Composite Membranes

Aldo Lopes
UFMG, [email protected]
Russell E. Howes, Brigham Young University
[email protected]
Cynthia Shepherd, California State University - Northridge
[email protected]
8th August 2006

1
Contents
1 Introduction 3

2 The Dirichlet Problem 4

2.1 Spectral Theorem . . . . . . . . . . . . . . . . . . . . . . . . . 6

3 Variations 7

4 Variations With Fixed Boundary 10

5 Critical Eigenvalues 12

2
1 Introduction
The study of vibrating membranes is an aspect of mathematical physics that
is primarily concerned with the geometries of the membrane corresponding to
its frequencies. Membranes consisting of two materials of different densities
are called composite membranes.

We are interested in determining the shape of membranes with minimal

eigenvalues; in particular the first eigenvalue, or fundamental frequency, of
such membranes.

In this paper1 we will be studying a membrane Ω divided into two subsets,

a lower density D and a higher density Dc . The boundary between D and
Dc is a hypersurface and will be called Γ. We will first look at the associated
Dirichlet boundary PDE problem.

We will review the Spectral Theorem, which is an important result re-

lated to our work and then consider variations that fix the boundary of Ω
and preserves the volume of D and show that such variations do exist under
our given specifications.

The main purpose of this paper is to determine conditions on D and its

corresponding first eigenfunction, u, for the critical points of the smallest
eigenvalue, λ. We will also prove that for any critical point of λ, Γ is a level
set of u.

1
Partially supported by NSF grant OISE: 0526008.

3
2 The Dirichlet Problem
The composite membrane eigenvalue problem for a bounded domain Ω ⊂ Rn
is the Dirichlet boundary PDE problem:

−∆u + αχD u = λu in Ω
(∗)Ω,α,D
u = 0 in ∂Ω
u is a normalised eigenfunction of the above problem, D is the region
of Ω previously defined, and we are interested in minimising λ0 , the lowest
eigenvalue or lowest possible value of λ for a given D. We will only worry
about the eigenfunction u associated with λ0 , and we want to find what
configurations of u and D will make λ00 = 0.
χD is the characteristic function:

0 if x ∈ Dc (= Ω − D)
χD (x) =
1 if x ∈ D

Let Ω ⊂ Rn be an open, connected and bounded set. We define N as the

boundary of Ω, i.e., ∂Ω = N . Here N is smooth and a C ∞ hypersurface. In
fact, N = f −1 (a), where a is a regular value of the function f : U → R with
N ⊂ U.
F : (−, ) × Ω̄ → Rn a C ∞ function is the variation of Ω. If is necessary,
we can consider the extension of F : F̄ : (−, ) × U → Rn a C ∞ function
and Ω̄ ⊂ U ⊂ Rn is an open set.
We define now the follwing function that for each time:

Ft = F (t, ·) : Ω̄ → Rn

Ft (Ω̄) = Ω̄t
Restricting the time if necessary, one always has that Ft : Ω̄ → Rn is a
diffeomorphism onto its image:

Ft : Ω̄ → Ω̄t

For each time we consider N = ∂Ω and Nt = ∂ Ω̄t also a C ∞ hypersurface.

And Γ = ∂D ∩ Ω.
Ft : N → Nt is also a diffeomorphism.
Let’s start by defining some functions. We previously defined F : (−, )×
Ω̄ −→ Ω̄.

Definition 1. For some fixed t ∈ (−, ), we define Ft : Ω̄ −→ Ω̄ as Ft =

F (t, x).

4
 Figure 1: The membrane

Let’s place some constraints on our Ft : For any t1 , t2 , (t1 + t2 ) ∈ (−, ):

Ft1 ◦ Ft2 = Ft1 +t2

This structure makes things nice for us–it makes the Ft s associative, and
gives us for each Ft an inverse, so that Ft ◦ F−t = F0 = F . We will refer to
F0 as F from here on out. Defining Ft leads to a natural definition of Dt and
thus χDt , and prompts us to redefine the first eigenfunction:

Definition 2. For some fixed t ∈ (−, ), we define ut : Ω̄ −→ R as ut =

u(t, x). We will newly define u = u0 .

