Operadores Elípticos - Julián López Gómez

February 21, 2013 9:29 World Scientific Book - 9in x 6in book-lsoeo
Publishers’ page
i
Publishers’ page
ii
Publishers’ page
iii
Publishers’ page
iv
Julián, you know...
no hard work, no glory...
and trust me...
if you want to be modern, you should be classical...
with all my love.
v
vi Linear Second Order Elliptic Operators

Preface
The main goal of this book is to provide a reasonably self-contained proof

of a scalar version of the characterization of the strong maximum principle
established by Theorem 2.1 of J. López-Gómez and M. Molina-Meyer [148],
which was extended by H. Amann and J. López-Gómez [13] to a generalized
class of non-classical mixed boundary conditions. Precisely, it will be shown
that, for any second order linear elliptic operator
L := −div (A∇·) + ⟨b, ·⟩ + c
in a regular bounded domain Ω, and any boundary operator B of non-
classical mixed type on ∂Ω, as discussed in the beginning of Chapter 4, the
following five conditions are equivalent:
i) (L, B, Ω) satisfies the strong maximum principle, in the sense that
{
Lu = f in Ω
and (f, g) > (0, 0) =⇒ u ≫ 0.
Bu = g on ∂Ω
By h > 0 it is meant that h ≥ 0 with h ̸= 0, and we write u ≫ 0 if u(x) > 0
for all x ∈ Ω and
∂u
(x) < 0 for all x ∈ u−1 (0) ∩ ∂Ω,
∂n
where n stands for the outward unit normal to Ω at x ∈ ∂Ω.
ii) (L, B, Ω) satisfies the maximum principle, in the sense that
{
Lu = f in Ω
and (f, g) ≥ (0, 0) =⇒ u ≥ 0.
Bu = g on ∂Ω
iii) (L, B, Ω) admits a positive strict supersolution h ∈ W 2,p (Ω), p > N ,
i.e.,
(Lh, Bh) > (0, 0) in Ω × ∂Ω.
vii
viii Linear Second Order Elliptic Operators
iv) The principal eigenvalue of the linear eigenvalue problem

{
Lφ = τ φ in Ω
(0.1)
Bφ = 0 on ∂Ω
is positive. It will be throughout denoted by σ0 .
v) The resolvent operator of the linear problem
{
Lu = f ∈ Lp (Ω) in Ω
Bu = 0 on ∂Ω
is strongly positive.
This characterization theorem, which is Theorem 7.10 of Chapter 7, has
tremendously facilitated the development of the modern theory of spatially
heterogeneous reaction diffusion systems. In particular, it has enhanced the
mathematical analysis of a great variety of nonlinear parabolic equations
whose dynamics are regulated by metasolutions (see R. Gómez-Reñasco
[85], R. Gómez-Reñasco and J. López-Gómez [86], J. López-Gómez [145],
J. López-Gómez and M. Molina-Meyer [151], and the list of references in
each of these monographs).
As a consequence from Theorem 1.1 of H. Berestycki, L. Nirenberg and
S. R. S. Varadhan [27], the equivalence of ii) and iv) holds true under no
regularity constraints on Ω when Bu = u for all u. Also, by Corollary 2.4
of [27], iii) implies ii) under Dirichlet boundary conditions. But, seemingly,
the lack of regularity of Ω did not allow H. Berestycki, L. Nirenberg and
S. R. S. Varadhan [27] to obtain the complete equivalence of J. López-
Gómez and M. Molina-Meyer [148]. Both papers, [27] and [148], appeared
simultaneously, early 1994, and both were submitted for publication in
1992.
More recently, H. Amann [11] completed the previous list of five items
by proving that each of them is also equivalent to each of the next two:
vi) (L, B, Ω) satisfies the weak maximum principle.

vii) (L, B, Ω) satisfies the very weak maximum principle.
But these natural extensions and refinements are left outside the general
scope of this book, because, instead of encyclopedic generality, we are pay-
ing special attention to simplicity and transparency in its exposition so that
it can be comfortably read by graduate and even undergraduate students.
Our proof of the previous characterization theorem is based on the clas-
sical minimum principles of E. Hopf [103, 104], M. H. Protter and H. F.
Weinberger [183] and J. M. Bony [28]. Consequently, though Theorem 7.10
Preface ix
is relatively recent, this is a rather classical textbook on linear second order

elliptic operators.
As this book focuses on linear operators, some important nonlinear ap-
plications of the classical minimum principles, like the moving plane method,
have not been included here. The moving plane method has shown to be a
milestone in analyzing the symmetry properties of the positive solutions in
a variety of spatially homogeneous nonlinear elliptic problems. Though it
was introduced by A. D. Alexandroff [5] and later refined by J. Serrin [201],
it became famous when B. Gidas, W. M. Ni and L. Nirenberg [78] used it
to prove the radial symmetry of the positive solutions of a special class of
spatially homogeneous semilinear elliptic problems in balls. Naturally, the
moving plane method has been covered by most of the existing literature
on nonlinear elliptic equations (see, e.g., T. Suzuki [216], P. Pucci and J.
Serrin [184], and the references therein).
Although this book might have shared title with the recent monograph
of P. Pucci and J. Serrin [184], both texts are really complementary, as
there is almost no overlapping between them. Essentially, [184] generalizes
the minimum principle of E. Hopf to a variety of quasilinear elliptic prob-
lems, whereas this monograph characterizes the range of applicability of
the generalized minimum principle of M. H. Protter and H. F. Weinberger
[183], and uses it to characterize whether or not the resolvent of a linear
elliptic operator is positive. As a consequence from this characterization, it
becomes apparent that the generalized minimum principle of [183] holds if,
and only if, σ0 > 0. As this result seems to be new, some of the contents of
P. Pucci and J. Serrin [184] might be substantially generalized in its light.
Anyway, in order to illustrate the degree of independence between the
book of P. Pucci and J. Serrin [184] and the present one, the reader should
be aware that some of the most fundamental references pioneering the
present monograph, like J. López-Gómez and M. Molina-Meyer [148], H.
Amann and J. López-Gómez [13], and the elegant paper of H. Berestycki,
L. Nirenberg and S. R. S. Varadhan [27], were not incorporated to the
bibliography of [184].
This book consists of nine chapters. The first seven provide us with a
self-contained proof of the characterization theorem that we have just es-
tablished above, which can be delivered at an undergraduate level, whereas
the last two chapters apply it to obtain some of the most relevant properties
of σ0 from the point of view of the applications.
Essentially, Chapter 1 studies the classical minimum principles of E.
Hopf [103, 104], and M. H. Protter and H. F. Weinberger [183], and Chap-
x Linear Second Order Elliptic Operators
ter 2 analyzes the more recent theorems of classification of supersolutions

of W. Walter [224] and J. López-Gómez [144]. As a result from these clas-
sification theorems, the generalized minimum principle of M. H. Protter
and H. F. Weinberger [183] has been completed and substantially polished.
We decided to begin this book by studying these classical topics, as they
can be easily delivered at an undergraduate level with no previous knowl-
edge of generalized solutions, Hilbert space techniques, or elliptic regularity
theory. Indeed, the first two chapters can be taught as part of an elemen-
tary course on classical minimum principles within the fruitful spirit of the
paradigmatic monograph of M. H. Protter and H. F. Weinberger [183].
Essentially, the next three chapters constitute an informal advanced
course on Hilbert space techniques for linear second order elliptic operators,
which may be taught independently of the more advanced materials treated
in the last four chapters. Actually, we have included these chapters to make
this book easily accessible for undergraduate students. Though most of the
contents of Chapters 3–5 are classical and can be found in a number of
well known textbooks, the reader should appreciate the generality of our
treatment, as we are not focusing our attention exclusively on the Dirichlet
problem, as it occurs in most of the existing textbooks (see, e.g., L. C.
Evans [60]), but on a rather general class of boundary value problems.
More precisely, Chapter 3 studies the theorem of G. Stampacchia [212]
and derives from it the celebrated theorem of P. D. Lax and A. N. Milgram
[122], much within the spirit of Chapter 6 of H. Brézis [29]. Besides polish-
ing some of the contents of Chapter 6 of the textbook of H. Brézis [29], this
book adds the construction of the projections on any closed convex set of
a uniformly convex Banach space, which might be a new result published
here for the first time.
Chapter 4 discusses the concept of weak solution for the problem
{
(L + ω)u = f in Ω
(0.2)
Bu = 0 on ∂Ω
and shows the existence of ω0 ∈ R such that, for every ω > ω0 and f ∈
L2 (Ω), (0.2) has a unique weak solution. The proof of this result is based
on the theorem of P. D. Lax and A. N. Milgram [122] for classic mixed
boundary conditions, but it goes back to J. López-Gómez [147] under the
general non-classical boundary conditions of this book. Astonishingly, the
existence of weak solutions in the general context of [147] is based on a
device introduced by M. H. Protter and H. F. Weinberger [183] to derive
their generalized minimum principle from the minimum principle of E. Hopf
[103].
Preface xi
Naturally, in order to establish the existence and the uniqueness of the

weak solution of (0.2) one must introduce some fundamentals on Sobolev
spaces and trace theorems. This is accomplished in the first part of Chap-
ter 4. However, the main theorems about Sobolev spaces are not proven
in this textbook, because here we are not focusing our attention on the
generalized Lebesgue spaces used in this book, but on the pivotal problem
of characterizing the positivity of the resolvent operator of (0.2). Actually,
most of these proofs are well documented in some recent undergraduate
textbooks, as, e.g., L. C. Evans [60], where the interested readers are sent
to study the proofs, if necessary.
Chapter 5 gives a first glance on the problem of the regularity of the
weak solutions. Essentially, it is a short introduction to this topic through
the Lp -theory of A. P. Calderón and A. Zygmund [36, 37], the ulterior
refinements of S. Agmon, A. Douglis and L. Nirenberg [4], and the method
of continuity. Obviously, the contents of this chapter might have been
expanded to generate an extremely specialized monograph on Lp -regularity
theory. But this is far from being our attempt in an introductory chapter
like Chapter 5, where we restrict ourselves to collect, without proofs, all
the regularity results used in the last four chapters of this book.
The materials of the last four chapters, the core of this book together
with Chapters 1 and 2, have been incorporated to a textbook for the first
time here with the generality adopted in this monograph.
Chapter 6 gives a self-contained proof of the theorem of M. G. Krein
and M. A. Rutman [120]. Basically, it polishes and completes the short
elementary proof of P. Takác [218].
Chapter 7 begins by establishing the generalized minimum principle of
J. M. Bony [28], as well as the weak counterparts of the classical minimum
principles already revisited in Chapters 1 and 2. The materials of Chapters
3–5 allow us to generalize all the classical results of Chapters 1 and 2 to the
weak context of the Sobolev spaces, providing them with its most powerful
versatility.
Alternatively, if we had wished to diffuse our own contributions to the
theory through a more specialized monograph, instead of writing down
a textbook appropriate for undergraduate students, we should have not
included in this book most of the materials of Chapters 3–5, but simply
completed from the very beginning the first two chapters with the minimum
principle of J. M. Bony [28] and invoked to [147] for the existence and
uniqueness of weak solutions.
xii Linear Second Order Elliptic Operators
Then, Chapter 7 combines the weak minimum principles with the the-
orem of M. G. Krein and M. A. Rutman [120] to establish the existence,
the uniqueness and the dominance of the principal eigenvalue σ0 , as well as
the scalar version of Theorem 2.1 of J. López-Gómez and M. Molina-Meyer
[148] and Theorem 2.4 of H. Amann and J. López-Gómez [13].
The last two chapters give a series of fundamental applications of the
main theorem of Chapter 7, which are imperative to study the dynam-
ics of wide classes of spatially heterogeneous reaction-diffusion equations
and systems. As a first application, Chapter 7 concludes by sharpening
the classical minimum principles of E. Hopf and M. H. Protter and H. F.
Weinberger through the characterization of its precise range of applicabil-
ity. Naturally, these refinements might enjoy a huge number of linear and
nonlinear applications.
Precisely, Chapter 8 studies some of the most fundamental properties of
σ0 . Among them, it establishes the monotonicity and continuity properties
of σ0 with respect to c, Ω, |Ω|, and the boundary operator B. Finally, Chap-
ter 9 characterizes the existence of principal eigenvalues for the weighted
linear boundary value problem
{
Lφ = τ W φ in Ω
(0.3)
Bφ = 0 on ∂Ω
where W ̸= 0 is a bounded and measurable function. Our results are
extremely sharp refinements of some pioneering results by R. Courant and
D. Hilbert [44], when W is bounded away from zero, and P. Hess and T.
Kato [97], when W changes sign. Most of the contents of the last two
chapters go back to J. López-Gómez [137] and S. Cano-Casanova and J.
López-Gómez [39].
This book is indebted to the fruitful scientific collaboration of the author
with Professors S. Cano-Casanova and M. Molina-Meyer, members of the
author’s research teams since the beginning of their scientific carriers, as
it is triggered by many of their own mathematical findings. Actually, the
Master thesis of Professor S. Cano-Casanova [38] really was a first attempt
to detail a self-contained proof of Theorem 2.1 of J. López-Gómez and M.
Molina-Meyer [148] under general boundary conditions.
Overall, the whole process of elaboration of this book has taken over ten
years. In the mean time, the author’s research teams have been supported
by a number of projects from the Spanish Government. Among them, we
should mention the following:
• BFM2000-0797, of the Ministry of Science and Technology.
Preface xiii
• REN2003-00707, of the National Plan of Natural Resources (Ministry

of Science and Technology).
• CGL2006–00524/BOS, of the National Plan of Global Change and Bio-
diversity (Ministry of Science and Innovation).
• MAT2009-08259, of the National Plan of Mathematics (Ministry of
Science and Innovation).
• MTM2012-30669, of the National Plan of Mathematics (Ministry of
Economy and Competitiveness).
Part of the research necessary to prepare this monograph has been tremen-
dously facilitated by them, though, rather astonishingly, only MAT2009-
08259 deserved the assignation of students in formation, which might be
explained by the high degree of corruption of the Spanish Administration.
Madrid, December 31st, 2012
J. López-Gómez
xiv Linear Second Order Elliptic Operators

Contents
Preface vii
1. The minimum principle 1

1.1 Concept of ellipticity. First consequences . . . . . . . . . 2
1.2 Minimum principle of E. Hopf . . . . . . . . . . . . . . . . 5
1.3 Interior sphere properties . . . . . . . . . . . . . . . . . . 12
1.4 Boundary lemma of E. Hopf . . . . . . . . . . . . . . . . . 19
1.5 Positivity properties of super-harmonic functions . . . . . 23
1.6 Uniform decay property of E. Hopf . . . . . . . . . . . . . 25
1.7 The generalized minimum principle of M. H. Protter and
H. F. Weinberger . . . . . . . . . . . . . . . . . . . . . . . 30
1.8 Appendix: Smooth domains . . . . . . . . . . . . . . . . . 32
1.9 Comments on Chapter 1 . . . . . . . . . . . . . . . . . . . 38
2. Classifying supersolutions 41
2.1 First classification theorem . . . . . . . . . . . . . . . . . 42
2.2 Existence of positive strict supersolutions . . . . . . . . . 47
2.3 Positivity of the resolvent operator . . . . . . . . . . . . . 52
2.4 Behavior of the positive supersolutions on Γ0 . . . . . . . 52
2.5 Second classification theorem . . . . . . . . . . . . . . . . 53
2.6 Appendix: Partitions of the unity . . . . . . . . . . . . . . 58
3. Representation theorems 63
3.1 The projection on a closed convex set . . . . . . . . . . . 65
3.2 The orthogonal projection on a closed subspace . . . . . . 69
xv
xvi Linear Second Order Elliptic Operators
3.3 The representation theorem of F. Riesz . . . . . . . . . . 71

3.4 Continuity and coercivity of bilinear forms . . . . . . . . . 75
3.5 The theorem of G. Stampacchia . . . . . . . . . . . . . . . 76
3.6 The theorem of P. D. Lax and A. N. Milgram . . . . . . . 78
3.7 Projecting on a closed convex set of a u.c. B-space . . . . 78
3.7.1 Basic concepts and preliminaries . . . . . . . . . . 79
3.7.2 The projection theorem . . . . . . . . . . . . . . . 82
3.7.3 The projection on a closed linear subspace . . . . 85
3.7.4 The projection on a closed hyperplane . . . . . . 87
4. Existence of weak solutions 91

4.1 Preliminaries. Sobolev spaces . . . . . . . . . . . . . . . . 93
4.1.1 Test functions . . . . . . . . . . . . . . . . . . . . 93
4.1.2 Weak derivatives. Sobolev spaces . . . . . . . . . 94
4.1.3 Hölder spaces of continuous functions . . . . . . . 97
4.1.4 Sobolev’s imbeddings . . . . . . . . . . . . . . . . 98
4.1.5 Compact imbeddings . . . . . . . . . . . . . . . . 101
4.2 Trace operators . . . . . . . . . . . . . . . . . . . . . . . . 102
4.3 Weak solutions . . . . . . . . . . . . . . . . . . . . . . . . 114
4.4 Continuity of the associated bilinear form . . . . . . . . . 117
4.5 Invertibility of (4.4) when β ≥ 0 . . . . . . . . . . . . . . 118
4.5.1 Coercivity of the associated bilinear form . . . . . 118
4.5.2 Existence of weak solutions. The resolvent operator 120
4.6 Invertibility of (4.4) for arbitrary β . . . . . . . . . . . . . 122
5. Regularity of weak solutions 129

5.1 Lp (RN )-estimates for the Laplacian . . . . . . . . . . . . 131
5.2 Lp (Ω)-estimates for the Laplacian . . . . . . . . . . . . . 135
5.3 General elliptic Lp (Ω)-estimates when Γ1 = ∅ . . . . . . . 138
5.4 The method of continuity . . . . . . . . . . . . . . . . . . 139
5.5 Regularity of weak solutions when Γ1 = ∅ . . . . . . . . . 141
5.6 A first glance to the general case when Γ1 ̸= ∅ . . . . . . . 147
6. The Krein–Rutman theorem 155

6.1 Orderings. Ordered Banach spaces . . . . . . . . . . . . . 155
Contents xvii
6.2 Spectral theory of linear compact operators . . . . . . . . 161

6.3 The Krein–Rutman theorem . . . . . . . . . . . . . . . . . 164
6.4 Preliminaries of the proof of Theorem 6.3 . . . . . . . . . 166
6.5 Proof of Theorem 6.3 . . . . . . . . . . . . . . . . . . . . . 171
7. The strong maximum principle 187

7.1 Minimum principle of J. M. Bony . . . . . . . . . . . . . . 189
7.2 The existence of the principal eigenvalue . . . . . . . . . . 195
7.3 Two equivalent weak eigenvalue problems . . . . . . . . . 206
7.4 Simplicity and dominance of σ[L, B, Ω] . . . . . . . . . . . 208
7.4.1 Proof of the strict dominance in case Γ0 = ∅ . . . 210
7.4.2 Proof of the strict dominance in case Γ1 = ∅ . . . 212
7.4.3 Proof of the strict dominance in the general case . 215
7.5 The strong maximum principle . . . . . . . . . . . . . . . 215
7.6 The classical minimum principles revisited . . . . . . . . . 217
8. Properties of the principal eigenvalue 225

8.1 Monotonicity properties . . . . . . . . . . . . . . . . . . . 226
8.2 Point-wise min-max characterizations . . . . . . . . . . . 229
8.3 Concavity with respect to the potential . . . . . . . . . . 232
8.4 Stability of Ω along the Dirichlet components of ∂Ω . . . 234
8.4.1 Proof of Proposition 8.5 . . . . . . . . . . . . . . 236
8.4.2 Proof of Theorem 8.4 . . . . . . . . . . . . . . . . 240
8.5 Continuous dependence with respect to Ω . . . . . . . . . 240
8.6 Continuous dependence with respect to β(x) . . . . . . . 254
8.7 Asymptotic behavior of σ(β) as min β ↑ ∞ . . . . . . . . . 260
Γ1
8.8 Lower estimates of σ[L, D, Ω] in terms of |Ω| . . . . . . . . 264
9. Principal eigenvalues of linear weighted boundary value problems 273

9.1 General properties of the map Σ(λ) . . . . . . . . . . . . . 274
9.2 Characterizing the existence of a principal eigenvalue . . . 278
9.3 Ascertaining limλ→∞ σ[L + λV, B, Ω] when V ≥ 0 . . . . . 284
9.3.1 The simplest case . . . . . . . . . . . . . . . . . . 285
9.3.2 The admissible V ’s satisfying the main theorem . 287
9.3.3 The main theorem . . . . . . . . . . . . . . . . . . 290
xviii Linear Second Order Elliptic Operators
9.4 Characterizing the existence of principal eigenvalues for

admissible potentials . . . . . . . . . . . . . . . . . . . . . 307
Bibliography 319
Index 331
Chapter 1
The minimum principle
This chapter considers second order differential operators of the form

∑
N
∂2 ∑ N
∂
L := − aij (x) + bj (x) + c(x), x ∈ Ω, (1.1)
i,j=1
∂xi ∂xj j=1 ∂xj
where N ∈ N, N ≥ 1, Ω is a domain (open and connected by arcs) of RN ,

and
aij , bj , c ∈ L∞
loc (Ω), i, j ∈ {1, ..., N }, (1.2)
i.e., all coefficients of L are measurable functions, as discussed by Lebesgue,
bounded on compact subsets of Ω. Without loss of generality, we can
assume that
aji = aij , i, j ∈ {1, ..., N }.
Then, the matrix of the coefficients
( )
Ax := aij (x) 1≤i,j≤N , x ∈ Ω, (1.3)
is symmetric. Subsequently, the operator
∑
N
∂2
Lp := − aij (x)
i,j=1
∂xi ∂xj
is referred to as the principal part of L, and aij , 1 ≤ i, j ≤ N , are said to

be the principal coefficients of L.
Throughout this book, := means equality by definition and I stands
for the identity map of the underlying linear space (in principle, anyone).
Also, for a given measurable function f : Ω → R, it is said that f ≥ 0 if
f (x) ≥ 0 almost everywhere in Ω, and, given another measurable function
g : Ω → R, it is said that f ≥ g if f − g ≥ 0. Most precisely, it is said
1
2 Linear Second Order Elliptic Operators
that f > 0 when f ≥ 0 and f > 0 on a set of positive measure, and we

write f > g if f − g > 0. Similarly, for every f, g ∈ C(∂Ω), it is said that
f ≥ 0 (resp. f > 0) when f (x) ≥ 0 for all x ∈ ∂Ω (resp. f ≥ 0 and there
exists x ∈ ∂Ω such that f (x) > 0), whereas we write f ≥ g (resp. f > g) if
f − g ≥ 0 (resp. f − g > 0).
Moreover, given two arbitrary Banach spaces U and V , we denote by
L(U, V ) the Banach space of the linear and continuous operators between
U and V , by Iso(U, V ) the set of isomorphisms between U and V , and by
U ′ the topological dual space of U , i.e.,
U ′ := L(U, R).
Notations are shortened by setting
L(U ) := L(U, U ), Iso(U ) := Iso(U, U ).
Further, MN (K) will stand for the linear space of matrices of order N ≥ 1
with entries in the numerical field K ∈ {R, C}, i.e.,
MN (K) = L(KN ),
and, for every L ∈ L(U, V ), N [L] and R[L] will stand for the kernel (null
space) and the image (range) of L, respectively.
This chapter introduces the concept of ellipticity for L and gives a series
of sufficient conditions so that it satisfies the classical minimum principle
in the domain Ω.
Besides L∞ ∞
loc (Ω), in this book we consider the Banach space L (Ω) of all
measurable and bounded functions in Ω, and, for every integer k ≥ 0, the
Banach space C k (Ω̄) of all real functions in Ω̄ with continuous derivatives
of order k in Ω̄. By simplicity, we write C(Ω̄) := C 0 (Ω̄). Similarly, we also
consider
∩
C k (Ω) := C k (D̄).
D open
D̄⊂Ω
1.1 Concept of ellipticity. First consequences
Throughout this book, | · | stands for the Euclidean norm of RN , N ≥ 1,

v
uN
u∑
|ξ| := t ξj2 , ξ = (ξ1 , ..., ξN )T ∈ RN ,
j=1
The minimum principle 3
where T means transposition, as coordinate vectors of RN will always be

written vertically. Also, for every R > 0 and x0 ∈ RN , we denote by BR (x0 )
the open ball of radius R centered at x0 ,
BR (x0 ) := {x ∈ RN : |x − x0 | < R}.
The following definition introduces the concept of ellipticity.
Definition 1.1 (Elliptic operator). Given x ∈ Ω, it is said that L is

elliptic at the single point x if there exists a constant µx > 0 such that
∑
N
aij (x)ξi ξj ≥ µx |ξ|2 for every ξ ∈ RN ,
i,j=1
i.e., when the bilinear form associated to the matrix Ax introduced in (1.3),
B(ξ, η) := ξ T Ax η, (ξ, η) ∈ RN × RN ,
is positive definite; in other words, when all the eigenvalues of Ax are pos-
itive. In such case, µx is said to be an ellipticity constant of L at x.
The operator L is said to be elliptic in Ω when it is elliptic at every
point x ∈ Ω. In such case, L is said to be uniformly elliptic in Ω when
the ellipticity constant µx can be chosen independently of x ∈ Ω, i.e., when
there exists µ > 0 such that
∑
N
aij (x)ξi ξj ≥ µ |ξ|2 for all (x, ξ) ∈ Ω × RN .
i,j=1
The largest µ for which this condition holds is called the ellipticity con-
stant of L in Ω.
The paradigm of elliptic operator is the ‘minus Laplacian’,

∑N
∂2
L = −∆ := − .
j=1
∂x2j
Although this operator seems to be of a very special nature, the next result
shows that the principal part of any elliptic operator can be transformed,
through a local change of coordinates, into −∆.
Proposition 1.1. Suppose L is elliptic at x0 ∈ Ω. Then, there exists

an invertible matrix M ∈ MN (R) such that, for every u ∈ C 2 (Ω), the
transformed function
v(y) := u(x), y := M x, (1.4)
satisfies  
∑
N
∂2v
Lp u|x=x0 = −∆v|y=y0 = − (y0 ) (1.5)
j=1
∂yj2
where Lp is the principal part of L and y0 := M x0 .
Proof. Let x0 ∈ Ω, u ∈ C 2 (Ω), and an invertible matrix
M = (mij )1≤i,j≤N ∈ MN (R).
Then,
∑
N
∂2u
Lp u|x=x = − aij (x0 ) = − ∇Tx Ax0 ∇x u (1.6)
0
i,j=1
∂xi ∂xj x=x0
x=x0
where ∇x stands for the column gradient vector in the x-coordinates. By

(1.4), we find that
∂u ∑N
∂v ∂yi ∑N
∂v
= = mij = (m1j , ..., mN j )∇y v
∂xj i=1
∂y i ∂x j i=1
∂y i
for all 1 ≤ j ≤ N , and hence,

∇x = M T ∇y .
Consequently, (1.6) can be expressed in the form
Lp u|x=x = − ∇Ty M Ax0 M T ∇y v
0 y=y0
and, therefore, to prove (1.5) it suffices to construct a matrix M satisfying

M Ax0 M T = I. (1.7)
To carry out this construction, let λj (x), 1 ≤ j ≤ N , denote, for every
x ∈ Ω, the eigenvalues of the matrix Ax counted according to their alge-
braic multiplicities. As Ax is symmetric, RN possesses a basis consisting
of orthogonal eigenvectors of Ax . Let Cx be the matrix whose columns are
the coordinates of these eigenvectors. Then, Cx is orthogonal, i.e.,
Cx−1 = CxT ,
and, by construction,
( )
Jx := diag λj (x) 1≤j≤N = CxT Ax Cx (1.8)
is the Jordan form of Ax . As L has been assumed to be elliptic at x0 ,
λj (x0 ) > 0 for all 1 ≤ j ≤ N , and hence, by making the choice
( )
1
M := DCx0 ,T
D := diag √ ,
λj (x0 ) 1≤j≤N
condition (1.7) holds. Indeed, owing to (1.8), it becomes apparent that
M Ax0 M T = DCxT0 Ax0 Cx0 D = DJx0 D = I.
This concludes the proof.
As a consequence from Proposition 1.1, the next result holds.
Theorem 1.1 (Classical minimum principle). Suppose the differen-

tial operator L is elliptic in Ω with c ≥ 0, and u ∈ C 2 (Ω) satisfies
Lu(x) > 0 for every x ∈ Ω.
Then, u cannot attain a non-positive local minimum in Ω.
Proof. On the contrary, assume that there exist x0 ∈ Ω and R > 0 such
that BR (x0 ) ⊂ Ω,
m := u(x0 ) ≤ 0,
and
u(x0 ) ≤ u(x) for all x ∈ BR (x0 ).
By Proposition 1.1, there exists a linear change of coordinates y = M x such
that
Lp u|x=x = −∆v(y0 ),
0
where
v(y) := u(x), y0 := M x0 .
As x0 is a critical point of u, necessarily ∇x u(x0 ) = 0 and, hence,
0 < Lu(x0 ) = Lp u(x0 ) + c(x0 )u(x0 ) = −∆v(y0 ) + c(x0 )u(x0 ).
Therefore,
∆v(y0 ) < c(x0 )u(x0 ) ≤ 0
because c(x0 ) ≥ 0 and u(x0 ) ≤ 0. This is impossible, for as y0 is a local
minimum of v and, hence, ∆v(x0 ) ≥ 0.
1.2 Minimum principle of E. Hopf
The following result extends Theorem 1.1 to cover the more general case
when Lu ≥ 0 vanishes somewhere in Ω.
Theorem 1.2 (Minimum principle of E. Hopf ). Suppose L is uni-

formly elliptic in Ω with c ≥ 0, and u ∈ C 2 (Ω) satisfies
Lu ≥ 0 in Ω, and m := inf u ∈ (−∞, 0].
Ω
Then, either u = m in Ω, or u(x) > m for all x ∈ Ω. In other words, u
cannot attain m in Ω, unless u = m in Ω.
In particular, when Ω is bounded and u ∈ C 2 (Ω) ∩ C(Ω̄), then
inf u = inf u = m.
Ω̄ ∂Ω
Throughout this book, a function u ∈ C 2 (Ω) is said to be a harmonic

function of L in Ω if
Lu = 0 in Ω,
while it is said to be superharmonic if
Lu ≥ 0 in Ω,
and subharmonic when −u is superharmonic.
According to this terminology, Theorem 1.2 establishes that no non-
constant superharmonic function u can reach a non-positive absolute mini-
mum in Ω. Consequently, interchanging u by −u, no non-constant subhar-
monic function can attain a non-negative absolute maximum in Ω. There-
fore, no non-constant harmonic function can attain its absolute maximum
or its absolute minimum in Ω.
Proof. It proceeds by contradiction. Suppose there exist x0 , x1 ∈ Ω such
that
m = u(x0 ) = inf u ≤ 0, u(x1 ) > m. (1.9)
Ω
Establishing a contradiction concludes the proof. Let γ ∈ C([0, 1]; Ω) be a

continuous curve connecting x0 with x1 in Ω, i.e., such that γ(t) ∈ Ω for
each t ∈ [0, 1] and
γ(0) = x0 , γ(1) = x1 .
According to (1.9), by the continuity of t 7→ u(γ(t)), there exists tm ∈ [0, 1)
such that
u(γ(tm )) = m and u(γ(t)) > m for all t ∈ (tm , 1],
i.e., y0 := γ(tm ) is the first point along the arc of curve γ from x1 to x0
where u attains m. Note that tm = 0 if u(γ(t)) > m for every t ∈ (0, 1].
Though in such case y0 = x0 , in general, y0 ̸= x0 . This is the situation
illustrated by Figure 1.1. Now, set
Tray γ := {γ(t) : 0 ≤ t ≤ 1}, d := dist (Tray γ, ∂Ω) > 0,
and pick
y1 ∈ {γ(t) : tm < t < 1}
such that
|y0 − y1 | < d/2.
x1
y0 y1
x
0
Fig. 1.1 The construction of y0 and y1
By construction, u(y1 ) > m. Thus, by the continuity of u, there exists

r > 0 such that
u(x) > m for each x ∈ Br (y1 ). (1.10)
Moreover, since u(y0 ) = m, necessarily
r ≤ |y0 − y1 | < d/2. (1.11)
Let ρ denote the maximal r > 0 satisfying (1.10). By (1.11), ρ > 0 is well
defined and it satisfies
ρ ≤ |y0 − y1 | < d/2. (1.12)
In particular, B̄ρ (y1 ) ⊂ Ω (see Figure 1.2). By the continuity of u and the
maximality of ρ, there exists
y2 ∈ ∂Bρ (y1 )
such that
u(y2 ) = m; (1.13)
we might take y2 = y0 if ρ = |y0 −y1 |. Subsequently, we consider the middle
point of the segment linking y1 to y2 ,
y1 + y2
z := ,
2
and the ball B ρ2 (z). This ball is tangent to Bρ (y1 ) at y2 and it satisfies
B̄ ρ2 (z) \ {y2 } ⊂ Bρ (y1 ) (1.14)
(see Figure 1.2). As u(x) > m for all x ∈ Bρ (y1 ), (1.14) implies
u(x) > m for all x ∈ B̄ ρ2 (z) \ {y2 }. (1.15)
y
1
y0
z
y2
Fig. 1.2 The construction of B ρ (y2 )

4
Finally, consider the ball of radius ρ4 centered at y2 , B ρ4 (y2 ). Figure 1.2

sketches the previous construction. Note that (1.12) implies
B̄ ρ4 (y2 ) ⊂ Ω
and, hence, by (1.2), the coefficients of L are bounded in B̄ ρ4 (y2 ). The rest
of the proof consists in constructing a non-constant function
( )
w ∈ C 2 B̄ ρ4 (y2 )
such that
w(y2 ) = m, (1.16)
w(x) > m for every x ∈ ∂B ρ4 (y2 ), (1.17)

and
Lw(x) > 0 for every x ∈ B ρ4 (y2 ). (1.18)
Thanks to (1.16) and (1.17), w attains its absolute minimum (necessarily
below m = w(y2 ) ≤ 0) in B ρ4 (y2 ). According to Theorem 1.1, such a
function w cannot satisfy (1.18). This contradiction concludes the proof of
the theorem.
Subsequently, we consider the functions
ρ2
v(x) := e−α|x−z| − e−α
2
4 , x ∈ RN , (1.19)
and
w(x) := u(x) − ϵv(x), x ∈ Ω,
for some appropriate constants α > 0 and ϵ > 0 to be chosen later. By
construction, w ∈ C 2 (B̄ ρ4 (y2 )). Thus, to complete the proof of the theorem
it suffices to show that there exist ϵ > 0 and α > 0 for which w satisfies
(1.16), (1.17) and (1.18).
Since |y2 − z| = ρ/2, we have that v(y2 ) = 0 and, hence, (1.13) implies
w(y2 ) = u(y2 ) − ϵv(y2 ) = u(y2 ) = m.
Thus, (1.16) holds.
Subsequently, we will prove that, for sufficiently large α > 0,
Lv(x) < 0 for every x ∈ B̄ ρ4 (y2 ). (1.20)
Since Lu ≥ 0 in Ω, (1.20) implies that
Lw(x) = Lu(x) − ϵLv(x) > 0
for all x ∈ B̄ ρ4 (y2 ) and ϵ > 0, and, consequently, (1.18) holds. Indeed, for
each j ∈ {1, ..., N } and x ∈ RN , we find from (1.19) that
∂v
(x) = −2α (xj − zj ) e−α|x−z| ,
2
∂xj
∂2v [ ]
(x) = 4α2 (xj − zj )2 − 2α e−α|x−z| ,
2
2
∂xj
where xi and zi , i ∈ {1, ..., N }, stand for the i-th coordinates of x and z,
respectively. Moreover, for every i, j ∈ {1, ..., N }, with i ̸= j, and x ∈ RN ,
∂2v
(x) = 4α2 (xi − zi ) (xj − zj ) e−α|x−z|
2
∂xi ∂xj
and hence,
∑
N
∂2v ∑ N
∂v
Lv(x) = − aij (x) (x) + bj (x) (x) + c(x)v(x)
i,j=1
∂xi ∂xj j=1
∂xj

ρ2
 ∑N
= −c(x)e−α 4 + −4α2 aij (x)(xi − zi )(xj − zj )

i,j=1

∑
N 
[ajj (x) − bj (x)(xj − zj )] + c(x) e−α|x−z| .
2
+2α

j=1
As L is uniformly elliptic in Ω, there exists a constant µ > 0 such that

∑
N
aij (x)(xi − zi )(xj − zj ) ≥ µ |x − z|2 for all x ∈ Ω,
i,j=1
and hence,
∑
N
ρ2
aij (x)(xi − zi )(xj − zj ) ≥ µ for all x ∈ B ρ4 (y2 ),
i,j=1
16
because x ∈ B ρ4 (y2 ) implies |x − z| ≥ ρ4 (see Figure 1.2). Moreover, since
B̄ ρ4 (y2 ) is a compact subset of Ω, it follows from (1.2) that there exist two
constants C1 > 0 and C2 > 0 such that
∑
N
|ajj (x) − bj (x)(xj − zj )| ≤ C1 , |c(x)| ≤ C2 ,
j=1
for all x ∈ B̄ ρ4 (y2 ). Thus, for every α > 0 and x ∈ B̄ ρ4 (y2 ),

( )
ρ2
Lv(x) ≤ −µ α2 + 2αC1 + C2 e−α|x−z| ,
2
4
since c ≥ 0. Therefore, (1.20) holds for sufficiently large α > 0. Throughout
the rest of this proof, we assume that α > 0 has been chosen in this way.
To complete the proof of the theorem it remains to show that (1.17)
holds for sufficiently small ϵ > 0. Indeed, setting
S1 := ∂B ρ4 (y2 ) ∩ B̄ ρ2 (z), S2 := ∂B ρ4 (y2 ) \ S1 ,
it is apparent that S1 is a compact subset of B̄ ρ2 (z) \ {y2 } (see Figure 1.3)
and, in particular, according to (1.15), we have that
u(x) > m for all x ∈ S1 .
Thus, since u is continuous, there exists ξ > 0 such that
u(x) ≥ m + ξ for all x ∈ S1 . (1.21)
Subsequently, we consider any ϵ > 0 satisfying
ξ
0<ϵ< ρ2
.
1 − e−α 4
According to (1.19), it is apparent that


 v(x) > 0 if and only if x ∈ B ρ2 (z),




v(x) = 0 if and only if x ∈ ∂B ρ2 (z), (1.22)





 v(x) < 0 if and only if x ̸∈ B̄ ρ (z).
2
y1
z
S1
S2
y
2
Fig. 1.3 S1 is the thicker portion of ∂B ρ (y2 )

4
By (1.22), for every x ∈ S1 , we have that

ρ2 ρ2
0 ≤ v(x) = e−α|x−z| − e−α < 1 − e−α
2
4 4
and hence, by the choice of ϵ,

( )
ρ2
0 ≤ ϵv(x) < ϵ 1 − e−α 4 < ξ .
Consequently, (1.21) implies that

w(x) = u(x) − ϵv(x) > m + ξ − ξ = m for all x ∈ S1 .
Finally, by (1.22), for every x ∈ S2 := ∂B ρ4 (y2 ) \ S1 we have that v(x) < 0
and, consequently,
w(x) = u(x) − ϵv(x) > u(x) ≥ inf u = m.
Ω
Thus, since
S1 ∪ S2 = ∂B ρ4 (y2 ),
condition (1.17) is fulfilled, and the proof is complete.
Remark 1.1.
(a) Although condition c ≥ 0 is not strictly necessary for the validity of
Theorem 1.2, when c is negative somewhere in Ω the theorem might not
be true.
(b) Suppose L is uniformly elliptic in Ω with c ≥ 0, u ∈ C 2 (Ω) satisfies

Lu ≥ 0 in Ω, and there exist x0 ∈ Ω and r > 0 such that u(x0 ) ≤ 0,
B̄r (x0 ) ⊂ Ω, and
u(x0 ) ≤ u(x) for every x ∈ Br (x0 ).
Then, by Theorem 1.2, u = u(x0 ) in Br (x0 ). Consequently, u must be
locally constant at any non-positive local minimum in Ω.
(c) Suppose c = 0. Then, for every m ∈ R , u − m is superharmonic if u is
superharmonic, since Lm = 0. Consequently, in this case, Theorem 1.2
holds independently of the sign of m. Indeed, if Lu ≥ 0, then u−m ≥ 0
is superharmonic and, therefore, by Theorem 1.2, u(x) > m for all
x ∈ Ω if u ̸= m.
1.3 Interior sphere properties
The next definition introduces some important geometrical properties of

∂Ω that will be used throughout this book.
Definition 1.2. Ω is said to satisfy the interior sphere property at a
single point x ∈ ∂Ω if there are zx ∈ Ω and r > 0 such that
|x − zx | = r, Br (zx ) ⊂ Ω.
Now, let Γ0 be a closed and open subset of ∂Ω. Then:
(a) Ω is said to satisfy the uniform interior sphere property on Γ0 if
there exists r > 0 such that for every x ∈ Γ0 there is a point zx ∈ Ω
for which
|x − zx | = r, Br (zx ) ⊂ Ω.
In such case, it is said that Ω satisfies the uniform interior sphere
property on Γ0 with parameter r. When Γ0 = ∂Ω, it is simply said that
Ω satisfies the uniform interior sphere property.
(b) Ω is said to satisfy the uniform interior sphere property in the
strong sense on Γ0 if there exists r > 0 such that for every z ∈ Ω
with dist (z, Γ0 ) ≤ r there is a point xz ∈ Γ0 for which
( )
z − xz
dist (z, ∂Ω) = |z − xz |, Br xz + r ⊂ Ω.
|z − xz |
In such case, it is said that Ω satisfies the uniform interior sphere
property in the strong sense on Γ0 with parameter r. When Γ0 = ∂Ω,
it is simply said that Ω satisfies the uniform interior sphere property in
the strong sense.
Note that if Ω satisfies the interior sphere property at x ∈ ∂Ω, then

there exist zx ∈ Ω and r > 0 such that
Br (zx ) ⊂ Ω, B̄r (zx ) ∩ ∂Ω = {x}. (1.23)
Indeed, by definition, there are z̃x ∈ Ω and r̃ > 0 such that
|x − z̃x | = r̃, Br̃ (z̃x ) ⊂ Ω.
Obviously, (1.23) holds by making the choice
r := r̃/2, zx := (x + z̃x )/2.
If Ω satisfies the uniform interior sphere property on Γ0 , then it satisfies the
interior sphere property at every x ∈ Γ0 , of course. Moreover, the following
result holds.
Lemma 1.1. Let Ω ⊂ RN , N ≥ 1, be an open set with boundary ∂Ω, and

Γ0 a closed and open subset of ∂Ω. Then, the following assertions are true:
(a) If Ω satisfies the uniform interior sphere property in the strong sense
on Γ0 with parameter r > 0, then it also satisfies the uniform interior
sphere property on Γ0 with the same parameter r.
(b) If Γ0 is of class C 1 (see Definition 1.5 and Theorem 1.8, if necessary)
and it satisfies the uniform interior sphere property on Γ0 with param-
eter r, then it also satisfies the uniform interior sphere property in the
strong sense on Γ0 with the same parameter r.
Therefore, Definitions 1.2(a) and (b) are equivalent when Γ0 is of class C 1 .
Proof. Suppose Ω satisfies the uniform interior sphere property in the

strong sense on Γ0 with parameter r > 0. According to Definition 1.2(b),
for every z ∈ Ω with dist(z, Γ0 ) = r there exists xz ∈ Γ0 such that
( )
z − xz
dist (z, ∂Ω) = |z − xz |, Br xz + r ⊂ Ω.
|z − xz |
As Γ0 ⊂ ∂Ω and xz ∈ Γ0 , this implies
r = dist(z, Γ0 ) ≥ dist (z, ∂Ω) = |z − xz | ≥ dist(z, Γ0 ) = r
and hence,
r = dist(z, Γ0 ) = dist (z, ∂Ω) = |z − xz |.
Thus,
z − xz z − xz
xz + r = xz + r =z
|z − xz | r
and, therefore,
Br (z) ⊂ Ω.
Let x ∈ Γ0 and n0 ∈ N such that n0 ≥ 1/r, and for any integer n ≥ n0 pick
zn ∈ Ω such that
1 1
|x − zn | ≤ ≤ ≤ r.
n n0
As
dist(zn , Γ0 ) ≤ |x − zn | ≤ r,
it follows from Definition 1.2(b) that, for every n ≥ n0 , there is a point
xn := xzn ∈ Γ0
for which
( )
zn − xn
dist (zn , ∂Ω) = |zn − xn |, Br xn + r ⊂ Ω.
|zn − xn |
Setting
zn − xn
z̃n := xn + r , n ≥ n0 ,
|zn − xn |
we have that Br (z̃n ) ⊂ Ω, |z̃n − xn | = r, and xn ∈ Γ0 . Hence,
dist(z̃n , Γ0 ) = dist(z̃n , ∂Ω) = r, n ≥ n0 .
Thus, as the set
Γr := {z ∈ Ω : dist(z, Γ0 ) = dist(z, ∂Ω) = r}
is compact, there exists z̃ω ∈ Γr such that along some subsequence of
{z̃n }n≥1 , say {z̃nm }m≥1 , the following holds
lim z̃nm = z̃ω .
m→∞
By construction, it becomes apparent that

Br (z̃ω ) ⊂ Ω,
because z ∈ Br (z̃ω ) implies that z ∈ Br (z̃nm ) ⊂ Ω for sufficiently large m.
Moreover,
1
znm ∈ Br (z̃nm ), |x − znm | ≤ , m ≥ 1,
nm
and hence, letting m → ∞ shows that
lim znm = x, x ∈ B̄r (z̃ω ).
m→∞
Consequently, setting zx := z̃ω , we conclude that

|x − zx | = r, Br (zx ) ⊂ Ω,
because x ∈ Γ0 . Therefore, Ω satisfies the uniform interior sphere property
on Γ0 with parameter r.
Now, suppose that Γ0 is of class C 1 , as discussed in Definition 1.5, and
that Ω satisfies the uniform interior sphere property on Γ0 with parameter
r. Then, by Theorem 1.8, it is apparent that Ω has a tangent hyperplane
at every x0 ∈ Γ0 .
Let z ∈ Ω such that
dist(z, Γ0 ) ≤ r
and pick a point x ∈ Γ0 with
dist(z, Γ0 ) = |z − x| ≤ r.
Necessarily, the tangent hyperplane to Γ0 at x must be orthogonal to z − x.
By Definition 1.2(a), there exists zx ∈ Ω such that
|x − zx | = r, Br (zx ) ⊂ Ω.
As Br (zx ) ⊂ Ω touches Γ0 at x, the vector zx − x must be orthogonal to
∂Ω at x and
dist(zx , ∂Ω) = dist(zx , Γ0 ) = |zx − x| = r.
Consequently, z must lie in the line segment [x, zx ] and hence,
z−x
zx = x + r .
|z − x|
Therefore,
( )
z−x
dist (z, ∂Ω) = |z − x|, Br x + r ⊂ Ω,
|z − x|
and, consequently, Ω satisfies the uniform interior sphere property in the
strong sense on Γ0 with parameter r. This ends the proof.
According to Lemma 1.1, in the special case when Ω is of class C 1 , Ω
satisfies the uniform interior sphere property with parameter r > 0 if and
only if it satisfies this property in the strong sense with the same parameter
r. But, in the general case when Ω is not of class C 1 , the fact that Ω satisfies
the uniform interior sphere property with parameter r > 0 and a uniform
interior sphere property in the strong sense do not necessarily entail the
strong condition to be satisfied with the same r > 0. As a counterexample,

consider two points z0 , z1 ∈ RN , N ≥ 2, and R > 0 with
dist(z0 , z1 ) < 2R
for which 2R − dist(z0 , z1 ) is sufficiently small. Then, the union of the two
balls
Ω := BR (z0 ) ∪ BR (z1 ),
is not of class C , it satisfies the uniform interior sphere property with
1
parameter r = R (optimal), as well as the corresponding property in the

strong sense for sufficiently small r ∈ (0, R), but it does not satisfy the
property in the strong sense with parameter r = R. Indeed, for every
x ∈ ∂Ω there exists i ∈ {0, 1} such that x ∈ ∂BR (zi ) and, hence, choosing
zx := zi , we have that
|x − zx | = |x − zi | = R, BR (zx ) = BR (zi ) ⊂ Ω.
Consequently, Ω satisfies the uniform interior sphere property with parame-
ter r = R. Obviously, this property cannot be satisfied if r > R and, conse-
quently, r = R is optimal. Now, pick two points x0 , x1 ∈ ∂BR (z0 )∩∂BR (z1 )
x
0
z
z0 z1
x1
Fig. 1.4 The special case Ω := BR (z0 ) ∪ BR (z1 )
such that
|x0 − x1 | = diam [B̄R (z0 ) ∩ B̄R (z1 )],
as illustrated by Figure 1.4. Note that |x0 − x1 | < 2R. Then, setting
x0 + x1
z := ,
2
we have that
|x0 − x1 |
|z − x0 | = |z − x1 | =
2
and it can be easily seen that Ω satisfies the uniform interior sphere property
in the strong sense with
|x0 − x1 |
r := < R,
2
though, obviously, the property fails for greater values of r.
In general, a domain Ω satisfying the uniform interior sphere property
does not necessarily satisfy it in the strong sense, as the following example
shows. Let Ω ⊂ R2 be the open set
Ω := B2 (0) \ {(0, 0), (1/n, 0) : n ≥ 1} , (1.24)
which consists of the ball B2 (0) perforated by (0, 0) and the sequence
(1/n, 0), n ≥ 1, which converges to (0, 0) as n → ∞. As
∂Ω = {x ∈ R2 : |x| = 2} ∪ {(0, 0), (1/n, 0) : n ≥ 1} ,
Ω cannot satisfy a uniform interior sphere property in the strong sense,
because of the special structure of its boundary around (0, 0). To prove
this we will argue by contradiction. So, suppose Ω satisfies the uniform
interior sphere property in the strong sense with parameter r > 0 and pick
an integer n ≥ 1 such that
1 1 1
− = < r,
n n+1 n(n + 1)
and a real number ϵ ∈ R satisfying
1 1 1 1
<ϵ< , ϵ− < − ϵ.
n+1 n n+1 n
Then, z := (ϵ, 0) ∈ Ω and
1
dist(z, ∂Ω) = ϵ − .
n+1
Moreover, xz := (1/(n + 1), 0) provides us with the unique point of ∂Ω such
that
dist(z, ∂Ω) = |z − xz |.
Thus, according to Definition 1.2(b), we should have that
( )
z − xz
Br xz + r ⊂ Ω,
|z − xz |
which is impossible because
( )
z − xz 1 1 1
xz + r = + r, 0 , +r > ,
|z − xz | n+1 n+1 n
and hence,
( )
z − xz
(1/n, 0) ∈ Br xz + r ∩ ∂Ω = ∅.
|z − xz |
Nevertheless, Ω satisfies the uniform interior sphere property with param-
eter r := 1/2. Indeed, for every x ∈ ∂B2 (0) the point zx := 34 x satisfies
1
zx ∈ Ω, |x − zx | = , B 12 (zx ) ⊂ Ω.
2
Moreover, setting
z(0,0) := (0, 1/2), z(1/n,0) := (1/n, 1/2), n ≥ 1,
it is apparent that
(0, 0) ∈ ∂B 12 (z(0,0) ), (1/n, 0) ∈ ∂B 21 (z(1/n,0) ), n ≥ 1,
and
∪
B 12 (z(0,0) ) ∪ B 12 (z(1/n,0) ) ⊂ Ω.
n≥1
Therefore, Ω satisfies the uniform interior sphere condition with r = 1/2.

Theorem 1.9 in the Appendix of this chapter shows that Ω satisfies the
uniform interior sphere property in the strong sense on Γ0 whenever Γ0 is
of class C 2 . Undoubtedly, this is the most paradigmatic class of domains
satisfying Definition 1.2(b).
Suppose Ω is of class C 2 around x0 ∈ ∂Ω. Then, owing to the proof of
Theorem 1.8, Ω satisfies a uniform interior sphere property around x0 and
any interior sphere Br (z) at x0 must be tangent to the tangent hyperplane
of ∂Ω at x0 . Thus, the vector x0 − z is orthogonal to ∂Ω at x0 and, hence,
x0 − z
n :=
|x0 − z|
is the outward unit normal at x0 . Note that the tangent hyperplane to
any interior sphere Br (z) at x0 does not vary with (r, z) and, hence, n is
uniquely determined. In this context, a given ν ∈ RN \ {0} is said to be an
outward pointing vector at x0 whenever
⟨ν, n⟩ > 0 (1.25)
(see Figure 1.5). As ∂Ω and n are orthogonal at x0 , (1.25) implies that
x0 + tν ∈ RN \ Ω̄ for sufficiently small t > 0.
Consequently, the nomenclature of outward pointing vector at x0 ∈ ∂Ω is
unambiguous. Figure 1.6 illustrates the existence of non-smooth boundaries
n
υ
Ω x0
Fig. 1.5 Smooth boundary around x0 ∈ ∂Ω
Ω
x0
Fig. 1.6 Non-smooth boundary at x0 ∈ ∂Ω
satisfying an interior sphere property. In these cases, the previous concept

of outward normal does not make sense, but yet a concept of outward
pointing vector can be given through the next definition.
Definition 1.3. Suppose Ω satisfies an interior sphere property at x0 ∈ ∂Ω.

Then, ν ∈ RN \ {0} is said to be an outward pointing vector at x0 if
x0 + tν ∈ RN \ Ω̄ for sufficiently small t > 0,
and there exist z ∈ Ω and r > 0 such that
|z − x0 | = r, Br (z) ⊂ Ω, ⟨ν, x0 − z⟩ > 0.
1.4 Boundary lemma of E. Hopf
The main result of this section improves Theorem 1.2 by establishing that
any non-constant superharmonic function u ∈ C 2 (Ω) ∩ C(Ω̄) of L in Ω must
decay linearly at any point x0 ∈ ∂Ω where

u(x0 ) = inf u ≤ 0
Ω̄
along any outward pointing vector for which u admits a directional deriva-
tive. Most precisely, it reads as follows.
Theorem 1.3 (Boundary lemma of E. Hopf ). Suppose the differen-

tial operator L is uniformly elliptic in Ω with c ≥ 0,
aij , bj , c ∈ L∞ (Ω), i, j ∈ {1, ..., N }, (1.26)
and u ∈ C 2 (Ω) is a non-constant function satisfying
Lu ≥ 0 in Ω, and m := inf u ∈ (−∞, 0].
Ω
Assume, in addition, that there exist x0 ∈ ∂Ω and R > 0 such that
u(x0 ) = m, u ∈ C(BR (x0 ) ∩ Ω̄),
and Ω satisfies an interior sphere property at x0 .
Then, for any outward pointing vector ν ∈ RN \ {0} at x0 for which
∂u
(x0 ) := lim ⟨ν, ∇u(x)⟩
∂ν x∈Ω
x→x0
exists, necessarily
∂u
(x0 ) < 0. (1.27)
∂ν
Thanks to Theorem 1.2, under the assumptions of Theorem 1.3, we have
that u(x) > m for all x ∈ Ω, as we are assuming that u is not a constant.
Therefore, Theorem 1.3 establishes that any non-constant superharmonic
function u(x) decays linearly towards its minimum, m = u(x0 ), as x ∈ Ω
approximates x0 ∈ ∂Ω, if m ≤ 0. In Figure 1.7 we have represented the
profile of u along the straight line passing through x0 in the direction of
the vector ν; we are denoting
∂u
tan α := (x0 ) < 0.
∂ν
Proof. Let z ∈ Ω and a sufficiently small r > 0 such that

Br (z) ⊂ Ω, B̄r (z) ∩ ∂Ω = {x0 }, ⟨ν, x0 − z⟩ > 0,
and
u ∈ C 2 (Br (z)) ∩ C(B̄r (z)).
0
α
m ν
x
0
Fig. 1.7 Linear decay guaranteed by (1.27)
Now, consider the domain

D := B r2 (x0 ) ∩ Br (z)
and set
S1 := ∂B r2 (x0 ) ∩ B̄r (z), S2 := B r2 (x0 ) ∩ ∂Br (z).
Then, ∂D = S1 ∪ S2 and S1 is a compact subset of Ω (see Figure 1.8).
By (1.26), the proof of (1.20) can be adapted to show that for sufficiently
large α > 0 the function
v(x) := e−α|x−z| − e−αr ,
2 2
x ∈ RN ,
satisfies
Lv(x) < 0 for all x ∈ Br (z). (1.28)
Throughout the remaining of the proof we suppose that α has been chosen
to satisfy (1.28). Note that (1.26) is needed in order to obtain (1.28),
because x0 ∈ ∂Br (z) ∩ ∂Ω; in the Proof of Theorem 1.2 the sign of Lv was
fixed within a compact subset of Ω, where the coefficients of L are always
bounded under condition (1.2). Also, note that v(x) > 0 if x ∈ Br (z), while
v(x) = 0 if x ∈ ∂Br (z) and v(x) < 0 for all x ∈ RN \ B̄r (z).
Thanks to Theorem 1.2,
u(x) > m for each x ∈ Ω. (1.29)
Thus, since S1 is a compact subset of Ω, there exists ξ > 0 such that
u(x) ≥ ξ + m for all x ∈ S1 . (1.30)
Now, fix ϵ > 0 such that
ξ
0<ϵ<
1 − e−αr2
Ω r
z
D
x0
ν
Fig. 1.8 The construction of D
and consider the auxiliary function

w(x) := u(x) − ϵv(x), x ∈ Ω.
Note that ∂w
∂ν (x0 ) is well defined if
∂u
∂ν (x0 ) exists, because v ∈ C ∞ (RN ).
Moreover, in such case,
∂w ∂u ∂v
(x0 ) = (x0 ) − ϵ (x0 ). (1.31)
∂ν ∂ν ∂ν
It follows from (1.28) that
Lw(x) = Lu(x) − ϵLv(x) ≥ −ϵLv(x) > 0 for all x ∈ D,
since D ⊂ Br (z) ⊂ Ω. Moreover, for each x ∈ S1 we have that
0 ≤ v(x) < 1 − e−αr
2
and hence,
( )
0 ≤ ϵv(x) < ϵ 1 − e−αr < ξ.
2
Thus, we find from (1.30) that

w(x) = u(x) − ϵv(x) ≥ m + ξ − ϵv(x) > m for all x ∈ S1 .
Further, since v(x) = 0 for each x ∈ S2 ⊂ ∂Br (z), we have that
w(x) = u(x) − ϵv(x) = u(x).
Thus, by (1.29),
w(x) = u(x) > m for each x ∈ S2 \ {x0 } ⊂ Ω
while
w(x0 ) = u(x0 ) = m.
Consequently, w ∈ C (D) ∩ C(D̄) and it satisfies the following properties:
2
(1) Lw(x) > 0 for all x ∈ D,

(2) w(x) > m for all x ∈ ∂D \ {x0 },
(3) w(x0 ) = m.
Therefore, Theorem 1.2 implies that
w(x) > m for every x ∈ D̄ \ {x0 }
and, necessarily,
∂w
(x0 ) ≤ 0. (1.32)
∂ν
It should be noted that w(x) > m, x ∈ D, cannot be obtained straight-
away from the definition of w, as v(x) > 0 for all x ∈ D. By (1.31) and
(1.32) it becomes apparent that
∂u ∂v
(x0 ) ≤ ϵ (x0 ) = ϵ⟨ν, ∇v(x0 )⟩.
∂ν ∂ν
On the other hand, by a direct calculation,
∇v(x0 ) = −2α(x0 − z)e−αr
2
and, consequently,
∂u
(x0 ) ≤ ϵ⟨ν, ∇v(x0 )⟩ = −2αϵe−αr ⟨ν, x0 − z⟩ < 0,
2
∂ν
because ⟨ν, x0 − z⟩ > 0. This shows (1.27) and ends the proof.
1.5 Positivity properties of super-harmonic functions
The following corollaries from Theorems 1.2 and 1.3, provide us with some
important positivity properties of the superharmonic functions of L in Ω.
Theorem 1.4. Suppose Ω is bounded, L is uniformly elliptic in Ω with

c ≥ 0, and u ∈ C 2 (Ω) ∩ C(Ω̄) \ {0} satisfies
{
Lu ≥ 0 in Ω,
u≥0 on ∂Ω.
Then,
u(x) > 0 for all x ∈ Ω.
If, in addition, u ∈ C 1 (Ω̄) and (1.26) is satisfied, then, for every
x ∈ ∂Ω ∩ u−1 (0)
where Ω satisfies the interior sphere property, and any outward pointing
vector ν at x, we have that
∂u
(x) < 0.
∂ν
Proof. Set
m := inf u .
Ω̄
If m > 0, then u(x) > 0 for all x ∈ Ω̄ and we are done. So, suppose m ≤ 0.
Then, thanks to Theorem 1.2,
inf u = inf u = m ≤ 0
Ω̄ ∂Ω
and, hence, m = 0, since u|∂Ω ≥ 0. Moreover, u(x) > 0 for all x ∈ Ω,

because u ̸= 0. The remaining assertions follow from Theorem 1.3.
Corollary 1.1. Suppose Ω is bounded, L is uniformly elliptic in Ω with

c ≥ 0, f ∈ C(Ω) and g ∈ C(∂Ω). Then, the boundary value problem
{
Lu = f in Ω,
(1.33)
u=g on ∂Ω,
admits, at most, a unique solution u ∈ C 2 (Ω) ∩ C(Ω̄). In particular, u = 0

is the unique function u ∈ C 2 (Ω) ∩ C(Ω̄) solving
{
Lu = 0 in Ω,
(1.34)
u=0 on ∂Ω.
Proof. On the contrary, assume that there exist u, v ∈ C 2 (Ω) ∩ C(Ω̄),

u ̸= v, solving (1.33). Then, the auxiliary function
w := u − v ∈ C 2 (Ω) ∩ C(Ω̄) \ {0}
provides us with a solution of (1.34) and, according to Theorem 1.4, w(x) >
0 for every x ∈ Ω. Similarly, −w provides us with a solution of (1.34) such
that −w(x) > 0 for each x ∈ Ω. This contradiction ends the proof.
1.6 Uniform decay property of E. Hopf
Essentially, the next result provides us with a lower estimate for the decay
rate of all positive superharmonic functions of L in Ω at the points of ∂Ω
where they vanish.
Theorem 1.5 (Uniform decay property of E. Hopf ). Let L be a

uniformly elliptic operator in Ω whose coefficients satisfy (1.26). Then,
for each R > 0 there exists a constant
M := M (L, R) > 0
such that for every x0 ∈ Ω with BR (x0 ) ⊂ Ω and any function u ∈

C 2 (BR (x0 )) ∩ C(B̄R (x0 )) satisfying
u(x) > 0 for all x ∈ BR (x0 ) (1.35)
and
Lu(x) ≥ 0 for all x ∈ BR (x0 ) (1.36)
the following estimate holds:

( )
u(x) ≥ M min u dist (x, ∂BR (x0 )) ∀ x ∈ B̄R (x0 ). (1.37)
B̄ R (x0 )
2
Proof. Let R > 0 and x0 ∈ Ω such that BR (x0 ) ⊂ Ω, and suppose u ∈

C 2 (BR (x0 )) ∩ C(B̄R (x0 )) satisfies (1.35) and (1.36). Consider the auxiliary
functions
2
−|x−x0 |2 )
E(x) := eα(R , v(x) := E(x) − 1, x ∈ RN ,
where α > 0 is a constant to be chosen later. The function v satisfies


 v(x) > 0 if and only if |x − x0 | < R,
v(x) = 0 if and only if |x − x0 | = R,

v(x) < 0 if and only if |x − x0 | > R.
Now, set
{ }
R
D := x ∈ RN : < |x − x0 | < R ⊂Ω
2
and
c+ := max{c, 0} ≥ c, L+ := L − c + c+ .
Then,
∑
N
∂2E ∑ ∂E
N
L+ v = − aij + bj + c+ v
i,j=1
and, differentiating and rearranging terms, shows that, for every x ∈ Ω,


 ∑N
L+ v(x) = −4α2 aij (x)(xi − x0i )(xj − x0j )

i,j=1

∑N 
+2α [ajj (x) − bj (x)(xj − x0j )]+c+ (x) E(x)−c+ (x)

j=1
{
= −4α2 (x − x0 )T Ax (x − x0 ) + 2α[tr Ax − ⟨b(x), x − x0 ⟩]
}
+ c+ (x) E(x) − c+ (x),
where Ax is the matrix of the principal coefficients of L (see (1.3)), tr Ax
stands for the trace of Ax , and b := (b1 , ..., bN ). Let µ > 0 be the ellipticity
constant of L in Ω (see Definition 1.1). Then, for every x ∈ D,
R2
(x − x0 )T Ax (x − x0 ) ≥ µ |x − x0 |2 > µ .
4
Moreover,
|tr Ax − ⟨b(x), x − x0 ⟩| ≤ |tr Ax | + |b(x)| R
and hence, thanks to (1.26), there exists a constant
( )
C := C ∥ajj ∥L∞ (Ω) , ∥bj ∥L∞ (Ω) , R = C(L, R) > 0
such that
|tr Ax − ⟨b(x), x − x0 ⟩| ≤ C for all x ∈ D,
independently of x0 . Thus, for every x ∈ D, we have that
( )
L+ v(x) ≤ −α2 µR2 + 2αC + ∥c+ ∥L∞ (Ω) E(x) − c+ (x)
and, therefore, there exists
α := α(L, R) > 0
such that
L+ v(x) < 0 for all x ∈ D. (1.38)
Throughout the rest of this proof, we suppose α > 0 has been chosen to
satisfy (1.38).
According to (1.35), u(x) > 0 for each x ∈ B̄ R (x0 ) and hence, by the
2
continuity of u,
uR := inf u > 0.
B̄ R (x0 )
2
Subsequently, we consider the auxiliary function

u
w := u − ϵv, ϵ := αR2R , (1.39)
e −1
which is well defined in B̄R (x0 ). Thanks to (1.39), it is apparent that
2
w ≥ uR − ϵ(eαR − 1) = 0 in B̄ R (x0 ),
2
because
2
u ≥ uR and v ≤ v(x0 ) = eαR − 1.
In particular,
w≥0 on ∂B R (x0 ).
2
Moreover, since v = 0 on ∂BR (x0 ) and u ≥ 0 in B̄R (x0 ), we have that

w≥0 on ∂BR (x0 ).
Consequently,
w≥0 on ∂D = ∂BR (x0 ) ∪ ∂B R (x0 ). (1.40)
2
Furthermore, thanks to (1.38), we find that, in D,

L+ w = L+ u − ϵL+ v > L+ u = (L − c + c+ )u ≥ (c+ − c)u ≥ 0,
because
Lu ≥ 0, c+ ≥ c, and u ≥ 0.
Thus, w is a superharmonic function of L+ in D satisfying (1.40) and,
therefore, Theorem 1.4 yields to
w = u − ϵv ≥ 0 in D. (1.41)
On the other hand, for every x ∈ D, we have that
−|x−x0 |2 )
− 1 = eα(R+|x−x0 |)(R−|x−x0 |) − 1
2
v(x) = eα(R
3
≥ e 2 αR(R−|x−x0 |) − 1 ≥ αR (R − |x − x0 |)
3
2
3
= αR dist (x, ∂BR (x0 )) .
2
Thus, it follows from (1.39) and (1.41) that, for every x ∈ D,

3αR
u(x) ≥ ϵv(x) ≥ ( αR2 ) u dist (x, ∂BR (x0 )). (1.42)
2 e −1 R
Finally, taking into account that, for any x ∈ B̄ R (x0 ),
2
uR
u(x) ≥ uR ≥ dist (x, ∂BR (x0 )),
R
it becomes apparent that (1.42) implies
{ }
3αR 1
u ≥ min ( ), uR dist (·, ∂BR (x0 )) in B̄R (x0 ).
2 eαR2 − 1 R
As the constant
{ }
3αR 1
M (L, R) := min ( ),
2 eαR2 − 1 R
satisfies (1.37) for all admissible u, the proof is complete.
For bounded domains satisfying the uniform interior sphere property in
the strong sense, the following sharp version of Theorem 1.5 holds.
Theorem 1.6. Suppose Ω is a bounded domain satisfying the uniform in-

terior sphere property in the strong sense on a closed and open subset Γ0
of ∂Ω, and L is a uniformly elliptic operator in Ω whose coefficients are
measurable and bounded in Ω.
Let u ∈ C 2 (Ω) ∩ C(Ω̄) be a superharmonic function of L in Ω such that
u(x) > 0 for all x ∈ Ω̄ \ Γ0 .
Then, there exists δ > 0 such that
u(x) ≥ δ dist (x, Γ0 ) for all x ∈ Ω.
Proof. Suppose Ω satisfies the uniform interior sphere property in the

strong sense on Γ0 with parameter R > 0, and consider the compact subset
of Ω̄ defined by
{ }
KR := x ∈ Ω̄ : dist (x, Γ0 ) ≥ R/2 .
Let u ∈ C 2 (Ω) ∩ C(Ω̄) such that Lu ≥ 0 in Ω and u(x) > 0 for every
x ∈ Ω̄ \ Γ0 . Then,
uL := min u > 0.
KR
Let x ∈ Ω with dist (x, Γ0 ) ≤ R and consider yx ∈ Γ0 such that

dist (x, Γ0 ) = |x − yx | and BR (x0 ) ⊂ Ω,
where
x − yx
x0 := yx + R .
|x − yx |
Since B̄ R (x0 ) ⊂ KR , we have that
2
min u ≥ uL ,
B̄ R (x0 )
2
and hence, according to Theorem 1.5, there exists a constant M = M (L, R)

(independent of x) such that
u(z) ≥ M uL dist (z, ∂BR (x0 )) for all z ∈ B̄R (x0 ).
( )
Therefore, for every x ∈ Ω ∩ Γ0 + B̄R , we find that
u(x) ≥ M uL dist (x, ∂BR (x0 )) = M uL |x − yx | = M uL dist (x, Γ0 ).
Finally, as
u(x)
η := inf > 0,
x∈KR dist (x, Γ0 )
it becomes apparent that making the choice
δ := min{η, M uL }
concludes the proof.
As the class of domains for which Theorem 1.6 holds true will play a
significant role throughout this book, it is appropriate to introduce the
following concept.
Definition 1.4. Suppose Ω is a bounded domain, Γ0 is a closed and open

subset of ∂Ω, and L is a uniformly elliptic operator in Ω satisfying (1.26).
It is said that (L, Ω) satisfies the decay property of E. Hopf on Γ0 if
for any superharmonic function u ∈ C 2 (Ω) ∩ C(Ω̄) of L in Ω with u(x) > 0
for all x ∈ Ω̄ \ Γ0 there exists δ > 0 such that
u(x) ≥ δ dist (x, Γ0 ) for all x ∈ Ω. (1.43)
Using this concept, Theorem 1.6 can be equivalently stated as follows.
Corollary 1.2. Suppose Ω is a bounded domain, Γ0 is a closed and open

subset of ∂Ω, and L is a uniformly elliptic operator in Ω whose coefficients
are measurable and bounded in Ω. Then, (L, Ω) satisfies the decay property
of E. Hopf on Γ0 if Ω satisfies the uniform interior sphere property in the
strong sense on Γ0 .
1.7 The generalized minimum principle of M. H. Protter

and H. F. Weinberger
The next result provides us with a sharp generalization of Theorems 1.2

and 1.3 to cover the general case when the coefficient function c(x) is not
necessarily non-negative.
Theorem 1.7 (Generalized minimum principle). Suppose L is a uni-

formly elliptic operator in Ω, and (L, Ω) admits a strictly positive superhar-
monic function h ∈ C 2 (Ω) ∩ C(Ω̄), in the sense that
i) h(x) > 0 for all x ∈ Ω̄,

ii) Lh ≥ 0 in Ω.
Then, for any superharmonic function u ∈ C 2 (Ω) of L in Ω such that

u
m := inf ∈ (−∞, 0], (1.44)
Ω h
either
u(x) > mh(x) for all x ∈ Ω, (1.45)
or
u = mh in Ω. (1.46)
Further, suppose (1.26), (1.45), and the following three conditions:
(a) h ∈ C 1 (Ω̄);
(b) u(x0 ) = mh(x0 ) for some x0 ∈ ∂Ω, and there exists R > 0 such that
u ∈ C(BR (x0 ) ∩ Ω̄);
(c) Ω satisfies the interior tangent sphere property at x0 and there is an
outward pointing vector ν ∈ RN \ {0} for which ∂(u/h)
∂ν (x0 ) exists.
Then,
∂(u/h)
(x0 ) < 0.
∂ν
Remark 1.2. When c ≥ 0, the function h := 1 satisfies (i) and (ii) and,
hence, it provides us with a strict positive superharmonic function of (L, Ω).
Consequently, in this special case, Theorem 1.7 provides us with Theo-
rems 1.2 and 1.3 simultaneously, for as u/h = u. Theorem 1.7 is substan-
tially sharper than these results, because it does not impose any restriction
on the sign of c ∈ L∞ (Ω).
Proof. Subsequently, we set

u(x)
v(x) := for every x ∈ Ω.
h(x)
A direct calculation shows that, in Ω,
∑
N
∂ 2 (hv) ∑ ∂(hv)
N
Lu = L(hv) = − aij + bj + chv
i,j=1
( )
∑N
∂2v ∑N
2∑
N
∂h ∂v
= −h aij +h bj − aij + vLh
i,j=1
∂xi ∂xj j=1
h i=1 ∂xi ∂xj
 ( ) 
∑N 2
∂ v ∑N
2 ∑N
∂h ∂v Lh 
= h − aij + bj − aij + v
i,j=1
∂x i ∂x j j=1
h i=1
∂x i ∂x j h
and hence,
Lu = hLh v in Ω, (1.47)
where
( )
∑
N
∂2 ∑N
2∑
N
∂h ∂ Lh
Lh := − aij + bj − aij + . (1.48)
i,j=1
∂x i ∂xj j=1
h i=1
∂x i ∂x j h
As h ∈ C 2 (Ω) ∩ C(Ω̄), we have that
2∑
N
Lh ∂h
∈ L∞
loc (Ω), bj − aij ∈ L∞
loc (Ω), 1 ≤ j ≤ N.
h h i=1 ∂xi
Moreover,
Lh
≥0 in Ω,
h
and, consequently, Theorems 1.2 and 1.3 can be applied to the operator Lh
defined by (1.48).
Since Lu ≥ 0 in Ω and h(x) > 0 for all x ∈ Ω̄, (1.47) implies that
Lh v ≥ 0 in Ω.
On the other hand, (1.44) can be equivalently expressed in the form
m = inf v ∈ (−∞, 0].
Ω
Therefore, according to Theorem 1.2, either

v(x) > m for every x ∈ Ω, (1.49)
or
v=m in Ω. (1.50)
As (1.49) and (1.50) imply (1.45) and (1.46), respectively, the proof of the
alternative of the first part of the theorem is complete.
Now, suppose (1.26), (1.45), and conditions (a), (b) and (c). Then, all
the coefficients of Lh live in L∞ (Ω), and
v(x0 ) = m, v ∈ C(BR (x0 ) ∩ Ω̄).
Consequently, owing to Theorem 1.3, we find that

∂v
(x0 ) < 0,
∂ν
which concludes the proof.
1.8 Appendix: Smooth domains
The following definition fixes the concept of regularity for open sets.
Definition 1.5. Let Ω ⊂ RN , N ≥ 1, be an open set with boundary ∂Ω,

x0 ∈ ∂Ω, and consider an integer k ≥ 1. Then, Ω is said to be of class C k
at x0 if there exist R := R(x0 ) > 0, an open neighborhood D of zero in RN ,
and a bijection Φ : BR (x0 ) → D such that Φ(x0 ) = 0 and
i) DΦ(x0 ) ∈ Iso(RN ),
ii) Φ(BR (x0 ) ∩ Ω) = {(x1 , ..., xN ) ∈ D : xN > 0},
iii) Φ(BR (x0 ) ∩ ∂Ω) = {(x1 , ..., xN ) ∈ D : xN = 0},
iv) Φ ∈ C k (BR (x0 ); D) and Φ−1 ∈ C k (D; BR (x0 )).
In such case, Φ is said to be a diffeomorphism straightening ∂Ω at x0 into

the hyperplane xN = 0.
Given an open subset Γ ⊂ ∂Ω, Γ is said to be of class C k if Ω is of class
C at each x0 ∈ Γ. The whole domain Ω is said to be of class C k when ∂Ω
k
is of class C k .
Given an open subset Γ ⊂ ∂Ω of class C k , a function ψ : Γ → R is said to
be of class C k if, for every x0 ∈ Γ and any diffeomorphism Φ straightening
∂Ω at x0 , the following condition holds
ψ ◦ Φ−1 ∈ C k (D ∩ {(x1 , ..., xN ) ∈ RN : xN = 0}).

x
N
Φ
x
0
Γ 0 D x =0
−1
N
Φ
Fig. 1.9 The straightening diffeomorphism Φ (Γ := ∂Ω)
Remark 1.3.
(a) According to the inverse function theorem, Property (i) implies Prop-
erty (iv) if Φ ∈ C k (BR (x0 ); RN ).
(b) Suppose ∂Ω is of class C k for some integer k ≥ 1. Then, ψ ∈ C k (Ω̄)
implies ψ|∂Ω ∈ C k (∂Ω). Conversely, according to the H. F. F. Tietze
extension theorems, every function ψ ∈ C k (∂Ω) admits an extension to
a function of class C k (Ω̄).
The next result shows that Ω is of class C k at x0 ∈ ∂Ω if and only if in a

neighborhood of x0 the boundary ∂Ω is the graph of a C k function of N −1
variables. Subsequently, the coordinates of a vector x ∈ RN are denoted by
x = (x1 , ..., xN ),
and, for a given i ∈ {1, ..., N −1}, we will denote by x[i] the vector
x[i] := (x1 , ..., xi−1 , xi+1 , ..., xN ) ∈ RN −1 ,
while
x[N ] := (x1 , ..., xN −1 ) ∈ RN −1 .
Also, for every η > 0 and x0 ∈ RN , we denote by Bη ((x0 )[i] ) the ball of
RN −1 of radius η centered at (x0 )[i] ∈ RN −1 . Using these notations, the
following result holds.
Theorem 1.8. Let Ω ⊂ RN , N ≥ 1, be an open set with boundary ∂Ω,

x0 ∈ ∂Ω, and k ∈ N, k ≥ 1. Then, the following conditions are equivalent:
(a) Ω is of class C k at x0 .
(b) There exist η > 0, i ∈ {1, ..., N }, f ∈ C k (B̄η ((x0 )[i] ); R), and an open
neighborhood N0 of x0 in RN such that
N0 ∩ ∂Ω = {(x1 , ..., xi−1 , f (x[i] ), xi+1 , ..., xN ) : x[i] ∈ Bη ((x0 )[i] )},
i.e., ∂Ω is the graph of the function xi = f (x[i] ) in N0 , and either
x ∈ N0 ∩ Ω if and only if xi > f (x[i] ), (1.51)
or else
x ∈ N0 ∩ Ω if and only if xi < f (x[i] ). (1.52)
Proof. Suppose (b) with, e.g., (1.51). Then, by Remark 1.3(a), it is easy
to see that the map
( )
Φ(x) := (x − x0 )[i] , xi − f (x[i] ) , x ∈ BR (x0 ),
provides us with a diffeomorphism straightening ∂Ω at x0 for sufficiently
small R > 0, as discussed in Definition 1.5. Therefore, (b) implies (a).
Suppose (a) and let Φ be a diffeomorphism straightening ∂Ω at x0 ,
as discussed in Definition 1.5. Let ΦN denote its N -th coordinate map.
Necessarily, ∇ΦN (x0 ) ̸= 0, because DΦ(x0 ) ∈ Iso(RN ). Thus,
∂ΦN
(x0 ) ̸= 0 for some i ∈ {1, ..., N }. (1.53)
∂xi
Moreover, by definition,
BR (x0 ) ∩ ∂Ω = Φ−1
N (0).
Consequently, (b) should follow by applying the implicit function theorem

to the equation ΦN = 0 at x0 , by expressing xi as a function of the remain-
ing variables, xi = f (x[i] ). Indeed, consider the one-parametric family of
equations
ΦN (x1 , ..., xN ) − c = 0, c ∼ 0.
According to the implicit function theorem, we find from (1.53) that there
exist an open neighborhood M of
(c, x[i] ) = (0, (x0 )[i] ) ∈ RN
and a (unique) function F ∈ C k (M; R) such that
F (0, (x0 )[i] ) = (x0 )i
and, for every (c, x[i] ) ∈ M,
ΦN (x1 , ..., xi−1 , F (c, x[i] ), xi+1 , ..., xN ) − c = 0. (1.54)
By the uniqueness obtained as an application of the implicit function the-

orem, it is apparent that ∂Ω is given through the graph of the function
f := F (0, ·).
Moreover, differentiating (1.54) with respect to c, yields to
∂ΦN ∂F
1= (x0 ) (0, (x0 )[i] ). (1.55)
∂xi ∂c
∂ΦN
Thus, if, e.g., ∂xi (x0 ) > 0, then, by continuity, (1.55) implies that
∂F
(c, x[i] ) > 0
∂c
for c ∼ 0 and x[i] ∼ (x0 )[i] . According to Definition 1.5 (ii), for every x ∈ Ω
sufficiently close to x0 , we have that
c := ΦN (x) > 0
and hence,
xi = F (c, x[i] ) > F (0, x[i] ) = f (x[i] ),
whereas if x ̸∈ Ω̄, then c := ΦN (x) < 0 and, so,
xi = F (c, x[i] ) < F (0, x[i] ) = f (x[i] ).
Consequently, in such case, condition (1.51) holds. Similarly, (1.52) holds
when ∂Φ
∂xi (x0 ) < 0. This concludes the proof.
N

The following result shows that Ω satisfies the uniform interior sphere
property in the strong sense if it is of class C 2 .
Theorem 1.9. Let Ω ⊂ RN , N ≥ 1, be an open set with boundary ∂Ω,

and Γ0 a closed and open subset of ∂Ω of class C 2 . Then, Ω satisfies the
uniform interior sphere property in the strong sense on Γ0 .
Proof. By Lemma 1.1(b), it suffices to prove that Ω satisfies the uniform

interior sphere property on Γ0 .
Pick x0 ∈ Γ0 , and let η > 0, i ∈ {1, ..., N }, and f ∈ C 2 (B̄η ((x0 )[i] ); R),
satisfying Theorem 1.8(b) with k = 2. Subsequently, we set
K := B̄η ((x0 )[i] ) × B̄η ((x0 )[i] )
and consider the function R ∈ C 2 (K; R) defined through
1
R(x̃, ỹ) := f (x̃) − f (ỹ) − Df (ỹ)(x̃ − ỹ) − D2 f (ỹ)(x̃ − ỹ, x̃ − ỹ),
2
for all (x̃, ỹ) ∈ K, where Df and D2 f stand for the first and the second
order differentials of f . It is well known that Df (ỹ) is given by the gradient
∇f (ỹ), while D2 f (ỹ) is given through the Hessian matrix of f at ỹ. By our
regularity assumptions on f , we have that
R(x̃, ỹ)
lim = 0, |ỹ − (x0 )[i] | ≤ η.
x̃→ỹ |x̃ − ỹ|2
Therefore, the quotient function Q defined by

{
R(x̃,ỹ)
2, if x̃ ̸= ỹ,
Q(x̃, ỹ) := |x̃−ỹ|
0, if x̃ = ỹ,
is continuous in K. As K is compact, Q is uniformly continuous in K and,
therefore, for every h > 0 there exists δ > 0 such that
|x̃ − ỹ| ≤ δ =⇒ |R(x̃, ỹ)| ≤ h|x̃ − ỹ|2 (1.56)
for every x̃, ỹ ∈ K. The importance of this estimate will become apparent
later.
By the choice of f and i ∈ {1, ..., N }, the tangent hyperplane to ∂Ω at
x0 is given through
( )
xi = x0i + ∇f ((x0 )[i] ) x[i] − (x0 )[i]
and hence,
( )
∂f ∂f ∂f ∂f
ν := , ..., , −1, , ...,
∂x1 ∂xi−1 ∂xi+1 ∂xN
((x0 )[i] )
is the normal vector to ∂Ω at x0 if i < N , whereas
( )
∂f ∂f
ν := , ..., , −1
∂x1 ∂xN −1
((x0 )[i] )
if i = N . Subsequently, the sign ± is chosen so that the unit normal vector
ν
n := ±
|ν|
satisfy x0 − ϵn ∈ Ω for sufficiently small ϵ > 0. For this choice, n provides
us with the outward unit normal to Ω at x0 ∈ ∂Ω. Next, we will show that
Bϵ (x0 − ϵn) ⊂ Ω
for sufficiently small ϵ > 0. Setting n := (n1 , ..., nN ), it is easy to see that
this holds from the estimate
[ ]2 ∑N
2
D0 := f (x[i] ) + ϵni − x0i + (xj + ϵnj − x0j ) ≥ ϵ2 . (1.57)
j=1
j̸=i
Indeed, by Theorem 1.8(b), we have that

( )
x1 , ..., xi−1 , f (x[i] ), xi+1 , ..., xN ∈ ∂Ω
for x[i] ∼ (x0 )[i] , and (1.57) implies that
( )
x1 , ..., xi−1 , f (x[i] ), xi+1 , ..., xN ∈ RN \ Bϵ (x0 − ϵn).
Therefore, Bϵ (x0 − ϵn) ⊂ Ω because x0 − ϵn ∈ Ω for sufficiently small ϵ > 0.
The estimate (1.57) can be derived from the next chain of identities
[ ( )( )
D0 = ϵni + ∇f (x0 )[i] x[i] − (x0 )[i]
1 ( )( ) ( )]2
+ D2 f (x0 )[i] x[i] − (x0 )[i] , x[i] − (x0 )[i] + R x[i] , (x0 )[i]
2
∑
N
2
∑
N ∑N
+ (xj − x0j ) + ϵ2 n2j + 2ϵ nj (xj − x0j )
j=1 j=1 j=1
j̸=i j̸=i j̸=i
∑
N
( )( ) ∑
N
=ϵ 2
n2j +2ϵni ∇f (x0 )[i] x[i] − (x0 )[i] +2ϵ nj (xj − x0j )
j=1 j=1
j̸=i
[ ( )( )]2 ∑N
2
+ ∇f (x0 )[i] x[i] − (x0 )[i] + (xj − x0j )
j=1
j̸=i
( )( ) ( )
+ϵni D2 f (x0 )[i] x[i] −(x0 )[i] , x[i] −(x0 )[i] +R x[i] , (x0 )[i] ,
where R is a certain C 2 function satisfying
( )
R x[i] , (x0 )[i]
lim 2 = 0, (1.58)
x[i] →(x0 )[i] x − (x )
[i] 0 [i]
whose explicit knowledge is not important for our purposes here. Since
∑
N
[ ( )( )]2
n2j = |n|2 = 1, ∇f (x0 )[i] x[i] − (x0 )[i] ≥ 0,
j=1
and
( )
∂f ( ) −1 ∂f ( )
ni (x0 )[i] + nj = ± (x0 )[i] + nj = 0, j ̸= i,
∂xj |ν| ∂xj
we find from the previous identity that
∑
N
2 ( )
D0 ≥ ϵ 2 + (xj − x0j ) + R x[i] , (x0 )[i]
j=1
j̸=i
( )( )
+ ϵni D2 f (x0 )[i] x[i] −(x0 )[i] , x[i] −(x0 )[i] .
Therefore, owing to (1.58), in a neighborhood of x0 we obtain that

1
D0 ≥ ϵ2 + |x[i] − (x0 )[i] |2 ≥ ϵ2
2
for sufficiently small ϵ > 0. Actually, by (1.56), this estimate is satisfied
uniformly in a neighborhood of x0 , because f is of class C 2 . Consequently,
(1.57) holds and, therefore, Ω satisfies the uniform interior sphere property
on Γ0 . This completes the proof.
1.9 Comments on Chapter 1
Most of the contents covered in this chapter were elaborated from the classi-
cal textbook of M. H. Protter and H. F. Weinberger [183], as well as many of
the subsequent historical remarks; it is the most paradigmatic book about
the maximum principle for elliptic operators, and, undoubtedly, it counts
among the best monographs on Partial Differential Equations of the 20th
Century. However, the reader should pay special attention to all necessary
changes for re-stating all classical results in terms of minimum principles,
instead of maximum principles, as it has been done in the present mono-
graph.
In the special case when L = −∆, Theorem 1.1 goes back to C. F.
Gauss [75] and S. Earnshaw [58], and to A. Paraf [174] when N = 2 (two
dimensions) and c > 0. The version of A. Paraf was extended by E. Picard
[178] and L. Lichtenstein [125], [126], to cover the more general case when
c ≥ 0. Later, T. Moutard [169] generalized the Paraf result to more than
two dimensions. These results were used by M. Picone [179] to obtain a
generalized minimum principle. Theorem 1.2 of Section 1.2 goes back to
E. Hopf [103]; it was the first result where the continuity assumptions on
the coefficients of L were removed. The proof of Theorem 1.2 given in this
chapter is based upon the proof of Theorems 5 and 6 of Section 3 in M. H.
Protter and H. F. Weinberger [183].
Section 1.3 is a non-trivial re-elaboration of the contents of the two
paragraphs after Theorem 1 of W. Walter [224]. The example (1.24) is due
to R. Redheffer [187].
Under some additional continuity properties on the coefficients of the
operator, Theorem 1.3 goes back to G. Giraud [80], [81]. The version in-
cluded in this chapter is attributable to E. Hopf [104] and O. A. Oleinik
[172]. It is a re-elaboration from Theorems 7 and 8 of Section 3 in M. H.
Protter and H. F. Weinberger [183].
Theorems 1.5 and 1.6 are based on Lemma 1 of W. Walter [224]. The
relevance of these uniform versions of Theorem 1.3 will become apparent
in Chapter 2.
Theorem 1.7 goes back to Theorem 10 of Chapter 2 of M. H. Protter and
H. F. Weinberger [183]; it generalizes substantially Theorems 1.2 and 1.3.
In Chapter 7 we will show that the existence of h satisfying conditions (i)
and (ii) is, actually, equivalent to the positivity of the principal eigenvalue
of L in Ω under Dirichlet boundary conditions. This characterization will
provide us with a sharp substantial generalization of all classical results of
this chapter.
Section 1.8 reviews some very basic concepts and results, which, how-
ever, might be difficult to document in the available literature. For example,
though L. C. Evans in the Remark on p. 330 of [60] claimed that “ the
interior ball condition automatically holds if ∂Ω is C 2 ”, the reader should
recognize that, being certainly elementary, the proof of Theorem 1.9 is far
from ‘ automatic ’.

