Sobolev spaces and applications by IITK

Sobolev Spaces and Applications
T. Muthukumar
tmk@iitk.ac.in
December 17, 2023

ii
Contents
Notations vii
1 Theory of Distributions 1
1.1 History . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.1 Case 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.2 Case 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Space of Test Functions . . . . . . . . . . . . . . . . . . . . . 4
1.3.1 Smooth Non-Analytic (Bump) Functions . . . . . . . . 6
1.3.2 Mollifiers . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3.3 Topology on C(Ω), C0 (Ω) and Cc (Ω) . . . . . . . . . . 12
1.3.4 Algebraic and Topological Dual of Cc (Ω) . . . . . . . . 13
1.3.5 Topology on C ∞ (Ω) . . . . . . . . . . . . . . . . . . . 15
1.3.6 Inductive Limit Topology on Cc∞ (Ω) . . . . . . . . . . 17
1.3.7 Regularization and Cut-off Technique . . . . . . . . . . 19
1.4 Space of Distributions . . . . . . . . . . . . . . . . . . . . . . 25
1.4.1 Functions as Distributions . . . . . . . . . . . . . . . . 28
1.4.2 Measures as Distributions . . . . . . . . . . . . . . . . 31
1.4.3 Multipole Distributions . . . . . . . . . . . . . . . . . . 35
1.4.4 Infinite Order Distributions . . . . . . . . . . . . . . . 38
1.4.5 Topology on Distributions . . . . . . . . . . . . . . . . 38
1.4.6 Principal Value Distribution . . . . . . . . . . . . . . . 42
1.4.7 Functions, but not Distributions . . . . . . . . . . . . . 47
1.5 Operations With Distributions . . . . . . . . . . . . . . . . . . 47
1.5.1 Differentiation . . . . . . . . . . . . . . . . . . . . . . . 48
1.5.2 Product . . . . . . . . . . . . . . . . . . . . . . . . . . 59
1.5.3 Support . . . . . . . . . . . . . . . . . . . . . . . . . . 64
1.5.4 Singular Support . . . . . . . . . . . . . . . . . . . . . 67
iii
CONTENTS iv
1.5.5 Shifting and Scaling . . . . . . . . . . . . . . . . . . . 68

1.5.6 Convolution . . . . . . . . . . . . . . . . . . . . . . . . 69
1.6 Tempered Distributions . . . . . . . . . . . . . . . . . . . . . . 72
1.7 Fundamental Solution . . . . . . . . . . . . . . . . . . . . . . 75
1.8 Further Reading . . . . . . . . . . . . . . . . . . . . . . . . . . 76
2 Sobolev Spaces 77
2.1 Hölder Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
2.2 W k,p (Ω) Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 85
2.3 Smooth Approximations . . . . . . . . . . . . . . . . . . . . . 90
2.4 Characterisation of W0k,p (Ω) . . . . . . . . . . . . . . . . . . . 95
2.5 Extension Operators . . . . . . . . . . . . . . . . . . . . . . . 96
2.6 Topological Dual of Sobolev Spaces . . . . . . . . . . . . . . . 101
2.7 Fractional Order Sobolev Space . . . . . . . . . . . . . . . . . 103
2.8 Imbedding Results . . . . . . . . . . . . . . . . . . . . . . . . 108
2.8.1 Sobolev Inequality (1 ≤ p < n) . . . . . . . . . . . . . 109
2.8.2 Poincaré Inequality . . . . . . . . . . . . . . . . . . . . 116
2.8.3 Equality case, p = n . . . . . . . . . . . . . . . . . . . 118
2.8.4 Morrey’s Inequality (n < p < ∞) . . . . . . . . . . . . 119
2.8.5 Generalised Sobolev Imbedding . . . . . . . . . . . . . 122
2.8.6 Compact Imbedding . . . . . . . . . . . . . . . . . . . 124
2.9 Trace Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
2.10 Further reading . . . . . . . . . . . . . . . . . . . . . . . . . . 142
3 Bounded Variation Functions 143

3.1 De Giorgi Perimeter . . . . . . . . . . . . . . . . . . . . . . . 144
4 Second Order Elliptic PDE 147

4.1 Homogeneous Dirichlet Problem . . . . . . . . . . . . . . . . . 147
4.2 Inhomogeneous Dirichlet problem . . . . . . . . . . . . . . . . 150
4.3 Linear Elliptic Operators . . . . . . . . . . . . . . . . . . . . . 151
4.4 Periodic Boundary Conditions . . . . . . . . . . . . . . . . . . 159
4.5 Coercive Neumann Problem . . . . . . . . . . . . . . . . . . . 160
4.6 Semi-coercive Neumann Problem . . . . . . . . . . . . . . . . 164
4.7 Mixed and Robin Problem . . . . . . . . . . . . . . . . . . . . 169
4.8 p-Laplacian Operator . . . . . . . . . . . . . . . . . . . . . . . 171
4.9 Stokes Problem . . . . . . . . . . . . . . . . . . . . . . . . . . 172
4.10 Elasticity System . . . . . . . . . . . . . . . . . . . . . . . . . 173
CONTENTS v
4.11 Leray-Lions or Nonlinear Monotone Operators . . . . . . . . . 173

4.12 Solutions for Measure data . . . . . . . . . . . . . . . . . . . . 174
4.12.1 Stampacchia Solutions (p = 2 and n > 2) . . . . . . . . 175
4.12.2 Renormalized Solutions (1 ≤ p < n) . . . . . . . . . . . 175
Appendices 177
A Cantor Set and Cantor Function 179
Bibliography 185
Index 187
CONTENTS vi
Notations
Symbols
Ω denotes an open subset of Rn , not necessarily bounded
R denotes the real line
Rn denotes the n-dimensional Euclidean space
|E| will denote the Lebesgue measure of E ⊂ Rn
Function Spaces
C k,γ (Ω) all functions in C k (Ω) whose k-th partial derivatives (k ≥ 0) are
Hölder continuous with exponent γ
Lip(E) denotes the space of all Lipschitz functions on E
C(Ω) is the class of all continuous functions on Ω
C k (Ω) is the class of all k-times (k ≥ 1) continuously differentiable functions

on Ω
C ∞ (Ω) denotes class of C ∞ (Ω) functions such that all its derivatives can be
extended continuously to Ω
Cc∞ (Ω) is the class of all infinitely differentiable functions on Ω with compact
support
∞
CK (Ω) is the class of all infinitely differentiable functions on Ω with support
contained in the compact subset K ⊂ Ω
C0 (Ω) is the class of all continuous functions on Ω that vanishes at boundary

(for bounded sets) or ∞
vii
NOTATIONS viii
H k (Ω) is the space W k,2 (Ω)
H0k (Ω) is the space W0k,2 (Ω)
E(Ω) is the class of all infinitely differentiable functions on Ω, also denoted

as C ∞ (Ω), endowed with topology of uniform convergence on compact
subsets
D(Ω) is the class of all infinitely differentiable functions on Ω with compact

support endowed with inductive limit topology
W k,p (Ω) is the class of all functions in Lp (Ω) whose distributional derivative
upto order k are also in Lp (Ω)
General Conventions
M(α) denotes, for α > 0, the class of all n × n matrices, A = A(x), with
L∞ (Ω) entries such that,
α|ξ|2 ≤ A(x)ξ.ξ a.e. x ∀ξ ∈ Rn

Chapter 1
Theory of Distributions
1.1 History
In late 1920’s P. A. M. Dirac (cf. [Dir49]) derived the equation
d 1
ln(x) = − iπδ(x) (1.1.1)
dx x
in the study of quantum theory of collision processes, where δ is a “function”
defined and continuous in real line R satisfying the following properties:
1. δ(x) = 0 for x 6= 0.
R∞
2. −∞ δ(x) dx = 1.
R∞
3. For any continuous function defined on R, f (a) = −∞
f (x)δ(a − x) dx
for all a ∈ R.
4. δ is infinitely differentiable and for

R ∞any k-times continuously differen-
(k) (k)
tiable function f on R, f (a) = −∞ f (x)δ (a − x) dx for all a ∈ R.
5. Given the Heaviside function

(
1 if x > 0
H(x) = (1.1.2)
0 if x < 0,
then the delta function is the “derivative” of H, i.e., δ(x) = H 0 (x).
1
CHAPTER 1. THEORY OF DISTRIBUTIONS 2
The δ became known as Dirac’s delta function. It had caused unrest among
the mathematicians of the era, because δ did not adhere to the classical
notion of function.
H. Lebesgue had laid the foundations of measure theory in his doctoral
thesis in 1902. With this knowledge, δ can be viewed as a set function on
the σ-algebra(a collection of subsets of R). δ can be viewed as a positive
measure, called Dirac measure, defined on subsets E from the σ-algebra of
R as, (
1 if 0 ∈ E
δ(E) =
0 if 0 ∈/ E.
Thus, mathematically one has given sense to properties in 1, 2 and 3. How-
ever, the differentiability of δ and H was yet to be made precise, since clas-
sically, every differentiable function is continuous.
In 1944, G. Choquet and J. Deny published a work on polyharmonic
functions in two dimensions (cf. [CD44]). Laurent Schwartz, in an attempt
to generalise the work of Choquet and Deny in higher dimensions, published
an article in 1945 (cf. [Sch45]). With this article, the theory of distributions
was discovered and also settled the issue raised in properties 4 and 5, in
addition to 1, 2 and 3.
Dirac won the physics Nobel prize (1933) and Schwartz won the Fields
medal (1950) for their respective work.
1.2 Motivation
The theory of distribution is a concept that generalises the notion of function,
hence is also called generalised functions. But the beauty of the theory lies in
the fact that this new notion admits the concept of differentiation and every
distribution is infinitely differentiable. The need for such a generalisation
was felt at various instances in the history of mathematics.
1.2.1 Case 1
Recall that the wave equation utt (x, t) = c2 uxx (x, t) on R × (0, ∞), describ-
ing the vibration of an infinite string, has the general solution u(x, t) =
F (x + ct) + G(x − ct). In reality, it happens that even if F and G are not
twice differentiable, they are “solution” to the wave equation. For instance,
consider F to be the Heaviside function H (cf. (1.1.2)) and G ≡ 0 or vicev-

ersa. Now, how do we mathematically describe this nature of “solution”?
1.2.2 Case 2
Consider the Burger’s equation

ut (x, t) + u(x, t)ux (x, t) = 0 x ∈ R and t ∈ (0, ∞)
u(x, 0) = φ(x) x ∈ R
and the corresponding characteristic equations:
dx(r, s) dt(r, s) dz(r, s)

= z, = 1, and =0
ds ds ds
with initial conditions,
x(r, 0) = r, t(r, 0) = 0, and z(r, 0) = φ(r).
Solving the ODE corresponding to z, we get z(r, s) = c3 (r) with initial

conditions z(r, 0) = c3 (r) = φ(r). Thus, z(r, s) = φ(r). Using this in the
ODE of x, we get
dx(r, s)
= φ(r).
ds
Solving the ODE’s, we get
x(r, s) = φ(r)s + c1 (r), t(r, s) = s + c2 (r)
with initial conditions

x(r, 0) = c1 (r) = r
and
t(r, 0) = c2 (r) = 0.
Therefore,
x(r, s) = φ(r)s + r, and t(r, s) = s.
Eliminating s, we get the characteristic curves on the x − t plane on which
we know u is constant. Thus, the characteristic curves are x = φ(x0 )t + x0
and value of u on this curve is φ(x0 ). Hence u(φ(x0 )t + x0 , t) = φ(x0 ).
Suppose we choose φ to be the function


1
 x<0
φ(x) = 1 − x 0 ≤ x ≤ 1

0 1 ≤ x.

Then the characteristic curves are


t + x 0
 x0 < 0
x = (1 − x0 )t 0 ≤ x0 ≤ 1

x0 1 ≤ x0 .

Note that the characteristic curves are intersecting. This is no issue, as long
as, the solution u which is constant on these curves takes the same constant
in each characteristic curve. Unfortunately, that is not the case here:

1
 x<t
x
u(x, t) = t+1 0 ≤ 1 − xt ≤ 1

0 1 ≤ x.

This situation is, usually, referred to as shock waves.
1.3 Space of Test Functions

The space of test functions will be the argument for the ‘generalised function’
or distributions, a notion to be introduced subsequently. Let (a, b) ⊂ R be
a non-empty open subset, not necessarily bounded. Let Cc∞ (a, b) denote the
class of all infinitely differentiable functions on Ω with compact support in
(a, b). The genius of L. Schwartz is the choice of this function space. The
motivation of this choice is the following: If f ∈ C 1 (Ω) and φ ∈ Cc∞ (a, b),
then by classical integration by parts we have
Z b Z b
0
φ(x)f (x) dx = − f (x)φ0 (x) dx. (1.3.1)
a a
There are no boundary integrals above because φ has compact support in

(a, b) and vanishes at the end-points a and b. This explains the choice of
compact support. Suppose f was chosen from C k (a, b), then above integra-
tion by parts could be repeated k times to get
Z b Z b
(k) k
φ(x)f (x) dx = (−1) f (x)φ(k) dx.
a a
Rb Rb
Observe that the maps φ 7→ a f φ dx and φ 7→ a f φ0 dx are linear on
Cc∞ (a, b). If there exists a suitable complete topology on Cc∞ (a, b) such that
these linear maps are continuous, then f can be identified with a continuous
linear functional on Cc∞ (a, b) given by
Z b
φ 7→ f φ dx (1.3.2)
a
and the derivative of f , f 0 , can be identified with the mapping φ mapped

to the right hand side of (1.3.1), even if f 6∈ C 1 (Ω), as long as the integrals
make sense.
Let Ω ⊂ Rn be a non-empty open subset. Let Cc∞ (Ω) denote the class
of all infinitely differentiable functions on Ω with compact support. Observe
that Cc∞ (Ω) is a vector space under usual addition and scalar multiplication
of real-valued functions. Of course, the zero function is in Cc∞ (Ω). We know
that polynomials, trigonometric functions, exponential are all C ∞ functions.
Any function in Cc∞ (Ω) falls to zero within a compact set.
Exercise 1. Show that any non-zero function in Cc∞ (Ω), for Ω ⊂ Rn , cannot
be an analytic function.
Proof. If φ were analytic then, for all a ∈ Ω, we have the Taylor expansion
P∞ φ(k) (a)
φ(x) = k=0 k!
(x − a)k . Analyticity of φ implies that the radius of
convergence of the Taylor series at a is infinite. For φ ∈ Cc∞ (Ω), choose
a 6∈ supp(φ). The Taylor’s expansion around a imply that φ is zero function,
which is a contradiction.
Therefore, no non-zero analytic function can sit in Cc∞ (Ω) because zero
function is the only analytic function with compact support. The function
f (x) = eax is a smooth analytic function on R for all a ∈ R. Do we have
non-zero smooth (non-analytic) functions in Cc∞ (Ω)?
1.3.1 Smooth Non-Analytic (Bump) Functions

We are looking for a smooth non-analytic functions in Cc∞ (Ω). This can be
viewed as gluing of the zero function, along the boundary of a compact set,
with smooth function inside the compact set. The question is : can they be
glued such that the smoothness is preserved in all of domain.
Two continuous maps can be glued together to result in a continuous map.
In a different view point: consider a connected subset (interval) E = (a, b) of
R and the function f on E c which takes zero on the connected component,
x ≤ a and 1 on the connected component x ≥ b. This is a step function. A
continuous extension is trivial: Consider the linear map x 7→ x−ab−a
from [a, b]
to [0, 1] which glues. However, the function is not differentiable at a and b.
Similarly, gluing of the smooth maps x 7→ −x on (−∞, 0] and x 7→ x on [0, ∞)
yields the non-differentiable map |x| on R. Thus the smoothness property
need not be preserved when glued. A tool to accomplish the gluing of smooth
functions in a smooth way is the partition of unity (cf. Appendix ??).
Recall that we wish to define f on E such that the extended function is
smooth or C ∞ on R. We wish to glue a function in (a, b) such that on R it is
smooth. Note that the smoothness can break only at the end-points a and b.
Thus, if we can find a smooth function whose zero derivatives are at a point,
then we hope to transit smoothly from 0 to 1 in (a, b).
Recall that the positive function f (x) = e1/x behaves badly at x = 0.
From the right side it approaches +∞ and from left side it approaches zero.
However, for x 6= 0, e1/x is infinitely differentiable (smooth). In fact, the
k-th derivative of e1/x is Pk (1/x)e1/x . The proof is by induction. For k = 0,
P0 (t) ≡ 1, the constant function 1. Let us assume f (k) (x) = Pk (1/x)e1/x for
x 6= 0. Then,

(k+1) −1 1/x 1/x 0 −1
f (x) = Pk (1/x) 2
e + e Pk (1/x) = Pk+1 (1/x)e1/x ,
x x2
where Pk+1 (t) = −t2 (Pk (t) + Pk0 (t)).
Exercise 2. Show that the k-th derivative of e1/x for x 6= 0 is
" k
#
dk 1/x

1/x k
X k k−1 −k−i
(e ) = e (−1) (k − i)!x .
dxk i=1
i i−1
The coefficients of the polynomial appearing in the derivative are called the
Lah numbers, after Ivo Lah who encountered these numbers in actuarial
science.
Example 1.1. Consider the non-negative function f0 : (−∞, 0] → [0, 1] de-

fined as (
exp(1/x) if x < 0
f0 (x) =
0 if x = 0.
Example 1.2. A variant of the above example is
(
exp(−1/|x|) if x 6= 0
f1 (x) =
0 if x = 0.
Example 1.3. Another variant is the function f2 : R → R defined as

(
exp(−1/x) if x > 0
f2 (x) =
0 if x ≤ 0.
It is clear that 0 ≤ fi (x) < 1 and fi is infinitely differentiable for all x 6= 0

and i = 0, 1, 2. We need to check differentiability only at x = 0. We shall do
it only for f = f2 as the proof is similar for other cases. Of course, the left
side limit of f and its derivative is zero at x = 0. Thus, we consider
f (h) − f (0) exp(−1/h)

f 0 (0) = lim+ = lim+ .
h→0 h h→0 h
For all h > 0 and any fixed positive integer m > 0, we have
∞ i
1 1 X1 1
= hm m+1 ≤ hm (m + 1)! = hm (m + 1)!exp(1/h).
h h i=0
i! h
Thus,
exp(−1/h)
≤ hm (m + 1)!
h
and the larger the m, the sharper the estimate. Hence,
exp(−1/h)
f 0 (0) = lim+ = 0.
h→0 h
The same argument follows for any k + 1 derivative of f because
f (k) (h) − f (k) (0) Pk−1 (h)exp(−1/h)

f (k+1) (0) = lim+ = lim+ = 0.
h→0 h h→0 h2k+1
Therefore, f ∈ C ∞ (R). We shall now observe that f is not analytic at x = 0.

The Taylor series of f at x = 0,
∞
X f (k) (0)
xk = 0,
k=0
k!
converges to the zero function for all x ∈ R. But for x > 0, we know that
f (x) > 0 and hence do not converge to the Taylor series at x = 0. Thus, f
is not analytic.
Example 1.4. The function f : R → R defined as
(
exp(−1/x2 ) if x 6= 0
f3 (x) =
0 if x = 0
is infinitely differentiable but not analytic.

Example 1.5. The function f : R → R, referred to as the Cauchy’s exponen-
tial function, defined as
(
exp(−1/x2 ) if x > 0
f4 (x) =
0 if x ≤ 0
is infinitely differentiable but not analytic.

One can, in fact, construct bump functions where the function takes the
value 1 on a given subset of the support of f .
Lemma 1.3.1 (Smooth Cut-off Function). For any interval (a, b) of R, there
is a decreasing smooth function g : R → R such that 0 < g(x) < 1 for all
x ∈ (a, b) and (
1 x≤a
g(x) =
0 x ≥ b.
Proof. Choose f = fi , for i = 2 or i = 4 in the above examples. This

is precisely the choice of function whose zero derivatives are all at a point.
Note that the function f (b−x) = 0 for all x ≥ b and the function f (x−a) = 0
for all x ≤ a. Consider the non-zero positive function
h(x) = f (b − x) + f (x − a)
on R which is (
f (b − x) x ≤ a
h(x) =
f (x − a) x ≥ b.
Since f (b − x) < f (b − x) + f (x − a), addition of positive function. Moreover,
h is smooth. Define the function
f (b − x)
g(x) =
f (b − x) + f (x − a)
and it satisfies all the desired properties.
In particular, when f = f2 , we have
e−1/(b−x)
g(x) = .
e−1/(b−x) + e−1/(x−a)
Note that the increasing function
f (x − a)
1 − g(x) =
f (b − x) + f (x − a)
is identically zero for x ≤ a and one on x ≥ b satisfying the other properties
of above lemma.
Lemma 1.3.2 (Smooth Bump Function). For any positive a, b ∈ R such
that a < b, there is a h ∈ Cc∞ (R) such that 0 ≤ h(x) ≤ 1, for all x ∈ R,
h ≡ 1 on [−a, a] and supp(h)= [−b, b].
Proof. Recall the function g obtained in above lemma. Then the function
x 7→ g(|x|) is identically one in [−a, a] and zero outside (−b, b). The desired
function h is obtained by setting h(x) = g(|x|). The function h is smooth
because it is composed with the function |x|, which is smooth except at x = 0.
But h ≡ 1 around the origin.
Lemma 1.3.3 (Smooth Bump Function on Rn ). For any 0 < a < b, there
is a h ∈ Cc∞ (Rn ) with range in [0, 1] such that 0 ≤ h(x) ≤ 1, for all x ∈ R,
h ≡ 1 on the closed ball Ba (0) and supp(h) is the closure of Bb (0).
Proof. The same function h(x) = g(|x|) works.
Theorem 1.3.4. For any compact subset K of Rn , then there exists a φ ∈
Cc∞ (Rn ) such that φ ≡ 1 on K.
Proof. Since K is compact, there exists a a > 0 such that Ba (0) ⊃ K and
we apply previous Lemma. Alternately, an existential proof is as follows:
One can chose an open set U ⊃ K such that U is compact. Then, U and
K c form an open cover of Rn . By Theorem ??, there is a partition of unity
subordinate to the given cover of Rn . Thus, there are smooth non-negative
functions φ and ψ on Rn such that supp(φ) is in U , supp(ψ) is in K c and
φ + ψ = 1 on Rn . On K, ψ = 0 and hence φ = 1. Also, supp(φ) is a
closed subset of a compact subset U , hence φ has compact set. Thus, φ is
the desired function.
1.3.2 Mollifiers
The functions f2 and f4 in the above examples can be tweaked to construct
a function in Cc∞ (R) such that supp(f )= [−b, b] for any positive b ∈ R. Note
that this time we do not demand that the function takes constant value 1 in
a subset of the support. Consider the transformation g2b (x) = f2 (1 − |x|/b).
Thus,
(
−b
exp b−|x| if |x| < b
g2b (x) =
0 if |x| ≥ b.
Also, using the function

(
−1

exp 1−x2
if |x| < 1
f5 (x) =
0 if |x| ≥ 1
one defines the function

(
−b2
exp b2 −|x|2 if |x| < b
g5b (x) =
0 if |x| ≥ b.
In fact, the functions g2b , g5b can be extended to n dimensions and are in
Cc∞ (Rn ) with support in B(0; b), the disk with centre at origin and radius b.
We shall now introduce an important sequence of functions in Cc∞ (Rn ),
called mollifiers. Recall the Cc∞ (Rn ) functions g2b and g5b introduced in the
previous section. For ε > 0, we set b = ε in g5b (or g2b ) and
−ε2
Z Z
ε
g5 (x) dx = exp 2 dx
Rn |x|<ε ε − |x|2

−1
Z
n
= ε exp dy (by setting y = x/ε)
|y|<1 1 − |y|2
= εn c−1 ,
where
−1
Z
−1
c = exp dy.
|y|≤1 1 − |y|2
Now, set ρε (x) = cε−n g5ε (x), equivalently,
( 2

cε−n exp ε2−ε
−|x|2
if |x| < ε
ρε (x) = (1.3.3)
0 if |x| ≥ ε.
Note that ρε ≥ 0 and is in Cc∞ (Rn ) with support in B(0; ε). The sequence
{ρε } is an example of mollifiers, a particular case of the Dirac Sequence.
Definition 1.3.5. A sequence of functions {ρk }, say on Rn , is said to be a

Dirac Sequence if
(i) ρk ≥ 0 for all k.

R
(ii) Rn ρk (x) dx = 1 for all k.
(iii) For every given r > 0 and ε > 0, there exists a N0 ∈ N such that
Z
ρk (x) dx < ε, ∀k > N0 .
Rn \B(0;r)
The connection between the sequence of mollifiers and Dirac delta func-
tion will become evident in the sequel. The notion of mollifiers is also an
example for the approximation of identity concept in functional analysis and
ring theory.
Definition 1.3.6. An approximate identity is a sequence (or net) {ρk in a

Banach algebra or ring (possible with no identity), (X, ?) such that for any
element a in the algebra or ring, the limit of a ? ρk (or ρk ? a) is a.
1.3.3 Topology on C(Ω), C0 (Ω) and Cc (Ω)

Recall that (cf. § ??) for an open subset Ω ⊂ Rn and for any φ ∈ Cb (Ω),
kφk = supx∈Ω |φ(x)| defines a norm, called the uniform norm or sup-norm, in
Cb (Ω). The space Cb (Ω) endowed with the uniform norm is a Banach space
and a sequence φk converges to φ in the uniform norm is said to converge
uniformly in Ω.
If Ω is a compact subset, the norm on the space of continuous functions
on Ω, C(Ω), is defined as
kφk∞,Ω := max |φ(x)|

x∈Ω
is complete. Thus C(Ω) is a Banach space w.r.t the uniform norm. For a non-
compact subset Ω of Rn , the space C(Ω) is not normable but are metrizable1 .
They form a locally convex complete metric space called Fréchet space. We
describe this metric briefly in the following paragraph.
For any open subset Ω of Rn , there is a sequence Kj of non-empty compact
subsets of Ω such that Ω = ∪∞ j=0 Kj and Kj ⊂ Int(Kj+1 ), for all j (exhaustion
of an open set by compact sets, cf. Lemma ??). This property is called the
σ-compactness of Ω. We define a countable family of semi-norms on C(Ω) as
pj (φ) = kφk∞,Kj .
Note that p0 ≤ p1 ≤ p2 ≤ . . . . The sets
{φ ∈ C(Ω) | pj (φ) < 1/j}
form a local base for C(Ω). The metric induced by the family of semi-norms
on C(Ω) is
1 pj (φ − ψ)
d(φ, ψ) = max j
j∈N∪{0} 2 1 + pj (φ − ψ)
and the metric is complete and C(Ω) is a Fréchet space. This is precisely the
topology of compact convergence (uniform convergence on compact sets) or
the compact-open topology in this case. If {φm } is a Cauchy sequence w.r.t
d then pj (φm − φ` ) → 0, for all j and as m, ` tends to infinity. Thus {φm }
converges uniformly on Kj to some φ ∈ C(Ω). Then it is easy to see that
1
Any norm space is a metric space but the converse is not true always because d(x, 0)
may fail to satisfy the properties of norm. A metric d induces a norm if d(x + z, y + z) =
d(x, y) and d(αx, αy) = |α|d(x, y).
d(φ, φm ) → 0. The metric defined above is called the Fréchet metric and is
equivalent to the metric
∞
X 1 pj (φ − ψ)
d(φ, ψ) = j 1 + p (φ − ψ)
.
j=0
2 j
The Fréchet space may be seen as a countable limit of Banach spaces. In our
case, the Banach spaces are C(Ki ) w.r.t the uniform norm, the restriction of
the semi-norm.
Exercise 3. Show that the topology given in C(Ω) is independent of the
choice the exhaustion compact sets {Kj } of Ω.
We say a function φ : Ω → R vanishes at infinity (or on the boundary),
if for every ε > 0 there exists a compact set K ⊂ Ω (depending on ε) such
that |φ(x)| < ε for all x ∈ Ω \ K. Let C0 (Ω) be the set of all continuous
functions on Ω vanishing at infinity (or on the boundary). Let Cc (Ω) be the
set of all continuous functions with compact support in Ω. Let Cb (Ω) be
the space of all bounded continuous functions on Ω. We have the inclusion
Cc (Ω) ⊂ C0 (Ω) ⊂ Cb (Ω) ⊂ C(Ω). One may assign the uniform norm on
Cb (Ω), C0 (Ω) and Cc (Ω). Under the uniform norm Cc (Ω) is dense in C0 (Ω)
which is a closed subspace of the Banach space Cb (Ω).
We shall construct a complete (non-metrizable) topology on Cc (Ω). For
every compact subset K ⊂ Ω, let CK (Ω) denote the class of all continuous
functions in Ω such that their support is in K. The space CK (Ω) is a Banach
space w.r.t the uniform norm. Note that
Cc (Ω) = ∪K⊂Ω CK (Ω)
where the union is over all compact subsets of Ω. We declare a map T on
Cc (Ω) is continuous if T restricted to CK (Ω), for each compact K ⊂ Ω, is
continuous. Such a topology is called the inductive limit topology of Cc (Ω)
with uniform norm. We say a sequence {φm } converges to φ w.r.t inductive
limit topology if there exists a compact set K such that supp(φm ) ⊂ K and
φm converges uniformly to φ. The space Cc (Ω) is complete with respect to
the inductive limit topology.
1.3.4 Algebraic and Topological Dual of Cc (Ω)

Definition 1.3.7. A linear map T : Cc (Ω) → R is said to be positive if
T (φ) ≥ 0, for all φ ≥ 0.
Note that given a measure µ, the map

Z
T (φ) := φ dµ (1.3.4)
Ω
defines a positive, linear functional on Cc (Ω).

Theorem 1.3.8. Let µ be a positive Radon measure on Ω and T be as defined
in (1.3.4). Let, for every open subset U ⊂ Ω, SU := {φ ∈ D(U ) | 0 ≤ φ ≤ 1}
and, for every compact subset K ⊂ Ω, SK := {φ ∈ Cc (Ω) | φ ≥ χK }. Then
µ(U ) = sup T (φ) ∀U ⊂ Ω (1.3.5)
φ∈SU
and
µ(K) = inf T (φ) ∀K ⊂ Ω. (1.3.6)
φ∈SK
Theorem 1.3.9. Let T be a positive, linear functional on Cc (Ω). Then there

exists a unique positive Radon measure defined as in (1.3.5) or (1.3.6) and
satisfies (1.3.4).
Definition 1.3.10. A linear map T : Cc (Ω) → R is locally bounded if, for
every compact subset K ⊂ Ω, there exists a positive constant CK > 0 such
that
|T (φ)| ≤ CK kφk∞ ∀φ ∈ D(K).
Note that every positive linear functional on Cc (Ω) is locally bounded.
Theorem 1.3.11. Let T be a locally bounded, linear functional on Cc (Ω).
Then there exists two positive, linear functionals T + and T − on Cc (Ω) such
that T = T + − T − .
Theorem 1.3.12. Let T be a locally bounded, linear functional on Cc (Ω).
Then
R thereR exists two Radon measures µ1 and µ2 on Ω such that T (φ) =
Ω
φ dµ1 − Ω φ dµ2 , for all φ ∈ Cc (Ω).
Note that any linear continuous (bounded) functional Cc (Ω) is also locally
bounded and, hence, the above result is true.
For a locally compact Hausdroff space Ω, recall (cf. § 1.3.3) that the
spaces Cc (Ω) ⊂ C0 (Ω) ⊂ Cb (Ω) can be endowed with the uniform norm
where the first inclusion is dense and C0 (Ω) is closed subspace of Cb (Ω).
Further, one may also endow the inductive limit topology on Cc (Ω) which
completes it.
Theorem 1.3.13. Consider Cb (Ω) endowed with the uniform topology. Then
there is an isometric isomorphism between the dual of Cb (Ω) and the space
of bounded finitely additive (not necessarily countably additive) measures,
i.e. for any continuous linear functional T : Cb (Ω) → R there is a unique
bounded, finitely additive measure µ such that
Z
T (φ) = φ(x) dµ ∀φ ∈ Cb (Ω).
Ω
This association T 7→ µ defines an isometry, i.e., kT k = |µ|.
Note that the class of bounded countably additive measures are a sub-
space of the class of bounded finitely additive measures.
Theorem 1.3.14 (Riesz-Alexandrov). Consider C0 (Ω) endowed with the

uniform topology. Then there is an isometric isomorphism between the dual of
C0 (Ω) and Mb (Ω), i.e. for any continuous linear functional T : C0 (Ω) → R
there is a unique finite regular Borel measure µ ∈ Mb (Ω) such that
Z
T (φ) = φ(x) dµ ∀φ ∈ C0 (Ω).
Ω
Note that due to the closed inclusion of C0 (Ω) in Cb (Ω) the dual space
inclusion are not reversed. In fact they preserve the inclusion.
Theorem 1.3.15 (Riesz-Markov). Consider Cc (Ω) endowed with the induc-

tive limit topology. Then there is an isometric isomorphism between the dual
of Cc (Ω) and R(Ω), i.e. for any continuous linear functional T : Cc (Ω) → R
there is a unique Radon measure µ ∈ R(Ω) such that
Z
T (φ) = φ(x) dµ ∀φ ∈ Cc (Ω).
Ω
1.3.5 Topology on C ∞ (Ω)

Recall that our aim is to define a suitable topology in Cc∞ (Ω) such that the
operations described in (1.3.2) are continuous. For a compact Ω, the space
of k-times differentiable functions on Ω, C k (Ω), is a Banach space w.r.t the

norm
Xk
kφkk,Ω := kDα φk∞,Ω , ∀k ≥ 1.
|α|=0
For any open subset Ω of Rn , there is a sequence Kj of non-empty compact

subsets of Ω such that Ω = ∪∞j=0 Kj and Kj ⊂ Int(Kj+1 ), for all j (exhaustion
of an open set by compact sets, cf. Lemma ??). Similar to the Frëchet metric
constructed on the space C(Ω), we shall construct one for C ∞ (Ω). We define
a countable family of semi-norms on C ∞ (Ω) as
j
X
pj (φ) = kDα φk∞,Kj = kφkj,Kj .
|α|=0
Again, as before, the family of semi-norms induces a locally convex complete

metric on C ∞ (Ω) making it a Fréchet space.
Exercise 4. Show that the topology given in C ∞ (Ω) is independent of the
choice the exhaustion compact sets {Kj } of Ω.
The space Cc∞ (Ω) is a subset of C ∞ (Ω) and the semi-norms defined in
C ∞ (Ω) restricted to Cc∞ (Ω) becomes a norm. For any φ ∈ Cc∞ (Ω), the family
of norms ( j = 0, 1, 2, . . .)
j
X
kφkj = kDα φk∞,Ω = kφkj,Ω
|α|=0
induces the same topology as the one inherited from C ∞ (Ω). However, this
norm induced topology on Cc∞ (Ω) is not complete and its completion is
C ∞ (Ω). For more details and proof of the semi-norm induced locally convex
topology refer to [Rud91].
Exercise 5. Let φ ∈ Cc∞ (R) with supp(φ)= [0, 1] and φ > 0 in (0, 1). Then
the sequence
m
X 1 1 1
ψm (x) = φ(x − i) = φ(x − 1) + φ(x − 2) + . . . + φ(x − m)
i=1
i 2 m
is Cauchy in the topology induced by the norms, but lim ψm 6∈ Cc∞ (R).
1.3.6 Inductive Limit Topology on Cc∞ (Ω)

We shall construct a complete (non-metrizable) topology on Cc∞ (Ω) different
from the one inherited from C ∞ (Ω). For every compact subset K ⊂ Ω, let
∞
CK (Ω) denote the class of all functions in C ∞ (Ω) such that their support
∞
is in K. The space CK (Ω) is given the topology inherited from C ∞ (Ω), the
same induced by the family of norms defined at the end of previous section,
j
X
kφkj,K = kDα φk∞,K ∀j ≥ 0.
|α|=0
∞
Exercise 6. CK (Ω) is a closed subspace of C ∞ (Ω) under the inherited topol-
∞
ogy of C (Ω).
Proof. For each x ∈ Ω, define the functional Tx : C ∞ (Ω) → R defined as
Tx (φ) = φ(x). For each x ∈ Ω, there is a j0 such that x ∈ Kj for all j ≥ j0 .
Then,
|Tx (φ)| = |φ(x)| ≤ pj (φ) ∀j ≥ j0 .
The functional Tx is continuous2 because uniform convergence implies point-
wise convergence. The topology on C ∞ (Ω) is uniform convergence on com-
pact subsets. Therefore the kernel of Tx ,
ker(Tx ) := {φ ∈ C ∞ (Ω) | Tx (φ) = 0},
is a closed subspace of C ∞ (Ω). Note that ker(Tx ) is precisely those φ ∈

C ∞ (Ω) such that φ(x) = 0. We claim that
∞
CK (Ω) = ∩x∈K c ker(Tx ).
∞
If φ ∈ CK (Ω) then φ is in the intersection because φ(x) = 0 for all x ∈
K . Conversely, If φ(x) = 0 for all x ∈ K c , then supp(φ)⊆ K. Thus, for
c
∞
any compact subset K of Ω, we have our claim that CK (Ω) is an arbitrary
∞ ∞
intersection of closed sets. Thus, CK (Ω) is closed in C (Ω).
Recall that Kj is a sequence of non-empty compact subsets of Ω such that
Ω = ∪∞
j=1 Kj and Kj ⊂ Int(Kj+1 ), for all j. Note that
Cc∞ (Ω) = ∪∞ ∞
j=1 CKj (Ω)
2
Compare the functional Tx with Dirac distribution to be introduced later
∞ ∞
and CK `
(Ω) ⊂ CK m
(Ω) for all ` < m. With the given topology on these
spaces, inherited from C ∞ (Ω), the inclusion map Ik` : CK
∞
`
∞
(Ω) → CKm
(Ω) is
∞
continuous. This is because the local base in CKm (Ω) is

∞ 1
φ ∈ CKm (Ω) | kφkm < .
m
For any such φ in the local base, we have kφk` < 1/` and is in the local base
∞
of CK `
(Ω). Thus, we endow Cc∞ with the finest3 topology that makes the
∞
inclusion maps Ij : CK j
(Ω) → Cc∞ (Ω) continuous, for all j. In other words, a
set U in Cc∞ (Ω) is said to be open if and only if Ij−1 (U ) is open in CK
∞
j
(Ω) for
all j ≥ 1. Such a topology is called the inductive limit topology with respect
∞
to CK j
(Ω) and the maps Ik` . The space Cc∞ (Ω) is complete with respect
to the inductive limit topology because any Cauchy sequence is Cauchy in
∞ ∞
CK j
(Ω), for some j. Since CK j
(Ω) is closed, the space Cc∞ (Ω) is complete
∞
w.r.t the inductive limit topology. Though each CK (Ω) is metrizable, the
∞
space Cc (Ω) is not metrizable (cf. Exercise 8).
Exercise 7. Every proper subspace of a topological vector space has empty
interior.
Proof. Let X be a topological vector space and V ( X be a vector subspace
of X. We need to show that for x ∈ V there is an open set U containing
x such that U ⊂ V . It is enough to show the claim for x = 0 (0 ∈ V )
because if U contains 0, then U + {x} ⊂ V is an open set (due to continuity
of addition) containing x. Due to vector space structure U + {x} is in V .
For every x ∈ X, we can define a function fx : R → X as fx (λ) = λx. The
function fx is continuous because the scalar multiplication map from R × X
to X is continuous. Suppose that there is an open set V containing 0 such
that V ⊂ X. Then fx−1 (V ) will be an open set containing 0 ∈ R. Thus, for
any λ ∈ fx−1 (V ), we have λx ∈ V . Thus, x ∈ V . This argument is true for
all x ∈ X, thus X ⊂ V . This implies X = V , a contradiction.
Exercise 8. The inductive limit topology on Cc∞ (Ω) is not metrizable.
Proof. Recall that Cc∞ (Ω) = ∪∞ ∞ ∞
j=1 CKj (Ω), where each closed set CKj (Ω) has
empty interior (cf. Exercise 7). Therefore, the complete space Cc∞ (Ω) is a
countable union of no-where dense sets. If Cc∞ (Ω) was metrizable then it
would contradict the Baire’s category theorem.
3
strongest or largest topology, the one with more open sets
Definition 1.3.16. The space Cc∞ (Ω) endowed with the inductive limit topol-
ogy, and denoted as D(Ω), is called the space of test functions.
Exercise 9 (cf. Exercise 4). Show that the topology defined on D(Ω) is
independent of the choice of (exhaustion sets) the sequence of compact sets
Kj of Ω.
The sequential characterisation of the inductive limit topology on D(Ω)
is given below. We define only for the zero converging sequence due to
continuity of the addition operation.
Proposition 1.3.17. A sequence of functions {φm } ⊂ D(Ω) converges to

zero iff there exists a compact set K ⊂ Ω such that supp(φm ) ⊂ K, for all
m, and φm and all its derivatives converge uniformly to zero on K.
1.3.7 Regularization and Cut-off Technique

In this section, we show that that space of test functions is densely contained
in Lp space. This observation is very useful since most conjectures can be
checked for test functions and then carried on to the required function in Lp .
As usual, let Ω be an open subset of Rn . We already know the following
results from the theory of Lebesgue measure on Rn .
Theorem 1.3.18 (cf. Theorem ??). For 1 ≤ p < ∞, the class of all simple4
functions S(Ω) is dense in Lp (Ω).
Theorem 1.3.19 (cf. Theorem ??). For 1 ≤ p < ∞, Cc (Ω) is dense in

Lp (Ω).
Further, a step ahead, we show that the space of test functions, Cc∞ (Ω) is
densely contained in Lp (Ω). This result is established by using the technique
of regularization by convolution introduced by Leray and Friedrichs.
4
By our definition, simple function is non-zero on a finite measure. A simple function
φ is a non-zero function on Rn having the (canonical) form
k
X
φ(x) = ai 1Ei
i=1
with disjoint measurable subsets Ei ⊂ Rn with µ(Ei ) < +∞ and ai 6= 0, for all i, and
ai 6= aj for i 6= j.
Definition 1.3.20. Let f, g ∈ L1 (Rn ). The convolution f ∗ g is defined as,

Z
(f ∗ g)(x) = f (x − y)g(y) dy ∀x ∈ Rn .
Rn
The integral on RHS is well-defined, since by Fubini’s Theorem and the

translation invariance of the Lebesgue measure, we have
Z Z Z
|f (x − y)g(y)| dx dy = |g(y)| dy |f (x − y)| dx = kgk1 kf k1 .
Rn ×Rn Rn Rn
Thus, for a fixed x, f (x − y)g(y) ∈ L1 (Rn ).
Theorem 1.3.21. The convolution operation on L1 (Rn ) is both commutative

and associative.
Theorem 1.3.22 (Young’s inequality). Let 1 ≤ p, q, r < ∞ such that (1/p)+

(1/q) = 1 + (1/r). If f ∈ Lp (Rn ) and g ∈ Lq (Rn ), then the convolution
f ∗ g ∈ Lr (Rn ) and
kf ∗ gkr ≤ kf kp kgkq .
In particular, for 1 ≤ p < ∞, if f ∈ L1 (Rn ) and g ∈ Lp (Rn ), then the
convolution f ∗ g ∈ Lp (Rn ) and
kf ∗ gkp ≤ kf k1 kgkp .
5
Proposition 1.3.23. Let f ∈ L1 (Rn ) and g ∈ Lp (Rn ), for 1 ≤ p ≤ ∞.
Then
supp(f ∗ g) ⊂ supp(f ) + supp(g)
If both f and g have compact support, then support of f ∗ g is also

compact. The convolution operation preserves smoothness.
Theorem 1.3.24. Let Ω ⊂ Rn be an open subset of Rn and let
Ωε := {x ∈ Ω | dist(x, ∂Ω) > ε}.
If f ∈ L1loc (Ω) then fε := ρε ∗ f is in C ∞ (Ωε ).