A few observations: We want to rewrite χDt as a deformation of χD . We

notice that if x = Ft (y), y = F−t (x) = Ft−1 (x) and

χD (y) = χDt ◦ Ft (y)

χD ◦ (Ft )−1 (x) = χDt (x) = χD ◦ F−t

5
2.1 Spectral Theorem
Usually one also has a boundary condition coming from the original problem.
We will consider the case where the membrane is fixed along ∂Ω,i.e., Ft (x) =
0 for all x ∈ ∂Ω and all t ∈ R. This gives rise to the same condition for u,
the Dirichlet boundary condition. Thus, one is bound to study the possible
solutions of the PDE problem

−∆u = λu in Ω
u = 0 in ∂Ω
Analogous to the linear algbra concepts, a nontrivial u satisfyng such
conditions will be called an eigenfunction of the Laplacian and the associated
real number λ an eigenvalue.
Recall that given two functions φ and ψ in Ω wich go to zero at its
boundary, their L2 - inner product is given by
Z
hφ, ψi = ϕψdx dy
Ω

One then has the classical

Theorem 2.1. (Spectral Theorem for the Laplacian with the Dirich-
let boundary condition). There is an unbounded discrete set of (positive)
eigenvalues
(0 <)λ1 < λ2 ≤ λ3 ≤ . . .
and associated set of eigenfunctions ui , which may be taken mutually orthog-
onal and with norm equal to 1, with respect to the L2 inner product in Ω,
which is complete in the sense that any smooth function v : Ω → R such that
v = 0 at ∂Ω may be written as the Fourier series
∞
X
v= hv, ui iui
i=1

Remark. In the case of one space variable, i.e., of a vibrating string, the
u0i sare given by sine functions, the original Fourier series situation.
Observe that the (orthonormal) system of eigenfunctions plays the usual
role of an orthonormal basis in finite dimensional linear algebra, when we use
orthogonal projection to write a vector in that basis (in that case one has a
finite sum of projected vectors). It is also called a Hilbert basis for the space
os functions in Ω (with the Dirichlet boundary condition).
The numbers λi give the frequency of the basic solutions (called harmon-
ics) for the vibrating membrane. For this reason, λ1 , the first eigenvalue, is
also called the fundamental frequency or pitch of the membrane

6
3 Variations
We say that F preserves the volume if

vol(Dt ) = vol(D) ∀ t ∈ (−, )

Since Γ bounds D, there exists ν : Γ → Rn , such that:

1. kν(p)k = 1

2. ν points to exterior of D

3. ν(p) ⊥ Tp Γ

Where Tp Γ is the tangent plane at p to Γ. ν is the exterior unit normal

vector field along Γ.
Now, let F : (−, ) × Ω̄ and p ∈ Ω̄. We will denote by

∂F
X(p) = (0, p) ∈ Rn
∂t
the variation vector of F . X(p) is the initial velocity of variation at points
of Ω̄. X : N → Rn is a C ∞ vector field on N .
We have a formula for the volume of Dt :
Z
v(t) = dx
Dt

Now we have the following important lemma.

Lemma 3.1. Z Z
0
v (0) = divX dx = hX, νi dx
D Γ

Proof. We will show two different proofs.

