Full-Note FPR Partition of Unity P-32 Thm2.7

*************************************
Applied Analysis I - (Advanced PDE I)

(Math 940, Fall 2014)
by
Baisheng Yan
Department of Mathematics
Michigan State University
[email protected]
Contents
Chapter 1. Preliminaries 1
§1.1. Banach Spaces 1
§1.2. Bounded Linear Operators 5
§1.3. Weak Convergence and Compact Operators 7
§1.4. Spectral Theory for Compact Linear Operators 9
§1.5. Nonlinear Functional Analysis 11
Chapter 2. Sobolev Spaces 23
§2.1. Weak Derivatives and Sobolev Spaces 23
§2.2. Approximations and Extensions 27
§2.3. Sobolev embedding Theorems 39
§2.4. Compactness 45
§2.5. Additional Topics 47
§2.6. Spaces of Functions Involving Time 50
Chapter 3. Second-Order Linear Elliptic Equations 55
§3.1. Differential Equations in Divergence Form 55
§3.2. The Lax-Milgram Theorem 58
§3.3. Energy Estimates and Existence Theory 59
§3.4. Fredholm Alternatives 64
§3.5. Regularity 69
§3.6. Regularity for Linear Systems* 76
Chapter 4. Linear Evolution Equations 97
§4.1. Second-order Parabolic Equations 97
§4.2. Second-order Hyperbolic Equations 111
§4.3. Hyperbolic Systems of First-order Equations 121
§4.4. Semigroup Theory 126
Index 143
iii
Chapter 1
Preliminaries
1.1. Banach Spaces

1.1.1. Vector Spaces. A (real) vector space is a set X, whose elements are called
vectors, and in which two operations, addition and scalar multiplication, are defined
as follows:
(a) To every pair of vectors x and y corresponds a vector x + y in such a way that
x+y =y+x and x + (y + z) = (x + y) + z.
X contains a unique vector 0 (the zero vector or origin of X) such that x+0 = x
for every x ∈ X, and to each x ∈ X corresponds a unique vector −x such that
x + (−x) = 0.
(b) To every pair (α, x), with α ∈ R and x ∈ X, corresponds a vector αx in such a
way that
1x = x, α(βx) = (αβ)x
and such that the two distributive laws
α(x + y) = αx + αy, (α + β)x = αx + βx
hold.
A nonempty subset M of a vector space X is called a subspace of X if αx + βy ∈ M

for all x, y ∈ M and all α, β ∈ R. A subset M of a vector space X is said to be convex if
tx + (1 − t)y ∈ M whenever t ∈ (0, 1), x, y ∈ M . (Clearly, every subspace of X is convex.)
Let x1 , . . . , xn be elements of a vector space X. The set of all α1 x1 + · · · + αn xn , with
αi ∈ R, is called the span of x1 , . . . , xn and is denoted by span{x1 , . . . , xn }. The elements
x1 , . . . , xn are said to be linearly independent if α1 x1 +· · ·+αn xn = 0 implies that αi = 0
for each i. If, on the other hand, α1 x1 + · · · + αn xn = 0 does not imply αi = 0 for each
i, the elements x1 , . . . , xn are said to be linearly dependent . An arbitrary collection of
vectors is said to be linearly independent if every finite subset of distinct elements is linearly
independent.
The dimension of a vector space X, denoted by dim X, is either 0, a positive integer
or ∞. If X = {0} then dim X = 0; if there exist linearly independent {u1 , . . . , un } such
1
2 1. Preliminaries
that each x ∈ X has a (unique) representation of the form

x = α1 u1 + · · · + αn un with αi ∈ R
then dim X = n and {u1 , . . . , un } is a basis for X; in all other cases dim X = ∞.
1.1.2. Normed Spaces. A (real) vector space X is said to be a normed space if to

every x ∈ X there is associated a nonnegative real number kxk, called the norm of x, in
such a way that
(a) kx + yk ≤ kxk + kyk for all x and y in X (Triangle inequality)
(b) kαxk = |α|kxk for all x ∈ X and all α ∈ R
(c) kxk > 0 if x 6= 0.
Note that (b) and (c) imply that kxk = 0 iff x = 0. Moreover, it easily follows from (a) that
|kxk − kyk| ≤ kx − yk for all x, y ∈ X.
1.1.3. Completeness and Banach Spaces. A sequence {xn } in a normed space X is

called a Cauchy sequence if, for each > 0, there exists an integer N such that kxm −
xn k < for all m, n ≥ N . We say xn → x in X if limn→∞ kxn − xk = 0 and, in this case, x
is called the limit of {xn }. X is called complete if every Cauchy sequence in X converges
to a limit in X.
A complete (real) normed space is called a (real) Banach space. A Banach space
is separable if it contains a countable dense set. It can be shown that a subspace of a
separable Banach space is itself separable.
Example 1.1. Let Ω be an open subset of Rn , n ≥ 1. The set C(Ω) of (real-valued)
continuous functions defined on Ω is an infinite dimensional vector space with the usual
definitions of addition and scalar multiplication:
(f + g)(x) = f (x) + g(x) for f, g ∈ C(Ω), x ∈ Ω
(αf )(x) = αf (x) for α ∈ R, f ∈ C(Ω), x ∈ Ω.
C(Ω̄) consists of those functions which are uniformly continuous on Ω. Each such function
has a continuous extension to Ω̄. C0 (Ω) consists of those functions which are continuous
in Ω and have compact support in Ω. (The support of a function f defined on Ω is the
closure of the set {x ∈ Ω : f (x) 6= 0} and is denoted by supp(f ).) The latter two spaces are
clearly subspaces of C(Ω).
For each n-tuple α = (α1 , . . . , αn ) of nonnegative integers, we denote by Dα the partial
derivative
D1α1 · · · Dnαn , Di = ∂/∂xi
of order |α| = α1 + · · · + αn . If |α| = 0, then D0 = I(identity).
For integers m ≥ 0, let C m (Ω) be the collection of all f ∈ C(Ω) such that Dα f ∈ C(Ω)
for all α with |α| ≤ m. We write f ∈ C ∞ (Ω) iff f ∈ C m (Ω) for all m ≥ 0. For
m ≥ 0, define C0m (Ω) = C0 (Ω) ∩ C m (Ω) and let C0∞ (Ω) = C0 (Ω) ∩ C ∞ (Ω). The spaces
C m (Ω), C ∞ (Ω), C0m (Ω), C0∞ (Ω) are all subspaces of the vector space C(Ω). Similar defini-
tions can be given for C m (Ω̄) etc.
For m ≥ 0, define X to be the set of all f ∈ C m (Ω) for which
X
kf km,∞ ≡ sup |Dα f (x)| < ∞.
Ω
|α|≤m
1.1. Banach Spaces 3
Then X is a Banach space with norm k · km,∞ . To prove, for example, the completeness
when m = 0, we let {fn } be a Cauchy sequence in X, i.e., assume for any ε > 0 there is a
number N (ε) such that for all x ∈ Ω
sup |fn (x) − fm (x)| < ε if m, n > N (ε).

x∈Ω
But this means that {fn (x)} is a uniformly Cauchy sequence of bounded continuous func-
tions, and thus converges uniformly to a bounded continuous function f (x). Letting m → ∞
in the above inequality shows that kfn − f km,∞ → 0.
Note that the same proof is valid for the set of bounded continuous scalar-valued func-
tions defined on a nonempty subset of a normed space X.
Example 1.2. Let Ω be a nonempty Lebesgue measurable set in Rn . For p ∈ [1, ∞), we
denote by Lp (Ω) the set of equivalence classes of Lebesgue measurable functions on Ω for
which
Z 1
p
p
kf kp ≡ |f (x)| dx < ∞.
Ω
(Two functions belong to the same equivalence class, i.e., are equivalent, if they differ
only on a set of measure 0.) Let L∞ (Ω) denote the set of equivalence classes of Lebesgue
measurable functions on Ω for which
kf k∞ ≡ ess-supx∈Ω |f (x)| < ∞.
Then Lp (Ω), 1 ≤ p ≤ ∞, are Banach spaces with norms k · kp . For p ∈ [1, ∞] we write
f ∈ Lploc (Ω) iff f ∈ Lp (K) for each compact set K ⊂ Ω.
For the sake of convenience, we will also consider Lp (Ω) as a set of functions. With this
convention in mind, we can assert that C0 (Ω) ⊂ Lp (Ω). In fact, if p ∈ [1, ∞), then as we shall
show later, C0 (Ω) is dense in Lp (Ω). The space Lp (Ω) is also separable if p ∈ [1, ∞). This
follows easily, when Ω is compact, from the last remark and the Weierstrass approximation
theorem.
Finally we recall that if p, q, r ∈ [1, ∞] with p−1 + q −1 = r−1 , then Hölder’s inequality
implies that if f ∈ Lp (Ω) and g ∈ Lq (Ω), then f g ∈ Lr (Ω) and
kf gkr ≤ kf kp kgkq .
Example 1.3. The Cartesian product X × Y , of two vector spaces X and Y , is itself a
vector space under the following operations of addition and scalar multiplication:
[x1 , y1 ] + [x2 , y2 ] = [x1 + x2 , y1 + y2 ]
α[x, y] = [αx, αy].

If in addition, X and Y are normed spaces with norms k · kX , k · kY respectively, then X × Y
becomes a normed space under the norm
k[x, y]k = kxkX + kykY .
Moreover, under this norm, X × Y becomes a Banach space provided X and Y are Banach
spaces.
4 1. Preliminaries
1.1.4. Hilbert Spaces. Let H be a real vector space. H is said to be an inner product
space if to every pair of vectors x and y in H there corresponds a real-valued function
(x, y), called the inner product of x and y, such that
(a) (x, y) = (y, x) for all x, y ∈ H
(b) (x + y, z) = (x, z) + (y, z) for all x, y, z ∈ H
(c) (λx, y) = λ(x, y) for all x, y ∈ H, λ ∈ R
(d) (x, x) ≥ 0 for all x ∈ H, and (x, x) = 0 if and only if x = 0.
For x ∈ H we set
(1.1) kxk = (x, x)1/2 .
Theorem 1.4. If H is an inner product space, then for all x and y in H, it follows that
(a) |(x, y)| ≤ kxk kyk (Cauchy-Schwarz inequality);
(b) kx + yk ≤ kxk + kyk (Triangle inequality);
(c) kx + yk2 + kx − yk2 = 2(kxk2 + kyk2 ) (Parallelogram law).
Proof. (a) is obvious if x = 0, and otherwise it follows by taking δ = −(x, y)/kxk2 in

0 ≤ kδx + yk2 = |δ|2 kxk2 + 2δ(x, y) + kyk2 .
This identity, with δ = 1, and (a) imply (b). (c) follows easily by using (1.1).
Furthermore, by (d), equation (1.1) defines a norm on an inner product space H. If H

is complete under this norm, then H is said to be a Hilbert space.
Example 1.5. The space L2 (Ω) is a Hilbert space with inner product
Z
(f, g) = f (x)g(x) dx for all f, g ∈ L2 (Ω).
Ω
Theorem 1.6. Every nonempty closed convex subset S of a Hilbert space H contains a
unique element of minimal norm.
Proof. Choose xn ∈ S so that kxn k → d ≡ inf{kxk : x ∈ S}. Since (1/2)(xn + xm ) ∈ S,

we have kxn + xm k2 ≥ 4d2 . Using the parallelogram law, we see that
(1.2) kxn − xm k2 ≤ 2(kxn k2 − d2 ) + 2(kxm k2 − d2 )
and therefore {xn } is a Cauchy sequence in H. Since S is closed, {xn } converges to some
x ∈ S and kxk = d. If y ∈ S and kyk = d, then the parallelogram law implies, as in (1.2),
that x = y.
If (x, y) = 0, then x and y are said to be orthogonal, written sometimes as x ⊥ y. For

M ⊂ H, the orthogonal complement of M , denoted by M ⊥ , is defined to be the set of
all x ∈ H such that (x, y) = 0 for all y ∈ M . It is easily seen that M ⊥ is a closed subspace
of H. Moreover, if M is a dense subset of H and if x ∈ M ⊥ , then in fact, x ∈ H ⊥ which
implies x = 0.
Theorem 1.7. (Projection) Suppose M is a closed subspace of a Hilbert space H. Then
for each x ∈ H there exist unique y ∈ M , z ∈ M ⊥ such that x = y + z. The element y is
called the projection of x onto M .
1.2. Bounded Linear Operators 5
Proof. Let S = {x − y : y ∈ M }. It is easy to see that S is convex and closed. Theorem 1.6
implies that there exists a y ∈ M such that kx − yk ≤ kx − wk for all w ∈ M . Let z = x − y.
For an arbitrary w ∈ M , w 6= 0, let α = (z, w)/kwk2 and note that
kzk2 ≤ kz − αwk2 = kzk2 − |(z, w)/kwk|2
which implies (z, w) = 0. Therefore z ∈ M ⊥ . If x = y 0 + z 0 for some y 0 ∈ M , z 0 ∈ M ⊥ , then
y 0 − y = z − z 0 ∈ M ∩ M ⊥ = {0}, which implies uniqueness.
Remark. In particular, if M is a proper closed subspace of H, then there is a nonzero

element in M ⊥ . Indeed, for x ∈ H\M , let y be the projection of x on M . Then z = x − y
is a nonzero element of M ⊥ .
1.2. Bounded Linear Operators

1.2.1. Operators on a Banach Space. Let X, Y be real vector spaces. A map T : X →
Y is said to be a linear operator from X to Y if
T (αx + βy) = αT x + βT y
for all x, y ∈ D(T ) and all α, β ∈ R.
Let X, Y be normed spaces. A linear operator T from X to Y is said to be bounded if
there exists a constant m > 0 such that
(1.3) kT xk ≤ mkxk for all x ∈ X.
We define the operator norm kT k of T by
(1.4) kT k = sup kT xk = sup kT xk.
x∈X, kxk=1 x∈X, kxk≤1
The collection of all bounded linear operators T : X → Y will be denoted by B(X, Y ). We

shall also set B(X) = B(X, X) when X = Y . Observe that
kT Sk ≤ kT kkSk if S ∈ B(X, Y ), T ∈ B(Y, Z).
Theorem 1.8. If X and Y are normed spaces, then B(X, Y ) is a normed space with norm
defined by equation (1.4). If Y is a Banach space, then B(X, Y ) is also a Banach space.
Proof. It is easy to see that B(X, Y ) is a normed space. To prove completeness, assume
that {Tn } is a Cauchy sequence in B(X, Y ). Since
(1.5) kTn x − Tm xk ≤ kTn − Tm kkxk
we see that, for fixed x ∈ X, {Tn x} is a Cauchy sequence in Y and therefore we can define
a linear operator T by
T x = lim Tn x for all x ∈ X.
n→∞
If ε > 0, then the right side of (1.5) is smaller than εkxk provided that m and n are large
enough. Thus, (letting n → ∞)
kT x − Tm xk ≤ εkxk for all large enough m.
Hence, kT xk ≤ (kTm k + ε)kxk, which shows that T ∈ B(X, Y ). Moreover, kT − Tm k < ε
for all large enough m. Hence, limn→∞ Tn = T .
The following theorems are important in linear functional analysis; see, e.g., [?].
Theorem 1.9. (Banach-Steinhaus) Let X be a Banach space and Y a normed space. If
A ⊂ B(X, Y ) is such that supT ∈A kT xk < ∞ for each fixed x ∈ X, then supT ∈A kT k < ∞.
6 1. Preliminaries
Theorem 1.10. (Bounded Inverse) If X and Y are Banach spaces and if T ∈ B(X, Y )
is one-to-one and onto, then T −1 ∈ B(Y, X).
1.2.2. Dual Spaces and Reflexivity. When X is a (real) normed space, the Banach
space B(X, R) will be called the (normed) dual space of X and will be denoted by X*. El-
ements of X* are called bounded linear functionals or continuous linear functionals
on X. Frequently, we shall use the notation hf, xi to denote the value of f ∈ X* at x ∈ X.
Using this notation we note that |hf, xi| ≤ kf k kxk for all f ∈ X*, x ∈ X.
Example 1.11. Suppose 1 < p, q < ∞ satisfy 1/p + 1/q = 1 and let Ω be a nonempty
Lebesgue measurable set in Rn . Then Lp (Ω)∗ = Lq (Ω). The case of p = ∞ is different. The
dual of L∞ is much larger then L1 .
The following results can be found in [?].

Theorem 1.12. (Hahn-Banach) Let X be a normed space and Y a subspace of X. As-
sume f ∈ Y ∗ . Then there exists a bounded linear functional f˜ ∈ X ∗ such that
hf˜, yi = hf, yi ∀ y ∈ Y, kf˜kX ∗ = kf kY ∗ .
Corollary 1.13. Let X be a normed space and x0 6= 0 in X. Then there exists f ∈ X ∗
such that
kf k = 1, hf, x0 i = kx0 k.
The dual space X ∗∗ of X ∗ is called the second dual space of X and is again a Banach
space. Note that to each x ∈ X we can associate a unique Fx ∈ X ∗∗ by Fx (f ) = hf, xi for
all f ∈ X ∗ . From Corollary 1.13, one can also show that kFx k = kxk. Thus, the (canonical)
mapping J : X → X**, given by Jx = Fx , is a linear isometry of X onto the subspace
J(X) of X ∗∗ . Since J is one-to-one, we can identify X with J(X).
A Banach space X is called reflexive if its canonical map J is onto X ∗∗ . For example,
all Lp spaces with 1 < p < ∞ are reflexive.
We shall need the following properties of reflexive spaces.

Theorem 1.14. Let X and Y be Banach spaces.
(a) X is reflexive iff X* is reflexive.
(b) If X is reflexive, then a closed subspace of X is reflexive.
(c) Let T : X → Y be a linear bijective isometry. If Y is reflexive, then X is
reflexive.
1.2.3. Bounded Linear Functionals on a Hilbert Space.

Theorem 1.15. (Riesz Representation) If H is a Hilbert space and f ∈ H*, then there
exists a unique y ∈ H such that
f (x) = hf, xi = (x, y) for all x ∈ H.
Moreover, kf k = kyk.
Proof. If f (x) = 0 for all x, take y = 0. Otherwise, there is an element z ∈ N (f )⊥ such

that kzk = 1. (Note that the linearity and continuity of f implies that N (f ) is a closed
subspace of H.) Put u = f (x)z −f (z)x. Since f (u) = 0, we have u ∈ N (f ). Thus (u, z) = 0,
which implies
f (x) = f (x)(z, z) = f (z)(x, z) = (x, f (z)z) = (x, y),
1.3. Weak Convergence and Compact Operators 7
where y = f (z)z. To prove uniqueness, suppose (x, y) = (x, y 0 ) for all x ∈ H. Then in
particular, (y − y 0 , y − y 0 ) = 0, which implies y = y 0 . From the Cauchy-Schwarz inequality
we get |f (x)| ≤ kxkkyk, which yields kf k ≤ kyk. The reverse inequality follows by choosing
x = y in the representation.
Corollary 1.16. H is reflexive.
Let T : H → H be an operator on the Hilbert space H. We define the Hilbert space

adjoint T * : H → H as follows:
(T x, y) = (x, T *y) for all x, y ∈ H.
The adjoint operator is easily seen to be linear.
Theorem 1.17. Let H be a Hilbert space. If T ∈ B(H), then T * ∈ B(H) and kT k = kT *k.
Proof. For any y ∈ H and all x ∈ H, set f (x) = (T x, y). Then it is easily seen that
f ∈ H ∗ . Hence by the Riesz representation theorem, there exists a unique z ∈ H such that
(T x, y) = (x, z) for all x ∈ H, i.e., D(T ∗ ) = H. Moreover, kT ∗ yk = kzk = kf k ≤ kT kkyk,
i.e., T ∗ ∈ B(H) and kT ∗ k ≤ kT k. The reverse inequality follows easily from kT xk2 =
(T x, T x) = (x, T ∗ T x) ≤ kT xkkT ∗ kkxk.
1.3. Weak Convergence and Compact Operators

1.3.1. Weak Convergence. Let X be a normed space. A sequence xn ∈ X is said to
be weakly convergent to an element x ∈ X, written xn * x, if hf, xn i → hf, xi for all
f ∈ X ∗.
Theorem 1.18. Let {xn } be a sequence in X.
(a) Weak limits are unique.
(b) If xn → x, then xn * x.
(c) If xn * x, then {xn } is bounded and kxk ≤ lim inf kxn k.
Proof. To prove (a), suppose that x and y are both weak limits of the sequence {xn } and
set z = x − y. Then hf, zi = 0 for every f ∈ X ∗ and by Corollary 1.13, z = 0. To prove (b),
let f ∈ X ∗ and note that xn → x implies hf, xn i → hf, xi since f is continuous. To prove (c),
assume xn * x and consider the sequence {Jxn } of elements of X ∗∗ , where J : X → X ∗∗
is the bounded operator defined above. For each f ∈ X ∗ , sup |Jxn (f )| = sup |hf, xn i| < ∞
(since hf, xn i converges). By the Banach-Steinhaus Theorem, there exists a constant c such
that kxn k = kJxn k ≤ c which implies {xn } is bounded. Finally, for f ∈ X ∗
|hf, xi| = lim |hf, xn i| ≤ lim inf kf kkxn k = kf k lim inf kxn k
which implies the desired inequality since kxk = supkf k=1 |hf, xi|.
We note that in a Hilbert space H, the Riesz representation theorem implies that xn * x
means (xn , y) → (x, y) for all y ∈ H. Moreover, we have
(xn , yn ) → (x, y) if xn * x, yn → y.
This follows from the estimate
|(x, y) − (xn , yn )| = |(x − xn , y) − (xn , yn − y)| ≤ |(x − xn , y)| + kxn kky − yn k
and the fact that kxn k is bounded.
8 1. Preliminaries
The main result of this section is given by:

Theorem 1.19. If X is a reflexive Banach space, then the closed unit ball is weakly
compact, i.e., the sequence {xn }, with kxn k ≤ 1 has a subsequence which converges weakly
to an x with kxk ≤ 1.
1.3.2. Compact Operators. Let X and Y be normed spaces. An operator T : X → Y is

said to be compact if it maps bounded sets in X into relatively compact sets in Y , i.e., if
for every bounded sequence {xn } in X, {T xn } has a subsequence which converges to some
element of Y .
Since relatively compact sets are bounded, it follows that a compact operator is bounded.
On the other hand, since bounded sets in finite-dimensional spaces are relatively compact, it
follows that a bounded operator with finite dimensional range is compact. It can be shown
that the identity map I : X → X (kIxk = kxk) is compact iff X is finite-dimensional.
Finally we note that the operator ST is compact if (a) T : X → Y is compact and S : Y → Z
is continuous or (b) T is bounded and S is compact.
One of the main methods of proving the compactness of certain operators is based upon
the Ascoli theorem.
Let Ω be a subset of the normed space X. A set S ⊂ C(Ω) is said to be equicontinuous
if for each ε > 0 there exists a δ > 0 such that |f (x) − f (y)| < ε for all x, y ∈ Ω with
kx − yk < δ and for all f ∈ S.
Theorem 1.20. (Ascoli) Let Ω be a relatively compact subset of a normed space X and
let S ⊂ C(Ω). Then S is relatively compact if it is bounded and equicontinuous.
Remark. In other words, every bounded equicontinuous sequence of functions has a uni-
formly convergent subsequence.
Theorem 1.21. Let X and Y be Banach spaces. If Tn : X → Y are linear and compact
for n ≥ 1 and if limn→∞ kTn − T k = 0, then T is compact. Thus, compact operators form
a closed, but not a dense, subspace of B(X, Y ).
Proof. Let {xn } be a sequence in X with M = supn kxn k < ∞. Let A1 denote an infinite
set of integers such the sequence {T1 xn }n∈A1 converges. For k ≥ 2 let Ak ⊂ Ak−1 denote
an infinite set of integers such that the sequence {Tk xn }n∈Ak converges. Choose n1 ∈ A1
and nk ∈ Ak , nk > nk−1 for k ≥ 2. Choose ε > 0. Let k be such that kT − Tk kM < ε/4
and note that
kT xni − T xnj k ≤ k(T − Tk )(xni − xnj )k + kTk xni − Tk xnj k < ε/2 + kTk xni − Tk xnj k.
Since {Tk xni }∞ ∞
i=1 converges, {T xni }i=1 is a Cauchy sequence.
Theorem 1.22. Let X and Y be normed spaces.
(a) If T ∈ B(X, Y ), then T is weakly continuous, i.e.,
xn * x implies T xn * T x.
(b) If T : X → Y is weakly continuous and X is a reflexive Banach space, then T
is bounded.
(c) If T ∈ B(X, Y ) is compact, then T is strongly continuous, i.e.,
xn * x implies T xn → T x.
1.4. Spectral Theory for Compact Linear Operators 9
(d) If T : X → Y is strongly continuous and X is a reflexive Banach space, then T

is compact.
Proof. (a) Let xn * x. Then for every g ∈ Y ∗

hg, T xn i = hT ∗ g, xn i → hT ∗ g, xi = hg, T xi.
(b) If not, there is a bounded sequence {xn } such that kT xn k → ∞. Since X is reflexive,
{xn } has a weakly convergent subsequence, {xn0 }, and so {T xn0 } also converges weakly. But
then {T xn0 } is bounded, which is a contradiction.
(c) Let xn * x. Since T is compact and {xn } is bounded, there is a subsequence {xn0 }
such that T xn0 → z, and thus T xn0 * z. By (a), T xn * T x, and so T xn0 → T x. Now
it is easily seen that every subsequence of {xn } has a subsequence, say {xn0 }, such that
T xn0 → T x. But this implies the whole sequence T xn → T x (See the appendix).
(d) Let {xn } be a bounded sequence. Since X is reflexive, there is a subsequence {xn0 }
such that xn0 * x. Hence T xn0 → T x, which implies T is compact.
Theorem 1.23. Let H be a Hilbert space. If T : H → H is linear and compact, then T * is
compact.
Proof. Let {xn } be a sequence in H satisfying kxn k ≤ m. The sequence {T *xn } is therefore
bounded, since T * is bounded. Since T is compact, by passing to a subsequence if necessary,
we may assume that the sequence {T T *xn } converges. But then
kT *(xn − xm )k2 = (xn − xm , T T *(xn − xm ))
≤ 2mkT T *(xn − xm )k → 0 as m, n → ∞.
Since H is complete, the sequence {T *xn } is convergent and hence T * is compact.
1.4. Spectral Theory for Compact Linear Operators

1.4.1. Fredholm Alternative.
Theorem 1.24. (Fredholm Alternative) Let T : H → H be a compact linear operator
on the Hilbert space H. Then equations (I − T )x = 0, (I − T ∗ )x∗ = 0 have the same finite
number of linearly independent solutions. Moreover,
(a) For y ∈ H, the equation (I − T )x = y has a solution iff (y, x∗ ) = 0 for every
solution x∗ of (I − T ∗ )x∗ = 0.
(b) For z ∈ H, the equation (I − T ∗ )x∗ = z has a solution iff (z, x) = 0 for every
solution x of (I − T )x = 0.
(c) The inverse operator (I − T )−1 ∈ B(H) whenever it exists.
1.4.2. Spectrum of Compact Operators. A subset S of a Hilbert space H is said to be

an orthonormal set if each element of S has norm 1 and if every pair of distinct elements
in S is orthogonal. It easily follows that an orthonormal
P set is linearly independent. An
orthonormal set S is said to be complete if x = φ∈S (x, φ)φ for all x ∈ H. It can be shown
that (x, φ) 6= 0 for at most countably many φ ∈ S. This series is called the Fourier series
for x with respect to the orthonormal set {φ}. Let {φi }∞
i=1 be a countable orthonormal set
PN 2
in H. Upon expanding kx − n=1 (x, φn )φn k , we arrive at Bessel’s inequality:
∞
X
|(x, φn )|2 ≤ kxk2 .
n=1
10 1. Preliminaries
Let T : D(T ) ⊂ H → H be a linear operator on the real Hilbert space H. The set ρ(T )
of all scalars λ ∈ R for which (T − λI)−1 ∈ B(H) is called the resolvent set of T . The
operator R(λ) = (T − λI)−1 is known as the resolvent of T . σ(T ) = R \ ρ(T ) is called the
spectrum of T . It can be shown that ρ(T ) is an open set and σ(T ) is a closed set. The
set of λ ∈ R for which there exists a nonzero x ∈ N (T − λI) is called the point spectrum
of T and is denoted by σp (T ). The elements of σp (T ) are called the eigenvalues of T and
the nonzero members of N (T − λI) are called the eigenvectors (or eigenfunctions if X
is a function space) of T .
If T is compact and λ 6= 0, then by the Fredholm alternative, either λ ∈ σp (T ) or
λ ∈ ρ(T ). Moreover, if H is infinite-dimensional, then 0 6∈ ρ(T ); otherwise, T −1 ∈ B(H)
and T −1 T = I is compact. As a consequence, σ(T ) consists of the nonzero eigenvalues of T
together with the point 0. The next result shows that σp (T ) is either finite or a countably
infinite sequence tending to zero.
Theorem 1.25. Let T : X → X be a compact linear operator on the normed space X.
Then for each r > 0 there exist at most finitely many λ ∈ σp (T ) for which |λ| > r.
1.4.3. Symmetric Compact Operators. The next result implies that a symmetric com-
pact operator on a Hilbert space has at least one eigenvalue. On the other hand, an arbi-
trary bounded, linear, symmetric operator need not have any eigenvalues. As an example,
let T : L2 (0, 1) → L2 (0, 1) be defined by T u(x) = xu(x).
Theorem 1.26. Suppose T ∈ B(H) is symmetric, i.e., (T x, y) = (x, T y) for all x, y ∈ H.
Then
kT k = sup |(T x, x)|.
kxk=1
Moreover, if H 6= {0}, then there exists a real number λ ∈ σ(T ) such that |λ| = kT k. If
λ ∈ σp (T ), then in absolute value λ is the largest eigenvalue of T .
Proof. Clearly m ≡ supkxk=1 |(T x, x)| ≤ kT k. To show kT k ≤ m, observe that for all
x, y ∈ H
2(T x, y) + 2(T y, x) = (T (x + y), x + y) − (T (x − y), x − y)
≤ m(kx + yk2 + kx − yk2 )
= 2m(kxk2 + kyk2 )
where the last step follows from the paralleogram law. Hence, if T x 6= 0 and y =
(kxk/kT xk)T x, then
2kxkkT xk = (T x, y) + (y, T x) ≤ m(kxk2 + kyk2 ) = 2mkxk2
which implies kT xk ≤ mkxk. Since this is also valid when T x = 0, we have kT k ≤ m. To
prove the ‘moreover’ part, choose xn ∈ H such that kxn k = 1 and kT k = limn→∞ |(T xn , xn )|.
By renaming a subsequence of {xn }, we may assume that (T xn , xn ) converge to some real
number λ with |λ| = kT k. Observe that
k(T − λ)xn k2 = kT xn k2 − 2λ(T xn , xn ) + λ2 kxn k2
≤ 2λ2 − 2λ(T xn , xn ) → 0.
We now claim that λ ∈ σ(T ). Otherwise, we arrive at the contradiction
1 = kxn k = k(T − λ)−1 (T − λ)xn k ≤ k(T − λ)−1 k k(T − λ)xn k → 0.
1.5. Nonlinear Functional Analysis 11
Finally, we note that if T φ = µφ, with kφk = 1, then |µ| = |(T φ, φ)| ≤ kT k which implies
the last assertion of the theorem.
Finally we have the following result.
Theorem 1.27. Let H be a separable Hilbert space and suppose T : H → H is linear,

symmetric and compact. Then there exists a countable complete orthonormal set in H
consisting of eigenvectors of T .
1.5. Nonlinear Functional Analysis

In this final preliminary section, we list some useful results in nonlinear functional anal-
ysis. Lots of the proof and other results can be found in the volumes of Zeidler’s book
[?].
1.5.1. Contraction Mapping Theorem. Let X be a normed space. A map T : X → X

is called a contraction if there exists a number k < 1 such that
(1.6) kT x − T yk ≤ kkx − yk for all x, y ∈ X.
Theorem 1.28. (Contraction Mapping) Let T : S ⊂ X → S be a contraction on the

closed nonempty subset S of the Banach space X. Then T has a unique fixed point, i.e.,
there exists a unique solution x ∈ S of the equation T x = x. Moreover, x = limn→∞ T n x0
for any choice of x0 ∈ S.
Proof. To prove uniqueness, suppose T x = x, T y = y. Since k < 1, we get x = y from

kx − yk = kT x − T yk ≤ kkx − yk.
To show that T has a fixed point we set up an iteration procedure. For any x0 ∈ S set
xn+1 = T xn , n = 0, 1, ...
Note that xn+1 ∈ S and xn+1 = T n+1 x0 . We now claim that {xn } is a Cauchy sequence.
Indeed, for any integers n, p
n+p−1
X
n+p n
kxn+p − xn k = kT x0 − T x0 k ≤ kT j+1 x0 − T j x0 k
j=n
n+p−1
X kn
≤ k j kT x0 − x0 k ≤ kT x0 − x0 k.
1−k
j=n
Hence as n → ∞, kxn+p − xn k → 0 independently of p, so that {xn } is a Cauchy sequence

with limit x ∈ S. Since T is continuous, we have
T x = lim T xn = lim xn+1 = x
n→∞ n→∞
and thus x is the unique fixed point. Note that the fixed point x is independent of x0 since
x is a fixed point and fixed points are unique.
Our main existence result will be based upon the following so-called method of con-
tinuity or continuation method.
12 1. Preliminaries
Theorem 1.29. Let T0 , T1 ∈ B(X, Y ), where X is a Banach space and Y is a normed

space. For each t ∈ [0, 1] set
Tt = (1 − t)T0 + tT1
and suppose there exists a constant c > 0 such that for all t ∈ [0, 1] and x ∈ X
(1.7) kxkX ≤ ckTt xkY .
Then R(T1 ) = Y if R(T0 ) = Y .
Proof. Set S = {t ∈ [0, 1] : R(Tt ) = Y }. By hypothesis, 0 ∈ S. We need to show that

1 ∈ S. In this direction we will show that if τ > 0 and τ c(kT1 k + kT0 k) < 1, then
(1.8) [0, s] ⊂ S implies [0, s + τ ] ⊂ S.
(Note that any smaller τ works.) Since τ can be chosen independently of s, (1.8) applied
finitely many times gets us from 0 ∈ S to 1 ∈ S.
Let s ∈ S. For t = s + τ, Tt x = f is equivalent to the equation
(1.9) Ts x = f + τ T0 x − τ T1 x.
By (1.7), Ts−1 : Y → X exists and kTs−1 k ≤ c. Hence (1.9) is equivalent to
(1.10) x = Ts−1 (f + τ T0 x − τ T1 x) ≡ Ax
and for A : X → X we have for all x, y ∈ X
kAx − Ayk ≤ τ c(kT1 k + kT0 k)kx − yk.
By the contraction mapping theorem, (1.10) has a solution and this completes the proof.
1.5.2. Nemytskii Operators. Let Ω be a nonempty measurable set in Rn and let f :

Ω × R → R be a given function. Assume
(i) for every ξ ∈ R, f (x, ξ) (as a function of x) is measurable on Ω
(ii) for almost all x ∈ Ω, f (x, ξ) (as a function of ξ) is continuous on R
(iii) for all (x, ξ) ∈ Ω × R
|f (x, ξ)| ≤ a(x) + b|ξ|p/q
where b is a fixed nonnegative number, a ∈ Lq (Ω) is nonnegative and 1 < p, q < ∞, 1/p +
1/q = 1. Note that p/q = p − 1. Then the Nemytskii operator N is defined by
N u(x) = f (x, u(x)), x ∈ Ω.
We have the following result needed later.
Lemma 1.30. N : Lp (Ω) → Lq (Ω) is continuous and bounded with
(1.11) kN ukq ≤ const (kakq + kukp/q
p ) for all u ∈ Lp (Ω)
and
Z
(1.12) hN u, vi = f (x, u(x))v(x)dx for all u, v ∈ Lp (Ω).
Ω
Here h·, ·i denotes the duality pairing between Lp (Ω) and Lq (Ω).
Proof. Since u ∈ Lp (Ω), the function u(x) is measurable on Ω and thus, Pn by (i) and P
(ii), the
function f (x, u(x)) is also measurable on Ω. From the inequality ( i=1 ξi ) ≤ c ni=1 ξir
r
and (iii) we get

|f (x, u(x))|q ≤ const(|a(x)|q + |u(x)|p ).
Integrating over Ω and applying the above inequality once more yields (1.11), which shows
that N is bounded.
To show that N is continuous, let un → u in Lp (Ω). Then there is a subsequence {un0 }
and a function v ∈ Lp (Ω) such that un0 (x) → u(x) a.e. and |un0 (x)| ≤ v(x) a.e. for all n.
Hence
Z
q
kN un0 − N ukq = |f (x, un0 (x)) − f (x, u(x))|q dx
Ω
Z
≤ const (|f (x, un0 (x))|q + |f (x, u(x))|q )dx
ZΩ
≤ const (|a(x)|q + |v(x)|p + |u(x)|p )dx.
Ω
By (ii), f (x, u (x)) − f (x, u(x)) → 0 as n → ∞ for almost all x ∈ Ω. The dominating
n0
convergence theorem implies that kN un0 − N ukq → 0. By repeating this procedure for any
subsequence of un0 , it follows that kN un − N ukq → 0 which implies that N is continuous.
Since N u ∈ (Lp (Ω))*, the integral representation (1.12) is clear.
Remarks. (a) The following remarkable statement can be proved: If f satisfies (i) and (ii)
above and if the corresponding Nemytskii operator is such that N : Lp (Ω) → Lq (Ω), then
N is continuous, bounded and (iii) holds.
(b) If (iii) is replaced by
(iii)0 for all (x, ξ) ∈ Ω × R
|f (x, ξ)| ≤ a(x) + b|ξ|p/r
where b is a fixed nonnegative number, a ∈ Lr (Ω) is nonnegative and 1 < p, r < ∞, then
the above results are valid with q replaced by r. (i.e., N : Lp (Ω) → Lr (Ω).)
(c) We say that f satisfies the Caratheodory property, written f ∈ Car, if (i) and
(ii) above are met. If in addition (iii) is met, then we write f ∈ Car(p).
1.5.3. Differentiability. Let S be an open subset of the Banach space X. The functional
f : S ⊂ X → R is said to be Gateaux differentiable (G-diff) at a point u ∈ S if there
exists a functional g ∈ X* (often denoted by f 0 (u)) such that
d f (u + tv) − f (u)
f (u + tv) = lim = [f 0 (u)]v for all v ∈ X.
dt t=0 t→0 t
The functional f 0 (u) is called the Gateaux derivative of f at the point u ∈ S. If f is
G-diff at each point of S, the map f 0 : S ⊂ X → X* is called the Gateaux derivative of
f on S. In addition, if f 0 is continuous at u (in the operator norm), then we say that f
is C 1 at u. Note that in the case of a real-valued function of several real variables, the
Gateaux derivative is nothing more than the directional derivative of the function at u in
the direction v.
Let X, Y be Banach spaces and let A : S ⊂ X → Y be an arbitrary operator. A is

said to be Frechet differentiable (F-diff) at the point u ∈ S if there exists an operator
14 1. Preliminaries
B ∈ B(X, Y ) such that

lim kA(u + v) − Au − Bvk/kvk = 0.
kvk→0
The operator B, often denoted by A0 (u), is called the Frechet derivative of A at u. Note
that if A is Frechet differentiable on S, then A0 : S → B(X, Y ). In addition, if A0 is
continuous at u (in the operator norm), we say that A is C 1 at u.
Remark. If the functional f is F-diff at u ∈ S, then it is also G-diff at u, and the two
derivatives are equal. This follows easily from the definition of the Frechet derivative. The
converse is not always true as may be easily verified by simple examples from several variable
calculus. However, if the Gateaux derivative exists in a neighborhood of u and if f ∈ C 1 at
u, then the Frechet derivative exists at u, and the two derivatives are equal.
Example 1.31. (a) Let f (ξ) ∈ C(R). Then for k ≥ 0, the corresponding Nemytskii operator
N : C k (Ω̄) → C(Ω̄) is bounded and continuous. If in addition f (ξ) ∈ C 1 (R), then N ∈ C 1
and the Frechet derivative N 0 (u) is given by
[N 0 (u)v](x) = f 0 (u(x))v(x).
Note that for u, v ∈ C k (Ω̄), |N 0 (u)v|0 ≤ |f 0 (u)|0 |v|k and so N 0 (u) ∈ B(C k (Ω̄), C ( Ω̄)) with
kN 0 (u)k ≤ |f 0 (u)|0 . Clearly N 0 (u) is continuous at each point u ∈ C k (Ω̄). Moreover,
Z 1
0 d
|N (u + v) − N u − N (u)v|0 = sup | [ f (u(x) + tv(x)) − f 0 (u(x))v(x)]dt|
x 0 dt
Z 1
≤ |v|0 sup |f 0 (u(x) + tv(x)) − f 0 (u(x))|dt.
x 0
The last integral tends to zero since f0 is uniformly continuous on compact subsets of R.
More generally, let f (ξ) ∈ C k (R). Then the corresponding Nemytskii operator N :
C k (Ω̄)
→ C k (Ω̄) is bounded and continuous. If in addition f (ξ) ∈ C k+1 (R), then N ∈ C 1
with Frechet derivative given by [N 0 (u)v](x) = f 0 (u(x))v(x). Note that |uv|k ≤ |u|k |v|k for
u, v ∈ C k (Ω̄), and since C k (Ω̄) ⊂ C(Ω̄), the Frechet derivative must be of the stated form.
(b) Let f (ξ) ∈ C k+1 (R), where k > n/2. Then we claim that the corresponding Ne-
mytskii operator N : H k (Ω) → H k (Ω) is of class C 1 with Frechet derivative given by
[N 0 (u)v](x) = f 0 (u(x))v(x).
First, suppose u ∈ C k (Ω̄). Then N (u) ∈ C k (Ω̄) by the usual chain rule. If u ∈ H k (Ω),
let um ∈ C k (Ω̄) with kum − ukk,2 → 0. Since the imbedding H k (Ω) ⊂ C(Ω̄) is continuous,
um → u uniformly, and thus f (um ) → f (u) and f 0 (um ) → f 0 (u) uniformly and hence in L2 .
Furthermore, Di f (um ) = f 0 (um )Di um → f 0 (u)Di u in L1 . Consequently, by Theorem 2.11,
we have
Di f (u) = f 0 (u)Di u.
In a similar fashion we find
Dij f (u) = f 00 (u)Di uDj u + f 0 (u)Dij u
with corresponding formulas for higher derivatives.
1.5.4. Implicit Function Theorem. The following lemmas are needed in the proof of
the implicit function theorem.
Lemma 1.32. Let S be a closed nonempty subset of the Banach space X and let M be a
metric space. Suppose A(x, λ) : S × M → S is continuous and there is a constant k < 1
such that, uniformly for all λ ∈ M
kA(x, λ) − A(y, λ)k ≤ kkx − yk for all x, y ∈ S.
Then for each λ ∈ M, A(x, λ) has a unique fixed point x(λ) ∈ S and moreover, x(λ)
depends continuously on λ.
Proof. The existence and uniqueness of the fixed point x(λ) is of course a consequence of
the contraction mapping theorem. To prove continuity, suppose λn → λ. Then
kx(λn ) − x(λ)k = kA(x(λn ), λn ) − A(x(λ), λ)k
≤ kA(x(λn ), λn ) − A(x(λ), λn )k + kA(x(λ), λn ) − A(x(λ), λ)k
≤ kkx(λn ) − x(λ)k + kA(x(λ), λn ) − A(x(λ), λ)k.
Therefore
1
kx(λn ) − x(λ)k ≤ kA(x(λ), λn ) − A(x(λ), λ)k.
1−k
By the assumed continuity of A, the right side tends to zero as n → ∞, and therefore
x(λn ) → x(λ).
Lemma 1.33. Suppose X, Y are Banach spaces. Let S ⊂ X be convex and assume A :
S → Y is Frechet differentiable at every point of S. Then
kAu − Avk ≤ ku − vk sup kA0 (w)k.
w∈S
In other words, A satisfies a Lipschitz condition with constant q = supw∈S kA0 (w)k.
Proof. For fixed u, v ∈ S, set g(t) = A(u + t(v − u)), where t ∈ [0, 1]. Using the definition
of Frechet derivative, we have

0 A(u + (t + h)(v − u)) − A(u + t(v − u))
g (t) = lim
h→0 h
0
hA (u + t(v − u))(v − u) + kh(v − u)kE
= lim
h→0 h
0
= A (u + t(v − u))(v − u).
Hence
kg(0) − g(1)k = kAu − Avk ≤ sup kg 0 (t)k
t∈[0,1]
which implies the desired result.
Lemma 1.34. Let X be a Banach space. Suppose A : B(u0 , r) ⊂ X → X is a contraction,
with Lipschitz constant q < 1, where
r ≥ (1 − q)−1 kAu0 − u0 k.
Then A has a unique fixed point u ∈ B(u0 , r).
Proof. For u ∈ B(u0 , r)

kAu − u0 k ≤ kAu − Au0 k + kAu0 − u0 k ≤ qku − u0 k + (1 − q)r.
Since ku − u0 k ≤ r, A maps the ball B(u0 , r) into itself, and the result follows from the
contraction mapping theorem.
16 1. Preliminaries
We now consider operator equations of the form A(u, v) = 0, where A maps a subset
of X × Y into Z. For a given [u0 , v0 ] ∈ X × Y we denote the Frechet derivative of A (at
[u0 , v0 ]) with respect to the first (second) argument by Au (u0 , v0 ) (Av (u0 , v0 )).
Theorem 1.35. (Implicit Function) Let X, Y, Z be Banach spaces. For a given [u0 , v0 ] ∈
X × Y and a, b > 0, let S = {[u, v] : ku − u0 k ≤ a, kv − v0 k ≤ b}. Suppose A : S → Z
satisfies the following:
(i) A is continuous.
(ii) Av (·, ·) exists and is continuous in S (in the operator norm)
(iii) A(u0 , v0 ) = 0.
(iv) [Av (u0 , v0 )]−1 exists and belongs to B(Z, Y ).
Then there are neighborhoods U of u0 and V of v0 such that the equation A(u, v) = 0 has
exactly one solution v ∈ V for every u ∈ U . The solution v depends continuously on u.
Proof. If in S we define
B(u, v) = v − [Av (u0 , v0 )]−1 A(u, v)
it is clear that the solutions of A(u, v) = 0 and v = B(u, v) are identical. The theorem will
be proved by applying the contraction mapping theorem to B. Since
Bv (u, v) = I − [Av (u0 , v0 )]−1 Av (u, v)
Bv (·, ·) is continuous in the operator norm. Now Bv (u0 , v0 ) = 0, so for some δ > 0 there is
a q < 1 such that
kBv (u, v)k ≤ q
for ku − u0 k ≤ δ, kv − v0 k ≤ δ. By virtue of Lemma 1.33, B(u, ·) is a contraction. Since
A is continuous, B is also continuous. Therefore, since B(u0 , v0 ) = v0 , there is an ε with
0 < ε ≤ δ such that
kB(u, v0 ) − v0 k ≤ (1 − q)δ
for ku − u0 k ≤ ε. The existence of a unique fixed point in the closed ball B(v0 , δ) follows
from Lemma 1.34 and the continuity from Lemma 1.32.
Example 1.36. Let f (ξ) ∈ C 1,α (R), f (0) = f 0 (0) = 0, g(x) ∈ C α (Ω̄) and consider the
boundary value problem
(1.13) ∆u + f (u) = g(x) in Ω, u|∂Ω = 0.
Set X = Z = C α (Ω̄), Y = {u ∈ C 2,α (Ω̄) : u|∂Ω = 0} and
A(g, u) = ∆u + N (u) − g
where N is the Nemytskii operator corresponding to f . The operator A maps X × Y into
the space Z. Clearly A(0, 0) = 0 (A is C 1 by earlier examples) and
Au (0, 0)v = ∆v, v ∈ Y.
It is easily checked that all the conditions of the implicit function theorem are met. In
particular, condition (iv) is a consequence of the bounded inverse theorem. Thus, for a
function g ∈ C α (Ω̄) of sufficiently small norm (in the space C α (Ω̄)) there exists a unique
solution of (1.13) which lies near the zero function. There may, of course, be other solutions
which are not close to the zero function. (Note that the condition f 0 (0) = 0 rules out linear
functions.)
Remark. Note that the choice of X = Z = C(Ω̄), Y = {u ∈ C 2 (Ω̄) : u|∂Ω = 0} would fail
above since the corresponding linear problem is not onto. An alternate approach would be
to use Sobolev spaces. In fact, if we take X = Z = W k−2 (Ω), Y = W k (Ω) ∩ H01 (Ω) with k
sufficiently large, and if f (ξ) ∈ C k+1 (R), then as above, we can conclude the existence of a
unique solution u ∈ W k (Ω) provided kgkk−2,2 is sufficiently small. Hence, we get existence
for more general functions g; however, the solution u ∈ W k (Ω) is not a classical (i.e., C 2 )
solution in general.
1.5.5. Generalized Weierstrass Theorem. In its simplest form, the classical Weier-
strass theorem can be stated as follows: Every continuous function defined on a closed ball
in Rn is bounded and attains both its maximum and minimum on this ball. The proof
makes essential use of the fact that the closed ball is compact.
The first difficulty in trying to extend this result to an arbitrary Banach space X is
that the closed ball in X is not compact if X is infinite dimensional. However, as we shall
show, a generalized Weierstrass theorem is possible if we require a stronger property for the
functional.
A set S ⊂ X is said to be weakly closed if {un } ⊂ S, un * u implies u ∈ S, i.e.,
S contains all its weak limits. A weakly closed set is clearly closed, but not conversely.
Indeed, the set {sin nx}∞ 2
1 in L (0, π) has no limit point (because it cannot be Cauchy) so
it is closed, but zero is a weak limit that does not belong to the set. It can be shown that
every convex, closed set in a Banach space is weakly closed.
A functional f : S ⊂ X → R is weakly continuous at u0 ∈ S if for every sequence
{un } ⊂ S with un * u0 it follows that f (un ) → f (u0 ). Clearly, every functional f ∈ X*
is weakly continuous. A functional f : S ⊂ X → R is weakly lower semicontinu-
ous(w.l.s.c.) at u0 ∈ S if for every sequence {un } ⊂ S for which un * u0 it follows that
f (u0 ) ≤ lim inf n→∞ f (un ). According to Theorem 1.18, the norm on a Banach space is
w.l.s.c.. A functional f : S ⊂ X → R is weakly coercive on S if f (u) → ∞ as kuk → ∞
on S.
Theorem 1.37. Let X be a reflexive Banach space and f : C ⊂ X → R be w.l.s.c. and
assume
(i) C is a nonempty bounded weakly closed set in X or
(ii) C is a nonempty weakly closed set in X and f is weakly coercive on C.
Then
(a) inf u∈C f (u) > −∞;
(b) there is at least one u0 ∈ C such that f (u0 ) = inf u∈C f (u).
Moreover, if u0 is an interior point of C and f is G-diff at u0 , then f 0 (u0 ) = 0.
Proof. Assume (i) and let {un } ⊂ C be a minimizing sequence, i.e., limn→∞ f (un ) =
inf u∈C f (u). The existence of such a sequence follows from the definition of inf. Since X is
reflexive and C is bounded and weakly closed, there is a subsequence {un0 } and a u0 ∈ C
such that un0 * u0 . But f is w.l.s.c. and so f (u0 ) ≤ lim inf n→∞ f (un0 ) = inf u∈C f (u),
which proves (a). Since by definition, f (u0 ) ≥ inf u∈C f (u), we get (b).
Assume (ii) and fix u0 ∈ C. Since f is weakly coercive, there is a closed ball B(0, R) ⊂ X
such that u0 ∈ B ∩ C and f (u) ≥ f (u0 ) outside B ∩ C. Since B ∩ C satisfies the conditions
of (i), there is a u1 ∈ B ∩ C such that f (u) ≥ f (u1 ) for all u ∈ B ∩ C and in particular for
u0 . Thus, f (u) ≥ f (u1 ) on all of C.
18 1. Preliminaries
To prove the last statement we set ϕv (t) = f (u0 + tv). For fixed v ∈ X, ϕv (t) has a
local minimum at t = 0, and therefore hf 0 (u0 ), vi = 0 for all v ∈ X.
The point u0 ∈ X is called a critical point of the functional f defined on X if f 0 (u0 )v =

0 for every v ∈ X.
Remark. Even though weakly continuous functionals on closed balls attain both their inf
and sup (which follows from the above theorem), the usual functionals that we encounter
are not weakly continuous, but are w.l.s.c.. Hence this explains why we seek the inf and
not the sup in variational problems.
1.5.5.1. Convex sets. A set C in the real normed space X is called convex if (1−t)u+tv ∈ C
for all t ∈ [0, 1], u, v ∈ C. The following result is needed later (see, e.g., [?]).
Theorem 1.38. A closed convex set in a Banach space is weakly closed.
1.5.6. Monotone Operators and Convex Functionals. Let A : X → X* be an oper-

ator on the real Banach space X.
A is monotone if
hAu − Av, u − vi ≥ 0 for all u, v ∈ X.
A is strongly monotone if
hAu − Av, u − vi ≥ cku − vkpX for all u, v ∈ X
where c > 0 and p > 1.
A is coercive if
hAu, ui
lim = +∞.
kuk→∞ kuk
Remark. A strongly monotone operator is coercive. This follows immediately from hAu, ui =
hAu − A0, ui + hA0, ui ≥ ckukpX − kA0kkukX .
Let C be a convex set in the real normed space X. A functional f : C ⊂ X → R is said

to be convex if
f ((1 − t)u + tv) ≤ (1 − t)f (u) + tf (v) for all t ∈ [0, 1], u, v ∈ C.
In the following we set
ϕ(t) = f ((1 − t)u + tv) = f (u + t(v − u))
for fixed u and v.
Lemma 1.39. Let C ⊂ X be a convex set in a real normed space X. Then the following
statements are equivalent:
(a) The real function ϕ : [0, 1] → R is convex for all u, v ∈ C.
(b) The functional f : C ⊂ X → R is convex.
(c) f 0 : C ⊂ X → X* (assuming f is G-diff on C) is monotone.
Proof. Assume ϕ is convex. Then

ϕ(t) = ϕ((1 − t) · 0 + t · 1) ≤ (1 − t)ϕ(0) + tϕ(1)
for all t ∈ [0, 1], which implies (b).
Similarly, if f is convex, then for t = (1 − α)s1 + αs2 , with α, s1 , s2 ∈ [0, 1], we have
ϕ(t) = f (u + t(v − u)) ≤ (1 − α)f (u + s1 (v − u)) + αf (u + s2 (v − u))
for all u, v ∈ C, which implies (a).
Fix u, v ∈ C. Then ϕ0 (t) = hf 0 (u + t(v − u)), v − ui. If f is convex, then ϕ is convex
and therefore ϕ0 is monotone. From ϕ0 (1) ≥ ϕ0 (0) we obtain
hf 0 (v) − f 0 (u), v − ui ≥ 0 for all u, v ∈ C
which implies (c).
Finally, assume f 0 is monotone. Then for s < t we have
1
ϕ0 (t) − ϕ0 (s) = hf 0 (u + t(v − u)) − f 0 (u + s(v − u)), (t − s)(v − u)i ≥ 0.
t−s
Thus ϕ0 is monotone, which implies ϕ, and thus f is convex.
Theorem 1.40. Consider the functional f : C ⊂ X → R, where X is a real Banach space.
Then f is w.l.s.c. if any one of the following conditions holds:
(a) C is closed and convex; f is convex and continuous.
(b) C is convex; f is G-diff on C and f 0 is monotone on C.
Proof. Set
Cr = {u ∈ C : f (u) ≤ r}.
It follows from (a) that Cr is closed and convex for all r, and thus is weakly closed (cf.
Theorem 1.38). If f is not w.l.s.c., then there is a sequence {un } ⊂ C with un * u and
f (u) > lim inf f (un ). Hence, there is an r and a subsequence {un0 } such that f (u) > r and
f (un0 ) ≤ r (i.e., un0 ∈ Cr ) for all n0 large enough. Since Cr is weakly closed, u ∈ Cr , which
is a contradiction.
Assume (b) holds and set ϕ(t) = f (u + t(v − u)). Then by Lemma 1.39, ϕ : [0, 1] → R
is convex and ϕ0 is monotone. By the classical mean value theorem,
ϕ(1) − ϕ(0) = ϕ0 (θ) ≥ ϕ0 (0), 0<θ<1
i.e.,
f (v) ≥ f (u) + hf 0 (u), v − ui for all u, v ∈ C.
If un * u, then hf 0 (u), un − ui → 0 as n → ∞. Hence, f is w.l.s.c.
1.5.7. Lagrange Multipliers. Let f, g : X → R be two functionals defined on the Banach

space X and let
Mc = {u ∈ X : g(u) = c}
for a given constant c. A point u0 ∈ Mc is called an extreme of f with respect to Mc if
there exists a neighborhood of u0 , U (u0 ) ⊂ X, such that
f (u) ≤ f (u0 ) for all u ∈ U (u0 ) ∩ Mc
or
f (u) ≥ f (u0 ) for all u ∈ U (u0 ) ∩ Mc .
20 1. Preliminaries
In the first case we say that f is (local) maximal at u0 with respect to Mc , while in the
second case f is (local) minimal at u0 with respect to Mc . A point u0 ∈ Mc is called an
ordinary point of the manifold Mc if its F-derivative g 0 (u0 ) 6= 0.
Let u0 be an ordinary point of Mc . Then u0 is called a critical point of f with respect
to Mc if there exists a real number λ, called a Lagrange multiplier, such that
f 0 (u0 ) = λg 0 (u0 ).
As we shall see, if u0 is an extremum of f with respect to Mc , and if u0 is an ordinary point,
then u0 is a critical point of f with respect to Mc . Note that if u0 is an extremum of f with
respect to X, then we can choose λ = 0, which implies the usual result.
Lemma 1.41. Let X be a Banach space. Suppose the following hold:
(i) f, g : X → R are of class C 1
(ii) For u0 ∈ X, we can find v, w ∈ X such that
(1.14) f 0 (u0 )v · g 0 (u0 )w 6= f 0 (u0 )w · g 0 (u0 )v.