5
Refer Brezis for proof
Proof. Fix x ∈ Ωε . Consider

fε (x + hei ) − fε (x)
Z
1
= [ρε (x + hei − y) − ρε (x − y)] f (y) dy
h h Ω
Z
1
= [ρε (x + hei − y) − ρε (x − y)]f (y) dy.
Bε (x) h
Now, taking limh→0 both sides, we get
Z
∂fε (x) 1
= lim [ρε (x + hei − y) − ρε (x − y)]f (y) dy
∂xi h→0 B (x) h
ε
∂ρε (x − y)
Z
= f (y) dy
Bε (x) ∂xi
(interchange of limits is due to the uniform convergence)
∂ρε (x − y)
Z
∂ρε
= f (y) dy = ∗ f.
Ω ∂xi ∂xi
Similarly, one can show that, for any tuple α, Dα fε (x) = (Dα ρε ∗ f )(x).
Thus, uε ∈ C ∞ (Ωε ).
Proposition 1.3.25. 6 Let f ∈ Cck (Rn ) (k ≥ 1) and let g ∈ L1loc (Rn ). Then
f ∗ g ∈ C k (Rn ) and for all |α| ≤ k
Dα (f ∗ g) = Dα f ∗ g = f ∗ Dα g.
Theorem 1.3.26 (Regularization technique). C ∞ (Rn ) is dense in C(Rn )
under the uniform convergence on compact sets topology.
Proof. Let g ∈ C(Rn ) and K ⊂ Rn be a compact subset. Note that g is
uniformly continuous on K. Hence, for every η > 0, there exist a δ > 0
(independent of x and dependent on K and η) such that |g(x − y) − g(x)| <
η whenever |y| < δ for all x ∈ K. For each m ∈ N, set ρm := ρ1/m ,
the sequence of mollifiers. Define gm := ρm ∗ g. Note that gm ∈ C ∞ (Rn )
(Dα gm = Dα ρm ∗ g). Now, for all x ∈ Rn ,
Z Z
|gm (x) − g(x)| = g(x − y)ρm (y) dy − g(x) ρm (y) dy
|y|≤1/m |y|≤1/m
Z
≤ |g(x − y) − g(x)|ρm (y) dy
|y|≤1/m
6
Refer Brezis for proof
Hence, for all x ∈ K and m > 1/δ, we have

Z
|gm (x) − g(x)| ≤ |g(x − y) − g(x)|ρm (y) dy
|y|<δ
Z
≤ η ρm (y) dy = η
|y|<δ
Since the δ is independent of x ∈ K, we have kgm − gk∞ < η for all m > 1/δ.
Hence, gm → g uniformly on K.
Theorem 1.3.27. For any Ω ⊆ Rn , Cc∞ (Ω) is dense in Cc (Ω) under the
uniform topology.
Proof. Let g ∈ Cc (Ω) and K := supp(g). One can view Cc (Ω) as a subset of
Cc (Rn ) under the following identification: Each g ∈ Cc (Ω) is extended to Rn
as g̃ (
g(x) x ∈ K
g̃(x) =
0 x ∈ Rn \ K.
By Theorem 1.3.26, the sequence gm := ρm ∗ g̃ in C ∞ (Rn ) converges to g̃

uniformly on every compact subsets of Rn . Note that supp(gm ) ⊂ K +
B(0; 1/m) is compact because K is compact. Since we want gm ∈ Cc∞ (Ω),
we choose m0 ∈ N such that 1/m0 < dist(K, Ωc ). Thus, supp(gm ) ⊂ Ω and
gm ∈ Cc∞ (Ω), for all m ≥ m0 . The proof of the uniform convergence of gm
to g on Ω is same as in Theorem 1.3.26.
Corollary 1.3.28. For any Ω ⊆ Rn , Cc∞ (Ω) is dense in C(Ω) under the
uniform convergence on compact sets topology.
Theorem 1.3.29 (Regularization technique). The space C ∞ (Rn ) is dense

in Lp (Rn ), for 1 ≤ p < ∞, under the p-norm.
Proof. Let f ∈ Lp (Rn ). For each m ∈ N, set ρm := ρ1/m , the sequence of

mollifiers. Then the sequence fm := ρm ∗ f is in C ∞ (Rn ). Since ρm ∈ L1 (Rn ),
by Young’s inequality, fm ∈ Lp (Rn ). We shall prove that fm converges to f
in p-norm. For any given ε > 0, by Theorem ??, we choose a g ∈ Cc (Rn ) such
that kg −f kp < ε/3. Therefore, by Theorem 1.3.27, there is a compact subset
K ⊂ Rn such that kρm ∗g −gk∞ < ε/3(µ(K))1/p . Hence, kρm ∗g −gkp < ε/3.
Thus, for sufficiently large m, we have
kfm − f kp ≤ kρm ∗ f − ρm ∗ gkp + kρm ∗ g − gkp + kg − f kp

2ε 2ε
< kρm ∗ (f − g)kp + ≤ kf − gkp kρm k1 +
3 3
ε 2ε
< + = ε.
3 3
The first term has been handled using Young’s inequality.
Theorem 1.3.30 (Cut-Off Technique). For 1 ≤ p < ∞ and Ω ⊆ Rn , Cc∞ (Ω)
is dense in Lp (Ω).
Proof. Any f ∈ Lp (Ω) can be viewed as an element in Lp (Rn ) under the
extension (
f (x) x ∈ Ω
f˜(x) =
0 x ∈ Ωc .
By Theorem 1.3.29, the sequence fm := ρm ∗ f˜ converges to f˜ in p-norm. The
sequence {fm } may fail to have compact support in Ω because support of f˜ is
not necessarily compact in Ω. To fix this issue, we shall multiply the sequence
with suitable choice of test functions in Cc∞ (Ω). Choose the sequence of
exhaustion compact sets {Km } in Ω (cf. Lemma ??). In particular, for
Ω = Rn , we can choose Km = B(0; m). Note that Ω = ∪m Km . Consider7
{φm } ⊂ Cc∞ (Ω) such that φm ≡ 1 on Km and 0 ≤ φm ≤ 1, for all m. We
extend φm by zero on Ωc . Define Fm := φm fm and, hence, Fm ∈ Cc∞ (Ω).
Also, Fm = fm on Km and |Fm | ≤ |fm | in Rn . Thus,
kFm − f kp,Ω = kFm − f˜kp,Rn ≤ kφm fm − φm f˜kp,Rn + kφm f˜ − f˜kp,Rn

≤ kfm − f˜kp,Rn + kφm f˜ − f˜kp,Rn .
The first term converges to zero by Theorem 1.3.29 and the second term
converges to zero by Dominated convergence theorem.
The case p = ∞ is ignored in the above results, because the L∞ -limit of
ρm ∗ f is continuous and we do have discontinuous functions in L∞ (Ω).
Theorem 1.3.31 (Kolmogorov Compactness Criteria). Let p ∈ [1, ∞) and
let A be a subset of Lp (Rn ). Then A is relatively compact in Lp (Rn ) iff the
following conditions are satisfied:
7
The type of functions, φk , are called cut-off functions
(i) A is bounded in Lp (Rn );

R
(ii) limr→+∞ {|x|>r} |f (x)|p dx = 0 uniformly with respect to f ∈ A;
(iii) limh→0 kτh f − f kp = 0 uniformly with respect to f ∈ A, where τh f is

the translated function (τh f )(x) := f (x − h).
Proof. We shall prove the ‘only if’ part, i.e, (i), (ii), (iii) implies that A
is relatively compact in Lp (Rn ). Equivalently, we have to prove that A is
precompact, which means that for any ε > 0, there exists a finite number of
balls Bε (f1 ), . . . , Bε (fk ) which cover A. Let us choose ε > 0. By (ii) there
exists a r > 0 such that
Z
|f (x)|p dx < ε ∀f ∈ A.
|x|>r
Let (ρn )n∈N be a mollifier. It follows from Theorem 1.3.30 that, for all n ≥ 1
and f ∈ Lp (Rn )
Z
p
kf − f ∗ ρn kp ≤ ρn (y)kf − τy f kpp dy.
Rn
Hence
kf − f ∗ ρn kp ≤ sup kf − τy f kp .
1
|y|≤ n
By (iii), there exists an integer N (ε) ∈ N such that, for all f ∈ A,
kf − f ∗ ρN (ε) kp < ε.
On the other hand, for any x, z ∈ Rn , f ∈ Lp (Rn ) and n ∈ N,

Z
|(f ∗ ρn )(x) − (f ∗ ρn )(z)| ≤ |f (x − y) − f (z − y)|ρn (y) dy
Rn
≤ kτx fˇ − τz fˇkp kρn kq
≤ kτx−z f − f kp kρn kq .
The last inequality follows from the invariance property of the Lebesgue
measure. Moreover,
|(f ∗ ρn )(x)| ≤ kf kp kρn kq .
Let us consider the family A = {f ∗ ρN (ε) : Br (0) → R | f ∈ A}. By
using (i) and (iii), and Ascoli-Arzela result, we observe that A is relatively
compact w.r.t the uniform topology on Br (0). Hence, there exists a finite set
{f1 , . . . , fk } ⊂ A such that
A ⊂ ∪ki=1 Bεr−n/p (fi ∗ ρN (ε) ).
Thus, for all f ∈ A, there exists some j ∈ {1, 2, . . . , k} such that, for all
x ∈ Br (0)
|f ∗ ρN (ε) (x) − fj ∗ ρN (ε) (x)| ≤ ε|Br (0)|−1/p .
Hence,
Z 1/p Z 1/p
p p
kf − fj kp ≤ |f | dx + |fj | dx
|x|>r |x|>r
+kf − f ∗ ρN (ε) kp + kfj − fj ∗ ρN (ε) kp
+kf ∗ ρN (ε) − fj ∗ ρN (ε) kp,Br (0) .
The last term may be treated as follows:

Z p1
k(f − fj ) ∗ ρN (ε) kp,Br (0) = |f ∗ ρN (ε) (x) − fj ∗ ρN (ε) (x)|p dx
Br (0)
≤ ε|Br (0)|−1/p |Br (0)|1/p = ε.
Finally,
kf − fj kp ≤ 5ε
and, hence, A is precompact in Lp (Rn ).
1.4 Space of Distributions

Definition 1.4.1. A linear functional T on D(Ω) is said to be continuous if
inverse image of open sets of R are open in D(Ω). A linear functional T on
D(Ω) is said to be sequential continuous if T φm → 0 in R whenever φm → 0
in D(Ω).
∞
Exercise 10. For any compact set K ⊂ Ω, the restriction to CK (Ω) of any
∞ ∞
continuous map on Cc (Ω) is also continuous on CK (Ω). Similarly, for any
∞
compact set K ⊂ Ω, the restriction to CK (Ω) of any sequentially continuous
∞ ∞
map on Cc (Ω) is also sequentially continuous on CK (Ω).
Hint. For each compact set K there is an positive integer i0 such that K ⊂
Ki , for all i ≥ i0 .
It is enough to define for zero convergent sequences because addition op-
eration is continuous. For a first countable space the notion of continuity and
sequential continuity are equivalent. A Hausdorff topological vector space is
metrizable iff it is first countable. We know that D(Ω) is not metrizable
(cf. Exercise 8) and hence cannot be first countable.
Theorem 1.4.2. Let T : D(Ω) → R be a linear map. Then the following are
equivalent:
(i) T is continuous, i.e., inverse image of open set in R, under T , is open

in D(Ω).
(ii) For every compact subset K ⊂ Ω, there exists a constant CK > 0 and
an integer NK ≥ 0 (both depending on K) such that
∞
|T (φ)| ≤ CK kφkNK , ∀φ ∈ CK (Ω).
(iii) T is sequentially continuous.
Proof. ((i) =⇒ (ii)): Let T be continuous on D(Ω). Then, for any compact
∞
subset K ⊂ Ω, the restriction of T to CK (Ω) is continuous (cf. Exercise 10).
∞
The inverse image of (−c, c) under T is open set in CK (Ω) containing origin.
∞
Since CK (Ω) is first countable (normed space), there is a local base at 0.
∞
Thus, there is a NK and for all φ ∈ CK (Ω) such that kφkNK ≤ 1/NK , we
∞
have |T (φ)| < c. Thus, for φ ∈ CK (Ω),

φ
T <c
NK kφkNK
and hence |T (φ)| < NK ckφkNK .

((ii) =⇒ (iii)): Let φm → 0 in D(Ω). By definition of the test function
convergence, there is a compact set K such that kφm kj → 0 for all j. Using
the fact that |T (φ)| ≤ CK kφkNK , we get |T (φm )| → 0. Thus T (φm ) → 0.
((iii) =⇒ (i)): Let T be sequentially continuous. Then, by Exercise 10,
∞
the restriction of T on CK (Ω), for every compact subset K ⊂ Ω, is also
∞ ∞
sequentially continuous. But CK (Ω) is metrizable and hence T on CK (Ω) is
continuous, for all compact K of Ω. Thus, T is continuous on D(Ω).
Definition 1.4.3. A linear functional T on D(Ω) is said to be a distribution

on Ω, if for every compact subset K ⊂ Ω, there exists a constant CK > 0
and an integer NK ≥ 0 (both depending on K) such that
∞
|T (φ)| ≤ CK kφkNK , ∀φ ∈ CK (Ω).
By Theorem 1.4.2, the above definition is saying that any continuous
linear functional on D(Ω) is a distribution. The space of all distributions in
Ω is denoted by D0 (Ω).
Definition 1.4.4. If the NK is independent of K, i.e., the same N is enough
for all compact sets K, then the smallest such N is called the order of T . If
there exist no such N , we say T is of infinite order.
Exercise 11. Which of the following are distributions? If your answer is
affirmation, give the order of the distribution. If your answer is in negation,
give reasons. For φ ∈ D(R), T (φ) is defined as:
(i) φ0 (1) − φ00 (−2).
P∞ (k)
(ii) k=0 φ (π).
P∞ (k)
(iii) k=0 φ (k).
P∞ 1 (k)
(iv) k=1 k φ (k).
R
(v) R φ2 (x) dx.
Proof. (i) This is a distribution of order 2.
(ii) This is not a distribution because it is not defined for all φ ∈ D(R). For
instance, choose φ(x) = ex−π ψ(x), where ψ ∈ D(R) such that ψ(x) = 1
in a neighbourhood of π. Note that φ(k) (π) = 1, for all k and the sum
is infinite.
(iii) This is a distribution of infinite order.
(iv) This is a distribution of infinite order.
(v) This is not a distribution, because it is nonlinear.
Exercise 12. Show that if T ∈ D0 (Ω) and ω is an open subset of Ω, then

T ∈ D0 (ω).
Definition 1.4.5. Two distributions S, T ∈ D0 (Ω) are said to be equal if
S(φ) = T (φ) for all φ ∈ D(Ω).
1.4.1 Functions as Distributions

Definition 1.4.6. We say a functionRf is locally integrable in Ω, denoted
as f ∈ L1loc (Ω), if f is measurable and K |f (x)| dx < +∞, for every compact
set K ⊂ Ω.
In other words, f ∈ L1loc (Ω) if f ∈ L1 (K) for all compact subsets K of Ω.
In particular, L1 (Ω) ⊂ L1loc (Ω). This inclusion is strict because, if Ω is not
of finite measure, then constant functions do not belong to L1 (Ω) but they
are in L1loc (Ω).
Example 1.6. The function ln |x| ∈ L1loc (R). To see this it is enough to show
R1
that limε→0+ ε | ln x| dx exists and is finite. Consider
Z 1
lim+ | ln x| dx = lim+ −(x ln x − x) |1ε = −[−1 − lim+ (ε ln(ε) − ε)] = 1.
ε→0 ε ε→0 ε→0
Because ε ln(ε) → 0 as ε → 0+ (use L’Hospital’s rule after proper substitu-

tion).
Example 1.7. The function 1/x ∈ / L1loc (R) (after assigning a real value at
x = 0). Similarly, e1/x ∈/ L1loc (R). In both these cases the integral is not
finite on a compact set containing origin.
Example 1.8. For any integer n ≥ 1 and β > 0, the function |x|−β ∈ L1loc (Rn ),
for all 0 < β < n, because
Z Z Z ρ
−β
|x| dx = r−β+n−1 dr dσ.
Bρ (0) Sρ (0) 0
Thus, for −β + n − 1 > −1 or β < n, the integral is finite and is equal to

ωn
ρ2n−β−1 n−β , where ωn is the surface measure of the unit ball. Note that the
function |x|−β has a blow-up near 0.
The notion of locally integrable functions can be extended to Lp spaces,
for all p ≥ 1.
Definition 1.4.7. For 1 ≤ p < ∞, we say a function f is
R locally pp-integrable
p
in Ω, denoted as f ∈ Lloc (Ω), if f is measurable and K |f (x)| dx < +∞,
for every compact set K ⊂ Ω.
We, in general, will not treat separately the notion of locally p-integrable
functions because any locally p-integrable function is locally integrable. i.e.,
if f ∈ Lploc (Ω) for all 1 < p < ∞, then f ∈ L1loc (Ω). Because the Hölder’s
inequality implies that for any compact subset K of Ω,
Z Z 1/p
p
|f (x)| dx ≤ |f | dx (µ(K))1/q < +∞.
K K
Thus, f ∈ L1loc (Ω). Since any Lp (Ω), for 1 ≤ p < ∞, is in Lploc (Ω) and hence
are in L1loc (Ω).
Exercise 13. If f is continuous on Ω, then f ∈ L1loc (Ω).
We shall now observe that to every locally integrable function one can
associate a distribution. For any f ∈ L1loc (Ω), we define the functional Tf on
D(Ω) defined as, Z
Tf (φ) = f (x)φ(x) dx.
Ω
The functional Tf is continuous on D(Ω) (and, hence, is in D0 (Ω)) because

for every compact set K in Ω,
Z
∞
|Tf (φ)| ≤ |f | dx kφk0 , ∀φ ∈ CK (Ω).
K
The distribution Tf is of zero order. We usually identify the distribution Tf

with the function f that induces it. Hence, every continuous function and
every Lp function, for p ≥ 1, induces the distribution described above.
Exercise 14. If f : Ω → R is such that
Z
|f φ| dx < +∞ ∀φ ∈ D(Ω)
Ω
then f ∈ L1loc (Ω).
Proof. Let K be a compact subset of Ω. Then there is a φK ∈ D(Ω) such

that φK ≡ 1 on K and 0 ≤ φK ≤ 1. Hence,
Z Z Z
|f (x)| dx = |f (x)| · 1K dx ≤ |f (x)φK (x)| dx < +∞.
K Ω Ω
Thus, f ∈ L1loc (Ω). The function 1K denotes the characteristic function which
takes 1 on K and zero on the K c .
The constant function 0 on Ω induces the zero distribution (the zero

functional on D(Ω)).
Exercise 15. If f ∈ C(Ω) such that Tf ≡ 0, i.e., for all φ ∈ D(Ω),
Z
f φ dx = 0
Ω
then f ≡ 0 in Ω.
Proof. Let us assume f is continuous. If f 6= 0 in Ω, then there is a x0 ∈
Ω such that f (x0 ) = λ 6= 0. Define the function R g = f /λ which is also
continuous on Ω such that g(x0 ) = 1. Note the gφ = 0 for all φ ∈ D(Ω).
By the continuity of g, there is an r > 0 such that, for all x ∈ Br (x0 ) ⊂ Ω,
g(x) R> 1/2. Choose a test function ψ ∈ D(Ω) such that supp(ψ)⊂ Br (x0 )
and Ω ψ(x) dx = 1. For instance, choose ψ to be the mollifier function
ψ(x) = ρr (x − x0 ). Then, in particular,
Z Z Z
1 1
0= gψ dx = gψ dx > ψ dx = > 0,
Ω Br (x0 ) 2 Br (x0 ) 2
a contradiction. Thus, g ≡ 0 on Ω and hence f ≡ 0 on Ω.

Exercise 16. If f ∈ L1loc (Ω) is non-negative, f ≥ 0, such that Tf ≡ 0, i.e., for
all φ ∈ D(Ω), Z
f φ dx = 0
Ω
then f = 0 a.e. in Ω.
Proof. Consider a compact subset K of Ω. By outer regularity of Lebesgue
measure, for every ε > 0, there is an open set Uε ⊃ K such that µ(Uε \K) < ε.
Let {φε } be a sequence of functions in D(Ω) such that φε ≡ 1 on K, φε ≡ 0
on Uεc and 0 ≤ φε ≤ 1. Note that φε converge point-wise to the characteristic
function 1K , as ε goes to 0. Therefore, by dominated convergence theorem,
Z Z Z
0 = lim f φε dx = f · 1K dx = f dx.
ε→0 Ω Ω K
R R R
Since f ≥ 0, K f dx = K |f | dx. Thus, K f dx = kf k1,K = 0 for all compact
subsets K of Ω. Hence f = 0 a.e. on all compact K of Ω and hence f = 0
a.e. in Ω.
The non-negative hypthesis on f in the above result can be relaxed.

Exercise 17. If f ∈ L1loc (Ω), such that Tf ≡ 0, i.e., for all φ ∈ D(Ω),
Z
f φ dx = 0
Ω
then f = 0 a.e. in Ω.
Proof. Note that it is enough to prove that f = 0 a.e. on all compact subsets
K of Ω. For any compact subset K of Ω, choose φK ∈ D(Ω) such that
φK ≡ 1. Define, for each x ∈ K,
Z
fε (x) := f φK ∗ ρε (x) = f (y)φK (y)ρε (x − y) dy.
K
Since y 7→ φK (y)ρε (x − y) is in D(Ω), we have fε ≡ 0 on K. Since fε → f in

L1 (K),
Z Z
0 = lim kfε − f k1,K = lim |fε − f | dx = |f | dx = kf k1,K
ε→0 ε→0 K K
and hence f = 0 a.e. in K.

A consequence of the above observations is that the association f 7→ Tf
is well-defined because if f = g a.e. in Ω, then the distributions Tf = Tg .
Thus, we have seen L1loc (Ω) ⊂ D0 (Ω). In particular, D(Ω) ⊂ D0 (Ω). Any
distribution that is induced by a locally integrable function is called a regular
distribution. Otherwise, the distribution is called singular.
1.4.2 Measures as Distributions

To each Radon measure µ ∈ R(Ω), we associate a linear functional on D(Ω)
as follows: Z
Tµ (φ) = φ dµ, ∀φ ∈ D(Ω).
Ω
The functional Tµ is continuous on D(Ω) (and is in D0 (Ω)) because for every
compact set K in Ω,
Z
∞
|Tµ (φ)| ≤ dµ kφk∞,K = |µ|(K)kφk∞,Ω , ∀φ ∈ CK (Ω),
K
where |µ| denotes the total variation of the measure. The distribution Tµ is
of zero order.
Example 1.9. The Lebesgue measure on Ω restricted to the Borel σ-algebra

of Ω is a Radon measure. The distribution induced by the Lebesgue measure
is same as that induced by the locally integrable constant function 1 on Ω.
Thus, the distribution induced by Lebesgue measure is a regular distribution.
Example 1.10 (Dirac Measure). The Dirac measure is a Radon measure that
assigns a total mass of 1 at a point a ∈ Ω. Thus, on the Borel measurable
subsets of Ω, (
1 if a ∈ E
δa (E) =
0 if a ∈
/ E.
If a = 0 then the corresponding measure δ0 = δ is the Dirac measure dis-
cussed in §1.1. The distribution induced by Dirac measure δa , called Dirac
distribution, is: Z
δa (φ) = φ(x) dδa .
Ω
Example 1.11. Let γ ⊂ Ω be a curve of length one (|γ| = 1). Given β ∈ R,

we define on the Borel measurable subsets of Ω, µγ (E) = β × |γ ∩ E|. This
measure is a generalization of the Dirac measure which signifies that β units
of mass are uniformly distributed on a curve of length one.
Exercise 18. Show that, for any8 function g on Ω,
Z
g(x) dδa = g(a).
Ω
Proof. We shall give two proofs: The first proof is the usual measure theory
technique, If g = 1E , the characteristic function of E, then
Z Z
1E dδa = dδa = δa (E) = 1E (a).
Ω E
Pk
For any simple function g = i=1 αi 1Ei , then
Z k
X
g(x) dδa = αi δa (Ei ) = g(a).
Ω i=1
8
Yes! we mean “any” function, because the class of Dirac measurable sets is the power
set and hence any function is Dirac measurable. Though, in the above context we have
restricted the Dirac measure to the class of Borel σ-algebra
For any non-negative g, by choosing an increasing sequence of simple func-

tions gi converging to g, one can show that
Z
g(x) dδa = g(a).
Ω
Further any g = g + − g − , which ends the first proof.

The second proof is quite elegant. Consider the constant function h on Ω
defined as h(x) = g(a). Let
E := {x ∈ Ω/g(x) 6= h(x)}.
Hence, a ∈
/ E and therefore δa (E) = 0. This implies that g = h almost
everywhere w.r.t δa and, hence, their integrals coincide
Z Z
g(x) dδa = h(x)dδa .
Ω Ω
But Z
h(x)dδa = g(a)δa (Ω) = g(a).
Ω
With the above observations, we note that the Dirac distribution is just
δa (φ) = φ(a) ∀φ ∈ D(Ω). (1.4.1)
Recall that the distribution induced by Lebesgue measure is same as the

one induced by the constant function 1 ∈ L1loc (Ω). If all measure induced
distributions were also induced by some locally integrable functions, then we
have not achieved anything different by considering measures as distribution.
But it turns out that not all measure induced distribution are induced by
locally integrable functions.
Proposition 1.4.8. The Dirac distribution is a singular distribution, i.e.,
there is no f ∈ L1loc (Ω) such that Tf = δa .
Proof. Note that for any a ∈ Ω, since Ω is open, there is a ε0 > 0 such that
B(a, ε0 ) ⊂ Ω. For each 0 < ε < ε0 , we choose φε ∈ D(Ω) with support in
B(a; ε), 0 ≤ φε ≤ 1 and φε = 1 in B(a; 2ε ). Thus,
δa (φε ) = φε (a) = 1 for all 0 < ε < ε0 .

Suppose f is a locally integrable function such that Tf = δa , then

Z Z Z
δa (φε ) = Tf (φε ) = f φε dx = f φε dx ≤ |f | dx.
Ω B(a;ε) B(a;ε)
R
Therefore, 1 ≤ B(a;ε) |f | dx. Since f ∈ L1loc (Ω), the quantity on RHS is finite.
Hence, as ε → 0, we get a contradiction 1 ≤ 0.
There are many more measures that induce singular distribution. In fact,
one can identify the class of measures which induce regular distribution.
They are, precisely, the absolutely continuous9 measures w.r.t the Lebesgue
measure. We shall not dwell on this topic, but a quick summary is as follows:
Note that for each f ∈ L1loc (Ω), one can define a signed measure
Z
µf (E) := f dx (1.4.2)
E
for all measurable subsets E of Ω. Conversely, the Radon-Nikodym theo-

rem states that, for any absolutely continuous measure µ w.r.t the Lebesgue
measure, there is a f ∈ L1loc (Ω) such that
Z
µ(E) = f dx
E
for all µ-measurable subsets of Ω. This is a generalisation of the Fundamental

theorem of Calculus (FTC).
With results of this section we have the following inclusions
D(Ω) ( Cc (Ω) ⊆ C0 (Ω) ( L1loc (Ω) ( R(Ω) ⊂ D0 (Ω).
Note that so far we have only seen examples of distributions which are of
zero order. This is because of the following result:
Theorem 1.4.9. A distribution T ∈ D0 (Ω) is of zero order iff T is a distri-

bution induced by a Radon measure.
9
A measure µ is absolutely continuous w.r.t another measure ν if for every element E
of the σ-algebra µ(E) = 0 whenever ν(A) = 0, denoted as µ ν. The Lebesgue measure
is absolutely continuous w.r.t counting measure but the converse is not true.
Proof. The implication that a Radon measure induced distribution is of zero

order is already shown. Conversely, let T be a distribution of zero order.
Then, for every compact set K ⊂ Ω, there is a CK > 0 such that
∞
|T (φ)| ≤ CK kφk0 ∀φ ∈ CK (Ω).
The idea is to continuously extend T to Cc (Ω), in a unique way, and then

consider the Radon measure obtained from Theorem 1.3.15. Let φ ∈ Cc (Ω)
and define φε := φ ∗ ρε . By Theorem 1.3.28, φε converges uniformly to φ.
The sequence {T (φε )} is Cauchy in R because
|T (φε ) − T (φδ )| = |T (φε − φδ )| ≤ CK kφε − φδ k0 .
Hence the sequence {T (φε )} converges, thus we extend T uniquely on Cc (Ω)

as T (φ) = limε→0 T (φε ). Thus, Theorem ??, there is a Radon measure
associated to the extended T .
Definition 1.4.10. A distribution T ∈ D0 (Ω) is said to be positive if T (φ) ≥

0 for all φ ≥ 0 in D(Ω).
Exercise 19. Any positive functional on D(Ω) is a distribution of order zero

and hence corresponds to a positive Radon measure.
∞
Proof. For any compact set K ⊂ Ω, consider φ ∈ CK (Ω) and ψ ∈ D(Ω), a
non-negative function, such that ψ = 1 on K. Set λ = kφk0 , then λ−φ(x) ≥ 0
on K. Thus, λψ(x) − φ(x) ≥ 0 is non-negative on Ω. Therefore,
0 ≤ T (λψ(x) − φ(x))
T (φ) ≤ λT (ψ) = T (ψ)kφk0 .
Hence, T is of order zero. Thus, there is a Radon measure associated to

T.
1.4.3 Multipole Distributions

Recall the formulation of Dirac distribution as given in the equation (1.4.1).
This motivates, for a fixed non-negative integer k and a ∈ R, the linear
functional on D(R)
δa(k) (φ) = φ(k) (a), ∀φ ∈ D(R),

where φ(k) denotes the k-th derivative of φ. This functional is continuous on

D(R) because, for all φ ∈ C ∞ (K),
|δa(k) (φ)| = |φ(k) (a)| ≤ kφkk .
The order of this distribution is, at most, k. The situation k = 0 corresponds

to the Dirac distribution. The situation k = 1 is called the dipole or doublet
distribution.
Example 1.12. We show in this example that the order of the dipole distribu-
∞
tion cannot be zero. Suppose that for all compact set K ⊂ R and φ ∈ CK (Ω)
we have
(1)
|δ0 (φ)| ≤ Ckφk0 .
∞
Now, choose φ ∈ CK (Ω) such that 0 ∈ Int(K) and φ0 (0) 6= 0. For each
integer m ≥ 1, set φm (x) := (1/m)φ(mx). Note that φ0m (x) = φ0 (mx) and
hence
(1) C
0 6= |φ0 (0)| = |φ0m (0)| = |δ0 (φm )| ≤ Ckφm k0 = kφk0 .
m
This is a contradiction because RHS converges to zero, as m increases, and
LHS is strictly positive.
Example 1.13. The example above can be tweaked to show that any k-th
(k)
multipole distribution, δ0 cannot have order less than k and hence is of
∞
order k. Suppose that for all compact set K ⊂ R and φ ∈ CK (Ω) we have
(k)
|δ0 (φ)| ≤ Ckφkj for some 0 ≤ j < k.
∞
Now, choose φ ∈ CK (Ω) such that 0 ∈ Int(K) and φ(k) (0) 6= 0. For each
(j)
integer m ≥ 1, set φm (x) := m−k φ(mx). Note that φm (x) = mj−k φ(j) (mx).
Thus,
j
(k)
X
(k) (k)
0 6= |φ (0)| = |φm (0)| = |δ0 (φm )| ≤ Ckφm kj = C kφ(i)
m k0
i=0
j j
X X
6= C mi−k kφ(i) k0 ≤ Cmj−k kφ(i) k0 = Cmj−k kφkj .
i=0 i=0
This is a contradiction because j < k and RHS converges to zero, as m

increases whereas LHS is strictly positive.
One may generalise these notions appropriately to higher dimensions. For

any open subset Ω of Rn , we define the distribution
δaα (φ) = Dα φ(a), ∀φ ∈ D(Ω),
where α is the n tuple of non-negative integers. Just as Dirac distribution

models charge density at a point charge, the dipole distribution models charge
density for an electric dipole. In fact, any dipole behaviour, viz., the magnetic
dipole layer on a surface. For instance, given a smooth surface Γ and f is a
continuous function on Γ representing the density of the magnetic moment,
one can define the distribution corresponding to the magnetic dipole layer
on Γ as, Z
dφ(x)
T (φ) = f (x) ds(x)
Γ dn
where ds denotes the surface element of Γ and d/dn is the derivation in the
direction of normal of Γ.
The dipole distribution is singular. In fact, the situation is much worse.
The dipole distribution is not induced by any Radon measure.
(1)
Proposition 1.4.11. The dipole distribution δa ∈ D(R) is not induced by
(1)
any Radon measure, i.e., there is no Radon measure µ such that Tµ = δa .
Proof. Let Ω = R and a = 0. Choose φ ∈ D(R) with support in [−1, 1]

such that 0 ≤ φ ≤ 1 and φ ≡ 1 on [−1/2, 1/2]. Set φm (x) = sin(mx)φ(x).
Thus, supp(φm ) ⊂ [−1, 1] and |φm | ≤ 1. Note that the derivative of φm
is φ0m (x) = sin(mx)φ0 (x) + m cos(mx)φ(x) and, hence, φ0m (0) = m for all
positive integer m > 0. Therefore, the dipole distribution takes the value,
(1)
δ0 (φm ) = φ0m (0) = m, ∀m ∈ N.
(1)
Suppose there exists a Radon measure µ inducing the dipole distribution δ0 ,
then
Z 1
(1)
m = |δ0 (φm )| = |Tµ (φm )| = φ(x) sin(mx) dµ ≤ |µ|(B(0; 1)).
−1
The inequality |µ|(B(0; 1)) ≥ m, for all m ∈ N, implies that |µ|(B(0; 1))
is infinite which contradicts the fact that the Radon measure µ is finite on
compact subsets of R.
We shall see later that the multipole distribution δaα is, in fact, the α-th
distributional derivative of the Dirac measure δa up to a sign change. This
motivation was behind the choice of notation of multipole distribution. In
this section, we have noted that the space of Radon measures is properly
contained in the space of distributions, i.e., R(R) ( D0 (R).
Exercise 20. Do we have singular distributions which are neither induced by
Radon measures nor distributional derivatives of Radon measures? (Hyper-
functions?)
1.4.4 Infinite Order Distributions

We have introduced distributions of finite order thus far. Does there exist
distributions of infinite order? In this section, we given an example of an
infinite order distribution.
Example 1.14. Define the functional T on D(R) as T (φ) := ∞ (k)
P
k=0 φ (k).
∞
Without loss of generality, consider the compact set [0, m]. If φ ∈ C[0,m] (R),
then m m
X X
(k)
|T (φ)| ≤ |φ (k)| = kφ(k) k0 = kφkm .
k=0 k=0
The larger the compact set becomes, the higher the derivatives of φ needs to
be taken in. Thus, there is no fixed m for all compact subsets of R.
Example 1.15. Define the functional T ∈ D0 (0, ∞) as
∞
X
T (φ) = φ(k) (1/k).
k=1
Let φ ∈ D(0, ∞) be such that supp(φ)⊂ [1/m, m] (exhaustion sets), then

m
X m
X m
X
(k) (k)
T (φ) = φ (1/k) ≤ kφ k0 ≤ kφ(k) k0 = kφkm .
k=1 k=1 k=0
Thus, T ∈ D0 (0, ∞) is a distribution of infinite order.
1.4.5 Topology on Distributions

In this section, we give a suitable topology on the space of distributions.
Recall that the space of distributions D0 (Ω) is the (topological) dual of D(Ω).
For every φ ∈ D(Ω), one can define the linear functional Λφ : D0 (Ω) → R as
follows, Λφ (T ) = T (φ). The linear functionals Λφ are included in the second
(algebraic) dual of D(Ω). We consider the coarsest10 topology on D0 (Ω) such
that all the linear maps Λφ : D0 (Ω) → R, corresponding to each φ ∈ D(Ω),
are continuous. This is called the weak-* topology on D0 (Ω). For every open
subset of V ⊂ R, consider the collection of subsets {Λ−1 0
φ (V )} of D (Ω), for all
φ ∈ D(Ω). The weak-* topology is the topology generated by this collection
of subsets in D0 (Ω). The space D0 (Ω) is sequentially weak-* complete but
not weak-* complete.
Definition 1.4.12. We say a sequence of distributions {Tm } is Cauchy in

D0 (Ω), if Tm (φ) is Cauchy in R, for all φ ∈ D(Ω). A sequence {Tm } ⊂ D0 (Ω)
converges to T in weak-* topology of D0 (Ω) or in the distribution sense if
Tm (φ) → T (φ) for all φ ∈ D(Ω), denoted as Tm * T .
Exercise 21. If Sm and Tm are sequences of distributions converging to S

and T , respectively, then show that λSm + µTm converges to λS + µT , for
all λ, µ ∈ R.
Proposition 1.4.13 (Sequential Completeness). Let {Tm } be a Cauchy se-

quence (of distributions) in D0 (Ω) and let T (φ) := limm→∞ Tm (φ) (T is well-
defined because R is complete). Then T ∈ D0 (Ω).
Proof. To show T ∈ D0 (Ω), it is enough to show that T : CK ∞

(Ω) → R is
continuous for all compact sets K in Ω. Fix a compact set K in Ω and
φ ∈ Ω. Since {Tm (φ)} is convergent in R it is bounded and hence there is a
real constant Cφ > 0 (may depend on φ) such that supm |Tm (φ)| ≤ Cφ . The
∞
family {Tm } is point-wise bounded on CK (Ω). Then, by Banach-Steinhaus
∞
theorem (uniform boundeness principle) of CK (Ω), T is uniformly bounded,
i.e., there is a constant C > 0 (independent of φ) such that supm |Tm (φ)| ≤ C
∞ ∞
for all φ ∈ CK (Ω). Therefore, for all φ ∈ CK (Ω),
∞
|T (φ)| ≤ sup |Tm (φ)| ≤ C ∀φ ∈ CK (Ω).
m
∞
and hence T is continuous on CK (Ω) for all compact subsets K of Ω. Thus,
T is continuous on D(Ω) and hence T ∈ D0 (Ω).
10
weakest or smallest topology, one with fewer open sets
A natural question is the relation between distributional convergence and

poin-wise, uniform or Lp convergence of regular distributions. The weak-*
convergence in D0 (Ω) is weaker than any other topology that preserves the
continuity of the functionals Λφ . For instance, this topology is weaker that
Lp -topology or uniform norm topology etc., i.e., convergence in p-norm or
uniform norm implies convergence in distribution sense of classical functions.
Example 1.16. Consider the sequence fm (x) = eimx on R which converges
point-wise to 0 iff x ∈ 2πZ. The sequence of distributions corresponding to
fm converges to zero in the distributional sense because
Z Z
1
imx
e φ(x) dx = − eimx φ0 (x) dx.
im
We have used integration by parts to get the second integral and since the
integrand is bounded in the second integral, it converges to zero as m → ∞.
R
Example 1.17. Let ρ ∈ Cc (Rn ) such that supp(ρ)⊂ B(0; 1) and Rn ρ(y) dy =
1. Now set ρε (x) := ε−n ρ( xε ). This sequence is called the Dirac Sequence. In
particular, the sequence of mollifiers (cf. (1.3.3)) is one example. Another
trivial example is (
2 |x| < 1
ρ(x) =
0 |x| ≥ 1.
We claim that the sequence ρε converges to the Dirac distribution δ0 , in the
distribution sense. Let Tε denote the distribution corresponding to ρε . For
any φ ∈ D(Rn ) , consider
Z Z x Z
−n
Tε (φ) = ρε (x)φ(x) dx = ε ρ φ(x) dx = ρ(y)φ(εy) dy.
Rn Rn ε Rn
Taking limit ε → 0 both sides we get
Z
lim Tε (φ) = ρ(y)φ(0) dy = φ(0) = δ0 (φ).
ε→0 Rn
The interchange of limit is possible due to uniform continuity of φ which
induces the uniform convergence of φ(εy) → φ(0). Thus, Tε → δ0 in the
distribution sense, whereas the point-wise limit at x = 0 did not exist for ρε .
Example 1.18. Consider the sequence of functions

m x
 2
0 ≤ x < m1
fm x = m2 n2 − x 1
< x ≤ m2

 m
0 otherwise.

Note that they converge point-wise to zero for all x ∈ R. Let Tm denote the
distribution corresponding to fm and hence
Z 1/m Z 2/m
2 2 2
Tm (φ) = m xφ(x) dx + m − x φ(x) dx.
0 1/m n
Both the integral above converges to φ(0)/2 and hence the sequence of dis-
tributions converges to δ0 . Let us give the proof for the first integral. Since
Z 1/m
2 1
m x dx =
0 2
we consider
Z 1/m Z 1/m
2 1 2
m xφ(x) dx − φ(0) = m x[φ(x) − φ(0)] dx
0 2 0
Z 1/m
2
≤ m |x||φ(x) − φ(0)| dx.
0
Since φ is continuous at x = 0, for every ε > 0, there is a δ > 0 such that

|φ(x) − φ(0)| < ε whenever |x| < δ. Thus, for all m such that 1/m < δ, we
have
Z 1/m Z 1/m
2 1 2 ε
m xφ(x) dx − φ(0) ≤ m dx < ε.
0 2 m 0
Hence,
Z 1/m
2 1
lim m xφ(x) dx = φ(0).
m→∞ 0 2
Similarly,
R one can show that the second integral is also (1/2)φ(0) and hence
fm (x)φ(x) dx = φ(0).
The following theorem gives the condition that is violated by the above
example for the point-wise limit to coincide with the distributional limit.
Theorem 1.4.14 (Weak Dominated Convergence Theorem). Let {fm } ⊂

L1loc (Ω) such that fm (x) → f (x) point-wise for a.e. x ∈ Ω and there is a
g ∈ L1 (Ω) such that for every compact set K ⊂ Ω |fm | ≤ g for all m. Then
f ∈ L1loc (Ω) and fm * f in D0 (Ω).
Proof. Let Tm denote the distribution corresponding to fm and fix φ ∈ D(Ω).

Set hm (x) = fm (x)φ(x) and h(x) = f (x)φ(x). Note that hm (x) → h(x)
point-wise a.e. in Ω (in fact in the supp(φ)). Also, |hm | ≤ g|φ| and g|φ| ∈
L1 (Ω). By classical Lebesgue’s dominated convergence theorem, h ∈ L1 (Ω)
and Z Z Z
h(x) dx = lim hm (x) dx = lim fm (x)φ(x) dx.
Ω m→∞ Ω m→∞ Ω
In particular, choosing φ ≡ 1 on a given compact set K of Ω and since for

this φ, h ∈ L1 (Ω), we have f ∈ L1loc (Ω) because
Z Z Z
∞> |h| dx = |f φ| dx ≥ |f | dx.
Ω Ω K
Let Tf be the distribution corresponding to f . Moreover,

Z Z
lim Tm (φ) = lim fm φ dx = f φ dx = Tf (φ).
m→∞ m→∞ Ω Ω
Corollary 1.4.15. Let {fm } ⊂ C(Ω) be a sequence of continuous functions

that converges uniformly on compact subsets of Ω to f . Then fm * f in
D0 (Ω).
By the vector space structure of D0 (Ω), we P

already have the finite sum
of distributions. If {Ti }1 ⊂ D (Ω) then T := ki=1 Ti ∈ D0 (Ω) defined as
k 0
T (φ) = ki=1 Ti (φ). The topology on D0 (Ω) can be used to give the notion
P
of series of distributions.
Definition 1.4.16.
P∞ For any countable collection of distributions {Ti }∞
1 ⊂
0 0
D (Ω) the series i=1 T
Pm i is said to converge to S ∈ D (Ω) if the sequence of
0
partial sums Tm := i=1 Ti converges to S in D (Ω).
1.4.6 Principal Value Distribution

In practice, we encounter functions whose integral diverges. This motivates
other ways of studying divergent integrals, like principal part, finite part of
divergent integral.
Definition 1.4.17 (Singularity at infinity). We say that the integral of a

function f (with singularity at ∞) exists in the generalised sense if the limit
Z b
lim f (x) dx
a,b→+∞ −a
exists. We say that the principal value of the integral exists if the limit (with
b = a)
Z b
lim f (x) dx
a→+∞ −a
exists.
Definition 1.4.18 (Singularity at point). We say that the integral of a com-

pact supported function f (with singularity at a point 0) exists in the gener-
alised sense if the limit
Z −a Z ∞
lim f (x) dx + f (x) dx
a,b→0+ −∞ b
exists. We say that the principal value of the integral exists if the limit (with
b = a) Z
lim+ f (x) dx
a→0 |x|>a
exists.
We have already seen that every locally integrable function can be identi-
/ L1loc (R).
fied with a distribution. We have already noted, in §1.4.1, that 1/x ∈
Consider the function fε : R → C, for each ε > 0, defined as
1
fε (x) = .
x + iε
What is the limit of fε , as ε → 0? Of course, when x 6= 0, fε (x) → 1/x.

What is the limit when x = 0? Classically
R1 we cannot make sense of this limit.
Similarly, note that the integral −1 (1/x) dx does not converge. However,
Z 1
lim fε (x) dx = lim ln(x + iε) |1−1 = ln(1) − ln(−1) = 0 − (iπ) = −iπ.
ε→0 −1 ε→0
R1
Thus, we expect −1 (1/x) dx = −iπ. This cannot be made sense classically11 .
Using the theory of distributions one can give a meaning to this observation.
We do know that 1/x ∈ L1loc (R \ {0}) and hence induces a distribution
in D(R \ {0}). One can extend this distribution to yield a distribution cor-
1

responding 1/x on R. We define the linear functional PV x on D(R) as
Z
1 1
PV (φ) = lim φ(x) dx. (1.4.3)
x ε→0 |x|≥ε x
As defined before, the limit on RHS is called the Cauchy’s Principal Value
and hence the use of ‘PV’ in the notation. Note that P V (1/x) is defined as
a distributional limit of the sequence of distributions corresponding to the
locally integrable functions
(
1
|x| > ε,
fε (x) := x
0 |x| ≤ ε,
1
in D0 (Ω).

i.e., fε * P V x
Lemma 1.4.19. The principal value of 1/x (limit on the RHS) exists.
Proof. Let φ ∈ D(R). Without loss of generality, let supp(φ)⊂ [−a, a] for
some real a > 0. For ε > 0 small enough, consider
Z Z −ε Z a
φ(x) φ(x) φ(x)
dx = dx + dx
|x|≥ε x −a x ε x
Z a Z a Z a
φ(x) φ(−x) φ(x) − φ(−x)
= dx − dx = dx.
ε x ε x ε x
Therefore, taking limit both sides as ε → 0,

Z a
1 φ(x) − φ(−x)
PV (φ) = dx
x 0 x
Z a Z x
1 0
= φ (s) ds dx (Using FTC)
0 x −x
Z a Z 1
1 0
= xφ (xt) dt dx.
0 x −1
11
recall (1.1.1)
R1
Set ψ(x) = −1
φ0 (xt) dt, then
Z a Z a Z a
1 1
PV (φ) = xψ(x) dx = ψ(x) dx ≤ |ψ(x)| dx.
x 0 x 0 0
But |ψ(x)| ≤ 2kφ0 k0 ≤ 2kφk1 . Hence,

1
PV (φ) ≤ 2akφk1 .
x
Since RHS is finite, the limit is finite.
We have shown that the principal value functional is a distribution, called

the Principal Value (PV) Distribution corresponding to 1/x. The 1/x in
(1.1.1) derived by Dirac will make sense as a principal value distribution.
The proof also highlights that, at most, the PV distribution can have order
one. We show in the example below that the PV distribution cannot be of
order zero.
Example 1.19. The PV distribution cannot be of zero order and hence its
order is one. Observe the following characterisation of principal value distri-
bution in the above argument when supp(φ)⊂ [−a, a],
Z a
1 φ(x) − φ(−x)
PV (φ) = dx.
x 0 x
Choose φ such that supp(φ)⊂ [0, a], 0 ≤ φ ≤ 1 and φ ≡ 1 on [c, d] ⊂ [0, a]

with c 6= 0 and d 6= a. Then,
Z c Z d Z a
φ(x) φ(x) φ(x)
P V (1/x)(φ) = dx + dx + dx
0 x c x d x
Z d
1
> dx first and third integral being non-negative
c x
1 d
Z
> dx since x < a
a c
1 d−c d−c
= (d − c) ≥ sup |φ| = kφk0 .
a a x∈[0,a] a
Thus, the distribution cannot be of zero order.

