Appendix A

Compact Operators

In this appendix we give some general properties about compact operators. We first
recall the Ascoli-Arzelà Theorem.

Theorem A.1 (Ascoli-Arzelà Theorem). Let K be a compact metric space and let F
be a bounded subset of CpK, Rq. We assume that F is equicontinuous:

@ε ą 0, Dδ ą 0, @f P F, @x, y P K, dpx, yq ď δ ùñ |f pxq ´ f pyq| ď ε.

Then the closure F of F in CpKq is compact.

A.1 Compact operators

A.1.1 Definition and first properties
Définition A.2. Let X and Y be Banach spaces. A bounded linear operator T : X Ñ Y
is said to be compact if for any bounded sequence pun qnPN P X N , the sequence pT un qnPN
has a convergent subsequence in Y . Equivalently, T is compact if T pBX q is compact in
Y , where BX is the unit ball in X.

Given two Banach spaces X, Y we denote by KpX, Y q the set of compact operators
from X to Y . We also write KpXq “ KpX, Xq.

Example A.3. Finite rank operators are compact.

Example A.4. We denote by pen qnPN˚ the canonical basis of �2 pN˚ q. We consider on
�2 pN˚ q the linear map A such that Aen “ enn for all n P N˚ . Then A is compact on
�2 pN˚ q.

Proposition A.5. Let X and Y be two Banach spaces.

(i) Let K P KpX, Y q and let pxn qnPN be a sequence in X which converges weakly to
some x P X (i.e. for any ϕ P X˚ we have ϕpxn q Ñ ϕpxq). Then Kpxn q converges
(in norm) to Kpxq.

(ii) KpX, Y q is a closed subspace of LpX, Y q.

(iii) For K P KpX, Y q, B1 P BpX1 , Xq and B2 P BpY, Y2 q we have K ˝ B1 P KpX1 , Y q

and B2 ˝ K P KpX, Y2 q.

69
M2RI - Elliptic PDEs and Evolution Equations

(iv) For K P KpX, Y q we have K ˚ P KpY ˚ , X ˚ q.

Proof. We prove the first and last statements.

‚ The sequence pxn qnPN is weakly convergent, so it is bounded in X (see Proposition
3.5.(iii) in [Brézis]). By continuity, a convergent subsequence of pKpxn qqnPN necessarily
goes to Kpxq. This implies that Kpxn q goes strongly to Kpxq.
‚ Let pϕn qnPN be a bounded sequence in Y ˚ . We denote by BX the unit ball in X.
Since K is compact, KpBX q is a compact metric space, and the functions ϕn , n P
N, are equicontinuous thereon. Then, by the Ascoli-Arzelà Theorem, there exists a
subsequence pϕnk qkPN convergent in C 0 pKpBX qq. We denote by ϕ P C 0 pKpBX qq the
limit. In particular we have

sup |ϕnk pKpxqq ´ ϕpKpxqq| ÝÝÝÝÑ 0.

}x}X ď1 kÑ`8

We deduce that pϕnk ˝ Kq is a Cauchy sequence in X ˚ . Since X ˚ is a Banach space, it

has a limit in X ˚ . This proves that K ˚ P KpY ˚ , X ˚ q.

We finish this paragraph with more examples of compact operators.

Let Ω be an open subset of Rd . For k P N we denote by Cbk pΩq the set of functions u
of class C k on Ω such that B α u is bounded on Ω for all |α| ď k. Then Cbk pΩq is endowed
by the norm defines by ÿ
}u}C k pΩq “ }B α u}L8 pΩq .
b
|α|ďk

Proposition A.6. Let Ω be an open bounded and subset of Rd and k P N. Then Cbk`1 pΩq
is compactly embedded in Cbk pΩq.