(1)
As in the formula above, we have:
Z  Z Z
0 d
v (t) = 0
dx = (χDt ) dx = (χDt ◦ F−t )0 dx
dt Dt Ω Ω

The function χD is not smooth neither continuous. Then we consider a

smooth function gδ : Γ → R. dgδ : Rn → R is a linear application. Also we
have:
lim gδ = χD
δ→0

7
 Now we can compute this integral
Z
(gδ ◦ F−t )0 dx
Ω

Remember that X = ∂F ∂t
is the variation vector. Then we have:
Z  Z
∂  ∂F
gδ (F (−t, x)) dx = dgδ (x) (0, x).(−1) dx
Ω ∂t t=0 ∂t

Ω
Z Z
= − h∇gδ (x), X(x)i dx = − X(gδ )(x) dx
Ω Ω
But we have
div(gδ X) = h∇gδ , Xi + gδ divX
And Z Z
div(gδ X) = gδ hX, νi = 0
Ω ∂Ω
because X is zero along the boundary ∂Ω. As we have limδ→0 gδ = χD ,
then the integral converges too. It means
Z Z Z
0
(χD ◦ F−t ) dx = (χD )divXdx = hX, νi dS
Ω Ω Γ

(2)
Using change of variables we have:
Z
v(t) = JFt dy
D

Where JFt = |det(dFt )|. Then

Z Z Z
0 d d d
v (t) = dy = JFt dx = JFt dx
dt Dt dt D D dt

Now, we will work only with dtd JFt . dFt is the Jacobian matrix for the
function Ft .
Ft : Ω ⊂ Rn → Rn
Ft (x1 , x2 , . . . , xn ) = (Ft1 x, Ft2 x, . . . , Ftn x)
Where x = (x1 , x2 , . . . , xn ) and t ∈ (−ε, ε).
 
(Ft1 )1 (Ft1 )2 · · · (Ft1 )n
 (F 2 )1 (Ft2 )2 · · · (Ft2 )n 
t
dFt = 
 
.. .. .. 
 . . . 
(Ftn )1 n n
(Ft )2 · · · (Ft )n

8
 ∂F i
Where (Fti )j = ∂xjt and i, j = 1, . . . , n.
If we look at the determinant of dFt by its columns, we get:
det(dFt ) = det[(Ft )1 (Ft )2 · · · (Ft )n ]
Now, we can differentiate with respect to t:
d
det(dFt ) = det[(Ft )01 (Ft )2 · · · (Ft )n ] + · · · + det[(Ft )1 (Ft )2 · · · (Ft )0n ]
dt
Where (Ft )0k is the derivative of the k-th column with respect to t. We
can take a look at one of the terms of the sum above:
 
(Ft1 )1 · · · (Ft1 )0k · · · (Ft1 )n
 (F 2 )1 · · · (F 2 )0 · · · (F 2 )n 
1 0 t t k t
det[(Ft )1 · · · (Ft )k · · · (Ft )n ] = det 
 
.. .. .. 
 . . . 
n n 0 n
(Ft )1 (Ft )k · · · (Ft )n

We can compute the determinant of the matrix above as a sum as below,

where P is the set of all permutations of the index i and σ is its permutation.
X
εσ (Ft1 )σ(1) (Ft2 )σ(2) · · · (Ftk )0σ(k) · · · (Ftn )σ(n)
σ∈P

Where εσ = ±1 is the signal of the permutation.We need this computation

only when t = 0. In this time we have for i, j = 1, ..., n:

(F0i )j = 0 i 6= j
i
(F0 )i = 1
This means that F0 = Id|Ω̄ . So, it is clear that:
d
det dFt |t=o = (F01 )0 + (F02 )0 + · · · + (F0n )0
dt
It is easy to see that
  1
dF0 dF02 dF0n
 
∂ ∂ ∂
, ,··· , · , ,··· , = ∇ · F0
∂x1 ∂x2 ∂xn dt dt dt
The end of this proof it is made by the Gauss-Green Theorem. If we
called F0i = Xi for i = 1, ..., n and div X = X10 + ... + Xn0 :
Z Z
0
v (0) = div X dx = hX, νi dS
D Γ