Then f cannot have a local extremum with respect to the level set Mc at u0 .
Proof. Fix v, w ∈ X, and for s, t ∈ R consider the real-valued functions

F (s, t) = f (u0 + sv + tw), G(s, t) = g(u0 + sv + tw) − c.
Then
∂F ∂F
(0, 0) = f 0 (u0 )v, (0, 0) = f 0 (u0 )w
∂s ∂t
∂G ∂G
(0, 0) = g 0 (u0 )v, (0, 0) = g 0 (u0 )w
∂s ∂t
so that condition (1.14) is simply that the Jacobian |∂(F, G)/∂(s, t)| is nonvanishing at
(s, t) = (0, 0). Since F, G ∈ C 1 on R2 , we may apply the implicit function theorem to
conclude that a local extremum cannot occur at u0 . More precisely, assume w.l.o.g. that
Gt (0, 0) 6= 0. Since G(0, 0) = 0, the implicit function theorem implies the existence of a C 1
function φ such that φ(0) = 0 and G(s, φ(s)) = 0 for sufficiently small s. Moreover,
Gs (0, 0)
φ0 (0) = − .
Gt (0, 0)
Set z(s) = F (s, φ(s)) = f (u0 + sv + φ(s)w) and note that g(u0 + sv + φ(s)w) = c. Hence,
if to the contrary f has an extremum at u0 , then z(s) has a local extremum at s = 0. But,
an easy computation shows that Gt (0, 0)z 0 (0) = f 0 (u0 )v · g 0 (u0 )w − f 0 (u0 )w · g 0 (u0 )v 6= 0,
which is a contradiction.
Theorem 1.42. (Lagrange) Let X be a Banach space. Suppose the following hold:
(i) f, g : X → R are of class C 1
(ii) g(u0 ) = c.
(iii) u0 is a local extremum of f with respect to the constraint Mc
Then either
(a) g 0 (u0 )v = 0 for all v ∈ X, or
(b) There exists λ ∈ R such that f 0 (u0 )v = λg 0 (u0 )v for all v ∈ X.
Proof. If (a) does not hold, then fix w ∈ X with g 0 (u0 )w 6= 0. By hypothesis and the above
lemma, we must have
f 0 (u0 )v · g 0 (u0 )w = f 0 (u0 )w · g 0 (u0 )v for all v ∈ X.
If we define λ = (f 0 (u0 )w)/(g 0 (u0 )w), then we obtain (b).
More generally, one can prove the following:

Theorem 1.43. (Ljusternik) Let X be a Banach space. Suppose the following hold:
(i) g0 : X → R is of class C 1
(ii) gi : X → R are of class C 1 , i = 1, . . . , n
(iii) u0 is an extremum of g0 with respect to the constraint C:
C = {u : gi (u) = ci (i = 1, . . . , n)}
where the ci are constants.
Then there are numbers λi (not all zero) such that
n
X
(1.15) λi gi0 (u0 ) = 0.
i=0
As an application of Ljusternik’s theorem we have

Theorem 1.44. Let f, g : X → R be C 1 functionals on the reflexive Banach space X.
Suppose
(i) f is w.l.s.c. and weakly coercive on X ∩ {g(u) ≤ c}
(ii) g is weakly continuous
(iii) g(0) = 0, g 0 (u) = 0 only at u = 0.
Then the equation f 0 (u) = λg 0 (u) has a one parameter family of nontrivial solutions (uR , λR )
for all R 6= 0 in the range of g(u) and g(uR ) = R. Moreover, uR can be characterized as
the function which minimizes f (u) over the set g(u) = R.
Proof. Since g(u) is weakly continuous, it follows that MR = {u : g(u) = R} is weakly

closed. If MR is not empty, i.e., if R belongs to the range of g, then by Theorem 1.37,
there is a uR ∈ MR such that f (uR ) = inf f (u) over u ∈ MR . If R 6= 0 then it cannot
be that g 0 (uR ) = 0. Otherwise by (iii), uR = 0 and hence R = g(uR ) = 0, which is a
contradiction. Thus, by Ljusternik’s theorem, there exist constants λ1 , λ2 , λ21 + λ22 6= 0 such
that λ1 f 0 (uR ) + λ2 g 0 (uR ) = 0. Since uR is an ordinary point, it follows that λ1 6= 0, and
therefore λR = −λ2 /λ1 .
Remark. In applying this theorem one should be careful and not choose g(u) = kuk, since
this g is not weakly continuous.
The following interpolation inequality, which is frequently referred to as Ehrling’s

inequality, will be needed in the next result.
Theorem 1.45. Let X, Y, Z be three Banach spaces such that
X ⊂ Y ⊂ Z.
22 1. Preliminaries
Assume that the imbedding X ⊂ Y is compact and that the imbedding Y ⊂ Z is continuous.
Then for each ε > 0, there is a constant c(ε) such that
(1.16) kukY ≤ εkukX + c(ε)kukZ for all u ∈ X.
Proof. If for a fixed ε > 0 the inequality is false, then there exists a sequence {un } such
that
(1.17) kun kY > εkun kX + nkun kZ for all n.
Without loss of generality we can assume kun kX = 1. Since the imbedding X ⊂ Y is
compact, there is a subsequence, again denoted by {un }, with un → u in Y . This implies
un → u in Z. By (1.17), kun kY > ε and so u 6= 0. Again by (1.17), un → 0 in Z, i.e., u = 0,
which is a contradiction.
Chapter 2
Sobolev Spaces
This chapter is devoted to the study of the necessary Sobolev function spaces which permit
a modern approach to partial differential equations.
2.1. Weak Derivatives and Sobolev Spaces

2.1.1. Weak Derivatives. Let Ω be a nonempty open set in Rn . Suppose u ∈ C m (Ω) and
ϕ ∈ C0m (Ω). Then by integration by parts
Z Z
α |α|
(2.1) uD ϕdx = (−1) vϕdx, |α| ≤ m
Ω Ω
where v = Dα u. Motivated by (2.1), we now enlarge the class of functions for which the
notion of derivative can introduced.
Let u ∈ L1loc (Ω). A function v ∈ L1loc (Ω) is called the αth weak derivative of u if it
satisfies
Z Z
α |α| |α|
(2.2) uD ϕdx = (−1) vϕdx for all ϕ ∈ C0 (Ω).
Ω Ω
It can be easily shown that the weak derivative is unique. Thus we write v = Dα u to indicate
that v is the αth weak derivative of u. If a function u has an ordinary αth derivative lying
in L1loc (Ω), then it is clearly the αth weak derivative.
In contrast to the corresponding classical derivative, the weak derivative Dα u is defined

globally on all of Ω by (2.2). However, in every subregion Ω0 ⊂ Ω the function Dα u will
also be the weak derivative of u. It suffices to note that (2.2) holds for every function
|α|
ϕ ∈ C0 (Ω0 ), and extended outside Ω0 by assigning to it the value zero. In particular, the
weak derivative (if it exists) of a function u having compact support in Ω has itself compact
support in Ω and thus belongs to L1 (Ω).
We also note that in contrast to the classical derivative, the weak derivative Dα u is
defined at once for order |α| without assuming the existence of corresponding derivatives
of lower orders. In fact, the derivatives of lower orders may not exist as we will see in a
forthcoming exercise. However, it can be shown that if all weak derivatives exist of a certain
order, then all lower order weak derivatives exist.
23
24 2. Sobolev Spaces
Example 2.1. (a) The function u(x) = |x1 | has in the ball Ω = B(0, 1) weak derivatives
ux1 = sgn x1 , uxi = 0, i = 2, . . . , n. In fact, we apply formula (2.2) as follows: For any
ϕ ∈ C01 (Ω)
Z Z Z
|x1 |ϕx1 dx = x1 ϕx1 dx − x1 ϕx1 dx
Ω Ω+ Ω−
where Ω+ = Ω ∩ (x1 > 0), Ω− = Ω ∩ (x1 < 0). Since x1 ϕ = 0 on ∂Ω and also for x1 = 0,
an application of the divergence theorem yields
Z Z Z Z
|x1 |ϕx1 dx = − ϕdx + ϕdx = − (sgn x1 )ϕdx.
Ω Ω+ Ω− Ω
Hence |x1 |x1 = sgn x1 . Similarly, since for i ≥ 2

Z Z Z
|x1 |ϕxi dx = (|x1 |ϕ)xi dx = − 0ϕdx
Ω Ω Ω
|x1 |xi = 0 for i = 2, . . . , n. Note that the function |x1 | has no classical derivative with
respect to x1 in Ω.
(b) By the above computation, the function u(x) = |x| has a weak derivative u0 (x) =
sgn x on the interval Ω = (−1, 1). On the other hand, sgn x does not have a weak derivative
on Ω due to the discontinuity at x = 0.
(c) Let Ω = B(0, 1/2) ⊂ R2 and define u(x) = ln(ln(2/r)), x ∈ Ω, where r = |x| =
+ x22 )1/2 . Then u 6∈ L∞ (Ω) because of the singularity at the origin. However, we will
(x21
show that u has weak first partial derivatives.
First of all u ∈ L2 (Ω), for
Z Z 2π Z 1/2
2
|u| dx = r[ln(ln(2/r))]2 dr dθ
Ω 0 0
and a simple application of L’hopitals rule shows that the integrand is bounded and thus
the integral is finite. Similarly, it is easy to check that the classical partial derivative
− cos θ
ux1 = , where x1 = r cos θ
r ln(2/r)
also belongs to L2 (Ω). Now we show that the defining equation for the weak derivative is
met.
Let Ωε = {x : ε < r < 1/2} and choose ϕ ∈ C01 (Ω). Then by the divergence theorem
and the absolute continuity of integrals
Z Z Z Z
uϕx1 dx = lim uϕx1 dx = lim − ux1 ϕdx + uϕn1 ds
Ω ε→0 Ωε ε→0 Ωε r=ε
where n = (n1 , n2 ) is the unit outward normal to Ωε on r = ε. But (ds = εdθ)

Z Z 2π
| uϕn1 ds| ≤ |u| |ϕ|εdθ ≤ 2πεc ln(ln(2/ε)) → 0
r=ε 0
as ε → 0. Thus Z Z
uϕx1 dx = − ux1 ϕdx.
Ω Ω
The same analysis applies to ux2 . Thus u has weak first partial derivatives given by the
classical derivatives which are defined on Ω\{0}.
2.1. Weak Derivatives and Sobolev Spaces 25
2.1.2. Sobolev Spaces. For p ≥ 1 and k a nonnegative integer, we let

W k,p (Ω) = {u : u ∈ Lp (Ω), Dα u ∈ Lp (Ω), 0 < |α| ≤ k}
where Dα u denotes the αth weak derivative. When k = 0, W k,p (Ω) will mean Lp (Ω). It is
clear that W k,p (Ω) is a vector space. A norm on W k,p (Ω) is introduced by defining
(R P
( |Dα u|p dx)1/p if 1 ≤ p < ∞,
(2.3) kukk,p = kukW k,p (Ω) = PΩ |α|≤kα
|α|≤k kD ukL∞ (Ω) if p = ∞.
The space W k,p (Ω) is known as a Sobolev space of order k and power p.
We define the space W0k,p (Ω) to be the closure of the space C0k (Ω) with respect to the
norm k · kk,p . As we shall see shortly, W k,p (Ω) 6= W0k,p (Ω) for k ≥ 1. (Unless Ω = Rn .)
Remark 2.1. The spaces W k,2 (Ω) and W0k,2 (Ω) are special since they become a Hilbert
space under the inner product
Z X
(u, v)k,2 = (u, v)W k,2 (Ω) = Dα uDα vdx.
Ω |α|≤k
Since we shall be dealing mostly with these spaces in the sequel, we introduce the special
notation:
H k (Ω) = W k,2 (Ω), H0k (Ω) = W0k,2 (Ω).
Theorem 2.2. W k,p (Ω) is a Banach space under the norm (2.3). If 1 < p < ∞, it is
reflexive, and if 1 ≤ p < ∞, it is separable.
Proof. 1. We first prove that W k,p (Ω) is complete with respect to the norm (2.3). We
prove this for 1 ≤ p < ∞; the case p = ∞ is similar. Let {un } be a Cauchy sequence of
elements in W k,p (Ω), i.e.,
p
X Z
kun − um kk,p = |Dα un − Dα um |p dx → 0 as m, n → ∞.
|α|≤k Ω
Then for any α, |α| ≤ k, when m, n → ∞

Z
|Dα un − Dα um |p dx → 0
Ω
and, in particular, when |α| = 0
Z
|un − um |p dx → 0.
Ω
Since Lp (Ω) is complete, it follows that there are functions uα ∈ Lp (Ω), |α| ≤ k such that
D un → u (in Lp (Ω)). Since each un (x) has weak derivatives (up to order k) belonging
α α
to Lp (Ω), a simple limit argument shows that uα is the αth weak derivative of u0 . In fact,
Z Z Z Z
α α |α| α |α|
uD ϕdx ← un D ϕdx = (−1) ϕD un dx → (−1) uα ϕdx.
Ω Ω Ω Ω
Hence u0
∈ W k,p (Ω)
and kun − u0 k
k,p → 0 as n → ∞. This proves the completeness of
k,p
W (Ω); hence it is a Banach space.
2. Consider the map T : W 1,p (Ω) → (Lp (Ω))n+1 defined by
T u = (u, D1 u, . . . , Dn u).
If we endow the latter space with the norm

n+1
X
kvk = ( kvi kpp )1/p
i=1
for v = (v1 , . . . , vn+1 ) ∈ (Lp (Ω))n+1 , then T is a (linear) isometry. Now (Lp (Ω))n+1 is
reflexive for 1 < p < ∞ and separable for 1 ≤ p < ∞. Since W 1,p (Ω) is complete, its image
under the isometry T is a closed subspace of (Lp (Ω))n+1 which inherits the corresponding
properties as does W 1,p (Ω) (see Theorem 1.14). Similarly, we can handle the case k ≥ 2.
The following result is of independent importance. We omit the proof.

Theorem 2.3. un * u in W k,p (Ω) if and only if Dα un * Dα u in Lp (Ω) for all |α| ≤ k.
Example 2.4. Let Ω be a bounded open connected set in Rn . Divide Ω into N open disjoint
subsets Ω1 , Ω2 , . . . , ΩN . Suppose the function u : Ω → R has the following properties:
(i) u is continuous on Ω̄.
(ii) For some i, Di u is continuous on Ω1 , Ω2 , . . . , ΩN , and can be extended contin-
uously to Ω̄1 , Ω̄2 , . . . , Ω̄N , respectively.
(iii) The surfaces of discontinuity are such that the divergence theorem applies.
Define wi (x) = Di u(x) if x ∈ ∪Ni=1 Ωi . Otherwise, wi can be arbitrary. We now claim that
wi ∈ Lp (Ω) is a weak partial derivative of u on Ω. Indeed, for all ϕ ∈ C01 (Ω), the divergence
theorem yields
Z XZ
uDi ϕdx = uDi ϕdx
Ω j Ωj
!
X Z Z
= uϕni dS − ϕDi udx
j ∂Ωj Ωj
Z
= − ϕDi udx.
Ω
Note that the boundary terms either vanish, since ϕ has compact support, or cancel out
along the common boundaries, since u is continuous and the outer normals have opposite
directions. Similarly, if u ∈ C k (Ω̄) and has piecewise continuous derivatives in Ω of order
k + 1, then u ∈ W k+1,p (Ω).
Remark 2.2. More generally, by using a partition of unity argument, we can show the
following: If O is a collection of nonempty open sets whose union is Ω and if u ∈ L1loc (Ω)
is such that for some multi-index α, the αth weak derivative of u exists on each member of
O, then the αth weak derivative of u exists on Ω.
Exercise 2.3. (a) Consider the function u(x) = sgnx1 + sgnx2 in the ball B(0, 1) ⊂ R2 .
Show that the weak derivative ux1 does not exist, yet the weak derivative ux1 x2 does exist.
(b) Let Ω be the hemisphere of radius R < 1 in Rn :
n
X
2
r ≡ x2i ≤ R2 , xn ≥ 0 (n ≥ 3).
i=1
Show that u = (r(n/2−1) ln r)−1 ∈ H 1 (Ω).

2.2. Approximations and Extensions 27
(c) Let B = B(0, 1) be the open unit ball in Rn , and let

u(x) = |x|−α , x ∈ B.
For what values of α, n, p does u belong to W 1,p (B)?
2.2. Approximations and Extensions

2.2.1. Mollifiers. Let x ∈ Rn and let B(x, h) denote the open ball with center at x and
radius h. For each h > 0, let ωh (x) ∈ C ∞ (Rn ) satisfy
ωh (x) ≥ 0; ωh (x) = 0 for |x| ≥ h
Z Z
ωh (x)dx = ωh (x)dx = 1.
Rn B(0,h)
Such functions are called mollifiers. For example, let
(
k exp [(|x|2 − 1)−1 ], |x| < 1,
ω(x) =
0, |x| ≥ 1,
R
where k > 0 is chosen so that Rn ω(x) dx = 1. Then, a family of mollifiers can be taken as
ωh (x) = h−n ω(x/h) for h > 0.
Let Ω be a nonempty open set in Rn and let u ∈ L1 (Ω). We set u = 0 outside Ω. Define
for each h > 0 the mollified function
Z
uh (x) = ωh (x − y)u(y)dy
Ω
where ωh is a mollifier.
Remark 2.4. There are two other forms in which uh can be represented, namely
Z Z
(2.4) uh (x) = ωh (x − y)u(y)dy = ωh (x − y)u(y)dy
Rn B(x,h)
the latter equality being valid since ωh vanishes outside the (open) ball B(x, h). Thus
the values of uh (x) depend only on the values of u on the ball B(x, h). In particular, if
dist(x, supp(u)) ≥ h, then uh (x) = 0.
Theorem 2.5. Let Ω be a nonempty open set in Rn . Then
(a) uh ∈ C ∞ (Rn ).
(b) If supp(u) is a compact subset of Ω, then uh ∈ C0∞ (Ω) for all h sufficiently
small.
Proof. Since u is integrable and ωh ∈ C ∞ , the Lebesgue theorem on differentiating integrals

implies that for |α| < ∞
Z
α
D uh (x) = u(y)Dα ωh (x − y)dy
Ω
i.e., uh ∈ C ∞ (Rn ). Statement (b) follows from the remark preceding the theorem.
With respect to a bounded set Ω we construct another set Ω(h) as follows: with each
point x ∈ Ω as center, draw a ball of radius h; the union of these balls is then Ω(h) . Clearly
Ω(h) ⊃ Ω. Moreover, uh can be different from zero only in Ω(h) .
Corollary 2.6. Let Ω be a nonempty bounded open set in Rn and let h > 0 be any number.
Then there exists a function η ∈ C ∞ (Rn ) such that
0 ≤ η(x) ≤ 1; η(x) = 1, x ∈ Ω(h) ; η(x) = 0, x ∈ (Ω(3h) )c .
Such a function is called a cut-off function for Ω.
Proof. Let χ(x) be the characteristic function of the set Ω(2h) : χ(x) = 1 for x ∈ Ω(2h) , χ(x) =
0 for x 6∈ Ω(2h) and set
Z
η(x) ≡ χh (x) = ωh (x − y)χ(y)dy.
Rn
Then Z
η(x) = ωh (x − y)dy ∈ C ∞ (Rn ),
Ω(2h)
Z
0 ≤ η(x) ≤ ωh (x − y)dy = 1,
Rn
and
(R
B(x,h) ωh (x − y)dy = 1, x ∈ Ω(h) ,
Z
η(x) = ωh (x − y)χ(y)dy =
B(x,h) 0, x ∈ (Ω(3h) )c .
In particular, we note that if Ω0 ⊂⊂ Ω, there is a function η ∈ C0∞ (Ω) such that η(x) = 1
for x ∈ Ω0 , and 0 ≤ η(x) ≤ 1 in Ω.
Henceforth, the notation Ω0 ⊂⊂ Ω means that Ω0 , Ω are open sets, Ω0 is bounded, and
that Ω0 ⊂ Ω.
We need the following well-known result.
Theorem 2.7. (Partition of Unity) Assume Ω ⊂ Rn is bounded and Ω ⊂⊂ ∪N i=1 Ωi ,
where each Ωi is open. Then there exist C ∞ functions ψi (x) (i = 1, . . . , N ) such that
(a) 0 ≤ ψi (x) ≤ 1
(b) ψi has its support in Ωi
PN
(c) i=1 ψi (x) = 1 for every x ∈ Ω.
2.2.2. Approximation Theorems.

Lemma 2.8. Let Ω be a nonempty bounded open set in Rn . Then every u ∈ Lp (Ω) is
p-mean continuous, i.e.,
Z
|u(x + z) − u(x)|p dx → 0 as z → 0.
Ω
Proof. Choose a > 0 large enough so that Ω is strictly contained in the ball B(0, a). Then
the function
u(x) if x ∈ Ω,
U (x) =
0 if x ∈ B(0, 2a) \ Ω
belongs to Lp (B(0, 2a)). For ε > 0, there is a function Ū ∈ C(B̄(0, 2a)) which satisfies the
inequality kU − Ū kLp (B(0,2a)) < ε/3. By multiplying Ū by an appropriate cut-off function,
it can be assumed that Ū (x) = 0 for x ∈ B(0, 2a) \ B(0, a). Therefore for |z| ≤ a,
kU (x + z) − Ū (x + z)kLp (B(0,2a)) = kU (x) − Ū (x)kLp (B(0,a)) ≤ ε/3.
Since function Ū is uniformly continuous in B(0, 2a), there is a 0 < δ < a such that
kŪ (x + z) − Ū (x)kLp (B(0,2a)) ≤ ε/3 whenever |z| < δ. Hence for |z| < δ we easily see that
ku(x + z) − u(x)kLp (Ω) = kU (x + z) − U (x)kLp (B(0,2a)) ≤ ε.
Theorem 2.9. Let Ω be a nonempty open set in Rn . If u ∈ Lp (Ω) (1 ≤ p < ∞), then
(a) kuh kp ≤ kukp
(b) kuh − ukp → 0 as h → 0.
If u ∈ C k (Ω̄) and Ω̄ is compact, then, for all Ω0 ⊂⊂ Ω,
(c) kuh − ukC k (Ω̄0 ) → 0 as h → 0.
1/p 1/q
Proof. 1. If 1 < p < ∞, let q = p/(p − 1). Then ωh = ωh ωh and Hölder’s inequality
implies
Z Z p/q
p p
|uh (x)| ≤ ωh (x − y)|u(y)| dy ωh (x − y)dy
Ω Ω
Z
≤ ωh (x − y)|u(y)|p dy
Ω
which obviously holds also for p = 1. An application of Fubini’s Theorem gives
Z Z Z Z
p p
|uh (x)| dx ≤ ωh (x − y)dx |u(y)| dy ≤ |u(y)|p dy
Ω Ω Ω Ω
which implies (a).

2. To prove (b), let ω(x) = hn ωh (hx). Then ω(x) ∈ C ∞ (Rn ) and satisfies
ω(x) ≥ 0; ω(x) = 0 for |x| ≥ 1
Z Z
ω(x)dx = ω(x)dx = 1.
Rn B(0,1)
Using the change of variable z = (x − y)/h we have
Z
uh (x) − u(x) = [u(y) − u(x)]ωh (x − y)dy
B(x,h)
Z
= [u(x − hz) − u(x)]ω(z)dz.
B(0,1)
Hence by Hölder’s inequality

Z
p
|uh (x) − u(x)| ≤ d |u(x − hz) − u(x)|p dz
B(0,1)
and so by Fubini’s Theorem

Z Z Z
|uh (x) − u(x)|p dx ≤ d ( |u(x − hz) − u(x)|p dx)dz.
Ω B(0,1) Ω
The right-hand side goes to zero as h → 0 since every u ∈ Lp (Ω) is p-mean continuous.
3. We now prove (c) for k = 0. Let Ω0 , Ω00 be such that Ω0 ⊂⊂ Ω00 ⊂⊂ Ω. Let h0 be the
shortest distance between ∂Ω0 and ∂Ω00 . Take h < h0 . Then
Z
uh (x) − u(x) = [u(y) − u(x)]ωh (x − y)dy.
B(x,h)
If x ∈ Ω̄0 , then in the above integral y ∈ Ω̄00 . Now u is uniformly continuous in Ω̄00 and
ωh ≥ 0, and therefore for an arbitrary ε > 0 we have
Z
|uh (x) − u(x)| ≤ ε ωh (x − y)dy = ε
B(x,h)
provided h is sufficiently small. The case k ≥ 1 is handled similarly and is left as an

exercise.
Remark 2.5. In (c) of the theorem above, we cannot replace Ω0 by Ω. Let u ≡ 1 for x ∈
R1 Rh
[0, 1] and consider uh (x) = 0 ωh (x − y)dy, where ωh (y) = ωh (−y). Now −h ωh (y)dy = 1
and so uh (0) = 1/2 for all h < 1. Thus uh (0) → 1/2 6= 1 = u(0). Moreover, for x ∈ (0, 1)
R x+h
and h sufficiently small, (x − h, x + h) ⊂ (0, 1) and so uh (x) = x−h ωh (x − y)dy = 1 which
implies uh (x) → 1 for all x ∈ (0, 1).
Corollary 2.10. Let Ω be a nonempty open set in Rn . Then C0∞ (Ω) is dense in Lp (Ω) for
all 1 ≤ p < ∞.
Proof. Suppose first that Ω is bounded and let Ω0 ⊂⊂ Ω. For a given u ∈ Lp (Ω) set
u(x), x ∈ Ω0

v(x) =
0, x ∈ Ω\Ω0 .
Then Z Z
|u − v|p dx = |u|p dx.
Ω Ω\ Ω0
By the absolute continuity of integrals, we can choose Ω0 so that the integral on the right
is arbitrarily small, i.e., ku − vkp < ε/2. Since supp(v) is a compact subset of Ω, Theorems
2.5(b) and 2.9(b) imply that for h sufficiently small, vh (x) ∈ C0∞ (Ω) with kv − vh kp < ε/2,
and therefore ku − vh kp < ε. If Ω is unbounded, choose a ball B large enough so that
Z
|u|p dx < ε/2
Ω\Ω0
0
where Ω = Ω ∩ B, and repeat the proof just given.
We now consider the following local approximation theorem.

Theorem 2.11. Let Ω be a nonempty open set in Rn and suppose u, v ∈ L1loc (Ω). Then v =
Dα u iff there exists a sequence of C |α| (Ω) functions {uh } with kuh − ukL1 (S) → 0, kDα uh −
vkL1 (S) → 0 as h → 0, for all compact sets S ⊂ Ω.
Proof. 1. (Necessity) Suppose v = Dα u. Let S ⊂ Ω, and choose d > 0 small enough so

that the sets Ω0 ≡ S (d/2) , Ω00 ≡ S (d) satisfy Ω0 ⊂⊂ Ω00 ⊂⊂ Ω. For x ∈ Rn define
Z Z
uh (x) = ωh (x − y)u(y)dy, vh (x) = ωh (x − y)v(y)dy.
Ω00 Ω00
Clearly, uh , vh ∈∞ n
C (R ) for h > 0. Moreover, from Theorem 2.9 we have kuh − ukL1 (S) ≤
kuh −ukL1 (Ω00 ) → 0. Now we note that if x ∈ Ω0 and 0 < h < d/2, then ωh (x−y) ∈ C0∞ (Ω00 ).
Thus by Theorem 2.5 and the definition of weak derivative,
Z Z
α α |α|
D uh (x) = u(y)Dx ωh (x − y)dy = (−1) u(y)Dyα ωh (x − y)dy
00 00
ZΩ Ω
= ωh (x − y) · v(y)dy = vh (x).
Ω00
Thus, kDα uh − vkL1 (S) → 0.

|α|
2. (Sufficiency) Choose ϕ ∈ C0 (Ω) and consider a compact set S ⊃ supp(ϕ). Then as
h→∞
Z Z Z Z
α α |α| α |α|
uD ϕdx ← uh D ϕdx = (−1) ϕD uh dx → (−1) vϕdx
S S S S
which is the claim.
Theorem 2.12. Let Ω be a domain in Rn . If u ∈ L1loc (Ω) has a weak derivative Dα u = 0
whenever |α| = 1, then u =const. a.e. in Ω.
Proof. Let Ω0 ⊂⊂ Ω. Then for x ∈ Ω0 and with uh as in Theorem 2.11, Dα uh (x) =

(Dα u)h (x) = 0 for all h sufficiently small. Thus uh = const = c(h) in Ω0 for such h. Since
kuh − ukL1 (Ω0 ) = kc(h) − ukL1 (Ω0 ) → 0 as h → 0, it follows that
kc(h1 ) − c(h2 )kL1 (Ω0 ) = |c(h1 ) − c(h2 )|mes(Ω0 ) → 0 as h1 , h2 → 0.
Consequently, c(h) = uh converges uniformly and thus in L1 (Ω0 ) to some constant. Hence
u = const (a.e.) in Ω0 and therefore also in Ω, by virtue of it being connected.
We now note some properties of W k,p (Ω) which follow easily from the results of this
and the previous section.
(a) If Ω0 ⊂ Ω and if u ∈ W k,p (Ω), then u ∈ W k,p (Ω0 ).
(b) If u ∈ W k,p (Ω) and |a(x)|k,∞ < ∞, then au ∈ W k,p (Ω). In this case any weak
derivative Dα (au) is computed according to the usual rule of differentiating the
product of functions.
(c) If u ∈ W k,p (Ω) and uh is its mollified function, then for any compact set S ⊂
Ω, kuh − ukW k,p (S) → 0 as h → 0. If in addition, u has compact support in Ω, then
kuh − ukk,p → 0 as h → 0.
More generally, we have the following global approximation theorems. (See Meyers and
Serrin H = W . The proofs make use of a partition of unity argument.)
Theorem 2.13. Assume Ω is bounded and let u ∈ W k,p (Ω), 1 ≤ p < ∞. Then there exist
functions um ∈ C ∞ (Ω) ∩ W k,p (Ω) such that
um → u in W k,p (Ω).
In other words, C ∞ (Ω) ∩ W k,p (Ω) is dense in W k,p (Ω).
Theorem 2.14. Assume Ω is bounded and ∂Ω ∈ C 1 . Let u ∈ W k,p (Ω), 1 ≤ p < ∞. Then
there exist functions um ∈ C ∞ (Ω̄) such that
um → u in W k,p (Ω).
In other words, C ∞ (Ω̄) is dense in W k,p (Ω).
Exercise 2.6. Prove the product rule for weak derivatives:
Di (uv) = (Di u)v + u(Di v)
where u, Di u are locally Lp (Ω), v, Di v are locally Lq (Ω) (p > 1, 1/p + 1/q = 1).
Exercise 2.7. (a) If u ∈ W0k,p (Ω) and v ∈ C k (Ω̄), prove that uv ∈ W0k,p (Ω).
(b) If u ∈ W k,p (Ω) and v ∈ C0k (Ω), prove that uv ∈ W0k,p (Ω).
2.2.3. Chain Rules.

Theorem 2.15. Let Ω be an open set in Rn . Let f ∈ C 1 (R), |f 0 (s)| ≤ M for all s ∈ R
and suppose u has a weak derivative Dα u for |α| = 1. Then the composite function f ◦ u
has a weak derivative Dα (f ◦ u) = f 0 (u)Dα u. Moreover, if f (0) = 0 and if u ∈ W 1,p (Ω),
then f ◦ u ∈ W 1,p (Ω).
Proof. 1. According to Theorem 2.11, there exists a sequence {uh } ⊂ C 1 (Ω) such that
kuh − ukL1 (Ω0 ) → 0, kDα uh − Dα ukL1 (Ω0 ) → 0 as h → 0, where Ω0 ⊂⊂ Ω. Thus
Z Z
0
|f (uh ) − f (u)|dx ≤ sup |f | |uh − u|dx → 0 as h → 0
Ω0 Ω0
Z Z
0 α 0 α 0
|f (uh )D uh − f (u)D u|dx ≤ sup |f | |Dα uh − Dα u|dx
Ω0 Ω0
Z
+ |f 0 (uh ) − f 0 (u)||Dα u|dx.
Ω0
Since kuh − ukL1 (Ω0 ) → 0, there exists a subsequence of {uh }, which we call {uh } again,
which converges a.e. in Ω0 to u. Moreover, since f 0 is continuous, {f 0 (uh )} converges to f 0 (u)
a.e. in Ω0 . Hence the last integral tends to zero by the dominated convergence theorem.
Consequently, the sequences {f (uh )}, {f 0 (uh )Dα uh } tend to f (u), f 0 (u)Dα u respectively,
and the first conclusion follows by an application of Theorem 2.11 again.
2. If f (0) = 0, the mean value theorem implies |f (s)| ≤ M |s| for all s ∈ R. Thus,
|f (u(x))| ≤ M |u(x)| for all x ∈ Ω and so f ◦ u ∈ Lp (Ω) if u ∈ Lp (Ω). Similarly,
f 0 (u(x))Dα u ∈ Lp (Ω) if u ∈ W 1,p (Ω), which shows that f ◦ u ∈ W 1,p (Ω).
Corollary 2.16. Let Ω be a bounded open set in Rn . If u has an αth weak derivative
Dα u, |α| = 1, then so does |u| and
 Dα u

if u > 0
α
D |u| = 0 if u = 0
−Dα u if u < 0

i.e., Dα |u| = (sgn u)Dα u for u 6= 0. In particular, if u ∈ W 1,p (Ω), then |u| ∈ W 1,p (Ω).
Proof. The positive and negative parts of u are defined by

u+ = max{u, 0}, u− = min{u, 0}.
If we can show that Dα u+ exists and that
α
D u if u > 0
Dα u+ =
0 if u ≤ 0
then the result for |u| follows easily from the relations |u| = u+ − u− and u− = −(−u)+ .
Thus, for h > 0 define
1
(u2 + h2 ) 2 − h if u > 0
fh (u) =
0 if u ≤ 0.
Clearly fh ∈ C 1 (R) and fh0 is bounded on R. By Theorem 2.15, fh (u) has a weak derivative,
and for any ϕ ∈ C01 (Ω)
uDα u
Z Z Z
α α
fh (u)D ϕdx = − D (fh (u))ϕdx = − ϕ 1 dx.
Ω Ω u>0 (u2 + h2 ) 2
Upon letting h → 0, it follows that fh (u) → u+ , and so by the dominating convergence

theorem
Z Z Z
+ α α
u D ϕdx = − ϕD udx = − vϕdx
Ω u>0 Ω
where
Dα u if u > 0

v=
0 if u ≤ 0
which establishes the desired result for u+ .
The next result extends the result on |u|, u+ and u− .
Theorem 2.17. Let f : R → R be Lipschitz continuous with f (0) = 0. Then if Ω is a

bounded open set in Rn , 1 < p < ∞ and u ∈ W01,p (Ω), we have f ◦ u ∈ W01,p (Ω).
Proof. Given u ∈ W01,p (Ω), let un ∈ C01 (Ω) with kun − uk1,p → 0 and define vn = f ◦ un .
Since un has compact support and f (0) = 0, vn has compact support. Also vn is Lipschitz
continuous, for
|vn (x) − vn (y)| = |f (un (x)) − f (un (y))|

≤ c|un (x) − un (y)| ≤ cn |x − y|.
Hence vn ∈ Lp (Ω). Since vn is absolutely continuous on any line segment in Ω, its par-
tial derivatives (which exist almost everywhere) coincide almost everywhere with the weak
derivatives. Moreover, we see from above that |∂vn /∂xi | ≤ cn for 1 ≤ i ≤ n, and as Ω is
bounded, ∂vn /∂xi ∈ Lp (Ω). Thus vn ∈ W 1,p (Ω) and has compact support, which implies
vn ∈ W01,p (Ω). From the relation
|vn (x) − f (u(x))| ≤ c|un (x) − u(x)|
it follows that kvn − f ◦ ukp → 0. Furthermore, if ei is the standard ith basis vector in Rn ,
we have
|vn (x + hei ) − vn (x)| |un (x + hei ) − un (x)|
≤c
|h| |h|
and so
∂vn ∂un
lim sup k kp ≤ c lim sup k kp .
n→∞ ∂xi n→∞ ∂xi
But, {∂un /∂xi } is a convergent sequence in Lp (Ω) and therefore {∂vn /∂xi } is bounded in
Lp (Ω) for each 1 ≤ i ≤ n. Since kvn k1,p is bounded and W01,p (Ω) is reflexive, a subsequence
of {vn } converges weakly in W 1,p (Ω), and thus weakly in Lp (Ω) to some element of W01,p (Ω).
Thus, f ◦ u ∈ W01,p (Ω).
Corollary 2.18. Let u ∈ W01,p (Ω). Then |u|, u+ , u− ∈ W01,p (Ω).
Proof. We apply the preceding theorem with f (t) = |t|. Thus |u| ∈ W01,p (Ω). Now
u+ = (|u| + u)/2 and u− = (u − |u|)/2. Thus u+ , u− ∈ W01,p (Ω).
2.2.4. Extensions. If Ω ⊂ Ω0 , then any function u(x) ∈ C0k (Ω) has an obvious extension
U (x) ∈ C0k (Ω0 ). From the definition of W0k,p (Ω) it follows that the function u(x) ∈ W0k,p (Ω)
and extended as being equal to zero in Ω0 \Ω belongs to W0k,p (Ω0 ). In general, a function
u ∈ W k,p (Ω) and extended by zero to Ω0 will not belong to W k,p (Ω0 ). (Consider the function
u(x) ≡ 1 in Ω.) However, if u ∈ W k,p (Ω) has compact support in Ω, then u ∈ W0k,p (Ω) and
thus the obvious extension belongs to W0k,p (Ω0 ).
We now consider a more general extension result.