The principal value distribution is not the only choice for the function
1/x. Consider the linear functional on D(R) corresponding to 1/x,
Z −ε Z ∞
φ(x) φ(x)
T (φ) = lim dx + dx.
ε→0 −∞ x 2ε x
Exercise 22. Show that the T defined above is a distribution. Also, compare
T and PV(1/x).
Proof. Let φ ∈ D(R). Without loss of generality, let supp(φ)⊂ [−a, a] for
some real a > 0. For ε > 0 small enough, consider
Z −ε Z ∞ Z −2ε Z −ε
φ(x) φ(x) φ(x) φ(x)
dx + dx = dx + dx
−∞ x 2ε x −a x −2ε x
Z a
φ(x)
+ dx
2ε x
Z a Z −ε
φ(x) − φ(−x) φ(x)
= dx + dx
2ε x −2ε x
Z a Z 2
φ(−εx)
= ψ(x) dx − dx.
2ε 1 x
Therefore,
Z a Z 2 Z a
φ(0)
T (φ) = ψ(x) dx − dx = ψ(x) dx − φ(0) ln 2.
0 1 x 0
Thus, |T (φ)| ≤ 2akφ0 k0 + kφk0 ln 2 ≤ Ckφk1 where C = max(2a, ln 2). T is
again a distribution of order one. Now, consider
Z 2ε Z 2
φ(x) φ(εx)
(PV(1/x) − T )(φ) = lim dx = lim dx = φ(0) ln 2.
ε→0 ε x ε→0 1 x
Remark 1.4.20. Yet another way to defining P V (1/x) is as follows, this

time from the complex plane. Consider the distributions Tε corresponding
to
1
fε (x) := .
x + iε
We shall show that Tε * P V (1/x) − iπδ0 . Similarly, if Sε are distributions
corresponding to
1
fε (x) := .
x − iε
We shall show that Sε * P V (1/x) + iπδ0 .
1.4.7 Functions, but not Distributions

It is not always possible to extend the notion of distributions to non-locally
integrable functions as done above.
Example 1.20. The function e1/x ∈ / L1loc (R) but is in L1loc (R \ {0}). Let
T ∈ D0 (0, ∞) be the distribution corresponding to e1/x on (0, ∞). We shall
show that there is no distribution S ∈ D0 (R) corresponding to e1/x on R
whose restriction coincides with the distribution T on (0, ∞).
Example 1.21. Consider the distribution T ∈ D0 (0, ∞) defined in Exam-
ple 1.15, i.e.,
X∞
T (φ) = φ(k) (1/k).
k=1
We shall show that there is no distribution S ∈ D0 (R) whose restriction

to (0, ∞) coincides with T . Suppose there exists a S ∈ D0 (R) such that
T = S on (0, ∞) then for the compact set [−1, 1] there exists a C > 0 and
N ∈ N ∪ {0} such that
∞
|S(φ)| ≤ CkφkN ∀φ ∈ C[−1,1] (R).
Now, choose m > N and ψ such that supp(ψ)⊂ (1/(m+1), 1/(m−1)). Then
T (ψ) = ψ (m) (1/m). Since T and S coincide for ψ ∈ D(0, ∞), we have
(m)
|δ1/m (ψ)| = |ψ (m) (1/m)| = |T (ψ)| = |S(ψ)| ≤ CkψkN .
(m)
Thus, we have that the order of the multipole distribution δ1/m is at most
N , a quantity smaller than m. This is a contradiction. Hence, we can have
no extension S of T to R.
1.5 Operations With Distributions

The space of distributions D0 (Ω) is a vector space over R (or C). Thus,
we already have the operation of addition of distribution and multiplication
by reals induced from the property of linear functionals. Hence, for any two
S, T ∈ D0 (Ω), (S + T )(φ) = S(φ) + T (φ). Also, for any λ ∈ R and T ∈ D0 (Ω),
the distribution λT is defined as (λT )(φ) = λT (φ). We begin by introducing
the concept of derivative of a distribution.
1.5.1 Differentiation
Recall the discussion leading to (1.3.1). For any f ∈ C ∞ (R), φ ∈ D(R) and
all k ∈ N, using integration by parts we have
Z Z
(k) k
f φ dx = (−1) f φ(k) dx.
This motivates the following definition of derivative of a distribution.
Definition 1.5.1. For any T ∈ D0 (Ω) and n-tuple α, the derivative Dα T of

T is defined as
(Dα T )(φ) = (−1)|α| T (Dα φ), φ ∈ D(Ω).
The derivative Dα T is also a distribution because for all compact subsets

K of Ω there are CK > 0 and non-negative integer NK = N such that
∞
|T (φ)| ≤ CK kφkN for all φ ∈ CK (Ω). Thus, for any compact subset K of Ω
∞
and φ ∈ CK (Ω),
|(Dα T )(φ)| = |(−1)|α| T (Dα φ)| = |T (Dα φ)| ≤ CK kDα φkN ≤ CK kφkN +|α| .
Thus, Dα T ∈ D0 (Ω). For a finite order distribution T , the derivative Dα T

has order more than T , i.e., differentiation operation increases the order of
the distribution. Observe that every distribution is infinitely differentiable
and the mixed derivatives are equal because the same holds for test functions.
The following result highlights some of these properties of the differentiation
operation:
Exercise 23. Let S, T ∈ D0 (Ω). Show that
(i) Dβ (Dα T ) = Dα (Dβ T ) = Dα+β T for all multi-indices α, β.
(ii) Dα (λS + µT ) = λDα S + µDα T , for each λ, µ ∈ R.
Proof. (i) We see that
Dβ (Dα T )(φ) = (−1)|β| Dα T (Dβ φ) = (−1)|β|+|α| T (Dα Dβ φ)

= (−1)|β|+|α| T (Dα+β φ) = (Dα+β T )(φ).
Similarly for Dα (Dβ T ).

(ii) Consider
[Dα (λS + µT )](φ) = (−1)|α| (λS + µT )(Dα φ)

= (−1)|α| [λS(Dα φ) + µT (Dα φ)]
= (λDα S + µDα T )(φ).
Exercise 24. Let n = 1, Ω = (0, 2) and

(
x if 0 < x ≤ 1
u(x) =
1 if 1 ≤ x < 2.
Compute the first distributional derivative of u.
Proof. For any φ ∈ Cc∞ (Ω), consider

Z 2
Du(φ) = − u(x)φ0 (x) dx
Z0 1 Z 2
0
= − xφ dx − φ0 dx
Z 10 1
Z 2
= φ dx − φ(1) + φ(1) = v(x)φ(x) dx,
0 0
where (
1 if 0 < x ≤ 1
v(x) =
0 if 1 < x < 2.
Thus, Du = v.
Exercise 25. question Let n = 1, Ω = (0, 2) and

(
x if 0 < x ≤ 1
u(x) =
2 if 1 < x < 2.
Compute the first distributional derivative of u.

Proof. Consider
Z 2 Z 1 Z 2
0 0
Du(φ) = − uφ dx = − xφ dx − 2 φ0 dx
0 0 1
Z 1
= φ dx + φ(1).
0
0
Then, Du = v + δ1 in D (Ω) where
(
1 if 0 < x ≤ 1
v=
0 if 1 < x < 2.
Definition 1.5.2. A function f ∈ L1loc (Ω) is said to be α-weakly differen-

tiable if Dα Tf is a regular distribution. In other words, for any given multi-
index α, a function f ∈ L1loc (Ω) is said to be α-weakly differentiable if there
exists a gα ∈ L1loc (Ω) such that
Z Z
|α|
gα φ dx = (−1) f Dα φ dx ∀φ ∈ D(Ω).
Ω Ω
Example 1.22. Consider the continuous function f : R → R defined as f (x) =
|x|. Classically, its derivative exists a.e. and is the function
(
−1 −∞ < x < 0
f 0 (x) =
1 0 < x < ∞.
The distribution corresponding to f 0 is
Z 0 Z ∞
Tf 0 (φ) = − φ(x) dx + φ(x) dx
−∞ 0
However, for φ ∈ D(R), the distributional derivative of Tf is

Z 0 Z ∞
0 0
DTf (φ) = −Tf (φ ) = − (−x)φ (x) dx − xφ0 (x) dx
−∞ 0
Z 0 Z ∞
= xφ(x) |0−∞ − φ(x) dx + φ(x) dx − xφ(x) |∞
0
−∞ 0
Z ∞ Z 0
= φ(x) dx − φ(x) dx
0 −∞
= T (φ).
f0
Therefore, DTf = Tf 0 . Moreover, |x| is weakly differentiable.

Example 1.23. Consider the everywhere discontinuous function

(
1 x ∈ Qc
f (x) =
0 x ∈ Q.
The function f is not differentiable at any point. The Lebesgue measure of
Q is zero, hence,
Z
DTf (φ) = −Tf (φ ) = − φ0 (x) dx = 0.
0
R
Thus, DTf = 0 and f is weakly differentiable because DTf = T0 .

Example 1.24. Consider the everywhere discontinuous function
(
sin x x ∈ Qc
f (x) =
0 x ∈ Q.
The function f is not differentiable at any point. The Lebesgue measure of
Q is zero, hence,
Z Z
0 0
DTf (φ) = −Tf (φ ) = − sin xφ (x) dx = cos xφ(x) dx.
R R
Thus, DTf = Tcos x and hence f is weakly differentiable.

Example 1.25. We shall now give an example of a function for which the
classical and distributional derivative do not coincide. Consider the locally
integrable function Ha , for every a ∈ R,
(
1 if x > a
Ha (x) =
0 if x ≤ a.
When a = 0, the function H0 is the Heaviside function H as defined in
(1.1.2). Classically, the function Ha is differentiable a.e. and Ha0 (x) = 0 a.e.
(except at a). Therefore, THa0 = 0. We shall now compute the distributional
derivative of the regular distribution THa , induced by Ha . For φ ∈ R, the
derivative of THa is computed as,
Z ∞
0
DTHa (φ) = −THa (φ ) = − φ0 (x) dx = φ(a) = δa (φ).
a
Thus, the distributional derivative of the function Ha is the Dirac distribu-

tion, a singular distribtuion. Therefore, DTHa = δa (φ) 6= 0 = THa0 . In fact,
Ha is not weakly differentiable.
Observe above that the function Ha had a jump discontinuity and Dirac
measure appeared as its derivative at the point of “jump”. This is the feature
of Dirac measure. It appears as a derivative at points of jump. However,
note that in Example 1.23 and 1.24, the jump at every point did not give
rise to a Dirac measure because outside the set of jump points (in that case
Q) the function was continuous.
Example 1.26. Consider the locally integrable function f : R → R defined as
(
|x| |x| < a
f (x) =
0 |x| > a.
Classically, its derivative exists a.e. and is the function


−1 −a < x < 0

0
f (x) = 1 0<x<a

0 |x| > a.


Z 0 Z a
Tf 0 (φ) = − φ(x) dx + φ(x) dx
−a 0
However, for φ ∈ D(R), the distributional derivative of Tf is

Z 0 Z a
0 0
DTf (φ) = −Tf (φ ) = − (−x)φ (x) dx − xφ0 (x) dx
−a 0
Z 0 Z a
= xφ(x) |0−a − φ(x) dx + φ(x) dx − xφ(x) |a0
−a 0
Z a Z 0
= aφ(−a) − aφ(a) + φ(x) dx − φ(x) dx
0 −a
= a(δ−a (φ) − δa (φ)) + Tf 0 (φ).
Therefore, DTf = aδ−a −aδa +Tf 0 . The function f is not weakly differentiable
(compare this with Example 1.22).
Example 1.27. Consider the discontinuous function
(
x2 + x x < 1
f (x) =
e−5x x > 1.
Classically, its derivative exists a.e. and is the function
(
2x + 1 x<1
f 0 (x) =
−5e−5x x > 1.
Z 1 Z ∞
Tf 0 (φ) = (2x + 1)φ(x) dx − 5 e−5x φ(x) dx.
−∞ 1
The distributional derivative of f (Tf ) is
Z 1 Z ∞
0 0
DTf (φ) = −Tf (φ ) = − 2
(x + x)φ (x) dx − e−5x φ0 (x) dx
−∞ 1
Z 1 Z ∞
2 1
= −(x + x)φ(x) |−∞ + (2x + 1)φ(x) dx − 5 e−5x φ(x) dx
−∞ 1
− e−5x φ(x) |∞
1
−5
= (e − 2)φ(1) + Tf 0 (φ).
Therefore, DTf = (e−5 − 2)δ1 + Tf 0 . The function f is not weakly differen-

tiable.
Example 1.28. Consider the Cantor function fC : [0, 1] → [0, 1] (cf. Ap-
pendix A) extended continuously to R by setting
(
0 x≤0
fC (x) =
1 x ≥ 1.
Note that fC0 = 0 for all x ≤ 0 and x ≥ 1. Since C is of Lebesgue measure

zero and fC is constant on each interval removed from [0, 1], we have fC0 = 0
a.e. in R. Note that

Z 1 Z ∞ Z
0 0 0
DTfC (φ) = −TfC (φ ) = − fC φ dx − φ dx = φ(1) − fC φ0 dx
0 1 [0,1]\C
Z
= φ(1) − fC φ0 dx (cf. Appendix A)
∪∞ c
i=1 Ci
∞
X Z
= φ(1) − fC φ0 dx (since the union is disjoint)
i=1 Cic
X∞ Z bk
= φ(1) − ck φ0 dx
k=1 ak
(ak , bk are end-points of intervals removed)

X∞ ∞
X
= φ(1) + ck (φ(ak ) − φ(bk )) = ck (δak − δbk )(φ) + δ1 (φ).
k=1 k=1
Note that fC is not weakly differentiable. Moreover, the classical derivative

and distributional derivative do not coincide.
Example 1.29. The k-th distributional derivative of the Dirac distribution is
the k-th multipole distribution. For φ ∈ D(Ω),
Dβ δaα (φ) = (−1)|β| δaα (Dβ φ) = (−1)|β| Dα (Dβ φ(a))
= (−1)|β| Dα+β φ(a)) = (−1)|β| δaα+β (φ).
In particular, the distributional derivative of the Dirac distribution is the
dipole distribution, up to a factor of sign.
Example 1.30. Let us compute the distributional derivative of the distribu-
tion induced by ln |x|. For supp(φ)⊂ [−a, a],
Z
DTln |x| (φ) = −Tln |x| (φ ) = − ln |x|φ0 dx
0
R
Z −ε Z a
0 0
= − lim+ ln |x|φ dx + ln |x|φ dx
ε→0 −a ε
Z −ε Z a
φ(x) φ(x)
= lim+ (φ(ε) − φ(−ε)) ln ε + dx + dx
ε→0 −a x ε x
Z a
φ(x) − φ(−x)
Z
φ(x)
= lim+ dx = lim+ dx
ε→0 ε x ε→0 |x|>ε x

1
= PV (φ).
x
Exercise 26. For any β ∈ R, we have already noted that the function f (x) =
|x|−β is in L1loc (Rn ) for β < n. However, the function is weakly differentiable
for β + 1 < n and its weak partial derivative is fxi (x) = − |x|ββ+1 |x|
xi
.
R
Exercise 27. Let φ ∈ D(R). Show that R φ(x) dx = 0 iff there is a ψ ∈ D(R)
such that φ(x) = ψ 0 (x).
Proof. Let there exist a ψ ∈ D(R) such that φ(x) = ψ 0 (x). Then,
Z Z a
φ(x) dx = lim ψ 0 (x) dx = lim [ψ(a) − ψ(−a)] = 0.
R a→∞ −a a→∞
Rx
Conversely, let R φ(x) dx = 0. Then, set ψ(x) := −∞ φ(t) dt and ψ 0 = φ.
R
It only remains to show that ψ ∈ D(R). ψ ∈ C ∞ (R) since φ ∈ C ∞ (R).

Suppose that supp(φ) ⊂ [−M, M ]. Then for all x ∈ (∞, −M ), ψ(x) = 0
RM
since φ(x) = 0. For all x ∈ (M, ∞), ψ(x) = 0 because −M φ = 0. Thus,
supp(ψ) ⊂ [−M, M ].
Exercise 28. Let T ∈ D0 (R) be such that DT = 0, then show that T is a
regular distribution generated by a constant (a.e.) function.
Proof. Let T ∈ D0 (R) be such that DT (φ) = 0 for all φ ∈ D(R). Therefore,
0 = DT (φ) = −T (φ0 )
for all φ ∈ D(R). Consider the subset E ⊂ D(R), defined as
E := {ψ ∈ D(R) | ψ = φ0 for some φ ∈ D(R)}.
WeR have T (ψ) = 0 for all ψ ∈ E. Using Exercise 27,R we know that ψ ∈ E
iff R ψ(x) dx = 0. Choose any χ ∈ D(R)R such that R χ = 1 and, for each
φ ∈ D(R), set ψ := φ − βχ where β = R φ. Then ψ ∈ E and T (ψ) = 0.
Therefore, Z
T (φ) = βT (χ) = T (χ)φ(x) dx.
R
Thus, T = Tλ where λ = T (χ), a constant.
An alternate proof of above exercise is given in Exercise 45. By “regular
distribution generated by a constant function” we mean the function is con-
stant except on a set of measure zero. (cf. Example 1.23). The same is not
true with weak derivative. If the weak derivative is zero, then the function
is constant on each connected component of Ω.
Exercise 29. Solve the equation DT = S, for any given distribution S ∈

D0 (R).
Note that the distributional derivative of a locally integrable functions
that are also differentiable in classical sense, admit two notions of differen-
tiability. We have already seen in the examples above that the two notions
of differentiability need not coincide. For smooth functions the two notions
do coincide. Let f ∈ C ∞ (Ω) and Tf be its induced distribution. For any
n-tuple α, g := Dα f ∈ C ∞ (Ω) and let Tg be its induced distribution. Then
Z
α |α| α |α|
D Tf (φ) = (−1) Tf (D φ) = (−1) f Dα φ dx
Z Z
α
= D f φ dx = gφ = Tg (φ).
Thus, the two notions of derivatives coincide for all functions which respect
integration by parts.
Proposition 1.5.3. Let f, g be integrable functions on [a, b] and c, d be some

real constants. Set
Z x Z x
F (x) := c + f (t) dt; and G(x) := d + g(t) dt.
a a
Then,
Z b
[F (x)g(x) + G(x)f (x)] dx = F (b)G(b) − F (a)G(a).
a
Proof. Note that c = F (a) and d = G(a). Now, consider

Z b Z b Z t
[F g + Gf ] dt = g(t) F (a) + f (s) ds dt
a a a
Z b Z t
+ f (t) G(a) + g(s) ds dt
a a
Z b Z t
= F (a)[G(b) − G(a)] + g(t) f (s) ds dt
a a
Z b Z t
+ G(a)[F (b) − F (a)] + f (t) g(s) ds dt.
a a
Set s = x, t = y in the second integral and s = y, t = x in the last integral.

Then,
Z b Z bZ y Z bZ x
[F g + Gf ] dt = f (x)g(y) dx dy + f (x)g(y) dy dx
a a a a a
+F (a)G(b) + G(a)F (b) − 2F (a)G(a).
Consider the square [a, b] × [a, b] in R2 and its bisection by the line y = x in
R2 . Then the first integral is evaluated in the region above the bisecting line
and the second integral is evaluated below the bisecting line. Thus, combined
together the region of computation is the square [a, b] × [a, b]. Hence,
Z b Z bZ b
[F g + Gf ] dt = f (x)g(y) dx dy + F (a)G(b) + G(a)F (b)
a a a
− 2F (a)G(a)
Z b Z b
= f (x) dx g(y) dy + F (a)G(b) + G(a)F (b)
a a
− 2F (a)G(a)
= [F (b) − F (a)][G(b) − G(a)] + F (a)G(b) + G(a)F (b)
− 2F (a)G(a)
= F (b)G(b) − F (a)G(a).
We know that F and G are bounded variation and F 0 = f and G0 = g.

Thus, we have the following consequence:
Corollary 1.5.4. If f, g ∈ AC(R) then, for each subinterval [a, b] ⊂ R,
Z b
[f (x)g 0 (x) + f 0 (x)g(x)] dx = f (b)g(b) − f (a)g(a).
a
In particular, if f ∈ AC(R) then, for all φ ∈ D(R),

Z Z
f (x)φ (x) = − f 0 (x)φ(x) dx.
0
R R
The above proposition shows that, in one dimensional case, the inte-
gration by parts formula can be extended, precisely, to the class of abso-
lutely continuous functions. Every absolutely continuous function on R is of
bounded variation and admits a derivative almost everywhere. We have the

following inclusions:
AC(R) ( C(R) ∩ BV (R) ( BV (R) ( L1loc (R) ( R(R) ( D0 (R).
Rb
We know that f ∈ AC(R) iff f (b) − f (a) = a f 0 (t) dt for all subintervals
[a, b] ⊂ R.
Proposition 1.5.5. Conversely, if f : R → R is an integrable function such
that Z Z
0
f (x)φ (x) dx = − g(x)φ(x) dx,
R R
for some integrable function g : R → R and for all φ ∈ D(R), then there is a
h ∈ AC(R) such that h = f a.e. and h0 = g.
Remark 1.5.6. Every f ∈ AC(R) can be associated with a measure µf ,
defined as µf (a, b) = f (b) − f (a), for every subinterval (a, b) ⊂ R. Note that
this measure is different from the measure µf 0 induced by f 0 ∈ L1loc (R) as in
(1.4.2). Observe that Tf 6= Tµf . In fact, Tµf = Tf 0 .
We have already seen in Example 1.28 that for the Cantor function fC ,
which is not absolutely continuous, the two derivatives do not coincide.
Exercise 30. Compute and compare the classical and distributional derivative
of (
x2 sin x12 x 6= 0
f (x) =
0 x = 0.
Exercise 31. Compute the distributional derivative of |u|.
Proof. Consider (√
t2 2 − if t > 0
G (t) =
0 if t ≤ 0.
Notice that for every test function ϕ ∈ Cc∞ (Ω) we have
Z Z
u(x)Du(x)
G (u(x))Dϕ(x) dx = − p ϕ(x) dx.
Ω Ω∩{u>0} u(x)2 2
From this we get that
Z Z Z
+
u (x)Dϕ(x) dx = lim G (u(x))Dϕ(x) dx = − Du(x)ϕ(x) dx,
Ω →0 Ω Ω∩{u>0}
so that Du+ = χ{u>0} Du in the sense of distributions. Similarly Du− =

−χ{u<0} Du. Notice that if {u = 0} has non zero measure we have Du = 0
a.e. on this set. Thus,
Z Z Z
|u|(x)Dϕ(x) dx = u (x)Dϕ(x) dx + u− (x)Dϕ(x) dx
+
Ω ΩZ ΩZ
= − Du+ (x)ϕ(x) dx − Du− (x)ϕ(x) dx

ZΩ Ω
= − (Du+ (x) + Du− (x))ϕ(x) dx.

Ω
This shows that |u| has a weak gradient and the formula is obtained.
Theorem 1.5.7. The differential operator Dα , for each α, is a continuous

map from D0 (Ω) to D0 (Ω).
Proof. It is enough to show that when Tm → T in D0 (Ω) then Dα Tm → Dα T

in D0 (Ω), for all multi-index α. Consider
lim Dα Tm (φ) = lim (−1)α Tm (Dα φ) = (−1)α T (Dα φ) = Dα T (φ).

m→∞ m→∞
Example 1.31. The above theorem is very special to the space of distributions
D0 (Ω). For ∞
√ instance, the result is not true in C (Ω). Consider the sequence
fm = (1/ m) sin mx that converges uniformly to 0 and hence converges√ to
0
0 in the distribution sense too. However, it’s derivative fm (x) = m cos mx
does not converge pointwise but converges to 0 in the distribution sense.
(Term-by term differentiation of series). If S := ∞

P
Corollary 1.5.8 i=1 Ti ,
then Dα S = ∞ α
P
i=1 D Ti .
1.5.2 Product
Now that we have addition of distributions, a natural question is whether one
can define product of any two distributions. The answer is in negation. The
space of distribution cannot be made an algebra which extends the classical
notion of point-wise multiplication. This is called the Schwartz impossibility
result. L. Schwartz himself showed that it is not possible to define product
of Dirac distributions, i.e., δ0 · δ0 . However, one may define product of a
distribution with a C ∞ function which is a generalisation of the point-wise

multiplication of two functions. To motivate this definition, we note that if
f ∈ L1loc (Ω) and ψ ∈ C ∞ (Ω) then
Z Z
[ψ(x)f (x)]φ(x) dx = f (x)[ψ(x)φ(x)] dx.
Ω Ω
If Tf is the distribution induced by f , then above equality is same as saying

Tψf (φ) = Tf (ψφ). But Tf (ψφ) makes sense only when ψφ ∈ D(Ω). This is
the reason we demand ψ ∈ C ∞ (Ω) because then supp(ψφ)⊆ supp(φ) and ψφ
has compact support. Also, by Leibniz’ rule,
X α!
Dα (ψφ) = Dβ ψ Dα−β φ
β≤α
β!(α − β)!
and hence ψφ ∈ D(Ω).

Definition 1.5.9. Let ψ ∈ C ∞ (Ω) and T ∈ D0 (Ω), then we define the func-
tional ψT : D(Ω) → R as ψT (φ) = T (ψφ).
Proposition 1.5.10. For every ψ ∈ C ∞ (Ω), ψT ∈ D0 (Ω).
Proof. We need to show that the functional ψT is continuous on D(Ω). For
∞
any φ ∈ CK (Ω),
NK
X
|ψT (φ)| = |T (ψφ)| ≤ CK kψφkNK = CK kDα (ψφ)k0
|α|=0
NK
X X α!
≤ CK kDβ ψk0 kDα−β φk0
β!(α − β)!
|α|=0 β≤α
≤ CK C0 kφkNK .
If T is a distribution of order k, then it is enough to demand ψ ∈ C k (Ω).

Example 1.32. Let ψ(x) = x on R and T = δa ∈ D0 (R), then xδa ∈ D0 (R)
and xδa (φ) = δa (xφ) = aφ(a).
Theorem 1.5.11 (Leibniz’ formula). If ψ ∈ C ∞ (Ω), then
X α!
Dα (ψT ) = Dβ ψ Dα−β T.
β≤α
β!(α − β)!
Exercise 32. Let φ ∈ D(R). Show that φ(k) (0) = 0 for k = 0, 1, 2, . . . , m − 1

iff there is a ψ ∈ D(R) such that φ(x) = xm ψ(x).
Proof. Let ψ ∈ D(R) be such that φ(x) = xm ψ(x). Then
k
(k)
X k! di (xm ) (k−i)
φ (x) = ψ
i=0
i!(k − i)! dxi
di (xm ) m!
where dxi
= (m−i)!
xm−i . Thus, for all i = 0, 1, . . . , m − 1, the value of
di (xm ) dm (xm )
dxi
at x = 0 is 0 and = m! 6= 0. Therefore, φ(k) (0) = 0 for
dxm
all k = 0, 1, 2, . . . , m − 1, irrespective of the value of ψ(0). Conversely, let
φ(k) (0) = 0 for all k = 0, 1, 2, . . . , m − 1. Let k = 0, then
Z x
φ(x) = φ(0) + φ(1) (t) dt.
0
Since φ(0) = 0 and setting t = xs, we get

Z x Z 1
(1)
φ(x) = φ (t) dt = x φ(1) (xs) ds = xψ0 (x),
0 0
R1 (1)
where ψ0 (x) := 0 φ(1) (xs) ds is in D(R). Since φ(1) (x) = xψ0 (x) + ψ0 (x),
we have ψ(0) = 0 and hence ψ0 (x) = xψ1 (x) for some ψ1 ∈ D(R). Thus,
φ(x) = x2 ψ1 (x). Using the fact that φ(2) (0) = 0, we get ψ1 (0) = 0 and
hence φ(x) = x3 ψ2 (x) for some ψ2 ∈ D(R). Proceeding this way, using
φ(m−1) (0) = 0, we get ψm−1 ∈ D(R) such that φ(x) = xm ψm−1 (x).
Exercise 33. (i) Find all real valued functions f on R which is a solution
to the equation xf (x) = 0 for all x ∈ R.
(ii) Find all distributions T ∈ D0 (R) that solve the equation xT = 0 in the
distribution sense.
(iii) Find all distributions T ∈ D0 (R) that solve the equation x2013 T = 0 in
the distribution sense.
Proof. (i) Consider the function f : R → R such that
(
0 x 6= 0
f (x) =
1 x = 0.
Then f solves xf (x) = 0 for all x ∈ R. Moreover, for any scalar λ ∈ R,
the function λf is also a solution of xf (x) = 0.
(ii) For any φ ∈ D(R), since 0 = xT (φ) = T (xφ). By Exercise 32, any
solution T ∈ D0 (R) satisfies T (ψ) = 0 for all ψ ∈ D(R) such that
ψ(0) = 0. Obviously, T ≡ δ0 is a solution. In fact, for any scalar
λδ0 ∈ D0 (R) is also a solution. We need to show that any solution is a
scalar multiple of δ0 . Let T ∈ D0 (R) be a solution of xT = 0. Hence
T (ψ) = 0 for all test functions on R such that ψ(0) = 0. Let χ ∈ D(R)
such that χ(0) = 1, then for any φ ∈ D(R), ψ(x) := φ(x) − φ(0)χ(x) is
a test function such that ψ(0) = 0. Therefore,
0 = T (ψ) = T (φ) − T (χ)φ(0) = T (φ) − T (χ)δ0 (φ).
Thus, T (φ) = λδ0 (φ) where the choice of λ depends on the choice of χ.
(iii) By Exercise 32, any solution T ∈ D0 (R) satisfies T (ψ) = 0, for all
ψ ∈ D(R), such that ψ (k) (0) = 0 for all k = 0, 1, 2, . . . , 2012. Therefore,
(k)
δ0 is a solution for all k = 0, 1, 2, . . . , 2012. Thus,
2012
X (k)
T ≡ λk δ0
k=0
is a solution for every possible choice of λk ∈ R.
Example 1.33. Let ψ(x) = x on R and T = P V (1/x) ∈ D0 (R), then

xP V (1/x) ∈ D0 (R) and
Z
1 1 1
xP V (φ) = P V (xφ) = lim+ xφ(x) dx
x x ε→0 |x|≥ε x
Z
= φ(x) dx = T1 (φ)
R
where T1 is the distribution corresponding to the constant function 1.

Exercise 34. Find all distributions T ∈ D0 (R) that solve the equation xT = 1
in the distribution sense.
Proof. The principal value distribution T ≡ P V (1/x) is one particular so-
lution of xT = 1. Also any solution S of xS = 0 when added to P V (1/x)
is also a solution. Thus, the general solution is P V (1/x) + λδ0 for every
λ ∈ R.
In light of the above example, we come back to the issue of not being
able to have a product in the space of distributions. Suppose we give a
notion of product of distributions, the product cannot be associative because
P V (1/x) · (x · δ0 ) = 0 6= δ0 = (P V (1/x) · x) · δ0 . Further, recall that
f (x) = |x|−1/2 ∈ L1loc (R), but g(x) = |x| 6∈ L1loc (R). How do we give a
notion of product, in a suitable way, such that (Tf )2 coincides with classical
pointwise product. The failure to define a suitable notion of product is
what makes the theory of distributions unsuitable for nonlinear differential
equations.
Exercise 35. For any given ψ ∈ C ∞ (R), find the solutions of the equation
DT + ψT = 0 in D0 (R).
Proof. If ψ ≡ 0, we know from Exercise 28 that T is a regular distribution

induced by a constant function. Therefore, we assume w.l.o.g that ψ 6≡ 0. We
copy the idea of integrating factor from ODE course. We shall now prove the
existence of Ψ ∈ C ∞ (R) (nowhere zero) such that D(ΨT ) = Ψ(DT + ψT ).
Multiply Ψ on both sides of the given equation to get ΨDT + ΨψT = 0.
If we wantRx
Ψ such that D(ΨT ) = Ψ(DT + ψT ), then Ψψ = Ψ0 . Thus,
Ψ(x) = e −∞ ψ(s) ds . By construction, Ψ is a non-zero function and the equa-
tion implies that D(ΨT ) = 0. By Exercise 28, for each scalar µ ∈ R, ΨT = Tµ
is a solution. Thus, T = Ψ−1 Tµ , for each µ ∈ R is the required solution of
the equation. The solution makes sense because Ψ−1 ∈ C ∞ (R).
Exercise 36. For any given ψ ∈ C ∞ (R) and S ∈ D0 (R), find the solutions of
the equation DT + ψT = S in D0 (R).
Recall the topology on C ∞ (Ω) described in §1.3.3. The sequential de-
scription of the topology is the uniform convergence on compact sets. Thus,
a sequence ψm is said to converge to 0 in C ∞ (Ω) if Dα ψm , for all α, uniformly
converges to 0 on all compact subsets of Ω.
Lemma 1.5.12. If ψm → 0 in C ∞ (Ω) then ψm φ → 0 in D(Ω).
Proof. Since Dα ψm → 0 uniformly on all compact subsets of Ω, in partic-

ular on K = supp(φ). Thus, Dα ψm φ → 0 uniformly on K, for all α, and
supp(ψm φ) ⊆ K, for all m.
Theorem 1.5.13. If Tm → T in D0 (Ω) and ψm → ψ in C ∞ (Ω) then

ψm Tm → ψT in D0 (Ω).
1.5.3 Support
The fact that we cannot have a product on D0 (Ω), generalising point-wise
multiplication, that makes D0 (Ω) an algebra leads us to look for other ways
to make D0 (Ω) an algebra. One such choice is the convolution operation and
whether this can be extended to distributions. We shall see later than one
can define a notion of convolution for distributions provided one of them
has compact support! This motivates us to understand the support of a
distribution which coincides for classical functions. Classically, support of a
function is complement of the largest open set on which the function vanishes.
Definition 1.5.14 (Localisation). The restriction of T ∈ D0 (Ω) to an open

subset ω ⊂ Ω, denoted as T |ω , is defined as
T |ω (φ) = T (φ) ∀φ ∈ D(ω) ⊂ D(Ω).
We say a distribution T vanishes on an open set ω ⊂ Ω if T |ω = 0. More

generally, two distributions S and T coincide on a open set ω if S |ω = T |ω .
The inclusion D(ω) ⊂ D(Ω) is continuous because any φ ∈ D(ω) can be

extended by zero outside ω which belongs to D(Ω). Since T |ω ∈ D0 (ω), T
vanishing on ω is same as the zero distribution in D0 (ω).
Example 1.34. The restriction of the Dirac distribution δa to any open set
ω ⊂ Ω \ {a} is zero.
Proposition 1.5.15. If T ∈ D0 (Ω) vanishes on the open subsets ωi of Ω,

then T vanishes on the union ∪i∈I ωi .
Proof. Let ω = ∪i∈I ωi . We choose the C ∞ locally Pfinite partition of unity

{φi } subordinate to {ωi }, i.e., supp(φi ) ⊂ ωi and i φ = 1 (the summation
is finite for each x ∈ ω). Fix φ ∈ D(ω) and K = supp(φ) then Pk there exists
finitely many i1 , i2 , . . . , ik such that K ∩ φim 6= ∅. Thus, φ = m=1 φφim and,
by linearity of T , T (φ) = km=1 T (φφim ). But supp(φφim ) ⊂ ωim and hence
P
T (φφim ) = 0. Therefore, T (φ) = 0 for all φ ∈ D(ω) and hence T |ω = 0.
Definition 1.5.16. The support of a distribution T is the complement of

the largest open set ω such that T |ω = 0. Equivalently, the support of a
distribution is the relative complement in Ω of the union of all open sets ω
of Ω such that T |ω = 0.
If E is the support of a distribution T , then T (φ) = 0 for all φ ∈ D(Ω)

such that E ∩ supp(φ) = ∅ because all such φ is in D(E c ). Obviously, the
support of the zero distribution is the empty set.
Exercise 37. Show that for any continuous function f , supp(Tf )=supp(f ).
More generally, for any f ∈ L1loc (Ω), there exists a measurable function g
such that g = f a.e. and supp(Tf ) = supp(g).
Example 1.35. The support of the Dirac distribution δa is the singleton set
{a}. Similarly, the support of the derivatives of Dirac distribution δaα is also
the singleton set {a}. The open set Ω \ {a} = ∪i Bi , is the union of punctured
open balls of rational radius ri with centre {a} removed. δaα vanishes on Bi
for each i and hence vanishes on Ω \ {a}.
Exercise 38. Show that the support of DT ∈ D(R) is contained in the support
of T ∈ D(R).
Proof. Let E and F denote the support of T and DT , respectively. We
need to show that F ⊆ E, or equivalently, E c ⊆ F c . For any open set
ω ⊂ E c , then T |ω = 0, i.e., for all φ ∈ D(ω) we have T (φ) = 0. Consider
DT (φ) = −T (φ0 ) = 0. The second equality is due to the fact that if φ ∈ D(ω)
then φ0 ∈ D(ω) with supp(φ0 ) ⊆ supp(φ).
A distribution is said to have compact support, if its support is a compact
subset of Ω. A nice characterisation of distributions with compact support,
which is a subclass of D0 (Ω), is that it can be identified as a dual of C ∞ (Ω)
endowed with the topology of uniform convergence on compact sets. This
observation is actually quite natural. Recall the motivation for the choice
of Cc∞ (Ω) as a test function. The compact support of test function was
necessary to kill the boundary evaluation while applying integration by parts.
If the function f itself is compact, to begin with, then we can expand the
space of test functions to C ∞ (Ω). For convenience sake (for ease of notation),
we set C ∞ (Ω) = E(Ω), the space of infinitely differentiable functions on Ω
with the topology of uniform convergence on compact subsets. The inclusion
D(Ω) ⊂ E(Ω) is continuous w.r.t the respective topology. This is because
if φm → 0 in D(Ω) then φm → 0 in E(Ω), as well. The following theorem
asserts that any distribution T : D(Ω) → R can be continuously extended as
a functional to E(Ω). Thus T ∈ E 0 (Ω), for compact support T .
Theorem 1.5.17. The dual space E 0 (Ω) is the collection of all distributions
with compact support.
Proof. Let T ∈ D0 (Ω) such that support of T is compact. Set K = supp(T ).

Thus, T (φ) = 0 for all φ ∈ D(Ω) such that supp(φ) ⊂ K c . Choose χ ∈ D(Ω)
such that χ ≡ 1 on K. For all φ ∈ D(Ω), consider
T (φ) − χT (φ) = T (φ − χφ) = T ((1 − χ)φ).
Since 1 − χ = 0 on K, supp(1 − χ)φ ⊂ K c and hence T = χT for any
χ such that χ ≡ 1 on K. This fact helps us to extend T as a functional
on E(Ω), because for all ψ ∈ E(Ω), ψχ ∈ D(Ω). We extend T to E(Ω) by
setting T (ψ) = T (χψ). The assignment is well-defined (it is independent
of choice of χ). Suppose χ1 ∈ D(Ω) is such that χ1 6= χ and χ1 ≡ 1
on K, then (χ − χ1 )ψ = 0 on K. Therefore, supp((χ − χ1 )ψ) ⊂ K c and
T ((χ − χ1 )ψ) = 0, By linearity of T , T (χψ) = T (χ1 ψ). The conitnuity of T
on E(Ω) is a consequence of Lemma 1.5.12. Thus, T ∈ E 0 (Ω).
Conversely, let T ∈ E 0 (Ω). We need to show that T ∈ D0 (Ω) and has a
compact support. The fact that T ∈ D0 (Ω) is obvious due to the continuous
inclusion of D(Ω) in E(Ω). Let {Ki } be the collection of compact subsets of
Ω such that Ki ⊂ Ki+1 and Ω = ∪∞ i=1 Ki (exhaustion subsets). Suppose that
the support of T is not compact in Ω, then the support of T intersects Ω \ Ki ,
for all i. Thus, for each i, there is a φi ∈ D(Ω) such that supp(φi ) ⊂ Ω \ Ki
and T (φi ) 6= 0. Say T (φi ) = αi 6= 0. Then T (αi−1 φi ) = 1. Set ψi = αi−1 φi .
Thus, supp(ψi ) ⊂ Ω \ Ki and T (ψi ) = 1. Hence, the sequence ψi → 0 in
E(Ω) because their support tend to empty set. Now, by continuity of T on
E(Ω), T (ψi ) → 0 which contradicts the fact that T (ψi ) → 1. Hence, T has a
compact support.
We have proved that E 0 (Ω) ⊂ D0 (Ω) is the class of all distributions whose
support is compact. Distributions with compact support are always of finite
order.
Proposition 1.5.18. If T ∈ E 0 (Ω) then order of T is finite.
Proof. Since T ∈ E 0 (Ω), T has compact support, say K. Let χ ∈ D(Ω) be
such that χ ≡ 1 on K, then we know that (cf. proof of last theorem) T = χT .
Let K 0 be the support of χ and hence K ⊂ K 0 . Now, for any compact subset
∞
K ⊂ Ω and φ ∈ CK (Ω), supp(φχ) ⊂ K 0 . Consider,
|T (φ)| = |χT (φ)| = |T (χφ)| ≤ CK 0 kχφkNK 0 ≤ C0 kφkNK 0 .
The NK 0 is independent of all compact subsets K. Thus, T has finite order.
1.5.4 Singular Support

The singular support of a distribution is the measure of the set on which the
distribution fails to be smooth or C ∞ .
Definition 1.5.19. A distribution T ∈ D0 (Ω) is said to be C ∞ on ω ⊂ Ω if
there is a f ∈ C ∞ (ω) such that T = Tf on D(ω).
Example 1.36. The Dirac distribution, δa , is C ∞ in Ω \ {a} because δa re-
stricted to the open set Ω \ {a} is the zero function.
Local behaviour of “compatible” distributions can be patched up to get
a global description of the distribution. This concept will facilitate the defi-
nition of singular support.
Theorem 1.5.20. Let Ti ∈ D0 (ωi ), an arbitrary collection of distributions,
such that Ti |ωi ∩ωj = Tj |ωi ∩ωj then there is a unique distribution T ∈ D0 (Ω),
where Ω = ∪i ωi , such that T |ωi = Ti for all i.
Proof. Let {φi } ⊂ D(Ω) be a C ∞ locally finite
P partition of unity subordinate
to {ωi }. Thus, supp(φi ) ⊂ ωP i and 1 = i φi . We define the functional
T : D(Ω) → R as T (φ) = T
i i (φφi ). We first check the continuity of
T . One way is to show the sequential continuity. Alternately, For any φ ∈
∞
CK (Ω) (compact subset of Ω), there exist finitely many i1 , i2 , . . . , ik such
that K ∩ supp(φim ) 6= ∅. Then
k
X X
|T (φ)| ≤ |Tim (φφim )| ≤ Ck kφφim kNk ≤ C0 kφkN0 .
m=1 m
It only remains to show that the restriction of T to ωk is Tk . Let φ ∈ D(ωk ),

then φφi ∈ D(ωi ∩ ωk ) for all i. Thus,
X X X
T (φ) = Ti (φφi ) = Tk (φφi ) = Tk ( φφi ) = Tk (φ).
i i i
The uniqueness of T follows from Proposition 1.5.15.

Corollary 1.5.21. If T ∈ D0 (Ω) is C ∞ on an arbitrary collection of open
subsets ωi , then T is C ∞ on the union ∪i∈I ωi .
A consequence of the above corollary is the following definition.
Definition 1.5.22. The singular support of T ∈ D0 (Ω) is the complement
of the largest open set on which T is C ∞ , denoted as sing.supp(T ).
A simple observation is that sing.supp(T ) ⊂ supp(T ).
1.5.5 Shifting and Scaling

Given any function f ∈ L1loc (Rn ) and a fixed a ∈ Rn , we introduce the shift
operator τa f (x) := f (x − a). In particular, τ0 f = f for all f . Observe that
τa τb f (x) = τa f (x − b) = f (x − b − a) = f (x − (b + a)) = τa+b f (x).
Note that if f ∈ L1loc (Rn ) then τa f ∈ L1loc (Rn ). We have the following
relation between Tf and Tτa f . For any φ ∈ D(Rn ), consider
Z Z Z
Tτa f (φ) = τa f (x)φ(x) dx = f (x − a)φ(x) dx = f (y)φ(y + a) dy
R n Rn Rn
Z
= f (y)τ−a φ(y) dy = Tf (τ−a φ).
Rn
The last equality is valid because τ−a φ ∈ D(Rn ) whenever φ ∈ D(Rn ). We
wish to extend the notion of shift to distributions such that for τa Tf = Tτa f .
This motivates the following definition of a shift of a distribution by a ∈ Rn .
Definition 1.5.23. For any T ∈ D0 (Rn ), we define its shift by a ∈ Rn as
τa T (φ) = T (τ−a φ).
Given any function f ∈ L1loc (Rn ) and a fixed scalar λ ∈ (0, ∞), we intro-
duce the scaling function fλ (x) := f (λx). If f = g + h then fλ (x) = f (λx) =
g(λx) + h(λx) = gλ (x) + hλ (x). Also, τa (f + g) = τa f + τa g. Note that if
f ∈ L1loc (Rn ) then fλ ∈ L1loc (Rn ). We have the following relation between Tf
and Tfλ . For any φ ∈ D(Rn ), consider
Z Z y
−n
Tfλ (φ) = f (λx)φ(x) dx = λ f (y)φ dy
Rn Rn λ
Z
= λ−n f (y)φ1/λ (y) dy = λ−n Tf (φ1/λ ).
Rn
The last equality is valid because φ1/λ ∈ D(Rn ) whenever φ ∈ D(Rn ). Simi-
larly, one can argue for λ ∈ (−∞, 0) and deduce that
Tfλ (φ) = (−1)n λ−n Tf (φ1/λ ).
Thus for any λ ∈ R \ {0}, we have Tfλ (φ) = |λ|−n Tf (φ1/λ ). We wish to
extend the notion of scaling to distributions such that for (Tf )λ = Tfλ . This
motivates the following definition of a scaling of a distribution by λ ∈ R\{0}.
Definition 1.5.24. For any T ∈ D0 (Rn ), we define its scaling by λ ∈ R\{0}
as Tλ (φ) = |λ|−n T (φ1/λ ). In particular, when λ = −1, we denote T−1 as Ť
and hence Ť (φ) = T (φ̌). We say T ∈ D0 (Rn ) is even if Ť = T .
Exercise 39. Show that Ťˇ = T .
1.5.6 Convolution
Recall the definition of convolution of functions given in Definition 1.3.20.
Thus, for any x ∈ Rn ,
Z
(f ∗ g)(x) := f (x − y)g(y) dy.
Rn
In this section, we wish to extend the convolution operation to distributions

such that for L1 distributions the notion coincide. An obvious extension is
visible when the convolution integral is rewritten using the shift and scaling
operator. Note that, for a fixed x ∈ Rn , f (x − y) = f (−(y − x)) = fˇ(y − x) =
τx fˇ(y). With this notation the convolution is rewritten as
Z Z
(f ∗ g)(x) = τx fˇ(y)g(y) dy = f (y)τx ǧ(y) dy.
Rn Rn
The second equality is due to the commutativity of convolution operation.