9
4 Variations With Fixed Boundary
Lemma 4.1. Let f : Γ → R be a piecewise smooth function such that
Z
f dS = 0
Γ

then there exists a volume-preserving normal variation whose variation vector

is f N . If, in addition, f ≡ 0 on ∂Ω, the variation can be chosen so that it
fixes the boundary ∂Ω.
R
Proof. Let g = 0 on ∂Ω and D gdM 6= 0. We can let g = 1 on Γ. Now, in
order to extend tf + t̄g over all of Ω̄, Take any point q ∈ Ω̄. We will define
d(Γ, q) to be the length of the normal vector to Γ passing through q (touching
Γ at point p; thus we also define a function p(q) mapping Ω̄ to Γ.) We also
define d(p, ∂Ω) for any point p ∈ Γ as the length of the vector normal to Γ
at p whose endpoint is on ∂Ω (we will say that the length is negative if q is
inside Γ.) We will also define the ”bump” function,

B : [0, 1] → R

so that B(x) = 1 if x ≤ 31 , B(x) = 0 if x ≥ 32 , and B(x) is C ∞ on [0, 1].

We then define our extension function,
 
d(Γ, q)
φ(q) = B
d(p(q), ∂Ω)
It is easily shown that φ(q) is uniformly 1 on Γ, 0 on ∂Ω, and C ∞ on Ω̄.
Now, set x(t, t̄) = x0 + φ(q)(tf + t̄g)N , g : Ω̄ → R is a piecewise smooth
function.
The volume function is
Z
VD (t, t̄) = hx, N i dS
Γ

Since we are considering the case where the volume is constant, we have

VD (t, t̄) = VD (0, 0)

Then we have  Z
∂VD 
= dM 6= 0
∂ t̄ (0,0) D

Now we want to apply the Implicit Function Theorem.

10
 We have VD (t, t̄) = VD (0, 0) so, let

G:R×R→R

G := VD (t, t̄) − VD (0, 0)

continuously differentiable on an interval containing (0, 0) and G(0, 0) = 0.
The Jacobian is  
∂VD ∂VD
dVD =
∂t ∂ t̄
 ∂V 
So, our M = ∂ t̄D
 Z
∂VD 
det(M ) |t̄=0 = = gdM 6= 0 by definition.
∂ t̄ t̄=0 D

Then, by the Implicit Function Theorem, there is an open set (−ε, ε) ⊂ R

containing t̄ = 0 such that ∀t(−ε, ε) ∃! ϕ(t)(−ε, ε) such that ϕ(t) = t̄ and
G(t, ϕ(t)) = 0 ⇒ V (t, ϕ(t)) = VD (0, 0).

11
5 Critical Eigenvalues
Theorem 5.1. Let D ⊂ Ω ⊂ Rn be bounded, measurable and with a smooth
boundary Γ,u,f as defined above. Then
Z
0
λ (0) = α u2 f dS
Γ

Proof. Observe that

−∆u0 + αχD u0 = λu0

−∆ut + αχDt ut = λut

We multiply the first equation by −ut and the second by u, then add
them and integrate both side over Ω̄ to obtain

Z Z Z
ut ∆u0 − u0 ∆ut + α (χDt − χD ) u0 ut = (λ(t) − λ) u 0 ut
Ω Ω Ω

The first integrand on the left-hand side of the equation is equal to the
divergence of u∇ut − ut ∇u. By the divergence theorem, this integral has
value 0 since both u and ut are uniformly 0 on ∂Ω. Our equation reduces to
Z Z
α (χDt − χD ) u0 ut = (λ(t) − λ) u0 ut
Ω Ω
We take the derivative of both sides with respect to t, and then terms
cancel out when we evaluate at t = 0:
Z Z Z 
d 0

0 0
α χDt u0 ut + χDt u0 ut − χD u0 ut =
Ω dt Ω Ω
Z Z Z
0 0 0
= λ (t) u0 ut + λ(t) u0 ut − λ0 u0 u0t
Ω Ω Ω
Evaluate at t = 0:
Z Z
d 0
χDt t=0 u = λ (0) u2 = λ0 (0)
2 0


α
Ω dt Ω

since the L2 -norm of u is 1. The question remains, how can we take the
derivative of a non-differentiable function? Well, we can approximate χD by
a C ∞ function, g , by defining g to be equal to χD on all points p ∈ Ω where
d(p, Γ) ≥  and C ∞ on all points (on the ring around Γ, g slopes from 0 to

12
1). Clearly as  → 0, g → χD , but g is still a C ∞ function which we can
differentiate.