Theorem 2.19. Let Ω be a bounded open set in Rn with Ω ⊂⊂ Ω0 and assume k ≥ 1.
(a) If ∂Ω ∈ C k , then any function u(x) ∈ W k,p (Ω) has an extension U (x) ∈
W k,p (Ω0 ) into Ω0 with compact support. Moreover,
kU kW k,p (Ω0 ) ≤ ckukW k,p (Ω)
where the constant c > 0 does not depend on u.
(b) If ∂Ω ∈ C k , then any function u(x) ∈ C k (Ω̄) has an extension U (x) ∈ C0k (Ω0 )
into Ω0 with compact support. Moreover,
kU kC k (Ω̄0 ) ≤ ckukC k (Ω̄) , kU kW k,p (Ω0 ) ≤ ckukW k,p (Ω)
(c) If ∂Ω ∈ C k , then any function u(x) ∈ C k (∂Ω) has an extension U (x) into Ω
which belongs to C k (Ω̄). Moreover
kU kC k (Ω̄) ≤ ckukC k (∂Ω)
Proof. 1. Suppose first that u ∈ C k (Ω̄). Let y = ψ(x) define a C k diffeomorphism that
straightens the boundary near x0 = (x01 , . . . , x0n ) ∈ ∂Ω. In particular, we assume there is a
ball B = B(x0 , r) such that ψ(B ∩ Ω) ⊂ Rn+ (i.e., yn > 0), ψ(B ∩ ∂Ω) ⊂ ∂Rn+ . (e.g., we
could choose yi = xi − x0i for i = 1, . . . , n − 1 and yn = xn − ϕ(x1 , . . . , xn−1 ), where ϕ is of
class C k . Moreover, without loss of generality, we can assume yn > 0 if x ∈ B ∩ Ω.)
2. Let G and G+ = G ∩ Rn+ be respectively, a ball and half-ball in the image of ψ such
that ψ(x0 ) ∈ G. Setting ū(y) = u ◦ ψ −1 (y) and y = (y1 , . . . , yn−1 , yn ) = (y 0 , yn ), we define
an extension Ū (y) of ū(y) into yn < 0 by
k+1
X
Ū (y 0 , yn ) = ci ū(y 0 , −yn /i), yn < 0
i=1
where the ci are constants determined by the system of equations
k+1
X
(2.5) ci (−1/i)m = 1, m = 0, 1, . . . , k.
i=1
Note that the determinant of the system (2.5) is nonzero since it is the Vandemonde deter-
minant. One verifies readily that the extended function Ū is continuous with all derivatives
up to order k in G. For example,
k+1
X
lim Ū (y) = ci ū(y 0 , 0) = ū(y 0 , 0)
y→(y 0 ,0)
i=1
by virtue of (2.5) with m = 0. A similar computation shows that

lim Ūyi (y) = ūyi (y 0 , 0), i = 1, . . . , n − 1.
y→(y 0 ,0)
Finally
k+1
X
lim Ūyn (y) = ci (−1/i)ūyn (y 0 , 0) = ūyn (y 0 , 0)
y→(y 0 ,0)
i=1
by virtue of (2.5) with m = 1. Similarly we can handle the higher derivatives. Thus
w = Ū ◦ ψ ∈ C k (B 0 ) for some ball B 0 = B 0 (x0 ) and w = u in B 0 ∩ Ω, (If x ∈ B 0 ∩ Ω, then
ψ(x) ∈ G+ and w(x) = Ū (ψ(x)) = ū(ψ(x)) = u(ψ −1 ψ(x)) = u(x)) so that w provides a C k
extension of u into Ω ∪ B 0 . Moreover,
sup |ū(y)| = sup |u(ψ −1 (y))| ≤ sup |u(x)|
G+ G+ Ω
and since x ∈ B 0 implies ψ(x) ∈ G

sup |Ū (ψ(x))| ≤ c sup |ū(y)| ≤ c sup |u(x)|.
B0 G+ Ω
Since a similar computation for the derivatives holds, it follows that there is a constant
c > 0, independent of u, such that
kwkC k (Ω̄∪B 0 ) ≤ ckukC k (Ω̄) .
3. Now consider a finite covering of ∂Ω by balls Bi , i = 1, . . . , N , such as B in the

preceding, and let {wi } be the corresponding C k extensions. We may assume the balls Bi
are so small that their union with Ω is contained in Ω0 . Let Ω0 ⊂⊂ Ω be such that Ω0 and
the balls Bi provide a finite open covering of Ω. Let {ηi }, i = 1, . . . , N , be a partition of
unity subordinate to this covering and set
X
w = uη0 + wi η i
with the understanding that wi ηi = 0 if ηi = 0. Then w is an extension of u into Ω0 and

has the required properties. Thus (b) is established.
4. We now prove (a). If u ∈ W k,p (Ω), then by Theorem 2.14, there exist functions
um ∈ C ∞ (Ω̄) such that um → u in W k,p (Ω). Let Ω ⊂ Ω00 ⊂ Ω0 , and let Um be the
extension of um to Ω00 as given in (b). Then
kUm − Ul kW k,p (Ω00 ) ≤ ckum − ul kW k,p (Ω)
which implies that {Um } is a Cauchy sequence and so converges to a U ∈ W0k,p (Ω00 ), since
Um ∈ C0k (Ω00 ). Now extend Um , U by 0 to Ω0 . It is easy to see that U is the desired
extension.
5. We now prove (c). At any point x0 ∈ ∂Ω let the mapping ψ and the ball G be defined
as in (b). By definition, u ∈ C k (∂Ω) implies that ū = u ◦ ψ −1 ∈ C k (G ∩ ∂Rn+ ). We define
Φ̄(y 0 , yn ) = ū(y 0 ) in G and set Φ(x) = Φ̄ ◦ ψ(x) for x ∈ ψ −1 (G). Clearly, Φ ∈ C k (B̄) for
some ball B = B(x0 ) and Φ = u on B ∩ ∂Ω. Now let {Bi } be a finite covering of ∂Ω by
balls such as B and let Φi be the corresponding C k functions defined on Bi . For each i, we
define the function Ui (x) as follows: in the ball Bi take it equal to Φi , outside Bi take it
equal to zero if x 6∈ ∂Ω and equal to u(x) if x ∈ ∂Ω. The proof can now be completed as in
(b) by use of an appropriate partition of unity.
2.2.5. Trace Theorem. Unless otherwise stated, Ω will denote a bounded open connected
set in Rn , i.e., a bounded domain. Let Γ be a surface which lies in Ω̄ and has the represen-
tation
xn = ϕ(x0 ), x0 = (x1 , . . . , xn−1 )
where ϕ(x0 ) is Lipschitz continuous in Ū . Here U is the projection of Γ onto the coordinate
plane xn = 0. Let p ≥ 1. A function u defined on Γ is said to belong to Lp (Γ) if
Z
1
kukLp (Γ) ≡ ( |u(x)|p dS) p < ∞
Γ
where
Z Z n−1
p 0 0 p
X ∂ϕ 0 2 1 0
|u(x)| dS = |u(x , ϕ(x ))| [1 + ( (x )) ] 2 dx .
Γ U ∂xi
i=1
Thus Lp (Γ) reduces to a space of the type Lp (U ) where U is a domain in Rn−1 .
For every function u ∈ C(Ω̄), its values γ0 u ≡ u|Γ on Γ are uniquely given. The function
γ0 u will be called the trace of the function u on Γ. Note that u ∈ Lp (Γ) since γ0 u ∈ C(Γ).
On the other hand, if we consider a function u defined a.e. in Ω (i.e., functions are
considered equal if they coincide a.e.), then the values of u on Γ are not uniquely determined
since meas(Γ) = 0. In particular, since ∂Ω has measure 0, there exist infinitely many
extensions of u to Ω̄ that are equal a.e. We shall therefore introduce the concept of trace for
functions in W 1,p (Ω) so that if in addition, u ∈ C(Ω̄), the new definition of trace reduces
to the definition given above.
Lemma 2.20. Let ∂Ω ∈ C 0,1 . Then for u ∈ C 1 (Ω̄),
(2.6) kγ0 ukLp (∂Ω) ≤ ckuk1,p
Proof. For simplicity, let n = 2. The more general case is handled similarly. In a neigh-
borhood of a boundary point x ∈ ∂Ω, we choose a local (ξ, η)-coordinate system, where the
boundary has the local representation
η = ϕ(ξ), −α ≤ ξ ≤ α
with the C 0,1 function ϕ. Then there exists a β > 0 such that all the points (ξ, η) with
−α ≤ ξ ≤ α, ϕ(ξ) − β ≤ η ≤ ϕ(ξ)
belong to Ω̄. Let u ∈ C 1 (Ω̄). Then
Z ϕ(ξ)
u(ξ, ϕ(ξ)) = uη (ξ, η)dη + u(ξ, t)
t
where ϕ(ξ) − β ≤ t ≤ ϕ(ξ). Applying the inequality (a + b)p ≤ 2p−1 (ap + bp ) together with
Hölder’s inequality we have
Z ϕ(ξ)
|u(ξ, ϕ(ξ))|p ≤ 2p−1 β p−1 |uη (ξ, η)|p dη + 2p−1 |u(ξ, t)|p .
ϕ(ξ)−β
An integration with respect to t yields

Z ϕ(ξ)
p p−1
β|u(ξ, ϕ(ξ))| ≤ 2 [β p |uη (ξ, η)|p + |u(ξ, η)|p ]dη.
ϕ(ξ)−β
Finally, integration over the interval [−α, α] yields

Z α Z
p p−1
(2.7) β|u(ξ, ϕ(ξ))| dξ ≤ 2 (β p |uη |p + |u|p )dξdη
−α S
where S denotes a local boundary strip. Suppose ϕ(·) is C 1 . Then the differential of arc
length is given by ds = (1 + ϕ02 )1/2 dξ. Addition of the local inequalities (2.7) yields the
assertion (2.6). Now if ϕ(·) is merely Lipschitz continuous, then the derivative ϕ0 exists a.e.
and is bounded. Thus we also obtain (2.6).
Since C 1 (Ω̄) = W 1,p (Ω), the bounded linear operator γ0 : C 1 (Ω̄) ⊂ W 1,p (Ω) → Lp (∂Ω)
can be uniquely extended to a bounded linear operator γ0 : W 1,p (Ω) → Lp (∂Ω) such that
(2.6) remains true for all u ∈ W 1,p (Ω). More precisely, we obtain γ0 u in the following
way: Let u ∈ W 1,p (Ω). We choose a sequence {un } ⊂ C 1 (Ω̄) with kun − uk1,p → 0. Then
kγ0 un − γ0 ukLp (∂Ω) → 0.
The function γ0 u (as an element of Lp (∂Ω)) will be called the trace of the function
u ∈ W 1,p (Ω) on the boundary ∂Ω. (kγ0 ukLp (∂Ω) will be denoted by kukLp (∂Ω) .) Thus the
trace of a function is defined for any element u ∈ W 1,p (Ω).
The above discussion partly proves the following:

Theorem 2.21. (Trace) Suppose ∂Ω ∈ C 1 . Then there is a unique bounded linear operator
γ0 : W 1,p (Ω) → Lp (∂Ω) such that γ0 u = u|∂Ω for u ∈ C(Ω̄)∩W 1,p (Ω), and γ0 (au) = γ0 a·γ0 u
for a(x) ∈ C 1 (Ω̄), u ∈ W 1,p (Ω). Moreover, N (γ0 ) = W01,p (Ω) and R(γ0 ) = Lp (∂Ω).
Proof. 1. Suppose u ∈ C(Ω̄) ∩ W 1,p (Ω). Then by Theorem 2.19, u can be extended into
Ω0 (Ω ⊂⊂ Ω0 ) such that its extension U ∈ C(Ω̄0 ) ∩ W 1,p (Ω0 ). Let Uh (x) be the mollified
function for U . Since Uh → U as h → 0 in both the norms k · kC(Ω̄) , k · kW 1,p (Ω) , we find that
as h → 0, Uh |∂Ω → u|∂Ω uniformly and Uh |∂Ω → γ0 u in Lp (∂Ω). Consequently, γ0 u = u|∂Ω .
2. Now au ∈ W 1,p (Ω) if a ∈ C 1 (Ω̄), u ∈ W 1,p (Ω) and consequently, γ0 (au) is defined.
Let {un } ⊂ C 1 (Ω̄) with kun − uk1,p → 0. Then
γ0 (aun ) = γ0 a · γ0 un
and the desired product formula follows by virtue of the continuity of γ0 .
3. If u ∈ W01,p (Ω), then there is a sequence {un } ⊂ C01 (Ω) with kun − uk1,p → 0.
But un |∂Ω = 0 and as n → ∞, un |∂Ω → γ0 u in Lp (∂Ω) which implies γ0 u = 0. Hence
W01,p (Ω) ⊂ N (γ0 ). Now suppose u ∈ N (γ0 ). If u ∈ W 1,p (Ω) has compact support in Ω,
then by an earlier remark, u ∈ W01,p (Ω). If u does not have compact support in Ω, then it
can be shown that there exists a sequence of cut-off functions ηk such that ηk u ∈ W 1,p (Ω)
has compact support in Ω, and moreover, kηk u − uk1,p → 0. By using the corresponding
mollified functions, it follows that u ∈ W01,p (Ω) and N (γ0 ) ⊂ W01,p (Ω). Details can be found
in Evans’s book.
4. To see that R(γ0 ) = Lp (∂Ω), let f ∈ Lp (∂Ω) and let ε > 0 be given. Then there is a
u ∈ C 1 (∂Ω) such that ku − f kLp (∂Ω) < ε. If we let U ∈ C 1 (Ω̄) be the extension of u into Ω̄,
then clearly kγ0 U − f kLp (∂Ω) < ε, which is the desired result since U ∈ W 1,p (Ω).
Remark 2.8. We note that the function u ≡ 1 belongs to W 1,p (Ω) ∩ C(Ω̄) and its trace
on ∂Ω is 1. Hence this function does not belong to W01,p (Ω), which establishes the earlier
assertion that W01,p (Ω) 6= W 1,p (Ω).
Let u ∈ W k,p (Ω), k > 1. Since any weak derivative Dα u of order |α| < k belongs to
W 1,p (Ω), this derivative has a trace γ0 Dα u belonging to Lp (∂Ω). Moreover
kDα ukLp (∂Ω) ≤ ckDα uk1,p ≤ ckukk,p
for constant c > 0 independent of u.
Assuming the boundary ∂Ω ∈ C 1 , the unit outward normal vector n to ∂Ω exists and
is bounded. Thus, the concept of traces makes it possible to introduce, for k ≥ 2, ∂u/∂n
for u ∈ W k,p (Ω). More precisely, for k ≥ 2, there exist traces of the functions u, Di u so
that, if ni are the direction cosines of the normal, we may define
Xn
γ1 u = (γ0 (Di u))ni , u ∈ W k,p (Ω), k ≥ 2.
i=1
The trace operator γ1 : W k,p (Ω) → Lp (∂Ω) is continuous and γ1 u = (∂u/∂n)|∂Ω for u ∈
C 1 (Ω̄) ∩ W k,p (Ω).
For a function u ∈ C k (Ω̄) we define the various traces of normal derivatives given by
∂j u
|∂Ω , 0 ≤ j ≤ k − 1.
γj u =
∂nj
Each γj can be extended by continuity to all of W k,p (Ω) and we obtain the following:
Theorem 2.22. (Higher-order traces) Suppose ∂Ω ∈ C k . Then there is a unique con-
Qk−1 k−1−j,p
tinuous linear operator γ = (γ0 , γ1 , . . . , γk−1 ) : W k,p (Ω) → j=0 W (∂Ω) such that
k
for u ∈ C (Ω̄)
∂j u
γ0 u = u|∂Ω , γj u = |∂Ω , j = 1, . . . , k − 1.
∂nj
Moreover, N (γ) = W0k,p (Ω) and R(γ) = k−1 k−1−j,p (∂Ω).
Q
j=0 W
The Sobolev spaces W k−1−j,p (∂Ω), which are defined over ∂Ω, can be defined locally.
2.2.6. Green’s Identities. In this section we assume that p = 2 and we continue to

assume Ω is a bounded domain.
Theorem 2.23. (Integration by Parts) Let u, v ∈ H 1 (Ω) and let ∂Ω ∈ C 1 . Then for
any i = 1, . . . , n
Z Z Z
(2.8) vDi udx = (γ0 u · γ0 v)ni dS − uDi vdx.
Ω ∂Ω Ω
(Di u, Di v are weak derivatives.)
Proof. Let {un } and {vn } be sequences of functions in C 1 (Ω̄) with kun − ukH 1 (Ω) →
0, kvn − vkH 1 (Ω) → 0 as n → ∞. Formula (2.8) holds for un , vn
Z Z Z
vn Di un dx = un vn ni dS − un Di vn dx
Ω ∂Ω Ω
and upon letting n → ∞ relation (2.8) follows.
Corollary 2.24. Let ∂Ω ∈ C 1 .
(a) If v ∈ H 1 (Ω) and u ∈ H 2 (Ω) then
Z Z Z
v∆udx = γ0 v · γ1 udS − (∇u · ∇v)dx (Green’s 1st identity).
Ω ∂Ω Ω
2.3. Sobolev embedding Theorems 39
(b) If u, v ∈ H 2 (Ω) then

Z Z
(v∆u − u∆v)dx = (γ0 v · γ1 u − γ0 u · γ1 v)dS (Green’s 2nd identity).
Ω ∂Ω
In these formulas ∇u ≡ (D1 u, . . . , Dn u) is the gradient vector and ∆u ≡ ni=1 Dii u is the
P
Laplace operator.
Proof. If in (2.8) we replace u by Di u and sum from 1 to n, then Green’s 1st identity is
obtained. Interchanging the roles of u, v in Green’s 1st identity and subtracting the two
identities yields Green’s 2nd identity.
Exercise 2.9. Establish the following one-dimensional version of the trace theorem: If
u ∈ W 1,p (Ω), where Ω = (a, b), then
kukLp (∂Ω) ≡ (|u(a)|p + |u(b)|p )1/p ≤ const kukW 1,p (Ω)
where the constant is independent of u.
2.3. Sobolev embedding Theorems

We consider the following question: If a function u belongs to W k,p (Ω), does u automatically
belong to certain other spaces? The answer will be yes, but which other spaces depend upon
whether 1 ≤ kp < n, kp = n, n < kp < ∞.
The general embedding theorem can be stated as follows.

Theorem 2.25. (Sobolev Inequalities) Let Ω ⊂ Rn be bounded and open with ∂Ω ∈ C 1 .
Assume 1 ≤ p < ∞ and k is a positive integer.
(a) If kp < n and 1 ≤ q ≤ np/(n − kp), then
W k,p (Ω) ⊂ Lq (Ω)
is a continuous embedding; that is,
(2.9) kukLq (Ω) ≤ CkukW k,p (Ω)
where the constant C depends only on k,p,n and Ω.
(b) If kp = n and 1 ≤ r < ∞, then
W k,p (Ω) ⊂ Lr (Ω)
and
(2.10) kukLr (Ω) ≤ CkukW k,p (Ω)
where the constant depends only on k,p,n and Ω.
(c) If kp > n and 0 ≤ α ≤ k − m − n/p, then
W k,p (Ω) ⊂ C m,α (Ω̄)
is a continuous embedding; that is,
(2.11) kukC m,α (Ω̄) ≤ CkukW k,p (Ω)
where the constant C depends only on k,p,n,α and Ω.
(d) Let 0 ≤ j < k, 1 ≤ p, q < ∞. Set d = 1/p − (k − j)/n. Then
W k,p (Ω) ⊂ W j,q (Ω)
is a continuous embedding for d ≤ 1/q.
The above results are valid for W0k,p (Ω) spaces on arbitrary bounded domains Ω.
A series of special results will be needed to prove the above theorem. Only selected
proofs will be given to illustrate some of the important techniques.
2.3.1. Gagliardo-Nirenberg-Sobolev Inequality. Suppose 1 ≤ p < n. Do there exist

constants C > 0 and 1 ≤ q < ∞ such that
(2.12) kukLq (Rn ) ≤ Ck∇ukLp (Rn )
for all u ∈ C0∞ (Rn )? The point is that the constants C and q should not depend on u.
We shall show that if such an inequality holds, then q must have a specific form. For
this, choose any u ∈ C0∞ (Rn ), u 6≡ 0, and define for λ > 0
uλ (x) ≡ u(λx) (x ∈ Rn ).
Now Z Z Z
q 1 q
|uλ | dx = |u(λx)| dx = n |u(y)|q dy
Rn Rn λ Rn
and
λp
Z Z Z
p p p
|∇uλ | dx = λ |∇u(λx)| dx = n |∇u(y)|p dy.
Rn Rn λ Rn
Inserting these inequalities into (2.12) we find
1 λ
kukLq (Rn ) ≤ C k∇ukLp (Rn )
λn/q λn/p
and so
(2.13) kukLq (Rn ) ≤ Cλ1−n/p+n/q k∇ukLp (Rn ) .
But then if 1 − n/p + n/q > 0 (or < 0), we can upon sending λ to 0 (or ∞) in (2.13) obtain
a contradiction (u = 0). Thus we must have 1 − n/p + n/q = 0; that is, q = p*, where
np
(2.14) p* =
n−p
is called the Sobolev conjugate of p. Note that then
1 1 1
(2.15) ∗
= − , p∗ > p.
p p n
Next we prove that the inequality (2.12) is in fact correct.
Lemma 2.26. (Gagliardo-Nirenberg-Sobolev Inequality) Assume 1 ≤ p < n. Then
there is a constant C, depending only on p and n, such that
(2.16) kukLp∗ (Rn ) ≤ Ck∇ukLp (Rn )
for all u ∈ C01 (Rn ).
Proof. First assume p = 1. Since u has compact support, for each i = 1, . . . , n we have
Z xi
u(x) = uxi (x1 , . . . , xi−1 , yi , xi+1 , . . . , xn )dyi
−∞
and so Z ∞
|u(x)| ≤ |∇u(x1 , . . . , yi , . . . , xn )|dyi (i = 1, . . . , n).
−∞
Consequently
n Z 1
n Y ∞
n−1
(2.17) |u(x)| n−1 ≤ |∇u(x1 , . . . , yi , . . . , xn )|dyi .
i=1 −∞
Integrate this inequality with respect to x1 :

Z ∞ Z ∞Y n Z ∞ 1
n n−1
|u(x)| n−1 dx1 ≤ |∇u|dyi dx1
−∞ −∞ i=1 −∞
1 n Z 1
Z ∞
n−1
Z ∞ Y ∞
n−1
= |∇u|dy1 |∇u|dyi dx1
−∞ −∞ i=2 −∞
1
1 n Z
!
Z ∞
n−1 Y ∞ Z ∞ n−1
≤ |∇u|dy1 |∇u|dx1 dyi

−∞ i=2 −∞ −∞
the last inequality resulting from the extended Hölder inequality in the appendix.
We continue by integrating with respect to x2 , . . . , xn and applying the extended Hölder
inequality to eventually find (pull out an integral at each step)
Z n Z ∞ Z ∞ 1
n Y n−1
|u(x)| n−1 dx ≤ ··· |∇u|dx1 . . . dyi . . . dxn
Rn i=1 −∞ −∞
Z n
n−1
= |∇u|dx
Rn
which is estimate (2.16) for p = 1.
Consider now the case that 1 < p < n. We shall apply the last estimate to v = |u|γ ,
where γ > 1 is to be selected. First note that
(γuγ−1 Di u)2

if u ≥ 0
(Di |u|γ )2 = = (γ|u|γ−1 Di u)2 .
(−γ(−u)γ−1 Di u)2 if u ≤ 0
Thus v ∈ C01 (Rn ), and
Z n−1 Z
γn n
|u(x)| n−1 dx ≤ |∇|u|γ |dx
Rn Rn
Z
= γ |u|γ−1 |∇u|dx
Rn
Z p−1 Z 1
p(γ−1) p p
p
≤ γ |u| p−1 dx |∇u| dx .
Rn Rn
We set
p(n − 1)
γ= >1
n−p
in which case
γn p(γ − 1) np
= = = p∗ .
n−1 p−1 n−p
Thus, the above estimate becomes
Z 1∗ Z 1
p p
p∗ p
|u| dx ≤C |∇u| dx .
Rn Rn

Theorem 2.27. Let Ω ⊂ Rn be bounded and open, with ∂Ω ∈ C 1 . Assume 1 ≤ p < n, and
∗
u ∈ W 1,p (Ω). Then u ∈ Lp (Ω) and
(2.18) kukLp∗ (Ω) ≤ CkukW 1,p (Ω)
where the constant C depends only on p,n and Ω.
Proof. Since ∂Ω ∈ C 1 , there exists an extension U ∈ W 1,p (Rn ) such that U = u in Ω, U

has compact support and
(2.19) kU kW 1,p (Rn ) ≤ CkukW 1,p (Ω) .
Moreover, since U has compact support, there exist mollified functions um ∈ C0∞ (Rn ) such
that um → U in W 1,p (Rn ). Now according to Lemma 2.26,
kum − ul kLp∗ (Rn ) ≤ Ck∇um − ∇ul kLp (Rn )
∗
for all l, m ≥ 1; whence um → U in Lp (Rn ) as well. Since Lemma 2.26 also implies
kum kLp∗ (Rn ) ≤ Ck∇um kLp (Rn )
we get in the limit that
kU kLp∗ (Rn ) ≤ Ck∇U kLp (Rn ) .
This inequality and (2.19) complete the proof.
Theorem 2.28. Let Ω ⊂ Rn be bounded and open. Assume 1 ≤ p < n, and u ∈ W01,p (Ω).
Then u ∈ Lq (Ω) and
kukLq (Ω) ≤ Ck∇ukLp (Ω)
for each q ∈ [1, p∗ ], the constant C depending only on p,q,n and Ω.
Proof. Since u ∈ W01,p (Ω), there are functions um ∈ C0∞ (Ω) such that um → u in W 1,p (Ω).
We extend each function um to be 0 in Rn \Ω̄ and apply Lemma 2.26 to discover (as above)
kukLp∗ (Ω) ≤ Ck∇ukLp (Ω) .
Since |Ω| < ∞, we furthermore have
kukLq (Ω) ≤ CkukLp∗ (Ω)
for every q ∈ [1, p∗ ].
2.3.2. Morrey’s Inequality. We now turn to the case n < p < ∞. The next result shows
that if u ∈ W 1,p (Ω), then u is in fact Hölder continuous, after possibly being redefined on
a set of measure zero.
Theorem 2.29. (Morrey’s Inequality) Assume n < p < ∞. Then there exists a constant
C, depending only on p and n, such that
(2.20) kuk 0,1− n ≤ CkukW 1,p (Rn ) , ∀ u ∈ C 1 (Rn ).
C p (Rn )
Proof. We first prove the following inequality: for all x ∈ Rn , r > 0 and all u ∈ C 1 (Rn ),
rn |∇u(y)|
Z Z
(2.21) |u(y) − u(x)| dy ≤ dy.
B(x,r) n B(x,r) |x − y|n−1
To prove this, note that, for any w with |w| = 1 and 0 < s < r,
Z s
d
|u(x + sw) − u(x)| = u(x + tw)dt
0 dt
Z s
= ∇u(x + tw) · wdt
0
Z s
≤ |∇u(x + sw)| dt.
0
Now we integrate w over ∂B(0, 1) to obtain
Z Z sZ
|u(x + sw) − u(x)| dS ≤ |∇u(x + sw)| dSdt
∂B(0,1) 0 ∂B(0,1)
|∇u(y)|
Z
= dy
B(x,s) |x − y|n−1
|∇u(y)|
Z
≤ dy.
B(x,r) |x − y|n−1
Multiply both sides by sn−1 and integrate over s ∈ (0, r) and we obtain (2.21).
To establish the bound on kukC 0 (Rn ) , we observe that, by (2.21), for x ∈ Rn ,
Z Z
1 1
|u(x)| ≤ |u(y) − u(x)| dy + |u(y)| dy
|B(x, 1)| B(x,1) |B(x, 1)| B(x,1)
! p−1
Z Z1/p (1−n)p
p
≤ C |∇u(y)|p dy |y − x| p−1 dy + CkukLp (Rn )

Rn B(x,1)
≤ CkukW 1,p (Rn ) .

To establish the bound on the semi-norm [u]γ , γ = 1 − np , take any two points x, y ∈ Rn .
Let r = |x − y| and W = B(x, r) ∩ B(y, r). Then
Z Z
1 1
(2.22) |u(x) − u(y)| ≤ |u(x) − u(z)| dz + |u(y) − u(z)| dz.
|W | W |W | W
Note that |W | = βrn , r = |x − y| and W ≤ min{ B(x,r) , B(y,r) }. Hence, using (2.21), by
R R R
Hölder’s inequality, we obtain

rn
Z Z Z
|u(x) − u(z)| dz ≤ |u(x) − u(z)| dz ≤ |Du(z)||x − z|1−n dz
W B(x,r) n B(x,r)
!1/p Z ! p−1
p
rn
Z (1−n)p
≤ |∇u(z)|p dz |z − x| p−1 dz
n B(x,r) B(x,r)
Z r p−1
(1−n)p p
n n−1
≤ C r k∇ukLp (Rn ) s p−1 s ds
0
n+γ
≤ Cr k∇ukLp (Rn ) ,
n
where γ = 1 − p; similarly,
Z
|u(y) − u(z)| dz ≤ C rn+γ k∇ukLp (Rn ) .
W
Hence, by (2.22),
|u(x) − u(y)| ≤ C |x − y|γ k∇ukLp (Rn ) .
This inequality and the bound on kukC 0 above complete the proof.
Theorem 2.30. (Estimates for W 1,p , n < p ≤ ∞) Let Ω ⊂ Rn be bounded and open,
with ∂Ω ∈ C 1 . Assume n < p < ∞, and u ∈ W 1,p (Ω). Then, after possibly redefining u on
0,1− n
a null set, u ∈ C p (Ω̄) and
kuk 0,1− n
p
≤ CkukW 1,p (Ω)
C (Ω̄)
where the constant C depends only on p, n and Ω.
Proof. Since ∂Ω ∈ C 1 , there exists an extension U ∈ W 1,p (Rn ) such that U = u in Ω, U

has compact support and
(2.23) kU kW 1,p (Rn ) ≤ CkukW 1,p (Ω) .
Moreover, since U has compact support, there exist mollified functions um ∈ C0∞ (Rn ) such
that um → U in W 1,p (Rn ) (and hence on compact subsets). Now according to Morrey’s
inequality,
kum − ul kC 0,1−n/p (Rn ) ≤ Ckum − ul kW 1,p (Rn )
for all l, m ≥ 1; whence there is a function u∗ ∈ C 0,1−n/p (Rn ) such that um → u∗ in

C 0,1−n/p (Rn ). Thus u∗ = u a.e. in Ω. Since we also have
kum kC 0,1−n/p (Rn ) ≤ Ckum kW 1,p (Rn )
we get in the limit that

ku∗ kC 0,1−n/p (Rn ) ≤ CkU kW 1,p (Rn ) .
This inequality and (2.23) complete the proof.
2.3.3. General Cases. We can now concatenate the above estimates to obtain more com-
plicated inequalities.
Assume kp < n and u ∈ W k,p (Ω). Since Dα u ∈ Lp (Ω) for all |α| ≤ k, the Sobolev-
Nirenberg-Gagliardo inequality implies
kDβ ukLp∗ (Ω) ≤ CkukW k,p (Ω)

∗
if |β| ≤ k − 1, and so u ∈ W k−1,p (Ω). Moreover, kukk−1,p∗ ≤ ckukk,p . Similarly, we find
∗∗
u ∈ W k−2,p (Ω), where
1 1 1 1 2
∗∗
= ∗− = − .
p p n p n
Moreover, kukk−2,p∗∗ ≤ ckukk−1,p∗ . Continuing, we find after k steps that u ∈ W 0,q (Ω) =
Lq (Ω) for
1 1 k
= − .
q p n
The stated estimate (2.9) follows from combining the relevant estimates at each stage of
the above argument. In a similar manner the other estimates can be established.
2.4. Compactness 45
2.4. Compactness
We now consider the compactness of the embeddings. Note that if X and Y are Banach
spaces with X ⊂ Y then we say that X is compactly embedded in Y , written X ⊂⊂ Y ,
provided
(i) kukY ≤ CkukX (u ∈ X) for some constant C; that is, the embedding is continuous;
(ii) each bounded sequence in X has a convergent subsequence in Y .
Before we present the next result we recall some facts that will be needed. A subset
S of a normed space is said to be totally bounded if for each ε > 0 there is a finite set
of open balls of radius ε which cover S. Clearly, a totally bounded set is bounded, i.e., it
is contained in a sufficiently large ball. It is not difficult to see that a relatively compact
subset of a normed space is totally bounded, with the converse being true if the normed
space is complete. Moreover, a totally bounded subset of a normed space is separable.
Theorem 2.31. (Rellich-Kondrachov) Let Ω ⊂ Rn be bounded and open. Then for

1 ≤ p < n:
(a) The embedding W01,p (Ω) ⊂ Lq (Ω) is compact for each 1 ≤ q < np/(n − p).
(b) Assuming ∂Ω ∈ C 1 , the embedding W 1,p (Ω) ⊂ Lq (Ω) is compact for each 1 ≤
q < np/(n − p).
(c) Assuming ∂Ω ∈ C 1 , γ0 : W 1,p (Ω) → Lp (∂Ω) is compact.
If p > n, then
(d) Assuming ∂Ω ∈ C 1 , the embedding W 1,p (Ω) ⊂ C 0,α (Ω̄) is compact for each
0 ≤ α < 1 − (n/p).
Proof. We shall just give the proof for p = q = 2. The other cases are proved similarly. (a)
Since C01 (Ω) is dense in H01 (Ω), it suffices to show that the embedding C01 (Ω) ⊂ L2 (Ω) is
compact. Thus, let S = {u ∈ C01 (Ω) : kuk1,2 ≤ 1}. We now show that S is totally bounded
in L2 (Ω).
For h > 0, let Sh = {uh : u ∈ S}, where uh is the mollified function for u. We claim
that Sh is totally bounded in L2 (Ω). Indeed, for u ∈ S, we have
Z
|uh (x)| ≤ ωh (z)|u(x − z)|dz ≤ (sup ωh )kuk1 ≤ c1 (sup ωh )kuk1,2
B(0,h)
and
|Di uh (x)| ≤ c2 sup |Di ωh |kuk1,2 , i = 1, . . . , n
so that Sh is a bounded and equicontinuous subset of C(Ω̄). Thus by the Ascoli Theorem,
Sh is relatively compact (and thus totally bounded) in C(Ω̄) and consequently also in L2 (Ω).
Now, by earlier estimates, we easily obtain
Z Z
kuh − uk22 ≤ ωh (z) 2
|u(x − z) − u(x)| dx dz
B(0,h) Ω
and
1 2
du(x − tz)
Z Z Z
2
|u(x − z) − u(x)| dx = dt dx
Ω Ω 0 dt
Z Z 1 2
= (−∇u(x − tz) · z)dt dx
Ω 0
Z Z 1
2
≤ |z| |∇u(x − tz)| dt dx ≤ |z|2 kuk21,2 .
2
Ω 0
Consequently, kuh − uk2 ≤ h. Since we have shown above that Sh is totally bounded in
L2 (Ω) for all h > 0, it follows that S is also totally bounded in L2 (Ω) and hence relatively
compact.
(b) Suppose now that S is a bounded set in H 1 (Ω). Each u ∈ S has an extension
U ∈ H01 (Ω0 ) where Ω ⊂⊂ Ω0 . Denote by S 0 the set of all such extensions of the functions
u ∈ S. Since kU kH 1 (Ω0 ) ≤ ckuk1,2 , the set S 0 is bounded in H01 (Ω0 ). By (a) S 0 is relatively
compact in L2 (Ω0 ) and therefore S is relatively compact in L2 (Ω).
(c) Let S be a bounded set in H 1 (Ω). For any u(x) ∈ C 1 (Ω̄), the inequality (2.7) with
p = 2 yields
c1
(2.24) kuk2L2 (∂Ω) ≤ kuk22 + c2 βkuk21,2
β
where the constants c1 , c2 do not depend on u or β. By completion, this inequality is
valid for any u ∈ H 1 (Ω). By (b), any infinite sequence of elements of the set S has a
subsequence {un } which is Cauchy in L2 (Ω): given ε > 0, an N can be found such that for
all m, n ≥ N, kum − un k2 < ε. Now we choose β = ε. Applying the inequality (2.24) to
um − un , it follows that the sequence of traces {γ0 un } converges in L2 (∂Ω).
(d) By Morrey’s inequality, the embedding is continuous if α = 1 − (n/p). Now use the
fact that C 0,β is compact in C 0,α if α < β.
Remarks. (a) When p = n, we can easily show that the embedding in (a) is compact for
all 1 ≤ q < ∞. Hence, it follows that the embedding W01,p (Ω) ⊂ Lp (Ω) is compact for all
p ≥ 1. However, when p = n, we do not have embedding W 1,n (Ω) ⊂ L∞ (Ω). For example,
1
u = ln ln(1 + |x| ) ∈ W 1,n (B(0, 1)) but not to L∞ (B(0, 1)) if n ≥ 2.
(b) The boundedness of Ω is essential in the above theorem. For example, let I = (0, 1)
and Ij = (j, j + 1). Let f ∈ C01 (I) and define fj to be the same function defined on Ij by
translation. We can normalize f so that kf kW 1,p (I) = 1. The same is then true for each fj
and thus {fj } is a bounded sequence in W 1,p (R). Clearly f ∈ Lq (R) for every 1 ≤ q ≤ ∞.
Further, if
kf kLq (R) = kf kLq (I) = a > 0
then for any j 6= k we have
Z j+1 Z k+1
kfj − fk kqLq (R) = q
|fj | + |fk |q = 2aq
j k
and so fi cannot have a convergent subsequence in Lq (R). Thus none of the embeddings
W 1,p (R) ⊂ Lq (R) can be compact. This example generalizes to n dimensional space and to
open sets like a half-space.
2.5. Additional Topics 47
2.5. Additional Topics

2.5.1. Equivalent Norms of W 1,p (Ω). Two norms k · k and | · | on a vector space X are
equivalent if there exist constants c1 , c2 ∈ (0, ∞) such that
kxk ≤ c1 |x| ≤ c2 kxk for all x ∈ X.
Note that the property of a set to be open, closed, compact, or complete in a normed space
is not affected if the norm is replaced by an equivalent norm.
A seminorm q on a vector space has all the properties of a norm except that q(u) = 0
need not imply u = 0.
Theorem 2.32. Let ∂Ω ∈ C 1 and let 1 ≤ p < ∞. Set
Z X n
!1/p
kuk = |Di u|p dx + (q(u))p
Ω i=1
where q : W 1,p (Ω) → R is a seminorm with the following two properties:

(i) There is a positive constant d such that for all u ∈ W 1,p (Ω)
q(u) ≤ dkuk1,p .
(ii) If u = constant, then q(u) = 0 implies u = 0.
Then k · k is an equivalent norm on W 1,p (Ω).
Proof. First of all, it is easy to check that k · k defines a norm. Now by (i), it suffices to
prove that there is a positive constant c such that
(2.25) kuk1,p ≤ ckuk for all u ∈ W 1,p (Ω).
Suppose (2.25) is false. Then there exist vn (x) ∈ W 1,p (Ω) such that kvn k1,p > nkvn k. Set
un = vn /kvn k1,p . So
(2.26) kun k1,p = 1 and 1 > nkun k.
According to Theorem 2.31, there is a subsequence, call it again {un }, which converges to u
in Lp (Ω). From (2.26) we have kun k → 0 and therefore ∇un → 0 in Lp (Ω) and q(un ) → 0.
From un → u, ∇un → 0 both in Lp (Ω), we have ∇u = 0 a.e. in Ω and hence u = C, a
constant. This also implies un → C in W 1,p (Ω) and, by kun k1,p = 1, C 6= 0. By continuity,
q(C) = 0, which implies C = 0 by (ii). We thus derive a contradiction.
Example 2.33. Let ∂Ω ∈ C 1 . Assume a(x) ∈ C(Ω), σ(x) ∈ C(∂Ω) with a ≥ 0 (6≡ 0), σ ≥
0 (6≡ 0). Then the following norms are equivalent to k · k1,p on W 1,p (Ω):
n p 1/p
Z X Z !
p
R
(2.27) kuk = |Di u| dx + udx with q(u) = Ω udx .
Ω i=1 Ω
n
!1/p
Z X Z p
|Di u|p dx +
R
(2.28) kuk = γ0 udS with q(u) = ∂Ω γ0 udS .
Ω i=1 ∂Ω
n
Z X Z !1/p
p p
R p
1/p
(2.29) kuk = |Di u| dx + σ|γ0 u| dS with q(u) = ∂Ω σ|γ0 u| dS .
Ω i=1 ∂Ω
n
Z X Z !1/p
p p
R p
1/p
(2.30) kuk = |Di u| dx + a|u| dx with q(u) = Ω a|u| dx .
Ω i=1 Ω
Clearly property (ii) of Theorem 2.32 is satisfied for these semi-norms q(u). In order to
verify condition (i), one uses the trace theorem in (2.28) and (2.29).
R
2.5.2. Poincaré’s Inequalities. Using u − (u)Ω , where (u)Ω = −Ω udx, in the equivalent
norm (2.27) we obtain that
Z n
Z X
p
(2.31) |u(x) − (u)Ω | dx ≤ c |Di u|p dx, u ∈ W 1,p (Ω)
Ω Ω i=1
where the constant c > 0 is independent of u. This inequality is often referred to as

Poincaré’s inequality.
We also note that if u ∈ W01,p (Ω) then the equivalent norm (2.28) implies
Z Z X n
(2.32) p
|u(x)| dx ≤ c |Di u|p dx, u ∈ W01,p (Ω),
Ω Ω i=1
where the constant c > 0 is independent of u. This is also called a Poincaré’s inequality.
Therefore !1/p
Z X n
p
kuk1,p,0 = |Di u| dx
Ω i=1
defines an equivalent norm on W01,p (Ω).
2.5.3. Difference Quotients. For later regularity theory, we will be forced to study the
difference quotient approximations to weak derivatives.
Assume u : Ω → RN is locally integrable. Let {e1 , · · · , en } be the standard basis of Rn .
Define the ith-difference quotient of size h of u by
u(x + hei ) − u(x)
Dih u(x) = , h 6= 0.
h
Then Dih u is defined on Ωh,i = {x ∈ Ω | x + hei ∈ Ω}. Note that
Ωh = {x ∈ Ω | dist(x; ∂Ω) > |h|} ⊂ Ωh,i .
We have the following properties of Dih u.
1) If u ∈ W 1,p (Ω; RN ) then Dih u ∈ W 1,p (Ωh,i ; RN ) and
D(Dih u) = Dih (Du) on Ωh,i .
2) If either u or v has compact support Ω0 ⊂⊂ Ω then the integration-by-parts

formula for difference quotient holds:
Z Z
u · Di v dx = − v · Di−h u dx ∀ |h| < dist(Ω0 ; ∂Ω).
h
Ω Ω
3) Dih (φ u)(x) = φ(x) Dih u(x) + u(x + hei ) Dih φ(x).

Theorem 2.34. (Difference quotient and weak derivatives) (a) Let u ∈ W 1,p (Ω).
Then Dih u ∈ Lp (Ω0 ) for any Ω0 ⊂⊂ Ω satisfying |h| < dist(Ω0 ; ∂Ω). Moreover, we have
kDih ukLp (Ω0 ) ≤ kDi ukLp (Ω) .
2.5. Additional Topics 49
(b) Let u ∈ Lp (Ω), 1 < p < ∞, and Ω0 ⊂⊂ Ω. If there exists a constant K > 0 such that
lim inf kDih ukLp (Ω0 ) ≤ K,
h→0
then the weak derivative Di u exists and satisfies kDi ukLp (Ω0 ) ≤ K.
Proof. (a) Let us suppose initially that u ∈ C 1 (Ω) ∩ W 1,p (Ω). Then, for h > 0,
1 h
Z
h
Di u(x) = Di u(x + thei ) dt,
h 0
so that by Hölder’s inequality
Z h
1
|Dih u(x)|p ≤ |Di u(x + thei )|p dt,
h 0
and hence Z Z hZ Z
1
|Dih u(x)|p dx ≤ p
|Di u| dx dt ≤ |Di u|p dx,
Ω0 h 0 Bh (Ω0 ) Ω
where Bh (Ω0 ) = {x ∈ Ω | dist(x; Ω0 ) < h}. The extension of this inequality to arbitrary
functions in W 1,p (Ω) follows by a straight-forward approximation argument.
(b) Since 1 < p < ∞, there exists a sequence {hm } tending to zero and a function
v ∈ Lp (Ω0 ) with kvkp ≤ K such that Dihm u * v in Lp (Ω0 ) as m → ∞. This means for all
φ ∈ C0∞ (Ω0 )
Z Z
lim φ Dihm u dx = φ v dx.
m→∞ Ω0 Ω0
Now for |hm | < dist(supp φ; ∂Ω0 ), we have
Z Z Z
φ Dihm u dx =− u Di−hm φ dx →− u Di φ dx.
Ω0 Ω0 Ω0
Hence Z Z
φ v dx = − u Di φ dx,
Ω0 Ω0
which shows v = Di u ∈ Lp (Ω0 ) and kDi ukLp (Ω0 ) ≤ K.
Remark 2.10. Variants of Theorem 2.34 can be valid even if it is not the case Ω0 ⊂⊂ Ω.
For example if Ω is the open half-ball B(0, 1) ∩ {xn > 0}, Ω0 = B(0, 1/2) ∩ {xn > 0}, and if
u ∈ W 1,p (Ω), then we have the bound
kDih ukLp (Ω0 ) ≤ kDi ukLp (Ω)
for i = 1, 2, · · · , n − 1 and 0 < |h| < dist(Ω0 , ∂Ω).
We will need this remark for boundary regularity later.
2.5.4. Fourier Transform Methods. For a function u ∈ L1 (Rn ), we define the Fourier
transform of u by
Z
1
û(y) = e−ix·y u(x) dx, ∀ y ∈ Rn ,
(2π)n/2 Rn
and the inverse Fourier transform by
Z
1
ǔ(y) = eix·y u(x) dx, ∀ y ∈ Rn .
(2π)n/2 Rn
Theorem 2.35. (Plancherel’s Theorem) Assume u ∈ L1 (Rn ) ∩ L2 (Rn ). Then û, ǔ ∈

L2 (Rn ) and
kûkL2 (Rn ) = kǔkL2 (Rn ) = kukL2 (Rn ) .
Since L1 (Rn ) ∩ L2 (Rn ) is dense in L2 (Rn ), we can use this result to extend the Fourier
transforms on to L2 (Rn ). We still use the same notations for them. Then we have
Theorem 2.36. (Property of Fourier Tranforms) Assume u, v ∈ L2 (Rn ). Then

(i) Rn uv̄ dx = Rn ûv̂¯ dy,
R R
(ii) Ddα u(y) = (iy)α û(y) for each multiindex α such that D α u ∈ L2 (Rn ),
ˇ
(iii) u = û.
Next we use the Fourier transform to characterize the spaces H k (Rn ).
Theorem 2.37. Let k be a nonnegative integer. Then, a function u ∈ L2 (Rn ) belongs to

H k (Rn ) if and only if
(1 + |y|k )û(y) ∈ L2 (Rn ).
In addition, there exists a constant C such that
C −1 kukH k (Rn ) ≤ k(1 + |y|k ) ûkL2 (Rn ) ≤ C kukH k (Rn )
for all u ∈ H k (Rn ).
Using the Fourier transform, we can also define fractional Sobolev spaces H s (Rn ) for
any 0 < s < ∞ as follows
H s (Rn ) = {u ∈ L2 (Rn ) | (1 + |y|s ) û ∈ L2 (Rn )},
and define the norm by

kukH s (Rn ) = k(1 + |y|s ) ûkL2 (Rn ) .
From this we easily get the estimate
kukL∞ (Rn ) ≤ kûkL1 (Rn )

= k(1 + |y|s )û (1 + |y|s )−1 kL1 (Rn )
≤ k(1 + |y|s )ûkL2 (Rn ) k(1 + |y|s )−2 k2L1 (Rn )
≤ C kukH s (Rn ) ,
where C = k(1 + |y|s )−2 k2L1 (Rn ) < ∞ if and only if s > n2 . Therefore we have an easy
embedding, which is known valid for integers s by the previous Sobolev embedding theorem,
H s (Rn ) ⊂ L∞ (Rn ) if s > n2 .
2.6. Spaces of Functions Involving Time

We study spaces of functions mapping time into Banach spaces. These will be essential for
the study of weak solutions to evolution equations later.
2.6. Spaces of Functions Involving Time 51
2.6.1. Calculus of Abstract Functions. Let X be a real Banach space with norm k · k
and let I be any interval of the real line R.
Definition 2.11. (i) A function u : I → X is called an abstract function.
(ii) An abstract function u : I → X is said to be continuous at a point t0 ∈ I if
lim ku(t) − u(t0 )k = 0.
t→t0
(If t0 is an end point of I, the continuity at t0 is defined through the one-sided limit.) If
u(t) is continuous at each point of I, then we write u ∈ C(I; X).
(iii) Abstract function u : I → X is said to be differentiable at the point t0 ∈ I if
there exists an element l = u0 (t0 ) ∈ X such that
lim k[u(t0 + h) − u(t0 )]/h − u0 (t0 )k = 0.
h→0
We say u(t) is differentiable on I if it is differentiable at each point of I.
Remark 2.12. (i) If u : I → X is continuous at t0 ∈ I then real-valued function ku(t)k is
continuous at t0 .
(ii) If I = [a, b] is a compact interval, then C([a, b]; X) becomes a Banach space with
norm
kukC([a,b];X) := max ku(t)k.
t∈[a,b]
Theorem 2.38. (Mean Value Theorem) Let u(t) ∈ C([a, b]; X) and suppose u0 (t) exists
for every t ∈ (a, b). Then
ku(a) − u(b)k ≤ (b − a) sup ku0 (t)k.
a<t<b
Proof. We use a standard device which reduces the problem to the classical case. Namely,
consider the real-valued function φ(t) = f (u(t)), where f ∈ X ∗ . Since f is continuous and
linear we have
0 u(t + h) − u(t)
φ (t) = lim f = f (u0 (t)).
h→0 h
Now apply the classical mean value theorem to φ(·) to get
ku(b) − u(a)k = sup f (u(b) − u(a))
kf k=1
= sup (φ(b) − φ(a))
kf k=1
= sup (b − a)f (u0 (t0 )) t0 ∈ (a, b)

kf k=1
≤ (b − a) sup ku0 (t)k.

a<t<b

Definition 2.13. Let u : [a, b] → X be an abstract function and define the partial sums
Xn
SZ = u(t¯i )(ti − ti−1 ), ti−1 ≤ t¯i ≤ ti ,
i=1
where Z is the partition a = t0 < t1 < · · · < tn = b of [a, b] and ∆Z = maxi (ti − ti−1 ) is the
mesh of the partition. We define the Riemann integral
Z b
u(t)dt = lim SZk
a k→∞
if such a common limit value exists for all sequences of partitions {Zk } for which ∆Zk → 0
as k → ∞.
Theorem 2.39. If u(t) ∈ C([a, b]; X), then the Riemann integral exists.
Proof. As in the classical proof, one uses the uniform continuity of u(t) together with the
completeness of X. We shall omit the details.
Theorem 2.40. Let u(t) : [a, b] → X be continuous. Then the following hold:
(a)
Z b Z b
k u(t)dtk ≤ ku(t)kdt
a a
(b)
Z b Z b
f u(t)dt = f (u(t))dt for all f ∈ X∗
a a
(c)
d t
Z
u(s)ds = u(t) for all a ≤ t ≤ b
dt a
(d) If u0 (t) ∈ C((a, b); X), then for any α, β ∈ (a, b)
Z β
u0 (s)ds = u(β) − u(α)
α
Proof. (a) and (b) follow by passing to the limit in the corresponding relations of Riemann
sums.
Rt
(c) Set v(t) = a u(s)ds. Since u(t) is uniformly continuous on [a, b], we have
Z t+h
−1
k[v(t + h) − v(t)]/h − u(t)k = kh [u(s) − u(t)]dsk
t
≤ max ku(s) − u(t)k → 0 as h → 0.
|s−t|≤|h|
(d) Let φ(t) = f (u(t)), where f ∈ X ∗ . Then by using (b), we obtain

Z β
f u(β) − u(α) − 0
u (t)dt = 0 for all f ∈ X ∗ .
α
The result now follows easily.
2.6.2. Measurable Functions and Sobolev Spaces. We now extend the continuous
functions to measurable functions. In what follows, we assume I is a bounded interval in R
and X is a Banach space.
Definition 2.14. (i) A function s : I → X is called simple if it has the form
m
X
s(t) = χEi (t)ui (t ∈ I),
i=1
where Ei is a Lebesgue measurable subset of I and ui ∈ X for i = 1, 2, · · · , m. In this case,
we define
Z m
X
s(t) dt = |Ei |ui ∈ X.
I i=1
2.6. Spaces of Functions Involving Time 53
(ii) A function f : I → X is strongly measurable if there exist simple functions

sk : I → X such that
sk (t) → f (t) ∀ a.e. t ∈ I.
(iii) A function f : I → X is weakly measurable if for each u∗ ∈ X ∗ the function
t 7→ hu∗ , f (t)i is Lebesgue measurable on I.
(iv) A function f : I → X is almost separably valued if there exists a subset N ⊂ I
with |N | = 0 such that the set {f (t) | t ∈ I \ N } is separable (that is, has a countable dense
subset).
(v) A strongly measurable function f : I → X is (Bochner) integrable if there exists
a sequence of simple functions sk : I → X such that
Z
lim ksk (t) − f (t)k dt = 0.
k→∞ I
In this case, we define Z Z

f (t) dt = lim sk (t) dt ∈ X.
I k→∞ I
Theorem 2.41. (Bochner-Pettis Theorem) (i) A function f : I → X is strongly mea-

surable if and only if f is weakly measurable and almost separably valued.
(ii) A strongly measurable function f : I → X is Bochner integrable if and only if kf (t)k
is Lebesgue integrable; that is,
Z
kf kL1 (I;X) := kf (t)k dt < ∞.
I
Remark 2.15. (i) The space Lp (I; X) consists of all strongly measurable functions
u : I → X with
Z 1/p
kukLp (I;X) := ku(t)kp dt <∞
I
if 1 ≤ p < ∞ and
kukL∞ (I;X) := esssupt∈I ku(t)k < ∞
if p = ∞. As in the usual Lebesgue space cases, we identify functions that are almost
everywhere equal. Then Lp (I; X) becomes a Banach space for all 1 ≤ p ≤ ∞.
(ii) If X is reflexive, then we have
p
(Lp (I; X))∗ ≈ Lq (I; X ∗ ) (1 < p < ∞, q = ).
p−1
However, usually, (L1 (I; X))∗ 6≈ L∞ (I; X ∗ ). In fact, if X is a separable Banach space, then
(L1 (I; X))∗ ≈ L∞ ∗
w (I; X ),
where L∞ ∗ ∗
w (I; X ) consists of functions g : I → X such that for each u ∈ X the function
t 7→ hg(t), ui is Lebesgue measurable and essentially bounded on I with the norm
kgkw := sup khg(t), uikL∞ (I) < ∞.
u∈X,kuk≤1
¯ X) consists of all continuous functions u : I¯ → X with

(iii) The space C(I;
kukC(I;X)
¯ := max ku(t)k < ∞.
t∈I¯
Definition 2.16. (i) Let u, v ∈ L1 (I; X). We say v is the weak derivative of u, written
u0 = v, provided Z Z
0
φ (t)u(t) dt = − φ(t)v(t) dt
I I
holds in X for all scalar test functions φ ∈ Cc∞ (I).
(ii) The Sobolev space W 1,p (I; X) consists of all functions u ∈ Lp (I; X) such that weak
derivative u0 exists and belongs to Lp (I; X). The norm is defined by
( R 1/p
I (ku(t)kp + ku0 (t)kp )dt (1 ≤ p < ∞),
kukW 1,p (I;X) = 0
esssupt∈I (ku(t)k + ku (t)k) (p = ∞).
We write H 1 (I; X) = W 1,2 (I; X).
Theorem 2.42. (Calculus in W 1,p (I; X)) Let u ∈ W 1,p (I; X) for some 1 ≤ p ≤ ∞. Then
¯ X) (after being redefined on a null set of time), with
(i) u ∈ C(I;
kukC(I;X)
¯ ≤ CkukW 1,p (I;X)
for a constant C depending on I.
Rt
(ii) u(t) = u(s) + s u0 (τ ) dτ for all s ≤ t in I.
Proof. Extend u outside of I by 0 on t ∈ R, and then set uε = ωε ? u, ωε denoting the

usual mollifier on R. We have (uε )0 = ωε ? u0 on Iε := (a + ε, b − ε) if I = (a, b). Then the
proof can be completed by standard approximation method upon ε → 0.
Theorem 2.43. (More calculus) Suppose u ∈ L2 (I; H01 (Ω)), with u0 ∈ L2 (I; H −1 (Ω)).
Then
¯ L2 (Ω)) (after being redefined on a null set of time), with
(i) u ∈ C(I;
0
kukC(I;L
¯ 2 (Ω)) ≤ C(kukL2 (I;H 1 (Ω)) + ku kL2 (I;H −1 (Ω)) )
0
for a constant C depending on I.

¯ with
(ii) The mapping t 7→ ku(t)k2L2 (Ω) is absolutely continuous on I,
d
ku(t)k2L2 (Ω) = 2hu0 (t), u(t)i
dt
for a.e. t ∈ I, where “h · i” is the pairing in H −1 (Ω) × H01 (Ω).
For use later in the regularity study, we will need an extension of Theorem 2.43.
Theorem 2.44. (Mapping into better spaces) Let Ω be a bounded domain with smooth
∂Ω and m a nonnegative integer. Suppose u ∈ L2 (I; H m+2 (Ω)), with u0 ∈ L2 (I; H m (Ω)).
¯ H m+1 (Ω)) (after being redefined on a null set of time), with
Then u ∈ C(I;
0
kukC(I;H
¯ m+1 (Ω)) ≤ C(kukL2 (I;H m+2 (Ω)) + ku kL2 (I;H m (Ω)) )
for a constant C depending on I, Ω and m.

Remark 2.17. In most of the study of evolution equations later, the interval I = (0, T ).
In this case, we write Lp (I; X) as Lp (0, T ; X); other spaces are to be denoted similarly.
Chapter 3
Second-Order Linear
Elliptic Equations
3.1. Differential Equations in Divergence Form

Henceforth, Ω ⊂ Rn denotes a bounded domain with boundary ∂Ω ∈ C 1 .
3.1.1. Linear Elliptic Equations. We study the (Dirichlet) boundary value problem
(BVP)
(3.1) Lu = f in Ω, u = 0 on ∂Ω.
Here f is a given function in L2 (Ω) (or more generally, an element in the dual space of
H01 (Ω)) and L is a second-order differential operator having either the divergence
form
Xn n
X
(3.2) Lu ≡ − Di (aij (x)Dj u) + bi (x)Di u + c(x)u
i,j=1 i=1
or else
n
X n
X
Lu ≡ − aij (x)Dij u + bi (x)Di u + c(x)u
i,j=1 i=1
with given real coefficients aij (x), bi (x) and c(x). We also assume
aij (x) = aji (x) (i, j = 1, . . . , n).
Definition 3.1. The partial differential operator L is said to be uniformly elliptic in
Ω if there exists a number θ > 0 such that for every x ∈ Ω and every real vector ξ =
(ξ1 , . . . , ξn ) ∈ Rn
n
X n
X
(3.3) aij (x)ξi ξj ≥ θ |ξi |2 .
i,j=1 i=1
We will assume aij , bi , c ∈ L∞ (Ω). Define the bilinear form

Z n n
X X
B1 [u, v] ≡ aij Dj uDi v + ( bi Di u + cu)v dx
Ω i,j=1 i=1
55
56 3. Second-Order Linear Elliptic Equations
Definition 3.2. Let f ∈ L2 (Ω). A function u ∈ H01 (Ω) is called a weak solution of (3.1)
with L given by (3.2) if B1 [u, v] = (f, v)L2 for all v ∈ H01 (Ω); that is, the following holds
Z n n Z
X X
f vdx ∀ v ∈ H01 (Ω).