The reformulation above motivates the following definition of convolution
between a distribution and a test function.
Definition 1.5.25. Let T ∈ D0 (Rn ) and φ ∈ D(Rn ). The convolution T ∗ φ :
Rn → R is defined as
(T ∗ φ)(x) = T (τx φ̌).
The above definition is meaningful because τx φ̌ ∈ D(Rn ), for all x ∈ Rn ,
whenever φ ∈ D(Rn ). In particular, (T ∗ φ̌)(x) = T (τx φ) since φ̌ˇ = φ. Thus,
T (φ) = (T ∗ φ̌)(0). As a consequence, if T ∗ φ = 0 for all φ ∈ D(Rn ) then
T = 0 in D0 (Rn ).
Exercise 40. (i) Show that T ∗ (φ1 + φ2 ) = T ∗ φ1 + T ∗ φ2 .
(ii) Show that (S + T ) ∗ φ = S ∗ φ + T ∗ φ.

Proof. (i) Consider
T ∗(φ1 +φ2 )(x) = T (τx (φ1 +φ2 )ˇ) = T (τx φˇ1 +τx φˇ2 ) = (T ∗φ1 +T ∗φ2 )(x).
(ii) Consider
(S + T ) ∗ φ = (S + T )(τx φ̌) = S ∗ φ + T ∗ φ.
Exercise 41. Let T ∈ D0 (Rn ) and φ ∈ D(Rn ), then
(i) for any a ∈ Rn , show that τa (T ∗ φ) = τa T ∗ φ = T ∗ τa φ.
(ii) show that, for any multi-index α, Dα (T ∗ φ) = Dα T ∗ φ = T ∗ Dα φ. In

particular, T ∗ φ ∈ E(Rn ).
(iii) for χ ∈ D(Rn ), T ∗ (φ ∗ χ) = (T ∗ φ) ∗ χ.
For any ψ ∈ E(Rn ), τx ψ̌ ∈ E(Rn ). Thus, the convolution notion can be

extended to any ψ ∈ E(Rn ) and T ∈ E 0 (Rn ).
Definition 1.5.26. Let T ∈ E 0 (Rn ) and ψ ∈ E(Rn ). The convolution T ∗ ψ :

Rn → R is defined as
(T ∗ ψ)(x) = T (τx ψ̌).
Exercise 42. Let T ∈ E 0 (Rn ) and ψ ∈ E(Rn ), then
(i) for any a ∈ Rn , show that τa (T ∗ ψ) = τa T ∗ ψ = T ∗ τa ψ.
(ii) show that, for any multi-index α, Dα (T ∗ ψ) = Dα T ∗ ψ = T ∗ Dα ψ. In

particular, T ∗ φ ∈ E(Rn ). Further, if φ ∈ D(Rn ) then T ∗ φ ∈ D(Rn )
and
T ∗ (ψ ∗ φ) = (T ∗ ψ) ∗ φ = (T ∗ φ) ∗ ψ.
Definition 1.5.27. Let S, T ∈ D0 (Rn ) such that one of them has compact
support, i.e. either T or S is in E 0 (Rn ). We define the convolution S ∗ T as,
(S ∗ T )(φ) = (S ∗ (T ∗ φ̌))(0), ∀φ ∈ D(Rn ).
Equivalently, one can also define the convolution S ∗ T to be
(S ∗ T ) ∗ φ = S ∗ (T ∗ φ) ∀φ ∈ D(Rn ).
Exercise 43. Let S, T ∈ D0 (Rn ) such that one of them is in E 0 (Rn ).
(i) Show that T ∗ S = S ∗ T .
(ii) supp(S ∗ T ) ⊂ supp(S) + supp(T ).
(iii) For any multi-index α, Dα (T ∗ S) = Dα T ∗ S = T ∗ Dα S.

Exercise 44. For any three distributions T1 , T2 , T3 ∈ D0 (Rn ) such that at least
two of them have compact support then
T1 ∗ (T2 ∗ T3 ) = (T1 ∗ T2 ) ∗ T3 .
Theorem 1.5.28. Let T ∈ D0 (Rn ), then T = T ∗ δ0 = δ0 ∗ T . Also, for any

multi-index α, Dα T = Dα δ0 ∗ T .
Proof. Since δ0 has compact support {0}, the convolution makes sense. By
commutativity, T ∗ δ0 = δ0 ∗ T . For φ ∈ D(Rn ), consider (T ∗ δ0 )(φ) =
(T ∗ (δ0 ∗ φ̌))(0). Set ψ := δ0 ∗ φ̌. Then (T ∗ δ0 )(φ) = T (ψ̌). Note that
ψ̌(x) = ψ(−x) = (δ0 ∗ φ̌)(−x) = δ0 (τ−x φ) = τ−x φ(0) = φ(x).
Hence, (T ∗ δ0 )(φ) = T (φ).
Theorem 1.5.29. D(Ω) is dense in D0 (Ω).
The following exercise is same as Exercise 28. However, using convolution

one can give a simpler proof to Exercise 28.
Exercise 45. Let T ∈ D0 (R) be such that DT = 0, then show that T is a
regular distribution generated by a constant function.
Proof. For a given sequence of mollifiers, let Tε (x) = (T ∗ ρε )(x). Thus,
Tε ∈ E(R) and DTε = DT ∗ ρε = 0. Hence, Tε = λε , where λε is a constant
0
R each ε. Also, λε * T in D (R). In particular, choose φ ∈ D(R)
function for
such that φ = 1, then the sequence of real numbers λε converges to some
λ. Thus, T = Tλ .
Exercise 46. Let T denote the distribution corresponding to the constant
(1)
function 1, δ0 be the dipole distribution at 0 and H be the Heaviside function
(
1 x>0
H(x) =
0 x ≤ 0.
(1) (1)
Show that (T ∗ δ0 ) ∗ H 6= T ∗ (δ0 ∗ H).
Proof. Consider
Z
(1) (1)
(1 ∗ δ0 )(φ) =1∗ (δ0 ∗ φ̌)(0) = (1 ∗ ψ)(0) = T1 (ψ̌) = ψ̌(x) dx.
R
(1) (1)
But ψ̌(x) = δ0 ∗ φ̌(−x) = δ0 (τ−x φ) = −τ−x φ0 (0) = −φ0 (x). Hence (1 ∗
(1)
δ0 )(φ) = − R φ0 (x) dx = 0. Further, 0 ∗ H = 0. Now, consider
R
(1) (1) (1) (1)

(δ0 ∗ H)(φ) = δ0 ∗ (H ∗ φ̌)(0) = (δ0 ∗ ψ)(0) = δ0 (ψ̌) = (ψ̌)0 (0) = (ψ̌)0 (0).
But (ψ̌)0 (0) = (H 0 ∗ φ̌)(0) = (δ0 ∗ φ̌)(0) = δ0 (φ) = φ(0). Since δ0 is the
identity distribution under convolution operation, 1 ∗ δ0 = 1.
(1)
Alternately, notice that by derivative of convolution (1 ∗ δ0 ) ∗ H = (D1 ∗
(1)
δ0 ) ∗ H = D1 ∗ H = 0 ∗ H = 0. On the other hand, 1 ∗ (δ0 ∗ H) =
1 ∗ (δ0 ∗ DH) = 1 ∗ (δ0 ∗ δ0 ) = 1 ∗ δ0 = 1.
1.6 Tempered Distributions

Recall the definition of Schwartz space from Definition ??. By definition,
D(Rn ) ⊂ S(Rn ). The inclusion is also dense because, for any f ∈ S(Rn ),
φm f → f in S where φm (x) = φ(x/m) and φ ∈ D(Rn ) is such that φ ≡ 1 on
the unit ball in Rn . Hence the dual of S, denoted S 0 , can be identified with
a subspace of D0 (Rn ).
Definition 1.6.1. The space of distributions S 0 is called the space of tem-
pered distributions.
Example 1.37. Any distribution with compact support is tempered, i.e., E 0 ⊂
S 0 because D ⊂ S ⊂ E and all inclusions are dense.
Example 1.38. Any slowly increasing measure induces a tempered distribu-
tion. A measure µ on Rn is said to be slowly increasing if, for some integer
k ≥ 0, Z
dµ(x)
2 k
< +∞.
Rn (1 + |x| )
For instance, any bounded measure is slowly increasing. Set
Z
Tµ (f ) = f (x) dµ(x), f ∈ S(Rn ).
Rn
Then Tµ is a linear functional on S(Rn ) and

Z
2 k dµ(x)
|Tµ (f )| ≤ sup |f (x)(1 + |x| ) | .
x∈Rn Rn (1 + |x|2 )k
Further, it follows that if fm → 0 in S(Rn ) then Tµ (fm ) → 0. Thus, Tµ ∈
S 0 (Rn ).
Example 1.39. If 1 ≤ p ≤ ∞ then Lp (Rn ) ⊂ S 0 (Rn ). If f ∈ Lp (Rn ), define

Z
Tf (φ) = f φ dx ∀φ ∈ S(Rn ).
Rn
If 1 < p < ∞, let q be the conjugate exponent of p given as (1/p)+(1/q) = 1.

n
Then for k > 2q , g(x) = (1+|x|2 )−k ∈ Lq (Rn ). Hence, by Hölder’s inequality,

2 k
|Tf (φ)| ≤ sup |(1 + |x| ) φ(x)| kf kLp (Rn ) kgkLq (Rn ) .
x∈Rn
Thus, Tf is tempered, and f 7→ Tf is continuous on Lp (Rn ). If p = 1,
|Tf (φ)| ≤ kφk∞,Rn kf k1,Rn
and if p = ∞,
|Tf (φ)| < sup (1 + |x|2 )k φ(x) kf k∞ Rn kgk1,(Rn )

x∈Rn
where g is as above with k > n2 .
Definition 1.6.2. Let T ∈ S 0 (Rn ). The Fourier transform of T , denoted as

T̂ , is defined as T̂ (f ) = T (fˆ) for all f ∈ S(Rn ).
Since f 7→ fˆ is continuous on S, it follows that T̂ ∈ S 0 .
Remark 1.6.3. Since S(Rn ) ⊂ L1 (Rn ) ⊂ S 0 (Rn ). It seems that, a priori,

there are two definitions of the Fourier transform on S(Rn ), each inherited
from L1 (Rn ) and S 0 (Rn ), respectively. But they are one and the same because
if f ∈ S(Rn ) then, for any g ∈ S(Rn ), by Parseval relation,
Z Z
ˆ
Tf (g) = f ĝ = fˆg = Tfˆ(g).
Rn Rn
Hence Tˆf = Tfˆ and, hence, both definitions of the Fourier transform coincide
on S(Rn ).
Recall that F : S(Rn ) → S(Rn ), defined as F(f ) = fˆ, is an isomorphism

on S(Rn ), it extends as a bijection to S 0 (Rn ), as well.
Remark 1.6.4. Since S(Rn ) ⊂ L2 (Rn ) ⊂ S 0 (Rn ), then there are two defi-
nitions of Fourier transform on L2 (Rn ) inherited from S 0 (Rn ) and the other
by extending from S(Rn ). It turns out that both are same. Consider the
weak-* convergence of tempered distributions, i.e., Tm → T weak-* sense if,
for every φ ∈ S(Rn ), Tm (φ) → T (φ). If Tm → T in S 0 (Rn ) then Tˆm → T̂ in
S 0 because
T̂m (φ) = Tm (φ̂) → T (φ̂) = T̂ (φ).
Let f ∈ L2 (Rn ) and let fk ∈ S(Rn ) such that fk → f in L2 (Rn ). Since the
inclusion of L2 (Rn ) ⊂ S 0 (Rn ) is continuous, fk → f in S 0 (Rn ), as well. Hence
fˆk → fˆ in S 0 (Rn ). On the other hand, F(fk ) → F(f ) in L2 (Rn ) and hence in
S 0 (Rn ). But as fk ∈ S(Rn ), we know that F(fk ) = fˆk . Now, by uniqueness
of the weak-* limit fˆ = F(f ).
By Proposition ??, for any polynomial P and mulit-index α, if T ∈ S 0 (Rn )

then P T ∈ S 0 (Rn ) and Dα T ∈ S 0 (Rn ).
We now extend the result of Theorem ?? to tempered distribution.
Theorem 1.6.5. Let T ∈ S 0 (Rn ) and let α be a multi-index. Then
Dα T̂ = (−ı)|α| (x
\ αT )
and
\
(D α T ) = (ı)|α| ξ α T̂ .
Proof. Let f ∈ S. Then
\
(x α T )(f ) = (xα T )(fˆ) = T (xα fˆ(x))
1 1
= |α|
T (Ddαf ) = T̂ (Dα f )
(ı) (ı)|α|
1
= |α|
Dα T̂ (f ).
(−ı)
Thus, the first relation is proved. The second relation can be proved similarly.
\
∂δ0
Example 1.40. δ̂0 = 1, ∂xk
(ξ) = ıξk and 1̂ = δ0 . Let φ ∈ S(Rn ). Then
Z
δˆ0 (φ) = δ0 (φ̂) = φ̂(0) = φ(x) dx.
Rn
Thus, δˆ0 = 1, i.e., the distribution induced by the constant function 1. Fur-
ther,
\
∂δ0
(ξ) = ıξk δˆ0 = ıξk .
∂xk
The 1̂ = δ0 follows by the Fourier inversion formula.
1.7 Fundamental Solution

Let L = |α|≤n cα Dα be a n-th order linear differential operator with con-
P
stant coefficients. Thus, any differential equation is of the form LT = S, in
the distribution sense for any given distribution S and T is the distribution
solution of the differential equation.
Definition 1.7.1. A distribution K is said to be a fundamental solution of

the operator L if L(K) = δ0 , where δ0 is the Dirac distribution at {0}.
The existence of fundamental solution of a linear differential operator for

constant coefficients is the famous result of Malgrange and Ehrenpreis (cf.
[Rud91]).
Suppose we wish to solve LT = S for a given distribution S, we first find
the fundamental solution LK = δ0 . Then K ∗ S (as long as the convolution
makes sense) is a solution of the equation L(·) = S, because L(K ∗ S) =
LK ∗ S = δ0 ∗ S = S. The fundamental solution is not unique because
any solution U of the homogeneous equation L(U ) = 0 can be added to
the fundamental solution K to obtain other fundamental solutions, since
L(U + K) = L(U ) + L(K) = δ0 .
Definition 1.7.2. A differential operator L = |α|≤n cα Dα with cα ∈ C ∞ is

P
said to be hypoelliptic whenever Lu ∈ C ∞ implies that u is C ∞ .
Theorem 1.7.3. For a differential operator with constant coefficients, the

following are equivalent:
(i) L is hypoelliptic.
(ii) Every fundamental solution of L is C ∞ on Rn \ {0}.
(iii) Some fundamental solution of L is C ∞ on Rn \ {0}.

1.8 Further Reading

Most of the concepts discussed in this chapter can be found in detail in
[Kes89, Rud91, Hor70, Tre70].
Chapter 2
Sobolev Spaces
In this chapter we develop a class of functional spaces which form the right
setting to study partial differential equations. Recall that Ω ⊂ Rn is open
and not necessarily bounded. The situations where Ω needs to be bounded
will be specified clearly. Before we describe Sobolev spaces in detail, we
introduce a classical functional space very useful in the theory of Sobolev
spaces.
2.1 Hölder Spaces

In this section, we introduce some class of function spaces in C k (Ω), for all
integers k ≥ 0, which can viewed in some sense as spaces of “fractional”
derivatives. For any γ ∈ (0, ∞) and x0 ∈ Ω, a function u : Ω → R satisfies
at x0 , for all x ∈ Ω, the estimate

|u(x) − u(x0 )| γ |u(x) − u(x0 )|
|u(x) − u(x0 )| = |x − x0 | ≤ sup |x − x0 |γ .
|x − x0 |γ x∈Ω |x − x 0 |γ
x6=x0
However, it is not always necessary that the supremum is finite. Note that
the modulus in the numerator and denominator are in R and Rn , respectively.
Definition 2.1.1. Let γ ∈ (0, 1]. We say a function u : Ω → R is Hölder
continuous of exponent γ at x0 ∈ Ω, if

|u(x) − u(x0 )|
pγ (u)(x0 ) := sup < +∞.
x∈Ω |x − x0 |γ
x6=x0
77
CHAPTER 2. SOBOLEV SPACES 78
Note that any Hölder continuous function at x0 ∈ Ω satisfies the estimate
|u(x) − u(x0 )| ≤ pγ (u)(x0 )|x − x0 |γ , for x ∈ Ω. (2.1.1)

The constant pγ (u)(x0 ) may depend on u, Ω, γ and x0 . It follows from
(2.1.1) that any Hölder continuous function at x0 , for any exponent γ, is also
h i1/γ
ε
continuous at x0 . Because, for every ε > 0, we can choose δ = pγ (u)(x0 ) .
The case when γ = 1 corresponds to the Lipshitz continuity of u at x0 .
Definition 2.1.2. Let γ ∈ (0, 1]. We say a function u : Ω → R is uniformly
Hölder continuous of exponent γ, if

|u(x) − u(y)|
pγ (u) := sup < +∞ (2.1.2)
x,y∈Ω, |x − y|γ
x6=y
and is denoted by C 0,γ (Ω). The quantity pγ (u) is called the γ-th Hölder
coefficient of u. If the Hölder coefficient is finite on every compact subsets of
Ω, then u is said to be locally Hölder continuous with exponent γ, denoted
0,γ
as Cloc (Ω).
Note that pγ (u) = supx0 ∈Ω pγ (u)(x0 ). In other words, uniformly Hölder
continuous function do not include all those that are Hölder continuous at
all points of Ω. However, bounded locally Hölder continuous functions are
Hölder continuous at all points of Ω.
Example 2.1. Note that, for all β < 0, u(x) = |x|β on the open ball B1 (0) ⊂
Rn do not belong to C(B1 (0)). For β ∈ [0, ∞), |x|β ∈ C(B1 (0)) and, for
β ∈ [2, ∞) ∪ {0}, |x|β ∈ C 1 (B1 (0)). For β ∈ (0, 1], |x|β is Hölder continuous,
with exponent γ, for all 0 < γ ≤ β, but is not Hölder continuous for γ > β. In
particular, |x|β ∈ C 0,β (B1 (0)) for each β ∈ [0, 1]. Thus, we have one example
from each of the space C 0,β (B1 (0)). We first check the Hölder continuity at
x0 = 0. The γ-th Hölder coefficient is
|x|β
sup γ
= |x|β−γ ≤ 1 for γ ≤ β.
x∈B1 (0) |x|
For γ > β the γ-th Hölder coefficient blows up. More generally, for x 6= 0
and |x| > |y| (wlog),
β
β−γ |y|
|x| 1 −
|x|β − |y|β |x|
= γ .
|x − y|γ x
− y
|x| |x|
β
|y| |y| |y|
Since 0 < 1 − |x|
< 1 and 0 < β ≤ 1, 1 − |x|
≤ 1− |x|
. Thus,
β
|y|
|x|β−γ 1 − |x| |x|β−γ 1 − |y|
|x| ||x| − |y||
γ ≤ γ ≤ .
x
− y x
− y |x|1−γ |x − y|γ
|x| |x| |x| |x|
The last inequality is true for γ ≤ β and we have used |x|β−γ ≤ 1. For one
dimension, the last quantity is equal to 1 because
|y|
||x| − |y|| |x| 1 − |x| 1 − xy
= ≤ = 1.
|x|1−γ |x − y|γ |x||1 − xy |γ |1 − xy |
For higher dimensions, INCOMPLETE!!!

Example 2.2. The Cantor function fC ∈ C 0,γ ([0, 1]), i.e., is Hölder contin-
uous with exponent 0 < γ ≤ log3 2.1 Geometrically, the graph of Hölder
continuous function have fractal appearance which increases with smaller γ.
The case γ = 0 corresponds to bounded functions on Ω and hence is
ignored as an possible exponent. In fact, continuous functions in C 0,0 (Ω)
can be identified with the space Cb (Ω) of bounded continuous functions on
0,0
Ω. The space C(Ω) can be identified with continuous functions in Cloc (Ω),
the space of all locally Hölder continuous with exponent γ = 0. The case
γ > 1 is also ignored because only constant functions, on each connected
component of Ω, can satisfy Hölder estimate with exponent greater than
one.
Exercise 47. If u ∈ C 0,γ (Ω) with exponent γ > 1, then u is constant in each
of the connected component of Ω.
Proof. Let u ∈ C 0,γ (Ω). For any x ∈ Ω,
1 pγ (u)(x) γ
|uxi (x)| = lim |u(x+tei )−u(x)| ≤ lim |t| = lim pγ (u)(x)|t|γ−1 = 0.
t→0 |t| t→0 |t| t→0
Therefore, the gradient Du(x) = 0 for all x ∈ Ω and hence u is constant in

each connected components of Ω.
1
log3 2 ≈ 0.6309 . . .
The case γ = 1 corresponds to u being Lipschitz continuous. The space

of all Lipschitz functions is denoted as Lip(Ω) = C 0,1 (Ω).
Exercise 48. If Ω has finite diameter, then for any 0 < γ < δ ≤ 1 we have
C 0,1 (Ω) ⊆ C 0,δ (Ω) ( C 0,γ (Ω) ( C(Ω).
Proof. The last inclusion is trivial because every u ∈ C 0,γ (Ω) is uniformly
h i1/γ
ε
continuous by choosing δ = pγ (u) , for any given ε > 0. Thus, C 0,γ (Ω) (
C(Ω) for all 0 < γ ≤ 1. On the other hand, let u ∈ C 0,δ (Ω). Consider

|u(x) − u(y)|
pγ (u) = sup
x,y∈Ω |x − y|γ
x6=y

|u(x) − u(y)| δ−γ
= sup |x − y|
x,y∈Ω |x − y|δ
x6=y
≤ diam(Ω)δ−γ pδ (u).
Thus, pγ (u) < +∞ and u ∈ C 0,γ (Ω). The inclusions are strict by Exam-
ple 2.1. If γ <√δ, then u(x) = |x|γ is in C 0,γ (−1, 1) but not in C 0,δ (−1, 1).
In particular, x ∈ C 0,1/2 (−1, 1) but is not Lipschitz, i.e. do not belong to
C 0,1 (−1, 1).
A natural question at this juncture is: what is the relation between the
spaces C 1 (Ω) and C 0,1 (Ω). Of course |x| ∈ C 0,1 (−1, 1) and not in C(−1, 1).
Given the inclusion relation in above exercise, we would tend to believe that
C 1 (Ω) is contained in C 0,1 (Ω), but this is not true. In fact, only a subclass
of C 1 (Ω) belong to Lipschitz class.
Theorem 2.1.3. Let Ω be a convex domain. If u ∈ C 1 (Ω) and Dα u is
bounded on Ω, for each |α| = 1, then u ∈ C 0,1 (Ω).
Proof. Let u ∈ C 1 (Ω) such that there is a C0 > 0 such that |∇u(x)| ≤ C0
for all x ∈ Ω. For any two given points x, y ∈ Ω, we define the function
F : [0, 1] → R as F (t) = u((1 − t)x + ty). F is well defined due to the
convexity of Ω. Since u is differentiable, F is differentiable in (0, 1). Thus, for
all t ∈ (0, 1), F 0 (t) = ∇u[(1−t)x+ty]·(y −x). Therefore, |F 0 (t)| ≤ C0 |y −x|.
By mean value theorem, there is a ξ ∈ (0, 1) such that F 0 (ξ) = F (1) − F (0).
Thus,
|u(y) − u(x)| = |F (1) − F (0)| = |F 0 (ξ)| ≤ C0 |y − x|.

Hence, u is Lipschitz continuous.

What fails in the converse of the above theorem is the fact that a Lipschitz
function can fail to be differentiable. If we assume, in addition to Lipschitz
continuity, that u is differentiable then u ∈ C 1 and derivative is bounded.
Exercise 49. Prove the above theorem when Ω is path connected.
Theorem 2.1.4. For any γ ∈ (0, 1], the space C 0,γ (Ω) can be identified with
the space C 0,γ (Ω).
Proof. The space C 0,γ (Ω) can be identified with a subset of C 0,γ (Ω), by
identifying any u ∈ C 0,γ (Ω) with u |Ω ∈ C 0,γ (Ω). On the other hand, for any
u ∈ C 0,γ (Ω) we wish to extend it uniquely to a function ũ ∈ C 0,γ (Ω) such
that ũ |Ω = u. Let u ∈ C 0,γ (Ω). Set ũ(x) = u(x) for all x ∈ Ω. Now, for any
x0 ∈ ∂Ω, choose a sequence {xm } ⊂ Ω such that limm xm = x0 . Then, we
have
|u(xk ) − u(xl )| ≤ C|xk − xl |γ → 0 as k, l → ∞.
Thus, {u(xm )} is a Cauchy sequence and converges in R. Set ũ(x0 ) =
limm u(xm ). We now show that the definition of ũ(x0 ) is independent of
the choice of the sequence. If {ym } ⊂ Ω is any other sequence converging to
x0 , then
|u(xm ) − u(ym )| ≤ C|xm − ym |γ → 0 as m → ∞.
Thus, ũ(x0 ) is well-defined for all x0 ∈ ∂Ω. Moreover, for all x, y ∈ Ω, we
have
|ũ(x) − ũ(y)| = lim |u(xm − u(ym )| ≤ C lim |xm − ym |γ = C|x − y|γ .

m m
Hence, ũ satisfies the Hölder estimate with exponent γ and is in C 0,γ (Ω).
The above result is very special to uniformly continuous spaces and,
hence, Hölder spaces. For instance, a similar result is not true between
C(Ω) and C(Ω). If Ω is compact then every function in C(Ω) is bounded,
where as C(Ω) might have unbounded functions. An interesting point in the
proof of above result is that the Hölder coefficient are same for the extension,
i.e., pγ (u) = pγ (ũ). An advantage of above theorem is that, without loss of
generality, we can use the space C 0,γ (Ω) and C 0,γ (Ω) interchangeably.
Exercise 50. Show that the Hölder coefficient, pγ , defines a semi-norm on
C 0,γ (Ω).
Proof. For any constant function u, pγ (u) = 0. Also,

|u(x) + v(x) − u(y) − v(y)|
pγ (u + v) = sup
x6=y

|u(x) − u(y) + v(x) − v(y)|
= sup
x6=y

|u(x) − u(y)| + |v(x) − v(y)|
≤ sup
x6=y
≤ pγ (u) + pγ (v).
Further pγ (λu) = |λ|pγ (u).

Note that for bounded open subsets Ω, the space C 0,γ (Ω) inherits the
uniform topology from C(Ω).
Exercise 51. For any bounded open set Ω ⊂ Rn , show that C 0,γ (Ω) is dense
in C(Ω).
To make the space C 0,γ (Ω) complete on bounded Ω, we define the γ-
Hölder norm on C 0,γ (Ω) as
kukC 0,γ (Ω) := kuk∞ + pγ (u),
where kuk∞ := supx∈Ω |u(x)|.

Exercise 52. For any bounded open set Ω ⊂ Rn , show that k · kC 0,γ (Ω) is a
norm on C 0,γ (Ω).
Theorem 2.1.5. For any bounded open set Ω ⊂ Rn , the space C 0,γ (Ω) is a
Banach space with norm k · kC 0,γ (Ω) .
Proof. We need to prove the completeness of the space C 0,γ (Ω) w.r.t the norm
k · kC 0,γ (Ω) . Let {um } be a Cauchy sequence in C 0,γ (Ω), then {um } ⊂ C(Ω)
is Cauchy w.r.t the supremum norm. Thus, there is a u ∈ C(Ω) such that
kum − uk∞ → 0, as m → ∞. We first show that u ∈ C 0,γ (Ω). For x, y ∈ Ω
with x 6= y, consider
|u(x) − u(y)| |um (x) − um (y)|

= lim ≤ lim sup pγ (um ) ≤ lim kum kC 0,γ (Ω) .
|x − y|γ m |x − y|γ m m
Since {um } is Cauchy, limm kum kC 0,γ (Ω) < ∞. Hence u ∈ C 0,γ (Ω). Finally,
we show that the sequence {um } converges to u in C 0,γ (Ω) w.r.t the γ-Hölder
norm. Consider,
|um (x) − u(x) − um (y) + u(y)| |um (x) − uk (x) − um (y) + uk (y)|
= lim
|x − y|γ k |x − y|γ
≤ lim sup pγ (um − uk ).
k
Therefore, limm pγ (um − u) ≤ limm lim supk pγ (um − uk ) and the RHS con-
verges to 0 since the sequence is Cauchy. Hence, kum − ukC 0,γ (Ω) → 0.
Exercise 53. The space C 0,γ (Ω) norm is not separable.2
For an unbounded open set Ω, we consider a sequence of exhaustion
compact sets, as described in §1.3.3, and make the space C 0,γ (Ω) a Fréchet
space.
Definition 2.1.6. Let X and Y be Banach spaces. A continuous (bounded)

linear operator L : X → Y is said to be compact if L maps every bounded
subset of X to precompact (closure compact) subsets of Y . Equivalently,
L maps bounded sequences of X to sequences in Y that admit convergent
subsequences.
Definition 2.1.7. Let X and Y be Banach spaces such that X ⊂ Y . We

say X is continuously imbedded in Y (denoted as X ,→ Y ), if there is a
constant such that
kxkY ≤ CkxkX ∀x ∈ X.
Further, we say X is compactly imbedded in Y (denoted as X ⊂⊂ Y ) if in
addition to being continuously imbedded in Y , every bounded set in X (w.r.t
the norm in X) is compact in Y (w.r.t the norm in Y ).
Theorem 2.1.8. Let Ω be a bounded open subset of Rn . For any 0 < γ <
δ ≤ 1, the inclusion map I : C 0,δ (Ω) → C 0,γ (Ω) is continuous and compact.
Further, the inclusion map I : C 0,γ (Ω) → C(Ω) is compact, for all 0 < γ ≤ 1.
Proof. The continuity of I follows from the inequality proved in Exercise 48

and hence I is a bounded linear operator. We need to show I is compact.
Let {um } be a sequence bounded in C 0,δ (Ω). Without loss of generality, we
2
The uniform norm in C(Ω) is separable
can assume that kum kC 0,δ (Ω) ≤ 1. Therefore, pδ (um ) ≤ 1 for all m which
implies that {um } is an equicontinuous sequence in C(Ω). By Arzelà-Ascoli
theorem, there exists a subsequence {umk } of {um } and a u ∈ C(Ω) such that
kumk − uk∞ → 0, as k → ∞. The u ∈ C(Ω) is, in fact, in C 0,δ (Ω) ⊂ C 0,γ (Ω)
because
|u(x) − u(y)| |umk (x) − umk (y)|
δ
= lim ≤1
|x − y| k |x − y|δ
and, hence, pδ (u) ≤ 1. We now show that the umk converges to u in C 0,γ (Ω).
For simplicity, set vk := umk − u and we will show that {vk } converges to 0 in
C 0,γ (Ω). Obviously, kvk k∞ → 0 thus it is enough to show that pγ (vk ) → 0.
Note that, for every given ε > 0, we have pγ (vk ) ≤ Skε + Tkε , where

ε |vk (x) − vk (y)|
Sk = sup
x6=y;|x−y|≤ε
and
|vk (x) − vk (y)|
Tkε = sup .
|x−y|>ε
Consider

|vk (x) − vk (y)|
Skε = sup |x − y|δ−γ ≤ εδ−γ pδ (vk ) ≤ 2εδ−γ .
x,y∈Ω |x − y|δ
x6=y;|x−y|≤ε
Similarly,
Tkε ≤ 2ε−γ kvk k∞ .
Hence, lim supk pγ (vk ) ≤ 2εδ−γ + 2ε−γ lim supk kvk k∞ = 2εδ−γ + 0. Since ε
can be made as small as possible, we have pγ (vk ) → 0.
We extend the Hölder continuous functions in to all differentiable class of
functions. We denote by C k,γ (Ω) the space of all C k (Ω), k-times continuously
differentiable, functions such that Dα u is Hölder continuous with exponent
γ, i.e., Dα u ∈ C 0,γ (Ω) for all |α| = k. For a bounded open set Ω, we give the
γ-th Hölder norm on C k,γ (Ω) as
k
X X
kukC k,γ (Ω) := kDα uk∞ + pγ (Dα u). (2.1.3)
|α|=0 |α|=k
It is enough to consider the Hölder coefficient of only the k-th derivative

because this is enough to complete the space.
Exercise 54. For any bounded open set Ω ⊂ Rn , show that k · kC k,γ (Ω) is
a norm on C k,γ (Ω). and the space C k,γ (Ω) is a Banach space with norm
k · kC k,γ (Ω) .
Proof. Let {um } be a Cauchy sequence in C k,γ (Ω). Then, {um } ⊂ C k (Ω)
is Cauchy in the supremum norm. Thus, there is a u ∈ C k (Ω) such that
kum − ukC k (Ω) → 0, as m → ∞. The fact that u ∈ C k,γ (Ω) and that u is a
limit, in the γ-Hölder norm, of the Cauchy sequence is similar to case k = 0
proved in Theorem 2.1.5.
As seen before, for an unbounded open set Ω, we consider a sequence of

exhaustion compact sets, as described in §1.3.3, and make the space C k,γ (Ω)
a Fréchet space. We end this section with a final result on the inclusion
between different order Hölder spaces.
Theorem 2.1.9. Let Ω be a bounded, convex open subset of Rn . For any

0 < γ, δ ≤ 1 and k, ` ∈ N ∪ {0} such that k + γ < ` + δ, then the inclusion
map I : C `,δ (Ω) → C k,γ (Ω) is continuous and compact.
Proof. If γ = δ and k = ` then there is nothing to prove. Alternately,

without loss of generality we assume 0 < γ < δ ≤ 1 and k < `. Then,
we know from Theorem 2.1.8 that C `,δ (Ω) ⊂ C `,γ (Ω) and the inclusion is
continuous and compact. Let u ∈ C `,γ . We need to show that u ∈ C k,γ (Ω).
i.e., pγ (Dk u) < ∞. Note that it is enough to show that p1 (Dk u) < ∞. The
fact that u ∈ C `,γ implies that u ∈ C ` (Ω) and D`−1 u is bounded. Thus,
by Theorem 2.1.3, we have u ∈ C `−1,1 (Ω) and hence is in C `−1,γ (Ω). By
repeating the argument for each derivative of D`−i for each i = 1, 2, . . . , k,
we have the required result. The inclusion C `,γ (Ω) ⊂ C k,γ (Ω) is continuous.
The composition of a continuous and compact operator is compact. Thus,
the inclusion I is compact.
2.2 W k,p(Ω) Spaces

For each 1 ≤ p ≤ ∞, we define its conjugate exponent q to be,
 p
 p−1
 if 1 < p < +∞
q = +∞ if p = 1

1 if p = +∞.

Note that for, 1 < p < +∞, q is the number for which 1/p + 1/q = 1. In
this section, we shall define spaces that are analogous to C k (Ω) using the
generalised notion of derivative, viz. weak derivative (cf. Definition 1.5.2).
We know that every u ∈ Lp (Ω) being locally integrable induces a distribu-
tion Tu . Further, the distribution Tu is differentiable for all multi-indices α.
However, we have already seen that Tu need not be weakly differentiable.
For a fixed multi-index α, if there exists a vα ∈ Lp (Ω) such that Tvα = Dα Tu ,
then we denote the vα as Dα u. We know that such a vα is unique up to a set
of measure zero. As usual, Ω is an open subset of Rn .
Definition 2.2.1. Let k ≥ 0 be an integer and 1 ≤ p ≤ ∞. The Sobolev

space W k,p (Ω) is the subclass of all u ∈ Lp (Ω) such that there exists a vα ∈
Lp (Ω), for all 0 ≤ |α| ≤ k, such that
Z Z
|α|
vα φ dx = (−1) uDα φ dx ∀φ ∈ D(Ω).
Ω Ω
Equivalently,
W k,p (Ω) := {u ∈ Lp (Ω) | Dα u ∈ Lp (Ω) ∀0 ≤ |α| ≤ k}.
With this convention, we have W 0,p (Ω) = Lp (Ω). Note that, by definition,
if u ∈ W k,p (Ω) then every function in the equivalence class of u is also in
W k,p (Ω) and they all have their derivative as vα upto a set of measure zero.
Example 2.3. Let Ω = (−a, a) ⊂ R, for a positive real number a. Recall that
the function u(x) = |x| is not in C 1 (−a, a). But u ∈ W 1,p (−a, a). Consider
the v ∈ Lp (−a, a) defined as
(
1 x ∈ [0, a)
v(x) =
−1 x ∈ (−a, 0).
We shall show that v is the weak derivative of u. Consider, for φ ∈ D(−a, a),
Z a Z 0 Z a Z 0 Z a Z a
0 0
vφ dx = − φ dx + φ dx = xφ dx − xφ dx = − |x|φ0 dx.
−a −a 0 −a 0 −a
More generally, any continuous function on [−a, a] which is pieceswise differ-

entiable, i.e., piecewise C 1 on (−a, a) is in W 1,p (−a, a).
Example 2.4. The function v defined in above example is not in W 1,p (−a, a).
The argument is similar to the function w defined as
(
1 x ∈ [0, a)
w(x) =
0 x ∈ (−a, 0).
Note that w ∈ Lp (−a, a), for all p. However, the distributional derivative
of w (and v) is the Dirac measure at 0, δ0 , which is a singular distribution.
Hence w 6∈ W k,p (Ω), for all k > 0.
Example 2.5. The function u defined as
(
x x ∈ (0, a)
u(x) =
0 x ∈ (−a, 0]
is in W 1,p (−a, a), for all p ∈ [1, ∞], since both u and its distributional deriva-
tive Du = w are in Lp (−a, a) (cf. previous example). But u 6∈ W k,p (Ω), for
all k ≥ 2.
Note that we have the inclusion W k,p (Ω) ( W `,p (Ω) for all ` < k. The
inclusion is strict as seen from above examples.
Exercise 55. Find values of β ∈ R such that |x|β ∈ W 1,p (B1 (0)) for a fixed
p ∈ [1, ∞) and B1 (0) is the open unit ball in Rn .
Proof. Let u(x) = |x|β . Note that Dei u(x) = βxi |x|β−2 and |∇u(x)| =
β|x|β−1 . Consider
Z Z 1
k∇ukpp =β p p(β−1)
|x| p
dx = β ωn rpβ−p+n−1 dr.
Ω 0
The last quantity is finite iff p(β −1)+n−1 > −1, i.e., β > 1− np (also β = 0,
if not already included in the inequality condition) and |x|β ∈ W 1,p (Ω).
Using the above exercise, we have an example of a function in Lp (B1 (0))

but not in W 1,p (B1 (0)). For instance, if Ω = B1 (0) in Rn and u(x) = |x|β ,
for non-zero β, such that − np < β ≤ 1 − np , is not in W 1,p (B1 (0)) but is in
Lp (B1 (0)).
Exercise 56. Find values of β ∈ R such that |x|β ∈ W 2,p (B1 (0)), for a fixed
p ∈ [1, ∞).
Proof. Let u(x) = |x|β . Note that
|uxi xj | ≤ |β(β − 2)||x|β−2 + |β||x|β−2 .

−n
Hence, uxi xj ∈ Lp (B1 (0)) if β − 2 > p
, or β > 2 − np . Thus u ∈ W 2,p (B1 (0))
if β > 2 − np and β = 0.
Exercise 57. If u ∈ Lp (Ω) is such that Dα u ∈ Lp (Ω) for |α| = k, then is it

true that Dα u ∈ Lp (Ω) for all |α| = 1, 2, . . . , k − 1?
Exercise 58. Show that the spaces W k,p (Ω) are all (real) vector spaces.
Exercise 59. Show that u 7→ |α|=k kDα ukp defines a semi-norm on W k,p (Ω)
P
for k > 0. We denote the semi-norm by |u|k,p,Ω .

For 1 ≤ p < ∞, we endow the space W k,p (Ω) with the natural norm,
k
X k Z
X 1/p
α α p
kukk,p,Ω := kD ukp = |D u| .
|α|=0 |α|=0 Ω
For p = ∞, we define the norm on W k,∞ (Ω) to be,
k
X
kukk,∞,Ω = ess supΩ |Dα u|.
|α|=0
Observe that kuk0,p,Ω = |u|0,p,Ω = kukp , the usual Lp -norm.

Exercise 60. Show that the norm kukk,p,Ω defined above is equivalent to the
norm
 1/p  1/p
Xk Xk Z
 kDα ukpp  =  |Dα u|p  , for 1 < p < ∞.
|α|=0 |α|=0 Ω
For p = 1, the norms are same.

Frequently, we may skip the domain subscript in the norm where there is
no confusion on the domain of function. Also, we shall tend to use the symbol
k · kp for k · k0,p in Lp -spaces. We set H k (Ω) = W k,2 (Ω). For u ∈ H k (Ω), we
shall denote the norm as k · kk,Ω .
Exercise 61. Show that

k Z
X
hu, vik,Ω = Dα uDα v, ∀u, v ∈ H k (Ω)
|α|=0 Ω
defines an inner-product in H k (Ω).
Theorem 2.2.2. For every 1 ≤ p ≤ ∞, the space W k,p (Ω) is a Banach

space. If 1 < p < ∞, it is reflexive and if 1 ≤ p < ∞, it is separable. In
particular, H k (Ω) is a separable Hilbert space.
Proof. Let {um } be a Cauchy sequence in W k,p (Ω). Thus, by the definition
of the norm on W k,p (Ω), {um } and {Dα um }, for 1 ≤ |α| ≤ k are Cauchy in
Lp (Ω). Since Lp (Ω) is complete, there exist function u, vα ∈ Lp (Ω) such that
um → u and Dα um → vα ∀1 ≤ |α| ≤ k.
To show that W k,p (Ω) is complete, it is enough to show that Dα u = vα , for

1 ≤ |α| ≤ k. Let φ ∈ D(Ω), then for each α such that 1 ≤ |α| ≤ k, we have
Z Z
α
uD φ dx = lim um Dα φ dx
Ω m→∞ Ω Z
|α|
= lim (−1) Dα um φ dx
m→∞ Ω
Z
= (−1)|α| vα φ dx
Ω
Thus, Dα u = vα , i.e., vα is the α weak derivative of u.

Note that we have an isometry from W k,p (Ω) to (Lp (Ω))β , where β :=
Pk P 1/p
i p β β p
i=0 n and the norm endowed on (L (Ω)) is kuk := i=1 kui kp with
u = (u1 , u2 , . . . , uβ ). The image of W k,p (Ω) under this isometry is a closed
subspace of (Lp (Ω))β . The reflexivity and separability of W k,p (Ω) is inherited
from reflexivity and separability of (Lp (Ω))β , for 1 < p < ∞ and 1 ≤ p < ∞,
respectively.
Remark 2.2.3. The proof above uses an important fact which is worth
noting. If um → u in Lp (Ω) and Dα um , for all |α| = 1, is bounded in Lp (Ω)
then u ∈ W 1,p (Ω).
2.3 Smooth Approximations

Recall that in Theorem 1.3.30, we proved that Cc∞ (Ω) is dense in Lp (Ω) for
any open subset Ω ⊆ Rn . We are now interested in knowing if the density of
Cc∞ (Ω) holds true in the Sobolev space W k,p (Ω), for any k ≥ 13 .
Theorem 2.3.1. For 1 ≤ p < ∞, Cc∞ (Rn ) is dense in W k,p (Rn ), for all
k ≥ 1.
Proof. Let k ≥ 1 and u ∈ W k,p (Rn ). The Dα u ∈ Lp (Rn ) for all 0 ≤ |α| ≤ k.
By Theorem 1.3.29, we know that {ρm ∗ u} ⊂ C ∞ (Rn ) converges to u in
the W k,p (Ω)-norm, for all k ≥ 0. Choose a function φ ∈ Cc∞ (Rn ) such that
φ ≡ 1 on B(0; 1), φ ≡ 0 on Rn \ B(0; 2) and 0 ≤ φ ≤ 1 on B(0; 2) \ B(0; 1).
Note that |Dα φ| is bounded. In fact, |Dα φ(x)| ≤ 1 for all 0 ≤ |α| ≤ k. Now
set φm = φ(x/m). This is one choice of cut-off functions as introduced in
Theorem 1.3.30. Set um := φm (ρm ∗ u) in Rn , then
X
kDα um − Dα ukp ≤ Cα (Dα−β φm )(ρm ∗ Dβ u) p
β<α
X
+ kφm (ρm ∗ Dα u) − Dα ukp
β=α
X
≤ |Cα |kDα−β φm k∞ kρm ∗ Dβ ukp
β<α
X
+ kφm (ρm ∗ Dα u − Dα u)kp
β=α
X
+ kφm Dα u − Dα ukp
β=α
M X
≤ |Cα |kDβ ukp
m|α−β| β<α
X
+ kρm ∗ Dα u − Dα ukp
β=α
X
+ kφm Dα u − Dα ukp .
β=α
The first term with 1/m goes to zero, the second term tends to zero by
Theorem 1.3.29 and the last term goes to zero by Dominated convergence
3
The case k = 0 is precisely the result of Theorem 1.3.30
theorem. Since
k
X
kum − ukk,p = kDα um − Dα ukp
|α|=0
k
X X
= Cα (Dα−β φm )(ρm ∗ Dβ u) − Dα u
|α|=0 β≤α p
and each term in the sum goes to zero as m increases, we have um converges
to u in W k,p (Rn ) norm.
The real trouble in approximating W k,p (Ω) by Cc∞ (Ω) functions is that
the approximation approach boundary continuously, if Ω has a boundary.
Theorem 2.3.2. Let 1 ≤ p < ∞ and Ω ⊂ Rn . Then Cc∞ (Rn ) is dense in

k,p
Wloc (Ω), i.e., for each u ∈ W k,p (Ω) there is a sequence vm ∈ Cc∞ (Rn ) such
that kvm − ukk,p,ω , for all k ≥ 1 and for all ω ⊂⊂ Ω (relatively compact in
Ω).
Proof. Let u ∈ W k,p (Ω). Fix ω relatively compact subset of Ω and a α such
that 1 ≤ |α| ≤ k. Let ũ denote the extension of u by zero in Ωc . Choose a
φ ∈ Cc∞ (Ω) such that φ ≡ 1 on ω and 0 ≤ φ ≤ 1 in Ω. Then supp(φu) ⊂ Ω
and we extend φu to all of Rn by zero in Ωc and denote it as φu.
f Set v = φuf
n k,p n
on R . Note that v ∈ W (R ). By Theorem 2.3.1, we have the sequence
vm := φm (ρm ∗ v) ∈ Cc∞ (Rn ) converging to v in α Sobolev norm. Since on ω,
v = u, we have
kvm − ukk,p,ω ≤ kvm − vkk,p,Rn → 0.
This is true for all relatively compact subset ω of Ω.
Note that the restriction ‘relatively compact set’ is only for k ≥ 1. For the
case k = 0, the restriction to Ω works, as seen in Theorem 1.3.30. The density
of Cc∞ (Ω) in W k,p (Ω) may fail to generalise for an arbitrary proper subset
Ω ⊂ Rn , because a “bad” derivative may be introduced at the boundary
while extending by zero outside Ω.
Example 2.6. Let Ω = (0, 1) ⊂ R and u ≡ 1 on Ω. Then u ∈ W 1,p (0, 1).
Setting ũ = 0 in R \ (0, 1), we see that ũ ∈ Lp (R) but not in W 1,p (R).
Because Dũ = δ0 − δ1 is not in L1loc (R).
Recall the inclusion Cc∞ (Ω) ⊂ C ∞ (Ω) ⊂ C ∞ (Ω). The C ∞ (Ω) denotes all
functions in C ∞ (Ω) such that all its derivatives can be extended continuously
to Ω. For an arbitrary subset Ω ⊆ Rn the best one can do is the following
density result.
Theorem 2.3.3 (Meyers-Serrin). Let 1 ≤ p < ∞ and Ω ⊆ Rn . For u ∈
W k,p (Ω) and any ε > 0, there is a φ ∈ C ∞ (Ω) such that kφkk,p,Ω < ∞ and
ku − φkk,p,Ω < ε.
Proof. For each m ∈ N, consider the sets
1
ωm := {x ∈ Ω | |x| < m and dist(x, ∂Ω) > }
m
and set ω0 = ∅. Define the collection of open sets {Um } as Um := ωm+1 ∩
(ω m−1 )c . Note that Ω = ∪m Um is an open covering of Ω. Thus, we choose the
C ∞ locally finite partition of unity {φm } ⊂ Cc∞ (Ω) such that supp(φm ) ⊂ Um ,
k,p
P
0 ≤ φm ≤ 1 and m φm = 1. For the given u ∈ W (Ω), note that
φm u ∈ W (Um ) with supp(φm u) ⊂ Um . We extend φm u to all of Rn by zero
k,p
outside Um , i.e.,
(
φm u(x) x ∈ Um
φg
m u(x) =
0 x ∈ Rn \ Um .
k,p
Observe that φg mu ∈ W (Rn ). Let ρδ be the sequence of mollifiers and
∞ n
consider the sequence ρδ ∗φgm u in C (R ). Support of ρδ ∗φm u ⊂ Um +B(0; δ).
g
Note that for all x ∈ Um , 1/(m + 1) < dist(x, ∂Ω) < 1/(m − 1). Thus,
for all 0 < δ < 1/(m + 1)(m + 2)4 , supp(ρδ ∗ φg c
m u) ⊂ ωm+2 ∩ (ω m−2 ) ,
which is compactly contained in Ω. Since φm u ∈ W k,p (Ω), we can choose a
subsequence {δm } going to zero in (0, 1/(m + 1)(m + 2) such that
ε
kρδm ∗ φm u − φm ukk,p,Ω = kρδm ∗ φgm u − φm ukk,p,Rn < m .
2
∞
P
Set φ = m ρδm ∗ φm u. Note that φ ∈ C (Ω) and kφkk,p,Ω < ∞. Since every
Um intersects Um−1 and Um+1 , we at most have three non-zero terms in the
sum for each x ∈ Um , i.e.,
1
X
φ(x) = (ρδm+i ∗ φm+i )(x) x ∈ Um .
i=−1

4 1 1
The choice of this range for δ is motivated from the fact that m+1 − (m+1)(m+2) =
1
(m+2)
Therefore,
X X
ku − φkk,p,Ω = k (uφm − ρδm ∗ φm u)kk,p,Ω ≤ kφm u − ρδm ∗ φm ukk,p,Ω < ε.
m m
This proves the density of C ∞ (Ω) in W k,p (Ω).

