Throughout this lecture, Ω is a connected bounded domain of class C∞.
The eigenvalue problem
We now turn our attention to the eigenvalue problem for L, i.e. the problem of finding λ∈R and u not identically zero solving the following boundary value problem:
Luu∣∂Ω=λuon Ω≡0
By Theorem 9.3, there are at most countably many eigenvalues λ (and correponding eigenfunctions u). We now obtain more refined information on the solutions to this eigenvalue problem.
Assume L is in divergence form with the same setup as in Week 8 and that aij∈C∞(Ω). We shall focus on the case where
Lu=−i,j=1∑n∂i(aij⋅∂ju),
in which case L∗u≡Lu, since BL(u,v)=∫Ω∑i,jaij∂iu∂jv=BL∗(u,v) for any u,v∈H01,2(Ω); we say that L is symmetric. We shall require the following fundamental result from functional analysis.
Theorem 12.1 (eigenvectors of symmetric compact operators). If X is a separable Hilbert space (i.e. X has a countable dense subset) and A:X→X is a compact symmetric linear mapping (for all x,y∈X⟨x,Ay⟩=⟨Ax,y⟩), then the eigenvalues {μi}i=1∞ (counted with multiplicity) of A are real and there exists an orthonormal basis {φi}i=1∞ of X such that for each i∈N, Aφi=μiφi.
Theorem 12.2 (eigenvalues of symmetric elliptic operators). If b≡0 and c≡0, then each eigenvalue λ of L is real and positive. Moreover, writing {λk}k=1∞ for the eigenvalues of L counted with multiplicity, we have that 0<λ1≤λ2≤… with λkk→∞∞. Finally, there exists a sequence {wk}k=1∞⊂H01,2(Ω) with Lwk=λkwk for each k∈N such that {wk}k=1∞ is an orthonormal basis of L2(Ω).
Proof. First note that if Lu=λu for some λ∈C and real-valued u not equivalently zero, then
Now, since BL is coërcive, Lax-Milgram implies that there is a unique continuous linear mapping A:L2(Ω)→H01,2(Ω)⊂L2(Ω) such that
BL(Af,v)=⟨f,v⟩
for all f∈L2(Ω) and v∈H01,2(Ω). Just as in Week 8, we deduce that A:L2(Ω)→L2(Ω) is compact, and we see that it is symmetric as well, since whenever f∈L2(Ω) and g∈H01,2(Ω),
⟨f,Ag⟩=BL(Af,Ag)=BL(Ag,Af)=⟨g,Af⟩
so that by approximation, ⟨f,Ag⟩=⟨Af,g⟩ for all f,g∈L2(Ω). Now,
u∈H01,2(Ω),Lu=λu⇔u∈L2(Ω),Au=λ−1u
so that as in Weeks 8 and 9, we deduce that the set of all eigenvalues of L may be written as monotone nondecreasing sequence {λi}i=1∞ with λii→∞∞. By Theorem 11.10, the (orthonormal) eigenfunctions {wi}i=1∞ associated with {λi}i=1∞ span L2(Ω).
Remark 12.3. By higher regularity, the {uk}k=1∞ are smooth on Ω (up to a set of measure zero), and if ∂Ω is smooth, they are smooth on Ω.
By Theorem 12.2, any u∈L2(Ω) may be written in terms of the eigenfunctions {wk}k=1∞ as
u=L2−N→∞limi=1∑Nuiwi,
where ui=⟨wi,u⟩ for each i∈N. We now turn our attention to the convergence of this series in H01,2(Ω) for a given u∈H01,2(Ω). We first note the following.
Lemma 12.4. The bilinear form BL associated to L defines an inner product on H01,2(Ω). Moreover, the associated norm is equivalent to ∣∣⋅∣∣1,2, i.e. there exist constants c0,c1>0 depending only on Ω such that for any u∈H01,2(Ω),
c0∣∣u∣∣1,2≤BL(u,u)≤c1∣∣u∣∣1,2.
Proof. BL is bilinear and symmetric by assumption. Moreover, by ellipticity,
Remark 12.5. BL is commonly known as the energetic inner product and the associated norm the energetic norm associated to L, the motivation being the Dirichlet-type energy
E(u)=21∫Ωi,j=1∑naij∂iu⋅∂ju=21BL(u,u)
from which BL naturally arises (cf. Lemma 7.10).
Corollary 12.6. The eigenfunctions {wk}k=1∞ form a basis of H01,2(Ω) and for any u∈H01,2(Ω),
u=W1,2−N→∞limi=1∑Nuiwi,
where ui=⟨wi,u⟩ for each i∈N.
