ECOLE POLYTECHNIQUE

Applied Mathematics Master Program

MAP 562 Optimal Design of Structures (G. Allaire)
Answers to the written exam of March 4th, 2015.

1 Parametric optimization: 14 points

1. Writing v = hu′ (h), ki, Λ = hλ′ (h), ki and differentiating problem (1)
yields

−div (h∇v) − div (k∇u) = λ ρ h v + λ ρ k u + Λ ρ h u in Ω,
v=0 on ∂Ω.

Differentiating the normalization condition Ω ρhu2 dx = 1 we obtain

R

Z
ρ u(2hv + ku) dx = 0.
Ω

2. Multiplying the above problem for v by u and integrating by parts

leads to
Z Z
λ ρ h v u + λ ρ k u2 + Λ ρ h u2 dx.


(h∇v · ∇u + k∇u · ∇u) dx =
Ω Ω

From the equation for u we know that

Z Z
h∇v · ∇u dx = λ ρ h v u dx.
Ω Ω

Therefore, from the normalization Ω ρhu2 dx = 1, we deduce

R

Z
k |∇u|2 − λ ρ u2 dx.


Λ=
Ω

In other words, we found λ′ (h) = |∇u|2 − λ ρ u2 which is a function in

L1 (Ω).

3. From the previous question and since, by assumption, ∇u 6= 0 on ∂Ω

and u = 0 on ∂Ω, we find λ′ (h) > 0 on ∂Ω (and thus by continuity on a
neighborhood of ∂Ω). Therefore, for an optimal thickness h, the lower
constraint h ≥ hmin must be saturated in this region, i.e., h = hmin
near ∂Ω. On the contrary, at the point where u is maximum, we have
∇u = 0 and u > 0, thus λ′ (h) < 0. It implies that, in the vicinity of
the maximum of u, we must have h = hmax .

1
4. The objective function is
 
u(h)
Z
J(h) = j dx.
Ω ku(h)k

It is clear that, since j is even, i.e., j(−w) = j(w), we have for any
t 6= 0    
tu(h) u(h)
j =j ,
ktu(h)k ku(h)k
so that the objective function is independent from the normalization
of the eigenfunction.
For h ∈ Uad , λ̂ ∈ R, û ∈ H01 (Ω) and p̂ ∈ H01 (Ω) we define the La-
grangian
Z  
û
Z  
L(h, λ̂, û, p̂) = j dx + h∇û · ∇p̂ − λ̂ρhûp̂ dx.
Ω kûk Ω

5. For û ∈ L2 (Ω) we define

 
û
Z
F (û) = j dx.
Ω kûk

By the chain rule lemma, since the derivativeR of kûk in the direction
of φ is hû, φi/kûk with the notation hû, φi = Ω û φ dx, we obtain
  
û φ hû, φi
Z Z
hF ′ (û), φi = F ′ (û) φ dx = j′ − û dx.
Ω Ω kûk kûk kûk3

Clearly we find hF ′ (û), ûi = 0. Equivalently,

  Z   Z
û φ û û
Z Z
′ ′ ′
F (û) φ dx = j dx− j dx û φ dx
Ω Ω kûk kûk Ω kûk kûk3 Ω

which implies
  Z   
′ 1 ′ û ′ û û
F (û) = j − j dx û .
kûk kûk Ω kûk kûk3

6. The variational formulation of the adjoint equation is, by definition,

L
h (h, λ, u, p), φi = 0 ∀ φ ∈ H01 (Ω).
∂ û
We compute
L
Z Z
′
h (h, λ, u, p), φi = F (u) φ dx + (h∇φ · ∇p − λρhφp) dx.
∂ û Ω Ω

2
 7. By integration by parts we find that the adjoint p is a solution of
(  
1 ′ u
−div (h∇p) − λ ρ h p = − kuk j kuk + α u in Ω,
p=0 on ∂Ω,

with Z   
′ u u
α= j dx
Ω kuk kuk3
We check that the right hand side is orthogonal to u since
 
u ′ u
Z Z
j dx = α u2 dx = αkuk2 .
Ω kuk kuk Ω

Clearly, if p is a solution, then p + Cu is another possible solution. To

determine the value of the constant C we use
L
Z
(h, λ, u, p) = − ρhup dx = 0
∂ λ̂ Ω
R
which implies that C = − Ω ρhup dx.

8. The derivative satisfies

∂L
hJ ′ (h), ki = h (h, λ, u, p), ki.
∂h
We compute
∂L
Z
h (h, λ, u, p), ki = (k∇u · ∇p − λρkup) dx,
∂h Ω

which implies
J ′ (h) = ∇u · ∇p − λρup.

2 Geometric optimization: 6 points

1. To define the Lagrangian we introduce two Lagrange multipliers q ∈
H 1 (RN ) and µ ∈ H 1 (RN ), which, together with v ∈ H 1 (RN ), are the
arguments of L
Z Z Z
L(Ω, v, q, µ) = j(v) dx + (∆v + f )q dx + (v − g)µ ds.
Ω Ω ∂Ω

2. To get the adjoint problem we differentiate the Lagrangian with re-

spect to v and set this partial derivative equal to 0. Before that we
perform two successive integration by parts
Z  
∂v
Z Z
L(Ω, v, q, µ) = j(v) dx+ (f q−∇q·∇v) dx+ q + (v − g)µ ds
Ω Ω ∂Ω ∂n

3
  
∂v ∂q
Z Z Z
L(Ω, v, q, µ) = j(v) dx+ (f q+∆qv) dx+ q− v + (v − g)µ ds
Ω Ω ∂Ω ∂n ∂n

For any φ ∈ H 1 (RN ) we have

∂L
Z Z
′
h (Ω, v, q, µ), φi = j (v)φ dx + ∆qφ dx (1)
∂v Ω Ω
 
∂φ ∂q
Z
+ q− φ + φµ ds
∂Ω ∂n ∂n

We first take a test function φ with compact support in Ω, so we

deduce that the optimal value of q, the adjoint p, satisfies

−∆p = j ′ (u) in Ω.
∂φ
Then we take φ = 0 on ∂Ω but with no restriction on the value of ∂n
on ∂Ω, so that
p = 0 on ∂Ω.
This yields the adjoint problem

−∆p = j ′ (u)

in Ω,
p=0 on ∂Ω.

Eventually, varying the trace of φ on ∂Ω gives the optimal value of the

Lagrange multiplier
∂p
λ= on ∂Ω.
∂n
3. Formally we know that the shape derivative is given by
∂L
J ′ (Ω)(θ) = (Ω, u, p)(θ).
∂Ω
We compute the partial derivative
∂L ∂h
Z   Z  
(Ω, v, q, µ)(θ) = j(v) + f q − ∇q · ∇v θ · n ds + + Hh θ · n ds,
∂Ω ∂Ω ∂Ω ∂n

∂v
with h = ∂n q + (v − g)µ. Taking into account p = 0 and u = g on ∂Ω,
we deduce
∂u ∂p ∂(u − g) 
Z   Z 
′
J (Ω)(θ) = j(u)−∇p·∇u θ·n ds+ +µ θ·n ds.
∂Ω ∂Ω ∂n ∂n ∂n
∂p
Since λ = ∂non ∂Ω and ∇t p = 0 on ∂Ω, it leads to

∂p ∂(u − g) 
Z 
′
J (Ω)(θ) = j(u) + θ · n ds.
∂Ω ∂n ∂n