Let C k,p (Ω) denote the closure of E := {φ ∈ C k (Ω) | kφkk,p,Ω < ∞} w.r.t
the Sobolev norm k · kk,p,Ω . The space C k,p (Ω) is a subspace of W k,p (Ω) be-
cause classical derivative (continuous) and distributional derivative coincide
(because integraion by parts is valid). Since W k,p (Ω) is a Banach space, we
have C k,p (Ω) is a closed subspace of W k,p (Ω). In fact, a consequence of above
result is that the Sobolev space W k,p (Ω) = C k,p (Ω), a result due to Meyers
and Serrin (cf. [MS64]) proved in 1964.
Corollary 2.3.4 (Meyers-Serrin). Let 1 ≤ p < ∞ and Ω ⊆ Rn , then

C k,p (Ω) = W k,p (Ω).
Proof. It is enough to show that E is dense in W k,p (Ω) because, as a con-

sequence, W k,p (Ω) = E = C k,p (Ω). For each given ε > 0 and u ∈ W k,p (Ω),
we need to show the existence of φ ∈ E such that ku − φkk,p,Ω < ε. By
Theorem 2.3.3 there is a φ ∈ C ∞ (Ω) ∩ E such that ku − φkk,p,Ω < ε. Hence
proved.
The density of C ∞ (Ω) is not true, in general, and fails for some “bad”
domains as seen in examples below. This, in turn, means that Cc∞ (Ω) cannot,
in general, be dense in W k,p (Ω).
Example 2.7. Let Ω := {(x, y) ∈ R2 | 0 < |x| < 1 and 0 < y < 1}. Consider
the function u : Ω → R defined as
(
1 x>0
u(x, y) =
0 x < 0.
Then for any fixed ε > 0 there exists no φ ∈ C 1 (Ω) such that ku − φk1,p < ε.
Example 2.8. Let Ω := {(r, θ) ∈ R2 | 1 < r2 < 2 and θ 6= 0}. Consider
u(r, θ) = θ. Then there exists no φ ∈ C 1 (Ω) such that ku − φk1,1 < 2π.
The trouble with domains in above examples is that they lie on both
sides of the boundary ∂Ω which becomes the main handicap while trying to
approximate W k,p (Ω) by C ∞ (Ω) functions.
Definition 2.3.5. A subset Ω ⊂ Rn is said to satisfy the segment property

if for every x ∈ ∂Ω, there is a open set Bx , containing x, and a non-zero
unit vector ex such that z + tex ∈ Ω, for all 0 < t < 1 whenever z ∈ Ω ∩ Bx .
Observe that a domain with segment property cannot lie on both sides
of its boundary.
Theorem 2.3.6. Let 1 ≤ p < ∞ and Ω ⊆ Rn satisfy the segment property.
Then, for every u ∈ W k,p (Ω) there exists a sequence {vm } ⊂ Cc∞ (Rn ) such
that kvm − ukk,p,Ω → 0, for all k ≥ 1, i.e., Cc∞ (Rn ) is dense in W k,p (Ω).
Theorem 2.3.1 is a particular case of the above result, because Rn trivially
satisfies the segment property as it has ‘no boundary’. A stronger version of
above result is Theorem 2.5.2 proved in next section.
Corollary 2.3.7. Let 1 ≤ p < ∞ and Ω ⊆ Rn satisfy the segment property.
Then C ∞ (Ω) is dense in W k,p (Ω).
In general, the closure of Cc∞ (Ω) is a proper subspace of W k,p (Ω) w.r.t
the k · kk,p,Ω .
Definition 2.3.8. Let W0k,p (Ω) be the closure of Cc∞ (Ω) in W k,p (Ω). For
p = 2, we denote W0k,p (Ω) by H0k (Ω).
Exercise 62. Show that W0k,p (Ω) is a closed subspace of W k,p (Ω).
In general, W0k,p (Ω) is a strict subspace of W k,p (Ω). However, as observed
in Theorem 2.3.1, for Ω = Rn , W k,p (Rn ) = W0k,p (Rn ) for all 1 ≤ p < ∞ and
k ≥ 0. In fact, W k,p (Ω) = W0k,p (Ω) iff capp (Rn \ Ω) = 05 .
The density results discussed in this section is not true for p = ∞ case as
seen from the example below.
Example 2.9. Let Ω = (−1, 1) and
(
0 x≤0
u(x) =
x x ≥ 0.
Then its distributional derivative is u0 (x) = 1(0,∞) . Let φ ∈ C ∞ (Ω) such that
kφ0 − u0 k∞ < ε. Thus, if x < 0, |φ0 (x)| < ε and if x > 0, |φ0 (x) − 1| < ε.
In particular, φ0 (x) > 1 − ε. By continuity, φ0 (0) < ε and φ0 (0) > 1 − ε
which is impossible if ε < 1/2. Hence, u cannot be approximated by smooth
functions in W 1,∞ (Ω) norm.
5
(k,p) polar sets
2.4 Characterisation of W0k,p(Ω)

Theorem 2.4.1 (Chain rule). Let G ∈ C 1 (R) such that G(0) = 0 and
|G0 (x)| ≤ M for all x ∈ R. Also, let u ∈ W 1,p (Ω). Then G ◦ u ∈ W 1,p (Ω)
and
∂ ∂u
(G ◦ u) = (G0 ◦ u) , 1 ≤ i ≤ n.
∂xi ∂xi
Recall the definition of W0k,p (Ω) from §2.2. We have already characterised
W0k,p (Ω) when Ω = Rn (cf. Theorem 2.3.1).
Theorem 2.4.2. Let 1 ≤ p < ∞ and let u ∈ W 1,p (Ω) such that u vanishes
outside a compact set contained in Ω. Then u ∈ W01,p (Ω).
Theorem 2.4.3. Let 1 ≤ p ≤ ∞ and u ∈ W 1,p (Ω) ∩ C(Ω). If u = 0 on ∂Ω
then u ∈ W01,p (Ω).
Theorem 2.4.4 (Stampacchia). Let G : R → R be a Lipschitz continuous
function such that G(0) = 0. For 1 < p < ∞, if Ω is bounded and u ∈
W01,p (Ω), then we have G ◦ u ∈ W01,p (Ω).
Exercise 63. Let G : R → R be a Lipschitz continuous function such that
G(0) = 0. Let Ω be a bounded open subset of Rn which admits an extension
operator. Then, for 1 < p < ∞, u ∈ W 1,p (Ω) implies that G ◦ u ∈ W 1,p (Ω).
Corollary 2.4.5. Let Ω be a bounded open set of Rn . If u ∈ H01 (Ω) then
|u|, u+ and u− belong to H01 (Ω), where
u+ (x) = max{u(x), 0} (2.4.1)
u− (x) = max{−u(x), 0} = min{u(x), 0}. (2.4.2)
Lemma 2.4.6. Let Ω ⊂ Rn be an open set and let u ∈ W 1,p (Ω). If K ⊂ Ω
is a closed set and u vanishes outside K, then the function
(
u(x) if x ∈ Ω
ũ(x) =
0 if x ∈ Rn \ Ω
belongs to W 1,p (Rn ).

Exercise 64. In the above Lemma, if K is compact then u ∈ W01,p (Ω).
Theorem 2.4.7. Let 1 < p < ∞ and Ω be an open subset of Rn . If u ∈
W01,p (Ω) then ũ ∈ W 1,p (Rn ). Further, for any 1 ≤ i ≤ n, ∂x
∂ ũ ∂u
= ∂x .
f
i i
2.5 Extension Operators

In general, many properties of W k,p (Ω) can be inherited from W k,p (Rn ) pro-
vided the domain is “nice”. For instance, we will see that if W k,p (Rn ) is
continuously imbedded in Lq (Rn ) then, for nice domains, W k,p (Ω) is also
continuously imbedded in Lq (Ω). In this section we classify such classes of
“nice” domain.
Definition 2.5.1. Let Ω be an open subset of Rn . We say P is an (k, p)-
extension operator for Ω, if P : W k,p (Ω) → W k,p (Rn ) is a bounded linear
operator, i.e., there is a constant C > 0 (depending on Ω, k and p) such that
kP ukk,p,Rn ≤ Ckukk,p,Ω ∀u ∈ W k,p (Ω)
and P u |Ω = u a.e. for every u ∈ W k,p (Ω). If P is same for all 1 ≤ p < ∞
and 0 ≤ m ≤ k, then P is called strong k-extension operator. If P is a strong
k-extension operator for all k then P is called total extension operator.
Example 2.10. There is a natural extension operator P : W0k,p (Ω) → W k,p (Rn )
which is the extension by zero. Define
(
u(x) x ∈ Ω
P u := ũ =
0 x ∈ Rn \ Ω.
Obviously, ũ |Ω = u and kũk0,p,Rn = kuk0,p,Ω . We shall show that kũkk,p,Rn =

kukk,p,Ω . Since u ∈ W0k,p (Ω), there is a sequence {φm } ⊂ Cc∞ (Ω) converging
to u in k · kk,p,Ω . For any φ ∈ Cc∞ (Rn ) and |α| ≤ k, we have
Z Z
α |α| α |α|
D ũ(φ) = (−1) ũD φ dx = (−1) uDα φ dx
Rn Ω Z
Z
= (−1)|α| lim φm Dα φ dx = lim Dα φm φ dx
m→∞ Ω m→∞ Ω
Z
= Dα uφ dx = D gα u(φ).
Ω
Thus, Dα ũ = D
g α u and therefore
k
X k
X k
X
α
kũkk,p,Rn = kD ũkp,Rn = kD
g α uk
p,Rn = kDα ukp,Ω = kukk,p,Ω .
|α|=0 |α|=0 |α|=0
Theorem 2.5.2. Let 1 ≤ p < ∞. If Ω is an open subset of Rn such that

there is a (k, p)-extension operator P : W k,p (Ω) → W k,p (Rn ), then Cc∞ (Rn )
is dense in W k,p (Ω). In particular, C ∞ (Ω) is dense in W k,p (Ω) w.r.t k·kk,p,Ω .
Proof. For each u ∈ W k,p (Ω), we choose the sequence φm (ρm ∗P u) in Cc∞ (Rn )
which converges to P u in W k,p (Rn ) and their restriction to Ω is in C ∞ (Ω)
and converges in W k,p (Ω).
Above result is a particular case of Theorem 2.3.6. A natural question
provoked by Theorem 2.5.2 is: For what classes of open sets Ω can one expect
an extension operator P .
Definition 2.5.3. For an open set Ω ⊂ Rn we say that its boundary ∂Ω is
C k (k ≥ 1), if for every point x ∈ ∂Ω, there is a r > 0 and a C k diffeomor-
phism γ : Br (x) → B1 (0) ( i.e. γ −1 exists and both γ and γ −1 are k-times
continuously differentiable) such that
1. γ(∂Ω ∩ Br (x)) ⊂ B1 (0) ∩ {x ∈ Rn | xn = 0} and
2. γ(Ω ∩ Br (x)) ⊂ B1 (0) ∩ {x ∈ Rn | xn > 0}
We say ∂Ω is C ∞ if ∂Ω is C k for all k = 1, 2, . . . and ∂Ω is analytic if γ is
analytic.
Equivalently, a workable definition of C k boundary would be the follow-
ing: if for every point x ∈ ∂Ω, there exists a neighbourhood Bx of x and a
C k function γ : Rn−1 → R such that
Ω ∩ Bx = {x ∈ Bx | xn > γ(x1 , x2 , . . . , xn−1 )}.
To keep the illustration simple, we shall restrict ourselves to k = 1. For

more general results we refer to [Ada75]. We begin by constructing an exten-
sion operator for the half-space and then use it along with partition of unity
to construct an extension operator for C 1 boundary domains.
Theorem 2.5.4. Let Rn+ := {x = (x0 , xn ) ∈ Rn | xn > 0} where x0 =
(x1 , x2 , . . . , xn−1 ). Given u ∈ W 1,p (Rn+ ), we define the extension to Rn as
(
u(x0 , xn ) xn > 0
P u = u∗ (x) := 0
u(x , −xn ) xn < 0.
Then ku∗ k1,p,Rn ≤ 2kuk1,p,Rn+ and u∗ ∈ W 1,p (Rn ).

Proof. Observe that u∗ ∈ Lp (Rn ) because

Z Z
∗ p 0
ku kp,Rn = p
|u(x , xn )| dx + |u(x0 , −xn )|p dx
Rn
+ Rn
−
Z Z Z ∞
0
= p
|u(x , xn )| dx + |u(x0 , yn )|p dx0 dyn
Rn
+ Rn−1 0
Z
= 2 |u(x)|p dx = 2kukpp,Rn+ .
Rn
+
Thus, ku∗ kp,Rn = 21/p kukp,Rn+ . We now show that for α = ei , 1 ≤ i ≤ n − 1,

Dα u∗ = (Dα u)∗ . Consider, for φ ∈ D(Rn ),
Z
α ∗
D u (φ) = − u∗ Dα φ dx
n
ZR Z
0
= − α
u(x , xn )D φ dx − u(x0 , −xn )Dα φ dx
Rn
+ Rn
−
Z Z
0 0
= − α
u(x , xn )D φ(x , xn ) dx − u(x0 , xn )Dα φ(x0 , −xn ) dx.
Rn
+ Rn
+
Hence, Z
α ∗
D u (φ) = − uDα ψ(x) dx, (2.5.1)
Rn
+
where ψ(x0 , xn ) = φ(x0 , xn )+φ(x0 , −xn ) when xn > 0. In general, ψ ∈

/ D(Rn+ ).
Thus, we shall multiply ψ by a suitable cut-off function so that the product
is in D(Rn+ ). Choose a {ζm } ∈ C ∞ (R) such that
(
0 if t < 1/2m
ζm (t) =
1 if t > 1/m,
then ζm (xn )ψ(x) ∈ D(Rn+ ). Since ζm is independent of xi , for 1 ≤ i ≤ n − 1,

we have
Z Z
α α
D u(ζm ψ) = − u(x)D (ζm (xn )ψ(x)) dx = − u(x)ζm (xn )Dα ψ(x) dx.
Rn
+ Rn
+
Passing to limit, as m → ∞ both sides, we get

Z Z
α
D u(x)ψ(x) dx = − u(x)Dα ψ(x) dx.
Rn
+ Rn
+
The RHS in above equation is same as the RHS obtained in (2.5.1). Hence,
α ∗
R α
we have D u (φ) = Rn D u(x)ψ(x) dx. By setting
+
(
Dα u(x0 , xn ) xn > 0
(Dα u)∗ (x0 , xn ) := α 0
D u(x , −xn ) xn < 0,
we get Dα u∗ = (Dα u)∗ . We now show a similar result for α = en . Consider,

for φ ∈ D(Rn ),
Z
en ∗
D u (φ) = − u∗ Den φ dx
n
ZR Z
0
= − en
u(x , −xn )D φ dx − u(x0 , xn )Den φ dx
Rn
− Rn
+
Z Z
0 0
= en
u(x , xn )D φ(x , −xn ) dx − u(x0 , xn )Den φ(x0 , xn ) dx.
Rn
+ Rn
+
Hence, Z
en ∗
D u (φ) = − uDen ψ(x) dx, (2.5.2)
Rn
+
where ψ(x0 , xn ) = φ(x0 , xn )−φ(x0 , −xn ) when xn > 0. Note that ψ(x0 , 0) = 0.
Thus, by Mean value theorem, |ψ(x0 , xn )| ≤ C|xn |. As before, in general,
ψ∈ / D(Rn+ ), so ζm (xn )ψ(x) ∈ D(Rn+ ). One such choice of ζm is by choosing
ζm (t) = ζ(mt) where (
0 if t < 1/2
ζ(t) =
1 if t > 1.
Consider,
Z
en
D u(ζm ψ) = − u(x)Den (ζm (xn )ψ(x)) dx
Rn
+
Z Z
0
= − u(x)ζm (xn )ψ(x) dx − u(x)ζm (xn )Den ψ(x) dx.
Rn
+ Rn
+
Passing to limit, as m → ∞ both sides, we get

Z Z Z
en 0
D u(x)ψ(x) dx = − lim u(x)ζm (xn )ψ(x) dx − u(x)Den ψ(x) dx.
Rn m→∞ Rn Rn
+ + +
Let us handle the first term in RHS. Note that

Z Z
0
u(x)ζm (xn )ψ(x) dx = m u(x)ζ 0 (mxn )ψ(x) dx
Rn
+ Rn
+
Z
≤ mC |u(x)||ζ 0 ||xn | dx
Rn
+
Z Z
0
≤ Ckζ k∞,[0,1] |u(x)| dx.
Rn−1 0<xn <1/m
0
R
Therefore, limm→∞ Rn
u(x)ζm (xn )ψ(x) dx = 0 and hence
+
Z Z
en
D u(x)ψ(x) dx = − u(x)Den ψ(x) dx.
Rn
+ Rn
+
The RHS in above equation is same as the RHS obtained in (2.5.2). Hence,
en ∗
R en
we have D u (φ) = Rn D u(x)ψ(x) dx. By setting
+
(
Den u(x0 , xn ) xn > 0
(Den u)] (x0 , xn ) :=
−Den u(x0 , −xn ) xn < 0,
we get Den u∗ = (Den u)] . Therefore, we have the estimate,

X
ku∗ k1,p,Rn = ku∗ kp,Rn + kDα u∗ kp,Rn
|α|=1
n−1
X
= 2 1/p
kuk p,Rn
+
+ k(Dei u)∗ kp,Rn + k(Den u)] kp,Rn
i=1
1/p
= 2 kuk1,p,Rn+ .
Hence, u∗ ∈ W 1,p (Rn ).
Theorem 2.5.5. Let Ω be a bounded open subset of Rn with C 1 boundary.

Then there is an extension operator P : W 1,p (Ω) → W 1,p (Rn ).
Corollary 2.5.6. For 1 ≤ p < ∞ and Ω be bounded open set with C 1

boundary, then C ∞ (Ω) is dense in W k,p (Ω), for all k ≥ 1.
2.6 Topological Dual of Sobolev Spaces

We shall now introduce the topological dual of Sobolev spaces and motivate
its notation. For any 1 ≤ p < ∞, k = 0, 1, 2, . . . and Ω be an open subset of
Rn . Let Xk,p (Ω) be the topological dual of W0k,p (Ω). Thus, if F ∈ Xk,p (Ω)
then F is a continuous linear functional on W0k,p (Ω) and the norm of F is
given as,
|F (u)|
kF kXk,p (Ω) := sup .
u∈W
k,p
(Ω)
kuk k,p
0
u6=0
Note that the dual space is considered for W0k,p (Ω) and not for W k,p (Ω). The
reason is that D(Ω) is dense in W0k,p (Ω) and hence W0k,p (Ω) will have a unique
continuous extension (by Hahn-Banach) for any continuous linear functional
defined on D(Ω).
Example 2.11. In general the dual of W k,p (Ω) may not even be a distribution.
Note that, in general, D(Ω) is not dense in W k,p (Ω). Thus, its dual [W k,p (Ω)]∗
is not in the space of distributions D0 (Ω). Of course, the restriction to D(Ω)
of every T ∈ [W k,p (Ω)]∗ is a distribution but this restriction may not identify
with T . For instance, consider f ∈ [L2 (Ω)]n with |f | ≥ c > 0 a.e. and
div(f ) = 0. Define Z
T (φ) := f · ∇φ dx.
Ω
Since |T (φ)| ≤ kf k2 k∇φk2 , we infer that T ∈ [H 1 (Ω)]∗ . However, the restric-
tion of T to D(Ω) is the zero operator of D0 (Ω) because
hT, φi = −hdiv(f ), φi = 0 ∀φ ∈ D(Ω).
Theorem 2.6.1 (Characterisation of X1,p (Ω)). Let 1 ≤ p < ∞ and let
F ∈ X1,p (Ω). Then there exist functions f0 , f1 , . . . , fn ∈ Lq (Ω) such that
Z n Z
X ∂u
F (u) = f0 u dx + fi ∀u ∈ W01,p (Ω)
Ω i=1 Ω
∂x i
and kF kX1,p (Ω) = max0≤i≤n kfi kq . Further, if Ω is bounded, one may assume
f0 = 0.
Proof. Recall that (cf. proof of Theorem 2.2.2) T : W01,p (Ω) → (Lp (Ω))n+1
defined as
∂u ∂u
T (u) = u, ,...,
∂x1 ∂xn
p 1/p
Pn+1
is an isometry. The norm in (Lp (Ω))n+1 is defined as kuk := i=1 kui kp ,
where u = (u1 , u2 , . . . , un+1 ). Let E = T (W0 (Ω)) ⊂ (L (Ω)) . Observe
1,p p n+1
that F ◦ T −1 is a continuous linear functional on E. By Hahn Banach

theorem, there is a continuous extension S of F ◦ T −1 to all of (Lp (Ω))n+1 .
Now, by Riesz representation theorem, there exist f0 , f1 , . . . , fn ∈ Lq (Ω) such
that
Z Xn Z
S(v) = f0 v0 dx + fi vi dx ∀v = (v0 , v1 , . . . , vn ) ∈ (Lp (Ω))n+1
Ω i=1 Ω
and kSk = kF ◦ T −1 k. Now, for any u ∈ W01,p (Ω),

∂u ∂u
F (u) = F ◦ T −1
u, ,...,
∂x1 ∂xn

∂u ∂u
= S u, ,...,
∂x1 ∂xn
Z XZn
∂u
= f0 u dx + fi dx.
Ω i=1 Ω
∂xi
Also, kF k = kF ◦ T −1 k, by isometry of T and hence kF k = kSk =

max0≤i≤n kfi kq .
Further, if Ω is bounded we have by Poincaré inequality that the seminorm
k∇ukp becomes a norm in W01,p (Ω), thus by using the gradient map as an
isometry from W01,p (Ω) to (Lp (Ω))n and arguing as above, we see that f0 can
be chosen to be zero.
Remark 2.6.2. Let F ∈ X1,p (Ω) and φ ∈ D(Ω), then
Z n Z
X ∂φ
F (φ) = f0 φ dx + fi dx
Ω i=1 Ω ∂x i
Z n Z
X ∂fi
= f0 φ dx − φ dx.
Ω i=1 Ω
∂xi
Now, since D(Ω) is dense in W01,p (Ω), the extension of F to W01,p (Ω) should
be unique. Henceforth, we shall identify any element F ∈ X1,p (Ω) with the
distribution n
X ∂fi
f0 − .
i=1
∂x i
The above remark motivates the right notation for the space Xk,p (Ω).
∂u
Observe that if u ∈ W k,p (Ω), then the first derivatives of ∂x i
, for all i, are
k−1,p
in W . To carry forward this feature in our notation, the above remark
motivates to rewrite Xk,p (Ω) as W −k,q (Ω), where q is the conjugate exponent
corresponding to p.
Let us observe that the representation of F in terms fi is not unique. Let
1,q
Pn ∂ ∂g
g ∈ W (Ω) such that ∆g = 0, i.e., i=1 ∂xi ∂xi = 0. Then

n ∂ f + ∂g
X i ∂xi
F = f0 −
i=1
∂xi
is also a representation of F .
Exercise 65. Show that
 !1/q 
n
 X 
kF kW −1,q (Ω) = inf kfi kqq .
P n ∂fi
F =f0 − i=1 ∂xi

i=0

2.7 Fractional Order Sobolev Space

Let 1 ≤ p < ∞ and 0 < σ < 1. The Sobolev spaces W σ,p (Ω), for non-integral
σ, is defined as:
( )
|u(x) − u(y)|
W σ,p (Ω) := u ∈ Lp (Ω) | σ+( n )
∈ Lp (Ω × Ω) (2.7.1)
|x − y| p
with the obvious norm. For any positive real number s, set k := bsc, integral
part and σ := s is the fractional part. Note that 0 < σ < 1.
W s,p (Ω) = {u ∈ W k,p (Ω)|Dα u ∈ W σ,p (Ω) for all |α| = k}. (2.7.2)
0
We denote by W0s,p (Ω), the closure of D(Ω) in W s,p (Ω) and W −s,p (Ω) is dual
of W0s,p (Ω).
We begin by giving a characterisation of the space H 1 (Rn ) in terms of
Fourier transform. Recall that for any u ∈ L1 (Rn ) its Fourier transform û is
defined as Z
û(ξ) = e−ı2πξ·x u(x) dx. (2.7.3)
Rn
Recall that û ∈ C0 (Rn ) and if û ∈ L1 (Rn ), one can invert the Fourier trans-
form to obtain u from û by the following formula:
Z
u(x) = eıξ·x û(ξ) dξ. (2.7.4)
Rn
In particular, by Fourier transform, if u ∈ L1 (Rn ) ∩ L2 (Rn ) then û ∈ L2 (Rn )

and
kûkL2 (Rn ) = kukL2 (Rn ) . (2.7.5)
Theorem 2.7.1. The Sobolev space
1
H 1 (Rn ) = {u ∈ L2 (Rn ) | (1 + |ξ|2 ) 2 û ∈ L2 (Rn )}.
Further, the H 1 is given as

1
kukH 1 (Rn ) = k(1 + |ξ|2 ) 2 ûk2,Rn .
Proof. Consider u ∈ D(Rn ). By definition,

Z
∂u
c ∂u
(ξ) = e−ı2πξ·x (x) dx.
∂xk Rn ∂xk
Using integration by parts and the fact that u has compact support, we
obtain
∂u
c
(ξ) = 2πıξk û(ξ). (2.7.6)
∂xk
By the density of D(Rn ) in H 1 (Rn ) (cf. Theorem 2.3.1), for any u ∈ H 1 (Rn ),
there exists a sequence {um }m∈N ⊂ D(Rn ) such that um → u in the norm
topology of H 1 (Rn ). So, for each m ∈ N,
∂u
dm
(ξ) = 2πıξk ûm (ξ). (2.7.7)
∂xk
∂um ∂u
Since um → u in the H 1 (Rn ) norm, um → u in L2 (Rn ) and ∂xk
→ ∂xk
in
L2 (Rn ), we can extract a subsequence um` such that
ûm` (ξ) → û(ξ) for a.e. ξ ∈ Rn
and
∂u
d m` ∂u
d
(ξ) → (ξ) for a.e. ξ ∈ Rn .
∂xk ∂xk
By the continuity of Fourier transform, we obtain
∂u
d
(ξ) = 2πıξk û(ξ) for a.e. ξ ∈ Rn . (2.7.8)
∂xk
∂u
Now, if u ∈ H 1 (Rn ) then ∂x k
∈ L2 (Rn ), for all k = 1, 2, . . . , n. Thus, the
∂u
Fourier transform of ∂x k
is well-defined and
∂u
d
(ξ) = 2πıξk û(ξ).
∂xk
Hence, ξk û(ξ) ∈ L2 (Rn ), for all k = 1, 2, . . . , n. Conversely, if u ∈ L2 (Rn )

∂u
such that ξk û(ξ) ∈ L2 (Rn ), for all k = 1, 2, . . . , n, then ∂x k
∈ L2 (Rn ) for all
k = 1, 2, . . . , n and, hence, u ∈ H 1 (Rn ). Therefore, u ∈ H 1 (Rn ) if and only
if û ∈ L2 (Rn ) and ıξk û ∈ L2 (Rn ), for all k = 1, 2, . . . , n. This is equivalent
1
to saying that (1 + |ξ|2 ) 2 û ∈ L2 (Rn ) (cf. Lemma 2.7.2). Further,
n
!
Z 2
X ∂u
kuk2H 1 (Rn ) = u2 (x) + (x) dx
R n
k=1
∂xk
 
Z n 2
|û(ξ)|2 +
X ∂u
d
= (ξ)  dξ
R n
k=1
∂x k
Z
1
= (1 + |ξ|2 )|û(ξ)|2 dξ = k(1 + |ξ|2 ) 2 ûk2L2 (Rn ) .
Rn
Lemma 2.7.2. There exist positive constants C1 and C2 depending only on

k and n such that
X
C1 (1 + |ξ|2 )k ≤ |ξ α |2 ≤ C2 (1 + |ξ|2 )k ∀ξ ∈ Rn .
|α|≤k
Proof. Note that |ξ|2 = ξ12 + . . . + ξn2 and |ξ α | = |ξ1 |α1 . . . |ξn |αn . By induction
argument on k, we can see that same powers of ξ occur in (1 + |ξ|2 )k and
α 2
P
|α|≤k |ξ | , albeit with different coefficients, which depend only on n and
k. Since the number of terms is finite and depends again only on n and k,
the result follows.
Owing to the above lemma one can define the space H k (Rn ) as follows:
k
H k (Rn ) = {u ∈ L2 (Rn )|(1 + |ξ|2 ) 2 û(ξ) ∈ L2 (Rn )}
and Z 21
2 k 2
kukH k (Rn ) = (1 + |ξ| ) |û(ξ)| dξ .
Rn
A major interest of the above approach is that it suggests a natural def-

inition of the space H s (Rn ), for s ∈ R. The central point being that the
integral Z
(1 + |ξ|2 )s |û(ξ)|2 dξ
Rn
is finite, for any s ∈ R, since as makes sense for any s ∈ R when a > 0.
In this case, a = 1 + |ξ|2 is positive. Thus, we are motivated to give the
following definition.
Definition 2.7.3. Let s ≥ 0 be non-negative real number. We define

s
H s (Rn ) = {u ∈ L2 (Rn ) | (1 + |ξ|2 ) 2 û ∈ L2 (Rn )},
equipped with the scalar product, for any u, v ∈ H s (Rn )

Z
hu, viH s (Rn ) := (1 + |ξ|2 )s û(ξ)v̂(ξ) dξ
Rn
and the corresponding norm,

Z 21
2 s 2
kukH s (Rn ) := (1 + |ξ| ) |û(ξ)| dξ .
Rn
Theorem 2.7.4. For any s ∈ [0, ∞), H s (Rn ) is a Hilbert space. If s = k ∈ N

then H s (Rn ) = H k (Rn ) = W k,2 (Ω) is the classical Sobolev space.
Proof. Recall that the Fourier transfrom is an isomorphism from H s (Rn )

onto Lebesgue space L2µ (Rn ) where µ := (1 + |ξ|2 )s dx is the measure with
density (1 + |ξ|2 )s w.r.t the Lebesgue measure dx on Rn . Moreover, the
Fourier transform is an isometry and the Hilbert structure of the weighted
Lebesgue space L2µ (Rn ) is passed on to H s (Rn ). Hence, H s (Rn ) is isomorphic
to L2µ (Rn ).
For s > 0, we define H −s (Rn ) as the dual of H s (Rn ). The negative order
Sobolev spaces has the following characterization:
Theorem 2.7.5. Let s ∈ (0, ∞). Then
s
H −s (Rn ) = {u ∈ S 0 (Rn ) | (1 + |ξ|2 )− 2 û(ξ) ∈ L2 (Rn )} (2.7.9)
Proof. We shall give the proof for s = 1. If u ∈ H −1 (Rn ) then
n
X ∂fi
u = f0 + , f0 , f1 . . . , fn ∈ L2 (Rn ).
i=1
∂xi
Hence, u is a tempered distribution and
n
X
û = fˆ0 + (2πı)ξi f̂i .
i=1
1
Then (1 + |ξ|2 )− 2 û ∈ L2 (Rn ), proving one inclusion in (2.7.9). To prove the
1
reverse inclusion, consider u ∈ S 0 (Rn ) such that (1 + |ξ|2 )− 2 û(ξ) ∈ L2 (Rn ).
Let φ ∈ D(Rn ). Then there exists ψ ∈ S(Rn ) such that φ = ψ̂. Set
1 1
k(ξ) := (1 + |ξ|2 ) 2 and k−1 (ξ) := (1 + |ξ|2 )− 2 .
Note that both k and k−1 are in C ∞ (Rn ), we write
u(φ) = u(ψ̂) = û(ψ) = (kk−1 )û(ψ) = k−1 û(kψ).
But k−1 û ∈ L2 (Rn ) and, hence,
Z
1 1
u(φ) = (1 + |ξ|2 )− 2 û(ξ)(1 + |ξ|2 ) 2 ψ(ξ) dξ.
Rn
Therefore,
1 1
|u(φ)| ≤ |(1 + |ξ|2 )− 2 û(ξ)|0,Rn |(1 + |ξ|2 ) 2 ψ(ξ)|0,Rn .
But
2
Z
1
2
(1 + |ξ| ) ψ(ξ) 2 = (1 + |ξ|2 )ψ 2 (ξ) dξ
0,Rn Rn
Z
= (1 + |ξ|2 )ψ 2 (−ξ) dξ
Z Rn
ˆ
= (1 + |ξ|2 )ψ̂ 2 (ξ) dξ
n
ZR
= (1 + |ξ|2 )(φ̂(ξ))2 dξ = kφk2H 1 (Rn ) .
Rn
Thus u defines a continuous linear functional on H 1 (Rn ) and so u ∈ H −1 (Rn ).

Also,
1
kukH −1 (Rn ) = |(1 + |ξ|2 )− 2 û(ξ)|0,Rn .
For an open subset Ω of Rn , we may define the Sobolev spaces H s (Ω), for
real s, as the restrictions to Ω of elements of H s (Rn ).
Example 2.12. If δ0 is the Dirac distribution, we know that δˆ0 ≡ 1 and, hence,
s
δ0 ∈ H −s (Rn ) if and only if (1 + |ξ|2 )− 2 ∈ L2 (Rn ). This is true for s > n2
since the integral in polar coordinates is
∞
rn−1
Z
dr
0 (1 + r2 )s
is finite only when s > n2 .
2.8 Imbedding Results

In this section, we shall prove various inequalities which, in turn, implies
the imbedding of Sobolev spaces in to classical spaces. By the definition
of Sobolev spaces and their norms, we observe that W k,p (Ω) ,→ Lp (Ω) is
continuously imbedded (cf. Definition 2.1.6). Let us make a remark on the
intuitive importance of Sobolev norms. Recall the uniform norm or essen-
tial supremum on the space of bounded continuous functions Cb (Ω). One
can, intuitively, think of essential supremum norm as capturing the “height”
of the function. In a similar sense, the Lp norms, for p < ∞, capture the
“height” and “width” of a function. In mathematical terms, “width” is same
as the measure of the support of the function. The Sobolev norms captures
“height”, “width” and “oscillations”. The Fourier transform measures oscil-
lation (or frequency or wavelength) by the decay of the Fourier transform,
i.e., the “oscillation” of a function is translated to “decay” of the its Fourier
transform. Sobolev norms measures “oscillation” via its derivatives (or reg-
ularity). Thus, Sobolev imbedding results are precisely statements about
functions after incorporating its “oscillation” information.
We shall restrict ourselves to W 1,p , for all 1 ≤ p ≤ ∞, to make the
presentation clear and later state the results for the spaces W k,p , k ≥ 2.
Recall that the Lp spaces are actually equivalence classes of functions with
equivalence relation being “equality almost everywhere”. This motivates the
following definition.
Definition 2.8.1. For any u ∈ Lp (1 ≤ p ≤ ∞), we say u? is a representa-

tive6 of u if u = u? a.e.
Theorem 2.8.2 (One dimensional case). Let (a, b) ⊆ R be an open interval

and 1 ≤ p ≤ ∞. If u ∈ W 1,p (a, b) then there is a representative of u, u? ,
which is absolutely continuous (is in AC(a, b)).
Proof. Since u ∈ W 1,p (a, b), u is weakly differentiable and u0 ∈ Lp (a, b). For
each x ∈ (a, b), we define v : (a, b) → R as
Z x
v(x) := u0 (t) dt.
a
By definition, v ∈ BV (a, b) and is differentiable a.e. Hence v 0 = u0 and

(v − u)0 = 0 a.e. Thus, u = v − c a.e., where c is some constant. Set
u? := v − c. We claim that u? is absolutely continuous because
Z b Z b Z b
0 0
? ?
u (b) − u (a) = v(b) − v(a) = u (t) dt = v (t) dt = (u? )0 (t) dt.
a a a
Thus, u? is the absolutely continuous representative of u.

Example 2.13. In the above result, we have shown that W 1,p (a, b) ⊂ AC(a, b),
in the sense of representatives. Let us give an example of a continuous
function on a bounded interval I ⊂ R which does not belong to H 1 (I).
Take I = (−1, 1). For non-zero γ in the range − 21 < γ ≤ 12 , the function
|x|γ 6∈ H 1 (I) but is in L2 (I). If 0 < γ ≤ 12 then |x|γ is continuous function
and not in H 1 (I).
2.8.1 Sobolev Inequality (1 ≤ p < n)

If there exists a positive constant C > 0 such that
kukr,Rn ≤ Ck∇ukp,Rn ∀u ∈ W 1,p (Rn ) (2.8.1)

6
This is not a standard usage in literature and is introduced by the author of this
manuscript for convenience sake
for some r ∈ [1, ∞) and 1 ≤ p < ∞, then we have a continuous imbedding

of W 1,p (Ω) in to Lr (Rn ) because
kukr,Rn ≤ Ck∇ukp,Rn ≤ Ck∇ukp,Rn + Ckukp,Rn = Ckuk1,p,Rn .
However, it is obvious that (2.8.1) is not a sufficient condition for continuous

imbedding. The equation (2.8.1) is a Sobolev inequality, a stronger necessary
condition for continuous imbedding.
Before we prove an inequality like (2.8.1), let’s check it validity. When can
we expect such an inequality? If u ∈ W 1,p (Rn ) satisfies (2.8.1) for some C > 0
and r, then uλ (x) := u(λx), for any λ > 0, also satisfies (2.8.1) for the same
1 λ
C and r. Since, for 1 ≤ p < ∞, kuλ kp = λn/p kukp and k∇uλ kp = λn/p k∇ukp ,
we have
1 λ
kukr ≤ C k∇ukp
λn/r λn/p
1+ n −n
kukr ≤ Cλ r p k∇ukp .
The above obtained inequality being true for λ > 0 would contradict (2.8.1)
except when 1 + nr − np = 0. Consequently, to expect an inequality of the
kind (2.8.1), we need to have 1r = p1 − n1 . Therefore, p1 − n1 > 0 and hence
np
1 ≤ p < n and r = n−p . This discussion motivates the definition of Sobolev
conjugate, p? , corresponding to the index p.
Definition 2.8.3. If 1 ≤ p < n, the Sobolev conjugate of p is defined as
np
p? :=
n−p
1 1
Equivalently, p?
= p
− n1 . Also, p? > p.
Lemma 2.8.4 (Loomis-Whitney Inequality). Let n ≥ 2. Let f1 , f2 , . . . , fn ∈

Ln−1 (Rn−1 ). For x ∈ Rn , set x̂i = (x1 , x2 , . . . , xi−1 , xi+1 , . . . , xn ) ∈ Rn−1 for
1 ≤ i ≤ n. Define f (x) = f1 (xˆ1 ) . . . fn (xˆn ) for x ∈ Rn . Then f ∈ L1 (Rn )
and
Y n
kf k1,Rn ≤ kfi kn−1,Rn−1 .
i=1
Proof. Let n = 2, then

Z Z
kf k1,R2 = |f (x1 , x2 )| dx1 dx2 = |f1 (x2 )||f2 (x1 )| dx1 dx2
R2 ZR 2
Z
= |f2 (x1 )| dx1 |f1 (x2 )| dx2
R R
= kf2 k1,R kf1 k1,R .
Let n = 3, then using Cauchy-Schwarz inequality twice we get,
Z
kf (x)k1,R3 = |f1 (x2 , x3 )||f2 (x1 , x3 )||f3 (x1 , x2 )| dx1 dx2 dx3
R3
Z Z
= |f3 (x1 , x2 )| |f1 (x2 , x3 )||f2 (x1 , x3 )| dx3 dx1 dx2
R2 R
Z " 2 Z
Y 21 #
≤ |f3 (x1 , x2 )| |fi (x̂i )|2 dx3 dx1 dx2
R2 i=1 R
Z
= |f3 (x1 , x2 )|g(x2 )1/2 h(x1 )1/2 dx1 dx2
R2
Z 1/2 Z 1/2
2
≤ |f3 (x1 , x2 )| dx1 dx2 g(x2 )h(x1 ) dx1 dx2
R2 R2
= kf3 k2,R2 kf2 k2,R2 kf1 k2,R2 .
The general case will be proved by induction. Assume the result for n.
Let x ∈ Rn+1 . Fix xn+1 and x0 = (x1 , . . . , xn ), then by Hölder’s inequality,
Z Z 1/n0
0 n0
|f (x)| dx ≤ kfn+1 kn,Rn |f1 f2 . . . fn | dx1 dx2 . . . dxn (2.8.2)
Rn Rn
0 n
where n = n−1
is the conjugate exponent of n. Recall that f1 , . . . , fn ∈
0 0
L (R ). Thus, by treating xn+1 as a fixed parameter, |f1 |n , . . . , |fn |n ∈
n n
Ln−1 (Rn−1 ). Therefore, by induction hypothesis,

Z n Z 1/(n−1)
n0 n0 (n−1)
Y
|f1 f2 . . . fn | dx1 . . . dxn ≤ |fi | dx1 . . . dxn
Rn i=1 Rn−1
Yn Z n0 /n
n
= |fi | dx1 . . . dxn
i=1 Rn−1
n
0
Y
= kfi knn,Rn−1
i=1
Now, substituting above inequality in (2.8.2), we get

Z n
Y
0
|f (x)| dx ≤ kfn+1 kn,Rn kfi kn,Rn−1 .
Rn i=1
Integrate both sides with respect to xn+1 . We get,

Z n
Y
|f (x)| dx ≤ kfn+1 kn,Rn kfi kn,Rn .
Rn+1 i=1
Theorem 2.8.5 (Gagliardo-Nirenberg-Sobolev Inequality). Let 1 ≤ p < n.