Proof. We shall employ the energetic inner product. First of all, note that by definition,
so that u1=0 in L2(Ω), thus also in H01,2(Ω). Therefore, u∈X and we may expand u in terms of the orthonormal basis {λiwi}i=1∞ (limit in W1,2(Ω)):
Example 12.7. A ubiquitous concrete instance of the above considerations is n=1, Ω=]0,1[ and Lu:=−Δu=−u′′. In this case, the eigenvalue problem boils down to the problem of solving the ODE
−u′′(x)=λu(x),x∈]0,1[
for u and λ>0 subject to the boundary conditions u(0)=u(1)=0. From ODE theory, we know that any C2 solution to this equation is of the form
u(x)=Acos(λx)+Bsin(λx),
where A,B∈R. Using the boundary conditions, we obtain constraints on A, B and λ:
0=u(0)=A0=u(1)=Bsin(λ)
The last condition implies that either B=0, in which case u is the trivial solution, or sinλ=0⇔λ=jπ for j∈N. Therefore, we see that the eigenvalues of L are given by {λj=j2π2}j=1∞ with corresponding orthonormal eigenfunctions {x↦wj(x)=2sin(jπx)}j=1∞ so that any u∈L2(]0,1[) may be expanded in the form
u(x)=j=1∑∞uj⋅2sin(jπx);
such a series is called a Fourier series.
The first eigenvalue
We now take a closer look at the smallest eigenvalue λ1>0 of Lu=−∑i,j=1n∂i(aij∂ju), which we call the principal eigenvalue. It turns out it may be characterised variationally.
Theorem 12.8 (variational principle for principal eigenvalue). The principal eigenvalue of L is characterised by the condition
λ1=inf{BL(u,u):u∈H01,2(Ω),∣∣u∣∣2=1};(*)
moreover, the infimum is attained at a positive function w1∈H01,2(Ω)∩C∞(Ω) solving the equation Lw1=λ1w1 , and if u∈H01,2(Ω) is any weak solution to the equation Lu=λ1u, then u=k⋅w1 a.e. for some k∈R.
Remark 12.9. This theorem states in particular that the eigenvalue λ1 is simple so that the eigenvalues of L satisfy the inequality
0<λ1<λ2≤λ3≤….
Remark 12.10. (*) is known as Rayleigh's formula and may be written as
λ1=u∈H01,2(Ω)\{0}min∣∣u∣∣22BL(u,u).
If in particular L=−Δ, then the lowest eigenvalue is given by
with equality iff ⟨wj,u⟩=0 for all j>1, i.e. iff u=w1. This establishes the first claim. For the second claim, we shall show that if u∈H01,2(Ω) is a nontrivial weak solution to Lu=λ1u, then it must be either positive or negative on Ω. To this end, write u+(x):=max{u(x),0} and u−(x):=max{−u(x),0} so that u may be written as
u=u+−u−,
and we have that u+⋅u−=0 so that ∣∣u∣∣2=∣∣u+∣∣2+∣∣u−∣∣2. Moreover, it may be shown that u±∈H01,2(Ω) and
Du±(x)={±Du(x),0,x∈{±u≥0}x∈{±u≤0}
a.e. so that ∂iu+⋅∂ju−≡0 a.e., which in turn implies that BL(u+,u−)=0 so that BL(u,u)=BL(u+,u+)+BL(u−,u−). Now, by the proof of the first part, BL(u±,u±)≥λ1∣∣u±∣∣22. It turns out that equality holds in both cases, for if we had a strict inequality, then
which is impossible. Therefore, by the first part, u+∈H01,2(Ω) is a weak solution to Lu=λ1u. By regularity theory, u+∈C∞(Ω), and since
Lu+=λ1u+≥0,
u+ is a supersolution to the equation Lu=0 so that by the strong maximum principle, since u+∣∂Ω≡0 and u+≥0, either u+≡0 on Ω or u+>0, in which case u−≡0. A similar argument applies to u− as well, thus establishing the second claim. To establish the last claim, let u,v∈H01,2(Ω) be any nontrivial solutions to Lu=λ1u. Then by the preceding part, both ∫Ωu and ∫Ωv are nonzero so that we may find an α∈R such that
∫Ωu−αv=0.
But since u−αv is a solution to the same equation, the preceding part now implies that u−αv≡0.
All of the above considerations were for the symmetric elliptic operator
Lu=−i,j=1∑n∂i(aij∂ju).
More generally, consider the more general nondivergence form elliptic operator
Lu:=−i,j=1∑naij∂i∂ju+i=1∑nbi∂iu+cu
with aij,bi,c∈C∞(Ω) for all i,j∈{1,…,n} and c≥0. This operator is not symmetric in general; therefore, its eigenvalues will not necessarily be real. However, we can still say something about its principal eigenvalue, which is both real and positive.
Theorem 12.11. There exists a λ1>0 such that
λ1 is an eigenvalue of L: There exists a u∈H01,2(Ω)\{0} such that Lu=λ1u weakly.
λ1 is simple: If v∈H01,2(Ω) is any weak solution to Lv=λ1v, then v=α⋅u for some α∈R.