Proof. Let pun qnPN be a bounded sequence in C k`1 pΩq. Let M be such that }un }C k`1 ď
b
M . Let α P Nd with |α| ď k and j P �1, d�. Let x P Ω and r ą 0 such that Bpx, rq Ă
Ω. Since }∇B α un }L8 pΩq ď M , the sequence pun q is uniformly Lipschitz in Bpx, rq. In
particular, the sequence pB α un q is uniformly equicontinuous on Ω. By the Ascoli-Arzelà
Theorem, it has a subsequence which converges to some vα in C 0 pΩq. Then there exists
an increasing sequence pnk q such that B α unk goes to vα when n Ñ `8 for all |α| ď k.
Let α P Nd with |α| ď k. Let x P Ω. For t P R small enough we have

vα px ` tej q ´ vα pxq “ lim B α unk px ` tej q ´ B α unk pxq

kÑ`8
żt
“ lim B α`ej unk px ` sej q ds.
kÑ`8 0

Since the map s ÞÑ B α`ej unk px ` sej q converges uniformly to s ÞÑ vα`ej px ` sej q on r0, ts
we get
żt
vα px ` tej q ´ vα pxq “ vα`ej px ` sej q ds.
0

This proves that Bj vα “ vα`ej . Finally for all |α| ď k we have B α v “ vα and we have

}unk ´ v}C k pΩq ÝÝÝÝÑ 0.

b kÑ`8

70 J. Royer - Université Toulouse 3

 Compact Operators

Exercise 42. Let Ω be a bounded subset of Rd . Let k P N and θ Ps0, 1r. We recall
that C k,θ is the set of functions of class C k whose derivatives are bounded and moreover
the derivatives of ordre k are Hölder-continuous of exponent θ. It is endowed with the
norm defined by
ÿ ÿ |B α upxq ´ B α upyq|
}u}C k,θ pΩq “ }B α u}L8 pΩq ` sup .
αďk |α|“k
x,yPΩ |x ´ y|θ
x‰y

Prove that C k,θ pΩq is compactly embedded in Cbk pΩq.

Example A.7. Let K P C 0 pr0, 1s2 q. For u P C 0 pr0, 1sq and x P r0, 1s we set
ż1
pT uqpxq “ Kpx, yqupuq dy.
0

Let M ą 0 and let pun qnPN be a sequence in C 0 pr0, 1sq such that }un }8 ď M for all
n P N. Let x P r0, 1s and ε ą 0. Since K is uniformly continuous there exists δ ą 0
such that for all px1 , y1 q, px2 , y2 q P r0, 1s2 we have
ε
|x1 ´ x2 | ` |y1 ´ y2 | ď δ ùñ |Kpx1 , y1 q ´ Kpx2 , y2 q| ď .
M
Then for n P N and x1 P r0, 1s such that |x ´ x1 | ď δ we have
ż1
1 ε
|pT un qpxq ´ T un px q| ď |Kpx, yq ´ Kpx1 , yq| |un pyq| dy ď
0 .
This proves that the family pT un qnPN is equicontinuous. By the Ascoli-Arzelà Theorem it
has a convergent subsequence in C 0 pr0, 1sq, which proves that T is compact on C 0 pr0, 1sq.

A.2 Fredholm Alternative

We consider a Hilbert space H.
Theorem A.8. Let K P KpHq. Then pId ´Kq is injective if and only if it is surjective,
and in this case its inverse defines a bounded operator on H. In any case we have

dimpKerpId ´Kqq “ dimpKerpId ´K ˚ qq ă `8.

Moreover RanpId ´Kq is always closed, and in particular

RanpId ´Kq “ KerpId ´K ˚ qK .

Remark A.9. We recall that for any A P LpHq we have

RanpAq “ KerpA˚ qK .

Proof. ‚ Assume by contradiction that dimpKerpId ´Kqq “ `8. Then we can find a
sequence pun qnPN in H such that �un , um � “ δn,m and Kun “ un for all n, m P N. This
is in particular a bounded sequence but, for n ‰ m,

}Kun ´ Kum }2H “ }un ´ um }2H “ 2,

2021-2022 71
M2RI - Elliptic PDEs and Evolution Equations

so the sequence pKun qnPN cannot have a convergent subsequence. This gives a contra-
diction and prove that dimpKerpId ´Kqq ă `8.
‚ Then we prove that there exists γ ą 0 such that

@u P KerpId ´KqK , }u ´ Ku}H ě γ }u}H . (A.1)