Then there exists a constant C > 0 (depending on p and n) such that
kukp? ,Rn ≤ Ck∇ukp,Rn ∀u ∈ W 1,p (Rn ).
In particular, we have the continuous imbedding

?
W 1,p (Rn ) ,→ Lp (Rn ).
n
Proof. We begin by proving for the case p = 1. Note that p? = 1? = n−1 .
We first prove the result for the space of test functions which is dense in
W 1,p (Rn ). Let φ ∈ D(Rn ), x ∈ Rn and 1 ≤ i ≤ n, then
Z xi
φ(x) = Dei φ(x1 , . . . , xi−1 , t, xi+1 , . . . , xn ) dt
Z−∞
∞
|φ(x)| ≤ |Dei φ(x1 , . . . , xi−1 , t, xi+1 , . . . , xn )| dt =: fi (x̂i )
−∞
n
Y
|φ(x)|n ≤ fi (x̂i )
i=1
Yn
|φ(x)|n/n−1 ≤ |fi (x̂i )|1/n−1 .
i=1
Now, integrating both sides with respect to x, we have

Z Z n
Y
n/n−1
|φ| dx ≤ |fi |1/n−1 dx. (2.8.3)
Rn Rn i=1
Observe that gi := |fi |1/n−1 ∈ Ln−1 (Rn−1 ) for each 1 ≤ i ≤ n. Hence, by

Lemma 2.8.4, we have
Z n
Y Z n
Y n Z
Y 1/n−1
1/n−1 n−1
|fi | dx = |gi | dx ≤ |gi | dx
Rn i=1 Rn i=1 i=1 Rn−1
Yn Z 1/n−1
= |fi | dx
i=1 Rn−1
Yn n
Y
1/n−1 1/n−1
= kfi k1,Rn−1 = kDei φk1,Rn .
i=1 i=1
Thus, substituting above inequality in (2.8.3), we get

n
Y 1/n
kφk1? ,Rn ≤ kDei φk1,Rn
i=1
and consequently, we get

n
Y n
Y
ei 1/n 1/n
kφk1? ,Rn ≤ kD φk1,Rn ≤ k∇φk1,Rn = k∇φk1,Rn .
i=1 i=1
Hence the result proved for p = 1. Let ψ := |φ|γ , where γ > 1 will be chosen
appropriately during the subsequent steps of the proof. Also, if φ ∈ D(Rn ),
then ψ ∈ Cc1 (Rn ). We shall apply the p = 1 result to ψ. Therefore,
kψk1? ,Rn ≤ k∇ψk1,Rn

Z n−1/n Z
nγ/n−1
|φ| dx ≤ |∇|φ|γ | dx
Rn Rn
Z
= γ |φ|γ−1 |∇φ| dx
Rn
Z 1/q Z 1/p
(γ−1)q p
≤ γ |φ| dx |∇φ| dx
Rn Rn
(using Hölder’s Inequality).
Since we want only the gradient term on the RHS, we would like to bring the
nγ
q norm term to LHS. If we choose γ such that n−1 = (γ − 1)q, then we can
club their powers. Thus, we get γ := p(n−1)

n−p
. The fact that p > 1 implies that
γ > 1, as we had demanded. Thus, the inequality obtained above reduces to
p(n − 1)
kφkp? ,Rn ≤ k∇φkp,Rn ∀φ ∈ D(Rn ).
n−p
Now, for any u ∈ W 1,p (Rn ), there is a sequence {um } ⊂ D(Rn ) (cf.
Theorem 2.3.1) such that um → u in W 1,p (Rn ). Therefore, um is Cauchy
?
in W 1,p (Rn ) and hence is Cauchy in Lp (Rn ) by above inequality. Since
? ?
Lp (Rn ) is complete, um converges in Lp (Rn ) and should converge to u,
?
since um = φm (ρm ∗ u) (cf. Theorem 2.3.1). Thus, u ∈ Lp (Rn ) and the
inequality is satisfied. Hence the theorem is proved for any 1 ≤ p < n. In
fact, in the proof we have obtained the constant C to be C = p(n−1)
n−p
(and
this is not the best constant).
Corollary 2.8.6. For any 1 ≤ p < n, W 1,p (Rn ) ,→ Lr (Rn ) is continuously
imbedded, for all r ∈ [p, p? ].
?
Proof. Let u ∈ W 1,p (Rn ). By Theorem 2.8.5, we have u ∈ Lp (Rn ). We need
to show that u ∈ Lr (Rn ) for any r ∈ (p, p? ). Since 1/r ∈ [1/p? , 1/p], there is
a 0 ≤ λ ≤ 1 such that 1/r = λ/p + (1 − λ)/p? . Consider,
Z Z
r r
kukr = |u| dx = |u|λr |u|(1−λ)r dx
Rn Rn
Z λrp Z (1−λ)r
p?
p p?
≤ |u| dx |u| dx (by Hölder inequality)
Rn Rn
kukr ≤ kukλp kuk1−λ
p?
≤ λkukp + (1 − λ)kukp? (By generalised AM-GM inequality)
≤ kukp + kukp?
≤ kukp + Ck∇ukp (By Theorem 2.8.5)
≤ max{C, 1}kuk1,p .
Hence the continuous imbedding is shown for all r ∈ [p, p? ].
We now extend the Sobolev inequality for a proper subset Ω ⊂ Rn with
smooth boundary, using the extension operator.
Theorem 2.8.7 (Sobolev Inequality for a Subset). Let Ω be an open subset
?
of Rn with C 1 boundary. Also, let 1 ≤ p < n. Then W 1,p (Ω) ,→ Lp (Ω) where
the constant C obtained depends on p, n and Ω. Further, W 1,p (Ω) ,→ Lr (Ω)
for all r ∈ [p, p? ].
Proof. Let C denote a generic constant in this proof. Since Ω has a C 1

boundary, by Theorem 2.5.5, there is an extension operator P such that for
some constant C > 0 (depending on Ω and p)
kP uk1,p,Rn ≤ Ckuk1,p,Ω ∀u ∈ W 1,p (Ω).
Moreover, by Theorem 2.8.5, there exists a constant C > 0 (depending on p
and n) such that
kP ukp? ,Rn ≤ Ck∇(P u)kp,Rn ∀u ∈ W 1,p (Ω).
Let u ∈ W 1,p (Ω), then
kukp? ,Ω ≤ kP ukp? ,Rn ≤ Ck∇(P u)kp,Rn ≤ CkP uk1,p,Rn ≤ Ckuk1,p,Ω .
where the final constant C is dependent on p, n and Ω. Similar ideas work
to prove the continuous imbedding in Lr (Ω) for r ∈ [p, p? ].
Note that the above result says
kukp? ≤ Ckuk1,p ∀u ∈ W 1,p (Ω)
and not
kukp? ≤ Ck∇ukp ∀u ∈ W 1,p (Ω)
because constant functions may belong to W 1,p (Ω) (as happens for bounded
subsets Ω) whose derivatives are zero.
Corollary 2.8.8 (For W01,p (Ω)). Let Ω be an open subset of Rn (not neces-
sarily bounded) and 1 ≤ p < n. Then there is a constant C > 0 (depending
on p and n) such that
kukp? ≤ Ck∇ukp ∀u ∈ W01,p (Ω)
and
kukr ≤ Ckuk1,p ∀u ∈ W01,p (Ω), r ∈ [p, p? ].
Proof. Use the fact that D(Ω) is dense in W01,p (Ω) and follow the last step
of the proof of Theorem 2.8.5 to get
kukp? ≤ Ck∇ukp ∀u ∈ W01,p (Ω).
The proof of the second inequality remains unchanged from the case of Rn .
One can also extend functions in W01,p (Ω) by zero outside Ω and use the
results proved for Rn and then restrict back to Ω.
Note that the second inequality in the statement of the above corollary
involves the W 1,p -norm of u and not the Lp -norm of the gradient of u. How-
ever, for bounded sets one can hope to get the inequality involving gradient
of u.
Corollary 2.8.9. Let Ω be a bounded open subset of Rn . For 1 ≤ p < n

there is a constant C (depending on p, n, r and Ω) such that
kukr ≤ Ck∇ukp ∀u ∈ W01,p (Ω) ∀r ∈ [1, p? ].
Proof. Let 1 ≤ p < n and u ∈ W01,p (Ω). Then, by previous corollary, there
is a constant C > 0 (depending on p and n) kukp? ≤ Ck∇ukp . For any
1 ≤ r ≤ p? , there is a constant C > 0 (depending on r and Ω) such that
kukr ≤ Ckukp? (since Ω is bounded). Therefore, we have a constant C > 0
(depending on p, n, r and Ω) kukr ≤ Ck∇ukp for all r ∈ [1, p? ].
Exercise 66 (Poincaré Inequality). Using the corollary proved above show

that, for 1 ≤ p ≤ ∞, then there is a constant C (depending on p and Ω) such
that
kukp ≤ Ck∇ukp ∀u ∈ W01,p (Ω).
2.8.2 Poincaré Inequality

The Poincaré inequality proved in previous section can be directly proved
without using the results of previous section, as shown below.
Theorem 2.8.10 (Poincaré Inequality). Let Ω be a bounded open subset of

Rn , then there is a constant C (depending on p and Ω) such that
kukp,Ω ≤ Ck∇ukp,Ω ∀u ∈ W01,p (Ω). (2.8.4)
Proof. Let a > 0 and suppose Ω = (−a, a)n . Let u ∈ D(Ω) and x = (x0 , xn ) ∈
Rn . Then
Z xn
∂u 0
u(x) = (x , t) dt.
−a ∂xn
Moreover, u(x0 , −a) = 0. Thus, by Hölder’s inequality,

Z xn p 1/p
∂u 0
|u(x)| ≤ (x , t) dt |xn + a|1/q
−a ∂x n
Z a p
p p/q ∂u 0
|u(x)| ≤ |xn + a| (x , t) dt
−a ∂xn
Z a p
p p/q ∂u 0
|u(x)| ≤ (2a) (x , t) dt.
−a ∂xn
First integrating w.r.t x0 and then integrating w.r.t xn we get

p
∂u
kukpp,Ω ≤ (2a) (p/q)+1
, ∀u ∈ D(Ω)
∂xn p,Ω
and taking (1/p)-th power both sides, we get
∂u
kukp,Ω ≤ 2a , ∀u ∈ D(Ω).
∂xn p,Ω
Thus,
∂u
kukp,Ω ≤ 2a ≤ 2ak∇ukp,Ω , ∀u ∈ D(Ω).
∂xn p,Ω
By the density of D(Ω) in W01,p (Ω), we get the result for all u ∈ W01,p (Ω).
Now, suppose Ω is not of the form (−a, a), then Ω ⊂ (−a, a) for some
a > 0, since Ω is bounded. Then any u ∈ W01,p (Ω) can be extended to
W01,p (−a, a) and use the result proved above.
The Poincaré inequality makes the norm k∇ukp equivalent to kuk1,p in

W01,p (Ω).
Remark 2.8.11. Poincaré inequality is not true for u ∈ W 1,p (Ω). For in-
stance, if u ≡ c, a constant, then ∇u = 0 and hence k∇ukp = 0 while
kukp > 0. However, if u ∈ W 1,p (Ω) such that u = 0 on Γ ⊂ ∂Ω, then
Poincaré inequality remains valid for such u’s.
Remark 2.8.12. Poincaré inequality is not true for unbounded domains.

However, one can relax the bounded-ness hypothesis on Ω to bounded-ness
along one particular direction, as seen from the proof.
2.8.3 Equality case, p = n

Theorem 2.8.13. Let n ≥ 2, W 1,n (Rn ) ,→ Lr (Rn ) for all r ∈ [n, ∞).
Proof. Let u ∈ D(Rn ). Observe that the conjugate exponent of n is same as
1? , the Sobolev conjugate of 1. Thus, we adopt the initial part of the proof
of Theorem 2.8.5 to obtain
kuk1? ,Rn ≤ k∇uk1,Rn

n
where 1? = n−1 , the conjugate exponent of n. Applying this inequality to
γ
|u| , for some γ > 1 as in Theorem 2.8.5, we get
Z n−1
n
Z n−1
n
Z n1
nγ (γ−1)n
n
|u| n−1 dx ≤ γ |u| n−1 dx |∇u| dx
Rn Rn Rn
kukγnγ γ−1
≤ γkuk n(γ−1) k∇ukn
n−1 n−1
γ
≤ kuk n(γ−1) + k∇ukn
n−1
Using (a + b)γ ≥ γaγ−1 b for a, b ≥ 0

kuk n−1
nγ ≤ kuk n(γ−1) + k∇ukn .
n−1
Now, by putting γ = n, we get
kuk n2 ≤ kukn + k∇ukn = kuk1,n .

n−1
and extending the argument, as done in Corollary 2.8.6, we get

n2

kukr ≤ Ckuk1,n ∀r ∈ n, .
n−1
Now, repeating the argument for γ = n + 1, we get
2
n n(n + 1)
kukr ≤ Ckuk1,n ∀r ∈ , .
n−1 n−1
Thus, continuing in similar manner for all γ = n + 2, n + 3, . . . we get
kukr ≤ Ckuk1,n ∀r ∈ [n, ∞).
The result extends to W 1,n (Rn ) by similar density arguments of D(Rn ), as

done for the 1 ≤ p < n case.
We now extend the results to proper subsets of Rn . The proofs are similar
to the equivalent statements from previous section.
Corollary 2.8.14 (For Subset). Let Ω be a bounded open subset of Rn with

C 1 boundary and let n ≥ 2. Then W 1,n (Ω) ,→ Lr (Ω) for all r ∈ [n, ∞).
Corollary 2.8.15. Let Ω be an open subset of Rn (not necessarily bounded)

and n ≥ 2. Then W01,n (Ω) ,→ Lr (Ω) for all r ∈ [n, ∞). Further, if Ω is
bounded, W01,n (Ω) ,→ Lr (Ω) for all r ∈ [1, ∞).
2.8.4 Morrey’s Inequality (n < p < ∞)

Theorem 2.8.16 (Morrey’s Inequality). Let n < p < ∞, then W 1,p (Rn ) ,→
L∞ (Rn ). Moreover, for any u ∈ W 1,p (Rn ), there is a representative of u,
u? , which is Hölder continuous with exponent 1 − n/p and there is a constant
C > 0 (depending only on p and n) such that
ku? kC 0,γ (Rn ) ≤ Ckuk1,p
where γ := 1 − n/p.
Proof. Let u ∈ D(Rn ) and let E be a cube of side a containing the origin
and each of its sides being parallel to the coordinate axes of Rn . Let x ∈ E.
We have
Z 1 Z 1
d
|u(x) − u(0)| = (u(tx)) dt = ∇u(tx) · x dt
0 dt 0
Z 1 Z 1 n
X ∂
≤ |∇u(tx)| · |x| dt = |xi | u(tx) dt
0 0 i=1
∂xi
Z 1 n Z 1 n
X ∂ X ∂
≤ a u(tx) dt = a u(tx) dt
0 i=1
∂xi 0 i=1
∂xi
Let u denote the average of u over the cube E,

Z Z
1 1
u= u(x) dx = n u(x) dx.
|E| E a E
Consider,
Z Z
1 1
|u − u(0)| = u(x) dx − u(0) ≤ n |u(x) − u(0)| dx
an E a E
Z Z 1X n
a ∂
≤ u(tx) dt dx
an E 0 i=1 ∂xi
n Z
1 X 1
Z
∂
= n−1
u(tx) dx dt (Fubini’s Theorem)
a i=1 0 E ∂xi
n Z
1 X 1 −n
Z
∂
= n−1
t u(y) dy dt (Change of variable)
a i=1 0 tE ∂xi
n Z
1 X 1 −n ∂u
≤ t (|tE|)1/q dt
an−1 i=1 0 ∂xi p,E
(By Hölder inequality and tE ⊂ E for 0 ≤ t ≤ 1)
n/q Z 1
a
= n−1 k∇ukp,E t−n tn/q dt (q is conjugate exponent of p)
a 0
Z 1
a1−n/p
= a1−n/p k∇ukp,E t−n/p dt = k∇ukp,E .
0 1 − n/p
The above inequality is then true for any cube E of side length a with sides
parallel to axes, by translating it in Rn . Therefore, for any cube E of side a
and x ∈ E, we have
aγ
|u − u(x)| ≤ k∇ukp,E (2.8.5)
γ
where γ := 1 − n/p. Consider,
1
|u(x)| = |u(x) − u + u| ≤ k∇ukp,E + kuk1,E
γ
1
≤ k∇ukp,E + kukp,E (by Hölder’s inequality)
γ
where we have used (2.8.5), in particular, for a unit cube E. Then
1
kuk∞,Rn ≤ k∇ukp,Rn + kukp,Rn .
γ
Therefore,
kuk∞,Rn ≤ Ckuk1,p,Rn , (2.8.6)
where C = max{1, γ1 }, and hence W 1,p (Rn ) ,→ L∞ (Rn ).

Further, it follows from (2.8.5) that for any x, y ∈ E,
2aγ
|u(x) − u(y)| ≤ k∇ukp,E .
γ
Now, for any given x, y ∈ Rn , one can always choose a cube E whose side
a = 2|x − y| and applying the above inequality, we get
2γ+1 γ 2γ+1
|u(x) − u(y)| ≤ |x − y| k∇ukp,E ≤ |x − y|γ k∇ukp,Rn .
γ γ
Thus, u is Hölder continuous and its Hölder seminorm pγ (.) (cf.(2.1.2)) is
bounded as below,
2γ+1
pγ (u) ≤ k∇ukp,Rn .
γ
and this together with (2.8.6) gives the bound for the γ-th Hölder norm,
kukC 0,γ (Rn ) ≤ Ckuk1,p,Rn .
By the density of D(Rn ) in W 1,p (Rn ) we have a sequence uε → u in W 1,p (Rn ).

By the bound on Hölder norm, we find the sequence is also Cauchy in
C 0,γ (Rn ) and should converge to a representative of u, u? , in the γ-th Hölder
norm.
Remark 2.8.17. As usual, the results can be extended to W 1,p (Ω) for Ω
bounded with C 1 smooth boundary and to W01,p (Ω) for any open subset Ω.
Theorem 2.8.18 (Characterisation of W 1,∞ ). For any u ∈ W 1,∞ (Rn ) there

is a representative u? which is Lipschitz continuous from Rn to R.
Example 2.14. Let p < n. Consider the function |x|δ , for any choice of δ in
1 − np < δ < 0, is in W 1,p (Ω) which has no continuous representative.
Example 2.15. Let p = n and n ≥ 2. We shall give an example of a function in
W 1,n (Ω) which has no continuous representative. We shall given an example
for the case n = 2. Let Ω := {x ∈ Rn : |x| < R} and u(x) = (− ln |x|)δ for
x 6= 0. We have, using polar coordinates,
Z Z R
n n−1
u dx = R ωn (− ln r)nδ r dr.
BR (0) 0
Using the change of variable t = − ln r, we get

Z Z +∞
n
u dx = R n−1
ωn tnδ e−2t dt < ∞ for all δ.
BR (0) − ln R
Thus, u ∈ Lp (Ω) for every δ ∈ R. Further, for each i = 1, 2, . . . , n,
uxi = −δxi |x|−2 (− ln |x|)δ−1
and, therefore,
|∇u| = δ(− ln |x|)δ−1 |x|−1 .
Thus, using polar coordinates, we get
Z Z R
n n−1 n
|∇u| dx = R ωn |δ| | ln r|nδ−n r1−n dr.
BR (0) 0
Using the change of variable t = − ln r, we get

Z Z ∞
n n−1 n
|∇u| dx = R ωn |δ| |t|n(δ−1) et(n−2) dt.
BR (0) − ln R
If n = 2 then the integral is finite iff n(1 − δ) > 1 or δ < 1 − n1 = 12 . In

particular, ∇u represents the gradient of u in the distribution sense as well.
We conclude that u ∈ H 1 (Ω) iff δ < 1/2. We point out that when δ > 0,
u is unbounded near 0. Thus, by taking 0 < δ < 21 , we obtain a function
belonging to H 1 (Ω) which blows up to +∞ at the origin and which does not
have a continuous representative.
2.8.5 Generalised Sobolev Imbedding

We shall now generalise the results of previous section to all derivative orders
of k ≥ 2.
Theorem 2.8.19. Let k ≥ 1 be an integer and 1 ≤ p < ∞. Then
1. If p < n/k, then W k,p (Rn ) ,→ Lr (Rn ) for all r ∈ [p, np/(n − pk)].
2. If p = n/k, then W k,n/k (Rn ) ,→ Lr (Rn ) for all r ∈ [n/k, ∞).

3. If p > n/k, then W k,p (Rn ) ,→ L∞ (Rn ) and further there is a represen-
tative of u, say u? , whose k-th partial derivative is Hölder continuous
with exponent γ and there is a constant C > 0 (depending only on p, k
and n) such that
ku? kC k−1−[n/p],γ (Rn ) ≤ Ckukk,p
where γ := k − n/p − [k − n/p] and [l] is the largest integer such that
[l] ≤ l.
Proof. 1. If k = 1 we know the above results are true. Let k = 2 and
p < n/2. Then, for any u ∈ W 2,p (Rn ), both u, D1 u ∈ W 1,p (Rn ) and
since p < n/2 < n, using Sobolev inequality, we have both u, D1 u ∈
? ?
Lp (Rn ) with p? = np/(n − p). Thus, u ∈ W 1,p (Rn ). Now, since
p < n/2, we have p? < n. Thus, using Sobolev inequality again again,
? ?
we get u ∈ L(p ) (Rn ). But 1/(p? )? = 1/p − 2/n. Extending similar
arguments for each case, we get the result. Note that when we say D1 u,
we actually mean Dα u for each |α| = 1. Such convention will be used
throughout this proof.
2. We know the result for k = 1. Let k = 2 and u ∈ W 2,n/2 (Rn ).

Then, u, D1 u ∈ W 1,n/2 (Rn ). Since n/2 < n and (n/2)? = n, we
have, using Sobolev inequality, u ∈ W 1,r (Rn ) for all r ∈ [n/2, n].
Thus, u ∈ W 1,n (Rn ), which is continuously imbedded in Lr (Rn ) for
all r ∈ [n, ∞).
3. Let k = 2 and p > n/2. W 2,p (Rn ) is continuously imbedded in

W 1,p (Rn ). If p > n > n/2, using Morrey’s inequality, we have W 1,p (Rn )
is continuously imbedded in L∞ (Rn ) and the Hölder norm estimate is
true for both u and ∇u. If n/2 < p < n then p? > n and, by Sobolev
inequality, any u ∈ W 2,p (Rn ) is also in W 1,r (Rn ) for all r ∈ [p, p∗ ].
?
Now, by Morrey’s inequality, W 1,p (Rn ) is continuously imbedded in
L∞ (Rn ) and the Hölder norm estimate is true for both u and ∇u. Let
p = n > n/2 then for any u ∈ W 2,n (Rn ), we have u ∈ W 1,r (Rn ) for
all r ∈ [n, ∞). Thus for r > n, we have the required imbedding in
L∞ (Rn ). It now only remains to show the Hölder estimate.
Remark 2.8.20. As usual, the results can be extended to W 1,p (Ω) for Ω
bounded with C k smooth boundary and to W01,p (Ω) for any open subset Ω.
2.8.6 Compact Imbedding

Our aim in this section is to isolate those continuous imbedding which are
also compact. We first note that we cannot expect compact imbedding for
unbounded domains.
Example 2.16. We shall construct a bounded sequence in W 1,p (R) and show
that it can not converge in Lr (R), for all those r for which the imbedding is
continuous. Let I = (0, 1) ⊂ R and Ij := (j, j+1), for all j = 1, 2, . . .. Choose
any f ∈ C 1 (R) with support in I and set fj (x) := f (x−j). Thus, fj is same as
f except that its support is now contained in Ij . Hence kf k1,p = kfj k1,p for all
f
j. Now, set g := kf kf1,p and gj := kfj kj1,p . Note that {gj } is a bounded sequence
(norm being one) in W 1,p (R). We know that W 1,p is continuously imbedded
in L∞ (R) for 1 < p < ∞ and with p = 1 is imbedded in Lr (R) for all
1 ≤ r < ∞. Since gj have compact support, {gj } ⊂ Lr (R) for all 1 ≤ r ≤ ∞
(depending on p), by the continuous imbedding. We will show that {gj } do
not converge strongly in Lr (R). Note that kgj kr,R = kgj kr,Ij = kgkr,I = c > 0.
Consider for any i 6= j,
kgi − gj krr,R = kgi krr,Ii + kgj krr,Ij + kgi − gj krr,R\(Ii ∪Ij ) = 2cr .
Thus, the sequence is not Cauchy in Lr (R) (as seen by choosing ε < 21/r c
for all i, j). The arguments can be generalised to Rn .
Theorem 2.8.21. Let 1 ≤ p ≤ ∞. For all u ∈ W 1,p (Rn ) the following

inequality holds:
kτh u − ukp ≤ k∇ukp |h|, ∀h ∈ Rn .
Proof. It is enough to prove the inequality for u ∈ D(Rn ) due to the density
of D(Rn ) in W 1,p (Rn ). Consider
(τh u)(x) − u(x) = u(x − h) − u(x)

Z 1
= − [∇u(x − th)] h dt.
0
Hence, by the Cauchy-Schwarz inequality

Z 1
|(τh u)(x) − u(x)| ≤ |∇u(x − th)||h| dt,
0
and then, by Hölder’s inequality,

Z 1
p p
|(τh u)(x) − u(x)| ≤ |h| |∇u(x − th)|p dt.
0
Integrating over Rn we have

Z Z Z 1
p p p
|(τh u)(x) − u(x)| dx ≤ |h| |∇u(x − th)| dt dx.
Rn Rn 0
By Fubini-Tonelli theorem and the invariance, under translation, of the

Lebesgue measure in Rn to obtain
Z Z
p p
|τh u − u| dx ≤ |h| |∇u(x)|p dx.
Rn Rn
Theorem 2.8.22 (Rellich-Kondrasov). Let 1 ≤ p < ∞ and let Ω be a

bounded domain with C 1 boundary, then
(i) If p < n, then W 1,p (Ω) ⊂⊂ Lr (Ω) for all r ∈ [1, p? ).
(ii) If p = n, then W 1,n (Ω) ⊂⊂ Lr (Ω) for all r ∈ [1, ∞).
(iii) If p > n, then W 1,p (Ω) ⊂⊂ C(Ω).
Proof. (i) We first prove the case p < n. Let B be the unit ball in W 1,p (Ω).
We shall verify conditions (i) and (ii) of Theorem 1.3.31. Let 1 ≤ q < p∗ .
Then choose α such that 0 < α ≤ 1 and
1 α 1−α
= + .
q 1 p∗
If u ∈ B, Ω0 ⊂⊂ Ω and h ∈ Rn such that |h| < dist(Ω0 , Rn \ Ω),
kτ−h u − ukq,Ω0 ≤ kτ−h u − ukα1,Ω0 kτ−h u − uk1−α

p? ,Ω0
≤ (|h|α k∇ukα1,Ω )(2kukp? ,Ω )1−α
≤ C|h|α .
We choose h small enough such that C|h|α < ε. This will verify condi-
tion (i) of Theorem 1.3.31. Now, if u ∈ B and Ω0 ⊂⊂ Ω, it follows by
Hölder’s inequality that
0 ?)
kukq,Ω\Ω0 ≤ kukp? ,Ω\Ω0 |Ω \ Ω |1−(q/p
0 ?)
≤ C|Ω \ Ω |1−(q/p
which can be made less than any given ε > 0 by choosing Ω0 ⊂⊂ Ω to

be ’as closely filling Ω’ as needed. This verifies condition (ii). Thus B
is relatively compact in Lq (Ω) for 1 ≤ q < p? .
(ii) Assume for the moment that the result is true for p < n. Notice that
as p → n, p? → ∞. Since Ω is bounded, W 1,n (Ω) ⊂ W 1,n−ε (Ω), for
every ε > 0. Since n − < n, using the p < n case, we get W 1,n−ε (Ω)
is compactly imbedded in Lr (Ω) for all r ∈ [1, (n − ε)? ). Note that as
ε → 0, (n − ε)? → ∞. Therefore, for any r < ∞ we can find small
enough ε > 0 such that 1 ≤ r < (n − )∗ . We deduce that W 1,n (Ω) is
compactly imbedded in Lr (Ω) for any 1 ≤ r < ∞.
(iii) For p > n, the functions of W 1,p (Ω) are Hölder continuous. If B is the
unit ball in W 1,p (Ω) then the functions in B are uniformly bounded and
equicontinuous in C(Ω̄). Thus B is relatively compact in C(Ω̄) by the
Ascoli-Arzela Theorem.
Remark 2.8.23. Note that the continuous inclusion for the r = p? case is
not compact. The above result can be extended to W01,p (Ω) provided Ω is
bounded and is a connected open subset (bounded domain) of Rn .
Corollary 2.8.24 (Compact subsets of W 1,p (Rn )). Let A be a subset of

W 1,p (Rn ), 1 ≤ p < +∞ which satisfies the two following conditions:
(i) A is bounded in W 1,p (Rn ), i.e., supf ∈A kf kW 1,p (Rn ) < +∞.
R
(ii) A is Lp -equi-integrable at infinity, i.e., limr→+∞ {|x|>r} |f (x)|p dx = 0
uniformly with respect to f ∈ A. Then, A is relatively compact in
Lp (Rn ).
Exercise 67. If Ω is connected subset of Rn and u ∈ W 1,p (Ω) such that

∇u = 0 a.e. in Ω, then show that u is constant a.e. in Ω
Theorem 2.8.25 (Poincaré-Wirtinger Inequality). Let Ω be a bounded, con-

nected open subset of Rn with C 1 smooth boundary and let 1 ≤ p ≤ ∞. Then
there is a constant C > 0 (depending on p, n and Ω) such that
ku − ukp ≤ Ck∇ukp ∀u ∈ W 1,p (Ω),

1
R
where u := |Ω| Ω
u(x) dx is the average of u in Ω.
Proof. Suppose the inequality was false, then for each positive integer m we
have a um ∈ W 1,p (Ω) such that
kum − um kp > mk∇um kp ∀u ∈ W 1,p (Ω). (2.8.7)
Set for all m,

um − um
vm := .
kum − um kp
Thus, kvm kp = 1 and vm = 0. Hence, by (2.8.7), we have
k∇vm kp < 1/m.
Therefore, {vm } are bounded in W 1,p (Ω) and, by Rellich-Kondrasov compact

imbedding, there is a subsequence of {vm } (still denoted by m) and a function
v ∈ Lp (Ω) such that vm → v in Lp (Ω). Therefore, v = 0 and kvkp = 1. Also,
for any φ ∈ D(Ω),
Z Z
∂φ ∂φ
v dx = lim vm dx
Ω ∂xi m→∞ Ω ∂xi
Z
∂vm
= − lim φ dx
m→∞ Ω ∂xi
→ 0.
Thus, ∇v = 0 a.e. and v ∈ W 1,p (Ω). Moreover v is constant, since Ω is

connected. But v = 0 implies that v ≡ 0 which contradicts the fact that
kvkp = 1.
n for measures). R(Ω) is compactly imbedded

Theorem 2.8.26 (Compactness
−1,r
in W (Ω) for all r ∈ 1, n−1 .
Proof. Let {µk }∞

1 be a bounded sequence in R(Ω). By weak compactness,
we extract a subsequence {µkj }∞ ∞
j=1 ⊂ {µk }k=1 such that µkj * µ weak-*
r
converges in R(Ω), for some measure µ ∈ R(Ω). Set s = r−1 and denote by
1,s n
B the closed unit ball in W0 (Ω). Since 1 ≤ r < n−1 , we have s > n and so
N (ε)
B is compact in C0 (Ω). Thus, given ε > 0 there exist {φi }i=1 ⊂ C0 (Ω) such
that
min kφ − φi kC(Ω) < ε
1≤i≤N (ε)
for each φ ∈ B. Therefore, if φ ∈ B,

Z Z Z Z
φ dµkj − φ dµ ≤ 2ε sup |µkj |(Ω) + φi dµkj − φi dµ
Ω Ω j Ω Ω
for some index 1 ≤ i ≤ N (ε). Consequently,

Z Z
lim sup φ dµkj − φ dµ = 0
j→∞ φ∈B Ω Ω
and, hence, µkj → µ strongly in W −1,r (Ω).
2.9 Trace Theory

Let u : Ω → R be a continuous function, where Ω is an open subset of Rn . The
trace of u on the boundary ∂Ω is the continuous function γ(u) : ∂Ω → R
defined by γ(u)(x) = u(x) for all x ∈ ∂Ω. Thus, we have a linear map
γ : C(Ω) → C(∂Ω) which is the restriction to the boundary. If u ∈ L2 (Ω)
then there is no sufficient information to talk about u on ∂Ω, because the
Lebesgue measure of ∂Ω is zero. However, an additional information on u,
∂u
viz., “ ∂xi
belongs to L2 (Ω) for any i = 1, . . . , n”, can give meaning to u
restricted on ∂Ω. Thus, in the Sobolev space W 1,p (Ω), where Ω is a domain
(open connected set) in Rn , the notion of trace or restriction to boundary
can be defined on ∂Ω, even for functions not continuous on Ω. The basis for
this extension is the following observation: Let Ω be a domain in Rn and let
1 ≤ p < n. Then one can show that the linear mapping
]
γ : C ∞ (Ω) → Lp (∂Ω)
is well defined and continuous if the space C ∞ (Ω) is endowed with the W 1,p -
norm where p] > 1 is defined as
1 1 p−1
:= − for 1 ≤ p < n.
p] p p(n − 1)
]
Since the space C ∞ (Ω) is dense in W 1,p (Ω) (cf. Corollary 2.5.6) and Lp (∂Ω)
is complete, there exists a unique continuous linear extension of (cf. The-
orem ??) γ from C ∞ (Ω) to W 1,p (Ω). This extension, still denoted as γ, is
]
called the trace operator and each γ(u) ∈ Lp (∂Ω) is called the trace of the
function u ∈ W 1,p (Ω).
Theorem 2.9.1. Let Ω be a domain in Rn and 1 ≤ p < ∞. The trace

operator γ satisfies the following:
]
(a) γ : W 1,p (Ω) → Lp (∂Ω) is a continuous linear operator, for 1 ≤ p < n.
(b) For p = n, γ : W 1,n (Ω) → Lq (∂Ω) is a continuous linear operator, for

all q ∈ [1, ∞).
(c) For p > n, γ : W 1,p (Ω) → C(∂Ω) is a continuous linear operator.
Further, if 1 < p < n then the trace operator γ : W 1,p (Ω) → Lq (∂Ω) is
compact for all q ∈ [1, p] ).
A consequence of the above Theorem is that, independent of the dimen-

sion n, the trace map γ : W 1,p (Ω) → Lp (∂Ω) is a continuous linear map for
all 1 ≤ p < ∞.
Lemma 2.9.2. Let Ω := Rn+ = {x = (x0 , xn ) ∈ Rn−1 × R : xn > 0}. Then,

n
for any 1 ≤ p < +∞, and u ∈ D(R̄+ )
1
kγ(u)kLp (Rn−1 ) ≤ p p kukW 1,p (Rn+ ) .
Proof. Let u ∈ D(R̄n+ ). For any x0 ∈ Rn−1 we have

Z +∞
0 ∂
p
|u(x , 0)| = − |u(x0 , xn )|p dxn
∂xn
Z 0+∞
∂u 0
≤ p |u(x0 , xn )|p−1 (x , xn ) dxn .
0 ∂xn
By Young’s convexity inequality
1 1
ab ≤ ap + bq
p q
1 1
with p
+ q
= 1 to the following situation:
∂u 0
a= (x , xn ) and b = |u(x0 , xn )|p−1 .
∂xn
We obtain
Z +∞ p
0 1 ∂u 0 1
|u(x , 0)| ≤ p p
(x , xn ) + |u(x0 , xn )|(p−1)q dxn .
0 p ∂xn q
By using the relation (p − 1)q = p we obtain

Z +∞ Z +∞ p
0 p 0 p ∂u 0
|u(x , 0)| ≤ (p − 1) |v(x , xn )| dxn + (x , xn ) dxn .
0 0 ∂xn
Integrating over Rn−1 yields

Z Z Z p
0 p 0 p ∂u
|u(x , 0)| dx ≤ (p − 1) |u(x)| dx + (x) dx
Rn−1 Rn
+ Rn
+
∂xn
Z Z n p
p
X ∂u
≤ (p − 1) |u(x)| dx + dx
Rn
+ Rn
+ i=1
∂xi
n
!
Z p
X ∂u
≤ p |u(x)|p + dx.
Rn
+ i=1
∂xi
Hence
1
ku(·, 0)kLp (Rn−1 ) ≤ p p kukW 1,p (Rn+ ) .
Theorem 2.9.3. Let Ω be bounded and ∂Ω is in the class C k . Then there is

a bounded linear operator γ : W k,p (Ω) → Lp (∂Ω) such that
1. γ(u) = u |∂Ω if u ∈ W k,p (Ω) ∩ C(Ω),
2. kγ(u)kp,∂Ω ≤ Ckuk1,p,Ω for all u ∈ W k,p (Ω) and C depending on p and

Ω.
We call γ(u) to be the trace of u on ∂Ω.

Proof. By the regularity property of the boundary ∂Ω, we know that D(Ω̄)
is dense in W 1,p (Ω). For any v ∈ D(Ω̄), we can define the restriction of u to
∂Ω, setting γ(u) := u|∂Ω . Suppose we prove that
γ : (D(Ω̄), k · kW 1,p (Ω) ) −→ (Lp (∂Ω), k · kLp (∂Ω) )
is continuous, then, by the linearity of γ, it is uniformly continuous. The

space Lp (∂Ω) is a Banach space. Therefore, there exists a unique linear and
continuous extension of γ
γ : W 1,p (Ω) −→ Lp (∂Ω).
Thus, we just need to prove that γ is continuous. By Lemma 2.9.2, the result
is true for half-space Ω = Rn+ . Let Ω̄ ⊂ ki=0 Gi with Ḡ0 ⊂ Ω, Gi open for all
S
i = 0, . . . , k, while φi : B(0, 1) → Gi , i = 1, 2, . . . , k, are the local coordinates
{α
Pk0 , . . . , αi }, is an associated partition of unity, i.e., αi ∈ D(Gi ), αi ≤ 0,
i=0 αi = 1 on Ω̄.
(
(αi v) ◦ φi on B+ ,
wi =
0 on Rn+ \B+ .
Clearly wi belongs to D(Rn+ ). By Lemma 2.9.2, we have

1
kwi (·, 0)kLp (Rn−1 ) ≤ p p kwi kW 1,p (Rn+ ) . (2.9.1)
By using classical differential calculus rules (note that all the functions αi , v, φi
are continuously differentiable), one obtains the existence, for any i = 1, . . . , k,
of a constant Ci such that
kwi kW 1,p (Rn+ ) ≤ Ci kvkW 1,p (Ω) . (2.9.2)
Combining the two inequalities (2.9.1) and (2.9.2), we obtain

1
kwi (·, 0)kLp (Rn−1 ) ≤ Ci p p kvkW 1,p (Ω) . (2.9.3)
We now use the definition of the Lp (∂Ω) norm which is based on the use of
local coordinates. One can show that an equivalent norm to the Lp (∂Ω) norm
can be obtained by using local coordinates: denoting by ∼ the extension by
zero outside of Rn−1 \{y ∈ Rn−1 | |y| < 1}, we have that
Lp (∂Ω) = {u : ∂Ω → R | (α
] p
i u) ◦ φi (·, 0) ∈ L (R
n−1
), 1 ≤ i ≤ k}
and
k
! p1
X
u 7→ αi u) ◦ φi kpLp (Rn−1 )
k(g (2.9.4)
i=1
is an equivalent norm to the Lp (∂Ω) norm.

This definition of the Lp (∂Ω) norm and the inequality (2.9.3) (note that
wi = (αfi v) ◦ φi ) yield
kukLp (∂Ω) ≤ C(p, n, Ω)kukW 1,p (Ω)
for some constant C(p, n, Ω). Thus, γ0 is continuous.
Theorem 2.9.4 (range of γ). Let Ω be an open bounded set in Rn whose

boundary ∂Ω is of class C1 . Then the trace operator γ is linear continuous
1
and surjective from H 1 (Ω) → H 2 (∂Ω).
1
Proof. The definition of H 2 (∂Ω) is obtained by local coordinates. Thus, it
is enough to prove when Ω = Rn+ and ∂Ω = Rn−1 . We do the proof in three
steps.
First step: Let w(x0 ) = v(x0 , 0). We first relate the Fourier transform (in
Rn−1 ) of w to that of v (in Rn ). We denote by w e the Fourier transform of w
in Rn−1 . By the Fourier inversion formula, if v ∈ D(Rn ),
Z
0 0 0
v(x , 0) = e2πıx ·ξ v̂(ξ) dξ, ξ = (ξ 0 , ξn )
n
ZR Z ∞
2πıx0 ·ξ 0
= e v̂(ξ) dξn dξ 0 .
Rn −∞
Now applying the same formula in Rn−1 , we get

Z
0 0 0 0
v(x , 0) = w(x ) = e2πıx ·ξ w(ξ
e 0 ) dξ 0 .
Rn−1
By the uniqueness of this formula (since w, w e ∈ S(Rn−1 ) if v ∈ D(Rn )) we

deduce that Z ∞
0
w(ξ
e )= v̂(ξ) dξn .
−∞
1
Second step: To show that v(x0 , 0) = w(x0 ) ∈ H 2 (Rn−1 ) we need to show
1
e 0 )|2 is integrable. But
that (1 + |ξ 0 |2 ) 2 |w(ξ
Z
1
(1 + |ξ 0 |2 ) 2 |w(ξ
e 0 )|2 dξ 0
Rn−1
Z Z ∞ 2
0 2 12
= (1 + |ξ | ) v̂(ξ) dξn dξ 0
Rn−1 −∞
Z Z ∞ 2
1 −1 1
= (1 + |ξ 0 |2 ) 2 (1 + |ξ|2 ) 2 dξn dξ 0
v̂(ξ)(1 + |ξ|2 ) 2
n−1
ZR Z−∞ ∞ Z ∞
0 2 21
≤ (1 + |ξ | ) 2 2
(1 + |ξ| )|v̂(ξ)| dξn 2 −1
(1 + |ξ| ) dξn dξ 0
Rn−1 −∞ −∞
Z
= π (1 + |ξ|2 )|v̂(ξ)|2 dξ = πkvk2H 1 (Rn )
Rn
since
Z ∞ Z ∞
2 −1 dξn 1
(1 + |ξ| ) dξn = 0 2 2
= π(1 + |ξ 0 |2 )− 2 .
−∞ −∞ 1 + |ξ | + ξn
The above formula is obtained by introducing the change of variable ξn =

1
(1 + |ξ 0 |2 ) 2 tan θ. Then
π/2 π/2
sec2 θ
Z Z
1 −1 −1
1 2
dθ = (1 + |ξ 0 |2 ) 2 dθ = π(1 + |ξ 0 |2 ) 2 .
(1 + |ξ 0 |2 ) 2 −π/2 1 + tan θ −π/2
1
Thus, if v ∈ D(Rn ), v(x0 , 0) ∈ H 2 (Rn−1 ) and by density the result follows
for v ∈ H 1 (Rn+ ).
1 1
Third step: We now show that γ is onto H 2 (Rn−1 ). Let h(x0 ) ∈ H 2 (Rn−1 ).
h(ξ 0 ) be its Fourier transform. We define u(x0 , xn ) by
Let e
0
e(ξ 0 , xn ) = e−(1+|ξ |)xn e
u h(ξ 0 ). (2.9.5)
We must first show that u then belongs to H 1 (Rn+ ). To see this extend u by
zero outside Rn+ . Now

Z
û(ξ) = e−2πıx·ξ u(x) dx
n
ZR∞ Z
0 0
= e−2πıx ·ξ u(x0 , xn )e−2πıxn ξn dx0 dxn
n−1
Z0 ∞ R
= e−2πıxn ξn u
e(ξ 0 , xn ) dxn
0
Z ∞
0 0
= eh(ξ ) e−(1+|ξ |+2πıξn xn ) dxn
0
h(ξ 0 )
e
= .
1 + |ξ 0 | + 2πıξn
Now,
(1 + |ξ 0 |)2 |e
h(ξ 0 )|2
Z Z
0 2 2
(1 + |ξ | )|û(ξ)| dξ = dξ
Rn Rn (1 + |ξ 0 |)2 + 4π 2 ξn2
(1 + |ξ 0 |)2 |e
h(ξ 0 )|2
Z
≤ dξ
Rn 1 + |ξ 0 |2 + ξn2
Z
1
= π (1 + |ξ 0 |2 ) 2 |e
h(ξ 0 )|2 dξ 0 < +∞
Rn−1
1 ∂u
since h ∈ H (Rn−1 ). This proves that u (extended by zero) and ∂x
2
i
are in
2 n ∂u 2 n
L (R ), for all 1 ≤ i ≤ n−1. Hence, u, ∂xi are in L (R+ ), for all 1 ≤ i ≤ n−1.
For the case i = n, notice that by differentiating under the integral sign,
^
∂u ∂e
u 0
(ξ 0 , xn ) = (ξ , xn ) = −(1 + |ξ 0 |)e
u (ξ 0 , xn ).
∂xn ∂xn
∂u
Extend ∂xn
by zero outside Rn+ . Then as before we get
−(1 + |ξ 0 |)e
h (ξ 0 )
\
∂u
(ξ) = .
∂xn 1 + |ξ 0 | + 2πıξn
Then
2
(1 + |ξ 0 |2 )|e
h (ξ 0 )|2
Z \ Z
∂u
(ξ) dξ ≤ 2 dξ
Rn ∂xn Rn 1 + |ξ 0 |2 + ξn2
Z
1
= 2π (1 + |ξ 0 |2 ) 2 |e
h (ξ 0 )|2 dξ 0 < ∞.
Rn
∂u ∂u
Thus ∂x n
(extended by zero) is in L2 (Rn ) and so ∂x n
∈ L2 (Rn+ ) and hence
u ∈ H 1 (Rn+ ). Now (by the Fourier inversion formula)
e(ξ 0 , 0) = e
u h(ξ 0 )
implies that u(x0 , 0) = h(x0 ) and so γ(u) = h.

Remark 2.9.5. If v ∈ W 2,p (Ω), by a similar argument one can give a mean-
∂v
ing to ∂ν .Note that ∇v ∈ [W 1,p (Ω)]n , and hence the trace of ∇v on ∂Ω
belongs to [Lp (∂Ω)]n . One defines
∂v
:= γ(∇v) · ν,
∂ν
which belongs to Lp (∂Ω). Indeed, one can show that
∂v 1
∈ W 1− p ,p (∂Ω).
∂ν
∂v 1
For p = 2, for v ∈ H 2 (Ω) we have ∂ν ∈ H 2 (∂Ω). One can also show that
∂v
the operator v 7→ {v|∂Ω , ∂ν } is linear continuous and onto from W 2,p (Ω) onto
1 1
W 2− p ,p (∂Ω) × W 1− p ,p (∂Ω).
]
The relation γ (W 1,p (Ω)) ( Lp (∂Ω) is the basis for defining the trace
spaces:
1 ]
W 1− p ,p (∂Ω) := {γ(v) ∈ Lp (∂Ω) | v ∈ W 1,p (Ω)} for 1 ≤ p < n
which is the traces of all the functions in W 1,p (Ω), 1 ≤ p < n.

Remark 2.9.6. By appropriate modifications one can easily prove that γ
maps H k (Rn+ ) onto H k−1/2 (Rn−1 ). In the same way, if u ∈ H 2 (Rn+ ) one can
1
∂u
prove that ∂x n
(x0 , 0) is in L2 (Rn−1 ) and is, in fact, in H 2 (Rn−1 ). Similarly,
one can extend − ∂x ∂u
n
(x0 , 0) to a bounded linear map γ1 : H 2 (Rn+ ) → L2 (Rn−1 )
1
whose range is H 2 (Rn−1 ). More generally, we have a collection of continuous
linear maps {γj } into L2 (Rn−1 ) such that the map γ = (γ0 , γ1 , · · · , γm−1 )
maps H k (Rn+ ) into [L2 (Rn−1 )]k and the range in the space
k−1
Y
H k−j−1/2 (Rn−1 ).
j=0
k
Definition 2.9.7. We shall denote the range of the map T to be W 2 ,p (∂Ω).
k k
For p = 2, we denote W 2 ,2 as H 2 (∂Ω).
Thus, instead of defining the fractional power Sobolev spaces using Fourier
transform, one can define them as range of trace operator γ.
k
Theorem 2.9.8. W 2 ,p (∂Ω) is dense in Lp (∂Ω).
Theorem 2.9.9 (Trace zero). Let Ω ⊂ Rn be bounded and ∂Ω is in C k class.