Chapter 2
Classifying supersolutions
This chapter will be developed under the following general hypotheses:

H1 Ω is a bounded domain of RN , N ≥ 1, whose boundary consists of two
disjoint open and closed subsets, Γ0 and Γ1 ,
∂Ω := Γ0 ∪ Γ1 ,
and it satisfies the interior sphere property at every x ∈ ∂Ω. Necessar-
ily, Γ0 and Γ1 must possess finitely many components. Either Γ0 , or
Γ1 , might be empty.
H2 The differential operator
∑
N
∂2 ∑ N
∂
L := − aij + bj +c (2.1)
i,j=1
is uniformly elliptic in Ω and

aij , bj , c ∈ L∞ (Ω), i, j ∈ {1, ..., N }.
H3 β ∈ C(Γ1 ; R) is a continuous function and ν = (ν1 , ..., νN ) ∈ C(∂Ω; RN )
is an outward pointing nowhere tangent vector field, in the sense that
ν(x) is an outward pointing vector at x for all x ∈ ∂Ω.
Under these assumptions, we can introduce the boundary operator
B : C(Γ0 ) ⊗ C 1 (Γ1 ) → C(∂Ω)
defined by
{
ψ on Γ0 ,
Bψ := ∂ψ for all ψ ∈ C(Γ0 ) ⊗ C 1 (Γ1 ). (2.2)
∂ν + βψ on Γ1 ,
Note that B is the Dirichlet boundary operator on Γ0 , subsequently denoted
by D, the Neumann boundary operator on Γ1 if ν = n and β = 0, denoted
41
by N, and a first order regular oblique derivative boundary operator on Γ1

in any other case.
The following concepts are pivotal.
Definition 2.1. Suppose u ∈ C 2 (Ω) ∩ C 1 (Ω ∪ Γ1 ) ∩ C(Ω̄). Then:

(a) u is said to be a supersolution of (L, B, Ω) if
{
Lu ≥ 0 in Ω,
Bu ≥ 0 on ∂Ω.
When some of these inequalities is strict, u is said to be a strict su-
persolution of (L, B, Ω).
(b) u is said to be a subsolution of (L, B, Ω) if
{
Lu ≤ 0 in Ω,
Bu ≤ 0 on ∂Ω.
When some of these inequalities is strict, u is said to be a strict sub-
solution of (L, B, Ω).
According to Definition 2.1, the main assumption of Theorem 1.7 can

be expressed by simply saying that h ∈ C 2 (Ω) ∩ C(Ω̄) is a supersolution
of (L, D, Ω) everywhere positive in Ω̄. Consequently, it must be a positive
strict supersolution of (L, D, Ω), because Dh > 0 on Γ0 = ∂Ω.
The main goal of this chapter is ascertaining all admissible canonical
behaviors of the supersolutions of (L, B, Ω) from the existence of a single
positive supersolution. This objective will be accomplished through The-
orems 2.1 and 2.4. Later, we will derive from Theorem 2.1 the strong
positivity of the resolvent operator of the linear boundary value problem
(2.31) associated to (L + ω, B, Ω) for sufficiently large ω, which will be
established by Theorem 2.2. In Chapter 7, using the theory of positive op-
erators developed in Chapter 6, we will infer from these results the existence
and uniqueness of the principal eigenvalue of (L + ω, B, Ω).
2.1 First classification theorem
The following result provides us with all admissible canonical behaviors of

the supersolutions of (L, B, Ω) in the presence of a supersolution every-
where positive in Ω̄. It should be remembered that the existence of such a
supersolution was the main assumption of Theorem 1.7 in the special case
when Γ1 = ∅ (B = D).
Classifying supersolutions 43
Theorem 2.1. Suppose (L, B, Ω) admits a supersolution h ∈ C 2 (Ω) ∩

C 1 (Ω̄) such that
h(x) > 0 for all x ∈ Ω̄. (2.3)
Then, any supersolution u ∈ C 2 (Ω) ∩ C 1 (Ω̄) of (L, B, Ω) must satisfy some
of the following alternatives:
A1 u = 0 in Ω.
A2 u(x) > 0 for every x ∈ Ω ∪ Γ1 , and
∂u
(x) < 0 for all x ∈ u−1 (0) ∩ Γ0 .
∂ν
A3 There exists a constant m < 0 such that
u = mh in Ω̄.
In such case, u(x) < 0 for all x ∈ Ω̄, Γ0 = ∅, and
{
Lh = 0 in Ω,
(2.4)
Bh = 0 on ∂Ω.
In other words, τ = 0 must be an eigenvalue to a positive eigenfunction
(h itself ) of the linear eigenvalue problem
{
Lψ = τ ψ in Ω,
(2.5)
Bψ = 0 on ∂Ω.
Proof. Let u ∈ C 2 (Ω) ∩ C 1 (Ω̄) be a supersolution of (L, B, Ω). Then,

thanks to (2.3),
u
v := ∈ C 2 (Ω) ∩ C 1 (Ω̄).
h
Moreover, arguing as in the proof of Theorem 1.7, we find that
Lu = hLh v, (2.6)
where
∑
N
∂2 ∑ ∂
N
Lh := − aij + bhj + ch , (2.7)
i,j=1
∂xi ∂x j j=1
∂x j
with
2∑
N
Lh ∂h
ch := , bhj := bj − aij , 1 ≤ j ≤ N.
h h i=1 ∂xi
As h ∈ C 1 (Ω̄), we find from the hypothesis H2 that

bhj , ch ∈ L∞ (Ω), 1 ≤ j ≤ N.
Moreover,
ch ≥ 0
because h is a supersolution of (L, B, Ω) and, hence, Lh ≥ 0 in Ω, by
Definition 2.1.
As u is a supersolution of (L, B, Ω), (2.3) and (2.6) imply
Lh v ≥ 0 in Ω. (2.8)
Moreover, Bu ≥ 0 on ∂Ω, and, in particular, Bu = u ≥ 0 on Γ0 . Thus,
v≥0 on Γ0 . (2.9)
Similarly, on Γ1 we have that
∂(hv) ∂v ∂h
0 ≤ Bu = B(hv) = + βhv = h +v + βhv
( ∂ν ) (∂ν ∂ν )
∂v ∂h ∂v Bh
=h + + βh v = h + v
∂ν ∂ν ∂ν h
and, consequently,
0 ≤ Bu = hBh v on Γ1 , (2.10)
where Bh stands for the boundary operator
{
ψ on Γ0 ,
Bh ψ := ∂ψ ψ ∈ C(Γ0 ) ⊗ C 1 (Γ1 ), (2.11)
∂ν + βh ψ on Γ 1 ,
with
Bh
βh := ≥0 on Γ1 , (2.12)
h
because Bh ≥ 0 on ∂Ω. Note that βh ∈ C(Γ1 ). Incidentally, (2.12) holds
independently of the sign of β ∈ C(Γ1 ).
Combining (2.8), (2.9) and (2.10), it becomes apparent that v provides
us with a supersolution of (Lh , Bh , Ω).
Now, we will show that Alternative A3 occurs if u(x0 ) < 0 for some
x0 ∈ Ω. Indeed, in this case we have that
u(x0 )
v(x0 ) = <0
h(x0 )
and hence,
m := min v < 0.
Ω̄
Thus, as ch ≥ 0 and v is a superharmonic function of Lh in Ω, it follows
from Theorem 1.2 that either
v(x) > m for all x ∈ Ω, (2.13)
or
v=m in Ω. (2.14)
Suppose (2.13). Then, since m < 0, (2.9) implies that
v(x) > m for all x ∈ Ω ∪ Γ0 . (2.15)
Let x1 ∈ Ω̄ be such that
m = v(x1 ).
By (2.15), x ∈ Γ1 . Consequently, owing to Theorem 1.3, we have that
∂v
(x1 ) < 0.
∂ν
Thus, according to (2.10) and (2.11),
∂v
0 ≤ Bh v(x1 ) = (x1 ) + βh (x1 )v(x1 ) < βh (x1 )v(x1 ) = βh (x1 )m
∂ν
and, consequently,
βh (x1 ) < 0,
which contradicts (2.12). Therefore, (2.14) holds, and hence,
u = mh in Ω. (2.16)
By continuity, (2.16) must be satisfied in Ω̄, and, so, Alternative A3 occurs.
The remaining assertions of Alternative A3 follow very easily from (2.16).
Indeed, suppose Γ0 ̸= ∅ and pick x0 ∈ Γ0 . Then, h(x0 ) > 0 and hence,
u(x0 ) = mh(x0 ) < 0,
which is impossible, for as u ≥ 0 on Γ0 . Thus, Γ0 = ∅. Moreover, the
estimates
0 ≤ Lu = mLh ≤ 0 in Ω
imply
Lu = 0,
while the estimates
0 ≤ Bu = mBh ≤ 0 on ∂Ω = Γ1 ,
entail
Bu = 0 on ∂Ω,
and hence (2.4) holds.
To complete the proof of the theorem, it suffices to show that some of

the first two alternatives occurs if u ≥ 0 in Ω̄. So, suppose
u≥0 in Ω̄.
Then, either
u(x0 ) = 0 for some x0 ∈ Ω, (2.17)
or
u(x) > 0 for all x ∈ Ω. (2.18)
Suppose (2.17). Then, v ≥ 0 in Ω̄ and v(x0 ) = 0. Thus, as v is a supersolu-
tion of (Lh , Bh , Ω), we find from Theorem 1.2 that v = 0 in Ω̄. Therefore,
u = 0 in Ω̄ and Alternative A1 holds.
In case (2.18), v(x) > 0 for all x ∈ Ω, and hence, since v is a supersolu-
tion of (Lh , Bh , Ω), it follows from Theorem 1.3 that
∂v
(x) < 0 for all x ∈ v −1 (0) ∩ ∂Ω. (2.19)
∂ν
Suppose Γ1 ̸= ∅ and v(x1 ) = 0 for some x1 ∈ Γ1 . Then,
∂v ∂v
0 ≤ Bh v(x1 ) = (x1 ) + βh (x1 )v(x1 ) = (x1 ),
∂ν ∂ν
which contradicts (2.19). Thus, v(x1 ) > 0, and so, u(x1 ) > 0 for all x1 ∈ Γ1 .
Moreover, for every
x0 ∈ Γ0 ∩ u−1 (0),
we have that v(x0 ) = 0 and hence, it follows from (2.19) that
∂u ∂(hv)
(x0 ) = (x0 )
∂ν ∂ν
∂v ∂h
= h(x0 ) (x0 ) + v(x0 ) (x0 )
∂ν ∂ν
∂v
= h(x0 ) (x0 ) < 0.
∂ν
Therefore, Alternative A2 holds. The proof is complete.
As an immediate consequence from Theorem 2.1, the following result
holds. Essentially, it shows the strong positivity of any smooth superso-
lution u ̸= 0 of (L, B, Ω) from the existence of a strict supersolution h
satisfying (2.3).
Corollary 2.1. Suppose (L, B, Ω) possesses a strict supersolution h ∈
C 2 (Ω) ∩ C 1 (Ω̄) such that h(x) > 0 for all x ∈ Ω̄. Then, any superso-
lution u ∈ C 2 (Ω) ∩ C 1 (Ω̄), u ̸= 0, of (L, B, Ω) (in particular, any strict
supersolution) satisfies u(x) > 0 for every x ∈ Ω ∪ Γ1 , and
∂u
(x) < 0 for all x ∈ u−1 (0) ∩ Γ0 .
∂ν
Proof. Since h is a strict supersolution of (L, B, Ω), (2.4) cannot be

satisfied and therefore, Alternative A3 of Theorem 2.1 cannot occur. As
Alternative A1 is excluded, because u ̸= 0, Alternative A2 holds.
As a consequence from Corollary 2.1, the next uniqueness result holds.
Corollary 2.2. Suppose (L, B, Ω) possesses a strict supersolution h ∈

C 2 (Ω) ∩ C 1 (Ω̄) such that h(x) > 0 for all x ∈ Ω̄. Then, u = 0 is the
unique function u ∈ C 2 (Ω) ∩ C 1 (Ω̄) satisfying
{
Lu = 0 in Ω,
(2.20)
Bu = 0 on ∂Ω.
Therefore, for every (f, g) ∈ C(Ω̄) × C(∂Ω), the boundary value problem
{
Lu = f in Ω,
(2.21)
Bu = g on ∂Ω,
admits, at most, a unique solution u ∈ C 2 (Ω) ∩ C 1 (Ω̄).
Proof. Suppose u ∈ C 2 (Ω) ∩ C 1 (Ω̄), u ̸= 0, solves (2.20). Then, according

to Corollary 2.1, u(x) > 0 for all x ∈ Ω ∪ Γ1 . Similarly, since −u ̸= 0
provides us with another solution of (2.20), we have that −u(x) > 0 for all
x ∈ Ω ∪ Γ1 , which is impossible. Consequently, u = 0 is the unique function
of C 2 (Ω) ∩ C 1 (Ω̄) solving (2.20). The uniqueness of the solution of (2.21)
follows easily from the previous uniqueness result.
2.2 Existence of positive strict supersolutions
The following result provides us with a sufficient condition for the existence
of a strict supersolution satisfying (2.3).
Proposition 2.1. Suppose there are ψ ∈ C 2 (Ω̄) and γ > 0 such that
∂ψ
(x) ≥ γ for all x ∈ Γ1 . (2.22)
∂ν
Then, there exists ω0 ∈ R such that, for every ω > ω0 , (L + ω, B, Ω)
possesses a strict supersolution h ∈ C 2 (Ω̄) with h(x) > 0 for all x ∈ Ω̄.
Proof. Suppose h ∈ C 2 (Ω̄), with h(x) > 0 for all x ∈ Ω̄, is a strict
supersolution of (L + ω0 , B, Ω) for some ω0 ∈ R. Then, for every ω > ω0 ,
(L + ω)h = (L + ω0 )h + (ω − ω0 )h ≥ (ω − ω0 )h > 0
and hence, h is a strict supersolution of (L + ω, B, Ω). Therefore, it suffices

to show that there are ω0 ∈ R and M > 0 for which the function h defined
through
h(x) := eM ψ(x) , x ∈ Ω̄, (2.23)
is a strict supersolution of (L + ω0 , B, Ω). Indeed, by definition, h(x) > 0

for all x ∈ Ω̄. In particular,
Bh(x) = h(x) > 0 for all x ∈ Γ0 .
Moreover, on Γ1 we have that

( )
∂ψ ∂ψ
Bh = M h + βh = M +β h
∂ν ∂ν
and hence, according to (2.22), we find that
Bh ≥ (M γ + β)h ≥ 0 on Γ1
for sufficiently large M > 0. Suppose M > 0 has been chosen in this way.
Then, since Lh ∈ L∞ (Ω) and h(x) > 0 for all x ∈ Ω̄, it becomes apparent
that
(L + ω0 )h > 0 in Ω
for sufficiently large ω0 ∈ R. This concludes the proof.

The next result provides us with a simple sufficient condition for the
existence of a function ψ satisfying the requirements of Proposition 2.1.
Lemma 2.1. Suppose Ω is of class C 2 . Then, there exist ψ ∈ C 2 (Ω̄) and

γ > 0 satisfying (2.22).
Proof. Fix x0 ∈ ∂Ω. Then, by Definition 1.5, there exist R > 0, an open
subset D ⊂ RN , and a bijection
Φ : BR (x0 ) → D = Φ(BR (x0 ))
such that:
i) Φ(x0 ) = 0 and DΦ(x0 ) ∈ Iso (RN ),

ii) Φ(BR (x0 ) ∩ Ω) = {(x1 , ..., xN ) ∈ D : xN > 0},
iii) Φ(BR (x0 ) ∩ ∂Ω) = {(x1 , ..., xN ) ∈ D : xN = 0},
iv) Φ ∈ C 2 (BR (x0 ); D) and Φ−1 ∈ C 2 (D; BR (x0 )).
As ν(x0 ) is an outward pointing vector at x0 ∈ ∂Ω,

x0 + tν(x0 ) ∈ BR (x0 ) \ Ω̄
for sufficiently small t > 0, and hence
Φ(x0 + tν(x0 )) ∈ {(x1 , . . . , xN ) ∈ D : xN < 0}. (2.24)
Moreover,
⟨ν(x0 ), n⟩ > 0,
where n stands for the outward unit normal to Ω at x0 ∈ ∂Ω.
Let Tx0 denote the tangent hyperplane to ∂Ω at x0 . Then,
{ }
Tx0 = x0 + v ∈ RN : ⟨v, n⟩ = 0
and
RN = Tx0 ⊕ span [ν(x0 )].
Subsequently, we denote by PN the projection of RN onto the N -th coor-
dinate. According to (i)–(iii), it becomes apparent that
DΦ(x0 ) (Tx0 ) = {v ∈ RN : PN v = 0} (2.25)
(see Figure 2.1) and
RN = DΦ(x0 ) (Tx0 ) ⊕ span [ν̃] , ν̃ := DΦ(x0 )ν(x0 ). (2.26)
x
N
Ω
x0
Φ
n
ν (x ) Φ (x ) x =0
0 0 N
Φ
−1
ν
Tx
0 (0,0,...,0,−1)
Fig. 2.1 Sketch of the construction
Also, since Φ ∈ C 2 (BR (x0 ); D) and Φ(x0 ) = 0, we have that

Φ(x0 + tν(x0 )) = tν̃ + o(t) as t ↓ 0. (2.27)
Due to (2.24),
PN Φ(x0 + tν(x0 )) < 0
for sufficiently small t > 0. Thus, (2.27) implies that PN ν̃ ≤ 0. On the
other hand, thanks to (2.25) and (2.26), it is apparent that PN ν̃ ̸= 0 and,
therefore,
PN ν̃ < 0. (2.28)
Consequently, ν̃ points outward the open set Φ(BR (x0 ) ∩ Ω) at Φ(x0 ) = 0,
as illustrated by Figure 2.1.
Now, consider the function
Ψ(x) := −PN Φ(x), x ∈ BR (x0 ).
Clearly, Ψ ∈ C 2 (BR (x0 )) and, for every x ∈ ∂Ω sufficiently close to x0 , we
have that
∂Ψ
(x) = DΨ(x)ν(x) = −PN DΦ(x)ν(x).
∂ν(x)
Consequently, by (2.26), (2.28), and the continuity of the map
x 7→ DΦ(x)ν(x),
it becomes apparent that R can be shortened, if necessary, so that
∂Ψ
(x) ≥ γ for all x ∈ BR (x0 ) ∩ ∂Ω,
∂ν(x)
for some constant γ > 0.
Note that, thanks to condition (iii), we also have that
Ψ(x) = 0 for all x ∈ BR (x0 ) ∩ ∂Ω.
As the previous argument is valid for all x0 ∈ ∂Ω and ∂Ω is a compact
manifold, there are m ≥ 1 points xj0 ∈ ∂Ω, m positive real numbers Rj > 0,
and m functions
Ψj ∈ C 2 (Uj ) , Uj := BRj (xj0 ), 1 ≤ j ≤ m,
such that
∪
m
(Uj ∩ ∂Ω) = ∂Ω,
j=1
Ψj = 0 in Uj ∩ ∂Ω, 1 ≤ j ≤ m, (2.29)
and, for some positive constant γ > 0,

∂Ψj
(x) ≥ γ for all x ∈ Uj ∩ ∂Ω, 1 ≤ j ≤ m. (2.30)
∂ν(x)
Clearly, there exists ϵ > 0 such that
∪
m
{x ∈ Ω : dist (x, ∂Ω) ≤ ϵ} ⊂ Uj
j=1
and hence, setting

Um+1 := {x ∈ Ω : dist (x, ∂Ω) > ϵ},
the open neighborhoods Uj , 1 ≤ j ≤ m + 1, provide us with a covering of
Ω̄. By Proposition 2.2 in the Appendix, there exist
ψj ∈ C0∞ (RN ), 1 ≤ j ≤ m + 1,
with
supp ψj ⊂ Ūj , 1 ≤ j ≤ m + 1,
such that
∑
m+1
ψj = 1 in Ω̄.
j=1
It remains to check that the function ψ defined through

∑
m
ψ(x) := Ψj (x)ψj (x) + ψm+1 (x), x ∈ Ω̄,
j=1
satisfies the requirements of the lemma. Indeed, ψ ∈ C 2 (Ω̄) and

∂ψ ∑m
∂Ψj ∑m
∂ψj
(x) = (x)ψj (x) + Ψj (x) (x)
∂ν(x) j=1
∂ν(x) j=1
∂ν(x)
for all x ∈ ∂Ω, because ψm+1 = 0 on a neighborhood of ∂Ω. Thus, by

(2.29) and (2.30), we have that
∂ψ ∑m
∂Ψj ∑m
(x) = (x)ψj (x) ≥ γ ψj (x) = γ
∂ν(x) j=1
∂ν(x) j=1
for all x ∈ ∂Ω, since ψm+1 = 0 on ∂Ω. This ends the proof.
Essentially, the function ψ(x) constructed in the proof of Lemma 2.1
behaves like −dist (x, ∂Ω) for all x ∈ Ω sufficiently close to ∂Ω.
2.3 Positivity of the resolvent operator
As an immediate consequence from Corollary 2.2 and Proposition 2.1, the

following result holds.
Theorem 2.2. Assume that there exist ψ ∈ C 2 (Ω̄) and γ > 0 satisfying
(2.22). Then, there exists ω0 ∈ R such that, for every ω > ω0 , u = 0 is the
unique function u ∈ C 2 (Ω) ∩ C 1 (Ω̄) solving the problem
{
(L + ω)u = 0 in Ω,
Bu = 0 on ∂Ω.
Therefore, for every ω > ω0 and (f, g) ∈ C(Ω̄) × C(∂Ω), the problem
{
(L + ω)u = f in Ω,
(2.31)
Bu = g on ∂Ω,
has, at most, a unique solution u ∈ C 2 (Ω) ∩ C 1 (Ω̄). Moreover, if such
solution exists and f ≥ 0, g ≥ 0, with (f, g) ̸= (0, 0), then
u(x) > 0 for all x ∈ Ω ∪ Γ1 (2.32)
and
∂u
(x) < 0 for all x ∈ u−1 (0) ∩ Γ0 . (2.33)
∂ν
Proof. According to Proposition 2.1, there exists ω0 ∈ R such that, for
every ω > ω0 , (L + ω, B, Ω) possesses a strict supersolution h ∈ C 2 (Ω̄) such
that h(x) > 0 for all x ∈ Ω̄. Consequently, the uniqueness follows from
Corollary 2.2.
Suppose ω > ω0 , f ≥ 0, g ≥ 0, (f, g) ̸= (0, 0), and (2.31) admits a
solution u ∈ C 2 (Ω) ∩ C 1 (Ω̄). Then, u provides us with a strict supersolu-
tion of (L + ω, B, Ω) and, therefore, by Corollary 2.1, (2.32) and (2.33) are
satisfied. The proof is complete.
2.4 Behavior of the positive supersolutions on Γ0
The following result provides us with the point-wise behavior of the positive
supersolutions of (L, B, Ω).
Theorem 2.3. Suppose there are ψ ∈ C 2 (Ω̄) and γ > 0 satisfying (2.22).
Let u ∈ C 2 (Ω) ∩ C 1 (Ω̄), u > 0, be a supersolution of (L, B, Ω). Then, u
satisfies (2.32) and (2.33).
If, in addition, (L, Ω) satisfies the decay property of E. Hopf on Γ0 , as

discussed in Definition 1.4, then, there exists a constant δ := δ(u) > 0 such
that
u(x) ≥ δ dist (x, ∂Ω) for all x ∈ Ω. (2.34)
Proof. By Proposition 2.1, there exist ω > 0 and h ∈ C 2 (Ω̄) such that
h(x) > 0 for all x ∈ Ω̄ and h is a supersolution of (L + ω, B, Ω). As u is a
supersolution of (L, B, Ω), we find that
(L + ω)u = Lu + ωu ≥ ωu > 0 in Ω,
since u > 0 in Ω. Also, Bu ≥ 0 on ∂Ω. Thus, u ∈ C 2 (Ω) ∩ C 1 (Ω̄), u > 0, is

a supersolution of (L + ω, B, Ω), and, therefore, according to Corollary 2.1,
u must satisfy (2.32) and (2.33). In particular, u is a superharmonic func-
tion of (L, Ω) such that u(x) > 0 for all x ∈ Ω̄ \ Γ0 . Consequently, by
Definition 1.4, there exists δ > 0 satisfying (2.34).
By the hypotheses H1, H2 and H3, it follows from Corollary 1.2 that
(L, Ω) satisfies the decay property of E. Hopf on Γ0 if Ω satisfies the uniform
interior sphere property in the strong sense on Γ0 . Moreover, according to
Theorem 1.9, this occurs if Ω is of class C 2 . Consequently, owing to Lemma
2.1, the next result holds.
Corollary 2.3. Suppose Ω is of class C 2 and u ∈ C 2 (Ω) ∩ C 1 (Ω̄), u > 0,

is a supersolution of (L, B, Ω). Then, u satisfies (2.32), (2.33), and (2.34)
for some δ = δ(u) > 0.
2.5 Second classification theorem
The main result of this section is the following sharp improvement of The-
orem 2.1, where the positive supersolution h can vanish on Γ0 .
Theorem 2.4. Suppose:
i) (L, Ω) satisfies the decay property of E. Hopf on Γ0 ;

ii) There exist ψ ∈ C 2 (Ω̄) and γ > 0 satisfying (2.22);
iii) (L, B, Ω) has a positive supersolution h ∈ C 2 (Ω) ∩ C 1 (Ω̄), h > 0.
Then, any supersolution u ∈ C 2 (Ω) ∩ C 1 (Ω̄) of (L, B, Ω) must satisfy some

of the Alternatives A1, A2, or A3, of Theorem 2.1.
Remark 2.1. By Corollary 1.2, Theorem 1.9, and Lemma 2.1, conditions
(i) and (ii) are satisfied when Ω is of class C 2 .
Proof. Thanks to Theorem 2.3, the supersolution h satisfies

h(x) > 0 for all x ∈ Ω ∪ Γ1 , (2.35)
∂h
(x) < 0 for all x ∈ h−1 (0) ∩ Γ0 , (2.36)
∂ν
and there exists a constant δ := δ(h) > 0 such that
h(x) ≥ δ dist (x, ∂Ω) for all x ∈ Ω. (2.37)
Let u ∈ C (Ω) ∩ C (Ω̄) be a supersolution of (L, B, Ω). Then, Lu ≥ 0 in Ω
2 1
and Bu ≥ 0 on ∂Ω. Suppose u > 0. Then, thanks again to Theorem 2.3,

u satisfies (2.32) and (2.33), and, therefore, Alternative A2 holds. Conse-
quently, in case u ≥ 0 some of the first two alternatives must occur. To
complete the proof of the theorem it remains to show that Alternative A3
is satisfied if u is somewhere negative. So, suppose
u(x− ) < 0 for some x− ∈ Ω. (2.38)
Subsequently, for every λ ≥ 0, we consider the function vλ defined by
vλ (x) := u(x) + λh(x) , x ∈ Ω̄.
We claim that vλ ≥ 0 in Ω for sufficiently large λ > 0. On the contrary,
assume that, for each integer k ≥ 1, there exists xk ∈ Ω such that
vk (xk ) = u(xk ) + kh(xk ) < 0. (2.39)
As Ω̄ is compact, there exist x0 ∈ Ω̄ and a subsequence of {k}k≥1 , say
{km }m≥1 , such that
lim xkm = x0 .
m→∞
Thanks to (2.39), we have that

1
u(xkm ) + h(xkm ) < 0, m ≥ 1. (2.40)
km
Moreover, by the continuity of u in Ω̄,
1
lim u(xkm ) = 0.
m→∞ km
Thus, letting m → ∞ in (2.40) gives h(x0 ) ≤ 0, and hence
h(x0 ) = 0,
because we are assuming that h ≥ 0 in Ω. Therefore, owing to (2.35),

x0 ∈ h−1 (0) ∩ Γ0 .
This entails Γ0 ̸= ∅. Consequently, if Γ0 = ∅, then vλ ≥ 0 for sufficiently
large λ > 0. So, suppose Γ0 ̸= ∅, and, for each m ≥ 1, let
ykm ∈ Γ0
such that
dist (xkm , Γ0 ) = |xkm − ykm |.
As u is a supersolution of (L, B, Ω), we have that u ≥ 0 on Γ0 and hence,
−u(xkm ) ≤ u(ykm ) − u(xkm ) , m ≥ 1.
Thus, since u ∈ C 1 (Ω̄), this implies that
−u(xkm ) ≤ L|ykm − xkm | = L dist (xkm , Γ0 ), m ≥ 1, (2.41)
where L ≥ 0 is the Lipschitz constant of u in Ω̄.
Now, combining (2.39) and (2.41) yields
km h(xkm ) < −u(xkm ) ≤ L dist (xkm , Γ0 ), m ≥ 1,
and, therefore, (2.37) implies that
δkm dist (xkm , ∂Ω) < L dist (xkm , Γ0 ), m ≥ 1. (2.42)
Finally, as for sufficiently large m
dist (xkm , ∂Ω) = dist (xkm , Γ0 ),
because xkm → x0 as m → ∞, we find from (2.42) that
δkm < L
for sufficiently large m ≥ 1, which is impossible, because
lim km = ∞.
m→∞
This contradiction shows that vλ ≥ 0 for sufficiently large λ > 0.

Let Λ denote the set of λ’s, λ > 0, for which vλ ≥ 0. We have just
proven that Λ ̸= ∅. According to (2.38), λ ̸∈ Λ for sufficiently small λ > 0.
Moreover, since h ≥ 0, we have that [λ, ∞) ⊂ Λ for all λ ∈ Λ. In particular,
µ := inf Λ > 0.
Clearly,
vµ ≥ 0.
Indeed, if vµ (xµ ) < 0 for some xµ ∈ Ω, then, for sufficiently small ϵ > 0,
there exists xµ+ϵ ∈ Ω, xµ+ϵ ∼ xµ , such that
vµ+ϵ (xµ+ϵ ) < 0,
which is impossible, by the definition of µ. Also,
Lvµ = Lu + µLh ≥ 0 in Ω,
and
Bvµ = Bu + µBh ≥ 0 on ∂Ω.
Thus, vµ ∈ C 2 (Ω) ∩ C 1 (Ω̄) is a non-negative supersolution of (L, B, Ω).
Consequently, according to Theorem 2.3, either
vµ = 0, (2.43)
or

 v (x) > 0 ∀ x ∈ Ω ∪ Γ1 ,
 µ
(2.44)

 ∂vµ (x) < 0 ∀ x ∈ v −1 (0) ∩ Γ .
µ 0
∂ν
Suppose (2.43). Then, setting m := −µ < 0, we obtain that
u = mh
and Alternative A3 holds; (2.4) follows easily from this identity taking into
account that u and h are supersolutions of (L, B, Ω) with m < 0. Con-
sequently, to complete the proof it suffices to show that (2.44) contradicts
the minimality of µ. Indeed, by the definition of µ, for every k ≥ 1 there
exists a point xk ∈ Ω such that
( )
1 h(xk )
vµ− k1 (xk ) = u(xk ) + µ − h(xk ) = vµ (xk ) − < 0. (2.45)
k k
Arguing as above, there exist x0 ∈ Ω̄ and a subsequence of {k}k≥1 , say
{km }m≥1 , such that
lim xkm = x0 .
m→∞

h(xkm )
vµ (xkm ) < , m ≥ 1. (2.46)
km
Moreover, by the continuity of h in Ω̄,
h(xkm )
lim = 0.
m→∞ km
Thus, letting m → ∞ in (2.46) shows that

vµ (x0 ) ≤ 0
and, therefore, according to (2.44), we find that
x0 ∈ vµ−1 (0) ∩ Γ0 .
As above, this entails Γ0 ̸= ∅ and ends the proof of the theorem when
Γ0 = ∅. So, suppose Γ0 ̸= ∅, and, for every m ≥ 1, let ykm ∈ Γ0 such that
dist (xkm , Γ0 ) = |xkm − ykm |.
The same argument used above shows that (2.41) holds.
On the other hand, by (2.45), we have that
( )
1
−u(xkm ) > µ − h(xkm ), m ≥ 1,
km
and hence (2.41) yields
( )
1
µ− h(xkm ) < L dist (xkm , Γ0 ), m ≥ 1.
km
Thus, since
lim km = ∞,
m→∞
we have that
Lkm
h(xkm ) < dist (xkm , Γ0 )
µkm − 1
for sufficiently large m. Hence, going back to (2.45) we find that, for suffi-
ciently large m,
h(xkm )
0 > vµ (xkm ) −
km
L
≥ δµ dist (xkm , Γ0 ) − dist (xkm , Γ0 )
µkm − 1
( )
L
= δµ − dist (xkm , Γ0 ),
µkm − 1
because (L, Ω) possesses the decay property of E. Hopf on Γ0 and vµ is a
positive supersolution of (L, B, Ω). We have denoted by δµ > 0 the constant
of the decay property corresponding to vµ . The previous inequality cannot
be satisfied, because it entails
dist (xkm , Γ0 ) < 0
for sufficiently large m. Therefore, vµ = 0. This ends the proof.
2.6 Appendix: Partitions of the unity
Subsequently, we denote by C0∞ (RN ) the set of functions ψ : RN → R of

class C ∞ with compact support
supp ψ := ψ −1 (R \ {0}) .
Also, for every subset A ⊂ RN , we will denote by χA the characteristic
function of A, i.e.,

 1, if x ∈ A,
χA (x) =

0, if x ∈ RN \ A.
Lemma 2.2. Let D, Ω, be two bounded domains of RN , N ≥ 1, with

D̄ ⊂ Ω. Then, there exists ψ ∈ C0∞ (RN ) such that 0 ≤ ψ ≤ 1 in RN , and
ψ=1 in D̄, supp ψ ⊂ Ω̄.
Proof. For every δ > 0, let Eδ denote the function defined through
{ δ2
− δ2 −|x|
Eδ (x) := e
2
, if |x| ≤ δ,
0, if |x| > δ,
and set
(∫ )−1
Cδ := Eδ (x) dx .
RN
Then, the function
ωδ := Cδ Eδ
satisfies
ωδ ∈ C0∞ (RN ), ωδ ≥ 0 in RN , supp ωδ ⊂ B̄δ (0),
and ∫
ωδ (x) dx = 1.
RN
Subsequently, we set
ϵ := dist (∂D, ∂Ω), δ := ϵ/3,
and, for any η > 0,
Dη := D + Bη (0) = {x ∈ RN : dist (x, D) < η}.
It is easy to check that the function
∫
ψ(x) := χD2δ (y) ωδ (x − y) dy, x ∈ RN ,
RN
satisfies all the requirements of the statement.
The following result establishes the existence of a partition of the unity

subordinated to each covering of Ω̄ by open connected neighborhoods.
Proposition 2.2. Let Ω be a bounded domain of RN , N ≥ 1, and m ≥ 1
open connected subsets Uj ⊂ RN , 1 ≤ j ≤ m, such that
∪
m
Ω̄ ⊂ Uj . (2.47)
j=1
Then, there exist ψj ∈ C0∞ (RN ), 1 ≤ j ≤ m, such that
supp ψj ⊂ Ūj , 1 ≤ j ≤ m,
and
∑m
ψj = 1 in Ω̄.
j=1
Any set of m functions ψj , 1 ≤ j ≤ m, satisfying all the requirements

of Proposition 2.2 will be refereed to as a partition of the unity in Ω
subordinated to the covering Uj , 1 ≤ j ≤ m.
Proof. According to (2.47), for sufficiently small ϵ > 0, the open sets
Dj := {x ∈ Uj : dist (x, ∂Uj ) > ϵ}, 1 ≤ j ≤ m,
satisfy
∪
m
Ω̄ ⊂ Dj ,
j=1
because Ω̄ is compact. By Lemma 2.2, for every 1 ≤ j ≤ m, there exists
Ψj ∈ C0∞ (RN ) such that 0 ≤ Ψj ≤ 1 in RN , and
Ψj = 1 in Dj , supp Ψj ⊂ Ūj .
∞
Let Ψ ∈ C0 (R ) be the function defined through
N
∑m
Ψ(x) := Ψj (x), x ∈ RN .
j=1
By construction, the functions
Ψj (x)
ψj (x) := , x ∈ RN , 1 ≤ j ≤ m,
Ψ(x)
satisfy
0 ≤ ψj ≤ 1 in RN , ψj ∈ C ∞ (RN ), supp ψj ⊂ Ūj ,
for all 1 ≤ j ≤ m, and
∑ m
ψj = 1 in Ω̄.
j=1
The main results of this chapter are Theorems 2.1 and 2.4. Both establish
that all non-zero supersolutions of (L, B, Ω) must have constant sign in Ω
if (L, B, Ω) admits a positive supersolution h. Theorem 2.4 is an extremely
sharp version of Theorem 2.1, because in Theorem 2.4 the supersolution h
can vanish on some piece of ∂Ω, while h must be separated away from zero
in Ω̄ for the validity of Theorem 2.1, which is the classical condition of M.
H. Protter and H. F. Weinberger [183] in their Theorem 10 of Chapter 2,
which is Theorem 1.7 of Chapter 1.
Theorem 2.2, which is a straightforward consequence from Theorem
2.1, establishes the positivity of the inverse for the boundary value problem
(2.31) associated to (L + ω, B, Ω) for sufficiently large ω. Theorem 2.2 goes
back to Theorem 6.1 of H. Amann [9] in the special case when Ω is of class
C 2 , as no sign restriction for β was imposed therein. Note that, previously,
M. H. Protter and H. F. Weinberger in Chapter 2 of [183] assumed that
β ≥ 0. But, according to the last paragraph of the proof of Theorem 6.1
on page 239 of H. Amann [9], where it was claimed that
“the assertion follows by an obvious combination of the generalized maximum
principle of Protter and Weinberger [183] with Bony’s maximum principle [28]”.
the reader might believe that Theorem 2.1 is a direct consequence of The-
orem 1.7. Although this is certainly true, we estimate that Theorem 2.1
is far from being an obvious consequence of Theorem 1.7. Nevertheless, it
should be remarked that the last assertion of H. Amann in the statement
of Theorem 6.1 of [9], where it is asserted that
∂u
(x) < 0 for x ∈ Γ0 ,
∂ν
does not seem to be true unless u(x) = 0.
Theorem 2.4 goes back to Theorem 2 of W. Walter [224] in the special
case when Γ1 = ∅. It was substantially refined by J. López-Gómez in
Theorem 5.2 of [144] to cover the general case treated in this chapter.
Theorem 2.1 should be weighted against Theorem 1 of W. Walter [224],
whose proof is, according to Section 4 of W. Walter [224],
“closely related to the idea of families of upper solutions which goes back to
A. McNabb [160] and is also known under the name Serrin’s sweeping principle.”
Some of the regularity assumptions imposed in the results of this chapter

can be substantially relaxed. For example, in Lemma 2.1 the regularity of

∂Ω can be relaxed up to assume that Γ1 is of class C 2 . The regularity of Γ0
is far from necessary, because ψ can be taken to be positive on Γ0 .
The exposition of this chapter has been elaborated from the materials
of J. López-Gómez [144].
In Chapter 7, using the maximum principle of J. M. Bony [28], it will
become apparent that most of the results of Chapters 1 and 2 are still valid
in the weak sense within the context of the Sobolev spaces.

Chapter 3
Representation theorems
Let H be a real vector space. A scalar product ⟨·, ·⟩ in H is a bilinear form

⟨·, ·⟩ : H × H → R
symmetric and positive definite. Every scalar product ⟨·, ·⟩ induces a canon-
ical norm in H through
√
∥u∥ := ⟨u, u⟩, u ∈ H,
which satisfies the Cauchy–Schwarz inequality
|⟨u, v⟩| ≤ ∥u∥ ∥v∥, u, v ∈ H. (3.1)
Indeed, for every u, v ∈ H and t ∈ R, we have that
0 ≤ ∥tu + v∥2 = ⟨tu + v, tu + v⟩ = t2 ∥u∥2 + 2t⟨u, v⟩ + ∥v∥2 .
Thus, the polynomial
P (t) := ∥u∥2 t2 + 2⟨u, v⟩t + ∥v∥2 , t ∈ R,
admits at most a unique real root. Consequently,
⟨u, v⟩2 ≤ ∥u∥2 ∥v∥2 , u, v ∈ H,
and, extracting square roots in this inequality, (3.1) holds. Moreover, the
parallelogram identity
2 2
u+v u−v ∥u∥2 + ∥v∥2
+ = , u, v ∈ H, (3.2)
2 2 2
is satisfied. Indeed,
∥u + v∥2 + ∥u − v∥2 = ⟨u + v, u + v⟩ + ⟨u − v, u − v⟩
= ∥u∥2 + ∥v∥2 + 2⟨u, v⟩ + ∥u∥2 + ∥v∥2 − 2⟨u, v⟩
= 2(∥u∥2 + ∥v∥2 ).
63
u+v
u−v
Fig. 3.1 The parallelogram of sides u and v
Geometrically, (3.2) establishes that the sum of the squares of the lengths
of the diagonals of a parallelogram equals the sum of the squares of the
lengths of all sides, as illustrated by Figure 3.1.
The real vector space H with the scalar product ⟨·, ·⟩ is said to be a
Hilbert space if the normed vector space (H, ∥ · ∥) is complete, i.e., if it
is a Banach space. Throughout this chapter, we will suppose that H is a
Hilbert space with scalar product ⟨·, ·⟩ and associated norm ∥ · ∥.
In this chapter we are going to study the theorem of G. Stampacchia
[212], which has been a milestone for the development of the calculus of
variations and its applications. As a byproduct, we will derive from it
the representation theorem of P. D. Lax and A. N. Milgram [122], which
is a substantial extension of the representation theorem of F. Riesz [191].
Essentially, the theorem of G. Stampacchia [212] is an abstract nonlinear
counterpart of the representation theorem of P. D. Lax and A. N. Milgram
[122], which is utterly linear.
This chapter is distributed as follows. Section 3.1 establishes the exis-
tence of the projection operator associated to each closed and convex subset
K of H, Section 3.2 shows how the projection operator equals the orthog-
onal projection when K is a linear subspace of H, Section 3.3 studies the
representation theorem of F. Riesz, Section 3.4 introduces some basic con-
cepts of the theory of bilinear forms, Section 3.5 studies the theorem of G.
Stampacchia, Section 3.6 derives from it the theorem of P. D. Lax and A.
N. Milgram, and, finally, Section 3.7 establishes the existence of continuous
projections on any convex and closed subset of a uniformly convex Banach
space.
Representation theorems 65
3.1 The projection on a closed convex set
The following result constructs the projection operator on any closed and
convex subset of the Hilbert space H.
Theorem 3.1. Let K ⊂ H be a closed and convex set. Then, for every
u ∈ H, there exists a unique PK u ∈ K such that
∥u − PK u∥ = dist (u, K) = min ∥u − k∥.
k∈K
Necessarily,
PK k = k if k ∈ K,
and, for every u ∈ H, PK u is the unique element of K for which
⟨u − PK u, k − PK u⟩ ≤ 0 for all k ∈ K. (3.3)
The underlying map
P
H −→
K
K
u 7→ PK u
is said to be the projection of H on K.
As illustrated by Figure 3.2, Theorem 3.1 has an obvious geometrical

meaning. For every u ∈ H \ K, the orthogonal hyperplane to the vector
u − PK u through the point PK u, given by
T := {h ∈ H : ⟨u − PK u, h − PK u⟩ = 0} ,
divides the whole space H into the two closed and convex subspaces
H+ := {h ∈ H : ⟨u − PK u, h − PK u⟩ ≥ 0} ,
H− := {h ∈ H : ⟨u − PK u, h − PK u⟩ ≤ 0} ,
in such a way that, according to (3.3), K ⊂ H− and u ∈ H+ . Consequently,
(3.3) does actually provide us with a separation property. As the closed
convex K is the set of fixed points of PK , K and PK determine each other.
Proof. Fix u ∈ H. Then, the quantity
d := dist (u, K) = inf ∥u − k∥ ≥ 0
k∈K
is well defined. Actually, d = 0 if and only if u ∈ K, since K is closed. In

such case, we define
PK u := u.
PKu
k
T
Fig. 3.2 The projection operator
Now, suppose u ∈ H \ K. Then, d > 0 and there exists a sequence {kn }n≥1
in K such that
d := lim ∥u − kn ∥. (3.4)
n→∞
Setting
dn := ∥u − kn ∥, n ≥ 1,
we find from (3.2) that, for every n, m ≥ 1,
2 2
u−kn +u−km u−kn −(u−km ) ∥u−kn ∥2 +∥u−km ∥2
+ =
2 2 2
and hence,
2 2
kn + km km − kn d2n + d2m
u− + = .
2 2 2
As K is convex, we have that
kn + km
kn , k m ∈ K =⇒ ∈K
2
for all n, m ≥ 1, and hence,
2
kn + km
d2 = dist2 (u, K) ≤ u − .
2
Thus,
2
km − kn d2n + d2m
d2 + ≤ , n, m ≥ 1,
2 2
and, therefore,
2
km − kn d2 + d2m
≤ n − d2 → 0 as n, m → ∞.
2 2
This shows that {kn }n≥1 is a Cauchy sequence in K and, since K is closed,
there exists PK u ∈ K such that
lim kn = PK u.
n→∞
By (3.4), it is apparent that
d = ∥u − PK u∥
and, consequently, d is attained at PK u.
Now, we will show that, for any m ∈ K, the next identity holds
∥u − m∥ = dist (u, K) = min ∥u − k∥ (3.5)
k∈K
if and only if
⟨u − m, k − m⟩ ≤ 0 for all k ∈ K. (3.6)
Suppose m ∈ K satisfies (3.5) and let k ∈ K. As K is convex,
(1 − t)m + tk ∈ K for all t ∈ [0, 1].
Thus, owing to (3.5),
∥u − m∥ ≤ ∥u − (1 − t)m − tk∥ = ∥u − m + t(m − k)∥
for every t ∈ [0, 1], and hence,
∥u − m∥2 ≤ ∥u − m + t(m − k)∥2
= ∥u − m∥2 + t2 ∥m − k∥2 + 2t⟨u − m, m − k⟩.
Consequently,
t2 ∥m − k∥2 + 2t⟨u − m, m − k⟩ ≥ 0 for all t ∈ [0, 1],
and, therefore, for every t ∈ (0, 1],
2⟨u − m, k − m⟩ ≤ t∥m − k∥2 .
Letting t ↓ 0 in this estimate, (3.6) holds.
Conversely, let m ∈ K satisfying (3.6) and pick a k ∈ K. Then,
∥u − m∥2 −∥u − k∥2 = ⟨u − m, u − m⟩ − ⟨u − k, u − k⟩
( )
= ∥u∥2 + ∥m∥2 − 2⟨u, m⟩ − ∥u∥2 + ∥k∥2 − 2⟨u, k⟩
= ∥m∥2 − ∥k∥2 + 2⟨u, k − m⟩
= 2⟨u − m, k − m⟩ + 2⟨m, k − m⟩ + ∥m∥2 − ∥k∥2
= 2⟨u − m, k − m⟩ + 2⟨m, k⟩ − ∥m∥2 − ∥k∥2
= 2⟨u − m, k − m⟩ − ∥m − k∥2 .
Thus, according to (3.6), we obtain that

∥u − m∥2 − ∥u − k∥2 ≤ −∥m − k∥2 ≤ 0
and, consequently,
∥u − m∥ ≤ ∥u − k∥ for all k ∈ K.
Equivalently,
∥u − m∥ = min ∥u − k∥.
k∈K
Therefore, (3.5) and (3.6) are indeed equivalent.

It remains to prove that PK u is the unique k ∈ K for which
d = ∥u − k∥.
Let u ∈ H and m1 , m2 ∈ K such that
∥u − m1 ∥ = ∥u − m2 ∥ = min ∥u − k∥.
k∈K
Then, by the equivalence of (3.5) and (3.6), we have that

⟨u − mj , k − mj ⟩ ≤ 0 for all (k, j) ∈ K × {1, 2},
and, in particular,
⟨u − m1 , m2 − m1 ⟩ ≤ 0 and ⟨u − m2 , m1 − m2 ⟩ ≤ 0.
Thus, adding these inequalities, yields
0 ≥ ⟨u − m1 , m2 − m1 ⟩ + ⟨u − m2 , m1 − m2 ⟩
= ⟨u, m2 − m1 ⟩ − ⟨m1 , m2 − m1 ⟩ + ⟨u, m1 − m2 ⟩ − ⟨m2 , m1 − m2 ⟩
= ∥m2 − m1 ∥2
and, therefore, m1 = m2 , which concludes the proof.
The next result reveals that the projection operator
PK : H → K
is globally Lipschitz with constant one.
Proposition 3.1. Let K ⊂ H be a closed and convex subset of H, and

PK : H → K the projection operator of H on K. Then,
∥PK u − PK v∥ ≤ ∥u − v∥ for all u, v ∈ H. (3.7)
As PK k = k for all k ∈ K, the Lipschitz constant of PK equals 1.
Proof. According to Theorem 3.1, and, most precisely, due to (3.3), we

have that
⟨u − PK u, k − PK u⟩ ≤ 0 and ⟨v − PK v, k − PK v⟩ ≤ 0
for all u, v ∈ H and k ∈ K. In particular, for every u, v ∈ H,
⟨u − PK u, PK v − PK u⟩ ≤ 0, ⟨v − PK v, PK u − PK v⟩ ≤ 0,
and, hence, adding these inequalities shows that
0 ≥ ⟨u − PK u, PK v − PK u⟩ + ⟨v − PK v, PK u − PK v⟩
= ⟨u − v, PK v − PK u⟩ + ⟨PK v − PK u, PK v − PK u⟩
= ⟨u − v, PK v − PK u⟩ + ∥PK u − PK v∥2 .
Consequently, for every u, v ∈ H, we find that
∥PK u − PK v∥2 ≤ ⟨u − v, PK u − PK v⟩
and, therefore, according to (3.1), we conclude that
∥PK u − PK v∥2 ≤ ∥PK u − PK v∥ ∥u − v∥.
This shows (3.7) and ends the proof.
3.2 The orthogonal projection on a closed subspace
Subsequently, given a closed subspace N of H, we denote by N ⊥ the or-

thogonal of N in H
N ⊥ := {u ∈ H : ⟨u, n⟩ = 0 for all n ∈ N }.
As any subspace of H is convex, Theorem 3.1 guarantees that the projection
PN : H → N is well defined. The next result identifies PN as the orthogonal
projection of H on N .
Theorem 3.2. Let N be a closed subspace of H, and denote by PN the

projection of H on N . Then,
PN ∈ L(H, N ), ∥PN ∥L(H,N ) = 1,
and
PN |N = IN , R[IH − PN ] ⊂ N ⊥ .
Consequently, PN is the orthogonal projection of H on N .
N
P u
N
Fig. 3.3 The orthogonal projection
Proof. Let u ∈ H. By Theorem 3.1, we have that

⟨u − PN u, n − PN u⟩ ≤ 0 for all n ∈ N,
and hence, for every n ∈ N and λ > 0,
0 ≥ ⟨u − PN u, λn − PN u⟩ = λ⟨u − PN u, n⟩ − ⟨u − PN u, PN u⟩,
since λn ∈ N for all λ ∈ R. Thus, for every n ∈ N and λ > 0,
1
⟨u − PN u, n⟩ ≤⟨u − PN u, PN u⟩, (3.8)
λ
and hence, letting λ ↑ ∞, (3.8) implies that
⟨u − PN u, n⟩ ≤ 0 for all n ∈ N.
As N is a linear subspace, we also have that
⟨u − PN u, −n⟩ ≥ 0 for all n ∈ N.
Consequently,
⟨u − PN u, n⟩ = 0 for all n ∈ N (3.9)
and, therefore,
R[IH − PN ] ⊂ N ⊥ .
The fact that
PN n = n, n ∈ N,
is a direct consequence from Theorem 3.1.
By Proposition 3.1, to conclude the proof it suffices to prove that PN is
a linear operator. Indeed, due to (3.9), for every λ, µ ∈ R, u, v ∈ H, and
n ∈ N , we have that
⟨λu + µv − PN (λu + µv), n⟩ = 0.
Similarly,
⟨λu+µv−λPN u−µPN v, n⟩ = λ⟨u−PN u, n⟩+µ⟨v−PN v, n⟩ = 0.
Thus, subtracting both identities shows that
⟨PN (λu + µv) − λPN u − µPN v, n⟩ = 0
for all n ∈ N . In particular, making the choice
n := PN (λu + µv) − λPN u − µPN v
it becomes apparent that, for every u, v ∈ H and λ, µ ∈ R,
∥PN (λu + µv) − λPN u − µPN v∥2 = 0,
which implies
PN (λu + µv) = λPN u + µPN v
and concludes the proof.
3.3 The representation theorem of F. Riesz
In this section, we will obtain the representation theorem of F. Riesz from

the theory already developed in Sections 3.2 and 3.3. It identifies any
Hilbert space H with its dual space
H ′ = L(H, R).
Consequently, any real Hilbert space is reflexive, in the sense that
H ′′ = H.
The abstract theory of Hilbert spaces relies on this feature.
Theorem 3.3 (of representation of F. Riesz). Let H be a real Hilbert

space. Then, for every φ ∈ H ′ there exists a unique uφ ∈ H such that
φ(u) = ⟨uφ , u⟩ for all u ∈ H. (3.10)
Moreover,
∥φ∥H ′ = ∥uφ ∥ for all φ ∈ H ′ ,
and, actually, the map
D
H ′ −→ H
(3.11)
φ 7→ D(φ) := uφ
establishes an isometric isomorphism between H ′ and H.
Proof. If φ = 0, then (3.10) is satisfied if and only if uφ = 0. So, suppose

φ ∈ H ′ \ {0} and consider
N := N [φ] = {u ∈ H : φ(u) = 0}.
Then, N is a closed proper subspace of H.
Next, we will show that there exists p ∈ N ⊥ with ∥p∥ = 1 such that
H = N ⊕ span [p]. (3.12)
Indeed, pick u ∈ H \ N , and let PN be the orthogonal projection of H on
N . Then, by the definition of N , φ(u) ̸= 0 and, hence,
φ(u − PN u) = φ(u) − φ(PN u) = φ(u) ̸= 0.
Thus,
u − PN u
q := ∈ H \ N.
φ(u − PN u)
On the other hand, owing to Theorem 3.2,
u − PN u ∈ R[IH − PN ] ⊂ N ⊥
and, consequently,
⟨q, n⟩ = 0 for all n ∈ N.
Therefore, the vector
q
p := ∈H \N
∥q∥
satisfies
∥p∥ = 1, p ∈ N ⊥. (3.13)
Moreover, it satisfies (3.12). Indeed, for every v ∈ H, we have that
φ(v) φ(v)
v=v− p+ p
φ(p) φ(p)
and, obviously,
φ(v)
φ(v − p) = 0.
φ(p)
Thus,
H = N + span [φ].
Moreover, if
n + λp = m + µp
for some λ, µ ∈ R and n, m ∈ N , then

n − m = (µ − λ)p
and hence, it follows from (3.13) that
∥n − m∥2 = (µ − λ)⟨n − m, p⟩ = 0.
Consequently,
n = m and (µ − λ)p = 0,
which implies λ = µ. Therefore, (3.12) holds. The algebraic direct sum
is actually topological because the orthogonal projections of H on N and
span [p] are continuous.
Next, we will show that
uφ := φ(p)p
satisfies (3.10). Indeed, according to (3.12), for each u ∈ H there exist
n ∈ N and λ ∈ R (unique) such that
u = n + λp.
Thus, since φ(n) = 0, we find from (3.13) that
⟨uφ , u⟩ = φ(p)⟨p, n + λp⟩
= φ(p)⟨p, n⟩ + λφ(p)∥p∥2
= λφ(p) = φ(n + λp)
= φ(u),
which concludes the proof of (3.10).
To prove the uniqueness of uφ , assume that there are u1φ , u2φ ∈ H such
that
⟨u1φ , u⟩ = ⟨u2φ , u⟩ for all u ∈ H.
Then, for every u ∈ H,
⟨u1φ − u2φ , u⟩ = 0
and, in particular,
0 = ⟨u1φ − u2φ , u1φ − u2φ ⟩ = ∥u1φ − u2φ ∥2 ,
which implies u1φ = u2φ and shows the uniqueness of uφ . Consequently, the
map (3.11) is well defined.
To complete the proof of the theorem it remains to show that the map
D defined by (3.11) is an isometric isomorphism.
Let λ, µ ∈ R and φ, ψ ∈ H ′ . Then, for every u ∈ H, we have that
(λφ + µψ)(u) = λφ(u) + µψ(u)

= λ⟨uφ , u⟩ + µ⟨uψ , u⟩
= ⟨λuφ + µuψ , u⟩
and hence,
uλφ+µψ = λuφ + µuψ ,
by uniqueness. In other words,
D(λφ + µψ) = λD(φ) + µD(ψ),
and, consequently, D is linear.

Now, suppose
D(φ) = uφ = 0.
Then, φ(u) = 0 for all u ∈ H and, so, φ = 0. Therefore, D is injective. To

show that it is surjective, for a given u ∈ H, let φu : H → R denote the
map defined by
φu (v) = ⟨u, v⟩, v ∈ H.
The map φu is linear and continuous, because
|φu (v)| ≤ ∥u∥ ∥v∥ for all v ∈ H.
Consequently, φu ∈ H ′ ,
∥φu ∥H ′ ≤ ∥u∥,
and, necessarily, by uniqueness,
D(φu ) = u.
Therefore, D establishes an isomorphism between H and H ′ .

Finally, owing to (3.10), we have that, for every φ ∈ H ′ and u ∈ H,
|φ(u)| ≤ ∥uφ ∥ ∥u∥, |φ(uφ )| = ∥uφ ∥2 .
Thus,
∥φ∥H ′ = ∥uφ ∥ = ∥D(φ)∥,
and, therefore, D is isometric. The proof is complete.

3.4 Continuity and coercivity of bilinear forms
The following concept has shown to be pivotal for the development of the
theory of partial differential equations.
Definition 3.1. Let H be a real Hilbert space and a : H × H → R a

bilinear form. Then:
i) a is said to be continuous if there is a constant β > 0 such that

|a(u, v)| ≤ β ∥u∥ ∥v∥ for all u, v ∈ H. (3.14)
ii) a is said to be coercive if there is a constant α > 0 such that
a(u, u) ≥ α ∥u∥2 for all u ∈ H. (3.15)
If the bilinear form a : H × H → R is continuous and coercive, then

α ∥u∥2 ≤ a(u, u) ≤ β ∥u∥2
for all u ∈ H. Actually, the following result holds.
Proposition 3.2. Let a : H × H → R be a continuous and coercive

bilinear form satisfying (3.14) and (3.15). Then, there exists A ∈ Iso(H)
such that
α ∥u∥ ≤ ∥Au∥ ≤ β∥u∥ for all u ∈ H, (3.16)
and
a(u, v) = ⟨Au, v⟩ for all u, v ∈ H. (3.17)
Proof. Fix u ∈ H and consider the map φu : H → R defined through

φu (v) := a(u, v), v ∈ H.
As a is bilinear, φu is a linear map. Moreover, by (3.14), we have that
|φu (v)| = |a(u, v)| ≤ β ∥u∥ ∥v∥ for every v ∈ H
and hence φu ∈ H ′ . Therefore, according to Theorem 3.3, there exists a
unique ũ ∈ H such that
φu (v) = ⟨ũ, v⟩ for all v ∈ H.
Let A : H → H be the map defined by
Au := ũ, u ∈ H.
By construction, A satisfies (3.17). To complete the proof it remains to

show that A ∈ Iso(H). Indeed, for every u, v, w ∈ H and λ, µ ∈ R,
⟨A(λu + µv), w⟩ = a(λu + µv, w)
= λa(u, w) + µa(v, w)
= λ⟨Au, w⟩ + µ⟨Av, w⟩
= ⟨λAu + µAv, w⟩
and hence,
⟨A(λu + µv) − (λAu + µAv), w⟩ = 0.
Thus,
A(λu + µv) = λAu + µAv
for all u, v ∈ H and λ, µ ∈ R. Therefore, A is linear.
Thanks to (3.15) and (3.17), it becomes apparent that
α ∥u∥2 ≤ |a(u, u)| = |⟨Au, u⟩| ≤ ∥Au∥ ∥u∥
for all u ∈ H, and, consequently,
α ∥u∥ ≤ ∥Au∥,
which provides us with the lower estimate of (3.16). Moreover, it follows
from (3.14) that, for every u ∈ H,
∥Au∥2 = |⟨Au, Au⟩| = |a(u, Au)| ≤ β ∥u∥ ∥Au∥
and, therefore,
∥Au∥ ≤ β ∥u∥ for all u ∈ H,
which concludes the proof of (3.16). This ends the proof, because (3.16)
entails that A ∈ Iso(H).
3.5 The theorem of G. Stampacchia
The main result of this section reads as follows.
Theorem 3.4 (of G. Stampacchia). Suppose a : H × H → R is a con-

tinuous and coercive bilinear form, and K is a closed and convex subset of
H. Then, for each φ ∈ H ′ there exists a unique f := f (φ) ∈ K such that
a(f, k − f ) ≥ φ(k − f ) for all k ∈ K.
Proof. Fix φ ∈ H ′ . Then, according to Theorem 3.3, there exists a

unique uφ ∈ H such that
φ(u) = ⟨uφ , u⟩ for all u ∈ H.
Consequently, it suffices to prove that there is a unique f ∈ K for which
a(f, k − f ) ≥ ⟨uφ , k − f ⟩ for all k ∈ K. (3.18)
According to Proposition 3.2, there exists A ∈ Iso(H) for which (3.17)
holds. Obviously, owing to (3.17), the inequality (3.18) can be expressed as
⟨Af, k − f ⟩ ≥ ⟨uφ , k − f ⟩ for all k ∈ K.
Equivalently, for some ρ > 0,
⟨ρ uφ − ρ Af + f − f, k − f ⟩ ≤ 0, for all k ∈ K. (3.19)
Let PK denote the projection of H on K. According to Theorem 3.1, for
every u ∈ H, PK u is characterized through
⟨u − PK u, k − PK u⟩ ≤ 0 for all k ∈ K.
Thus, condition (3.19) holds if
f = PK (ρ uφ − ρ Af + f ).
Therefore, it suffices to prove that there exists ρ > 0 for which the operator
S : K → K defined by
S(k) := PK (ρ uφ − ρAk + k), k ∈ K,
has a fixed point. By Proposition 3.1, we have that, for every k1 , k2 ∈ K
and ρ > 0,
∥S(k1 )−S(k2 )∥ = ∥PK (ρ uφ −ρAk1 +k1 )−PK (ρ uφ − ρAk2 +k2 )∥
≤ ∥ρ uφ − ρAk1 + k1 − ρ uφ + ρAk2 − k2 ∥
= ∥k1 − k2 − ρA(k1 − k2 )∥
≤ ∥IH − ρA∥L(H) ∥k1 − k2 ∥.
Thus, thanks to the contracting mapping theorem, it suffices to show that
there exists ρ > 0 such that
∥IH − ρA∥L(H) < 1. (3.20)
Indeed, for every ρ > 0 and u ∈ H, we have that
∥(IH − ρA)u∥2 = ⟨u − ρAu, u − ρAu⟩ = ∥u∥2 +ρ2 ∥Au∥2 −2ρ⟨Au, u⟩,
and hence, it becomes apparent from (3.15)–(3.17) that
∥(IH − ρA)u∥2 = ∥u∥2 + ρ2 ∥Au∥2 − 2ρa(u, u)
( )
≤ 1 + ρ2 β 2 − 2ρα ∥u∥2 .
Consequently, √
∥IH − ρA∥L(H) ≤ 1 + ρ2 β 2 − 2ρα
for every ρ > 0, and, therefore, (3.20) holds provided
0 < ρ < 2α/β 2 .
The proof is complete.
3.6 The theorem of P. D. Lax and A. N. Milgram
When K = H, Theorem 3.4 provides us with a very celebrated result of P.

D. Lax and A. N. Milgram, which is a substantial improvement of Riesz’
representation Theorem 3.3, as the underlying bilinear form is not required
to be symmetric and, so, it might not define a scalar product.
Theorem 3.5 (of P. D. Lax and A. N. Milgram). Let a : H 2 → R be

a continuous and coercive bilinear form. Then, for every φ ∈ H ′ there
exists a unique uφ ∈ H such that
φ(u) = a(uφ , u) for all u ∈ H. (3.21)
Proof. Choose K := H and let φ ∈ H ′ . Then, according to Theorem 3.4,

there exists a unique uφ ∈ H such that
a(uφ , u − uφ ) ≥ φ(u − uφ ) for all u ∈ H.
Consequently,
a(uφ , v) ≥ φ(v), for all v ∈ H,
and hence,
−a(uφ , v) = a(uφ , −v) ≥ φ(−v) = −φ(v).
Therefore, (3.21) holds. The proof is complete.
If, in addition, the bilinear form a(·, ·) is symmetric, i.e.,
a(u, v) = a(v, u) for every u, v ∈ H,
then, Theorem 3.5 follows directly from Theorem 3.3, as the map
⟨·,·⟩a
H × H −→ R
(u, v) 7→ ⟨u, v⟩a := a(u, v)
provides us with a scalar product in H and hence we can apply Theorem
3.3 to get Theorem 3.5. Therefore, the importance of Theorem 3.5 relies
on the fact that it does not require the symmetry of a.
3.7 Projecting on a closed convex set of a u.c. B-space
Essentially, this section generalizes Theorem 3.1 to the more general context
of uniformly convex Banach spaces (u.c. B-spaces). But the projection on
a closed subspace of a u.c. B-space, though continuous, is not necessarily
linear, unless the subspace is a hyperplane. This is in strong contrast with

the situation already described in the context of Hilbert spaces.
This section is divided into four parts. The first one introduces some
basic concepts and preliminaries. The second one shows the existence and
continuity of the projection operator on any closed and convex subset of a
u.c. B-space. The third section shows some linear properties of the projec-
tion on a closed subspace. Among them, it proves that it is homogeneous of
degree one. Finally, the fourth section shows that the projection on a closed
hyperplane is linear. It remains an open problem to ascertain whether this
property holds in the context of u.c. B-spaces. Throughout this section, X
will stand for a real Banach space with norm ∥ · ∥.
3.7.1 Basic concepts and preliminaries

This section collects some basic concepts and results whose proofs will not
be given here, as they are outside the general scope of this book.
The next concept involves a geometrical property of the unit ball that
might not be satisfied for any equivalent norm. Basically, it establishes that
the unit ball is sufficiently curved.
Definition 3.2 (Uniformly convex B-space). The Banach space X is

said to be uniformly convex if for every ϵ > 0 there exists δ = δ(ϵ) > 0 such
that

x, y ∈ X 
x+y
∥x∥ ≤ 1, ∥y∥ ≤ 1 =⇒ ∥ ∥ < 1 − δ. (3.22)
 2
∥x − y∥ > ϵ
To simplify notations, X is said to be a u.c. B-space if it is a uniformly

convex real Banach space. Suppose X = RN . Then, though all norms in
X are equivalent, and, in particular, so are the norms
  p1
∑N
∥x∥p =  |xj |p  , x = (x1 , · · · , xn ) ∈ RN ,
j=1
for all p ∈ [1, ∞), it is easily seen that (X, ∥ · ∥p ) is uniformly convex if
p ∈ (1, ∞), while this property fails for p = 1. More generally, any Hilbert
space is a u.c. B-space.
The concept of uniform convexity goes back to J. A. Clarkson [42],
where it was shown that Lp (Ω) is a u.c. B-space for all p ∈ (1, ∞). As a by-
product, the spaces W 2,p (Ω), 1 < p < ∞, which are going to be introduced
in Chapter 4, are u.c. B-spaces, because there is a natural isometry between

W 2,p (Ω) and a certain product of Lp (Ω)’s. A short time later, the following
result was established by D. P. Milman, [162] and, independently, by B. J.
Pettis [176].
Theorem 3.6 (of D. P. Milman and B. J. Pettis ). Every u.c. B-

space X is reflexive, i.e., X ′′ = X.
This result establishes an astonishing connection between a geometrical

property, the uniform convexity of the unit ball, and a rather topological
property, the reflexivity. Other proofs of this result, some of them very
short, were given by S. Kakutani [110], J. R. Ringrose [192], J. Linden-
strauss and L. Tzafriri [128] (see the bottom of p. 127 in [128]), and H.
Brézis [29] (see the proof of Theorem III.29). According to M. M. Day [48],
there are reflexive B-spaces which are not equivalent to any u.c. B-space.
We now collect some classical results on weak topologies that will be
used in the proof of the main theorem of this section. Subsequently, for
any f ∈ X ′ , we denote by φf : X → R the map defined by
φf (x) = f (x), x ∈ X.
Then, the weak topology σ(X, X ′ ) is defined as the weakest topology in
X for which every functional φf , f ∈ X ′ , is continuous.
The next result establishes that, for convex sets, the concepts of strong
and weak closeness coincide; it is Theorem III.7 of H. Brézis [29]. A set
C ⊂ X is said to be weakly closed if it is closed in the weak topology
σ(X, X ′ ).
Theorem 3.7. Let C ⊂ X be a convex subset of the Banach space X.

Then, C is weakly closed if and only if C is closed in the strong topology.
The next result is a consequence from Theorem 3.7; it is Corollary III.8

of H. Brézis [29]. Subsequently, given a subset Y ⊂ X and a functional
Φ : Y → (−∞, ∞], Φ is said to be lower semi continuous (l.s.c.) if, for
every yn , y ∈ Y , n ≥ 1,
lim yn = y implies Φ(y) ≤ lim inf Φ(yn ).
n→∞ n→∞
Obviously, Φ is l.s.c. if it is continuous.
Corollary 3.1. Let Φ : X → (−∞, ∞] be a l.s.c. convex functional. Then,

lim xn = x in σ(X, X ′ ) implies Φ(x) ≤ lim inf Φ(xn ).
n→∞ n→∞
The following result provides us with a pivotal characterization of the

reflexivity attributable to W. F. Eberlein [59] and V. L. Shmulyan [202];
it is Theorem III.16 of H. Brézis [29]. The interested reader might wish to
have a look at p. 141 of K. Yosida [227] too.
Theorem 3.8 (of W. F. Eberlein and V. L. Shmulyan). A B-space

X is reflexive if and only if the unit ball
B := {x ∈ X : ∥x∥ ≤ 1}
is compact in the weak topology σ(X, X ′ ).
The next theorem is a classical pivotal result in the calculus of variations;

it is Corollary III.20 of H. Brézis [29].
Theorem 3.9. Let X be a reflexive B-space, K ⊂ X a non-empty closed

and convex set, and Φ : K −→ (−∞, ∞] a convex l.s.c. functional, Φ ̸= ∞,
such that
lim Φ(k) = ∞ (3.23)

k∈K
∥k∥→∞
if K is unbounded. Then, there exists kmin ∈ K such that
Φ(kmin ) = min Φ(k). (3.24)

k∈K
Any point kmin ∈ K satisfying (3.24) is called a minimizer of Φ in K.

In general, the minimizer is not unique. In the context of the calculus of
variations, any functional Φ satisfying (3.23) is said to be coercive on K.
Finally, we need the following useful property of the u.c. B-spaces; it is
Proposition III.30 of H. Brézis [29]
Proposition 3.3. Let X be a u.c. B-space, and {xn }n≥1 a sequence of X

such that
lim xn = x in σ(X, X ′ ) and lim sup ∥xn ∥ ≤ ∥x∥.

n→∞ n→∞
Then,
lim ∥xn − x∥ = 0.
n→∞
Naturally, the reader is referred to Chapter III of H. Brézis [29] for the
proofs of all the previous results.
3.7.2 The projection theorem

Theorem 3.10. Let X be a u.c. B-space and K ⊂ X a closed and convex

proper subset. Then, for every x ∈ X, there exists a unique PK x ∈ K such
that
∥x − PK x∥ = dist (x, K) = min ∥x − k∥.
k∈K
Necessarily, PK k = k for all k ∈ K. Moreover, the associated map

P
X −→
K
K
x 7→ PK x
is continuous. The operator PK is called the projection of X on K.
Proof. Fixed x ∈ X, let Φ : K → [0, ∞) denote the distance map

Φ(k) := ∥x − k∥ for all k ∈ K.
As, for every k1 , k2 ∈ K and t ∈ [0, 1], we have that
Φ(tk1 + (1 − t)k2 ) = ∥x − tk1 − (1 − t)k2 ∥
= ∥t(x − k1 ) + (1 − t)(x − k2 )∥
≤ t∥x − k1 ∥ + (1 − t)∥x − k2 ∥
= tΦ(k1 ) + (1 − t)Φ(k2 ),
the map Φ is convex. Moreover, since
|Φ(k1 ) − Φ(k2 )| = |∥x − k1 ∥ − ∥x − k2 ∥| ≤ ∥k1 − k2 ∥,
Φ is continuous and hence lower semi-continuous. The coercivity of Φ is a
direct consequence from the estimate
|∥k∥ − ∥x∥| ≤ ∥k − x∥ = Φ(k),
which implies (3.23).
According to Theorem 3.6, X is reflexive. Therefore, by Theorem 3.9,
Φ admits a minimizer in K.
It should be noted that the existence of the minimizer is guaranteed
as soon as X is reflexive. The uniform convexity of X is necessary for
the uniqueness of the minimizer, as it will become apparent soon. The
uniqueness is obvious if x ∈ K. So, suppose that x ∈ X \ K and there are
k1 , k2 ∈ K, k1 ̸= k2 , such that
0 < ∥x − k1 ∥ = ∥x − k2 ∥ = dist(x, K) ≤ ∥x − k∥ (3.25)
for all k ∈ K. Then, setting

x kj
x̃ := , k̃j := , j ∈ {1, 2},
∥x − kj ∥ ∥x − kj ∥
and dividing (3.25) by ∥x − k1 ∥, we are led to the identity
k
1 = ∥x̃ − k̃1 ∥ = ∥x̃ − k̃2 ∥ ≤ ∥x̃ − ∥
∥x − k1 ∥
for all k ∈ K. In particular, as K is convex, we obtain that
tk1 + (1 − t)k2
1 = ∥x̃ − k̃1 ∥ = ∥x̃ − k̃2 ∥ ≤ ∥x̃ − ∥
∥x − k1 ∥
for all t ∈ [0, 1]. Consequently,
1 ≤ ∥x̃ − tk̃1 − (1 − t)k̃2 ∥ = ∥t(x̃ − k̃1 ) + (1 − t)(x̃ − k̃2 )∥
≤ t∥x̃ − k̃1 ∥ + (1 − t)∥x̃ − k̃2 ∥ = t + 1 − t = 1
and, therefore,
∥x̃ − tk̃1 − (1 − t)k̃2 ∥ = 1 ∀ t ∈ [0, 1].
In particular,
x̃ − k̃1 x̃ − k̃2
+ ∥ ∥ = 1. (3.26)
2 2
By construction, we already know that
x̃ − k̃1 x̃ − k̃2 1
∥ ∥=∥ ∥=
2 2 2
and
∥x̃ − k̃1 − x̃ + k̃2 ∥ = ∥k̃1 − k̃2 ∥ > 0,
because k1 ̸= k2 . Thus, as X is a u.c. B-space, we should have
x̃ − k̃1 x̃ − k̃2
∥ + ∥≤1−δ
2 2
for some δ > 0. Obviously, this contradicts (3.26) and, consequently, it
shows the uniqueness of the minimizer. Subsequently, for every x ∈ X, we
denote by PK x ∈ K the unique minimizer of Φ in K.
To conclude the proof of the theorem, it remains to prove that PK is
continuous. Let {xn }n≥1 be a sequence in X such that
lim ∥xn − x∥ = 0
n→∞
for some x ∈ X. Then, as the map

dist (·, K) : X −→ [0, ∞)
is continuous, we have that
lim ∥xn − PK xn ∥ = lim dist (xn , K)
n→∞ n→∞
(3.27)
= dist (x, K) = ∥x − PK x∥.
As a by-product, the sequence {PK xn }n≥1 is bounded, because, for suffi-
ciently large n ≥ 1,
∥PK xn ∥ ≤ ∥xn ∥ + ∥x − PK x∥ + 1
and. moreover, {xn }n≥1 is bounded.
By Theorem 3.6, X is reflexive. Thus, owing to Theorem 3.8, there
exists w ∈ X and a subsequence of {xn }n≥1 , say {xnm }m≥1 , such that
lim PK xnm = w in σ(X, X ′ ).
m→∞
As, due to Theorem 3.7, K is closed for the weak topology σ(X, X ′ ), nec-
essarily w ∈ K and
lim (xnm − PK xnm ) = x − w in σ(X, X ′ ). (3.28)
m→∞
Consequently, we obtain from Corollary 3.1 that
∥x − w∥ ≤ lim inf ∥xnm − PK xnm ∥, (3.29)
m→∞
because ∥ · ∥ : X → [0, ∞) is l.s.c. and convex. Thus, we find from (3.27)
and (3.29) that
∥x − w∥ ≤ ∥x − PK x∥
and, hence, w = PK x, by the definition of PK . Therefore, (3.28) does
actually provide us with
lim (xnm − PK xnm ) = x − PK x in σ(X, X ′ ). (3.30)
m→∞
Finally, by (3.27) and (3.30), Proposition 3.3 implies that
lim (xnm − PK xnm ) = x − PK x in X,
m→∞
and, consequently,
lim PK xnm = PK x in X.
m→∞
As this scheme can be repeated along any subsequence, we find that
lim PK xn = PK x in X.
n→∞
3.7.3 The projection on a closed linear subspace

Let N be a closed subspace of a u.c. B-space X. According to Theorem
3.2, when X is a Hilbert space, PN is a linear operator and hence N admits
a topological complement in X. Consequently, a natural question arises.
Should PN be a linear operator for any closed subspace N of a u.c. B-space
X? The next theorem by J. Lindenstrauss and L. Tzafriri [127] shows that,
in general, PN is non-linear. It actually shows that all the projections PN
are linear if and only if X is Hilbertizable.
Theorem 3.11 (of J. Lindenstrauss and L. Tzafriri). A B-space X

is Hilbertizable, i.e., it admits an equivalent norm associated to an inner
product, if and only if every closed subspace possesses a topological comple-
ment.
Nevertheless, when X is an arbitrary u.c. B-space and N is a closed

subspace of X, the projection PN constructed by Theorem 3.10 satisfies the
properties collected in the next result.
Proposition 3.4. Suppose X is a u.c. B-space and N is a closed subspace

of X. Then,
PN (λx) = λPN x (3.31)
for all λ ∈ R and x ∈ X. Moreover,
PN (x − PN x) = 0 for all x ∈ X, (3.32)
and hence, PN is linear if and only if, for every x, y ∈ X,
PN x = PN y = 0 implies PN (x + y) = 0. (3.33)
Proof. According to Theorem 3.10,

∥x − PN x∥ < ∥x − n∥ (3.34)
for all x ∈ X and n ∈ N \ {PN x}. Pick λ > 0. Then, (3.34) implies that
∥λx − λPN x∥ < ∥λx − n∥ ∀ x ∈ X, n ∈ N \ {PN x}.
Thus,
PN (λx) = λPN x (3.35)
for all x ∈ X and λ > 0. Moreover, (3.34) also shows that
∥ − x + PN x∥ ≤ ∥ − x − n∥ ∀ x ∈ X, n ∈ N.
Hence,
PN (−x) = −PN x (3.36)
for all x ∈ X. Thanks to (3.35) and (3.36), we find that, for every λ < 0
and x ∈ X,
PN (λx) = PN (−λ(−x)) = −λPN (−x) = λPN x.
Moreover, PN 0 = 0, by definition. Consequently, (3.31) holds.
Now, we will prove (3.32). By definition, we have that
∥x − PN x − PN (x − PN x)∥ ≤ ∥x − PN x − n∥ (3.37)
for all x ∈ X and n ∈ N . Similarly,
∥x − PN x∥ ≤ ∥x − ñ∥ (3.38)
for all x ∈ X and ñ ∈ N . Thus, particularizing (3.37) at n = 0 and (3.38)
at
ñ = PN x + PN (x − PN x) ∈ N,
it becomes apparent that
∥x − PN x − PN (x − PN x)∥ ≤ ∥x − PN x∥
≤ ∥x − PN x − PN (x − PN x)∥
and, therefore,
∥x − PN x − PN (x − PN x)∥ = ∥x − PN x∥,
which implies (3.32), by the uniqueness of PN (x − PN x).
Obviously, (3.33) holds if PN is linear. To show the converse, suppose
(3.33) holds and let x, y ∈ X. Then, owing to (3.32),
0 = PN (x − PN x) = PN (y − PN y) = PN (x + y − PN (x + y)).
Thus, it follows from (3.33) that
0 = PN (x + y − PN x − PN y) = PN (x + y − PN (x + y))
and hence, by (3.31) and (3.33), we find that
PN (x + y − PN x − PN y − x − y + PN (x + y)) = 0.
Equivalently,
PN (PN (x + y) − PN x − PN y) = 0,
and, therefore, since
PN (x + y) − PN x − PN y ∈ N,
we obtain that
PN (x + y) = PN x + PN y.
Combining this identity with (3.31) shows that PN is linear and ends the
proof.
3.7.4 The projection on a closed hyperplane

The next result shows that PN is indeed linear when N is a closed hyper-
plane.
Theorem 3.12. Let X be a u.c. B-space and N ⊂ X a closed hyperplane.