If this is not the case, we can find a sequence pun qnPN in KerpId ´KqK such that }un }H “
1 and }un ´ Kun }H ď 2´n for all n P N. Since pun qnPN is bounded, there exists a
subsequence punk qkPN and u P H such that unk goes weakly to u as k Ñ `8 . By
Proposition A.5, Kunk goes to Ku as k Ñ `8. Then

unk “ Kunk ` punk ´ Kunk q ÝÝÝÝÑ Ku.

kÑ`8

This implies that u “ Ku, so u P KerpId ´Kq. In particular, for all n P N we have
�u, unk �H “ 0 so, taking the limit, }u}H “ 0. This gives a contradiction and proves
(A.1).
‚ We deduce from (A.1) that RanpId ´Kq is closed in H. Indeed, let pvn qnPN be a
sequence in RanpId ´Kq which goes to some v in H. Then for all n P N there exists
un P KerpId ´KqK such that vn “ pId ´Kqun . By (A.1), pun qnPN is a Cauchy sequence in
H, and hence it has a limit u P H, By continuity, we have v “ pId ´Kqu P RanpId ´Kq,
which proves that RanpId ´Kq is closed.
‚ Now assume that pId ´Kq is injective, and assume by contradiction that H1 “
pId ´KqpHq is not equal to H. Since H1 is closed, it is a Hilbert space with the structure
inherited from H, and by restriction, K defines a compact operator on H1 . We set H2 “
pId ´KqpH1 q. Then H2 is closed, and since pId ´Kq is injective, we have H2 Ł H1 (take
u P HzH1 , then pId ´Kqu belongs to H1 zH2 ). By induction we set Hk “ pId ´KqpHk´1 q
for all k ě 2. Then Hk is closed and Hk`1 Ł Hk for all k P N˚ . In particular, for all
k P N˚ we can find uk P Hk such that }uk }H “ 1 and uk P Hk`1 K
. Then for k P N˚ and
j ą k we have

Kuj ´ Kuk “ ´puj ´ Kuj q ` puk ´ Kuk q ` uj ´ uk .

Since ´puj ´ Kuj q ` puk ´ Kuk q ` uj P Hk`1 this yields

}Kuj ´ Kuk } ě 1.

This gives a contradiction since K is compact. Thus, if pId ´Kq is injective, then it is
also surjective.
‚ Conversely, assume that RanpId ´Kq “ H. Then KerpId ´K ˚ q “ t0u. Since K ˚ is
also compact, we deduce that pId ´K ˚ q is surjective, and finally

KerpId ´Kq “ KerpId ´K ˚˚ q “ RanpId ´K ˚ qK “ t0u .

This proves that pId ´Kq is injective if and only if it is surjective. Moreover, in
this case, (A.1) proves that the inverse pId ´Kq´1 defines a bounded operator with
}pId ´Kq´1 }LpHq ď γ ´1 .
‚ It remains to prove that KerpId ´Kq and KerpId ´K ˚ q have the same dimension.
Assume by contradiction that dimpKerpId ´Kqq ă dimpRanpId ´KqK q. There exists a
bounded operator A : KerpId ´Kq Ñ RanpId ´KqK injective but not surjective. We
extend A by 0 on KerpId ´KqK . This defines an operator A on H which has a finite

72 J. Royer - Université Toulouse 3

 Compact Operators

dimensional range included in RanpId ´KqK . In particular it is compact, and so is

K̃ “ K ` A. Let u P KerpId ´K̃q. We have u ´ Ku “ Au. Since u ´ Ku P RanpId ´Kq
and Au P RanpId ´KqK , we have u ´ Ku “ 0. Therefore u “ 0 since A is injective
on KerpId ´Kq. Then pId ´K̃q is injective, and hence surjective. However for v P
RanpId ´KqK zRanpAq the equation

u ´ pKu ` Auq “ v

cannot have a solution. This gives a contradiction and proves that

dimpKerpId ´Kqq ě dimpRanpId ´KqK q “ dimpKerpId ´K ˚ qq.

We get the opposite inequality by interchanging the roles of K and K ˚ , and the proof
is complete.

Exercise 43. Let K P LpHq. Prove that

˜ ¸
ď ˘
dim KerppId ´Kq k
ă `8.
kPN

A.3 Spectral properties

In this section we discuss the spectral properties of a compact operator. We first recall
the definition of the spectrum of a general operator.

