed bounded subset ofRd and F is a worst-case functional given by
F(Ω) = sup{F (Ω, f + g) : ‖g‖Lp(D) ≤ δ
},
being F a given shape functional.
9
We start by considering as F the energy func-tional
E(Ω, f) = infu∈H10(Ω)
∫Ω
(1
2|∇u|2 − fu
)dx.
The worst-case functional F is:F(Ω) = sup
‖g‖Lp(D)≤δE(Ω, f + g)
= infu∈H10(Ω)
∫D
(1
2|∇u|2 − fu
)dx+ δ‖u‖
Lp′(D)
and the worst-case shape optimization prob-lem becomes
min{F(Ω) : Ω ⊂ D, |Ω| ≤ m
}.
10
Before stating the existence of an optimaldomain for the problem above let us recall avery general existence theorem based on theDal Maso-Mosco (AMO 1987) characteriza-tion of relaxed Dirichlet problems.
Theorem [Buttazzo-Dal Maso (ARMA 1993)]Let F (Ω) be such that:• F is γ-lower semicontinuous;• F is decreasing for set inclusion.Then the shape optimization problem
min{F (Ω) : |Ω| ≤ m
}admits a solution.
11
The assumptions above are verified in the
worst-case shape optimization problem, and
so we have that for every δ and m there exists
an optimal domain Ωδ,m solving
min{Fδ(Ω) : Ω ⊂ D, |Ω| ≤ m
}where
Fδ(Ω) = infu∈H10(Ω)
∫D
(1
2|∇u|2−fu
)dx+δ‖u‖
Lp′(D)
.
12
Radial case
We consider the case of a right-hand side
f of radial type; more precisely, we assume
f = f(|x|) with f(r) decreasing.
Theorem If D is large enough (to contain
a ball of measure m) the optimal domain
Ωδ,m is a ball of measure m (centered at the
origin).
13
Uncertainty only in the state equation
We consider the case of a shape optimal con-trol problem
max|Ω|≤m
∫Ωh(x)uΩ dx
where h ≥ 0 and uΩ is the solution of
−∆u = f in Ω, u ∈ H10(Ω).We assume that h is perfectly known, whilef is uncertain.
Example: best shape for the average tem-perature under partially known heat sources.
14
Writing f + g instead of f and taking the
worst-case situation, denoting by R the re-
solvent operator of −∆, we have the worstcase functional
Fδ(Ω) = sup‖g‖p≤δ
−∫DhR(f + g) dx
= sup‖g‖p≤δ
−∫D
(fR(h) + gR(h)
)dx
=∫D−f(x)wΩ dx+ δ‖wΩ‖Lp′(D)
where
−∆wΩ = h in Ω, wΩ ∈ H10(Ω).
15
Notice that Fδ is still γ-lower semicontinu-ous but it is not monotone decreasing. Then
the Buttazzo-Dal Maso theorem for the ex-
istence of an optimal shape cannot be used.
Nevertheless, the following result holds.
Theorem Assume:
• h ≥ 0 and h ∈ Ld(D);• f ∈ Lp(D) with p ≥ 2d/(d+ 2);• f ≥ c > 0 on D.Then, there exists δ̄ > 0 such that for every
0 < δ ≤ δ̄, there exists a solution Ωδ to theworst-case shape optimal control problem.
16
A numerical example
D = [0,1]× [0,1], p = 2, δ = 0.25
f =
1 on [0, 12]× [0,1]2 on [12,1]× [0,1]It is numerically convenient to simulate a do-
main Ω by a potential V (x) taking the value
0 in Ω and +∞ outside. The measure |Ω| isthen simulated through the quantity∫
De−αV (x) dx with α small.
17
More precisely this approximation has to be
stated in terms on Γ-convergence, proved in
[BGRV, JEP 2014].
The simulation has been made by J.C. Bel-
lido using:
• FreeFEM++• the Method of Moving Asymptotes (a kindof gradient method widely used for Topology
and Structural Optimization problems)
• a mesh of 50× 50 elements.
18
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
0
200
400
600
800
Optimal potential for the unperturbed case
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
0
200
400
600
800
Results for the perturbed case with δ = 0.25
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
0
2
4
6
·10−2
Optimal state for the unperturbed case
0 0.2 0.4 0.6 0.8 10
0.2
0.4
0.6
0.8
1
0
1
2
3
4
·10−2
Optimal state for the perturbed case with δ = 0.25
0
0.20.4
0.60.8
1 00.2
0.40.6
0.8
10
500
1,000
0
200
400
600
800
Optimal potential (3D view) for the unperturbed case
0
0.20.4
0.60.8
1 00.2
0.40.6
0.8
10
500
0
200
400
600
800
Optimal potential (3D view) for the case with δ = 0.25
In progress: It would be very interesting to
make an asymptotic analysis (often called Γ
development) of the sets Ωδ for δ small.
The expected result is that Ωδ is (asymptot-
ically) equal to Ω with a boundary layer Σδof local thickness δh(σ)
Σδ ={x = tν(σ), σ ∈ ∂Ω, −δh−(σ) < t < δh+(σ)
}for a suitable function h to be characterized,
with∫∂Ω h dσ = 0.
25