Then, PN ∈ L(X, N ).
The proof of this result is based on the following
Lemma 3.1. Let X be a u.c. B-space and N ⊂ X a closed proper subspace.

Then, there exists z ∈ X such that
∥z∥ = 1 and PN z = 0. (3.39)
Proof. Let x ∈ X \ N . Then,

x ̸= PN x ∈ N
and hence,
x − PN x
z :=
∥x − PN x∥
is well defined and it satisfies ∥z∥ = 1. Moreover, according to (3.31) and
(3.32), we have that
1
PN z = P (x − PN x) = 0.
∥x − PN x∥ N
Proof. [Proof of Theorem 3.12] By Theorem 3.10 and Proposition 3.4, it
remains to prove that PN satisfies (3.33). Let z ∈ X satisfying (3.39). As
N is a closed hyperplane,
X = N ⊕ span[z]. (3.40)
Let x, y ∈ X be such that
PN x = PN y = 0.
According to (3.40), there exist λ, µ ∈ R and nx , ny ∈ N (unique) such that
x = λz + nx , y = µz + ny .
As PN x = 0, we have that
∥x∥ = ∥λz + nx ∥ ≤ ∥λz + n∥
for all n ∈ N . In particular,
∥λz + nx ∥ ≤ ∥λz∥.
On the other hand, by (3.31) and (3.39),

PN (λz) = λPN z = 0.
Thus,
∥λz∥ = ∥λz + nx ∥
and, consequently, nx = 0, by the definition of PN . As this is valid for all
x ∈ X, we also have that ny = 0. So,
x + y = (λ + µ)z
and, therefore, (3.31) implies that
PN (x + y) = PN ((λ + µ)z) = (λ + µ)PN z = 0.
This shows (3.33) and concludes the proof.
The abstract concept of Hilbert space goes back to J. Von Neumann [171],
where an axiomatic definition of Hilbert space was given for the first time in
the context of separable spaces. The original Hilbert space of D. Hilbert [99]
was nothing more than ℓ2 ! In 1935, P. Jordan and J. Von Neumann [109]
proved that the parallelogram identity (3.2) characterizes all Hilbertian
norms.
Theorem 3.3 goes back to F. Riesz [191], where, rather remarkably, the
separability requirement of J. Von Neumann [171] and predecessors was
overcome. In his original paper, published in 1934, F. Riesz stressed that
the theory of Hilbert spaces should be founded upon his representation
theorem.
Twenty years later, in 1954, P. D. Lax and A. N. Milgram [122] for-
mulated the variant of the Riesz’ representation theorem established by
Theorem 3.5. As it will become apparent in Chapter 4, Theorem 3.5 is
an extremely useful device to get the existence of weak solutions in wide
classes of linear boundary value problems of elliptic type.
Naturally, the importance of the concepts introduced in Definition 3.1
was revealed by P. D. Lax and A. N. Milgram [122] themselves, though,
apparently, the word coercive was coined later (see p. 92 of K. Yosida
[227], whose first edition goes back to 1964, and p. 298 of L. C. Evans
[60]).
Ten years later, in 1964, G. Stampacchia [212] generalized Theorem
3.5 up to obtain Theorem 3.4, which has shown to be a milestone for the
development of the calculus of variations (see D. Kinderlehrer and G. Stam-

pacchia [113]). The proof of Theorem 3.4 given here goes back to J. L.
Lions and G. Stampacchia [131]. It substantially differs from the original
one given by G. Stampacchia [212].
The existence of the orthogonal projection on a closed subspace is a
very classical result attributable to E. Schmidt [198], which allowed him
to tidy up and generalize some abstract previous results of D. Hilbert on
linear systems.
The existence of the projection operator on a closed and convex subset
of a Hilbert space goes back, at least, to Th. Motzkin [168], who, in 1935,
characterized the convexity of a closed set K in RN by establishing that
K is convex if and only if to each point in RN there corresponds a unique
nearest point in K. His demonstrations adapt mutatis mutandis to cover
the general infinite dimensional case. Among weakly closed sets, convex
sets are the unique ones admitting projections (see Theorem 4.8 of F. A.
Valentine [221]). These existence results can be applied to analyze many
problems of the calculus of variations, and, in particular, to solve a number
of obstacle problems (see p. 4 of D. Kinderlehrer and G. Stampacchia [113]).
Theorem 3.1 was the key tool of J. L. Lions and G. Stampacchia [131]
to prove Theorem 3.4. The variational characterization of the projection
operator established by (3.3) is attributable to them.
Theorems 3.10, 3.12 and Proposition 3.4 might be new. Anyway, they
should be weighted against other related results on geometry of Banach
spaces, as, e.g., B. Beauzamy [24], J. Diestel [53], J. Lindenstrauss and L.
Tzafriri [128], and L. Schwartz [200]. The fact that the projection on a
general closed subspace of a Banach space might not be linear had been
observed independently by T. Kato [112] (see Lemma 2.3 therein).
Lemma 3.1 is a sharp version, for uniformly convex Banach spaces, of a
classical result of F. Riesz [190] about the existence of “nearly orthogonal”
projections in general Banach spaces (e.g., see Section III.2 of K. Yosida
[227], Lemma 5.4 of D. Gilbarg and N. Trudinger [79], and Lemma VI.1
of H. Brézis [29]) with a number of important applications in functional
analysis. Indeed, according to (3.39),
dixt (z, N ) = ∥z − PN z∥ = ∥z∥ = 1
and, therefore, for any ϵ ∈ (0, 1), there exists an xϵ ∈ X such that
∥xϵ ∥ = 1 and dist (xϵ , N ) ≥ 1 − ϵ.
This chapter has been strongly inspired by Chapter VI of H. Brézis [29],
though most of its contents, are, certainly, folklore and, consequently, very
well documented in many textbooks.
The author expresses his deepest gratitude to Professor F. Bombal, at

Complutense University, for a very fruitful telephone discussion about the
materials of Section 3.7.3.
Chapter 4
Existence of weak solutions
This chapter considers a second order uniformly elliptic differential operator

of the form
L := − div (A∇ · ) + ⟨b, ∇ · ⟩ + c (4.1)
in a bounded domain Ω of RN , N ≥ 1, where ‘div’ stands for the divergence
operator
∑N
∂uj
div (u1 , . . . , uN ) = ,
j=1
∂xj
⟨·, ·⟩ is the Euclidean inner product of RN , and

 sym
 A = (aij )1≤i,j≤N ∈ MN (W
1,∞
(Ω)),
(4.2)

 b = (b , ..., b ) ∈ (L∞ (Ω))N ,
1 N c ∈ L∞ (Ω).
For a given Banach space X, we are denoting by Msym N (X) the space of
the symmetric square matrices of order N with entries in X, and W 1,∞ (Ω)
stands for the Sobolev space of all bounded and measurable functions in Ω
with weak derivatives in L∞ (Ω) (see Section 4.1 for its precise definition).
According to (4.2), the differential operator (4.1) fits within the abstract
setting of Chapters 1 and 2. Actually, it can be expressed in the form
∑N
∂2 ∑
N
∂
L=− aij (x) + b̃j (x) + c(x), x ∈ Ω,
i,j=1
where
∑
N
∂aij
b̃j := bj − ∈ L∞ (Ω), 1 ≤ j ≤ N.
i=1
∂xi
Therefore, L satisfies Assumption H2 of Chapter 2.
Throughout this chapter, we are imposing the following general assump-
tions:
91
B1 Ω is a bounded domain of RN , N ≥ 1, whose boundary consists of two

disjoint open and closed subsets, denoted by Γ0 and Γ1 , of class C 1
∂Ω := Γ0 ∪ Γ1 .
Necessarily, Γ0 and Γ1 must possess finitely many components. Either
Γ0 , or Γ1 , might be empty.
By Theorem 1.9, Ω satisfies the uniform interior sphere property in the
strong sense if, in addition, it is of class C 2 . In particular, Assumption
H1 of Chapter 2 holds if Ω is of class C 2 .
B2 β ∈ C(Γ1 ) is a continuous real function, n denotes the outward unit
normal vector field of Ω, and
ν := An
is the conormal vector field, i.e.,
∂u
= ⟨∇u, An⟩ = ⟨A∇u, n⟩ for all u ∈ C 1 (Γ1 ).
∂ν
The vector field ν satisfies Assumption H3 of Chapter 2. Indeed,
⟨ν, n⟩ = ⟨An, n⟩ ≥ µ|n|2 = µ > 0,
where µ is the ellipticity constant of L in Ω. Therefore, (1.25) holds.
Summarizing, under conditions (4.2) and B1, B2, Assumptions H2 and H3

of Chapter 2 hold, as well as Assumption H1 if, in addition, Ω is of class
C2.
Under Assumption B2, we denote by
B : C(Γ0 ) ⊗ C 1 (Γ1 ) → C(∂Ω)
the boundary operator
{
ψ on Γ0 ,
Bψ := ∂ψ ψ ∈ C(Γ0 ) ⊗ C 1 (Γ1 ). (4.3)
∂ν + βψ on Γ1 ,
Essentially, this chapter introduces the concept of weak solution for the
linear boundary value problem
{
(L + ω)u = f in Ω,
(4.4)
Bu = 0 on ∂Ω,
and it shows the existence of an ω0 ∈ R such that, for every ω > ω0
and f ∈ L2 (Ω), (4.4) has a unique weak solution. Precisely, the distribu-
tion of this chapter is as follows. Section 4.1 begins by introducing the
Existence of weak solutions 93
concepts of weak derivatives, Sobolev spaces, and Hölder spaces of contin-

uous functions. Then, it collects the Sobolev imbedding theorems and the
compactness result of F. Rellich and V. I. Kondrachov. As the contents
of Section 4.1 can be thought of as a series of well-known properties of
Lp -spaces, we are not giving the proofs of these results here. Section 4.2
constructs the trace operators and studies their main properties. Section
4.3 introduces the concept of weak solution for (4.4) and studies some of its
properties through the associated bilinear form, denoted by a(·, ·). Section
4.4 shows the continuity of a, and Section 4.5 shows the coercivity of a
for sufficiently large ω when β ≥ 0. The existence of a weak solution in
this case is derived from the theorem of P. D. Lax and A. N. Milgram (see
Theorem 3.5). Finally, Section 4.6 shows the existence of a weak solution
for a general β ∈ C(Γ1 ) and sufficiently large ω. This result might go back
to [147].
Throughout the rest of this book, we will use the concept of classical
solution introduced by the next definition.
Definition 4.1 (Classical solution). A function u ∈ C 1 (Ω̄) is said to be

a classical solution of (4.4) if it is twice classically differentiable almost
everywhere in Ω,
(L + ω)u = f almost everywhere in Ω,
and
Bu = 0 on ∂Ω.
4.1 Preliminaries. Sobolev spaces
4.1.1 Test functions

Extending the concept already introduced in the beginning of Section 2.6,
we will subsequently denote by C0∞ (Ω) the set of functions ϕ : Ω̄ → R of
class C ∞ with compact support
supp ϕ := ϕ−1 (R \ {0}) ⊂ Ω.

As we are assuming that Ω is bounded, it is apparent that
C0∞ (Ω) = {ϕ ∈ C ∞ (Ω̄) : supp ϕ ⊂ Ω}.
More generally, for every
Γ ∈ {Γ0 , Γ1 , ∂Ω},
we will denote by CΓ∞ (Ω) the set

CΓ∞ (Ω) := {ϕ ∈ C ∞ (Ω̄) : supp ϕ ⊂ Ω̄ \ Γ}.
Consequently, a function ϕ ∈ C ∞ (Ω̄) belongs to CΓ∞ (Ω) if and only if it
vanishes on some open neighborhood of Γ. According to these notations,
∞
C∂Ω (Ω) = C0∞ (Ω).
All those functions will be referred to as test functions.
4.1.2 Weak derivatives. Sobolev spaces

Besides L∞ (Ω), which has already been introduced in the beginning of
Chapter 1, we will subsequently consider, for every p ∈ [1, ∞), the Banach
space Lp (Ω) of all measurable real functions u : Ω → R, as discussed by
Lebesgue, such that
∫
|u|p < ∞.
Ω
It is folklore that Lp (Ω) is a Banach space with norm

(∫ ) p1
∥u∥p = ∥u∥Lp (Ω) := |u|p , u ∈ Lp (Ω),
Ω
which is usually referred to as the Lp -norm. Actually, L2 (Ω) is a Hilbert

space, as ∥ · ∥2 is the norm induced by the scalar product
∫
⟨u, v⟩L2 := u(x)v(x) dx, u, v ∈ L2 (Ω).
Ω
Similarly, for every p ∈ [1, ∞), we will denote by Lploc (Ω) the set of all
measurable real functions u : Ω → R such that
∫
|u|p < ∞ for all compact subset K ⊂ Ω.
K
A fundamental property of these spaces establishes that, for every p, q ∈

[1, ∞] with
1 1
+ = 1,
p q
and u ∈ Lp (Ω), v ∈ Lq (Ω), the product function uv belongs to L1 (Ω) and
∥uv∥1 ≤ ∥u∥p ∥v∥q . (4.5)
This estimate is usually referred to as the Hölder inequality.
Subsequently, for any multi-index
α = (α1 , ..., αN ) ∈ NN
we denote
∑
N
∂ |α|
|α| := αj , Dα := .
j=1
∂xα
1
1
· · · ∂xα
N
N
Integrating by parts, it becomes apparent that

∫ ∫
uDα ϕ = (−1)|α| ϕDα u (4.6)
Ω Ω
for all u ∈ C |α| (Ω) and ϕ ∈ C0∞ (Ω). Actually, the left-hand side of (4.6)
makes sense even if u ∈ L1loc (Ω), as Dα ϕ has compact support in Ω. Con-
sequently, the concept of weak derivatives, or derivatives in the weak sense,
introduced by the next definition is rather natural.
Definition 4.2 (Weak derivatives). Let u, v ∈ L1loc (Ω) and α ∈ NN .

Then, it is said that
v = Dα u in the weak sense
or, equivalently, that v is the weak derivative of order α of u, if

∫ ∫
|α|
α
uD ϕ = (−1) ϕv for all ϕ ∈ C0∞ (Ω).
Ω Ω
This concept is consistent because the weak derivative Dα u is uniquely

determined if it exists. Indeed, if v, w ∈ L1loc (Ω) satisfy
∫ ∫
vϕ = wϕ for all ϕ ∈ C0∞ (Ω),
Ω Ω
then, v = w almost everywhere in Ω.

For every p ∈ [1, ∞] and k ∈ N, the Sobolev space W k,p (Ω) is defined as
the set of functions u ∈ L1loc (Ω) such that the weak derivative Dα u exists
for every α ∈ NN with |α| ≤ k and Dα u ∈ Lp (Ω), endowed with the norm
  p1



 ∑

 ∥Dα u∥Lp (Ω) 
p
if p < ∞,


0≤|α|≤k
∥u∥W k,p (Ω) :=



 ∑

 ∥Dα u∥L∞ (Ω) if p = ∞,


0≤|α|≤k
for all u ∈ W k,p (Ω). Equivalently, one might define these norms through
∑
∥u∥W k,p (Ω) := ∥Dα u∥Lp (Ω)
0≤|α|≤k
if p < ∞. It is folklore that all these normed spaces are Banach spaces.
Actually, they are Hilbert spaces if p = 2 with inner product
∑ ∫
⟨u , v⟩W k,2 (Ω) := Dα u Dα v, u, v ∈ W k,2 (Ω),
0≤|α|≤k Ω
because L2 (Ω) is a Hilbert space. In particular,

∫ ∫
⟨u , v⟩W 1,2 (Ω) := ⟨∇u, ∇v⟩ + uv, u, v ∈ W 1,2 (Ω).
Ω Ω
Applying (4.5) with p = q = 2, it becomes apparent that all these inner

products are well defined.
As usual, these features might induce us to adopt the notation
H k (Ω) := W k,2 (Ω), k ∈ N. (4.7)
According to it, H 0 (Ω) = L2 (Ω). More generally, W 0,p (Ω) = Lp (Ω) for all
k,p
p ∈ [1, ∞]. Also, for every k ∈ N and p ∈ [1, ∞], we will denote by Wloc (Ω)
the set
k,p
∩
Wloc (Ω) := W k,p (D).
D open
D̄⊂Ω
Note that D̄ is compact, since Ω is bounded. Naturally, the weak derivatives

satisfy the algebraic properties collected in the next result.
Lemma 4.1. Let k ∈ N, p ∈ [1, ∞], and u, v ∈ W k,p (Ω). Then,
i) For every α, β ∈ NN with |α| + |β| ≤ k, Dα u ∈ W k−|α|,p (Ω) and
Dβ (Dα u) = Dα (Dβ u) = Dα+β u.
ii) For every A, B ∈ R and α ∈ NN with |α| ≤ k, Au + Bv ∈ W k,p (Ω) and
Dα (Au + Bv) = ADα u + BDα v.
iii) For every ζ ∈ C0∞ (Ω) and α ∈ NN with |α| ≤ k, ζu ∈ W k,p (Ω) and
∑ (α) ( )
α |α|!
α β α−β
D (ζu) = D ζD u, := .
β β |β|! |α − β|!
β≤α
Based on these properties and the completeness of Lp (Ω), it is easy to

see that the spaces W k,p (Ω), k ∈ N, p ∈ [1, ∞], are indeed Banach spaces.
The next result establishes that, for every p ∈ [1, ∞) and k ∈ N, W k,p (Ω) is
the completion of C ∞ (Ω̄) with respect to the W k,p (Ω)-norm. It should not
be forgotten that, according to Assumption B1, throughout this chapter we
are assuming that Ω is a connected bounded open set of class C 1 .
Theorem 4.1 (Global approximation by smooth functions).

Let k ∈ N, p ∈ [1, ∞), and u ∈ W k,p (Ω). Then, there exists a sequence
ϕn ∈ C ∞ (Ω̄), n ≥ 1, such that
lim ∥ϕn − u∥W k,p (Ω) = 0.

n→∞
When u ∈ W k,p (Ω) ∩ C(Ω̄), then {ϕn }n≥1 can be chosen so that
( )
lim ∥ϕn − u∥W k,p (Ω) + ∥ϕn − u∥C(Ω̄) = 0.
n→∞
4.1.3 Hölder spaces of continuous functions

The following definition introduces some important basic concepts.
Definition 4.3. Let u ∈ C(Ω) and ν ∈ (0, 1]. Then,
i) u is said to be (globally) Lipschitz continuous in Ω if there exists a

constant C ≥ 0 such that
|u(x) − u(y)| ≤ C|x − y| for all x, y ∈ Ω.
ii) u is said to be locally Lipschitz continuous in Ω if for every x ∈ Ω

there exists R > 0 such that BR (x) ⊂ Ω and u is (globally) Lipschitz
continuous in BR (x).
iii) u is said to be (globally) Hölder continuous in Ω, with exponent ν, if
there exists a constant C ≥ 0 such that
|u(x) − u(y)| ≤ C|x − y|ν for all x, y ∈ Ω.
iv) u is said to be locally Hölder continuous in Ω, with exponent ν, if for

every x ∈ Ω there exists R > 0 such that BR (x) ⊂ Ω and u is (globally)
Hölder continuous in BR (x) with exponent ν.
Obviously, u is Hölder continuous with exponent ν = 1 if and only if it

is Lipschitz continuous.
Throughout the rest of this book, for a given ν ∈ (0, 1], we denote
by C 0,ν (Ω̄) the Banach space of all functions u ∈ C(Ω̄) that are Hölder
continuous in Ω with exponent ν, endowed with the Hölder-norm
|u(x) − u(y)|
∥u∥C0,ν (Ω̄) := ∥u∥C(Ω̄) + sup , u ∈ C 0,ν (Ω̄).
x,y∈Ω̄ |x − y|ν
x̸=y
More generally, for every k ∈ N and ν ∈ (0, 1], we denote by C k,ν (Ω̄) the
Banach subspace of C k (Ω̄) consisting of all functions u ∈ C k (Ω̄) such that
Dα u ∈ C 0,ν (Ω̄) for all α ∈ NN with |α| = k,
equipped with the norm
∑ |Dα u(x)−Dα u(y)|
∥u∥Ck,ν (Ω̄) := ∥u∥Ck (Ω̄) + sup , u ∈ C k,ν (Ω̄).
x,y∈Ω̄ |x − y| ν
|α|=k
x̸=y
In other words, the space C k,ν (Ω̄) consists of all functions u that are k times
continuously differentiable and whose k th -partial derivatives are Hölder con-
tinuous of exponent ν.
Eventually, for every ν ∈ (0, 1], we will also consider the set of functions
k,ν −
C (Ω̄) defined through
− ∩
C k,ν (Ω̄) = C k,θ (Ω̄).
0<θ<ν
Also, for every k ∈ N and ν ∈ (0, 1], we will consider the set of functions
C k,ν (Ω) defined by
∩
C k,ν (Ω) := C k,ν (D̄).
D open
D̄⊂Ω
The W (Ω) spaces are analogous in a certain sense to the C k,ν (Ω̄) spaces.
k,p
In the W k,p (Ω) spaces, classical differentiability is replaced by weak dif-

ferentiability and Hölder continuity by Lp -integrability. The results in the
next sections make precise this idea.
4.1.4 Sobolev’s imbeddings

In the next result we collect the main Sobolev imbedding theorems. Sub-
sequently, for every real number r ∈ R, r ≥ 0, we denote by [r] the entire
part of r, i.e., the unique integer [r] ∈ N such that
[r] ≤ r < [r] + 1.
Theorem 4.2 (Sobolev’s imbeddings). Let k ∈ N, k ≥ 1, and p ∈

(1, ∞).
i) If kp < N , then
Np
N −kp
W k,p (Ω) ⊂ L (Ω)
and the associated injection is continuous, i.e., there exists a constant
C := C(k, p, N, Ω) such that
∥u∥ Np
≤ C ∥u∥W k,p (Ω)
N −kp
L (Ω)
for all u ∈ W k,p (Ω).

ii) If kp > N , then
[ ]
k− N −1,ν
W k,p (Ω) ⊂ C
p
(Ω̄), (4.8)
where
[ ] [ ]

 p +1− ∈ (0, 1),
N N N N
 p if p < p,
ν= [ ] (4.9)


 1− , if N
= N
p p.
Moreover, the associated injection is continuous, i.e., there exists a

constant C := C(k, p, N, Ω) such that
∥u∥ ≤ C∥u∥W k,p (Ω)
C
p[ ]
k− N −1,ν
(Ω̄)
for all u ∈ W k,p (Ω). When ν = 1− , this estimate should be understood

in the sense that, for every θ ∈ (0, 1), there exists a constant C :=
C(k, p, N, Ω, θ) such that
∥u∥ ≤ C∥u∥W k,p (Ω)
C
p[ ]
k− N −1,θ
(Ω̄)
for all u ∈ W k,p (Ω).
In the special –but relevant– case when k = 1, Theorem 4.2 provides us

with the next result.
Corollary 4.1. Suppose 1 < p < ∞. Then,

 Np

 N −p
L (Ω), if p < N,
W 1,p (Ω) ⊂ (4.10)


 0,1− Np
C (Ω̄), if p > N,
with continuous embeddings.
[ ]
N
Note that p = 0 if p > N and hence, (4.8) becomes
0,ν N
W 1,p (Ω) ⊂ C (Ω̄) with ν =1− ∈ (0, 1).
p
According to (4.10), we also have that
∩ −
W 1,∞ (Ω) ⊂ C 0,θ (Ω̄) = C 0,1 (Ω̄).
0<θ<1
Actually, the following characterization holds.
Theorem 4.3. u ∈ W k,∞ (Ω) if and only if u ∈ C k−1,1 (Ω̄), i.e., if Dα u is

Lipschitz continuous in Ω for every α ∈ NN with |α| = k −1. Consequently,
k,∞
u ∈ Wloc (Ω) if and only if Dα u is locally Lipschitz continuous in Ω for all
these α’s. In other words,
k,∞
W k,∞ (Ω) = C k−1,1 (Ω̄) and Wloc (Ω) = C k−1,1 (Ω).
In particular, u ∈ W 1,∞ (Ω) if and only if u is Lipschitz continuous in

1,∞
Ω, and u ∈ Wloc (Ω) if and only if u is locally Lipschitz continuous in Ω.
According to Corollary 4.1, u ∈ C 0,ν (Ω̄) with ν = 1 − N/p ∈ (0, 1) if
u ∈ W 1,p (Ω) with p > N , but the converse fails to be true, as a function
u ∈ C 0,ν (Ω̄) with ν ∈ (0, 1) does not necessarily belong to
∪
W 1,p (Ω).
p>N
The following regularity result clarifies this fact.
Theorem 4.4. Suppose

∪ 1,p
u∈ Wloc (Ω).
N <p≤∞
Then, u is differentiable almost everywhere (a.e.) in Ω, and its classical

gradient equals its weak gradient a.e. in Ω. Consequently, if
∪ 2,p
u∈ Wloc (Ω),
N <p≤∞
then u ∈ C 1 (Ω) and it is twice classically differentiable a.e. in Ω. Moreover,

the classical derivative Dα u equals the corresponding weak derivative a.e.
in Ω for all multi-index α ∈ NN with |α| ≤ 2.
In particular, by Theorem 4.3, any locally Lipschitz continuous function

must be differentiable almost everywhere. Note that the inclusion
∪ 2,p
Wloc (Ω) ⊂ C 1 (Ω)
N <p≤∞
2,p
follows from (4.8) and Theorem 4.3. Moreover, u ∈ Wloc (Ω) implies
∂u 1,p
∈ Wloc (Ω), 1 ≤ i ≤ N,
∂xi
and hence the second assertion of Theorem 4.4 is a direct consequence from
the first one.
Letting p ↑ N in (4.10) one might expect the validity of the injection
W 1,N
(Ω) ⊂ L∞ (Ω), but this fails to be true if N > 1. Indeed, if N ≥ 2
and Ω = B1 (0), then the function
( )
1
u(x) := Log Log 1 + , x ∈ Ω,
|x|
belongs to W 1,N (Ω) \ L∞ (Ω). Obviously, the limiting inclusion obtained
by letting p ↓ N in (4.10) cannot be true either, because C(Ω̄) ⊂ L∞ (Ω)
and we have just seen that W 1,N (Ω) * L∞ (Ω).
4.1.5 Compact imbeddings

Given two Banach spaces X, Y and a linear continuous operator T ∈
L(X, Y ), it is said that T is compact if T (A) is a pre-compact subset of
Y (i.e., the closure T (A) is a compact subset of Y ) for every bounded
subset A of X. In such case, it will simply said that
T ∈ K(X, Y ).
If X ⊂ Y and Ā is compact in Y for all bounded set A ⊂ X, we will simply
write that
X ,→ Y.
The main compactness result concerning Sobolev spaces is the next one.
Theorem 4.5 (of F. Rellich and V. I. Kondrachov). Suppose k ∈

N, k ≥ 1, and 1 ≤ p < ∞.
i) If kp < N , then, for every
Np
1≤q< ,
N − kp
the imbedding of W k,p (Ω) into Lq (Ω) is a compact operator, i.e.,
W k,p (Ω) ,→ Lq (Ω).
ii) If kp > N and ν is given through (4.9), then,

[ ]
k− N −1,β
W k,p (Ω) ,→ C
p
(Ω̄)
for all β < ν.
Moreover,
W 1,p (Ω) ,→ Lp (Ω) for all p ∈ [1, ∞]. (4.11)
Certainly, (4.11) is reminiscent from the classical theorem of G. Ascoli

[18] and C. Arzela [17].
4.2 Trace operators
Whenever u ∈ C(Ω̄) the trace restriction of u to the boundary ∂Ω, denoted

by u|∂Ω , is well defined in the classical sense. But a function u ∈ W 1,p (Ω)
does not need to be continuous in Ω. Actually, it is only defined almost
everywhere in Ω. Moreover, according to Assumption B1, ∂Ω is an (N − 1)-
dimensional surface of class C 1 and, hence, it has zero N -dimensional
Lebesgue measure. Therefore, it does not make sense to define the re-
striction of u ∈ W 1,p (Ω) to any portion of ∂Ω. Nevertheless, the next trace
theorem holds.
Theorem 4.6 (trace theorem). Let p ∈ [1, ∞) and Γ ∈ {Γ0 , Γ1 , ∂Ω}.

Then, there exists a unique linear continuous operator
( )
TΓ ∈ L W 1,p (Ω), Lp (Γ)
such that
TΓ u = u|Γ for all u ∈ W 1,p (Ω) ∩ C(Ω̄);
TΓ will be called the trace operator of W 1,p (Ω) on Γ, and, for every
u ∈ W 1,p (Ω), TΓ u ∈ Lp (Γ) will be referred to as the trace of u on Γ.
By the uniqueness, we also have that
T∂Ω = TΓ0 ⊗ TΓ1 .
Proof. First, we will show that there exists a constant C > 0 such that
∥u∥Lp (Γ) ≤ C∥u∥W 1,p (Ω) (4.12)
for all u ∈ C 1 (Ω̄). Then, the theorem will follow through a density argu-
ment. As Γ is a compact subset of Ω̄, to prove (4.12) it suffices to show
that for each x0 ∈ Γ there exist an open neighborhood Γ̃ of x0 in Γ and a

constant C = C(x0 ) > 0 such that
∫ (∫ ∫ )
|u|p dS ≤ C |u|p + |∇u|p (4.13)
Γ̃ Ω Ω
for all u ∈ C (Ω̄). Indeed, fix x0 ∈ Γ and suppose, in addition, that Γ is

1
flat in a neighborhood of x0 ∈ Γ, in the sense that it lies on the plane

Π := {x ∈ RN : xN = 0}.
As ∂Ω is of class C 1 , this can be reached after a local change of coordinates
of class C 1 , but the details of the proof in the general case are postponed.
Then, there exists r = r(x0 ) > 0 such that
Br+ := Br (x0 ) ∩ {x ∈ RN : xN > 0} ⊂ Ω,
Br− := Br (x0 ) ∩ {x ∈ RN : xN < 0} ⊂ RN \ Ω̄.
Necessarily, Br (x0 ) ∩ Γ ⊂ Π (see Figure 4.1).
xN
Ω Γ
0
Γ1
r r/2 x0 Π
Fig. 4.1 The case when Γ (= Γ0 ) is flat around x0 ∈ Γ
Now, consider Br/2 (x0 ), the concentric open ball of radius r/2, and
select ζ ∈ C0∞ (Br (x0 )) such that ζ ≥ 0 in Br (x0 ) and ζ = 1 in Br/2 (x0 ).
Obviously, ζ can be chosen to be radially symmetric. Figure 4.2 represents
one of those ζ’s. Subsequently, for any δ ∈ (0, r], we will denote by
Γδ := Γ ∩ Bδ (x0 )
x0−r x0 −r/2 x0 x0+r/2 x0+r
Fig. 4.2 A slice of an admissible ζ
the portion of Γ within Bδ (x0 ), and, using the notations already introduced
in Section 1.8, we set
x[N ] := (x1 , ..., xN −1 ) ∈ RN −1 ≡ Π.
Then, it is apparent that
∫ ∫ ∫
|u|p dS ≤ ζ|u|p dS = ζ|u|p dS.
Γr/2 Γr ∂Br+
Thus, by the divergence theorem, we find that

∫ ∫
∂
|u| dS ≤ −
p
(ζ|u|p ) dx
+ ∂xN
Γr/2 Br
∫ ∫
∂ζ ∂u
=− |u|p dx − ζp|u|p−1 sign u dx
Br+ ∂xN Br+ ∂xN
because
n = (0, ..., 0, −1)
at the points of ∂Br+ where ζ|u|p does not vanish. Consequently, there exist
C1 > 0 and C2 > 0, independent of u ∈ C 1 (Ω̄), such that
∫ ∫ ∫
|u|p dS ≤ C1 |u|p dx + C2 |u|p−1 |∇u| dx. (4.14)
Γr/2 Br+ Br+
On the other hand, according to a classical inequality of W. H. Young [228],

for every a > 0, b > 0, and 1 < p, q < ∞, with p−1 + q −1 = 1, one has that
ap bq
ab ≤ + ,
p q
and hence,
1 p−1 p
|∇u||u|p−1 ≤ |∇u|p + |u| in Ω,
p p
because q = p/(p − 1). Thus, setting
{ }
p−1 C2
C := max C1 + C2 , ,
p p
we obtain from (4.14) that

∫ (∫ ∫ )
|u|p dS ≤ C |u|p dx + |∇u|p dx . (4.15)
Γr/2 Br+ Br+
As C > 0 is independent of u, the estimate (4.13) is an easy consequence

from (4.15).
When Γ is not flat around x0 , we can proceed as follows. By Definition
1.5, there exist R := R(x0 ) > 0, an open neighborhood D of 0 ∈ RN , and
a bijection Φ : BR (x0 ) → D such that Φ(x0 ) = 0, DΦ(x0 ) ∈ Iso(RN ),
Φ ∈ C 1 (BR (x0 ); D), Φ−1 ∈ C 1 (D; BR (x0 )), and
Φ(BR (x0 ) ∩ Ω) = {x ∈ D : xN > 0},
Φ(BR (x0 ) ∩ ∂Ω) = {x ∈ D : xN = 0}.
In other words, Φ is a local diffeomorphism straightening ∂Ω at x0 into the
hyperplane xN = 0. Now, let r = r(x0 ) > 0 for which B̄r (0) ⊂ D satisfies
all the properties of the special case just dealt with before (in the present
case x0 = 0), and, for every δ ∈ (0, r], let Uδ denote the open neighborhood
of x0 defined by
Uδ := Φ−1 (Bδ (0)).
Subsequently, for every δ ∈ (0, r], we also set
Γδ := Uδ ∩ Γ
and
v(x) := u(Φ−1 (x)), x ∈ D = Φ(BR (x0 )).
By (4.15), there exists a constant C1 > 0, independent of v, such that
∫ (∫ ∫ )
|v| dS ≤ C1
p
|v| dx +
p
|∇v| dx .
p
Φ(Γr/2 ) Φ(Ur ∩Ω) Φ(Ur ∩Ω)
Consequently, performing the change of variable x = Φ(y), it becomes

apparent that there exists C2 = C2 (x0 , Ω) > 0 such that
∫ (∫ ∫ )
|u| dS ≤ C2
p
|u| dy +
p
|∇u| dy
p
Γr/2 Ur ∩Ω Ur ∩Ω
(∫ ∫ )
≤ C2 |u|p dy + |∇u|p dy
Ω Ω
for all u ∈ C (Ω̄). This ends the proof of (4.13).

1
As Γ is compact, there exist m ∈ N and m points

x0,j ∈ Γ, 1 ≤ j ≤ m,
x0
Ur
U r/2
BR (x0)
Fig. 4.3 Ur/2 = Φ−1 (Br/2 (0)) and Ur = Φ−1 (Br (0))
such that
∪
m
Γ= Γr(x0,j )/2 .
j=1
Therefore,
∫ m ∫
∑
|u|p dS ≤ |u|p dS
Γ j=1 Γr(x0,j )/2
∑
m (∫ ∫ )
≤ C2 (x0,j , Ω) |u| dx +
p
|∇u| dx
p
j=1 Ω Ω
(∫ ∫ )
≤ C3 |u| dx +
p
|∇u| dx ,
p
Ω Ω
where
C3 := m max C2 (x0,j , Ω).
1≤j≤m
As the constant C3 only depends on the geometry of Ω, the proof of (4.12)
is complete.
Now, we will construct the trace operator TΓ . Naturally, for every u ∈
C (Ω̄), we define
1
TΓ u := u|Γ
as the restriction of u to the component Γ of the boundary ∂Ω.

Suppose u ∈ W 1,p (Ω). Then, according to Theorem 4.1, there exists a
sequence ϕn ∈ C ∞ (Ω̄), n ≥ 1, such that
lim ϕn = u in W 1,p (Ω).
n→∞
By (4.12), we have that
∥TΓ ϕn − TΓ ϕm ∥Lp (Γ) = ∥ϕn − ϕm ∥Lp (Γ) ≤ C∥ϕn − ϕm ∥W 1,p (Ω)
and hence, the limit
TΓ u := lim ϕn |Γ
n→∞
is well defined in Lp (Γ). This definition is consistent. Indeed, if ψn ∈
C ∞ (Ω̄), n ≥ 1, is another sequence for which
lim ψn = u in W 1,p (Ω),
n→∞
then, it follows from (4.12) that
∥TΓ ϕn − TΓ ψn ∥Lp (Γ) ≤ C∥ϕn − ψn ∥W 1,p (Ω)
for all n ≥ 1, and, therefore,
lim ∥ϕn − ψn ∥Lp (Γ) = 0.
n→∞
Consequently,
TΓ u := lim ϕn |Γ = lim ψn |Γ in Lp (Γ)
n→∞ n→∞
and the definition of TΓ is indeed consistent. Obviously, TΓ is linear, by
definition. Moreover, owing to (4.12),
∥ϕn ∥Lp (Γ) ≤ C∥ϕn ∥W 1,p (Ω) , n ≥ 1,
and hence, letting n → ∞ provides us with the estimate
∥TΓ u∥Lp (Γ) ≤ C∥u∥W 1,p (Ω)
for all u ∈ W 1,p (Ω), which shows the continuity of TΓ .
Finally, let u ∈ W 1,p (Ω) ∩ C(Ω̄). Then, according to Theorem 4.1, there
exists a sequence ϕn ∈ C ∞ (Ω̄), n ≥ 1, such that
( )
lim ∥ϕn − u∥W 1,p (Ω) + ∥ϕn − u∥C(Ω̄) = 0.
n→∞
Consequently,
TΓ u := lim TΓ ϕn = lim ϕn |Γ = u|Γ .
n→∞ n→∞
The uniqueness of TΓ is based upon the fact that
TΓ ϕ = ϕ|Γ for all ϕ ∈ C ∞ (Ω̄),
as C ∞ (Ω̄) is dense in W 1,p (Ω). The proof is complete.
Subsequently, for every Γ ∈ {Γ0 , Γ1 , ∂Ω} and p ∈ [1, ∞), we will consider
the closed subspaces
WΓ1,p (Ω) := TΓ−1 (0) = N [TΓ ], (4.16)
where TΓ is the trace operator of W 1,p (Ω) on Γ. Then, the following density
result holds.
Theorem 4.7. Let Γ ∈ {Γ0 , Γ1 , ∂Ω} and p ∈ [1, ∞). Then,

W 1,p (Ω)
WΓ1,p (Ω) = CΓ∞ (Ω) . (4.17)
In other words, for any given u ∈ W 1,p (Ω), one has that u ∈ WΓ1,p (Ω), i.e.,
TΓ u = 0, if and only if, for some sequence ϕn ∈ CΓ∞ (Ω), n ≥ 1,
lim ∥ϕn − u∥W 1,p (Ω) = 0. (4.18)
n→∞
Remark 4.1. According to (4.17), we can extend the definition (4.16) up

to cover the case p = ∞ by simply setting
W 1,∞ (Ω)
WΓ1,∞ (Ω) := CΓ∞ (Ω) .
Proof. Let u ∈ W 1,p (Ω) for which there exists a sequence ϕn ∈ CΓ∞ (Ω),
n ≥ 1, satisfying (4.18). Then, by Theorem 4.6,
TΓ u = lim TΓ ϕn = lim ϕn |Γ = 0
n→∞ n→∞
and, therefore, u ∈ WΓ1,p (Ω).

It remains to prove that if u ∈ W 1,p (Ω) satisfies TΓ u = 0, then, there
exists a sequence ϕn ∈ CΓ∞ (Ω), n ≥ 1, satisfying (4.18). This property is
based on the following estimate
∫ ∫
2p−1 p
|u|p dx ≤ ϵ |∇u|p , ϵ > 0, ϵ ∼ 0, (4.19)
Aϵ p Aϵ
where we are denoting

Aϵ := {x ∈ Ω : dist (x, Γ) < ϵ}
for sufficiently small ϵ > 0. For these ϵ’s,
∂Aϵ = Γ ∪ Γϵ with Γϵ := {x ∈ Ω : dist(x, Γ) = ϵ}.
To prove (4.19), let ϕn ∈ C ∞ (Ω̄), n ≥ 1, satisfying (4.18). Such a sequence
exists by Theorem 4.1. Then, due to Theorem 4.6,
0 = TΓ u = lim TΓ ϕn = lim ϕn |Γ . (4.20)
n→∞ n→∞
For sufficiently small ϵ > 0, say 0 < ϵ ≤ ϵ0 , and every x ∈ Γϵ , there exists
some point yx := PΓ x ∈ Γ such that
ϵ = |x − yx | = dist (x, Γ)
(see Figure 4.4). Fix 0 < ϵ ≤ ϵ0 . Then, for every n ≥ 1, 0 < δ < ϵ, and
x ∈ Γδ , the following identities hold
∫ δ ( )
d x − yx
ϕn (x) = ϕn (yx ) + ϕn yx + t dt
0 dt δ
∫ δ⟨ ( ) ⟩
x − yx x − yx
= ϕn (yx ) + ∇ϕn yx + t , dt
0 δ δ
and hence,
∫ ( )
δ
x − yx
|ϕn (x)| ≤ |ϕn (yx )| + ∇ϕn yx + t dt. (4.21)
0 δ
x
ε
yx
Γε
Fig. 4.4 The projection of Γϵ into Γ, 0 < ϵ ≤ ϵ0
On the other hand, for every a ≥ 0, b ≥ 0, and p ≥ 1,
(a + b)p ≤ 2p−1 (ap + bp ). (4.22)
Indeed, (4.22) is obvious if ab = 0, for as 1 ≤ 2p−1 . So, suppose ab > 0.

Then, dividing by ap , it becomes apparent that (4.22) holds if and only if
(1 + x)p ≤ 2p−1 (1 + xp )
for all x ≥ 0. This estimate holds true because the auxiliary function
(1 + x)p
g(x) := , x ≥ 0,
1 + xp
satisfies
g ≥ 0, g(0) = 1, lim g(x) = 1, g(1) = 2p−1 ≥ 1,
x→∞
and g ′ (x) = 0 if and only if x = 1. Consequently,

g(x) ≤ g(1) = 2p−1 for all x ≥ 0.
Applying (4.22), we find from (4.21) that
[ (∫ ( ) )p ]
p p
δ
x − yx
|ϕn (x)| ≤ 2 p−1
|ϕn (yx )| + ∇ϕn yx + t dt
0 δ
and, according to Hölder inequality,
∫ (∫ ) p1
δ p−1
δ
p
|∇ϕn (z(x, t))| dt ≤ δ p |∇ϕn (z(x, t))| dt ,
0 0
where we have denoted

x − yx
z(x, t) := yx + t.
δ
Thus, for every 0 < δ < ϵ and x ∈ Γδ ,
( ∫ )
δ
p p p
|ϕn (x)| ≤ 2 p−1
|ϕn (yx )| + δ p−1
|∇ϕn (z(x, t))| dt
0
and, therefore, integrating in the (N − 1)-dimensional C 1 -surface Γδ and,

later, in (0, ϵ), we are led to
∫ (∫ ∫ ϵ ∫ )
|ϕn |p ≤ 2p−1 |ϕn (yx )|p dx+ δ p−1 |∇ϕn (z)|p dz dδ
Aϵ Aϵ 0 Aδ
because Aϵ can be parametrized through the union of the Γδ ’s for δ ranging

in (0, ϵ). By the theorem of dominated convergence of Lebesgue, it follows
from (4.20) that
∫
lim |ϕn (yx )|p dx = 0,
n→∞ Aϵ
since yx ∈ Γ for all x ∈ Aϵ . Consequently, letting n → ∞ in the previous

inequality, (4.18) implies that
∫ ∫ ϵ (∫ ) ∫
ϵp
|u|p ≤ 2p−1 δ p−1 |∇u|p dδ ≤ 2p−1 |∇u|p ,
Aϵ 0 Aδ p Aϵ

Let n0 ∈ N such that 1/n < ϵ0 if n ≥ n0 , and, for every n ≥ n0 , pick
ζn ∈ CΓ∞ (Ω) such that 0 ≤ ζn ≤ 1,
{
1 in Ω̄ \ A n1 ,
ζn = and ∥∇ζn ∥C(Ω̄) ≤ C n,
0 in A 2n1 ,
for some constant C > 0 independent of n. By using an appropriate par-

tition of the unity and straightening ∂Ω locally, it is easy to realize the
existence of such a sequence of test functions. Subsequently, we consider
the auxiliary sequence
un := ζn u, n ≥ n0 .
Obviously,
{
u in Ω̄ \ A n1 ,
un ∈ W 1,p
(Ω), and un = (4.23)
0 in A 2n
1 ,
for all n ≥ n0 , by construction. Moreover,

∥un − u∥pW 1,p (Ω) = ∥u(ζn − 1)∥pW 1,p (Ω)
∫ ∫
= |u(ζn − 1)| +p
|∇ [u(ζn − 1)] |p
A1 A1
n
∫ n
≤ ∥u∥p p + |∇ [u(ζn − 1)] |p .

L (A1/n ) A1
n
Thus, letting n → ∞ shows that

∫
lim sup ∥un − u∥pW 1,p (Ω) ≤ lim sup |∇ [u(ζn − 1)] |p . (4.24)
n→∞ n→∞ A1
n
On the other hand, by Lemma 4.1 and (4.22), we find that

∫ ∫
|∇ [u(ζn − 1)] | =
p
|(ζn − 1)∇u + u∇ζn |p
A1 A1
n n
∫
≤2 p−1
(|(ζn − 1)∇u|p + |u∇ζn |p )
A1
n
∫
≤2 p−1
∥∇u∥ p
+2 p−1
|u∇ζn |p
(
Lp A1/n ) A1
n
for all n ≥ n0 . Thus, since

lim ∥∇u∥p p = 0,
n→∞ L (A1/n )
it follows from (4.24) that

∫
lim sup ∥un − u∥W 1,p (Ω) ≤ 2
p p−1
lim sup |u∇ζn |p . (4.25)
n→∞ n→∞ A1
n
Finally, by the choice of ζn , (4.19) implies that

∫ ∫
|u∇ζn |p ≤ C p np |u|p
A1 A1
∫
n n
p−1
2
≤ C p np |∇u|p
p np A1
n
C p p−1
= 2 ∥∇u∥p p →0
p L (A1/n )
as n → ∞. Therefore, (4.25) shows that
lim ∥un − u∥W 1,p (Ω) = 0. (4.26)

n→∞
It remains to prove that the un ’s can be approximated in W 1,p (Ω) by test

functions ϕn ∈ CΓ∞ (Ω), n ≥ 1. This can be accomplished because, according
to (4.23), un = 0 in A 2n1 for all n ≥ n0 . Indeed, consider the auxiliary
function
{ 1
Ce |x|2 −1 if |x| ≤ 1,
η(x) :=
0 if |x| ≥ 1,
where the constant C is given through

(∫ )−1
1
C := e |x|2 −1 dx .
B1 (0)
Then,
∫
η∈ C0∞ (RN ), η ≥ 0, supp η = B1 (0), η = 1. (4.27)
RN
Now, for every ϵ > 0, we introduce

(x)
ηϵ (x) := ϵ−N η , x ∈ RN .
ϵ
According to (4.27), it becomes apparent that
∫
ηϵ ∈ C0∞ (RN ), ηϵ ≥ 0, supp η = Bϵ (0), ηϵ = 1, (4.28)
RN
for all ϵ > 0. Further, for every n ≥ n0 and 0 < ϵ < 4n 1

, we consider the
function ∫
ψn,ϵ (x) = ηϵ (x − y)un (y) dy
∫Ω
= ηϵ (x − y)un (y) dy, x ∈ Ω.
Ω\A 1
4n
By (4.28), it readily follows that

ψn,ϵ ∈ CΓ∞ (Ω), lim ∥ψn,ϵ − un ∥W 1,p (Ω) = 0
ϵ↓0
for all n ≥ n0 . Fix n ≥ n0 . Then, there exists ϵ := ϵ(n) such that the test
function
ϕn := ψn,ϵ(n) ∈ CΓ∞ (Ω)
satisfies
1
∥ϕn − un ∥W 1,p (Ω) ≤ .
n
Therefore, for every n ≥ n0 , we obtain that
∥ϕn − u∥W 1,p (Ω) ≤ ∥ϕn − un ∥W 1,p (Ω) + ∥u − un ∥W 1,p (Ω)
1
≤
+ ∥u − un ∥W 1,p (Ω) .
n
Consequently, (4.26) implies that
lim ∥ϕn − u∥W 1,p (Ω) = 0.
n→∞
This section concludes with another important property of the trace
operator.
Proposition 4.1. Let Γ ∈ {Γ0 , Γ1 , ∂Ω}, p ∈ [1, ∞), and g ∈ C 1 (Ω̄). Then,
for every u ∈ W 1,p (Ω),
TΓ (gu) = gTΓ u.
Proof. Let u ∈ W 1,p (Ω). According to Theorem 4.1, there exists a se-
quence ϕn ∈ C ∞ (Ω̄), n ≥ 1, such that
lim ∥ϕn − u∥W 1,p (Ω) = 0.
n→∞
As g ∈ C (Ω̄), it is easy to see that
1
lim ∥gϕn − gu∥W 1,p (Ω) = 0.

n→∞
Hence, by Theorem 4.6, it becomes apparent that
TΓ (gu) = lim TΓ (gϕn ) = lim (gϕn ) = gTΓ u
n→∞ n→∞
in Lp (Γ).
Remark 4.2. To simplify the notation, throughout the rest of this book
we will denote
u|Γ := TΓ u ∈ Lp (Γ)
for all p ≥ 1, u ∈ W 1,p (Ω), and Γ ∈ {Γ0 , Γ1 , ∂Ω}, if there is no ambiguity.
4.3 Weak solutions
This section considers the linear boundary value problem (4.4) for an arbi-
trary ω ∈ R.
Definition 4.4. Let 1 ≤ p < ∞, f ∈ Lp (Ω), and u ∈ W 2,p (Ω). Then, u is

said to be a solution of (4.4) if (L + ω)u = f in Ω and Bu = 0 on ∂Ω in
the sense of traces, i.e., TΓ0 u = 0, or, equivalently, u ∈ WΓ1,p
0
(Ω) and
⟨TΓ1 ∇u, ν⟩ + βTΓ1 u = 0 on Γ1 .
Let p ∈ [2, ∞), f ∈ Lp (Ω), and suppose u ∈ W 2,p (Ω) ∩ WΓ1,p

0
(Ω) is a
solution of problem (4.4). Then, according to Definition 4.2 and Lemma
4.1, for every ϕ ∈ CΓ∞0 (Ω), we have that
∫ ∫
f ϕ = ϕ (L + ω)u
Ω Ω
∫ ∫ ∫
= − ϕ div (A∇u)+ ϕ ⟨b, ∇u⟩+ (c + ω)uϕ
∫Ω ∫Ω ∫
Ω
∫
= − div (ϕA∇u)+ ⟨∇ϕ, A∇u⟩+ ϕ⟨b, ∇u⟩+ (c+ω)uϕ
∫Ω Ω
∫ Ω
∫ Ω
= − ⟨ϕA∇u, n⟩ dS + ⟨A∇u, ∇ϕ⟩+ (⟨b, ∇u⟩+cu+ωu) ϕ.

∂Ω Ω Ω
Moreover,
∫ ∫
⟨ϕA∇u, n⟩ dS = ⟨ϕA∇u, n⟩ dS,
∂Ω Γ1
since ϕ vanishes in a neighborhood of Γ0 , and hence,

∫ ∫ ∫
∂u
⟨ϕA∇u, n⟩ dS = ϕ dS = − βuϕ dS,
∂Ω Γ1 ∂ν Γ1
because
∂u
0 = Bu = + βu on Γ1
∂ν
(see Remark 4.2). Consequently,

∫ ∫ ∫ ∫
f ϕ = ⟨A∇u, ∇ϕ⟩+ (⟨b, ∇u⟩ + cu + ωu) ϕ+ βuϕ dS
Ω Ω Ω Γ1
for all ϕ ∈ CΓ∞0 (Ω), and, therefore,

a(u, ϕ) = ⟨f, ϕ⟩L2 (Ω) ∀ ϕ ∈ CΓ∞0 (Ω), (4.29)
where
a : WΓ1,2
0
(Ω) × WΓ1,2
0
(Ω) −→ R
stands for the bilinear form defined by
∫ ∫ ∫
a(u, v) := ⟨A∇u, ∇v⟩+ (⟨b, ∇u⟩+cu+ωu) v+ βuv dS
Ω Ω Γ1
for all u, v ∈ WΓ1,2

0
(Ω) (see Remark 4.2). Throughout the rest of this book,
a(·, ·) will be referred to as the bilinear form associated to (4.4), or, equiv-
alently, to the tern (L + ω, B, Ω).
The following concept plays a fundamental role in the theory of partial
differential equations.
Definition 4.5 (Weak solution). For any f ∈ L2 (Ω), a function u ∈

WΓ1,2
0
(Ω) is said to be a weak solution of (4.4) if (4.29) is satisfied.
We have just seen that every solution u ∈ W 2,p (Ω), p ≥ 2, of (4.4) provides
us with a weak solution. Actually, the following characterization holds.
Proposition 4.2. Let f ∈ L2 (Ω) and u ∈ W 2,p (Ω) ∩ WΓ1,p 0

(Ω) for some
p ≥ 2. Then, u solves (4.4) in the sense of Definition 4.4 if and only if u
is a weak solution of (4.4).
Proof. We already know that u is a weak solution if it solves (4.4). So,

suppose u is a weak solution of (4.4). Then, it satisfies (4.29) and, conse-
quently, for every ϕ ∈ CΓ∞0 (Ω), we have that
∫
f ϕ = a(u, ϕ)
Ω
∫ ∫ ∫
= ⟨A∇u, ∇ϕ⟩+ (⟨b, ∇u⟩+cu+ωu) ϕ+ βuϕ dS
Ω Ω Γ1
∫ ∫ ∫
= ϕ (L + ω)u+ div (ϕA∇u)+ βuϕ dS
Ω Ω Γ1
∫ ∫ ∫
= ϕ (L + ω)u+ ⟨ϕA∇u, n⟩ dS + βuϕ dS,
Ω ∂Ω Γ1
whence
∫ ∫ ∫ ( )
∂u
fϕ = ϕ (L + ω)u + ϕ + βu dS. (4.30)
Ω Ω Γ1 ∂ν
Now, fix ψ ∈ C ∞ (Γ1 ) and let ϕn ∈ CΓ∞0 (Ω), n ≥ 1, be a sequence of test

functions such that
{ }
1
ϕn |Γ1 = ψ and supp ϕn ⊂ x ∈ Ω : dist (x, Γ1 ) ≤
n
for sufficiently large n, and
∥ϕn ∥L∞ (Ω) ≤ C, n ≥ 1,
for some positive constant C. By (4.30),

∫ ∫ ∫ ( )
∂u
f ϕn = ϕn (L + ω)u + ϕn + βu dS
Ω Ω Γ1 ∂ν
for all n ≥ 1. Hence, letting n → ∞ in this identity, the theorem of

dominated convergence of Lebesgue shows that
∫ ( )
∂u
ψ + βu dS = 0.
Γ1 ∂ν
As this identity holds for every ψ ∈ C ∞ (Γ1 ), it becomes apparent that
∂u
+ βu = 0 on Γ1 (4.31)
∂ν
and, therefore,
Bu = 0 on ∂Ω.
Moreover, substituting (4.31) into (4.30), yields

∫ ∫
fϕ = ϕ (L + ω)u ∀ ϕ ∈ CΓ∞0 (Ω)
Ω Ω
and, consequently,
(L + ω)u = f in Ω.

4.4 Continuity of the associated bilinear form
This section shows the continuity of the bilinear form a(·, ·) associated to
(4.4). The main result reads as follows.
Proposition 4.3. For every ω ∈ R there exists a constant C := C(ω) > 0

such that
|a(u, v)| ≤ C∥u∥W 1,2 (Ω) ∥v∥W 1,2 (Ω) (4.32)

Γ0 Γ0
for all u, v ∈ WΓ1,2

0
(Ω).
Proof. Let u, v ∈ WΓ1,2

0
(Ω). Then, by the Cauchy–Schwarz inequality,
and according to Remark 4.2, we find that
∫ ∫ ∫
|a(u, v)| ≤ |⟨A∇u, ∇v⟩|+ |⟨b, ∇u⟩+cu+ωu| |v|+ |β| |u| |v| dS
Ω Ω Γ1
∫ ∫
( )
≤ |A∇u| |∇v|+ |b| |∇u|+∥c+ω∥L∞ (Ω) |u| |v|
Ω Ω
∫
+∥β∥L∞ (Γ1 ) |u| |v| dS.
Γ1
Thus, setting
{ }
∥A∥∞ := max ∥aij ∥L∞ (Ω) ,
1≤i,j≤N
{ } (4.33)
∥b∥∞ := max ∥bj ∥L∞ (Ω) ,
1≤j≤N
and applying the Hölder inequality, we are led to
|a(u, v)| ≤ ∥A∥∞ ∥∇u∥L2 (Ω) ∥∇v∥L2 (Ω) + ∥b∥∞ ∥∇u∥L2 (Ω) ∥v∥L2 (Ω)
+ ∥c + ω∥L∞ (Ω) ∥u∥L2 (Ω) ∥v∥L2 (Ω)
+ ∥β∥L∞ (Γ1 ) ∥TΓ1 u∥L2 (Γ1 ) ∥TΓ1 v∥L2 (Γ1 ) .
On the other hand, owing to Theorem 4.6, we have that
∥TΓ1 w∥L2 (Γ1 ) ≤ ∥TΓ1 ∥L(W 1,2 (Ω),L2 (Γ ∥w∥W 1,2 (Ω)
1 ))
for all w ∈ W 1,2 (Ω). Therefore, (4.32) holds with

(
C(ω) = 2 ∥A∥∞ + ∥b∥∞ + ∥c∥L∞ (Ω) + |ω|
)
+ ∥β∥L∞ (Γ1 ) ∥TΓ1 ∥2L(W 1,2 (Ω),L2 (Γ .
1 ))

4.5 Invertibility of (4.4) when β ≥ 0
Throughout this section we suppose that β ≥ 0.
4.5.1 Coercivity of the associated bilinear form

The next result shows the coercivity of a(·, ·) for sufficiently large ω.
◦
Theorem 4.8 (of Garding). Suppose β ≥ 0. Then, there exists ω0 ∈ R
such that for every ω ≥ ω0 there is a constant α := α(ω) > 0 for which
|a(u, u)| ≥ α∥u∥2W 1,2 (Ω) ∀ u ∈ WΓ1,2
0
(Ω). (4.34)
Γ0
Consequently, a(·, ·) is coercive in the Hilbert space

H := WΓ1,2
0
(Ω) (4.35)
for all ω ≥ ω0 .
Proof. Remember that the scalar product of H is given by

∫ ∫
⟨u , v⟩H := ⟨∇u, ∇v⟩ + uv, u, v ∈ H.
Ω Ω
As β ≥ 0 implies
∫
βu2 dS ≥ 0, u ∈ H,
Γ1

∫ ∫
a(u, u) ≥ ⟨A∇u, ∇u⟩ + (⟨b, ∇u⟩ + cu + ωu) u (4.36)
Ω Ω
for all u ∈ H. Let µ > 0 denote the ellipticity constant of L in Ω. Then,
⟨A∇u, ∇u⟩ ≥ µ |∇u|2 in Ω,
and hence,
∫
⟨A∇u, ∇u⟩ ≥ µ ∥∇u∥2L2 (Ω) . (4.37)
Ω
Moreover,
∫ ∫
| ⟨b, ∇u⟩u| ≤ ∥b∥∞ |u| |∇u|,
Ω Ω
where ∥b∥∞ is given by (4.33), and
|∇u| ϵ2 |∇u|2
|u| |∇u| = ϵ|u| ≤ |u|2 + in Ω
ϵ 2 2ϵ2
for all ϵ > 0. Thus, setting η := ϵ2 , shows that

∫ ( )
∥b∥∞ 1
| ⟨b, ∇u⟩u| ≤ η ∥u∥2L2 (Ω) + ∥∇u∥2L2 (Ω)
Ω 2 η
for all η > 0. In particular,

∫ ( )
∥b∥∞ 1
⟨b, ∇u⟩u ≥ − η ∥u∥2L2 (Ω) + ∥∇u∥2L2 (Ω) . (4.38)
Ω 2 η
Also,
∫ ( )
(cu + ωu) u ≥ ω + inf c ∥u∥2L2 (Ω) . (4.39)
Ω Ω
Consequently, substituting (4.37), (4.38) and (4.39) into (4.36), we find

that, for every η > 0 and u ∈ H,
( )
∥b∥∞ 1
a(u, u) ≥ µ ∥∇u∥2L2 (Ω) − η ∥u∥2L2 (Ω) + ∥∇u∥2L2 (Ω)
2 η
( )
+ ω + inf c ∥u∥2L2 (Ω)
Ω
( ) ( )
∥b∥∞ ∥b∥ ∞
= µ− ∥∇u∥L2 (Ω) + ω + inf c −
2
η ∥u∥2L2 (Ω) .
2η Ω 2
Pick δ ∈ (0, µ), let η > 0 sufficiently large so that
∥b∥∞
µ− > µ − δ,
2η
and set
∥b∥∞
ω0 := − inf c + η + µ − δ.
Ω 2
Then, for every ω ≥ ω0 and u ∈ H, we have that
( ) ( )
∥b∥∞ ∥b∥∞
a(u, u) ≥ µ − ∥∇u∥L2 (Ω) + ω + inf c −
2
η ∥u∥2L2 (Ω)
2η Ω 2
( )
≥ (µ − δ) ∥∇u∥2L2 (Ω) + ∥u∥2L2 (Ω)
2
= (µ − δ)∥u∥W 1,2 (Ω) .
Γ0

4.5.2 Existence of weak solutions. The resolvent operator

This section applies Theorem 3.5 to obtain the next fundamental result.
Theorem 4.9 (of invertibility of (L, B, Ω)). Suppose the bilinear form
a(·, ·) associated to the problem (4.4) is coercive for some ω ∈ R. Then,
(4.4) possesses a unique weak solution u ∈ WΓ1,2
0
(Ω). Moreover, if we denote
it by
u := (L + ω)−1 f ∈ WΓ1,2
0
(Ω),
then, the resolvent operator
(L+ω)−1
L2 (Ω) −→ WΓ1,2
0
(Ω)
f 7→ (L + ω)−1 f
is linear and continuous, i.e.,
( )
(L + ω)−1 ∈ L L2 (Ω), WΓ1,2
0
(Ω) . (4.40)
Therefore, if we denote by J the compact imbedding

J : WΓ1,2
0
(Ω) ,→ L2 (Ω),
then, the composition
J(L + ω)−1 : L2 (Ω) −→ L2 (Ω)
is a linear and compact operator.
Remark 4.3. Owing to Theorem 4.8, when β ≥ 0, there exists ω0 > 0

such that a(·, ·) is coercive for all ω ≥ ω0 . Therefore, the conclusions of
Theorem 4.9 hold true for all ω ≥ ω0 .
Proof. Consider the Hilbert space H := WΓ1,2 0

(Ω) and the bilinear form
a : H → R associated to (4.4). By Proposition 4.3, a(·, ·) is continuous,
2
and we are assuming that it is coercive for some ω ∈ R. Moreover, for every
f ∈ L2 (Ω), the map
∫
1,2
⟨f, ·⟩L2 (Ω) : WΓ0 (Ω) −→ R, u 7→ ⟨f, u⟩L2 (Ω) = f u,
Ω
is linear and continuous, and hence it belongs to the topological dual H ′ of

H. Therefore, according to Theorem 3.5, for every f ∈ L2 (Ω) there exists
a unique
u := (L + ω)−1 f ∈ H
such that
a(u, v) = ⟨f, v⟩L2 (Ω) ∀ v ∈ H.
Obviously, u is the unique weak solution of (4.4).

Subsequently, we consider arbitrary A, B ∈ R, f, g ∈ L2 (Ω), and v ∈ H.
Then,
⟨Af + Bg, v⟩L2 (Ω) = A⟨f, v⟩L2 (Ω) + B⟨g, v⟩L2 (Ω)
( ) ( )
= A a (L + ω)−1 f, v + B a (L + ω)−1 g, v
( )
= a A(L + ω)−1 f + B(L + ω)−1 g, v .
Consequently, by the uniqueness of the weak solution,
(L + ω)−1 (Af + Bg) = A(L + ω)−1 f + B(L + ω)−1 g
and, therefore, (L + ω)−1 is a linear operator.

Now, let {fn }n≥1 be a sequence in L2 (Ω) such that
lim ∥fn − f ∥L2 (Ω) = 0 (4.41)

n→∞
and set
u := (L + ω)−1 f, un := (L + ω)−1 fn , n ≥ 1.
By definition, we have that
a(un − u, un − u) = ⟨fn − f, un − u⟩L2 (Ω)
for all n ≥ 1. Consequently, as a is coercive, there exists a constant C > 0

such that
∥un − u∥2H ≤ C |a(un − u, un − u)|

= C |⟨fn − f, un − u⟩L2 (Ω) |
≤ C ∥fn − f ∥L2 (Ω) ∥un − u∥H
for all n ≥ 1. Therefore, (4.41) implies
lim ∥un − u∥H = 0

n→∞
and shows (4.40).

The compactness of the imbedding J was already established by Theo-
rem 4.5. The proof is complete.
4.6 Invertibility of (4.4) for arbitrary β
The main goal of this section is to extend the result of Remark 4.3 up to
cover the general case when β changes sign on Γ1 . To accomplish this task,
besides Assumptions B1 and B2, we will impose
B3. There exist ψ ∈ C 2 (Ω̄) and γ > 0 such that
∂ψ
(x) ≥ γ for all x ∈ Γ1 .
∂ν
By Lemma 2.1, B3 holds if ∂Ω is of class C 2 . According to Proposition 2.1,
under these assumptions, there exist ω0 ∈ R and h ∈ C 2 (Ω̄), with h(x) > 0
for all x ∈ Ω̄, such that h is a strict supersolution of (L + ω, B, Ω) for all
ω ≥ ω0 . Actually, going back to the proof of Proposition 2.1, it becomes
apparent that we can make the choice
h(x) = eM ψ(x) , x ∈ Ω̄, (4.42)
for any M > 0 satisfying
Mγ + β > 0 on Γ1 . (4.43)
2 Lh
bh := b − A∇h, ch := , (4.44)
h h
and consider the operator
Lh := −div (A∇ ·) + ⟨bh , ∇ ·⟩ + ch . (4.45)
Then,
bh ∈ (L∞ (Ω)) , ch ∈ L∞ (Ω),
N
because h ∈ C 2 (Ω̄), and, much like in the proof of Theorem 1.7, a straight-
forward computation shows that
u
Lu = h Lh in Ω (4.46)
h
for all function u almost everywhere twice classically differentiable in Ω.
Similarly, we introduce
Bh
βh := (4.47)
h
and the boundary operator
{
D on Γ0 ,
Bh := ∂ (4.48)
∂ν + βh on Γ1 .
By Assumption B3, (4.42) and (4.43) imply that

Bh ∂ψ
βh = =M + β ≥ Mγ + β > 0 on Γ1 . (4.49)
h ∂ν
Moreover, βh ∈ C(Γ1 ) and
u
Bu = h Bh on ∂Ω (4.50)
h
for all u ∈ C 1 (Ω̄). Therefore, if u is a classical solution of (4.4) in Ω, as
discussed in Definition 4.1, and we perform the change of variable
u
v := (4.51)
h
then, v is twice classically differentiable a.e. in Ω, and, thanks to (4.46)
and (4.50), it is a classical solution of
{
(Lh + ω)v = fh in Ω,
(4.52)
Bh v = 0 on ∂Ω.
As the converse is as well true, (4.51) establishes a bijection between the

classical solutions of (4.4) and (4.52). According to (4.48) and (4.49), the
transformed problem (4.52) fits within the framework of Section 4.5, as
βh ≥ 0. Not surprisingly, (4.51) also establishes a bijection between the
corresponding weak solutions. Indeed, let
a , ah : WΓ1,2
0
(Ω) × WΓ1,2
0
(Ω) −→ R
denote the bilinear forms associated to (4.4) and (4.52), respectively, i.e.,
∫ ∫ ∫
a(u, w) := ⟨A∇u, ∇w⟩+ (⟨b, ∇u⟩+cu+ωu) w+ βuw dS
Ω Ω Γ1
∫ ∫ ∫
ah (u, w) := ⟨A∇u, ∇w⟩+ (⟨bh , ∇u⟩+ch u+ωu) w+ βh uw dS,
Ω Ω Γ1
for all u, w ∈ WΓ1,2

0
(Ω), where the integrals on Γ1 should be understood in
the sense of traces. The following result establishes a fundamental relation-
ship between these bilinear forms.
Lemma 4.2. For every v, w ∈ WΓ1,2

0
(Ω),
( w )
a hv, = ah (v, w).
h
Proof. Let v, w ∈ WΓ1,2

0
(Ω). Then, by definition,
( ) ∫ ∫ ∫
w w w
a hv, = ⟨A∇(hv), ∇ ⟩ + ⟨b, ∇(hv)⟩ + (c + ω)vw
h h h
Ω Ω
∫ Ω
w
+ βTΓ1 (hv)TΓ1 dS,
Γ1 h
where TΓ1 is the trace operator of W 1,p (Ω) on Γ1 . Moreover, by Proposition
4.1 and Remark 4.2,
∫ ∫
w
βTΓ1 (hv)TΓ1 dS = βvw dS.
Γ1 h Γ1
Thus, expanding the underlying derivatives we are led to

( ∫
w) h∇w − w∇h
a hv, = ⟨hA∇v + vA∇h, ⟩
h h2
Ω
∫ ∫ ∫
w
+ ⟨b, h∇v + v∇h⟩ + (c + ω)vw + βvw dS
Ω h Ω Γ1
∫ ∫ ∫
w v
= ⟨A∇v, ∇w⟩ − ⟨A∇v, ∇h⟩ + ⟨A∇h, ∇w⟩
h h
Ω
∫ Ω
∫ Ω
∫
vw vw
− ⟨A∇h, ∇h⟩ 2 + ⟨b, ∇v⟩w + ⟨b, ∇h⟩
h h
∫Ω ∫ Ω Ω
+ (c + ω)vw + βvw dS.

Ω Γ1
On the other hand, it follows from (4.44) that

∫ ∫ ∫
w A∇h
⟨b, ∇v⟩w− ⟨A∇v, ∇h⟩ = ⟨b − , ∇v⟩w
h h
Ω Ω
∫Ω
∫
w
= ⟨bh , ∇v⟩w+ ⟨A∇h, ∇v⟩
Ω Ω h
and hence,
( ∫ ∫ ∫
w) w
a hv, = ⟨A∇v, ∇w⟩ + ⟨bh , ∇v⟩w + ⟨A∇h, ∇v⟩
h h
∫Ω Ω
∫ Ω
∫
v vw vw
+ ⟨A∇h, ∇w⟩ − ⟨A∇h, ∇h⟩ 2 + ⟨b, ∇h⟩
h h h
∫Ω ∫ Ω Ω
+ (c + ω)vw + βvw dS.

Ω Γ1
Thus, taking into account that

∫ ∫ ∫
w v ∇h
⟨A∇h, ∇v⟩ + ⟨A∇h, ∇w⟩ = ⟨A , ∇(vw)⟩,
Ω h Ω h Ω h
∫ ∫ ∫
∇h vw vw
⟨A , ∇(vw)⟩ − ⟨A∇h, ∇h⟩ = ⟨A∇h, ∇ ⟩,
Ω h Ω h2 Ω h
and
∫ ∫ ) ∫
( vw
vw vw
⟨A∇h, ∇ ⟩= div A∇h − div (A∇h)
h h h
Ω
∫Ω ∫ Ω
vw ∂h vw
= dS − div (A∇h) ,
Γ1 h ∂ν Ω h
we find that
( ∫ ∫ ∫
w) vw
a hv, = ⟨A∇v, ∇w⟩+ ⟨bh , ∇v⟩w− div (A∇h)
h h
∫Ω ∫ Ω ∫ Ω( )
vw 1 ∂h
+ ⟨b, ∇h⟩ + (c + ω)vw+ β+ vw dS
Ω h Ω Γ1 h ∂ν
and, consequently,
( w)
a hv,
∫ h ∫ ∫ ∫
(L+ω)h Bh
= ⟨A∇v, ∇w⟩+ ⟨bh , ∇v⟩w+ vw+ vw dS
Ω Ω Ω h Γ h
∫ ∫ ∫ ∫ 1
= ⟨A∇v, ∇w⟩+ ⟨bh , ∇v⟩w+ (ch +ω)vw+ βh vw dS
Ω Ω Ω Γ1
= ah (v, w),
The next result is a corollary from Lemma 4.2.
Theorem 4.10. Suppose B1–B3, and let h be the function (4.42), with M
satisfying (4.43). Then, v ∈ WΓ1,20
(Ω) is a weak solution of (4.52) if and
only if u := hv ∈ WΓ1,2
0
(Ω) is a weak solution of (4.4).
Proof. By definition, v is a weak solution of (4.52) if and only if

f
ah (v, ϕ) = ⟨ , ϕ⟩L2 (Ω) (4.53)
h
for all ϕ ∈ CΓ∞0 (Ω). By Lemma 4.2, (4.53) can be expressed in the form
ϕ ϕ
a(hv, ) = ⟨f, ⟩L2 (Ω) ,
h h
or, equivalently,
ϕ ϕ
a(u, ) = ⟨f, ⟩L2 (Ω) for all ϕ ∈ CΓ∞0 (Ω). (4.54)
h h
As CΓ∞0 (Ω) is dense in the subspace of C 2 (Ω̄) defined by
CΓ20 (Ω) := { u ∈ C 2 (Ω̄) : supp ϕ ⊂ Ω ∪ Γ1 },

it follows from Proposition 4.3 that (4.54) holds if and only if
a(u, ψ) = ⟨f, ψ⟩L2 (Ω) for all ψ ∈ CΓ∞0 (Ω),
i.e., if and only if u is a weak solution of (4.4).
Theorem 4.11. Suppose B1–B3. Then, there exists ω0 ∈ R such that, for
every ω ≥ ω0 , the problem (4.4) possesses a unique weak solution
u := (L + ω)−1 f ∈ WΓ1,2
0
(Ω).
Moreover, all the conclusions of Theorem 4.9 hold true.
Proof. Let h be the function (4.42), with M satisfying (4.43). Then, by

the choice of h, βh ≥ 0. Thus, according to Theorem 4.9 and Remark 4.3,
there exists ω0 > 0 such that, for each ω ≥ ω0 , the problem (4.52) possesses
a unique weak solution
f
v := (Lh + ω)−1 ∈ WΓ1,2 (Ω).
h 0
By Theorem 4.10, the function

f
u := hv = h(Lh + ω)−1 ∈ WΓ1,2 (Ω)
h 0
provides us with the unique weak solution of (4.4). The remaining asser-
tions of the theorem can be easily obtained from the identity
.
(L + ω)−1 = h(Lh + ω)−1
h
by applying Theorem 4.9 to the transformed problem (4.52).
The exposition of Section 4.1 is based on Chapter 7 of D. Gilbarg and N.