Let H be a real (or complex) Hilbert. An operator A on H is a linear map from a

dense subset D of H to H. We say that D is the domain of A.
Let λ P R (or λ P C). We say that λ is in the resolvent set ρpAq of A if the operator
pA ´ λ Idq : D Ñ H is bijective and if its inverse pA ´ λ Idq´1 defines a bounded operator
on H. We usually write pA ´ λq instead of pA ´ λ Idq. The spectrum σpAq of A is the
complement of ρpAq in R (or C).
We recall that if H is of finite dimension, a linear map is bijective if and only if it
is injective, and in this case the inverse is always continuous, so the spectrum of A is
exactly the set of eigenvalues. This is not the case in general.

If λ is an eigenvalue of A, then its geometric multiplicity is

` ˘
dim KerpA ´ λq ,

and its algebraic multiplicity is

˜ ¸
ď ` ˘ ` ˘
dim Ker pA ´ λqk “ lim dim KerpA ´ λqk .
kÑ`8
kPN

In particular, the geometric multiplicity is smaller than or equal to the algebraic multi-
plicity.

2021-2022 73
M2RI - Elliptic PDEs and Evolution Equations

A.3.1 Spectrum of compact operators

For compact operators, we have the following result.

Theorem A.10. Let K P KpHq.

(i) If dimpHq “ `8 then 0 belongs to the spectrum of K.

(ii) λ ‰ 0 belongs to the spectrum of K if and only if it is an eigenvalue of K. In this

case it is an eigenvalue of finite geometric (and algebraic) multiplicity.

(iii) σpKqz t0u is finite or is given by a sequence of eigenvalues tending to 0.

Proof. ‚ Assume that 0 belongs to the resolvent set of K. Then Id is the composition
of the compact operator K with the bounded operator K ´1 , so Id is a compact operator.
This implies that dimpHq ă `8.
‚ Let λ P R˚ (or C˚ ). Then we have K ´ λ “ λpλ´1 K ´ Idq. Since λ´1 K is compact,
Theorem A.8 shows that pK ´ λq is bijective (with bounded inverse) if and only if it is
injective, so λ is in the resolvent set of K if and only if it is not an eigenvalue. Moreover,
if λ is an eigenvalue of K we have dimpKerpK ´ λqq “ dimpKerpλ´1 K ´ Idqq ă `8.
More generally, Exercise 43 shows that 1 is an eigenvalue of finite algebraic multiplicity
for λ´1 K.
‚ Since K is a bounded operator, the set of eigenvalues of K is bounded in R (C).
Assume that pλn qnPN is a sequence of distinct non-zero eigenvalues of K tending to some
λ. We prove that λ “ 0. For n P N we consider wn P Hz t0u such that Kwn “ λn wn .
Then for n P N we set Hn “ spanpw0 , . . . , wn´1 q and we consider un P Hn such that
}un } “ 1 and un P Hn´1 K
if n ě 1. Then for j P N and k ą j we have
› › › ›
› Kuk Kuj › › Kuk ´ λk uk Kuj ´ λj uj ›
› ´ › “ › ´ ` u ´ u › ě 1,
› λk λj ›H › λk λj
k j ›
H

since Kuk ´ λk uk , Kuj ´ λj uj , uj P Hk´1 . If λ ‰ 0 we obtain a contradiction with the

compactness of K.

A.3.2 The case of symmetric operators

Let A be a bounded operator on H. We assume that A is symmetric:

@ϕ, ψ P H, �Aϕ, ψ�H “ �ϕ, Aψ�H .

In particular, even if H is a complex Hilbert space, we have �Au, u� P R for all u P H. In

particular, the eigenvalues of A are real. Moreover, two eigenspaces of A corresponding
to two distinct eigenvalues are orthogonal.

Lemma A.11. Let A be a bounded symmetric operator on H. Let

m “ inf �Au, u�H and M “ sup �Au, u�H .

uPH uPH
}u}“1 }u}“1

Then σpAq Ă rm, M s and m, M P σpAq.

74 J. Royer - Université Toulouse 3

 Compact Operators

Proof. We consider the case where H is a real Hilbert space. We prove that sM, `8rĂ
ρpAq and that M P σpAq. Let λ ą M . For u P H we have

�λu ´ Au, u�H ě pλ ´ M q }u}2H .