Then u ∈ W0k,p (Ω) iff γ(u) = 0 on ∂Ω. In particular, ker(γ) = W0k,p (Ω).
Proof. We first show the inclusion W01,p (Ω) ⊂ kerγ. Take v ∈ W01,p (Ω). By
definition of W01,p (Ω), there exists a sequence of functions (vn )n∈N , vn ∈ D(Ω)
such that vn → v in W 1,p (Ω). Since γ0 (vn ) = vn |∂Ω = 0, by continuity of γ0
we obtain that γ0 (v) = 0, i.e., v ∈ kerγ0 .
The other inclusion is a bit more involved. We shall prove it in a sequence
of lemma for H 1 (Rn+ ). The idea involved is following: Take v ∈ W 1,p (Rn+ )
such that γ0 (v) = 0. Prove that v ∈ W01,p (Rn+ ). Let us first extend v by zero
outside of Rn+ . By using the information γ0 (v) = 0 one can verify that the
so-obtained extension ṽ belongs to W 1,p (Rn ). Then, let us translate ṽ, and
consider for any h > 0
τh ṽ(x0 , xn ) = ṽ(x0 , xn − h).
Finally, one regularizes by convolution the function τh ṽ. We have that for ε
sufficiently small, ρε ? (τh ṽ) belongs to D(Rn+ ) and ρε ? (τh ṽ) tends to v in
W 1,p (Rn+ ) as h → 0 and ε → 0. Hence v ∈ W01,p (Rn+ ).
Lemma 2.9.10 (Green’s formula). . Let u, v ∈ H 1 (Rn+ ). Then

Z Z
∂v ∂u
u =− v if 1 ≤ i ≤ (n − 1) (2.9.6)
Rn
+
∂x i Rn ∂xi
+
Z Z Z
∂v ∂u
u =− v− γ0 (u)γ0 (v). (2.9.7)
Rn
+
∂x n Rn
+
∂xn Rn−1
Proof. If u, v ∈ D(Rn ), then the relations (2.9.6) and (2.9.7) follow by

integration by parts. The general case follows by the density of the re-
strictions of functions of D(Rn ) in H 1 (Rn+ ) and the continuity of the map
γ0 : H 1 (Rn+ ) → L2 (Rn−1 ).
Corollary 2.9.11. If u, v ∈ H 1 (Rn+ ) and at least one of them is in ker (γ0 )

then (2.9.6) holds for all 1 ≤ i ≤ n.
Lemma 2.9.12. Let v ∈ ker(γ0 ). Then its extension by zero outside Rn+ ,
denoted ve, is in H 1 (Rn ), and
^
∂e
v ∂v
= , 1 ≤ i ≤ n.
∂xi ∂xi
Proof. Let φ ∈ D(Rn ). Then for 1 ≤ i ≤ n,
Z Z Z Z ^
∂φ ∂φ ∂v ∂v
ve = v =− φ= φ
Rn ∂xi Rn
+
∂x n Rn
+
∂xi Rn ∂xi
by the above corollary. Let h > 0 and consider h̄ = hen ∈ Rn where en is the
unit vector (0, 0, ..., 0, 1). Consider the function τh̄ ve, where ve is the extension
by zero outside Rn+ of v ∈ ker(γ0 ). Then τh̄ ve vanishes for all x ∈ Rn such
that xn < h. (Recall that τh̄ ve(x) = ve(x − h̄).)
Lemma 2.9.13. Let 1 ≤ p < ∞ and h̄ ∈ Rn . Then if f ∈ Lp (Rn )
lim |τh̄ f − f |0,p, Rn = 0

h̄→0
Proof. By the translation invariance of the Lebesgue measure, τh̄ f ∈ Lp (Rn )

as well. Let > 0 be given and choose φ ∈ D(Rn ) such that

|f − φ|0,p, Rn < . (2.9.8)
3
Let a > 0 such that supp (φ) ⊂ [−a, a]n . Since φ is uniformly continuous,
there exists δ < 0 be given and choose φ ∈ D(Rn ) such that
n
|φ(x − h̄) − φ(x)| < (2(a + 1))− p .
3
Then
Z Z p
p
|φ(x − h̄) − φ(x)| dx = |φ(x − h̄) − φ(x)|p dx < .
Rn [−(a+1),(a+1)]n 3
Thus for |h̄| < δ,

|τh̄ φ − φ|0,p, Rn < . (2.9.9)
3
Finally, again by the translation invariance of the Lebesgue measure, we have

|τh̄ f − τh̄ φ|0,p,Rn = f − φ|0,p,Rn < . (2.9.10)
3
The result now follows on combining (2.9.8), (2.9.9) and (2.9.10) by the
triangle inequality.
Corollary 2.9.14. If v ∈ H 1 (Rn ) then
lim kτh̄ v − vk1,Rn = 0.

h→0
Proof. Clearly by the preceding lemma τh̄ v → v in L2 (Rn ). Also it is easy

to check that for any 1 ≤ i ≤ n,
∂ ∂v
(τh̄ v) = τh̄ .
∂xi ∂xi
∂(τh̄ v) ∂v
Thus again by the lemma, ∂xi
→ ∂xi
in L2 (Rn ).
Theorem 2.9.15. ker (γ0 ) = H01 (Rn+ ).
Proof. We already have seen that H01 (Rn+ ) ⊂ ker(γ0 ). Let now v ∈ ker(γ0 ).
Then we have seen that its extension ve by zero is in H 1 (Rn ). Using the
cut-off functions {ζk } (cf. Theorem 2.1.3) we have that ζk ve → ve as k → ∞
in H 1 (Rn ). The functions ζk ve have compact support in Rn and vanish for
xn < 0. Now fix such a k so that
ke
v − ζk vek1,Rn < η,
where η > 0 is a given positive quantity. Again we can choose h small enough
so that if h̄ = hen , then
kτh̄ (ζk ve) − ζk vek1, Rn < η.
Now τh̄ (ζk ve) has compact support in Rn+ and vanishes for all x ∈ Rn with
xn < h. Let {ρ } be the family of mollifers. If > 0 is chosen small enough
then ρ ∗ τh̄ (ζk ve) will have support contained in the set
B(0; ) + K ∩ {x|xn ≥ h > 0}
where K = supp(τh̄ (ζk ve)) is compact. Thus
ρ ∗ τh̄ (ζke) ∈ D(Rn+ )

and we know that as ↓ 0, ρ∗τh̄ (ζk ve) → τh̄ (ζk ve). Thus we can choose small
enough such that
kρ ∗ τh̄ (ζk ve) − τh̄ (ζk ve)k1,Rn < η.
Thus we have found a function φη ∈ D(Rn+ ),
φη = ρ ∗ τh̄ (ζk ve)
such that
kφη − vk1,Rn+ ≤ kφη − vek1,Rn < 3η .
Thus, as η is arbitrary, it follows that
ker(γ0 ) ⊂ D(Rn+ ) = H01 (Rn+ ).
Similarly, it can be proved that if γ = (γ0 , γ1 , · · · , γm−1 ), then the kernel

of γ in H m (Rn+ ) is precisely the set H01 (Rn+ ).
Theorem 2.9.16. Let 1 ≤ p < ∞. Then

( n
)
X ∂u
W02,p (Ω) = u∈W 2,p
(Ω) | γ(u) = 0 and νi γ( )=0 ,
i=1
∂xi
where (νi )ni=1 denotes the unit outer normal vector field along ∂Ω.
Theorem 2.9.17 (Trace Theorem). Let Ω ⊂ Rn be a bounded open set of

class C m+1 with boundary ∂Ω. Then there exists a trace map γ = (γ0 , γ1 , · · · , γm−1 )
from H m (Ω) into (L2 (Ω))m such that
(i) If v ∈ C ∞ (Ω̄), Then γ0 (v) = v|∂Ω , γ1 (v) = ∂ν

∂v
|∂Ω , · · · , and γm−1 (v) =
∂ m−1
∂ν m−1
(v)|∂Ω , where ν is the unit exterior normal to the boundary ∂Ω
(ii) The range of γ is the space

m−1
1
Y
H m−j− 2 (∂Ω).
j=0
(iii) The kernel of γ is H0m (Ω)

Proof. Let us now turn to the case of a bounded open set Ω of class C 1 . Let
{Uj , Tj }kj=1 be an associated local chart for the boundary ∂Ω and let {ψj }kj=1
be a partition of unity subordinate to the cover {Uj } of ∂Ω. If u ∈ H 1 (Ω),
then (ψj u|Uj∩ Ω)o Tj ∈ H 1 (Rn+ ) and so we can define its trace asTan element of
H 21 (Rn−1 ). Coming back by Tj−1 we can define the trace on Uj ∂Ω. Piecing
these together we get the trace γ0 u in L2 (∂Ω) and the image (by definition of
1
the spaces) will be precisely H 2 (∂Ω). Similarly if the boundary is smoother
we can define the higher order traces γj .
Now that we have given a meaning to the functions restricted to the
boundary of the domain, we intend to generalise the classical Green’s identi-
ties (cf. Appendix ??) to H 1 (Ω). The trace theorem above helps us to obtain
Green’s theorem for functions in H 1 (Ω), Ω of class C 1 . If ν(x) denotes the
unit outer-normal vector on the boundary ∂Ω (which is defined uniquely a.e.
on ∂Ω), we denote its components along the coordinate axes by νi (x). Thus
we write generically,
ν = (ν1 , · · · , νn ).
For example, if Ω = B(0; 1) then ν(x) = x for all |x| = 1. Thus νi (x) = xi
in this case. If Ω = B(0; R), then ν(x) = Rx . If Ω has a part of its boundary,
say, xn = 0, then the unit outer normal is ±en depending on the side on
which Ω lies.
Theorem 2.9.18 (Green’s Theorem). . Let Ω be a bounded open set of Rn
set of class C 1 lying on the same side of its boundary ∂Ω. Let u, v ∈ H 1 (Ω).
Then for 1 ≤ i ≤ n,
Z Z Z
∂v ∂u
u =− v+ (γ0 u) (γ0 v) νi . (2.9.11)
Ω ∂xi Ω ∂xi ∂Ω
Proof. Recall that C ∞ (Ω̄) is dense in H 1 (Ω). If um , vm ∈ C ∞ (Ω̄) then we

have by the classical Green’s Theorem
Z Z Z
∂vm ∂um
um =− vm + um vm νi
Ω ∂xi Ω ∂xi ∂Ω
and choosing um → u, vm → v in H 1 (Ω) we deduce (2.9.11) by the continuity

of the trace map γ0 .
Usually, we rewrite γ0 (v) as just v on ∂Ω and understand it as the trace
∂v
of v on ∂Ω. Similarly if γ1 v is defined we will write it as ∂ν .
Theorem 2.9.19. Let Ω be a bounded open subset of Rn with a C 1 boundary.

Given a vector field V = (v1 , . . . , vn ) on Ω such that vi ∈ H 1 (Ω) for all
1 ≤ i ≤ n, then Z Z
∇ · V dx = V · ν dσ.
Ω ∂Ω
Proof. Setting u ≡ 1 and v = vi ∈ H 1 (Ω), we get from (2.9.11)

Z Z
∂vi
= vi νi .
Ω ∂xi Ω
and if V = (vi ) ∈ (H 1 (Ω))n , we get on summing with respect to i,

Z Z
div (V ) = V · ν dσ
Ω ∂Ω
which is the Gauss Divergence Theorem.

Corollary 2.9.20. Let Ω be a bounded open subset of Rn with a C 1 boundary.
Let u, v ∈ H 2 (Ω), then
(i) Z Z
∂u
(v∆u + ∇v · ∇u) dx = v dσ,
Ω ∂Ω ∂ν
∂u
where ∂ν
:= ∇u · ν and ∆ := ∇ · ∇.
(ii) Z Z
∂u ∂v
(v∆u − u∆v) dx = v −u dσ.
Ω ∂Ω ∂ν ∂ν
∂u
Proof. If we have u ∈ H 2 (Ω) and use ∂x i
in place of u in (2.9.11), we get
Z Z 2 Z
∂u ∂v ∂ u ∂u
=− 2
v+ vνi .
Ω ∂xi ∂xi Ω ∂xi ∂Ω ∂xi
If u were smooth then ni=1 ∂x ∂u

νi = ∂u
P
i ∂ν
. Thus by continuity of the trace γ1 ,
2 1
we get, for u ∈ H (Ω), v ∈ H (Ω),
Z Z Z
∂u
∇u · ∇v = − (∆u) v + v .
Ω Ω ∂Ω ∂ν
Remark 2.9.21. The Green’s formula holds for u ∈ W 1,p (Ω) and a ∈
W 1,q (Ω) with p1 + 1q = 1
2.10 Further reading

For a detailed study of Sobolev spaces, we refer to [Ada75, Maz85, Kes89].
Chapter 3
Bounded Variation Functions
Let Φ : Ω → Rn and Φ = (Φ1 , . . . , Φn ) be the coordinates where Φi : Ω → R

for all i. Recall that
Pm the divergence of any Φ ∈ C 1 (Ω; Rn ) is defined as
∂Φi
div(Φ) = ∇ · Φ = i=1 ∂xi . Also, the uniform norm of Φ is given as
n
!1/2
X
kΦk∞ = sup Φ2i (x) .
x∈Ω
i=1
Recall that any u ∈ L1loc (Ω) is a distribution and admits a distributional

derivative.
Definition 3.0.1. The total variation of u ∈ L1loc (Ω) is defined as

Z
V (u, Ω) = sup u(∇ · Φ) dx
Φ∈D(Ω;Rn ) Ω
kΦk∞ ≤1
Definition 3.0.2. For any open subset Ω ⊂ Rn , we define the space of

bounded variations BV(Ω) := {u ∈ L1 (Ω) | V (u, Ω) < ∞} and the norm on
it is given as
kukBV(Ω) = kuk1 + V (u, Ω).
Theorem 3.0.3. The normed space BV(Ω) is complete.
Definition 3.0.4. We define the metric of strict convergence , d, on BV(Ω)

as
d(u, v) = ku − vk1 + |V (u, Ω) − V (v, Ω)|.
143
CHAPTER 3. BOUNDED VARIATION FUNCTIONS 144
Theorem 3.0.5. Let Ω be an open subset of Rn . Given any u ∈ BV(Ω),

there exist functions {uk }∞ ∞
k=1 ⊂ BV(Ω) ∩ Cc (Ω) such that
1. uk → u in L1 (Ω) and
2. V (uk , Ω) → V (u, Ω) as k → ∞
Theorem 3.0.6. Let Ω be an open subset of Rn . For any function u ∈
Cc∞ (Ω), the total variation of u satisfies the identity V (u, Ω) = k∇uk1,Ω .
1,1
Corollary 3.0.7. Let u ∈ Wloc (Ω). Then ∇u ∈ L1 (Ω; Rn ) iff V (u, Ω) < ∞.
Further, in that case, k∇uk1,Ω = V (u, Ω).
Theorem 3.0.8 (Sobolev inequality for BV). There exists a constant C > 0
such that
n
kuk n−1
n
,Rn ≤ CV (u, R ) ∀u ∈ BV(Rn ).
Proof. Choose uk ∈ Cc∞ (Rn ) so that uk → u in L1 (Rnn) and V (uk , Rn ) →

V (u, Rn ) as k → ∞. Applying Fatou’s lemma on {|uk | n−1 }, we get
Z Z
n n
|u| n−1 dx ≤ lim inf |uk | n−1 dx.
Rn k→∞ Rn
1,1
Thus, kuk n−1
n
,Rn ≤ lim inf k→∞ kuk k n−1
n
,Rn . Since uk ∈ W (Rn ), applying
GNS inequality for p = 1, we obtain
kf k n−1
n
,Rn ≤ lim Ck∇uk k1,Rn .
k→∞
But k∇uk k1,Rn = V (uk , Rn ), for all k. Hence,

n n
kuk n−1
n
,Rn ≤ lim CV (uk , R ) = CV (u, R ).
k→∞
3.1 De Giorgi Perimeter

Theorem 3.1.1. Let Ω be an open subset of Rn with C 1 boundary and the
boundary is bounded. Then
Z
dσ = V (χΩ , Rn ).
∂Ω
Definition 3.1.2. Let E be a measurable subset of Ω. The De Giorgi perime-

ter of E is defined as PΩ (E) = V (χE , Ω). The set E is said to have finite
perimeter in Ω if χE ∈ BV (Ω).
Corollary 3.1.3 (Isoperimetric Inequality). For any bounded set E ⊂ Rn of

finite perimeter, we have
|E|1−1/n ≤ CPRn (E).
Proof. Let u = χE . Then u ∈ BV(Rn ) and hence kuk n−1

n n
,Rn ≤ CV (u, R ).
1−1/n
But kuk n−1
n
,Rn = |E| and V (u, Rn ) = PRn (E).
Theorem 3.1.4 (Morse-Sard). Let Ω be an open subset of Rn and u ∈

C ∞ (Ω). Then the set
{t ∈ R | u(x) = t and ∇u(x) = 0}
has Lebesgue measure zero.
Theorem 3.1.5 (Co-area Formula). Let u ∈ D(Rn ) and f ∈ C 1 (Rn ). Then

Z Z ∞ Z
f |∇u| dx = dt f dσ.
Rn 0 {|u|=t}
Moreover, for every open subset Ω ⊂ Rn ,

Z Z ∞ Z
|∇u| dx = dt χΩ dσ.
Ω 0 {|u|=t}
Chapter 4
Second Order Elliptic PDE
4.1 Homogeneous Dirichlet Problem

Let Ω be an open subset of Rn and let ∂Ω be the topological boundary of
Ω. Recall the homogeneous Dirichlet problem, i.e., given f : Ω → R, find
u : Ω → R such that
−∆u = f in Ω
(4.1.1)
u = 0 on ∂Ω.
Note that the boundary condition imposed above is referred to as the Dirich-
let boundary condition. The minus (−) sign before Laplacian has physical
motivation.
Definition 4.1.1. We say u : Ω → R is a classical solution of the Dirichlet
problem (4.1.1), if u ∈ C 2 (Ω) and satisfies (4.1.1) pointwise for each x ∈ Ω.
Let u be a classical solution of (4.1.1) and suppose Ω is bounded with
C -smooth boundary (cf. Corollary ??). Then, multiplying any φ ∈ C 1 (Ω)
1
and integrating both sides in (4.1.1), we get

Z Z
− (∆u)φ dx = f φ dx
Ω Ω
Z Z Z
∂u
∇φ · ∇u dx − φ dσ = f φ dx
Ω ∂Ω ∂ν Ω
(By Green’s identity)
Now, if we impose the Dirichlet boundary condition on u, we get
Z Z
∇φ · ∇u dx = f φ dx,
Ω Ω
147
CHAPTER 4. SECOND ORDER ELLIPTIC PDE 148
since u = 0 implies that ∂u∂ν

= 0 on ∂Ω. Thus, any classical solution u of
(4.1.1) solves the following formulation of the problem: for any given f , find
u such that Z Z
∇u · ∇φ dx = f φ dx ∀φ ∈ C 1 (Ω),
Ω Ω
as long as the integrals make sense. Note that, in the above problem, it
is enough for u to be in C 1 (Ω). Also, in particular, any classical solution
satisfies the identity Z Z
|∇u|2 dx = f u dx.
Ω Ω
Since C 1 (Ω) which vanishes on boundary is dense in H01 (Ω), we have the
classical solution u will also solve
Z Z
∇u · ∇v dx = f v dx ∀v ∈ H01 (Ω),
Ω Ω
as long as the integrals make sense. The integrals will make sense if f ∈
L2 (Ω). In fact, more generally, if f ∈ H −1 (Ω) then the integral on the
right side can be replaced with the duality product. The arguments above
motivates the notion of weak solution
Definition 4.1.2. Given f ∈ H −1 (Ω), we say u ∈ H01 (Ω) is a weak solution

of (4.1.1) if
Z
∇u · ∇v dx = hf, viH −1 (Ω),H01 (Ω) , ∀v ∈ H01 (Ω). (4.1.2)
Ω
We have already observed that every classical solution is a weak solution

for a bounded open set Ω with C 1 -smooth boundary.
Theorem 4.1.3. Let Ω be a bounded open subset of Rn and f ∈ H −1 (Ω).

Then there is a unique weak solution u ∈ H01 (Ω) satisfying (4.1.2). Moreover,
u minimizes the functional J : H01 (Ω) → R defined as,
Z
1
J(v) := |∇v|2 dx − hf, viH −1 (Ω),H01 (Ω)
2 Ω
in H01 (Ω).
Proof. Recall that H 1 (Ω) is a Hilbert space endowed with the inner-product
Z Z
(v, w) := vw dx + ∇v∇w dx.
Ω Ω
Since Ω is bounded, by Poincaré inequality (cf. (2.8.4)), there is a constant
C > 0 such that
kvk2 ≤ Ck∇vk2 ∀v ∈ H01 (Ω).
Thus, H01 (Ω) is a Hilbert space endowed with the equivalent norm kvkH01 (Ω) :=
k∇vk2 and the corresponding inner product is given as
Z
((v, w)) := ∇v · ∇w dx.
Ω
We shall use the Lax-Milgram result (cf. Theorem ??). We define the map
a : H01 (Ω) × H01 (Ω) :→ R as
Z
a(v, w) := ((v, w)) = ∇v · ∇w dx.
Ω
Note that a is a bilinear. We need to show that a is continuous and coercive.
Consider,
Z
|a(v, w)| ≤ |∇v||∇w| dx
Ω
Z 1/2 Z 1/2
2 2
≤ |∇v| dx |∇w| dx
Ω Ω
= kvkH01 (Ω) kwkH01 (Ω) .
Thus, a is a bilinear continuous form. Now, consider
Z
a(v, v) = |∇v|2 dx = k∇vk22 .
Ω
Hence a is coercive, bilinear continuous form. By Lax-milgram theorem,
there is a u ∈ H01 (Ω) such that a(u, v) = hf, viH −1 (Ω),H01 (Ω) . This is equivalent
to (4.1.2) and also u minimizes the functional
Z
1
= |∇v|2 dx − hf, viH −1 (Ω),H01 (Ω) .
2 Ω
We have the shown the existence of weak solution to the homogeneous

Dirichlet problem under the assumption that Ω is bounded. We now turn
our attention to the inhomogeneous Dirichlet problem.
4.2 Inhomogeneous Dirichlet problem

Ω. The inhomogeneous Dirichlet problem is the following: Given f : Ω → R
and g : ∂Ω → R, find u : Ω → R such that

−∆u = f in Ω
(4.2.1)
u = g on ∂Ω.
Definition 4.2.1. We say u : Ω → R is a classical solution of (4.2.1), if

u ∈ C 2 (Ω) and satisfies (4.2.1) pointwise for each x ∈ Ω.
We wish to give a weak notion of solution to the inhomogeneous Dirichlet
problem. It follows from trace theorem (cf. Theorem 2.9.3) that for u = g on
∂Ω to make sense for any u ∈ H 1 (Ω), g should be in H 1/2 (∂Ω). Conversely,
for any g ∈ H 1/2 (∂Ω), there is a vg ∈ H 1 (Ω) such that T vg = g. However,
note that the choice of vg is not unique.
Let f ∈ L2 (Ω) and g ∈ H 1/2 (∂Ω). For any given g ∈ H 1/2 (∂Ω) there is a
vg ∈ H 1 (Ω) such that T vg = g. Let
Kg := {v ∈ H 1 (Ω) | v − vg ∈ H01 (Ω)}.
Exercise 68. Show that Kg is a closed convex non-empty subset of H 1 (Ω)

and Kg = vg + H01 (Ω). (Hint: Use the linearity and continuity of the trace
map T ).
Note that if u is a classical solution of (4.2.1) and g ∈ H 1/2 (∂Ω), then
w := u − vg is a weak solution of the homogeneous Dirichlet problem

−∆w = f + ∆vg in Ω
(4.2.2)
w =0 on ∂Ω.
Definition 4.2.2. Given f ∈ H −1 (Ω) and g ∈ H 1/2 (∂Ω), we say u ∈ Kg is

a weak solution of (4.2.1) if w := u − vg ∈ H01 (Ω) is the weak solution of the
homogeneous Dirichlet problem (4.2.2), i.e., u ∈ Kg is such that
Z
∇u · ∇v dx = hf, viH −1 (Ω),H01 (Ω) dx ∀v ∈ H01 (Ω). (4.2.3)
Ω
Theorem 4.2.3. Let Ω be a bounded open subset of Rn , g ∈ H 1/2 (∂Ω) and

f ∈ H −1 (Ω). Then there is a unique weak solution u ∈ Kg of (4.2.1).
Moreover, if f ∈ L2 (Ω) then u minimizes the functional J : Kg → R defined

as, Z Z
1 2
J(v) := |∇v| dx − f v dx
2 Ω Ω
in Kg .
Proof. For any given g ∈ H 1/2 (Ω), there is a vg ∈ H 1 (Ω). Note that the
choice of vg is not unique. We have already seen (cf. Theorem 4.1.3) the
existence of weak solution for the homogeneous Dirichlet problem. Thus,
there is a unique w ∈ H01 (Ω) satisfying (4.2.2). Then u := w + vg is a weak
solution of (4.2.1). However, the uniqueness of u does not follow from the
uniqueness of w, since the choice of vg is not unique. If u1 , u2 ∈ Kg are two
weak solutions of (4.2.1) then, by (4.2.3),
Z
∇(u1 − u2 ) · ∇v dx = 0 ∀v ∈ H01 (Ω).
Ω
In particular, for v = u1 − u2 , we have k∇(u1 − u2 )k22 = 0. But, by Poincaré

inequality, ku1 − u2 k22 ≤ Ck∇(u1 − u2 )k22 = 0 implies that u1 = u2 .
It now only remains to show that u minimizes J in Kg . Since Kg is a
closed convex subset of H 1 (Ω), by Theorem ??, there is a unique u? ∈ Kg
such that Z
? ?
a(u , v − u ) ≥ f (v − u? ) dx, ∀v ∈ Kg ,
Ω
R
where a(v,Rw) := Ω ∇v · ∇w dx. But v ∈ Kg iff v − u? ∈ H01 (Ω). Thus,
a(u? , v) ≥ Ω f v dx for
R all v ∈ H01 (Ω). Since H01 (Ω) is a subspace of H 1 (Ω),
we have a(u? , v) = Ω f v dx for all v ∈ H01 (Ω). Thus, u? satisfies (4.2.3)
which by uniqueness of u ∈ Kg we have u? = u.
4.3 Linear Elliptic Operators

In this section, we introduce the generalization of the Laplace operator, called
the elliptic operators. The elliptic operators come in two forms, divergence
and non-divergence form, and we shall see that the notion of weak solution
can be carried over to elliptic operator in divergence form.
Let Ω be an open subset of Rn . Let A = A(x) = (aij (x)) be any given
n × n matrix of functions, for 1 ≤ i, j ≤ n. Let b = b(x) = (bi (x)) be any
given n-tuple of functions and let c = c(x) be any given function.
Definition 4.3.1. A second order operator L is said to be in divergence

form, if L acting on some u has the form
Lu := −∇ · (A(x)∇u) + b(x) · ∇u + c(x)u.
Equivalently,
Lu := −div(A(x)∇u) + b(x) · ∇u + c(x)u.
(Hence the name divergence form). On the other hand, a second order op-
erator L is said to be in non-divergence form, if L acting some u has the
form
n
X ∂ 2u
Lu := − aij (x) + b(x) · ∇u + c(x)u.
i,j=1
∂x i ∂x j
Observe that the operator L will make sense in the divergence form only
if aij (x) ∈ C 1 (Ω). Thus, if aij (x) ∈ C 1 , then a divergence form equation can
be rewritten in to a non-divergence form because
n n
!
X ∂ 2u X ∂aij
∇ · (A(x)∇u) = aij (x) + · ∇u.
i,j=1
∂x i ∂x j j=1
∂x j
i
∂a
Now, by setting ebi (x) = bi (x) − nj=1 ∂xijj , we have written a divergence L in
P
non-divergence form. Also, note that for A(x) = I, b(x) = 0 and c(x) = 0,
L = −∆ is the usual Laplacian.
Definition 4.3.2. We say a second order operator L is elliptic or coercive

if there is a positive constant α > 0 such that
α|ξ|2 ≤ A(x)ξ.ξ a.e. in x, ∀ξ = (ξi ) ∈ Rn .
The second order operator L is said to be degenerate elliptic if
0 ≤ A(x)ξ.ξ a.e. in x, ∀ξ = (ξi ) ∈ Rn .
We shall now extend the notion of weak solution introduced for the Lapla-
cian operator to a general second order elliptic equation in divergence form.
We remark that for the integrals in the defintion of weak solution to make
sense, the minimum hypotheses on A(x), b and c is that aij , bi , c ∈ L∞ (Ω).
Definition 4.3.3. Let aij , bi , c ∈ L∞ (Ω) and let f ∈ H −1 (Ω), we say u ∈

H01 (Ω) is a weak solution of the homogeneous Dirichlet problem

−div(A(x)∇u) + b(x) · ∇u + c(x)u = f in Ω
(4.3.1)
u = 0 on ∂Ω
whenever
Z Z Z
A∇u·∇v dx+ (b·∇u)v dx+ cuv dx = hf, viH −1 (Ω),H 1 (Ω) , ∀v ∈ H01 (Ω).
0
Ω Ω Ω
(4.3.2)
We define the map a(·, ·) : H01 (Ω) × H01 (Ω) → R as

Z Z Z
a(v, w) := A∇v · ∇w dx + (b · ∇v)w dx + cvw dx.
Ω Ω Ω
It is easy to see that a(·, ·) is bilinear.
Lemma 4.3.4. If aij , bi , c ∈ L∞ (Ω) then the bilinear map a(·, ·) is continuous
on H01 (Ω) × H01 (Ω), i.e., there is a constant c1 > 0 such that
|a(v, w)| ≤ c1 kvkH01 (Ω) kwkH01 (Ω) .
Also, in addition, if Ω is bounded and L is elliptic, then there are constants

c2 > 0 and c3 ≥ 0 such that
c2 kvk2H 1 (Ω) ≤ a(v, v) + c3 kvk22 .

0
Proof. Consider,
Z Z
|a(v, w)| ≤ |A(x)∇v(x) · ∇w(x)| dx + |(b(x) · ∇v(x))w(x)| dx
ΩZ Ω
+ |c(x)v(x)w(x)| dx
Ω
≤ max kaij k∞ k∇vk2 k∇wk2 + max kbi k∞ k∇vk2 kwk2 + kck∞ kvk2 kwk2
i,j i
≤ mk∇vk2 (k∇wk2 + kwk2 ) + mkvk2 kwk2 ,

where m = max max kaij k∞ , max kbi k∞ , kck∞
i,j i
≤ c1 kvkH01 (Ω) kwkH01 (Ω) .

Since L is elliptic, we have

Z
2
αk∇vk2 ≤ A(x)∇v · ∇v dx
Ω Z Z
= a(v, v) − (b · ∇v)v dx − cv 2 dx
Ω Ω
≤ a(v, v) + max kbi k∞ k∇vk2 kvk2 + kck∞ kvk22 .
i
If b = 0, then we have the result with c2 = α and c3 = kck∞ . If b 6= 0, we

choose a γ > 0 such that
2α
γ< .
maxi kbi k∞
Then, we have
αk∇vk22 ≤ a(v, v) + max kbi k∞ k∇vk2 kvk2 + kck∞ kvk22

i
kvk2
= a(v, v) + max kbi k∞ γ 1/2 k∇vk2 1/2 + kck∞ kvk22
i γ
kvk22

maxi kbi k∞ 2
≤ a(v, v) + γk∇vk2 + + kck∞ kvk22
2 γ
2 2
(using ab ≤ a /2 + b /2)

γ 1
α− max kbi k∞ k∇vk22 ≤ a(v, v) + max kbi k∞ + kck∞ kvk22 .
2 i 2γ i
By Poincaré inequality there is a constant C > 0 such that 1/Ckvk2H 1 (Ω) ≤

0
k∇vk22 . Thus, we have
c2 kvk2H 1 (Ω) ≤ a(v, v) + c3 kvk22 .

0
Theorem 4.3.5. Let Ω be a bounded open subset of Rn , aij , c ∈ L∞ (Ω),

b = 0, c(x) ≥ 0 a.e. in Ω and f ∈ H −1 (Ω). Also, let A satisfy ellipticity
condition. Then there is a unique weak solution u ∈ H01 (Ω) satisfying
Z Z
A∇u · ∇v dx + cuv dx = hf, viH −1 (Ω),H 1 (Ω) , ∀v ∈ H01 (Ω).
0
Ω Ω
Further, if A is symmetric then u minimizes the functional J : H01 (Ω) → R

defined as,
Z Z
1 1
J(v) := A∇v · ∇v dx + cv 2 dx − hf, viH −1 (Ω),H01 (Ω)
2 Ω 2 Ω
in H01 (Ω).
Proof. We define the bilinear form as

Z Z
a(v, w) := A∇v · ∇w dx + cvw dx.
Ω Ω
It follows from Lemma 4.3.4 that a is a continuous. Now ,consider

Z
2
αk∇vk2 ≤ A(x)∇v · ∇v dx
ZΩ Z
≤ A(x)∇v · ∇v dx + cv 2 dx (since c(x) ≥ 0)
Ω Ω
= a(v, v).
Thus, a is coercive in H01 (Ω), by Poincaré inequality. Hence, by Lax Milgram

theorem (cf. Theorem ??), u ∈ H01 (Ω) exists. Also, if A is symmetric, then
u minimizes the functional J on H01 (Ω).
Theorem 4.3.6. Let Ω be a bounded open subset of Rn , aij , bi , c ∈ L∞ (Ω)

and f ∈ L2 (Ω). Also, let A satisfy ellipticity condition. Consider L as in
(4.3.1). The space of solutions {u ∈ H01 (Ω) | Lu = 0} is finite dimensional.
For non-zero f ∈ L2 (Ω), there exists a finite dimensional subspace S ⊂ L2 (Ω)
such that (4.3.1) has solution iff f ∈ S ⊥ , the orthogonal complement of S.
Proof. It is already noted in Lemma 4.3.4 that one can find a c3 > 0 such
that a(v, v) + c3 kvk22 is coercive in H01 (Ω). Thus, by Theorem 4.3.5, there is
a unique u ∈ H01 (Ω) such that
Z Z
a(u, v) + c3 uv dx = f v dx ∀v ∈ H01 (Ω).
Ω Ω
Set the map T : L2 (Ω) → H01 (Ω) as T f = u. The map T is a compact

operator on L2 (Ω) because it maps u into H01 (Ω) which is compactly con-
tained in L2 (Ω). Note that (4.3.1) is equivalent to u = T (f + c3 u). Set
v := f + c3 u. Then v − c3 T v = f . Recall that T is compact and c3 > 0.

Thus, I − c3 T is invertible except when c−1 −1
3 is an eigenvalue of T . If c3 is
not an eigenvalue then there is a unique solution v for all f ∈ L2 (Ω). If c−1 3
is an eigenvalue then it has finite geometric multiplicity (T being compact).
Therefore, by Fredhölm alternative (cf. Theorems ?? and ??), solution ex-
ists iff f ∈ N (I − c3 T ∗ )⊥ and the dimension of S := N (I − c3 T ∗ ) is same as
N (I − c3 T ).
Theorem 4.3.7 (Regularity of Weak Solution). Let Ω be an open subset of
class C 2 . Let aij , bi , c ∈ L∞ (Ω) and f ∈ L2 (Ω). Let u ∈ H01 (Ω) be such that
it satisfies (4.3.2). If aij ∈ C 1 (Ω), bi ∈ C(Ω) and f ∈ L2 (Ω) then u ∈ H 2 (Ω).
More generally, for m ≥ 1, if aij ∈ C m+1 (Ω), bi ∈ C m (Ω) and f ∈ H m (Ω)
then u ∈ H m+2 (Ω).
Exercise 69. Let {fk }∞
1 be a bounded sequence in W
−1,p
(Ω), for some p > 2.
If fk = gk + hk , where {gk } is relatively compact in H −1 (Ω) and {hk } is
bounded in R(Ω), then {fk } is relatively compact in H −1 (Ω).
Proof. For k = 1, 2, ... let uk ∈ H01 (Ω) be the weak solution of

−∆uk = fk in Ω,
uk = 0 on ∂Ω.
Then uk = vk + wk , where

−∆vk = gk in Ω,
vk = 0 on ∂Ω
and
−∆wk = hk in Ω,
wk = 0 on ∂Ω.
We observe that {vk } is relatively compact in H01 (Ω). By the compact imbed-
ding of measures (above theorem), {wk } is relatively compact in W01,r (Ω) for
n
each 1 ≤ r < n−1 . Thus, {uk } is relatively compact in W01,r (Ω), and {fk } is
relatively compact in W −1,r (Ω). As {fk } is bounded in W −1,p (U ) for some
p > 2, the result follows.
Theorem 4.3.8 (Weak Maximum Principle). Let Ω be a bounded open subset
of Rn with sufficient smooth boundary ∂Ω. Let aij , c ∈ L∞ (Ω), c(x) ≥ 0 and
f ∈ L2 (Ω). Let u ∈ H 1 (Ω) ∩ C(Ω) be such that it satisfies (4.3.2) with b ≡ 0.
Then the following are true:
(i) If f ≥ 0 on Ω and u ≥ 0 on ∂Ω then u ≥ 0 in Ω.
(ii) If c ≡ 0 and f ≥ 0 then u(x) ≥ inf y∈∂Ω u(y) for all x ∈ Ω.
(iii) If c ≡ 0 and f ≡ 0 then inf y∈∂Ω u(y) ≤ u(x) ≤ supy∈∂Ω u(y) for all
x ∈ Ω.
Proof. Recall that if u ∈ H 1 (Ω) then |u|, u+ and u− are also in H 1 (Ω).
(i) If u ≥ 0 on ∂Ω then u = |u| on ∂Ω. Hence, u− ∈ H01 (Ω). Thus, using
v = u− in (4.3.2), we get
Z Z Z
− − − 2
− A∇u · ∇u dx − c(x)(u ) dx = f (x)u− (x) dx
Ω Ω Ω
because u+ and u− intersect on {u = 0} and, on this set, u+ = u− = 0

and ∇u+ = ∇u− = 0 a.e. Note that RHS is non-negative because both
f and u− are non-negative. Therefore,
Z Z
0≥ A∇u · ∇u dx + c(x)(u− )2 dx ≥ αk∇u− k22 .
− −
Ω Ω
Thus, k∇u− k2 = 0 and, by Poincarè inequality, ku− k2 = 0. This implies

u− = 0 a.e and, hence, u = u+ a.e. on Ω.
(ii) Let m = inf y∈∂Ω u(y). Then u − m ≥ 0 on ∂Ω. Further, c ≡ 0 implies

that u − m satisfies (4.3.2) with b = 0. By previous case, u − m ≥ 0 on
Ω.
(iii) If f ≡ 0 then −u satisfies (4.3.2) with b = 0. By previous case, we have

the result.
Theorem 4.3.9 (Spectral Decomposition). Let A be such that aij (x) =

aji (x), i.e., is a symmetric matrix and c(x) ≥ 0. There exists a sequence of
positive real eigenvalues {λm } and corresponding orthonormal basis {φm } ⊂
C ∞ (Ω) of L2 (Ω), with m ∈ N, such that

−div[A(x)∇φm (x)] + c(x)φm (x) = λm φm (x) in Ω
(4.3.3)
φm = 0 on ∂Ω
and 0 < λ1 ≤ λ2 ≤ . . . diverges.

Proof. Let T : L2 (Ω) → H01 (Ω) defined as T f = u where u is the solution of

−div[A(x)∇u(x)] + c(x)u(x) = f (x) in Ω
φm = 0 on ∂Ω.
Thus,
Z Z Z
A(x)∇(T f ) · ∇v(x) dx + c(x)(T f )(x)v(x) dx = f (x)v(x) dx.
Ω Ω Ω
Note that T is a compact operator on L2 (Ω) and T is self-adjoint because,

for every g ∈ L2 (Ω),
Z Z Z
(T f )(x)g(x) dx = A(x)∇(T g) · ∇(T f ) dx + c(x)(T g)(x)(T f )(x) dx
Ω ZΩ Ω
= (T g)(x)f (x) dx.

Ω
Further, T is positive definite because, for f 6≡ 0,

Z Z Z
(T f )(x)f (x) dx = A(x)∇(T f ) · ∇(T f ) dx + c(x)(T f )2 (x) dx
Ω Ω Ω
≥ αk∇T f k22 > 0.
Thus, there exists an orthonormal basis of eigenfunctions {φm } in L2 (Ω) and

a sequence of positive eigenvalues µm decreasing to zero such that T φm =
µm φm . Set λm = µ−1 m . Then φm = λm T φm = T (λm φm ). Thus, φm ∈ H0 (Ω)
1
because range of T is H01 (Ω). Hence, φm satisfies (4.3.3). It now only remains
to show that φm ∈ C ∞ (Ω). For any x ∈ Ω, choose Br (x) ⊂ Ω. Since
φm ∈ L2 (Br (x)) and solves the eigen value problem, by interior regularity
(cf. Theorem 4.3.7), φm ∈ H 2 (Br (x)). Arguing similarly, we obtain φm ∈
H k (Br (x)) for all k. Thus, by Sobolev imbedding results, φm ∈ C ∞ (Br (x)).
Since x ∈ Ω is arbitrary, φm ∈ C ∞ (Ω).
Remark 4.3.10. Observe that if H01 (Ω) is equipped with the inner product
R −1/2
Ω
∇u·∇v dx, then λm φm is an orthonormal basis for H01 (Ω) where (λm , φm )
is the eigen pair corresponding to A(x) = I and c ≡ 0. With the usual inner
product Z Z
uv dx + ∇u · ∇v dx
Ω Ω
in H01 (Ω), (λm + 1)−1/2 φm forms an orthonormal basis of H01 (Ω). The set
{φm } is dense in H01 (Ω) w.r.t both the norms mentioned above. Suppose
f ∈ H01 (Ω) is such that hf, φmRi = 0 in H01 (Ω), for all m. Then, from the
eigenvalue problem, we get λm Ω φm f dx = 0. Since φm is a basis for L2 (Ω),
f = 0.
Definition 4.3.11. The Rayleigh quotient map R : H01 (Ω) \ {0} → [0, ∞)
is defined as
R R
Ω
A(x)∇v · ∇v dx + Ω c(x)v 2 (x) dx
R(v) = .
kvk22,Ω
4.4 Periodic Boundary Conditions

Let Y = [0, 1)n be the unit cell of Rn and let, for each i, j = 1, 2, . . . , n,
aij : Y → R and A(y) = (aij ). For any given f : Y → R, extended Y -
periodically to Rn , we want to solve the problem

−div(A(y)∇u(y)) = f (y) in Y
(4.4.1)
u is Y − periodic.
The condition u is Y -periodic is equivalent to saying that u takes equal values
on opposite faces of Y . One may rewrite the equation on the n-dimensional
unit torus Tn without the periodic boundary condition.
∞
Let us now identify the solution space for (4.4.1). Let Cper (Y ) be the
∞ n 1
set of all Y -periodic functions in C (R ). Let Hper (Y ) denote the closure
∞
of Cper (Y ) in the H 1 -norm. Being a second order equation, in the weak
1
formulation, we expect the weak solution u to be in Hper (Y ). Note that if
u solves (4.4.1) then u + c, for any constant c, also solves (4.4.1). Thus, the
1
solution will be unique up to a constant in the space Hper (Y ). Therefore, we
1
define the quotient space Wper (Y ) = Hper (Y )/R as solution space where the
solution is unique.
Solving (4.4.1) is to find u ∈ Wper (Y ), for any given f ∈ (Wper (Y ))? in
the dual of Wper (Y ), such that
Z
A∇u · ∇v dx = hf, vi(Wper (Y ))? ,Wper (Y ) ∀v ∈ Wper (Y ).
Y
The requirement that f ∈ (Wper (Y ))? is equivalent to saying that

Z
f (y) dy = 0
Y
because f defines a linear functional on Wper (Y ) and f (0) = 0, where 0 ∈

1
Hper (Y )/R. In particular, the equivalence class of 0 is same as the equivalence
class 1 and hence Z
f (y) dy = hf, 1i = hf, 0i = 0.
Y
Theorem 4.4.1. Let Y be unit open cell and let aij ∈ L∞ (Ω) such that the
matrix A(y) = (aij (y)) is elliptic with ellipticity constant α > 0. For any
f ∈ (Wper (Y ))? , there is a unique weak solution u ∈ Wper (Y ) satisfying
Z
A∇u · ∇v dx = hf, vi(Wper (Y ))? ,Wper (Y ) ∀v ∈ Wper (Y ).
Y
Note that the solution u we find from above theorem is an equivalence

class of functions which are all possible solutions. Any representative element
from the equivalence class is a solution. All the elements in the equivalence
differ by a constant. Let u be an element from the equivalence class and let
c be the constant Z
1
c= u(y) dy.
|Y | Y
R
Thus, we have u−c is a solution with zero mean value in Y , i.e., Y u(y) dy =
0. Therefore, rephrasing (4.4.1) as

 −div(A(y)∇u(y)) = f (y) in Y
u is Y − periodic
1
R

|Y | Y
u(y) dy = 0
we have unique solution u in the solution space

Z
1 1
Vper (Y ) = u ∈ Hper (Y ) | u(y) dy = 0 .
|Y | Y
4.5 Coercive Neumann Problem

Let Ω be an open subset of Rn and let ∂Ω be the topological boundary of Ω.
The inhomogeneous Neumann problem, i.e., given f : Ω → R, g : ∂Ω → R,
aij ∈ L∞ (Ω) and c ∈ L∞ (Ω) such that c 6≡ 0 and c(x) ≥ 0 for a.e. x ∈ Ω,
find u : Ω → R such that

−div(A(x)∇u(x)) + c(x)u(x) = f (x) in Ω
(4.5.1)
A(x)∇u · ν = g on ∂Ω
where ν = (ν1 , . . . , νn ) is the unit outward normal. It is customary to denote

∂u
A∇u · ν as ∂ν A
. Note that the boundary condition imposed above is referred
to as the Neumann boundary condition.
Definition 4.5.1. We say u : Ω → R is a classical solution of the Neumann
problem (4.5.1), if u ∈ C 2 (Ω) and satisfies (4.5.1) pointwise for each x.
C 1 -smooth boundary (cf. Corollary ??). Then, multiplying any φ ∈ C 1 (Ω)
Z Z Z
− (∇ · (A∇u))φ dx + cuφ dx = f φ dx
Ω
Z Z Z Ω ZΩ
∂u
A∇u · ∇φ dx − φ dσ + c(x)uφ dx = f φ dx
Ω ∂Ω ∂νA Ω Ω
Z Z Z Z
A∇u · ∇φ dx + c(x)uφ dx = f φ dx + gφ dσ
Ω Ω Ω ∂Ω
∂u
(since = g on ∂Ω).
∂νA
Thus, any classical solution u of (4.5.1) solves the problem: For any given f ,
find u such that
Z Z Z Z
A∇u · ∇φ dx + c(x)uφ dx = f φ dx + gφ dσ ∀φ ∈ C 1 (Ω),
Ω Ω Ω ∂Ω
as long as the integrals make sense. Note that, in the above problem, it
is enough for u to be in C 1 (Ω). Also, in particular, any classical solution
satisfies the identity
Z Z Z Z
2
A∇u · ∇u dx + c(x)u dx = f u dx + gu dσ.
Ω Ω Ω ∂Ω
1 1
By the density of C (Ω) in H (Ω), the classical solution u will also solve
Z Z Z Z
A∇u · ∇v dx + c(x)uv dx = f v dx + gv dσ ∀v ∈ H 1 (Ω),
Ω Ω Ω ∂Ω
where v inside the boundary integral is in the trace sense, as long as the
integrals make sense. The integrals will make sense if f ∈ L2 (Ω) and g ∈
L2 (∂Ω). In fact, one can consider g ∈ H −1/2 (∂Ω), if we choose to replace
the boundary integral with the duality product h·, ·iH −1/2 (∂Ω),H 1/2 (∂Ω) . The
arguments above motivates the notion of weak solution for (4.5.1).
Definition 4.5.2. Given f ∈ L2 (Ω) and g ∈ H −1/2 (∂Ω), we say u ∈ H 1 (Ω)

is a weak solution of (4.5.1) if
Z Z Z
A∇u·∇v dx+ c(x)uv dx = f v dx+hg, viH −1/2 (∂Ω),H 1/2 (∂Ω) ∀v ∈ H 1 (Ω).
Ω Ω Ω
(4.5.2)
We have already observed that every classical solution is a weak solution
for a bounded open set Ω with C 1 -smooth boundary. Let M(α) denote the
space of all n × n matrix A = (aij ) such that aij ∈ L∞ (Ω) for all i, j and
α|ξ|2 ≤ A(x)ξ.ξ a.e. in x, ∀ξ = (ξi ) ∈ Rn .
Theorem 4.5.3. Let Ω be a bounded open subset of Rn , f ∈ L2 (Ω) and

g ∈ H −1/2 (∂Ω). Also, let A ∈ M(α) and γ > 0 be such that c(x) ≥ γ for a.e.
x ∈ Ω. Then there is a unique weak solution u ∈ H 1 (Ω) satisfying (4.5.2).
Moreover, if A is symmetric, then u minimizes the functional J : H 1 (Ω) → R
defined as,
Z Z Z
1 1 2
J(v) := A∇v · ∇v dx + cv dx − f v dx − hg, viH −1/2 (∂Ω),H 1/2 (∂Ω)
2 Ω 2 Ω Ω
in H 1 (Ω).
Proof. Recall that H 1 (Ω) is a Hilbert space endowed with the inner-product
Z Z
(v, w) := vw dx + ∇v∇w dx.
Ω Ω
We shall use the Lax-Milgram result (cf. Theorem ??). We define the map
a : H 1 (Ω) × H 1 (Ω) :→ R as
Z Z
a(v, w) := A∇v · ∇w dx + cvw dx.
Ω Ω
Consider,
Z Z
|a(v, w)| ≤ max kaij k∞ |∇v||∇w| dx + kck∞ |vw| dx
i,j Ω Ω
≤ C (k∇vk2 k∇wk2 + kvk2 kwk2 )

(where C := max max kaij k∞ , kck∞ )
i,j
≤ 2CkvkH 1 (Ω) kwkH 1 (Ω) .