Trudinger [79], and on Chapter 5 of L. C. Evans [60]. Some monographs and
classical textbooks about Sobolev spaces are R. A. Adams [1], A. Friedman
[67], J. L. Lions and E. Magenes [130], P. Malliavin [156], V. G. Maz’ja
[159], C. B. Morrey [167], J. Nec̆as [170], H. Triebel [220], and J. Wloka
[226], among others.
Although the basic concepts of measure and integral used in this book
go back to H. Lebesgue [124], it was F. Riesz [189] who first introduced and
studied the spaces Lp (I), I being an interval, for 1 < p < ∞. These contri-
butions have shown to be a milestone for the development of mathematical
analysis. Though inequality (4.5) is attributed to O. Hölder [101], it might
be attributed to L. J. Rogers [193] too.
As in most of the specialized literature, we have attributed the spaces
k,p
W (Ω) to S. L. Sobolev [210], though a number of spaces of weakly
differentiable functions had been previously introduced by C. B. Morrey
[166, 167]. Later, L. Schwartz [199] established the foundations of the ab-
stract theory of distributions.
The most pioneering approximation results in the sprit of Theorem 4.1
go back to K. O. Friedrichs [69]. Theorem 4.1 goes back to Theorem 2.3 of
J. Deny and J. L. Lions [52] for k = 1, and it was later extended by N. G.
Meyers and J. Serrin [161] to cover the general case when k ≥ 1.
In Definition 4.3, the concept of Lipschitz function goes back to R.
Lipschitz [133], and the concept of Hölder function seem to go back to O.
Hölder [101], though the Hölder spaces introduced in Section 4.1.3 might
go back to A. Korn [116] and J. Schauder [195, 196]. The interested reader
is advised to refer to Chapters 4 and 6 of D. Gilbarg and N. Trudinger [79]
for further details.
Except for the Hölder condition on the maximal order derivative in (4.8),
which goes back to C. B. Morrey [166], Theorem 4.2 is attributed to S. L.
Sobolev [210, 211]. Theorem 4.3 goes back, at least, to Problem 7.7 of D.
Gilbarg and N. Trudinger [79] (see Section 5.8.2 of L. C. Evans [60] for a
proof). Naturally, Theorem 4.4 is based upon the Lebesgue differentiation
theorem. It goes back to Theorem 12 of A. P. Calderón and A. Zygmund
[37] and Theorem VIII.1 of E. M. Stein [215] (cf. Section 5.8.3 of L. C. Evans
[60]). The particular feature that any locally Lipschitz continuous function
must be differentiable almost everywhere is attributed to H. Rademacher
[90]. The compactness results collected by Theorem 4.5 are attributed to
F. Rellich [188] in case p = 2 and to V. I. Kondrachov [115] in the general
case p > 1. Theorem 4.5 was later extended by E. Gagliardo [74] to cover
the general case p ≥ 1.
The exposition of Section 4.2 is based on Section 5.5 of L. C. Evans [60].
It includes the detailed proofs of the trace theorems because the results of
L. C. Evans [60] were proven in the very special case when Γ = ∂Ω, as in
most of the available literature. Although the proof of Theorem 4.6 is a
detailed re-elaboration of the proof of Theorem 1 in Section 5.5 of L. C.
Evans [60], our proof of Theorem 4.7 differs substantially from the proof of
Theorem 2 in Section 5.5 of L. C. Evans [60], which was suggested by W.
Schlag [197] (see p. 292 of [60]). Sharper results about the exact range of
the trace operator can be found on p. 196 of the classical textbook of H.
Brézis [29], and in J. L. Lions and E. Magenes [130].
According to Chapter 9 of D. R. Adams and L. I. Hedberg [2], the
results of Section 4.2 go back at least to S. L. Sobolev’s fundamental paper
[209], which appeared in 1937, but was translated to English only in 1963.
In his celebrated paper, S. L. Sobolev generalized some earlier work by K.
O. Friedrichs [68].
The variational approach to problem (4.4) studied in Sections 4.3-4.5
can be traced back as far as the works of D. Hilbert [98] and H. Lebesgue
[123], at least in the very special case when L = −∆ and Γ1 = ∅. Theorem
◦
4.8 goes back to L. Garding [77]. In its full generality, Theorem 4.9 should
be attributed to P. D. Lax and A. N. Milgram [122].
Although H. Amann [9] found some “inverse positivity” results for the
classical solutions of (4.4) when Ω is of class C 2 and β ∈ C 1 (Γ1 ), the ex-
istence results of Section 4.6 seem to go back to J. López-Gómez [147] for
arbitrary β.
In a number of circumstances along this chapter, the condition that
∂Ω is of class C 1 is far from necessary. For instance, for introducing the
Hölder spaces of continuous functions. While, in many others, it is a crucial
hypothesis for the validity of the results. For instance, Theorem 4.1 can
fail when ∂Ω is not of class C 1 .
A careful reading of this chapter reveals that β ∈ C(Γ1 ) can be relaxed
up to consider β ∈ L∞ (Γ1 ). Moreover, for the validity of the last assertion
of Theorem 4.6 it suffices to impose u ∈ W 1,2 (Ω) ∩ C(Ω ∪ Γ). But the
‘principle of greatest generality’ remains outside the scope of this book.
Chapter 5
Regularity of weak solutions
Throughout this chapter, besides the general assumptions of Chapter 4, we

assume that ∂Ω is of class C 2 . Then, according to Lemma 2.1, Assumptions
B1, B2 and B3 of Chapter 4 hold, as well as H1, H2 and H3 of Chapter 2.
It should be remembered that B3 was stated at the beginning of Section
4.6. Consequently, thanks to Theorem 4.11, there exists ω0 ∈ R such that
(4.4) possesses a unique weak solution
u := (L + ω)−1 f ∈ WΓ1,2
0
(Ω)
for every ω ≥ ω0 . The main goal of this chapter is to sketch a proof of
the fact that the weak solution u must be a classical solution of (4.4), in
the sense of Definition 4.1, if f ∈ L∞ (Ω). By the weak counterpart of the
classical theory developed in Chapter 2 that will be studied in Chapter 7,
this regularity result entails the strong positivity of the resolvent operators
(L + ω)−1 , ω ≥ ω0 ,
associated to (L, B, Ω). More precisely, our main goal in this chapter is
to provide the reader with a first informal approach to the problem of
the regularity of the weak solutions of (4.4) from a classical very fruitful
perspective. Consequently, this chapter has an expository nature.
Essentially, the main goal is accomplished through the elliptic Lp theory.
According to it, every weak solution u of (4.4) lies in the Sobolev space
W 2,p (Ω) if f ∈ Lp (Ω), p ≥ 2. Consequently,
∩
u∈ W 2,p (Ω) if f ∈ L∞ (Ω)
p≥2
−
and hence, by Theorems 4.2(ii) and 4.4, u ∈ C 1,1 (Ω̄) and it is twice clas-
sically differentiable almost everywhere in Ω. By Proposition 4.2, u must
be a solution of (4.4) in the sense of Definition 4.4. Moreover, owing to
129
Theorem 4.4, for any α ∈ NN with |α| ≤ 2, the classical derivative Dα u

equals the corresponding weak derivative of u and hence
(L + ω)u = f a.e. in Ω,
in the classical sense. Also, as u ∈ WΓ1,2
0
(Ω),
we find from (4.16) that
TΓ0 u = 0 and, consequently, thanks to Theorem 4.6,
u=0 on Γ0 ,
because u ∈ W 1,p
(Ω) ∩ C (Ω̄). Similarly,
1
∂u
+ βu = 0 on Γ1 ,
∂ν
and, therefore, any weak solution u of (4.4) must be a classical solution
−
satisfying u ∈ C 1,1 (Ω̄).
The fact that any weak solution of (4.4) belongs to the Sobolev space
W 2,p (Ω) if f ∈ Lp (Ω), p ≥ 2, will be proven through the method of continu-
ity, from the classical elliptic Lp -estimates (see Theorem 9.14 of D. Gilbarg
and N. Trudinger [79]) and the corresponding result for the Laplace oper-
ator −∆, which is a very classical result in potential theory. Basically, the
Lp regularity is a consequence from the theory of singular integrals of A. P.
Calderón and A. Zygmund [36], attributable to K. O. Friedrichs [70] and
to S. Agmon, A. Douglis, and L. Nirenberg [4]. The pioneering results of
K. O. Friedrichs [70] extended a classical result of H. Weyl [225], usually
referred to as the Weyl lemma, establishing that any weak solution of the
Laplace equation
−∆u = f ∈ L2
must be of class C ∞ almost everywhere in any domain D where f ∈ C ∞ (D).
The method of continuity was introduced by J. Schauder [195, 196] to ob-
tain maximal Hölder regularity for general second order elliptic operators
with Hölder continuous coefficients. Essentially, it consists in establishing
a continuous deformation, or homotopy, between the original elliptic differ-
ential operator and the classical Laplace operator for transferring all the
available regularity results from classical potential theory to the original
setting. The underlying ideas have been a milestone for the development of
modern nonlinear analysis and its applications to the theory of nonlinear
partial differential equations (see [140] and Chapter 12 of [153] for some
very recent advances in this direction).
This chapter pays special attention to the case of Dirichlet boundary
conditions (Γ1 = ∅). Consequently, as we shall use the method of continu-
ity to derive the general Lp regularity result, this chapter might actually
Regularity of weak solutions 131
be regarded as the resolution of Problem 9.8 of [79]. Getting general Lp

regularity results under the general boundary operator B dealt with in this
book is much harder technically and it remains outside the general scope
of this book. Some recent general Lp a priori estimates are now available
with “modern” proofs in R. Denk, M. Hiebber and J. Prüss [50, 51].
This chapter will provide with a general regularity scheme which might
be substantially expanded, completed, and polished up to generate a more
specialized monograph on Lp regularity theory. It is distributed as fol-
lows. Section 5.1 collects a fundamental theorem of A. P. Calderón [35]
establishing the Lp regularity in RN . Section 5.2 discusses the validity of
the underlying regularity result in bounded domains. Section 5.3 collects
the Lp a priori estimates established by Theorem 9.14 of D. Gilbarg and
N. Trudinger [79]. Section 5.4 studies the fundamental theorem support-
ing the method of continuity. Although it is Theorem 5.2 on p. 75 of D.
Gilbarg and N. Trudinger [79], the reader should be aware of the fact that
the proof of Theorem 5.2 of [79] contains a gap, as the operator T intro-
duced there might not be a contraction. Section 5.5 contains the resolution
of Problem 9.8 of [79]. Finally, Section 5.6 provides a scheme for extending
the previous regularity results to the general case when Γ1 ̸= ∅. As, in our
approach here, the underlying Banach spaces do actually change when the
deformation parameter of the differential operator varies, one must over-
come an additional highly technical difficulty which did not arise in the
previous analysis of the Dirichlet problem.
5.1 Lp (RN )-estimates for the Laplacian
For every f ∈ L1 (RN ), the Fourier transform of f is the bounded and

continuous function
∫
Ff (ξ) := e−i ⟨x,ξ⟩ f (x) dx, ξ ∈ RN ,
RN
while the inverse Fourier transform of f is defined by
∫
1
F−1 f (x) := ei ⟨x,ξ⟩ f (ξ) dξ, x ∈ RN .
(2π)N RN
If f ∈ L1 (RN ) and Ff ∈ L1 (RN ), then the Fourier inversion formula
f = F−1 Ff
holds. A fundamental property of the Fourier transform is the product
formula, establishing that if f, g ∈ L1 (RN ), then the convolution of f and
g, denoted by f ∗ g and defined as

∫ ∫
f ∗ g(x) := f (x − y)g(y) dy = f (y)g(x − y) dy
RN RN
satisfies f ∗ g ∈ L1 (RN ) and

F(f ∗ g) = F(f )F(g). (5.1)
Another important property of the Fourier transform establishes that
∂f
F (ξ) = iξj Ff (ξ), 1 ≤ j ≤ N,
∂xj
for sufficiently regular f , and hence,
F∆f (ξ) = −|ξ|2 Ff (ξ), ξ ∈ RN . (5.2)
According to (5.2), the problem of solving the Laplace equation
(−∆ + 1)u = f in RN (5.3)
can be transformed into the problem of finding some sufficiently regular
function u such that
Ff
Fu = in RN . (5.4)
1 + |ξ|2
The function
( ) ∫
−1 1 1 ei ⟨x,ξ⟩
G(x) := F (x) = dξ (5.5)
1 + |ξ|2 (2π)N RN 1 + |ξ|2
is known as the Bessel kernel. According to Section V.3.1 of E. M. Stein
[215],
∫ ∞
1 |x|2
t− 2 e−π t − 4π dt,
N t
G(x) = x ∈ RN .
4π 0
Thus, combining the Fourier inversion formula with (5.1), (5.4) and (5.5),
it becomes apparent that the best candidate to solve (5.3) is
∫
u(x) = G ∗ f (x) = G(x − y)f (y) dy, x ∈ RN .
RN
The next fundamental result of A. P. Calderón [35] confirms it.
Theorem 5.1 (of Lp (RN )-regularity). Let 1 < p < ∞. Then, for every
f ∈ Lp (RN ), the function
u := G ∗ f
satisfies u ∈ W 2,p (RN ) and it solves (5.3) a.e. in RN . Moreover, the map
(−∆+1)−1
Lp (RN ) −→ W 2,p (RN ) (5.6)
f 7→ G∗f
is a linear topological isomorphism. In particular, there exists a constant
C > 0 such that
C −1 ∥f ∥Lp (RN ) ≤ ∥G ∗ f ∥W 2,p (RN ) ≤ C∥f ∥Lp (RN ) (5.7)
for all f ∈ Lp (RN ).
Theorem 5.1 is a consequence from the fact that the Riesz transform
( )
ξj
Rj f := F−1 −i F(f ) , 1 ≤ j ≤ N,
|ξ|
can be extended from L2 (RN ) to a bounded operator on Lp (RN ), for all
1 < p < ∞, which is a fundamental consequence of the theory of singular
integrals of A. P. Calderón and A. Zygmund [36]. Different proofs of Theo-
rem 5.1 are given in A. P. Calderón [35] and in Theorem V.3 of E. M. Stein
[215].
As a byproduct from (5.7), we find that
( )
∥u∥W 2,p (RN ) ≤ C ∥ − ∆u∥Lp (RN ) + ∥u∥Lp (RN ) (5.8)
for all u ∈ W 2,p (RN ). Indeed, for every u ∈ W 2,p (RN ), there exists f ∈
Lp (RN ) such that u = G ∗ f and hence,
∥u∥W 2,p (RN ) = ∥G ∗ f ∥W 2,p (RN ) ≤ C∥f ∥Lp (RN )
= C∥ − ∆u + u∥Lp (RN )
( )
≤ C ∥ − ∆u∥Lp (RN ) + ∥u∥Lp (RN ) .
Theorem 5.1 can be generalized to cover the more general case when −∆
is substituted by a second order elliptic operator with constant coefficients
of the form
L = − div (A∇ · ), A ∈ Msym
N (R).
In such case, the generalized Bessel kernel

( )
−1 1
GA := F
ξ T Aξ + 1
plays a similar role as the former G = GI .
More generally, for every s > 0, one can consider the Bessel kernel
( ) ∫
−1 1 1 ei ⟨x,ξ⟩
Gs (x) := F s/2
= N s/2
dξ
(1 + |ξ|2 ) (2π) RN (1 + |ξ|2 )
∫ ∞
1 s−N
−1 −π
|x|2
− 4π
t
= t 2 e t dt, x ∈ RN ,
(4π)s/2 Γ(s/2) 0
where Γ stands for gamma of L. Euler
∫ ∞
Γ(α) := tα−1 e−t dt, α > 0,
0
as well as the associated Bessel potential space
Ls,p (RN ) := { Gs ∗ f : f ∈ Lp (RN ) }, 1 < p < ∞,
equipped with the norm
∥Gs ∗ f ∥Ls,p (RN ) := ∥f ∥Lp (RN ) .
Adopting this perspective, A. P. Calderón [35] actually proved that
Lk,p (RN ) = W k,p (RN ), 1 < p < ∞, (5.9)
with equivalent norms, for all integer k ≥ 0. Theorem 5.1 has already estab-
lished (5.9) for k = 2. By (5.9), the Bessel potential spaces can be regarded
as generalized Sobolev spaces incorporating fractional order derivatives of
arbitrary order.
Another fruitful approach to the problem of the construction of gen-
eralized Sobolev spaces incorporating fractional order derivatives was pro-
posed by E. N. Slobodeckii [205]. According to it, for every s ∈ (0, 1) and
p ∈ [1, ∞), the Slobodeckii space W s,p (Ω) stands for the Banach space of
all functions u ∈ Lp (Ω) such that
∫
|u(x) − u(y)|p
Ip (u) := dx dy < ∞
Ω×Ω |x − y|
N +ps
equipped with the Slobodeckii norm

∥u∥W s,p (Ω) = [Ip (u)]1/p .
When s ∈ R \ Z satisfies s > 1, the Sobolev–Slobodeckii space W s,p (Ω) is
defined as the set of functions u ∈ W [s],p (Ω) such that Dα u ∈ W s−[s],p (Ω)
for all α with |α| = [s]. Naturally, in this case the Sobolev–Slobodeckii
norm is defined by
∑
∥u∥W s,p (Ω) = ∥u∥W [s],p (Ω) + ∥Dα u∥W s−[s],p (Ω) .
|α|=[s]
By using local coordinate charts, the Sobolev–Slobodeskii spaces can be ex-

tended to any sufficiently smooth manifold. In particular, it makes sense to
consider the spaces W s,p (Γ) for all s > 0, p ∈ [1, ∞), and Γ ∈ {Γ0 , Γ1 , ∂Ω}.
The next remark collects some useful properties of these spaces.
Remark 5.1.
(a) For every s > 0 and Γ ∈ {Γ0 , Γ1 , ∂Ω}, the imbedding W s,2 (Γ) ,→ L2 (Γ)
is compact (see, e.g., Theorem 7.10 of J. Wloka [226]).
(b) For every Γ ∈ {Γ0 , Γ1 , ∂Ω}, k ∈ {1, 2} and p > 1, the trace operator TΓ
of W 1,p (Ω) on Γ satisfies
( 1
)
TΓ ∈ L W k,p (Ω), W k− p ,p (Γ)
and hence
( 1 1
)
B ∈ L W 2,p (Ω), W 2− p ,p (Γ0 ) × W 1− p ,p (Γ1 ) for all p > 1
(see, e.g., Theorem 8.7 of J. Wloka [226]).
5.2 Lp (Ω)-estimates for the Laplacian
Subsequently, we denote by ψ(r), r > 0, the function defined by

 2−N
 r
 (2−N )ωN , N > 2,
ψ(r) :=

 log r ,
2π N = 2,
where ωN stands for the surface “area” of the unit sphere in RN
2π N/2
ωN := .
Γ(N/2)
Then, the function
K(x, y) := ψ(|x − y|), (x, y) ∈ RN × RN , x ̸= y,
provides us with the normalized fundamental solution of the Laplace equa-
tion, and the Green representation formula establishes that
∫ ∫ [ ]
∂K ∂u
u(y) = K(x, y)∆u(x) dx + u(x) (x, y) − K(x, y) dSx
Ω ∂Ω ∂nx ∂nx
for all u ∈ C 2 (Ω̄) and y ∈ Ω (see, e.g., Chapter 4 of F. John [108], or Section
2.4 of D. Gilbarg and N. Trudinger [79]). For a measurable function f , the
integral
∫
N (f ) := K(x, ·)f (x) dx
Ω
is called the Newtonian potential with density f , and, independently of the

values of u and ∂u/∂nx on ∂Ω, the function
∫ [ ]
∂K ∂u
H(u) := u(x) (x, ·) − K(x, ·) (x) dSx
∂Ω ∂nx ∂nx
is harmonic in Ω. Thus, the Green representation formula actually estab-
lishes that any function u ∈ C 2 (Ω̄) differs from the Newtonian potential
with density ∆u in a harmonic function. Precisely,
u = N (∆u) + H(u). (5.10)
As H(u) is real analytic in Ω, because it inherits the regularity of K out-

side its singularity, (5.10) entails that u and N (∆u) must have the same
regularity. This is an extremely important observation, because of the next
consequence of the theory of A. P. Calderón and A. Zygmund [36].
Theorem 5.2. For every 1 < p < ∞ and f ∈ Lp (Ω), the function
∫
w := − K(x, ·)f (x) dx
Ω
2,p
belongs to W (Ω) and it satisfies
−∆w = f a.e. in Ω.
According to Theorem 5.2, the Green representation formula makes

sense as soon as u ∈ W 2,p (Ω) for some 1 < p < ∞, because, in such case,
the traces of u and ∂u/∂nx on ∂Ω are well defined in Lp (∂Ω). Consequently,
by (5.10) and Theorem 5.2, the most natural space for the solutions of the
Laplace equation
−∆u = f ∈ Lp (Ω)
is W 2,p (Ω), much like in the context of Theorem 5.1. The reader is advised
to read the proof of Theorem 9.9 of D. Gilbarg and N. Trudinger [79] for a
proof of Theorem 5.2.
Now, suppose h ∈ C 2 (Ω̄) is harmonic in Ω. Then, by the second Green
identity, we obtain that
∫ ∫ ( )
∂h ∂u
0= h∆u dx + u −h dS (5.11)
Ω ∂Ω ∂n ∂n
for all u ∈ C 2 (Ω̄). Thus, setting
G(x, y) := K(x, y) + h(x), x, y ∈ Ω̄, x ̸= y,

and adding up (5.10) and (5.11) yields

∫ ∫ ( )
∂G ∂u
u(y) = G(·, y)∆u dx+ u(x) (x, y)−G(x, y) dSx (5.12)
Ω ∂Ω ∂nx ∂nx
for all u ∈ C 2 (Ω̄) and y ∈ Ω. Note that G(x, y) again is a fundamental
solution of the Laplace equation and, therefore, (5.12) provides us with a
generalized version of the Green representation formula (5.10). Now, for a
given y ∈ Ω, let choose the harmonic function h := h(·, y) to satisfy
h(x, y) = −K(x, y) ∀ x ∈ ∂Ω.
Then, G(x, y) = 0 for all x ∈ ∂Ω, and hence it follows from (5.12) that
∫ ∫
∂G
u(y) = G(x, y)∆u(x) dx + u(x) (x, y) dSx
Ω ∂Ω ∂nx
for all y ∈ Ω and u ∈ C 2 (Ω̄). For such choice of h, G is called the Green
function of the Dirichlet problem in the domain Ω, and the function
∂G
P (x, y) := (x, y), x ∈ ∂Ω, y ∈ Ω,
∂nx
is usually referred to as the Poisson kernel of the problem. Naturally,
∫ ∫
u(y) = G(x, y)f (x) dx + P (x, y)g(x) dSx
Ω ∂Ω
provides us with the unique solution of the Dirichlet problem
{
∆u = f in Ω,
u=g on ∂Ω.
In particular,
∫
u(y) := − G(x, y)f (x) dx, y ∈ Ω, (5.13)
Ω
is the unique solution of
{
−∆u = f in Ω,
(5.14)
u=0 on ∂Ω.
The following result of S. Agmon, A. Douglis, and L. Nirenberg [4] ascer-
tains the regularity of u. It should be weighted against Theorem 5.1.
Theorem 5.3. Let 1 < p < ∞ and f ∈ Lp (Ω). Then, the function u
defined by (5.13) satisfies u ∈ W 2,p (Ω) ∩ W01,p (Ω) and it is the unique
solution of (5.14) in the sense of Definition 4.4. Moreover, the map
(−∆)−1
Lp (Ω) −→ W 2,p (Ω) ∩ W01,p (Ω)
(5.15)
−1
f 7→ (−∆) f := u
is a linear topological isomorphism. Therefore, there exists a constant C > 0

such that
C −1 ∥f ∥Lp (Ω) ≤ ∥u∥W 2,p (Ω) ≤ C∥f ∥Lp (Ω) (5.16)
for all f ∈ Lp (Ω).
Suppose p ≥ 2 and u ∈ W 2,p (Ω)∩W01,p (Ω) satisfies (5.14) in the sense of

Definition 4.4. Then, by Proposition 4.2, u provides us with a weak solution
of (5.14). Consequently, the uniqueness assertion in Theorem 5.3 is a by-
product of the uniqueness of the weak solution obtained from Theorem
4.9.
Suppose u ∈ W 2,p (Ω) ∩ W01,p (Ω). Then,
f := −∆u ∈ Lp (Ω)
and hence, thanks to Theorem 5.3,
( )
∥u∥W 2,p (Ω) ≤ C ∥ − ∆u∥Lp (Ω) + ∥u∥Lp (Ω) . (5.17)
This inequality extends the validity of the global elliptic estimate (5.8) to
any bounded domain Ω of class C 2 .
The proof of Theorem 5.3 is based upon the representation formula
(5.13). Unfortunately, though (5.13) should imply
∫
∂2u ∂2G
:= − (x, ·)f (x) dx in Ω
∂yi ∂yj Ω ∂yi ∂yj
for all i, j ∈ {1, ..., N }, it turns out that

∂2G
(x, ·) ̸∈ L1 (Ω),
∂yi ∂yj
as a consequence of the singularity of G at y = x. Consequently, once
again, the theory of singular integrals of A. P. Calderón and A. Zygmund
[36] is imperative to overcome such a technical difficulty in the proof of the
theorem, whose technical details are omitted here.
5.3 General elliptic Lp (Ω)-estimates when Γ1 = ∅
This section collects a general version of the global elliptic estimate (5.17).
Such estimate will show the solvability of (4.4) in Lp (Ω), under homoge-
neous Dirichlet boundary conditions, when L + ω is coercive. To state it
with the appropriate generality, consider a family A of symmetric matrices
A = (aij )1≤i,j≤N ∈ Msym
N (W
1,∞
(Ω))
with
sup ∥A∥L∞ (Ω) < ∞,
A∈A
a family B of vectors fields b ∈ (L∞ (Ω))

N
and a family C of functions
c ∈ L∞ (Ω) such that
sup ∥b∥L∞ (Ω) < ∞, sup ∥c∥L∞ (Ω) < ∞.
b∈B c∈C
Then, the next result holds.
Theorem 5.4. Let A, B, C as above and suppose that there is a constant

µ > 0 such that
ξ T A(x) ξ ≥ µ |ξ|2 (5.18)
for all A ∈ A, x ∈ Ω, and ξ ∈ RN . Then, for every 1 < p < ∞, there exists
a constant C := C(p) > 0 such that
( )
∥u∥W 2,p (Ω) ≤ C ∥Lu∥Lp (Ω) + ∥u∥Lp (Ω) (5.19)
for all u ∈ W 2,p (Ω) ∩ W01,p (Ω), a ∈ A, b ∈ B and c ∈ C, where

L := − div (A∇ · ) + ⟨b, ∇ · ⟩ + c.
Theorem 5.4 is a direct consequence from Theorem 9.14 of D. Gilbarg

and N. Trudinger [79], where we refer for the proof.
5.4 The method of continuity
It relies upon the following theorem.
Theorem 5.5 (of continuity). Let E be a Banach space, F a normed

linear space, and L0 , L1 ∈ L(E, F ) two linear continuous operators from E
into F . For every t ∈ [0, 1], set
Lt := (1 − t)L0 + tL1 (5.20)
and suppose that there is a constant C > 0 such that
∥x∥E ≤ C∥Lt x∥F (5.21)
for all t ∈ [0, 1] and x ∈ E. Then, the following conditions are equivalent:
(a) R[L0 ] = F ,
(b) R[Lt ] = F for all t ∈ [0, 1].
In other words, L0 is onto if and only if Lt is onto for all t ∈ [0, 1].
Equivalently, by the open mapping theorem, L0 ∈ Iso(E, F ) if and only
if Lt ∈ Iso(E, F ) for all t ∈ [0, 1], because (5.21) entails the injectivity of
Lt for all t ∈ [0, 1].
Proof. Obviously, (b) implies (a). Now, suppose

R[Ls ] = F
for some s ∈ [0, 1]. According to (5.21), Ls is injective and, hence, by the
open mapping theorem,
Ls ∈ Iso(E, F ).
According to (5.20), we have that
Ls x = Lt x + (Ls − Lt ) x = Lt x + (t − s) (L0 − L1 ) x
for all x ∈ E and t ∈ [0, 1]. Thus,
x = L−1 −1
s Lt x + (t − s)Ls (L0 − L1 )x
for all x ∈ E and t ∈ [0, 1]. Therefore, for any given y ∈ F , there exists
x ∈ E such that
y = Lt x
if and only if
x = L−1 −1
s y + (t − s)Ls (L0 − L1 )x,
which can be equivalently expressed as
[ ]
IE − (t − s)L−1 −1
s (L0 − L1 ) x = Ls y. (5.22)
Owing to (5.21), we have that
∥L−1
s z∥E ≤ C∥z∥F
for all z ∈ F , and hence,
( )
∥(t − s)L−1s (L0 − L1 )∥L(E,F ) ≤ |t − s|C ∥L0 ∥L(E,F ) + ∥L1 ∥L(E,F ) .
Consequently, if
1
|t − s| < δ := ( )
C ∥L0 ∥L(E,F ) + ∥L1 ∥L(E,F )
then,
IE − (t − s)L−1s (L0 − L1 ) ∈ Iso(E, E)
and we find from (5.22) that
[ ]−1 −1
x = IE − (t − s)L−1 s (L0 − L1 ) Ls y
satisfies Lt x = y. Therefore,
Lt ∈ Iso(E, F ) if |t − s| < δ.
As δ does not depend on s and we can divide the interval [0, 1] in subinter-
vals with length lesser than δ, the proof is complete.
5.5 Regularity of weak solutions when Γ1 = ∅
Throughout this section we suppose that p ∈ [2, ∞) and consider the Ba-
nach spaces
E := W 2,p (Ω) ∩ W01,p (Ω), F := Lp (Ω),
the operators
L0 := −∆ + ω : E → F, (5.23)
L1 := − div (A∇ · ) + ⟨b, ∇·⟩ + c + ω : E → F, (5.24)

and the homotopy
Lt := (1 − t)L0 + tL1 : E → F, t ∈ [0, 1], (5.25)
where ω > 0 is sufficiently large to guarantee the coercivity of the bilinear
form associated to (L1 , D, Ω). According to (5.23) and (5.24), it is apparent
that
Lt := − div ([tA + (1 − t)IRN ] ∇ · ) + ⟨tb, ∇·⟩ + tc + ω
for all t ∈ [0, 1]. Subsequently, for every t ∈ [0, 1], we consider the problem
{
Lt u = f in Ω,
(5.26)
u=0 on ∂Ω,
as well as its associated bilinear form
∫ ∫
at (u, v) := ⟨[tA + (1 − t)IRN ]∇u, ∇v⟩ + (⟨tb, ∇u⟩ + tcu + ωu) v.
Ω Ω
The next result collects some important properties of this homotopy.
Theorem 5.6. Let ω ∈ R be such that a1 is coercive and consider the

family {Lt }t∈[0,1] defined through (5.25). Let µ > 0 denote the ellipticity
constant of L1 . Then:
(a) For every t ∈ [0, 1], Lt is uniformly elliptic in Ω with constant

min{1, µ} > 0.
(b) There exists a constant α > 0 such that
|at (u, u)| ≥ α ∥u∥2W 1,2 (Ω)
0
for all u ∈ W01,2 (Ω) and t ∈ [0, 1]. In particular, the bilinear form at is
coercive for all t ∈ [0, 1].
(c) There exists a constant C > 0 such that

( )
∥u∥W 2,p (Ω) ≤ C ∥Lt u∥Lp (Ω) + ∥u∥Lp (Ω)
for all u ∈ W 2,p (Ω) ∩ W01,p (Ω) and t ∈ [0, 1].
Proof. For every t ∈ [0, 1], the matrix of the coefficients of the principal
part of Lt is given by
At := tA + (1 − t)IRN .
Obviously, for every ξ ∈ RN and t ∈ [0, 1], we have that
ξ T At ξ = tξ T Aξ + (1 − t)|ξ|2
≥ (1 − t + tµ)|ξ|2
≥ min{1, µ}|ξ|2 .
This ends the proof of Part (a).

Subsequently, we denote by αj the coercivity constant of Lj , j ∈ {0, 1}.
Obviously,
at = (1 − t)a0 (u, v) + ta1 for all t ∈ [0, 1].
Consequently,
at (u, u) = (1 − t)a0 (u, u) + ta1 (u, u)

≥ (1 − t)α0 ∥u∥2 1,2
+ tα1 ∥u∥2 1,2
W0 (Ω) W0 (Ω)
≥ min{α0 , α1 }∥u∥ 2
1,2
W0 (Ω)
for all u ∈ W01,2 (Ω) and t ∈ [0, 1]. The proof of Part (b) is complete.
Finally, thanks to Theorem 5.4, Part (c) is a direct consequence from
Part (a). This ends the proof.
Obviously, the operators Lt : E → F defined by (5.25) are linear and

continuous, as there is a constant C > 0 such that
∥Lt u∥Lp (Ω) ≤ C∥u∥W 2,p (Ω)
for all t ∈ [0, 1] and u ∈ W 2,p (Ω). The next result establishes that they are
injective.
Proposition 5.1. Lt ∈ L(E, F ) is injective for all t ∈ [0, 1].

Proof. Let u, v ∈ E such that

Lt u = Lt v
for some t ∈ [0, 1]. Then, w := u − v ∈ E satisfies Lt w = 0 in Ω and, hence,
by Proposition 4.2, w is a weak solution of
{
Lt w = 0 in Ω,
w=0 on ∂Ω.
On the( other hand,
) according to Theorems 5.6 and 4.9, for every f ∈
L (Ω) ⊂ L (Ω) and t ∈ [0, 1], the problem (5.26) has a unique weak
p 2
solution. Therefore, w = 0, and hence, u = v.

Thanks to Proposition 5.1,
Lt ∈ Iso(E, Lt (E)) for all t ∈ [0, 1].
Therefore, for every t ∈ [0, 1], there exists a constant Ct > 0 such that
∥u∥W 2,p (Ω) ≤ Ct ∥Lt u∥Lp (Ω) for all u ∈ E.
The next result shows that Ct can be chosen so that
sup Ct < ∞
t∈[0,1]
and infers a fundamental property from this estimate.
Theorem 5.7. There exists a constant C > 0 such that

∥u∥W 2,p (Ω) ≤ C∥Lt u∥Lp (Ω) (5.27)
for all u ∈ E and t ∈ [0, 1]. Therefore, according to Theorem 5.5,
R[Lt ] = R[L0 ] for all t ∈ [0, 1]. (5.28)
Proof. The proof of (5.27) proceeds by contradiction. Suppose that there

exists a sequence
( )
(tn , un ) ∈ [0, 1] × W 2,p (Ω) ∩ W01,p (Ω) \ {0} , n ≥ 1,
such that
∥un ∥W 2,p (Ω)
lim = ∞. (5.29)
n→∞ ∥Ltn un ∥Lp (Ω)
As [0, 1] is compact, we can assume, without loss of generality, that
lim tn = t∗ ∈ [0, 1].
n→∞
un
vn := , n ≥ 1. (5.30)
∥un ∥W 2,p (Ω)
Then,
vn ∈ E, ∥vn ∥W 2,p (Ω) = 1, n ≥ 1, (5.31)
and (5.29) can be expressed as
lim ∥Ltn vn ∥Lp (Ω) = 0. (5.32)
n→∞
According to (5.31) and (4.11), we can suppose, without loss of generality,

that there is v ∗ ∈ Lp (Ω) such that
lim ∥vn − v ∗ ∥Lp (Ω) = 0. (5.33)
n→∞
Owing to Theorem 5.6(c), there exists a constant C > 0 such that

( )
∥vn −vm ∥W 2,p (Ω) ≤ C ∥Ltn (vn −vm )∥Lp (Ω) +∥vn −vm ∥Lp (Ω) (5.34)
for all n, m ≥ 1. By (5.33),
lim ∥vn −vm ∥Lp (Ω) = 0.
n,m→∞
Moreover, for every n, m ≥ 1,

Ltn (vn − vm ) = Ltn vn − Ltm vm + (Ltm − Ltn ) vm , (5.35)
and, due to (5.25),
Ltm − Ltn = (tm − tn )(L1 − L0 ).
Thus, we find from (5.31) that
∥ (Ltm − Ltn ) vm ∥F ≤ |tm − tn |∥L0 − L1 ∥L(E,F )
and, therefore, it follows from (5.32) and (5.35) that
lim ∥Ltn (vn −vm )∥Lp (Ω) = 0.
n,m→∞
Consequently, by (5.34), {vn }n≥1 is a Cauchy sequence in W 2,p (Ω). Thus,

according to (5.33), v ∗ does actually belong to E and
lim ∥vn − v ∗ ∥E = 0. (5.36)
n→∞
In particular, by (5.31),
∥v ∗ ∥W 2,p (Ω) = 1. (5.37)
Obviously, (5.36) implies that
lim ∥Lj vn − Lj v ∗ ∥Lp (Ω) = 0, j ∈ {0, 1},

n→∞
and, consequently,
lim ∥Ltn vn − Lt∗ v ∗ ∥Lp (Ω) = 0. (5.38)

n→∞
Indeed, we have that
∥Ltn vn −Lt∗ v ∗ ∥Lp (Ω) ≤ ∥Ltn vn −Lt∗ vn ∥Lp (Ω) +∥Lt∗ vn −Lt∗ v ∗ ∥Lp (Ω)
for all n ≥ 1. Moreover, by (5.25) and (5.31),
lim ∥Ltn vn − Lt∗ vn ∥Lp (Ω) = 0,

n→∞
and, due to (5.36),
lim ∥Lt∗ vn − Lt∗ v ∗ ∥Lp (Ω) = 0.

n→∞
Therefore, (5.38) holds, and we conclude from (5.32) and (5.38) that
Lt∗ v ∗ = 0.
Consequently, thanks to Proposition 5.1, v ∗ = 0. This contradicts (5.37)

and concludes the proof.
A further application of Theorem 5.5 to the operators
M0 := −∆ : E → F,
M1 := −∆ + ω : E → F,
through the homotopy
Mt := (1 − t)M0 + tM1 , t ∈ [0, 1],
reveals that
R[L0 ] = R[−∆ + ω] = R[−∆].
Therefore, according to Theorems 5.3 and 5.7, we find that
R[Lt ] = Lp (Ω) for all t ∈ [0, 1], (5.39)
and, consequently, the next fundamental result holds.

Theorem 5.8 (of regularity of weak solutions). Let ω ∈ R be such

that a1 is coercive. Then, for every p ∈ [2, ∞) and f ∈ Lp (Ω), the unique
weak solution of
{
(L + ω)u = f in Ω,
(5.40)
u=0 on ∂Ω,
satisfies
u ∈ W 2,p (Ω) ∩ W01,p (Ω). (5.41)
Therefore, by Proposition 4.2, u solves (5.40) in the sense of Definition 4.4.
−
Furthermore, if f ∈ L∞ (Ω), then u ∈ W 2,∞ (Ω), where
− ∩
W 2,∞ (Ω) := W 2,p (Ω)
p>1
and, consequently,
−
i) u ∈ C01 (Ω̄) ∩ C 1,1 (Ω̄).
ii) u is twice classically differentiable almost everywhere in Ω.
iii) u is a classical solution of (5.40), in the sense of Definition 4.1.
Proof. By Theorem 4.9, the problem (5.40) has a unique weak solution.
According to (5.39), there exists u ∈ W 2,p (Ω)∩W01,p (Ω) satisfying (5.40) in
the sense of Definition 4.4. By Proposition 4.2, u must be a weak solution
of (5.40). Therefore, the unique weak solution of (5.40) satisfies (5.41) and
it solves (5.40) in the sense of Definition 4.4.
The remaining assertions of the theorem are direct consequences from
Theorems 4.2(ii), 4.4 and 4.6. Indeed, let f ∈ L∞ (Ω). Then, the unique
weak solution of (5.40) satisfies
∞ [
∩ ] −
u∈ W 2,p (Ω) ∩ W01,p (Ω) ⊂ W 2,∞ (Ω).
p=2
On the other hand, by Theorem 4.2(ii),

N
W 2,p (Ω) ⊂ C 1,1− p (Ω̄)
for all p > N . Thus,
− ∩ N −
u ∈ W 2,∞ (Ω) ⊂ C 1,1− p (Ω̄) = C 1,1 (Ω̄).
p>N
Moreover, by Theorem 4.4, u is twice classically differentiable a.e. in Ω and

the classical derivative Dα u equals the corresponding weak derivative a.e.
in Ω for all multi-index α ∈ NN with |α| ≤ 2. Consequently,
(L + ω)u = f almost everywhere in Ω
in the classical sense, as u solves (5.40) in the sense of Definition 4.4.

Finally, since u ∈ C(Ω̄), Theorem 4.6 implies that
0 = T∂Ω u = u|∂Ω
and, hence, u is a classical solution of (5.40), as discussed by Definition 4.1.
This ends the proof.
5.6 A first glance to the general case when Γ1 ̸= ∅
As for the existence of weak solutions in Section 4.6, the regularity of the
weak solutions of (4.4) in the general case when β changes of sign can
be easily derived from Theorem 4.10 through the regularity result for the
special case when
β ≥ 0. (5.42)
Therefore, (5.42) will be assumed throughout this section.
When Γ1 ̸= ∅, Theorem 5.5 cannot be applied straightaway to get the
W (Ω)-regularity of the weak solutions of (4.4) for f ∈ Lp (Ω), p ≥ 2,
2,p
as we have done in the previous section for the special case when B = D
(Γ1 = ∅), because the method of continuity now involves a homotopy of
both differential operators and boundary operators on Γ1 and, as a result,
the Banach space E where the operators Lt , t ∈ [0, 1], are defined depends
on t. This technical difficulty makes the analysis of the underlying regu-
larity problem extraordinarily more involved than in case Γ1 = ∅, however
Theorem 5.5 also applies to get the regularity result in the general case. As
giving all the technical details of the proofs of the necessary results to get
the W 2,p (Ω)-regularity of the weak solutions of (4.4) in the general case
when Γ1 ̸= ∅ lies outside the scope of this book, in this section we will
restrict ourselves to sketch an approach to this problem and to state the
main result which will be used later.
Subsequently, we fix p ∈ [2, ∞) and consider, for every t ∈ [0, 1], the
boundary operator
Bt : W 2,p (Ω) → Lp (Γ1 )
defined through
Bt u := ⟨∇u, tAn + (1 − t)n⟩ + βu, u ∈ W 2,p (Ω), (5.43)
in the sense of traces. According to Theorem 4.6,
( )
Bt ∈ L W 2,p (Ω), Lp (Γ1 ) for all t ∈ [0, 1].
Thus,
N [Bt ] = B−1
t (0)
is a closed linear subspace of W 2,p (Ω) and hence, we can consider the curve
of Banach subspaces of W 2,p (Ω) defined by
Et := W 2,p (Ω) ∩ WΓ1,p
0
(Ω) ∩ N [Bt ], t ∈ [0, 1], (5.44)
as well as the associated homotopy
Lt : Et → F := Lp (Ω)
defined by
Lt u := −div ([tA + (1 − t)IRN ]∇u) + t⟨b, ∇u⟩ + tcu + ωu, (5.45)
for all u ∈ Et and t ∈ [0, 1]; Lt is well defined in W 2,p
(Ω) and
Lt u = tL1 u + (1 − t)L0 u (5.46)
for all u ∈ W (Ω) and t ∈ [0, 1]. The homotopy Lt : Et → F between
2,p
L0 : E0 → F and L1 : E1 → F establishes a natural deformation between

the solutions of (4.4) and the solutions of

 (−∆ + ω)u = f in Ω,
u=0 on Γ0 , (5.47)
 ∂u
∂n + βu = 0 on Γ 1 .
By (5.42), it follows from Theorem 4.8 that there exists ω0 > 0 such that
the bilinear forms associated to L0 and L1 are coercive for all ω ≥ ω0 .
Throughout the rest of this section we will assume that ω ≥ ω0 .
According to L. N. Slobodeckii [205], F. E. Browder [30–32], S. Agmon,
A. Douglis, and L. Nirenberg [4] (see Section 7 of H. Amann [9] too), it is
well known that
R[L0 ] = Lp (Ω). (5.48)
Moreover, the next result holds. Essentially, it establishes the existence of
a uniformly bounded family of isomorphisms between every pair of Et ’s.
Theorem 5.9. For every t ∈ [0, 1], there exists Φt ∈ Iso(E0 , Et ) such that
Φ0 = I, the map
[0, 1] −→ L(E0 , W 2,p (Ω))
t 7→ Φt
is continuous, and there is a constant C > 0 such that
∥Φ−1
t ∥L(Et ,E0 ) ≤ C for all t ∈ [0, 1]. (5.49)
Proof. By (5.43), we have that

B0 = Bt − t (B1 − B0 ) in W 2,p (Ω) ∩ WΓ1,p
0
(Ω) (5.50)
for all t ∈ [0, 1]. By some advanced results in trace theory (e.g., R. A.
Adams [1], P. L. Butzer and H. Berens [34], J. Nec̆as [170], H. Triebel [220]
and J. Wloka [226]), we have that
R[Bt ] = W 1−1/p,p (Γ1 ) ∀ t ∈ [0, 1],
and, in particular,
R[Bt ] = R[B0 ] for all t ∈ [0, 1]
(see Remark 5.1). By some sophisticated inverse trace theorems (see, e.g.,
Lemma 5.1 of H. Amann [9], Lemma II.5.8 of J. Nec̆as [170], and J. L.
Lions [129]), it is apparent that N [Bt ] possesses a topological complement
in W 2,p (Ω) ∩ WΓ1,p0
(Ω) for all t ∈ [0, 1]. Equivalently, Bt admits a right
inverse for all t ∈ [0, 1]. In other words, there exists a linear continuous
operator
Rt : R[Bt ] → W 2,p (Ω) ∩ WΓ1,p
0
(Ω)
such that
Bt Rt q = q for all q ∈ R[Bt ].
Consequently, (5.50) can be written in the form
B0 = Bt [I − tRt (B1 − B0 )] in W 2,p (Ω) ∩ WΓ1,p
0
(Ω)
and, hence, for every t ∈ [0, 1], we find that
u ∈ E0 ⇐⇒ u − tRt (B1 − B0 ) u ∈ Et .
Therefore, the most reasonable choice for Φt is
Φt u := u − tRt (B1 − B0 ) u, u ∈ E0 , t ∈ [0, 1]. (5.51)
The most delicate part of the proof of Theorem 5.9 is the construction of the
right inverses Rt so that the operators (5.51) satisfy (5.49). The technical
details are left outside of this book.
As W 2,p (Ω) ∩ WΓ1,p
0
(Ω) is a u.c. B-space, according to Theorem 3.10,
there exists a continuous projection PN [Bt ] of W 2,p (Ω) ∩ WΓ1,p
0
(Ω) on N [Bt ].
Unfortunately, it is unknown whether PN [Bt ] is a linear operator and, con-
sequently, in the proof of Theorem 5.9, one must invoke to an inverse trace
theorem to get the existence of a linear continuous projection on N [Bt ].
Next, we will use Theorem 5.9 to show how (5.48) implies that
R[Lt ] = Lp (Ω) for all t ∈ [0, 1]. (5.52)
Adapting the proof of Proposition 5.1, it becomes apparent that
Lt : Et → Lp (Ω)
is injective for all t ∈ [0, 1]. Thus, there exists a constant Ct > 0 such that
∥xt ∥W 2,p (Ω) ≤ Ct ∥Lt xt ∥Lp (Ω) for all xt ∈ Et , 0 ≤ t ≤ 1.
Much within the spirit of Theorems 5.4 and 5.7, the constant Ct can be
chosen so that
sup Ct < ∞.
t∈[0,1]
Indeed, the next result holds.
Theorem 5.10. There is a constant C > 0 such that

∥xt ∥W 2,p (Ω) ≤ C∥Lt xt ∥Lp (Ω) (5.53)
for all t ∈ [0, 1] and xt ∈ Et .
The proof of Theorem 5.10 combines the global interior elliptic estimates
of the Appendix of H. Amann and J. López-Gómez [13] with the boundary
estimates of H. Amann [12] (see Remark A3.2 of [13]). The details of the
proof are not included.
Subsequently, for every t ∈ [0, 1], we denote by Mt the operator
Mt = Lt Φt : E0 → Lp (Ω).
By construction, Mt ∈ L(E0 , Lp (Ω)) and it is injective for all t ∈ [0, 1].
Also, since Φ0 = I, we find from (5.48) that
R[M0 ] = R[L0 ] = Lp (Ω),
and hence, by the open mapping theorem,
M0 ∈ Iso(E0 , Lp (Ω)).
Moreover, according to Theorem 5.9 and (5.53), there exists a constant
C > 0 such that
∥x∥E0 ≤ C∥Mt x∥Lp (Ω) (5.54)
for all t ∈ [0, 1] and x ∈ E0 . More generally, suppose that
Ms ∈ Iso(E0 , Lp (Ω))
for some s ∈ [0, 1]. Then,

Ms x = Mt x + (Ms − Mt ) x
for all t ∈ [0, 1] and x ∈ E0 . On the other hand,
Ms − Mt = Ls Φs − Lt Φt
= (Ls − Lt ) Φs + Lt (Φs − Φt )
= (s − t)(L1 − L0 )Φs + Lt (Φs − Φt ) .
Thus, for a given y ∈ Lp (Ω), there exists x ∈ E0 such that Mt x = y if and
only if
[Ms − (s − t)(L1 − L0 )Φs − Lt (Φs − Φt )] x = y. (5.55)
As the map t 7→ Φt is uniformly continuous, there exists δ > 0 such that
|t − s| ≤ δ implies
1
∥(s − t)(L1 − L0 )Φs + Lt (Φs − Φt ) ∥ ≤ ,
2C
where C is the constant of (5.54). As, due to (5.54), ∥M−1 s ∥ ≤ C, for such
a choice of δ, the operator on the left-hand side of (5.55) is invertible and,
therefore, there is a unique x satisfying (5.55) whenever |s−t| ≤ δ. As [0, 1]
can be divided into a finite number of subintervals with length less than δ,
R[Mt ] = Lp (Ω) for all t ∈ [0, 1].
Consequently, (5.52) holds. As a by-product, the next counterpart of The-
orem 5.8 holds. Note that, thanks to (5.42) and Theorem 4.8, there exists
ω0 such that the bilinear form a1 associated to L1 is coercive for all ω ≥ ω0 .
Theorem 5.11 (of regularity of weak solutions). Let ω ∈ R be such

that a1 is coercive. Then, for every p ≥ 2 and f ∈ Lp (Ω), the unique weak
solution of
{
(L + ω)u = f in Ω,
(5.56)
Bu = 0 on ∂Ω,
satisfies
u ∈ W 2,p (Ω) ∩ WΓ1,p
0
(Ω) ∩ N [B1 ].
Therefore, by Proposition 4.2, u solves (5.56) in the sense of Definition 4.4.
−
Furthermore, if f ∈ L∞ (Ω), then u ∈ W 2,∞ (Ω) and, consequently,
−
i) u ∈ CΓ1 0 (Ω̄) ∩ C 1,1 (Ω̄).
ii) u is twice classically differentiable almost everywhere in Ω.

iii) u is a classical solution of (5.56), in the sense of Definition 4.1.
In the statement of Theorem 5.11, we are denoting by CΓ1 0 (Ω̄) the closure
of CΓ∞0 (Ω) in C 1 (Ω̄) (see Section 4.1.1, if necessary).
It should be noted that Theorem 5.11 also follows from the abstract
theory of R. Denk, M. Hieber and J. Prüss [51].
Far reaching results about the Weyl lemma were given by L. Hörmander
[105].
Section 5.1 was inspired by Chapter 1 of D. R. Adams and L. I. Hedberg
[2]. The Bessel potential spaces were systematically studied by N. Aron-
szajn and coworkers (see N. Aronszajn and K. T. Smith [15], N. Aronszajn
F. Mulla and P. Szeptycki [16], and N. Aronszajn [14]). As already com-
mented in Chapter 4, specialized monographs and textbooks discussing the
Sobolev–Slobodeskii spaces are those of R. A. Adams [1], A. Friedman [67],
J. L. Lions and E. Magenes [130], P. Malliavin [156], V. G. Maz’ja [159],
C. B. Morrey [167], J. Nec̆as [170], H. Triebel [220], and J. Wloka [226],
among others.
Section 5.2 is based on Chapter 4 of F. John [108], Theorem 9.9 of D.
Gilbarg and N. Trudinger [79], and Theorem IX.32 of H. Brézis [29]. Section
5.3 provides a statement of Theorem 9.14 of D. Gilbarg and N Trudinger
[79].
Section 5.4 is based on Section 5.2 of D. Gilbarg and N Trudinger [79],
however the proof of Theorem 5.2 of [79] is wrong, because the map
T x := L−1 −1
s y + (t − s)Ls (L0 − L1 )x
defined on p. 75 of [79] is far from being contractive if

|s − t| < δ = [C∥L0 ∥ + ∥L1 ∥]−1 ,
as claimed by D. Gilbarg and N Trudinger therein. Theorem 5.4 does
actually ascertain whether the interval [L0 , L1 ] of L(E, F ) satisfies
[L0 , L1 ] ⊂ Iso(E, F ).
This is a very classical result, which might be attributed to S. Banach
[22]. As already mentioned in the presentation of this chapter, Section 5.5
consists of the resolution of Problem 9.8 of [79].
The underlying ideas outlined in Section 5.6 might be considerably ex-

panded up to span a more specialized monograph on regularity of weak
solutions. Within the context of Hölder regularity theory, the method of
continuity for rather general classes of mixed boundary value problems had
already been used by D. Gilbarg and N Trudinger [79] in the proof of The-
orem 6.31 of [79].
Besides D. Gilbarg and N Trudinger [79], other classical texts covering
the most fundamental features about the regularity of the weak solutions
of elliptic partial differential equations are those of L. Hörmander [106]
and Morrey [167]. Some optimal Lp a priori estimates are available in R.
Denk, M. Hiebber and J. Prüss [50, 51], where the reader should refer to
for further details.

Chapter 6
The Krein–Rutman theorem
Essentially, the theorem of M. G. Krein and M. A. Rutman culminated a

series of generalizations of the celebrated results of O. Perron [175] and G.
Frobenius [71, 73] about the spectrum of matrices with non-negative en-
tries, establishing that the spectral radius of an n × n matrix with positive
entries is an eigenvalue with a component-wise positive eigenvector in Rn .
Essentially, the Krein–Rutman theorem extended the Perron–Frobenius
theorem to deal with strongly positive compact endomorphisms in ordered
Banach spaces.
For a historical account and a bibliography of earlier work on compact
positive operators, the reader should refer to the original monograph of M.
G. Krein and M. A. Rutman [120].
Not surprisingly, compactness and positivity are crucial in all these ex-
tensions. Basically, the compactness localizes the spectrum and provides
with the Fredholm alternative, while the positivity entails the spectral ra-
dius of the operator to be the unique eigenvalue on the spectral circle. Com-
bining both, compactness and positivity, provides with the convergence of
the monotone schemes approximating the spectral radius and its associated
positive eigenvector.
6.1 Orderings. Ordered Banach spaces
Given a nonempty set X, an ordering in X is a relation in X which is

reflexive, transitive and anti-symmetric. In this book, these relations are
denoted by ≤. A nonempty set X together with an ordering, (X, ≤), will be
called an ordered set. Given an ordered set (X, ≤) and a point (x, y) ∈ X 2 ,
it is said that y ≥ x if x ≤ y, while we write x < y (or, equivalently, y > x)
155
if x ≤ y and x ̸= y, The set

[x, y] := { z ∈ X : x ≤ z ≤ y }
is called the order interval defined by x and y. Note that [x, y] ̸= ∅ if, and
only if, x ≤ y.
Let B be a subset of X. Then,
• B is said to be bounded if B ⊂ [x, y] for some x, y ∈ X.
• B is said to be convex if [x, y] ⊂ B for every x, y ∈ B.
Given two arbitrary ordered sets (X, ≤) and (Y, ≤) (both ordering are de-
noted by ≤) and a map f : X → Y , then
• f is said to be increasing if f (x) ≤ f (y) for every x, y ∈ X with x ≤ y.
• f is said to be strictly increasing if f (x) < f (y) for every x, y ∈ X with
x < y.
• f is said to be decreasing if f (x) ≥ f (y) for every x, y ∈ X with x ≤ y.
• f is said to be strictly decreasing if f (x) > f (y) for every x, y ∈ X with
x < y.
Given a real vector space V , an ordering ≤ in V is said to be linear if
x, y ∈ V, x≤y =⇒ x + z ≤ y + z and λx ≤ λy
for all z ∈ V and
λ ∈ R+ := [0, ∞) ⊂ R.
A real vector space V together with a linear ordering, (V, ≤), will be called
an ordered vector space (OVS).
Given a real vector space V , a subset P ⊂ V is said to be a cone if
P + P ⊂ P, R+ P ⊂ P, and P ∩ (−P ) = {0}. (6.1)
According to (6.1), every cone is convex. When (V, ≤) is an OVS, then it
is easy to see that the set P defined through
P := { x ∈ V : x ≥ 0 }
is a cone, which is referred to as the positive cone associated to the ordering
≤. Conversely, given a real vector space V together with a cone P ⊂ V ,
the relation defined by
x≤y if and only if y−x∈P (6.2)
is a linear ordering in V whose associated positive cone is P . Actually, this
correspondence establishes a bijection between the set of linear orderings
The Krein–Rutman theorem 157
of a real vector space and its set of cones, in such a way that each cone is
the positive cone of its induced linear ordering through (6.2). The vectors
of the punctate cone
Ṗ := P \ {0} = { x ∈ V : x > 0 }
are called positive.
Let E = (E, ∥ · ∥) be a Banach space ordered by a cone P through (6.2).
Then,
E := (E, ∥ · ∥, P )
is said to be an ordered Banach space (OBS) if P is closed, i.e., if
P̄ = P.
In such case, we will denote by
◦
P := int P
the interior of the cone P , which might be empty.
Throughout the rest of this book, given an OBS, (E, ∥ · ∥, P ), with
int P ̸= ∅, and x, y ∈ E, it is said that x ≪ y, or, equivalently, y ≫ x, when
◦
y − x ∈ P.
In particular,
◦
x≫0 if and only if x ∈ P .
The following results will be extremely useful later.
◦
Lemma 6.1. Let (E, ∥ · ∥, P ) be an OBS with P ̸= ∅. Then,
x ≫ 0, ρ > 0 =⇒ ρx ≫ 0.
Proof. Suppose x ≫ 0 and ρ > 0. Then, there exists ϵ > 0 such that
x+y ∈P if ∥y∥ ≤ ϵ.
Thus, owing to (6.1),
ρ(x + y) ∈ P if ∥y∥ ≤ ϵ,
and hence,
ρx + z ∈ P if ∥z∥ ≤ ρϵ.
Therefore, ρx ≫ 0, because ρϵ > 0.
◦
Lemma 6.2. Let (E, ∥ · ∥, P ) be an OBS with P ̸= ∅. Then,
E = P − P := { x − y : x, y ∈ P }. (6.3)
◦
Proof. Fix y ∈ P , and let x ∈ E arbitrary. Then, by Lemma 6.1,
( )
z := λy − x = λ y − λ−1 x ∈ P
for sufficiently large λ > 0. Therefore,
x = λy − z with z ≥ 0 and λy ≫ 0,
Definition 6.1 (generating and total cones). Let (E, ∥ · ∥, P ) be an

OBS. Then:
• P is said to be generating if E = P − P .
• P is said to be total if the set of all finite linear combinations of its
elements is dense in E.
Obviously, P is total if it is generating.
According to Definition 6.1, Lemma 6.2 can be equivalently stated by

simply saying that P is generating if it has nonempty interior.
Total cones induce a canonical dual ordering in the dual space
E ′ := L(E, R).
Indeed, given an OBS (E, ∥ · ∥, P ), let P ∗ denote the subset of E ′ defined

through
P ∗ := { x′ ∈ E ′ : x′ (x) ≥ 0 ∀ x ∈ P }. (6.4)
By definition,
P∗ + P∗ ⊂ P∗ and R+ P ∗ ⊂ P ∗ . (6.5)
Moreover, the following result holds.
Lemma 6.3. P ∗ is a cone of E ′ if and only if P is total. In such case, P ∗

is said to be the dual cone of P and
E ′ := (E ′ , ∥ · ∥E ′ , P ∗ )
is an OBS — the dual OBS of E.

Proof. According to (6.5), to prove the first assertion it suffices to show

that
P ∗ ∩ (−P ∗ ) = {0} if and only if P is total.
Indeed, since
P ∗ ∩ (−P ∗ ) = {x′ ∈ E ′ : x′ (x) = 0 ∀ x ∈ P },
P ∗ ∩ (−P ∗ ) = {0}
if and only if
x′ ∈ E ′ and x′ (x) = 0 ∀ x∈P imply x′ = 0.
As x′ is linear and continuous, it is obvious that this occurs if the set of all
finite linear combinations of elements of P is dense in E. Consequently, P ∗
is a cone if P is total.
Now, suppose that P is not total. Then, the linear manifold F generated
by all finite linear combinations of vectors of P satisfies
F̄ ̸= E.
Thus, according to the Hahn–Banach theorem (see, e.g., Corollary 1.8 of
H. Brézis [29]), there exists x′ ∈ E ′ \ {0} such that x′ (x) = 0 for all x ∈ F .
In particular, x′ (x) = 0 for all x ∈ P , and hence,
x′ ∈ P ∗ ∩ (−P ∗ ).
Consequently, P ∗ is not a cone. Therefore, P ∗ is a cone if and only if P is
total.
To conclude the proof of the lemma, it remains to show that P ∗ is closed
in E ′ . Indeed, let {x′n }n≥1 be a sequence in P ∗ ⊂ E ′ such that
lim ∥x′n − x′ ∥E ′ = 0
n→∞
for some x′ ∈ E ′ . Then,

lim x′n (x) = x′ (x)
n→∞
for all x ∈ E. Moreover, as x′n ∈ P ∗ for all n ≥ 1, we have that x′n (x) ≥ 0
for all x ∈ P and n ≥ 1. Consequently, x′ (x) ≥ 0 for all x ∈ P and,
therefore, x′ ∈ P ∗ . The proof is complete.
As a consequence from Lemma 6.3, P ∗ is a cone if P is generating.
According to Lemma 6.2, this occurs if P has non-empty interior. The
following concept will play an important role later.
Definition 6.2 (normal cone). Let (E, ∥·∥, P ) be an OBS with total pos-
itive cone P . Then, P is said to be normal if the cone P ∗ is generating.
Thanks to Lemma 6.3, the set P ∗ is a cone if P is total. Due to Definition

6.2, P is normal if it is total and
E′ = P ∗ − P ∗.
According to the first result of Section V.3.4 and Corollary V.3.5 of H. H.
Schaefer [194], within the context of ordered Banach spaces the concept
of normal cone introduced by Definition 6.2 coincides with the concept
introduced on p. 215 of H. H. Schaefer [194]. Thus, by the first theorem of
Section V.3.1 of H. H. Schaefer [194], the positive cone P is normal if and
only if the norm ∥ · ∥ of E is equivalent to some monotone norm ||| · |||. By
monotone, it means that
0≤x≤y implies |||x||| ≤ |||y|||.
More generally, the following characterization holds.
Theorem 6.1 (of characterization of normal cones). Let (E, ∥ · ∥, P )

be an ordered Banach space. Then, the following properties are mutually
equivalent:
(a) P is a normal cone.
(b) The norm ∥ · ∥ is semi-monotone, i.e., there exists a constant δ > 0
such that
0≤x≤y implies ∥x∥ ≤ δ ∥y∥.
(c) There exists an equivalent monotone norm for E.
(d) For every x, y ∈ P , the interval [x, y] is bounded.
(e) The order convex hull of every bounded set is bounded.
(f) There exists a positive real number α > 0 such that
}
x, y ∈ P
=⇒ ∥x + y∥ ≥ α.
∥x∥ = ∥y∥ = 1
Geometrically, when E is a Hilbert space, Theorem 6.1(f) entails that the

angle between two arbitrary positive unit vectors must be bounded away
from π. So, normal cones cannot be too large.
The proof of Theorem 6.1 can be accomplished from Theorem 5.1 of H.
Amann [8], M. A. Krasnoselskij [118], H. H. Schaefer [194], and G. Jameson
[107]. The technical details of the proof are omitted here.
6.2 Spectral theory of linear compact operators
This section collects a series of fundamental results about linear compact

operators that will be used in the proof of the Krein–Rutman theorem.
The proofs of these results will not be given here, as they can be easily
reconstructed from available monographs on linear operator theory.
Let (E, ∥·∥) be a real Banach space. Then, for every operator T ∈ L(E),
the following limit exists:
1/n
spr (T ) := lim ∥T n ∥L(E) (6.6)
n→∞
and it is called the spectral radius of T . Clearly,
spr T ≤ ∥T ∥L(E)
because
∥T n ∥L(E) ≤ ∥T ∥nL(E) , n ≥ 1.
Moreover, for every ζ ∈ C with |ζ| > spr T , we have that
T
spr <1
ζ
and hence the resolvent operator
( )−1
R(ζ; T ) := (ζI − T )−1 = ζ −1 I − ζ −1 T (6.7)
is given through the series
∞
∑
R(ζ; T ) := ζ −(n+1) T n (6.8)
n=0
which converges uniformly on compact subsets of |ζ| > spr T (see Theorem
3 on p. 211 of Section VIII.2 of K. Yosida [227]).
Subsequently, when ζ ∈ C, in the expression ζI − T , I stands for the
identity map of EC , the canonical complexification of E,
EC := E + iE,
where i is the complex imaginary unit, and T denotes the canonical exten-
sion of T to EC , i.e.,
T (x + iy) := T x + iT y, x, y ∈ E.
For a given T ∈ L(E), we denote by σ(T ) the spectrum of T . It consists
of the set of values λ ∈ C for which λI − T is not an isomorphism of EC .
Also, we denote by ϱ(T ) the resolvent set of T , which is defined by
ϱ(T ) := C \ σ(T ).
It is well known that
spr T = sup |λ| (6.9)

λ∈σ(T )
(see Theorem 4 on p. 212 of Section VIII.2 of K. Yosida [227]).

Throughout the rest of this chapter, for any real Banach space E and
T ∈ L(E), we denote by T ∗ the adjoint operator of T , which is defined as
the unique operator T ∗ ∈ L(E ′ ) such that
T ∗ x′ (x) = x′ (T x) for every (x, x′ ) ∈ E × E ′ .
An operator T ∈ L(E) is said to be compact if T (B) is a compact subset

of E for every bounded subset B ⊂ E. The class of compact operators of
L(E) will be denoted throughout by K(E). By a celebrated result of J.
Schauder,
T ∈ K(E) if and only if T ∗ ∈ K(E ′ ) (6.10)
(see Theorem of Schauder on p. 282 of Section X.4 of K. Yosida [227]).

Moreover, in such case, by the Riesz–Schauder Theory, we have that, for
every ζ ∈ C \ {0},
R[ζI − T ] is closed, dimN [ζI − T ] < ∞,
and
 ∗ ⊥
 R[ζI − T ] = N [ζI − T ] ,
(6.11)

dim N [ζI − T ] = dim N [ζI − T ∗ ],
where, for every M ⊂ E ′ , we have denoted
M ⊥ := { x ∈ E : x′ (x) = 0 ∀ x′ ∈ M }
(see Section X.5 of K. Yosida [227]). The following result collects some of
the most fundamental spectral properties of the compact operators.
Theorem 6.2 (of spectral structure). Suppose T ∈ K(E). Then, the

following properties hold:
(a) 0 ∈ σ(T ) if dim E = ∞, σ(T ) \ {0} is a discrete set, and there exists
λ ∈ σ(T ) such that |λ| = spr T . Therefore, by (6.9),
spr T = max |λ|. (6.12)

λ∈σ(T )
(b) For every λ ∈ σ(T ) \ {0}, λ is an eigenvalue of T and there exists a

minimal integer number ν(λ) ≥ 1 such that
N [(λI − T )ν(λ)+n ] = N [(λI − T )ν(λ) ]
for all n ∈ N. The integer ν(λ) ≥ 1 is the algebraic ascent of λ, the
linear space N [(λT − T )ν(λ) ] is the ascent generalized eigenspace
associated to λ, and the algebraic multiplicity of λ is defined through
m(λ; T ) := dim N [(λI − T )ν(λ) ].
An eigenvalue λ ∈ σ(T ) \ {0} is said to be algebraically simple if
ν(λ) = m(λ; T ) = 1.
(c) The resolvent operator
ϱ(T ) −→ L(E)
ζ 7→ R(ζ; T )
is holomorphic. Moreover, for each λ ∈ σ(T ) \ {0}, λ is a pole of
order ν(λ) of R(ζ; T ). Therefore, it possesses a (unique) Laurent
development at λ,
∞
∑
R(ζ; T ) = (ζ − λ)n Tn , ζ ∼ λ, ζ ̸= λ, (6.13)
n=−ν(λ)
where, for every n ≥ −ν(λ) and sufficiently small ϵ > 0,

∫
1
Tn := (ζ − λ)−(n+1) R(ζ; T ) dζ ∈ L(E). (6.14)
2πi |ζ−λ|=ϵ
Furthermore, T−ν(λ) = ̸ 0, Tn is a finite rank operator for every n ∈
{−ν(λ), ..., −1}, and T−1 is a projection of E onto N [(λI − T )ν(λ) ].
Consequently,
tr T−1 := dim R[T−1 ] = dim N [(λI − T )ν(λ) ] = m(λ; T ).
The proof of Theorem 6.2(a) can be found in Theorem VI.8 of H. Brézis

[29]. The finite-dimensional counterpart of Theorem 6.2(b) is folklore, but,
surprisingly, it might not be trivial to give a reference containing a self-
contained proof of it in the general infinite-dimensional setting. Part (c) is
a more specialized material, though, naturally, it is folklore for experts. A
self-contained proof of Theorem 6.2(b)(c) may be built up from the results
of Chapter I of T. Kato [112], through a preliminary reduction of dimension,
and the results of Section VIII.8 of K. Yosida [227]. The reader should
refer to Chapters 6 and 7 of [141] and Chapter 7 of [142] for the finite-
dimensional theory, and to the more recent monograph of J. López-Gómez
and C. Mora-Corral [153] for the general infinite-dimensional case, where,
besides a complete self-contained proof of Theorem 6.2, the reader will
find some very recent advances in the theory of algebraic multiplicities of
eigenvalues of linear operators, as the axiomatization theorem. Classical
textbooks covering most of these materials, easily accessible to beginners,
are I. C. Göhberg, P. Lancaster and L. Rodman [84], and I. C. Göhberg, S.
Goldberg and M. A. Kaashoek [82].
By (6.10) and (6.11), it becomes apparent from Theorem 6.2 that
σ(T ) = σ(T ∗ )
and that
m(λ; T ) = m(λ; T ∗ ), λ ∈ σ(T ) \ {0}, (6.15)
for all T ∈ K(E). Moreover, the following fundamental identity, going back
to R. S. Phillips [177], holds
R(ζ; T )∗ = R(ζ; T ∗ ), ζ ∈ ϱ(T ) = ϱ(T ∗ ), (6.16)
(see Theorem 2 of Section VIII.6 of K. Yosida [227]).
6.3 The Krein–Rutman theorem
Let (X, PX ) and (Y, PY ) be two ordered vector spaces and T : X → Y a

linear operator. Then,
• T is said to be positive if
T (PX ) ⊂ PY .
• T is said to be strictly positive if
T (PX \ {0}) ⊂ PY \ {0}.
• T is said to be strongly positive if Y possesses a norm ∥ · ∥Y for which
(Y, ∥ · ∥Y , PY ) is an ordered Banach space with
◦ ◦
P Y ̸= ∅ and T (PX \ {0}) ⊂ P Y .
Naturally, T is said to be strongly negative if −T is strongly positive.
The main result of this chapter can be stated as follows.

Theorem 6.3 (generalized of Krein–Rutman). Let (E, ∥ · ∥, P ) be an

◦
ordered Banach space with P ̸= ∅, and T ∈ K(E) a compact strongly positive
operator, i.e.,
◦
T (P \ {0}) ⊂ P . (6.17)
Then, the following assertions are true:
(a) spr T > 0 is an algebraically simple eigenvalue of T with
N [spr T I − T ] = span [x0 ]
◦
for some x0 ∈ P .
(b) spr T is the unique real eigenvalue of T to an eigenvector in P \ {0}.
(c) spr T is the unique eigenvalue of T in the spectral circle
|ζ| = spr T.
In other words,
|λ| < spr T for all λ ∈ σ(T ) \ {spr T }.
(d) For every real number λ > spr T , the resolvent operator
R(λ; T ) := (λI − T )−1 ∈ L(E)
is strongly positive, i.e.,
◦
R(λ; T )(P \ {0}) ⊂ P .
(e) There exist ϵ > 0 and x > 0 such that
R(λ; T )x ≪ 0 for all λ ∈ (spr T − ϵ, spr T ).
(f) Suppose, in addition, that P is a normal cone. Then, the following
assertions are true:
(a) There exists x′0 ∈ P ∗ \ {0} such that
x′0 (x0 ) > 0 and N [spr T I − T ∗ ] = span [x′0 ].
(b) For every x ∈ P \ {0},
x′0 (x) > 0
and the equation
spr T u − T u = x
cannot admit a solution u ∈ E.
Remark 6.1.
(a) By the general properties of adjoint operators revisited in Section 6.2,

it is apparent that T ∗ ∈ K(E ′ ) satisfies
σ(T ) = σ(T ∗ ) and spr T = spr T ∗ .
Thus, by (6.11) and Theorem 6.3(a)(c), spr T must be a simple eigen-
value of T ∗ , and it is the unique eigenvalue of T ∗ in the spectral circle
|ζ| = spr T .
(b) According to Lemma 6.2, P is generating and, hence, it is total, because
intP ̸= ∅. Thus, by Lemma 6.3, (E ′ , ∥·∥E ′ , P ∗ ) is an OBS. By definition,
for every x ∈ P and x′ ∈ P ∗ , the following estimate holds
T ∗ x′ (x) = x′ (T x) ≥ 0
and hence T ∗ x′ ∈ P ∗ . Therefore, T ∗ is positive with respect to the
cone P ∗ . But T ∗ might not be strongly positive. Actually, it is far
from obvious whether P ∗ has empty interior, though we are assuming
that P ∗ is generating, because P is normal in Part (f).
6.4 Preliminaries of the proof of Theorem 6.3
This section collects some important properties needed in the proof of The-
orem 6.3. Therefore, throughout it, we will assume that (E, ∥ · ∥, P ) is an
OBS with int P ̸= ∅, and that T ∈ K(E) is strongly positive. Also, we will
consider the sets
B := { x ∈ E : ∥x∥ < 1 }, ∂B := { x ∈ E : ∥x∥ = 1 },
K := T (∂B ∩ P ),
and, for every γ > 0,
Mγ := { x ∈ P \ {0} : T x ≥ γx }
= (T − γI)−1 (P ) ∩ (P \ {0}),
and
Σγ := { x ∈ K ∩ Mγ : ∥x∥ ≥ γ }
= K ∩ (T − γI)−1 (P ) ∩ P ∩ { x ∈ E : ∥x∥ ≥ γ }.
The following result collects some useful properties of these sets which are
going to be used in the proof of Theorem 6.3.
Proposition 6.1. K is a compact subset of P , and Σγ is a compact subset

of K such that
Σγ ⊂ Mγ
for all γ > 0. Moreover, the following properties are satisfied:
(a) If γ̃ > γ > 0, then
Mγ̃ ⊂ Mγ and Σγ̃ ⊂ Σγ .
(b) The cone P can be expressed through
∪
P \ {0} = Mγ .
γ>0
(c) If T x > γx for some γ > 0 and x > 0, then, T x ∈ Mγ̃ for some γ̃ > γ
and, in particular, T x ∈ Mγ , by Part (a).
(d) Let γ > 0 and any sequence {γn }n≥1 such that
lim γn = γ, 0 < γn < γn+1 < γ, and Σγn ̸= ∅,
n→∞
for all n ≥ 1. Then,
∩
Σγ = Σγn ̸= ∅.
n≥1
(e) Σγ = Mγ = ∅ if
( γ > ∥T ∥L(E)
] (> 0).
(f) For every γ ∈ 0, ∥T ∥L(E) , Σγ ̸= ∅ if and only if Mγ ̸= ∅.
Proof. By (6.17), T (P ) ⊂ P and hence,

T (∂B ∩ P ) ⊂ T (P ) ⊂ P,
which implies
K = T (∂B ∩ P ) ⊂ P̄ = P.
Moreover, since ∂B∩P is bounded and T is compact, T (∂B∩P ) is relatively
compact. Therefore, K is a compact subset of P .
Now, fix a γ > 0. Then, since K, P , (T − γI)−1 (P ) and {x ∈ E : ∥x∥ ≥
γ} are closed sets, it is apparent that Σγ is a closed subset of K. Thus, as
K is compact, Σγ is a compact subset of K. By definition, Σγ ⊂ Mγ for
all γ > 0.
Next, we will prove each of the properties listed in the statement.
Proof of Part (a): Suppose γ̃ > γ > 0 and Mγ̃ ̸= ∅. Let x ∈ Mγ̃ . Then, by
definition,
x>0 and T x ≥ γ̃ x.
Moreover, since γ̃ −γ > 0 and x > 0, we find from R+ P ⊂ P that (γ̃ −γ)x >
0 and hence,
x > 0 and T x ≥ γ̃ x > γ x.
Therefore, x ∈ Mγ and, consequently, Mγ̃ ⊂ Mγ . Obviously, this also
implies that
Σγ̃ = { x ∈ K ∩ Mγ̃ : ∥x∥ ≥ γ̃ }
⊂ { x ∈ K ∩ Mγ : ∥x∥ ≥ γ } = Σγ ,
which concludes the proof of Part (a).
Proof of Part (b): By definition, Mγ ⊂ P \ {0} for all γ > 0 and hence,
∪
Mγ ⊂ P \ {0}.
γ>0
◦
To prove the converse, let x > 0. Then, by (6.17), T x ∈ P and hence,
◦
lim (T x − γx) = T x ∈ P .
γ→0
Thus,
T x − γx ∈ P
for sufficiently small γ > 0, and, therefore, x ∈ Mγ . This ends the proof of
Part (b).
Proof of Part (c): Let γ > 0 and x > 0 such that T x > γx. Then, according
to (6.17), we have that
◦ ◦
Tx ∈ P, T 2 x − γT x ∈ P ,
and hence, T x > 0 and
◦
lim (T 2 x − γT x − ϵT x) = T 2 x − γT x ∈ P .
ϵ→0
Thus, for sufficiently small ϵ > 0, we find that

Tx > 0 and T 2 x ≥ (γ + ϵ)T x.
Equivalently, T x ∈ Mγ+ϵ . This concludes the proof of Part (c).
Proof of Part (d): Let γ > 0 and a sequence {γn }n≥1 such that
lim γn = γ, 0 < γn < γn+1 < γ, and Σγn ̸= ∅,
n→∞
for all n ≥ 1. According to Part (a), we have that

Σγ ⊂ Σγn+1 ⊂ Σγn
for all n ≥ 1, and hence {Σγn }n≥1 provides us with a non-increasing se-
quence of non-empty compact subsets of K. Therefore, by the Cantor
principle,
∩
Σγ ⊂ Σγn ̸= ∅.
n≥1
To show the converse inclusion, let
∩
x∈ Σγn .
n≥1
Then, x ∈ Σγn for all n ≥ 1, and hence,
x ∈ K, ∥x∥ ≥ γn , T x ≥ γn x.
Thus, since K is a compact subset of P , letting n → ∞, we find that
∥x∥ ≥ γ, T x ≥ γ x.
Therefore, x ∈ Σγ , which completes the proof of Part (d).
Proof of Parts (e) and (f ): Fix γ > 0. By definition, Σγ = ∅ if Mγ = ∅.

So, suppose
Mγ ̸= ∅
and pick x ∈ Mγ . Then,
x > 0 and T x ≥ γ x > 0.
In particular, T ̸= 0 and ∥T ∥L(E) > 0.
Subsequently, for every ζ ∈ C, we consider the series
∑∞
S(ζ) = ζ n T n x.
n=0
It is absolutely convergent in E provided
( )−1
∥T n+1 x∥
|ζ| < R(x) := lim sup ≤ ∞.
n→∞ ∥T n x∥
Thus, for every ζ ∈ [0, R(x)), we have that
∑∞ ∞
∑
S(ζ) = ζ nT nx = x + ζ nT nx
n=0 n=1
∞
∑ ∞
∑
= x + ζT ζ n−1 T n−1 x = x + ζT ζ nT nx
n=1 n=0
∞
∑
≥ x + ζγ ζ n T n x = x + ζγS(ζ)
n=0
and, consequently,
(1 − ζγ)S(ζ) ≥ x > 0.
Moreover,
0 < x ≤ S(ζ)
and hence,
1 − ζγ > 0
for all ζ ∈ [0, R(x)). Therefore,
∥T n+1 x∥
0 < γ ≤ R−1 (x) := lim sup ≤ ∥T ∥L(E) . (6.18)
n→∞ ∥T n x∥
When P is a normal cone, owing to Theorem 6.1, the norm ∥ · ∥ can be
chosen to be monotone and, in such case, (6.18) can be inferred directly
from
◦
γ T n−1 x ≤ T n x ∈ P , n ≥ 2,
by taking norms on both sides of these inequalities, which gives rise to
∥T n+1 x∥
γ≤ ≤ ∥T ∥L(E) , n ≥ 1.
∥T n x∥
But in the general case when P is not normal, the previous argument cannot
be shortened according to these patterns.
According to (6.18), it is apparent that Mγ = ∅, and hence, Σγ = ∅ if
γ > ∥T ∥L(E) , which concludes the proof of Part (e).
Subsequently, we consider the sequence of vectors
( n )
T x T n+1 x
yn := T = ∈ E, n ≥ 1.
∥T x∥
n ∥T n x∥
By construction,
T nx T nx
>0 and ∥ ∥ = 1, n ≥ 1.
∥T n x∥ ∥T n x∥
Thus,
T nx
∈ ∂B ∩ P, n ≥ 1,
∥T n x∥
and hence,
yn ∈ T (∂B ∩ P ) ⊂ T (∂B ∩ P ) = K, n ≥ 1.
Actually, according to (6.17), we have that

◦
yn ∈ P ∩ K, n ≥ 1. (6.19)
Note that (6.18) can be rewritten as
0 < γ ≤ R−1 (x) := lim sup ∥yn ∥ ≤ ∥T ∥L(E) .
n→∞
As K is compact, we find from (6.19) that there exist y ∈ K and a subse-

quence {ynk }k≥1 of {yn }n≥1 such that
y = lim ynk and R−1 (x) = lim ∥ynk ∥.
k→∞ k→∞
Necessarily,
0 < γ ≤ R−1 (x) = ∥y∥
and hence, y ∈ (P \ {0}) ∩ K. Also, for every k ≥ 1,
( n +1 ) n +1
T k x T k x
T y nk = T ≥ γ = γ ynk ,
∥T k x∥ ∥T k x∥
n n
because T x ≥ γx. Therefore, letting k → ∞ yields

T y ≥ γ y.
Consequently, y ∈ Σγ and hence Σγ ̸= ∅. The converse is always true,
because Σγ ⊂ Mγ for all γ > 0 and hence Mγ ̸= ∅ if Σγ ̸= ∅. This
concludes the proof.
6.5 Proof of Theorem 6.3
This section consists of the proof of Theorem 6.3. Therefore, it is through-

out assumed that (E, ∥ · ∥, P ) is an OBS with int P ̸= ∅ and that T ∈ K(E)
is strongly positive. All notations introduced in the previous section will
be maintained here. The proof of the theorem will follow after a series of
lemmas.
◦
Lemma 6.4. There exist x0 ∈ P and ρ > 0 such that
T x0 = ρ x0 . (6.20)
Proof. Subsequently, we consider the set

G := { γ > 0 : Σγ ̸= ∅ }.
By Proposition 6.1(e),
( ]
∥T ∥L(E) > 0 and G ⊂ 0, ∥T ∥L(E) .
Moreover, due to Proposition 6.1(b), there exists γ > 0 such that Mγ ̸= ∅.
By Proposition 6.1(f), we have that Σγ ̸= ∅ and, therefore, according to
Proposition 6.1(a), we find that
(0, γ] ⊂ G.
In particular, G ̸= ∅. Actually, combining Parts (a) and (d) of Proposition
6.1, it becomes apparent that there is ρ ∈ [γ, ∥T ∥L(E) ] such that
G := { γ > 0 : Σγ ̸= ∅ } = (0, ρ]. (6.21)
As Σρ ̸= ∅, there exists x0 > 0 such that
T x0 ≥ ρ x0 .
Suppose T x0 ̸= ρ x0 . Then,
T x 0 > ρ x0
and, owing to Proposition 6.1(c), there exists ϵ > 0 such that
T x0 ∈ Mρ+ϵ .
In particular, this implies Mρ+ϵ ̸= ∅ and hence, by Parts (e) and (f) of
Proposition 6.1, Σρ+ϵ ̸= ∅, which entails ρ + ϵ ∈ G and contradicts (6.21).
Consequently, (6.20) holds. Moreover, since x0 > 0, it follows from (6.17)
that
T x 0 = ρ x0 ≫ 0
and, therefore, according to Lemma 6.1, we also obtain that
x0 = ρ−1 ρx0 ≫ 0,
because ρ−1 > 0. This concludes the proof.
Throughout the rest of this section, we fix a real number ρ > 0 and a
vector x0 ≫ 0 satisfying (6.20).
Lemma 6.5. The algebraic multiplicity of ρ as an eigenvalue of the oper-

ator T equals one. Precisely,
N [T − ρIE ] = span [x0 ] (6.22)
and
2
N [(T − ρIE ) ] = N [T − ρIE ] . (6.23)
Proof. The proof of (6.22) proceeds by contradiction. Suppose there is

another vector
x ∈ N [T − ρIE ] \ span[x0 ].
Then, since
( ) ( ) ◦
lim x0 − µ−1 x = lim x0 + µ−1 x = x0 ∈ P ,
µ↑∞ µ↑∞
we obtain that, for sufficiently large µ > 0,

x0 − µ−1 x ∈ P and x0 + µ−1 x ∈ P
and hence, multiplying by µ yields
µx0 − x ∈ P and µx0 + x ∈ P.
Therefore, for sufficiently large µ > 0, we have that
−µ x0 ≤ x ≤ µ x0 . (6.24)
Obviously, (6.24) fails at µ = 0, since x ̸= 0. Therefore, the minimum
value of µ satisfying (6.24) must be positive. Let us denote it by µ0 . Then,
µ0 > 0,
−µ0 x0 ≤ x ≤ µ0 x0 ,
and, since x cannot be a multiple of x0 , we actually have that
−µ0 x0 < x < µ0 x0 .
Consequently, it follows from (6.17) that
−µ0 ρ x0 = −µ0 T x0 ≪ T x = ρ x ≪ µ0 T x0 = µ0 ρ x0 .
Therefore, according to Lemma 6.1,
−µ0 x0 ≪ x ≪ µ0 x0 .
As these estimates are strict, they imply that
−(µ0 − ϵ)x0 ≤ x ≤ (µ0 − ϵ)x0
for sufficiently small ϵ > 0. Consequently, they contradict the minimality
of µ0 and show (6.22).
The proof of (6.23) also proceeds by contradiction. Suppose there exists
x ∈ E such that
2
x ∈ N [(T − ρIE ) ] \ N [T − ρIE ] .
Then, by (6.22),
0 ̸= (T − ρIE )x ∈ N [T − ρIE ] = span [x0 ]
and hence, there is ξ ∈ R \ {0} such that
T x − ρ x = ξx0 . (6.25)
Actually, by choosing −x, instead of x, if necessary, we can assume, without
loss of generality, that (6.25) holds with
ξ > 0.
Arguing as in the proof of (6.22) it is apparent that
−µ x0 ≤ x ≤ µ x0
for sufficiently large µ > 0. Thus, the following minimum is well defined
µ0 := min{ µ ∈ R : x ≤ µ x0 } ∈ R.
In order to show this, we will argue by contradiction. Suppose there exists
a sequence {µn }n≥1 such that
lim µn = ∞ and x ≤ −µn x0
n→∞
for all n ≥ 1. Then,
µ−1
n x ≤ −x0
for all n ≥ 1 and hence, letting n → ∞ shows that −x0 ≥ 0. On the other
hand, we already know that x0 ≥ 0. Thus, x0 = 0, which is a contradiction.
By construction, we have that
x ≤ µ0 x0
and, since
x ̸∈ N [T − ρIE ] = span [x0 ],
necessarily
x < µ0 x0
and hence, it follows from (6.17) and (6.25) that
T x = ρx + ξx0 ≪ µ0 T x0 = µ0 ρ x0 .
Consequently,
( )
ξ
x≤ µ0 − x0 ,
ρ
with
ξ
µ0 −
< µ0 ,
ρ
which contradicts the minimality of µ0 and ends the proof.
Subsequently, we denote by EC the canonical complexification of E,

EC := E + iE,
and TC ∈ L(EC ) stands for the operator
TC (x + iy) = T x + iT y, x, y ∈ E.
By definition,
TC = T in E
and
Re (TC z) = T (Re z) for all z ∈ EC .
Moreover, the following complex counterpart of Lemma 6.5 holds.
Lemma 6.6. One has that

[ ]
2
N (TC − ρIEC ) = N [TC − ρIEC ] = span [x0 ].
Proof. Let z = x + iy ∈ EC be such that

TC z = ρz.
Then,
T x + iT y = ρ(x + iy)
and hence,
x, y ∈ N [T − ρIE ] .
Thus, by (6.22), there are λ, µ ∈ R such that
x = λx0 and y = µx0 .
Therefore,
z = x + iy = (λ + iµ)x0 ∈ span [x0 ].
Consequently,
N [TC − ρIEC ] = span [x0 ]. (6.26)
Now, let x, y ∈ E be such that
[ ]
2
z := x + iy ∈ N (TC − ρIEC ) .
Then,
(TC − ρIEC ) z ∈ N [TC − ρIEC ]
and hence, thanks to (6.26), there are λ, µ ∈ R such that

T x + iT y − ρ(x + iy) = (λ + iµ)x0 .
Equivalently,
(T − ρIE ) x = λx0 , (T − ρIE ) y = µx0 ,
and, so, according to Lemma 6.5, we find that
2
x, y ∈ N [(T − ρIE ) ] = span [x0 ].
Consequently, there exist α, β ∈ R such that
x = αx0 , y = βx0 ,
and, therefore,
z = x + iy = (α + iβ)x0 ∈ N [TC − ρIEC ] = span [x0 ].
The next lemma establishes an important property of the eigenvalues
λ ̸= ρ of TC .
Lemma 6.7. Suppose

λ ∈ C \ {ρ}, z ∈ N [TC − λIEC ].
Then,
( )
Re eiθ z ̸= µ x0 for all (θ, µ) ∈ R × (R \ {0}). (6.27)
Proof. It proceeds by contradiction. Suppose

( )
Re eiθ z = µ x0 (6.28)
for some θ ∈ R, µ ∈ R \ {0}, and z ∈ EC \ {0} such that
TC z = λ z. (6.29)
Note that (6.28) cannot be satisfied if z = 0, because µx0 ̸= 0, and, there-
fore, (6.27) is obvious if z = 0.
According to (6.20), (6.28) and (6.29), it is apparent that
( ( ))
µρx0 = µT x0 = T Re eiθ z
( ) ( ) ( )
= Re TC eiθ z = Re eiθ TC z = Re eiθ λz .
Thus, since ρ > 0, we find that
( ) ( )
Re eiθ ρz = Re eiθ λz ,
which can be expressed in the form

(( ) )
λ
Re − 1 eiθ z = 0,
ρ
or, equivalently,
( )
λ
− 1 eiθ z ∈ iE. (6.30)
ρ
As λ ̸= ρ, there is τ ∈ R for which
λ λ
0 ̸= −1= − 1 eiτ
ρ ρ
and, so, (6.30) becomes
ei(θ+τ − 2 ) z ∈ E.
π
Consequently, there exists t ∈ R such that

x := eit z ∈ E.
Now, going back to (6.28), yields
( ) ( )
µx0 = Re eiθ z = Re ei(θ−t) eit z
( ) ( )
= Re ei(θ−t) x = Re ei(θ−t) x = cos (θ − t)x
and hence,
cos (θ − t) ̸= 0,
because µ ̸= 0 and x0 ≫ 0. Therefore,
µe−it
z = e−it x = x0 .
cos (θ − t)
This is impossible, since this identity implies
λz = TC z = ρz
and, consequently, λ = ρ. This contradiction ends the proof.
The previous lemmas together with the next one complete the proofs of
Parts (a) and (c) of Theorem 6.3.
Lemma 6.8. One has that |λ| < ρ for all λ ∈ σ(T ) \ {ρ}. Therefore,
ρ = spr T.
Proof. Obviously, the estimate holds if λ = 0. So, suppose λ ̸= 0 and let

z ∈ EC \ {0} such that
TC z = λz.
Fix θ ∈ R. Then, since
( ( ))
lim x0 − µ−1 Re eiθ z = x0 ≫ 0,
µ↑∞
there exists µθ > 0 such that

( )
Re eiθ z ≪ µθ x0 ≪ µx0 for all µ > µθ .
Moreover, as the map
( )
θ 7→ Re eiθ z
is continuous, there exists ϵθ > 0 such that
( )
Re eit z ≪ µθ x0 for all t ∈ Iθ := [θ − ϵθ , θ + ϵθ ].
Consequently,
( )
Re eit z ≪ µx0 (6.31)
for all t ∈ Iθ and µ ≥ µθ . As [0, 2π] is compact, there exist an integer p ≥ 1
and p points
θj ∈ [0, 2π], j ∈ {1, ..., p},
such that
∪
p
[0, 2π] ⊂ Iθ j .
j=1
Set
µ̃ := max µθj > 0.
1≤j≤p
Then, by construction, the estimate (6.31) holds for all t ∈ [0, 2π] and
µ ≥ µ̃. Therefore, by 2π-periodicity, we obtain that
( )
Re eiθ z ≤ µx0 ∀ θ ∈ R, µ ≥ µ̃. (6.32)
Now, we will show that (6.32) fails if we take µ = 0. The proof will proceed
by contradiction. Suppose
( )
Re eiθ z ≤ 0 ∀ θ ∈ R.
Then, setting
z = x + iy, x, y ∈ E,
we obtain that
x cos θ ≤ y sin θ, θ ∈ R. (6.33)
In particular, for (6.33) at θ = ± π4 , we find that
x=y
and hence, (6.33) implies that
(sin θ − cos θ)x ≥ 0, θ ∈ R. (6.34)
For (6.34) at θ = π
2 and θ = 0 gives x ≥ 0 and −x ≥ 0. Thus,
x ∈ P ∩ (−P ) = {0}
and, therefore, x = 0, which implies z = 0. This is a contradiction. Conse-
quently, (6.32) fails at µ = 0, and, therefore, there exists a minimal µ0 > 0
such that
( )
Re eiθ z ≤ µ0 x0 ∀ θ ∈ R.
According to Lemma 6.7, we must have
( )
Re eiθ z < µ0 x0 ∀ θ∈R
and, hence, due to (6.17) and (6.20), we find that
( ( ))
T Re eiθ z ≪ µ0 ρx0 ∀ θ ∈ R. (6.35)
As λ ̸= 0, |λ| > 0 and there exists τ ∈ R such that
λ = |λ|eiτ .
( ( )) ( )
µ0 ρx0 ≫ T Re eiθ z = Re eiθ TC z
( ) ( )
= Re eiθ λz = |λ| Re ei(θ+τ ) z
for all θ ∈ R. Consequently,
( ) ρ
Re eit z ≪ µ0 x0 ∀ t ∈ R.
|λ|
By the minimality of µ0 , we must have
ρ ≥ |λ|.
Moreover, if ρ = |λ|, then
( )
Re eit z ≪ µ0 x0 ∀ t ∈ R,
and, so, µ0 could be shortened. Therefore, ρ > |λ|, as requested.
Finally, thanks to Theorem 6.2(a), it is obvious that
0 < ρ = spr T.
The next lemma shows Part (b) of Theorem 6.3.
Lemma 6.9. ρ = spr T is the unique real eigenvalue of T to a positive

eigenvector x > 0.
Proof. Let λ ∈ [R ∩ σ(T )] \ {ρ} such that

T x = λx
for some x > 0. Then, since T is strongly positive, we have that T x ≫ 0
and hence λ > 0. Thus, thanks to Lemma 6.8,
0 < λ < ρ.
Also, by Lemma 6.1, x ≫ 0. Thus,
0 ≪ x0 ≤ µ x
for sufficiently large µ. Let µ0 > 0 be the minimum among all these µ’s.
Then,
0 ≪ x0 < µ0 x, (6.36)
because x0 = µ0 x implies
ρ x0 = T x0 = µ0 T x = µ0 λ x = λ x0 ,
and hence λ = ρ, which is a contradiction. This shows (6.36) and, conse-
quently,
T x0 = ρ x0 ≪ µ0 T x = µ0 λ x.
Therefore,
λ
x0 ≪ µ0
x,
ρ
which contradicts the minimality of µ0 , as λ < ρ, and ends the proof.
To complete the proof of Theorem 6.3, it remains to prove Parts (d),
(e) and (f). First, we will prove Part (d). According to (6.8), for every
λ > spr T , we have that
∞
∑
R(λ; T ) := λ−(n+1) T n .
n=0
Therefore, R(λ; T ) must be strongly positive, because λ−(n+1) > 0 and T n

is strongly positive for all n ≥ 1. Indeed, let x > 0. Then,
∑∞
y := λ−1 x + λ−(n+1) T n x ≥ 0
n=2
and
R(λ; T )x = λ−2 T x + y.
As T is strongly positive, Lemma 6.1 implies that
◦
λ−2 T x ∈ P .
Thus,
λ−2 T x + z ∈ P
for sufficiently small z ∈ E, and, therefore, for such z’s,
R(λ; T )x + z = λ−2 T x + z + y ∈ P,
because P + P ⊂ P . This shows that
◦
R(λ; T )x ∈ P
and concludes the proof of Part (d).
Now, we will prove Part (e). We already know that spr T is a simple
eigenvalue of T . Thus, by Theorem 6.2(c), there exists ϵ > 0 such that
∞
∑
R(ζ; T ) = (ζ − spr T )n Tn , 0 < |ζ − spr T | ≤ ϵ,
n=−1
where
∫
1
Tn := (ζ − spr T )−(n+1) R(ζ; T ) dζ
2πi |ζ−spr T |=δ
for all n ≥ −1 and δ ∈ (0, ϵ]. Moreover, T−1 is a linear projection of E onto
R[T−1 ] = N [spr T IE − T ] = span [x0 ] (6.37)
and, therefore,
tr T−1 := dim R[T−1 ] = 1.
On the other hand, we have that
T−1 = lim [(λ − spr T ) R(λ; T )]
λ∈R
λ↓spr T
and, due to Part (d), R(λ; T ) is strongly positive for all λ > spr T . Thus,
T−1 ≥ 0 in the sense that
T−1 x ≥ 0 for all x ∈ P, (6.38)
because P̄ = P . Moreover, there exists x ∈ P such that
T−1 x > 0.
Indeed, if T−1 = 0 in P , then, since P is generating, T−1 = 0 in E, which

is impossible, by (6.37). Again by (6.37), there exists ξ > 0 such that
T−1 x = ξx0 ≫ 0.
Consequently, for sufficiently small λ − spr T < 0, we find that
∑∞
ξ
R(λ; T )x = x0 + (λ − spr T )n Tn x ≪ 0,
λ − spr T n=0
which concludes the proof of Part (e).