By the Lax-Milgram Theorem, the operator λ ´ A is bijective with bounded inverse on

H, so λ P ρpAq.
Now let pun qnPN be a sequence in H such that }un }H “ 1 for all n P N and

�Aun , un � ÝÝÝÝÑ M.
nÑ`8

The quadratic form u ÞÑ �pM ´ Aqu, u� is non-negative, so by the Cauchy-Schwarz

inequality we have for all u, v P H

|�pM ´ Aqu, v�H |2 ď �pM ´ Aqu, u�H �pM ´ Aqv, v�H

Applied with u “ un and v “ pM ´ Aqun this gives

� �
}pM ´ Aqun }2H ď �pM ´ Aqun , un �H pM ´ Aq3 un , pM ´ Aqun ÝÝÝÝÑ 0.
H nÑ`8

This proves that M P σpAq.

Theorem A.12. Let H be a separable Hilbert space and let K be a compact and sym-
metric operator on H. Then there exists an orthonormal basis pen qnPN consisting of
eigenvectors of K.

Proof. Let pλn q1ďnďN for N P NYt`8u be the sequence of distinct non-zero eigenvalues
of K. For n P �1, N � we set Hn “ KerpK ´ λn q. Then we have dimpHn q P N˚ . We also
set H0 “ KerpKq. ` Ť ˘
We set H̃ “ span N
n“0 H n . We have KpH̃q Ă H̃ and hence KpH̃K q Ă H̃K . Assume
by contradiction that H̃K ‰ t0u. The restriction of K to H̃K is compact and symmetric,
and it has no eigenvalue, so its spectrum is included in t0u. By Lemma A.11, we have
�Ku, u� “ 0 for all u P H̃K . We deduce that K “ 0 on H̃K , and hence H̃K Ă KerpKq Ă
H̃. This gives a contradiction and proves that H̃K “ t0u, so H̃ is dense.
It only remains to choose an orthonormal basis of each Hn for n P �1, N �, and a
countable orthonormal basis of H0 (it exists since H is separable).

A.3.3 Operators with compact resolvent

We finish we operators which are not compact but have a compact resolvent.

Theorem A.13. Let A be an operator on H with domain D. Assume that there exists z0
such that pA ´ z0 q is bijective and pA ´ z0 q´1 : H Ñ D Ă H defines a compact operator
on H. Then the spectrum of A consists of a discrete set of eigenvalues with finite
(geometric and algebraic) multiplicities (in particular the spectrum of A is countable
without accumulation points).

Proof. Let B “ A ´ z0 : D Ñ H. We have 0 P ρpBq and B ´1 defines a compact operator

on H. Let λ P C˚ . Assume that λ P ρpBq. We have

B ´1 ´ λ´1 “ ´λ´1 pB ´ λqB ´1 ,

2021-2022 75
M2RI - Elliptic PDEs and Evolution Equations

so B ´1 ´ λ´1 : H Ñ H is invertible, with bounded inverse pB ´1 ´ λ´1 q´1 “ ´BpB ´

λq´1 λ “ ´λ ´ λ2 pB ´ λq´1 . Similarly, on D we have

B ´ λ “ ´λpB ´1 ´ λ´1 qB. (A.2)

If λ´1 P ρpB ´1 q then B´λ : D Ñ H is invertible and its inverse pB´λq´1 “ ´B ´1 pB ´1 ´

λ´1 q´1 λ´1 defines a bounded operator on H. Thus λ P ρpBq. This proves that the map
λ ÞÑ λ´1 is a bijection between the spectrum of B and the non-zero spectrum of B ´1 .
In particular, the spectrum of B is discrete. Moreover, if λ P σpBq then pB ´1 ´ λ´1 q is
not injective. By (A.2), λ is an eigenvalue of B, with finite geometric multiplicity. More
precisely, since B and B ´1 commute, we see that for k P N˚ we have
` ˘ ` ˘
Ker pB ´ λqk “ Ker pB ´1 ´ λ´1 qk ,

so the eigenvalues of B have finite algebraic multiplicities. After translation, the operator
A shares the same properties and the proof is complete.

76 J. Royer - Université Toulouse 3