(3.4) aij Dj uDi v + ( bi Di u + cu)v dx =
Ω i,j=1 i=1 Ω
If f ∈ H −1 (Ω) = (H01 (Ω))∗ , the dual space of H01 (Ω), then weak solutions are defined by
replacing the right-hand side by hf, ui, h, i being the dual pairing on H −1 (Ω) × H01 (Ω).
Exercise 3.3. Consider the following weak formulation: Given f ∈ L2 (Ω). Find u ∈ H 1 (Ω)
satisfying
Z Z
∇u · ∇vdx = f vdx for all v ∈ H 1 (Ω).
Ω Ω
Find the boundary value problem solved by u. What is the necessary condition for the
existence of such a u?
3.1.2. General Systems in Divergence Form. For N unknown functions, u1 , · · · , uN ,

we can write u = (u1 , · · · , uN ) and say u ∈ X(Ω; RN ) if each uk ∈ X(Ω), where X is a
symbol of any function spaces we learned. If u ∈ W 1,p (Ω; RN ) then we use Du to denote
the N × n Jacobi matrix
Du = (∂uk /∂xi )1≤k≤N,1≤i≤n .
The (Dirichlet) BVP for a most general system of second-order (quasilinear) par-
tial differential equations in divergence form can be written as follows:
(3.5) − div A(x, u, Du) + b(x, u, Du) = F in Ω, u = 0 on ∂Ω,
where A(x, s, ξ) = (Aki (x, u, ξ)), 1 ≤ i ≤ n, 1 ≤ k ≤ N, and b(x, s, ξ) = (bk (x, u, ξ)),
1 ≤ k ≤ N, are given functions of (x, u, ξ) ∈ Ω × RN × MN ×n , and F = (f k ), 1 ≤ k ≤ N ,
with each f k being a given functional in the dual space of W01,p (Ω).
The coefficients A, b usually satisfy certain structural conditions that will generally
0
assure that both |A(x, u, Du)| and |b(x, u, Du)| belong to Lp (Ω) for all u ∈ W 1,p (Ω; RN ),
where p0 = p−1 p
. In such cases, a function u ∈ W01,p (Ω; RN ) is called a weak solution of
(3.5) if the following holds
Z "X n
#
(3.6) Ai (x, u, Du)Di ϕ + b (x, u, Du)ϕ dx = hf k , ϕi
k k
Ω i=1
for all ϕ ∈ W01,p (Ω) and each k = 1, 2, · · · , N.
Definition 3.4. The system (3.5) is said to be linear if both A and b are linear in the
variables (u, ξ); that is,
X N
X
Aki (x, u, Du) = akl
ij (x) Dj u
l
+ dkl l
i (x) u ,
1≤l≤N, 1≤j≤n l=1
(3.7)
X N
X
k
b (x, u, Du) = bkl
j (x) Dj u
l
+ ckl (x) ul .
1≤j≤n, 1≤l≤N l=1
3.1. Differential Equations in Divergence Form 57
For linear systems, the suitable space is Hilbert space H01 (Ω; RN ) equipped with the
inner product defined by
X Z
(u, v) ≡ Di uk Di v k dx.
1≤i≤n, 1≤k≤N Ω
The pairing between H01 (Ω; RN ) and its dual is given by

N
X
hF, ui = hf k , uk i if F = (f k ) and u = (uk ).
k=1
The bilinear form in this case is defined by
Z
B2 [u, v] ≡ akl l k kl l k kl l k kl l k
ij Dj u Di v + di u Di v + bj Dj u v + c u v dx;
Ω
here the conventional summation notation is used. With this B2 [u, v], a weak solution to
the Dirichlet problem of linear system (3.5) is then a function u ∈ H01 (Ω; RN ) such that
(3.8) B2 [u, v] = hF, vi ∀ v ∈ H01 (Ω; RN ).
Ellipticity Conditions. There are several ellipticity conditions for the system (3.5) in
terms of the leading coefficients A(x, u, ξ). Assume A is smooth on ξ and define
∂Aki (x, u, ξ)
Akl
ij (x, u, ξ) = , ξ = (ξjl ).
∂ξjl
The system (3.5) is said to satisfy the (uniform, strict) Legendre ellipticity condition
if there exists a ν > 0 such that, for all (x, s, ξ), it holds
n X
X N
(3.9) Akl k l
ij (x, s, ξ) ηi ηj ≥ ν |η|
2
for all N × n matrix η = (ηik ).
i,j=1 k,l=1
A weaker condition, obtained by setting η = q ⊗ p = (q k pi ) with p ∈ Rn , q ∈ RN , is the

following (uniform) Legendre-Hadamard condition:
n X
X N
(3.10) Akl k l 2
ij (x, s, ξ) q q pi pj ≥ ν |p| |q|
2
∀ p ∈ Rn , q ∈ RN .
i,j=1 k,l=1
For systems with linear leading terms A given by (3.7), the Legendre condition and
Legendre-Hadamard condition become, respectively,
n X
X N
(3.11) akl k l
ij (x) ηi ηj ≥ ν |η|
2
∀ η;
i,j=1 k,l=1
n X
X N
(3.12) akl k l 2
ij (x) q q pi pj ≥ ν |p| |q|
2
∀ p, q.
i,j=1 k,l=1
Exercise 3.5. If N > 1, the Legendre-Hadamard condition does not imply the Legendre
ellipticity condition. For example, let n = N = 2 and ε > 0. Define constants akl
ij by
2
X
akl k l 2
ij ξi ξj ≡ det ξ + ε |ξ| .
i,j,k,l=1
Show that the Legendre-Hadamard condition holds for all ε > 0. But, the Legendre condi-
tion holds for this system if and only if ε > 1/2.
Exercise 3.6. Let u = (v, w) and x = (x1 , x2 ) = (x, y) ∈ R2 . Then the system of differential
equations defined by akl
ij given above is
∆v + wxy = 0, ∆w − vxy = 0.
This system reduces to two fourth-order equations for v, w (where ∆f = fxx + fyy ):
2 ∆2 v − vxxyy = 0, 2 ∆2 w + wxxyy = 0.
We can easily see that both equations are elliptic if and only if > 1/2.
Exercise 3.7. Formulate the biharmonic equation ∆2 u = f as a linear system and find the
appropriate bilinear form B[u, v] in the definition of weak solutions.
3.2. The Lax-Milgram Theorem

Let H denote a real Hilbert space with inner product (·, ·) and norm k · k.
A map B : H × H → R is called a bilinear form if
B[αu + βv, w] = αB[u, w] + βB[v, w],
B[w, αu + βv] = αB[w, u] + βB[w, v]
for all u, v, w ∈ H and all α, β ∈ R.
Our first existence result is frequently referred to as the Lax-Milgram Theorem.
Theorem 3.1. (Lax-Milgram Theorem) Let B : H → H be a bilinear form. Assume
(i) B is bounded; i.e., |B[u, v]| ≤ αkukkvk, and
(ii) B is strongly positive; i.e., B[u, u] ≥ βkuk2 ,
where α, β are positive constants. Let f ∈ H ∗ . Then there exists a unique element u ∈ H
such that
(3.13) B[u, v] = hf, vi, ∀ v ∈ H.
1
Moreover, the solution u satisfies kuk ≤ β kf k.
Proof. For each fixed u ∈ H, the functional v 7→ B[u, v] is in H ∗ , and hence by the Riesz
Representation Theorem, there exists a unique element w = Au ∈ H such that
B[u, v] = (w, v) ∀ v ∈ H.
It can be easily shown that A : H → H is linear. From (i), kAuk2 = B[u, Au] ≤ αkukkAuk,
and hence kAuk ≤ αkuk for all u ∈ H; that is, A is bounded. Furthermore, by (ii),
βkuk2 ≤ B[u, u] = (Au, u) ≤ kAukkuk and hence kAuk ≥ βkuk for all u ∈ H. By the
Riesz Representation Theorem again, we have a unique w0 ∈ H such that hf, vi = (w0 , v)
for all v ∈ H and kf k = kw0 k. We will show that the equation Au = w0 has a (unique)
solution. There are many different proofs for this, and here we use the Contraction Mapping
Theorem. Note that the solution u to equation Au = w0 is equivalent to the fixed-point
of the map T : H → H defined by T (v) = v − tAv + tw0 (v ∈ H) for any fixed t > 0. We
will show for t > 0 small enough T is a contraction. Note that for all v, w ∈ H we have
kT (v) − T (w)k = k(I − tA)(v − w)k. We compute that for all u ∈ H
k(I − tA)uk2 = kuk2 + t2 kAuk2 − 2t(Au, u)
≤ kuk2 (1 + t2 α2 − 2βt).
3.3. Energy Estimates and Existence Theory 59
We now choose t such that 0 < t < α2β2 . Then the expression in parentheses is positive and
less than 1. Thus the map T : H → H is a contraction on H and therefore has a fixed point.
This fixed point u solves Au = w0 and thus is the unique solution of (3.13); moreover, we
have kf k = kw0 k = kAuk ≥ βkuk and hence kuk ≤ β1 kf k. The proof is complete.
3.3. Energy Estimates and Existence Theory

We study the bilinear forms B1 , B2 defined above. In the following, we assume all coefficients
involved in the problems are in L∞ (Ω). One can easily show the boundedness:
|Bj [u, v]| ≤ αkukkvk
for all u, v in the respective Hilbert spaces H = H01 (Ω) or H = H01 (Ω; RN ) for j = 1, 2.
The strong positivity (also called coercivity) for both B1 and B2 is not always guar-
anteed and involves estimating on the quadratic form Bj [u, u], usually called Gårding’s
estimates. We will derive these estimates for both of them and state the corresponding
existence theorems below.
3.3.1. Gårding’s estimate for B1 [u, u].

Theorem 3.2. Assume the ellipticity condition (3.3) holds. Then, there are constants
β > 0 and γ ≥ 0 such that
(3.14) B1 [u, u] ≥ βkuk2 − γkuk2L2 (Ω) ∀ u ∈ H = H01 (Ω).
Proof. Note that, by the ellipticity,

Z X n n
Z X
B1 [u, u] − ( bi Di u + cu)u dx ≥ θ |Di u|2 dx.
Ω i=1 Ω i=1
Let m = max{kbi kL∞ (Ω) | 1 ≤ i ≤ n} and k0 = kckL∞ (Ω) . Then

|(bi Di u, u)2 | ≤ mkDi uk2 kuk2
≤ (m/2)(εkDi uk22 + (1/ε)kuk22 )
where in the last step we used the arithmetic-geometric inequality |αβ| ≤ (ε/2)α2 +(1/2ε)β 2 .
Combining the estimates we find
B1 [u, u] ≥ (θ − mε/2)kDuk2L2 (Ω) − (k0 + mn/2ε)kuk2L2 (Ω) .
By choosing ε > 0 so that θ − mε/2 > 0 we arrive at the desired inequality, using the
Poincare inequality: kukH 1 (Ω) ≤ CkDukL2 (Ω) for all u ∈ H01 (Ω).
Theorem 3.3. (First Existence Theorem for weak solutions) There is a number
γ ≥ 0 such that for each λ ≥ γ and for each function f ∈ L2 (Ω), the boundary value
problem
Lu + λu = f in Ω, u = 0 on ∂Ω
has a unique weak solution u ∈ H = H01 (Ω) which satisfies
kukH ≤ ckf kL2 (Ω)
where the positive constant c is independent of f . Then result also holds for all f ∈ H −1 (Ω),
with kf kL2 (Ω) replaced by kf kH −1 (Ω) .
Proof. Take γ from (3.14), let λ ≥ γ and define the bilinear form
B λ [u, v] ≡ B1 [u, v] + λ(u, v)2 for all u, v ∈ H
which corresponds to the operator Lu + λu. Then B λ [u, v] satisfies the hypotheses of the
Lax-Milgram Theorem.
Example 3.4. Consider the Neumann boundary value problem
∂u
(3.15) −∆u(x) = f (x) in Ω, = 0 on ∂Ω.
∂ν
A function u ∈ H 1 (Ω) is said to be a weak solution to (3.15 if
Z Z
(3.16) ∇u · ∇v dx = f v dx, ∀ v ∈ H 1 (Ω).
Ω Ω
Obviously,
R taking v ≡ 1 ∈ H 1 (Ω), a necessary condition to have a weak solution is
Ω f (x) dx = 0. We show this is also a sufficient condition for existence of the weak so-
lutions. Note that, if u is a weak solution, then u + c, for all constants c, is also a weak
solution. Therefore, to fix the constants, we consider the vector space
Z
1
H = u ∈ H (Ω) u(x) dx = 0
Ω
equipped with inner product
Z
(u, v)H = ∇u · ∇v dx.
Ω
By the theorem on equivalent norms, it follows that H with this inner product, is indeed a
Hilbert space, and (f, u)L2 (Ω) is a bounded linear functional on H:
|(f, u)L2 (Ω) | ≤ kf kL2 (Ω) kvkL2 (Ω) ≤ kf kL2 (Ω) kvkH .
Hence the Riesz Representation Theorem implies that there exists a unique u ∈ H such
that
(3.17) (u, w)H = (f, w)L2 (Ω) , ∀ w ∈ H.
It follows that u is a weak solutionR to the Neumann problem since for any Rv ∈ H 1 (Ω) we
1
take w = v − c ∈ H, where c = |Ω| Ω vdx, in (3.17) and obtain (3.16) using Ω f dx = 0.
Example 3.5. Let us consider the nonhomogeneous Dirichlet boundary value problem
(3.18) −∆u = f in Ω, u|∂Ω = ϕ
where f ∈ L2 (Ω) and ϕ is the trace of a function w ∈ H 1 (Ω). Note that it is not sufficient
to just require that ϕ ∈ L2 (∂Ω) since the trace operator is not onto. If, for example,
ϕ ∈ C 1 (∂Ω), then ϕ has a C 1 extension to Ω̄, which is the desired w.
The function u ∈ H 1 (Ω) is called a weak solution of (3.18) if u − w ∈ H01 (Ω) and if
Z Z
∇u · ∇vdx = f vdx for all v ∈ H01 (Ω).
Ω Ω
Let u be a weak solution of (3.18) and set u = z + w. Then z ∈ H01 (Ω) satisfies
Z Z
(3.19) ∇z · ∇vdx = (f v − ∇v · ∇w)dx for all v ∈ H01 (Ω).
Ω Ω
Since the right hand side belongs to the dual space H −1 (Ω) = H01 (Ω)∗ , the Lax-Milgram
theorem yields the existence of a unique z ∈ H01 (Ω) which satisfies (3.19). Hence (3.18) has
a unique weak solution u.
Example 3.6. Now let us consider the boundary value (also called Dirichlet) problem for
the fourth order biharmonic operator:
∂u
∆2 u = f in Ω, u|∂Ω = |∂Ω = 0.
∂n
We take H = H02 (Ω). By the general trace theorem, H = H02 (Ω) = {v ∈ H 2 (Ω) : γ0 v =
γ1 v = 0}. Therefore, this space H is the right space for the boundary conditions.
Accordingly, for f ∈ L2 (Ω), a function u ∈ H = H02 (Ω) is a weak solution of the
Dirichlet problem for the biharmonic operator provided
Z Z
∆u∆vdx = f vdx ∀ v ∈ H.
Ω Ω
Consider the bilinear form Z
B[u, v] = ∆u∆vdx.
Ω
Its boundedness follows from the Cauchy-Schwarz inequality
|B[u, v]| ≤ k∆uk2 k∆vk2 ≤ dkuk2,2 kvk2,2 .
Furthermore, it can be shown that k∆uk2 defines a norm on H02 (Ω) which is equivalent to
the usual norm on H 2 (Ω). (Exercise!) Hence
B[u, u] = k∆uk22 ≥ ckuk22,2
and so, by the Lax-Milgram theorem (in fact, just the Riesz Representation Theorem), there
exists a unique weak solution u ∈ H.
Exercise 3.8. Denote by Hc1 the space
Hc1 = {u ∈ H 1 (Ω) : γ0 u = const}.
Note that the constant may be different for different u’s.
(a) Prove that Hc1 is complete.
(b) Let f ∈ C(Ω̄). Prove existence of a unique u ∈ Hc1 satisfying
Z Z
(∇u · ∇v + uv)dx = f vdx ∀ v ∈ Hc1 .
Ω Ω
(c) If u ∈ C 2 (Ω̄) satisfies the equation in (b), find the underlying BVP.
Exercise 3.9. Let Ω = (1, +∞). Show that the BVP −u00 = f ∈ L2 (Ω), u ∈ H01 (Ω) does
not have a weak solution.
3.3.2. Gårding’s estimate for B2 [u, u]. We will derive the Gårding estimate for B2 [u, u].
For simplicity, let H = H01 (Ω; RN ) and let (u, v)H and kukH be the equivalent inner product
and norm defined above on H. Define the bilinear form of the leading terms by
Xn X N Z
A[u, v] = akl l k
ij (x) Dj u Di v dx.
i,j=1 k,l=1 Ω
Theorem 3.7. Assume that either coefficients akl kl

ij satisfy the Legendre condition or aij
are all constants and satisfy the Legendre-Hadamard condition. Then
A[u, u] ≥ ν kuk2H , ∀ u ∈ H.
Proof. In the first case, the conclusion follows easily from the Legendre condition. We
prove the second case when Aklij are constants satisfying the Legendre-Hadamard condition
n X
X N
akl k l 2 2
ij q q pi pj ≥ ν |p| |q| , ∀p ∈ Rn , q ∈ RN .
i,j=1 k,l=1
We prove
n X
X N Z Z
A[u, u] = akl l k
ij Dj u Di u dx ≥ ν |Du|2 dx
i,j=1 k,l=1 Ω Ω
for all u ∈ C0∞ (Ω; RN ). For these test functions u, we extend them onto Rn by zero outside
Ω and thus consider them as functions in C0∞ (Rn ; RN ). Define the Fourier transforms
for such functions u by
Z
−n/2
û(y) = (2π) e−i y·x u(x) dx; y ∈ Rn .
Rn
Then, for any u, v ∈ C0∞ (Rn ; RN ),

Z Z
u(x) · v(x) dx = û(y) · v̂(y) dy,
Rn Rn
D[ k ck
j u (y) = i yj u (y);
the last identity can also be written as Du(y) c = i û(y) ⊗ y. Now, using these identities, we
have Z Z
kl k l
aij Di u (x) Dj u (x) dx = akl [k [l
ij Di u (y) Dj u (y) dy
Rn R n
Z Z
= akl y y
ij i j
ck (y) ubl (y) dy = Re
u akl
y y
ij i j
ck (y) ubl (y) dy .
u
Rn Rn
Write û(y) = η + iξ with η, ξ ∈ RN . Then

Re u (y) u (y) = η k η l + ξ k ξ l .
ck bl
Therefore, by the Legendre-Hadamard condition,

Xn X N
Re aij yi yj u (y) u (y) ≥ ν |y|2 (|η|2 + |ξ|2 ) = ν |y|2 |û(y)|2 .
kl ck bl
i,j=1 k,l=1
Hence,
n X
X N Z
A(u, u) = akl k l
ij Di u (x) Dj u (x) dx
i,j=1 k,l=1 Rn
n
X N Z
X
= Re akl
ij yi yj
ck (y) ubl (y) dy
u
i,j=1 k,l=1 Rn
Z Z
2 2
≥ν |y| |û(y)| dy = ν |iû(y) ⊗ y|2 dy
Rn Rn
Z Z
2
=ν |Du(y)|
c dy = ν |Du(x)|2 dx.
Rn Rn
The proof is complete.
Theorem 3.8. (Gårding’s estimate for system) Let B2 [u, v] be defined by (3.8). Assume
1) akl
ij ∈ C(Ω̄),
2) the Legendre-Hadamard condition holds for all x ∈ Ω; that is,
akl k l 2 2
ij (x) q q pi pj ≥ ν |p| |q| , ∀p ∈ Rn , q ∈ RN .
3) bkl kl kl ∞
i , c , di ∈ L (Ω).
Then, there exist constants λ0 > 0 and λ1 ≥ 0 such that
B2 [u, u] ≥ λ0 kuk2H − λ1 kuk2L2 , ∀ u ∈ H01 (Ω; RN ).
Proof. 1. By uniform continuity, we can choose a small > 0 such that

ν
|akl kl
ij (x) − aij (y)| ≤ , ∀ x, y ∈ Ω̄, |x − y| ≤ .
2
We claim
Z Z Z
ν ν X
(3.20) akl k l
ij (x) Di u Dj u dx ≥ |Du(x)|2 dx = |Di uk (x)|2 dx
Ω 2 Ω 2 1≤i≤n Ω
1≤k≤N
for all test functions u ∈ C0∞ (Ω; RN ) with diam(supp u) ≤ . To see this, we choose any
point x0 ∈ supp u. Then
Z Z
kl k l
aij (x) Di u Dj u dx = akl k l
ij (x0 ) Di u Dj u dx
Ω Ω
Z
akl kl k l

+ ij (x) − aij (x0 ) Di u Dj u dx
supp u
Z Z
ν 2
≥ν |Du(x)| dx − |Du(x)|2 dx,
Ω 2 Ω
which proves (3.20).
2. Now assume u ∈ C0∞ (Ω; RN ), with arbitrary compact support. We cover Ω̄ with
finitely many open balls {B/4 (xm )} with xm ∈ Ω and m = 1, 2, ..., M. For each m, let
ζm ∈ C0∞ (B/2 (xm )) with ζm (x) = 1 for x ∈ B/4 (xm ). Since for any x ∈ Ω̄ we have at least
one m such that x ∈ B/4 (xm ) and thus ζm (x) = 1, we may therefore define
ζm (x)
ϕm (x) = PM 2 1/2 , m = 1, 2, ..., M.
j=1 ζj (x)
Then M 2
P
m=1 ϕm (x) = 1 for all x ∈ Ω. (This is a special case of partition of unity.) We
have thus
M
X
akl k l
ij (x) Di u Dj u = akl 2 k
ij (x) ϕm Di u Dj u
l
m=1
M
X
(3.21) = akl k l
ij (x) Di (ϕm u ) Dj (ϕm u )
m=1
M
X
− akl
ij (x) ϕm ul
D ϕ D
i m i uk
+ ϕm uk
D ϕ D
i m j ul
+ uk l
u Di m j m .
ϕ D ϕ
m=1
Since ϕm u ∈ C0∞ (Ω ∩ B/2 (xm ); RN ) and diam(Ω ∩ B/2 (xm )) ≤ , we have by (3.20)
Z Z
kl k l ν X
aij (x) Di (ϕm u )Dj (ϕm u ) dx ≥ |Di (ϕm uk )|2 dx
Ω 2 1≤i≤n Ω
1≤k≤N
Z
ν X
= ϕ2m |Di uk |2 + |Di ϕm |2 |uk |2 + 2ϕm uk Di ϕm Di uk dx
2 1≤i≤n Ω
1≤k≤N
Z
ν X
≥ ϕ2m |Di uk |2 dx + 2ϕm uk Di ϕm Di uk dx
2 1≤i≤n Ω
1≤k≤N
Z Z
ν X 2 k 2 2 ν
≥ ϕ |Di u | dx − CkukL2 (Ω) = ϕ2 |Du|2 dx − Ckuk2L2 (Ω) ,
4 1≤i≤n Ω m 4 Ω m
1≤k≤N
where we have used the Cauchy inequality with . Then by (3.21) and the fact that
PM 2
m=1 ϕm = 1 on Ω,
Z Z
kl k l ν
aij (x) Di u Dj u dx ≥ |Du|2 dx − CM kuk2L2 (Ω) − C1 kukL2 (Ω) kDukL2 (Ω) .
Ω 4 Ω
The terms in B2 [u, u] involving b, c and d can all be estimated by
C2 (kukL2 (Ω) kDukL2 (Ω) + kuk2L2 (Ω) ).
Finally, using the Cauchy inequality with again, we have
ν
B2 [u, u] ≥ kuk2H 1 (Ω) − C3 kuk2L2 (Ω) ∀ u ∈ C0∞ (Ω; RN )
8 0
and, by density, for all u ∈ H01 (Ω; RN ). This completes the proof.
Note that the bilinear form B λ [u, v] = B2 [u, v] + λ (u, v)L2 satisfies the condition of
the Lax-Milgram theorem on H = H01 (Ω; RN ) for all λ ≥ λ1 ; thus, by the Lax-Milgram
theorem, we easily obtain the following existence result.
Theorem 3.9. Under the hypotheses of the previous theorem, for λ ≥ λ1 , the Dirichlet
problem for the system (3.5) with linear coefficients (3.7) has a unique weak solution u in
H01 (Ω; RN ) for any bounded linear functional F on H. Moreover, the solution u satisfies
kukH ≤ C kF k with a constant C depending on λ, and the coefficients.
Corollary 3.10. Given λ ≥ λ1 as in the theorem, then the operator K : L2 (Ω; RN ) →
L2 (Ω; RN ), where, for each F ∈ L2 (Ω; RN ), u = KF is the unique weak solution to the
BVP above, is a compact linear operator.
Proof. By the theorem, kukH01 (Ω;RN ) ≤ CkF kL2 (Ω;RN ) . Hence K is a bounded linear opera-
tor from L2 (Ω; RN ) to H01 (Ω; RN ), which, by the compact embedding theorem, is compactly
embedded in L2 (Ω; RN ). Hence, as a linear operator from L2 (Ω; RN ) to L2 (Ω; RN ), K is
compact.
3.4. Fredholm Alternatives

We study the general linear system Lu whose bilinear form is given by B2 [u, v] defined above
on H = H01 (Ω; RN ). We need some necessary results on the spectral theory of compact linear
operators given in §1.4 of Chapter 1.
3.4. Fredholm Alternatives 65
Definition 3.10. The adjoint bilinear form B2∗ of B2 is defined by

B2∗ [u, v] = B2 [v, u] ∀ u, v ∈ H = H01 (Ω; RN ).
This bilinear form B2∗ [u, v] is associated to the formal adjoint of Lu of the form
(3.22) L∗ u = − div A∗ (x, u, Du) + b∗ (x, u, Du),
with linear coefficients A∗ (x, u, Du) = (Ãkl ∗ k
ij ) and b (x, u, Du) = (b̃ ) given by
X N
X
Ãki (x, u, Du) = ãkl l
ij (x) Dj u + d˜kl l
i (x) u ,
1≤l≤N, 1≤j≤n l=1
(3.23)
X N
X
b̃k (x, u, Du) = b̃kl l
j (x) Dj u + c̃kl (x) u,
1≤j≤n, 1≤l≤N l=1
where
ãkl lk
ij = aji , d˜kl lk
i = bi , b̃kl lk
j = di , c̃kl = clk (1 ≤ i, j ≤ n, 1 ≤ k, l ≤ N ).
The ellipticity condition of L∗ u is the same as that of Lu, and also B2∗ [u, u] = B2 [u, u].
Theorem 3.11. (Second Existence Theorem for weak solutions) Assume the ellip-
ticity and boundedness of the coefficients of Lu.
(i) Precisely one of the following statements holds:
either
(
for each F ∈ L2 (Ω; RN ) there exists a unique
(3.24)
weak solution u ∈ H01 (Ω; RN ) of Lu = F,
or else
(3.25) there exists a weak solution u 6= 0 in H01 (Ω; RN ) of Lu = 0.
(ii) Furthermore, should case (3.25) hold, the dimension of the subspace N ⊂ H01 (Ω; RN )
of weak solutions of Lu = 0 is finite and equals the dimension of the subspace
N ∗ ⊂ H01 (Ω; RN ) of weak solutions of adjoint problem L∗ u = 0.
(iii) Finally, the problem Lu = F has a weak solution if and only if
(F, v)L2 (Ω;RN ) = 0 ∀ v ∈ N ∗.
The dichotomy (3.24), (3.25) is called the Fredholm alternatives.
Proof. We assume the Gårding inequality holds (see Theorem 3.8 for sufficient conditions):
(3.26) B2 [u, u] ≥ σkuk2H 1 − µkuk2L2 , ∀ u ∈ H01 (Ω; RN ),
0
where σ > 0 and µ ∈ R are constants. We also assume µ > 0. For each F ∈ L2 (Ω; RN ),
define u = KF to be the unique weak solution in H01 (Ω; RN ) of the BVP
Lu + µu = F in Ω, u|∂Ω = 0.
By Theorem 3.9 and Corollary 3.10, this K is well defined and is a compact linear oper-
ator on L2 (Ω; RN ). We write K = (L + µI)−1 . Here I denotes the identity on L2 (Ω; RN )
and also the identity embedding of H01 (Ω; RN ) into L2 (Ω; RN ). Furthermore, given F ∈
L2 (Ω; RN ), u ∈ H01 (Ω; RN ) is a weak solution of Lu = F if and only if Lu + µu = F + µu,
which is equivalent to the equation u = K(F + µu) = KF + µKu; that is, (I − µK)u = KF.
Hence, we have N = N (I − µK) and similarly, N ∗ = N (I − µK∗ ); moreover, Lu = F if
and only if KF ∈ R(I − µK) = (N (I − µK∗ ))⊥ = (N ∗ )⊥ . The proof of (iii) follows as, for
all v ∈ N ∗ , v = µK∗ v and so
(F, v) = (F, µK∗ v) = µ(KF, v);
hence KF ∈ (N ∗ )⊥ if and only if F ∈ (N ∗ )⊥ .
3.4.1. Symmetric Elliptic Operators. In what follows, we assume Ω is a bounded

domain. We consider the operator
Lu = − div A(x, Du) + c(x)u, u ∈ H01 (Ω; RN ),
where A(x, Du) is a linear system defined with A(x, ξ), ξ ∈ MN ×n , given by
X
Aki (x, ξ) = akl l
ij (x)ξj .
1≤l≤N, 1≤j≤n
Here akl ∞
ij (x) and c(x) are given functions in L (Ω). The bilinear form associated to L is
 
Z XN X n
(3.27) B[u, v] =  akl l k
ij (x)Dj u Di v + c(x)u · v
 dx, u, v ∈ H01 (Ω; RN ).
Ω k,l=1 i,j=1
We assume that B is symmetric on H01 (Ω; RN ), that is,

B[u, v] = B[v, u] ∀ u, v ∈ H01 (Ω; RN ).
In this case B ∗ = B and Lu is self-adjoint: L∗ u = Lu. This condition is equivalent to the
following symmetry condition:
(3.28) akl lk
ij (x) = aji (x), ∀ i, j = 1, 2, · · · , n; k, l = 1, 2, · · · , N.
We also assume the Gårding inequality holds (see Theorem 3.8 for sufficient conditions):
(3.29) B[u, u] ≥ σkuk2H 1 − µkuk2L2 , ∀ u ∈ H01 (Ω; RN ),
0
where σ > 0 and µ ∈ R are constants.

For each F ∈ L2 (Ω; RN ), define u = KF to be the unique weak solution in H01 (Ω; RN )
of the BVP
Lu + µu = F in Ω, u|∂Ω = 0.
By Theorem 3.9 and Corollary 3.10, this K is well defined and is a compact linear operator
on L2 (Ω; RN ). Sometime, we write K = (L+µI)−1 . Here I denotes the identity on L2 (Ω; RN )
and also the identity embedding of H01 (Ω; RN ) into L2 (Ω; RN ).
Theorem 3.12. K : L2 (Ω; RN ) → L2 (Ω; RN ) is symmetric and positive; that is,
(KF, G)L2 = (KG, F )L2 , (KF, F )L2 ≥ 0, ∀ F, G ∈ L2 (Ω; RN ).
Furthermore, given λ ∈ R and F ∈ L2 (Ω; RN ), u ∈ H01 (Ω; RN ) is a weak solution of
Lu − λu = F if and only if [I − (λ + µ)K]u = KF.
Proof. Let u = KF and v = KG. Then

(u, G)L2 = B[v, u] + µ(v, u)L2 = B[u, v] + µ(v, u)L2 = (v, F )L2 ,
proving the symmetry. Also, by (3.29),
(3.30) (KF, F )L2 = (u, F )L2 = B[u, u] + µkuk2L2 ≥ σkuk2H 1 = σkKF k2H 1 ≥ 0.
0 0
3.4. Fredholm Alternatives 67
Finally, u ∈ H01 (Ω; RN ) is a weak solution of Lu − λu = F if and only if Lu + µu =

F + (λ + µ)u, which is equivalent to the equation u = K[F + (λ + µ)u] = KF + (λ + µ)Ku;
that is, [I − (λ + µ)K]u = KF.
3.4.2. Eigenvalue Problems. A number λ ∈ R is called a (Dirichlet) eigenvalue of

operator L if the BVP problem
Lu − λu = 0, u|∂Ω = 0
has nontrivial weak solutions in H01 (Ω; RN ); these nontrivial solutions are called the eigen-
functions corresponding to eigenvalue λ.
Theorem 3.13. (Eigenvalue Theorem) Assume (3.28) and (3.29). Then the eigenvalues
of L consist of a countable set Σ = {λk }∞
k=1 , where
−µ < λ1 ≤ λ2 ≤ λ3 ≤ · · ·
are listed repeatedly the same times as the multiplicity, and
lim λk = ∞.
k→∞
Let wk be an eigenfunction corresponding to λk satisfying kwk kL2 (Ω;RN ) = 1. Then {wk }∞

k=1
forms an orthonormal basis of L2 (Ω; RN ).
The first (smallest) eigenvalue λ1 , which is called the (Dirichlet) principal eigenvalue
of L, is characterized by
(3.31) λ1 = min B[u, u].
u∈H01 (Ω;RN )
kuk 2 =1
L (Ω)
Moreover, if u ∈ H01 (Ω; RN ), u 6≡ 0, then u is an eigenfunction corresponding to λ1 if and

only if
B[u, u] = λ1 kuk2L2 (Ω) .
Proof. 1. From Theorem 3.12, we see that λ is an eigenvalue of L if and only if equation
(I − (λ + µ)K)u = 0 has nontrivial solutions u ∈ L2 (Ω; RN ); this exactly says that λ 6= −µ
1
and λ+µ is an eigenvalue of operator K. Since, by (3.30), K is strictly positive, all eigenvalues
of K consist of a countable set of positive numbers tending to zero and hence the eigenvalues
of L consist of a set of numbers {λj }∞j=1 with
−µ < λ1 ≤ λ2 ≤ · · · ≤ λj → ∞.
2. We now prove the second statement. If u is an eigenfunction corresponding to λ1

with kukL2 (Ω) = 1, then easily B[u, u] = λ1 (u, u) = λ1 kuk2L2 (Ω) = λ1 . We now assume
u ∈ H01 (Ω), kukL2 (Ω) = 1.

Let {wk } be the orthonormal basis of L2 (Ω) consisting of eigenfunctions. Then
(
B[wk , wl ] = λk (wk , wl ) = 0 (k 6= l),
B[wk , wk ] = λk (wk , wk ) = λk .
Set w̃k = (λk + µ)−1/2 wk , and consider the inner product on H = H01 (Ω; RN ) defined by
((u, v)) := Bµ [u, v] = B[u, v] + µ(u, v)L2 (Ω) (u, v ∈ H).
Then ((w̃k , w̃l )) = δkl . Let dk = (u, wk )L2 (Ω) . We have

∞
X ∞
X ∞
X
(3.32) d2k = kuk2L2 (Ω) = 1, u= dk wk = d˜k w̃k ,
k=1 k=1 k=1
√
where d˜k = dk λk + µ, in the sense of norm-convergence in L2 (Ω; RN ). We claim that the
series for u converges also in the norm defined by the inner product (( , )) in H. To prove
this claim, for m = 1, 2, · · · , define
m
X m
X
um = dk wk = d˜k w̃k ∈ H.
k=1 k=1
From ((w̃k , u)) = B[w̃k , u] + µ(w̃k , u) = (λk + µ)(w̃k , u) = d˜k , we have

m
X
((um , u)) = d˜2k = ((um , um )) (m = 1, 2, · · · ).
k=1
This implies ((um , um )) ≤ ((u, u)) for all m = 1, 2, · · · . Hence, {um } is bounded in H and
so, by a subsequence, um * ũ in H as m → ∞. Since um → u in L2 , we must have ũ = u
and so
((u, u)) ≤ lim inf ((um , um )),
m→∞
which, combined with ((u − um , u − um )) = ((u, u)) + ((um , um )) − 2((u, um )) = ((u, u)) −
((um , um )), implies that um → u in H, and the claim is proved.
3. Now, by (3.32), we have
∞
X ∞
X ∞
X
B[u, u] = dk B[wk , u] = d2k λk ≥ d2k λ1 = λ1 .
k=1 k=1 k=1
Hence (3.31) is proved. Moreover, if in addition B[u, u] = λ1 , then we have

∞
X
(λk − λ1 )d2k = 0; so dk = 0 if λk > λ1 .
k=1
Pm
Assume λ1 has multiplicity m, with Lwk = λ1 wk (k = 1, 2, · · · , m). Then u = k=1 dk wk ,
and so Lu = λ1 u; that is, u is an eigenfunction corresponding to λ1 .
We consider a special case when N = 1 and the operator Lu is given by

n
X
Lu = − Dj (aij (x)Di u) + c(x)u,
i,j=1
where the uniform ellipticity condition is satisfied, ∂Ω is smooth, and aij , c are smooth
functions satisfying
aij (x) = aji (x), c(x) ≥ 0 (x ∈ Ω̄).
Theorem 3.14. The principal eigenvalue λ1 > 0. Let w1 be an eigenfunction corresponding
to the principal eigenvalue λ1 of L above. Then, either w1 (x) > 0 for all x ∈ Ω or w1 (x) < 0
for all x ∈ Ω. Moreover, the eigenspace corresponding to λ1 is one-dimensional.
Proof. 1. Since in this case the bilinear form B is positive: B[u, u] ≥ σkuk2H 1 (Ω) , we have
0
λ1 > 0. Let w1 be an eigenfunction corresponding to λ1 with kw1 kL2 (Ω) = 1, and set
w1+ = max{0, w1 }, w1− = min{0, w1 }.
3.5. Regularity 69
Then w1± ∈ H01 (Ω), w1 = w1+ + w1− , kw1+ k2L2 (Ω) + kw1− k2L2 (Ω) = kw1 k2L2 (Ω) = 1, and
∇w1+ = χ{w1 ≥0} ∇w1 , ∇w1− = χ{w1 ≤0} ∇w1 .
Hence we have B[w1+ , w1− ] = 0, and

λ1 = B[w1 , w1 ] = B[w1+ , w1+ ] + B[w1− , w1− ] ≥ λ1 kw1+ k2L2 (Ω) + λ1 kw1− k2L2 (Ω) = λ1 .
But then the inequality must be equality. So
B[w1+ , w1+ ] = λ1 kw1+ k2L2 (Ω) , B[w1− , w1− ] = λ1 kw1− k2L2 (Ω) .
Therefore, u = w1± are both solutions to the elliptic equation

(
Lu = λ1 u in Ω,
u=0 on ∂Ω.
2. Since the coefficients of L and Ω are smooth, u = w1± are smooth solutions. (See
Theorem 3.19 below.) Note that Lw1+ = λ1 w1+ ≥ 0 in Ω. By Strong Maximum Principle,
either w1+ ≡ 0 or else w1+ > 0 in Ω; similarly, either w1− ≡ 0 or else w1− < 0 in Ω. This
proves that either w1 < 0 in Ω or else w1 > 0 in Ω.
3. We now prove the eigenspace of λ1 is one-dimensional. Let w be another eigenfunc-
tion. Then, either w(x) > 0 for all x ∈ Ω or w(x) < 0 for all x ∈ Ω. Let t ∈ R be such
that Z Z
w(x) dx = t w1 (x)dx.
Ω Ω
Note that u = w − tw1 is also a solution to Lu = λ1 u. We claim u ≡ 0 and hence w = tw1 ,
proving the eigenspace is one-dimensional. Suppose u 6≡ 0. Then u is another eigenfunction
corresponding to λ1 . Then, by the theorem, weR would have either u(x) > 0 for all x ∈ Ω or
u(x) < 0 for all x ∈ Ω; hence, in either case, Ω u(x)dx 6= 0, which is a contradiction.
Remark 3.11. Let L = −∆. Then, there exists an orthonormal basis {wk }∞ k=1 of L (Ω)
2
1
consisting eigenfunctions wk of −∆ in H0 (Ω). We can see that {wk } is also orthogonal in
H01 (Ω); in fact,
Z
∇wk (x) · ∇wl (x) dx = B[wk , wl ] = λk (wk , wl )L2 (Ω) = λk δkl (k, l = 1, 2, · · · ).
Ω
Furthermore, wk ∈ C ∞ (Ω). If ∂Ω is smooth, then each wk is smooth on Ω̄. See Theorem

3.19 below.
3.5. Regularity
We now address the question as to whether a weak solution u of the PDE
Lu = f in Ω
is smooth or not. This is the regularity problem for weak solutions.
We first study second-order linear differential equations of the divergence form
n
X n
X
(3.33) Lu ≡ − Di (aij (x)Dj u) + bi (x)Di u + c(x)u
i,j=1 i=1
To see that there is some hope that a weak solution may be better than a typical function
in H01 (Ω), let us consider the model problem −∆u = f in Rn . Assume u is smooth enough
to justify the following calculations.
Z Z Xn Z
2 2
f dx = (∆u) dx = uxi xi uxj xj dx
Rn Rn i,j=1 R
n
Xn Z
= − uxi xi xj uxj dx
n
i,j=1 R
n
X Z Z
= uxi xj uxi xj dx = |D2 u|2 dx.
i,j=1 Rn Rn
Thus the L2 norm of the second derivatives of u can be estimated by the L2 norm of f .
Similarly, if we differentiate the PDE with respect to xk , we see that the L2 norm of the
third derivatives of u can be estimated by the L2 norm of the first derivatives of f , etc. This
suggests that we can expect a weak solution u ∈ H01 (Ω) to belong to H m+2 (Ω) whenever
f ∈ H m (Ω).
The above calculations do not really constitute a proof, since we assumed that u was
smooth in order to carry out the calculation. If we merely start with a weak solution in
H01 (Ω), we cannot justify the above computations.
3.5.1. Difference Quotient Method. One can instead rely upon an analysis of certain
difference quotients to obatin higher regularity of weak solutions in H 1 (Ω). Our first regu-
larity result provides the interior H 2 -regularity for weak solutions of the equation Lu = f
based on the difference quotient method.
Theorem 3.15. (Interior H 2 -regularity) Let L be uniformly elliptic, with aij ∈ C 1 (Ω), bi
and c ∈ L∞ (Ω). Let f ∈ L2 (Ω). If u ∈ H 1 (Ω) is a weak solution of (3.33), then for any
Ω0 ⊂⊂ Ω we have u ∈ H 2 (Ω0 ), and
(3.34) kukH 2 (Ω0 ) ≤ C (kukL2 (Ω) + kf kL2 (Ω) )
where the constant C depends only on n, Ω0 , Ω and the coefficients of L.
Proof. Set q = f − ni=1 bi Di u − cu. Since u is a weak solution of (3.33), (by the similar
P
definition as above), this means that
Z X n Z
(3.35) aij Di uDj ϕdx = qϕdx ∀ ϕ ∈ H01 (Ω), supp ϕ ⊂⊂ Ω.
Ω i,j=1 Ω
Step 1: (Interior H 1 -estimate). Take any Ω00 ⊂⊂ Ω. Choose a cutoff function ζ ∈ C0∞ (Ω)
with 0 ≤ ζ ≤ 1 and ζ|Ω00 = 1. We take ϕ = ζ 2 u in (3.35) and perform elementary calculations
using the ellipticity condition, to discover
Z Z
ζ 2 |∇u|2 dx ≤ C (f 2 + u2 )dx.
Ω Ω
Thus
(3.36) kukH 1 (Ω00 ) ≤ C(kf kL2 (Ω) + kukL2 (Ω) ),
where the constant C depends on Ω00 .
3.5. Regularity 71
Step 2: (Interior H 2 -estimate). Take Ω0 ⊂⊂ Ω1 ⊂⊂ Ω2 ⊂⊂ Ω00 ⊂⊂ Ω. Let v ∈ H01 (Ω)

be any function with supp (v) ⊂⊂ Ω1 . Let
1 n o
δ = min dist (supp (v), ∂Ω1 ), dist (Ω1 , ∂Ω2 ), dist (Ω2 , ∂Ω00 ) .
2
h
Let Dk be the difference quotient operator defined above. For 0 < |h| < δ, we choose
the test function ϕ = Dk−h v in (3.35) and obtain, using integration by parts for difference
quotient,
Z n
X Z
Dkh ( aij Di u)Dj vdx = − qDk−h vdx.
Ω i,j=1 Ω
P
Notice that the integrals are in fact over domain Ω1 . Henceforth, we omit the sign. Using
the definition of q and the equality
Dkh (aij Di u) = ahij Dkh Di u + Di uDkh aij ,
where ahij (x) = aij (x + hek ), we get
Z Z
ahij Di Dkh uDj vdx = − Dkh aij Di uDj v + qDk−h v dx
Ω Ω

≤ C kukH 1 (Ω1 ) + kf kL2 (Ω1 ) k∇vkL2 (Ω2 ) .
Take η ∈ C0∞ (Ω1 ) such that η(x) = 1 for x ∈ Ω0 and choose v = η 2 Dkh u. Then
Z Z
η aij Di Dk uDj Dk udx ≤ −2 ηahij Di Dkh u(Dj η)Dkh udx
2 h h h
Ω Ω

+C kukH 1 (Ω1 ) + kf kL2 (Ω1 ) kη∇Dkh ukL2 (Ω2 ) + 2kDkh u∇ηkL2 (Ω2 ) .
Using the ellipticity condition and Cauchy’s inequality, we obtain
Z Z
θ
|ηDkh ∇u|2 dx ≤ C |∇η|2 |Dkh u|2 dx + C kuk2H 1 (Ω00 ) + kf k2L2 (Ω00 ) .
2 Ω Ω
Hence
kηDkh ∇uk2L2 (Ω) ≤ C kuk2H 1 (Ω00 ) + kf k2L2 (Ω00 ) .
Since η = 1 on Ω0 , by using Theorem 2.34, we derive that Dk ∇u ∈ L2 (Ω0 ). This proves
that u ∈ H 2 (Ω0 ) and

(3.37) kukH 2 (Ω0 ) ≤ C kukH 1 (Ω00 ) + kf kL2 (Ω00 ) ,
where C depends on Ω0 . Combining with (3.36) it follows u satisfies (3.34).
Remark 3.12. (i) The result holds if the coefficients aij are only (locally) Lipschitz con-
tinuous in Ω, since the proof above only used the fact that Dkh aij is bounded.
(ii) The proof shows that Dk ∇u ∈ L2 (Ω0 ) as long as the function ϕ = Dk−h (η 2 Dkh u) is a
function in H 1 (Ω) with compact support in Ω even when Ω0 ∩ ∂Ω 6= ∅. This is used in the
boundary regularity theory later.
By using an induction argument, we can also get higher regularity for the solution.
Theorem 3.16. (Higher interior regularity) Let L be uniformly elliptic, with aij ∈
C k+1 (Ω), bi , c ∈ C k (Ω), and f ∈ H k (Ω). If u ∈ H 1 (Ω) is a weak solution of Lu = f , then
for any Ω0 ⊂⊂ Ω we have u ∈ H k+2 (Ω0 ) and
(3.38) kukH k+2 (Ω0 ) ≤ C (kukL2 (Ω) + kf kH k (Ω) )
where the constant C depends only on n, Ω0 , Ω and the coefficients of L.
Proof. Suppose we have proved this theorem for k. Now assume aij ∈ C k+2 (Ω), bi , c ∈
C k+1 (Ω), f ∈ H k+1 (Ω) and u ∈ H 1 (Ω) is a weak solution of Lu = f. Then, by the induction
k+2 k+3
assumption, u ∈ Hloc (Ω), with the estimate (3.38). We want to show u ∈ Hloc (Ω). Fix
0 00
Ω ⊂⊂ Ω ⊂⊂ Ω and a multiindex α with |α| = k + 1. Let
ũ = Dα u ∈ H 1 (Ω00 ).
Given any ṽ ∈ C0∞ (Ω00 ), let ϕ = (−1)|α| Dα ṽ be put into the identity B1 [u, ϕ] = (f, ϕ)L2 (Ω)
and perform some elementary integration by parts, and eventually we discover
B1 [ũ, ṽ] = (f˜, ṽ)L2 (Ω) ,
where
 
n n
X α  X α−β X
f˜ := Dα f − − (D aij Dβ uxi )xj + Dα−β bi Dβ uxi + Dα−β cDβ u .
β
β≤α,β6=α i,j=1 i=1
That is, ũ ∈ H 1 (Ω00 ) is a weak solution of Lũ = f˜ on Ω00 . (This is equivalent to differentiating
the equation Lu = f with Dα -operator.) We have f˜ ∈ L2 (Ω00 ), with, in light of the induction
assumption on the H k+2 (Ω00 )-estimate of u,
kf˜kL2 (Ω00 ) ≤ C(kf kH k+1 (Ω00 ) + kukH k+2 (Ω00 ) ) ≤ C(kf kH k+1 (Ω) + kukL2 (Ω) ).
Therefore, by Theorem 3.15, ũ ∈ H 2 (Ω0 ), with the estimate
kũkH 2 (Ω0 ) ≤ C(kf˜kL2 (Ω00 ) + kũkL2 (Ω00 ) ) ≤ C(kf kH k+1 (Ω) + kukL2 (Ω) ).
This exactly proves u ∈ H k+3 (Ω0 ) and the corresponding estimate (3.38) with k + 1.
3.5.2. Boundary Regularity. We now study the regularity up to the boundary. For this
purpose we need certain smoothness of the boundary. We have the following result.
Theorem 3.17. (Global H 2 -regularity) Assume in addition to the assumptions of The-
orem 3.15 that aij ∈ C 1 (Ω̄) and ∂Ω ∈ C 2 . If u ∈ H01 (Ω) is a weak solution to Lu = f , then
u ∈ H 2 (Ω), and
(3.39) kukH 2 (Ω) ≤ C(kukL2 (Ω) + kf kL2 (Ω) )
where the constant C depends only on n, kaij kW 1,∞ (Ω) , kbi kL∞ (Ω) , kckL∞ (Ω) and ∂Ω.
Proof. 1. First investigate the special case that Ω is a half ball