Finally, we will prove Theorem 6.3(f). By the results on Section 6.2,
T ∗ ∈ K(E ′ ), σ(T ) = σ(T ∗ ), and
spr T ∗ = spr T
is a simple eigenvalue of T ∗ . Consequently, applying Theorem 6.2(c) to T ∗ ,
there exists ϵ > 0 such that
∞
∑
R(ζ; T ∗ ) = (ζ − spr T )n T̃n , 0 < |ζ − spr T | ≤ ϵ,
n=−1
where
∫
1
T̃n := (ζ − spr T )−(n+1) R(ζ; T ∗ ) dζ, n ≥ −1.
Thus, owing to (6.16), we have that

∫
1
T̃n = (ζ − spr T )−(n+1) R(ζ; T )∗ dζ = T∗n
for all n ≥ −1 and δ ∈ (0, ϵ]. Moreover, T∗−1 is a linear projection of E ′
onto
R[T∗−1 ] = N [spr T IE ′ − T ∗ ] (6.39)
and hence, by (6.11) and Theorem 6.3(a),
tr T∗−1 := dim R[T∗−1 ] = 1. (6.40)
Also, by definition, for every x′ ∈ P ∗ , we have that x′ (x) ≥ 0 for all x ∈ P
and
T∗−1 x′ (x) = x′ (T−1 x).
T∗−1 x′ (x) ≥ 0 ∀ x ∈ P, x′ ∈ P ∗ ,
and, therefore,
T∗−1 (P ∗ ) ⊂ P ∗ .
As P is normal, i.e., P ∗ is generating, and, due to (6.40), T∗−1 ̸= 0, there
exists y0′ ∈ P ∗ \ {0} such that
x′0 := T∗−1 (y0′ ) ∈ P ∗ \ {0}.
By (6.39), we infer that
T ∗ x′0 = spr T x′0 .
Thus,
N [spr T IE ′ − T ∗ ] = span [x′0 ]
and (6.11) implies that
R[spr T IE − T ] = span [x′0 ]⊥ = ker x′0 . (6.41)
Next, we will show that
/ R[spr T IE − T ] = ker x′0 .
x0 ∈ (6.42)
The proof of (6.42) proceeds by contradiction. Suppose
spr T x − T x = x0
for some x ∈ E. Then,
(spr T IE − T )2 x = (spr T IE − T )x0 = 0
and, therefore,
x ∈ N [(spr T IE − T )2 ] \ N [spr T IE − T ]
which contradicts the fact that spr T is a simple eigenvalue of T . This shows
(6.42) and hence x′0 (x0 ) ̸= 0. Necessarily, x′0 (x0 ) > 0, because x′0 ∈ P ∗ .
Finally, let x ∈ P \ {0} be arbitrary. Then,
spr T x′0 (x) = T ∗ x′0 (x) = x′0 (T x) > 0.
◦
Indeed, since T x ∈ P , there exists ϵ > 0 such that
◦
T x − ϵx0 ∈ P
and hence,
x′0 (T x − ϵx0 ) ≥ 0,
which implies
x′0 (T x) ≥ ϵx′0 (x0 ) > 0
and, therefore, x′0 (x) > 0. Consequently, thanks to (6.41), we conclude that
x∈
/ R[spr T IE − T ],
which ends the proof of Theorem 6.3.
Section 6.1 introduces some basic concepts of the theory of ordered Banach
spaces. It strictly contains the necessary background to prove the gener-
alized version of the Krein–Rutman theorem studied in this chapter. As
we learned most of these materials from H. Amann [8] and H. H. Schae-
fer [194], we have adapted the contents of these references to our purposes
here, though we tidied up considerably some of the materials of H. Amann
[8] (compare our proof of Lemma 6.2 with the sketch of the proof of Propo-
sition 1.7 of H. Amann [8] through the — unnecessary — concept of open
decomposition of E). The abstract theory of H. H. Schaefer [194], in the
framework of topological vector spaces, widely surpasses the limitations
imposed to this chapter.
According to Theorem V.5.5 of H. H. Schaefer [194], if E is an ordered
Banach space with generating positive cone P , then, every positive linear
form on E is continuous. This property, going back to V. L. Klee [114],
reveals the existence of extremely sharp connections between topology and
algebra within the theory of positive operators.
Although H. H. Scheafer [194] attributed to M. G. Krein [119] the proof
of the fact that P ∗ is generating if P is normal (see Lemma V.3.2.1 of
[194]), the whole characterization, P ∗ is generating if and only if P is
normal, might be attributed to J. Grosberg and M. G. Krein [89].
The Krein–Rutman theorem provides us with the most natural exten-
sion of the Frobenius–Perron theorem to the general context of ordered
Banach spaces. Although the original setting of the Krein–Rutman theo-
rem was widely extended to cover abstract topological vectors spaces (see
H. H. Schaefer [194] and its list of references), M. G. Krein himself and
his collaborators, instead, begun a systematic program to study general
spectral properties of Fredholm operators of index zero, whose highest cli-
max was reached with the publication of the celebrated monograph of I. C.
Göhberg and M. G. Krein [83] (see J. López-Gómez and C. Mora-Corral
[153] for the most recent advances in this field).
The proof of Theorem 6.3(a)-(c) is based upon P. Takác [218]. Takác’s
proof, besides it admits a number of geometrical and analytical interpreta-
tions, as it relies upon a quite natural iterative method, it has the tremen-
dous advantage over other available proofs that it does not make use of any
sophisticated mathematical tool. As a consequence, it can be comfortably
taught at undergraduate level. There are other elementary proofs in the lit-
erature, as, for instance, the dynamical one of N. D. Alikakos and G. Fusco
[6], but yet the students should be familiar with the most basic concepts
of the theory of dynamical systems, like ω-limits and their properties, for
reading it comfortably.
The proof of Theorem 6.3(d), though of a different nature, is also el-
ementary. But the proof of Theorem 6.3(e) is far from elementary, as it
uses some advanced mathematical tools and results from operator theory.
Theorem 6.3(e) provides us with a rather abstract anti-maximum principle
within the spirit of Theorem 4.1 of P. Takác [219]. The proofs of Theorem
6.3(e) and (f) in this book have been elaborated from the proof of the gen-
eralized version of the Krein–Rutman theorem collected on p. 265 of the
Appendix of H. H. Schaefer [194], which reads as follows.
Theorem 6.4 (of M.G. Krein and M. A Rutman). Let E be an or-

dered Banach space with total positive cone P , and let T be a compact
positive endomorphism of E. If T has a spectral radius spr T > 0, then
spr T is a pole of the resolvent of maximal order on the spectral circle
|ζ| = spr T,
with an eigenvector in P . A corresponding result holds for the adjoint T ∗
in E ′ .
Essentially, the strong positivity on T imposed in Theorem 6.3 is nec-

essary for obtaining the simplicity of spr T . But it is unnecessary for the
existence of a positive eigenvector associated to spr T , as it becomes appar-
ent from Theorem 6.4.
The fact that, in the context of Theorem 6.4, spr T is a pole of the
resolvent follows from a classical theorem by Pringsheim (see p. 262 of H.
H. Schaefer [194]). The fact that T and T ∗ admit a positive eigenvector
in P and P ∗ , respectively, associated to spr T , can be obtained by easily
adapting our proof of Theorem 6.3(f), as in H. H. Scheafer [194].
The proof of the fact that x′0 (x) > 0 for all x ∈ P \ {0} in the final part
of the proof of Theorem 6.3 was communicated by H. Amann to the author
through an electronic mail.

Chapter 7
The strong maximum principle
Throughout this chapter we work under the general assumptions of Chapter

5. Thus, besides the general assumptions of Chapter 4, we suppose that
∂Ω is of class C 2 .
The main goal of the chapter is to study the scalar version of Theo-
rem 2.1 of J. López-Gómez and M. Molina-Meyer [148], which was estab-
lished for a general class of cooperative systems under Dirichlet boundary
conditions. Such result characterized whether, or not, (L, B, Ω) satisfies
the strong maximum principle, or, equivalently, the maximum principle, in
terms of the positivity of the principal eigenvalue of the linear eigenvalue
problem
{
Lφ = τ φ in Ω,
(7.1)
Bφ = 0 on ∂Ω,
as well as in terms of the existence of a positive strict supersolution h of
(L, B, Ω). A function h is said to be a strict supersolution of (L, B, Ω) if

 Lh ≥ 0 in Ω,
Bh ≥ 0 on ∂Ω,

(Lh, Bh) ̸= (0, 0) in Ω × ∂Ω,
and (L, B, Ω) is said to satisfy the strong maximum principle if any strict
supersolution h of (L, B, Ω) satisfies
∂h
h(x) > 0 ∀ x ∈ Ω ∪ Γ1 and (x) < 0 ∀ x ∈ h−1 (0) ∩ Γ0 ,
∂ν
while it is said to satisfy the maximum principle when every supersolution
h of (L, B, Ω) satisfies h ≥ 0. The principal eigenvalue of (7.1) is the unique
τ ∈ R for which (7.1) admits a positive eigenfunction φ > 0. Throughout
the rest of this book it will be denoted by
σ0 := σ[L, B, Ω].
187
The theorem of characterization of the strong maximum principle has shown

to be a milestone for the development of the modern theory of nonlinear
second order elliptic and parabolic equations via bifurcation theory and the
method of sub- and supersolutions (cf. J. López-Gómez and R. M. Pardo
[154], J. López-Gómez [137], M. Molina-Meyer [163–165], J. M. Fraile et al.
[66], J. Garcı́a-Melián et al. [76], H. Amann and J. López-Gómez [13], M.
Delgado et al. [49], J. López-Gómez [138, 139, 143, 144], H. Amann [11],
Y. Du [56], and the list of references therein).
As an application of the theorem of characterization of the strong max-
imum principle, when B = D (Γ1 = ∅) we will be able to re-formulate
the generalized minimum principle of M. H. Protter and H. F. Weinberger
(Theorem 1.7) in terms of the positivity of the principal eigenvalue σ0 , in-
stead of on the existence of a supersolution h such that h(x) > 0 for all
x ∈ Ω̄. This feature provides us with an extremely sharp improvement of
the classical theory covered by Chapter 1, as we are actually characterizing
whether the supersolution h exists. More precisely, as a result of the charac-
terization theorem, such an h exists if and only if σ0 > 0 and, in such case,
there are infinitely many h’s for which all the conclusions of Theorem 1.7
hold. Indeed, if σ0 > 0, then, for every f ≥ 0 and g ≥ 0 with inf ∂Ω g > 0,
the unique solution of
{
Lh = f in Ω,
h=g on ∂Ω,
provides us with an admissible strict supersolution h.
This chapter is distributed as follows. Section 7.1 adapts the classical
theory of Chapters 1 and 2 to deal with supersolutions h, u ∈ W 2,p (Ω),
p > N , through the main theorem of J. M. Bony [28]. Section 7.2 derives
the existence and the uniqueness of the principal eigenvalue of (7.1) from
Theorem 6.3. Section 7.3 shows that any weak eigenvalue τ must actually
be a classical eigenvalue, in the sense that any eigenfunction associated to
τ in the weak sense must be a classical eigenfunction. Section 7.4 shows
that σ0 is simple and dominant. By dominant, it is meant that
Re τ ≥ σ0
for any other eigenvalue τ of (7.1). This result is imperative for ascertaining
the local stability character of the solutions of large classes of parabolic
quasi-linear problems through the signs of the principal eigenvalues of the
associated variational problems. Section 7.5 states and proves the main
theorem of this chapter. Precisely, it establishes that the following five
conditions are equivalent:
The strong maximum principle 189
• σ0 := σ[L, B, Ω] > 0.
• (L, B, Ω) admits a positive strict supersolution h ∈ W 2,p (Ω), p > N .
• (L, B, Ω) satisfies the strong maximum principle.
• (L, B, Ω) satisfies the maximum principle.
• The resolvent of (L, B, Ω) is strongly positive.
Finally, Section 7.6 goes back to Chapter 1 for discussing the range of
validity and applicability of the classical minimum principles of E. Hopf
(Theorem 1.2) and M. H. Protter and H. F. Weinberger (Theorem 1.7).
Essentially, Section 7.6 polishes and updates these classical results in the
light of the main theorem of this chapter.
7.1 Minimum principle of J. M. Bony
This section extends the results of Chapters 1 and 2 to cover the case when
u, h ∈ W 2,p (Ω) for some p > N , instead of u, h ∈ C 2 (Ω) ∩ C 1 (Ω̄). Note that,
according to Theorem 4.2,
N
W 2,p (Ω) ⊂ C 1,1− p (Ω̄)
for all p > N . Also, by Theorem 4.4, any function u ∈ W 2,p (Ω), with
p > N , is twice classically differentiable almost everywhere in Ω.
The results of this section are valid for a general uniformly elliptic op-
erator L whose coefficients satisfy
aij ∈ C(Ω̄), bj , c ∈ L∞ (Ω), ∀ i, j ∈ {1, ..., N }, (7.2)
though, eventually, (7.2) could be relaxed. So, throughout this section we
will impose (7.2).
A careful reading of Chapters 1 and 2 reveals that Theorems 2.1 and 2.4
are based upon Theorem 1.2, which is a direct consequence from Theorem
1.1. Consequently, thanks to the next theorem, which is the bulk of this
section, the regularity requirements in the results of Chapters 1 and 2 can
be substantially relaxed up to deal with supersolutions u, h ∈ W 2,p (Ω),
with p > N .
Theorem 7.1 (Minimum principle of J. M. Bony). Suppose c ≥ 0

and u ∈ W 2,p (Ω), p > N , satisfies
inf ess Lu > 0 for every compact subset K ⊂ Ω.
K
Then, u cannot attain a local minimum m ≤ 0 in Ω.

The proof of Theorem 7.1 follows after a series of preliminary technical

results. Essentially, the next lemma establishes that any sufficiently regu-
lar map sends negligible sets into negligible sets. As usual, for any given
measurable set M , |M | stands for the measure of M , as discussed by H.
Lebesgue [124].
Lemma 7.1. Let
f : Ω → RN , f = (f1 , ..., fN ),
with fi ∈ W 1,p (Ω), p > N , for all 1 ≤ i ≤ N . Then,
M ⊂Ω and |M | = 0 =⇒ |f (M )| = 0.
Proof. Subsequently, for every x0 ∈ RN and γ > 0, we denote by Cγ (x0 )

the closed γ-cube centered at x0
[ γ γ ]N
Cγ (x0 ) := x0 + − , .
2 2
By Theorem 4.2, there exists a constant K > 0 such that
(∫ ) p1
max |g(x) − g(y)| ≤ K |∇g|p
x,y∈C1 (0) C1 (0)
for every g ∈ W 1,p (C1 (0)).

Let x0 ∈ Ω and γ > 0 such that Cγ (x0 ) ⊂ Ω. Then, the functions
gi (x) := fi (γx + x0 ), x ∈ C1 (0), 1 ≤ i ≤ N,
satisfy gi ∈ W 1,p (C1 (0)) for all 1 ≤ i ≤ N . Thus,
max |fi (x) − fi (y)| = max |fi (γx + x0 ) − fi (γy + x0 )|

x,y∈Cγ (x0 ) x,y∈C1 (0)
= max |gi (x) − gi (y)|

x,y∈C1 (0)
(∫ ) p1
≤K |∇gi |p
C1 (0)
(∫ ) p1
= Kγ |∇fi (γ · +x0 )| p
C1 (0)
(∫ ) p1
1− N
=γ p K |∇fi |p
Cγ (x0 )
for all 1 ≤ i ≤ N . Moreover, since

  12
∑N
∂f
2 ∑N
∂fi ∑N
∂fi
|∇fi | = 
i 
≤ ≤ ,
j=1
∂xj j=1
∂xj i,j=1
∂xj
we find that
  p  p1
∫ ∑N
1− N   ∂f i  
max |fi (x) − fi (y)| ≤ γ p K
x,y∈Cγ (x0 ) Cγ (x0 ) i,j=1
∂x j
for all 1 ≤ i ≤ N . Therefore,

  p  Np
∫ ∑N
)K N   ∂f i  
|f (Cγ (x0 ))| ≤ γ N ( 1− N
p . (7.3)
Cγ (x0 ) i,j=1
∂xj
As |M | = 0, for every ϵ > 0 there exists a sequence of disjoint cubes, say

Cn := Cγn (xn ), n ≥ 1,
such that
∪ ∑ ∑
M⊂ Cn ⊂ Ω and |Cn | = γnN ≤ ϵ.
n≥1 n≥1 n≥1
Hence, from (7.3) and the Hölder inequality it is apparent that

  p  Np
∑ ∑ N (1− N ) ∫ ∑N
∂fi  
|f (M )| ≤ |f (Cn )| ≤ K N γn p  
Cn i,j=1
∂x j
n≥1 n≥1
 1− Np   p  Np
∑ ∫ ∑N
∂fi  
≤ KN  γnN   
Ω i,j=1
∂xj
n≥1
  p  Np
∫ ∑N
∂fi  
≤ ϵ1− p K N  
N
Ω i,j=1
∂xj
for all ϵ > 0. Consequently, |f (M )| = 0.

The next result establishes the positivity of the quadratic form D2 u at
any local strict minimum when u ∈ W 2,p (Ω) for some p > N .
Lemma 7.2. Let u ∈ W 2,p (Ω), p > N , and x0 ∈ Ω such that u attains a
local strict minimum m ∈ R at x0 , i.e., there exists δ > 0 such that
u(x) > u(x0 ) = m ∀ x ∈ B̄δ (x0 ) \ {x0 }. (7.4)
Then, for every ϵ > 0, there exists a measurable subset M of Bϵ (x0 ) with
|M | > 0 such that the quadratic form
( 2 )
2 ∂ u(x)
D u(x) :=
∂xi ∂xj 1≤i,j≤N
is positive definite for almost every x ∈ M .
Proof. Let S denote the C 1 surface of RN × R determined by the graph

of y = u(x)
S := {(x, u(x)) : x ∈ Ω}.
Fix ϵ > 0 such that B̄ϵ (x0 ) ⊂ Ω and let M denote the set of points
x ∈ Bϵ (x0 ) for which S lies above the tangent hyperplane Tx of S at x
in a neighborhood of x. Obviously, M is a closed set and hence it is mea-
surable and x0 ∈ M , because u attains a local strict minimum at x0 . Also,
∇u(x0 ) = 0 and, so, Tx0 is given through y = m. Now, we will prove that
the next property holds:
(P) There exists η > 0 such that for every h ∈ Bη := Bη (0) ⊂ RN , there is
some p ∈ RN for which the hyperplane
y = ⟨h, x⟩ + p
is tangent to S at some point of M .
Indeed, thanks to (7.4), for every

δ̃ ∈ (0, min{ϵ, δ}),
there exists η := η(δ̃) such that
⟨h, x − x0 ⟩ + m < u(x), δ̃ < |x − x0 | < δ,
for all h ∈ Bη . By construction, for every h ∈ Bη , there must exist a
parallel hyperplane to
y = ⟨h, x − x0 ⟩ + m (7.5)
tangent to S at some point x ∈ B̄δ̃ ∩ M . Indeed, it is the unique hyperplane
parallel to (7.5) supporting S. This shows Property (P).
Let f : Ω → RN denote the gradient map
f (x) := ∇u(x), x ∈ Ω.
According to Property (P), for every h ∈ Bη , there exists x ∈ M such that
f (x) = ∇u(x) = h.
Thus, Bη ⊂ f (M ) and, consequently,

|f (M )| > 0.
On the other hand, u ∈ W (Ω) implies that
2,p
∂u
fi = ∈ W 1,p (Ω), 1 ≤ i ≤ N,
∂xi
and, therefore, owing to Lemma 7.1, we find that
|M | > 0.
Furthermore, as u ∈ W (Ω) with p > N , u is twice classically differen-
2,p
tiable almost everywhere in Ω. Consequently, the bilinear form D2 u(x)

must be positive definite for almost every x ∈ M . This ends the proof.
Finally, Theorem 7.1 is a corollary from the next result.
Proposition 7.1. Suppose c ≥ 0, and u ∈ W 2,p (Ω), p > N , possesses a
local minimum m ≤ 0 at some x0 ∈ Ω. Then,
lim sup ess Lu(x) ≤ 0. (7.6)
x→x0
Proof. Suppose u has a local strict minimum at x0 . By Lemma 7.2,

lim inf ess ∆u(x) ≥ 0. (7.7)
x→x0
Thus, by performing the linear change of coordinates (1.4) of the proof of
Proposition 1.1, it is apparent that (7.7) entails
 
∑N
∂ 2
u
lim sup ess − aij (x) (x) ≤ 0.
x→x0
i,j=1
∂x i ∂x j
Note that, according to (7.2), aij ∈ C(Ω̄) and, therefore,

 
∑N
∂ 2
u
lim sup ess Lu(x) ≤ lim sup ess − aij (x) (x)
x→x0 x→x0
i,j=1
∂xi ∂xj
+ lim sup ess (c(x)u(x)) ≤ 0,

x→x0
because ∇u(x0 ) = 0, c ≥ 0 and m ≤ 0. Consequently, (7.6) holds.
In the general case when x0 is not a strict local minimum of u, one can
apply the previous result to the auxiliary function
v(x) := u(x) + |x − x0 |4 , x ∈ Ω,
to get
lim sup ess Lv(x) ≤ 0,
x→x0
which implies (7.6) and concludes the proof.
By inter-exchanging the roles of Theorem 7.1 and Theorem 1.1, the proof
of Theorem 1.2 can be adapted, mutatis mutandis, to obtain the following
generalized version of the minimum principle of E. Hopf.
Theorem 7.2. Suppose c ≥ 0, and u ∈ W 2,p (Ω), p > N , satisfies
Lu ≥ 0 in Ω, and m := inf u ∈ (−∞, 0].

Ω
Then, either u = m in Ω, or u(x) > m for all x ∈ Ω. In other words, u

cannot attain m in Ω, unless u = m in Ω. Moreover,
inf u = inf u = m
Ω̄ ∂Ω
when Ω is bounded.
Similarly, Theorem 1.3 admits the next counterpart. Note that u ∈

C (Ω̄) if u ∈ W 2,p (Ω) with p > N .
1
Theorem 7.3. Suppose c ≥ 0 and u ∈ W 2,p (Ω), p > N , is a non-

constant function satisfying
inf ess Lu ≥ 0 in Ω, and m := inf u ∈ (−∞, 0].

Ω
Assume, in addition, that there exists x0 ∈ ∂Ω such that u(x0 ) = m, with

Ω satisfying an interior sphere property at x0 . Then,
∂u
(x0 ) < 0.
∂ν
Based on Theorems 7.2 and 7.3, the proofs of Theorems 2.1 and 2.4 can
be easily adapted to get the next counterparts in W 2,p (Ω), p > N .
Theorem 7.4. Suppose (L, B, Ω) admits a supersolution h ∈ W 2,p (Ω),

p > N , such that
h(x) > 0 for all x ∈ Ω̄.
Then, any supersolution u ∈ W 2,p (Ω) of (L, B, Ω) must satisfy some of the
following alternatives:
A1. u = 0 in Ω.
A2. u(x) > 0 for every x ∈ Ω ∪ Γ1 , and
∂u
(x) < 0 for all x ∈ u−1 (0) ∩ Γ0 .
∂ν
A3. There exists a constant m < 0 such that

u = mh in Ω̄.
In such case, much like in Theorem 2.1, u(x) < 0 for all x ∈ Ω̄, Γ0 = ∅,
and τ = 0 must be an eigenvalue to a positive eigenfunction (h itself )
of (7.1).
Theorem 7.5. Suppose Ω is of class C 2 and (L, B, Ω) has a positive su-

persolution h ∈ W 2,p (Ω), p > N . Then, any supersolution u ∈ W 2,p (Ω)
of (L, B, Ω) must satisfy some of the Alternatives A1, A2, or A3, of Theo-
rem 7.4.
Note that, thanks to Remark 2.1, Conditions i) and ii) of Theorem 2.4
hold when Ω is of class C 2 .
7.2 The existence of the principal eigenvalue
The main goal of this section is to show the existence and the uniqueness
of τ ∈ R for which (7.1) admits a positive eigenfunction φ. This funda-
mental property will be derived by combining Theorem 6.3 with the next
invertibility and positivity result.
Throughout the rest of this book, for every ν ∈ (0, 1), we denote by
1,ν
CB (Ω̄) the Banach subspace of the Hölder space C 1,ν (Ω̄) (see Section 4.1.3)
consisting of all functions u ∈ C 1,ν (Ω̄) such that Bu = 0 on ∂Ω. Similarly,
CB
1
(Ω̄) := { u ∈ C 1 (Ω̄) : Bu = 0 on ∂Ω }.
Theorem 7.6. There exists ω0 ∈ R such that for every ω ≥ ω0 and f ∈

C(Ω̄) the boundary value problem
{
(L + ω)u = f in Ω,
(7.8)
Bu = 0 on ∂Ω,
possesses a unique weak solution u ∈ WΓ1,2

0
(Ω). Moreover,
− −
u ∈ W 2,∞ (Ω) ⊂ C 1,1 (Ω̄),
u is twice classically differentiable almost everywhere in Ω, and u is a clas-
sical solution of (7.8) (in the sense of Definition 4.1).
Also, ω0 can be chosen so that f > 0 (f ≥ 0 with f ̸= 0) implies
∂u
u(x) > 0 ∀ x ∈ Ω ∪ Γ1 and (x) < 0 ∀ x ∈ Γ0 . (7.9)
∂ν
Furthermore, for every ν ∈ (0, 1), the resolvent operator

(L+ω)−1 1,ν
C(Ω̄) −→ CB (Ω̄)
f 7→ (L + ω)−1 f := u
is linear and continuous. Consequently, the operator
(L+ω)−1
C(Ω̄) −→ CB
1
(Ω̄)
(7.10)
−1
f 7→ (L + ω) f := u
is linear, continuous and compact.
Proof. The existence and the uniqueness of the weak solutions for suf-
ficiently large ω follows from Theorem 4.11. The regularity of the weak
solution is a consequence from Theorem 5.11. The positivity properties
established by (7.9) follow straightaway from Lemma 2.1, Proposition 2.1
and Theorem 7.4. Indeed, by Lemma 2.1 and Proposition 2.1, there exists
ω0 ∈ R such that, for every ω > ω0 , (L + ω, B, Ω) possesses a strict su-
persolution h ∈ C 2 (Ω̄) with h(x) > 0 for all x ∈ Ω̄. Thus, for sufficiently
large ω ≥ ω0 , the unique weak solution of (7.8) must satisfy some of the
alternatives of Theorem 7.4 if f > 0. Clearly, Alternative A1 cannot occur
because u ̸= 0. Similarly, if u = mh with m < 0, then
f = (L + ω)u = m(L + ω)h ≤ 0,
which is impossible. Therefore, Alternative A2, and hence (7.9) holds.
The fact that (L + ω)−1 is linear follows from the linear structure of
the problem and the uniqueness of the weak solution. Its continuity follows
from Theorem 4.11 and the uniform elliptic estimates of Chapter 5. Indeed,
let {fn }n≥1 be a sequence in C(Ω̄) such that
lim ∥fn − f ∥C(Ω̄) = 0 (7.11)
n→∞
for some f ∈ C(Ω̄). Set

u := (L + ω)−1 f, un = (L + ω)−1 fn , n ≥ 1;
u is the unique weak solution of (7.8) and, for every n ≥ 1, un is the unique
weak solution of
{
(L + ω)un = fn in Ω,
Bun = 0 on ∂Ω.
By Theorem 4.11,
( )
(L + ω)−1 ∈ L L2 (Ω), WΓ1,2
0
(Ω) ,
and, due to (7.11),

lim ∥fn − f ∥Lp (Ω) = 0 (7.12)
n→∞
for all p ≥ 2. Thus,

lim ∥un − u∥W 1,2 (Ω) = 0. (7.13)
n→∞ Γ0
On the other hand, for every n, m ≥ 1, we have that

{
(L + ω)(un − um ) = fn − fm in Ω,
B(un − um ) = 0 on ∂Ω,
and hence, thanks to Theorem 5.10, for every p ≥ 2, there exists a constant
C = C(p) > 0 such that
∥un − um ∥W 2,p (Ω) ≤ C∥fn − fm ∥Lp (Ω)
for all n, m ≥ 1. According to (7.12), these estimates imply that {un }n≥1
is a Cauchy sequence in W 2,p (Ω) for all p ∈ [2, ∞). By (7.13), we must
have u ∈ W 2,p (Ω) and
lim ∥un − u∥W 2,p (Ω) = 0 for all p ≥ 2.
n→∞
Therefore, due to Theorem 4.2, we find that

lim ∥un − u∥C 1,ν (Ω̄) = 0
n→∞ B
for all ν ∈ (0, 1), which concludes the proof of the continuity.
1,ν
The compactness of (7.10) is a consequence from the fact that CB (Ω̄)
is compactly embedded in CB 1
(Ω̄) for all ν ∈ (0, 1). Indeed, fix ν ∈ (0, 1)
1,ν
and let {un }n≥1 be a bounded sequence of CB (Ω̄). Then, there exists a
constant C > 0 such that
∂un ∂un
∥un ∥C 1 (Ω̄) ≤ C, (x) − (y) ≤ C|x − y|ν , (7.14)
∂xj ∂xj
for all n ≥ 1, x, y ∈ Ω̄, and 1 ≤ j ≤ N . As {un }n≥1 is bounded in C 1 (Ω̄),
it is bounded and equicontinuous in C(Ω̄). Thus, by the theorem of Arzela-
Ascoli, there exists u ∈ C(Ω̄) such that, along some subsequence relabeled
by n, we have that
lim ∥un − u∥C(Ω̄) = 0.
n→∞
According to (7.14), the sequence {∂un /∂x1 }n≥1 is also bounded and
equicontinuous in C(Ω̄). Thus, there exists v1 ∈ C(Ω̄) such that, along
some subsequence relabeled by n,
∂un
lim ∥ − v1 ∥C(Ω̄) = 0.
n→∞ ∂x1
Repeating this argument N times, it is apparent that there exist vj ∈ C(Ω̄),
2 ≤ j ≤ N , such that, along some subsequence relabeled by n,
∂un
lim ∥ − vj ∥C(Ω̄) = 0, 1 ≤ j ≤ N.
n→∞ ∂xj
On the other hand, for every x ∈ Ω̄ and n ≥ 1, we have that
∫ 1
un (x + h) − un (x) = ⟨∇un (x + th), h⟩ dt
0
for sufficiently small h ∈ R . Consequently, letting n → ∞, we find that

N
∫ 1
u(x + h) − u(x) = ⟨(v1 (x + th), ..., vN (x + th)) , h⟩ dt
0
for all x ∈ Ω̄ and sufficiently small h. From these identities it becomes

apparent that u ∈ C 1 (Ω̄) and that
∇u = (v1 , ..., vN ) in Ω̄.
Therefore,
lim ∥un − u∥C 1 (Ω̄) = 0
n→∞
and, in particular, Bu = 0 on ∂Ω, because Bn u = 0 on ∂Ω for all n ≥ 1.

Throughout the rest of this chapter, we always take ω ≥ ω0 . By the
theorem of Arzela-Ascoli, the canonical injection
J : CB
1
(Ω̄) ,→ C(Ω̄)
is a linear compact operator and, therefore,
Rω := J(L + ω)−1 : C(Ω̄) −→ C(Ω̄)
is a compact endomorphism of C(Ω̄). The space C(Ω̄) is an ordered Ba-
nach space with the natural ordering induced by its cone of non-negative
functions
P := { u ∈ C(Ω̄) : u ≥ 0 in Ω̄ }.
Obviously, the interior of P is given by

◦
P = { u ∈ C(Ω̄) : u(x) > 0 ∀ x ∈ Ω̄ } =
̸ ∅,
and, according to Theorem 6.1, P is normal, because the norm ∥ · ∥C(Ω̄) is
monotone. But the resolvent operator Rω is not strongly positive in C(Ω̄)
if Γ0 ̸= ∅, because, in such case, Rω f = 0 on Γ0 for all f ∈ P and, hence,
◦
Rω f ∈
/ P.
The Banach space CB
1
(Ω̄) is also an ordered Banach space with positive
cone
P1 := { u ∈ CB
1
(Ω̄) : u ≥ 0 in Ω̄ } = P ∩ CB
1
(Ω̄).
◦
Moreover, the interior of P1 , P 1 , consists of the set of functions u ∈ P1
satisfying (7.9). Thus,
◦
Rω f ∈ P 1 ∀ f ∈ P1 \ {0}
and, therefore, according to Theorem 7.6, the resolvent operator
Rω,1 := Rω | : CB
1
(Ω̄) −→ CB
1
(Ω̄)
C 1 (Ω̄)
B
satisfies
◦
Rω,1 (P1 \ {0}) ⊂ P 1 .
Consequently, in this case, Theorem 6.3 applies for inferring the existence
of the principal eigenvalue of (7.1), though, owing to Theorem 6.1, the cone
P1 is not normal, because the norm ∥ · ∥CB 1 (Ω̄) is not monotone, and hence
Theorem 6.3(f) cannot be guaranteed to hold in this ordered Banach space.

The lack of monotonicity of ∥ · ∥CB1 (Ω̄) is inherent to the fact that it involves
the derivatives of the functions on which it acts.

To overcome this shortcoming, one can proceed as follows. First, let e
be the unique weak solution of
{
(L + ω)e = 1 in Ω,
(7.15)
Be = 0 on ∂Ω.
According to Theorem 7.6, we have that
∂e
e(x) > 0 ∀ x ∈ Ω ∪ Γ1 and (x) < 0 ∀ x ∈ Γ0 . (7.16)
∂ν
Then, consider the Banach space
{ }
Ce (Ω̄) := u ∈ C(Ω̄) : ∃λ > 0 such that −λe ≤ u ≤ λe in Ω̄ (7.17)
equipped with the Minkowski norm

∥u∥e := inf { λ > 0 : −λe ≤ u ≤ λe } , u ∈ Ce (Ω̄). (7.18)
( )
It is easily seen that Ce (Ω̄), ∥ · ∥e is a Banach space whose positive cone
{ }
Pe := u ∈ Ce (Ω̄) : u ≥ 0 in Ω̄ = P ∩ Ce (Ω̄) (7.19)
equips it with a structure of ordered Banach space. As ∥ · ∥e is monotone,
by Theorem 6.1, Pe is a normal cone. Moreover, the canonical injection
Je : Ce (Ω̄) −→ C(Ω̄)
is continuous, because
∥u∥C(Ω̄) ≤ ∥e∥C(Ω̄) ∥u∥e for all u ∈ Ce (Ω̄).
Thus, according to Theorem 7.6, the resolvent operator
Rω,e := (L + ω)−1 Je : Ce (Ω̄) −→ CB
1
(Ω̄) (7.20)
is linear and continuous. The next result shows that
Rω,e f ∈ Ce (Ω̄) for all f ∈ Ce (Ω̄)
and infers some pivotal consequences from this feature.
Proposition 7.2. R[Rω,e ] ⊂ Ce (Ω̄) ∩ CB1

(Ω̄). Therefore, Rω,e can be re-
garded as an endomorphism of Ce (Ω̄). Let denote it by Rω . Then, for every
f ∈ Ce (Ω̄),
u := Rω f ∈ Ce (Ω̄)
is the unique weak solution of (7.8) and Rω is linear, continuous and com-
pact.
Proof. Let f ∈ Ce (Ω̄). By definition,

u := Rω,e f
is the unique weak solution of (7.8). By Theorem 7.6, u is a classical
solution. Moreover, since f ∈ Ce (Ω̄), we have that
−∥f ∥e ∥e∥C(Ω̄) ≤ −∥f ∥e e ≤ f ≤ ∥f ∥e e ≤ ∥f ∥e ∥e∥C(Ω̄)
and hence,
−∥f ∥e ∥e∥C(Ω̄) ≤ (L + ω)u ≤ ∥f ∥e ∥e∥C(Ω̄) in Ω. (7.21)
λ := ∥f ∥e ∥e∥C(Ω̄) , v1 := λe − u, v2 := λe + u.
Then, owing to (7.15) and (7.21), we obtain that

{
(L + ω)vj ≥ 0 in Ω,
(7.22)
Bvj = 0 on ∂Ω,
for each j = 1, 2. Thus, according to Theorem 7.6, vj ≥ 0 in Ω̄, for j = 1, 2,
and, therefore,
−λe ≤ u ≤ λe in Ω̄. (7.23)
Consequently, u ∈ Ce (Ω̄) and
∥u∥e ≤ λ,
which concludes the proof of the first assertion of the statement.
Now, let
Rω : Ce (Ω̄) −→ Ce (Ω̄)
be the linear operator such that Rω f is the unique weak solution of (7.8)
for every f ∈ Ce (Ω̄). The estimate (7.23) actually shows that
∥Rω f ∥e ≤ λ = ∥e∥C(Ω̄) ∥f ∥e for all f ∈ Ce (Ω̄)
and hence,
∥Rω ∥L(Ce (Ω̄)) ≤ ∥e∥C(Ω̄) .
Consequently, Rω is continuous. To prove the compactness of Rω , let
{fn }n≥1 be a bounded sequence of Ce (Ω̄). Then, there exists a constant
λ > 0 such that
∥fn ∥e ≤ λ, n ≥ 1. (7.24)
As Ce (Ω̄) is continuously embedded in C(Ω̄), {fn }n≥1 is also bounded in
C(Ω̄). Subsequently, for every n ≥ 1, we denote by un ∈ Ce (Ω̄) ∩ CB
1
(Ω̄) the
unique weak solution of
{
(L + ω)un = fn in Ω,
(7.25)
Bun = 0 on ∂Ω.
Thanks to Theorem 7.6, {un }n≥1 is relatively compact in CB 1
(Ω̄). Thus,
there exists u ∈ CB (Ω̄) such that, along some subsequence relabeled by n,
1
lim ∥un − u∥CB

1 (Ω̄) = 0. (7.26)
n→∞
On the other hand, according to (7.23), it follows from (7.25) that

−∥fn ∥e ∥e∥C(Ω̄) e ≤ un ≤ ∥fn ∥e ∥e∥C(Ω̄) e for all n ≥ 1.
Thus, owing to (7.24), it is apparent that

−λ∥e∥C(Ω̄) e ≤ un ≤ λ∥e∥C(Ω̄) e, n ≥ 1.
Hence, letting n → ∞, we find that
−λ∥e∥C(Ω̄) e ≤ u ≤ λ∥e∥C(Ω̄) e
and, therefore, u ∈ Ce (Ω̄). Consequently, thanks to (7.26), to complete
the proof of the proposition, it suffices to show that CB
1
(Ω̄) is continuously
embedded in Ce (Ω̄), because this implies
lim ∥un − u∥Ce (Ω̄) = 0.
n→∞
To prove this property, we first show that CB 1
(Ω̄) ⊂ Ce (Ω̄). Let u ∈ CB
1
(Ω̄).
Then, u = 0 on Γ0 . Moreover, by (7.15), e = 0 on Γ0 and, thanks to (7.16),
∂e(x)/∂ν < 0 for all x ∈ Γ0 . Thus, there exists λ0 > 0 such that
∂e ∂u ∂e
λ (x) < (x) < −λ (x)
∂ν ∂ν ∂ν
for all x ∈ Γ0 and λ ≥ λ0 . Hence, there exists ϵ > 0 such that
−λ0 e(x) < u(x) < λ0 e(x) ∀ x ∈ Uϵ , (7.27)
where
Uϵ := Ω ∩ [Γ0 + Bϵ (0)] = { x ∈ Ω : dist (x, Γ0 ) < ϵ }.
Moreover, according to (7.16), e must be separated away from zero in
Ω̄ \ Uϵ = { x ∈ Ω̄ : dist (x, Γ0 ) ≥ ϵ }
and, therefore, there exists λ1 ≥ λ0 such that
−λ1 e(x) < u(x) < λ1 e(x) ∀ x ∈ Ω̄ \ Uϵ . (7.28)
Obviously, (7.27) and (7.28) imply that
−λ1 e ≤ u < λ1 e in Ω̄,
and, consequently, u ∈ Ce (Ω̄). Now, let J denote the canonical injection
J : CB
1
(Ω̄) −→ Ce (Ω̄).
We claim that J is a linear closed operator. Indeed, let {un }n≥1 be a
sequence of CB
1
(Ω̄), and u ∈ CB
1
(Ω̄), v ∈ Ce (Ω̄), such that
lim ∥un − u∥CB
1 (Ω̄) = 0, lim ∥un − v∥e = 0.
n→∞ n→∞
The first limit implies that {un }n≥1 is point-wise convergent to u in Ω̄.
Moreover, by definition, we have that
−∥un − v∥e e ≤ un − v ≤ ∥un − v∥e e in Ω̄
for all n ≥ 1. In particular, {un }n≥1 must be point-wise convergent to v in
Ω̄. Consequently, v = u and, therefore, J is a closed linear operator. By
the closed graph theorem (see, e.g., Theorem II.7 of H. Brézis [29]), J is a
linear continuous operator. This concludes the proof.
The next result shows that the operator Rω introduced in Proposition

7.2 is strongly positive.
Proposition 7.3. For every ω ≥ ω0 , the operator

Rω : Ce (Ω̄) −→ Ce (Ω̄)
is strongly positive. In other words,
◦
Rω (Pe \ {0}) ⊂ intPe := P e .
Proof. Let ω ≥ ω0 and f ∈ Pe \ {0}. Then, f ∈ Ce (Ω̄), f ≥ 0, and f ̸= 0

in Ω. Moreover, by Theorem 7.6, the solution of (7.8)
u := Rω f ∈ CB
1
(Ω̄) ⊂ Ce (Ω̄)
satisfies (7.9). To complete the proof, it remains to show that
◦
u ∈ P e.
Indeed, from (7.9) and (7.16), there must exist ϵ > 0 such that
u > ϵe in Ω̄.
Now, let h ∈ Ce (Ω̄) be such that
∥h∥e ≤ ϵ.
Then, by definition,
−ϵe ≤ ∥h∥e e ≤ h ≤ ∥h∥e e ≤ ϵe,
and hence,
u − ϵe ≤ u + h ≤ u + ϵe in Ω̄.
In particular,
u + h ≥ u − ϵe > 0
and, therefore, u + h ∈ Pe for all h ∈ Ce (Ω̄) with ∥h∥e ≤ ϵ. Consequently,
u ∈ intPe and the proof is complete.
As Pe is a normal cone in Ce (Ω̄), the next result is a direct consequence
from Propositions 7.2, 7.3, Theorem 6.3, and Remark 6.1.
Corollary 7.1. For every ω ≥ ω0 the following properties are satisfied:

(a) spr Rω > 0 is an algebraically simple eigenvalue of Rω and there exists
φ0 ∈ intPe such that
N [spr Rω I − Rω ] = span [φ0 ],
where I stands for the identity map of Ce (Ω̄).
(b) spr Rω is the unique real eigenvalue of Rω to an eigenvector in Pe \{0}.

(c) |λ| < spr Rω for all λ ∈ σ(Rω ) \ {spr Rω }.
(d) For every λ ∈ R with λ > spr Rω , the resolvent operator
−1 ( )
(λI − Rω ) ∈ L Ce (Ω̄)
is strongly positive. Moreover, there exist ϵ > 0 and u > 0 such that
−1
(λI − Rω ) u≪0 for all λ ∈ (spr Rω − ϵ, spr Rω ) .
(e) σ(Rω ) = σ(R∗ω ) and spr Rω = spr R∗ω is a simple eigenvalue of R∗ω .
Moreover, it is the unique eigenvalue of R∗ω in the spectral circle
|ζ| = spr Rω .
(f) There exist φ∗0 ∈ Pe∗ \ {0} such that
N [spr Rω I − R∗ω ] = span [φ∗0 ],
where I stands for the identity map of Ce′ (Ω̄). Moreover, for every
x ∈ Pe \ {0}, φ∗0 (x) > 0 and
spr Rω u − Rω u = x
cannot admit a solution u ∈ Ce (Ω̄).
By Corollary 7.1(a),
Rω φ0 = spr Rω φ0 in Ce (Ω̄)
and hence, by the definition of Rω , the eigenfunction φ0 > 0 provides us
with a weak solution of
{
(L + ω)φ = spr1Rω φ in Ω,
(7.29)
Bφ = 0 on ∂Ω.
−
According to Theorem 7.6, φ0 ∈ C 1,1 (Ω̄) is twice classically differentiable
almost everywhere in Ω and, actually, it is a classical solution of (7.8).
Moreover, since φ0 > 0, we find that
1
(L + ω)φ0 = φ0 > 0
spr Rω
and, therefore,
∂φ0
φ0 (x) > 0 ∀ x ∈ Ω ∪ Γ1 and (x) < 0 ∀ x ∈ Γ0 . (7.30)
∂ν
As (7.29) can be equivalently expressed as
{ ( )
Lφ0 = spr1Rω − ω φ0 in Ω,
Bφ0 = 0 on ∂Ω,
it becomes apparent that there exists a value of τ for which (7.1) admits a
positive eigenfunction. Namely,
1
τ= − ω.
spr Rω
Most precisely, the following fundamental result holds. As a byproduct, the
previous value of τ must be independent of ω ≥ ω0 .
Theorem 7.7 (Existence of the principal eigenvalue). There is a

unique value of τ for which the linear boundary value problem (7.1) ad-
mits a weak solution φ ∈ Pe \ {0}. Such a value of τ will be throughout
denoted by
σ0 := σ[L, B, Ω]
and called the principal eigenvalue of (7.1), or, equivalently, the princi-
pal eigenvalue of (L, B, Ω). Any associated weak solution φ0 ∈ Pe \ {0} will
be referred to as a principal eigenfunction of (7.1), or, equivalently, of
(L, B, Ω).
The principal eigenfunction φ0 is unique, up to a positive multiplicative
− −
real number, it satisfies φ0 ∈ W 2,∞ (Ω) ⊂ C 1,1 (Ω̄) and (7.30), and it is a
classical solution of (7.1), in the sense of Definition 4.1.
Proof. First, we will prove the uniqueness of the principal eigenvalue.

Let σ1 , σ2 ∈ R and φ1 , φ2 ∈ Pe \ {0} such that
{
Lφj = σj φj in Ω,
Bφj = 0 on ∂Ω,
for j = 1, 2. Then, for sufficiently large ω ≥ ω0 and j = 1, 2, we have that
{
(L + ω) φj = (σj + ω) φj > 0 in Ω,
Bφj = 0 on ∂Ω,
and hence,
1
Rω φj := φj .
σj + ω
Consequently, by Corollary 7.1(b),
1 1
spr Rω = =
σ1 + ω σ2 + ω
and, therefore, σ1 = σ2 , which shows the uniqueness of the principal eigen-
value.
The uniqueness of the principal eigenfunction follows by rather similar

patterns. Indeed, let φ1 , φ2 ∈ Pe \ {0} such that
{
Lφj = σ[L, B, Ω]φj in Ω,
Bφj = 0 on ∂Ω,
for j = 1, 2. Then, for sufficiently large ω ≥ ω0 and j = 1, 2, we have that
{
(L + ω) φj = (σ[L, B, Ω] + ω) φj > 0 in Ω,
Bφj = 0 on ∂Ω,
and hence,
1
Rω φj := φj .
σ[L, B, Ω] + ω
By Corollary 7.1(b), we have that
1
spr Rω = .
σ[L, B, Ω] + ω
Therefore,
Rω φj := spr Rω φj , j = 1, 2,
and, consequently, by Corollary 7.1(a), there must exist λ > 0 such that
φ1 = λφ2 ,
which concludes the proof of the uniqueness of the principal eigenfunction.
The regularity of the principal eigenfunction and (7.30) follow straight-
away from Theorem 7.6. This ends the proof.
7.3 Two equivalent weak eigenvalue problems
The next result establishes a bijection between the set of eigenvalues of

problem (7.1) and σ(Rω )\{0}, ω ≥ ω0 . It also shows that any eigenfunction
of (7.1) in the weak sense must be classical.
Proposition 7.4. Let ω ≥ ω0 and

[ ]
φ ∈ WΓ1,2
0
(Ω) ∩ Ce (Ω̄) + i WΓ1,2
0
(Ω) ∩ Ce (Ω̄) , φ ̸= 0.
Then, the following assertions hold:
(a) If τ ∈ C is an eigenvalue of (7.1) and φ is a weak solution of (7.1),
then, φ is a classical solution of (7.1), τ + ω ̸= 0, and
1
∈ σ(Rω ) \ {0}.
τ +ω
(b) Conversely, if σ ∈ σ(Rω ) \ {0} and φ is an eigenfunction of Rω asso-

ciated to σ, then, φ is a classical solution of (7.1) with
τ = σ −1 − ω.
In particular, if we denote by Σ(L, B, Ω) the set of weak eigenvalues of
(7.1), then, the map
Σ(L, B, Ω) −→ σ(Rω ) \ {0}
τ 7→ 1
τ +ω
is a bijection, and all associated eigenfunctions are classical.
Proof. Suppose τ ∈ C and φ is a weak solution of (7.1). Then, φ is a
weak solution of{
(L + ω)φ = (τ + ω)φ ∈ Ce (Ω̄) in Ω,
(7.31)
Bφ = 0 on ∂Ω,
−
and, thanks to Theorem 7.6, φ ∈ C 1,1 (Ω̄) and it is a classical solution of
(7.1). Moreover, it is the unique weak solution of
{
(L + ω)u = f := (τ + ω)φ ∈ Ce (Ω̄) in Ω,
Bu = 0 on ∂Ω,
and hence τ + ω ̸= 0, because φ ̸= 0 and 0 is a weak solution if τ + ω = 0.
Furthermore, by the definition of Rω , it becomes apparent that
1
Rω φ = φ,
τ +ω
which concludes the proof of Part (a).
Conversely, assume that
Rω φ = σ φ
for some σ ̸= 0. Then,
{ by the definition of Rω , φ is a weak solution of
(L + ω)φ = σ −1 φ ∈ Ce (Ω̄) in Ω,
Bφ = 0 on ∂Ω,
−
and hence, thanks to Theorem 7.6, φ ∈ C 1,1 (Ω̄) and it is a classical solution
of this problem. Therefore, φ is a classical solution of (7.1) for τ = σ −1 −ω.
Throughout the rest of this book, we will use the following concept.
Definition 7.1. Given τ ∈ Σ(L, B, Ω), it is said that φ is an eigenfunction
associated to (7.1) if [ ]
φ ∈ WΓ1,2
0
(Ω) ∩ C e ( Ω̄) + i W 1,2
Γ0 (Ω) ∩ C e (Ω̄)
and φ is a weak solution of (7.1).
According to Theorem 7.6 and Proposition 7.4, all these weak eigen-
−
functions are classical solutions of (7.1) lying in C 1,1 (Ω̄).
7.4 Simplicity and dominance of σ[L, B, Ω]
The following concept plays a central role in the theory of nonlinear elliptic
problems involving second order operators.
Definition 7.2 (Simple eigenvalue). Let

τ ∈ R ∩ Σ(L, B, Ω)
for which (7.1) admits a unique eigenfunction φ, up to a multiplicative
constant. Then, τ is said to be a simple eigenvalue of (7.1) if
{
(L − τ )u = φ in Ω,
(7.32)
Bu = 0 on ∂Ω,
does not admit a weak solution u ∈ WΓ1,2
0
(Ω) ∩ Ce (Ω̄).
As for every ω ≥ ω0 , any weak solution u ∈ WΓ1,2

0
(Ω) ∩ Ce (Ω̄) of (7.32)
must be a weak solution of
{
(L + ω)u = (ω + τ )u + φ ∈ Ce (Ω̄) in Ω,
Bu = 0 on ∂Ω,
due to Theorem 7.6, these weak solutions are classical solutions of (7.32).
Now, we can state the main result of this section.
Theorem 7.8 (Dominance of σ0 ). The principal eigenvalue

σ0 := σ[L, B, Ω]
is a simple eigenvalue of (7.1). Moreover, it is dominant, in the sense
that
Re τ ≥ σ0 for all τ ∈ Σ(L, B, Ω). (7.33)
In particular,
τ > σ0 if τ ∈ R ∩ [Σ(L, B, Ω) \ {σ0 }] .
Proof. Let φ ∈ Pe \ {0} be a principal eigenfunction of (L, B, Ω) and

suppose
{
(L − σ0 )u = φ in Ω,
(7.34)
Bu = 0 on ∂Ω,
admits a weak solution u ∈ WΓ1,2
0
(Ω) ∩ Ce (Ω̄). Let ω ≥ ω0 (see Theorem
7.6) such that
ω + σ0 > 0.
Clearly, u must be a weak solution of

{
(L + ω)u = (ω + σ0 )u + φ in Ω,
Bu = 0 on ∂Ω,
and hence,
u = (ω + σ0 )Rω u + Rω φ. (7.35)
On the other hand, we find from Corollary 7.1(a)(b) that
1
Rω φ = spr Rω φ = φ
σ0 + ω
and, so, dividing (7.35) by ω + σ0 we are driven to
1
(spr Rω I − Rω ) u = φ.
(ω + σ0 )2
Consequently,
φ ∈ R [spr Rω I − Rω ] .
As this contradicts Corollary 7.1(f), σ0 must be a simple eigenvalue of (7.1).
Now, we will prove (7.33). Let
τ ∈ Σ(L, B, Ω) \ {σ0 }.
Then, according to Proposition 7.4,
1
∈ σ(Rω ) for all ω ≥ ω0 ,
τ +ω
and hence, by Corollary 7.1(c), we find that
1 1
< spr Rω = for all ω ≥ ω0 .
τ +ω σ0 + ω
Equivalently,
(Re τ + ω)2 + (Im τ )2 > (σ0 + ω)2 for all ω ≥ ω0 .
Thus, rearranging terms, it is apparent that
(Re τ )2 + 2ω(Re τ − σ0 ) + (Im τ )2 − σ02 > 0
for all ω ≥ ω0 , and, therefore, dividing by ω and letting ω → ∞ we obtain
that
Re τ ≥ σ0 .
This proves (7.33). The fact that any real eigenvalue τ ̸= σ0 must satisfy
τ > σ0 is a byproduct of (7.33). The proof is complete.
The next result sharpens Theorem 7.8.
Theorem 7.9 (Strict dominance of σ0 ). The principal eigenvalue σ0 is

strictly dominant, in the sense that
Re τ > σ0 for all τ ∈ Σ(L, B, Ω) \ {σ0 }. (7.36)
7.4.1 Proof of the strict dominance in case Γ0 = ∅

Suppose Γ0 = ∅, pick an eigenvalue
τ ∈ Σ(L, B, Ω) \ {σ0 },
and let u be an eigenfunction associated to (7.1), in the sense of Definition
7.1. By Theorems 4.10 and 5.11,
u ∈ W 2,p (Ω) for all p > N.
Now, consider the auxiliary function
u
v := in Ω, (7.37)
φ0
where φ0 ≫ 0 is a principal eigenfunction associated to σ0 . By (7.30),
φ0 (x) > 0 ∀ x ∈ Ω̄,
− −
because ∂Ω = Γ1 . Hence, v ∈ W 2,∞ (Ω) and, in particular, v ∈ C 1,1 (Ω̄)
and it is twice classically differentiable almost everywhere in Ω̄. Arguing as
in the proof of Theorem 1.7, a direct calculation from (7.37) shows that
τ u = Lu = φ0 L0 v in Ω, (7.38)
where
( )
∑
N
∂2 ∑N
2 ∑
N
∂φ0 ∂ Lφ0
L0 := − aij + bj − aij +
i,j=1
∂x i ∂x j j=1
φ 0 i=1
∂x i ∂x j φ0
( )
∑N
∂2 ∑N
2 ∑
N
∂φ0 ∂
=− aij + bj − aij + σ0 .
i,j=1
∂x i ∂x j j=1
φ 0 i=1 ∂x i ∂x j
Thus, dividing by φ0 the identity (7.38), yields to

L0 v = τ v
and hence,
(L0 − σ0 ) v = (τ − σ0 )v and (L0 − σ0 ) v̄ = (τ̄ − σ0 )v̄,
where τ̄ and v̄ stand for the complex conjugates of τ and v, respectively.
Therefore, by a chain of direct calculations, we are driven to the identities
(L0 − σ0 ) |v|2 = (L0 − σ0 ) (vv̄)
∑
N
∂v ∂v̄
= v̄ (L0 − σ0 ) v + v (L0 − σ0 ) v̄ − 2 aij
i,j=1
∂xi ∂xj
∑
N
∂v ∂v̄
= v̄ (τ − σ0 ) v + v (τ̄ − σ0 ) v̄ − 2 aij
i,j=1
∂xi ∂xj
and, consequently,
∑
N
∂v ∂v̄
(L0 − σ0 ) |v|2 = 2 (Re τ − σ0 ) |v|2 − 2 aij . (7.39)
i,j=1
∂xi ∂xj
On the other hand, setting

v = ψ + iξ,
we have that
∑
N ∑N ( )( )
∂v ∂v̄ ∂ψ ∂ξ ∂ψ ∂ξ
aij = aij +i −i
i,j=1
∂xi ∂xj i,j=1
∂xi ∂xi ∂xj ∂xj
∑
N [ ( )]
∂ψ ∂ψ ∂ξ ∂ξ ∂ξ ∂ψ ∂ξ ∂ψ
= aij + +i −
i,j=1
∂xi ∂xj ∂xi ∂xj ∂xi ∂xj ∂xj ∂xi
∑
N ( )
∂ψ ∂ψ ∂ξ ∂ξ
= aij + ,
i,j=1
∂xi ∂xj ∂xi ∂xj
because aij = aji for all i, j ∈ {1, ..., N }. Thus, by the strong ellipticity of
L, we find that
∑
N
∂v ∂v̄ ( )
aij ≥ µ |∇ψ|2 + |∇ξ|2 , (7.40)
i,j=1
∂xi ∂xj
where µ > 0 is the ellipticity constant of L in Ω.

To show that Re τ > σ0 we proceed by contradiction. Suppose
Re τ ≤ σ0 .
Then, owing to (7.33), we must have
Re τ = σ0
( ) ( )
(L0 − σ0 ) −|v|2 ≥ 2µ |∇ψ|2 + |∇ξ|2 . (7.41)
On the other hand, on ∂Ω, we have that
∂v ∂v
0 = Bu = B(φ0 v) = φ0 + vBφ0 = φ0
∂ν ∂ν
and hence,
∂v
=0 on ∂Ω, (7.42)
∂ν
because φ0 (x) > 0 for all x ∈ Ω̄. Consequently, setting

ϕ := −|v|2 in Ω̄, (7.43)
we obtain from (7.41) and (7.42) that
∂ϕ
(L0 − σ0 ) ϕ ≥ 0 in Ω and
= 0 on ∂Ω. (7.44)
∂ν
As the zero order term of L0 − σ0 vanishes and ϕ ≤ 0 in Ω̄, we find from
Theorem 7.2 that
m := inf ϕ ≤ 0
Ω̄
cannot be reached in Ω, unless
ϕ=m in Ω̄. (7.45)
Also, by Theorem 7.3, if m is attained at some x ∈ ∂Ω, then
∂ϕ
(x) < 0,
∂ν
which is impossible, by (7.44). Therefore, (7.45) holds and, hence, going
back to (7.41), it becomes apparent that
( )
0 = (L0 − σ0 ) ϕ ≥ µ |∇ψ|2 + |∇ξ|2 .
Consequently,
∇ψ = ∇ξ = 0 in Ω̄,
and, so, v ̸= 0 is constant in Ω. Therefore,
σ0 v = L0 v = τ v
and hence, σ0 = τ , which is impossible, because we have taken τ ̸= σ0 from
the beginning. This contradiction concludes the proof of (7.36).
7.4.2 Proof of the strict dominance in case Γ1 = ∅

Suppose Γ1 = ∅. Then B = D is the Dirichlet boundary operator.
Subsequently, we consider the product open set Ω̃ := Ω × Ω, with points
(x, y) ∈ Ω̃ for all x, y ∈ Ω, the second order differential operator
∑N ( )
∂2 ∂2
L̃ := − aij (x) + aij (y)
i,j=1
∂xi ∂xj ∂yi ∂yj
N (
∑ )
∂ ∂
+ bj (x) + bj (y) + c(x) + c(y)
j=1
∂xj ∂yj
in Ω̃, and the boundary operator

D̃u = u|∂ Ω̃ for all u ∈ C(Ω × Ω).
Obviously, (L̃, D̃, Ω̃) fits into the general setting of this chapter, as (L, D, Ω),
and, in particular, it possesses a principal eigenvalue, denoted by
σ̃0 := σ[L̃, D̃, Ω̃].
Pick a principal eigenfunction φ0 ≫ 0 associated to σ0 and consider the
product function defined by
φ̃0 (x, y) := φ0 (x)φ0 (y), (x, y) ∈ Ω × Ω.
Then,
φ̃0 = 0 on ∂ Ω̃,
and
L̃φ̃0 (x, y) = φ0 (y)Lφ0 (x) + φ0 (x)Lφ0 (y) = 2σ0 φ̃0 (x, y)
for all x, y ∈ Ω. Therefore, by the uniqueness of the principal eigenvalue
already established by Theorem 7.7, we have that
σ̃0 = 2σ0 .
Now, pick
τ ∈ Σ (L, B, Ω) \ {σ0 }
and, given any eigenfunction φ := ψ + iξ ̸= 0 associated to τ , consider the
auxiliary function φ̃ defined by
φ̃(x, y) := φ(x)φ̄(y) + φ̄(x)φ(y)
= 2 Re [φ(x)φ̄(y)]
= 2[ψ(x)ψ(y) + ξ(x)ξ(y)]
for all x, y ∈ Ω̄. Then, φ̃ is a real function such that φ̃ = 0 on ∂ Ω̃ and

L̃φ̃(x, y) = φ(x)Lφ̄(y) + φ̄(y)Lφ(x) + φ(y)Lφ̄(x) + φ̄(x)Lφ(y)
= τ̄ φ(x)φ̄(y) + τ φ̄(y)φ(x) + τ̄ φ(y)φ̄(x) + τ φ̄(x)φ(y)
= (τ + τ̄ )φ(x)φ̄(y) + (τ + τ̄ )φ(y)φ̄(x)
= (τ + τ̄ )φ̃(x, y) = 2 Re τ φ̃(x, y)
for all x, y ∈ Ω. As, for every x ∈ Ω,
φ̃(x, x) = 2φ(x)φ̄(x) = 2|φ(x)|2 ≥ 0
and φ ̸= 0, we have that φ̃ ̸= 0. Thus, 2 Re τ is an eigenvalue of (L̃, D̃, Ω̃)

with associated eigenfunction φ̃ ̸= 0.
Suppose φ̃ > 0 in Ω̃. Then, according to Theorem 7.7,
2 Re τ = σ̃0 = 2σ0
and, consequently,
Re τ = σ0 .
Moreover, there must exist λ ∈ C such that
φ̃ = λφ̃0 .
Without loss of generality, we can take λ = 2 by changing φ by an appro-
priate multiple, if necessary. Hence, by the definitions of φ̃ and φ̃0 ,
φ(x)φ̄(y) + φ̄(x)φ(y) = 2φ0 (x)φ0 (y) ∀ x, y ∈ Ω,
and, consequently, at x = y yields
|φ(x)| = φ0 (x) ∀ x ∈ Ω.
Thus, there exists θ ∈ C(Ω, [0, 2π)) such that
φ(x) = eiθ(x) φ0 (x), ∀ x ∈ Ω.
Going back to the previous identity in Ω̃, we obtain that
eiθ(x) φ0 (x)e−iθ(y) φ0 (y) + e−iθ(x) φ0 (x)eiθ(y) φ0 (y) = 2φ0 (x)φ0 (y)
and so,
cos (θ(x) − θ(y)) = 1 ∀ x, y ∈ Ω.
Consequently, θ must be constant, say θ0 ∈ [0, 2π), and hence,
φ = eiθ0 φ0 .
Thus,
τ φ = Lφ = eiθ0 Lφ0 = eiθ0 σ0 φ0 = σ0 φ
and, therefore, τ = σ0 , which is a contradiction. As this contradiction
comes from the assumption that φ̃ > 0 in Ω̃, necessarily, φ̃ changes sign in
Ω̃. Then, by Theorem 7.7,
[ ( ) ]
2 Re τ ∈ R ∩ Σ L̃, B̃, Ω̃ \ {σ̃0 }
and, therefore, thanks to Theorem 7.8, we find that
2 Re τ > σ̃0 = 2σ0 ,
which implies Re τ > σ0 and ends the proof.
7.4.3 Proof of the strict dominance in the general case

According to Theorem 7.8, any eigenvalue
τ ∈ Σ(L, B, Ω) \ {σ0 }
satisfies
Re τ ≥ σ0 .
Suppose that, for some of these eigenvalues,
Re τ = σ0 .
Then, Im τ ̸= 0 and it follows from Theorem 2.4 of G. Greiner [88] that
σ0 + ik Im τ ∈ Σ(L, B, Ω)
for all k ∈ Z. But this is impossible, because Σ(L, B, Ω) is contained in
a symmetric sector around the real axis in the complex plane with a total
angle less than π, because an appropriate Lp realization of −L generates a
holomorphic semigroup in Lp (Ω) and, therefore, L is a sectorial operator,
as discussed by D. Henry [93] (see Section II.3 of J. López-Gómez [134]
for a rather general proof of this fact in the special case Γ1 = ∅, and the
argument on p. 41 of H. Amann [11] for the general case).
7.5 The strong maximum principle
This section establishes the main result of this chapter. To state it, we need
to introduce some preliminary concepts.
Definition 7.3.
(a) A function h ∈ W 2,p (Ω), p > N , is said to be a supersolution of
(L, B, Ω) if
{
Lh ≥ 0 in Ω,
Bh ≥ 0 on ∂Ω.
The function h is said to be a strict supersolution of (L, B, Ω) if, in
addition, some of these inequalities is strict on a measurable set with
positive measure.
(b) It is said that (L, B, Ω) satisfies the strong maximum principle
(SMP) if any supersolution u ∈ W 2,p (Ω), p > N , u ̸= 0, of (L, B, Ω)
— in particular, any strict supersolution — satisfies
∂u
u(x) > 0 ∀ x ∈ Ω ∪ Γ1 and (x) < 0 ∀ x ∈ u−1 (0) ∩ Γ0 .
∂ν
(c) It is said that (L, B, Ω) satisfies the maximum principle (MP) if any
supersolution u ∈ W 2,p (Ω), p > N , of (L, B, Ω) satisfies u(x) ≥ 0 for
all x ∈ Ω̄.
The main theorem of this chapter reads as follows.
Theorem 7.10 (of characterization of the SMP). The following as-

sertions are equivalent:
i) σ0 := σ[L, B, Ω] > 0.
−
ii) (L, B, Ω) possesses a positive strict supersolution h ∈ W 2,∞ (Ω).
iii) (L, B, Ω) satisfies the strong maximum principle.
iv) (L, B, Ω) satisfies the maximum principle.
v) The resolvent of the linear boundary value problem
{
Lu = f ∈ Ce (Ω̄) in Ω,
(7.46)
Bu = 0 on ∂Ω,
subsequently denoted by R0 : Ce (Ω̄) → Ce (Ω̄), is well defined and it is

strongly positive.
Proof. Suppose σ0 > 0, and let h = φ0 , where φ0 > 0 is a principal

eigenfunction associated to σ0 . Then, Bh = 0 on ∂Ω,
Lh = σ0 h > 0 in Ω,
and, according to Theorems 4.10 and 5.11, h ∈ W 2,p (Ω) for all p > N .
−
Thus, h ∈ W 2,∞ (Ω) and it provides us with a positive strict supersolution
of (L, B, Ω). Therefore, i) implies ii).
−
Suppose (L, B, Ω) has a positive strict supersolution h ∈ W 2,∞ (Ω) and
let u ∈ W 2,p (Ω) \ {0}, p > N , be a supersolution of (L, B, Ω). Then, the
assumptions of Theorem 7.5 are fulfilled. Hence, some of the alternatives
A1, A2, or A3, of Theorem 7.4 holds. As u ̸= 0, Alternative 1 cannot occur.
As h is a strict supersolution, Alternative 3 cannot occur either. There-
fore, Alternative 2 occurs and, consequently, ii) implies iii). Obviously, iii)
implies iv).
Suppose iv) and σ0 ≤ 0. Then,
L(−φ0 ) = −σ0 φ0 ≥ 0 in Ω
and B(−φ0 ) = 0 on ∂Ω. Thus, according to iv), we should have −φ0 ≥ 0,

which is a contradiction. This contradiction concludes the proof of the
equivalence of i), ii), iii) and iv).
Now, suppose σ0 > 0 and pick ω > max{ω0 , 0}. Then, Lu = f if, and
only if,
(L + ω)u = ωu + f
and hence, u solves (7.46) if and only if
( )
1 1
− Rω u = Rω f.
ω ω
On the other hand, as σ0 > 0,
1 1
spr Rω = < ,
σ0 + ω ω
and, consequently, by Corollary 7.1(d), the resolvent operator
( )−1
1 ( )
− Rω ∈ L Ce (Ω̄)
ω
is well defined and strongly positive. In particular, the resolvent of the
problem (7.46) must be given through
( )−1
1 1
R0 := − Rω Rω . (7.47)
ω ω
As, according to Proposition 7.3, Rω is also strongly positive, it follows
from (7.47) that R0 must be strongly positive. Consequently, i) implies v).
Conversely, if R0 is well defined and it is strongly positive, then, h :=
R0 f > 0 for all f > 0 and, hence, any of these functions provides us with an
admissible positive strict supersolution of (L, B, Ω). Therefore, v) implies
ii). This ends the proof.
7.6 The classical minimum principles revisited
Throughout this section we suppose that B = D is the Dirichlet boundary

operator and, as usual, we denote
σ0 := σ[L, D, Ω].
The next consequence from Theorem 7.10 shows that the assumption that
L admits a superharmonic function h in Ω such that
h(x) > 0 for all x ∈ Ω̄
in the generalized minimum principle of M. H. Protter and H. F. Weinberger
(Theorem 1.7) is equivalent to the positivity of σ0 .
Corollary 7.2. Suppose B = D. Then, conditions i)–v) of Theorem 7.10

are equivalent to the next two:
−
vi) (L, D, Ω) admits a positive supersolution h ∈ W 2,∞ (Ω) such that
h(x) > 0 for all x ∈ Ω̄.
−
vii) (L, D, Ω) admits a positive strict supersolution h ∈ W 2,∞ (Ω) such
that h = 0 on ∂Ω.
Proof. Suppose σ0 > 0. Then, by Theorem 7.10, the unique solution of

{
Lh = 0 in Ω,
h=1 on ∂Ω,
provides us with a strict supersolution satisfying vi). Note that
h = 1 − R0 c,
where c is the zero order term of L. Moreover, any principal eigenfunction
φ0 > 0 provides us with a positive strict supersolution satisfying vii).
Conversely, under any of the conditions vi) or vii), h provides us with
a positive strict supersolution of (L, D, Ω) and hence, thanks to Theorem
7.10, σ0 > 0. The proof is complete.
When c ≥ 0, the constant function h := 1 provides us with a superso-
lution satisfying condition vi), and hence σ0 > 0. Consequently, the next
result provides us with a substantial generalization of the classical theorems
of E. Hopf (Theorems 1.2, 1.3), and M. H. Protter and H. F. Weinberger
(Theorem 1.7).
Theorem 7.11. Suppose σ0 > 0 and u ∈ W 2,p (Ω), p > N , satisfies

Lu ≥ 0 in Ω and inf u ≥ 0. (7.48)
Ω̄
Then, either u = 0, or
∂u
u(x) > 0 ∀x∈Ω and (x) < 0 ∀ x ∈ u−1 (0) ∩ ∂Ω.
∂ν
If, instead of (7.48), u satisfies
Lu ≥ 0 in Ω and inf u < 0, (7.49)
Ω̄
then, for every h ∈ W 2,p (Ω) such that

Lh ≥ 0 in Ω and inf h > 0, (7.50)
Ω̄
the quotient function

u(x)
v(x) := , x ∈ Ω̄, (7.51)
h(x)
satisfies
m := inf v < 0,
Ω̄
∂v
v(x) > m ∀ x ∈ Ω and (x) < 0 ∀ x ∈ v −1 (m) ∩ ∂Ω, (7.52)
∂ν
unless v = m in Ω̄. In particular, for every f ∈ L∞ (Ω), f > 0, the unique
weak solution h of the problem
{
Lh = f in Ω,
(7.53)
h=1 on ∂Ω,
satisfies
u u
inf = inf = inf u < 0, (7.54)
Ω̄ h ∂Ω h
( ) ∂Ω
u(x) > inf u h(x) for all x ∈ Ω, (7.55)

∂Ω
and ( )
∂ u
(x0 ) < 0 for all x0 ∈ u−1 inf u ∩ ∂Ω. (7.56)
∂ν h ∂Ω
Proof. As σ0 > 0, by Theorem 7.10, (L, D, Ω) satisfies the strong maxi-

mum principle. Suppose u ̸= 0 satisfies (7.48). Then, u provides us with a
non-zero non-negative supersolution of (L, B, Ω) and, therefore, thanks to
Theorem 7.5, u satisfies the required properties.
Subsequently, we suppose that u and h satisfy (7.49) and (7.50), re-
spectively. By Theorem 7.10, the solution of (7.53) provides us with one of
these functions h for every f > 0. Note that
h = 1 + R0 (f − c).
Next, we consider the quotient function v defined by (7.51). As
inf u < 0 and inf h > 0,
Ω̄ Ω̄
we have that
m := inf v < 0.
Ω̄
Moreover, by Theorem 7.5, (7.52) holds, unless v = m in Ω̄. In any of these
circumstances, we have that
inf v = inf v. (7.57)
Ω̄ ∂Ω
Subsequently, we fix f > 0, f ∈ L∞ (Ω), and suppose that h is the unique
solution of (7.53). Then, since h = 1 on ∂Ω, we have that v = u on ∂Ω and
hence (7.57) implies (7.54).
Suppose v = m in Ω̄. Then, u = mh and, hence,
0 ≤ Lu = mLu = mf < 0,
which is impossible. Therefore, (7.55) and (7.56) follow from (7.52). The
proof is complete.
Seemingly, the characterization of the strong maximum principle given by

Theorem 7.10 goes back to Theorem 2.1 of J. López-Gómez and M. Molina-
Meyer [148]. Although when [148] appeared, in early 1994, there were
already available a number of preliminary results trying to establish the
hidden connections between the sign of σ0 , the validity of the maximum
principle, the validity of the strong maximum principle, and the existence
of a positive supersolution (see G. Sweers [217], D. G. de Figueiredo and E.
Mitidieri [62, 63], Lemma 3.2 of J. López-Gómez and R. M. Pardo [154],
and J. Fleckinger, J. Hernández and F. de Thélin [64]), the theorem estab-
lishing the equivalence between the following five conditions
• (L, D, Ω) possesses a positive strict supersolution.