Z Z Z
d d d 2
(χDt )t=0 u2 = α (χD ◦ F−t )t=0 u2 ≈ α


α u g ◦ F−t t=0
Ω dt Ω dt Ω dt
Z Z
2
= −α u X (g ) = −α hu2 X, ∇g i
Ω Ω
where X is the vector field associated with F . Now, by the properties of
divergence,

div(g ◦ u2 X) = hu2 X, ∇g i + g div(u2 X)

Z Z Z
2 2
g div(u X) = div(g ◦ u X) − hu2 X, ∇g i
Ω Ω Ω
Z Z Z
= g u2 hX, µi − hu2 X, ∇g i = − hu2 X, ∇g i
∂Ω Ω Ω
since u ≡ 0 on ∂Ω. Thus
Z Z
0
λ (0) = −α hu X, ∇g i = α g div(u2 X)
2
Ω Ω
Z Z Z
2 2
≈α χD u divX = α u divX = α u2 hX, νi
Ω D ∂D

by the Divergence Theorem. We previously stated that f = hX, νi. ∂D

is made up of two parts, Γ and ∂Ω. On ∂Ω, u ≡ 0, so the previous result
reduces to
Z
0
λ (0) = α u2 f dS
Γ

which is a nice simple expression for λ0 (0).

R
Definition 3. (Ω, D) is a critical point if for all variations f where Γ
f dS =
0 we have λ0 (0) = 0.

Proposition 5.2. (Ω, D) is a critical point for our problem if and only if
Γ = u−1 (c) where c is a constant.

13
Proof. (⇒) Assume, in order to prove the contra-positive, that u is not con-
stant over Γ. ⇒ u2 is not constant over Γ. Then define
R 2
2 u dS
ū = ΓR
Γ
dS

Now define positive, smooth functions φ, ψ such that

φ = 0 when u2 − ū2 ≤ 0 and,

ψ = 0 when u2 − ū2 ≥ 0
Then let f = (φ + ψ)(u2 − ū2 )
Z
⇒ (φ + ψ)(u2 − ū2 ) = 0
Γ
R
Consider Γ u2 f dS
Z Z Z Z
u f dS = u (φ+ψ)(u − ū ) = (φ+ψ)(u − ū ) + ū2 (φ+ψ)(u2 − ū2 )
2 2 2 2 2 2 2
Γ Γ Γ Γ

Then since ū2 is constant, we can factor it out and that term goes to zero by
our definition of f , leaving
Z Z
u f dS = (φ + ψ)(u2 − ū2 )2 6= 0
2
Γ Γ

(⇐) Trivial, since if u = c constant, λ0 (0) is clearly 0.

References
[1] Pedrosa, H.L. Renato, Some Results Regarding Symmetry and
Symmetry-breaking Proprieties of Optimal Composite Membranes,
Progress in Nonlinear Differential Equations and Their Applications,
Vol 66, 429-442 (2005)

[2] Barbosa, Joäo Lucas and do Carmo, Manfredo Stability of Hypersurfaces

with Constant Mean Curvature, Mathematische Zeitschrift 185, 339-353
(1984)

14
 This research was conducted in the IRES program at UNICAMP, Brazil,
funded by NSF and CNPq, and was advised by Renato Pedrosa of the
UNICAMP Mathematics Department. The authors would like to thank the
Department of Mathematics for their hospitality.

15