Ω = B(0, r) ∩ {xn > 0}.
and u ≡ 0 along plane {xn = 0} in the sense of trace. Set Ω0 = B(0, s) ∩ {xn > 0}, where
0 < s < r. Then select a smooth cutoff function ζ ∈ C0∞ (B(0, r)) with
0 ≤ ζ ≤ 1, ζ|B(0,s) ≡ 1.
So ζ ≡ 1 on Ω0 and ζ = 0 near the curved part of ∂Ω.
2. Now fix k ∈ {1, 2, · · · , n − 1}. For h > 0 sufficiently small, let
ϕ = Dk−h v, v = ζ 2 Dkh u.
Let us note carefully that if x ∈ Ω then
ζ 2 (x − hek )[u(x) − u(x − hek )] − ζ 2 (x)[u(x + hek ) − u(x)]
ϕ(x) = .
h2
3.5. Regularity 73
Since u = 0 along {xn = 0} and ζ = 0 near the curved portion of ∂Ω, we see ϕ ∈ H01 (Ω).
Therefore, we can use this ϕ as a test function in (3.35) as we did in the Step 2 above. The
end result is that
Dk u ∈ H 1 (Ω0 ) (k = 1, 2, · · · , n − 1)
with the estimate
X n
(3.40) kDkl ukL2 (Ω0 ) ≤ C(kf kL2 (Ω) + kukL2 (Ω) ).
k,l=1,k+l<2n
3. We must estimate kDnn ukL2 (Ω0 ) . We now use the fact that the assumption aij ∈ C 1
and the interior regularity imply the equation Lu = f is satisfied almost everywhere in
Ω; in this case we say u ∈ Hloc 2 (Ω) is a strong solution. Since the ellipticity implies
ann (x) ≥ θ > 0, we can actually solve Dnn u from the equation Lu = f in terms of Dij u
and Di u with i + j < 2n, i, j = 1, 2, · · · , n. This way we deduce the pointwise estimate
 
X n
|Dnn u| ≤ C  |Dij u| + |∇u| + |u| + |f | .
i,j=1,i+j<2n
Hence, by (3.40),
kukH 2 (Ω0 ) ≤ C(kf kL2 (Ω) + kukL2 (Ω) ).
4. Now it is standard to treat smooth domains by locally flattening the boundary. Since
∂Ω is C 2 , at each point x0 ∈ ∂Ω, we have a small ball B(x0 , r) and a C 2 map y = Φ(x),
with Φ(x0 ) = 0, that maps B(x0 , r) bijectively onto a domain in the y space such that
Φ(Ω ∩ B(x0 , r)) ⊂ {y ∈ Rn | yn > 0}.
We assume the inverse of this map is x = Ψ(y). Both Ψ and Φ are C 2 . Choose s > 0 so small
that the half-ball V := B(0, s) ∩ {yn > 0} lies in Φ(Ω ∩ B(x0 , r)). Set V 0 = B(0, s/2) ∩ {yn >
0}. Finally define
v(y) = u(Ψ(y)) (y ∈ V ).
1
Then v ∈ H (V ) and v = 0 on ∂V ∩ {yn = 0} (in the sense of trace). Moreover, u(x) =
v(Φ(x)) and hence
n
X
uxi (x) = vyk (Φ(x))Φkxi (x) (i = 1, 2, · · · , n).
k=1
5. We now show that v is a weak solution of a linear PDE M v = g in V. To find this

PDE, let I(y) = det ∂Ψ(y)∂y be the Jacobi matrix of x = Ψ(y); since I(y) 6= 0 and Ψ ∈ C 2 ,
−1
we have |I|, |I| ∈ C (V̄ ). Let ζ ∈ H 1 (V ) with supp ζ ⊂⊂ V and let ϕ = ζ/|I|. Then
1
ϕ ∈ H 1 (V ) with supp ϕ ⊂⊂ V. Let w(x) = ϕ(Φ(x)) for x ∈ Ω0 = Ψ(V ). Then w ∈ H 1 (Ω0 )

and supp w ⊂⊂ Ω0 . We use the weak formulation of Lu = f : B1 [u, w] = (f, w)L2 (Ω) and the
change of variable x = Ψ(y) to compute
Z Z
(3.41) (f, w)L2 (Ω) = f (x)w(x)dx = f (Ψ(y))ϕ(y)|I(y)|dy := (g, ζ)L2 (V ) ,
Ω0 V
where for g(y) = f (Ψ(y)). We also compute
Z

B1 [u, w] = aij (x)uxi (x)wxj (x) + bi (x)uxi w(x) + c(x)u(x)w(x) dx
Ω0
Z
= aij (x)vyk (Φ(x))Φkxi (x)ϕyl (Φ(x))Φlxj (x) + bi (x)vyk (Φ(x))Φkxi (x)w(x) + c(x)u(x)w(x) dx
Ω0
Z
= aij (Ψ(y))vyk (y)Φkxi (Ψ(y))ϕyl (y)Φlxj (Ψ(y))
V

+ bi (Ψ(y))vyk (y)Φkxi (Ψ(y))ϕ(y) + c(Ψ(y))v(y)ϕ(y) |I(y)|dy.
|I|y
Since ϕyl |I| = ζyl − |I|l ζ, we have
Z
(3.42) B1 [u, w] = ãij (y)vyk (y)ζyk (y) + b̃i (y)vyk (y)ζ(y) + c̃(y)v(y)ζ(y) dy := B̃[v, ζ],
V
where c̃(y) = c(Ψ(y)),

n
X
ãkl (y) = aij (Ψ(y))Φkxi (Ψ(y))Φlxj (Ψ(y)) (k, l = 1, 2, · · · , n),
i,j=1
and
n n
X X |I|yl
b̃k (y) = bi (x)Φkxi (x) − aij (Ψ(y))Φkxi (Ψ(y))Φlxj (Ψ(y)) (k = 1, 2, · · · , n).
|I|
i=1 i,j,l=1
By (3.41), (3.42), it follows that

B̃[v, ζ] = (g, ζ)L2 (V ) for all ζ ∈ H 1 (V ) with supp ζ ⊂⊂ V ;
hence, v ∈ H 1 (V ) is a weak solution of M v = g in V , where
n
X n
X
M v := − Dyl (ãkl (y)Dyk v) + b̃k (y)Dyk v + c̃(y)v.
k,l=1 k=1
6. We easily have that ãkl ∈ C 1 (V̄ ), b̃k , c̃ ∈ L∞ (V ). We now check that the operator
M is uniformly elliptic in V . Indeed, if y ∈ V and ξ ∈ Rn , then, again with x = Ψ(y),
n
X n X
X n n
X
ãkl (y)ξk ξl = ars (x)Φkxr Φlxs ξk ξl = ars (x)ηr (x)ηs (x) ≥ θ|η(x)|2 ,
k,l=1 r,s=1 k,l=1 r,s=1
where η(x) = (η1 (x), · · · , ηn (x)), with

n
X
ηr (x) = Φkxr (x)ξk (r = 1, 2, · · · , n).
k=1
That is, η(x) = ξDΦ(x). Hence ξ = η(x)DΨ(y) with y = Φ(x). So |ξ| ≤ C|η(x)| for some
constant C. This shows that
n
X
ãkl (y)ξk ξl ≥ θ|η(x)|2 ≥ θ0 |ξ|2
k,l=1
for some constant θ0 > 0 and all y ∈ V and ξ ∈ Rn . By the result with flat boundary in
Step 1, we have
kvkH 2 (V 0 ) ≤ C(kgkL2 (V ) + kvkL2 (V ) ).
Consequently, with O0 = Ψ(V 0 ), using the fact Φ, Ψ are of C 2 , we deduce
(3.43) kukH 2 (O0 ) ≤ C(kf kL2 (Ω) + kukL2 (Ω) ).
Note that x0 ∈ Ψ(B(0, s/2)) := G0 , which is an open set containing open set O0 .
3.5. Regularity 75
7. Since ∂Ω is compact, we can find finitely many open sets Oi0 ⊂ G0i (i = 1, 2, · · · , k)
such that ∂Ω ⊂ ∪ki=1 G0i . Then there exists a δ > 0 such that
k
[
F := {x ∈ Ω | dist(x, ∂Ω) ≤ δ} ⊂ Oi0 .
i=1
Then U = (Ω \ F ) ⊂⊂ Ω. By (3.43), we have
kukH 2 (F ) ≤ C(kf kL2 (Ω) + kukL2 (Ω) ).
By interior regularity,
kukH 2 (U ) ≤ C(kf kL2 (Ω) + kukL2 (Ω) ).
Combining these two estimates, we deduce (3.39).
Theorem 3.18. (Higher global regularity) Let L be uniformly elliptic, with aij ∈
C k+1 (Ω̄), bi , c ∈ C k (Ω̄), f ∈ H k (Ω), and ∂Ω ∈ C k+2 . Then a weak solution u of Lu = f
satisfying u ∈ H01 (Ω) belongs to H k+2 (Ω), and
(3.44) kukH k+2 (Ω) ≤ C(kukL2 (Ω) + kf kH k (Ω) )
where the constant C is independent of u and f . Furthermore, if the only weak solution
u ∈ H01 (Ω) of Lu = 0 is u ≡ 0, then, whenever Lu = f ,
(3.45) kukH k+2 (Ω) ≤ Ckf kH k (Ω)
where C is independent of u and f .
Proof. 1. As above, we first investigate the special case

Ω = B(0, 1) ∩ {xn > 0}.
Set Ωt = B(0, t) ∩ {xn > 0} for each 0 < t < 1. We intend to show by induction on k that
whenever u = 0 along {xn = 0} (always in the sense of trace), we have u ∈ H k+2 (Ωt ) and
(3.46) kukH k+2 (Ωt ) ≤ C(kukL2 (Ω) + kf kH k (Ω) ).
Suppose this is proved with k. Now assume aij ∈ C k+2 (Ω̄), bi , c ∈ C k+1 (Ω̄), f ∈ H k+1 (Ω),
and u is a weak solution of Lu = f in Ω, which vanishes along {xn = 0}. Fix any 0 < t <
r < 1. By induction assumption, u ∈ H k+2 (Ωr ), with
(3.47) kukH k+2 (Ωr ) ≤ C(kukL2 (Ω) + kf kH k (Ω) ).
k+3
Furthermore, according to the interior regularity, u ∈ Hloc (Ω).
2. Let α be any multiindex with |α| = k + 1 and αn = 0. Then ũ := Dα u ∈ H 1 (Ω)
and vanishes along {xn = 0}. (For example, this can be shown by induction on |α| using
the difference quotient operator Djh .) Furthermore, as in the proof of the interior higher
regularity theorem, ũ is a weak solution of Lũ = f˜ in Ω, with the same f˜ as above. This f˜
belongs to L2 (Ωr ) and
kf˜kL2 (Ω ) ≤ C(kf k k+1 + kukL2 (Ω) ).
r H (Ω)
Consequently, ũ ∈ H 2 (Ωt ), with

kũkH 2 (Ωt ) ≤ C(kf˜kL2 (Ωr ) + kũkL2 (Ωr ) ) ≤ C(kf kH k+1 (Ω) + kukL2 (Ω) ).
This proves
(3.48) kDβ ukL2 (Ωt ) ≤ C(kf kH k+1 (Ω) + kukL2 (Ω) )
for all β with |β| = k + 3 and βn = 0, 1, 2.
3. We need to extend (3.48) to all β with |β| = k + 3. Fix k, we prove (3.48) for all β
by induction on j = 0, 1, · · · , k + 2 with βn ≤ j. We have already shown it for j = 0, 1, 2.
Assume we have shown it for j. Now assume β with |β| = k + 3 and βn = j + 1. Let us
k+3
write β = γ + δ, for δ = (0, · · · , 0, 2) and so |γ| = k + 1. Since u ∈ Hloc (Ω) and Lu = f in
Ω, we have Dγ Lu = Dγ f a.e. in Ω. Now
Dγ Lu = ann Dβ u + R,
where R is the sum of terms involving at most j derivatives of u with respect to xn and
at most k + 3 derivatives in all. Since ann ≥ θ, we can solve Dβ u in terms of R and Dγ f ;
hence,
kDβ ukL2 (Ωt ) ≤ C(kf kH k+1 (Ω) + kukL2 (Ω) ).
By induction, we deduce (3.48), which proves
kukH k+3 (Ωt ) ≤ C(kf kH k+1 (Ω) + kukL2 (Ω) ).
This estimate in turn completes the induction process on k, begun in step 2. This proves
(3.46).
4. As above, we can cover the domain Ω̄ by finitely many small balls and use the method
of flattening the boundary to eventually deduce (3.44).
5. We prove the last statement (3.45) for k = 0; the more general case is similar. In
view of (3.44), it suffices to show that
kukL2 (Ω) ≤ Ckf kL2 (Ω)
If to the contrary this inequality is false, there would exist sequences un ∈ H 2 (Ω) ∩ H01 (Ω)
and fn ∈ L2 (Ω) for which kun k2 = 1 and kfn k2 → 0. By (3.44) we have kun k2,2 ≤ C. Thus
we can assume that un converges weakly to u in H 2 (Ω) and strongly in L2 (Ω). For fixed
v ∈ H01 (Ω), the functional l(u) = B[u, v] ∈ (H 2 (Ω))∗ , and so by passing to the limit in
Z
l(un ) = B[un , v] = fn vdx for all v ∈ H01 (Ω)
Ω
we see that B[u, v] = 0 for all v ∈ H01 (Ω) and thus u is a weak solution of Lu = 0. Hence,
u ≡ 0 by weak uniqueness. This contradicts kuk2 = limn→∞ kun k2 = 1.
Finally, we iterate this regualrity theorem to obtain

Theorem 3.19. (Infinite smoothness) Let L be uniformly elliptic, with aij , bi , c, f ∈
C ∞ (Ω̄), and ∂Ω ∈ C ∞ . Then a weak solution u ∈ H01 (Ω) of Lu = f belongs to C ∞ (Ω̄).
3.6. Regularity for Linear Systems*

In this extra section, we study the regularity problem for weak solutions of elliptic second-
order linear systems. The methods include some direct generalization of linear equations
and some new techniques.
Let A(x, ξ) = A(x) ξ be a linear matrix function of ξ given by
(A(x, ξ))iα = (A(x)ξ)iα = Aαβ j
ij (x)ξβ ,
where Aαβ ∞
ij ∈ L (Ω). Consider the linear partial differential system
−Dα Aαβ j
= g i − Dα fαi , i = 1, 2, · · · , N,

(3.49) ij (x) Dβ u
3.6. Regularity for Linear Systems* 77
where we assume g i , fαi ∈ L2loc (Ω). We also write this system as

− div(A(x)Du) = g − div f,
where g = (g i ), f = (fαi ). Recall that u ∈ Hloc
1 (Ω; RN ) is a weak solution of (3.49) if
Z Z
(3.50) A(x)Du · Dφ(x) dx = (g(x) · φ(x) + f (x) · Dφ(x)) dx
Ω Ω
holds for all φ ∈ C0∞ (Ω; RN ). Since Aαβ ∞

ij ∈ L (Ω), the test function φ in (3.50) can be
chosen in H01 (Ω0 ; RN ) for any subdomain Ω0 ⊂⊂ Ω.
The regularity for system (3.49) relies on certain ellipticity conditions. We shall assume
one of the following conditions holds: with some constant ν > 0,
(H1) Aαβ i j 2
ij (x) ξα ξβ ≥ ν |ξ| .
(H2) Aαβ αβ i j 2 2
ij are constants, Aij pα pβ q q ≥ ν |p| |q| .
(H3) Aαβ αβ i j 2 2
ij ∈ C(Ω̄), Aij (x) pα pβ q q ≥ ν |p| |q| .
Under such a condition, we shall have the following Gårding inequality holds:
Z Z Z
2
(3.51) A(x)Dψ · Dψ ≥ ν0 |Dψ| − ν1 |ψ|2 , ∀ ψ ∈ H01 (BR ; RN ),
BR BR BR
where ν0 > 0, ν1 ≥ 0 are constants. For, under the hypothesis (H1) or (H2) the Gårding
inequality (3.51) holds with ν0 = ν, ν1 = 0, and under (H3) the inequality (3.51) also holds
(see Theorem 3.8).
3.6.1. Caccioppoli-type Estimates. Assume u ∈ Hloc 1 (Ω; RN ) is a weak solution of
(3.49). Almost all the estimates pertaining to regularity of u are derived using test functions
of the form φ = η (u−λ), where η is a cut-off function which belongs to W01,∞ (Ω0 ) for certain
Ω0 ⊂⊂ Ω. Let Bρ ⊂⊂ BR ⊂⊂ Ω be concentric balls with center a ∈ Ω. Let

1
 if 0 ≤ t ≤ ρ,
R−t
θ(t) = R−ρ if ρ ≤ t ≤ R,

0 if t > R.

Let ζ = ζρ,R (x) = θ(|x − a|). Then ζ ∈ W01,∞ (Ω) with supp ζ ⊆ B̄R and
χρ,R
(3.52) 0 ≤ ζ ≤ 1, ζ|Bρ ≡ 1, |Dζ| ≤ ,
R−ρ
where χρ,R = χBR \Bρ (x) is the characteristic function of BR \ Bρ . Define
ψ = ζ (u − λ), φ = ζ 2 (u − λ) = ζ ψ.
Then ψ, φ ∈ H01 (BR ; RN ) and
Dφ = ζDψ + ψ ⊗ Dζ, Dψ = ζDu + (u − λ) ⊗ Dζ.
Using φ as a test function in (3.50) yields
Z Z
(g · φ + f · Dφ) = A(x)Du · Dφ
BR BR
Z Z
= A(x)Du · ζDψ + A(x)Du · ψ ⊗ Dζ.
BR BR
Note that
A(x)Dψ · Dψ = A(x)Du · ζDψ + A(x)(u − λ) ⊗ Dζ · Dψ,
A(x)Du · ψ ⊗ Dζ = A(x)ζDu · (u − λ) ⊗ Dζ
= A(x)Dψ · (u − λ) ⊗ Dζ
− A(x)(u − λ) ⊗ Dζ · (u − λ) ⊗ Dζ.
Therefore, it follows that
Z Z
A(x)Dψ · Dψ = A(x)(u − λ) ⊗ Dζ · Dψ
BR BR
Z Z
+ (g · φ + f · Dφ) − A(x)Dψ · (u − λ) ⊗ Dζ
BR BR
Z
+ A(x)(u − λ) ⊗ Dζ · (u − λ) ⊗ Dζ.
BR
(Note that the first and third terms on the righthand side would cancel out if A(x)ξ · η is
symmetric in ξ, η.) Then, by (3.51), it follows that
Z Z Z
2
ν0 |Dψ| ≤ A(x)Dψ · Dψ + ν1 |ψ|2
BR B BR
ZR
χρ,R |u − λ|
Z Z
≤ g·φ + |f | · |Dψ| + |f | ·
BR BR BR R−ρ
χρ,R |u − λ| χρ,R |u − λ|2
Z Z Z
+ C |Dψ| · +C + ν1 |u − λ|2 .
BR R−ρ BR (R − ρ)2 BR
Using the Cauchy inequality with , we deduce
"Z #
|u − λ|2
Z Z Z Z
2 2 2
(3.53) |Dψ| ≤ C 2
+ g·φ + |f | + ν1 |u − λ| .
BR BR \Bρ (R − ρ) BR BR BR
Since ψ|Bρ = u − λ, this last estimate (3.53) proves the following theorems.
1 (Ω; RN ) be a weak solution of (3.49). Assume either condition
Theorem 3.20. Let u ∈ Hloc
(H1) or (H2) holds. Then
"Z #
Z
|u − λ| 2 Z Z
(3.54) |Du|2 ≤ C 2
+ g · ζ 2 (u − λ) + |f |2
Bρ BR \Bρ (R − ρ) BR BR
for all concentric balls Bρ ⊂⊂ BR ⊂⊂ Ω and constants λ ∈ RN , where ζ = ζρ,R and C > 0
is a constant depending on the L∞ -norm of Aαβ
ij .
Theorem 3.21. Assume condition (H3) holds. Then

"Z #
|u − λ|2
Z Z Z
2 2 2 2
(3.55) |Du| ≤ C 2
+ g · ζ (u − λ) + (|f | + |u − λ| )
Bρ BR \Bρ (R − ρ) BR BR
for all concentric balls Bρ ⊂⊂ BR ⊂⊂ Ω and constants λ ∈ RN .

1 (Ω; RN ) be a weak solution of (3.49). Assume either condition
Corollary 3.22. Let u ∈ Hloc
(H1) or (H2) holds. Then
"Z #
|u − λ|2
Z Z
2 2
(3.56) |Du| ≤ C + |u − λ| · |g| + |f |
BR/2 BR \BR/2 R2 BR
for all balls BR ⊂⊂ Ω and constants λ ∈ RN .

Remark 3.13. 1) The estimates (3.54), (3.55) and (3.56) are usually referred to as the
Caccioppoli-type inequalities or Caccioppoli estimates.
2) In both (3.54), (3.55), we keep the term BR g · ζ 2 (u − λ) in the estimates. We shall
R
see later that this term needs a special consideration when we deal with higher regularity
for weak solutions, especially when g is of the form of quotient difference.
As an application of these Caccioppoli estimates, we prove the following results of

Liouville’s theorem.
1 (Rn ; RN ) is a weak solution of
Corollary 3.23. Suppose u ∈ Hloc
(3.57) −Dα (Aαβ j

ij (x)Dβ u ) = 0,
where coefficients Aαβ 2 n

ij (x) satisfy (H1) or (H2). If |Du| ∈ L (R ), then u is a constant.
Proof. By (3.56), it follows that

Z Z
C2
|Du| ≤ 2 |u − λ|2 .
BR/2 R BR \BR/2
We choose Z
1
λ= u(x) dx.
|BR \ BR/2 | BR \BR/2
Then the Poincaré inequality shows that

Z Z
2 2
|u − λ| ≤ c(n) R |Du|2 .
BR \BR/2 BR \BR/2
Therefore Z Z
2
|Du| ≤ C |Du|2 .
BR/2 BR \BR/2
Adding C B |Du|2 to both sides of this inequality (a.k.a. the hole-filling technique of
R
R/2
Widman), we obtain
Z Z
2 C
|Du| ≤ |Du|2 .
BR/2 C + 1 BR
Letting R → ∞ we have
Z Z
C
|Du|2 dx ≤ |Du|2 dx.
Rn C +1 Rn
C
|Du|2 = 0 and thus Du ≡ 0; hence u ≡ constant.
R
Since C+1 < 1 we have Rn
Corollary 3.24. Assume either condition (H1) or (H2) holds. Then any bounded weak
1 (R2 ; RN ) to (3.57) for n = 2 must be constant.
solution u ∈ Hloc
Proof. Let |u| ≤ M ; then by the Caccioppoli inequality (3.56) with λ = 0 we have
Z
|Du|2 dx ≤ CM < ∞, ∀ R > 0.
BR/2
This implies |Du| ∈ L2 (R2 ); hence by the result above, u is a constant.

3.6.2. Regularity – Difference Quotient Method. This is completely similar to the

scalar case studied above.
1 (Ω; RN ) is a weak solution of linear system
Assume u ∈ Hloc
− div(A(x) Du) = g − div f.
This means
Z Z Z
A(x) Du(x) · Dφ(x) dx = g(x) · φ(x) dx + f (x) · Dφ(x) dx
Ω Ω Ω
holds for all φ ∈ W01,2 (Ω0 ; RN ). If 0 < |h| < dist(Ω0 ; ∂Ω) we have
Z
A(x + hes ) Du(x + hes ) · Dφ(x) dx
Ω
Z Z
= g(x + hes ) · φ(x) dx + f (x + hes ) · Dφ(x) dx.
Ω Ω
Subtract two equations and divide by h to get
Z Z
h
A(x + hes ) DDs u · Dφ = Dsh g(x) · φ(x) dx
Ω
ZΩ Z
+ Dsh f (x) · Dφ(x) dx − Dsh A(x) Du(x) · Dφ(x) dx.
Ω Ω
This shows that v = Dsh u is a weak solution of system

(3.58) − div(A(x + hes ) Dv) = Dsh g − div(Dsh f ) + div(Dsh A Du) on Ω0 .
Assume that Gårding’s inequality (3.51) holds. Then we can invoke the estimate (3.53)
with λ = 0, ρ = R/2 to obtain
Z hZ Z
2 1 h 2
(3.59) |Dψ| ≤ C 2
|Ds u| + Dsh g · φ
BR BR R BR
Z
i
+ |Dsh f |2 + ν1 |Dsh u|2 + |Dsh A|2 |Du|2 ,
BR
where ψ = ζ Dsh u, φ= ζ 2 Dsh u

and ζ = ζR/2,R is defined as before. Note that
Z Z Z
− h
Ds g · φ = h
Ds g · φ = g · Ds−h φ
BR Ω Ω
Z Z
−h
= g · ζ(x − hes ) Ds ψ + g · ψ Ds−h ζ
Ω Ω
≡ I + II.
We estimate I, II as follows.
Z Z Z
−h −h 2
|I| ≤ |g| · |Ds ψ| ≤ |Ds ψ| + C |g|2
Ω0 Ω0 Ω0
Z Z Z Z
2 2 2
≤ |Ds ψ| + C |g| ≤ |Dψ| + C |g|2 ,
Ω Ω0 BR Ω0
where Ω0 ⊂⊂ Ω is a domain containing B̄R .

Z Z Z
h −h C 2
|II| ≤ |g| |Ds u| |Ds ζ| ≤ 2 |g| + C |Dsh u|2 .
BR R Ω0 BR
Combining these estimates with (3.59) yields

Z Z
2
|Dsh u|2 + |g|2 + |Dsh f |2 + |Dsh A|2 |Du|2 .

|Dψ| ≤ C(R)
BR Ω0
Since ψ = Dsh u and Dψ = Dsh Du on BR/2 we have

Z Z
|Dsh Du|2 ≤ C(R) |Dsh u|2 + |g|2 + |Dsh f |2 + |Dsh A|2 |Du|2 .

(3.60)
BR/2 Ω0
1 (Ω; MN ×n ) and A(x) is Lipschitz continuous with Lipschitz

Finally, if we assume f ∈ Hloc
constant K then Z Z
|Dsh f |2 ≤ |Df |2 , |Dsh A(x)| ≤ K,
Ω0 Ω00
where Ω0 ⊂⊂ Ω00 ⊂⊂ Ω, and hence by (3.60) we have
Z
|Dsh Du|2 (x) dx ≤ M < ∞ ∀ |h| << 1,
BR/2
and thus Ds Du exists and belongs to L2 (BR/2 ; MN ×n ) for all s = 1, 2, · · · , n. This implies
2 (Ω; RN ). Therefore, we have proved the following theorem.
u ∈ Hloc
Theorem 3.25. Suppose A ∈ C(Ω̄) is Lipschitz continuous and the Gårding inequality
1 (Ω; MN ×n ) and u ∈ H 1 (Ω; RN ) is a weak solution
(3.51) holds. If g ∈ L2loc (Ω; RN ), f ∈ Hloc loc
of the system
− div(A(x) Du) = g − div f
2 (Ω; RN ).
then u ∈ Hloc
The following higher regularity result can be proved by the standard bootstrap method.
1 (Ω; RN ) is a weak solution of the system
Theorem 3.26. Suppose u ∈ Hloc
with A ∈ C k,1 (Ω̄) (that is, Dk A is Lipschitz continuous) satisfying the Gårding inequality
(3.51) and g ∈ Hlock (Ω; RN ), f ∈ H k+1 (Ω; MN ×n ). Then u ∈ H k+2 (Ω; RN ).
loc loc
Proof. Let ψ ∈ C0∞ (Ω; RN ); then we use φ = Ds ψ as a test function for the system to
obtain Z Z Z
Ds (A(x) Du) · Dψ dx = Ds g · ψ + Ds f · Dψ.
Ω Ω Ω
Since Ds (A(x) Du) = (Ds A) Du + A(x) DDs u we thus have
Z Z Z

A(x) D(Ds u) · Dψ = Ds g · ψ + Ds f − (Ds A) Du · Dψ.
Ω Ω Ω
This shows v = Ds u ∈ 1 (Ω; RN )

Hloc is a weak solution of
− div(A(x) Dv) = Ds g − div(Ds f − (Ds A) Du),
2 (Ω; RN ); that is, u ∈ H 3 (Ω; RN ). The result for general k then follows
and hence v ∈ Hloc loc
from induction.
Remark 3.14. Note that if A, g, f are all of C ∞ then any weak solution u ∈ Hloc
1 (Ω; RN )
∞ N
must be in C (Ω; R ). We also have the following result.
1 (Ω; RN ) be a weak solution of

Theorem 3.27. Assume (H2) holds. Let u ∈ Hloc
(3.61) −Dα (Aαβ j
ij Dβ u ) = 0, i = 1, 2, ..., N.
k (Ω; RN ) for all k = 1, 2, ... and
Then u ∈ Hloc
kukH k (BR/2 ;RN ) ≤ C(k, R) kukL2 (BR ;RN )
for any ball BR ⊂⊂ Ω.
Proof. By the Caccioppoli-type inequality, we have for any weak solution u of system (3.61)
Z Z
2 C
|Du| dx ≤ |u|2 dx.
Bρ (R − ρ)2 BR
k,2
The regularity result shows that u ∈ Wloc (Ω; RN ) for all k and then it follows that any
derivative Dk u is also a weak solution of (3.61). Therefore, the conclusion will follow from
a successive use of the above Caccioppoli inequality with a finite number of R/2 = ρ1 <
ρ2 < · · · < ρK = R.
In exactly the same way as the scalar equations, we obtain the global H 2 -regularity.
Theorem 3.28. (Global H 2 -regularity) Let ∂Ω be C 2 . Suppose A ∈ C 1 (Ω̄) and the
Gårding inequality (3.51) holds. If g ∈ L2 (Ω; RN ), f ∈ H 1 (Ω; MN ×n ) and u ∈ H 1 (Ω; RN )
is a weak solution of the system
then u ∈ H 2 (Ω; RN ) and we have
(3.62) kukH 2 (Ω) ≤ C(kgkL2 (Ω) + kf kH 1 (Ω) + kukL2 (Ω) ),
where the constant C depends only on Ω and the coefficients.
3.6.3. Morrey and Campanato Spaces. Let Ω ⊂ Rn be a bounded open domain. For
x ∈ Rn , ρ > 0 let
Ω(x, ρ) = {y ∈ Ω | |y − x| < ρ}.
Definition 3.15. For 1 ≤ p < ∞, λ ≥ 0 we define the Morrey space Lp,λ (Ω; RN ) by
 

 Z 

p,λ N p N −λ p
L (Ω; R ) = u ∈ L (Ω; R ) sup ρ |u(x)| dx < ∞ .

 a∈Ω Ω(a,ρ) 

0<ρ<diamΩ
We define a norm by
Z !1/p
−λ p
kukLp,λ (Ω;RN ) = sup ρ |u(x)| dx .
a∈Ω Ω(a,ρ)
0<ρ<diamΩ
Theorem 3.29. Lp,λ (Ω; RN ) is a Banach space.

Lemma 3.30. (Lebesgue Differentiation Theorem) If v ∈ L1loc (Ω) then
Z
lim − |v(x) − v(y)| dy = 0
ρ→0 Bρ (x)
for almost every x ∈ Ω.

Theorem 3.31. (a) If λ > n then Lp,λ (Ω; RN ) = {0}.

(b) Lp,0 (Ω; RN ) ∼
= Lp (Ω; RN ); Lp,n (Ω; RN ) ∼
= L∞ (Ω; RN ).
n−λ n−µ
(c) If 1 ≤ p ≤ q < ∞, p ≥ q , then Lq,µ (Ω; RN ) ⊂ Lp,λ (Ω; RN ).
Proof. (a) By Lebesgue’s differentiation theorem,

Z
(3.63) |u(a)| = lim − |u(x)| dx, ∀ a.e. a ∈ Ω.
ρ→0 Ω(a,ρ)
Now, by Hölder’s inequality,
Z Z !1/p
λ−n
(3.64) − |u(x)| dx ≤ − |u(x)|p dx ≤Cρ p kukLp,n (Ω;RN ) .
Ω(a,ρ) Ω(a,ρ)
If λ > n, letting ρ → 0 we have u(a) = 0 for almost every a ∈ Ω; thus u ≡ 0.

(b) That Lp,0 (Ω; RN ) ∼
= Lp (Ω; RN ) easily follows from the definition. We prove Lp,n (Ω; RN ) ∼
=
L∞ (Ω; RN ). If u ∈ L∞ (Ω; RN ), then
Z
ρ−n |u(x)|p dx ≤ C kukp∞
Ω(a,ρ)
so that kukLp,n ≤ C kuk∞ . Suppose now u ∈ Lp,n (Ω; RN ). Then by (3.63), (3.64)
Z
|u(a)| = lim − |u| ≤ C kukLp,n (Ω;RN ) .
ρ→0 Ω(a,ρ)
Hence kukL∞ (Ω;RN ) ≤ C kukLp,n (Ω;RN ) . Therefore Lp,n (Ω; RN ) ∼

= L∞ (Ω; RN ).
(c) We first note that u ∈ Lp,λ (Ω; RN ) if and only if Ω(a,ρ) |u(x)|p dx ≤ C ρλ for all
R
a ∈ Ω and 0 < ρ < ρ0 = min{1, diamΩ}. Suppose u ∈ Lq,µ (Ω; RN ). Then, by Hölder’s
inequality, for all a ∈ Ω, 0 < ρ < ρ0 < 1,
!p
Z Z q
p
1−
|u|p dx ≤ |Ω(a, ρ)| q |u|q dx
Ω(a,ρ) Ω(a,ρ)
n− np p
≤ Cρ q (kukqLq,µ (Ω;RN ) ρµ ) q
µp
+n− np
≤ Cρ q q kukpLq,µ (Ω;RN )
≤ C ρλ kukpLq,µ (Ω;RN ) ,
µp np
where we have used the assumption q +n− q ≥ λ and the fact 0 < ρ < 1. Therefore,
u ∈ Lp,λ (Ω; RN ).
Definition 3.16. For 1 ≤ p < ∞, λ ≥ 0 we define the Campanato space Lp,λ (Ω; RN ) by
 

 Z 

p,λ N p N −λ p
L (Ω; R ) = u ∈ L (Ω; R ) sup ρ |u − ua,ρ | dx < ∞ ,

 a∈Ω Ω(a,ρ) 

0<ρ<diamΩ
where ua,ρ is the average of u on Ω(a, ρ). Define the seminorm and norm by
Z !1/p
−λ p
[u]Lp,λ (Ω;RN ) = sup ρ |u − ua,ρ | dx ,
a∈Ω Ω(a,ρ)
0<ρ<diamΩ
kukLp,λ (Ω;RN ) = kukLp (Ω;RN ) + [u]Lp,λ (Ω;RN ) .
Remark 3.17. For the estimates on linear elliptic systems of second-order partial differen-
tial equations, only spaces L2,λ (Ω; RN ) and L2,λ (Ω; RN ) are needed. The spaces with other
p ≥ 1 are useful for nonlinear problems; however, we shall not study the nonlinear problems
in this course.
For 0 < α ≤ 1, we define the Hölder space C 0,α (Ω̄; RN ) by

n o
C 0,α (Ω̄; RN ) = v ∈ L∞ (Ω; RN ) |v(x) − v(y)| ≤ C |x − y|α , ∀ x, y ∈ Ω
and define the seminorm and norm by
|v(x) − v(y)|
[v]C 0,α (Ω;RN ) = sup ,
x, y∈Ω |x − y|α
x6=y
kvkC 0,α (Ω;RN ) = kvkL∞ (Ω;RN ) + [v]C 0,α (Ω;RN ) .
Theorem 3.32. Both Lp,λ (Ω; RN ) and C 0,α (Ω̄; RN ) are Banach spaces.
Theorem 3.33. (a) For any p ≥ 1, λ ≥ 0, Lp,λ (Ω; RN ) ⊂ Lp,λ (Ω; RN ).
(b) For any 0 < α ≤ 1, C 0,α (Ω̄; RN ) ⊂ Lp,n+pα (Ω; RN ).
Proof. (a) Note that

Z !1/p
p
|v(y) − va,ρ | dx ≤ kvkLp (Ω(a,ρ) + |va,ρ | · |Ω(a, ρ)|1/p .
Ω(a,ρ)
It turns out that we can exactly estimate the two terms on the right-hand side by
kvkLp (Ω(a,ρ)) ≤ ρλ/p kvkLp,λ (Ω;RN ) ,
|va,ρ | · |Ω(a, ρ)|1/p ≤ ρλ/p kvkLp,λ (Ω;RN )

so that it follows that
[v]Lp,λ (Ω;RN ) ≤ 2 kvkLp,λ (Ω;RN ) .
Hence Lp,λ (Ω; RN ) ⊂ Lp,λ (Ω; RN ).
(b) Assume v ∈ C 0,α (Ω̄; RN ). Then
Z
|v(x) − va,ρ | = | − (v(x) − v(y)) dy|
Ω(a,ρ)
Z
≤ [v]C 0,α (Ω̄;RN ) · − |x − y|α dy
Ω(a,ρ)
≤ [v]C 0,α (Ω̄;RN ) · (2ρ)α .
Hence Z
|v(x) − va,ρ |p dx ≤ C [v]pC 0,α (Ω̄;RN ) · ρn+pα
Ω(a,ρ)
and hence
(3.65) [v]Lp,n+pα (Ω;RN ) ≤ C [v]C 0,α (Ω̄;RN ) .
In order to study the properties of Campanato functions, we need a condition on domain

Ω introduced by Campanato.
Definition 3.18. We say that Ω ⊂ Rn is of type A if there exists a constant A > 0 such
that
(3.66) |Ω(a, ρ)| ≥ A ρn , ∀ a ∈ Ω, 0 < ρ < diamΩ.
This condition excludes that Ω may have sharp outward cusps; for instance, all Lipschitz
domains are of type A.
Lemma 3.34. Assume Ω is of type A and u ∈ Lp,λ (Ω; RN ). Then for any 0 < r < R <
∞, a ∈ Ω it follows that
− p1 λ
−n
|ua,R − ua,r | ≤ 2 A Rp r p · [u]Lp,λ (Ω;RN ) .
Proof.
1
|ua,R − ua,r | · |Ω(a, r)| p = kua,R − ua,r kLp (Ω(a,r))
≤ ku − ua,R kLp (Ω(a,r)) + ku − ua,r kLp (Ω(a,r))
≤ ku − ua,R kLp (Ω(a,R)) + ku − ua,r kLp (Ω(a,r))
λ λ
≤ [u]Lp,λ (Ω;RN ) R p + [u]Lp,λ (Ω;RN ) r p
λ
≤ 2 [u]Lp,λ (Ω;RN ) R p .
Hence the lemma follows from the assumption that |Ω(a, r)| ≥ A rn .
Theorem 3.35. If Ω is of type A then Lp,λ (Ω; RN ) ∼
= Lp,λ (Ω; RN ) for 0 ≤ λ < n.
Proof. We only need to show Lp,λ (Ω; RN ) ⊂ Lp,λ (Ω; RN ). Let u ∈ Lp,λ (Ω; RN ). Given any
a ∈ Ω, 0 < ρ < diamΩ, we have
kukLp (Ω(a,ρ)) ≤ ku − ua,ρ kLp (Ω(a,ρ)) + kua,ρ kLp (Ω(a,ρ))
λ n
≤ [u]Lp,λ (Ω;RN ) ρ p + C |ua,ρ | ρ p .
We choose an integer k large enough so that Ω(a, 2k ρ) = Ω. By Lemma 3.34, we have
k−1
X
|ua,ρ | ≤ |ua,2k ρ | + |ua,2j+1 ρ − ua,2j ρ |
j=0
k−1
− p1 λ
−n
X
≤ |uΩ | + 2A (2j+1 ρ) p (2j ρ) p · [u]Lp,λ (Ω;RN )
j=0
k
λ−n X
≤ |uΩ | + C ρ p [u]Lp,λ (Ω;RN ) · 2j(λ−n)/p
j=0
λ−n
≤ |uΩ | + C [u]Lp,λ (Ω;RN ) ρ p ,
where uΩ is the average of u on Ω and therefore |uΩ | ≤ C(Ω) kukLp (Ω;RN ) . Combining these
estimates, we deduce
λ n
kukLp (Ω(a,ρ)) ≤ C [u]Lp,λ (Ω;RN ) ρ p + C kukLp (Ω;RN ) ρ p
λ
and, by dividing both sides by ρ p and noting λ < n,
−λ
ρ p kukLp (Ω(a,ρ)) ≤ C [u]Lp,λ (Ω;RN ) + C(Ω) kukLp (Ω;RN ) .
This proves
kukLp,λ (Ω;RN ) ≤ C(Ω) kukLp,λ (Ω;RN ) .

Remark 3.19. Note that Lp,n (Ω; RN ) ∼ 6 Lp,n (Ω; RN ) ∼
= = L∞ (Ω; RN ). For example, let
p = λ = 1, n = N = 1 and Ω = (0, 1). Then u = ln x is in L1,1 (0, 1) but not in L1,1 (0, 1) ∼
=
∞ p,n N ∼ N
L (0, 1). In fact, L (Ω; R ) = BM O(Ω; R ), which is called the John-Nirenberg space.
Theorem 3.36. (Campanato ’63) Let Ω be of type A. Then for n < λ ≤ n + p,
λ−n
Lp,λ (Ω; RN ) ∼
= C 0,α (Ω̄; RN ), α = ,
p
whereas for λ > n + p we have Lp,λ (Ω; RN ) = {constants}.
Proof. 1. Assume λ > n and v ∈ Lp,λ (Ω; RN ). For any x ∈ Ω and R > 0 we define
ṽ(x) = lim vx, R .
k→∞ 2k
We claim ṽ is well-defined and independent of R > 0. We first show the limit defining ṽ(x)
exists. We need to show the sequence {vx, R } is Cauchy. For h > k we have, by Lemma
2k
3.34,
h−1
X
|vx, R − vx, R | ≤ |vx, R − vx, R |
2h 2k 2j 2j+1
j=k
h−1
λ−n j(n−λ)
− p1
X
≤ 2A [v]Lp,λ (Ω;RN ) R p · 2 p ,
j=k
which, since λ > n, tends to zero if k, h → ∞. Therefore {vx, R } is Cauchy and the limit
2k
ṽ(x) exists. Also in the inequality above, if k = 0 and h → ∞ we also deduce
λ−n
(3.67) |vx,R − ṽ(x)| ≤ C [v]Lp,λ (Ω;RN ) · R p .
We now prove ṽ(x) is independent of R > 0. This follows easily since by Lemma 3.34
lim |vx, R − vx, r | = 0.
k→∞ 2k 2k
2. By Lebesgue’s differentiation theorem, we also have ṽ(x) = v(x) for almost every
x ∈ Ω. Therefore ṽ = v in Lp,λ (Ω; RN ). We claim ṽ ∈ C 0,α (Ω̄; RN ), where α = λ−n
p . To
show this, let x, y ∈ Ω and x 6= y. Let R = |x − y|. By (3.67) it follows that
|ṽ(x) − ṽ(y)| ≤ |ṽ(x) − vx,2R | + |ṽ(y) − vy,2R | + |vx,2R − vy,2R |
≤ C [v]Lp,λ (Ω;RN ) · Rα + |vx,2R − vy,2R |.
We need to estimate |vx,2R −vy,2R |. To this end, let S = Ω(x, 2R)∩Ω(y, 2R). Then Ω(x, R) ⊂
S and hence
|S| ≥ |Ω(x, R)| ≥ A Rn .
On the other hand, we have
1
|S| p · |vx,2R − vy,2R | = kvx,2R − vy,2R kLp (S)
≤ kvx,2R − vkLp (S) + kvy,2R − vkLp (S)
≤ kvx,2R − vkLp (Ω(x,2R)) + kvy,2R − vkLp (Ω(y,2R))
≤ 2 [v]Lp,λ (Ω;RN ) · (2R)λ/p .
Combining the above two estimates we have

λ−n
|vx,2R − vy,2R | ≤ C [v]Lp,λ (Ω;RN ) · R p = C [v]Lp,λ (Ω;RN ) · Rα
and hence
|ṽ(x) − ṽ(y)| ≤ C [v]Lp,λ (Ω;RN ) · |x − y|α .
This shows
[ṽ]C 0,α (Ω̄;RN ) ≤ C [v]Lp,λ (Ω;RN ) .
Finally, observe that, by (3.67) with R = diamΩ,
kṽkL∞ (Ω;RN ) ≤ |vΩ | + C [v]Lp,λ (Ω;RN ) · Rα

≤ C(Ω) kvkLp (Ω;RN ) + C(Ω) [v]Lp,λ (Ω;RN )
= C(Ω) kvkLp,λ (Ω;RN ) .
3. We have thus proved that if λ > n then every v ∈ Lp,λ (Ω; RN ) has a representation
ṽ which belongs to C 0,α (Ω̄; RN ) with α = (λ − n)/p. If λ > n + p then α > 1 and any u ∈
C 0,α (Ω̄; RN ) must be a constant (why?). The proof of Campanato’s theorem is complete.
In order to use the Campanato spaces for elliptic systems, we also need some local
version of these spaces. To disperse some technicalities, we prove the following lemma.
Lemma 3.37. Let p = 1, 2 and u ∈ Lploc (Ω; RN ). Then the map E 7→ E |u − uE |p is

R
nondecreasing in subsets E ⊂⊂ Ω.
Proof. We prove the case p = 2 first. Let E ⊂ F ⊂⊂ Ω. Then

Z Z
|u − uE |2 = |u − uF + uF − uE |2
E
ZE Z
= |u − uF | + 2 (uF − uE ) · (u − uF ) + |E| · |uF − uE |2
2
ZE E
= |u − uF |2 − |E| · |uF − uE |2
E
Z
≤ |u − uF |2 .
F
We now prove the case p = 1. Note that

Z Z
|u − uE | = |u − uF + uF − uE |
E E
Z Z
≤ |u − uF | + |uF − uE |
E E
Z Z
= |u − uF | − |u − uF | + |E| · |uF − uE |.
F F \E
Thus we need to prove

Z
(3.68) |E| · |uF − uE | ≤ |u − uF |.
F \E
Note that, by Jensen’s inequality,

Z Z
− |u − uF | ≥ − (u − uF )
F \E F \E
Z
1
= |F \ E| · uF − u
|F \ E| F \E
Z Z
1
= |F | · uF − |E| · uF − u+ u
|F \ E| F E
1
= |F | · uF − |E| · uF − |F | · uF + |E| · uE
|F \ E|
|E|
= |uF − uE |,
|F \ E|
and hence (3.68) follows.
Theorem 3.38. Let p = 1, 2 and u ∈ Lploc (Ω; RN ). Assume there exists a constant Cu > 0
and α > 0 such that Z
|u − uBρ |p dx ≤ Cu ρλ
Bρ
holds for all balls Bρ ⊂⊂ Ω. Then for any subdomain Ω0 ⊂⊂ Ω we have u ∈ Lp,λ (Ω0 ; RN )
and moreover
kukLp,λ (Ω0 ;RN ) ≤ C(Ω0 ) [Cu1/p + kukLp (Ω0 ;RN ) ].
Proof. Let Ω0 ⊂⊂ Ω be given. We will show u ∈ Lp,λ (Ω0 ; RN ). Let d = dist(Ω0 ; ∂Ω). Given
any a ∈ Ω0 and 0 < ρ < diam(Ω0 ). If ρ < dist(a; ∂Ω) we have by the previous lemma,
Z Z
p
|u − uΩ0 (a,ρ) | dx ≤ |u − uBρ (a) |p dx ≤ Cu ρλ .
Ω0 (a,ρ) Bρ (a)
If ρ ≥ dist(a; ∂Ω), then ρ ≥ d > 0 and hence

Z Z 2p kukpLp (Ω0 ;Rn )
p p p
|u − uΩ0 (a,ρ) | dx ≤ 2 |u| dx ≤ ρλ .
Ω0 (a,ρ) Ω0 (a,ρ) dλ
Therefore, for all a ∈ Ω0 , 0 < ρ < diam(Ω0 ) it follows that
2p kukpLp (Ω0 ;Rn )
Z " #
p
|u − uΩ0 (a,ρ) | ≤ Cu + λ
ρλ ,
0
Ω (a,ρ) d
and hence by definition u ∈ Lp,λ (Ω0 ; RN ) and moreover

[u]Lp,λ (Ω0 ;RN ) ≤ C(Ω0 ) [Cu1/p + kukLp (Ω0 ;RN ) ].
1,p
Theorem 3.39. (Morrey) Let u ∈ Wloc (Ω; RN ). Suppose for some β > 0 we have
Z
|Du|p dx ≤ C ρn−p+β , ∀ Bρ ⊂⊂ Ω.
Bρ
0, βp
Then for any Ω0 ⊂⊂ Ω of type A, we have u ∈ C (Ω̄0 ; RN ).
Proof. Using the Poincaré type inequality

Z Z
(3.69) |u − uBR | dx ≤ Cn R |Du| dx,
BR BR
we have for all balls Bρ ⊂⊂ Ω,

Z Z
|u − uBρ | dx ≤ Cn ρ |Du| dx
Bρ Bρ
1− p1
≤ Cn ρ kDukLp (Bρ ;MN ×n ) · |Bρ |
n−p+β
n(1− p1 )
≤ Cρ·ρ p ·ρ
n+ βp
= Cρ .
1,n+ βp β
(Ω0 ; RN ) ∼
0,
Therefore, by Theorem 3.38, u ∈ L = C p (Ω̄0 ; RN ).
When β = 0 Morrey’s theorem has to be replaced by the John-Nirenberg estimate; see

G-T, P. 166, Theorem 7.21.
Theorem 3.40. (John-Nirenberg) Let u ∈ W 1,1 (Ω; RN ) where Ω is convex. Suppsose
there exists a constant K such that
Z
(3.70) |Du| dx ≤ K Rn−1 ∀ a ∈ Ω, R < diamΩ.
Ω(a,R)
Then there exist positive constants σ0 and C depending only on n such that
Z σ
(3.71) exp |u − uΩ | dx ≤ C (diamΩ)n ,
Ω K
where σ = σ0 |Ω| (diamΩ)−n .
Remark 3.20. The set of all functions u ∈ W 1,1 (Ω; RN ) satisfying (3.70) is the space
BM O(Ω; RN ) introduced by John and Nirenberg, and for Ω cubes or balls it follows that
BM O(Ω; RN ) ∼
= Lp,n (Ω; RN ), ∀ p ≥ 1.
For the proof of all these results and more on BM O-spaces, we refer to Gilbarg-Trudinger’s
book for a proof based on the Riesz potential, and Giaquinta’s book on the Calderon-
Zygmund cube decomposition.
3.6.4. Estimates for systems with constant coefficients. We consider systems with
constant coefficients. Let A = Aαβ
ij be constants satisfying hypothesis (H2). We first have
some Campanato estimates for homogeneous systems.
1 (Ω; RN ) be a weak solution of
Theorem 3.41. Let u ∈ Hloc
(3.72) Dα (Aαβ j
ij Dβ u ) = 0, i = 1, 2, ..., N.
Then there exists a constant c depending on Aαβ

ij such that for any concentric balls Bρ ⊂⊂
BR ⊂⊂ Ω,
Z Z
2 ρ n
(3.73) |u| dx ≤ c · ( ) |u|2 dx,
Bρ R BR
Z Z
2 ρ n+2
(3.74) |u − uBρ | dx ≤ c · ( ) |u − uBR |2 dx.
Bρ R BR
Proof. We do scaling first. Let BR = BR (a), where a ∈ Ω. Define
v(y) = u(a + Ry),
where y ∈ D ≡ {y ∈ Rn | a + Ry ∈ Ω}, which includes B̄1 (0) in the y-space. Note that
1 (D; RN ) is a weak solution of
v ∈ Hloc
Dyα (Aαβ j
ij Dyβ v ) = 0.
Then the Caccioppoli estimates show that
kvkW k,2 (B1/2 (0);RN ) ≤ C(k) kvkL2 (B1 (0);RN ) , ∀ k = 1, 2, ...
and hence for all 0 < t ≤ 1/2 it follows that

Z
|v|2 dy ≤ c(n) tn sup |v(y)|2
Bt (0) y∈B1/2 (0)
≤ c(n) tn kvk2W k,2 (B1/2 (0);RN )

≤ c(n, k) tn kvk2L2 (B1 (0);RN ) ,
where we have chosen integer k > n/2 and used the Sobolev embedding W k,2 (B1/2 (0); RN ) ,→
C 0,α (B1/2 (0); RN ) for some 0 < α < 1. Now if t ≥ 1/2 we easily have
Z Z
2 n n
|v| dy ≤ 2 t |v|2 dy.
Bt (0) B1 (0)
Therefore we have proved

Z Z
|v|2 dy ≤ C(n) tn |v|2 dy, ∀ 0 < t < 1.
Bt (0) B1 (0)
Rescaling back to u(x) and letting ρ = tR we have

Z Z
ρ
|u|2 dx ≤ C(n) ( )n · |u|2 dx, ∀ ρ < R < dist(a; ∂Ω);
Bρ (a) R BR (a)
this proves (3.73). Note that Du is also a weak solution of (3.72); therefore, by (3.73) it
follows that
Z Z
2 ρ n
|Du| dx ≤ C(n) ( ) · |Du|2 dx, ∀ ρ < R < dist(a; ∂Ω).
Bρ (a) R BR (a)
Suppose 0 < ρ < R/2. Then we use the Poincaré inequality, the previous estimate and the
Caccioppoli inequality to obtain
Z Z
2 2
|u − uBρ | dx ≤ c(n) ρ · |Du|2 dx
Bρ Bρ
Z
ρ
≤ C(n) ρ2 ( )n · |Du|2 dx
R BR/2
Z
ρ
≤ C(n) ( )n+2 · |u − uBR |2 dx.
R BR
Now if ρ ≥ R/2 we easily have

Z Z
|u − uBρ |2 dx = |u − uBR |2 dx − |Bρ | · |uBρ − uBR |2
Bρ Bρ
Z
≤ |u − uBR |2 dx
BR
Z
n+2 ρ n+2
≤ 2 ( ) · |u − uBR |2 dx.
R BR
Therefore, for all 0 < ρ < R < dist(a; ∂Ω),

Z Z
ρ
|u − uBρ |2 dx ≤ C(n) ( )n+2 · |u(x) − uBR |2 dx.
Bρ R BR
The proof of both (3.73) and (3.74) is now complete.
In (3.73) and (3.74), if we let R → dist(a; ∂Ω), we see that both estimates also hold for
all balls Bρ ⊂⊂ BR ⊂ Ω. We state this fact as follows.
Corollary 3.42. Both estimates (3.73) and (3.74) hold for all balls Bρ ⊂⊂ BR ⊂ Ω.
In the following, we consider the nonhomogeneous elliptic systems with constant coef-
ficients:
(3.75) Dα (Aαβ j i
ij Dβ u ) = Dα fα , i = 1, 2, · · · .
Theorem 3.43. Let Aαβ 1,2 N

ij satisfy hypothesis (H2) and u ∈ Wloc (Ω; R ) be a weak solution
of (3.75). Suppose f ∈ L2,λ
loc (Ω; M
N ×n ) and 0 ≤ λ < n + 2. Then Du ∈ L2,λ (Ω; MN ×n ).
loc
0,µ
Corollary 3.44. Under the same assumptions, if f ∈ Cloc (Ω; MN ×n ) and 0 < µ < 1 then
0,µ
Du ∈ Cloc (Ω; MN ×n ).
Proof of Theorem 3.43. Let Ω0 ⊂⊂ Ω00 ⊂⊂ Ω. Let a ∈ Ω0 and BR (a) = BR ⊂ Ω00 . We

write u = v + w = v + (u − v), where v ∈ H 1 (BR ; RN ) is the solution of the Dirichlet
problem (
div(A Dv) = 0 in BR ,
v|∂BR = u.
The existence of solution v follows by the Lax-Milgram theorem. We now by Corollary 3.42
have for all ρ < R
Z Z
2 ρ n+2
(3.76) |Dv − (Dv)Bρ | dx ≤ c · ( ) |Dv − (Dv)BR |2 dx.
Bρ R BR
From this we have

Z
|Du − (Du)Bρ |2 dx
Bρ
Z
= |Dv + Dw − (Dv)Bρ − (Dw)Bρ |2 dx
Bρ
Z Z
ρ
≤ C · ( )n+2 2
|Dv − (Dv)BR | dx + |Dw − (Dw)Bρ |2 dx
R B Bρ
ZR Z
ρ
≤ C1 · ( )n+2 2
|Du − (Du)BR | dx + C2 |Du − Dv|2 dx.
R BR BR
Since u − v ∈ W01,2 (BR ; RN ), we use the Legendre-Hadamard condition to have

Z Z
ν |Du − Dv|2 dx ≤ A D(u − v) · D(u − v) dx
BR BR
Z
= (f − fBR ) · D(u − v) dx
BR
Z Z
ν
≤ |Du − Dv|2 dx + Cν |f − fBR |2 dx
2 BR BR
and hence
Z Z
|Du − Dv|2 dx ≤ Cν |f − fBR |2 dx ≤ Cν [f ]2L2,λ (Ω00 ;MN ×n ) · Rλ .
BR BR
Combining what we proved above, we have

Z Z
2 ρ n+2
|Du − (Du)Bρ | dx ≤ C1 · ( ) |Du − (Du)BR |2 dx
Bρ R BR
+ C3 [f ]2L2,λ (Ω00 ;MN ×n ) · Rλ .
Let Z
Φ(ρ) = |Du − (Du)Bρ |2 dx.
Bρ
Using the Campanato lemma below, it follows that

ρ λ 2 λ
Φ(ρ) ≤ C4 Φ(R) + [f ]L2,λ (Ω00 ;MN ×n ) · ρ .
R
Now we have
Z Z
2
|Du − (Du)Ω0 (a,ρ) | ≤ |Du − (Du)Bρ (a) |2
Ω0 (a,ρ) Ω0 (a,ρ)
Z
≤ |Du − (Du)Bρ |2 = Φ(ρ)
Bρ
≤ C5 ρλ kDuk2L2 (Ω00 ;MN ×n ) + [f ]2L2,λ (Ω00 ;MN ×n ) .