• The resolvent of (L, D, Ω) is well defined and it is strongly positive.
• (L, D, Ω) satisfies the strong maximum principle.
• (L, D, Ω) satisfies the maximum principle.
• The principal eigenvalue of (L, D, Ω), denoted by σ0 , is positive.
seems to go back to Theorem 2.1 of [148], not only for a single second order
elliptic operator, but, more generally, for a rather general class of linear
elliptic systems of cooperative type.
Almost simultaneously, but in this case for the scalar operator, without
any regularity constraint on ∂Ω, Theorem 1.1 of H. Berestycki, L. Nirenberg
and S. R. S. Varadhan [27] established that (L, D, Ω) satisfies the maximum
principle if and only if σ0 > 0. Some precursors of this result had already
been given by S. Agmon [3]. They also found that the maximum principle
holds if (L, D, Ω) admits a positive strict supersolution (see Corollary 2.4
of [27]), but the authors did not infer from this result the strong maximum
principle, possibly because of the lack of regularity of ∂Ω. Incidentally, H.
Berestycki, L. Nirenberg and S. R. S. Varadhan [27] might have not realized
the importance of their Corollary 2.4, as it was left outside of Section 1 of
[27], where the main results of their paper were collected.
The fact that the characterization of the strong maximum principle in
terms of the existence of a strict positive supersolution had been left out-
side the general scope of H. Berestycki, L. Nirenberg and S. R. S. Varadhan
[27], prompted J. López-Gómez to give a short proof of Theorem 7.10 for
Dirichlet boundary conditions, in Theorem 2.5 of [137], as the author real-
ized that even the simplest version of Theorem 2.1 of J. López-Gómez and
M. Molina-Meyer [148] for the scalar operator was unknown for the more
recognized specialists. The proof of Theorem 2.5 of [137] is based on Theo-

rem 2 of W. Walter [224]; an old version, for Dirichlet boundary conditions,
of our Theorem 2.4 here. All the materials covered by [137] had already
been delivered by J. López-Gómez in his Ph.D. course on Bifurcation The-
ory at the University of Zürich during the summer semester of 1994 (see
the Acknowledgements of [137]).
As it will become apparent throughout Chapters 8 and 9, from the point
of view of the applications, the most crucial feature of Theorem 7.10 is the
fact that the existence of a positive strict supersolution characterizes the
strong maximum principle, as this is the usual strategy adopted in the
applications to make sure that σ0 > 0, or, equivalently, that the strong
maximum principle holds. This provides Theorem 2.1 of J. López-Gómez
and M. Molina-Meyer [148] with its greatest significance when it is weighted
versus the results of H. Berestycki, L. Nirenberg and S. R. S. Varadhan [27].
Three years later, in March 1997, Theorem 2.4 of H. Amann and J.
López-Gómez [13] generalized Theorem 2.5 of J. López-Gómez [137] up to
cover the general boundary operators considered in this book. Actually, in
[13], not only the existence of a smooth positive strict supersolution, but
also of a positive strict supersolution in W 2,p , p > N , was shown to be
necessary and sufficient for the validity of the strong maximum principle.
The proof of H. Amann and J. López-Gómez relies on Theorem 6.1 of
H. Amann [9]. Naturally, the proof of Theorem 6.1 of [9] is based upon
the construction of a strict supersolution h ∈ W 2,p , p > N , satisfying
h(x) > 0 for all x ∈ Ω̄, which is the classical assumption of the generalized
minimum principle of M. H. Protter and H. F. Weinberger (Theorem 1.7),
together with the minimum principle of J. M. Bony [28] (Theorem 7.1). But,
incidentally, it turns out that the technical details of that construction do
actually rely on the inverse trace theorem of H. Amann [9] (see Lemma 5.1
and Sections 5.5 and 5.7 of [9]) whose proof is, in words of H. Amann [11]
(see p. 17 and Remark 36(a) of [11])
somewhat involved and complicated and perhaps not too transparent.
This lack of transparency prompted J. López-Gómez (see Theorem 6.1 of

[144]) to give a new self-contained elementary proof of Theorem 2.4 of H.
Amann and J. López-Gómez [13] through Theorem 2.4 of Chapter 2. At
the end of the day, this has been the proof of Theorem 7.10 included in this
chapter.
The relevance of Theorem 7.10 has been recognized and, possibly, en-
hanced by the recent monographs of H. Amann [11] and Y. Du [56]. Indeed,
Theorem 13 of H. Amann [11] replaced the W 2,p -supersolutions of Theo-

rem 2.4 of H. Amann and J. López-Gómez [13] by Lp -supersolutions and
extended to a very general class of linear cooperative systems the Theorem
2.1 of J. López-Gómez and M. Molina-Meyer [148], whereas Y. Du [56] ded-
icated Chapter 2, among 7 chapters, to Theorem 7.10. Incidentally, neither
J. López-Gómez and M. Molina-Meyer [148], nor J. López-Gómez [144],
were incorporated to the bibliography of Y. Du [56], however the main the-
orem of Chapter 2 of [56] goes back to [148], and the proof given by Y. Du
in Chapter 2 of [56] is based on J. López-Gómez [144].
The reader should be aware of the extremely significant fact that the
statements of all these characterization theorems have remained unchanged
since J. López-Gómez and M. Molina-Meyer established Theorem 2.1 of
[148], except for the brilliant, though natural, observation of H. Amann
[11] that the classical strong maximum principle is equivalent to his weak
and very weak formulations.
The results of Section 2 go back to J. M. Bony [28], whose proofs have
been adapted in this book. Theorem 7.1 was later extended by P. L. Lions
[132] to cover the case when p = N .
The proof of the existence and the uniqueness of the principal eigenvalue
established by Theorem 7.7 is deeply indebted with the general theory de-
veloped by H. Amann in his monographs [7] and [8] and by P. H. Rabinowitz
in his course of Paris [185]. The pivotal Proposition 7.2 is attributed to H.
Amann [7] (see Lemma 5.3 there in). Corollary 7.1(d) provides us with a
version of the anti-maximum principle of Ph. Clément and L. A. Peletier
[43] (see P. Takác [219]).
Up to the best of our knowledge, the more pioneering results on principal
eigenvalues for the Dirichlet problem associated to the Laplace equation go
back to J. B. J. Fourier [65], though he possibly never referred to π 2 as the
principal eigenvalue of
{
−φ′′ = τ φ in (0, 1),
φ(0) = φ(1) = 0.
As H. Poincaré [180] established the analyticity of the resolvent of the
Dirichlet problem for the Laplace operator and could then infer from it
the fact that it has discrete spectrum, H. Poincaré surely knew that the
lowest eigenvalue of the problem plays a significant role in potential theory,
as well as J. W. S. Rayleigh [186]. Incidentally, H. Poincaré was teaching
an advanced course on Potential Theory when D. Hilbert visited Paris for
the first time after defending his Ph.D. Thesis. Later, R. Courant and D.
Hilbert [44] clearly realized the importance of the principal eigenvalues in

Mathematical Physics.
The simplicity of the principal eigenvalue, as it has been delivered here,
is indebted to M. G. Crandall and P. H. Rabinowitz [45]. Once chosen
an appropriate realization for (L, B, Ω) on some subspace X ⊂ WΓ1,2 0
(Ω),
the fact that τ is a simple eigenvalue according to Definition 7.2 can be
equivalently stated by simply saying that
N [L − τ ] = span[φ], φ∈
/ R[L − τ ].
Thus, if, in addition, L − τ is a Fredholm operator of index zero, i.e.,
codim R[L − τ ] = dim N [L − τ ] = 1,
then
N [L − τ ] ⊕ R[L − τ ] = X,
and, therefore, τ is indeed an algebraically simple eigenvalue of (7.1) in the
classical sense.
In H. Amann [11], the fact that σ0 is a simple eigenvalue of (7.1) was
stated in Theorem 12 of [11] and proven on p. 40 of [11]. In some moment
of the proof, after (64), it was claimed, and then used, that
The Krein–Rutman theorem guarantees also that there exists an eigenvector
φ of the dual T ′ ∈ L(E ′ ) to the eigenvalue r = spr T satisfying ⟨φ, v⟩ > 0 for
v ∈ E + \ {0}.
But H. Amann [11] did not provide the readers with a reference for this
result and we could not find it in the literature. Thus, apparently, the proof
of the simplicity of the principal eigenvalue of H. Amann [11] contains a gap.
After reading this book, H. Amann sent the author, through an electronic
mail, a short proof of this fact, which is a slight modification of the one
included in the final part of the proof of Theorem 6.3.
The proof of the dominance of σ0 in Theorem 7.8 seems to be new,
whereas the proof of Theorem 7.9 in case Γ0 = ∅ is based on the proof of
Theorem 1 of M. H. Protter and H. F. Weinberger [182], the proof of Theo-
rem 7.9 in case Γ1 = ∅ has been borrowed from H. Berestycki, L. Nirenberg
and S. R. S. Varadhan [27], and the proof of the strict dominance in the
general case has been taken from p. 41 of H. Amann [11]. Incidentally, the
proof on p. 7 of Y. Du [56] is wrong, since
Kϕ ≥ 2(Re λ − λ1 )ϕ
cannot imply Kϕ ≥ 0 if Re λ < λ1 .
Apparently, Theorem 1 of M. H. Protter and H. F. Weinberger [182] is

the first dominance result available in the literature for general nonselfad-
joint operators. Using the notations of this chapter, Theorem 1 of [182]
established that, for any eigenvalue τ of (7.1),
Lu
Re τ ≥ inf (7.58)
Ω u
for all u ∈ C 2 (Ω) ∩ C 1 (Ω̄) such that inf Ω u > 0 and Bu ≥ 0 on ∂Ω. In
particular, Σ(L, B, Ω) must be contained in a half plane. Naturally, if we
approximate a principal eigenfunction φ0 > 0 by those functions, (7.58)
will provide us with
Re τ ≥ σ0 .
Therefore, Re z ≥ σ0 is the optimal half-plane containing the set of eigen-
values of (7.1). Some precursors of these results, for selfadjoint operators,
go back to J. Barta [23], R. J. Duffin [57], J. Hersch [94], W. W. Hooker
[102] and M. H. Protter [181].
In its greatest generality, Theorem 7.9 goes back to Theorem 12.1 of
H. Amann [9]. But, incidentally, according to the comments of H. Amann
after the statement of Theorem 12 on p. 15 of [11], the proof of Theorem
12.1 of [9]
contained a gap since it had not been asserted that the spectrum is nonempty.
Motivated by this, B. de Pagter [173] derived a general theorem on irreducible
compact positive operators on Banach lattices implying that such an operator
has a strictly positive spectral radius. That theorem can be used to fill the gap
(cf. the proof of Theorem 2.2 of H. Amann and J. López-Gómez [13]).
Indeed, in the proof of Theorem 12.1 of H. Amann [9], the book of H. H.

Schaefer [194] was invoked to assert that a positive compact irreducible
linear operator on a Banach lattice has a strictly positive spectral radius,
however this result cannot be found in H. H. Schaefer [194], but it is The-
orem 3 of B. de Pagter [173].
Under Dirichlet boundary conditions, J. P. Gossez and E. Lami-Dozo
[87] also established the strict dominance of σ0 .
The results of Section 7.6 go back to J. López-Gómez [146].
Chapter 8
Properties of the principal eigenvalue
This chapter applies Theorem 7.10 to obtain some useful properties, from
the point of view of the applications, of the principal eigenvalue of (7.1)
σ0 := σ[L, B, Ω].
Among them, count the monotonicity properties of σ0 with respect to c(x),
β(x), and Ω, its continuity and concavity with respect to c(x), its continuous
dependence with respect to the variations of Ω along Γ0 , as well as with
respect to β, and the crucial property that (L, B, Ω) satisfies the strong
maximum principle for sufficiently large β and small |Ω|.
Throughout the rest of this book, we will denote
− ∩
W 2,∞ (Ω) := W 2,p (Ω),
p>1
as in Theorem 5.8,
{ }
WB2,p (Ω) := u ∈ W 2,p (Ω) : Bu = 0 , p > 1,
and
− ∩
WB2,∞ (Ω) := WB2,p (Ω).
p>1
Since β ∈ C(Γ1 ), by Theorem 4.6, we have that

( )
B ∈ L W 2,p (Ω), Lp (∂Ω) for all p > 1.
Thus, WB2,p (Ω) is a closed subspace of W 2,p (Ω) for all p > 1.
Also, we will denote by
φ0 := φ[L,B,Ω] > 0
the principal eigenfunction associated to σ0 , normalized so that
∫
φ20 = 1.
Ω
225
Note that φ0 satisfies (7.30).

Eventually, we shall emphasize the dependence of the boundary operator
B on the weight function β ∈ C(Γ1 ) by setting
B[β] := B.
If not strictly necessary, such dependency will be dropped.
Suppose Γ1 ̸= ∅. Then, for every proper subdomain Ω0 of class C 2 of Ω
with
dist (Γ1 , ∂Ω0 ∩ Ω) > 0, (8.1)
we shall denote by B[β, Ω0 ], or simply by B[Ω0 ], the boundary operator
{
u on ∂Ω0 ∩ Ω,
B[Ω0 ]u := (8.2)
Bu on ∂Ω0 ∩ ∂Ω,
for all u ∈ W 2,p (Ω) with p > 1. If Ω̄0 ⊂ Ω, then ∂Ω0 ⊂ Ω and hence,
B[Ω0 ]u = u
for all u ∈ W (Ω) with p > 1, i.e., B[Ω0 ] becomes the Dirichlet boundary
2,p
operator D in Ω0 .
If Γ1 = ∅, i.e., B = D, then, we define
B[Ω0 ] := D (8.3)
for all subdomain Ω0 ⊂ Ω of class C . Note that (8.2) becomes (8.3) if
2
Γ1 = ∅. Finally, we also allow Ω0 = Ω by setting

B[Ω] := B,
and, throughout the rest of this book, we shall denote
∂
∂ν := and W01,p (Ω) = W∂Ω1,p
(Ω)
∂ν
for all p ≥ 1 (see (4.16)).
8.1 Monotonicity properties
This section applies Theorem 7.10 to obtain some fundamental monotonic-

ity properties of the principal eigenvalue. Among them, its monotonicity
with respect to the coefficients c and β, and with respect to the support
domain Ω.
As an easy consequence of the uniqueness given by Theorem 7.7, for
every subdomain Ω0 ⊂ Ω of class C 2 the following identity holds
σ[L + s, B, Ω0 ] = σ[L, B, Ω0 ] + s for all s ∈ R.
Properties of the principal eigenvalue 227
This identity will be used very often throughout the remaining of this book.
From the point of view of the theory of dynamical systems, the next
result shows that, among all boundary operators B, the Dirichlet operator
D is the one with the strongest stabilizing effects.
Proposition 8.1. Suppose Γ1 ̸= ∅. Then,
σ[L, B[β], Ω] < σ[L, D, Ω] for all β ∈ C(Γ1 ).
Proof. Let φ[L,B,Ω] and φ[L,D,Ω] denote the normalized principal eigen-
functions associated with σ[L, B, Ω] and σ[L, D, Ω], respectively. By (7.30),
we have that
φ[L,B,Ω] (x) > 0 for all x ∈ Ω ∪ Γ1
and hence
Dφ[L,B,Ω] = φ[L,B,Ω] > 0 on ∂Ω.
Thus, φ[L,B,Ω] provides us with a positive strict supersolution of
(L − σ[L, B, Ω], D, Ω)
and, therefore, according to Theorem 7.10, we find that
0 < σ[L − σ[L, B, Ω], D, Ω] = σ[L, D, Ω] − σ[L, B, Ω].
This completes the proof.
The following result shows the monotonicity of the principal eigenvalue
with respect to Ω.
Proposition 8.2. Let Ω0 be a proper subdomain of Ω of class C 2 satisfying
(8.1) if Γ1 ̸= ∅. Then,
σ[L, B, Ω] < σ[L, B[Ω0 ], Ω0 ],
where B[Ω0 ] is the boundary operator defined by (8.2).
Proof. Let φ[L,B,Ω] be the normalized principal eigenfunction associated
with σ[L, B, Ω]. Then, using (7.30), it is easy to see that


 (L − σ[L, B, Ω])φ[L,B,Ω] = 0 in Ω0 ,

φ[L,B,Ω] (x) > 0 if x ∈ ∂Ω0 ∩ Ω,

 φ (x) = 0 if x ∈ ∂Ω0 ∩ Γ0 ,
 [L,B,Ω]
∂ν φ[L,B,Ω] (x) + β(x)φ[L,B,Ω] (x) = 0 if x ∈ ∂Ω0 ∩ Γ1 .
Moreover, ∂Ω0 ∩ Ω ̸= ∅, as Ω0 is a proper subdomain of Ω. Therefore,
φ[L,B,Ω] is a positive strict supersolution of
(L − σ[L, B, Ω], B[Ω0 ], Ω0 )
and, consequently, it follows from Theorem 7.10 that
0 < σ[L − σ[L, B, Ω], B[Ω0 ], Ω0 ] = σ[L, B[Ω0 ], Ω0 ] − σ[L, B, Ω].
The next result establishes the monotonicity of the principal eigenvalue

with respect to the potential.
Proposition 8.3. Let P1 , P2 ∈ L∞ (Ω) such that P1 < P2 . Then,

σ[L + P1 , B, Ω] < σ[L + P2 , B, Ω].
Proof. Set
φ1 := φ[L+P1 ,B,Ω] .
Then, by (7.30),
(L + P2 − σ[L + P1 , B, Ω])φ1 = (P2 − P1 )φ1 > 0
in Ω, and hence φ1 is a positive strict supersolution of
(L + P2 − σ[L + P1 , B, Ω], B, Ω).
Therefore, thanks to Theorem 7.10, we find that
0 < σ[L+P2 −σ[L+P1 , B, Ω], B, Ω] = σ[L+P2 , B, Ω]−σ[L+P1 , B, Ω].
As an immediate consequence, from this result we can get the continuous
dependence of the principal eigenvalue with respect to the potential.
Corollary 8.1. Let Pn ∈ L∞ (Ω), n ≥ 1, be a sequence of potentials such

that
lim Pn = P in L∞ (Ω).
n→∞
Then,
lim σ[L + Pn , B, Ω] = σ[L + P, B, Ω].
n→∞
Proof. For every ϵ > 0 there exists a natural number n(ϵ) ≥ 1 such that
P − ϵ ≤ Pn ≤ P + ϵ in Ω
for all n ≥ n(ϵ). Therefore, owing to Proposition 8.3,
σ[L + P, B, Ω] − ϵ ≤ σ[L + Pn , B, Ω] ≤ σ[L + P, B, Ω] + ϵ
for all n ≥ n(ϵ). This ends the proof.
The next proposition shows the monotonicity of the principal eigenvalue
with respect to β ∈ C(Γ1 ).
Proposition 8.4. Suppose Γ1 ̸= ∅ and let β1 , β2 ∈ C(Γ1 ) with β1 < β2 .

Then,
σ[L, B[β1 ], Ω] < σ[L, B[β2 ], Ω].
Proof. Set
φ1 := φ[L,B[β1 ],Ω] .
Then,
(L − σ[L, B[β1 ], Ω])φ1 = 0 in Ω,
φ1 = 0 on Γ0 , and, due to (7.30),
∂ν φ1 + β2 φ1 = (β2 − β1 )φ1 > 0 on Γ1 .
Thus, φ1 is a positive strict supersolution of
(L − σ[L, B[β1 ], Ω], B[β2 ], Ω)
and, therefore, Theorem 7.10 implies that
0 < σ[L − σ[L, B[β1 ], Ω], B[β2 ], Ω] = σ[L, B[β2 ], Ω] − σ[L, B[β1 ], Ω].
Corollary 8.2. Suppose Γ1 ̸= ∅ and Ω0 is a subdomain of class C 2 of Ω

satisfying (8.1). Let β1 , β2 ∈ C(Γ1 ) with β1 < β2 . Then,
σ[L, B[β1 , Ω], Ω] < σ[L, B[β2 , Ω0 ], Ω0 ]. (8.4)
The same conclusion holds if β1 ≤ β2 and Ω0 ( Ω.
Proof. If Ω = Ω0 , then (8.4) becomes into

σ[L, B[β1 ], Ω] < σ[L, B[β2 ], Ω],
which is guaranteed by Proposition 8.4.
Suppose Ω0 ( Ω. Then, according to Propositions 8.4 and 8.2, we have
that
σ[L, B[β1 , Ω], Ω] ≤ σ[L, B[β2 , Ω], Ω] < σ[L, B[β2 , Ω0 ], Ω0 ].
8.2 Point-wise min-max characterizations
As a consequence from Theorem 7.10, the next point-wise min-max char-

acterization of σ[L, B, Ω] follows.
−
Theorem 8.1. Let P denote the set of functions ψ ∈ W 2,∞ (Ω) such that
ψ(x) > 0 for all x ∈ Ω and Bψ ≥ 0 on ∂Ω. Then,
Lψ Lψ
σ0 := σ[L, B, Ω] = sup inf = max inf . (8.5)
ψ∈P Ω ψ ψ∈P Ω ψ
Proof. Fix λ < σ0 . Then,

σ[L − λ, B, Ω] = σ0 − λ > 0
and hence, according to Theorem 7.10, (L − λ, B, Ω) satisfies the strong
maximum principle. Thus, the problem
{
(L − λ)ψ = 1 in Ω,
Bψ = 0 on ∂Ω,
−
2,∞
has a unique solution in WB (Ω), denoted by ψ1 , and ψ1 is strongly
positive, in the sense that it satisfies (7.30). In particular, ψ1 ∈ P and, so,
P ̸= ∅. As ψ1 (x) > 0 for all x ∈ Ω ∪ Γ1 , we have that
Lψ1
λ< in Ω.
ψ1
Thus,
Lψ1 Lψ
λ ≤ inf ≤ sup inf . (8.6)
Ω ψ1 ψ∈P Ω ψ
As (8.6) holds for every λ < σ0 , it becomes apparent that
Lψ
σ0 ≤ sup inf .
ψ∈P Ω ψ
To prove the equality we argue by contradiction. Suppose
Lψ
σ0 < sup inf .
ψ∈P Ω ψ
Then, there exist ϵ > 0 and ψ ∈ P such that
Lψ(x)
σ0 + ϵ < for all x ∈ Ω.
ψ(x)
Hence,
{
(L − σ0 − ϵ)ψ > 0 in Ω,
Bψ ≥ 0 on ∂Ω,
and, consequently, ψ provides us with a positive strict supersolution of
(L − σ0 − ϵ, B, Ω). Therefore, thanks to Theorem 7.10,
0 < σ[L − σ0 − ϵ, B, Ω] = −ϵ,
which is impossible. This contradiction shows that
Lψ
σ0 = sup inf .
ψ∈P Ω ψ
As the normalized principal eigenfunction φ0 associated with σ0 lies in
2,∞−
WB (Ω) and Lφ0 = σ0 φ0 , we also have that
Lφ0
σ0 = inf ,
Ω φ0
Similarly, the following min-max characterization holds.

−
Theorem 8.2. Let P> denote the set of functions ψ ∈ W 2,∞ (Ω) such
that ψ(x) > 0 for all x ∈ Ω̄ and Bψ ≥ 0 on ∂Ω. Then,
Lψ
σ0 := σ[L, B, Ω] = sup inf . (8.7)
ψ∈P> Ω ψ
Proof. Fix λ < σ0 . Then,

σ[L − λ, B, Ω] = σ0 − λ > 0
and hence, according to Theorem 7.10, (L − λ, B, Ω) satisfies the strong
maximum principle. Subsequently, we consider the auxiliary problem
{
(L − λ)ψ = 1 in Ω,
(8.8)
Bψ = 1 on ∂Ω.
Let h ∈ C 2 (Ω̄) be such that
Bh = 1 on ∂Ω.
Then, the change of variable
ψ =h+w
transforms (8.8) into
{
(L − λ)w = 1 − (L − λ)h in Ω,
Bw = 0 on ∂Ω.
Therefore, according to Theorem 7.10, the function
ψ1 := h + (L − λ)−1 (1 − (L − λ)h)
−
provides us with the unique solution of (8.8) in W 2,∞ (Ω). Moreover,
ψ1 (x) > 0 for all x ∈ Ω̄ and hence ψ1 ∈ P> . Thus, we have that
Lψ1
λ< in Ω
ψ1
and
Lψ1 Lψ
λ ≤ inf ≤ sup inf . (8.9)
Ω ψ1 ψ∈P> Ω ψ
As (8.9) holds for every λ < σ0 , it becomes apparent that

Lψ
σ0 ≤ sup inf .
ψ∈P> Ω ψ
The proof of (8.7) will be completed by contradiction. Suppose

Lψ
σ0 < sup inf .
ψ∈P> Ω ψ
Then, there exist ϵ > 0 and ψ ∈ P> such that
Lψ(x)
σ0 + ϵ <
ψ(x)
for all x ∈ Ω. Hence,
{
(L − σ0 − ϵ)ψ > 0 in Ω,
Bψ ≥ 0 on ∂Ω,
and, consequently, ψ provides us with a positive strict supersolution of
(L − σ0 − ϵ, B, Ω). Therefore, thanks to Theorem 7.10,
0 < σ[L − σ0 − ϵ, B, Ω] = −ϵ,
which is impossible. This contradiction concludes the proof.
8.3 Concavity with respect to the potential
This section establishes the concavity of the map

L∞ (Ω) 7−→ R
P → σ(P ) := σ[L + P, B, Ω]
with respect to the potential P .
Theorem 8.3 (Concavity of σ0 with respect to the potential).

For every P1 , P2 ∈ L∞ (Ω) and t ∈ [0, 1] the following inequality holds
σ(tP1 + (1 − t)P2 ) ≥ t σ(P1 ) + (1 − t) σ(P2 ).
Proof. Subsequently, we set ξ = (ξ1 , ..., ξN ) for all ξ ∈ RN . As L is

strongly uniformly elliptic in Ω, for every x ∈ Ω̄ the bilinear form
∑N
⟨ξ, ψ⟩ := aij (x)ξi ψj , ξ, ψ ∈ RN ,
i,j=1
defines a scalar product in RN . Hence, by the Cauchy–Schwarz inequality,

we find that
∑N
2⟨ξ, ψ⟩ = 2 aij (x)ξi ψj ≤ 2 |ξ| |ψ| ≤ |ξ|2 + |ψ|2
i,j=1
∑
N ∑
N
= aij (x)ξi ξj + aij (x)ψi ψj
i,j=1 i,j=1
for all ξ, ψ ∈ RN and x ∈ Ω̄, where | · | stands for the norm associated to the
scalar product ⟨·, ·⟩. From this inequality, using (4.2) and Corollary 4.1, it
is easy to see that the map
− −
N : W 2,∞ (Ω) → C 0,1 (Ω̄)
defined by
∑
N
∂u ∂u −
N(u) := − aij = −⟨∇u, ∇u⟩, u ∈ W 2,∞ (Ω),
i,j=1
∂xi ∂xj
−
is concave. Indeed, for every u1 , u2 ∈ W 2,∞ (Ω) and t ∈ [0, 1], the following
chain of inequalities holds
N(tu1 + (1 − t)u2 ) = −⟨t∇u1 + (1 − t)∇u2 , t∇u1 + (1 − t)∇u2 ⟩
= t2 N(u1 ) + (1 − t)2 N(u2 ) − 2t(1 − t)⟨∇u1 , ∇u2 ⟩
≥ t2 N(u1 ) + (1 − t)2 N(u2 ) + t(1 − t) (N(u1 ) + N(u2 ))
= tN(u1 ) + (1 − t)N(u2 ).
Thus, the map G defined by
−
G(u) := (L − c)u + c + N(u), u ∈ W 2,∞ (Ω),
is concave, because N(u) is concave and the mapping u 7→ (L − c)u is linear
and hence concave. Our interest in G comes from the fact that, for every
ψ ∈ P> (see Theorem 8.2), the following relationship holds
Lψ
= G(log ψ).
ψ
Subsequently, we consider P1 , P2 ∈ L∞ (Ω), t ∈ [0, 1], and ψ1 , ψ2 ∈ P>
arbitrary. Taking into account that ψ ∈ P> implies ψ, 1/ψ ∈ L∞ (Ω) and
∇ψ ∈ L∞ (Ω, RN ), it is easily seen that ψ1t ψ21−t ∈ P> . Thus,
L(ψ1t ψ21−t )
[L + tP1 + (1 − t)P2 ](ψ1t ψ21−t )/ψ1t ψ21−t = tP1 +(1−t)P2 +
ψ1t ψ21−t
= tP1 + (1 − t)P2 + G(log(ψ1t ψ21−t ))
= tP1 + (1 − t)P2 + G(t log ψ1 + (1 − t) log ψ2 )
≥ t P1 + (1 − t) P2 + t G(log ψ1 ) + (1 − t) G(log ψ2 )
(L + P1 )ψ1 (L + P2 )ψ2
=t + (1 − t)
ψ1 ψ2
(L + P1 )ψ1 (L + P2 )ψ2
≥ t inf + (1 − t) inf
Ω ψ1 Ω ψ2
and, therefore, it follows from Theorem 8.2 that

(L + P1 )ψ1 (L + P2 )ψ2
σ(tP1 + (1 − t)P2 ) ≥ t inf + (1 − t) inf .
Ω ψ1 Ω ψ2
As this inequality holds for all ψ1 , ψ2 ∈ P> , taking the supreme on its
right-hand side with respect to ψ1 and ψ2 yields
σ(tP1 + (1 − t)P2 ) ≥ tσ(P1 ) + (1 − t)σ(P2 ),
8.4 Stability of Ω along the Dirichlet components of ∂Ω
In the next section we shall show that if Γ0 ̸= ∅ and Ω perturbs in such a

way that Γ1 is kept fixed, then σ[L, B, Ω] varies continuously with Ω. This
is why we adopt the following concepts, where the portion of the boundary
Γ1 is kept fixed. If Γ1 varies, then the results might not be true in general.
Definition 8.1 (Convergence of domains). Let Ω0 be a bounded do-

main of RN with boundary
∂Ω0 = Γ00 ∪ Γ1
such that Γ00 ∩ Γ1 = ∅, and Ωn , n ≥ 1, a sequence of bounded domains of
RN with boundaries ∂Ωn = Γn0 ∪ Γ1 of class C 2 such that
Γn0 ∩ Γ1 = ∅ ∀ n ≥ 1.
Then:
(E) It is said that Ωn converges to Ω0 as n → ∞ from its exterior if
Ω0 ⊂ Ωn+1 ⊂ Ωn
for all n ≥ 1 and
∞
∩
Ω̄n = Ω̄0 .
n=1
(I) It is said that Ωn converges to Ω0 as n → ∞ from its interior if

Ωn ⊂ Ωn+1 ⊂ Ω0
for all n ≥ 1 and
∞
∪
Ωn = Ω0 .
n=1
(C) It is said that Ωn converges to Ω0 as n → ∞ if there are two sequences

of bounded domains of class C 2 , say ΩIn and ΩE
n , n ≥ 1, such that Ωn
I
E
converges to Ω0 from its interior, Ωn converges to Ω0 from its exterior,
ΩIn ⊂ Ω0 ∩ Ωn and Ω0 ∪ Ωn ⊂ ΩE
n for all n ≥ 1.
The main result of this section reads as follows. It introduces the concept
of stability and establishes the stability of any smooth domain.
Theorem 8.4. Let Ω0 be a bounded domain of RN with boundary ∂Ω0 =
Γ00 ∪ Γ1 of class C 1 such that Γ00 ∩ Γ1 = ∅. Then, Ω0 is stable along Γ00
in the sense that for any sequence of bounded domains Ωn , n ≥ 1, of class
C 2 converging to Ω0 from its exterior (see Definition 8.1(E)), the following
relation holds
∞
∩
WΓ1,2 1,2
n (Ωn ) = W 0 (Ω0 ),
Γ
(8.10)
0 0
n=1
where
∞
∩ { }
WΓ1,2
n (Ωn ) := u ∈ WΓ1,2
1 (Ω 1 ) : u| Ωn ∈ W 1,2
Γ n (Ω n ), ∀ n ≥ 2 .
0 0 0
n=1
According to (4.16) and Theorem 4.7,

W 1,2 (Ωn )
∞
WΓ1,2
n (Ωn ) = N [TΓn ] = CΓn (Ωn ) ∀ n ≥ 0.
0 0 0
Thus, essentially, Theorem 8.4 shows that, for every u ∈ WΓ1,2

1 (Ω1 ),
0
TΓn u = 0 ∀n≥1 =⇒ TΓ0 u = 0.
0 0
So, it provides us with the stability of the null space N [TΓn ] as n ↑ ∞.

0
The inclusion
∩∞
WΓ1,2
0 (Ω0 ) ⊂ WΓ1,2
n (Ωn )
0 0
n=1
is understood in the sense that the function
{
u in Ω̄0 ,
ũ :=
0 in Ω̄1 \ Ω̄0 ,
lies in WΓ1,2 1,2
n (Ωn ) for all n ≥ 1 if u ∈ W 0 (Ω0 ).
Γ0
0
The proof of Theorem 8.4 is based on the next result of technical nature.
Proposition 8.5. Let Ω be a bounded domain of RN of class C 1 with bound-
ary ∂Ω = Γ0 ∪ Γ1 , Γ0 ∩ Γ1 = ∅, and consider any proper subdomain Ω0 ⊂ Ω
of class C 1 with boundary ∂Ω0 = Γ00 ∪ Γ1 , Γ00 ∩ Γ1 = ∅. Then,
{ }
WΓ1,2
0 (Ω0 ) = u ∈ W 1,2 (Ω) : supp u ⊂ Ω̄0 .
0
As above, the inclusion

{ }
WΓ1,2
0 (Ω0 ) ⊂ u ∈ W 1,2 (Ω) : supp u ⊂ Ω̄0
0
is understood in the sense that the function

{
u in Ω̄0 ,
ũ :=
0 in Ω̄ \ Ω̄0 ,
lies in W 1,2 (Ω) for all u ∈ WΓ1,2
0 (Ω0 ).
0
8.4.1 Proof of Proposition 8.5

Let u ∈ W 1,2 (Ω) be such that
supp u ⊂ Ω̄0 .
Since Γ1 is a common set of components of ∂Ω and ∂Ω0 , and Ω0 is a
subdomain of Ω, setting
U ϵ := Γ1 + Bϵ ,
there exists ϵ > 0 such that
U ϵ ∩ Ω ⊂ Ω0 ,
where Bϵ stands for the ball of radius ϵ centered at the origin. Moreover, ϵ
can be chosen sufficiently small so that
U ϵ ∩ Γ00 = ∅ (8.11)
because Γ00 ∩ Γ1 = ∅. Let η ∈ C0∞ (U ϵ ) such that
ϵ
η(x) = 1 for each x ∈ U 2 .
Since u ∈ W 1,2 (Ω), we have that u ∈ W 1,2 (Ω0 ). Moreover, by Theorem
4.1, there exists a sequence ψn ∈ C ∞ (Ω̄), n ≥ 1, such that
lim ∥ψn − u∥W 1,2 (Ω) = 0. (8.12)
n→∞
Subsequently, we consider the auxiliary functions
ξn := (ηψn )|Ω̄0 , n ≥ 1.
According to (8.11), ξn vanishes in a neighborhood of Γ00 , as η ∈ C0∞ (U ϵ ).
Hence,
ξn ∈ CΓ∞0 (Ω̄0 ) for all n ≥ 1.
0
Moreover, by (8.12), we have that

lim ∥ηψn − ηu∥W 1,2 (Ω) = 0
n→∞
and, in particular,
lim ∥ξn − ηu∥W 1,2 (Ω0 ) = 0. (8.13)
n→∞
On the other hand, since Ω0 is of class C 1 , it is easy to see that it satisfies

the next segment property. For every x ∈ ∂Ω0 there exist a neighborhood
Ux of x and a vector vx ∈ RN \ {0} such that
Ūx ∩ Ω̄0 + tvx ⊂ Ω0 (8.14)
for all t ∈ (0, 1). By the compactness of Γ00 , there exist a natural number
m ≥ 1 and m points xj ∈ Γ00 , 1 ≤ j ≤ m, such that
∪
m
Γ00 ⊂ Uxj .
j=1
Let Um+1 be an open set such that Ūm+1 ⊂ Ω0 for which

C := { Ux1 , ..., Uxm , Um+1 }
ϵ
is a covering of Ω̄0 \ U 2 and set
Uj := Uxj , 1 ≤ j ≤ m.
Thanks to Proposition 2.2, there exists a partition of the unity
{ β1 , ..., βm , βm+1 }
in Ω0 \ Ū subordinated to the covering C. Then, βj ∈ C0∞ (RN ) and
ϵ
2
supp βj ⊂ Ūj for all 1 ≤ j ≤ m + 1, and

∑
m+1
ϵ
βj = 1 in Ω̄0 \ U 2 . (8.15)
j=1
γj := βj (1 − η)u, 1 ≤ j ≤ m + 1,
and consider the translations
γjt := γj (· − tvxj ), 1 ≤ j ≤ m, 0 < t < 1.
By construction, we have that βm+1 ̸= 0 if γm+1 ̸= 0, and hence,
supp γm+1 ⊂ supp βm+1 ⊂ Ūm+1 ⊂ Ω0 .
Consequently,
1,2
γm+1 ∈ W∂Ω 0
(Ω0 )
and, therefore, there exists a sequence ξnm+1 ∈ C0∞ (Ω0 ), n ≥ 1, such that
lim ∥ξnm+1 − γm+1 ∥W 1,2 (Ω0 ) = 0. (8.16)
n→∞
Moreover, as we are assuming that supp u ⊂ Ω̄0 , it is apparent that

supp γj ⊂ supp βj ∩ supp u ⊂ Ūj ∩ Ω̄0
for all 1 ≤ j ≤ m. Thus, thanks to (8.14),
supp γjt ⊂ Ūj ∩ Ω̄0 + tvxj ⊂ Ω0 , 1 ≤ j ≤ m, 0 < t < 1,
and hence,
1,2
γjt ∈ W∂Ω 0
(Ω0 ), 1 ≤ j ≤ m, 0 < t < 1. (8.17)
By the continuity of the translation operator, for each natural number n ≥ 1
there exists tn ∈ (0, 1) such that
1
∥γjtn − γj ∥W 1,2 (Ω0 ) ≤ , 1 ≤ j ≤ m. (8.18)
n
Moreover, thanks to (8.17), for each n ≥ 1 and 1 ≤ j ≤ m, there exists
ξnj ∈ C0∞ (Ω0 ) such that
1
∥γjtn − ξnj ∥W 1,2 (Ω0 ) ≤ . (8.19)
n
Therefore, according to (8.18) and (8.19),
lim ∥ξnj − γj ∥W 1,2 (Ω0 ) = 0, 1 ≤ j ≤ m. (8.20)
n→∞
Now, consider the sequence

∑
m+1
φn := ξn + ξnj , n ≥ 1.
j=1
By construction,
φn ∈ CΓ∞0 (Ω̄0 ) ∀ n ≥ 1.
0
Moreover, by (8.13), (8.16) and (8.20), we have that

∑
m+1 ∑
m+1
lim φn = ηu + βj (1 − η)u = ηu + (1 − η)u βj in W 1,2 (Ω0 ).
n→∞
j=1 j=1
Consequently, according to (8.15), we have that

ϵ
lim φn = u in W 1,2 (Ω0 \ Ū 2 )
n→∞
ϵ
and, since η = 1 in Ū 2 , it becomes apparent that, actually,
lim φn = u in W 1,2 (Ω0 ).
n→∞
Therefore, u ∈ WΓ1,2
0 (Ω0 ), which shows the validity of the inclusion
0
{ }
u ∈ W 1,2 (Ω) : supp u ⊂ Ω̄0 ⊂ WΓ1,2 0 (Ω0 ).
0
To prove the converse inclusion let u ∈ WΓ1,2

0 (Ω0 ). By Theorem 4.7, there
0
exists a sequence
φn ∈ CΓ∞0 (Ω̄0 ), n ≥ 1, (8.21)
0
such that
lim ∥φn − u∥W 1,2 (Ω0 ) = 0. (8.22)
n→∞
In particular,
lim φn = u almost everywhere in Ω̄0 . (8.23)
n→∞
Next, we consider the auxiliary sequence
{
φn in Ω̄0 ,
ψn := n ≥ 1.
0 in Ω̄ \ Ω̄0 ,
∞
ψn ∈ C∂Ω\Γ 1
(Ω̄), n ≥ 1.
Moreover, owing to (8.22), it is apparent that ψn , n ≥ 1, is a Cauchy
sequence in W 1,2 (Ω). Thus, there exists ψ ∈ W 1,2 (Ω) such that
lim ∥ψn − ψ∥W 1,2 (Ω) = 0.
n→∞
In particular,
lim ψn = ψ almost everywhere in Ω̄ (8.24)
n→∞
and hence ψ = 0 in Ω̄\Ω̄0 , since ψn = 0 in Ω̄\Ω̄0 for all n ≥ 1. Consequently,

supp ψ ⊂ Ω̄0 . (8.25)
Also, from (8.23) and (8.24), it is apparent that
ψ=u in Ω̄0 ,
because φn = ψn in Ω̄0 for all n ≥ 1, by definition. Therefore, the extension
function
{
u in Ω̄0 ,
ψ=
0 in Ω̄ \ Ω̄0 ,
satisfies ψ ∈ W 1,2 (Ω) and (8.25). This concludes the proof.
8.4.2 Proof of Theorem 8.4

Let Ωn , n ≥ 1, be a sequence of bounded domains of class C 2 converging
to Ω0 from its exterior in the sense of Definition 8.1 (E). We have to prove
(8.10). If Ω1 = Ω0 , then Ωn = Ω0 for all n ≥ 1 and, therefore, (8.10)
holds true. So, for the rest of the proof we shall assume that Ω0 is a proper
subdomain of Ω1 . Let
∩∞
u∈ WΓ1,2
n (Ωn ).
0
n=1
Then,
u|Ωn ∈ WΓ1,2
n (Ωn ) for all n ≥ 1,
0
by definition. Thus, according to Proposition 8.5,

∞
∩
supp u ⊂ Ω̄n = Ω̄0
n=1
and, therefore, thanks again to Proposition 8.5, it follows that
u ∈ WΓ1,2
0 (Ω0 ).
0
Conversely, let u ∈ WΓ1,2

0 (Ω0 ) and consider the function
0
{
u in Ω̄0 ,
ũ :=
0 in Ω̄1 \ Ω̄0 .
Owing to Proposition 8.5, it is apparent that
ũ ∈ WΓ1,2
n (Ωn )
0
for all n ≥ 1. This completes the proof.
8.5 Continuous dependence with respect to Ω
This section establishes the continuous dependence of σ[L, B, Ω] with re-

spect to a regular class of perturbations of the domain Ω around its Dirichlet
boundary Γ0 .
The continuous dependence of the principal eigenvalue with respect to
the domain is based on the next result, which provides us with the contin-
uous dependence from the exterior.
Theorem 8.5. Suppose

aij ∈ C 1 (Ω̄), bi ∈ C(Ω̄), 1 ≤ i, j ≤ N, (8.26)
and Ω0 is a proper subdomain of Ω with boundary of class C 2 such that

∂Ω0 = Γ00 ∪ Γ1 , Γ00 ∩ Γ1 = ∅.
Let Ωn , n ≥ 1, be a sequence of bounded domains of RN of class C 2 con-
verging to Ω0 from the exterior, in the sense of Definition 8.1 (E), and such
that Ωn ⊂ Ω for all n ≥ 1. For each n ≥ 0, let Bn denote the boundary
operator defined by
{
u on Γn0 := ∂Ωn \ Γ1 ,
Bn u := (8.27)
∂ν u + βu on Γ1 ,
and let φn > 0 be the principal eigenfunction associated to σ[L, Bn , Ωn ]
normalized so that
∥φn ∥W 1,2 (Ωn ) = 1, n ≥ 0. (8.28)
−
2,∞
Then, φ0 ∈ WB 0
(Ω0 ),
lim σ[L, Bn , Ωn ] = σ[L, B0 , Ω0 ]
n→∞
and
lim ∥φn |Ω0 − φ0 ∥W 1,2 (Ω0 ) = 0.
n→∞
Proof. Note that ν := An ∈ C 1 (Γ1 ), because aij ∈ C 1 (Ω̄), 1 ≤ i, j ≤ N ,

and Γ1 is of class C 2 . Also,
Bn [Ωn+1 ] = Bn+1 , Bn [Ω0 ] = B0 , n ≥ 0, (8.29)
where we have used the notation introduced in (8.2).
The existence and the uniqueness of σ[L, Bn , Ωn ] and φn , n ≥ 0, are
guaranteed by Theorem 7.7. By construction, we have that
Ω0 ⊂ Ωn+1 ⊂ Ωn ⊂ Ω
for all n ≥ 0, and hence, by (8.29) and Proposition 8.2, we find that
σ[L, Bn , Ωn ] ≤ σ[L, Bn+1 , Ωn+1 ] ≤ σ[L, B0 , Ω0 ], n ≥ 1.
Thus, the limit
σ E := lim σ[L, Bn , Ωn ] (8.30)
n→∞
is well defined. We have to prove that
σ E = σ[L, B0 , Ω0 ].
Thanks to Theorem 7.7,
− −
φn ∈ WB2,∞
n
(Ωn ) ⊂ W 2,2 (Ωn ) ∩ C 1,1 (Ω̄n ), n ≥ 0.
Now, for every n ≥ 0, let φ̃n denote the function

{
φn in Ωn ,
φ̃n := n ≥ 0.
0 in Ω \ Ωn ,
Since φn ∈ WΓ1,2n (Ωn ), for each n ≥ 0 we have that φ̃n ∈ W
1,2
(Ω). More-
0
over, by (8.28),
∥φ̃n ∥W 1,2 (Ω) = ∥φn ∥W 1,2 (Ωn ) = 1, n ≥ 0. (8.31)
Thus, by Theorem 4.5, there exist a subsequence of φ̃n , n ≥ 1, say
{φ̃nm }m≥1 , and a function φ̃ ∈ L2 (Ω) such that
lim ∥φ̃nm − φ̃∥L2 (Ω) = 0. (8.32)
m→∞
In particular,
lim φ̃nm (x) = φ̃(x) almost everywhere in Ω. (8.33)
m→∞
We claim that
supp φ̃ ⊂ Ω̄0 . (8.34)
Indeed, pick
∞
∩
x ̸∈ Ω̄0 = Ω̄n .
n=1
Then, since Ω̄n , n ≥ 1, is a decreasing sequence of compact sets, there

exists a natural number n0 ≥ 1 such that x ̸∈ Ω̄n for all n ≥ n0 . Thus,
φ̃n (x) = 0, n ≥ n0 ,
and hence,
lim φ̃n (x) = 0 if x ̸∈ Ω̄0 .
n→∞
Therefore, by the uniqueness of the limit in (8.33), we have that
φ̃ = 0 in Ω \ Ω̄0 ,
which shows (8.34). Note that φn (x) > 0 for all x ∈ Ωn ∪ Γ1 and n ≥ 0. In
particular, φn (x) > 0 for every x ∈ Ω0 ∪ Γ1 and n ≥ 0. Hence, it follows
from (8.33) that
φ̃ ≥ 0 in Ω0 . (8.35)
Now, we will analyze the limiting behavior of the traces of φnm , m ≥ 1,
on Γ1 . Since Ω0 is of class C 2 , by Theorem 4.6, the trace operator TΓ1 of
W 1,2 (Ω0 ) on Γ1 is well defined and, actually, according to Remark 5.1,
( 1
)
TΓ1 ∈ L W 1,2 (Ω0 ), W 2 ,2 (Γ1 ) .
For each n ≥ 1, let in denote the canonical injection
in : W 1,2 (Ωn ) → W 1,2 (Ω0 )
defined by
in u = u|Ω0 for all u ∈ W 1,2 (Ωn ).
Then,
∥in ∥L(W 1,2 (Ωn ),W 1,2 (Ω0 )) ≤ 1, n ≥ 1, (8.36)
and, setting
Tn := TΓ1 ◦ in , n ≥ 1,
we find from (8.36) that
∥Tn ∥ 1 ,2 ≤ ∥TΓ1 ∥ 1 ,2 (8.37)

L(W 1,2 (Ωn ),W 2 (Γ1 )) L(W 1,2 (Ω0 ),W 2 (Γ1 ))
for all n ≥ 1. Therefore, the sequence of operators {Tn }n≥1 is uniformly

− −
bounded. Moreover, since φn ∈ W 2,∞ (Ωn ) ⊂ C 1,1 (Ω̄n ) for all n ≥ 1, we
find that
1
φn |Γ1 = Tn φn ∈ W 2 ,2 (Γ1 )
∥φn |Γ1 ∥ 1 ,2 = ∥Tn φn ∥ 1 ,2 ≤ ∥TΓ1 ∥ 1 ,2

W 2 (Γ1 ) W 2 (Γ1 ) L(W 1,2 (Ω0 ),W 2 (Γ1 ))
for all n ≥ 1. Consequently, the sequence of traces {φn |Γ1 }n≥1 is bounded
1
in W 2 ,2 (Γ1 ). Thus, as according to Remark 5.1, the embedding
1
W 2 ,2 (Γ1 ) ,→ L2 (Γ1 )
is compact, because Γ1 is compact, there exist a subsequence of φnm , m ≥ 1,

relabeled by nm , and a function φ∗ ∈ L2 (Γ1 ) such that
lim ∥φnm |Γ1 − φ∗ ∥L2 (Γ1 ) = 0. (8.38)

m→∞
Next, we will show that {φ̃nm }m≥1 is a Cauchy sequence in W 1,2 (Ω). By
(8.32), this entails
lim ∥φ̃nm − φ̃∥W 1,2 (Ω) = 0. (8.39)

m→∞
Indeed, suppose that k and m are natural numbers such that 1 ≤ k ≤

m. Then, Ωnm ⊂ Ωnk and, since L is strongly uniformly elliptic in Ω,
integrating by parts and taking into account that φn = 0 on Γn0 for all
n ≥ 1 shows that
N ∫
∑
2 ∂(φ̃nk − φ̃nm ) ∂(φ̃nk − φ̃nm )
µ∥∇(φ̃nk − φ̃nm )∥L2 (Ω) ≤ aij
i,j=1 Ω ∂xi ∂xj
(∫ ∫
∑
N
∂φnk ∂φnk ∂φnm ∂φnm
= aij + aij
i,j=1 Ωnk ∂xi ∂xj Ωnm ∂xi ∂xj
) ∫
∂φnk ∂φnm
−2 aij
Ωnm ∂xi ∂xj
[∫ ( ) ∫ ( )
∑N
∂ ∂φnk ∂ ∂φnm
=− aij φnk + aij φnm
i,j=1 Ωnk ∂xj ∂xi Ωnm ∂xj ∂xi
∫ ( ) ]
∂ ∂φnk
−2 aij φnm
Ωnm ∂xj ∂xi
∑N ∫ ( )
∂φnk ∂φnm ∂φnk
+ aij φnk + φn m − 2 φnm nj ,
i,j=1 Γ1
∂xi ∂xi ∂xi
where µ > 0 stands for the ellipticity constant of L in Ω and
n = (n1 , . . . , nN )
is the outward unit normal on Γ1 . From this relation, taking into account
that φn is the principal eigenfunction associated with σ[L, Bn , Ωn ] for all
n ≥ 0, we find that
2
µ∥∇(φ̃nk − φ̃nm )∥L2 (Ω)
∫
≤ (σ[L, Bnk , Ωnk ]φnk − ⟨b, ∇φnk ⟩ − cφnk ) φnk
Ωnk
∫
+ (σ[L, Bnm , Ωnm ]φnm − ⟨b, ∇φnm ⟩ − cφnm ) φnm
Ωnm
∫
−2 (σ[L, Bnk , Ωnk ]φnk − ⟨b, ∇φnk ⟩ − cφnk ) φnm
Ωnm
N ∫
∑ ( )
∂φnk ∂φnm ∂φnk
+ aij φnk + φnm − 2 φn m nj .
i,j=1 Γ1 ∂xi ∂xi ∂xi
Hence, rearranging terms yields

∫
2
µ∥∇(φ̃nk − φ̃nm )∥L2 (Ω) ≤ σ[L, Bnk , Ωnk ] φnk (φnk − φ̃nm )
Ωnk
∫
+ (σ[L, Bnm , Ωnm ] − σ[L, Bnk , Ωnk ]) φ2nm
Ωnm
∫
+ σ[L, Bnk , Ωnk ] φnm (φnm − φnk )
Ωnm
∫ ∫ (8.40)
+ ⟨b, ∇φnk ⟩(φ̃nm − φnk ) + ⟨b, ∇(φnk − φnm )⟩φnm
Ωnk Ωnm
∫ ∫
+ cφnk (φ̃nm − φnk ) + cφnm (φnk − φnm )
Ωnk Ωnm
N ∫
∑ [ ]
∂φnk ∂(φnm − φnk )
+ aij (φnk − φnm ) + φn m nj .
i,j=1 Γ1 ∂xi ∂xi
Now, we will estimate each of the terms on the right-hand side of (8.40).
By (8.31), the following estimates hold
∥φ̃n ∥L2 (Ω) ≤ 1 , ∥∇φ̃n ∥L2 (Ω) ≤ 1, for all n ≥ 0. (8.41)
Moreover, by construction, σ[L, Bn , Ωn ], n ≥ 1, is increasing and bounded

above by σ[L, B0 , Ω0 ]. Thus, by Hölder inequality, (8.41) implies that
∫
σnk φnk (φnk − φ̃nm ) ≤ |σ0 | ∥φ̃nk − φ̃nm ∥L2 (Ω) , (8.42)
Ωnk
where we denote
σh := σ[L, Bh , Ωh ], h ≥ 0.
Similarly,
∫
(σnm − σnk ) φ2nm ≤ |σnm − σnk |,
Ωnm
∫
σnk φnm (φnm − φnk ) ≤ |σ0 | ∥φ̃nm − φ̃nk ∥L2 (Ω) ,
Ωnm
∫
⟨b, ∇φnk ⟩(φ̃nm − φnk ) ≤ ∥b∥∞ ∥φ̃nm − φ̃nk ∥L2 (Ω) ,
Ωnk
where
v
uN
u∑
∥b∥∞ := max t b2 (x), j
x∈Ω̄
j=1
and
∫
cφnk (φ̃nm − φnk ) ≤ ∥c∥L∞ (Ω) ∥φ̃nm − φ̃nk ∥L2 (Ω) ,
Ωnk
∫
cφnm (φnk − φnm ) ≤ ∥c∥L∞ (Ω) ∥φ̃nm − φ̃nk ∥L2 (Ω) .
Ωnm
In order to estimate the integrals on Γ1 , we will use the identities

∂ν φn + β φn = 0 on Γ1 , n ≥ 0,
where
∑
N
νi := aij nj , 1 ≤ i ≤ N.
j=1
According to them, we have that, on Γ1 ,

∑
N
∂φn ∑ ∂φn N
aij nj = νi = ⟨∇φn , ν⟩ = ∂ν φn = −β φn
i,j=1
∂xi i=1
∂xi
for all n ≥ 0 and hence,

∑
N
∂(φnm − φnk )
aij nj = −β (φnm − φnk ).
i,j=1
∂xi
Therefore,
∑
N ∫ ∫
∂φnk
aij (φnk − φnm ) nj = βφnk (φnm − φnk )
i,j=1 Γ1 ∂xi Γ1
≤ ∥β∥L∞ (Γ1 ) ∥φnk |Γ1 ∥L2 (Γ1 ) ∥(φnk − φnm )|Γ1 ∥L2 (Γ1 )
and
∑
N ∫ ∫
∂(φnm − φnk )
aij φnm nj = βφnm (φnm − φnk )
i,j=1 Γ1 ∂xi Γ1 (8.43)
≤ ∥β∥L∞ (Γ1 ) ∥φnm |Γ1 ∥L2 (Γ1 ) ∥(φnk − φnm )|Γ1 ∥L2 (Γ1 ) .
It remains to estimate the integrals

∫
Imk := ⟨b, ∇(φnk − φnm )⟩φnm . (8.44)
Ωnm
Since bi ∈ C(Ω̄), in order to integrate by parts in (8.44) we must first

approach each of the components bi , 1 ≤ i ≤ N , by a sequence of smooth
coefficients, say bin , n ≥ 1. This regularization can be done as follows.
Fix δ > 0 and consider the δ-neighborhood of Ω
Ωδ := Ω̄ + Bδ (0).
For every 1 ≤ i ≤ N , let b̂i be a continuous extension of bi to RN such that
b̂i ∈ C0 (Ωδ ), ∥b̂i ∥L∞ (RN ) = ∥bi ∥L∞ (Ω) . (8.45)
Now, consider the function
{ 1
e |x|2 −1 if |x| < 1,
ρ(x) :=
0 if |x| ≥ 1,
as well as its associated approximation of the identity
∫
ρn := ( ρ)−1 nN ρ(n ·), n ∈ N \ {0}.
RN
The function ρn satisfies
ρn ∈ C0∞ (RN ), supp ρn ⊂ B̄1/n (0), ρn ≥ 0, ∥ρn ∥L1 (RN ) = 1,
for all n ≥ 1. Moreover, for every 1 ≤ i ≤ N , the new sequence
∫
bin := ρn ∗ b̂i = ρn (· − y)b̂i (y) dy, n ≥ 1,
RN
is of class C0∞ (RN )

and it approximates b̂i uniformly in compact subsets of
RN as n ↑ ∞ (see, e.g., Theorem 8.1.3 of M. Guzmán and B. Rubio [91]).
In particular,
lim ∥bin |Ω − bi ∥L∞ (Ω) = 0, 1 ≤ i ≤ N, (8.46)
n→∞
because b̂i |Ω = bi . Moreover, thanks to (8.45), it follows from Young’s

inequality that
∥bin ∥L∞ (RN ) ≤ ∥ρn ∥L1 (RN ) ∥b̂i ∥L∞ (RN ) = ∥bi ∥L∞ (Ω) (8.47)
and
∂bin ∂ρn
∥ ∥L∞ (RN ) ≤ ∥ ∥ 1 N ∥bi ∥L∞ (Ω) (8.48)
∂xj ∂xj L (R )
for all 1 ≤ i, j ≤ N and n ≥ 1, since

∂bin ∂ρn
= ∗ b̂i .
∂xj ∂xj
On the other hand, we find from the definition of ρn that
(∫ )−1
∂ρn ∂ρ
∥ ∥L1 (RN ) = ρ n∥ ∥ 1 N
∂xj RN ∂xj L (R )
for all 1 ≤ j ≤ N and n ≥ 1, and hence, it is apparent from (8.48) that
(∫ )−1
∂bin ∂ρ
∥ ∥L∞ (RN ) ≤ ρ n∥ ∥ 1 N ∥bi ∥L∞ (Ω) (8.49)
∂xj RN ∂xj L (R )
for all 1 ≤ i, j ≤ N and n ≥ 1.
Now, going back to (8.44), it follows that
∫ ∫
Imk = ⟨b−bn , ∇(φnk −φnm )⟩φnm + ⟨bn , ∇(φnk −φnm )⟩φnm , (8.50)
Ωnm Ωnm
where
bn := (b1n , b2n , ..., bN n ) , n ≥ 1.
We now estimate each of the terms on the right-hand side of (8.50). Sub-
sequently, we will set
∫
Imk,1 := ⟨b − bn , ∇(φnk − φnm )⟩φnm ,
Ωnm
∫
Imk,2 := ⟨bn , ∇(φnk − φnm )⟩φnm .
Ωnm
By Hölder inequality,
∫
|Imk,1 | ≤ (|b − bn | |∇(φnk − φnm )| φnm )
Ωnm
≤ ∥b − bn ∥∞ ∥∇(φ̃nk − φ̃nm )∥L2 (Ω) ∥φ̃nm ∥L2 (Ω) ,

( )1/2
∑
N
2
∥b − bn ∥∞ := max (bi (x) − bin (x)) .
x∈Ω̄
i=1
Thus, according to (8.41), we find that
|Imk,1 | ≤ 2∥b − bn ∥∞ , n ≥ 1. (8.51)

To estimate Imk,2 we first integrate by parts

∑ N ∫
∂(φnk − φnm )
Imk,2 = bin φnm
i=1 Ωnm
∂xi
∑N ∫
∂(bin φnm )
=− (φnk − φnm )
i=1 Ωnm
∂xi
∫
+ φnm (φnk − φnm )⟨bn , n⟩
Γ1
and, then, we estimate
(N )
∑
|Imk,2 | ≤ ∥bin ∥L∞ (RN ) ∥∇φ̃nm ∥L2 (Ω) ∥φ̃nk − φ̃nm ∥L2 (Ω)
i=1
( )
∑N
∂bin
+ ∥ ∥ ∞ N ∥φ̃nm ∥L2 (Ω) ∥φ̃nk − φ̃nm ∥L2 (Ω)
i=1
∂xi L (R )
+ ∥bn ∥∞ ∥φnm ∥L2 (Γ1 ) ∥φnk − φnm ∥L2 (Γ1 ) .
Consequently, taking into account (8.41), (8.47) and (8.49), it becomes
apparent that
(N )
∑
|Imk,2 | ≤ ∥bi ∥L∞ (Ω) ∥φ̃nk − φ̃nm ∥L2 (Ω)
i=1
( )
n ∑N
∂ρ
+∫ ∥ ∥L1 (RN ) ∥bi ∥L∞ (Ω) ∥φ̃nk − φ̃nm ∥L2 (Ω)
RN
ρ i=1
∂xi
+ ∥bn ∥∞ ∥φnm ∥L2 (Γ1 ) ∥φnk − φnm ∥L2 (Γ1 ) .

Therefore, combining this estimate with (8.51) we obtain that
|Imk | ≤ 2∥b − bn ∥∞ + ∥bn ∥∞ ∥φnm ∥L2 (Γ1 ) ∥φnk − φnm ∥L2 (Γ1 )
∑N ( )
n ∂ρ (8.52)
+ 1+ ∫ ∥ ∥L1 (RN ) ∥bi ∥L∞ (Ω) ∥φ̃nk − φ̃nm ∥L2 (Ω)
i=1 RN
ρ ∂xi
for all n ≥ 1.
Now, fix ϵ > 0. Thanks to (8.46), there is n = n(ϵ) ≥ 1 such that
ϵ
2∥b − bn ∥∞ ≤ . (8.53)
6
Subsequently, we fix one of those n’s. According to (8.32), there exists
n0 ≥ 1 such that
∑N ( )
n ∂ρ ϵ
1+ ∫ ∥ ∥L1 (RN ) ∥bi ∥L∞ (Ω) ∥φ̃nk − φ̃nm ∥L2 (Ω) ≤ (8.54)
i=1 RN
ρ ∂x i 6
for all n0 ≤ k ≤ m. By (8.38), n0 ≥ 1 can be enlarged so that

ϵ
∥bn ∥∞ ∥φnm ∥L2 (Γ1 ) ∥φnk − φnm ∥L2 (Γ1 ) ≤ (8.55)
6
for all n0 ≤ k ≤ m. Thus, substituting (8.53), (8.54) and (8.55) into (8.52)
we find that
ϵ
|Imk | ≤ for all n0 ≤ k ≤ m. (8.56)
2
Therefore, substituting all the estimates found in between (8.42) and (8.43),
as well as (8.56), in (8.40), and using (8.30), (8.32), (8.38) and (8.56), it
becomes apparent that n0 can be enlarged, if necessary, so that
2
µ∥∇ (φ̃nk − φ̃nm ) ∥L2 (Ω) ≤ ϵ
for all m ≥ k ≥ n0 , which concludes the proof of (8.39).
According to (8.39), we find from (8.31) that
∥φ̃∥W 1,2 (Ω) = lim ∥φ̃nm ∥W 1,2 (Ω) = 1. (8.57)
m→∞
Moreover, if we denote by TΓΩ1 the trace operator of W 1,2 (Ω) on Γ1 , we also

obtain that
lim φnm |Γ1 = lim TΓΩ1 (φ̃nm ) = TΓΩ1 (φ̃) in L2 (Γ1 )
m→∞ m→∞
and, hence, thanks to (8.38), it is apparent that

TΓΩ1 (φ̃) = φ∗ .
φ := φ̃|Ω0 .
Since φ̃ ∈ W 1,2 (Ω) and, due to (8.34), supp φ̃ ⊂ Ω̄0 , it follows from Propo-
sition 8.5 that
φ̃ ∈ WΓ1,2
0 (Ω0 )
0
and, therefore,
φ := φ̃|Ω0 ∈ WΓ1,2
0 (Ω0 ).
0
Moreover, according to (8.57), we have that

∥φ∥W 1,2 (Ω0 ) = ∥φ̃∥W 1,2 (Ω) = 1,
and, consequently, by (8.35),
φ>0 in Ω0 .
Also, note that, thanks to (8.39),

lim ∥φnm |Ω0 − φ∥W 1,2 (Ω0 ) = 0, (8.58)
m→∞
because φn |Ω0 = φ̃n |Ω0 for all n ≥ 1.
Next, we will show that φ is a weak solution of
{
Lφ = σ E φ in Ω0 ,
(8.59)
B0 φ = 0 on ∂Ω0 ,
where σ E is the limit (8.30). We already know that
−
φn ∈ WB2,∞
n
(Ωn )
provides us with a classical — and, hence, weak — solution of
{
Lφn = σn φn in Ωn ,
Bn φn = 0 on ∂Ωn ,
for all n ≥ 1. Equivalently, ∫
an (φn , ϕ) = σn φn ϕ for all ϕ ∈ CΓ∞n0 (Ω̄n ) (8.60)
Ωn
and any n ≥ 1, where∫ ∫ ∫
an (u, v) := ⟨A∇u, ∇v⟩+ (⟨b, ∇u⟩+cu) v+ βuv dS
Ωn Ωn Γ1
for every u, v ∈ WΓ1,2
n (Ω). Now, pick
0
ξ ∈ CΓ∞0 (Ω̄0 )
0
and set {
ξ in Ω̄0 ,
ϕ = ϕm := m ≥ 1,
0 in Ωnm \ Ω0 ,
in (8.60). Then,
∫ ∫
⟨A∇φnm , ∇ξ⟩ + (⟨b, ∇φnm ⟩ + cφnm ) ξ
Ω0 Ω0
∫ ∫ (8.61)
+ βφnm ξ dS = σnm φnm ξ
Γ1 Ω0
for all m ≥ 1. According to (8.58), letting m → ∞ in (8.61) we find from
the Lebesgue
∫ dominated
∫ convergence theorem∫ that ∫
⟨A∇φ, ∇ξ⟩ + (⟨b, ∇φ⟩ + cφ) ξ + βφξ dS = σ E φξ
Ω0 Ω0 Γ1 Ω0
for all ξ ∈ CΓ∞0 (Ω̄0 ). Therefore, φ > 0 is a weak solution of (8.59). By
0
Theorem 7.7,
σ E := σ[L, B0 , Ω0 ]
and φ must be a principal eigenfunction associated to it. This concludes
the proof of the theorem, as the same compactness argument works out
along any subsequence of Ωn , n ≥ 1.
The following result provides us with the continuous dependence of the

principal eigenvalue from the interior of the domain.
Theorem 8.6. Suppose (8.26), and let Ω0 be a proper subdomain of Ω with

boundary of class C 2 such that
∂Ω0 = Γ00 ∪ Γ1 , Γ00 ∩ Γ1 = ∅.
Let Ωn , n ≥ 1, be a sequence of bounded domains of RN of class C 2 con-
verging from the interior to Ω0 as n → ∞ in the sense of Definition 8.1(I).
For every n ≥ 0, let Bn denote the boundary operator defined by (8.27) and
φn the unique principal eigenfunction associated to σ[L, Bn , Ωn ] for which
2,∞−
(8.28) holds. Then, φ0 ∈ WB 0
(Ω0 ),
lim σ[L, Bn Ωn ] = σ[L, B0 , Ω0 ] and lim ∥φ̃n − φ0 ∥W 1,2 (Ω0 ) = 0,
n→∞ n→∞
where
{
φn in Ωn ,
φ̃n := n ≥ 1.
0 in Ω0 \ Ωn ,
Proof. As in the proof of Theorem 8.5, the existence and the uniqueness
of σ[L, Bn , Ωn ] and φn , n ≥ 0, are guaranteed by Theorem 7.7.
According to Definition 8.1(I), we have that
Ωn ⊂ Ωn+1 ⊂ Ω0 for all n ≥ 1
and hence, by Proposition 8.2, we find that
σ[L, Bn , Ωn ] ≥ σ[L, Bn+1 , Ωn+1 ] ≥ σ[L, B0 , Ω0 ], n ≥ 1,
because
Bn+1 [Ωn ] = Bn , B0 [Ωn ] = Bn , n ≥ 1.
Consequently, the limit
σ I := lim σ[L, Bn , Ωn ]
n→∞
is well defined. One has to prove that

σ I = σ[L, B0 , Ω0 ].
The proof of Theorem 8.5 can be adapted to show that there exist φ̃ ∈
W 1,2 (Ω0 ), φ̃ > 0, and a subsequence φ̃nm , m ≥ 1, of φ̃n , n ≥ 1, such that
lim ∥φ̃nm − φ̃∥W 1,2 (Ω0 ) = 0. (8.62)
m→∞
Since φn ∈ WΓ1,2 1,2

n (Ωn ) for all n ≥ 1, it becomes apparent that φ̃n ∈ W 0 (Ω0 )
Γ0
0
for all n ≥ 1 and, consequently, we can infer from (8.62) that
φ̃ ∈ WΓ1,2
0 (Ω0 ).
0
Similarly, it is easily seen that φ̃ is a weak positive solution of

{
Lφ̃ = σ I φ̃ in Ω0 ,
B0 φ̃ = 0 on ∂Ω0 .
Therefore, owing to Theorem 7.7,
φ̃ = φ0 , σ I = σ[L, B0 , Ω0 ],
which concludes the proof, as the same argument works out along any
subsequence of φn , n ≥ 1.
As an immediate consequence from Theorems 8.5 and 8.6 one can infer
the continuous dependence of the principal eigenvalue with respect to the
domain, which can be stated as follows.
Theorem 8.7. Suppose (8.26), and let Ω0 be a proper subdomain of Ω

with boundary of class C 2 such that
∂Ω0 = Γ00 ∪ Γ1 , Γ00 ∩ Γ1 = ∅.
Let Ωn , n ≥ 1, be a sequence of bounded domains of Ω of class C 2 converging
to Ω0 as n → ∞ in the sense of Definition 8.1(C). For every n ≥ 0, let Bn
denote the boundary operator defined by (8.27). Then,
lim σ[L, Bn , Ωn ] = σ[L, B0 , Ω0 ]. (8.63)
n→∞
Proof. By Definition 8.1(C), there are two sequences of bounded domains

of class C 2 , ΩIn and ΩE
n , n ≥ 1, such that Ωn converges to Ω0 from its
I
interior, Ωn coverges to Ω0 from its exterior, as n → ∞, and

E
ΩIn ⊂ Ω0 ∩ Ωn , Ω0 ∪ Ωn ⊂ ΩE
n, n ≥ 1.
In particular,
ΩIn ⊂ Ωn ⊂ ΩE
n
for all n ≥ 1. Subsequently, for every n ≥ 1, we denote by BE I

n and Bn the
boundary operators defined through (8.27) from ΩE I
n and Ωn , respectively.
Then, owing to Proposition 8.2, we find that
σ[L, BIn , ΩIn ] ≥ σ[L, Bn , Ωn ] ≥ σ[L, BE E
n , Ωn ] (8.64)
for all n ≥ 1, because

Bn [ΩIn ] = BIn and BE
n [Ωn ] = Bn .
Moreover, thanks to Theorems 8.5 and 8.6, it becomes apparent that

lim σ[L, BIn , ΩIn ] = σ[L, B0 , Ω0 ],
n→∞
(8.65)
lim σ[L, BE E
n , Ωn ] = σ[L, B0 , Ω0 ].
n→∞
Finally, letting n → ∞ in (8.64), (8.63) holds from (8.65).
8.6 Continuous dependence with respect to β(x)
This section varies the coefficient β ∈ C(Γ1 ) in the boundary operator

{
D on Γ0 ,
B[β] := B =
∂ν + β on Γ1 ,
to analyze the continuous dependence of
σ(β) := σ[L, B[β], Ω] (8.66)
with respect to β.
Subsequently, we denote by
( )
σ L∞ (Γ1 ), L1 (Γ1 )
the weak ∗ topology of L∞ (Γ1 ). Let β, βn ∈ L∞ (Γ1 ), n ≥ 1. Then,
according to Proposition III.12(i) of H. Brézis [29], it is well known that
( )
lim βn = β in σ L∞ (Γ1 ), L1 (Γ1 )
n→∞
if, and only if,

∫ ∫
lim βn ξ = βξ for every ξ ∈ L1 (Γ1 ). (8.67)
n→∞ Γ1 Γ1
When this occurs, it is simply said that

∗
βn ⇀ β as n → ∞.
Obviously, when
lim ∥βn − β∥L∞ (Ω) = 0,
n→∞
∗
then, βn ⇀ β as n → ∞.
Theorem 8.8. Suppose Γ1 ̸= ∅ and (8.26), and let β, βn ∈ C(Γ1 ), n ≥ 1,

∗
such that βn ⇀ β as n → ∞. For every n ≥ 1, let φn denote the unique
principal eigenfunction associated with σ(βn ) normalized so that
∥φn ∥W 1,2 (Ω) = 1, n ≥ 1. (8.68)
Then,
lim σ(βn ) = σ(β) and lim ∥φn − φ∥W 1,2 (Ω) = 0,
n→∞ n→∞
where φ stands for the unique principal eigenfunction associated with σ(β)
normalized so that ∥φ∥W 1,2 (Ω) = 1.
Proof. The existence and the uniqueness of (σ(βn ), φn ), n ≥ 1, and

∗
(σ(β), φ) are guaranteed by Theorem 7.7. Moreover, since βn ⇀ β as
n → ∞, the Banach–Steinhaus theorem shows the existence of a constant
C > 0 such that
∥βn ∥L∞ (Γ1 ) ≤ C for all n≥1 (8.69)
(see, e.g., Proposition III.12(iii) of H. Brézis [29]). Thus, by Proposition
8.4, we find that
σ(−C) ≤ σ(βn ) ≤ σ(C) (8.70)
for all n ≥ 1 and hence there exist a subsequence of σ(βn ), n ≥ 1, relabeled
by n, and a σ ∞ ∈ R such that
σ ∞ := lim σ(βn ).
n→∞
Moreover, by Theorem 7.7, we also have that
− −
2,∞
φn ∈ WB[β n]
(Ω) ⊂ W 2,2 (Ω) ∩ C 1,1 (Ω̄)
for all n ≥ 1, and, so, since W 1,2 (Ω) is compactly embedded in L2 (Ω),
there exist a subsequence of φn , n ≥ 1, again labeled by n, and a function
φ∞ ∈ L2 (Ω) such that
lim ∥φn − φ∞ ∥L2 (Ω) = 0. (8.71)
n→∞
As the previous argument is valid along any subsequence and, due to The-
orem 7.7, the principal eigenpair (σ(β), φ) is unique, to complete the proof
of this theorem it suffices to show that
σ ∞ = σ(β), φ∞ = φ,
and that, actually,
lim ∥φn − φ∥W 1,2 (Ω) = 0.
n→∞

lim φn = φ∞ a.e. in Ω and φ∞ ≥ 0 in Ω,
n→∞
because φn > 0 for all n ≥ 1.
As far as the traces of φn , n ≥ 1, on Γ1 are concerned, the same argu-
ment of the proof of Theorem 8.5 shows that there exist a subsequence of
φn , n ≥ 1, relabeled by n, and a function φ∗ ∈ L2 (Γ1 ) such that
lim ∥φn |Γ1 − φ∗ ∥L2 (Γ1 ) = 0. (8.72)
n→∞
In particular, there exists a constant C1 > 0 such that
∥φn |Γ1 ∥L2 (Γ1 ) ≤ C1 , n ≥ 1. (8.73)
Next, we will show that φn , n ≥ 1, is a Cauchy sequence in W (Ω). 1,2
Indeed, arguing as in the proof of Theorem 8.5, we find that, for any natural
numbers 1 ≤ k ≤ m and every n ≥ 1,
∫ ∫
2
µ∥∇(φk −φm )∥L2 (Ω) ≤ σ(βk ) φk (φk −φm )+[σ(βm )−σ(βk )] φ2m
∫ Ω
∫ Ω
+ σ(βk ) φm (φm − φk ) + ⟨b, ∇φk ⟩(φm − φk )

∫ Ω Ω
∫
+ ⟨b−bn , ∇(φk −φm )⟩φm + ⟨bn , ∇(φk −φm )⟩φm
(8.74)
∫Ω ∫ Ω
+ cφk (φm − φk ) + cφm (φk − φm )

Ω Ω
N ∫
∑ [ ]
∂φk ∂(φm − φk )
+ aij (φk − φm ) + φm nj ,
i,j=1 Γ1 ∂xi ∂xi
where bn ∈ C0∞ (RN , RN ), n ≥ 1, is any sequence satisfying

lim ∥bn − b∥L∞ (Ω,RN ) = 0.
n→∞
Actually, we suppose that bn , n ≥ 1, has been constructed as in the proof
of Theorem 8.5. Thanks to (8.68), we have that
∥φn ∥L2 (Ω) ≤ 1 and ∥∇φn ∥L2 (Ω) ≤ 1 (8.75)
for all n ≥ 1. Therefore, arguing as in the proof of Theorem 8.5, we find
from (8.70) and (8.75) the following estimates
∫
σ(βk ) φk (φk − φm ) ≤ σ(C)∥φk − φm ∥L2 (Ω) , (8.76)
Ω
∫
[σ(βm ) − σ(βk )] φ2m ≤ |σ(βm ) − σ(βk )|,
Ω
∫
σ(βk ) φm (φm − φk ) ≤ σ(C)∥φm − φk ∥L2 (Ω) ,
Ω
∫
⟨b, ∇φk ⟩(φm − φk ) ≤ ∥b∥∞ ∥φm − φk ∥L2 (Ω) ,
Ω
∫
cφk (φm − φk ) ≤ ∥c∥L∞ (Ω) ∥φm − φk ∥L2 (Ω) ,
Ω
∫
cφm (φk − φm ) ≤ ∥c∥L∞ (Ω) ∥φm − φk ∥L2 (Ω) ,
Ω
∫
⟨b − bn , ∇(φk − φm )⟩φm ≤ 2∥b − bn ∥∞ ,
Ω
and
∫
⟨bn , ∇(φk − φm )⟩φm | ≤ ∥bn ∥∞ ∥φm ∥L2 (Γ1 ) ∥φk − φm ∥L2 (Γ1 )
Ω
N (
∑ ) (8.77)
n ∂ρ
+ 1+ ∫ ∥ ∥L1 (RN ) ∥bi ∥L∞ (Ω) ∥φk − φm ∥L2 (Ω)
i=1 RN
ρ ∂xi
for all n ≥ 1.
In order to estimate the integrals over Γ1 , we will use the identities
∂ν φ n + β n φ n = 0 on Γ1 , n ≥ 1,
where
∑
N
νi := aij nj , 1 ≤ i ≤ N.
j=1
According to them, we have that

∑
N
∂φk ∑ ∂φk N
aij nj = νi = ⟨ν, ∇φk ⟩ = ∂ν φk = −βk φk ,
i,j=1
∂xi i=1
∂xi
∑
N
∂(φm − φk )
aij nj = −βm φm + βk φk .
i,j=1
∂xi
Thus,
N ∫
∑ ∫
∂φk
aij (φk − φm ) nj = βk φk (φm − φk ), (8.78)
i,j=1 Γ1 ∂xi Γ1
N ∫
∑ ∫
∂(φm − φk )
aij φm nj = (βk φk − βm φm )φm . (8.79)
i,j=1 Γ1 ∂xi Γ1
Moreover, owing to (8.69) and (8.73), we obtain that

∫
βk φk (φm − φk ) ≤ C2 ∥(φm − φk )|Γ1 ∥L2 (Γ1 ) , (8.80)
Γ1
for some constant C2 > 0 independent of k and m. Similarly,
∫ ∫ ∫
(βk φk −βm φm )φm ≤ βk φm (φk −φm ) + φ2m (βk − βm )
Γ1 Γ1 Γ1
∫ (8.81)
≤ C2 ∥φm − φk ∥L2 (Γ1 ) + φ2m (βk −βm )
Γ1
and, according to (8.69), we also have that

∫ ∫ ∫
φ2m (βk −βm ) ≤ (φ2m −φ2∗ )(βk −βm ) + φ2∗ (βk −βm )
Γ1 Γ1 Γ1
∫
≤ 2C∥φm +φ∗ ∥L2 (Γ1 ) ∥φm −φ∗ ∥L2 (Γ1 )+ φ2∗ (βk −βm ) .
Γ1
Thus, according to (8.67), (8.72) and (8.73), we find from (8.81) that, for
any ϵ > 0, there exists a natural number n0 ≥ 1 such that
∫
ϵ
(βk φk − βm φm )φm ≤ (8.82)
Γ1 2
for all k, m ≥ n0 . Finally, owing to (8.78) and (8.79) and substituting the
estimates between (8.76) and (8.77), as well as (8.80) and (8.82), in (8.74),
it is easily seen that there exists k0 ≥ n0 such that
2
µ∥∇(φk − φm )∥L2 (Ω) ≤ ϵ for all k , m ≥ k0 .
Thanks to (8.71), this shows that
lim ∥φn − φ∞ ∥W 1,2 (Ω) = 0. (8.83)
n→∞
Subsequently, we denote by TΓ1 the trace operator of W 1,2 (Ω) on Γ1 . Then,

there exists a constant C3 > 0 such that
∥φn − TΓ1 φ∞ ∥L2 (Γ1 ) = ∥TΓ1 (φn − φ∞ ) ∥L2 (Γ1 ) ≤ C3 ∥φn − φ∞ ∥W 1,2 (Ω)
for all n ≥ 1. Hence, (8.83) implies that
lim ∥φn − TΓ1 φ∞ ∥L2 (Γ1 ) = 0 (8.84)
n→∞
and, therefore, thanks to (8.72), we find that

TΓ1 φ∞ = φ∗ .
Similarly, since φn |Γ0 = 0 for all n ≥ 1, by the continuity of the trace

operator TΓ0 of W 1,2 (Ω) on Γ0 we find that TΓ0 φ∞ = 0 and, consequently,
φ∞ ∈ WΓ1,2
0
(Ω).
Moreover, thanks to (8.83), (8.68) guarantees that
∥φ∞ ∥W 1,2 (Ω) = 1.
Thus, since φn > 0 for all n ≥ 1, we have that
φ∞ ≥ 0, φ∞ ̸= 0.
Next, we will show that φ∞ provides us with a weak solution of

{
Lφ∞ = σ ∞ φ∞ in Ω,
(8.85)
B(β)φ∞ = 0 on ∂Ω;
it should be remembered that
σ ∞ = lim σ(βn ).
n→∞
By the assumptions, φn is a weak solution of

{
Lφn = σ(βn )φn in Ω,
B(βn )φn = 0 on ∂Ω,
for all n ≥ 1. Thus, for every ξ ∈ WΓ1,2

0
(Ω), we have that
∫ ∫ ∫ ∫
⟨A∇φn , ∇ξ⟩+ (⟨b, ∇φn ⟩ + cφn )ξ + βn φn ξ = σ(βn ) φn ξ (8.86)
Ω Ω Γ1 Ω
for all n ≥ 1. According to (8.67), we find from (8.84) that

∫ ∫
lim (βn ξφn ) = (βξTΓ1 φ∞ ) .
n→∞ Γ1 Γ1
Therefore, letting n → ∞ in (8.86), the dominated convergence theorem of

Lebesgue implies
∫ ∫ ∫ ∫
⟨A∇φ∞ , ∇ξ⟩+ (⟨b, ∇φ∞ ⟩+ cφ∞ )ξ + βφ∞ ξ = σ ∞ φ∞ ξ
Ω Ω Γ1 Ω
for all ξ ∈ WΓ1,2

0
(Ω) and, therefore, φ∞ ∈ WΓ1,2
0
(Ω) is a weak positive solution
of (8.85). Theorem 7.7 concludes the proof, as it shows that σ ∞ = σ(β)
and that φ∞ is a principal eigenfunction associated to it.
8.7 Asymptotic behavior of σ(β) as min β ↑ ∞

Γ1
This section analyzes the behavior of the principal eigenvalue (8.66) as

minΓ1 β ↑ ∞. Its main result establishes that it converges to the principal
eigenvalue of the Dirichlet problem in Ω and it reads as follows.
Theorem 8.9. Suppose Γ1 ̸= ∅ and (8.26), and let βn ∈ C(Γ1 ), n ≥ 1, be

an arbitrary sequence such that
lim min βn = ∞. (8.87)
n→∞ Γ1
Then,
lim σ[L, B[βn ], Ω] = σ[L, D, Ω]. (8.88)
n→∞
Moreover, if φn stands for the principal eigenfunction associated with

σ(βn ) := σ[L, B[βn ], Ω], n ≥ 1,
normalized so that
∥φn ∥W 1,2 (Ω) = 1, n ≥ 1, (8.89)
then
lim ∥φn − φ0 ∥W 1,2 (Ω) = 0, (8.90)
n→∞
where φ0 is the corresponding principal eigenfunction of σ[L, D, Ω].
Proof. By (8.87), we can assume, without loss of generality, that

ℓn := min βn > 0, n ≥ 1. (8.91)
Γ1
Clearly, according to Propositions 8.1 and 8.4, it follows from (8.91) that
σ(0) := σ[L, B[0], Ω] < σ[L, B[βn ], Ω] < σ[L, D, Ω] (8.92)
for all n ≥ 1. Thus, there exist
σ ∞ ∈ [σ(0), σ[L, D, Ω]]
and a subsequence of βn , n ≥ 1, relabeled by n, such that
σ ∞ := lim σ[L, B[βn ], Ω].
n→∞
To prove (8.88), we should show that

σ ∞ = σ[L, D, Ω]. (8.93)
By Theorem 7.7, we have that

− −
2,∞
φn ∈ WB[β n]
(Ω) ⊂ W 2,2 (Ω) ∩ C 1,1 (Ω̄) (8.94)
for all n ≥ 1, and, in particular, since W 1,2 (Ω) is compactly embedded in

L2 (Ω), there exist a subsequence of φn , n ≥ 1, again labeled by n, and a
function φ ∈ L2 (Ω) such that
lim ∥φn − φ∥L2 (Ω) = 0. (8.95)
n→∞
Necessarily,
lim φn = φ a.e. in Ω and φ ≥ 0.
n→∞
Consequently, to complete the proof of the theorem it suffices to show

(8.93), φ = φ0 and (8.90), since the previous argument is valid along any
subsequence of βn , n ≥ 1.
By the definition of the φn ’s, we have that
Lφn = σ(βn )φn in Ω
for all n ≥ 1. Moreover, owing to (8.89) and (8.92), it becomes apparent
that
∥σ(βn )φn ∥L2 (Ω) ≤ |σ(βn )| ≤ C1 , n ≥ 1, (8.96)
for some constant C1 > 0. On the other hand, thanks to Theorem 5.10,
there exists a constant C2 > 0 such that
∥φn ∥W 2,2 (Ω) ≤ C2 ∥Lφn ∥L2 (Ω)
for all n ≥ 1. Therefore, by (8.96), we find that
∥φn ∥W 2,2 (Ω) ≤ C2 C1 for all n ≥ 1. (8.97)
Subsequently, we denote by j1 and j2 the compact injections
3 1
j1 : W 2 ,2 (Γ1 ) ,→ L2 (Γ1 ), j2 : W 2 ,2 (Γ1 ) ,→ L2 (Γ1 ),
and by
3 1
T1 ∈ L(W 2,2 (Ω), W 2 ,2 (Γ1 )), T2 ∈ L(W 1,2 (Ω), W 2 ,2 (Γ1 )),
the corresponding trace operators on Γ1 . According to (8.94), we obtain
that
3 j1
φn |Γ1 = T1 φn ∈ W 2 ,2 (Γ1 ) ,→ L2 (Γ1 ),
1 j2
∇φn |Γ1 = T2 ∇φn ∈ W 2 ,2 (Γ1 ) ,→ L2 (Γ1 ),
for all n ≥ 1. Hence, it follows from (8.97) that

∥φn |Γ1 ∥L2 (Γ1 ) = ∥j1 T1 φn ∥L2 (Γ1 )
≤ ∥j1 T1 ∥L(W 2,2 (Ω),L2 (Γ1 )) ∥φn ∥W 2,2 (Ω)
≤ ∥j1 T1 ∥L(W 2,2 (Ω),L2 (Γ1 )) C2 C1
for all n ≥ 1. Similarly,
∥∇φn |Γ1 ∥L2 (Γ1 ) ≤ ∥j2 T2 ∥L(W 1,2 (Ω),L2 (Γ1 )) ∥∇φn ∥W 1,2 (Ω)
≤ ∥j2 T2 ∥L(W 1,2 (Ω),L2 (Γ1 )) ∥φn ∥W 2,2 (Ω)
≤ ∥j2 T2 ∥L(W 1,2 (Ω),L2 (Γ1 )) C2 C1
for all n ≥ 1. Consequently, there exists a constant C3 > 0 such that
∥φn |Γ1 ∥L2 (Γ1 ) ≤ C3 and ∥∇φn |Γ1 ∥L2 (Γ1 ) ≤ C3 ∀ n ≥ 1. (8.98)
On the other hand, by (8.91), we find from
∂ν φn = −βn φn on Γ1 , n ≥ 1,
that
(∂ν φn )2 = βn2 φ2n ≥ ℓ2n φ2n on Γ1 , n ≥ 1,
and hence,
φ2n |Γ1 ≤ ℓ−2 2 −2
n (∂ν φn ) |Γ1 ≤ ℓn |ν| |∇φn |Γ1 | ,
2 2
n ≥ 1.
Thus, (8.98) implies that
∥φn |Γ1 ∥L2 (Γ1 ) ≤ ℓ−2 −2
2 2
n |ν| ∥∇φn |Γ1 ∥L2 (Γ1 ) ≤ ℓn |ν| C3 , n ≥ 1,
2 2 2
and, consequently, we find from (8.87) that

lim ∥φn |Γ1 ∥L2 (Γ1 ) = 0. (8.99)
n→∞
In particular,
lim φn |Γ1 = 0 a.e. in Γ1 .
n→∞
Also, by construction, we already know that φn |Γ0 = 0 for all n ≥ 1.