Thus, a is a bilinear continuous form. Now, consider

Z Z
a(v, v) = A∇v · ∇v dx + cv 2 dx
Ω Ω
≥ αk∇vk22 + γkvk22
≥ min{α, γ}kvk2H 1 (Ω) .
Hence a is coercive, bilinear continuous form. Also, note that the map
Z
v 7→ f v dx + hg, viH −1/2 (∂Ω),H 1/2 (∂Ω)
Ω
is continuous linear functional on H 1 (Ω), since

Z
f v dx ≤ kf k2 kvk2 ≤ kf k2 kvkH 1 (Ω)
Ω
and the linearity and continuity of the trace map T . Thus, by Lax-milgram
theorem, there is a u ∈ H 1 (Ω) such that
Z
a(u, v) = f v dx + hg, viH −1/2 (∂Ω),H 1/2 (∂Ω) .
Ω
This is equivalent to (4.5.2) and if A is symmetric then u minimizes the

functional
Z Z Z
1 1 2
J(v) := A∇v · ∇v dx + cv dx − f v dx − hg, viH −1/2 (∂Ω),H 1/2 (∂Ω)
2 Ω 2 Ω Ω
in H 1 (Ω).
The term coercive refers to the fact that both A and c are positive definite.
In contrast to the Dirichlet boundary condition, observe that the Neumann
boundary condition is not imposed a priori, but gets itself imposed naturally.
In other words, if u is a weak solution of (4.5.1) and is in H 2 (Ω), then for
any v ∈ H 1 (Ω),
Z Z Z Z
A∇u · ∇v dx + c(x)uv dx = f v dx + gv dσ
Ω
Z Z ZΩ ZΩ Z∂Ω
∂u
− ∇ · (A∇u)v dx + v dσ + c(x)uv dx = f v dx + gv dσ
Ω ∂Ω ∂νA Ω Ω ∂Ω
implies that
Z
∂u
v −g dσ = 0 ∀v ∈ H 1 (Ω).
∂Ω ∂νA
But v |∂Ω ∈ H 1/2 (∂Ω), since v ∈ H 1 (Ω). Now, by the density of H 1/2 (∂Ω) in
∂u
L2 (∂Ω) (cf. Theorem 2.9.8), we have ∂ν A
= g in L2 (∂Ω). Thus, note that we
have obtained the Neumann boundary condition as a consequence of u being
a weak solution to the Neumann problem on the domain Ω. Thus, Dirichlet
conditions are called essential boundary condition and Neumann conditions
are called natural boundary conditions.
An astute reader should be wondering that if H01 (Ω) was the precise solu-
tion space for the Dirichlet problem (4.1.1), then the corresponding solution
space for the Neumann problem (4.5.1) should be the closure of
Vg := {φ ∈ D(Ω) | A∇φ · ν = g on ∂Ω}
in H 1 (Ω). The guess is right, solve the following exercise!
Exercise 70. H 1 (Ω) is the closure of Vg in H 1 (Ω)-norm.
4.6 Semi-coercive Neumann Problem

Ω. We consider the inhomogeneous Neumann problem with c ≡ 0 (thus,
the operator is semi-coercive), i.e., given f : Ω → R, g : ∂Ω → R and
aij ∈ L∞ (Ω), find u : Ω → R such that

−div(A(x)∇u(x)) = f (x) in Ω
(4.6.1)
A(x)∇u · ν = g on ∂Ω
where ν = (ν1 , . . . , νn ) is the unit outward normal.

In contrast to the coercive Neumann problem, (4.6.1) is not well posed.
This can be seen as follows:
C -smooth boundary (cf. Corollary ??). Then, multiplying any 1 ∈ C 1 (Ω)
1

Z Z
− (∇ · (A∇u)) dx = f dx
Ω Ω
Z Z
∂u
− dσ = f dx
∂Ω ∂νA Ω
Z Z
f dx + g dσ = 0
Ω ∂Ω
∂u
(since = g on ∂Ω).
∂νA
Thus, any classical solution u of (4.6.1) necessarily satisfies
Z Z
f dx + g dσ = 0.
Ω ∂Ω
This is called the compatibility condition for the semi-coercive Neumann

problem. We shall now see that the compatibility condition is also a suf-
ficient condition. Moreover, observe that if u is a solution of (4.6.1), then
for any constant C, u + C is also a solution of (4.6.1). Thus, one can expect
uniqueness only up R to a constant. In particular, if u is a solution
R of (4.6.1),
1
then w := u − |Ω| Ω
u dx is also a solution of (4.6.1) such that Ω
w dx = 0.
Let us introduce the space
Z
1
V := v ∈ H (Ω) | v dx = 0 .
Ω
Exercise 71. Show that V is a closed subspace of H 1 (Ω).

We expect the solution of (4.6.1) to be unique in V . We endow V with
the inner-product of H 1 (Ω), i.e.,
Z Z
(v, w) := vw dx + ∇v∇w dx.
Ω Ω
Definition 4.6.1. Let f ∈ L2 (Ω) and g ∈ H −1/2 (∂Ω) be such that

Z
f dx + hg, 1iH −1/2 (∂Ω),H 1/2 (∂Ω) = 0, (4.6.2)
Ω
then we say u ∈ V is a weak solution of (4.6.1) if

Z Z
A∇u · ∇v dx = f v dx + hg, viH −1/2 (∂Ω),H 1/2 (∂Ω) ∀v ∈ V. (4.6.3)
Ω Ω
As usual, it is easy to check that every classical solution is a weak solution

for a bounded open set Ω with C 1 -smooth boundary.
Theorem 4.6.2. Let Ω be a bounded, connected open subset of Rn with C 1
smooth boundary. Let f ∈ L2 (Ω) and g ∈ H −1/2 (∂Ω) be given such that
(4.6.2) is satisfied. Also, let A ∈ M(α), then there is a unique weak solution
u ∈ V satisfying (4.6.3). Moreover, if A is symmetric, then u minimizes the
functional J : V → R defined as,
Z Z
1
J(v) := A∇v · ∇v dx − f v dx − hg, viH −1/2 (∂Ω),H 1/2 (∂Ω) .
2 Ω Ω
Proof. We shall use the Lax-Milgram result (cf. Theorem ??). We define the
map a : V × V :→ R as
Z
a(v, w) := A∇v · ∇w dx.
Ω

Consider,
Z
|a(v, w)| ≤ max kaij k∞ |∇v||∇w| dx
i,j Ω
≤ max kaij k∞ k∇vk2 k∇wk2
i,j
≤ max kaij k∞ kvkH 1 (Ω) kwkH 1 (Ω) .
i,j
Thus, a is a bilinear continuous form. Consider

Z
a(v, v) = A∇v · ∇v dx
Ω
≥ αk∇vk22 .
Now, by (by Poincaré-Wirtinger, cf. Theorem 2.8.25) in V , there is a constant
C > 0 such that
kvk2 ≤ Ck∇vk2 ∀v ∈ V.
Thus,
a(v, v) ≥ αk∇vk22
α
≥ kvk22
C2
α
≥ kvk2H 1 (Ω) .
C2
Therefore, a is a coercive, bilinear continuous form on V . Also, note that

the map Z
v 7→ f v dx + hg, viH −1/2 (∂Ω),H 1/2 (∂Ω)
Ω
is continuous linear functional on H 1 (Ω), since

Z
f v dx ≤ kf k2 kvk2 ≤ kf k2 kvkH 1 (Ω)
Ω
and the linearity and continuity of the trace map T . Thus, by Lax-milgram
theorem, there is a u ∈ V such that
Z
a(u, v) = f v dx + hg, viH −1/2 (∂Ω),H 1/2 (∂Ω) .
Ω
This is equivalent to (4.6.3) and if A is symmetric then u minimizes the

functional
Z Z
1
J(v) := A∇v · ∇v dx − f v dx − hg, viH −1/2 (∂Ω),H 1/2 (∂Ω)
2 Ω Ω
in V .
The term semi-coercive refers to the fact that A is positive definite and
c is identically zero.
We introduce a different approach to prove the existence of weak solution
for the semi-coercive Neumann problem (4.6.1). This approach is called
the Tikhonov Regularization Method. The basic idea is to obtain the semi-
coercive case as a limiting case of the coercive Neumann problem.
Theorem 4.6.3. Let Ω be a bounded, connected open subset of Rn with C 1

smooth boundary. Let f ∈ L2 (Ω) and g ∈ H −1/2 (∂Ω) be given such that
(4.6.2) is satisfied. Let A ∈ M(α) and, for any ε > 0, if uε ∈ H 1 (Ω) is the
weak solution of (4.5.1) with c ≡ ε, i.e.,
Z Z Z
A∇uε ·∇v dx+ εuε v dx = f v dx+hg, viH −1/2 (∂Ω),H 1/2 (∂Ω) ∀v ∈ H 1 (Ω),
Ω Ω Ω
(4.6.4)
1
then uε → u weakly in H (Ω) where u ∈ V is the weak solution of (4.6.1),
as obtained in Theorem 4.6.2.
Proof. Firstly, note that by choosing v ≡ 1 in (4.6.4), we have

Z
uε dx = 0.
Ω
Thus, uε ∈ V for all ε. Now consider,

Z
2
αk∇uε k2 ≤ A∇uε · ∇uε dx
ZΩ
≤ A∇uε · ∇uε dx + εkuε k22

ZΩ
= f uε dx + hg, uε iH −1/2 (∂Ω),H 1/2 (∂Ω)
Ω
≤ kf k2 kuε k2 + kgkH −1/2 (∂Ω) kT uε kH 1/2 (∂Ω)

≤ C kf k2 + kgkH −1/2 (∂Ω) k∇uε k2
(Using Poincaré-Wirtinger and continuity of T )

k∇uε k2 ≤ C kf k2 + kgkH −1/2 (∂Ω)
Hence, again by Poincaré-Wirtinger inequality (cf. Theorem 2.8.25), {uε } is
a bounded sequence in H 1 (Ω), in fact a bounded sequence in V . Therefore,
one can extract a weakly convergence subsequence of {uε } (still denoted by
uε ) in V such that uε * u for some u ∈ V (since V is closed). Thus, passing
to the limit in (4.6.4), we get
Z Z
A∇u · ∇v dx = f v dx + hg, viH −1/2 (∂Ω),H 1/2 (∂Ω) ∀v ∈ V.
Ω Ω
Thus, u ∈ V is a weak solution of (4.6.1). Suppose u1 and u2 are two solutions

of (4.6.1), then w := u1 − u2 is a solution to (4.6.1) with f, g ≡ 0. Thus, w
is equal to some constant. Also, w ∈ V because both u1 and u2 are in V .
Hence, w = 0, since |Ω| 6= 0. Since the subsequence all converge to a unique
u, the entire sequence {uε } converges weakly to u in H 1 (Ω). Moreover, u is
the weak solution of (4.6.1).
Observe that by choosing v = uε in (4.6.4) and v = u in (4.6.3), and
passing to limit, we get
Z Z
lim A∇uε · ∇uε dx = f u dx + hg, uiH −1/2 (∂Ω),H 1/2 (∂Ω)
ε→0 Ω
ZΩ
= A∇u · ∇u dx.
Ω
Therefore, if A is identity matrix, then k∇uε k22 → k∇uk22 and hence uε

converges strongly to u in H 1 (Ω). In fact, in general, we have the identity
Z
2 1
lim sup k∇uε k2 ≤ A∇u · ∇u dx
ε→0 α Ω
maxi,j kaij k∞
≤ k∇uk22
α
maxi,j kaij k∞
≤ lim inf k∇uε k22 .
α ε→0
Corollary 4.6.4. If A is symmetric, then the u obtained in the above The-

orem minimizes the functional J : V → R defined as,
Z Z
1
J(v) := A∇v · ∇v dx − f v dx − hg, viH −1/2 (∂Ω),H 1/2 (∂Ω) .
2 Ω Ω
Proof. If A is symmetric, then, for each ε > 0, uε minimizes the functional

Jε : V → R defined as
Z Z Z
1 ε 2
Jε (v) := A∇v · ∇v dx + v dx − f v dx − hg, viH −1/2 (∂Ω),H 1/2 (∂Ω) .
2 Ω 2 Ω Ω
Equivalently, Jε (uε ) ≤ Jε (v) for all v ∈ V . It easy to note that Jε converges

to J pointwise. Now taking limits both sides, we get,
lim Jε (uε ) ≤ lim Jε (v) ∀v ∈ V

ε→0 ε→0
Z Z
1
lim A∇uε · ∇uε dx + ε u2ε dx −
ε→0 2
Z Ω Ω

lim f uε dx + hg, uε iH −1/2 (∂Ω),H 1/2 (∂Ω) ≤ J(v) ∀v ∈ V
ε→0 Ω
J(u) ≤ J(v) ∀v ∈ V.
Hence u minimizes J in V .
4.7 Mixed and Robin Problem

Ω. Let Γ1 ⊂ ∂Ω such that Γ = 6 ∅. Let Γ2 := ∂Ω \ Γ1 . We consider the
mixed Dirichlet-Neumann problem, i.e., given f : Ω → R, g : Γ2 → R and

aij ∈ L∞ (Ω), find u : Ω → R such that

 −div(A(x)∇u(x)) = f (x) in Ω
u =0 on Γ1 (4.7.1)
A(x)∇u · ν =g on Γ2

where ν = (ν1 , . . . , νn ) is the unit outward normal. Let Vm := {v ∈ H 1 (Ω) |

T v = 0 on Γ1 }.
Exercise 72. Show that Vm is a closed subspace of H 1 (Ω) endowed with the
H 1 (Ω)-norm. (Hint: use the continuity of the trace operator T ).
Definition 4.7.1. Given f ∈ L2 (Ω) and g ∈ H −1/2 (Γ2 ), we say u ∈ Vm is a

weak solution of (4.7.1) if
Z Z
A∇u · ∇v dx = f v dx + hg, viH −1/2 (Γ2 ),H 1/2 (Γ2 ) ∀v ∈ Vm . (4.7.2)
Ω Ω
Theorem 4.7.2. Let Ω be a bounded open subset of Rn , f ∈ L2 (Ω) and

g ∈ H −1/2 (Γ2 ). Also, let A ∈ M(α), then there is a unique weak solution
u ∈ Vm satisfying (4.7.2). Moreover, if A is symmetric, then u minimizes
the functional J : Vm → R defined as,
Z Z
1
J(v) := A∇v · ∇v dx − f v dx − hg, viH −1/2 (Γ2 ),H 1/2 (Γ2 ) .
2 Ω Ω
We skip the proof of above theorem because the arguments are similar
to those already introduced while studying homogeneous Dirichlet problem.
The Poincaré inequality is valid in Vm (since u vanishes on section of the
boundary, i.e., Γ1 ). Thus, all earlier arguments can be carried over smoothly.
Similarly, one can also study the mixed Dirichlet-Neumann problem with
inhomogeneous condition on Γ1 , say h. In this case the arguments are similar
to the Inhomogeneous Dirichlet problem with Vm replaced with its translate
Vm + h.
Let c : ∂Ω → R+ be given. We consider the Robin problem, i.e., given
f : Ω → R, g : ∂Ω → R and aij ∈ L∞ (Ω), find u : Ω → R such that

−div(A(x)∇u(x)) = f (x) in Ω
(4.7.3)
c(x)u + A(x)∇u · ν = g on ∂Ω
where ν = (ν1 , . . . , νn ) is the unit outward normal. Observe that the Robin
problem incorporates most of the problems introduced in this chapter. For
instance, if c ≡ 0, then we have the semi-coercive Neumann problem. If
c ≡ +∞, then (formally) we have the homogeneous Dirichlet problem. If
c ≡ +∞ on Γ1 ⊂ ∂Ω and c ≡ 0 in ∂Ω \ Γ1 , then we have the mixed Dirichlet-
Neumann problem.
Definition 4.7.3. Given f ∈ L2 (Ω) and g ∈ H −1/2 (∂Ω), we say u ∈ H 1 (Ω)
is a weak solution of (4.7.3) if
Z Z Z
A∇u·∇v dx+ cuv dσ = f v dx+hg, viH −1/2 (∂Ω),H 1/2 (∂Ω) ∀v ∈ H 1 (Ω).
Ω ∂Ω Ω
(4.7.4)
Theorem 4.7.4. Let Ω be a bounded, connected open subset of R with C 1 n
smooth boundary. Let f ∈ L2 (Ω), g ∈ H −1/2 (∂Ω) and A ∈ M(α). Also,

c(x) ≥ γ for a.e. x ∈ ∂Ω, for some γ > 0. Then there is a unique weak
solution u ∈ H 1 (Ω) satisfying (4.7.4). Moreover, if A is symmetric, then u
minimizes the functional J : H 1 (Ω) → R defined as,
Z Z Z
1 1 2
J(v) := A∇v · ∇v dx + cv dσ − f v dx − hg, viH −1/2 (∂Ω),H 1/2 (∂Ω) .
2 Ω 2 ∂Ω Ω
Proof. The only non-trivial part of the proof is to show the coercivity of the
bilinear map Z Z
a(v, w) := A∇v · ∇w dx + cvw dσ
Ω ∂Ω
in H 1 (Ω).
4.8 p-Laplacian Operator

Let 1 < p < +∞. We define the p-Laplacian operator ∆p , for 1 < p < +∞,
as
∆p v := div(|∇v|p−2 ∇v).
Note that when p = 2, we get the usual Laplacian operator, ∆. Let Ω be
an open subset of Rn and let ∂Ω be the topological boundary of Ω. Let
us first consider the Dirichlet problem for the p-Laplacian operator. Given
f : Ω → R, find u : Ω → R such that

−∆p u = f in Ω
(4.8.1)
u = 0 on ∂Ω.
Definition 4.8.1. Let q be the conjugate exponent of p. Given f ∈ W −1,q (Ω),

we say u ∈ W01,p (Ω) is a weak solution of (4.8.1) if
Z
|∇u|p−2 ∇u · ∇v dx = hf, viW −1,q (Ω),W 1,p (Ω) , ∀v ∈ W01,p (Ω). (4.8.2)
0
Ω
Theorem 4.8.2. Let Ω be a bounded open subset of Rn and f ∈ W −1,q (Ω).

Then there is a unique weak solution u ∈ W01,p (Ω) satisfying (4.8.2). More-
over, u minimizes the functional J : W01,p (Ω) → R defined as,
Z
1
J(v) := |∇v|p dx − hf, viW −1,q (Ω),W 1,p (Ω) .
p Ω 0
Proof. First prove J is coercive and continuous on W01,p (Ω), then use Theo-
rem ?? to show the existence of a minimizer u ∈ W01,p (Ω).
Secondly, show J is convex, in fact strictly convex on W01,p (Ω) and con-
sequently conclude the uniqueness of minimizer.
Lastly, show that J is Gâteaux differentiable and use (i) of Proposition ??
to show that the minimizer u is a weak solution of (4.8.1).
4.9 Stokes Problem

Let Ω be an open subset of Rn and let ∂Ω be the topological boundary of Ω.
The Stokes problem for a viscous fluid flow is described as follows: Given n
functions f1 , f2 , . . . , fn : Ω → R, find u1 , u2 , . . . , un : Ω → R and p : Ω → R
such that for each i = 1, 2, . . . , n,
∂p

−µ∆ui + ∂x = fi in Ω
i (4.9.1)
ui = 0 on ∂Ω,
where µ > 0 is the viscosity coefficient (it is a scaler inversely proportional to
Reynolds number). Observe that we have to find n+1 unknowns (ui ’s and p)
with n equations. The n + 1 equation is introduced as the incompressibility
condition of the fluid, i.e.,
∇ · u = 0, (4.9.2)
where u = (u1 , u2 , . . . , un ). Let f = (f1 , f2 , . . . , fn ). Thus, the Stokes system
can be written in the compressed form,

 −µ∆u + ∇p = f in Ω
div(u) = 0 in Ω (4.9.3)
u =0 on ∂Ω.

4.10 Elasticity System

Let Ω ⊂ R3 denote an elastic homogeneous with boundary ∂Ω = Γ1 ∪ Γ2
such that Γ1 ∩ Γ2 = ∅. Assume that the body is fixed along Γ2 and the
surface measure of Γ2 is strictly positive. Let f = (f1 , f2 , f3 ) denote the
force applied on the body Ω and g = (g1 , g2 , g3 ) denote the force applied
on the surface of the body Γ1 . Let u = (u1 , u2 , u3 ) denote the displacement
vector for each point of Ω. One can derive the strain (deformation) tensor
ij as, for 1 ≤ i, j ≤ 3,

1 ∂uj ∂ui
ij (u) = + .
2 ∂xi ∂xj
Note that the strain tensor is 3 × 3 symmetric matrix. Let σij (u) denote the
stress (measure of internal forces arising due to force f acting on Ω) tensor.
Then the constitutive law (Hooke’s law) states that
3
!
X
σij (u) = λ kk (u) δij + 2µij (u),
k=1
where λ ≥ 0 and µ > 0 are called Lame’s coefficients and δij is the Kronecker
delta defined as (
0 i 6= j
δij =
1 i = j.
Note that the stress tensor (σij ) is also a 3 × 3 symmetric matrix. Now, for
each i = 1, 2, 3, the displacement vector is given as a solution of the BVP
 P
3 ∂
 − j=1 ∂xj (σij (u)) = fi in Ω

P3
j=1 σij (u)νj = gi on Γ1 (4.10.1)

 ui = 0 on Γ2
4.11 Leray-Lions or Nonlinear Monotone Op-

erators
Definition 4.11.1. A function a : Ω×Rn → Rn is said to be a Carathéodory
function if a(·, ξ) is measurable on Ω for every ξ ∈ Rn and a(x, ·) is contin-
uous on Rn for almost every x ∈ Ω.
Let a(x, ξ) be a Carathéodory function satisfying the following hypotheses

for almost every x ∈ Ω:
H 1 (Coercive) There is a constant α > 0 such that a(x, ξ).ξ ≥ α|ξ|p , for
every ξ ∈ Rn .
H 2 (Bounded) There is a constant β > 0 and a non-negative function h ∈

Lp (Ω) such that |a(x, ξ)| ≤ β (h(x) + |ξ|p−1 ) for every ξ ∈ Rn .
H 3 (Monotone) (a(x, ξ) − a(x, ξ 0 )) .(ξ − ξ 0 ) > 0 for every ξ, ξ 0 ∈ Rn and

ξ 6= ξ 0 .
Let H(α, β, h) denote the set of all Carathéodory functions satisfying the
above three hypotheses. For every coercive (hypotheis H1) Carathéodory
function, a(x, 0) = 0 for almost every x ∈ Ω.
Definition 4.11.2. Given f ∈ W −1,q (Ω), we say u ∈ W01,p (Ω) is a weak

solution of

−div(a(x, ∇u)) = f in Ω
(4.11.1)
u = 0 on ∂Ω
whenever
Z
a(x, ∇u) · ∇φ dx = hf, φiW −1,q (Ω),W 1,p (Ω) , ∀φ ∈ W01,p (Ω).
0
Ω
The map u 7→ −div(a(x, ∇u)) is a coercive, continuous, bounded and

monotone operator from W01,p (Ω) to W −1,q (Ω). The existence and uniqueness
of solution is a standard theory of monotone operators ([LL65, Lio69]).
4.12 Solutions for Measure data

If p > n then, by Theorem 2.8.26, B(Ω) is compactly imbedded in W −1,q (Ω)
n
for all q ∈ [1, n−1 ). Hence, for p > n the notion of weak solution carries
forward to measure data, as well. However, for p < n, one cannot expect the
solution to belong to W01,p (Ω). For example, consider the Laplace equation
in a ball, with µ as the Dirac mass at the centre. existence and uniqueness
of sul
4.12.1 Stampacchia Solutions (p = 2 and n > 2)

For n > 2, we introduce the notion of solution for measure data introduced
by G. Stampacchia in [Sta65, Definition 9.1].
Definition 4.12.1. Given µ ∈ B(Ω) with finite variation, we say u ∈ L1 (Ω)

is a stampacchia solution of

−div(A(x)∇u) = µ in Ω
(4.12.1)
u = 0 on ∂Ω
whenever Z Z
uf dx = v dµ, ∀f ∈ L∞ (Ω),
Ω Ω
where v solves
−div(At (x)∇v) = f in Ω
(4.12.2)
v =0 on ∂Ω
The existence of v follows from Definition 4.3.3 and, by classical regularity

results, v is in the class of Hölder continuous functions. The existence and
uniqueness of stampacchia solution was shown in [Sta65, Theorem 9.1]. Also,
one has
n
kukW 1,r (Ω) ≤ C0 |µ|, ∀r ∈ 1,
0 n−1
where the constant C0 depends on n, α and Ω.
4.12.2 Renormalized Solutions (1 ≤ p < n)

For every k > 0, we define the truncation function (truncation at height k)
Tk : R → R by (
s if |s| ≤ k
Tk (s) = ks
|s|
if |s| ≥ k.
An equivalent way of stating the truncation at height k is Tk (s) = max(−k, min(k, s)).
Let T01,p (Ω) be the set all measurable functions u : Ω → R on Ω and
finite almost everywhere in Ω, such that Tk (u) ∈ W01,p (Ω) for every k > 0.
Every function u ∈ T01,p (Ω) will be identified with its p-quasi continuous
representative.
Lemma 4.12.2 (Lemma 2.1 of [BBG+ 95]). For every u ∈ T01,p (Ω), there
exists a measurable function v : Ω → Rn such that ∇Tk (u) = vχ{|u|≤k} almost
everywhere in Ω, for every k > 0. Also v is unique up to almost everywhere
equivalence.
Thus, by above lemma, one can define a generalised gradient ∇u of u ∈

T01,p (Ω)by setting ∇u = v. If u ∈ W 1,1 (Ω), then this gradient coincides
with the usual one. However, for u ∈ L1loc (Ω), then it may differ from the
distributional gradient of u (cf. Remark 2.9 of [DMMOP99]).
Let Lip0 (R) be the set of all Lipschitz continuous functions l : R → R
whose derivative l0 has compact support. Thus, every function l ∈ Lip0 (R)
is constant outside the support of its derivative, so that we can define the
constants l(+∞) = lims→+∞ l(s) and l(−∞) = lims→−∞ l(s).
Appendices
177
Appendix A
Cantor Set and Cantor

Function
Let us construct the Cantor set which plays a special role in analysis.
Consider C0 = [0, 1] and trisect C0 and remove the middle open interval
to get C1 . Thus, C1 = [0, 1/3] ∪ [2/3, 1]. Repeat the procedure for each
interval in C1 , we get
C2 = [0, 1/9] ∪ [2/9, 1/3] ∪ [2/3, 7/9] ∪ [8/9, 1].
Repeating this procedure at each stage, we get a sequence of subsets Ci ⊆

[0, 1], for i = 0, 1, 2, . . .. Note that each Ck is a compact subset, since it is a
finite union of compact sets. Moreover,
C0 ⊃ C1 ⊃ C2 ⊃ . . . ⊃ Ci ⊃ Ci+1 ⊃ . . . .
The Cantor set C is the intersection of all the nested Ci ’s, C = ∩∞

i=0 Ci .
Lemma A.0.1. C is compact.

Proof. C is countable intersection of closed sets and hence is closed. C ⊂
[0, 1] and hence is bounded. Thus, C is compact.
The Cantor set C is non-empty, because the end-points of the closed
intervals in Ci , for each i = 0, 1, 2, . . ., belong to C. In fact, the Cantor set
cannot contain any interval of positive length.
Lemma A.0.2. For any x, y ∈ C, there is a z ∈
/ C such that x < z < y.
(Disconnected)
179
APPENDIX A. CANTOR SET AND CANTOR FUNCTION 180
Proof. If x, y ∈ C are such that z ∈ C for all z ∈ (x, y), then we have the
open interval (x, y) ⊂ C. It is always possible to find i, j such that

j j+1
, ⊆ (x, y)
3i 3i
but does not belong Ci ⊃ C.
We show in example ??, that C has length zero. Since C is non-empty,

how ‘big’ is C? The number of end-points sitting in C are countable. But C
has points other than the end-points of the closed intervals Ci for all i. For
instance, 1/4 (not an end-point) will never belong to the the intervals being
removed at every step i, hence is in C. There are more! 3/4 and 1/13 are
all in C which are not end-points of removed intervals. It is easy to observe
these by considering the ternary expansion characterisation of C. Consider
the ternary expansion of every x ∈ [0, 1],
∞
X ai
x= = 0.a1 a2 a3 . . .3 where ai = 0, 1 or 2.
i=1
3i
The decomposition of x in ternary form is not unique1 . For instance, 1/3 =

0.13 = 0.022222 . . .3 , 2/3 = 0.23 = 0.1222 . . .3 and 1 = 0.222 . . .3 . At the C1
stage, while removing the open interval (1/3, 2/3), we are removing all num-
bers whose first digit in ternary expansion (in all possible representations)
is 1. Thus, C1 has all those numbers in [0, 1] whose first digit in ternary
expansion is not 1. Carrying forward this argument, we see that for each i,
Ci contains all those numbers in [0, 1] with digits upto ith place, in ternary
expansion, is not 1. Thus, for any x ∈ C,
∞
X ai
x= = 0.a1 a2 a3 . . .3 where ai = 0, 2.
i=1
3i
Lemma A.0.3. C is uncountable.
Proof. Use Cantor’s diagonal argument to show that the set of all sequences
containing 0 and 2 is uncountable.
1
This is true for any positional system. For instance, 1 = 0.99999 . . . in decimal system
Cantor Function
We shall now define the Cantor function fC : C → [0, 1] as,
∞
! ∞
X ai X ai −i
fC (x) = fC i
= 2 .
i=1
3 i=1
2
Since ai = 0 or 2, the function replaces all 2 occurrences with 1 in the

ternary expansion and we interpret the resulting number in binary system.
Note, however, that the Cantor function fC is not injective. For instance,
one of the representation of 1/3 is 0.0222 . . .3 and 2/3 is 0.2. Under fC
they are mapped to 0.0111 . . .2 and 0.12 , respectively, which are different
representations of the same point. Since fC is same on the end-points of the
removed interval, we can extend fC to [0, 1] by making it constant along the
removed intervals.
Alternately, one can construct the Cantor function step-by-step as we
remove middle open intervals to get Ci . Consider f1 to be a function which
takes the constant value 1/2 in the removed interval (1/3, 2/3) and is linear
on the remaining two intervals such that f1 is continuous. In the second
stage, the function f2 coincides with f1 in 1/3, 2/3, takes the constant value
1/4 and 3/4 on the two removed intervals and is linear in the remaining
four intervals such that f2 is continuous. Proceeding this way we have a
sequence of monotonically increasing continuous functions fk : [0, 1] → [0, 1].
Moreover, |fk+1 (x)−fk (x)| < 2−k for all x ∈ [0, 1] and fk converges uniformly
to fC : [0, 1] → [0, 1].
Exercise 73. The Cantor function fC : [0, 1] → [0, 1] is uniformly continuous,
monotonically increasing and is differentiable a.e. and fC0 = 0 a.e.
Exercise 74. The function fC is not absolutely continuous.
Generalised Cantor Set

We generalise the idea behind the construction of Cantor sets to build Cantor-
like subsets of [0, 1]. Choose a sequence {ak } such that ak ∈ (0, 1/2) for all
k. In the first step we remove the open interval (a1 , 1 − a1 ) from [0, 1] to get
C1 . Hence C1 = [0, a1 ] ∪ [1 − a1 , 1]. Let
C11 := [0, a1 ] and C12 := [1 − a1 , 1].

Hence, C1 = C11 ∪ C12 . Note that C1i are sets of length a1 carved out from
the end-points of C0 . We repeat step one for each of the end-points of C1i of
length a1 a2 . Therefore, we get four sets
C21 := [0, a1 a2 ] C22 := [a1 − a1 a2 , a1 ],
C23 := [1 − a1 , 1 − a1 + a1 a2 ] C24 := [1 − a1 a2 , 1].
Define C2 = ∪4i=1 C2i . Each C2i is of length a1 a2 . Note that a1 a2 < a1 .
Repeating the procedure successively for each term in the sequence {ak },
we get a sequence of sets Ck ⊂ [0, 1] whose length is 2k a1 a2 . . . ak . The
“generalised” Cantor set C is the intersection of all the nested Ck ’s, C =
2k
∩∞ i
k=0 Ck and each Ck = ∪i=1 Ck . Note that by choosing the constant sequence
ak = 1/3 for all k gives the Cantor set defined in the beginning of this
Appendix. Similar arguments show that the generalised Cantor set C is
compact. Moreover, C is non-empty, because the end-points of the closed
intervals in Ck , for each k = 0, 1, 2, . . ., belong to C.
Lemma A.0.4. For any x, y ∈ C, there is a z ∈
/ C such that x < z < y.
Lemma A.0.5. C is uncountable.
We show in example ??, that C has length 2k a1 a2 . . . ak .
The interesting fact about generalised Cantor set is that it can have non-
zero “length”.
Proposition A.0.6. For each α ∈ [0, 1) there is a sequence {ak } ⊂ (0, 1/2)
such that
lim 2k a1 a2 . . . ak = α.
k
Proof. Choose a1 ∈ (0, 1/2) such that 0 < 2a1 −α < 1. Use similar arguments
to choose ak ∈ (0, 1/2) such that 0 < 2k a1 a2 . . . ak − α < 1/k.
Generalised Cantor Function

We shall define the generalised Cantor function fC on the generalised Cantor
set C. Define the function f0 : [0, 1] → [0, 1] as f0 (x) = x. f0 is continuous
on [0, 1]. We define f1 : [0, 1] → [0, 1] such that f is linear and continuous
on C1i , and 1/2 on [a1 , 1 − a1 ], the closure of removed open interval at first
stage. We define fk : [0, 1] → [0, 1] continuous fk (0) = 0, fk (1) = 1 such that
fk (x) = i/2k on the removed interval immediate right to Cki .
Theorem A.0.7. Each fk is continuous, monotonically non-decreasing and

uniformly converges to some fC : [0, 1] → [0, 1].
Thus, fC being uniform limit of continuous function is continuous and is

the called the generalised Cantor function.
Bibliography
[Ada75] Robert A. Adams, Sobolev spaces, Academic Press, Boston,

1975.
[BBG+ 95] P Bénilan, L Boccardo, T Gallouët, R Gariepy, M Pierre, and

J. L. Vazquez, An L1 -theory of existence and uniqueness of
solutions of nonlinear elliptic equations, Ann. Scuola Normale
Sup. Pisa Cl. Sci. 22 (1995), 241–273.
[CD44] G. Choquet and J. Deny, Sur quelques propriétés de

moyenne caractéristiques des fonctions harmoniques et poly-
harmoniques, Bull. Soc. Math. de France 72 (1944), 118–140.
[Dir49] P. A. M. Dirac, The principles of quantum mechanics, 3 ed.,

University Press, Oxford, 1949.
[DMMOP99] Gianni Dal Maso, F. Murat, L. Orsina, and A. Prignet, Renor-

malized solutions for elliptic equations with general measure
data, Ann. Scuola Normale Sup. Pisa Cl. Sci. 28 (1999), 741–
808.
[Hor70] John Horvath, An introduction to distributions, American

Mathematical Monthly 77 (1970), 227–240.
[Kes89] S. Kesavan, Topics in functional analysis and applications,

New Age International, 1989, New edition/reprint might be
avalaible.
[Lio69] J. L. Lions, Quelques méthodes de résolution des problémes aux

limites non linéaires, Dunod et Gauthier-Villars, Paris, 1969.
185
BIBLIOGRAPHY 186
[LL65] J. Leray and J. L. Lions, Quelques résultats de visik sur les

problémes elliptiques non linéaires par les méthodes de Minty-
Browder, Bull. Soc. Math. France 93 (1965), 97–107.
[Maz85] Vladimir Maz’ya, Sobolev spaces, Springer-Verlag, 1985.
[MS64] N. Meyers and J. Serrin, h = w, Proc. Nat. Acad. Sci., USA

51 (1964), 1055–1056.
[Rud91] Walter Rudin, Functional analysis, second ed., McGraw-Hill

Book Company, 1991.
[Sch45] L. Schwartz, Généralisation de la notion de fonction, de

transformation de fourier et applications mathématiques et
physiques, Ann. Univ. Grenoble 21 (1945), 57–74.
[Sta65] Guido Stampacchia, Le problème de Dirichlet pour les

équations elliptiques du second ordre à coefficients discontinus,
Annales de l’institut Fourier 15 (1965), no. 1, 189–257.
[Tre70] F. Treves, Applications of distributions to pde theory, American

Mathematical Monthly 77 (1970), 241–248.
Index
σ-algebra, 2 Neumann Problem, 160, 164

σ-compact, 12 Non-divergence form operator, 152
p-Laplacian, 171
Robin Problem, 170
Coercive operator, 152
Schwartz, 2
Dirac, 1, 2 sequence
delta function, 2 mollifiers, 10
measure, 2 solution
sequence, 11 fundamental, 75
Dirac sequence, 40 Stokes Problem, 172
Dirichlet boundary condition, 147
Dirichlet Problem, 147
inhomogeneous, 150
Divergence form operator, 152
Elliptic operator, 151, 152
function
Cauchy’s exponential, 8
Dirac delta, 2
Heaviside, 1
test, 19
hypoelliptic operator, 75
inequality
Young’s, 20
Lebesgue, 2
Mixed Dirichlet-Neumann Problem, 170
Neumann boundary condition, 161
187

Sobolev spaces and applications by IITK

Uploaded by

Sobolev spaces and applications by IITK

Uploaded by

Sobolev Spaces and Applications

December 17, 2023

1.5.5 Shifting and Scaling . . . . . . . . . . . . . . . . . . . 68

3 Bounded Variation Functions 143

4 Second Order Elliptic PDE 147

4.11 Leray-Lions or Nonlinear Monotone Operators . . . . . . . . . 173

A Cantor Set and Cantor Function 179

Ω denotes an open subset of Rn , not necessarily bounded

R denotes the real line

Rn denotes the n-dimensional Euclidean space

|E| will denote the Lebesgue measure of E ⊂ Rn

Lip(E) denotes the space of all Lipschitz functions on E

C(Ω) is the class of all continuous functions on Ω

C k (Ω) is the class of all k-times (k ≥ 1) continuously differentiable functions

C0 (Ω) is the class of all continuous functions on Ω that vanishes at boundary

H k (Ω) is the space W k,2 (Ω)

H0k (Ω) is the space W0k,2 (Ω)

E(Ω) is the class of all infinitely differentiable functions on Ω, also denoted

D(Ω) is the class of all infinitely differentiable functions on Ω with compact

α|ξ|2 ≤ A(x)ξ.ξ a.e. x ∀ξ ∈ Rn

4. δ is infinitely differentiable and for

5. Given the Heaviside function

then the delta function is the “derivative” of H, i.e., δ(x) = H 0 (x).

consider F to be the Heaviside function H (cf. (1.1.2)) and G ≡ 0 or vicev-

and the corresponding characteristic equations:

dx(r, s) dt(r, s) dz(r, s)

x(r, 0) = r, t(r, 0) = 0, and z(r, 0) = φ(r).

Solving the ODE corresponding to z, we get z(r, s) = c3 (r) with initial

x(r, s) = φ(r)s + c1 (r), t(r, s) = s + c2 (r)

with initial conditions

Suppose we choose φ to be the function

Then the characteristic curves are

This situation is, usually, referred to as shock waves.

1.3 Space of Test Functions

There are no boundary integrals above because φ has compact support in

and the derivative of f , f 0 , can be identified with the mapping φ mapped

1.3.1 Smooth Non-Analytic (Bump) Functions

Example 1.1. Consider the non-negative function f0 : (−∞, 0] → [0, 1] de-

Example 1.3. Another variant is the function f2 : R → R defined as

It is clear that 0 ≤ fi (x) < 1 and fi is infinitely differentiable for all x 6= 0

f (h) − f (0) exp(−1/h)

f (k) (h) − f (k) (0) Pk−1 (h)exp(−1/h)

Therefore, f ∈ C ∞ (R). We shall now observe that f is not analytic at x = 0.

is infinitely differentiable but not analytic.

is infinitely differentiable but not analytic.

Proof. Choose f = fi , for i = 2 or i = 4 in the above examples. This

Also, using the function

one defines the function

previous section. For ε > 0, we set b = ε in g5b (or g2b ) and

Definition 1.3.5. A sequence of functions {ρk }, say on Rn , is said to be a

(i) ρk ≥ 0 for all k.

Definition 1.3.6. An approximate identity is a sequence (or net) {ρk in a

1.3.3 Topology on C(Ω), C0 (Ω) and Cc (Ω)

kφk∞,Ω := max |φ(x)|

Note that p0 ≤ p1 ≤ p2 ≤ . . . . The sets

{φ ∈ C(Ω) | pj (φ) < 1/j}

1.3.4 Algebraic and Topological Dual of Cc (Ω)

Note that given a measure µ, the map

defines a positive, linear functional on Cc (Ω).

Theorem 1.3.9. Let T be a positive, linear functional on Cc (Ω). Then there

This association T 7→ µ defines an isometry, i.e., kT k = |µ|.

Theorem 1.3.14 (Riesz-Alexandrov). Consider C0 (Ω) endowed with the

This association T 7→ µ defines an isometry, i.e., kT k = |µ|.

Theorem 1.3.15 (Riesz-Markov). Consider Cc (Ω) endowed with the induc-

This association T 7→ µ defines an isometry, i.e., kT k = |µ|.

1.3.5 Topology on C ∞ (Ω)

of k-times differentiable functions on Ω, C k (Ω), is a Banach space w.r.t the

For any open subset Ω of Rn , there is a sequence Kj of non-empty compact

Again, as before, the family of semi-norms induces a locally convex complete

1.3.6 Inductive Limit Topology on Cc∞ (Ω)

ker(Tx ) := {φ ∈ C ∞ (Ω) | Tx (φ) = 0},

is a closed subspace of C ∞ (Ω). Note that ker(Tx ) is precisely those φ ∈

Proposition 1.3.17. A sequence of functions {φm } ⊂ D(Ω) converges to

1.3.7 Regularization and Cut-off Technique

Theorem 1.3.19 (cf. Theorem ??). For 1 ≤ p < ∞, Cc (Ω) is dense in

Definition 1.3.20. Let f, g ∈ L1 (Rn ). The convolution f ∗ g is defined as,