Therefore

[Du]L2,λ (Ω0 ;MN ×n ) ≤ C kDukL2 (Ω00 ;MN ×n ) + [f ]L2,λ (Ω00 ;MN ×n ) .
Theorem 3.45. (Campanato Lemma) Let Φ(t) be a nonnegative nondecreasing function.

Consider the inequality
h ρ α i
(3.77) Φ(ρ) ≤ A + Φ(R) + B Rβ ∀ ρ ≤ R ≤ R0 ,
R
where A, B, α, β, are positive constants with α > β. Then there exists 0 = 0 (A, α, β)
such that if (3.77) holds for some 0 ≤ ≤ 0 then

ρ β β
Φ(ρ) ≤ C Φ(R) + B ρ ∀ ρ ≤ R ≤ R0 ,
R
where C is a constant depending only on α, β, A.
Proof. For 0 < τ < 1 and R ≤ R0 , (3.77) is equivalent to

Φ(τ R) ≤ A τ α (1 + τ −α ) Φ(R) + B Rβ .
Let γ ∈ (β, α) be fixed and choose τ ∈ (0, 1) so that 2Aτ α ≤ τ γ . Let 0 = τ α . Then, if
(3.77) holds for some 0 ≤ ≤ 0 , we have for every R ≤ R0
Φ(τ R) ≤ τ γ Φ(R) + B Rβ
and therefore for all k = 1, 2, · · ·
Φ(τ k+1 R) ≤ τ γ Φ(τ k R) + B τ kβ Rβ
k
X
(k+1)γ kβ β
≤ τ Φ(R) + B τ R τ j(γ−β)
j=0
(k+1)β β
≤ Cτ (Φ(R) + B R ).
Since Φ(t) is nondecreasing and τ k+2 R < ρ ≤ τ k+1 R for some k, we have
ρ β
β ρ β β
Φ(ρ) ≤ C (Φ(R) + B R ) = C Φ(R) + B ρ ,
R R
as desired. The proof is complete.
3.6.5. Schauder estimates for systems with variable coefficients. We now study
the local regularity of weak solutions of systems with variable coefficients. We first prove
the regularity in the Morrey space L2,λ
loc (Ω) for the gradient of the weak solutions.
Theorem 3.46. Let Aαβ 1,2 N

ij (x) satisfy the hypothesis (H3) and u ∈ Wloc (Ω; R ) be a weak
solution of system
(3.78) Dα (Aαβ j i
ij (x) Dβ u ) = Dα fα .
Suppose f ∈ L2,λ
loc (Ω; M
N ×n ) and 0 ≤ λ < n. Then Du ∈ L2,λ (Ω; MN ×n ).
loc
Proof. Let Ω0 ⊂⊂ Ω00 ⊂⊂ Ω. Let a ∈ Ω0 and BR (a) = BR ⊂ Ω00 . Using the standard Korn’s
freezing coefficients device, u is a weak solution of system with constant coefficients
div(A(a) Du) = div F, F = f + (A(a) − A(x)) Du.
Let v ∈ H 1 (BR ; RN ) be the solution of the Dirichlet problem
(
div(A(a) Dv) = 0 in BR ,
v|∂BR = u.
Then, as before, using (3.73) instead of (3.74) we have
Z Z Z
2 ρ n 2
|Du| ≤ c · ( ) |Du| + C |D(u − v)|2
Bρ R BR BR
Z Z
ρ n 2
≤ c·( ) |Du| + C |F |2
R BR BR
Z Z Z
ρ
≤ c · ( )n |Du|2 + C |f |2 + C ω(R) |Du|2
R BR BR BR
h ρ iZ
≤ c ( )n + ω(R) |Du|2 + C kf k2L2,λ (Ω00 ;MN ×n ) Rλ ,
R BR
where ω(R) is the uniform modulus of continuity of A(x) :

ω(R) = sup |A(x) − A(y)|.
|x−y|≤R
We choose R0 > 0 sufficiently small so that ω(R) < 0 for all R < R0 , where 0 is the
constant appearing in the Campanato lemma above. Therefore,
Z
|Du|2 dx ≤ C(Ω0 , Ω00 ) kDuk2L2 (Ω00 ;MN ×n ) + kf k2L2,λ (Ω00 ;MN ×n ) ρλ .
Bρ
This, by a local version similar to the Campanato space, we have

kDukL2,λ (Ω0 ;MN ×n ) ≤ C(Ω0 , Ω00 ) kDukL2 (Ω00 ;MN ×n ) + kf kL2,λ (Ω00 ;MN ×n ) ,
which proves the theorem.
We now study the regularity of the gradient of weak solutions in the Hölder spaces.
2,n+2µ
This is done by proving the regularity of gradient in the Campanato space Lloc (Ω) for
some µ ∈ (0, 1).
Theorem 3.47. Let Aαβ ij ∈ C

0,µ (Ω) with some 0 < µ < 1 satisfy the hypothesis (H3) and
1,2
u ∈ Wloc (Ω; RN ) be a weak solution of system
(3.79) Dα (Aαβ j i
0,µ 0,µ
Suppose f ∈ Cloc (Ω; MN ×n ). Then Du ∈ Cloc (Ω; MN ×n ).
Proof. Similarly as above, we have

Z Z Z
2 ρ n+2 2
|Du − (Du)Bρ | ≤ c · ( ) |Du − (Du)BR | + C |F − fBR |2
Bρ R BR BR
Z Z
ρ n+2 2
≤ c·( ) |Du − (Du)BR | + C |f − fBR |2
R BR BR
Z
+ C [A]2C 0,µ R2µ |Du|2
BR
Z
ρ n+2
≤ c·( ) |Du − (Du)BR |2 + C [f ]2C 0,µ (Ω00 ) Rn+2µ
R BR
Z
2 2µ
+ C [A]C 0,µ R |Du|2 .
BR
Since by the previous theorem Du ∈ L2,n−

loc (Ω; M
N ×n ) for all > 0 we obtain
Z Z
2 ρ n+2
|Du − (Du)Bρ | ≤ A ( ) |Du − (Du)BR |2 + B Rn+2µ− .
Bρ R BR
Using Campanato’s lemma, we have

Z
|Du − (Du)Bρ |2 ≤ C ρn+2µ−
Bρ
and hence Du ∈ L2,n+2µ−

loc
0,β
(Ω; MN ×n ) for all > 0. This implies Du ∈ Cloc (Ω; MN ×n )
for β = µ − 2 . In particular, Du is locally bounded. Therefore, again, using the above
estimates, it follows that

Z Z
ρ
|Du − (Du)Bρ |2 ≤ c · ( )n+2 |Du − (Du)BR |2
Bρ R BR
Z
+ C [f ]2C 0,µ (Ω00 ) Rn+2µ + C [A]2C 0,µ R2µ |Du|2
BR
Z
ρ n+2
≤ c·( ) |Du − (Du)BR |2
R BR
+ C [f ]2C 0,µ (Ω00 ) Rn+2µ + C [A]2C 0,µ Rn+2µ
and using Campanato’s lemma again we have Du ∈ L2,n+2µ

loc (Ω; MN ×n ) and hence Du ∈
0,µ
Cloc (Ω; MN ×n ).
Finally we remark that the following higher order regularity result can be easily deduced.
Theorem 3.48. Let k ≥ 0, 0 < µ < 1 and Aαβ ij ∈ C
k,µ (Ω) satisfy the hypothesis (H3) and
1,2
u ∈ Wloc (Ω; RN ) be a weak solution of system
(3.80) Dα (Aαβ j i
k,µ k+1,µ
Suppose f ∈ Cloc (Ω; MN ×n ). Then u ∈ Cloc (Ω; RN ).
3.6.6. Systems in non-divergence form and boundary estimates. In this section,

we show that the Campanato estimates can also be proved for systems that are not in
divergence form and also the global estimates are valid if the boundary ∂Ω of the domain
Ω ⊂ Rn satisfies certain smoothness condition.
We first prove the interior estimates for systems in the following form:
Aαβ j i
ij (x) Dαβ u = f ; i = 1, 2, · · · , N.
2,2
By a weak solution u to this system we mean a function u ∈ Wloc (Ω; RN ) such that the
system is satisfied almost everywhere in Ω.
Theorem 3.49. Let Aαβ i 0,µ 2,2 N
ij , f ∈ Cloc (Ω) and 0 < µ < 1. If u ∈ Wloc (Ω; R ) is a weak
2,µ
solution to the system above, then u ∈ Cloc (Ω; RN ).
We now consider the regularity up to the boundary. In what follows, we assume the
boundary ∂Ω of the domain Ω is of C 1,µ ; that is, for any x0 ∈ ∂Ω, there exist an open set
U ⊂ Rn containing x0 and a C 1,µ -diffeomorphism y = G : U → Rn such that
G(x0 ) = 0, G(U ∩ Ω) = B1+ = {y ∈ Rn | |y| < 1, yn > 0};
G(U ∩ ∂Ω) = Γ1 = {y ∈ Rn | |y| < 1, yn = 0}.

This G is called (locally) flattening the boundary. As in the scalar case studied above, we
have the following global regularity theorems for linear systems.
Theorem 3.50. Let ∂Ω be of C 1,µ with 0 < µ < 1 and Aαβ i
ij , fα ∈ C
0,µ (Ω̄) and g j ∈ C 1,µ (Ω̄).
Let A(x) satisfy the condition (H3). If u ∈ H 1 (Ω; RN ) is a weak solution to the problem
div(A(x) Du) = div f, u|∂Ω = g,
then u ∈ C 1,µ (Ω̄; RN ).
Theorem 3.51. Let ∂Ω be of C 1,µ with 0 < µ < 1 and Aαβ

ij ∈ C
0,µ (Ω̄) satisfy (H3).
i 0,µ j 2,µ 2,2 N
Assume f ∈ C (Ω̄), g ∈ C (Ω̄). If u ∈ W (Ω; R ) is a weak solution to the problem
Aαβ j i
ij (x) Dαβ u = f ; uj |∂Ω = g j ,
then u ∈ C 2,µ (Ω̄; RN ).
Chapter 4
Linear Evolution
Equations
This chapter studies various linear partial differential equations that involve time. We call
these equations linear evolution equations. We will study two major types of evolution
equations of second-order: parabolic and hyperbolic equations. Two methods will be used:
Galerkin method and Semigroup method.
4.1. Second-order Parabolic Equations

For this chapter we assume Ω to be an open bounded subset in Rn and set ΩT = Ω×(0, T ] for
some fixed time T > 0; the parabolic boundary ∂ 0 ΩT of ΩT is defined by ∂ 0 ΩT = ΩT \ΩT .
We will study the initial-boundary value problem

ut + Lu = f in ΩT ,

(4.1) u=0 on ∂Ω × [0, T ],

u=g on Ω × {t = 0},

where f : ΩT → R and g : Ω → R are given and u : ΩT → R is the unknown function,

u = u(x, t).
The operator Lu denotes for each time t a second-order partial differential operator,
having either the divergence form
n
X n
X
(4.2) Lu = − Dj (aij (x, t)Di u) + bi (x, t)Di u + c(x, t)u
i,j=1 i=1
or else the nondivergence form

n
X n
X
(4.3) Lu = − aij (x, t)Dij u + bi (x, t)Di u + c(x, t)u,
i,j=1 i=1
for given coefficients aij , bi , c (i, j = 1, 2, · · · , n).
97
98 4. Linear Evolution Equations
∂
Definition 4.1. We say the operator ∂t + L is (uniformly) parabolic on ΩT if there
exists a constant θ > 0 such that
n
X
(4.4) aij (x, t)ξi ξj ≥ θ|ξ|2 for all (x, t) ∈ ΩT and ξ ∈ Rn .
i,j=1
Note that for each fixed time t ∈ [0, T ] the operator Lu is uniformly elliptic in x ∈ Ω.
4.1.1. Weak Solutions. We consider the case that Lu has the divergence form (4.2). We
assume
aij , bi , c ∈ L∞ (ΩT ) (i, j = 1, 2, · · · , n), f ∈ L2 (ΩT ), g ∈ L2 (Ω).
We will also assume aij = aji for i, j = 1, 2, · · · , n.
Introduce the time-dependent bilinear form
n
Z X n
X
(4.5) B[u, v; t] = aij (x, t)Di uDj v + bi (x, t)Di uv + c(x, t)uv dx
Ω i,j=1 i=1
for u, v ∈ H01 (Ω) and almost every t ∈ [0, T ].
Definition 4.2. A weak solution to Problem (4.1) is a function u ∈ L2 (0, T ; H01 (Ω)) with
weak time-derivative u0 ∈ L2 (0, T ; H −1 (Ω)) such that
(i) hu0 (t), vi + B[u(t), v; t] = (f (t), v) for each v ∈ H01 (Ω) and a.e. time t ∈ [0, T ], and
(ii) u(0) = g. (Note that u ∈ C([0, T ]; L2 (Ω)) and thus u(0) is well-defined in L2 (Ω).)
Remark 4.3. (Motivation for definition of weak solutions.)

Suppose u = u(x, t) is a smooth solution of (4.1). Then u defines a function, still
denoted by u: [0, T ] → H01 (Ω) by u(t)(x) = u(x, t). In other words, we consider u not as
a function of (x, t) but rather as a function of t into the space H01 (Ω). We also consider
f : [0, T ] → L2 (Ω) in terms of f (t)(x) = f (x, t).
Fix v ∈ H01 (Ω) and multiply the PDE by v and integrate by parts, and we find
(4.6) (ut (t), v) + B[u(t), v; t] = (f (t), v) ∀ t ∈ [0, T ],
pairing ( , ) is the inner product in L2 (Ω). Note that the PDE can be written as
where the P
ut = g + ni=1 gxi i , with
0
n
X n
X
g0 = f − bi Di u − cu, gi = aij Di u (i = 1, 2, · · · , n).
i=1 i=1
Hence, with G = (g 0 , g 1 , · · · , g n ) ∈ L2 (Ω; Rn+1 ),
(4.7) kut kH −1 (Ω) ≤ kGkL2 (Ω) ≤ C(kukH 1 (Ω) + kf kL2 (Ω) ) a.e. t ∈ [0, T ].
This estimate suggests it is reasonable to look for weak solutions u ∈ L2 (0, T ; H01 (Ω))
with weak time-derivative u0 (t) ∈ H −1 (Ω) for a.e. t ∈ [0, T ], which, by (4.7), also satisfies
u0 ∈ L2 (0, T ; H −1 (Ω)). In this case the first term in (4.6) should be reexpressed as hu0 (t), vi,
as the pairing of H −1 (Ω) and H01 (Ω).
4.1. Second-order Parabolic Equations 99
4.1.2. Galerkin Method – Existence of weak solutions. We intend to build a weak

solution to (4.1) by constructing approximate solutions uk that, for each t ∈ [0, T ], lie in a
finite-dimensional space Vk of H01 (Ω) such that condition (i) in the definition above holds for
all v ∈ Vk . Then we pass to the limit as k → ∞. This is the so-called Galerkin’s method.
Assume the functions wi = wi (x) are smooth and
(4.8) {wi }∞ 1 2
i=1 forms an orthogonal basis of H0 (Ω) and an orthonormal basis of L (Ω).
(For instance, we could take {wi } to be the complete set of appropriately normalized eigen-
functions for −∆ in H01 (Ω).)
Fix now a positive integer k. Let Vk be the linear span of {w1 , · · · , wk } and we look for
a function uk : [0, T ] → Vk of the form
k
X
(4.9) uk (t) = di (t)wi ,
i=1
where the coefficient functions di (t) is selected so that
(4.10) (u0k (t), wi ) + B[uk (t), wi ; t] = (f (t), wi ), di (0) = (g, wi ),
for almost every t ∈ [0, T ] and i = 1, 2, · · · , k.
Here and what follows, ( , ) denotes the inner product in L2 (Ω), and 0 denotes the
time-derivative of a function (whenever it is well-defined in classical or weak sense).
Theorem 4.1. (Construction of approximate solutions) For each k = 1, 2, · · · there
exists a unique function uk of the form (4.9) satisfying (4.10).
Proof. Note that

k
αij (t)dj (t),
X
(u0k (t), wi ) = d0i (t), B[uk (t), wi ; t] =
j=1
where αij (t) = B[wj , wi ; t] (i, j = 1, 2, · · · , k). Hence condition (4.10) becomes the initial
value problem for the ODE system on d(t) = (d1 (t), · · · , dk (t)):
k
αij (t)dj (t) = fi (t) ≡ (f (t), wi ),
X
d0i (t) + di (0) = (g, wi ) (i = 1, 2, · · · , k).
j=1
Note that the coefficients αij belong to L∞ (0, T ) and fi ∈ L2 (0, T ). The existence of a
unique solution d ∈ H 1 (0, T ) ⊂ C([0, T ]) is guaranteed by the (not so) standard existence
theory for ODE (think of approximating αij and fi by smooth functions first and then pass
to limits).
4.1.3. Energy Estimates.

Theorem 4.2. Assume the uniform parabolicity condition. There exists a constant C,
depending only on Ω, T, and the coefficients of L, such that, for all k = 1, 2, · · · ,
(4.11)
max kuk (t)kL2 (Ω) + kuk kL2 (0,T ;H01 (Ω)) + ku0k kL2 (0,T ;H −1 (Ω)) ≤ C(kf kL2 (ΩT ) + kgkL2 (Ω) ).
t∈[0,T ]
Proof. 1. From the uniform parabolicity, as in the elliptic case, there exist constants
β > 0, γ ≥ 0 such that
(4.12) βkvk2H 1 (Ω) ≤ B[v, v; t] + γkvk2L2 (Ω) ∀ v ∈ H01 (Ω), a.e. t ∈ [0, T ].
0
Multiply (4.10) by di (t) and sum for i = 1, 2, · · · , k to find

(4.13) (u0k (t), uk (t)) + B[uk (t), uk (t); t] = (f (t), uk (t)) ∀ a.e. t ∈ [0, T ].
Hence
d
(4.14) kuk (t)k2L2 (Ω) + 2βkuk (t)k2H 1 (Ω) ≤ C1 kuk (t)k2L2 (Ω) + C2 kf (t)k2L2 (Ω)
dt 0
for a.e. t ∈ [0, T ], and appropriate constants C1 , C2 .

2. Now write
η(t) = kuk (t)k2L2 (Ω) , ξ(t) = kf (t)k2L2 (Ω) .
Then
η 0 (t) ≤ C1 η(t) + C2 ξ(t) ∀ a.e. t ∈ [0, T ].
Thus, by Gronwall’s inequality,
Z t
C1 t
η(t) ≤ e η(0) + C2 ξ(s) ds (0 ≤ t ≤ T ).
0
Since η(0) = kuk (0)k2L2 (Ω) ≤ kgk2L2 (Ω) , we obtain the estimate
max kuk (t)k2L2 (Ω) ≤ C(kgk2L2 (Ω) + kf k2L2 (ΩT ) ).

t∈[0,T ]
3. Integrating (4.14) over t ∈ [0, T ], we have

Z T
2
kuk kL2 (0,T ;H 1 (Ω)) = kuk (t)k2H 1 (Ω) dt ≤ C(kgk2L2 (Ω) + kf k2L2 (ΩT ) ).
0 0
0
4. Finally we need to estimate ku0k kL2 (0,T ;H −1 (Ω)) . So, fix any v ∈ H01 (Ω), with kvkH01 (Ω) ≤
1. We write v = v 1 + v 2 , where v 1 ∈ Vk , and (v 2 , wi ) = 0 for all i = 1, 2, · · · , k. (That is, v 2
is in the L2 orthogonal complement of Vk .) Since {wi } are orthogonal in H01 (Ω), we have
kv 1 kH01 (Ω) ≤ kvkH01 (Ω) ≤ 1.
Using (4.10), we have

(u0k (t), v 1 ) + B[uk (t), v 1 ; t] = (f (t), v 1 ).
Then
hu0k (t), vi = (u0k (t), v) = (u0k (t), v 1 ) = (f (t), v 1 ) − B[uk (t), v 1 ; t]
and consequently
|hu0k (t), vi| ≤ C(kf (t)kL2 (Ω) + kuk kH01 (Ω) ).
This implies
ku0k (t)k2H −1 (Ω) ≤ C(kf (t)k2L2 (Ω) + kuk (t)k2H 1 (Ω) ) ∀ t ∈ [0, T ].
0
Integrate over t ∈ [0, T ] to finally obtain

Z T
0 2
kuk kL2 (0,T ;H −1 (Ω)) = ku0k (t)k2H −1 (Ω) dt ≤ C(kf k2L2 (ΩT ) + kuk k2L2 (0,T ;H 1 (Ω)) ),
0
0
which, combining with the estimate in Step 3, derives the desired estimate.
4.1.4. Existence and Uniqueness.
Theorem 4.3. There exists a unique weak solution to (4.1).
Proof. (Existence.) 1. According to the energy estimate (4.11), we see that {uk } is
bounded in L2 (0, T ; H01 (Ω)) and {u0k } is bounded in L2 (0, T ; H −1 (Ω)). Consequently there
exists a subsequence {ukm } of {uk } with km → ∞ and functions u ∈ L2 (0, T ; H01 (Ω)),
w ∈ L2 (0, T ; H −1 (Ω)) such that
(
ukm * u in L2 (0, T ; H01 (Ω)),
(4.15)
u0km * w in L2 (0, T ; H −1 (Ω)).
Moreover, u0 exists in L2 (0, T ; H −1 (Ω)) and u0 = w. From the regularity result, u ∈

C([0, T ]; L2 (Ω)).
2. Fix any integer N and let ψ ∈ C 1 ([0, T ]; H01 (Ω)) have the form
N
X
ψ(t) = ζi (t)wi ,
i=1
where ζi ∈ C 1 ([0, T ]; R). Let k ≥ N , multiply (4.10) by ζi , sum i = 1, 2, · · · , N and then

integrate over t ∈ [0, T ] to find
Z T Z T
0

(4.16) huk (t), ψ(t)i + B[uk (t), ψ(t); t] dt = (f (t), ψ(t))dt.
0 0
Now let k = km → ∞ and we have

Z T Z T
0

(4.17) hu (t), ψ(t)i + B[u(t), ψ(t); t] dt = (f (t), ψ(t))dt.
0 0
This equality then holds for all functions ψ ∈ L2 (0, T ; H01 (Ω)), as functions ζ of the given
form are dense in this space. We then take ψ(t) = ζ(t)v with ζ ∈ L2 (0, T ) and v ∈ H01 (Ω)
in (4.17) to obtain
Z T Z T
0

ζ(t) hu (t), vi + B[u(t), v; t] dt = ζ(t)(f (t), v)dt.
0 0
This holding for all ζ ∈ L2 (0, T ) yields that

(4.18) hu0 (t), vi + B[u(t), v; t] = (f (t), v) ∀ v ∈ H01 (Ω), a.e. t ∈ [0, T ].
3. We need to show the initial data u(0) = g. In (4.16), (4.17), take ψ(t) = ζ(t)v with
ζ ∈ C 1 [0, T ], ζ(T ) = 0, ζ(0) = −1 and v ∈ H01 (Ω). Note that, since ψ 0 (t) = ζ 0 (t)v, we have
Z T Z T
0
huk (t), ψ(t)idt = (uk (0), v) − ζ 0 (t)(uk (t), v) dt,
0 0
(uk (0), v) → (g, v) as k → ∞, and

Z T Z T
0
hu (t), ψ(t)idt = (u(0), v) − ζ 0 (t)(u(t), v) dt.
0 0
Hence, in (4.16), let k = km → ∞, we eventually obtain (u(0), v) = (g, v), for all v ∈ H01 (Ω);
hence u(0) = g.
(Uniqueness.) It suffices to prove that a weak solution u with f = g = 0 must be

zero. To show this, set v = u(t) ∈ H01 (Ω) in (4.18) with f = 0 to have

d 1 2
ku(t)kL2 (Ω) + B[u(t), u(t); t] = 0 (a.e. t ∈ [0, T ]).
dt 2
Since the Gårding’s inequality above implies −B[u(t), u(t); t] ≤ γku(t)k2L2 (Ω) , it follows that

d 1
ku(t)kL2 (Ω) ≤ γku(t)k2L2 (Ω) ∀ a.e. t ∈ [0, T ].
2
dt 2
So Gronwall’s inequality implies ku(t)kL2 (Ω) = 0 as u(0) = 0. Hence u ≡ 0.
4.1.5. Regularity. We now discuss the regularity of weak solutions when the initial data
and coefficients are more regular. Our eventual goal is to prove that the weak solution is
smooth, as long as the coefficients and initial data and the domain are all smooth. This
mirrors the regularity of elliptic equations.
Before we proceed, we prove the following useful result which is Problem 9 of Chapter
7 in Evans’s book; the proof follows a paper by Brezis and Evans (Arch. Rational Mech.
Analysis 71 (1979), 1–13).
Lemma 4.4. If Lu is uniformly elliptic with smooth coefficients, then there exist constants
β > 0, γ ≥ 0 such that
(4.19) βkuk2H 2 (Ω) ≤ (Lu, −∆u) + γkuk2L2 (Ω) ∀ u ∈ H01 (Ω) ∩ H 2 (Ω).
Proof. 1. Given a function u ∈ H 2 (Ω) ∩ H01 (Ω), we claim that there exists a sequence of
functions um in C 3 (Ω̄) vanishing on ∂Ω such that kum − ukH 2 (Ω) → 0; therefore, we only
need to prove (4.19) for functions u ∈ C 3 (Ω̄) vanishing on ∂Ω. To prove this claim, let
f = ∆u ∈ L2 (Ω) and choose fm ∈ C ∞ (Ω̄) so that kfm − f kL2 (Ω) → 0. Let um ∈ H01 (Ω)
be the weak solution of ∆um = fm in Ω. Then the global regularity theorem shows that
um ∈ H k (Ω) if ∂Ω is of C k and k ≥ 2. By the general Sobolev inequalities we know that if
n
k > n2 then um ∈ C k−[ 2 ]−1,γ (Ω̄) for some 0 < γ < 1. Hence if ∂Ω is of C k with k = 4 + [ n2 ],
then um ∈ C 3 (Ω̄). Clearly um = 0 on ∂Ω. Finally, since ∆(um − u) = fm − f , by the global
H 2 -estimate, we have
kum − ukH 2 (Ω) ≤ Ckfm − f kL2 (Ω) → 0.
2. In the following, assume u ∈ C 3 (Ω̄) and u = 0 on ∂Ω. Then we can write Lu in

non-divergence form (using the new smooth coefficients) as
Lu = −aij (x)uxi xj + bi (x)uxi + c(x)u(x).
Let x0 ∈ ∂Ω and assume a smooth homeomorphism y = Φ(x) from small ball B(x0 , r) to
a domain D in y-space satisfies that Φ(x0 ) = 0 and Φ(B(x0 , r) ∩ Ω) = D ∩ {yn > 0} and
Φ(B(x0 , r) ∩ ∂Ω) = D ∩ {yn = 0}. Let x = Ψ(y) be the inverse map of y = Φ(x). (This
is the standard technique of locally flattening the boundary.) Let ζ ∈ Cc∞ (B(x0 , r) and
0 ≤ ζ ≤ 1. Then
Z Z
ζ(x)(Lu, −∆u)dx = ψ(y)(L1 v, L2 v)dy,
Ω D∩{yn >0}
where v(y) = u(Ψ(y)), ψ(y) = ζ(Ψ(y))|I(y)| with I(y) the Jacobian determinant of x =
Ψ(y), and L1 , L2 are the new differential operators of v(y) from Lu and ∆u, both of the
form
Lk v = −akij (y)vyi yj + bki (y)vyi + ck (y)v(y) (k = 1, 2).
The coefficients ck (y) are bounded (in fact, c2 (y) ≡ 0) and the coefficients akij (y), bki (y) are
smooth functions on D ∩ {yn ≥ 0} and satisfy for a constant θ > 0, with k = 1, 2,
akij (y) = akji (y),
akij (y)ξi ξj ≥ θ|ξ|2 (y ∈ D ∩ {yn ≥ 0}, ξ ∈ Rn ).
From this, an easy algebra proof shows that for all symmetric matrices C ∈ Mn×n
n
X n
X
(4.20) a1ij (y)cik a2kl (y)cjl ≥ θ2 c2ij (y ∈ D ∩ {yn ≥ 0}).
i,j,k,l=1 i,j=1
ψ(y)(L1 v, L2 v)dy.
R
We need to estimate D∩{yn >0}
3. We write the integrand as
ψ(y)(L1 v, L2 v) = ψa1ij vyi yj a2kl vyk yl + ψR,
where R is the term of the form
X X X
R= Aijk vyi yj vyk + Bij vyi vyj + Cij vyi yj v + c1 c2 v 2 ,
with bounded coefficients A, B, C. The leading term can be written as
ψa1ij vyi yj a2kl vyk yl = ψa1ij vyi yk a2kl vyj yl
+ (ψa1ij a2kl )yj vyi yk vyl − (ψa1ij a2kl )yk vyi yj vyl
+ (ψa1ij a2kl vyi yj vyl )yk − (ψa1ij a2kl vyi yk vyl )yj .
Therefore, by (4.20), we have
Z n Z
X
1 2 2
ψ(y)(L v, L v)dy ≥ θ ψ(y)vy2i yj dy
D∩{yn >0} i,j=1 D∩{yn >0}
n
X Z
(4.21) + ψa1ij a2kl (vyi yj vyl νk − vyi yk vyl νj ) dS
i,j,k,l=1 ∂(D∩{y n >0})
X n Z Z
− ε vy2i yj dy − Cε (|∇v|2 + v 2 ) dy,
i,j=1 D∩{yn >0} D∩{yn >0}
where ν = (ν1 , · · · , νn ) is the outer unit normal on the boundary.

4. Since ψ = 0 on ∂(D ∩ {yn > 0}) \ (D ∩ {yn = 0}), we have the boundary integral
Xn Z
ψa1ij a2kl (vyi yj vyl νk − vyi yk vyl νj ) dS
i,j,k,l=1 ∂(D∩{yn >0})
n Z
X n
X Z
0
=− ψa1ij a2nl vyi yj vyl dy + ψa1in a2kl vyi yk vyl dy 0 .
i,j,l=1 D∩{yn =0} i,k,l=1 D∩{yn =0}
Now since v(y 0 , 0) = 0 we have vyi = 0, vyi yj = 0 for 1 ≤ i, j ≤ n − 1. So
n Z
X n Z
X
ψa1ij a2nl vyi yj vyl dy 0 = ψa1ij a2nn vyi yj vyn dy 0
i,j,l=1 D∩{yn =0} i,j=1 D∩{yn =0}
Z n−1
XZ
= ψa1nn a2nn vyn yn vyn dy 0 + 2 ψa1in a2nn vyi yn vyn dy 0 ,
D∩{yn =0} i=1 D∩{yn =0}
and
n
X Z n Z
X
0
ψa1in a2kl vyi yk vyl dy = ψa1in a2kn vyi yk vyn dy 0
i,k,l=1 D∩{yn =0} i,k=1 D∩{yn =0}
Z n−1
XZ
0
= ψa1nn a2nn vyn yn vyn dy + ψa1in a2nn vyi yn vyn dy 0
D∩{yn =0} i=1 D∩{yn =0}
n−1
XZ
+ ψa1nn a2kn vyk yn vyn dy 0 .
k=1 D∩{yn =0}
Therefore
n
X Z
i,j,k,l=1 ∂(D∩{yn >0})
n−1
XZ
= ψ(a1nn a2in − ψa1in a2nn )vyi yn vyn dy 0
i=1 D∩{yn =0}
n−1 Z
1X
= ψ(a1nn a2in − ψa1in a2nn )(vy2n )yi dy 0
2 D∩{yn =0}
i=1
n−1 Z
1X
=− [ψ(a1nn a2in − ψa1in a2nn )]yi vy2n dy 0 .
2 D∩{yn =0}
i=1
Hence
n
X Z
i,j,k,l=1 ∂(D∩{yn >0})
n−1 Z Z
1X
= [ψ(a1nn a2in − ψa1in a2nn )]yi vy2n dy 0 ≤ C |∇v|2 dS
2 D∩{yn =0} ∂(D∩{yn >0})
i=1
n Z
X Z
≤ε vy2i yj dy + Cε |∇v|2 dy,
i,j=1 D∩{yn >0} D∩{yn >0}
by the trace inequality: for ε > 0,

k∇vkL2 (∂U ) ≤ εkD2 vkL2 (U ) + Cε k∇vkL2 (U ) ∀ v ∈ H 2 (U ).
(Prove this inequality!)

5. Putting all inequalities above together in (4.21), we eventually obtain
Z Xn Z
1 2 2
ψ(y)(L v, L v)dy ≥ θ ψ(y)vy2i yj dy
D∩{yn >0} i,j=1 D∩{yn >0}
n
X Z Z
−ε vy2i yj dy − Cε (|∇v|2 + v 2 ) dy,
i,j=1 D∩{yn >0} D∩{yn >0}
where ε > 0 is arbitrary and Cε depends on ψ.

6. Switching back to the domain Ω, we have

Z n Z
X
2
ζ(x)(Lu, −∆u) dx ≥ θ ζ(x)u2xi xj dx
Ω i,j=1 Ω
(4.22) n Z Z
X
−ε u2xi xj dx − Cε (|∇u|2 + u2 ) dx,
i,j=1 Ω Ω
where Cε depends on ζ. Clearly this estimate also holds when ζ ∈ Cc∞ (Ω).
7. We now cover ∂Ω by finitely many balls Bk := B(xk , rk ) with xk ∈ ∂Ω, k =
1, 2, · · · , N , that the local flattening of the boundary works. Let BN +1 := Ω\∪N k
k=1 B̄(x , rk /2).
PN +1
We find a partition of unity k=1 ζk = 1 subordinate to {B1 , B2 , · · · , BN +1 } with 0 ≤ ζk ≤
1 and supp ζk ⊂⊂ Bk . Then using (4.22) and a choice of small ε > 0 we deduce that
θ2
Z Z
(4.23) (Lu, −∆u) dx ≥ kD ukL2 (Ω) − C (|∇u|2 + u2 ) dx.
2 2
Ω 2 Ω
From this, the estimate (4.19) follows since kD2 ukL2 (Ω) is an equivalent norm for H 2 (Ω) ∩
H01 (Ω) in H 2 (Ω) and
k∇ukL2 (Ω) ≤ εkD2 ukL2 (Ω) + Cε kukL2 (Ω) ∀ u ∈ H 2 (Ω) ∩ H01 (Ω),
which can be seen from one of the homework problem.
Remark 4.4. Clearly, from the proof, the estimate (4.19) holds if we replace −∆u by
another uniformly elliptic operator M u.
In the following we assume the coefficients aij , bi , c are as smooth as we need on ΩT . As

usual, we always assume the uniform parabolicity.
Theorem 4.5. (Improved regularity) (i) Assume g ∈ H01 (Ω), f ∈ L2 (ΩT ). Suppose u
is the weak solution of (4.1). Then
u ∈ L2 (0, T ; H 2 (Ω)) ∩ L∞ (0, T ; H01 (Ω)), u0 ∈ L2 (0, T ; L2 (Ω)),
with the estimate
kukL∞ (0,T ;H01 (Ω)) + kukL2 (0,T ;H 2 (Ω)) + ku0 kL2 (0,T ;L2 (Ω))
(4.24)
≤ C(kgkH01 (Ω) + kf kL2 (ΩT ) ),
where C depends only on Ω, T and the coefficients of L.
(ii) If, in addition, g ∈ H 2 (Ω), f 0 ∈ L2 (0, T ; L2 (Ω)), then
u ∈ L∞ (0, T ; H 2 (Ω)), u0 ∈ L∞ (0, T ; L2 (Ω)) ∩ L2 (0, T ; H01 (Ω)), u00 ∈ L2 (0, T ; H −1 (Ω)),
with the estimate
kukL∞ (0,T ;H 2 (Ω)) + ku0 kL∞ (0,T ;L2 (Ω)) + ku0 kL2 (0,T ;H01 (Ω))
(4.25)
+ ku00 kL2 (0,T ;H −1 (Ω)) ≤ C(kgkH 2 (Ω) + kf kH 1 (0,T ;L2 (Ω) ),
Proof. Let {uk } be the Galerkin approximations satisfying (4.10) constructed as above
with {wi } being the complete collection of eigenfunctions with eigenvalues {λi } for −∆ on
H01 (Ω). As before, we assume {wi } is orthogonal on H01 (Ω) and orthonormal on L2 (Ω).
By the uniqueness theorem, the weak solution u is obtained as the limit of the Galerkin
approximations {uk }. We prove the theorem by deriving the same estimates for the ap-
proximate solutions uk independent of k.
1. We first claim the following estimate for {uk }: for each t ∈ [0, T ],
(4.26) kuk (t)k2H 2 (Ω) ≤ C(kf (t)k2L2 (Ω) + ku0k (t)k2L2 (Ω) + kuk (t)k2L2 (Ω) ),
where C is a constant independent of k. To prove (4.26), note
(Luk (t), wi ) = B[uk (t), wi ; t] = (f (t) − u0k (t), wi ) (i = 1, 2, · · · , k).
Multiply this equation by λi di (t) and sum over i = 1, 2, · · · , k to deduce
(4.27) (Luk (t), −∆uk (t)) = B[uk (t), −∆uk (t); t] = (f (t) − u0k (t), −∆uk (t)),
since −∆uk (t) ∈ H01 (Ω). Then (4.26) follows from (4.27) and Lemma 4.4.
2. We multiply the equation in (4.10) by d0i (t) and sum i = 1, 2, · · · , k, to discover
(4.28) (u0k (t), u0k (t)) + B[uk (t), u0k (t); t] = (f (t), u0k (t)).
Write B[u, v; t] = A[u, v; t] + C[u, v; t], where
 
Z Xn
A[u, v; t] =  aij (x, t)Di uDj v  dx,
Ω i,j=1
(4.29)
n
Z !
X
C[u, v; t] = bi (x, t)(Di u)v + c(x, t)uv dx.
Ω i=1
Note that A[u, v; t] is a symmetric bilinear form on H01 (Ω) and for any functions u ∈
C 1 ([0, T ]; H01 (Ω)),

0 0 1 d
(4.30) A[u (t), u(t); t] = A[u(t), u (t); t] = A[u(t), u(t); t] − Ã[u(t), u(t); t] ,
2 dt
where  
Z n
X
Ã[u, v; t] =  a0ij (x, t)Di uDj v  dx.
Ω i,j=1
The equation (4.28) above can be written as

1d
ku0k (t)k2L2 (Ω) + (A[uk (t), uk (t); t])
(4.31) 2 dt
1
= (f (t), u0k (t)) − C[uk (t), u0k (t); t] + Ã[uk (t), uk (t); t].
2
Moreover, for all > 0,
(4.32) |C[u, v; t]| ≤ kvk2L2 (Ω) + C kuk2H 1 (Ω) (u, v ∈ H01 (Ω)).
0
Therefore, by (4.31), we have

d
(4.33) ku0k (t)k2L2 (Ω) + (A[uk (t), uk (t); t]) ≤ C(kf (t)k2L2 (Ω) + kuk k2H 1 (Ω) ).
dt 0
Integrate over t ∈ [0, T ] to have

Z T
ku0k (t)k2L2 (Ω) dt + max A[uk (t), uk (t); t]
0 t∈[0,T ]
≤ A[uk (0), uk (0); 0] + C(kuk k2L2 (0,T ;H 1 (Ω)) + kf k2L2 (ΩT ) )

0
≤ C(kgk2H 1 (Ω) + kf k2L2 (ΩT ) ),

0
where we have used (4.11) and the estimate A[uk (0), uk (0); 0] ≤ Ckuk (0)k2H 1 (Ω) ≤ Ckgk2H 1 (Ω) ,
0 0
since
k
X k
X
kuk (0)k2H 1 (Ω) = d2i (0)kwi k2H 1 (Ω) = (g, wi )2 kwi k2H 1 (Ω)
0 0 0
i=1 i=1
X∞
≤ (g, wi )2 kwi k2H 1 (Ω) = kgk2H 1 (Ω) .
0 0
i=1
Therefore, using A[v, v; t] ≥ θkvk2H 1 (Ω) for all v ∈ H01 (Ω),

0
ku0k k2L2 (0,T ;L2 (Ω)) + kuk k2L∞ (0,T ;H 1 (Ω)) ≤ C(kgk2H 1 (Ω) + kf k2L2 (ΩT ) ).
0 0
From this and (4.26),
kuk k2L2 (0,T ;H 2 (Ω)) ≤ C(kgk2H 1 (Ω) + kf k2L2 (ΩT ) ).

0
Note that (4.24) follows from the two estimates above.

3. Assume now the hypotheses of assertion (ii). We differentiate the equation in (4.10)
with respect to t and set ũk := u0k to obtain
(4.34) (ũ0k (t), wi ) + B[ũk (t), wi ; t] = (f 0 (t), wi ) − B̃[uk (t), wi ; t],
where B̃ is the bilinear form defined by

n
Z X n
X
B̃[u, v; t] = a0ij (x, t)Di uDj v + b0i (x, t)(Di u)v + c0 (x, t)uv dx.
Ω i,j=1 i=1
Multiplying (4.34) by d0i (t) and summing over i = 1, 2, · · · , k, we discover
(ũ0k (t), ũk (t)) + B[ũk (t), ũk (t); t] = (f 0 (t), ũk (t)) − B̃[uk (t), ũk (t); t].
So,
1d
(4.35) kũk (t)k2L2 (Ω) + B[ũk (t), ũk (t); t] = (f 0 (t), ũk (t)) − B̃[uk (t), ũk (t); t].
2 dt
Notice that, for u ∈ H01 (Ω) ∩ H 2 (Ω), v ∈ H01 (Ω), integration by parts yields
Z n
X n
X
B̃[u, v; t] = v (−Dj a0ij Di u − a0ij Dij u) + b0i (x, t)Di u 0
+ c (x, t)u dx,
Ω i,j=1 i=1
and, hence, for a.e. t ∈ [0, T ] and all u ∈ H 2 (Ω), v ∈ H01 (Ω),
|B̃[u, v; t]| ≤ C(kvk2H 1 (Ω) + kuk2H 2 (Ω) ).

0
Using this estimate and Gårding’s inequality, we deduce from (4.67)

d
kũk (t)k2L2 (Ω) + β kũk (t)k2H 1 (Ω)
dt
≤ C(kũk (t)k2L2 (Ω) + kuk (t)k2H 2 (Ω) + kf 0 (t)k2L2 (Ω) ).
Hence Gronwall’s inequality implies

Z T
sup kũk (t)k2L2 (Ω) +β kũk (t)k2H 1 (Ω) dt
t∈[0,T ] 0
Z T
2 0
(4.36) ≤ C kũk (0)kL2 (Ω) + (kuk (t)k2H 2 (Ω) + kf (t)k2L2 (Ω) )dt
0
≤ C(kũk (0)k2L2 (Ω) + kf k2H 1 (0,T ;L2 (Ω)) + kgk2H 1 (Ω) )
0
≤ C(kuk (0)k2H 2 (Ω) + kf k2H 1 (0,T ;L2 (Ω)) + kgk2H 1 (Ω) ),

0
where, recall that ũk = u0k and by equation (4.10),

kũk (0)kL2 (Ω) = ku0k (0)kL2 (Ω) ≤ C(kf (0)kL2 (Ω) + kuk (0)kH 2 (Ω) )
≤ C(kf kH 1 (0,T ;L2 (Ω)) + kuk (0)kH 2 (Ω) ).
4. We must estimate kuk (0)kH 2 (Ω) . This is a little tricky. Recall that {wi } is the
complete set of smooth eigenfunctions of −∆ on H01 (Ω). Since both uk and ∆uk are in
H01 (Ω) ∩ H 2 (Ω), we have
kuk (0)k2H 2 (Ω) ≤ Ck∆uk (0)k2L2 (Ω) = C(uk (0), ∆2 uk (0)) = C(g, ∆2 uk (0)),
since (uk (0), wi ) = (g, wi ), and
(g, ∆2 uk (0)) = (∆g, ∆uk (0)) ≤ εkuk (0)k2H 2 (Ω) + Cε kgk2H 2 (Ω) .
From these, with sufficiently small ε > 0, we have
kuk (0)k2H 2 (Ω) ≤ C(g, ∆2 uk (0)) ≤ Ckgk2H 2 (Ω) .
Hence, by (4.36),
ku0k kL∞ (0,T ;L2 (Ω)) + ku0k kL2 (0,T ;H 1 (Ω)) ≤ C(kf k2H 1 (0,T ;L2 (Ω)) + kgk2H 2 (Ω) ).
From this and (4.26), noting maxt∈[0,T ] kf (t)kL2 (Ω) ≤ kf kH 1 (0,T ;L2 (Ω)) , we deduce
kuk k2L∞ (0,T ;H 2 (Ω)) ≤ C(kf k2H 1 (0,T ;L2 (Ω)) + kgk2H 2 (Ω) ).
5. It remains to show u00 ∈ L2 (0, T ; H −1 (Ω))). To do so, take v ∈ H01 (Ω) with kvkH01 (Ω) ≤
1 and set v = v 1 + v 2 , as above with v 1 ∈ Vk and (v 2 , wi ) = 0 for i = 1, 2, · · · , k. Then, for
a.e. t ∈ [0, T ], by (4.34),
hu00k (t), vi = (u00k (t), v) = (u00k (t), v 1 ) = (f 0 (t), v 1 ) − B[u0k (t), v 1 ; t] − B̃[uk (t), v 1 ; t].
Hence, since kv 1 kH01 (Ω) ≤ 1,
|hu00k (t), vi| ≤ C(kf 0 (t)kL2 (Ω) + ku0k (t)kH01 (Ω) + kuk (t)kH01 (Ω) ).
This proves
ku00k (t)kH −1 (Ω) ≤ C(kf 0 (t)kL2 (Ω) + ku0k (t)kH01 (Ω) + kuk (t)kH01 (Ω) ).
So, squaring, integrating over t ∈ [0, T ] and using the estimates obtained above, we have
ku00k k2L2 (0,T ;H −1 (Ω)) ≤ C(kf 0 k2L2 (0,T ;L2 (Ω)) + ku0k k2L2 (0,T ;H 1 (Ω)) + kuk k2L2 (0,T ;H 1 (Ω)) )
0 0
≤ C(kf k2H 1 (0,T ;L2 (Ω)) + kgk2H 2 (Ω) ).