Therefore, (8.99) entails that
lim ∥φn |∂Ω ∥L2 (∂Ω) = 0. (8.100)
n→∞
Now, we show that φn , n ≥ 1, is a Cauchy sequence in W 1,2 (Ω). Combining

this fact with (8.89) and (8.95) yields
lim ∥φn − φ∥W 1,2 (Ω) = 0, ∥φ∥W 1,2 (Ω) = 1. (8.101)
n→∞
Indeed, arguing as in the proof of Theorems 8.5 and 8.8, for every 1 ≤ k ≤ m
the estimate (8.74) holds, as well as the estimates (8.75)–(8.77). Moreover,
by (8.26) and (8.98), there exists a constant C4 > 0 such that
N ∫
∑ ∂φk
aij (φk − φm ) nj ≤ C4 ∥(φk − φm )|Γ1 ∥L2 (Γ1 )
i,j=1 Γ1 ∂xi
and
N ∫
∑ ∂(φm − φk )
aij φm nj ≤ C4 ∥φm |Γ1 ∥L2 (Γ1 ) .
i,j=1 Γ1 ∂xi
Therefore, owing (8.100), for every ϵ > 0 there exists a natural number
n0 = n0 (ϵ) ≥ 0 such that
∑N ∫ [ ]
∂φk ∂(φm − φk ) ϵ
aij (φk − φm ) + φm nj ≤ (8.102)
i,j=1 Γ1
∂xi ∂xi 2
for all k, m ≥ n0 . Finally, substituting (8.76)–(8.77) and (8.102) into (8.74),

it is easily realized that there exists k0 ≥ n0 such that
2
µ∥∇(φk − φm )∥L2 (Ω) ≤ ϵ, k, m ≥ k0 .
This completes the proof of (8.101) and, in particular, it shows that φ ∈
W 1,2 (Ω) and that φ > 0.
We now ascertain the behavior of φ on ∂Ω. Let j denote the compact
injection
1
j : W 2 ,2 (∂Ω) ,→ L2 (∂Ω)
and
1
T ∈ L(W 1,2 (Ω), W 2 ,2 (∂Ω))
the trace operator on ∂Ω. Since φn − φ ∈ W 1,2 (Ω), we have that
1
T (φn − φ) = φn − T φ ∈ W 2 ,2 (∂Ω)
for all n ≥ 1, and hence,
∥φn − T φ∥L2 (∂Ω) ≤ ∥jT ∥L(W 1,2 (Ω),L2 (∂Ω)) ∥φn − φ∥W 1,2 (Ω) .
Thus, by (8.101), we find that
lim ∥φn − T φ∥L2 (∂Ω) = 0
n→∞
and, therefore, thanks to (8.100), we obtain that T φ = 0. Equivalently,

φ ∈ W01,2 (Ω).
Finally, the same argument used in the proofs of Theorems 8.5 and 8.8
shows that φ is a weak positive solution of
{
Lφ = σ ∞ φ in Ω,
φ=0 on ∂Ω.
Therefore, according to Theorem 7.7, we find that
(σ ∞ , φ) = (σ[L, D, Ω], φ0 ).
As an immediate consequence of Theorem 8.9, the next result holds.
Corollary 8.3. Suppose (8.26). Then,

σ[L, D, Ω] = sup σ[L, B[β], Ω].
β∈C(Γ1 )
Proof. Thanks to Proposition 8.1,

sup σ[L, B[β], Ω] ≤ σ[L, D, Ω].
β∈C(Γ1 )
Moreover, according to Theorem 8.9,

lim σ[L, B(n), Ω] = σ[L, D, Ω].
n→∞
8.8 Lower estimates of σ[L, D, Ω] in terms of |Ω|
As usual, we denote by | · | the Lebesgue measure of RN . The main result

of this section establishes that if |Ω| is sufficiently small and minΓ1 β is
sufficiently large, then
σ[L, B, Ω] > 0.
Consequently, according to Theorem 7.10, (L, B, Ω) satisfies the strong
maximum principle. This result is based on Theorem 8.9 and on the lower
estimates of σ[L, D, Ω] in terms of |Ω| given by Proposition 8.6 below.
In the special case when
L = −∆
the celebrated inequality of C. Faber [61] and E. Krahn [117] establishes
that, among all domains with a fixed Lebesgue measure, |Ω|, the ball BR :=
BR (0) has the smallest principal eigenvalue under homogeneous Dirichlet

boundary conditions. In other words, for every R > 0 and Ω such that
|Ω| = |BR | = RN |B1 | (8.103)
the following estimate holds
σ[−∆, D, Ω] ≥ σ[−∆, D, BR ] = σ[−∆, D, B1 ]R−2 . (8.104)
Therefore, setting
Σ := σ[−∆, D, B1 ],
it becomes apparent from (8.103) and (8.104) that
σ[−∆, D, Ω] ≥ Σ|B1 | N |Ω|− N .

2 2
(8.105)
Consequently,
( 2
) 2
lim inf σ[−∆, D, Ω]|Ω| N ≥ Σ|B1 | N . (8.106)
|Ω|↓0
The next result provides us with a substantial extension of these lower

estimates to cover the case when L is a general second order uniformly
elliptic operator in Ω.
Proposition 8.6. Suppose

( √ )N
µ Σ
|Ω| ≤ |B1 |, (8.107)
∥b∥∞
where µ > 0 is the ellipticity constant of L in Ω, and

( N )1/2
∑
∥b∥∞ := sup b2i .
Ω i=1
Then,
√
σ[L, D, Ω] ≥ µΣ|B1 | N |Ω|− N − ∥b∥∞ Σ|B1 | N |Ω|− N + inf c.
2 2 1 1
(8.108)
Ω
In particular,
( 2
) 2
lim inf σ[L, D, Ω]|Ω| N ≥ µΣ|B1 | N , (8.109)
|Ω|↓0
which is optimal in the light of (8.106).

Proof. Let φ > 0 denote a principal eigenfunction of σ[L, D, Ω]. Then,

multiplying by φ the differential equation
Lφ = σ[L, D, Ω]φ,
integrating in Ω and
∫ applying
∫ the formula of∫integration by∫ parts yields
σ[L, D, Ω] φ2 = ⟨A∇φ, ∇φ⟩ + ⟨b, ∇φ⟩φ + cφ2
Ω
∫ Ω
∫ Ω
∫
Ω
≥µ |∇φ| + ⟨b, ∇φ⟩φ + inf c

2
φ2 .
Ω Ω Ω Ω
Moreover,
∫ ∫
⟨b, ∇φ⟩φ ≤ |b||∇φ|φ ≤ ∥b∥∞ ∥φ∥L2 (Ω) ∥∇φ∥L2 (Ω)
Ω Ω
and hence, ∫
⟨b, ∇φ⟩φ ≥ −∥b∥∞ ∥φ∥L2 (Ω) ∥∇φ∥L2 (Ω) .
Ω
Consequently,
∫ ∫ ∫
σ[L, D, Ω] φ2 ≥ µ |∇φ|2 − ∥b∥∞ ∥φ∥L2 (Ω) ∥∇φ∥L2 (Ω) + inf c φ2 ,
Ω Ω Ω Ω
or, equivalently,
( )
∥∇φ∥L2 (Ω) ∥∇φ∥L2 (Ω)
σ[L, D, Ω] ≥ µ − ∥b∥∞ + inf c. (8.110)
∥φ∥L2 (Ω) ∥φ∥L2 (Ω) Ω
On the other hand, according to the variational characterization of the
principal eigenvalue σ[−∆, D, Ω] as ∫
Ω∫
|∇ψ|2
σ[−∆, D, Ω] = inf (8.111)
ψ∈W01,2 (Ω) Ω
ψ2
(see, e.g., p. 336 ∫of L. C. Evans [60]), it follows from (8.105) that
|∇φ|2
Ω∫
≥ σ[−∆, D, Ω] ≥ Σ|B1 | N |Ω|− N .
2 2
2
(8.112)
Ω
φ
Therefore, it follows from (8.112) and (8.107) that
∥∇φ∥L2 (Ω) √
≥ µ Σ|B1 | N |Ω|− N ≥ ∥b∥∞ .
1 1
µ
∥φ∥L2 (Ω)
Finally, (8.108) follows by substituting (8.112) into (8.110).
Finally, from Theorem 8.9 and Proposition 8.6, the next result holds.
Corollary 8.4. Suppose (8.26). Then,
( 2
) 2
lim inf lim σ[L, B[β], Ω]|Ω| N ≥ µΣ|B1 | N .
|Ω|↓0 β∈C(Γ1 )
min β ↑ ∞
Γ1
In particular, σ[L, B[β], Ω] can be as large as we wish by choosing Ω

with |Ω| sufficiently small and β ∈ C(Γ1 ) with minΓ1 β sufficiently large.
Although the monotonicity properties established in Section 8.1 are folklore

under Dirichlet boundary conditions, in the general setting of this chapter
they might go back to S. Cano-Casanova and J. López-Gómez [39]. The
proofs of the monotonicity results in case B = D, through Theorem 7.10,
seem to go back to J. López-Gómez [137]. Most of the available proofs in
the literature dealt with the selfadjoint case (b = 0) by means of the J. W.
S. Rayleigh quotients (8.111).
Theorem 1 of M. H. Protter and H. F. Weinberger [182] might be con-
sidered not only as a precursor of Theorem 7.10 but also of the min-max
characterization of the principal eigenvalue provided by Theorem 8.1. Some
years later, M. D. Donsker and S. R. S Varadhan [54, 55], and M. Venturino
[222] gave some closely related general variational formulas for σ[L, D, Ω]
through (8.5). The proof of the min-max characterization (8.5) given in this
chapter, based on Theorem 7.10, has been adapted from the proofs of The-
orem 3.1 of J. López-Gómez [137] and Theorem 4.1 of S. Cano-Casanova
and J. López-Gómez [39].
Seemingly, the most pioneering version of the concavity of the principal
eigenvalue established by Theorem 8.3 goes back to T. Kato [111]. Some
further extensions of Kato’s result were given by A. Beltramo and P. Hess
[25], for some periodic-parabolic counterparts of the original problem, by
P. Hess [96], and in Proposition 2.1 of H. Berestycki, L. Nirenberg and S.
R. S. Varadhan [27] in the special case when B = D. In its full generality,
Theorem 8.3 goes back to Theorem 5.1 of S. Cano-Casanova and J. López-
Gómez [39]. The proof of Theorem 8.3 given here follows the general scheme
of the proof of Theorem 3.3 in J. López-Gómez [137] and it uses a device
coming from the proof of Lemma 5.1 of H. Berestycki, L. Nirenberg and S.
R. S. Varadhan [27]. Based on (iii) of p. 102 in P. Hess [95], one might
think that the original idea of the concavity of the principal eigenvalue
with respect to the potential goes back, actually, to an observation of H.
Berestycki and P. L. Lions.
The concept of stability of a domain through (8.10), as discussed in the
statement of Theorem 8.4, goes back to I. Babus̆ka [19] and I. Babus̆ka and
R. Vyborny [20], where it was introduced to generalize the pioneering results
of R. Courant and D. Hilbert [44] on the continuous variation with respect
to the domain Ω of the eigenvalues of a self-adjoint differential operator L
(b = 0) under homogeneous Dirichlet boundary conditions (B = D). The
concept of stability has shown to play a central rol in potential theory, as
it provides with all domains for which the Dirichlet problem is well posed
(cf. D. R. Adams and L. I. Hedberg [2] and the references therein). Rather
surprisingly, H. Berestycki, L. Nirenberg and S. R. S. Varadhan [27] did not
impose any stability condition on Ω.
Proposition 8.5 is a sharp version of Theorem 3.7 in J. Wloka [226]. It
is a pivotal result to get the continuous variation of the principal eigenvalue
with respect to the exterior approximations of Ω along Γ0 , which has been
established by Theorem 8.5. More general results, within the spirit of
Proposition 8.5, for Dirichlet boundary conditions, have been established
by J. L. Lions and E. Magenes [130].
As R. Courant and D. Hilbert [44] and I. Babus̆ka and R. Vyborny [20]
focused their attention on the special case when b = 0 and B = D, and E.
N. Dancer [46] dealt with the special case when
L = −∆ + ⟨b, ·⟩ + c
under Dirichlet boundary conditions, it seems that, in case B = D, The-
orem 8.7 goes back to Theorem 4.2 of J. López-Gómez [137], though the
regularity requirements on the coefficients of L in this book are substan-
tially weaker than those of [137]. In its greatest generality, Theorem 8.7
goes back to Theorem 7.4 of S. Cano-Casanova and J. López-Gómez [39].
We conjecture that, under the general assumptions of Theorem 8.7,
the spectrum of (L, Bn , Ωn ) does actually approximate the spectrum of
(L, B0 , Ω0 ) as n → ∞. R. Courant and D. Hilbert [44] observed that the
continuous dependence of the spectrum with respect to the domain may
fail when dealing with Neumann boundary conditions (Γ0 = ∅, β = 0 and
ν = n). This explains why in the results of Section 8.5 the portion Γ1 of
∂Ω remained unchanged. The celebrated example of R. Courant and D.
Hilbert [44] is the following. For any σ > 0, let
Ωσ := { (x, y) ∈ R2 : |x| < σ/2, |y| < σ/2 }
the square of area σ 2 /4 centered at the origin. Now, for any ϵ > 0 and
τ > 0, let
Rϵ,τ := { (x, y) ∈ R2 : 0 < x < ϵ, |y| < τ /2 }
and consider
Ωϵ,τ := Ω1 ∪ [Rϵ,τ + (1/2, 0)] ∪ [Ωϵ + (1/2 + ϵ, 0)].
For τ = ϵ , the domain Ωϵ,ϵ4 can be viewed as a C 0 -perturbation of Ω0 , but
4
not a C 1 -perturbation, as discussed by J. K. Hale [92]. For every ϵ > 0, let

( )
Aϵ : D(Aϵ ) → L2 Ωϵ,ϵ4
the operator defined by

Aϵ u := −∆u for all u ∈ D(Aϵ ),
where
D(Aϵ ) := {u ∈ W 1,2 (Ωϵ,ϵ4 ) : ∂n u = 0 on ∂Ωϵ,ϵ4 }.
Let σkϵ , k ≥ 0, denote the ordered eigenvalues of Aϵ . Then, σ0ϵ = 0 for all
ϵ ≥ 0, σ2ϵ > 0 for all ϵ > 0, and it is shown by R. Courant and D. Hilbert
[44] that
lim σ2ϵ = 0.
ϵ↓0
Consequently, the eigenvalues exhibit a singular behavior at ϵ = 0, in the

sense that the second eigenvalue for ϵ > 0, σ2ϵ , is bounded away from the
second eigenvalue for ϵ = 0, σ20 > 0, because σ2ϵ ∼ 0 if ϵ ∼ 0. In this
example, the principal eigenvalues for ϵ ≥ 0 equal zero.
Even if in the statement of Theorem 8.6 no regularity requirement is
imposed to Γ00 , the limit
σ I := lim σ[L, Bn , Ωn ]
n→∞
is well defined and it provides us with an eigenvalue to a weak positive so-

lution of (L, B0 , Ω0 ). This follows easily by adapting the proof of Theorem
8.6 (see Theorem 7.2 of S. Cano-Casanova and J. López-Gómez [39]). But,
except when Γ00 is of class C 2 , it is unknown whether or not σ E is the unique
eigenvalue of (L, B0 , Ω0 ) to a weak positive eigenfunction.
Although, in order to define WΓ1,p 0 (Ω0 ), p ≥ 1, through (4.16), this book
0
requires Γ00 to be of class C 1 , it is also possible to define
W 1,p (Ω0 )
WΓ1,p ∞
0 (Ω0 ) = CΓ0 (Ω0 )
0 0
and, in such case, no a priori regularity requirement on Γ00 is needed to

introduce the concept of stability of Ω0 along Γ00 by (8.10). Suppose Ω0 is
stable in this sense. Then, the value
σ E := lim σ[L, Bn , Ωn ]
n→∞
constructed in the proof of Theorem 8.5 also provides us with an eigenvalue

to a weak positive eigenfunction of (L, B0 , Ω0 ). According to Theorem 8.7,
we have that
σI = σE (8.113)
if Γ00 is of class C 2 . Naturally, it would be of the greatest interest to charac-

terize the class of domains Ω0 saisfying (8.113). We conjecture that (8.113)
holds if, and only if, Ω0 is stable along Γ00 .
Theorem 8.8 is Theorem 8.2 of S. Cano-Casanova and J. López-Gómez
[39]. The proof of Theorem 8.8 also provides us with the existence of
a principal eigenvalue, not necessarily unique, for a large class of weight
functions β ∈ L∞ (Γ1 ). Indeed, if β is the point-wise limit of a (uniformly)
bounded sequence βn ∈ C(Γ1 ), n ≥ 1, then, it follows from the dominated
∗
convergence theorem of Lebesgue that βn ⇀ β as n → ∞ and that
σ ∞ := lim σ(βn )
n→∞
is well defined. Moreover, the argument of the proof of Theorem 8.8 also
shows that σ ∞ is an eigenvalue of (L, B[β], Ω) to a positive eigenfunction.
Theorem 8.9 goes back to Theorem 9.1 of S. Cano-Casanova and J.
López-Gómez [39]. When, instead of (8.87), the following holds
lim max βn = −∞,
n→∞ Γ1
then, the behavior of the principal eigenvalue may change drastically as the
next simple one-dimensional example shows. Suppose
d2
N = 1, Ω = (0, 1), Γ0 = {0}, Γ1 = {1}, , L=−
dx2
and β ∈ R. Then, λ is an eigenvalue of (L, B[β], Ω) if, and only if, it is an
eigenvalue of problem
{
−u′′ (x) = λu(x), 0 < x < 1,
′ (8.114)
u(0) = 0, u (1) + βu(1) = 0.
Let us denote by σk [β], k ≥ 0, the increasing sequence of eigenvalues of
problem (8.114). According to Theorem 8.9, we have that
lim σ0 [β] = π 2
β↑∞
equals the principal eigenvalue of

{
−u′′ (x) = λu(x), 0 < x < 1,
(8.115)
u(0) = 0, u(1) = 0.
But
lim σ0 [β] = −∞,
β↓−∞
and
lim σn [β] = (nπ)2 for all n ≥ 1
β↓−∞
provide us with the eigenvalues of the limiting problem (8.115) (see Ex-
ercise 9.3 of J. López-Gómez [142]). Consequently, the second eigenvalue
of (8.114), σ1 [β], instead of σ0 [β], approximates the principal eigenvalue of
(8.115), π 2 , as β ↓ −∞.
According to H. Berestycki, L. Nirenberg and S. R. S. Varadhan [27],
the fact that
σ[L, D, Ω] > 0,
or, equivalently, that (L, D, Ω) satisfies the strong maximum principle for
sufficiently small |Ω|, was first noted by I. J. Bakelman [21] and then used
extensively by H. Berestycki and L. Nirenberg [26], though D. Gilbarg and
N. Trudinger, after the proof of Lemma 8.4 of [79], had already established
that L is coercive for sufficiently small |Ω|. Indeed, on pp. 52–53 of [27],
H. Berestycki, L. Nirenberg and S. R. S. Varadhan claim that
This paper was motivated by the following observation which was first noted
by Bakelman [2] and used extensively in [5].
PROPOSITION 1.1. Let L satisfy conditions (1.2) and (1.3) in Ω. Assume
diam Ω ≤ d. There exists δ > 0 depending only on n, c0 , b and d [the spatial
dimension and the coefficients of L] such that the maximum principle holds for
L in Ω if the measure of Ω, |Ω|, satisfies |Ω| < δ.
Naturally, the low estimates (8.105) are attributable to C. Faber [61]
and E. Krahn [117]. They were later refined to include some general classes
of second order elliptic operators L, under Dirichlet boundary conditions,
in Theorem 2.5 and Remark 2.2 of H. Berestycki, L. Nirenberg and S. R.
S. Varadhan [27], and in Theorem 5.1 of J. López-Gómez [137]. In their
greatest generality, Proposition 8.6 and Corollary 8.4 go back to Section 10
of S. Cano-Casanova and J. López-Gómez [39]. Essentially, they established
that (L, B[β], Ω) satisfies the strong maximum principle if β is sufficiently
large and |Ω| sufficiently small.

Chapter 9
Principal eigenvalues of linear

weighted boundary value problems
This chapter studies the existence and multiplicity of principal eigenvalues

for the linear weighted boundary value problem
{
Lφ = λW (x)φ in Ω,
(9.1)
Bφ = 0 on ∂Ω,
where L and B satisfy the general requirements of Chapter 4, ∂Ω is of class
C 2 , W ∈ L∞ (Ω), and λ ∈ R. Setting
V := −W (9.2)
and
Σ(λ) := σ[L + λV, B, Ω], λ ∈ R, (9.3)
a given value λ∗ ∈ R is said to be a principal eigenvalue of (9.1) if
Σ(λ∗ ) = 0.
This chapter characterizes the existence of principal eigenvalues of (9.1)
in all possible cases. The notations (9.2) and (9.3) will be maintained
throughout it.
Subsequently, we will outline the distribution and main contents of this
chapter. In Section 9.1 we study some general properties of the map Σ(λ).
Precisely, we show that it is real analytic and strictly concave if W changes
sign, while it is strictly monotone if W does not change sign. As a result of
these facts, collected in Theorem 9.1, the results of Section 9.2 characterize
the existence of principal eigenvalues for (9.1). Precisely, in the special, but
important, case when W ∈ C(Ω̄), the next results hold:
• Suppose W ≥ 0, or W ≤ 0, and W ̸= 0. Then, (9.1) has a principal
eigenvalue if, and only if, (L−λ0 W, B, Ω) satisfies the strong maximum
principle for some λ0 ∈ R. In such case, the principal eigenvalue is
unique and algebraically simple.
273
• Suppose W ̸= 0 changes sign in Ω. Then, (9.1) admits a principal

eigenvalue if, and only if, some of the following conditions hold:
(a) (L − λ0 W, B, Ω) satisfies the strong maximum principle for some
λ0 ∈ R. In such case, (9.1) possesses two principal eigenvalues,
which are algebraically simple.
(b) For every λ ∈ R, (L − λW, B, Ω) does not satisfy the strong maxi-
mum principle, but there exists λ0 ∈ R such that
σ[L − λ0 W, B, Ω] = 0.
In such case, λ0 must be the unique principal eigenvalue of (9.1),
and it is a double eigenvalue.
In Section 9.3, we introduce a general class of potentials V ≥ 0 for which
lim σ[L + λV, B, Ω] = σ[L, B[Ω0 ], Ω0 ],
λ↑∞
where
Ω0 := int V −1 (0).
For that class of potentials, the next characterization holds:
• Suppose W ≥ 0, or W ≤ 0, and W ̸= 0. Then, (9.1) has a principal
eigenvalue if, and only if,
σ[L, B[Ω0 ], Ω0 ] > 0.
According to Corollary 8.4, this holds if |Ω0 | is sufficiently small and β is
sufficiently large on Γ1 ∩ ∂Ω0 .
Finally, Section 9.4 uses these results to give a series of rather explicit
sufficient conditions so that (9.1) possess two principal eigenvalues when
W changes sign. Naturally, one has to obtain sufficient conditions so that
(L − λ0 W, B, Ω) satisfy the maximum principle for some λ0 ∈ R.
9.1 General properties of the map Σ(λ)
The next result collects some important properties of the map Σ(λ).
Theorem 9.1. Suppose V ∈ L∞ (Ω). Then, the map Σ(λ) defined by (9.3)
satisfies the following properties:
(a) Σ(λ) is real analytic and concave, in the sense that Σ′′ (λ) ≤ 0 for all
λ ∈ R. Therefore, either Σ′′ = 0 in R, or there exists a discrete set
Z ⊂ R such that Σ′′ (λ) < 0 for all λ ∈ R \ Z. By discrete it means that
Z ∩ K is finite for all compact subset K ⊂ R.
Principal eigenvalues of linear weighted boundary value problems 275
(b) Assume there exist x+ ∈ Ω and R > 0 such that

B+ := BR (x+ ) ⊂ Ω and inf V > 0. (9.4)
B+
Then,
lim Σ(λ) = −∞. (9.5)
λ↓−∞
Suppose, in addition, that V ≥ 0 in Ω. Then, Σ′ (λ) > 0 for all λ ∈ R.

(c) Assume there exist x− ∈ Ω and R > 0 such that
B− := BR (x− ) ⊂ Ω and sup V < 0. (9.6)
B−
Then,
lim Σ(λ) = −∞. (9.7)
λ↑∞
Suppose, in addition, that V ≤ 0 in Ω. Then, Σ′ (λ) < 0 for all λ ∈ R.

(d) Suppose there are x+ , x− ∈ Ω and R > 0 satisfying (9.4) and (9.6).
Then, (9.5) and (9.7) hold and, therefore, there exists λ0 ∈ R such that
Σ(λ0 ) = max Σ(λ). (9.8)
λ∈R
Moreover, Σ′ (λ0 ) = 0, Σ′ (λ) > 0 if λ < λ0 , and Σ′ (λ) < 0 if λ > λ0 .

Consequently, λ0 must be unique.
Proof. We will prove each of them separately.

Proof of Part (a): Subsequently, we set
L(λ) := L + λV, λ ∈ R,
and regard L(λ), λ ∈ R, as a family of closed operators with common
domain
−
D(L(λ)) = WB2,∞ (Ω)
and values in L2 (Ω). Then, L(λ) is a real holomorphic family of type (A)
as discussed in Section VII.2 of T. Kato [112], because
∫
λ 7→ vL(λ)u
Ω
−
is real analytic in λ for all v ∈ L (Ω) and u ∈ WB2,∞ (Ω). Therefore, by
2
Theorems 1.7 and 1.8 in Section VII.1.3 of [112], Σ(λ) is real analytic in
λ. Moreover, if φ(λ) ≫ 0 stands for the unique principal eigenfunction of
Σ(λ) such that
∫
φ2 (λ) = 1,
Ω
the map
R −→ L2 (Ω)
λ 7→ φ(λ)
is also real analytic. Next, we will show that
Σ′′ (λ) ≤ 0 for all λ ∈ R. (9.9)
According to Theorem 8.3, we have that
Σ(tλ1 + (1 − t)λ2 ) = σ[L + tλ1 V + (1 − t)λ2 V, B, Ω]
≥ tσ[L + λ1 V, B, Ω] + (1 − t)σ[L + λ2 V, B, Ω]
= tΣ(λ1 ) + (1 − t)Σ(λ2 )
for all t ∈ [0, 1] and λ1 , λ2 ∈ R. Thus,
Σ(λ2 + t(λ1 − λ2 )) ≥ Σ(λ2 ) + t (Σ(λ1 ) − Σ(λ2 ))
and hence,
Σ(λ2 + t(λ1 − λ2 )) − Σ(λ2 )
≥ Σ(λ1 ) − Σ(λ2 )
t
for all t ∈ (0, 1] and λ1 , λ2 ∈ R. Consequently,
Σ(λ2 + t(λ1 − λ2 )) − Σ(λ2 ) Σ(λ1 ) − Σ(λ2 )
≥ (9.10)
t(λ1 − λ2 ) λ1 − λ2
for all t ∈ (0, 1] and λ1 , λ2 ∈ R with λ1 > λ2 . Letting t ↓ 0 in (9.10), it
becomes apparent that
Σ(λ1 ) − Σ(λ2 )
Σ′ (λ2 ) ≥
λ1 − λ2
for all λ1 > λ2 . Thus, by the mean value theorem, we find that for every
λ1 , λ2 ∈ R such that λ1 > λ2 there exists λ ∈ (λ2 , λ1 ) for which
Σ′ (λ2 ) ≥ Σ′ (λ). (9.11)
Clearly, this is impossible if there exists λ ∈ R for which Σ′′ (λ) > 0, because
Σ′ should be increasing in a neighborhood of such λ. Therefore, (9.9) holds.
Finally, since Σ′′ is real analytic, owing to the identity principle, it
becomes apparent that either Σ′′ = 0, or Σ′′ vanishes, at most, in a discrete
set, possibly empty. This concludes the proof of Part (a).
Proof of Part (b): According to Proposition 8.2,
Σ(λ) = σ[L + λV, B, Ω] < σ[L + λV, D, B+ ],
because B[B+ ] = D. Thus, by Proposition 8.3, we have that

Σ(λ) < σ[L, D, B+ ] + λ inf V for all λ < 0.
B+
Clearly, (9.5) follows from (9.4) and this estimate by letting λ ↓ −∞.
Now, suppose V ≥ 0. Then, by Proposition 8.3, λ 7→ Σ(λ) is increasing
and hence Σ′ (λ) ≥ 0 for all λ ∈ R. By analyticity, either Σ′ = 0 in R,
or Σ′ (λ) > 0 for all λ ∈ R except in a discrete set. According to (9.5),
Σ(λ) cannot be a constant. Therefore, the second option occurs. Suppose
Σ′ (λ0 ) = 0 for some λ0 ∈ R. Then,
∫ λ
′ ′ ′
Σ (λ) = Σ (λ) − Σ (λ0 ) = Σ′′ ≤ 0 for all λ ≥ λ0
λ0
and, consequently, Σ′ = 0 in [λ0 , ∞), which is impossible. Therefore,
Σ′ (λ) > 0 for all λ ∈ R. This ends the proof of Part (b).
Proof of Part (c): As in the proof of Part (b), we have that
Σ(λ) = σ[L + λV, B, Ω] < σ[L + λV, D, B− ],
for all λ ∈ R and hence,
Σ(λ) < σ[L, D, B− ] + λ sup V for all λ > 0.
B−
Obviously, (9.7) follows from (9.6) and this estimate by letting λ ↑ ∞.

Now, suppose V ≤ 0. Then, by Proposition 8.3, λ 7→ Σ(λ) is decreasing
and hence Σ′ (λ) ≤ 0 for all λ ∈ R. By analyticity, either Σ′ = 0 in R, or
Σ′ (λ) < 0 for all λ ∈ R except in a discrete set. According to (9.7), Σ(λ)
cannot be a constant and, so, the second option occurs. Suppose Σ′ (λ0 ) = 0
for some λ0 ∈ R. Then,
∫ λ0
−Σ′ (λ) = Σ′ (λ0 ) − Σ′ (λ) = Σ′′ ≤ 0 for all λ ≤ λ0
λ
and, consequently, Σ′ = 0 in (−∞, λ0 ], which is impossible. Therefore,
Σ′ (λ) < 0 for all λ ∈ R. This shows Part (c).
Proof of Part (d): The existence of λ0 is a direct consequence from the
continuity of Σ(λ) and of (9.5) and (9.7). Necessarily, Σ′ (λ0 ) = 0. Next,
we will show that Σ′ (λ) > 0 for all λ < λ0 . On the contrary, suppose there
exists λ1 < λ0 such that
Σ′ (λ1 ) ≤ 0.
Then,
∫ λ1 ∫ λ0
0 ≥ Σ′ (λ1 ) = Σ′′ (λ) dλ = − Σ′′ ≥ 0,
λ0 λ1
because Σ′′ ≤ 0. Hence,

∫ λ0
′
Σ (λ1 ) = − Σ′′ = 0
λ1
and, therefore, Σ′′ = 0 in [λ1 , λ0 ]. By Part (a), Σ′′ = 0 in R and, so, there
exist a, b ∈ R such that
Σ(λ) = aλ + b for all λ ∈ R.
By (9.5) and (9.7), this cannot occur. Thus, Σ′ (λ) > 0 for all λ < λ0 .
Finally, we will prove that Σ′ (λ) < 0 for all λ > λ0 . On the contrary,
suppose that there exists λ2 > λ0 such that
Σ′ (λ2 ) ≥ 0.
Then,
∫ λ2
′
0 ≤ Σ (λ2 ) = Σ′′ ≤ 0
λ0
and so,
∫ λ2
Σ′ (λ2 ) = Σ′′ = 0
λ0
which implies Σ′′ = 0 in [λ0 , λ2 ] and, consequently, by analyticity, Σ′′ = 0

in R, which is impossible. The proof is complete.
9.2 Characterizing the existence of a principal eigenvalue
This section characterizes the existence and multiplicity of principal eigen-

values for the linear weighted boundary value problem (9.1). According to
the sign of the weight function W , we will distinguish three different cases.
When W ≥ 0, the main result reads as follows. Throughout this section, it
should not be forgotten that
V := −W.
Theorem 9.2. Suppose W ≥ 0 and there are x− ∈ Ω and R > 0 such that
B− := BR (x− ) ⊂ Ω and inf W > 0. (9.12)
B−
Then, (9.1) possesses a principal eigenvalue if and only if

Σ(−∞) := lim Σ(λ) > 0. (9.13)
λ↓−∞
Moreover, it is unique if it exists and if we denote it by λ∗ , then, λ∗ is a

simple eigenvalue of (L − λW, W ) as discussed by M. G. Crandall and P.
H. Rabinowitz [45], i.e.,
W φ∗ ∈
/ R[L − λ∗ W ] (9.14)
∗ ∗
for any principal eigenfunction φ ≫ 0 of (9.1) associated to λ .
Proof. By (9.12), the function V := −W ≤ 0 satisfies (9.6) and hence,

according to Theorem 9.1(c), (9.7) holds. Moreover,
Σ′ (λ) < 0 for all λ ∈ R (9.15)
and, therefore, the limit (9.13) is well defined. Note that Σ(−∞) might be
finite or infinity. Indeed,
Σ(−∞) = ∞ if inf W > 0,
Ω
because, in such case, for every λ < 0 we have that
Σ(λ) = σ[L − λW, B, Ω] ≥ σ[L, B, Ω] − λ inf W,
Ω
while, due to Proposition 8.2, when W = 0 in some open set Ω0 ⊂ Ω with
Ω̄0 ⊂ Ω, then,
Σ(λ) = σ[L − λW, B, Ω] ≤ σ[L, D, Ω0 ] for all λ ∈ R
and, consequently,
Σ(−∞) ≤ σ[L, D, Ω0 ].
In (9.13) we use the natural convention that ∞ > 0 if Σ(−∞) = ∞.
Suppose Σ(−∞) > 0. Then, Σ(λ1 ) > 0 for some λ1 ∈ R and hence,
by (9.7), there exists λ∗ > λ1 such that Σ(λ∗ ) = 0. Conversely, if there
exists λ∗ ∈ R such that Σ(λ∗ ) = 0, then, by (9.15), Σ′ (λ∗ ) < 0 and hence
Σ(λ) > 0 for all λ < λ∗ . Therefore, Σ(−∞) > 0. This shows that λ∗ exists
if and only if (9.13) holds.
The uniqueness of λ∗ can be obtained, for example, with the following
argument. Suppose
Σ(λ∗1 ) = Σ(λ∗2 ) = 0
for some λ∗1 < λ∗2 . Then, by the mean value theorem, there exists λ ∈
(λ∗1 , λ∗2 ) such that Σ′ (λ) = 0, which contradicts (9.15). Therefore, λ∗ is
unique.
It remains to prove (9.14). Let φ(λ) be the principal eigenfunction
associated to Σ(λ) normalized so that
∫
φ2 (λ) = 1. (9.16)
Ω
According to the proof of Theorem 9.1(a), the map λ 7→ φ(λ) is real ana-
lytic. Thus, differentiating the identity
(L − λW )φ(λ) = Σ(λ)φ(λ), λ ∈ R,
with respect to λ yields
(L − λW )φ′ (λ) − W φ(λ) = Σ′ (λ)φ(λ) + Σ(λ)φ′ (λ), λ ∈ R,
and, consequently, particularizing at λ = λ∗ shows that
(L − λ∗ W )φ′ (λ∗ ) = W φ(λ∗ ) + Σ′ (λ∗ )φ(λ∗ ). (9.17)
Suppose
W φ(λ∗ ) ∈ R[L − λ∗ W ].
Then, since Σ′ (λ∗ ) < 0, (9.17) implies that
φ(λ∗ ) ∈ R[L − λ∗ W ],
which is impossible, because
N [L − λ∗ W ] = span [φ(λ∗ )]
and, according to Theorem 7.8, Σ(λ∗ ) = 0 is a simple eigenvalue of (L −
λ∗ W, B, Ω). The proof is complete.
Figure 9.1 shows a genuine graph of the map λ 7→ Σ(λ) in the special
case when W ≥ 0 and Σα := Σ(−∞) > 0. Note that, under the assump-
Σ(λ)
Σα
λ
λ*
Fig. 9.1 A genuine graph of λ 7→ Σ(λ) when W ≥ 0
tions of Theorem 9.2, λ∗ > 0 if

Σ(0) = σ[L, B, Ω] > 0,
λ∗ = 0 if σ[L, B, Ω] = 0, and λ∗ < 0 if σ[L, B, Ω] < 0.
Remark 9.1. The proof of (9.14) is based on the fact that Σ′ (λ∗ ) ̸= 0,
rather than on the sign of W . Therefore, the last assertion of Theorem 9.2
holds true as soon as
Σ(λ∗ ) = 0 and Σ′ (λ∗ ) ̸= 0.
Similarly, the next result holds in the special case when W ≤ 0.
Theorem 9.3. Suppose W ≤ 0 and there are x+ ∈ Ω and R > 0 such that
B+ := BR (x+ ) ⊂ Ω and sup W < 0. (9.18)
B+
Then, (9.1) possesses a principal eigenvalue if and only if

Σ(∞) := lim Σ(λ) > 0. (9.19)
λ↑∞
Moreover, it is unique if it exists and, if we denote it by λ∗ , then, λ∗ is a

simple eigenvalue of (L − λW, W ) as discussed by M. G. Crandall and P.
H. Rabinowitz [45].
Proof. By (9.18), the function V := −W ≤ 0 satisfies (9.4) and hence,

according to Theorem 9.1(b), (9.5) holds. Moreover,
Σ′ (λ) > 0 for all λ ∈ R (9.20)
and, consequently, the limit (9.19) is well defined. It might be finite or
infinite. Indeed,
Σ(∞) = ∞ if sup W < 0,
Ω
because, in such case, for every λ > 0 we have that

Σ(λ) = σ[L − λW, B, Ω] ≥ σ[L, B, Ω] − λ sup W,
Ω
while, owing to Proposition 8.2, if W = 0 in some open subset Ω0 ⊂ Ω with

Ω̄0 ⊂ Ω, then,
Σ(λ) = σ[L − λW, B, Ω] ≤ σ[L, D, Ω0 ] for all λ ∈ R,
and, consequently,
Σ(∞) ≤ σ[L, D, Ω0 ].
Naturally, in (9.19) we use the convention ∞ > 0 if Σ(∞) = ∞.
Suppose Σ(∞) > 0. Then, Σ(λ1 ) > 0 for some λ1 ∈ R and hence,
by (9.5), there exists λ∗ < λ1 such that Σ(λ∗ ) = 0. Conversely, if there
is λ∗ ∈ R such that Σ(λ∗ ) = 0, then, by (9.20), Σ′ (λ∗ ) > 0 and hence

Σ(λ) > 0 for all λ > λ∗ . Therefore, Σ(∞) > 0. This shows that λ∗ exists
if and only if (9.19) holds.
The uniqueness of λ∗ follows as in the proof of Theorem 9.2 and, so,
we omit the details here. The last assertion follows from Remark 9.1, by
(9.20). The proof is complete.
Figure 9.2 shows a genuine graph of the map λ 7→ Σ(λ) in the special
case when W ≤ 0 and Σω := Σ(∞) > 0. Under the assumptions of Theorem
9.3, λ∗ > 0 if
Σ(0) = σ[L, B, Ω] < 0,
λ∗ = 0 if Σ(0) = 0, and λ∗ < 0 if Σ(0) > 0. Figure 9.2 shows an admissible
case when λ∗ < 0.
Σ(λ)
Σω
λ
λ*
Fig. 9.2 A genuine graph of λ 7→ Σ(λ) when W ≤ 0
Finally, the next result holds when the weight function W changes of
sign in Ω.
Theorem 9.4. Suppose W changes sign in Ω, in the sense that there are
x+ , x− ∈ Ω and R > 0 for which (9.12) and (9.18) hold. Then,
lim Σ(λ) = lim Σ(λ) = −∞, (9.21)
λ↓−∞ λ↑∞
there exists a unique λ0 ∈ R such that

Σ(λ0 ) = max Σ(λ), (9.22)
λ∈R
and Σ′ (λ0 ) = 0, Σ′ (λ) > 0 if λ < λ0 , and Σ′ (λ) < 0 for all λ > λ0 .
Therefore, (9.1) possesses a principal eigenvalue if, and only if, Σ(λ0 ) ≥ 0.
Moreover, λ0 provides us with the unique principal eigenvalue of (9.1) if

Σ(λ0 ) = 0, while (9.1) possesses two principal eigenvalues, say λ∗− < λ∗+ if
Σ(λ0 ) > 0. Furthermore, in such case,
λ∗− < λ0 < λ∗+
and λ∗− , λ∗+ are simple eigenvalues of (L − λW, W ) as discussed by M. G.
Crandall and P. H. Rabinowitz [45].
Proof. This result is an immediate consequence from Theorem 9.1(d) and

Remark 9.1. It should be noted that Σ′ (λ0 ) = 0 and, therefore, the principal
eigenvalue cannot be a simple eigenvalue of (L − λW, W ) if Σ(λ0 ) = 0.
Figure 9.3 shows a genuine graph of the map λ 7→ Σ(λ) in the general
case when W changes sign in Ω and Σ(λ0 ) > 0.
Σ(λ)
λ
λ*- λ*
+
Fig. 9.3 A genuine graph of λ 7→ Σ(λ) when W changes sign in Ω
Under the assumptions of Theorem 9.4, suppose Σ(λ0 ) > 0. Then,

λ∗− < 0 < λ∗+ if
Σ(0) = σ[L, B, Ω] > 0, (9.23)
0 = λ∗− < λ∗+ if
Σ(0) = σ[L, B, Ω] = 0 and Σ′ (0) > 0,
λ∗− < λ∗+ = 0 if
Σ(0) = σ[L, B, Ω] = 0 and Σ′ (0) < 0,
0 < λ∗− < λ∗+ if
Σ(0) = σ[L, B, Ω] < 0 and Σ′ (0) > 0,
and, finally, λ∗− < λ∗+ < 0 if

Σ(0) = σ[L, B, Ω] < 0 and Σ′ (0) < 0. (9.24)
The very special case when condition (9.23) holds is the classical situation
covered by the theorem of P. Hess and T. Kato [97] (see P. Hess [96]).
In Figure 9.3, (9.24) holds and hence λ∗− < λ∗+ < 0. By changing the
sign of the weight function W , i.e., working with −W , instead of W , the
corresponding Σ(λ) has two positive zeros. Namely, −λ∗− and −λ∗+ .
9.3 Ascertaining limλ→∞ σ[L + λV, B, Ω] when V ≥ 0
Throughout this section we will assume (8.26) and V ≥ 0 with V ̸= 0.

Imposing (8.26) is imperative for applying the results of Section 8.5 on
continuity of the principal eigenvalue with respect to the domain.
In the special case when
inf V > 0,
Ω
we have that
Σ(λ) = σ[L + λV, B, Ω] > σ[L, B, Ω] + λ inf V for all λ>0
Ω
and, therefore,
lim Σ(λ) = ∞. (9.25)
λ↑∞
Consequently, throughout this section, we will assume that

inf V = 0. (9.26)
Ω
This section consists of three parts. In the first one, we will ascertain
lim σ[L + λV, D, Ω]
λ↑∞
for a special class of potentials V for which characterizing this limit does
not require so many technicalities as the proof of the most general case
covered by the main theorem. Then, we will introduce the most general
class of potentials V for which we will characterize this limit. Finally,
we will establish and prove the main result, through a rather natural but
extremely technical and lengthy proof.
Naturally, all results of this section can be easily adapted to cover the
case when V ≤ 0, but the corresponding details are not given here.
9.3.1 The simplest case

Theorem 9.5. Suppose (8.26) and the open set

Ω0 := int V −1 (0) (9.27)
is connected and of class C , and it satisfies Ω̄0 ⊂ Ω.
2
For sufficiently small δ > 0, let Ωδ be the open δ-neighborhood of Ω0

Ωδ := { x ∈ Ω : dist (x, Ω0 ) < δ }, (9.28)
and suppose, in addition, that
inf V > 0 for sufficiently small δ > 0. (9.29)
Ω̄\Ωδ
Then,
lim σ[L + λV, D, Ω] = σ[L, D, Ω0 ]. (9.30)
λ↑∞
As Ω0 is of class C 2 , ∂Ω0 necessarily has finitely many components.

Note that a sufficient condition for (9.29) is
V ∈ C(Ω̄) and V (x) > 0 ∀ x ∈ Ω̄ \ Ω̄0 ,
because Ω̄ \ Ωδ is a compact subset of Ω̄ \ Ω̄0 , for δ ∼ 0, and V is continuous
and positive therein. Figure 9.4 shows an admissible situation where ∂Ω
has one component and ∂Ω0 two. The darker area is the region where
V > 0. Its complement is Ω0 , the region where V = 0.
Ω
0
Ω
Γ0
Fig. 9.4 A nodal configuration within the setting of Theorem 9.5
Proof. According to Proposition 8.2,

σ[L + λV, D, Ω] < σ[L + λV, D, Ω0 ] = σ[L, D, Ω0 ]
for all λ ∈ R and hence,

Σ(∞) := lim σ[L + λV, D, Ω] ≤ σ[L, D, Ω0 ].
λ↑∞
Thus, to complete the proof of (9.30) we should prove that for every ϵ > 0
there exists λ1 = λ1 (ϵ) such that
σ[L + λV, D, Ω] > σ[L, D, Ω0 ] − ϵ ∀ λ ≥ λ1 . (9.31)
Subsequently, we fix ϵ > 0 and consider the δ-neighborhoods of Ω0 defined
by (9.28). According to Definition 8.1(E), Ωδ converges to Ω0 from its
exterior as δ ↓ 0. Thus, thanks to Theorems 8.4 and 8.5, we find that
lim σ[L, D, Ωδ ] = σ[L, D, Ω0 ].
δ↓0
Therefore, for sufficiently small δ > 0,

σ[L, D, Ω0 ] − ϵ < σ[L, D, Ωδ ] < σ[L, D, Ω0 ].
Fix one of these δ’s. Then, the estimate
σ[L + λV, D, Ω] > σ[L, D, Ωδ ] ∀ λ ≥ λ1 , (9.32)
implies (9.31) and hence, to complete the proof, it suffices to show (9.32),
or, equivalently,
σ [L + λV − σ[L, D, Ωδ ], D, Ω] > 0 ∀ λ ≥ λ1 . (9.33)
This will be a consequence from Theorem 7.10 through the construction of
a positive strict supersolution for
(L + λV − σ[L, D, Ωδ ], D, Ω) .
Let φδ ≫ 0 denote the unique principal eigenfunction of σ[L, D, Ωδ ] with
maxΩ φδ = 1, and consider the function h defined by
{
φδ (x) if x ∈ Ωδ/2 ,
h(x) := (9.34)
ψ(x) if x ∈ Ω̄ \ Ωδ/2 ,
where ψ is any positive smooth extension of φδ outside Ωδ/2 separated from
zero. It exists because φδ is positive and bounded away from zero along
∂Ωδ/2 , as ∂Ωδ/2 ⊂ Ωδ and φδ ≫ 0 in Ωδ . We claim that, for sufficiently
large λ > 0, the function h is a strict supersolution of
Lλ := L + λV − σ[L, D, Ωδ ]
in Ω under homogeneous Dirichlet boundary conditions. Indeed, by con-
struction, h(x) > 0 for all x ∈ Ω̄, and, in Ωδ/2 , we have that
Lλ h = Lλ φδ = (L − σ[L, D, Ωδ ])φδ + λV φδ = λV φδ ≥ 0,
while, in the complement of Ωδ/2 , for sufficiently large λ > 0,
Lλ h = Lλ ψ = (L − σ[L, D, Ωδ ])ψ + λV ψ
≥ (L − σ[L, D, Ωδ ])ψ + λ inf V inf ψ > 0
Ω̄\Ωδ/2 Ω̄\Ωδ/2
because, thanks to (9.29) and the construction of ψ,
inf V > 0 and inf ψ > 0.

Ω̄\Ωδ/2 Ω̄\Ωδ/2
This shows the claim above. Theorem 7.10 ends the proof.
As a consequence from Theorem 9.3, Proposition 8.6 and (9.30), the

next result holds.
Corollary 9.1. Suppose V satisfies the requirements of Theorem 9.5.

Then, (9.1) admits a principal eigenvalue if, and only if,
σ[L, D, Ω0 ] > 0.
Moreover, this holds for sufficiently small |Ω0 |.
9.3.2 The admissible V ’s satisfying the main theorem

In this section we introduce a class of non-negative potentials V ∈ L∞ (Ω),
denoted by A, for which
lim σ[L + λV, B, Ω] = σ[L, B[Ω0 ], Ω0 ], (9.35)

λ↑∞
where Ω0 is the maximal open subset of Ω where V = 0 and B[Ω0 ] stands

for the boundary operator introduced in (8.2). The class A is introduced
through the next definition.
Definition 9.1. Suppose V ∈ L∞ (Ω) and V ≥ 0. It is said that V is an

admissible potential for (9.35), or, shortly, that V ∈ A, if there exist an
open subset Ω0 of Ω and a compact subset K of Ω̄ with Lebesgue measure
zero such that
K ∩ (Ω̄0 ∪ Γ1 ) = ∅, (9.36)
Ω+ := { x ∈ Ω : V (x) > 0} = Ω \ (Ω̄0 ∪ K) , (9.37)
for which the following conditions are satisfied:

(a) Ω0 possesses finitely many components of class C 2 , say Ωj0 , 1 ≤ j ≤ m,

such that
Ω̄i0 ∩ Ω̄j0 = ∅ if i ̸= j,
and
dist (Γ1 , ∂Ω0 ∩ Ω) > 0. (9.38)
Thus, if we denote by Γi1 , 1 ≤ i ≤ n1 , the components of Γ1 , then, for

each 1 ≤ i ≤ n1 , either Γi1 is a component of ∂Ω0 , or Γi1 ∩ ∂Ω0 = ∅.
Indeed, if Γi1 ∩ ∂Ω0 ̸= ∅ but Γi1 is not a component of ∂Ω0 , then
dist(Γi1 , ∂Ω0 ∩ Ω) = 0,
which contradicts (9.38).

(b) Let {i1 , ..., ip } denote the subset of {1, ..., n1 } for which
Γi1 ∩ ∂Ω0 = ∅ ⇐⇒ i ∈ {i1 , ..., ip }.
Then, V is bounded away from zero in any compact subset of

∪
p
i
Ω+ ∪ Γ1j .
j=1
When Γ1 ⊂ ∂Ω0 , i.e., {i1 , ..., ip } = ∅, then we are only imposing that
V is bounded away from zero in any compact subset of Ω+ .
Now, we will explain the meaning of this condition in the special,
but important, case when V ∈ C(Ω̄). In such case, by continuity, V is
bounded away from zero in any compact subset of Ω+ and, therefore,
condition (b) holds if either Γ1 ⊂ ∂Ω0 , or {i1 , ..., ip } =
̸ ∅ and
∪
p
i
V (x) > 0 for all x ∈ Γ+
1 := Γ1j , (9.39)
j=1
because Γ+1 is compact and V continuous. According to condition (a),

Γi1 must be a component of ∂Ω0 for all
i ∈ {1, ..., n1 } \ {i1 , ..., ip }
and hence V = 0 on all these components.

(c) Let Γi0 , 1 ≤ i ≤ n0 , denote the components of Γ0 , and let {i1 , ..., iq } be
the subset of {1, ..., n0 } for which
(∂Ω0 ∪ K) ∩ Γi0 ̸= ∅ ⇐⇒ i ∈ {i1 , ..., iq }.

Then, V is bounded away from zero in any compact subset of

∪q
i
Ω+ ∪ Γ0j \ (∂Ω0 ∪ K).
j=1
When (∂Ω0 ∪ K) ∩ Γ0 = ∅, we are only imposing that V is bounded

away from zero in any compact subset of Ω+ .
Now, we will explain this condition in case V ∈ C(Ω̄). In such
situation, by continuity, V is bounded away from zero in Ω+ , and,
therefore, condition (c) holds if (∂Ω0 ∪ K) ∩ Γ0 = ∅. So, suppose
(∂Ω0 ∪ K) ∩ Γ0 ̸= ∅.
Then, according to condition (a), V = 0 in ∂Ω0 ∪ K and hence,
∪
q
i
V =0 in Γ0j ∩ (∂Ω0 ∪ K).
j=1
Consequently, the condition (c) holds if and only if

∪
q
i
V (x) > 0 for all x ∈ Γ0j \ (∂Ω0 ∪ K). (9.40)
j=1
(d) For every η > 0 there exist a natural number ℓ(η) ≥ 1 and ℓ(η) open
subsets of RN , Gηj , 1 ≤ j ≤ ℓ(η), with |Gηj | < η, 1 ≤ j ≤ ℓ(η), such
that
∪
ℓ(η)
Ḡηi ∩ Ḡηj = ∅ if i ̸= j, K⊂ Gηj ,
j=1
and Gηj ∩ Ω is connected and of class C for all 1 ≤ j ≤ ℓ(η).

2
When Ω0 satisfies all the requirements of Definition 9.1(a), one can

introduce the following concept.
Definition 9.2. Let Ω0 be an open subset of Ω satisfying the requirements

of Definition 9.1(a). Then, the principal eigenvalue of (L, B[Ω0 ], Ω0 ) is
defined through
σ[L, B[Ω0 ], Ω0 ] := min σ[L, B[Ωj0 ], Ωj0 ].
1≤j≤m
Remark 9.2. As (9.38) is (8.1) and Ω0 is of class C 2 , the principal eigen-

values
σ[L, B[Ω0 ], Ω0 ] and σ[L, B[Ωj0 ], Ωj0 ], 1 ≤ j ≤ m,
are well defined.
9.3.3 The main theorem

Now, we are ready to state and prove the main result of this section.
Theorem 9.6. Suppose (8.26) and V ∈ A satisfies int Ω+ ̸= ∅. Then,

(9.35) holds.
Proof. Thanks to Proposition 8.2, since V = 0 in Ω0 , we have that

σ[L + λV, B, Ω] < σ[L, B[Ωj0 ], Ωj0 ]
for all 1 ≤ j ≤ m and λ ∈ R. Thus, according to Definition 9.2,
σ[L + λV, B, Ω] < σ[L, B[Ω0 ], Ω0 ] ∀ λ ∈ R. (9.41)
Note that V satisfies (9.4) if V ∈ A with int Ω+ ̸= ∅. Hence, by Theorem
9.1(a), the map
λ 7→ Σ(λ) = σ[L + λV, B, Ω]
is increasing and, consequently,
Σ(∞) = lim Σ(λ) ≤ σ[L, B[Ω0 ], Ω0 ].
λ↑∞
Therefore, to complete the proof of (9.35) it remains to show that for every
ϵ > 0 there exists λ1 = λ1 (ϵ) ∈ R such that
Σ(λ) > σ[L, B[Ω0 ], Ω0 ] − ϵ ∀ λ ≥ λ1 ,
or, equivalently,
σ[L + λV − σ[L, B[Ω0 ], Ω0 ] + ϵ, B, Ω] > 0 ∀ λ ≥ λ1 . (9.42)
According to Theorem 7.10, (9.42) holds if and only if
(L + λV − σ[L, B[Ω0 ], Ω0 ] + ϵ, B, Ω) (9.43)
possesses a positive strict supersolution for all λ > λ1 . Therefore, much
like in the proof of Theorem 9.5, the rest of the proof is devoted to the con-
struction of a positive strict supersolution of (9.43) for sufficiently large λ.
In the construction of the supersolution we will distinguish several different
cases according to the structure of Ω0 . First, we will consider the simplest
cases when Ω0 is connected and K = ∅. Then, we shall consider the most
general cases.
Step 1: Suppose
m = 1, K = ∅, and Γ0 ∩ ∂Ω0 = ∅. (9.44)
Then, Definition 9.1(c) only requires V to be bounded away from zero in

any compact subset of Ω+ , which had already been imposed by Definition
9.1(b).
For each k ∈ {0, 1}, let Γjk , 1 ≤ j ≤ nk , denote the components of Γk ,
and let {i1 , ..., ip } be the subset of {1, ..., n1 } for which
Γj1 ∩ ∂Ω0 = ∅ ⇐⇒ j ∈ {i1 , ..., ip }. (9.45)
According to Definition 9.1(a), Γi1 is a component of ∂Ω0 for all i ∈

{1, ..., n1 } \ {i1 , ..., ip }.
Figure 9.5 shows an admissible example satisfying (9.44). In this case,
Γ0 consists of two components, Γ10 and Γ20 , as well as Γ1 , whose compo-
nents have been named by Γ11 and Γ21 . The boundary of Ω0 , ∂Ω0 , has two
components too. One of them is Γ21 . The other one lies within Ω.
1
Γ
0
2
Γ
0
2
Γ
1
Γ1
1 Ω
0
Ω+
Fig. 9.5 A nodal configuration satisfying (9.44)
In Figure 9.5, Ω+ is the darker subdomain of Ω, while Ω0 is its comple-

ment, so that Ω+ = Ω \ Ω̄0 . Any potential V ∈ C(Ω̄) such that V = 0 in
Ω0 and V (x) > 0 for all x ∈ Ω+ satisfies Definition 9.1(a), (c) and (d), and
it satisfies Definition 9.1(b) if, and only if, V (x) > 0 for all x ∈ Γ11 .
Subsequently, we fix ϵ > 0 and, for sufficiently small δ > 0, consider the
open δ-neighborhoods
Ωδ := (Ω0 + Bδ ) ∩ Ω,
Nδ0,j := (Γj0 + Bδ ) ∩ Ω, 1 ≤ j ≤ n0 , (9.46)
Nδ1,j := (Γj1 + Bδ ) ∩ Ω, j ∈ {i1 , ..., ip }.
Recall that Bδ ⊂ RN is the ball of radius δ centered at the origin. Figure

9.6 shows these neighborhoods, for sufficiently small δ > 0, for the example
of Figure 9.5.
0,2 1
N Γ
δ 0
Γ2
0
0,1
N
δ
Γ12
1 Ω0
Γ
1
Ω
δ
1,1
N Ω+
δ
Fig. 9.6 The δ-neighborhoods defined in (9.46)
As Ω0 is of class C 2 , ∂Ω0 possesses finitely many components and, there-

fore, Ω0 possesses, at most, finitely many holes. By (9.44), there exists
δ0 > 0 such that, for every 0 < δ < δ0 ,
∪
n0 ∪
n0
∂Ωδ \ (Γ1 ∩ ∂Ω0 ) ⊂ Ω+ , Ω̄δ ∩ N̄δ0,j = ∅, N̄δ0,j \ Γ0 ⊂ Ω+ .
j=1 j=1
Moreover, since Γjk ∩ Γiℓ = ∅ if i ̸= j, there exists δ1 ∈ (0, δ0 ) such that, for
each 0 < δ < δ1 ,
N̄δk,j ∩ N̄δℓ,i = ∅ if (i, ℓ) ̸= (j, k), k , ℓ ∈ {0, 1}.
Also, according to (9.45), we have that

∪
p
i
∂Ω0 ∩ Γ1j = ∅
j=1
and hence, there exists δ2 ∈ (0, δ1 ) such that

∪
p
1,ij
Ω̄δ ∩ N̄δ =∅
j=1
for all 0 < δ < δ2 .

By construction, Ω0 is a proper subdomain of Ωδ such that

lim Ωδ = Ω0
δ↓0
in the sense of Definition 8.1(E). Thus, owing to (9.38), Proposition 8.2

implies that
σ[L, B[Ωδ ], Ωδ ] < σ[L, B[Ω0 ], Ω0 ], 0 < δ < δ2 ,
and, thanks to Theorems 8.4 and 8.5,
lim σ[L, B[Ωδ ], Ωδ ] = σ[L, B[Ω0 ], Ω0 ].
δ↓0
Therefore, there exists δ3 ∈ (0, δ2 ) such that

σ[L, B[Ωδ ], Ωδ ] < σ[L, B[Ω0 ], Ω0 ] < σ[L, B[Ωδ ], Ωδ ] + ϵ (9.47)
On the other hand, since
lim |Nδ0,j | = 0, 1 ≤ j ≤ n0 ,
δ↓0
it follows from Proposition 8.6 that

lim σ[L, D, Nδ0,j ] = ∞, 1 ≤ j ≤ n0 ,
δ↓0
and, consequently, there exists δ4 ∈ (0, δ3 ) such that

σ[L, D, Nδ0,j ] > σ[L, B[Ω0 ], Ω0 ], 1 ≤ j ≤ n0 , (9.48)
Subsequently, we fix δ ∈ (0, δ4 ). Let φδ , ψδi , i ∈ {i1 , .., ip }, and ξδj ,
j ∈ {1, ..., n0 }, denote three arbitrary principal eigenfunctions associated
with
σ[L, B[Ωδ ], Ωδ ], σ[L, B[Nδ1,i ], Nδ1,i ], i ∈ {i1 , ..., ip },
and
σ[L, D, Nδ0,j ], j ∈ {1, ..., n0 },
respectively, and consider the function Φ : Ω̄ → [0, ∞) defined through


 φδ in Ω̄δ/2 ,








ij 1,i
in N̄δ/2j , 1 ≤ j ≤ p,
 ψδ
Φ := (9.49)

 j 0,j

 ξ in N̄ , 1 ≤ j ≤ n ,


δ δ/2 0



 ∪p ∪n0
ζ 1,i
in Ω̄ \ (Ω̄δ/2 ∪ j=1 N̄δ/2j ∪ j=1 0,j
N̄δ/2 ),
δ
i
where ζδ is any smooth positive extension of φδ , ψδj , 1 ≤ j ≤ p, and ξδj ,
1 ≤ j ≤ n0 , from
∪p
1,i
∪
n0
0,j
Kδ := Ω̄δ/2 ∪ N̄δ/2j ∪ N̄δ/2
j=1 j=1
to Ω̄ which is bounded away from zero in Ω̄ \ Kδ . The function ζδ does exist

because the restrictions
i
φδ |∂Ωδ/2 ∩Ω , ψδj |∂N 1,ij \Γij , 1 ≤ j ≤ p,
δ/2 1
and ξδj |∂N 0,j \Γj , 1 ≤ j ≤ n0 , are positive and bounded away from zero. By
δ/2 0
construction,
Φ(x) > 0 for each x ∈ Ω.
Naturally, if Γ1 ⊂ ∂Ω0 , i.e., {i1 , ..., ip } = ∅, then, in the definition of Φ the
i
ψδj ’s should be deleted.
Next, we will show that there exists λ1 = λ1 (ϵ) such that Φ provides us
with a strict supersolution of (9.43) for all λ ≥ λ1 . By Theorem 7.10, this
completes the proof under condition (9.44).
By (9.47), the following estimate holds in Ωδ/2
(L + λV − σ[L,B[Ω0 ], Ω0 ] + ϵ)Φ = (L + λV − σ[L, B[Ω0 ], Ω0 ] + ϵ)φδ
= (σ[L, B[Ωδ ], Ωδ ] − σ[L, B[Ω0 ], Ω0 ] + ϵ)φδ > 0
0,j
for all λ ∈ R. Similarly, by (9.48), the next estimate holds in Nδ/2
(L + λV − σ[L,B[Ω0 ], Ω0 ] + ϵ)Φ = (L + λV − σ[L, B[Ω0 ], Ω0 ] + ϵ)ξδj
> (σ[L, D, Nδ0,j ] − σ[L, B[Ω0 ], Ω0 ] + ϵ)ξδj > 0
for all 1 ≤ j ≤ n0 . Now, note that
 
∪
n0
Ω̄ \ Ωδ/2 ∪ 0,j 
Nδ/2
j=1
is a compact subset of
∪
p
i
Ω+ ∪ Γ1j
j=1
and hence, according to Definition 9.1(b), there exists a constant ω > 0

such that
 
∪
n0
V ≥ω>0 in Ω̄ \ Ω̄δ/2 ∪ 0,j 
N̄δ/2 . (9.50)
j=1
1,i
Thanks to (9.50), for every 1 ≤ j ≤ p, in N̄δ/2j we have that
i
(L + λV − σ[L, B[Ω0 ], Ω0 ] + ϵ)Φ = (L + λV − σ[L, B[Ω0 ], Ω0 ] + ϵ)ψδj
1,ij 1,ij i
≥ (σ[L, B[Nδ ], Nδ ] − σ[L, B[Ω0 ], Ω0 ] + ϵ + λω)ψδj > 0
provided
{ ( ) }
ω −1 σ[L, B[Ω0 ], Ω0 ] − ϵ − σ[L, B[Nδ j ], Nδ j ] , 0 ,
1,i 1,i
λ > max
whereas in
 
∪
p ∪
n0
Ω̄ \ Ω̄δ/2 ∪ 0,j 
1,i
N̄δ/2j ∪ N̄δ/2
j=1 j=1
we have that
(L + λV − σ[L, B[Ω0 ], Ω0 ] + ϵ)Φ = (L + λV − σ[L, B[Ω0 ], Ω0 ] + ϵ)ζδ
≥ (L − σ[L, B[Ω0 ], Ω0 ] + ϵ)ζδ + λωζδ > 0
for sufficiently large λ > 0, because the function
(L − σ[L, B[Ω0 ], Ω0 ] + ϵ)ζδ
does not depend on λ, and ζδ is positive and bounded away from zero.
Finally, by construction,
BΦ = Dξδj = 0 on Γj0 , 1 ≤ j ≤ n0 ,
i i
BΦ = Bψδj = 0 on Γ1j , 1 ≤ j ≤ p,
and
BΦ = Bφδ = 0 on ∂Ω0 ∩ Γ1 .
This completes the proof of the theorem under condition (9.44).
Step 2: Suppose
m = 1, K = ∅, and Γ0 ∩ ∂Ω0 ̸= ∅, (9.51)
instead of (9.44). Let Γi0 , 1 ≤ i ≤ n0 , denote the components of Γ0 , and
{i1 , ..., iq } be the subset of {1, ..., n0 } for which
∂Ω0 ∩ Γj0 ̸= ∅ ⇐⇒ j ∈ {i1 , ..., iq }.
Figure 9.7 illustrates a possible nodal configuration of V satisfying these
requirements with q = 1 and i1 = 2.
Γ1
0
2
Γ
0
Ω
0
1
Γ 2
1 Γ
1
Ω
+
Fig. 9.7 A nodal configuration satisfying (9.51)
Fix ϵ > 0 and, for sufficiently small η > 0, consider the auxiliary domain
 
∪q
Gη := Ω ∪  Γ0j + Bη  .
i
j=1
In the special case illustrated by Figure 9.7, it becomes apparent that
Gη = Ω ∪ (Γ20 + Bη );
equivalently, Gη consists of Ω plus the set of points x ∈ RN such that

dist (x, Γ20 ) < η.
By (8.26), we have that
aij ∈ C 1 (Ω̄), bi ∈ C(Ω̄), 1 ≤ i, j ≤ N.
Fix η > 0 and let
ãij = ãji ∈ C 1 (Ḡη ), b̃i ∈ C(Ḡη ), c̃ ∈ L∞ (Gη ), 1 ≤ i, j ≤ N,
be regular extensions from Ω̄ to Ḡη of the coefficients aij = aji , bi , and c,

1 ≤ i, j ≤ N , respectively.
Now, consider the differential operator
L̃ := − div (Ã∇ · ) + ⟨b̃, ∇ · ⟩ + c̃
in Gη , where we have denoted
Ã := (ãij )1≤i,j≤N , b̃ = (b̃1 , ..., b̃N ).

Ω+
Gη
Ω0
Fig. 9.8 Enlarging Ω through Γ20
As L is uniformly elliptic in Ω with constant µ > 0, there exists η̃ ∈ (0, η)

such that L̃ is uniformly elliptic in Gη̃ with constant µ/2. Subsequently,
we set
Ω̃ := Gη̃
and consider the extended potential
{
1 in Ω̃ \ Ω ,
Ṽ :=
V in Ω ,
and the boundary operator
{
D on ∂ Ω̃ \ Γ1 ,
B̃ :=
B, on Γ1 .
By construction, it is easy to check that Ṽ ∈ Ã, where Ã stands for the
class of admissible potentials in the open set Ω̃. Moreover,
Ω̃0 = Ω0 , Γ̃1 = Γ1 , (∂ Ω̃ \ Γ̃1 ) ∩ ∂ Ω̃0 = ∅.
Thus, the extended problem in Ω̃ satisfies (9.44) and, consequently, by Step
1, there exist λ̃1 > 0 and a function
¯ → [0, ∞)
Φ̃ : Ω̃
such that Φ̃(x) > 0 for all x ∈ Ω̃, and it is a strict supersolution of
( )
L̃ + λṼ − σ[L̃, B̃[Ω̃0 ], Ω̃0 ] + ϵ, B̃, Ω̃
for all λ > λ̃1 . Note that

σ[L̃, B̃[Ω̃0 ], Ω̃0 ] = σ[L̃, B̃[Ω0 ], Ω0 ] = σ[L, B[Ω0 ], Ω0 ],
because Ω̃0 = Ω0 , L̃ = L in Ω̄0 , and B̃[Ω0 ] = B[Ω0 ]. Therefore, Φ̃ provides

us with a strict supersolution of
( )
L̃ + λṼ − σ[L, B[Ω0 ], Ω0 ] + ϵ, B̃, Ω̃
for all λ > λ̃1 . Subsequently, we consider the restriction

Φ := Φ̃|Ω̄ .
Obviously, Φ(x) > 0 for all x ∈ Ω. Moreover, in Ω, we have that
(L + λV − σ[L, B[Ω0 ], Ω0 ] + ϵ)Φ = (L̃ + λṼ − σ[L, B[Ω0 ], Ω0 ] + ϵ)Φ̃ ≥ 0
for all λ > λ̃1 . Also,
∪
q
i
Φ(x) = Φ̃(x) > 0 for all x ∈ Γ0j ,
j=1
because, by construction,
∪
q
i
Γ0j ⊂ Ω̃ and Φ̃(x) > 0 for all x ∈ Ω̃.
j=1
In addition,
∪
q
i
Φ = Φ̃ = 0 on Γ0 \ Γ0j
j=1
and
BΦ = (∂ν + β)Φ̃ ≥ 0 on Γ1 ,
by the construction of Φ̃. It should be noted that, in a neighborhood of Γ1 ,
Φ̃ = Φ and, hence, not only βΦ = β Φ̃, but also ∂ν Φ = ∂ν Φ̃.
Consequently, BΦ > 0 on ∂Ω, and, therefore, Φ provides us with a
positive strict supersolution of (9.43) for all λ > λ̃1 . This completes the
proof of Step 2.
Step 3: Now, suppose
m ≥ 1, K ̸= ∅, and Γ0 ∩ (∂Ω0 ∪ K) = ∅. (9.52)
Figure 9.9 shows an admissible situation where (9.52) holds. In this ex-
ample, Γ0 consists of two components, Γ10 and Γ20 , as well as Γ1 , whose
components have been named Γ11 and Γ21 , and Ω0 , whose components are
Ω10 and Ω20 . So, in this example, m = 2. In Figure 9.9, V = 0 in
Ω0 := Ω10 ∪ Ω20 ,
2
Ω
K 0
2
Γ0
1
Γ
0
2
Γ
1
1
Γ1
1
Ω
0
Ω+
Fig. 9.9 V (x) > 0 for all x ∈ Ω+ ∪ Γ11 and V = 0 in Ω10 ∪ Ω20 ∪ K
and in K, which is a negligible set. As due to (9.36),

K ∩ (Ω̄0 ∪ Γ1 ) = ∅, (9.53)
all structural requirements of Definition 9.1 are fulfilled, as well as (9.52),
as soon as, for instance, V ∈ C(Ω̄) satisfies V (x) > 0 for all x ∈ Ω+ ∪ Γ11
and V (x) = 0 for all x ∈ Ω10 ∪ Ω20 ∪ K.
By (9.52) and (9.53), we have that K ∩ Γ1 = ∅ and K ∩ Γ0 = ∅,
respectively. Thus, K ∩ ∂Ω = ∅ and hence, K ⊂ Ω. Moreover, by (9.52),
Ω̄0 ⊂ Ω ∪ Γ1 , and, due to (9.53), K ∩ Ω̄0 = ∅. Summarizing,
K ⊂ Ω, Ω̄0 ⊂ Ω ∪ Γ1 , K ∩ Ω̄0 = ∅, (9.54)
and, therefore,
dist(Γ0 , Ω̄0 ∪ K) > 0, dist(Γ1 , K) > 0, dist(K, Ω̄0 ) > 0, (9.55)
because each of the pairs consists of two disjoint compact subsets.
Subsequently, we denote by Ωi0 , 1 ≤ i ≤ m, the components of Ω0 .
Without loss of generality, they can be relabeled so that
σ[L, B, Ωi0 ] ≤ σ[L, B, Ωi+1
0 ], 1 ≤ i ≤ m − 1,
and, consequently, according to Definition 9.2, we have that
σ[L, B[Ω0 ], Ω0 ] = σ[L, B[Ω10 ], Ω10 ].
Fix η > 0. Thanks to Definition 9.1(d), there exist a natural number
ℓ(η) ≥ 1 and ℓ(η) open sets Gηj ⊂ RN , 1 ≤ j ≤ ℓ(η), with |Gηj | < η,
1 ≤ j ≤ ℓ(η), such that
∪
ℓ(η)
K⊂ (Gηj ∩ Ω), Ḡηi ∩ Ḡηj = ∅ if i ̸= j,
j=1
and Gηj ∩ Ω is connected and of class C 2 for all 1 ≤ j ≤ ℓ(η). Thanks to

(9.54), we can choose the Gηj ’s so that
∪
ℓ(η)
∪
ℓ(η)
K⊂ Ḡηj ⊂ Ω, Ḡηj ∩ Ω̄0 = ∅. (9.56)
j=1 j=1
Indeed, since
dist(K, Ω̄0 ∪ Γ0 ∪ Γ1 ) > 0,
there exists an open set G such that
K ⊂ G, Ḡ ⊂ Ω, Ḡ ∩ Ω̄0 = ∅,
and hence, to get (9.56) it suffices to take G ∩ Gηj , instead of Gηj , 1 ≤ j ≤
ℓ(η). In the particular case covered by Figure 9.9, we can take ℓ(η) = 1 for
all η > 0, because K is connected, and, actually,
Gη1 = K + Bδ
for an appropriate sufficiently small δ > 0.
Owing to Proposition 8.6, there exists η0 > 0 such that, for every η ∈
(0, η0 ) and 1 ≤ j ≤ ℓ(η),
√
σ[L, D, Gηj ] ≥ µΣ1 |B1 | N η − N − ∥b∥∞ Σ|B1 | N η − N + inf c.
2 2 1 1
Therefore, there exists η1 ∈ (0, η0 ) such that

σ[L, B[Ωm m
0 ], Ω0 ] < min σ[L, D, Gηj ] for all η ∈ (0, η1 ). (9.57)
1≤j≤ℓ(η)
Without loss of generality, by rearranging the Gηj ’s, if necessary, we can

assume that
σ[L, D, Gηj ] ≤ σ[L, D, Gηj+1 ], 1 ≤ j ≤ ℓ(η) − 1.
Fix η ∈ (0, η1 ) and consider the δ-neighborhoods of the components Ωi0 ’s
Ωiδ := (Ωi0 + Bδ ) ∩ Ω
for all 1 ≤ i ≤ m and sufficiently small δ > 0. As Ω̄i0 ∩ Ω̄j0 = ∅ if i ̸= j,
there exists δ0 > 0 such that
Ω̄iδ ∩ Ω̄jδ = ∅ if i ̸= j (9.58)
for all 0 < δ < δ0 . Moreover, by (9.56), there exists δ1 ∈ (0, δ0 ) such that
  ( )
∪ η
ℓ(η)
∪
m
 Ḡj  ∩ Ω̄iδ = ∅ for all 0 < δ < δ0 . (9.59)
j=1 i=1
Subsequently, we consider the m potentials

{
V in Ω̄iδ
Vi := 1 ≤ i ≤ m. (9.60)
1 in Ω̄ \ Ω̄iδ ,
We claim that
Vi ∈ A, 1 ≤ i ≤ m. (9.61)
Note that, for each 1 ≤ i ≤ m, the vanishing set associated to Vi is Ωi0 ,

which is connected, and that the corresponding K, say Ki , is the empty
set, since (9.56) and (9.59) imply that
K ∩ (Ω̄iδ ∪ Γ1 ) = ∅, 1 ≤ i ≤ m,
and Vi = 1 in the complement of Ω̄iδ . Therefore, if (9.61) holds, necessarily

mi = 1, Ki = ∅ and, due to (9.52), ∂Ωi0 ∩ Γ0 = ∅, where mi stands for
the number of components of Ωi0 . Consequently, Vi would be an admissible
potential to apply the previous Step 1. To prove (9.61) we proceed as
follows. Thanks to (9.37), (9.56), (9.58) and (9.59), we have that
∪
m
( )
Ω∩ Ω̄iδ \ Ω̄i0 ⊂ Ω+
i=1
and hence, for every 1 ≤ i ≤ m, Vi is bounded away from zero in any

compact subset of Ω+ , as V is bounded away from zero in any compact
subset of Ω+ . Let Γj1 , 1 ≤ j ≤ n1 , denote the components of Γ1 and, for
every 1 ≤ i ≤ m, let {j1 , ..., jpi } denote the subset of {1, ..., n1 } for which
Γj1 ∩ ∂Ωi0 = ∅ ⇐⇒ j ∈ {j1 , ..., jpi }.
Then, for each 1 ≤ i ≤ m we have that

∪
pi
∂Ωi0 ∩ Γj1k = ∅
k=1
and hence,
∪
pi
dist(∂Ωi0 , Γj1k ) > 0, 1 ≤ i ≤ m.
k=1
Thus, there exists δ2 ∈ (0, δ1 ) such that

(p )
∪ i
jk
Γ1 + Bδ ∩ Ω̄iδ = ∅ (9.62)
k=1
for all 1 ≤ i ≤ m and 0 < δ < δ2 . Fix δ ∈ (0, δ2 ). Then, it follows from
(9.60) and (9.62) that
(p )
∪i
jk
Vi = 1 in Γ1 + Bδ ∩ Ω
k=1
for all 1 ≤ i ≤ m. Consequently, for every 1 ≤ i ≤ m, Vi is bounded away

from zero in any compact subset of
∪
pi
Ω+ ∪ Γj1k
k=1
and, therefore, (9.61) indeed holds.

By the result of Step 1, for every ϵ > 0 there exist λ1 = λ1 (ϵ) > 0 and
m smooth functions
Φi : Ω̄ → [0, ∞), 1 ≤ i ≤ m,
such that
Φi (x) > 0 for all x ∈ Ω and 1 ≤ i ≤ m, (9.63)
and Φi is a strict supersolution of
(L + λVi − σ[L, B[Ωi0 ], Ωi0 ] + ϵ, B, Ω)
for all 1 ≤ i ≤ m and λ > λ1 .
Similarly, we will consider the ℓ(η) potentials
{
0 in Gηj ,
V̂j := 1 ≤ j ≤ ℓ(η). (9.64)
1 in Ω̄ \ Gηj ,
By definition, for every 1 ≤ j ≤ ℓ(η), the vanishing set of V̂j equals Gηj ,
which is connected and of class C 2 . Moreover, by (9.56), Ḡηj ⊂ Ω. Thus,
there exists ρ > 0 such that
V̂j = 1 in (Γ1 + Bρ ) ∩ Ω̄,
and hence,
V̂j ∈ A, 1 ≤ j ≤ ℓ(η).
Actually, since
Γ0 ∩ Gηj = ∅, 1 ≤ j ≤ ℓ(η),
each of these potentials fits into the abstract framework of Step 1. Conse-
quently, there are λ2 = λ2 (ϵ) > λ1 (ϵ) and ℓ(η) smooth functions
Φ̂j : Ω̄ → [0, ∞), 1 ≤ j ≤ ℓ(η),
such that
Φ̂j (x) > 0 for all x ∈ Ω, 1 ≤ j ≤ ℓ(η), (9.65)
and Φ̂j is a strict supersolution of
(L + λV̂j − σ[L, B[Gηj ], Gηj ] + ϵ, B, Ω)
for all 1 ≤ j ≤ ℓ(η) and λ > λ2 . Note that Ḡηj ⊂ Ω, 1 ≤ j ≤ ℓ(η), implies
B[Gηj ] = D, 1 ≤ j ≤ ℓ(η),
and, therefore, Φ̂j is a strict supersolution of
(L + λV̂j − σ[L, D, Gηj ] + ϵ, B, Ω)
for all 1 ≤ j ≤ ℓ(η) and λ > λ2 .
Let Γj1 , 1 ≤ j ≤ n1 , be the components of Γ1 and let {i1 , ..., ip } denote
the subset of {1, ..., n1 } for which
Γj1 ∩ ∂Ω0 = ∅ ⇐⇒ j ∈ {i1 , ..., ip }.
According to Definition 9.1(a), Γj1 is a component of ∂Ω0 for all j ∈
{1, ..., n1 } \ {i1 , ..., ip }. Moreover,
∪
p
i
Γ1j ∩ ∂Ω0 = ∅
j=1
and hence,
∪
p
i
dist ( Γ1j , ∂Ω0 ) > 0. (9.66)
j=1
Now, consider the δ-neighborhoods Nδ0,j and Nδ1,j defined in (9.46). Thanks
to (9.52), (9.56) and (9.66), there exists δ3 ∈ (0, δ2 ) such that
   
∪p ∪
n0 ∪
m ∪
ℓ(η)
 1,i
N̄δ j ∪ N̄δ0,j  ∩  Ω̄jδ ∪ Ḡηj  = ∅ (9.67)
j=1 j=1 j=1 j=1
for all 0 < δ < δ3 . Moreover, since

Γjk ∩ Γiℓ = ∅ if (i, ℓ) ̸= (j, k),
there exists δ4 ∈ (0, δ3 ) such that
N̄δk,j ∩ N̄δℓ,i = ∅ if (i, ℓ) ̸= (j, k), k , ℓ ∈ {0, 1}, (9.68)
for each 0 < δ < δ4 . Furthermore, since
lim |Nδ0,j | = 0 for each 1 ≤ j ≤ n0 ,
δ↓0
by Proposition 8.6, there exists δ5 ∈ (0, δ4 ) such that

σ[L, D, Nδ0,j ] > σ[L, B[Ω0 ], Ω0 ], 1 ≤ j ≤ n0 , (9.69)
for all 0 < δ < δ5 . Finally, let i ∈ {i1 , ..., ip }, and
ψδi , ξδj , 1 ≤ j ≤ n0 ,
denote the principal eigenfunctions associated to
σ[L, B[Nδ1,i ], Nδ1,i ], i ∈ {i1 , ..., ip }, and σ[L, D, Nδ0,j ], 1 ≤ j ≤ n0 ,
respectively.
Thanks to (9.59), (9.67) and (9.68), the next function is well defined


 Φi in Ω̄iδ , 1 ≤ i ≤ m,







 Φ̂j in Ḡηj , 1 ≤ j ≤ ℓ(η),





i 1,i
Φ := ψδj in N̄δ/2j , 1 ≤ j ≤ p, (9.70)






 ξj

0,j
in N̄δ/2 , 1 ≤ j ≤ n0 ,

 δ





ζδ in Ω̄ \ Kδ,η ,
∪
m ∪
ℓ(η)
∪
p
1,i
∪
n0
Kδ,η := Ω̄iδ ∪ Ḡηj ∪ N̄δ/2j ∪ 0,j
N̄δ/2
i=1 j=1 j=1 j=1
and ζδ is any positive regular extension of Φi , 1 ≤ i ≤ m, Φ̂j , 1 ≤ j ≤ ℓ(η),

i
ψδj , 1 ≤ j ≤ p, and ξδj , 1 ≤ j ≤ n0 , from Kδ,η to Ω̄ which is bounded away
from zero in Ω̄ \ Kδ,η . The function ζδ does exit because, thanks to (9.56),
(9.63) and (9.65), the restrictions
Φi |∂Ωiδ \Γ1 , Φ̂j |∂Gηj , 1 ≤ i ≤ m, 1 ≤ j ≤ ℓ(η),
i
ψδj |∂N 1,ij \Γ , ξδi |∂N 0,i \Γ0 , 1 ≤ j ≤ p, 1 ≤ i ≤ n0 ,
δ/2 1 δ/2
i
are positive and bounded away from zero. As in Step 1, the functions ψδj ’s
should not appear in (9.70) if Γ1 ⊂ ∂Ω0 .
According to the definition of Φi , for each 1 ≤ i ≤ m, the following
estimates hold in Ωiδ
( )
(L+λV −σ[L, B[Ω0 ], Ω0 ]+ϵ) Φ = L+λVi −σ[L, B[Ω10 ], Ω10 ]+ϵ Φi
( )
≥ L+λVi −σ[L, B[Ωi0 ], Ωi0 ]+ϵ Φi
≥0
for all λ > λ2 , because V = Vi in Ωiδ . Moreover, thanks to (9.57), for

every 1 ≤ j ≤ ℓ(η), it follows from the definition of Φ̂j that the following
estimates are satisfied in Gηj
(L+λV −σ[L, B[Ω0 ], Ω0 ]+ϵ) Φ = (L+λV̂j −σ[L, B[Ω10 ], Ω10 ]+ϵ)Φ̂j

> (L+λV̂j −σ[L, D, Gηj ]+ϵ)Φ̂j
≥0
for all λ > λ2 , because V ≥ V̂j = 0 in Gηj and λ > 0.
On the other hand, by Definition 9.1(b), V is positive and bounded
away from zero in any compact subset of
∪
p
i
Ω+ ∪ Γ1j
j=1
and hence there exists ω > 0 such that

∪
p
1,i
V ≥ω>0 in Nδ/2j .
j=1
1,i
Thus, for every 1 ≤ j ≤ p, the following estimates hold in Nδ/2j
i
(L+λV −σ[L, B[Ω0 ], Ω0 ]+ϵ)Φ = (L+λV −σ[L, B[Ω0 ], Ω0 ]+ϵ)ψδj
1,ij 1,ij i
= (σ[L, B[Nδ ], Nδ ] + λV − σ[L, B[Ω0 ], Ω0 ] + ϵ)ψδj
1,ij 1,ij i
> (σ[L, B[Nδ ], Nδ ] + λω − σ[L, B[Ω0 ], Ω0 ] + ϵ)ψδj
>0
for sufficiently large λ > 0. Also, due (9.69), for each 1 ≤ j ≤ n0 , we find
from the definitions of Φ and ξδj that the next estimates hold in Nδ/2
0,j
(L + λV − σ[L, B[Ω0 ], Ω0 ] + ϵ)Φ = (L + λV − σ[L, B[Ω0 ], Ω0 ] + ϵ)ξδj

= (σ[L, D, Nδ0,j ] + λV − σ[L, B[Ω10 ], Ω10 ] + ϵ)ξδj
> (σ[L, D, Nδ0,j ] − σ[L, B[Ω10 ], Ω10 ] + ϵ)ξδj > 0
for all λ > 0, while, in Ω̄ \ Kδ,η , we have that
(L + λV − σ[L, B[Ω0 ], Ω0 ] + ϵ)Φ = (L + λV − σ[L, B[Ω0 ], Ω0 ] + ϵ)ζδ
≥ (L − σ[L, B[Ω0 ], Ω0 ] + ϵ)ζδ + λV ζδ > 0
for sufficiently large λ > 0, because (L−σ[L, B[Ω0 ], Ω0 ]+ϵ)ζδ is independent
of λ and V ζδ is separated away from zero in Ω̄ \ Kδ,η , since the closure of
this set is a compact subset of Ω+ .
Finally, by construction,
BΦ = DΦ = ξδj = 0 on Γj0 , 1 ≤ j ≤ n0 ,
i i
BΦ = (∂ν + b)Φ = (∂ν + b)ψδj = 0 on Γ1j , 1 ≤ j ≤ p,
and, for every 1 ≤ j ≤ m such that ∂Ωj0 ∩ Γ1 ̸= ∅,
BΦ = (∂ν + b)Φ = (∂ν + b)Φj ≥ 0 on ∂Ωjδ ∩ Γ1 = ∂Ωj0 ∩ Γ1 .
Therefore,
BΦ ≥ 0 on ∂Ω
and, consequently, the function Φ defined by (9.70) provides us with a

positive strict supersolution of (9.43) for sufficiently large λ > 0. This
completes the proof of Step 3.
Step 4: Finally, suppose
m ≥ 1, K ̸= ∅, and Γ0 ∩ (∂Ω0 ∪ K) ̸= ∅. (9.71)
Let Γj0 , 1 ≤ j ≤ n0 , be the components of Γ0 , and let {i1 , ..., iq } denote the
subset of {1, ..., n0 } for which
Γj0 ∩ (∂Ω0 ∪ K) ̸= ∅ ⇐⇒ j ∈ {i1 , ..., iq }.
As in the proof of Step 2, for sufficiently small η > 0, we consider the

extended open set
 
∪q
Ω̃ := Gη := Ω ∪  Γ0j + Bη  .
i
j=1
The rest of the proof consists in constructing L̃, Ṽ and B̃, as in the proof
of Step 2, satisfying
m̃ ≥ 1, K̃ ̸= ∅, and Γ̃0 ∩ (∂ Ω̃0 ∪ K̃) = ∅.
Reasoning as in the proof of the second part of Step 2, but this time using
the result of Step 3, instead of the result of Step 1, ends the proof.
9.4 Characterizing the existence of principal eigenvalues for

admissible potentials
The main result for sign definite potentials W reads as follows.
Theorem 9.7. Suppose W ∈ L∞ (Ω) and W ≥ 0. Then, the following

assertions are true:
(a) If
inf W > 0, (9.72)
Ω
then (9.1) possesses a unique principal eigenvalue.
(b) Suppose
inf W = 0, (9.73)
Ω
W ∈ A in the sense of Definition 9.1 with int Ω+ ̸= ∅, and the coeffi-
cients of L satisfy (8.26). Then, (9.1) admits a principal eigenvalue if,
and only if,
σ[L, B[Ω0 ], Ω0 ] > 0, (9.74)
and it is unique if it exists.
Similarly, (9.1) possesses a (unique) principal eigenvalue if
sup W < 0, (9.75)
Ω
whereas, if
sup W = 0, (9.76)
Ω
−W ∈ A with Ω+ ̸= ∅ and L satisfies (8.26), then (9.1) possesses a prin-
cipal eigenvalue if, and only if, (9.74) holds, and it is unique if it exists.
Proof. Suppose (9.72). Then, for every λ < 0,

Σ(λ) = σ[L − λW, B, Ω] ≥ σ[L, B, Ω] − λ inf W,
Ω
and hence,
lim Σ(λ) = ∞.
λ↓−∞
Thus, (9.13) holds and, therefore, Theorem 9.2 ends the proof of Part (a).
Now, suppose W satisfies (9.73), W ∈ A, and L satisfies (8.26). Then,
according to Theorem 9.6, we have that
lim σ[L − λW, B, Ω] = σ[L, B[Ω0 ], Ω0 ]
λ↓−∞
and, consequently, Part (b) again is a corollary from Theorem 9.2.