By limit, this proves u00 ∈ L2 (0, T ; H −1 (Ω)) with the desired norm estimate.
We now study the higher regularity. For simplicity, we assume the coefficients of L
are smooth and independent of time t; we also assume the uniform parabolicity and the
smoothness of the domain Ω.
Theorem 4.6. (Higher regularity) Assume m ≥ 0 in an integer and
dk f
g ∈ H 2m+1 (Ω), ∈ L2 (0, T ; H 2m−2k (Ω)) (k = 0, 1, · · · , m).
dtk
Suppose the following m-th-order compatibility conditions hold:
g0 := g ∈ H01 (Ω), g1 := f (0) − Lg0 ∈ H01 (Ω),
(4.37) dm−1 f
· · · , gm := (0) − Lgm−1 ∈ H01 (Ω).
dtm−1
Then the weak solution u to (4.1) satisfies
dk u
∈ L2 (0, T ; H 2m−2k+2 (Ω)) (k = 0, 1, 2, · · · , m + 1),
dtk
with the estimate
m+1 m
!
X dk u X dk f
(4.38) k k kL2 (0,T ;H 2m−2k+2 (Ω)) ≤ C k k kL2 (0,T ;H 2m−2k (Ω)) + kgkH 2m+1 (Ω) .
dt dt
k=0 k=0
Proof. The proof is an induction on m, the case m = 0 being the conclusion (i) of Theorem
4.5 above. Assume now the theorem is valid for some integer m ≥ 0, and suppose then
dk f
g ∈ H 2m+3 (Ω), ∈ L2 (0, T ; H 2m+2−2k (Ω)) (k = 0, 1, · · · , m + 1)
dtk
and the (m + 1)-th-order compatibility conditions hold. Let ũ = u0 . Then the previous
theorem implies that ũ ∈ L2 (0, T ; H01 (Ω)) and ũ0 ∈ L2 (0, T ; H −1 (Ω)). Also ũ is the weak
solution to
˜

ũt + Lũ = f in ΩT ,

ũ = 0 on ∂Ω × [0, T ],

ũ = g̃ on Ω × {t = 0},

where f˜ = f 0 , g̃ = f (0)−Lg = g1 . In particular, f˜ and g̃ satisfy the m-th order compatibility

conditions. Then we use induction on m; details are referred to Evans’s book.
Remark 4.5. The condition on f implies
f (0) ∈ H 2m−1 (Ω), f 0 (0) ∈ H 2m−3 (Ω), · · · , f (m−1) (0) ∈ H 1 (Ω)
and consequently
g ∈ H 2m+1 (Ω), g1 ∈ H 2m−1 (Ω), · · · , gm ∈ H 1 (Ω).
The compatibility conditions are precisely the requirements that, each of these functions
gk is zero on ∂Ω. This is because the homogeneous boundary condition u(x, t) = 0 foron
k
x ∈ ∂Ω (in trace), making ddtku (x, t) = 0 on x ∈ ∂Ω (in trace) for each k = 0, 1, · · · , m.
Theorem 4.7. (Smoothness of weak solution) Assume g ∈ C ∞ (Ω̄), f ∈ C ∞ (ΩT ), and

the m-th-order compatibility conditions hold for all m = 0, 1, · · · . Then problem (4.1) has a
unique solution u ∈ C ∞ (ΩT ).
4.1.6. Maximum Principle. We now study some properties for classical (smooth) so-
lutions of parabolic equations. We include such a study here in order to compare with
the properties for classical solutions of hyperbolic equations we shall study later in Section
4.2.6. Although the methods we use here to treat both parabolic and hyperbolic equations
are very similar, we shall see that their solutions behave quite different.
Let us denote by C 2,1 (ΩT ) functions satisfying ut , uxi xj ∈ C(ΩT ). We shall consider
general inequalities of the form
(4.39) ut + Lu ≤ 0 in ΩT
where we assume L is of nondivergence form:

n
X ∂2u X ∂u
Lu = − aij (x, t) + bi (x, t) + c(x, t)u
∂xi ∂xj ∂xi
i,j=1
and satisfies the ellipticity condition

n
X
aij (x, t)ξi ξj ≥ λ|ξ|2 for (x, t) ∈ ΩT
i,j=1
where λ > 0 is a constant, and the coefficients aij , bi , c are all bounded functions in ΩT .
Theorem 4.8. (Weak Maximum Principle) Suppose u ∈ C 2,1 (ΩT ) ∩ C(ΩT ) satisfies
(4.39) and c ≥ 0. Then
(4.40) max u ≤ max

0
u+ .
ΩT ∂ ΩT
Theorem 4.9. (Parabolic Harnack inequality) Assume u ∈ C 2,1 (ΩT ) solves ut + Lu =

0 in ΩT and u ≥ 0 in ΩT . Suppose Ω0 ⊂⊂ Ω is connected. Then, for each 0 < t1 < t2 ≤ T,
there exists a constant C depending only on V, t1 , t2 , and the coefficients of L, such that
(4.41) sup u(·, t1 ) ≤ C inf u(·, t2 ).

V V
Theorem 4.10. (Strong Maximum Principle) Suppose u ∈ C 2,1 (ΩT ) ∩ C(ΩT ) satisfies
(4.39). Let
M = max u = u(x0 , t0 ).
ΩT
Assume one of the following conditions holds:
(a) c(x, t) ≡ 0; (b) c(x, t) ≥ 0 and M ≥ 0; (c) c(x, t) arbitrary and M = 0.
Then we have the strong maximum principle:

(i) If (x0 , t0 ) ∈ ΩT , then u(x, t) ≡ M for all (x, t) ∈ Ωt0 .
(ii) If x0 ∈ ∂Ω and 0 < t0 < T , but u(x, t) < M for all x ∈ Ω, 0 < t < t0 , then
∂u
(4.42) (x0 , t0 ) > 0
∂ν
provided the exterior normal derivative exists at (x0 , t0 ).
4.2. Second-order Hyperbolic Equations 111
4.2. Second-order Hyperbolic Equations

We will study the initial-boundary value problem

utt + Lu = f in ΩT ,

(4.43) u=0 on ∂Ω × [0, T ],

u = g, ut = h on Ω × {t = 0},

where f : ΩT → R and g, h : Ω → R are given and u : ΩT → R is the unknown function,

u = u(x, t).
The operator Lu denotes for each time t a second-order partial differential operator,
having either the divergence form
n
X n
X
(4.44) Lu = − Dj (aij (x, t)Di u) + bi (x, t)Di u + c(x, t)u
i,j=1 i=1
or else the nondivergence form

n
X n
X
(4.45) Lu = − aij (x, t)Dij u + bi (x, t)Di u + c(x, t)u,
i,j=1 i=1
for given coefficients aij , bi , c (i, j = 1, 2, · · · , n).

∂ 2
Definition 4.6. We say the operator ∂t 2 + L is called (uniformly) hyperbolic on ΩT if
there exists a constant θ > 0 such that
n
X
(4.46) aij (x, t)ξi ξj ≥ θ|ξ|2 for all (x, t) ∈ ΩT and ξ ∈ Rn .
i,j=1
Note that for each fixed time t ∈ [0, T ] the operator Lu is uniformly elliptic on Ω.
4.2.1. Weak Solutions. We consider the case that Lu has the divergence form (4.44).
Let B[u, v; t] be the time-dependent bilinear form defined as above.
Definition 4.7. A weak solution to Problem (4.43) is a function u ∈ L2 (0, T ; H01 (Ω))
having weak time-derivatives u0 ∈ L2 (0, T ; L2 (Ω)) and u00 ∈ L2 (0, T ; H −1 (Ω)) such that
(i) hu00 (t), vi + B[u(t), v; t] = (f (t), v) for each v ∈ H01 (Ω) and a.e. time t ∈ [0, T ], and
(ii) u(0) = g, u0 (0) = h. (Note that u ∈ C([0, T ]; L2 (Ω)) and u0 ∈ C([0, T ]; H −1 (Ω)),
and thus u(0) and u0 (0) are well-defined.)
4.2.2. Galerkin Approximations. We assume

aij , bi , c ∈ C 1 (ΩT ) (i, j = 1, 2, · · · , n),
f ∈ L2 (ΩT ), g ∈ H01 (Ω), h ∈ L2 (Ω).
We will also assume aij = aji for i, j = 1, 2, · · · , n.
Again, assume the functions wi = wi (x) are smooth and
(4.47) {wi }∞ 1 2
i=1 forms an orthogonal basis of H0 (Ω) and an orthonormal basis of L (Ω).
(For instance, we could take {wi } to be the complete set of appropriately normalized eigen-
functions for −∆ in H01 (Ω).)
Fix now a positive integer k. Let Vk be the linear span of {w1 , · · · , wk } and we look for
a function uk : [0, T ] → Vk of the form
k
X
(4.48) uk (t) = di (t)wi ,
i=1
where the coefficient functions di (t) is selected so that

(
(u00k (t), wi ) + B[uk (t), wi ; t] = (f (t), wi ),
(4.49)
di (0) = (g, wi ), d0i (0) = (h, wi ),
for almost every t ∈ [0, T ] and i = 1, 2, · · · , k. Here ( , ) denotes the inner product in L2 (Ω).
Theorem 4.11. (Construction of approximate solutions) For each k = 1, 2, · · · there
exists a unique function uk of the form above satisfying (4.49).
Proof. Note that

k
αij (t)dj (t),
X
(u00k (t), wi ) = d00i (t), B[uk (t), wi ; t] =
j=1
where αij (t)

= B[wj , wi ; t] (i, j = 1, 2, · · · , k). Hence condition (4.10) becomes the initial
value problem for the ODE system on d(t) = (d1 (t), · · · , dk (t)):
(
d00i (t) + kj=1 αij (t)dj (t) = fi (t) ≡ (f (t), wi ),
P
di (0) = (g, wi ), d0i (0) = (h, wi ) (i = 1, 2, · · · , k).
Note that the coefficients αij belong to L∞ (0, T ) and fi ∈ L2 (0, T ). The existence of a
unique solution d ∈ H 2 (0, T ) ⊂ C 1 ([0, T ]) is guaranteed by the (not so) standard existence
theory for ODE (think of approximating αij and fi by smooth functions first and then pass
to limits).
4.2.3. Energy Estimates.

Theorem 4.12. Assume the uniform hyperbolicity condition. There exists a constant C,
depending only on Ω, T, and the coefficients of L, such that, for all k = 1, 2, · · · ,
max (kuk (t)kH01 (Ω) + ku0k (t)kL2 (Ω) ) + ku00k kL2 (0,T ;H −1 (Ω))
t∈[0,T ]
(4.50)
≤ C(kf kL2 (ΩT ) + kgkH01 (Ω) + khkL2 (Ω) ).
Proof. 1. Multiply (4.49) by d0i (t) and sum for i = 1, 2, · · · , k to find

(4.51) (u00k (t), u0k (t)) + B[uk (t), u0k (t); t] = (f (t), u0k (t)) ∀ a.e. t ∈ [0, T ].
Observe (u00k , u0k ) = d 1 0 2
dt ( 2 kuk kL2 (Ω) ); furthermore, as above,
B[uk (t), u0k (t); t] = A[uk (t), u0k (t); t] + C[uk (t), u0k (t); t] := B1 + B2 ,
where,
Z X n
d 1 1
B1 = A[uk (t), uk (t); t] − a0ij Di uk Dj uk dx
dt 2 2 Ω
i,j=1

d 1
≥ A[uk (t), uk (t); t] − Ckuk (t)k2H 1 (Ω) .
dt 2 0
We also note that

|B2 | ≤ C(kuk (t)k2H 1 (Ω) + ku0k (t)k2L2 (Ω) ).
0
Hence, in view of (4.51), we discover

(4.52)
d 0
kuk (t)k2L2 (Ω) + A[uk (t), uk (t); t] ≤ C(ku0k (t)k2L2 (Ω) + kuk (t)k2H 1 (Ω) + kf (t)k2L2 (Ω) )
dt 0
≤ C(ku0k (t)k2L2 (Ω) + A[uk (t), uk (t); t] + kf (t)k2L2 (Ω) ),

where we used the Gårding’s inequality A[u, u; t] ≥ θkuk2H 1 (Ω) .
0
2. Now write
η(t) = ku0k (t)k2L2 (Ω) + A[uk (t), uk (t); t], ξ(t) = kf (t)k2L2 (Ω) .
Then
η 0 (t) ≤ Cη(t) + Cξ(t) ∀ a.e. t ∈ [0, T ].
Thus, by Gronwall’s inequality,
Z t
Ct
η(t) ≤ e η(0) + C ξ(s) ds (0 ≤ t ≤ T ).
0
Since
η(0) = ku0k (0)k2L2 (Ω) + A[uk (0), uk (0); 0] ≤ C(khk2L2 (Ω) + kgk2H 1 (Ω) ),
0
according to the initial data in (4.49) and kuk (0)kH01 (Ω) ≤ kgkH01 (Ω) , we thus obtain
max (ku0k (t)k2L2 (Ω) + A[uk (t), uk (t); t]) ≤ C(khk2L2 (Ω) + kgk2H 1 (Ω) + kf k2L2 (ΩT ) ).
t∈[0,T ] 0
This proves
(4.53) max (kuk (t)kH01 (Ω) + ku0k (t)kL2 (Ω) ) ≤ C(kf kL2 (ΩT ) + kgkH01 (Ω) + khkL2 (Ω) ).
t∈[0,T ]
3. Finally we need to estimate ku00k kL2 (0,T ;H −1 (Ω)) . So, fix any v ∈ H01 (Ω), with kvkH01 (Ω) ≤
1. We write v = v 1 + v 2 , where v 1 ∈ Vk , and (v 2 , wi ) = 0 for all i = 1, 2, · · · , k. (That is, v 2
is in the L2 orthogonal complement of Vk .) Since {wi } are orthogonal in H01 (Ω), we have
kv 1 kH01 (Ω) ≤ kvkH01 (Ω) ≤ 1.
Using (4.49), we have
(u00k (t), v 1 ) + B[uk (t), v 1 ; t] = (f (t), v 1 ).
Then
hu00k (t), vi = (u00k (t), v) = (u00k (t), v 1 ) = (f (t), v 1 ) − B[uk (t), v 1 ; t]
and consequently
|hu00k (t), vi| ≤ C(kf (t)kL2 (Ω) + kuk kH01 (Ω) ).
This implies
ku00k (t)k2H −1 (Ω) ≤ C(kf (t)k2L2 (Ω) + kuk (t)k2H 1 (Ω) ) ∀ t ∈ [0, T ].
0
Integrate over t ∈ [0, T ] to finally obtain

Z T
00 2
kuk kL2 (0,T ;H −1 (Ω)) = ku0k (t)k2H −1 (Ω) dt ≤ C(kf k2L2 (ΩT ) + kuk k2L2 (0,T ;H 1 (Ω)) ),
0
0
which, combined with the estimate (4.53), derives the desired estimate.
4.2.4. Existence and Uniqueness of Weak Solutions.

Theorem 4.13. (Existence) There exists a weak solution to (4.1).
Proof. 1. According to the energy estimate (4.50), we see that {uk } is bounded in
L2 (0, T ; H01 (Ω)), {u0k } is bounded in L2 (0, T ; L2 (Ω)), and {u00k } is bounded in L2 (0, T ; H −1 (Ω)).
Consequently there exists a subsequence {ukm } of {uk } with km → ∞ and functions
u ∈ L2 (0, T ; H01 (Ω)), with u0 ∈ L2 (0, T ; L2 (Ω)), u00 ∈ L2 (0, T ; H −1 (Ω)), such that

ukm * u
 in L2 (0, T ; H01 (Ω)),
(4.54) u0km * u0 in L2 (0, T ; L2 (Ω)),

 00
ukm * u00 in L2 (0, T ; H −1 (Ω)).
In fact, by estimate (4.50), we also have u ∈ L∞ (0, T ; H01 (Ω)) ∩ C([0, T ]; L2 (Ω)) and u0 ∈
L∞ (0, T ; L2 (Ω)); moreover, u0 ∈ C([0, T ]; H −1 (Ω)).
2. Fix any integer N and let ψ ∈ C 1 ([0, T ]; H01 (Ω)) have the form
N
X
ψ(t) = ζi (t)wi ,
i=1
where ζi ∈ C 1 ([0, T ]; R). Let k ≥ N , multiply (4.10) by ζi (t), sum i = 1, 2, · · · , N and then
integrate over t ∈ [0, T ] to find
Z T Z T
00

(4.55) huk (t), ψ(t)i + B[uk (t), ψ(t); t] dt = (f (t), ψ(t))dt.
0 0
Now let k = km → ∞ and we have
Z T Z T
00

(4.56) hu (t), ψ(t)i + B[u(t), ψ(t); t] dt = (f (t), ψ(t))dt.
0 0
This equality then holds for all functions ψ ∈ L2 (0, T ; H01 (Ω)), as functions ζ of the given
form are dense in this space. We then take ψ(t) = ζ(t)v with ζ ∈ L2 (0, T ) and v ∈ H01 (Ω)
in (4.56) to obtain
Z T Z T
ζ(t) hu00 (t), vi + B[u(t), v; t] dt =

ζ(t)(f (t), v)dt.
0 0
This holding for all ζ ∈ L2 (0, T ) yields that
(4.57) hu00 (t), vi + B[u(t), v; t] = (f (t), v) ∀ v ∈ H01 (Ω), a.e. t ∈ [0, T ].
3. We need to show the initial data u(0) = g, u0 (0) = h. In (4.55), (4.56), we take
ψ(t) = α(t)v + β(t)w with α, β ∈ C 2 [0, T ] and v, w ∈ H01 (Ω) arbitrarily given such that
ψ(T ) = ψ 0 (T ) = 0, ψ(0) = v and ψ 0 (0) = w. Note that
Z T Z T
hu00k (t), ψ(t)idt = −(u0k (0), v) + (uk (0), w) + (uk (t), ψ 00 (t)) dt,
0 0
(uk (0), w) → (g, w), (u0k (0), v) → (h, v) as k = km → ∞, and
Z T Z T
00 0
hu (t), ψ(t)idt = −hu (0), vi + (u(0), w) + (u(t), ψ 00 (t)) dt,
0 0
Hence, in (4.55), let k = km → ∞, we eventually obtain
−hu0 (0), vi + (u(0), w) = −(h, v) + (g, w) ∀ v, w ∈ H01 (Ω);
hence u(0) = g and u0 (0) = h.
Theorem 4.14. (Uniqueness) A weak solution of (4.43) is unique.
Proof. 1. It suffices to prove that a weak solution u with f = g = 0 must be zero. Unlike
the parabolic case, the proof here is tricky because we cannot insert v = u0 (t) in (4.49) since
u0 (t) ∈
/ H01 (Ω). We instead consider, for each fixed s ∈ [0, T ], the function
(R s
t u(τ )dτ if t ∈ [0, s],
v(t) =
0 if t ∈ [s, T ].
Then for each t ∈ [0, T ], v(t) ∈ H01 (Ω), and so, by the weak solution definition,
Z s
(hu00 (t), v(t)i + B[u(t), v(t); t]) dt = 0.
0
Since u0 (0) = v(s) = 0, we obtain by integration by parts
Z s
(4.58) (−hu0 (t), v 0 (t)i + B[u(t), v(t); t]) dt = 0.
0
As above, we write B[u, v; t] = A[u, v; t] + C[u, v; t]. Note, for all u, v ∈ H01 (Ω) and t ∈ [0, T ],
Z Xn Z X n
C[u, v; t] = bi Di uv + cuv dx = − (Di bi uv + bi uDi v) + cuv dx.
Ω i=1 Ω i=1
(The trick here is to avoid the Di u terms.) Since v 0 (t)

= −u(t) on [0, s], using (4.58), we
can write Z s Z s
0 0
(hu (t), u(t)i − A[v (t), v(t); t])dt = − C[u(t), v(t); t]dt.
0 0
0 d 1 2
Note that (u (t), u(t)) = dt ( 2 ku(t)kL2 (Ω) )
and, since A is symmetric,
Z Xn
0 d 1 1
A[v (t), v(t); t] = A[v(t), v(t); t] − a0ij Di vDj vdx.
dt 2 2 Ω
i,j=1
Hence
Z s
1 d
ku(t)k2L2 (Ω) − A[v(t), v(t); t] dt
2 0 dt
(4.59) Z s n
1 s
Z Z X
=− C[u(t), v(t); t]dt − a0ij Di vDj vdxdt,
0 2 0 Ω
i,j=1
and consequently,
Z s Z sZ n
X
ku(s)k2L2 (Ω) + A[v(0), v(0); 0] = −2 C[u(t), v(t); t]dt − a0ij Di vDj vdxdt
0 0 Ω i,j=1
Z s
≤C (kv(t)k2H 1 (Ω) + ku(t)k2L2 (Ω) ) dt.
0
0
By Gårding’s inequality for A[u, u; t], we obtain
Z s
2 2
(4.60) ku(s)kL2 (Ω) + kv(0kH 1 (Ω) ≤ C (kv(t)k2H 1 (Ω) + ku(t)k2L2 (Ω) ) dt.
0 0
0
2. Now let Z t
w(t) = u(τ )dτ (t ∈ [0, T ]).
0
Then v(0) = w(s) and v(t) = w(s) − w(t), and hence (4.60) becomes
Z s
2 2
ku(s)kL2 (Ω) + kw(s)kH 1 (Ω) ≤ C (kw(t) − w(s)k2H 1 (Ω) + ku(t)k2L2 (Ω) )dt.
0 0
0
But
kw(t) − w(s)k2H 1 (Ω) ≤ 2kw(t)k2H 1 (Ω) + 2kw(s)k2H 1 (Ω) ,
0 0 0
so we have
Z s
ku(s)k2L2 (Ω) + (1 − 2sC1 )kw(s)k2H 1 (Ω) ≤ C1 (kw(t)k2H 1 (Ω) + ku(t)k2L2 (Ω) )dt.
0 0
0
Choose 0 < T1 < T so small that 1 − 2T1 C1 ≥ 21 . Then if 0 ≤ s ≤ T1 , we have

Z s
2 2
ku(s)kL2 (Ω) + kw(s)kH 1 (Ω) ≤ C2 (kw(t)k2H 1 (Ω) + ku(t)k2L2 (Ω) )dt.
0 0
0
Hence Gronwall’s inequality implies u ≡ 0 on [0, T1 ].
3. Apply the same argument on the intervals [T1 , 2T1 ], [2T1 , 3T1 ], etc, to eventually
deduce u ≡ 0 on [0, T ].
4.2.5. Regularity. We now study the regularity of weak solutions when the initial data
and coefficients are more regular. Our eventual goal is to prove that the weak solution
is smooth, as long as the coefficients and initial data and the domain are all smooth.
Although the methods and results are similar to the ones used for the regularity study of
parabolic equations as above, as we shall see later, there are some quite essential differences
of regularity concerning these two classes of evolution equations.
We assume the coefficients aij , bi , c are smooth on ΩT . As usual, we assume the uniform
hyperbolicity.
Theorem 4.15. (Improved regularity) (i) Assume g ∈ H01 (Ω), h ∈ L2 (Ω), f ∈ L2 (ΩT ).
Suppose u is the weak solution of (4.43). Then
u ∈ L∞ (0, T ; H01 (Ω)), u0 ∈ L∞ (0, T ; L2 (Ω)),
with the estimate
(4.61) kukL∞ (0,T ;H01 (Ω)) + ku0 kL∞ (0,T ;L2 (Ω)) ≤ C(kgkH01 (Ω) + khkL2 (Ω) + kf kL2 (ΩT ) ),
(ii) If, in addition, g ∈ H 2 (Ω), h ∈ H01 (Ω), f 0 ∈ L2 (0, T ; L2 (Ω)), then
u ∈ L∞ (0, T ; H 2 (Ω)), u0 ∈ L∞ (0, T ; H01 (Ω)),
u00 ∈ L∞ (0, T ; L2 (Ω)), u000 ∈ L2 (0, T ; H −1 (Ω)),
with the estimate
kukL∞ (0,T ;H 2 (Ω)) + ku0 kL∞ (0,T ;H01 (Ω)) + ku00 kL∞ (0,T ;L2 (Ω))
(4.62)
+ ku000 kL2 (0,T ;H −1 (Ω)) ≤ C(kgkH 2 (Ω) + khkH01 (Ω) + kf kH 1 (0,T ;L2 (Ω) ),
Proof. Let {uk } be the Galerkin approximations satisfying (4.49) constructed as above
with {wi } being the complete collection of eigenfunctions with eigenvalues {λi } for −∆ on
H01 (Ω). As before, we assume {wi } is orthogonal on H01 (Ω) and orthonormal on L2 (Ω).
By the uniqueness theorem, the weak solution is obtained as the limit of the Galerkin
approximations {uk }. We prove the theorem by deriving the same estimates for these
approximate solutions independent of k.
1. By energy estimates (4.50), we have
(4.63) max (kuk (t)kH01 (Ω) + ku0k (t)kL2 (Ω) ) ≤ C(kf kL2 (ΩT ) + kgkH01 (Ω) + khkL2 (Ω) )
t∈[0,T ]
and thus we deduce (4.61).

2. Similar to the parabolic case, we claim the following estimate for {uk }: for each
t ∈ [0, T ],
(4.64) kuk (t)k2H 2 (Ω) ≤ C(kf (t)k2L2 (Ω) + ku00k (t)k2L2 (Ω) + kuk (t)k2L2 (Ω) ),
where C is a constant independent of k. We would easily obtain this if uk was itself a
weak solution of (4.43), since then we could use the elliptic estimate to the elliptic equation
Luk (t) = f (t) − u00k (t) on Ω. To prove (4.64), write the equation in (4.49) as
B[uk (t), wi ; t] = (f (t) − u00k (t), wi ) (i = 1, 2, · · · , k),
multiply this equation by λi di (t) and sum over i = 1, 2, · · · , k to deduce
(4.65) (Luk (t), −∆uk (t)) = B[uk (t), −∆uk (t); t] = (f (t) − u00k (t), −∆uk (t)),
since −∆uk (t) ∈ H01 (Ω). Then (4.64) follows from (4.65) by using Lemma 4.4.
3. Assume now the hypotheses of assertion (ii). We differentiate the equation in (4.49)
with respect to t and set ũk := u0k to obtain
(ũ00k (t), wi ) + B[ũk (t), wi ; t] = (f 0 (t), wi ) − B̃[uk (t), wi ; t],
where B̃ is the bilinear form defined by
Z X n n
X
0
B̃[u, v; t] = aij (x, t)Di uDj v + b0i (x, t)Di uv + c0 (x, t)uv dx.
Ω i,j=1 i=1
Multiplying by d00i (t) and summing over i = 1, 2, · · · , k, we discover

(ũ00k (t), ũ0k (t)) + B[ũk (t), ũ0k (t); t] = (f 0 (t), ũ0k (t)) − B̃[uk (t), ũ0k (t); t].
So
(ũ00k (t), ũ0k (t)) + A[ũk (t), ũ0k (t); t] =(f 0 (t), ũ0k (t)) − B̃[uk (t), ũ0k (t); t]
− C[ũk (t), ũ0k (t); t],
where, as above, we set B[u, v; t] = A[u, v; t] + C[u, v; t], with A being the symmetric part
of B and C the term involving no Di v terms. We write this equation as
1Z X n
1d 0 2
kũk (t)kL2 (Ω) + A[ũk (t), ũk (t); t] = a0ij Di ũk (t)Dj ũk (t)dx
(4.66) 2 dt 2 Ω i,j=1
+ (f 0 (t), ũ0k (t)) − B̃[uk (t), ũ0k (t); t] − C[ũk (t), ũ0k (t); t].
Notice that, for u ∈ H01 (Ω) ∩ H 2 (Ω), v ∈ H01 (Ω), integration by parts yields
Z X n n
X
B̃[u, v; t] = v (−Dj a0ij Di u − a0ij Dij u) + b0i (x, t)Di u + c0 (x, t)u dx,
Ω i,j=1 i=1
and, hence, for a.e. t ∈ [0, T ] and all u ∈ H 2 (Ω), v ∈ H01 (Ω),
|B̃[u, v; t]| ≤ C(kuk2H 2 (Ω) + kvk2L2 (Ω) ).
Moreover, for all u, v ∈ H01 (Ω),

|C[u, v; t]| ≤ C(kuk2H 1 (Ω) + kvk2L2 (Ω) ) ≤ C(A[u, u; t] + kvk2L2 (Ω) ).
0
Using these estimates in (4.66), we deduce from (4.64)

d 0
kũk (t)k2L2 (Ω) + A[ũk (t), ũk (t); t]
dt
(4.67) ≤ C(kũ0 (t)k2 2 + A[ũk (t), ũk (t); t] + kuk (t)k2 2 + kf 0 (t)k2 2 )
k L (Ω) H (Ω) L (Ω)
≤ C(kũ0k (t)k2L2 (Ω) + A[ũk (t), ũk (t); t] + kf (t)k2L2 (Ω) + kuk (t)k2L2 (Ω) + kf 0 (t)k2L2 (Ω) ).
Hence Gronwall’s inequality implies
kũ0k (t)k2L2 (Ω) + A[ũk (t), ũk (t); t]
Z T
0 2 0
≤ C kũk (0)kL2 (Ω) + A[ũk (0), ũk (0); 0]) + (kf (t)k2L2 (Ω) + kuk (t)k2L2 (Ω) + kf (t)k2L2 (Ω) )dt
0
≤ C(kũ0k (0)k2L2 (Ω) + kũk (0)k2H 1 (Ω) + kf k2H 1 (0,T ;L2 (Ω)) + kgk2H 1 (Ω) + khk2L2 (Ω) ).
0 0
Recall that ũk = u0k and

ku0k (0)kH01 (Ω) ≤ CkhkH01 (Ω) , ku00k (0)kL2 (Ω) ≤ C(kf (0)kL2 (Ω) + kuk (0)kH 2 (Ω) )
to simplify the previous estimate as
ku00k (t)k2L2 (Ω) + ku0k (t)k2H 1 (Ω) ≤ C(kuk (0)k2H 2 (Ω) + kf k2H 1 (0,T ;L2 (Ω)) + kgk2H 1 (Ω) + khk2H 1 (Ω) ).
0 0
Finally as above, kuk (0)kH 2 (Ω) ≤ CkgkH 2 (Ω) , from which and (4.64), we deduce
sup (ku00k (t)k2L2 (Ω) + kuk (t)k2H 2 (Ω) + ku0k (t)k2H 1 (Ω) )
0
t∈[0,T ]
(4.68)
≤ C(kf k2H 1 (0,T ;L2 (Ω)) + kgk2H 1 (Ω) + khk2H 1 (Ω) ).
0
4. As in the earlier proof for the parabolic equations, we can deduce the estimate for
u000 ∈ L2 (0, T ; H −1 (Ω)) in terms of the right-hand side of (4.68).
Theorem 4.16. (Higher regularity) Let m ∈ {0, 1, 2, · · · }. Assume
dk f
g ∈ H m+1 (Ω), h ∈ H m (Ω), ∈ L2 (0, T ; H m−k (Ω)) (k = 0, 1, · · · , m).
dtk
Suppose the following m-th order compatibility conditions hold:

∈ H01 (Ω), h1 := h ∈ H01 (Ω), · · · ,
g0 := g 2l−2

(4.69) g2l := dt2l−2f (0) − Lg2l−2 ∈ H01 (Ω) ( if m = 2l),
d
 2l−1
h2l+1 := ddt2l−1f (0) − Lh2l−1 ∈ H01 (Ω) ( if m = 2l + 1).

Then
dk u
∈ L∞ (0, T ; H m+1−k (Ω)) (k = 0, 1, 2, · · · .m + 1),
dtk
with the estimate
(4.70)
m+1 m
!
X dk u X dk f
k k kL∞ (0,T ;H m−k+1 (Ω)) ≤ C k k kL2 (0,T ;H m−k (Ω)) + kgkH m+1 (Ω) + khkH m (Ω) .
dt dt
k=0 k=0
Proof. Again use induction on m and differentiate the equation with respect to t. Details
are referred to Evans’s book.
Remark 4.8. The condition on f implies

f (0) ∈ H m−1 (Ω), f 0 (0) ∈ H m−2 (Ω), · · · , f (m−2) (0) ∈ H 1 (Ω)
and consequently
g0 ∈ H m+1 (Ω), h1 ∈ H m (Ω), g2 ∈ H m−1 (Ω), h3 ∈ H m−2 (Ω),
· · · , g2l ∈ H 1 (Ω) (if m = 2l), h2l+1 ∈ H 1 (Ω) (if m = 2l + 1).

The compatibility conditions are precisely the requirements that each of these functions is
zero on ∂Ω.
Theorem 4.17. (Smoothness of weak solution) Assume g, h ∈ C ∞ (Ω̄), f ∈ C ∞ (ΩT ),
and the m-th-order compatibility conditions hold for all m = 0, 1, · · · . Then problem (4.43)
has a unique solution u ∈ C ∞ (ΩT ).
4.2.6. Propagation of Disturbances. So far our study of hyperbolic equations has much
paralleled our treatment of parabolic equations, using the Galerkin method. However, we
learned that the classical solution to a second-order parabolic equation has the maximum
principle, which implies an infinite propagation speed of initial disturbances for such
equations. We now study a property for second-order hyperbolic equations that is totally
the opposite phenomenon, namely the finite propagation speed of initial disturbances.
For simplicity, we study the operator of nondivergence form
n
X
Lu = − aij (x)Dij u,
i,j=1
where the coefficients aij are smooth, independent of time,aij = aji , and satisfy the usual
uniform ellipticity condition.
Assume q(x) is a continuous function on Rn and smooth in Rn \ {x0 }, satisfying
 n
 q(x) > 0 in R \ {x0 }, q(x0 ) = 0,

n
(4.71) X

 aij (x)Di qDj q ≤ 1 in Rn \ {x0 }.
i,j=1
Given a t0 > 0, define the cone-like domain with vertex (x0 , t0 )

K = {(x, t) ∈ Rn × (0, t0 ) | q(x) < t0 − t}.
For each 0 < t < t0 , define
Kt = {x ∈ Rn | q(x) < t0 − t}.
Theorem 4.18. (Finite propagation speed) Let u = u(x, t) be a smooth solution of
utt + Lu = 0 in Rn × (0, ∞).
If u = ut ≡ 0 on K0 , then u ≡ 0 within K.
Proof. 1. Define the energy

 
Z n
1 u2t +
X
e(t) = aij Di uDj u dx (0 ≤ t ≤ t0 ).
2 Kt i,j=1
We compute e0 (t). In order to do so, note that if f (x, t) is continuous in x and smooth in t
then Z Z Z
d f (x, t)
f (x, t)dx = f (x, t)dx − dS,
dt Kt Kt ∂Kt |∇q(x)|
according to the co-area formula.
2. Therefore, we compute
Z n Z n
X 1 X 1
e0 (t) = ut utt + aij Di uDj ut dx − u2t + aij Di uDj u dS
Kt 2 ∂Kt |∇q|
i,j=1 i,j=1
:= A − B.
Using aij Di uDj ut = Dj (aij ut Di u) − ut Dj (aij Di u) and integration by parts, we have
Z X n Z Xn
A= ut utt − Dj (aij Di u) dx + aij ν j ut Di udS
Kt i,j=1 ∂Kt i,j=1
(4.72) Z n Z n
X X
=− ut Dj aij Di u dx + aij ν j ut Di udS,
Kt i,j=1 ∂Kt i,j=1
where ν = (ν 1 , · · · , ν n ) is the outer unit normal to ∂Kt . Since on ∂Kt , q(x) = t0 − t, we

∇q
have ν = |∇q| on ∂Kt ; that is, ν j = Dj q/|∇q| on ∂Kt . Since matrix (aij ) is symmetric and
positive definite, for each x ∈ Rn , the form hξ, ηi = ni,j=1 aij (x)ξi ηj (ξ, η ∈ Rn ) defines
P
an inner product on Rn , with norm kξk = hξ, ξi1/2 ; hence, the Cauchy-Schwarz inequality
|hξ, ηi| ≤ kξkkηk implies
 1/2  1/2
Xn Xn n
X
aij (x)ν j Di u ≤  aij (x)Di uDj u  aij (x)ν i ν j 
i,j=1 i,j=1 i,j=1
 1/2  1/2
n n
X X Di qDj q 
= aij (x)Di uDj u  aij (x)
|∇q|2
i,j=1 i,j=1
 1/2
n
X 1
≤ aij (x)Di uDj u (x ∈ ∂Kt ),
|∇q|
i,j=1
by (4.71). Returning to (4.72), we have

 1/2
n
|ut |
Z X
|A| ≤ Ce(t) +  aij (x)Di uDj u dS
∂Kt |∇q|
i,j=1
 
Z n
1 u2t +
X 1
≤ Ce(t) + aij (x)Di uDj u dS
2 ∂Kt |∇q|
i,j=1
= Ce(t) + B.
3. Therefore, we deduce
e0 (t) = A − B ≤ Ce(t) + B − B = Ce(t) (0 < t < t0 ).
Since e(0) = 0 and e(t) ≥ 0, we deduce from Gronwall’s inequality, that e(t) ≡ 0 for all
0 ≤ t ≤ t0 . This proves u ≡ 0 in K.
4.3. Hyperbolic Systems of First-order Equations 121
4.3. Hyperbolic Systems of First-order Equations

We broaden our study of hyperbolic PDE to the first-order PDE systems.
4.3.1. Notations and Definitions. Consider systems of linear first-order PDE having
the form
n
X
(4.73) ut + Bj (x, t)Dj u = f in Rn × (0, ∞),
j=1
subject to the initial condition
(4.74) u=g on Rn × {t = 0}.
The unknown is u : Rn × [0, ∞) → Rm , u = (u1 , · · · , um ), and the functions Bj : Rn ×
[0, ∞) → Mm×m (j = 1, 2, · · · , n), f : Rn × [0, ∞) → Rm and g : Rn → Rm are given,
Definition 4.9. (i) The system of PDE (4.73) is called a hyperbolic system if the m × m
matrix X
B(x, t; ξ) = ξj Bj (x, t)
j=1
is diagonalizable for each x, ξ ∈ Rn , t ≥ 0. In other words, (4.73) is a hyperbolic system
if for each x, ξ, t the matrix B(x, t; y) defined above has m real eigenvalues
λ1 (x, t; ξ) ≤ λ2 (x, t; ξ) ≤ · · · ≤ λm (x, t; ξ)
and corresponding eigenvectors {rk (x, t; ξ)}m m
k=1 that form a basis of R .
(ii) We say (4.73) is a symmetric hyperbolic system if Bj (x, t) is symmetric for each
x ∈ Rn , t ≥ 0.
(iii) The system is called strictly hyperbolic if if for each x, ξ, t the matrix B(x, t; ξ)
defined above has m distinct real eigenvalues
λ1 (x, t; ξ) < λ2 (x, t; ξ) < · · · < λm (x, t; ξ).
4.3.2. Vanishing Viscosity Method. We study the initial value problem (4.73),(4.74),
with
Bj ∈ C 2 (Rn × [0, T ]; Mm×m ) is symmetric,
2
(4.75) sup (|Bj | + |Dx,t Bj | + |Dx,t Bj |) < ∞,
Rn ×[0,T ]
g ∈ H 1 (Rn ; Rm ), f ∈ H 1 (Rn × (0, T ); Rm ).

In this section, we do not need the hyperbolicity of the system. We define the bilinear
form
Xn Z X n
(4.76) B[u, v; t] := Bj (·, t)Dj u, v = (Bj (x, t)Dj u(x)) · v(x) dx
j=1 Rn j=1
for all u, v ∈ H 1 (Rn ; Rm ), t ∈ [0, T ].

Definition 4.10. We say a function
u ∈ L2 (0, T ; H 1 (Rn ; Rm )), with u0 ∈ L2 (0, T ; L2 (Rn ; Rm )),
is a weak solution of the initial value problem (4.73), (4.74) provided
(i) (u0 , v) + B[u, v; t] = (f , v) for each v ∈ H 1 (Rn ; Rm ) and a.e. t ∈ [0, T ], and
(ii) u(0) = g. Again, by regularity, u ∈ C([0, T ]; L2 (Rn ; Rm )), so u(0) is well-defined.
We shall use the vanishing viscosity method to prove the existence of weak solution. To
this end, we approximate the initial value problem by the parabolic problem
(
ut − ε∆u + nj=1 Bj Dj u = f in Rn × (0, T ],
P
(4.77)
u = g on Rn × {t = 0},
for 0 < ε ≤ 1, gε := ηε ? g. The second-order term −ε∆u is called the viscosity term,
which tends to regularize the original first-order system.
Theorem 4.19. (Existence of approximate solutions) For each 0 < ε ≤ 1, there exists
a unique solution u = uε of (4.77), with
uε ∈ L2 (0, T ; H 3 (Rn ; Rm )), u0ε ∈ L2 (0, T ; H 1 (Rn ; Rm )).
Proof. 1. Set X = L∞ (0, T ; H 1 (Rn ; Rm )). For each v ∈ X, consider the linear system
(
ut − ε∆u = f − nj=1 Bj Dj v in Rn × (0, T ],
P
(4.78)
u = g on Rn × {t = 0}.
The right-hand side is bounded in L2 , there exists a unique solution u ∈ L2 (0, T ; H 2 (Rn ; Rm )),
with u0 ∈ L2 (0, T ; L2 (Rn ; Rm )). This solution u can be expressed by the Duhamel formula
using the heat kernel. From this we can also show that u ∈ X = L∞ (0, T ; H 1 (Rn ; Rm )).
Hence we define a map S : X → X by setting u = S(v). Let v1 ∈ X and u1 = S(v1 ). Set
w = u − u1 and z = v − v1 . Then
(
wt − ε∆w = − nj=1 Bj Dj z in Rn × (0, T ],
P
w = 0 on Rn × {t = 0}.
From the representation of w in terms of the heat kernel and nj=1 Bj Dj z, we have
P
n
X
kwkL∞ (0,T ;H 1 (Rn ;Rm )) ≤ C(ε)k Bj Dj zkL2 (0,T ;L2 (Rn ;Rm ))
j=1
≤ C(ε)kzkL2 (0,T ;H 1 (Rn ;Rm ))
≤ C(ε)T 1/2 kzkL∞ (0,T ;H 1 (Rn ;Rm )) .
Thus kwkX ≤ C(ε)T 1/2 kzkX ; that is,

(4.79) kS(v) − S(v1 )kX ≤ C(ε)T 1/2 kv − v1 kX .
3. If C(ε)T 1/2 < 1 then S is a strict contraction on X; hence it has a unique fixed point
u = uε : S(uε ) = uε . Then u = uε solves (4.77) for such a T > 0. If C(ε)T 1/2 ≥ 1, then
1/2
we choose 0 < T1 < T so that C(ε)T1 < 1 and repeat the above argument on the time
intervals [0, T1 ], [T1 , 2T1 ], etc, to obtain a weak solution u = uε for all T > 0. Finally the
high regularity of such a solution u follows from parabolic regularity theory.
Theorem 4.20. (Energy estimates) There exists a constant C, depending only on n and
the coefficients,such that
max (kuε (t)kH 1 (Rn ;Rm ) + ku0ε (t)kL2 (Rn ;Rm ) )
t∈[0,T ]
(4.80)
≤ C(kgkH 1 (Rn ;Rm ) + kf kL2 (0,T ;H 1 (Rn ;Rm ) + kf 0 kL2 (0,T ;L2 (Rn ;Rm ) )
for all 0 < ε ≤ 1.
Proof. 1. We compute
 
n
d 1 X
kuε (t)k2L2 (Rn ;Rm ) = (uε (t), u0ε (t)) = uε , f (t) + ε∆uε (t) − Bj Dj uε (t) .
dt 2
j=1
Note that
(uε (t), ε∆uε (t)) = −εk∇uε (t)k2L2 (Rn ) ≤ 0 (0 < t ≤ T ).
2. Suppose v ∈ C0∞ (Rn ; Rm ). Then, by the symmetry of Bj (this is the only place the
symmetry assumption is used: if B is symmetric then Ba · b = Bb · a),
X n Z X n
B[v, v; t] = v, Bj Dj v = (Bj Dj v) · v dx
j=1 Rn j=1
n Z n Z
1 X 1X
= Dj [(Bj v) · v]dx − [(Dj Bj )v) · v]dx
2 n 2 Rn
j=1 R j=1
n Z
1X
=− [(Dj Bj )v) · v]dx.
2 Rn
j=1
Hence, for all v ∈ C0∞ (Rn ; Rm ),

X n
(4.81) |B[v, v; t]| = | v, Bj Dj v | ≤ Ckvk2L2 (Rn ;Rm ) .
j=1
By approximation, (4.81) holds for all v ∈ H 1 (Rn ; Rm ). Hence

n
X
| uε (t), Bj Dj uε (t) | ≤ Ckuε (t)k2L2 (Rn ;Rm ) .
j=1
We therefore deduce

d 1
kuε (t)k2L2 (Rn ;Rm ) ≤ C kuε (t)k2L2 (Rn ;Rm ) + kf (t)k2L2 (Rn ;Rm ) .
dt 2
So Gronwall’s inequality and kgε kL2 ≤ kgkL2 will yield

(4.82) max kuε (t)k2L2 (Rn ;Rm ) ≤ C kgk2L2 (Rn ;Rm ) + kf k2L2 (0,T ;L2 (Rn ;Rm )) .
t∈[0,T ]
3. Differentiating the equation with respect xk , estimating Dk uε and summing k =

1, 2, · · · , n, we deduce

(4.83) max k∇uε (t)k2L2 (Rn ) ≤ C k∇gk2L2 (Rn ) + kf k2L2 (0,T ;H 1 (Rn ;Rm )) ,
t∈[0,T ]
where we have used k∇gε kL2 ≤ Ck∇gkL2 .

4. Next differentiating the equation with respect to t and setting v = u0ε , we have
(
vt − ε∆v + nj=1 Bj Dj v = f 0 − nj=1 B0j Dj Dj uε in Rn × (0, T ],
P P
(4.84)
v = f + ε∆gε − nj=1 Bj Dj gε on Rn × {t = 0}.
P
Reasoning as above, we deduce

max ku0ε (t)k2L2 (Rn ;Rm ) ≤ C k∇gk2L2 (Rn ) + ε2 k∆gε k2L2 (Rn ;Rm )
t∈[0,T ]
(4.85)
+ kf (0)k2L2 (Rn ;Rm ) + kf k2L2 (0,T ;H 1 (Rn ;Rm )) + kf 0 k2L2 (0,T ;L2 (Rn ;Rm )) .
Now, using gε = ηε ? g, we have

C
k∆gε k2L2 (Rn ;Rm ) ≤ k∇gk2L2 (Rn ) .
ε2
Furthermore,
kf (0)k2L2 (Rn ;Rm ) ≤ Ckf k2H 1 (0,T ;L2 (Rn ;Rm )) .
Combining all these estimates completes the proof.
Theorem 4.21. (Existence of weak solution by vanishing viscosity) There exists a
weak solution to problem (4.77) as certain limit of {uε } along a sequence ε → 0+ .
Proof. This follows from the energy estimates above in exactly the same fashion as before.

Theorem 4.22. (Uniqueness of weak solution) A weak solution to problem (4.77) is
unique.
Proof. It suffices to show that the only weak solution to (4.77) with f ≡ g ≡ 0 is u ≡ 0.
To verify this, note that (u0 (t), u(t)) + B[u(t), u(t); t] = 0 for a.e. t ∈ [0, T ] and, by (4.81),
|B[u(t), u(t); t]| ≤ Cku(t)k2L2 (Rn ;Rm ) ,
so we have
d
ku(t)k2L2 (Rn ;Rm ) ≤ Cku(t)k2L2 (Rn ;Rm ) ,
dt
whence Gronwall’s inequality forces u ≡ 0.
4.3.3. Systems with Constant Coefficients. In this section, we study

(
ut + nj=1 Bj Dj u = 0 in Rn × (0, ∞),
P
(4.86)
u=g on Rn × {t = 0}.
Here we assume that the coefficients Bj are constant m × m matrices and that, for each
ξ ∈ Rn , the matrix
Xn
B(ξ) = ξj Bj
j=1
has all real eigenvalues
λ1 (ξ) ≤ λ2 (ξ) ≤ · · · ≤ λm (ξ).
There is no hypothesis concerning the eigenvectors and there is no symmetry assumption
on Bj . The weak notion of hyperbolicity lies in the assumption that matrix B(ξ) has all real
eigenvalues for all ξ ∈ Rn . We will apply the Fourier transform to solve the corresponding
problem (4.86).
Theorem 4.23. (Existence of solution by Fourier transform) Assume g ∈ H s (Rn ; Rm )
with s > n2 + m. There exists a unique weak solution u ∈ C 1 (Rn × [0, ∞); Rm ) to the initial
value problem (4.86).
Proof. If u ∈ L2 (Rn ; Rm ), u = (u1 , · · · , um ), we use

û(ξ) = (û1 (ξ), · · · , ûm (ξ)),
where v̂(ξ) stands for the Fourier transform a function v ∈ L2 (Rn ). If u = u(x, t) ∈
L2x (Rn Rm ) for each t, then we use û(ξ, t) to denote the Fourier transform of u(·, t); we do
not transform with respect to t.
1. Taking the spatial Fourier transform, System (4.86) becomes

(4.87) ût (ξ, t) + iB(ξ)û(ξ, t) = 0, û(ξ, 0) = ĝ(ξ).
For fixed ξ ∈ Rn , we can solve (4.87) to find
(4.88) û(ξ, t) = e−itB(ξ) ĝ(ξ) (ξ ∈ Rn , t ≥ 0).
Consequently, via the Fourier inverse transform,
Z
1
(4.89) u(x, t) = eix·ξ e−itB(ξ) ĝ(ξ) dξ (x ∈ Rn , t ≥ 0).
(2π)n/2 Rn
2. We need to verify that the formula (4.89) indeed defines a function u ∈ C 1 (Rn ×
[0, ∞); Rm ) that is a solution to (4.86). Since g ∈ H s (Rn ; Rm ), we have
h(ξ) := (1 + |ξ|s )ĝ(ξ) ∈ L2 (Rn ; Rm ).
To show the integral in (4.89) converges, we must estimate ke−itB(ξ) k.
3. Let λ1 (ξ), · · · , λm (ξ) be the eigenvalues of B(ξ). Let Γ be the circle ∂B(0, r) in the
complex plane, traversed counterclockwise, with radius r > 0 so large that all λj (ξ) lie
inside Γ and dist(λj (ξ); Γ) ≥ 2. By Cauchy’s theorem in complex analysis, we have
Z
−itB(ξ) 1
e = e−itz (zI − B(ξ))−1 dz.
2πi Γ
4. Define a new path ∆ in the complex plane by
 
[m
∆=∂ B(λj (ξ), 1) ,
j=1
traversed counterclockwise. Deforming Γ to ∆, we have

Z
−itB(ξ) 1
e = e−itz (zI − B(ξ))−1 dz.
2πi ∆
Note that
|e−itz | ≤ et (z ∈ ∆),
and
| det(zI − B(ξ))| = |(z − λ1 (ξ))(z − λ2 (ξ)) · · · (z − λm (ξ))| ≥ 1 (z ∈ ∆).
So we can estimate the inverse matrix by
k(zI − B(ξ))−1 k ≤ kcof(zI − B(ξ))k ≤ C(1 + |ξ|m−1 + kB(ξ)km−1 ) ≤ C(1 + |ξ|m−1 ),
where we used the fact |λk (ξ)| ≤ C|ξ| since B(ξ) is linear in ξ. Combining these estimates,
we have
ke−itB(ξ) k ≤ et (1 + |ξ|m−1 ) (ξ ∈ Rn ).
5. Therefore,
ke−itB(ξ) kh(ξ)
Z Z
|eix·ξ e−itB(ξ) ĝ(ξ)| dξ ≤ C dξ
Rn Rn 1 + |ξ|s
Z
≤ Cet h(ξ)(1 + |ξ|−s )(1 + |ξ|m−1 ) dξ
Rn
≤ Cet khkL2 (Rn ) k(1 + |ξ|m−1−s )kL2 (Rn ) < ∞,
since s > n2 + m > n2 + m − 1. Hence the integral in (4.89) converges, and it follows easily
that the function u(x, t) defined is continuous on Rn × [0, ∞).
6. To show u is C 1 (Rn × [0, ∞); Rm ), observe for 0 < |h| ≤ 1,

u(x, t + h) − u(x, t)
Z
1
= ex·ξ (e−i(t+h)B(ξ) − e−itB(ξ) )ĝ(ξ) dξ.
h (2π)n/2 h Rn
Since
Z t+h
−i(t+h)B(ξ) −itB(ξ)
e −e = −i B(ξ)e−isB(ξ) ds,
t
we can easily estimate
1 −i(t+h)B(ξ)
e − e−itB(ξ) ≤ Cet+1 (1 + |ξ|m ).
h
From this we have
u(x, t + h) − u(x, t)
Z
≤ Cet+1 h(ξ)(1 + |ξ|m )(1 + |ξ|s )−1 dξ < ∞,
h Rn
since s > n2 + m. Therefore ut exists and is continuous on Rn × [0, ∞). A similar argument
shows that Dj u exists and is continuous for all j = 1, 2, · · · , n (the proof is similar since ξ
is like B(ξ)). Furthermore, we can differentiate inside the integral in (4.89) to confirm that
u solves the system (4.86).
4.4. Semigroup Theory

Semigroup theory is an abstract study of first-order ordinary differential equations with
values in Banach spaces, driven by linear, but possibly unbounded operators. The method
provides an elegant alternative to some of the existence theory for evolution equations set
forth above.
The whole idea of semigroups springs from properties of solutions of the elementary
initial value problem in u ∈ Rm :
du
(4.90) = Au (t > 0), u(0) = u0
dt
where A is a constant m × m matrix. The solution of course is
(4.91) u(t) = T (t)u0 = etA u0 .
This operator T (t) has some obvious properties:
(a) limt→0 T (t) = 1, (b) T (t1 )T (t2 ) = T (t1 + t2 ).
Effectively, T (t) maps the initial data u0 into the current value of the solution u(t).
Before we discuss the general theory for (4.90) with A being a linear operator defined
in a subspace of a Banach space, we review some definitions and elementary properties.
4.4.1. Definitions and Elementary Properties. Let X, Y be normed spaces. A linear

operator T : D(T ) ⊂ X → Y is said to be closed if whenever {xn } ⊂ D(T ) is a sequence
satisfying
xn → x, T xn → y
then x ∈ D(T ) and T x = y.
4.4. Semigroup Theory 127
Example 4.24. Let L : D(L) ⊂ L2 (0, 1) → L2 (0, 1) be the differential operator L = d/dx,
where D(L) = H01 (0, 1). To show that L is closed, let un → u, u0n → f in L2 (0, 1), where
un ∈ D(L). Passing to the limit in
Z 1 Z 1
un v 0 dx = − u0n vdx for all v ∈ C0∞ (0, 1)
0 0
we see, by the definition of weak derivative, that u0 = f and un → u in H 1 (0, 1). Since
H01 (0, 1) is closed, we have u ∈ H01 (0, 1).
Example 4.25. Let Ω be a bounded domain in Rn and let X, Y = L2 (Ω). Let

D(L) = H 2 (Ω) ∩ H01 (Ω)
Lu = ∆u, u ∈ D(L).
Note that we are considering L as an operator on L2 (Ω). Clearly L is densely defined. It is
unbounded, for if we consider {ϕn }, the sequence of eigenfunctions of −∆, then kϕn k2 = 1
while kLϕn k2 = λn → ∞ as n → ∞.
To see that L : D(L) ⊂ L2 (Ω) → L2 (Ω) is a closed operator, let un ∈ D(L) with
un → u, Lun → f . Applying the estimate kuk2,2 ≤ ckLuk2 to un − um , it follows that {un }
is a Cauchy sequence in H 2 (Ω) and thus kun − vk2,2 → 0 for some v ∈ H 2 (Ω). Clearly u = v
and u ∈ D(L). Since L : H 2 (Ω) → L2 (Ω) is continuous, Lun → Lu which yields Lu = f .
Hence L is closed.
Theorem 4.26. (Closed Graph Theorem) If X, Y are Banach spaces and if T : D(T ) ⊂
X → Y is a closed linear operator with closed domain, then T is bounded.
Corollary 4.27. If X, Y are Banach spaces and if T : D(T ) ⊂ X → Y is a closed linear

operator which is one-to-one, then T −1 is bounded iff R(T ) is closed. (In particular, if T
is 1-1 and onto, then T −1 is bounded.)
Definition 4.11. Let X be a Banach space and let T : D(T ) ⊂ X → X be a closed operator
on X.
(i) We say a real number λ belongs to the resolvent set ρ(T ) of T , provided the
operator
λI − T : D(T ) → X
is one-to-one and onto.
(ii) If λ ∈ ρ(T ), the resolvent operator Rλ : X → X is defined by
Rλ u := (λI − T )−1 u (u ∈ X);
that is, Rλ = (λI − T )−1 .
Remark 4.12. (i) Note that, since T is closed, by (4.27), Rλ is a bounded linear operator
on X if λ ∈ ρ(T ). Furthermore,
T Rλ u = Rλ T u (u ∈ D(T )).
(ii) We have the resolvent identity:

Rλ − Rµ = (µ − λ)Rλ Rµ , Rλ Rµ = Rµ Rλ (λ µ ∈ ρ(T )).
4.4.2. C0 Semigroups of Operators.
Definition 4.13. Let X be a Banach space. A family {T (t)} ⊂ B(X) (0 ≤ t < ∞) is called
a strongly continuous semigroup of operators if
(i) T (t)T (s) = T (t + s), t, s ≥ 0 (semigroup property)
(ii) T (0) = I
(iii) For all u ∈ X, T (t)u is strongly continuous in t ∈ [0, ∞), i.e.,
kT (t + h)u − T (t)uk → 0 as h → 0.
For simplicity we say that T (t) is a C0 semigroup.
Moreover, for fixed u ∈ X, we write T (·)u ∈ C([0, ∞); X). If in addition the map
t → T (t) is continuous in the uniform operator topology, i.e., kT (t + h) − T (t)k → 0 for
t, t + h ≥ 0, then the family {T (t)} is called a uniformly continuous semigroup. If
the C0 semigroup {T (t)} satisfies the property kT (t)k ≤ 1 for t ≥ 0, then it is called a
contraction semigroup.
Note that T (t) and T (s) commute as a consequence of (i), and that T (t)u is also strongly
continuous in u for each fixed t ≥ 0. (kT (t)u − T (t)vk ≤ kT (t)kku − vk.)
Example 4.28. Let A ∈ B(X) where X is a Banach space. Then the series ∞ n n
P
n=0 (A /n!)t
converges in the uniform operator topology for any real number t. In fact, set
n
X
Sn = (Ak /k!)tk
k=0
and observe that for m < n

n
X
kSn − Sm k ≤ (kAkk /k!)|t|k → 0 as m, n → ∞.
k=m+1
Thus {Sn } converges to a bounded linear operator, in the uniform operator topology, which
we denote by etA . From the above estimate we see that
ketA k ≤ e|t|kAk .
Let u(t) = etA u0 where u0 ∈ X. Since e(t+s)A = etA esA , it follows that
hA
u(t + h) − u(t) e −I
− Au = − A u.
h h
However,
∞
ehA − I X
k − Ak ≤ (kAkk /k!)|h|k−1
h
k=2
e|h|kAk −1
= − kAk → 0 as h → 0.
|h|
Hence u0 (t) exists for all t and equals Au, and so we have shown that u(t) = etA u0 satisfies
the Cauchy problem
du
= Au (t > 0), u(0) = u0 .
dt
Finally, we show that {etA }, t ≥ 0, is a uniformly continuous semigroup. In fact

ke(t+h)A − etA k = ketA (ehA − I)k
≤ e|t|kAk (e|h|kAk − 1) → 0 as h → 0.
Lemma 4.29. If {T (t)} is a C0 semigroup, then there exist constants ω ≥ 0 and M ≥ 1
such that
kT (t)k ≤ M eωt , t ≥ 0
i.e., kT (t)k grows slower than an exponential.
Proof. In fact, since the function g(t) = kT (t)uk is continuous on [0, 1] for each fixed
u ∈ X, we have supt∈[0,1] kT (t)uk < ∞. Hence, by the Uniform Boundedness Principle,
there is a constant M > 0 such that kT (t)k ≤ M for t ∈ [0, 1]. Let ω = log M . Then ω ≥ 0,
since M ≥ 1 by virtue of T (0) = I. Now let t be given and let n be the least integer greater
than or equal to t. By virtue of the semigroup property (T (t/n))n = T (t) and thus
kT (t)k = k(T (t/n))n k ≤ M n ≤ M t+1 = M eωt .

4.4.3. Infinitesimal Generator. Let {T (t)} be a C0 semigroup on the Banach space X.

For h > 0 we define the linear operator Ah by the formula
T (h)u − u
Ah u = , u ∈ X.
h
Definition 4.14. (i) Let D(A) be the set of all u ∈ X for which limh→0+ Ah u exists. Define
the operator A on D(A) by the relation
d
Au = lim Ah u = T (h)u|h=0+ (u ∈ D(A)).
h→0+ dh
The operator A is called the infinitesimal generator of the semigroup {T (t)}.
(ii) Given an operator A on D(A), we say that it generates a C0 semigroup {T (t)} if A
coincides with the infinitesimal generator of {T (t)}.
Example 4.30. Clearly A ∈ B(X) is the infinitesimal generator of {etA }, t ≥ 0.
Theorem 4.31. Let {T (t)} be a C0 semigroup on the Banach space X and let A : D(A) →
X be its infinitesimal generator. Then the following hold:
(a) D(A) is a subspace of X and A is a linear operator.
(b) If u ∈ D(A), then T (t)u ∈ D(A), 0 ≤ t < ∞, is strongly differentiable in t and
(4.92) (d/dt)T (t)u = AT (t)u = T (t)Au, t ≥ 0

(c) If u ∈ D(A), then
Z t
T (t)u − T (s)u = T (h)Au dh, t, s ≥ 0
s
(d) If f (t) is a real-valued continuous function for t ≥ 0, then
Z t+h
−1
lim h f (s)T (s)u ds = f (t)T (t)u, u ∈ X, t ≥ 0
h→0 t
Rt Rt
(e) 0 T (s)u ds ∈ D(A) and T (t)u = u + A 0 T (s)u ds, u ∈ X
(f) D(A) = X and A is a closed operator.
Proof. (a) This follows directly from the definition since Ah is linear.
(b) Since T (t) and T (h) commute and kT (t)k < ∞, we have
Ah T (t)u = T (t)Ah u → T (t)Au, as h → 0+ .
Hence
T (t)u ∈ D(A), AT (t)u = T (t)Au = D+ T (t)u.
Next we consider D− T (t)u if t > 0. Note that

T (t)u − T (t − h)u T (h)u − u
− T (t)Au = T (t − h) − Au
h h
+ (T (t − h) − T (t))Au.
But
T (h)u − u T (h)u − u
kT (t − h) − Au k ≤ M eωt k − Auk → 0
h h
as h → 0+ and
k(T (t − h) − T (t))Auk → 0 as h → 0+
which implies the desired result.
(c) The abstract function T (t)u is differentiable by (b) and its derivative T (t)Au is
continuous in t. The conclusion follows by integrating (4.92).
(d)
Z t+h Z t+h
−1 −1
kh f (s)T (s)u ds − f (t)T (t)uk = kh (f (s)T (s)u − f (t)T (t)u) dsk
t t
Z t+h Z t+h
−1 −1
≤h |f (s)|k(T (s)u − T (t)uk ds + h kT (t)uk|f (s) − f (t)| ds.
t t
Since the functions T (t)u and f (t) are continuous in t, by choosing h small enough so that
k(T (s)u − T (t)uk < ε and |f (s) − f (t)| < ε for |t − s| < h, the result easily follows.
(e) Let h > 0 and consider
Z t t
T (h) − I
Z
1
T (s)u ds = (T (s + h)u − T (s)u)ds
h 0 h 0
Z t+h Z h
1
= T (s)u ds − T (s)u ds
h t 0
→ T (t)u − u
Rt
by (d). Thus T (s)u ds ∈ D(A) and (e) follows by the definition of A.
0
Rh Rh
(f) Let u ∈ X. Then 0 T (s)u ds ∈ D(A) and by (d), limh→0 h−1 0 T (s)u ds =
T (0)u = u. Thus D(A) = X. Let un ∈ D(A) with un → u, Aun → v in X. Now
Z h
T (h)un − un −1
Ah u = lim = lim h T (s)Aun ds
n→∞ h n→∞ 0
where the last term follows from (c). But Aun → v and so
Z h
−1
Ah u = h T (s)v ds → v as h → 0
0
by virtue of (d). Thus u ∈ D(A) and Au = v.
Remark 4.15. Since the map t → T (t)Au is continuous, it follows from (b) that T (t)u is
continuously differentiable from [0, ∞) with values in X. Also T (t)u ∈ D(A) as proved, so
it has values in D(A) as well. Further, the continuous differentiability into X also proves
the continuity into D(A) (with the graph norm).
4.4.4. Application to Abstract Cauchy Problems. As an application we shall prove

that the abstract Cauchy problem
du
(4.93) = Au (t ≥ 0), u(0) = u0 , u0 ∈ D(A)
dt
has a unique solution.
Theorem 4.32. Let A : D(A) → X be the infinitesimal generator of a C0 semigroup
{T (t)}. Then the Cauchy problem (4.93) has the unique solution
u(t) = T (t)u0 , t ≥ 0.
Proof. The existence is a consequence of Theorem 4.31 (b). To prove uniqueness, let v(t)
be any solution of the Cauchy problem and set F (s) = T (t − s)v(s). Then
F (s + h) − F (s) = T (t − s − h) (v(s + h) − v(s)) − (T (t − s) − T (t − s − h)) v(s).
Since v(s) ∈ D(A), Theorem 4.31 implies that F (s) is strongly differentiable in s and
(d/ds)F (s) = −AT (t − s)v(s) + T (t − s)v 0 (s)
= −AT (t − s)v(s) + T (t − s)Av(s) = 0, 0 ≤ s ≤ t.
Hence by the mean value theorem, F (s) =constant for 0 ≤ s ≤ t. In particular, F (t) = F (0)
or v(t) = T (t)u0 = u(t).
Remark 4.16. (i) If T (t)u0 is differentiable for every u0 ∈ X and t ≥ 0, then in particular,
(d/dt)(T (t)u0 )|t=0 = Au0 . Hence, D(A) = X and A (being a closed operator) must be
bounded. Thus if A is unbounded, then T (t)u0 is not differentiable for all u0 ∈ X. We can
however consider u(t) = T (t)u0 as a generalized solution of the Cauchy problem.
(ii) From the uniqueness in Theorem 4.32 we have that a linear operator A : D(A) → X
which is densely defined can be the infinitesimal generator of at most one C0 semigroup
{T (t)}. Moreover, if {T (t)} is a C0 semigroup whose infinitesimal generator A is bounded,
then T (t) = eAt since then A is the infinitesimal generator of T (t) and etA .
4.4.5. Characterization of Generators. We know the generator A of a C0 semigroup is

a closed, densely defined linear operator. Now given a closed, densely defined linear operator
A on a Banach space X, we would like to know whether A generates a C0 semigroup on X;
if it does, then the abstract Cauchy problem (4.93) has a unique solution.
Suppose A generates a C0 contraction semigroup {T (t)}, i.e., kT (t)k ≤ 1 for all t ≥ 0.
For u ∈ X, λ > 0, the integral
Z t
e−λτ T (τ )udτ
s
is well defined, and as kT (τ )uk ≤ kuk for all τ , we deduce that this integral tends to zero
as t, s → ∞. Hence the integral
Z ∞
R(λ; A)u = e−λτ T (τ )udτ
0
exists as an improper Riemann integral. Since

Z ∞
kR(λ; A)uk ≤ kuk e−λτ dτ = (1/λ)kuk
0
it follows that R(λ; A) is a bounded linear operator and kR(λ; A)k ≤ 1/λ for all λ > 0.
Lemma 4.33. If A is the infinitesimal generator of a contraction semigroup {T (t)}, then
(λI − A) is invertible for every λ > 0 and
(λI − A)−1 = R(λ; A).
In particular, for every λ > 0, k(λI − A)−1 k ≤ λ−1 . (Thus the resolvent operator Rλ =
(λI − A)−1 is the Laplace transform of the semigroup.)
Proof. Let h > 0. Then for u ∈ X

Z ∞
T (h) − I
R(λ; A)u = (1/h) e−λτ (T (τ + h)u − T (τ )u)dτ
h
Z0 ∞ Z ∞
−λ(τ −h)
= (1/h) e T (τ )udτ − (1/h) e−λτ T (τ )udτ
h 0
λτ Z ∞ −λh Z h
e −1 e
= e−λh T (τ )udτ − e−λτ T (τ )udτ
h 0 h 0
→ λR(λ; A)u − u
Thus R(λ; A)u ∈ D(A) and AR(λ; A)u = λR(λ; A)u − u, i.e., (λI − A)R(λ; A)u = u.
Now let u ∈ D(A). Then
Z ∞ Z ∞
−λτ d
R(λ; A)Au = e T (τ )Audτ = e−λτ (T (τ )u)dτ
0 0 dτ
Z ∞
= λ e−λτ T (τ )udτ − u
0
= λR(λ; A) − u
i.e., R(λ; A)(λI − A)u = u. This proves R(λ; A) = (λI − A)−1 = Rλ .
Theorem 4.34. (Hille-Yosida Theorem) A linear operator A : D(A) → X is the gen-
erator of a C0 contraction semigroup if and only if A is closed, densely defined and
(4.94) (0, ∞) ⊂ ρ(A), k(λI − A)−1 k ≤ λ−1 ∀ λ > 0.
Proof. 1. If A is the generator, (4.94) follows from the previous lemma.

2. Now assume A is closed, densely defined, and satisfies (4.94). We must build a C0
contraction semigroup whose infinitesimal generator is A. For this, fix λ > 0 and defined
Aλ := −λI + λ2 Rλ = λARλ .
The operator Aλ is a kind of regularization of A.
3. We first claim
(4.95) lim Aλ u = Au (u ∈ D(A)).
λ→∞
Indeed, for each u ∈ D(A), since λRλ u − u = ARλ u = Rλ Au, we have

1
kλRλ u − uk ≤ kRλ kkAuk ≤ kAuk → 0 as λ → ∞.
λ
Hence λRλ u → u as λ → ∞ if u ∈ D(A). But since kλRλ k ≤ 1 and D(A) is dense, we have
lim λRλ u = u (u ∈ X).
λ→∞
So, if u ∈ D(A),
lim Aλ u = lim λARλ u = lim λRλ Au = Au.
λ→∞ λ→∞ λ→∞
4. Define
∞
2 tR
X (λ2 t)k
Tλ (t) := etAλ = e−λt eλ λ
= e−λt Rλk .
k!
k=0
Observe, since kRλ k ≤ 1/λ,
∞
−λt
X (λ2 t)k
kTλ (t)k ≤ e =1 (t ≥ 0).
k!
k=0
Consequently, {Tλ (t)}t≥0 is a contraction semigroup, with generator Aλ .

5. By the resolvent identity above, we see Aλ Aµ = Aµ Aλ for all λ, µ > 0. So
Aµ Tλ (t) = Tλ (t)Aµ (t > 0).
We thus compute
Z t Z t
d
Tλ (t)u − Tµ (t)u = [Tµ (t − τ )Tλ (τ )u]dτ = Tµ (t − τ )Tλ (τ )(Aλ u − Aµ u)dτ.
0 dτ 0
Consequently, if u ∈ D(A), then
kTλ (t)u − Tµ (t)uk ≤ tkAλ u − Aµ uk → 0 as λ, µ → ∞.
This proves that {Tλ (t)u}λ>0 is a Cauchy sequence in X. Hence define
T (t)u := lim Tλ (t)u (u ∈ D(A), t ≥ 0).
λ→∞
From kTλ (t)k ≤ 1, it is straightforward to show that {T (t)}t≥0 is a C − 0 contraction

semigroup.
6. Finally we show the generator of {T (t)}t≥0 is A. Write B to denote this generator.
Now Z t
Tλ (t)u − u = Tλ (τ )Aλ udτ
0
and
Tλ (τ )Aλ u − T (τ )Auk ≤ kTλ (τ )kkAλ u − Auk + kTλ (τ )Au − T (τ )Auk → 0
as λ → ∞ for all u ∈ D(A). Thus we have
Z t
T (t)u − u = T (τ )Aλ udτ (u ∈ D(A)).
0
This, recalling the definition of D(B), proves D(A) ⊆ D(B) and, for all u ∈ D(A),
T (h)u − u
Bu = lim = Au.
h→0+ h
To show D(B) ⊆ D(A), note (0, ∞) ⊂ ρ(A) ∩ ρ(B) and also (λI − B)(D(A)) = (λI −
A)(D(A)) = X. So, if y ∈ D(B), then there exists a x ∈ D(A), such that (λI − B)y =
λy − By = (λI − A)x = λx − Ax = λx − Bx and hence y = x ∈ D(A). This proves
D(B) ⊆ D(A). So A = B.
Remark 4.17. A C0 semigroup {T (t)} is called a C0 ω-contraction semigroup, provided

for some ω ∈ R,
(4.96) kT (t)k ≤ eωt (t ≥ 0).
In this case, T1 (t) = e−ωt T (t) will be a C0 contraction semigroup. If A is the generator of
T , then A − ωI is the generator of T1 and if A is the generator of T1 , then A + ωI is the
generator of T . Thus we can deduce the following result:
Theorem 4.35. A linear operator A : D(A) → X is the generator of a C0 ω-contraction
semigroup {T (t)} if and only if A is closed, densely defined and satisfies
1
(ω, ∞) ⊂ ρ(A), k(λI − A)−1 k ≤ ∀ λ > ω.
λ−ω
To characterize the infinitesimal generators of general C0 semigroups, we usually renorm
the Banach space so that {T (t)} becomes a C0 contraction semigroup in the new (equivalent)
norm. We just state the following general result without proof.
Theorem 4.36. (Hille-Yosida-Phillips) A linear operator A : D(A) → X is the gener-
ator of a C0 semigroup {T (t)} if and only if A is closed, densely defined and there exist
constants M ≥ 1, ω ∈ R such that λ ∈ ρ(A) for each λ > ω and
M
k(λI − A)−n k ≤ , ∀ λ > ω, n = 1, 2, . . . .
(λ − ω)n
In this case kT (t)k ≤ M eωt .
4.4.6. Another Characterization of C0 Contraction Semigroups. We now give a

different characterization of generators of semigroups of contractions in a Hilbert space H.
Definition 4.18. (i) A linear operator A : D(A) → H is said to be dissipative if
Re(Au, u) ≤ 0 for all u ∈ D(A).
(Note that if H is a real Hilbert space, then A is dissipative iff −A is monotone.)
(ii) We say that A is maximal dissipative if in addition R(I − A) = H. It is maximal
in the sense that there exists no linear operator B with the same properties and such that B
is an extension of A. Indeed, suppose such an extension exists. Let u ∈ D(B) and v ∈ D(A)
be such that (I − A)v = (I − B)u. Since Av = Bv, we have (I − B)(v − u) = 0. Multiplying
this by (v − u) we get
0 ≤ kv − uk2 = Re(B(v − u), v − u) ≤ 0
whence u = v or u ∈ D(A). Thus B = A.
Lemma 4.37. If A is dissipative, then
(4.97) k(λI − A)uk ≥ λkuk for all λ > 0, u ∈ D(A).
Proof.
k(λI − A)ukkuk ≥ Re((λI − A)u, u) = λkuk2 − Re(Au, u) ≥ λkuk2 .

Theorem 4.38. (Lumer-Phillips) Let A : D(A) → H be a densely defined linear operator.
(a) If A is a generator of a C0 contraction semigroup {T (t)}, then A is dissipative
and R(λI − A) = H for all λ > 0.
(b) If A is dissipative and there exists a λ0 > 0 such that R(λ0 I − A) = H, then
A is a generator of a C0 contraction semigroup. In particular, A is maximal
dissipative.
Proof. (a) By the Hille-Yosida Theorem, (0, ∞) ⊂ ρ(A) and therefore λI − A is onto H
for all λ > 0. Furthermore,
|(T (t)u, u)| ≤ kT (t)kkuk2 ≤ kuk2 .
Hence
T (t)u − u 1
Re , u = (Re(T (t)u, u) − kuk2 ) ≤ 0.
t t
+
Let u ∈ D(A) and let t → 0 obtaining Re(Au, u) ≤ 0.
(b) Since R(λ0 I − A) = H, it follows from (4.97) that λ0 ∈ ρ(A) and A is closed. If
R(λI − A) = H for all λ > 0, then (0, ∞) ⊂ ρ(A) and kR(λ; A)k ≤ λ−1 by (4.97). The
desired result then follows from the Hille-Yosida Theorem.
In order to prove that R(λI − A) = H for all λ > 0, consider the open set
Γ = {λ : λ > 0 and R(λI − A) = H}.
Note that λ ∈ Γ implies λ ∈ ρ(A), and since ρ(A) is open, there is a neighborhood of λ
whose intersection with the real line is in Γ. By hypothesis, λ0 ∈ Γ. Hence if Γ is closed in
(0, ∞), then Γ = (0, ∞). Let λn → λ > 0, λn ∈ Γ. For every v ∈ H, there exist un ’s such
that
(4.98) λn un − Aun = v for all n.
From (4.97) it follows that kun k ≤ λ−1
n kvk ≤ c. Now by (4.97) again,
λn kun − um k ≤ kλn (un − um ) − A(un − um )k

= kv − λn um + Aum k = kv − λn um + λm um − vk
= |λn − λm |kum k ≤ c|λn − λm |.
Therefore {un } is a Cauchy sequence with limit, say, u. Thus by (4.98), Aun → λu − v and
since A is closed, u ∈ D(A) and λu − Au = v. Thus λ ∈ Γ.
Corollary 4.39. Let A : D(A) → H be a densely defined closed linear operator. If both A
and A* are dissipative, then A is a generator of a C0 contraction semigroup.
Proof. In view of (b) of Theorem 4.38, it suffices to show that R(I − A) = H. Since A is
closed so is I − A. Moreover, (4.97) and Corollary 4.27 imply R(I − A) is a closed subspace
of H and therefore R(I − A) = (N (I − A*))⊥ . Consequently, it suffices to show that I −A*
is one-to-one. But this follows from the fact that A* is dissipative.
While we could solve the initial value problem in [0, ∞) when u0 ∈ D(A) and not for
general u0 ∈ H, one can show that if A is self-adjoint then we can solve the problem for
any initial data u0 ∈ H. The price we pay for this is the lack of differentiability at t = 0.
Theorem 4.40. Let A be a self-adjoint maximal dissipative operator and let u0 ∈ H. Then
there exists a unique u such that
u ∈ C([0, ∞); H) ∩ C 1 ((0, ∞); H) ∩ C((0, ∞); D(A))
and u satisfies the initial value problem (4.93). (Continuity in D(A) is with respect to the
graph norm.)
Proof. We only prove uniqueness. Let u1 , u2 be two solutions of (4.93). Set v(t) = ku1 (t)−
u2 (t)k2 which is continuous and such that v(0) = 0. Taking the inner product of u0 = Au
with u1 (t) − u2 (t) for u = u1 and u = u2 and using the dissipativity of A, we get
d
v(t) ≤ 0
dt
whence it follows that v(t) ≡ 0 and so u1 (t) = u2 (t) for t ≥ 0.
4.4.7. Applications. We demonstrate in this section that certain second-order parabolic

and hyperbolic PDE can be realized within the semigroup framework. The solvability of the
Cauchy problem for u0 = Au is reduced to the study of the solvability and a-priori estimate
of the solutions of the usually simpler problem λu − Au = v, u ∈ D(A) where v ∈ X and
λ > ω.
A. Second-order Parabolic Equations. We consider the following initial-boundary

value problem (IBVP):

ut + Lu = 0 in ΩT ,

(4.99) u=0 on ∂Ω × [0, T ],

u=g on Ω × {t = 0},

which a special case studied above. We assume L has the divergence structure, satisfies the
uniform ellipticity condition, and has coefficients that are smooth and independent of time
t. We assume domain Ω is bounded and has smooth boundary ∂Ω.
We propose problem (4.99) as the flow determined by a semigroup on X = L2 (Ω). For
this we set
D(A) = H01 (Ω) ∩ H 2 (Ω)
and define
Au := −Lu (u ∈ D(A)).
Recall the Gårding’s inequality
(4.100) βkuk2H 1 (Ω) ≤ B[u, u] + γku2L2 (Ω) .
0
Theorem 4.41. (Second-order parabolic PDE as semigroup) The operator A gen-

erates a C0 γ-contraction semigroup {T (t)}t≥0 on L2 (Ω). Therefore u(t) = T (t)g defines
the solution to (4.99) for any given g ∈ H01 (Ω) ∩ H 2 (Ω).
Proof. We must verify the hypotheses of the Hille-Yosida-Phillips Theorem. Therefore

the evolution problem becomes a study of the spectral properties of the elliptic operator A
defined above.
1. D(A) is clearly dense in L2 (Ω).
2. We show A is closed. Indeed, let {uk } be a sequence in D(A) with
uk → u, Auk → f in L2 (Ω).
According the global regularity,
kuk − ul kH 2 (Ω) ≤ C(kAuk − Aul kL2 (Ω) + kuk − ul kL2 (Ω) )
for all k, l. This shows that {uk } is a Cauchy sequence in H 2 (Ω) and so
uk → u ∈ H 2 (Ω) in H 2 (Ω).
Moreover, u ∈ H01 (Ω). (Trace operator is continuous from H 1 (Ω) to L2 (∂Ω).) Therefore
u ∈ D(A). Furthermore the strong convergence uk → u in H 2 (Ω) implies Auk → Au in
L2 (Ω), and thus f = Au. By definition, A is closed.
3. We next prove the condition (γ, ∞) ⊂ ρ(A); that is, λI − A is one-to-one and onto
for all λ > γ. By Lax-Milgram’s Theorem, for all λ ≥ γ, the BVP
(
Lu + λu = f in Ω,
(4.101)
u=0 on ∂Ω
has a unique weak solution u ∈ H01 (Ω) for each f ∈ L2 (Ω). The global regularity theory
shows that u ∈ D(A), and
λu − Au = f.
Thus (λI − A) : D(A) → X is one-to-one and onto, for all λ ≥ γ. This proves [γ, ∞) ⊂ ρ(A).
4. Let Rλ = R(λ, A) = (λI − A)−1 . We will show
1
kRλ k ≤ (λ > γ)
λ−γ
as required for generating a C0 γ-contraction semigroup. To show this, consider the weak
form of problem (4.101):
B[u, v] + λ(u, v) = (f, v) ∀ v ∈ H01 (Ω),
where, as usual, (, ) stands for the L2 -inner product. Set v = u and recall the Gårding’s
inequality to compute
(λ − γ)kuk2L2 (Ω) ≤ kf kL2 (Ω) kukL2 (Ω) .
This implies, as u = Rλ f ,
1
kRλ f kL2 (Ω) = kukL2 (Ω) ≤ kf kL2 (Ω) .
λ−γ
This bound is valid for all f ∈ L2 (Ω), which proves the desired claim.
Example 4.42. We study the heat equation:

ut (x, t) − ∆u(x, t) = 0 in Ω × [0, ∞),

(4.102) u(x, t) = 0 on ∂Ω × [0, ∞),

u(x, 0) = u0 (x), x ∈ Ω.

Theorem 4.43. Let u0 ∈ L2 (Ω). Then there exists a unique solution u of (4.102) such that
u ∈ C([0, ∞); L2 (Ω)) ∩ C 1 ((0, ∞); L2 (Ω)) ∩ C((0, ∞); H 2 (Ω) ∩ H01 (Ω)).
Proof. Let H = L2 (Ω) and define A : D(A) ⊂ H → H by

D(A) = H 2 (Ω) ∩ H01 (Ω); Au = ∆u for u ∈ D(A).
Using the concept of an abstract function, the IBVP can be posed as the abstract Cauchy
problem
du
= Au (t ≥ 0), u(0) = u0 ∈ H.
dt
Since for u ∈ H 2 (Ω) ∩ H01 (Ω)
Z Z
(Au, u) = u∆u dx = − |∇u|2 dx ≤ 0,
Ω Ω
we have that A is dissipative. It is maximal dissipative; that is, R(I − A) = H, for, by the
Lax-Milgram, there exists a unique u ∈ H01 (Ω) such that
(I − A)u = u − ∆u = f for all f ∈ L2 (Ω).
Also by elliptic regularity, u ∈ H 2 (Ω) and so u ∈ D(A); hence R(I − A) = H. Finally, we
saw earlier that A is self-adjoint. Hence we can apply Theorem 4.40 to deduce the desired
conclusions.
Remark 4.19. Note that however badly behaved u0 ∈ L2 (Ω) may be, u(x, t) is very smooth
for all t > 0. This is known as the strong regularizing effect of the heat operator. In
particular, this shows that the heat equation is irreversible in time, i.e., we cannot always
solve the problem 
ut (x, t) − ∆u(x, t) = 0 in Ω × [0, T ),

u(x, t) = 0 on ∂Ω × (0, T ),

u(x, T ) = uT (x), x ∈ Ω.

B. Second-order Hyperbolic Equations. We consider the following initial-boundary

value problem (IBVP):

utt + Lu = 0 in ΩT ,

(4.103) u=0 on ∂Ω × [0, T ],

u = g, ut = h on Ω × {t = 0},

which a special case studied above. We assume domain Ω is bounded and has smooth
boundary ∂Ω and L has the symmetric form:
Xn
Lu = − Dj (aij (x)Di u) + c(x)u,
i,j=1
where aij (x) = aji (x) and c(x) ≥ 0 are all smooth functions and (aij (x)) ≥ θI. Hence
(4.104) B[u, u] ≥ θkuk2H 1 (Ω) (u ∈ H01 (Ω)).
0
We recast problem (4.103) as a first-order system by setting v := ut . Then (4.103) reads


ut = v, vt + Lu = 0 in ΩT ,

(4.105) u=0 on ∂Ω × [0, T ],

u = g, v = h on Ω × {t = 0},

Now take Hilbert space X = H01 (Ω) × L2 (Ω), with the inner product
hh(u, v), (f, g)ii = B[u, f ] + (v, g)L2 (Ω)
and the norm
k(u, v)k := (B[u, u] + kvk2L2 (Ω) )1/2 .
Define
D(A) := [H 2 (Ω) ∩ H01 (Ω)] × H01 (Ω)
and define
A(u, v) := (v, −Lu) ∀ (u, v) ∈ D(A).
Theorem 4.44. (Second-order hyperbolic PDE as semigroup) The operator A gen-
erates a C0 contraction semigroup {T (t)}t≥0 on H01 (Ω) × L2 (Ω). Therefore (u(t), v(t)) =
T (t)(g, h) defines the solution to (4.105) for any given (g, h) ∈ [H01 (Ω) ∩ H 2 (Ω)] × H01 (Ω).
Proof. We must verify the hypotheses of the Hille-Yosida-Phillips Theorem. Therefore the
evolution problem again becomes a study of the spectral properties of the linear operator
A defined above.
1. D(A) is clearly dense in L2 (Ω).
2. We show A is closed. Indeed, let {(uk , vk )} be a sequence in D(A) with
(uk , vk ) → (u, v), A(uk , vk ) → (f, g) in X = H01 (Ω) × L2 (Ω).
Since A(uk , vk ) = (vk , −Luk ), we conclude f = v and Luk → −g in L2 (Ω). As before,

by elliptic estimates and regularity, uk → u in H 2 (Ω) and g = −Lu. Thus (u, v) ∈
D(A), A(u, v) = (v, −Lu) = (f, g). By definition, A is closed.
3. We next prove (0, ∞) ⊂ ρ(A); that is, λI − A is one-to-one and onto for all λ > 0.
Now given λ > 0 and (f, g) ∈ X = H01 (Ω) × L2 (Ω), consider the operator equation
(λI − A)(u, v) = λ(u, v) − A(u, v) = (f, g).
This is equivalent to the two scalar equations:

(
λu − v = f (u ∈ H 2 (Ω) ∩ H01 (Ω)),
(4.106)
λv + Lu = g (v ∈ H01 (Ω)).
But this implies

λ2 u + Lu = λf + g (u ∈ H 2 (Ω) ∩ H01 (Ω)).
Since λ2 > 0, by existence and regularity of elliptic PDE, this problem has a unique solution
u ∈ H 2 (Ω) ∩ H01 (Ω). Once u is found, define v = λu − f ∈ H01 (Ω). We have shown that
(4.106) has a unique solution (u, v). This proves that (λI − A) : D(A) → X is one-to-one
and onto, for each λ > 0; hence (0, ∞) ⊂ ρ(A).
4. Let Rλ = R(λ, A) = (λI − A)−1 . We will show
1
kRλ k ≤ (λ > 0),
λ
as required for generating a C0 contraction semigroup. Note that Rλ (f, g) = (u, v) if and
only if (4.106) holds. From the second equation in (4.106), we deduce
λkvk2L2 (Ω) + B[u, v] = (g, v)L2 (Ω) .
Putting v = λu − f, we obtain
λ(kvk2L2 (Ω) + B[u, u]) = (v, g)L2 (Ω) + B[u, f ] = hh(u, v), (f, g)ii ≤ k(u, v)kk(f, g)k.
This implies
1
k(u, v)k ≤ k(f, g)k;
λ
so
1
kRλ (f, g)k = k(u, v)k ≤ k(f, g)k.
λ
1
This bound is valid for all (f, g) ∈ X, which proves the desired claim kRλ k ≤ λ for all
λ > 0.
4.4.8. Nonhomogeneous Problems. We consider the nonhomogeneous Cauchy problem

du
(4.107) − Au = f (t) (t > 0), u(0) = u0 , u0 ∈ D(A).
dt
A function u(t) ∈ D(A) is called a classical solution of (4.107) if it continuous for
t ≥ 0, continuously differentiable for t > 0 and satisfies (4.107).
Theorem 4.45. Let A : D(A) → X be the infinitesimal generator of a C0 semigroup {T (t)}.
Let f : [0, ∞) → X be continuously differentiable. Then the Cauchy problem (4.107) has a
unique solution
Z t
(4.108) u(t) = T (t)u0 + T (t − s)f (s) ds, t ≥ 0.
0
Proof. Obviously u(0) = u0 . Define the function

Z t Z t
v(t) = T (t − s)f (s) ds = T (s)f (t − s) ds.
0 0
Since T (t) is bounded for each t ≥ 0 and f (s) is continuous, the above integrals exist. Now
Z t+h Z t
[v(t + h) − v(t)]/h = h−1 T (s)f (t + h − s)ds − h−1 T (s)f (t − s)ds
0 0
Z t
= T (s)[f (t + h − s) − f (t − s)]/h ds
0
Z t+h
−1
+ h T (s)f (t + h − s) ds.
t
Hence v 0 (t) exists and

Z t
v 0 (t) = T (s)f 0 (t − s) ds + T (t)f (0).
0
On the other hand, for h > 0 we have

Z t+h Z t
−1 −1
[v(t + h) − v(t)]/h = h T (t + h − s)f (s)ds − h T (t − s)f (s)ds
0 0
Z t
= [T (h) − I]/h T (t − s)f (s) ds
0
Z t+h
+ h−1 T (t + h − s)f (s) ds.
t
Since the limit on the left exists and also

Z t+h
lim h−1 T (t + h − s)f (s)ds = f (t)
h→0 t
it follows that Z t Z t
lim Ah T (t − s)f (s) ds = A T (t − s)f (s) ds
h→0 0 0
which yields
Z t
v 0 (t) = A T (t − s)f (s) ds + f (t).
0
As a result we get
Z t
du
= AT (t)u0 + A T (t − s)f (s) ds + f (t)
dt 0
= Au(t) + f (t)
and the proof is complete.
Remark 4.20. The expression (4.108) is called the variation of constants or Duhamel
formula. If the function f is integrable, then (4.108) still makes sense. Then u defined
by that formula is called a generalized solution or mild solution of (4.107). It can be
shown that a generalized solution always exists, but need not be a classical solution.
Index
C0 ω-contraction semigroup, 134 convex, 18

ith-difference quotient of size h, 48 critical point, 18, 20
m-th order compatibility conditions, 118
m-th-order compatibility conditions, 109 Difference quotient and weak derivatives Theorem, 48
(Bochner) integrable, 53 differentiable, 51
(uniformly) hyperbolic, 111 differentiable on I, 51
(uniformly) parabolic, 98 differential operator
uniformly elliptic, 55
abstract function, 51 divergence form, 55
Riemann integral, 51 dual space, 6
adjoint bilinear form, 65 second dual space, 6
almost separably valued, 53 Duhamel formula, 141
Ascoli Theorem, 8
Ehrling’s inequality, 21
Banach space eigenfunctions, 10, 67
reflexive, 6 eigenspace, 68
separable, 2 eigenvalue, 67
Banach-Steinhaus Theorem, 5 Eigenvalue Theorem Theorem, 67
Bessel’s inequality, 9 eigenvalues, 10
biharmonic, 61 eigenvectors, 10
bilinear form, 55, 58 Energy estimates Theorem, 122
bounded, 58 equicontinuous, 8
strongly positive, 58 Estimates for W 1,p , n < p ≤ ∞ Theorem, 44
Bochner-Pettis Theorem Theorem, 53 evolution equations, 97
Bounded Inverse Theorem, 6 Existence of approximate solutions Theorem, 122
Existence of solution by Fourier transform Theorem,
Calculus in W 1,p (I; X) Theorem, 54 124
Campanato ’63 Theorem, 86 Existence of weak solution by vanishing viscosity
Campanato Lemma Theorem, 92 Theorem, 124
Caratheodory property, 13 Existence Theorem, 114
Cartesian product, 3 extreme, 19
Cauchy sequence, 2 maximal, 20
Cauchy-Schwarz inequality, 4 minimal, 20
classical solution, 140
Closed Graph Theorem Theorem, 127 finite propagation speed, 119
co-area formula, 120 Finite propagation speed Theorem, 119
coercivity, 59 First Existence Theorem for weak solutions Theorem,
compact linear operator, 65 59
compactly embedded, 45 fixed point, 11
Construction of approximate solutions Theorem, 99, formal adjoint, 65
112 Fourier series, 9
continuation method, 11 Fourier transform, 49
continuous, 51 Fourier transforms, 62
Contraction Mapping Theorem, 11 Frechet differentiable, 13
143
144 Index
Fredholm Alternative Theorem, 9 More calculus Theorem, 54

Fredholm alternatives, 65 Morrey Theorem, 88
function Morrey’s Inequality Theorem, 42
cut-off function, 28
mollified function, 27 Neumann boundary, 60
p-mean continuous, 28 nonlinear functional analysis, 11
support, 2 norm, 2
trace, 36, 37 equivalent, 47
functional
convex, 18 operator
weakly coercive, 17 bounded, 5
weakly continuous, 17 closed, 126
weakly lower semicontinuous, 17 compact, 8
contraction, 11
Gårding’s estimate for system Theorem, 63 contraction semigroup, 128
Gårding’s estimates, 59 dissipative, 134
Gagliardo-Nirenberg-Sobolev Inequality Lemma, 40 Hilbert space adjoint, 7
Galerkin method, 97 infinitesimal generator, 129
Galerkin’s method, 99 Laplace operator, 39
Gateaux differentiable, 13 linear operator, 5
generalized solution, 131, 141 maximal dissipative, 134
Global H 2 -regularity Theorem, 72, 82 Nemytskii operator, 12
Green’s 1st identity, 38 operator norm, 5
Green’s 2nd identity, 39 resolvent, 10
strongly continuous, 8
Hahn-Banach Theorem, 6 strongly continuous semigroup of operators, 128
Higher global regularity Theorem, 75 uniformly continuous semigroup, 128
Higher interior regularity Theorem, 71 weakly continuous, 8
Higher regularity Theorem, 109, 118 ordinary point, 20
Higher-order traces Theorem, 38 orthogonal, 4
Hilbert space, 4 orthogonal complement, 4
Hille-Yosida Theorem Theorem, 132 orthonormal, 9
Hille-Yosida-Phillips Theorem, 134 complete, 9
hyperbolic system, 121 orthonormal basis, 67
Implicit Function Theorem, 16 parabolic boundary, 97

Improved regularity Theorem, 105, 116 Parabolic Harnack inequality Theorem, 110
infinite propagation speed, 119 Parallelogram law, 4
Infinite smoothness Theorem, 76 partition of unity, 63
initial-boundary value problem, 97, 111 Partition of Unity Theorem, 28
inner product, 4 Plancherel’s Theorem Theorem, 50
integration-by-parts formula, 48 Poincaré’s inequality, 48
Interior H 2 -regularity Theorem, 70 point spectrum, 10
inverse Fourier transform, 49 principal eigenvalue, 67
projection, 4
John-Nirenberg Theorem, 89 Projection Theorem, 4
Property of Fourier Tranforms Theorem, 50
Lagrange multiplier, 20
Lagrange Theorem, 20 regularity, 69
Lax-Milgram Theorem, 58 Rellich-Kondrachov Theorem, 45
Lax-Milgram Theorem Theorem, 58 resolvent identity, 127, 133
Lebesgue Differentiation Theorem Lemma, 82 resolvent operator, 127
Legendre ellipticity condition, 57 resolvent set, 10, 127
Legendre-Hadamard condition, 57 Riesz Representation Theorem, 6
linear functionals
bounded linear functionals, 6 Second Existence Theorem for weak solutions
continuous linear functionals, 6 Theorem, 65
Liouville’s theorem, 79 second-order differential operator, 55
Ljusternik Theorem, 21 Second-order hyperbolic PDE as semigroup Theorem,
Lumer-Phillips Theorem, 134 138
Second-order parabolic PDE as semigroup Theorem,
Mapping into better spaces Theorem, 54 136
Mean Value Theorem Theorem, 51 self-adjoint, 66
method of continuity, 11 Semigroup method, 97
mild solution, 141 seminorm, 47
Index 145
sequence
weakly convergent, 7
set
convex, 1
totally bounded, 45
weakly closed, 17
simple, 52
Smoothness of weak solution Theorem, 109, 119
Sobolev conjugate, 40
Sobolev Inequalities Theorem, 39
Sobolev space of order k and power p, 25
spectrum, 10
strictly hyperbolic, 121
Strong Maximum Principle, 69
strong maximum principle, 110
Strong Maximum Principle Theorem, 110
strong solution, 73
strongly measurable, 53
structural conditions, 56
symmetric hyperbolic system, 121
system of second-order (quasilinear) partial
differential equations in divergence form, 56
Trace Theorem, 37
Triangle inequality, 2, 4
Uniqueness of weak solution Theorem, 124

Uniqueness Theorem, 115
vanishing viscosity method, 122

variation of constants, 141
vector space, 1
Banach space, 2
basis, 2
complete, 2
dimension, 1
inner product space, 4
normed space, 2
subspace, 1
viscosity term, 122
weak derivative, 23, 54

Weak Maximum Principle Theorem, 110
weak solution, 56, 98, 111, 121
weakly compact, 8
weakly measurable, 53

Full-Note FPR Partition of Unity P-32 Thm2.7

Uploaded by

Copyright:

Available Formats

Full-Note FPR Partition of Unity P-32 Thm2.7

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

Full-Note FPR Partition of Unity P-32 Thm2.7

Uploaded by

Copyright:

Available Formats

*************************************

Applied Analysis I - (Advanced PDE I)

1.1. Banach Spaces

A nonempty subset M of a vector space X is called a subspace of X if αx + βy ∈ M

that each x ∈ X has a (unique) representation of the form

1.1.2. Normed Spaces. A (real) vector space X is said to be a normed space if to

1.1.3. Completeness and Banach Spaces. A sequence {xn } in a normed space X is

sup |fn (x) − fm (x)| < ε if m, n > N (ε).

kf k∞ ≡ ess-supx∈Ω |f (x)| < ∞.

[x1 , y1 ] + [x2 , y2 ] = [x1 + x2 , y1 + y2 ]

α[x, y] = [αx, αy].

k[x, y]k = kxkX + kykY .

Proof. (a) is obvious if x = 0, and otherwise it follows by taking δ = −(x, y)/kxk2 in

Furthermore, by (d), equation (1.1) defines a norm on an inner product space H. If H

Proof. Choose xn ∈ S so that kxn k → d ≡ inf{kxk : x ∈ S}. Since (1/2)(xn + xm ) ∈ S,

If (x, y) = 0, then x and y are said to be orthogonal, written sometimes as x ⊥ y. For

Remark. In particular, if M is a proper closed subspace of H, then there is a nonzero

1.2. Bounded Linear Operators

The collection of all bounded linear operators T : X → Y will be denoted by B(X, Y ). We

The following results can be found in [?].

We shall need the following properties of reflexive spaces.

1.2.3. Bounded Linear Functionals on a Hilbert Space.

Proof. If f (x) = 0 for all x, take y = 0. Otherwise, there is an element z ∈ N (f )⊥ such

Let T : H → H be an operator on the Hilbert space H. We define the Hilbert space

1.3. Weak Convergence and Compact Operators

The main result of this section is given by:

1.3.2. Compact Operators. Let X and Y be normed spaces. An operator T : X → Y is

(d) If T : X → Y is strongly continuous and X is a reflexive Banach space, then T

Proof. (a) Let xn * x. Then for every g ∈ Y ∗

1.4. Spectral Theory for Compact Linear Operators

1.4.2. Spectrum of Compact Operators. A subset S of a Hilbert space H is said to be

Finally we have the following result.

Theorem 1.27. Let H be a separable Hilbert space and suppose T : H → H is linear,

1.5. Nonlinear Functional Analysis

1.5.1. Contraction Mapping Theorem. Let X be a normed space. A map T : X → X

Theorem 1.28. (Contraction Mapping) Let T : S ⊂ X → S be a contraction on the

Proof. To prove uniqueness, suppose T x = x, T y = y. Since k < 1, we get x = y from

Hence as n → ∞, kxn+p − xn k → 0 independently of p, so that {xn } is a Cauchy sequence

Theorem 1.29. Let T0 , T1 ∈ B(X, Y ), where X is a Banach space and Y is a normed

Proof. Set S = {t ∈ [0, 1] : R(Tt ) = Y }. By hypothesis, 0 ∈ S. We need to show that

1.5.2. Nemytskii Operators. Let Ω be a nonempty measurable set in Rn and let f :

and (iii) we get

Let X, Y be Banach spaces and let A : S ⊂ X → Y be an arbitrary operator. A is

B ∈ B(X, Y ) such that

Proof. For u ∈ B(u0 , r)

The point u0 ∈ X is called a critical point of the functional f defined on X if f 0 (u0 )v =

1.5.6. Monotone Operators and Convex Functionals. Let A : X → X* be an oper-

Let C be a convex set in the real normed space X. A functional f : C ⊂ X → R is said

Proof. Assume ϕ is convex. Then

1.5.7. Lagrange Multipliers. Let f, g : X → R be two functionals defined on the Banach

(1.14) f 0 (u0 )v · g 0 (u0 )w 6= f 0 (u0 )w · g 0 (u0 )v.

Proof. Fix v, w ∈ X, and for s, t ∈ R consider the real-valued functions

More generally, one can prove the following:

As an application of Ljusternik’s theorem we have

Proof. Since g(u) is weakly continuous, it follows that MR = {u : g(u) = R} is weakly

The following interpolation inequality, which is frequently referred to as Ehrling’s

2.1. Weak Derivatives and Sobolev Spaces

In contrast to the corresponding classical derivative, the weak derivative Dα u is defined

Hence |x1 |x1 = sgn x1 . Similarly, since for i ≥ 2