Subsequently, we suppose W ≤ 0, instead of W ≥ 0. If, in addition,
(9.75) is satisfied, then,
Σ(λ) = σ[L − λW, B, Ω] ≥ σ[L, B, Ω] − λ sup W
Ω
for all λ > 0 and hence,

lim Σ(λ) = ∞.
λ↑∞
Consequently, (9.19) holds and Theorem 9.3 ends the proof.

Finally, suppose W satisfies (9.73), −W ∈ A with Ω+ ̸= ∅ and L satisfies
(8.26). Then, thanks to Theorem 9.6,
lim σ[L − λW, B, Ω] = σ[L, B[Ω0 ], Ω0 ]
λ↑∞
and, consequently, in such case, the result also follows from Theorem 9.3.
When W changes sign in Ω, i.e., there are x+ , x− ∈ Ω and R > 0 for
which (9.12) and (9.18) are satisfied, and
σ[L, B, Ω] > 0,
then, it follows from Theorem 9.4 that (9.1) possesses two principal eigen-
values λ∗− < 0 < λ∗+ . If (L, B, Ω) does not satisfy the maximum principle,
then, the next result holds.
Theorem 9.8. Suppose L satisfies (8.26) and

σ[L, B, Ω] < 0. (9.77)
∞
Let W ∈ L (Ω) satisfy (9.12) and (9.18) for some x+ , x− ∈ Ω and R > 0.
Set
W + := max{W, 0}, W − := W + − W, (9.78)
and assume one of the following two conditions holds:
(a) W + ∈ A and
Σ+ (λ)
σ[L, B[Ω+ +
0 ], Ω0 ] > 0, ∥W − ∥L∞ (Ω) < max+ , (9.79)
λ<λ −λ
where Ω+
0 is the set Ω0 associated to W
+
through Definition 9.1(a),
Σ+ (λ) := σ[L − λW + , B, Ω], λ ∈ R, (9.80)
and λ+ < 0 is the unique zero of Σ+ .
(b) W − ∈ A and
Σ− (λ)
σ[L, B[Ω− −
0 ], Ω0 ] > 0, ∥W + ∥L∞ (Ω) < max , (9.81)
λ>λ− λ
where Ω−
0 is the set Ω0 associated to W
−
through Definition 9.1(a),
Σ− (λ) := σ[L + λW − , B, Ω], λ ∈ R, (9.82)
and λ− > 0 is the unique zero of Σ− .
Then, (9.1) possesses exactly two principal eigenvalues. Moreover, in case
(a) both are negative, while both are positive in case (b).
−
Owing to Corollary 8.4, condition (9.79) holds if |Ω+ 0 | and ∥W ∥∞ are
sufficiently small and β is sufficiently large on Γ1 ∩ ∂Ω0 . Similarly, (9.81)
+
holds if |Ω−
0 | and ∥W ∥∞ are sufficiently small and β is sufficiently large
+
−
on Γ1 ∩ ∂Ω0 . Naturally, the restrictions on the size of β are unnecessary if
−
Γ1 ∩ ∂Ω+ 0 = ∅, or Γ1 ∩ ∂Ω0 = ∅.
Proof. First, we will make sure that (9.79) and (9.81) make sense. As
W satisfies (9.12), it becomes apparent that
+ ̸= ∅
int Ω+
if W + ∈ A, where Ω++ is the set Ω+ associated to W
+
through Definition
9.1(a). Consequently, if W + ∈ A satisfies
σ[L, B[Ω+ +
0 ], Ω0 ] > 0,
then, according to Theorems 9.1 and 9.6,

lim Σ+ (λ) = σ[L, B[Ω+ +
0 ], Ω0 ] > 0, lim Σ+ (λ) = −∞,
λ↓−∞ λ↑∞
and, therefore, there exists a unique value of λ, denoted by λ+ , such that

Σ+ (λ+ ) = 0. By (9.77), Σ(0) < 0. Thus, since Σ+ is decreasing, necessarily
λ+ < 0. Subsequently, we set
Σ+ (λ)
q+ (λ) := , λ ≤ λ+ .
−λ
By construction, q+ (λ+ ) = 0. Moreover,
Σ+ (λ) > Σ+ (λ+ ) = 0
for all λ < λ+ < 0 and hence, q+ (λ) > 0 for all λ < λ+ . As
lim Σ+ (λ) = σ[L, B[Ω+ +
0 ], Ω0 ],
λ↓−∞
we have that
lim q+ (λ) = 0
λ↓−∞
and, consequently,
Σ+ (λ)
max+ = max+ q+ (λ) ∈ (0, ∞)
λ<λ −λ λ<λ
is well defined and condition (9.79) makes sense.

Now, suppose W − ∈ A satisfies
σ[L, B[Ω− −
0 ], Ω0 ] > 0.
According to (9.18), we have that

int Ω−
+ ̸= ∅
where Ω− + is the set Ω+ associated to W

−
through Definition 9.1(a). More-
over, arguing as above, we have that
lim Σ− (λ) = −∞, lim Σ− (λ) = σ[L, B[Ω− −
0 ], Ω0 ] > 0,
λ↓−∞ λ↑∞
and hence, there exists a unique λ− such that Σ− (λ− ) = 0. By (9.77),

Σ(0) < 0. Thus, since Σ− is increasing, necessarily λ− > 0. From these
features, it becomes apparent that the auxiliary function defined by
Σ− (λ)
q− (λ) := , λ ≥ λ− ,
λ
satisfies
q− (λ− ) = 0, lim q− (λ) = 0,
λ↑∞
and q− (λ) > 0 for all λ > λ− . Therefore, (9.81) is consistent too.
Suppose W + ∈ A satisfies (9.79). Then, there exists λ̃ < λ+ < 0 such
that
Σ+ (λ̃)
∥W − ∥L∞ (Ω) <
−λ̃
and hence, by (9.80),
Σ+ (λ̃) := σ[L − λ̃W + , B, Ω] > −λ̃∥W − ∥L∞ (Ω) .
Therefore, since λ̃ < 0,
Σ(λ̃) = σ[L − λ̃W, B, Ω] = σ[L − λ̃W + + λ̃W − , B, Ω]
≥ σ[L − λ̃W + , B, Ω] + λ̃∥W − ∥L∞ (Ω) > 0
and, consequently, the conclusions hold from Theorem 9.4.
Similarly, if W − ∈ A satisfies (9.81), there is λ̃ > λ− > 0 such that
Σ− (λ̃)
∥W + ∥L∞ (Ω) <
λ̃
and hence,
Σ− (λ̃) := σ[L + λ̃W − , B, Ω] > λ̃∥W + ∥L∞ (Ω) .
Therefore,
Σ(λ̃) = σ[L − λ̃W, B, Ω] = σ[L − λ̃W + + λ̃W − , B, Ω]

≥ σ[L + λ̃W − , B, Ω] − λ̃∥W + ∥L∞ (Ω) > 0
and Theorem 9.4 ends the proof.
The analysis of the classical case when the potential W has definite sign
and it is bounded away from zero does not entail any special difficulty and
it goes back, at least, to R. Courant and D. Hilbert [44]. Seemingly, the
analysis of (9.1) in the most general and interesting situation when the
potential W changes sign goes back to the works of A. Manes and A. M.
Micheletti [157], P. Hess and T. Kato [97], and K. J. Brown and S. S. Lin
[33].
In the special case when Γ1 = ∅ (B = D) and L is self-adjoint with
coercive associated bilinear form, A. Manes and A. M. Micheletti [157]
established the existence of two principal eigenvalues, λ− < 0 < λ+ , from
a strong maximum principle of G. Stampacchia [214]. This result was later
extended by P. Hess and T. Kato [97] to cover a more general class of
operators L, not necessarily self-adjoint, with c ≥ 0, by establishing that
if W is continuous and W (x+ ) > 0 for some x+ ∈ Ω, then (9.1) admits a
unique positive principal eigenvalue. Almost simultaneously, K. J. Brown
and S. S. Lin [33] extended these results to cover the special case when
L = −∆, Γ0 = ∅ and β = 0 (B = N). In such case, their main result
established that (9.1) possesses a positive principal eigenvalue, λ+ > 0, if,
and only if,
∫
W < 0. (9.83)
Ω
Moreover, λ+ is unique and algebraically simple if it exists.

In the settings of A. Manes and A. M. Micheletti [157] and P. Hess and

T. Kato [97], Γ1 = ∅ and
σ[L, B, Ω] = σ[L, D, Ω] > 0. (9.84)
Therefore, the existence and the uniqueness of a positive and a negative
eigenvalue is a consequence from Theorem 9.4.
Incidentally, in the framework of K. J. Brown and S. S. Lin [33],
σ[L, B, Ω] = σ[−∆, N, Ω] = 0, (9.85)
since 0 is the principal eigenvalue of −∆ under Neumann boundary con-
ditions in Ω. Therefore, 0 must be a principal eigenvalue of (9.1) and,
according to Theorem 9.4, (9.1) admits a positive eigenvalue if, and only
if, Σ′ (0) > 0, where
Σ(λ) := σ[−∆ − λW, N, Ω], λ ∈ R.
Indeed, 0 must be the unique eigenvalue of (9.1) if Σ′ (0) = 0, while the
second eigenvalue of (9.1) must be negative if Σ′ (0) < 0. Note that
Σ(0) = 0 and Σ′ (0) ̸= 0 imply max Σ(λ) > 0.
λ∈R
Moreover, adapting the last part of the proof of Theorem 9.2, with the
choice φ(0) = 1, it is easy to check that
∫
1
Σ′ (0) = − W
|Ω| Ω
and, consequently,
∫
′
Σ (0) > 0 ⇐⇒ W < 0,
Ω
which provides us with the main theorem of K. J. Brown and S. S. Lin [33].
Summarizing, the most pioneering results established that if either
σ[L, B, Ω] > 0, (9.86)
or
σ[L, B, Ω] = 0 and Σ′ (0) > 0, (9.87)
then (9.1) has two principal eigenvalues
λ− ≤ 0 < λ+ .
According to Theorem 7.10, we already know that (9.86) occurs if, and only
if, (L, B, Ω) satisfies the strong maximum principle, which is a very severe
restriction. Actually, this chapter has developed a general abstract theory

to overcome it.
Even in the special case when W has definite sign, the results of this
chapter provide with some substantial and very deep extensions of all clas-
sical available results. Indeed, according to Theorem 9.6, if W ≥ 0 and
Ω0 := int W −1 (0)
is a nice open subset of Ω, then (9.1) admits a principle eigenvalue, neces-
sarily unique, if and only if
σ[L, B[Ω0 ], Ω0 ] > 0. (9.88)
Moreover, by Theorem 7.10, this occurs if, and only if, (L, B[Ω0 ], Ω0 ) satis-
fies the strong maximum principle, and, due to Corollary 8.4, this holds true
whenever |Ω0 | is sufficiently small and β is sufficiently large on Γ1 ∩ ∂Ω0 .
Theorem 9.6 goes back to J. López-Gómez [135, 137] when Γ1 = ∅, and
to S. Cano-Casanova and J. López-Gómez [39] in the general setting covered
in this chapter, as well as the general abstract theory of this chapter in its
full generality. Essentially, this chapter polishes the contents of [39].
Besides Theorem 9.6 is a pivotal result for characterizing the existence of
principal eigenvalues of (9.1), it has a number of fundamental applications
in a variety of scientific fields.
In the context of the semiclassical analysis of Schrödinger operators,
one is interested in the problem of analyzing the behavior as h ↓ 0 of the
principal eigenvalue and associated eigenfunctions of the linear boundary
value problem
{ 2
h Lφ + V (x)φ = σφ in Ω,
(9.89)
φ=0 on ∂Ω.
In quantum physics,
h ∼ 6.624 × 10−27 erg/sec
is Planck’s constant. Except at atomic scales, h is very small; the length
of the scales over which quantum effects are important being dependent
on it. This fact provides us with a way of transition from classical to
quantum mechanics, by comparing the classical limit when h = 0, where
all quantum effects are neglected, with the original quantum system through
the semiclassical regime, where h is assumed to be arbitrarily small.
When V ∈ C ∞ is a non-negative potential bounded away from zero at
infinity and having a finite number of non-degenerate zeros, it is a rather
classical result that the fundamental energy of −h2 ∆ + V (x), subsequently

denoted by E(h), satisfies
E(h) := σ[−h2 ∆ + V, D, Ω] = E1 h + O(h2 ) as h ↓ 0, (9.90)
where E1 stands for the fundamental energy of the associated harmonic
oscillator localized at the wells of V (x). Moreover, the associated prin-
cipal eigenfunctions, or ground states, concentrate in V −1 (0), and either
there is a rapid eigenvalue degeneracy, or the limiting ground states reside
asymptotically in a single well (see Theorem 1.2 of B. Simon [204]).
It was the intention of B. Simon on p. 22 of [203] to study the case
when V −1 (0) is a manifold, instead of a discrete set, but he only observed
that if some of the zeros of V (x) are degenerate, then, the fundamental
energy E(h) might decay to zero faster than linearly, and that, under the
appropriate hypothesis on the shape of V (x) at the degenerate minima,
it should be possible to get lower bounds for that decay. Theorem 9.6
provides us with an extremely satisfactory answer to these open questions,
as it reveals that
lim h−2 σ[−h2 ∆ + V, D, Ω] = σ[−∆, D, Ω0 ] (9.91)
h↓0
if Ω0 := int V −1 (0) is a nice open subdomain of Ω and, therefore,

E(h) = σ[−∆, D, Ω0 ] h2 + o(h2 ) as h ↓ 0. (9.92)
Consequently, the fundamental energy decays quadratically in these de-
generate situations. The interested reader is sent to E. N. Dancer and J.
López-Gómez [47] for further details.
In the context of population dynamics, Theorem 9.6 has tremendously
facilitated some recent studies about the range of validity of the principle of
competitive exclusion. Actually, Theorem 9.6 has revealed that, in the pres-
ence of refuge areas, the most paradigmatic spatial competing species mod-
els predict permanence even when the level of the aggressions between the
antagonist species grows up. Basically, because in the presence of refuges
the species can segregate into them to avoid the aggressions from com-
petitors. Naturally, as soon as the refuges can maintain the corresponding
species in isolation (see J. López-Gómez [136], J. López-Gómez and J. C.
Sabina [155], S. Cano-Casanova and J. López-Gómez [40], and J. López-
Gómez and M. Molina-Meyer [150]). Consequently, in spatial models, the
principle of competitive exclusion fails in the presence of refuges. To prove
this result, we take as a model for competing species the next spatially het-
erogeneous evolutionary model of Lotka–Volterra type, which incorporates
diffusion and transport effects,



 ∂t u+L1 u = λu−A(x)u2 −b B(x)uv x ∈ Ω, t > 0,

∂t v+L2 v = µv−c C(x)uv −D(x)v 2 x ∈ Ω, t > 0,
(9.93)

 B u(x, t) = B2 v(x, t) = 0 x ∈ ∂Ω, t > 0,
 1
(u(x, 0), v(x, 0)) = (u0 (x), v0 (x)) x ∈ Ω,
where:
i) Ω is a bounded domain of RN , N ≥ 1, of class C 2 .
ii) λ, µ ∈ R, b, c > 0, and L1 , L2 are second order uniformly elliptic
operators in Ω
Lj := − div (Aj ∇ · ) + ⟨bj , ∇ · ⟩ + cj , j ∈ {1, 2}.
iii) For each j ∈ {1, 2}, Bj stands for the boundary operator
{
D on Γj0
Bj :=
∂ν j + β j on Γj1
where Γj0 and Γj1 are two disjoint open and closed subsets of ∂Ω with
Γj0 ∪ Γj1 = ∂Ω, βj ∈ C(Γj1 ), and νj = Aj n is the co-normal vector-field.
iv) A, B, C, D ∈ C(Ω̄) are non-negative functions such that
A(x) > 0 and D(x) > 0 for all x ∈ Ω̄,
B, C ∈ A, with respect to the partitions of ∂Ω induced by Γj0 and Γj1 ,
j ∈ {1, 2}, respectively, and
+ ̸= ∅,
int ΩB + ̸= ∅,
int ΩC
where ΩB C
+ and Ω+ stand for the respective regions Ω+ associated to
B(x) and C(x) through Definition 9.1(a).
In population dynamics, (9.93) models the evolution of two competing
species u and v in the inhabiting area Ω when the individuals disperse
randomly within Ω according to the patterns
Lj := − div (Aj ∇ · ) + ⟨bj , ∇ · ⟩, j ∈ {1, 2},
where the bj ’s represent the transport effects of each of the species. Typi-
cally, for every x ∈ Ω and t > 0, u(x, t) and v(x, t) measure the densities of
the populations at the point x ∈ Ω at time t > 0, and λ−c1 (x) and µ−c2 (x)
measure the intrinsic growth (or death) rates of u and v, respectively, while
the function coefficients A(x) and D(x) stand for the normalized carrying
capacities of the species, and B(x) and C(x) fix the nature of the compe-
tition between u and v. Precisely, the region where B > 0 (resp. C > 0)
provides us with the patches of the territory where u (resp. v) receives

aggressions from v (resp. u), while the region where B = 0 (resp. C = 0),
denoted by ΩB C
0 (resp. Ω0 ), is the spatial refuge of the species u (resp. v).
In model (9.93), the parameters b > 0 and c > 0 measure the intensity of
the aggressions between u and v, and u0 and v0 stand for the initial pop-
ulation densities. Actually, in the most classical non-spatial models where
L1 = L2 = 0 and
A = B = C = D = 1,
the competition is considered to be of low intensity if bc < 1, whereas it
is of high intensity if bc > 1. The book of R. S. Cantrell and C. Cosner
[41] is an excellent monograph on spatial ecology through reaction diffusion
systems.
For every initial data u0 , v0 ∈ C(Ω̄), (9.93) has a unique global solution
(u(x, t; u0 , v0 ), v(x, t; u0 , v0 )) (see, e.g., H. Amann [10]), and, due to the
parabolic maximum principle,
0 ≤ u(·, t; u0 , v0 ) ≤ T1 (t)u0 and 0 ≤ v(·, t; u0 , v0 ) ≤ T2 (t)v0
for all t > 0, where
T1 (t) = et(λ−L1 ) , T2 (t) = et(µ−L2 ) .
According to the pioneering results of M. W. Hirsch [100] and H. Matano
[158], and the synthesis of H. L. Smith and H. R. Thieme [208], the limiting
profiles as t ↑ ∞ of the positive solutions of (9.93) are, generically, non-
negative steady-states (see the monograph of H. L. Smith [207]), which are
the non-negative solutions of

 L1 u = λu − A(x)u − b B(x)uv
2
in Ω,
L2 v = µv − c C(x)uv − D(x)v 2 in Ω, (9.94)

B1 u = B2 v = 0 on ∂Ω.
Besides (0, 0), the problem (9.94) admits three types of component-wise
non-negative solutions. Namely, the solutions having one component van-
ishing, (u, 0) or (0, v), known as the semi-trivial positive solutions, and the
solutions having both component positive, known as the coexistence states.
According to Theorem 2.14 of [40], (9.94) possesses a semi-trivial posi-
tive solution of the form (u, 0) if, and only if,
λ > σ[L1 , B1 , Ω]
and, in such case, (Uλ , 0) is the unique one, where Uλ is the unique positive
solution of
{
L1 u = λu − A(x)u2 in Ω,
B1 u = 0 on ∂Ω.
Similarly, (9.94) possesses a semi-trivial positive solution of the form (0, v)

if, and only if,
µ > σ[L2 , B2 , Ω]
and, in such case, (0, Vµ ) is the unique one, where Vµ is the unique positive
solution of {
L2 v = µv − D(x)v 2 in Ω,
B2 v = 0 on ∂Ω.
Moreover, according to Proposition 5.1 [40], for every λ > σ[L1 , B1 , Ω],
(Uλ , 0) is linearly unstable if and only if
µ > σ[L2 + cC(x)Uλ , B2 , Ω],
and, for any µ > σ[L2 , B2 , Ω], (0, Vµ ) is linearly unstable if and only if
λ > σ[L1 + bB(x)Vµ , B1 , Ω].
Therefore, under the estimates
{
µ > σ[L2 + cC(x)Uλ , B2 , Ω],
(9.95)
λ > σ[L1 + bB(x)Vµ , B1 , Ω],
both semi-trivial positive solutions, (Uλ , 0) and (0, Vµ ), are linearly unsta-
ble. It turns out that, in such case, the model is permament in the sense
that no species is driven to extinction by the other (see, e.g., Theorem 5.3
of [40]). Moreover, (9.94) has a coexistence state.
On the other hand, according to Theorem 9.6, we have that
lim σ[L2 + cC(x)Uλ , B2 , Ω] = σ[L2 , B2 [ΩC C
0 ], Ω0 ],
c↑∞
lim σ[L1 + bB(x)Vµ , B1 , Ω] = σ[L1 , B1 [ΩB B

0 ], Ω0 ],
b↑∞
and, therefore, the estimates
µ > σ[L2 , B2 [ΩC C
0 ], Ω0 ], λ > σ[L1 , B1 [ΩB B
0 ], Ω0 ], (9.96)
guarantee the permanence of u and v for all b > 0 and c > 0. Consequently,
under condition (9.96), the species u and v are permanent independently
of the intensity of their mutual aggressions; measured by b and c.
As conditions (9.96) utterly mean that the refuge areas of u and v, ΩB 0
and ΩC0 , respectively, can maintain the corresponding species in the absence
of antagonists, Theorem 9.6 establishes how the principle of competitive
exclusion fails in spatial models in the presence of refuges for each of the
competitors.
The previous ideas are on the foundations of the mathematical analysis
of the effects of facilitation in competitive environments, as well as in the
analysis of the effects of strategic alliances in management (see J. López-
Gómez and M. Molina-Meyer [149, 151, 152]).

Bibliography
[1] Adams, R. A. (1975). Sobolev Spaces (Academic Press, New York).

[2] Adams, D. R. and Hedberg, L. I. (1996). Function Spaces and Potential
Theory, Grundlehren der Mathematischen Wissenschaften 314 (Springer,
Berlin).
[3] Agmon, S. (1983). On positivity and decay of solutions of second order
elliptic equations on Riemanian manifolds, in D. Greco (ed.) Methods of
Functional Analysis and Theory of Elliptic Equations (Liguori, Napples),
pp. 187–204.
[4] Agmon, S., Douglis, A. and Nirenberg, L. (1959). Estimates near the bound-
ary for solutions of elliptic partial differential equations satisfying general
boundary value conditions I, Comm. Pure Appl. Math. 12, pp. 623–727.
[5] Alexandroff, A. D. (1962). A characterization property of the spheres, Ann.
Mat. Pura Appl. 58, pp. 303–354.
[6] Alikakos, N. D. and Fusco, G. (1991). A dynamical systems proof of the
Krein–Rutman theorem and an extension of the Perron theorem, Proc. Roy.
Soc. Edinburgh 117A, pp. 209–214.
[7] Amann, H. (1972). On the number of solutions of nonlinear equations in
ordered Banach spaces, J. Funct. Anal. 11, pp. 346–384.
[8] Amann, H. (1976). Fixed point equations and nonlinear eigenvalue problems
in ordered Banach spaces, SIAM Rev. 18, pp. 620–709.
[9] Amann, H. (1983). Dual semigroups and second order linear elliptic bound-
ary value problems, Israel J. Math. 45, pp. 225–254.
[10] Amann, H. (1995). Linear and Quasilinear Parabolic Problems, Vol. I: Ab-
stract Linear Theory (Birkhäuser, Bassel).
[11] Amann, H. (2005). Maximum principles and principal eigenvalues, in J. Fer-
rera, J. López-Gómez and F. R. Ruiz del Portal (eds.) 10 Mathematical
Essays on Approximation in Analysis and Topology (Elsevier, Amsterdam),
pp. 1–60.
[12] Amann, H. (In preparation). Linear and Quasilinear Parabolic Problems,
Volume II: Function Spaces and Linear Differential Operators.
[13] Amann, H. and López-Gómez, J. (1998). A priori bounds and multiple so-
lutions for superlinear indefinite elliptic problems, J. Diff. Eqns. 146, pp.
319
336–374.
[14] Aronszajn, N. (1965). Potentiels Besseliens, Ann. Inst. Fourier (Grenoble)
15, pp. 43–58.
[15] Aronszajn, N. and Smith, K. T. (1961). Theory of Bessel potentials I, Ann.
Inst. Fourier (Grenoble) 11, pp. 385–475.
[16] Aronszajn, N., Mulla, F. and Szeptycki, P. (1963). On spaces of potentials
connected with Lp -classes, Ann. Inst. Fourier (Grenoble) 13, pp. 211–306.
[17] Arzela, C. (1889). Funzioni di linee, Atti R. Accad. Lincei Rend. 4, pp.
342–348.
[18] Ascoli, G. (1883). Le curve limiti di una varietà data di curve, Mem. R.
Accad. Lincei 18, pp. 521–586.
[19] Babus̆ka, I. (1961). Stability of the domain with respect to the fundamental
problem in the theory of partial differential equations, mainly in connection
with the theory of elasticity, I, II, Czechoslovak Math. J. 11, pp. 76–105,
165–203.
[20] Babus̆ka, I. and Vyborny, R. (1965). Continuous dependence of eigenvalues
on the domain, Czechoslovak Math. J. 15, pp. 169–178.
[21] Bakelman, I. J. (1994). In D. Taliaferro (ed.) Convex Analysis and Nonlinear
Geometric Elliptic Equations. With an obituary for the author by William
Rundell (Springer, New York).
[22] Banach, S. (1932). Théorie des opérations linéaires (Monografje Matematy-
czne, Warsaw).
[23] Barta, J. (1937). Sur la vibration fondamentale d’une membrane, C. R.
Acad. Sci. Paris 204, pp. 472–473.
[24] Beauzamy, B. (1985). Introduction to Banach Spaces and their Geometry,
2nd edn. North-Holland Mathematic Studies 68 (North-Holland, Amster-
dam).
[25] Beltramo, A. and Hess, P. (1984). On the principal eigenvalue of a periodic-
parabolic operator, Comm. Part. Diff. Eqns. 9, pp. 919–941.
[26] Berestycki, H. and Nirenberg, L. (1991). On the method of moving planes
and the sliding method, Boll. Soc. Brasil Mat. (Nova Ser.) 22, pp. 1–37.
[27] Berestycki, H., Nirenberg, L. and Varadhan, S. R. S. (1994). The princi-
pal eigenvalue and maximum principle for second order elliptic operators in
general domains, Comm. Pure Appl. Math. XLVII, 1, pp. 47–92.
[28] Bony, J. M. (1967). Principe du maximum dans les espaces de Sobolev, C.
R. Acad. Sci. Paris 265, pp. 333–336.
[29] Brézis, H. (1983). Analyse Fontionnelle (Masson, Paris).
[30] Browder, F. E. (1960). A priori estimates for solutions of elliptic boundary
value problems I, Neder. Akad. Wetensch. Indag. Math. 22, pp. 149–159.
value problems II, Neder. Akad. Wetensch. Indag. Math. 22, pp. 160–169.
value problems III, Neder. Akad. Wetensch. Indag. Math. 23, pp. 404–410.
[33] Brown, K. J. and Lin, C. C. (1980). On the existence of positive eigenfunc-
tions for an eigenvalue problem with indefinite weight function, J. Math.
Anal. Appl. 75, pp. 112–120.
Bibliography 321
[34] Butzer, P. L. and Berens, H. (1967). Semi-Groups of Operators and Approx-

imation (Springer, Berlin).
[35] Calderón, A. P. (1961). Lebesgue spaces of differentiable functions and dis-
tributions, in Partial Differential Equations, Proc. Sympos. Pure Math. 4,
(Amer. Math. Soc., Providence), pp. 33–49.
[36] Calderón, A. P. and Zygmund, A. (1952). On the existence of certain singular
integrals, Acta Math. 88, pp. 85–139.
[37] Calderón, A. P. and Zygmund, A. (1961). Local properties of solutions of
elliptic partial differential equations, Studia Math. 20, pp. 171–225.
[38] Cano-Casanova, S. (1998). Caracterización del principio del máximo para
problemas elı́pticos con condiciones de contorno generales, Master’s thesis,
Complutense University of Madrid, Spain.
[39] Cano-Casanova, S. and López-Gómez, J. (2002). Properties of the principal
eigenvalues of a general class of non-classical mixed boundary value prob-
lems, J. Diff. Eqns. 178, pp. 123–211.
[40] Cano-Casanova, S. and López-Gómez, J. (2003). Permanence under strong
aggressions is possible, Ann. Inst. H. Poincaré Anal. Non Linéaire 20, 6,
pp. 999–1041.
[41] Cantrell, R. S. and Cosner, C. (2003). Spatial Ecology via Reaction-Diffusion
Equations, Wiley Series in Mathematical and Computational Biology (John
Wiley and Sons, Chichester).
[42] Clarkson, J. A. (1936). Uniformly convex spaces, Trans. Amer. Math. Soc.
40, pp. 396–414.
[43] Clément, Ph. and Peletier, L. A. (1979). An anti-maximum principle for
second-order elliptic operators, J. Diff. Eqns. 34, pp. 218–229.
[44] Courant, R. and Hilbert, D. (1962). Methods of Mathematical Physics I-II
(Wiley-Interscience, New York).
[45] Crandall, M. G. and Rabinowitz, P. H. (1971). Bifurcation from simple eigen-
values, J. Funct. Anal. 8, pp. 321–340.
[46] Dancer, E. N. (1990). The effect of domain shape on the number of positive
solutions of certain nonlinear equations II, J. Diff. Eqs. 87, pp. 316–339.
[47] Dancer, E. N. and López-Gómez, J. (2000). Semiclassical analysis of general
second order elliptic operators on bounded domains, Trans. Amer. Math.
Soc. 352, pp. 3723–3742.
[48] Day, M. M. (1941). Reflexive Banach spaces not isomorphic to uniformly
convex spaces, Bull. Amer. Math. Soc. 47, pp. 313–317.
[49] Delgado, M., López-Gómez, J. and Suárez, A. (2000). On the symbiotic
Lotka–Volterra model with diffusion and transport effects, J. Diff. Eqns.
160, pp. 175–262.
[50] Denk, R., Hieber, M. and Prüss, J. (2003), R-boundedness. Fourier multi-
pliers and problems of elliptic and parabolic type, Mem. Amer. Math. Soc.
166, no. 788.
[51] Denk, R., Hieber, M. and Prüss, J. (2007). Optimal Lp − Lq estimates for
parabolic boundary value problems with inhomogeneous data, Math. Z. 257,
pp. 193–224.
[52] Deny, J. and Lions, J. L. (1953). Les espaces de type Beppo Levi, Ann. Inst.
Fourier, Grenoble 5, pp. 305–370.

[53] Diestel, J. (1975). Geometry of Banach Spaces: Selected Topics, Lecture
Notes in Mathematics 485 (Springer, Berlin).
[54] Donsker, M. D. and Varadhan, S. R. S. (1975). On a variational formula for
the principal eigenvalue for operators with maximum principle, Proc. Nat.
Acad. USA 72, pp. 780–783.
[55] Donsker, M. D. and Varadhan, S. R. S. (1976). On the principal eigenvalue
of second order elliptic differential operators, Comm. Pure Appl. Math. 29,
pp. 595–621.
[56] Du, Y. (2006). Order Structure and Topological Methods in Nonlinear Partial
Differential Equations, Series on Partial Differential Equations and Applica-
tions 2 (World Scientific, Singapore).
[57] Duffin, R. J. (1947). Lower bounds for eigenvalues, Phys. Rev. 71, pp. 827–
828.
[58] Earnshaw, S. (1839). On the nature of the molecular forces which regulate
the constitution of the luminiferous ether, Cambridge Phil. Soc. Trans. 7,
pp. 97–112.
[59] Eberlein, W. F. (1947). Weak compactness in Banach spaces, Proc. Nat.
Acad. Sci. USA 38, pp. 51–53.
[60] Evans, L. C. (1998). Partial Differential Equations, Graduate Studies in
Mathematics 19 (Amer. Math. Soc., Providence, RI).
[61] Faber, C. (1923). Beweis das unter allen homogenen Membranen von gleicher
Fläche und gleicher Spannung die kreisdörmige den tiefsten Grundton gibt,
Sitzungsber. Bayer. Akad. der Wiss. Math. Phys. pp. 169–171.
[62] Figueiredo, D. G. and Mitidieri, E. (1990), Maximum principles for cooper-
ative elliptic systems, C. R. Acad. Sci. Paris 310, pp. 49–52.
[63] Figueiredo, D. G. and Mitidieri, E. (1992). Maximum principles for linear
elliptic systems, Rend. Istit. Mat. Univ. Trieste 22, pp. 36–66.
[64] Fleckinger, J. Hernández, J. and Thélin, F. (1993). A maximum principle
for linear cooperative elliptic systems, in Differential Equations with Ap-
plications to Mathematical Physics, Math. Sci. Eng. 192, (Academic Press,
Boston, MA), pp. 79–86.
[65] Fourier, J. B. J. (1888-90). Oeuvres I-II (Gauthier-Villars, Paris).
[66] Fraile, J. M., Koch-Medina, P., López-Gómez, J. and Merino, S. (1996).
Elliptic eigenvalue problems and unbounded continua of positive solutions
of a semilinear elliptic equation, J. Diff. Eqns. 127, pp. 295–319.
[67] Friedman, A. (1969). Partial Differential Equations (Holt, Rinehart and
Winston, New York).
[68] Friedrichs, K. O. (1928). Die Randwert und Eigenwertprobleme aus der The-
orie des elastichen Platten, Math. Ann. 98, pp. 205–247.
[69] Friedrichs, K. O. (1944). The identity of weak and strong extensions of dif-
ferential operators, Trans. Amer. Math. Soc. 55, pp. 131–151.
[70] Friedrichs, K. O. (1953). Differentiability of solutions of elliptic partial dif-
ferential equations, Comm. Pure and Appl. Maths. 5, pp. 299–326.
[71] Frobenius, G. (1908). Über matrizen aus positiven elementen I, S. B. Preuss.
Akad. Wiss. Berlin, pp. 471–476.
Bibliography 323
[72] Frobenius, G. (1909). Über matrizen aus positiven elementen II, S. B.

Preuss. Akad. Wiss. Berlin, pp. 514–518.
[73] Frobenius, G. (1912). Über matrizen aus nicht negativen elementen, S. B.
Preuss. Akad. Wiss. Berlin, pp. 456–477.
[74] Gagliardo, E. (1958). Proprieta di alcune classi di funzioni in piu variabili,
Ric. Mat. 7, pp. 102–137.
[75] Gauss, C. F. (1839). Allgemeine Theorie des Erdmagnetismus, Beobachtun-
gen des magnetischen Vereins im Jahre 1838, (Leipzig).
[76] Garcı́a-Melián, J., Gómez-Reñasco, R., López-Gómez, J. and Sabina, J. C.
(1998). Point-wise growth and uniqueness of positive solutions for a class
of sublinear elliptic problems where bifurcation from infinity occurs, Arch.
Rat. Mech. Anal. 145, pp. 261–289.
◦
[77] Garding, L. (1953). Dirichlet’s problem for linear elliptic partial differential
equations, Math. Scand. 1, pp. 55–72.
[78] Gidas, B. Ni, W. M. and Nirenberg, L. (1979). Symmetry and related prop-
erties via the maximum principle, Comm. Math. Phys. 68, pp. 209–243.
[79] Gilbarg, D. and Trudinger, N. (2001). Elliptic Partial Differential Equations
of Second Order, Classics in Mathematics (Springer, Berlin and Heildelberg).
[80] Giraud, G. (1932). Généralizations des problèmes sur les opérations du type
elliptiques, Bull. des Sciences Math. 56, pp. 248–272, 281–312, 316–352.
[81] Giraud, G. (1933). Problèmes de valeurs á la frontière relatifs á certainnes
données discontinues, Bull. de la Soc. Math. de France 61, pp. 1–54.
[82] Göhberg, I. C., Goldberg, S. and Kaashoek, M. A. (1990). Classes of Linear
Operators, Operator Theory: Advances and Applications 49 (Birkhäuser,
Bassel).
[83] Göhberg, I. C. and Krein, M. G. (1957). The basis properties on defect
numbers, root numbers and indices of linear operators, Usp. Mat. Nauk. 12,
pp. 43–118.
[84] Göhberg, I. C., Lancaster, P. and Rodman, L. (1982). Matrix Polynomials,
Comp. Sci. Appl. Mathematics (Academic Press, New York).
[85] Gómez-Reñasco, R. (1999). The effect of varying coefficients in semilinear
elliptic boundary value problems. From classical solutions to metasolutions,
Ph.D. thesis, University of La Laguna, Tenerife, Spain.
[86] Gómez-Reñasco, R. and López-Gómez, J. (2002). On the existence and nu-
merical computation of classical and non-classical solutions for a family of
elliptic boundary value problems, Nonl. Anal. TMA 48, pp. 567–605.
[87] Gossez, J. P. and Lami-Dozo, E. (1985). On the principal eigenvalue of a
second order linear elliptic problem, Arch. Rat. Mech. Anal. 89, pp. 169–
175.
[88] Greiner, G. (1981). Zur Perron-Frobenius Theorie stark stetiger Halbgrup-
pen, Math. Z. 177, pp. 401–423.
[89] Grosberg, J. and Krein, M. G. (1939). Sur la Décomposition des fonction-
nelles en composants positives, C.R. (Doklady) Acad. Sci. CCCR (N.S.) 25,
pp. 723–726.
[90] Grosswald, E. (1974). Collected Papers of Hans Rademacher, (MIT Press,
Cambridge).
[91] Guzmán, M. and Rubio, B. (1979). Integración: Teorı́a y Técnicas (Alham-
bra, Madrid).
[92] Hale, J. K. (2005). Eigenvalues and perturbed domains, in J. Ferrera, J.
López-Gómez and F. R. Ruiz del Portal (eds.) 10 Mathematical Essays on
Approximation in Analysis and Topology (Elsevier, Amsterdam), pp. 95–
123.
[93] Henry, D. (1981). Geometric Theory of Semi-linear Parabolic Equations,
Lecture Notes in Mathematics 840 (Springer, Berlin).
[94] Hersch, J. (1960). Sur la fréquence fondamentale d’une membrane vibrante:
évaluations par déefaut et principe de maximum, Z. Angew. Math. Phys. 11,
pp. 387–413.
[95] Hess, P. (1986). On the spectrum of elliptic operators with respect to indef-
inite weights, Lin. Alg. and its Appns. 84, pp. 99–109.
[96] Hess, P. (1991). Periodic-Parabolic Boundary Value Problems and Positivity,
Pitman Research Notes in Mathematics 247 (Longman, London).
[97] Hess, P. and Kato, T. (1980). On some linear and nonlinear eigenvalue prob-
lems with an indefinite weight function, Comm. Partial Diff. Eqns. 5, pp.
999—1030.
[98] Hilbert, D. (1900). Über das Dirichletsche Prinzip, Jber. Deutsch. Math.
Verein 8, pp. 184–188.
[99] Hilbert, D. (1909). Wesen und Ziele einer Analysis der unendlich vielen un-
abhängigen variablen, Rend. Circ. Mat. Palermo 27, pp. 59–74.
[100] Hirsch, M. W. (1988). Stability and convergence in strongly monotone dy-
namical systems, J. Reine Angew. Math. 383, pp. 1–53.
[101] Hölder, O. (1889). Über einen Mittelwerthsatz, Göttingen Nachr. pp. 38–
47.
[102] Hooker, W. W. (1960). Lower bounds for the first eigenvalue of elliptic
equations of order two and four, Tech. Rep. Univ. California at Berkeley
10, AF49(638) p. 398.
[103] Hopf, E. (1927). Elementare Bemerkungen über die Lösungen partieller Dif-
ferentialgleichungen zweiter Ordnung vom elliptischen Typus, Sitzungsber.
Preuss. Akad. Wiss. 19, pp. 147–152.
[104] Hopf, E. (1952). A remark on linear elliptic differential equations of the
second order, Proc. of the Amer. Math. Soc. 3, pp. 791–793.
[105] Hörmander, L. (1955). On the theory of general partial differential opera-
tors, Acta Math. 94, pp. 161–248.
[106] Hörmander, L. (1983). The Analysis of Linear Partial Differential Opera-
tors I-IV (Springer, Berlin and New York).
[107] Jameson, G. (1970). Ordered Linear Spaces, Lectures Notes in Mathematics
141 (Springer, New York).
[108] John, F. (1978). Partial Differential Equations, Applied Mathematical Sci-
ences 1 (Springer, New York).
[109] Jordan, P. and von Neumann, J. (1935). On inner products in linear, metric
spaces, Ann. of Math. 36, pp. 719–723.
[110] Kakutani, S. (1939). Weak topologies and regularity of Banach spaces,
Bibliography 325
Proc. Imp. Acad. Tokyo 15, pp. 169–173.

[111] Kato, T. (1982). Superconvexity of the spectral radius and convexity of the
spectral bound and type, Math. Z. 180, pp. 265–273.
[112] Kato, T. (1995). Perturbation Theory for Linear Operators, Classics in
Mathematics (Springer, Berlin).
[113] Kinderlehrer, D. and Stampacchia, G. (1980). An Introduction to Varia-
tional Inequalities and their Applications (Academic Press, London).
[114] Klee, V. L. (1955). Boundness and continuity of linear functionals, Duke
Math. J. 22, pp. 263–270.
[115] Kondrachov, V. I. (1945). On some properties of functions from the space
Lp , Dokl. Akad. Nauk SSSR 48, pp. 533–538.
[116] Korn, A. (1909). Über Minimalflächen deren Randkurven wenig von ebenen
Kurven abweichen, Adhandl. Königl. Preuss. Akad. Wiss. 2, (Preuss. Akad.
Press, Berlin).
[117] Krahn, E. (1925). Über eine von Rayleigh formulierte Minimaleigenschaft
des Kreises, Math. Ann. 91, pp. 97–100.
[118] Krasnoselskij, M. A. (1964). Positive Solutions of Operator Equations (No-
ordhoff, Groninggen).
[119] Krein, M. G. (1940). Propriétés fondamentales des ensembles coniques nor-
maux dans l’espace de Banach, C. R. (Doklady) Acad. Sci. CCCR (N.S.)
28, pp. 13–17.
[120] Krein, M. G. and Rutman. M. A. (1948). Linear operators leaving invariant
a cone in a Banach space (in Russian), Usp. Mat. Nauk. (N.S.) 3, pp. 3–95.
[121] Kresin, G. and Maz’ja, V. G. (2013). Maximum Principles and Sharp Con-
stants for Solutions of Elliptic and Parabolic Systems (Amer. Math. Soc.,
Providence).
[122] Lax, P. D. and Milgram, A. N. (1954). Contributions to the theory of partial
differential equations, Ann. Maths. Studies 33, pp. 167–190.
[123] Lebesgue, H. (1907). Sur le problème de Dirichlet, Rend. Circ. Mat.
Palermo 24, pp. 371–404.
[124] Lebesgue, H. (1910). Sur l’integration des fonctions discontinues, Ann. Ec.
Norm. Sup. XXVII, pp. 361–450.
[125] Lichtenstein, L. (1912). Beiträge zur Theorie der linearen partiellen Dif-
ferentialgleichungen zweiter Ordnung vom elliptischen Typus, Rend. Circ.
Mat. Palermo 33, pp. 201–211.
[126] Lichtenstein, L. (1924). Neue Beiträge zur Theorie der linearen partiellen
Differentialgleichungen zweiter Ordnung vom elliptischen Typus, Math. Z.
20, pp. 194–212.
[127] Lindenstrauss, J. and Tzafriri, L. (1971). On complemented subspace prob-
lem, Israel J. Math. 9, pp. 263–269.
[128] Lindenstrauss, J. and Tzafriri, L. (1973). Classical Banach Spaces, Lectures
Notes in Mathematics 338 (Springer, Berlin).
[129] Lions, J. L. (1963). Théorèmes de traces et d’interpolation IV, Math. Ann.
151, pp. 42–56.
[130] Lions, J. L. and Magenes, E. (1968). Problèmes aux Limites non Homogènes
I-III (Dunod, Paris).
[131] Lions, J. L. and Stampacchia, G. (1967). Variational inequalities, Comm.

Pure Appl. Maths. 20, pp 493–519.
[132] Lions, P. L. (1982). A remark on Bony maximum principle, Proc. Amer.
Math. Soc. 88, pp. 503–508.
[133] Lipschitz, R. (1876). Sur la possibilité d’intégrer complétement un système
donné d’équations différentielles, Bull. Sci. Math. Astro. 10, pp. 149–159.
[134] López-Gómez, J. (1981). Problemas de contorno y valor inicial para ecua-
ciones parabólicas lineales y semilineales, Master’s thesis, Complutense Uni-
versity of Madrid, Spain.
[135] López-Gómez, J. (1994). On linear weighted boundary value problems, in
Partial Differential Equations. Models in Physics and Biology, Mathematical
Research 82 (Akademie-Verlag, Berlin), pp. 188–203.
[136] López-Gómez, J. (1995). Permanence under strong competition, Dynamical
systems and applications, World Sci. Ser. Appl. Anal. 4, pp. 473–488.
[137] López-Gómez, J. (1996). The maximum principle and the existence of prin-
cipal eigenvalues for some linear weighted boundary value problems, J. Diff.
Eqns. 127, pp. 263–294.
[138] López-Gómez, J. (2000). Large solutions, metasolutions, and asymptotic
behaviour of the regular positive solutions of sublinear parabolic problems,
El. J. Diff. Eqs. Conf. 05, pp. 135–171.
[139] López-Gómez, J. (2001). Approaching metasolutions by classical solutions,
Diff. Int. Eqs. 14, pp. 739–750.
[140] López-Gómez, J. (2001). Spectral Theory and Nonlinear Functional Anal-
ysis, Research Notes in Mathematics Vol. 426 (Chapman and Hall/CRC
Press, Boca Raton).
[141] López-Gómez, J. (2001). Ecuaciones Diferenciales y Variable Compleja
(Prentice-Hall, Madrid).
[142] López-Gómez, J. (2002). Ecuaciones Diferenciales y Variable Compleja,
Problemas y Ejercicios Resueltos (Prentice-Práctica, Madrid).
[143] López-Gómez, J. (2003). Coexistence and metacoexistence states in com-
peting species models, Houston J. Math. 29, pp. 485–538.
[144] López-Gómez, J. (2003). Classifying smooth supersolutions for a general
class of elliptic boundary value problems, Adv. Diff. Eqns. 8, pp. 1025–1042.
[145] López-Gómez, J. (2005). Metasolutions: Malthus versus Verhulst in popu-
lation dynamics. A dream of Volterra, in M. Chipot and P. Quittner (eds.)
Stationary Partial Differential Equations II (Elsevier, Amsterdam), pp. 211–
309.
[146] López-Gómez, J. (2009). The strong maximum principle, RIMS Kôkyûroku
Bessatsu B15, pp. 113–123.
[147] López-Gómez, J. (2011). The existence of weak solutions for a general class
of mixed boundary value problems, Disc. and Cont. Dyn. Syst. Supp. 2011,
pp. 1015–1024.
[148] López-Gómez, J. and Molina-Meyer, M. (1994). The maximum principle for
cooperative weakly elliptic systems and some applications, Diff. Int. Eqs. 7,
pp. 383–398.
[149] López-Gómez, J. and Molina-Meyer, M. (2004). Singular perturbations in
Bibliography 327
Economy and Ecology. The effect of strategic symbiosis in random compet-

itive environments, Adv. Math. Sci. Appns. 14, pp. 87–107.
[150] López-Gómez, J. and Molina-Meyer, M. (2006). The competitive exclusion
principle versus biodiversity through segregation and further adaptation to
spatial heterogeneities, Theor. Population Biol. 69, pp. 94–109.
[151] López-Gómez, J. and Molina-Meyer, M. (2006). Superlinear indefinite sys-
tems: Beyond Lotka-Volterra models, J. Diff. Eqns. 221, pp. 343–411.
[152] López-Gómez, J. and Molina-Meyer, M. (2007). Biodiversity through co-
opetition, Discrete Contin. Dyn. Syst. Ser. B 8, pp. 187–205.
[153] López-Gómez, J. and Mora-Corral, C. (2007). Algebraic Multiplicities of
Eigenvalues for Linear Operators, Operator Theory: Advances and Appli-
cations Vol. 177 (Birkhäuser, Bassel).
[154] López-Gómez, J. and Pardo, R. M. (1993). The existence and the unique-
ness for the predator-prey model with diffusion, Diff. Int. Eqs. 6, pp. 1025–
1031.
[155] López-Gómez, J. and Sabina, J. C. (1995). Coexistence states and global
attractivity for some convective diffusive competing species models, Trans.
Amer. Math. Soc. 347, 3797–3833.
[156] Malliavin, P. (1982). Intégration et Probabilitiés, Analyse de Fourier et
Analyse Spectrale (Masson, Paris).
[157] Manes, A. and Micheletti, A. M. (1973). Un’estensiones della teoria vari-
azionale classica degli autovalori per operatori ellittici del secondo ordine,
Boll. Un. Mat. Ital. 7, pp. 285–301.
[158] Matano, H. (1984). Existence of nontrivial unstable sets for equilibriums of
strongly order-preserving systems, J. Fac. Sci. Univ. Tokyo Sect. IA Math.
30, pp. 645–673.
[159] Maz’ja, V. G. (1985). Sobolev Spaces (Springer, Berlin).
[160] McNabb, A. (1961). Strong comparison theorems for elliptic equations of
second order, J. Math. Mech. 10, pp. 431–440.
[161] Meyers, N. G. and Serrin, J. (1964). H = W , Proc. Nat. Acad. Sci. USA
51, pp. 1055–1056.
[162] Milman, D. P. (1938). On some criteria for the regularity of spaces of type
(B), Doklady Akad. Nauk. SSSR 20, pp. 234.
[163] Molina-Meyer, M. (1995). Existence and uniqueness of coexistence states
for some nonlinear elliptic systems, Nonl. Anal. T.M.A. 25, pp. 279–296.
[164] Molina-Meyer, M. (1996). Global attractivity and singular perturbation for
a class of nonlinear cooperative systems, J. Diff. Eqns. 128, pp. 347–378.
[165] Molina-Meyer, M. (1997). Uniqueness and existence of positive solutions
for weakly coupled general sublinear systems, Nonl. Anal. T.M.A. 30, pp.
5375–5380.
[166] Morrey, C. B. (1938). On the solutions of quasi-linear elliptic partial dif-
ferential equations, Trans. Amer. Math. Soc. 43, pp. 126–166.
[167] Morrey, C. B. (1966). Multiple Integrals in the Calculus of Variations
(Springer, Berlin).
[168] Motzkin, Th. (1935). Sur quelques propriétés charactéristiques des ensem-
bles convexes, Atti Acad. Naz. Lincei Rend. 6, pp. 562–567.
[169] Moutard, T. (1894). Notes sur les équations aux dérivées partielles, J. de
l’École Polytechnique 64, pp. 35–69.
[170] Nec̆as, J. (1967). Les méthodes directes en théorie des équations elliptiques
(Éditeurs Academia, Prague).
[171] von Neumann, J. (1929). Allgemeine Eigenwerttheorie Hermitescher Funk-
tionaloperatoren, Math. Ann. 102, pp. 49–131.
[172] Oleinik, O. A. (1952). On properties of some boundary problems for equa-
tions of elliptic type, Math. Sbornik, N. S. 30, pp. 695–702.
[173] Pagter, B. (1986). Irreducible compact operators, Math. Z. 192, pp. 149–
153.
[174] Paraf, A. (1892). Sur le problème de Dirichlet at son extension au cas de
l’équation linéaire générale du second ordre, Ann. Fac. Sci. Toulouse Ser. I,
6, pp. 1–75.
[175] Perron, O. (1907). Zur Theorie der matrices, Math. Ann. 64, pp. 248–263.
[176] Pettis, B. J. (1939). A proof that every uniformly convex space is reflexive,
Duke Math. J. 5, pp. 249–253.
[177] Phillips, R. S. (1955). The adjoint semi-group, Pacific J. Math. 5, pp. 269–
283.
[178] Picard, E. (1905). Traité d’Analyse Vol. 2 (Gauthier Villars, Paris).
[179] Picone, M. (1927). Maggiorazione degli integrali di equazioni lineari ellitico-
paraboliche alle derivate parziali del secondo ordine, Atti Accad. Naz. dei
Lincei 5, pp. 138–143.
[180] Poincaré, H. (1916-1956). Oeuvres Vol. 1-11 (Gauthier-Villars, Paris).
[181] Protter, M. H. (1960). Lower bounds for the first eigenvalue of elliptic
equations, Ann. of Math. 71, pp. 423–444.
[182] Protter, M. H. and Weinberger, H. F. (1966). On the spectrum of general
second order operators, Bull. Amer. Math. Soc. 72, pp. 251–255.
[183] Protter, M. H. and Weinberger, H. F. (1967). Maximum Principles in Dif-
ferential Equations (Prentice-Hall, Englewood Cliffs).
[184] Pucci, P. and Serrin, J. (2007). The Maximum Principle, Progress in Non-
linear Differential Equations and Their Applications Vol. 73 (Birkhäuser,
Bassel).
[185] Rabinowitz, P. H. (1975). Théorie de degré topologique et applications à des
problémes aux limites non linéaires, Lecture Notes Lab. Analyse Numerique
(Université Paris VI, Paris).
[186] Rayleigh, J. W. S. (1945). The Theory of Sound, 2nd edn. (Dover Publica-
tions, New York).
[187] Redheffer, R. Personal communication to W. Walter (see p. 295 [224]).
[188] Rellich, F. (1930). Ein satz über mittlere konvergenz, Math. Nachr. 31, pp.
30–35.
[189] Riesz, F. (1910). Untersuchungen über Systeme integrierbarer Funktionen,
Math. Ann. LXIX, pp. 449–497.
[190] Riesz, F. (1918). Über lineare Funktionalgleichungen, Acta Math. 41, pp.
71–98.
[191] Riesz, F. (1934). Zur Theorie des Hilbertschen Raumes, Acta Sci. Math.
Szeged 7, pp. 34–38.
Bibliography 329
[192] Ringrose, J. R. (1959). A note on uniformly convex spaces, J. London Math.

Soc. 34, pp. 92.
[193] Rogers, L. J. (1888). An extension of a certain theorem in inequalities,
Messenger of Math. 17, pp. 145–150.
[194] Schaefer, H. H. (1967). Topological Vector Spaces, Macmillan Series in Ad-
vanced Mathematics and Theoretical Physics (Macmillan, New York).
[195] Schauder, J. (1934). Über lineare elliptische Differentialgleichungen zweiter
Ordnung, Math. Z. 38, 257–282.
[196] Schauder, J. (1935). Numerische Abschätzungen in elliptischen linearen
Differentialgleichungen, Studia Math. 5, pp. 34–42.
[197] Schlag, W. Personal communication to L. C. Evans (see [60]).
[198] Schmidt, E. (1907). Zur Theorie der linearen und nichtlinearen Integralgle-
ichungen, Math. Ann. LXIII, pp. 433–476.
[199] Schwartz, L. (1950–51). Théorie des Distributions Vols. I-II (Hermann,
Paris).
[200] Schwartz, L. (1981). Geometry and Probability in Banach Spaces, Based
on notes taken by P. R. Chernoff, Lecture Notes in Mathematics Vol. 852
(Springer, Berlin and New York).
[201] Serrin, J. (1971). A symmetry problem in potential theory, Arch. Rat. Mech.
Anal. 43, pp. 304–318.
[202] Shmulyan, V. L. (1940). Über lineare topologische Räume, Math. Sbornik,
N. S. 7, 49, pp. 425–448.
[203] Simon, B. (1983). Semiclassical analysis of low lying eigenvalues I: Nonde-
generate minima: asymptotic expansions, Ann. Inst. H. Poincaré A 38, pp.
12–37.
[204] Simon, B. (1984). Semiclassical analysis for low lying eigenvalues II: Tun-
neling, Ann. of Math. 120, pp. 89–118.
[205] Slobodeckii, L. N. (1958). Generalized Sobolev spaces and their application
to boundary problems for partial differential equations, Leningrad Gos. Ped.
Inst. Ucen. Zap. 197, pp. 54–112.
[206] Slobodeckii, L. N. (1958). Estimate in Lp of solutions of elliptic systems,
Dokl. Akad. Nauk SSSR 123, pp. 616–619.
[207] Smith, H. L. (1995). Monotone Dynamical Systems. An Introduction to the
Theory of Competitive and Cooperative Systems, Mathematical Surveys and
Monographs Vol. 41. (Amer. Math. Soc., Providence).
[208] Smith, H. L. and Thieme, H. R. (1991). Convergence for strongly order-
preserving semiflows, SIAM J. Math. Anal. 22, 4, pp. 1081–1101.
[209] Sobolev, S. L. (1937). On a boundary value problem for polyharmonic
equations (in russian), Mat. Sbornik, N.S. 2, 44, pp. 465–499.
[210] Sobolev, S. L. (1938). Sur un théorème d’analyse fonctionelle, Math.
Sbornik, 4, 46, pp. 471–497.
[211] Sobolev, S. L. (1945). Certaines Applications de l’Analyse Fonctionelle à
la Physique Mathématique (University of Leningrad, Leningrad).
[212] Stampacchia, G. (1964). Formes bilinéaires coercitives sur les ensembles
convexes, C. R. Acad. Sci. Paris 258, pp. 4413–4416.
[213] Stampacchia, G. (1966). Èquations elliptiques du second ordre à coefficients
discontinus (Les Presses de l’Université de Montréal, Montreal).

[214] Stampacchia, G. (1965). Le problème de Dirichlet pour les équations ellip-
tiques du second ordre à coefficients discontinus, Ann. Inst. Fourier 15, 1,
pp. 189–258.
[215] Stein, E. M. (1970). Singular Integrals and Differentiability Properties of
Functions (Princeton University Press, Princeton).
[216] Suzuki, T. (1994). Semilinear Elliptic Equations (Gakkotosho, Tokyo).
[217] Sweers, G. (1992). Strong positivity in C(Ω̄) for elliptic systems, Math. Z.
209, pp. 251–271.
[218] Takác, P. (1994). A short elementary proof of the Krein–Rutman theorem,
Houston J. of Maths. 20, pp. 93–98.
[219] Takác, P. (1996). An abstract form of maximum and anti-maximum prin-
ciples of Hopf’s type, J. Math. Anal. Appns. 201, pp. 339–364.
[220] Triebel, H. (1978). Interpolation Theory, Function Spaces, Differential Op-
erators (North-Holland, Amsterdam).
[221] Valentine, F. A. (1964). Convex Sets (McGraw-Hill, New York).
[222] Venturino, M. (1978). Primo autovalore di operatore lineari ellittici in forma
non-variazionale, Boll. Un. Math. It. 5, pp. 576–591.
[223] Walter, W. (1970). Differential and Integral Inequalities, Ergebnisse der
Mathematik und ihrer Grenzgebiete Vol. 55 (Springer, Berlin).
[224] Walter, W. (1989). A theorem on elliptic differential inequalities and ap-
plications to gradient bounds, Math. Z. 200, pp. 293–299.
[225] Weyl, H. (1940). The method of orthogonal projection in potential theory,
Duke Math. J. 7, pp. 414–444.
[226] Wloka, J. (1987). Partial Differential Equations (Cambridge University
Press, Cambridge).
[227] Yosida, K. (1995). Functional Analysis, Classics in Mathematics
(Springer,Berlin).
[228] Young, W. H. (1912). On classes of summable functions and their Fourier
series, Proc. Roy. Soc. London Ser. A 87, pp. 225–229.
Index
Adams D. R., 128, 152, 268, 319 potential space, 134

Adams R. A., 126, 149, 152, 319 Bessel F. W., 132–134, 152, 320
adjoint operator, 162 bilinear form
Agmon S., xi, 130, 137, 148, 220, 319 coercive, 75
Alexandroff A. D., ix, 319 continuous, 75
algebraic ascent, 163 Bombal F., 90
algebraic multiplicity, 163 Bony J. M., viii, xi, 60, 61, 188, 189,
Alikakos N. D., 184, 319 221, 320, 326
Amann H., viii, 60, 128, 148–150, minimum principle, 189
160, 184, 185, 188, 215, 221–224, boundary lemma
316, 319 of E. Hopf, 20
Aronszajn N., 152, 320 weak of E. Hopf, 194
Arzela C., 102, 197, 198, 320 Brézis H., x, 80, 81, 89, 128, 152, 159,
Ascoli G., 102, 197, 198, 320 163, 202, 254, 255, 320
Browder F. E., 320
Babus̆ka I., 267, 268, 320 Brown K. J., 311, 312, 320
Bakelman I. J., 320 Butzer P. L., 149, 321
Banach S., x, 2, 64, 78–80, 89, 91, 94,
96–98, 134, 141, 147, 148, 152, 155, Calderón A. P., xi, 127, 130–134, 136,
157, 159–162, 164, 165, 184, 185, 138, 321
195, 198–200, 224, 255, 319–322, Cano-Casanova S., xii, 267–271, 313,
324, 325, 329 314, 321
Banach space, 64 Cantor G., 169
Barta J., 224, 320 Cantrell R. S., 316, 321
Beauzamy B., 320 Cauchy A. L., 63, 67, 117, 197, 232,
Beltramo A., 267, 320 239, 243, 256, 262
Berens H., 149, 321 Cauchy–Schwarz inequality, 63
Berestycki H., viii, ix, 220, 221, 223, characteristic function, 58
267, 268, 271, 320 characterization of the strong
Bessel maximum principle, 216
generalized kernel, 133 Chernoff P. R., 329
kernel, 132 Chipot M., 326
331
Clarkson J. A., 79, 321 Eberlein W. F., and Shmulyan S. L.,

classical solution, 93 theorem, 81
Clement Ph., 222, 321 eigenvalue
coercive bilinear form, 75 dominant, 208
coercive functional, 81 principal, 205
compact operator, 162 strictly dominant, 209
concavity of the principal eigenvalue elliptic
with respect to the potential, 232 operator, 3
cone, 156 uniformly, 3
conormal vector field, 92 ellipticity constant, 3
continuous bilinear form, 75 Euclides, 91
continuous dependence Euler L., 134
of the principal eigenvalue with Evans L. C., x, xi, 39, 88, 126, 127,
respect to the potential, 228 266, 322, 329
convergence of domains, 234
convolution, 131 Faber, C., 264, 271, 322
Ferrera J., 319, 324
Cosner C., 316, 321
Figueiredo D. G., 220, 322
Courant R., xii, 222, 267–269, 311,
Fleckinger J., 220, 322
321
Fourier
Crandall M. G., 223, 279, 281, 283,
inverse transform, 131
321
inversion formula, 131
product formula, 131
Dancer E. N., 268, 314, 321
transform, 131
Day M. M., 80, 321
Fourier J. B. J., 131, 132, 222, 322,
decay property of E. Hopf, 29
327
decreasing, 156
Fraile J. M., 188, 322
Delgado M., 188, 321 Fredholm I., 155, 184, 223
Denk R., 131, 152, 153, 321 Friedman A., 126, 152, 322
Deny J., 127, 321 Friedrichs K. O., 127, 128, 130, 322
Diestel J., 89, 322 Frobenius G., 155, 184, 322, 323
differential operator, 1 function
Dirichlet G. L., viii, x, 39, 41, 130, characteristic, 58
131, 137, 138, 187, 212, 217, harmonic, 6
220–222, 224, 226, 227, 234, 240, subharmonic, 6
260, 265, 267, 268, 271, 286, superharmonic, 6
323–325 functional
dominant eigenvalue, 208 coercive, 81
Donsker M. D., 267, 322 minimizer, 81
Douglis A., xi, 130, 137, 148, 319 fundamental solution, 135
Du Y., 188, 221–223, 322 Fusco G., 184, 319
dual cone, 158
Duffin R. J., 224, 322 Gagliardo E., 127, 323
Garcia-Melian J., 188, 323
Earnshaw S., 38, 322 Garding L., 118, 128, 323
Eberlein W. F., 81, 322 Garding theorem, 118
Index 333
Gauss C. F., 38, 323 Hooker W. W., 224, 324

generalized minimum principle, 30 Hopf E., viii–x, xii, 5, 19, 20, 25, 29,
generating cone, 158 38, 53, 57, 189, 194, 218, 324, 330
Gidas B., ix, 323 boundary lemma, 20
Gilbarg D., 89, 126, 127, 130, 131, decay property, 29
135, 136, 139, 152, 153, 271, 323 minimum principle, 5
Giraud G., 38, 323 uniform decay property, 25
Gohberg I. C., 164, 184, 323 weak boundary lemma, 194
Goldberg S., 164, 323 weak minimum principle, 194
Gomez-Reñasco R., viii, 323 Hormander L., 152, 153, 324
Gossez J. P., 224, 323
Greco D., 319 increasing map, 156
Green interior sphere
first representation formula, 135 property, 12
function, 137 uniform property, 12
second representation formula, 136 uniform property in the strong
Green G., 135–137 sense, 12
Greiner G., 215, 323 invertibility of (L, B, Ω), 120
Grosberg J., 184, 323
Grosswald E., 323 Jameson G., 160, 324
Guzmán M., 247, 324 John F., 135, 152, 324
Jordan P., 88, 324
Hahn H., 159
Hale J. K., 268, 324 Kaashoek M. A., 164, 323
Hedberg L. I., 128, 152, 268, 319 Kakutani S., 80, 324
Henry D., 215, 324 Kato T., xii, 89, 163, 267, 275, 284,
Hernández J., 220, 322 311, 312, 324, 325
Hersch J., 224, 324 Kinderlehrer D., 89, 325
Hess P., xii, 267, 284, 311, 312, 320, Klee V. L., 184, 325
324 Koch-Medina P., 322
Hesse L. O., 36 Kondrachov V. I., 93, 101, 127, 325
Hiebber M., 131, 152, 153, 321 Korn A., 127, 325
Hilbert D., x, xii, 64, 65, 71, 75, 79, Krahn, E., 264, 271, 325
85, 88, 89, 94, 96, 118, 120, 128, Krasnoselskij M. A., 160, 325
160, 222, 223, 267–269, 311, 321, Krein M. G., xi, xii, 155, 161, 164,
324 165, 184, 185, 223, 323, 325
Hilbert space, 64 Kresin G., 325
Hirsch M. W., 316, 324
Holder Lami-Dozo E., 224, 323
inequality, 94 Lancaster P., 164, 323
norm, 98 Laplace operator, 3
spaces, 97 Laplace P. S., 3, 130–132, 135–137,
Holder E., 127 222
Holder O., 93, 94, 97, 98, 110, 117, Laurent development of the resolvent
127, 128, 130, 153, 191, 195, 245, operator, 163
248, 324 Laurent P. A., 163
Lax P. D., x, 64, 78, 93, 128, 325 minimum principle

Lax P. D. and Milgram A. N. classical, 5
theorem of, 78 generalized of M. H. Protter and
Lebesgue H., xi, 1, 94, 102, 110, 116, H. F. Weinberger, 30
126–128, 251, 259, 264, 270, 287, of E. Hopf, 5
325 of J. M. Bony, 189
Levi B., 321 weak of E. Hopf, 194
Lichtenstain E., 38 Minkowski H., 200
Lichtenstein L., 325 Mitidieri E., 220, 322
Lin C. C., 320 Molina-Meyer M., vii–ix, xii, 187,
Lin S. S., 311, 312 188, 220–222, 314, 317, 326, 327
Lindenstrauss J., 80, 85, 89, 325 monotone norm, 160
Lindenstrauss J. and Tzafriri L. monotonicity
theorem, 85 of the principal eigenvalue with
linear ordering, 156 respect to β, 228
Lions J. L., 89, 126–128, 149, 152, of the principal eigenvalue with
268, 321, 325, 326 respect to Ω, 227
Lions P. L., 267, 326 of the principal eigenvalue with
Lipschitz R., 55, 68, 97, 100, 101, respect to the potential, 228
127, 326 Mora-Corral C., 164, 184, 327
Lopez-Gomez J., vii–x, xii, 60, 61, Morrey C. B., 126, 127, 152, 153, 327
128, 150, 164, 184, 187, 188, 215, Motzkin Th., 89, 327
220–222, 224, 267–271, 313, 314, Moutard T., 38, 328
317, 319, 321–324, 326, 327 Mulla F., 152, 320
Lotka A. J., 314
lower semi-continuous functional, 80 Nec̆as J., 126, 149, 152, 328
Neumann J. von, 88, 324, 328
MacNabb A., 327 Neumann K. G., 41, 268, 312
Magenes E., 126, 128, 152, 268, 325 Newton I., 136
Malliavin P., 126, 152, 327 Newtonian potential, 136
Manes A., 311, 312, 327 Ni W. M., ix, 323
Matano H., 316, 327 Nirenberg L., viii, ix, xi, 130, 137,
maximum principle, 216 148, 220, 221, 223, 267, 268, 271,
strong, 215 319, 320, 323
Maz’ja V. G., 126, 152, 325, 327 normal cone, 160
McNabb A., 60
Merino S., 322 Oleinik O. A., 38, 328
Meyers N. G., 127, 327 order interval, 156
Micheletti A. M., 311, 312, 327 ordered
Milgram A. N., x, 64, 78, 93, 128, 325 Banach space, 157
Milman D. P., 80, 327 set, 155
Milman D. P. and Pettis B. J. vector space, 156
theorem, 80 ordering, 155
min-max characterization of the orthogonal
principal eigenvalue, 229 of a closed subspace, 69
minimizer of a functional, 81 projection, 69
Index 335
outward theorem of existence and

pointing vector, 19 uniqueness, 205
unit normal, 18 Pringsheim, 185
product formula, 131
Pagter B., 224, 328 projection
Paraf A., 38, 328 on a closed convex set, 65
parallelogram identity, 63 on a closed convex set of a u.c.
Pardo R. M., 188, 220, 327 B-space, 82
partition of the unity, 59 orthogonal, 69
Peletier L. A., 222, 321 Protter M. H., viii–x, xii, 30, 38, 39,
Perron O., 155, 184, 328 60, 188, 189, 217, 218, 221, 223,
Pettis B. J., 80, 328 224, 267, 328
Phillips R. S., 164, 328 Protter M. H. and Weinberger H. F.
Picard E., 38, 328 generalized minimum principle, 30
Picone M., 38, 328 Prüss J., 131, 152, 153, 321
Planck M., 313 Pucci P., ix, 328
Poincaré H., 222, 328
Poisson Quittner P., 326
kernel, 137
Poisson S. D., 137 Rabinowitz P. H., 222, 223, 279, 281,
positive 283, 321, 328
cone, 156 Rademacher H., 127, 323
operator, 164
Rayleigh J. W. S., 222, 267, 328
vector, 157
Redheffer R., 38, 328
principal
Rellich F., 93, 101, 127, 328
eigenfunction, 205
Rellich F. and Kondrachov V. I.
eigenvalue, 205
compact imbeddings, 101
principal eigenvalue
resolvent operator, 120, 161
concavity with respect to the
potential, 232 resolvent set, 161
continuous dependence with Riemann G. F. B., 319
respect to β, 255 Riesz F., 64, 71, 78, 88, 89, 126, 162,
continuous dependence with 328
respect to the domain, 253 representation theorem, 71
continuous dependence with Riesz M., 133
respect to the potential, 228 Riesz transform, 133
dominance, 208 Ringrose J. R., 80, 329
min-max characterization, 229 Rodman L., 164, 323
monotonicity with respect to β, 228 Rogers L. J., 127, 329
monotonicity with respect to Ω, Rubio B., 247, 324
227 Ruiz del Portal F. R., 319, 324
monotonicity with respect to the Rutman M. A., xi, xii, 155, 161, 164,
potential, 228 165, 184, 185, 223, 325
of a linear weighted boundary
value problem, 273 Sabina J. C., 314, 323, 327
strict dominance, 209 scalar product, 63
Schaefer H. H., 160, 184, 185, 224, dominant eigenvalue, 209

329 increasing map, 156
Schauder J., 127, 130, 162, 329 positive operator, 164
Schlag W., 128, 329 strong maximum principle, 215
Schmidt E., 89, 329 strongly
Schrödinger E., 313 negative operator, 164
Schwartz L., 89, 127, 329 positive operator, 164
Schwarz H.A., 63, 117, 232 Suárez A., 321
Serrin J., ix, 60, 127, 327–329 subsolution, 42
Shmulyan V. L., 81, 329 strict, 42
Simon B., 314, 329 supersolution, 42, 215
simple eigenvalue, 163, 208 first classification theorem, 43
Slobodeckii second classification theorem, 53
norm, 134 strict, 42, 187, 215
space, 134 Suzuki T., ix, 330
Slobodeckii L. N., 134, 135, 148, 152, Sweers G., 220, 330
329 Szeptycki P., 152, 320
Smith H. L., 316, 329
Smith K. T., 152, 320 Takác P., xi, 184, 185, 222, 330
Sobolev Taliaferro D., 320
approximation by smooth Thélin F., 220, 322
functions, 97 theorem
imbeddings, 99 generalized of M. G. Krein and M.
spaces, 95 A. Rutman, 165
Sobolev S. L., xi, 61, 91, 93–95, 98, of D. P. Milman and B. J. Pettis,
99, 101, 126–130, 134, 135, 152, 80
327, 329 of W. F. Eberlein and V. L.
Sobolev–Slobodeckii Shmulyan, 81
norm, 134 boundary lemma of E. Hopf, 20
space, 134 classical minimum principle, 5
solution continuous dependence of the
classical, 93 principal eigenvalue with
weak, 115 respect to β, 255
spectral radius, 161 continuous dependence of the
spectrum, 161 principal eigenvalue with
stability of a smooth domain, 235 respect to the domain, 253
Stampacchia G., x, 64, 76, 88, 89, existence of orthogonal projection,
311, 325, 326, 329, 330 69
theorem of, 76 existence of projection on closed
Stein E. M., 127, 132, 133, 330 convex sets, 65
Steinhaus H., 255 existence of projection on closed
strict convex sets of u.c.
subsolution, 42 B-spaces, 82
supersolution, 42, 215 first classification of supersolutions,
strictly 43
decreasing, 156 generalized minimum principle of
Index 337
M. H. Protter and H. W. weak minimum principle of E.

Weinberger, 30 Hopf, 194
imbeddings of S. L. Sobolev, 99 Thieme H. R., 316, 329
minimum principle of E. Hopf, 5 Tietze H. F. F., 33
minimum principle of J. M. Bony, total cone, 158
189 trace
of Lp (RN )-regularity, 132 of a function, 102
of characterization of normal cones, operator, 102
160 theorem, 102
of characterization of the strong Triebel H., 126, 149, 152, 330
maximum principle, 216 Trudinger N., 89, 126, 127, 130, 131,
of concavity of σ0 with respect to 135, 136, 139, 152, 153, 271, 323
the potential, 232 Tzafriri L., 80, 85, 89, 325
of continuity, 139
of dominance of the principal u.c. B-space, 79
eigenvalue, 208 uniform decay property of E. Hopf, 25
of existence and uniqueness of the uniformly convex Banach space, 79
principal eigenvalue, 205 unity partition, 59
of F. Rellich and V. I. Kondrachov,
Valentine F. A., 89, 330
101
◦ Varadhan S. R. S., viii, ix, 220, 221,
of Garding, 118
223, 267, 268, 271, 320, 322
of G. Stampacchia, 76
Venturino M., 267, 330
of global approximation by smooth
Volterra V., 314
functions in Sobolev spaces,
Vyborny R., 267, 268, 320
97
of invertibility of (L, B, Ω), 120
Walter W., x, 38, 39, 60, 328, 330
of J. Lindenstrauss and L. Tzafriri, weak
85 derivative, 95
of M. G. Krein and M. A. Rutman, minimum principle of E. Hopf, 194
185 solution, 115
of P. D. Lax and A. N. Milgram, 78 topology, 80
of regularity of weak solutions, 146, Weinberger H. F., viii–x, xii, 30, 38,
151 39, 60, 188, 189, 217, 218, 221, 223,
of representation of F. Riesz, 71 224, 267, 328
of spectral structure, 162 Weyl H., 130, 152, 330
of stability of smooth domains, 235 Wloka J., 126, 135, 149, 152, 268, 330
of strict dominance of the principal
eigenvalue, 209 Yosida K., 81, 88, 89, 161–164, 330
of traces, 102 Young inequality, 104
second classification of Young W. H., 104, 247, 330
supersolutions, 53
uniform decay property of E. Hopf, Zygmund A., xi, 127, 130, 133, 136,
25 138, 321
weak boundary lemma of E. Hopf,
194

Operadores Elípticos - Julián López Gómez

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Operadores Elípticos - Julián López Gómez

Uploaded by

Copyright:

Available Formats

February 21, 2013 9:29 World Scientific Book - 9in x 6in book-lsoeo

Julián, you know...

no hard work, no glory...

and trust me...

if you want to be modern, you should be classical...

with all my love.

vi Linear Second Order Elliptic Operators

The main goal of this book is to provide a reasonably self-contained proof

viii Linear Second Order Elliptic Operators

iv) The principal eigenvalue of the linear eigenvalue problem

vi) (L, B, Ω) satisﬁes the weak maximum principle.

is relatively recent, this is a rather classical textbook on linear second order

x Linear Second Order Elliptic Operators

ter 2 analyzes the more recent theorems of classiﬁcation of supersolutions

Naturally, in order to establish the existence and the uniqueness of the

xii Linear Second Order Elliptic Operators

• REN2003-00707, of the National Plan of Natural Resources (Ministry

Madrid, December 31st, 2012

xiv Linear Second Order Elliptic Operators

1. The minimum principle 1

xvi Linear Second Order Elliptic Operators

3.3 The representation theorem of F. Riesz . . . . . . . . . . 71

4. Existence of weak solutions 91

5. Regularity of weak solutions 129

6. The Krein–Rutman theorem 155

6.2 Spectral theory of linear compact operators . . . . . . . . 161

7. The strong maximum principle 187

8. Properties of the principal eigenvalue 225

9. Principal eigenvalues of linear weighted boundary value problems 273

xviii Linear Second Order Elliptic Operators

9.4 Characterizing the existence of principal eigenvalues for

The minimum principle

This chapter considers second order diﬀerential operators of the form

where N ∈ N, N ≥ 1, Ω is a domain (open and connected by arcs) of RN ,

is referred to as the principal part of L, and aij , 1 ≤ i, j ≤ N , are said to

2 Linear Second Order Elliptic Operators

that f > 0 when f ≥ 0 and f > 0 on a set of positive measure, and we

Notations are shortened by setting

L(U ) := L(U, U ), Iso(U ) := Iso(U, U ).

1.1 Concept of ellipticity. First consequences

Throughout this book, | · | stands for the Euclidean norm of RN , N ≥ 1,

The minimum principle 3

where T means transposition, as coordinate vectors of RN will always be

Deﬁnition 1.1 (Elliptic operator). Given x ∈ Ω, it is said that L is

The paradigm of elliptic operator is the ‘minus Laplacian’,

Proposition 1.1. Suppose L is elliptic at x0 ∈ Ω. Then, there exists

4 Linear Second Order Elliptic Operators

where ∇x stands for the column gradient vector in the x-coordinates. By

for all 1 ≤ j ≤ N , and hence,

and, therefore, to prove (1.5) it suﬃces to construct a matrix M satisfying

The minimum principle 5

As a consequence from Proposition 1.1, the next result holds.

Theorem 1.1 (Classical minimum principle). Suppose the diﬀeren-

1.2 Minimum principle of E. Hopf

Theorem 1.2 (Minimum principle of E. Hopf ). Suppose L is uni-

6 Linear Second Order Elliptic Operators

Throughout this book, a function u ∈ C 2 (Ω) is said to be a harmonic

Establishing a contradiction concludes the proof. Let γ ∈ C([0, 1]; Ω) be a

The minimum principle 7

Fig. 1.1 The construction of y0 and y1

By construction, u(y1 ) > m. Thus, by the continuity of u, there exists

8 Linear Second Order Elliptic Operators

Fig. 1.2 The construction of B ρ (y2 )